Improved bound for rank revealing LU factorizations

In many applications it is necessary to determine the rank (or numerical rank) of a matrix. Many of these situations involve matrices that are very large order or that are sparse or that may undergo some form of modification (rank-k update, row or column appended or removed). In these cases the singular value decomposition's cost may be prohibitively high or the decomposition may not be computationally feasible (especially for large sparse problems). We thus examine the theoretical merits of rank revealing LU (RRLU) factorizations. We find that in those cases where the nullity is small and the gap is well defined, an RRLU factorization could be a very useful tool.