Comparison Graphs: a Unified Method for Uniformity Testing

TL;DR

This paper introduces a unified framework using comparison graphs for uniformity testing, enabling the development of nearly-optimal testers under various computational constraints and simplifying existing methods.

Contribution

It formalizes comparison graphs for uniformity testing, providing a structural theorem and a generic method to create efficient testers under different computational models.

Findings

Developed a generalized collision-based tester using comparison graphs

Proved a structural theorem for good uniformity testers

Created nearly-optimal testers under various constraints

Abstract

Distribution testing can be described as follows: $q$ samples are being drawn from some unknown distribution $P$ over a known domain $[n]$. After the sampling process, a decision must be made about whether $P$ holds some property, or is far from it. The most studied problem in the field is arguably uniformity testing, where one needs to distinguish the case that $P$ is uniform over $[n]$ from the case that $P$ is $\epsilon$-far from being uniform (in $\ell_1$). In the classic model, it is known that $\Theta\left(\sqrt{n}/\epsilon^2\right)$ samples are necessary and sufficient for this task. This problem was recently considered in various restricted models that pose, for example, communication or memory constraints. In more than one occasion, the known optimal solution boils down to counting collisions among the drawn samples (each two samples that have the same value add one to the count), an idea that dates back to the first uniformity tester, and was coined the name "collision-based tester". In this paper, we introduce the notion of comparison graphs and use it to formally define a generalized collision-based tester. Roughly speaking, the edges of the graph indicate the tester which pairs of samples should be compared (that is, the original tester is induced by a clique, where all pairs are being compared). We prove a structural theorem that gives a sufficient condition for a comparison graph to induce a good uniformity tester. As an application, we develop a generic method to test uniformity, and devise nearly-optimal uniformity testers under various computational constraints. We improve and simplify a few known results, and introduce a new constrained model in which the method also produces an efficient tester. The idea behind our method is to translate computational constraints of a certain model to ones on the comparison graph, which paves the way to finding a good graph.