The Social Physics of Polarization

The problem

Take a population of agents, each holding an opinion on a line, and let each one move towards the average of the neighbours whose opinion is within some distance ε of its own. Nothing else. Depending on ε the population ends in consensus, in two camps, or in many fragments, and the passage from one regime to the next is sharp. This is the bounded-confidence model, the simplest mechanism that produces polarisation from purely local, symmetric interactions. It is a statistical-mechanics problem in disguise: a many-body system with a control parameter, an order parameter, and transitions between phases, on which the tools of the degree apply directly.

Basis paper

R. Hegselmann and U. Krause, “Opinion dynamics and bounded confidence: models, analysis and simulation”, Journal of Artificial Societies and Social Simulation 5(3), 2 (2002), jasss.org/5/3/2.html. Open access, with every simulation described in enough detail to be rerun.

Companion: G. Deffuant, D. Neau, F. Amblard and G. Weisbuch, “Mixing beliefs among interacting agents”, Advances in Complex Systems 3, 87 (2000), the pairwise version of the same idea. Map of the field: C. Castellano, S. Fortunato and V. Loreto, “Statistical physics of social dynamics”, Reviews of Modern Physics 81, 591 (2009). The Medium article by Michael Ng is a readable first contact, not a source.

The path

1. Study. The statistical physics of social dynamics as laid out in the Castellano review: voter, Ising-like and bounded-confidence models, order parameters, and what “phase transition” means in a finite system of agents. Network science with networkx: lattices, random graphs, small-world and scale-free topologies and the measures that distinguish them. Good practice for stochastic simulation: seeds, ensembles, error bars, finite-size scaling.

2. Reproduce. Implement the Hegselmann–Krause model (synchronous averaging) and the Deffuant–Weisbuch model (random pairwise encounters) in Python. Reproduce the time-evolution plots, the number of final clusters as a function of ε, the approximate 1/(2ε) rule for the number of clusters, and the asymmetric-confidence cases. Matching the published figures is the checkpoint.

3. Find the gaps and go beyond.

• The original papers use complete mixing. Put both models on small-world and scale-free networks and measure how topology moves the thresholds.
• Add noise to the opinions and heterogeneity to ε across agents; look for the regimes where noise restores consensus or destroys it.
• Introduce repulsion (agents move away from distant opinions) and see whether polarisation becomes stable rather than transient.
• Treat the consensus-to-fragmentation passage as a phase transition: define an order parameter, run ensembles at several system sizes, and attempt finite-size scaling.
• If time allows, confront the model with a real opinion dataset (published survey time series) to see which regime a real population sits in.

What you need

Statistical physics from the degree; Python with NumPy, networkx and Matplotlib. No laboratory work. Suitable for one student.