Sharp indistinguishability bounds from non-uniform approximations

TL;DR

This paper establishes precise bounds on the ability to distinguish between symmetric distributions over bits using limited samples, employing novel non-uniform approximation techniques to improve upon previous results.

Contribution

It introduces sharp bounds on statistical distinguishability for all sample sizes by developing non-uniform approximation methods, extending prior uniform approximation approaches.

Findings

Sharp upper and lower bounds for all sample sizes

Non-uniform approximations outperform uniform ones

Improved understanding of distribution distinguishability

Abstract

We study the problem of distinguishing between two symmetric probability distributions over $n$ bits by observing $k$ bits of a sample, subject to the constraint that all $k-1$-wise marginal distributions of the two distributions are identical to each other. Previous works of Bogdanov et al. and of Huang and Viola have established approximately tight results on the maximal statistical distance when $k$ is at most a small constant fraction of $n$ and Naor and Shamir gave a tight bound for all $k$ in the case of distinguishing with the OR function. In this work we provide sharp upper and lower bounds on the maximal statistical distance that holds for all $k$. Upper bounds on the statistical distance have typically been obtained by providing uniform low-degree polynomial approximations to certain higher-degree polynomials; the sharpness and wider applicability of our result stems from the construction of suitable non-uniform approximations.