[Paper Review] The random 2-SAT partition function

Dimitris Achlioptas, Amin Coja‐Oghlan+5 more|arXiv (Cornell University)|Feb 10, 2020

Constraint Satisfaction and Optimization48 references4 citations

TL;DR

This paper proves that the normalized logarithm of the number of satisfying assignments (the partition function) in a random 2-SAT formula converges in probability to an expression derived from the cavity method in statistical physics. The key result establishes that Belief Propagation correctly computes the marginal probabilities of variables under uniform random satisfying assignments, and the limiting partition function is given by the Bethe free energy functional applied to a solution of a stochastic fixed-point equation.

ABSTRACT

We show that throughout the satisfiable phase the normalised number of satisfying assignments of a random $2$-SAT formula converges in probability to an expression predicted by the cavity method from statistical physics. The proof is based on showing that the Belief Propagation algorithm renders the correct marginal probability that a variable is set to `true' under a uniformly random satisfying assignment.

Motivation & Objective

• To resolve the long-standing open question of the typical number of satisfying assignments in random 2-SAT formulas.
• To rigorously validate the cavity method's prediction for the 2-SAT partition function using probabilistic and combinatorial techniques.
• To establish that Belief Propagation computes the correct variable marginal probabilities under uniform random satisfying assignments in the satisfiable phase.
• To derive a precise asymptotic expression for the normalized logarithm of the partition function in terms of a stochastic fixed-point equation.

Proposed method

• Uses a stochastic fixed-point equation on the unit interval to model the message-passing dynamics of Belief Propagation in random 2-SAT formulas.
• Defines a Bethe free energy functional $\mathfrak{B}_\beta(\pi_d)$ that approximates the log-partition function and is shown to converge as $\beta \to \infty$.
• Introduces a tempered partition function $Z_\beta(\boldsymbol{\Phi})$ to control convergence and enable the use of concentration inequalities.
• Employs extremal boundary conditions tailored to the random formula structure to ensure convergence of Belief Propagation to correct marginals.
• Applies the monotone convergence theorem and coupling arguments to justify the limit $\mathfrak{B}_\infty(\pi_d)$ as the asymptotic value of $n^{-1}\mathbb{E}[\log Z(\boldsymbol{\Phi})]$.
• Uses concentration bounds and tail estimates to show that $n^{-1}\log Z(\boldsymbol{\Phi})$ concentrates around $\mathfrak{B}_\infty(\pi_d)$ with high probability.

Experimental results

Research questions

• RQ1 Does the cavity method’s prediction for the 2-SAT partition function hold with high probability in the satisfiable phase?
• RQ2 Can Belief Propagation be rigorously shown to compute the correct variable marginal probabilities in random 2-SAT?
• RQ3 Is the limiting value of the normalized log-partition function given by the Bethe free energy functional applied to a solution of the stochastic fixed-point equation?
• RQ4 What is the precise asymptotic behavior of the number of satisfying assignments in a random 2-SAT formula with clause-to-variable ratio below the satisfiability threshold?

Key findings

• The normalized log-partition function $n^{-1}\log Z(\boldsymbol{\Phi})$ converges in probability to $\mathfrak{B}_\infty(\pi_d)$, the limit of the Bethe free energy functional as $\beta \to \infty$.
• Belief Propagation correctly computes the marginal probability that a variable is set to 'true' under a uniformly random satisfying assignment, even under hard constraints.
• The limiting expression $\mathfrak{B}_\infty(\pi_d)$ is given by $\mathbb{E}\left[\log\left(\prod_{i=1}^{\boldsymbol{d}^-}\boldsymbol{\mu}_{\pi_d,i} + \prod_{i=1}^{\boldsymbol{d}^+}\boldsymbol{\mu}_{\pi_d,i+\boldsymbol{d}^-}\right) - \frac{d}{2}\log(1 - \boldsymbol{\mu}_{\pi_d,1}\boldsymbol{\mu}_{\pi_d,2})\right]$, where $\pi_d$ solves the stochastic fixed-point equation.
• The convergence is sharp: for any $\varepsilon > 0$, $\mathbb{P}[n^{-1}\log(Z(\boldsymbol{\Phi}) \vee 1) > \mathfrak{B}_\infty(\pi_d) + \varepsilon] = o(1)$ as $n \to \infty$.
• The result confirms the physics-based cavity method prediction for the 2-SAT partition function, resolving a long-standing open problem in random constraint satisfaction.