Lecture 1
Lecture 1
Mechanism Design
Fabio Castiglionesi
CentER, and Tilburg University
1 Mechanism Design
If we have some theory which allows us to predict the outcome of a given
game, we may also pose the reverse question: If we want to obtain a certain
outcome, how should we design the rules of the game? This is what the mech-
anism design literature studies. Formally, the problem can be represented as
1
follows.
Consider a society composed of a set I of members, with i denoting the
generic member. There is a set of possible states of nature Θ, and each agent
is endowed with a partition PΘi of Θ. Furthermore, each agent i can take an
action ai in some set Ai . We denote by A = Πi∈I Ai the set of all possible
action profiles, with generic element a = {ai }i∈I . Action profiles may be
partially observable, and we will assume that each agent is endowed with a
partition PAi of A.
There is a set D of decisions that the society can take. A collective de-
cision rule (sometimes called a social welfare correspondence) is a mapping
from the possible states of the world and action profiles to (subsets of) de-
cisions, that is d : Θ × A → D. In other words, the society would like,
for each state of the world θ, and action profile a to have the decision d be
implemented.
Each agent is endowed with a utility function ui : Θ × D × A → , with
ui (θ, d, a) denoting the utility when the state of the world is θ, the decision
is d and the action profile is a.
Special cases of this general problem can be considered according to the
following criteria: whether the private information bears on who the agent
is (or what his characteristics are) or on what the agent does. The former
are called hidden information models,1 the latter are called hidden action (or
moral hazard) models.
2
from the agents, and then take the appropriate decision. In this case we can
write utilities as ui (θ, d), and a collective decision rule as d (θ) ⊂ D.
3
the agents can be (at least partially) observed by other players. One ex-
treme case is the one in which all agents observe perfectly the action profile,
PAi (a) = {a} for each i and a. In principle, in this case there is no private in-
formation on actions, and therefore the trade-off between incentive provision
and risk-sharing disappears. However, nontrivial problems of multiplicity of
equilibrium may arise.
4
For example, assume that (M, g) induces a game of incomplete informa-
tion among players, and that the solution notion we adopt is that of Bayesian
equilibrium. Assume further that the collective decision rule is a function,
so that d (θ) is a singleton for each θ. Then we say that the mechanism
(M, g) realizes d (θ) if there is a Bayesian equilibrium σ : Θ → M such that
g (σ (θ)) = d (θ) for each θ.
The term ‘weak implementation’ refers to the fact that we do not insist
on the mechanism having only outcomes which we deem desirable (that is,
belonging to d (θ)). In fact, we only require that elements in d (θ) be among
the possible outcomes of the mechanism, leaving open the possibility of other
outcomes of the mechanism which may not be desirable. Of course, we
may want to look at a stronger form of implementation, requiring that only
desirable outcomes should occur. This is what motivates the next definition.
In this case we require that an outcome can result from the mechanism
if and only if it is in d (θ). The ‘only if’ part is what differentiates the
notion of (strong) implementation from the one of weak implementation. The
difference with weak implementation is that with (strong) implementation we
want all the equilibria of the mechanism Γ to deliver d (θ), rather than just
some (or one) of the equilibria of the mechanism to deliver d (θ). However,
there are significant differences between the set of collective decision rules
which are weakly implementable and the set of collective decision rules which
are implementable. We will only focus our attention on weak implementation.
5
The direct mechanism is a mechanism in which each agent has a message
space M i = {PΩi (θ)}θ∈Θ , that is each agent simply announces the information
possessed. The planner collects all the information issued by the agents and
determines
the state of the world θ = ∩i∈I PΘi (θ). Then, the decision selected
is d θ . The Revelation Principle tells us that, only if truthtelling is an
equilibrium in the direct mechanism, it is possible to weakly-implement the
decision rule d(θ). In a truthful mechanism each agent finds it optimal to
announce the true value of her information (or ‘type’).
Summing up, in the case of weak implementation (in dominant strategies
or in Bayesian equilibrium), the revelation principle allows us to restrict
attention to a particular class of mechanisms, the class of direct (truthful)
mechanisms.
The Revelation Principle. Suppose that there is a mechanism Γ = (M, g)
which weakly implements d(θ) in Bayesian equilibrium. Then d(θ) is
also implemented by the direct mechanism.
Proof. Since Γ = (M, g) weakly implements d(θ), the mechanism Γ has a
Bayesian equilibrium σ such that g (σ (θ)) = d (θ) for each θ. Define now the
new mechanism Γ = M , g
such that M i = {P i (θ)}
θ∈Θ and g = g(σ(.)) =
Θ
i = {P i (θ)}
d(.). Since M
Θ θ∈Θ the mechanism Γ is a direct mechanism. We
now show that truthtelling is an equilibrium of the direct mechanism.
Suppose not. Then there is at least one
agent i and state ω
such that the
i i
agent prefers to announce PΘ θ = PΘ θ . Let:
E ui g σ i PΘi θ , σ −i PΘ−i (θ) , θ PΘi θ
be the expected
utility of agent i when the other
agents
tell the truth, infor-
i i
mation PΘ θ is observed and information PΘ θ is announced by agent i.
Then, since PΘi θ is a profitable deviation from truthtelling:
E ui g σ i PΘi θ , σ −i PΘ−i (θ) , θ PΘi θ > (1)
E ui g σ i PΘi θ , σ −i PΘ−i (θ) , θ PΘi θ
Consider now the original mechanism Γ, and let mi = σ i PΘi θ i =
and m
σi PΘi θ . Then, inequality [1] implies:
E ui g mi , σ −i PΘ−i (θ) , θ PΘi θ >
6
E ui g m
i , σ −i PΘ−i (θ) , θ PΘi θ
The last inequality says
that
agent i prefers to use the message intended to
reveal information PΘi θ under the strategy profile σ (even if she observes
PΘi θ ). In other words, when the state of the world is given by PΘi θ ,
agent
i prefer to use the message intended to reveal the state of the world
PΘi θ under the strategy profile σ (i.e., agent i finds convenient to deviate).
This contradicts the fact that σ is an equilibrium in the original mechanism
(M, g).
Remark. A version of the Revelation Principle also holds for dominant
strategy implementation, that is when we predict that agents only select
dominant strategies (the proof is easily adapted from the one above).
The revelation principle can be interpreted as follows. Each agent dele-
gates her choice of a strategy to a planner. The planner asks to the player
what her type is, and with this information chooses a strategy in order to
maximize the player’s playoff. The revelation principle states that no agent
has incentive to fool the mediator by telling false information, because this
would result in the mediator picking up not the best strategy.
The revelation principle allows us to focus our attention on the properties
of the collective decision rule d, rather than on the institutional details which
allow the implementation of the rule. It says that, as long as we are satisfied
with weak implementation, either the simplest institution (i.e. the direct
mechanism) works, or no institution works. No complicated ways of eliciting
the agents’ information need to be taken into account.2
The following property is crucial for weak-implementation.
7
A decision rule satisfies incentive compatibility (IC) if at each state of the
world, each agent prefers to announce the truth when the other agents are
announcing the truth.3 Then we immediately have the following result.
Proof. For the ‘if’ part (sufficient condition), simply observe that when d
satisfies incentive compatibility, the direct mechanism weakly implements d.
For the ‘only if’ part (necessary condition) we invoke the revelation princi-
ple. If d does not satisfy incentive compatibility then truthtelling is not an
equilibrium of the direct mechanism. But then, by the revelation principle,
no other mechanism can weakly implement d.
Then, in case of weak implementation (i.e., it is enough to have a mech-
anism in which one of the equilibria is the desired outcome), the incentive
compatibility condition is a necessary and sufficient condition for implemen-
tation. All we have to do is to check whether truthtelling is an equilibrium in
the direct mechanism. If this is not the case, then the outcome is not weakly
implementable.
Incentive compatibility does not assure instead (strong) implementation.
The difference on the conditions between weak implementation and (strong)
implementation is due to the fact that the truthful strategy may not be
the only equilibrium strategy in the direct mechanism, even if the incentive
compatibility constraint is satisfied. In other words, there can be several
equilibria in the direct mechanism: In one of them players are using truth-
ful strategies, but in the other equilibria the players are using untruthful
strategies. If such untruthful equilibria exist, unless there is a good reason
to discard all but the truthful equilibrium, truthful implementation does not
imply (strong) implementation.
In principle, there are no good reasons to consider only the truthful equi-
librium in a direct mechanism. However, this is the attitude that has pre-
3
Notice that we have written the function d as a function of reported information
rather than as a function of the state of the world ω. This is necessary in order to take
care of situations in which the information reported by the agents is incompatible, that
is the intersection of the reported sets PΩi is empty. When we describe uncertainty with
the ‘type’ structure, this is equivalent to agents announcing a profile of types having
probability zero. In order to ensure incentive compatibility, the function d has to specify
the decision to be taken in those cases.
8
vailed in the literature because the results are clearer. If we want to elimi-
nate all the equilibria that are not desired then the direct mechanism is not
enough. More complex mechanisms are needed.
3 Dominant Strategies
We would like to be able to implement a collective decision rule by imposing
the weakest behavioral requirements on the agents. One weak requirement is
that agents only pick dominant strategies. This implies that we do not have
to worry about the conjectures that each agent may hold about other agents’
behavior. So why do we choose to use Bayesian equilibrium as solution
concept instead of dominant strategy to analyze weak implementation?
The reason is that weakly implementing a decision rule in dominant
strategies may be very hard in some contexts. An important result due
to Gibbard and Sattherwaite states that if a decision rule is implementable
in dominant strategies and the agents may rank elements in D in all possible
ways, then the only decision rules which are weakly implementable in domi-
nant strategies are the ones which are dictatorial, that is rules which always
select the alternative preferred by a particular individual (the dictator).
There are two ways out of this negative result. The first is to use a
different solution concept to predict the outcome of the mechanisms proposed
to the agents (obvious candidates are less restrictive solution concepts such
as Bayesian equilibrium as done before). The other possibility is to retain the
requirement that agents play dominant strategies, but limit the variability of
the environment. We will focus on this second alternative in the next section.
9
denoting the transfer to agent i. Then D = X × M 1 × . . . M n .
The utility function of agent i is ui (d, mi ) = xv i + mi . The cost of the
project is C, and a typical state of the world is a vector θ = (v 1 , . . . , v n ).
We would like to weakly implement the efficient rule, stating that x (θ) = 1
iff ni=1 vi ≥ C. The problem is that the agents may have an incentive to
understate or overstate their valuations, in order to increase or decrease the
probability of producing the public good. Therefore, in order to implement
the rule we need to devise a system of transfers m1 (θ) , . . . , mn (θ) such that
every agent has an incentive to tell the truth. Notice that here we depart
somewhat from the framework spelled out in the previous section. In princi-
ple, a decision rule describes both the production and the transfer decision.
Here we focus only on the production decision, permitting any kind of trans-
fers. This is important because, as we will shortly see, implementation in
dominant strategies becomes impossible as soon as we impose some reason-
able restriction on the class of transfers allowed.
At any rate, here is a mechanism that implements in dominant strategies
the efficient rule. Let vi be the announcement by agent i. Then:
n
j − C + h (v
j=i v −i ) if j ≥ C
j=1 v
mi v1 , . . . , vn =
n
h (v−i ) if j=1 vj < C
Under this transfer scheme, for any vector of announcements v−i by the other
players the agent solves:
max x vi , v−i v i + vj − C + h v−i
vi j=i
It is clear that by announcing the truth vi = vi the agent maximizes the
objective function. We therefore conclude that truthtelling is implemented
in dominated strategies.
The main problem with the Clarke-Groves scheme is that there may be
no way to make sure that the budget balances. In particular, notice that we
can exploit the quasi-linearity assumption and the lack of IR constraints to
make sure that the sum of transfers is always negative, but we cannot make
sure that it is zero. A solution to the problem can be found if we give up
dominant strategies and accept to use Bayesian equilibrium as the solution
notion.
10
4 Appendix: Review of Game Theory
This appendix does not represent an exhaustive treaty of game theory. A
complete and in deep treaty is given by Fudenberg and Tirole (1991). A more
introductory level of game theory is provided by Gibbons (1992). Salanié’s
appendix (1997) provides a summary of the main game theoretic concepts
used in contract theory.
11
4.1.1 Normal Form Games with Complete Information
These class of games are of the following simple form: first the players simul-
taneously choose actions; then the players receive payoffs that depend on the
combination of actions just chosen. Each player’s payoff function, i.e. the
function that determines the player’s payoff from the combination of actions
chosen by the players, is common knowledge among players.
Let I = {1, . . . , n} be a set of players. An n−person normal form game
of complete information G is an array:
G = {A1 , . . . , An , u1 , . . . un }
where:
We will frequently use the notation A−i = ×j=i Aj to denote the set of
possible action profiles that can be undertaken by agents other than i. The
typical element of A−i will be denoted a−i = (a1 , . . . ai−1 , ai+1 , . . . an ), and
when we want to emphasize in a given action profile a ∈ A the action of
agent i we write a = (ai , a−i ).
In the case of normal form games, a strategy σi by agent i can be described
as a probability distribution on the set Ai . In general, given any set X we
will denote by ∆X the set of probability distributions over X. Thus, a
strategy is an element of ∆Ai . Given any action ai , the symbol σi (ai ) denotes
the probability that agent i will choose action ai . More generally, given
any subset Ai ⊂ Ai , the probability of taking an action in the set Ai is
σi (Ai ) = ai ∈Ai σi (ai ). A strategy by agent i is called pure if the support
of σi is a singleton set (i.e., {ai }); it is called mixed otherwise. Since in
a normal form game it is assumed that agents take their action in isolation
and independently, a given collection of strategies (σ1 , . . . , σn ) can be used to
define a probability distribution over A. We simply have that the probability
of an action profile a = (a1 , ...an ) is given by σ(a) = σ1 (a1 )·σ2 (a2 )·. . .·σn (an ).
12
If agent i chooses action ai and the other players are using the strategies
σ−i then the expected utility of agent i is:
Ui (ai , σ−i ) = ui (ai , a−i ) σ−i (a−i )
a−i ∈A−i
13
precedes n if n ≺ n . Equivalently, we say that n is a successor of n. If a
node precedes all other nodes then that node is called the origin (or root) of
the game. If a node has no successors then it is called a terminal node. We
will call Z the set of terminal nodes. A finite n−person extensive form game
of complete information Γ is an array:
Γ = {{N, ≺} , ι, B1 , . . . , Bn , C1 , . . . Cn , u1 , . . . un }
where:
The semantic convention has prevailed that an extensive form game is said
to be of perfect information if each information set is singleton, otherwise it
is called of imperfect information. A subgame of an extensive form game Γ is
obtained when a certain node n0 can be used as the root of a tree. We don’t
give here the formal definition. Notice that, by definition, Γ is a subgame of
itself. A subgame Γ excluding some nodes of Γ is called a proper subgame.
The definition of strategy is slightly more complicated for extensive form
games. Remember that Bi = {b1i , . . . , bm i } is the collection of information
i
sets of agent i, and Ci bki is the set of possible actions at information set
bki . Then
a strategy for agent i is a collection of probability distributions
σi = σb1i , . . . , σbm i , where each probability distribution σbk is defined on
i i
the set Ci bki . For any subgame Γ of the extensive form game Γ we can
define the restriction of the strategy of player i to the subgame Γ simply as
14
the collection of probability distributions σbki such that the information set
bki belongs to the subgame. The restriction is denoted σi |Γ .
When a strategy profile σ = (σ1 , . . . , σn ) is given, we obtain a probability
distribution over the set of final nodes Z. Similarly, for any subgame Γ , we
can define the restricted strategy profile σ|Γ = (σ1 |Γ , . . . , σn |Γ ), and such
restricted strategy profile delivers a probability distribution over Z , the set
of terminal nodes of Γ .
A subgame perfect equilibrium is a collection of strategies σ ∗ = (σ1∗ , . . . , σn∗ )
such that for each subgame Γ the restriction σ ∗ |Γ is a Nash equilibrium. No-
tice that the definition implies that a subgame perfect equilibrium is a Nash
equilibrium, since the game Γ is a subgame of itself. Therefore, the set of
subgame perfect equilibria of a given game is contained in the set of Nash
equilibria of that game. The two sets coincide if the game has no proper
subgames. There are however games having Nash equilibria which are not
subgame perfect equilibria. The reason is that the concept of Nash equilib-
rium does not put any restriction on strategies adopted on subgames which
are not reached with positive probability, while subgame perfection insists
on rational behavior on unreached subgames as well.
Gµ = {Θ1 , . . . , Θn , µ, A1 , . . . An , u1 , . . . , un }
where:
15
• µ is a probability distribution over Θ = ×ni=1 Θi . The set Θ is called
the set of the states of the world, and it has typical element θ. The
notations Θ−i and θ−i have obvious meaning.
16
is given by:
Ui (ai , σ−i | θi ) = ui ((ai , a−i ) , (θi , θ−i )) σ−i ( a−i | θ−i ) µ (θ−i | θi )
θ−i ∈Θ−i a−i ∈A−i
where σ−i (a−i | θ−i ) is the probability that action profile a−i is taken by agents
other than i when their type profile is θ−i , and µ ( θ−i | θi ) is the conditional
probability of θ−i given θi . In order to compute µ (θ−i | θi ), it is assumed that
it is common knowlwdge that nature draws the type vector θ = (θ1 , ..., θn )
according to the prior distribution µ (θ). Then when nature reveals θi to
player i, she can compute the belief µ ( θ−i | θi ) using Bayes’ rule:
µ (θ−i , θi ) µ (θ−i , θi )
µ ( θ−i | θi ) = = .
µ (θi ) θ−i ∈Θ−i µ (θ−i , θi )
That is, no player wants to change her startegy, even if the change involves
only one action by one type.
17
Given a strategy profile σ we say that a probability assessment Ψ is consistent
with σ if, for each information set b reached with positive probability when
σ is used, the probability µb is obtained using the Bayes’ rule. At each
information set bki , the agent who has to move can compute her expected
utility for each possible action ai ∈ C bki . In order to do this, she needs a
probability distribution over the nodes of bki and a probability distribution
over the terminal nodes Z. The first element is provided by the probability
distribution µbki , while the second element can be computed using the strategy
profile σ. For each node n ∈ bki define π (z| n, ai , σ−i ) the probability of
reaching terminal node z starting from node n given that agent i takes action
ai and the other agents use the strategy profile σ−i . Then an action ai gives
to a player i at information set bki and holding a probability distribution µbki
the following expected utility:
Ui ai , σ−i | bki , µbki = ui (z) π (z| n, ai , σ−i ) µbki (n)
n∈bki z∈Z
18
4.2 Signaling Games
A class of extensive form games of incomplete information which has been
frequently used in finance, as well as in other fields, is that of signaling games
(also called ‘sender-receiver games’).
The simplest signaling game has the following structure:
• The type of sender is not known to the receiver, while the type of the
receiver is common knowledge. Let Θ be the set of sender’s types, with
typical element θ, and µ the probability distribution on Θ.
• Dividend games. The sender is the management of the firm, the re-
ceivers are investors. The private information held by the management
is the value of the firm. The action taken by the management is the
determination of the level of dividends. The action taken by investors
is the provision of capital to the firm.
• Capital structure games. The situation is as before, but now the man-
agement decides the mix of debt and equity to be used to finance the
firm.
19
A perfect Bayesian equilibrium in a signaling game is given by a strategy
profile σ ∗ = (σ1∗ , σ2∗ ) and a probability assessment Ψ = { µ ( ·| a1 )| a1 ∈ A1 }.
Here σ1∗ : Θ → ∆A1 gives the probability with which each action is chosen
by each type of player 1, σ2∗ : A1 → ∆A2 gives the probability with which
player 2 chooses each action given the move of player 1, and µ ( ·| a1 ) gives
the revised probability distribution on Θ given a1 . The probability µ ( ·| a1 )
is computed using the Bayes’ rule whenever possible, that is:
µ (θ) σ1∗ ( a1 | θ)
µ ( θ| a1 ) =
θ ∈Θ µ (θ ) σ (a1 | θ )
∗
whenever θ ∈Θ µ (θ ) σ ∗ ( a1 | θ ) > 0. Notice here that θ ∈Θ µ (θ ) σ ∗ (a1 | θ )
is the total probability that action a1 will be chosen by agent 1, while the
numerator is the joint probability that the type is θ and action a1 is selected.
There are two types of equilibria in signaling games: separating equilibria
and (partially) pooling equilibria. An equilibrium is separating if the action
chosen by the sender completely reveals her type. Formally:
4.2.1 Refinement
Signaling games have usually a multiplicity of perfect Bayesian equilibria.
The main reason for this multiplicity is the relative freedom in specifying
beliefs on information sets which are reached out-of-equilibrium. Imposing
reasonable restrictions on out-of-equilibrium beliefs, can help in reducing the
number of equilibria. A classic example is the quiche-beer game, proposed
5
Notice however that even in a totally pooling equilibrium, the probability distribution
may change after a certain action a1 ∈suppσ1 is taken. This happens when different types
take the action with different probability.
20
by Cho and Kreps (1986).
What are the perfect Bayesian equilibria of this game? Let us indicate
with p the probability that the second player assigns to the weak type (tw )
after observing the action ‘quiche’, and with 1 − p the probability that player
2 assigns to the strong type (ts ) after observing the action ‘quiche’. Accord-
ingly, let us indicate with q the probability that the second player assigns to
the weak type after observing the action ‘beer’, and with 1−q the probability
that player 2 assigns to the strong type after observing the action ‘beer’.
First observe that there cannot be a separating equilibrium. In a sepa-
rating equilibrium the two types of agent 1 take different actions, and agent
2 knows the type of agent 1 when it is her turn to move. Thus agent 2 will
choose N (don’t duel) when the type is strong and F (duel) when the type
is weak. But this cannot be an equilibrium, since the weak type of 1 can
pretend to be strong (use the action of strong) and be better off.
There are two pooling perfect Bayesian equilibria:
21
type. In both equilibria, agent 2 chooses N after observing the equilibrium
move (i.e., after observing Q in the first case and B in the second case). This
action is sustained in both equilibrium by assigning a probability to the weak
type of p = 0.1 and q = 0.1, respectively. For example, in the first pooling
equilibrium, N is better than F (for agent 2) since with p = 0.1 the expected
payoff of N (i.e., 0.9) is greater than the expected payoff of F (i.e., 0.1).
If an out-of-equilibrium move is observed (i.e. B in the first equilibrium
and Q in the second), then agent 2 chooses F . The way in which it is
sustained, in the first pooling equilibrium, is by assigning a high probability
to the weak type when B is observed (i.e., q ≥ 0.5) so that agent 2 chooses F .
In the same way, in the second equilibrium it is sustained by assigning a high
probability to the weak type when Q is observed (i.e., p ≥ 0.5). For example,
in the first equilibrium when the out-of-equilibrium move B is observed, the F
is preferred to N if and only if q(1)+(1−q)(0) ≥ q(0)+(1−q)(1) =⇒ q ≥ 0.5.
The first equilibrium appears to be particularly non-intuitive. The reason
why this is non-intuitive is that weak is surely worse off by choosing B with
respect to the equilibrium; the payoff in equilibrium is 3, while at best she gets
2 after the deviation. Then the weak type cannot improve on the equilibrium
payoff choosing B rather than Q. The same reasoning does not apply to type
strong. In fact, the strong type could improve on the equilibrium payoff of
2, by receiving the payoff of 3. Hence, it seems reasonable to increase the
probability of strong when B is observed. In particular, it seems reasonable
to put q = 0 that is incompatible with the pooling equilibrium (Q, Q).
On the other hand, the beliefs in the second pooling equilibrium are not
suspicious. In fact, the weak type can be better off by choosing Q with respect
to the equilibrium (in equilibrium she gets 2, while she can get 3 after the
deviation). The strong type cannot improve choosing Q with respect to the
equilibrium (she gets 3, while at most she can get 2 after the deviation).
Hence, in this case it is reasonable to assign a high probability to the weak
type when the out-of-equilibrium move is observed (i.e., Q).
Cho and Kreps (1986) have proposed the following intuitive criterion to
reduce the set of equilibria in a signaling game.
22
It can be checked that in the quiche-beer game the equilibrium in which both
types choose Q does not survive the intuitive criterion. Consequently, the
set of perfect Bayesian equilibria of the game is made of only one equilibrium
once we impose the refinement proposed by Cho and Kreps.
References
[1] Cho, I.K. and D. Kreps (1987), ‘Signaling Games and Stable Equilibria’,
Quarterly Journal of Economics, 102: 179 -221.
[2] Fudenberg, D. and J. Tirole (1991), ‘Game Theory’, MIT press, Cam-
bridge MA.
[3] Gibbons, R. (1992), ‘Game theory for applied economists’, Princeton Uni-
versity press, Princeton.
23