The central problem here is figuring out how well you can approximate the probability distribution of a sum of many dependent random quantities using a normal (Gaussian) distribution, and doing so with a precise, quantitative error bound. The random quantities in question come from a Markov chain, a process where each step depends only on the previous one, and the sums are built from "martingale differences," which are increments that have zero average when you condition on the past. The authors measure approximation quality using something called Wasserstein distance, which captures not just whether probabilities match but also how far apart outcomes are geometrically. The main achievement is showing that even with this stronger notion of closeness, and even in many dimensions, the approximation error shrinks at the rate one divided by the square root of the number of steps, which is the best possible rate you could hope for.
Getting this result required overcoming a key technical tension: the tools normally used to prove Gaussian approximation work best when observations are independent, but Markov chain observations are correlated across time. The authors handle this using an approach based on something called Stein's method, which turns the approximation problem into a question about how certain operators act on smooth functions. They extend a framework involving the Ornstein-Uhlenbeck operator, a mathematical tool that connects diffusion processes to Gaussian distributions, and they carefully track how dependencies in the chain affect higher-order terms in the error.
The second main technical contribution is a new coupling construction called "refresh-then-maximal coupling." Couplings are ways of linking two probability processes together so their paths stay close. The idea here is to first resample the chain from scratch at one step (the "refresh"), which breaks temporal correlation and preserves a key algebraic identity needed for Stein's method, and then immediately apply a "maximal coupling" that makes the restarted chain and the original chain agree as often as mathematically possible. Together these two ingredients let the authors control the error tightly without the usual penalties that arise from dependence. Both techniques are presented in a way that should be useful for other researchers working on approximation problems involving time-dependent data.