Quantitative approximation for Fourier neural operators in controlled ordinary differential equations
Year: 2026
Type: preprint
Abstract: Fourier neural operators (FNOs) avoid the curse of parametric complexity when approximating the solution operator of certain partial differential equations (PDEs). While recent studies primarily focus on PDEs, in this article we derive upper error bounds for approximating the solution operator of controlled ordinary differential equations (CODEs) $\dot{y}=f(y,u)$ using a FNO. Under a Carathéodory setting and suitable regularity assumptions, including a smallness condition, we prove that the size of the FNO grows at most algebraically in $1/\varepsilon$, namely like $\varepsilon^{-(1+d/(2(s-1)))}$ up to logarithmic factors for GELU networks, and is thus sub-quadratic for sufficiently smooth $f$, where $d=1+n_y+n_u$ and $s$ denotes the Sobolev regularity of $f$. Our proof is based on emulating the Picard iteration through the layers of the FNO while exploiting the differentiation formula of the discrete Fourier transform. This result contributes to the quantitative approximation theory of FNOs for physically motivated problems and provides a theoretical foundation for leveraging a FNO as efficient forward surrogate for the solution operator of CODEs, especially in inverse parameter estimation where repeated evaluations of the forward operator are needed.
Source: arXiv (Cornell University)
Authors: Zenner, Dean, Rupp, Andreas, Schuster, Thomas
Institution: Saarland University
Language: English
Subfield: Statistical and Nonlinear Physics
Field: Physics and Astronomy
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7235243697
