Minimax posterior convergence rates and model selection consistency in high-dimensional DAG models based on sparse Cholesky factors (Q2284379)

Consider a sample of data \(X_1,\ldots,X_n\) which has the \(p\)-dimensional normal distribution with the \(p\)-dimensional mean vector \(\mu\) and \(p \times p\)-dimensional covariance matrix \(\Sigma_n\). For every positive definite matrix, the modified Cholesky decomposition (MCD) guarantees the existence and unique Cholesky factor \(A_n\) and diagonal matrix \(D_n\) such that \(\Omega_n:=\Sigma_n^{-1}=(I_p - A_n)^{\tau} D_n^{-1} (I_p -A_n)\). The sparsity of a Gaussian directed acyclic graph (DAG) can be uniquely encoded by the Cholesky factor \(A_n\) through the structure of the graph. In this paper the required conditions on the dimensionality, sparsity, structure of the Cholesky factor \(A_n\) and the lower bound of the nonzero elements in \(A_n\) are significantly weakened. The minimax or nearly minimax posterior convergence rates are obtained for precision matrices with respect to the spectral norm and matrix \(l_{+\infty}\) norm. The proposed method significantly improves the model selection performance in practice.