CENTER FOR COMPUTATIONAL MATHEMATICS COLLOQUIUM
UNIVERSITY OF COLORADO AT DENVER
TITLE: On High Relative Accuracy in Matrix Singular Value and
Symmetric Eigenvalue Problems -- from Perturbation Theory to
Accurate Algorithms
SPEAKER: Zlatko Drmac
DATE: Monday, December 15, 1997
PLACE: Mathematics Conference Room 626
UCD Building, 1250 14th St., Denver
TIME: noon (Refreshments served at 11:45 am)
ABSTRACT
We present a unified approach to accurate floating--point computation
of the singular value decomposition (SVD) of products and quotients of
matrices and related spectral decompositions of positive definite matrix
pencils. The list of problems of interest includes:
(i) the generalized SVD of two full rank matrices;
(ii) the product induced SVD of matrix pairs and triplets;
(iii) the positive definite generalized eigenvalue problems
$ H x = \lambda M x $, $ H M x = \lambda x $
($H$, $M$ symmetric positive definite);
(iv) the $(H,K)$--SVD and $(H^{-1},K)$--SVD of general matrices with
applications to weighted least squares and canonical correlations.
We develop perturbation analysis and new algorithms which are capable
of achieving optimal theoretical accuracy. For example, in the symmetric
positive definite $n\times n$ eigenvalue problems $Hx=\lambda Mx$ and
$HMx=\lambda x$, each eigenvalue $\lambda$ is computed with high relative
accuracy. More precisely, the relative error $|\delta\lambda|/\lambda$
is, up to a factor of $n$, of the order
$\varepsilon \{ \min_{D\in{\cal D}}\kappa_2(D H D) +
\min_{D\in{\cal D}}\kappa_2(D M D) \}$,
where ${\cal D}$ denotes diagonal nonsingular matrices, $\kappa_2(\cdot)$
is the spectral condition number and $\varepsilon$ is the roundoff unit.
Furthermore, the backward errors $\delta H$, $\delta M$ are element--wise
small:
$|\delta H_{ij}|/\sqrt{H_{ii}H_{jj}}$ and $|\delta M_{ij}|/\sqrt{M_{ii}M_{jj}}$
are $O(n\varepsilon)$ for all $i$, $j$.