Graph Neural Multilevel Preconditioners for Iterative Solvers

DOI10.48550/arXiv.2607.28456MaRDI QIDQ7284478arXivFDO

Authors Yousef Saad, Zechen Zhang, Rui Peng Li

Publication date 30 July 2026

Published in arXiv

Tldr This paper proposes a Graph Neural Multilevel Preconditioner (GMP) that adopts an algebraic multigrid hierarchy as a structural prior and learns smoothing, restriction, and interpolation operators using a unified framework based on graph neural networks. We demonstrate GMP’s effectiveness for general sparse matrices, including indefinite or nonsymmetric systems, achieving improved convergence rates compared to state-of-the-art single-level preconditioners in certain regimes.

Full work available at URL https://doi.org/10.48550/arXiv.2607.28456

arXiv classification

cs.LGmath.NA

Cites work

• A general interpolation strategy for algebraic multigrid using energy minimization
• A review of algebraic multigrid
• Adam: A Method for Stochastic Optimization
• Approximate Inverse Preconditioners via Sparse-Sparse Iterations
• Bootstrap AMG
• Generalized Optimal AMG Convergence Theory for Nonsymmetric and Indefinite Problems
• Graph Neural Preconditioners for Iterative Solutions of Sparse Linear Systems
• Nonsymmetric algebraic multigrid based on local approximate ideal restriction (AIR)
• The University of Florida sparse matrix collection

This page was built for publication: Graph Neural Multilevel Preconditioners for Iterative Solvers

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q7284478)