A Cautionary Note on Quantum Oracles

In recent years, the quantum oracle model introduced by Aaronson and Kuperberg (2007) has found a lot of use in showing oracle separations between complexity classes and cryptographic primitives. It is generally assumed that proof techniques that do not relativize with respect to quantum oracles will also not relativize with respect to classical oracles. In this note, we show that this is not the case: specifically, we show that there is a quantum oracle problem that is contained in the class QMA, but not in a class we call polyQCPH. The class polyQCPH is equal to PSPACE with respect to classical oracles, and it is a well-known result that QMA is contained in PSPACE (also with respect to classical oracles). We also show that the same separation holds relative to a distributional oracle, which is a model introduced by Natarajan and Nirkhe (2024). We believe our findings show the need for some caution when using these non-standard oracle models, particularly when showing separations between quantum and classical resources.

💡 Research Summary

The paper revisits the role of quantum oracles in establishing separations between complexity classes, challenging the widely held belief that any non‑relativizing technique for quantum oracles automatically implies a non‑relativizing technique for classical oracles. The authors introduce a new class, polyQCPH, which extends the quantum‑classical polynomial hierarchy (QCPH) by allowing a polynomial number of alternating classical proofs rather than a constant number. They observe that, with respect to classical oracles, polyQCPH coincides with PSPACE, and therefore the well‑known containment QMA ⊆ PSPACE (and thus QMA ⊆ polyQCPH) holds in the classical relativized world.

The core contribution is a separation in the quantum‑oracle setting. Using the same unitary quantum oracle U originally employed by Aaronson and Kuperberg (2007) to separate QMA from QCMA, the authors show that QMA is not contained in polyQCPH_U. Intuitively, a QMA verifier can exploit the unitary oracle to validate a quantum witness, while any polyQCPH verifier—restricted to polynomially many alternating classical proofs and bounded error—fails to simulate the same power under this oracle. This yields the first bounded‑error quantum class separation that does not relativize with respect to a quantum oracle, thereby resolving an open problem posed by Aaronson (2009).

The paper further extends the result to the recently introduced distributional oracle model (Natarajan & Nirkhe, 2024). A distributional oracle D supplies a randomly chosen function from a prescribed distribution for each query. The authors prove that QMA ⊈ polyQCPH_D, demonstrating that the same phenomenon occurs when the oracle’s behavior is probabilistic rather than deterministic. Consequently, both quantum and distributional oracles must be handled with caution when they are used to argue about classical‑quantum separations.

In addition to the main separation, the authors discuss the relationship between polyQCPH and BQPSPACE (bounded‑error quantum polynomial space). While BQPSPACE equals PSPACE in the unrelativized and classical‑oracle settings, the paper argues that BQPSPACE can be strictly stronger than polyQCPH in the presence of quantum oracles, indicating that the classical equivalence PSPACE = BQPSPACE does not survive quantum relativization.

Key contributions:

1. Definition of polyQCPH and proof of its equivalence to PSPACE for all classical and distributional oracles.
2. Construction of a quantum unitary oracle U and a distributional oracle D for which QMA ⊈ polyQCPH, establishing a bounded‑error non‑relativizing separation.
3. Resolution of Aaronson’s open problem concerning bounded‑error quantum class separations.
4. Insight into the altered relationship between BQPSPACE and polyQCPH under quantum relativization.

The work relies heavily on previously known oracle constructions (the AK07 unitary oracle and the LLPY24 distributional oracle), so the technical novelty lies more in the reinterpretation of these constructions within the polyQCPH framework rather than in inventing new oracle techniques. Nonetheless, the paper provides a valuable meta‑theoretical perspective: quantum oracles can exhibit non‑relativizing behavior that does not manifest with classical oracles, and the same caution applies to distributional oracles. Future research directions include clarifying the exact inclusion hierarchy between polyQCPH and BQPSPACE, exploring similar separations for more general quantum oracle models such as CPTP‑map oracles, and investigating whether natural problems (outside artificially constructed oracles) exhibit the same disparity.