Detecting causality using symmetry transformations

Detecting causality between variables in a time series is a challenge, particularly when the relationship is nonlinear and the dataset is noisy. Here, we present a novel tool for detecting causality that leverages the properties of symmetry transformations. The aim is to develop an algorithm with the potential to detect both unidirectional and bidirectional coupling for nonlinear systems in the presence of significant sampling noise. Most of the existing tools for detecting causality can make determinations of directionality, but those determinations are relatively fragile in the presence of noise. The novel algorithm developed in the present study is robust and very conservative in that it reliably detects causal structure with a very low rate of error even in the presence of high sampling noise. We demonstrate the performance of our algorithm and compare it with two popular model-free methods, namely transfer entropy and convergent cross map. This first implementation of the method of symmetry transformations is limited in that it applies only to first-order autonomous systems.