Matrices, moments, and rational quadrature (Q952044)

Let us consider a functional of the form \(F_u(A)= u^T f(A)u\), where \(A\in\mathbb{R}^{n\times n}\) is a large, sparse or structured symmetric matrix, \(u\in\mathbb{R}^n\) and \(f\) is a nonlinear function. This kind of functionals arises in many applications. The paper describes a quite simple approach to evaluate with fairly little computational effort upper and lower bounds of \(F_u(A)\) for functions \(f\) sufficiently many times differentiable in an interval containing the spectrum of \(A\). The approach extends to rational approximations of \(f\) a method (proposed by Golub and Meurant) for polynomial approximations of \(f\), based on the connection between the Lanczos procedure, orthogonal polynomials and Gauss-type quadrature rules.