{"id":"arxiv/2610.29532","url":"https://arxiv.org/abs/2610.29532","tags":["math.ST"],"score":0,"title":"Minimax Rates for imperfect Physics-Informed Regression","domain":"math.ST","source":"arxiv","authors":["Nathan Doumèche"],"reviews":0,"subject":"math.ST","summary":"Physics-informed machine learning incorporates physical knowledge into data-driven models, often through a partial differential equation (PDE) that the target is assumed to satisfy. However, how much this knowledge improves statistical performance remains unclear, especially when the physical model is imperfect. Since satisfying a PDE implies being differentiable, a natural benchmark for physics-informed regression is the best achievable performance when smoothness is the only prior information about the target function. We show that there is a characteristic threshold for the modeling error…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-10-21","reactions":{}}