Neural operators are popular surrogates for PDE solution maps, but quantitative guarantees for nonlinear reaction-diffusion systems have been scarce. In a new arXiv preprint, Takashi Furuya, Ryo Ozawa, and Jenn-Nan Wang address this gap by analyzing Laplacian-based neural operators that map initial conditions to time-dependent solutions.
Using the spectral representation of the diffusion semigroup, the authors derive polynomial complexity bounds on network depth, width, and spectral rank. This shows that Laplacian eigenfunction-based operators can exploit the PDE structure and avoid the curse of parametric complexity that plagues generic operator learning.
Numerical experiments back the theory and indicate that Laplacian eigenfunction representations are more effective than Fourier-based spectral representations for these systems. The work provides a step toward more principled design of neural operators for reaction-diffusion problems.