Improved Analysis of PINNs: Alleviate the CoD for Compositional Solutions

Improved Analysis of PINNs: Alleviate the CoD for Compositional Solutions

Year:    2023

Author:    Yuling Jiao, Xiliang Lu, Jerry Zhijian Yang, Cheng Yuan, Pingwen Zhang

Annals of Applied Mathematics, Vol. 39 (2023), Iss. 3 : pp. 239–263

Abstract

In this paper, we present an improved analysis of the Physics Informed Neural Networks (PINNs) method for solving second-order elliptic equations. By assuming an intrinsic sparse structure in the underlying solution, we provide a convergence rate analysis that can overcome the curse of dimensionality (CoD). Specifically, using some approximation theory in Sobolev space together with the multivariate Faa di Bruno formula, we first derive the approximation error for composition functions with a small degree of freedom in each compositional layer. Furthermore, by integrating several results on the statistical error of neural networks, we obtain a refined convergence rate analysis for PINNs in solving elliptic equations with compositional solutions. We also demonstrate the benefits of the intrinsic sparse structure with two simple numerical examples.

You do not have full access to this article.

Already a Subscriber? Sign in as an individual or via your institution

Journal Article Details

Publisher Name:    Global Science Press

Language:    English

DOI:    https://doi.org/10.4208/aam.OA-2023-0021

Annals of Applied Mathematics, Vol. 39 (2023), Iss. 3 : pp. 239–263

Published online:    2023-01

AMS Subject Headings:    Global Science Press

Copyright:    COPYRIGHT: © Global Science Press

Pages:    25

Keywords:    Composition function deep neural network approximation.

Author Details

Yuling Jiao

Xiliang Lu

Jerry Zhijian Yang

Cheng Yuan

Pingwen Zhang