{"id":"arxiv/2603.17785","url":"https://arxiv.org/abs/2603.17785","tags":["math.OC","cs.AI"],"score":0,"title":"A Dual Certificate Approach to Sparsity in Infinite-Width Shallow Neural Networks","domain":"math.OC","source":"arxiv","authors":["Leonardo Del Grande","Christoph Brune","Marcello Carioni"],"reviews":0,"subject":"math.OC","summary":"In this paper, we study total variation (TV)-regularized training of infinite-width shallow ReLU neural networks, formulated as a convex optimization problem over measures on the unit sphere. Our approach leverages the duality theory of TV-regularized optimization problems to establish rigorous guarantees on the sparsity of the solutions to the training problem. Our analysis further characterizes how and when this sparsity persists in a low noise regime and for small regularization parameter. The key observation that motivates our analysis is that, for ReLU activations, the associated dual…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-03-18","reactions":{}}