A FEAST variant incorporated with a power iteration

Abstract

We present a variant of the FEAST matrix eigensolver for solving restricted real and symmetric eigenvalue problems. The method is derived from a combination of a variant of the FEAST method, which employs two contour integrals per iteration, and a power subspace iteration process. Compared with the original FEAST method, our new method does not require that the search subspace dimension must be greater than or equal to the number of eigenvalues inside a search interval, and can deal with narrow search intervals more effectively. Empirically, the FEAST iteration and the power subspace iteration are in a mutually beneficial collaboration to make the new method stable and robust.