Numerical behaviour of the modified Gram-Schmidt GMRES implementation (Q1371674)

The authors study the implementation of the generalized minimal residual (GMRES) method, where the modified Gram-Schmidt process is used for computing the Arnoldi basis vectors. It is proved that these basis vectors begin to lose their linear independence only after the GMRES residual norm has been reduced to an accuracy, which is proportional to \(\kappa(A) \varepsilon\) (\(\kappa(A)\) is the condition number of the matrix of the considered linear algebraic system \(Au=f\), and \(\varepsilon\) denotes the machine precision).