{"id":"palomar/PALOMAR-2026-11-14-000001","url":"https://github.com/lowrank/last-iterate-sgd-bound/tree/82241483b73d53c8084edb33a7d946d070674562","json":"https://data.palomar-registry.org/entries/PALOMAR-2026-11-14-000001-v1.json","repo":"https://github.com/lowrank/last-iterate-sgd-bound","tags":["stat.ML","cs.LG","msc:90C15","msc:90C25","msc:68T05"],"score":0,"title":"Improved last-iterate bounds for SGD with and without replacement","trust":"high","domain":"stat.ML","source":"palomar","authors":["Yimin Zhong"],"reviews":0,"subject":"stat.ML","summary":"For a finite family of differentiable convex beta-smooth losses with a population minimizer, this formalization proves an expected last-iterate population-loss bound for stochastic gradient descent under uniform sampling with replacement and uniform sampling without replacement. With delta = min(1, 2 - eta beta), the bound is 280000 times log(T+2)^3 / log(2+log(T+2)) times [R^2 / (delta^3 eta T) + eta sigma_star^2 / delta^4]. It holds for all 0 < eta beta < 2, every positive WR horizon, and every WOR horizon through one full pass, including T = n. Under interpolation at eta = 1/beta, this…","category":"research","comments":0,"disputed":false,"division":"stat","mentions":0,"verified":true,"discussed":0,"published":"2026-11-14","reactions":{},"formalized":true,"formalizes":[{"title":"Improved last-iterate bounds for SGD with and without replacement","identifier":"","relationship":"formalizes"}]}