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 Category: research Comments: 0 Disputed: false Division: math Mentions: 0 Discussed: 0 Published: 2026-03-18 Page: https://delphi.science/r/arxiv/2603.17785 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…