Static Games of Incomplete Information . Example 1 • Two firms: incumbent (player 1) & entrant (player 2) • Player 1 decides whether to build new plant; player 2 decides whether to enter • Player 1’s cost of building is 1.5 (wp 1-p1) or 3 (wp p1) 2 Enter 1 Don’t enter 2 Enter Don’t enter 1 Build 0, -1 2, 0 Build 1.5, -1 3.5, 0 Don’t build 2, 1 3, 0 Don’t build 2, 1 3, 0 1’s cost is high 1’s cost is low • How is this game to be played? • It depends on 2’s belief about 1’s probability of building when cost is low (x), and 1’s belief of 2’s probability of entering (y) Example 1 • Harsanyi’s Key Idea (1967-68): i. Player 1’s ‘type’ is determined by a prior move of nature ii. Transforms a game of incomplete info (2 doesn’t know 1’s cost) into one of imperfect info • Equilibrium: 1. Player 2 enters (y=1) if x<1/[2(1-p1], and stays out (y=0) if x>1/[2(1-p1], 2. Low cost player 1 builds (x=1) if y<1/2, and not (x=0) if y>1/2. Bayesian equilibrium • There are i єI players I • Players types { }i 1 are drawn from dist p(θ1, θ2,… θI,), where θi єΘi, Θi is finite and p(θ-i| θi) is i’s conditional probability about his rival’s types θ-i • Pure strategies for i are si єSi, and payoff is ui(s1, s2, …,sI, θI, θ2,…,θI,), • A Bayesian equilibrium for above game of incomplete information is a Nash equil of the ‘expanded game’ where each player i’s space of pure strategies is the S set i of maps from Θi to Si. Given strategy profile s(.), and s/i(.) є Si , s(.) is a Bayesian equilibrium if i i si (.) arg max p(i , i )ui ( si/ (i ), si ( i ), (i , i )) si/ (.)Sii i i Cournot competition with incomplete info • Let firm i’s profit be ui= qi(θi- qi- qj), where θi=α-ci, α being the intercept of the linear demand function • Let θ1=1, and θ2=3/4 w.p. ½ and 5/4 w.p. ½ • Denote q2 as q2H if θ2=3/4 , and as q2L if θ2=5/4 • Firm 2’s equil choice must satisfy: q2(θ2)=(θ2- q1)/2 • Firm 1 does not know 2’s type, so its expected payoff is: 1 1 H q1 arg max { q1 (1 q1 q2 ) q1 (1 q1 q2L )} 2 2 q1 • This gives q1=(2- q2L - q2H )/4 • Plugging in for q2(θ2) we get (q1=1/3, q2L =11/24, q2H = 5/24) • This is the Bayesian equilibrium War of attrition • Each player i chooses a number si in [0, +∞], • Payoffs are: si ui i s j if s j si if s j si • i’s type θi takes values in [0, +∞], with distribution function P and density p • We look for pure-strategy Bayesian equil (s1(.), s2(.)) • For each θi, si(θi ) must satisfy: si ( i ) arg max { si Pr( s j ( j ) si ) si ( i s j ( j )) p j ( j )d j } { j s j ( j ) si } War of attrition • A. Key step: Look for monotonic strategies that are strictly increasing and continuous in a player’s type 1. To show that: If θi// > θi/ implies si //> si / where si //=si (θi //) and si /=si (θi /) Proof: If θi/ prefers si (θi /) to si (θi //) then s i/ Pr( s j ( j ) si/ ) si/ Pr( s j ( j ) si/ ) j ( j ) p j ( j ) d j { j s j ( j ) si/ } i/ Pr( s j ( j ) si// ) si// Pr( s j ( j ) si// ) s j ( j ) p j ( j ) d j { j s j ( j ) si// } A similar inequality obtains when θi// prefers si (θi //) to si (θi /) Subtracting the two inequalities gives (i// i/ )[Pr( s j ( j ) si// ) Pr( s j ( j ) si/ )] 0 Since, θi// > θi/ it must be that si //> si / . War of attrition 2. To show that strategies, si (θi) and sj (θj), are strictly increasing Proof: If not, there would be an “atom” for j at s>0, i.e. Pr(sj (θj)=s)>0 Then i assigns probability 0 to [s-є, s] Then any type of j planning to play at s is better-off playing at s-є Thus, no atom at s after all 3. To show that si (θi) is continuous in θi Proof: Similar to above B. Let Φi be inverse function of si : Φ-1i (θi)= si War of Attrition C. Transforming variable of integration from θj to sj, si si (i ) arg max{si (1 Pj ( j ( si )) (i s j ) p j ( j ( s j )) /j ( s j )ds j } si 0 D. The FOC: If si ≡ si (θi), then i cannot benefit by playing si +dsi instead of si Cost is dsi if j plays above si +dsi which has probability 1-Pj(Φj(si +dsi )) Expected cost, to first order in dsi , is [1-Pj(Φj(si ))] dsi . Gain is θi= Φi(si) if j plays in [si, si +dsi ], i.e. if θi є [Φj(si ), Φj(si +dsi )] This has probability, pj(Φj(si ))Φ/j(si ) dsi . Equating costs and benefits, the FOC is Φi(si) pj(Φj(si ))Φ/j(si ) = 1-Pj(Φj(si )) War of Attrition • Impose symmetry P1= P2=P • Substitute θ= Φ(s), and use Φ/=1/s/ to get, p ( ) s ( ) 1 P( ) / • Integrating, xp( x) s( ) dx 0 1 P( x) • If P(.) is exponential, P(θ)=1-exp(-θ), then, S(θ)= θ2/2 Double auction • A seller and buyer trade a unit of good • Seller (player 1) has cost c and buyer (player 2) has valuation v . v, c є[0, 1] • Players simultaneously bid b1, b2 є[0, 1] • If b1≤ b2 , they trade at price t= (b1+b2)/2 • With trade 1 gets u1=(b1+b2)/2-c; 2 gets u2=v(b1+b2)/2. Without trade both get 0 • c distributed as P1 and v distributed as P2 • Let F1(.) and F2(.) be cumulative dist of b1 and b2 • Find the Bayes-Nash equilibrium (s1(.), s2(.)), where si(.):[0, 1]→[0, 1] Double auction • To show that: Bids increase in type. That is, c// > c/ implies b1//>b1/ where b1 //=s1(c//) and b1/=s1(c /) Proof: Optimization by the seller requires // 1 b b b1/ b2 / / 1 2 ( c ) dF ( b ) ( c )dF2 (b2 ) 2 2 b1/ 2 b1// 2 1 and / 1 b b b1// b2 // // 1 2 ( c ) dF ( b ) ( c )dF2 (b2 ) 2 2 b1// 2 b1/ 2 1 // / // / Combining these inequalities, (c c )[ F2 (b1 ) F2 (b1 )] 0 Since, c// > c/ it must be that b1//>b1/ . • The bids are strictly increasing and continuous in types • Similar things hold for the buyer Double auction • Maximization problem for type-c seller b1 b2 max ( c)dF2 (b2 ) b1 b1 2 1 • The FOC is: ½ [1-F2(s1(c))]-(s1(c)-c)f2(s1(c))=0 • For the buyer, b2 max (v b2 0 b1 b2 )dF1 (b1 ) 2 • And FOC is: (v – s2(v)f1(s2(v))= ½ F1(s2(v)) Double auction • Specific case: - P1 and P2 are uniform dist on [0, 1], and strategies are linear in types s1(c)= α1+β1c and s2(v)= α2+β2v - Then, Fi(b)=Pi(s-1i(b))= s-1i(b)=(b- αi)/βi, so fi(b)=1/ βi, - Plugging into FOCs, we get 2[α1+(β1-1)c]/ β2 = [β2 – (α1+β1c)+ α2]/β2 2[(1-β2)v –α2]/ β1 = [α2+β2v- α1]/β1 - Solving this system, β1= β2= 2/3; α1=1/4; α2=1/12 - In equilibrium parties trade only if, α2+β2v ≥ α1+β1c - Thus trade occurs only if, v ≥c+1/4 • Too little trading in equilibrium!! First price auction with a continuum of types • Two bidders with a unit of good to trade • Player i’s valuation is θi and belongs to [ , ] • Players have beliefs P, with density p, about rival’s valuation • Seller imposes reservation price s0> • Player i bids si; gets ui = θi - si if si > sj & ui =0, si < sj • If si = sj both get good w.p. ½ and ui = (θi - si )/2 • Let si (.) be the pure strategy of player i First price auction (continuum of types) A. Show that strategies are monotonic, strictly increasing, and continuous in type B. To show that: si ( ) s j ( ) s Proof: If si ( ) s j ( ) then type of player i could slightly lower his bid and still win w.p. 1 C. Let Φi be inverse function of si (.): Φ-1i (θi)= s on [s0, s ], i.e. player i bids s if his valuation is Φi(s) D. Type θi maximizes (θi - s)P(Φj(s)) over s E. This gives, P(Φj(s)) = [Φi(s)-s]p(Φj(s)) Φj/(s) F. There is a similar FOC by switching i and j First price auction (continuum of types) D. To show that: There cannot be an asymmetric solution, Φ1(s) ≠ Φ2(s) for all s E. Using Φ1 = Φ2 = Φ, in FOC, and integrating, we get, s dx ln[ P(( s))] ( x) x s F. This will give Φ(.), and the inverse function gives s(.) First price auction (with two types) • Each bidder can have types , with < • Corresponding probabilities are p and p • Seller’s reservation bid is lower than • Key idea: Look for mixed-strategy equilibrium • Type bids , and type randomizes according to a continuous distribution F(s) on [ s, s ] • Argue that: s • For i of type to play a mixed strategy with support [ s , s ] it must be that s [ s, s], ( s)[ p pF ( s)] constant First price auction (with two types) • Because F( )=0, the constant is ( ) p • Thus, F(.) is given by ( s)[ p pF (s)] ( ) p • Let G(s)≡ p pF (s) be distribution of bids. Above can be written as ( s)G(s) ( ) p • Since F( s )=1, implies s p p • Each bidders net utility is 0 when his type is and p( ) when his type is Bayesian equil can justify mixed equil • Harsanyi, 1973: A mixed strategy equil of a complete info game can be interpreted as the limit of pure strategy equil of perturbed games of incomplete info • Example: “Grab the Dollar”- complete info version - At times t=0, 1, 2…, two players want to grab a $1 - If only one grabs, he gets 1 and other gets 0 - If both grab at once, dollar destroyed & each gets -1 - If neither grabs, both get 0 - Players have a common discount factor δ - The only symmetric strategy is a mixed strategy, where both grab w.p. p*=1/2 in each period Bayesian equil can justify mixed equil • Example: “Grab the dollar”- complete info version -Consider player 1. By grabbing at time t, he gets δt(1- p*)+ δt p*(-1). By not grabbing, he gets 0. He is indifferent, so δt(1- p*)+ δt p*(-1)=0, and so, p*=1/2 • Consider ‘perturbed’ version of the above game • Example: “Grab the dollar”- incomplete info version -If player i wins, he gets 1+θi, θi is uniform on [-є, є] -Consider symmetric strategy: “si(θi<0)=do not grab; si(θi≥0)=grab” -This is a pure strategy Bayesian equilibrium! Why? -What happens when є→0? Bayesian & mixed equil: 1st price auctions 1. Consider FOC for first-price auctions with continuum of types: P(Φj(s)) = [Φi(s)-s]p(Φj(s))Φj/(s) Let Gj(s)= P(Φj(s)) be dist of bids s. gj(s)=p(Φj(s))Φj/(s) Then FOC becomes: Gj(s)= [Φi(s)-s] gj(s), s> 2. Equivalent condition for two-type case: ( s)G(s) ( ) p Differentiating w.r.t s, G ( s) ( s) g ( s) 0, n P ( ) p, Consider sequence, Pn(θ), s.t. lim n 1, [ , ) n ( s) , s If Φn(.) is equil strategy for Pn(.), then lim n