Efficient spectral integral-based scheme for fast matrix-vector product algorithm

A fast algorithm to accelerate the matrix-vector product arising in conjugate gradient methods is proposed. The algorithm reduces the operational count to O(N/sup 3/2/) keeping the degree of accuracy to an arbitrarily high level. The technique is based on an efficient computation of a plane wave spectral integral along a steepest descent path using a Gauss-Hermite quadrature integration method.