The rapid advancement in Quantum Computing, particularly through Noisy-Intermediate Scale Quantum (NISQ) devices, has spurred significant interest in Quantum Machine Learning (QML) applications. Despite their potential, fully-quantum algorithms remain impractical due to the limitations of current NISQ devices. Hybrid quantum-classical neural networks (HQNNs) have emerged as a viable alternative, leveraging both quantum and classical computations to enhance machine learning capabilities. However, the constrained resources of NISQ devices, particularly the limited number of qubits, pose significant challenges for executing large-scale quantum circuits. This work addresses these current challenges by proposing a novel and practical methodology for quantum circuit cutting of HQNNs, allowing large quantum circuits to be executed on limited-qubit NISQ devices. Our approach not only preserves the accuracy of the original circuits but also supports the training of quantum parameters across all subcircuits, which is crucial for the learning process in HQNNs. We propose a cutting methodology for HQNNs that employs a greedy algorithm for identifying efficient cutting points, and the implementation of trainable subcircuits, all designed to maximize the utility of NISQ devices in HQNNs. The findings suggest that quantum circuit cutting is a promising technique for advancing QML on current quantum hardware, since the cut circuit achieves comparable accuracy and much lower qubit requirements than the original circuit. The code is available at https://github.com/eBrain4Everyone/QNN-Cutting.
Circuit cutting partitions a large quantum circuit into smaller subcircuits that can be executed independently and recombined by classical post-processing. In classical state-vector simulation with full-state reconstruction, the runtime is governed by a trade-off between reduced subcircuit size and the overheads of exponentially many subcircuits and full-state reconstruction. For equal partitioning, we derive threshold conditions on the number of cuts below which cutting reduces the wall-clock time. State-vector experiments validate the predicted speedup boundary up to 24 qubits, and a runtime breakdown up to 30 qubits identifies crossovers at
q≈18 and
q≈22 where merging overtakes first preprocessing and then subcircuit simulation. As a practical guideline, we show that under a 10-minute wall-clock budget, two-way cutting extends the maximum feasible qubit count by 4 to 6 qubits relative to simulation without cutting.
The rapid growth of computer vision and increasingly complex image recognition tasks has exposed fundamental computational limitations of classical machine learning models, motivating the exploration of quantum computing as an emerging new paradigm. This paper presents a comprehensive benchmarking study of classical and quantum machine learning models for image recognition on the MNIST handwritten digit dataset, evaluating both traditional models, a Classical Support Vector Machine (CSVM) and a Quantum Support Vector Machine (QSVM), and deep neural network models, a Classical Convolutional Neural Network (CCNN) and a Quantum Convolutional Neural Network (QCNN), across four performance dimensions: classification accuracy, computational runtime, parameter count, and memory requirements. Experiments are conducted as functions of both feature dimensionality and sample size, and across CPU and GPU execution environments, providing a controlled, multidimensional comparison to address gaps in prior work. For the SVM-based models, QSVM consistently outperforms CSVM in accuracy, reaching
∼ 0.90 versus
∼ 0.85 at 1,000 samples, with a higher computational cost. A feature count of 10 qubits and a sample size in the range of 200 -- 500 emerge as practical operating points that balance accuracy and runtime. For the neural network models, CCNN and QCNN achieve comparable classification accuracy, both exceeding 0.96 at 64 features and 60,000 samples, yet QCNN offers superior parameter and memory efficiency at higher feature counts, while incurring higher runtime. Across both model families, quantum models consistently outperform classical models by greater margins in accuracy as feature dimensionality or sample size increases.
Reporting the Fisher geometry of a trained variational quantum model is routine; quoting the shot budget that would establish it is not. Certifying an empirical Fisher matrix to relative Frobenius error
ε under coordinate-wise parameter shift costs
Θ(Bp2V/(ε2G)) circuit executions, where
V is the measured readout variance and
G the measured squared gradient norm, with uniform allocation optimal in that class. One constant reproduces the cost of two circuit families whose exponents differ by a full power of
p. The exponent is an identity in how
nV and
nG scale with the register, holding family by family to
0.001 across 624 matrix-product-state cells once the finite-
p prefactor is removed. The cubic cost is therefore a finite-size window, set by whether the readout light cone grows with the register. A product family to 256 qubits gives
1.966 (95% CI
1.934--
1.997); a brickwork entangler falls from
2.853 below ten qubits to
1.715 beyond sixty-four; a blocked entangler gives
1.984 at a fixed cone width against
3.034 at a proportional one. Fixed device connectivity fixes the cone, so a cubic budget from a small simulation overestimates a large machine, on top of hardware multipliers
2.07× (
1.41--
3.02),
2.38× and
1.91× on ibm_marrakesh, ibm_fez and ibm_kingston. Cost-optimal readout weights cut the measured shot budget by
2.67× (
1.33--
4.00) on hardware, flat from four to twelve qubits. A discrepancy model fitted on cheap circuits transfers its mean inside the calibration grid and, at six larger sizes named before the data, does not: nominal 90% intervals cover 36%, and split conformal is the only rung that stays near nominal.
Today's Quantum Computers (QCs) face significant engineering challenges that limit their size and fidelity. To execute realistic applications, researchers must develop ways to move past these constraints. CutQC enables small QCs to run larger quantum circuits by cutting circuits into smaller subcircuits and then utilizing classical resources and Kronecker products to reproduce the desired output. In this work, we enhance the reconstruction process. In particular, we develop a distributed PyTorch implementation of CutQC's classical post-processing step, which can run across multiple nodes on either GPU or CPU devices. Our results show that our PyTorch implementation executes circuits of up to 35 qubits on 10- and 15-qubit QCs with efficient reconstruction. We use single-node computation as a baseline and then compare the results to the runtimes of our PyTorch workflows. By utilizing parallelism and specified devices, the reconstruction step of hybrid quantum-classical circuit execution can be improved. The original CutQC paper demonstrated a 60X to 8600X runtime speedup over classical simulation, and our work increases this speedup. These results show that large quantum circuits can be run on smaller QCs by leveraging circuit cutting with data parallelism techniques to maintain a reasonable runtime. This allows researchers to maintain the high fidelity of smaller machines while executing larger circuits.
Quantum kernel methods are the most concrete near-term proposal for quantum machine learning on classical data, yet published benchmarks disagree on whether they outperform tuned classical kernels, and the disagreement cannot be resolved from the literature because the comparisons are not protocol-controlled. We argue that the obstacle is one of measurement, not theory: the usual setup fixes the quantum feature map's data scaling while grid-searching the classical kernel's hyperparameters, recomputes nothing per fold, and reports cost in simulator time. We therefore treat the comparison itself as the object of study and impose three controls. The bandwidth is tuned on the same grid budget as the classical
γ, the kernel is recomputed inside every fold from fold-fitted preprocessing, and cost is counted in hardware circuit evaluations. On top of these we run a factorial design of 324 cells in which dimensionality, training size, class separability and data structure are explicit covariates, together with four standard datasets, a shots-versus-gate-noise ablation, and a cost sweep. The dominant effect we measure is not quantum: tuning the feature map's bandwidth is worth 0.120 accuracy on synthetic data and 0.339 on real data, exceeding every quantum-classical difference we observe. Tuned quantum kernels are then statistically indistinguishable from a tuned RBF on all four real datasets. The advantage on synthetic data is confined almost entirely to data generated by the feature map itself, and it shrinks with dimensionality (-0.0045 per qubit) while growing with sample size. Finite sampling costs an order of magnitude more accuracy than gate and readout noise, and a fidelity kernel needs
5×107 circuits at n = 10,000. Protocol variance is thus sufficient to explain the field's contradictory conclusions.