A Krylov subspace generalization of the improved Wilkinson's algorithm for ill-conditioned linear systems (Q260135)

To solve the system \(Ax=b\), this note proposes to start with an initial \(x^{(0)}\) and iterate with \(x^{(n+1)}=x^{(n)}+e^{(n)}\) where \(e^{(n)}=(D+pA)^{-1}p(b-Ax^{(n)})\). \(D\) is a symmetric positive definite matrix and \(p>0\). \textit{R. W. Freund} et al. poposed previously in [in: Acta Numerica 1992. Cambridge: Cambridge University Press. 57--100 (1992; Zbl 0762.65019)] a special case with \(D=qI\) as an improved Wilkinson iteration (originally Wilkinson chose \(q=0\) and \(p=1\)). In this note, the authors show that they may choose \(p\) such that the spectral radius of the iteration matrix is less than 1 so that convergence is guaranteed and moreover that the residue \(r^{(n)}=b-Ax^{(n)}\) belongs to the Krylov subspace \(K_n(C,r^{(0)})\) and \(x^{(n)}\in x^{(0)}+EK_{n-1}(C,r^{(0)})\) with \(E=p(D+pA)^{-1}\) and \(C=AE\), hence identifying the iteration as a preconditioned Krylov iteration.