A Multiclass Quantum Aligned Centroid Kernel

AI summary

Overview

Research area: Quantum machine learning (QML), specifically quantum kernel methods for supervised classification, with an emphasis on benchmarking and trainability.

Technical level: Intermediate. The paper assumes familiarity with kernel methods (Gram matrices, RBF kernels, support vector machines) and basic quantum computing concepts (qubit counts, fidelities, variational circuits). The appendices supply the circuit-level detail and a QKM introduction for readers who need it.

Scope: The paper introduces and empirically evaluates McQuack, a trainable multiclass quantum kernel that replaces the full training-set Gram matrix with a sample-to-class-centroid fidelity matrix, and tests it in simulation across more than 150 datasets and on 124-qubit subgraphs of two IBM devices.

What This Paper Is About

Standard kernel methods in machine learning carry three burdens: building the kernel matrix scales quadratically with the number of training samples, commonly used kernels such as RBF are fixed and cannot adapt to data, and most kernel formulations are binary, so multiclass problems require one-vs-one or one-vs-all expansions with extra cost. The authors propose a quantum kernel that sidesteps all three by comparing each data point only to a small set of trainable class centroids in Hilbert space, giving linear scaling in both training set size and number of classes. The goal is to determine whether such a model is competitive with existing quantum and classical baselines, and whether it can run on current large-scale quantum hardware.

Key Contributions

1. The McQuack algorithm. A multiclass quantum kernel method that optimizes a trainable feature map via quantum kernel target alignment (KTA) and achieves runtime complexity of O(n_train · M) in the number of training samples and classes, rather than the O(n_train²) cost of full-Gram kernels or the O(M²) cost of one-vs-one multiclass extensions.

2. Hardware inference at 124 qubits. An evaluation of the untrained model on 124-qubit subgraphs of two 156-qubit IBM devices, demonstrating that fidelity evaluation remains feasible at that scale and yielding accuracy in the same range as an untrained RBF baseline.

3. Extensive benchmarking. Evaluation on the QML Benchmark Suite comprising 133 synthetic binary datasets across five classification tasks, plus 19 real-world datasets with two to ten classes, against seven quantum models from the benchmark suite, an MLPClassifier, a NystroemSVM, an SVC, a variational classifier ensemble, and a two-layer neural network.

4. Trainability analysis. A study of initialization and regularization showing no evidence of barren plateaus for up to 13 qubits, with parameter initialization scale and distribution identified as by far the dominant factor in successful optimization.

Main Findings

• Hardware accuracy (untrained model): 0.83 ± 0.02 on ibm_pittsburgh and 0.79 ± 0.02 on ibm_aachen, compared with 0.86 for the SVC and 0.80 for SimpleKernel (an untrained RBF kernel). Differences were not statistically significant at α = 0.05 given the limited number of runs, so the authors frame these results as a feasibility demonstration rather than evidence of hardware competitiveness. Statistics were collected from three runs of each model and dataset.

• Device-dependent trend: The better result on ibm_pittsburgh is attributed to its more advanced processor and lower error rates; on that device McQuack was in the same range as the untrained RBF baseline.

• Synthetic benchmark ranking: On the 133 synthetic datasets, the classical MLPClassifier achieved the best average normalized rank of 0.21, the DressedQuantumCircuitClassifier ranked 0.27, McQuack ranked 0.33, and NystroemSVM ranked 0.65. McQuack therefore ranked as the strongest "pure" quantum model (models without trainable classical pre- or post-processing), ahead of the DataReuploadingClassifier and the other quantum models, though the DressedQuantumCircuitClassifier is a hybrid model that was not evaluated on all 133 datasets.

• Centroids versus random landmarks: McQuack ranked better than NystroemSVM, which the authors interpret as evidence that trainable centroids in Hilbert space give a more effective representation than a small number of randomly sampled RBF landmark points. In the synthetic benchmark NystroemSVM used six landmark points, giving roughly 97% probability that the approximation includes at least one point from each class.

• Real-world benchmark ranking: On the 19 real-world datasets, the two-layer neural network had the best overall rank of 0.17 for AUC and 0.10 for balanced accuracy. McQuack had mean ranks of 0.50 (AUC) and 0.49 (balanced accuracy), ranking below the SVC on AUC but above it on balanced accuracy despite the SVC's quadratic scaling. McQuack also ranked better than NystroemSVM (which used 3M landmark points there) and better than an ensemble of variational classifiers.

• Shared parameters help: Outperforming the variational classifier ensemble indicates that partially shared trainable parameters confer an advantage over purely individual parameters for multiclass problems.

• No barren plateaus observed: Across experiments with up to 13 qubits, the authors found no evidence of barren plateaus. The fANOVA analysis identified initialization scale and distribution (init_small) as the dominant factor with relative importance exceeding 0.7, while all remaining hyperparameters contributed less than 0.2. Small, normally distributed initial parameters produced stable gradients and improved convergence; large initializations led to oscillatory behavior. Effects of regularization and identity-based initialization were less pronounced.

• Hardware training is currently impractical: Training one epoch (one KAO iteration plus one CO iteration) for the hardware dataset and architecture would require approximately 1.5 × 10⁸ shots by the authors' shot-count formula. At roughly two seconds per 100 shots, that is approximately 827 hours, or over 34 days of QPU time, which is why hardware experiments were restricted to inference without training.

• Kernel alignment bound: The alignment measure used for training is upper bounded by 1/√M for M classes, and higher values indicate stronger alignment between the estimated sample-to-centroid matrix and the ideal label matrix.

Methodology in Plain English

The model keeps a single trainable data-encoding circuit that maps each input into a quantum state, plus one trainable "centroid" state per class. For a given sample and a given class, the model computes the fidelity (the squared overlap) between the sample's encoded state and that class's centroid state; the predicted label is simply the class whose centroid has the highest fidelity with the sample.

Rather than building an n_train × n_train matrix of comparisons between all pairs of training samples, the method builds only an n_train × M matrix of sample-to-centroid fidelities. Training alternates between two steps. In the kernel alignment optimization step, the shared embedding weights and biases are updated to maximize kernel target alignment against an ideal ±1 matrix derived from the labels, with the centroids held fixed. In the centroid optimization step, the centroid parameters are updated with the embedding fixed. This alternation repeats over a number of epochs. Because the encoding re-encodes features cyclically when there are more rotation gates than input dimensions, the circuit performs data re-uploading, which the authors note connects the model to a truncated Fourier series whose accessible frequencies grow linearly with the number of re-uploading repetitions.

The embedding uses hardware-efficient layers, each consisting of general single-qubit rotation gates followed by a ring of CNOTs. The authors also map a single classifier onto the variational classifier formalism, showing that McQuack differs from a standard variational classifier ensemble by sharing the embedding parameters across all classes while keeping only centroid parameters class-specific.

Experiments proceed in three stages: inference without training on 124-qubit hardware subgraphs using dynamical decoupling, measurement twirling, and a bit flip tolerance procedure with a calibration tailored to the algorithm; simulation-based benchmarking against classical and quantum baselines on the QML Benchmark Suite and 19 real-world datasets; and a trainability study varying regularization and initialization strategies.

Why This Matters

Impact on research. The paper targets a structural weakness in quantum kernel methods rather than a narrow accuracy gain: it removes the quadratic Gram-matrix bottleneck and provides a native multiclass formulation, while pairing the proposal with the kind of large-scale, critical benchmarking the QML community has been urged to adopt. The negative and mixed results are reported openly, including the failure to beat simple classical baselines, the absence of hardware training, and the use of a noiseless simulator. The trainability study and the observation about shared parameters give other researchers concrete design guidance.

Real-world applications.

• Security-sensitive classification tasks, which the authors name explicitly as a motivation for future adversarial-robustness work, consistent with the authors' affiliation at Fraunhofer AISEC.
• Multiclass pattern recognition where training sets are large and quadratic kernel construction is the bottleneck, since the method scales linearly in training samples and classes.
• Resource-constrained deployment scenarios in which a kernel needs few trainable parameters and few reference points per class.
• Near-term quantum hardware workloads, where fidelity estimation between encoded states is the only primitive required at inference time, as demonstrated at 124 qubits.

Industry relevance. The results give a realistic picture of where quantum kernels sit relative to classical methods: superior to other pure quantum models, competitive with an SVC on balanced accuracy, but still behind feedforward neural networks and other hybrid approaches. The explicit accounting of QPU time for training (over 34 days for a single epoch) is directly useful for anyone planning hardware experiments, and the dependence on error rates across two devices illustrates how hardware quality, not just algorithm design, governs near-term viability.

Future Directions

• Reducing the cost of gradient estimation so that training, not just inference, becomes feasible on quantum hardware; the parameter-shift rule's shot overhead is identified as the main blocker.
• Systematic comparison of McQuack's kernel with other quantum kernels, and deeper study of parameter initialization, specifically disentangling the contributions of initialization magnitude from initialization distribution, which the authors leave as an open question.
• Further investigation of variational classifier ensembles with partially shared parameters for multiclass classification, a direction the authors explicitly encourage based on the equivalence sketched in the mapping section.
• Assessing robustness against adversarial perturbations to judge suitability for security-sensitive settings, and extending the trainability analysis beyond 13 qubits, since optimization landscapes may differ drastically at larger scale.

Target Audience

This paper suits quantum machine learning researchers and graduate students working on quantum kernels, kernel target alignment, or multiclass QML formulations; benchmarking-focused researchers who want an example of rigorous comparison against the QML Benchmark Suite; and practitioners at the boundary of quantum algorithms and applied security who need a realistic assessment of what current hardware can and cannot do. Readers without a kernel-methods background will need the background and appendix sections, while specialists can go directly to the benchmarking and trainability results.

Authors’ abstract

Kernel methods are powerful tools in machine learning but commonly used full-Gram kernels face three key limitations: (1) quadratic scaling with training set size; (2) the use of fixed, non-trainable kernels; and (3) the absence of an intrinsic formulation for multiclass classification. We present McQuack, a trainable quantum kernel method for multiclass problems that achieves linear scaling in the number of training samples. This is accomplished by replacing the full training-set Gram matrix with a trainable sample-to-(class-centroid) fidelity matrix. We evaluate the model in simulation and on 124 qubits of two IBM devices, across more than 150 datasets. In simulation, McQuack outperforms existing "pure" quantum baselines, while results from hardware inference -- obtained without training -- achieve performance similar to an RBF kernel. Finally, we study the trainability of the model and observe no evidence of barren plateaus in our experiments with up to 13 qubits, and highlight the importance of parameter initialization for successful optimization.