Nonlinear Sherman-type inequalities (Q2311821)

Let \(J\subset \mathbb{R}\) be an interval. A theorem by \textit{S. Sherman} [Proc. Natl. Acad. Sci. USA 37, 826--831 (1951; Zbl 0044.27801)] states that if \(a\in\mathbb{R}_+^n\), \(b\in\mathbb{R}_+^m\) and \(x\in J^n, y\in J^m\) are such that \(y=Sx\) and \(a=S^T b\) for some \(n\times m\) row stochastic matrix \(S\), then \[\sum_{j=1}^m b_j f(y_j) \leq \sum_{i=1}^n a_i f(x_i)\] for any convex function \(f:J\to \mathbb{R}\). In the present paper, a non-trivial generalization of this result is obtained. In particular, for \(a,b,x,y\in\mathbb{R}^n\), \(\Psi:\mathbb{R}^n\times\mathbb{R}^n\to\mathbb{R}^l\) and convex \(f:\mathbb{R}\to\mathbb{R}\) such that several conditions are met, the inequality \[\Psi(b,(f(y_1),\ldots,f(y_n))^T) \leq \Psi(a,(f(x_1),\ldots,f(x_n))^T)\] is derived, where \(\leq\) denotes the componentwise order on \(\mathbb{R}^l\). Among others, the map \(\Psi\) needs to satisfy the so-called \textit{simplified Leon-Proschan property}. That is, for each \(n\times n\) permutation matrix \(g\) there exists another permutation matrix \(h\) such that \[\Psi(x,gy)=\Psi(hx,y)\] for every \(x,y\in\mathbb{R}^n\). Note, that while Sherman's inequality can be seen as a relation between two inner products, the inequality stated here involves a more general, possibly nonlinear map \(\Psi\). Furthermore, the author derives corollaries for interesting special cases as well as Sherman-type inequalities that are induced by directional derivatives and gradients of Schur-convex functions.