Coherence Resonance in Influencer Networks

The study addresses why many real-world networks evolve to contain highly connected hubs (influencers) when such heterogeneity is known to hinder synchronization in deterministic settings. Influencers couple more strongly to the network than typical nodes and often mediate function, yet their presence can raise the coupling required for synchronization and sometimes prevent it entirely. The authors investigate the effect of stochastic forcing on influencers in networks of coupled phase oscillators. They propose and test the hypothesis that appropriately tuned noise in influencer nodes can counteract desynchronizing structural heterogeneity and induce optimal collective synchronization across the network. The work situates this hypothesis within the broader context that small amounts of noise can enhance system response in complex systems and information processing, raising the question of how noise interacts with network structure in heterogeneous (influencer) networks.

Prior work shows hubs strongly influence dynamics and structure across domains (neuroscience, social/political networks, scale-free technological networks). Deterministic analyses indicate influencer (heterogeneous) networks are harder to synchronize than homogeneous ones, sometimes making synchrony unattainable. Constructive roles of noise are well-documented: stochastic resonance, coherence resonance, and noise-induced synchronization in excitable systems, homogeneous networks, near Hopf bifurcations, and in small oscillator systems. However, how coherence resonance manifests in heterogeneous influencer networks remained unclear. The authors build on the Kuramoto–Sakaguchi framework for weakly coupled oscillators, the Ott–Antonsen reduction for heterogeneous frequency distributions, and slow–fast/adiabatic theories to bridge microscopic interactions and macroscopic order parameters in networks with structural heterogeneity.

Model: A network of phase oscillators with Kuramoto–Sakaguchi coupling g(θm−θn)=sin(θm−θn)+c0 (equivalently sin(θm−θn−α) after shifting frequencies), subject to independent Gaussian white noise. The phase dynamics for node n: θ̇n=ωn+λn Σm Wnm g(θm−θn)+√(2Dn) ξn. Influencers have stronger coupling intensity λn=βλ0; followers have λn=λ0. Influencer noise strength is D (identical across influencers), followers have noise D0 and a Lorentzian frequency distribution for ωi with mean ω0 and width γ0. The influencer mean frequency is ω, with gap Δω=ω−ω0. Network structure: influencer networks where followers predominantly connect to a small set of influencer hubs; followers are partitioned into sets Pσ based on which influencers they connect to. Order parameters: For each partition σ, define Zσ(t)=(1/|Pσ|) Σi∈Pσ e^{iθi}, with Rσ=|Zσ|; global order parameter Z and R defined across all followers. Dimensionless reduction: Time rescaled by 1/λ0; frequency shift from shear c0 absorbed into ω; move to a co-rotating frame where the followers’ mean frequency is zero. Effective parameters (Table 1): ΔΩ/λ0=dynamical frequency gap; Λ=βλ0/ΔΩ (dimensionless coupling of influencers); influencer effective noise q=D/ΔΩ; followers’ effective noise D0/λ0; followers’ heterogeneity γ0/λ0. Regime of interest: ΔΩ/λ0>1 (slow–fast separation), Λ<1, q=O(1), and small D0/λ0, γ0/λ0. Mean-field and reductions: Use partition mean-fields to capture collective dynamics. For followers with Cauchy-distributed frequencies and small noise, apply the Ott–Antonsen (OA) approach to obtain a low-dimensional complex Riccati equation for Zσ in the thermodynamic limit. Develop a slow–fast/adiabatic theory exploiting ΔΩ/λ0≫1: influencers evolve fast relative to followers; replace fast influencer phases by their expected values Gk given quasi-static follower fields, yielding an averaged interaction H that drives the slow partition mean-fields. This results in an effective hyper-graph where partitions are nodes and influencers mediate higher-order coupling via functions Gk(A,q). An explicit expression for Gk is obtained from the stationary distribution of the noisy Adler equation (Fokker–Planck approach), expressible via continued fractions or confluent hypergeometric limit functions. Under symmetry (identical influencers and partitions), derive an amplitude equation for the global order parameter R with growth term Re[G(R;A,q)]/(1−R^2), enabling analytic prediction of R versus q and parameter thresholds when the synchronization manifold is stable. Simulations: Numerical integration of the stochastic phase equations in the dimensionless canonical form using Euler–Maruyama with small timestep dt≈1e−4 to resolve the fast influencer timescale. Network cases: (i) Toy influencer network with two influencers and three follower partitions (each 300 followers), no follower–follower or influencer–influencer links; (ii) Undirected scale-free network (largest connected component; top-5 degree nodes as influencers); (iii) Directed C. elegans neuronal network (N=268 after removing zero in-degree nodes; top-K=15 out-degree nodes as influencers). Connectivity weights: Wmn=1 for edges incident to influencers; Wmn=0.01 for follower–follower edges. Parameters: Example settings include β=10 with ΔΩ/λ0=18 (A=10/18) and β=100 with ΔΩ/λ0=198 (A=100/198); followers typically identical (γ0/λ0=0) with small noise D0/λ0=0.02; in theory curves, heterogeneity γ0/λ0=0.02 is considered; q is varied by changing influencer noise D for fixed ΔΩ. For each q, long simulations (T≈10^4 in dimensionless time) produce histograms/densities of R to estimate expected order parameter and its distribution.

• Coherence resonance in influencer networks: Varying the influencer effective noise q=D/ΔΩ reveals a nonmonotonic response of network synchrony. The global order parameter R increases from low values at small q, reaches a maximum near q≈1, and decreases for large q.
• Two-step mechanism: (1) The network acts as a nonlinear filter of stochastic influencer dynamics; at optimal q, influencers induce local synchronization in directly connected follower partitions. (2) These synchronized partitions interact indirectly via influencers, giving rise to macroscopic collective oscillations describable by an emergent hyper-graph coupling.
• Optimal noise scale: Analytical slow–fast mean-field theory predicts maximal synchronization when q≈1 (i.e., D≈ΔΩ) and A close to but below 1, consistent with simulations across toy, scale-free, and C. elegans networks. The resonance persists for both moderate (ΔΩ/λ0=18) and large (ΔΩ/λ0=198) frequency gaps.
• Threshold conditions: The coherence resonance effect requires very small follower heterogeneity/noise; above a Λ-dependent threshold in γ0/λ0 or D0/λ0, the effect disappears. Frequency heterogeneity and follower noise have qualitatively and quantitatively similar desynchronizing impacts.
• Theoretical–numerical agreement: Analytical predictions for R(q) in the thermodynamic limit (with OA-based reduction and adiabatic averaging) match finite-size simulations, particularly evident in Fig. 1 where the solid curve closely tracks simulated R-density maxima.
• Loss of synchrony at extremes: When influencer noise is too weak (q≪1) or too strong (q≫1), effective coupling vanishes due to bounded amplitude of stochastic drives and timescale separation, leading to reduced R.
• Realistic networks: In scale-free and C. elegans networks, time series of R show low coherence at small q, strong coherent oscillations (R near 1) at q≈1, and reduced coherence at large q, demonstrating robustness beyond idealized topologies.

The findings resolve the paradox of influencers hindering synchronization by showing that stochastic forcing of influencers can render them facilitators of global synchrony. An optimal noise level compensates for structural heterogeneity that would preclude synchronization deterministically, thereby addressing the central question of why influencer architectures may be beneficial dynamically. The mechanism differs from classical coherence resonance in excitable homogeneous systems, where noise merely excites oscillations; here, noise and heterogeneity synergize to create effective macroscopic interactions across follower partitions via influencers, forming an emergent hyper-network of higher-order interactions. This reconfiguration at the macroscopic level underscores how network structure and stochasticity can jointly reshape collective dynamics. Practically, the work suggests that tuning noise on a small subset of key nodes can control emergent order in complex oscillator systems.

The study introduces coherence resonance in influencer networks: optimally tuned noise applied to highly connected influencer nodes maximizes global synchronization that otherwise would be unattainable in deterministic heterogeneous networks. Through a combination of slow–fast averaging, OA-based mean-field reduction, and simulations on idealized and real-world networks, the authors show that influencers mediate higher-order interactions among follower partitions, yielding a hyper-graph description and a resonance peak in synchrony at q≈1. These results highlight a control paradigm where regulating stochasticity in a few hubs can enhance collective function. Potential future directions include extending the theory to directed, repulsive/attractive, and intra/inter-partition couplings, allowing heterogeneous influencers, characterizing stability of the synchronization manifold under broader conditions, and exploring applications to empirical systems where hub noise can be modulated.

• Regime constraints: The effect requires a large dynamical frequency gap (ΔΩ/λ0≫1) to justify slow–fast separation and adiabatic averaging, and small follower heterogeneity/noise (γ0/λ0, D0/λ0 ≪ 1); above a Λ-dependent threshold in follower heterogeneity/noise, coherence resonance vanishes.
• Modeling assumptions: Influencers are often treated as identical (same β, ω, D), followers’ frequencies follow a Lorentzian distribution, and initial analyses consider simplified influencer–follower-only coupling (limited or no strong follower–follower or influencer–influencer links). Weight schemes (W=1 to/from influencers; W=0.01 otherwise) are stylized.
• Analytical reductions: The OA manifold is exact for Cauchy frequency distributions without white Gaussian noise; with small Gaussian noise, invariance is broken and corrections are approximate. Mean-field and thermodynamic-limit predictions may deviate for small partitions or strong finite-size fluctuations.
• Parameter sensitivity: Optimal synchronization occurs near q≈1 and A close to but below 1; performance may degrade away from this regime, and effective coupling disappears at very low or very high q.
• Generality: While demonstrated on toy, scale-free, and C. elegans networks, broader topologies and more complex directional/weighted structures require further validation and may exhibit additional dynamical phenomena.