arXiv:2610.04190
Reduced Cost Quantum Kernel Training
Daniel R. Germain, Benjamin Goldweber, Kayla J. Rodriguez, Kristen Rhinehardt
Abstract
Quantum kernel methods are promising for near-term quantum machine learning, but their training costs are driven by the quadratic scaling of Gram-matrix construction and the measurement cost of estimating individual kernel values. In this work, we investigate one-class anomaly detection as a resource-efficient operating regime for training non-variational quantum kernel models. Using the ULB Credit Card Fraud Detection dataset, we evaluate three quantum feature maps across training-set sizes and qubit counts. Power-law saturation models reveal strong diminishing returns with increasing training data, with 500-sample models retaining approximately 90--93\% of their fitted asymptotic average precision (AP). We further observe no severe exponential concentration among the evaluated feature maps over the investigated qubit range, thereby avoiding the exponentially increasing measurement requirements associated with that regime. Finally, Nyström approximation reduces kernel construction from $\mathcal{O}(N^2)$ to $\mathcal{O}(Nm)$ evaluations, with QSVR retaining 94--97\% of its exact-kernel AP using only five landmarks. For ZZ-QSVR, downsampling from 4,000 to 500 training samples and applying a five-landmark approximation decreases the number of required kernel evaluations from approximately $8.0$ million to $2,485$, a reduction of more than 3,200-fold, while AP decreases from $0.812$ to $0.737$. Under the minimum IonQ pricing assumption considered here, this corresponds to an illustrative reduction in quantum kernel training cost from approximately \$1.34 billion to \$417{,}480. These results demonstrate that training-set reduction and low-rank approximation can substantially reduce the resources required to train non-variational quantum kernel anomaly detectors while retaining meaningful anomaly-discrimination performance.
AI Summary
The paper investigates whether training-set reduction and low-rank approximation can reduce the resource cost of non-variational quantum-kernel anomaly detection while preserving detection performance. It studies one-class SVM and reconstruction-error support vector regression (QSVR) on the ULB credit-card fraud dataset, using angle encoding, a fully connected ZZ feature map, and three-layer ZZ data re-uploading, with a classical RBF baseline. Experiments use exact noiseless simulation over 1–20 qubits, training sets of 500–4,000 nominal samples, 12 sampling seeds, and a fixed 2,000-sample test set containing 5% anomalies. Three-parameter power-law fits to four training sizes estimate that 500-sample models retain 89.79–93.19% of asymptotic average precision (AP). At 20 qubits, median off-diagonal fidelities are 0.458, 0.112, and 0.385 for angle encoding, ZZ, and ZZ re-uploading, respectively; no severe concentration is observed within this range. Nyström approximation reduces kernel-evaluation scaling from $\mathcal{O}(N^2)$ to $\mathcal{O}(Nm)$. With five landmarks, QSVR retains approximately 94–97% of exact 500-sample AP. Combining downsampling and Nyström reduces ZZ-QSVR evaluations from 7,998,000 to 2,485, while AP falls from 0.812 to 0.737. Under an assumed $168 per evaluation, illustrative training costs fall from $1.34 billion to $417,480. The concentration analysis excludes hardware noise and does not establish the hypothesized role of encoded-state localization. Saturation thresholds below 500 samples are extrapolations rather than measurements; performance asymptotes are likewise model-based. Results concern one dataset and at most 20 qubits, with approximation sensitivity depending on the kernel–detector combination.
The authors used Claude (Anthropic) to assist with editing and formatting tables and with identifying relevant references.
Error found: The absence of tiny off-diagonal fidelities in noiseless training kernels through 20 qubits does not establish the measurement precision or shot budget required to preserve anomaly-detection performance. The illustrative hardware costs and simulated AP are not demonstrated jointly, because the assumed $168 per kernel evaluation is not connected to the estimation accuracy required for the reported AP.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.