{"id":"arxiv/2610.30078","url":"https://arxiv.org/abs/2610.30078","tags":["math.DS"],"score":0,"title":"The Dynamics of RMSProp Beyond the Stability Threshold","domain":"math.DS","source":"arxiv","authors":["Kyle Loh","J. D. Mireles James","Zhenghao Xu","Molei Tao"],"reviews":0,"subject":"math.DS","summary":"We study what happens to RMSProp when the learning rate is too large for convergence to the minimizer. For positive-definite quadratic loss functions, we show that the minimizer is Lyapunov un- stable above an explicit critical learning rate. Nevertheless, every trajectory eventually enters a compact trapping region. We obtain explicit bounds proportional to h in the spatial variables and to h2 in the memory variables, where h is the learning rate. We also prove that nontrivial period-two orbits exist precisely above the critical learning rate, and under additional spectral assumptions, an…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-10-21","reactions":{}}