Quantum Kernel Classifiers

• Quantum kernel classifiers are classification algorithms that map classical data into quantum state space using fidelity-based or Hilbert–Schmidt kernels.
• They employ quantum feature maps—such as angle encoding and IQP-style circuits—to create Gram matrices for robust classical SVM decision making.
• Innovations like Hamming quantum kernels and trainable variants address scalability challenges by optimizing kernel geometry and reducing shot complexity.

Quantum kernel classifiers are classification algorithms that rely on quantum-generated kernels. They map an input $x$ to a quantum state or operator, evaluate a similarity such as $$K(x,x')=\bigl|\langle \phi(x')\mid \phi(x)\rangle\bigr|^2$$ or $K(\rho,\sigma)=\mathrm{Tr}(\rho\,\sigma)$, assemble a Gram matrix from these pairwise overlaps, and then apply a classical decision rule, most commonly a support vector machine (SVM), in the induced feature space (Agnihotri et al., 29 May 2026, Park et al., 2020). In contemporary usage, the term covers not only fidelity-based quantum support vector machines (QSVMs), but also multiclass extensions, ensemble constructions, trainable-kernel variants, and classifiers derived from quantum state discrimination.

1. Mathematical foundation

The standard binary formulation mirrors the classical kernel method. Given training data $\{(x_i,y_i)\}_{i=1}^N$ with $y_i\in\{+1,-1\}$, the decision function has the form

$$f(x)=\mathrm{sgn}\!\left(\sum_{i=1}^N \alpha_i y_i\,K(x_i,x)+b\right),$$

where the coefficients $\alpha_i$ and bias $b$ are obtained from a classical optimization in dual form (Park et al., 2020). The distinctive element is therefore not the SVM optimization itself, but the kernel evaluation step.

For pure-state embeddings $|\psi(x)\rangle$, the most common kernel is the squared overlap,

$$K(\psi(x),\psi(x'))=\bigl|\langle\psi(x)\mid\psi(x')\rangle\bigr|^2.$$

For mixed states or operator-valued encodings, a more general choice is the Hilbert–Schmidt inner product,

$$K(x,x')=\bigl|\langle \phi(x')\mid \phi(x)\rangle\bigr|^2$$0

which reduces to the squared overlap on pure states (Park et al., 2020). A common misconception is that quantum kernel classifiers are restricted to pure-state fidelities. The mixed-state construction shows that the kernel formalism extends naturally to density operators and operator-valued data.

The legitimacy of these kernels as Mercer kernels is explicit in the literature. The function $$K(x,x')=\bigl|\langle \phi(x')\mid \phi(x)\rangle\bigr|^2$$1 is symmetric and positive semidefinite, and $$K(x,x')=\bigl|\langle \phi(x')\mid \phi(x)\rangle\bigr|^2$$2 is likewise positive semidefinite; accordingly, Mercer's theorem guarantees the existence of an RKHS in which the feature map can be written as $$K(x,x')=\bigl|\langle \phi(x')\mid \phi(x)\rangle\bigr|^2$$3 (Park et al., 2020). This places quantum kernel classifiers on the same functional-analytic footing as classical kernel methods, while allowing the feature space to be realized by quantum state space rather than by an explicitly stored vector representation.

2. Quantum feature maps and kernel estimation

A quantum kernel classifier begins with a feature map $$K(x,x')=\bigl|\langle \phi(x')\mid \phi(x)\rangle\bigr|^2$$4. Several constructions recur across the literature. One minimal instance uses angle encoding with

$$K(x,x')=\bigl|\langle \phi(x')\mid \phi(x)\rangle\bigr|^2$$5

after scaling each feature into $$K(x,x')=\bigl|\langle \phi(x')\mid \phi(x)\rangle\bigr|^2$$6 (Pinheiro et al., 12 Sep 2025). Other families include IQP-style feature maps with single-qubit $$K(x,x')=\bigl|\langle \phi(x')\mid \phi(x)\rangle\bigr|^2$$7 layers and pairwise $$K(x,x')=\bigl|\langle \phi(x')\mid \phi(x)\rangle\bigr|^2$$8 entanglers, as well as product kernels generated by Pauli-$$K(x,x')=\bigl|\langle \phi(x')\mid \phi(x)\rangle\bigr|^2$$9, Pauli-$K(\rho,\sigma)=\mathrm{Tr}(\rho\,\sigma)$0, or Pauli-$K(\rho,\sigma)=\mathrm{Tr}(\rho\,\sigma)$1 rotations; in the latter case the kernel takes the closed form $K(\rho,\sigma)=\mathrm{Tr}(\rho\,\sigma)$2 (Ding et al., 2024).

The canonical kernel-evaluation circuit is a Loschmidt-echo or overlap circuit. One prepares $K(\rho,\sigma)=\mathrm{Tr}(\rho\,\sigma)$3, applies $K(\rho,\sigma)=\mathrm{Tr}(\rho\,\sigma)$4, then $K(\rho,\sigma)=\mathrm{Tr}(\rho\,\sigma)$5, measures in the computational basis, and interprets the probability of the all-zero outcome as the kernel estimate,

$K(\rho,\sigma)=\mathrm{Tr}(\rho\,\sigma)$6

(Pinheiro et al., 12 Sep 2025). This projective-measurement strategy replaces more costly SWAP tests in many QSVM implementations, although swap-test and Hadamard-test classifiers remain important alternative constructions (Park et al., 2020).

The same abstract procedure has been specialized to distinct hardware modalities. In gate-based neutral-atom architectures, a $K(\rho,\sigma)=\mathrm{Tr}(\rho\,\sigma)$7-qubit ZZ-feature map was built from Raman-channel single-qubit rotations and Rydberg-blockade entangling gates, with the method generalized to $K(\rho,\sigma)=\mathrm{Tr}(\rho\,\sigma)$8 qubits by exploiting flexible neutral-atom connectivity (Russo et al., 2023). In NMR, a $K(\rho,\sigma)=\mathrm{Tr}(\rho\,\sigma)$9-qubit star-topology register was used to encode classical data via data-dependent unitaries acting on multiple quantum coherence orders, and a double-layered star register was proposed to define kernels on non-parametrized operator inputs such as entangling versus non-entangling quantum operations (Sabarad et al., 2024). These realizations illustrate that quantum kernel estimation is a platform-level primitive rather than a single-circuit recipe.

3. Classifier constructions beyond the basic QSVM

The dominant realization is the quantum-kernel SVM. In the binary case, the quantum device provides $\{(x_i,y_i)\}_{i=1}^N$0, and a classical SVM solver produces the separating hyperplane in feature space. For multiclass problems, one-vs-rest and two-step hierarchical decompositions are standard: the former trains one binary QSVM per class and selects the largest decision score, whereas the latter isolates an “easy” or majority class first and delegates the remainder to a second classifier (Pinheiro et al., 12 Sep 2025).

Several architectures generalize the kernel-based paradigm rather than the kernel itself. A quantum random forest constructs a decision directed acyclic graph of QSVM nodes, using bagging and information-gain-based splits; to reduce kernel-evaluation cost, it employs a low-rank Nyström approximation and derives explicit generalization-error and finite-sampling bounds (Srikumar et al., 2022). A hybrid quantum-classical radial basis function network replaces the classical radial kernel by a quantum fidelity kernel $\{(x_i,y_i)\}_{i=1}^N$1, solves for the output weights by pseudoinverse, and performs multiclass classification natively through a vector-valued output rather than one-vs-all or one-vs-one decomposition (Micklethwaite et al., 23 Dec 2025).

A distinct line of work derives classifiers from quantum state discrimination. In the Pretty Good Measurement (PGM) construction, each datum is embedded as $\{(x_i,y_i)\}_{i=1}^N$2, class averages define an ensemble $\{(x_i,y_i)\}_{i=1}^N$3, and the decision rule is

$\{(x_i,y_i)\}_{i=1}^N$4

with $\{(x_i,y_i)\}_{i=1}^N$5 (Cruzeiro et al., 2023). Because the associated kernel becomes $\{(x_i,y_i)\}_{i=1}^N$6 for $\{(x_i,y_i)\}_{i=1}^N$7 tensor copies, this approach shows that “quantum kernel classifier” need not mean “QSVM”: the kernel can mediate state-discrimination-based multiclass decision rules as well.

Trainable-kernel variants extend the same logic to the feature map itself. A parameterized quantum feature map $\{(x_i,y_i)\}_{i=1}^N$8 can be optimized through a clustering objective so that same-class states align with orthonormal label states, after which the trained fidelity kernel is used inside an improved QSVM or an iterative multiclass framework (Xu et al., 7 May 2025). This suggests a shift from fixed-feature-map kernel evaluation toward kernel learning in the stricter sense.

4. Expressivity, concentration, and scalable kernel design

Two lines of theory govern expressivity. The first is constructive and problem-dependent. A feature-map analysis based on “minimum accuracy” $\{(x_i,y_i)\}_{i=1}^N$9 provides a lower bound on exact hard-margin SVM training accuracy from the Pauli-decomposition coordinates of $y_i\in\{+1,-1\}$0; empirically, selecting the encoding with maximal $y_i\in\{+1,-1\}$1 reproduced the best-performing classifier on the tested $y_i\in\{+1,-1\}$2-qubit datasets (Suzuki et al., 2019). The same work also shows that weighted sums of quantum kernels,

$y_i\in\{+1,-1\}$3

correspond to a direct-sum enlargement of feature space and can improve separability (Suzuki et al., 2019).

The second line is complexity-theoretic. By embedding the $y_i\in\{+1,-1\}$4-Forrelation promise problem into a classification task, one can construct a feature map and a quantum kernel such that QSVMs solve a PromiseBQP-complete problem efficiently. Under the stated assumptions, this implies that there exists a feature map and a quantum kernel that make quantum-kernel SVMs efficient solvers for any BQP problem (Jäger et al., 2022). This establishes universal expressiveness in a formal worst-case sense, though not a guarantee of advantage on typical empirical datasets.

The central scalability obstacle in practical QSVMs is exponential concentration of the fidelity kernel. For the standard fidelity kernel

$y_i\in\{+1,-1\}$5

the mean off-diagonal overlap for generic states scales as $y_i\in\{+1,-1\}$6, and fluctuations are exponentially suppressed. Under a polynomial shot budget, the kernel matrix therefore tends toward the identity and classification becomes near-random at larger $y_i\in\{+1,-1\}$7 (Agnihotri et al., 29 May 2026). This phenomenon is now a primary reference point in the design of scalable quantum kernels.

A recent response is the Hamming quantum kernel (HQK), which keeps the same circuit $y_i\in\{+1,-1\}$8 but uses the full output distribution rather than only the $y_i\in\{+1,-1\}$9 outcome. Writing $$f(x)=\mathrm{sgn}\!\left(\sum_{i=1}^N \alpha_i y_i\,K(x_i,x)+b\right),$$0 for the probability of bitstring $$f(x)=\mathrm{sgn}\!\left(\sum_{i=1}^N \alpha_i y_i\,K(x_i,x)+b\right),$$1 and $$f(x)=\mathrm{sgn}\!\left(\sum_{i=1}^N \alpha_i y_i\,K(x_i,x)+b\right),$$2 for its Hamming weight, the kernel is

$$f(x)=\mathrm{sgn}\!\left(\sum_{i=1}^N \alpha_i y_i\,K(x_i,x)+b\right),$$3

It interpolates between the trivial constant kernel at $$f(x)=\mathrm{sgn}\!\left(\sum_{i=1}^N \alpha_i y_i\,K(x_i,x)+b\right),$$4 and the fidelity kernel at $$f(x)=\mathrm{sgn}\!\left(\sum_{i=1}^N \alpha_i y_i\,K(x_i,x)+b\right),$$5, requires no additional quantum resources, and can be estimated with polynomial shot complexity by classical post-processing of the same measurement outcomes used for fidelity estimation (Agnihotri et al., 29 May 2026).

A related mitigation strategy is bandwidth control inside the feature map. In large-scale hyperspectral classification, fidelity-kernel SVMs were simulated on inputs with hundreds of spectral bands by tensor-network contraction and GPU acceleration, with quadratic scaling $$f(x)=\mathrm{sgn}\!\left(\sum_{i=1}^N \alpha_i y_i\,K(x_i,x)+b\right),$$6 in the number of qubits for the chosen linear-entanglement circuit family. The study emphasizes that bandwidth optimization is crucial for mitigating exponential concentration and ensuring generalization (Delilbasic et al., 17 May 2026). Taken together, these results indicate that scalable quantum kernel design is increasingly driven by kernel geometry and post-processing, not only by deeper or more entangling circuits.

5. Sampling complexity, reliability, noise, and adversarial robustness

Quantum kernel classification is often limited less by training optimization than by the cost of estimating overlaps. In the Hadamard-test and swap-test classifiers, the measured observable has variance $$f(x)=\mathrm{sgn}\!\left(\sum_{i=1}^N \alpha_i y_i\,K(x_i,x)+b\right),$$7, so $$f(x)=\mathrm{sgn}\!\left(\sum_{i=1}^N \alpha_i y_i\,K(x_i,x)+b\right),$$8 shots are required to estimate the classification score to additive precision $$f(x)=\mathrm{sgn}\!\left(\sum_{i=1}^N \alpha_i y_i\,K(x_i,x)+b\right),$$9. For fixed circuit structure, this variance is already optimal; further reduction is possible only through data-related parameter choices such as weight optimization (Park et al., 2021). The same work also shows that the two-qubit readout can be reduced to a single-qubit measurement by inserting a CNOT before measurement, and that under several Pauli-noise models the noisy expectation value is only rescaled rather than sign-flipped (Park et al., 2021).

A more aggressive route is to replace repeated sampling by Quantum Amplitude Estimation (QAE). In simplified Hadamard and swap classifiers, the classification score is encoded into one ancilla amplitude, and QAE reduces the precision cost from $\alpha_i$0 circuit executions to $\alpha_i$1 applications of the Grover iterate. The same construction also removes one qubit and reduces class-labeling depth linearly with respect to the number of sample data (Lee et al., 2024). This is a genuine algorithmic improvement in shot complexity rather than merely a hardware-level measurement refinement.

Another important correction concerns the performance target itself. Exact kernel estimation is not necessary if the goal is correct classification. The Shot-frugal and Robust (ShofaR) framework therefore defines reliability

$\alpha_i$2

and derives shot bounds in terms of margins rather than absolute kernel-entry precision. Using chance-constrained programming and subgaussian kernel-noise bounds, ShofaR produces convex robust programs that can reduce measurement counts substantially; on the circles dataset, reliability $\alpha_i$3 was reached at $\alpha_i$4 for ShofaR-Primal, whereas a naive stochastic-kernel classifier needed $\alpha_i$5 (Shastry et al., 2022). A common misconception is therefore that quantum kernel methods must estimate the full kernel matrix accurately entry by entry; margin-based reliability bounds show that classification can be far more shot-frugal than kernel reconstruction.

Adversarial robustness has emerged as a separate issue. Hybrid quantum classifiers based on quantum kernels and SVMs are vulnerable to white-box evasion attacks: small engineered perturbations can flip the sign of the decision function. However, simple adversarial retraining with a small number of crafted perturbations can restore robustness; in Medical MNIST simulations, adding about $\alpha_i$6 adversarial points to the training set returned test accuracy on both clean and adversarial held-out points to $\alpha_i$7 (Montalbano et al., 2024). By contrast, HHL-based least-squares SVM variants have been reported as highly noise-sensitive on real NISQ hardware, even when QSVMs remain comparatively stable (Pinheiro et al., 12 Sep 2025).

6. Empirical performance across domains

Empirical studies now span classical vision tasks, astronomical data, remote sensing, speech, synthetic quantum data, and classification of quantum operations. The resulting picture is mixed: some benchmarks show parity with strong classical baselines, some show moderate gains, and some show that quantum kernels are useful primarily when the feature map matches the structure of the data.

Setting	Quantum-kernel result	Comparator
MNIST $\alpha_i$8 vs $\alpha_i$9	$b$0 remains above $b$1 for all $b$2	$b$3 drops below $b$4 for $b$5 (Agnihotri et al., 29 May 2026)
Reduced SDSS, two-step	QSVM accuracy $b$6	classical SVM $b$7; HHL LS-QSVM $b$8 (Pinheiro et al., 12 Sep 2025)
Georgian spoken commands	Gaussian-QKL $b$9	DNN $|\psi(x)\rangle$0; QCNN-DNN $|\psi(x)\rangle$1 (Yang et al., 2022)
Quantum entanglement classification	$|\psi(x)\rangle$2 accuracy on $|\psi(x)\rangle$3 uniform samples	classical Gaussian RBF fails outside training region (Sabarad et al., 2024)
Indian Pines, $|\psi(x)\rangle$4-class, $|\psi(x)\rangle$5 bands	Quantum SVM $|\psi(x)\rangle$6	Classical RBF SVM $|\psi(x)\rangle$7 (Delilbasic et al., 17 May 2026)

The Hamming-kernel study is particularly important because it isolates a failure mode of the standard fidelity kernel and remedies it without changing the quantum circuit. On MNIST $|\psi(x)\rangle$8 vs $|\psi(x)\rangle$9, all kernels reach $$K(\psi(x),\psi(x'))=\bigl|\langle\psi(x)\mid\psi(x')\rangle\bigr|^2.$$0 accuracy at $$K(\psi(x),\psi(x'))=\bigl|\langle\psi(x)\mid\psi(x')\rangle\bigr|^2.$$1, but $$K(\psi(x),\psi(x'))=\bigl|\langle\psi(x)\mid\psi(x')\rangle\bigr|^2.$$2 falls below $$K(\psi(x),\psi(x'))=\bigl|\langle\psi(x)\mid\psi(x')\rangle\bigr|^2.$$3 once $$K(\psi(x),\psi(x'))=\bigl|\langle\psi(x)\mid\psi(x')\rangle\bigr|^2.$$4, while $$K(\psi(x),\psi(x'))=\bigl|\langle\psi(x)\mid\psi(x')\rangle\bigr|^2.$$5 stays near $$K(\psi(x),\psi(x'))=\bigl|\langle\psi(x)\mid\psi(x')\rangle\bigr|^2.$$6; on synthetic quantum data, $$K(\psi(x),\psi(x'))=\bigl|\langle\psi(x)\mid\psi(x')\rangle\bigr|^2.$$7 remains within $$K(\psi(x),\psi(x'))=\bigl|\langle\psi(x)\mid\psi(x')\rangle\bigr|^2.$$8 of the theoretical distinguishability $$K(\psi(x),\psi(x'))=\bigl|\langle\psi(x)\mid\psi(x')\rangle\bigr|^2.$$9 across the full range $$K(x,x')=\bigl|\langle \phi(x')\mid \phi(x)\rangle\bigr|^2$$00 (Agnihotri et al., 29 May 2026).

On multiclass astronomy data, the comparison is more conservative. For a reduced SDSS dataset with $$K(x,x')=\bigl|\langle \phi(x')\mid \phi(x)\rangle\bigr|^2$$01 features, QSVM outperforms HHL LS-SVM in most cases, but quantum and classical methods achieve broadly similar performance, with classical models performing slightly ahead overall (Pinheiro et al., 12 Sep 2025). This is representative of a broader empirical pattern: current quantum kernels are often competitive rather than uniformly dominant.

Beyond SVMs, ensemble and multiclass constructions have also produced nontrivial gains. On quantum-relabeled binary tasks, a quantum random forest with $$K(x,x')=\bigl|\langle \phi(x')\mid \phi(x)\rangle\bigr|^2$$02–$$K(x,x')=\bigl|\langle \phi(x')\mid \phi(x)\rangle\bigr|^2$$03, $$K(x,x')=\bigl|\langle \phi(x')\mid \phi(x)\rangle\bigr|^2$$04, $$K(x,x')=\bigl|\langle \phi(x')\mid \phi(x)\rangle\bigr|^2$$05, and $$K(x,x')=\bigl|\langle \phi(x')\mid \phi(x)\rangle\bigr|^2$$06 outperformed a single QSVM by up to $$K(x,x')=\bigl|\langle \phi(x')\mid \phi(x)\rangle\bigr|^2$$07–$$K(x,x')=\bigl|\langle \phi(x')\mid \phi(x)\rangle\bigr|^2$$08 and classical random forests by $$K(x,x')=\bigl|\langle \phi(x')\mid \phi(x)\rangle\bigr|^2$$09–$$K(x,x')=\bigl|\langle \phi(x')\mid \phi(x)\rangle\bigr|^2$$10, while using $$K(x,x')=\bigl|\langle \phi(x')\mid \phi(x)\rangle\bigr|^2$$11–$$K(x,x')=\bigl|\langle \phi(x')\mid \phi(x)\rangle\bigr|^2$$12 fewer shots (Srikumar et al., 2022). In multiclass kernel simulations on six real-world datasets, the optimal quantum kernel achieved $$K(x,x')=\bigl|\langle \phi(x')\mid \phi(x)\rangle\bigr|^2$$13 accuracy on Iris, $$K(x,x')=\bigl|\langle \phi(x')\mid \phi(x)\rangle\bigr|^2$$14 on Ecoli, and $$K(x,x')=\bigl|\langle \phi(x')\mid \phi(x)\rangle\bigr|^2$$15 on Vowel, with the reported best quantum kernel outperforming the best classical kernel on each dataset by $$K(x,x')=\bigl|\langle \phi(x')\mid \phi(x)\rangle\bigr|^2$$16–$$K(x,x')=\bigl|\langle \phi(x')\mid \phi(x)\rangle\bigr|^2$$17 percentage points (Ding et al., 2024).

7. Limitations, controversies, and research directions

The most persistent controversy concerns quantum advantage. On the one hand, quantum kernels can be embedded in PromiseBQP-complete classification problems and thereby exhibit universal expressiveness in principle (Jäger et al., 2022). On the other hand, some practically studied fidelity-kernel families with shallow or low-entangling circuits remain classically simulable, so no provable quantum advantage follows from those families alone (Delilbasic et al., 17 May 2026). The gap between worst-case expressiveness and experimentally accessible advantage remains central.

A second limitation is resource scaling. QSVM training requires $$K(x,x')=\bigl|\langle \phi(x')\mid \phi(x)\rangle\bigr|^2$$18 kernel evaluations, with additional dependence on feature count because the kernel circuit itself scales with the number of encoded features; in direct empirical comparisons, this quadratic sample scaling is one reason large datasets remain difficult on NISQ hardware (Pinheiro et al., 12 Sep 2025). Even when kernel estimation is feasible, many studies still rely on noiseless simulation or classically emulated kernels, and real-device performance depends strongly on readout noise, gate fidelity, and state-preparation quality (Micklethwaite et al., 23 Dec 2025).

A third issue is kernel geometry. Fidelity kernels may collapse through concentration, while broader families such as the Hamming kernel require hyperparameter tuning, here through $$K(x,x')=\bigl|\langle \phi(x')\mid \phi(x)\rangle\bigr|^2$$19, to avoid degenerating either to a constant kernel or back to the fidelity limit (Agnihotri et al., 29 May 2026). This suggests that future progress may come less from a single canonical quantum kernel and more from systematic kernel design, including post-processing families beyond Hamming weighting, projected or partial-measurement kernels, randomized feature maps, and trainable embeddings.

Finally, several directions remain open in the sources. These include analytical bounds on concentration and generalization for distribution-informed kernels, extensions to multi-class and regression settings, approximate Gram-matrix inversion and Nyström-style compression, implementation on hardware with error mitigation, and better integration of quantum kernels with state-discrimination readouts and multiclass native architectures (Agnihotri et al., 29 May 2026, Park et al., 2020). The accumulated literature indicates that quantum kernel classifiers are now best understood not as a single algorithm, but as a family of hybrid methods whose empirical value depends on feature-map choice, kernel-estimation strategy, robustness criterion, and the extent to which the underlying kernel geometry resists concentration and noise.