{"id":"arxiv/2607.07778","url":"https://arxiv.org/abs/2607.07778","tags":["cs.LG","math.PR","stat.ML"],"score":0,"title":"A law of robustness for two-layer neural networks with arbitrary weights","domain":"cs.LG","source":"arxiv","authors":["Yitzchak Shmalo"],"reviews":0,"subject":"cs.LG","summary":"Bubeck, Li and Nagaraj conjectured that, for generic data, any two-layer neural network with $m$ neurons that fits $n$ noisy labels must have Lipschitz constant at least of order $\\sqrt{n/m}$, with no restriction on the size of the weights. Bubeck and Sellke proved a universal version of this law for Lipschitz-parameterized classes, but under a polynomial bound on the parameters; at depth three that boundedness hypothesis is genuinely necessary. The two-layer unbounded-weight case requires a different argument. We prove the conjectured law, up to one logarithmic factor, for every continuous…","category":"research","comments":0,"disputed":false,"division":"cs","mentions":0,"discussed":0,"published":"2026-07-08","reactions":{}}