ECO 472 - Modern Value Theory

Class Notes - Feb 2

1 Game Theory

A game is a description of a strategic situation. We have several decision makers,

called players, and each such player faces a set of possible choices. Once all players
make their choices, an outcome ensues. We assume that players have preferences
over outcomes, and proceed to analyze the decisions of the players. That’s game
theory. The crucial feature is that outcomes depend on everyone’s choices, so a
player alone cannot control what s/he will end up getting. In turn, it will be on
the player’s best interest to forecast and/or influence what other players will do,
that is, to strategize. By studying decisions of players, game theory is the study of
strategies.

2 Games in Normal Form

The most basic game form is the normal from. It is given by a tuple

(Si , ui )ni=1 ,

where {1, ..., n} is the set of players, Si and ui : S → R are the strategy set and
utility function of a generic player i ∈ {1, ..., n}. The normal form combines two
ideas. First the representation of a strategic situation as a game form. This is
captured by the player set, the strategy sets, and an outcome function g : S → X,
where S = ×ni=1 Si is the set of profiles s = (s1 , ..., sn ) of strategies and X is the set
of possible outcomes. The sets Si , i = 1, ..., n, and X are completely general and as
such capture just about any strategic situation, from simple static problems where
a choice of a player is to pick one out of many available actions, to complicated
dynamic problems where a choice of a player is to pick one out of many available
contingent plans of actions. The strategic component is captured by the outcome
function, as it’s defined over profiles of strategies. We envision each player making
a choice si ∈ Si , yielding a profile s ∈ S of strategies. The implied actions are
then carried on and an outcome x ∈ X ensues, which is captured by g(s) = x ∈ X.
Outcomes are what matter for a player, and outcomes depend on the choices of
every player, so a player would do well to anticipate and influence the choices of the
others. That is, strategic thinking takes center stage. The second idea encapsulated
in the strategic form is our reliance on utility theory: since what matters for a player

1
are the outcomes, we assume each player i has a preference relation %i defined over
X. In fact, we assume %i satisfies all the nice properties that allow us to represent
it by a numerical function ui , which will shortly take the more restricted form of
an expected utility model (once we extend preferences to lotteries over outcomes.)
Differently from decision theory, there’s still a lot left to be determined after utility
functions have been “obtained”. (Ideally one would estimate such functions; we will
simply make plausible assumptions about them and proceed with our analysis. Of
course, our analyzes can be applied to generic utility functions, so whatever they
happen to be, we will know how to proceed.)

Observe now that the outcome function and the outcome space are just an extra
burden in terms of notation, because the players’ choices can be analyzed directly
by their numerical consequences ui (g(s)). So we abuse notation, suppress g and X,
write ui (s) = ui (g(s)), and arrive at the normal form representation (Si , ui )ni=1 .1

Every aspect of (Si , ui )ni=1 is common knowledge: every player knows the
tuple (Si , ui )ni=1 , and knows that everyone knows it, and knows that everyone knows
it, and so on. This is certainly a strong assumption in most settings and we will
relax it below when we introduce games of incomplete information. As a notational
convention, we use s−i ∈ ×j6=i Sj to denote the strategies of every player but player
i, and sometimes write s = (si , s−i ), when we need to highlight the i-th coordinate
of the profile s.

Example. The Prisoner’s Dilemma game is as follows. The players are two
prisoners. The D.A. puts each prisoner in a separate room and tells each separately
that if s/he cooperates s/he goes to jail for 10 years if the other defects, and for
1 year if the other cooperates; and if s/he defects s/he goes to jail for 2 years if
the other defects, and is freed if the other cooperates. Then, because they cannot
communicate, each prisoner makes his or her choice without knowing the choice of
the other. The player set is {1, 2}. Using C for cooperate and D for defect, we have
Si = {C,D} for i = 1, 2. The outcome space is X = {(1, 1), (10, 0), (0, 10), (2, 2)}
where (10, 0) means that prisoner 1 (resp. 2) goes to jail for 10 (resp. 0) years, and
the outcome function is g(C,C) = (1, 1), g(C,D) = (10, 0), g(D,C) = (0, 10), and
g(D,D) = (2, 2). The standard assumption is that a prisoner will prefer less years in
jail to more years in jail. This can be captured by the utility function (for prisoner
1): u1 : X → R given by u1 (1, 1) = 2, u1 (10, 0) = 0, u1 (0, 10) = 3, and u1 (2, 2) = 2.
Prisoner 2’s utility function is similarly defined.2 It is clear that we can simply
1
A common misuse of terminology, that we will nevertheless often employ, is to call ui player i’s
payoff function. It should be clear from the description above that ui is a subjective utility function
representing player i’s preferences.
2
Obviously, there’s nothing special about the numbers chosen; we simply normalized the worst
outcome to 0 and added one util as we moved up the preference of the player. Any other function

2
suppress the outcome space and function and work directly with utility functions
taking S as the domain. In short, we arrive at the following payoff matrix:

C D
C 2, 2 0, 3
D 3, 0 1, 1

Example. Sealed-Bid Auctions: There’s one indivisible good for sale and n > 1
bidders who bid to get the good. Bidders write their bids on envelopes, seal them,
and submit to the auctioneer. The auctioneer opens all envelopes and awards the
good to the highest bidder; in case of a tie at the top, the winner is decided by some
tie-breaking rule. The winner gets the good and pays a price, everyone else does
not get the good and does not pay anything. The price is a function of the bids: it
can be the bid itself (first price) or the highest among the other bids (second price),
or the second highest among the other bids (third price), etc. In principle any bid
is allowed, so Si = R (or maybe just positive integers) for i = 1, ..., n. A standard
assumption is that preferences take a very specific form: each bidder i is described
by his or her willingness to pay vi ∈ R and the utility of the winner is equal to vi
minus the price paid (that is, it is quasilinear). The utility of a bidder who does
not win is normalized to zero. The outcome space is the set describing the winner
and the price paid, as a function of the profile of bids (the winner is the one with
the highest bid, the price paid depends on the format of the auction.) Suppressing
directly the outcome space and function and assuming that tie-breaking rule is a
simple coin-flip, we write the utility function of bidder i as
(
1
ui (s) = m (vi − p(s)) if i ∈ W (s)
0 otherwise

where p(s) is the price paid by the winner, W (s) is the set of players whose bids are
the highest, and m is its cardinality.

These are just two simple examples to illustrate the ideas involved. When we
have at most 3 players and a few strategies per player, it is often convenient to
represent a game in normal form by its payoff matrix. We cannot do so when we
have many players and many strategies, as in the sealed-bid auction example.

The following are the classical two-player, two-strategy games in normal form.

Example. Battle of the Sexes is an example of a coordination game. The

players are a married couple, and they are to meet either at a Football game or at
the Opera, but they cannot communicate so all either can do is to show up at either
respecting this order would do.

3
of the places (and hope that the other will pick the same venue.) Player 1 is the
husband and player 2 is the wife. Husband would rather meet at the Football game,
wife at the Opera; they prefer meeting to not meeting. The payoff matrix is thus
(again normalizing the worst outcome to zero):
F O
F 2, 1 0, 0
O 0, 0 1, 2

Example. An example of Game of Chicken (or Hawk-Dove) is the following:

two young people, old enough to drive, but not to know better, decide to drive their
cars towards one another in a narrow street. Their options are to Continue or Not.
N guarantees not crashing, but if the other continues, the given player will be the
chicken and lose face among their peers; C gives you a chance of being the brave
one, in case the other swerves, but if you both continue, you’ll crash and get severely
injured. The following matrix captures this story:
N C
N 0, 0 −1, 1
C 1, −1 −2, −2

Example. Matching Pennies is a children’s game where each of two children

have a penny and is supposed to place it at the table. If the sides match, player
1 takes the penny of player 2; otherwise player 2 gets player 1’s penny. Here’s the
payoff matrix (H for Head, T for Tail):
H T
H 1, −1 −1, 1
T −1, 1 1, −1

The last example above is a situation where it is necessary to allow for players to
use randomized strategies. That is, instead of simply choosing an element si ∈ Si ,
a player may chose a probability distribution over such pure strategies (the basic
intuition is that a player may want to conceal their real play, to avoid being an easy
prey, and an effective way of doing so is to delegate choices to “coin-flips.”) Indeed,
by tossing the coin a child is effectively leaving the other child in the dark as to what
will turn up. Formally, a mixed strategy is a probability distribution σi ∈ ∆(Si )
(or a probability measure over the measurable subsets of Si , in case Si is not a finite
set.) For a given profile σ = (σ1 , ..., σn ) of mixed strategies, we compute expected
payoffs by taking mathematical expectation of ui with respect to the product of the
σi ’s. That is,
XY n
ui (σ) = σj (sj )ui (s),
s∈S j=1

4
in case Si ’s are finite, and in general
Z
ui (σ) = ui (s)σ(ds), with σ = ⊗ni=1 σi .
S

(Observe the abuse of notation, which will be repeatedly applied in the course, of
using the same letter ui to denote the Bernoulli index ui : S → R and its expectation
ui : ∆(S) → R with respect to the product probability measure σ; the argument of
the function will make it clear which function we are referring to.) As we know, what
lies underneath this computation is the assumption that players have preferences
over lotteries over S that satisfy the usual vN-M axioms.

3 Dominance

Consider a normal form game (Si , ui )ni=1 . A strategy s∗i ∈ Si is dominant if it is

the best option regardless of what others may do: that is, s∗i is dominant if, for each
si 6= s∗i
ui (s∗i , s−i ) > ui (si , s−i ) for all s−i ∈ S−i .

For instance, strategy D in the prisoner’s dilemma is dominant, for conditional

on the choices of the other, D yields 1 and 3, whereas C yields 0 and 2. Notice
that in choosing D there is no strategic thinking: it simply makes no sense to play
anything else (unilaterally). Beware of a common misconception here: the utility
numbers capture the preferences of the players over the outcomes. For instance,
we postulated that a player is happier if he is freed and the other is sent for 10
years of jail time than if both are sent for 1 year of jail time. This means that the
player compared these two situations and decided that the possible personal cost of
sending the partner to 10 years of jail time is not enough to offset the benefit of
being freed. So it indeed does not make any sense to play anything other than D -
all the other-regarding considerations have already been incorporated in the utility
numbers.

Even if a strategy is weakly dominant (same inequalities as above, but some,

not all, are not strict), it makes sense to play it, and again we can recommend a
choice that is weakly dominant. For instance, in the second-price sealed-bid auction,
for a bidder with valuation vi , the strategy si = vi is weakly dominant: if a higher
bid wins when it should win (when the second highest bid is lower than vi ), so
would si = vi ; if a higher bid wins when it should not (when the second highest bid
is higher than vi ), then si = vi would lose; a symmetric logic applies for bids lower
than vi . The argument just used assumed complete information. But even with

5
incomplete information (that is, with vi as i’s private information) the same logic
shows that the strategy si (vi ) = vi is weakly dominant.

Most games do not have dominant strategies. The next best thing is to look for
strategies that would not be chosen. A strategy si is dominated if there exists s′i
such that
ui (s′i , s−i ) > ui (si , s−i ) for all s−i ∈ S−i

A strategy is weakly dominated if some, not all, of the strict inequalities above
are equalities. A strategy that is not weakly dominated is called an admissible
strategy. It does not make sense too much sense to choose a dominated strategy, for
it makes more sense to choose the strategy that dominates it. Observe that we need
one strategy s′i that is uniformly better than si , not one strategy per each s−i .

In a first-price sealed-bid auction with complete information, the strategy si = vi

is weakly dominated (provided that vi is greater than the lowest possible valuation.)
It yields zero for any profile of bids of the others, whereas the payoff using si < vi is
always nonnegative, and is positive whenever it is the winning bid. Even with incom-
plete information, the same argument shows that the strategy si (vi ) = vi is weakly
dominated. There is a dual argument that sometimes is more transparent, and in
any case it provides the decision theoretic foundations for the term “rationality” in
games. This is what we do next

4 Rationality

In decision theoretic terms, a choice is rational if it can be justified by a subjective

belief, where a belief is a probability distribution over the relevant space of uncer-
tainty. What is behind the concept is Savage’s notion of subjective probabilities. A
choice of a strategy si is in fact a choice of a random variable that maps S−i to R:
s−i 7→ ui (si , s−i ) ∈ R. It is a rational choice if there exists a conjecture µi ∈ ∆(S−i )
(note that player i may or may not consider that his opponents will correlate their
choices) such that
ui (si , µi ) ≥ ui (s′i , µi ) for all s′i ∈ Si ,
P
where again we abuse notation and use ui (si , µi ) = s−i ∈S−i µi (s−i )ui (si , s−i ), in
the finite case; in the infinite case, ui (si , µi ) is the expected value of ui (si , · ) with
respect to µi ∈ ∆(S−i ). In other words, si is rational if it is a best response to some
conjecture. We denote this by si ∈ BRi (µi ).

It is clear that a dominated strategy cannot be a rational choice – it is dominated

pointwise, there’s no way to find a weighted average that will do at least as well as

6
the weighted average associated with the dominating strategy. The converse is true
as well. To prove this central result, let us use the following useful result, which is
an implication of Farkas’ Lemma. Let A be an (m × n) matrix and b be an m-
dimensional vector. Then either the system Ax ≤ b has a solution x ≥ 0 or yA ≥ 0,
yb < 0 has a solution y ≥ 0, but not both.3

Assume that S is a finite set for now.

Proposition 4.1. A strategy s∗i is rational if and only if there does not exist a mixed
strategy σi ∈ ∆(Si ) such that ui (σi , s−i ) > ui (s∗i , s−i ) for all s−i ∈ S−i .

Proof. Let A be a matrix with |S−i | rows and |Si | columns, with typical entry given
by ui (s∗i , s−i ) − ui (s′i , s−i ), where s′i runs over Si and s−i runs over S−i :

ui (s∗i , s−i ) − ui (s′i , s−i ) · · · ui (s∗i , s−i ) − ui (ŝi , s−i )

 

A=
 .. 
. 
ui (s∗i , ŝ−i ) − ui (s′i , ŝ−i ) · · · ui (s∗i , ŝ−i ) − ui (ŝi , ŝ−i )

and let b = (−1, ..., −1). If there exists x ≥ 0 such that Ax ≤ b, then there exists
a probability distribution σi ∈ ∆(Si ) such that ui (σi , s−i ) > ui (s∗i , s−i ) for all s−i .
In fact, x 6= 0 because Ax < 0, so we can normalize x to obtain a probability
)
distribution σi ∈ ∆(Si ) (σi (si ) = P ′ x(si x(s ′ ) ). Likewise, if there exists y ≥ 0 such
s ∈Si i
i
that yA ≥ 0 and yb < 0, then s∗i is rational. If fact, y 6= 0 because yb < 0, so we
again can normalize y to obtain a probability distribution µi ∈ ∆(S−i ). And yA ≥ 0
then implies that ui (s∗i , µi ) ≥ ui (s′i , µi ), for all s′i . Now apply Farkas’ Lemma.

Here’s an example (the numbers are player 1’s utility numbers)

L R
U 3 0
M 0 3
D 1 1

showing that D, which is never a best response (that is, there’s no conjecture for
which D is best response), is dominated by a coin-flip between U and M.

From now on we say that a strategy is dominated if it is dominated by a mixed

strategy. As for weak dominance, you can verify the following:
3
Farkas’ Lemma states that either the system Ax = b has a solution x ≥ 0 or the system yA ≥ 0,
yb < 0 has a solution, not necessarily non-negative. The variant used here follows from applying
the Lemma to the matrix [I A], where I is the identity matrix.

7
Proposition 4.2. A strategy is weakly dominated by a mixed strategy if and only if
it is not a best response to a conjecture with full support (that is, a conjecture that
assigns positive probability to every s−i ∈ S−i .)

A strategy that is a best response to a conjecture with full support is called an

admissible strategy. Here’s an example (again the numbers are player 1’s utility
numbers)
L C R
U 2 0 2
M 0 2 0
D 1 1 0
showing that D, which is weakly dominated by a coin-flip between U and M, is never
a best response to a conjecture what puts positive probability to L, C, and R. That
is, D is not admissible.

5 Common Knowledge of Rationality

It is apparent that we can safely predict and recommend that a dominated (i.e.
not rational) strategy not be chosen. But this hardly solves the problem in most
strategic situations: if you have many choices to make and only one is not rational,
you are still far from a decision. But think about your opponents: if some of their
strategies are not rational, then it makes sense to assume that those are not going
to be chosen. And once some strategies of others are excluded, some of your own
strategies may become dominated. Eliminating these may reveal that some of others’
strategies become dominated. And so on.

Formally, the set of rationalizable strategies of player i is given by Ri∞ =

T∞ k 0
k=0 Ri , where Ri = Si and for each k > 0

Rik = {si ∈ Si : ∃µi ∈ ∆(R−i

k−1
) s.t. si ∈ BRi (µi )}.

Let Si∞ = ∞ k
T
k=0 Si be the set of strategies that survive iterated elimination
of dominated strategies (IEDS): Si0 = Si and for each k > 0

Sik = {si ∈ Sik−1 : ∄σi ∈ ∆(Sik−1 ) s.t. ui (σi , s−i ) > ui (si , s−i )∀s−i ∈ S−i
k−1
}.

From Proposition 4.1 we have Ri∞ = Si∞ .

The important point here is the assumption of common knowledge of rationality:

everyone knows that everyone knows that ... everyone is rational. In order to make

8
this statement precise, we use type structures, much in the same way that we
used them to handle incomplete information. Let Ti be a space of possible types of
player i, with typical element ti ∈ Ti , and let λi : Ti → ∆(S−i × T−i ) be a belief
mapping. A game with a type structure is the tuple (Si , ui , Ti , λi )ni=1 . We view
points in S × T (where T = ×ni=1 Ti ) as states: a state determines the profile s ∈ S
played, and also the profile of types t ∈ T . Given the belief mappings λi , we can
obtain the conjecture of type ti (namely margS−i λi (ti )), and also what players know:
for a given event E ⊂ S−i × T−i , the set

Ki (E) = {ti ∈ Ti : λi (ti )(E) = 1}

is the set of types where player i is sure of E. (We sometimes abuse terminology
and say that Ki (E) are the types that “know” the event E, but it is more precise
to say that these types are certain of E.)

Let
R̃i1 = {(si , ti ) : si ∈ BRi (margS−i λi (ti ))}
be the strategy-type pairs of player i that are rational. For each k > 1, let

R̃ik+1 = R̃ik ∩ [Si × Ki (R̃−i

k
)]

and put R̃i∞ = k≥1 R̃ik . Then R̃∞ = ×ni=1 R̃i∞ is the set of states where rationality
T

is common knowledge: for any (s, t) ∈ R̃∞ , players are rational because (si , ti ) ∈
R̃i1 for each i; also players know that they are rational because (si , ti ) ∈ R̃i2 for each
i; and they know that they know that they are rational because (si , ti ) ∈ R̃i3 for each
i; and so on ad infinitum.

Assume that λi is surjective. This means that we assume that the players do
not know much about each other, as they do not exclude any belief a priori. Then
by construction we have Ri1 = projSi R̃i1 . By induction, the same is true for each
k > 1: (si , ti ) ∈ R̃ik means that si is a best response to a conjecture supported
k−1 k−1
in projS−i R̃−i = R−i . That is, si ∈ Rik . Conversely, if si ∈ Rik then it is a best
k−1 k−1 k−1 k−1 k−2
response to a conjecture µi ∈ ∆(R−i ). Because projS−i R̃−i = R−i , R̃−i ⊂ R̃−i ,
k−1
and λi is surjective, there is ti with λ(ti )(R̃−i ) = 1 and margS−i λi (ti ) = µi , which
then means that (si , ti ) ∈ R̃ik . So we conclude that

Ri∞ = projSi R̃i∞

for each i, verifying that in fact the rationalizable strategies (which are the strategies
surviving IEDS) are the ones arising when there is common knowledge of rationality.

Example. 2/3 of Average: Each of n people is asked to write down a num-

ber from the set {1, 2, ..., 100} in a piece of paper, without knowing the numbers

9
written by the others. The n numbers are collected, and the average is computed.
The person who writes the number closest to 23 of this average wins. The unique
rationalizable choice is to write down the number 1: the best conjecture to justify
writing down 100 is that everyone would do the same; but the best response to this
conjecture is to write down 67; eliminating 100 for everyone, the same logic shows
that 99 is never a best response; and so on until all that’s left is to write down 1.

The example above is perhaps the best example of how quickly “recommenda-
tions” face a dead-end: in experiments, subjects usually write numbers in the range
[15, 25], perhaps because no one is actually sure about the rationality of the others.
If I am not confident that everybody will go through the same mental process, and
some may well write down numbers well above 1, I myself would do well to write a
number, say 20, that is more likely to be the winning number than 1. The challenge
here is to come up with a theory behind this informal statement, for writing number
20 is not a sustainable outcome because the others will eventually anticipate it. That
is, it is a boundedly rational outcome, which will likely be recognized by intelligent
players. Nevertheless, as in the 2/3 of average game, rationalizability picks up a
single profile of strategies in some particular games:

Example. Hotelling Location Game: Two players choose locations in the

set {0, k1 , k2 , ..., 1}, where k is even. If xi < xj then player i captures the portion
x +x
[0, i 2 j ] and player j captures the remainder of the interval [0, 1]; if xi = xj then
each player gets 12 . Assume that the payoff is equal to the size of the interval
captured. The unique rationalizable profile is ( 12 , 12 ): 0 and 1 are dominated by k1
1
and k−1k , respectively. Eliminating 0 and 1 for both players, k and k become
k−1

dominated by k2 and k−2 k , and so on until we reach the mid point.

Example. Cournot Duopoly: Two identical firms compete in a market, each

chooses a quantity qi ∈ [0, q̄] not knowing the choice of the other. The price that
prevails depends on total quantity supplied, given by Q = q1 + q2 . That is, the
inverse demand is p(Q) = (a − Q), where a > 0 is large enough. Each firm has a
constant marginal cost c, so that firm i’s profit is πi (qi , qj ) = (a − Q)qi − cqi . As it
is linear in qj , the best response to a conjecture is the best response to the expected
a−c−qj
qj . As BRi (qj ) = 2 for i = 1, 2 and j 6= i, plotting such functions we see that
firm i will not choose any qi above a−c 2 , as these would only be BR’s to negative qj ’s;
eliminating these, we see that qj < a−c 4 can be eliminated as well, as they’d only
a−c
be BR’s to qi ’s above 2 . Continuing in this fashion we conclude that the unique
rationalizable profile is the pair (q1 , q2 ) where the best response functions intersect,
that is, q1 = q2 = a−c 3 .

Example. Duopoly with Differentiated Products: Two firms compete in

a market, each chooses a price pi ∈ [c, p̄], where c is the (common) marginal cost and

10
p̄ > c, not knowing the choice of the other. The products are imperfect substitutes:
the demand for good i is given by a − pi + pj , for some large a > 0. Following the
same logic of the previous example, we find and plot the best response functions
a+c+pj
BRi (pj ) = 2 , eliminate the never best response prices and continue until we
reach the intersection p1 = p2 = a + c.

It is clear that a dominated strategy si need not be dominated by a pure strategy.

It follows that the rationalizable strategies also survive a process of elimination of
strategies dominated by some pure strategy. Let us call a strategy si serially
undominated if it survives this weaker elimination process. Formally, given Ŝ =
×N ′ ′
i=1 Ŝi , let Ui (Ŝ) = {si ∈ Si : ∀si ∈ Si , ∃ŝ ∈ Ŝ such that ui (si , ŝ−i ) ≥ ui (si , ŝ−i )},
and put U (Ŝ) = ×N i=1 Ui (Ŝ). Then the set of serially undominated strategies is
D∞ = ∞ k , where D 0 = S and, for all k > 0, D k = U (D k−1 ). Observe that U
T
k=0 D
is monotone nondecreasing (U (Ŝ) ⊂ U (S ′ ) whenever Ŝ ⊂ S ′ ), and that S ∞ ⊂ D ∞ .

5.1 Supermodular Games

There is a wide class of normal form games where rationalizability goes quite far. It
is the class of supermodular games, particularly relevant in economic applications.
A partially ordered space is a set X with a reflexive, transitive and anti-symmetric
binary relation ≥ defined on it. For any subset X ′ ⊂ X the supremum sup(X ′ )
is the least upper bound, where an upper bound for X ′ is an element b ∈ X such
that b ≥ x′ for all x′ ∈ X ′ . The infimum (greatest lower bound) inf(X ′ ) is defined
analogously. A lattice is a partially ordered set such that the supremum x ∨ y and
the infimum x ∧ y of every two point subset {x, y} exist. A lattice is a complete
lattice if it contains the supremum and infimum of each of its nonempty subsets.
A sublattice is a subset of a lattice which is closed under pairwise ∨ and ∧. A
sublattice X ′ of a lattice X is complete if supX X ′′ ∈ X ′ and inf X X ′′ ∈ X ′ for every
nonempty X ′′ ⊂ X ′ . Finally, let [x, y] be the set of points z such that z ≥ x and
y ≥ z.

A function f : X → R is called supermodular if

f (x) + f (y) ≤ f (x ∨ y) + f (x ∧ y)

When X ⊂ Rk and f is twice continuously differentiable, supermodularity is

2f
captured by ∂x∂i ∂x j
≥ 0 for every 1 ≤ i, j ≤ k.

Supermodularity captures the economic notion of complementary inputs. Given

two lattices X and Y , a function f : X × Y → R has increasing (decreasing) dif-
ferences in x and y if for all x′ ≥ x the difference f (x′ , y) − f (x, y) is nondecreasing

11
(nonincreasing) in y. Interpreting y as the choices of the other players, increasing
differences capture the idea of strategic complements, and decreasing differences cap-
ture the idea of strategic substitutes. It is easy to see that a supermodular function
f : X × Y → R has increasing differences in x and y, but the converse is not neces-
sarily true. When X ⊂ ×ki=1 Xi we say that f : X → R has increasing differences in
X when it has increasing differences in each pair (xi , xj ) when all other coordinates
are held fixed, and then it is not difficult to show that f : X → R has increasing
differences in X ⊂ Rk if and only if it is supermodular. Hence, when X ⊂ Rk and
f : X → R is twice continuously differentiable, increasing (decreasing) diferences is
2f
captured by ∂x∂i ∂x j
≥ 0 (≤ 0) for every 1 ≤ i, j ≤ k, i 6= j.

Let X and Y be lattices, and f : X × Y → R be supermodular in X and has

increasing differences in (x, y). Let b(y) = arg maxx f (x, y), and assume for now
it is non-empty for every y. For y ′ ≥ y, let x ∈ b(y) and x′ ∈ b(y ′ ). Then it
must be that x ∨ x′ ∈ b(y ′ ) and x ∧ x′ ∈ b(y). Indeed, as f (x, y) ≥ f (x ∧ x′ , y),
supermodularity implies that f (x ∨ x′ , y) ≥ f (x′ , y), and then increasing differences
gives us f (x ∨ x′ , y ′ ) ≥ f (x′ , y ′ ), meaning that x ∨ x′ ∈ b(y ′ ). The other claim is
similarly proved (this result is known as Topkis monotonicity theorem). In particular,
setting y = y ′ , we get that b(y) is a sublattice of X. When f is in addition order
continuous,4 bounded above and X is a complete lattice, then it can be shown that
b(y) is a non-empty complete sublattice of X for every y. In this case, we can
then define b(y) = sup(b(y)) and b(y) = inf(b(y)) as the “top” and the “bottom”
of b(y), and Topkis’ result then implies that b : Y → X and b : Y → X are
nondecreasing functions (that is, b(y ′ ) ≥ b(y) for y ′ ≥ y, and similarly for b). In
fact, b(y ′ ) ∨ b(y) ∈ b(y ′ ), so b(y ′ ) ≥ b(y ′ ) ∨ b(y) ≥ b(y), and similarly for the other
case.

A lattice game with strategic complements (substitutes) is a game (Si , ui )ni=1

such that, for each i

• Si is a complete lattice

• ui is order-continuous and bounded above

• ui is supermodular in Si

• ui has increasing (decreasing) differences in si and s−i .

Lattice games with strategic complements are also called supermodular games.
4
Order-continuity is convergence along every chain, where a set C is a chain if x, y ∈ C implies
that either x ≥ y or y ≥ x. In fact, for the result I am about to mention, all that is required is
order-upper semicontinuity.

12
 Example. The multi-firm extension of the duopoly with differentiated products
is supermodular: n firms compete in a market, each chooses a price pi ∈ [ci , p̄], where
ci is firm i’s marginal cost and p̄ > ci , not knowing the choice of the others. The
demand for good i, D i (p1 , ..., pn ) is twice continuous differentiable, with Dii < 0,
Dji > 0, Diii ≤ 0 and Dij i ≥ 0 for all i and j 6= i. Firm i’s profit is given by

πi (p) = (pi − ci )D i (p).

2
In fact, ∂p∂ i ∂p
πi
j
i + D i ≥ 0, so the game has increasing differences,
= (pi − ci )Dij j
and other conditions defining a supermodular game are immediately verified.

Let BRi : S−i ⇒ Si be given by BRi (s−i ) = arg maxŝi ui (ŝi , s−i ) and let BR :
S ⇒ S be given by BR(s) = ×ni=1 BRi (s−i ). From what we established above,
BRi (s−i ) is a complete sublattice of Si whenever the game is supermodular.

For each s ∈ S, let bi (s) = sup(BRi (s−i )) and bi (s) = inf(BRi (s−i )), and let
b(s) = (b1 (s), ..., bn (s)), b(s) = (b1 (s), ..., bn (s)). These are well defined because
BRi (s−i ) is a complete sublattice of Si . As we established above, both b and b
are monotone non-decreasing functions, so we see that a supermdular game has
the property that the top and the bottom of the best response correspondence are
nondecreasing. Finally, let U (Ŝ) = [inf U (Ŝ), sup U (Ŝ)]. We can now prove the
following general result:

Proposition 5.1. In a lattice game with strategic complements, D ∞ ⊂ [s, s], where
s and s are fixed points of BR.

Proof. Let s0 = sup(S) and s0 = inf(S), and for k > 0 let sk = b(sk−1 ) and
sk = b(sk−1 ). We show by induction that U k (S) ⊂ [sk , sk ]. It is true for k = 0. If it
is true for some k ≥ 0 then U k+1 (S) ⊂ U ([sk , sk ]) because U is monotone, so we need
to show that U ([sk , sk ]) ⊂ [sk+1 , sk+1 ]. Assume, by way of contradiction, that there
is ŝ ∈ U ([sk , sk ]) and ŝ ∈/ [sk+1 , sk+1 ]. In particular, say that ŝi  sk+1 i . Because
ŝi ∈ Ui ([sk , sk ]), there must exist s′ ∈ [sk , sk ] such that ui (ŝi , s′−i ) ≥ ui (sk+1
i ∧ŝi , s′−i ).
But for any s′ ∈ [sk , sk ] we have

ui (ŝi , s′−i ) − ui (sk+1

i ∧ ŝi , s′−i ) ≤ ui (ŝi , sk−i ) − ui (sk+1
i ∧ ŝi , sk−i )
≤ ui (sk+1
i ∨ ŝi , sk−i ) − ui (sk+1
i , sk−i )
< 0,

a contradiction. A similar reasoning applies for the other side of the interval (ŝi 
sk+1
i ).

Now, because sk+1 , sk+1 ∈ U ([sk , sk ]), we have U ([sk , sk ]) = [sk+1 , sk+1 ]. We
now claim that [sk+1 , sk+1 ] ⊂ [sk , sk ] for every k ≥ 0. It is true for k = 0. Assume

13
it is true up to k − 1, and note that

[sk+1 , sk+1 ] = U ([sk , sk ]) ⊂ U ([sk−1 , sk−1 ]) = [sk , sk ]

verifying the claim.

It follows that {sk } is monotone non-decreasing and {sk } is monotone non-

increasing, so s = lim sk = sup{sk } and s = lim sk = inf{sk } exist, and D ∞ ⊂ [s, s].

Alternatively, we could directly have established monotonicity of the sequences

from monotonicity of b and b: s0 ≥ s1 implies that s1 = b(s0 ) ≥ b(s1 ) = s2 , and so
on to establish that {sk } is a non-increasing sequence (and similarly we could have
used b to establish that {sk } is a non-decreasing sequence.)

Now, if s were not a fixed point of BR, there would exist i and si such that
ui (si , s−i ) > ui (s). But order-continuity would imply that ui (si , sk−i ) > ui (sk+1
i , sk−i )
for some finite k, a contradiction. Similarly, we can show s must be a fixed point as
well, so we are done.

As for strategic substitutes we have the following:

Proposition 5.2. In a lattice game with strategic substitutes, D ∞ ⊂ [s, s], where s
and s are fixed points of BR2 , where BR2 = BR ◦ BR.

There are ordinal generalizations of supermodularity and increasing (decreasing)

differences. For a lattice X, a function f : X → R is called quasisupermodular
if f (x ∨ y) ≥ f (y) whenever f (x) ≥ f (x ∧ y), and f (x ∨ y) > f (y) whenever
f (x) > f (x∧y). It is immediate that a supermodular function is quasisupermodular.
Likewise, for two partially ordered sets X and T , a function f : X × T → R has the
single crossing property with increasing (decreasing) returns (abbreviated
as SCP-IR (SCP-DR)) if whenever we have x′ ≥ x and t′ ≥ t, it must be that
f (x′ , t) > f (x, t) implies f (x′ , t′ ) > f (x, t′ ) (f (x′ , t) < f (x, t) implies f (x′ , t′ ) <
f (x, t′ )), and also that f (x′ , t) ≥ f (x, t) implies f (x′ , t′ ) ≥ f (x, t′ ) (f (x′ , t) ≤ f (x, t)
implies f (x′ , t′ ) ≤ f (x, t′ )). Again, a function satisfying increasing (decreasing)
differences in X and T satisfies SCP-IR (SCP-DR).

Using quasisupermodularity in the place of supermodularity and SCP-IR (SCP-

DR) in the place of increasing (decreasing) differences in the definition of lattice
games we obtain lattice games with ordinal strategic complementarities
(substitutes). Propositions 5.1 and 5.2 remain valid using these weaker notions of
complementarity/substitutability.

Summing up, whenever BR (resp. BR2 ) has a unique fixed point in a lattice

14
game with (ordinal) strategic complements (resp. substitutes), there is a unique
rationalizable strategy. These two conditions are met in several applications.

5.2 Admissibility

We can extend the ideas involved above to weakly dominated strategies. That is,
let Wi∞ = ∞ k 0
T
k=0 Wi , where Wi = Si and for each k > 0

Wik = {si ∈ Wik−1 : ∄σi ∈ ∆(Wik−1 ) that weakly dominates si }.

The profiles in W ∞ = ×ni=1 Wi∞ are those that survive Iterated Elimination
of Weakly Dominated Strategies (IEWDS). The dual procedure is as follows. Let
A0i = Si for i = 1, ..., n and for each k > 0

Aki = {si ∈ Aik−1 : ∃µi ∈ ∆0 (A−i

k−1
) with si ∈ BRi (µi )}.

where ∆0 (X) is the space of probability distributions in X with full support. Put
T∞
A∞
i =
k
k=0 Ai . The profiles in A
∞ = ×n A∞ are those that survive Iterated
i=1 i
Admissibility (IA). From Exercise 4.2, we know that W ∞ = A∞ .

Although the IEWDS (IA) procedure is quite similar to IEDS and rationalizabil-
ity, the need for fully supported conjectures makes its epistemic characterization a
bit problematic. Consider the following simple game

L R
U 1, 1 0, 1
D 0, 2 1, 0

where player 2 has a weakly dominated strategy, R. If we assume that players play
only admissible strategies, then player 1’s conjecture must assign positive probability
to R. But R is not admissible for player 2, so player 1 should assign probability zero
to R. Hence, player 1 should at the same time exclude R (that is, know that it will
not be played) and include R (that is, use a conjecture with full support.) If we
capture knowledge as probability 1 belief, player 1 cannot know that player 2 plays
admissible, for player 1 does not assign probability 1 to L.

Indeed, if one tries to mimic the analysis that we made above to model common
knowledge of rationality, one runs into problems, as follows. Let (Ti , λi )ni=1 be a
type structure appended to a game (Si , ui )ni=1 . From Proposition 4.2, a strategy
is admissible iff it is a best response to a fully supported conjecture. A strategy-
type pair (si , ti ) is admissible if si is a best response to the conjecture margS−i λi (ti )
and margS−i λi (ti )(s−i ) > 0 for all s−i ∈ S−i . Let Ãi denote the set of admissible

15
strategy type pairs. Common knowledge of admissibility would then be captured by
T∞
Ã∞ = ×ni=1 Ã∞ ∞
i , with Ãi =
k
k=0 Ãi and

Ãki = Ãik−1 ∩ [Si × Ki (Ã−i

k−1
)],

for all k > 0. But say that the profile ŝ−i is not admissible. Then ŝ−i ∈ / projS−i Ã1−i .
Now if (si , ti ) ∈ Ã1i , then (si , ti ) ∈
/ Ã2i , for margS−i λi (ti )(ŝ−i ) > 0, so ti ∈
/ Ki (Ã1−i ).
In short, Â∞ = ∅, so the “language” used above does not describe common knowl-
edge of admissibility.

References:

Farkas’ Lemma can be found is most linear algebra books. I particularly like the
treatment in the book “Theory of Linear and Integer Programming”, by A. Schri-
jver. The result that a strategy is dominated if and only if it is not a best response
to a conjecture holds in more general settings and follows from classical separating
hyperplane ideas. See for instance A. Zimper “Equivalence between Best Responses
and Undominated Strategies: A Generalization from Finite to Compact Strategy
Sets,” Economics Bulletin (2005), 3(7). The classical reference for the knowledge
foundations of IEDS is T. Tan and S. Werlang “The Bayesian Foundation of Solution
Concepts of Games,” Journal of Economic Theory (1988), 45. See also “A. Bran-
denburger, A. Friedenberg, and J. Keisler “Admissibility in Games,” Econometrica
(2008), 76(2) for ideas related to admissibility. The material on supermodular games
comes from P. Milgrom and J. Roberts “Rationalizability, Learning, and Equilibrium
in Games with Strategic Complementarities,” Econometrica (1990), 58(6), and P.
Milgrom and C. Shannon “Monotone Comparative Statics,” Econometrica (1994),
62(1).

16