On directional derivatives of trace functionals of the form \(A \mapsto \mathrm{Tr}({Pf}(A))\) (Q2418977)

Let $\mathbf{H}_n$ and $\mathbf{P}_n$ be the set of $n\times n$ Hermitian matrices and positive semidefinite matrices respectively. The \textit{directional derivative} of a real valued function $f\!:\!\mathcal{A}\;(\subseteq \mathbf{H}_n)\to \mathbb{R}$ at $A\in\mathcal{A}$ in the direction $B\in\mathbf{H}_n$ is defined as \[ df(A;B)=\lim_{t\to 0^+}\frac{f(A+tB)-f(A)}{t}. \] For a given differentiable function $f:(0,\infty)\to\mathbb{R}$, the author studies the induced trace functional $f_P:\mathbf{P}_n\to \mathbb{R}$ with $P\in\mathbf{P}_n$ defined as follows: for every $A\in\mathbf{P}_n$ has rank $r$ and admits the spectral decomposition $A=\sum_{i=1}^r\alpha_i v_iv_i^*$, \[ f_P(A)=\sum_{i=1}^r f(\alpha_i)\mathrm{Tr}(Pv_iv_i^*). \] The induced trace functional arises in the study of quantum information theory and the directional derivatives is connected to some corresponding optimization problems. \par In the paper, directional derivatives of $f_P$ induced by a special class of functions $f$ at $A\in\mathbf{P}_n$ in some specific directions $B\in\mathbf{H}_n$ are computed using the first divided difference of $f$ at every pair of eigenvalues of $A$. The function $f$ to be considered satisfies the condition that $\lim_{t\to 0^+}tf(t)=0$ and $B$ is restricted to have $A+tB\in\mathbf{P}_n$ for $t \in[0,\varepsilon)$, $\varepsilon>0$. The proof of the main result relies on the well-known spectral perturbation theory of Hermitian matrices. More specifically, the eigenvalues and eigenvectors of $A+tB$ are analytic in the real variable $t$. In addition, when $f$ admits an integral representation such as $f(x)=\log(x)$ or $f(x)=x^p$ with $-1<p<1$, the author gives an alternative proof without using the perturbation. Finally the function $f(t)=t^{-1}$ has been considered and a lower bound for the directional derivatives of the induced trace functional is obtained.