Constraint satisfaction problems with isolated solutions are hard

We study the phase diagram and the algorithmic hardness of the random `locked’ constraint satisfaction problems, and compare them to the commonly studied ’non-locked’ problems like satisfiability of boolean formulas or graph coloring. The special property of the locked problems is that clusters of solutions are isolated points. This simplifies significantly the determination of the phase diagram, which makes the locked problems particularly appealing from the mathematical point of view. On the other hand we show empirically that the clustered phase of these problems is extremely hard from the algorithmic point of view: the best known algorithms all fail to find solutions. Our results suggest that the easy/hard transition (for currently known algorithms) in the locked problems coincides with the clustering transition. These should thus be regarded as new benchmarks of really hard constraint satisfaction problems.

💡 Research Summary

The paper investigates random “locked” constraint satisfaction problems (CSPs), a class of CSPs distinguished by the fact that each solution forms an isolated point in configuration space. In a locked problem, the constraint word A contains no consecutive “1” bits and every variable participates in at least two constraints. Consequently, moving from one solution to another requires flipping an entire closed loop of variables; on locally tree‑like random graphs this loop typically involves O(log N) variables. This structural property makes the solution space dramatically simpler than in standard CSPs such as K‑SAT or graph coloring, where clusters contain exponentially many solutions.

Using the cavity method and belief propagation (BP), the authors derive the replica‑symmetric (RS) equations for random occupation models, both for regular degree ensembles and for truncated Poissonian degree distributions. Because messages factorize on regular graphs, the RS solution can be obtained analytically, yielding explicit formulas for the entropy and for the critical constraint density α at which the system transitions from a satisfiable (SAT) phase to an unsatisfiable (UNSAT) phase. More importantly, the paper studies the reconstruction problem on trees: when information about the root variable can be inferred from the leaves, the RS ansatz breaks down and a clustering (or dynamical glass) transition occurs. For locked CSPs the clustering transition coincides exactly with the freezing transition—every variable becomes frozen within a cluster, and each cluster contains a single solution.

The authors then turn to algorithmic performance. They test the state‑of‑the‑art random CSP solver belief‑propagation‑reinforcement (BPR) on instances of locked problems across a range of constraint densities. Empirically, BPR (and other algorithms such as WalkSAT or Survey Propagation) succeeds easily for α below the clustering threshold α_c, but its success probability collapses sharply just above α_c. In the clustered phase, none of the tested algorithms can find a solution in polynomial time, indicating that the clustering/freezing transition marks the onset of genuine algorithmic hardness for all known methods. This contrasts with non‑locked CSPs, where a “hard” region exists but is typically narrower and sometimes surmountable by specialized heuristics.

The paper also discusses “balanced” locked occupation problems, where each variable appears in exactly L constraints and the constraint word satisfies A_0 = A_K = 0. These models retain the same isolated‑cluster structure and exhibit identical easy‑hard behavior, reinforcing the claim that the phenomenon is intrinsic to the locked nature rather than to particular degree distributions.

In conclusion, locked CSPs provide a rare combination of mathematical tractability and extreme computational difficulty. Their point‑like clusters allow precise analytical determination of phase diagrams without invoking full replica‑symmetry‑breaking machinery, yet the clustered phase is empirically intractable for all existing algorithms. This makes locked problems excellent benchmarks for testing new algorithms and for probing the fundamental relationship between solution‑space geometry (clustering, freezing) and algorithmic performance. Future work may explore algorithmic strategies that explicitly target loop flips, develop refined RSB analyses for locked models, or use locked CSPs as a testing ground for complexity‑theoretic conjectures about average‑case hardness.