{"id":"arxiv/2603.02782","url":"https://arxiv.org/abs/2603.02782","tags":["math.OC","math.DS","math.NA"],"score":0,"title":"A non-autonomous center-stable set theorem for saddle avoidance in optimization","domain":"math.OC","source":"arxiv","authors":["Andreea-Alexandra Muşat","Nicolas Boumal"],"reviews":0,"subject":"math.OC","summary":"Optimization algorithms are unlikely to converge to strict saddle points. Proofs to that effect rely on the Center-Stable Manifold Theorem (CSMT), casting algorithms as dynamical systems: $x_{k+1} = g_k(x_k)$. In its standard form, the CSMT is limited to autonomous systems (the maps $g_k$ are all the same). To study algorithms such as gradient descent with non-constant step-size schedules, we need a non-autonomous CSMT. There are a few, but they are unable to handle, for example, vanishing step sizes. To cover such scenarios, we establish a new Center-Stable Set Theorem (CSST) for…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-03-03","reactions":{}}