Joint estimation of multiple graphical models (i.e., multiple precision matrices) has emerged as an important topic in statistics. Unlike separate estimation, joint estimation can leverage shared structural patterns across multiple graphs to yield more accurate results. In this paper, we present an efficient and tuning-free method named MIGHT (Multi-task Iterative Graphical Hard Thresholding) to jointly estimate multiple graphical models. We reformulate the joint model into a series of multi-task learning problems in a column-by-column manner, and then solve these problems by using an iterative algorithm based on the hard thresholding operator. Theoretically, we derive the non-asymptotic error bound for our method. We prove that, under proper signal conditions, our method attains selection consistency and an improved error bound, and also exhibits asymptotic normality -- properties rarely explored in existing joint graphical model estimation literature. The performance of our method is validated through numerical simulations and real data analysis of a cancer gene-expression RNA-seq dataset.
翻译:暂无翻译