Characterizing Multivariate Information Flows

One of the crucial steps in scientific studies is to specify de pen-dent relationships among factors in a system of interest. Givenlittle knowledge of a system, can we characterize the underlyingdependent relationships through observation of its temporal be-haviors? In multivariate systems, there are potentially many pos-sible dependentstructures confusable with each other, and it maycause false detection of illusory dependency between unrelatedfactors. The present study proposes a new information-theoreticmeasure with consideration to such potential multivariate rela-tionships. The proposed measure, called multivariate transferentropy, is an extension of transfer entropy, a measure of tem-poral predictability. In the simulations and empirical studies, wedemonstrated that the proposedmeasure characterized the latentdependentrelationships in unknowndynamical systems more ac-curately than its alternative measure.