Towards a Doubly Efficient IP=PSPACE

Abstract

We show that every language in PSPACE decidable by a Turing machine in time $T(n)=n^{O(\log n)}$ admits a doubly efficient interactive proof system: the prover runs in time polynomial in T(n), and the verifier runs in time polynomial in n. This extends the best previously known regime for such proof systems from $T(n)=n^{O(\sqrt{\log n / \log\log n})}$, established by Berger, Goyal, Hong, and Kalai (FOCS 2025), to $T(n)=n^{O(\log n)}$. Beyond improving the range of T, our protocol is substantially simpler than previous doubly efficient proofs for time-bounded PSPACE. Earlier constructions proceed indirectly: they first build batch interactive proofs and then invoke them as a black box to obtain doubly efficient protocols. In contrast, we give a direct construction. This not only simplifies the proof but also points to a more promising route for future improvements.