Data Re-Uploading Scheme in Quantum ML

• Data re-uploading is a quantum machine-learning technique that repeatedly encodes classical inputs into quantum states, enhancing circuit expressivity.
• It alternates data-dependent and trainable rotations, leveraging depth in circuits to maximize performance with limited quantum resources.
• Empirical and theoretical studies show its effectiveness in applications like traffic forecasting, image recognition, and reinforcement learning while addressing trainability challenges.

Data re-uploading is a quantum machine-learning scheme in which the same classical input is encoded into a quantum state multiple times, interleaved with trainable unitaries, so that a shallow parametrized circuit acquires a richer functional dependence on the input than a single-upload feature map typically permits (Pérez-Salinas et al., 2019). The paradigm was introduced as a route to universal single-qubit classification and has since been extended to multi-qubit quantum neural networks, qudit models, bosonic photonic circuits, reinforcement learning, neural quantum kernels, and time-series forecasting, while also becoming a focal point for questions of trainability, generalization, and approximation complexity (Rodriguez-Grasa et al., 2024, Schetakis et al., 22 Jan 2025, Liu et al., 24 Jun 2026).

1. Origins and core idea

The original formulation presents data re-uploading as the mechanism that allows a single qubit, assisted by a classical optimization loop, to act as a universal quantum classifier (Pérez-Salinas et al., 2019). Instead of encoding an input vector once and then applying a purely variational circuit, the model alternates data-dependent rotations and trainable rotations. In the compressed single-qubit form used in the original proposal, one writes

$$\mathcal U(\boldsymbol{\theta},\boldsymbol{w},\boldsymbol{x}) = \prod_{i=1}^{N} U\!\bigl(\boldsymbol{\theta}_i+\boldsymbol{w}_i\circ\boldsymbol{x}\bigr),$$

so the same $\boldsymbol{x}$ reappears at every layer through different trainable offsets and weights (Pérez-Salinas et al., 2019).

Two motivations recur across the literature. First, repeated encoding boosts expressivity: the output expectation becomes a progressively more intricate nonlinear function of the input. Second, re-uploading trades width for depth: it reuses the same quantum degrees of freedom rather than demanding more qubits. Later work reframed the same mechanism in task-specific ways. In traffic forecasting, for example, repeated uploads are interpreted not only as depth but also as a recurrent memory mechanism, because each upload acts on an already evolved state and therefore preserves compressed information from earlier uploads (Schetakis et al., 22 Jan 2025).

2. Circuit structure and architectural realizations

A generic re-uploading circuit can be written as

$$U(x,\theta)=\prod_{r=1}^{N_{\mathrm{ru}}} U^{(r)}_{\mathrm{ent}}(\theta^{(r)})\,U^{(r)}_{\mathrm{enc}}(x),$$

with prediction extracted through an observable expectation,

$$y_{\mathrm{pred}}(x)=\langle 0|U^\dagger(x,\theta)\,O\,U(x,\theta)|0\rangle.$$

In the traffic-forecasting implementation, $U^{(r)}_{\mathrm{enc}}(x)$ is realized with angle encoding using $R_x$ and $R_y$, while each processing block contains trainable $R_zR_yR_z$ rotations and a CNOT entangling layer, followed by Pauli-$Z$ measurement (Schetakis et al., 22 Jan 2025). In the teacher–student comparison of re-uploading and quantum perceptrons, the basic re-uploading layer is written as

$$L_i(\mathbf{x},\mathbf{w}_i)=U(\mathbf{w}_i)\,U_{\mathrm{enc}}(\mathbf{x}),$$

with $\boldsymbol{x}$0 and trainable $\boldsymbol{x}$1 plus CZ gates in the processing unitary (Aikaterini et al., 2021).

Although angle encoding with Pauli rotations is the dominant pattern, the architecture admits several non-equivalent realizations. Multi-qubit neural-kernel work employs repeated $\boldsymbol{x}$2 blocks, local $\boldsymbol{x}$3 rotations, and controlled-$\boldsymbol{x}$4 entanglers, together with an iterative warm-start strategy that adds qubits while preserving previous solutions (Rodriguez-Grasa et al., 2024). Single-qudit re-uploading replaces Pauli rotations by angular-momentum generators $\boldsymbol{x}$5, $\boldsymbol{x}$6, and the squeezing operator $\boldsymbol{x}$7, thereby treating labels as orthogonal qudit basis states rather than non-orthogonal qubit label states (Wach et al., 2023). Bosonic and photonic realizations translate the same alternating pattern into two-mode interferometers and phase shifters, so that data enter as optical phases and the measurement is a coincidence probability in the relevant Fock sector (Ono et al., 2022, Mauser et al., 7 Jul 2025).

A particularly important later distinction is between the original separated architecture and the compressed variant in which data and trainable angles are merged into a single gate. The integrated-photonics analysis argues that these two are not merely implementation choices: the original scheme keeps encoding and processing as separate gates, whereas the compressed scheme changes both the learning-theoretic properties and the geometry of the loss landscape (Mauser et al., 7 Jul 2025).

3. Expressivity, Fourier structure, and approximation theory

The standard theoretical explanation for re-uploading is spectral. Several papers describe expectation values of encoded parametrized circuits as partial Fourier series in the input, with the accessible frequency set determined by the encoding Hamiltonians and enlarged by repeated uploads (Coelho et al., 2024, Rodriguez-Grasa et al., 2024). In that picture, a single Pauli rotation introduces a restricted set of frequencies, while re-uploading increases the number of accessible harmonics and thereby the complexity of functions that can be represented.

Empirical teacher–student benchmarks support this view. When re-uploading circuits are used as teachers, they generate prediction maps with non-trivial inner structure that a single quantum perceptron cannot reproduce faithfully; conversely, re-uploading students can learn perceptron-generated targets essentially perfectly (Aikaterini et al., 2021). The same study finds that the decisive variable is not simply “more trainable gates,” but the number of times the encoding unitaries appear. Adding more processing gates without additional encodings does not materially close the expressive gap, whereas architectures with comparable encoding counts exhibit much closer behavior (Aikaterini et al., 2021).

Learning-theoretic analyses sharpen this picture. The photonic implementation that preserves separated encoding and processing proves that a one-layer original scheme yields a hypothesis class

$\boldsymbol{x}$8

and more generally that the VC dimension of the original one-qubit architecture scales as $\boldsymbol{x}$9 for $$U(x,\theta)=\prod_{r=1}^{N_{\mathrm{ru}}} U^{(r)}_{\mathrm{ent}}(\theta^{(r)})\,U^{(r)}_{\mathrm{enc}}(x),$$0 layers, while the compressed one-layer scheme has infinite VC dimension in the analysis developed there (Mauser et al., 7 Jul 2025). The same work also reports a sharp contrast in Hessian spectra at minima: the original scheme exhibits much flatter minima than the compressed scheme (Mauser et al., 7 Jul 2025). This suggests that “re-uploading” is not a single invariant object; the exact factorization of encoding and trainable operations matters.

A complementary theoretical line asks what happens if trainable upload frequencies are removed. For tunable upload circuits, a fixed-upload circuit can approximate the target with depth

$$U(x,\theta)=\prod_{r=1}^{N_{\mathrm{ru}}} U^{(r)}_{\mathrm{ent}}(\theta^{(r)})\,U^{(r)}_{\mathrm{enc}}(x),$$1

up to a target-dependent constant overhead, and mismatch-class targets obey logarithmic lower bounds

$$U(x,\theta)=\prod_{r=1}^{N_{\mathrm{ru}}} U^{(r)}_{\mathrm{ent}}(\theta^{(r)})\,U^{(r)}_{\mathrm{enc}}(x),$$2

The result gives a quantitative answer to the cost of replacing tunable frequencies by fixed uploads: the lost tunability can be transferred into circuit depth with only polylogarithmic growth in the approximation error $$U(x,\theta)=\prod_{r=1}^{N_{\mathrm{ru}}} U^{(r)}_{\mathrm{ent}}(\theta^{(r)})\,U^{(r)}_{\mathrm{enc}}(x),$$3 (Liu et al., 24 Jun 2026).

4. Trainability, generalization, and failure modes

Re-uploading is often introduced as an expressivity device, but several studies emphasize its optimization behavior. In variational deep Q-learning, models with data re-uploading display gradient norms and gradient variances that remain substantial throughout training, and increasing the number of qubits up to 12 in the tested regime does not induce the exponential vanishing expected from standard barren-plateau arguments (Coelho et al., 2024). That paper further reports that re-uploading tends to increase gradient magnitude and variance relative to single-encoding baselines, especially when combined with trainable input and output scaling (Coelho et al., 2024). In traffic forecasting, the authors likewise note that re-uploading circuits tend to be better in terms of gradient behavior and observe that the hybrid recurrent-like models converge at least as fast, and often faster, than matched classical baselines in epochs (Schetakis et al., 22 Jan 2025).

Generalization, however, is not uniformly benign. A major recent result shows that for high-dimensional classical data processed by limited-qubit re-uploading models, increasing the number of encoding layers can drive the average encoded state toward the maximally mixed state, so that predictive performance on unseen data degenerates toward random guessing in classification and toward trivial baselines in regression (Wang et al., 24 May 2025). The key claim is explicit: repeated uploading does not rescue predictive performance in the deep, narrow, high-dimensional regime (Wang et al., 24 May 2025). The same work therefore advocates wider, shallower architectures over deeper, narrower ones for high-dimensional inputs.

This tension has motivated hybrid strategies. Neural quantum kernels use a trained re-uploading QNN as a problem-adapted embedding and then freeze it to define embedding or projected quantum kernels, with numerical evidence that the resulting neural kernels alleviate exponential concentration and improve generalization relative to problem-agnostic kernels (Rodriguez-Grasa et al., 2024). A plausible implication is that re-uploading can be most effective when it is used to learn a feature map whose downstream readout is subsequently regularized by kernel machinery, rather than asked to shoulder the entire prediction problem end to end.

5. Empirical domains and representative results

The scheme has moved well beyond two-dimensional toy classification. It now appears in transport forecasting, image recognition, offline reinforcement learning, particle identification, photonic classification, and hardware-native pulse control.

Domain	Implementation	Reported outcome
Traffic forecasting	Hybrid quantum recurrent-like layer with $$U(x,\theta)=\prod_{r=1}^{N_{\mathrm{ru}}} U^{(r)}_{\mathrm{ent}}(\theta^{(r)})\,U^{(r)}_{\mathrm{enc}}(x),$$4 qubits and $$U(x,\theta)=\prod_{r=1}^{N_{\mathrm{ru}}} U^{(r)}_{\mathrm{ent}}(\theta^{(r)})\,U^{(r)}_{\mathrm{enc}}(x),$$5 re-uploading blocks	Hybrid models outperform classical LSTMs from about $$U(x,\theta)=\prod_{r=1}^{N_{\mathrm{ru}}} U^{(r)}_{\mathrm{ent}}(\theta^{(r)})\,U^{(r)}_{\mathrm{enc}}(x),$$6 and show lower variance across CV folds (Schetakis et al., 22 Jan 2025)
Superconducting image recognition	4-qubit hybrid classifier with multiple encoding layers	Around 90% digit-recognition accuracy on the MNIST task; around 95% on simpler supervised tasks (Tolstobrov et al., 2023)
Bosonic photonic classification	Two-mode two-photon optical circuit with re-uploading	Proof-of-principle classification with a reproduction rate of approximately 94% (Ono et al., 2022)
Calorimetric particle identification	Single-qubit QRU on three calorimetric features	Mean test accuracy around 0.98 in repeated runs; a learning-rate sweep reports 0.9854 (Cassé et al., 2024)
Offline reinforcement learning	4-qubit BCQQ with cyclic data re-uploading	Average reward 500 on CartPole with random-policy buffers of $$U(x,\theta)=\prod_{r=1}^{N_{\mathrm{ru}}} U^{(r)}_{\mathrm{ent}}(\theta^{(r)})\,U^{(r)}_{\mathrm{enc}}(x),$$7, $$U(x,\theta)=\prod_{r=1}^{N_{\mathrm{ru}}} U^{(r)}_{\mathrm{ent}}(\theta^{(r)})\,U^{(r)}_{\mathrm{enc}}(x),$$8, and $$U(x,\theta)=\prod_{r=1}^{N_{\mathrm{ru}}} U^{(r)}_{\mathrm{ent}}(\theta^{(r)})\,U^{(r)}_{\mathrm{enc}}(x),$$9 transitions (Periyasamy et al., 2023)

These results are not uniform in what produces the gain. In transport forecasting, re-uploading is used as a quantum analogue of recurrence, with the number of qubits deliberately matched to the number of re-uploads to mirror the number of LSTM time steps (Schetakis et al., 22 Jan 2025). In the superconducting MNIST experiment, the advantage is instead tied to multiple encoding layers combined with a classical convolutional front-end that generates feature angles reused across the circuit (Tolstobrov et al., 2023). In calorimetric particle identification, the decisive factors are shallow single-qubit depth, appropriate normalization to $$y_{\mathrm{pred}}(x)=\langle 0|U^\dagger(x,\theta)\,O\,U(x,\theta)|0\rangle.$$0 or $$y_{\mathrm{pred}}(x)=\langle 0|U^\dagger(x,\theta)\,O\,U(x,\theta)|0\rangle.$$1, and a three-parameter-per-input sandwich block such as $$y_{\mathrm{pred}}(x)=\langle 0|U^\dagger(x,\theta)\,O\,U(x,\theta)|0\rangle.$$2 (Cassé et al., 2024).

Synthetic studies continue to function as a diagnostic testbed. Strategic classifier benchmarks on line and circle datasets report that deeper re-uploading improves accuracy substantially, with single-qubit five-layer settings reaching 97.7% on a fixed linear task and 88.8% on a random circular task under the specific optimizer and sample settings reported there, while entangled two-qubit variants improve peak accuracy on random datasets (Aminpour et al., 2024). Such results do not settle real-world utility, but they do expose the dependence of re-uploading behavior on optimizer choice, loss function, label-state geometry, and circuit topology.

6. Variants, limitations, and current research directions

A major theme in recent work is that the placement and organization of uploads can matter as much as their mere count. Incremental Data-Uploading redistributes a fixed set of encoding gates throughout the circuit instead of repeatedly uploading a compressed global representation. Under an equal parameter budget on downscaled $$y_{\mathrm{pred}}(x)=\langle 0|U^\dagger(x,\theta)\,O\,U(x,\theta)|0\rangle.$$3 images, the IDU$$y_{\mathrm{pred}}(x)=\langle 0|U^\dagger(x,\theta)\,O\,U(x,\theta)|0\rangle.$$4 architecture reports 56.7% test accuracy on MNIST versus 33.2% for the DRU baseline, and 56.9% versus 43.5% on Fashion-MNIST (Periyasamy et al., 2022). In offline reinforcement learning, cyclic data re-uploading modifies the standard strategy by cyclically shifting the feature-to-qubit assignment from layer to layer so that every qubit is exposed to every feature, and the paper reports a slight increase in effective dimension relative to standard DRU (Periyasamy et al., 2023). This suggests that “how data are re-uploaded” is itself a hyperparameter family.

Another theme is hardware alignment. On a simulated superconducting transmon processor with realistic noise, a pulse-native variant of data re-uploading embeds trainable parameters directly into single-qubit pulse blocks and CR-like two-qubit pulses, while retaining the re-uploading architecture at the high level. Under equivalent noise conditions, that pulse-based model reports higher test accuracy and improved generalization than its gate-based counterpart, together with slower degradation as noise strength is increased (Acedo et al., 11 Dec 2025). In parallel, integrated photonics now provides an experimental realization of the original separated scheme, together with a proof that the implementation is both a universal classifier and an effective learner with finite-depth VC-dimension control (Mauser et al., 7 Jul 2025).

The principal limitations remain depth, noise, and data geometry. Classical simulation becomes expensive quickly: in traffic forecasting, each additional re-uploading block roughly doubles training time under simulation (Schetakis et al., 22 Jan 2025), and the four-transmon MNIST experiment estimates roughly 100 hours of wall-clock training time for about 100 iterations on the reported hardware stack (Tolstobrov et al., 2023). Performance also depends strongly on alignment between architecture and problem structure. Single-qudit work shows that when labels, qudit states, and encoding operators are well aligned, performance can improve substantially over qubit-based circuits; when that alignment is broken, the advantage can disappear (Wach et al., 2023). This suggests that re-uploading should not be treated as a generic depth heuristic alone, but as a structured encoding design problem involving operator choice, feature order, label geometry, and hardware-native constraints.

Across these threads, a consistent picture has emerged. Data re-uploading is no longer merely a claim that “repeating the data helps.” It is a family of architectures whose behavior depends on the interplay among encoding multiplicity, trainable processing, measurement design, and approximation regime. Its strongest results arise when that interplay is engineered deliberately; its sharpest limitations appear when deep encoding is used as a substitute for width or when circuit structure and data geometry are poorly matched (Wang et al., 24 May 2025, Liu et al., 24 Jun 2026).