jacobian-geometry-robustness-qnn

name: jacobian-geometry-robustness-qnn description: "JGRA framework for assessing robustness in NISQ noise-aware Quantum Neural Networks via Jacobian geometry. Captures model sensitivity to parameter perturbations induced by noise through entropy-matched noise calibration, noise-aware training, and noise-conditioned Jacobian extraction. Accepted at IEEE qCCL 2026. Activation: QNN robustness, NISQ noise, Jacobian geometry, quantum neural network, noise-aware training, robustness assessment, decoherence, parameter perturbation, geometric descriptor, noise calibration" metadata: arxiv_id: "2606.09964" published: "2026-06-08" authors: "Gianluca Scanu, Luca Barletta, Stefano Rini"

Context

NISQ-era quantum computation is fundamentally limited by noise and decoherence. Classical DNNs exhibit robustness through structural redundancy; analogous principles for QNNs are underdeveloped. JGRA (Jacobian Geometry Robustness Assessment) bridges this gap by analyzing QNN robustness through the geometry of the Jacobian matrix.

Core Methodology

1. Entropy-Matched Noise Calibration

• Characterize NISQ noise channels via entropy matching: match the entropy of simulated noise to hardware-measured noise profiles
• Calibrate noise strength to ensure realistic perturbation modeling
• Supports depolarizing, amplitude damping, and phase damping channels

2. Noise-Aware Training

• Inject calibrated noise during training phase (not just inference)
• Optimize loss function under noisy forward pass: L_noisy(θ) = E_noise[L(θ, noise)]
• Produces noise-resilient parameters that generalize to unseen noise conditions

3. Noise-Conditioned Jacobian Extraction

• Compute Jacobian J_ij = ∂output_i/∂θ_j under both clean and noisy conditions
• Extract geometric descriptors:
  • Jacobian singular values: measure sensitivity to parameter directions
  • Condition number κ(J): ratio of largest to smallest singular value → robustness indicator
  • Jacobian spectrum entropy: H(J) = -Σ σ_i log σ_i / Σ σ_j → information distribution across parameter directions
  • Clean-noisy Jacobian distance: ||J_clean - J_noisy||_F → perturbation magnitude

4. Predictive Robustness Descriptors

• Geometric descriptors from clean regime encode predictive information about noisy inference behavior
• High condition number → fragile model (small parameter changes cause large output changes)
• Low Jacobian spectrum entropy → concentrated sensitivity (few dominant directions)

Implementation Steps

1. Define QNN architecture (parameterized quantum circuit)
2. Characterize target hardware noise profile (T1, T2, gate fidelities)
3. Calibrate noise model via entropy matching
4. Train QNN with noise injection (noise-aware training)
5. Extract Jacobian at converged parameters under clean conditions
6. Extract Jacobian under calibrated noise conditions
7. Compute geometric descriptors (singular values, condition number, spectrum entropy)
8. Correlate descriptors with actual robustness on held-out noise conditions
9. Use descriptors as early-warning indicators for model fragility

Pitfalls

• Noise model accuracy: Entropy-matched calibration is necessary but not sufficient — verify noise model against actual hardware benchmarks
• Jacobian dimensionality: For large QNNs, full Jacobian extraction is O(n_params × n_outputs) — use randomized SVD or Hutchinson trace estimation for scalability
• Training-inference mismatch: Noise-aware training only protects against noise types seen during training — test on genuinely unseen noise channels
• Classical vs quantum redundancy: QNNs lack the overparameterization of classical DNNs — structural redundancy is limited by qubit count and circuit depth

Verification

• JGRA descriptors should correlate (r > 0.7) with actual performance degradation under unseen noise
• Condition number should increase monotonically with noise strength
• Noise-aware trained models should show lower clean-noisy Jacobian distance than noise-unaware models