A Statistical Formulation Gap for Nonlinear Multiscale Physics-Informed Learning

DOI10.48550/arXiv.2607.15702MaRDI QIDQ7284182arXivFDO

Authors Ronald Katende

Publication date 17 July 2026

Published in arXiv

Tldr We formulate a statistical gap between finite-width models learning nonlinear multiscale elliptic equations and their exact solutions with sample complexity lower bounds depending on scale $ε$. We also show that the variational approach to physics-informed learning can remove this gap.

Full work available at URL https://doi.org/10.48550/arXiv.2607.15702

arXiv classification

cs.LGmath.APmath.NA

Cites work

• A Statistical Formulation Gap for Nonlinear Multiscale Physics-Informed Learning
• Estimates on the generalization error of Physics Informed Neural Networks (PINNs) for approximating PDEs
• Physics informed neural networks for elliptic equations with oscillatory differential operators
• Random test functions, $H^{-1}$ norm equivalence, and stochastic variational physics-informed neural networks
• The deep Ritz method: a deep learning-based numerical algorithm for solving variational problems
• Variational Physics-Informed Neural Networks For Solving Partial Differential Equations

Cited in

(1)

• A Statistical Formulation Gap for Nonlinear Multiscale Physics-Informed Learning

This page was built for publication: A Statistical Formulation Gap for Nonlinear Multiscale Physics-Informed Learning

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q7284182)