An Efficient, Sparsity-Preserving, Online Algorithm for Data Approximation

The Singular Value Decomposition (SVD) is a longstanding standard for data approximation because it is optimal in the 2 and Frobenius norms. The SVD, nevertheless, suffers from many setbacks, including computational cost, loss of sparsity in the decomposition, and the inability to be updated easily when new information arrives. Additionally, the SVD provides limited information on data features and variables that best represent the data. In this work, we present a truncated LU factorization called Spectrum-Revealing LU (SRLU) for effective low-rank matrix approximation, and develop the first algorithm to compute an SRLU factorization, which is both efficient and reliable. Our algorithm uses randomization and a novel LU updating technique with partial pivoting, which is more stable than any other known LU updating algorithm. We provide both approximation error bounds and singular value bounds for the SRLU approximation computed by our algorithm. Our analysis suggests that SRLU is competitive with the best low-rank matrix approximation methods, deterministic or randomized, in both computational complexity and approximation quality. Numeric experiments illustrate that SRLU preserves sparsity, highlights important data features and variables, can be efficiently updated, and calculates data approximations nearly as accurately as the SVD. To the best of our knowledge this is the first practical variant of the LU decomposition for efficient and effective low-rank matrix approximation.