{"id":"arxiv/2607.22478","url":"https://arxiv.org/abs/2607.22478","tags":["math.NA"],"score":0,"title":"ReLU$^k$ Neural de Rham Complexes","domain":"math.NA","source":"arxiv","authors":["Kaibo Hu","Jindong Wang","Jinchao Xu"],"reviews":0,"subject":"math.NA","summary":"We construct finite-dimensional de Rham subcomplexes generated by fixed-neuron shallow ReLU$^k$ neural networks, a class of spaces known to provide optimal approximation rates. For neurons of the form $s_i(x)=ω_i\\cdot x+b_i$, we introduce spaces of neural differential forms: differential $p$-forms whose coefficients are the ReLU$^k$ ridge functions $σ_{k-p}(s_i)$. These spaces are compatible with the exterior derivative because differentiating a ReLU power lowers its order by one, and for each fixed neuron, differentiation amounts to exterior multiplication by the fixed one-form $ d s_i$.…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-07-24","reactions":{}}