Indexed text attachment
games-and-information-in-higher-categories.tex
Source for Games and Information in Higher Categories
View paper · Download original file
Lines 4259–4358 of 6559. Source text is displayed literally.
4259\end{exercise}4260%% ---- end ch/14-mechanism.tex ----4261%% ---- begin ch/15-signalling.tex ----4262\chapter{Signalling}\label{ch:signalling}4263 4264If the informed party moves first, her action can reveal what she knows. Whether it does, and how much, is4265determined in equilibrium: the meaning of a signal is the Bayesian inverse of the strategy that produces4266it. This chapter treats four models of strategic information transmission, ordered by how much the sender4267can commit: costly signals, costless messages, signals whose meaning the sender designs in advance, and4268signals that the sender distorts in secret.4269 4270\section{Signalling games}4271 4272Nature draws a type $\theta\in\Theta$ from a prior $\pi$. The sender observes $\theta$ and chooses a signal $e\in E$;4273the receiver observes $e$ and chooses an action $a$. A sender strategy is a kernel4274$\sigma\colon\Theta\to E$. The receiver's belief after $e$ is $\mu(e)\in\Dist\Theta$, and in a perfect Bayesian4275equilibrium $\mu$ is the Bayesian inverse $\sigma^\dagger_\pi$ wherever that is defined, the receiver best-responds4276to $\mu$, and the sender best-responds to the receiver.4277 4278The classical trichotomy of equilibria is a trichotomy of kernels. In a \term{separating} equilibrium $\sigma$4279is an injective function, so the inverse is deterministic and the receiver learns the type. In a4280\term{pooling} equilibrium $\sigma$ is constant, the inverse is the prior, and the receiver learns nothing. In4281a \term{hybrid} equilibrium $\sigma$ is properly stochastic. Equilibria are partially ordered by the Blackwell4282order of \cref{sec:informativeness} on the experiments $\sigma$: the question ``how much information is4283transmitted?'' has a precise meaning and, usually, several answers.4284 4285\section{Education as a signal}4286 4287\begin{example}[Spence]\label{ex:spence}\Scl[\cite{Spe73}]4288A worker's productivity is $\theta_L=1$ or $\theta_H=2$; the fraction of high types is $\lambda$. Before4289entering the labour market she chooses a level of education $e\ge0$ at cost $e/\theta$. Education does not4290affect productivity. Competing employers observe $e$ and pay the expected productivity given $e$.4291 4292\emph{Separating equilibria.} The low type chooses $e=0$ and is paid $1$; the high type chooses $e^*$ and is4293paid $2$. The low type must not wish to imitate, $2-e^*\le1$, and the high type must not wish to give up,4294$2-e^*/2\ge1$. So every $e^*\in[1,2]$ gives a separating equilibrium, supported by the belief that any4295$e<e^*$ comes from the low type.4296 4297\emph{Pooling equilibria.} Both types choose the same $e_p$ and are paid $1+\lambda$, with deviations4298attributed to the low type; this is an equilibrium iff $e_p\le\lambda$.4299 4300\emph{Welfare.} In the least wasteful separating equilibrium, $e^*=1$, the high type nets $1.5$ and the low4301type $1$. In the pooling equilibrium at $e_p=0$ both net $1+\lambda$. If $\lambda>\tfrac12$, pooling is better4302for \emph{both} types: the high types burn half a unit each to distinguish themselves from a small minority.4303\end{example}4304 4305What makes separation possible is that the cost of the signal falls with type: the \emph{single-crossing}4306condition of \cref{thm:MS}. The high type's indifference curves in the plane of education and wage are4307flatter, so there are contracts she prefers and the low type does not. What makes signalling socially4308costly is that the signal is otherwise useless. Education in this model produces nothing and reveals4309everything. The private return to signalling is the wage difference; the social return is zero, or the4310value of a better match if the information is used. As in the war of attrition of \cref{ch:mixing}, a4311contest for a fixed prize dissipates it.4312 4313\section{Refinements}4314 4315The example has a continuum of separating equilibria and a continuum of pooling ones. They differ only in4316beliefs after signals that are never sent, which Bayes' rule does not restrict. Sequential equilibrium4317does not help; every belief at an unreached information set following a single deviation is consistent.4318Refinements for signalling games restrict off-path beliefs by asking which types could conceivably gain4319from a deviation.4320 4321\begin{definition}[Intuitive criterion]\Scl[\cite{CK87}]4322Fix an equilibrium with payoffs $u^*(\theta)$. A type $\theta$ is \emph{equilibrium-dominated} at an4323unsent signal $e$ if her payoff from $e$ is below $u^*(\theta)$ for every best response of the receiver to4324every belief. The equilibrium \emph{fails the intuitive criterion} if for some unsent $e$ some type would4325gain from sending $e$ whenever the receiver's belief is concentrated on the types not equilibrium-dominated4326at $e$.4327\end{definition}4328 4329\begin{proposition}\label{prop:intuitive}\Scl[\cite{CK87}]4330In \cref{ex:spence} the only equilibrium outcome satisfying the intuitive criterion is the separating4331equilibrium with $e^*=1$.4332\end{proposition}4333 4334\begin{proof}4335At a separating equilibrium with $e^*>1$, consider $e'\in(1,e^*)$. The low type cannot gain from $e'$ even4336if believed to be high, since $2-e'<1$; the high type gains if believed high. So beliefs after $e'$ must be4337concentrated on the high type, and she deviates. At a pooling equilibrium at $e_p$, a signal $e'$ with4338$1-\lambda<e'-e_p<2(1-\lambda)$ is equilibrium-dominated for the low type and attractive to the high type if4339believed high.4340\end{proof}4341 4342The surviving outcome is the \emph{Riley outcome}: the separating equilibrium with the least signalling.4343The argument is a form of forward induction, and in generic signalling games the outcomes that pass it4344include those of Kohlberg--Mertens stable sets (\cref{sec:thompson}); stronger criteria of the same family4345select the Riley outcome with any number of types \cite{BS87}. The selection has an awkward feature. As4346$\lambda\to1$ the Riley outcome does not change: when almost everyone is a high type, each of them still4347spends $e^*=1$ to avoid being taken for one of the vanishing few, although pooling at zero would be better4348for everyone. And at $\lambda=1$ exactly, there is nothing to signal and $e=0$. The refined equilibrium4349correspondence is discontinuous in the prior. This is the phenomenon of \cref{thm:WY} in miniature, and a4350reason for caution about any refinement that is insensitive to probabilities.4351 4352\section{Screening}4353 4354If instead the uninformed party moves first, offering a menu of contracts from which the informed party4355chooses, the model is one of \emph{screening}. The monopoly version is the nonlinear pricing of4356\cref{ch:mechanism}; the competitive version is the insurance market of \cref{thm:RS}. The comparison with4357signalling is instructive. The two games have the same boxes composed in a different order, and the4358equilibrium sets differ sharply. When the informed party moves first, the uninformed party's response to