Network Inference and Maximum Entropy Estimation on Information Diagrams

摘要: Maximum entropy estimation is of broad interest for inferring properties systems across many disciplines. Using a recently introduced technique estimating the maximum set random discrete variables when conditioning on bivariate mutual informations and univariate entropies, we show how this can be used to estimate direct network connectivity between interacting units from observed activity. As generic example, consider phase oscillators that our approach typically superior simply using information. In addition, propose nonparametric formulation connected informations, test explanatory power description in general. We give an illustrative example showing agrees with existing parametric formulation, demonstrate its applicability advantages resting-state human brain networks, which also discuss effective connectivity. Finally, generalize continuous vastly expand types information-theoretic quantities one condition on. This allows us establish significant over ones. Not only does method perform favorably undersampled regime, where methods fail, but it dramatically less computationally expensive as cardinality increases.