VibeMathedMath problems solved with AI

Rosenthal-Yuen question: unitaryQIP = unitaryPSPACE

Rosenthal and Yuen introduced the unitary synthesis classes unitaryQIP\mathsf{unitaryQIP} (unitaries synthesizable by a polynomial-time quantum verifier interacting with an untrusted prover) and unitaryPSPACE\mathsf{unitaryPSPACE} (unitaries implementable in polynomial space), and left both inclusions open: is every unitaryPSPACE\mathsf{unitaryPSPACE} family in unitaryQIP\mathsf{unitaryQIP} with nontrivial completeness and soundness, and is every unitaryQIP\mathsf{unitaryQIP} family implementable by space-uniform circuits? Does unitaryQIP=unitaryPSPACE\mathsf{unitaryQIP}=\mathsf{unitaryPSPACE}, the unitary analogue of QIP=PSPACE\mathsf{QIP}=\mathsf{PSPACE} and stateQIP=statePSPACE\mathsf{stateQIP}=\mathsf{statePSPACE}?

Result
Proved(see note)
Status
Candidate (review pending)
AI contribution
AI co-developed
Method
Argument
Field
Quantum complexity theory; unitary synthesis
Posed by
Gregory Rosenthal and Henry Yuen, Interactive proofs for synthesizing quantum states and unitaries (arXiv 2108.07192; ITCS 2022), open problems
Year posed
2021
Years open
5y
Solved
2026-11-02
Model
GPT-6
Vendor
OpenAI
Collaborators
Entong He, Fengning Ou, Mingnan Zhao
Verification
Unreviewed
Publication
Preprint
Significance
20 / 100
Disclosed cost
—
Wikipedia
No dedicated article

What was actually shown

With unitaryQIP\mathsf{unitaryQIP} requiring perfect completeness, soundness 1/21/2 and every inverse-polynomial synthesis error, and unitaryPSPACE\mathsf{unitaryPSPACE} every inverse-polynomial error, unitaryQIP=unitaryPSPACE\mathsf{unitaryQIP}=\mathsf{unitaryPSPACE}. One inclusion simulates any protocol in polynomial space via an acceptance-rigidity lemma and the Metger-Yuen prover; the other synthesizes unitaryPSPACE\mathsf{unitaryPSPACE} families in polynomial time with a PSPACE\mathsf{PSPACE} oracle, building on OpenAI's unitary synthesis construction, and replaces the oracle by statePSPACE\mathsf{statePSPACE} states obtained interactively. Also: constant-round protocols given polynomially many input copies, and a barrier for EFI pairs. Caveat: relies on an arbitrary-precision form of OpenAI's construction proved in the appendix.

What the AI did

AI disclosure in the paper: 'OpenAI GPT-6 was used extensively throughout the preparation of this manuscript, including for mathematical formulation, checking proof arguments, typesetting, and language polishing. The idea behind our proof of unitaryPSPACE in unitaryQIP originated in an early discussion among the authors that assumed a positive answer to the unitary synthesis problem. Prompted by the authors, GPT-6 helped sketch the argument that, for unitaryPSPACE families, the oracles in the unitary synthesis problem can be replaced by a PSPACE language. The authors have reviewed and verified the manuscript and take full responsibility for its content.' The reverse inclusion also builds on OpenAI's AI-produced unitary synthesis construction.

Verification

No independent mathematician has checked this yet. Checked here on 15 November 2026 against the arXiv v1 TeX source: introduction, main theorems, AI disclosure, and the restatement of OpenAI's unitary synthesis theorem, read against the TeX source of Rosenthal and Yuen (arXiv 2108.07192), whose open problems ask for unitaryPSPACE⊆unitaryQIP\mathsf{unitaryPSPACE}\subseteq\mathsf{unitaryQIP} with nontrivial completeness and soundness, and for space-uniform circuits implementing unitaryQIP\mathsf{unitaryQIP} families. The equality answers both, with classes defined as in Rosenthal-Yuen and Metger-Yuen (inverse-polynomial synthesis error). One discrepancy: the paper restates OpenAI's Theorem 1.1 with precision 2−β−12^{-\beta-1}, but OpenAI's Theorem 1.1 (identical PDF at the catalog's commit and on main) gives diamond-norm error at most 1/2; the arbitrary-precision version is proved in this paper's own appendix. Neither the appendix nor the rest of the proof was refereed here.

Sources

Submitted by MellowGander562 on

Changelog2 changes

Discussion