Peaks, plateaus, numerical instabilities in a Galerkin minimal residual pair of methods for solving \(Ax=b\) (Q1917421)

This paper discusses a pair of bidiagonalization methods, referred to as BLanczos and BMinres, for solving unsymmetric linear equation systems \(Ax = b\). Convergence is monitored through the residual norm \(|r_k|= |b - Ax_k|\), and it has been noted that this quantity does not decrease monotonically. In a Galerkin-type method such as BLanczos, irregular peaks occur, whilst in a residual minimization method such as BMinres, plateaus occur; in either case convergence is hard to identify. The main purpose of the paper is discussing possible reasons for these peak and plateau formations. It is shown in particular that, if the linear system is sufficiently well conditioned, numerical instabilities play no role in peak formations in BLanczos, and that peak or plateau production can occur in either finite precision or exact arithmetic, though more peaks are likely in the former. The question of whether there are correlations between the residual norms generated by each of the two methods when used on the same linear system is discussed. Detailed numerical examples are used to complement the discussions.