Abstract
We investigate two types of the improvement techniques for the GMRES(m) method to solve nonsymmetric linear systems: the deflation-type restart and the Look-Back-type restart. From the analysis based on the residual polynomials, we show in this paper that these restart techniques modify the convergence behavior of the GMRES(m) method by different mathematical backgrounds. Then under the knowledge from the analysis, we propose an efficient improvement of the GMRES(m) method with these restart techniques. The numerical experiments indicate that the proposed method shows the efficient convergence behavior.
