On the Convergence of GMRES with Invariant-Subspace Deflation

We consider the solution of large and sparse linear systems of equations by GMRES. Due to the appearance of unfavorable eigenvalues in the spectrum of the coefficient matrix, the convergence of GMRES may hamper. To overcome this, a deflated variant of GMRES can be used, which treats those unfavorable eigenvalues effectively. In the literature, several deflated GMRES variants are applied successfully to various problems, while a theoretical justification is often lacking. In contrast to deflated CG, the convergence of deflated GMRES seems to be harder to analyze and to understand. This paper presents some new theoretical insights into deflated GMRES based on A-invariant deflation subspaces. Fundamental results regarding the convergence of deflated GMRES are proved in order to show the effectiveness and robustness of this method. Numerical experiments are provided to illustrate the theoretical results and to show some further properties of deflated GMRES. Consequently, practical variants of deflated GMRES from the literature can be better understood based on the results presented in this paper. University of Wyoming, Department of Mathematics, 1000 East University Avenue Laramie, WY 82071, USA ([email protected]), This research was supported by Flittie Sabbatical Augmentation Award, University of Wyoming. University of Minnesota, Department of Computer Science and Engineering, 200 Union Street S.E, Minneapolis, MN 55455, USA ([email protected]), Work supported in part by DOE under grant DE-FG 08ER 25841. Delft University of Technology, Faculty of Electrical Engineering, Mathematics and Computer Science, Delft Institute of Applied Mathematics, P.O. Box 5031, 2600 GA Delft, The Netherlands, ([email protected])