Overcoming the Curse of Dimensionality: Structural Connectivity Reconstruction via Pairwise Information Flow in Nonlinear Networks

A team led by Kai Chen (Shanghai Jiao Tong University, NYU Shanghai) published a preprint on arXiv introducing Pairwise Delayed Information Flow (PDIF), a new method to infer the structural connectivity of nonlinear networks from observed dynamics. The paper, "Overcoming the Curse of Dimensionality: Structural Connectivity Reconstruction via Pairwise Information Flow in Nonlinear Networks," addresses a fundamental open problem in complex systems: how to deduce who is connected to whom when only the noisy time-series of node activities are available. Existing approaches either require prior knowledge of the underlying equations (model-based) or suffer from the curse of dimensionality as the network size grows (model-free methods).

PDIF works by computing a time-delayed mutual information between pairs of nodes and then applying a theoretical quadratic relationship between this PDIF and the actual coupling strength. Crucially, the researchers demonstrated that indirect interactions (e.g., A→B→C) are suppressed at leading order, meaning simple pairwise measurements are sufficient to reconstruct the true graph without needing to condition on all other nodes—a first for nonlinear systems. They validated PDIF on three tiers: synthetic nonlinear oscillators, biologically realistic neuronal network models (with up to 10,000 neurons), and large-scale electrophysiological recordings from macaque visual cortex. In all cases, PDIF achieved higher reconstruction accuracy and noise robustness compared to standard methods like Granger causality, transfer entropy, and correlation-based approaches.