We discuss a versatile method of constructing directed networks frommultivariate time series. While most common methods widely accepted at presentutilize the concept of cross correlation between pairs of time series, themethod presented here is based on the linear modeling technique in time seriesanalysis. Since linear models generally contain terms representing feedbackeffects of different time delays, constructed networks reveal the intrinsicdynamical nature of the system under consideration such as complicatedentanglement of different periodicities, which we referred to as "timestructure." The method enables us to construct networks even if a givenmultivariate time series do not have sufficiently large values of crosscorrelation, the case in which the approach using cross correlation is notapplicable. We explicitly show a simple example where the method of crosscorrelation cannot reproduce the relationship among multivariate time series.The method we propose is demonstrated for numerical data generated by a knownsystem and applied to two actual systems to see its effectiveness.
展开▼