The paper studies a family of algorithms for sampling from probability distributions, which is a core task in statistics, machine learning, and scientific computing. The key idea involves equipping ordinary space with a curved geometry derived from a convex function, turning it into what mathematicians call a Hessian manifold. On this curved space, one can define a natural random process called a Mirror Langevin diffusion that, if run long enough, will produce samples from a target distribution of interest. The central practical question is: how quickly does this process converge, and can one cleverly choose the geometry to make convergence faster, even when the target distribution has awkward shapes that would slow down standard methods?
The main theoretical contributions involve establishing conditions under which the Mirror Langevin diffusion converges exponentially fast to the target distribution. The authors use tools called Lyapunov functions to derive inequalities (Poincare and log-Sobolev inequalities) that control how quickly information is lost and equilibrium is approached. Importantly, they show that by choosing the geometry carefully, one can sometimes achieve fast exponential convergence even for distributions that are not strongly log-concave, a property that standard Langevin methods typically require for similar guarantees.
Because continuous diffusion processes must be approximated in practice, the authors also introduce a discrete Markov chain that mimics the Mirror Langevin diffusion. This chain is a two-step Gibbs sampler, meaning it alternates between updating different components of the state in a structured way, and it is related to a previously studied algorithm called the Sinkhorn Markov chain. The authors prove that this discrete chain converges at a rate consistent with what the continuous diffusion theory predicts, using ideas borrowed from the field of optimal transport. This bridges the gap between the clean theory of continuous processes and algorithms that can actually be run on a computer.