From Fields to Random Trees

ICLR 2026 Conference Submission14951 Authors

Keywords: MAP estimation, Markov Random Fields, random spanning trees

Abstract: This study introduces a novel method for performing Maximum A Posteriori (MAP) estimation on Markov Random Fields (MRFs) that are defined on locally and sparsely connected graphs, broadly existing in real-world applications. We address this long-standing challenge by sampling uniform random spanning trees(SPT) from the associated graph. Such a sampling procedure effectively breaks the cycles and decomposes the original MAP inference problem into overlapping sub-problems on trees, which can be solved exactly and efficiently. We demonstrate the effectiveness of our approach on various types of graphical models, including grids, cellular/cell networks, and Erdős–Rényi graphs. Our algorithm outperforms various baselines on synthetic, UAI inference competition, and real-world PCI problems, specifically in cases involving locally and sparsely connected graphs. Furthermore, our method achieves comparable results to these methods in other scenarios.The code of our model can be accessed at \url{https://anonymous.4open.science/r/From-fields-to-trees-iclr-EB75}.

Primary Area: optimization

Submission Number: 14951