The paper develops a mathematical framework for measuring statistical dependencies among multiple random variables. Classical probability theory has good tools for describing pairwise correlations, but capturing higher-order interactions, where three or more variables are jointly dependent in ways that cannot be explained by looking at smaller groups, is much harder. The authors use a structure called a partition lattice, which organizes the different ways a set of variables can be grouped or split, and they use a classical tool from combinatorics called the Mobius function to assign numerical "interaction" values that detect these higher-order dependencies. The focus is on symmetric partition lattices, meaning the grouping structure does not change when you relabel the variables, which is a natural assumption when the variables play interchangeable roles.
A central result is a characterization of which symmetric partition lattices are possible, described through integer partitions (ways of writing a number as a sum of smaller numbers), along with a recursive formula for computing the key Mobius coefficients. The authors also settle a foundational question: if an interaction value is invariant under relabeling variables, does that mean the underlying lattice structure must also be symmetric? They show the answer is yes, as long as the individual variable spaces are large enough, by constructing probability distributions whose "partition products" are linearly independent. They also identify which parts of a lattice force the interaction to be zero regardless of the probability distribution chosen, and build explicit examples of distributions that have zero interaction without factoring into independent pieces in an obvious way.
The second half of the paper applies these ideas to specific families of distributions and lattices. The authors study generalized Lancaster distributions, which are built from orthogonal polynomial expansions, and define two new families called Streitberg and size-limited partition lattices, for which they derive explicit formulas. For Gaussian distributions, a clean result emerges: the interaction associated with a given lattice vanishes if and only if the distribution factorizes according to some nontrivial partition in the lattice, linking the abstract combinatorics back to familiar statistical independence. They also give conditions under which interactions are guaranteed to be nonzero for a broad class of distributions defined through their characteristic functions, providing practical tools for distinguishing genuine higher-order dependence from coincidental zeros.