Time-Memory Trade-Offs Using Sparse Matrix Methods for Large-Scale Eigenvalue Problems

Iterative methods such as Lanczos and Jacobi-Davidson are typically used to compute a small number of eigenvalues and eigenvectors of a sparse matrix. However, these methods are not effective in certain large-scale applications, for example, “global tight binding molecular dynamics.” Such applications require all the eigenvectors of a large sparse matrix; the eigenvectors can be computed a few at a time and discarded after a simple update step in the modeling process. We show that by using sparse matrix methods, a direct-iterative hybrid scheme can significantly reduce memory requirements while requiring less computational time than a banded direct scheme. Our method also allows a more scalable parallel formulation for eigenvector computation through spectrum slicing. We discuss our method and provide empirical results for a wide variety of sparse matrix test problems.