Bayes Correlated Equilibrium and Comparing Information Structures

Transcription

1 Bayes Correlated Equilibrium and Comparing Information Structures Dirk Bergemann and Stephen Morris Spring 2013: 521 B

2 Introduction game theoretic predictions are very sensitive to "information structure" a.k.a. higher order beliefs" information structure is hard to observe what can we say about outcomes if we do not know exactly what the information structure is?

3 Basic Question fix a game of incomplete information which distributions over actions and states could arise in Bayes Nash equilibrium in that game of incomplete information or one in which players observed additional information

4 Bayes correlated equilibrium set of outcomes consistent with equilibrium given any additional information the players may observe = set of outcomes that could arise if a mediator who knew the payoff state could privately make action recommendations

5 Bayes correlated equilibrium set of outcomes consistent with equilibrium given any additional information the players may observe = set of outcomes that could arise if a mediator who knew the payoff state could privately make action recommendations set of incomplete information correlated equilibrium outcomes

6 Bayes correlated equilibrium set of outcomes consistent with equilibrium given any additional information the players may observe = set of outcomes that could arise if a mediator who knew the payoff state could privately make action recommendations set of incomplete information correlated equilibrium outcomes we will define this very permissive version of incomplete information correlated equilibrium as "Bayes correlated equilibrium (BCE)"

7 Bayes correlated equilibrium set of outcomes consistent with equilibrium given any additional information the players may observe = set of outcomes that could arise if a mediator who knew the payoff state could privately make action recommendations set of incomplete information correlated equilibrium outcomes we will define this very permissive version of incomplete information correlated equilibrium as "Bayes correlated equilibrium (BCE)" and we will prove formal equivalence result between BCE and set of outcomes consistent with equilibrium given any additional information the players may observe

8 Many Applied Uses for Equivalence Result robust predictions robust identification tractable solutions optimal information structures

9 Many Applied Uses for Equivalence Result Robust Predictions in Games with Incomplete Information (linear best response games with continuum of agents), Extremal Information Structures in First Price Auctions (joint with Ben Brooks) Price Discrimination etc...

10 Today s Paper and Talk: Foundational Issues 1 basic equivalence result 2 more information can only reduce the set of possible outcomes... what is the formal ordering on information structures that supports this claim? "individual suffi ciency" generalizes Blackwell s (single player) ordering on experiments 3 literature: how does Bayes correlated equilibrium relate to other definitions of incomplete correlated equilibrium? how does our novel ordering on information structures relate to other orderings?

11 Outline of Talk: Single Player Case Bayes correlated equilibrium with single player: what predictions can we make in a one player game ("decision problem") if we have just a lower bound on the player s information structure ("experiment")? we suggest a partial order on experiments: one experiment is more incentive constrained than another if it gives rise to smaller set of possible BCE outcomes across all decision problems

12 Single Player Ordering an experiment S is suffi cient for experiment S if signals in S are suffi cient statistic for signals in S an experiment S is more incentive constrained than experiment S if, for every decision problem, S supports fewer Bayes correlated equilibria

13 Connection to Blackwell (1951/1953) an experiment S is more permissive than experiment S if more distributions over actions and states are supported by S than by S an experiment S is more valuable than experiment S if, in every decision problem, ex ante utility is higher under S than under S

14 Blackwell s Theorem Plus: One Player Theorem The following are equivalent: 1 Experiment S is suffi cient for experiment S (statistical ordering); 2 Experiment S is more incentive constrained than experiment S (incentive ordering); 3 Experiment S is more permissive than experiment S (feasibility ordering); 4 Experiment S is more valuable than experiment S

15 Blackwell s Theorem Plus: Many Players Theorem The following are equivalent: 1 Information structure S is individually suffi cient for information structure S (statistical ordering); 2 Information structure S is more incentive constrained than information structure S (incentive ordering); 3 Information structure S is more permissive than information structure S (feasibility ordering); 4 Information structure S is more valuable than information structure S

16 Related Literature 1 Forges (1993, 2006): many notions of incomplete information correlated equilibrium 2 Lehrer, Rosenberg and Shmaya (2010, 2012): many multi-player versions of Blackwell s Theorem 3 Liu (2005, 2012): one more (important for us) version of incomplete information correlated equilibrium and a characterization of correlating devices that relates to our ordering

17 Single Person Setting single decision maker finite set of payoff states Θ a decision problem G = (A, u, ψ), where A is a finite set of actions, u : A Θ R is the agent s (vnm) utility and is a prior. ψ (Θ) an experiment S = (T, π), where T is a finite set of types (i.e., signals) and likelihood function π : Θ (T ) a choice environment (one player game of incomplete information) is (G, S)

18 Consistency behavior in (G, S) is described by a joint distribution of actions, signals, and states: ν (A T Θ) Definition A distribution ν (A T Θ) is consistent for (G, S) if ν (a, t, θ) = ψ (θ) π (t θ) a A for each t T and θ Θ.

19 Feasibility / Belief Invariance Definition A distribution ν (A T Θ) is belief invariant in (G, S) if ν (θ t, a) is independent of a.

20 Statistical Fact Consider a joint distribution of any three variables, like ν (A T Θ) Say that ν (θ t, a) is independent of a if ν (θ t, a) = ν (θ t, a ) for all θ, t, a, a The following three statements are equivalent: 1 ν (θ t, a) is independent of a 2 ν (a t, θ) is independent of θ 3 ν (a, t, θ) = ν (t) ν (a t) ν (θ t)

21 Statistical Fact The following three statements are equivalent: 1 ν (θ t, a) is independent of a 2 ν (a t, θ) is independent of θ 3 ν (a, t, θ) = ν (t) ν (a t) ν (θ t) Proof that (1) implies (3): ν (a, t, θ) = ν (t) ν (a t) ν (θ t, a), w.l.o.g. = ν (t) ν (a t) ν (θ t), by (1)

22 Equivalent Definition Definition A distribution ν (A T Θ) is belief invariant in (G, S) if ν (θ t, a) is independent of a. Definition A distribution ν (A T Θ) is belief invariant in (G, S) if ν (a t, θ) is independent of θ. Belief invariance captures distributions that can arise from a decision maker randomizing conditional on his signal.

23 Obedience Definition A distribution ν (A T Θ) is obedient for (G, S) if θ Θ ν (a, t, θ) u (a, θ) θ Θ for each a A, t T and a A. ν (a, t, θ) u ( a, θ ) Definition A distribution ν (A T Θ) is a Bayes Nash equilibrium (single agent optimal) if it is consistent, belief invariant and obedient.

24 consider separate experiments, S 1 = ( T 1, π 1), S 2 = ( T 2, π 2) Combining Experiments join the experiments S 1 and S 2 into S = (T, π ) : T = T 1 T 2, π : Θ ( T 1 T 2) Definition S is a combined experiment of S 1 and S 2 if: 1 T = T 1 T 2, π : Θ ( T 1 T 2) 2 marginal of S 1 is preserved: π (( t 1, t 2) ) θ = π 1 ( t 1 θ ), t 1, θ. t 2 T 2 3 marginal of S 2 is preserved: π (( t 1, t 2) θ ) = π 2 ( t 2 θ ), t 2, θ.

25 Combining Experiments and Expanding Information there are multiple combined experiments S for any pair of experiments, since only the marginals have to match If S is combination of S and another experiment S, we say that S is an expansion of S.

26 (One Person) Robust Predictions Question fix (G, S) which outcomes can arise under optimal decision making in (G, S ) where S is any expansion of S? as a special case, information structure may be the null information structure: S = {T = {t }, π (t θ ) = 1}

27 (One Person) Bayes Correlated Equilibrium Definition Distribution ν (A T Θ) is a Bayes correlated equilibrium (BCE) of (G, S) if it is obedient and consistent.... not necessarily belief invariant

28 (One Person) Bayes Correlated Equilibrium Definition Distribution ν (A T Θ) is a Bayes correlated equilibrium (BCE) of (G, S) if it is obedient and consistent. Definition An outcome is a distribution µ (A Θ) over actions and states. distribution ν (A T Θ) induces outcome µ (A Θ) if ν (a, t, θ) = µ (a, θ), a A, θ Θ. t T Definition An outcome µ is a BCE outcome of (G, S) if there is a BCE ν (A T Θ) of (G, S) that induces µ.

29 (One Person) Bayes Correlated Equilibrium Theorem An outcome µ is a BCE outcome of (G, S) if and only if there is an expansion S of S such that µ is a Bayes Nash equilibrium outcome for (G, S ) Idea of Proof: ( ) S has "more" obedience constraints than S ( ) let ν be BCE of (G, S) supporting µ and consider expansion S with T = T A and π (t, a θ) = ν (t, a θ).

30 Example: Bank Run A bank is solvent or insolvent: Each event is equally likely: Θ = {θ I, θ S } ψ ( ) = ( 1 2, 1 ) 2 running (r) gives payoff 0 not running (n) gives payoff 1 if insolvent, y if solvent: 0 < y < 1 G = (A, u) with A = {r, n} and u given by θ S θ I r 0 0 n y 1

31 Bank Run: Common Prior Only suppose we have the prior information only - the null information structure: S = (T, π ), T = {t } parameterized consistent outcomes: µ (a, θ) θ S θ I r 1 2 ρ S 1 2 ρ I n 1 2 (1 ρ S ) 1 2 (1 ρ I ) ρ S : (conditional) probability of running if solvent ρ I : (conditional) probability of running if insolvent

32 Bank Run: Obedience agent may not necessarily know state θ but makes choices according to µ ( ) if "advised" to run, run has to be a best response: 0 ρ S y ρ I ρ I ρ S y if "advised" not to run, not run has to be a best response (1 ρ S ) y (1 ρ I ) 0 ρ I (1 y) + ρ S y here, not to run provides binding constraint: ρ I (1 y) + ρ S y never to run, ρ I = 0, ρ S = 0, cannot be a BCE

33 set of BCE described by (ρ I, ρ S ) Bank Run: Equilibrium Set never to run, ρ I = 0, ρ S = 0, is not be a BCE

34 Bank Run: Extremal Equilibria BCE minimizing the probability of runs has: ρ I = 1 y, ρ S = 0 Noisy stress test T = { t I, t S } implements BNE via informative signals: π (t θ ) θ I θ S t I 1 y 0 t S y 1 the bank is said to be healthy if it is solvent (always) and if it is insolvent (sometimes) solvent and insolvent banks are bundled

35 Bank Run: Positive Information suppose player observes conditionally independent private binary signal of the state with accuracy: q > 1 2 S = (T, π) where T = { t S, t I } : π θ S θ I t S q 1 q t I 1 q q strictly more information than null information q = 1 2

36 Bank Run: Obedience Constraints conditional probability of running now depends on the signal: t { t S, t I } ρ I, ρ S become ( ρ I I, ) ( ρi S, ρ S I, ρ S ) S conditional distribution of outcomes for t S : t S θ S θ I r qρ S S (1 q) ρ S I n q ( 1 ρ S ) S (1 q) ( 1 ρ S ) I conditional distribution of outcomes for t I : t I θ S θ I r (1 q) ρ I S qρ I I n (1 q) ( 1 ρ I ) S q ( 1 ρ I ) I

37 Bank Run: Additional Obedience Constraints possibly distinct running probabilities for different types: ( ) ( ) ρ I I, ρi S, ρ S I, ρs S conditional obedience constraints, say for t S : r : 0 qρ S S y (1 q) ρs I ( ) ( ) n : q 1 ρ S S y (1 q) 1 ρ S I 0 or r : ρ S I q 1 q ρs S y n : ρ S I 1 q 1 q y + q 1 q ρs S y

38 set of BCE described by (ρ I, ρ S ) Bank Run: Equilibrium Set ρ I = 1, ρ S = 0, is complete information BCE

39 Incentive Compatibility Ordering Write BCE (G, S) for the set of BCE outcomes of (G, S) Definition Experiment S is more incentive constrained than experiment S if, for all decision problems G, BCE (G, S) BCE ( G, S ). Note that "more incentive constrained" corresponds, intuitively, to having more information

40 Permissiveness outcome µ (A Θ) is feasible for (G, S) if there exists a consistent and belief invariant ν (A T Θ) inducing µ write F (G, S) for the set of feasible outcomes of (G, S) Definition Experiment S is more permissive than experiment S if, for all decision problems G, F (G, S) F ( G, S )

41 More Valuable Than write V (G, S) for the maximum attainable utility in (G, S) V (G, S) = max µ F (G,S) a A,θ Θ µ (a, θ) u (a, θ) Definition Experiment S is more valuable than experiment S if, for all decision problems G, V (G, S) V ( G, S )

42 Back to the Example: Feasibility suppose we have the prior information only - the null information structure: S 0 = (T 0, π), T 0 = {t 0 } feasible outcomes µ (a, θ) can be described by (ρ I, ρ S ):

43 Back to the Example: Feasibility suppose player observes conditionally independent private binary signal of the state with accuracy q 1 2 : feasible outcomes µ (a, θ) can be described by (ρ I, ρ S ):

44 Statistical Ordering: Suffi ciency Experiment S is suffi cient for experiment S if there exists a combination S of S and S such that is independent of t. Pr ( t t, θ ) = π (t, t θ) π ( t, t θ ) t Θ

45 Suffi ciency: Two Alternative Statements 1 (following from statistical fact): for any ψ ++ (Θ), Pr ( θ t, t ) = ψ (θ) π (t, t θ) ψ ( θ ) π ( t, t θ ). θ Θ is independent of t. 2 (naming the θ-independent conditional probability) there exists φ : T (T ) such that π ( t θ ) = t T φ ( t t ) π (t θ).

46 Aside: Belief Invariance = Suffi ciency of Signals An outcome µ (A Θ) embeds an experiment (A, π ν ) where ν (a, θ) π ν (a θ) = ν (ã, θ) An outcome is induced by a distribution which is belief invariant and consistent for (G, S) if and only if S is suffi cient for (A, π ν ). ã

47 Blackwell s Theorem Plus Theorem The following are equivalent: 1 Experiment S is suffi cient for experiment S (statistical ordering); 2 Experiment S is more incentive constrained than experiment S (incentive ordering); 3 Experiment S is more permissive than experiment S (feasibility ordering); 4 Experiment S is more valuable than experiment S

48 Proof of Blackwell s Theorem Plus Equivalence of (1) "more permissive" and (3) "suffi cient for" is due to Blackwell (3) (4) by definition Simple direct arguments from (4) (1) (2) "suffi cient for" (3) "more incentive constrained": 1 take the stochastic transformation φ that maps S into S 2 take any BCE ν (A T Θ) of (G, S) and use φ to construct ν (A T Θ) 3 show that ν is a BCE of (G, S )

49 Proof of Blackwell s Theorem Plus (3) "more incentive constrained" (2) "more permissive" by contrapositive suppose S is not more permissive than S so F (G, S) F (G, S ) for some G so there exists G and ν (A T Θ) which is feasible for (G, S ) and gives outcome µ (A Θ), with µ not feasible for (G, S) can choose G so that the value V of ν in (G, S ) is V and the value every feasible µ of (G, S) is less than V now every there all BCE of (G, S ) will have value at least V and some BCE of (G, S) will have value strictly less than V so BCE (G, S) BCE (G, S )

50 Basic Game players i = 1,..., I (payoff) states Θ actions (A i ) I i=1 utility functions (u i ) I i=1, each u i : A Θ R state distribution ψ (Θ) ) G = ((A i, u i ) I i=1, ψ "decision problem" in the one player case

51 Information Structure signals (types) (T i ) I i=1 signal distribution π : Θ (T 1 T 2... T I ) ) S = ((T i ) I i=1, π "experiment" in the one player case

52 Key Properties Definition [UNCHANGED] A distribution ν (A T Θ) is consistent for (G, S) if ν (a, t, θ) = ψ (θ) π (t θ) a A for each t T and θ Θ. Definition A distribution ν (A T Θ) is obedient for (G, S) if a i A i,t i T i,θ Θ a i A i,t i T i,θ Θ for each i, t i T i and a i, a i A. ν ((a i, a i ), (t i, t i ), θ) u ((a i, a i ), θ) ν ((a i, a i ), (t i, t i ), θ) u (( a i, a i ), θ )

53 Robust Predictions Question fix (G, S). which outcomes can arise in Bayes Nash Equilibrium in (G, S ) where S is an expansion of S? Definition Distribution ν (A T Θ) is a Bayes correlated equilibrium (BCE) of (G, S) if it is obedient and consistent. Outcome µ (A Θ) is BCE outcome of (G, S) if a BCE ν (A T Θ) gives outcome µ.

54 Bayes Correlated Equilibrium Theorem An outcome µ is a BCE outcome of (G, S) if and only if there is an expansion S of S such that µ is a BNE outcome for (G, S ) Proof Idea: ( ) feasibility of S has "more" obedience constraints than S ( ) let ν be BCE of (G, S) supporting µ and consider augmented information structure S with T = T A and π (t, a θ) = ν (t, a θ).

55 Statistical Ordering: Individual Suffi ciency Experiment S is individually suffi cient for experiment S if there exists a combination S of S and S such that π ( t, ( t i, ) ) t i θ Pr ( t i t i, t i, θ ) = is independent of (t i, θ). t i T i t i T i t i T i π ( t, ( t i, ) ) t i θ

56 Suffi ciency: Two Alternative Statements following from statistical fact applied to triple (t i, t i, (t i, θ)) after integrating out t i for any ψ ++ (Θ), Pr ( t i, θ t i, t i ) = is independent of t i. t i T i t i T i θ Θ t i T i ψ (θ) π ( (t i, t i ), ( t i, ) ) t i θ ) ( (ti ) ( ) ) ψ ( θ π, t i, t i, t i θ

57 Suffi ciency: Two Alternative Statements letting φ : T Θ (T ) be conditional probability for combined experiment π there exists φ : T Θ (T ) such that π ( t θ ) = t T φ ( t t, θ ) π (t θ) and ( Pr t i t i, t i, θ ) = φ (( t i, t ) i (ti, t i ), θ ) φ t i T i is independent of (t i, θ)

58 Nice Properties of Ordering Transitive Neither weaker or stronger than suffi ciency (i.e., treating signal profiles as multidimensional signals) Two information structures are each suffi cient for each other if and only if they share the same higher order beliefs about Θ S is individually suffi cient for S if and only if S is higher order belief equivalent to an expansion of S S is individually suffi cient for S if and only if there exists a combined experiment equal to S plus a correlation device

59 Example Dekel, Fudenberg, Morris (2007), Liu (2011) Let Θ = {0, 1}, I = 2 Let G have A 1 = A 2 = {0, 1} and θ = , 1 0, 0 1 0, 0 1, 1 θ = , 0 1, 1 1 1, 1 0, 0

60 Example Compare null information structure S......with information structure S with T 1 = T 2 = {0, 1} π ( 0) π ( 1) Each information structure is individually suffi cient for the other

61 Feasibility Definition A distribution ν (A T Θ) is belief invariant in (G, S) if Pr (t i, θ t i, a i ) is independent of a i. Definition An outcome µ is feasible for (G, S) if and only if it is induced by a consistent and belief invariant action-signal-state distribution Write F (G, S) for the set of feasible outcomes for (G, S) Write V (G, S) for the highest feasible payoff in a common interest game Define orderings as before...

62 Blackwell s Theorem Plus Theorem The following are equivalent: 1 Information structure S is individually suffi cient for information structure S (statistical ordering); 2 Information structure S is more incentive constrained than information structure S (incentive ordering); 3 Information structure S is more permissive than information structure S (feasibility ordering); 4 Information structure S is more valuable than information structure S

63 Proof of Blackwell s Theorem Plus Equivalence of (1) and (4) can be proved from results of LRS discussed below (1) (3) directly constructive argument (3) (4) by definition (1) "suffi cient for" (2) "more incentive constrained" works as in the single player case 1 take the stochastic transformation φ that maps S into S 2 take any BCE ν (A T Θ) of (G, S) and use φ to construct ν (A T Θ) 3 show that ν is a BCE of (G, S ) need a new argument to show (3) (2)

64 New Argument: Game of Belief Elicitation Suppose that S is more incentive constrained than S Consider game where players report types in S Construct payoffs such that (i) truthtelling is a BCE of (G, S) (ii) actions corresponding to reporting beliefs over T i Θ with incentives to tell the truth In order to induce the truth-telling outcome of (G, S), there must exist φ : T Θ (T ) corresponding to one characterization of individual suffi ciency

65 Incomplete Information Correlated Equilibrium Forges (1993): "Five Legitimate Definitions of Correlated Equilibrium" BCE = set of outcomes consistent with (common prior assumption plus) common knowledge of rationality and that players have observed at least information structure S. Not a solution concept for a fixed information structure as information structure is in flux

66 Other Definitions: Stronger Feasibility Constraints Belief invariance: information structure cannot change, so players cannot learn about the state and others types from their action recommendations Liu (2011) - belief invariant Bayes correlated equilibrium: consistency, obedience and belief invariance captures common knowledge of rationality and players having exactly information structure S. Join Feasibility: equilibrium play cannot depend on things no one knows Forges (1993) - Bayesian solution: consistency, obedience and join feasibility captures common knowledge of rationality and players having at least information structure S and a no correlation restriction on players conditional beliefs belief invariant Bayesian solution - imposing both belief invariance and join feasibility - has played prominent role in the literature

67 Other Definitions: the rest of the Forges Five 1 More feasibility restrictions: agent normal form correlated equilibrium 2 More incentive constraints: communication equilibrium: mediator can make recommendations contingent on players types only if they have an incentive to truthfully report them. 3 Both feasibility and incentive constraints: strategic form correlated equilibrium

68 () November 1, / 3

69 () November 1, / 3

70 Generalizing Blackwell s Theorem we saw - in both the one and the many player case - that "more information" helps by relaxing feasibility constraints and hurts by imposing incentive constraints Lehrer, Rosenberg, Shmaya (2010, 2011) propose family of partial orders on information structures, refining suffi ciency LRS10 kill incentive constraints by showing orderings by focussing on common interest games. Identify right information ordering for different solution concepts LRS 11 kill incentive constraints by restriction attention to info structures with the same incentive constraints. Identify right information equivalence notion for different solution concepts We kill feasibility benefit of information by looking at BCE. Thus we get "more information" being "bad" and incentive constrained ordering characterized by individual suffi ciency. Same ordering corresponds to a natural feasibility ordering (ignoring incentive constraints)

71 Conclusion a permissive notion of correlated equilibrium in games of incomplete information: Bayes correlated equilibrium BCE renders robust predicition operational, embodies concern for robustness to strategic information leads to a natural multi-agent generalization of Blackwell s single agent information ordering