Extreme eigenvalues of large sparse matrices by Rayleigh quotient and modified conjugate gradients (Q1067355): Difference between revisions

The extreme eigenvalues of a symmetric positive-definite matrix A may be obtained as the solution to an extremum problem, namely through the minimization or the maximization of the Rayleigh quotient by the conjugate gradients. While this procedure works well for the upper bound \(\lambda_ 1\), its rate of convergence proves too slow for the lower bound \(\lambda_ N\). For large sparse matrices the iteration can be extraordinarily accelerated by the aid of a preconditioning matrix derived from the incomplete Cholesky factorization of A. The new scheme has been applied to determine the smallest eigenvalue of finite element matrices of size N, witn N between 150 and 2220 taken from the engineering practice. The results show that a good estimate of \(\lambda_ N\) is achieved after very few iterations and that the Rayleigh quotient/modified conjugate gradient technique is more than one order of magnitude faster than the reverse power/conjugate gradient algorithm recently developed by the authors for the same problem.