Boltzmann social learning with heterogeneous rationality

Sylvie Chaddad, Yezekael Hayel, Vineeth Satheeskumar Varma, Nicolas GastSPIRE 2026

TL;DR

This paper introduces a new model for social learning where agents have varying levels of "rationality" (how strongly they stick to their preferences). The authors show that even with these differences, agents tend to converge to a consensus, meaning they all eventually prefer the same actions. This work is important because it provides a more realistic way to model how groups make decisions, accounting for individual differences in decision-making.

What Problem Does This Paper Address?

Imagine a group of people trying to decide on the best way to do something, like choosing a restaurant or a new technology to adopt. They observe what others are doing and adjust their own preferences. Traditional models often assume everyone in the group is equally "rational" – meaning they all weigh their options and new information in the same way. However, in reality, people have different levels of rationality. Some are very decisive and stick to their guns, while others are more easily swayed or explore more options. This paper tackles the problem of modeling social learning when this "rationality heterogeneity" exists. The authors aim to understand how these individual differences affect the group's final decision and whether they can still reach a common agreement (consensus).

Methods

The core of this paper is a novel social learning model that uses a Boltzmann (softmax) rule for agents to choose actions based on their private preferences. What's new is the introduction of agent-specific rationality parameters ($\gamma_i$). This means each agent $i$ has its own $\gamma_i$, reflecting how strongly they influence their action choice based on their preference scores.

To analyze the behavior of a large number of agents, the authors employ a deterministic mean-field approximation (MFA). Instead of simulating every single agent's stochastic choices, they use the expected values of these choices to create a deterministic update rule. This MFA allows for analytical tractability, especially for large populations. They then mathematically analyze the fixed points of this deterministic system, which represent stable states or equilibria. Key analytical tools include:

  • Fixed-point analysis: Determining conditions for the existence and uniqueness of stable states.
  • Jacobian analysis: Studying the eigenvalues of the system's Jacobian matrix to assess local asymptotic stability.
  • Contraction mapping theorem: Used to prove global convergence and uniqueness of the equilibrium under certain conditions.

The model is studied on a complete graph, meaning every agent can observe every other agent's actions.

Datasets

Not discussed in the paper. The paper focuses on theoretical modeling and simulation-based validation rather than using external datasets.

Key Insights and Ideas

  • Rationality Heterogeneity Doesn't Prevent Consensus: Even with agents having different levels of rationality, the system tends to converge to a state where all agents share the same preference vector. This is a significant finding, as it suggests that individual differences in decision-making don't necessarily lead to fragmentation or indecision in a group.
  • The "Consensus State" is Key: On a complete graph, any stable equilibrium (fixed point) must be a "consensus state," meaning all agents agree on the relative preference for each action. This is a strong structural property that simplifies understanding the system's long-term behavior.
  • Critical Threshold for Stability: The stability of the consensus equilibrium depends on the sum of the rationality parameters. When this sum is below a certain threshold, the consensus is locally asymptotically stable. Beyond this threshold, the system can exhibit multiple non-uniform equilibria.
  • Bifurcation Behavior: The paper demonstrates a bifurcation phenomenon. As the homogeneous rationality parameter increases, the system transitions from a single, stable uniform consensus to multiple distinct, non-uniform consensus states. This highlights how a change in a single parameter can qualitatively alter the system's behavior.
  • Mean-Field Approximation is Accurate: The simulations show that the deterministic mean-field approximation closely tracks the stochastic dynamics, especially as the number of agents increases. This validates the analytical approach and suggests the MFA is a powerful tool for understanding large-scale social learning.

Key Results

  • Existence of Fixed Points: The mean-field dynamics always admit at least one fixed point, which is the uniform point (all actions equally preferred).
  • Consensus Property: On complete graphs, all fixed points are consensus states, meaning agents share identical preference vectors.
  • Local Asymptotic Stability: If the sum of rationality parameters ($\sum \gamma_j$) is less than $2n$, the consensus fixed point is locally asymptotically stable.
  • Global Convergence: Under the same condition ($\sum \gamma_j < 2n$), the mean-field map is a contraction, guaranteeing a unique fixed point and global convergence from any initial state.
  • Multiple Equilibria: When rationality is homogeneous and sufficiently high ($\gamma > m$), there exist at least $m$ distinct non-uniform consensus fixed points.
  • Error Bounds: The expected error of the stochastic dynamics from the mean-field fixed point is shown to be of order $O(1/\sqrt{n})$, and the expectation of the state itself is closer to the fixed point, $O(1/n)$, justifying the MFA's accuracy.

Limitations and Future Work

  • Complete Graph Assumption: The primary equilibrium structure results are derived for a complete graph, where all agents observe each other. The paper acknowledges that real-world networks are often sparse and non-complete.
  • Fixed Action Set: The model assumes a fixed set of actions available to all agents.
  • Non-Dynamic Population: The number of agents ($n$) is assumed to be fixed over time.

Future work suggested by the authors includes:

  • Non-complete Graphs: Studying the dynamics and equilibrium structure on more realistic network topologies (sparse, connected, disconnected graphs).
  • Dynamic Action Sets: Exploring scenarios where agents can learn to create or abandon actions.
  • Agent Population Dynamics: Investigating models where the number of agents can change.
  • More Complex Heterogeneity: Extending the model to account for other forms of heterogeneity beyond just rationality parameters.

Summary updated

Abstract

This paper analyzes a novel social learning model in which, at each discrete time step, agents with private preferences repeatedly select actions via a softmax (Boltzmann) rule, and update their preferences based on public observations of others' choices. This work addresses a critical gap by introducing rational heterogeneity through agent-specific rationality parameters gamma-i. Unlike previous models, our approach accounts for the diverse ways individuals process social information by using a discrete-time deterministic mean-field approximation map. We establish fundamental equilibrium properties that were previously unexplored. In particular, we prove the existence of fixed points and show that, on complete graphs, every mean-field equilibrium is a consensus state, where all agents share identical preferences. We further derive sufficient conditions for the uniqueness of this equilibrium and its local asymptotic stability. Numerical simulations validate our theoretical findings and illustrate how rational heterogeneity and network structure interact to shape collective behavior in social learning systems.
BibTeX
@inproceedings{chaddad2026boltzmann,
  author = {Sylvie Chaddad and Yezekael Hayel and Vineeth Satheeskumar Varma and Nicolas Gast},
  title = {{Boltzmann social learning with heterogeneous rationality}},
  booktitle = {SPIRE},
  year = {2026},
}
Figure