The paper tackles a fundamental question in probability theory: when does a complicated random quantity behave like a Poisson distribution in the limit? Poisson distributions count rare events, like the number of typos on a page or radioactive decays per second. The authors work in two structured mathematical settings, called Poisson and Rademacher chaoses, which are spaces of random variables built from underlying random inputs in a layered, polynomial-like fashion. The central result gives a precise bound on how far a random variable in these settings is from a Poisson distribution, measured in total variation distance, which captures the worst-case difference in probabilities between two distributions.
The key finding is a "four-moment criterion": under the right conditions, checking just the first four moments of a random variable is enough to determine whether it converges to a Poisson law. This mirrors celebrated results in the normal distribution setting, where matching four moments can guarantee convergence to a Gaussian. The paper shows that in the Poisson chaos setting, the first four moments essentially tell the whole story. In the Rademacher setting, one extra condition is needed, related to how much any single input variable influences the outcome. The authors prove this extra condition is genuinely necessary by constructing an explicit example where it fails and Poisson convergence breaks down.
The proofs combine several powerful tools: the Chen-Stein method for Poisson approximation, exchangeable pairs, and a spectral perspective on chaos decompositions. A notable structural discovery is what the authors call "unit-jump rigidity": the moment conditions not only push the distribution toward Poisson but simultaneously rule out large jumps in the random variable, which are incompatible with a Poisson limit. The paper also constructs examples showing that certain rigidity properties do not hold universally once natural lattice constraints are removed, carefully mapping the boundaries of what four-moment conditions can and cannot guarantee.