Network games and diffusion

Overview

Network games and diffusion analyze how individual decisions and behaviors interact through social or technological networks. In network games, each player’s payoff depends on their own action and the actions of their neighbors in a graph (the network). Strategic complements (neighbors’ higher actions increase one’s best response) amplify behavior; strategic substitutes dampen it. Diffusion studies how behaviors, technologies, information, or pathogens spread across the same networks via learning, imitation, or contagion. This lens matters for product adoption, pricing with peer effects, platform growth, cybersecurity, and public health, where outcomes depend not just on averages but on who influences whom.

Origins and Credit

Foundational work includes Mark Granovetter (1978) on threshold models of collective behavior and Everett Rogers on diffusion of innovations. In economics, Matthew Jackson (network economics; 2008) synthesized the field. Ballester, Calvó-Armengol, and Zenou (2006) linked equilibrium behavior in linear network games to Bonacich centrality and introduced the “key player” concept. Bramoullé and Kranton (2007) studied public goods on networks with strategic substitutes. On diffusion and contagion, influential contributions include Morris (2000) on contagion with strategic complementarities, and Kempe, Kleinberg, and Tardos (2003) on influence maximization in networks. Recent work integrates spectral methods, learning, and platform design.

Core Idea and Mechanics

Network games. Players i = 1,…,n choose actions x_i (e.g., adopting a tool, setting effort, price). Links G = [g_ij] encode who influences whom. A tractable benchmark is a linear-quadratic payoff: an agent’s marginal benefit is increasing in a weighted sum of neighbors’ actions. The best responses are linear, for example:

x_i = α + β · ∑_j g_ij x_j, where β > 0 implies strategic complements and β < 0 strategic substitutes.

In vector form, x = (I − βG)⁻¹ α1 if the matrix inverse exists. Existence and uniqueness require a “small-gain” condition: |β| times the network’s largest eigenvalue (its spectral radius) is less than 1. Equilibrium actions are proportional to Bonacich centrality, which weights connections by how many paths (walks) lead from a node to the rest of the network, discounting longer paths by β. Intuition: more central players act more in complements games and less in substitutes games.

Diffusion. Two common models:

• Threshold models: Each node adopts when the share or number of adopted neighbors exceeds a threshold (heterogeneous across nodes). Strategic complements generate cascades once a critical mass is reached.
• Epidemic models (SIS/SIR): Adoption or infection transmits probabilistically along edges; a diffusion “takes off” if the effective transmission rate exceeds an epidemic threshold inversely related to the network’s largest eigenvalue. Hubs and dense clusters lower the threshold.

These mechanics generate clear intervention levers: change β (strength of peer effects) via incentives, change seeds (who you start with), or change the graph (connectivity, recommendations).

Key Assumptions and Conditions

• Fixed network: Links are given during the time horizon (who influences whom is stable).
• Local externalities: Payoffs depend on neighbors’ actions, not on distant nodes directly.
• Monotone best responses (for complements): Higher neighbor actions raise best responses, enabling existence and comparative statics via supermodularity.
• Information and updating: Players observe neighbor actions or signals; diffusion models specify thresholds or transmission probabilities.
• Parameter bounds: For linear models, |β|·ρ(G) < 1 ensures a unique stable equilibrium.

Implications

• Targeting central nodes: In complements, subsidizing or seeding highly central users yields outsized spillovers; removing “key players” (for deterrence) has large effects.
• Multiple equilibria and tipping: Strong complements can create low- and high-activity equilibria; small nudges near thresholds trigger cascades (or collapses).
• Policy design: Epidemic thresholds guide vaccination or patching priorities; threshold models inform when to use broad discounts versus influencer programs.
• Competition and pricing: Peer effects shift optimal prices downward initially (to build installed base) and upward later (harvest), especially when early adopters are central.
• Measurement and inference: Observed correlations in neighbor behavior need not be causal peer effects; identification is hard due to homophily (similar people link) and common shocks.

Example in Practice

Enterprise SaaS rollout on an internal collaboration network. A firm plans to deploy a new project-management tool. Employees are nodes in a collaboration graph (email/slack co-mentions). Adoption raises each user’s productivity more when teammates also adopt (strategic complements). The deployment team estimates node centrality and identifies “champions.”

• Seeding: Offer training and temporary fee waivers to the top 5% by Bonacich centrality.
• Local incentives: For the first quarter, teams with at least 60% adoption receive priority IT support (a threshold incentive to overcome coordination barriers).
• Monitoring: Track diffusion; if certain clusters stall, add targeted training (increasing effective β) or connect them with already-adopted teams through cross-team pilots (rewiring edges).

Result: adoption spreads from champions through dense subgraphs; once thresholds are crossed in key teams, a firmwide cascade completes without blanket subsidies.

Limitations and Common Misunderstandings

• Endogenous networks: Links often change in response to adoption (people collaborate more after adopting), blurring cause and effect. Static models may mispredict.
• Identification challenges: Peer effects are hard to separate from shared preferences (homophily) and common shocks (reflection problem). Randomized seeding or instrumental variables are needed for credible inference.
• Model misspecification: Assuming linear responses or separable CTR-like influence may overlook saturation, congestion, or multi-dimensional interactions.
• Equity and fairness: Targeting hubs can bypass peripheral groups, entrenching disparities; designs may include fairness constraints.
• Privacy and governance: Building accurate graphs (emails, location, purchases) raises privacy issues; consent and minimization are prerequisites.
• Overreliance on influencers: In substitutes or competitive contexts (e.g., limited capacity), targeting hubs can crowd out others; optimal targeting depends on the game’s sign (β) and constraints.

Related Concepts (names only)

Bonacich Centrality; Strategic Complementarities/Substitutes; Threshold Models (Granovetter); Epidemic Threshold and Spectral Radius; Influence Maximization; Peer Effects Identification.