Reduced and Mixed-Precision Block Householder QR Factorization

DOI10.48550/ARXIV.2610.38537MaRDI QIDQ7409130arXivFDO

Authors Peter Benner, Martin Köhler, Eda Oktay, Xiaobo Liu

Publication date 28 October 2026

Published in arXiv

Tldr We analyze and optimize the block Householder QR factorization using the WY representation and present a new mixed-precision algorithm mpWYQR that avoids many precision conversions by performing norm computations with single precision accumulation to ensure numerical stability. This method retains accuracy comparable to an existing state-of-the-art mixed-precision approach, while running up to 18x faster and using less memory.

Full work available at URL https://doi.org/10.48550/ARXIV.2610.38537

Mathematics Subject Classification ID

Multilinear algebra, tensor calculus (15A69) Roundoff error (65G50) Parallel numerical computation (65Y05) Complexity and performance of numerical algorithms (65Y20)

arXiv classification

math.NA

Cites work

• A mixed precision algorithm for the matrix square root
• A mixed precision Jacobi SVD algorithm
• A Storage-Efficient WY Representation for Products of Householder Transformations
• Accuracy and Stability of Numerical Algorithms
• Block Householder transformation for parallel QR factorization
• Communication-optimal parallel and sequential QR and LU factorizations
• Efficient CP rounding using alternating least squares with QR decomposition
• Finding structure with randomness: probabilistic algorithms for constructing approximate matrix decompositions
• LAPACK Users' Guide
• Matrix Computations
• Mixed precision algorithms in numerical linear algebra
• Mixed precision block fused multiply-add: error analysis and application to GPU tensor cores
• Mixed-precision algorithms for solving the Sylvester matrix equation
• Probabilistic Rounding Error Analysis of Householder QR Factorization
• Rounding error analysis of mixed precision block Householder QR algorithms
• Rounding Errors in Algebraic Processes
• The WY Representation for Products of Householder Matrices
• Three-precision GMRES-based iterative refinement for least squares problems
• Unitary Triangularization of a Nonsymmetric Matrix

This page was built for publication: Reduced and Mixed-Precision Block Householder QR Factorization

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q7409130)