LU factoring of non-invertible matrices

Abstract

The definition of the LU factoring of a matrix usually requires that the matrix be invertible. Current software systems have extended the definition to non-square and rank-deficient matrices, but each has chosen a different extension. Two new extensions, both of which could serve as useful standards, are proposed here: the first combines LU factoring with full-rank factoring, and the second extension combines full-rank factoring with fraction-free methods. Amongst other applications, the extension to full-rank, fraction-free factoring is the basis for a fractionfree computation of the Moore-Penrose inverse.