Cumulative distribution functions for the five simplest natural exponential families (Q2338097)

Let \(X_a\) be a random variable supported on the non-negative integers with distribution function of the form \[ \mathrm{Pr}(X_a\leq k)=c_k\int_a^\infty u^k\,\mu(du)\,, \] for some sequence \((c_k)\) and positive measure \(\mu\), each independent of \(a\), where \(a\) is a parameter. The main result of the paper is that this implies that \(X_a\) must be either binomial, Poisson or negative binomial (up to translation, and ignoring trivial cases). An analogous result is also established in the continuous case: a random variable supported on \(\mathbb{R}\) whose distribution function is of a form analogous to that considered in the discrete case above must have either a Gaussian or gamma distribution. The proofs of both these results use that cross ratios are preserved only by Möbius functions.