Indexed text attachment
games-and-information-in-higher-categories.tex
Source for Games and Information in Higher Categories
View paper · Download original file
Lines 4759–4858 of 6559. Source text is displayed literally.
4759\section{Revenue equivalence}4760 4761In the first-price auction the winner pays her own shaded bid; in the second-price auction she pays the4762runner-up's unshaded value. For uniform values the expected revenue is $\frac{n-1}{n}\cdot\frac{n}{n+1}$ in the4763first and $\frac{n-1}{n+1}$ in the second. They are equal. This is not a coincidence of the uniform4764distribution.4765 4766\begin{theorem}[Revenue equivalence]\label{thm:RET}\Scl[\cite{Vic61,Mye81,RS81}]4767Consider risk-neutral bidders with independent private values on an interval. Any two auctions that, in4768equilibrium, (i) allocate the object in the same way as a function of the values and (ii) give the lowest4769type the same expected utility, yield the same expected payment from every type of every bidder, and the4770same expected revenue.4771\end{theorem}4772 4773\begin{proof}4774From the standpoint of bidder $i$, an auction with its equilibrium is a direct mechanism in which a4775report $v$ leads to a probability $P(v)$ of winning and an expected payment $m(v)$. Her type space is an4776interval and her valuation is $vP$. By \cref{cor:myerson}, $m$ is determined by $P$ up to a constant:4777$m(v)=m(0)+\int_0^vs\,dP(s)$.4778\end{proof}4779 4780All of the content is in \cref{thm:rochet}: on a connected type space the incentive category is a4781groupoid, and a potential on a connected groupoid is unique up to a constant. The theorem is as useful4782for what it computes as for what it compares. In any efficient auction $P(v)=F(v)^{n-1}$, so4783$m(v)=\int_0^vs\,dF(s)^{n-1}$. In a first-price auction the bidder pays her bid when she wins, so4784$b(v)=m(v)/P(v)$, which is the formula above without solving a differential equation. In an \emph{all-pay}4785auction she pays her bid regardless, so $b(v)=m(v)$; for uniform values, $b(v)=\frac{n-1}{n}v^n$. The same4786method gives equilibrium behaviour in wars of attrition, contests and lobbying games, which are auctions4787in which losers pay.4788 4789\section{Optimal auctions}4790 4791Revenue equivalence says that the seller cannot gain by changing the payment rule while keeping the4792allocation. She can gain by changing the allocation---specifically, by sometimes not selling.4793 4794\begin{theorem}[Myerson]\label{thm:optimal-auction}\Scl[\cite{Mye81,RS81}]4795With independent private values and increasing virtual values $\psi(v)=v-\frac{1-F(v)}{\phi(v)}$, expected4796revenue equals the expected virtual value of the winner, and the revenue-maximizing auction awards the4797object to the bidder with the highest virtual value provided it exceeds the seller's own value $v_0$. With4798symmetric bidders this is any standard auction with the reserve price $r$ solving $\psi(r)=v_0$.4799\end{theorem}4800 4801The reserve price does not depend on the number of bidders. For uniform values and $v_0=0$ it is4802$\tfrac12$: the seller refuses bids below one half although every bidder values the object more than she4803does. This is the monopolist's distortion of \cref{ch:mechanism}, exclusion at the bottom to reduce the4804rent at the top. With two bidders the optimal reserve raises revenue from $\tfrac13$ to $\tfrac5{12}$.4805 4806How much is this worth? Less than one more bidder. With three bidders and no reserve, expected revenue is4807$\tfrac12>\tfrac5{12}$, and the comparison is general.4808 4809\begin{theorem}[Bulow--Klemperer]\Scl[\cite{BK96}]4810Under the assumptions of \cref{thm:optimal-auction} with symmetric bidders, an English auction with $n+1$4811bidders and no reserve yields at least as much expected revenue as the optimal auction with $n$ bidders.4812\end{theorem}4813 4814A seller does better to spend effort on attracting participation than on designing the mechanism. It is4815one of the few results in the subject that says what theory is worth compared with marketing.4816 4817\section{Risk aversion}4818 4819Revenue equivalence needs risk neutrality, because it needs payoffs to be quasi-linear in the payment. If4820bidders are risk averse, bidding one's value remains dominant in the second-price auction, since the4821argument does not involve probabilities. In the first-price auction a higher bid buys insurance against4822losing, and risk-averse bidders bid more than risk-neutral ones would. So the first-price auction yields4823strictly higher expected revenue \cite{MR84b}. A risk-averse \emph{seller}, by contrast, prefers the4824first-price auction for a different reason: its revenue, being a function of the highest value alone, is4825less variable.4826 4827\section{Common values}4828 4829Now let the object have a common value $V$ and let each bidder observe a signal $x_i$ correlated with it.4830 4831\begin{example}[The wallet game]\label{ex:wallet}\Scl[\cite{Kle98}]4832Two bidders each know the amount $x_i$ in their own wallet, independent and uniform on $[0,1]$; the prize is4833the sum $V=x_1+x_2$, sold by second-price auction. A bidder's unconditional estimate of the prize is4834$x_i+\tfrac12$. But if she wins against an opponent who bids increasingly in her signal, the opponent's4835signal is below her own, and the estimate \emph{conditional on winning} is $\tfrac32x_i$, which is lower4836unless $x_i=1$. The symmetric equilibrium is $b(x)=2x$: each bids what the prize would be worth if the4837opponent's signal were equal to her own. Then the winner pays $2x_j$ for a prize worth $x_i+x_j$ and earns4838$x_i-x_j>0$, and the loser has no regret. Expected revenue is $\tfrac23$, out of an expected value of $1$.4839\end{example}4840 4841The \term{winner's curse} is the failure to condition on the event of winning. The bidder who wins is the4842one with the most optimistic signal, so winning is bad news about the value; a bidder who bids her4843unconditional estimate will, on average, regret winning. Rational bidders shade, and shade more the more4844competitors there are, since beating more rivals is worse news. The curse is a statement about selection4845of exactly the kind met in \cref{ch:adverse}: the event on which the payoff is realized is correlated with4846the unknown, and the relevant expectation is the one conditional on that event.4847 4848\begin{theorem}[Milgrom--Weber]\label{thm:MW}\Scl[\cite{MW82}]4849Suppose the bidders' signals and the value are \emph{affiliated}: their joint density $g$ satisfies4850$g(z\vee z')\,g(z\wedge z')\ge g(z)\,g(z')$. Then4851\begin{enumerate}[label=(\roman*)]4852\item the second-price auction has the symmetric equilibrium4853 $b(x)=\E\bigl[V\mid x_i=x,\ \max_{j\ne i}x_j=x\bigr]$;4854\item expected revenues are ranked: English $\ge$ second-price $\ge$ first-price;4855\item a seller with private information affiliated with the value raises expected revenue by committing to4856 disclose it.4857\end{enumerate}4858\end{theorem}