The Complexity of Agreement

TL;DR

This paper demonstrates that rational Bayesian agents can reach agreement efficiently and with limited computational resources, strengthening Aumann's theorem by providing bounds on communication and computation needed for agreement.

Contribution

It establishes bounds on the amount of communication required for agents to agree within epsilon, and introduces practical protocols that can be simulated efficiently.

Findings

Agreement within epsilon can be achieved with O(1/epsilon^2) bits exchanged.

Standard protocols nearly saturate the communication bound.

Agents can be simulated with limited computational resources, indistinguishable from perfect Bayesians.

Abstract

A celebrated 1976 theorem of Aumann asserts that honest, rational Bayesian agents with common priors will never "agree to disagree": if their opinions about any topic are common knowledge, then those opinions must be equal. Economists have written numerous papers examining the assumptions behind this theorem. But two key questions went unaddressed: first, can the agents reach agreement after a conversation of reasonable length? Second, can the computations needed for that conversation be performed efficiently? This paper answers both questions in the affirmative, thereby strengthening Aumann's original conclusion. We first show that, for two agents with a common prior to agree within epsilon about the expectation of a [0,1] variable with high probability over their prior, it suffices for them to exchange order 1/epsilon^2 bits. This bound is completely independent of the number of bits n of relevant knowledge that the agents have. We then extend the bound to three or more agents; and we give an example where the economists' "standard protocol" (which consists of repeatedly announcing one's current expectation) nearly saturates the bound, while a new "attenuated protocol" does better. Finally, we give a protocol that would cause two Bayesians to agree within epsilon after exchanging order 1/epsilon^2 messages, and that can be simulated by agents with limited computational resources. By this we mean that, after examining the agents' knowledge and a transcript of their conversation, no one would be able to distinguish the agents from perfect Bayesians. The time used by the simulation procedure is exponential in 1/epsilon^6 but not in n.