Data Re-Uploading Quantum Classifier

• The paper demonstrates that repeated data re-uploading with interleaved trainable quantum gates enables universal function approximation, validated through theoretical proofs and experimental benchmarks.
• It employs versatile quantum circuit architectures—from single-qubit and qudit setups to bosonic and quantum-input schemes—to robustly encode classical and quantum data.
• Resource-accuracy trade-offs are quantified by analyzing depth scaling and optimized training methods, showing competitive performance compared to classical shallow networks.

A data re-uploading universal quantum classifier is a quantum machine learning model that achieves universal function approximation on classical or quantum data by repeatedly encoding (“re-uploading”) the same data into a quantum circuit—typically via parameterized rotations or, in more general constructions, through entangling or Hamiltonian evolutions—interleaved with layers of trainable quantum gates. This architectural motif is robust across qubit, qudit, bosonic, and even direct quantum-input settings, and has been realized experimentally on both superconducting and photonic quantum hardware. Universality is mathematically established by showing that arbitrary continuous decision functions can be approximated arbitrarily well by choosing sufficient circuit depth and parameter resolution.

1. Mathematical Foundations and Circuit Architecture

The core building block of a data re-uploading universal quantum classifier is the alternation of data-encoding unitaries and variational (trainable) gates. For an input vector $x \in \mathbb{R}^d$ (often normalized), the layered ansatz for a single-qubit model takes the form:

$$U(x;\Theta) = \prod_{l=1}^L U_{\text{train}}^{(l)}(\theta_l)\, U_{\text{data}}^{(l)}(x)$$

where $U_{\text{data}}^{(l)}(x)$ encodes $x$ into the circuit (e.g. using $R_z$ and $R_x$ rotations for each feature or as affine-linear forms in the data), and $U_{\text{train}}^{(l)}$ is a trainable single- or multi-qubit unitary (Pérez-Salinas et al., 2019, Tapia et al., 2022, Tolstobrov et al., 2023). For multi-qubit or higher-dimensional architectures, these layers can include entangling gates and, in the case of qudits or bosonic modes, unitaries generated by $su(d)$ or $su(M)$ algebra elements, including squeezing operators for expressivity (Wach et al., 2023, Ono et al., 2022).

Hamiltonian embedding approaches further generalize this scheme by defining the data-encoding as evolution under a Hamiltonian function of the datapoint (e.g., for an image $M$, $$U(x;\Theta) = \prod_{l=1}^L U_{\text{train}}^{(l)}(\theta_l)\, U_{\text{data}}^{(l)}(x)$$0), where $$U(x;\Theta) = \prod_{l=1}^L U_{\text{train}}^{(l)}(\theta_l)\, U_{\text{data}}^{(l)}(x)$$1 reflects the structure of the data (such as symmetrized pixel matrices for images) (Wang et al., 2024).

In scenarios processing quantum input states, the classifier becomes a composition of completely positive trace-preserving maps (CPTP), where an ancilla qubit interacts with fresh copies of the quantum input at each layer, and data re-uploading is implemented via controlled entangling gates and mid-circuit resets (Cha et al., 23 Sep 2025).

Table: Representative Data Re-Uploading Classifier Circuits

Platform	Data Encoding	Trainable Layer	Measurement
Qubit	$$U(x;\Theta) = \prod_{l=1}^L U_{\text{train}}^{(l)}(\theta_l)\, U_{\text{data}}^{(l)}(x)$$2, arctan	$$U(x;\Theta) = \prod_{l=1}^L U_{\text{train}}^{(l)}(\theta_l)\, U_{\text{data}}^{(l)}(x)$$3 rotations	$$U(x;\Theta) = \prod_{l=1}^L U_{\text{train}}^{(l)}(\theta_l)\, U_{\text{data}}^{(l)}(x)$$4, Bloch label
Qudit	$$U(x;\Theta) = \prod_{l=1}^L U_{\text{train}}^{(l)}(\theta_l)\, U_{\text{data}}^{(l)}(x)$$5	$$U(x;\Theta) = \prod_{l=1}^L U_{\text{train}}^{(l)}(\theta_l)\, U_{\text{data}}^{(l)}(x)$$6	Computational basis
Bosonic	$$U(x;\Theta) = \prod_{l=1}^L U_{\text{train}}^{(l)}(\theta_l)\, U_{\text{data}}^{(l)}(x)$$7 (mode ops)	Phase shifts/beamsplitters	Fock basis projector
Hamiltonian	$$U(x;\Theta) = \prod_{l=1}^L U_{\text{train}}^{(l)}(\theta_l)\, U_{\text{data}}^{(l)}(x)$$8	$$U(x;\Theta) = \prod_{l=1}^L U_{\text{train}}^{(l)}(\theta_l)\, U_{\text{data}}^{(l)}(x)$$9 variational	Computational basis, softmax
Quantum input	Ancilla–input entangling SU(2d) operations	Per-layer parameterized unitaries	Ancilla observable post CPTP cascade

2. Universality and Expressivity

Universality is established through explicit constructive theorems. For a single-qubit layered circuit with sufficient depth $U_{\text{data}}^{(l)}(x)$0, the predicted measurement probability $U_{\text{data}}^{(l)}(x)$1 can approximate any bounded continuous function $U_{\text{data}}^{(l)}(x)$2 to arbitrary precision (Pérez-Salinas et al., 2019, Mauser et al., 7 Jul 2025, Tapia et al., 2022). This follows from the trigonometric polynomial expansion of the quantum output (i.e., each layer incrementing accessible Fourier frequencies), and by the Stone–Weierstrass theorem.

For the general single-qudit case, the inclusion of a squeezing gate $U_{\text{data}}^{(l)}(x)$3 alongside $U_{\text{data}}^{(l)}(x)$4 and $U_{\text{data}}^{(l)}(x)$5 in each layer ensures $U_{\text{data}}^{(l)}(x)$6 universality, with expressivity governed by the total number of variational generators and circuit depth (Wach et al., 2023). For quantum inputs, repeated ancilla–input interactions allow the output expectation to become an arbitrary multivariate polynomial of the input Bloch parameters—yielding universal expressivity over quantum data distributions (Cha et al., 23 Sep 2025).

Hamiltonian embedding strategies further enhance expressiveness by including nontrivial global polynomial nonlinearities in input pixel matrices, enriching the class of representable hypotheses beyond those accessible by simple angle-encoding (Wang et al., 2024).

3. Training Methods and Cost Functions

Data re-uploading classifiers are trained via classical optimizers wrapped around quantum circuit evaluations. Common loss functions include:

• Cross-entropy or log-loss (for multi-class tasks)
• Rescaled logistic loss plus $U_{\text{data}}^{(l)}(x)$7 regularization (to stabilize optimization and enforce parameter norm control), e.g.,

$U_{\text{data}}^{(l)}(x)$8

where $U_{\text{data}}^{(l)}(x)$9 and $x$0 are experimentally tuned (Tolstobrov et al., 2023)

• Fidelity-based loss and trace distance (for robust optimization and avoidance of barren plateaus in cost landscapes) (Aminpour et al., 2024)

Gradients are computed via the parameter-shift rule, which enables unbiased quantum differentiation by evaluating the circuit at shifted parameter values. For single-parameter updates, the analytic dependence of the circuit output on phase angles facilitates efficient sequential minimal optimization (SMO) (Mauser et al., 7 Jul 2025, Ono et al., 2022). For multi-qubit or Hamiltonian-embedded models, Adam or Nesterov-accelerated SGD are standard (Tolstobrov et al., 2023, Wang et al., 2024).

Resource efficiency of the approach is notable: certain photonic and ion-trap hardware realizations require only a single physical qubit, minimal ancillae, and O(L) single-qubit gates, minimizing coherence and parallelism requirements (Mauser et al., 7 Jul 2025, Dutta et al., 2021, Abe et al., 7 Jul 2025).

4. Experimental Realizations and Performance Benchmarks

Data re-uploading classifiers have been implemented on various hardware platforms:

• Superconducting qubits: Four-qubit chains with nearest-neighbor coupling, up to $x$1 trainable parameters, achieving $x$295% test accuracy on classical binary and multi-class benchmarks, $x$390% on image recognition tasks (downsized MNIST) (Tolstobrov et al., 2023).
• Photonic chips: Silicon photonic discrete-variable circuits, using heralded single photons, achieved $x$499% accuracy on 2D “circles” and “moons” datasets and $x$585% (MS-Overhead MNIST) (Mauser et al., 7 Jul 2025, Abe et al., 7 Jul 2025). Bosonic extensions reached $x$694% accuracy on a two-photon, two-mode device (Ono et al., 2022).
• Ion-trap quantum devices: Single trapped Ba$x$7 ions, arbitrary $x$8, $x$9 rotations, multi-class tasks with 4 layers; experimental accuracies closely match simulators, e.g., 96% (binary circle), 91–93% (multi-class) (Dutta et al., 2021).
• Room-temperature diamond NV-qubit hardware: Achieved near-ideal test performance (difference $R_z$02.5% relative to noiseless simulators) on unseen binary and trinary classification problems without hardware-aware retraining (Herrmann et al., 2023).

Optimization regimes were validated with comparative studies, showing e.g., Adam and L-BFGS-B yielding maximal generalization with moderate sample sizes, while layer-wise SMO is optimal in photon-limited scenarios (Aminpour et al., 2024, Abe et al., 7 Jul 2025).

Empirical performance is competitive with classical shallow neural networks and support vector machines at matched parameter budgets, with multi-qubit/entangled circuits further closing the gap and exceeding classical analogs on expressive multiclass boundaries (Pérez-Salinas et al., 2019, Tapia et al., 2022, Wach et al., 2023).

5. Architectural Variants: Qudit, Bosonic, and Quantum-Input Schemes

Qudit data re-uploading classifiers generalize the single-qubit model by employing $R_z$1 algebra generators ($R_z$2, $R_z$3, $R_z$4) and, crucially, a squeezing operator for universality at $R_z$5 (Wach et al., 2023). Expressivity is maximized by aligning measurement label states with the intrinsic “ladder” structure of the Hilbert space.

Bosonic classifiers extend the principle to photonic Fock spaces, realizing data encoding with passive linear-optical circuits and multi-photon input states. Theoretical analysis and experiment confirm that multilayered bosonic models, with both uncorrelated and entangled photon inputs, inherit universality and achieve high accuracy on non-trivial function classes (Ono et al., 2022).

Quantum-input data re-uploading architectures process general input density matrices by repeated ancilla–data interactions, implementing CPTP maps and leveraging mid-circuit resets for noise resilience and compact qubit usage. Universality is again established via polynomial expansions in input state parameters (Cha et al., 23 Sep 2025).

6. Resource–Accuracy Tradeoffs, Depth Scaling, and Tunability

Universality in data re-uploading quantum classifiers is robust to architectural constraints. Recent results establish that fixed-frequency (non-tunable) upload circuits, though constrained, retain universality: the expressiveness lost from removing frequency tunability can be recovered with only a polylogarithmic increase in circuit depth (Liu et al., 24 Jun 2026). The key result is that for target accuracy $R_z$6, a fixed-upload circuit of depth $R_z$7 for any $R_z$8 suffices, while lower bounds show that mismatch targets force at least $R_z$9.

Thus, the data re-uploading paradigm constitutes a universal function approximator family for both classical and quantum data, whose resource-accuracy scaling (in depth and trainable parameters) can be made quantitative and near-optimal under realistic constraints (Liu et al., 24 Jun 2026).

7. Design Principles, Limitations, and Research Directions

Emergent guidelines for building efficient data-reuploading classifiers include: (1) focus on classical performance first, seeking speedup only in subsequent hardware mapping; (2) preserve the intrinsic structure of the data when embedding (e.g., encode images as Hermitian matrices); (3) minimize classical preprocessing to avoid hybrid architectures that offload feature extraction; (4) avoid direct quantization of classical layers in favor of exploiting native quantum operations (e.g., direct Hamiltonian evolution); (5) realize nonlinearity via embedding, not solely via quantum measurement; (6) be aware of encoding-geometry mismatch that may introduce inductive bias (Wang et al., 2024).

Outstanding challenges include scaling to high qubit number while avoiding barren plateaus, hybrid quantum–classical co-design for best-in-class accuracy and resource usage, and the mapping from classical data structure to optimal quantum embedding schemes. Additionally, further study of the trainability landscape, generalization capacity (e.g., VC-dimension scaling O(dL)), and architecture-dependent noise robustness remains active (Mauser et al., 7 Jul 2025, Wach et al., 2023, Aminpour et al., 2024).

Data re-uploading universal quantum classifiers thus supply a versatile, resource-efficient, and hardware-compatible variational platform for quantum machine learning, with strong theoretical guarantees, mature experimental support, and a principled pathway toward larger-scale deployment (Tolstobrov et al., 2023, Pérez-Salinas et al., 2019, Wang et al., 2024, Mauser et al., 7 Jul 2025, Liu et al., 24 Jun 2026).