A hybrid block GMRES method for nonsymmetric systems with multiple right-hand sides (Q1919396)

Given a set of linear equations with several right-hand sides (rhs) but the same coefficient matrix, the block GMRES method proposed by the authors in an earlier paper [Convergence properties of block GMRES and matrix polynomials, Linear Algebra Appl. (to appear)] computes iterates for all rhs from a Krylov subspace spanned by all initial residuals. Although these iterates satisfy an optimality condition, `pure' block GMRES is hardly computationally efficient. The authors propose a hybrid variant of block GMRES whose performance is by far better. This hybrid variant executes some initial steps of block GMRES and then extracts information on the systems based on the underlying block Arnoldi process, particularly via the latent roots of an associated matrix polynomial. Two different approaches to exploit this information for a subsequent semi-iterative phase are investigated. The Richardson-type approach turns out to be particularly useful. Computational experiments with up to 20 (random) rhs show that the new method can be up to 3 times faster than applying GMRES or TFQMR to all rhs individually, and much faster than `pure' block GMRES. The authors also discuss extensions to the case of several rhs together with several shifted coefficient matrices.