52.04.02 · economics / game-theory

Signaling games and Bayesian Nash equilibrium: Spence, Akerlof, and the intuitive criterion

shipped3 tiersLean: none

Anchor (Master): Akerlof 1970 The Market for Lemons (QJE 84); Spence 1973 Job Market Signaling (QJE 87); Cho-Kreps 1987 Signaling Games and Stable Equilibria (Handbook of Game Theory); Crawford-Sobel 1982 Strategic Information Transmission (Econometrica 50)

Intuition Beginner

Suppose two kinds of worker exist in roughly equal numbers: highly productive workers who generate $$300, 000$ of value a year for an employer, and less productive workers who generate $$150, 000$ . The workers know which kind they are, but employers cannot tell them apart by looking. If employers cannot tell them apart, they pay everyone the average wage, $$225, 000$ — and the high-productivity workers are underpaid. Now let workers choose whether to attend college. College costs the high-productivity worker $$40, 000$ in effort and fees, but costs the less-productive worker $$120, 000$ because the work is harder for them. College does not raise anyone's productivity. Should anyone go?

The answer is surprising. In the equilibrium Michael Spence described in 1973, the high-productivity workers go to college and the less-productive ones do not. Employers pay college graduates $$300, 000$ and non-graduates $$150, 000$ , because they have figured out the pattern. The high-productivity workers spend $$40, 000$ to gain $$75, 000$ in wages. The less-productive workers do not go to college, because $$75, 000$ in extra wages is not worth the $$120, 000$ cost. The signal sorts the types — even though college raised no one's productivity.

Why this matters: a costly action can carry information precisely because it is costly. The signal works because the cost is different for the two types. If college cost the same for both, it could not sort them, because either both would go or neither would. Spence's insight founded the economics of asymmetric information, the field that also includes Akerlof's "lemons" market (where bad-quality used cars drive out good ones when buyers cannot tell them apart) and earned both authors, with Joseph Stiglitz, the 2001 Nobel Prize.

Visual Beginner

The picture shows the signaling game as an extensive-form tree. At the top, Nature draws the worker's type (high or low, with equal probability). The worker observes their own type and chooses College or No College. The employer observes the worker's choice (but not the type) and chooses a wage.

Two outcomes are highlighted: in the left branch (College), the employer pays a high wage, and only the high type chooses this path. In the right branch (No College), the employer pays a low wage, and only the low type chooses this path. A side panel shows the cost lines crossing: the cost of college rises more steeply for the low type than for the high type, the single-crossing property that makes the signal work.

The crossing cost lines are the structural condition for signaling: the signal must be cheaper for the high-ability type than for the low-ability type, so that only the high type finds it worth paying.

Worked example Beginner

Solve the Spence job-market signaling game with the numbers in the intuition, step by step, and verify the separating equilibrium.

Step 1. Set up. Two types: high $H$ (productivity $$300, 000$ ) and low $L$ (productivity $$150, 000$ ), in equal proportions. College cost: $c_{H} = $40, 000$ for high, $c_{L} = $120, 000$ for low. College raises no productivity. Employers observe only the college choice.

Step 2. Propose a separating equilibrium: $H$ chooses College, $L$ chooses No College; employer pays college graduates $$300, 000$ and non-graduates $$150, 000$ .

Step 3. Check that no type wants to deviate. The $H$ worker going to College earns $$300, 000 - $40, 000 = $260, 000$ . If they deviated to No College they would earn $$150, 000$ , which is worse. So $H$ does not deviate. The $L$ worker going to No College earns $$150, 000$ . If they deviated to College they would earn $$300, 000 - $120, 000 = $180, 000$ , which is worse. So $L$ does not deviate.

Step 4. Verify the employer's beliefs and wages. The employer sees College and infers (correctly) that the worker is type $H$ , paying the expected productivity $$300, 000$ . The employer sees No College and infers type $L$ , paying $$150, 000$ . The wages match the inferred productivity, and the inferences match the workers' actual choices. The equilibrium is consistent.

What this tells us: a separating equilibrium exists when the signal's cost differs enough between types. The high type signals; the low type does not; the wage gap exactly covers the high type's cost but not the low type's larger cost. The signal sorts the types without raising productivity — a profound and unsettling result.

Check your understanding Beginner

Exercise (easy, multiple choice).

A separating equilibrium in a signaling game is one in which:

A. All types of sender choose the same action B. Different types of sender choose different actions, so the receiver can infer the sender's type from the action C. The receiver ignores the sender's action D. No one has a strict incentive

Hint

"Separating" means the types separate: each type chooses a distinct action, making the type observable through the action.

Answer

B. In a separating equilibrium, each type of sender chooses a different action, and the receiver's belief (after seeing the action) identifies the type. The receiver then conditions the response on the type. In the Spence game, $H$ chooses College, $L$ chooses No College, and the employer reads the choice as a perfect indicator of type. The opposite is a pooling equilibrium, where all types choose the same action and the receiver learns nothing.

Exercise (easy, multiple choice).

What is the single-crossing property (the Spence-Mirrlees condition) that makes signaling possible?

A. The signal must be free for all types B. The signal must cost less for the high-ability type than for the low-ability type (so the marginal cost of the signal crosses once when plotted against type) C. The signal must be impossible to fake D. The signal must be the same for all types

Hint

The single-crossing condition is what makes the high type willing to send the costly signal while the low type is not. The cost difference is the sorting mechanism.

Answer

B. The single-crossing property requires that the marginal cost of the signal be lower for the high-ability type than for the low-ability type. Geometrically, the cost-of-signal curves for the two types cross at most once, so above the crossing point only the high type is willing to send the signal. The condition is the load-bearing structural assumption of every signaling model and the formal version of the slogan "the signal must be costly, and the cost must differ by type."

Exercise (easy, true-false).

In the Spence job-market signaling model, college raises workers' productivity, and the signal works because graduates are more skilled.

Hint

Spence's disturbing point is the opposite: college need not raise productivity at all for the signal to work. The signal sorts the types even if college is pure waste.

Answer

False. In the Spence model, college raises no one's productivity. The signal works purely because it is costly and the cost differs by type. High-ability workers attend because the wage premium covers their lower cost; low-ability workers do not because the same wage premium does not cover their higher cost. The signal is informational, not productive. This is the unsettling content of the signaling model: a vast amount of expenditure on credentials can be socially wasteful even though it is individually rational.

Exercise (easy, numeric).

In a Spence game, the high type's college cost is $$30, 000$ and the low type's is $$80, 000$ . The productivity gap (the difference between high and low productivity wages) is $$60, 000$ . Is a separating equilibrium sustainable where high types go to college and low types do not? Answer "yes" or "no" with one sentence of justification.

Hint

Check the two no-deviation conditions: the high type must prefer college ( $c_{H} <$ wage gap), and the low type must prefer no college ( $c_{L} >$ wage gap).

Answer

Yes. The high type's college cost ( $$30, 000$ ) is less than the wage gap ( $$60, 000$ ), so the high type will attend. The low type's college cost ( $$80, 000$ ) exceeds the wage gap ( $$60, 000$ ), so the low type will not attend. Both no-deviation conditions hold, so the separating equilibrium is sustainable.

Formal definition Intermediate+

A signaling game is a Bayesian extensive-form game in which an informed player (the sender) chooses an action that conveys information to an uninformed player (the receiver), who then chooses a response. Its core notions are the type, the message, the belief, and the perfect Bayesian equilibrium.

Definition (signaling game). A signaling game is a tuple $(T, M, A, p, u_{S}, u_{R})$ where $T$ is the set of sender types, $M$ is the set of messages the sender can send, $A$ is the set of actions the receiver can take, $p$ is the prior distribution over types, and $u_{S}, u_{R}$ are the sender's and receiver's payoff functions. The order of play: Nature draws $t \in T$ for the sender; the sender observes $t$ and chooses $m \in M$ ; the receiver observes $m$ (not $t$ ) and chooses $a \in A$ ; payoffs are realised.

Definition (separating and pooling equilibria). A pure-strategy perfect Bayesian equilibrium is separating if each type $t$ chooses a distinct message $m (t)$ , so that the receiver's posterior belief after seeing $m (t)$ places probability one on type $t$ . The equilibrium is pooling if all types choose the same message, so the receiver's posterior equals the prior.

Definition (perfect Bayesian equilibrium). A perfect Bayesian equilibrium (PBE) is a strategy profile $(m (\cdot), a (\cdot))$ together with a system of beliefs $μ (\cdot ∣ m)$ such that (i) the sender's strategy maximises expected payoff given the receiver's strategy, for each type; (ii) the receiver's strategy maximises expected payoff given the beliefs; (iii) the beliefs are derived from Bayes' rule wherever possible (on the equilibrium path) and unrestricted off the path.

Counterexamples to common slips Intermediate+

Treating the receiver's belief as exogenous. Beliefs are endogenous in equilibrium: the receiver's posterior is derived by Bayes' rule from the sender's strategy. A strategy-belief pair is an equilibrium only if the beliefs are consistent with the strategy on the equilibrium path.
Conflating Bayesian Nash and perfect Bayesian equilibrium. A Bayesian Nash equilibrium (BNE) is a Nash equilibrium of the one-shot Bayesian game; a perfect Bayesian equilibrium adds sequential rationality at every information set and belief consistency on the equilibrium path. The two coincide for static games but diverge in dynamic games, where the PBE refines away non-credible threats.
Forgetting the off-path beliefs. In signaling games, off-path messages are observed with probability zero in equilibrium, so Bayes' rule does not pin down the receiver's belief. The unrestricted off-path beliefs support many equilibria, which is why refinements like the intuitive criterion matter.

Economic theory Intermediate+

Theorem (existence of separating equilibrium under single-crossing, Spence 1973). Consider a signaling game with two types $H, L$ , two messages (signal $m = 1$ or $m = 0$ ), and the single-crossing cost structure: cost $c_{H} < c_{L}$ of sending $m = 1$ , identical productivity wages $w_{H} > w_{L}$ , with $w_{H} - w_{L}$ the wage gap. A separating equilibrium in which $H$ sends $m = 1$ and $L$ sends $m = 0$ exists if and only if $c_{H} \leq w_{H} - w_{L} \leq c_{L}$ .

Proof. In a separating equilibrium the receiver infers the type from the message: $m = 1$ implies type $H$ and earns wage $w_{H}$ ; $m = 0$ implies type $L$ and earns wage $w_{L}$ . The high type's no-deviation condition is that sending $m = 1$ beats sending $m = 0$ : $w_{H} - c_{H} \geq w_{L}$ , equivalently $c_{H} \leq w_{H} - w_{L}$ . The low type's no-deviation condition is that sending $m = 0$ beats sending $m = 1$ : $w_{L} \geq w_{H} - c_{L}$ , equivalently $c_{L} \geq w_{H} - w_{L}$ . Together, $c_{H} \leq w_{H} - w_{L} \leq c_{L}$ is necessary and sufficient for both no-deviation conditions, and the receiver's wages match the inferred productivity. $□$

Theorem (Cho-Kreps intuitive criterion, 1987). Among the many perfect Bayesian equilibria of a signaling game, the intuitive criterion eliminates any equilibrium in which some type could not possibly benefit from deviating to an off-path message, regardless of the receiver's response. The refinement selects the least-cost separating equilibrium in the Spence game, ruling out pooling equilibria supported by pessimistic off-path beliefs.

Reconstruction. In the Spence game, a pooling equilibrium (all types choose No College, the employer pays the average wage) can be supported by the off-path belief that anyone attending College is the low type — and so the employer pays a low wage even to College graduates. The high type then does not deviate because the low wage does not cover the college cost. Cho and Kreps's intuitive criterion asks: could the low type ever benefit from deviating to College, even under the most favourable belief? If the low type's cost of College exceeds the maximum wage, the low type could never benefit, so a College signal must come from the high type. The receiver should then form the belief that College implies $H$ , pay the high wage, and the pooling equilibrium collapses. The criterion selects the separating equilibrium and is the standard refinement in applied signaling theory.

Bridge. The separating equilibrium builds toward 52.06.01 behavioral economics by showing how costly actions can convey private information even when boundedly rational agents follow simple heuristics, and appears again in 49.06.01 decision theory and Bayesian reasoning as the game-theoretic extension of belief updating. The foundational reason signaling works is the single-crossing property, which makes the signal differentially costly across types, so the action sorts the type by revealed preference. This is exactly the structure that identifies the sender's action with a credible report of their private information, and the bridge is from a single observed message to the receiver's updated posterior belief, with the cost differential as the credibility mechanism. The pattern generalises across the major signaling applications — Spence's job market, Akerlof's lemons, Zahavi's biological handicaps, Crawford-Sobel's cheap talk — each of which treats a costly (or, in cheap talk, free) action as a carrier of private information, and the central insight is that an informationally rational receiver should read each action under the constraint that the sender would not have taken it unless it was in their interest given their type.

Exercises Intermediate+

Exercise 2 (easy, multiple choice).

In Akerlof's lemons model (used cars), the market for good-quality cars can collapse because:

A. Buyers prefer bad cars B. Buyers cannot distinguish good from bad cars, so they offer the average price, which is below the reservation price of good-car sellers, who then withdraw — leaving only the bad cars C. Sellers prefer to keep their good cars D. Government regulation bans good-car sales

Hint

The adverse-selection logic: when quality is unobservable, the price reflects average quality. Above-average-quality cars are worth more than the price, so their sellers withdraw.

Answer

B. Akerlof's lemons model: when buyers cannot distinguish good from bad cars, they offer a price equal to the expected value of a randomly selected car, which is below the good-car seller's reservation price. The good-car sellers therefore withdraw from the market, leaving only the bad cars ("lemons"). The remaining market is even worse on average, so buyers lower their offers further, and the unraveling can continue until only the worst cars trade. The model is the foundation of the economics of adverse selection.

Exercise 3 (medium, symbolic).

In the Spence game, derive the wage the employer pays in a pooling equilibrium where all types attend college. Show that this wage equals the average productivity across types weighted by the prior.

Hint

In a pooling equilibrium, the receiver's posterior equals the prior, so the wage equals the prior-weighted average of the types' productivities.

Answer

In a pooling equilibrium both types choose the same message (College). The receiver's posterior after seeing College equals the prior $p$ on type $H$ and $1 - p$ on type $L$ . The wage that maximises the receiver's expected payoff (assuming competitive employers earning zero profit) is the expected productivity

w^{pool} = p \cdot θ_{H} + (1 - p) \cdot θ_{L},

with $θ_{H}, θ_{L}$ the two productivities. For $p = 0.5$ , $θ_{H} = 300$ , $θ_{L} = 150$ , this gives $w^{pool} = 225$ . The pooling wage is the average productivity, and neither type earns their true marginal product.

Exercise 4 (medium, numeric).

In the Spence game with type-H productivity $$300, 000$ , type-L productivity $$150, 000$ , equal proportions, and college costs $c_{H} = $40, 000$ , $c_{L} = $120, 000$ , compute the welfare loss from the separating equilibrium relative to the first-best (no signaling, each type paid their productivity).

Hint

In the first-best, neither type attends college (since college raises no productivity) and each is paid their productivity. In the separating equilibrium, the high type attends college and incurs $c_{H}$ . The welfare loss is the total signaling cost.

Answer

$$20, 000$ per high-type worker. In the first-best, total welfare per high type is $$300, 000$ (productivity) and per low type is $$150, 000$ . In the separating equilibrium, the high type attends college and total welfare is $$300, 000 - $40, 000 = $260, 000$ . The low type does not attend, so their welfare is unchanged at $$150, 000$ . The welfare loss is $$40, 000$ per high-type worker, or $$20, 000$ averaged across the population (since half are high-type). The loss is the social cost of signaling: college is pure waste from a productivity standpoint.

Exercise 5 (medium, multiple choice).

The intuitive criterion of Cho and Kreps refines the set of equilibria by:

A. Eliminating equilibria in which some type could not benefit from deviating to an off-path message even under the most favourable belief B. Eliminating all separating equilibria C. Requiring players to be risk-averse D. Restricting the message set to two messages

Hint

The intuitive criterion constrains off-path beliefs. A deviation that some type could never benefit from must be attributed to the type(s) that could benefit, which then constrains the receiver's belief.

Answer

A. The intuitive criterion eliminates equilibria supported by off-path beliefs that ascribe a deviation to a type who could not possibly benefit from it. In the Spence game, a pooling equilibrium supported by the belief that "anyone attending College is the low type" is eliminated: the low type's college cost exceeds the high wage, so the low type could never benefit from deviating, regardless of the receiver's response. The receiver should infer that a College signal must come from the high type, and the pooling equilibrium collapses. The criterion selects the least-cost separating equilibrium in the Spence game and is the standard refinement in applied work.

Exercise 6 (medium, multiple choice).

In a cheap-talk game (Crawford-Sobel), communication is possible between sender and receiver but messages are costless and unverifiable. The central finding is:

A. No communication is ever informative B. Informative communication is possible but bounded: the more the sender and receiver's interests diverge, the less information can be credibly transmitted C. The sender always tells the truth D. The receiver ignores all messages

Hint

When interests align, cheap talk can convey a lot; when interests diverge, only coarse information is credible. Crawford and Sobel quantified this trade-off.

Answer

B. Crawford and Sobel (1982) showed that even costless, unverifiable communication can be informative when the sender's and receiver's interests are sufficiently aligned. As interests diverge (measured by a bias parameter $b$ ), the maximum number of distinct messages that can be credibly transmitted falls: when $b = 0$ , fully informative communication is possible; as $b$ grows, only coarser partitions of the state space can be communicated; and beyond a threshold, no information is transmitted. The model is the foundation of the economics of strategic communication.

Exercise 7 (hard, symbolic).

Prove that, in a signaling game satisfying the single-crossing condition and continuous message space, the least-cost separating equilibrium is unique up to the off-path beliefs, and is selected by the intuitive criterion. Sketch the argument for the two-type case.

Hint

The least-cost separating equilibrium has the high type choosing the minimum signal that the low type would not want to mimic. Any separating equilibrium with a higher signal is dominated. The intuitive criterion eliminates equilibria with non-credible off-path beliefs that would otherwise support pooling or higher-signal separating equilibria.

Answer

In the two-type Spence game with continuous message $m \in [0, \infty)$ , single-crossing costs $c_{H} (m) < c_{L} (m)$ for $m > 0$ , and wages $w_{H}, w_{L}$ equal to inferred productivities, a separating equilibrium has the high type choosing $m^{*}$ such that the low type's no-mimic condition binds: $w_{H} - c_{L} (m^{*}) = w_{L}$ , so $m^{*} = c_{L}^{- 1} (w_{H} - w_{L})$ . The high type's payoff is $w_{H} - c_{H} (m^{*})$ , which is greater than $w_{L}$ by single-crossing. Any signal $m^{'} > m^{*}$ is also a separating equilibrium (the low type still does not want to mimic, the high type is worse off but does not deviate), but $m^{*}$ is the least-cost separating equilibrium and Pareto-dominates the others. The intuitive criterion eliminates the higher-signal equilibria: at an off-path signal $m^{'} < m^{*}$ , the low type could not benefit (the wage gap is below its cost), so the receiver should infer $H$ , pay $w_{H}$ , and the high type would deviate — collapsing any equilibrium with $m > m^{*}$ . The unique intuitive-criterion equilibrium is the least-cost separating one.

Exercise 8 (hard, multiple choice).

The Zahavi handicap principle in evolutionary biology holds that:

A. Animals signal honestly because they are moral B. Costly, wasteful traits (the peacock's tail, the stag's antlers) are honest signals of genetic quality precisely because they are costly and only high-quality individuals can afford them C. All animal signals are deceptive D. Animal signals carry no information

Hint

The handicap principle is the biological analogue of Spence signaling: cost is the credibility mechanism. Amotz Zahavi argued that costly handicaps are honest signals of quality.

Answer

B. Zahavi's handicap principle (1975) holds that costly, wasteful traits like the peacock's tail are honest signals of genetic quality precisely because they are costly. A low-quality individual cannot afford to grow and carry a large tail, so only high-quality individuals display large tails. The principle is the biological analogue of Spence signaling, with the costliness of the trait providing the same single-crossing credibility mechanism. Alan Grafen formalised the principle in game-theoretic terms in 1990, showing that honest handicap signaling is an evolutionarily stable strategy under standard conditions.

Advanced results Master

Result 1 (Akerlof's lemons, 1970). In a market where product quality is privately known to sellers and unobservable to buyers, the equilibrium price reflects average quality. Above-average-quality sellers, whose reservation price exceeds the average, withdraw from the market, lowering the average further, and the unraveling can continue until only the worst quality trades. The model is the foundation of the economics of adverse selection and earned Akerlof the 2001 Nobel Prize jointly with Spence and Stiglitz.

Result 2 (Spence's job-market signaling, 1973). When one party (a worker) has private information about a productive trait (ability), a costly action (education) that is cheaper for the high-ability type can credibly signal the trait to an uninformed party (an employer). The separating equilibrium, in which high types signal and low types do not, sorts the types without raising productivity. The model is the foundation of the economics of signaling.

Result 3 (Cho-Kreps intuitive criterion, 1987). Among the many perfect Bayesian equilibria of a signaling game, the intuitive criterion eliminates any equilibrium supported by off-path beliefs that ascribe a deviation to a type who could not possibly benefit from it. The refinement selects the least-cost separating equilibrium in the Spence game and is the standard equilibrium-refinement tool in applied signaling theory.

Result 4 (Crawford-Sobel cheap talk, 1982). Even costless, unverifiable communication between a sender and a receiver with partially aligned interests can convey information. The maximum amount of information conveyed depends on the bias parameter $b$ measuring the interest divergence: when $b = 0$ , fully informative communication is possible; as $b$ grows, only coarser partitions of the state space can be credibly transmitted; and beyond a threshold, no information is conveyed. The model is the foundation of strategic-communication economics.

Result 5 (Riley 1975, reactive equilibrium). In signaling games with a continuum of types, the Riley equilibrium is the unique separating equilibrium satisfying a stability condition (continuity in the type distribution) and is selected by standard refinements. The equilibrium is the canonical solution concept for continuous-type signaling models and is the basis of the modern theory of screening and nonlinear pricing (Mussa-Rosen 1978, Maskin-Riley 1984).

Result 6 (Stiglitz-Weiss 1981, credit rationing). In a credit market where borrower risk is privately known, the interest rate acts as an adverse-selection device: higher rates drive out safe borrowers, leaving only risky ones, so lenders rationally ration credit below the market-clearing rate rather than raise rates. The model is a leading application of the adverse-selection framework to financial markets and earned Stiglitz the 2001 Nobel Prize jointly with Akerlof and Spence.

Synthesis. The signaling model builds toward 52.06.01 behavioral economics by showing how boundedly rational agents can credibly convey private information through costly actions, and appears again in 49.06.01 decision theory and Bayesian reasoning as the game-theoretic extension of belief updating under strategic uncertainty. The foundational reason signaling works is the single-crossing property, which makes the cost of the signal differentially lower for the high-ability type, so the choice of whether to signal sorts the type by revealed preference. This is exactly the structure that identifies the sender's action with a credible report of their private information, and putting these together with the Cho-Kreps intuitive criterion and the Crawford-Sobel cheap-talk bounds, the bridge is from a single observed message to the receiver's updated posterior belief, with the cost differential (or, in cheap talk, the interest alignment) as the credibility mechanism. The pattern generalises across the major signaling applications — Spence's job market, Akerlof's lemons, Zahavi's biological handicaps, Crawford-Sobel's cheap talk, Stiglitz-Weiss credit rationing — each of which treats an action or message as a carrier of private information, and the central insight is that an informationally rational receiver should read each action under the constraint that the sender would not have taken it unless it was in their interest given their type.

Full proof set Master

Proposition (Separating-equilibrium condition). Consider a two-type signaling game with single-crossing costs $c_{H} (m) < c_{L} (m)$ for all $m > 0$ , $c_{H} (0) = c_{L} (0) = 0$ , and competitive-receiver wages equal to inferred productivity ( $w_{H}$ for type $H$ , $w_{L}$ for type $L$ ). A separating perfect Bayesian equilibrium in which $H$ chooses $m_{H} > 0$ and $L$ chooses $m_{L} = 0$ exists if and only if $c_{H} (m_{H}) \leq w_{H} - w_{L}$ and $c_{L} (m_{H}) \geq w_{H} - w_{L}$ .

Proof. Necessity: the high type's no-deviation condition requires that sending $m_{H}$ (and earning $w_{H} - c_{H} (m_{H})$ ) beats sending $m_{L} = 0$ (and earning $w_{L}$ ), giving $c_{H} (m_{H}) \leq w_{H} - w_{L}$ . The low type's no-deviation condition requires that sending $m_{L} = 0$ beats mimicking $m_{H}$ (which would earn $w_{H} - c_{L} (m_{H})$ ), giving $c_{L} (m_{H}) \geq w_{H} - w_{L}$ . Sufficiency: when both inequalities hold, neither type wants to deviate given the receiver's belief-driven wages. The receiver's beliefs on the equilibrium path are pinned down by Bayes' rule (each message is sent by exactly one type); off-path beliefs can be set to deter deviations (any $m \neq = m_{H}, m_{L}$ is met with wage $w_{L}$ , ensuring no deviation is profitable). The strategy-belief pair is a PBE. $□$

Proposition (Least-cost separating equilibrium under intuitive criterion). Under the intuitive criterion, the unique selected equilibrium is the least-cost separating equilibrium, in which the high type chooses the minimum signal $m^{*}$ satisfying the low type's no-mimic condition with equality.

Proof. The least-cost separating signal $m^{*}$ satisfies $c_{L} (m^{*}) = w_{H} - w_{L}$ (the low type's no-mimic binds). At any signal $m^{'} < m^{*}$ , the low type would mimic, so $m^{*}$ is the minimum separating signal. Any separating equilibrium with $m > m^{*}$ is also a PBE under suitable off-path beliefs, but the intuitive criterion eliminates it: at an off-path signal $m^{''} \in (m^{*}, m)$ , the low type's cost exceeds $w_{H} - w_{L}$ (because $c_{L}$ is increasing), so the low type could not benefit from deviating to $m^{''}$ regardless of the receiver's response. The intuitive criterion then forces the receiver to attribute $m^{''}$ to the high type, pay $w_{H}$ , and the high type would deviate from $m$ to $m^{''}$ (saving $c_{H} (m) - c_{H} (m^{''}) > 0$ ) — collapsing the equilibrium. Only the least-cost separating equilibrium at $m^{*}$ survives. $□$

Connections Master

Game theory — Nash equilibrium and strategic interaction 52.04.01. Signaling games are the central extension of complete-information game theory to settings of asymmetric information. Bayesian Nash equilibrium, perfect Bayesian equilibrium, and the intuitive criterion are all developed to handle the strategic inference problem that signaling poses.
Behavioral economics — bounded rationality, heuristics, and biases 52.06.01. The signaling framework interacts with bounded rationality in two directions: real senders and receivers use heuristics that sometimes violate the equilibrium predictions, and the cost-differential credibility mechanism suggests design principles for institutions that elicit honest reporting even from boundedly rational agents.
Decision theory and Bayesian reasoning 49.06.01. The receiver's belief update in a signaling game is an application of Bayes' rule under the strategic constraint that the sender's message is itself chosen to influence the belief. The interplay of strategic and probabilistic reasoning is the bridge between decision theory and game theory.

Historical & philosophical context Master

George Akerlof's 1970 Quarterly Journal of Economics paper on the market for lemons ^{[Akerlof1970]} is the founding document of the economics of asymmetric information. Akerlof showed that when product quality is privately known to sellers, the market can unravel: the average price drives out high-quality goods, leaving only low-quality goods to trade. The paper was famously rejected by three journals before publication and is now one of the most cited papers in economics.

Michael Spence's 1973 paper on job-market signaling ^[Spence1973], published in the Quarterly Journal of Economics, introduced the signaling model. Spence's insight was that a costly action (education) with no direct productivity effect could still convey private information because the cost differed by type. The model founded the signaling literature and is the foundation of the modern analysis of education, advertising, warranties, and other costly credentials.

The equilibrium-refinement literature is anchored by In-Koo Cho and David Kreps's 1987 Handbook of Game Theory chapter ^{[ChoKreps1987]} on signaling games and stable equilibria, which introduced the intuitive criterion. Vincent Crawford and Joel Sobel's 1982 Econometrica paper ^{[CrawfordSobel1982]} on strategic information transmission introduced the cheap-talk model. The 2001 Nobel Prize in Economics was awarded jointly to Akerlof, Spence, and Joseph Stiglitz for their work on asymmetric information. The deeper lineage runs through von Neumann-Morgenstern game theory (1944) and John Harsanyi's 1967-68 work on Bayesian games, which introduced the type-space formalism that signaling games depend on.

Bibliography Master

@article{Akerlof1970,
  author = {Akerlof, G. A.},
  title = {The Market for ``Lemons'': Quality Uncertainty and the Market Mechanism},
  journal = {Quarterly Journal of Economics},
  volume = {84},
  number = {3},
  pages = {488--500},
  year = {1970},
}

@article{Spence1973,
  author = {Spence, M.},
  title = {Job Market Signaling},
  journal = {Quarterly Journal of Economics},
  volume = {87},
  number = {3},
  pages = {355--374},
  year = {1973},
}

@incollection{ChoKreps1987,
  author = {Cho, I.-K. and Kreps, D. M.},
  title = {Signaling Games and Stable Equilibria},
  booktitle = {Essays in Honor of Joseph Stiglitz},
  publisher = {MIT Press},
  address = {Cambridge, MA},
  year = {1987},
}

@article{CrawfordSobel1982,
  author = {Crawford, V. P. and Sobel, J.},
  title = {Strategic Information Transmission},
  journal = {Econometrica},
  volume = {50},
  number = {6},
  pages = {1431--1451},
  year = {1982},
}

@book{MasColellWhinstonGreen1995,
  author = {Mas-Colell, A. and Whinston, M. D. and Green, J. R.},
  title = {Microeconomic Theory},
  publisher = {Oxford University Press},
  address = {New York},
  year = {1995},
}

@book{FudenbergTirole1991,
  author = {Fudenberg, D. and Tirole, J.},
  title = {Game Theory},
  publisher = {MIT Press},
  address = {Cambridge, MA},
  year = {1991},
}

@article{StiglitzWeiss1981,
  author = {Stiglitz, J. E. and Weiss, A.},
  title = {Credit Rationing in Markets with Imperfect Information},
  journal = {American Economic Review},
  volume = {71},
  number = {3},
  pages = {393--410},
  year = {1981},
}

@article{Riley1975,
  author = {Riley, J. G.},
  title = {Competitive Signaling},
  journal = {Journal of Economic Theory},
  volume = {10},
  number = {2},
  pages = {175--186},
  year = {1975},
}

Prerequisites

52.04.01

Tier anchors

beginner: Gibbons — Game Theory for Applied Economists (Princeton 1992) Ch. 3-4 on dynamic and Bayesian games; Spence — Job Market Signaling (Quarterly Journal of Economics 87, 1973) introductory section
intermediate: Mas-Colell-Whinston-Green — Microeconomic Theory (Oxford 1995) Ch. 13 on adverse selection and signaling; Fudenberg-Tirole — Game Theory (MIT 1991) Ch. 8-9 on Bayesian and extensive-form games
master: Akerlof 1970 The Market for Lemons (QJE 84); Spence 1973 Job Market Signaling (QJE 87); Cho-Kreps 1987 Signaling Games and Stable Equilibria (Handbook of Game Theory); Crawford-Sobel 1982 Strategic Information Transmission (Econometrica 50)

References

Spence — Job Market Signaling · Quarterly Journal of Economics 87, 355-374 (1973); originator of the signaling model and the job-market application
Akerlof — The Market for Lemons: Quality Uncertainty and the Market Mechanism · Quarterly Journal of Economics 84, 488-500 (1970); originator of the adverse-selection analysis
Cho and Kreps — Signaling Games and Stable Equilibria · Handbook of Game Theory with Economic Applications, Volume 1, North-Holland 1987; originators of the intuitive criterion equilibrium refinement
Crawford and Sobel — Strategic Information Transmission · Econometrica 50, 1431-1451 (1982); originators of the cheap-talk model of communication
Mas-Colell, Whinston, and Green — Microeconomic Theory · Oxford University Press 1995, Ch. 13 on information economics and signaling
Fudenberg and Tirole — Game Theory · MIT Press 1991, Ch. 8 (Bayesian Nash equilibrium) and Ch. 9 (perfect Bayesian equilibrium and refinements)

Estimated time

beginner: 15m
intermediate: 45m
master: 80m