On the convergence rate of the conjugate gradients in presence of rounding errors (Q1326451)

The paper investigates rounding error effects on the convergence rate of the conjugate gradients. Both theoretical and experimental analyses are performed to examine how known bounds on the number of iterations are affected by finite precision arithmetic. Special attention is paid to the case where the spectrum of the system matrix presents small or large isolated eigenvalues. It is shown that the number of iterations needed to eliminate error modes associated with interior and small isolated eigenvalues is only weakly affected by rounding errors; whereas, in the case of large isolated eigenvalues, this number is directly proportional to the logarithm of the inverse machine precision. Reliable bounds on the number of iterations are derived in each case.