Could you test surrogate robustness by intentionally corrupting training data with geometric noise before fitting?

Your question is about a technique in machine learning robustness testing, but it's asking whether a classical data-corruption strategy would work in the context of quantum machine learning surrogate models. The archive materials address quantum noise and robustness, so let me unpack what we know.

Yes, intentionally corrupting training data with geometric noise before fitting is a recognized robustness-testing approach — though the papers don't explicitly frame it that way. What they do show is the inverse: they *add* noise to test whether models remain robust.

For quantum machine learning surrogates specifically, the closest analogue is adversarial training. VeriQR demonstrates this: the tool "detect[s] adversarial examples and utilize[s] them for further analysis and to enhance the local robustness through adversarial training" [AG-2024.07-2163]. This means you deliberately generate problematic inputs (or in your case, corrupted training data), then retrain to resist them — a form of regularization through intentional perturbation.

The broader context: noise corruption during training has been studied in equivariant quantum neural networks. Research on "Symmetry breaking in geometric quantum machine learning" [AG-2024.01-1688] shows that Pauli-channel noise, when present during training, causes symmetry breaking that "grows linearly in the number of layers and noise strength." This suggests that geometric noise in training data has measurable, predictable effects on model behavior.

For practical robustness certification, you'd want to combine this with formal verification. VeriQR supports both local and global robustness checks [AG-2024.07-2163], so you could test whether a surrogate trained on corrupted data actually generalizes under the same noise distribution.

However: the archive focuses on quantum hardware noise channels and adversarial robustness, not on general surrogate model validation — so these papers don't directly compare geometric noise corruption strategies for non-quantum surrogates.