arXiv:2609.21916
Asymptotics for Frequency Redundancies in Quantum Machine Learning Models
Felix Paul, Bhilahari Jeevanesan, Peter Jung
Abstract
The redundancy distribution of the frequency spectrum has been shown in the literature to impact the expressivity and trainability of Quantum Fourier Models (QFMs). In this work, we address the question of how this redundancy spectrum is shaped by the choice of eigenvalues of the data re-uploading Hamiltonians, using a simple mathematical formalism based on generating functions. We derive exact and asymptotic redundancy profiles for several structured eigenvalue choices identical for every layer, including arithmetic progressions and single-qubit Pauli encodings, and show that both approach a Gaussian profile as the number of layers grows. We then show that this Gaussian limit is not specific to these constructions but is a generic feature of QFMs built from integer eigenvalues that are identical in every layer. These results can be tied to the central limit theorem for random walks. These results clarify why generic or unstructured encoding choices give rise to a redundancy bias that favours low frequencies, highlighting the importance of well-thought-out encoding strategies when constructing a QFM.
AI Summary
Quantum Fourier models represent data-dependent circuit outputs as finite Fourier series. This paper asks how the eigenvalues of data-encoding Hamiltonians determine frequency redundancy—the number of eigenvalue-index pairs generating each frequency—and how repeated encoding shapes that distribution. The motivation is existing links between redundancy, Fourier-coefficient expressivity, and training gradients. For integer-valued frequencies, redundancy is expressed as a Laurent-polynomial coefficient: $r_\omega=[z^\omega]\prod_{\ell=1}^L f_\ell(z)f_\ell(z^{-1})$, where $f_\ell(z)=\sum_k z^{\lambda_k^{(\ell)}}$. Fourier inversion gives an exact integral representation. For identical Hamiltonians with eigenvalues $1,\ldots,d$, the single-layer profile is triangular, $r_\omega=d-|\omega|$ for $|\omega|<d$, and the large-layer profile is Gaussian. Products of weighted single-qubit Pauli rotations are also treated; binary weights $a_k=2^{k-1}$ reproduce the equally spaced spectrum. More generally, for a fixed, nonconstant integer eigenvalue multiset repeated across layers, with eigenvalue-difference greatest common divisor one, the central asymptotic profile is $r_\omega\sim d^{2L}(4\pi Ls^2)^{-1/2}\exp[-\omega^2/(4Ls^2)]$, where $s^2$ is the eigenvalue variance. Thus the normalized redundancy has variance $2Ls^2$. The derivation uses Laplace expansion of the generating-function integral and is interpreted through central-limit behaviour of sums of independent eigenvalue differences. The general asymptotic result assumes identical encodings across layers and fixed eigenvalues; other lattice spacings require rescaling. The analysis concerns redundancy rather than directly establishing training performance. Designing eigenvalues to realize a target redundancy profile remains an open follow-up problem.
Error found: The general Gaussian formula is stated for nonzero eigenvalue variance and lattice span one; in the original variables, span g requires a prefactor g on gZ and zero redundancy elsewhere, while zero variance gives a point mass. The scaled arithmetic-spectrum asymptotic incorrectly replaces d²−1 by d² without assuming large d, so it is not asymptotically equivalent at fixed d as the layer count grows. The binary-weighted Pauli example prints eigenvalue endpoints ±(d+1)/2 instead of the correct ±(d−1)/2. The appendix's argument excluding additional maxima incorrectly infers integer a from integer eigenvalue differences, although its gcd-one conclusion can be established directly using Bézout's identity.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 9 pages, 4 figures