No label defined (Q4940822)

The authors consider the two-level method mentioned in the title that was studied is several previous papers. They present two convergence results. First, global convergence is proved under the assumption that a sufficiently high number of relaxations on the fine level is performed. Second, it is proved that for local convergence (the starting vector sufficiently close to the result) an arbitrary number of relaxations (especially only one) is sufficient. No restrictive requirements are made on the stochastic matrix. The authors study also the dependence of the fine level error on the coarse level error. The results are illustrated by numerical experiments. Matrices with tridiagonal blocks are considered. The blocks are nonzero either in all positions or only on three block diagonals. Clusters of eigenvalues are present. The order of the matrices goes up to 1000.