UMR MIP 5640, Département de Mathématiques, INSA, Complexe Scientifique de Rangueil, 31077 Toulouse Cedex, France

Department of Computer Science and Engineering, University of Minnesota, 200 Union Street S.E., Minneapolis, MN 55455

Department of Computer Science, University of Minnesota - Duluth, 320 Heller Hall, 10 University Drive, Duluth, MN 55812-2496

Abstract

We present a class of preconditioning techniques which exploit rational function approximations to the original matrix. The matrix is first shifted and then an incomplete LU factorization of the resulting matrix is computed. The resulting factors are then used to compute a better preconditioner to the original matrix. Since the incomplete factorization is made on a shifted matrix, a good LU factorization is obtained without allowing much fill-in. The result needs to be extrapolated to the non-shifted matrix. Thus, the main motivation for this process is to save memory. The method is useful for matrices whose incomplete LU factorizations are poor, e.g., unstable. An error analysis for the conjugate gradient algorithm gives some guidance for choosing the shift of the matrix, in the special case where the shifted system is solved exactly. For several diffcult real-world problems, we provide a few examples of the practical application of the proposed techniques.