{"id":"arxiv/2604.13870","url":"https://arxiv.org/abs/2604.13870","tags":["math.OC","cs.LG"],"score":0,"title":"Gradient Descent's Last Iterate is Often (slightly) Suboptimal","domain":"math.OC","source":"arxiv","authors":["Guy Kornowski","Ohad Shamir"],"reviews":0,"subject":"math.OC","summary":"We consider the well-studied setting of minimizing a convex Lipschitz function using either gradient descent (GD) or its stochastic variant (SGD), and examine the last iterate convergence. By now, it is known that standard stepsize choices lead to a last iterate convergence rate of $\\log T/\\sqrt{T}$ after $T$ steps. A breakthrough result of Jain et al. [2019] recovered the optimal $1/\\sqrt{T}$ rate by constructing a non-standard stepsize sequence. However, this sequence requires choosing $T$ in advance, as opposed to common stepsize schedules which apply for any time horizon. Moreover, Jain…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-04-15","reactions":{}}