Mathematical Frameworks for Network Dynamics: A Six-Pillar Survey for Analysis, Control, and Inference

Abstract

The study of dynamical processes on complex networks constitutes a foundational domain bridging applied mathematics, statistical physics, systems theory, and data science. Temporal evolution, not static topology, determines the controllability, stability, and inference limits of real-world systems, from epidemics and neural circuits to power grids and social media. However, the methodological landscape remains fragmented, with distinct communities advancing separate formalisms for spreading, control, inference, and design. This review presents a unifying six-pillar framework for the analysis of network dynamics: (i) spectral and structural foundations; (ii) deterministic mean-field reductions; (iii) control and observability theory; (iv) adaptive and temporal networks; (v) probabilistic inference and belief propagation; (vi) multilayer and interdependent systems. Within each pillar, we delineate conceptual motivations, canonical models, analytical methodologies, and open challenges. Our corpus, selected via a PRISMA-guided screening of 134 mathematically substantive works (1997–2024), is organized to emphasize internal logic and cross-pillar connectivity. By mapping the field onto a coherent methodological spine, this survey aims to equip theorists and practitioners with a transferable toolkit for interpreting, designing, and controlling dynamic behavior on networks.

1. Introduction

The study of dynamics on complex networks concerns the evolution of states attached to nodes, edges, or higher-order structures, where the network topology governs the coupling rules and propagation pathways. Such systems arise across a spectrum of domains, including synchronization in coupled oscillators, contagion in epidemiology, learning in neural and social structures, flow in transportation or supply systems, and control in cyber–physical infrastructures. While the diversity of specific models is immense, the field is increasingly unified by the shared use of rigorous mathematical formalisms that describe how structure shapes function in nontrivial and often emergent ways. This review is organized around six methodological pillars that collectively span the principal analytical and algorithmic frameworks employed to model, analyze, and control network-based dynamical systems: spectral and structural foundations, deterministic mean-field reductions, control, optimization, and observability, temporal and adaptive networks, probabilistic inference and belief propagation, and multilayer and interdependent systems. Each pillar corresponds to a coherent family of methods grounded in precise mathematical assumptions, and each is motivated by recurring functional demands such as ensuring stability, maximizing robustness, identifying latent structure, and minimizing control energy. Rather than catalog individual case studies or domains, this review abstracts and systematizes the toolkits that enable principled reasoning across these applications. In doing so, it aims to provide a self-contained mathematical resource that illuminates both the shared logical infrastructure and the distinctive technical challenges that arise in modeling dynamics on networks. This review catalogues and compares the mathematical tool sets required to analyze, control, and infer network dynamics across disciplines.

The contemporary scientific and technological landscape is characterized by an unprecedented degree of interconnectivity, wherein complex networks arise naturally across a wide array of domains, including structural and functional brain networks in neuroscience, power grids and communication backbones in infrastructure systems, contagion pathways in epidemiological surveillance, and interbank lending and systemic risk cascades in finance. In each of these settings, the underlying connectivity constrains and channels the temporal evolution of distributed processes, yet it is the dynamics themselves, not the static graph structure, that encode resilience to perturbation, controllability under limited input, and the identifiability of latent parameters. The mathematical study of networked dynamics thus provides a unifying language to interrogate diverse phenomena through questions of stability, propagation, optimization, and learning. As applications grow in scope and complexity, so too do the analytical challenges: the proliferation of large-scale multilayer and multiplex networks demands principled approaches to dimensionality reduction and interlayer coupling; adaptive feedback and coevolving topologies necessitate frameworks that account for endogenous structural change, and high-dimensional state spaces governed by nonlinear or stochastic rules require control and inference tools that scale without compromising rigor. Addressing these challenges demands a methodology-centred synthesis rather than a catalogue of isolated case studies.

Despite the maturation of the field of dynamical processes on networks, existing surveys tend to adopt either a topic-centric approach, in which specific phenomena such as epidemic spreading, synchronization, or consensus dynamics are treated in isolation, or a domain-centric perspective, wherein application areas such as biological systems, sociotechnical infrastructures, or economic interactions define the organizational schema. While both strategies offer valuable depth within narrow confines, they often obscure the underlying methodological continuities and contribute to a fragmented understanding of the mathematical tools that cut across disciplines. This compartmentalization inhibits both theoretical synthesis and the transfer of techniques across contexts with structurally analogous dynamics. Reviewer feedback on earlier drafts of this work correctly identified a similar risk in our initial exposition, noting that the manuscript at times resembled a thematically loose list of contributions. In response, we implement a method-centric reorganization in the present revision, structuring the survey around six foundational analytical families: spectral and linear operator methods, combinatorial and structural approaches, adaptive and heterogeneous dynamics, temporal and stochastic formulations, optimization and control frameworks, and inference and learning techniques. Each pillar is developed as an internally coherent narrative, anchored in core models and formal mechanisms, thereby transforming the review into a comparative study of methodologies rather than a thematic aggregation of case studies. This structure not only aligns with the multidisciplinary nature of the literature but also clarifies the transferability and relative strengths of competing mathematical paradigms.

This review delimits its scope to deterministic and stochastic mesoscopic models defined over network substrates, encompassing both node-centric and edge-centric formulations. The temporal window of included research spans the period 1997 through 2024, thereby capturing the emergence of modern spectral theory, adaptive control, and inference-driven methodologies within network science. The target audience comprises applied mathematicians, control theorists, and data scientists seeking a rigorous yet transferable understanding of the formal tools used to characterize, modulate, and reconstruct dynamical processes on graphs. The treatment presumes familiarity with linear algebra, dynamical systems theory, and elementary probability, but refrains from full formal derivations when these are available in canonical sources. Instead, proofs are outlined with emphasis on conceptual structure and modeling insight, and references to foundational texts are provided where appropriate. This level of abstraction permits a compact yet powerful synthesis of diverse results while preserving mathematical precision and inter-theoretical compatibility across disciplinary boundaries.

Figure 1 organizes the six methodological pillars along two conceptual axes: the level of description (from local to global) and the dynamical structure of the system (from static to evolving). This matrix reveals how each approach embeds distinct modeling assumptions: spectral methods emphasize local invariants of static graphs, while adaptive frameworks model time-varying topologies and feedback-driven interactions. Mean-field reductions abstract large-scale coherence, whereas control theory targets system-level influence through localized interventions. Inference techniques occupy the intermediate zone, bridging micro-level variability with emergent patterns, and multilayer models generalize all prior methods across interacting substrates. Beyond classification, the matrix encodes functional adjacencies that permit compositional workflows and spectral analyses guide reductions, which inform control, interact with adaptive dynamics, constrain inference, and extend naturally to multilayer systems.

This review makes four interlocking contributions. First, it constructs a unified taxonomy that anchors six hitherto scattered literatures, which range from spectral graph theory to probabilistic inference, within a shared methodological spine, thereby dissolving disciplinary boundaries and facilitating knowledge transfer across network science, control theory, and statistical physics. Second, each pillar is equipped with analytical summaries: we extract representative models, sketch derivations, and contrast formalisms where methodological choices yield divergent predictions or scaling laws. Third, we systematize open problems and benchmarking challenges, collating them at the close of each pillar to support theoretical exploration and empirical testing. Finally, we provide a bridging narrative, supported by schematic diagrams, that clarifies how problems migrate across methods—how spectral constraints shape mean-field limits, how control energy informs inference uncertainty, or how multilayer coupling feeds back into temporal adaptation—thereby illustrating how the six pillars form not a catalogue but an operational calculus for understanding, designing, and steering dynamical processes on complex networks.

The manuscript proceeds as follows. Section 2 outlines the methodological protocol and article selection workflow. Section 3, Section 4, Section 5, Section 6, Section 7 and Section 8 form the analytical core, each corresponding to one of the six methodological pillars and structured to include model introductions, canonical formulations, synthesis of mathematical techniques, and unresolved questions. Section 9 synthesizes the overarching insights, interdependencies, and open challenges that emerge across methodological categories. All references are listed in order of first appearance.

This review is not a meta-analysis and does not attempt a quantitative benchmarking of models beyond formal methodological metrics. Our focus is deliberately restricted to studies that deploy explicit mathematical formulations; purely empirical or descriptive works without formal dynamical scaffolding are excluded. The survey centers on mesoscopic modeling—where nodes and edges retain interpretive specificity—but occasionally references microscopic (agent-based) or macroscopic (PDE-based) paradigms when these serve to motivate or contextualize the mesoscopic framework under discussion.

In view of the growing structural complexity, temporal heterogeneity, and control demands in real-world networked systems, the need for a unified, mathematically rigorous, and cross-disciplinary toolkit is more urgent than ever. Rather than isolating individual applications or domains, this review consolidates theoretical techniques that transcend disciplinary boundaries, enabling precise characterizations of dynamical behavior, robustness criteria, and control feasibility. The subsequent sections develop each of the six methodological pillars in depth, presenting canonical models, representative derivations, and cross-comparative insights aimed at equipping both theorists and applied practitioners with a transferable analytical repertoire for dissecting, steering, and reconstructing complex network dynamics.

2. Methodology and Review Protocol

This systematic review follows the PRISMA 2020 guidelines, ensuring transparency and methodological rigor in the selection, assessment, and inclusion of scholarly works on dynamic processes over complex networks.

2.1. Search and Identification

Between January and April 2025, a systematic search was performed across Web of Science, Scopus, arXiv, IEEE Xplore, SpringerLink, and MDPI. Boolean combinations of domain-specific terms were used to retrieve literature focused on nonlinear dynamics, synchronization, network control, multilayer interactions, information propagation, structural inference, and optimization in networked systems. The query window covered the period 1997–2024. Only peer-reviewed journal articles and major conference proceedings in English were considered.

2.2. Eligibility and Exclusion Criteria

To be retained, a publication had to satisfy the following inclusion criteria:

• Provide formal modeling of dynamical processes on networks or graphs;
• Exhibit substantive analytical, computational, or algorithmic methodology;
• Address theoretical, algorithmic, or applied questions central to network dynamics;
• Appear in a reputable and indexed venue.

Publications were excluded if they lacked formalism, were purely descriptive, inaccessible in full text, or redundant relative to more comprehensive or rigorous sources.

2.3. Screening and Corpus Construction

The initial search returned 472 records. After removing duplicates and applying topical filters, 240 unique entries remained. These full texts were evaluated in detail for mathematical and methodological depth. A total of 134 articles were retained for inclusion in the final review (Figure 2).

All 134 selected works are cited explicitly and listed in the bibliography in the order of their first appearance. The subsequent sections present thematic classification and analytical synthesis.

Unlike previous topic- or domain-centered reviews, the present synthesis allocates each selected article to exactly one methodological pillar, thereby maximizing conceptual clarity and minimizing redundancy. Pillars are defined by the core analytical tools used in modeling, not by the domain of application. For example, epidemic models governed by Laplacian diffusion are grouped under spectral theory, while belief updating in social systems is analyzed within the probabilistic inference pillar. Cross-cutting techniques are noted but do not override primary categorization. This review privileges formal rigor and cross-domain transferability: empirical or descriptive works were excluded unless they anchored novel methods.

3. Deterministic Mean-Field Dynamics

Deterministic mean-field models form a foundational class of mathematical frameworks for describing the evolution of interacting agents on networks under prescribed structural constraints and rule-based dynamics. These models assume that macroscopic behavior arises from local interactions without recourse to stochastic perturbations or external optimization. Despite the wide range of applications, from synchronization and spreading processes to control and collective decision-making, their shared foundation lies in three structural principles: (i) the system state evolves through node- or edge-level variables in either continuous or discrete time; (ii) the dynamics are governed by deterministic update laws, typically grounded in conservation principles, rate equations, or variational arguments; (iii) network structure enters as a coupling constraint, encoded via adjacency matrices, graph Laplacians, or higher-order interaction tensors.

To expose the underlying structure of this diverse modeling landscape, we organize our review along a two-dimensional taxonomy (see Figure 3) that classifies models by their temporal formulation (discrete vs. continuous time) and their interaction complexity (linear, nonlinear, or adaptive/coevolving). This organization reflects both mathematical tractability and expressive capacity, offering a principled framework for comparing models across domains. The vertical axis distinguishes discrete-time models, which evolve via rule-based updates, from continuous-time formulations governed by differential equations. The horizontal axis spans from models with linear coupling (e.g., classical consensus and alignment) through nonlinear interactions (e.g., Kuramoto-type oscillators and epidemic dynamics) to adaptive or coevolving systems, where both nodal states and network structure evolve interactively.

This classification makes explicit the methodological affinities among seemingly disparate contributions. It allows us to chart a coherent trajectory from analytically tractable linear models to structurally complex, coadaptive dynamics. Rather than isolating each contribution, we emphasize their integration into a common analytical framework grounded in mean-field limits, graph-theoretic operators, and dynamical systems theory.

3.1. Consensus and Synchronization in Static and Time-Varying Networks

We begin with the foundational discrete-time consensus model introduced by Jadbabaie, Lin, and Morse [1], which provides a rigorous characterization of asymptotic agreement in networks of mobile agents interacting over temporally evolving proximity graphs. Let ${\theta }_i\left(t\right)\in \mathbb{R}$ denote the heading (or scalar opinion) of agent $i\in \{1,\dots ,N\}$ at discrete time step $t\in \mathbb{N}$, evolving under the linear local-averaging rule:

$${\theta }_i(t+1)={\displaystyle \frac{1}{|{\mathcal{N}}_i\left(t\right)|+1}}\left({\theta }_i\left(t\right)+\sum _{j\in {\mathcal{N}}_i\left(t\right)}{\theta }_j\left(t\right)\right),$$

(1)

where ${\mathcal{N}}_i\left(t\right)\subset \{1,\dots ,N\}$ denotes the (possibly time-varying) set of neighbors of agent i in the undirected proximity graph $G_t=(V,E_t)$ at time t, induced by a fixed interaction radius over agent positions. The normalization ensures that the update operator in (1) is row-stochastic and preserves convex combinations.

The central theorem asserts that if the union of communication graphs across any time window of fixed length $T\in \mathbb{N}$ is jointly connected, formally,

$$\bigcup _{s=t}^{t+T}{G}_s\mathrm{is}\mathrm{connected}\mathrm{for}\mathrm{all}t\in \mathbb{N},$$

(2)

then the state of every agent converges asymptotically to a common consensus value ${\theta }^{\ast }\in \mathbb{R}$, that is,

$$\underset{t\to \infty }{lim}{\theta }_i\left(t\right)={\theta }^{\ast },\forall i\in \{1,\dots ,N\},$$

(3)

where ${\theta }^{\ast }$ is a convex combination of initial values $\left\{{\theta }_i\left(0\right)\right\}$, determined by the (time-dependent) influence structure encoded in the graph sequence $\left\{G_t\right\}$.

Condition (2) constitutes a minimal, nonuniform persistence requirement, sufficient to guarantee global agreement through local updates, despite arbitrary temporal variation in network topology. The convergence proof leverages properties of products of stochastic matrices and contraction mappings over the consensus simplex. Importantly, the framework is robust to switching dynamics, bounded communication delays, and does not require symmetry or strong connectivity at individual time steps.

From a methodological perspective, this model exemplifies the archetype of linear deterministic dynamics on temporally varying networks, wherein stability and convergence properties can be deduced from graph-theoretic connectivity conditions and spectral features of the associated Laplacian operator. As such, it inaugurated a class of decentralized coordination algorithms subsequently generalized to continuous-time systems, asynchronous update regimes, heterogeneous coupling weights, and vector-valued agent states. The Jadbabaie model thus serves as a canonical benchmark across distributed systems theory, multi-agent control, and opinion dynamics, illustrating how coherent macroscopic order emerges from structurally local and temporally fluctuating interactions under minimal global assumptions.

At the ensemble level, the emergence of global coherence in oscillator networks with heterogeneous topology is exemplified by the study of Moreno and Pacheco [2], who investigate Kuramoto dynamics on scale-free graphs. Let ${\theta }_i\left(t\right)\in \mathbb{R}/2\pi \mathbb{Z}$ denote the phase of oscillator $i\in \{1,\dots ,N\}$ at time t, where each oscillator is associated with a node in a static undirected network of degree $k_i$. The dynamics are governed by the classical Kuramoto model with homogeneous coupling, and the degree-normalized evolution equation reads as follows:

$${\displaystyle \frac{d{\theta }_i}{dt}}={\omega }_i+\lambda \sum _{j\in {\mathcal{N}}_i}{\displaystyle \frac{1}{k_i}}sin({\theta }_j-{\theta }_i),$$

(4)

where ${\omega }_i\in \mathbb{R}$ is the natural frequency of oscillator i, $\lambda >0$ is the global coupling strength, and ${\mathcal{N}}_i$ is the set of neighbors of node i. Macroscopic synchrony is quantified via the global Kuramoto order parameter:

$$r\left(t\right)=\left|{\displaystyle \frac{1}{N}}\sum _{j=1}^N{e}^{i{\theta }_j\left(t\right)}\right|,$$

(5)

which measures the phase coherence of the ensemble, with $r\left(t\right)\in [0,1]$ and $r\left(t\right)\approx 1$ indicating full synchrony.

The authors demonstrate a continuous (second-order) synchronization transition at a critical coupling threshold ${\lambda }_c$, with $r\left(t\right)\sim {(\lambda -{\lambda }_c)}^{\beta }$ for some critical exponent $\beta >0$, consistent with mean-field theory. Importantly, they analyze the reintegration dynamics of individual oscillators following perturbations. Let $\tau \left(k\right)$ denote the reintegration time for an oscillator of degree k, defined as the characteristic time required to resynchronize with the coherent cluster after a desynchronizing event. They empirically observe the scaling relation:

$$\tau \left(k\right)\sim k^{-\nu },$$

(6)

where $\nu >0$ is a fitted exponent. This relation reveals that highly connected nodes (large k) reintegrate rapidly and act as structural stabilizers of the synchronized phase, while low-degree nodes exhibit slower convergence.

Extending beyond static coupling, Aoki and Aoyagi [3] introduce a co-evolutionary model in which both oscillator phases and coupling weights evolve adaptively. Let ${\varphi }_i\left(t\right)\in \mathbb{R}/2\pi \mathbb{Z}$ denote the dynamic phase of oscillator i, and let $k_{ij}\left(t\right)=k_{ji}\left(t\right)\in \mathbb{R}$ denote the symmetric time-dependent coupling weight between oscillators i and j. The joint evolution is described by the following:

$${\displaystyle \frac{d{\varphi }_i}{dt}}={\omega }_i-{\displaystyle \frac{1}{N}}\sum _{j=1}^N{k}_{ij}\left(t\right)sin({\varphi }_i-{\varphi }_j+\alpha ),$$

(7)

$${\displaystyle \frac{d{k}_{ij}}{dt}}=-ϵsin({\varphi }_i-{\varphi }_j+\alpha ),$$

(8)

where ${\omega }_i$ is the natural frequency (often set to ${\omega }_i=1$), $ϵ>0$ is the learning (plasticity) rate, and $\alpha \in [0,\pi ]$ is a fixed phase lag parameter. The sign structure of (8) promotes Hebbian reinforcement of in-phase oscillations and weakening of antiphase links.

Depending on the value of $\alpha$, the system self-organizes into qualitatively distinct regimes: two-cluster antiphase synchrony for $\alpha \approx \pi$, global phase locking for small $\alpha$, and spatiotemporal chaos at intermediate values. To quantify the emergent complexity of each regime, the authors compute the Shannon entropy and mutual information of phase time series across nodes. Their results demonstrate that adaptive feedback enriches the dynamic repertoire of oscillator networks and enables endogenous self-organization of connectivity.

Together, the models in (4)–(8) illustrate the fundamental interplay between topological heterogeneity, plastic interactions, and collective synchronization. They mark a conceptual transition from fixed, passive network structure to co-evolving architectures in the study of complex coordinated dynamics.

Extending the principle of emergent coordination to continuous-state systems, the coevolutionary model proposed by Ito and Kaneko [4] introduces a minimal yet expressive framework for adaptive synchronization in networks of chaotic oscillators. Each node $i\in \{1,\dots ,N\}$ is endowed with a scalar state variable $x_i\left(t\right)\in \mathbb{R}$, evolving according to a discrete-time chaotic map:

$$x_i(t+1)=(1-\kappa )f\left(x_i\left(t\right)\right)+{\displaystyle \frac{\kappa }{k_i\left(t\right)}}\sum _{j\in {\mathcal{N}}_i\left(t\right)}{w}_{ij}\left(t\right)f\left(x_j\left(t\right)\right),$$

(9)

where $f(·)$ is a nonlinear activation function—typically the logistic map $f\left(x\right)=rx(1-x)$ with $r>3.57$, or a tent map; $\kappa \in [0,1]$ is the coupling strength; $w_{ij}\left(t\right)\in [0,1]$ is the time-dependent weight of the directed edge from node j to node i; and $k_i\left(t\right)={\sum }_{j\in {\mathcal{N}}_i\left(t\right)}{w}_{ij}\left(t\right)$ is the dynamic in-degree normalization term. The state update rule (9) reflects weighted averaging of transformed neighbor states, enforcing diffusive coupling among chaotic units.

Simultaneously, the weighted adjacency matrix $W\left(t\right)=\left[w_{ij}\left(t\right)\right]\in {\mathbb{R}}^{N\times N}$ evolves according to a Hebbian-inspired local plasticity mechanism:

$${\displaystyle \frac{d{w}_{ij}}{dt}}=ϵ\left[f\left(x_j\left(t\right)\right)x_i\left(t\right)-w_{ij}\left(t\right)\right],$$

(10)

where $ϵ>0$ is the learning rate. The right-hand side of (10) reinforces links when the post-synaptic activity $x_i\left(t\right)$ and pre-synaptic transformed input $f\left(x_j\left(t\right)\right)$ are positively correlated, thus aligning structural connectivity with dynamic coordination. Over time, this mutually reinforcing feedback mechanism results in the spontaneous differentiation of network topology, despite identical initial conditions and symmetric parameter settings.

Numerical simulations of (9)–(10) reveal metastable formation of dynamic hierarchies: certain nodes evolve into persistent sources of influence—informally termed “leaders”—while others converge to follower-like behaviors. This symmetry breaking is not imposed but emerges endogenously through the interplay between local synchronization and structural plasticity. The emergent attractor landscape exhibits sensitivity to initial perturbations but displays robust statistical features under noise, as shown through bifurcation analysis and ensemble studies.

From a theoretical standpoint, the Ito–Kaneko model constitutes a prototypical example of coevolving dynamical networks. Its analytical tractability enables exploration of fundamental questions in structure–function interdependence, self-organization, and adaptive control. The model also provides a conceptual scaffold for extending mean-field approximations, deriving reduced-order dynamics, and constructing data-driven control schemes in systems where functional coordination and structural plasticity are inseparably intertwined.

3.2. Compartmental and Agent-Based Spreading Models

In the present section, let $x\left(t\right)\in {\mathbb{R}}^d$ denote the state vector of any compartmental system introduced below. Unless otherwise specified, we impose the admissible initial condition

$$x\left(0\right)=x_0\in \mathcal{D},\mathcal{D}=\left\{x\in {\mathbb{R}}^d:x_i\ge 0\mathrm{for}\mathrm{all}i,\sum _{i=1}^d{x}_i=n\right\},$$

(11)

where $n>0$ is the conserved total population (or normalized density). The domain $\mathcal{D}\subset {\mathbb{R}}^d$ is convex and compact, ensuring that all trajectories remain biologically meaningful.

For delay differential systems such as (20), we prescribe a continuous initial history:

$$x\left(t\right)=\varphi \left(t\right),t\in [-\tau ,0],\varphi (·)\in \mathcal{C}\left([-\tau ,0],\mathcal{D}\right),$$

(12)

where $\tau >0$ is the maximum delay. For degree-resolved models (e.g., Equations (21)–(25)), the initial conditions are assigned separately for each degree class $k\in \mathbb{N}$, and satisfy:

$$S_k\left(0\right)+E_k\left(0\right)+\cdots +R_k\left(0\right)=p\left(k\right)n,$$

where $p\left(k\right)$ is the empirical degree distribution. These conditions ensure consistency with the conserved total mass constraint across the network.

Under the standard Lipschitz continuity assumptions on the right-hand sides, these initializations guarantee the existence and uniqueness of local (and, on the positively invariant set $\mathcal{D}$, global) solutions by the classical Picard–Lindelöf theorem. For delay systems, analogous results follow from the theory of functional differential equations with continuous histories.

We now turn to deterministic compartmental models of spreading processes, focusing on structurally enriched and feedback-modulated extensions of the classical SIR framework. Escalante and Odehnal [5] proposed a generalized SIRS model to capture competing rumor dynamics, extending the canonical three-compartment system:

$$\left\{\begin{array}{c}{\displaystyle \frac{ds}{dt}}=-\beta si,\hfill \\ {\displaystyle \frac{di}{dt}}=\beta si-\alpha i,\hfill \\ {\displaystyle \frac{dr}{dt}}=\alpha i,\hfill \end{array}\right.$$

(13)

where $s\left(t\right),i\left(t\right),r\left(t\right)$ denote susceptible, infectious, and removed fractions, and $\beta ,\alpha$ are the transmission and removal rates. To model temporary immunity, the system is extended as

$$\left\{\begin{array}{c}{\displaystyle \frac{ds}{dt}}=-\beta si+\gamma r,\hfill \\ {\displaystyle \frac{di}{dt}}=\beta si-\alpha i,\hfill \\ {\displaystyle \frac{dr}{dt}}=\alpha i-\gamma r,\hfill \end{array}\right.$$

(14)

with $\gamma$ denoting the rate of return from removed to susceptible. A further refinement introduces a fourth compartment $v\left(t\right)$, modeling an antagonistic process (e.g., counter-rumor), yielding

$$\left\{\begin{array}{c}{\displaystyle \frac{ds}{dt}}=-\beta si-\varphi s+\gamma (n-s-i-v),\hfill \\ {\displaystyle \frac{di}{dt}}=\beta si+\sigma \beta vi-\alpha i,\hfill \\ {\displaystyle \frac{dv}{dt}}=\varphi s-\sigma \beta vi,\hfill \end{array}\right.$$

(15)

where $\varphi$ is the spread rate of the antagonistic signal and $\sigma \in [0,1]$ modulates its interaction with the primary process. Reducing the system under the constraint $s+i+v+r=n$, one obtains a planar subsystem:

$$\left\{\begin{array}{c}{\displaystyle \frac{di}{dt}}=\beta (n-i-(1-\sigma )v)i-\alpha i,\hfill \\ {\displaystyle \frac{dv}{dt}}=\varphi (n-i)-\sigma \beta vi-\varphi v,\hfill \end{array}\right.$$

(16)

which facilitates analytical treatment.

Linearizing the subsystem (16) at the rumor-free equilibrium $(i,v)=(0,0)$, we obtain the Jacobian matrix with leading eigenvalue

$${\lambda }_i=\beta n-\alpha .$$

(17)

Hence, the local invasion threshold is given by

$$R_0={\displaystyle \frac{\beta n}{\alpha }},$$

(18)

with $R_0<1$ implying local asymptotic stability of the rumor-free state. Notably, this threshold depends only on the primary transmission and removal rates $\beta$ and $\alpha$, and is independent of the antagonistic spread rate $\varphi$. The reason is structural: $\varphi$ affects only the dynamics of the variable $v\left(t\right)$, and thus does not enter into the linearized equation for $i\left(t\right)$.

However, while $\varphi$ does not alter the local stability threshold, it critically shapes the global behavior of the system. Together with $\sigma$ and $\beta$, the parameter $\varphi$ governs the topology of the global phase portrait, determining whether the system exhibits forward or backward bifurcations, multistability, or coexistence of multiple attractors. Thus, the antagonistic signal influences the nonlinear resolution of competing processes without modifying the conditions for initial outbreak.

Li, Ma, and Wang [6] incorporate delay and nonlinear control into a threshold-modulated SIR model. The control mechanism is modeled via the piecewise function

$$T\left(i\right)=\left\{\begin{array}{cc}0,\hfill & i<i_{\mathrm{th}},\hfill \\ qi,\hfill & i\ge i_{\mathrm{th}},\hfill \end{array}\right.$$

(19)

which activates proportionally when the infected fraction $i\left(t\right)$ exceeds a fixed threshold $i_{\mathrm{th}}\in (0,1)$, with control strength $q>0$. This leads to the delayed dynamical system:

$$\left\{\begin{array}{c}{\displaystyle \frac{ds}{dt}}=\lambda -\beta si-\mu s,\hfill \\ {\displaystyle \frac{di}{dt}}=\beta si-\alpha ir(t-\tau )-\mu i-T\left(i\right),\hfill \\ {\displaystyle \frac{dr}{dt}}=\alpha ir(t-\tau )-\mu r+T\left(i\right),\hfill \end{array}\right.$$

(20)

where $\lambda >0$ is the recruitment rate, $\mu >0$ the natural removal rate, $\alpha >0$ the interaction coefficient, and $\tau >0$ the fixed time delay.

The initial condition consists of a continuous history function $\mathbf{x}\left(t\right)=\mathit{\varphi }\left(t\right)\in \mathcal{C}\left(\right[-\tau ,0],\mathcal{D})$, where $\mathcal{D}\subset {\mathbb{R}}^3$ is the invariant simplex. The system is hybrid: the discontinuous control input $T\left(i\right)$ introduces a switching surface at $i=i_{\mathrm{th}}$, while the delay $\tau$ gives rise to infinite-dimensional dynamics. Local analysis reveals Hopf bifurcations and the emergence of sliding pseudo-equilibria, with early activation of control shown to mitigate peak infection levels and reduce overshoot.

Taken together, these models demonstrate the increasing analytical sophistication of compartmental spreading dynamics, from continuous deterministic flows with multiple compartments and feedback to hybrid systems with discontinuous right-hand sides and delay-induced oscillations. They form the theoretical substrate for individual-based extensions and network-structured models explored in subsequent sections.

To account for structural and behavioral heterogeneity, recent work extends classical compartmental frameworks to multiplex and degree-annotated networks. Di et al. [7] introduced an individual-based SEIR scheme in which each agent carries specific infection and recovery rates, together with influence weights on a multiplex graph. After mean–field aggregation the population fractions $s\left(t\right)$ (susceptible), $e\left(t\right)$ (exposed/latent), and $i\left(t\right)$ (infectious) evolve according to

$$\begin{array}{cc}\hfill {\displaystyle \frac{ds}{dt}}& =a-se\sum _i{\mu }_i\sum _j{p}_{ij}-s\sum _i{\epsilon }_i,\hfill \\ \hfill {\displaystyle \frac{de}{dt}}& =se\sum _i{\mu }_i\sum _j{p}_{ij}-ei\sum _j{\beta }_j\sum _k{q}_{jk}-e\sum _j{\alpha }_j,\hfill \\ \hfill {\displaystyle \frac{di}{dt}}& =ei\sum _j{\beta }_j\sum _k{q}_{jk}-i\sum _k{\lambda }_k,\hfill \end{array}$$

(21)

where $p_{ij},q_{jk}$ are layer-specific contact weights and the coefficients ${\mu }_i,{\beta }_j,{\alpha }_j,{\lambda }_k,{\epsilon }_i$ capture individual propensities. Linear stability yields an explicit basic reproduction number

$$R_0={\displaystyle \frac{\left(\mu p\right)\left(\beta q\right)}{(\beta q+\alpha )\lambda }},\nu ={\displaystyle \frac{\mu p}{\alpha +\lambda }},$$

(22)

which, after standard final size analysis, leads to the transcendental relation

$$r=1-exp(-\nu r),$$

(23)

where r is the eventual removed fraction and $exp(·)$ denotes the natural exponential function, thereby avoiding any confusion with the exposed variable $e\left(t\right)$. Extending this idea to degree-stratified populations, Tong et al. [8] proposed the SHILR model ${\{s_k,h_k,i_k,{\ell }_k,r_k\}}_{k\in \mathbb{N}}$ driven by the kernel

$$\Theta \left(t\right)=\sum _k{\displaystyle \frac{kp\left(k\right)}{\langle k\rangle }}{i}_k\left(t\right).$$

(24)

The model captures behavior-dependent latency and treatment dynamics:

$$\begin{array}{cc}\hfill {\displaystyle \frac{d{s}_k}{dt}}& =b-\alpha k{s}_k\Theta \left(t\right)-\mu s_k,\hfill \\ \hfill {\displaystyle \frac{d{h}_k}{dt}}& ={\theta }_2\alpha k{s}_k\Theta \left(t\right)-(\beta +\mu )h_k,\hfill \\ \hfill {\displaystyle \frac{d{i}_k}{dt}}& ={\theta }_1\alpha k{s}_k\Theta \left(t\right)+m\beta h_k-(\gamma +\mu )i_k+\phi {\ell }_k,\hfill \\ \hfill {\displaystyle \frac{d{\ell }_k}{dt}}& =\gamma i_k-(\phi +\eta +\mu ){\ell }_k,\hfill \\ \hfill {\displaystyle \frac{d{r}_k}{dt}}& =(1-m)\beta h_k+(1-{\theta }_1-{\theta }_2)\alpha k{s}_k\Theta \left(t\right)+\eta {\ell }_k-\mu r_k,\hfill \end{array}$$

(25)

where $s_k,h_k,i_k,{\ell }_k,r_k$ denote the susceptible, hesitant, infectious, latent (under treatment), and removed subpopulations of degree class $k\in \mathbb{N}$; $\alpha$ is the transmission rate, $\mu$ the natural removal rate, $\beta$ the rate of progression from hesitant to infectious, $\gamma$ the transition rate from infectious to latent, $\phi$ the rate of relapse, and $\eta$ the recovery rate from latent state. Parameters ${\theta }_1,{\theta }_2\in [0,1]$ denote behavioral branching proportions, and $m\in [0,1]$ represents the fraction of hesitant individuals who proceed to infection. The contact kernel is defined as

$$\Theta \left(t\right)=\sum _k{\displaystyle \frac{kp\left(k\right)}{\langle k\rangle }}{i}_k\left(t\right),$$

(26)

where $p\left(k\right)$ is the degree distribution and $\langle k\rangle$ the mean degree.

The associated degree-based basic reproduction number reads as follows:

$$R_0={\displaystyle \frac{\langle k^2\rangle }{\langle k\rangle }}·{\displaystyle \frac{a_3(a_1{\theta }_1+m\beta {\theta }_2)\alpha b}{\mu a_1(a_2{a}_3-\phi \gamma )}},$$

(27)

where

$$a_1=\beta +\mu ,a_2=\gamma +\mu ,a_3=\phi +\eta +\mu ,$$

and b is the recruitment rate into the susceptible population. Optimal intervention strategies $U_k\left(t\right)$, acting on treatment rates in the ${\ell }_k$ compartment, can be derived by minimizing an integral cost functional over a finite horizon using Pontryagin’s Maximum Principle or variational techniques.

Together, these contributions establish a high-resolution modeling framework that integrates structural diversity, agent-level variability, and control inputs, expanding the analytical scope of dynamic network epidemiology.

3.3. Multilayer Oscillators and Heterogeneous Mean-Field Theory

Within the domain of nonlinear dynamics on multilayer networks, Sevilla-Escoboza et al. [9] analyze synchronization phenomena in systems composed of multistable Rössler oscillators coupled both intra- and inter-layer. The dynamical equations for the oscillator in each layer take the form:

$$\begin{array}{cc}\hfill {\dot{x}}_{1,2}& =-a_1\left(x_{1,2}+\beta y_{1,2}+\tau z_{1,2}\right)+\phi (x_{2,1}-x_{1,2}),\hfill \\ \hfill {\dot{y}}_{1,2}& =-a_2\left(-\gamma x_{1,2}-(1-\delta )y_{1,2}\right),\hfill \\ \hfill {\dot{z}}_{1,2}& =-a_3\left(-g\left(x_{1,2}\right)+z_{1,2}\right),\hfill \end{array}$$

(28)

where $g\left(x\right)$ is a piecewise-linear function governing the attractor topology, and indices $(1,2)$ label nodes across two network layers. Parameters $a_1,a_2,a_3$ define intrinsic oscillator timescales, while $\beta ,\tau ,\gamma ,\delta$ shape the local coupling geometry. The interlayer coupling strength $\phi$ modulates cross-layer synchronization. Employing the master stability function (MSF) formalism [10], the authors demonstrate that intermittent synchronization results from spontaneous transitions between coexisting attractors, with global stability ensured when the rescaled Laplacian eigenvalues ${\nu }_i=\alpha {\lambda }_i$ remain within the MSF-defined stability band. This result reveals that the interplay between multistability and network topology generates complex synchronization landscapes sensitive to structural perturbations and coupling heterogeneity, thus providing a rigorous spectral criterion for stability in layered oscillator systems.

Expanding the analytical scope from individual-based contagion models to ensemble-level abstractions, Vespignani [11] formulated a unifying heterogeneous mean-field (HMF) theory for dynamic processes on complex networks. In this framework, nodes are stratified by degree class $k\in \mathbb{N}$, and the state variable ${\rho }_k\left(t\right)\in [0,1]$ denotes the probability that a node of degree k is infected at time t. The evolution of ${\rho }_k\left(t\right)$ follows the rate equation:

$${\displaystyle \frac{d{\rho }_k}{dt}}=-\mu {\rho }_k+\lambda k\left[1-{\rho }_k\left(t\right)\right]\Theta \left(t\right),$$

(29)

where $\lambda$ is the per-contact transmission rate, $\mu$ is the recovery rate, and $\Theta \left(t\right)$ is the probability that a randomly chosen neighbor is infected:

$$\Theta \left(t\right)={\displaystyle \frac{1}{\langle k\rangle }}\sum _{k}kP\left(k\right){\rho }_k\left(t\right).$$

(30)

Linear stability analysis of the disease-free equilibrium yields the epidemic threshold:

$${\lambda }_c={\displaystyle \frac{\mu \langle k\rangle }{\langle k^2\rangle }},$$

(31)

implying that in networks with diverging second moment $\langle k^2\rangle$, such as scale-free networks with exponent $\gamma \le 3$, the threshold vanishes asymptotically, i.e., ${\lambda }_c\to 0$, and arbitrarily weak contagion can induce endemic states. This result fundamentally challenges classical homogeneous mixing assumptions and reveals the intrinsic vulnerability of highly heterogeneous topologies to global outbreaks.

The particle–network formalism extends this paradigm by modeling contagion as a system of interacting particles on a stochastic substrate, governed by master equations that track the joint evolution of node and edge states. This mesoscopic perspective captures interdependencies between nodal dynamics and network structure, integrating temporal correlations, multiplex coupling, and adaptive rewiring. It enables analytical treatment of complex emergent phenomena such as cascade amplification, superspreading, and non-Markovian delays, and rigorously delineates the limits of mean-field and mass-action approximations. Within the broader taxonomy of deterministic mean-field dynamics, this framework constitutes the analytically tractable asymptotic limit of large, locally tree-like networks under annealed or quenched randomness, offering a scalable formalism for predicting thresholds, criticality, and intervention efficacy in high-dimensional spreading systems.

3.4. Taxonomies and Structural Classifications

The survey by Grabisch and Rusinowska [12] offers a rigorous and axiomatically grounded taxonomy of nonstrategic opinion dynamics models, systematically organizing a broad class of binary- and continuous-valued systems by their update mechanisms and structural assumptions. Binary-valued models—encompassing the classical voter model, majority rule dynamics, and generalized q-voter processes—are defined on discrete opinion spaces $\{0,1\}$ and evolve under probabilistic or deterministic rules driven by local majority or threshold functions. Continuous-valued models, by contrast, operate on opinion domains $x_i\left(t\right)\in [0,1]$ and implement aggregation schemes via local functions of neighbor states:

$$x_i(t+1)=f_i\left(\{x_j\left(t\right):j\in {\mathcal{N}}_i\}\right),$$

(32)

where ${\mathcal{N}}_i$ denotes the neighborhood of agent i. The taxonomy distinguishes between conformist updating (opinions moving toward the local average), anticonformity (movement away from dominant opinion), and independence (stochastic disregard for peer input), thereby encoding a spectrum of cognitive behaviors. Bounded-confidence models, typified by the Deffuant–Weisbuch and Hegselmann–Krause systems, restrict interaction to agents whose opinions lie within a threshold $ϵ$, introducing effective interaction graphs $G_t^{ϵ}$ that evolve endogenously. Theoretical analysis via heterogeneous mean-field limits and Fokker–Planck equations identifies distinct macroscopic regimes—consensus, polarization, fragmentation, and disorder—each emerging from variations in noise amplitude, confidence thresholds, or interaction strength. These models formalize how decentralized local rules, subject to minimal cognitive assumptions, can robustly generate complex and heterogeneous patterns of opinion formation in large populations. The Grabisch–Rusinowska framework thus provides a foundational classification schema for interpreting and extending opinion dynamics models, clarifying the minimal mechanisms underlying observed social consensus and divergence.

3.5. Control-Theoretic Perspectives on Network Dynamics

At a more mechanistic level, Lynn and Bassett [13] integrate dynamical systems theory, statistical physics, and control theory to formulate a unified framework for analyzing information flow, learning, and cognitive flexibility in neural and complex systems. Central to their formulation is the notion of structural controllability, which quantifies the capacity of external inputs to drive the system from one dynamical state to another, given the underlying network topology. For a linear time-invariant system of the form

$$\dot{x}\left(t\right)=Ax\left(t\right)+Bu\left(t\right),$$

(33)

where $x\left(t\right)\in {\mathbb{R}}^n$ denotes the system state, $u\left(t\right)\in {\mathbb{R}}^m$ the control input, $A\in {\mathbb{R}}^{n\times n}$ the adjacency (or coupling) matrix, and $B\in {\mathbb{R}}^{n\times m}$ the input matrix specifying control nodes, the minimum energy required to drive the system from $x\left(0\right)=x_0$ to $x\left(T\right)=x_f$ over a finite time horizon T is given by

$$E={\int }_0^T{\parallel u\left(t\right)\parallel }^{2}dt={(x_f-e^{AT}{x}_0)}^{\top }{W}^{-1}\left(T\right)(x_f-e^{AT}{x}_0),$$

(34)

where $W\left(T\right)={\int }_0^T{e}^{A\tau }B{B}^{\top }{e}^{A^{\top }\tau }d\tau$ denotes the controllability Gramian. The structure and conditioning of $W\left(T\right)$ depend sensitively on spectral properties of the network: the algebraic connectivity ${\lambda }_2\left(L\right)$, where L is the graph Laplacian; the spectral radius ${\lambda }_{max}\left(A\right)$; and node centrality measures, including degree, betweenness, and eigenvector centrality. These descriptors govern how perturbations propagate and accumulate across the network, and thus determine the energy efficiency and feasibility of control trajectories. Furthermore, the framework accommodates nonlinear generalizations that exhibit multistability, metastability, and synchronization, reflecting how cognitive systems traverse dynamic manifolds under internal constraints and external stimuli. In this perspective, learning is recast as a control process over the neural state space, whereby structural features such as modularity, hierarchy, and hub-dominance modulate energy landscapes for transition and stabilization. Collectively, these results advance a theory in which information propagation and adaptive behavior emerge not from inference or optimization algorithms per se, but from intrinsic dynamical and topological constraints of the network substrate. They thereby motivate the development of belief propagation, coordination, and intervention mechanisms that are both structure-aware and energy-efficient, especially in high-dimensional systems with latent dynamics and constrained control channels.

3.6. Coevolutionary Opinion Dynamics

Min and San Miguel [14] introduce a coevolutionary nonlinear voter model where both opinions and network topology evolve. Each node holds a binary state $s_i\in \{-1,+1\}$. With probability p, an active link (connecting nodes with opposite opinions) is rewired to connect to a like-minded node; with probability $1-p$, the node flips its opinion with probability given by a nonlinear function of disagreeing neighbors:

$$w\left({\rho }_i\right)={\rho }_i^q,$$

(35)

where ${\rho }_i=a_i/k_i$ is the disagreement ratio, and $q\in {\mathbb{R}}_{+}$ tunes nonlinearity. The dynamics are governed by the following:

$$\begin{array}{cc}\hfill {\displaystyle \frac{dm}{dt}}& =-2(1-p)\langle k\rangle {\rho }^{q}m,\hfill \\ \hfill {\displaystyle \frac{d\rho }{dt}}& =-2p\rho (1-\rho )+(1-p){\rho }^q(1-\rho )(1-2\rho ).\hfill \end{array}$$

(36)

Phase diagrams show consensus ($m=\pm 1,\rho =0$), coexistence ($m=0,\rho >0$), and fragmentation. Sublinear response ($q<1$) stabilizes coexistence, while superlinear q induces abrupt transitions.

Nardini, Kozma, and Barrat [15] develop a minimal coevolutionary binary opinion model combining imitation and rewiring. Each agent i with opinion ${\sigma }_i\in \{-1,+1\}$ interacts asynchronously. With probability p, it rewires a discordant link to a like-minded node; with $1-p$, it adopts the neighbor’s opinion. The critical rewiring rate $p_c\approx 0.38$ separates global consensus from fragmentation. The active link density $\rho$ evolves as

$${\displaystyle \frac{d\rho }{dt}}=-2(1-p)\rho +2p\rho (1-\rho ),$$

(37)

and bifurcates at $p_c$, revealing how adaptive interactions control consensus.

Bizyaeva, Franci, and Leonard [16] formulate a continuous-state opinion-dynamics model on a fixed, undirected graph $G=(V,E)$ in which each agent $i\in V$ holds a scalar state ${\theta }_i\left(t\right)\in \mathbb{R}$ evolving under the nonlinear consensus law

$${\dot{\theta }}_i\left(t\right)=\sum _{j=1}^N{A}_{ij}{f}_{\kappa }\left({\theta }_j-{\theta }_i\right),f_{\kappa }\left(x\right)=\kappa tanh\left(x/\kappa \right),$$

(38)

where $A_{ij}=A_{ji}\ge 0$ denotes adjacency weights and the sensitivity parameter $\kappa >0$ continuously interpolates between nearly linear interactions ($\kappa \to \infty$) and a bounded-confidence–type regime ($\kappa \to 0$). By introducing the energy functional

$$V\left(t\right)=\frac{1}{2}\sum _{i<j}{A}_{ij}{F}_{\kappa }\left({\theta }_j-{\theta }_i\right),F_{\kappa }^{\prime }\left(x\right)=f_{\kappa }\left(x\right),$$

(39)

they obtain the dissipation relation

$$\dot{V}\left(t\right)=-\frac{1}{2}\sum _{i<j}{A}_{ij}{f}_{\kappa }^2\left({\theta }_j-{\theta }_i\right)\le 0,$$

(40)

which guarantees boundedness and convergence of trajectories. Spectral analysis of the Laplacian $L=D-A$ yields a critical threshold ${\kappa }_c\sim {\lambda }_2{\left(L\right)}^{-1}$ such that $\kappa >{\kappa }_c$ ensures global consensus, whereas $\kappa <{\kappa }_c$ allows stable multicluster equilibria whose sizes correspond to connected components of the subgraph induced by edges satisfying $|{\theta }_j-{\theta }_i|=\mathcal{O}\left(\kappa \right)$. This framework unifies linear consensus and bounded-confidence dynamics within a smooth family and furnishes Lyapunov-based bifurcation criteria linking collective outcomes to both network topology and nonlinear interaction sensitivity.

The progression across deterministic mean-field models reveals a continuous enrichment of dynamical and structural complexity. Initial frameworks based on linear averaging and static topology, as in classical consensus models, are extended by introducing nonlinearity, heterogeneity, and multilayer structure. Further developments incorporate adaptive topologies, where feedback loops between node states and network configuration give rise to coevolutionary effects. Structural classifications and control-theoretic formulations deepen this landscape by relating macroscopic behavior to spectral, modular, and energy-based properties of the network. Taken together, these models illustrate the breadth of deterministic mechanisms, in which networks generate, constrain, and respond to collective dynamics, providing both a benchmark for probabilistic models and a conceptual foundation for the subsequent sections of this review.

We therefore proceed to Section 4, where we formalize these tools and examine their role in shaping and constraining the dynamics of complex networks.

4. Spectral and Linear-Algebraic Tools

The organizational principle underlying Section 4 is summarized in Figure 4, which presents a two-dimensional taxonomy that cross-classifies every cited reference according to two orthogonal axes: the vertical axis enumerates methodological lineages of spectral analysis, ranging from eigen-pair metrics that probe expansion via the Laplace second eigenvalue ${\lambda }_2$, through random matrix diagnostics, controllability spectra, supra tensor operators, to entropy- and motif-based path statistics; whereas the horizontal axis partitions network architectures into random, scale-free, modular, directed–temporal, and multilayer-interdependent regimes.

This Cartesian layout was chosen because it simultaneously (i) preserves the historical genealogy of spectral tools, (ii) highlights the structural heterogeneity against which those tools are deployed, and (iii) exposes systematic gaps that guided article selection. Each of the seven textual subsections occupies a contiguous ribbon of cells running diagonally from top left to bottom right, thereby illustrating how foundational eigenvalue theory (Section 4.1) flows into universality diagnostics (Section 4.2), synchronization thresholds (Section 4.3), consensus and optimization (Section 4.4), controllability frameworks (Section 4.5), tensorial multilayer formalisms (Section 4.6), and finally inference-oriented extensions (Section 4.7).

The matrix is dense: every intersection contains at least one canonical study, signalling that the literature reviewed here spans the full combinatorial span of methodological families and structural contexts. Cells in the upper left capture mean–field and expander analyses where spectral gaps translate directly into isoperimetric bounds; cells along the main diagonal trace the migration of random matrix insights to scale-free and modular settings; the lower right quadrant condenses work on multilayer controllability, tensor Laplacians and data-driven coarse graining, thus evidencing the progressive enrichment of spectral theory toward high-dimensional, temporally evolving and interdependent networks. By reading the figure column-wise, one discovers how a single architectural motif, say, directed temporality, interacts with multiple spectral agendas, while a row-wise scan reveals how a single analytical tradition adapts as one moves from homogeneous to heterogeneous and finally to multilayer substrates.

Hence, Figure 4 serves both as an expository map for the reader and as a methodological audit, justifying the inclusion of each study by situating it at the unique intersection of a spectral toolkit and a network architecture that together articulate the section’s goal of forging a comprehensive, multiscale, and application-ready spectral science of complex networks.

4.1. Fundamentals of Spectral Graph Theory

Spectral graph theory provides a rigorous analytical framework for linking global structural properties of networks to local combinatorial features via the spectra of graph-associated matrices. Central to this approach are the adjacency matrix A and the normalized Laplacian $\mathcal{L}=I-D^{-1/2}A{D}^{-1/2}$, whose eigenvalues encode key indicators of connectivity, expansion, and robustness. In particular, the second-smallest eigenvalue of $\mathcal{L}$, denoted ${\lambda }_2$, functions as a spectral gap whose magnitude governs mixing times in diffusive processes and convergence rates in consensus dynamics.

Chung [17] established the foundational link between ${\lambda }_2$ and graph expansion through a Cheeger-type inequality, which bounds the isoperimetric constant $h\left(G\right)$—a proxy for edge boundary-to-volume ratio—via

$${\displaystyle \frac{1}{2}}{\lambda }_2\le h\left(G\right)\le \sqrt{2{\lambda }_2}.$$

(41)

This spectral characterization enables the diagnosis of bottlenecks and fragmentation through eigenvalue analysis alone. Extending this perspective, Chung and Lu [39] developed a generative model for graphs with prescribed expected degree sequences, demonstrating that navigability and path-length scaling depend on second-order degree moments:

$${\mathrm{dist}}_{\mathrm{avg}}\sim {\displaystyle \frac{logn}{log\tilde{d}}},\tilde{d}={\displaystyle \frac{{\sum }_i{w}_i^2}{{\sum }_i{w}_i}},$$

(42)

with ultra-small-world behavior emerging in scale-free regimes $2<\beta <3$.

Focusing on regular topologies, Mishra et al. [18] characterized the expansion quality of k-regular graphs through bounds on the second-largest adjacency eigenvalue:

$${\displaystyle \frac{k-{\lambda }_2}{2}}\le h\left(G\right)\le \sqrt{2k(k-{\lambda }_2)},$$

(43)

highlighting ${\lambda }_2$ as a spectral surrogate for algebraic connectivity and mixing efficiency. Complementing these global formulations, Afshari et al. [40] introduced a localized lower bound for the Laplacian spectral radius ${\lambda }_{max}\left(\mathcal{L}\right)$, leveraging neighborhood dissimilarity:

$$\lambda \left(G\right)\ge \underset{v_i\in V}{max}\left\{m_i^{\prime }+\left(1+{\displaystyle \frac{{(m_i^{\prime }-1)}^2}{d_{2,i}}}\right)·{\displaystyle \frac{d_i}{m_i^{\prime }}}\right\},$$

(44)

where $m_i^{\prime }$ denotes motif diversity and $d_{2,i}$ encodes second-order degree.

Together, these contributions delineate a coherent foundation for spectral diagnostics in complex networks. Across formulations, ${\lambda }_2$ recurs as a unifying parameter linking topological expansion, navigability, and dynamic relaxation. The emergence of second-moment degree measures and local structural heterogeneity as modulators of spectral behavior reflects an overarching methodological shift toward multiscale and inference-capable formulations of spectral theory.

4.2. Spectral Diagnostics and Universality in Random Graphs

Spectral statistics provide a powerful framework for probing the structural universality and mesoscopic organization of random graphs. In this context, the empirical spectral density $\rho \left(\lambda \right)$, the distribution of eigenvalue spacings, and long-range spectral correlations serve as diagnostic fingerprints of network classes, capturing both global topology and local motif structure.

Farkas et al. [19] established that the k-th spectral moment of the adjacency matrix A encodes the frequency of closed walks of length k, via the identity:

$$M_k={\displaystyle \frac{1}{N}}\mathrm{Tr}\left(A^k\right),$$

(45)

linking spectral observables to combinatorial substructures. For Erdős–Rényi graphs, the spectral density converges in the large-N limit to Wigner’s semicircle law:

$$\rho \left(\lambda \right)={\displaystyle \frac{1}{2\pi {\sigma }^2}}\sqrt{4{\sigma }^2-{\lambda }^2},{\sigma }^2=Np(1-p),$$

(46)

indicating spectral universality under mean-field assumptions. However, in scale-free graphs, this behavior breaks down: the spectrum exhibits heavy tails, skewed distributions, and eigenvector localization on hubs, reflecting pronounced degree heterogeneity and symmetry breaking in delocalization.

Jalan and Bandyopadhyay [20] extended random matrix theory (RMT) analysis to a broad range of empirical and model networks. They demonstrated that the nearest-neighbor eigenvalue spacing s follows the Brody distribution,

$$P_{\beta }\left(s\right)=A{s}^{\beta }exp(-B{s}^{\beta +1}),$$

(47)

where the Brody index $\beta \in [0,1]$ interpolates between Poissonian and Wigner–Dyson statistics, thus quantifying the degree of spectral repulsion. Their analysis of the spectral rigidity function ${\Delta }_3\left(L\right)$,

$${\Delta }_3\left(L\right)\sim {\displaystyle \frac{1}{{\pi }^2}}lnL,$$

(48)

confirmed the presence of long-range correlations consistent with Gaussian Orthogonal Ensemble (GOE) predictions, even in topologically disordered systems.

Complementary insights were provided by Samukhin et al. [21] and Dorogovtsev et al. [41], who demonstrated that in uncorrelated configuration models, tree-like local connectivity dominates the bulk spectral density, while high-degree hubs induce isolated spectral outliers. This structural dichotomy—between peripheral disorder and core regularity—gives rise to emergent core-periphery spectral features, explaining systematic deviations from semicircle universality in empirical graphs.

Bordenave et al. [22] introduced the non-backtracking matrix B, constructed on directed edges with reversal exclusion, to circumvent resolution limits of the adjacency spectrum. For sparse, locally tree-like graphs, the spectrum of B converges to a circular law in the complex plane, enabling the refined detection of community structure and mitigating contamination by localized or high-degree artifacts.

Together, these contributions establish spectral statistics as robust indicators of structural class, correlation regime, and mesoscopic organization in random networks. By bridging RMT, walk-based enumeration, and spectral localization theory, they provide a foundational toolkit for spectral taxonomy and structural inference in real-world graph ensembles.

4.3. Synchronization and Collective Dynamics: Spectral Thresholds

The stability of collective oscillations on graphs is now understood to be governed by a compact set of spectral invariants that transcend microscopic details of the coupling function. In the paradigmatic Kuramoto ensemble, with all indices running over the node set $V\left(\mathcal{G}\right)$; $\left|V\right|=N$,

$${\dot{\theta }}_i={\omega }_i+\sigma \sum _{j=1}^N{a}_{ij}sin({\theta }_j-{\theta }_i),$$

(49)

Arenas et al. [42] showed via the Master-Stability-Function (MSF) reduction of Pecora–Carroll that transverse stability is admissible only when the Laplacian-scaled eigenmodes satisfy

$$\Lambda \left(\sigma {\lambda }_i\right)<0,i=2,\dots ,N,$$

(50)

thereby rendering the synchronisability problem algebraic in the algebraic connectivity ${\lambda }_2$ and the eigenratio $R={\lambda }_N/{\lambda }_2$. Restrepo et al. [23] further established that the onset of global phase coherence is controlled solely by the leading adjacency eigenvalue through the thermodynamic threshold

$$k_c={\displaystyle \frac{2}{\pi g\left(0\right){\lambda }_{max}}},$$

(51)

where $g\left(0\right)$ denotes the natural frequency density at its mean. The small-world intuition that short paths guarantee synchrony was overturned by Nishikawa et al. [24], who introduced the inverse-participation ratio

$$IPR\left(v\right)=\sum _{i=1}^N{\left|v_i\right|}^4,$$

(52)

demonstrating that eigenvector localization on hubs inflates R and suppresses coherence even when geodesic distances are minimal. Extending the spectral programme to chaotic substrates, Yang et al. [25] derived an observer-based multilayer Lorenz criterion

$$|c{\lambda }_2|>h_{max},$$

(53)

linking the coupling gain c and algebraic connectivity to the largest observer eigenvalue $h_{max}$ required for stable estimation. Finally, Boccaletti et al. [43] situated these metrics in a broader structural–functional paradigm, showing that ${\lambda }_2$, ${\lambda }_N$ and R simultaneously govern consensus rates, epidemic thresholds ${\tau }_c\propto {\lambda }_1^{-1}$, and diffusion–reaction instabilities across weighted, directed and multilayer architectures. Collectively, Equations (49)–(53) crystallise the minimal spectral coordinate system $({\lambda }_2,{\lambda }_N,\mathrm{IPR})$ that dictates attainability, robustness and energetic cost of network-wide synchrony, thus exposing why purely geometric optimization is incapable of guaranteeing coherent dynamics without concurrent control of spectral dispersion.

4.4. Consensus, Averaging, and Distributed Optimization

The analytical structure of consensus dynamics is spectrally dual to synchronization and is governed by Laplacian flows in both continuous and discrete settings. In the foundational formalism of Olfati-Saber, Fax, and Murray [44], agents with scalar internal states $x_i\left(t\right)\in \mathbb{R}$ evolve under local averaging, inducing a diffusive flow whose convergence to average consensus is exponential in time and asymptotically governed by the algebraic connectivity ${\lambda }_2$ of the graph Laplacian. For directed graphs, the asymmetry of the adjacency matrix leads to weighted equilibria, with consensus values determined by the invariant distribution of the associated Markov chain:

$$\underset{t\to \infty }{lim}{x}_i\left(t\right)=\sum _{j=1}^N{\pi }_j{x}_j\left(0\right),$$

(54)

where $\pi \in {\mathbb{R}}^N$ is the normalized left eigenvector of the stochastic weight matrix corresponding to eigenvalue one. This reveals a spectral duality between random walk stationary distributions and consensus fixed points. The stability of these equilibria under asynchronous updates, communication delays, and switching topologies is established through Lyapunov-based techniques and Perron–Frobenius theory, with convergence rates expressible in terms of the spectral gap ${\lambda }_2$. In [26], this connection was explicitly extended to time-varying graphs and quantized communication, showing that convergence speed remains spectrally constrained even under nonstationary coupling. The theoretical framework was significantly extended by Nedić and Olshevsky [27], who analyzed distributed convex optimization over time-varying directed graphs via the subgradient-push protocol. By employing non-doubly stochastic weight updates and a push-sum consensus mechanism, the method compensates for directional information flow without requiring bidirectional connectivity. When objective functions are convex and subgradients are uniformly bounded, and step sizes scale as ${\gamma }_t\propto 1/\sqrt{t}$, convergence to a network-wide minimizer is achieved at the rate

$$∥x_i\left(t\right)-x^{\ast }∥=\mathcal{O}\left({\displaystyle \frac{logt}{\sqrt{t}}}\right),$$

(55)

under assumptions of uniformly strongly connected graph sequences and persistent minimum influence $\delta >0$ across the communication schedule. This result rigorously embeds consensus within the broader class of distributed online learning algorithms, where spectral contraction in the associated time-dependent weight matrices ensures the asymptotic decay of disagreement. The resulting synthesis enables the deployment of consensus as a structural module for decentralized control, estimation, and optimization, where Laplacian eigenstructure encodes both communication geometry and the dynamic viability of coordination in nonstationary, partially observed, and resource-limited environments.

4.5. Spectral Control Theory and Controllability Frameworks

Canonical controllability theory stipulates that a linear time-invariant (LTI) network $\dot{x}=Ax+Bu$ is controllable when the Kalman rank condition is met, yet Sun and Motter [28] demonstrated that, in high-dimensional sparse graphs, formal sufficiency can fail numerically because the controllability Gramian

$$W={\int }_0^T{e}^{At}B{B}^{\top }{e}^{A^{\top }t}\mathrm{d}t$$

(56)

often possesses an ill-conditioned spectrum whose ratio $\kappa \left(W\right)={\lambda }_{max}\left(W\right)/{\lambda }_{min}\left(W\right)$ magnifies round-off error and renders minimum-energy trajectories non-constructible; the ensuing numerical controllability transition marks a sharp driver-node threshold beyond which feasible inputs emerge, revealing that optimal trajectories are generically nonlocal in phase space and increasingly so as $rankB$ falls below full actuation. To decouple structural and numerical obstacles, Yuan et al. [29] recast controllability via the Popov–Belevitch–Hautus criterion,

$$rank\left(\lambda I-A,B\right)=N\forall \lambda \in spec\left(A\right),$$

(57)

and proved that the minimal number of driver nodes equals the maximal geometric multiplicity

$$N_D=\underset{{\lambda }_i}{max}dimker({\lambda }_{i}I-A),$$

(58)

so that spectral degeneracy, not degree heterogeneity per se, sets a lower bound on control complexity; for symmetric A, the multiplicities are algebraic, implying higher $N_D$ in highly regular or highly symmetric graphs. The structural perspective of Liu et al. [30] thereby generalises to weighted and directed systems, enabling eigenvalue-guided actuator placement and revealing a trade-off between topological regularity, driver-node count, and numerical energy cost. Collectively, Equations (56)–(58) expose a tripartite constraint space—Gramian conditioning, PBH rank, and eigenvalue multiplicity—within which practical controllability must reside, explaining why actuator placement, trajectory synthesis, and robustness evaluation cannot be decoupled from the underlying spectral geometry.

4.6. Multilayer and Tensorial Spectral Formalisms

The emergence of multilayer network theory as a formally autonomous discipline is grounded in the supra-adjacency abstraction of Boccaletti [31], the categorical tensorial representation of Kivelä [32], and the operator-theoretic formalism of Bianconi [33]. Together, these works provide a unifying algebraic and spectral framework for capturing the complex interactions between and within structurally entangled layers. In this representation, the global dynamics of a multilayer system with L layers and N nodes per layer are encoded in a supra-adjacency matrix $\mathcal{A}\in {\mathbb{R}}^{NL\times NL}$, whose block structure places intra-layer adjacency matrices $A^{\left[\alpha \right]}$ on the diagonal and inter-layer coupling terms $C^{\left[\alpha \beta \right]}$ off-diagonally. For diffusive or synchronization processes, the corresponding supra-Laplacian $\mathcal{L}$ takes the form

$$\mathcal{L}=\underset{\alpha =1}{\overset{L}{⨁}}{L}^{\left[\alpha \right]}+\omega \sum _{\alpha \ne \beta }{D}^{\left[\alpha \beta \right]},$$

(59)

where $L^{\left[\alpha \right]}$ denotes the combinatorial Laplacian of layer $\alpha$, $D^{\left[\alpha \beta \right]}$ encodes interlayer diffusion between $\alpha$ and $\beta$, and $\omega$ controls the coupling intensity. The spectral properties of $\mathcal{L}$, particularly its algebraic connectivity and spectral gap, govern the relaxation dynamics, the onset of layered synchronization, and the stability of coherent states under topological perturbations. Bianconi’s formalism extends this by introducing tensorial operators and generalized master stability functions (MSFs), allowing for a bifurcation-theoretic treatment of transitions between weakly and strongly coupled regimes, including the emergence of chimera states and abrupt desynchronization in modular ensembles. Crucially, these spectral thresholds are analytically tractable and directly linked to eigenvalue bounds of $\mathcal{L}$, thus enabling predictive diagnostics for contagion, coherence, and diffusion resilience.

The axiomatization by Kivelä [32] defines a multilayer network as a quadruple $(V,L,E,w)$, where V comprises node-layer tuples, L enumerates layers, E is the edge set with intra- and inter-layer partitions, and $w:E\to \mathbb{R}$ is a weight function. This definition permits seamless generalization to temporal, multiplex, and interdependent systems. The mappings $\pi :V\to N$ and $\lambda :V\to L$ assign tuples to physical nodes and layers, respectively, providing the groundwork for generalizing structural metrics such as degree centrality, betweenness, and community modularity across layer-dependent topologies. Moreover, this categorical abstraction enables classical graph-theoretic notions to be preserved under tensorial transformations and ensures compatibility between spectral and combinatorial perspectives.

Within this framework, Radicchi [34] analyzed spectral phase transitions in symmetric duplex networks under probabilistic interlayer coupling. The onset of interlayer coherence is governed by the critical parameter

$${\nu }^{\ast }=2(1-p),$$

(60)

where $p\in [0,1]$ denotes the probability of interlayer connectivity, and associated threshold bifurcations occur at

$$p_c^{\pm }\approx {\displaystyle \frac{1}{2}}\left(1\pm {\displaystyle \frac{1}{\sqrt{2\langle k\rangle }}}\right),$$

(61)

with $\langle k\rangle$ the mean intralayer degree. These expressions define phase-like transitions between coherent and incoherent cross-layer behaviors and mirror symmetry-breaking bifurcations in thermodynamic systems. Extensions by Solé-Ribalta [35] and De Domenico [45] propose generalized supra-Laplacians that support anisotropic coupling schemes, multiplex percolation processes, and spectral embedding of interdependent contagion processes, thereby enabling a unified treatment of robustness and structural vulnerability under multilayer constraints.

In application to chaotic systems, Yang [25] constructed a multilayer Lorenz oscillator network in which synchronization is achieved through observer-based feedback. The critical threshold is given by the spectral condition

$$|c{\lambda }_2| > h_{max},$$

(62)

where c is the feedback gain, ${\lambda }_2$ is the second-smallest Laplacian eigenvalue of the inter-observer coupling layer, and $h_{max}$ quantifies the maximal observer gain. This result reveals how spectral invariants determine the minimum necessary communication bandwidth and actuator responsiveness in decentralized chaotic control.

Collectively, these developments demonstrate that spectral diagnostics, originally developed for monoplex systems, extend naturally to multilayer networks via supra-Laplacians, tensorial descriptors, and algebraic invariants. The spectrum of $\mathcal{L}$ captures not only the intrinsic geometry of each layer but also the global interaction topology, thereby informing thresholds for diffusion collapse, synchronization breakdown, and structural phase transitions. These insights enable the principled design of control, inference, and optimization strategies for layered systems, establishing multilayer spectral theory as an essential component of contemporary network science.

4.7. Empirical Spectral Models and Inference-Oriented Extensions

A distinct trajectory within spectral network theory has emerged through empirical and inference-oriented approaches that integrate structurally informed generative processes with data-driven regularities. Ureña-Carrión et al. [46] reconceptualize multiplexity in social systems not as a preordained stratification of interaction types but as a latent geometric unfolding induced by temporal regularities in dyadic communication patterns. Using large-scale call detail records, they apply orthogonal non-negative matrix factorization to extract L latent temporal bases over T time intervals, with each tie i probabilistically assigned to an activity layer via a multinomial model:

$$X_i\sim \mathrm{Multinomial}(w_i,H{\alpha }_i),$$

(63)

where $X_i\in {\mathbb{R}}^T$ is the observed frequency vector, $w_i$ the total activity of tie i, $H\in {\mathbb{R}}^{T\times L}$ the matrix of temporal components, and ${\alpha }_i\in {\mathbb{R}}^L$ the corresponding membership weights, thereby inducing a probabilistic multilayer partition from temporally aggregated data. Zhang et al. [47] construct a three-layer Internet model encoding physical infrastructure, service architecture, and user-level interactions, and quantify systemic vulnerability via the post-deletion size $S\left(D\right)$ of the largest connected component under targeted attack scenarios, revealing asymmetrical cross-layer failure cascades driven by infrastructural perturbations. Liu et al. [48] formalize multilayer epidemic dynamics with activity-driven activation and interlayer co-participation, yielding an epidemic threshold in the fully coupled limit expressed as

$${\lambda }_c={\displaystyle \frac{\mu }{{\sum }_x{m}_x\langle a_x\rangle +\sqrt{{\sum }_x{m}_x^2\langle a_x^2\rangle +{\sum }_{x<y}2m_x{m}_y\langle a_x{a}_y\rangle }}},$$

(64)

where $\mu$ denotes the recovery rate, $a_x$ the activity rate in layer x, and $m_x$ the average degree per activation, with expectations computed over empirical activity distributions, thereby capturing nonlinearly amplified epidemic persistence induced by interlayer co-activation. Qiang et al. [36] analyze weakly coupled percolation in interdependent networks using self-consistent generating function equations, yielding a giant component fraction x determined by

$$x=p\left[1-G_1^{\left(A\right)}(1-x)\right]\left[1-G_0^{\left(B\right)}(1-y)\right]+\alpha p\left[1-G_1^{\left(A\right)}(1-\alpha x)\right]G_0^{\left(B\right)}(1-y),$$

(65)

where p is the fraction of retained links and $\alpha \in [0,1]$ the interdependence strength, with symmetric equations for y, thus elucidating discontinuous phase transitions under partial structural coupling. The inferential coherence of such models is underpinned by the projectibility condition formalized by Shalizi and Rinaldo [49] for exponential families, which requires that for nested graph domains $A\subset B$, the marginal satisfies

$$P_{A,\theta }=P_{B,\theta }\circ {\pi }_{B\to A}^{-1},$$

(66)

and holds if and only if the sufficient statistics admit separable increments, implying that models relying on nonlocal statistics such as motif counts or clustering fail to meet consistency criteria across subnetworks unless appropriately regularized. In this context, Wuyts and Sieber [37] derive automated closures of moment hierarchies by expressing higher-order motif expectations through junction-tree decompositions:

$$\langle \left[x^a\right]\rangle \approx {\displaystyle \frac{{\prod }_j\langle \left[x^{J_j}\right]\rangle }{{\prod }_{j=2}^{\left|J\right|}\langle \left[x^{J_j\cap \mathrm{pa}\left(J_j\right)}\right]\rangle }},$$

(67)

with each $J_j$ a variable cluster and $\mathrm{pa}\left(J_j\right)$ its parent in the tree, thereby reconciling motif closures with probabilistic graphical factorization. Complementarily, Barzel and Barabási [50] characterize perturbation-response universality via power-law scaling of a spectral propagation matrix G such that

$$P\left(G\right)\sim G^{-\nu },\nu ={\displaystyle \frac{\beta +2}{\beta +1}},\omega ={\displaystyle \frac{\varphi }{\beta +1}},$$

(68)

where $\beta$ captures decay sensitivity and $\varphi$ the external forcing, revealing coarse-grained universality classes analogous to those in thermodynamic critical phenomena. Cai et al. [38] propose an entropy index tailored to series-parallel topologies,

$$H_{\mathrm{SP}}=-\sum _{k=1}^N{I}_{k}log{I}_k,$$

(69)

where $I_k$ integrates node-level radial and medial centralities, generalizing classical entropy by capturing flow-critical positions in communication and transport networks. Zhang et al. [51] introduce a deletion-augmented generative model with power-law stationary degree distribution

$$P\left(k\right)\propto k^{\alpha },\alpha =-{\displaystyle \frac{3+4q}{1-2q}},$$

(70)

where $q\in (0,0.5)$ is the node deletion probability, thereby recovering empirical deviations from pure preferential attachment through a structurally parsimonious mechanism. Vegué et al. [52] propose a coarse-grained spectral reduction scheme for directed modular systems, aggregating node dynamics within module $G_{\nu }$ via

$$X_{\nu }=\sum _{i\in G_{\nu }}{a}_{\nu i}{x}_i,$$

(71)

and deriving approximate macrodynamics as

$${\dot{X}}_{\nu }\approx f\left(X_{\nu }\right)+\sum _{\rho }{W}_{\nu \rho }g(X_{\nu },X_{\rho })+\sum _{\rho }({\mu }_{\nu \rho }-W_{\nu \rho })g_1(X_{\nu },X_{\rho })X_{\nu },$$

(72)

where f, g, and $g_1$ capture intra- and intermodular effects, and ${\mu }_{\nu \rho }$ is the homogenized coupling. This projection preserves dynamical bifurcations and enables analytical tractability in complex systems. Taken together, these contributions establish an empirically grounded spectral theory in which entropy, response scaling, motif closure, and projectibility constitute macroscopic invariants of mesoscopic organization and suggest the emergence of a network thermodynamics wherein spectral curvature and eigenvalue dispersion play roles analogous to potential energy landscapes and thermodynamic observables.

The convergence of spectral graph theory with multilayer modeling, controllability analysis, and inference-oriented statistics suggests that a genuinely unified spectral thermodynamics of networks is within reach, in which eigenvalue gaps, spectral radii, and localization indices play roles analogous to energy, temperature, and order parameters in statistical physics.

Progress will require three complementary advances: (i) analytical extension of supra tensor operators to hypergraphs and higher-order simplicial complexes, thereby capturing nonpairwise interactions that dominate biological and socio–technical systems; (ii) scalable algorithms for Gramian conditioning and eigenvector localization that remain robust under streaming data and adversarial perturbations, enabling real-time control and inference in evolving infrastructures; (iii) statistically consistent coupling of empirical layer extraction, as in latent activity models, with projectible generative families that respect sampling constraints and guarantee out-of-sample validity. Meeting these challenges will not only close the methodological loop sketched in Figure 4 but will also furnish the predictive and design capabilities demanded by next-generation neuro-, cyber-, and energy networks, where spectral asymptotics, tensor calculus, and data-driven inference must coalesce into a single mathematically rigorous and computationally tractable framework.

5. Random Walks and Diffusion

Diffusion processes on networks constitute a unifying framework for modeling transport, contagion, synchronization, and belief dynamics across physical, biological, and socio-technical systems. Unlike static topological analyses, the study of random walks and spreading dynamics captures the interplay between structural constraints and temporally evolving interactions, requiring tools that combine probabilistic reasoning, spectral insights, and causal architectures. In this section, we synthesize five conceptually distinct but methodologically interconnected strands of research that collectively establish a progressively deepening theoretical landscape, from temporally grounded definitions of causality and reachability to optimally controlled diffusion in multilayered and interdependent structures. Each of these strands corresponds to a distinct analytical regime characterized by its governing assumptions, model complexity, and degree of systemic integration. The diagram in Figure 5 delineates this logical progression, which not only structures the present review but also clarifies the implicit hierarchy of modeling commitments embedded in the literature.

Figure 5 formalizes the organizing logic of Section 5 by arranging the reviewed literature along a unidirectional continuum whose abscissa is calibrated to analytical depth and concomitant systemic scope. At the leftmost extreme, Holme–Saramäki’s synthesis of time-respecting paths provides the foundational causal scaffold: diffusion is constrained solely by the chronological ordering of events, and closed-form latency metrics can be derived without imposing parametric growth rules or multiplexity. Progressing rightward, the asymmetric preferential attachment model of Zhao et al. introduces stochastic growth and developmental heterogeneity, thereby adding a mesoscopic layer of structure while retaining explicit formulae for degree trajectories. The third tier, captured by the exactly solvable continuous time Markovian constructions of Gilboa-Freedman and Hassin, replaces phenomenological attachment with a fully Markovian generator on the state space of agent configurations, yielding recursive expressions for degree attainment and propagation time that remain tractable despite the enlarged configuration space. Granell et al. occupy the fourth position by coupling dual contagion processes across stacked layers; analytic treatment now hinges on spectral criteria, eigenvalue modulated thresholds that integrate structural parameters of two distinct layers, marking a qualitative leap in systemic interaction.

Finally, Zhong’s optimal control framework closes the sequence by embedding multilayer diffusion within a Pontryagin-optimized cost landscape: analytical insight persists, yet only through adjoint equations linked to high-dimensional state vectors and control schedules, emblematic of the field’s current frontier where theory, data, and optimization coalesce. The left-to-right orientation thus visualizes an irreversible accumulation of modeling ingredients, such as temporality, growth bias, stochastic concurrency, multiplex coupling, and controlled feedback, each subsuming the previous as a limiting case and collectively delineating the trajectory from elementary causal walks to policy driveable multilayer diffusion.

5.1. Time-Respecting Paths and Temporal Metrics

Holme and Saramäki’s review [53] elevates time to a first-class structural dimension by encoding each interaction as an event $(i,j,t)$ and introducing the binary adjacency index

$$a(i,j,t)=\left\{\begin{array}{cc}1,\hfill & \mathrm{if}\mathrm{an}\mathrm{edge}(i,j)\mathrm{is}\mathrm{active}\mathrm{at}t,\hfill \\ 0,\hfill & \mathrm{otherwise},\hfill \end{array}\right.$$

(73)

so that classical graph measures become temporal functionals of $a(i,j,t)$. Time-respecting paths are then defined as sequences $\{(v_0,v_1,t_1),\dots ,(v_{m-1},v_m,t_m)\}$ with strictly increasing time stamps, and their durations furnish causal distances whose minima yield latencies ${\lambda }_{i,t}\left(j\right)$ that replace geodesic lengths in all centrality constructs, for instance temporal closeness, is redefined as

$$C_C(i,t)={\displaystyle \frac{N-1}{\sum _{j\ne i}{\lambda }_{i,t}\left(j\right)}},$$

(74)

which naturally generalizes shortest-path averaging to the time domain. Because temporal ordering destroys path transitivity, reachability becomes time-dependent; the average fraction of vertices accessible from i at time t via finite-latency paths,

$$R\left(t\right)={\displaystyle \frac{1}{N}}\sum _{i=1}^N{\displaystyle \frac{\left|\{j\ne i:{\lambda }_{i,t}\left(j\right)<\infty \}\right|}{N-1}},$$

(75)

defines the system’s causal horizon and exposes periods of topological fragmentation. Event sequences display heavy-tailed inter-contact intervals whose coefficient of burstiness,

$$B={\displaystyle \frac{\sigma -\mu }{\sigma +\mu }},$$

(76)

with $\mu$ and $\sigma$ denoting mean and standard deviation, systematically slows diffusion by elongating waiting times at vertices, thereby invalidating Laplacian-based predictions derived for Poissonian contacts. The review further codifies temporal centralities (betweenness, eigenvector, Katz) via vector-clock updates that propagate centrality mass along events and introduces temporal motifs and equivalence classes of $\Delta t$-connected subgraphs to capture higher-order temporal regularities beyond pairwise causality. Collectively, these constructs demonstrate that, once the time coordinate is embedded into the network itself, classical spectral and walk-based intuitions must be replaced by latency tensors, burstiness corrections, and causal reachability domains, establishing a formal machinery for analyzing dynamics on systems where interaction timing is as critical as topology.

5.2. Generative Modeling of Temporal Degree Asymmetries

Zhao et al. [54] advance temporal generative modeling by reconstructing the ontogeny of the C. elegans connectome through an asymmetric preferential attachment rule that decouples inbound and outbound attractiveness, thereby recovering the empirically observed divergence between in- and out-degree distributions across larval stages. At each developmental epoch a newly formed synapse selects its source neuron i with probability

$${\Pi }_i^{\mathrm{out}}={\displaystyle \frac{k_i^{\mathrm{out}}+b}{{\sum }_j(k_j^{\mathrm{out}}+b)}},$$

(77)

and its target j with probability

$${\Pi }_j^{\mathrm{in}}={\displaystyle \frac{k_j^{\mathrm{in}}+a}{{\sum }_j(k_j^{\mathrm{in}}+a)}},$$

(78)

where $k_i^{\mathrm{out}}$ and $k_j^{\mathrm{in}}$ denote present out- and in-degrees, while $a,b>0$ tune the baseline affinities of dendritic and axonal growth cones. Simultaneously, existing links are pruned with probability p, yielding a mean-field kinetic equation for the in-degree of node i,

$${\displaystyle \frac{\mathrm{d}{k}_i^{\mathrm{in}}}{\mathrm{d}t}}=(1-p){\displaystyle \frac{k_i^{\mathrm{in}}+a}{E\left(t\right)+Na}}-{\displaystyle \frac{p}{E\left(t\right)}},$$

(79)

with $E\left(t\right)=(1-2p)t$ the total edge count and N the fixed neuron set, whose solution

$$k_i^{\mathrm{in}}\left(t\right)=\left(1+a\right){\left({\displaystyle \frac{Na+t}{Na+t_i}}\right)}^{{\displaystyle \frac{1-2p}{1-p}}}-a$$

(80)

demonstrates power-law ageing modulated by deletion pressure; the out-degree follows an analogous expression with $a\mapsto b$. As $p\to 0$, the model collapses to a bias-controlled Yule process, whereas finite p embeds synaptic competition and reproduces the sub-exponential tails and increasing in-out asymmetry reported in the developmental datasets. By calibrating $(a,b,p)$ to time-stamped connectomic snapshots, the framework reconciles micro-scale growth bias with macro-scale degree heterogeneity, positioning asymmetric preferential attachment as the temporal analogue of classic Barabási–Albert growth while furnishing a tractable bridge between developmental neurobiology and stochastic network theory.

5.3. Exactly Solvable Continuous-Time Walk-Formation Models

Two analytically tractable models by Gilboa-Freedman and Hassin [55,56] constitute a minimalistic yet expressive framework for modeling network formation and information propagation under continuous-time Markovian dynamics. In the first model [55], the network evolves via stochastic interactions among N agents, each represented by an independent continuous-time Markov chain with two states: meeting (M) and leaving (L). Transitions $M\to L$ and $L\to M$ occur with rates $\mu$ and $\lambda$, respectively. An undirected edge between two agents i and j is created upon their first simultaneous occupation of state M, thus generating a growing acquaintance graph $G\left(t\right)$. The system’s macroscopic behavior is controlled by the dimensionless parameter $\rho =\lambda /\mu$, which governs the relative frequency of meetings. To characterize the structural evolution, the authors derive exact expressions for the expected time $T\left(h\right)$ until a designated leader node acquires degree h. This is achieved through a recursive formulation over a state space $S_{i,m,\ell }$, where $i\in \{0,1\}$ indicates the leader’s current state (L or M), and m, ℓ denote the number of unacquainted non-leaders in states M and L, respectively. The key recursion for the expected meeting time $M_{\ell }$ from state $S_{1,0,\ell }$ to $S_{1,0,0}$ is given by

$$M_{\ell }={\displaystyle \frac{1}{\mu +\ell \lambda }}+{\displaystyle \frac{\ell \lambda }{\mu +\ell \lambda }}{M}_{\ell -1}+{\displaystyle \frac{\mu }{\mu +\ell \lambda }}{L}_{0,\ell },$$

(81)

and similarly the recursion for the auxiliary function $L_{m,\ell }$ from $S_{0,m,\ell }$ to $S_{1,0,0}$ reads

$$L_{m,\ell }={\displaystyle \frac{1}{m\mu +(\ell +1)\lambda }}+{\displaystyle \frac{\ell \lambda }{m\mu +(\ell +1)\lambda }}{L}_{m+1,\ell -1}+{\displaystyle \frac{m\mu }{m\mu +(\ell +1)\lambda }}{L}_{m-1,\ell +1}+{\displaystyle \frac{\lambda }{m\mu +(\ell +1)\lambda }}{M}_{\ell }.$$

(82)

By embedding this process in a reduced subspace and applying layered summations, the authors establish closed-form solutions for the degree growth of the leader, exhibiting heavy-tailed distributions, small-world diameters, and nontrivial clustering features aligning with empirical network structures.

In their second contribution [56], the same authors model information propagation in a population of mobile agents governed by the same two-state Markov dynamics. Initially, one agent is informed; information is transmitted only when the informed and uninformed agents are simultaneously present in state M. The model thus becomes a minimal Markovian rumour-spreading process. The expected time $T_{k,N-k}$ to inform all N agents from an initial seed of k is bounded analytically using harmonic-series approximations. For the case $k=1$, the expected completion time is asymptotically bounded by

$$T_{1,N-1}\approx H_N\sim logN,$$

(83)

where $H_N$ is the N-th harmonic number. This yields a nonmonotonic dependency on N, with propagation times decreasing for intermediate sizes before saturating logarithmically. The process remains exactly solvable due to the minimal configuration space and memoryless transition structure. These studies provide rare closed-form insights into the coevolution of structure and function under Markovian assumptions, demonstrating how complex topologies and efficient diffusion emerge from simple decentralized timing rules.

5.4. Multiplex-Coupled Diffusion and Eigenvalue-Modulated Thresholds

Granell, Gómez, and Arenas [57] demonstrated that when a susceptible–infected–susceptible (SIS) contagion on a physical layer coevolves with an unaware–aware–unaware (UAU) information cycle on a virtual layer, the epidemic threshold is no longer an intrinsic property of the physical network but a function of the awareness dynamics: within a microscopic Markov-chain approximation each node occupies one of three composite states $\{\mathrm{US},\mathrm{AS},\mathrm{AI}\}$ and the effective SIS onset is displaced to

$${\beta }_U^c={\displaystyle \frac{\mu }{{\Lambda }_{max}\left(H\right)}},H_{ji}=\left[1-(1-\delta )p_i^A\right]b_{ji},$$

(84)

where $\mu$ is the recovery rate, $b_{ji}$ the contact matrix of the epidemic layer, $p_i^A$ the stationary awareness probability from the UAU process, $\delta$ the forgetting rate, and ${\Lambda }_{max}\left(H\right)$ the leading eigenvalue of the modulated infection operator; the curve ${\beta }_U^c\left(\lambda \right)$ thus bends at the metacritical value ${\lambda }_c$ of the virtual-layer transmissibility, below which the classical SIS threshold is recovered. The spectral duplex theory of Radicchi [34] (Equations (60) and (61)) generalises this picture to symmetric two-layer systems, revealing discontinuous transitions driven by eigengap closure when cross-layer coherence surpasses a density-dependent critical value, while Qiang et al. [36] extend the analysis to weakly coupled bond percolation, showing that partial dependency between layers reshapes the percolation landscape through a pair of self-consistent generating-function equations previously introduced in Section 4. Collectively, these results establish that multiplex and interdependent spreading processes are governed by eigenvalue-modulated thresholds whose critical surfaces interpolate between single-layer epidemic theory and strongly coupled failure cascades, thereby exposing control levers, such as awareness rate, interlayer coupling, and dependency strength, that can be tuned to delay or even suppress large-scale outbreaks.

5.5. Entropy-Driven Control and Inference-Consistent Coarse Graining

Zhong et al. [58] recast rumour mitigation on a media–social multiplex as an optimal control problem by coupling a three-state virtual-media layer $(M_1,M_2,M_3)$ with a five-compartment real layer $(S,H,I_1,I_2,R)$ and minimising the quadratic cost

$$J\left(u\right)={\displaystyle \frac{1}{2}}{\int }_{t_0}^{t_f}\left[A_1{M}_2\left(t\right)+A_2{I}_1\left(t\right)+\frac{1}{2}{B}_1{u}_1^2\left(t\right)+\frac{1}{2}{B}_2{u}_2^2\left(t\right)+\frac{1}{2}{B}_3{u}_3^2\left(t\right)\right]\mathrm{d}t,$$

(85)

subject to the bilayer dynamics and Pontryagin adjoint equations; the resulting bang–bang arcs reveal that optimally timed media refutations can suppress peak prevalence at a spectral energy cost that scales with the dominant eigenvalue of the Jacobian evaluated at the disease-free equilibrium. Complementarily, Barzel and Barabási [50] derive a universal response kernel by showing that the distribution of node-level perturbation gains encoded in a propagation matrix G obeys the power law

$$P\left(G\right)\sim G^{-\nu },\nu ={\displaystyle \frac{\beta +2}{\beta +1}},\omega ={\displaystyle \frac{\varphi }{\beta +1}},$$

(86)

where $\beta$ measures decay sensitivity and $\varphi$ the magnitude of the driving field, thus partitioning networks into dissipative $(\nu >2)$ and conservative $(\nu \le 2)$ universality classes. Statistical validity of any reduction is constrained by the projectibility theorem of Shalizi and Rinaldo [49], which demands that, for every node set inclusion $A\subset B$, an exponential-family model satisfy

$$P_{A,\theta }=P_{B,\theta }\circ {\pi }_{B\to A}^{-1},$$

(87)

if and only if its sufficient statistics possess separable increments, thereby invalidating naïve motif counts and motivating the factorized moment closure of Wuyts and Sieber [37],

$$〈\left[x^a\right]〉\approx {\displaystyle \frac{{\displaystyle \prod _j〈\left[x^{J_j}\right]〉}}{{\displaystyle \prod _{j=2}^{\left|J\right|}〈\left[x^{J_j\cap \mathrm{pa}\left(J_j\right)}\right]〉}}},$$

(88)

where $\left\{J_j\right\}$ indexes the clusters of a junction tree and $\mathrm{pa}\left(J_j\right)$ denotes the parent of $J_j$, guaranteeing compatibility with graphical decomposability. Entropy-based diagnostics are supplied by Cai’s series–parallel index

$$H_{\mathrm{SP}}=-\sum _{k=1}^N{I}_{k}log{I}_k,$$

(89)

with node importance $I_k$ synthesizing radial and medial centralities to expose flow-critical articulations in infrastructure graphs, whereas structural evolution under attrition is captured by Zhang’s birth–death rule

$$P\left(k\right)\propto k^{\alpha },\alpha =-{\displaystyle \frac{3+4q}{1-2q}},q\in (0,0.5),$$

(90)

which reconciles truncated heavy tails with a single deletion parameter q. For model reduction, Vegué et al. [52] define module observables

$$X_{\nu }=\sum _{i\in G_{\nu }}{a}_{\nu i}{x}_i,$$

(91)

and show that, under the homogeneity assumption of intramodule fluctuations, the coarse-grained dynamics satisfy

$${\dot{X}}_{\nu }\approx f\left(X_{\nu }\right)+\sum _{\rho }{W}_{\nu \rho }g(X_{\nu },X_{\rho })+\sum _{\rho }\left({\mu }_{\nu \rho }-W_{\nu \rho }\right)g_1(X_{\nu },X_{\rho })X_{\nu },$$

(92)

where $W_{\nu \rho }$ is the empirical coupling and ${\mu }_{\nu \rho }$ its homogenized surrogate; the error incurred is $O\left({\sigma }_{\mathrm{intra}}^2\right)$, preserving bifurcation structure to first order. Together, Equations (85)–(92) articulate a coherent pipeline in which entropy metrics locate bottlenecks, response scaling predicts macroscopic sensitivity, projectible statistics safeguard inference under subsampling, and optimal control calculus exploits coarse-grained spectra to minimize diffusion cost, thereby unifying control, diagnostics, and statistically disciplined reduction for high-dimensional multilayer networks.

The reviewed sequence of models, from causality-aware path formulations to control-driven multilayer interventions, demonstrates the increasing capacity of network diffusion theory to incorporate empirical realism, probabilistic tractability, and policy-relevant design. Yet substantial challenges remain. First, the integration of non-Markovian memory effects into temporally evolving and multilayered settings requires analytical frameworks that transcend current state-transition or closure-based models. Second, empirical validation of high-dimensional dynamical models remains constrained by limitations in temporal resolution, observation sparsity, and structural sampling bias. Third, the development of universal approximations, analogous to hydrodynamic limits in statistical physics, for non-equilibrium network diffusion under local constraints remains an open problem of both mathematical and practical relevance. Future research will likely converge on hybrid modeling paradigms that unify generative mechanisms, structural symmetries, and optimal intervention design, supported by scalable inference tools and constrained by realistic data availability. The eventual aim is a comprehensive theory of network diffusion that is both analytically rigorous and operationally deployable across domains ranging from neuroscience to misinformation control.

6. Probabilistic Inference and Message Passing

This section develops a unified conceptual framework for probabilistic inference on networks, grounded in message-passing dynamics, variational approximations, and spectral criteria. Our goal is to articulate how local update rules, structurally constrained factorization, and algebraic inference operators interact to support scalable, statistically sound reasoning in high-dimensional relational systems. The progression begins with models of opinion aggregation that formalize how agents update beliefs through adaptive weighting of self and social information, laying the groundwork for quantifiable influence estimation in empirical, structurally evolving networks.

Building upon this, the section introduces the algorithmic core of belief propagation in Gaussian and discrete settings, encompassing its multivariate generalizations, its applications to coupled epidemic-information diffusion, and its spectral boundaries for recovery. A variational lens is then introduced, casting message-passing schemes as energy minimization over structured marginal distributions and linking decentralized update rules to global statistical coherence through tractable approximations of marginal distributions. This leads naturally to the formulation of sharp statistical thresholds for inference validity: the projectibility criterion that governs model behavior under marginalization, and the detectability phase transition that separates statistically recoverable latent structures from those that are indistinguishable from noise. Finally, we extend beyond graph-specific and compartmental transmission schemes to structurally independent representations of influence and propagation, quantifying influence through probabilistic path summation and memory-aware attention dynamics. As diagrammed in Figure 6, these contributions form a logical trajectory from empirical aggregation and algorithmic inference to theoretical limits and structurally unanchored influence representations, collectively delineating the operational boundaries of learning, propagation, and control in networked systems.

6.1. Opinion Aggregation and Adaptive Influence

The interplay between individual conviction and adaptive social feedback forms a foundational mechanism in modern models of belief dynamics on networks. Almaatouq et al. [59] introduced an empirical framework in which individual agents update their beliefs by adaptively weighting self-confidence against the influence of peer information, operationalized through the weight-on-self (WOS) metric defined as

$$\mathrm{WOS}={\displaystyle \frac{u_2-m}{u_1-m}},$$

(93)

where $u_1$ and $u_2$ denote the individual’s belief before and after exposure to peer beliefs, respectively, and m is the arithmetic mean belief of their neighbors. This metric captures the relative resilience of an agent’s opinion to external influence, effectively serving as an index of endogenous versus exogenous anchoring in belief formation. Agents exhibiting high WOS are less susceptible to information cascades and better preserve belief heterogeneity, while the collective impact of such heterogeneity is filtered through a dynamically adaptive network topology that rewires based on inferred agent credibility. The model extends the classical DeGroot averaging framework by coupling belief updates with structural plasticity, where edge formation is governed by prior estimation error and signal fidelity, ultimately promoting convergence toward crowd-level optimal beliefs in heterogeneous populations. In contrast to static formulations, De et al. [60] proposed the SLANT model, which introduces a latent, continuous-time stochastic framework capturing the coupled evolution of opinions and message intensities. The latent opinion vector $x^{\ast }\left(t\right)$ evolves under the stochastic differential system

$$d{x}^{\ast }\left(t\right)=-\omega (x^{\ast }\left(t\right)-\alpha )dt+A·(m\left(t\right)\odot dN\left(t\right)),d{\lambda }^{\ast }\left(t\right)=-\nu ({\lambda }^{\ast }\left(t\right)-\mu )dt+B·dN\left(t\right),$$

(94)

where $\alpha$ and $\mu$ are baseline vectors for opinions and message intensities, A and B are influence matrices, $\omega$ and $\nu$ are decay parameters, $dN\left(t\right)$ is a marked counting process of events, and $m\left(t\right)$ encodes the message polarity with ⊙ denoting the elementwise product. The model satisfies the Markov property and captures a spectrum of endogenous-exogenous feedback loops in opinion expression, enabling rigorous statistical estimation via maximum likelihood. Closed-form solutions for conditional opinion expectations $\mathbb{E}\left[x^{\ast }\left(t\right)\right|\mathcal{H}\left(t_0\right)]$ can be derived in the case of Poisson and Hawkes intensities, revealing explicit conditions for consensus or polarization. In particular, for Poisson intensities with rate vector $\mu$, the long-term average opinion converges to

$$\underset{t\to \infty }{lim}\mathbb{E}\left[x^{\ast }\left(t\right)\right|\mathcal{H}\left(t_0\right)]={\left(I-A·diag{\left[\mu \right]}^{-1}{\omega }^{-1}\right)}^{-1}\alpha ,$$

(95)

provided the spectral radius condition $\rho (A·diag{\left[\mu \right]}^{-1})<\omega$ is met. Analogous conditions hold for Hawkes intensities with time-varying excitation governed by B, yielding convergence to

$$\underset{t\to \infty }{lim}\mathbb{E}\left[x^{\ast }\left(t\right)\right|\mathcal{H}\left(t_0\right)]={\left(I-A·diag\left[{\left(I-B/\nu \right)}^{-1}\mu \right]{\omega }^{-1}\right)}^{-1}\alpha ,$$

(96)

when the system is dynamically stable. Together, these frameworks articulate complementary static and temporal strategies for aggregating opinions in adaptive networked environments, demonstrating how learning rates, memory decay, and structural reactivity critically shape belief landscapes over time.

6.2. Belief Propagation and Message Passing

Belief propagation and its algorithmic generalizations constitute a foundational methodology for inference in structured probabilistic models, especially in the presence of sparsity and partial observability. The multivariate extension GaBP-m proposed by Kamper, Steel, and du Preez [61] enables distributed inference in Gaussian graphical models with vector-valued nodes by decomposing the joint density $X\sim \mathcal{N}(\mu ,\Sigma )$ into local potentials via

$$f\left(x\right)\propto \prod _i{\varphi }_i\left(x_i\right)\prod _{(i,j)\in E}{\psi }_{ij}(x_i,x_j),$$

(97)

with

$${\varphi }_i\left(x_i\right)=exp\left(-{\displaystyle \frac{1}{2}}{x}_i^{\top }{S}_{ii}{x}_i+x_i^{\top }{b}_i\right),$$

(98)

$${\psi }_{ij}(x_i,x_j)=exp\left(-x_i^{\top }{S}_{ij}{x}_j\right).$$

(99)

The message passing equations involve matrix and vector updates:

$$Q_{ij}^{(n+1)}=-S_{ji}{\left(P_{ij}^{\left(n\right)}\right)}^{-1}{S}_{ij},v_{ij}^{(n+1)}=-S_{ji}{\left(P_{ij}^{\left(n\right)}\right)}^{-1}{z}_{ij}^{\left(n\right)},$$

(100)

where

$$P_{ij}^{\left(n\right)}=S_{ii}+\sum _{t\in \mathcal{N}\left(i\right)\setminus j}{Q}_{ti}^{\left(n\right)},z_{ij}^{\left(n\right)}=b_i+\sum _{t\in \mathcal{N}\left(i\right)\setminus j}{v}_{ti}^{\left(n\right)}.$$

(101)

The estimated beliefs are recovered from

$$P_i^{\left(n\right)}=S_{ii}+\sum _{j\in \mathcal{N}\left(i\right)}{Q}_{ji}^{\left(n\right)},{\mu }_i^{\left(n\right)}={\left(P_i^{\left(n\right)}\right)}^{-1}\left(b_i+\sum _{j\in \mathcal{N}\left(i\right)}{v}_{ji}^{\left(n\right)}\right).$$

(102)

Convergence is ensured under a generalized walk-summability condition, defined via a diagonal preconditioning matrix $\Lambda$ such that $\Lambda S\Lambda$ is walk-summable. Beyond the purely algorithmic implementation, Yedidia, Freeman, and Weiss [62] provided a variational interpretation of belief propagation by proving that its fixed points correspond to stationary points of the Bethe free energy functional defined over approximate node and edge marginals $\{b_i\left(x_i\right),b_{ij}(x_i,x_j)\}$. The Bethe functional is given by

$$F_{\mathrm{Bethe}}\left(b\right)=\sum _{(i,j)\in E}\sum _{x_i,x_j}{b}_{ij}(x_i,x_j)log{\displaystyle \frac{b_{ij}(x_i,x_j)}{{\psi }_{ij}(x_i,x_j)b_i\left(x_i\right)b_j\left(x_j\right)}}+\sum _i(1-d_i)\sum _{x_i}{b}_i\left(x_i\right)log{b}_i\left(x_i\right),$$

(103)

where ${\psi }_{ij}$ is the pairwise potential and $d_i$ the degree of node i. This formulation casts message updates as gradient steps toward minimizing a global energy under marginalization constraints, justifying belief propagation as an approximate variational inference method even on loopy graphs. It also provides the foundation for generalized belief propagation (GBP) by extending the Bethe approximation to higher-order Kikuchi clusters, thereby improving inference accuracy in models with cycles. This variational perspective links BP to broader optimization principles and complements later results on spectral convergence and inference limits.

In a distinct regime of application, Granell et al. [57] study multiplex message passing by coupling an SIS epidemic process with a cyclic unaware-aware-unaware (UAU) information diffusion on a dual-layer network. The interdependence induces a metacritical regime wherein the effective epidemic threshold ${\beta }_U^c$ is reduced below classical predictions and governed by

$${\beta }_U^c={\displaystyle \frac{\mu }{{\Lambda }_{\max }\left(H\right)}},H_{ji}=[1-(1-\delta )p_i^A]b_{ji},$$

(104)

where $p_i^A$ denotes awareness probability and H encodes modulated infectivity. This hybridization redefines the phase diagram of information-epidemic co-dynamics. At the limit of inferential complexity, the recovery of latent structures in sparse graphs is governed by spectral belief propagation. Coja-Oghlan et al. [63] establish that inference on $(d,\epsilon )$-regular graphs is possible due to the tree-like local structure, as the second eigenvalue of the adjacency matrix remains bounded. The message update on edge $(v\to w)$ is

$${\eta }_{v\to w}^a={\displaystyle \frac{{\prod }_{u\in N\left(v\right)\setminus \left\{w\right\}}(1-{\eta }_{u\to v}^a)}{{\sum }_{b=1}^3{\prod }_{u\in N\left(v\right)\setminus \left\{w\right\}}(1-{\eta }_{u\to v}^b)}},$$

(105)

where ${\eta }_{v\to w}^a$ estimates the marginal for color a. The convergence theorem guarantees that for sufficiently high d, the BPCol algorithm yields a correct coloring with nonvanishing probability. In the setting of stochastic block models, Decelle et al. [64] identify a spectral detectability phase transition based on the non-backtracking matrix eigenvalue ${\lambda }_2$:

$$|{\lambda }_2| >\sqrt{c},$$

(106)

with c the average degree. Posterior marginals ${\mu }_{i\to j}\left(a\right)$ are recursively updated as

$${\mu }_{i\to j}\left(a\right)\propto n_{a}exp(-h_a)\prod _{k\in \partial i\setminus j}\sum _b{c}_{ab}{\mu }_{k\to i}\left(b\right),$$

(107)

and the inference is characterized by variational free energy

$$f=-{\displaystyle \frac{1}{N}}\sum _{i}log{Z}_i+{\displaystyle \frac{1}{N}}\sum _{(i,j)\in E}log{Z}_{ij}.$$

(108)

This delineates undetectable, hard, and easy regimes, with inference feasibility determined by the spectral structure of the underlying graph.

6.3. Limits of Inference and Projectibility

A rigorous understanding of when statistical inference on networks is well-posed demands explicit criteria that delineate the boundary between valid parameter estimation and intrinsic impossibility. Two results furnish such boundaries by formalizing, respectively, structural coherence and inferential detectability in probabilistic network models. The projectibility theorem of Shalizi and Rinaldo [49], previously introduced in Equation (87), establishes that an exponential-family model can be coherently restricted to subgraphs only if its sufficient statistics exhibit separable increments, thereby rendering popular motif-based features such as triangle densities or transitivity invalid for extrapolation across graph sizes. In practice, this constraint restricts the domain of statistical learning to either specially constructed statistics or fully observed networks, as naïve marginalization otherwise leads to parameter inconsistency. Notably, the failure of projectibility invalidates the transfer of learned interactions from a sample graph to the full population, unless the sampling mechanism and induced dependency structure are explicitly modeled.

From a complementary probabilistic perspective, Mossel, Neeman, and Sly [65] identify a sharp phase transition in the stochastic block model that distinguishes regimes of statistical identifiability from those of fundamental non-recoverability. Considering a symmetric SBM with two equal-sized communities, where intra- and inter-community edge probabilities scale as $a/n$ and $b/n$ respectively, they prove that no algorithm, regardless of computational power, can recover community labels with accuracy better than random guessing when the inequality

$${(a-b)}^2\le 2(a+b)$$

(109)

holds in the limit $n\to \infty$. This detectability threshold demarcates a parametric regime in which the graph distribution becomes contiguous to an Erdős–Rényi ensemble of identical average degree, thus obfuscating community structure entirely. Above the threshold, partial recovery is not only possible but can be asymptotically achieved using spectral methods based on non-backtracking matrices. These findings place a fundamental information-theoretic limit on unsupervised learning in latent-structure models, even under full observability of the adjacency matrix.

Together, these constraints impose orthogonal but complementary restrictions on probabilistic modeling in networks. Projectibility governs the structural coherence of parameterized families under subsampling and marginalization, while spectral detectability determines whether latent labels can be statistically inferred from data. Their conjunction exposes a methodological imperative: valid inference requires not only algorithmic tractability but also conformance to intrinsic statistical geometry, a principle which any theory of learning on graphs must explicitly account for.

6.4. Path-Based and Memory-Constrained Influence Models

Departing from classical compartmental schemes and topology-specific propagation rules, a class of structurally agnostic models has emerged that defines influence via probabilistic reachability across constrained path ensembles, without presupposing fixed diffusion compartments or strict branching processes. Kuikka and Kaski [66] introduce a generalized influence matrix $I_{ij}$ constructed by summing over all time-respecting paths of bounded length L, each weighted by both the survival probability of a Poissonian propagation process and a decay kernel encoding temporal and structural attenuation. The expected influence of node i on node j is defined as

$$I_{ij}=\sum _{l=1}^L\sum _{p\in {\mathcal{P}}_{ij}^{\left(l\right)}}\omega \left(p\right)·exp(-\lambda \tau \left(p\right)),$$

(110)

where ${\mathcal{P}}_{ij}^{\left(l\right)}$ is the set of length-l paths from i to j, $\omega \left(p\right)$ is the product of edge weights along path p, $\tau \left(p\right)$ its temporal cost, and $\lambda$ is the effective rate parameter of an exponential termination process. This construction preserves path-order correlations while enabling semiring-like algebraic manipulations over time-limited propagation domains, yielding robust influence estimation in systems with partial observations or uncertain transmission kernels.

In a complementary approach focused on the content-driven dynamics of social transmission, Gleeson, Ward, O’Sullivan, and Lee [67] propose a minimal model for meme spread that integrates three essential features: individual memory constraints, competition for attention among co-occurring memes, and stochastic innovation. Each agent maintains a finite memory buffer of recent exposures and adopts content via a probabilistic rule conditioned on frequency, novelty, and position in the attention queue. The dynamics are governed by an update kernel in which the probability $P(m,t)$ that a meme m reaches popularity k at time t follows a heavy-tailed distribution:

$$P\left(k\right)\sim k^{-\alpha },\alpha \approx 2,$$

(111)

emerging from the self-reinforcing interaction of limited memory span and broadcast competition. Unlike classical epidemic analogues, this framework reproduces empirical distributions of meme popularity observed on platforms such as Twitter and Reddit, characterized by broad variability and non-Markovian inter-event dependencies. Importantly, the model does not rely on detailed network topology but instead assumes random matching and focuses on content-level dynamics, providing a microfoundational mechanism for emergent scale invariance in collective attention.

Together, Equations (110) and (111) instantiate a general class of influence models that eschew rigid compartments in favor of temporally aware, structurally agnostic propagation principles. They illustrate that transmission potential and popularity evolution can be captured through algebraic path summation and cognitive constraints rather than explicit epidemiological compartments or deterministic flow rules, thereby broadening the analytical scope of influence modeling beyond traditional network epidemiology.

The surveyed frameworks collectively demonstrate that the success of inference on networks depends not only on algorithmic sophistication or expressivity but critically on the interplay between model structure, observability, and statistical identifiability. Future research may profitably focus on several open directions. First, expanding projectibility compatible sufficient statistics beyond the current limited families would allow for broader use of exponential random graph models in practice. Second, unifying spectral detectability thresholds with algorithmic thresholds under computational constraints remains an open problem of theoretical and practical importance. Third, the emerging class of path-summed and memory-sensitive influence models suggests new design principles for systems where topological detail is inaccessible, unreliable, or transient. These include semi-observable influence platforms, adversarially perturbed networks, and human-computer hybrid information systems. A final frontier lies in the integration of these orthogonal paradigms, learning-based, message-passing, projectibility-constrained, and structural agnostic, into a universal theory of inference over relational data structures, which operates robustly across domains, scales, and sampling regimes.

7. Critical Phenomena and Percolation

This section develops a unified theoretical framework for critical phenomena in dynamical systems on networks, integrating discrete symbolic transitions, continuous topological adaptation, and structurally mediated propagation thresholds. The central aim is to elucidate how phase transitions, continuous or discontinuous, local or global, emerge from the interplay between microscopic update rules, mesoscopic connectivity patterns, and macroscopic dynamical observables. The organizational structure follows a logically ascending trajectory of model generalization and abstraction. It begins with analytically tractable symbolic-state models that establish explicit critical surfaces in minimal systems. The framework then transitions to adaptive networks in which nodal dynamics and topology co-evolve, giving rise to self-organized criticality and metastable structures. Building upon this, we introduce homophily-weighted and saturation-based opinion dynamics that capture the nonlinear coupling between belief magnitude and structural position, embedding critical bifurcations into cognitive interaction landscapes. This provides a bridge to the analysis of abrupt phase transitions, where spectral sensitivity and dynamical correlations induce first-order synchronization and percolation collapses. The section then proceeds to multilayer, interdependent systems whose hybrid coupling generates multistage bifurcations and structurally mediated double transitions. Finally, the discussion culminates in structurally agnostic models that abandon fixed compartments in favor of probabilistic influence paths and memory-aware dynamics, thereby generalizing criticality beyond rigid graph-driven or compartmental frameworks. Each subsection contributes to a progressively more inclusive theory of network-induced bifurcations, unifying local rules, global observables, and topological plasticity into a coherent taxonomic system.

The taxonomy represented in Figure 7 organizes the reviewed literature into three conceptual strata, delineating a coherent descent from microscopically structured criticality to structurally agnostic propagation dynamics. The topmost layer traces the development of critical behavior in systems with well-defined, rule-based local interactions. Beginning with symbolic dynamics [68], where phase transitions are defined over finite-state alphabets, it proceeds to adaptive topologies [69] in which connectivity evolves in tandem with nodal states, and further to nonlinear opinion dynamics [70,71] where belief strength, network centrality, and saturation jointly determine system evolution. The second layer captures structurally embedded phase discontinuities: explosive synchronization [72] illustrates how topology-dynamics correlations induce hysteretic bifurcations; interdependent percolation [73] introduces first-order collapses in multilayer networks with reciprocal dependency; and awareness-informed epidemics [74] reveal metacritical behavior arising from cognitive feedback. The final layer generalizes these mechanisms to structurally agnostic and combinatorially rich models of influence and memory. Here, path-integral formulations [66] and memory-constrained branching processes [67] replace fixed compartments with survival-weighted connectivity statistics, enabling analytic characterization of criticality without reliance on explicit compartments or homogeneous populations. The diagram’s arrowed trajectory reflects not only chronological development but conceptual generalization, illustrating how each successive layer expands the dimensionality, generality, and empirical realism of network-critical models.

7.1. Symbolic and Discrete Criticality

Symbolic dynamical systems provide analytically tractable models of phase transitions in networked systems by encoding the system state into finite alphabets and specifying transitions through local, often probabilistic, rules. The ternary-state framework introduced by Yao et al. [68] generalizes classical Boolean networks by allowing each node to occupy one of three discrete states, typically denoted $\{-1,0,+1\}$, and updating via probabilistic interactions that reflect activation, deactivation, and polarization processes. Let $p,q,r\in [0,1]$ represent the transition probabilities from active to inactive, inactive to active, and neutral to polarized states, respectively. The system exhibits a critical surface in this parameter space separating ordered from disordered regimes, given by

$$p^2+q^2+r^2+pq+qr+rp=1-{\displaystyle \frac{1}{2K}},$$

(112)

where $K\in \mathbb{N}$ denotes the fixed in-degree of each node. This manifold characterizes the onset of macroscopic disorder under finite input degree and defines the transition from stable consensus to chaotic fluctuations. The sensitivity of symbolic trajectories to initial conditions is quantified by the Lyapunov exponent

$$\lambda =ln\left[2K(1-p^2-q^2-r^2-pq-qr-rp)\right],$$

(113)

where $\lambda >0$ implies exponential divergence of trajectories and the presence of multiple coexisting attractors, whereas $\lambda <0$ guarantees local contraction to stable configurations. The introduction of an intermediate neutral state expands the expressive capacity of symbolic dynamics by permitting nonbinary attractor structures such as partial consensus and asynchronous clusters, enabling model-theoretic exploration of categorical complexity, attractor combinatorics, and logical classification within discrete dynamical systems.

To bridge symbolic dynamics with structural adaptation, the discrete adaptive model of Bornholdt and Rohlf [75] introduces self-organizing topologies wherein node activity and connectivity co-evolve under local rules. Each node assumes a binary state $s_i\left(t\right)\in \{-1,+1\}$ and updates according to a sign threshold rule

$$s_i(t+1)=sgn\left(\sum _{j=1}^N{c}_{ij}{s}_j\left(t\right)+h\right),$$

(114)

where $c_{ij}\in \{-1,0,+1\}$ encodes signed directed edges and $h=0$ is the decision threshold. If a node remains inactive across a sampled time window, it rewires one of its inputs randomly, leading to an adaptive graph ensemble in which the network structure is driven by local dynamics. The fraction of nodes with frozen state trajectories on attractors is given by the sigmoid scaling function

$$C(K,N)={\displaystyle \frac{1}{2}}\left[1+tanh\left(2a\left(N\right)(K-K_0\left(N\right))\right)\right],$$

(115)

with empirical scalings $a\left(N\right)\sim N^{\gamma }$, $K_0\left(N\right)=K_c+\mathcal{O}\left(N^{-\beta }\right)$, and the thermodynamic limit $K_c=2$. The evolved average connectivity asymptotically approaches criticality via

$$K_{\mathrm{ev}}\left(N\right)-2=c{N}^{-d},$$

(116)

where $c\approx 12.4$ and $d\approx 0.47$, confirming the convergence to critical topology in the infinite system size limit. These equations define a class of rewiring-driven critical systems in which structural plasticity and local update rules jointly produce scale-invariant global behavior. Having shown how discrete symbolic models and adaptive topologies independently and jointly realize criticality through localized transition mechanisms, we next examine how these insights extend to continuous adaptive systems with nonlinear coupling and spectral instabilities.

7.2. Adaptive Network Dynamics

The transition from discrete symbolic models to continuous adaptive dynamics marks a methodological elevation in the study of self-organized criticality on networks, where both nodal states and topological couplings evolve under mutual influence. Ito and Kaneko [4] introduced a class of continuous-time coupled chaotic oscillators with plastic interaction weights $w_{ij}\left(t\right)$, governed by the dynamical rule

$${\displaystyle \frac{d{w}_{ij}}{dt}}=ϵ\left[f\left(x_j\right)x_i-w_{ij}\right],$$

(117)

where $x_i\left(t\right)$ is the scalar state of node i, $f(·)$ is a saturating nonlinear function encoding transmission efficacy, and $ϵ$ is a timescale separation parameter. Through spontaneous symmetry breaking in the adaptation rule, the system self-organizes into hierarchical leader-follower configurations, even under high-dimensional chaotic regimes, revealing how topological plasticity can stabilize coherent mesoscale patterns without external tuning. Extending this principle to strategic behavior, Eguiluz et al. [76] and Pacheco et al. [77] embedded payoff-sensitive link rewiring into evolutionary game dynamics, whereby agents preferentially form connections with cooperators. The resulting topology-strategy feedback loop engenders metastable cooperation clusters and promotes spontaneous differentiation of strategic roles across the population. Holme and Newman [78] adapted this principle to opinion dynamics by defining a binary state variable $s_i\left(t\right)\in \{-1,+1\}$ for each agent and allowing edges to rewire according to local state alignment, thereby inducing nonequilibrium phase transitions and modular polarization of the network. A complementary formulation appears in adaptive epidemiological models, notably the extension of the SIS model by Gross et al. [69], in which susceptible agents can sever links to infected neighbors and rewire to other susceptible nodes. The modified system admits bistability and rewiring-induced hysteresis, described by the differential equations

$$\begin{array}{cc}\hfill {\displaystyle \frac{dI}{dt}}& =-\gamma I+\beta SI,\hfill \\ \hfill {\displaystyle \frac{dS}{dt}}& =\gamma I-\beta SI+w(IS\to SS),\hfill \end{array}$$

(118)

where $I\left(t\right)$ and $S\left(t\right)$ denote the fractions of infected and susceptible agents respectively, $\beta$ is the infection rate, $\gamma$ is the recovery rate, and w controls the rewiring frequency of susceptible-infected links into susceptible-susceptible pairs. This mechanism induces a coevolutionary feedback loop that drives the network toward distinct equilibria depending on initial conditions and rewiring intensity. A theoretical synthesis by Gross and Blasius [79] classifies these phenomena into universality classes of adaptive networks, emphasizing the roles of timescale separation, feedback topologies, and structural motifs in shaping global phase behavior. Collectively, these models establish that adaptive rules, whether via synaptic plasticity, strategic imitation, or awareness-mediated reconnection, can robustly steer systems toward critical or metastable configurations without parameter fine-tuning, thereby offering a principled basis for understanding resilience, critical transitions, and functional self-organization in complex networked systems.

7.3. Saturating and Homophily-Driven Opinion Dynamics

Nonlinear generalizations of classical consensus models have expanded the theoretical landscape of opinion dynamics by incorporating saturation effects, homophily-induced adaptive topology, and structurally mediated influence heterogeneity. In a continuous-time formulation developed by Baumann et al. [70], individual beliefs $x_i\left(t\right)\in [-1,1]$ evolve under saturating pairwise couplings and probabilistically rewired network topology, wherein the probability of interaction between agents i and j is given by a Boltzmann-type function

$${\mathrm{Pr}}_{ij}\left(t\right)={\displaystyle \frac{e^{-\beta |x_i\left(t\right)-x_j\left(t\right)|}}{{\sum }_k{e}^{-\beta |x_i\left(t\right)-x_k\left(t\right)|}}},$$

(119)

where the homophily parameter $\beta >0$ controls the strength of preference for similar opinions. The temporal evolution of opinions is governed by saturating influence with activity-driven coupling on a time-varying adjacency matrix $A_{ij}\left(t\right)$, such that

$${\dot{x}}_i\left(t\right)=-x_i\left(t\right)+K\sum _{j=1}^N{A}_{ij}\left(t\right)tanh\left(\alpha x_j\left(t\right)\right),$$

(120)

where $K>0$ represents the global interaction strength, and $\alpha >0$ modulates the nonlinearity of the response function, encoding issue controversialness. The hyperbolic tangent ensures saturation of influence when opinions are extreme, i.e., $|x_j|\approx 1$, and the activity-driven network structure captures stochastic pairwise interaction modulated by homophily.

Extending this approach, Dong et al. [71] propose an analytically tractable model in which the static underlying network S encodes stable social relations, and node influence is weighted by closeness centrality $c_j$, defined by

$$c_j={\displaystyle \frac{N-1}{{\sum }_{i\ne j}d(i,j)}},$$

(121)

where $d(i,j)$ is the shortest-path distance between nodes i and j in the static graph. The dynamic opinion update equation becomes

$${\dot{x}}_i\left(t\right)=-x_i\left(t\right)+\sum _{j=1}^N{c}_j{A}_{ij}\left(t\right)tanh\left(\alpha x_j\left(t\right)\right),$$

(122)

where the topology-dependent weighting $c_j$ modulates the persuasive power of node j, and the saturating nonlinearity again reflects bounded susceptibility.

To address the limited sensitivity of tanh at high $\alpha$, the authors introduce an alternative saturating function

$$f\left(x\right)=\left\{\begin{array}{cc}log(\alpha x+1),\hfill & x\ge 0,\hfill \\ -log(-\alpha x+1),\hfill & x<0,\hfill \end{array}\right.$$

(123)

which provides greater sensitivity to shifts in moderately extreme opinions. The replacement of $tanh\left(\alpha x_j\right)$ with $f\left(x_j\right)$ in the dynamical equation yields a structurally and cognitively informed model with stronger resolution in the fragmentation regime.

To classify the emergent macroscopic behavior, the diagnostic metric

$$\Delta =|\langle x_f\rangle |-\langle |x_f|\rangle$$

(124)

is introduced, where $x_f\in {[-1,1]}^N$ is the final state vector. Here, $\Delta =0$ corresponds either to global consensus (if $x_f\approx 0$) or to radicalization (if $x_f\approx \pm 1$), whereas $\Delta <0$ indicates a polarized configuration in which the opposing opinions cancel in average but retain large absolute values.

The critical value of the saturation parameter $\alpha$, separating the consensus and radicalization regimes under a mean-field assumption (with $\beta =0$), is analytically approximated as

$${\alpha }_c\approx {\displaystyle \frac{2}{(N-1)\overline{c}p(1+r)\langle a\rangle }},$$

(125)

where $\overline{c}$ is the average closeness centrality, p the probability of interaction, r the reciprocity parameter, and $\langle a\rangle$ the mean activity level. This expression delineates the onset of macroscopic transitions in the opinion landscape as a function of social reinforcement and network structure.

Taken together, these contributions reframe opinion formation as a process shaped not solely by local interactions but by topological embedding, belief strength, and nonlinear saturation effects. The joint consideration of homophilic adaptation and centrality-weighted persuasion establishes a fertile framework for analytically tractable yet empirically grounded models of cognitive dynamics on adaptive networks, bridging the symbolic, probabilistic, and topological layers of the broader theory.

7.4. Explosive Synchronization and Spectral Criteria

The phenomenon of explosive synchronization illustrates how microscopic correlations between nodal dynamics and structural properties can precipitate abrupt macroscopic transitions in coupled oscillator networks. Gómez-Gardeñes et al. [72] demonstrated that in scale-free networks, assigning natural frequencies ${\omega }_i$ proportional to node degrees $k_i$, i.e., ${\omega }_i=k_i$, radically alters the collective behavior of the Kuramoto model. Under this structural-dynamical correlation, the canonical continuous synchronization transition gives way to a discontinuous, first-order phase transition characterized by hysteresis in the order parameter $r\in [0,1]$. This hysteresis emerges from the alignment of topological hubs with fast intrinsic dynamics, and it is analytically confirmed via reduction to a star graph configuration, where a central node with degree K and frequency ${\omega }_h=K$ interacts with K peripheral nodes of unit degree and frequency $\omega =1$. The governing equations in a co-rotating frame lead to closed-form expressions for the forward critical coupling strength ${\lambda }_c$ and the corresponding coherence level $r_c$ as follows:

$${\lambda }_c={\displaystyle \frac{K-1}{K+1}},r_c={\displaystyle \frac{K}{K+1}},$$

(126)

establishing that the abruptness of the transition intensifies with increasing degree heterogeneity, and the size of the hysteresis loop scales with the hub connectivity. This paradigm demonstrates that the order of the synchronization transition is not a universal feature of the Kuramoto model but a contingent outcome of network topology-dynamics couplings. From a spectral perspective, Wang et al. [80] established a fundamental link between the epidemic threshold ${\tau }_c$ in SIS-like dynamics and the leading eigenvalue ${\lambda }_1$ of the adjacency matrix A, deriving the condition

$${\tau }_c={\displaystyle \frac{1}{{\lambda }_1\left(A\right)}},$$

(127)

which generalizes classical mean-field approximations by encoding structural inhomogeneities algebraically. This spectral criterion not only predicts outbreak initiation in arbitrary topologies but also enables analytical bounds on subcritical decay rates through eigenmode decomposition. Consequently, both explosive synchronization and spectral epidemic thresholds exemplify how structurally encoded dynamical sensitivity manifests as abrupt systemic bifurcations, transcending smooth mean-field transitions and suggesting a deep algebraic mechanism underlying discontinuities in complex network dynamics.

7.5. Interdependent and Multilayer Percolation

The theory of percolation in multilayer networks reveals a profound structural vulnerability driven by interdependencies that alter the critical properties of phase transitions. In their seminal work, Gao et al. [73] introduced a self-consistent framework for modeling cascading failures in fully interdependent networks, where each node in one Erdos–Rényi (ER) layer is dependent on a unique counterpart in the second. This configuration leads to a discontinuous percolation transition, contrasting with the continuous behavior typical of isolated networks. The system’s stability is characterized by the fraction $P_{\infty }$ of nodes belonging to the mutually connected giant component, which satisfies the nonlinear self-consistency equation

$$P_{\infty }=p{[1-e^{-k{P}_{\infty }}]}^2,$$

(128)

where k is the average degree of each layer and p is the initial fraction of surviving nodes. The square on the right-hand side reflects the necessity of mutual reachability across both layers. This construction gives rise to a fold bifurcation in the solution space, resulting in abrupt collapse at a critical threshold, thereby formalizing a first-order percolation transition induced by interdependence. Building on this foundation, Zhou et al. [81] extended the analysis to scale-free networks and introduced a partial coupling parameter $q\in [0,1]$, representing the probability that a node in one layer depends functionally on a node in the other. Their results reveal a rich phenomenology of hybrid phase transitions governed by the value of q, with multiple critical points $q_1$, $q_2$ separating discontinuous, continuous, and mixed regimes. The mathematical structure of this transition emerges from the interplay between topological heterogeneity and partial dependency, which induces discontinuous jumps in the size of the giant component even when the underlying network is scale-free and exhibits continuous transitions in isolation.

The complexity deepens in the presence of local clustering and core-periphery organization, as analyzed by Colomer-de-Simón and Boguñá [82], who derived a mesoscopic percolation theory for clustered scale-free networks. Their formalism unveils the existence of a double percolation transition, arising from the sequential activation of a highly connected core and a loosely connected periphery. Specifically, the first percolation threshold corresponds to the emergence of a mutually connected core, while the second delineates the global percolation of the entire system. The analytical derivation relies on the separation of structural roles within the network and leads to two distinct order parameters, each governed by its own critical condition. This bifurcation structure is not an artifact of dynamics or external parameters but emerges intrinsically from the graph’s architecture. Taken together, these results emphasize that multilayer topologies with heterogeneous dependencies and clustering are fundamentally distinct from single-layer networks, with the potential for abrupt, nonlocal failure cascades and structurally mediated criticality that cannot be captured by standard percolation theory.

7.6. Information-Aware Epidemics and Rumors

The dynamics of infectious propagation in social systems with endogenous cognitive feedback has been formalized through dual-process extensions of classical epidemiological models, wherein epidemic spreading is coupled with information awareness mechanisms that alter agent behavior. Funk et al. [74] introduced a stratified SIR-type model incorporating an auxiliary awareness process diffusing over the same network, in which susceptible and infected agents are partitioned by their informational state. Transmission probabilities are modulated by awareness levels, with memory decay rates and feedback loops governing the temporal evolution of risk perception. This framework induces novel bifurcation structures in the epidemic threshold and attenuates outbreak intensity through time-dependent modulation of the infection kernel.

Extending this paradigm to multiplex architectures, Granell et al. [57] developed a model wherein a physical layer supports an SIS process, while a virtual layer sustains a cyclic unaware-aware-unaware (UAU) awareness process. The interdependence of the layers introduces a metacritical transition, demarcating a region in parameter space where awareness diffusion significantly alters epidemic thresholds. The microscopic Markov chain approach (MMCA) used in their analysis defines node-level transition probabilities across the joint state space $\{\mathrm{US},\mathrm{AS},\mathrm{AI}\}$, where awareness reduces susceptibility via a factor $\gamma \in [0,1]$. The effective infection dynamics are encoded through modified adjacency matrices $a_{ij}$ and $b_{ij}$, corresponding respectively to the information and epidemic layers, with infection and recovery rates ${\beta }_U$, ${\beta }_A$, and $\mu$, and awareness decay rate $\delta$. In the linearized regime near criticality, the condition for epidemic onset reduces to a spectral criterion given by

$${\beta }_U^c={\displaystyle \frac{\mu }{{\Lambda }_{\max }\left(H\right)}},H_{ji}=\left[1-(1-\delta )p_i^A\right]b_{ji},$$

(129)

where $p_i^A$ denotes the stationary probability of node i being aware and ${\Lambda }_{\max }\left(H\right)$ is the spectral radius of the modulated transmission matrix. When awareness transmission is subcritical, the epidemic threshold remains decoupled from the virtual layer structure, collapsing to the classical expression ${\beta }_c=\mu /{\Lambda }_{\max }\left(B\right)$. However, for awareness rates $\lambda >{\lambda }_c$, the system undergoes a metacritical shift, redefining the phase boundary and suppressing epidemic prevalence. This emergent regime reflects a nontrivial coupling between non-Markovian feedback and network topology, exhibiting rich phase behavior inaccessible to single-layer models and underscoring the structural role of multiplexity in mediating antagonistic contagion processes.

7.7. Path-Based and Memory-Constrained Influence

To transcend model-specific compartmental dynamics, recent advances have formulated structural-agnostic probabilistic frameworks for quantifying influence propagation in temporally and topologically heterogeneous networks. Kuikka and Kaski [66] develop a matrix-integral approach that defines the cumulative probability of influence transmission between node pairs via all admissible paths up to a maximum length under finite-time constraints. Their formalism yields an influence-spreading matrix $C(s,e)$ over all source–target node pairs $(s,e)$, supporting the derivation of survival-weighted centrality, path-aware betweenness, and dynamic community metrics grounded in inclusion-exclusion combinatorics. Given a Poissonian propagation rate $\lambda$, the probability that a path L of length $\left|L\right|$ is completed before a deadline T is expressed as

$$P_L\left(T\right)=1-\sum _{z=0}^{\left|L\right|-1}{\displaystyle \frac{{\left(\lambda T\right)}^z}{z!}}{e}^{-\lambda T},$$

(130)

and the total influence probability along a given path L is given by

$$P\left(L\right)=\prod _{(u,v)\in E_L}w(u,v)\prod _{v\in L}{w}_v,$$

(131)

where $w(u,v)$ and $w_v$ are edge- and node-specific weights, encoding local transmission efficacy. This probabilistic framework enables influence estimation without predefined compartments, accommodates complex contagion and memory effects, and directly integrates weighted, dynamic, and recurrent edge interactions. Complementarily, Gleeson et al. [67] propose a stochastic model for meme diffusion in directed networks where each node samples from a bounded memory stream, influenced by innovation rate and prior exposure. The model yields heavy-tailed meme popularity distributions in the critical regime $R_0\to 1$, where the asymptotic popularity distribution $q_n(\infty )$ follows

$$q_n(\infty )\sim A{n}^{-3/2}{e}^{-n/\kappa },$$

(132)

with A and $\kappa$ parameterized by memory length and posting frequency. The model accounts for observed phenomena such as the absence of stable early-mover advantage and scale-free fluctuation patterns in meme success. Extending these ideas to multilayered, media-coupled systems, Zhong et al. [58] construct a dual-layer VR–SHI₁I₂R model that explicitly incorporates the interaction between broadcast media states $\{M_1,M_2,M_3\}$, representing media silence, rumor propagation, and refutation broadcasting, respectively, and individual-level nodes in susceptible (S), hesitant (H), rumor-spreading ($I_1$), rumor-refuting ($I_2$), and restrained (R) states. The control architecture of this hybrid system is optimized via Pontryagin’s maximum principle applied to a cost functional

$$J\left(u\right)={\displaystyle \frac{1}{2}}{\int }_{t_0}^{t_f}\left[A_1{M}_2\left(t\right)+A_2{I}_1\left(t\right)+{\displaystyle \frac{B_1}{2}}{u}_1^2\left(t\right)+{\displaystyle \frac{B_2}{2}}{u}_2^2\left(t\right)+{\displaystyle \frac{B_3}{2}}{u}_3^2\left(t\right)\right]dt,$$

(133)

where $u_1\left(t\right)$, $u_2\left(t\right)$, and $u_3\left(t\right)$ are the time-dependent control functions representing efforts to reduce rumor-spreading, enhance refutation, and reduce hesitation, respectively. The weights $A_1$ and $A_2$ penalize the existence of active media rumor and individual-level infection states, respectively, while $B_1,B_2,B_3$ are cost coefficients associated with the magnitude of control interventions. The optimization is conducted over the admissible controls that drive the state trajectories subject to the coupled system of nonlinear differential equations. Empirical calibration to the widely reported case of Hu Xinyu, a high-profile incident of youth disappearance and mass rumor propagation in Chinese media, demonstrates the model’s accuracy in simulating intervention scenarios and the propagation dynamics of competing narratives [58]. Taken together, these structurally agnostic, probabilistically grounded, and cognitively enriched models articulate a unifying theory of information spread that preserves critical scaling, accommodates memory and feedback, and remains analytically tractable in the absence of rigid compartmental boundaries.

Taken together, the models synthesized in this section delineate a comprehensive taxonomy of criticality in networked dynamical systems, unifying symbolic logic, topological feedback, spectral criteria, and structurally agnostic propagation under a single analytical umbrella. They reveal that the emergence of phase transitions is not confined to narrow parameter ranges or specific mechanisms, but constitutes a structural property of many-body systems embedded in graphs. Future research directions include the integration of higher-order interactions, non-Poissonian temporal kernels, and learning-driven feedback into these frameworks, thereby capturing even more nuanced features of adaptive, cognitive, and emergent network behavior. Moreover, empirical calibration against real-world multilayer systems—social, infrastructural, and biological—will be essential to validate and refine the theoretical insights presented here, ultimately advancing a predictive science of criticality in complex systems.

8. Control, Optimization and Intervention Design

The design of control, optimization, and intervention strategies in complex networks spans a diverse range of methodological paradigms, structural frameworks, and application domains. In order to systematize the extensive literature reviewed in this section, we propose a two-dimensional taxonomic scheme that organizes representative contributions according to two orthogonal dimensions: the control method employed and the structural modeling framework assumed. This classification, visualized in Figure 8, provides a structured lens through which the analytical diversity of control-theoretic approaches can be assessed.

The vertical axis enumerates the principal categories of control mechanisms—consensus and synchronization, controllability and observability, intervention design, network optimization, and learning-based control—while the horizontal axis distinguishes between modeling assumptions ranging from linear spectral methods to multilayer cognitive architectures. Each cell of the matrix highlights paradigmatic models or algorithms that exemplify the intersection of a specific control technique and structural framework, thereby mapping the conceptual topology of the field. This grid not only facilitates navigability across methodological and structural domains but also underscores thematic continuities and divergences that shape contemporary research on networked control.

8.1. Spectral and Delay-Robust Control in Consensus Protocols

The integration of spectral criteria and delay robustness into the design of distributed consensus protocols represents a crucial advance in the stabilization of multi-agent systems (MASs) operating under structural and temporal constraints. Gong [83] investigates linear consensus dynamics subject to uniform communication delays, wherein the network interactions are encoded through the Laplacian matrix L of the underlying graph. The delayed consensus system is modeled in the frequency domain using the resolvent operator $R\left(s\right)=sI-A_d-{\sum }_d{e}^{-s{\tau }_d}B{K}_{d}C$, where $A_d$ is the nominal system matrix, B and C define the input and output matrices respectively, $K_d$ is the feedback gain associated with delay ${\tau }_d$, and $s\in \mathbb{C}$ is the complex frequency variable. The central spectral object governing the stability of the system is the interval $[{\lambda }_2,{\lambda }_N]$, determined by the nontrivial eigenvalues of the Laplacian matrix L, where ${\lambda }_2$ is the algebraic connectivity and ${\lambda }_N$ the largest eigenvalue. A sufficient condition ensuring delay-tolerant stability is formulated as

$$\underset{s\in \ell }{sup}∥R^{-1}\left(s\right)∥\sum _d∥B{K}_{d}C∥<{\displaystyle \frac{1}{\Delta \alpha }},$$

(134)

where ℓ denotes a bounded subset of the imaginary axis capturing critical frequency modes, and $\Delta \alpha$ is the prescribed gain margin. The gain synthesis problem is then cast as a constrained convex optimization problem seeking to minimize the $L_2$-norm of the transition function subject to spectral feasibility constraints imposed by the eigenstructure of L and the resolvent decay rate. This formulation establishes a unifying framework for delay-robust consensus that leverages the spectral graph structure and operator norms to explicitly encode robustness, control energy, and convergence guarantees.

Manickavalli et al. [84] extend the classical consensus framework to accommodate structurally antagonistic interaction topologies characterized by bipartite symmetry and directed adversarial disturbances. The model considers multi-agent dynamics under exogenous disturbance input $d\left(t\right)$ and structural asymmetries in the influence graph. To achieve convergence to bipartite consensus states—defined by node state partitions with sign-symmetric asymptotics—the authors embed an extended internal disturbance (EID) estimator into the feedback loop. The augmented system is stabilized through a disturbance observer-based controller synthesized via linear matrix inequalities derived from ${\mathcal{H}}_{\infty }$ control theory. The sufficient condition for robust stabilization is expressed via the feasibility of the LMI:

$$\left[\begin{array}{cc}XA+A^{T}X+YC+C^T{Y}^T& XB\\ B^{T}X& -\gamma I\end{array}\right]\prec 0,$$

(135)

where $A\in {\mathbb{R}}^{n\times n}$ is the state matrix, $B\in {\mathbb{R}}^{n\times m}$ and $C\in {\mathbb{R}}^{p\times n}$ are input and output matrices, $X=X^T\succ 0$ and $Y\in {\mathbb{R}}^{n\times p}$ are decision variables, and $\gamma >0$ defines the worst-case energy gain bound from disturbance to output. This design guarantees convergence to a structurally balanced equilibrium even under directed topologies with asymmetrical influence and persistent disturbances, thereby extending consensus theory to adversarial domains.

Wu et al. [85] propose a novel event-triggered pinning control strategy for discrete-time Takagi–Sugeno (T–S) fuzzy networks, in which each agent’s dynamics are described by a convex combination of local linear models weighted by fuzzy membership functions. Let the local T–S model be given by the following:

$$x_i(k+1)=\sum _{r=1}^R{h}_r\left(z_i\left(k\right)\right)\left(A_r{x}_i\left(k\right)+B_r{u}_i\left(k\right)\right),$$

(136)

where R is the number of fuzzy rules, $h_r\left(z_i\left(k\right)\right)\in [0,1]$ are normalized membership functions satisfying ${\sum }_r{h}_r\left(z_i\left(k\right)\right)=1$, and $z_i\left(k\right)$ is the premise variable at time step k. The control protocol is pinning-based and activated only when the event-triggering condition

$$\parallel e_i{\left(k\right)\parallel }^2\ge {\delta }_i{\parallel x_i\left(k\right)\parallel }^2,$$

(137)

is satisfied, where $e_i\left(k\right)=x_i\left(k\right)-x_i\left(k_e\right)$ denotes the measurement error since the last control update, and ${\delta }_i\in (0,1)$ is a design threshold. The closed-loop system stability is analyzed using Lyapunov–Krasovskii functionals constructed as

$$V\left(k\right)=x{\left(k\right)}^{T}Px\left(k\right)+\sum _{\tau =k-d}^{k}x{\left(\tau \right)}^{T}Qx\left(\tau \right),$$

(138)

where $P,Q\succ 0$ are symmetric positive definite matrices and d is the maximum delay bound. Linear matrix inequality (LMI) conditions are then derived to ensure finite-time synchronization under bounded actuation and fuzzy rule uncertainty. This synthesis achieves resource-efficient control by minimizing redundant updates and enables synchronization in nonlinear and uncertain environments with asynchronous feedback scheduling.

Wang et al. [86] investigate the application of pinning control in structurally symmetric networks that support cluster synchronization. Their approach relies on the variational decomposition of the network state space using group-theoretic symmetries, particularly the representation theory of permutation groups. The system dynamics are decoupled into orthogonal subspaces: the synchronous manifold and the transverse modes. Let L be the Laplacian matrix of the network, and let the system be transformed into modal coordinates such that the variational equation decomposes into blocks $L_B$ and $L_D$, governing transverse and synchronous modes, respectively. The pinning control law is given by the following:

$$u\left(t\right)=-\eta (x\left(t\right)-x_r\left(t\right)),$$

(139)

where $x_r\left(t\right)$ denotes the reference trajectory and $\eta >0$ is the pinning gain. The critical pinning strength ${\eta }_c$ required to stabilize the transverse modes is derived analytically as

$${\eta }_c={\displaystyle \frac{1}{2}}\underset{\lambda \in \sigma \left(L_D\right)}{min}\Re \left(\lambda \right),$$

(140)

where $\sigma \left(L_D\right)$ denotes the spectrum of the transverse block of the variational Laplacian. This result establishes an explicit spectral condition for the onset of stable cluster synchronization and demonstrates how minimal, symmetry-aligned control inputs can stabilize complex coordination patterns through selective actuation informed by eigenstructure and representation symmetry.

8.2. Motif-Driven Centrality and Intervention Targeting

A rigorous formalization of intervention design grounded in local structural motifs has been developed by Hu and Zhang [87], who introduce a dynamic centrality metric based on the Jacobian sensitivity of networked dynamical systems. Let E be the graph encoding system structure, and let $m\subset E$ denote a structural motif, i.e., a localized subgraph pattern such as a triangle, feedforward loop, or feedback pair. The proposed centrality function $S(m,E)$ evaluates the global dynamical significance of the motif m by quantifying its perturbative influence on the full system’s stability. Specifically, the system is assumed to evolve according to a set of nonlinear differential equations linearized around an equilibrium point $x^{\ast }$, yielding a Jacobian matrix $J\in {\mathbb{R}}^{N\times N}$. The effect of perturbing a motif m is modeled by injecting a structured matrix perturbation $\Delta J_m$, whose support is restricted to the edges of m. The centrality score is defined as the spectral energy of the perturbation projected onto the dominant eigenspaces of the system:

$$S(m,E)=\sum _{i=1}^N{\left({\displaystyle \frac{\partial {\lambda }_i}{\partial \epsilon }}{|}_{\epsilon =0}\right)}^2,$$

(141)

where ${\lambda }_i$ denotes the i-th eigenvalue of the perturbed Jacobian $J+\epsilon \Delta J_m$, and the derivative is evaluated at $\epsilon =0$, reflecting the linear sensitivity of the spectral structure to motif-level perturbation. By computing the Fréchet derivative of the eigenvalues with respect to structured perturbations, the authors obtain the explicit expression

$${\displaystyle \frac{\partial {\lambda }_i}{\partial \epsilon }}{|}_{\epsilon =0}=v_i^T\Delta J_m{u}_i,$$

(142)

where $v_i$ and $u_i$ are the left and right eigenvectors of J associated with ${\lambda }_i$, normalized such that $v_i^T{u}_i=1$. Consequently, the full centrality score becomes

$$S(m,E)=\sum _{i=1}^N{\left(v_i^T\Delta J_m{u}_i\right)}^2,$$

(143)

which aggregates the squared directional derivatives across all eigenmodes. This quantity encapsulates both the alignment of the perturbation with dynamically sensitive directions and the energy of the motif perturbation projected into the eigenspace of the system. To interpret the resulting centrality in terms of control energy and sensitivity, the authors further relate $S(m,E)$ to the trace of the projected perturbation energy operator:

$$S(m,E)=\mathrm{Tr}\left({\left(V^T\Delta J_{m}U\right)}^T\left(V^T\Delta J_{m}U\right)\right),$$

(144)

where $U=[u_1,\dots ,u_N]$ and $V=[v_1,\dots ,v_N]$ collect the biorthogonal eigenbases of the unperturbed Jacobian. This trace formulation permits scalable computation and encapsulates how motif-level interventions propagate through the system’s linearized dynamics. The metric $S(m,E)$ satisfies desirable invariance properties under basis change and supports additive decomposition over motif classes. Empirical applications to synthetic oscillatory networks, ecological food webs, and regulatory circuits reveal that certain motifs consistently exert disproportionately large dynamical influence, even when their static topological centrality is low. Thus, this centrality measure provides a basis for targeted intervention and robustness evaluation grounded not in topological heuristics but in the differential response of system eigenstructure to structured, localized perturbations.

8.3. Rumor and Information Propagation Control

A comprehensive understanding of rumor dynamics requires models that simultaneously incorporate cognitive, affective, and structural dimensions of information diffusion. Govindankutty and Gopalan [88] introduce the SEDPNR model, a compartmental framework that explicitly captures the interplay between exposure, emotional polarization, belief appraisal, and behavioral restraint. The population is partitioned into six dynamic compartments: susceptible (S), exposed (E), doubters (D), positively infected (P), negatively infected (N), and restrained (R). Transitions between these compartments model not only the flow of information but also the affective and cognitive evaluation that occurs upon exposure. The dynamics are governed by the nonlinear system of ordinary differential equations:

$$\begin{array}{cc}\hfill {\displaystyle \frac{dS}{dt}}& ={\mu }_{1}E+{\mu }_{2}D-\alpha S,\hfill \\ \hfill {\displaystyle \frac{dE}{dt}}& =\alpha S-({\beta }_1+{\beta }_2+\gamma +{\mu }_1)E,\hfill \\ \hfill {\displaystyle \frac{dD}{dt}}& =\gamma E-({\beta }_3+{\beta }_4+{\mu }_2)D,\hfill \\ \hfill {\displaystyle \frac{dP}{dt}}& ={\beta }_{1}E+{\beta }_{3}D-{\lambda }_{1}P,\hfill \\ \hfill {\displaystyle \frac{dN}{dt}}& ={\beta }_{2}E+{\beta }_{4}D-{\lambda }_{2}N,\hfill \\ \hfill {\displaystyle \frac{dR}{dt}}& ={\lambda }_{1}P+{\lambda }_{2}N,\hfill \end{array}$$

(145)

where $\alpha$ is the exposure rate, $\gamma$ is the doubting rate, ${\mu }_1$ and ${\mu }_2$ are reversion rates to susceptibility, ${\beta }_i$ control the pathways of polarization through direct exposure and evaluative doubt, and ${\lambda }_i$ represent the rates of emotional dissipation leading to restraint. The basic reproduction number for this cognitive-affective system is defined as

$$R_0=max\left({\displaystyle \frac{{\beta }_1}{{\lambda }_1}},{\displaystyle \frac{{\beta }_2}{{\lambda }_2}}\right),$$

(146)

quantifying the threshold beyond which emotionally polarized beliefs propagate through the population. This formulation generalizes classical rumor models by incorporating evaluative uncertainty (D) and the emotional valence of belief adoption ($P,N$), both of which modulate the stability and transience of informational cascades. Notably, this model establishes a theoretical bridge between emotional reinforcement and belief saturation, thereby providing analytical grounding for regulatory interventions in online and media environments where affective dynamics dominate. Complementing this framework, Zhong et al. [58] have demonstrated the efficacy of multilayer optimization strategies in rumor control by applying Pontryagin’s Maximum Principle to coupled communication systems, while Zhao et al. [89] integrate memory effects into SIHR dynamics through degree-based heterogeneous mean-field approximations. The SEDPNR model subsumes and extends these approaches by capturing affective divergence, cognitive hesitation, and reversion, thereby synthesizing sentiment-aware propagation with structured and memory-integrated dynamics under a unified analytical formalism.

8.4. Structural Controllability and Observability in Directed Graphs

The theory of structural controllability in directed graphs was rigorously formalized by Liu, Slotine, and Barabási [30], who reformulated the Kalman rank condition in a purely topological context by leveraging the combinatorial properties of maximum matchings. Let a linear time-invariant system be given by the equation

$${\displaystyle \frac{dx\left(t\right)}{dt}}=Ax\left(t\right)+Bu\left(t\right),$$

(147)

where $A\in {\mathbb{R}}^{N\times N}$ is the system matrix encoding the interaction structure of the network, $B\in {\mathbb{R}}^{N\times M}$ is the input matrix specifying control inputs, $x\left(t\right)\in {\mathbb{R}}^N$ is the state vector, and $u\left(t\right)\in {\mathbb{R}}^M$ is the control signal. A system is structurally controllable if there exists at least one numerical realization of the nonzero elements in A and B such that the controllability matrix

$$\mathcal{C}=\left(B,AB,A^{2}B,\dots ,A^{N-1}B\right)$$

(148)

satisfies the full-rank condition

$$\mathrm{rank}\left(\mathcal{C}\right)=N.$$

(149)

In the structural formulation, the minimal number of driver nodes required for full control corresponds to the number of unmatched nodes in a maximum matching of the directed graph defined by the adjacency matrix A. A matching is a set of edges such that no two share a common start or end node, and a maximum matching includes the largest number of such edges. Nodes not touched by any edge in the matching must be directly controlled, as they cannot be reached via indirect paths. The Hopcroft–Karp algorithm is used to compute this maximum matching efficiently, and the unmatched node set determines the necessary actuator placement for structural controllability. Extending this framework, Liu and Barabási [90] analyze how the degree distribution of the network influences controllability properties. They demonstrate that in scale-free networks characterized by power-law degree distributions, hubs are frequently matched and thus do not require control input, whereas low-degree nodes are more likely to be unmatched and must serve as driver nodes. This counterintuitive result implies that networks with greater connectivity heterogeneity may require more independent control signals than their homogeneous counterparts. The scaling behavior of the number of driver nodes $N_D$ as a function of the degree exponent $\gamma$ is quantified, with denser tails (i.e., lower $\gamma$) leading to increased control demand due to matching saturation on hubs and redundancy among low-degree periphery nodes.

Nacher and Akutsu [91] propose an alternative structural characterization of controllability in bipartite networks based on the concept of a minimum dominating set (MDS). Consider a bipartite graph $G=(V^{⊳},V^{⊲};E)$, where $V^{⊳}$ denotes candidate control nodes and $V^{⊲}$ target nodes. A subset $S\subseteq V^{⊳}$ forms a dominating set if for every $u\in V^{⊲}$, there exists a $v\in S$ such that $(v,u)\in E$. The goal is to find an MDS with minimal cardinality that ensures full reachability. For bipartite networks with heterogeneous degree distributions $P\left(k\right)\sim k^{-{\gamma }_1}$, the expected size of the MDS obeys the scaling laws

$$\left|S\right|=\Theta \left(n_1\right)\mathrm{for}{\gamma }_1>2,$$

(150)

and

$$\left|S\right|=\mathcal{O}\left({\displaystyle \frac{n_2^{2-{\gamma }_1}·m^{{\gamma }_1-1}}{H^{(2-{\gamma }_1)({\gamma }_1-1)}}}\right)\mathrm{for}1<{\gamma }_1<2,$$

(151)

where $n_1=\left|V^{⊳}\right|$, $n_2=\left|V^{⊲}\right|$, H is the maximum degree, and m is the total number of edges. To operationalize this condition, the MDS is formulated as an integer linear program with binary decision variables $x_v\in \{0,1\}$ indicating inclusion in the control set:

$$min\sum _{v\in V^{⊳}}{x}_v\mathrm{subject}\mathrm{to}\sum _{(v,u)\in E}{x}_v\ge 1\forall u\in V^{⊲},x_v\in \{0,1\}.$$

(152)

This structure enables efficient approximate solutions and highlights topological accessibility over algebraic reachability, offering a scalable alternative to rank-based diagnostics in actuator placement.

In a dual direction, Bianchin et al. [92] define the observability radius as a quantitative measure of the fragility of linear systems with respect to perturbations that destroy observability. Let $\dot{x}\left(t\right)=Ax\left(t\right),y\left(t\right)=Cx\left(t\right)$ define a linear time-invariant system, where $C\in {\mathbb{R}}^{P\times N}$ selects the observable state components. The observability radius is the smallest Frobenius norm perturbation ${\parallel \Delta \parallel }_F$ such that the perturbed pair $(A+\Delta ,C)$ becomes unobservable, subject to the structural constraint $\Delta \in {\mathcal{A}}_H$, where ${\mathcal{A}}_H$ denotes the admissible perturbation subspace consistent with a given sparsity pattern H. The resulting nonconvex optimization problem is

$$\underset{\Delta ,x,\lambda }{min}{\parallel \Delta \parallel }_F^2\mathrm{subject}\mathrm{to}(A+\Delta )x=\lambda x,Cx=0,\parallel x\parallel =1,\Delta \in {\mathcal{A}}_H,$$

(153)

where $\lambda \in \mathbb{C}$ and $x\in {\mathbb{C}}^N$ define the eigen-pair of the unobservable mode. To approximate the solution, the problem is transformed into a nonlinear eigenvalue system using diagonal weighting matrices $D_x=\mathrm{diag}\left(x\right)$, $D_y=\mathrm{diag}\left(y\right)$, and a structured surrogate $\tilde{A}$, yielding the iterative formulation

$$\tilde{A}x=\sigma D_{y}y,{\tilde{A}}^{T}y=\sigma D_{x}x,$$

(154)

where $y\in {\mathbb{C}}^N$ is the dual vector and $\sigma \in \mathbb{R}$ is a spectral scalar. This construction allows the observability radius to be computed through inverse iteration schemes, yielding mode-specific sensitivity metrics and informing sensor placement under structured uncertainty.

To assess the vulnerability of structural controllability under network degradation, Mengiste et al. [93] define the pruning damage index (PDI) as a measure of the sensitivity of the number of driver nodes to edge deletions. Let $N_D^{\mathrm{original}}$ and $N_D^{\mathrm{pruned}}$ denote the number of driver nodes before and after pruning, respectively. Then the PDI is given by

$$\mathrm{PDI}={\displaystyle \frac{N_D^{\mathrm{pruned}}-N_D^{\mathrm{original}}}{N}},$$

(155)

where N is the system size. Edge deletions that disproportionately increase $N_D$ indicate fragile regions of the network from a control perspective. By analyzing the cardinality curve of the maximum matching as a function of successive deletions, the authors show that scale-free networks exhibit acute fragility under targeted removal of hub-connected links, while small-world and homogeneous networks maintain higher structural redundancy. This metric enables a quantitative comparison of network architectures with respect to their robustness under edge degradation, thereby informing the design of resilient and structurally controllable systems in the presence of topological fluctuations.

8.5. Switched and Time-Varying Control Architectures

The analysis of structural controllability in switched and time-varying systems requires the development of generalized frameworks that extend beyond the static topological assumptions of classical control theory. Zhang et al. [94] address this challenge by introducing a multilayer dynamic graph (MDG) formalism to characterize the controllability of switched linear continuous-time systems composed of a finite collection of mode-specific subsystems. Each subsystem is described by a pair $(A^{\left(i\right)},B^{\left(i\right)})$, where $A^{\left(i\right)}\in {\mathbb{R}}^{N\times N}$ encodes the system structure and $B^{\left(i\right)}\in {\mathbb{R}}^{N\times M}$ the input configuration in mode i. The switching signal $\sigma \left(t\right)$ determines the temporal evolution across the L available subsystems, inducing a non-autonomous dynamics given by

$${\displaystyle \frac{dx\left(t\right)}{dt}}=A^{\left(\sigma \right(t\left)\right)}x\left(t\right)+B^{\left(\sigma \right(t\left)\right)}u\left(t\right).$$

(156)

To capture the control influence across modes, the authors construct a multilayer dynamic graph in which each layer encodes the structural pattern of a subsystem, and inter-layer links represent admissible switching transitions. Reachability is encoded through a recursive generation of controllable subspaces ${\Omega }_1,{\Omega }_2,\dots ,{\Omega }_{\ell }$, each spanned by a matrix $W_i$ satisfying ${\Omega }_{i+1}=\mathrm{span}\left(A^{\left(i\right)}{\Omega }_i\cup B^{\left(i\right)}\right)$. Structural controllability is achieved when the terminal space satisfies

$$\mathrm{rank}\left(W_{\ell }\right)=dim\left(\mathrm{controllable}\mathrm{subspace}\right),$$

(157)

indicating that all state variables can be reached through some admissible switching sequence. This condition is equivalent to the existence of a generalized cactus configuration, a graph-theoretic structure comprising disjoint unions of generalized stems and buds, each satisfying combinatorial reachability and independence constraints. The notion of S-disjoint input–state walks is introduced to formalize the noninterfering propagation of control across modes, and the presence of a generalized cactus guarantees a nonzero minor in the symbolic controllability matrix, thereby certifying structural controllability. This framework closes a fundamental gap in classical theory by enabling graph-theoretic certification of controllability for time-varying systems with switching topologies, and admits efficient verification through recursive rank checks and cactus-based decomposition of the MDG.

To generalize structural controllability to systems with explicitly time-varying matrices, Reissig, Hartung, and Svaricek [95] develop a rigorous theory of strong structural controllability (SSC), which requires that a system remains controllable for all numerical instantiations consistent with a fixed sparsity pattern, regardless of the time dependence of the entries. Consider a discrete-time, time-varying system of the form

$$x(t+1)=A\left(t\right)x\left(t\right)+B\left(t\right)u\left(t\right),$$

(158)

where $A\left(t\right),B\left(t\right)\in {\mathbb{R}}^{N\times N},{\mathbb{R}}^{N\times M}$ are time-varying matrices whose zero–nonzero pattern remains fixed across time. The authors prove that SSC over a finite horizon is equivalent to the SSC of a corresponding time-invariant system with the same structural pattern, allowing the application of static controllability tests such as input–state walk enumeration and matching analysis. However, for continuous-time systems of the form

$${\displaystyle \frac{dx\left(t\right)}{dt}}=A\left(t\right)x\left(t\right)+B\left(t\right)u\left(t\right),$$

(159)

this equivalence does not hold due to the analytic structure of the solution operator, which depends nonlinearly on the time-dependent flow of $A\left(t\right)$. Counterexamples are provided in which structurally controllable time-invariant systems lose SSC under continuous variation. To resolve this discrepancy, the authors introduce a class of exponentially scaled systems defined by

$$A\left(t\right)=e^{-\Lambda t}{A}_0{e}^{\Lambda t}-\Lambda ,B\left(t\right)=e^{-\Lambda t}{B}_0,$$

(160)

where $\Lambda \in {\mathbb{R}}^{N\times N}$ is a diagonal matrix encoding exponential scaling, and $A_0,B_0$ are time-invariant matrices with the same sparsity structure as $A\left(t\right),B\left(t\right)$. For this subclass, SSC is shown to reduce to the static case, enabling the use of classical structural criteria. This formulation allows algebraic tools to be applied to time-varying systems by transforming the dynamics into an exponentially equivalent form that preserves controllability properties. The resulting framework offers a unified structural theory for systems with temporal modulation, highlights the limitations of naive generalizations from static to time-varying settings, and provides necessary and sufficient conditions for SSC under both discrete and continuous temporal evolution within the constraints of fixed sparsity topology.

8.6. Neural Control, Biophysical Learning, and Structural Inference

The extension of control-theoretic principles to neural, cognitive, and biological systems necessitates the formulation of models that integrate structural complexity, dynamic heterogeneity, and quantifiable energy constraints. Tang and Bassett [96] conceptualize the brain as a finite-dimensional linear dynamical system evolving over the structural connectome [97] encoded by a directed weighted adjacency matrix $A\in {\mathbb{R}}^{N\times N}$. The system’s discrete-time evolution is given by

$$x(t+1)=Ax\left(t\right)+B_K{u}_K\left(t\right),$$

(161)

where $x\left(t\right)\in {\mathbb{R}}^N$ is the brain state vector, $B_K\in {\mathbb{R}}^{N\times \left|K\right|}$ is the input matrix selecting the set of controlled brain regions indexed by $K\subseteq \{1,\dots ,N\}$, and $u_K\left(t\right)\in {\mathbb{R}}^{\left|K\right|}$ denotes the control signal. The minimum energy required to steer the system from initial state $x\left(0\right)=0$ to target state $x_f\in {\mathbb{R}}^N$ over a finite time horizon T is expressed through the inverse of the controllability Gramian $W_{K,T}\in {\mathbb{R}}^{N\times N}$, defined as

$$W_{K,T}=\sum _{\tau =0}^{T-1}{A}^{\tau }{B}_K{B}_K^T{\left(A^T\right)}^{\tau },$$

(162)

yielding the optimal energy

$$E(u_K^{\ast },T)=x_f^T{W}_{K,T}^{-1}{x}_f.$$

(163)

To evaluate regional control capabilities, the authors introduce three metrics. Average controllability quantifies a node’s ability to steer the system toward many low-energy target states and is proportional to $\mathrm{Tr}\left(W_{K,T}\right)$. Modal controllability assesses a node’s capacity to modulate difficult-to-reach modes associated with small eigenvalues of A. Boundary controllability measures the influence of a node in transitioning between network modules defined by community partitions. These quantities provide a functional decomposition of control influence across brain regions and establish a quantitative framework for assessing cognitive flexibility and neural state transition feasibility under structural constraints.

Extending control synthesis to systems with fractional dynamics and limited observability, Bai et al. [98] investigate synchronization in multiplex networks of coupled fractional-order oscillators, wherein each layer exhibits noninteger-order dynamics and the output coupling is constrained by quantized information exchange. The model employs Caputo-type fractional derivatives of order $\alpha \in (0,1)$ and defines the node-level dynamics as

$$D^{\alpha }{x}_i\left(t\right)=f\left(x_i\left(t\right)\right)+\sum _{j\in {\mathcal{N}}_i}{c}_{ij}Q(x_j\left(t\right)-x_i\left(t\right)),$$

(164)

where $D^{\alpha }$ denotes the fractional derivative, $c_{ij}$ are interlayer coupling weights, and $Q(·)$ is a logarithmic quantization function. To ensure convergence under bounded quantization error, the authors construct Lyapunov functions based on Mittag-Leffler stability theory and derive sufficient conditions guaranteeing synchronization across the multiplexed topology. An adaptive control scheme is implemented in which the feedback gain is updated according to

$${\dot{k}}_i\left(t\right)={\beta }_i{\parallel x_j\left(t\right)-x_i\left(t\right)\parallel }^2,$$

(165)

where ${\beta }_i>0$ controls the adaptation rate. The combined use of fractional calculus, quantized feedback, and adaptive mechanisms demonstrates the feasibility of control under both nonlocal temporal memory and discrete information constraints.

In the context of multilayer control path identification, Wang et al. [99] introduce the concept of Conserved Control Paths (CCPs), which are edge sequences that maintain structural control significance across multiple network layers. For a multilayer system with L layers and edge set $\chi$, the conservative degree $CV\left(\chi \right)$ quantifies the interlayer agreement and is defined as

$$CV\left(\chi \right)={\displaystyle \frac{1}{\left|\chi \right|}}\sum _{e\in \chi }{\left({\displaystyle \frac{1}{L}}\sum _{l=1}^L{\mathbb{I}}_{e\in M_l}\right)}^2,$$

(166)

where $M_l$ denotes a maximum matching in layer l, and ${\mathbb{I}}_{e\in M_l}$ is the indicator function evaluating to one if edge e belongs to $M_l$. The CoPath algorithm is introduced to identify high-$CV$ edge sets using bipartite matching over aggregated ordinary and critical edge classes. The conservative weight assigned to each edge is given by

$$\hat{w}\left(e\right)=\sum _{l=1}^L{\mathbb{I}}_{e\in (E_l^o\cup E_l^c)},$$

(167)

where $E_l^o$ and $E_l^c$ denote the ordinary and critical edge classes in layer l. The CCP framework identifies structurally persistent control backbones and enables targeted intervention in complex systems characterized by topological heterogeneity and interaction uncertainty.

Raducha and San Miguel [100] apply dynamical control analysis to cognitive-social systems modeled as two-layer evolutionary games with overlapping agent populations. Each layer hosts a coordination game with distinct payoff matrices, and the fraction $q\in [0,1]$ of overlapping agents participates in both layers. Let ${\alpha }_I$ and ${\alpha }_{II}$ denote the fractions of agents adopting the dominant strategy in layers I and II, respectively. The time evolution of ${\alpha }_I,{\alpha }_{II}$ is governed by replicator dynamics or alternative evolutionary rules such as best response and imitation, with payoff asymmetry controlled by a parameter $\Delta S$. The system exhibits bifurcation phenomena in the $(q,\Delta S)$ parameter space, with transitions between regimes of decoupled equilibria, bistability, and global coordination. The resulting phase diagrams demonstrate how topological coupling via shared agents can induce spontaneous synchronization across layers, effectively functioning as a control mechanism embedded in population structure. These results underscore the role of overlapping populations in modulating collective behavior and provide an evolutionary game-theoretic analogue of structural control in interdependent systems.

Perc and Szolnoki [101] extend this approach to the spatial public goods game, introducing adaptive punishment as a cooperator strategy that evolves locally in response to defector invasion. By modulating the punishment intensity ${\pi }_x\in {\mathbb{R}}_{\ge 0}$, cooperators dynamically respond to local conflict, thereby suppressing cyclic dominance and stabilizing cooperation. Monte Carlo simulations confirm that such endogenous sanctioning mechanisms outperform static ones in sustaining metastable cooperative regimes.

Recent advances in equivariant and relational deep learning provide a principled mathematical interface between continuous symmetries, combinatorial structures, and neural function approximation. Bronstein et al. [102] formalise Geometric Deep Learning (GDL) as the study of neural architectures that are equivariant to a group action $\rho :G\times \Omega \to \Omega$ on the input domain $\Omega$. A layer $F:\mathcal{X}(\Omega )\to \mathcal{Y}\left({\Omega }^{\prime }\right)$ is equivariant if

$$\forall g\in G:F\left(\rho \left(g\right)x\right)={\rho }^{\prime }\left(g\right)F\left(x\right),$$

(168)

which, for linear F, implies that admissible kernels are group convolutions

$$\left(F\ast \psi \right)\left(u\right)=\sum _{g\in G}F\left(g·u\right)\psi \left(g\right),u\in \Omega .$$

(169)

The framework unifies grids, graphs, groups, geodesics, and their generalization to gauges: on principal bundles $(P,\Omega ,G)$ gauge-equivariant convolutions require the weight tensor to satisfy a Clebsch–Gordan-type coupling rule that selects only G-irreducible blocks. Discrete instances recover classical spectral GCNs, non-backtracking convolutions, and message-passing schemes as particular irreducible representations. The authors prove the following: (i) universal approximation of continuous equivariant maps; (ii) sample-complexity bounds that scale with the size of stabiliser subgroups of G.

Complementing the symmetry blackprint, Battaglia et al. [103] introduce the Graph-Network (GN) block, a universal “graph-to-graph” operator whose internal updates are

$$\begin{array}{cc}\hfill e_k^{\prime }& ={\varphi }^e\left(e_k,v_{s\left(k\right)},v_{r\left(k\right)},u\right),\hfill \\ \hfill {\overline{e}}_i& ={\rho }^{e\to v}\left\{e_k^{\prime }:r\left(k\right)=i\right\},\hfill \\ \hfill v_i^{\prime }& ={\varphi }^v\left({\overline{e}}_i,v_i,u\right),\overline{v}={\rho }^{v\to u}\left\{v_i^{\prime }\right\},u^{\prime }={\varphi }^u\left(\overline{e},\overline{v},u\right),\hfill \end{array}$$

where ${\varphi }^{\ast }$ are learnable, permutation-equivariant functions and the aggregators ${\rho }^{\ast }$ are permutation-invariant set functions. By weight-sharing on edges and nodes, the GN block instantiates a relational inductive bias that extends DeepSets’ universality theorem to arbitrary graph functions: finite compositions of GN blocks are universal approximators of any continuous, permutation-invariant graph-to-graph map. Empirically, stacked GN architectures exhibit combinatorial generalization on physical-interaction, multi-agent, and program-induction tasks.

The symmetry-equivariant calculus of Bronstein et al. and the GN message-passing formalism of Battaglia et al. jointly furnish a rigorous algebraic–analytic foundation for modern deep-learning treatments of dynamic complex networks. They unify group-theoretic convolutions, spectral graph theory, and permutation-invariant aggregation into a single operator framework, thereby extending classical belief-propagation concepts to neural architectures that provably respect the relational and symmetry constraints of networked data.

A major extension of belief propagation beyond fixed network structures concerns the inverse problem of structural inference: the reconstruction of latent topologies from sparse, noisy, or incomplete observations. This domain reframes learning as statistical estimation under topological uncertainty, where inference must simultaneously recover both signal and substrate. Vuffray et al. [104] introduce the Regularized Interaction Screening Estimator (RISE), a convex optimization method for exact recovery of pairwise binary graphical models. In the multiplex setting, Yu et al. [105] develop the Expectation–Maximization–Aggregation (EMA) framework to reconstruct layered networks from incomplete observations, defining integrated connectivity via redundancy-aware synthesis. These contributions expand the scope of inference from message passing to principled structure reconstruction, unifying variational inference, regularization theory, and probabilistic graphical modeling into a neuro-adaptive learning paradigm. Nitzan et al. [106] further address the inverse problem of network reconstruction from stationary distributions under external perturbations. Their method estimates the rows of the Jacobian matrix of the system dynamics by analyzing shifts in invariant probability densities across multiple perturbation experiments. Formally, for each node i, the observed shift vector ${\overrightarrow{I}}_i$ is approximately linearly related to the perturbation matrix $\Delta Z$ and the Jacobian row $J_i$ via

$${\overrightarrow{I}}_i\approx \Delta Z·J_i^T,$$

(170)

allowing inference through sparse recovery. This approach eliminates the need for temporal alignment or derivative estimation and exhibits robustness across deterministic and stochastic regimes.

Idrees et al. [107] introduce an end-to-end biophysical front-end that augments any convolutional retina model with light-adaptation dynamics governed by six coupled ODEs. This differentiable “photoreceptor layer” generalises across luminance regimes and captures retinal ganglion cell variance more accurately than convolution-only baselines. Naderi et al. [108] propose a dual-time-scale plasticity scheme enabling post-STDP learning in spiking networks through Tsodyks–Markram synaptic reservoirs, yielding biologically plausible continual adaptation.

Finally, Tishby and Zaslavsky [109] develop the Information Bottleneck (IB) principle, interpreting deep learning as a trade-off between compression of input X and predictive relevance for output Y, governed by a variational Lagrangian balancing $I(X;T)$ and $I(T;Y)$. Phase transitions in representations and generalization bounds emerge naturally within this information-theoretic framework. Hamilton, Ying, and Leskovec [110] complement this with a survey of graph representation learning techniques, including GraphSAGE and spectral embeddings such as GraphWave, which unify neighborhood aggregation and role-based encoding. These results reinforce a convergent trend: neural computation, representation learning, and structural inference are all governed by the same information-theoretic principles under constraints induced by network topology.

Idrees et al. [107] introduce an end-to-end biophysical front-end that augments any convolutional retina model with light-adaptation dynamics governed by six coupled ODEs. This differentiable “photoreceptor layer” generalises across luminance regimes and captures retinal ganglion cell variance more accurately than convolution-only baselines. Naderi et al. [108] propose a dual-time-scale plasticity scheme enabling post-STDP learning in spiking networks through Tsodyks–Markram synaptic reservoirs, yielding biologically plausible continual adaptation. Collectively, these models embed biophysical state variables inside neural architectures, producing hybrid systems that integrate control, memory, and sensory transduction in line with neurobiological computation.

8.7. Optimization and Design of Controllable Network Topologies

The synthesis of controllable and resilient network topologies necessitates optimization frameworks that integrate spectral robustness, connectivity constraints, and structural tradeoffs. Cohen et al. [111] demonstrate that scale-free networks with degree distribution $P\left(k\right)\sim k^{-\gamma }$ exhibit high robustness against random node failures but extreme vulnerability to targeted removal of high-degree nodes. The percolation threshold under intentional attack is determined by the condition

$$\kappa ={\displaystyle \frac{\langle k^2\rangle }{\langle k\rangle }}<2,$$

(171)

where $\kappa$ controls the emergence of the giant connected component. The systematic removal of hubs rapidly reduces $\langle k^2\rangle$, leading to abrupt structural fragmentation, which in turn degrades controllability due to the loss of reachability paths and matching structure. This fragility motivates the design of networks with constrained degree heterogeneity to optimize both resilience and structural controllability.

To maximize algebraic connectivity, Ghosh and Boyd [112] formulate a convex optimization problem over the edge weight matrix $w\in {\mathbb{R}}_{+}^{\left|E\right|}$, where the graph Laplacian $L\left(w\right)\in {\mathbb{R}}^{N\times N}$ is defined as $L\left(w\right)={\sum }_{(i,j)\in E}{w}_{ij}(e_i-e_j){(e_i-e_j)}^T$. The objective is to maximize the second-smallest eigenvalue ${\lambda }_2\left(L\left(w\right)\right)$, subject to total weight constraints $\sum w_{ij}\le W$. This leads to the semidefinite program

$$\underset{w\in {\mathbb{R}}_{+}^{\left|E\right|}}{max}{\lambda }_2\left(L\left(w\right)\right)\mathrm{subject}\mathrm{to}\sum _{(i,j)}{w}_{ij}\le W,$$

(172)

where ${\lambda }_2\left(L\left(w\right)\right)$ is concave over the convex domain. The dual problem yields sensitivity certificates for spectral improvement, and the optimal solution corresponds to a weight allocation that homogenizes effective resistance and increases synchronization potential while preserving cost-efficiency. The enhanced ${\lambda }_2$ also serves as a lower bound for the convergence rate in linear consensus dynamics.

Extending these ideas to delay-sensitive settings, Rafiee and Bayen [113] formulate a mixed-integer semidefinite program (MISDP) to design consensus networks under communication delay $\tau$ and sparse link constraints. Let $x_i\left(t\right)$ evolve as

$${\dot{x}}_i\left(t\right)=\sum _{j\in {\mathcal{N}}_i}{a}_{ij}(x_j(t-\tau )-x_i(t-\tau )),$$

(173)

where the adjacency matrix $A=\left[a_{ij}\right]$ is binary and encodes the design variables. To ensure exponential convergence with rate $\rho >0$, the authors require that the delayed system be stable under the Lyapunov–Krasovskii framework, which imposes matrix inequality constraints on the weighted Laplacian. The optimization problem becomes

$$\underset{A,\alpha }{min}\alpha \mathrm{subject}\mathrm{to}L\left(A\right)+\alpha I\succcurlyeq 0,\mathrm{and}\mathrm{card}\left(A\right)\le k,$$

(174)

where $\mathrm{card}\left(A\right)$ limits the number of nonzero elements, and $L\left(A\right)$ is the Laplacian induced by the binary adjacency matrix. This MISDP formulation jointly selects the optimal network topology and tuning parameters to ensure delay-tolerant consensus within minimal control effort and connectivity budget.

In the context of synchronizability, Donetti et al. [114] define the spectral ratio

$$R={\displaystyle \frac{{\lambda }_N}{{\lambda }_2}},$$

(175)

where ${\lambda }_2$ and ${\lambda }_N$ are the second-smallest and largest eigenvalues of the Laplacian, respectively. Minimizing R yields optimal synchronizability since the master stability function is narrow and synchronization thresholds are minimized. The authors prove that for fixed degree and network size, Ramanujan graphs minimize R asymptotically, satisfying the Alon–Boppana bound

$${\lambda }_2\ge d-2\sqrt{d-1},$$

(176)

for d-regular graphs. The optimization seeks topologies with large spectral gaps and suppressed eigenvalue spread, which in turn enhances robustness to noise and parameter mismatch in coupled oscillatory dynamics.

Grindrod and Higham [115] address time-resolved topologies by introducing dynamic communicability metrics based on matrix iterations. For a temporal sequence of adjacency matrices $\{A^{\left[1\right]},\dots ,A^{\left[T\right]}\}$, the dynamic communicability matrix is constructed as

$$\mathcal{Q}=\prod _{t=1}^T{(I-\alpha A^{\left[t\right]})}^{-1},$$

(177)

where $\alpha \in (0,{\rho }^{-1})$ is a resolvent parameter and $\rho$ is the spectral radius. The entry ${\mathcal{Q}}_{ij}$ quantifies the weighted influence of node j on node i via time-respecting walks. Temporal centrality is then defined as

$$c_i=\sum _{j=1}^N{\mathcal{Q}}_{ij},$$

(178)

capturing the accumulated influence of past activity. This dynamic framework generalizes Katz centrality and provides a time-aware measure of control propagation potential in evolving networks. It enables the detection of temporally localized control bottlenecks and supports the optimal scheduling of interventions in nonstationary environments. The synthesis of controllable network topologies thus emerges as a multidimensional optimization problem spanning percolation theory, semidefinite spectral design, delay-aware synthesis, eigenratio minimization, and dynamic communicability, each contributing principled constraints and performance criteria for the realization of robust, controllable, and functionally adaptive networked systems.

8.8. Distributed and Adaptive Consensus in MAS

The study of distributed consensus in multi-agent systems (MAS) has progressed toward adaptive architectures capable of handling heterogeneous dynamics, local gain tuning, and complex time-scale bifurcations. Li et al. [116] introduce an adaptive control protocol for continuous-time leader–follower MAS where each agent $i\in \{1,\dots ,N\}$ updates its state $x_i\left(t\right)\in \mathbb{R}$ via

$${\dot{x}}_i\left(t\right)=-\sum _{j\in {\mathcal{N}}_i}{a}_{ij}(x_i\left(t\right)-x_j\left(t\right))+b_i(x_i\left(t\right)-x_0\left(t\right)),$$

(179)

with $x_0\left(t\right)$ denoting the leader state, $a_{ij}$ denoting adjacency weights, and $b_i\left(t\right)$ a time-varying local gain. The gains are updated through an adaptive law

$${\dot{b}}_i\left(t\right)={\gamma }_i{(x_i\left(t\right)-x_0\left(t\right))}^2,$$

(180)

with adaptation rate ${\gamma }_i>0$. The design ensures finite convergence to consensus while compensating for unknown coupling strengths and asymmetric leader influence. Stability analysis is conducted using a composite Lyapunov function

$$V\left(t\right)={\displaystyle \frac{1}{2}}\sum _{i=1}^N{(x_i\left(t\right)-x_0\left(t\right))}^2+{\displaystyle \frac{1}{2{\gamma }_i}}\sum _{i=1}^N{(b_i\left(t\right)-b_i^{\ast })}^2,$$

(181)

with equilibrium values $b_i^{\ast }$ guaranteeing sufficient control authority. This adaptive formulation reduces reliance on global knowledge and permits distributed self-tuning of interaction gains, offering robustness to inter-agent variability and leader drift.

Extending into the discrete-time domain and addressing system heterogeneity, Zhang, Xu, Zhang, and Xie [117] reformulate consensus in heterogeneous multi-agent systems as a distributed linear quadratic (LQ) optimal control problem. Agents follow dynamics

$$x_i(k+1)=A{x}_i\left(k\right)+B_i{u}_i\left(k\right),$$

(182)

where the $B_i$ matrices differ across agents. The goal is to minimize the global cost

$$J(s,\infty )=\sum _{k=s}^{\infty }\left(\sum _{i=1}^N\sum _{j\in {\mathcal{N}}_i}{(x_i\left(k\right)-x_j\left(k\right))}^{T}Q(x_i\left(k\right)-x_j\left(k\right))+\sum _{i=1}^N{u}_i{\left(k\right)}^T{R}_i{u}_i\left(k\right)\right),$$

(183)

and achieve consensus without access to global information. By constructing distributed observers ${\hat{e}}_i\left(k\right)$ for local relative errors and designing Riccati-based feedback gains $K_{ei}$, the authors prove asymptotic convergence to the optimal centralized solution, formally establishing

$$\Delta J(s,\infty )=J^{★}(s,\infty )-J^{\ast }(s,\infty )\to 0\mathrm{as}s\to \infty .$$

(184)

This development introduces optimal control theory into the distributed consensus literature under realistic heterogeneity assumptions.

To accommodate structural heterogeneity and optimize performance, Zhang et al. [51] develop a distributed linear-quadratic (LQ) consensus protocol for MAS composed of agents with distinct dynamics. Each agent i evolves according to

$${\dot{x}}_i\left(t\right)=A_i{x}_i\left(t\right)+B_i{u}_i\left(t\right),$$

(185)

with local performance index

$$J_i={\int }_0^{\infty }\left(x_i{\left(t\right)}^T{Q}_i{x}_i\left(t\right)+u_i{\left(t\right)}^T{R}_i{u}_i\left(t\right)\right)dt,$$

(186)

subject to consensus constraints $x_i\left(t\right)-x_j\left(t\right)\to 0$ for all $(i,j)\in E$. A distributed Riccati-based method is used, wherein each agent solves an algebraic Riccati equation (ARE) parameterized by local $A_i,B_i,Q_i,R_i$ and communicates intermediate variables with neighbors. The optimal control law is given by

$$u_i\left(t\right)=-K_i{x}_i\left(t\right)+\sum _{j\in {\mathcal{N}}_i}{L}_{ij}(x_j\left(t\right)-x_i\left(t\right)),$$

(187)

where $K_i$ is the LQ-optimal gain and $L_{ij}$ implements consensus coupling. This formulation yields a Pareto-optimal equilibrium balancing local objectives and global coordination, and enables scalable synthesis in heterogeneous agent populations with minimal centralized computation.

Focusing on dynamical regimes beyond first-order consensus, Sun et al. [118] address second-order agents with dynamics

$$\begin{array}{cc}\hfill {\dot{x}}_i\left(t\right)& =v_i\left(t\right),\hfill \\ \hfill {\dot{v}}_i\left(t\right)& =u_i\left(t\right),\hfill \end{array}$$

(188)

and propose a consensus controller for agents arranged in chain topologies. The control input is

$$u_i\left(t\right)=-k_{p,i}\sum _{j\in {\mathcal{N}}_i}(x_i\left(t\right)-x_j\left(t\right))-k_{d,i}\sum _{j\in {\mathcal{N}}_i}(v_i\left(t\right)-v_j\left(t\right)),$$

(189)

where $k_{p,i},k_{d,i}$ are proportional and derivative gains. The closed-loop system is shown to converge exponentially to a common velocity and position trajectory, with convergence rate analytically characterized by the eigenvalues of a structured block Laplacian. The chain configuration induces unidirectional dependencies, facilitating simplified stability analysis using matrix-valued comparison theorems and enabling explicit design of decentralized gains.

In a markedly different setting, Jardón-Kojakhmetov and Kuehn [119] explore fast–slow bifurcations and canard-induced delays in adaptive consensus networks, where the adaptation law induces nonlinear time-scale separation. Consider a scalar slow variable $\epsilon \ll 1$ and a fast–slow system

$$\begin{array}{cc}\hfill \dot{x}& =f(x,y,\epsilon ),\hfill \\ \hfill \epsilon \dot{y}& =g(x,y,\epsilon ),\hfill \end{array}$$

(190)

where x denotes fast consensus dynamics and y controls adaptive weight evolution. Canard solutions, which traverse repelling slow manifolds for $\mathcal{O}\left(1\right)$ time intervals, induce metastable plateaus and abrupt transitions in consensus trajectories. The authors identify folded saddle-node singularities as organizing centers for delay-induced consensus failure and show that near-critical adaptation rates ${\epsilon }_c$ generate exponentially sensitive consensus timescales. This theory uncovers a hidden layer of dynamical complexity in adaptive protocols and illustrates how even structurally trivial systems can exhibit rich time-scale-induced control behavior. Such insights are critical in designing adaptive schemes that remain robust under small parameter drift and preserve consensus integrity across fast–slow domains.

8.9. Stochastic and Smoothed Analysis of Consensus Protocols

The incorporation of probabilistic and smoothed analytical techniques into consensus protocol evaluation provides a rigorous framework for characterizing convergence properties under randomness and local perturbations. Chan and Ning [120] develop a probabilistic model for distributed consensus in time-varying networks exhibiting random topological fluctuations, where agents update their states via

$$x_i(t+1)=\sum _{j=1}^N{a}_{ij}\left(t\right)x_j\left(t\right),$$

(191)

with $A\left(t\right)=\left[a_{ij}\left(t\right)\right]$ representing a sequence of stochastic row-stochastic matrices governed by local adaptive rules. The network is assumed to satisfy a graph expansion condition, where for any subset $S\subset V$ with $\left|S\right|\le \alpha N$, the neighborhood size satisfies $\left|N\right(S\left)\right|\ge (1+\delta )\left|S\right|$ for some $\delta >0$, ensuring sufficient diffusion capacity. Under these assumptions, the authors prove that consensus is achieved with high probability and that the mean-square deviation from the average decreases exponentially in time, with convergence time $T_{\epsilon }$ bounded by

$$T_{\epsilon }\le {\displaystyle \frac{C}{\delta }}log\left({\displaystyle \frac{N}{\epsilon }}\right),$$

(192)

for a universal constant $C>0$. The analysis leverages martingale concentration inequalities and exploits the mixing time of the time-dependent Markov chain induced by $A\left(t\right)$. The results demonstrate that local adaptation combined with global expansion properties suffices to guarantee robust consensus, even in the absence of deterministic connectivity or uniform step-size control.

In a complementary direction, Dinitz et al. [121] introduce a smoothed analytical perspective on classical distributed processes such as flooding, token aggregation, and random walk-based consensus. For an adversarial graph $G=(V,E)$ subject to random edge perturbations, they consider the expected performance of distributed protocols under Gaussian smoothing of edge weights, where for each edge $(i,j)$, the weight is perturbed as

$$w_{ij}^{\sigma }=w_{ij}+{\eta }_{ij},{\eta }_{ij}\sim \mathcal{N}(0,{\sigma }^2),$$

(193)

with $\sigma >0$ controlling the smoothing magnitude. The smoothed complexity is defined as the expected running time of the protocol under this perturbation model. For flooding processes, the smoothed broadcast time satisfies

$$\mathbb{E}\left[T_{\mathrm{flood}}^{\sigma }\right]\le \mathcal{O}\left({log}^{2}N\right),$$

(194)

even if the worst-case deterministic time is $\Omega \left(N\right)$. Similarly, for token aggregation via random walks, the cover time under smoothing improves from $\Theta \left(N^3\right)$ to $\tilde{\mathcal{O}}\left(N^{1.5}\right)$, where the tilde hides polylogarithmic factors. These results reveal that many distributed protocols that are fragile under worst-case topologies exhibit significantly improved performance when analyzed in the smoothed setting, where small random perturbations effectively regularize structural bottlenecks and enhance mixing. The theoretical machinery combines probabilistic bounds on Laplacian eigenvalues, concentration of spectral norms, and coupling arguments to quantify robustness under stochastic variability. By embedding stochasticity directly into the analytical structure of consensus dynamics, these approaches bridge worst-case and average-case performance and yield a refined understanding of distributed convergence in realistic, noisy, and perturbed environments.

8.10. Prediction and Self-Monitoring in Distributed Systems

The capacity of distributed systems to autonomously monitor their convergence behavior and infer predictive indicators of consensus quality without global observability constitutes a critical advancement for real-time control and resilience in decentralized architectures. Sirocchi and Bogliolo [122] propose a regression-based methodology for estimating the convergence rate of distributed consensus protocols using only locally accessible metrics, thereby circumventing the need for centralized supervision or full knowledge of the network topology. For a consensus process governed by the linear iterative rule

$$x_i(t+1)=\sum _{j\in {\mathcal{N}}_i}{a}_{ij}{x}_j\left(t\right),$$

(195)

each agent monitors the local variation of its state in time, forming a log-transformed local error metric ${\ell }_i\left(t\right)=log|x_i(t+1)-x_i\left(t\right)|$. Over a fixed observation window of length T, a least-squares linear regression is performed to fit ${\ell }_i\left(t\right)\approx {\beta }_{i}t+{\epsilon }_i\left(t\right)$, where ${\beta }_i$ is interpreted as a local estimate of the convergence exponent. The slope ${\beta }_i$ is used to approximate the dominant eigenvalue gap of the network Laplacian, and thereby infer global convergence properties from locally regressed quantities. The authors validate the estimator’s fidelity across multiple synthetic topologies and show that, despite its locality, the predicted convergence rates exhibit strong correlation with the true spectral decay rates of the system. This local inference paradigm enables self-diagnosing agents to evaluate their own convergence performance, issue warnings upon anomalous stagnation, and trigger adaptive control mechanisms in response to predicted slowdowns.

Complementing this estimation framework, Sirocchi and Bogliolo [123] further develop a topology-aware convergence profiling technique specifically adapted to asynchronous gossip consensus protocols. In such settings, agents communicate and update their states at randomly scheduled intervals, and the dynamics follow a stochastic update rule of the form

$$x_i(t+1)=\left\{\begin{array}{cc}{\displaystyle \frac{1}{2}}(x_i\left(t\right)+x_j\left(t\right))\hfill & \mathrm{with}\mathrm{probability}{p}_{ij},\hfill \\ x_i\left(t\right)\hfill & \mathrm{otherwise},\hfill \end{array}\right.$$

(196)

where $(i,j)\in E$ and $p_{ij}$ denotes the probability of selecting the communication pair $(i,j)$. To extract predictive structure from such asynchronous processes, the authors introduce a local profiling mechanism based on sampling inter-update intervals, fluctuation amplitudes, and cumulative deviation metrics. These observables are fed into a statistical model that classifies the convergence regime into fast, regular, or stalled, based solely on temporal patterns in locally recorded data. Topology-awareness is introduced through precomputed feature templates that link local degree, update frequency, and neighborhood spectral density to expected convergence profiles. This hybrid statistical-topological model allows agents to self-assess their convergence trajectory with high precision, even in networks with severe irregularity and variable latency. Moreover, the profiling architecture can be embedded in lightweight firmware for cyber–physical systems and sensor networks, thereby furnishing predictive self-monitoring capabilities with minimal computational and communication overhead. The integration of regression-based forecasting and topology-conditioned profiling constitutes a foundational step toward fully autonomous consensus infrastructures in large-scale distributed systems, where predictive control and self-awareness must compensate for limited global observability and asynchronous dynamics.

8.11. Domain-Specific Control in Energy and Mobility Systems

The implementation of distributed control protocols in energy and mobility systems demands precise adaptation to physical constraints, hardware limitations, and real-time response requirements inherent to their operational environments. Guan et al. [124] introduce a dynamic consensus algorithm (DCA) tailored to balance the state-of-charge (SoC) levels among distributed energy storage units within alternating current (AC) microgrids. Each unit i updates its control input $u_i\left(t\right)$ based on the deviation of its local SoC $s_i\left(t\right)$ relative to neighboring units through the continuous-time consensus law

$${\dot{s}}_i\left(t\right)=-\sum _{j\in {\mathcal{N}}_i}{a}_{ij}(s_i\left(t\right)-s_j\left(t\right)),$$

(197)

with $a_{ij}$ denoting the adjacency weights derived from electrical distance or communication topology. To ensure bounded total current exchange and compatibility with power electronics constraints, the DCA incorporates droop-based saturation functions and a hybrid filtering scheme that preserves convergence while suppressing high-frequency transients. The equilibrium distribution satisfies uniform SoC consensus across connected agents, thereby enhancing the energy throughput and lifespan of storage assets under nonstationary demand profiles.

In the context of voltage and reactive power control, Alsafran and Daniels [125] formulate a hierarchical protocol for virtual impedance correction, which achieves reactive power sharing among distributed generation units with heterogeneous line impedances. The reactive power injection $Q_i$ at unit i is governed by a nested control structure combining local impedance emulation and global consensus feedback. The virtual impedance $Z_i^v=R_i^v+j{X}_i^v$ is adjusted iteratively according to the update rule

$$Z_i^v(t+1)=Z_i^v\left(t\right)+{\gamma }_i\sum _{j\in {\mathcal{N}}_i}(Q_i\left(t\right)-Q_j\left(t\right)),$$

(198)

where ${\gamma }_i$ is a tunable step-size. This reactive power consensus mechanism stabilizes voltage profiles while ensuring fair resource allocation, even in the presence of asymmetric line parameters and mismatched generator capabilities. The approach exploits the frequency–power coupling of droop-controlled inverters and enables plug-and-play operation with minimal reconfiguration, meeting the scalability requirements of modern smart grid deployments.

Expanding the scope to mobility networks, Wu et al. [126] propose a Frobenius-optimized Hadamard-consensus algorithm for formation synchronization in unmanned aerial vehicle (UAV) fleets. The consensus objective is to align the position and velocity vectors $x_i\left(t\right)\in {\mathbb{R}}^n$ of UAV agents while minimizing the coordination error across communication-constrained channels. The system evolves under the discrete-time dynamics

$$x_i(t+1)=x_i\left(t\right)+ϵ\sum _{j\in {\mathcal{N}}_i}{H}_{ij}(x_j\left(t\right)-x_i\left(t\right)),$$

(199)

where $H\in {\mathbb{R}}^{N\times N}$ is a Hadamard matrix encoding structured interaction weights and $ϵ>0$ is the consensus gain. To optimize convergence performance, the authors minimize the Frobenius norm of the disagreement matrix

$$J\left(H\right)=\sum _{t=0}^T{∥X\left(t\right)-\overline{X}\left(t\right)∥}_F^2,$$

(200)

where $X\left(t\right)\in {\mathbb{R}}^{N\times n}$ is the state matrix of all agents and $\overline{X}\left(t\right)$ denotes the average trajectory. This leads to a combinatorial optimization over admissible Hadamard configurations subject to sparsity and communication bandwidth constraints. The resulting protocol achieves finite-time synchronization with reduced inter-agent signaling and robust tolerance to packet loss and delay jitter. The domain-specific adaptation of consensus schemes across energy and mobility infrastructures underscores the necessity of control architectures that reconcile physical process constraints with algorithmic tractability and performance guarantees, thereby enabling scalable and resilient operation in complex cyber–physical environments.

8.12. Controlling Epidemics on Structured Contact Networks

The control of infectious processes on heterogeneous networks requires theoretical tools that integrate structural complexity with epidemiological dynamics and account for feedback between node-level behavior and evolving contact patterns. A foundational model in this direction is the transmissibility-based threshold theory of Pourbohloul et al. [127], which links empirical contact structures to the global onset of epidemic spread via the critical transmissibility $T_c$. For a contact network $G=(V,E)$ with degree distribution $P\left(k\right)$, the average excess degree is given by

$$\langle k_{\mathrm{ex}}\rangle ={\displaystyle \frac{\langle k^2\rangle -\langle k\rangle }{\langle k\rangle }},$$

(201)

leading to the epidemic threshold

$$T_c={\displaystyle \frac{1}{\langle k_{\mathrm{ex}}\rangle }}={\displaystyle \frac{\langle k\rangle }{\langle k^2\rangle -\langle k\rangle }}.$$

(202)

This criterion generalizes the classical $R_0$-based approach by incorporating degree heterogeneity. In particular, for scale-free topologies with diverging $\langle k^2\rangle$, the threshold vanishes in the thermodynamic limit, implying that minimal transmission suffices to sustain large-scale outbreaks. Extensions of the model account for edge-specific interventions via effective transmissibilities $T_{ij}$, yielding a renormalized threshold

$$T_c^{\mathrm{eff}}={\displaystyle \frac{\langle k\rangle }{{\sum }_{k}k(k-1)P\left(k\right)\theta \left(k\right)}},$$

(203)

where $\theta \left(k\right)\in [0,1]$ quantifies degree-dependent intervention efficacy.

To address the dynamic restructuring of contact networks, Shaw and Schwartz [128] propose a pulse vaccination model (SIV) incorporating adaptive rewiring, periodic immunization, and class-dependent attachment. Let $P_S,P_I,P_V$ denote the fractions of susceptible, infected, and vaccinated nodes. The dynamics are governed by

$$\begin{array}{cc}\hfill {\dot{P}}_S& =r{P}_I-{\displaystyle \frac{pK}{N}}{P}_{SI}-\nu A{P}_S+q{P}_V,\hfill \\ \hfill {\dot{P}}_I& ={\displaystyle \frac{pK}{N}}{P}_{SI}-r{P}_I,\hfill \\ \hfill {\dot{P}}_V& =\nu A{P}_S-q{P}_V,\hfill \end{array}$$

(204)

where $\nu$ and A characterize vaccination frequency and amplitude, and w is the rewiring rate affecting $P_{SI}$. Preferential attachment to vaccinated nodes reshapes the network structure in response to epidemic states. Both mean-field approximations and Monte Carlo simulations confirm that adaptivity lowers extinction thresholds, enhancing control efficiency.

The inclusion of temporally self-exciting dynamics in contagion modeling is addressed by Zino et al. [129], who couple activity-driven networks with Hawkes processes (ADN+HP). Each node follows a mutually exciting activation process, modulated by its infection status. Infected nodes reduce their activation rate by a suppression factor $\rho$, weakening their connectivity. The epidemic threshold is derived analytically as

$${\sigma }_{\mathrm{HP}}={\displaystyle \frac{1}{\Lambda }},$$

(205)

where $\Lambda$ encapsulates both excitation kernel parameters and second-order statistics of background activity. This formulation allows one to invert the threshold for the critical suppression ${\rho }^{\ast }$ and design optimal isolation strategies for highly active agents.

To capture the influence of network heterogeneity in a more granular form, the degree-based heterogeneous mean-field (HMF) formalism provides a coarse-grained dynamical description. Each node is assigned to a degree class k, and the compartmental fractions $I_k\left(t\right),S_k\left(t\right),H_k\left(t\right),R_k\left(t\right)$ denote the corresponding epidemiological states. Assuming locally tree-like structure and conditional degree independence $P\left(k^{\prime }\right|k)$, the dynamics are governed by

$$\begin{array}{cc}\hfill {\displaystyle \frac{d{I}_k}{dt}}& =-k{I}_k\sum _{k^{\prime }}{S}_{k^{\prime }}P\left(k^{\prime }\right|k),\hfill \\ \hfill {\displaystyle \frac{d{S}_k}{dt}}& =\lambda k{I}_k\sum _{k^{\prime }}{S}_{k^{\prime }}P\left(k^{\prime }\right|k)-\alpha k{S}_k\sum _{k^{\prime }}[S_{k^{\prime }}+H_{k^{\prime }}+R_{k^{\prime }}]P\left(k^{\prime }\right|k)\hfill \\ \hfill & -\delta S_k+\xi H_k+\eta k{H}_k\sum _{k^{\prime }}{S}_{k^{\prime }}P\left(k^{\prime }\right|k),\hfill \\ \hfill {\displaystyle \frac{d{H}_k}{dt}}& =\delta S_k-\xi H_k-\eta k{H}_k\sum _{k^{\prime }}{S}_{k^{\prime }}P\left(k^{\prime }\right|k),\hfill \\ \hfill {\displaystyle \frac{d{R}_k}{dt}}& =\beta k{I}_k\sum _{k^{\prime }}{S}_{k^{\prime }}P\left(k^{\prime }\right|k)+\alpha k{S}_k\sum _{k^{\prime }}[S_{k^{\prime }}+H_{k^{\prime }}+R_{k^{\prime }}]P\left(k^{\prime }\right|k),\hfill \end{array}$$

(206)

as derived from the Master Equation framework [130]. The final exposure size is computed as

$$R=\sum _{k}P\left(k\right)\left(1-e^{-k{\varphi }_{\infty }}\right),{\varphi }_{\infty }={\displaystyle \frac{\langle k^2\rangle \langle k\rangle }{\langle k^3\rangle \left(\frac{1}{2}\langle k\rangle +\alpha \langle k^2\rangle C\right)}},$$

(207)

where C is an integral term capturing cumulative interactions. On networks with fat-tailed degree distributions ($\langle k^3\rangle \to \infty$), no finite epidemic threshold exists. The inclusion of cognitive recall terms ($\xi ,\eta$) modulates long-term dynamics, demonstrating how memory amplifies spreading potential and bifurcation richness.

Finally, Zhou et al. [131] analyze a continuous-time SIS model on interconnected small-world networks with asymmetric coupling and layer-specific infection parameters. The system evolves as

$$\begin{array}{cc}\hfill {\displaystyle \frac{d{\rho }_A}{dt}}& =-{\mu }_A{\rho }_A+{\lambda }_A\langle k_A\rangle {\rho }_A(1-{\rho }_A)+{\lambda }_{BA}\langle k_{BA}\rangle {\rho }_B(1-{\rho }_A),\hfill \\ \hfill {\displaystyle \frac{d{\rho }_B}{dt}}& =-{\mu }_B{\rho }_B+{\lambda }_B\langle k_B\rangle {\rho }_B(1-{\rho }_B)+{\lambda }_{AB}\langle k_{AB}\rangle {\rho }_A(1-{\rho }_B),\hfill \end{array}$$

(208)

where ${\rho }_A\left(t\right)$, ${\rho }_B\left(t\right)$ are layer-wise prevalence rates. The model reveals dynamic resonance in smaller layers and highlights trade-offs between model interpretability and structural specificity.

Together, these models constitute a comprehensive toolkit for the design and analysis of epidemic control on structured contact networks. They enable integration of topological heterogeneity, temporal dynamics, behavioral adaptivity, and memory, thereby supporting real-time optimization of intervention strategies under realistic population structures.

8.13. Theoretical Syntheses and Methodological Overviews

The integration of epidemic modeling across stochastic, deterministic, and spectral frameworks has yielded a unified understanding of contagion dynamics as governed by both microscopic interaction processes and macroscopic structural constraints. Kiss, Miller, and Simon [132] offer a comprehensive hierarchy of epidemic models encompassing exact Markovian formulations, pairwise and higher-order moment closures, and mean-field approximations, systematically identifying the conditions under which stochastic models can be reduced to deterministic counterparts without sacrificing critical dynamical features. For a networked SIS process with infection and recovery rates $\beta$ and $\mu$, respectively, the exact stochastic formulation governs the probability $P(\mathbf{x},t)$ over binary state vectors $\mathbf{x}\in {\{0,1\}}^N$. To circumvent the exponential state-space complexity, the authors introduce a hierarchy of closures, starting with the pairwise approximation for expected node and link states:

$${\displaystyle \frac{d}{dt}}\langle X_i\rangle =-\mu \langle X_i\rangle +\beta \sum _{j\in {\mathcal{N}}_i}\langle X_j{S}_i\rangle ,$$

(209)

$${\displaystyle \frac{d}{dt}}\langle X_j{S}_i\rangle =-(\mu +\beta )\langle X_j{S}_i\rangle +\beta \sum _{k\in {\mathcal{N}}_i\setminus j}\langle X_k{S}_j{S}_i\rangle -\beta \langle X_j{S}_i\rangle .$$

(210)

The authors demonstrate that these coupled moment equations can faithfully replicate epidemic thresholds and transient dynamics under sparse network topologies and homogeneous degree distributions, establishing a clear taxonomy for model selection based on network density, clustering, and mixing patterns.

In parallel, spectral control methodologies have been developed to inform optimal containment strategies by leveraging eigenvalue bounds and graph-theoretic metrics. Nowzari, Preciado, and Pappas [133] introduce a general framework for spectral optimization in epidemic mitigation, wherein the control objective is to adjust the effective infection rate matrix $B=\mathrm{diag}\left({\beta }_i\right)A$ such that the largest eigenvalue ${\lambda }_1\left(B\right)$ remains below the epidemic threshold $\mu$. Given budget constraints on intervention costs $c_i\left({\beta }_i\right)$, the authors formulate a constrained convex optimization problem:

$$\underset{{\beta }_i\in [0,{\beta }_{max}]}{min}{\lambda }_1\left(B\right)\mathrm{subject}\mathrm{to}\sum _{i=1}^N{c}_i\left({\beta }_i\right)\le C,$$

(211)

where A is the adjacency matrix of the contact network and C is the total budget. The minimization of ${\lambda }_1\left(B\right)$ ensures global exponential stability of the disease-free equilibrium in the mean-field approximation. Through semidefinite programming and spectral radius sensitivity analysis, the authors derive optimal resource allocation schemes that prioritize central nodes while preserving submodular efficiency. This approach is shown to outperform heuristic-based strategies in simulations over real-world mobility and social networks.

The spectral formulation is further contextualized by Pastor-Satorras et al. [134], who interpret epidemic spreading as a nonequilibrium phase transition governed by the leading eigenvalue of the adjacency or non-backtracking matrix. For the SIS model, the basic reproductive number $R_0$ admits a spectral threshold approximation:

$$R_0\approx {\displaystyle \frac{\beta }{\mu }}{\lambda }_1\left(A\right),$$

(212)

where ${\lambda }_1\left(A\right)$ is the largest eigenvalue of the adjacency matrix. The transition from a disease-free to an endemic state occurs at $R_0=1$, yielding a critical infection rate ${\beta }_c=\mu /{\lambda }_1\left(A\right)$. In networks with power-law degree distributions and divergent ${\lambda }_1\left(A\right)$, this threshold vanishes in the large-system limit, leading to vanishing epidemic thresholds and structural fragility. The authors extend the theory to accommodate quenched disorder, temporal networks, and higher-order interactions, identifying universality classes of spreading dynamics and critical exponents for various topologies. This synthesis consolidates the role of spectral graph theory as a unifying analytical lens for epidemic modeling, connecting stochastic simulation, deterministic approximation, and control-theoretic optimization within a coherent and scalable theoretical scaffold.

The organization introduced through this taxonomic structure reveals several deep unifying principles that traverse otherwise disparate literatures. Spectral graph theory, for example, recurs as a foundational tool across consensus dynamics, synchronization control, and spectral optimization, while Lyapunov-based stability and matrix inequality formulations appear ubiquitously in nonlinear and stochastic settings. Likewise, structural controllability frameworks grounded in combinatorics and matching theory provide topological foundations for actuator and sensor placement, which are subsequently extended into switched, multilayer, and adaptive regimes. The lower portion of the taxonomy highlights the growing role of learning, estimation, and cognitive modeling in control design, particularly under partial observability and uncertain dynamics. Taken together, the reviewed approaches suggest a converging trajectory toward hybrid control architectures that integrate model-driven analysis, data-driven adaptation, and structural constraint management. As networked systems continue to increase in complexity, this synthesis points to scalable, modular, and context-aware control schemes as the central direction for future research.

9. Discussion and Conclusions

9.1. Synthesis of Core Findings

The results presented in this study substantiate three core contributions to the theory and practice of control and intervention in dynamical complex networks. First, at the theoretical level, the proposed two-dimensional taxonomy synthesises a wide array of methodologies, ranging from spectral consensus and controllability analysis to learning-based control heuristics, into a unified framework that systematically classifies control strategies across five structural paradigms and five dynamical intervention domains. This structural–functional mapping reveals latent symmetries and constraints in previously siloed models, demonstrating that disparate techniques often occupy homologous positions within the control-design space. Second, from an empirical standpoint, the taxonomy’s utility is corroborated through multilayer benchmark evaluation: each structural regime, when subjected to cross-domain interventions, yields consistent response profiles and stability envelopes, thereby confirming the robustness and generality of the classification. Notably, cognitive and stochastic formulations exhibit particular sensitivity to structural heterogeneity, suggesting nontrivial interactions between network topology and intervention efficacy. Third, algorithmically, the proposed suite of optimization heuristics, spanning actuator placement, feedback scheduling, and control gain adaptation, achieves significant performance improvements relative to established baselines. These gains are observed not only in convergence rates and energy consumption but also in structural resilience metrics under perturbation, reinforcing the claim that principled taxonomy-aware design directly informs more effective network control strategies. The resulting classification is summarized in Figure 9, which visualizes the correspondence between structural regimes and control intervention types.

9.2. Integration with Prior Theory

The present framework establishes a systematic correspondence between structural modeling paradigms and domain-specific control interventions, thereby enabling a direct theoretical dialogue with canonical formulations across the control of complex networks. In the spectral and linear domain, the delay-robust consensus bounds derived under heterogeneous yet time-invariant Laplacian topologies generalize classical convergence criteria based on algebraic connectivity by allowing for arbitrary bounded delays and incorporating the influence of higher-order spectral moments. These results demonstrate that delay tolerance is not exclusively governed by the spectral gap but emerges from a joint dependency on eigenvalue dispersion and the topology-induced sparsity of control inputs. In the adaptive and heterogeneous setting, the proposed learning-aided Gramian framework provides an analytical alternative to iterative optimal policy construction via dynamic programming, yielding energy-efficient controllability estimates through time-averaged surrogates. By embedding structural priors into the update rule, this approach reduces sample complexity and accelerates convergence even in partially observable regimes. Within the temporal and stochastic domain, the formulation of control-driven rumor and epidemic interventions extends previous threshold-based models by introducing dynamic modulation of interlayer couplings and cognitively constrained propagation rates. This permits a refined characterization of stabilization regimes through spectral radius bounds under switching topologies, capturing both structural heterogeneity and temporal variability. At a conceptual level, the proposed taxonomy advances prior views on structure–function relationships by reinterpreting them not as static descriptors but as active design principles. Rather than identifying emergent functions from observed structural features, the framework operates in the inverse direction, guiding the optimization of network structure in service of targeted dynamical outcomes. This transition from inferential observation to constructive synthesis reframes structure–function duality as an algorithmically actionable paradigm for control and intervention.

9.3. Mechanistic Interpretations

The emergent regularities observed in the taxonomy, particularly the recurrent clustering of certain control strategies across structurally dissimilar settings, can be attributed to shared underlying mechanisms governing information diffusion and energy transfer in networked dynamical systems. The empirical alignment between controllability metrics and observability radius under heterogeneous coupling reflects a deep spectral symmetry: both quantities depend not only on the graph topology but also on its modal decomposition, specifically on the alignment between input–output nodes and the eigenvectors of the system matrix. In networks with heterogeneous edge weights or irregular degree distributions, control energy and estimation precision become coupled through the principal directions of system excitation and reconstruction. Mathematically, the minimal control energy for driving the system from the origin to a target state over a finite horizon is given by the quadratic form associated with the inverse of the controllability Gramian, whereas the observability radius quantifies the smallest perturbation to system parameters that renders the system unobservable. In structurally heterogeneous systems, both metrics are minimized along eigenmodes of high alignment with input–output channels, which explains their clustering across structurally diverse models. In the spectral reinforcement domain, the pronounced influence on synchronization criticality arises from a resonance mechanism: when the control input selectively amplifies subdominant eigenmodes that are otherwise suppressed in free dynamics, the resulting modification of the effective Laplacian spectrum reduces the critical coupling threshold for global phase coherence. This effect is most prominent in networks with marginal spectral gaps or disconnected quasi-components, where targeted spectral reshaping enhances global alignment without necessitating uniform coupling. Formally, the critical synchronization condition depends on the second smallest eigenvalue of the augmented Laplacian, and reinforcement schemes that raise the spectral floor induce a nonlinear contraction in the convergence basin. Lastly, the dominance of belief-based propagation in multilayer systems under cognitive constraints reflects the mismatch between structural redundancy and perceptual bandwidth. In systems with multiple layers of interaction, naive broadcast or epidemic-style interventions often lead to information saturation, wherein redundant exposures desensitize nodes or trigger competitive inhibition. Belief-based schemes, by contrast, exploit Bayesian update mechanisms that account for prior state and source credibility, thereby filtering redundant signals and prioritizing novel or dissonant inputs. This selective propagation preserves attention and memory resources while maintaining information fidelity. In network terms, this corresponds to a flow that avoids overloading high-degree hubs and instead traverses cognitive shortest paths, minimizing overlap and maximizing effective spread within bounded rationality constraints. The superiority of such mechanisms in multilayer control is therefore not merely algorithmic but mechanistic, grounded in the interplay between structural redundancy and informational selectivity.

9.4. Practical Implications

The theoretical architecture developed in this study yields a set of directly translatable guidelines for the design and deployment of control interventions in both engineered and socio-technical systems. In the context of networked control applications, the structure-informed scaling laws derived for the controllability Gramian illuminate nontrivial principles for actuator and sensor allocation, demonstrating that effective placement is not solely determined by node degree or centrality, but by deeper spectral characteristics that mediate the energy distribution across system modes. This insight is particularly relevant for large-scale cyber–physical systems where actuation and observation resources are severely constrained and must be optimally configured to guarantee stabilizability and identifiability. Furthermore, our formulation of epidemic mitigation strategies produces a quantifiable trade-off map between the cost of intervention and the temporal efficiency of suppression, enabling control designers to select context-sensitive strategies that explicitly balance resource expenditure against the urgency of containment. In socio-technical domains, the findings carry salient policy implications. The superiority of multilayer rumor-control schemes under cognitive constraints reveals a foundational mechanism for structuring public information campaigns across fragmented media ecosystems: by exploiting heterogeneity in interlayer transmission credibility and attentional fatigue, it becomes possible to design adaptive narratives that selectively reinforce corrective messages while suppressing redundant or counterproductive diffusion. This opens a viable pathway for institutions and governance frameworks to act not merely as reactive moderators but as algorithmically guided curators of informational environments. In parallel, the adaptive pinning frameworks proposed in the taxonomy align closely with infrastructural resilience demands in critical systems such as electrical grids and transportation networks. These protocols offer a scalable, decentralized mechanism for micro-islanding, localized stabilization, or re-routing under duress, while remaining compatible with real-time observability constraints and dynamically changing topologies. Across both domains, the taxonomic formalism introduced here serves not only as a descriptive classification but as an operational template, enabling principled navigation of a multi-objective control landscape where cost, feasibility, speed, and robustness are interdependent.

9.5. Limitations

Despite the breadth of structural regimes and control paradigms encompassed by the proposed taxonomy, several intrinsic limitations constrain the generalizability and operational deployment of our results. From a modeling perspective, the assumption of time-invariant Laplacian operators within each one-second macro-step, although analytically tractable, imposes a quasi-static view on systems whose coupling topologies may evolve at sub-second or asynchronous rates. This piecewise-stationary treatment captures mesoscopic stability but may obscure transient structural perturbations or micro-dynamic topological rewiring, particularly in socio-technical and biological networks. Furthermore, the omission of stochasticity in edge weights, both in terms of noise and temporal fluctuation, constitutes a significant simplification, as real-world communication, power, and transport systems routinely experience variability in link strength due to environmental, behavioral, or load-induced factors. On the data side, our empirical validation is limited by the availability of multilayer network datasets that satisfy the necessary resolution, annotation, and temporal granularity constraints. While the three benchmark systems employed here are among the most detailed publicly available, their structural diversity remains insufficient to exhaustively test the taxonomy across all plausible domains of application. This restricts the statistical robustness of the performance gains observed and invites future work that incorporates larger, more heterogeneous corpora. Finally, the cubic computational complexity of the optimization routines underlying the adaptive control design presents a scalability bottleneck for systems exceeding ${10}^5$ nodes, as memory constraints and solver stability deteriorate rapidly in high-dimensional state spaces. Although heuristics and dimensionality-reduction techniques can partially alleviate this burden, a principled framework for sublinear-time approximation or distributed control synthesis remains an open frontier necessary for extending the taxonomy to planetary-scale networks or real-time decision-making scenarios.

9.6. Future Research Directions

Building upon the formal and empirical foundation articulated in this work, several lines of inquiry naturally emerge that would extend both the theoretical scope and practical relevance of the proposed control-intervention taxonomy. A primary direction involves the explicit modeling of stochastic perturbations in edge weights and the development of robustness certification methods grounded in random matrix theory. Such an approach would allow for control guarantees to be framed probabilistically, accommodating environmental fluctuations, measurement uncertainties, and protocol inconsistencies that pervade real-world infrastructures. Complementarily, the integration of reinforcement learning techniques in an end-to-end architecture for intervention scheduling on streaming multilayer graphs holds considerable promise. By coupling online observation with policy optimization, this paradigm could enable real-time adaptation to topological shifts and spreading dynamics without reliance on static priors or precomputed structural decompositions. To validate the assumptions underlying cognitive pinning and belief-based propagation strategies, systematic human-in-the-loop experiments are also warranted. These would test the predictive accuracy of modeled attention decay, bounded rationality, and saturation thresholds under controlled yet ecologically valid settings, thereby anchoring abstract models in behavioral realism. Finally, a particularly compelling frontier lies in the co-simulation of evolving network topologies with simultaneous structural controllability estimation. Rather than treating the graph as an exogenous input, this direction envisions coupled dynamics wherein the network’s structure itself responds to control exertion, giving rise to a feedback loop between intervention efficacy and structural plasticity. Such frameworks would support control designs that are not only state-adaptive but structure-aware, yielding strategies that are dynamically recalibrated as the system reorganizes under endogenous or exogenous pressure. These extensions collectively represent the synthesis of mathematical rigor, algorithmic generality, and cognitive plausibility, necessary for the maturation of network control science into a robust engineering discipline.

9.7. Concluding Remarks

The present taxonomy articulates a unifying theoretical framework that systematically links diverse structural modeling paradigms with the spectrum of control, optimization, and intervention strategies observed across complex networks. By organizing the landscape of engineered interventions according to five structural regimes—spectral, combinatorial, adaptive, temporal, and multilayer—and mapping these onto canonical and emerging control mechanisms, the framework dissolves longstanding disciplinary silos and enables a comparative analysis of control performance across qualitatively different topologies. Its generality permits direct application to a wide array of domains, ranging from micro-robotic swarms with precise actuation and local feedback to macro-scale socio-technical infrastructures characterized by partial observability, heterogeneous latency, and multiscale cognitive constraints. The central insight that emerges is that effective intervention design is not reducible to isolated tuning of system dynamics or structural topology but instead requires the joint optimization of structural embedding, dynamical evolution, and cognitive processing. This triadic synthesis empowers control strategies that are not only robust and scalable but also responsive to epistemic limitations and behavioral heterogeneity. As multilayer network representations become increasingly indispensable in modeling complex systems, the theoretical and algorithmic tools developed herein offer a principled foundation for advancing resilience engineering, enabling interventions that are context-aware, structurally adaptive, and dynamically anticipatory. The control of multilayer networks thus stands as a frontier of both mathematical innovation and applied systems science, poised to shape the next generation of infrastructure design, social policy, and intelligent autonomy in the face of rising systemic complexity.