| name | coherence-law-noisy-equivariant-qnn |
| description | Coherence law for trainability in noisy equivariant quantum neural networks. Proves that readout-visible sector coherence determines gradient survival under decoherence, not just symmetry structure. |
| category | ai_collection |
| tags | ["quantum-neural-networks","equivariant-circuits","noisy-qml","trainability","coherence-law","barren-plateaus"] |
| trigger_words | ["coherence law trainability QNN","equivariant quantum neural network noise","noisy QNN gradient decay","sector coherence trainability","light-cone reduction quantum","open-system quantum trainability"] |
Coherence Law for Trainability in Noisy Equivariant QNNs
Background
Symmetry provides structure for quantum neural networks, but does not guarantee trainability once noise is present. The expressivity-trainability paradox shows that unstructured QML architectures suffer from quantum underfitting driven by barren plateaus. This skill provides a coherence law that determines whether gradients survive decoherence.
Core Methodology
1. Light-Cone Reduction
- Causality fixes where the gradient can live — confined to the backward light cone of the readout
- Light-cone reduction pins the noiseless gradient to the sector-restricted cone
- Lower bound independent of total qubit number
2. Coherence Contraction Law
- Coherence determines how fast the gradient decays through contraction of off-diagonal sector modes
- Readout-visible aligned coherence rate defined as Rayleigh quotient of noise generator along gradient-carrying mode
- Perturbative open-system analysis converts this rate into a leading-order training law
3. Training Law
- Finite-noise degradation follows a single accumulated variable built from noise depth and coherence contraction
- Coefficient of determination R² = 0.979 between predicted and observed gradient degradation
- Sector coherence outperforms every standard channel diagnostic
4. Key Insight
- For correlated-dephasing channels with large worst-case rate but near-zero aligned rate:
- The law predicts no gradient loss → none is observed
- Sector coherence is the quantity linking equivariant architecture, open-system dynamics, and noisy trainability
Implementation Guidelines
def coherence_aligned_rate(noise_generator, gradient_mode, readout_projector):
"""Calculate readout-visible aligned coherence rate as Rayleigh quotient"""
projected_noise = readout_projector @ noise_generator @ readout_projector
rate = (gradient_mode.T @ projected_noise @ gradient_mode) / (gradient_mode.T @ gradient_mode)
return rate
def predict_gradient_decay(coherence_rate, circuit_depth, noise_strength):
"""Leading-order training law for gradient degradation"""
accumulated = coherence_rate * circuit_depth * noise_strength
decay_factor = np.exp(-accumulated)
return decay_factor
Applications
- Noisy quantum neural network training
- Equivariant quantum circuit design
- Open-system quantum dynamics
- Quantum machine learning robustness
- Barren plateau mitigation strategies
Pitfalls
- Symmetry alone does NOT guarantee trainability under noise — coherence is the deciding factor
- The aligned coherence rate can be near-zero even when worst-case channel rate is large
- Light-cone confinement means gradients only exist in specific sectors — design readout accordingly
- This law applies to U(1)-equivariant brickwork circuits — verify applicability for other symmetries
References
- arXiv:2606.30688 — "A Coherence Law for Trainability in Noisy Equivariant Quantum Neural Networks"
- Ugail, Howard (2026)