In the world of game theory, we often assume that when players randomize their strategies, they use a "fair coin"—a device where everyone agrees on the odds. In this seminal 1974 paper, Robert J. Aumann (who later won the Nobel Prize in Economics) asks a radical question: What happens if players disagree on the probabilities of the random events they use to pick their moves? By introducing subjectivity and correlation into randomized strategies, Aumann demonstrates that the fundamental nature of competition changes, allowing players to achieve payoffs that were previously thought impossible.
TL;DR
Aumann shows that when players use "subjective" random devices (like horse races or elections) where they disagree on the odds, even 2-person zero-sum games lose their strictly competitive nature. Furthermore, by allowing strategies to be "correlated" through shared or partial information, players can reach new, self-enforcing equilibria that offer higher payoffs than any standard Nash equilibrium.
What Problem Does This Paper Address?
Traditional game theory relies on "objective" mixed strategies. If a player decides to bluff 30% of the time, it is assumed that every player in the game agrees that the probability is exactly 0.3. Aumann identifies a gap: in the real world, people often have different subjective probabilities for the same event (e.g., "I think there's a 70% chance of rain, but you think it's 40%").
The paper explores how this "agreement to disagree" affects the equilibrium of a game. It also tackles the concept of correlation. While cooperative game theory often uses correlated strategies, Aumann investigates their role in non-cooperative games, where players cannot make binding contracts but might still benefit from coordinating their moves based on external signals.
Methods
Aumann develops a formal mathematical model called a randomizing structure. This framework includes:
- States of the World ($\Omega$): A set of possible outcomes for random events.
- Information Structures ($\mathcal{F}_i$): What each player $i$ knows about the state of the world.
- Subjective Probabilities ($p_i$): Each player’s personal probability measure over the states.
Unlike standard models where the outcome function depends only on the players' choices, here the players' choices are functions of the state of the world. Aumann uses this to define a new kind of equilibrium—what we now call Correlated Equilibrium—where no player has an incentive to deviate from their strategy, given the information they receive from a random device.
Datasets
As a theoretical paper in mathematical economics, no empirical datasets were used. Instead, Aumann employs several ingenious "toy games" (like Matching Pennies or 3-person coordination games) to prove his points through logical deduction and matrix algebra.
Key Insights and Ideas
- Zero-Sum Games Aren't Always Zero-Sum: If two players disagree on the probability of an event $D$, they can bet on it through their game strategies. In a "Matching Pennies" game, if Player 1 thinks $D$ is likely and Player 2 thinks it’s unlikely, they can agree to a strategy where both players' subjective expected payoffs are higher than the game's objective value.
-
The Power of Secrecy: In $n$-person games ($n \ge 3$), new equilibrium points appear if some players have information that others do not.
A 3-person game where subjective probabilities lead to a (3,3,3) payoff, far exceeding the standard (1,1,1) Nash equilibrium. In the example above (Figure 1), if Player 3 knows whether an event $D$ occurred but Players 1 and 2 do not, they can coordinate to reach a payoff of (3,3,3), which is impossible in a standard objective Nash equilibrium.
-
Correlation Beyond the Convex Hull: It was previously thought that correlation could only achieve payoffs that were weighted averages (the convex hull) of existing Nash equilibria. Aumann proves that "partial correlation"—where players receive different, but related, signals—can reach points entirely outside that hull.
Key Results
- Proposition 5.1: In 2-person games, subjectivity alone (without correlation) does not change the set of equilibrium payoffs, provided the players agree on what is "impossible" (absolute continuity of probabilities).
- Correlated Equilibrium: Aumann demonstrates that by using a "public roulette" or a shared signal, players can achieve any payoff in the convex hull of Nash equilibria. More importantly, he shows that private, correlated signals can lead to payoffs that dominate all Nash equilibria.
- Zero-Sum Cooperation: In 2-person zero-sum games, if players have any area of subjective disagreement, they can enter into "mutually advantageous" agreements that yield payoffs $(H_1, H_2)$ where $H_1 > v$ and $H_2 > -v$ (where $v$ is the value of the game).
As shown in Figure 2, the standard Nash equilibria for this game yield payoffs like (2,7) or (7,2). Aumann’s correlated strategy allows players to reach (5,5), a more "fair" and stable outcome that isn't a simple average of the others.
Limitations and Future Work
- The Harsanyi Doctrine: Aumann acknowledges the "Harsanyi doctrine," which argues that differences in subjective probability must stem from differences in information. Aumann defends his "inconsistent case" (where priors differ) as a useful "small world" approximation for complex real-world scenarios.
- A Posteriori Equilibria: The paper notes a technical difficulty: a player might agree to a strategy beforehand but want to "renege" once they see the actual outcome of the random event. Aumann briefly introduces "a posteriori equilibria" to address this, but suggests a more complete theory is needed.
- Extensive Form: The paper focuses on strategic (normal) form games. Aumann suggests that extending these ideas to extensive form games (games with a sequence of moves) is a promising area for future research.
Summary updated