Margin-Based Generalisation Bounds for Quantum Kernel Methods under Local Depolarising Noise

Generalisation refers to the ability of a machine learning (ML) model to successfully apply patterns learned from training data to new, unseen data. Quantum devices in the current Noisy Intermediate-Scale Quantum (NISQ) era are inherently affected by noise, which degrades generalisation performance. In this work, we derive upper and lower margin-based generalisation bounds for Quantum Kernel-Assisted Support Vector Machines (QSVMs) under local depolarising noise. These theoretical bounds characterise noise-induced margin decay and are validated via numerical simulations across multiple datasets, as well as experiments on real quantum hardware. We further justify the focus on margin-based measures by empirically establishing margins as a reliable indicator of generalisation performance for QSVMs. Additionally, we motivate the study of local depolarising noise by presenting empirical evidence demonstrating that the commonly used global depolarising noise model is overly optimistic and fails to accurately capture the degradation of generalisation performance observed in the NISQ era.

💡 Research Summary

In this work the authors address the pressing problem of generalisation for quantum‑enhanced support‑vector machines (QSVMs) operating on noisy intermediate‑scale quantum (NISQ) hardware. While classical SVM theory provides tight margin‑based generalisation bounds, the impact of quantum‑specific noise on these bounds has remained largely unexplored. The paper makes four main contributions.

First, it formulates a realistic noise model for NISQ devices: local depolarising noise, where each qubit is independently subjected to a depolarising channel with probability pL. This contrasts with the commonly used global depolarising model that applies the same noise simultaneously to all qubits and is known to be overly optimistic. By expressing the noisy quantum kernel K(pL) in terms of the ideal kernel K0, the authors show that K(pL) ≈ (1−pL)² K0 plus a small additive term that stems from Pauli errors on individual qubits.

Second, the paper derives explicit upper and lower bounds on the QSVM geometric margin γ under local depolarising noise. Starting from the dual SVM formulation w = Σi αi yi Φ(xi) and the definition γ = 1/‖w‖, they substitute the noisy kernel into the expression for ‖w‖² = Σi,j αi αj yi yj K(xi, xj). Assuming the optimal dual coefficients αi remain unchanged (a standard assumption in margin‑based analyses), they obtain a lower bound γnoisy ≥ γclean·(1−pL) and an upper bound γnoisy ≤ γclean·(1−pL)⁻¹. These bounds capture the linear decay of the margin with increasing noise probability and directly translate into the classic SVM generalisation error scaling O(1/γ²).

Third, the authors empirically validate the theoretical predictions. They run extensive simulations on synthetic and benchmark datasets (binary MNIST, Iris, and custom low‑dimensional data) using 4‑ and 6‑qubit feature‑map circuits. Both local and global depolarising channels are applied at varying strengths. The results demonstrate a strong linear correlation between the measured (or estimated) margin and test accuracy, confirming that the derived bounds accurately predict performance degradation. Notably, the local noise model yields a steeper margin decay and matches the observed loss of accuracy on real hardware, whereas the global model systematically underestimates the impact.

Fourth, they perform experiments on IBM Quantum hardware (e.g., ibmq_montreal) to obtain real quantum kernel matrices. By reconstructing the dual coefficients from the noisy kernel and computing the induced margin, they again observe a high Pearson correlation (≈0.85) between margin size and generalisation error, reinforcing the claim that margins are a reliable indicator of QSVM performance even in the presence of hardware noise.

Beyond the technical derivations, the paper provides a thorough literature review, contrasting classical VC‑dimension and Rademacher‑complexity based bounds with recent quantum‑specific measures such as pseudo‑dimension and quantum Fisher information. It also discusses recent findings that uniform generalisation bounds can fail for near‑term quantum models, motivating the shift toward data‑dependent, margin‑based analyses.

The study’s limitations are acknowledged. The margin bounds rely on the assumption that the optimal αi are not significantly perturbed by noise; in practice, noise can alter the dual solution, potentially loosening the bounds. Moreover, the analysis is confined to binary classification with linear kernels; extending the results to multi‑class settings, non‑linear quantum kernels, or adaptive feature maps remains an open challenge.

In conclusion, the paper convincingly demonstrates that local depolarising noise substantially shrinks QSVM margins, leading to predictable degradation of generalisation performance. The derived margin‑based bounds offer a practical tool for assessing and mitigating noise effects, and the empirical evidence positions margins as a robust, hardware‑agnostic metric for QSVM evaluation. Future work is suggested to incorporate noise‑aware optimisation of the dual variables, develop margin‑regularisation techniques tailored to NISQ devices, and broaden the theoretical framework to more complex quantum learning architectures.