# Opinion dynamics
Opinion dynamics is the study, using statistical-physics methods, of how the opinions of interacting agents evolve under simple local rules, producing collective outcomes such as consensus, polarization, or fragmentation. Each agent holds an opinion, either a discrete state or a continuous value, and updates it based on interactions with neighbors; the question is what macroscopic order emerges from the microscopic rule. The field goes back at least 50 years to Weidlich's 1971 sociodynamics model and the [voter model](https://www.edgepedia.org/voter-model) of Holley and Liggett (1975)[1](https://ar5iv.labs.arxiv.org/html/2201.01322), and was codified for physicists in the 2009 review *Statistical physics of social dynamics* by Castellano, Fortunato and Loreto[2](http://www.socialdynamics.it/wp-content/uploads/2010/11/RevModPhys.81.591.pdf).
| Key fact | Value |
|---|---|
| Critical confidence ε between polarization and consensus | 0.27 for the Deffuant–Weisbuch model, 0.19 for the Hegselmann–Krause model[3](https://ar5iv.labs.arxiv.org/html/1909.12089) |
| Consensus limit for a homogeneous Deffuant–Weisbuch model | Confidence radius 0.5, above which consensus occurs across a variety of network topologies[3](https://ar5iv.labs.arxiv.org/html/1909.12089) |
| Voter-model consensus time on a d-dimensional lattice | \(T_N \sim N^2\) (\(d = 1\)), \(T_N \sim N \ln N\) (\(d = 2\)), \(T_N \sim N\) (\(d > 2\))[2](http://www.socialdynamics.it/wp-content/uploads/2010/11/RevModPhys.81.591.pdf) |
| Voter-model consensus on scale-free networks | Linear in N for degree exponent γ ≥ 3, sublinear for γ < 3[2](http://www.socialdynamics.it/wp-content/uploads/2010/11/RevModPhys.81.591.pdf) |
| Majority-rule consensus time in the mean-field limit | Scales as ln N in the number of agents[4](https://journals.aps.org/prl/abstract/10.1103/PhysRevLett.90.238701) |
| Deffuant-model phases as ε decreases | Consensus, then two-group polarization, then fragmentation into three or more stationary opinions[1](https://ar5iv.labs.arxiv.org/html/2201.01322) |
| Empirical validation | Bounded confidence is one of the few mechanisms with some degree of empirical validation, via computerized experiments (Chacoma and Zanette 2015)[1](https://ar5iv.labs.arxiv.org/html/2201.01322) |
## The canonical models
The 2009 review organizes the field into distinct families: voter-model variants, majority-rule models, social impact theory models, the [Sznajd model](https://www.edgepedia.org/sznajd-model), and bounded-confidence models[2](http://www.socialdynamics.it/wp-content/uploads/2010/11/RevModPhys.81.591.pdf).
**Discrete models.** In the voter model, on a finite lattice the process always reaches consensus, with the winning opinion determined by the initial conditions[5](https://arpi.unipi.it/retrieve/e0d6c92c-cc61-fcf8-e053-d805fe0aa794/opDynChapterarxiv.pdf). In the majority-rule model, a group of agents is specified whose members then all adopt the local majority state[4](https://journals.aps.org/prl/abstract/10.1103/PhysRevLett.90.238701); this group-based updating can produce faster consensus than the voter model and has been used to study polarization and the impact of group size[6](https://link.springer.com/article/10.1140/epjb/s10051-025-01053-7). The Sznajd model belongs to the same discrete family and also shows phase transitions between consensus and polarization[6](https://link.springer.com/article/10.1140/epjb/s10051-025-01053-7).
**Continuous bounded-confidence models.** In bounded-confidence dynamics, agents hold opinions on a continuous interval and interact only with others whose opinions lie within a confidence radius ε. The justification is homophily: people are influenced by those whose opinions are close enough to their own[3](https://ar5iv.labs.arxiv.org/html/1909.12089). Two update rules dominate. In the Deffuant model, opinions lie in [0, 1] and individuals interact pairwise only if their opinion difference is below ε, with each interaction shrinking the gap proportionally to a convergence parameter μ[1](https://ar5iv.labs.arxiv.org/html/2201.01322). In the Hegselmann–Krause (HK) model, agent i holding opinion \(x_{i}\) interacts with all neighbors with opinions in \([x_{i} - \varepsilon,\, x_{i} + \varepsilon]\) simultaneously, averaging over every compatible neighbor at once[5](https://arpi.unipi.it/retrieve/e0d6c92c-cc61-fcf8-e053-d805fe0aa794/opDynChapterarxiv.pdf). Deffuant interactions are binary[3](https://ar5iv.labs.arxiv.org/html/1909.12089), whereas in HK agents interact with all compatible neighbors at the same time[5](https://arpi.unipi.it/retrieve/e0d6c92c-cc61-fcf8-e053-d805fe0aa794/opDynChapterarxiv.pdf).
## Consensus, polarization, and fragmentation
The confidence bound ε is the control parameter that decides the outcome. For high enough ε, Deffuant dynamics drive the system toward consensus, a state where all individuals share the same opinion value near the initial average. Below a certain value of ε the system polarizes into two opinion groups, and for still lower ε it fragments into three or more stationary opinions[1](https://ar5iv.labs.arxiv.org/html/2201.01322).
The transition points are known from simulation: the critical value between polarization and consensus is 0.27 for the Deffuant–Weisbuch (DW) model and 0.19 for the HK model[3](https://ar5iv.labs.arxiv.org/html/1909.12089). For a homogeneous DW model, where all agents share the same confidence level, the confidence radius 0.5 (half the opinion scale) has been demonstrated as the limit above which consensus occurs for a variety of network topologies[3](https://ar5iv.labs.arxiv.org/html/1909.12089).
The magnet analogy is explicit in the physics literature: opinion dynamics models such as majority rule, Sznajd, and several voter-model variants have demonstrated phase transitions showing that societies can switch between consensus and polarization states, and spontaneous symmetry breaking in sociophysics resembles continuous phase transitions in magnets[6](https://link.springer.com/article/10.1140/epjb/s10051-025-01053-7).
## By the numbers: consensus times
How long convergence takes depends sharply on dimension and topology.
**Voter model on lattices.** The time needed for a finite system to reach consensus scales as \(T_N \sim N^2\) for \(d = 1\), \(T_N \sim N \ln N\) for \(d = 2\), and \(T_N \sim N\) for \(d > 2\)[2](http://www.socialdynamics.it/wp-content/uploads/2010/11/RevModPhys.81.591.pdf) • [5](https://arpi.unipi.it/retrieve/e0d6c92c-cc61-fcf8-e053-d805fe0aa794/opDynChapterarxiv.pdf). The dimension also decides whether consensus happens at all: for \(d \leq 2\) the voter model undergoes coarsening leading to complete consensus, while for \(d > 2\) an infinite system retains a finite density of interfaces, so domains of opposite opinions coexist indefinitely[2](http://www.socialdynamics.it/wp-content/uploads/2010/11/RevModPhys.81.591.pdf).
**Voter model on scale-free networks.** For networks with scale-free degree distribution of exponent \(\gamma\), \(T_{\mathrm{N}}\) scales linearly in \(N\) for \(\gamma \geq 3\) and sublinearly for \(\gamma < 3\), in good agreement with simulations[2](http://www.socialdynamics.it/wp-content/uploads/2010/11/RevModPhys.81.591.pdf).
**Majority rule.** In the mean-field limit, where groups consist of randomly selected agents, majority-rule consensus is reached in a time scaling as ln N[4](https://journals.aps.org/prl/abstract/10.1103/PhysRevLett.90.238701).
## Networks, zealots, and noise
Most modern opinion dynamics models unfold on complex networks, including Erdős–Rényi, scale-free, and modular community-structured graphs, rather than on mean-field complete graphs or lattices; the complete graph remains useful as a mean-field baseline where full analytical tractability is possible[7](https://arxiv.org/html/2507.11521v2).
**Zealots and leaders.** Stubbornness is a standard extension: full stubbornness, where some agents never change opinion, is an important special case of openness and close-mindedness in models like Deffuant[7](https://arxiv.org/html/2507.11521v2). A single quenched zealot breaks the magnetization conservation of the voter model, and the effect is dimension-dependent: in d ≤ 2 the zealot influences all agents and induces general consensus with its opinion, while in higher dimensions consensus is still not reached[2](http://www.socialdynamics.it/wp-content/uploads/2010/11/RevModPhys.81.591.pdf). In a Deffuant variant with fixed-opinion leaders, the dynamics show a rapid first stage toward the average opinion followed by a slow second stage dominated by the leaders; when leaders disagree, the final state lacks consensus[1](https://ar5iv.labs.arxiv.org/html/2201.01322).
**Noise and heterogeneity.** Bounded-confidence dynamics depend strongly on initial conditions; noise decreases this dependence while adding dispersion to the final opinion distribution[1](https://ar5iv.labs.arxiv.org/html/2201.01322). Heterogeneity in confidence can work in surprising directions: mixing open-minded and closed-minded agents with two different confidence radii yields consensus below the critical radius of 0.5, counterintuitively[3](https://ar5iv.labs.arxiv.org/html/1909.12089).
## Recent developments
Recent reviews have pushed the field toward phenomena that earlier model families treated only implicitly.
**Echo chambers as a model ingredient.** Echo chambers, environments where individuals are primarily exposed to information aligned with their preexisting beliefs through selective exposure or algorithmic curation, can be formalized as increased opinion-based homophily in the interaction network. They have frequently been associated with both the emergence of polarization and the spread of misinformation, with the misinformation link drawing on 2023 work[7](https://arxiv.org/html/2507.11521v2).
**Algorithmic curation as dynamics.** [Algorithmic bias](https://www.edgepedia.org/algorithmic-bias) arising from content filtering is now treated as an extension of bounded-confidence models[1](https://ar5iv.labs.arxiv.org/html/2201.01322), and a 2025 review frames computational opinion dynamics, including bounded-confidence and adaptive-network models, as describing how network structure can amplify confirmation bias, filter information diversity, and lead to ideological extremization[8](https://arxiv.org/html/2510.15053).
**Taxonomy by phenomenon.** A 2025 review proposes categorizing opinion-dynamics models by the macroscopic phenomena they describe, such as consensus, fragmentation, and polarization, rather than by model class[7](https://arxiv.org/html/2507.11521v2).
## References
1. Peralta, Kertész, Iñiguez, *Opinion dynamics in social networks: From models to data*, Handbook of Computational Social Science (2023). https://ar5iv.labs.arxiv.org/html/2201.01322
2. Castellano, Fortunato, Loreto, *Statistical physics of social dynamics*, Rev. Mod. Phys. 81, 591 (2009). http://www.socialdynamics.it/wp-content/uploads/2010/11/RevModPhys.81.591.pdf
3. *From classical to modern opinion dynamics*, arXiv:1909.12089. https://ar5iv.labs.arxiv.org/html/1909.12089
4. Galam, *Dynamics of Majority Rule in Two-State Interacting Spin Systems*, Phys. Rev. Lett. 90, 238701 (2003). https://journals.aps.org/prl/abstract/10.1103/PhysRevLett.90.238701
5. *Opinion dynamics: models, extensions and applications*, book chapter, Università di Pisa repository. https://arpi.unipi.it/retrieve/e0d6c92c-cc61-fcf8-e053-d805fe0aa794/opDynChapterarxiv.pdf
6. *Sociophysics models inspired by the Ising model*, Eur. Phys. J. B (2025). https://link.springer.com/article/10.1140/epjb/s10051-025-01053-7
7. *Opinion dynamics: Statistical physics and beyond*, arXiv:2507.11521 (2025). https://arxiv.org/html/2507.11521v2
8. *The Physics of News, Rumors, and Opinions*, arXiv:2510.15053 (2025). https://arxiv.org/html/2510.15053
---
*Topic: Encyclopedia › Physical world and mathematics › Physics › Physics methods, practice, and community › Applied and interdisciplinary physics › Biophysics and cross-disciplinary physics › Econophysics and social physics › Opinion dynamics and consensus models*
*Initially written Sep 17, 2026 · Edited: Oct 11, 2026; Oct 15, 2026*
*Copyright 2026 EdgeChat AI, a subsidiary of Biostate AI.*
License: Edgepedia Community License 1.0, https://www.edgepedia.org/edgepedia/license