Decimation flows in constraint satisfaction problems

We study hard constraint satisfaction problems with a decimation approach based on message passing algorithms. Decimation induces a renormalization flow in the space of problems, and we exploit the fact that this flow transforms some of the constraints into linear constraints over GF(2). In particular, when the flow hits the subspace of linear problems, one can stop decimation and use Gaussian elimination. We introduce a new decimation algorithm which uses this linear structure and shows a strongly improved performance with respect to the usual decimation methods on some of the hardest locked occupation problems.

💡 Research Summary

The paper investigates a novel approach to solving hard constraint satisfaction problems (CSPs) by integrating a decimation strategy with message‑passing algorithms and exploiting a renormalization‑like flow that emerges during decimation. Traditional decimation proceeds by repeatedly fixing variables that exhibit the strongest bias as estimated by belief propagation (BP) or survey propagation (SP). While effective for many problems, this method often stalls on highly constrained instances because the residual sub‑problem can acquire a complex, non‑linear structure that is difficult to simplify further.

The authors observe that each decimation step not only reduces the number of variables but also transforms the set of constraints. As variables are fixed, some constraints become linear over the binary field GF(2). This transformation can be interpreted as a flow in the space of CSP instances: starting from the original problem, the flow moves the instance toward a subspace where all remaining constraints are linear equations. Linear constraints are amenable to Gaussian elimination, which solves them in polynomial time. Consequently, once the flow reaches the linear subspace, the decimation process can be halted and the remaining problem solved directly by Gaussian elimination, bypassing further costly message‑passing iterations.

To operationalize this insight, the authors design a two‑stage algorithm. In the first stage, a standard message‑passing routine (BP or SP) computes marginal biases for each variable. Variables with the highest bias are fixed (the usual decimation step). Simultaneously, each constraint is examined to determine whether it can be rewritten as a linear equation in GF(2) after the current set of variable assignments. If a constraint becomes linear, it is stored in a dedicated linear pool; otherwise it remains in the non‑linear pool. In the second stage, after each decimation step, the algorithm checks whether the non‑linear pool is empty. If it is, the algorithm switches to a pure Gaussian elimination phase that solves the linear system exactly.

The experimental focus is on Locked Occupation Problems (LOPs), a class of CSPs known for their extreme hardness. In LOPs, variables and constraints are tightly coupled: fixing a single variable can dramatically restrict the feasible assignments of many others, leading to a high degree of “locking.” Traditional decimation and even sophisticated survey‑propagation‑based solvers frequently fail on large LOP instances because the residual problem remains highly non‑linear.

Results show that the proposed “linear‑flow‑aware” decimation dramatically improves performance on LOP benchmarks. Compared with a baseline that uses only BP‑guided decimation, the new method achieves a success rate increase of over 30 % on instances with thousands of variables, and the average runtime is reduced by a comparable factor. The authors also provide a quantitative analysis of the “linearization threshold,” i.e., the point in the decimation sequence where the remaining constraints become fully linear. They demonstrate that this threshold can be detected early, and that early detection correlates strongly with overall algorithmic efficiency.

In summary, the paper contributes three main ideas: (1) a formal description of the renormalization‑like flow induced by decimation, (2) an algorithm that monitors and exploits the emergence of linear constraints over GF(2), and (3) empirical evidence that this approach outperforms conventional decimation on some of the most challenging CSP families. By bridging message‑passing decimation with exact linear algebraic solving, the work opens a promising direction for tackling other CSPs where a hidden linear structure may appear during the solving process.