Information Entropy-Based Penalty for PARAFAC Analysis of Resting EEG

PARAFAC analysis of time-varying EEG spectra data provides a parsimonious representation in terms of topographic, temporal and spectral signatures, which allows for the identification of functional neural networks. This networks work in a critical state where short and long range connections coexist. Although PARAFAC is unique through usual least squares estimation, the use of penalized least squares allows to incorporate this kind of knowledge into the decomposition. Here we propose the use of Information Entropy-based penalties for obtaining atoms with minimum spectral entropy. They offer sparse spectral signatures corresponding to networks oscillating in a well-defined (sharp) frequency band.