arXiv:2610.39350
Provably efficient experimental classification of quantum data with quantum kernels
Keitaro Anai, Akimoto Nakayama, Shintaro Ooka, Hiroko Tomoda, Yuto Morohoshi, Yohji Chin, Shigehito Miki, Masahiro Yabuno, Hirotaka Terai, Peter van Loock, Kosuke Mitarai, Shuntaro Takeda
Abstract
Analyzing quantum systems is essential for quantum information processing, but can require exponentially many measurements without exploiting the structure of the system or task. Quantum machine learning offers a promising way to reduce this cost by learning from quantum data. Here, we theoretically establish a framework for efficient quantum-state classification and experimentally demonstrate the resulting scheme on a programmable photonic processor. Our analysis links learning efficiency to the structure of quantum data, without relying on a physical description tailored to each task. We derive sufficient conditions for accurate classification with a number of measurements polynomial in system size. We apply this guarantee to a resource problem in continuous-variable (CV) quantum computing: classifying stellar rank, a measure of non-Gaussianity. Experimentally, we implement the classification of single-mode CV quantum states with stellar rank zero and one. This is enabled by developing a programmable CV photonic processor capable of preparing and processing diverse Gaussian and non-Gaussian states and performing virtual projection onto such states, going beyond the Gaussian-only capabilities of most existing photonic platforms. The experiment demonstrates accurate classification despite experimental imperfections. Our results establish a theory-guided and experimentally validated framework for quantum-state classification that can be implemented on quantum devices available today.
AI Summary
The paper asks when fidelity-based quantum kernels can classify quantum states with polynomial measurement resources, rather than reconstructing them tomographically. It analyzes a support-vector machine using the RBF kernel $K(\psi,\phi)=\exp[-(1-|\langle\psi|\phi\rangle|^2)/\sigma^2]$ and demonstrates binary stellar-rank classification on a programmable continuous-variable photonic processor. The guarantee assumes one class has $M$ representative states: every same-class state has kernel similarity at least $\kappa$ to some representative, while opposite-class similarities are at most $\varepsilon_{\mathrm{sep}}$. An inverse-polynomial positive margin $\eta=\kappa/M-\varepsilon_{\mathrm{sep}}$, together with a noncollapse condition, suffices for polynomial-resource learning. The proof constructs an RKHS separator and combines generalization bounds with finite-shot kernel estimation and positive-semidefinite projection. For target error $\epsilon$, sufficient training-set size is $\widetilde O(\eta^{-2}\epsilon^{-2})$; training and per-prediction measurement counts are $\widetilde O(\sigma^{-4}\eta^{-12}\epsilon^{-12})$ and $\widetilde O(\sigma^{-4}\eta^{-6}\epsilon^{-4})$. Stellar-rank-zero/one simulations cover one to three modes, using datasets constructed around $9m(m+1)$ representatives per class with within-representative fidelity at least $0.9$. The experiment measures kernels for 180 single-mode Gaussian-transformed vacuum and single-photon states. RBF classification accuracy approaches unity with increasing training size despite imperfections. The stated guarantees concern ideal state families and sufficient theoretical bandwidth choices, not the cross-validated bandwidths used in practice. The experimental homodyne estimator also differs from the bounded-outcome model analyzed theoretically. The demonstrations involve small, simple states and establish no quantum advantage; baseline comparisons do not compare measurement-shot costs.
Error found: The experiment demonstrates classification accuracy, but its homodyne estimator, noisy kernels, and cross-validated kernel width are not covered by the paper's polynomial-measurement guarantee, a limitation explicitly acknowledged in the supplement.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 9 pages, 5 figures for the main text, and 22 pages and 6 figures for the supplementary information