Message-Passing GNNs Fail to Approximate Sparse Triangular Factorizations

Vladislav Trifonov [email protected], Ekaterina Muravleva [email protected], Ivan Oseledets [email protected] · 2025

Read Message-Passing GNNs Fail to Approximate Sparse Triangular Factorizations by Vladislav Trifonov [email protected], Ekaterina Muravleva [email protected], Ivan Oseledets [email protected] (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-01397.txt.

Description

Trifonov, Muravleva, and Oseledets show that message-passing graph neural networks cannot learn a key matrix factorization, in a 2025 paper. Sparse triangular factors serve as preconditioners that speed up linear solvers in science and engineering. Graph networks have been proposed to predict them cheaply. The catch is that good factors depend on far-apart entries, while message passing only mixes local neighborhoods.

The team back the theory with tests on synthetic matrices and real ones from the SuiteSparse collection. Across attention networks and graph transformers, predicted factors matched the true ones poorly. Even a fully global transformer failed to reach simple baselines. The authors conclude that scientific computing needs architectures built for non-local structure, not borrowed message passing.

Catalog record

CREATORVladislav Trifonov [email protected], Ekaterina Muravleva [email protected], Ivan Oseledets [email protected]
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE39 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2502-01397

Browse the full library · About the Tower of Babel