Characterizations of Sparsifiability for Affine CSPs and Symmetric CSPs

CSP sparsification, introduced by Kogan and Krauthgamer (ITCS 2015), considers the following question: when can an instance of a constraint satisfaction problem be sparsified (by retaining a weighted subset of the constraints) while still roughly capturing the weight of constraints satisfied by {\em every} assignment. CSP sparsification generalizes and abstracts other commonly studied problems including graph cut-sparsification, hypergraph cut-sparsification and hypergraph XOR-sparsification. A central question here is to understand what properties of a constraint predicate $P:\Sigma^r \to {0,1}$ (where variables are assigned values in $\Sigma$) allow for nearly linear-size sparsifiers (in the number of variables). In this work (1) we significantly extend the class of CSPs for which nearly linear-size, and other non-trivial, sparsifications exist and give classifications in some broad settings and (2) give a polynomial-time algorithm to extract this sparsification. Our results captured in item (1) completely classify all symmetric Boolean predicates $P$ (i.e., on the Boolean domain $\Sigma = {0,1}$) that allow nearly-linear-size sparsifications. Symmetric Boolean CSPs already capture all the special classes of sparisifcation listed above including hypergraph cut-sparsification and variants. Our study of symmetric CSPs reveals an inherent, previously undetected, number-theoretic phenomenon that determines near-linear size sparsifiability. We also completely classify the set of Boolean predicates $P$ that allow non-trivial ($o(n^r)$-size) sparsifications, thus answering an open question from the work of Kogan and Krauthgamer.