{"id":"arxiv/2606.05242","url":"https://arxiv.org/abs/2606.05242","tags":["stat.ML","cs.LG","math.NA","math.PR"],"score":0,"title":"Deterministic Envelopes for Tamed SGLD: Decoupling Stochastic Gradient Noise and Localizing Taming","domain":"stat.ML","source":"arxiv","authors":["Yiwei Zhou","Ziheng Chen"],"reviews":0,"subject":"stat.ML","summary":"Stochastic gradient Langevin algorithms often use tamed denominators to stabilize superlinear drifts. This paper shows that when the denominator depends on the current stochastic gradient, the transformed update can have a biased conditional mean even if the original stochastic gradient is unbiased. This creates a stationary mean-shift channel that is absent for deterministic denominators.We propose a structure-preserving framework for designing tamed denominators. The construction keeps the denominator deterministic given the current state, and uses localized deterministic envelopes to…","category":"research","comments":0,"disputed":false,"division":"stat","mentions":0,"discussed":0,"published":"2026-06-03","reactions":{}}