arXiv:2610.31260
Can Entanglement-Induced Feature-Map Designs Improve Projected Quantum Kernels?
Ravi Kumar Jha, Hsuan-Yin Lin, Eirik Rosnes
Abstract
Projected quantum kernels (PQKs) have emerged as a notable approach to circumvent the hardness of machine learning tasks arising from the exponential dimensionality of the quantum Hilbert space. However, recent PQK studies often rely heavily on extensive classical hyperparameter tuning of the radial basis function (RBF) kernel parameter and the support vector machine regularization parameter. The extensive tuning can mask the true representational power of the underlying PQKs induced by quantum feature maps, which has not been explored in detail. Thus, this work aims to expose the structural capabilities of PQKs by isolating and evaluating their native inductive bias through comparisons between untuned and tuned PQKs against the RBF baseline. Furthermore, we propose novel quantum feature maps that leverage two-qubit entanglement to capture multivariate correlations in IoT data. The results show that extensive tuning can lead to competitive performance architectures, while our proposed entanglement topologies exhibit a significantly stronger native inductive bias. This reveals that superior quantum architectural design can yield structurally better PQKs with improvements arising from quantum-mechanical properties rather than from classical hyperparameter tuning.
AI Summary
The paper asks whether projected quantum kernels (PQKs) benefit from feature-map design independently of classical kernel and SVM hyperparameter tuning. It compares rotation-based and IQP feature maps with two proposed variants using selected, feature-dependent two-qubit interactions. Entangling topologies are chosen by empirical performance over candidate pairwise interaction structures. Experiments use noiseless six-qubit PennyLane simulations on one room-occupancy dataset containing 2,865 samples and six sensor features. PQKs apply an RBF kernel to 18 single-qubit or adjacent-pair Pauli expectation values, or their 36-dimensional concatenation. Evaluation uses 10-fold cross-validation with reported 95% confidence intervals, comparing fixed regularization and variance-based bandwidth heuristics against grid-search tuning. Classical comparisons include RBF SVMs, random forests, and neural networks. The proposed entangled maps reach tuned accuracy of $0.903\pm0.011$ and $0.903\pm0.008$, compared with $0.893\pm0.008$ for the tuned classical RBF and $0.902\pm0.006$ for random forests. The strongest untuned proposed map reaches $0.868\pm0.014$, versus $0.826\pm0.017$ for the untuned RBF; standard rotation maps reach up to $0.864$. Generic circular-CNOT and IQP designs do not improve on standard rotation maps. An unentangled ablation is also reported. Smaller tuned–untuned performance differences are interpreted as evidence of task-specific inductive bias under the prescribed defaults. The stated limitations are heuristic feature-map design and evaluation on a single, relatively small dataset. Performance comparisons may depend on baseline hyperparameter choices; broader dataset and baseline robustness tests are left for future work.
Central claim may not hold: The proposed Rot(X)^en map has equal or lower untuned accuracy than standard Rot(X) in all three measurement settings, contradicting stronger native performance for both proposed architectures. The reported results do not establish the abstract's statistically suggestive claim of a 'significantly stronger' native inductive bias. Both proposed maps have larger combined-measurement tuning gains than their corresponding standard rotation maps, so reduced tuning dependence is not demonstrated consistently relative to existing PQKs. The purportedly untuned architectures were selected through empirical topology optimization, which remains a model-selection contribution not isolated by fixing SVM and bandwidth parameters. The ablation supports a task-specific benefit of selected entangling interactions, but does not establish the broader attribution to intrinsically quantum properties rather than nonlinear representation design and empirical selection.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.