We improve both upper and lower bounds for the distributionfree testing of monotone conjunctions. Given oracle access to an unknown Boolean function f : {0, 1} n → {0, 1} and sampling oracle access to an unknown distribution D over {0, 1} n , we present anÕ(n 1/3 /ǫ 5 )-query algorithm that tests whether f is a monotone conjunction versus ǫ-far from any monotone conjunction with respect to D. This improves the previous best upper bound ofÕ(n 1/2 /ǫ) by Dolev and Ron [DR11], when 1/ǫ is small compared to n. For some constant ǫ0 > 0, we also prove a lower bound ofΩ(n 1/3 ) for the query complexity, improving the previous best lower bound ofΩ(n 1/5 ) by Glasner and Servedio [GS09]. Our upper and lower bounds are tight, up to a polylogarithmic factor, when the distance parameter ǫ is a constant. Furthermore, the same upper and lower bounds can be extended to the distribution-free testing of general conjunctions, and the lower bound can be extended to that of decision lists and linear threshold functions.
IntroductionThe field of property testing analyzes the resources an algorithm requires to determine whether an unknown object satisfies a certain property versus far from satisfying the property. It was introduced in [RS96], after prior work in [BFL91,BLR93], and has been studied extensively during the past two decades (e.g., see surveys in [Gol98,Fis01,Ron01,AS05,Rub06]). For the purpose of this paper we consider a Boolean function f : {0, 1} n → {0, 1}, and a class C of Boolean functions, viewed as a property. The distance between f and C in the standard testing model is measured with respect to the uniform distribution. Equivalently, it is the smallest fraction of entries of f one needs to flip to make it a member of C. A natural generalization of the standard model, called distribution-free property testing, was first introduced by Goldreich, Goldwasser and Ron [GGR98] and has been studied in [AC06,HK07,HK08a, HK08b,GS09,DR11].For the distribution-free model, there is an unknown distribution D over {0, 1} n in addition to the unknown f . The goal of an algorithm is to determine whether f is in C versus far from C with respect to D, given both blackbox access to f and sampling access to D. The model of distribution-free property testing is well motivated by scenarios where the distance being of interest is measured * Supported in part by NSF grants CCF-1149257 and CCF-1423100.† xichen@cs.columbia.edu, Columbia University ‡ jinyu@cs.columbia.edu, Columbia University with respect to an unknown distribution D. It is also inspired by similar models in computational learning theory (e.g., the distribution-free PAC learning model [Val84] with membership queries). It was observed [GGR98] that any proper distribution-free PAC learning algorithm can be used for distribution-free property testing.In this paper we study the distribution-free testing of monotone conjunctions (or monotone monomials): f is a monotone conjunction if f (z) = i∈S z i , for some S ⊆ [n]. We first obtain an efficient algorithm that is one-s...