Independent scholarship · Open accessThe monthly digest

Indexed text attachment

games-and-information-in-higher-categories.tex

Source for Games and Information in Higher Categories

View paper · Download original file

Lines 4359–4458 of 6559. Source text is displayed literally.

4359unexpected signals is unconstrained, and there are many equilibria. When the uninformed party moves first4360she commits to a response to \emph{every} choice in advance, there are no off-path beliefs to play with,4361and the equilibrium is unique or fails to exist. Commitment by the uninformed side replaces multiplicity4362by non-existence. The Riley outcome and the Rothschild--Stiglitz separating pair are the same allocation,4363reached from the two sides.4364 4365\section{Cheap talk}4366 4367Suppose the signal is a costless, unverifiable message. Nothing then distinguishes the types' costs of4368sending it, and single crossing in the signal is unavailable. Information can still be transmitted, in4369proportion to the alignment of interests.4370 4371\begin{theorem}[Crawford--Sobel]\label{thm:CS}\Scl[\cite{CS82}]4372Let the sender's type $\theta$ be uniform on $[0,1]$, the receiver choose $a\in\R$ with payoff $-(a-\theta)^2$,4373and the sender have payoff $-(a-\theta-b)^2$ with bias $b>0$. Then every equilibrium is equivalent to a4374\emph{partition equilibrium}: there are cutoffs $0=c_0<c_1<\dots<c_N=1$; types in the same interval send the4375same message; the receiver responds with the midpoint. The cutoffs satisfy4376\[ c_{i+1}-c_i=c_i-c_{i-1}+4b, \]4377and an equilibrium with $N$ intervals exists iff $2N(N-1)b<1$.4378\end{theorem}4379 4380\begin{example}4381With $b=\tfrac1{20}$ the largest $N$ with $2N(N-1)<20$ is $3$. The interval lengths are $d$, $d+\tfrac15$,4382$d+\tfrac25$ with sum $1$, so $d=\tfrac2{15}$ and the partition is4383$[0,\tfrac2{15}],[\tfrac2{15},\tfrac7{15}],[\tfrac7{15},1]$. The one-interval \emph{babbling} equilibrium, in which4384messages are ignored, also exists, as it does for every $b$.4385\end{example}4386 4387An equilibrium of the cheap-talk game is an \emph{endogenous information structure}. The sender's strategy4388is a quotient map from the type space onto a set of messages, its fibres are the intervals, and the4389receiver's knowledge is the operator $p^*\forall_p$ of \cref{def:K} for that quotient. The players choose,4390in equilibrium, how coarse the receiver's partition will be. The more divergent the interests, the coarser4391it must be, because a fine partition would give the sender too many profitable ways to shade her report;4392and the partition is coarser at the top, where the sender's incentive to exaggerate has more room. The4393language is vague because the speaker is biased, and vaguest where the bias matters most.4394 4395\section{Persuasion}\label{sec:persuasion}4396 4397Now give the sender commitment. Before learning the state she chooses an experiment, a kernel4398$\sigma\colon\Omega\to S$ from states to signals, and its realization is observed by the receiver. She cannot lie4399about the realization; she can only choose what will be learned.4400 4401A signal realization induces a posterior, so an experiment induces a distribution of posteriors4402$\tau\in\Dist(\Dist\Omega)$. Which $\tau$ arise?4403 4404\begin{proposition}[Bayes plausibility]\label{prop:bayes-plausible}\Srf[\cite{KG11,AM95}]4405Let $m\colon\Dist\Dist\Omega\to\Dist\Omega$ be the multiplication of the distribution monad. A distribution of4406posteriors $\tau$ is induced by some experiment under the prior $\mu_0$ if and only if $m(\tau)=\mu_0$. The4407experiments, up to relabelling of signals, are the points of the fibre $m^{-1}(\mu_0)$, and the Blackwell4408order on experiments is the order of mean-preserving spreads on this fibre.4409\end{proposition}4410 4411The multiplication $m$ averages a distribution over distributions; the condition says that the posteriors4412average to the prior. That this is the only constraint is the ``splitting lemma'' of Aumann and Maschler.4413Given the receiver's best response, let $\hat v(\mu)$ be the sender's expected payoff when the receiver4414holds the posterior $\mu$. The sender's problem is to maximize the linear functional $\tau\mapsto\int\hat v\,d\tau$4415over the fibre $m^{-1}(\mu_0)$.4416 4417\begin{theorem}[Kamenica--Gentzkow]\label{thm:KG}\Scl[\cite{KG11}]4418The sender's optimal value is $(\cav\hat v)(\mu_0)$, where $\cav\hat v$ is the smallest concave function4419above $\hat v$. She benefits from persuasion iff $(\cav\hat v)(\mu_0)>\hat v(\mu_0)$.4420\end{theorem}4421 4422\begin{example}\label{ex:persuasion}4423A regulator approves a drug iff she believes it effective with probability at least $0.6$; the prior is4424$0.3$; the firm wants approval regardless. Without information the drug is rejected. Full disclosure gets4425it approved with probability $0.3$. The optimal experiment induces posteriors $0$ and $0.6$, with4426probabilities $\tfrac12$ each: an effective drug always passes the test, and an ineffective one passes4427with probability $\tfrac27$. The drug is approved half the time, although it works only three times in4428ten. Here $\hat v=1_{[0.6,1]}$ and $\cav\hat v(\mu)=\min\{\mu/0.6,1\}$. The receiver is exactly as well off as4429with no information: the sender extracts the entire value of the information she designs.4430\end{example}4431 4432Persuasion is a mechanism-design problem in which the instrument is information rather than money, and its4433solution has the same form as those of \cref{ch:mechanism}: a hull. There the hull was the convex envelope4434that irons the virtual type; here it is the concave envelope of the value over the simplex of beliefs.4435With several receivers who play a game among themselves, the set of outcomes an information designer can4436induce is the set of Bayes correlated equilibria \cite{BM16} of \cref{sec:correlated}, and the designer is4437the mediator of that section with an objective of her own.4438 4439\section{Signal jamming}4440 4441Finally, let the sender influence the signal by an action that the receiver does not observe.4442 4443\begin{example}[Career concerns]\label{ex:career}\Scl[\cite{Hol99}]4444A manager's output is $y=\eta+a+\varepsilon$, where $\eta\sim N(m,\sigma_\eta^2)$ is her unknown talent, $a$ her4445unobserved effort at cost $c(a)$, and $\varepsilon\sim N(0,\sigma_\varepsilon^2)$. After seeing $y$ the market pays4446her next-period wage equal to its expectation of $\eta$. If the market conjectures effort $a^*$, it computes4447\[ \E[\eta\mid y]=m+\kappa\,(y-m-a^*),\qquad\kappa=\frac{\sigma_\eta^2}{\sigma_\eta^2+\sigma_\varepsilon^2}. \]4448The manager, taking the conjecture as given, gains $\kappa$ in future wage per unit of effort, so with4449discount factor $\delta$ she chooses $c'(a)=\delta\kappa$. In equilibrium $a=a^*$: the market correctly4450subtracts her effort, and her wage is exactly what it would have been had she and everyone else known she4451would not exert it. She works to bias an inference that is not biased.4452\end{example}4453 4454The logic is the lens structure of \cref{ch:open,ch:belief} made vivid. The market's inference is the4455backward pass, and it is computed from the \emph{conjectured} forward pass. A deviation by the manager4456changes the actual forward pass and leaves the backward pass fixed. So although no one is fooled on the4457equilibrium path, the equilibrium effort is pinned down by the return to fooling off it. The same4458structure gives \emph{limit pricing} \cite{MR82b}: an incumbent whose cost is private information prices