{"id":"arxiv/2610.37858","url":"https://arxiv.org/abs/2610.37858","tags":["math.ST","math.NA","stat.CO"],"score":0,"title":"Wasserstein Bounds for Euler and Randomized Midpoint Langevin Sampling with Noisy Gradients","domain":"math.ST","source":"arxiv","authors":["Yifei Xiong","Vinayak Rao"],"reviews":0,"subject":"math.ST","summary":"Stochastic-gradient Langevin methods are sampling algorithms that replace the gradient of the log probability density by a noisy, possibly biased evaluation. We study how such gradient-evaluation errors affect the Euler and randomized midpoint discretizations of Langevin dynamics by comparing each noisy chain with its exact-gradient counterpart. For continuously differentiable, strongly log-concave targets with globally Lipschitz gradients, we derive comparison bounds under a sub-Gaussian condition on the gradient noise. The resulting 2-Wasserstein bounds hold uniformly over iterations, and…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-10-27","reactions":{}}