arXiv:2610.09947
On the Expressive Power and Capacity of Quantum Data Reuploaders
Lord Sen, Shyamapada Mukherjee
Abstract
Data reuploading---repeatedly encoding classical inputs at multiple layers of a parameterized quantum circuit---is a central mechanism for enhancing the expressivity of quantum neural networks (QNNs). Building on the Fourier-theoretic framework of Schuld et al.~\cite{schuld2021fourier}, we study the frequency spectrum of the general multivariate, $R$-reupload architecture (QDR). (i) We make explicit that the accessible frequencies form the set $K_R=S^{(R)}-S^{(R)}$, which equals the $R$-fold sumset $T^{(R)}$ of the difference set $T=S-S$ of the encoding spectrum, and we characterise it exactly: $K_R=\{k\in\mathcal L:\ \ell_T(k)\le R\}$, where $\mathcal L=\mathrm{span}_{\mathbb Z}(T)$ is an integer lattice and $\ell_T$ is word length in $T$. (ii) We show that the \emph{directions} of the spectrum stabilise already at $R=1$ (the real span of $K_R$ equals that of $\mathcal L$), whereas the spectrum itself never saturates: the maximal frequency grows linearly in $R$ and the number of frequencies grows polynomially, as $Θ(R^{m})$ with $m=\mathrm{rank}\,\mathcal L$. (iii) Using the spectrum size, we give a Rademacher-complexity bound on the capacity of QDRs of order $\sqrt{|K_R|/n}$. (iv) We report experiments on function approximation and classification with train/test splits, several seeds and parameter-matched classical baselines. The simulated spectra match the theory exactly. A single-qubit QDR beats a matched MLP on several smooth or periodic univariate targets and on the spiral task, but it is clearly worse on multivariate regression ($d\ge4$), where it is highly sensitive to the input-encoding scale; it beats the MLP only on equations with $d\le3$ inputs.
AI Summary
The paper studies how repeated data encoding changes the Fourier expressivity and statistical capacity of quantum data-reuploading circuits, separating available frequencies from constraints on their coefficients. For an encoding with finite phase spectrum $S$, it writes the possible frequency set after $R$ reuploads as $K_R=S^{(R)}-S^{(R)}=(S-S)^{(R)}$. Under a commensurability assumption, $K_R$ is exactly the radius-$R$ word-metric ball in the lattice $\mathcal L=\mathrm{span}_{\mathbb Z}(S-S)$. Elementary sumset arguments establish that its real span is fixed at $R=1$, every lattice frequency appears at finite depth, and the spectrum never saturates for nontrivial encodings. Maximum frequency grows linearly with $R$; an existing sumset theorem gives $|K_R|=\Theta(R^m)$, where $m=\mathrm{rank}\,\mathcal L$. Parseval’s identity and a standard feature-norm argument yield empirical Rademacher complexity at most $B\sqrt{|K_R|/n}$ for observable norm bounded by $B$. Single-qubit simulations verify the predicted spectra and compare regression and classification against approximately parameter-matched MLPs. Input scaling substantially changes performance: halving the scale reduces ReLU test RMSE from $0.178$ to $0.030$. Across 27 Feynman equations, QDR wins occur only for $d\le3$; at 45 parameters, median test RMSE is $0.147$, versus $0.076$ for MLPs. The spectral results do not characterize reachable coefficients or establish quantum advantage. The capacity bound can be loose. Experiments use small circuits, limited tuning, and scale searches frequently ending at the grid boundary; coefficient constraints and optimization effects remain unresolved.
Error found: The abstract's general Fourier-spectrum description omits the finite-phase-decomposition assumption, which excludes general noncommuting sum encodings. The abstract calls the generated frequency group an integer lattice without stating the commensurability restriction imposed in the body. The abstract's unconditional non-saturation and linear-frequency-growth statements fail for singleton encoding spectra, a case correctly excluded by the theorem.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.