Non commutative Khintchine inequalities and Grothendieck s theo

Similar documents
Preliminaries on von Neumann algebras and operator spaces. Magdalena Musat University of Copenhagen. Copenhagen, January 25, 2010

Magdalena Musat University of Memphis

On Some Local Operator Space Properties

Lecture 1 Operator spaces and their duality. David Blecher, University of Houston

arxiv:math/ v1 [math.fa] 14 May 2005

Finite-dimensional spaces. C n is the space of n-tuples x = (x 1,..., x n ) of complex numbers. It is a Hilbert space with the inner product

Operator Space Approximation Properties for Group C*-algebras

Chapter 8 Integral Operators

von Neumann algebras, II 1 factors, and their subfactors V.S. Sunder (IMSc, Chennai)

Functional Analysis I

MAT 578 FUNCTIONAL ANALYSIS EXERCISES

On positive maps in quantum information.

p-operator Spaces Zhong-Jin Ruan University of Illinois at Urbana-Champaign GPOTS 2008 June 18-22, 2008

INF-SUP CONDITION FOR OPERATOR EQUATIONS

THE ALTERNATIVE DUNFORD-PETTIS PROPERTY FOR SUBSPACES OF THE COMPACT OPERATORS

Introduction. In C -algebra theory, the minimal and maximal tensor products (denoted by A 1 min A 2 and A 1 max A 2 ) of two C -algebras A 1 ; A 2, pl

Functional Analysis HW #1

The numerical index of Banach spaces

Norms of Elementary Operators

Where is matrix multiplication locally open?

Functional Analysis. Martin Brokate. 1 Normed Spaces 2. 2 Hilbert Spaces The Principle of Uniform Boundedness 32

Spectral theory for compact operators on Banach spaces

FREE PROBABILITY THEORY

1 Math 241A-B Homework Problem List for F2015 and W2016

Outline. Multipliers and the Fourier algebra. Centralisers continued. Centralisers. Matthew Daws. July 2009

Diffun2, Fredholm Operators

Recall that any inner product space V has an associated norm defined by

COUNTEREXAMPLES TO THE COARSE BAUM-CONNES CONJECTURE. Nigel Higson. Unpublished Note, 1999

ORDERED INVOLUTIVE OPERATOR SPACES

The numerical index of Banach spaces

Grothendieck s Theorem, past and present UNCUT updated and still expanding...

Fréchet differentiability of the norm of L p -spaces associated with arbitrary von Neumann algebras

Spectral Continuity Properties of Graph Laplacians

Boundedly complete weak-cauchy basic sequences in Banach spaces with the PCP

David Hilbert was old and partly deaf in the nineteen thirties. Yet being a diligent

SEPARABLE LIFTING PROPERTY AND EXTENSIONS OF LOCAL REFLEXIVITY

I teach myself... Hilbert spaces

A COMMENT ON FREE GROUP FACTORS

Multi-normed spaces and multi-banach algebras. H. G. Dales. Leeds Semester

Functional Analysis. Franck Sueur Metric spaces Definitions Completeness Compactness Separability...

The projectivity of C -algebras and the topology of their spectra

Overview of normed linear spaces

Math 350 Fall 2011 Notes about inner product spaces. In this notes we state and prove some important properties of inner product spaces.

Mixed q-gaussian algebras and free transport

arxiv:math/ v1 [math.fa] 24 Aug 1993

Acta Mathematica Academiae Paedagogicae Nyíregyháziensis 24 (2008), ISSN

What are operator spaces?

Banach Spaces II: Elementary Banach Space Theory

C.6 Adjoints for Operators on Hilbert Spaces

Grothendieck s works on Banach spaces. their surprising recent repercussions (parts 1 and 2)

SOME BANACH SPACE GEOMETRY

CLOSED RANGE POSITIVE OPERATORS ON BANACH SPACES

Measurable functions are approximately nice, even if look terrible.

arxiv:math.fa/ v1 2 Oct 1996

4 Hilbert spaces. The proof of the Hilbert basis theorem is not mathematics, it is theology. Camille Jordan

Mathematical foundations - linear algebra

SYMPLECTIC GEOMETRY: LECTURE 1

BERNARD RUSSO. University of California, Irvine

Math 123 Homework Assignment #2 Due Monday, April 21, 2008

A note on the σ-algebra of cylinder sets and all that

Combinatorics in Banach space theory Lecture 12

Lipschitz p-convex and q-concave maps

Elliott s program and descriptive set theory I

Spectrally Bounded Operators on Simple C*-Algebras, II

Spectral theorems for bounded self-adjoint operators on a Hilbert space

Functional Analysis (2006) Homework assignment 2

PROBLEMS. (b) (Polarization Identity) Show that in any inner product space

Functional Analysis II held by Prof. Dr. Moritz Weber in summer 18

Morita equivalence of dual operator algebras

Analysis Preliminary Exam Workshop: Hilbert Spaces

ON SPECTRA OF LÜDERS OPERATIONS

Gilles Pisier. 1 The Theory of Operator Spaces

Lecture 5: Peak sets for operator algebras and quantum cardinals. David Blecher. December 2016

Spectral Theory, with an Introduction to Operator Means. William L. Green

Upper triangular forms for some classes of infinite dimensional operators

2.2 Annihilators, complemented subspaces

MAT 570 REAL ANALYSIS LECTURE NOTES. Contents. 1. Sets Functions Countability Axiom of choice Equivalence relations 9

arxiv:math/ v1 [math.oa] 4 Jan 2007 JAN M. CAMERON

SOME APPROXIMATION PROPERTIES OF BANACH SPACES AND BANACH LATTICES

The Polynomial Numerical Index of L p (µ)

Subsequences of frames

On Riesz-Fischer sequences and lower frame bounds

arxiv:math/ v1 [math.fa] 29 Apr 1999

BERNARD RUSSO. University of California, Irvine

COMPACT OPERATORS. 1. Definitions

ABELIAN SELF-COMMUTATORS IN FINITE FACTORS

Elementary linear algebra

Lecture 3. Random Fourier measurements

The Real Grassmannian Gr(2, 4)

A non-commutative notion of separate continuity

Lecture Notes 1: Vector spaces

Extensions of Lipschitz functions and Grothendieck s bounded approximation property

Decomposition of Riesz frames and wavelets into a finite union of linearly independent sets

COMMON COMPLEMENTS OF TWO SUBSPACES OF A HILBERT SPACE

GENERALIZED COVARIATION FOR BANACH SPACE VALUED PROCESSES, ITÔ FORMULA AND APPLICATIONS

6 Classical dualities and reflexivity

A 3 3 DILATION COUNTEREXAMPLE

Banach spaces without local unconditional structure

SPECTRAL THEORY EVAN JENKINS

Mathematical foundations - linear algebra

Transcription:

Non commutative Khintchine inequalities and Grothendieck s theorem Nankai, 2007

Plan Non-commutative Khintchine inequalities 1 Non-commutative Khintchine inequalities 2

µ = Uniform probability on the set { 1, 1} IN Notation ε = (ε n ) { 1, 1} IN Classical Khintchine inequalities : For any 0 < p < there are constants A p > 0 and B p > 0 such that for any sequence x = (x n ) in l 2 we have A p ( xn 2) 1/2 ( xn ε n p dµ(ε) ) 1/p B p ( xn 2) 1/2. (1)

If B is a Banach space and x k B, the square function must be replaced by ( x n 2 ) 1/2 ( sup ) n ε 2 1/2 kx k dµ (2) n D 0 B where D = { 1, 1} IN equipped with its usual probability µ and where ε n = D { 1, 1} denotes the n-th coordinate. When B is a Hilbert space, for all x k in B we have n 0 ε kx k 2 = n 0 x k 2 and hence we recover the square function, but in general this is not possible and we must work with (2).

Theorem (Kahane) For any 0 < p < q < there is a constant K (p, q) such that for any Banach space B and any finite subset x 1,..., x n in B we have εk x k Lp(B) εk x k Lq(B) K (p, q) εk x k Lp(B). In particular ε k x k L2 (B) is equivalent to ε k x k Lp(B) for any 0 < p <.

We will concentrate on the very important example of B-valued martingales of the form M n = n 0 ε kx k with B = S p or more generally B = L p (M, τ) when (M, τ) is a semi-finite non-commutative measure space (due mainly to Lust-Piquard). This had a major influence on all the subsequent developments of non-commutative martingale theory.

Theorem Let (M, τ) be a semi-finite generalized measure space, let 2 p < and let (x n ) be a finite sequence in L p (τ). There are constants A p > 0 and B p > 0 such that A p 1/2 ( x n x n + x n xn) L p(τ) In particular, if x n = xn we have ( xn 2 ) ( 1/2 p 2 A p ( IE p ε n x n B p L p(τ) ) 1/p 1/2 ( x n x n + x n xn) IE ε n x n p L p(τ)) 1/p 2 B p L p(τ) ( xn 2 ) 1/2 p..

The case p < 2 is different from that of p > 2. We emphasize that even if x n = x n the expression defined as [(x n )] p which replaces ( x 2 n ) 1/2 p cannot be simplified significantly. Theorem Now assume 1 p 2. For any finite sequence (x n ) in L p (τ) we define { ) } 1/2 [(x n )] p = inf ( b 1/2 n b n + ( cn cn) where the infimum runs over all possible decompositions x n = b n + c n with b n, c n in L p (τ). Then α > 0 such that ( α [(x n )] p IE p ε n x n p L p(τ) ) 1/p [(x n )] p. (3) p

The dual norm to the norm [(x n )] p is NOT [(x n )] p it is the following norm y = (y n ), y n L p (τ) { ) 1/2 p [y] p = max ( yn yn ) 1/2, ( y n y n We will concentrate on case p = 1, p = so we define { ( ) 1/2 y RC = max yn yn ) } 1/2, ( y n y n. p }.

Proof of Non-com Khintchine for p = 1 Very recently (to appear in JFA) Haagerup and Musat found a short proof of the case p = 1 which gives optimal constant for the complex Gaussian case. We will sketch this proof for (ε n ) for which they obtain the constant 3, i.e. [(x n )] 1 L1 3IE εn x n. (τ)

Recall B non-commutative L 1, M = B von Neumann algebra. By duality the following are equivalent (for a fixed constant C) Indeed : x B N y M N [(x n )] 1 C ε n x n L1 (B) inf ϕ [εn] ε n y n + ϕ L (M) C y RC [(x n )] 1 and y RC are dual to each other ε n x n L1 (B) and inf ϕ [ε n] ε n y n + ϕ L (M) are dual to each other

Assume B non-commutative L 1, i.e. M = B is a von Neumann algebra. Key Lemma Let y = (y n ) 1 n N ; y n M. Assume y RC 1, where : Then, ŷ y RC = max{ y n y n 1/2, y n y n 1/2 }. [ε n ] = {ϕ L (M) ϕ [ε n ] such that ε n ŷ n + ϕ L (M) 3/2 y ŷ RC 1/2. ϕε n = 0, n}

Key Lemma implies C 3 Indeed, we have C 3/2 + C/2 and hence C/2 3/2.

Proof of Key Lemma (with constant 4 3 for short) : Can assume y n = y n. Let S = ε n y n. Let S t = S1 S t. Clearly S t L (M) t Moreover, if ŷ n = ε n S t we may write a priori S t = ε n ŷ n + ϕ with ϕ [ε n ] CLAIM : y ŷ RC 3/t Assuming this claim, if we choose t = 2 3 then 3/t = 1/2, we obtain y ŷ RC 1/2 and S t L (M) = ε n ŷ n + ϕ L (M) t = 2 3.

So it remains to prove CLAIM : y ŷ RC = ( (y n ŷ n ) 2 ) 1/2 3/t this follows from Lemma (P. JFA1978) y RC 1 ( S 4 ) 1/4 3 1/4 y RC 3 1/4 Indeed, y n ŷ n = ε n (S S t ) = ε n S1 S >t therefore (yn ŷ n ) 2 (y n ŷ n ) 2 + ϕ 2 = S 2 1 S >t t 2 S 4 and hence (y n ŷ n ) 2 S 2 1 S >t t 2 S 4 t 2 3 = 1/4

GT Non-commutative Khintchine inequalities In 1956, Grothendieck published the fundamental theorem of the metric theory of tensor products", now known as Grothendieck s theorem (GT in short), or Grothendieck s inequality, it has played a major role both in Banach space theory and in C -algebra theory. PROBLEM : extend to operator spaces"

A, B C -algebras with subspaces E A, F B. Previous versions : Any bounded linear map T : A B factors boundedly through a Hilbert space as : A T 1 H T 2 B More precisely : H comes from GNS-Hilbert spaces constructed using states f, g on A, B, and : < x, y >= f (x y) << x, y >>= g(yx )

Commutative case (Grothendieck 1956) (Banach spaces) K, L compact sets A = C(K ) B = C(L) T : A B There are P and Q probabilities on K and L such that A J P T L 2 (P) L b 2 (Q) J Q B J P : C(K ) L 2 (P) J Q : C(L) L 2 (Q) canonical inclusions T K G T. 1.338 < K C G < K IR G < 1.783 (unknown exact value of) K G

Key inequality (Grothendieck s inequality) : For any finite sequence (a i, b i ) in C(K ) C(L) Ta i, b i K G T a i 2 1/2 bi 2 1/2 (a, b) C(K ) C(L) ( Ta, b K G T ) 1/2 ( a 2 dp ) 1/2 b 2 dq

Non-commutative case (Banach spaces) f, g states on A, B, we define L 2 (f ) and L 2 (g) as completion of A and B respectively for scalar products < x, y >= [f (x y)+f (yx )]/2 << x, y >>= [g(x y)+g(yx )]/2 Then we obtain analogous factorisation with A J f T L 2 (f ) L b 2 (g) J g B T K T. Key inequality : For any finite sequence (a i, b i ) in A B Tai, b i K T [a a i i + a i ai ]/2 1/2 [b b i i + b i bi ]/2 1/2 Reference : P. JFA 1978, Haagerup Adv. Math. 1985

Equivalent form of Key inequality : (K T ) 1 Tai, b i { a i a i 1/2 + ai ai 1/2} { b i b i 1/2 + bi b i 1/2} = a 1/2 i a i b 1/2 i b i + a 1/2 i a i bi bi 1/2 + ai a i 1/2 b i b i 1/2 + ai a i 1/2 bi b i 1/2.

New (operator space) version of non-com GT completely bounded linear maps T : A B and completely bounded factorizations. cf. paper with D. Shlyakhtenko, Inv. Math (2002). Previously Conjectured by Effros-Ruan and Blecher

OS version of GT (under suitable assumption, and say in the separable case) Any completely bounded T : A B factors completely boundedly through the direct sum of two very simple building blocks" : the row and column Hilbert operator spaces, as : A T 1 R C T 2 B New feature (without Banach Space analogue) : same factorization valid for c.b. maps T : E F when E, F are EXACT operator spaces.

Equivalently, any T : E F decomposes as a sum T = T 1 + T 2 with T 1 factoring through R and T 2 factoring through C. So we will prove K T : E F inf {γ R (T 1 ) + γ C (T 2 )} K T cb. T =T 1 +T 2 Equivalently, this gives a factorization" for jointly CB bilinear forms ϕ : E F C as a sum ϕ = ϕ 1 + ϕ 2 with ϕ 1 (E h F ) and ϕ 2 (F h E). Left Open : Extension to the case of a B(H)-valued bilinear form E F B(H).

Exact operator spaces Let d SK (E) = inf{d cb (E, F) F K } By a simple perturbation argument, it can be shown that d SK (E) = inf{d cb (E, F) F M N, N dim E}. An operator space X is exact iff d SK (X) def = sup{d SK (E) E X, dim E < } <. We call this number the exactness constant of X.

The terminology exact" (due to Kirchberg) comes from the following FACT : Consider the short exact sequence {0} K B(l 2 ) B(l 2 )/K {0}. An operator space X is exact IFF this sequence remains exact after tensoring by X, more precisely if {0} K min X B(l 2 ) min X (B(l 2 )/K ) min X {0} is an exact sequence.

Theorem Theorem Let E A, F B be exact operator spaces (A, B C -algebras). Let C = d SK (E)d SK (F). Any c.b. map T : E F with T cb = 1 satisfies the KEY INEQUALITY : (a n, b n ) A B and w n > 0 ( weights") Tan, b n C [ w 2 n ana 1/2 n + w 2 n a n an [ w 2 n bnb 1/2 n + w 2 n b n bn 1/2] 1/2]

Theorem Consequently : Tan, b n 2C a n an 1/2 b 1/2 n b n + a 1/2 n a n bn bn 1/2. Conversely any map satisfying such an inequality must be c.b. with T cb 2C. So we obtain a characterization of c.b. maps from E to F.

Corollary Recall E A, F B. Any T CB(E, F ) extends" to a mapping T CB(A, B). Meaning : T viewed as bilinear form on A B extends T viewed as bilinear form on E F. A B E F T T Case of C -algebras : If E = A and F = B enough to assume one of them exact, or T suitably approximable by finite rank maps. NOTE : Very recently (announced in June 2007), Haagerup and Musat showed that this holds without any restriction on C -algebras A, B. C

Corollary T : A B with T cb 1, f 1, f 2, g 1, g 2 states on A, B respectively such that (a, b) A B Ta, b 2 3/2 [(f 1 (aa )g 1 (b b)) 1/2 + (f 2 (a a)g 2 (bb )) 1/2 ]. Corollary Assume E A is exact. Then a map u : E OH is c.b. IFF there is a constant K and states f, g such that a A ua 2 K f (a a) 1/2 g(aa ) 1/2. Note : Not true if assumption exact" is removed... If OH is replaced by (C, R) θ the characterization becomes a A ua 2 K f (a a) 1 θ g(aa ) θ.

Corollary T : A B with T cb 1, f 1, f 2, g 1, g 2 states on A, B respectively such that (a, b) A B Ta, b 2 3/2 [(f 1 (aa )g 1 (b b)) 1/2 + (f 2 (a a)g 2 (bb )) 1/2 ]. Corollary Assume E A is exact. Then a map u : E OH is c.b. IFF there is a constant K and states f, g such that a A ua 2 K f (a a) 1/2 g(aa ) 1/2. Note : Not true if assumption exact" is removed... If OH is replaced by (C, R) θ the characterization becomes a A ua 2 K f (a a) 1 θ g(aa ) θ.

Corollary A map u : A OH is c.b. IFF there is a constant K and states f, g such that a A ua 2 K f (a a) 1/2 g(aa ) 1/2. More generally, u : A (C, R) θ is c.b. IFF there is a constant K and states f, g such that a A ua 2 K f (a a) 1 θ g(aa ) θ.

Corollary Reformulation as a factorization (already mentioned) Any T CB(A, B ) can be factorized completely boundedly through an operator space of the form R C (or more generally R I C J for abstract sets I, J in the non separable case). More precisely there are c.b. maps T 1 : A R C T 2 : R C B such that T = T 2 T 1 and T 1 cb T 2 cb 2 3/2 T cb.

Applications In classical Banach space theory, GT has several well known consequences. Our result automatically allows to transfer some of these to the operator space setting. One of them is Corollary For a Banach space E {E L 1 and E L 1 } E Hilbert Proof One can assume (unessential) E reflexive for simplicity. We simply factorize through H the map L quotient E Id E L 1 and conclude that the identity of E factors through a subspace of a quotient of H, hence E Hilbert

If we now follow the analogy we find : Corollary For an operator space E {E Non.Com.L 1 and E Non.Com.L 1 } E SQ(R C) where SQ(R C) denotes the class of all quotients of subspaces of R C. Thus the single Hilbert space is now replaced by a class of spaces namely the class of all subspaces of quotients of R C denoted by SQ(R C). Note that this class is self-dual : indeed, (R C) R C C R R C and for any X,

Ingredients of the proof Kirchberg s ideas on Exactness Kirchberg s theorem that there is only one C -norm on C (IF ) B(H) The analysis of Voiculescu s free circular elements (free analog of Gaussian random variables) as generalized by Shlyakhtenko This last ingredient replaces the Gaussian isometric embedding in the classical GT

Gaussian Embeddings (g n ) independent Gaussian random variables, standard (mean zero, variance 1) on (Ω, A, P) J : l 2 L p (P) J(x) = x n g n because Jx p = x l2 g 1 p Jx dist = x l2 g 1 Hence x g 1 1 p Jx is an isometric embedding l 2 L p (P) (0 < p < ) In particular, if E = l 2 : we have both E L 1 and E L 1 CONVERSELY : Grothendieck s isomorphic charact. of Hilbert space : {E L 1 and E L 1 } E Hilbert

Problem : what remains for operator spaces? Does OH embed in a non-commutative L p -space for p <? Note : Using operator space duality such a non-commutative L 1 -space actually becomes an operator space, then using complex interpolation any non-commutative L p -space (1 < p < ) becomes an operator space.

Some preliminary answers The above Gaussian embedding produces a different space than OH in the operator space setting The embedding OH non commutative L p is impossible for 2 < p < (Junge, 1996) The spaces R and C (and hence their direct sum R C) embed in a non-commutative L 1 -space, namely the Schatten class S 1, the predual of B(l 2 ). Indeed their natural embedding in B(l 2 ) is completely complemented", i.e. there is a c.b. projection from B(l 2 ) onto either R or C. One can show that OH is completely isometric to a subspace of a quotient of R C. Thus to embed OH into a non-commutative L 1 -space it suffices to embed any quotient of R C.

Some preliminary answers The above Gaussian embedding produces a different space than OH in the operator space setting The embedding OH non commutative L p is impossible for 2 < p < (Junge, 1996) The spaces R and C (and hence their direct sum R C) embed in a non-commutative L 1 -space, namely the Schatten class S 1, the predual of B(l 2 ). Indeed their natural embedding in B(l 2 ) is completely complemented", i.e. there is a c.b. projection from B(l 2 ) onto either R or C. One can show that OH is completely isometric to a subspace of a quotient of R C. Thus to embed OH into a non-commutative L 1 -space it suffices to embed any quotient of R C.

Some preliminary answers The above Gaussian embedding produces a different space than OH in the operator space setting The embedding OH non commutative L p is impossible for 2 < p < (Junge, 1996) The spaces R and C (and hence their direct sum R C) embed in a non-commutative L 1 -space, namely the Schatten class S 1, the predual of B(l 2 ). Indeed their natural embedding in B(l 2 ) is completely complemented", i.e. there is a c.b. projection from B(l 2 ) onto either R or C. One can show that OH is completely isometric to a subspace of a quotient of R C. Thus to embed OH into a non-commutative L 1 -space it suffices to embed any quotient of R C.

Some preliminary answers The above Gaussian embedding produces a different space than OH in the operator space setting The embedding OH non commutative L p is impossible for 2 < p < (Junge, 1996) The spaces R and C (and hence their direct sum R C) embed in a non-commutative L 1 -space, namely the Schatten class S 1, the predual of B(l 2 ). Indeed their natural embedding in B(l 2 ) is completely complemented", i.e. there is a c.b. projection from B(l 2 ) onto either R or C. One can show that OH is completely isometric to a subspace of a quotient of R C. Thus to embed OH into a non-commutative L 1 -space it suffices to embed any quotient of R C.

Some preliminary answers The above Gaussian embedding produces a different space than OH in the operator space setting The embedding OH non commutative L p is impossible for 2 < p < (Junge, 1996) The spaces R and C (and hence their direct sum R C) embed in a non-commutative L 1 -space, namely the Schatten class S 1, the predual of B(l 2 ). Indeed their natural embedding in B(l 2 ) is completely complemented", i.e. there is a c.b. projection from B(l 2 ) onto either R or C. One can show that OH is completely isometric to a subspace of a quotient of R C. Thus to embed OH into a non-commutative L 1 -space it suffices to embed any quotient of R C.

Theorem (Junge s Theorem Invent. 2005) There is a von Neumann algebra M and a subspace of the predual E M that is completely 2-isomorphic to OH (i.e. d cb (E, OH) 2). Actually the same is true for any subspace of a quotient of R C. Moreover, OH embeds into a non-commutative L p for any 1 p 2. Open question Can 2 be replaced by 1 (completely isometric)?

Note : The above M is not semi-finite... Theorem (P. Bull. LMS 2004) If OH M, then M cannot be semi-finite. More precisely, if M is semi-finite and E M then any factorization through E of the identity OH E OH must be compact. Remark. In (IMRN 2004) I give a characterization of the (infinite dimensional) quotients of R C that embed into M with M semi-finite. There are precisely 7 isomorphic types! These are obtained by direct sum from simply R, C and G = span{g n n IN} L 1 (Ω, A, P) the span of standard Gaussians (g n ) as before. R, C, R C, G, R G, C G, R C G.

Note Moreover, where by definition G = span{g n n IN} span{ε n n IN} G R + C R + C = (R C)/{(x, t x)}. Let (e n ) denote the biorthogonal basis to (g n ) in G. Then for any y = (y n ) K N (or B(H) N ) { ( ) 1/2 y n e n = max yn yn ) } 1/2, ( y n y n.

Key ingredient : Shlyakhtenko s generalization of Voiculescu s FREE GAUSSIAN VARIABLES" Let H be any Hilbert space. Let H n = H 2 2 H (n times). We denote by F (H) the full Fock space over H, i.e. F (H) = C H H 2 H n We denote by V (for vacuum") the unit in C viewed as an element of F (H). For any h in H, we denote by l(h): F (H) F (H) the left creation operator, defined by : l(h)v = h and for any x in H n with n > 0 : l(h)x = h x. We will assume that H admits an orthonormal basis which can be split in two parts with equal cardinality, {e n n I} and {e n n I} so that the union {e n n I} {e n n I} is an orthonormal system.

Key ingredient : Shlyakhtenko s generalization of Voiculescu s FREE GAUSSIAN VARIABLES" Let H be any Hilbert space. Let H n = H 2 2 H (n times). We denote by F (H) the full Fock space over H, i.e. F (H) = C H H 2 H n We denote by V (for vacuum") the unit in C viewed as an element of F (H). For any h in H, we denote by l(h): F (H) F (H) the left creation operator, defined by : l(h)v = h and for any x in H n with n > 0 : l(h)x = h x. We will assume that H admits an orthonormal basis which can be split in two parts with equal cardinality, {e n n I} and {e n n I} so that the union {e n n I} {e n n I} is an orthonormal system.

We will denote l n = l(e n ) and l n = l(e n) Then we define, for any weights w n > 0, the generalized circular elements as follows : (1) c n = w n l n + wn 1 l n These are bounded operators on F (H).

When w n = 1, we recover circular elements -also called Free Gaussian"- in Voiculescu s sense. (2) f n = l n + l n If we use the symmetric Fock space instead of the full one, then we recover the standard (complex) Gaussians (but now these are unbounded operators of course) (3) g n = l n + l n

Let M be the von Neumann algebra generated by one of these three Gaussian systems. In (1) in general M is not semi-finite In (2) it is semi-finite, actually even finite with trace τ(t ) = TV, V. In (3) M L (Ω, A, P) again finite (actually commutative!) Our problem is to produce a quotient map Q : M OH or for any subspace S R C a quotient map Q : M S and then use an S admitting OH as a quotient.

In case (3) what works is essentially the orthogonal projection up to a normalization constant c! Q(x) = c g n, x e n then we obtain for any λ l 2 ( λ n 2 ) 1/2 = inf{ x x M Q(x) = λ} showing : l 2 is isometric to a quotient of M = L.

The corresponding operator space analogue would be the following : we need to find a Q such that for any (finite) family of coefficients (λ i ) in B(H) we have 1/2 λ n λ n inf { X X M B(H) [Q Id](X) = } e n λ n Then if we use either (2) ( Free Gaussians ") or (3) (standard Gaussians) we find inf { X } max { λ n λ n 1/2, λ nλ n 1/2} which produces G as a quotient of M.

So what works is finally to use (1) : we then find in that case { inf { X } max wn 2 λ n λ n 1/2, wn 2 λ nλ n 1/2} which gives us a subspace S R C spanned by or equivalently after rescaling (w n e 1n w 1 n e n1 ) (w 2 n e 1n e n1 ) but it is easy to see (using the rotational invariance of each of R and C) that it is essentially the form of the generic subspace S R C.

Conclusion. Any subspace S R C is a quotient of M, and hence, by duality, any quotient of R C is a subspace of M (and actually of M ). Finally : any E in SQ(R C) embeds in M!