Estimating Satisfiability

The problem of estimating the proportion of satisfiable instances of a given CSP (constraint satisfaction problem) can be tackled through weighting. It consists in putting onto each solution a non-negative real value based on its neighborhood in a way that the total weight is at least 1 for each satisfiable instance. We define in this paper a general weighting scheme for the estimation of satisfiability of general CSPs. First we give some sufficient conditions for a weighting system to be correct. Then we show that this scheme allows for an improvement on the upper bound on the existence of non-trivial cores in 3-SAT obtained by Maneva and Sinclair (2008) to 4.419. Another more common way of estimating satisfiability is ordering. This consists in putting a total order on the domain, which induces an orientation between neighboring solutions in a way that prevents circuits from appearing, and then counting only minimal elements. We compare ordering and weighting under various conditions.

💡 Research Summary

The paper “Estimating Satisfiability” develops a unified framework for estimating the proportion of satisfiable instances of constraint satisfaction problems (CSPs) by means of weighting, and it systematically compares this approach with the more traditional ordering (or “solution selection”) technique. The authors start from the first‑moment method, where one needs a non‑negative random variable X that is at least 1 on every satisfiable instance; the expected value of X then yields an upper bound on the satisfiability probability via Markov’s inequality. The naïve choice X = number of solutions is useless because it is typically huge, leading to very loose bounds. To overcome this, the paper proposes to assign a non‑negative weight to each solution, designed so that the total weight of all solutions of any satisfiable instance is at least 1.

The core of the weighting scheme is built from three components: (1) a weighting seed s_F(x,a) that assigns a base weight to each variable‑value pair; (2) a weight generator ω_F(x,a,Δ) that, given a variable x, its current value a, and the set Δ of values taken by the variable among the solutions adjacent to a given solution σ, produces the actual weight contributed by x in σ; (3) a dispatcher d_F(x,a) that redistributes the seed weight of forbidden values to allowed values. When the seed is unitary (for each variable x, Σ_a s_F(x,a)=1), the “unladen” weight U_F(v)=∏_x s_F(x,v(x)) summed over all full assignments v equals exactly 1. The actual weight of a solution σ is W_F(σ)=∏_x ω_F(x,σ(x),A_F(σ,x)), where A_F(σ,x) is the set of values taken by x in the x‑adjacent clique of σ. The authors introduce a decomposition family δ_F,σ,x,a such that ω_F can be expressed as a sum over a∈D of δ_F,σ,x,a; this leads to a transfer function T_F,σ→v=∏_x δ_F,σ,x,v(x).

The Weight Conservation Theorem states that if (i) the seed is unitary, (ii) the actual weight is decomposable via δ_F, and (iii) for a chosen set S of solutions the collection of transfers {T_F,σ→v | σ∈S} covers the unladen weight (i.e., Σ_{σ∈S} T_F,σ→v ≥ U_F(v) for every full assignment v), then the total weight of S satisfies W_F(S)≥1. Consequently, taking X = W_F(S) yields a correct first‑moment estimator. The theorem provides three sufficient conditions; the paper also shows that they are not necessary by constructing counter‑examples.

Using this general machinery, the authors design two concrete weighting schemes. The first is a generic construction applicable to any CSP: choose any unitary seed, define a dispatcher that spreads the weight of forbidden values proportionally among allowed values, and obtain ω_F via the formula (7) in the paper. The second scheme refines the weighting used by Maneva and Sinclair (2008) for 3‑SAT. By extending the domain to three values (true, false, “*”) and carefully selecting seed and dispatcher functions that respect the three conditions of the Weight Conservation Theorem, the authors improve the upper bound on the existence of non‑trivial cores in random 3‑SAT from 4.453 (Maneva‑Sinclair) to 4.419. The proof reuses the combinatorial analysis of Maneva‑Sinclair, showing that the improvement stems solely from the more efficient weighting.

The paper then turns to ordering (solution selection). An ordering imposes a total order on the domain D, which induces an orientation on each edge of the solution network (two solutions are adjacent if they differ on a single variable). By orienting edges from the larger to the smaller value according to the domain order, the resulting directed graph is acyclic; minimal solutions (those with no outgoing edge) are counted. This method also respects the locality condition because the orientation of an edge depends only on the two adjacent solutions.

A systematic comparison between weighting and ordering is carried out under three scenarios:

1. Instance‑dependent weighting: when the seed and dispatcher may vary from instance to instance. The authors prove (Theorem 32) that in this most general setting a well‑chosen ordering can dominate any such weighting; i.e., there exists an ordering that yields a smaller (or equal) first‑moment bound than any instance‑dependent weighting scheme.

2. Homogeneous weighting (seed = dispatcher, same for all instances) and instance‑independent ordering (the domain order does not depend on the particular instance). For families of instances closed under value renaming (which includes essentially all standard random CSP models), the paper shows (Theorem 38) that the expected weight of a random solution under the homogeneous weighting equals the expected number of minimal solutions under the ordering. Hence, on average the two methods are equivalent.

3. Hybrid cases where some structural parameters (e.g., variable occurrence frequencies, clause‑variable ratios) guide the design of the seed or the order. The authors discuss how such syntactic information can be exploited to tailor a weighting that outperforms a naïve ordering, but they also demonstrate that the advantage disappears when the weighting is forced to be homogeneous.

Throughout, the authors emphasize the locality condition: both weighting and ordering must be computable using only the immediate neighborhood of a solution (i.e., solutions that differ on a single variable). This reflects realistic algorithmic constraints where each “processor” (representing a solution) only knows its adjacent solutions.

In summary, the paper contributes (i) a rigorous, general framework for constructing correct weighting schemes via the Weight Conservation Theorem, (ii) an improved bound for non‑trivial cores in random 3‑SAT using a refined weighting, and (iii) a nuanced comparison showing when weighting is strictly stronger, when it is equivalent, and when ordering can dominate. The results clarify the relationship between two major techniques for estimating CSP satisfiability and provide concrete guidelines for researchers designing probabilistic analyses of random CSP models.