arXiv:2610.01874
Fourier Symmetrization for Geometric Quantum Machine Learning
Letao Wang, Abdel Lisser, Sreejith Sreekumar, Zeno Toffano
Abstract
Geometric quantum machine learning incorporates symmetry into quantum models, but how symmetry shapes their expressivity and guides effective model design remains insufficiently understood. We address this question through the Fourier representation of quantum Fourier models (QFMs). Symmetry organizes the frequency spectrum into orbits and sums the Fourier coefficients within each orbit into a symmetrized coefficient. When QFM trainable layers form independent exact 2-designs, the variance of each symmetrized coefficient equals the sum of the coefficient variances in its orbit. For single-layer QFMs with $\varepsilon$-approximate 2-design trainable layers, we bound the deviation from this identity. The resulting bound for individual Fourier coefficients can be exponentially tighter than an existing bound. The hyperoctahedral group provides an example of orbit growth that can mitigate vanishing expressivity of the symmetrized coefficients. Symmetrization of QFMs over an elementary abelian 2-group also yields pure multivariate Chebyshev polynomial basis functions. We introduce randomized encoding, which implements invariant models without ancilla qubits or the additional circuit depth for quantum twirling. We evaluate the models as quantum physics-informed neural networks (QPINNs) on two-dimensional screened Poisson and stationary viscous Hamilton-Jacobi equations. Under hard boundary constraints, QPINNs using exact symmetrization and randomized encoding achieve the lowest mean errors in the two benchmarks, respectively.
AI Summary
The paper asks how imposing symmetry changes the Fourier expressivity of data-reuploading quantum models, and how to implement invariant predictors without additional quantum circuitry. For finite orthogonal group actions, it decomposes the group-averaged predictor into normalized frequency-orbit basis functions, with each coefficient equal to the sum of the original coefficients in that orbit. Assuming independent exact unitary 2-design trainable layers, distinct Fourier coefficients have zero covariance, so $\operatorname{Var}(a_\nu)=\sum_{\omega\in[\nu]}\operatorname{Var}(c_\omega)$. Large orbits can therefore offset coefficient-variance decay in this basis. For single-uploading models with spectral-norm $\varepsilon$-approximate 2-designs, the deviation from the exact-design variance is $O(\varepsilon\|O\|_F^2(\sqrt{R}/4^n+\varepsilon))$, where $R$ is total orbit redundancy. The individual-coefficient bound removes a redundancy factor from the previous bound’s quadratic-error term. Sign-flip averaging of Chebyshev-encoded inputs also eliminates sine-containing terms, leaving pure multivariate Chebyshev polynomials. “Randomized encoding” samples input transformations to estimate the averaged predictor without extra qubits or depth. Its worst-case shot bound is independent of group size, although its input-averaged measurement variance can exceed the baseline’s. Two-qubit simulator benchmarks use two PDEs and ten runs per setting. With hard boundary constraints, the lowest mean relative errors are $0.259\%$ for exact symmetrization on screened Poisson and $3.68\%$ for randomized symmetrization on Hamilton–Jacobi. Coefficient-variance gains do not guarantee larger gradients or faster optimization. Approximate-design theory covers only one upload; hardware derivative estimation remains untested, and sampled loss gradients can be biased. Classical PINNs perform best under soft boundary constraints.
GPT-5.5 Pro was used to assist in implementing the code for the numerical experiments.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 47 pages, 9 figures