The Communication Complexity of the Coin Problem

Abstract

We study the two-player communication complexity of the \textsf{Coin} problem. Here, two players, denoted Alice and Bob, are given up to $n$ individual i.i.d. random coins, and asked to output whether there are more heads or tails. An immediate approach is for Alice to send Bob her exact numbers of heads and tails, yielding a deterministic protocol with one-way communication cost $O(\log n)$ bits. Indeed, this is provably optimal in the related multi-player and streaming settings [Braverman, Garg, Woodruff, FOCS'20]. Our contribution is to show that this is not the right answer in the one-way two-player setting. Instead, we prove a $Ω(\log\log n)$ bit lower bound for the constant error-rate regime, and a $Ω(\log(1/δ))$ bit lower bound for arbitrarily small error rates, $δ\geq n^{-0.1}$. We also complement these with a tight one-way protocol using $O(\log\log n + \log(1/δ))$ bits of communication. Unlike the previous works, our lower bounds avoid intricate direct information-theoretic arguments. Instead, our $Ω(\log(1/δ))$ bound is via a standard fooling-set style argument, whilst our $Ω(\log\log n)$ bound is via a novel yet simple reduction from the \textsf{Gap-Greater-Than} problem. This reduction is our main technical contribution, and can be seen as loosely inspired by the \emph{Robust Communication Complexity} arguments of [Chakrabarti, Cormode, McGregor, STOC'08]. Finally, we also show that our constant-error lower bound and protocol can be adapted to the multi-round setting: up to polynomial factors in $r$, the $r$-round communication complexity of the coin problem is $Θ\left((\log\log n)^{1/r}\right)$ bits.