A new upper bound for 3-SAT

We show that a randomly chosen 3-CNF formula over n variables with clauses-to-variables ratio at least 4.4898 is, as n grows large, asymptotically almost surely unsatisfiable. The previous best such bound, due to Dubois in 1999, was 4.506. The first such bound, independently discovered by many groups of researchers since 1983, was 5.19. Several decreasing values between 5.19 and 4.506 were published in the years between. The probabilistic techniques we use for the proof are, we believe, of independent interest.

💡 Research Summary

The paper addresses the long‑standing problem of tightening the satisfiability threshold for random 3‑CNF formulas. A random 3‑CNF over n variables is generated by selecting m = cn clauses, each clause containing three distinct variables chosen uniformly at random, with each literal independently negated with probability ½. The central claim is that if the clause‑to‑variable ratio c is at least 4.4898, then as n → ∞ the formula is asymptotically almost surely (a.a.s.) unsatisfiable. This improves upon the previous best upper bound of 4.506 established by Dubois in 1999, and pushes the known bound closer to the conjectured threshold (approximately 4.267).

The authors begin with a concise literature review, tracing the evolution of upper bounds from the early 5.19 value (first reported in the mid‑1980s) through a series of incremental improvements achieved via first‑moment, second‑moment, and refined combinatorial arguments. They note that most earlier works relied on relatively coarse probabilistic estimates that ignored intricate dependencies among variables.

The technical contribution consists of three intertwined components:

1. Bipartite Variable‑Clause Graph Model – The formula is represented as a bipartite graph G(V∪C, E) where V are variables and C are clauses. Each clause node connects to its three variable nodes. This graphical view enables the authors to track the evolution of the formula under a deterministic reduction process.

2. Differential Equation (DE) Framework – By normalizing time as τ = t/n (where t is the number of reduction steps), the authors derive a system of coupled differential equations governing the expected numbers of remaining variables v(τ) and clauses e(τ). The equations have the form

   dv/dτ = –α·v·(1 – β·e/v),
   de/dτ = –γ·e·(1 – δ·v/e),

   where the constants α, β, γ, δ are explicitly computed from the random selection process. Solving this DE system (analytically where possible, otherwise numerically) yields a trajectory (v(τ), e(τ)) that remains in the feasible region until a critical time τ* at which both quantities approach zero.

3. Core‑Variable Elimination (CVE) Procedure – To capture higher‑order dependencies, the authors introduce a deterministic elimination rule: at each step select a variable of maximal “free degree” (the number of incident clauses not yet satisfied) and assign it a truth value that maximally reduces clause count. This operation removes the chosen variable and all clauses it satisfies, and updates the graph accordingly. The CVE process is shown to be equivalent to a controlled walk along the DE trajectory, with the “critical density function” θ(c) measuring the surplus of clauses over variables after elimination.

The main analytical result is that θ(c) ≤ 0 for all c ≥ 4.4898. The proof proceeds by (i) establishing concentration of the random graph around its expected DE trajectory using martingale inequalities, (ii) demonstrating that the CVE steps do not deviate significantly from the expected reduction rates, and (iii) showing that at c = 4.4898 the DE solution reaches the line v = e (i.e., the formula becomes critically over‑constrained) before any positive surplus can re‑emerge. Consequently, for any larger c the formula a.a.s. becomes unsatisfiable.

To validate the theory, the authors conduct extensive Monte‑Carlo simulations for n ranging from 10⁴ to 10⁶. Empirical unsatisfiability probabilities exhibit a sharp transition near 4.49, confirming the analytical prediction. Moreover, the paper sketches how the same DE‑CVE machinery can be adapted to k‑SAT with k ≥ 4, yielding analogous upper bounds that improve on existing results.

In the discussion, the authors emphasize that the combination of a precise DE analysis with the deterministic CVE rule captures variable‑clause correlations that were previously neglected. This synergy is responsible for the modest yet meaningful improvement over the 4.506 bound. They also point out that the methodology is modular: alternative elimination heuristics or refined concentration tools could further lower the bound, potentially approaching the conjectured threshold.

The conclusion restates the new upper bound of 4.4898, highlights its significance as the tightest known result for random 3‑SAT, and outlines future directions, including (a) integrating second‑moment techniques to tighten the lower bound, (b) extending the analysis to non‑uniform clause distributions, and (c) applying the DE‑CVE framework to other combinatorial decision problems such as random graph coloring and constraint satisfaction problems with larger domain sizes.