next up previous
Next: About this document ...

Karsten Kahl
Least-squares Based Multilevel Eigensolvers for Markov Chain Applications

Department of Mathematics
University of Wuppertal
Germany
kkahl@math.uni-wuppertal.de
Achi Brandt
UCLA
James Brannick
Department of Mathematics
Pennsylvania State University

Irene Livshits
Department of Mathematical Sciences
Ball State University

To guarantee efficient performance of algebraic multilevel methods for solving the eigenproblem Bx = x in Markov Chain applications, we are using an adaptive setup that captures the nature of the eigenmode of interest. The least-squares based approach computes appropriate testfunctions and fits them to define highly accurate interpolation that leads to a robust multilevel performance. The algorithm iteratively improves the ability of the multi-level approximation To the lowest mode of interest until it can be efficiently resolved using this hierarchy as preconditioner to an additive GMRES step. We test the proposed approach for a variety of applications, showing promising results over a wide range of tests.





root 2009-03-24