Wishart laws on colored matrix cones
Gérard Letac and Hélène Massam initiated in their 2007 Annals of Statistics article [1] "Wishart distributions for decomposable graphs" a systematic mathematical study of Wishart matrices subject to natural statistical constraints.
Classical Wishart matrices are supported on the cone Sym+(p,R) of symmetric positive definite matrices. In modern statistics, Wishart matrices are considered on:
- cones of matrices with obligatory zeros (sparsity)
- cones of matrices with subsets of equal terms (coloring).
These multiple-shape parameter Wishart distributions are useful in high dimensional statistical inference.
In this talk I will present some recent results obtained
in the paper [2], jointly with H. Ishi(Osaka), B. Kolodziejek(Warsaw) and H. Massam.
In [2] we consider multivariate Gaussian models N(0,Σ) for the random variable Z=(Z1,…,Zp), invariant under the action of a subgroup Γ of the group Sp of permutations on {1,…,p}.
We compute the normalization constants for Wishart laws on colored subcones of Sym+(p,R).
The statistical objective is a Bayesian model selection in the class C of such complete Gaussian models invariant by the action of a subgroup Γ of the symmetric group Sp, also called saturated RCOP models.
Using the representation theory of the symmetric group Sp on the field of reals, we derive the distribution of the maximum likelihood estimate of the covariance parameter Σ and also the analytic expression of the normalizing constant of the Diaconis-Ylvisaker conjugate prior for the precision parameter K=Σ−1. We can thus perform Bayesian model selection in the class C.
We illustrate our results with Frets' Heads example of dimension 4 and a high-dimensional example in the case of cyclic groups.
[1] G. Letac, H. Massam, Wishart distributions for decomposable graphs, Ann. Statist. Volume 35, Number 3(2007), 1278-1323.
[2] P. Graczyk, H. Ishi, B. Kolodziejek, H. Massam,
Model selection in the space of Gaussian models invariant by symmetry, to appear in Annals of Statistics, 2022. arXiv:2004.03503