Volume 21, pp. 28-46, 2005.

Obtaining bounds on the two norm of a matrix from the splitting lemma

Doron Chen, John R. Gilbert, and Sivan Toledo

Abstract

The splitting lemma is one of the two main tools of support theory, a framework for bounding the condition number of definite and semidefinite preconditioned linear systems. The splitting lemma allows the analysis of a complicated system to be partitioned into analyses of simpler systems. The other tool is the symmetric-product-support lemma, which provides an explicit spectral bound on a preconditioned matrix.
The symmetric-product-support lemma shows that under suitable conditions on the null spaces of A and B, the finite eigenvalues of the pencil (A,B) are bounded by W22, where U=VW, A=UUT, and B=VVT. To apply the lemma, one has to construct a W satisfying these conditions, and to bound its 2-norm.
In this paper we show that in all its existing applications, the splitting lemma can be viewed as a mechanism to bound W22 for a given W. We also show that this bound is sometimes tighter than other easily-computed bounds on W22, such as WF2 and W1W. The paper shows that certain regular splittings have useful algebraic and combinatorial interpretations. In particular, we derive six separate algebraic bounds on the 2-norm of a real matrix; to the best of our knowledge, these bounds are new.

Full Text (PDF) [262 KB], BibTeX

Key words

matrix norm bounds, two-norm, norm bounds for sparse matrices, splitting lemma, support theory, support preconditioning

AMS subject classifications

15A60, 65F10, 65F35, 65F50.