Random matrices and subexponential operator spaces (Q476502)

The classical Grothendieck inequality has deep consequences for the theory of tensor norms of Banach spaces. It asserts that there exists a universal constant \(K_G\) such that for every two commutative \(C^*\)-algebras \(A_i=C(K_i)\) and every bounded bilinear form \(u:A_1 \times A_2 \rightarrow {\mathbb C}\), there exist probability measures \(\mu_i\) on \(K_i\), or equivalently states \(\mu_i\in A_i^*\), such that \(| u(f_1,f_2)| \leq K_G \| u\| \| f_1\|_{L^2(\mu_1)} \| f_2 \|_{L^2(\mu_2)}\). Non-commutative generalizations, proved for C*-algebras with a certain approximation property by \textit{G. Pisier} [J. Funct. Anal. 29, 397--415 (1978; Zbl 0388.46043)] and for general C*-algebras \(A_1,A_2\) by \textit{U. Haagerup} [Adv. Math. 56, 93--116 (1985; Zbl 0593.46052)], represented an important achievement in the theory of operator algebras. One consequence is that every bounded linear operator \(T:A_1 \rightarrow A_2^*\) can be factorized through a Hilbert space \(H\) as \(T=SR\) with \(R:A_1 \rightarrow H\), \(S:H \rightarrow A_2^*\) linear bounded maps such that \(\| R\| \| S\| \leq 2\| T\|\). Major developments in the theory of operator spaces led Effros and Ruan, and independently Blecher, to conjecture in the early 90's that an analogous result should hold for jointly completely bounded (j.c.b.) bilinear forms \(u:A_1 \times A_2 \rightarrow {\mathbb C}\), with \(u\) viewed as a c.b.~linear map \(u:A_1 \rightarrow A_2^*\) with appropriate norm \(\| u\|_{cb}\). This was proved by \textit{G. Pisier} and \textit{D. Shlyakhtenko} [Invent. Math. 150, No. 1, 185--217 (2002; Zbl 1033.46044)] when one of \(A_1,A_2\) is exact (actually they considered j.c.b.~bilinear forms \(u:E_1 \times E_2 \rightarrow {\mathbb C}\) where \(E_i \subset A_i\) are operator spaces), and by \textit{U. Haagerup} and \textit{M. Musat} [Invent. Math. 174, No. 1, 139--163 (2008; Zbl 1188.46034)] for arbitrary C*-algebras. As a consequence, c.b.~linear maps \(u:E_1 \rightarrow E_2^*\) can be factorized, through a direct sum of a row Hilbert operator space and a column Hilbert operator space, as \(u=wv\) with c.b.~linear maps \(v:E_1 \rightarrow H_r \oplus K_c\), \(w:H_r \oplus K_c \rightarrow E_2^*\) such that \(\| v\|_{cb} \| w\|_{cb} \leq C\| u\|_{cb}\) for some constant \(C\). An alternative approach, inspired by embezzlement states in quantum information theory and providing new quantitative estimates, was discovered by \textit{O. Regev} and \textit{T. Vidick} [J. Oper. Theory 71, No. 2, 491--506 (2014; Zbl 1349.46062)]. The paper under review shows that the previous factorization result holds for a more general class of operator spaces, called \textit{subexponential}. To define it, the author considers, for every finite-dimensional operator space \(E\), every \(C>0\) and every positive integer \(N\), the quantity \(K_E(N,C)=\min\{ K\mid \exists F\subset M_K\) operator subspace such that \(d_N(E,F) \leq C\}\), where \(d_N(E,F)=\inf \{ \| u_N \| \| (u^{-1})_N \| \mid u:E\rightarrow F\) linear isometry\(\}\). An operator space \(X\) is called: {(i)} \textit{\(C\)-exact} if \(\sup_N K_E(N,C) <\infty\) for every finite-dimensional subspace \(E\subset X\), {(ii)} \textit{exact} if it is \(C\)-exact for some \(C\geq 1\). The infimum of such \(C\)'s is denoted by \(ex(X)\), {(iii)} \textit{\(C\)-subexponential} if \(\limsup_{N\rightarrow \infty} N^{-1} \log K_E(N,C)=0\) for every finite-dimensional subspace \(E\subset X\). One denotes \(C(X)=\inf\{ C>0\mid X\) is \(C\)-subexponential\(\}\), {(iv)} \textit{subexponential} if it is \(C\)-subexponential for some \(C\geq 1\). For subexponential operator spaces \(E_1,E_2\) it is proved that every c.b.~map \(u:E_1\rightarrow E_2^*\) can be factorized as \(u=wv\) with c.b.~linear maps \(v:E_1 \rightarrow H_r \oplus K_c\), \(w:H_r \oplus K_c \rightarrow E_2^*\) such that \(\| v\|_{cb} \| w\|_{cb} \leq 8C(E) C(F)\| u\|_{cb}\). The proof employs a Khintchin type inequality for Gaussian random matrices and the concentration of measure method. The paper contains a large number of very interesting examples. For instance, it is shown that the operator spaces \(OH\), \(R+C\) and \(\text{max} (\ell_2)\) are not subexponential. In a different direction, the random \(C^*\)-algebra generated by the block direct sum of a sequence of i.i.d. random matrices is used to produce examples of \(C^*\)-algebras which are \((1+\varepsilon)\)-subexponential for any \(\varepsilon >0\) but are not exact. Finally, for any \(n\geq 1\) and \(\varepsilon >0\), a continuum of \(n\)-dimensional subexponential operator spaces with constant \(2+\varepsilon\), exactness constant \(\geq \sqrt{n}/(2+\varepsilon)^3\) and mutual \(cb\)-distance \(\geq n/(2+\varepsilon)^4\) is produced.