A new look at survey propagation and its generalizations

  • Authors:
  • Elitza Maneva;Elchanan Mossel;Martin J. Wainwright

  • Affiliations:
  • UC Berkeley, CA;UC Berkeley, CA;UC Berkeley, CA

  • Venue:
  • SODA '05 Proceedings of the sixteenth annual ACM-SIAM symposium on Discrete algorithms
  • Year:
  • 2005

Quantified Score

Hi-index 0.06

Visualization

Abstract

We study the survey propagation algorithm [19, 5, 4], which is an iterative technique that appears to be very effective in solving random k-SAT problems even with densities close to threshold. We first describe how any SAT formula can be associated with a novel family of Markov random fields (MRFs), parameterized by a real number ρ. We then show that applying belief propagation---a well-known "message-passing technique---to this family of MRFs recovers various algorithms, ranging from pure survey propagation at one extreme (ρ = 1) to standard belief propagation on the uniform distribution over SAT assignments at the other extreme (ρ = 0). Configurations in these MRFs have a natural interpretation as generalized satisfiability assignments, on which a partial order can be defined. We isolate cores as minimal elements in this partial ordering, and prove that any core is a fixed point of survey propagation. We investigate the associated lattice structure, and prove a weight-preserving identity that shows how any MRF with p 0 can be viewed as a "smoothed" version of the naive factor graph representation of the k-SAT problem (p = 0). Our experimental results show that message-passing on our family of MRFs is most effective for values of ρ ≠ 1 (i.e., distinct from survey propagation); moreover, they suggest that random formulas may not typically possess non-trivial cores. Finally, we isolate properties of Gibbs sampling and message-passing algorithms that are typical for an ensemble of k-SAT problems. We prove that the space of cores for random formulas is highly disconnected, and show that for values of p sufficiently close to one, either the associated MRF is highly concentrated around the all-star assignment, or it has exponentially small conductance. Similarly, we prove that for p sufficiently close to one, the all-star assignment is attractive for message-passing when analyzed in the density-evolution setting.