“On the Equivalence of Bayesian and Dominant Strategy Implementation,” A. Gershkov et al (2012)

Let’s take a quick break from the job market. Bayesian incentive compatibility is not totally satisfying when trying to implement some mechanism, as each agent must care about the rationality and actions of other agents. Dominant strategy incentive compatibility (DSIC) is much cleaner: no matter what other agents do, the action that makes me best off is just to state my preferences truthfully. For example, the celebrated Vickrey-Clarke-Groves mechanism is dominant strategy incentive compatible. Even if other agents lie or make mistakes, the transfers I pay (or receive) are set so that I have just enough incentive to report my true type.

Many clever mechanisms are not DSIC, however. Consider the Cremer-McLean auction with correlated types. There are a bunch of bidders with correlated private values for an oil field. If the values were independent, in the optimal mechanism, the winning bidder gets an information rent. Consider a first price auction, where all of our values are distributed uniformly on [0,1]. I draw .7 for my value. I won’t actually bid .7, because bidding slightly lower increases my profit when I win, but only slightly decreases my probability of winning. If the auctioneer wants to use a scheme that reveals my true value, he better give me some incentive to reveal; the second-price auction does just that, by making me pay only the second-highest bid if I win.

Cremer-McLean does the following. If I want to enter the auction, I need to make a side bet that gives me zero expected payoff if both I and everyone else report our true types, and a very negative payoff if somebody lies. In particular, charge me an entry fee that depends only on what I think other people’s bids will be. Conditional on everyone else telling the truth, I am perfectly happy as a Bayesian to report truthfully. My willingness to accept the bet reveals, because of correlated private values, something about my own private value. But now the auctioneer knows our true values, and hence can extract the full surplus in the auction without paying any information rents. Many, many people – Milgrom and Wilson, famously – find this mechanism rather unsatisfying in the real world. Certainly, it’s tough to think of any social choice mechanism that uses the Cremer-McLean strategy, or even something similar. This may be because it relies heavily on knife-edge Bayesian reasoning and common knowledge of rationality among the players, much stronger conditions than we need for dominant strategy incentive compatibility.

Manelli and Vincent have a 2010 Econometrica with a phenomenal result: in many simple cases, Bayesian IC gives me nothing that I can’t get from a DSIC. This makes sense in some ways: consider the equivalence of the Bayesian mechanism of a sealed-bid auction and the dominant strategy mechanism of a second price auction. What Gershkov et al do is extent that equivalence to a much broader class of social choice implementation. In particular, take any social choice function with one dimensional independent types, and quasi-linear utility for each agent. If there is a Bayesian IC mechanism to implement some social choice, then I can write down allocations and transfers which are DSIC and give the exact same interim (meaning after individuals learn their private types) expected utility for every agent. That is, in any auction with independent private values and linear utility, there is nothing I can do with Bayesian mechanisms that I can’t do with much more plausible mechanisms.

How does this work? Recall that the biggest difference between Bayesian IC and DSIC is that a mechanism is Bayesian IC (on a connected type space) if expected utility from an allocation rule is non-decreasing in my own type, and DSIC if utility from the allocation rule is non-decreasing in my own type no matter what the types of the other agents. Gershkov et al give the following example. I want to give an object to the agent with the highest value, as long as her value is not more than .5 higher than the other agent. Both agent’s values are independent drawn from U[0,1]. If the difference between the two is more than .5, I want to allocate the good to no one. Just giving an agent the good with probability 1 if his type is higher than the other agent’s report and lower than the other agent’s report plus .5 is Bayesian incentive compatible (the marginal of expected utility is nondecreasing in my type, so there must exist transfer payments that implement), but not DSIC: if the other agent reports his type minus .1, then I want to shade an equal amount. However, consider just giving an agent the good with probability equal to the minimum of his report and .5. If my type is .7, then I get the object with probability .5. This is exactly the interim probability I would get the object in the Bayesian mechanism. Further, the allocation probability is increasing in my own type no matter what the other agent’s type, so there must exist transfer payments that implement in dominant strategies. The general proof relies on extending a mathematical proof from the early 1990s: if a bounded, non-negative function of several variables generates monotone, one-dimensional marginals, then there must exist a non-negative function with the same bound, and the same marginals, that is monotone is each coordinate. The first function looks a lot like the condition on allocation rules for Bayesian IC, and the second the condition on allocation rules for DSIC…

Final Econometrica preprint (IDEAS version)

About these ads

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s


Get every new post delivered to your Inbox.

Join 169 other followers

%d bloggers like this: