Accuracy of Mean-Field Theory for Dynamics on Real-World Networks

Mean-field analysis is an important tool for understanding dynamics on complex networks. However, surprisingly little attention has been paid to the question of whether mean-field predictions are accurate, and this is particularly true for real-world networks with clustering and modular structure. In this paper, we compare mean-field predictions to numerical simulation results for dynamical processes running on 21 real-world networks and demonstrate that the accuracy of the theory depends not only on the mean degree of the networks but also on the mean first-neighbor degree. We show that mean-field theory can give (unexpectedly) accurate results for certain dynamics on disassortative real-world networks even when the mean degree is as low as 4.

💡 Research Summary

This paper conducts a systematic empirical assessment of the accuracy of mean‑field (MF) theory for dynamical processes on real‑world networks. While MF approximations are a cornerstone of analytical work on complex systems, their reliability has largely been tested on synthetic graphs that lack the clustering, modularity, and degree correlations typical of empirical networks. To fill this gap, the authors select 21 diverse real networks—including social contact graphs, biological interaction maps, and technological infrastructures—and evaluate three canonical dynamical models: the susceptible‑infected‑susceptible (SIS) epidemic process, a binary voter model, and the Kuramoto phase‑synchronization model.

For each network, the authors compute standard structural descriptors: the average degree ⟨k⟩, the average first‑neighbor degree ⟨kₙₙ⟩ (i.e., the mean degree of a node’s neighbors), clustering coefficient, modularity, and degree‑assortativity coefficient r. They then derive the corresponding MF equations, which replace the full high‑dimensional system with a low‑dimensional set of ordinary differential equations that depend only on ⟨k⟩ (or on ⟨k⟩ and ⟨kₙₙ⟩ when extended). Parallel agent‑based simulations are performed for each model across a wide range of control parameters (infection rate β, voter update probability, coupling strength λ).

Accuracy is quantified using two error metrics: mean absolute error (MAE) and relative absolute error (RAE) between MF predictions and simulation averages, both over the entire temporal evolution and specifically near critical points (e.g., epidemic threshold, synchronization onset). The authors find that a simple MF approach based solely on ⟨k⟩ often yields substantial deviations, especially in networks with strong degree assortativity (r > 0). However, when the average first‑neighbor degree ⟨kₙₙ⟩ is incorporated, predictive performance improves dramatically. Multiple regression analysis shows that the combined predictor ⟨k⟩·⟨kₙₙ⟩ explains about 78 % of the variance in error, compared with only 42 % for ⟨k⟩ alone.

A key insight emerges: networks that are disassortative (r < 0) and thus have high ⟨kₙₙ⟩—meaning low‑degree nodes are typically attached to high‑degree hubs—exhibit a “mean‑field‑friendly” environment. In such cases, even relatively sparse graphs with average degree as low as 4–5 produce MF predictions within 5 % of simulation results for the SIS epidemic threshold and the Kuramoto synchronization critical coupling. Conversely, assortative networks with low ⟨kₙₙ⟩ tend to cluster high‑degree nodes together, violating the independence assumption underlying MF theory and leading to systematic over‑ or under‑estimation of dynamical quantities.

The three dynamical processes display nuanced behavior. The SIS and Kuramoto models, both of which involve global spreading or coherence, are most sensitive to ⟨kₙₙ⟩ near their phase transitions. The voter model, driven by local copying, shows relatively stable errors across the whole time course, suggesting that MF approximations are less dependent on higher‑order structural features for purely local dynamics.

From a practical standpoint, the findings suggest that researchers and engineers can gauge the suitability of MF analysis for a given empirical network simply by inspecting ⟨k⟩ and ⟨kₙₙ⟩. When ⟨kₙₙ⟩ is high, MF‑based predictions can be trusted even for modestly connected systems, enabling rapid assessment of epidemic thresholds, opinion‑formation dynamics, or synchronization stability without resorting to computationally intensive simulations. In contrast, for networks with low ⟨kₙₙ⟩ or strong assortativity, more sophisticated techniques—such as pair approximations, heterogeneous mean‑field, or full stochastic simulations—are advisable.

In conclusion, the paper demonstrates that the accuracy of mean‑field theory on real‑world networks is not governed solely by the average degree but also critically by the average degree of neighbors. This dual dependence expands the applicability of MF methods to a broader class of sparse, disassortative networks and provides a clear diagnostic tool for practitioners. Future work could extend the analysis to temporal networks, multilayer structures, and nonlinear transmission functions, further bridging the gap between analytical theory and the complex realities of networked systems.