Exhaustive enumeration unveils clustering and freezing in random 3-SAT

We study geometrical properties of the complete set of solutions of the random 3-satisfiability problem. We show that even for moderate system sizes the number of clusters corresponds surprisingly well with the theoretic asymptotic prediction. We locate the freezing transition in the space of solutions which has been conjectured to be relevant in explaining the onset of computational hardness in random constraint satisfaction problems.

💡 Research Summary

The paper investigates the geometric structure of the full solution space of random 3‑SAT instances by exhaustive enumeration, providing a direct numerical test of predictions from statistical‑physics methods. Random K‑SAT formulas consist of N Boolean variables and M clauses, each clause containing K distinct variables chosen uniformly at random; the constraint density is α = M/N. As α grows, the probability of satisfiability drops sharply, leading to a well‑studied satisfiability threshold α_s ≈ 4.267 for 3‑SAT.

The authors generate random 3‑SAT instances with N ranging from 25 to 150 using the makewf generator, then enumerate all satisfying assignments with the exact solver relsat. They define a graph whose vertices are the solutions and whose edges connect pairs that differ in exactly one variable. Connected components of this graph are called “clusters”. Although this definition is not identical to the cavity‑method notion of “cavity‑clusters”, it reproduces many of their statistical properties and is applicable to any finite instance.

For each instance they count the number of clusters S, compute the complexity Σ(N) = (1/N) log S, and average over many random formulas (median instance selection is used to avoid outliers). The resulting Σ(N) curves for different N match the asymptotic prediction obtained from survey‑propagation equations (theoretical non‑zero complexity for α > 3.92) remarkably well, especially near the satisfiability threshold. This demonstrates that the asymptotic theory is quantitatively accurate even for moderate system sizes.

To study the freezing transition, the authors employ the whitening (or peeling) procedure. Starting from a solution, variables that belong only to already‑satisfied clauses are iteratively replaced by a joker symbol “”. The fixed point of this process is the whitening core. A variable that remains non‑ in the core is frozen: it takes the same Boolean value in every solution of the cluster. The freezing transition α_f is defined as the smallest α at which all solutions have a non‑trivial (i.e., not all‑) whitening core. By progressively deleting random clauses and monitoring when the first all‑ core appears, they estimate α_f = 4.254 ± 0.009. This value lies just below the satisfiability threshold and coincides with the empirical performance limits of the best known algorithms: stochastic local search works reliably up to α ≈ 4.21, while Survey Propagation (SP) decimation remains effective up to α ≈ 4.252. The close agreement supports the conjecture that the frozen phase is intrinsically hard for all known polynomial‑time heuristics.

The paper also discusses finite‑size scaling. Earlier works suggested a width exponent ν_s ≈ 1.5 for the satisfiability transition, but rigorous results indicate ν_s ≥ 2 and numerical studies show a crossover only for N ≈ 10⁴. In the present size range (N ≤ 150) the crossing point for the unfrozen probability P_f(α,N) shows negligible size dependence, indicating that the freezing transition is less affected by finite‑size effects than the satisfiability transition.

Overall, the study makes three main contributions: (1) it validates the statistical‑physics predictions for cluster complexity on finite, moderate‑size instances; (2) it provides the first direct numerical estimate of the freezing transition in random 3‑SAT, showing that it aligns with algorithmic hardness thresholds; and (3) it demonstrates that exhaustive enumeration, despite its computational limits, is a powerful tool for probing the detailed structure of solution spaces, offering insights complementary to Monte‑Carlo or message‑passing approaches. The authors suggest future work on 2‑SAT (where solutions are abundant), higher‑K SAT (where freezing separates more clearly from satisfiability), and real‑world CSP instances, as well as a deeper investigation of condensation and clustering transitions.