Causal network inference by optimal causation entropy

The broad abundance of time series data, which is in sharp contrast to limited knowledge of the underlying network dynamic processes that produce such observations, calls for a rigorous and efficient method of causal network inference. Here we develop mathematical theory of causation entropy, an information-theoretic statistic designed for model-free causality inference. For stationary Markov processes, we prove that for a given node in the network, its causal parents form the minimal set of nodes that maximizes causation entropy, a result we refer to as the optimal causation entropy principle. Furthermore, this principle guides us in developing computational and data efficient algorithms for causal network inference based on a two-step discovery and removal algorithm for time series data for a network-coupled dynamical system. Validation in terms of analytical and numerical results for Gaussian processes on large random networks highlights that inference by our algorithm outperforms previous leading methods, including conditional Granger causality and transfer entropy. Interestingly, our numerical results suggest that the number of samples required for accurate inference depends strongly on network characteristics such as the density of links and information diffusion rate and not necessarily on the number of nodes.