{"id":"arxiv/2607.19335","url":"https://arxiv.org/abs/2607.19335","tags":["math.NA","cs.LG"],"score":0,"title":"1-Lipschitz Neural Networks on Hadamard Manifolds","domain":"math.NA","source":"arxiv","authors":["Davide Murari","Marta Ghirardelli","Ben Adcock","Elena Celledoni","Brynjulf Owren","Carola-Bibiane Schönlieb"],"reviews":0,"subject":"math.NA","summary":"Controlling the Lipschitz constant of a neural network is a standard way to promote robustness and stability. Most existing constraining strategies are designed for Euclidean spaces. In this work, we construct and analyze a class of 1-Lipschitz neural networks on Hadamard manifolds. Our layers are of gradient-descent type, $1$-Lipschitz, and quasi-$α$-firmly nonexpansive. The core building blocks of the proposed architecture are Busemann functions, and we exploit the properties of Busemann gradient flows to design $1$-Lipschitz geometry-preserving layers. We provide explicit constructions…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-07-21","reactions":{}}