Minimum Rates of Approximate Sufficient Statistics

TL;DR

This paper introduces the concept of approximate sufficient statistics, demonstrating that allowing small estimation errors can significantly reduce the code length needed to represent data from parametric distributions.

Contribution

It develops a Shannon-theoretic framework for approximate sufficient statistics, providing bounds on code length reductions and establishing strong converses for various error measures.

Findings

Code length can be reduced to (d/2) log n + O(1) with approximate statistics.

Established bounds using relative entropy and variational distance.

Proved strong converses even with non-vanishing errors.

Abstract

Given a sufficient statistic for a parametric family of distributions, one can estimate the parameter without access to the data. However, the memory or code size for storing the sufficient statistic may nonetheless still be prohibitive. Indeed, for $n$ independent samples drawn from a $k$-nomial distribution with $d=k-1$ degrees of freedom, the length of the code scales as $d\log n+O(1)$. In many applications, we may not have a useful notion of sufficient statistics (e.g., when the parametric family is not an exponential family) and we also may not need to reconstruct the generating distribution exactly. By adopting a Shannon-theoretic approach in which we allow a small error in estimating the generating distribution, we construct various {\em approximate sufficient statistics} and show that the code length can be reduced to $\frac{d}{2}\log n+O(1)$. We consider errors measured according to the relative entropy and variational distance criteria. For the code constructions, we leverage Rissanen's minimum description length principle, which yields a non-vanishing error measured according to the relative entropy. For the converse parts, we use Clarke and Barron's formula for the relative entropy of a parametrized distribution and the corresponding mixture distribution. However, this method only yields a weak converse for the variational distance. We develop new techniques to achieve vanishing errors and we also prove strong converses. The latter means that even if the code is allowed to have a non-vanishing error, its length must still be at least $\frac{d}{2}\log n$.