The paper revisits a classical topic in probability theory concerning what happens when you take a large collection of random events and independently keep each one with some small probability, a procedure called "thinning." A well-known result called the Law of Thin Numbers says that when you thin a sum of independent, identically distributed random variables in just the right way, the resulting distribution converges to a Poisson distribution. This is a probabilistic analogue of the Central Limit Theorem, which describes convergence to a Gaussian distribution in a different setting. The paper also addresses Yu's Thinning Lemma, a technical result that controls how information-theoretic quantities like relative entropy behave under thinning.
The main novelty is the proof technique. Rather than using the methods from previous work, the authors employ a stochastic variational formula for relative entropy, which is a way of expressing how different two probability distributions are using an optimization problem over stochastic processes. The specific process they use is called the Poisson-Follmer process, a continuous-time random process that serves as a probabilistic tool for studying discrete, count-based distributions. This approach draws inspiration from techniques developed in the continuous setting, where analogous Gaussian results are often proved using Brownian motion.
By using this new approach, the authors obtain explicit quantitative bounds, meaning they can say not just that convergence happens but how fast it happens. These new convergence rates complement and in some cases extend what was previously known in the literature. The work fits into a broader research program connecting information theory, optimal transport, and probability, showing that tools originally developed for continuous probability distributions can be adapted to yield new insights about discrete distributions as well.