Hybrid quantum neural networks (QNNs) pair a classical feature extractor with a variational quantum circuit, and the effective dimension of the circuit's quantum Fisher information matrix (QFIM) is widely used to assess such models. This scalar is often interpreted as an indicator of trainability or accuracy, although these readings are rarely tested under controlled conditions. We therefore investigate what it actually measures. We define the effective dimension as the participation ratio of the QFIM diagonal, which lies between one and the parameter count, and we evaluate it across two image datasets, four frozen backbones including a random-projection control, four circuit depths on eight qubits, and three seeds. During training, the effective dimension remains well below the parameter count, and its correlation with test accuracy ranges from -0.72 to +0.87, changing sign across seeds in half of the configurations. Increasing depth activates a larger share of the circuit directions, from 0.38 to 0.91 of the parameter count, while the initialization gradient variance remains between 1e-5 and 1e-4, so no barren plateau (BP) appears at eight qubits. At a fixed circuit, the four backbones share an initial effective dimension of 6 to 9, even though CIFAR-10 accuracy ranges from 23.14% to 86.44%. These results show that the effective dimension characterizes circuit geometry rather than model quality. We conclude with a practical protocol that assigns each diagnostic to the question it can answer.
AI Summary
The paper asks whether a quantum-Fisher-information diagnostic can assess accuracy or trainability in hybrid image classifiers. It studies the participation ratio of the QFIM diagonal, $d_{\mathrm{eff}}=(\sum_k F_{kk})^2/\sum_k F_{kk}^2$, rather than a spectral or global Fisher effective dimension. For nonzero, nonnegative diagonals, it establishes $1\le d_{\mathrm{eff}}\le r\le p$, where $r$ counts positive diagonal entries, and scale invariance. Separately, $\operatorname{rank}(F)\le r$; the diagonal statistic does not determine the rank.
The experiments comprise 42 noise-free statevector training runs: two image datasets, four frozen feature maps including a dimension-matched random-projection control, three seeds, and an eight-qubit circuit with depths one through four. At fixed depth, initial $d_{\mathrm{eff}}$ clusters around 6–9 while CIFAR-10 accuracy spans 23.14%–86.44%. Within-run correlations between $d_{\mathrm{eff}}$ and test accuracy range from −0.72 to +0.87, changing sign across seeds in seven of fourteen configurations. Across the depth sweep, the thresholded active-coordinate fraction rises from approximately 0.38 to 0.91, while initialization gradient variance remains roughly $10^{-5}$–$10^{-4}$, without a collapse over the tested depths. These measurements motivate separating circuit-sensitivity diagnostics from feature-quality and accuracy assessments.
Stated limitations include one small ansatz family, a fixed eight-angle bottleneck, three seeds, diagonal-only geometry, and no hardware noise or finite-shot estimation. No qubit-number scaling is tested, so the gradient results remain finite-size observations. Full-QFIM checks, classical feature probes, wider-bottleneck ablations, and correlations corrected for shared training trends are left for future work.
Error found: Initialization gradient variances at eight qubits and depths one through four support an observation of no gradient collapse in that sweep, not a conclusion excluding barren-plateau scaling with qubit count.
The diagonal QFIM participation ratio can equal the parameter count even for a rank-one QFIM, so its interpretation as a count of independent circuit-state directions is not justified by the proved bounds.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
Hybrid quantum neural networks (QNNs) pair a classical feature extractor with a variational quantum circuit, and the effective dimension of the circuit's quantum Fisher information matrix (QFIM) is widely used to assess such models. This scalar is often interpreted as an indicator of trainability or accuracy, although these readings are rarely tested under controlled conditions. We therefore investigate what it actually measures. We define the effective dimension as the participation ratio of the QFIM diagonal, which lies between one and the parameter count, and we evaluate it across two image datasets, four frozen backbones including a random-projection control, four circuit depths on eight qubits, and three seeds. During training, the effective dimension remains well below the parameter count, and its correlation with test accuracy ranges from -0.72 to +0.87, changing sign across seeds in half of the configurations. Increasing depth activates a larger share of the circuit directions, from 0.38 to 0.91 of the parameter count, while the initialization gradient variance remains between 1e-5 and 1e-4, so no barren plateau (BP) appears at eight qubits. At a fixed circuit, the four backbones share an initial effective dimension of 6 to 9, even though CIFAR-10 accuracy ranges from 23.14% to 86.44%. These results show that the effective dimension characterizes circuit geometry rather than model quality. We conclude with a practical protocol that assigns each diagnostic to the question it can answer.
The paper asks whether a quantum-Fisher-information diagnostic can assess accuracy or trainability in hybrid image classifiers. It studies the participation ratio of the QFIM diagonal, deff=(∑kFkk)2/∑kFkk2, rather than a spectral or global Fisher effective dimension. For nonzero, nonnegative diagonals, it establishes 1≤deff≤r≤p, where r counts positive diagonal entries, and scale invariance. Separately, rank(F)≤r; the diagonal statistic does not determine the rank.
The experiments comprise 42 noise-free statevector training runs: two image datasets, four frozen feature maps including a dimension-matched random-projection control, three seeds, and an eight-qubit circuit with depths one through four. At fixed depth, initial deff clusters around 6–9 while CIFAR-10 accuracy spans 23.14%–86.44%. Within-run correlations between deff and test accuracy range from −0.72 to +0.87, changing sign across seeds in seven of fourteen configurations. Across the depth sweep, the thresholded active-coordinate fraction rises from approximately 0.38 to 0.91, while initialization gradient variance remains roughly 10−5–10−4, without a collapse over the tested depths. These measurements motivate separating circuit-sensitivity diagnostics from feature-quality and accuracy assessments.
Stated limitations include one small ansatz family, a fixed eight-angle bottleneck, three seeds, diagonal-only geometry, and no hardware noise or finite-shot estimation. No qubit-number scaling is tested, so the gradient results remain finite-size observations. Full-QFIM checks, classical feature probes, wider-bottleneck ablations, and correlations corrected for shared training trends are left for future work.
Hybrid quantum neural networks are increasingly explored for classification, yet it remains unclear how their performance and quantum behavior scale with circuit depth and qubit count. We present a controlled scaling study of hybrid quantum-classical classifiers along two axes: (1) increasing the number of quantum layers L at fixed qubits Q, and (2) increasing the number of qubits Q at fixed depth L. Across multiple datasets, we evaluate predictive performance using Accuracy, PR-AUC, Precision, Recall, and F1, and track quantum-specific metrics (QCE, EEE, QGN) to characterize how quantum properties evolve under scaling. Our results summarize scaling trends, saturation regimes, and dataset-dependent sensitivity, and further analyze how quantum metrics relate to predictive performance. This study provides practical guidance for selecting (Q,L) in hybrid QNN classifiers and establishes a consistent evaluation protocol for scaling analysis.
The practical value of quantum machine learning must be derived from efficiency rather than scale, as current NISQ processors support only small, shallow circuits. This paper evaluated whether variational quantum circuits (VQCs) can reduce the trainable parameter count in an image classification scenario while preserving accuracy. Experimentation compared quantum classifier heads to parameter-matched classical heads on an identical set of frozen features drawn from the Fashion-MNIST dataset by a CNN. 9 configurations of a quantum circuit crossed 4, 6, and 8 qubits with 1, 2, and 4 entangling layers, each matched to a multilayer perceptron with an approximately equal parameter count. The linear probes served as a floor baseline for all 105 runs across 5 random seed values. 7 out of 9 quantum-based configurations met the mean test accuracy of no more than 1 percentage point below its matched classical head, enforced by the pre-specified criterion. The 2 failures were single-layer circuits, with each falling roughly 8 points below its matched classical head and 12 points below the circuit-free linear probe; in total, 6 configs with 2+ layers passed, with the smallest at 74 trainable parameters. The general trend noticed was that accuracy increased progressively with depth. The 4-layer circuits exceeded the linear probe at every qubit count and produced the highest accuracy at 6 and 8 qubits. The simulation ran noiselessly; VQCs preserved accuracy at the matched parameter budget when depth was given, showcasing depth rather than qubit count gating performance.
Introduction: Variational quantum algorithms can show unstable trainability and uneven generalisation that are not fully explained by circuit depth or raw parameter count alone. We study the data quantum Fisher information matrix (DQFIM) as a data-dependent effective-capacity diagnostic for controlled supervised unitary-learning tasks. Materials and methods: We used exact-state, classically simulated, 4-qubit, supervised unitary-learning benchmarks with two ansatz families: a hardware-efficient ansatz (HEA) for the baseline trainability study and a symmetry-preserving ansatz for symmetry-controlled comparisons. The analysis covered four linked experiments: E0 trainability phase boundaries, E1 symmetry-controlled data-regime comparisons, E2 trainability versus generalisation, and E3 pre-training prediction of optimisation success. For the main analyses we used the support-basis DQFIM, 30 random seeds per main configuration, grouped cross-validation in the predictive benchmark, and additional diagnostics for threshold sensitivity, phase indeterminacy, Hamming-sector generalisation, and test set size sensitivity. Results: In E0, the empirical trainability boundary increased from M c = 32 at L = 1 to M c = 192 at L = 8 , and the support-basis DQFIM-predicted boundary matched the empirical boundary on the resolved scanned grid under the main analysis setting. In E1, the sector-preserving and sector-broken conditions showed no resolved empirical boundary shift on the scanned grid, with only mild low-L asymmetry in the DQFIM-predicted boundary. In E2, trainability and generalisation separated clearly: some regimes generalised well, while others reached near-zero training loss but retained high test loss. The phase diagnostic supported relative phase indeterminacy as a mechanism for failure on in-sector superpositions after basis-state training. In E3, parameter count alone was a weak predictor of optimisation success, with receiver operating characteristic area under the curve (ROC AUC) = 0.648 ; the structural baseline was stronger, with ROC AUC = 0.958 ; and the DQFIM-enhanced model performed best, with ROC AUC = 0.989 . Conclusions: In these small, idealised, classically simulated matched-family tasks, the support-basis DQFIM provides a useful data-dependent pre-training diagnostic of effective capacity on the retained data support. It tracks resolved trainability boundaries and contributes predictive information beyond raw parameter count and structural metadata. The conclusions remain scoped to controlled small-system simulations; larger-qubit, noisy, hardware-executed, and less-matched settings require further validation before claiming practical scalability.
Hybrid quantum neural networks (HQNNs) integrate parameterized quantum circuits (PQCs) within classical networks, where the behavior of the underlying PQCs is often the primary focus of analysis. In this context, expressibility and trainability are widely used to characterize PQC's performance and are commonly assumed to exhibit a trade-off, where highly expressive circuits are more susceptible to barren plateaus. However, the validity of this relationship in HQNNs remains unclear. In this paper, we systematically analyze the expressibility--trainability relationship in HQNNs across varying circuit depths, qubit counts, entanglement topologies. We consider different training configurations, including pure PQCs, quantum-only training in hybrid setting, and full end-to-end training of hybrid models. Our results show that pure PQCs exhibit only a weak and regime-dependent trade-off, while hybrid architectures increasingly disrupt and can eliminate this relationship under full hybrid training. This indicates that classical components reshape the optimization landscape, decoupling trainability from PQC expressibility. We further propose a multi-objective neural architecture search (NAS) framework that jointly optimizes expressibility, trainability, and task performance over a combined classical--quantum design space, revealing different Pareto-optimal solutions under full end-to-end and quantum only training in hybrid setting. different trainability definitions. Our results suggest that hybridization is not just an implementation detail, but a defining factor in the performance of quantum machine learning models.
This study investigates classification performance and trainability of variational quantum classifiers on near-term noisy quantum hardware across circuit depth and qubit count regimes within the context of quantum computing and machine learning, an area of growing scientific importance given its implications for hybrid quantum-classical algorithm design and practical NISQ hardware benchmarking for near-term quantum advantage assessment. Using gradient-based parameter optimization of parameterized quantum circuits with hardware noise modeling via density matrix simulation and physical execution on IBM Eagle processors, we examine expressibility of parameterized circuits enabling nonlinear decision boundaries in feature Hilbert space, with trainability constrained by barren plateau gradients at large qubit counts in 96 circuit architectures (4 qubit counts x 4 depths x 6 ansatz families) evaluated across 5 classification datasets with 20 random initialization seeds each drawn from IBM Quantum Eagle (127-qubit) and Falcon (27-qubit) processors with noise mitigation via zero-noise extrapolation. Results indicate that 10-qubit, 4-layer VQC achieves 84.2% classification accuracy on the best benchmark, within 3.8 percentage points of classical SVM, but gradient variance decays exponentially with qubit count (variance halving per 2 added qubits), confirming barren plateau scaling (p < 0.001), with 84.2% accuracy, 3.8 pp gap to classical SVM baseline as the primary quantitative benchmark. Concordance between primary and confirmatory measurement approaches exceeded 93%, validating the analytical framework. These findings contribute empirically to quantum computing and machine learning and carry actionable implications for the design of programs and policies targeting hybrid quantum-classical algorithm design and practical NISQ hardware benchmarking for near-term quantum advantage assessment.
Hybrid Quantum-Classical Neural Networks (HQNNs) offer a practical approach for near-term quantum machine learning; however, their performance on real-world, high-dimensional, and imbalanced tabular datasets remains underexplored. In this work, we benchmark HQNNs against parameter-equivalent classical models on two data-driven classification tasks: the Santander Customer Transaction dataset (200 features, 9:1 class imbalance) and the UCI Adult dataset (105 features, 3:1 class imbalance). Across 48 independent experimental trials, we evaluate 24 hybrid quantum configurations alongside 24 classical baselines by varying qubit count (N ∈ {4,6}), circuit depth (L ∈ {3,4,5}), and entanglement topology (Strongly Entangling Layers vs. Random Layers). Results show that select hybrid configurations reduce training epochs by a factor of 2–4 while maintaining competitive AUC; for example, on Santander, peak AUC is reached at 15 epochs compared to 60 for classical models, with only a 0.0166 difference in final AUC. We further observe that structured entanglement yields more stable training, and while deeper (5-layer) configurations can exhibit training variance, successful runs still yield strong convergence efficiencies. These findings highlight HQNNs as an efficient and practical approach for data-driven imbalanced classification tasks.