Q-Capsule: A Localized Capsule-Based Quantum Neural Architecture for Barren Plateau Mitigation

1 Introduction

• Q-Capsule partitions the register into fixed-width capsules with independent parameters and local readouts, giving each capsule a gradient bound governed by csize rather than total width N.This restricts the dominant support of each local objective as qubits are added.Source: §1 · pp. 2–3
• Sparse periodic boundary interactions let neighboring capsules share information while limiting global entanglement, retaining most measured inter-capsule mutual information with substantially fewer coupling gates.Coupling occurs at selected layers rather than every layer.Source: §1 · p. 3
• QFIM-guided growth increases each capsule's depth independently when its expressibility score begins to saturate, allocating capacity according to local optimization geometry rather than a fixed-depth schedule.Trainable data re-uploading and structured initialization are also combined with adaptive growth.Source: §1 · p. 3
• The evaluation examines gradient scaling, optimization landscapes, parameter efficiency, component ablations, noise robustness, measurement cost, and classification performance.The paper also reproduces gradient statistics in PennyLane, Qiskit, and Cirq.Source: §1 · p. 3
• Across binary and four-class digit classification, the adaptive policy remains competitive across task complexity without requiring circuit depth to be selected in advance.Gradient statistics are reported as consistent across PennyLane, Qiskit, and Cirq.Source: §1 · p. 3

2.1 Quantum Neural Networks and Structured Architectures

• Quantum neural network work explores data re-uploading for capacity and structured alternatives such as quantum convolutional networks, while fixed pooling schedules may limit adaptation to input complexity.Prior studies also characterize capacity using the Quantum Fisher Information Matrix and effective dimension.Source: §2.1 · p. 4

2.2 Barren Plateaus in Variational Quantum Circuits

• Barren-plateau mitigation includes structured circuits and optimization schedules, but circuit organization and capacity allocation are generally chosen before optimization or follow predetermined schedules.Q-Capsule's capsule-based measurement framework and alternating entangling layers target the balance between local trainability and information exchange.Source: §2.2 · pp. 4–5

3.1 Capsule Design

• The N-qubit register is divided into disjoint capsules of width csize, each with an independent depth and local observable; when uncoupled or before coupling reaches its readout, gradient support remains capsule-sized.With coupling every k layers, the observable's backward light cone expands by one capsule on either side per coupling round; the paper uses local capsule-wise measurements for classification.Source: §3.1 · p. 6

3.2 Adaptive growth

• Each capsule's depth grows independently when its QFIM-derived expressibility score approaches saturation, rather than following a uniform depth schedule.The score uses the QFIM spectrum, and a nearly constant score over a window signals saturation; an added layer uses near-identity initialization.Source: §3.2 · pp. 6–7

3.3 Sparse Inter-Capsule Coupling and the Entanglement Budget

• A CNOT links neighboring capsules every k layers in an alternating even–odd pattern, with k = 1 coupling each layer and k →∞ leaving capsules isolated.The coupling interval is treated as an entanglement regulator; the paper motivates k = O(N) to keep coupling weak as system size increases.Source: §3.3 · p. 7

3.4 Layer Structure and Trainable Re-Uploading

• Each layer encodes the input and applies trainable rotations; nearest-neighbor CNOTs entangle qubits within each capsule, while the trainable encoding scale adapts feature strength across re-uploading layers.Repeated input encoding between trainable SU(2) rotations lets each capsule represent more complex input functions as the number of re-uploading layers increases.Source: §3.4 · pp. 7–8

4.1 Experimental setup

• The MNIST experiments use two-class and four-class tasks, encoding the first N principal components into qubit rotations after min-max normalization to [0, π].The default configuration uses N = 12, capsule size csize = 3, four capsules, and initial depth 2; the classification pipeline concatenates capsule features for a classical softmax output layer.Source: §4.1 · p. 8

4.2 Baseline

• The five baselines are Layer, qMPS, ResQNet, Global, and Glob96, spanning layerwise growth, structured circuits, and globally entangling circuits.The comparison assesses gradient-variance scaling, fixed-width gradient distributions, cost-landscape geometry, and parameter efficiency.Source: §4.2 · p. 8

4.3 Gradient Variance Scaling with Qubit Count and Circuit Depth

• Q-Capsule maintains gradient variance on the order of 10^-1 across register widths, with a fitted slope of +0.01 per qubit; globally coupled baselines have slopes near -0.69 per qubit.At N = 20, Q-Capsule’s gradient variance is approximately five orders of magnitude above the globally coupled baselines; Layer stabilizes near 10^-3 at larger widths.Source: §4.3 · pp. 8–9
• Across depths L = 1, 5, 10, 15, 25, Q-Capsule maintains gradient variance of approximately 10^-1, while globally entangling baselines decline further at larger widths.The paper attributes Q-Capsule’s depth stability to sparse inter-capsule coupling that constrains global entanglement while allowing additional layers within capsules.Source: §4.3 · p. 9

4.4 Gradient Distributions at Fixed Width

• At N = 16 and depth 25, Q-Capsule gradients have empirical variance 8.3 × 10^-2, while Global, Glob96, qMPS, and ResQNet have variances between 3 × 10^-6 and 7 × 10^-6.Q-Capsule gradients span [−0.5, 0.5], with individual components above magnitude 0.7; the baseline gradient mass is concentrated near zero.Source: §4.4 · p. 10, Figure 4 · p. 11
• With 100 shots per gradient estimate, the roughly four-order-of-magnitude variance gap implies a comparable difference in measurement cost at this width.The paper relates the shot requirement for resolving a descent direction inversely to gradient variance.Source: §4.4 · p. 10
• Figure 4 reports empirical gradient distributions over random initializations at N = 16 and maximum depth 25, annotating each model’s variance and mean.The distributions are based on gradient estimates evaluated using 100 shots.Source: §4.4 · pp. 9–10, Figure 4 · p. 11

4.5 Cost-Landscape Geometry

• Q-Capsule’s two-parameter cost-landscape slice spans a total loss range of approximately 0.75, with structured variation and well-separated minima.The slice varies θ1 and θ2 while holding all remaining parameters fixed.Source: §4.5 · p. 12
• The Global baseline’s total loss variation is approximately 2.6 × 10^-4, while qMPS, Glob96, and ResQNet show similarly limited variation.The globally coupled landscapes are described as nearly flat at the evaluated width and depth.Source: §4.5 · p. 12
• The approximately three-order-of-magnitude difference in landscape variation is consistent with the gradient-variance results.The paper concludes that Q-Capsule preserves meaningful local cost variations while globally coupled architectures operate in effectively flat regions.Source: §4.5 · p. 12

4.6 Parameter Efficiency

• At N = 12, Q-Capsule achieves η = 1.06 × 10^-5, more than two orders of magnitude above the next-best baselines.The table reports η = Var[∂θL]/P and the advantage relative to each baseline.Source: §4.6 · p. 12, Table 1 · p. 12
• Q-Capsule’s advantage is driven primarily by gradient variance of 5.08 × 10^-2, versus O(10^-4)–O(10^-5) for the baselines, despite 4,800 trainable parameters at N = 12.The paper states that the ranking cannot be attributed to a smaller parameter budget.Source: §4.6 · p. 12
• The adaptive growth mechanism introduces additional capacity only after the existing capsule structure approaches saturation.Table 2 describes the influence of capsule size on capsule count, gradient variance, and classification accuracy at N = 12.Source: §4.6 · p. 13, Table 2 · p. 13

4.7.1 Capsule Partitioning and Capsule Size

• At N = 18, global gradient variance falls to 5.1 × 10−7, while capsule-partitioned variance stays between O(10−3) and O(10−2) over N = 4–18.The separation exceeds three orders of magnitude at N = 18.Source: §4.7.1 · p. 13
• At csize = 3, accuracy peaks at 0.972; gradient variance falls from 1.45 × 10−2 at csize = 2 to 2.37 × 10−4 at csize = 6.Smaller capsules preserve stronger gradients but provide less within-capsule expressivity; larger capsules provide more capacity but weaker gradients.Source: §4.7.1 · p. 13

4.7.2 Information Shared by Sparse Coupling

• At k = 8, two boundary CNOTs retain 1.92 bits, or 84% of the k = 1 setting’s 2.29 bits at 22% of its coupling cost.The highest measured mutual information is 3.06 bits at k = 2 with five CNOTs; the isolated setting has 0.00 bits.Source: §4.7.2 · p. 14

4.7.3 Component Knock-Out: Re-Uploading, Readout Locality, Initialization

• Removing data re-uploading reduces accuracy from 0.972 to 0.509 despite increased gradient variance, while removing structured initialization reduces accuracy to 0.639 with nearly unchanged variance.The results indicate that gradient magnitude alone does not establish whether descent directions support useful representations.Source: §4.7.3 · pp. 14–15
• Replacing local readout with the global observable reduces gradient variance from 2.42 × 10−3 to 9.77 × 10−5, while accuracy decreases from 0.972 to 0.963.The reported accuracy change is small at the experiment’s moderate register width.Source: §4.7.3 · p. 15

4.7.4 Adaptive Growth of Representational Capacity

• Mean S-score rises from 2.7 at d = 1 to 13.9 at d = 6 and approximately 31 at d = 10 as adaptive capsule depth increases.Each depth increment is followed by a sustained S-score rise; depth–S-score correlations range from 0.85 to 0.98.Source: §4.7.4 · pp. 15–16, Figure 9 · p. 15

4.7.5 Trainability Under Depolarizing Noise and Measurement Cost

• At p = 0.1, mean gradient magnitude is 5.1 × 10−2 for Q-Capsule and 1.5 × 10−2 for the global block; Q-Capsule gradients decay more slowly across tested noise levels.As register width grows, Q-Capsule’s measurement requirement remains nearly constant, while the global block’s required shots grow by several orders of magnitude.Source: §4.7.5 · p. 16, Figure 11 · p. 16

4.8 Classification Performance

• The fixed-deep policy reaches 0.944 binary and 0.963 four-class accuracy using 49 two-qubit gates and 192 parameters, and is slowest to train.Its accuracy is reliable on both tasks, but its resource cost is much higher than the adaptive policy’s.Source: §4.8 · p. 17
• At d = 1, the shallow policy reaches 0.491 on binary classification and 0.935 on four-class classification, so its chance-level result is specific to the binary task.Both tasks use an N = 6 register divided into two capsules of size csize = 3.Source: §4.8 · p. 17
• 0.981 binary accuracy and 0.977 four-class accuracy are reached by the adaptive policy using 13 two-qubit gates, 73% fewer than the fixed-deep policy.The adaptive policy uses 48 parameters; the accuracy margins over the deep policy are small and should be read as a tie.Source: §4.8 · p. 17, Figure 12 · p. 18

4.9 Cross-Framework Implementation Benchmark

• Across tested widths, gradient variance remains in O(10−2) across PennyLane, Qiskit, and Cirq; at N = 6, PennyLane and Qiskit yield 1.8918 × 10−2 versus Cirq’s 1.8909 × 10−2.PennyLane and Qiskit use parameter shift; Cirq uses central finite differences with ε = 10−3.Source: §4.9 · p. 18
• At N = 18, gradient-evaluation times are approximately 1.36 × 10^3 s for PennyLane, 7.73 s for Qiskit, and 1.50 s for Cirq, but the evaluated gradient components differ.PennyLane computes the full gradient vector, while Qiskit and Cirq evaluate a single component, so these times are not a direct simulator-speed comparison.Source: §4.9 · p. 18

5 Conclusion

• At N = 16 and depth 25, Q-Capsule has gradient variance 8.3 × 10−2, compared with approximately 10−6 for globally read baselines.The conclusion attributes stable gradient variance to the capsule-based architecture and identifies capsule locality, sparse coupling, re-uploading, and structured initialization as important in ablations.Source: §5 · p. 19
• The adaptive policy reaches 0.981 binary and 0.977 four-class accuracy while using about 73% fewer two-qubit gates than the fixed-deep model on the multiclass task.The conclusion also reports structured optimization landscapes, improved depolarizing-noise robustness, and lower measurement requirements.Source: §5 · p. 19
• Periodic QFIM evaluations, data re-uploading, and adaptive circuit reconstruction add training overhead, while the fixed final circuit leaves inference unaffected.Future work is directed toward reducing overhead, scaling to more complex datasets and larger systems, and validating on real noisy processors.Source: §5 · p. 19

• The methods do not involve human subjects, personal data, or sensitive datasets.Source: Ethics and Privacy Statement · p. 19
• The proposed techniques target improved training stability, convergence, and scalability in variational quantum machine learning models.Source: Ethics and Privacy Statement · p. 19
• The work presents general-purpose optimization methods, with no application-specific privacy, fairness, or safety risks beyond those inherent to quantum machine learning research.The authors note that improved quantum learning algorithms may accelerate future scientific and engineering applications.Source: Ethics and Privacy Statement · p. 19