Compressing Complexity: A Critical Synthesis of Structural, Analytical, and Data-Driven Dimensionality Reduction in Dynamical Networks

TL;DR

This paper reviews methods for reducing high-dimensional dynamical networks into simpler models, categorizing approaches into structural, analytical, and data-driven techniques, and discusses their trade-offs and future directions.

Contribution

It provides a comprehensive classification and synthesis of existing dimensionality reduction methods for dynamical networks, highlighting their principles, limitations, and emerging hybrid approaches.

Findings

Structural coarse-graining contracts network graphs using spectral methods.

Analytical reduction employs ansatzes and moment closures for differential equations.

Data-driven methods utilize manifold learning and operator theory to infer dynamics.

Abstract

The contemporary scientific landscape is characterized by a "curse of dimensionality," where our capacity to collect high-dimensional network data frequently outstrips our ability to computationally simulate or intuitively comprehend the underlying dynamics. This review provides a comprehensive synthesis of the methodologies developed to resolve this paradox by extracting low-dimensional "macroscopic theories" from complex systems. We classify these approaches into three distinct methodological lineages: Structural Coarse-Graining, which utilizes spectral and topological renormalization to physically contract the network graph; Analytical-Based Reduction, which employs rigorous ansatzes (such as Watanabe-Strogatz and Ott-Antonsen) and moment closures to derive reduced differential equations ; and Data-Driven Reduction, which leverages manifold learning and operator-theoretic frameworks (e.g., Koopman analysis) to infer latent dynamics from observational trajectories. We posit that the selection of a reduction strategy is governed by a fundamental "No Free Lunch" theorem, establishing a Pareto frontier between computational tractability and physical fidelity. Furthermore, we identify a growing epistemological schism between equation-based derivations that preserve causal mechanisms and black-box inference that prioritizes prediction. We conclude by discussing emerging frontiers, specifically the necessity of Higher-Order Laplacian Renormalization for simplicial complexes and the development of hybrid "Scientific Machine Learning" architectures-such as Neural ODEs-that fuse analytical priors with deep learning to solve the closure problem.