New bounds for the Moser-Tardos distribution (1610.09653v7)

Abstract: The Lovasz Local Lemma (LLL) is a probabilistic tool which has been used to show the existence of a variety of combinatorial structures with good "local" properties. The "LLL-distribution" can be used to show that the resulting structures have good global properties in expectation. The simplest, variable-based setting of the LLL was covered by the seminal algorithm of Moser & Tardos (2010). This has since been extended to other probability spaces including random permutations. One can similarly define an "MT-distribution" for these algorithms, that is, the distribution of the configuration they produce. Haeupler et al. (2011) showed bounds on the MT-distribution which essentially match the LLL-distribution for the variable-assignment setting; Harris & Srinivasan showed similar results for the permutation setting. In this work, we show new bounds on the MT-distribution which are significantly stronger than those known to hold for the LLL-distribution. In the variable-assignment setting, we show a tighter bound on the probability of a disjunctive event or singleton event. As a consequence, in $k$-SAT instances with bounded variable occurrence, the MT-distribution satisfies an $\epsilon$-approximate independence condition asymptotically stronger than the LLL-distribution. We use this to show a nearly tight bound on the minimum implicate size of a CNF boolean formula. Another noteworthy application is constructing independent transversals which avoid a given subset of vertices; this provides a constructive analogue to a result of Rabern (2014). In the permutation LLL setting, we show a new type of bound which is similar to the cluster-expansion LLL criterion of Bissacot et al. (2011), but is stronger and takes advantage of the extra structure in permutations. We illustrate with improved bounds on weighted Latin transversals and partial Latin transversals.