MCC Seminar
Inderjit Dhillon,
University of Texas, Austin
DATE: Thursday, February 12, 2004
TIME: 11:00 A.M.
PLACE: 1003 MRL
TITLE: Fast Eigenvalue/Eigenvector Computation for Dense Symmetric
Matrices
ABSTRACT
Computing the eigenvalues/eigenvectors of a dense symmetric matrix is a
classical problem and usually proceeds in 3 phases: (i) reduce the
matrix to tridiagonal form, (ii) solve the tridiagonal eigenproblem and
(iii) back-transform to obtain eigenvectors of the original matrix. In
this talk, we present the first O(n^2), numerically stable,
embarrassingly parallel algorithm for phase (ii). All previous
algorithms for this purpose were O(n^3) in complexity, and in practice,
often accounted for 80% of the total time for solving large dense
eigenproblems. The new algorithm achieves O(n^2) complexity by avoiding
all Gram-Schmidt orthogonalization.
The crucial ingredients of the new algorithm are:
- using factored forms to achieve provably high relative accuracy,
- using "twisted factorizations" to compute approximate eigenvectors
with provably small relative residual norms.
- using multiple factored forms to "separate" clustered eigenvalues,
thereby automatically obtaining numerically orthogonal approximations
to eigenvectors (no Gram-Schmidt).
An interesting facet of our work is that high accuracy in intermediate
leads to a much faster overall algorithm.
Impressive speedups are obtained on matrices from the real-life
applications of computational quantum chemistry (PNNL) and
finite-element modeling (from Jeff Bennighof). The largest dense
problem solved "in-core" on a 256-node parallel computer is of size
128,000 x 128,000 which takes about 8 hours of CPU time. A preliminary
version of the sequential software is now included in the LAPACK
public-domain library.
The work on the sequential algorithm is joint with Beresford Parlett of
UC Berkeley, while the parallel algorithm has been developed jointly
with Paolo Bientinesi and Robert van de Geijn of UT Austin.