Title: Network induced large covariance matrix estimation

Abstract: In this paper, we consider to estimate network/community induced large covariance matrices. The massive biomedical data (e.g. gene expression or neuroimaging data) often include networks where features are correlated with each other, and thus the covariance matrix in a complex yet organized topology. Although, current large covariance matrix and precision matrix estimation methods using thresholding or shrinkage strategies could provide satisfactory overall covariance/precision matrix estimation, they are limited for automatic network/topology detection and network induced covariance matrix estimation. To fill the gap, we propose a new network induced covariance estimation method (NICE) to simultaneously detect highly correlated networks and estimate the covariance matrix by leveraging an adaptive and graph topology oriented thresholding strategy. The novel thresholding strategy can reduce both false positive and false negative discovery rates by using the whole graph topological information which allows edges to borrow with each other. Moreover, we propose a novel `quality and quantity' objective function to shrink the covariance matrix towards a parsimonious model while retaining most of the information. Simulation study results show that our method outperform the competing thresholding and shrinkage methods. We further illustrate the application of our new method by analysis of a serum mass spectrometry proteomics data set.