Computing orthogonal decompositions of block tridiagonal or banded matrices

Abstract

A method for computing orthogonal URV/ULV decompositions of block tridiagonal (or banded) matrices is presented. The method discussed transforms the matrix into structured triangular form and has several attractive properties: The block tridiagonal structure is fully exploited; high data locality is achieved, which is important for high efficiency on modern computer systems; very little fill-in occurs, which leads to no or very low memory overhead; and in most practical situations observed the transformed matrix has very favorable numerical properties. Two variants of this method are introduced and compared. © Springer-Verlag Berlin Heidelberg 2005.