Game Theory Sequential Equilibria Ruitian Lang ANU October 11, 2021 Ruitian Lang (ANU) Game Theory October 11, 2021 1 / 38 Overview Consider a player’s decision making at a non-singleton information set under some conjecture about the opponents’ strategies (like they playing the equilibrium strategies). In order to evaluate his options, the player needs to assign probabilities to the histories contained in the information set. In SPE, we do not consider decision making at these information sets specifically. In PPE, by considering public strategies, such probabilities are irrelevant. Ruitian Lang (ANU) Game Theory October 11, 2021 2 / 38 Table of contents 1 Conditional Probabilities 2 Definition and the First Example 3 Some More Examples Ruitian Lang (ANU) Game Theory October 11, 2021 3 / 38 Terminology of probability theory We usually have a big set Ω containing all possible outcomes. In extensive form games, that set is usually the set of all histories or the set of histories in a particular information set. A probability measure P assigns a non-negative number to each (measurable) subset of outcomes, such that P(Ω) = 1 and for any countably many disjoint subsets A1 , A2 , ..., of Ω, ∞ ∞ [ X P Aj = P(Aj ). j=1 j=1 The above equation implies the requirement that the infinite series on the right hand side converges. We only need the case for finitely many Aj ’s. Ruitian Lang (ANU) Game Theory October 11, 2021 4 / 38 Terminology of probability theory In talking about probability measures on a finite set I, it is convenient to represent a measure P by a probability P mass function p : I → R such that p(j) ≥ 0 for every j ∈ I and P j∈I p(j) = 1. The probability measure of a J ⊂ I is then P(J) = j∈J p(j). However, we often misuse notation and write P({j}) = p(j) as “P(j)” and talk about the probability of “j” instead of that of {j}. In game theory, we often need to talk about different probability measures on the same set, like the different probability measures on Y depending on the action profile in public monitoring games. Ruitian Lang (ANU) Game Theory October 11, 2021 5 / 38 Some motivation I Focus on Player i and fix the opponents’ strategies. Imagine that the opponents’ (including Nature’s) play lands Player i in one of the following disjoint information sets: I1 , ..., Im . At Ij , Player i chooses an action aj ∈ Aj and there is no future decision for Player i to make. Let P be the probability measure on I1 ∪ I2 ∪ ... ∪ Im so that P(A) is the probability that a subset A of histories is reached based on the opponents’ strategies. For each j, Player i needs a probability measure Pj on Ij to evaluate his options at the information set. Ruitian Lang (ANU) Game Theory October 11, 2021 6 / 38 Some motivation II For a history h ∈ Ij and action a ∈ aj , let v(a, h) be Player i’s expected payoff from playing a (assuming that the opponents play according to their strategy profiles in the future). Then P Player i should choose aj ∈ Aj to maximize h∈Ij v(aj , h)Pj ({h}). After an aj is chosen for each j,Pwe hope that Player i’s expected m P payoff from the entire game is j=1 h∈Ij v(aj , h)Pj ({h})P(Ij ). However, Pm P Player i’s expected payoff from playing this strategy is j=1 h∈Ij v(aj , h)P({h}) by definition. Ruitian Lang (ANU) Game Theory October 11, 2021 7 / 38 Some motivation III The correct way of specifying Pj for each information set Ij should be independent of the payoff function. Therefore, we require that Pj ({h})P(Ij ) = P({h}) for every h ∈ Ij . To accommodate the case where Ij is infinite, we generalize the requirement to the condition that for every (measurable) subset A of Ij , Pj (A)P(Ij ) = P(A). When Ij is reached with positive probability, the condition determines Pj : Pj (A) = P(A)/P(Ij ) for every (measurable) A ⊂ Ij . When Ij is reached with zero probability, our entire derivation (and the entire probability theory) says nothing. Ruitian Lang (ANU) Game Theory October 11, 2021 8 / 38 Conditional probabilities Definition Let P be a probability measure on Ω and A and B be (measurable) subsets of Ω such that P(B) > 0. The probability of A conditional on B is defined as P(A|B) = P(A ∩ B)/P(B). Theorem (Bayes) Let P be a probability measure on Ω, A, B1 , ..., Bm be (measurable) subsets of Ω such that (a) Bi ∩ Bj = ∅ for i 6= j; (b) Sm B = Ω; (c) P(A) > 0; and (d) P(Bj ) > 0 for j = 1, ..., m. Then j=1 j P(A|Bj )P(Bj ) P(Bj |A) = Pm , for j = 1, ..., m. k=1 P(A|Bk )P(Bk ) Ruitian Lang (ANU) Game Theory October 11, 2021 9 / 38 Assessments Definition In an extensive form game, an assessment of Player i assigns to each information I of Player i a mix of actions σI on AI (where AI is the set of actions available at I) and a probability measure PI on I. Definition Fix an extensive form game that is continuity at infinity and an assessment profile. Fix a Player i and one of his information sets I, where his set of available actions is denoted by AI . For each a ∈ AI and h ∈ I, denote by vi (a, h) his expected payoff when every player (including i himself) follows the strategy profile after the history (h, a). Player i’s assessment is called sequentially rational if for each of his information set I, each action a in the P support of σI maximizes EI [vi (a, h)] = h∈I vi (a, h)PI (h). Ruitian Lang (ANU) Game Theory October 11, 2021 10 / 38 Table of contents 1 Conditional Probabilities 2 Definition and the First Example 3 Some More Examples Ruitian Lang (ANU) Game Theory October 11, 2021 11 / 38 Equilibrium refinements For an information set I which is reached with positive probability in equilibrium, the probability measure PI is given by conditional probabilities. Probabilists do not care about information sets I that are not reached with positive probabilities as decisions made there do not affect expected payoffs. Economists do care about such information sets as decisions made there affect other players’ decision making. Different proposals have been made about restrictions on PI when P(I) = 0. Each proposal leads to a different solution concept. We focus on a standard one called “sequential equilibrium” and will briefly mention another (specifically for signaling games). Ruitian Lang (ANU) Game Theory October 11, 2021 12 / 38 The scandal game A journalist learns a scandal of a celebrity and demands a payment from the celebrity to hide that scandal. 1 2 3 The reader decides whether to read the newspaper. Without observing whether the public will read the newspaper, the celebrity decides whether to pay the journalist. Without observing the public’s action but after observing whether the celebrity pays, the journalist chooses whether to publicize the scandal. Ruitian Lang (ANU) Game Theory October 11, 2021 13 / 38 The scandal game (cont.) Payoffs: Reader 0 if not reading; 1 if reading a scandal, and -1 if reading a newspaper with no scandal. Celebrity 0 if scandal not publicized, -5 if scandal publicized without being read, -10 if scandal publicized and read. Suffers additional 2 if making payment. Journalist 0 if hiding the scandal, 1 if scandal publicized and read, -5 if scandal publicized without being read. Gains 2 if payment is made, but suffers -5 if publicizing scandal after payment (through bad reputation or the celebrity’s retaliation). Ruitian Lang (ANU) Game Theory October 11, 2021 14 / 38 The scandal game: surprising move Consider the following strategy profile: the reader does not read, the celebrity pays, and the journalist publicizes if and only if no payment is made. The reader’s and the celebrity’s moves are obviously sequentially rational. The question is whether the journalist’s move without payment is sequentially rational. That movement can be justified by the belief that the reader will read the newspaper. Is that belief reasonable? Ruitian Lang (ANU) Game Theory October 11, 2021 15 / 38 Consistency Definition A behavior strategy is called fully mixed if its support at every information set is the entire set of the available actions at that information set. Definition An assessment of Player i which assigns to each information set I of Player i the probability measure PI is called consistent with a strategy profile s if there exists a sequence {s (k) }∞ k=1 of fully mixed strategy profiles converging (k) to s such that the corresponding conditional probabilities PI converges to PI as k → ∞, for every information set I of Player i. Ruitian Lang (ANU) Game Theory October 11, 2021 16 / 38 Consistency (cont.) In case every information set is finite: A fully mixed behavior strategy assigns a positive probability to every action. For every history h ∈ I, PI ({h}) → PI ({h}) as k → ∞. (k) Definition An assessment profile of an extensive form game (continuous at infinity) is called a sequential equilibrium if it is consistent and sequentially rational. Ruitian Lang (ANU) Game Theory October 11, 2021 17 / 38 Sequential equilibria It can be shown that every finite extensive form game has a sequential equilibrium; actually, each of them has a more refined solution, a (trembling-hand) perfect equilibrium. There is no mechanical procedure of finding all sequential equilibrium. The proof of the existence theorem involves two non-constructive steps. To check that a given assessment profile is a sequential equilibrium, we first check consistency by constructing a suitable sequence of fully mixed strategy profiles and then check sequential rationality. Ruitian Lang (ANU) Game Theory October 11, 2021 18 / 38 Perfect Bayesian Equilibria Many textbooks use a solution concept called “Perfect Bayesian Equilibria” (PBE), which I do not recommend you to use. There is no unified definition of PBE: different textbooks use different definitions. One version is to define PBE as a synonym of sequential equilibria. Therefore, if you give a presentation and somebody asks you what you mean by a sequential equilibrium, just says “PBE”. Ruitian Lang (ANU) Game Theory October 11, 2021 19 / 38 The certificate game There is a seller of a used car and a buyer. The car may be roadworthy or not roadworthy; the seller knows this but the buyer does not. The cost of a roadworthy inspection is c > 0, the value of a roadworthy car is v > c and the value of a non-roadworthy car is zero. The seller may sell a roadworthy car with a certificate (by ordering the inspection himself) or sell a car without a certificate. Either way, the seller makes a take-it-or-leave-it offer. A buyer of a car without certificate needs to order the inspection. The seller’s payoff is payment (if any) minus cost of inspection (if any). The buyer’s payoff is the value of the car (if purchased) minus payment (if any) minus cost of inspection (if any). Ruitian Lang (ANU) Game Theory October 11, 2021 20 / 38 Persuasion models There is a seller and a buyer. Nature chooses the seller’s type from a finite set Θ = {0, 1, ..., m}. There is a set M = {1, 2, ..., m} of certificates. For every k ∈ M, only types θ ≥ k can obtain Certificate k. For simplicity, assume that obtaining an available certificate is free. The seller chooses which certificate to obtain and present to the buyer. There is a continuation game after that. The model is called voluntary disclosure of verifiable information or persuasion. Ruitian Lang (ANU) Game Theory October 11, 2021 21 / 38 A persuasion game For simplicity, assume that after a certificate (if any) is presented, the seller makes a take-it-or-leave-it offer to the buyer. The seller’s payoff is payment from the buyer. The buyer’s payoff is the value vθ of the product (if purchased) minus payment (if any). Assume that v0 < v1 < ... < vm . Ruitian Lang (ANU) Game Theory October 11, 2021 22 / 38 Table of contents 1 Conditional Probabilities 2 Definition and the First Example 3 Some More Examples Ruitian Lang (ANU) Game Theory October 11, 2021 23 / 38 Overview In three classes of models, a player takes an action to actively influence other players’ beliefs in order to improve his future payoff. Career concerns (Signal jamming) The player does not know what he is. Signaling The player knows what he is and tries to portrait himself truthfully. Reputation The player knows what he is and tries to pretend to be something else. Ruitian Lang (ANU) Game Theory October 11, 2021 24 / 38 A simple career concern model There is a worker and two employers. Nature chooses a worker’s talent θ ∈ {H, L}. The probability that θ = H is µ ∈ (0, 1). There are two periods. In each period t, the following events happen: 1 2 3 Each employer i makes a wage offer wi,t (without observing the other employer’s offer). The worker chooses which offer to accept (or accepts neither). If working, the worker chooses a binary effort et ∈ {0, 1} and a binary outcome yt ∈ {R, 0} is generated and publicly observed. Ruitian Lang (ANU) Game Theory October 11, 2021 25 / 38 A simple career concern model (cont.) The probability that yt = R is qθ + aet where 0 < a < 1 − qH . When hiring the worker, Employer i’s stage payoff is yt − wi,t and the worker’s stage payoff is wi,t − cet with c > 0. All players discount future payoffs by δ ∈ (0, 1). The two employers essentially play a Bertrand competition, so they both offer wage equal to the expected revenue. Ruitian Lang (ANU) Game Theory October 11, 2021 26 / 38 Career concerns There is no contractual or discretionary bonus, so there is no direct incentive for effort. However, the worker may exert effort to improve the employers’ beliefs about his talent. When inferring about the worker’s talent from his output, both employers take into account the “distortion” created by the effort. Ruitian Lang (ANU) Game Theory October 11, 2021 27 / 38 Signal jamming Signal jamming refers to a situation where a player exerts a costly effort to distort a signal without observing the signal or the true state. Career concern is one example of signal jamming where the effort is socially beneficial. There are other models where the effort is a social waste such as advertisement or manipulating report to supervisor. A rational receiver of the signal should “deduct” the expected distortion from the observed signal. In the mostly commonly used functional form, the deduction is perfect. When the deduction is perfect, the equilibrium outcome (apart from the cost and benefit of the effort) is the same as when the sender has no opportunity to distort. Ruitian Lang (ANU) Game Theory October 11, 2021 28 / 38 Spence’s theory of education Nature chooses a student’s type θ ∈ {H, L}. The student observes his type and decides whether to pursue an advanced degree m ∈ {D, N}. An employer observes the student’s degree choice but not his type and chooses whether to offer a job. The cost of getting the advanced degree depends on the student’s type. Ruitian Lang (ANU) Game Theory October 11, 2021 29 / 38 Signaling There is a sender and a receiver. Nature chooses the sender’s type θ ∈ Θ. The sender observes his type and chooses a message m ∈ M from a pre-defined set of messages M. The receiver observes m but not θ and makes some decision d ∈ D. The receiver’s payoff uR depends on θ, m and d; crucially, the sender’s payoff uS depends on m as well as d and θ. Ruitian Lang (ANU) Game Theory October 11, 2021 30 / 38 Signaling (cont.) Compared with persuasion, the message is not a verifiable piece of information about the sender’s type. In theory, every message can be sent by all types. However, the “cost” of sending a message depends on the sender’s type. This is different from cheap talk, where all messages are costless. Ruitian Lang (ANU) Game Theory October 11, 2021 31 / 38 Separating vs pooling In a separating equilibrium, any two different types send different messages. In a pooling equilibrium, all types send the same message. An equilibrium may also be mixed when the sender uses a non-degenerate behavior strategy, or partially separating when the set space is partitioned into several groups and all types in the same group send the same message. Since every message is available to all types (albeit at different costs), consistency imposes no restriction on the receiver’s belief upon seeing a message that is not supposed to be sent by any type. Ruitian Lang (ANU) Game Theory October 11, 2021 32 / 38 The intuitive criterion Let Ã(m) be the set of receiver’s actions that is optimal under some probability measure on Θ when she receive message m. (Ã(m) is independent of m if the receiver’s payoff function is so.) Fix an assessment profile and denote by vS (θ) the expected payoff of a sender with type θ ∈ Θ. Intuitive criterion: for every message m ∈ M, if there is a θ ∈ Θ such that uS (a, m, θ) ≥ vS (θ) for no a ∈ Ã(m) and another θ0 ∈ Θ such that uS (a, m, θ0 ) ≥ vS (θ0 ) for some a ∈ Ã(m), then zero probability should be assigned to this θ by the receiver upon seeing m. In words, if sending a message m may (weakly) benefit θ0 but always makes θ strictly worse off, then we rule out θ upon seeing m. Ruitian Lang (ANU) Game Theory October 11, 2021 33 / 38 Remarks (optional) The criterion rules out “counterintuitive” beliefs of types that cannot benefit from sending the message under consideration. The criterion does not follow from trembling-hand perfection. Even with supermodularity, the criterion does not unravel the game; a stronger and much less intuitive criterion (called “divinity I”) does so. Ruitian Lang (ANU) Game Theory October 11, 2021 34 / 38 The idea of reputation There is a behavioral type that commits to a particular strategy and a rational (opportunistic) type that maximizes some payoff function. The rational type’s payoff is increasing in other players’ belief that he is behavioral. In early stages of the game, the rational type may imitate the behavioral type to build “reputation” that he is behavioral, in order to show his true face and earn high payoff later on. Ruitian Lang (ANU) Game Theory October 11, 2021 35 / 38 A concession game This is a hugely simplified version of Abreu and Gul (2000) “Bargaining and Reputation”. There are two players spliting a unit rent. Each player i has announced a demand αi ∈ ((1 + δ)−1 , 1). In Period t = 0 or 2, Player 1 decides whether to concede (receiving δ t (1 − α2 ) and ending the game);In Period 1, Player 2 decides whether to concede (receiving δ(1 − α1 ) and ending the game). If nobody concedes by the end of Period 2, the game ends with both players receiving zero payoff. Ruitian Lang (ANU) Game Theory October 11, 2021 36 / 38 A concession game (cont.) Under the above assumptions, the game has perfect information and can be solved backward. In equilibrium, Player 1 concedes immediately. Now assume that Player i is of a commited type with probability qi ∈ (0, 1) (independent of the other player’s type). Each player knows his own type at the beginning but not the opponent’s type. A commited type never concedes. A non-commited (rational or opportunistic) type maximizes his expected payoff. Would a rational Player 1 concede immediately? Ruitian Lang (ANU) Game Theory October 11, 2021 37 / 38 A concession game (cont.) Unless q2 is sufficiently high, there is no equilibrium in which a rational Player 1 concedes immediately. The result can be interpreted as follows: a rational Player 1 wants to pretend to be committed (for as long as possible) in the hope that it will convince Player 2 to concede. The full model considers a continuous-time version of the concession game and both players concede with some positive rate after a while. This outcome is referred to as a war of attrition. Ruitian Lang (ANU) Game Theory October 11, 2021 38 / 38