TCSLib

85 Communication Complexity — Derandomization

85.1 Overview

This module establishes the Chernoff-plus-union-bound derandomization step at the heart of Newman’s theorem. Starting from a public-coin protocol that \(\varepsilon \)-computes a function \(f : X \times Y \to \alpha \), it produces \(t = O\! \left(\log (|X|\cdot |Y|)/((c-1)^2\varepsilon ^2)\right)\) explicit random seeds that collectively cover every input: for every \((x,y)\), at most a \(c\varepsilon \) fraction of the seeds cause an incorrect output.

85.2 Declarations

Theorem 85.1 Hoeffding tail bound for many events

Let \((\Omega ', \mu )\) be a probability space and let \(Y_0, \dots , Y_{t-1} : \Omega ' \to \mathbb {R}\) be mutually independent, a.e. \([0,1]\)-valued random variables satisfying \(\mathbb {E}[Y_i] \le \varepsilon \) for all \(i\), with \(\varepsilon \ge 0\) and \(c \gt 1\). Then

\[ \mu \! \left\{ \omega \; \middle |\; c\, \varepsilon \, t \le \sum _{i=0}^{t-1} Y_i(\omega )\right\} \; \le \; \exp \! \left(-2(c-1)^2\varepsilon ^2 t\right). \]
Definition 85.2 Derandomization sample count

Given finite types \(X\) and \(Y\) and real parameters \(\varepsilon , c\), the number of random samples required for derandomization via Chernoff and union bound is

\[ t(X, Y, \varepsilon , c) \; =\; \left\lceil \frac{\log (|X|\cdot |Y|)}{2\, (c-1)^2\, \varepsilon ^2} \right\rceil _{\! +} + 1, \]

where \(\lceil \cdot \rceil _{+}\) denotes the natural-number ceiling.

Let \(\Omega \) be a finite probability space, let \(p\) be a public-coin protocol on \(X \times Y\) with output type \(\alpha \), and let \(f : X \to Y \to \alpha \). Suppose \(c \gt 1\) and \(p\) \(\varepsilon \)-computes \(f\) (i.e. for every \((x, y)\), the probability that \(p\) produces an incorrect output is at most \(\varepsilon \)). Then there exist \(t = \texttt{CommunicationComplexity.PublicCoin.FiniteMessage.Protocol.derandomizationSamples}(X, Y, \varepsilon , c)\) random seeds \(\omega _0, \dots , \omega _{t-1} \in \Omega \) such that for every \((x, y)\),

\[ \frac{\# \{ \, i \mid p.\text{rrun}(x, y, \omega _i) \ne f(x, y)\, \} }{t} \; \le \; c\, \varepsilon . \]