Entropy and the Law of Small Numbers

Ioannis Kontoyiannis, Peter Harremoes, Oliver Johnson

Published 2002-11-01, updated 2004-11-17Version 2

Two new information-theoretic methods are introduced for establishing Poisson approximation inequalities. First, using only elementary information-theoretic techniques it is shown that, when $S_n=\sum_{i=1}^nX_i$ is the sum of the (possibly dependent) binary random variables $X_1,X_2,...,X_n$, with $E(X_i)=p_i$ and $E(S_n)=\la$, then \ben D(P_{S_n}\|\Pol)\leq \sum_{i=1}^n p_i^2 + \Big[\sum_{i=1}^nH(X_i) - H(X_1,X_2,..., X_n)\Big], \een where $D(P_{S_n}\|{Po}(\la))$ is the relative entropy between the distribution of $S_n$ and the Poisson($\la$) distribution. The first term in this bound measures the individual smallness of the $X_i$ and the second term measures their dependence. A general method is outlined for obtaining corresponding bounds when approximating the distribution of a sum of general discrete random variables by an infinitely divisible distribution. Second, in the particular case when the $X_i$ are independent, the following sharper bound is established, \ben D(P_{S_n}\|\Pol)\leq \frac{1}{\lambda} \sum_{i=1}^n \frac{p_i^3}{1-p_i}, % \label{eq:abs2} \een and it is also generalized to the case when the $X_i$ are general integer-valued random variables. Its proof is based on the derivation of a subadditivity property for a new discrete version of the Fisher information, and uses a recent logarithmic Sobolev inequality for the Poisson distribution.