Speaker
Description
Quantum fidelity kernels are often judged by circuit architecture, but their first test is simpler: does the encoding create the right similarity geometry for the task? We answer this question in cross-quality face verification using fixed AdaFace embeddings and the official XQLFW fold structure. A default four-qubit ZZ fidelity kernel at the unscaled input bandwidth is near chance. Calibrating only the input bandwidth moves the same circuit family into a competitive regime, and the recovery persists under nested, training-only bandwidth selection. The effect is therefore a genuine model-selection result, not a post-hoc choice from held-out test scores. We then test two common architectural instincts: add entangling connectivity, or increase qubit/PCA width. In this benchmark, denser ZZ connectivity does not improve performance, and scaling from four to eight qubits produces a plateau rather than a gain. Matched linear and RBF SVMs remain at least as strong as the best quantum-kernel configurations, giving a hard classical reference point. The message is direct: for quantum kernels on frozen classical embeddings, bandwidth is a first-order design choice, while neither more entanglement nor more qubits improves the inductive bias here.
| Topic | Quantum kernel methods and their applications |
|---|---|
| Additional Topic | Encoding and processing of data in quantum systems |