Transformability of Higher-order Network Dynamics

Shibo He, Ming Xie, Aming Li, Zi‐Ke Zhang, Qiang Huang, Youxian Sun, Jiming ChenResearch Square 2025 · DOI

TL;DR

Researchers have developed a unified framework to bridge the gap between simple pairwise network models (like traditional graphs) and complex higher-order networks (like hypergraphs and simplicial complexes). By using information entropy to measure "system instability," the study demonstrates that higher-order dynamics are not independent phenomena but can be systematically transformed into simpler ones by accounting for two additive factors: dynamical nonlinearity and structural complexity.

What Problem Does This Paper Address?

In network science, we often model processes like the spread of a virus or the evolution of political opinions. Traditionally, these are modeled on pairwise networks, where interactions happen between two nodes at a time. However, real-world systems often involve group interactions—think of a group chat or a research collaboration—which are better represented by higher-order networks (hypergraphs and simplicial complexes).

While we know higher-order interactions produce unique behaviors (like sudden "explosive" transitions), we haven't had a clear way to relate these complex models back to simpler ones. This "representation gap" makes it difficult to know when we actually need a complex higher-order model and when a simpler graph-based model would suffice. The authors tackle this by asking: Can we mathematically transform dynamics across different network orders?

Methods

The authors propose a "Transformation System of Dynamics" based on three pillars:

  1. Network Projection: They define rigorous ways to project higher-order structures (hypergraphs) into lower-order ones (simplicial complexes and dyadic graphs) while preserving as much information as possible.
  2. Information Entropy as a Metric: To compare different models, they use information entropy ($\Phi$) to quantify "system instability." This provides a global, time-varying measure of how much a process (like an infection) has spread and how much uncertainty remains in the system.
  3. Factor Decomposition: The core of their framework is the discovery that the difference in instability between a higher-order model and a pairwise model ($\Delta\Phi^*$) can be decomposed into two additive components:
  4. Dynamical Factor ($\Delta\Phi_1$): Arising from the nonlinear nature of group interactions.
  5. Structural Factor ($\Delta\Phi_2$): Arising from the presence of multiple, overlapping contacts within those groups.

They validated this framework using Mean-Field Theory to derive master equations and used Support Vector Machines (SVM) to test if dynamics on different structures could be statistically distinguished.

Datasets

The study utilizes a mix of synthetic and empirical datasets to ensure robustness:

  • Empirical Social Networks: High-school social networks, primary school interactions, and the Email-Enron dataset.
  • Synthetic Networks: Generated using models like HyperCL (for scale-free hypergraphs), Barabási–Albert (BAH), and Erdős–Rényi (ERH) hypergraphs to test varying degrees of node heterogeneity.
  • Metadata-rich sets: The NDC-classes-full dataset was also used for broader validation.

Key Insights and Ideas

  • The Additivity Principle: Perhaps the most significant "aha" moment is that the total difference in system behavior is simply the sum of dynamical and structural differences ($\Delta\Phi^* = \Delta\Phi_1 + \Delta\Phi_2$). This suggests that complexity in networks is modular and predictable.
  • Dynamics are Interconnected: Higher-order dynamics aren't "new" physics; they are systematic extensions of pairwise dynamics. The paper provides specific "Transformable Conditions" (C1, C2, C3) under which a complex hypergraph model becomes mathematically equivalent to a simple graph model.
  • Universality Across Domains: The framework isn't just for disease spreading (Contagion Dynamics). The authors successfully applied the same logic to Opinion Dynamics (using the Deffuant-Weisbuch and Voter models), showing that the principles of transformability hold whether you are tracking a pathogen or a point of view.

Key Results

  • Theoretical Alignment: The proposed mathematical conditions for transformability were confirmed by simulations. When the conditions were met, the infection density curves ($I(t)$) for hypergraphs and their projected simple graphs overlapped almost perfectly.
  • Classification Accuracy: Using SVM, the researchers showed that as the "Transformable Conditions" were satisfied, the F1-score for distinguishing between higher-order and pairwise dynamics dropped significantly. In other words, the dynamics became indistinguishable.
  • Robustness: The framework remained accurate across various infection rates, recovery rates, and network sizes, though it required refinement (using "effective degree") to handle highly heterogeneous networks where some nodes are much more connected than others.

Limitations and Future Work

  • Highly Heterogeneous Networks: While the authors introduced "effective degree" to improve accuracy, extremely skewed networks still pose a challenge for simple mean-field approximations.
  • Static vs. Temporal: The current work focuses on static networks. However, real-world group interactions (like a meeting) are often fleeting. Extending this "transformability" to temporal networks is a logical next step.
  • Model Variety: While they covered SI, SIS, SIR, and common opinion models, other complex processes like "threshold models" (where you only change your mind if enough friends do) or "evolutionary games" remain to be fully integrated into this framework.

This research provides a much-needed "Rosetta Stone" for network science, allowing researchers to translate complex higher-order interactions into more manageable pairwise terms without losing sight of the underlying physics.

Summary updated

Abstract

<title>Abstract</title> Recent studies on higher-order network dynamics have primarily concentrated on intrinsic properties such as bistability and discontinuous phase transitions, while largely overlooking transformability— the ability of dynamics adaptation across different network structures, which fundamentally captures the essence of the dynamical processes. Here, we present a holistic analysis framework that elucidates the interplays between dynamics in pairwise and higher-order networks, demonstrating that these processes are not independent but can undergo systematic transformations. By focusing on contagion dynamics, we identify and quantify dynamical and structural factors that drive the transformability of higher-order network dynamics, uncovering a universal model for system instability governed by these factors. Furthermore, we validate the findings from contagion dynamics to opinion dynamics, showcasing its broad applicability across diverse dynamical processes. Our findings uncover the fundamental nature of dynamic transformability for shaping collective behaviors in higher-order networks, offering profound implications for modeling, predicting, and controlling dynamics in a wide range of real-world networked intelligent systems.
BibTeX
@inproceedings{he2025transformability,
  author = {Shibo He and Ming Xie and Aming Li and Zi‐Ke Zhang and Qiang Huang and Youxian Sun and Jiming Chen},
  title = {{Transformability of Higher-order Network Dynamics}},
  booktitle = {Research Square},
  year = {2025},
  doi = {10.21203/rs.3.rs-7125491/v1},
}

Loading related content...

Figure