Published June 22, 2026 | Version v1
Preprint Open

A Novel CNN Ablation Reveals Limited Fourier Inductive Bias in LiteFNO: A Reproducibility Study

Authors/Creators

Description

We present a from-scratch reproducibility study and critical ablation of LiteFNO (Ahn et al., 2025), a lightweight Fourier Neural Operator for time-dependent PDEs.

Key contributions:
- From-scratch reimplementation with documented ambiguities
- Parameter-matched low-rank CNN baseline (absent in original paper)
- CNN matches or outperforms LiteFNO across 3 seeds on Gray-Scott (32×32)
- Confirms "compact beats dense" but shows gains come from compactness, not Fourier bias

Full code: https://github.com/AIscend-Research/litefno-repro

Files

LiteFNO_AI4Math.pdf

Files (1.1 MB)

Name Size Download all
md5:100fb23101b436ab77a1cb284f25cffb
1.1 MB Preview Download

Additional details

Software

Repository URL
https://github.com/AIscend-Research/litefno-repro/