🌀 nonlinear sciences

Modelling the spillover from online engagement to offline protest: stochastic dynamics and mean-field approximations on networks

This study proposes a coupled stochastic modeling framework to analyze how online social media engagement spills over into offline protests, demonstrating that offline surges depend on a critical transmission rate range and that mean-field approximation accuracy varies by network density.

Original authors: Moyi Tian, P. Jeffrey Brantingham, Nancy Rodríguez

Published 2026-06-08

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Moyi Tian, P. Jeffrey Brantingham, Nancy Rodríguez

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a rumor spreading through a high school. It starts with one person whispering to a friend, who tells another, and soon half the school is talking about it. Now, imagine that this rumor doesn’t just stay in the hallways—it spills out into the playground, causing a massive, organized protest.

This paper is essentially a mathematical recipe for understanding exactly how that spill-over happens. The authors, Tian, Brantingham, and Rodríguez, built a computer simulation to see how "online engagement" (like posting on social media) turns into "offline action" (like protesting in the streets).

Here is the breakdown of their findings, translated into everyday concepts:

1. The Two-Layer Cake

Think of society as a two-layer cake.

The Top Layer (Online): This is the digital world. People are either Uninterested (scrolling past), Engaged (liking, sharing, arguing), or Disengaged (lost interest and moved on).
The Bottom Layer (Offline): This is the physical world. People are either Non-Protesting (staying home) or Protesting (out on the streets).

The key question is: How does energy move from the top layer to the bottom?

2. The "Spillover" Valve

The paper identifies a crucial "valve" or transmission rate. Let’s call it the Action Button.

If the Action Button is too weak, people get angry online but never bother to leave their houses. The online buzz dies out without any real-world impact.
If the Action Button is too strong, everyone rushes to the streets immediately. But here’s the catch: if everyone goes at once, the "reservoir" of potential protesters empties out instantly. The protest burns out quickly because there’s no one left to keep the momentum going.
The Sweet Spot: For a big, sustained protest to happen, the rate at which online anger turns into offline action needs to be just right—not too slow, not too fast. It’s like pouring water from a pitcher; if you pour too fast, it splashes everywhere and wastes; if you pour too slow, it drips. You need a steady stream.

3. The Feedback Loop

It’s not a one-way street. The paper shows a feedback loop:

Online anger leads to offline protests.
But seeing people protest offline actually makes more people get angry online. It’s like a megaphone. The offline action amplifies the online noise, which in turn fuels more offline action.

4. The "R0" Number (The Viral Threshold)

In epidemiology, doctors use a number called $R_0$ to see if a virus will cause an epidemic. The authors borrowed this idea. They calculated a "Reproductive Number" for protests.

If this number is below 1, the protest fizzles out. It’s like a spark in the rain.
If this number is above 1, the protest can grow into a wildfire.
They found that this number depends heavily on how connected people are online and how quickly they lose interest.

5. Simple Maps vs. Detailed Maps (The Network Structure)

To predict these protests, the authors created different mathematical "maps" of how people are connected.

Simple Map (Single-Level): Assumes everyone is connected to everyone else equally. It’s like assuming every student in the school knows every other student.
Detailed Map (Pairwise): Tracks who is friends with whom more carefully. It’s like knowing that the basketball team talks to each other, and the drama club talks to each other, but they don’t mix much.

The Surprise Finding:

In synthetic (fake) computer networks that are sparse (people have few friends), the Detailed Map is much better at predicting what will happen.
However, when they tested this on real-world Facebook data, the Simple Map worked almost just as well as the Detailed Map.
Why? Real social networks are messy and complex in ways the math doesn’t fully capture (like cliques, clusters, and hidden relationships). So, adding more mathematical complexity didn’t necessarily make the prediction more accurate for real people. Sometimes, a simpler model is surprisingly good enough.

Summary in a Nutshell

This paper says that social media doesn’t just reflect society; it actively drives it. But for online anger to become offline protest, the "conversion rate" has to be in a specific Goldilocks zone. And while we can build complex mathematical models to predict this, real-world social networks are so chaotic that simpler models often do a surprisingly good job of capturing the big picture.

Technical Summary: Modelling the Spillover from Online Engagement to Offline Protest

Problem Statement
The paper addresses the mathematical modeling of how engagement on online social networks spills over into offline protest activities. While social media’s role in conflict and collective action is well-documented, existing mathematical frameworks for protest dynamics largely overlook the coupling between online information diffusion and offline physical mobilization. The authors aim to quantify this interplay, specifically investigating how online tension translates into offline action and how offline activity feeds back to influence online engagement. A central question is identifying the conditions under which small-scale online unrest escalates into large-scale offline protests.

Methodology
The authors propose a coupled, two-layer compartmental model represented as a Markov process on a multilayer network:

Online Layer (Information): Individuals are in one of three states: Uninterested ( $U$ ), Engaged ( $E$ ), or Disengaged ( $D$ ). Engagement spreads via social contacts with a transmission rate $\tau$ . Engaged individuals disengage at rate $\gamma_i$ .
Offline Layer (Physical): Individuals are in one of three states: Non-protesting ($NP$), Protesting ( $P$ ), or Recovered ( $R$ ). Only a fraction $\eta$ of online-engaged individuals transition to protesting. Protesters disengage at rate $\gamma_p$ .
Coupling Mechanisms:
- Online-to-Offline: Modeled by the rate $\eta$ , where engaged individuals become protesters. The model assumes a "tension-inhibitive regime," where participating in offline protest immediately moves an individual to the disengaged state online.
- Offline-to-Online: Modeled by a self-excitation rate $\theta$ , where the magnitude of offline protest activity motivates uninterested individuals to become engaged online, independent of direct social contact.

The authors derive deterministic mean-field approximations of varying complexity from the underlying stochastic process:

Fully-Mixed Population: A baseline model assuming random mixing.
Homogeneous Networks ( $k$ -regular): Two approximations are derived:
- Single-level (Node-level): Assumes uniform distribution of engaged nodes.
- Pairwise: Tracks edge dynamics ($[UE]$) and uses triplet closures to account for local network structure.
Heterogeneous Networks: Extensions of the single-level and pairwise approximations that track node states conditioned on degree, using closure assumptions based on degree distributions.

The accuracy of these approximations is evaluated against stochastic simulations (using the Gillespie algorithm) on synthetic networks (Erdős-Rényi, scale-free, Watts-Strogatz) and empirical networks (Facebook friendship and UC33 networks). The authors also analyze the basic reproductive number ( $R_0$ ) to determine thresholds for protest outbursts.

Key Contributions

Multilayer Framework: Development of a stochastic compartmental model that explicitly couples online information diffusion with offline protest dynamics, including bidirectional feedback.
Mean-Field Derivations: Derivation of single-level and pairwise mean-field approximations for both homogeneous and heterogeneous networks, allowing for analytical tractability.
Reproductive Number Analysis: Definition and analysis of $R_0$ for the stochastic and approximate models, revealing that the cross-layer transmission rate $\eta$ must fall within a critical range ("sweet spot") to sustain offline outbursts.
Empirical Validation: Comparison of approximation accuracy across synthetic and real-world networks, highlighting the limitations of mean-field methods in capturing complex structural features like clustering.

Results

Threshold Dynamics: In fully-mixed and homogeneous networks, an outburst occurs when $R_0 > 1$ . The analysis shows that if the online-to-offline transmission rate $\eta$ is too low, online flare-ups fail to mobilize offline protests; if $\eta$ is too high, the reservoir of online-engaged individuals is exhausted too quickly, also limiting offline activity. Thus, a moderate $\eta$ is required for substantial offline outbursts.
Approximation Accuracy on Synthetic Networks:
- On low-density homogeneous networks, the pairwise approximation significantly outperforms the single-level approximation.
- As network density (degree $k$ ) increases, the single-level approximation improves and converges toward the pairwise approximation’s accuracy.
- On heterogeneous synthetic networks (e.g., Erdős-Rényi), the pairwise approximation generally aligns more closely with stochastic simulations than the single-level model, particularly in low-average-degree regimes.
Approximation Accuracy on Real-World Networks:
- On two Facebook networks, the advantage of the pairwise approximation over the single-level approximation was minimal or negligible.
- Both approximations performed well when online engagement decayed monotonically (no outburst) but showed larger errors during outbursts.
- The authors attribute the limited gain from higher-order approximations in real-world networks to structural features not captured by mean-field methods, such as clustering.
Reproductive Number Comparison: The $R_0$ derived from the stochastic model and the single-level approximation closely overlap. The pairwise approximation’s $R_0$ threshold is more conservative (predicting outbursts in a smaller parameter region) at low degrees but converges with the others as degree increases.

Significance
The paper claims to provide a foundational mathematical framework for understanding the coupled dynamics of online and offline social movements. By deriving continuum models, the authors enable the use of dynamical systems tools to predict protest outbursts and identify critical parameter ranges. The study highlights that model selection should depend on network density: simpler models suffice for high-density networks, while more complex approximations are needed for low-density ones. However, the findings on real-world networks suggest that increasing model complexity does not necessarily enhance accuracy due to unmodeled structural features like clustering. The authors position this work as a first step toward integrating empirical data for parameter estimation and system identification to better understand how social media influences collective offline behavior.