On the real convergence rate of the conjugate gradient method (Q808614)

The author studies the convergence of the conjugate gradient (CG) algorithm for the solution of systems of linear equations in the presence of round-off errors. He describes very interesting numerical experiments with a parametrized set of matrices showing that a small change in the eigenvalue distribution (without change in the condition number) can cause a large change in the sensitivity of CG to rounding errors. In a theoretical part, a theorem is proved which gives equivalent conditions for ordering exact arithmetic CG processes for systems with different spectra according to the energy norm of the error in all iterations. The paper is supplied with some extensions of earlier results by \textit{C. C. Paige} [ibid. 34, 235-258 (1980; Zbl 0471.65017)] and \textit{A. Greenbaum} [ibid. 113, 7-63 (1989; Zbl 0662.65032)].