Bayesian Games

Definition and basic components

A Bayesian game (a game of incomplete information) extends the normal-form game by allowing players to hold private information about their own characteristics, preferences, or available actions. Formally, a Bayesian game consists of:

• •A finite set of players $I=\{1,\dots,n\}$.
• •For each player $i$, a finite set of actions $A_i$ (pure strategies). Example: $A_1=\{C,D\}$ where $C$ stands for "Cooperate" and $D$ for "Defect".
• •For each player $i$, a finite set of types $T_i$ representing private information. Example: $T_1=\{H,L\}$ meaning high or low valuation.
• •A common prior $p(t_1,\dots,t_n)$ over type profiles in $T\equiv\prod_i T_i$. This captures ex ante beliefs before types are realized. Example: for two players each with types $\{H,L\}$, one might have $p(H,H)=0.25,p(H,L)=0.25,p(L,H)=0.25,p(L,L)=0.25$ (uniform).
• •For each player $i$, a payoff function $u_i(a_1,\dots,a_n,t_i)$ which may depend on the action profile and player $i$'s own type; sometimes the payoff depends on the entire type profile, $u_i(a,t)$.

The crucial element is that players know the structure above (including the prior), know their own type $t_i\in T_i$, but do not observe other players' types. They form beliefs about others' types via conditional probabilities derived from the common prior using Bayes' rule when possible.

Why this matters: In many economic settings (auctions, contracting, signaling) players have private information that materially affects optimal actions. If you learned Nash equilibrium in "Nash Equilibrium" (a prerequisite), you know how to analyze strategic stability when payoffs are common knowledge. Bayesian games let you extend that analysis to private information by replacing actual payoffs with expected payoffs given beliefs.

Intuition: Replace each player by a collection of type-specific decision-makers. Each type chooses an action possibly contingent on the type. Players choose strategy functions $s_i:T_i\to\Delta(A_i)$ (possibly mixed) mapping types to distributions over actions. A strategy profile $s=(s_1,\dots,s_n)$ is a Bayesian Nash equilibrium if for every player $i$ and every type $t_i$, $s_i(t_i)$ maximizes expected payoff conditional on $t_i$ given other players' strategies and the prior.

Mathematical expression: Given strategies $s$ the expected payoff to player $i$ with type $t_i$ when she plays an action $a_i$ and others follow $s_{-i}$ is

Concrete numeric example: Two players 1 and 2; $T_1=\{H,L\}$, $T_2=\{0\}$ (player 2 has no private info). Actions: $A_i=\{0,1\}$. Prior: $p(T_1=H)=0.6, p(T_1=L)=0.4$. Suppose $u_1(a_1,a_2,t_1)=\begin{cases}2 &\text{if }a_1=1,a_2=1,t_1=H\\0 &\text{otherwise}\end{cases}$ and $u_2$ irrelevant. If player 1 of type $H$ contemplates $a_1=1$ while believing $s_2(1)=0.5$, then

This shows how conditional beliefs and strategy mappings determine expected payoff.

Relation to incomplete vs imperfect information

A Bayesian game models incomplete information (players have private types). Using the Harsanyi transformation (covered later), one can represent an incomplete-information game as an extensive-form game with imperfect information. This is central: "incomplete information" refers to missing payoff-relevant facts; "imperfect information" refers to players not observing previous moves. Harsanyi’s trick converts uncertainty about types into moves by Nature at the game’s start, turning incomplete into imperfect information while preserving solution concepts.

Comparison to prerequisites

• •In "Nash Equilibrium" we learned that equilibrium requires no profitable unilateral deviation given exact payoffs. In Bayesian games, exact payoffs aren't known to all players; we replace the payoff by the expectation conditional on each type and require no profitable deviation for each type separately.
• •In "Bayesian Inference" we learned to update priors into posteriors using Bayes’ rule. In Bayesian games, conditional beliefs $p(t_{-i}|t_i)$ are computed via Bayes’ rule from the common prior when types are independent or correlated; strategic behavior may cause off-path beliefs that require explicit specification in dynamic/ signaling contexts.

Summary: A Bayesian game formalizes strategic interaction under private information by turning each player's private information into "types" and requiring strategy functions that are best responses in expectation given beliefs derived from a common prior.

Definition and derivation

A Bayesian Nash equilibrium is the central solution concept for static Bayesian games. It is a profile of type-contingent strategies $s^*_i: T_i\to\Delta(A_i)$ such that for every player $i$ and every type $t_i\in T_i$, $s^*_i(t_i)$ maximizes expected payoff given other players’ strategies and the conditional distribution over other types:

where

Concretely, if strategies are pure deterministic, $s_i(t_i)\in A_i$, the condition reduces to: for all $a'_i\in A_i$,

Numerical toy calculation

Take two players, each with types $\{H,L\}$. Actions $A_i=\{0,1\}$. Prior independent: $p_1(H)=0.6,p_2(H)=0.7$. Payoffs for player 1 depend on his type only: if type $H$ then $u_1(a_1,a_2,H)=3$ if $a_1=a_2=1$ else $0$; if type $L$ then $u_1(a_1,a_2,L)=1$ if $a_1=1$ else $0$. Suppose player 2 always plays $a_2=1$. Evaluate whether player 1 of type $H$ should play $a_1=1$:

Conditional distribution $p(t_2|t_1)$ equals $p_2$ marginal since independence, so $p(t_2=H|t_1)=0.7$. The expected payoff from $a_1=1$ for type $H$ is

If instead $a_1=0$, payoff is $0$. So playing $1$ is optimal. Numeric check: if $p_2(H)$ were $0.2$, then $U_1(1|H)=0.2\cdot3=0.6<0$, so $1$ would not be optimal.

Existence, independence, and correlated types

If types are independent and action spaces are finite, a BNE in mixed strategies always exists by a fixed-point argument similar to Nash existence (Glicksberg). If types are correlated, the same existence results hold for finite games under standard mixed-strategy compactness and continuity conditions.

Derivation for two-player case

Consider two players 1 and 2 with finite type spaces $T_1,T_2$ and actions $A_1,A_2$. A strategy profile $s=(s_1,s_2)$ is a BNE if for all $t_1\in T_1$,

For finite supports, we can compute expected payoff vectors for each type and use linear programming to find best responses. Example: With two actions per player and two types per player, each strategy is a vector of four probabilities. Solve the best response equations using expected payoff formulas. A numeric illustration helps:

Example numeric derivation: Let $T_1=T_2=\{H,L\}$ with $p(H,H)=0.3,p(H,L)=0.3,p(L,H)=0.2,p(L,L)=0.2$. Suppose $A_i=\{0,1\}$. Let payoffs for player 1 be $u_1(a_1,a_2,t)=a_1(2a_2+\mathbf{1}_{t=H})$ (i.e. payoff increases with matching and with being high type). If player 2 uses pure strategy $s_2(H)=1,s_2(L)=0$ (play 1 when H, 0 when L), then for player 1 of type $H$ the conditional probability of $t_2=H$ is

Expected payoff from $a_1=1$ is

From $a_1=0$ payoff is $0$, so $a_1=1$ is optimal for type $H$.

Interpretation: BNE demands this optimality for every type simultaneously. In mixed strategies, solving requires equalizing expected payoffs across pure actions used with positive probability. The computations are conceptually identical to those in "Nash Equilibrium" but carried out conditionally for each type.

Belief updating and off-equilibrium path issues

In static Bayesian games with exogenous types and full-support priors, conditional probabilities $p(t_{-i}|t_i)$ are well-defined. In dynamic or signaling games (next section), Bayes' rule is applied to observed actions to update beliefs; if an observed action has zero probability under equilibrium, Bayes' rule does not pin down beliefs and one must specify a solution refinement (e.g., Perfect Bayesian Equilibrium).

Harsanyi transformation: converting incomplete into imperfect information

Harsanyi’s insight (1967–68) is that a game of incomplete information can be modeled as an extensive-form game where Nature moves first and selects a type profile according to the common prior. Each player observes her own type and then plays the rest of the game. Information sets capture that players do not observe others' types. This transformation is mechanical and preserves BNE as the counterpart of Nash equilibrium in the resulting imperfect-information extensive form.

Formal construction

Start with a Bayesian game (I, A_i, T_i, p, u_i). Create an extensive form:

1. 1)Nature (chance node) selects $t=(t_1,\dots,t_n)$ with probability $p(t)$. Example: for two players, Nature picks $(H,L)$ with probability 0.25.
2. 2)Player $i$ is informed of her own type $t_i$ but not others. So for each $t_i$ there is a decision node for player $i$; nodes for different $t_i$ are in different information sets (player distinguishes them), but for a fixed $t_i$, nodes corresponding to different $t_{-i}$ are in the same information set (player cannot distinguish other players' types).
3. 3)Players choose actions and payoffs are $u_i(a,t)$.

The extensive-form game's Nash equilibria correspond to BNEs of the Bayesian game when strategies are type-contingent actions. Thus, solving BNE reduces to finding Nash equilibria in an imperfect-information extensive form where Nature's move implements the prior.

Concrete numeric Harsanyi example

Two players, $T_1=\{H,L\}$, $T_2=\{\alpha,\beta\}$, prior: $p(H,\alpha)=0.2, p(H,\beta)=0.3, p(L,\alpha)=0.1, p(L,\beta)=0.4$. Nature draws a profile accordingly. Player 1 observes $H$ vs $L$; player 2 observes $\alpha$ vs $\beta$. If player 1 faces $H$, she has decision nodes for $(H,\alpha)$ and $(H,\beta)$ but these are in the same information set for her since she doesn't see player 2’s type. The expected payoff calculations in the extensive-form game for a strategy mapping $s_i:T_i\to A_i$ are identical to the BNE formulas earlier.

Signaling games: an important class

A signaling game is a two-stage Harsanyi-style extensive-form game where one informed player (the Sender) moves after Nature selects her type and then an uninformed player (the Receiver) observes Sender's action but not her type, updates beliefs via Bayes' rule, and chooses a response. This models persuasion, job-market signaling, and cheap-talk.

Simple signaling numeric example (job market):

Types: Worker can be High ($H$) or Low ($L$) ability with prior $p(H)=0.4,p(L)=0.6$. Sender (worker) chooses education level $e\in\{0,1\}$. Education is costless to $H$ (cost 0) and costly to $L$ (cost 1). Employer observes $e$, then sets wage $w\in\{0,2\}$ equal to expected productivity. The worker's payoff is $w-\text{cost}(e,t)$. If employer uses Bayesian updating and sets $w=2$ if posterior $P(H|e)\ge 0.5$ and $w=0$ otherwise, then calculate whether pooling or separating equilibria exist.

Compute posterior if both types choose $e=1$ (pooling):

Employer sets $w=0$, so high type gets $0-0=0$, low type gets $0-1=-1$. But if high deviates to $e=0$, employer still pays 0, so no profitable deviation. If high takes $e=1$ while employer pays nothing, maybe high can't gain.

Separating candidate: let $H$ choose $e=1$, $L$ choose $e=0$. Then after $e=1$, posterior $P(H|e=1)=1$ so $w=2$, and after $e=0$, $P(H|e=0)=0$ so $w=0$. Worker payoffs: $H$ gets $2-0=2$, $L$ gets $0-0=0$. Low won't mimic because if $L$ deviates to $e=1$, her payoff would be $2-1=1<0$? Wait numeric check: $0$ is better than $1$? Actually $1>0$, so low would want to mimic. So no separating equilibrium. But if cost for $L$ were $c=2$, then deviating yields $2-2=0$ which equals staying; further refinements needed. This numeric manipulation shows how costs and priors determine signals' credibility.

Off-equilibrium beliefs and refinement

Harsanyi transformation makes explicit when Bayes’ rule applies: only on-path signals have beliefs pinned down. In signaling games one often needs refinements (e.g., Perfect Bayesian Equilibrium, Sequential Equilibrium, or the Intuitive Criterion) to rule out equilibria supported by unreasonable beliefs about off-path actions.

Proof sketch that BNE corresponds to Nash equilibrium in Harsanyi game

Given a Bayesian game, take a strategy profile $s$ mapping $T_i$ to mixed actions. Treat $s$ as a strategy in the Harsanyi extensive form where each type's node prescribes the corresponding action distribution. Because Nature’s probabilities are the prior, each type-maximization condition in BNE is exactly the best-response condition in the extensive form given the same beliefs. Hence existence and characterization properties carry over.

Why Bayesian games matter: fields and canonical problems

Bayesian games are foundational in auction theory, mechanism design, contract theory, signaling in labor markets, political economy (e.g., elections with private valuations), cybersecurity (hidden types or states), and negotiation under asymmetric information. Key canonical models built on Bayesian games include:

• •Private-value auctions (first-price, second-price): bidders know their valuations (types) drawn from a common prior; BNE yields bidding functions. Example numeric: valuations iid Uniform[0,1]; in a second-price sealed-bid auction, truthful bidding is a dominant strategy; in a first-price auction the symmetric BNE bidding function is $b(v)=\frac{n-1}{n}v$. For $n=3$ and $v=0.9$ we have $b(0.9)=\frac{2}{3}\cdot0.9=0.6$.

• •Adverse selection: sellers and buyers have private information (e.g., used car market). Market unraveling and the Akerlof lemons model are naturally expressed with types and priors. Numeric illustration: Suppose high-quality cars fraction $0.7$, valuations for buyers depend on type; a pooling price will be set based on expected quality.

• •Mechanism design: Bayesian Incentive Compatibility (BIC) conditions generalize dominant-strategy incentive constraints by requiring truth-telling to maximize expected utility given beliefs over other types. For instance, in auctions revenue-maximization uses BIC constraints: Myerson’s optimal auction solves for virtual values and requires BIC. Numeric illustration: For a single bidder with value distribution $F(v)=v$ on $[0,1]$, virtual value is $\phi(v)=v-\frac{1-F(v)}{f(v)}=v-\frac{1-v}{1}=2v-1$. A reserve price solves where $\phi(v)=0$, i.e., $v=0.5$.

Advanced connections: equilibrium selection and refinements

In dynamic Bayesian games with signaling, Perfect Bayesian Equilibrium (PBE) refines BNE by specifying beliefs at every information set and requiring sequential rationality. Sequential Equilibrium further requires consistency of beliefs with strategies via Bayes' rule and limits of perturbed strategies. These refinements are essential when off-path beliefs can sustain multiple spurious equilibria.

Computational aspects

Computing BNE in games with continuous type spaces often reduces to solving functional fixed-point equations (for symmetric equilibria). For example, first-price auction with iid Uniform[0,1] leads to solving an ODE for the symmetric bidding function \(b(v)\). Derivation: Expected payoff for bidder with value $v$ bidding $b$ and other bidders using $b(\cdot)$ is

For Uniform[0,1] and symmetric increasing $b$, first-order condition yields $b(v)=v-\int_0^v F(t)^{n-1}dt/F(v)^{n-1}$ which simplifies to $b(v)=\frac{n-1}{n}v$. Numeric: with $n=2$ bidders and $v=0.8$, $b=0.4$.

Empirical and experimental applications

Structural estimation: Many empirical studies estimate players’ type distributions and strategy functions from observed behavior by inverting equilibrium conditions (e.g., structural auction estimation). Lab experiments test predictions on signaling or entry under incomplete information.

Practical modeling tips

• •Decide whether types are independent or correlated; correlation can dramatically change equilibrium (common-value vs private-value auctions).
• •Choose common prior carefully; many results depend sensitively on the prior's support.
• •Check on-path vs off-path beliefs: specify how Bayes’ rule applies and which equilibrium refinements are plausible.

Summary: Bayesian games turn private information into a formal object (types) and let you analyze strategic behavior under uncertainty by maximizing expected payoffs conditional on types. The Harsanyi transformation embeds such games into imperfect-information extensive forms, making dynamic analysis and signaling straightforward. This framework underlies auctions, mechanism design, signaling, and inference in strategic settings.