Gemini: graph estimation with matrix variate normal instances (Q2249840)

The paper under review presents a theoretical framework for estimating the row and column covariance and inverse covariance matrices using only one matrix from matrix-variate normal distribution. The author establishes consistency and rates of convergence in the operator and the Frobenius norm of the covariance matrices and their inverses, proves large deviation results for the sample correlation estimators (proposed for estimating both the row an column correlation and covariance matrices given a single matrix or multiple replicates of the matrix-normal data), and provides conditions that guarantee simultaneous estimation of the graphs for both rows and columns. Provided simulation evidence and a real data example show that the methodology allows to recover graphical structures and also to estimate the precision matrices effectively.