A one-round, two-prover, zero-knowledge protocol for NP

The model of zero-knowledge multi-prover interactive proofs was introduced by Ben-Or, Goldwasser, Kilian and Wigderson in [4]. A major open problem associated with this model is whether NP problems can be proven by one-round, two-prover, zero-knowledge protocols with exponentially small error probability (e.g. via parallel executions). A positive answer was claimed by Fortnow, Rompel and Sipser in [12], but its proof was later shown to be flawed by Fortnow who demonstrated that the probability of cheating inn independent parallel rounds can be much higher than the probability of cheating inn independent sequential rounds (with exponential ratio between them). In this paper we solve this problem: We show a new one-round two-prover interactive proof for Graph Hamiltonicity, we prove that it is complete, sound and perfect zeroknowledge, and thus every problem in NP has a one-round two-prover interactive proof which is perfectly zero knowledge under no cryptographic assumptions. The main difficulty is in proving the soundness of our parallel protocol namely, proving that the probability of cheating in this one-round protocol is upper bounded by some exponentially low threshold. We prove that this probability is at most 1/2n/9 (wheren is the number of parallel rounds), by translating the soundness problem into some extremal combinatorial problem, and then solving this new problem.