An Efficient Randomized QLP Algorithm for Approximating the Singular Value Decomposition

TL;DR

This paper presents Rand-QLP, a randomized matrix decomposition method that efficiently approximates the SVD, providing theoretical error bounds and demonstrating significant computational speedups over classical algorithms while maintaining comparable accuracy.

Contribution

The paper introduces Rand-QLP, a novel randomized QLP algorithm that leverages modern hardware and provides rigorous error bounds, improving efficiency over traditional SVD and QR-based methods.

Findings

Rand-QLP achieves up to 6.6x speedup over SVD on CPU.

Rand-QLP provides approximation quality comparable to SVD and pivoted QLP.

The method effectively utilizes random sampling and unpivoted QR for efficient computation.

Abstract

In this paper, we introduce a randomized QLP decomposition called Rand-QLP. Operating on a matrix $\bf A$, Rand-QLP gives ${\bf A}={\bf QLP}^T$, where $\bf Q$ and $\bf P$ are orthonormal, and $\bf L$ is lower-triangular. Under the assumption that the rank of the input matrix is $k$, we derive several error bounds for Rand-QLP: bounds for the first $k$ approximate singular values and for the trailing block of the middle factor $\bf L$, which show that the decomposition is rank-revealing; bounds for the distance between approximate subspaces and the exact ones for all four fundamental subspaces of a given matrix; and bounds for the errors of low-rank approximations constructed by the columns of $\bf Q$ and $\bf P$. Rand-QLP is able to effectively leverage modern computational architectures, due to the utilization of random sampling and the unpivoted QR decomposition, thus addressing a serious bottleneck associated with classical algorithms such as the singular value decomposition (SVD), column-pivoted QR (CPQR) and most recent matrix decomposition algorithms. To assess the performance behavior of different algorithms, we use an Intel Xeon Gold 6240 CPU running at 2.6 GHz with a NVIDIA GeForce RTX 2080Ti GPU. In comparison to CPQR and the SVD, Rand-QLP respectively achieves a speedup of up to 5 times and 6.6 times on the CPU and up to 3.8 times and 4.4 times with the hybrid GPU architecture. In terms of quality of approximation, our results on synthetic and real data show that the approximations by Rand-QLP are comparable to those of pivoted QLP and the optimal SVD, and in most cases are considerably better than those of CPQR.