Local Consistency and SAT-Solvers

Local consistency techniques such as k-consistency are a key component of specialised solvers for constraint satisfaction problems. In this paper we show that the power of using k-consistency techniques on a constraint satisfaction problem is precisely captured by using a particular inference rule, which we call negative-hyper-resolution, on the standard direct encoding of the problem into Boolean clauses. We also show that current clause-learning SAT-solvers will discover in expected polynomial time any inconsistency that can be deduced from a given set of clauses using negative-hyper-resolvents of a fixed size. We combine these two results to show that, without being explicitly designed to do so, current clause-learning SAT-solvers efficiently simulate k-consistency techniques, for all fixed values of k. We then give some experimental results to show that this feature allows clause-learning SAT-solvers to efficiently solve certain families of constraint problems which are challenging for conventional constraint-programming solvers.

💡 Research Summary

The paper establishes a precise correspondence between the well‑known local consistency technique of k‑consistency in constraint satisfaction problems (CSPs) and a specific inference rule on Boolean formulas called negative‑hyper‑resolution. The authors first formalize the direct encoding of a CSP into a set of propositional clauses: each CSP variable and each value in its domain are represented by a Boolean literal, and each constraint is translated into a collection of clauses that forbid illegal value combinations. Within this Boolean framework they introduce negative‑hyper‑resolution, a rule that combines a clause containing a single positive literal with several clauses each containing a negative literal of the same variable, producing a new clause that retains only the positive literal while eliminating the negatives. They prove that any inconsistency that can be derived by enforcing k‑consistency on the original CSP can also be derived by a sequence of negative‑hyper‑resolution steps whose intermediate resolvents never exceed size k. In other words, the domain‑pruning power of k‑consistency is exactly captured by bounded‑size negative‑hyper‑resolvents.

The second major contribution is an analysis of modern clause‑learning SAT solvers (e.g., MiniSat, Glucose, Lingeling) and how their conflict‑analysis and clause‑learning mechanisms implicitly perform negative‑hyper‑resolution. When a conflict is detected, the solver builds a conflict graph and extracts a learned clause using the first‑UIP (Unique Implication Point) strategy. The authors show that this learned clause is precisely the result of a negative‑hyper‑resolution inference whose size is bounded by the number of literals involved in the conflict. Consequently, if a CSP inconsistency can be expressed using a negative‑hyper‑resolvent of size at most k, the SAT solver will, with high probability, discover the corresponding learned clause in expected polynomial time. This establishes that clause‑learning SAT solvers automatically simulate k‑consistency for any fixed k, without any explicit CSP‑specific code.

To validate the theory, the authors conduct experiments on several families of CSP instances that are known to be difficult for conventional constraint‑programming (CP) solvers. These include variants of the pigeonhole principle, graph‑coloring problems on graphs with large girth, and Latin‑square‑like constraints that require high‑order consistency. Each instance is directly encoded into CNF and handed to state‑of‑the‑art SAT solvers. The empirical results demonstrate that the SAT solvers solve these benchmarks dramatically faster—often by orders of magnitude—than leading CP solvers such as Choco and Gecode. The speedup grows with problem size, confirming that the SAT solvers are indeed exploiting the simulated k‑consistency to prune the search space far more effectively than the CP systems, which must invoke explicit consistency algorithms.

The paper concludes by discussing the broader implications of this equivalence. It suggests that the boundary between CSP and SAT solving is more porous than previously thought, opening the door to hybrid architectures that combine the expressive modeling of CSPs with the powerful learning mechanisms of SAT solvers. Moreover, the authors propose that future work could explore other forms of local consistency (e.g., arc‑consistency, path‑consistency) in terms of alternative resolution‑style inference rules, and that specialized preprocessing or clause‑management strategies tailored to bounded‑size hyper‑resolvents might further boost SAT‑solver performance on structured combinatorial problems.

In summary, the authors prove two complementary facts: (1) k‑consistency on a CSP is exactly captured by bounded‑size negative‑hyper‑resolution on its direct Boolean encoding, and (2) modern clause‑learning SAT solvers discover such hyper‑resolvents in expected polynomial time, thereby implicitly performing k‑consistency for any fixed k. Experimental evidence confirms that this implicit simulation translates into concrete performance advantages on CSP families that challenge traditional CP techniques. The work thus bridges a theoretical gap and provides practical motivation for leveraging SAT technology in the realm of constraint satisfaction.