Graph quilting: graphical model selection from partially observed covariances

Abstract: Graphical model selection is a seemingly impossible task when many pairs of variables are never jointly observed; this requires inference of conditional dependencies with no observations of corresponding marginal dependencies. This under-explored statistical problem arises in neuroimaging, for example, when different partially overlapping subsets of neurons are recorded in non-simultaneous sessions. We call this statistical challenge the "Graph Quilting" problem. We study this problem in the context of sparse inverse covariance learning, and focus on Gaussian graphical models where we show that missing parts of the covariance matrix yields an unidentifiable precision matrix specifying the graph. Nonetheless, we show that, under mild conditions, it is possible to correctly identify edges connecting the observed pairs of nodes. Additionally, we show that we can recover a minimal superset of edges connecting variables that are never jointly observed. Thus, one can infer conditional relationships even when marginal relationships are unobserved, a surprising result! To accomplish this, we propose an $\ell_1$-regularized partially observed likelihood-based graph estimator and provide performance guarantees in population and in high-dimensional finite-sample settings. We illustrate our approach using synthetic data, as well as for learning functional neural connectivity from calcium imaging data.