Abstract

We unify the treatment of asymmetry of information in theoretical computer science and economics. We put forward a new type of proof system, where an unbounded Prover and a polynomial time Verifier interact, on inputs a string x and a function f, so that the Verifier may learn f(x). In our setting there no longer are “good” or “malicious” Provers, but only RATIONAL ones. In essence, (1) the Verifier gives the Prover a reward in [0,1] determined by the transcript of their interaction; (2) the Prover wishes to maximize his expected reward; and (3) the reward is maximized only if the verifier correctly learns f(x). Rational proofs are as powerful as interactive ones, but can be amazingly more efficient in the amount of communication involved: that is, the number of bits exchanged and the number of rounds of interaction. Joint work with Pablo Azar.