Source-linked AI summary
A Tutorial on Modeling and Analysis of Dynamic Social Networks. Part I
Anton V. Proskurnikov, Roberto Tempo
TL;DR
This tutorial addresses the limited application of modern control theory to social systems by examining classical models of opinion formation and their connections to multiagent control. It presents mathematical conditions for convergence and consensus, while also discussing persistent disagreement and model boundaries.
Problem
Modern control theory has remained almost untouched in social systems, motivating analysis of how models can represent both consensus and persistent community cleavage.
Method
The tutorial introduces preliminary concepts, examines four classical opinion-evolution models, and relates them to modern multiagent control through mathematical analysis.
Results
The analysis gives graph-theoretic conditions for convergence and consensus in the French-DeGroot model, including directed-spanning-tree conditions when self-weights are positive.
Takeaways & Limitations
Classical social-dynamics models provide mathematically analyzable cases for studying consensus, clustered opinions, and persistent disagreement within control theory.
Takeaways & Limitations
A directed spanning tree is not generally sufficient for consensus when the influence matrix has zero diagonal entries, and discontinuous mappings may lack classical solutions.
Abstract
from arXiv · showhide
In recent years, we have observed a significant trend towards filling the gap between social network analysis and control. This trend was enabled by the introduction of new mathematical models describing dynamics of social groups, the advancement in complex networks theory and multi-agent systems, and the development of modern computational tools for big data analysis. The aim of this tutorial is to highlight a novel chapter of control theory, dealing with applications to social systems, to the attention of the broad research community. This paper is the first part of the tutorial, and it is focused on the most classical models of social dynamics and on their relations to the recent achievements in multi-agent systems.
1. Introduction
The tutorial addresses the gap between social network analysis and control by reviewing mathematically mature models of dynamic social processes. Part I focuses on classical opinion-formation models and their connections to multi-agent control.
- Research gap: Social systems remained relatively untouched by modern control theory despite advances in controlling complex large-scale systems.
- Research gap: The gap reflects limited mathematical models and quantitative tools for analyzing and simulating large-scale social dynamics.
- Emerging field: Recent work has begun connecting social network analysis with dynamical systems through dynamic social network analysis, temporal networks, and sociodynamics.
- Tutorial scope: This tutorial concentrates on mature dynamic models and influential mathematical results, mainly concerning opinion formation under social influence.
- Tutorial scope: Part I introduces preliminary concepts, examines the French-DeGroot, Abelson, Friedkin-Johnsen, and Taylor models, and relates them to modern multi-agent control.
2. Opinions, Agents, Graphs and Matrices
This section introduces agent-based opinion dynamics, graph representations of social interactions, and matrix concepts used to analyze network structure. It emphasizes directed weighted graphs and the correspondence between graph connectivity and Laplacian properties.
- Opinion dynamics: Microscopic models describe how individual agents’ opinions evolve and can represent both small and large communities.
- Opinion dynamics: The tutorial studies agent-based models with real-valued scalar or vector opinions in a closed community whose size n ≥2 remains fixed.
- Graphs and matrices: Social interactions are represented by weighted directed graphs whose nodes correspond to agents and whose matrix entries encode influence gains.
- Graph connectivity: A quasi-strongly connected graph has at least one root and is equivalent to containing a directed spanning tree.
- Graph connectivity: A graph’s strong components partition its nodes, and a rooted graph has a unique closed strong component.
- Matrices and Laplacians: The Laplacian has a simple zero eigenvalue exactly when the associated graph is quasi-strongly connected.
3. The French-DeGroot Opinion Pooling
The French-DeGroot model represents opinion pooling through weighted influence among agents and connects opinion formation with social power and centrality. Its convergence and consensus depend on stochastic-matrix and graph properties, while stubborn agents can prevent consensus yet still determine the limiting opinions.
- 3.1. The French-DeGroot model of opinion formation: The French-DeGroot model updates agents’ opinions by x(k + 1) = Wx(k), where each stochastic weight wij measures agent j’s influence on agent i.The model also extends from scalar opinions to opinion matrices through X(k + 1) = WX(k).
- 3.2. History of the French-DeGroot model: French’s graph-based model is a special case in which each agent uniformly distributes influence among itself and connected neighbors.Its influence matrix has positive diagonal entries and equal nonzero weights within each row.
- 3.3. Algebraic convergence criteria: Convergence is equivalent to regularity of W, while consensus is equivalent to full regularity and requires λ = 1 to be the only unit-circle eigenvalue, with a one-dimensional eigenspace.For irreducible W, convergence is equivalent to primitivity, which also yields consensus.
- 3.4. Graph-theoretic conditions for convergence: The graph reaches consensus exactly when it is quasi-strongly connected and its only closed strong component is aperiodic.With positive self-weights, convergence is guaranteed and consensus requires a directed spanning tree; zero diagonal entries invalidate that sufficiency in general.
- 3.5. The dual Markov chain and social power: The limiting weight p∞i measures agent i’s social power, while stubborn agents can produce persistent disagreement and determine the final opinion vector.With multiple stubborn agents, consensus among them is impossible, but opinions typically converge; under the stated influence assumptions, the final vector is fully determined by stubborn agents’ opinions.
4. Abelson’s Models and Diversity Puzzle
Abelson’s continuous-time model recasts opinion adjustment as a consensus dynamics and is always convergent, with consensus characterized by graph connectivity. This raises the community cleavage problem: explaining persistent disagreement within models that otherwise tend toward agreement.
- Abelson’s continuous-time model: Abelson’s model is a continuous-time counterpart of the French-DeGroot model, obtained when opinion updates become arbitrarily frequent.Its dynamics are also rediscovered as a continuous-time consensus algorithm in multi-agent control.
- Convergence and stability: The linear Abelson model is Lyapunov stable and always convergent, unlike the French-DeGroot model, whose stability is not asymptotic.The distinction follows from the Laplacian spectrum: nonzero eigenvalues have positive real parts, while zero eigenvalue Jordan blocks are trivial.
- Convergence and stability: The limit matrix P∞ is a projection onto the Laplacian null space and is closely related to the graph structure.It determines the limiting opinions through the model’s equilibrium subspace.
- Consensus conditions: Consensus occurs if and only if the influence graph is quasi-strongly connected, meaning it has a directed spanning tree.In that case, opinions converge to a limit determined by the agents’ social-power vector.
- The community cleavage problem: Abelson’s diversity puzzle asks what assumptions can produce community cleavage when broad mathematical models instead yield universal agreement.The target phenomenon is persistent disagreement or opinion clustering, requiring models both socially realistic and mathematically tractable.
- Scope and caveat: The tutorial notes that discontinuous coupling functions can prevent the nonlinear Abelson system from having a classical solution.Nonlinear extensions are therefore not uniformly covered by the linear model’s analytical conclusions.
5. Cleavage and Prejudices: Taylor’s model
Taylor’s model extends opinion dynamics by adding static communication sources or prejudices, producing cleavage while typically yielding a unique asymptotically stable equilibrium. Its stability and containment-control interpretations depend on every agent being directly or indirectly influenced by a source.
- Taylor’s model: Taylor’s model adds static communication sources with opinions s1, ..., sm and represents their effects through persuasibility constants in matrix B.Agents may be unaffected by external sources or influenced by one or more of them.
- Cleavage and stability: External influence typically causes opinion cleavage, while the system usually becomes asymptotically stable and converges to a unique equilibrium determined by the source opinions.This contrasts with the Abelson model’s stability and convergence properties.
- Scope and caveat: The tutorial excludes Taylor’s nonlinear opinion dynamics from its scope because they were still awaiting rigorous mathematical examination.The stated scope is therefore centered on the linear model and its analyzed extensions.
- Model relations: Taylor’s model can be represented as an Abelson model augmented with stubborn virtual agents whose fixed opinions equal the communication-source opinions.This transformation explains why the model is convergent for arbitrary initial and source opinions.
- Prejudices and dependence: An agent is P-dependent when it is prejudiced or reachable from a prejudiced agent; otherwise it is P-independent.The resulting partition separates agents whose dynamics are stabilized by prejudice from those governed by an Abelson subsystem.
- Stability of the Taylor model: The dynamics of P-dependent agents are asymptotically stable, and their final opinions are convex combinations of prejudices and the final opinions of P-independent agents.The associated matrix M is stochastic.
- Stability of the Taylor model: The Taylor system is asymptotically stable if and only if every agent is directly or indirectly influenced by at least one communication source.Direct influence is represented by a positive source weight; indirect influence occurs through a chain of agents.
- Containment control: In the multidimensional extension, mobile agents reach the convex hull of static leaders exactly when every agent is directly or indirectly influenced by a leader.Theorem 20 equates Hurwitz stability, P-dependence, and reaching the target convex hull for all leader positions and initial conditions.
6. Friedkin-Johnsen Model
The Friedkin–Johnsen model extends opinion dynamics with agent susceptibilities and prejudices, yielding stability, convergence, and centrality results linked to network structure. Its limiting opinions combine prejudices with the final opinions of P-independent agents.
- Model formulation: The Friedkin–Johnsen model adds a diagonal susceptibility matrix Λ to stochastic social influences W and a prejudice vector u.Susceptibility complements determine how strongly prejudices affect each agent; Λ = 0 recovers the French-DeGroot model.
- Model formulation: P-dependent agents are prejudiced or reachable from prejudiced agents, while P-independent agents have λi = 1 and follow a French-DeGroot subsystem.The opinion dynamics decompose into a P-dependent subsystem and a stochastic W22 subsystem for P-independent agents.
- Convergence and stability: More generally, convergence requires either every agent to be P-dependent or regularity of the P-independent subsystem W22.The P-dependent subsystem is asymptotically stable, with ρ(Λ11W11) < 1.
- Convergence and stability: Final opinions are convex combinations of prejudices and the final opinions of P-independent agents because the matrix V is stochastic.The proof establishes V1n = 1r and nonnegativity through the M-matrix structure.
- Convergence and stability: The model is asymptotically stable exactly when all agents are P-dependent, including whenever Λ < In or Λ ≠ In with a strongly connected influence graph.Under these conditions, V = (I − ΛW)^−1(I − Λ).
- Influence centrality and PageRank: Friedkin’s construction produces centrality measures interpolating between uniform social power at α = 0 and French’s social power as α approaches 1 when it exists.For fully regular W, cα converges to French’s social power as α → 1−0; PageRank is a special case with Λ = (1 − m)In.
Concluding Remarks and Acknowledgements
The tutorial analyzes classical continuous- and discrete-time opinion-dynamics models and their relations to multi-agent systems. It leaves scalability, large-scale validation, convergence rates, and links between opinion behavior and community structure as open issues.
- Concluding remarks: The first tutorial part discusses several continuous- and discrete-time models for opinion formation and evolution.It provides rigorous convergence and stability analysis and relates the models to multi-agent-systems results.
- Limitations and open problems: Scalability issues remain outside the paper’s scope, including algorithmic convergence analysis, convergence rates, and experimental validation on big data.These omissions are attributed to the page limit.
- Limitations and open problems: An open problem is relating opinion behavior in the reviewed models to community or module structure in the network graph.The paper identifies this as an important unresolved connection.
- Future directions: The tutorial’s second part is planned to cover bounded confidence, antagonistic interactions, asynchronous gossip-based interactions, and future control perspectives.These topics are presented as extensions beyond the first part.