Why did the QNN fail specifically on the lowest-amplitude signal—noise floor or training data distribution?

Your question touches on a real diagnostic challenge in quantum machine learning, but the archive doesn't directly address the specific failure mode you're asking about—low-amplitude signal detection in QNNs.

That said, the preprints offer relevant context. The most likely culprit is training data distribution, not a hard noise floor. Here's why:

QNNs struggle with low-amplitude signals for a subtle reason tied to how they learn. When a quantum circuit is asked to distinguish tiny signal variations, the cost function landscape becomes extremely flat—a problem called "barren plateaus" [AG-2024.02-1624]. This means gradients vanish, and the network can't learn which parameters to adjust, regardless of the signal strength itself. It's not that the quantum hardware can't *measure* the signal; it's that the learning algorithm can't find a path through parameter space.

Noise does complicate this further. Different noise types have wildly different effects: phase and bit-flip errors become manageable at extreme probabilities (the network adapts), but depolarizing noise consistently degrades performance [AG-2024.02-1629]. However, for low-amplitude signals *specifically*, the barrier is usually the training landscape geometry, not noise tolerance.

The solution isn't to lower the noise floor—it's to choose the right measurement observable and noise-aware training strategy [AG-2024.02-1624]. This suggests the QNN's failure was structural (how it was set up to learn) rather than a fundamental detection limit.

To answer whether it was noise or training data distribution in *your* case, you'd need to test: did retraining on synthetic high-amplitude data first help? If yes, it's a distribution problem. Did switching observables help? If yes, it's architecture. The preprints don't contain a direct case study of low-amplitude signal failure.