A Classification of Reproducible Natural Exponential Families in the Broad Sense

Research paper by Shaul K. Bar-Lev, Muriel Casalis

Indexed on: 01 Jan '03Published on: 01 Jan '03Published in: Journal of Theoretical Probability


Let μ be a positive Radon measure on \(\mathbb{R}\) having a Laplace transform Lμ and F=F(μ) be a natural exponential family (NEF) generated by μ. A positive Radon measure μλ with λ>0 and its associated NEF Fλ=F(μλ) are called the λth convolution powers of μ and F, respectively, if \(L_{\mu _\lambda } = L_\mu ^\lambda\). Let fα,β(F) be the image of an NEF F under the affine transformation fα,β:x↦αx+β. If for a given NEF F, there exists a triple (α,β,λ) in \(\mathbb{R}^3 \) such that fα,β(F)=Fλ, we call F a reproducible NEF in the broad sense. In other words, an NEF F is reproducible in the broad sense if a convolution power of F equals an affine transformation of F. Clearly, for λ=1, F is reproducible in the broad sense if there exists an affinity under which F is invariant. In this paper we obtain a complete classification of the class of reproducible NEF's in the broad sense. We show that this class is composed of infinitely divisible NEF's and that it contains the class of NEF's having exponential and power variance functions as well as NEF's constituting discrete versions of the latter NEF's. We also provide a characterization of the reproducible NEF's in the broad sense in terms of their associated exponential dispersion models.