Quantum learning shows a steep slope but does not agree with the classical result at 0.05

Siavash Kakavand and colleagues present an exhaustive empirical study that benchmarks quantum kernel support vector machines (QSVMs) against classical baselines on tabular data, revealing important insights into the near-term feasibility of quantum computers for supervised learning. The team conducted 970 experiments utilizing nine datasets, four quantum feature maps, and three classical kernels with nested cross-validation. Their analysis, which includes hardware validation on IBM ibm_fez (Heron r2) demonstrating high kernel fidelity, shows no statistically significant advantage for QSVM at the 0.05 alpha level, even though steeper learning curves were observed in some cases. This finding highlights the important role of dataset selection in the variance of performance, suggests that current quantum feature maps lack the necessary spectral properties to outperform radial basis function kernels, and provides practical guidelines for future quantum kernel research.

High kernel fidelity and dataset selection dominate quantum machine learning performance

Kernel fidelity reached 0.976 on IBM quantum hardware, exceeding previous benchmarks and enabling reliable quantum computation of kernel matrices. This level of precision has not been previously achievable due to limitations in qubit coherence and gate fidelity. This high fidelity was sustained over six experiments, confirming reproducibility of the results with an average coefficient of variation of 1.4%. This is an important step towards building reliable quantum machine learning algorithms.

Nearly 1,000 experiments reveal that dataset selection explains 73% of the variance in performance, significantly outweighing the effect of kernel type, suggesting that data preprocessing and selection are critical to success. The quantum model showed steeper initial learning curves in six of the eight datasets, but ultimately failed to outperform the classical best results. The single quantum kernel training configuration achieved a competitively balanced accuracy of 0.968 on the breast cancer dataset, but required approximately 2,000 times more computation than traditional methods. Despite these advances, current quantum feature maps produce eigenspectra that lack the subtle profiles of optimal classical kernels, highlighting major hurdles to achieving practical quantum benefits.

Rigorous generalization evaluation with nested cross-validation and extensive experimentation

Nested cross-validation underpinned the entire study and functioned similar to repeatedly testing students with different exam questions to ensure true content understanding. This robust method extends beyond simple validation to estimate how well a machine learning model generalizes to new, unknown data by building multiple models on different training data subsets and testing them on independent data. One layer of cross-validation tunes the model parameters, and an outer layer evaluates the overall performance to prevent overly optimistic estimates. 0.970 Experiments were conducted using nine binary classification datasets and compared four quantum feature maps with three classical kernels and multiple noise models. Hardware validation was performed on an IBM_fez (Heron r2) processor and achieved a kernel fidelity of at least 0.976 across six experiments, confirming reproducibility with an average coefficient of variation of 1.4%. In this study, we rigorously evaluated generalization performance using a method that thoroughly evaluates model robustness. This approach provides a more reliable measure of how well a model performs on data not encountered during training, an important aspect of machine learning development.

Dataset characteristics currently limit the demonstrable benefits of quantum machine learning

Establishing kernel fidelity, or how accurately a quantum computer can reproduce complex mathematical relationships in data, represents a hard-won technical victory, but this benchmark does not lead to superior machine learning performance. The analysis identified that dataset selection was overwhelmingly dominant in determining the results, accounting for 73% of the variance in performance. This raises significant tensions, as focusing solely on improving quantum hardware can be misguided. Recognizing that dataset selection currently dwarfs the gains from quantum hardware does not lead to despair, but it does make it clear where focused research would benefit the most.

This detailed analysis establishes an important baseline for evaluating future quantum machine learning algorithms and hardware improvements, and pinpoints specific spectral properties that classical systems already exploit. Making the benchmark suite publicly available will accelerate progress by enabling broader reproducibility and collaborative development within the field and supporting a deeper understanding of quantum kernel functionality. Analysis of the eigenspectra reveals that current quantum feature maps lack subtle properties of optimal classical kernels, opening a clear path for future research to design more effective quantum data representations. Achieving consistently high kernel fidelity above 0.976 on IBM quantum hardware does not automatically result in superior performance on tabular datasets. This is because the characteristics of the dataset were found to have an overwhelming influence, explaining 73% of the performance variation, exceeding the influence of either the quantum or classical kernel of choice. This suggests that data preprocessing can provide greater benefits than hardware improvements alone.

This study demonstrated that a quantum kernel support vector machine did not outperform classical methods at a significance level of 0.05 across 970 experiments using nine datasets and three classical kernels. Dataset characteristics were found to be the main driver of performance, explaining 73% of the observed variance and highlighting the importance of data selection in machine learning. Spectral analysis reveals that current quantum feature maps do not reproduce the properties of valid classical kernels. The authors suggest that future research should focus on designing quantum data representations with improved spectral properties to improve performance.