Answers to Problem Set 5 Economics 703 Spring 2016 1. Let’s write down what a common prior, if it exists, must satisfy. So suppose p is a common prior. Then it must be true that the conditional probabilities it generates match the beliefs of the two types of the players. In particular, it must satisfy 1 p(a, c) = p1 (c | a) = p(a, c) + p(a, d) 4 p(b, c) 3 = p1 (c | b) = p(b, c) + p(b, d) 4 p(a, c) 1 = p2 (a | c) = p(a, c) + p(b, c) 4 3 p(a, d) = p2 (a | d) = p(a, d) + p(b, d) 4 We can rearrange these equations as 3p(a, c) = p(a, d) p(b, c) = 3p(b, d) 3p(a, c) = p(b, c) p(a, d) = 3p(b, d) The first three equations together imply 3p(a, c) = p(a, d) = p(b, c) = 3p(b, d). This plus the condition that the probabilities sum to 1 is enough to pin down the prior. Specifically, let x = p(a, d). Then we have p(b, c) = x and p(a, c) = p(b, d) = x/3. Hence the condition that the probabilities sum to 1 implies 2x + (2x/3) = 1 or 8x = 3, so x = 3/8. Hence p(a, d) = p(b, c) = 3/8 and p(a, c) = p(b, d) = 1/8. Note that these probabilities satisfy the fourth equation, so this gives a common prior. b) After seeing (a), you probably have a good idea why we’re not going to have a common prior in (b). In (a), the common prior was pinned down by the first three equations. Hence it was sheer luck (or clever construction of the problem) that it satisfied 1 the fourth. Here we won’t be so fortunate. To see this, suppose that there is a common prior p. Then it must satisfy p(a, c) 1 = p1 (c | a) = p(a, c) + p(a, d) 2 p(b, c) 1 = p1 (c | b) = p(b, c) + p(b, d) 2 p(a, c) 1 = p2 (a | c) = p(a, c) + p(b, c) 3 p(a, d) 2 = p2 (a | d) = p(a, d) + p(b, d) 3 We can rearrange these equations as p(a, c) = p(a, d) p(b, c) = p(b, d) 2p(a, c) = p(b, c) p(a, d) = 2p(b, d) The first three equations together imply 2p(a, c) = 2p(a, d) = p(b, d) = p(b, c). This plus the condition that the probabilities sum to 1 is enough to pin down the prior. Specifically, let x = p(b, d). Then we have p(b, c) = x and p(a, c) = p(a, d) = x/2. Hence the condition that the probabilities sum to 1 implies 3x = 1 or x = 1/3. Hence p(b, d) = p(b, c) = 1/3 and p(a, c) = p(a, d) = 1/6. However, plugging these numbers into the last equation gives 1/6 = 2(1/3) which is obviously false. Hence no common prior exists. 2. a) The simplest way to describe to do this one is to suppose that player 2 has two types, t1 and t2 , both of which have probability 1/2. He is type t1 when he knows the game is G1 and type t2 when he knows it is G2 . In other words, the utility function for player j given that player 2 is type ti is the one defined by the matrix Gi . Clearly, if when player 2 is type t1 , he has a strictly dominant strategy of b2 . Hence σ2 (t1 ) = b2 . Suppose that σ2 (t2 ) also equals b2 . Then player 1’s best strategy is a2 since he knows this will yield a payoff of 3. Since b2 is not the best reply to a2 for player 2 when his type is t2 , this is not an equilibrium. Hence if there is a pure strategy equilibrium, we must have σ2 (t2 ) = b1 . Since the two types are equally likely, it is not hard to show player 1’s best reply to this strategy is a3 . Since player 2’s best reply when his type is t2 is b1 , this is an equilibrium. Hence the unique pure strategy Bayes–Nash equilibrium is σ1 = a3 , σ2 (t1 ) = b2 , and σ2 (t2 ) = b1 . b) Now we have several types for player 1. The easiest way to describe the game is to i2 suppose that there are three possible types for player 1, say tu1 , ti1 1 , and t1 , and two for 2 1 i2 2 player 2, say t12 and t22 . The common prior is p(tu1 , t12 ) = p(tu1 , t12 ) = p(ti1 1 , t2 ) = p(t1 , t2 ) = i1 u 1/4. More intuitively, think of t1 as player 1’s type when he is uninformed, t1 as his type when he is informed that player 2 is his first type, and ti2 1 as his type when he is informed that player 2 is his second type. Just as in part (a), player 1 is of type t12 when he is informed that the game is G1 and type t22 when he is informed that the game is G2 . Again, we can define payoffs as a function of types and actions analogously to what we did in (a). You can use Bayes’ Rule to verify that if player 1 is told his type, he updates to 1/2 probability of t12 if his type is tu1 , updates to probability 1 on t12 if his type i2 1 is ti1 1 , and updates to probability 0 on t2 if his type is t1 . Similarly, player 2 always puts probability 1/2 on player 1 being informed and knows that if player 1 is uninformed, player 1 believes each game has probability 1/2. To find an equilibrium, again, note that σ2 (t12 ) must be b2 . As a result, we must have 2 σ1 (ti1 1 ) = a2 . Suppose σ2 (t2 ) = b2 . Then every type of player 1 chooses a2 . Since b2 is not then a best reply by t22 , this is not an equilibrium. So suppose σ2 (t22 ) = b1 . As in part (a), the best response for player 1 if his type is tu1 is a3 . If his type is ti2 2 , it is a1 . Hence this strategy for type t22 is optimal iff (1/2)(0) + (1/2)(2) ≥ (1/2)(1) + (1/2)(0) which is true. (Recall that if player 2 is type t22 , he puts probability 1/2 on player 1 being type tu1 and probability 1/2 on player 1 being type ti2 2 .) Therefore, this is the unique pure strategy Bayes–Nash equilibrium. 3. First, let’s look for separating equilibria. Try m(t1 ) = m1 and m(t2 ) = m2 . If this is 1’s strategy, 2’s strategy must be a(m1 ) = a2 and a(m2 ) = a1 . Given this strategy for 2, is 1’s strategy optimal? First consider type t1 . If he sends m1 , his payoff is −1, while if he sends m2 , his payoff is 0. Hence he prefers message m2 , so this is not an equilibrium. Now let’s try m(t1 ) = m2 and m(t2 ) = m1 . If this is 1’s strategy, 2’s best strategy is a(m1 ) = a1 and a(m2 ) = a2 . If this is 2’s strategy, what is optimal for 1? Type t1 gets 5 if he uses m1 and gets −1 if he uses m2 . Hence he prefers m1 so this is not an equilibrium. Hence there are no separating equilibria. Now let’s turn to pooling equilibria. First, suppose m(t1 ) = m(t2 ) = m1 . Then 2’s belief in response to m1 must be the same as the prior which is to put probability p < 1/2 on t1 . Hence his optimal action in response to m1 must be a1 . Suppose his response to m2 is also a1 . Then both types prefer m1 to m2 since m1 gives each type 5, while m2 would give each type 0. So if 2’s beliefs are that he puts probability p on t1 in response to either message, he’ll play a1 in response to either and 1’s optimal strategy will be to always send message m1 . Hence this is a pooling equilibrium. Could we have a pooling equilibrium with m(t1 ) = m(t2 ) = m1 , a(m1 ) = a1 , and a(m2 ) = a2 — that is, can we change 2’s reaction off the equilibrium path? Yes. This change lowers the incentives of player 1 to play m2 since this reaction gets him a payoff of −1. Hence this is another pooling equilibrium. 3 Can we have a pooling equilibrium where m(t1 ) = m(t2 ) = m2 ? The beliefs in response to this message would have to be the prior, so 2 would be putting probability p on t1 and would play a1 . If he also plays a1 in response to m1 , this is not an equilibrium since both types of player 1 would deviate to m1 . If he plays a2 in response to m1 , though, it is an equilibrium. Neither type would deviate to m1 since this gives a payoff of −1, while m2 gives a payoff of 0. Again, we can pick 2’s beliefs to make this sequentially rational for him. Summarizing, then, there are three pooling equilibria: one with m(t1 ) = m(t2 ) = m1 and a(m1 ) = a(m2 ) = a1 , one with m(t1 ) = m(t2 ) = m1 , a(m1 ) = a1 , and a(m2 ) = a2 , and one with m(t1 ) = m(t2 ) = m2 , a(m1 ) = a2 , and a(m2 ) = a1 . Some people would argue that this last equilibrium is not terribly intuitive. Since both types really hate message m2 , why are they pooling at that message? 4. Note first that there is no way to get type t1 to pick m2 — his payoff is −4, while at worst he gets 0 from m1 . Let’s start with separating. As noted above, we can’t have m(t1 ) = m2 , so the only possibility for separating is m(t1 ) = m1 and m(t2 ) = m2 . If this is 1’s strategy, 2’s best response to m1 is a1 since this gives 1 instead of 0. His best response to m2 is also a1 . So we’d have to have a(m1 ) = a(m2 ) = a1 . If this is 2’s strategy, then 1’s strategy is optimal: type t1 certainly wants to play m1 , while type t2 is indifferent and so is willing to play m2 . Now let’s try pooling. Again, we know we can’t have m(t1 ) = m2 , so the only possibility for pooling is m(t1 ) = m(t2 ) = m1 . If this is 1’s strategy, 2’s beliefs in response to m1 must be the same as the prior which puts probability p < 1/2 on t1 . Hence 2’s best action would be a2 . What should we make the reply to m2 be? If we make it a1 , type t2 will certainly deviate: he gets 1 from playing m1 and getting a response of a2 , while the deviation would then give him 3. So we must have a(m1 ) = a(m2 ) = a2 . If this is 2’s strategy, then it is optimal for both types of player 1 to choose m1 . Notice that the only beliefs that will get player 2 to respond to m2 with a2 are probability 1 on type t1 . Weak perfect Bayesian will certainly let us use these beliefs and even sequential will allow them (which you can verify). However, it seems very weird: type t1 is the type who would never use message m2 , but 2’s beliefs in this equilibrium have to put probability 1 on t1 when he sees m2 . 4