IRLB, a fast partial SVD
Authors: Baglama, Jim, University of Rhode Island; Kane, Michael, Yale
The singular value decomposition (SVD) is central to many important analysis methods and applications. Numerical implementations of the SVD are computationally intensive. Many applications of the SVD require only a subset of the largest or smallest singular values and corresponding singular vectors. The authors have prepared a Scikit-ready Python implementation of recent implicitly-restarted Lanczos (IRLB) methods for computing a few singular values and corresponding vectors of a matrix. The methods are compatible with dense or scipy.sparse matrices. The IRLB methods significantly outperform existing SVD implementations in computational and memory efficiency. Adaptation of IRLB to parallel computation for very large problems is straightforward.