Adversarial Satisfiability Problem

We study the adversarial satisfiability problem, where the adversary can choose whether variables are negated in clauses or not in order to make the resulting formula unsatisfiable. This is one case of a general class of adversarial optimization problems that often arise in practice and are algorithmically much harder than the standard optimization problems. We use the cavity method to compute large deviations of the entropy in the random satisfiability problem with respect to the negation-configurations. We conclude that in the thermodynamic limit the best strategy the adversary can adopt is extremely close to simply balancing the number of times every variable is and is not negated. We also conduct a numerical study of the problem, and find that there are very strong pre-asymptotic effects that are due to the fact that for small sizes exponential and factorial growth is hardly distinguishable.

💡 Research Summary

The paper introduces and studies the “adversarial satisfiability” (AdSAT) problem, a variant of the classic Boolean satisfiability (SAT) task in which an adversary is allowed to choose, for each clause, whether each variable appears negated or not. The adversary’s goal is to make the resulting formula unsatisfiable, thereby turning a standard optimization problem into a worst‑case, adversarial one. This formulation captures a broad class of adversarial optimization problems that appear in security, robust machine‑learning, and fault‑tolerant system design, and it is believed to be substantially harder than the corresponding non‑adversarial version.

The authors focus on random k‑SAT instances with N variables and M=αN clauses. For a given negation configuration σ (a binary vector indicating the presence of a negation for each literal), they define the entropy S(σ)=log |{assignments x that satisfy the formula F(x;σ)}|. The adversarial problem is then to minimize S(σ) over all possible σ. To analyze the typical behavior of this minimization in the thermodynamic limit (N→∞), they employ the cavity method, a powerful tool from statistical physics that extends mean‑field theory to sparse, random graphs.

The cavity analysis proceeds by first computing the average entropy under the replica‑symmetric (RS) ansatz, then introducing a large‑deviation (or “tilted”) partition function Z(λ)=∑_σ e^{−λS(σ)}. By sending the conjugate parameter λ→∞, the dominant contribution comes from the σ that yields the smallest entropy, i.e., the worst‑case negation pattern. The authors solve the resulting self‑consistency equations both at the RS level and with a one‑step replica‑symmetry‑breaking (1‑RSB) correction, which becomes necessary for clause densities above the dynamical transition α_d.

The analytical outcome is strikingly simple: in the N→∞ limit the optimal adversarial strategy is to balance, for each variable, the number of times it appears negated and non‑negated across all clauses. In other words, the optimal σ is essentially “balanced” (each variable’s negation frequency ≈½). This balanced configuration maximally spreads constraints, thereby minimizing the number of satisfying assignments. The result holds irrespective of the clause density as long as the system remains in the satisfiable phase under random negations.

To validate the theory, extensive numerical experiments were performed on random 3‑SAT instances with sizes ranging from N=50 to N=500. Exact enumeration (for the smallest sizes) and heuristic searches (simulated annealing, genetic algorithms) were used to approximate the minimal entropy configuration. The data confirm the theoretical prediction for large N, but they also reveal strong pre‑asymptotic effects: for moderate system sizes the entropy reduction curve deviates noticeably from the asymptotic line, and the distinction between exponential (∝e^{cN}) and factorial (∝N!) growth regimes is blurred. This “pre‑asymptotic” regime reflects the fact that the combinatorial explosion of possible negation patterns makes the balanced configuration only approximately optimal until N becomes very large.

The paper concludes that adversarial SAT exemplifies how granting an adversary a modest amount of control (choice of negations) dramatically reshapes the solution space, creating a problem that is algorithmically far more challenging than ordinary SAT. The cavity‑method large‑deviation framework introduced here provides a systematic way to quantify such worst‑case scenarios and can be extended to other adversarial combinatorial problems such as graph coloring, clustering, or constraint satisfaction on hypergraphs. Future work suggested includes higher‑order replica‑symmetry‑breaking analyses, development of message‑passing algorithms tailored to adversarial settings, and exploration of finite‑size scaling to better understand the transition from pre‑asymptotic behavior to the thermodynamic limit.