A characterization of the Wishart exponential families by an invariance property (Q1120225)

The natural exponential family (n.e.f.) is defined as a slight modification of linear exponential families [cf. \textit{O. Barndorff- Nielsen}: Information and exponential families in statistical theory. (1986; Zbl 0387.62011)]. The Wishart n.e.f. with parameter p is a set of probability distributions on \(\bar S\) defined by \[ (*)\quad F_ p=\{\exp [2^{-1}Tr(\Gamma x)](\det \Gamma)^ p\mu_ p(dx);\quad \Gamma \in S\} \] where \(E=space\) of \(d\times d\) real symmetric matrices; S and \(\bar S\) are subsets of E of positive definite and positive semi- definite matrices; b belongs to \(\Lambda\), \(\Lambda =\{i/2:\) \(i=1,2,...,d-1\}\cup (d-1,\infty)\); and \(\mu_ p\) is a suitable measure on \(\bar S.\) Let G be the subset of E of invertible matrices. For a in G define the automorphism \(g_ a\) of E by \(g_ a(x)=a'xa\) where \(a'=transpose\) of a. In Theorem 1, some known results are reproved to illustrate the interest in working with n.e.f. and without coordinates. The main result of the paper is Theorem 2: A n.e.f. F on E is invariant by \(g_ a\) for all a in G iff there exists a p in \(\Lambda\) such that either \(F=F_ p\) or F is the image of \(F_ p\) by \(x\mapsto -x\).