Abstract
Spectral networks derived from multivariate time series data arise in manydomains, from brain science to Earth science. Often, it is of interest to studyhow these networks change under different conditions. For instance, to betterunderstand epilepsy, it would be interesting to capture the changes in thebrain connectivity network as a patient experiences a seizure, usingelectroencephalography data. A common approach relies on estimating thenetworks in each condition and calculating their difference. Such estimates maybehave poorly in high dimensions as the networks themselves may not be sparsein structure while their difference may be. We build upon this observation todevelop an estimator of the difference in inverse spectral densities across twoconditions. Using an L1 penalty on the difference, consistency is establishedby only requiring the difference to be sparse. We illustrate the method onsynthetic data experiments, on experiments with electroencephalography data,and on experiments with optogentic stimulation and micro-electrocorticographydata.