From Spatial to Spectral: Network Renormalization via Dynamical Correlations

TL;DR

This work replaces adjacency-based coarse-graining with spectral-space renormalization, using diffusion on the Laplacian to drive RG-like transformations. By recasting the Gaussian model in Laplacian eigenspace, the authors derive scaling exponents $y_t=d_w$ and $y_h=(d_f+d_w)/2$ and show that diffusion constitutes an RG step, with all critical exponents expressible via the fractal and random-walk dimensions $d_f$ and $d_w$ alongside the spectral dimension $d_s$; the Alexander–Orbach relation $d_s/2 = d_f/d_w$ connects dynamics to topology. They then introduce a meta-graph reconstruction algorithm that maps dimensionally reduced spectral information back to explicit topology, yielding renormalized networks that preserve dynamical correlations and exhibit long-range meta-connections. Across Internet, yeast regulatory, and European power-grid networks, the method yields internally consistent dimensions ($d_f$, $d_s$, $d_w$) and reveals multiscaling and hidden vulnerability pathways, with practical implications for resilience and infrastructure risk assessment. The framework thus provides a unified language uniting structure and dynamics in complex networks and suggests extensions to time-varying, multilayer, and weighted systems for more realistic modeling of complex infrastructures and biological networks.

Abstract

Network renormalization has traditionally relied on spatial adjacency-grouping nearby nodes together, but this approach fails to capture the dynamical correlations that govern system-wide behavior in scale-free networks. We present a spectral-space renormalization framework that enables coarse-graining based on dynamical coherence rather than geometric proximity. Within this framework, diffusion processes naturally constitute renormalization transformations in spectral space, yielding scaling relations that connect network dimensions with critical exponents. Building on this foundation, we develop a meta-graph reconstruction algorithm that systematically maps spectral information back into explicit topology while preserving dynamical correlations. The resulting renormalized networks uncover organizational structures that remain invisible to adjacency-based methods, including long-range correlations between structurally distant nodes that reflect coherent dynamical responses. Applications to Internet topologies, yeast regulatory networks, and European power grids demonstrate the broad applicability of this framework. The algorithm consistently extracts fractal, spectral, and random-walk dimensions with theoretical consistency across diverse systems. In power grids, it further reveals hidden failure pathways, exposing transcontinental correlations that match documented cascade patterns. In Internet networks, it reveals multiscaling behavior as the topology evolves over time. By shifting network renormalization from spatial geometry to dynamical flow, this work provides a unified foundation for understanding how information, energy, and failures propagate through complex systems, with direct implications for infrastructure resilience and network vulnerability assessment.

Figures (15)

• Figure 1: Scaling analysis of Internet topology. Application of our meta-graph algorithm to Internet AS-level topology (1998) reveals complete theoretical consistency across all fundamental network dimensions. (a) Network exhibits power-law degree distribution $P(k) \sim k^{-2.37}$ for $2 \leq k \leq 36$. (b) Spectral dimension from specific heat plateau: $C = d_s/2 = 1.965$ for $\tau \in [6, 24]$ yields $d_s = 3.93$. (c) Random-walk dimension from supernode size scaling: slope $1/d_w = 0.53$ gives $d_w = 1.89$. (d) Fractal dimension from supernode number scaling: $N_s'$ vs $\langle \ell_s \rangle$ yields $d_f = 3.77$. (e) Degree scaling validation: $\langle k_J' / k \rangle_J$ vs $\langle \ell_s \rangle$ gives $d_k^{(\ell)} = 2.79$. (f) Theoretical consistency: measured $d_k/d_f = 0.74$ matches prediction $1/(\gamma_d - 1) = 0.73$. All independently measured dimensions satisfy the Alexander--Orbach relation ($d_f/d_w = 1.99 \approx d_s/2 = 1.98$), confirming our comprehensive approach captures both structural and dynamical scaling within a single consistent framework. Additional supernode distributions in Supplementary Figs. \ref{['fig:internet_degree']}--\ref{['fig:internet_size']}.
• Figure 2: Multi-scaling behavior in evolving networks. Application of our meta-graph algorithm to Internet AS-level topology (1999) demonstrates crossover between distinct scaling regimes, revealing coexisting network substructures. (a) Degree distribution with $\gamma_d \approx 1.86$ for $k \in [4, 66]$ shows heavier tail compared to 1998 data. (b) Spectral dimension crossover: $C = d_s/2 = 2.33$ (short $\tau \in [6, 9]$) vs $2.08$ (long $\tau \in [9, 24]$). (c) Random-walk dimension regime change: $1/d_w = 0.82$ (short $\tau$) vs $0.40$ (long $\tau$) indicates different transport mechanisms. (d) Fractal dimension transition: $d_f \approx 2.82$ (short $\tau$) vs $5.08$ (long $\tau$) reflects structural reorganization. (e) Degree scaling consistency within regimes: $d_k^{(\ell)} \approx 3.28$ (short $\tau$) and $3.70$ (long $\tau$). (f) Theoretical validation: $d_k^{(\ell)}/d_f = 1.17 \approx d_k^{(N)} = 1.17$ matches $1/(\gamma_d - 1) = 1.17$. The Alexander--Orbach relation remains valid in both regimes despite different exponent values, confirming that multiple coexisting scaling behaviors challenge universal network scaling assumptions.
• Figure 3: Non-recursive nature of the meta-graph algorithm in deterministic networks. Analysis of a three-generation deterministic scale-free tree demonstrates path dependence and irreversibility inherent in spectral space transformations. (a) Diffusion time domains mapped to accessible renormalized topologies shown in panels (c)--(g). (b) Original network with hierarchical tree structure across three generations. (c)--(g) Renormalized meta-graphs at different diffusion times $\tau$ show emergence of new connections (thick red edges) and structural reorganization. Critically, specific configurations (e.g., panel f) become inaccessible from intermediate renormalized states (e.g., panel d), even though they remain directly reachable from the original network. This path dependence demonstrates the fundamental non-recursive and irreversible nature of the meta-graph algorithm, distinguishing it from conventional geometric coarse-graining methods.
• Figure 4: Hidden dynamical correlations. Meta-graph analysis of the European power grid reveals latent long-range correlations that match documented transcontinental cascade patterns. (a) Original geographical network topology showing only local physical connections. (b)--(f) Renormalized meta-graphs at increasing diffusion times $\tau$ consistently reveal new long-range correlations (thick red edges), most prominently between Denmark and Spain. These dynamically coherent connections, absent from the physical topology, reflect shared spectral signatures that indicate synchronized responses to network perturbations. The Denmark-Spain correlation matches empirically observed cascade pathways where faults originating in Greece propagated through Northern Europe to destabilize the Spanish grid, validating our framework's predictive capability for infrastructure vulnerability assessment.
• Figure S1: Workflow for spectral renormalization and meta-graph reconstruction. It combines $d_s$ estimation, spectral space coarse-graining, crossover detection, and consistency validation via critical exponents.