Must players know each other's private values?

No. Their uncertainty about those values is central to the framework.

Can an equilibrium involve randomization?

Yes; strategies may specify probability distributions over actions for each type.

Does each outcome maximize total welfare?

No. Individual expected-payoff best responses need not maximize collective welfare.

AI Glossary / Statistics and probability / Advanced

Bayes-Nash equilibrium

Definition

A Bayes-Nash equilibrium, also called a Bayesian Nash equilibrium, is a profile of strategies in a game with incomplete information such that each player's strategy maximizes expected payoff given that player's private information, beliefs about others' types, and the other players' strategies. No player type can gain by changing its own action while the others' strategies remain fixed.

What it means

Players often have private information: bidders may know their own valuations, while suppliers know their own costs.

A Bayesian game represents private information using types. A strategy specifies what a player would do for each possible type, rather than naming only one observed action. Beliefs describe how likely the other types are, conditional on what the player knows.

Equilibrium requires mutually consistent best responses. It does not require players to know others' private information, achieve a socially best outcome, or receive the same payoff. It is a property of a specified model, not a guarantee about actual behavior.

How it works

Specify the players, available actions, possible types, payoffs, and beliefs. A common formulation supplies a joint prior distribution of types, allowing players to condition their beliefs on their own types. Correlated types must be handled with conditional distributions rather than assumed independent.

Propose a strategy for each player. For each type with relevant probability, calculate the expected payoff from its proposed action using the other players' strategies and the conditional distribution of their types.

Compare that payoff with every feasible unilateral alternative. The profile is a Bayes-Nash equilibrium if no such deviation improves expected payoff. Mixed strategies allow a type to randomize over actions; checking a single realized action or a single pair of types is not enough.

The concept extends Nash equilibrium to private information, but it does not by itself impose every sequential-rationality requirement needed in dynamic games. Refinements such as perfect Bayesian or sequential equilibrium address behavior and beliefs at later information sets. Existence, uniqueness, and computational difficulty depend on the game's mathematical structure.

Three practical examples

Auctions: analyze bidding incentives when participants privately know their valuations and have beliefs about competitors.
Market entry: study whether firms enter when costs are privately known and profitability depends on how many others enter.
Automated agents: evaluate strategic policies for agents that have private resources or preferences, within a carefully specified interaction model.

Think of it this way

Think of players each holding a private card that changes the rewards of available moves. A strategy is a complete plan for every card they could hold. Equilibrium means every card-holder prefers its prescribed move, given the other players' plans and beliefs about their cards.

Worked example

Consider two fictional firms choosing simultaneously whether to enter a small market. Staying out gives payoff zero. An entrant receives gross revenue 6 if alone and 3 if both enter, then pays its own cost. Each firm's cost is independently 2 or 5 with probability one-half, and each observes only its own cost.

Consider the strategy: enter with cost 2 and stay out with cost 5. If the other firm follows this strategy, it enters with probability one-half. Expected gross revenue from entry is therefore 0.5 × 3 + 0.5 × 6 = 4.5.

For the low-cost type, entry gives expected payoff 4.5 - 2 = 2.5, exceeding zero from staying out. For the high-cost type, entry gives 4.5 - 5 = -0.5, so staying out is better. The same reasoning applies to both firms.

Both types of both firms are best responding, so this strategy profile is a Bayes-Nash equilibrium. It does not imply that every realized outcome is best after all private information is revealed. For example, a high-cost firm would prefer entry if it somehow knew the other firm would stay out, but that is not its information when choosing.

Illustrative case study

Fictional case: Bracken Exchange develops a simulation of suppliers deciding whether to participate in a procurement event. Suppliers privately observe costs, and participation returns depend on the number of competitors.

Analysts specify cost distributions and test candidate type-dependent strategies by calculating deviation incentives. They discover that replacing conditional beliefs with a single average competitor cost can misrepresent incentives for suppliers whose own information is correlated with market conditions.

The exchange uses the model to compare possible event designs, then checks whether simulated assumptions resemble observed participation. It does not label a numerical outcome an equilibrium merely because repeated simulations appear stable; the best-response conditions still need verification.

Common misconceptions and mistakes

Do not confuse a strategy with one action. The strategy must describe choices across possible private types.

Do not assume Bayesian means players must run an explicit numerical inference algorithm. The model concerns beliefs and expected-payoff maximization, not a required software implementation.

Do not interpret equilibrium as truthful reporting, cooperation, or efficiency. Those properties require additional conditions and may fail.

Do not assume independent types without justification, or evaluate deviations using information the player does not possess. Likewise, convergence of a learning algorithm is not automatic, and multiple equilibria can make a single predicted outcome difficult to defend.

Business use

The framework helps organizations reason about incentives when counterparties possess private information. It can clarify why a rule that appears attractive under average costs produces different behavior across participant types.

Business use requires credible payoff assumptions, information timing, and beliefs. Sensitivity analysis should examine whether conclusions survive changes in these inputs and whether alternative equilibria exist. For AI systems interacting strategically, equilibrium analysis can complement empirical testing, but it does not establish how boundedly rational or learning participants will behave in deployment.

Check your understanding

Why does the high-cost firm stay out in the example?

Answer: Entry has expected payoff -0.5 under its information and the rival's strategy, below the zero payoff from staying out.

Is checking only low-cost types sufficient?

Answer: No. The strategy must also be optimal for the high-cost types and for each player.

Related terms

Bayesian game
Expected utility
Mechanism design
Nash equilibrium
private information
type space
auction
perfect Bayesian equilibrium

Learn next

Write a two-player game with two private types and verify every type's deviation incentives. Then compare Bayes-Nash equilibrium with sequential refinements in a game where later actions reveal information.

Sources

Daniel Quint, University of Wisconsin, Lecture 4: Bayesian Games and No Trade, including type-dependent strategies and Bayesian Nash equilibrium: https://users.ssc.wisc.edu/~dquint/econ698/lecture%204.pdf Terminology note: Bayes-Nash and Bayesian Nash refer to the same equilibrium concept here; the worked entry game is an original illustration.

Previous: Bayes linear statistics Next: Bayes risk