Verification of claimed quantum advantage in quantum metric learning

Determine rigorous and practical verification procedures to certify whether a purported quantum advantage in quantum metric learning holds; specifically, establish methods that allow an independent verifier to assess if a parameterized quantum circuit used for metric learning truly achieves the claimed performance benefits.

Background

Quantum metric learning optimizes a quantum feature map to maximize separation between data classes in Hilbert space, potentially reducing circuit depth and enabling robust classification. While hybrid quantum-classical training methods can learn such embeddings, validating that a claimed quantum advantage (e.g., improved separability or classifier performance attributable to quantum embeddings) actually exists is critical for reliability, especially on noisy NISQ hardware.

The paper motivates this gap by stating that, despite progress in training quantum embeddings, there is no established approach for verification. The authors then propose a black-box verification protocol aimed at addressing this need, but explicitly note the verification question as an open problem in the introduction.

References

Although recent work has demonstrated the feasibility of training quantum embeddings using hybrid quantum-classical optimization , the question of how to verify that the claimed quantum advantage in metric learning remains open.

— Efficient and Practical Black-Box Verification of Quantum Metric Learning Algorithms  (2603.28687 - Shokry et al., 30 Mar 2026) in Section 1: Introduction

Formalizing this into a hardware-agnostic, gaming-resistant application-level benchmark remains an important open problem for the field.

— CLOPS: Benchmarking System Speed at Utility Scale  (2608.18044 - Wack, 18 Aug 2026) in Section 5.4, "Implications for the Performance Hierarchy"; reiterated in the Conclusion

Second, the hybrid adds the entire QuFeX module (quantum circuit, encoding, and classical residual); although the representation width is matched, the module's capacity is not, so part of the gain may be attributable to added capacity rather than to the quantum circuit itself. A capacity-matched classical module is the natural ablation. We do not run it here.

— A Width-Matched Comparison of Hybrid Quantum-Classical Self-Supervised Learning for Fingerprint Recognition  (2609.39172 - Edwards et al., 30 Sep 2026) in Section 6.3, “Limitations”