Thursday, October 20, 9:30 am in room MTH 3206, University of Maryland,
College Park
Peaks and Plateaus in Residual Norm Plots Generated by Galerkin and
Norm Minimizing Procedures for Solving Ax=b
Dr. Jane Cullum
Visiting Professor, Department of Computer Science, University of
Maryland, College Park and
IBM Research,T.J. Watson Research Center,
Yorktown Heights, NY 10598
cullumj@cs.umd.edu
The convergence of an iterative method for solving a linear system of equations
Ax=b is usually monitored by tracking the size of the norms of the residuals
generated. Galerkin methods, such as biconjugate gradients (BCG), often
converge well. However, plots of the norms of the residuals generated by
such procedures often exhibit erratic behavior. The residual norms may
oscillate irregularly, forming sequences of peaks of varying size and
frequency, making the user feel insecure about the method, and making it
difficult to track the convergence. To correct this problem, methods such
as quasi-minimal residuals (QMR), which minimize a related residual norm at
each iteration, have been proposed. However, these methods exhibit a different
type of irregular behavior. Plateaus appear in the residual norm plots,
intervals of iterations over which there exist unacceptably small decreases
in the residual norms. We first present the results of a series of numerical
experiments on a Galerkin/norm minimizing pair of bidiagonalization methods
which indicate certain relationships between residual norms generated by
the two procedures, and between the Galerkin residual norm plots and the
spectra of the associated Galerkin iteration matrices. We then derive
specific relationships between the residual norms generated within each of
three pairs of methods, GMRES/Arnoldi, QMR/BCG, and the pair of
bidiagonalization methods. Using these relationships we can then make
statements comparing the convergence rates within each pair. The
arguments for the bidiagonalization procedures include the effects of finite
precision arithmetic.
|