2022arXiv (Cornell University)Open access

Randomized Cholesky QR factorizations

Oleg Balabanov

Open full text 3 citations

Abstract

This article proposes and analyzes several variants of the randomized Cholesky QR factorization of a matrix $X$. Instead of computing the R factor from $X^T X$, as is done by standard methods, we obtain it from a small, efficiently computable random sketch of $X$, thus saving computational cost and improving numerical stability. The proposed direct variant of the randomized Cholesky QR requires only half the flops and the same communication cost as the classical Cholesky QR. At the same time, it is more robust since it is guaranteed to be stable whenever the input matrix is numerically full-rank. The rank-revealing randomized Cholesky QR variant has the ability to sort out the linearly dependent columns of $X$, which allows to have an unconditional numerical stability and reduce the computational cost when $X$ is rank-deficient. We also depict a column-oriented randomized Cholesky QR that establishes the connection with the randomized Gram-Schmidt process, and a reduced variant that outputs a low-dimensional projection of the Q factor rather than the full factor and therefore yields drastic computational savings. It is shown that performing minor operations in higher precision in the proposed algorithms can allow stability with working unit roundoff independent of the dominant matrix dimension. This feature may be of particular interest for a QR factorization of tall-and-skinny matrices on low-precision architectures.

Open-access reader

About this research paper

What this paper is about

This article proposes and analyzes several variants of the randomized Cholesky QR factorization of a matrix $X$. Instead of computing the R factor from $X^T X$, as is done by standard methods, we obtain it from a small, efficiently computable random sketch of $X$, thus saving computational cost and improving numerical stability. The proposed direct variant of the randomized Cholesky QR requires only half the flops and the same communication cost as the classical Cholesky QR. At the same time, it is more robust since it is guaranteed to be stable whenever the input matrix is numerically full-rank. The rank-revealing randomized Cholesky QR variant has the ability to sort out the linearly dependent columns of $X$, which allows to have an unconditional numerical stability and reduce the computational cost when $X$ is rank-deficient. We also depict a column-oriented randomized Cholesky QR that establishes the connection with the randomized Gram-Schmidt process, and a reduced variant that outputs a low-dimensional projection of the Q factor rather than the full factor and therefore yields drastic computational savings. It is shown that performing minor operations in higher precision in the proposed algorithms can allow stability with working unit roundoff independent of the dominant matrix dimension. This feature may be of particular interest for a QR factorization of tall-and-skinny matrices on low-precision architectures.

Why it matters

OpenAlex reports 3 citations for this work. Citation counts describe recorded attention and do not establish research quality.

Key contribution

A contribution statement is not available in the OpenAlex record.

Method / approach

Method details are not available in the OpenAlex metadata.

Main findings

Findings are not separately available in the OpenAlex metadata.

Limitations

Limitations are not available in the OpenAlex metadata.

Applications

Application details are not available in the OpenAlex metadata.

Available abstract

This article proposes and analyzes several variants of the randomized Cholesky QR factorization of a matrix $X$. Instead of computing the R factor from $X^T X$, as is done by standard methods, we obtain it from a small, efficiently computable random sketch of $X$, thus saving computational cost and improving numerical stability. The proposed direct variant of the randomized Cholesky QR requires only half the flops and the same communication cost as the classical Cholesky QR. At the same time, it is more robust since it is guaranteed to be stable whenever the input matrix is numerically full-rank. The rank-revealing randomized Cholesky QR variant has the ability to sort out the linearly dependent columns of $X$, which allows to have an unconditional numerical stability and reduce the computational cost when $X$ is rank-deficient. We also depict a column-oriented randomized Cholesky QR that establishes the connection with the randomized Gram-Schmidt process, and a reduced variant that outputs a low-dimensional projection of the Q factor rather than the full factor and therefore yields drastic computational savings. It is shown that performing minor operations in higher precision in the proposed algorithms can allow stability with working unit roundoff independent of the dominant matrix dimension. This feature may be of particular interest for a QR factorization of tall-and-skinny matrices on low-precision architectures.

Key concepts: Cholesky decomposition, QR decomposition, Incomplete Cholesky factorization, Minimum degree algorithm, Mathematics, Factorization, Rank (graph theory), Matrix (chemical analysis)

Related papers

Back to paper searchBrowse research topicsOriginal source
Randomized Cholesky QR factorizations — Research Paper | ScholarLens