Integer-Quantized Hierarchical Matrices with Adaptive Bit Allocation

DOI10.48550/ARXIV.2610.40134MaRDI QIDQ7409105arXivFDO

Authors Erin Carson, Xinye Chen, Ritesh Khan

Publication date 29 October 2026

Published in arXiv

Tldr We propose an adaptive bitwidth allocation method for quantizing low-rank factors in hierarchical matrices while meeting a given constraint on their reconstruction error. We develop a greedy algorithm to optimize storage and show its effectiveness through simulations with kernel and BEM matrices, achieving up to 44.3 times reduction in storage compared to dense storage in FP64.

Full work available at URL https://doi.org/10.48550/ARXIV.2610.40134

Mathematics Subject Classification ID

Numerical aspects of computer graphics, image analysis, and computational geometry (65D18) Numerical methods for low-rank matrix approximation; matrix compression (65F55) Roundoff error (65G50) Analysis of algorithms and problem complexity (68Q25)

arXiv classification

cs.MScs.PFmath.NA

Cites work

• A Fast Solver for HSS Representations via Sparse Matrices
• Construction of data-sparse \(\mathcal{H}^2\)-matrices by hierarchical compression
• Hierarchical low-rank arithmetic with floating point compression
• Hierarchical matrices based on a weak admissibility criterion
• Hierarchical matrices: algorithms and analysis
• Hybrid hierarchical matrices with adaptive mixed precision storage
• Mixed precision HODLR matrices
• Mixed precision low-rank approximations and their application to block low-rank LU factorization

This page was built for publication: Integer-Quantized Hierarchical Matrices with Adaptive Bit Allocation

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q7409105)