Eigenvalue inequalities for convex and log-convex functions (Q884407)

For a vector \(x\in\mathbb R^n\), denote by \(x^{\downarrow}\) the vector obtained by rearranging the coordinates of \(x\) in decreasing order. Let \(x, y\in\mathbb R^n\). The weak majorization relation \(x\prec_w y\) means \(\sum_{i=1}^k x_i^{\downarrow}\leq \sum_{i=1}^k y_i^{\downarrow}\) for \(1\leq k\leq n\). If \(x, y\) also have positive coordinates, the weak log-submajorization \(x\prec_{w\text{log}} y\) means \(\prod_{i=1}^k x_i^{\downarrow}\leq \prod_{i=1}^k y_i^{\downarrow}\) for \(1\leq k\leq n\). Denote by \(M_n\) the set of \(n\times n\) complex matrices, and by \(\lambda(A)\) for \(A\in M_n\) the vector of the eigenvalues of \(A\). In this paper, among others, interesting eigenvalue inequalities for convex and log-convex functions are derived. It is shown that if \(A\) is an \(m\times m\) Hermitian matrix whose eigenvalues are contained in an interval \(I\) and \(f\) is a convex function on \(I\), then \[ \lambda^{\downarrow}(f(\Phi(A)))\prec_w\lambda^{\downarrow}(\Phi(f(A))), \] where \(\Phi\) is a unital positive map from \(M_m\) to \(M_n\). Moreover, if \(A, B\) are \(n\times n\) Hermitian matrices whose eigenvalues are contained in \(I\) and \(f\) is a log-convex function on \(I\), then \[ \lambda^{\downarrow}(f(\alpha A+(1-\alpha)B))\prec_{w\log} \lambda^{\downarrow} (f(A)^\alpha f(B)^{1-\alpha}) \] for \(0\leq \alpha \leq 1\).