Multivariate extensions of the Golden-Thompson inequality (Q2516304)

Assume that \(H_1,\dots,H_k\) are \(m \times n\) matrices such that \[ H_1^*H_1+\dots+H_k^*H_k=I_n. \] Then the following Golden-Thompson's trace inequality holds \[ \operatorname{Tr}\exp\left(L+\sum_{i=1}^kH_i^*B_iH_i\right) \leq \operatorname{Tr}\exp(L)\sum_{i=1}^kH_i^*\exp(B_i)H_i, \] where \(L\) is an arbitrary selfadjoint \(n \times n\) matrix and \(B_1,\dots,B_k\) are arbitrary \(m \times m\) matrices. In the paper, the author proves the following very nice statement. Theorem. If \(H_1,\dots,H_k\) are \(m \times n\) matrices such that \[ H_1^*H_1+\dots+H_k^*H_k \leq I_n, \] and \(L\) is a selfadjoint \(n \times n\) matrix, then the trace function \[ \varphi(A_1,\dots,A_k)=\operatorname{Tr}\exp\left(L+\sum_{i=1}^kH_i^*(\log A_i)H_i\right) \] is concave in \(k\)-tuples of positive definite \(m \times m\) matrices. So, by using this fact, if \(H_1^*H_1+\dots+H_k^*H_k=I_n\), then \[ \operatorname{Tr}\exp\left(\sum_{i=1}^kH_i^*(\log B_i-\log A_i)H_i\right) \leq \sum_{i+1}^k \operatorname{Tr}H_i^*\big(d\log(A_i)B_i\big)H_i, \] where \(A_1,\dots,A_k\) and \(B_1,\dots,B_k\) are positive definite \(m \times m\) matrices and \(d\) is the first Fréchet differential. Also, in continuation, the author concludes the above Golden-Thompson's trace inequality.