A Block Bidiagonalization Method for Fixed-Accuracy Low-Rank Matrix Approximation
Eric Hallman · SIAM Journal on Matrix Analysis and Applications · 2022
We present randUBV, a randomized algorithm for matrix sketching based on the block Lanzcos bidiagonalization process. Given a matrix $\mathbb{A}$, it produces a low-rank approximation of the form $\mathbb{UBV}^T$, where $\mathbb{U}$ and $\mathbb{V}$ have orthonormal columns in exact arithmetic and $\mathbb{B}$ is block bidiagonal. In finite precision, the columns of both $\mathbb{U}$ and $\mathbb{V}$ will be close to orthonormal. Our algorithm is closely related to the randQB algorithms of Yu, Gu, and Li [ SIAM J. Matrix Anal. Appl., 39 (2018), pp. 1339--1359]. in that the entries of $\mathbb{B}$ are incrementally generated and the Frobenius norm approximation error may be efficiently estimated. It is therefore suitable for the fixed-accuracy problem and so is designed to terminate as soon as a user input error tolerance is reached. Numerical experiments suggest that the block Lanczos method is generally competitive with or superior to algorithms that use power iteration, even when $\mathbb{A}$ has significant clusters of singular values.