Nonlinear Network Reconstruction by Pairwise Time-delayed Transfer Entropy

Abstract: Analyzing network structural connectivity is crucial for understanding dynamics and functions of complex networks across disciplines. In many networks, structural connectivity is not observable, which requires to be inferred via causal inference methods. Among them, transfer entropy (TE) is one of the most broadly applied causality measure due to its model-free property. However, TE often faces the curse of dimensionality in high-dimensional probability estimation, and the relation between the inferred causal connectivity and the underlying structural connectivity remains poorly understood. Here we address these issues by proposing a pairwise time-delayed transfer entropy (PTD-TE) method. We theoretically establish a quadratic relationship between PTD-TE values and node coupling strengths, and demonstrate its immunity to dimensionality issues and broad applicability. Tests on biological neuronal networks, nonlinear physical systems, and electrophysiological data show PTD-TE achieves consistent, high-performance reconstructions. Compared to a bunch of existing approaches for network connectivity reconstruction, PTD-TE outperforms these methods across various network systems in accuracy and robustness against noise. Our framework provides a scalable, model-agnostic tool for structural connectivity inference in nonlinear real-world networks.