Informational properties of the Nash solutions of two stochastic nonzero-sum games

Tamer Başar, Yu‐Chi HoJournal of Economic Theory 1974 · DOI

TL;DR

This paper investigates how the amount of information available to players in two types of stochastic games (LQG and duopoly) affects their equilibrium strategies and outcomes. It reveals that in LQG games, more information benefits both players by reducing their costs, while in duopoly games, it benefits the informed player at the expense of the less informed one, highlighting a "zero-sum" property within a non-zero-sum context.

What Problem Does This Paper Address?

In game theory, understanding how information asymmetry influences strategic decisions and outcomes is crucial. This paper delves into stochastic, two-person, non-zero-sum games, where players make decisions under uncertainty about the "state of nature" and receive noisy measurements about it. The core problem is to analyze how changes in the quality of these measurements (i.e., the amount of information each player has) impact the equilibrium strategies and, consequently, the payoffs (costs or profits) for both players. The authors specifically examine two well-established models: the Linear-Quadratic-Gaussian (LQG) model, often used in engineering, and the duopoly model, common in economics. They aim to understand if better information always leads to better outcomes for everyone, or if it can create winners and losers.

Methods

The authors employ a formal game-theoretic approach to analyze the Nash equilibrium solutions for the chosen stochastic games. They define "better information" for player i as a reduction in the variance of the noise corrupting their measurements, which leads to a more accurate estimate of the state of nature.

Key methodological aspects include:

  • Nash Equilibrium Analysis: They derive the conditions for a Nash equilibrium, where each player chooses their best strategy given the other player's strategy, aiming to minimize their expected cost (or maximize profit).
  • Information Structure Definition: Information quality is quantified by the variance of the additive Gaussian noise in the measurement model ($z_i = h_i(x, w_i)$). Lower noise variance implies better information.
  • General Framework: They first establish a general framework for stochastic games with quadratic cost functions and linear observation models.
  • Model Specialization: The LQG and duopoly models are then presented as specific instances of this general framework by identifying corresponding parameters.
  • Comparative Statics: The core of the analysis involves examining how the Nash equilibrium strategies and expected payoffs change as the noise variances ($s_{iL}$ or $s_{iD}$) are varied. This is done by taking partial derivatives of the players' expected costs/profits with respect to these noise variances.
  • Entropy Analysis: To formally justify the definition of "better information," they use Shannon's information flow, demonstrating that reduced noise variance corresponds to increased information flow.

Datasets

Not discussed in the paper. The paper focuses on theoretical models and analytical derivations rather than empirical evaluation on datasets.

Key Insights and Ideas

  • Divergent Information Effects: The most striking insight is the contrasting impact of improved information in the two models. In LQG, better information is a win-win, reducing costs for both players. In the duopoly, it's a zero-sum dynamic: one player's gain in profit comes directly from the other's loss.
  • "Odd Couple" Effect: For the LQG model, when information improves for one player, it disproportionately hurts the other player more than it helps the informed player, a phenomenon the authors playfully dub the "odd couple" effect. This suggests that even in cooperative-leaning scenarios, information asymmetry can lead to complex, non-intuitive trade-offs.
  • Zero-Sum Property in Duopoly: The duopoly model exhibits a "zero-sum" property in its response to information changes, even though the players' objectives are not strictly opposed. This means that the gains of one player are directly mirrored by the losses of the other when information levels shift.
  • Team Property in LQG: Conversely, the LQG model displays a "team" property, where improvements in information for one player lead to improvements for both, mirroring the behavior of a cooperative game where players share a common objective.
  • Generalizability: The paper proposes a general framework for analyzing such games, suggesting that the derived properties might extend to broader classes of quadratic cost functions and multidimensional models, despite the complexity of obtaining explicit solutions.

Key Results

  • LQG Model: For the LQG model, an increase in information for either player (i.e., a decrease in $s_{iL}$) leads to a decrease in the average Nash cost for both players. This is mathematically expressed as $\frac{\partial J_{iL}}{\partial s_{iL}} > 0$ for $i=1,2$ and $j=1,2$.
  • Duopoly Model: For the duopoly model, an increase in information for player i (i.e., a decrease in $s_{iD}$) leads to a decrease in their average Nash profit (or cost, if viewed as negative profit), but an increase in the average Nash profit for player j ($j \neq i$). This is expressed as $\frac{\partial J_{iD}}{\partial s_{iD}} < 0$ for $i=j$ and $\frac{\partial J_{jD}}{\partial s_{jD}} > 0$ for $i \neq j$.
  • Existence and Uniqueness: The paper proves the existence and uniqueness of affine Nash strategies for the general stochastic game framework, which then applies to the specific LQG and duopoly models.

Limitations and Future Work

  • Model Specificity: The authors acknowledge that the specific results are derived for LQG and duopoly models with quadratic cost functions and linear observation models. They are aware that these models are quite specific and that drawing broad conclusions for more general payoff functions and information structures requires further extensive analysis.
  • Computational Complexity: While the general method is applicable to more complex scenarios (multidimensional, general quadratic cases), obtaining explicit, concrete results for these general models is computationally intensive and may exceed current symbolic manipulation capabilities.
  • Future Directions: The paper suggests two main avenues for future research:
  • General Information Formulation: Developing a more general way to model information structures, perhaps by changing the information function itself rather than just the distribution of random variables.
  • Quality of Information: Investigating how the quality of information, beyond just its quantity, relates to payoff structures, which could lead to more counterintuitive results (like mutual ignorance leading to mutual bliss, as hinted in a cited work). They suggest that such counterintuitive outcomes might be more common than exceptions in specific models.

Summary updated

BibTeX
@article{basar1974informational,
  author = {Tamer Başar and Yu‐Chi Ho},
  title = {{Informational properties of the Nash solutions of two stochastic nonzero-sum games}},
  journal = {Journal of Economic Theory},
  year = {1974},
  doi = {10.1016/0022-0531(74)90110-0},
}
Figure