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

You've asked a sharp question that cuts across numerical methods and adversarial machine learning — but the archive provided focuses almost entirely on adversarial robustness in *machine learning*, not on discretization errors in *physics simulations*. Let me explain the distinction, and then say plainly what's missing.

The Core Difference

Discretization errors arise when you approximate continuous equations (like the Navier–Stokes equations or Einstein's field equations) using a finite grid or time step. They're *systematic approximations* baked into your numerical scheme — a trade-off between accuracy and computational cost. A finer grid or smaller time step reduces them predictably.

Adversarial perturbations, by contrast, are *worst-case input modifications* designed to fool a trained model. They're typically found by solving an optimization problem (e.g., maximizing the loss subject to a small perturbation budget). They exploit the geometry of the model's learned decision boundaries, which is very different from the geometry of a discretized PDE.

To be concrete: if you're simulating fluid flow on a coarse grid, the discretization error is the gap between your approximate solution and the true continuum solution. An adversarial perturbation, by contrast, is a tiny change to the *input data* that causes a neural network trained on flow simulations to misclassify the fluid state — it's not about approximating equations at all.

Why They're Fundamentally Different

Discretization errors scale with the mesh size and time step — they're *deterministic* and *inherent to the scheme*. Adversarial perturbations are *data-dependent* and *discovered* via optimization; they exist because neural networks carve up input space in ways that leave narrow adversarial manifolds.

One archive paper does touch on a related theme in high-energy physics: researchers found that adversarial training can *reduce overfitting to Monte Carlo artifacts* in Higgs classification, improving generalization to real data [AG-2024.11-1226]. But this is about correcting biases introduced by imperfect *simulation data*, not about discretization error per se.

What's Missing

The preprints provided don't discuss how discretization errors compound with adversarial robustness — for instance, whether a physics simulator that is discretized coarsely becomes more or less vulnerable to adversarial attacks, or how to certify robustness when your "ground truth" is itself an approximation. That gap isn't covered here.