arXiv:2610.00841
Neural Fourier Surrogates for Data Reuploading Quantum Neural Networks
Oliver Knitter, Jonathan Mei, Sang Hyub Kim, Chi Chen, Masako Yamada, Martin Roetteler
Abstract
For quantum machine learning, the exact boundary between classical and quantum advantage is still poorly understood. Direct comparison between quantum neural networks (QNNs) and existing classical models, which encompass fundamentally different function classes, often fails to provide broader insight into the difference between the two. Inspired by the techniques of Neural Quantum States and Random Fourier Features, this work introduces Neural Fourier Surrogates (NFS), a stochastic classical neural network architecture for efficiently learning coefficients over the same finite Fourier series support as quantum neural networks. Testing on a selection of tabular benchmark datasets, we find that NFS is an effective classifier architecture broadly competitive with established classical baselines, including a comparable Random Fourier Features model, and possessing comparable performance to data-reuploading QNNs; combined with additional analysis comparing the learned Fourier spectra of QNNs and NFS on synthetic data, these results establish NFS as a natural classical baseline for evaluating QNN performance.
AI Summary
Data-reuploading quantum neural networks (QNNs) represent finite Fourier series, motivating classical comparisons that share their frequency support rather than use unrelated function classes. This paper introduces Neural Fourier Surrogates (NFS), which learn Fourier coefficients implicitly instead of enumerating the exponentially large frequency space. NFS combines an autoregressive transformer for the normalized coefficient magnitudes, a feedforward phase network, and a trainable coefficient $\ell_1$ norm. Frequency sampling provides Monte Carlo estimates of outputs and gradients; symmetry constraints ensure real outputs. A Gaussian-prior regularizer approximates QNNs’ preference for lower frequencies. NFS shares the frequency support of angle-encoded QNNs but permits coefficients unavailable to the fixed-observable circuits tested. Experiments cover 15 tabular classification datasets with 97–13,000 training examples, reporting means and standard deviations over five seeds. NFS is compared with six classical baselines, including random Fourier features, without demonstrating superior performance. Comparisons with noiseless simulated QNNs use six datasets at encoding depth $L=4$, plus width sweeps over 4–12 encoded dimensions on three datasets. Performance is qualitatively similar, with neither architecture consistently outperforming the other. Synthetic QNN-generated targets test spectral capacity and recovery: increasing encoding depth generally improves fitting, while correlations between learned and target frequency distributions cluster near zero or one. Stated limitations include small-scale benchmarks, higher training cost than conventional classical baselines without detailed runtime analysis, unmatched parameter counts, and differing QNN/NFS inductive biases. Reusing samples for outputs and gradients introduces finite-sample bias. Comparisons with quantum-hardware implementations remain future work.
Generative AI (Anthropic Claude) was used by the authors for purposes of project and experiment ideation, benchmark code generation, and general manuscript outlining and proofreading.
Central claim may not hold: The claimed scalable efficiency is not established because Monte Carlo output accuracy depends on the Fourier coefficient 1-norm, while neither that norm nor gradient-estimator variance is controlled as dimension grows. The specified antisymmetric phase forces the zero-frequency coefficient to be nonnegative, contradicting the claimed full expressivity over real Fourier series and excluding negative constants that QNNs can represent. The spectral-recovery experiment does not establish that alternative spectra learned by NFS are inaccessible to the compared QNN architecture.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 10 pages, 5 figures, 2 tables