Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Preprint |
| Published: |
2025
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2504.06757 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866908309558132736 |
|---|---|
| author | Machado, David González-García, Jonathan Mulet, Roberto |
| author_facet | Machado, David González-García, Jonathan Mulet, Roberto |
| contents | Despite significant advances in characterizing the highly nonconvex landscapes of constraint satisfaction problems, the good performance of certain algorithms in solving hard combinatorial optimization tasks remains poorly understood. This gap in understanding stems largely from the lack of theoretical tools for analyzing their out-of-equilibrium dynamics. To address this challenge, we develop a system of approximate master equations that capture the behavior of local search algorithms in constraint satisfaction problems. Our framework shows excellent qualitative agreement with the phase diagrams of two paradigmatic algorithms: Focused Metropolis Search (FMS) and greedy-WalkSAT (G-WalkSAT) for random 3-SAT. The equations not only confirm the numerical observation that G-WalkSAT's algorithmic threshold is nearly parameter-independent, but also successfully predict FMS's threshold beyond the clustering transition. We also exploit these equations in a decimation scheme, demonstrating that the computed marginals encode valuable information about the local structure of the solution space explored by stochastic algorithms. Notably, our decimation approach achieves a threshold that surpasses the clustering transition, outperforming conventional methods like Belief Propagation-guided decimation. These results challenge the prevailing assumption that long-range correlations are always necessary to describe efficient local search dynamics and open a new path to designing efficient algorithms to solve combinatorial optimization problems. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2504_06757 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Local equations describe unreasonably efficient stochastic algorithms in random K-SAT Machado, David González-García, Jonathan Mulet, Roberto Disordered Systems and Neural Networks Despite significant advances in characterizing the highly nonconvex landscapes of constraint satisfaction problems, the good performance of certain algorithms in solving hard combinatorial optimization tasks remains poorly understood. This gap in understanding stems largely from the lack of theoretical tools for analyzing their out-of-equilibrium dynamics. To address this challenge, we develop a system of approximate master equations that capture the behavior of local search algorithms in constraint satisfaction problems. Our framework shows excellent qualitative agreement with the phase diagrams of two paradigmatic algorithms: Focused Metropolis Search (FMS) and greedy-WalkSAT (G-WalkSAT) for random 3-SAT. The equations not only confirm the numerical observation that G-WalkSAT's algorithmic threshold is nearly parameter-independent, but also successfully predict FMS's threshold beyond the clustering transition. We also exploit these equations in a decimation scheme, demonstrating that the computed marginals encode valuable information about the local structure of the solution space explored by stochastic algorithms. Notably, our decimation approach achieves a threshold that surpasses the clustering transition, outperforming conventional methods like Belief Propagation-guided decimation. These results challenge the prevailing assumption that long-range correlations are always necessary to describe efficient local search dynamics and open a new path to designing efficient algorithms to solve combinatorial optimization problems. |
| title | Local equations describe unreasonably efficient stochastic algorithms in random K-SAT |
| topic | Disordered Systems and Neural Networks |
| url | https://arxiv.org/abs/2504.06757 |