Reconstruction and clustering for a class of constraint satisfaction problem

Random instances of a huge class of Constraint Satisfaction Problems (CSP’s) appear to be hard for all known algorithms, when the number of constraints per variable lies in a certain interval.
Contributing to the general understanding of the structure of the solution space of a CSP in the satisfiable regime, we formulate a set of analytical conditions on a large family of (random) CSP’s, and prove bounds on three most interesting thresholds for the density of such an ensemble: namely, the satisfiability threshold, the threshold for clustering of the solution space, and the threshold for an appropriate reconstruction problem on the CSP’s. The bounds become asymptotically tight as the number of degrees of freedom in each clause diverges. The families are general enough to include commonly studied problems such as, random instances of Not-All-Equal-SAT, k-XOR formulae, hypergraph 2-coloring, and graph k-coloring. An important new ingredient is a condition involving the Fourier expansion of clauses, which characterizes the class of problems with a similar threshold structure. (Joint work with Andrea Montanari and Prasad Tetali).

Speaker Details

Ricardo Restrepo is a PhD student at the Mathematics Department at Georgia Tech under the supervision of Prasad Tetali. His education was focused on (classical) probability theory and stochastic calculus, and is currently interested in the threshold phenomena for discrete and continuous structures, particularly the relation between spatial and dynamic properties of spin systems.

Date:
Speakers:
Ricardo Restrepo
Affiliation:
GA Tech, Mathematics Dept