Imagine bidding in an auction without knowing how much your competitors value the item, or a firm deciding on pricing without being certain of its rival’s costs. These scenarios involve strategic decision-making under incomplete information, where players don’t have complete knowledge about key aspects of the game. Bayesian games provide a powerful framework for analyzing such situations, allowing economists and strategists to model real-world interactions where uncertainty is the norm rather than the exception.
Table of Contents
What is a Bayesian game?
A Bayesian game represents a strategic situation where players possess private information that affects their payoffs, and this information isn’t commonly known to all participants. Think of it as a regular game with a twist: not everyone knows exactly what game they’re playing. The name honors Hungarian economist John C. Harsanyi, who developed this groundbreaking concept in papers published between 1967 and 1968, work that eventually earned him the Nobel Prize in Economics in 1994.
What makes Bayesian games special is how they handle uncertainty. Rather than requiring infinite loops of “I think that you think that I think,” Bayesian games assign probability distributions to different possibilities, allowing players to calculate outcomes in a single step. This elegant approach transforms seemingly intractable problems into solvable mathematical models.
The formal structure of a Bayesian game
Every Bayesian game consists of five essential components that work together to capture strategic uncertainty. First, there’s the set of players, just like in any game. Second, each player has an action set containing all possible moves they can make. The third element introduces the uncertainty: type sets for each player. A type captures all the private information a player possesses, such as their cost structure, preferences, or capabilities.
The fourth component is the payoff function, which determines what each player receives based on everyone’s actions and types. Finally, there’s a common prior probability distribution over all possible type combinations. This common prior represents the shared understanding among players about how likely different scenarios are, even though they don’t know which specific scenario has actually occurred. For example, in an auction, all bidders might agree that competitor valuations are uniformly distributed between certain values, even though each bidder only knows their own valuation.
Consider a simple example: a sheriff confronting an armed suspect. The suspect knows whether they’re a criminal or civilian, but the sheriff doesn’t. Both players know there’s a probability p that the suspect is criminal and probability 1-p that they’re civilian. The suspect’s type (criminal or civilian) determines their payoffs from shooting or not shooting, creating different strategic situations that the sheriff must account for when making their decision.
Bayesian Nash equilibrium explained
The solution concept for Bayesian games extends the familiar Nash equilibrium to handle incomplete information. In a Bayesian Nash equilibrium (BNE), each player’s strategy must be optimal given their beliefs about opponents’ types and the strategies those types will employ. The key difference from standard Nash equilibrium is that optimization happens at the type level rather than the player level.
More precisely, a strategy profile constitutes a BNE when each type of each player maximizes their expected payoff. This expectation accounts for uncertainty about opponents’ types, weighted by the probabilities from the common prior. For instance, if you’re bidding in an auction and believe your competitor’s valuation is uniformly distributed between zero and one hundred, your optimal bid depends on this entire distribution, not just a single guess about their value.
Why BNE matters for strategic thinking
Bayesian Nash equilibrium generalizes the Nash equilibrium concept in a natural way. In traditional games with complete information, players know everything about the game and simply need to find mutually best responses. With incomplete information, players must form beliefs using Bayes’ rule and optimize based on those beliefs. This makes BNE particularly relevant for real-world applications where decision-makers rarely have perfect information about their competitors’ characteristics or intentions.
The power of BNE lies in how it transforms incomplete information problems. Rather than modeling endless chains of beliefs about beliefs, Harsanyi’s insight allows us to treat these games as complete but imperfect information games where nature moves first, selecting each player’s type according to the common prior. Players then observe their own type but not others’ types, and proceed to choose actions. This elegant reformulation makes solving these games tractable.
Computing pure strategy BNE
Finding a pure strategy Bayesian Nash equilibrium involves determining what action each type of each player should take. The process resembles finding Nash equilibria in complete information games, but with the added complexity of conditioning on types. Let’s walk through a concrete example to see how this works.
Imagine two faculty members deciding whether to call a prospective student. Player 1 knows their calling cost is low (say, c₁ less than one-half), while Player 2’s cost is either low (c with probability p) or high (c̄ with probability 1-p), where p is less than one-half. If both call, the student accepts and both get payoff 1 minus their cost. If only one calls, that person pays the cost but both benefit. If neither calls, the student declines and both get zero.
To find the BNE, we work backwards through types. For Player 2 with high cost c̄ greater than one, calling is strictly dominated regardless of what Player 1 does, because the cost exceeds any possible benefit. Therefore, this type must choose “Don’t Call.” Knowing this, Player 1 compares calling (payoff 1 – c₁) versus not calling (which yields payoff p times 1, plus 1-p times 0, equals p). Since 1 – c₁ exceeds p by assumption, Player 1 calls. Finally, Player 2 with low cost c compares calling (payoff 1 – c) versus not calling (payoff 1, since Player 1 calls). Therefore, Player 2’s low type also doesn’t call. The unique BNE is: Player 1 calls, and Player 2 never calls regardless of type.
Using best response analysis
A systematic approach to finding pure strategy BNE involves checking best responses for each type. Start by identifying any dominated strategies for particular types. These can be eliminated, simplifying the analysis. Then, for remaining strategies, calculate expected payoffs for each type conditional on their beliefs about opponents’ types.
Often, BNE strategies exhibit threshold properties. For example, in the faculty calling game with costs uniformly distributed on an interval, the equilibrium takes the form “call if cost is below c*, don’t call if above c*” for some cutoff c*. To find this cutoff, we use the indifference condition: at exactly c*, the player must be indifferent between calling and not calling. This gives us an equation that pins down c*. In the uniform case with costs between zero and two, the symmetric BNE has both players calling when their cost is below two-thirds.
Computing mixed strategy BNE
Sometimes pure strategy equilibria don’t exist in Bayesian games, just as in complete information games. In these cases, we look for mixed strategy equilibria where types randomize over their available actions. The classic example comes from auction theory, where players bid probabilistically based on their private valuations.
Consider a first-price sealed-bid auction where two bidders have valuations uniformly distributed on the interval from zero to one. Each bidder knows their own valuation but only knows the distribution of their opponent’s valuation. In this setting, the unique symmetric BNE has each type bidding exactly half their valuation. A bidder with valuation v bids v/2, randomizing in the sense that different types (different valuations) bid different amounts.
Solving for mixed strategy equilibria
To derive mixed strategy BNE, we typically assume strategies take a particular functional form and then verify that each type plays a best response. In the auction example, assume opponent strategies are strictly increasing functions mapping types to bids. Then, the probability that opponent j’s bid falls below some amount b equals the probability that their type falls below the inverse of their bidding function evaluated at b.
The first-order condition for optimization gives us a differential equation that characterizes the equilibrium bidding function. For bidder one facing bidder two’s strategy f₂, maximizing expected payoff (v₁ – b₁) times the probability of winning yields the condition that relates v₁ to the optimal bid b₁. In symmetric equilibrium where both players use the same strategy, this simplifies considerably. Solving the resulting differential equation produces the equilibrium strategy: bid half your valuation.
This approach extends to more complex settings. The key steps remain constant: conjecture a form for equilibrium strategies, write down each type’s optimization problem given beliefs about opponents’ strategies, derive first-order conditions, and solve the resulting equations. Verification then confirms that the proposed strategies indeed constitute a mutual best response.
What do you think? Can you identify situations in your own experience where you’ve made decisions without knowing crucial information about others’ preferences or characteristics? How might understanding Bayesian games change your approach to such situations?
Leave a Reply