Why do simulations introduce correlation artifacts that adversarial training alone cannot detect or eliminate?

Your question touches on a real tension in machine learning—but it's asking about *simulation artifacts* and adversarial training in a way that doesn't align cleanly with the preprints provided. Let me explain what the archive does cover, and where the gap is.

What the archive *does* address: adversarial training's fundamental limits

The preprints show that adversarial training, while helpful, has intrinsic blind spots. In quantum machine learning, [AG-2024.01-2126] proves that even adversarially trained quantum classifiers have generalization errors that depend on mutual information between data and the quantum embedding—meaning there are attack patterns the training process cannot fully erase because they're baked into how the model represents information. Similarly, [AG-2024.04-2186] demonstrates that adversarial attacks on classical and quantum models follow different patterns, suggesting that robustness learned against one threat class may not transfer to others.

The deeper insight comes from [AG-2024.02-1732]: neural network vulnerabilities are *systemic*, analogous to the quantum uncertainty principle. Adversarial perturbations emerge directly from the loss function's gradient structure. This suggests no amount of adversarial training alone can eliminate attacks that exploit the fundamental geometry of the loss landscape.

Why this matters

Adversarial training is reactive—it immunizes against known attack directions but cannot pre-emptively eliminate all correlated failure modes in the loss landscape. [AG-2024.05-2145] hints at why: quantum correlations in certain circuits can provide *provable* robustness guarantees that classical training cannot, suggesting the problem lies in the representational capacity, not just the training algorithm.

The gap

The archive does not directly address why *simulations specifically* introduce correlation artifacts that adversarial training misses—e.g., finite-size effects, discretization errors, or unobserved dependencies in numerical approximations that remain hidden from the adversarial loss function during training.