On the random satisfiable process

In this work we suggest a new model for generating random satisfiable k-CNF formulas. To generate such formulas – randomly permute all 2^k\binom{n}{k} possible clauses over the variables x_1, …, x_n, and starting from the empty formula, go over the clauses one by one, including each new clause as you go along if after its addition the formula remains satisfiable. We study the evolution of this process, namely the distribution over formulas obtained after scanning through the first m clauses (in the random permutation’s order). Random processes with conditioning on a certain property being respected are widely studied in the context of graph properties. This study was pioneered by Ruci'nski and Wormald in 1992 for graphs with a fixed degree sequence, and also by Erd\H{o}s, Suen, and Winkler in 1995 for triangle-free and bipartite graphs. Since then many other graph properties were studied such as planarity and H-freeness. Thus our model is a natural extension of this approach to the satisfiability setting. Our main contribution is as follows. For m \geq cn, c=c(k) a sufficiently large constant, we are able to characterize the structure of the solution space of a typical formula in this distribution. Specifically, we show that typically all satisfying assignments are essentially clustered in one cluster, and all but e^{-\Omega(m/n)} n of the variables take the same value in all satisfying assignments. We also describe a polynomial time algorithm that finds with high probability a satisfying assignment for such formulas.

💡 Research Summary

The paper introduces a novel random process for generating satisfiable k‑CNF formulas, extending the well‑studied “conditioned random process” paradigm from graph theory to the SAT domain. The process works as follows: list all possible k‑clauses over n variables (there are 2^k·C(n,k) of them), permute this list uniformly at random, and start with the empty formula. Scan the clauses in the random order, adding a clause only if the resulting formula remains satisfiable. After examining the first m clauses of the permutation, the distribution of the obtained formula is denoted by F(n,k,m).

The authors focus on the regime m ≥ c·n, where c=c(k) is a sufficiently large constant. Their main structural result is that a typical formula in this regime has an extremely concentrated solution space: almost all satisfying assignments belong to a single “cluster.” More precisely, with probability 1 − e^{‑Ω(m/n)} every variable takes the same value in all satisfying assignments, except for at most e^{‑Ω(m/n)}·n variables. Consequently, any two satisfying assignments differ on only a vanishing fraction of variables, and the Hamming diameter of the solution space is O(n·e^{‑Ω(m/n)}). This “single‑cluster” phenomenon sharply contrasts with the well‑known clustering transition in the standard random k‑SAT model, where multiple well‑separated clusters appear as the clause density approaches the satisfiability threshold.

To prove this, the authors perform a two‑stage probabilistic analysis. In the first stage they show that as clauses are added, the set of “forced” variables grows linearly, while the number of truly free variables shrinks exponentially in m/n. They use a martingale argument combined with a careful union bound over all possible partial assignments to control the probability that a clause addition creates a new degree of freedom. In the second stage they condition on the event that only a tiny set of variables remains free; they then demonstrate that any two extensions of the partial assignment to a full satisfying assignment must be close in Hamming distance, establishing the single‑cluster structure.

The second major contribution is an efficient algorithm that, with high probability, finds a satisfying assignment for a formula drawn from F(n,k,m). The algorithm proceeds greedily: repeatedly select the literal that appears most frequently among the currently admissible clauses, fix its value, and simplify the formula. After each fixation the algorithm checks that the reduced formula is still satisfiable; this check can be performed in polynomial time because the remaining formula is essentially a 2‑SAT instance or a collection of unit‑propagation steps. The single‑cluster property guarantees that the greedy choice never leads to a dead end: the high‑frequency literal must agree with the (essentially unique) global cluster, so fixing it preserves satisfiability. The algorithm runs in O(n·m) time and succeeds with probability 1 − e^{‑Ω(m/n)}.

An important conceptual point is that, unlike the classical random k‑SAT model where the clause‑to‑variable ratio controls a sharp SAT‑UNSAT threshold, the conditioned process deliberately avoids unsatisfiable formulas. Consequently, even for clause densities far above the usual threshold, the generated formulas remain satisfiable, yet their solution space collapses into a single dominant cluster. This demonstrates that conditioning on a global property (satisfiability) fundamentally reshapes the combinatorial landscape, mirroring similar phenomena observed in conditioned random graph processes such as triangle‑free or H‑free graph evolution.

The paper concludes with several directions for future work: extending the model to other logical frameworks (e.g., Horn clauses, XOR‑SAT), investigating the behavior when m is sublinear or super‑linear in n, and developing a general theory of conditioned random processes that unifies the graph and SAT settings. Overall, the work provides a rigorous characterization of the solution‑space geometry of the random satisfiable process and supplies a practical polynomial‑time algorithm for finding solutions, thereby opening a new line of inquiry at the intersection of probabilistic combinatorics and computational complexity.