A Crash Course in Implementation Theory

Soc Choice Welfare (2001) 18: 655708
2001
9
9
9
9
A crash course in implementation theory

Matthew O. Jackson
Humanities and Social Sciences 228-77, California Institute of Technology, Pasadena,
CA 91125, USA (e-mail: jacksonm@hss.caltech.edu)
Received: 26 March 2001/Accepted: 21 May 2001
Abstract. This paper is meant to familiarize the audience with some of the
fundamental results in the theory of implementation and provide a quick
progression to some open questions in the literature.
1 Introduction
There are many economic, social, and political situations where individuals interact to make decisions that aect them collectively. Examples range
from voting to elect representatives or choose a public policy, to trading in
a market. Based on their preferences over the possible outcomes of the interaction, individuals may act strategically in order to inuence the outcome to
their advantage. For instance, in an election a voter might vote for his or her
second ranked candidate if the voter's favorite candidate has no chance of
being elected; or in an auction a buyer may select a bid considering tradeos between the probability of winning the auction and the price to be paid.
The specic design of the institution through which individuals interact, for
instance the rules of the election or auction, can have a profound impact on
the strategic behavior of the members of the society and on the outcomes of
the process. Implementation theory is a study of the relationship between the
structure of the institution through which individuals interact and the outcome of that interaction.
This paper grew out of lectures given at the NATO Advanced Study Institute on
Game Theory and Resource Allocation: The Axiomatic Approach, which took place
at SUNY Stony Brook in July of 1997. I thank the organizers and especially William
Thomson for organizing the institute and the participants for feedback on the lectures.
I thank Salvador Barbera for detailed suggestions an earlier version of this manuscript,
and Maurice Salles and an anonymous referee for helpful comments.
656
M. O. Jackson
Game theory plays a central role in the modeling of the strategic interaction studied in implementation theory. In many applications of game theory,
the game modeling interaction is taken as given and analyzed to predict the
actions of individuals and the resulting outcome. In implementation theory,
instead of taking the game as given, it is something to be designed. Often, one
thinks of the desired outcomes as the given and analyses whether there exist
game forms for which the strategic properties induce individuals to (always)
choose actions that lead to the desired outcomes. An example of an implementation question is: how can we design an auction to be sure that the individual who most highly values an object is sure to be the winner of the auction?1 In this view, implementation theory is a normative branch and game
theory is a positive branch of the same tree, and implementation theory is the
design or reverse engineering process associated with game theory. Of course,
this view is a bit caricatured, but indicates that there is a close relationship
between the tools and understandings developed in implementation theory
and game theory.
To get a feeling for the type of questions that are analyzed in implementation theory let us start by looking at a classic example. Consider a society or
committee holding an election to select one out of a set of candidates. Each
member of the society has a preference ranking over the candidates. The society may have certain aspirations regarding which candidate should be
selected as a function of the preferences of the members of the society. For
instance, it may wish to avoid selecting a candidate who is Pareto dominated
by another candidate (i.e., a candidate ranked lower than another candidate
by all members of society). It may also wish to select a Condorcet winner (a
candidate who defeats any other candidate in a pairwise comparison according
to a majority of voters' preferences) if such a candidate exists. If these were the
goals of the society then the implementation question would be, ``Does there
exist an election procedure for which for each possible prole of preference
rankings of the voters, each equilibrium outcome of the election procedure
would be Pareto optimal and Condorcet consistent?''
In order to answer this question one has to make precise what an election
procedure is and what equilibrium outcomes are. This is the point at which
game theoretic tools are used. The election procedure is modeled as a game
form or what is commonly referred to as a mechanism in the literature. It
species a set of possible actions or messages that each member of society
can use, and then the outcome (in this case the candidate elected) as a function of the actions or messages sent by the members of society. For instance
one could have each member of society submit their ranking of the candidates. If there are m candidates, then one could award a candidate m points
for each voter whose submitted ranking places them highest, m 1 points
for each voter's submitted ranking places them second highest, and so on.
1 For an analysis of this particular problem from a mechanism design point of view,
see Dasgupta and Maskin (1997). I will discuss the relationship and dierence between
mechanism design and implementation.
Implementation theory
657
The elected candidate is the one who has the most points, with ties broken
according to some pre-specied rule. This is the mechanism corresponding to
Borda's (1781) scoring method. Of course, it may not be in a society member's
best interest to report their true preference ranking. This is where equilibrium
concepts from game theory are used to make predictions concerning strategic
behavior. A solution concept such as Nash equilibrium can be used to predict
the preference rankings that will be reported by voters as a function of the
voters' true preference rankings. Depending on the class of mechanisms that
are admitted (e.g., how complicated we allow the message spaces to be) and
which equilibrium concept that is used, we can end up with dierent answers
concerning the possibility of selecting Pareto optimal outcomes and being
Condorcet consistent.2
In the above example a very concrete question was posed. We asked
whether it was possible to design election procedures that possess specic
properties, or in the language of the theory, whether it is possible to implement correspondences (mapping preference proles into social outcomes) with
specic properties. Also, we focused our attention on a specic setting. While
this is one way to proceed in analyzing implementation, it does not directly3
provide a general understanding of implementation that moves across dierent problems. Another way to proceed, is to take a more abstract approach
to the implementation problem and attempt to fully characterize what can
be implemented. That is, we can search for properties that precisely identify
which correspondences are implementable and which are not, and also identify implementing mechanisms. These properties can then be applied to specic settings to check whether a given correspondence can be implemented.
Such general characterization results have been produced with remarkably
little in terms of assumptions on the structure of the set of alternatives or the
feasible preferences of individuals. As we shall see, the theory has alternated
between fairly general and abstract theorems that cover a wide variety of
potential applications, and more focused and detailed theorems that apply
to particular settings or correspondences.
In what follows, I concentrate on a specic progression in the literature
and in doing so provide a biased view that often reects my own perspective
on the literature. These lectures are meant to be an introduction to the theory
rather than a survey,4 and so I do this without apology. At points I provide
2 The Borda scoring mechanism will not satisfy Condorcet consistency. Moreover, as
we shall see in Example 3, if we use Nash equilibrium as a solution concept then it will
be impossible to always select Pareto optimal outcomes and be Condorcet consistent.
However, as shown for instance in Dutta and Sen (1993) and Jackson et al. (1994), under
the solution concepts of backwards induction or undominated Nash equilibrium it will
be possible to always select Pareto optimal outcomes and be Condorcet consistent.
3 Nevertheless, I will come back to discuss why an approach of cataloguing the
answers to such narrower questions may turn out to be a useful approach to developing
the theory.
4 Surveys of various aspects of the literature may be found in Moulin (1982), Moore
(1992), Palfrey (1992, 1995), Palfrey and Srivastava (1993), Allen (1997), and Corchon
658
M. O. Jackson
opinions concerning assumptions and results, which are quite critical. This is
meant to constructively point out limitations of some of the existing results
and suggest potential directions for future research.
In what follows I assume that the reader has an introductory level knowledge of game theory (and thus has seen notions such as Nash equilibrium) and
is also familiar with some concepts from microeconomics such as preference
relations and Walrasian equilibrium.
2 Denitions and an example
2.1 Individuals
A nite group of individuals interact. N denotes both the set of individuals
and its cardinality. Generic individuals are represented as i, j, and k.
2.2 Outcomes
The set of outcomes is denoted A, and generic elements are represented as
a, b, c, d.
The set of outcomes may be nite or innite depending on the application.
For example, consider the design of a voting procedure to elect one of a nite
number of candidates. In that case, N is the set of voters and A is the nite
set of candidates. As another example, consider the design of a market where
individuals interact to exchange l dierent goods. In that case N is the set
of economic agents, and A H RNl
represents the nal allocation of goods
(including labor, leisure, consumption goods, etc.) that are possible given the
endowments and production possibilities.
2.3 Preferences
Individual i's preferences are represented by a binary relation R i over A that
is complete and transitive. The notation aR i b indicates that i weakly prefers
alternative a to b. The strict preference relation associated with R i is denoted
Pi (where aPi b if and only if it is not the case that bR i a). The notation R denotes a prole R R1 ; . . . ; RN , and R i ; R i denotes the prole where the
i-th entry of R is replaced with R i .
The set of admissible proles of preferences is the set P. Depending on the
application, P may impose restrictions on the preferences. For instance, in the
context of the exchange of private goods it may be assumed that preferences
are convex, continuous, and non-decreasing. I will sometimes refer to a prole
of preferences as being the state of the environment.
(1998). While these lectures are not meant to survey the literature, the Bibliography
that I include here is fairly comprehensive. The exception is that I do not include references to the large literature related to dominant strategy implementation as Salvador
Barbera (2001) oers lectures on strategy-proofness in this same volume.
659
2.4 Social choice correspondences

A social choice correspondence, F, maps proles of preferences into subsets of
alternatives. For any R A P, F R H A represents the set of socially desirable
alternatives when preferences are given by R. It will be assumed throughout
that F is non-empty. When F is single-valued it is referred to as a social choice
function.
In many applications, F will be a well-known correspondence, such as the
Walrasian, Lindahl, or top-cycle correspondence, or will be a social choice
correspondence derived from some normative axioms, such as the correspondence of Pareto optimal, individually rational, and envy-free allocations in a
private good exchange setting.
Generally, implementation theory seeks to characterize the set of social
choice correspondences that are obtainable as equilibrium outcomes when
individuals interact through some game form making strategic use of their
knowledge of the preference prole R. There are several perspectives that one
can take on this problem. One perspective is to begin with a specic correspondence and setting, such as the Walrasian correspondence in a privategoods economic setting, and ask whether that specic correspondence can be
implemented and if so by what mechanism. A second perspective is to work
more abstractly and characterize the set of correspondences that are implementable, by identifying conditions that they must satisfy. Both of these (and
some other) perspectives have been taken in the literature, and we shall see
examples of each.
2.5 Mechanisms
A mechanism is a pair M; g, where M M1 MN is a cross product of
message spaces and g : M ! A is an outcome function. Thus, for each prole
of messages m m1 ; . . . ; mN , gm A A represents the resulting outcome or
allocation.
A mechanism is often also referred to in the literature as a game form. The
terminology game form distinguishes it from a game, as the consequence of
a prole of messages (or actions) is an outcome rather than a vector of utility payos. Once the preferences of the individuals are specied, then a game
form induces a game. Since in the implementation analysis the preferences of
individuals vary from state to state, this distinction between game forms and
games is critical.
2.6 Solution concepts
A solution concept5 species the strategic behavior of individuals faced with
a mechanism M; g given a preference prole R. Thus, it is a correspondence
5 I use the terminology ``solution concept'' rather than `èquilibrium concept'', as
some of the concepts employed in the literature (such as single or iterative removal of
dominated strategies or maximin) do not make explicit use of `èquilibrium'' ideas, and
can predict vectors of messages that are not stable in an equilibrium sense.
660
M. O. Jackson
S that identies a subset of M for any given M; g; R specication. For the

question of what a mechanism implements, the specic messages that are predicted by the solution concept are only of intermediate interest as the corresponding set of outcomes that result is the important concept. Thus, we pay
attention to the outcome correspondence associated with a solution S represented by OS M; g; R fa A A j bm A SM; g; R s:t: gm ag.
For example, for any given M; g; R a pure strategy Nash equilibrium is
a prole m A M such that gmR i gm i ; mi for all i and mi A Mi . Denote
this set by NEM; g; R. The set of associated outcomes is ONE M; g; R
fa A A j bm A NEM; g; R s:t: gm ag.
I dene other solution concepts as they arise and similarly defer denitions
of mixed strategies until they arise.
2.7 Implementation
A social choice correspondence F is implemented by the mechanism M; g
via the solution S, if OS M; g; R F R for all R A P. F is said to be implementable via the solution S if there exists a mechanism M; g which implements it.
The above form of implementation is sometimes referred to as full implementation as it requires the exact coincidence of the outcomes of a mechanism
with the social choice correspondence.
One can also reasonably argue for a weaker form of implementation where
one only requires that OS M; g; R be a non-empty subset of F R for every
R.6 This takes the point of view that any outcome in F R is socially desirable
and so any selection is ne. Note, however, that F is weakly implementable if
and only if some sub-correspondence of F is implementable. The literature has
thus proceeded in characterizing what is implementable, which then provides
an indirect answer to the question of weak implementability.7
One key aspect of implementation (and weak implementation) is the requirement that all equilibrium outcomes lie in the given social choice correspondence. This is dierent from designing a mechanism that has one desired
outcome as an equilibrium outcome, as there may also be undesired equilibrium outcomes that are not accounted for. This points to the important distinction between the `ìmplementation'' literature and the ``mechanism design''
literature as being one of worrying about multiple equilibria.8 The mechanism
design literature focuses on incentive compatibility issues, asking whether a
given outcome can be induced as an equilibrium of some mechanism and generally ignores whether there are other equilibria. In situations where this leads
to a negative result (e.g., Myerson and Satterthwaite 1983) showing that certain outcomes cannot be sustained as the equilibrium of any mechanism, this
6 See Thomson (1996) for an argument supporting full implementation over weak.
7 This may not be an entirely satisfactory approach, as there may be conditions
for weak implementability that are easier to verify than checking whether some subcorrespondence satises the conditions for implementability.
8 See Jackson (2000) for a recent survey of the mechanism design literature.
661
approach is ne. However, in situations where it leads to positive results so

that one identies a mechanism and one of its equilibrium as satisfying some
desired criteria, then one should be very careful in interpreting the result. Can
we be sure that the given equilibrium will be played and not some other equilibrium of the mechanism?9 Are there other mechanisms that are better from
the perspective of maintaining the desired equilibrium, but not having other
undesired ones? These questions are important and can be applied to almost all
of the mechanism design literature, including the auctions and the principalagent literatures.10 This point is well illustrated in an example of Demski and
Sappington (1984), where the `òptimal'' mechanism to a principal-multiple
agent problem has a second equilibrium which makes all the agents better o
and the principal worse o.11
The multiple equilibrium issue is why the implementation problem is
still non-trivial in environments with complete information, where each individual knows the other individuals' preferences. As an example, suppose
that we consider trying to implement the Walrasian correspondence in a
classical Edgeworth setting. If we only want to make sure that Walrasian
outcomes are Nash equilibrium outcomes, and do not care about what other
equilibrium outcomes may arise, there is a simple mechanism that works.
Individuals simultaneously announce a full vector of feasible allocations. If
they all announce the same vector of allocations then that is the outcome,
and otherwise the outcome is that they each keep their endowment. Clearly
everyone announcing the same Walrasian allocation is a Nash equilibrium of
this mechanism. Unfortunately, there are many other equilibria of this mechanism which are not Walrasian allocations. The implementation question is
whether there exists a mechanism whose entire set of Nash equilibrium outcomes coincides with the set of Walrasian allocations for each preference
prole.
2.8 On the timing and information structure
Much of the implementation literature uses an idiom that refers to a social
planner, who is benevolent and selects the mechanism to implement a social
choice correspondence with a society's best interest at heart. A question that
then naturally comes to mind when thinking about the implementation problem is why doesn't the social planner adjust the mechanism once the preference prole is realized.12 Alternatively, why do the individuals in society have
9 If so, why was this not built into the solution concept to begin with?
10 For further examples and a pointed discussion of this issue, see Palfrey and Srivastava (1993).
11 A defense that is often oered for focusing a single equilibrium is that society can
direct its members to play the desired equilibrium and then they should have no reason
to deviate. This argument loses some power when there are alternative equilibria that
are preferred by some (or even all) of the members of society, especially if there is any
possibility for the members to directly or indirectly coordinate on an equilibrium.
12 See Baliga et al. (1997) for the possibility of having the planner be a player who has
preferences and can aect the outcome function as part of equilibrium play.
662
M. O. Jackson
complete information concerning each others' preferences (or at least enough

information to arrive at an equilibrium) and the social planner not? An answer
is that mechanisms represent institutions (or in some cases even constitutions),
that are meant to be long-lived. These may be costly to adjust, or it even may
be better to commit to xing the institutions not to adjust to the specics
of each realized setting. The rules for a securities market are xed and then
traders arrive with their desired trades, and similarly the rules for an election
are (usually) set before preferences of the voters are realized. Thus, mechanisms
can be thought of as representing institutions, and the implementation problem
is to characterize institutions in terms of the correspondence of outcomes they
produce as a function of the realized state of individual preferences.
Finally, while incomplete information environments, where for instance individuals know their own preferences and hold beliefs over the preferences of
others, are arguably more realistic than complete information environments,
it has proven useful to attack the problem one piece at a time. Understanding
implementation in the complete information setting has helped signicantly in
developing characterizations of implementation in Bayesian settings.
2.9 An example.
Hurwicz laid much of the foundation of the implementation problem, and
presented a simple example that is a very good one for a rst analysis of some
of the issues.
Example 1 (Hurwicz 1972). Consider a two person, two good exchange economy, represented by an Edgeworth box. Individual 1 has an endowment e 1
0; 1 and individual 2 has an endowment e 2 1; 0. Let xli denote individual i's allocation of the l-th good. So here our space of feasible allocations is A fx 1 ; x 2 A R4 jx11 x12 x 12 x 22 1g.
Consider two states of the world. In the rst state R 1 ; R 2 , the preference
relation R i of each individual is represented by a Cobb-Douglas utility function ui x x1i x 2i . In the second state, R 1 ; R 2 , the preferences of agent 1, R 1 ,
1
are represented by the utility function u1 x x 12
.
1 x11
Suppose that we are interested in implementing the Walrasian correspondence. The unique Walrasian equilibrium allocation at R 1 , R 2 is x 1
x 2 1=2; 1=2, and at R 1 ; R 2 is x 1 1=2; 7=9 and x 2 1=2; 2=9.
Clearly, in either state agent 1 prefers to have the allocation associated
with R 1 ; R 2 to that associated with R 1 ; R 2 .13 This is seen in Fig. 1.
There are a few important conclusions to draw from this example. The rst
is that a direct mechanism, or the classic one associated with a ``Walrasian
auctioneer'' who asks individuals to report their demand functions, will not
implement the Walrasian correspondence. Individual 1 is better o acting as
13 Hurwicz actually makes a stronger point. Every allocation that is Pareto ecient
and individually rational for both individuals at R 1 ; R 2 , is preferred under R 1 (and R 1 )
to the Walrasian equilibrium allocation at R 1 ; R 2 .
663
Fig. 1.
if he had preferences R 1 instead of R 1 when the state is actually R 1 ; R 2 . This

also implies that any sort of incentive compatibility condition will be violated
by the Walrasian correspondence (or any selection from it) when agents' preferences are private information.
Hurwicz points out another consequence of this line of reasoning. A
mechanism that implements14 the Walrasian correspondence must make use
of information of agent 2 concerning the preferences of agent 1, as agent 1 can
otherwise pretend to be of type R 1 . Thus, Hurwicz argues that no ``privacy
preserving'' mechanism can implement the Walrasian correspondence, where
privacy preserving is an appropriate formalization of the idea that the mechanism only incorporates knowledge of an individual about his or her own
preferences.
Let us now explore a related question in the context of the same problem
in order to get a better feeling for the implementation problem. We know that
a mechanism that implements the Walrasian correspondence in this example
must make some use of individual 2's knowledge of individual 1's preferences.
So we might ask, is that enough? Is there a simple mechanism that Nash implements the Walrasian correspondence in this two state example?
The following is the direct mechanism associated with simply asking individuals to announce their own preferences. The rst entry in each row is
individual 1's allocation and the second is individual 2's allocation.
Clearly, there is a unique strictly dominant strategy equilibrium in either
state for individual 1, which is to say R 1 , and so such a direct mechanism
14 I am being vague about the solution concept at the moment, but the statement is
true for most solutions concepts.
664
M. O. Jackson
Table 1.
Individual 2
R2
x 1 12 ; 12, x 2 12 ; 12
R1
Individual 1
x 1 12 ; 79, x 2 12 ; 29
Table 2.
Individual 2
Left
Individual 1
R1
R
Right
x 1 12 ; 12, x 2 12 ; 12
1
x 0; 1, x 1; 0
2
2
7 1
x 1 11
18 ; 3, x 18 ; 3
x 1 12 ; 79, x 2 12 ; 29
cannot Nash implement the Walrasian correspondence. This is just an illustration of Hurwicz's point.15
Let us now expand individual 2's strategies to include two possible announcements ``Left'' and ``Right''. We will design the mechanism so that it
is a strictly dominant strategy for individual 1 to say R 1 when the state is R 1 ,
and R 1 when it is R 1 . Moreover, this can be done so that it is a strict best response for individual 2 to say ``Left'' if individual 1 is playing R 1 , and to
say ``Right'' if individual 1 is saying R 1 .
For the mechanism in Table 2, the unique Nash equilibrium in state R 1 ; R 2
is R 1 ; Left which results in the Walrasian allocation for that state, and similarly for the state R 1 ; R 2 the unique Nash equilibrium is R 1 ; Right. This
mechanism thus implements the Walrasian correspondence (for this simple
two state environment) in Nash equilibrium. The mechanism also implements
the desired correspondence in the iterative elimination of strictly dominated
strategies, and most other solution concepts.
As is consistent with Hurwicz's privacy preserving point, it was necessary
for us to have individual 2 play a role even though only 1's preferences vary.
One thing to note about the above mechanism, is that individual 2 does not
simply ``report'' what she knows about individual 1. Individual 2 would have
an incentive to say that it was state R 1 , even if it was state R 1 . Thus, the agents'
incentives in this regard are opposed. Instead, allowing individual 2 to have a
non-trivial message space allows us to create a richer mechanism; and in particular to exploit switches in individual 1's preferences between the two states.
For instance, the ``R 1 , Left'' entry is one that individual 1 prefers to ``R 1 , Left''
15 This also illustrates that no mechanism can implement the Walrasian correspondence in dominant strategies, as it would have to take this form when looking only at
the strategies used based on these preferences.
665
in state R 1 but not in state R 1 . Thus, the o-diagonal entries are critical to the
working of the mechanism and were not chosen by chance. This begins to hint
at an important necessary condition for Nash implementation that I discuss
next.
The contrast between the workings of the two mechanisms above points
out the importance of considering a large class of mechanisms for the implementation problem, and most notably to include mechanisms that do more
than simply ask individuals to report their preferences. Although the mechanism in Table 2 can be regarded as a direct mechanism if one thinks of asking each individual to report the state, the entries in the o-diagonals are still
critical to successful implementation. Moreover, there are problems where implementation can only be achieved by mechanisms that are more complicated.
In order to develop a deeper understanding of this, let us now analyze Nash
implementation.
3 Nash implementation
The seminal work on Nash implementation by Maskin (1999),16 not only
provides us with an understanding of what is implementable in Nash equilibrium, but it also provides a blueprint for the techniques and approach that
underlie many of the general characterization results in the literature. Thus, it
is useful to study Nash implementation in some detail.
3.1 Monotonicity two ways
Maskin identied an intuitive necessary condition for Nash implementation
that he called monotonicity.17 This condition may be expressed in two dierent ways. It is a trivial exercise to see that the statements are equivalent.
Nevertheless, I present each of them separately as these expressions correspond to dierent expressions in the implementation problem, and it is useful
to emphasize each of these points of view.
Suppose that a social choice correspondence F is Nash implementable.
What can we deduce about F ? Its implementability implies that there exists
an implementing mechanism M; g. So consider a preference prole R and
an alternative a A F R. Since F is Nash implementable there exists a prole
of actions m A M such that m is a Nash equilibrium at R and gm a.
Next, suppose that there exists another preference prole R such that a B
F R. The fact that M; g Nash implements F then implies that m cannot
be a Nash equilibrium at R. Thus, there must exist an agent i and a deviation mi that i prefers to mi at R i . That is, there must exist i and mi such
16 The paper circulated as a working paper from 1977 to 1998.
17 This condition previously appeared in the social choice literature under the name
strong positive association (e.g., see Muller and Satterthwaite 1977). The name monotonicity, however, stuck and has recently been referred to as Maskin-monotonicity by
several authors. This condition diers from the monotonicity condition discussed in the
context of social welfare orderings (e.g., see Moulin 1988 for a review of some results
regarding monotonic social welfare orderings).
666
M. O. Jackson
b
Fig. 2a,b.
that gm i ; mi Pi gm. Since m was a Nash equilibrium at R, it must be that

gmR i gm i ; mi . Letting b gm i ; mi , we have reasoned that F must satisfy the following condition.
A social choice correspondence F is monotonic if for any R A P, R A P, and
a A F R such that a B F R, there exist i and b such that aR i b and bPi a.
This condition is represented in Fig. 2, above.
In Fig. 2, the set of allocations is in R2 for individual i. The monotonicity
condition states that if a A F R but a B F R, then there exists some i for
whom either the indierence curves through a corresponding to R i and R i
cross (Fig. 2a)18 or the upper contour set of R i through a is a superset of the
upper contour set of R i through a (Fig. 2b).
We have reasoned the following result.
Theorem 1 (Maskin 1999). If a social choice correspondence F is Nash implementable, then F is monotonic.
To get some practice understanding monotonicity, the reader can return to
the implementing mechanism in Example 1 (Table 2) and verify that player 1
18 Although in the gure, the indierence curves cross at a, they can cross at any
point and may cross more than once. As noted in Maskin (1999), the Spence-Mirrlees
single-crossing condition is stronger than monotonicity. For a study of the relationship
between monotonicity and single crossing properties see Arya et al. (1998).
667
and the o-diagonal entries satisfy the monotonicity conditions that are necessary for implementation.
To develop a fuller understanding of monotonicity, let us examine the
condition from a dierent, but formally equivalent perspective.
^
Consider a preference prole R and alternative a A F R. Suppose that R
^
is such that for each i if aR i b, then aR i b. This says that a's standing in i's
^ so that a is still weakly
preference ranking has has not fallen from R to R,
^ i to each b it was weakly preferred to under R i . From Nash
preferred under R
implementability we know that there exists m which is a Nash equilibrium at
^ i for any i and potential m
^ i . Given that
R. This implies that gmR i gm i ; m
^ i b, it must be that gmR
^ i gm i ; m
^ i for any i and potential
aR i b implies aR
^ and so a A F R.
^ This leads to
^ i . Thus, m is also a Nash equilibrium at R
m
the following statement of monotonicity.
A social choice correspondence F is monotonic if for any R, a A F R, and
^ such that for each i and b aR i b implies aR
^ i b, a A F R.
^
R,
Thus, monotonicity requires that if a A F R and for each i the upper
^ i through a is a subset of the upper contour set of R i through
contour set of R
^
a, then a A F R.
This condition is represented in Fig. 3, below.
One statement of monotonicity follows the reasoning that if an alternative
is to be implemented at one prole but not another, then it must have fallen in
someone's rankings in order to break the Nash equilibrium via some deviation. The other statement of monotonicity follows the reasoning that if an alternative is implemented at one prole and rises in each individual's rankings
at another preference prole, then prole of actions leading to the alternative
which form a Nash equilibrium at the rst prole must still be a Nash equilibrium prole at the second prole. These conditions are obviously equivalent
as one is simply the contra-positive of the other. Nevertheless, it is still useful
to consider both statements. The rst emphasizes that there must exist some
preference reversal if an equilibrium at one prole is to be broken at another.
This suggests natural relationships with various preference crossing properties,
such as the Spence-Mirrlees condition. The second emphasizes that if the
Fig. 3.
668
M. O. Jackson
standing of an equilibrium alternative improves it must remain an equilibrium

outcome, which is often a useful way of checking whether monotonicity is
satised and provides useful bridges to understanding the connection to conditions such as strategy-proofness.
3.2 Sucient conditions for Nash implementation
Monotonicity alone is not a sucient condition for Nash implementation, but
it is sucient together with an auxiliary condition called no-veto power when
there are at least three individuals.
A social choice correspondence F satises no veto power if whenever i, R,
and a are such that aRj b for all j 0 i and all b A A, then a A F R.
No veto power is a condition that is quite weak in some contexts, such as
environments with some private goods when there are at least three (nonsatiated) individuals. In such a context, each individual would prefer to have
more of the private good and so there is never a single alternative that is most
preferred by more than one individual. However, the condition is more restrictive with 2 individuals, as then it requires that each individual's favorite
outcome be in the correspondence; and in contexts such as voting environments where there is no private good, as then it is possible to have a number
of individuals agree on a favorite outcome and no-veto power ignores how
poorly that outcome may rank under a remaining individual's preferences.
Theorem 2 (Maskin 1999). If N b 3 a social choice correspondence F satises
monotonicity and no veto power, then it is Nash implementable.
It is instructive to sketch a proof of this theorem, as it provides insight into
the approach used to prove many of the suciency theorems in the literature.19
Generally, it is easy to design a mechanism that has the desired outcomes
as Nash equilibria. It is more dicult to rule out undesired outcomes. To see
this, consider the following trivial mechanism: if at least N 1 individuals
name the same alternative then that is the outcome, otherwise some arbitrary
b is the outcome. Notice that a unanimous announcement of any a is a Nash
equilibrium for any preference prole. However, every alternative is an equilibrium outcome for such a mechanism regardless of the preference prole.
This points to the importance of the multiple equilibrium problem. One way
to view the mechanism below is to start with such a trivial mechanism and
then modify it using monotonicity to rule out undesired equilibria.
Consider a social choice correspondence F which satises monotonicity
and no veto power and assume that Nb3. Consider the following mechanism.
19 Although the theorem was rst stated in Maskin (1999) (circulating as a working
paper in 1977), the rst proof was by Williams (1986). William's proof required some
additional assumptions on the setting, and the rst complete proof of the theorem as
stated is in Saijo (1988). The proof given here is adapted from Repullo (1987) and
Moore and Repullo (1990). Karl Vind was the rst to suggest this structure of proof,
and a mechanism that was a precursor to this one, in a discussion of Maskin's work
in the 1970's (see Groves 1977, Footnote 6).
669
Mi P A N where N is the set of nonnegative integers. Thus, each

individual announces a preference prole, an alternative, and an integer.
Dene g as follows:
(1) If m1 m2 mN R; a; n and a A F R, then gm a.
(2) If there exists i such that mj R; a; n for all j 0 i, where a A F R, and
mi ; b; where mi 0 mj , then gm b if aR i b and gm a if bPi a.
(3) For any other m label mi ; ai ; ni and let i be the lowest indexed i
such that ni b nj for all j 0 i, and then gm ai .
To paraphrase the mechanism:
(1) applies if all individuals make exactly the same announcement, and
then the announced alternative a is the outcome of the mechanism.
(2) applies to situations where if all but one individual make exactly the
same announcement, and some individual i makes a dierent announcement.
In that case, the outcome is the b announced by that individual i if it is no
better than the alternative a announced by the others under the preference for
i announced by the other individuals. Otherwise the outcome is a.
(3) applies the remaining announcements that do not fall under (1) or
(2). Here we identify the individual who announced the highest integer (and
break ties according to individual labels) and the outcome is the alternative
announced by that individual.
Let us verify that this mechanism Nash implements the desired F. Let R
be the true preference prole.
First, let us check that for any a A F R it is a Nash equilibrium for all
agents to announce mi R; a; 0. A unilateral deviation by any i results in an
m i ; mi that falls in (2), and then can only result in a b such that aR i b. Thus,
no i can gain by deviating and so m is an equilibrium.
Next, let us check that every Nash equilibrium results in some a A F R.
If m is a Nash equilibrium that falls into (3), then it must be that gm is
the most preferred outcome of all agents, since otherwise any agent i could
deviate and announce an integer higher than the other agents and select any
outcome ai . Thus, by no veto power it must be that gm A F R. If m is a
Nash equilibrium that falls into (2), then any agent j 0 i (where i is as dened
in (2)) could unilaterally deviate and choose an action such that g would be
determined by (3) and j i (simply by announcing an integer higher than
any other agent and any desired aj ). Thus, for such an m to be an equilibrium, it must be that gm is most preferred by all agents j 0 i, and again
no veto power ensures that gm A F R. Finally, consider the case where m
is a Nash equilibrium and g is determined by (1). If the preference prole R
is announced truthfully, then the outcome must be in F R. So consider the
~ a; n and a A F R
~ where R
~ 0 R. Any
case where m1 m2 mN R;
~ i b (by simply
individual i could deviate to obtain any b 0 a such that aR
~i b
~ i under (2)). Thus, it must be that if aR
announcing ; b; which puts m i ; m
then aR i b. Thus, by monotonicity a A F R.
This basic structure of this mechanism underlies the construction of mechanisms in many of the constructive proofs for other solution concepts too. The
670
M. O. Jackson
basic idea is that there is a possibility for complete agreement which falls under (1). This allows outcomes in F R to be sustained as equilibrium outcomes. As it is possible for the individuals all to announce the same R; a; n
even when the state is R and a B F R, part (2) of the denition of g allows
monotonicity to work so that some i can deviate and announce R; b; n and
obtain b. Part (3) then allows any individual other than i to obtain any outcome they like, which rules out equilibria that fall under (2), unless b is most
preferred by all j 0 i in which case by no veto power, b A F R.
As discussed previously, no veto power is not a necessary condition and
can be restrictive. So it is worthwhile to identify a set of conditions that is
both necessary and sucient for Nash implementation. Such conditions have
been obtained by Moore and Repullo (1990). Although these conditions are
more complicated, they do provide additional insight and I refer the reader to
the Moore and Repullo (1990) paper for more detail.20
Theorem 2 requires that N b 3. The case of N 2 is of obvious importance as there are many bilateral interactions that one would like the theory
to handle. Interestingly, there are non-trivial dierences between the case of
N 2 and N b 3. To get a rough idea of the additional considerations that
appear for the case of N 2, notice that the mechanism used to prove Theorem
2 makes use of the fact that when N 1 agents announce the same thing, then
one can identify an i who is demanding a dierent alternative as in (2) of
the denition of g. With 2 agents this is not possible and an additional necessary condition appears, which arises from considerations of what happens
when agent 1 plays a strategy that should be played in one state and agent 2
plays a strategy that should be played in another state. Characterizations for
the case of N 2 appear in Dutta and Sen (1991b) and Moore and Repullo
(1990). While the full characterization is complex, an intuitive sucient condition called the ``non-empty lower intersection condition'' appears in Dutta
and Sen (1991b), and I refer the interested reader there for details.
3.3 The restrictiveness of monotonicity
There are aspects of monotonicity that make it a strong and restrictive
condition, and a condition that is more easily motivated as a necessary condition of Nash implementation than by direct normative argument. Consider
^ i b. Monotonicity
^ such that for each i and b aR i b implies aR
a A F R and R
^
requires that a A F R. Two aspects of the strength of this condition are as
follows. First, it is allowed for a to be strictly preferred to b under R i and in^ i . Thus, there is a limited sense in which a could have
dierent to b under R
^ and yet monotonicity still requires that
gotten worse in moving from R to R,
^
a lie in F R. The reasoning based on Nash implementation makes it clear as
to why this is a necessary condition, but this makes it a condition that will be
20 See Danilov (1989) and Yamato (1992) for elegant full characterizations in some
specic settings and Sjostrom (1991) for an algorithm for verifying the necessary and
sucient conditions.
671
Fig. 4.
violated by a number of social choice correspondences that have some strong

^ pays no attennormative backing. Second, the comparison between R and R
tion to what happens to the relative ranking of alternatives other than a,
(except those implied by comparisons to a). These aspects mean that monotonicity is violated by many natural voting methods as is illustrated below.
To get a more complete impression of the implications of monotonicity,
let us now explore in some detail whether it is satised by some specic wellknown social choice correspondences.
Let us begin with the classical exchange environment referred to in Hurwicz
(1972) and consider whether the Walrasian correspondence satises monotonicity. In Example 1 monotonicity was satised. However, that was a very
restrictive setting and in fact monotonicity is not satised by the Walrasian
correspondence on the full classic domain of preferences. This was shown in
the context of the following example by Hurwicz et al. (1995).
Example 2 (Hurwicz et al. 1995). Consider a two person, two good Edgeworth
economy.21 There are two states R 1 ; R 2 and R 1 ; R 2 , with the preferences as
pictured in Fig. 4 above. The allocation a is a Walrasian equilibrium at R 1 ; R 2 ,
but not at R 1 ; R 2 .
The diculties that appear in the gure above have to do with the boundary
of the Edgeworth box. In order to satisfy monotonicity, we need to nd b such
21 Although the gure is simply for 2 individuals, it is easily adapted for 3 or more
individuals.
672
M. O. Jackson
that aR 1 b and bP 1 a. The allocation b necessary to satisfy this requires an

amount of good 2 for player 1 that exceeds the total amount of good 2 available in the economy. The problem is that the crossing of preferences required
by monotonicity takes place outside of the Edgeworth box, and outside of the
set of allocations that are feasible for a mechanism to provide.
The way around this problem suggested by Hurwicz et al. (1995) is to
modify the Walrasian correspondence to include a as an equilibrium at R 1 ; R 2 .
This new correspondence, called the ``constrained Walrasian correspondence''
satises monotonicity and is Nash implementable when there are 3 or more
individuals.
P
P
To be precise let A fx A RNl j i xki a i eki Ekg. Say that x is a constrained Walrasian equilibrium allocation if there exists p A Rl such that p x i
a p e i for each i and xR i y for all y A A such that p y i a p e i .
The constraint in the denition is in the requirement that y A A. In the
usual denition of a demand correspondence, one requires that a demanded
point be weakly preferred to all other points in the budget set. In this constrained denition, it is only required that a demanded point be weakly preferred to the points in the budget set that are also feasible given the resources in
the economy. Thus, a 1 is not in the demand correspondence of player 1 at R 1 ,
but it is in the constrained demand correspondence. So, the allocation a is a
constrained Walrasian equilibrium at R 1 ; R 2 , and so the diculty with monotonicity is overcome. Hurwicz, Maskin, and Postlewaite (1995) show that the
constrained Walrasian correspondence is monotonic on a classical domain of
preferences, and it is contained in any upper-hemicontinuous social choice
correspondence that is Pareto ecient, individually rational, and monotonic.
Next, let us examine monotonicity in a voting context.
Example 3. Let A be a nite set of candidates, and consider the following
preference prole, with alternatives listed in order of preference.
R 1 a; b; c, R 2 c; a; b, and R 3 b; c; a.
Also consider R 3 c; b; a.
Let F R fa; b; cg and F R 3 ; R 3 fcg. F agrees exactly with the choice
of most scoring rules (e.g., Borda), and most any rule that is anonymous and
neutral (e.g., the top cycle correspondence, uncovered correspondence, Copeland rule, Simpson rule, etc.), as well as plurality rule. Unfortunately, F is not
monotonic since the standing of a has not changed at all in anyone's preference ranking and yet a is in F R, but not in F R 3 ; R 3 .
This means that none of the voting correspondences that are commonly
used in political elections are Nash implementable!22 This suggests that sequential rationality or other solution concepts are needed for the implementation of voting correspondences.
To see the restrictiveness of monotonicity in the voting context from
22 Some well-known (but often quite ``thick'') correspondences such as the Pareto
correspondence or union of favorite alternatives, however, are monotonic and Nash
implementable.
673
another perspective, note that the Muller-Satterthwaite (1977) theorem states

that if A is a nite set, P includes all strict preference orderings on A, and
N b 3 then any social choice function that is monotonic is necessarily dictatorial (where a dictator is a single agent who picks his favorite alternative
from the range of the function).
As we have seen, although monotonicity can be satised, it is in general
a strong condition. Many interesting and desirable social choice correspondences are not Nash implementable. In view of this, various alternatives to
Nash implementation have been investigated in detail. One possibility is to
use alternative equilibrium concepts, an idea we will turn to later. Generally,
using stronger equilibrium concepts (i.e., more rened ones) allows for more
things to be implemented. Usually the more dicult challenge in implementation comes from ruling out undesired equilibrium outcomes, rather than
ensuring that desired outcomes are equilibrium outcomes. Stronger solution
concepts aid in ruling out undesired equilibria. The second possibility that
has been explored is to place additional structure on the set of alternatives
and then to require some sort of approximate, or what has become known as
virtual implementation. I discuss this next.
4 Virtual implementation
Virtual implementation works under the assumption that if one cannot exactly
implement a desired correspondence, then one should be willing to implement
a correspondence that is arbitrarily close to the desired one. In order to dene
what is meant by ``close'', we need additional structure on the set of alternatives. The way in which this has been done in the virtual implementation
literature is to let A be the set lotteries over some primitive set of alternatives.
Then we can say that we are implementing something ``close'' to a desired
alternative if we end up with a lottery that places suciently high probability
on the desired alternative.
If we cannot Nash implement a correspondence we can ask whether there
is a correspondence suciently close to it, in the lottery sense, that can be
Nash implemented. More pragmatically, if a correspondence is not monotonic, is there a correspondence close to it that is monotonic? There is no
reason to expect that the answer to this question should be yes. However, the
work on virtual implementation has coupled this approximation question together with additional conditions. The answer turns out to be yes if one restricts attention to certain types of correspondences called `òrdinal'' ones, or
if one restricts attention to situations where individuals' preferences have von
Neumann-Morgenstern representations. Both of these additional conditions
place signicant structure on the problem as we shall see and discuss below.
Following with the literature on virtual implementation, let us assume in
this section that there is some underlying set of alternatives A 0 that is nite,
let K
and that the set alternatives A is the set of lotteries on A 0 . More formally,
P
denote the number of alternatives in A 0 . Then A fx A 0; 1 K s:t: k xk 1g:
674
M. O. Jackson
This structure on the set of alternatives allows us to dene distance among

two elements of A through Euclidean distance. This induces a notion of closeness of social choice correspondences:
Two social choice correspondences F and H are e-close if for every R A P
there exists a bijection (one to one and onto function) t : F R ! HR such
that ja taj < e for each a A F R.
So, two correspondences are e-close if for each preference prole there is
a one-to-one mapping so that each lottery in one correspondence is within e
of its corresponding lottery in the other correspondence. While this may seem
like an innocuous denition, it may be that two lotteries that are e close to
each other are not even roughly equivalent to a society that cares about expost realizations. This is a discussion that we will return to after seeing how
this denition is used.
A social choice correspondence F is virtually implementable if there exists
a social choice correspondence G that is e-close to F and is Nash implementable.23
We can now dene the ordinality condition mentioned above, which is an
invariance condition on F.
Any given preference relation, R i on A, allows us to compare pure outcomes, which correspond to lotteries that provide probability 1 to some alternative, thereby inducing a preference relation on the set of underlying alternatives A 0 . For any R i , denote this preference relation by R pi . In what follows,
assume that R pi is a strict ordering.24
F is ordinal if F R F R whenever R pi R pi for all i.25
A preference relation R i is monotone if there exists an ordering of A 0 ,
a1 ; . . . ; aK under R pi such that
(i) a1 Ppi a2 . . . Ppi aK , and
P
P
(ii) if x A A and y A A are such that kaM xk b kaM yk for any M a K
and x 0 y then xP i y.
Thus, R i is monotone (not to be confused with monotonic) if a lottery that
puts more weight on more preferred alternatives is preferred to one that puts
comparatively less weight on more preferred alternatives. This is a condition
that is satised by preferences that have a von Neumann-Morgenstern representation, but is more general and allows for a richer set of preferences.
Now we are ready to state a result that contains the main insight of the
virtual implementation literature.
23 One can dene other notions of virtual implementation by substituting some other
solution concept in the place of Nash equilibrium.
24 This simplies the exposition. One can weaken this somewhat, as should be clear
in what follows, although one cannot allow for complete indierence. Details can be
found in Abreu and Sen (1991).
25 Ordinal is a bit of a misnomer. It requires that F ignore information contained in
how preferences vary on A rather than just on A 0 . Such information is not necessarily
``cardinal'' information.
675
Theorem 3 (Matsushima 1988, Abreu and Sen 1991).
. Any ordinal social choice correspondence dened on a domain of monotone

.
preferences is e-close to a social choice correspondence that satises monotonicity.

Any social choice correspondence dened on a domain of von NeumannMorgenstern preferences26 is e-close to a social choice correspondence that
satises monotonicity.
Let me outline a proof for each of these results, as the ideas are fairly
straightforward. A formal proof, although notationally heavy, can be easily
eshed out by the reader.
1
First, let c^ be the lottery that places equal weight on every alternative in
K
A 0 . We construct G as follows. For any a A A consider the lottery 1 ea e^
c.
Transforming any lottery to this corresponding lottery we are sure to have an
`ìnterior'' lottery, or one that puts weight on each alternative. This mapping
denes a G that is e-close to F. The proof is completed by showing that given
either of the restrictions, such an interior G27 is necessarily monotonic. Consider any R, and R, and 1 ea e^
c A GR such that 1 ea e^
c B GR.
First let us treat the case where F is ordinal and preferences are monotone.
Since GR 0 GR, it follows from the ordinality of F that we can nd i such
that Rpi 0 Rpi . So there are alternatives d A A 0 and d A A 0 such that dPpi d and
1
1
d
dPpi d. From the denition of c^, the lottery b 1 ea e^
c
d
2K
2K
is well-dened and in fact interior.28 Since preferences are monotone, it follows that 1 ea e^
cP i b and bP i 1 e e^
c. Thus, G is monotonic.
For the case where preferences satisfy the von Neumann-Morgenstern
axioms, the proof is quite similar. Note that if GR 0 GR, then there is some
i for whom R i 0 R i . Then, given the linearity of von Neumann-Morgenstern
preferences, we can nd lotteries e A A and e A A such that ePi e and ePi e.
This is pictured in the following gure for the case where A 0 consists of three
alternatives.
1
1
e again fullls the requirements to
Letting b 1 ea e^
c
e
2K
2K
ensure that G is monotonic.
In this setting, given that every social choice correspondence satisfying
suitable conditions is close to one that is monotonic, is any such social choice
26 Here the social choice correspondence is dened on the domain of preferences and
not on the domain of utility functions. If one wants to use the set of von NeumannMorgenstern utility functions as the domain, then the social choice correspondence
must be constant across utility functions that are ane transformations of each other.
27 The reader can check that the following argument is easily adaptable to show that
any interior social choice correspondence is monotonic on these domains.
28 Here d and d represent the lotteries given weight one to d and d. These convex
combinations can be made directly as A is a Euclidean simplex.
676
M. O. Jackson
Fig. 5.
correspondence virtually implementable? The answer is yes when N b 3.29

One can easily modify the environment so that no veto power is always
satised. Simply take A to be the set of all lotteries on A 0 , but omitting each
pure lottery (that places weight one on some alternative). Then no veto power
is satised trivially, as there is no best lottery for any individual since with
monotone preferences their favorite lottery is a pure lottery. The above theorem then leads to the following corollary.
Corollary 1 (Matsushima 1988; Abreu and Sen 1991). Let N b 3. Any ordinal
social choice correspondence dened on a domain of monotone preferences is
virtually implementable. Any social choice correspondence dened on a domain
of von Neumann-Morgenstern monotone preferences is virtually implementable.
While virtual implementation provides for a remarkable conclusion, it
comes at the expense of strong assumptions.
First, in considering virtual implementation, one is implicitly assuming
that society is willing to settle for implementing something that is e-close to
the desired social choice correspondence. This raises the question of the expost rationality of virtual implementation.30 Virtual implementation requires
that we play with small probabilities on lotteries that may have nothing to
do with the desired social choice correspondence, but can be used to distinguish preferences (the d 's and e's in the above proof ). In particular, with small
29 For the case of N 2 one needs to add a non-empty lower intersection condition
which corresponds to the extra condition needed for two-person Nash implementation.
Details appear in Abreu and Sen (1991).
30 Many mechanisms for implementation can be criticized for the same reason. This
applies to the mechanisms for subgame perfect and undominated Nash implementation
presented in Sect. 6, and is something I discuss in more detail in Sect. 7.
677
probabilities these may turn out to be the outcome of the mechanism. In order
for the virtual implementation arguments to be valid, agents must take these
small probabilities seriously and base decisions on them, with full expectations
that these outcomes will be enforced if they happen to be selected by the lottery. It is perfectly possible that an outcome which is bad for all agents is randomly selected, and the arguments underlying virtual implementation require
that the agents cannot change this outcome (for example, by renegotiating to
some other outcome).
Second, the assumptions of von Neumann-Morgenstern preferences or of
ordinality coupled with monotone preferences are strong ones and critical to
the arguments. Von Neumann-Morgenstern preferences are assumed in many
economics models because of their nice linearity with respect to probabilities,
which allows for a very tractable analysis. However, we should always be cautious in interpreting results that rely critically on that linearity, rather than just
using it for tractability. In other words, if our results are robust to variations
in the set of preferences then working with von Neumann-Morgenstern preferences for tractability's sake is ne. However, if the results depend in a special way on a restriction to the domain of von Neumann-Morgenstern preferences, then we should be cautious in applying the results. Arguably, virtual
implementation relies critically on the linearity of preferences (or ordinality of
F coupled with monotone preferences31) because it implies that the crossing
conditions needed for monotonicity will be satised at all interior points! Once
one allows for slight amounts of non-linearity in preferences over lotteries,
then the crossing conditions are not automatically satised, and monotonicity
once again becomes restrictive.
5 Renements
Another avenue in the theory is implementation under various renements of
Nash equilibrium.32 As we saw in the Nash implementation theorems, it is
easy to ensure that a given desired outcome is an equilibrium outcome while
it is dicult to rule out undesired equilibria. This is where the necessity of
the monotonicity condition arises and what limits our ability to implement.
31 Ordinality coupled with the monotone preference condition plays a similar role to
the linearity. If F R 0 F R, then some agent's preferences must dier over at least
two underlying alternatives in A 0 . Then at any interior lottery there must be a crossing
of this agent's indierence curves, as substituting a slight amount of one of these two
underlying alternatives for the other will be ranked in opposite ways by the two preferences.
32 When I use the word renements, I use it literally to mean a solution concept which
always selects a subset of Nash equilibria, such as undominated Nash equilibrium, subgame perfect equilibrium, coalition-proof Nash equilibrium, strong equilibrium, and
trembling hand perfect equilibrium. Implementation in other solution concepts that are
not renements of Nash equilibrium, such as protective criterion or iterative elimination of weakly dominated strategies also appear in the literature (e.g., Barbera and
Dutta 1982 and Abreu and Matsushima 1994), but are not discussed here.
678
M. O. Jackson
Looking at renements helps in ruling out undesired equilibria and leads to

more permissive results. However, with such additional power will come some
cautions about how the renements are exploited.
Let us begin with a discussion of implementation in undominated Nash
equilibrium, as was rst explored by Palfrey and Srivastava (1991). This is a
simple and natural renement of Nash equilibrium, that adds to Nash equilibrium the requirement that no individual play a weakly dominated strategy.33
This solution concept is also useful for illustrating the power of implementation via a renement of Nash implementation, and some of the cautions that
go along with it.
Given a mechanism M; g, an individual i, and a preference relation
for i, R i , a message mi A Mi is weakly dominated by mi A Mi at R i if
gmi ; m i R i gmi ; m i for all m i A M i and gmi ; m i Pi gmi ; m i for some
m i A M i . A message mi A Mi is undominated at R i if there is no other message which weakly dominates it. Let
UDM; g; R fm j mi is undominated at R i Eig:
A prole of messages m A M is an undominated Nash equilibrium of
M; g at R A P if m A NEM; g; R X UDM; g; R. So let UNEM; g; R
NEM; g; R X UDM; g; R.
The denition of undominated Nash implementation is then that F R
OUNE M; g; R for all R A P.
In order to keep the discussion of undominated Nash implementation relatively uncluttered, let us impose the following assumption on the domain of
preferences. A slightly weaker value distinction condition appears in Palfrey
and Srivastava (1991), and the reader is referred there for details. This condition allows us to provide a unied exposition of a number of results.
Strict value distinction. The domain P satises strict value distinction if for
every R A P and R A P with R 0 R:
(I) For all i, there exists a A A and b A A such that aPi b,
(II) For all i, if R 0 R i there exists a A A and b A A such that aPi b and
bPi a.
(I) simply rules out complete indierence, and (II) says that if an individual's preferences change, then they change in a non-trivial way so that
some alternatives switch ranking (rather than just becoming indierent). So
(II) simply requires that if preferences change then indierence curves cross
somewhere, which is a condition that is vacuously satised in most contexts
of interest. For instance, strict value distinction is satised on a domain of
strict preferences (linear orders), as well as any setting where preferences are
33 Such equilibria are a superset of trembling hand perfect equilibria (except for N 2
where they coincide, see van Damme 1987), and thus have nice existence properties,
but also have the advantage of being easier to dene and work with in such abstract
environments as those arising in the implementation literature.
679
upper-semi continuous and locally non-satiated34 and so applies to almost

any setting of interest.
Theorem 4 (Palfrey and Srivastava 1991). If N b 3 and strict value distinction
is satised, then any social choice correspondence that satises no veto power is
implementable in undominated Nash equilibrium.
The contrast of this theorem with Theorem 2 is dramatic, as any social
choice correspondence in economic settings as well as most voting rules are
admitted. The renement to undominated Nash equilibrium has thus made a
big dierence in which social choice correspondence are implementable, completely eliminating the necessity of monotonicity. The main intuition is that
any switch in preferences (one is guaranteed to exist by value distinction) can
be used to change the set of strategies that are undominated from one state to
another. The central diculty in Nash implementation that is handled by the
monotonicity condition is that there can be an alternative a that one wants to
be a Nash equilibrium outcome in one state but not in another state. So one
needs a preference reversal between a and some other alternative b that can be
obtained through a deviation. When one considers undominated Nash implementation, one instead needs a to be supported as an undominated Nash
equilibrium outcome in one state R but not in another R. Let m be an undominated Nash equilibrium message prole at R of some mechanism. In
order to rule m out as an undominated Nash equilibrium at R, it must be that
either m is no longer a Nash equilibrium or some m i is weakly dominated at
R i (while it was undominated at R i ). One can assure that m i is weakly dominated at R i and not at R i , simply by taking advantage of any dierence in the
preferences between R i and R i . Of course, one has to be careful in working
out the details so that such dierences are accounted for in each state while
still producing the desired undominated Nash equilibria in each state, which
results in a complex mechanism.
The idea that domination arguments can take advantage of relatively slight
dierences in preferences,35 can be pushed further. In fact an even more permissive theorem holds with a weaker solution concept. We can drop no veto
power, weaken the solution concept to undominated strategies, and include
the case of N 2.
Theorem 5 (Jackson 1992). If strict value distinction is satised, then any social
choice correspondence is implementable in undominated strategies.
The proof of each of these theorems involves mechanisms that are quite
intricate. The underlying principles are similar to that of the mechanism used
to prove Theorem 2: if there is agreement in the announcements then the out34 See Jackson (1992) for a proof.
35 You may notice some analogies to way in which slight dierences in preferences
over lotteries are expoited by virtual implementation. The same criticisms regarding
the ex-post rationality of enforcing the outcomes can be made, and I discuss this in
more detail below.
680
M. O. Jackson
come is implemented (as in (1) in the mechanism proving Theorem 2), there
is a possibility for some individuals to deviate and allow for some ``test'' pair
of alternatives which serve the role of changing which strategies are weakly
dominated, and there is some use of an integer game (as in (3)) to rule out
certain undesired congurations of strategies. However, the mechanisms are
more complex in terms of how and when the test pairs apply in order to take
advantage of weak dominance arguments. Rather than exhaust the reader with
the details of the mechanisms, I present a simple example that illustrates an
implementing mechanism for Theorem 5 in a specic case. The example is
also useful for another purpose: it points out a serious caveat that we should
have regarding implementing mechanisms.
Example 4 (Jackson 1992). Let N 2 and A fa; bg. R1 R 2 is such that
aPi b, and R1 is such that bP1 a.
Consider F R1 ; R 2 fbg and F R1 ; R 2 fag. This is a peculiar social
choice correspondence as it goes exactly against the preferences of the agents.
But it is then convincing36 that if we can implement this social choice correspondence then we can implement any social choice correspondence. The following mechanism does the trick.
m1
m1
m2
b
b
b
b
..
.
~2
m
a
a
b
b
..
.
^2
m
a
a
a
b
..
.
a
a
a
a
..
.
a
a
a
a
..
.

a
b
b
b
..
.
a
b
b
b
..
.
a
b
b
b
..
.
a
b
b
b
..
.
a
b
b
b
..
.

a
a
a
a
..
.
b
a
a
a
..
.
b
a
a
a
..
.
b
a
a
a
..
.
b
a
a
a
..
.
...
...
...
...
b
a
a
a
..
.
b
b
a
a
..
.
b
b
b
a
..
.
b
b
b
b
..
.
b
b
b
b
..
.
...
...
...
...
.
In this table, the and .. indicate a countable continuation of the series
of entries and corresponding strategies.37 Let us check that the mechanism
implements the stated social choice correspondence. First, m 2 is undominated
for 2 since it is the only strategy that results in a against m 1 . Second, any
strategy in the right half of the mechanism for individual 2 is dominated by
~ 2 . Next, m
~ 2 is dominated by m
^ 2 . Next, m
^ 2 is domiindividual 2's strategy m
nated by the strategy immediately to its right; and so forth. This leaves m 2 as
36 The hard-to-convince reader is referred to Jackson (1992) for the proof of Theorem 5.
37 Each individual here has a message set that is divided into two sets that are both
countably innite. So for instance, m 2 gets b against every message in 1's rst set and a
against every message in 1's second set.
681
the only undominated strategy for individual 2. A similar analysis for player 1
leads to a unique undominated strategy of m 1 at R 1 , and of m 1 at R 1 . Thus,
the mechanism implements the stated F.
The elimination of dominated strategies is a questionable practice in the
~ 2, m
^ 2 ; . . . ; each strategy is
above mechanism. For instance, in the sequence m
eliminated by the next, but there is no undominated or `best' strategy in that
string. We end up predicting that individual 2 will not play any of these strategies, but will play m 2 instead, even though it does worse in some situations
than any of the eliminated strategies.38
As shown in Jackson (1992) this is not simply a problem when dealing with
implementation in undominated strategies, but also holds for undominated
Nash implementation. This example can be modied to have three individuals
and implement essentially the same social choice function in undominated Nash
equilibrium, via a mechanism with similar diculties. Moreover, in each of the
examples the same is true of every implementing mechanism.
To address the questionable removal of dominated strategies in Example
4, we can rule out the use of such mechanisms by requiring that the implementing mechanism be bounded. This requires that when a strategy is dominated, it be dominated by some undominated strategy.39
5.1 Bounded mechanisms
A mechanism M; g is bounded relative to P if for every i and R A P,
whenever mi A Mi is dominated at R i , then it is dominated by some message
mi A Mi that is undominated at R i .
This denition makes reference to the preference domain, and so boundedness is a condition relative to an environment. A nite mechanism is bounded
regardless of the environment due to the transitivity of the domination relation.
However, an innite mechanism's boundedness property may depend on the
domain of preferences in question.
If we require implementation via a bounded mechanism, then Theorem 5
no longer holds and necessary conditions arise.
F is strategy-resistant if for all i, R A P and R i such that R i ; R i A P
and for each b A F R i ; R i , there exists a A F R (possibly a b) such that
aR i b.
38 As an aside, the use of the integer game in the mechanism used for Nash implementation (i.e., proving Theorem 2) may have bothered you. As no domination arguments were used there, the issues are slightly dierent than that in Example 4. However,
the integer game used in proving Theorem 2 had a similar feature in that it could produce situations where agents have no best response (i.e., if they have beliefs that have
positive probability that some other individual will use any integer). I will return to
discuss this issue after discussing the domination issues.
39 An alternative approach, would be to modify the concept of domination so that
weakly dominated strategies are eliminated only if they are dominated by some undominated strategy. (Timothy Van Zandt suggested this alternative approach in a discussion of Jackson 1992). As is easily seen from the proof of Theorem 6, the same
restrictions on implementation ensue with either approach.
682
M. O. Jackson
This condition has a avor of strategy-proofness, but is dened for correspondences. If we think of i at R, considering a manipulation by acting in
accordance with R i , we nd that for whatever outcome i may hope for, b A
F R i ; R i , there is some outcome a A F R which is at least as good as b. This
condition is very strong: it reduces to strategy-proofness when F is a function.
The following theorem illustrates the impact of requiring boundedness in
implementation via undominated strategies.
Theorem 6 (Jackson 1992). If a social choice correspondence is implementable
in undominated strategies via a bounded mechanism, then it is strategy-resistant.
The proof of this theorem is quite easy. Let F be implemented in undominated strategies via the bounded mechanism M; g. Consider i, R A P
and R i such that R i ; R i A P and consider any b A F R i ; R i . Since F is
implemented, there exists m A UDM; g; R i ; R i such that gm b. Either
b A F R, in which case let a b, or it must be that mi is dominated by some
~ i ; m i . Since m
~ i ; m i is undomi~ i that is undominated at R i . Let a gm
m
nated at R, it follows that a A F R, and from the denition of domination, it
follows that that aR i b.
Thus, we end up with a stark contrast between what is implementable in
undominated strategies when we can use any mechanism and when we can
only use bounded mechanisms. This suggests it is very important to understand the restrictions imposed by considering only mechanisms for which a
given solution concept is appropriate.
Boundedness ends up restricting the class of social choice correspondences that are implementable in undominated Nash equilibrium as well. That
is studied in Jackson et al. (1994). It turns out to introduce a necessary condition (the chained condition) that is weaker than monotonicity and strategyresistance, but nonetheless rules out some well-known correspondences. The
reader is referred to that paper for details.
While boundedness is a condition that is natural to require when we examine elimination of dominated strategies, or solution concepts like undominated Nash equilibrium or iterative elimination of weakly dominated
strategies, it is not an obvious condition to require when considering Nash
equilibrium, or subgame perfect equilibrium, where the elimination of dominated strategies is never an issue. However, there are conditions that one
might like to impose on mechanisms so that such solution concepts are reasonable.
5.2 Mixed strategies
The denition of Nash equilibrium that we worked with in proving Theorem 2,
does not consider mixed strategies. Even giving a proper denition of mixed
strategy equilibrium requires additional structure as preferences have to be dened over lotteries on alternatives, and since message spaces may be uncountably innite one has to take care. Without tackling these technical issues, let us
683
consider an example that shows that worrying about the possibility of mixedstrategy equilibria makes a dierence.
What we can show is that if one Nash implements a social choice correspondence by a nite mechanism, then there may still exist mixed strategy
Nash equilibria that result in outcomes outside of the social choice correspondence and which may be preferable from the individuals' perspective.
Example 5 (Jackson 1992). Let N 2 and A fa; b; c; dg. Preferences are
described by aP 1 bP 1 cI 1 d (where I 1 indicates indierence), aP 2 bP 2 cP 2 d and
bP 2 aP 2 cI 2 d.
Consider F such that F R1 ; R 2 fag and F R1 ; R 2 fdg. F is Nash
implemented by the following mechanism.
1
m
^1
m
~1
m
m2
d
d
d
^2
m
c
a
b
~2
m
c
b
a
Note however, that at R 1 ; R 2 for most denitions of preferences over

lotteries, there also exists a mixed strategy equilibrium that results in a and b
with positive probability. This is true for every nite implementing mechanism.
Claim 1. Suppose that players' preferences over lotteries on A are continuous,
convex, and monotone. Any mechanism with a nite number of strategies for
each player has a Nash equilibrium in mixed strategies at R 1 ; R 2 resulting in
outcomes a or b with positive probability, and which strictly Pareto dominates
the outcome d.
The proof of the claim is simple. Suppose that a mechanism M; g Nash
^ i implies that there exists
^ i H M i such that m
î A M
implements F. Consider M
i
i
^
^ A fa; bg. Thus, M is found by eliminating strategies that
^ such that gm
m
can only ever lead to c or d. Since F is implemented, there must be some
^ is non-empty. Given the as^ such that gm
^ a, and so M
strategy prole m
sumptions on preferences and the niteness of the mechanism, there exists a
^ g at R 1 ; R 2 . Given the denition
Nash equilibrium in mixed strategies to M;
^
of M and the fact that preferences on lotteries are monotone, it follows that
this equilibrium must put positive probability on at least one of the outcomes
a and b, as these are reachable by some strategy of 1 given any (mixed) strategy of 2, and are preferred by 1 to c and d over which 1 is indierent. Note
^ is then also a Nash equilibrium of M; g at R 1 ; R 2 since strategies
that m
i î
in M nM only lead to c or d and cannot provide improving deviations. This
establishes the claim.
The above claim shows that the standard denition of Nash implementation that only considers pure strategies can be a problematic one. Let me make
two remarks about the above example.
First, the social choice correspondence that is being implemented is a bit
strange in that we are trying to implement d when it is a least preferred outcome for both of the individuals. However, this might be quite natural in the
684
M. O. Jackson
context of a problem such as a principal/multiple-agent problem, where

the two individuals above represent the agents, the principal designs the
mechanism and prefers that the agents take actions leading to outcome a in
the rst state and d in the second state.
Second, the niteness of the mechanism is critical to the result. If we
allowed for innite mechanisms, then we could ensure that there are no mixed
strategy equilibria, by employing an integer game.40 Maskin (1999) shows
that one can design a mechanism for Nash implementation that has no mixed
strategy equilibria (see the Appendix of his article); but his mechanism involves an integer game and thus an innite message space. Example 5 shows
that there are correspondences in nite settings that are Nash implementable
when one only considers pure strategies, but not Nash implementable (except
by innite mechanisms) when one allows for mixed strategies.
For Example 5 it is natural to require nite mechanisms as the setting is
completely nite. However, many settings of interest are naturally innite and
so requiring implementation in nite mechanisms might not always make sense.
So one might ask, ``What is the correct restriction on mechanisms is for Nash
implementation if one accounts for mixed strategies?'' It should be a condition that naturally rules out improper use of integer games, without unduly restricting the message space or outcome function when innite spaces might be
appropriate.
The bothersome aspect about integer games is that they are appended
to mechanisms precisely because there is no Nash equilibrium to them, even
though players could very well try to announce higher integers than others in
the hopes that they will get their most-preferred alternative. We have no prediction for how a player will act if he or she believed that the other players
would be announcing integers. A player's best response correspondence is not
well-dened when that player faces a mixed strategy of the others that places
weight on an innite set of integers. So, one way to rule this out is to look only
at mechanisms for which best response correspondences are well-dened.
Jackson et al. (1994) suggest such a requirement, calling it the ``best response
property.'' The best response property is satised by any nite mechanism
and also by many innite mechanisms. Nevertheless, it is a strong condition
and rules out some mechanisms that are well-behaved in the way that they
either admit or rule out mixed strategy equilibria.
The same point regarding mixed strategies in this example that was made
in Claim 1 above, can be made with regard to subgame perfect implementation.41
40 We saw such a construction in the mechanism used to prove Theorem 2. Under (3),
the outcome was that announced by the agent who named the highest integer. Clearly
as long as there is some disagreement over most preferred outcomes, then there is no
equilibrium in (pure or) mixed strategies to the integer game.
41 Moore and Repullo (1988) provide a mechanism for subgame perfect implementation that has no mixed strategy equilibria. However their mechanism also employs an
integer game and fails the best response property.
685
Claim 2. Suppose that players' preferences over lotteries on A are continuous,

convex, and monotone. Any extensive form mechanism42 that has a nite
number of terminal nodes and that implements F, has a subgame perfect equilibrium in mixed strategies at R 1 ; R 2 which results in outcomes a or b with positive probability and which strictly Pareto dominates the outcome d.
The proof of Claim 2 is similar to that of Claim 1. We rst need to trim the
^ g in the previous proof. In the extensive
tree to obtain a game form like M;
form this is a bit trickier, but the niteness allows us to do this inductively.
Start with information sets that only precede terminal nodes, and eliminate
actions that only lead to c or d at those information sets. If all actions are
eliminated, then replace the information set with terminal nodes that lead
to d. Next, proceed to information sets that lead to either terminal nodes or
information sets considered in the rst step, and perform the same trimming.
Proceed inductively until the root is reached. Find a subgame perfect equilibrium of the remaining extensive form. It must lead to a or b with positive
probability. Then construct a subgame perfect equilibrium of the original
game form leading to the same outcomes, by picking any action at information sets that were completely eliminated in the trimming procedure. This then
establishes the claim.
The diculties with mixed strategies pointed out in the claims above,
can be attacked from dierent directions. One approach to handling this
problem is to examine virtual implementation to see whether the diculties
can be avoided there. Abreu and Matsushima (1992a) have analyzed virtual
implementation via the iterative elimination of strictly dominated strategies.
Since the set of Nash equilibria (pure or mixed) has support in the set of
strategies that survives the iterative elimination of strictly dominated strategies, if one obtains a unique prole of actions via the iterative elimination of
strictly dominated strategies, then there must be a unique Nash equilibrium
(even considering mixed strategies). Abreu and Matsushima show that with
three or more individuals, any social choice function is virtually implementable in iteratively undominated strategies by a nite mechanism. Their construction of an implementing mechanism is very clever and quite intuitive. I
refer the reader to their paper for details. While they obtain a powerful result handling mixed strategies, it comes at the expense of the strength of the
assumptions in the virtual implementation setting (as discussed above).43
6 Specic environments
Another approach to accounting for mixed strategies, and also to providing
implementation by mechanisms that are bounded/and or simple and natural
42 The denition of extensive form mechanism is the same as that of an extensive
form game, except that there is an outcome associated with each terminal node rather
than a payo. As the denition is standard, available in most any game theory text,
and notationally cumbersome, I omit it here.
43 Glazer and Rosenthal (1992) criticize the Abreu and Matsushima mechanisms on
complexity grounds and Abreu and Matsushima (1992c) oer a reply.
686
M. O. Jackson
in other ways, is to examine implementation in specic environments. Part

of the reason that the mechanisms used in the suciency theorems are so
abstract and complicated is that they implement a given social choice correspondence without using any detailed information about the specic restrictions on preferences that may be satised in an environment. If we take
advantage of knowledge about the structure of the environment, the implementation problem can be simplied.
One important class of environments to be considered is that of economic
environments where there is a private good. In such environments there is a
disagreement between agents' preferences that allows a mechanism to play
agents against each other each agent prefers to have more of the private
good. Also, there is some possibility of imposing ``nes'' or punishments that
are agent specic. Let us examine two results on implementation in such settings that employ simple and intuitive mechanisms that are bounded and
handle mixed strategies. The rst such result applies to undominated Nash
implementation and environments that are called separable.
6.1 Separable environments
(1) there exists w A A such that aP i w for all i, R A P, and a A F R,
(2) for any a A A, J H N, and R A P, there exists a J such that a J I i w for any
i A J and a J I i a for i B J,
(3) R i 0 R i implies that there exist a and b such that aP i bP i w and bP i aP i w.
(1) is a requirement that there is an outcome that each agent considers to
be worse than any alternative in the range of the social choice correspondence.
(This means that the denition of a separable environment is relative to a
given social choice correspondence.) For instance, in an exchange economy
this would be satised by setting w to be the 0 allocation, provided the social
choice correspondence leaves each agent with some minimal consumption in
every state. (2) is the requirement that captures the separable nature of the
environment. It states that it is possible to reduce the allocations of some set
of individuals J while not aecting the remaining individuals. This is satised
in exchange economies that satisfy free disposal, as well as other settings such
as those with public goods and/or externalities provided there is at least one
disposable private good that agents value suciently. (3) is a version of strict
value distinction again, with the additional requirement that the test pair of
alternatives (a and b) be preferred to the bad outcome. Again, it is satised
in many economic settings.
Theorem 7 (Jackson et al. 1994; Sjostrom 199444). If the environment is
separable relative to a social choice function F, then F is implementable in
undominated Nash equilibrium by a bounded mechanism that has no mixed
strategy equilibria.
44 The denition of separable environments is from Jackson et al. (1994). Although
Sjostrom's (1994) setting is slightly dierent, the intuition behind his mechanism is
identical.
687
Theorem 7 is proven using the following mechanism.

Each individual i announces either a pair of preference relations m i
i
R ; R i1 (interpreted as being an announcement i and i 1's preferences respectively, letting N 1 1), or a pair of alternatives m i a; b. There are
three cases to consider.
(I) All individuals announce pairs of preferences. In this case, let J be the set
of i such that m2i m1i1 . This is the set of i whose announcement about
i 1 matched what i 1 announced about him or herself. The outcome
is then F m2 NnJ .45
(II) Some i announces m i R i ; R i1 and all other j 0 i announce m j
a; b such that aR i w and bR i w (with all j 0 i announcing the same
a; b). Then gm a Nnfig if aR i b and gm b Nnfig otherwise (where
a Nnfig and b Nnfig are as dened in (2) of separable environments).
(III) For any other conguration of messages gm w.
Since an individual's announcement about his or her preferences can
only aect his or her allocation in (II), it is clear that the only undominated
announcement of preferences is to announce one's own preference truthfully.
Also, since by announcing a pair of alternatives one always obtains an alternative equivalent to w, any announcement of a pair of alternatives is dominated by a truthful announcement of preferences. Finally, given that all individuals should be announcing pairs of preferences, and announcing their
own truthfully, it then follows from (I) that each individual should announce
their neighbor's preference truthfully. This is the unique undominated Nash
equilibrium. The mechanism is bounded because the dominated actions were
dominated by undominated actions.
Thus, in this separable class of environments we can obtain a strong implementation result with a relatively simple bounded mechanism. Before discussing some of the shortcomings of the mechanism above, let me discuss a
closely related setting and mechanism for implementation in subgame perfect
equilibrium.
A special case of separable environments is that of quasi-linear preferences.
That setting allows for a simple, clean, and intuitive implementation in subgame perfect equilibrium.
Let us say that an environment has quasi-linear preferences that are representable by bounded utility functions if there exists A 0 such that
P
(1) A fa; t 1 ; . . . ; t N A A 0 RN j i t i a 0g, and
(2) each R i is represented by a utility function u i : A 0 ! 0; 1 such that
a; t 1 ; . . . ; t N R i b; t 1 ; . . . ; t N if and only if u i a t i b u i b t i .
Condition (1) says that allocations can be written as a product space of
some primitive allocations as well as a vector of private transfers. Note that
45 So this is the outcome where individuals in J get F evaluated at the R corresponding
to that in the second entry of each agent's message (what they announced about their
neighbor), and individuals outside of J get an outcome equivalent to the bad outcome w.
688
M. O. Jackson
P i
the transfers are freely disposable since
i t a 0. Condition (2) states that
preferences are representable by a bounded quasi-linear utility function.
Let us consider social choice functions for which F R a; 0; . . . ; 0
for every R A P. This is almost without loss of generality, since A 0 is arbitrary and if one wishes to specify transfers that do not sum to 0, for instance specifying the sharing of the cost of a public good as a function of R,
then this can be included in the specication of A 0 and the t i 's can be thought
of as additional transfers. The only loss of generality comes from condition
(2) above, which implies that any transfers included in A 0 must be bounded
in size. This bound is needed for the implementing mechanism below to
function.
Theorem 8 (Moore and Repullo 1988). Consider an environment with a nite
number of states, quasi-linear preferences represented by utility functions that
are bounded, and N 2. Any social choice function is implementable in subgame perfect equilibrium by an extensive game form of perfect information of
nite length that has no mixed strategy equilibria.
Although the theorem is stated for N 2, the mechanism below can be
extended to N > 2 (see Moore and Repullo 1988 and Moore 1992 for details).
Theorem 8 is established through the mechanism pictured below. To understand the mechanism we note two points. First, the reader can easily check
that for any quasi-linear utility functions on A 0 , u i and v i that represent different preferences on A, we can nd a; t i and b; t i with 0 b t i and 0 b t i
such that
u i a t i > u i b t i
while v i a t i < v i b t i :
Thus, in the challenge phase of the mechanism pictured below, if agent i has
announced v i and is challenged by the other agent who counters with u i , then
such a pair is invoked. Second, we can rewrite the social choice function to be
a map from proles of bounded utility functions on A 0 into allocations. So,
the implemented social choice function can be written to depend directly on
the announcements of u 1 and u 2 .
The intuition behind the mechanism is straightforward. Suppose that the
true prole of preferences corresponds to u 1 ; u 2 . If 1 announces a false preference v 1 instead of u 1 , then individual 2 can challenge 1 and say u 1 . If 1 is
challenged then 1 will end up paying a large ne no matter what occurs in the
sequel. 1 is then oered a choice between the a; t i and b; t i that distinguish
u 1 and v 1 . 1 will choose honestly, as this is the end of the tree and 1 pays a ne
regardless of the choice. If 1 chooses a; t i then this indicates that 1 was not
truthful in the rst stage, and so 2 is rewarded by receiving the ne that 1 pays.
If 1 chooses b; t i then this indicates that 2 was not truthful (either in challenging or in 2's announcement of u 1 ) and then 2 also pays a ne. This ne is
suciently large (outweighing any implemented allocation) that 2 would like
to challenge 1 if 1 lies, and not otherwise. The same reasoning applies to the
other section of the tree and so the unique subgame perfect equilibrium out-
689
Fig. 6.
come is for both individuals to announce utility functions that represent their
true preferences and the correct outcome is implemented.46
While the results in this section have very optimistic implications, they
come at the expense of strong restrictions. Most importantly, one of the criticisms made of virtual implementation that it may not be credible to believe
that bad outcomes will stand is also particularly important here. In both of
46 The niteness of the state space is important in guaranteeing existence of equilibrium, as it ensures that there exists a well-dened best response of the ``challenging''
agent in the challenging subgames.
690
M. O. Jackson
the mechanisms above there is some ``bad'' outcome (w in the undominated

Nash implementation and a ne large enough to outweigh all other considerations in the subgame perfect implementation) and the mechanisms need to
commit to these bad outcomes in order to sustain the implementation. There
are papers that address this shortcoming, which I discuss below.
Finally, while the mechanisms above are relatively simple and intuitive,
bounded, and avoid mixed strategy equilibria; the mechanisms are still cumbersome in the sense that they require announcements of preferences (or utility
functions). Thus, although they are well-behaved in a theoretical sense they
are still far from ``natural'' in the sense that they could be easily used in practice. So, although they are useful in delineating the boundary of what can be
implemented, we should want to push further to require implementation by a
mechanism with a simple or natural message space.47 By adding even more
structure to the problem, one can obtain natural mechanisms with simple
action spaces. The tractability of specic environments for obtaining intuitive simple mechanisms for implementation has been taken advantage of in
allotment, bargaining, voting, principal-agent, and public goods environments
in a series of papers.48 The implementing mechanisms that are obtained
often have auction and voting like features that make them very simple and
natural.
One specic environment that has received a great deal of attention in the
implementation literature is that of the exchange of private goods, and the
implementability of the constrained Walrasian correspondence.49 Within this
context one can examine implementation with respect to certain message
spaces. Can the constrained Walrasian correspondence be implemented by
mechanisms that only involve announcements of prices and quantities or
announcements of prices and allocations? These questions and others are
addressed in recent papers by Saijo, Tatamitani and Yamato (1996), Dutta et
al. (1995), and Sjostrom (1995b), which map out exactly what sorts of message
spaces are needed for Nash implementation of the Walrasian correspondence.
47 One approach to addressing the complexity of announcements of the mechanism

is to examine the dimensionality of the message spaces and look for mechanisms
with minimal sized message spaces and to establish lower bounds on what is necessary for implementation in various environments. For explicit discussion of the size
and complexity of mechanisms for decentralization see the seminal paper of Mount
and Reiter (1974). Papers focusing on this issue in implementation (when incentives
are also an issue) include Reichelstein and Reiter (1988), Saijo (1988), and Hong and
Page (1994).
48 See, for example, Glazer and Ma (1989), Jackson and Moulin (1992), Rubinstein
and Wolinsky (1992), Dutta and Sen (1993), Glover (1993), Thomson (1994), Bag
(1996, 1997), Jackson and Palfrey (2001), and Brusco and Jackson (1999).
49 The large literature on this subject begins with Hurwicz (1972), and includes papers
that provide increasingly well-behaved mechanisms worrying about auxiliary properties of the mechanism like continuity of the outcome function, individual rationality,
and balance out of equilibrium. See for example, Schmeidler (1980), Hurwicz et al.
(1995), Wettstein (1992), Nakamura (1990), and Hurwicz (1996).
691
7 Other topics and open questions

While the discussion of implementation theory above exposes the reader to
some of the results, techniques, and themes in the literature, there is a substantial portion of the literature that I have not discussed. In this section I
briey discuss some of the issues that are addressed in the rest of the literature
as well as some interesting open problems.
7.1 Implementation under incomplete information
There is an extensive literature on implementation when individuals hold private information. Nash implementation has a natural analog in such settings
which is referred to as Bayesian implementation.50
When individuals may hold private information (that is not known collectively to the other individuals in the society), an incentive compatibility condition becomes necessary. In a complete information setting if I ask agents to
tell me the state and the announcements do not all agree, then I can be sure that
someone is lying. In an incomplete information setting, where only one person
knows the state, there is no way to check whether that person is lying and so it
must be in their interest to announce the state in a direct revelation mechanism.
This discussion is in terms of direct revelation mechanisms, but it implies that
an incentive compatibility condition is necessary in incomplete information
settings that is not necessary (or alternatively is vacuously satised) in complete information worlds. As we know from the mechanism design literature,
incentive compatibility conditions can turn out to be quite restrictive, rendering many ecient and desirable social choice functions non-implementable.51
Once one adds incentive compatibility as a necessary condition, we can use
our understanding of implementation in complete information settings to make
signicant progress on implementation in incomplete information settings. For
instance, the reasoning behind the necessity of monotonicity for Nash implementation extends to provide an analogous necessary condition for Bayesian
implementation that has been called Bayesian monotonicity in the literature
(e.g., see Postlewaite and Schmeidler 1986, Palfrey and Srivastava 1989a, and
Jackson 199152). Although the Bayesian monotonicity condition is more
complex, the intuition underlying it is similar to that behind monotonicity.
While the necessity of an incentive compatibility and a Bayesian monotonicity condition would be expected53, the Bayesian setting also introduces
50 Bayesian implementation may be seen as a generalization of Nash implementation,
as it reduces to Nash implementation in situations where agents are symmetrically informed.
51 See Palfrey and Srivastava (1987) and (1993) for a number of examples in the context of Bayesian implementation.
52 How one denes the setting and the Bayesian monotonicity condition is important
for identifying a condition which is both necessary and sucient for implementation.
See Jackson (1991) for a discussion of this point.
53 A closure condition which links play of dierent equilibrium strategies across the
common knowledge partition is also necessary, and quite intuitive.
692
M. O. Jackson
new considerations that render the task of nding sucient conditions signicantly more complex than in the complete information setting. For instance, one cannot state a direct analog of Theorem 2 in a Bayesian environment. In particular, a no-veto style condition and Bayesian monotonicity need
to be carefully intertwined respecting the information structure, as discussed
in Jackson (1991).54 There are other interesting dierences that arise in a
Bayesian setting: Dutta and Sen (1994) demonstrate a simple social choice
function in a nite setting that requires a mechanism with innite message
spaces for implementation.
While progress has been made in understanding the conditions that
characterize Bayesian implementation, the full implications of the Bayesian
monotonicity condition and the extent to which it may be satised are still less
well understood. Palfrey and Srivastava (1987, 1993) make signicant headway in showing that many analyses in mechanism design and agency theory
that use incentive compatibility and invoke the revelation principle suer from
multiple equilibrium problems and the identied social choice functions fail
to satisfy Bayesian monotonicity condition. They also show that in environments with transferable utility the conditions for Bayesian implementation
are more easily satised.55 A recent paper by Serrano and Vohra (1999) sheds
more light on the restrictiveness of Bayesian implementation by showing that
Bayesian monotonicity is essentially an ordinal condition, and so implemented
social choice correspondences must be constant across dierent cardinal representations of underlying ordinal preferences. What is left open is a detailed
understanding of what can be implemented in non-transferable utility settings
when we worry not only about incentive compatibility, but also about full
implementation and thus the multiple equilibrium problem and Bayesian
monotonicity.
While Nash implementation has an obvious generalization to environments with incomplete information,56 subgame perfect implementation does
not have a unique generalization, but instead several. This is due to the variety of alternative formulations of sequential rationality under incomplete
information. In particular, varying assumptions about how individuals update beliefs o the equilibrium path results in alternative solution concepts to
be used in implementation. The two incomplete information extensive form
notions of implementation that have been analyzed are perfect Bayesian
implementation (Brusco 1995, 1997, 1998) and implementation via sequential
equilibrium (Baliga 1999 and Bergin and Sen 1998). The interesting new aspect that arises in these settings is that preference reversals can arise not only
54 See Dutta and Sen (1991c) for more on the necessary and sucient conditions for
Bayesian implementation.
55 See also Matsushima (1993).
56 Other natural extensions of complete information implementation to incomplete
information settings have been analyzed as well, such as implementation in undominated Bayesian equilibrium (Palfrey and Srivastava 1989), virtual Bayesian implementation (Duggan 1997a), and virtual implementation in iteratively undominated strategies with incomplete information (Abreu and Matsushima 1992).
693
because of dierences in the primitive underlying preferences, but also from

the way in which information is revealed in equilibrium through the extensive
form. This is the focus of the work of Bergin and Sen (1998). While various
conditions for implementation have been identied, the role of information
revelation through an extensive form is not yet fully understood.57
7.2 Ex post individual rationality, renegotiation, and credibility
At several points I have mentioned that various forms of implementation rely
on the belief that the outcomes of the mechanism will be enforced, even if they
are ``bad'' from society's point of view ex-post. This can be problematic, to
the extent that the positive results depend on such outcomes being used by
the mechanism and such beliefs holding.58 If, for example, a mechanism is
constructed to assist bargainers in reaching an ecient agreement, then it is
questionable to assume that highly inecient outcomes will be allowed to
stand o (or on) the equilibrium path.
There are several papers that consider implementation in the face of
individual rationality and/or renegotiation on and o the equilibrium path.
Ma et al. (1988) were the rst to point out the importance of imposing an individual rationality constraint both in and out of equilibrium. They examined
a principal-agent model where the usual individual rationality constraint
(imposed only on the equilibrium path) was replaced by an `òpt-out,'' where
each player had the ability to decline the outcome of the mechanism and accept a status-quo outcome instead. Maskin and Moore (1999) examined a
more general implementation problem, and changed the opting out to a possibility of renegotiation. This takes the point of view that instead of settling
for some outside option or status-quo, the individuals involved are likely to
renegotiate to some ecient outcome. Maskin and Moore considered implementation where any outcome of a mechanism that suggests a Pareto dominated allocation is replaced by a Pareto ecient allocation according to an
exogenous renegotiation function.59 Given any such renegotiation function,
Maskin and Moore obtain characterizations of Nash and subgame perfect
implementation that have intuitive relationships to the standard characterizations. Although these two papers provide insight into ex-post individual
rationality and renegotiation, much depends on the exogenous specication
of the outside options or renegotiation function.
One would like to model the process that occurs when players opt-out,
rather than take as given a status quo or renegotiation function. Jackson and
Palfrey (1998) push in this direction, in the context of a dynamic bargain57 For instance, see Brusco (1998) for a puzzle on the necessity of multi-stage mechanisms.
58 See Hurwicz (1994) for a discussion of some other, related issues related to enforceability in mechanism design.
59 See Rubinstein and Wolinsky (1992) for a dierent approach. Their notion of
``renegotiation-proof '' implementation requires that the possibility for renegotiation
never arise on or o the equilibrium path.
694
M. O. Jackson
ing and matching model,60 by having a player be rematched with a new bargaining partner whenever he or she opts-out of a prescribed alternative. They
show that although such an endogenous individual rationality constraint is
compatible with eciency within individual matches, it can be incompatible
with eciency from society's point of view accounting for the overall evolution
of the market which requires specic exercise of the rematching option. In other
words, it can be impossible to implement the ecient rule in such a setting.
The above approaches are based on the idea that the individuals themselves
are not bound to the mechanism, but have the ability to opt out of the prescribed
outcome either to some status quo, renegotiated outcome, or replay of the
mechanism. There might be other contexts where such an opt-out can be prevented and the outcome can be made legally binding. Nevertheless, one still has
to worry about whether the planner (or society at large) will let inecient outcomes stand ex-post. Studies which address such ``credibility'', or the ability of
the mechanism designer or planner to commit to o-equilibrium-path outcomes
that are known to be undesirable, include Chakravorti et al. (1992), Baliga et al.
(1995) and Baliga and Sjostrom (1995). These studies include the planner as a
player in the mechanism,61 in which case one can explicitly account for the
planner's preferences and behavior with regards to enforcing an outcome.
In all of the above work there are two forces at work. On the one hand,
allowing for movement away from ex-post undesirable outcomes can be improving just by itself since truly undesirable outcomes are eliminated automatically. On the other hand, this limits out of equilibrium threats that (as
we saw in the last section) can play a strong role in selecting which outcomes
are implemented. Thus, although allowing for players or the planner to alter
outcomes builds in some minimal individual rationality or eciency, it can
come at the expense of selectivity of the mechanism. While the above cited
work makes progress in understanding enforceability and credibility issues, a
complete understanding of these issues requires full modeling (endogeneity)
of the process that occurs when an outcome is opted away from by the players
or the planner. Often this process may be situation specic, and so it should
be worthwhile to analyze this issue in a variety of specic environments, which
is something this branch of the literature is just beginning to do.
7.3 Mixed strategies
We do not yet have a good understanding of what is Nash or subgame perfect implementable when one accounts for mixed strategies, by mechanisms
satisfying, for instance, the best response property.62 I discussed one approach
60 See Jackson and Palfrey (2001) for a unied approach to dealing with renegotiation,
outside options, and replay of the mechanism, in more abstract settings, with some
applications to exchange economies.
61 Including the planner as a player has an interesting theoretical byproduct: the
planner is part of the equilibrium and can thus rule out undesired equilibria unilaterally.
62 As mentioned in the discussion following Example 5, there are mechanisms that
have no mixed strategy equilibria; but those mechanisms involve integer games and fail
to have well-dened best response correspondences.
695
to addressing this problem which is to analyze specic environments where the

additional structure can lead to simple implementing mechanisms that avoid
mixed strategy equilibria and other diculties altogether. Such an approach
has been very successful leading to some of the most natural mechanisms for
implementation to date. So it seems fruitful to simply break the implementation analysis into many subproblems by analyzing specic settings and developing a catalog of what is implementable setting by setting. However, a case
can still be made for attacking the problem head-on in the general setting. The
reason for complementing the setting-specic analyses with a general analysis
is that it would generate insight to the additional necessary (and sucient)
conditions that come with the consideration of mixed strategies. Given how
tailored the analyses are in specic environments, it is not clear that such an
understanding will be obtained by piecing together a catalog of results.
A number of possible head-on approaches suggest themselves, even simply
with respect to Nash implementation. One might ask what can be (exactly)
Nash implemented in nite environments by a nite mechanism that has no
mixed strategies.63 Alternatively, one could allow for mixed strategies and
consider implementing social choice correspondences that map into lotteries
and ask what is implementable.
7.4 Robustness of mechanisms
In order to push the theory towards applied use, one has to worry about how
sensitive a mechanism is to such obvious things as misspecication of the environment or ìrrational' behavior by the players.
Some work has been done on implementation allowing for variations in
the behavior of the individuals. Some of the rst work in this direction was on
the implementation of the Walrasian correspondence by mechanisms with a
continuous outcome function (e.g., see Postlewaite and Wettstein 1989). Requiring continuity of a mechanism results in an outcome close to the desired
one if some players deviate slightly from equilibrium strategies. More recent
attention to robustness with respect to behavior appears in studies of double
implementation (e.g., see Yamato 1993, 1994a,b). Double implementation
requires implementation in more than one solution concept by the same mechanism. When allowing for combinations such as Nash and strong equilibrium,
potentially unforeseen coalitional deviations are accounted for in implementation. One can also ask for robustness with regards to behavior that is completely unpredictable by some of the players. A new paper by Eliaz (1999)
examines this sort of problem asking that a mechanism rely only on a majority of the individuals (not knowing ex-ante which majority it will be), and using a solution concept that requires behavior of the players to be immune to
arbitrary behavior by some subsets of other players.
63 Again, Abreu and Matsushima (1992a) provide a good deal of insight into this
question for virtual implementation. The reason for considering the exact setting would
be to loosen assumptions on preferences and sensitivity to small probability events
under the lotteries.
696
M. O. Jackson
While the work described above has examined robustness with regards to
certain aspects of the behavior of players, there is almost no study of mechanisms that are immune to misspecications of the environment. The only
study in this direction that I am aware o is by Duggan and Roberts (1998),
who study how continuity of a mechanism can result in robustness with respect to small errors in specication, due to corresponding upper-hemi continuity of equilibrium correspondences.64 This suggests that an intuitive notion
of requiring some continuity of a mechanism in order to guarantee some
minimal robustness is well founded. Short of this, we are lacking even a basic
understanding of the exact limitations that are imposed by requiring implementation by mechanisms that are robust to misspecications of the domain
of preferences, beliefs, allocations, or even the number or role of various
individuals in a society.
7.5 Comparisons across solution concepts
As we have developed a deeper understanding of what is implementable in a
variety of contexts and solution concepts, it has become clear that the choice
of solution concept used to model behavior has an important bearing on what
is implementable. The class of social choice correspondences that are implementable in undominated Nash or subgame perfect equilibrium is signicantly
larger than the class of Nash implementable social choice correspondences.
What is it about solution concepts (or assumptions on individual behavior)
that accounts for this dierence? Jackson and Srivastava (1996) examine this
issue in the context of a voting setting with a nite number of strictly ranked
alternatives. They point to dierences in consideration of o-equilibrium path
behavior as the critical factor the aects how much is implementable via different solution concepts.65 While this works out cleanly in a voting context, it
is still unknown if the same is true more generally or even in other specic
contexts like exchange economies.
Tracing the characterizations of implementability back to assumptions
concerning individual behavior should be quite important. As modelers, we
often consider a solution concept as an object of choice; but ultimately it boils
down to understanding how people behave in a given context. If we understand how variations in behavior tie back to implementability, then we will
eventually be well-equipped to make prescriptions in various contexts. We are
still short of such an understanding.
7.6 Repeated and dynamic implementation
Perhaps one of the most interesting and relevant, and still least studied, questions in implementation theory is that of implementation in dynamic and re64 This holds provided one worries about mixed strategies, which Duggan and Roberts
point to as another reason for accounting for mixed strategies in implementation.
65 New work by Kaya and Koray (2000) provides interesting characterizations based
on identifying solutions that only implement monotonic social choice functions on a
given class of mechanisms.
697
peated contexts. We know from results on folk theorems, reputation, and

learning (as well as various experimental and applied studies), that behavior in
repeated games can dier in important ways from behavior in one-shot games.
As most applications of implementation are repeated ones, ranging from the
design of voting procedure to markets, it is very important that we understand
how mechanisms perform when used repeatedly by the same society. Work by
Kalai and Ledyard (1998) is the only that I am aware of that takes this repeated perspective. They draw on intuition from the reputation and learning
literatures to note that if a mechanism designer is more patient than the underlying population, then the society's characteristics can be teased out and
learned over time. Thus in the long run, a planner can avoid even incentive
compatibility requirements in a Bayesian setting.
There are also interesting contexts where a mechanism is not binding and
so may be used an arbitrary number of times until an agreement is reached.
This is the case in many negotiation and market settings, where a mechanism
outlines potential transactions but a transaction is concluded only when all
parties sign a binding agreement. This leads to a notion of implementation
that is dynamic in nature, but quite dierent from repeated implementation as
an outcome is only reached at one date which is endogenously determined.
Such implementation is studied in Jackson and Palfrey (2001), where a notion
of stationary implementation is examined that is roughly implementation in
Markov perfect equilibria when agents may ask for a replay of the mechanism
if they do not like the outcome it suggests. This notion of implementation has
a relatively simple characterization, and allows for the implementation of the
constrained Walrasian correspondence if agents discount the future.
While these papers take rst steps towards modeling dynamic versions of
implementation, the question of what is implementable in dynamic contexts is
still largely open. Given that most design questions deal with institutions, such
as markets or voting procedures, that will be used more than once, this is one
of the most important open areas in the literature.
7.7 Endogenous choice of mechanisms
The theory of implementation is agnostic on how a social choice correspondence is selected for implementation or how a society might choose a mechanism. These are critical questions in understanding the eventual outcome of the
process. These are also dicult questions to formulate, since at some point one
has to make an ad hoc assumption about the process of choice. How do people
choose how to choose? Nonetheless one can ask for some sort of consistency
properties in decision making. Koray (2000), in a voting context, asks for
consistency in that a social choice function chosen by a society should be
consistent with the manner in which it is selected. That is, the selection of a
social choice function can be formulated as a social choice problem, and the
social choice function selected should be one that would be selected by itself.
A clever formulation (see Koray 2000) allows social choice functions to be
both the objects of choice and the methods of choice at the same time. Although
Koray's result is negative (he nds that only a dictatorial social choice function
698
M. O. Jackson
is self-consistent), one can think of addressing such issues on restricted domains

and allowing for correspondences rather than functions.66
The question of endogenous determination of mechanisms is one of particular importance if we wish to push implementation theory to the full extent
of positive implications and not simply normative ones.67 This is yet another
area with largely unexplored questions and themes.
7.8 Mechanisms in a larger natural context
Having a useful theory of implementation requires not only that we develop an
understanding of implementation via mechanisms with a variety of nice properties, but also that we keep track of the fact that the mechanism will generally
sit in a larger ``natural'' context that the designer may have no control over.
This issue is raised by Hurwicz (1994) where he points out that individuals
may have strategies that are beyond the control of the designer.68 Here, almost
by denition, the issues will be context specic and so we should not expect to
develop general abstract models that encompass large sets of applications.
Instead the variety of issues that arise will be specic to the problem at hand.
One issue that has been looked at is endogenizing the set of alternatives in
economic settings. This was explored by Postlewaite (1979) (see also Thomson
1977) who examined the possibility of individuals withholding (or destroying)
part of their endowments, which can lead to improvements for individuals.
Systematic analyses of endogenous endowments or production possibilities in
implementation in economic settings have been provided by Hurwicz et al.
(1995) and Hong (1996, 1998).
Settings other than production and exchange economies also have alternatives that are to some extent endogenous beyond the control of the mechanism. This endogeneity can be critical to evaluating dierences in mechanisms
as dierent mechanisms may provide very dierent incentives with regards to
how individuals behave in the larger context, and thus to which outcomes are
eventually realized. Other work in this spirit includes studies where the set
of candidates in a voting context are endogenous as explored in Dutta et al.
(1998, 2001), Eraslan and McLennan (2000), and Rodriguez-Alvarez (2000),
where mechanisms may be selected as in Laguno (1992), and situations where
66 New work by Barbera and Jackson (2000) avoids this impossibility result by focussing on settings where voting is over two alternatives a status quo and an alternative.
Voting rules are simply characterized by a quota of how many supporters are needed to
have the status-quo defeated. Under simple assumptions on uncertainty, voters individuals have nicely behaved preference over voting procedures. Barbera and Jackson
examine the existence of self-stable voting procedures: a given voting procedure is selfstable if it is not defeated by some other voting procedure when the given voting procedure is used to make the choice.
67 A related branch of the mechanism design literature relates to the choice of mechanisms by competing sellers, as rst studied by McAfee (1993).
68 This can also be related back to the issue of credibility, renegotiation, and ex post
individual rationality as discussed above, and so that discussion could very well have
been framed in this section.
699
there are endogenous transfer payments from one player to another that may
aect the equilibrium structure of a mechanism, as in Jackson and Wilkie
(2000). The reason that this is an important issue is that the predictions of
what outcomes we might expect can be quite dierent if the individuals have
strategic choices (like deciding whether to be a candidate, or bribing other
individuals to take certain actions) that are outside of the standard implementation analysis. If such actions are available and unaccounted for then the
standard analysis that ignores such outside actions could be quite misleading.
A related matter is that an individual's preferences and behavior may depend on the mechanism itself. Glazer and Rubinstein (1998) present a context
where individuals providing messages (`òpinions'') care explicitly about how
their message ends up comparing to the ultimate social decision, as they would
like to have their recommendation match the social choice. This introduces an
added layer of feedback between the mechanism and equilibrium behavior,
that presents interesting and challenging issues for implementation theory.
Such issues are generally unexplored in the theory, and one can think of any
number of examples beyond the Glazer and Rubinstein context (for instance
in principal-agent problems) where they are important.
7.9 Testing implementing mechanisms
As the theory continues to generate mechanisms for implementation, we can
begin to evaluate and compare them through testing, simulation, experiments,
and ultimately empirical work. For instance Chen (1997) compares behavior
observed in experiments involving a series of public goods mechanisms. She
attributes dierences in performance to supermodularity properties of the
mechanisms, which then helps us appreciate another feature that we might
consider in designing mechanisms. One can also test some of the canonical
mechanisms directly as done by Sefton and Yavas (1996). They ran experiments on the Abreu and Matsushima (1992a) mechanism, and present some
formidable hurdles for the theory.69
Simulations can be also be used to study the same issues, as in Cabrales
(1999). He studies adaptive learning by players in some of the canonical mechanisms for implementation. The adaptive learning algorithm Cabrales analyzes
has features that allow for complete randomization over best responses which
can kick the process out of an integer game, and thus produce convergence
to equilibrium in Nash mechanisms. Ultimately, one would like a mechanism
to perform well under a variety of processes, and to oer fast convergence to
equilibrium as players learn or evolve.
69 These just touch the surface of what can be explored, and one can check whether
various criticisms of unnatural features of mechanisms are supported by experimental
evidence. For instance, how a mechanism with an integer game would be played?
Would players play equilibrium strategies? Would they end up trying to announce high
integers? (I am sure that Caltech undergraduates could be very creative in announcing
high integers.) How would behavior depend on the social choice correspondence being
implemented? Are they more likely to coordinate on the equilibrium play when there is
alignment of preferences?
700
M. O. Jackson
There is ample room for tests, simulations, experiments, and empirical research on mechanisms to add valuable guidance and insight to the theory of
implementation. Issues such as how well various mechanisms perform when
players are not at equilibrium but learning or adjusting according to some
dynamic are quite important and some of the rst issues that come to mind
when one takes an experimental point of view; and yet have not even been
touched by implementation theory.70
References
Abreu D, Matsushima H (1992a) Virtual implementation in iteratively undominated
strategies I: Complete information. Econometrica 60: 9931008
Abreu D, Matsushima H (1992b) Virtual implementation in iteratively undominated
strategies II: Incomplete information. mimeo
Abreu D, Matsushima H (1992c) A response to Glazer and Rosenthal. Econometrica
60: 14391442
Abreu D, Matsushima H (1994) Exact implementation. J Econ Theory 64: 119
Abreu D, Sen A (1991) Virtual implementation in Nash equilibria. Econometrica 59:
9971022
Abreu D, Sen A (1990) Subgame perfect implementation: A necessary and almost
sucient condition. J Econ Theory 50: 285299
Allen B (1997) Implementation theory with incomplete information. In: Hart S, MasColell A, Cooperation: Game theoretic approaches. Springer, Berlin Heidelberg
New York
Amaros P, Moreno B (2001) Implementation of optimal contracts under adverse selection. Rev Econ Design 6: 4162
Arya A, Glover J (1995) A simple forecasting mechanism for moral hazard settings. J
Econ Theory 66: 507521
Arya A, Glover J, Hughes J (1997) Implementing coordinated team play. J Econ
Theory 74: 218232
Arya A, Glover J, Rajan U (2000) Implementation in principal-agent models of adverse selection. J Econ Theory 93: 87109
Arya A, Glover J, Young R (1995) Virtual implementation separable Bayesian environments using simple mechanisms. Games Econ Behav 9: 127138
Bag PK (1996) Ecient allocation of a ``pie'': King Solomon's dilemma. Games Econ
Behav 12: 2141
Bag PK (1997) Public goods provision: Applying Jackson-Moulin mechanisms for
restricted agent characteristics. J Econ Theory 73: 460472
Bagnoli M, Lipman B (1989) Provision of public goods: Fully implementing the core
through private contributions. Rev Econ Stud 56: 583601
Baliga S (1999) Implementation in economic environments with incomplete information: The use of multi-stage games. Games Econ Behav 27: 173183
Baliga S, Brusco S (2000) Collusion, renegotiation, and implementation. Soc Choice
Welfare 17: 69 84
Baliga S, Corchon L, Sjostrom T (1997) The theory of implementation when the
planner is a player. J Econ Theory 77: 1533
70 There is a rich literature in game theory on this topic, but it has not been looked
at from the perspective of designing mechanism to have nice learning or dynamic
properties.
701
Baliga S, Sjostrom T (1998) Decentralization and collusion. J Econ Theory 83: 196
232
Baliga S, Sjostrom T (1999) Interactive implementation. Games Econ Behav 27: 3863
Bandyopadhyay T, Samuelson L (1992) Weakly implementable social choice rules.
Theory Decision 33: 135151
Barbera S (2001) An introduction to strategy-proof social choice functions. Soc Choice
Welfare 18: 619653
Barbera S, Dutta B (1982) Implementation via protective equilibria. J Math Econ 4:
4965
Barbera S, Jackson MO (2000) Choosing how to choose: Self-stable majority rules.
mimeo, Caltech
Bergin J (1993) On some recent results in incomplete information implementation.
mimeo, Queen's University
Bergin J, Sen A (1996) Implementation in generic environments. Soc Choice Welfare
13: 467478
Bergin J, Sen A (1998) Extensive form implementation in incomplete information
environments. J Econ Theory 80: 222256
Bergin J, Duggan J (1999) An implementation theoretic approach to non-cooperative
foundations. J Econ Theory 86: 5076
Bernheim D, Whinston M (1987) Coalition-proof Nash equilibrium II: Applications. J
Blume L, Easley D (1990) Implementation of Walrasian expectations equilibria. J
Boylan R (1998) Coalition-proof implementation. J Econ Theory 82: 132143
Brusco S (1995) Perfect Bayesian implementation. Econ Theory 5: 419444
Brusco S (1997) Implementing action proles when agents collude. J Econ Theory 73:
395424
Brusco S (1999) Implementation with extensive form games: One round of signaling
is not enough. J Econ Theory 87: 356378
Brusco S (1998a) Implementing action proles with sequential mechanisms. Rev Econ
Design 3: 271300
Brusco S (1998b) Unique implementation of the full surplus extraction outcome in
auctions with correlated types. J Econ Theory 80: 185200
Brusco S (1998c) Perfect Bayesian implementation in economic environments. mimeo,
Universidad Carlos III de Madrid
Brusco S, Jackson M (1999) The optimal design of a market. J Econ Theory 88: 139
Cabrales A (1999) Adaptive dynamics and the implementation problem with complete
information. J Econ Theory 86: 159184
Chakravorti B (1991) Strategy space reductions for feasible implementation of Walrasian performance. Soc Choice Welfare 8: 235246
Chakravorti B, Corchon L, Wilkie S (1992) Credible implementation. Games Econ
Behav (forthcoming)
Chattopadhyay S, Corchon L, Naeve J (2000) Contingent commodities and implementation. Econ Letters 68: 293298
Chen Y (1997) Dynamic stability of Nash-ecient public goods mechanisms: reconciling theory and experiments. mimeo, University of Michigan
Corchon L (1996) The theory of implementation of socially optimal decisions in economics. McMillan, London
Corchon L, Herrero C (1995) A decent proposal. mimeo, Universidad de Alicante
Corchon L, Ortuno-Ortin I (1995) Robust implementation under alternative information structures. Econ Design 1: 159172
Corchon L, Wilkie S (1996) Double implementation of the ratio correspondence by a
market mechanism. Econ Design 2: 325337
Dagan N, Serrano R, Volij O (1999) Feasible implementation of taxation methods.
Rev Econ Design 4: 5772
702
M. O. Jackson
Danilov V (1992) Implementation via Nash equilibrium. Econometrica 60: 4356

Dasgupta P, Hammond P, Maskin E (1979) The implementation of social choice
rules: Some general results on incentive compatibility. Rev Econ Stud 46: 185
216
Dasgupta P, Maskin E (1997) Notes on ecient auctions. mimeo, Harvard University
Demange G (1984) Implementing ecient egalitarian equivalent allocations. Econometrica 52: 11671178
Demski J, Sappington D (1984) Optimal incentive contracts with multiple agents. J
Duggan J (1995) Sequentially rational implementation with incomplete information.
mimeo, University of Rochester
Duggan J (1997a) Virtual Bayesian implementation. Econometrica 65: 11751199
Duggan J (1997b) Implementing social welfare optima when there is a private good.
Duggan J (1998) An extensive form solution to the adverse selection problem in
principal/multi-agent environments. Rev Econ Design 3: 167191
Duggan J, Roberts J (1997) Robust implementation. mimeo, University of Rochester
Dutta B, Jackson MO, Le Breton M (2001) Strategic candidacy and voting procedures.
Econometrica 69: 10131038
Dutta B, Jackson MO, Le Breton M (1998) Voting by successive elimination and
strategic candidacy. J Econ Theory (forthcoming)
Dutta B, Sen A (1991a) Implementation under strong equilibria: A complete characterization. J Math Econ 20: 4968
Dutta B, Sen A (1991b) Necessary and sucient conditions for 2-person Nash implementation. Rev Econ Stud 58: 121129
Dutta B, Sen A (1991c) Implementation in Bayesian equilibrium. mimeo, Indian
Statistical Institute
Dutta B, Sen A (1993) Implementing generalized condorcet social choice functions via
backward induction. Soc Choice Welfare 10: 149160
Dutta B, Sen A (1994) The necessity of innite mechanisms. J Econ Theory 64: 130
141
Dutta B, Sen A, Vohra R (1995) Nash implementation through elementary mechanisms in economic environments. Econ Design 1: 173204
Ehlers L (2000) Monotonic and implementable solutions in generalized matching
problems. mimeo, Maastricht University
Eliaz K (1999) Fault tolerant implementation. mimeo, Tel Aviv University
Eraslan H, McLennan A (2000) Strategic candidacy for multivalued voting procedures. mimeo, University of Minnesota
Gevers L (1986) Walrasian social choice: Some simple axiomatic approaches. In: Heller
W et al., Social choice and public decision making: Essays in honor of Kenneth J.
Arrow. Cambridge University Press, Cambridge
Gibbard A (1973) Manipulation of voting schemes: A general result. Econometrica
41: 587601
Giraud G, Rochon C (2001) Consistent collusion-proofness and correlation in exchange economies. mimeo, CORE
Glazer J, Ma A (1989) Ecient allocation of a ``Prize'': King Solomon's dilemma.
Games Econ Behav 1: 222233
Glazer J, Perry M (1996) Virtual implementation in backwards induction. Games Econ
Behav 15: 2732
Glazer J, Rosenthal R (1992) A note on AbreuMatsushima mechanisms. Econometrica 60: 14351438
Glazer J, Rubinstein A (1998) Motives and implementation: On the design of mechanisms to elicit opinions. J Econ Theory 79: 157173
Glover J (1994) A simpler mechanism that stops agents from cheating. J Econ Theory
62: 221229
703
Groves T (1977) Ecient collective choice with compensation. In: Laont J-J, Aggregation and revelation of preferences. North-Holland, Amsterdam
Hahn G, Yannelis N (1995) Coalitional Bayesian Nash implementation in dierential
information economies. mimeo, University of Illinois
Harsanyi JC (196768) Games with incomplete information played by `Bayesian'
Players. Management Science 14: 159189, 320334, 486502
Herrero M, Srivastava S (1992) Implementation via backward induction. J Econ
Theory 56: 7088
Hong L (1995) Nash implementation in production economies. J Econ Theory 5:
401417
Hong L (1996) Bayesian implementation in exchange economies with state dependent
feasible sets and private information. Soc Choice Welfare 13: 433444
Hong L (1998) Feasible Bayesian implementation with state dependent feasible sets. J
Hong L, Page S (1994) Reducing informational costs in endowment mechanisms. Econ
Design 1: 103117
Hurwicz L (1972) On informationally decentralized systems. In: McGuire CB, Radner
R, Decision and organization. North Holland, Amsterdam
Hurwicz L (1973) The design of mechanisms for resource allocation. Am Econ Rev
61: 130
Hurwicz L (1977) On allocations attainable through Nash equilibria. In: Laont J-J,
Aggregation and revelation of preferences. North-Holland, Amsterdam
Hurwicz L (1979) Outcome functions yielding Walrasian and Lindahl allocations at
Nash equilibrium points. Rev Econ Stud 46: 217225
Hurwicz L (1994) Economic design, adjustment processes, mechanisms and institutions. Econ Design 1: 114
Hurwicz L (1996) Feasible balanced outcome functions yielding constrained Walrasian
and Lindahl allocations at Nash equilibrium points in economies with two agents
when the designer knows the feasible set. mimeo, University of Minnesota
Hurwicz L, Schmeidler D (1978) Construction of outcome functions guaranteeing existence and pareto optimality of Nash equilibria. Econometrica 46: 14471474
Hurwicz L, Maskin E, Postlewaite A (1995) Feasible Nash implementation of social
choice correspondences when the designer does not know endowments or production sets. In: Ledyard J, The economics of information decentralization: Complexity,
eciency, and stability. Kluwer, Amsterdam
Jackson MO (1991) Bayesian implementation. Econometrica 59: 461478
Jackson MO (1992) Implementation in undominated strategies: A look at bounded
mechanisms. Rev Econ Stud 59: 757775
Jackson MO, Moulin H (1992) Implementing a public project and distributing its cost.
J Econ Theory 57: 125140
Jackson MO, Palfrey T (1998) Eciency and voluntary implementation in markets
with repeated pairwise bargaining. Econometrica 66: 13531388
Jackson MO, Palfrey T (2001) Voluntary implementation. J Econ Theory 98: 1
25
Jackson MO, Palfrey T, Srivastava S (1994) Undominated Nash implementation in
bounded mechanisms. Games Econ Behav 6: 474501
Jackson MO, Srivastava S (1992) On two person Nash implementable social choice
functions. Soc Choice Welfare 9: 263264
Jackson MO, Srivastava S (1996) Characterizations of game theoretic solution concepts
which lead to impossibility theorems. Rev Econ Stud 63: 2338
Jackson MO, Wilkie S (2000) Endogenous games and mechanisms: Side payments
among players. mimeo, Caltech
Kalai E, Ledyard J (1998) Repeated implementation. J Econ Theory 83: 308317
Kara T, Sonmez T (1996) Nash implementation of matching rules. J Econ Theory 68:
425439
704
M. O. Jackson
Kara T, Sonmez T (1997) Implementation of college admissions rules. Econ Theory

9: 197218
Kaya A, Koray S (2000) Two essays in social choice theory. mimeo, Bilkent University
Koray S (2000) Self-selective social choice functions verify arrow and GibbardSatterthwaite theorems. Econometrica 68: 981996
Kwasnica AM (1998) Bayesian implementable ecient and core allocations. mimeo,
Caltech
Laguno R (1992) Fully endogenous mechanism selection on nite outcome sets. Econ
Theory 2: 462480
Li Q, Nakamura S, Tian G (1995) Nash implementation of the Lindahl correspondence with decreasing returns to scale technologies. Int Econ Rev 36: 3752
Ma C (1988) Unique implementation of incentive contracts with many agents. Rev
Econ Stud 55: 555572
Ma C, Moore J, Turnbull S (1988) Stopping agents from `cheating'. J Econ Theory
46: 355372
Maniquet F (1994) On equity and implementation in economic environments. Ph.D.
dissertation, University of Namur
Maniquet F (1999) Implementation under perfect information in economic environments. mimeo, University of Namur
MasColell A, Vives X (1993) Implementation in economies with a continuum of
agents. Rev Econ Stud 60: 613629
Maskin E (1985) The theory of implementation in Nash equilibrium. In: Hurwicz L,
Schmeidler D, Sonnenschein H (eds) Social goals and social organization: Essays
in honor of Elisha A. Pazner. Cambridge University Press, Cambridge
Maskin E (1999) Nash equilibrium and welfare optimality. Rev Econ Stud 66: 2338
Maskin E, Moore J (1999) Implementation with renegotiation. Rev Econ Stud 66:
3956
Maskin E, Sjostrom T (2001) Implementation theory. mimeo, Harvard University and
Penn State
Matsushima H (1988) A new approach to the implementation problem. J Econ Theory
45: 128144
Matsushima H (1993) Bayesian monotonicity with side payments. J Econ Theory 59:
107121
McAfee RP (1993) Mechanism design by competing sellers. Econometrica 61: 1281
1312
McKelvey R (1989) Game forms for Nash implementation of general social choice
correspondences. Soc Choice Welfare 6: 139156
Miyagawa E (1997) Implementation of the normalized utilitarian bargaining solution.
Miyamoto H, Watanabe T, Mizuno S (1990) Implementation in admissible strategies.
Technical Report No. 28, Department of Management Science and Engineering,
Tokyo Institute of Technology
Mookherjee D, Reichelstein S (1990) Implementation via augmented revelation mechanisms. Rev Econ Stud 57: 453476
Mookherjee D, Reichelstein S (1992) Dominant strategy implementation of Bayesian
incentive compatible allocation rules. J Econ Theory 56: 378399
Moore J (1992) Implementation, contracts, and renegotiation environments with
complete information. In: Laont J-J, Advances in economic theory. Cambridge
University Press, Cambridge
Moore J, Repullo R (1988) Subgame perfect implementation. Econometrica 56: 1191
1220
Moore J, Repullo R (1990) Nash implementation: A full characterization. Econometrica 58: 10831100
Moulin H (1980b) Implementing ecient, anonymous and neutral social choice functions. J Math Econ 72: 249269
705
Moulin H (1981) Implementing just and ecient decision making. J Publ Econ 16:
193213
Moulin H (1982) Non-cooperative implementation: A survey of recent results. Math
Soc Sci 3: 243257
Moulin H (1988) Axioms of cooperative decision making. Cambridge University Press,
Cambridge
Mount K, Reiter S (1974) The informational size of message spaces. J Econ Theory 8:
161192
Muller E, Satterthwaite M (1977) On the equivalence of strong positive association and
strategyproofness. J Econ Theory 14: 412418
Myerson R, Satterthwaite M (1983) Ecient mechanisms for bilateral trading. J Econ
Theory 29: 265281
Nagahisa RI (1994) A necessary and sucient condition for Walrasian social choice. J
Nakamura S (1988) Feasible Nash implementation of competitive equilibria in an
economy with externalities. mimeo, University of Minnesota
Nakamura S (1990) A feasible Nash implementation of Walrasian equilibria in the two
agent economy. Econ Lett 34: 59
Nakamura S (1998) Impossibility of Nash implementation in two person economies.
Econ Design 3: 159166
Osana H (1997) Nash-implementation of the weak pareto choice rule for indecomposable environments. Rev Econ Design 3: 5774
Osana H (1990) Implementation of multi-agent incentive contracts with the principal's
renegotiation oer. Rev Econ Design 4: 161178
Palfrey T (1990) Implementation in Bayesian equilibrium: The multiple equilibrium
problem in mechanism design. In: Laont J-J, Advances in economic theory. Cambridge University Press, Cambridge
Palfrey T (1995) Implementation theory. In: Aumann R, Hart S, Handbook of game
theory (forthcoming)
Palfrey T, Srivastava S (1986) Private information in large economies. J Econ Theory
39: 3458
Palfrey T, Srivastava S (1987) On Bayesian implementable allocations. Rev Econ Stud
54: 193208
Palfrey T, Srivastava S (1989a) Implementation with incomplete information in exchange economies. Econometrica 57: 115134
Palfrey T, Srivastava S (1989b) Mechanism design with incomplete information: A
solution to the implementation problem. J Polit Econ 97: 668691
Palfrey T, Srivastava S (1991) Nash implementation using undominated strategies.
Palfrey T, Srivastava S (1993) Bayesian implementation. Harwood Academic Publishers, Switzerland
Peleg B (1996a) A continuous double implementation of the constrained Walrasian
equilibrium. Econ Design 2: 8997
Perry M, Reny P (1999) A general solution to King Solomon's dilemma. Games Econ
Behav 26: 279285
Peleg B (1996b) Double implementation of the Lindahl equilibrium by a continuous
mechanism. Econ Design 2: 311324
Piketty T (1993) Implementation of rst-best allocations via generalized tax schedules.
J Econ Theory 61: 2341
Postlewaite A (1985) Implementation via Nash equilibrium in economic environments.
In: Hurwicz L, Schmeidler D, Sonnenschein H (eds) Social goals and social organization: Essays in honor of Elisha A. Pazner. Cambridge University Press, Cambridge
Postlewaite A (1979) Manipulation via endowments. Rev Econ Stud 46: 255262
Postlewaite A, Schmeidler D (1986) Implementation in dierential information
economies. J Econ Theory 39: 1433
706
M. O. Jackson
Postlewaite A, Wettstein D (1989) Continuous and feasible implementation. Rev Econ

Stud 56: 1433
Rai A (1997) Ecient audit mechanisms to target the poor. Dissertation, University
of Chicago
Reichelstein S, Reiter S (1988) Game forms with minimal strategy spaces. Econometrica 56: 661692
Repullo R (1987) A simple proof of Maskin's theorem on Nash implementation. Soc
Choice Welfare 4: 3942
Rochon C (2000) Stability and eciency: Monetary policy and implementation theory.
Ph.D. Dissertation, Universite Catholique de Louvain
Rodriguez-Alvarez C (2000) Candidate stability and multi-valued voting procedures.
mimeo, Universitat Autonoma de Barcelona
Rubinstein A, Wolinsky A (1992) Renegotiation-proof implementation and time preferences. Am Econ Rev 82: 600614
Saglam I (1997) A note on Jackson's theorems in Bayesian implementation. mimeo,
Bilkent University
Saijo T (1987) On constant Maskin-monotonic social choice functions. J Econ Theory
42: 382386
Saijo T (1988) Strategy space reduction in Maskin's theorem: Sucient conditions for
Nash implementation. Econometrica 56: 693700
Saijo T, Tatimatani Y, Yamato T (1996) Toward natural implementation. Int Econ
Rev 37: 949980
Schmeidler D (1980) Walrasian analysis via strategic outcome functions. Econometrica
48: 15851593
Sefton M, Yavas A (1996) Abreu-Matsushima mechanisms: Experimental evidence.
Games Econ Behav 16: 280302
Sen A (1995) The implementation of social choice functions via social choice correspondences: A general formulation and a limit result. Soc Choice Welfare 12: 277
292
Serrano R (1993) Non-cooperative implementation of the nucleolus- the 3-player case.
Int J Game Theory 22: 345357
Serrano R (1995) A market to implement the core. J Econ Theory 67: 285294
Serrano R (1997) A comment on the Nash program and the theory of implementation.
Econ Lett 55: 203208
Serrano R, Vohra R (1997) Non-cooperative implementation of the core. Soc Choice
Welfare 14: 513525
Serrano R, Vohra R (2001) Some limitations of virtual Bayesian implementation.
Serrano R, Vohra R (2000) Type diversity and virtual Bayesian implementation.
mimeo, Brown University
Sertel M (1994) Manipulating Lindahl equilibrium via endowments. Econ Lett 46:
167171
Shin S, Suh S (1996) A mechanism implementing the stable rule in marriage problems.
Econ Lett 51: 169176
Shin S, Suh S (1997) Double implementation by a simple game form in the commons
problem. J Econ Theory 77: 205213
Sjostrom T (1991) On the necessary and sucient conditions for Nash implementation.
Soc Choice Welfare 8: 333340
Sjostrom T (1993) Implementation in perfect equilibrium. Soc Choice Welfare 10: 97
106
Sjostrom T (1994) Implementation in undominated Nash equilibria without integer
games. Games Econ Behav 6: 502511
Sjostrom T (1995a) Implementation in teams. Econ Design 1: 327342
Sjostrom T (1995b) Implementation by demand mechanisms. Econ Design 1: 343
354
707
Sjostrom T (1996) Credibility and renegotiation of outcome functions in implementation. Japanese Econ Rev 47: 157169
Sjostrom T (1999) Undominated Nash implementation with collusion and renegotiation. Games Econ Behav 26: 337352
Sonmez T (1996) Implementation in generalized matching problems. J Math Econ
26: 169176
Suh S-C (1995) A mechanism implementing the proportional solution. Econ Design
1: 301317
Suh S-C (1996) Implementation with coalition formation: A complete characterization.
J Math Econ 25: 109122
Suh S-C (1997a) An algorithm checking strong Nash implementability. J Math Econ
25: 109122
Suh S-C (1997b) Games implementing the stable rule of marriage problems in strong
Nash equilibria. mimeo, University of Windsor
Suh S-C (2001) An algorithm for verifying the double implementability in Nash and
strong Nash equilibria. Math Soc Sci 41: 103110
Tadenuma K, Toda M (1998) Implementable stable solutions to pure matching
problems. Math Soc Sci 35: 121132
Tatamitani Y (1991) Double implementation in Nash and undominated Nash equilibria in social choice environments. mimeo, University of Tsukuba
Thomson W (1977) Comment on: L. Hurwicz On allocations attainable through
Nash equilibria. In: Laont J-J, Aggregation and revelation of preferences. NorthHolland, Amsterdam
Thomson W (1979) Maximin strategies and elicitation of preferences. In: Laont J-J,
Aggregation and revelation of preferences. North Holland, Amsterdam
Thomson W (1996) Concepts of implementation. Japanese Econ Rev 47: 133143
Thomson W (1994) Divide and permute and the implementation of solutions to the
problem of fair division. mimeo, University of Rochester
Thomson W (1999) Monotonic extensions on economic domains. Rev Econ Design
4: 1334
Tian G (1989) Implementation of the Lindahl correspondence by a single-valued, feasible, and continuous mechanism. Rev Econ Stud 56: 613621
Tian G (1990) Completely feasible continuous implementation of the Lindahl correspondence with a message space of minimal dimensions. J Econ Theory 51: 443452
Tian G (1991) Implementation of the Lindahl allocations with nontotal-nontransitive
preferences. J Publ Econ 46: 247259
Tian G (1993) Implementing Lindahl allocations by a withholding mechanism. J Math
Econ 22: 169179
Tian G (1994) Implementation of linear cost share equilibrium allocations. J Econ
Theory 64: 568584
Tian G (1996) Continuous and feasible implementation of rational expectations
Lindahl allocations. Games Econ Behav 16: 135151
Tian G (1999) Bayesian implementation in exchange economies with state dependent
preferences and feasible sets. Soc Choice Welfare 16: 99120
Tian G, Li Q (1991) Completely feasible and continuous implementation of the
Lindahl correspondence with any number of goods. Math Soc Sci 21: 6779
Tian G, Li Q (1994) An implementable and informational ecient state-ownership
system with variable returns. J Econ Theory 64: 286297
Tian G, Li Q (1995) On Nash-implementation in the presence of withholding. Games
Econ Behav 9: 222233
Trockel W (1998) An exact implementation of the Nash bargaining solution in
dominant strategies. In: Abramovich Y, Avgerinos E, Yanellis N (eds) Functional
analysis and economic theory. Springer, Berlin Heidelberg New York
van Damme E (1987) Stability and perfection of Nash equilibria. Springer, Berlin
Heidelberg New York
708
M. O. Jackson
Vartiainen H (2000) Subgame perfect implementation: A full characterization for the

many-player case. mimeo, University of Helsinki
Vartiainen H (2000b) Subgame perfect implementation: A full characterization for the
two-player case. mimeo, University of Helsinki
Wettstein D (1990) Continuous implementation of constrained rational expectations
equilibria. J Econ Theory 52: 208222
Wettstein D (1992) Continuous implementation in economies with incomplete information. Games Econ Behav 4: 463483
Williams S (1984) Sucient conditions for Nash implementation. mimeo, Northwestern University
Williams S (1986) Realization and Nash implementation: Two aspects of mechanism
design. Econometrica 54: 139151
Yamato T (1992) On Nash implementation of social choice correspondences. Games
Econ Behav 4: 484492
Yamato T (1993) Double implementation in Nash and undominated Nash equilibria. J
Yamato T (1994) Equivalence of Nash implementability and robust implementability
with incomplete information. Soc Choice Welfare 11: 289303
Yamato T (1999) Nash implementability and double implementation: Equivalence
theorems. J Math Econ 31: 215238
Yoshihara N (1999) Natural and double implementation of public ownership solutions
in dierentiable production economies. Rev Econ Design 4: 127152
Yoshihara N (2000) A characterization of natural and double implementation in production economies. Soc Choice Welfare 17: 571600

A Crash Course in Implementation Theory

Uploaded by

Copyright:

Available Formats

A Crash Course in Implementation Theory

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

A Crash Course in Implementation Theory

Uploaded by

Copyright:

Available Formats

Soc Choice Welfare (2001) 18: 655708

A crash course in implementation theory

2.4 Social choice correspondences

S that identies a subset of M for any given M; g; R specication. For the

approach is ne. However, in situations where it leads to positive results so

complete information concerning each others' preferences (or at least enough

if he had preferences R 1 instead of R 1 when the state is actually R 1 ; R 2 . This

that gm i ; mi Pi gm. Since m was a Nash equilibrium at R, it must be that

standing of an equilibrium alternative improves it must remain an equilibrium

Mi P A N where N is the set of nonnegative integers. Thus, each

violated by a number of social choice correspondences that have some strong

that aR 1 b and bP 1 a. The allocation b necessary to satisfy this requires an

another perspective, note that the Muller-Satterthwaite (1977) theorem states

This structure on the set of alternatives allows us to dene distance among

Theorem 3 (Matsushima 1988, Abreu and Sen 1991).

. Any ordinal social choice correspondence dened on a domain of monotone

preferences is e-close to a social choice correspondence that satises monotonicity.

correspondence virtually implementable? The answer is yes when N b 3.29

Looking at renements helps in ruling out undesired equilibria and leads to

upper-semi continuous and locally non-satiated34 and so applies to almost

Note however, that at R 1 ; R 2 for most denitions of preferences over

context of a problem such as a principal/multiple-agent problem, where

Claim 2. Suppose that players' preferences over lotteries on A are continuous,

in other ways, is to examine implementation in specic environments. Part

Theorem 7 is proven using the following mechanism.

the mechanisms above there is some ``bad'' outcome (w in the undominated

47 One approach to addressing the complexity of announcements of the mechanism

7 Other topics and open questions

because of dierences in the primitive underlying preferences, but also from

to addressing this problem which is to analyze specic environments where the

peated contexts. We know from results on folk theorems, reputation, and

is self-consistent), one can think of addressing such issues on restricted domains

Danilov V (1992) Implementation via Nash equilibrium. Econometrica 60: 4356

Kara T, Sonmez T (1997) Implementation of college admissions rules. Econ Theory

Postlewaite A, Wettstein D (1989) Continuous and feasible implementation. Rev Econ

Vartiainen H (2000) Subgame perfect implementation: A full characterization for the

You might also like

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.