Residual smoothing and peak/plateau behavior in Krylov subspace methods (Q1917422)

When iterative methods are used for solving linear equation systems \(Ax = b\), the residual norm sequence \(|r_k|\) may behave in an irregular fashion; in particular, certain pairs of Krylov subspace methods may produce peaks and plateaus in their residual norm sequences. The author reviews recent theoretical results on residual smoothing techniques. Firstly, using the minimal residual smoothing due to \textit{R. Weiss} [Convergence behavior of generalized conjugate gradient methods, (1990; Zbl 0738.90074)], residuals produced by the orthogonal residual and minimum residual Krylov subspace methods are compared and the result shows that if one method performs well or badly, so does the other. Similar results using the quasi-minimal residual smoothing of \textit{L. Zhou} and \textit{H. F. Walker} [SIAM J. Sci. Comput. 15, No. 2, 297-312 (1994; Zbl 0802.65041)] relate residuals from the biconjugate gradient method and the quasi-residual minimum method, which indicate correlations of peaks in the former and plateaus in the latter, a fact observed experimentally.