REVIEW 3 cited by
Randomized Cholesky QR factorizations
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
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.
Forward citations
Cited by 3 Pith papers
-
Sketch low-rank dynamics: orthogonal vs. oblique projections
Orthogonal sketch DLRA preserves classical DLRA dynamics and stability; oblique sketching of the Galerkin condition fails on large perpendicular residuals such as Vlasov–Poisson.
-
GPU-Parallelizable Randomized Sketch-and-Precondition for Linear Regression using Sparse Sign Sketches
A GPU implementation of sparse-sign-sketch-and-precondition for linear regression scales well on up to 8 NVIDIA A100s, with a novel rejection-sampling sketch generator.
-
RCLUPPr: a new randomized CholeskyQR with LU preconditioning
RCLUPPr, a randomized CholeskyQR variant with LU preconditioning, is shown stable and more applicable to ill-conditioned tall-skinny matrices than prior CholeskyQR-type algorithms.
Discussion (0). Continue with ORCID to comment.