PaperlayerView exact revision on arXiv

Source-linked AI summary

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

James P. Gleeson, Sergey Melnik, Jonathan A. Ward, Mason A. Porter, Peter J. Mucha

arXiv:1011.3710v3physics.soc-phcond-mat.stat-mechcs.SI

TL;DR

Mean-field predictions are widely used, but their accuracy on real-world networks with clustering and modular structure has not been comprehensively explained. The paper compares mean-field theory with simulations across 21 networks and finds accuracy tracks mean first-neighbor degree as well as mean degree. It concludes that low mean degree need not prevent accurate predictions in the studied settings.

  • Problem

    Whether mean-field theory predicts dynamics accurately on real-world networks remains incompletely explained, despite its common use and the frequent violation of its assumptions by clustering and modularity.

    Source: §I · p. 1

  • Method

    The paper compares mean-field predictions with numerical simulations for multiple dynamical processes on 21 undirected, unweighted real-world networks.

    Source: §I · p. 2

  • Results

    Across networks, low d and low z coincide with inaccurate correlated mean-field predictions, while high-d cases are accurate; at fixed z=4.8, error is highest at the lowest d.

    Source: A. Mean-Field Accuracy for the SIS Model · p. 4, B. Isolating the effect of dynamical… · p. 6

  • Takeaways & Limitations

    A large mean degree is not necessary for accurate mean-field predictions, at least for ensemble-averaged quantities in the studied settings.

    Source: §V · pp. 9–10

    The reported error measures focus on behavior far from critical points, so they do not establish accuracy for calculating phase-transition points.

    Source: §V · p. 9

Abstract

from arXiv · show

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.

I. INTRODUCTION

Mean-field theory is widely used, although real-world clustering and modularity violate assumptions behind it. The paper tests how well it predicts network dynamics and identifies mean first-neighbor degree as a better accuracy indicator than mean degree alone.

I. INTRODUCTION

  • Mean-field theory remains widely used despite assumptions that are often violated in networks with clustering and modular structure.The paper notes that prior work had observed reasonable accuracy but had not comprehensively explained it.Source: §I · p. 1
  • Mean-field assumptions of independent neighbor states, non-modular degree classes, and absent dynamical correlations can fail in clustered or modular real-world networks.Pair approximations can better account for dynamical correlations, but are more complicated, so the paper mostly studies mean-field theory.Source: §I · p. 1
  • The study tests multiple dynamical processes on 21 real-world networks and finds that mean-field accuracy typically improves when mean first-neighbor degree d is sufficiently large.Mean degree z alone is not necessarily a good indicator of accuracy.Source: §I · p. 2

II. EXAMPLES

The section opens by contrasting examples where mean-field theory accurately predicts real-world network dynamics with examples where it performs poorly.

II. EXAMPLES

  • The section frames its examples as a contrast between accurate and poor mean-field predictions for dynamics on real-world networks.Source: §II · p. 2

A. Kuramoto Phase Oscillator Model

The section compares Kuramoto synchronization simulations with mean-field predictions on the Facebook Oklahoma and US Power Grid networks. Agreement differs sharply between these examples, and mean degree alone does not fully explain the broader accuracy question.

A. Kuramoto Phase Oscillator Model

  • The Kuramoto model assigns oscillators unit-variance Gaussian intrinsic frequencies, couples them through network links with strength K, and measures synchrony using an order parameter.Synchrony is expected to emerge for sufficiently large K.Source: A. Kuramoto Phase Oscillator Model · p. 2
  • Mean-field predictions agree very well with Kuramoto simulations on Facebook Oklahoma but very poorly on the US Power Grid.Source: A. Kuramoto Phase Oscillator Model · p. 2
  • Facebook Oklahoma has mean degree z=102, while the Power Grid has z≈2.67, but mean degree alone does not capture all the subtleties of mean-field accuracy.The text presents the idea that high-degree nodes experience something closer to a mean field as a plausible but incomplete explanation.Source: A. Kuramoto Phase Oscillator Model · p. 2

B. SIS Epidemic Model

The section compares mean-field predictions for the SIS epidemic model with simulations, using theories that either retain or ignore degree-degree correlations. Accuracy differs between the two example networks, including a low-mean-degree case.

B. SIS Epidemic Model

  • The correlated and uncorrelated mean-field curves are indistinguishable for the Electronic Circuit (s838) network in Fig. 2(b).Source: Figure 2 · p. 3
  • The section compares steady-state infected fractions against spreading rate β using correlated and uncorrelated mean-field predictions for the SIS process.Correlated mean-field theory retains degree-degree correlations through P(k′|k); uncorrelated theory replaces it with k′pk′/z.Source: B. SIS Epidemic Model · pp. 2–3
  • On the AS Internet, mean-field theory is very accurate despite mean degree 4, whereas it performs poorly on the Electronic Circuit (s838) network.The AS Internet result is notable although 96% of its nodes have degree 10 or less.Source: B. SIS Epidemic Model · p. 3, Figure 2 · p. 3

C. Voter Model

Voter-model survival probabilities show that mean-field accuracy depends on network structure: it is strong on the C. elegans network but poor on a highly clustered network and its degree-preserving rewiring.

C. Voter Model

  • Mean-field theory is very accurate for voter-model survival on C. elegans, but very inaccurate on the low-degree synthetic clustered network.The two networks each contain approximately 300 nodes.Source: Figure 3 · p. 3, C. Voter Model · p. 3
  • The comparison measures the fraction of voter-model realizations still disordered over time against mean-field and pair-approximation theory curves.The voter-model figure compares the C. elegans network with the synthetic clustered network.Source: C. Voter Model · p. 3, Figure 3 · p. 3
  • Rewiring the clustered network into a random 3-regular graph leaves the mean-field match poor, while pair approximation matches the rewired network well.Pair approximation does not match the original clustered network well.Source: C. Voter Model · p. 4

III. WHY IS MEAN-FIELD THEORY ACCURATE?

Across the examples discussed, mean-field theory matches some high-mean-degree networks well and can also be accurate on low-mean-degree networks. The section broadens those comparisons across three dynamics and six networks.

III. WHY IS MEAN-FIELD THEORY ACCURATE?

  • The examples include both high-mean-degree networks with good mean-field matches and low-mean-degree networks with accurate mean-field predictions.Source: §III · p. 4
  • The authors describe these examples as representative of behavior across different dynamical processes and a variety of real-world networks.Source: §III · p. 4
  • Figure 4 extends the comparisons to the Kuramoto, SIS, and voter models on each of the six networks used in Figures 1–3.Source: §III · p. 4

A. Mean-Field Accuracy for the SIS Model

Mean-field accuracy for SIS dynamics is not explained by mean degree alone: the mean degree of first neighbors is a useful predictor. The authors connect this pattern to neglected dynamical correlations and test it across real-world networks.

A. Mean-Field Accuracy for the SIS Model

B. Isolating the effect of dynamical correlations using synthetic networks

Synthetic-network tests show that, at fixed mean degree, cMF accuracy tracks mean first-neighbor degree: higher d corresponds to lower error. Tests isolating dynamical correlations support their dominant role in this pattern, while leaving the strongest source of error in real-world networks unresolved.

B. Isolating the effect of dynamical correlations using synthetic networks

  • Networks with negligible transitivity and community structure isolate violations of mean-field assumption (iii), since assumptions (i) and (ii) hold in that test.These SIS networks have N=10^4 nodes.Source: B. Isolating the effect of dynamical… · p. 6
  • The two synthetic families vary degree correlations and degree heterogeneity, allowing d to change while z remains fixed.For the uncorrelated family, d increases with the degree distribution's second moment at fixed z.Source: B. Isolating the effect of dynamical… · pp. 6–7
  • With z fixed at 4.8, cMF error is highest for fully assortative r=1 networks (lowest d) and falls for disassortative r=−1 networks.The error trends for degree-3 and degree-12 nodes also support lower error with higher-degree neighbors.Source: Figure 6 · p. 6, B. Isolating the effect of dynamical… · pp. 6–7
  • The relative error in the SI-edge fraction increases as d decreases, paralleling SIS infection-fraction error and supporting a dominant role for neglected dynamical correlations.The SI-edge comparison tests the independence assumption using simulated infected-node fractions.Source: Figure 8 · p. 7, B. Isolating the effect of dynamical… · p. 7
  • For degree-3 nodes, variation in the distribution of infection fractions across nodes is not the main source of the high-d accuracy effect.The standard deviation changes little with r, and replacing node values with their class mean leaves the Fig. 6 errors unchanged.Source: Figure 7 · p. 6, B. Isolating the effect of dynamical… · p. 7
  • For this network, SIS-model error magnitude |ES| increases monotonically with α, from 0.07 at α=0 (d=11) to 0.16 at α=1 (d=5), as d decreases linearly with α.Source: B. Isolating the effect of dynamical… · p. 7
  • In real-world networks, d may affect all three mean-field assumptions, and further work is needed to identify which assumptions most affect accuracy.The synthetic-network evidence isolates assumption (iii), but does not determine the strongest source of error in real-world settings.Source: B. Isolating the effect of dynamical… · pp. 7–8

C. Mean-Field Accuracy for Other Dynamical Processes

The section compares mean-field accuracy for Kuramoto and voter dynamics using errors over the (z, d) plane. Higher d corresponds to lower Kuramoto error, but not similarly to voter mean-field error; pair approximation improves voter survival-probability predictions.

C. Mean-Field Accuracy for Other Dynamical Processes

  • Kuramoto mean-field error falls with higher d, whereas the same high-d pattern does not appear for voter mean-field survival probability.The comparison concerns different predicted quantities: Kuramoto's order parameter versus voter-model survival probability.Source: C. Mean-Field Accuracy for Other… · p. 8
  • Voter mean-field theory predicts the ensemble-averaged fraction in either state, while survival probability measures ordering in individual finite-network realizations.Although simulations average to a voter-state fraction of 1/2, individual realizations eventually reach consensus.Source: C. Mean-Field Accuracy for Other… · pp. 8–9
  • Only networks with both low z and low d have relative errors significantly above 5% across SIS, Kuramoto, and voter pair approximation; higher d does not similarly improve voter mean-field error.The figure compares Kuramoto mean-field error, voter mean-field error, and voter pair-approximation error over the (z, d) plane.Source: C. Mean-Field Accuracy for Other… · pp. 8–9, Figure 9 · p. 8
  • For voter survival probability, pair approximation of dynamical correlations improves accuracy over mean-field, particularly on high-d networks; only networks with both low z and low d have relative errors that significantly exceed 5%.Source: C. Mean-Field Accuracy for Other… · p. 9

IV. COMPARISON WITH AN ALTERNATIVE MEASURE

The paper compares the first-neighbor degree d with q, a distance-based measure associated with bond-percolation pair-approximation error. Networks with higher d have lower q and lower reported bond-percolation PA error, while d is simpler to calculate and interpret.

IV. COMPARISON WITH AN ALTERNATIVE MEASURE

  • q is an alternative error measure based on mean inter-vertex distances in the clustered network and a rewired version that reduces clustering while preserving degree-degree correlations.The comparison also uses the network's mean degree z.Source: §IV · p. 9
  • High-d networks have low q, whereas low-d networks have high q and larger bond-percolation PA errors.The relationship is shown for the real-world networks examined in Ref..Source: Figure 10 · p. 9, §IV · p. 9
  • The first-neighbor degree d is appealing because it is simple to calculate and helps explain the underlying causes of mean-field inaccuracies.The paper notes that more complicated diagnostics may also help predict accuracy for particular dynamics.Source: §IV · p. 9

V. CONCLUSIONS

Mean-field theory works best when low-degree nodes connect to high-degree nodes, or when network mean degree is high; a large mean degree is not required for good ensemble-averaged results. The conclusions also suggest a similar condition for pair-approximation accuracy, while leaving phase-transition accuracy for future study.

V. CONCLUSIONS

  • Mean-field theory works best when low-degree nodes connect to high-degree nodes, including in disassortative or high-mean-degree networks.A large mean degree is not necessary for good mean-field accuracy on ensemble-averaged quantities.Source: §V · p. 9
  • The reported error measures address behavior far from critical points; mean-field accuracy for calculating phase-transition points remains future work.Source: §V · p. 9
  • Pair-approximation theories may follow a similar pattern, but remain vulnerable to clustering and modularity when d is low.The paper gives an example where PA works well only on a rewired, unclustered version of a low-d network.Source: §V · pp. 9–10
Loading 1011.3710v3…