An affine-invariant property over a finite field is a property of functions over F n p that is closed under all affine transformations of the domain. This class of properties includes such well-known beasts as low-degree polynomials, polynomials that nontrivially factor, and functions of low spectral norm. The last few years has seen rapid progress in characterizing the affine-invariant properties which are testable with a constant number of queries. We survey the current state of this project.
What Is Property Testing?A scientific experiment takes as input an unknown object 3 , and it aims to determine whether or not a certain statement about the object holds true. Often, it's not feasible to decide whether the statement is exactly true or not, and the experimenter is satisfied with knowing whether the statement is "approximately true" for the object. The experiment consists of making certain kinds of measurements on the object, and one usually wants to minimize the number of measurements needed to reach a conclusion. Property testing is an algorithmic formalization of this basic scientific endeavor.Let O be a set of objects, and let P be the subset of O that satisfies a certain desirable property. Given an unknown object o ∈ O, we wish to know whether o is "close to being a member of P" by making a small number of "measurements" on o. To make this precise, introduce a distance function dist between pairs of objects in O and a query model that specifies what the possible queries (measurements) into o are and how each query reveals information about o. Then, for a given parameter ε > 0 that parameterizes the level of approximation and a positive integer q, an (ε, q)-tester for P is an algorithm 4 A that makes q queries into an unknown input o, outputs yes if o is in P and no if dist(o, o ) > ε for every o in P. That is, if the algorithm A outputs yes, then there is a guarantee that dist(o, o ) ε for some o in P, and hence, P approximately holds true for o. 1 . arnabb@csa.iisc.ernet.in.3 Well, not completely unknown. The experimenter usually already has some partial information about the object. 4 The algorithm is randomized, and the output guarantees hold with constant probability over the randomness of the algorithm. We give a precise definition later.