On the joint convexity of the Bregman divergence of matrices (Q2351643): Difference between revisions

Let \(I\) denote an interval in the real line and let \(f:I \rightarrow \mathbb{R}\) be given. For \(A \in \mathbb{C}^{n \times n}\), let \(\sigma(A)\) denote its spectrum. The matrix function corresponding to \(f\) is the mapping \(f:\{A \in \mathbb{C}^{n \times n}: A\) is Hermitian and \(\sigma(A) \subseteq I \} \rightarrow \mathbb{C}^{n \times n}\), defined by \(f(A)=\sum_j f(\lambda_j) P_j\) whenever \(A\) has the spectral decomposition \(A=\sum_j \lambda_j P_j\). Let \(f\) be a continuously differentiable convex function on the positive real axis and let \(X,Y\) be two positive definite Hermitian matrices. The Bregman \(f\)-divergence of \(X\) and \(Y\) is defined by \(H_f(X,Y):=\mathrm{tr}(f(X)-f(Y)-{\mathbf D}f[Y](X-Y))\), where tr denotes the trace and \({\mathbf D}f[Y]\) denotes the Fréchet derivative of the standard matrix function \(f\). Now let \(f\) be a \(C^2\) function on the positive real axis such that its second derivative is strictly positive. The main result of the article is that the Bregman \(f\)-divergence is jointly convex (in both the arguments) if and only if the map \(\phi(X) =({\mathbf D}f'([X])^{-1}\) (whose domain is the set of all Hermitian positive definite matrices) is operator concave. An application to the quantum Tsallis entropy is presented.