Upper bounds for convergence rates of acceleration methods with initial iterations (Q1275439)

This paper is about the convergence of \(\text{GMRES}(n,k)\) methods for the numerical solution of large linear systems, with typical applications coming from discretization methods for the solution of partial differential equations. The \(\text{GMRES}(n,k)\) methods use a sequence of cycles, with each cycle consisting of \(k\) GMRES iterations and \(n\) Richardson iterations. This primary contribution of the paper is some analysis which shows how to derive tight bounds for the convergence rates of the \(\text{GMRES}(n,k)\) methods. These bounds depend on knowledge of the spectrum of the coefficient matrix for the linear system. The bounds imply that for many problems the \(\text{GMRES}(n,k)\) methods will converge much more rapidly than the \(\text{GMRES}(k)\) methods, and numerical results are given at the end of the paper to demonstrate this.