Is the optimal \(\omega\) best for the SOR iteration method? (Q2365670)

\textit{J. Dancis} claimed in his recent paper [ibid. 154-156, 819-845, (1991; Zbl 0731.65020)] that a smaller average spectral radius can be achieved by using a polynomial acceleration of the SOR iteration together with a suboptimal relaxation parameter \(\omega<\omega_{opt}=2/(1+\sqrt{1-\rho^ 2(B)})\), where \(\beta=\rho(B)\) denotes the spectral radius of the Jacobi iteration matrix \(B\). This result can only be achieved if more information on the spectrum \(\sigma(B^ 2)\) of \(B^ 2\) than (1) \(\sigma(B^ 2)\subset[0,\beta^ 2]\) is at hand. The authors show that, for any real \(\omega\), no polynomial acceleration of the SOR method is asymptotically faster than the SOR scheme with \(\omega=\omega_{opt}\) under the only assumption (1) on \(\sigma(B^ 2)\). The authors also carefully investigate the question of what can be gained from the additional information \(\sigma(B^ 2)\subset[0,\gamma^ 2]\cup\{\beta^ 2\}\), with \(\gamma<\beta\), used by J. Dancis [loc. cit.].