Reference priors for the orbit in a group model (Q2277692)

Let the Lie group G act freely (but not transitively) to the left on a differential manifold \(\Omega\) that is the parameter space of some statistical problem. The parameter \(\theta\) of interest indexes the orbits in \(\Omega\) under G. A prior measure \(p(\omega)d\omega\) is proposed by the requirement that conditional on given \(\theta\) it is the right Haar measure on G, and the marginal measure on \(\Omega\) /G has a density of the form \(\lim_{n\to \infty}(n^{-1}\det I_ n(\theta))^{1/2},\) where \(I_ n(\theta)\) is the Fisher information matrix of the maximal invariant in the sample space based on a random sample of size n from the distribution indexed by \(\omega\in \Omega.\) It turns out that \(p(\omega)\) is the Bernardo minimally \(\theta\)- informative prior and it is shown that \(p(\omega)=\tilde p(\omega)(J(\omega))^{-1/2}\), in which \(\tilde p(\omega)\) is Jeffrey's prior and J(\(\omega\)) will be defined next. Let \(A_ 1,...,A_ k\) be a basis of the tangent space of G at its identity element and put \(A^*_ i(\omega)=df_{\omega}(A_ i)\), \(i=1,...,k\), where for \(\omega\in \Omega\), \(df_{\omega}\) is the differential of the mapping \(f_{\omega}: G\to \Omega\) defined by \(g\to g\omega\). Then \(J(\omega)\) is defined as the determinant of the \(k\times k\) matrix whose (i,j)- element is \(A_ i^*{}'(\omega)I(\omega)A^*_ j(\omega)\), where \(I(\omega)\) is the Fisher information matrix (of one observation). A formula for \(J(\omega)\) in terms of expectations is also presented. The value of \(J(\omega)\) does not depend on the parametrization but the function J is determined only up to a positive multiplicative constant as a result of freedom of choice of the basis of the tangent space. Under some additional conditions the assumption of free action may be relaxed. Applications are given to a multivariate normal population with mean vector \(\mu\), covariance matrix V, and \(\theta\) is either the correlation matrix, or the set of eigenvalues of V, or the noncentrality parameter \(\mu '\mu /\sigma^ 2\) when \(V=\sigma^ 2I\).