Effective Dimension Governs Generalization in Quantum Kernel Vision Models

Abstract: Recent quantum vision models-quantum vision transformers and quantum convolutional networks-report two striking but unexplained empirical phenomena: (i) ansatze with more, or more uniformly distributed, entanglement generalize better, and (ii) injecting quantum noise can improve test accuracy rather than degrade it. These observations are currently treated as curiosities, discovered by grid search and explained, if at all, by hand. We show that both are manifestations of a single, measurable quantity: the \emph{effective dimension} $d_{\rm eff}$ of the (noise-shaped) quantum feature kernel. Working primarily with quantum-kernel vision models-a quantum feature map read out by a kernel classifier-we give a spectral account in which entanglement structure and quantum noise are two knobs that move $d_{\rm eff}$; in an overfitting regime, contracting $d_{\rm eff}$ acts as ridge-like regularization. We analyze the mechanism: an \emph{exact} decomposition of the depolarized kernel $K_p=(1-p)^2K+\tfrac{p(2-p)}{D}\mathbf{1}\mathbf{1}^\top$ with $d_{\rm eff}(K_p)\to1$, a contraction result (and its boundary) for amplitude damping, a kernel-machine capacity bound, and a capacity/alignment risk decomposition; the monotone contraction operative in our entangled experiments is verified empirically, not proven in general. Along the one-parameter depolarizing family the collapse is instead exact by construction; we use it only to confirm the kernel decomposition to machine precision and at up to $12$ qubits, not as evidence for $d_{\rm eff}$. Amplitude damping contracts $d_{\rm eff}$ and lifts test accuracy by up to $+13\%$ along an inverted-U sweet spot; the effect's sign flips between the over- and under-fitting regimes; noise injection matches an explicit spectral-filtering frontier. Our results organize two reported anecdotes into a single measurable principle for designing quantum-vision models.

• The paper establishes that the effective dimension, quantified by the participation ratio of kernel eigenvalues, unifies the roles of entanglement and noise in governing model generalization.
• The paper presents evidence that quantum noise acts as a ridge-like spectral regularizer, moderating overfitting by contracting the kernel spectrum and improving test accuracy.
• The paper shows that test accuracy consistently collapses onto a single curve as a function of effective dimension across varied circuit topologies and noise regimes, supporting a unified spectral generalization law.

Spectral Regularization and Generalization in Quantum Kernel Vision Models

Introduction

This paper establishes a precise spectral principle for understanding generalization in quantum vision models, specifically those employing quantum feature maps evaluated via kernel methods. Two persistent empirical observations in quantum vision architectures—(i) improved generalization with greater or more uniform circuit entanglement, and (ii) surprising increases in test accuracy upon injection of quantum noise—are unified under the governing role of the quantum feature kernel’s effective dimension. The effective dimension, quantified by the participation ratio of kernel eigenvalues, is shown to behave as a spectral regularizer in overfitting regimes, mediating the bias–variance tradeoff and connecting entanglement, noise, and generalization.

Quantum Feature Kernel and Effective Dimension

Quantum vision models encode classical data into quantum states via parameterized circuits, resulting in a feature map $\rho(x)$. The Hilbert–Schmidt kernel $k(x, x') = \operatorname{Tr}[\rho(x) \rho(x')]$ yields a Gram matrix with eigenvalues $\{\lambda_i\}$. The effective dimension, defined as

$$d_{\text{eff}} = \frac{(\operatorname{tr} K)^2}{\operatorname{tr}(K^2)},$$

summarizes the concentration of the kernel's spectrum and corresponds to the ridge effective dimension in classical kernel learning theory. This quantity is central to the capacity and generalization properties of kernel classifiers.

Entanglement and Noise as Spectral Controls

Entanglement topology and quantum noise are interpreted as complementary spectral "knobs":

• Entanglement structure: Acts as a prerequisite by providing high kernel–target alignment, enabling nontrivial correlation across features.
• Quantum noise: Contracts the kernel spectrum, acting as a ridge-like regularizer, and provably reduces the effective dimension, especially under global depolarizing channels.

This principle is visualized in an overview schematic connecting quantum feature encoding, kernel computation, spectrum analysis, and classifier performance, with entanglement and noise as upstream controls.

(Figure 1)

Figure 1: Overview schematic—entanglement and noise independently regulate the effective dimension, which governs test accuracy via spectral regularization.

Quantum noise channels (depolarizing and amplitude damping) are analytically shown to filter the spectrum toward rank-one structure, reducing $d_{\text{eff}}$ and enforcing monotonic contraction under suitable conditions.

Empirical Evidence: Collapse onto a Single Spectral Curve

The central empirical finding is that across distinct entangled circuit ansatze (chain, ring, all-to-all), test accuracy as a function of $d_{\text{eff}}$ collapses onto a single curve, regardless of whether $d_{\text{eff}}$ is manipulated via circuit topology or noise injection. This convergence is conditional on sufficient entanglement (kernel–target alignment), and is not observed for unentangled (product) maps, which exhibit low alignment and spectral factorization.

Figure 2: Noise reshapes the quantum feature geometry—quantum kernel-PCA embeddings visualize contraction of the feature spectrum with increasing noise.

Figure 3: Across a $4 \times 6$ grid (entangling topology $\times$ noise level), accuracy collapses onto a single function of $d_{\text{eff}}$.

Noise-Induced Regularization: Inverted-U and Sign Flip

Amplitude damping and depolarizing noise systematically contract $k(x, x') = \operatorname{Tr}[\rho(x) \rho(x')]$0, reduce training accuracy (i.e., memorization), and increase test accuracy—characteristic of ridge-like regularization in classical overfitting regimes. The regularization benefit exhibits an inverted-U dependence: excess noise eventually destroys signal-carrying eigen-directions, reducing generalization. The magnitude of the improvement grows with the severity of overfitting (fraction of corrupted labels), and the optimal noise rate is largely invariant to label corruption fraction.

This mechanistic effect is replicated on real IBM Heron hardware: device noise contracts the spectrum and increases test accuracy on overfitting tasks.

Figure 4: On IBM Heron hardware, intrinsic device noise contracts $k(x, x') = \operatorname{Tr}[\rho(x) \rho(x')]$1 and improves test accuracy—demonstrating noise-as-regularization in practice.

The predicted sign flip is observed: in overfitting regimes, spectral contraction via noise increases generalization, while in underfitting regimes, the identical operation diminishes test accuracy.

Figure 5: The $k(x, x') = \operatorname{Tr}[\rho(x) \rho(x')]$2–accuracy relation reverses between overfitting and underfitting regimes—noise-induced contraction benefits test accuracy only when the model overfits.

Robustness: Datasets, Architectures, and Training

The principle persists across multiple benchmarks (Digits, Fashion-MNIST, BloodMNIST), circuit depths, and widths, as well as in trained quantum vision transformers and quantum convolutional models. When kernel alignment is high, the effective dimension is the dominant predictor of generalization, with minimal dependence on the specifics of circuit topology or noise channel.

Figure 6: Representative images from Digits, Fashion-MNIST, and BloodMNIST—quantum feature maps constructed from PCA-reduced vision data.

Practical and Theoretical Implications

The spectral account transforms entanglement and quantum noise from ad-hoc hyperparameter heuristics into measurable, controllable design handles. Selection of the optimal noise level can be performed label-free based on training data alone, reducing computational search. On classical data with quantum kernels, noise acts on the spectrum directly and cannot be substituted by SVM regularization alone. The effective dimension is not a universal sufficient statistic; kernel–target alignment and spectral shape remain essential, especially in the empirical cross-ansatz collapse.

This framework opens avenues for principled quantum kernel model design, leveraging entanglement and noise as spectral controls and tying quantum architecture decisions to classical learning theory. Realizing spectral regularization natively on quantum hardware is especially relevant given unavoidable device noise.

Conclusion

This work organizes previously unexplained empirical behaviors in quantum vision kernel models into a single spectral generalization law governed by the effective dimension. Entanglement acts as an alignment precondition, and noise acts as a spectral regularizer. Accuracy, within suitably entangled regimes, collapses onto a function of effective dimension, with the sign of the effect set by the data/model regime. The findings provide a predictive, theoretically grounded principle for quantum vision system design and offer transferable diagnostics to classical kernel learning.