{"id":"arxiv/2609.01034","url":"https://arxiv.org/abs/2609.01034","tags":["cs.LG","math.OC","stat.ML"],"score":0,"title":"The Multiple Timescales of Gradient Descent on the Edge of Stability: A Perturbative Derivation of the Central Flow","domain":"cs.LG","source":"arxiv","authors":["Raphaël Berthier"],"reviews":0,"subject":"cs.LG","summary":"The central flow of Cohen et al. (2025) is an empirically accurate continuous-time model of gradient descent at the edge of stability in deep learning, However, its derivation is heuristic. We propose a perturbative regime in which the central flow is the limit of gradient descent: we assume that the loss decomposes as $f = g + \\varepsilon h$; in the limit $\\varepsilon \\to 0$, the dynamics of gradient descent with learning rate $η$ converge to the gradient flow of $h$ constrained to the minimizers of $g$ of sharpness at most $2/η$. Our approach is formal rather than rigorous; it treats…","category":"research","comments":0,"disputed":false,"division":"cs","mentions":0,"discussed":0,"published":"2026-09-01","reactions":{}}