On the Letac-Massam conjecture and existence of high dimensional Bayes estimators for graphical models (Q2293718)

Inference on graphical models has been a topic of contemporary interest. One aspect in such study is establishing sufficient and/or necessary conditions for the existence of high dimensional estimators for the model parameters. A useful notion for flexible high dimensional inference is the multiple-shape parameter concept. The families of Type I and Type II Wishart distributions, introduced by \textit{G. Letac} and \textit{H. Massam} [Ann. Stat. 35, No. 3, 1278--1323 (2007; Zbl 1194.62078)], have multiple-shaped parameters. Those distributions have been used for objective Bayesian model selection in Gaussian graphical models. But, since the involved distributions do not have the domain of their high dimensional parameters fully identified, it is not clear when these distributions yield well-defined and proper priors. In this direction there is the Letac-Massan (LM) conjecture. This conjecture is critical for understanding when those priors lead to well-defined Bayesian estimators, since their existence (of estimators) is not always guaranted in high-dimensional setting, with the sample size \(n\) being much smaller than \(p,\) the dimension of the distribution. Type I and Type II Wishart distributions when used for general decomposable graphs, except for homogeneous graphs and paths, do not have their (multiple-shape) parameter spaces fully identified. For non-homogeneous decomposable graph \(\mathcal{G}\), Letac and Massan [loc. cit.] first fix a perfect order \(\mathcal{P}\) of \(\mathcal{G}\) and identify two corresponding subsets \(A_\mathcal{P}\) and \(B_\mathcal{P}\) for the corresponding parameter space. By taking the unions of \(A_\mathcal{P}\) and \(B_\mathcal{P}\) over all perfect orders of \(\mathcal{G}\), they conjectured the union of the two just mentioned unions can fully cover, and hence, identify the parameter space. In the reviewed paper, the LM conjecture is approached by an appropriate acyclic directed graph (DAG) setting. The Type I and Type II Wishart distributions are transformed to distributions that correspond to appropriate perfect DAG models. Through the analysis of LM conjecture on the corresponding domain, the authors derive a necessary condition for the conjecture to hold, namely, for non--homogeneous decomposable graph \(\mathcal{G}\) no nested separators can occur in order to the LM conjecture to hold on \(\mathcal{G}\). Using their result, the authors can provide graphs that are counterexamples to the LM conjecture.