Probabilistic verifiers for asymmetric debates

We examine the power of silent constant-space probabilistic verifiers that watch asymmetric debates (where one side is unable to see some of the messages of the other) between two deterministic provers, and try to determine who is right. We prove that probabilistic verifiers outperform their deterministic counterparts as asymmetric debate checkers. It is shown that the membership problem for every language in NSPACE(s(n)) has a 2^{s(n)}-time debate where one prover is completely blind to the other one, for polynomially bounded space constructible s(n). When partial information is allowed to be seen by the handicapped prover, the class of languages debatable in 2^{s(n)} time contains TIME(2^{s(n)}), so a probabilistic finite automaton can solve any decision problem in P with small error in polynomial time with the aid of such a debate. We also compare our systems with those with a single prover, and with competing-prover interactive proof systems.

💡 Research Summary

The paper introduces a novel interactive model called an asymmetric debate, in which two deterministic provers exchange messages while a verifier with only constant‑size memory observes the exchange. The key twist is that one prover (the “handicapped” prover) cannot see any of the messages sent by the other prover, whereas the second prover has full visibility. The verifier is allowed to use randomness but must operate in constant space, essentially behaving like a probabilistic finite automaton (PFA).

The authors first show that for any language L that can be recognized by a nondeterministic Turing machine using s(n) space (where s(n) is polynomially constructible), there exists a debate of length 2^{s(n)} in which the blind prover can present a correct computation history while the visible prover supplies only the information required for the verifier to cross‑check the steps. The verifier randomly samples positions in the claimed history, checks consistency between the two streams, and rejects if any discrepancy is found. By standard amplification, the error probability can be reduced below any constant, and the overall running time remains O(2^{s(n)}). This result demonstrates that constant‑space probabilistic verifiers are strictly more powerful than their deterministic counterparts in the asymmetric setting, because deterministic constant‑space verifiers can only recognize regular languages.

Next, the paper relaxes the blindness constraint: the handicapped prover is allowed to see a limited, predetermined portion of the other prover’s messages (for example, a single bit per round). Under this partial‑visibility model the class of languages that admit a 2^{s(n)}‑time debate expands to include TIME(2^{s(n)}). Consequently, a PFA equipped with such a debate can solve any problem in P with arbitrarily small error in polynomial time, simply by encoding a polynomial‑time deterministic computation into the debate and letting the verifier perform random checks. This bridges the gap between the modest power of constant‑space deterministic verifiers and the full power of polynomial‑time deterministic machines.

The authors then compare the two‑prover asymmetric debate model with two well‑studied frameworks: single‑prover interactive proofs (IP) and competing‑prover interactive proofs (CIP). In the single‑prover case, a constant‑space verifier cannot achieve the same expressive power because it lacks an independent source of cross‑validation. In the CIP setting, both provers are usually assumed to have polynomial resources, and the verifier may also have polynomial memory. By contrast, the present model keeps the verifier’s memory at the absolute minimum while still achieving comparable or even stronger language‑recognition capabilities, thanks to the asymmetry of information.

Finally, the paper discusses practical implications. Many real‑world systems—such as low‑power sensor networks, embedded devices, or cryptographic protocols for constrained hardware—cannot afford a verifier with large memory. The asymmetric debate framework suggests a way to outsource verification to two mutually distrustful agents: one can be a powerful server that knows the full computation, the other a lightweight client that only observes a tiny slice of the exchange. The constant‑space verifier, using cheap random bits, can still guarantee with high confidence that the server’s claim is correct, without ever storing the entire transcript.

In summary, the work establishes that probabilistic constant‑space verifiers, when placed in an asymmetric debate with two deterministic provers, can recognize languages far beyond regular ones—up to NSPACE(s(n)) and even TIME(2^{s(n)})—and that this model both enriches the theoretical landscape of interactive proof systems and offers a promising blueprint for verification in severely resource‑constrained environments.