Efficient parallel solution of sparse eigenvalue and eigenvector problems

This paper gives a new algorithm for computing the characteristic polynomial of a symmetric sparse matrix. We derive an interesting algebraic version of nested dissection, which constructs a sparse factorization the matrix A-λ where A is the input matrix. While nested dissection is commonly used to minimize the fill-in in the solution of sparse linear systems, our innovation is to use the separator structure to bound also the work for manipulation of rational polynomials in the recursively factored matrices. We compute the characteristic polynomial sparse symmetric matrix in polylog time using O(n(n+P(s(n))))⩽O(n(n+s(n)^2.376)) processors, where the sparsity graph of the matrix has separator size s(n). Our method requires only that the matrix be symmetric and nonsingular (it need not be positive definite as usual for nested dissection techniques); we use perturbation methods to avoid singularities. For the frequently occurring case where the matrix has small separator size our polylog parallel algorithm requires work bounds competitive with the best known sequential algorithms (i.e. sparse Lanczos methods), for example: (1) when the sparsity graph is a planar graph, s(n)⩽√n, and we require only n^2.188 processors, and (2) in the case where the input matrix is b-banded, we require only O(nP(b))=O(n) processors, for constant b