The communication complexity of distributed estimation

TL;DR

This paper investigates the communication complexity involved in distributed estimation of expected function values between two parties, introducing new protocols and bounds that improve efficiency and establish optimality for various function classes.

Contribution

It presents a novel debiasing protocol reducing communication dependence on error parameter, and provides tight bounds and lower bound techniques for specific functions like equality and greater-than.

Findings

The debiasing protocol achieves linear dependence on 1/ε, improving over quadratic.

Upper bounds are improved for equality and greater-than functions.

Many protocols are proven to be optimal through spectral and discrepancy-based lower bounds.

Abstract

We study an extension of the standard two-party communication model in which Alice and Bob hold probability distributions $p$ and $q$ over domains $X$ and $Y$, respectively. Their goal is to estimate \[ \mathbb{E}_{x \sim p,\, y \sim q}[f(x, y)] \] to within additive error $\varepsilon$ for a bounded function $f$, known to both parties. We refer to this as the distributed estimation problem. Special cases of this problem arise in a variety of areas including sketching, databases and learning. Our goal is to understand how the required communication scales with the communication complexity of $f$ and the error parameter $\varepsilon$. The random sampling approach -- estimating the mean by averaging $f$ over $O(1/\varepsilon^2)$ random samples -- requires $O(R(f)/\varepsilon^2)$ total communication, where $R(f)$ is the randomized communication complexity of $f$. We design a new debiasing protocol which improves the dependence on $1/\varepsilon$ to be linear instead of quadratic. Additionally we show better upper bounds for several special classes of functions, including the Equality and Greater-than functions. We introduce lower bound techniques based on spectral methods and discrepancy, and show the optimality of many of our protocols: the debiasing protocol is tight for general functions, and that our protocols for the equality and greater-than functions are also optimal. Furthermore, we show that among full-rank Boolean functions, Equality is essentially the easiest.