Numerical results on the SD and CG methods for slightly non-symmetric matrices (Q2704280)

The author presents some theoretical and numerical results about the convergence of the steepest descent (SD) and the classical conjugate gradient (CG) methods when the coefficient matrix of the system of linear equations has to be solved is positive definite and ``almost'' symmetric, i.e. the matrix norm of the non-symmetric part of the matrix is less than some value related to the smallest and largest eigenvalues of the symmetric part of the given matrix. If \(A=A_s + A_n\) where \(A_s\) is the symmetric and \(A_n\) is the non-symmetric part of A, \(0 < \lambda_1 \leq \ldots \leq \lambda_n\) are the eigenvalues of \(A_s\) and \(\kappa = \lambda_n/ \lambda_1\) is the condition number of \(A_s\) the author proves that the SD method converges if \(\|A_n\|_s < \lambda_1 (-1 + \sqrt{(1+\kappa^{-1}})\) and that the CG method converges if \(\|A_n\|< \lambda_1 (-1 + \sqrt{(1+\kappa^{-1}})\). The proofs are done without use of the Krylov spaces, so they are much different from those that are done so far on CG-like methods for non-symmetric matrices. NEWLINENEWLINENEWLINEFinally, results from numerical experiments with SD and SG methods for a one-dimensional problem which describes the motion of a fluid with viscosity which leads to a system with slightly non-symmetric coefficient matrix under some conditions ar.