{"id":"arxiv/2610.06070","url":"https://arxiv.org/abs/2610.06070","tags":["math.OC","cs.LG"],"score":0,"title":"Last-Iterate Convergence Rate of Normalized Gradient Descent under Hölder Smoothness","domain":"math.OC","source":"arxiv","authors":["Yuki Takezawa","Eduard Gorbunov"],"reviews":0,"subject":"math.OC","summary":"Normalized gradient descent is a widely studied adaptive optimization method. Most existing analyses focus on the best iterate or a weighted average of the iterates, whereas practical implementations typically return the last iterate. In this paper, we study the last-iterate convergence of normalized gradient descent for convex, $(ν,M_ν)$-Hölder-smooth objectives. For a constant stepsize, we establish an upper bound of $\\mathcal{O}\\bigl((\\log^2(T)/T)^{(1+ν)/2}\\bigr)$, which contains a logarithmic overhead relative to the known $\\mathcal{O}\\bigl(T^{-(1+ν)/2}\\bigr)$ guarantees for the best and…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-10-05","reactions":{}}