{"id":"arxiv/2606.10559","url":"https://arxiv.org/abs/2606.10559","tags":["stat.ME","math.PR","stat.ML"],"score":0,"title":"Deterministic Denominator Design for Localized Tamed Stochastic Gradient Langevin Dynamics","domain":"stat.ME","source":"arxiv","authors":["Yiwei Zhou","Ziheng Chen"],"reviews":0,"subject":"stat.ME","summary":"If the denominator in a tamed stochastic gradient Langevin update uses the current stochastic-gradient draw, the conditional mean can be biased even when the stochastic-gradient oracle is unbiased. A state-dependent denominator fixed before that draw removes this coupling. We build practical deterministic denominators from a short pilot run. A log-scale proxy is fitted to the growth score $G_\\star(x)=\\|b(x)\\|/(1+\\|x\\|)$, and empirical pilot quantiles set the activation thresholds of a local proxy-quantile envelope. The reported proxy-quantile experiments use this local denominator.…","category":"research","comments":0,"disputed":false,"division":"stat","mentions":0,"discussed":0,"published":"2026-06-09","reactions":{}}