The Approximability of Learning and Constraint Satisfaction Problems

An α-approximation algorithm is an algorithm guaranteed to output a solutionthat is within an α ratio of the optimal solution. We are interested in thefollowing question: Given an NP-hard optimization problem, what is the bestapproximation guarantee that any polynomial time algorithm could achieve? We mostly focus on studying the approximability of two classes of NP-hardproblems: Constraint Satisfaction Problems (CSPs) and Computational Learning Problems. For CSPs, we mainly study the approximability of MAX CUT, MAX 3-CSP,MAX 2-LINR, VERTEX-PRICING, as well as serval variants of the UNIQUEGAMES.• The problem of MAX CUT is to find a partition of a graph so as to maximizethe number of edges between the two partitions. Assuming theUnique Games Conjecture, we give a complete characterization of the approximationcurve of the MAX CUT problem: for every optimum value ofthe instance, we show that certain SDP algorithm with RPR2 roundingalways achieve the optimal approximation curve.• The input to a 3-CSP is a set of Boolean constraints such that each constraintcontains at most 3 Boolean variables. The goal is to find an assignmentto these variables to maximize the number of satisfied constraints.We are interested in the case when a 3-CSP is satisfiable, i.e.,there does exist an assignment that satisfies every constraint. Assumingthe d-to-1 conjecture (a variant of the Unique Games Conjecture), weprove that it is NP-hard to give a better than 5/8-approximation for theproblem. Such a result matches a SDP algorithm by Zwick which givesa 5/8-approximation problem for satisfiable 3-CSP. In addition, our resultalso conditionally resolves a fundamental open problem in PCP theory onthe optimal soundness for a 3-query nonadaptive PCP system for NP withperfect completeness.• The problem of MAX 2-LINZ involves a linear systems of integer equations;these equations are so simple such that each equation contains atmost 2 variables. The goal is to find an assignment to the variables so asto maximize the total number of satisfied equations. It is a natural generalizationof the Unique Games Conjecture which address the hardness ofthe same equation systems over finite fields. We show that assuming theUnique Games Conjecture, for a MAX 2-LINZ instance, even that thereexists a solution that satisfies 1−ε of the equations, it is NP-hard to findone that satisfies ² of the equations for any ε > 0.