Counting with Combined Splitting and Capture-Recapture Methods

We apply the splitting method to three well-known counting problems, namely 3-SAT, random graphs with prescribed degrees, and binary contingency tables. We present an enhanced version of the splitting method based on the capture-recapture technique, and show by experiments the superiority of this technique for SAT problems in terms of variance of the associated estimators, and speed of the algorithms.

💡 Research Summary

The paper presents a novel combination of two stochastic estimation techniques—splitting and capture‑recapture—to approximate the counting of solutions for #P‑complete problems. The splitting method, originally devised for rare‑event simulation, decomposes the original solution set X* into a nested sequence of subsets X₀ ⊃ X₁ ⊃ … ⊃ Xₘ, where Xₘ = X*. At each level t the algorithm estimates the conditional probability cₜ = |Xₜ| / |Xₜ₋₁| by generating N uniformly distributed points in Xₜ₋₁, selecting the top ρ‑fraction (the elite set) according to a score function S(x) that counts satisfied constraints, and then using a Markov‑chain Monte‑Carlo (MCMC) sampler (typically a Gibbs sampler) to produce uniform samples in Xₜ. The product of the estimated ratios ∏̂cₜ multiplied by the known size of X₀ yields an estimator for |X*|. This adaptive splitting scheme avoids the exponential sample size required by naïve Monte‑Carlo when p = |X*|/|X₀| is extremely small.

The second component is the classic capture‑recapture (CAP‑RECAP) technique, originally used in ecology to estimate animal populations. After the splitting algorithm reaches the final level Xₘ = X*, the authors draw two independent subsamples of sizes N₁ and N₂ from the elite points, count the number R of points appearing in both subsamples, and compute the bias‑corrected estimator
  \hat M = (N₁+1)(N₂+1)/(R+1).
The variance of this estimator can be approximated analytically, and it is typically much smaller than the variance of the product estimator from splitting alone, especially when the total number of solutions lies in the range 10⁶–10⁹ and the overall sample budget N is limited (e.g., N ≈ 10⁴).

For extremely large solution spaces (|X*| > 10⁹) the authors propose an “extended capture‑recapture” scheme. Additional artificial constraints (auxiliary clauses) are introduced to shrink the feasible region, making the solution count manageable for capture‑recapture. After estimating the reduced count, a correction factor based on the number of added constraints restores an estimate for the original problem. This approach dramatically reduces variance compared with a crude Monte‑Carlo estimator for huge spaces.

The experimental evaluation covers three representative #P‑complete counting problems: (1) random 3‑SAT instances with 100–300 variables, (2) random graphs with prescribed degree sequences, and (3) binary contingency tables with given row/column sums. For each problem the authors compare three algorithms: (a) basic splitting, (b) splitting combined with capture‑recapture, and (c) the extended capture‑recapture for the largest instances. Results show that (i) the combined method achieves a 30–70 % reduction in relative standard error while incurring negligible extra runtime (the capture‑recapture step is just a set‑intersection), (ii) the extended method maintains a relative error below 0.22 even when |X*| exceeds 10⁹, and (iii) the splitting‑only estimator suffers from exploding variance as the solution space grows.

Key insights include: (1) splitting efficiently “amplifies” rare events by breaking the problem into a sequence of easier conditional estimations, (2) capture‑recapture leverages the final set of elite samples to correct bias and reduce variance without additional costly simulations, and (3) the combination is particularly effective for #SAT where the solution space grows exponentially and traditional MCMC‑based counting is memory‑intensive. The paper also discusses practical considerations such as the choice of elite fraction ρ (values between 0.1 and 0.2 work well), the required length of MCMC chains (≈20–30 sweeps were sufficient for convergence in the experiments), and the need to remove duplicate samples before applying capture‑recapture.

Limitations are acknowledged: the performance depends on proper mixing of the MCMC sampler; correlations between successive samples can introduce bias into the capture‑recapture estimate; and theoretical guidance for optimal ρ and N is still lacking. Future work is suggested on adaptive parameter tuning, more advanced MCMC techniques (e.g., Hamiltonian Monte‑Carlo), and multi‑stage capture‑recapture schemes to further improve accuracy.

Overall, the paper demonstrates that integrating splitting with capture‑recapture yields a practical, low‑variance, and computationally efficient framework for approximate counting in several important #P‑complete domains, and it opens avenues for further methodological refinements.