The analogues of entropy and of Fisher's information measure in free probability theory. VI: Liberation and mutual free information (Q1805830)

Let \(M\) be a W\(^{\ast}\)-algebra and \(\tau\) a n.f. trace on \(M\) with \(\tau (1)= 1\). If \(A_{1},\dots, A_{n}, B\) are \( \ast \)-subalgebras containing \(1\), the author defines the liberation Fisher information \(\varphi^{\ast} (A_{1} ;\dots ; A_{n}: B)\) as \(\sum_{k= 1}^{n} \|j (A_{k}: A_{1} \vee\dots \vee A_{k}^{\wedge} \vee\dots \vee A_{n} \vee B \|_{2}^{2}\) (\(\infty\) when some \(j\) does not exist), \(X \vee Y\) meaning the subalgebra generated by \(X \cup Y\) . In its turn, \(j (A: B)\) (the liberation gradient) is defined (when existing) by \(\tau (j (A: B) m) =\) (\(\tau \otimes \tau) (\delta_{A: B} m)\), \(m \in A \vee B\), \(\delta_{A: B}: A \vee B \rightarrow (A \vee B) \otimes (A \vee B)\) being the derivation taking \(a \in A\) into \(a \otimes 1\)-\(1 \otimes a\) and \(b \in B\) into \(0\). The mutual free information is defined by \(i^{\ast} (A_{1} ;\dots ; A_{n}: B)= (1/2) \int_0^\infty \varphi^{\ast} (U_{1} (t) A_{1} U_{1} (t)^{\ast} ;\dots ; U_{n} (t) A_{n} U_{n} (t)^{\ast}: B) d t\), \(U_{k}\) being free unitary Brownian motions constituting together with \(A_{1} \vee\dots \vee A_{n} \vee B\) a free family. Properties: \(A_{1},\dots, A_{n}, B\) are free if and only if \(\varphi^{\ast}= 0\) and if and only if \(i^{\ast}= 0\); if \(A_{1},\dots, A_{n}\) are \(B\)-free then \(\varphi^{\ast}= \sum \varphi^{\ast} (A_{i}: B)\) and the same for \(i^{\ast}\); inferior semicontinuity of \(\varphi^{\ast}\) and \(i^{\ast}\) with respect to a \( \ast \)-homomorphism of the full C\(^{\ast}\)-free product of \(A_{1},\dots, A_{n}, B\) into \(M\); \(\varphi^{\ast} (A: C)^{- 1} \geq \varphi^{\ast} (A: B)^{- 1} + \varphi^{\ast} (B: C)^{- 1}\) if \(A\), \(B\), \(C\) are freely Markovian (concept defined and studied in section 3 of the paper); \(\varphi^{\ast} (U_{1} (t) A_{1} U_{1} (t)^{\ast} ;\dots ; U_{n} (t) A_{n} U_{n} (t)^{\ast}: B)\) is right continuous, decreasing on \([ 0, \infty)\) and finite on \((0, \infty)\); the analogous function \(i\), defined with \(i^{\ast}\) instead of \(\varphi^{\ast}\), is continuous, decreasing, convex, not greater than \(- n \Sigma (U (t))\), (\(\Sigma (U (t))\) being \(\int \int \)log\(|z-u |d \mu (z) d \mu (u)\), where \(\mu\) is the distribution of \(U (t)\)) and its right derivative is \(- \varphi^{\ast}/2\). This describes the contents of sections 9,10. The sections 11,12 are devoted to the calculation of \(\varphi^{\ast}\) for a pair of subalgebras generated each by a subset of a semicircular system and for a pair of \(*\)-algebras generated each by a projector respectively. Section 4 contains heuristics for the above definitions, sections 5-8 are devoted to the study of \(j(A: B)\) (and \(\delta_{A: B}\) in section 6). Besides properties preceding those of \(\varphi^{\ast}\), let us mention the relation of \(j\) with \({\mathcal J}\) defined in the previous (V) paper of the author [Invent. Math. 132, No. 1, 189-227 (1998; Zbl 0930.46053)], the properties \(j(A: B)= j(A: B \vee C)\) for \(A\),\(B\),\(C\) freely Markovian, \(\|j(A: B)\|_{2} =\lim\|j(A_{k}: B_{k})\|_{2}\) if \(A_{k} \uparrow A\), \(B_{k} \uparrow B\), \(\sup\|j(A_{k}: B_{k})\|_{2} < \infty\), the inequalities (section 7) starting with \(|\tau(a b)|\leq \|j(A: B)\|(1 + \|j(A: B)\|^{2})^{- 1/2} \|a \|_{2} \|b \|_{2}\) for \(\tau(a)= \tau(b)= 0\), \(a \in A\), \(b \in B\), and in section 8: for an unitary \(U\) of infinite order \(j(U B U^{\ast}: B)= -E_{_{B \vee U B U} \ast} \partial_{U}^{\ast}(1 \otimes 1)\) if \(B( U , U^{- 1})\) is free, where \(\partial_{U}: C \rightarrow C \otimes C\), \(C =\) \textbf{C}\([ U , U^{- 1} ]\) is the derivation with \(\partial_{U} U= 1 \otimes 1\). If \(\tau(U)\neq 0\) and \(A\) contains a diffuse commutative W\(^{\ast}\)-algebra then it was proved in section 2 that \(U \in W^{\ast}(B \cup U B U^{\ast})\), i.e. the E disappears in the formula for \(j(U B U^{\ast}: B)\). Moreover, if the distribution of \(U\) is absolutely continuous on \(( |\cdot |= 1)\) and its density \(p\) is in \(L^{3}\), then \(\partial_{U}^{\ast}(1 \otimes 1)=\) \(ig(U)\) where \(g\) is the Hilbert transform of \(p\). In section 1 (and also in 9) some properties of the unitary free Brownian motion \(U(t)\) are established: the distribution of \(U(t)\) is absolutely continuous with a bounded density etc. The ``liberation process'' is \(\varrho_{t}\), where \(\varrho_{t}\) is the homomorphism of the free product \(A \ast B\) in \(M\) taking \(a \in A\) into \(U(t)a U(t)^{\ast}\) and \(b \in B\) into itself, where \(( U(t))\) is freely independent of \(A \cup B\); \(U(t)\) converges in distribution to a \(U_{\infty}\) and \(U_{\infty} A U_{\infty}^{\ast}\) and \(B\) are free. In section 12, it is proved that if \(T= T^{\ast} \in B '\), \(T\)-\(P \in A '\), where \(P\) is the projector on \textbf{C} in \(L^{2}(M)\), and \(B\) is a W\(^{\ast}\)-subalgebra, then \(j(A: B)= T-J T J\), \(J\) being the notation in the Tomita theory. The final section (14) contains four concluding remarks, we mention \(U \notin W^{\ast}(A \cup U B U^{\ast})\) if \(\varphi^{\ast}(A: B)< \infty\) and \(U\) and \(A \cup B\) are free and an algebraic generalization of a criterion for freeness involving \(\delta_{A: B}\).