# Problem set 7: Equilibria in extensive-form games

**1. (a) By backward induction, in period 2:1
**

∗ q2 ∈ arg max [q2 (1 − q1 − q2 )] q2 ≥0 1−q1 2

⇒ And thus in period 1:

∗ q2 (q1 )

=

0

if q1 ≤ 1 otherwise.

∗ ∗ q1 ∈ arg max [q1 (1 − q1 − q2 (q1 ))] q1 ≥0 ∗ ⇒ q1 =

1 2 1 ∗ ⇒ q2 = . 4

1 8 ∗ and π2 = 1 . 16

∗ This path yields π1 =

We also need to check the other paths since

the FOC for player 2 might yield negative payoﬀ on the other paths.2 For each q1 (each player 2’s history), if ﬁrm 2 chooses the quantity derived by the FOC, its payoﬀ is 1 − q1 2 1 − q1 − 1 − q1 2 = (1 − q1 )2 ≥ 0. 4

**Thus, both players have no incentive to deviate from the following strategy proﬁle:
**

∗ q1

1 ∗ = , q2 (q1 ) = 2

1−q1 2

0

if q1 ≤ 1 otherwise

,

**which is SPE. (b) We can write ﬁrm 2’s (discontinuous) proﬁt function:3 π2 (q1 , q2 ) =
**

1

q2 (1 − q1 − q2 ) − f 0

if q2 > 0 if q2 = 0

We assume q1 , q2 ≥ 0. If there is no restriction, the question will be easier. Of course, in this question, we do not need to worry about this by the form of the payoﬀ function, unlike part (b). 3 Here, we assume that q2 = 0 means “staying out.”

2

1

**Since FOC for q2 > 0 does not change, if ﬁrm 1 chooses q1 = chooses q2 =
**

1 4

1 2

as before, ﬁrm 2

and thus π2 =

1 16

− f . Firm 1 earns π = 1 . 8

However, ﬁrm 1 may prefer to deter entry, that is, we need to check the other paths as in part (a). For each q1 (each player 2’s history), if ﬁrm 2 chooses the quantity derived by the FOC, its payoﬀ is

∗ ∗ q2 (q1 )(1 − q1 − q2 (q1 )) − f =

(1 − q1 )2 − f. 4

If

(1−q1 )2 4

√ √ ≤ f , that is, 1 − 2 f ≤ q1 ≤ 1 + 2 f , ﬁrm 2 weakly prefers inaction. ≤ f , we can always ﬁnd such q1 in this case. Since √ < 1 − 2 f ≤ q1 . When q2 = 0, ﬁrm 1’s payoﬀ is globally

(1−q1 )2 4 (1−q1 )2 4

**Note that, since 0 ≤ f <
**

1 , 16

we have

1 2

maximized at q1 = 1 . This argument implies that if 2 √ action is q1 = 1 − 2 f ≤ 1, which yields

≤ f , ﬁrm 1’s optimal

**π1 = (1 − 2 f ) 1 − (1 − 2 f ) = 2 f (1 − 2 √ √ 1 2− 2 2+ 2 ≥ if ≤ f≤ . 8 8 8 If f is suﬃciently close to SPE is as follows:
**

∗ q1 1 , 16

f)

√ the above inequality is satisﬁed. Since 1 + 2 f ≥ 1,

=

√ 1−2 f

1 2

∗ q2 (q1 ) =

0

1−q1 2

√ ≤ f (< √ > f, √ if q1 ≥ 1 − 2 f √ if q1 < 1 − 2 f . if if

√ 2− 2 8 √ 2− 2 8

1 4

<

√ 2+ 2 ) 8

Intuitively, if the ﬁxed cost is high, the ﬁrm 1’s cost of deterring entry is low and so it is worthwhile for ﬁrm 1 to do so.4

√ When ﬁrm 1 chooses 1 − 2 f , it would hope that ﬁrm 2 chooses 0 but it is not strictly guaranteed since 1−q1 is also optimal for ﬁrm 2 and this is a strictly worse choice for ﬁrm 1 if f > 0. One might think about 2 √ slightly larger quantity of ﬁrm 1 than 1 − 2 f to guarantee the better choice, that is, q2 = 0, but, in this case, we can always ﬁnd a better action, which means this cannot constitute SPE.

4

2

2. First, it is straightforward to check that the strategy pair deﬁned in the question is a subgame perfect equilibrium as in Proposition 122.1. Actually, we can show this game satisﬁes the one deviation property (Exercise 123.1 in MOAR). In the proof of Lemma 98.2 in MOAR, we suppose one proﬁtable deviant strategy si of player i in ﬁnite subgame Γ(h ) but we can apply this logic to this game since any inﬁnite subgame in this game yields the worst outcome. This implies that we can ignore any inﬁnite path when you consider some proﬁtable deviation. Second, we show the uniqueness. Step 1. M1 (G1 ) = m1 (G1 ) = 1 and M2 (G2 ) = m2 (G2 ) = c1 . Let Mi (Gi ) and mi (Gi ) be as in the proof of Proposition 122.1 in MOAR for i = 1, 2. By the argument for (124.1) in MOAR with the roles of the players reversed we have M2 (G2 ) ≤ 1 − m1 (G1 ) + c1 , or m1 (G1 ) ≤ 1 − M2 (G2 ) + c1 . Now suppose that M2 (G2 ) ≥ c2 . Then by the argument for (123.2) in MOAR with the roles of the players reversed we have m1 (G1 ) ≥ 1 − M2 (G2 ) + c2 , a contradiction (since c1 < c2 ). Thus M2 (G2 ) < c2 . But now the argument for (123.2) implies that m1 (G1 ) ≥ 1, so that m1 (G1 ) = 1 and hence M1 (G1 ) = 1. Since (124.1) implies that M2 (G2 ) ≤ 1 − m1 (G1 ) + c1 we have M2 (G2 ) ≤ c1 ; by (123.2) we have m2 (G2 ) ≥ c1 , so that M2 (G2 ) = m2 (G2 ) = c1 . The remainder of the argument follows as in the proof of Proposition 122.1 in MOAR: Step 2. In every SPE of G1 player 1’s initial proposal is 1, which player 2 immediately accepts. Step 3. In every SPE of G1 player 2’s strategy accepts any proposal. 3. Suppose that player 2 adheres to tit-for-tat. Consider player 1’s behavior in subgames following histories that end in each of the following outcomes. (C, C) If player 1 adheres to tit-for-tat the outcome is (C, C) in every period, so that 3

her discounted average payoﬀ in the subgame is x. If she chooses D in the ﬁrst period of the subgame, then adheres to tit-for-tat, the outcome alternates between (D, C) and (C, D), and her discounted average payoﬀ is y/(1 + δ). Thus we need x ≥ y/(1 + δ), or δ ≥ (y − x)/x, for a one-period deviation from tit-for-tat not to be proﬁtable for player 1. (C, D) If player 1 adheres to tit-for-tat the outcome alternates between (D, C) and (C, D), so that her discounted average payoﬀ is y/(1 + δ). If she deviates to C in the ﬁrst period of the subgame, then adheres to tit-for-tat, the outcome is (C, C) in every period, and her discounted average payoﬀ is x. Thus we need y/(1 + δ) ≥ x, or δ ≤ (y − x)/x, for a one-period deviation from tit-for-tat not to be proﬁtable for player 1. (D, C) If player 1 adheres to tit-for-tat the outcome alternates between (C, D) and (D, C), so that her discounted average payoﬀ is δy/(1 + δ). If she deviates to D in the ﬁrst period of the subgame, then adheres to tit-for-tat, the outcome is (D, D) in every period, and her discounted average payoﬀ is 1. Thus we need δy/(1 + δ) ≥ 1 for a one-period deviation from tit-for-tat not to be proﬁtable for player 1. (D, D) If player 1 adheres to tit-for-tat the outcome is (D, D) in every period, so that her discounted average payoﬀ is 1. If she deviates to C in the ﬁrst period of the subgame, then adheres to tit-for-tat, the outcome alternates between (C, D) and (D, C), and her discounted average payoﬀ is δy/(1 + δ). Thus we need 1 ≥ δy/(1+δ) for a one-period deviation from tit-for-tat not to be proﬁtable for player 1. The same arguments apply to deviations by player 2, so we conclude that (tit-fortat, tit-for-tat) is a subgame perfect equilibrium if and only if δ = (y − x)/x and 4

δy/(1 + δ) = 1, or y − x = 1 and δ = 1/x. 4. First we will prove that there are no separating equilibria. I will call player 1 the ‘sender’ and player 2 the ‘receiver.’ Let the receiver’s beliefs be q ≡ Pr(t1 chooses m1 ) and r ≡ Pr(t1 chooses m2 ). Separating equilibria: In a separating equilibrium all types must choose to send diﬀerent messages, so there are two cases for us to check. First, let’s try this: m(t1 ) = m1 , m(t2 ) = m2 . If the sender plays this strategy, the best response for the receiver is this: a(m1 ) = a2 , a(m2 ) = a1 since q = 1 and r = 0 by Bayes’ rule. However, in that case the sender of type t1 prefers to send m2 rather than m1 , since then he gets a payoﬀ 0 rather than −1. This cannot be an equilibrium. Next let’s try this: m(t1 ) = m2 , m(t2 ) = m1 . The best response for the receiver is then: a(m1 ) = a1 , a(m2 ) = a2 since q = 0 and r = 1. However, then sender type t1 prefers to send m1 rather than m2 , earning 5 rather than −1. This cannot be an equilibrium. There are thus no separating equilibria. Pooling equilibria: In a pooling equilibrium all types choose to send the same message. There are three pooling equilibria here. First, let’s try this: m(t1 ) = m(t2 ) = m1 . 5

The receiver’s belief must put probability p < her best response is:

1 2

on type t1 after observing m1 . Thus

a(m1 ) = a1 . We also need to specify a(m2 ). First: a(m2 ) = a1 . Then both types of sender are happy to send m1 since this earns a payoﬀ of 5 rather than 0. If the receiver puts probability r ≤

1 2

on t1 , a1 is also a best response to m2 ,

and so with these supporting beliefs there is a ﬁrst pooling equilibrium at: m(t1 ) = m(t2 ) = m1 a(m1 ) = a(m2 ) = a1 1 1 q =p < , r ≤ . 2 2 There is a second pooling equilibrium at: m(t1 ) = m(t2 ) = m1 a(m1 ) = a1 , a(m2 ) = a2 1 1 q =p < , r ≥ . 2 2 This is similar to the ﬁrst except that the receiver responds to the oﬀ-equilibrium message m2 with a2 ; the only change is that the sender’s incentive to deviate to m2 is even lower than under a(m2 ) = a1 . The third pooling equilibrium features: m(t1 ) = m(t2 ) = m2 . In that case, the receiver’s beliefs must again be equal to the prior p and so her best response is a(m2 ) = a1 . 6

If the receiver plays a(m1 ) = a1 this cannot be an equilibrium, since then both types of sender would deviate to m1 . However, if the receiver plays a(m1 ) = a2 , neither type of sender would like to deviate to m1 . The third pooling equilibrium is then at: m(t1 ) = m(t2 ) = m2 a(m1 ) = a2 , a(m2 ) = a1 1 1 q≥ , r=p< . 2 2 5. Let the receiver’s beliefs be q ≡ Pr(t1 chooses m1 ) and r ≡ Pr(t1 chooses m2 ). First notice we get one thing for free: m2 is strictly dominated for the sender t1 , so m(t1 ) = m1 . Separating equilibria: Given this, the only possible separating equilibrium is at: m(t1 ) = m1 , m(t2 ) = m2 . The receiver’s best response to this is: a(m1 ) = a1 , a(m2 ) = a1 since q = 1 and r = 0. Given this, sender type t1 indeed prefers to send m1 as we saw above; send type t2 is indiﬀerent between sending m1 and sending m2 , since he earns 3 either way. He is thus willing to send m2 , and so there is indeed a separating equilibria with the preceding strategies. Pooling equilibria: Again, the only possible pooling equilibria is: m(t1 ) = m(t2 ) = m1 . The receiver’s belief in response to m1 must then be q = p < response is: a(m1 ) = a2 . 7

1 , 2

and so her best

Next see that a(m2 ) cannot be a1 in equilibrium, since then sender type t2 will deviate to m2 and earn 3 instead of 1. What about a(m2 ) = a2 ? For this to be part of an equilibrium, the receiver must put probability 1 on the oﬀ-equilibrium message m2 having come from player 1 - this is perfectly OK in weak perfect Bayesian equilibrium since this node is reached with probability zero, meaning that Bayes’ Rule is not applicable and the receiver can hold any beliefs. Given a(m2 ) = a2 , it is indeed optimal for both types of sender to choose m1 . There is indeed then a pooling equilibrium at: m(t1 ) = m(t2 ) = m1 a(m1 ) = a(m2 ) = a2 1 q =p < , r = 1. 2 This one does not survive Cho-Kreps intuitive criterion. As in the lecture note, let’s consider some oﬀ-equilibrium path, that is, the receiver’s information set given by m2 . Since m2 is strictly dominated for type t1 , type 2 can convince the receiver to believe that m2 is sent by t2 . Hence, if the receiver observed m2 , he would have r = 0 and choose a1 . However, this means type t2 would never choose m1 since a(m1 ) = a2 . 6. Let the buyer’s beliefs be λ ≡ Pr(B chooses b) and µ ≡ Pr(B chooses g and the government chooses n). i. Buyer must play ‘not purchase’ whenever a challenge is observed, since this reveals that the product is bad. ii. The government must thus challenge whenever they are called on to decide, since they get 1 instead of 0. iii. Then the seller must send message b when the product is bad B.

8

iv. Suppose the seller send message b for both states, so that he always sends b. At the information set for the buyer where b is observed, the buyer must put probability λ =

3 4

on the seller’s type being B. The buyer will therefore choose not to

purchase, since purchasing has an expected payoﬀ of − 1 . This is sequentially 2 rational for the seller when his type is B, since if he sends message g instead he would get −1 (by point ii.). It is sequentially rational for the type G seller only if the buyer would also choose not to purchase if he saw message g. This information set is never reached on the equilibrium path, so this can indeed form a weak PBE. The equilibrium is: ∗ Seller: play b when type B, play b when type G. ∗ Buyer: play ‘not purchase’ when observing b, when observing g and a challenge, and when observing g and no challenge. ∗ Government: challenge when called on to act. ∗ Beliefs: buyer puts probability λ =

3 4

and µ ≥ 1 . 2

v. Now suppose the seller plays g when of type G. In this case the buyer must put probability 1 on G when observing g and probability 1 on B when observing b. From there it is simple to see that there is another weak PBE at: ∗ Seller: play b when type B, play g when type G. ∗ Buyer: play ‘not purchase’ when observing b and when observing g and a challenge. Play ‘purchase’ when observing g and no challenge. ∗ Government: challenge when called on to act. ∗ Beliefs: buyer puts probability λ = 1 and µ = 0. vi. The second of these weak PBE must also be a sequential equilibrium since all information sets are reached in equilibrium. vii. The ﬁrst of the weak PBE is also a sequential equilibrium. Pick the totally mixed 9

strategies as follows. Given µ ∈ [ 1 , 1), let the seller’s probability of a mistake 2 when type B (i.e., playing g) be type G (i.e., playing g) be mistake (playing n) be

1 n 3−3µ 2n2 µ 2n

∈ (0, 1) and his probability of a mistake when

∈ (0, 1). Let the government’s probability of a

∈ (0, 1). Then we have µn =

3 4 3 4 µ · 2n2 +1· 4

·

µ 2n2

3−3µ 2n2

= µ,

**which converges to any µ ∈ [ 1 , 1) as n → ∞. 2 λn =
**

3 (1 4 µ 3 (1 − 2n ) 4 , µ ) + 1 (1 − 3−3µ ) 2 2n 4 2n

−

**which converges to 3/4 as n → ∞. For µ = 1, let those probabilities of the mistakes be
**

1 n

∈ (0, 1),

1 n3

∈ (0, 1), and µn =

3 4n2 3 4n2

1 n

∈ (0, 1), respectively. Then we have = 3n , 3n + 1

+

1 4n3

**which converges to 1 as n → ∞ and λn =
**

1 3 (1 − n ) 4 1 ) + 1 (1 n 4 1 , ) n3

3 (1 4

−

−

which converges to 3/4 as n → ∞. Thus, since the assessment is consistent, the ﬁrst of the weak PBE is a sequential equilibrium. 7. (a) Let player 2’s belief be π ≡ Pr(player 1 chooses a) and player 3’s beliefs be µ ≡ Pr(player 1 chooses b and player 2 chooses d) and λ ≡ Pr(player 1 chooses b and player 2 chooses c). The normal form is c d a 2, 2, 2 4, 3, 1 b 3, 0, 1 3, 1, 0 e c d a 2, 2, 2 0, 0, 0 b 0, 0, 0 1, 0, 1 f

The pure strategy Nash equilibria are (a, c, f ), (a, d, e), and (b, d, f ). Since the only subgame is the game itself, all are subgame perfect equilibria. Note that 10

(a, d, e) and (b, d, f ) reach every information set, so the weak perfect Bayesian equilibria are as follows: (a, d, e, π = 1, µ = 0, λ = 0) and (b, d, f, π = 0, µ = 1, λ = 0). So consider the equilibrium (a, c, f ). (a, c, f, π = 1, µ ≥

1 ) 2

is a weak perfect

Bayesian equilibrium. To see this, note that player 2’s belief must put probability 1 on the node where player 1 plays a (π = 1) and player 3’s belief is not pinned down by the strategies. By the sequential rationality, you could choose µ ≥ make f optimal for player 3. Turning to sequential equilibrium, we know that every sequential equilibrium is a weak perfect Bayesian equilibrium. Hence we only need to check the three weak perfect Bayesian equilibria: (a, d, e, π = 1, µ = 0, λ = 0), (b, d, f, π = 0, µ = 1, λ =

1 0), and (a, c, f, π = 1, µ ≥ 2 ). The ﬁrst two reach every information set, so they 1 2

to

must be sequential. To see this, simply note that any totally mixed strategies that converge to these must generate beliefs converging to the unique weakly consistent beliefs. Hence both satisfy consistency, not just weak consistency. However, (a, c, f, π = 1, µ ≥ 1 ) is not a sequential equilibrium. Clearly, the action 2

1 f is sequentially rational if and only if µ ≥ 2 . Is this consistent? Suppose it is.

Let pn be player 1’s probability on a and qn be player 2’s probability on c in our sequence of totally mixed strategies converging to this equilibrium. We must have pn → 1 and qn → 1. Let µn be the probability put on the node in question by Bayes’ Rule with these strategies. Then µn = (1 − pn )(1 − qn ) (1 − pn )(1 − qn ) + qn (1 − pn ) + pn (1 − qn ) (1 − pn )(1 − qn ) < = 1 − pn . (1 − pn )(1 − qn ) + pn (1 − qn )

Since pn → 1, we must have µn → 0. Hence we cannot converge to these beliefs, so 11

the beliefs are not consistent and the equilibrium is not a sequential equilibrium. (b) Let player 2’s belief be π ≡ Pr(player 1 chooses a) and player 3’s beliefs be µ ≡ Pr(player 1 chooses a and player 2 chooses d). The normal form is c d a 1, 2, 0 −1, 1, 0 b 2, 1, 3 0, 3, 2 e c d a 1, 2, 0 0, 3, 1 b 2, 1, 3 1, −1, 0 f

Note that 1 has a strictly dominant strategy of b. Hence he must play this in every Nash equilibrium. There are two Nash equilibria in pure strategies: (b, d, e) and (b, c, f ). Because there is only one subgame, all are subgame perfect equilibria. It is easy to see that (b, d, e, π = 0, µ = 0) is a weak perfect Bayesian equilibrium. We can also show that (b, c, f, π = 0, µ ≥ equilibrium by the sequential rationality. Are the weak PBE’s sequential equilibria? Since (b, d, e, π = 0, µ = 0) reaches

2 every information set, it must be sequential. Is (b, c, f, π = 0, µ ≥ 3 )? No – there 2 are no totally mixed strategies and beliefs converging to (b, c, f, π = 0, µ ≥ 3 ). To 2 ) 3

is also a weak perfect Bayesian

see this, let pn be the probability player 1 makes a mistake (plays a) and let qn be the probability player 2 makes a mistake in this equilibrium (plays d). Then by Bayes’ rule, the probability µ is pn qn pn q n = = pn . pn qn + (1 − pn )qn qn But this must go to zero as n → ∞. Hence player 3 must put probability zero on the left-hand node, meaning that f cannot be sequentially rational given any consistent belief. Therefore, this equilibrium is not a sequential equilibrium.

12