By how much can residual minimization accelerate the convergence of orthogonal residual methods? (Q5948215)

The authors study the relationship between pairs of orthogonal and minimal residual methods. Examples of such pairs are the conjugate gradient and the conjugate residual methods for Hermitian positive definite matrices, the Arnoldi method and GMRES for arbitrary nonsingular square matrices; the paper mentions several other examples, including solution by using the normal equations in two different ways. The analysis focuses on the minimal residual and quasi-minimal residual smoothing processes which provide the transition from the first method of such a pair of methods to the second method. In this analysis the roundoff error is ignored. With this restriction (reasonable in such a study) the analysis is very careful. From the proved relations between the norms of the residuals or their coordinate vectors (the quasi-residuals) of primary and smoothed methods, the authors derive several estimates for which can be gained in the 2-norm of the residual or quasi-residual by the transition from the original to the smoothed method. The results are illustrated by artificial examples. It would be nice to see applications to some of the standard problems in the literature.