Convex \(\mathrm{SO}(N) \times \mathrm{SO}(n)\)-invariant functions and refinements of von Neumann's inequality (Q2475502)

Let \(\xi, \eta\in M_n({\mathbb R})\). We denote by \(\lambda_1(\xi)\leq \lambda_2(\xi)\leq \cdots\leq \lambda_n(\xi)\) the increasingly ordered singular values of \(\xi\). Then, von Neumann's inequality is the following: \[ \text{tr}(\xi\eta^t)\leq \sum_{k=1}^n\lambda_k(\xi)\lambda_k(\eta). \] In the paper under review, the authors show that this inequality can be significantly refined. In other words, on denoting by \(\mu_1(\xi), \;\mu_2(\xi),\dots, \mu_n(\xi)\) the signed singular values, that is, \(\mu_1(\xi):=\text{sgn}(\det \xi)\lambda_1(\xi)\) and \(\mu_k(\xi):=\lambda_k(\xi)\) for \(k\geq 2\), they prove the following inequality: \[ \text{tr}(\xi\eta^t)\leq \sum_{k=1}^n\mu_k(\xi)\mu_k(\eta). \] In the proof of this inequality, they do not use von Neumann's inequality. Moreover, they establish the link between this inequality and some results on convex \(\text{SO}(n)\times \text{SO}(n)\)-invariant functions. They also consider analogous results for rectangular matrices.