Single-Qubit Quantum Classifier

• The single-qubit quantum classifier is a supervised model that uses repeated data re-uploading to approximate any continuous decision function on compact domains.
• Architectures range from a minimal R_X circuit to ancilla-based designs like Hadamard, SWAP-test, and dissipative models that execute kernel evaluations.
• Experimental implementations on ion-trap and silicon-photonic platforms demonstrate accuracies near 90%, validating the practical feasibility of these compact classifiers.

A single-qubit quantum classifier is a supervised-learning model in which classification is carried by a single qubit used either as the entire parametrized processor, as the effective decision-carrying ancilla or label qubit, or as the central subsystem of a dissipative device. In the variational setting, a single qubit initialized in $|0\rangle$ undergoes repeated data-dependent and trainable rotations, and data re-uploading makes the resulting model a universal quantum classifier in the sense of approximating arbitrary continuous decision functions on compact domains (Pérez-Salinas et al., 2019). Other constructions use a single Hadamard- or SWAP-test readout qubit to evaluate distance or fidelity kernels in quantum parallel (Schuld et al., 2017, Blank et al., 2022, Pillay et al., 2023), while dissipative designs encode data in Lindblad operators acting on auxiliaries and read out the label from the steady state of a central qubit (Wang et al., 2023). Experimental demonstrations on ion-trap and silicon-photonic hardware show that such models are not merely formal: the ion-trap implementation carried out non-trivial classification tasks (Dutta et al., 2021), and the photonic implementation reported nearly $90\%$ accuracy even when the average number of photon samples per input was reduced to approximately two, provided that the training dataset was sufficiently large (Abe et al., 7 Jul 2025).

1. Conceptual scope and model families

In the strictest usage, a single-qubit quantum classifier is literally a one-qubit circuit. The minimal example is a binary classifier with one qubit initialized in $|0\rangle$, one parametrized rotation gate $R_X(\phi)$, and one measurement observable $M$, where the rotation angle is loaded as a linear form $\phi=\vec\theta^{\,T}\vec X$ of the input features (Zhang et al., 2022). A more expressive one-qubit family is the data-re-uploading architecture, in which the same classical input is encoded repeatedly across several single-qubit layers, with trainable parameters interleaved between encodings (Pérez-Salinas et al., 2019). Experimental ion-trap and silicon-photonic realizations both adopt this re-uploading perspective, though with different physical encodings and training procedures (Dutta et al., 2021, Abe et al., 7 Jul 2025).

A broader usage attaches the phrase to architectures in which a single qubit is the decisive readout degree of freedom rather than the whole device. In the distance-based interference classifier, the nontrivial processing after state preparation is a single Hadamard gate on an ancilla qubit, and the class is inferred from the statistics of that qubit together with a class qubit (Schuld et al., 2017). In the compact kernel classifier, label information is encoded in relative phases and the decision is extracted from a single-qubit measurement on the ancilla (Blank et al., 2022). In the multi-class SWAP-test classifier, all class information is compressed into the Bloch vector of one label qubit, and single-qubit tomography suffices for prediction (Pillay et al., 2023). A third variant replaces unitary circuit dynamics by engineered open-system evolution: a central qubit coupled to dissipative auxiliaries relaxes to a data-dependent steady state whose $\sigma_z$ expectation determines the class (Wang et al., 2023).

This plurality of meanings is not accidental. It reflects a common design principle: a small quantum core—often one qubit—implements the nonlinear map or kernel evaluation, while classical preprocessing, data loading, or auxiliary quantum subsystems provide the surrounding structure.

2. Data encoding, re-uploading, and expressive power

The central technical mechanism in the variational literature is data re-uploading. A representative form is

$$|\psi(\vec x,\Theta)\rangle = \prod_{i=1}^{L} U(\vec x,\vec\theta_i)\,|0\rangle,$$

or, in the photonic notation,

$$|\psi_{\mathrm{out}}\rangle = \hat U(\varphi_1)\hat U(\varphi_2)\cdots \hat U(\varphi_N)|0\rangle,$$

with each layer depending on both the classical input and trainable parameters (Pérez-Salinas et al., 2019, Abe et al., 7 Jul 2025). The reason this matters is that repeated, non-commuting single-qubit rotations generate trigonometric functions of linear forms in the input. The data-re-uploading paper makes this explicit and connects it to a universal-approximation argument: finite sums of periodic functions with trainable linear arguments can approximate arbitrary continuous target functions on compact domains, so a single qubit becomes universal once depth is allowed to substitute for width (Pérez-Salinas et al., 2019).

The ion-trap implementation instantiated this program with two ansätze. For multidimensional data it used

$$U_A(\vec{x}, \vec\theta) = R_z(\varphi)\; R_y(\vec{\omega}\cdot \vec x + \alpha),$$

and for two-dimensional data

$90\%$0

followed by comparison with label states on the Bloch sphere (Dutta et al., 2021). In this formulation, classification is geometric: the circuit maps $90\%$1 to a point on the Bloch sphere, and the predicted class is the label state with maximal fidelity to the output state.

At the opposite extreme, the single-$90\%$2 classifier uses only

$90\%$3

so all trainable parameters live inside one angle (Zhang et al., 2022). The paper explicitly interprets this as a linear model followed by a sinusoidal quantum readout. This sharply limits expressivity relative to re-uploading models, but it also makes the optimization landscape analytically tractable.

Kernel and interference constructions realize a different type of expressive map. In the distance-based classifier, amplitude-encoded data and a single ancilla Hadamard implement

$90\%$4

so the classifier is a uniformly weighted kernel machine with a distance-derived kernel (Schuld et al., 2017). The cosine-similarity classifier similarly encodes data into amplitudes and reduces prediction to the sign of a weighted cosine-similarity sum, obtained from the probability of measuring one control qubit in state $90\%$5 (Pastorello et al., 2021). The compact kernel classifier retains the same kernel logic while packing one positive-class vector into the real part and one negative-class vector into the imaginary part of a single compact amplitude-encoded register (Blank et al., 2022).

In the dissipative model, the data are not injected through unitary rotations at all. Instead, each feature determines a dissipative mode $90\%$6, which specifies Lindblad jump operators on auxiliary qubits. Adiabatic elimination then induces an effective single-qubit Lindblad equation on the central qubit, and the steady state $90\%$7 becomes the learned feature representation (Wang et al., 2023). This suggests that “single-qubit classifier” is best understood as a statement about the location of the decision dynamics, not about a unique encoding formalism.

3. Training objectives and optimization strategies

Training procedures vary substantially across single-qubit classifier families, but most exploit the low dimensionality of the one-qubit state space. In the universal data-re-uploading classifier, training is framed through fidelities to label states. One cost function minimizes

$90\%$8

while a weighted variant introduces class weights $90\%$9 and target overlaps $|0\rangle$0 for all classes simultaneously (Pérez-Salinas et al., 2019). This is close in spirit to a classical mean-squared-error objective on class scores.

The ion-trap implementation used a two-step hybrid procedure. Parameters were first optimized in simulation using classical optimizers such as CMA-ES and L-BFGS-B, and then fine-tuned on hardware by locally scanning subsets of parameters to maximize experimental accuracy (Dutta et al., 2021). The rationale was operational rather than theoretical: global optimization is cheap in simulation, whereas hardware calls are expensive and mainly needed to absorb systematic calibration errors.

The minimal $|0\rangle$1 classifier adopts a gradient-free coordinate-wise optimizer. Because expectation values with respect to one rotation parameter are sinusoidal,

$|0\rangle$2

the optimizer samples four angles, reconstructs the sinusoid analytically, and updates one parameter at a time without parameter-shift gradients (Zhang et al., 2022). The paper reports that this gradient-free optimization reaches high classification accuracy faster than Adam and shows good performance under bit flip, depolarizing, phase damping, phase flip, and amplitude damping noise at $|0\rangle$3 gate noise probability, although Adam can achieve higher final accuracy for phase flip and amplitude damping (Zhang et al., 2022).

The photonic classifier uses a more hardware-specific analytic strategy. Training minimizes a mean-squared-error cost,

$|0\rangle$4

where $|0\rangle$5 is estimated directly from photon counts (Abe et al., 7 Jul 2025). For each phase parameter $|0\rangle$6, the dependence

$|0\rangle$7

is reconstructed from only three phase settings, $|0\rangle$8, and layer-wise Sequential Minimal Optimization then analytically minimizes the reconstructed cost (Abe et al., 7 Jul 2025). Training is performed directly on the chip for both pseudo-single-photon and heralded single-photon regimes.

The dissipative classifier again uses a distinct objective. For state preparation it minimizes Bloch-vector distance to a desired qubit state; for classification it uses cross-entropy on a sigmoid of $|0\rangle$9, with the trainable variables being the coupling matrices between the central qubit and dissipative auxiliaries (Wang et al., 2023). The optimization is fully classical, but the model itself is quantum and open-system.

4. Interference, kernels, and single-qubit readout

A major branch of the literature treats the single qubit not as a variational processor but as a readout core for kernel evaluation. In the 2017 interference classifier, a superposition of a test vector and all training vectors is prepared, a single Hadamard is applied to an ancilla, and ancilla postselection plus class-qubit measurement implements a distance-based kernel rule (Schuld et al., 2017). The importance of this architecture lies less in raw accuracy than in the fact that the entire “quantum part” after state preparation is constant depth and conceptually concentrated in one qubit.

The compact kernel-based binary classifier pushes this compression further. Using compact amplitude encoding, two training vectors of opposite classes are stored in a single register as

$R_X(\phi)$0

and label information is carried by the real-versus-imaginary decomposition rather than by a separate label qubit (Blank et al., 2022). After a final ancilla Hadamard, the decision function becomes

$R_X(\phi)$1

so the classifier reduces to a single-qubit measurement of the ancilla (Blank et al., 2022). The same paper shows that, if the index register is sampled classically rather than encoded coherently, the architecture can be reduced to what it calls the smallest kernel-based binary classifier: $R_X(\phi)$2 qubits for data and one qubit for measurement (Blank et al., 2022).

The multi-class SWAP-test classifier generalizes the single-qubit-readout idea beyond binary tasks. A SWAP-test-like circuit aggregates class-wise kernel contributions into the Bloch vector of a label qubit,

$R_X(\phi)$3

and prediction is obtained by

$R_X(\phi)$4

where $R_X(\phi)$5 are pre-chosen label vectors on the Bloch sphere (Pillay et al., 2023). The number of qubits, the measurement strategy, and the circuit topology are invariant to the number of classes, and only three single-qubit measurement settings are needed for tomography of the label qubit (Pillay et al., 2023).

The one-class variant, QOCC, strips the older Hadamard classifier down still further. It removes the class qubit, interprets the ancilla probability $R_X(\phi)$6 directly as a membership degree, and in its minimal form removes the index qubit by storing only one prototype sample (Oliveira et al., 2020). This makes explicit a recurring theme: much of the classification logic can survive aggressive circuit compression when the single measured qubit is designed to carry a similarity score rather than a full quantum state description.

5. Experimental implementations and empirical behavior

Several experimental platforms now realize single-qubit classifiers in materially different ways.

Platform and model	Task	Representative outcome
Ion-trap re-uploading classifier (Dutta et al., 2021)	Circle; 4D hypersphere	Circle: $R_X(\phi)$7, $R_X(\phi)$8, tuned $R_X(\phi)$9; hypersphere: $M$0, $M$1, tuned $M$2
Silicon-photonic path-encoded classifier (Abe et al., 7 Jul 2025)	2D circular boundary with $M$3 training points	$M$4 accuracy at $M$5 photons per input; accuracy saturates above $M$6 for large $M$7
IBM interference circuit (Schuld et al., 2017)	Tiny Iris subset	Two test vectors correctly classified as $M$8 with 8192 shots
Minimal QOCC on IBM (Oliveira et al., 2020)	One-class Iris and Skin tasks	Minimal QOCC reached $M$9 on Iris and $\phi=\vec\theta^{\,T}\vec X$0 on Skin in simulation

The ion-trap experiment is historically significant because it is the first experimental implementation of a classification algorithm based on the re-uploading scheme (Dutta et al., 2021). It demonstrated not only binary circle and non-convex tasks but also multi-class problems such as tricrown, three circles, squares, and wavy lines, with quantum-processor performance tracking ideal simulation reasonably closely after local hardware tuning (Dutta et al., 2021).

The photonic experiment adds a different dimension: sample-starved training. The logical qubit is path encoded in two spatial modes of a silicon photonic integrated circuit, the photon source is heralded spontaneous four-wave mixing in a silicon waveguide, and the decisive quantity is the empirical probability

$\phi=\vec\theta^{\,T}\vec X$1

estimated from coincidence counts (Abe et al., 7 Jul 2025). The central statistical result is that the variance of the cost estimate scales as

$\phi=\vec\theta^{\,T}\vec X$2

so increasing the number of training examples $\phi=\vec\theta^{\,T}\vec X$3 and increasing the number of photon samples per input $\phi=\vec\theta^{\,T}\vec X$4 enter symmetrically at the level of cost variance (Abe et al., 7 Jul 2025). This suggests a concrete resource trade-off: low quantum sampling budgets can be compensated, at least partly, by larger classical datasets.

The IBM interference and one-class experiments are proof-of-principle demonstrations rather than broad benchmarks, but they are important because they show that ancilla-centric classifiers fit within very small superconducting devices and can tolerate modest noise levels while preserving the intended class decision (Schuld et al., 2017, Oliveira et al., 2020).

6. Limitations, misconceptions, and current directions

A persistent misconception is that a single qubit is too small to support nontrivial classification. The literature refutes this in a qualified way. A single qubit with only one encoded rotation is indeed severely limited and behaves like a linear model followed by a sinusoidal readout (Zhang et al., 2022). But repeated data re-uploading changes the functional class substantially and yields universal approximation in the single-qubit setting (Pérez-Salinas et al., 2019). The operative qualification is depth: finite-depth architectures remain capacity-limited. The photonic three-layer classifier cannot represent all circle positions equally well, and its ideal accuracy depends on the center of the target circle (Abe et al., 7 Jul 2025). The ion-trap results likewise show degradation on higher-dimensional tasks when depth is held fixed (Dutta et al., 2021).

A second misconception is terminological: “single-qubit classifier” does not always mean “one physical qubit total.” In interference and SWAP-test classifiers, many qubits may be used for data registers or index registers, but one ancilla or label qubit carries the decisive measurement statistic (Schuld et al., 2017, Blank et al., 2022, Pillay et al., 2023). The term therefore spans both single-qubit ansätze and single-qubit readout architectures.

A third misconception is that low-shot or photon-limited training is necessarily infeasible. The photonic study explicitly shows good overall accuracy for $\phi=\vec\theta^{\,T}\vec X$5 and $\phi=\vec\theta^{\,T}\vec X$6, and near-ideal behavior for $\phi=\vec\theta^{\,T}\vec X$7 when $\phi=\vec\theta^{\,T}\vec X$8 in simulation, though whether degradation for $\phi=\vec\theta^{\,T}\vec X$9 can be fully overcome by pushing $\sigma_z$0 far beyond $\sigma_z$1 remains open (Abe et al., 7 Jul 2025). The significance is not that shot noise disappears, but that its effect on the cost landscape can be statistically averaged.

The current frontier is diversification rather than convergence to one canonical architecture. The dissipative classifier replaces unitary circuit depth with engineered relaxation and demonstrates high accuracy and near-perfect AUC on linear, quadratic, and cubic two-dimensional tasks (Wang et al., 2023). Quaternionic constructions reinterpret fractional single-qubit gates as weighted components of classifier ensembles and activation-like nonlinear adders, suggesting a route toward quantum analogues of ensemble methods and neural activations (Widdows, 2022). On the networking side, partially-blind single-qubit classification proposes delegated inference in which the server knows that a classification is being performed but the data and outcome remain hidden, while a two-qubit extension enables verification (Pasini et al., 2 Jul 2026). Taken together, these directions suggest that the research problem is no longer whether a single qubit can classify, but which notion of “single-qubit classifier” is most suitable for a given combination of expressivity, trainability, hardware cost, and cryptographic or physical constraints.