What Does the Effective Dimension of a Hybrid Quantum Neural Network Actually Measure?
Organizations: Qatar University, Doha, Qatar
Abstract
Hybrid quantum neural networks (QNNs) pass features from a classical backbone into a variational quantum circuit, and many studies log a scalar called the effective dimension as a sign of capacity or model quality. The effective dimension is usually defined as the participation ratio (PR) of the quantum Fisher information matrix (QFIM), but studies in practice log a PR of cheap parameter sensitivities instead, and the meaning of that scalar is rarely tested. We show that the logged scalar equals the PR of Pauli-Z output sensitivities, that each sensitivity is at most n times its QFIM diagonal entry on n qubits, and that a PR of the QFIM diagonal overstates the full-matrix PR by an exact factor set by the off-diagonal mass. We then test the scalar on CIFAR-10 and EuroSAT with four frozen backbones, one of them a random projection, over 4 to 16 qubits, 1 to 4 layers, three entangling patterns, and three seeds. At a fixed circuit, CIFAR-10 accuracy spans 23.14% to 86.44% while the scalar stays between 6.63 and 9.13, and its correlation with accuracy changes sign across seeds in 7 of the 14 configurations. Depth raises the share of active parameters from 0.39 to 0.91, near a ceiling we derive, and initial gradient variance falls by a factor of 34.5 from 4 to 16 qubits. Removing entanglement raises accuracy by 5.3 to 8.2 points at the same parameter count. The effective dimension therefore describes circuit geometry rather than model quality, so capacity, trainability, and accuracy each need their own diagnostic.
Figures & tables
| Symbol | Meaning |
|---|---|
| , | Pure-state QFIM of Section 3.3 and its diagonal part |
| Pauli- sensitivity vector of Eq. ( 7 ) | |
| , | Logged scalar and number of entries of above |
| , | Reference scalars and |
| Off-diagonal mass | |
| Unbiased variance of the circuit-parameter gradients of the loss at initialization |
| Sub-grid | Configurations | Logged quantities | Cells |
|---|---|---|---|
| A | ; four backbones at , and EffNet at ; ring | , , , gradient variance, accuracy | 42 |
| B | EffNet; ; ; ring; untrained | Gradient variance at initialization | 168 |
| C | EffNet ring at , ; three other backbones, the linear and product patterns, and at | Accuracy and train–test gap | 126 |
| D | EffNet at ; first ring CNOT gates, at and at | Accuracy | 36 |
| Statistic | CIFAR-10 | EuroSAT | All |
| Configurations with positive mean | 7/7 | 6/7 | 13/14 |
| Sign-stable configurations | 3/7 | 4/7 | 7/14 |
| Configurations whose range crosses zero | 4/7 | 3/7 | 7/14 |
| Sign-stable configurations with | 2/2 | 2/2 | 4/4 |
| Mean of configuration means | |||
| Lowest and highest single-run | and | ||
| Dataset | Final acc. (%) | Initial | Mean | Ceiling | |||
|---|---|---|---|---|---|---|---|
| CIFAR-10 | 1 | 24 | 72.40 3.15 | 5.76 0.30 | 9.3 0.1 | 0.39 | 0.67 |
| 2 | 48 | 74.25 1.05 | 9.13 0.68 | 38.2 0.5 | 0.80 | 0.83 | |
| 3 | 72 | 74.79 0.72 | 11.55 0.30 | 63.2 0.4 | 0.88 | 0.89 | |
| 4 | 96 | 75.70 0.55 | 15.23 2.28 | 87.2 0.2 | 0.91 | 0.92 | |
| EuroSAT | 1 | 24 | 74.12 0.95 | 6.03 0.71 | 9.0 0.1 | 0.38 | 0.67 |
| 2 | 48 | 77.71 1.20 | 9.28 0.23 | 38.9 0.2 | 0.81 | 0.83 |
| CIFAR-10 | EuroSAT | ||
|---|---|---|---|
| 4 | 24 | 42.78 11.14 | 25.57 19.23 |
| 6 | 36 | 46.04 15.92 | 62.11 70.45 |
| 8 | 48 | 17.51 13.60 | 42.92 22.30 |
| 10 | 60 | 18.70 9.82 | 30.96 7.16 |
| 12 | 72 | 2.95 1.41 | 16.06 5.95 |
| 14 | 84 | 2.97 1.66 | 7.26 2.78 |
| Statistic | CIFAR-10 | EuroSAT |
|---|---|---|
| Variance ratio, over | 34.5 | 9.0 |
| Exponential slope of on (decades per qubit) | 0.019 | 0.025 |
| of the exponential fit | 0.92 | 0.72 |
| Power-law exponent of on | 0.54 | 0.67 |
| of the power-law fit | 0.83 | 0.54 |
| Dataset | Backbone | Initial | Peak | Final acc. (%) | Best acc. (%) | |
|---|---|---|---|---|---|---|
| CIFAR-10 | RandProj | 1280 | 8.46 0.20 | 18.72 3.03 | 23.14 0.97 | 24.48 0.82 |
| EfficientNet-B0 | 1280 | 9.13 0.68 | 10.09 2.19 | 74.25 1.05 | 75.49 1.16 | |
| ResNet-50 | 2048 | 6.63 1.17 | 6.21 0.42 | 82.22 0.43 | 85.10 1.31 | |
| Swin-T | 768 | 9.08 1.29 | 9.78 2.15 | 86.44 0.61 | 86.88 0.73 | |
| EuroSAT | RandProj | 1280 | 8.03 1.08 | 10.80 3.13 | 34.69 4.86 | 36.35 4.89 |
| EfficientNet-B0 | 1280 | 9.28 0.23 | 9.79 1.07 | 77.71 1.20 | 78.94 1.29 |
| Width | CIFAR-10 | EuroSAT | |
|---|---|---|---|
| 2 | 77.26 0.83 | 82.54 0.47 | |
| 4 | 76.56 0.40 | 81.77 0.92 | |
| 6 | 75.45 1.34 | 80.25 1.19 | |
| 3 | 72.99 0.47 | 83.53 0.38 | |
| 6 | 72.91 0.10 | 81.97 0.72 | |
| 9 | 72.13 0.53 | 81.02 0.73 |
| Question | Quantity | Evidence in this paper |
|---|---|---|
| How many parameters move the outputs? | Active count against | within 0.04 of for (Table 4 ) |
| How evenly is the sensitivity shared? | , or when is computed | Initial between 0.16 and 0.25 (Table 4 ) |
| Can a diagonal scalar stand in for ? | and Eq. ( 10 ) | Overstates by (Proposition 3 ) |
| Is the initialization trainable as grows? | Gradient variance against | Falls by a factor of from to on CIFAR-10 (Table 6 ) |
| Which backbone should be used? | Accuracy or a feature probe | Initial spans to while accuracy spans points (Table 7 ) |
| Does the scalar forecast accuracy? | Accuracy curve | Sign unstable in 7 of the 14 configurations (Table 3 ) |
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
| Dataset | Backbone | Initial | Peak | Mean | Late | Stab. epoch | [min, max] | |
|---|---|---|---|---|---|---|---|---|
| CIFAR-10 | RandProj | 2 | 8.46 0.20 | 18.72 3.03 | 11.05 2.04 | 11.42 2.69 | 117 | [ , ] |
| EfficientNet-B0 | 2 | 9.13 0.68 | 10.09 2.19 | 6.97 1.46 | 8.94 2.55 | 43 | [ , ] | |
| ResNet-50 | 2 | 6.63 1.17 | 6.21 0.42 | 4.49 0.97 | 5.34 0.77 | 45 | [ , ] | |
| Swin-T | 2 | 9.08 1.29 | 9.78 2.15 | 6.77 1.22 | 7.19 2.64 | 25 | [ , ] | |
| EfficientNet-B0 | 1 | 5.76 0.30 | 2.17 0.27 | 1.63 0.26 | 1.77 0.50 | 20 | [ , ] | |
| EfficientNet-B0 | 3 | 11.55 0.30 | 16.30 1.67 | 10.64 1.94 | 12.96 0.48 | 32 | [ , ] |
| Dataset | Comparison | Gap (points) | ( -test) | (Wilcoxon) | |
|---|---|---|---|---|---|
| CIFAR-10 | EfficientNet-B0 RandProj | 58.47 | 0.0003 | 0.250 | |
| ResNet-50 RandProj | 77.00 | 0.0002 | 0.250 | ||
| Swin-T RandProj | 74.95 | 0.0002 | 0.250 | ||
| ResNet-50 EfficientNet-B0 | 17.32 | 0.0033 | 0.250 | ||
| Swin-T EfficientNet-B0 | 13.92 | 0.0051 | 0.250 | ||
| Swin-T ResNet-50 | 10.11 | 0.0096 | 0.250 |
Quantum Neural Networks
Quantum neural networks use quantum circuits and measurements to learn representations, make predictions, and model complex data.
See all 148 papers