How do discretization errors in physics simulations differ from adversarial perturbations that robustness training targets?

I appreciate the precise framing of your question, but I need to be direct: the preprints provided address adversarial robustness in machine learning, not discretization errors in physics simulations—these are fundamentally different problems.

Let me clarify what makes them distinct:

Discretization errors arise when you approximate continuous physics (differential equations, fields, trajectories) using a finite grid or timestep. Think of simulating a fluid by dividing space into boxes; the error comes from truncating Taylor series and from information lost between grid points. These errors are *deterministic, systematic, and tied to your numerical scheme*—using a finer grid or higher-order method reduces them predictably.

Adversarial perturbations, by contrast, are *crafted worst-case input perturbations* designed to fool a trained model, often by exploiting learned features that don't align with human intuition [AG-2024.11-1226]. They're studied in machine learning precisely because they're not random noise—they're malicious and often imperceptibly small [AG-2024.04-2204].

The provided papers examine adversarial robustness mainly in quantum and classical machine learning classifiers [AG-2024.05-2145, AG-2024.04-2186, AG-2024.01-2126]. One paper does mention Monte Carlo simulation artifacts in particle physics [AG-2024.11-1226], but treats them as a generalization problem solvable by adversarial training—not as a discretization error analysis.

To answer your question properly, you'd need papers on numerical stability, convergence rates, or error propagation in physics solvers—topics not represented in this collection.