EC992 - Advanced Microeconomics. Lecture 5: Bayesian Games
Transcription
EC992 - Advanced Microeconomics. Lecture 5: Bayesian Games
EC992 - Advanced Microeconomics. Lecture 5: Bayesian Games and Auctions Ludovic Renou Academic year: 2013-204 Games with Incomplete Information Bayesian games model strategic situations in which players are uncertain about the situation of other players. DEFINITION: A Bayesian game consists of • a finite set N (the set of players) • a set Ω (the set of states) • for each player i ∈ N a nonempty set Ai (the set of actions available to player i) • for each player i ∈ N a nonempty set Ti (the set of signals or types of player i) and a signal function τi : Ω → Ti • for each player i ∈ N a probability measure pi on Ω (the prior belief of player i) • for each player i ∈ N a payoff (or utility ) function ui : A → R (where A := ×j∈N Aj is the set of action profiles or outcomes); preferences over lotteries on A × Ω are given by expected utility. This definition allows players to have different prior beliefs. They are commonly assumed to be identical (common prior assumption). Frequently the state of nature is a profile of parameters that affect players’ payoffs. However the model is more general. Often the model is given in reduced form: player i knows his own type, and has a belief over the types of the other players pi (t−i |ti ), conditional on his type. Example 1. Two players must decide between two actions. Player 2 would like to play the same action than player 2, and prefers B to A. The preferences of player 1 are not known to player 2. Player 1 could be of one of two types. Type t1:1 has the preferences represented in the left table. Type t1:2 has the preferences represented on the right table. Player 2 has no private information and hence there is only one type, t2 , of player 2. Player 2 attaches probability p to Player 1 being of type t1:1 . A B A B 2, 1 0, 0 0, 0 1, 2 A B A B 0, 1 1, 0 2, 0 0, 2 Formally, N = {1, 2}, Ω = {ω1 , ω2 }, τ1 (ω1 ) = t1:1 , τ1 (ω2 ) = t1:2 , , τ2 (ω1 ) = τ2 (ω2 ) = t2 , p2 (ω1 ) = p. The following definition is most useful when, for all i, the set of types Ti is finite. DEFINITION: A Bayesian equilibrium of the Bayesian game hN, Ω, (Ai ), (Ti ), (τi ), (pi ), (ui )i is a Nash equilibrium of the game in which each type of each player is viewed as an independent player. An equivalent, alternative, approach is to view the strategy of player i as a function from i’s types to his action set, si : Ti → Ai . DEFINITION: A (pure strategy) Bayesian equilibrium of the Bayesian game hN, Ω, (Ai ), (Ti ), (τi ), (pi ), (ui )i is a profile s ∗ = (s1∗ , ..., sn∗ ) of ∗ , s ∗ ) ≥ U (s ∗ , s ) for all functions si∗ : Ti → Ai such that, Ui (s−i i −i i i si : Ti → Ai , where Ui (s−i , si ) is player i’s expected utility when the strategy profile is (s−i , si ). (This expected utility is computed using the beliefs pi of player i concerning the.type profile t. It is equivalent to each type ti of player i, for all i, maximizing his expected utility conditional on his type ti ). Example 1. Exercise: Find the pure strategy Bayesian equilibria. Example 2. Player 1 is an incumbent firm, player 2 is a potential entrant. Player 1 decides whether to build a new plant (B), or not (N), player 2 simultaneously decides whether to enter (E ) or stay out (O). The building cost of the incumbent, player 1, is not known to player 2. If the cost is high, payoffs are in the left table. If the cost is low, payoffs are as in the right table. Player 2 has no private information. The entrant, player 2 attaches probability p to the incumbent having high cost. E B N 0, −1 2, 1 O 2, 0 3, 0 E B N 1.5, −1 2, 1 Exercise: Find the pure strategy Bayesian equilibria. O 3.5, 0 3, 0 Purification of mixed strategy equilibrium:. Given a mixed strategy equilibrium, of a strategic game, one can construct a “close” Bayesian game with pure strategy Bayesian equilibrium which is “close” to the mixed strategy equilibrium of the original strategic game. Example 3. (A Bayesian version of matching pennies). The types t1 of player 1 and t2 of player 2 are elements of the interval [0, 1]. The beliefs of player i about player j’s type are uniformly distributed, i.e., the density of type t is f (t) = 1 and the distribution is F (t) = t. The payoffs are shown below, where ε > 0. H H T 1 + ε(2t1 − 1), −1 −1, 1 + ε(2t2 − 1) T −1, 1 1, −1 Exercise: Find the pure strategy Bayesian equilibrium. Show that it converges to the mixed strategy Nash equilibrium of matching pennies as ε → 0. Example 4. (Providing a public good). Two players simultaneously decide whether to contribute to a public good C , or not N (contributing is a 0-1 decision). Each player derives a benefit of 1 from the public good which is provided if at least one contributes. Players costs of contributing are ci and are privately know. The types c1 of player 1 and c2 of player 2 are elements of the interval [0, 2]. The beliefs of player i about player j’s type are uniformly distributed, i.e., the density of type t is f (t) = 1/2 and the distribution is F (t) = t/2. The payoffs are shown below. C C N 1 − c1 , 1 − c2 1, 1 − c2 N 1 − c1 , 1 0, 0 Exercise: Find the pure strategy Bayesian equilibrium. (Hint: Players contribute to providing the public good if and only if their cost is below a threshold.) Example 5. (A bargaining game). A buyer and a seller have private information about their valuations for an object, vB and vS . The (beliefs about the) valuations are uniformly distributed on [0, 1]. If trade does not take place, payoffs of both players are normalized to zero. If trade takes place at price p, then the buyer’s payoff is vB − p, while the seller’s payoff is p − vS . Buyer and seller simultaneously announce prices pB and pS . Trade only takes place if pB ≥ pS . If trade takes place, the price is (pB + pS )/2. There are very many Bayesian equilibria. Consider the following class of equilibria. For any value of x ∈ [0, 1] the buyer offers pB = x if vB ≥ x and offers pB = 0 otherwise. The sellers asks pS = x if vS ≤ x and asks pS = 1 otherwise. There is also a linear equilibrium in which: pB (vB ) = pS (vS ) = 2 1 vB + 3 12 2 1 vS + 3 4 Exercise: Prove this is a Bayesian equilibrium. Trade occurs in this equilibrium if and only if vB ≥ vS + 1/4. This turns out to be the “most efficient” equilibrium, that is the equilibrium with the highest gains from trade. (Note: equilibrium is not efficient, because sometimes trade does not take place even though the buyer values the object more than the seller.) Auction Theory In the standard symmetric, independent, private value model, there are N bidders for one object. Each bidder’s type x is independently and identically drawn (i.i.d) from the increasing cumulative distribution F (x) with support [0, 1] (this is just a normalization). F has a continuous density function f and E [x] < ∞. Bidders are risk neutral, so that a type x bidder’s value when winning the object at price p is x − p. Each bidder knows his own type, but not the types of the other bidders. If there are interdependent values (often called common values in the literature), then the value of the object to bidder i depends not only on bidder i’s signal, but also on the signals of all other bidders. In such a case the value to bidder i of winning at price p is v (xi , x−i ) − p where v is the value of winning and x−i is the profile of types of allPother bidders (as an example, think of v (xi , x−i ) = N j=1 xj this is an example of pure common values). In a first-price sealed bid auction, bidders simultaneously submit bids; the highest bidder wins and pay his own bid. In a second-price sealed bid auction or Vickrey auction, bidders simultaneously submit bids; the highest bidder wins and pays the second highest bid. In an open ascending auction, or English auction, the price starts low and it is raised continuously by the auctioneer; bidders decide when to drop out; once a bidder has dropped out, he cannot re-enter (this is the theoretically simplest way of modeling ascending open auctions). In an open descending auction, or Dutch auction, the price starts high and it is lowered continuously by the auctioneer; the first bidder to take the current price wins the auction at the current price. Proposition The Dutch auction is strategically equivalent to the sealed bid, first-price auction. Proposition Under private values (i.e., v = xi ), but not under interdependent values, the English auction is equivalent to the sealed bid, second-price auction, in the sense that the optimal strategies in the two auctions are the same. (But not strategically equivalent.) Proposition Under private values (i.e., v (xi , x−i ) = xi ), in a second-price and in an English auction it is a weakly dominant strategy for each bidder to bid b(x) = x (i.e., it is a Bayesian equilibrium for each bidder i of type xi to bid xi ). There are Bayesian equilibria in which players use weakly dominated strategy for example: b(x1 ) = 1 for all x1 ∈ [0, 1] and b(xi ) = 0 for all i 6= 1 and all xi ∈ [0, 1]. It is standard to disregard these equilibria. Let the order statistic Y1 be the highest of N − 1 random draws from the distribution F (x). The distribution of Y1 is F N−1 (x); its density is (N − 1)f (x)F N−2 (x). Proposition Under private values (i.e., v (xi , x−i ) = xi ), it is a symmetric equilibrium of a first-price auction to bid b(x) = E [Y1 |Y1 < x] Z x y (N − 1)f (y )F N−2 (y ) dy = F N−1 (x) 0 Z x N−1 F (y ) = x− dx. N−1 (x) 0 F With interdependent values, we can write v (xi , x−i ) = v (xi , Y1 , Y2 , ...) where the order statistics Y1 , Y2 ,... are the highest, second highest,... of N − 1 random draws from the distribution F (x). Proposition In a second-price auction with interdependent values, it is an equilibrium for each bidder i to bid b(xi ) = E [v (xi , Y1 , ...)|Y1 = xi ]. Proposition Under interdependent values, it is a symmetric equilibrium of a first-price auction for each bidder i to bid b(xi ) = E [v (Y1 , Y1 , ...)|Y1 < xi ]. Winner curse The winner’s curse under interdependent (or common) values. Bidders must condition on winning. If they didn’t they would end up overbidding. This is because winning conveys the bad news that all other bidders have lower signals. This is easily seen in a second-price auction, where E [v (xi , Y1 , ...)|Y P 1 = xi ] < E [v (xi , Y1 , ...)]. For example, if v (xi , x−i ) = N j=1 xj and the xj are uniformly distributed, then E [v (xi , Y1 , ...)|Y1 = xi ] = 2xi + (N − 2)xi /2, while E [v (xi , Y1 , ...)] = xi + (N − 1)/2. In equilibrium rational bidders take into account that winning is bad news, and hence do not suffer from the winner’s curse. Proposition Under interdependent values, it is a symmetric equilibrium of an English auction for each bidder i with signal xi to follow the following strategy: If nobody has dropped out yet, drop out at price p(xi ) = v (xi , xi , ..., xi ). If one player has already dropped out at price pN−1 = v (xN−1 , xN−1 , ..., xN−1 ), then drop out at price p(xi , xN−1 ) = v (xi , xi , ..., xi , xN−1 ). ... If k players have already dropped out at prices pN−1 = v (xN , xN , ..., xN ), ..., pN−k = v (xN−k , xN−k , ..., xN−k , xN−k−1 , ..., xN−1 ), then drop out at price p = v (xi , xi , ..., xi , xN−k , ..., xN−1 ). P Thus, for example, suppose v (xi , x−i ) = N j=1 xj and the xj are uniformly distributed. Suppose also that x1 > x2 > x3 ..., then the price in a second price auction is 2x2 + (N − 2)x2 /2, while the price in an English auction is 2x2 + x3 + x4 + .... Note: the expected price in a second-price and in an English auction are the same! This revenue equivalence result is true in general. With independent signals, all efficient auctions are revenue equivalent. Mechanism Design Every auction can be seen as a mechanism in which bidders send messages to a designer (their bids) and the designer commits to an outcome decision (e.g., who gets the item) and transfer function (who pays what). A mechanism can be used for any general allocation, or social decision, problem. In general a mechanism can be seen as a triple hM, D, T i where: M = ×i={1,...,N} Mi and Mi is the set of messages that player i can send to the designer; D : M → A, the decision function, maps messages into the set of possible allocations A (in the case of an auction of a single object A could be the set of vectors of probabilities (π1 , ..., πN ), where πi is the probability that i wins the object); T : M → RN , the transfer function, maps messages into a transfer payment from each of the bidders to the designer. A direct mechanism is a mechanism in which the message space Mi of each player is his set of types (in the auction described before, the set [0, 1]). It is standard to omit the message space and denote a direct mechanism by the pair hD, T i. Revelation principle The following is a fundamental result. Theorem (The Revelation Principle) Given any mechanism and a (Bayesian) equilibrium for that mechanism, there exists a direct mechanism in which (1) it is an equilibrium for each buyer to report her true type truthfully, and (2) the equilibrium outcomes (decision and transfers) are the same as in the given equilibrium of the original mechanism. Proof (sketch) Instead of each buyer i reporting message (bid) b = b(xi ) a direct mechanisms asks her to report xi and makes sure the outcome is the same as in the original mechanism. Intuitively, a direct mechanism does the equilibrium calculations in the original mechanism directly for the bidder. If it was an equilibrium for type xi to bid according to b(xi ) in the original mechanism (rather than, say, bid b b = b(z)), then it will be an equilibrium in the direct mechanism to report the true type xi (rather than, say, z). Let < π, p > be a direct mechanism, where πi (xi , x−i ) is the probability that i wins an object and pi (xi , x−i ) is i’s payment. Q Let g (x−i ) = f (xj ). Bidder i’ s expected payoff when his type is j6=i xi , but he reports zi and all other bidders report truthfully is Z Ui (zi ; xi ) = x Z x x [v (xi , x−i )πi (zi , x−i ) − pi (zi , x−i )] g (x−i )dx−i ... x Letting Ui (xi ) be type xi ’s expected payoff in the truthful equilibrium of hπ, pi, and using a standard envelope argument yields Z x Z x ∂V (xi , x−i ) 0 ... πi (xi , x−i )g (x−i )dx−i Ui (xi ) = ∂xi x x Bidder-payoff equivalence Theorem. Bidders’ expected payoffs are the same in any mechanism hπ, pi having the same outcome function π and yielding the same payoff to the lowest type. Bidder i’ s expected payoff is given by Z xi Z x Z x ∂V (x, x−i ) πi (x, x−i )g (x−i )dx−i dx. ... Ui (xi ) = u + ∂x x x x In particular, if u = 0, the outcome function π is efficient (so that πi = 1 if xi > YKI −1 and πi = 0 if xi < YKI −1 ), and there are no informational externalities, then the above expression reduces to R x (I −1) (y )dy . Ui (xi ) = x i FK Revenue Equivalence Theorem As a simple consequence of bidder’s payoff equivalence, we have the revenue equivalence theorem. Theorem. The seller’s expected revenue is the same in any mechanism hπ, pi having the same outcome function π and yielding the same payoff to the lowest type. In particular, in all efficient auctions in which the lowest type expected payoff is zero (these include the first-price, second-price and English auction), the seller’s expected revenue is the same. A mechanism design application: Example 1. (A general version of a bargaining game). A buyer and a seller have private information about their valuations for an object, vB and vS . The (beliefs about the) valuations are distributed on [0, 1] according to the distribution functions FB (vB ) and FS (vS ) with positive densities fB (vB ) and fS (vS ).If trade does not take place, payoffs of both players are normalized to zero. If trade takes place at price p, then the buyer’s payoff is vB − p, while the seller’s payoff is p − vS . A mechanism specifies the probability of trade π and the payment p from the buyer to the seller (which could be negative) as functions of reports. A mechanism is efficient if trade takes place whenever vB > vS . A direct mechanism is Bayesian incentive compatible if all players telling the truth is a Bayesian equilibrium. A mechanism is individually rational if no player’s type makes a negative expected payoff. Theorem (Myerson and Satterthwaite) In the bargaining model, there is no efficient, Bayesian incentive compatible and individual rational mechanism.