The central question of this paper is: if you observe only a few coordinates of an infinitely long random sequence where the order of elements does not matter (called an exchangeable sequence), can you uniquely recover the full probabilistic structure governing the entire sequence? By de Finetti's theorem, such sequences are always mixtures of simpler independent processes, so the question becomes whether a small number of observations pin down that mixing distribution exactly.
The authors find precise thresholds for how many coordinates you need. Under a mild "localization" condition on the underlying probability model, three coordinates are enough to determine everything; under a completeness condition alone, four suffice; and they show you cannot generally do better. For a classical family of models called exponential families, a number-theoretic condition on which values the sufficient statistic can take (related to something called the Muntz-Szasz theorem from approximation theory) can bring the threshold back down to three, and this is tight in one dimension. They also characterize exactly when a strong uniformity property holds across all possible mixing distributions using a density condition on certain weighted sums, revealing that simply knowing where the sufficient statistic lives is not enough to guarantee this property.
As a striking application, the paper sharpens a classical result by Freedman about sequences that are symmetric under rotations as well as reordering. The authors show that if just three coordinates of an exchangeable sequence look rotationally symmetric, the entire infinite sequence must be rotationally symmetric, whereas two coordinates are not enough to force this conclusion. They also work out complete concrete criteria for when certain multivariate distributions based on the lq norm can be extended to infinite exchangeable sequences, giving exact answers to a natural extendibility problem.