Class Notes Final Version
Uploaded by
Leonardo SimonciniClass Notes Final Version
Uploaded by
Leonardo SimonciniGame Theory and Microeconomics
(Classnotes)
Domenico Menicucci∗
December 6, 2021
Contents
1 An introduction to Choice Theory 2
2 Choice under uncertainty 4
2.1 Lotteries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.2 The Expected Utility Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3 The case of = R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.4 Stochastic dominance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.5 Risk attitudes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.5.1 Risk neutrality, Risk aversion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.5.2 Measuring risk aversion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3 Game Theory 17
3.1 Games with simultaneous moves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.1.1 Dominant strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.1.2 Nash Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.1.3 Mixed strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.1.4 Games with incomplete information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.2 Dynamic games . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.2.1 The extensive form of a game . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.2.2 Strategies, and Normal form representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.2.3 Behavior strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.2.4 Subgame perfection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.2.5 Refinements based on beliefs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4 Consumer Theory and General Equilibrium 54
4.1 The consumer’s problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
4.1.1 Preference relations and utility functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
4.1.2 Monotonicity, Convexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.1.3 Utility maximization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.2 General equilibrium for pure exchange economies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.2.1 Definition of Walrasian Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
∗I thank Geronimo Bartolomei for pointing out a few typos in a previous version of this document.
1
4.2.2 Pareto efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.2.3 The Fundamental Theorem of Welfare Economics . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.2.4 Appendix to Section 4: Existence of multiple Walrasian Equilibria . . . . . . . . . . . . . . . 75
This document is a summary of the classes for the course of Game Theory and Microeconomics, but Subsections
2.5.2, 3.2.5, the material in Subsection 4.1.3 related to optimal bundles for each ≥ 2, and the appendix to Section
4 are not relevant towards the exam. Useful books (perhaps at a high level) for this course are
• MWG: Microeconomic Theory, Mas-Colell A., Whinston M., Green J., Oxford University Press, 1995;
• JR: Advanced Microeconomic Theory, Jehle, G.A., Reny, P.J., Financial Time Prentice Hall, 2011.
1 An introduction to Choice Theory
Microeconomic theory is based on agents which pursue their objectives, and examines how these agents’ individual
actions determine collective outcomes. Here we describe a way to illustrate a generic choice problem for an agent,
called decision maker — dm henceforth — pursuing his own interest.
Preference relations Consider a dm facing a set of alternatives, 6= ∅. The dm must choose one element
in .
Example 1.1 Lunch choice: = {Students’ cafeteria, Pizza restaurant, Ordinary restaurant, Food from home,
...}.
A key assumption in this context is that the dm is endowed with a preference relation on , denoted with %,
which allows him to compare alternatives in . The interpretation of % is as follows: Given in , % means
that is at least as good as for dm, that is if % holds then dm views as no worse than .
Example 1.2(i) Suppose that = {1 2 3 4 } and
1 % 1 , 1 % 4 , 2 % 1 , 2 % 2 , 2 % 4 , 3 % 1 , 3 % 2 , 3 % 3 , 3 % 4 , 4 % 1 , 4 % 4 (1)
Since 1 % 4 , it follows that 1 is at least as good as 4 for dm. Likewise, 2 % 1 reveals that 2 is at least as
good as 4 . Conversely, 4 % 2 does not hold; therefore 4 is not at least as good as 2 , that is 4 is inferior to 2 .
From % we can derive two other relations on :
• the strict preference relation Â, defined as follows:
 if and only if % but not %
the interpretation of  is that is better than for dm: he prefers to .
• the indifference relation ∼, defined as follows:
∼ if and only if % and %
the interpretation of ∼ is that dm is indifferent between and : he views and as equally good.
Example 1.2(ii) From (1) we deduce that 3 Â 2 , 3 Â 1 , 3 Â 4 , 2 Â 1 , 2 Â 4 , 1 ∼ 4 . Hence the best
alternative in is 3 , and dm’s choice will be 3 .
Example 1.2(ii) illustrates how, starting from %, we derive  and ∼ which clarify the tastes of dm over . For
each in we have determined whether  ( is better than ), or  ( is better than ), or ∼ (they
are equally good). This reveals that 3 is the best element in .
2
This approach requires that dm is able to compare two alternatives at the time, any two alternatives, through
%. These binary comparisons generate  and ∼, which yield a ranking for all the alternatives in . In a sense, %
is a machine which produces a ranking. It is natural to inquire the origin of %, and the answer is that % comes
from dm himself: given , by looking into himself the dm understands which alternative is better for him between
and . Maybe his personal history matters, or his cognitive abilities, ..., but all these elements are summarized
by %.
We consider % as a primitive data which is not explained but is taken as given. However, we require that %
satisfies some properties which guarantee that dm’s preferences are well organized.
• Completeness Given and %, we say that % is complete if for each in we have % , or % , or
both.
This property requires that each alternative in is comparable with each other, that is it never happens that
% and % are both untrue. This implies that either  , or  , or ∼ : in no case dm is indecisive.1
Verify that % in (1) is complete.
• Transitivity Given and %, we say that % is transitive if for each in such that % and % , we
have % .
This is more or less a property of consistency among binary comparisons and is simple to interpret: If is at
least as good as , and is at least as good as , then must be at least as good as . Transitivity is a crucial
property because it rules out cases of indecisiveness generated by cycles. Verify that % in (1) is transitive.
We will consider only % complete and transitive, and this implies that  and ∼ are both transitive. However,
 and ∼ are typically not complete; see for instance  and ∼ derived from (1).
Utility functions The preference relation % is the starting point to describe dm’s preferences, but sometimes
it is possible to give a numeric representation of % which is more practical. Precisely, sometimes it is possible to
describe % through a function : → R which attaches to each ∈ a numeric value in such a way that higher
values are assigned to the best alternatives, as expressed by this definition.
Definition 1.3 We say that a function : → R represents % if for each in we have
 ⇔ () ()
% ⇔ () ≥ () which is equivalent to (2)
∼ ⇔ () = ()
If there exists which satisfies (2), then is called utility function for dm. The condition in (2) guarantees
that orders the alternatives in as % does. Hence, given in , the values () () reveal dm’s preferences
between and . In a sense, attaches a beauty index to each alternative in a way that the higher is the index,
the better is the alternative.
Example 1.2(iii) Given = {1 2 3 4 } and % in (1), which implies 3 Â 2 Â 1 ∼ 4 , a function which
represents % is : → R such that (1 ) = 8, (2 ) = 11, (3 ) = 17, (4 ) = 8 as it orders the alternatives as %.
The role of is to summarize the informations contained in % in order to produce the same ranking of alternatives
as %. The utility numbers matter only in determining a ranking, hence if there exists a which represents %,
then there exist many other functions which represent %: each other function which generates the same ranking of
alternatives as the one generated by %.2
1 In reality, for a dm it is sometimes difficult to compare two alternatives if they are not familiar to him. Referring to Example 1.1,
think of a restaurant serving only food which is typical for a certain country, of another restaurant serving only food which is typical
for another country, and of a dm that never tasted neither sort of food.
2 For instance, if : R → R is strictly increasing then the composite function u = ◦ , such that u() = (()) for each ∈ ,
represents %.
3
Example 1.2(iv) The following function represents % in (1): : → R such that (1 ) = −1, (2 ) = 0,
(3 ) = 4, (4 ) = −1.
If there exists which represents %, then the best element in is a max point for , an alternative where
takes on its max value: see Example 1.2(iii) for instance. Hence dm’s choice problem can be seen as the problem
of maximizing .
Since representing preferences through a utility function is often more practical than representing preferences
through a preference relation, it is natural to ask whether given % which is complete and transitive, there exists
which represents %. The answer depends on the context. If is a finite set, then necessarily which represents %
exists. But if consists of infinitely many alternatives, then in some cases the answer is affirmative, in other cases
it is negative.
For the case of finite , for each ∈ , define the set () = { ∈ : % }; this is the set of alternatives
which are no better than and is called the lower contour set of . Then consider : → R such that
() =cardinality of () for each ∈ ; since is a finite set, also () is finite and () is well defined. This
represents %.
Example 1.2(v) Given = {1 2 3 4 } and % in (1), we have (3 ) = {1 2 3 4 }, (2 ) = {1 2 4 },
(1 ) = (4 ) = {1 4 }. Hence such that (3 ) = 4, (2 ) = 3, (1 ) = (4 ) = 2 represents %. The best
alternative, 3 , is a max point for .
Multiple best alternatives In some cases, that is for some %, there exist multiple best elements rather than
only one.
Example 1.4(i) Given = {1 2 6 } and 3 ∼ 5 Â 2 Â 4 Â 1 ∼ 6 , there exist two best alternatives, 3
and 5 , and dm is equally happy with either of them.
A restricted set of feasible alternatives Sometimes, not all the alternatives in are available for dm, and
only the alternatives in a subset 0 of are feasible. Then the choice of dm is given by a best element in 0
according to %, or a max point for among the elements in 0 .
Example 1.4(ii) Suppose that 0 = {1 2 4 } is the set of feasible alternatives, that is dm cannot choose 3 ,
nor 5 , nor 6 . Then he will choose his most preferred element in 0 , which is 2 .
The framework introduced here is the basic structure to represent a dm’s choice problem: The ingredients are
0 %, and dm chooses a best element in 0 according to %, or (if representing % exists) a max point for in
0 .
2 Choice under uncertainty
The previous section has introduced a model of choice in which a dm selects an alternative from a set of feasible
alternatives. Now we consider a particular case of this model in which the dm’s choice leads to an outcome which
is uncertain at the time he takes his decision.
Example 2.1 If we invest today 6 000 euros in the Italian Stock market, it is quite possible that in one year from
now the investment’s value will be 5 300 euros, or 6 200, or 6 900, ... As of today, it is impossible to know the value
of this investment in one year.
2.1 Lotteries
We consider a setting in which dm must choose one among a few risky alternatives, and each of them may result
in one of a number of possible outcomes. We denote with the set of the possible outcomes, and to fix the ideas
4
we think of as a set of monetary prizes. We suppose initially that is finite, that is = {1 2 } for some
≥ 2. We use lotteries to represent risky alternatives.
Simple lotteries Simple lotteries are defined as follows:
Definition 2.2 Given = {1 2 }, a simple lottery is a probability distribution on , identified by the
numbers 1 ≥ 0 ≥ 0 such that 1 + + = 1 and, for = 1 , is the probability that the outcome
is realized given the lottery .
A simple lottery is just a probability distribution over which specifies the (objective) probability the lottery
assigns to each element of . In the terminology of probability theory, a simple lottery identified by 1 is a
random variable such that realized value of the random variable will be one element in , with having probability
, for = 1 .
Example 2.3(i) The lottery
−100 0 50 200
=
02 01 06 01
is interpreted as follows: the dm receives a prize of 200 with probability 01, a prize of 50 with probability 06, pays
100 with probability 02, and receives/pays nothing with probability 01.
In the setting of Example 2.3(i), with = {−100 0 50 200}, can be represented by the vector (02 01 06 01)
which specifies the probabilities of the prizes in . In the same setting, the vector (05 0 0 05) represents another
lottery, denoted 0 , which gives probability 12 to the prize −100, probability 12 to the prize 200, and zero probability
to each other prize.
When the set is unambiguous, in order to represent a simple lottery it suffices to specify the probabilities
1 , thus we write = (1 ); otherwise we describe both the set and the probabilities. The following
is a graphical representation for :
q1
z1
q2
z2
Figure 2.4:
qn
zn
We denote with L the set of all simple lotteries for which the set of possible prizes is included in , that is L
is the set of all the vectors (1 ) such that 1 ≥ 0, ..., ≥ 0 and 1 + + = 1. This includes the so-called
degenerate lotteries, which give probability one to a single prize. For instance, 2 denotes the lottery such that
2 = 1 and = 0 for each 6= 2; hence 2 = (0 1 0 0) yields the prize 2 with certainty. In a sense, 2 is
not a real lottery as it involves no uncertainty. The set L includes the lotteries 1 , 2 , ..., .
The dm’s problem is to select one lottery from a set L0 of lotteries which is a subset of L. For instance, in the
context of Example 2.3(i), L0 might consist of and 0 . In the terms of the choice model of Section 1, the set L
corresponds to the set of all conceivable alternatives, and L0 plays the role of the set 0 of feasible alternatives.
5
Compound lotteries Before we deal with the choice problem, we need to introduce lotteries which are not
simple lotteries. In a simple lottery the outcomes are certain: at the end of each limb in Figure 2.4 lies a prize, but
a more general lottery may be such that one (or more) of its outcomes is itself a lottery.
Example 2.5(i) The following lottery is not a simple lottery:
50 0 −100 50 200
= with 0 =
03 07 04 01 05
0.3
50
0.4
Figure 2.6: -100
0.7 0.1
50
0.5
200
The lottery selects with probability 03 the upper limb in Figure 2.6, and then the dm wins 50. But with
probability 07, selects the lower limb and then a new lottery 0 is entered.
The lottery in Example 2.5(i) is a compound lottery, that is a lottery in which at least one of its outcomes
is another lottery. In general, a compound lottery consists of ≥ 2 simple lotteries 1 2 and numbers
1 2 such that 1 ≥ 0 2 ≥ 0 ≥ 0, 1 + 2 + + = 1, and 1 1 + 2 2 + + is
the compound lottery which selects 1 with probability 1 , selects 2 with probability 2 , ..., selects with
probability .3
α1
L1
α2
L2
Figure 2.7:
αm
Ln
Example 2.5(ii) In Example 2.5(i) we have = 2, 1 = 03, 2 = 07 and 1 = 50 , 2 = 0 .
For the compound lottery of Example 2.5(i) we can compute the ultimate probability of each prize: the
probability of −100 is 07 · 04 = 028, the probability of 0 is 0, the probability of 50 is 03 + 07 · 01 = 037, the
probability of 200 is 07 · 05 = 035. Then consider the simple lottery = (028 0 037 035) which assigns to each
prize the same probability as the compound lottery . If a dm must choose between the compound lottery and
the simple lottery , we assume he views them as equivalent since they induce the same probability distribution
on prizes. Hence, he is indifferent between them.
More generally, given any compound lottery = 1 1 + 2 2 + + , we can identify the simple lottery
that assigns to each prize in the same probability as ,4 and we assume that the dm views as equivalent
to because he only cares about the effective probability distribution on , and not about details such as the
number of stages in the lottery. This is the so-called consequentialist approach.
3 This is a two-stage compound lottery, but we can conceive compound lotteries with more than two stages, that is such that at least
one among 1 is a compound lottery.
4 Precisely, if = ( ) for = 1 , then = ( ) with equal to 1 + 2 + + , for each = 1 .
1 1 1 2
6
2.2 The Expected Utility Theorem
Here we examine a dm’s preferences over the set L of simple lotteries. We focus on L as we have illustrated above
that for each compound lottery there exists an equivalent simple lottery.
We suppose that the dm’s preferences are described by a preference relation % on L which we take as given as
in Section 1. From % we can derive the strict preference relation  and the indifference relation ∼:
 0 if and only if % 0 but not 0 % , ∼ 0 if and only if % 0 and 0 %
As we have mentioned in Section 1, it would be useful to represent % using a utility function, that is a function
: L → R such that for each 0 in L we have
% 0 ⇔ () ≥ (0 )
The most important result for the theory of choice under uncertainty states that if % satisfies suitable properties,
then a utility function for % exists and it is relatively simple. The properties, often called axioms, are the following.
Axioms for a numerical representation of % The first axiom is a familiar one, whereas the others are peculiar
to this environment.
• A1: Completeness, Transitivity For each 0 in L, we have that % 0 , or 0 % , or both; for each
0 00 , if % 0 and 0 % 00 , then % 00 .
We have already introduced these properties in Section 1, and the interpretations provided there apply here as
well.
• A2: Continuity For each 0 00 in L such that  0 , we have that
+ (1 − )00 Â 0 for close to 1 (3)
 0 + (1 − )00 for close to 1 (4)
In (3), the lottery + (1 − )00 is a compound lottery which selects with probability , selects 00 with
probability 1 − . The key observation is that if is about equal to 1, then + (1 − )00 is about equal to ,
that is, it is a small modification of , and (3) says that if  0 , then  still holds if we replace with a lottery
close to , that is  still holds after a small perturbation in the superior lottery. Likewise, 0 + (1 − )00 in (4)
is almost equal to 0 if is close to 1, and (4) says that given  0 ,  still holds if we replace 0 with a lottery
close to it, that is after a small perturbation in the inferior lottery.
• A3: Independence For each 0 00 in L we have that
% 0 if and only if + (1 − )00 % 0 + (1 − )00 for each ∈ (0 1) (5)
Although (5) establishes an equivalence, the key implication is that if % 0 , then + (1 − )00 % 0 +
(1 − )00 for each ∈ (0 1). In order to illustrate this, it is convenient to pick = 12 and notice that 12 + 12 00 is
a compound lottery which can be seen as a coin toss followed by if Heads comes up, but followed by 00 if Tails
comes up. A similar interpretation applies to 12 0 + 12 00 .
α α
L L′
Figure 2.8:
1-α L″ 1-α L″
7
Then, conditional on Heads we have 12 + 12 00 % 12 0 + 12 00 because % 0 ; conditional on Tails we have
2 + 2 % 2 + 2 because % . Therefore % holds in each case after the coin toss; A3 requires that %
1 1 00 1 0 1 00 00 00
holds also before the coin toss. The same principle applies to the case in which 6= 12 (using a biased coin).
From another viewpoint, notice that (1 − )00 appears in both compound lotteries + (1 − )00 and
0 + (1 − )00 . A3 says that the comparison between them is independent of the common term (1 − )00 : it
depends only on the other terms, and 0 . Since % 0 , it follows that + (1 − )00 % 0 + (1 − )00 .
The main theorem The next theorem establishes that if % satisfies A1-A3, then it is possible to represent %
with a utility function as follows:
Proposition 2.9 (The Expected Utility Theorem, EUT) Suppose that a preference relation % on L satisfies
A1, A2, A3. Then there exist numbers 1 , one for each element of , with the following property. For
each = (1 ), define : L → R as () = 1 1 + 2 2 + + . The function represents % according
to Definition 1.3, that is, for each % 0 in L (with 0 = (10 0 )), we have
% 0 ⇔ () ≥ (0 ) (6)
Example 2.3(ii) Suppose that = {1 2 3 4 } = {−100 0 50 200}, and that % satisfies A1-A3. Then the
EUT can be applied, and suppose that 1 = 2, 2 = 4, 3 = 7, 4 = 9. For the two lotteries = (02 01 06 01)
and 0 = (05 0 0 05) we have () = 2 · 02 + 4 · 01 + 7 · 06 + 9 · 01 = 59, (0 ) = 2 · 05 + 4 · 0 + 7· 0 + 9 · 05 = 55.
This reveals that  0 .
The numbers 1 can be thought of as the range of a function : → R such that ( ) = for each
= 1 . This is called Bernoulli utility function and is defined on the set of prizes, not on the set L of
lotteries.
The function is defined on the set L, that is on simple lotteries, and is called vonNeumann&Morgenstern expected
utility function because it is the expectation of . Precisely, () = (1 )1 + (2 )2 + + ( ) is the ex-
pectation of according to . It is important to notice that represents %, not ; is a sort of input (whose
existence is guaranteed by the EUT) that is needed to define ().
Summary: The EUT says that if the dm’s preference relation % on L satisfies A1-A3, then % can be represented
P
by a utility function as follows: There exists : → R such that : L → R defined as () = =1 ( ) (the
expectation of according to ) represents %. The EUT is typically applied as follows: (i) assume that % satisfies
A1-A3; (ii) by the EUT, there exists : → R such that : L → R represents %; (iii) make assumptions on
(not on %) which allow to compute () for each , which reveals the dm’s preferences on L.
Invariance of with respect to linear increasing transformations It is simple to see that the Bernoulli
utility function is not uniquely determined by the EUT. Given , consider : → R such that for each ∈ ,
() = () + for some 0 and some ∈ R. Then it is possible to use to replace , as described in the
P
following. Define : L → R as () = =1 ( ) ; hence () is the expectation of according to . We claim
that represents %, just like , because
X
X
X
X
() ≥ (0 ) ⇔ ( ) ≥ ( )0 ⇔ (( ) + ) ≥ (( ) + )0
=1 =1 =1 =1
X
X
⇔ ( ) ≥ ( )0 ⇔ () ≥ (0 ) ⇔ % 0
=1 =1
Example 2.3(iii) In Example 2.3(ii) we have assumed that (−100) = 2, (0) = 4, (50) = 7, (200) = 9.
Now consider such that () = 3() − 5. Then (−100) = 1, (0) = 7, (50) = 16, (200) = 22 and
8
() = 02 · 1 + 01 · 7 + 06 · 16 + 01 · 22 = 127, (0 ) = 05 · 1 + 05 · 22 = 115, hence () (0 ), consistently
with () (0 ).5
However, if we consider : → R such that is not equal to + for some 0, , then the resulting
does not represent %.
Example 2.3(iv) Consider such that () = (())2 , hence (−100) = 4, (0) = 16, (50) = 49, (200) = 81.
Then () = 02 · 4 + 01 · 16 + 06 · 49 + 01 · 81 = 399, (0 ) = 05 · 4 + 05 · 81 = 425. Hence () (0 ) fails
to hold, which is inconsistent with () (0 ).6
Sketch of the proof of the EUT The whole proof of the EUT is too long to be given here, but some basic
ideas are simple to convey.
Step 1 (definition of ̄) Consider the degenerate lotteries 1 , and in this set of lotteries identify a worst
lottery and a best lottery ̄ ( and ̄ necessarily exist because {1 } is a finite set and % satisfies A1).
Then ̄ % % for each ∈ L.
Step 2 (definition of ) For = 1 , define ( ) (that is, in the statement of the EUT) as the unique ∈ [0 1]
such that ∼ ̄ + (1 − ).
P
Step 3 (definition of and proof that represents %) For each = (1 ), define () = =1 ( ) , and
notice that
X X
= ∼ (( )̄ + (1 − ( ))) = ()̄ + (1 − ())
=1 =1
that is ∼ ()̄ + (1 − ()). From this it follows that % 0 ⇔ () ≥ (0 ), that is represents %.
Remarks on the EUT Since each decision maker is typically happier the higher the prize he/she receives, it is
natural to suppose that is a strictly increasing function, that is 0 implies () ( 0 ).7
The EUT establishes an implication: If a preference relation % on L satisfies A1-A3, then it can be represented
by an expected utility function. In fact, also the converse implication is true: if % on L can be represented by an
expected utility function, then % satisfies A1-A3.
Normatively, the EUT suggests a simple approach to choice under uncertainty: If a dm believes his preferences
should satisfy A1-A3, then he can use introspection to identify : → R, and then use the expectation of to
choose among lotteries.
Example 2.10 Suppose that = {0 20 70}. First set (0) = 0, (70) = 1, without loss of generality.8 Then, in
order to determine (20), consider the lotteries 20 and = (1 − 0 ). Lottery pays 0 with probability 1 − ,
pays 70 with probability ; ∈ (0 1) is still to be determined. The expected utility of 20 is 1 · (20) = (20).
The expected utility of is (1 − )(0) + (70) = 0 + = . Thus, if the dm determines such that 20 ∼ ,
then the two expected utilities are the same, that is = (20). This identifies (20) for the dm.
Given that : → R has been fully identified, it can be used to rank lotteries. For instance, suppose that
= 25 and that the dm must choose one of the following two lotteries:
0 20 70 0 20 70
= 037 053 01 , 0 = 021 079 0
5 Precisely, () = 127 = 3 () − 5, (0 ) = 115 = 3 (0 ) − 5.
6 Here we consider a specific example, but we can prove that if is not equal to + for some 0, , then there exist 0
such that () = 0
=1 ( ) ≥ ( ) =
0
=1 ( ) but () =
0
=1 ( ) ( ) =
0
=1 ( ) . Therefore the EUT
determines up to linear increasing transformations.
7 Formally, this result can be obtained by adding a fourth axiom that specifies the following: 0 Â if 0 .
8 The invariance of the Bernoulli utility function with respect to linear increasing transformations allows to choose freely () for
two values of . Precisely, if (0) = 0, (70) = 1 do not both hold, then consider such that () = () + , and choose in
such a way that (0) = 0, (100) = 1, that is (0) + = 0, (100) + = 1. This is a linear system in the unknowns , and since
(100) (0) the system has a unique solution, such that 0. Thus the function that is obtained satisfies (0) = 0, (100) = 1
and the expectation of represents the same preferences as the expectation of .
9
2 2
Then () = 0 · 037 + 5 · 053 + 1 · 01 = 0312, (0 ) = 0 · 021 + 5 · 079 + 1 · 0 = 0316. Hence (0 ) ()
and 0 Â .
The EUT is very convenient analytically since it is easy to work with, and probably for this reason it is very
popular. However, sometimes it does not describe well the behavior of real world dm. The most famous piece of
adverse evidence is the so-called Allais’ paradox.
Example 2.11 Consider the following four lotteries:
0 500 000 2 500 000
1 = 500 000 , 2 = 001 089 010
0 500 000 2 500 000 0 500 000 2 500 000
3 = 089 011 0 , 4 = 09 0 010
and then consider (i) the choice between 1 and 2 ; (ii) the choice between 3 and 4 . A few individuals prefer
1 to 2 , and 4 to 3 . However, notice that
0 2 500 000
=
1 = 089500 000 + 011500 000 ; 2 = 089500 000 + 011 with 111 1011
Here A3 implies that the comparison between 1 and 2 is determined by 500 000 compared to because
089500 000 is common to both 1 and 2 . Moreover,
3 = 0890 + 011500 000 ; 4 = 0890 + 011
Here A3 implies that the comparison between 3 and 4 is determined by 500 000 compared to . Hence, 1 Â 2
and 4 Â 3 reveal that % violates A3.9
A few answers have been given to the Allais’ paradox. One is that the paradox is not important, as it involves
amounts of money which are not ordinary for ordinary people and probabilities close to 0 and 1. Another one
is that a dm should be ready to correct mistakes if they violate principles of choice that he considers important.
Other answers suggest to modify the theory by eliminating A3. This has been an active area of research in the last
few decades, although it has had a limited impact on the applied literature.
2.3 The case of = R
In the previous two subsections, the set of the possible prizes was a finite set. Now we consider the case in which
coincides with R. Recall from Subsection 2.1 that a lottery can be seen as a random variable, hence here is brief
reminder of a few basic notions from probability theory.
Cumulative distribution functions (c.d.f.s) A random variable can be described through its c.d.f., which
is a function : R → [0 1] such that
() = Pr{ ≤ } = Pr{ ∈ (−∞ ]} for each ∈ R
Hence, () is the probability that the realized value of the random variable (the prize) is smaller than or equal
to . Each c.d.f. is weakly increasing, continuous from the right, lim→−∞ () = 0, lim→+∞ () = 1.
⎧
⎪
⎨ 0 if 4
1
Example 2.12(i) Consider the random variable with c.d.f. () = ( − 4) if ∈ [4 13] , which has the
⎪
⎩
9
1 if 13
9 In terms of expected utility, 1 Â 2 is equivalent to (500 000) 001(0) + 089(500 000) + 01(2 500 000), that is to
011(500 000) 001(0) + 01(2 500 000). Likewise, 4 Â 3 is equivalent to 09(0) + 01(2 500 000) 089(0) + 011(500 000),
that is to 001(0) + 01(2 500 000) 011(500 000), which contrasts with the previous inequality.
10
following graph:
y 1.0
0.5
-2 2 4 6 8 10 12 14 16
x
This random variable can take on any real value in the interval [4 13], and for instance (2) = 0, (6) = 29 ,
(10) = 23 , (15) = 1.
A c.d.f. can be used also to represent a random variable which can take on a finite set of values.
⎧
⎪
⎪ 0 if −100
⎪
⎪
⎪
⎪
⎨ 02 if ∈ [−100 0)
Example 2.3(v) For the lottery in Example 2.3(i), the c.d.f. is () = 03 if ∈ [0 50) , with graph
⎪
⎪
⎪ 09 if ∈ [50 200)
⎪
⎪
⎪
⎩ 1 if ≥ 200
y 1.0
0.5
-100 100 200
x
Using the c.d.f. of a random variable we can compute the probability not only of the interval (−∞ ] for
each ∈ R, but also of many other sets. For instance, 1 − () = Pr{ } = Pr{ ∈ ( +∞)}, and
(2 ) − (1 ) = Pr{1 ≤ 2 } = Pr{ ∈ (1 2 ]}.
Example 2.12(ii) For the random variable in Example 2.12(i) we have Pr{ 8} = 1 − (8) = 59 , Pr{7 ≤
11} = (11) − (7) = 49 .
It is useful to notice that we can think of a lottery just as a c.d.f., since a c.d.f. reveals all the essential
informations about a lottery. Hence, with a little abuse of notation, sometimes in the following a lottery will be
denoted with .
Densities and expectations In some cases, given a c.d.f. there exists a function : R → R such that () ≥ 0
for each ∈ R and Z 2
(2 ) − (1 ) = () for each 1 2 in R (7)
1
If there exists which satisfies (7), then is called a density function for , and for each 1 2 the equality
R 2
1
() = (2 ) − (1 ) can be used to compute Pr{1 ≤ 2 }. If a density exists for , then (i) is
continuous in R; (ii) at each point 0 ∈ R where is continuous we have that is differentiable and 0 (0 ) = (0 ).
Example 2.3(vi) For the c.d.f. in Example 2.3(v) there exists no density because is not continuous.
Given a continuous , a sufficient condition for the existence of a density for is that is differentiable in R
(except at most in finitely many points) and its derivative is continuous where it exists; then the derivative of is
a density for .
11
Example 2.12(iii) For the c.d.f. in Example 2.12(i) there exists a density function, which is10
⎧
⎪
⎨ 0 if 4
1
() = if ∈ [4 13]
⎪
⎩
9
0 if 13
Given a function : R → R, with the notation
Z Z
() () or () () (8)
R
we represent the expectation of the function according to the c.d.f. . For a lottery which takes on finitely many
values, the expectation is computed as in the previous subsection.
Example 2.3(vii) Consider the lottery in Example 2.3(i) and () = 15 + 3. The expectation of according
to this lottery is (−100) · 02 + (0) · 01 + (50) · 06 + (200) · 01 = (−17) · 02 + 3 · 01 + 13 · 06 + 43 · 01 = 9.
R R
If has a density , then R () () is given by R () ().
Example 2.12(iv) Consider the lottery in Example 2.12(i) and () = 15 + 3. The expectation of according to
R 13
this lottery is 4 ( 15 + 3) 19 = 47
10 .
R
In the special case with () = , (8) reduces to R (), which is called expected value of the lottery and
is denoted with in the following.
R 13
Example 2.12(v) For the lottery in Example 2.12(i), is equal to 4 19 = 85.
Example 2.3(viii) For the lottery in Example 2.3(i), = (−100) · 02 + 0 · 01 + 50 · 06 + 200 · 01 = 30.
The Expected Utility Theorem when the set of prizes is R In this environment, the EUT holds under
suitable assumptions, such that the preference relation on lotteries satisfies A1, A2, A3, plus a topological condition
— of little economic relevance — which we neglect here. Then it is possible to prove that there exists a continuous
Bernoulli utility function : R → R (here R replaces the finite set in Proposition 2.9) such that given any two
lotteries and , Z Z
% ⇔ () () ≥
()() (9)
R R
R
Therefore % is represented by the expectation of , that is ( ) = R () () represents %, as in Proposition
2.9.
2.4 Stochastic dominance
In the comparison between two lotteries, it is often the case that the dm’s Bernoulli utility function determines
which lottery is superior.
0 20000 √
Example 2.13(i) Consider the two lotteries = 04 06 and 10000 . If () = , then 10000 Â because
√ √
10000 04 · 0 + 06 · 20000 = 8485, but if () = then  10000 because 04 · 0 + 06 · 20000 10000.
Here we show that in some cases two lotteries are ranked in the same way for any which is strictly increasing.
In order to simplifiy the arguments, we restrict the set of lotteries we consider by assuming that there exists two
numbers and ̄ such that ̄ and each lottery satisfies () = 0, (̄) = 1; that is the set of prizes
is a subset of ( ̄]. This means that the random variables we are considering cannot take on values outside the
interval ( ̄], but notice that ̄ can be chosen arbitrarily. The key definition is as follows:11
10 Notice that is differentiable in R, except at the points 4 and 13, and the derivative of is continuous where it exists, that is in
(−∞ 4) ∪ (4 13) ∪ (13 +∞). Hence, a density for exists and is the derivative of .
11 In fact, the definition below is often referred to as ”first order stochastic dominance”, as stochastic dominance can be defined in
terms different from (10) below. But since in this document no other stochastic ranking is considered, we simplify the terminology by
omitting the words ”first order”.
12
Definition 2.14 Given two lotteries , we say that stochastically dominates if
() ≤ () for each ∈ [ ̄] (10)
This property means that for each possible , the probability of a prize (weakly) lower than is (weakly) lower
with than with , or equivalently the probability of a prize higher than is (weakly) higher with than with
, and this is true for each as (10) implies (is equivalent to) 1 − () ≥ 1 − () for each ∈ [ ̄].
Example 2.15 We represent below three pairs of c.d.f.s. In each one, the graph of is a thin curve, the graph
of is a thick curve. In the first two cases from the left, stochastically dominates . In the third case, neither
c.d.f. stochastically dominates the other.
y 1.0 y 1.0 y 1.0
0.8 0.8 0.8
0.6 0.6 0.6
0.4 0.4 0.4
0.2 0.2 0.2
0.0 0.0 0.0
0 2 4 6 0 5 10 15 0 5 10 15
x x x
If (10) holds, then for each the c.d.f. gives a weakly higher probability than to prizes higher than .
This suggests that is a better lottery than for each dm, as indeed next proposition establishes.
Proposition 2.16 Condition (10) holds if and only if
Z ̄ Z ̄
() () ≥ ()() for each strictly increasing (11)
This proposition establishes a double implication, but the more interesting one is that (10) implies (11): this
means that if (10) is satisfied, then each dm (weakly) prefers to provided that his is strictly increasing.
Hence, the comparison between and is unambiguous and very robust as it does not depend on .12
Proof We first prove that (10) implies (11). In order to simplify the proof, we assume that there exists a density
for , a density for , and that is continuously differentiable. Then integration by parts reveals that
R ̄ ̄ R ̄ R ̄ R ̄ ̄
() () = [() ()] − 0 () () = (̄) − 0 () () and ()() = [()()] −
R ̄ 0 R ̄ R ̄ R ̄ R ̄
()() = (̄) − 0 ()(). Therefore () () ≥ ()() reduces to 0 ()() ≥
R ̄ 0 R ̄
() (), or 0 ()[() − ()] ≥ 0. This inequality is satisfied since 0 () ≥ 0 for each as is
strictly increasing, and () − () ≥ 0 for each because of (10).13
Now we suppose that (11) holds, but there exists ̂ ∈ ( ̄) such that (̂) (̂). Then consider
(
0 if ̂
() = (12)
1 if ̂
12 Notice that (10) implies ≥ . This can be seen by choosing () = in (10). But ≥ does not imply (10).
13 Notice that if there exists 0 ∈ ( ̄) such that (0 ) (0 ), then there exists an interval (1 2 ) ⊆ ( ̄) that includes 0 and
is such that () () for each ∈ (1 2 ). For each strictly increasing , there exists an interval included in (1 2 ) in which
0 () 0, thus (11) holds with strict inequality.
13
R ̄ R ̄ R ̄ R ̄
Given , we have () () = 1 − (̂) and ()() = 1 − (̂), hence () () ()(),
which violates (11), a contradiction.14 ¥
Notice that given two arbitrary c.d.f.s , it is not necessarily the case that (10), or the opposite of it, holds.
That is, there are (many) pairs of c.d.f.s such that neither c.d.f. stochastically dominates the other: for instance,
see the third pair of c.d.f.s in Example 2.15.15 In this case, the best lottery between and is determined by .
2.5 Risk attitudes
It is often the case that individuals display an aversion to risk in the sense that, ceteris paribus, they prefer lotteries
with little variation in prizes to lotteries with large variation.
2.5.1 Risk neutrality, Risk aversion
Risk attitudes are defined on the basis of the dm’s preference between a lottery and the lottery , the
degenerate lottery which yields for sure the expected value of .
Definition 2.17 A decision maker is said to be risk neutral if ∼ holds for each lottery .
Consider a dm having the choice between playing the lottery on one hand, and receiving with certainty the
expected value of , , on the other hand. A risk neutral dm is indifferent between these two alternatives, that
is for him playing is exactly as good as the lottery which gives him for sure.
Example 2.13(ii) Consider the lottery of Example 2.13(i), with = 12 000. A risk neutral dm is indifferent
between and 12 000 .
As a consequence of Definition 2.17, given two lotteries , a risk neutral dm chooses the lottery with the
highest expected value because ∼ , ∼ , hence % is equivalent to % , that is to
≥ .
The arguments above are expressed without reference to the EUT, but it is possible to express risk neutrality
R
in terms of expected utility. Precisely, risk neutrality is equivalent to () = because then () () = ,
R R R
()() = , hence () () ≥ ()() boils down to ≥ .16
However, many dm are not risk neutral, that is do not choose the lottery with the highest expected value. For
instance, between of Example 2.13(i) and 12 000 , many dm strictly prefer 12 000 .
Definition 2.18 A decision maker is said to be
• risk averse if % for any lottery ;
• strictly risk averse if  for each non-degenerate .
A strictly risk averse dm is such that when he faces a non-degenerate lottery (i.e., a lottery which involves
some risk) he prefers to receive for sure rather than playing . For instance, given from Example 2.13(i),
he prefers 12 000 to . For a risk averse dm, the preference holds weakly: risk aversion (without the adjective
”strict”) includes risk neutrality as a special case.
In the terms of expected utility, strict risk aversion is equivalent to
Z
( ) () () for each non-degenerate (13)
14 In fact, the Bernoulli utility function in (12) is not strictly increasing, but it can be approximated arbitrarily closely by a strictly
increasing in such a way that (11) is violated.
15 Hence, stochastic dominance induces an order between c.d.f.s which violates the property of completeness introduced in Section 1.
16 Of course, also () = + (with 0) is a Bernoulli utility function for a risk neutral dm, because of the invariance with
respect to linear increasing transformations.
14
and from mathematical analysis we know that (13) holds if and only if is a strictly concave function. Precisely, a
function : R → R is said to be strictly concave if for each 1 2 in R with 1 2 , the following inequality holds
for each ∈ (0 1):
(1 + (1 − )2 ) (1 ) + (1 − )(2 ) (14)
Graphically, this means that for each two points (1 (1 )), (2 (2 )) on the graph of , the segment connecting
them lies uniformly below the graph (except at the extremes of the segment).
100 500
Example 2.19(i) Consider = 05 05 , with = 300. In the Cartesian plane below, the graph of is the
continuous curve; hence is strictly concave. We represent (100), (500), (300), and 12 (100) + 12 (500), that is
R R
() (), which is at the level of the dot with label ”exp ut”. It is apparent that (300) () (), which
is indeed a consequence of being strictly concave:
u(500)
u(300)
exp ut
u(100)
100 300 500 x
In fact, it is simple to go beyond the above example and observe that the equivalence between strict risk
aversion of the dm and the strict concavity of is immediate for lotteries that can take on just two values. Indeed,
1 2
consider a lottery = 1− that pays the prize 1 with probability , the prize 2 with probability 1 − . Since
= 1 + (1 − )2 , it follows that (13) boils down to (14). Therefore strict risk aversion is satisfied if and only if
is strictly concave for lotteries with just two possible prizes. But this equivalence can be proved for more general
lotteries.
Given that strict risk aversion is equivalent to strictly concave, we can recognize if a dm is strictly risk averse
or not from the curvature of his . We do not prove this fact, but in next example we consider a strictly concave
R
and suggest that, for a specific lottery , strict concavity implies ( ) () ().
In order to recognize whether a certain is strictly concave, it is useful to know that if is differentiable, then
strictly concave is equivalent to 0 strictly decreasing. If is twice differentiable, then 00 () 0 for each implies
that is strictly concave.
√
Example 2.13(iii) Consider lottery of Example 2.13(i) and suppose that () = ; this can be used for
lotteries with non-negative prizes. Since 00 () = − 14 −32 0 for each 0, it follows that is strictly concave and
R √
the dm with this Bernoulli utility function is strictly risk averse. Indeed () () = 04 · 0 + 06 20000 = 8485
√
is smaller than ( ) = 12000 = 10954.
We now introduce a notion which provides a different perspective on risk attitudes.
Definition 2.20 Given a decision maker with Bernoulli utility function , the certainty equivalent of a lottery
for this decision maker is the amount of money ( ) such that the decision maker is indifferent between ()
and .
15
This notion tries to summarize a lottery using a single prize: ( ) is the amount of money such that the dm
is indifferent between the following two alternatives: receiving ( ) for sure, or playing . In terms of expected
utility, ( ) satisfies Z
(( )) = () ()
√ R √
Example 2.13(iv) Consider lottery of Example 2.13(i) and () = . Since () () = 06 20000, it
√ √
follows that ( ) solves = 06 20000, hence ( ) = 7200.
If the dm is risk neutral, then ( ) = since is worth to the dm exactly : see Definition 2.17. But if
the dm is strictly risk averse, then  ∼ () , hence ( ).17 In words,  implies that
is worth less to the dm than the riskless amount of money , hence ( ) . For instance, for and in
the Example 2.13(iv), ( ) = 7200 is smaller than = 12000.
Example 2.19(ii) Consider the lottery of Example 2.19(i), with = 300 and strictly concave. In the
R
Cartesian plane below, we represent (100), (500), and 12 (100) + 12 (500), that is () (), which is at the
R
level of the dot with label ”exp ut”. Since ( ) is the prize (on the horizontal axis) such that () = () (),
it follows that ( ) 300 = .
u(500)
exp ut
u(100)
100 c(F,u) 300 500 x
The next proposition summarizes the two characterizations we have given of strict risk aversion.
Proposition 2.21 A dm with a Bernoulli utility function is strictly risk averse if and only if is strictly concave,
if and only if ( ) for each non-degenerate .
The difference − ( ) is said risk premium of the lottery and can be interpreted as the reduction in value
of below due to the riskiness of . The stronger is the dm’s risk aversion, the more the risk associated to
is harmful to the dm, and the lower is ( ), that is the greater is the value reduction.
√
Example 2.13(v) Consider lottery in Example 2.13(i) and a dm with () = . Then − ( ) =
12000 − 7200 = 4800.
2.5.2 Measuring risk aversion
After introducing the notion of risk aversion, here we describe a way to measure how strongly a dm is risk averse.
00
Definition 2.22 Given (twice differentiable) the coefficient of risk aversion of at is ( ) = − 0 ()
()
.
In order to make sense of this, recall that 00 () 0 for each implies strict risk aversion, hence it may be
intuitive that risk aversion is stronger the more negative is 00 (), that is the greater is −00 (). But also recall
17 It is immediate to see that also the reverse implication holds: ( ) implies  () , hence  .
16
that the function such that () = () with 0 is equivalent to , and − 00 () = −00 () is greater the
greater is . Thus we cannot use the negative of the second derivative of the Bernoulli utility function to measure
the intensity of risk aversion because that can be increased or reduced even though the dm’s preferences do not
change.18 The definition of ( ) relies on −00 (), but avoids the problem of scale by dividing by 0 (); is thus
independent of linear increasing transformations because of 0 () at the denominator: ( ) = ( ) for each
0.19
The interpretation of is that the greater is , the more the dm is risk averse. Sometimes can be used to
compare two Bernoulli utility functions 1 2 : 2 is said to be more risk averse than 1 if ( 2 ) ≥ ( 1 ) for
each .
Example 2.23 Consider 1 () = −− , 2 () = −− with 0. Then ( 1 ) = , ( 2 ) = for each
, hence a dm with Bernoulli utility function 2 is more risk averse than a dm with Bernoulli utility function 1 .
The Bernoulli utility function () = −− with 0 is popular in applied work, and the parameter 0
measures the intensity of the dm’s risk aversion: the higher is , the more he is risk averse . If is close to 0, then
the dm is approximately risk neutral and ( ) is just sligthly smaller than . If is very large, then he is very
risk averse and ( ) is close to the lowest possible prize given ; in this case the best lottery is the one with the
highest lowest prize.
√ 1
Example 2.24 For lotteries with nonnegative prizes, consider 3 () = . Then ( 3 ) = 2 , and comparing
1 1
( 3 ) with ( 1 ) we find that ( 3 ) ( 1 ) if ∈ (0 2 ), ( 3 ) ( 1 ) if 2 . Thus we cannot
say that 3 is more risk averse than 1 , nor viceversa. But 3 is more risk averse than 1 if we consider only
1
lotteries with prizes in the interval (0 2 ); 1 is more risk averse than 3 if we consider only lotteries with prizes
1
greater than 2 .
Another natural way to compare two Bernoulli utility functions 1 2 is based on ( 1 ) ( 2 ): we may
say that 2 is more risk averse than 1 if ( 2 ) ≤ ( 1 ) for each . This means that for each given lottery
, risk aversion destroys (weakly) more of the value of the lottery given 2 than given 1 . It turns out that the
property ( 2 ) ≥ ( 1 ) for each is equivalent to the property ( 2 ) ≤ ( 1 ) for each . Therefore, the
comparison based on and the comparison based on certainty equivalents yield the same result.
3 Game Theory
A game is a decision problem with ≥ 2 agents/players in a setting of strategic interdependence. This means that
the well-being (utility, profit, ...) of each player depends not only on his action, but also on the actions of the other
players, and often this implies that the most convenient action for him depends on the actions of the other players.
That is, one player can identify his best action only if he knows the actions of the other players. Game theory
studies the behavior of players in settings with strategic interdependence.
Within economics (and social sciences, more generally), many settings are characterized by strategic interde-
pendence and can be analyzed using the tools of game theory. A typical example is a duopoly, that is an industry
in which two firms operate. Suppose each firm chooses the price at which it offers the own product, and the sales
of firm 1 depend both on the price 1 firm 1 chooses, and on the price 2 chosen by firm 2. Then, determining
the most convenient price for firm 1 requires that 1 knows 2 . Likewise, firm 2 needs to know 1 in order to figure
out its most convenient price. Thus each firm needs to have expectations about how the other firm plays, and this
circularity makes it difficult to formulate a prediction on firms’ prices.20 Notice that this context is significantly
18 Recall that 0 is irrelevant to represent % as () () represents % as () () does.
19 In fact, ( ) is often called coefficient of absolute risk aversion, in order to distinguish it from the coefficient of relative risk
00 ()
aversion, which is defined as − 0 () . Here we do not consider relative risk aversion, hence we do not need the qualifier.
20 More generally, a branch of economics called industrial organization uses game theory to study many different forms of competition
among firms, in which the firms are the players, each choosing for instance the price to charge or the quantity to offer, and the action
17
different from the one in which a single firm (a monopolist) is active in the market, as that firm can determine its
optimal price with no need to figure out the behavior of its rival(s), as no rival firm exists.
3.1 Games with simultaneous moves
We start with a class of games called games with simultaneous moves. In each game of this class, each player moves
only once and all players move at the same time. Let = {1 2 } denote the set of players, with ≥ 2. For
each player , denotes the set of his possible moves, which are called strategies; is called strategy set of player
, and is a generic strategy of player . With = (1 ) we denote a profile of strategies, that is specifies
one strategy for each player; = 1 × 2 × × is the set of all profiles of strategies.
In each game with simultaneous moves, each player chooses a strategy from , and the resulting strategy
profile determines the outcome of the game, but for simplicity we may think that itself is the outcome. We
assume that each player has a preference relation over the possible outcomes, that is over the elements of , and
the preference relation of player , % , is described by a utility function : → R which associates to each ∈
a number () such that % 0 if and only if () ≥ (0 ).
A game with simultaneous moves is denoted with and consists of the following elements: (set of players),
1 (players’ strategy sets), 1 (players’ utility functions):
= { 1 1 }
We use − to denote a profile of strategies for the players different from , that is − = (1 −1 +1 ).
Then can be written as ( − ) in order to emphasize the strategy played by player and the strategies played
by the other players. We use − = ×6= to denote the set of strategy profiles for the players different from .
Example 3.1(i) Two persons, for instance a boyfriend and a girlfriend, cannot communicate but must simultane-
ously decide where they will spend the evening; the possible choices for each player are a disco and a movie theater.
Hence, each player chooses either or , that is
1 = { }, 2 = { } (15)
Therefore = {( ) ( ) ( ) ( )} (there are four outcomes) and 1 associates a number to each
element of , such that
1 ( ) = 5 1 ( ) = 2 1 ( ) = 1 1 ( ) = 0 (16)
From (16) we infer that the preferred outcome for player 1 is the one such that both players go to the movie theater;
the second preferred outcome for 1 is such that both go to the disco ... About player 2, suppose that (the first
argument of 2 is 2 , the second argument is 1 )
2 ( ) = 2 2 ( ) = 5 2 ( ) = 1 2 ( ) = 0 (17)
From (17) we infer 2’s preferences over . This game is called Battle of Sexes and consists of the following elements:
= {1 2}, 1 2 in (15), 1 2 in (16)-(17)
Example 3.2(i) Consider an industry in which there are two active firms. Each firm simultaneous chooses the
price at which it offers the own product. Let ≥ 0 denote the price chosen by firm , for = 1 2, and a regulation
imposes ≤ 140. Given 1 2 , the quantity sold by firm 1 is 300 − 21 + 2 ; the quantity sold by firm 2 is
300 − 22 + 1 . Each firm has a constant marginal production cost, which is equal to 35 for firm 1, is equal to 10
for firm 2. Therefore
of each firm affects both the own profit and the profits of the other firms.
18
= {1 2} 1 = [0 140] 2 = [0 140]
and for each firm the utility function is the profit function:
1 (1 2 ) = (300 − 21 + 2 )(1 − 35) 2 (2 1 ) = (300 − 22 + 1 )(2 − 10)
When = 2 and the sets 1 2 have each a small number of elements, we can represent though a matrix
in which (i) each row (except the top row) represents a strategy of player 1; (ii) each column (except the leftmost
column) represents a strategy of player 2; (iii) given = (1 2 ), the matrix entry in the row representing 1 and
in the column representing 2 indicates the utilities of 1 and 2 when is played. This representation is possible for
the Battle of Sexes, but not for the duopoly game of Example 3.2(i).
Example 3.1(ii) The game of Example 3.1(i) can be represented by the following matrix
1\2
: 2 5 0 0
1 1 5 2
A fundamental assumption of game theory that
• (i) each player knows and wants to maximize the own utility;21
• (ii) each player knows that (i) is true;
• (iii) each player knows that (ii) is true;
and so on, ad infinitum. This infinite series of claims is typically expressed by saying that (i) is common knowledge
among players, that is it is common knowledge that each player knows and wants to maximize the own utility.
3.1.1 Dominant strategies
Strict domination The idea of strict domination is that sometimes a player’s strategy is superior to another
strategy (or to several other strategies) for each profile of strategies the other players may play, as in next example.
Example 3.3
1\2
9 3 5 4 6 2
1 :
6 5 1 7 3 5
0 4 2 5 2 3
In
1 we have 1 = { }, 2 = { }. For player 1, the best strategy is for each 2 ∈ 2 , that is 1
does not need to figure out how player 2 plays in order to realize that his best choice is (but 2’s strategy still
affects player 1’s utility).
A player’s best strategy typically depends on what he expects the other players are playing. But this is not
always the case: In some games a player can identify his best strategy even though he has no idea on how the other
players play. This occurs in
1 , for player 1 and also for player 2, since strategy is 2’s best strategy for each
1 ∈ 1 .
Definition 3.4 Given , we say that a strategy ∗ ∈ strictly dominates ̄ ∈ for player if
(∗ − ) (̄ − ) for each − ∈ −
21 Sometimes, the fact that a player wants to maximize the own utility is expressed by saying that the player is ”rational”.
19
This means that for player it is more convenient to play ∗ than ̄ , for each − ∈ − , that is regardless of what
the other players play. Therefore, player will certainly not play ̄ . Moreover, if ∗ strictly dominates for each
6= ∗ in (as does in ∗
1 for player 1) then we say that is a strictly dominant strategy for player , and
he will certainly play ∗ because ∗ is more profitable for him than any other strategy, regardless of the strategies
played by the other players. In this case player solves his decision problem even though he does not know what
the other players play. If for each player there exists a strictly dominant strategy, then we conclude that the profile
of strictly dominant strategies will be played; we call this profile an equilibrium with strictly dominant strategies
and we say that is dominance solvable. In particular, 1 is dominance solvable and ( ) is an equilibrium of
1 with strictly dominant strategies.
Example 3.5 The following game is called Prisoner’s Dilemma:
1\2
2 : −2 −2 −10 −1
−1 −10 −5 −5
Two individuals are arrested and charged of a serious crime. The district attorney has sufficient evidence to get
them sentenced for a minor crime, but he would like them to confess the more serious crime about which he has no
sufficient evidence to get a conviction. He talks to each individual and explains that if a single prisoner confesses,
then he will be sentenced to one year in jail, and the other to ten years; if both confess, each is going to jail for
five years; if no prisoner confesses, then each will stay in jail for two years (the numbers in the matrix represent
the negative of the number of years spent in jail by each player). For each player, strictly dominates , hence
( ) is an equilibrium of 2 with strictly dominant strategies.
It is important to notice that the equilibrium outcome in 2 is worse for the two players with respect to the
outcome obtained with ( ). Therefore, the choices of each player based on the maximization of the own
utility may generate a bad outcome from a collective point of view (that is, from the point of view of the two
players), as both would be better off with ( ). The reason is that when a player plays , he increases
his utility with respect to the utility he obtains by playing , but reduces the utility of the other player (here
strategic interdependence is at work). Hence, when both play each player benefits from the own move, but is
harmed by the move of the other player. Since the latter has a bigger impact, this makes each player worse off with
respect to the case in which both players play . If, before 2 is played, the players could commit to play ,
then both would earn a higher utility than from the equilibrium of 2 with strictly dominant strategies, but if
binding agreements are infeasible then each player has incentive to violate a non-binding agreement. To summarize,
for the players there are joint incentives not to confess, but each player has an individual incentive to confess.
Although this course is not about individuals charged with crimes, this game is relevant because it is a sort
of paradigm. In several situations players have incentives similar to those in 2 : there exist joint gains from
cooperating (playing in 2 ), but each player has an individual incentive not to cooperate (playing in
2 ).
Iterated elimination In fact, it is infrequent that for a player there exists a strategy that strictly dominates all
his other strategies, hence it is rare that an equilibrium with strictly dominant strategies exists. It is a bit more
frequent that for a player there exists one or more strictly dominated strategies, and if so this is relevant as he
should not play those strategies.
Example 3.6
1\2
1 0 3 2 1 3
3 :
2 3 1 1 3 2
0 3 2 3 0 1
20
In 3 , no strategy of player 1 strictly dominates each other strategy in 1 , but is strictly dominated by ; thus
1 will not play . If player 2 can reproduce the reasoning of player 1, then he learns that 1 will not play and this
simplifies the game from 2’s viewpoint: he can argue as if the strategy set for 1 was { }. In this simplified game,
is strictly dominated by for 2 (although is not so in the original game). If 1 can reproduce the reasoning
of 2, then he learns that 2 will not play . In the simplified game in which the strategy set for 2 is { }, is
strictly dominated by for 1 (although is not so in 3 ). If player 2 realizes that 1 plays , then is strictly
dominated by for 2. Hence, ( ) is the unique strategy profile that survives this deletion procedure, called
iterated deletion of strictly dominated strategies.
Notice that in 3 , does not strictly dominate each other strategy of player 1, and does not strictly
dominate each other strategy of player 2: actually, there is a single strictly dominated strategy in 3 , which is .
However, eliminating one strategy makes it possible that additional strategies become strictly dominated because
the fewer strategies a player’s opponents may play, the more likely is that a particular strategy of his is strictly
dominated.
In 3 , iterated deletion kills each strategy of player 1 except one, and each strategy of 2 except one. But this
is not always the case, and sometimes several strategies survive for each player. For instance, if 1 ( ) = 4 in
3 then two strategies for each player survive. For each player, the set of strategies which survive is called set
of non-iteratively strictly dominated strategies, and we expect that each player chooses a strategy from this set.
When it consists of a single strategy for each player, is said to be solvable by iterated elimination of strictly
dominated strategies. The set of profiles of strategies which survive iterated elimination is independent of the order
of deletion.
This procedure requires some assumptions about players’ informations. Precisely,
• in order to eliminate , player 1 needs to know his own utility function 1 ;
• in order to eliminate , player 2 needs to know (i) 2 1 ; (ii) that 1 knows 1 and 1 wants to maximize 1
(hence, 2 deduces that 1 will not play );
• in order to eliminate , player 1 needs to know (i) 1 2 ; (ii) that 2 knows 2 and that 2 wants to maximize
2 ; (iii) that 2 knows 1 and that 1 knows 1 and 1 wants to maximize 1 (hence, 1 infers that 2 deduces
that 1 will not play , and 2 will then choose not to play ).
The successive step requires even deeper information. If it is common knowledge that each player knows and
wants to maximize the own utility, then the procedure can be iterated for any number of steps.
Weak domination There exists another notion of domination which is useful to know.
Definition 3.7 Given , we say that a strategy ∗ ∈ weakly dominates ̄ ∈ for player if
(∗ − ) ≥ (̄ − ) for each − ∈ − with strict inequality for at least one −
This definition is somewhat similar to Definition 3.4 about strict domination, but does not require strict in-
equalities: weak inequalities suffice, but strict inequality needs to hold for at least one − . This means that ∗
does at least as well as ̄ for each − (i.e., regardless of what the other players play), and does strictly better than
̄ for at least one − .
Example 3.8(i) In the following game, both and are weakly dominated by for player 2:
1\2
4 : 1 5 0 6 4 6
0 4 1 3 4 4
21
Notice that if ∗ strictly dominates ̄ , then ∗ also weakly dominates ̄ , but the viceversa is untrue. Is it
legitimate to conclude that player will not play a weakly dominated strategy ̄ ? If ∗ does not strictly dominate
̄ , then there exists (at least one) − such that ∗ yields player the same utility as ̄ when the other players
play − . Thus, concluding that player will not play a weakly dominated strategy is less justified with respect to
a strictly dominated strategy. On the other hand, in no case ̄ yields to player a higher utility than ∗ , thus
may want not to play ̄ in order to avoid a risk that is costless to eliminate. As a consequence, there is not an
unanimous point of view on this issue in the literature on game theory.
It should be obvious that the iterated deletion of weakly dominated strategies is more powerful than iterated
deletion of strictly dominated strategies, but this procedure is even more controversial than the non-iterated elim-
ination of weakly dominated strategies. Moreover, in some games the surviving strategies depend on the order of
deletion.
Example 3.8(ii) In 4 , deleting first , then , then , shows that only ( ) survives iterated elimination
of weakly dominated strategies when we start from . But deleting first , then , then , shows only ( )
survives iterated elimination of weakly dominated strategies when we start from .
3.1.2 Nash Equilibrium
In many games the set of strategies which survive iterated elimination of strictly/weakly dominated strategies
provides little or no insight about the players’ behavior; in fact, sometimes iterated elimination does not eliminate
anything. Here we introduce a notion of solution of a game which is more widely applicable.
Definition 3.9 Given , we say that ∗ ∈ is a Nash Equilibrium (NE) of if for each player we have
(∗ ∗− ) ≥ ( ∗− ) for each ∈ (18)
Consider a game with two players to fix the ideas. A NE ∗ is a pair of strategies (∗1 ∗2 ), one for each player,
with the following property: 1 (∗1 ∗2 ) ≥ 1 (1 ∗2 ) for each 1 ∈ 1 , 2 (∗2 ∗1 ) ≥ 2 (2 ∗1 ) for each 2 ∈ 2 From
the point of view of player 1, given the strategy ∗2 of player 2, ∗1 must be a strategy of 1 that maximizes 1 . In
other words, given that 2 plays ∗2 , ∗1 must be a best strategy that player 1 can play: we say that ∗1 is a best reply
against ∗2 for player 1. Likewise, given that player 1 plays ∗1 , for player 2 the best strategy must be ∗2 , that is ∗2
is a best reply to ∗1 for 2.
Example 3.10 Consider the following game:
1\2
1
: 1 3 5 2 2 6
3 1 4 7 1 0
In this game, ( ) is not a NE because is not a best reply for player 1 to played by player 2. Likewise,
( ) is not a NE because is a best reply for player 1 to played by player 2, but is not a best reply for
player 2 to played by player 1. There exists a unique NE in 1 , which is ( ). Indeed, given that player 2
plays , for player 1 the strategy is a best reply; given that 1 plays , for 2 the strategy is a best reply. For
each other pair of strategies, at least one player wants to deviate. If 1 ( ) = 2, then ( ) would still be a NE.
Recall that a player needs to have expectations about how other players play, in order to identify his best
strategy (unless there exists a strictly dominant strategy for him). Then we can interpret a NE as follows: a NE
∗ = (∗1 ∗2 ) provides to each player a conjecture on how the other player plays. To player 1, it indicates to expect
that player 2 plays ∗2 ; to player 2, it indicates to expect that 1 plays ∗1 , and these conjectures mutually justify. If
1 expects that player 2 plays ∗2 , then he wants to play ∗1 , a best reply for him to ∗2 ; this justifies that player 2
expects that 1 plays ∗1 and gives an incentive to 2 to play ∗2 , which justifies that 1 expects ∗2 from 2, The word
22
”equilibrium” is meant to convey the idea that there is no tendency to move away from it because no player can
gain by deviating; this is indeed the property characterizing each NE.
Recall that a player needs to have expectations about how other players play, in order to identify his best
strategy (unless there exists a strictly dominant strategy for him). Then we can interpret a NE as follows: a NE
∗ = (∗1 ∗2 ) provides to each player a conjecture on how the other player plays. To player 1, it indicates to expect
that player 2 plays ∗2 ; to player 2, it indicates to expect that 1 plays ∗1 , and these conjectures mutually justify. If
1 expects that player 2 plays ∗2 , then he wants to play ∗1 , a best reply for him to ∗2 ; this justifies that player 2
expects that 1 plays ∗1 and gives an incentive to 2 to play ∗2 , which justifies that 1 expects ∗2 from 2, The word
”equilibrium” is meant to convey the idea that there is no tendency to move away from it because no player can
gain by deviating; this is indeed the property characterizing each NE.
Example 3.1(iii) The game has two NE: ( ) and ( ). For the first one, if player 1 expects that
player 2 plays , then it is a best reply for him to play , and if 2 expects that 1 plays , then it is a best reply for
him to play . A similar argument applies to ( ). Notice that within the set of NE, player 1 prefers ( ),
2 prefers ( ).
Example 3.1(iv) Consider the Battle of Sexes with the same preferences for the two players, such that both
players prefer to :
1\2
2 2 0 1
1 0 5 5
This game has two NE: ( ) and ( ), even if players have the same preferences. Both players prefer the
NE ( ) to ( ), but ( ) is nevertheless a NE. If a player (1 or 2) expects the other player to play ,
then it is a best reply for him to play .
It is important to notice that if ∗ is a NE, then not necessarily ∗1 is a best reply for 1 against 2 6= ∗2 ; Definition
3.9 only requires that ∗1 is a best reply for 1 against ∗2 . Likewise, ∗2 is best reply for player 2 against ∗1 , but it
may not be so against 1 6= ∗1 . In 1 , ( ) is a NE but is not a best reply for player 2 if 1 plays .
Example 3.2(ii) In the duopoly of Example 3.2(i), we can find the best reply of each firm given the strategy of
the other firm. The profit of firm 2 is maximized with respect to 2 at 2 = 14 1 + 80. Then consider firm 1, and
prove that its profit is maximized with respect to 1 at 1 = 14 2 + 925 (see the Cartesian plane below). A pair
(∗1 ∗2 ) is a NE if and only if ∗1 is a best reply for firm 1 to ∗2 , hence ∗1 = 14 ∗2 + 925; ∗2 is a best reply for firm 2
to ∗1 , hence ∗2 = 14 ∗1 + 80. Therefore, ∗1 ∗2 solves the system
1 1
2 = 1 + 80, 1 = 2 + 925
4 4
23
The unique solution is ∗1 = 120, ∗2 = 110, which is the unique NE of the duopoly game.
140
s2
120
s2=(1/4)s1+80
100
80
60
40 s1=(1/4)s2+92.5
20
0
0 20 40 60 80 100 120 140
s1
When there are more than two players, the same idea applies: A NE is a profile of strategies ∗ such that for
each player , strategy ∗ is a best reply for him to the strategies ∗− of the other players.
Example 3.11 In the three-player game below, player 1 chooses a row ( or ), player 2 chooses a column ( or
), player 3 chooses a matrix ( or ):
3: 3:
1\2 1\2
8 2 8 5 4 0 5 3 6 3 4 2
4 3 7 6 1 5 4 3 4 1 2 0
In this game, ( ) is not a NE because is not a best reply for player 3 to (1 2 ) = ( ). Likewise,
( ) is not a NE because is not a best reply for player 1 to (2 3 ) = ( ). The unique NE in this game
is ( ).
Relationship between NE and dominance Each equilibrium with strictly dominant strategies is a NE, and
indeed ( ) is a NE in the game of Example 3.3, ( ) is a NE in the game of Example 3.5. More generally,
each NE survives iterated deletion of strictly dominated strategies (this is untrue for iterated elimination of weakly
dominated strategies), hence if only one strategy profile survives iterated deletion of strictly dominated strategies,
then it is a NE — actually, it is the unique NE. In each NE, no strictly dominated strategy is played but weakly
dominated strategies may be played.
Example 3.12 In the following game, ( ) is a NE even though is weakly dominated by for player 2:
1\2
5 : 5 4 1 4 0 3
2 0 3 1 1 2
3.1.3 Mixed strategies
Each of the games considered up to now has at least one NE, but there exist games which have no NE.
24
Example 3.13(i) Consider the following game:
1\2
1 : 0 3 4 1
3 2 2 4
In 1 , no is a NE because if 2 = (for instance), then the best reply of player 1 is and then it is profitable
for player 2 to play rather than .
It is unsatisfactory to have a theory which tries to predict players’ behavior but in some cases does not give
any indication. One way to avoid this inconvenience consists in extending the choice set of each player by allowing
each player to select a strategy in a random way, using a so-called mixed strategy. In this subsection we focus on
the case in which is a finite set for each player .
Definition 3.14 Given , a mixed strategy for player is a probability distribution over .
A generic mixed strategy of player is denoted , and it consists of a list of probabilities, one for each strategy
P
in , such that ( ) denotes the probability that plays ; ( ) ≥ 0 and ∈ ( ) = 1. From now on
the strategies in will be called pure strategies, in order to distinguish them from mixed strategies.
Example 3.15(i) Consider the following game:
1\2
2 : 6 2 1 6 5 0
4 7 3 4 2 10
In 2 we have 1 = { } 2 = { }. For instance, a mixed strategy for player 1 is 1 such that
1 () = 16 1 () = 56 . For the sake of brevity, sometimes this mixed strategy is represented as 16 + 56 .
9 4 9 4
Another mixed strategy for 1 is 1 such that 1 () = 13 , 1 () = 13 , or 13 + 13 . For player 2, two examples
of mixed strategies are 2 such that 2 () = 6 2 () = 2 2 () = 3 , or 6 + 12 + 13 , and 2 = 47 + 37 .
1 1 1 1
A mixed strategy is the way through which a player randomizes his strategy choice. We let denote the set
P
of mixed strategies of player , hence = { ∈ R| | : ( ) ≥ 0 for each ∈ , ∈ ( ) = 1}. Notice
that includes mixed strategies which are equivalent to pure strategies. Precisely, given a fixed 0 ∈ , consider
such that (0 ) = 1 and ( ) = 0 for each 6= 0 . This mixed strategy assigns probability 1 to 0 and
probability 0 to each other pure strategy of player . If player plays this mixed strategy, then it is as if he is
playing the pure strategy 0 . The set includes this mixed strategy and each other mixed strategy which gives
probability 1 to a pure strategy.
Example 3.15(ii) In
2 , 2 includes 2 such that 2 () = 0, 2 () = 1, 2 () = 0, which is equivalent to
2 = .
When players can choose mixed strategies, for each player the strategy set is , not , but in practice
includes , that is can still play a pure strategy by selecting a suitable mixed strategy. With = (1 ) we
denote a profile of mixed strategies, and = 1 × 2 × × denotes the set of all profiles of mixed strategies.
With − we denote a profile of mixed strategies for the players different from , and − is the set of all − .
If players play , then the outcome of the game is a random variable and we assume that the mixed strategies
of different players are stochastically independent random variables.
1 2 3 2
Example 3.13(ii) In
1 , consider 1 = 3 + 3 , 2 = 5 + 5 . Then
1 2 2 4
Pr{( )|} = , Pr{( )|} = , Pr{( )|} = , Pr{( )|} = (19)
5 15 5 15
When mixed strategies are played, the outcome is determined by a lottery over , hence each player needs to
have a preference relation among lotteries over . A typical assumption in this case is that the EUT (Proposition
25
2.9) can be applied to each player. Precisely, we assume that for each player his preference relation % on lotteries
on satisfies A1-A3 of the EUT, hence there exists a Bernoulli utility function : → R such that % 0 if and
P P
only if ∈ Pr{|} () ≥ ∈ Pr{|0 } () and is just the function which describes player ’s preferences
on introduced at the beginning of Subsection 3.1. Then, given , the expected utility for player is
X
() = () Pr{|}
∈
1 2 3 2
Example 3.13(iii) In 1 , consider 1 = 3 + 3 , 2 = 5 + 5 . Then we use (19) to compute 1 () =
0 · 15 + 4 · 15
2
+ 3 · 25 + 2 · 15
4
= 34 1 2 2 4 13
15 , 2 () = 3 · 5 + 1 · 15 + 2 · 5 + 4 · 15 = 5 .
For each player , represents his preference relation on . However, a more imprecise notation is common
and is used also for the expected utility function defined on , therefore
X X X
Y
() = () Pr{|} = ()1 (1 ) · · · ( ) = () ( )
∈ ∈ ∈ =1
denotes the utility function of player when mixed strategies are allowed. We now describe an important property
of . First, notice that X Y
( − ) = ( − ) ( )
− ∈− 6=
is player ’s expected utility from playing the pure strategy if the other players play − , and
X
( − ) = ( − ) ( ) (20)
∈
In words, the expected utility of player given − can be computed as the expectation (according to ) of
the expected utility of each of his pure strategies given − .
1 2 3 2
Example 3.13(iv) In
1 , consider 1 = 3 + 3 , 2 = 5 + 5 . Then
3 2 8 3 2 13 1 8 2 13 34
1 ( 2 ) = 0 · +4· = , 1 ( 2 ) = 3 · +2· = and 1 () = · + · =
5 5 5 5 5 5 3 5 3 5 15
as we know from Example 3.13(iii). Likewise,
1 2 7 1 2 3 7 2 13
2 ( 1 ) = 3 · +2· = , 2 ( 1 ) = 1 · +4· =3 and 2 () = · + ·3=
3 3 3 3 3 5 3 5 5
as we know from Example 3.13(iii).
Since represents player ’s preferences, it follows that player chooses in order to maximize , and then
we can give the definition of NE.
Definition 3.16 Given , we say that ∗ ∈ is a Nash Equilibrium (NE) of if for each player we have
(∗ ∗− ) ≥ ( ∗− ) for each ∈ (21)
This is the same idea as for a pure-strategy NE (Definition 3.9). A strategy profile ∗ such that for each player
, ∗ is a best reply for to the strategies − of the other players. The difference with respect to the definition
for pure strategies is that for each player , the set of feasible strategies is rather than .
1 1 2 3
Example 3.13(v) In ∗
1 , the profile = ( 2 + 2 5 + 5 ) is a NE. In order to see why, consider the viewpoint
of player 1 and notice that 1 ( ∗2 ) = 1 ( ∗2 ) = 12 ∗ ∗ ∗
5 , hence 1 (1 2 ) = 1 ( 2 )1 ()+1 ( 2 )1 () =
12 ∗ ∗
5 for each 1 ∈ 1 . Therefore each 1 ∈ 1 , including 1 , is a best reply for 1 to 2 . From the viewpoint of
player 2, we have 2 ( ∗1 ) = 2 ( ∗1 ) = 52 , hence 2 (2 ∗1 ) = 2 ( ∗1 )2 () + 2 ( ∗1 )2 () = 52 for
each 2 ∈ 2 . Therefore each 2 ∈ 2 , including ∗2 , is a best reply for player 2 to ∗1 .
26
In 1 there exists no pure-strategy NE, but there is a NE in mixed strategies. Indeed, the reason mixed
strategies are introduced is that we can prove quite generally that at least one NE exists when mixed strategies are
feasible, even though no pure strategy NE exists.
Proposition 3.17 (Nash’s Theorem) Hp: is such that 1 are all finite sets; Ts: There exists at least
one NE in if mixed strategies are feasible.
Therefore, allowing for mixed strategies guarantees the existence of a NE, unlike in the case in which only pure
strategies are feasible. In order to understand NE in this context, we need to examine the set of best replies when
mixed strategies are feasible. For this reason we denote with
• (i) + ( ) the support of a mixed strategy of player , that is + ( ) = { ∈ : ( ) 0} is the set
of pure strategies in to which gives positive probability;
• (ii) (− ) the set of best replies for player given − .
1 1 1 + 4 1 +
Example 3.15(iii) In
2 , if 2 = 6 + 2 + 3 then 2 (2 ) = { }; for 2 = 5 + 5 , 2 (2 ) = { }.
Definition 3.18 Given and − ∈ − , the set (− ) ⊆ is the set of mixed strategies of player which
maximize ( − ), that is (− ) = { ∈ : ( − ) ≥ (0 − ) for each 0 ∈ }.
Example 3.19 Suppose that = { } and − is such that
( − ) = 6, ( − ) = 6, ( − ) = 4, ( − ) = 6, ( − ) = 3 (22)
Then (− ) includes, for instance, the following mixed strategies: 17 + 47 + 27 , 15 + 45 , , 10
9 1
+ 10 ; these
mixed strategies are best replies to − since from (20) it follows that is maximized by giving positive probability
only to the pure strategies in with the highest utility given − (which is 6 in this case), but not necessarily to all
of them. Conversely, the mixed strategies 14 + 14 + 12 , , 19 + 23 + 29 are not best replies because they give
a positive probability to some pure strategy which has expected utility lower than 6. Precisely, (− ) consists
of all mixed strategies such that and are played with probability 0, that is { ∈ : + ( ) ⊆ { }}.
Now suppose that (22) is replaced by
( − ) = 2, ( − ) = 5, ( − ) = 3, ( − ) = 5, ( − ) = 3
Then (− ) = { ∈ : + ( ) ⊆ { }}.
Now suppose that (22) is replaced by
( − ) = 2, ( − ) = 5, ( − ) = 3, ( − ) = 4, ( − ) = 3
Then (− ) = {}.
Example 3.19 suggests that (− ) is the set of mixed strategies which give zero probability to pure strategies
which are not best replies to − . However, in order to identify NE the following (equivalent) characterization is
more useful:
( )
(i) if ∈ + ( ) and 0 ∈
+ ( ), then ( − ) ≥ (0 − )
(− ) = ∈ : (23)
(ii) if and 0 are in + ( ), then ( − ) = (0 − )
The interpretation of (23) is as follows: (i) each pure strategy in the support is at least as good as each pure strategy
outside the support; (ii) all the pure strategies in the support yield the same payoff. We can use (23) to characterize
NE as follows. First notice that ∗ is a NE of if and only if ∗ is a best reply to ∗− for = 1 , that is if
∗ ∈ (∗− ) for = 1 . Then the next proposition is a consequence of (23).
Proposition 3.20 Given , ∗ is a NE of if and only if for each player
27
• (i) ( ∗− ) ≥ (0 ∗− ) for each ∈ + (∗ ) and 0 ∈
+ (∗ );
• (ii) ( ∗− ) = (0 ∗− ) for each 0 in + (∗ ).
This is a necessary and sufficient condition for NE: Each player , given ∗− , is indifferent among all pure
strategies he plays with positive probability, and all these pure strategies are at least as good as each pure strategy
plays with 0 probability.
Proposition 3.20 implies that if ∗ is a pure-strategy NE, then ∗ remains a NE when mixed strategies can be
used. Precisely, consider any player and notice that + (∗ ) = {∗ }. Hence, condition (i) in Proposition 3.20
reduces to (∗ ∗− ) ≥ (0 ∗− ) for each 0 ∈ , given that ∗− = ∗− , which is satisfied since ∗ is a NE:
see Definition 3.9 and (18). Finally, condition (ii) in Proposition 3.20 does not impose any restriction as + (∗ )
consists of a single strategy of player .
Example 3.15(iv) We apply Proposition 3.20 to 2 to look for all NE of this game; Proposition 3.17 guarantees
that at least one NE exists. No NE exists such that player 1 plays a pure strategy: if 1 = , then player 2’s best
reply is 2 = , but then 1 does not want to play ; if 1 = , then 2’s best reply is 2 = , but then 1 does not
want to play . Hence 1+ (∗1 ) = { } in any NE, thus ∼1 is necessary. We need to determine 2+ (∗2 ),
and we rule out 2+ (∗2 ) = {}, 2+ (∗2 ) = {}, 2+ (∗2 ) = {} because they violate ∼1 .
• If 2+ (∗2 ) = { }, then ∼2 is 21 () + 7(1 − 1 ()) = 61 () + 4(1 − 1 ()), hence 1 () = 37 ,
2 ( 1 ) = 34 40 34
7 = 2 ( 1 ), but 2 ( 1 ) = 7 7 .
• If 2+ (∗2 ) = { }, then ∼1 cannot hold.
• If 2+ (∗2 ) = { }, then ∼2 is 61 () + 4(1 − 1 ()) = 10(1 − 1 ()), hence 1 () = 12 and
2 ( 1 ) = 2 ( 1 ) = 5, 2 ( 1 ) = 92 . The condition ∼1 reduces to 2 () + 5(1 − 2 ()) =
32 () + 2(1 − 2 ()), thus 2 () = 35 . The pair (∗1 ∗2 ) = ( 12 + 12 35 + 25 ) is a NE.
• If 2+ (∗2 ) = { }, then ∼2 ∼2 reduces to 2 ( 1 ) = 2 ( 1 ) = 2 ( 1 ) and there is no
solution to these equations. Hence, (∗1 ∗2 ) is the unique NE of
2 .
Notwithstanding the existence result of Proposition 3.17, there are some problems with the interpretation of
1 1 3 2
NE in mixed strategies. For instance, consider ∗ ∗
2 , = ( 2 + 2 5 + 5 ), and player 2. Given 1 , we have
∗ ∗ ∗ ∗ ∗
2 (2 1 ) = 5 and also 2 ( 1 ) = 5 = 2 ( 1 ). Hence, 2 has no strict incentive to play 2 : he may simply
play or , and obtain the same utility. Then, why should he bother to randomize according to ∗2 ? Recall that
∗2 is determined by the condition ∼1 , which is needed for player 1 to play a mixed strategy, but from the
point of view of 2, if he expects 1 to play ∗1 there is no strict reason to play ∗2 rather than or . This makes
it somewhat unclear whether we can expect that a mixed strategy NE is played.
Dominance with mixed strategies When players can use mixed strategies, the definition of strictly dominated
strategy is as follows:
Definition 3.21 Given , we say that ∗ ∈ strictly dominates ̄ ∈ for player if
(∗ − ) (̄ − ) for each − ∈ − (24)
This definition relies on the same logic which applies to strict domination with pure strategies, in Definition
3.4. However, (24) is equivalent to (∗ − ) (̄ − ) for each − ∈ − , that is it suffices to compare ’s
utility from ∗ and ’s utility from ̄ when the other players play pure strategies:22
(∗ − ) (̄ − ) for each − ∈ −
22 This holds because ( − ) = − ∈− ( − ) Pr{− |− }, thus (∗ − ) − (̄ − ) =
∗ ) − (̄ )] Pr{ | }, therefore (∗ ) (̄ ) for each
− ∈− [ ( − − − − − − − ∈ − if and only if
(∗ − ) − (̄ − ) 0 for each − ∈ .
28
In particular, a pure strategy ∈ is strictly dominated for player if and only if there exists ∗ ∈ such that
(∗ − ) ( − ) for each − ∈ − .
• When we allow for mixed strategies, it is more likely that a pure strategy is strictly dominated because
there are more strategies that can dominate it. Precisely, even though is not strictly dominated by a pure
strategy, it may be strictly dominated by a mixed strategy. In 3 , is not strictly dominated by or ,
but is by 12 + 12 . Therefore, iterated elimination of strictly dominated strategies is more effective if mixed
strategies are feasible. For instance, in 3 only the strategy profile ( ) survives
1\2 1\2
6 2 3 5 4 4
3 :
4 :
3 0 6 1 7 0
4 3 4 1 0 7
• If is such that ( ) 0 for some which is strictly dominated, then is strictly dominated. For
4 1 2 4 1 2 1 1 5 2
instance, in
3 , 7 + 7 + 7 is strictly dominated by 7 + 7 + 7 ( 2 + 2 ) = 7 + 7 , in which
has been replaced by a mixed strategy, 12 + 12 , that strictly dominates .
• Even though is such that ( ) 0 only for non strictly dominated pure strategies, it is possible that
is strictly dominated. For instance, in
4 (only the utility of player 1 is described in 4 ) and are
1 1
not strictly dominated, but 2 + 2 has support { } and is strictly dominated by for 1.
Justification for the use of NE The notion of NE receives a large attention because the most widespread
method which is used to foresee how players will play in a game with simultaneous moves is to identify the NE of
the game and expect that players will play one of them. For instance, for the game in Example 3.10, we expect
that players play ( ); for the battle of Sexes, we expect either ( ) or ( ).
The typical justification for this practice is as follows. Suppose that = 2 for simplicity (the generalization
of this idea to the case of ≥ 3 is immediate) and assume that game theory can identify a rational behavior for
players, that is a strategy profile ∗ which is the right way to play : call it the solution for . Suppose that
everybody can identify ∗ , included the players. Then player 1 learns that the solution indicates player 2 to play
∗2 , hence he expects ∗2 from 2. Moreover, the solution indicates him to play ∗1 . Will he play ∗1 ? He will do so
only if ∗1 is a best reply to ∗2 for him, otherwise he will want to use a strategy that gives him a higher utility
than the utility he obtains with ∗1 . The same argument for player 2 implies that ∗ must be a NE, otherwise the
solution becomes invalid when players learn it.
This idea gives a clear indication for games in which a unique NE exists, but not when there are multiple NE.
There exist different approaches to handle multiplicity of NE (i.e., to select a unique NE).
• One approach relies on the so-called refinements of NE, for which a large literature exists. A refinement is a
solution concept which requires a strategy profile to be a NE, and also requires some additional properties,
like a best reply property more general than required by Definitions 3.9, 3.16. Some refinements are very
illuminating and effective at killing ”bad” NE (see Section 3.2), but no refinement identifies in a convincing
way a unique solution for each game.
• Another approach relies on modifying somewhat the game in such a way that uniqueness of NE obtains. If
the game is used to model a particular real world setting (like competition among firms, for instance), then
the modified game should still be a decent representation of reality.
• A further approach relies on the so-called focal points. A focal point in is a NE of which for some reasons
seems more plausible than other NE.
29
Example 3.22 The following game has two pure-strategy NE, ( ) and ( )
1\2
7 1 6 4 9 2
8 5 4 3 5 4
Both players prefer ( ) to ( ), and then in games of this sort it is frequent to read in the literature that
players should be expected to coordinate on the NE that both prefer. The reason is that since the NE ( ) makes
both players happier with respect to the NE ( ), ( ) attracts the players’ attention more than ( ), and
thus is a focal point. This may be somewhat convincing if players can communicate before the game is played.23
Example 3.23 In the following game, each player simultaneously selects one French city from a set of three
cities:
1\2 Dijon Lille Paris
Dijon 10 10 0 0 0 0
Lille 0 0 10 10 0 0
Paris 0 0 0 0 10 10
There exist three pure-strategy NE in this game: (Dijon,Dijon), (Lille,Lille), (Paris,Paris), and each player is
indifferent among all the NE. However, (Paris,Paris) looks like a focal point because it requires players to indicate
a city that is likely more familiar to them than the other two cities.
Similar arguments can be made in suitable contexts about the importance of social conventions, of players’
background, ... For instance, in the battle of sexes between a specific boyfriend and girlfriend, it may be relevant
if one of them — for instance the girlfriend — is typically the dominating partner. In this case, we may expect the
girlfriend’s preferred NE to be played.
In a sense, when a focal point exists it helps a player to form an expectation on the actions of the other players,
which solves the essential problem of that player in a game. However, it is important to notice that focal points
rely on ad hoc arguments, as they depend on details which are not part of the game. Although these details may
be relevant sometimes, they make it impossible to derive a general theory. Arguing that a particular NE in a
particular game is a focal point depends more on a subjective judgement than on a systematic principle.
3.1.4 Games with incomplete information
In the games considered in subsection 3.1, the preferences of each player are common knowledge among players,
and then we say the game has complete information. However, sometimes a player does not know the preferences
of another player; in these cases we say the game has incomplete information. For instance, in a duopoly each firm
may know the own cost function, but may not know the cost function of the rival firm. A situation of this sort is
modelled by assuming that each player knows what are the possible utility functions of the other players and has
some beliefs about them, as we now describe.
In a game with incomplete information, each player privately observes a signal before he plays; is called
type of player ; the set of possible values for (i.e., the set of possible types for player ) is denoted , a finite set.
The type of player affects the utility of , that is player ’s utility is ( − ; ), hence it depends on strategies
− as usual, and on ’s type .
The other players do not observe , and consider it as a realization of a random variable with support and
probability distribution denoted by such that for each ∈ , ( ) is the probability that is the type of player
. The random variables that determine the types of different players are stochastically independent. Moreover, let
23 Ingeneral, preplay communication among players may help them to reach an agreement, but the agreement should still be a NE,
otherwise some player will have an incentive to violate it.
30
= (1 ) represent the profile of types of all players; = 1 × 2 × × is the set of all possible profiles of
types, thus ∈ . Furthermore, − is the profile of types of the players different from , and − = ×6= is the
set of all possible − .
Example 3.2(iii) Consider the duopoly setting introduced in Example 3.2(i), in which each firm’s cost function
has constant marginal cost. Assume that each firm observes the own marginal cost, but does not observe the
marginal cost of the other firm. Here, a firm’s type coincides with its own marginal cost. Assume that the set of
possible marginal costs for firm 1 is 1 = {15 35}, the set of possible marginal costs for firm 2 is 2 = {10 20}.
Therefore, there exist two types of firm 1 and two types of firm 2. Moreover, suppose that 1 2 are such that
1 (15) = 04, 1 (35) = 06, 2 (10) = 025, 2 (20) = 075 (25)
A game with incomplete information is sometimes called Bayesian game, is denoted with and consists of the
following elements: 1 1 1 1 , in which : × → R for = 1 . With respect
to a game with complete information, the additional ingredients are the sets 1 , the probability distributions
1 , and the fact that the utility of each player depends also on the player’s type.
Definition 3.24 A game with incomplete information is = { 1 1 1 1 } such
that all the elements in are common knowledge among players.
Example 3.2(iv) For the duopoly introduced in Example 3.2(iii), we have = { 1 2 1 2 1 2 1 2 }
with
= {1 2}; 1 = [0 140], 2 = [0 140]; 1 = {15 35}, 2 = {10 20}
1 (1 2 ; 15) = (300 − 21 + 2 )(1 − 15) 1 (1 2 ; 35) = (300 − 21 + 2 )(1 − 35)
2 (1 2 ; 10) = (300 − 22 + 1 )(2 − 10) 2 (1 2 ; 20) = (300 − 22 + 1 )(2 − 20)
and 1 2 are specified in (25).
In a game with incomplete information it is natural to think that the strategy played by a player depends on his
type, and for this reason we consider each type of player as a separate player. We denote with ( ) the strategy
played by type of player , therefore a profile of strategies is {1 (1 )}1 ∈1 { ( )} ∈ .
Example 3.2(v) For the duopoly, we consider a game with four players, not two: firm 1 with marginal cost 15,
firm 1 with marginal cost 35, firm 2 with marginal cost 10, firm 2 with marginal cost 20. Hence, a strategy profile is
1 (15) 1 (35) 2 (10) 2 (20) [or {1 (1 )}1 ∈{1535} {2 (2 )}2 ∈{1020} ] that is a strategy for each of the four players.
Now we show how to compute the utility of a type of player . Suppose that he plays a strategy 0 ∈ and
that the other players play the strategies { ( )} ∈ for 6= . Then the expected utility of type of player is
X
− (− ) [1 (1 ) −1 (−1 ) 0 +1 (+1 ) ( ); ] (26)
− ∈−
in which − (− ) = 1 (1 ) · · · −1 (−1 )+1 (+1 ) · · · ( ) is the probability that − are the types of the players
different from . Type of player is uncertain about the types of the other players, and these types are impor-
tant
Xas they determine the strategies of the other players. For the sake of brevity, (26) is written henceforth as
− (− ) [0 − (− ); ].
− ∈−
Example 3.2(vi) Consider type 15 of firm 1. This firm knows that it is facing type 10 of firm 2 (which charges
price 2 (10)) with probability 14 , or type 20 of firm 2 (which plays 2 (20)) with probability 34 . Therefore, the
expected profit of type 15 of firm 1 from playing 1 (15) is
1 3
[300 − 21 (15) + 2 (10)](1 (15) − 15) + [300 − 21 (15) + 2 (20)](1 (15) − 15) (27)
4 4
31
For type 35 of firm 1 playing 1 (35), the expected profit is
1 3
[300 − 21 (35) + 2 (10)](1 (35) − 35) + [300 − 21 (35) + 2 (20)](1 (35) − 35) (28)
4 4
Next definition introduces a NE for a game with incomplete information as a NE of the game in which each
type of each player is viewed as a separate player.
Definition 3.25 Given a Bayesian game = { 1 1 1 1 }, a pure strategy Bayes-Nash
equilibrium (BNE) of is a profile of strategies {∗1 (1 )}1 ∈1 {∗ ( )} ∈ such that for each player and for
each type ∈ we have
X X
− (− ) [∗ ( ) ∗− (− ); ] ≥ − (− ) [0 ∗− (− ); ] for each 0 ∈ (29)
− ∈− − ∈−
The interpretation of (29) is analogous to that of (18) in the definition of NE in . Given the strategy profile
{∗1 (1 )}1 ∈1 {∗ ( )} ∈ , type of player expects the types of the other players to play as described by
{∗− (− )}− ∈− , thus his expected utility if he plays 0 is the right hand side of (29). The strategy profile
{∗1 (1 )}1 ∈1 {∗ ( )} ∈ prescribes him to play ∗ ( ), thus (29) requires that ∗ ( ) gives type of player
an expected utility not smaller than his utility if he plays another strategy in . In other words, each type of each
player takes as given the strategies of other players and chooses a strategy that is a best reply for him.24
Example 3.2(vii) In the duopoly game, consider type 15 of firm 1. From (27) we see that his best reply is
1 3
1 (15) = 16 2 (10) + 16 2 (20) + 165 1 3 185
2 . Likewise, from (28) we obtain that 1 (35) = 16 2 (10) + 16 2 (20) + 2 is the
best reply for type 35 of firm 1. The expected profits of type 10 of firm 2 and of type 20 of firm 2 are
2 3
(300 − 22 (10) + 1 (15)) (2 (10) − 10) + (300 − 22 (10) + 1 (35)) (2 (10) − 10)
5 5
2 3
(300 − 22 (20) + 1 (15)) (2 (20) − 20) + (300 − 22 (20) + 1 (35)) (2 (20) − 20)
5 5
1 3
Therefore, the best reply of type 10 of firm 2 is 2 (10) = 10 1 (15) + 20 1 (35) + 80, and the best reply of type 20
1 3
of firm 2 is 2 (20) = 10 1 (15) + 20 1 (35) + 85.
The system of these four conditions yields the solution 1 (15) = 1661 1811
15 ' 11073, 1 (35) = 15 ' 12073,
2 (10) = 6551 6851
60 ' 10918, 2 (20) = 60 ' 11418, which is the unique BNE of the game.
3.2 Dynamic games
In the previous section we have considered games with simultaneous moves. But in many interesting settings the
players’ moves are not simultaneous. Consider for instance bargaining with offers and counteroffers; competition
among firms after a stage in which firms may have invested to improve their production efficiency; repeated
competition in an oligopoly setting ...
A dynamic game is any game which is not a game with simultaneous moves; hence, sometimes dynamic games
are called games with sequential moves. We use Γ to denote a generic dynamic game.
3.2.1 The extensive form of a game
The typical way to represent a dynamic game is called extensive form. The extensive form of a game reveals who
moves when, what are the feasible actions for each player, what a player knows when he moves, what is the outcome
as a function of players’ actions, what are the players’ preferences over outcomes.
24 If mixed strategies are feasible, then a strategy profile is { ( )}
1 1 1 ∈1 { ( )} ∈ and we can define mixed strategy BNE
of by modifying (29) in a straigthforward way. If is finite for each , then Proposition 2.9 can be applied to conclude that in
there exists at least one BNE.
32
In detail, the extensive form represents a game using a tree with nodes and branches. The tree has a unique
starting node; the nodes which are not final nodes are decision nodes, where players make decisions; the branches
represent players’ possible moves; the tree’s final nodes represent possible outcomes. At the end of Subsection 3.2.1
we provide a formal description for extensive forms.
Example 3.27(i) Suppose that two agents have 100 euros to split, and they bargain according to the following
procedure: Player 1 makes to player 2 a proposal ( 100 − ), according to which player 1 receives , player 2
receives 100 − . After learning player 1’s proposal, player 2 either accepts it or rejects it: if player 2 accepts, then
player 1’s proposal is implemented; if player 2 rejects, then all the money is lost and each player receives zero.
Suppose that {30 80} is the set of the possible values for among which player 1 can choose,25 and that each
player wants to maximize the amount of money he receives. The extensive form for this game is
2 2
A R A R
30 0 80 0
70 0 20 0
The initial node is at the top of the tree, where player 1 chooses one of the two possible values of ; each of the two
branches starting from the initial node represents a move of player 1. After player 1’s move, player 2 accepts 1’s
proposal or rejects it; therefore at the end of each of the previous branches is a decision node for player 2, where
he selects or — each choice of player 2 is represented by a branch. Each -branch and -branch leads to a
terminal node, where the outcome is described by the amount of money each player receives.
Example 3.27(ii) Consider the same setting as in Example 3.27(i), but assume that if player 2 rejects player
1’s offer, then the money is not lost and 2 makes a counteroffer to 1. Precisely, player 2 proposes a new split
( 100 − ) of the 100 euros, still with ∈ {30 80}. After 2’s proposal, 1 accepts it or rejects it — in case that 1
25 If we allow to be any real number in [0 100], then we obtain a tree with infinitely many branches starting from the initial node,
which looks like a bad idea for a first example.
33
rejects, all the money is lost and each player receives zero. The extensive form for this game is
2 2
A R A R
2 2
30 80
70 20
1 1 1 1
A R A R A R A R
30 0 80 0 30 0 80 0
70 0 20 0 70 0 20 0
In the two games of Example 3.27(i) and Example 3.27(ii), when a player moves he knows all the moves that
have been made in the past. Next example considers a game in which a player moves without knowing a move
made by another player.
Example 3.27(iii) Consider again the setting of Example 3.27(i), but now assume that player 2 must decide to
accept or reject player 1’s offer without observing . The extensive form for this game is such that the decision
nodes of 2 lie within an oval, with the interpretation that 2 cannot distinguish among the nodes in the oval:
2
A R A R
30 0 80 0
70 0 20 0
In general, a set of nodes in a same oval is called information set. If an information set includes two or more
nodes and play reaches one of them, then the player who moves at the information set does not know at which
node he actually is, but knows that one of the nodes of the information set has been reached.
34
Example 3.28(i) A game with three players:
A B C
2 2
D E F G H G H
3 3
2 0 3
4 6 3
7 5 0 I J I J L M L M
6 5 2 9 1 6 3 1
0 2 7 4 1 5 2 6
8 1 6 3 2 3 8 3
In this game, player 2 recognizes if player 1 has played , but cannot distinguish the case in which 1 has played
from the case in which player 1 has played . If player 3 is called upon to play, then he knows the move played
by player 1, but not whether 2 has played or .
The extensive form can also be used to represent games with simultaneous moves, as next example illustrates.
Example 3.27(iv) Consider again the setting of Example 3.27(i), but assume that the players move simultaneously:
This means that player 1 chooses = 30 or = 80 at the same time as player 2 chooses or . This game
has the same extensive form as the game of Example 3.27(iii) because the players’ informations and moves are the
same in the two games. Precisely, in the game of Example 3.27(iii) player 1 moves before player 2, and 2 moves
without knowing the move of 1. But also in the game we are considering in this example, player 2 moves without
knowing the move of player 1 because moves are simultaneous. As long as player 2 cannot observe 1’s move, the
actual timing of moves is irrelevant: what determines the structure of the tree is the information a player has at
the time he moves.
Example 3.1(v) In the Battle of Sexes, moves are simultaneous. Each of the two extensive forms below represents
this game (in both extensive forms the first number after each terminal node is the utility of player 1):
1 2
D M D M
2 1
D M D M D M D M
2 0 1 5 2 1 0 5
5 0 1 2 5 1 0 2
35
It is useful to consider each decision node as part of an information set. For instance, each decision node of
player 2 in the game of Example 3.27(i) is part of a singleton information set which consists only of that node,
even though for simplicity information sets that include a single node are often not drawn. With this convention
in mind, we say that a game has perfect information if each information set in the extensive form of the game is a
singleton, otherwise the game has imperfect information. Thus, a game with perfect information is a game in which
there are no simultaneous moves and each player, when he moves, knows all the previous moves. For instance, the
games in Examples 3.27(i) and 3.27(ii) have perfect information. The games in Examples 3.27(iii) and 3.28(i) have
impefect information.
It is important to notice that both for the case of perfect information and for the case of imperfect information,
given any non-initial node there exists a unique path that connects the initial node to the given node, that is there
exists a unique sequence of previous moves which lead to the given node. For instance, given the final node in the
game of Example 3.28(i) in which the players’ utilities are (6 5 3), in order to reach that node it is necessary that
player 1 plays , that player 2 plays , and finally that player 3 plays .
Some games include stochastic elements, that is events which are not controlled by the players. We view these
events as moves made by Nature according to an exogenous probability distribution. That is, Nature does not have
preferences over the outcomes of the game and its move is a random variable with a given probability distribution.
Example 3.27(v) Consider again the setting of Example 3.27(i), but assume that if player 2 rejects player 1’s
proposal, then Nature allocates all the money to 1 with probability 40%, all the money to 2 with probability 60%.
The extensive form of this game is
2 2
A R A R
Nature Nature
30 [0.4] [0.6] 80 [0.4] [0.6]
70 20
the money is the money is
allocated to 1 allocated to 1
the money is the money is
allocated to 2 allocated to 2
100 0 100 0
0 100 0 100
A formal description of extensive forms The extensive form of a game consists of the following elements:
• A set of players (which sometimes includes Nature), a set of actions, a set of nodes.
• A single node 0 ∈ is called initial node of the game and to each node ∈ \{0 } is associated a
unique immediate predecessor (); the nodes for which is an immediate predecessor are called immediate
36
successors of and gathered in the set (), that is, () = {0 ∈ : = (0 )}. In order to avoid cycles, a
successor of a node cannot be a predecessor of the same node (even if not immediate). If () = ∅, then is
called terminal node, and denotes the set of terminal nodes; if () 6= ∅, then is called decision node.
• The set \ of decision nodes is partitioned into a collection of information sets. Each ∈ (\) belongs
to an information set ∈ , and if 0 are in the same information set , then the set of possible actions
at and 0 is the same, denoted .
• If ∈ and 0 ∈ (), then there exists 0 ∈ that from leads to 0 [if 00 ∈ () and 00 6= 0 , then
00 ∈ that leads to 00 from is different from 0 ].
• To each ∈ is assigned a player that moves in ; ⊆ is the set of information sets in which player
plays. If the set of players includes Nature, then at each node of Nature we specify a probability distribution
over the possible actions at that node.
• For each player, his preferences on are represented by a utility function which is a Bernoulli utility function
in order to allow for randomized outcomes.
A dynamic game Γ is typically identified with its extensive form, therefore we use Γ also to indicate the game’s
extensive form.
3.2.2 Strategies, and Normal form representation
Starting from the extensive form of a game, it is possible to derive an alternative representation for the game which
relies on the notion of strategy.
Definition 3.29 Given Γ, a strategy of player in Γ is a function : → that to each ∈ associates an
action () which is feasible at , that is () ∈ .
A strategy of player is a complete plan of action for him: In Γ, if player is called upon to play then one of
his information sets has been reached; a strategy tells player how to play in each of these circumstances. Thus,
a strategy for player is a complete description of how to play Γ because it specifies a move for player for every
possible situation in which it is his turn to play, that is for every ∈ it indicates a move for player at . The
set of strategies for player is denoted , and is a generic strategy in
Example 3.27(vi) In the game of Example 3.27(i), the strategy set of player 1 is 1 = {30 80}. Player 2 has two
information sets and 2 = { }. For instance, is the strategy of player 2 such that 2 plays if
1’s proposal is (30 70), plays if 1’s proposal is (80 20). In case that player 2 follows this strategy and 1 chooses
= 80, then 2 accepts since = 80 implies that only the second part of matters.
37
Example 3.28(ii) For the game in Example 3.28(i), next figure emphasizes the information sets:
1
h11
A B C
2 h21 2
D E F G H
h22 G H
3 3
2 0 3
4 6 3 h31 h32
7 5 0 I J I J L M L M
6 5 2 9 1 6 3 1
0 2 7 4 1 5 2 6
8 1 6 3 2 3 8 3
For player 1, 1 = {11 } and 1 = { }. For player 2, 2 = {21 22 } and 2 = { }:
each strategy of 2 indicates a move for him both if he is playing at 21 and if he is playing at 22 . For player 3,
3 = {31 32 } and 3 = { }.
Example 3.30(i) Consider the following game:
1
x0 h11
A B
2
1
4 C D
1
h12
2 x
3 E F
4 3
2 5
In this game, each information set is a singleton. There exists a unique information set for player 2, at which
he chooses or ; hence 2 = { }. There exist two information sets for player 1, 11 and 12 ; 11 consists
of the initial node 0 , 12 consists of the node . Hence, each strategy of player 1 indicates a move at both these
information sets. For instance, is a strategy of player 1 which prescribes that 1 plays at 0 , plays if play
reaches . Also is a strategy of 1; it indicates that 1 plays at 0 , plays if play reaches . It may be
unintuitive that prescribes a move at given that the strategy itself rules out that this node is reached (since
indicates at 0 ) but this is a consequence of the definition of strategy. Precisely, 1 = { }.
38
After a player has decided to follow a strategy, his decision problem in the game has been solved because
whenever it is his turn to move, he can consult the strategy and it indicates him a move given the information set
he is in. In fact, once a player has chosen a strategy he may even be replaced in the game by a representative who
just follows the strategy’s instructions. As for games with simultaneous moves, = (1 2 ) denotes a profile
of strategies, = 1 × 2 × × is the set of all strategy profiles, − = (1 −1 +1 ) is a profile of
strategies of players different from , and − = ×6= .
We now introduce a representation of a game which is alternative to the extensive form. The key observation is
that after each player has chosen a strategy, each player’s actions in the game are completely determined. In other
words, given a strategy profile ∈ , for each ∈ , determines the action which is played at , and therefore
identifies a path from the initial node to a terminal node. In particular, each determines a unique terminal node,
to which a profile of utilities is associated.
Example 3.28(iii) In the game of Example 3.28(i), consider = ( ). The resulting path in the tree
reaches the terminal node at which the players’ utilities are (3 2 8). Conversely, 0 = ( ) determines a
path that results in the utilities (3 3 0).
Example 3.30(ii) In the game of Example 3.30(i), the strategy profile ( ) identifies the terminal node with
utilities (3 5).
Since to each ∈ is associated a profile of utilities (1 () 2 () ()), we can think of the game as
follows: each player simultaneously chooses a strategy and the resulting strategy profile generates the utilities
(1 () 2 () ()). Therefore, we can view the game as a simultaneous move game { 1 1 } (see
Subsection 3.1), which is called normal form representation or strategic form representation.
Example 3.27(vii) For the game of Example 3.27(i), the normal form is as follows:
1\2
30 30 70 30 70 0 0 0 0
80 80 20 0 0 80 20 0 0
Example 3.30(iii) For the game of Example 3.30(i), we have 1 = { }, 2 = { } and the
normal form is
1\2
1 4 1 4
1 4 1 4
2 3 4 2
2 3 3 5
If we compare the extensive form with the normal form, it should be clear that the former is very detailed and
rich of informations, whereas the latter is simpler but poorer of informations, as it consists only of strategy sets
and utility functions.26 For instance, the normal form does not illustrate how a given identifies a path in the tree
and eventually a terminal node: it merely associates to a utility profile. One way to see that the normal form
includes fewer informations than the extensive form is to notice that for each given extensive form there exists a
unique normal form, but the same normal form may be obtained starting from different extensive forms.
Example 3.31 For each of the following two games, the normal form is the same (in both extensive forms the first
26 But notice that in games with simultaneous moves, the extensive form gives no additional information with respect to the normal
form. See for instance the game of Example 3.27(iii).
39
number after each terminal node is the utility of player 1)
1 2
A B C D
2 1
2
2
C D A B A B
3 1 2 3 2 1
1 0 2 1 2 0
3.2.3 Behavior strategies
We have seen in Subsection 3.1 that in simultaneous move games, NE may not exist if we do not allow players to
randomize their strategy choices; the same problem arises in dynamic games. We may still describe randomized
choices using mixed strategies, but when we use the extensive form to represent a game, there is another way to
conceive randomized choices.
Definition 3.32 Given Γ, a behavior strategy for player specifies for each information set ∈ a probability
distribution over the set of the available moves at . Hence, = { }∈ is such that for each ∈ ,
X
() ≥ 0 for each ∈ , () = 1
∈
A behavior strategy for player is a family of probability distributions, one for each information set of player
, and for each ∈ it specifies how he randomizes over the moves he can play at . That is, if play reaches
, then player chooses an action from the set according to the probability distribution .27 The set of
behavior strategies of player is denoted . A profile of behavior strategies is denoted = (1 2 ), and
= 1 × 2 × × is the set of behavior strategy profiles.
Example 3.30(iv) In the game of Example 3.30(i) there are two information sets for player 1, 11 and 12 . Hence,
a behavior strategy of player 1 specifies a probability distribution on { } and a probability distribution on
{ }, for instance (1 11 () 1 11 ()) = (03 07), (1 12 () 1 12 ( )) = (06 04). This is interpreted as follows:
(i) in 11 , player 1 plays with probability 0.3, plays with probability 07 (he plays 03 + 07); if play reaches
12 , then player 1 plays 06 + 04 .
Behavior strategies also allow for non-randomized choices at some information set of player . For instance, in
the game of Ex 3.30(i), 1 such that (1 11 () 1 11 ()) = (1 0), (1 12 () 1 12 ( )) = ( 13 23 ) prescribes that 1 plays
with certainty at 11 . Another example is 1 such that (1 11 () 1 11 ()) = (0 1), (1 12 () 1 12 ( )) = (1 0),
which is equivalent to the pure strategy .
The main difference between a behavior strategy a mixed strategy is the number of randomizations which the
strategy involves. A behavior strategy of player relies on several randomizations: a randomization occurs every
time it is player ’s turn to move. Conversely, a mixed strategy of player relies on a single randomization (over
), which occurs at the beginning of the game; then player follows throughout the game the pure strategy
which has been selected. For instance, in the game of Example 3.30(i), one mixed strategy of player 1 is 1 =
03 + 05 + 02 , according to which player 1 plays the pure strategy (, ) with probability 03
27 Although the superscript in
() is useful for reasons of clarity, sometimes no ambiguity results from suppressing it. Then, for
the sake of brevity sometimes we will skip .
40
(probability 05, 02). If the pure strategy is selected, then player 1 follows the indications of this strategy at
each of his information sets.28
When considering dynamic games, we will mostly use behavior strategies because they are more natural and
easier to interpret than mixed strategies. But notice that behavior strategies and mixed strategies are equivalent
for games with perfect recall, that is for games such that no player forgets what he once knew (included his own
past actions).29 We consider only games with perfect recall in the following, and the equivalence is defined as
follows.
Definition 3.33 Given Γ, we say that ∈ and ∈ are equivalent if, for each − ∈ − , and induce
the same probability distribution on the set of the terminal nodes of Γ.
Example 3.30(v) Consider the game of Example 3.30(i) and 1 such that (1 11 () 1 11 ()) = (03 07), (1 12 () 1 12 ( )) =
(06 04). We claim that 1 = 03 + 042 + 028 is a mixed strategy equivalent to 1 . In order to see why,
suppose that player 2 plays 2 = . Then the probability distribution over the four terminal nodes is 03 07 0 0
both if player 1 plays 1 and if he plays 1 . In case that player 2 plays , the probability distribution over is
03 0 042 028 both if 1 plays 1 and if he plays 1 (notice that also 1 = 02 + 01 + 042 + 028
is equivalent to 1 ).
Therefore, player 1 is indifferent between 1 and 1 in Example 3.30(v) because, regardless of how 2 plays, the
two strategies generate the same outcome, that is the same probability distribution over the set , hence they yield
the same expected utility to player 1.
Given any game with perfect recall and any player in the game, for each behavior strategy of player there
exists an equivalent mixed strategy of player , and for each mixed strategy of player there exists an equivalent
behavior strategy. Next example starts from a mixed strategy for player 1 in the game of Example 3.30(i) and
determines and equivalent behavior strategy.
Example 3.30(vi) For the game of Example 3.30(i), consider ̄1 = 08 + 02. Then ̄1 such that
(̄1 11 () ̄1 11 ()) = (08 02), (̄1 12 () ̄1 12 ( )) = (1 0) is equivalent to ̄1 because in either case the proba-
bility distribution on is 08 02 0 0 if 2 plays , is 08 0 02 0 if 2 plays .
3.2.4 Subgame perfection
One way to study a dynamic game consists in deriving its normal form and then applying to it the solution concepts
for games with simultaneous moves, for instance the notion of NE (or strict domination, ...). However, in some
cases this notion does not work well, as next example shows.
28 Notice that if for a player there exists a unique information set in the game, then for him there is no difference between mixed
strategies and behavior strategies, that is the set coincides with . For instance, see player 2 in the game of Example 3.30(i).
29 Formally, a game is said to have perfect recall when it satisfies the following conditions: (i) if and 0 are in ∈ , then is not
00
a predecessor nor a successor of 0 ; (ii) if and 0 are in ∈ , 00 in 00 ∈ is a predecessor of , and 00 ∈ is the action on
the path from 00 to , then there exists 000 in 00 which is a predecessor to 0 such that 00 is the action on the path from 000 to 0 .
41
Example 3.34(i) Consider the game with the following extensive form, and next to it its normal form:
A B
1\2
5
x 2 5 4 5 4
4 7 3 2 0
C D
7 2
3 0
From the normal form we see that two pure-strategy NE exist: ( ) and ( ).30 But ( ) is in fact
unsatisfactory, in the sense that we should not expect that players play ( ). In this NE, is a best reply for
player 1 because, given 2 = , player 1’s utility if he chooses is 2 instead of 5.31 But if we look at the extensive
form, does it make sense to believe that player 2 will play at node ? At node , player 2 prefers to play
rather than as 3 is greater than 0. This suggests that 1 should play rather than because then 2 will play
and 1’s utility will be 7 rather than 5. This makes ( ) unconvincing even though it is a NE.
In a NE, each player takes as given the strategies of other players, and if player 1 takes 2 = as given, then
1 = is his best reply. But since 1 plays before 2, he should not take 2 = as given and should expect that 2
will choose if he is called to move.
In a sense, in ( ) the strategy 2 = is like a threat of player 2 towards 1: ”If you dare to play , then I
will play and you will have a low utility equal to 2”. If this threat scares 1 and induces him to play , then 2
will be better off without actually having to act consistently with his threat. But player 2’s threat is not credible
(and 1 should not be scared by it) because if play reaches , then it is in 2’s own interest to play rather than .
Example 3.34(i) suggests that the notion of NE is not appropriate for dynamic games because it does not rule
out NE based on incredible threats. As a consequence, a stronger equilibrium notion is introduced below in order
to eliminate this sort of NE. This equilibrium notion relies on the so-called subgames of a game.
Definition 3.35 Given Γ, a subgame of Γ is a subset of Γ such that (i) it begins with a singleton information set;
(ii) it contains all the nodes which are successors of its initial node, and contains only these nodes; (iii) if a decision
node ∈ is in the subgame, then also all the other nodes in are in the subgame.
Hence, a subgame of Γ is a subset of the game which satisfies the properties (i)-(iii) in Definition 3.35. In
particular, property (iii) requires that a subgame does not break any information set. Notice that the whole game
is a subgame of itself, and it is called improper subgame. Any subgame of Γ which does not coincide with Γ is said
to be a proper subgame of Γ.
30 There also exist NE in mixed strategies: See footnote 32.
31 Furthermore, 2 = is a best reply for player 2 given 1 = .
42
Example 3.36 Consider the following game, for which it is not necessary to specify the actions and the utilities:
2 x 2
3 y 3 z 3
This game has three proper subgames: one starts at node , one starts at node , one start at node .
Example 3.34(ii) The game of Example 3.34(i) includes one proper subgame, which starts with node .
An important feature of a subgame is that, considered in isolation (i.e., independently of the rest of Γ), it is
itself a game. If play reaches the initial node of subgame, then it is common knowledge for all the players in the
subgame that this subgame is being played, and this makes irrelevant the rest of Γ; in particular, the past moves
do not matter. Since we expect that players play a NE in each game they find themselves in, it makes sense to
think that in this subgame they play a NE of the subgame. This argument is the basis for the following definition.
Definition 3.37 Given Γ, a strategy profile ∗ is said to be a Subgame Perfect Nash Equilibrium (SPNE) of Γ if
and only if ∗ induces a NE in each subgame of Γ.
Recall that a strategy of a player specifies how he plays at each of his information sets, hence a profile of
strategies indicates how the players play in each given subgame of Γ. The definition of SPNE requires that ∗
prescribes a NE in each subgame, when each subgame is considered in isolation. Since Γ is a subgame of itself, this
condition implies that any SPNE of Γ is a NE of Γ. Moreover, it requires that ∗ prescribes a NE in each proper
subgame of Γ.
Example 3.34(iii) In the game of Example 3.34(i) there exists a unique proper subgame, which starts with the
node and is a one-player game; its unique NE is . Since the NE ( ) indicates to player 2 to play in this
subgame, it does not induce a NE in the subgame; hence ( ) is not a SPNE of the whole game. Conversely,
the NE ( ) is a SPNE as it prescribes in the subgame, which is a NE in the subgame. Indeed, ( ) is the
unique SPNE of the game.32
We have observed above that each SPNE ∗ is necessarily a NE. From Example 3.34(iii) we see that the reverse
implication is untrue, that is if ∗ is a NE, then it is not necessarily a SPNE: ( ) is a NE in the game of Example
3.34(i) but is not a SPNE. Thus the notion of SPNE is a refinement of NE.
Example 3.34(iii) helps to see the basic idea behind Definition 3.37. Recall that in the NE ( ), player 1 plays
because he expects (fears) that player 2 plays if play reaches node . Even though is not reached along the
equilibrium path according to ( ), the action 1 expects 2 would play at affects 1’s action at the initial node
32 In this game there also exist NE in which 1 plays , and 2 plays with probability , 2 plays with probability 1 − , with
∈ (0 35 ]. Each of these NE prescribes to player 2 to play with positive probability in the subgame, hence it is not a SPNE.
43
in Γ. Definition 3.37 verifies whether 1’s expectation (or 2’s threat) makes sense, that is, whether is a NE in
the one-player subgame which starts with node . Since this is not the case, we conclude that 1’s expectation is
unreasonable and the NE ( ) fails to be a SPNE. For any other game, SPNE relies on the same principle.
Next, we introduce a property of each NE which helps to understand better the difference between NE and
SPNE. This property requires to distinguish whether, given a strategy profile , a subgame is entered with positive
probability or with zero probability. We use to denote a generic subgame of Γ.
Definition 3.38 Given Γ, a subgame and a NE ∗ , we denote with Pr{|∗ } the probability that play reaches
the initial node of when players play ∗ . Then is said to lie on the equilibrium path if Pr{|∗ } 0, is said
to lie off the equilibrium path if Pr{|∗ } = 0.
Example 3.39(i) Consider the following game:
A B
2
x
6
3 C D
2
1 x' x'' 1
E F G H
3 3
I J I J K L K L
5 12 8 10 6 2 3 5
10 4 6 0 9 1 7 5
8 12 9 7 3 7 4 5
in which denotes the subgame which starts with node , 0 denotes the subgame which starts with node 0 ,
00 denotes the subgame which starts with node 00 . Suppose that there exists a NE ∗ of the game such that
∗1 () = 1, ∗1 () = 0. Then Pr{ |∗ } = 0, Pr{ 0 |∗ } = 0, Pr{ 00 |∗ } = 0, and the three subgames , 0 , 00
all lie off the equilibrium path.
Now suppose there exists a NE ∗ such that ∗1 () = 27 , ∗1 () = 57 and ∗2 () = 1, ∗2 () = 0. Then Pr{ |∗ } = 57 ,
Pr{ 0 |∗ } = 57 , Pr{ 00 |∗ } = 0, hence 00 is the only subgame off the equilibrium path.
Finally, if a NE ∗ is such that ∗1 () 0, ∗2 () 0, ∗2 () 0, then , 0 , 00 are all on the equilibrium path.
Next proposition describes the difference between NE and SPNE.
Proposition 3.40 Given Γ, a strategy profile ∗ is a NE of Γ if and only if ∗ induces a NE in each subgame of Γ
which lies along the equilibrium path.
This says that each NE ∗ automatically satisfies the condition of Definition 3.37 in all subgames along the equilibrium path:
since ∗ is a NE, it follows that ∗ induces a NE in each subgame such that Pr{|∗ } 0. But this may not
occur in a subgame off the equilibrium path. Thus, the notion of SPNE is more restrictive than NE because of the
condition it imposes on subgames off the equilibrium path.
44
Therefore, in order to a be a SPNE, a NE ∗ is required to prescribe a NE (i.e., ”reasonable” play) in each
subgame of Γ such that Pr{|∗ } = 0. The reason of this requirement by Definition 3.38 is that if a certain
subgame is not entered when players play ∗ , it is because one player, say player , has chosen an action which
prevents from being entered, and he has done so under the expectation that players follow the prescriptions of
∗ if is entered. But if ∗ prescribes that players do not play a NE in , then the expectation of player seems
unsound, and the whole ∗ is unconvincing.
Example 3.41(i) Consider the following game Γ, for which we provide also the normal form:
A B
2 x x' 1 1\2
5 2 5 2 3 6 3 6
G H C D 5 2 5 2 3 6 3 6
0 4 1 5 0 4 1 5
2 4 3 2 2 4 3 2 2
5 3
2 6
E F E F
0 1 4 2
4 5 3 2
From the normal form we see that three pure-strategy NE exist: ( ), ( ), ( ), and from the
extensive form we see that only two proper subgames exist: which starts with node , 0 which starts with
node 0 . In order to verify whether ( ) is a SPNE, in view of Proposition 3.40, it suffices to verify whether
( ) induces a NE in 0 , which is the only subgame off the equilibrium path. This subgame seen in isolation
and its normal form are
1
C D
1\2
2
0 4 1 5
E F E F 4 3 2 2
0 1 4 2
4 5 3 2
Here, ( ) is not a NE since is not a best reply for player 1 when 2 plays ; hence ( ) is not a SPNE
of Γ. Using the terms described just before this example, in the NE ( ) player 1 plays at the beginning
of the game because he expects that ( ) is played if 0 is entered. This yields him utility 1, which is less 3, his
utility if is entered, given that 1 expects that is played in . But it is not reasonable that 1 expects that
( ) is played in 0 because ( ) is not a NE of 0
45
In order to see whether ( ) is a SPNE, we verify whether it induces a NE in 0 . In this subgame, ( )
is not a NE because is not a best reply for player 2 when 1 plays ; hence ( ) is not a SPNE of Γ.
Finally, we see whether ( ) is a SPNE by verifying whether it induces a NE in . This subgame is a
one-player game in which is the unique NE, thus ( ) is a SPNE of Γ.
Example 3.34(iv) The game of Example 3.34(i) shows that a player may benefit from reducing his own set of
moves. In this game, both ( ) and ( ) are NE, player 2 prefers ( ) to ( ), but ( ) is the unique
SPNE. Then it would be profitable for player 2 to reduce the set of his possible actions at node from { } to
{}, because in this case he would be forced to play if 1 plays ; this would induce 1 to play . It is remarkable
that an agent cannot gain from reducing his choice set in an individual decision problem (like those in Example
1.2(i) and Example 1.3(i)), but in a game this may be useful for a player if it induces the other players to modify
suitably their behavior.33
Backward induction We now introduce a procedure, called backward induction, which allows to identify all
the SPNE of a game (precisely, this procedure works for each game in which the set of actions is finite). The basic
idea relies on examining each subgame by itself and identifying all its NE, starting with the subgames which do
not include any other subgame — these are called terminal subgames.
In detail, backward induction works as follows, given a a dynamic game Γ.
• Step 1: For each terminal subgame, identify all the NE of the subgame;
• Step 2: In each terminal subgame, select one NE of subgame and replace the subgame with its equilibrium
utilities;
• Step 3: For the reduced extensive form game obtained after Step 2, repeat Steps 1 and 2 until a strategy
profile for Γ is derived.
Notice that if each subgame of Γ has a unique NE, then Γ has a unique SPNE. Otherwise, we need to consider
each possible selection of NE in each subgame, and as a result we derive all the SPNE of Γ.
Ex 3.39(ii) Consider the game of Ex 3.39(i). This game has three proper subgames, which are 0 00 ; the
last two are terminal subgames. The normal form for 0 is
1\3
5 8 12 12
8 9 10 7
In this game there exist three NE: ( ), ( ), and ( 13 + 23 25 + 35 ). The normal form for 00 is
1\3
6 3 2 7
3 4 5 5
In this game there exists a unique NE: ( ). Thus, there is a unique NE to select in 00 , and three NE in 0 .
If we select the NE ( ) in 0 , then the players’ utilities are (8 6 9) in case that play reaches 0 and we obtain
the following reduced game in which 0 has been replaced by (8 6 9), and 00 has been replaced by (5 5 5), the
33 Notice that for player 2, reducing the set of his feasible actions to {} is equivalent to committing ex ante to play in case node
is reached.
46
players’ utilities if 00 is entered:
1
A B
6 2
x
3
2 C D
8 5
6 5
9 5
In this reduced game, the unique subgame starts with node , and only player 2 plays in this subgame. The unique
NE of this subgame is because 6 5. Thus we obtain a further reduced game in which the subgame is replaced
by (8 6 9):
1
A B
6 8
3 6
2 9
The only player in this game is player 1, and the unique NE is since 8 6. This yields the SPNE ( )
for the whole game.
If instead we select the NE ( ) in 0 , then the players’ utilities are (12 4 12) in case that play reaches 0 , thus
we replace 0 with (12 4 12) and replace 00 with (5 5 5). The reduced game we obtain is
A B
6 2
x
3
2 C D
12 5
4 5
12 5
Arguing as above we identify the SPNE ( ) for the whole game.
Finally, if we select the NE ( 13 + 23 25 + 35 ) in 0 , then the players’ utilities are (465 5615 263) in case
that play reaches 0 , thus we replace 0 with (465 5615 263) and replace 00 with (5 5 5). The reduced game
47
we obtain is
1
A B
6 2
x
3
2 C D
46/5 5
56/15 5
26/3 5
Arguing as above we find a SPNE such that (1 () 1 ()) = (1 0), (1 () 1 ( )) = ( 13 23 ), (1 () 1 ()) =
(0 1); (2 () 2 ()) = (0 1); (3 () 3 ()) = ( 25 35 ), (3 () 3 ()) = (0 1). Thus, there exists three SPNE for
this game, two of which generate the same outcome.
Ex 3.41(ii) If we apply backward induction to the game of Ex 3.41(i), we see that is the unique NE in , and
( ) is the unique NE in 0 . This yields the following reduced one-player game
A B
3 4
6 3
for which the unique NE is . Hence, ( ) is the unique SPNE for the game, as we know from Example
3.41(i).34
Backward induction provides a perspective on SPNE which Definition 3.37 does not emphasize. When a player
moves at any point in the game, he must anticipate how the other players will play after his move. The notion of
SPNE implies that he expects that in each subgame which follows his move, a NE of the subgame will be played.
On the selective power of SPNE Since the notion of SPNE is a refinement of NE, it is natural to inquire
when it is very selective, and when it is not. In general, SPNE has a modest selective power in games with only
few subgames because then Definition 3.37 imposes few restrictions on a strategy profile. To the extreme, if Γ has
no proper subgame then each NE of Γ is a SPNE of Γ.
Conversely, SPNE is a very selective refinement in games with many subgames because then Definition 3.37
imposes many restrictions on a strategy profile. To the extreme, if Γ has perfect information then each terminal
subgame is a one-player game. If we apply backward induction, it is straightforward to find a NE in each terminal
subgame — for simplicity, assume that for each two nodes, no player is indifferent between the nodes, therefore the
NE is unique. Each reduced game still has perfect information, hence each terminal subgame in each reduced game
has a unique player and a unique NE. As a consequence, a unique SPNE exists, and it is a pure-strategy profile.
34 In fact, from Example 3.41(i) we know that ( ) is the unique SPNE in pure strategies. Backward induction reveals that
( ) is the unique SPNE even if behavior strategies are feasible.
48
3.2.5 Refinements based on beliefs
A motivating example In some games the notion of SPNE does not work well because it does not eliminate
all the NE that for some reasons are unsatisfactory.
Example 3.42(i) For the following game we represent the extensive form and the normal form:
A B C
2 1\2
4 x1 2 x2 2 4 2 4
h2 1 0 0 3
D E D E 0 1 3 2
1 0 0 3
0 3 1 2
This game has two pure-strategy NE: ( ) and ( ), and both are SPNE because in this game there exists no
proper subgame; thus each NE is a SPNE. However, the NE ( ) is implausible because if player 2 is called upon
to play at 2 , then his best move is rather than . Precisely, 2 does not know whether he is playing at node 1
or 2 , but in both nodes the move is superior to as 3 0 and 2 1. Hence, player 1 should expect that 2
will play , rather than , if 2 is reached. Then is a profitable deviation for 1.
The game in Example 3.42(i) is somewhat similar to the game of Example 3.34(i). There exists a NE, ( ),
such that player 2 is not called upon to move along the equilibrium path, but if 2 is reached then ( ) indicates
to player 2 to play , which is suboptimal for him because it does not maximize his utility once 2 has been
entered. However, player 1 chooses precisely because he fears that 2 will play if play reaches 2 ( is a sort of
threat of 2 to 1). In the game of Example 3.34(i), the notion of SPNE kills the NE which relies on an implausible
move off the equilibrium path. But in the game of Example 3.42(i) the notion of SPNE does not eliminate ( )
because it does not eliminate any NE.
We introduce now a new refinement of NE in order to eliminate undesirable NE such as ( ) in Example
3.42(i). This equilibrium notion requires that when a player moves at an information set , he has beliefs on the
past moves that led to , and his strategy prescribes an optimal action for him at given his beliefs.
Beliefs and Sequential rationality Beliefs play a crucial role in the equilibrium notion to introduce, and they
are defined as follows.
Definition 3.43 Given Γ, a system of beliefs = { }∈ for Γ specifies, for each ∈ , a probability distribution
P
on the nodes of such that () ≥ 0 for each ∈ and ∈ () = 1.
A system of beliefs consists of a family of probability distributions, one for each information set of the game,
denoted . The interpretation is that if play reaches , then for the player who moves at , () is the probability
of being at node , for each ∈ . Hence, represents the beliefs of this player about what happened in the past,
given that play has reached .
49
Example 3.44(i) Consider the following game:
A B C
1
6
3 x1 2 x2
h2
D E D E
x3 x4 3 x5 x6 3
h32
h31
F G F G F G H I L
2 3 4 5 6 7 8 9 10
7 4 9 1 5 2 6 4 5
9 1 3 7 3 2 7 5 4
In this game there exist two non-singleton information sets, one of player 2 denoted 2 , and one of player 3 denoted
31 . A system of beliefs is, for instance, such that
2 (1 ) = 03, 2 (2 ) = 07
31 (3 ) = 01, 31 (4 ) = 06, 31 (5 ) = 03
Then, if player 2 is called upon to play at 2 , he thinks he is at node 1 with probability 03, he is at node 2 with
probability 07. Likewise, if 31 is reached then player 3 attaches probability 01 to node 3 , probability 06 to node
4 , probability 03 to node 5 . There also exists a singleton information set 32 , and necessarily 32 (6 ) = 1.
The equilibrium notion to introduce requires that a strategy profile is such that for each player and for each
∈ , prescribes a utility maximizing move for player at given and given that the other players will
play − in the remaining of the game. Since the optimal move for player may depend on the node in where he
is playing, the probability distribution is used to compute player ’s expected utility from his possible moves.
Formally, given ∈ and ∈ , let ( − ) be the utility of player given that (i) he is playing at node ;
(ii) he plays and the other players play − . Let ( − ) be the expected utility of player given that (i)
he plays at and attaches probability () to each ∈ ; (ii) he plays and the other players play − ; thus
P
( − ) = ∈ () ( − ).
Example 3.44(ii) Consider the game of Example 3.44(i), the profile = ( ), and the beliefs described in
Example 3.44(i). Then 2 2 ( −2 ) = 03 · 4 + 07 · 2 = 26, 3 31 ( −3 ) = 01 · 1 + 06 · 7 + 03 · 2 = 49.
Definition 3.45 Given Γ and , a strategy profile is said to be sequentially rational at ∈ given if
( − ) ≥ (0 − ) for each 0 ∈ (30)
If this condition holds for each ∈ , then is said to be sequentially rational in Γ given .
50
Property (30) means that prescribes in an optimal move for player given . Thus, sequential rationality
tests a strategy profile at each ∈ and verifies whether it indicates an optimal action everywhere in the game,
given the beliefs.
Example 3.44(iii) Consider the game of Example 3.44(i) and = ( ). We inquire whether is sequentially
rational in this game, given the beliefs in Example 3.44(i).
Consider player 2 moving at 2 . We know from Example 3.44(ii) that 2 2 ( −2 ) = 26, hence now we compute
2 2 ( −2 ) = 03 · 1 + 07 · 5 = 38, which is greater than 2 2 ( −2 ). Hence, 2 = is not sequentially
rational at 2 given the beliefs (notice that would be sequentially rational at 2 if 2 (1 ) were equal to 08, for
instance).
Consider player 3 moving at 31 . We know from Example 3.44(ii) that 3 31 ( −3 ) = 49, hence now we compute
3 31 ( −3 ) = 9 · 01 + 3 · 06 + 3 · 03 = 36, which is smaller than 3 31 ( −3 ). Hence, 3 = is sequentially
rational at 31 given beliefs (notice that would not be sequentially rational at 31 if 31 (3 ) were equal to 1,
for instance).
Finally, consider player 3 moving at 32 , where beliefs are 32 (6 ) = 1. It is immediate to see that 3 = is not
sequentially rational at 32 as is superior to the move .
Weakly consistent beliefs As Example 3.44(iii) suggests, the sequential rationality of a strategy typically
depends on beliefs. Therefore we need to establish which beliefs must be used. Since in equilibrium the players
play a profile ∗ , beliefs should not be independent of ∗ and it is necessary to introduce a distinction among
information sets.
Definition 3.46 Given a Γ, a strategy profile , a node , an information set , we denote with
• Pr{|} the probability that the node is reached when players play ;
• Pr{|} the probability that the information set is reached when players play .
If ∗ is a NE of Γ, then we say that an information set is on the equilibrium path if Pr{|∗ } 0; we say that
is off the equilibrium path if Pr{|∗ } = 0.
Therefore, an information set is on the equilibrium path if is entered with positive probability given that
players play the equilibrium strategies.
Example 3.44(iv) Consider the game of Example 3.44(i), and = ( ). Then Pr{1 |} = 0, Pr{2 |} = 1,
Pr{3 |} = 0, Pr{4 |} = 0, Pr{5 |} = 0, Pr{6 |} = 1 and Pr{2 |} = 1, Pr{31 |} = 0, Pr{32 |} = 1. If were
a NE, then 31 would be an information set off the equilibrium path; 2 and 32 would be information sets on the
equilibrium path.
Now consider such that 1 = 13 + 12 + 16 , 2 = 57 + 27 , 3 = . Then Pr{1 |} = 12 , Pr{2 |} = 16 ,
5
Pr{3 |} = 14 , Pr{4 |} = 17 , Pr{5 |} = 42
5 1
, Pr{6 |} = 21 and Pr{2 |} = 23 , Pr{31 |} = 13 1
21 , Pr{32 |} = 21 . If
were a NE, then 2 , 31 , 32 would all be information sets on the equilibrium path.
Now we discuss how to determine beliefs, given a NE ∗ . Consider an ∈ on the equilibrium path, and
notice that if play reaches then player can reasonably believe that ∗ is being played because is reached with
positive probability when players play ∗ . Thus, he should compute () as the conditional probability of node ,
given that has been entered, and given that the players play ∗ . This conditional probability is computed using
Bayes’ rule:
Pr{|∗ }
() = for each ∈ (31)
Pr{|∗ }
If an ∈ off the equilibrium path is entered, then player deduces that ∗ has not been played — because
cannot be reached if players play ∗ — and Bayes’ rule cannot be applied at since Pr{|∗ } = 0. In this case there
is not an obvious way to determine , and next definition allows to choose arbitrarily. The adverb ”weakly”
51
appears in Definition 3.47 because the restrictions it imposes on beliefs are minimal, limited to information sets on
equilibrium path.
Definition 3.47 The system of beliefs is said to be weakly consistent with ∗ if and only if for each information
set along the equilibrium path, satisfies (31).
Thus, beliefs weakly consistent with ∗ are such that if is on the equilibrium path, then is the conditional
distribution on the nodes of given that play has reached and that players play ∗ . Conversely, if is off the
equilibrium path then Definition 3.47 leaves completely free.
Example 3.44(v) For the game of Example 3.44(i), consider = ( ). If were a NE, then we would
say that 2 is on the equilibrium path and (31) would imply 2 (1 ) = 01 = 0, 2 (2 ) = 11 = 1. For 31 ,
Definition 3.47 imposes no restriction since Pr{31 |} = 0, hence any 31 is weakly consistent with , for instance
31 (3 ) = 13 , 31 (4 ) = 13 , 31 (5 ) = 13 , or 31 (3 ) = 10
9 1
, 31 (4 ) = 10 , 31 (5 ) = 0, or ...
1 1 1 5 2
Now consider such that 1 = 3 + 2 + 6 , 2 = 7 + 7 , 3 = . Then 2 and 31 are on the equilibrium
path and (31) implies 2 (1 ) = 12 3 2 16
23 = 4 , (2 ) = 23 = 4 ;
1 31 514
(3 ) = 1321 = 15
26 ,
31 17
(4 ) = 1321 6
= 26 ,
542 5
31 (5 ) = 1321 = 26 .
Weak Sequential Equilibrium The notion of equilibrium we introduce here requires that ∗ is sequentially
rational for a system of beliefs which is weakly consistent with ∗ .
Definition 3.48 Given Γ, we say that (∗ ) is a Weakly Sequential Equilibrium (WSE) of Γ if and only if ∗ is
sequentially rational given , and is weakly consistent with ∗ .35
This definition summarizes the ideas introduced up to now. It requires that, for each player and each ∈ ,
∗ prescribes at an action that maximizes given . The beliefs are uniquely determined at information sets
on the equilibrium path, but they can be chosen freely off the equilibrium path.36
Example 3.42(ii) For the NE ∗ = ( ), can we find weakly consistent with ∗ such that (∗ ) is WSE? Since
Pr{2 |∗ } = 0, 2 can be chosen arbitrarily. We have that 2 2 ( ∗1 ) = (2 ), 2 2 ( ∗1 ) = 3(1 ) + 2(2 ),
and (2 ) 3(1 ) + 2(2 ) for each (1 ), (2 ); hence there exists no that makes ∗2 = sequentially rational
at 2 . As a consequence, for no the pair (∗ ) is a WSE. For the NE ̄ = ( ), we have that Pr{2 |̄} = 1 0,
hence (31) implies (2 ) = 1 and 2 2 ( ̄1 ) = 1 2 2 ( ̄1 ) = 2. Thus (̄ ) is a WSE if (2 ) = 1.
In a sense, the notion of WSE relies on the same principle as SPNE: it puts some discipline on the moves
the equilibrium strategies indicate off the equilibrium path by requiring that the strategies prescribe an optimal
action at each information set given beliefs. But the conditions imposed by Definition 3.37 apply only to subgames,
whereas sequential rationality applies to each information set.
Next proposition establishes a property of NE which is similar to Proposition 3.40: each NE ∗ is certainly
sequentially rational at an information set along the equilibrium path given weakly consistent beliefs.
Proposition 3.49 Given Γ, ∗ is a NE of Γ if and only if, for each information set on the equilibrium path, ∗
is sequentially rational in given the beliefs determined by (31).
This proposition implies that each WSE is a NE. Precisely, if (∗ ) is a WSE then ∗ is sequentially rational
at each given which is weakly consistent with ∗ , and in particular ∗ is sequentially rational at each on the
equilibrium path given determined by (31); then Proposition 3.49 applies to conclude that ∗ is a NE.
Since Proposition 3.49 establishes a double implication, it can also be used to shed light on the answer to the
following question: Given a NE ∗ of Γ, can we find such that (∗ ) is a WSE? In the jargon of game theory,
can ∗ extended to a WSE? Since Proposition 3.49 guarantees that ∗ is sequentially rational at each information
35 InMWG, this notion is called Weakly Perfect Bayesian Equilibrium.
36 Noticethat a WSE is not just a profile of strategies (like for NE and for SPNE), but is a pair which consists of a strategy profile
and a system of beliefs. That is, beliefs are part of the equilibrium.
52
set on the equilibrium path given determined by (31), it suffices to focus our attention to information sets off the
equilibrium path and inquire whether there exist beliefs which make ∗ sequentially rational at these information
sets.37 Although can be chosen freely at each off the equilibrium path, sometimes no exists such that ∗
is sequentially rational: see for instance ∗ = ( ) in the game of Example 3.42(i).
The notion of WSE works well in the game of Example 3.42(i), but in other games it is unsatisfactory: see next
example.
Example 3.50 For the following game we represent the extensive form and the normal form:
A B
x1 1 1\2
2 2 5 2 5
5 C D 2 5 2 5
0 2 3 1
x2 2 x3 1 2 4 3
h2
E F E F
0 3 1 4
2 1 2 3
In this game, ∗ = ( ) is a NE and (∗ ) is a WSE with such that 2 (2 ) = 1. In order to verify this,
notice that if play reaches 1 , then player 1 has no uncertainty about where he is playing, and is sequentially
rational given that 1 expects 2 to play at 2 . At 2 , beliefs such that 2 (2 ) = 1 are weakly consistent with
∗ and make it sequentially rational for 2 to play . However, ∗ is not a SPNE: In the subgame that starts with
node 1 , the unique NE is ( ), hence ∗ does not induce a NE in the subgame (backward induction shows that
( ) is the unique SPNE of this game).
This example shows that a WSE is not necessarily a SPNE. The reason is that Definition 3.47 does not impose
restrictions on beliefs off the equilibrium path and this freedom in some cases generates unpleasant WSE as in the
game of Example 3.50. More generally, this implies that the notion of WSE has a weak selective power. There is
a vast literature on the restrictions that can be imposed on beliefs off the equilibrium path. In particular, suitable
restrictions yield the notion of Sequential Equilibrium, SE, for which it is possible to prove that (i) each SE is a
SPNE; (ii) each SE is a WSE; (iii) in each game there exists at least one SE. Therefore, the relations among the
notions of NE, WSE, SPNE, SE are as follows:
37 Therefore, ∗ can certainly be extended to a WSE if each information set in the game is on the equilibrium path: (∗ ) is a WSE
with the beliefs in (31). For instance, this applies to the NE ̄ = ( ) in the game of Example 3.42(i), and guarantees that (̄ ) with
(2 ) = 1 is a WSE, even without the direct verification carried out in Example 3.42(ii).
53
NE
WSE
SE
SPNE
Strategy profiles
In next example we impose a restriction on beliefs based on a novel argument.
Example 3.51
1
A B C
2 1\2
2 x1 2 x2 2 2 2 2
h2 3 3 1 0
D E D E 0 0 1 1
3 1 0 1
3 0 0 1
In this game, ∗ = ( ) is a NE, and (∗ ) is a WSE if (1 ) ≤ 14 . But the following argument suggests that
(1 ) ≤ 14 is an unreasonable belief. Precisely, is strictly dominated by for player 1, hence it would be silly
for 1 to play . Thus, if player 2 is called upon to play, he should deduce that 1 (who should have played ), has
played , otherwise player 1 cannot do better than with . Hence, player 2’s beliefs should be such that (1 ) = 1
and 2 should play at 2 . Then it is profitable for 1 to deviate from and play , and therefore the considered
WSE seems not plausible.
The argument above relies on what is called forward induction reasoning: when a player has to move at an
information set which is off the equilibrium path, he should try to explain why this information set has been entered,
and his beliefs should reflect this.
4 Consumer Theory and General Equilibrium
4.1 The consumer’s problem
A consumer has a certain amount of money (his income, or wealth) 0 and there are ≥ 2 goods/services
available for purchase, each with a fixed unit price. The consumer has to decide how much he buys and consumes
of each good: This is called The consumer’s problem. Let denote the quantity the consumer buys of good , for
= 1 , and = (1 ) is the list of his purchases of the goods, often called the consumer’s consumption
bundle. Notice that is a vector with components, that is ∈ R , and it is useful to describe (briefly) the
operations of linear combination and inner product for vectors in R .
54
Two vectors in R , = (1 ) and = (1 ), can be combined linearly as follows: Given in R,
+ is the vector in R with components (1 + 1 + ).
Example 4.1 Given = (3 6 8), = (5 −2 3) in R3 and = −2, = 5, we have that the linear combination
+ is equal to −2 (3 6 8) + 5 (5 −2 3) = (19 −22 −1). If = 04, = 06, then + = (42 12 5).
A linear combination of two vectors with coefficients ∈ [0 1] and = 1 − is said to be a convex combination
of the vectors. Hence, given and in R , the convex combination of and with coefficients and 1 − , that
is + (1 − ), is the vector in R with component equal to + (1 − ) for = 1 . If and are in
R2 , then + (1 − ) lies on the segment (in the Cartesian plane) connecting to , and it is close to if is
close to 1, is close to if is close to 0.
Example 4.2 Let = (1 2), = (11 8). Then 08 + 02 = (3 3 2), 01 + 09 = (10 74), 05 + 05 = (6 5):
8 y
0.1x+0.9y
0.5x+0.5y
4
0.8x+0.2y
2
x
0
0 2 4 6 8 10 12
If and are in R with ≥ 3, then it is difficult or impossible to represent + (1 − ) graphically, but it is
still possible to compute + (1 − ): see for instance Example 4.1.
Definition 4.3 A set , subset of R , is said to be convex if it has the following property: For each in , each
convex combination of is in , that is + (1 − ) ∈ for each ∈ (0 1).
If is a subset of R2 , then is a convex set if and only if for each two points in the segment connecting
them lies entirely in .
Example 4.4 Let , , denote the circle, the quadrilateral, the pentagon in the Cartesian plane below,
respectively. Then and are convex sets, whereas is not.
B
A C
If is a subset of R with ≥ 3, then it is difficult or impossible to represent graphically, but it is still possible
to inquire whether is a convex set.
55
Example 4.5 The set R+ = { ∈ R : ≥ 0 for = 1 } is the set of vectors of R for which all components
are nonnegative, and R+ is a convex set: If and are in R+ , then also = + (1 − ) is in R+ for each
∈ (0 1) because the -th component of is + (1 − ) ≥ 0.
Given and in R , the inner product between and is denoted · and is defined as follows: · =
1 1 + + .
Example 4.6 Given = (3 6 8), = (5 −2 3) in R3 , the inner product between and is · = 15−12+24 = 27.
Typically, the possible consumption bundles are restricted by physical constraints, the most common of which
establishes the nonnegativity of consumption: for most goods, it is impossible to consume a negative quantity.
Therefore, we consider only consumption bundles such that ≥ 0 for = 1 , that is must be a vector in
R+ .
In addition to the constraint that is in R+ , the consumer needs to satisfy an economic constraint, that is he
can buy a bundle only if he can afford it. Precisely, suppose that for each good there exists a price 0 such
that each nonnegative quantity of good can be purchased at the unit price , and the consumer takes as
given: he is price taker. Then, if he buys the bundle , the consumer spends for good and his total expense
is 1 1 + + . It is useful to gather the prices in the vector = (1 ) ∈ R , because then the consumer’s
total outlay can be written as the inner product · . The consumer has a certain amount of money 0, thus
he can afford to buy if and only if costs no more than , that is if 1 1 + + ≤ , or · ≤ .
The arguments above establish that the set of the bundles the consumer can purchase is { ∈ R+ : · ≤ },
called budget set and denoted with . In the case of = 2, = {(1 2 ) ∈ R2 : 1 ≥ 0, 2 ≥ 0, 1 1 + 2 2 ≤ }
is a triangle with vertices given by the points with coordinates (0 0), ( 1 0), (0 2 ) (highlighted in the figure
below):
x2
w/p2
w/p1 x1
It is immediate that (i) an increase (decrease) of moves ’s diagonal edge outwards (inwards); (ii) an increase
(decrease) of 1 moves leftward (rightward) the vertex with coordinates ( 1 0); (iii) an increase (decrease) of 2
moves downward (upward) the vertex with coordinates (0 2 ); (iv) the triangle’s hypotenuse has slope − 12 . It is
also immediate to see that is a convex set when = 2, but in fact it is possible to prove that is convex for
each ≥ 2: If and are in , then = + (1 − ) is in for each ∈ (0 1) since (i) = + (1 − ) ≥ 0
as 0, ≥ 0, 1 − 0, ≥ 0; (ii) · = · ( + (1 − )) = · + (1 − ) · ≤ + (1 − ) = .
The consumer’s problem can be expressed as follows: Given which determine , what is the bundle in
the consumer buys?
4.1.1 Preference relations and utility functions
The set is the set of bundles the consumer can afford to buy, therefore we expect he will select the bundle in
he likes most. As in other cases, we assume that his tastes are described by a preference relation % on R+ which
56
is complete and transitive. From % it is possible to derive the strict preference relation  and the indifference
relation ∼ (see Section 1 in this document); these can be used to identify the consumer’s preferred bundle in .
However, when possible, it is useful to represent % through a utility function : R+ → R: see Definition 1.3. The
next example describes one case in which this is not possible.
Example 4.7 Consider the following preferences on R2+ :
 if and only if 1 1 or (1 = 1 and 2 2 )
(32)
∼ if and only if 1 = 1 and 2 = 2
The preferences described by (32) are such that two bundles = (1 2 ) and = (1 2 ) are compared first
according to the quantity of good 1. If 1 6= 1 , then the bundle with the higher quantity of good 1 is preferred
to the other bundle. If 1 = 1 , then the comparison between 2 and 2 determines the preference. Hence, the
quantity of good 1 has a higher priority in determining the preference, as a word’s first letter has a higher priority
in determining the word position in a dictionary. For this reason, these preferences are called lexicographic. For
instance, if = (3 5), = (2 9), = (3 7), = (4 2) then  ,  ,  . Even though they are simple to
describe, it is possible to prove that there exists no utility function which represents these preferences.38
Now we provide a sufficient condition on % which guarantees the existence of a utility function for %. To
this purpose, given = (1 ) and = (1 ) in R , let ( ) denote the Euclidean distance between
p
and , that is ( ) = (1 − 1 )2 + + ( − )2 . Then, given ∈ R+ and 0, let () = { ∈
R+ : ( ) } be the set of vectors in R+ which are less than away from . Typically, () is called
neighborhood of with radius . Actually, often there is no need to specify and then we write just () to
denote a neighborhood of with unspecified positive radius. A sufficient condition for the existence of a utility
function is as follows.
• Continuity Given % on R+ , we say that % is continuous if for each in R+ such that  , there exist
() and () such that each vector in () is preferred to each vector in ().
This property says that preferences do not change suddenly: If  and is close to , is close to , then
 .
Proposition 4.8 If a preference relation % on R+ is complete, transitive, and continuous, then there exists a
utility function which represents %; furthermore, is continuous.
This proposition establishes that continuity of % implies that a utility function for % exists. Conversely, if % is
not continuous then a utility function may exist or not exist. Of course, the lexicographic preference relation is not
continuous otherwise a utility function for it would exist. The proof of this fact is straightforward, as it suffices to
provide a counterexample: (i) consider = (3 5) and = (3 1), hence  ; (ii) take any two neighborhoods (),
(); (iii) pick in () the bundle (3 − 5) with a small 0, and pick in (); (iv) verify that (3 − 5) Â
is untrue, hence continuity is violated.
From now on we consider only preference relations that are complete, transitive, continuous. Therefore, for
each considered % there exists a continuous utility function : R+ → R such that  if and only if () (),
∼ if and only if () = (). Sometimes, it is useful to assume that is partially differentiable with respect to
each variable, and ()
denotes the partial derivative of with respect to the variable , for = 1 ;
39
()
38 In view of a contradiction, suppose that there exists that represents %. Then for each 1 ≥ 0 we have that (1 1) ≺ (1 2),
hence (1 1) (1 2). Therefore there exists a rational number (1 ) such that (1 1) (1 ) (1 2), and 01 001 implies
(01 ) (00 0 0 0 00 00 00
1 ) because (1 1) (1 ) (1 2) (1 1) (1 ) (1 2) Thus is strictly increasing, and therefore it is
one-to-one. Hence establishes a one-to-one correspondence between the set [0 +∞) and a subset of Q. This is impossible since Q is
a countable set, and so is any subset of Q, whereas [0 +∞) is a non-countable set. As a consequence, no that represents % exists.
39 Assuming that is partially differentiable is a further restriction on preferences in addition to continuity: Some continuous preference
relations cannot be represented by a partially differentiable utility function. See for instance the preference relation represented by the
function in Example 4.12 below.
57
is the vector of the partial derivatives of evaluated at , called gradient of : () = ( () ()
1 ).
Example 4.9 Consider : R2+ → R such that (1 2 ) = 71 − 452 + 231 22 . Then () 2 2 ()
1 = 7 + 61 2 , 2 =
−202 +41 2 , gathered in () = (7+61 2 −202 +41 2 ). If : R+ → R is (1 2 3 ) = 61 2 −(1+3 )42 ,
4 3 2 2 4 3 3 2
then () () 2
1 = 121 2 , 2 = 61 − 4(1 + 3 )
42 ()
, 3 = −42 .
Partial derivatives are useful to obtain informations on how changes when 1 change. In particular,
()
1. If 0 for each ∈ R+ , then is strictly increasing with respect to : When is increased, the value
()
of increases. If 0 for each ∈ R+ , then is strictly decreasing with respect to .
2. If partial derivatives exist, are continuous, and are in R+ with close to , then () − () is well
approximated by the inner product () · ( − ), that is
() − () ' () · ( − ) if is close to (33)
Hence, a small change in the variables from to increases if () · ( − ) 0, decreases if () ·
( − ) 0.
Example 4.10 For (1 2 ) = 71 − 452 + 231 22 we have that () 2
1 0 for each ∈ R+ , hence is strictly
increasing with respect to 1 .
Consider = (3 1), = (3 − 2 1 + ) for close to 0. Then − = (−2 ) and () ()
1 = 61, 2 = 88. Hence
() − () is about equal to 61 · (−2) + 88 · = −34. This reveals that a change in the variables from to
with a small 0 ( 0) reduces (increases) .
4.1.2 Monotonicity, Convexity
Solving the consumer’s problem requires to identify the consumer’s preferred consumption bundle in = { ∈
R+ : · ≤ }. Hence, we are interested in ∗ in such that ∗ % for each ∈ , but if there exists which
represents %, then an optimal bundle is a max point for in , that is ∗ satisfies ∗ ∈ and (∗ ) ≥ () for
each ∈ . Sometimes, this problem is written as
max () or max () such that · ≤ (34)
∈ ≥0
The assumptions made up to now do not yield any meaningful results, hence now we introduce some additional
assumptions.
Given and in R+ , the notation ≥ means that ≥ for = 1 , that is for each good , bundle
has a weakly higher quantity of good than bundle .
• Monotonicity Given a preference relation % on R+ , we say that % is monotone if ( ≥ and 6= ) implies
 .
The interpretation of this property is that more is better ; that is, larger amounts of goods are preferred to
smaller ones. For the case of = 2, if we fix and take such that 1 1 , 2 2 , or 1 1 , 2 = 2 , or
1 = 1 , 2 2 , then monotonicity implies ≺ .40 Conversely, if 1 1 and 2 2 (of if 1 1 and 2 2 )
then monotonicity does not imply anything. Given , we define three sets as follows:
() = { ∈ R+ : ∼ } () = { ∈ R+ : % } () = { ∈ R+ : % }
Hence, () is the set of bundles the consumer views exactly as good as (the indifference set of ); () (the
upper contour set of ) is the set of bundles which are at least as good as ; () (the lower contour set of )
40 Likewise, if 1 1 , 2 2 , or 1 1 , 2 = 2 , or 1 = 1 , 2 2 , then monotonicity implies ≺ .
58
is the set of bundles which are no better than . Notice that both () and () include (). If = 2 and
% is monotone, then () is a strictly decreasing curve called indifference curve through ; () consists of the
points which are in () or above, and () is the set of points which are in () or below.
Example 4.11 After fixing a particular ∗ , we provide below four examples of indifference curves through ∗ for
monotone preferences:
x2 x2
x* x*
x1 x1
(a) (b)
x2 x2
x* x*
x1 x1
(c) (d)
For the case of ≥ 3, the interpretation of monotonicity is similar to the case of = 2: a bundle is preferred
to if for each good the quantity in is at least as great as the quantity in , and strictly greater for at least one
good.
Monotonicity can be expressed in terms of as follows: ( ≥ and 6= ) implies () (), that is is
strictly increasing with respect to each variable. If is partially differentiable in R+ , then the condition
()
0 at each ∈ R+ , for = 1 (35)
implies that is strictly increasing with respect to each variable, hence the preference relation represented by is
monotone.
In the following we will often consider preference relations which are monotone, but many of the results for
consumer theory can be obtained under the following weaker condition.
• Local Nonsatiation Given a preference relation % on R+ , we say that % is locally nonsatiated if for each
∈ R+ and for each () there exists at least a in () such that  .
59
This property says that for any given and any neighborhood of , there exists at least one in the neighborhood
that the consumer prefers to . Loosely speaking, for each there exist bundles which are superior to and are
close to . Monotonicity implies local nonsatiation, that is if % is monotone then % is locally nonsatiated.41 The
proof is as follows: Given , () and % monotone, consider such that 1 = 1 + and 2 = 2 , ..., = ,
with 0 and small. Then belongs to () since is small, and  since % is monotone.
Example 4.12 Consider a two-good setting, and the following utility functions: (1 2 ) = (1 + 2)(2 + 1),
(1 2 ) = min{1 2 }.
It turns out that is strictly increasing with respect to 1 and with respect to 2 since ()
1 = 2 + 1 0,
()
2 = 1 + 2 0, hence the preference relation represented by is monotone (and also locally nonsatiated).
The function does not represent monotone preferences: Consider for instance = (7 4) and = (9 4), which
satisfy ≥ and 6= , but () = 4 is not greater than () = 4. However, for each = (1 2 ) and each
() there exists a in () such that () (): = (1 + 2 + ) with a small 0 belongs to () and
() = min{1 + 2 + } = min{1 2 } + = () + ().
Now we introduce two assumptions which formalize the notion that the consumer loves variety.
• Convexity Given a preference relation % on R+ , we say that % is convex if
( % , % ) implies + (1 − ) % for each ∈ (0 1) (36)
• Strict Convexity Given a preference relation % on R+ , we say that % is strictly convex if
( % , % and 6= ) implies + (1 − ) Â for each ∈ (0 1)
At least two interpretations are possible for the condition in (36). First, (36) states that if and are both
at least as good as , then each convex combination of and is at least as good as . This means that the set
() is a convex set for each . Strict convexity is a stronger property as it implies that any convex combination
of and (if 6= ) is strictly better than .42
Another interpretation requires the assumption that = 2 and % is monotone. Then each indifference curve
is a decreasing curve, and given an arbitrary , consider and in () such that 1 1 , 2 2 and 1 1 ,
2 2 . Then any convex combination of is a more balanced combination of goods than the extreme bundles
and , and convexity of preferences implies that the convex combination is not worse than and . This means that
more variety is at least as good for the consumers as the extreme bundles, and for this reason convexity expresses
a consumer’s inclination for diversification.
Example 4.13 Here we examine whether , in Example 4.12 represent convex/strictly convex preferences.
To this purpose, notice that an indifference curve for a preference relation % is a level curve of the utility function
representing %. Thus we can use to derive indifference curves by fixing an arbitrary ≥ 2 (because () ≥ 2 for
each ∈ R2+ ) and solving the equation () = with respect to 2 . We find 2 = 1+2 − 1, and examining the
set of points on the curve 2 = 1+2 − 1 or above it, we conclude that represents a strictly convex preference
relation. As varies, we obtain a family of indifference curves: see figure (a) below.
Conversely, does not represent a strictly convex preference relation: Consider = (7 4), = (9 4), = (7 4),
which satisfy % , % and 6= . But the bundle 12 + 12 = (8 4) is a convex combination of , yet it is not
preferred to since ( 12 + 12 ) = 4, () = 4. The indifference curves for (see figure (b)) reveal that represents
41 However, the reverse implication is untrue: If % is locally nonsatiated, then % is not necessarily monotone: See Example 4.12.
42 Therefore, with reference to Example 4.11, it is immediate to conclude that a preference relation is not convex (and is not strictly
convex) if (∗ ) is the curve in figure (b) or in figure (d).
60
a convex preference relation.
x2 x2
10 10
8 8
v(
x)
=7
6 6
v(
x)
=
u(
5
x)
4 4
u(
=3
x)
0
=2
v(
u(
x)
0
x
2 2
=2
)=
u(
10
x
)=
5
2 4 6 8 10 x1 2 4 6 8 10 x1
(a): indifference curves for (b): indifference curves for
4.1.3 Utility maximization
In this section we derive some results for the consumer’s problem in (34). First we notice that the budget set is
closed and bounded,43 that is compact. Then, given a continuous , we can apply the Extreme Value Theorem to
guarantee that an optimal bundle exists: that is there exists at least an ∗ ∈ such that (∗ ) ≥ () for each
∈ .
For the case of = 2, if % is monotone then each optimal bundle lies on the highest indifference curve that
intersects : see figures (a) and (b) below, and notice that two optimal bundles exist for the case described by
figure (b).44
x2 x2
x'
x*
x''
x1 x1
(a): ∗ is the unique optimal bundle (b): 0 and 00 are both optimal bundles
The assumptions introduced in the previous subsection allow to draw some general conclusions about optimal
bundles.
Proposition 4.14 Suppose that is continuous. Then
(a) if % is locally nonsatiated, then each optimal bundle satisfies · = ;
(b) if % is strictly convex, then there exists a unique optimal bundle;
43 Since includes its boundary and there exists a number 0 such that for each ∈ , 0 ≤ ≤ for = 1 .
44 Indeed, the Extreme Value Theorem guarantees the existence of at least one optimal bundle, but does not imply that the optimal
bundle is unique: Multiple optimal bundles may exist, and in this case the consumer is indifferent among all optimal bundles.
61
(c) if and are multiplied by the same 0, then the set of optimal bundles does not change;
(d) if % is locally nonsatiated and strictly convex, then the optimal bundle is a continuous function of .
Proof (a) In view of a contradiction, suppose that ∗ is an optimal bundle such that · ∗ . Then there exists
a small (∗ ) such that (∗ ) ⊆ , and local nonsatiation implies that (∗ ) includes at least one such that
 ∗ . Thus is in and is superior to ∗ ; this contradicts the assumption that ∗ is optimal.
(b) In view of a contradiction, suppose that there exist two optimal bundles and , such that 6= . Then ∼ ,
hence % , % (put = in the definition of strict convexity) and consider 12 + 12 . It is immediate that
( 12 + 12 ) ∈ since is a convex set, and strict convexity of preferences implies ( 12 + 12 ) Â ∼ ; this contradicts
the assumption that and are optimal.
(c) This claim is obvious because remains the same set for each 0, thus the set of optimal bundles remains
the same. ¥
Proposition 4.14(a,b) describes some properties of the solution(s) to the consumer’s problem, and part (b) is
quite important as it establishes that for strictly convex preferences a unique optimal bundle exists. We denote it
with ( ) since it depends on .
Next figure represents the optimal bundle, ∗ , given 1 2 which generate the thin budget line (the indifference
curve through ∗ is the thin curve). Then suppose that the price of good 2 increases to 02 2 . The new budget
line is the thick segment, and the new optimal bundle is 0 (the indifference curve through 0 is the thick curve).
Hence, (1 2 ) = ∗ and (1 02 ) = 0 .
x2
x*
x'
x1
Once the function ( ) is available, it is possible to move to Subsection 4.2 of this document, which is about
General Equilibrium. Before that, however, it is useful to describe some techniques to solve the consumer’s problem.
Identifying the optimal bundles when = 2 Suppose that % is locally nonsatiated. Then we know from
Proposition 4.14(a) that each optimal bundle satisfies 1 1 + 2 2 = , and we can solve this equation with respect
to 2 to obtain 2 = 2 − 12 1 . This allows to write the consumer’s utility as (1 2 − 12 1 ), which we denote
as ̂(1 ); ̂ is a function of the single variable 1 and is defined for 1 ∈ [0 1 ]. Maximizing ̂ with respect to 1
yields ∗1 , hence ∗2 = 2 − 12 ∗1 .
Example 4.15 Consider (1 2 ) = 1 + 2 (notice that represents a monotone, hence lns, preference relation),
and 1 = 4, 2 = 6, = 48. Then 2 = 8 − 23 1 and ̂(1 ) = 8 + 13 1 , defined for 1 ∈ [0 12]. Since ̂ is strictly
increasing, it follows that ∗1 = 12 and ∗2 = 0.
Consider still (1 2 ) = 1 + 2 , 1 = 4, = 48, but 2 = 3. Then 2 = 16 − 43 1 and ̂(1 ) = 16 − 13 1 , defined
for 1 ∈ [0 12]. Since ̂ is strictly decreasing, it follows that ∗1 = 0 and ∗2 = 16.
62
Consider still (1 2 ) = 1 + 2 , 1 = 4, = 48, but 2 = 4. Then 2 = 12 − 1 and ̂(1 ) = 12 for each
1 ∈ [0 12]. Since ̂ is a constant function, it follows that each 1 in [0 12] is a max point for ̂; hence, any pair
(1 2 ) is optimal if 1 ∈ [0 12] and 2 = 12 − 1 .45
Example 4.16 Consider (1 2 ) = 1 2 (notice that represents a lns preference relation) and generic values
for 1 2 . Using 2 = 2 − 12 1 we obtain ̂(1 ) = 1 ( 2 − 12 1 ), defined for 1 ∈ [0 1 ]. The unique max
point for ̂ is ∗1 = ∗
21 , hence 2 = 22 . Arguing likewise, it is possible to prove that if (1 2 ) = 1 2 with
0, 0, then ∗1 = + ∗
1 , 2 = + 2 . The utility function (1 2 ) = 1 2 is called Cobb-Douglas utility
function.
A result on optimal bundles for each ≥ 2 Now we allow for an arbitrary number ≥ 2 of goods. Given
an arbitrary natural number ≥ 2, we may assume that % is locally nonsatiated and then argue as for = 2.
That requires to solve · = with respect to , thus = 1 ( − 1 1 − − −1 −1 ), to express the
consumer’s utility as a function of − 1 variables. Unfortunately, in general this is not very useful when ≥ 3.
It is more convenient and insightful to assume that is partially differentiable with respect to each variable (with
continuous partial derivatives) and that (35) holds. This implies monotone preferences and allows to use calculus
to characterize each optimal bundle.
As a first step, we introduce the marginal rate of substitution between two goods: Given ∗ ∈ R+ (∗ is not
(∗ )
necessarily an optimal bundle) and given and , we define MRS (∗ ) as (∗ ) and now we provide an
interpretation. Starting from ∗ , modify the quantity of good by ∆ and the quantity of good by ∆ , both
∗
(∗ ) (∗ )
close to 0. Then, by (33), ∆ is about equal to ( )
∆ + ∆ and choose ∆ such that ∆ +
(∗ )
∆ = 0. Hence
(∗ )
∆ = − ∆ or ∆ = − (∗ )∆ (37)
(∗ )
More in detail, given the consumer’s initial bundle ∗ , reducing by one the quantity of good reduces the consumer’s
utility, and we inquire how many additional units of good he should receive in order to have the same utility
as with ∗ . Inserting ∆ = −1 in (37) reveals that the answer is ∆ = (∗ ), hence (∗ ) is the
amount of good the consumer must receive to be compensated after a unit reduction in good . In general, for
each ∆ close to 0, ∆ = − (∗ )∆ yields a bundle the consumer views as equally good as ∗ . Formally,
(∗1 ∗ ∗ ∗ ) ∼ (∗1 ∗ + ∆ ∗ − (∗ )∆ ∗ )
Hence, (∗ ) is the rate at which, starting from ∗ , goods and can be traded off in a way which does not
change the consumer’s utility. In the case of = 2 it is possible to prove that the slope at ∗ of the indifference
curve through ∗ is equal to − 12 (∗ ). This suggests that if % is strictly convex, then 12 () decreases as
moves on an indifference curve from north west to south east.
√ √
Example 4.17 Consider (1 2 ) = 4 1 + 10 2 and ∗ = (100 64). Then () = ( √21 √52 ), (∗ ) =
( 15 58 ), 12 (∗ ) = 25
8
. Therefore, starting from ∗ , a unit reduction in 1 (∆1 = −1) is compensated by
∆2 = 825; ∆1 = 4 (∆1 = − 12 ) is compensated by ∆2 = −225 (by ∆2 = 425).
1
Proposition 4.18 Suppose that is partially differentiable with respect to each variable, with continuous partial
derivatives; satisfies (35) and ∗ is a solution to (34). Then ∗ is such that · ∗ = and
(∗ ) ∗
(i) for any two goods such that ∗ 0, ∗ 0, we have = () , that is (∗ ) = ;
(∗ ) (∗ )
(ii) for any two goods such that ∗ 0, ∗ = 0 we have ≥ , that is (∗ ) ≥ .
45 In this case infinitely many optimal bundles exist. This does not conflict with Proposition 4.14(b), as the preference relation
represented by (1 2 ) = 1 + 2 is not strictly convex.
63
Proof (i) Assuming that ∗ 0, ∗ 0, we prove that if (∗ ) 6= then ∗ is not optimal. For instance, if
∗
( )
then consider modifying the purchases of goods and (only) as follows: ∆ = −, ∆ =
(∗ )
with 0 and small. The new bundle has a lower quantity of good , a higher quantity of good and is in
∗
(∗ )
as costs just like ∗ . Moreover, the consumer’s utility changes by () − (∗ ) ' ( )
(−) + =
(∗ ) (∗ ) (∗ )
( − (∗ ) ) 0, thus () (∗ ); this means that ∗ is not optimal. If instead (∗ ) ,
then a bundle in which is superior to ∗ is obtained by choosing ∆ = , ∆ = − 0.
(∗ )
(ii) Now consider such that ∗ 0, ∗ = 0. We prove that if (∗ ) , then ∗ is not optimal; hence
∗ ∗
( ) ( )
(∗ ) ≥ . The proof is just as the proof of part (i): If (∗ ) , then
can be increased by modifying
the consumer’s bundle with ∆ = −, ∆ = with 0 and small. ¥
The proof of Proposition 4.18 (i) shows that (∗ ) = by proving that if (∗ ) 6= , then it is
possible to find an in which is superior to ∗ by modifying the purchases of goods and . In particular, if
(∗ ) then it is convenient to slightly reduce and slightly increase in such a way to leave the total
expense unchanged, because then is increased at the rate which is higher than the rate (∗ ) which is
necessary to leave the consumer indifferent. The opposite modification, with ∆ 0 and ∆ 0, can be applied
if (∗ ) . But notice that these two arguments require ∗ 0, ∗ 0. In case that ∗ 0, ∗ = 0,
we cannot use an argument which relies on ∆ 0, hence we cannot rule out that (∗ ) . Indeed,
Proposition 4.18 (ii) establishes that (∗ ) ≥ when ∗ 0, ∗ = 0.
Proposition 4.18 applies also when = 2. In this case, if ∗ is such that ∗1 0, ∗2 0, then 12 (∗ ) = 12
and the slope of (∗ ) at ∗ , which is − 12 (∗ ), is equal to slope of the hypotenuse of the budget set, − 12 .
This means that (∗ ) intersects and is tangent to the hypotenuse at ∗ .
The conditions in Proposition 4.18(i-ii) are necessary (sometimes they are called necessary first order conditions
because they involve first order partial derivatives) but are not sufficient. However, if there exists a unique bundle,
∗ , which satisfies these conditions, then we can conclude that ∗ is the unique optimal bundle by the virtue of the
Extreme Value Theorem.
Example 4.19 Consider a three-good setting with (1 2 3 ) = (1 + 1 )(1 + 2 )(1 + 3 ), 1 = 2, 2 = 6, 3 = 10,
and = 8. First we inquire whether there exist bundles such that 1 0, 2 0, 3 0 which satisfy the
conditions in Proposition 4.18(i-ii). If that is the case, then
()1 ()3 ()2 ()3
= and =
2 10 6 10
that is
(1 + 2 )(1 + 3 ) (1 + 1 )(1 + 2 ) (1 + 1 )(1 + 3 ) (1 + 1 )(1 + 2 )
= and =
2 10 6 10
hence
5 2
1 = 53 + 4, 2 =
3 +
3 3
Since · = 8 (because represents a monotone preference relation), it follows that
5 2
2(53 + 4) + 6( 3 + ) + 103 = 8
3 3
2
and the unique solution to this equation is 3 = − 15 , which violates 3 0. Hence, there exists no bundle such
that 1 0, 2 0, 3 0 and which satisfies the conditions in Proposition 4.18(i-ii).
Now we inquire whether there exist bundles such that 1 0, 2 0, 3 = 0 which satisfy the conditions in
Proposition 4.18(i-ii). If that is the case, then
()1 ()2 ()1 ()3
= and ≥
2 6 2 10
that is
1 + 2 1 + 1 1 + 2 (1 + 1 )(1 + 2 )
= and ≥
2 6 2 10
64
hence
1 2
2 = 1 −
3 3
Since · = 8, it follows that 21 + 6( 13 1 − 23 ) = 8, thus 1 = 3 and 2 = 13 , both positive numbers. Finally, the
inequality 1+
2
2
≥ (1+110
)(1+2 )
reduces to 23 ≥ 15 8
, which is satisfied. Thus = (3 13 0) satisfies the conditions in
Proposition 4.18(i-ii).
We should then verify whether there exist other bundles which satisfy the conditions in Proposition 4.18(i-ii),
considering the cases of
1 0, 2 = 0, 3 0;
1 = 0, 2 0, 3 0;
1 0, 2 = 0, 3 = 0;
1 = 0, 2 0, 3 = 0;
1 = 0, 2 = 0, 3 0.
It turns out that no other bundle satisfies such conditions, hence = (3 13 0) is the unique optimal bundle for this
consumer.
An alternative proof of Proposition 4.18 can be obtained (maybe more quickly) by using the Kuhn-Tucker
theorem. The Lagrangian function for problem (34) is
( ) = () + ( − · )
with the constraints 1 ≥ 0, ..., ≥ 0. The Kuhn-Tucker theorem reveals46 that if ∗ is an optimal bundle, then
there exists ∗ ≥ 0 such that
(∗ ∗ )
≤ 0, with equality if ∗ 0, for = 1 ; = · ∗
We rewrite these conditions as = · ∗ and
(∗ ) (∗ )
− ∗ 1 ≤ 0, = 0 if ∗1 0; ...; − ∗ ≤ 0, = 0 if ∗ 0
1
(∗ ) ∗
If ∗ 0, ∗ 0, then (∗ ) = ∗ , (∗ ) = ∗ , hence = ∗ = () , as stated
(∗ )
by Proposition 4.18(i). If ∗ 0, ∗ = 0, then (∗ ) = ∗ , (∗ ) ≤ ∗ , hence = ∗ ≥
(∗ )
, as stated by Proposition 4.18(ii).
4.2 General equilibrium for pure exchange economies
General equilibrium views the economy as a set of interdependent markets in which several agents operate, and
all the variables of interest are determined simultaneously: Prices, consumption levels and production levels. In
fact, we consider an economy with no production opportunities, in which the only activities consist of trading and
consumption. The reason is that allowing for production complicates the model but yields analogous results.
In the economy there are ≥ 2 consumers and ≥ 2 goods, and each consumer initially owns some goods
called (initial) endowment: Each consumer owns quantity 1 ≥ 0 of good 1, 2 ≥ 0 of good 2, ..., ≥ 0 of
P
good , and the vector = ( 1 2 ) describes the entire endowment of consumer . The sum ∈ is
denoted with ̄; hence ̄ = (̄ 1 ̄ 2 ̄ ) is the vector of total endowments available in the economy: ̄ 1 0 is
the total quantity of good 1, ̄ 2 0 is the total quantity of good 2, ..., ̄ 0 is the total quantity of good . For
a generic vector , we write 0 to mean that each component of is positive. We suppose that ̄ 0, that
is in the economy there exists a positive quantity of each of the goods.
46 This theorem can be applied because the qualification condition on the constraints is satisfied, given that 0 for = 1 .
65
Example 4.20(i) Let = 2, = 3 and 1 = (7 15 0), 2 = (4 3 13). Therefore, this economy consists of
two consumers and three goods, such that initially consumer 1 owns a positive quantity of good 1 and good 2,
owns zero quantity of good 3. Consumer 2 owns a positive quantity of each good. The total endowment vector is
̄ = (11 18 13).
Consumers can exchange goods starting from their endowments. One exchange between consumer 1 and con-
sumer 2 in the context of Example 4.20(i) could be such that 1 gives to 2 some quantity of good 2 and receives
in exchange from 2 some quantities of good 1 and of good 3. The model to be introduced inquires the out-
come of these trades, that is the consumption of each consumer after the trading occurred. This model is called
pure exchange economy because the goods consumers eventually consume in aggregate are those which are in the
economy since the beginning.
With we denote the consumption bundle of consumer , that is = (1 2 ), in which ≥ 0 is the
quantity of good consumed by , for = 1 ; hence ∈ R+ .47
Definition 4.21 An allocation x = (1 2 ) ∈ R+ × R+ × × R+ is a list of consumption bundles, one for
each consumer, such that 1 + 2 + + = ̄.
An allocation specifies a bundle for each consumer, that is an allocation can be seen as a way to distribute the
available goods among consumers.
Example 4.20(ii) In the setting of Example 4.20(i), 1 = (3 16 4), 2 = (8 2 9) is an allocation because 1 ∈ R3+ ,
2 ∈ R3+ , and 1 + 2 = ̄. Another allocation is 1 = (0 9 8), 2 = (11 9 5).
The simplest exchange economy is such that there are two consumers and two goods: = 2, = 2. Then
an allocation is x = (1 2 ) with 1 = (11 21 ), 2 = (12 22 ) such that 11 + 12 = ̄ 1 and 21 + 22 = ̄ 2 .
Each allocation for this economy can be represented in the so-called Edgeworth Box (EB), which is the rectangle
[0 ̄ 1 ] × [0 ̄ 2 ] in the Cartesian plane’s first quadrant, in which the horizontal axis refers to the quantity of good 1
consumed by consumer 1, the vertical axis refers to the quantity of good 2 consumed by consumer 1; the origin of
these axes is denoted 1 . There is a one-to-one correspondence between allocations and points in the EB: for each
point in the EB, the point’s first coordinate is 11 , the point’s second coordinate is 21 , and these coordinates also
identify the consumption bundle of consumer 2, as 12 = ̄ 1 − 11 and 22 = ̄ 2 − 21 . Therefore, in addition to
the Cartesian axes for the quantities consumed by consumer 1, it is useful to have two more Cartesian axes, with
origin at point (̄ 1 ̄ 2 ), for the quantities consumed by consumer 2; the origin for these axes is denoted 2 . For
example, see Figure 15.B.1 in MWG, in which is the initial endowment and another allocation is represented
(indicated with ), which is reached, starting from , if consumer 1 gives up some quantity of good 2 in exchange
for some quantity of good 1 (hence, consumer 2 gives up some quantity of good 1 in exchange for some quantity of
good 2).
This model of pure exchange economy addresses the following question: Given that the goods in ̄ are in the
economy, and given the way they are initially distributed among consumers, what are the final consumption bundles
for consumers?
4.2.1 Definition of Walrasian Equilibrium
In order to answer the question above, we need to specify a mechanism for the distribution of goods. In fact,
there exist several mechanisms which may determine the final allocation. One is barter, which means that con-
sumers bargain until they reach an agreement. Another possibility is that consumers vote over several alternative
allocations, until a winning allocation is determined. One further possibility is dictatorship: consumers appoint a
dictator (the dictator himself may be one consumer), and he determines the final allocation. We do not consider
any of these mechanisms, but rather we assume that consumers can exchange the goods they initially own through
47 Notice that this notation is different from the one in Subsection 4.1, in which is the quantity of good purchased by the (unique)
consumer considered in Subsection 4.1.
66
a system of markets, essentially because this tries to mimic what goes on in reality, although some features of the
considered markets are not realistic. Precisely, for each good there exists a market and a going price 0 in
that market. Each consumer can buy or sell each quantity of good at the unit price ,48 and each consumer is
price taker, that is he considers as independent of his actions. Because of this assumption, the economy is said
to be competitive.
Given = (1 ), each consumer sells the own endowment and receives the amount of money · =
1 1 + 2 2 + + , which is analogous to in Subsection 4.1. But notice that here a consumer’s wealth
is not exogenous: it is given by the market value of his endowment given . With this wealth, he can buy each
bundle which is in the set
= { ∈ R+ : · ≤ · }
When = 2, we know from Subsection 4.1 that is a triangle. When = 2, 1 and 2 can be represented as
follows in the EB. Consider the endowment point in the EB, and the line through with slope −1 2 . Then
1 is given by the set of points which are below or on this line, and are in the first quadrant of the Cartesian plane
with axes 11 and 21 (and origin 1 ); 2 is given by the set of points which are above or on this line, and are
in the first quadrant of the Cartesian plane with axes 12 and 22 (and origin 2 ): See Figure 15.B.2 in MWG.
Notice that 1 is typically not a subset of the EB, but it includes some points outside the EB: For instance, in
Figure 15.B.2 in MWG it includes points in which 11 is greater than ̄ 1 . Likewise, 2 includes points such that
12 ̄ 1 .
For each consumer , his tastes are represented by a preference relation % on R+ , as in Subsection 4.1, which
is complete, transitive, continuous, monotone, and strictly convex. When = 2, = 2, the preference relation of
each of the two consumers can be represented in the EB through indifference curves: See Figure 15.B.3 in MWG.
Since % is complete, transitive, continuous, there exists a continuous utility function that represents % : See
Proposition 4.8. In this context, consumer is described by two objects: % (or ) and ; the whole economy is
described by {% }∈ , that is the preferences and the endowment of each consumer.
Each consumer selects his preferred bundle in , which is the solution to the problem
max ( ) subject to · ≤ · (38)
∈R
+
Notice that (38) coincides with (34), after setting equal to · . The unique optimal bundle for consumer is
a function of prices and is denoted (),49 that is () = (1 () 2 () ()) is the bundle demanded by
consumer given , and () is the quantity of good demanded by .
Figure 4.22 represents an economy with = 2, = 2, in which the price vector (1 2 ) determines the slope
of the budget line, − 12 , and the budget set of consumer 1, 1 , is the triangle with vertices 1 . The optimal
bundle for consumer 1, 1 (), is the point in 1 on the highest indifference curve of consumer 1 which intersects
1 ; such indifference curve is denoted 1 . The coordinates of 1 (), 11 () and 21 (), are read on the axes which
48 Infact, since is the quantity consumer initially owns of good , he cannot sell more than quantity of good .
49 Since% is strictly convex, there exists a unique optimal bundle for consumer : See Proposition 4.14(b). This bundle depends on
and also on % ; but % are considered as exogenous, not explained by this model.
67
have origin in 1 :
O2
N
x1(p)
x21(p)
Figure 4.22 : I1
B1
O1 x11(p) M
P
The sum ∈ () is the total demand for good in the economy, that is the sum of demands of all consumers
P
in market . The supply in market is just the total endowment of good in the economy, that is ̄ = ∈ ,
because ̄ is the total quantity of good consumers put on sale, and is independent of . We say that the market
P
of good clears if ∈ () = ̄ , hence all markets clear if and only if
X
() = ̄ for = 1
∈
To summarize, in this model each consumer (i) takes as given a vector of prices; (ii) sells his endowment to
receive some money, the market value of his endowment; (iii) uses this money to buy the bundle he prefers among
the bundles he can afford. Adding up the demanded bundles across consumers yields the vector of total demand
P
∈ () for the goods, and we say that all markets are in equilibrium if this vector coincides with the total
supply vector ̄.
Definition 4.23 Given an exchange economy {% }∈ , we say that a pair (∗ x∗ ), with 0, x∗ =
(∗1 ∗2 ∗ ) ∈ R+ × R+ × × R+ , is a Walrasian Equilibrium (WE) if
∗ solves max ( ) subject to ∗ · ≤ ∗ · for = 1 (39)
∈R
+
and
∗1 + ∗2 + + ∗ = ̄ that is
∗11 + ∗12 + + ∗1 = ̄ 1 , ∗21 + ∗22 + + ∗2 = ̄ 2 , ..., ∗1 + ∗2 + + ∗ = ̄ (40)
A WE consists of a price vector ∗ and an allocation x∗ = (∗1 ∗2 ∗ ) such that for each consumer , his
bundle in x∗ , ∗ , is optimal for him given ∗ , and the total consumption of good is equal to ̄ for each .
This means that (i) given ∗ , consumer has wealth ∗ · and ∗ is his preferred bundle among those in , for
= 1 ; (ii) all markets are in equilibrium.
General equilibrium takes this definition as a prediction of the outcome in this economy, that is we assume
that prices adjust to satisfy Definition 4.23. Notice that each consumer does not take into account what the other
consumers demand: he only needs to know the price vector and is confident that his demand of goods will be
satisfied. Thus, the allocation is determined in a decentralized way in the sense that the consumers’ bundles are
not chosen by a central planner. The allocation is determined by the WE prices which make consumers’ preferred
bundles consistent with total supply.
68
Figure 4.24 represents an economy with = 2, = 2 in which the budget set of consumer 1, 1 , is the triangle
with vertices 1 and the budget set of consumer 2, 2 , is the triangle with vertices 2
H x12(p) O2
N B2
x1(p)
x21(p)
Figure 4.24 : I2 I1
B1 x22(p)
x2(p)
L
O1 x11(p) M
The optimal bundle for consumer 1, 1 (), is the point in 1 on consumer 1’s highest indifference curve, 1 , which
intersects 1 . The optimal bundle for consumer 2, 2 (), is the point in 2 on consumer 2’s highest indifference
curve, 2 , which intersects 2 ; recall that the axes that measure the quantities consumed by consumer 2 have origin
in 2 , hence the coordinates of 2 (), 12 () and 22 (), are read on those axes. The price vector in Figure 4.24
is not an equilibrium price vector since the total demand for good 1, 11 () + 12 (), is smaller than ̄ 1 (recall
that ̄ 1 is the length of the base of the EB) and the total demand for good 2, 21 () + 22 (), is greater than ̄ 2
(recall that ̄ 2 is the length of the height of the EB). Hence, markets do not clear given . But if the quotient 1 2
changes, then the budget line rotates, pivoting in . As a consequence, 1 2 change and also the optimal bundle
for each consumer changes. The intuition suggests that 1 2 needs to decrease in order to obtain an equilibrium,
in such a way to decrease the demand of good 2 and increase the demand for good 1. Figure 4.25 represents a WE
x*
12 O2
I1
Figure 4.25 : x*
1
x*
21 x*
22
x*
2
I2
O1 x*
11
Remark One useful property of WE: If (∗ x∗ ) is a WE, then also (∗ x∗ ) is a WE for each 0 since
1 2 do not change if prices are ∗ instead of ∗ , hence each consumer’s optimal bundle does not change
[see Proposition 4.14(c)]. Thus, only relative prices are determined in WE, and in particular we can pick = 1∗
to obtain a WE in which the price of good is equal to one.
Example 4.26 Consider a two-consumer two-good setting such that
13 23 12 12
1 (11 21 ) = 11 21 , 1 = (3 1) and 2 (12 22 ) = 12 22 , 2 = (5 6)
69
hence ̄ = (8 7). By the virtue of the last remark above, we can focus on price vectors with 2 = 1, and then the
optimal bundle for each consumer depends only on 1 as follows:50
31 + 1 2 51 + 6 1
11 (1 ) = , 21 (1 ) = (31 + 1) 12 (1 ) = 22 (1 ) = (51 + 6)
31 3 21 2
The market clearing equations, from (40), are
11 (1 ) + 12 (1 ) = 8, 21 (1 ) + 22 (1 ) = 7 (41)
The unique solution to the first equation in (41) is 1 = 20 20
27 , and in fact 1 = 27 also solves the second equation
in (41).51 Hence, at 1 = 20 ∗ ∗ ∗ 20
27 both markets clear and = (1 2 ) = ( 27 1) is an equilibrium price vector. The
equilibrium allocation associated with ∗ is ∗1 = ( 20 27 ), ∗2 = ( 20 27 ), such that consumer 1 gives up 31
29 58 131 131
20 units
31 3127 ∗
of good 1 in exchange for 27 units of good 2 (it is not a coincidence that 3120 = ∗ ). 1
2
In this model, each consumer has an initial endowment , and he may simply consume his endowment without
engaging in any trade; formally, for each the bundle belongs to , hence it is feasible for consumer . But
trading may allow a consumer to increase his utility with respect to consuming his endowment. Whether trading is
strictly beneficial or not for a consumer depends on the endowment, on the preferences, and on the prices. In the
setting of Example 4.26, each consumer benefits from trade at the equilibrium prices as he is better off consuming
the equilibrium allocation than his initial endowment (why?). As a general principle, in no case a consumer’s utility
is decreased by the possibility to trade because the consumer can refuse to trade and consume his endowment.
After defining a notion of equilibrium, it is natural to inquire whether at least an equilibrium exists in each
economy. In the following we discuss this problem. Recall that, given , () denotes the optimal bundle for
consumer and () is the quantity wants to consume of good . Since he initially owns the quantity of
good , it is possible to interpret () = () − as the excess of demand (or net demand) of consumer in
market with respect to his endowment. Precisely, if () 0 then wants to consume of good more than his
endowment; if () 0, then wants to consume of good less than his endowment. Thus () (which may
have either sign) is the net quantity wants to buy in market .
P
It is useful to define () as the sum ∈ (). This is the total excess of demand in market : () =
P P P
∈ () = ∈ ( () − ) = ∈ () − ̄ . If () 0, then the total demand for good is greater
than the supply of good and there is excess of demand for good ; if instead () 0, then there is an excess
of supply in market . Market clears if and only if () = 0, hence a price vector (1 2 ) is a WE price
vector if and only if
1 (1 2 ) = 0, 2 (1 2 ) = 0, (1 2 ) = 0 (42)
that is (42) is equivalent to (40). This is a system of equations through which we can express the problem of
existence of WE: There exists at least a WE if and only if there exists at least a solution to (42). Notice that (42)
consists of equations with unknowns, but this property alone does not guarantee the existence of a solution to
(42).
An important property of the functions 1 2 is called Walras’ Law. For each , the following equality
holds:
1 1 () + 2 2 () + + () = 0 (43)
The left hand side in (43) is the market value of the excess of demand in market 1 (this value is positive if 1 () 0,
is negative if 1 () 0) plus the market value of the excess of demand for good 2 plus ... the market value of
50 Both and are Cobb-Douglas utility functions, and Example 4.16 describes how the optimal bundle for each consumer is
1 2
determined.
51 This is not a coincidence: See Footnote 52.
70
the excess of demand for good . Walras’ law establishes that the total market value of the excesses of demand is
equal to zero. In order to prove (43), let the vector () = (1 () 2 () ()) gather the excesses of demand
P
in the markets, and notice that (43) is equivalent to · () = 0. From () = ∈ ( () − ), it follows that
P
· () = ∈ ( · () − · ), and · () − · = 0 for each ∈ by Proposition 4.14(a) since % is monotone.
P
Hence, · () = ∈ 0 = 0.
Walras’ law has some useful consequences:
• When = 2, (43) reduces to 1 1 () + 2 2 () = 0. Hence, if 1 () = 0, that is if market 1 clears, then also
market 2 clears (and viceversa).52 If instead 1 () 0 (1 () 0), then 2 () 0 (2 () 0).
• When ≥ 3, (43) implies that if − 1 markets clear then also the remaining -th market clears. If
P
∈ () ̄ , then there is excess of demand in the market for good , therefore there exists at least one
P
market with an excess of supply, that is such that ∈ () ̄ .
Walras’ law reveals that in order to solve (42), it suffices to solve − 1 equations, for instance the first − 1
equations in (42). Moreover, from the remark just before Example 4.26 we know that we can set = 1, therefore
we can focus on the system
1 (1 2 1) = 0, 2 (1 2 1) = 0, −1 (1 2 1) = 0 (44)
In general, (44) is not much simpler than (42), but if = 2 then (44) reduces to a single equation:
1 (1 1) = 0
For this equation it is possible to prove as follows the existence of at least a solution: (i) 1 (1 1) 0 if 1 is
close to zero, say if 1 is equal to a certain 01 ; (ii) 1 (1 1) 0 if 1 is large, say if 1 is equal to a certain
001 ; (iii) 1 is a continuous function of 1 because 1 (1 1) = 11 (1 1) + 12 (1 1) + + 1 (1 1) − ̄ 1 and
11 (1 1) 12 (1 1) 1 (1 1) are all continuous functions of 1 (see Proposition 4.14(d)). Hence, the Inter-
mediate Value Theorem can be applied to show that there exists at least one ∗1 between 01 and 001 such that
1 (∗1 1) = 0. Therefore, the price vector in which 1 is equal to ∗1 and 2 is equal to 1 is a WE price vector.
An analogous results holds for each ≥ 3, that is a solution to (44) exists not only for = 2, but for each
≥ 2.
Proposition 4.27 Consider an exchange economy {% }∈ such that ̄ 0, and assume that the preference
relation % of each consumer is complete, transitive, continuous, monotone and strictly convex. Then there exists
at least one WE.
Proposition 4.27 has been a sought after result for a few decades in the past century, yet no simple proof for
it is known. Notice that Proposition 4.27 establishes existence of equilibrium, but not uniqueness: It is possible
that multiple WE exist. A unique WE exists if each consumer’s utility function is a Cobb-Douglas utility function
(Example 4.16 describes the Cobb-Douglas utility function for the case of = 2). The appendix to this section
describes an economy with three WE.
4.2.2 Pareto efficiency
After establishing the existence of at least one WE, it is natural to evaluate WE from a normative point of view:
Does a WE lead to a desirable allocation of goods among consumers? In order to answer this question, now we
introduce an important property for an allocation.
52 In Example 4.26, we have identified = 20 (given = 1) as the unique value of which clears market 1. Walras’ Law implies
1 27 2 1
that also market 2 clears at prices 1 = 20
27
, 2 = 1.
71
Definition 4.28 Given two allocations x = (1 2 ) and x∗ = (∗1 ∗2 ∗ ), we say that x Pareto dominates
x∗ if % ∗ for each , and  ∗ for at least one .
Therefore, x Pareto dominates x∗ if at least one consumer prefers x to x∗ and no consumer prefers x∗ to x.
Sometimes this property is expressed by saying that moving from x∗ to x is a Pareto improvement.
Definition 4.29 An allocation x∗ is said to be a Pareto Optimum (PO), or Pareto efficient, if there exists no
allocation which Pareto dominates x∗ .
An allocation x∗ is a PO if each way to redistribute the goods that makes some consumer better off with respect
to x∗ makes some other consumer worse off. In other words, if x∗ is a PO and a consumer wants to move from x∗
to x because he prefers x, then somebody else does not want to move to x. Conversely, if there exists x such that
everybody is happier with x, then x Pareto dominates x∗ and x∗ is not a PO (in fact, it suffices that somebody is
happier with x and nobody is hurt).53
In Figure 4.30, consider the allocation x∗ . The indifference curve of consumer 1 through x∗ , 1 , and the
indifference curve of consumer 2 through x∗ , 2 , reveal that there exist allocations which are Pareto superior to x∗ :
for instance, x. In fact, each allocation which is in the ”lens” (or ”eye”) determined by 1 and 2 Pareto dominates
x∗ because each consumer prefers such allocation to x∗ :
O2
I1
x*
Figure 4.30 :
x
I2
O1
Conversely, x∗ = (∗1 ∗2 ) in Figure 4.31 is a PO because each allocation that consumer 2 prefers to x∗ makes
consumer 1 worse off with respect to x∗ , and each allocation that consumer 1 prefers to x∗ makes consumer 2 worse
53 The notion of Pareto Optimum is not restricted to the setting of pure exchange economies. For instance, it can be applied to
games: In the Prisoner’s Dilemma (Example 3.5), the strategy profile ( ) leads to an outcome which is not a PO. In the game
of Example 3.10, the unique NE of the game, ( ), leads to an outcome which is Pareto dominated by the outcome of the strategy
profile ( ).
72
off with respect to x∗ .
O2
I1
Figure 4.31 : x*
I2
O1
In a sense, an allocation which is not Pareto optimal involves some waste: There exist opportunities which are
not exploited by the society to make somebody better off without damaging anybody. Therefore, Pareto optimality
looks like a minimal property that an allocation should satisfy to be considered a good allocation from a social
point of view. However, a Pareto optimal allocation is not necessarily a good allocation because it may be very
unfair.
Example 4.32 Consider the allocation x∗ such that ∗1 = ̄, and is equal to (0 0 0) ∈ R for each consumer
6= 1. This means that in x∗ , consumer 1 receives all the goods that are in the economy. Then, since %1 is
monotone, x∗ is a Pareto optimal allocation because each other allocation x is such that ∗1 Â1 1 .
Even though x∗ in Example 4.32 is a PO, x∗ is not a good allocation from a social point of view because it
involves a very unequal distribution. Therefore, being a PO is a nice property for an allocation, but not each Pareto
optimal allocation is satisfactory as the notion of PO neglects distributional issues among consumers.
In a typical exchange economy, many Pareto optimal allocations exist, and in particular if = 2, = 2 an
allocation x∗ = (∗1 ∗2 ) with ∗1 0 , ∗2 0 is a Pareto optimal allocation if and only if (given strictly convex
preferences) the indifference curve of consumer 1 through ∗1 is tangent to the indifference curve of consumer 2
through ∗2 . The set of Pareto optimal allocations is called Pareto set and is a curve in the EB that connects 1
with 2 , for instance like in Figure 4.33:
O2
x'
Figure 4.33 :
x
O1
Definition 4.28 introduces the relation of Pareto dominance between allocations. This relation is transitive but
is not complete because if, for instance, we consider the two Pareto optimal allocations x and x0 in Figure 4.33,
then neither of them Pareto dominates the other.
73
4.2.3 The Fundamental Theorem of Welfare Economics
Next theorem establishes a link between Pareto optimal allocations and WE allocations.
Proposition 4.34 (Fundamental Theorem of Welfare Economics) Consider an exchange economy {%
}∈ such that % is lns for = 1 , and (∗ x∗ ) is a WE. Then x∗ is a Pareto optimal allocation.
Proof The proof is by contradiction: We assume there exists an allocation x = (1 2 ) which Pareto
dominates x∗ = (∗1 ∗2 ∗ ), and from this we derive a contradiction. Given that x Pareto dominates x∗ , we
have that % ∗ for each and  ∗ for at least one , say for consumer 1: 1 Â1 ∗1 . The rest of the proof is
split into two steps.
Step 1 For consumer 1, 1 Â1 ∗1 implies ∗ · 1 ∗ · 1 . For each other consumer , % ∗ implies
∗ · ≥ ∗ · .
Proof of Step 1 Since (∗ x∗ ) is a WE, it follows that for each consumer , ∗ is an optimal bundle for consumer
in the set : See (39).
• First we consider consumer 1, and 1 ∗1 such that 1 Â1 ∗1 , and prove that ∗ · 1 ∗ · 1 ; this means
that 1 is not affordable for consumer 1. In order to see why, notice that if ∗ · 1 ≤ ∗ · 1 , then 1 ∈ 1
and this contradicts that ∗1 is an optimal bundle for consumer 1 in 1 .
• Now we consider any consumer different from 1, bundles ∗ such that % ∗ , and prove that ∗ · ≥
∗ · ; this means that has a cost that is greater or equal then the entire wealth of consumer , ∗ · . In
order to see why, notice that if ∗ · ∗ · , then there exists a small ( ) such that each bundle in ( )
costs less than ∗ · (that is, ( ) satisfies ( ) ⊆ , as in the proof of Proposition 4.14(a)). Since %
is lns, there exists ̂ in ( ) such that ̂ Â . Since % ∗ , it follows that ̂ Â ∗ . Moreover, ̂ ∈
because ̂ ∈ ( ) ⊆ . This contradicts that ∗ is an optimal bundle for consumer in .
Step 2 Deriving the contradiction.
Proof of Step 2 From Step 1 we know that
∗ · 1 ∗ · 1 , ∗ · 2 ≥ ∗ · 2 , ..., ∗ · ≥ ∗ · (45)
Adding up the inequalities in (45) yields
∗ · 1 + ∗ · 2 + + ∗ · ∗ · 1 + ∗ · 2 + + ∗ ·
that is ∗ · (1 + 2 + + ) ∗ · ( 1 + 2 + + ). Since x is an allocation, we know that 1 + 2 + + = ̄
(see Definition 4.21) furthermore, 1 + 2 + + = ̄. Thus we obtain ∗ · ̄ ∗ · ̄, which is a contradiction.
¥
This theorem reveals that if goods are allocated by competitive markets, then the outcome is a Pareto optimal
allocation, that is there are no missed opportunities. Figure 4.25 illustrates this result in an EB: x∗ = (∗1 ∗2 )
is the equilibrium allocation, and there exists no allocation (i.e., no point in the EB) which Pareto dominates x∗
because the allocations 1 prefers to x∗ are those above the curve 1 , and the allocations that 2 prefers to x∗ are
those below the curve 2 .
Proposition 4.34 is linked to a famous claim of Adam Smith, according to which in a market system, if each
individual pursues his own interest then an invisible hand guides the agents’ actions towards a desirable outcome
for the whole society. This means that the market system promotes the welfare of society even though this is not
the purpose of the agents which operate in the markets. In a sense, Proposition 4.34 formalizes Adam Smith’s idea,
as a market system leads to an allocation which is a PO. However, a Pareto optimal allocation is not necessarily
a desirable allocation for the society, since Pareto optimality neglects equity among consumers: Recall Example
4.32. Even though Proposition 4.34 is sometimes mentioned by supporters of free markets, it does not imply that a
74
government should not interfere with a market economy because a government may want to pursue distributional
corrections.
From the point of view of economic theory, Proposition 4.34 is an important result as it relies on relatively
weak assumptions, but we should notice some ”implicit” assumptions for this model: There exists a market for each
good; consumers are price takers; each consumer knows the quality of each good; there are no externalities (i.e.,
the utility of each consumer depends only on , not on the consumption bundles of the other agents). Without
these assumptions, a WE allocation is likely not a PO. In a sense, Proposition 4.34 is a benchmark for the study
of a market economy: If the outcome of a model displays a Pareto inefficiency, then necessarily some assumptions
above are violated, perhaps because some agents are not price takers and have the ability to affect prices (firms,
typically), or because there are some externalities, or there is some asymmetric information about the quality of
goods.54 Of course, in reality these assumptions are often violated and markets do not allocate goods efficiently,
but it is useful to see Proposition 4.34 as a benchmark.
4.2.4 Appendix to Section 4: Existence of multiple Walrasian Equilibria
Suppose that = 2, = 2, and
⎧
⎨ 1
1
if 11 0 and 21 0
+
1 (11 21 ) = 2
11 2
21 1 = (1 0)
⎩ 0 if 11 = 0 and/or 21 = 0
⎧
⎨
1
1 if 12 0 and 22 0
+
2 (12 22 ) = 2
12 2
22 2 = (0 1)
⎩ 0 if 12 = 0 and/or 22 = 0
with = ( 12 3
37 ) . Both 1 and 2 represent a monotone preference relation. Hence each optimal bundle for consumer 1
³ ´−1
satisfies 1 11 + 21 = 1 , and we can solve consumer 1’s problem by maximizing ̂1 (11 ) = 12 + (1 −11 11 )2
³ ´−2 ³ ´ 11
0 1 1 1 1
with respect to 11 ∈ (0 1). Then ̂1 (11 ) = 2 2 + 2 (1−11 )2 3
− 2 (1−11 )3 , which is positive if
11 1 11 1
23 23 23 12
1 1 1 37 1
11 23 , is negative if 11 is between 23 and 1. Hence 11 (1 ) = 23 and 21 (1 ) = 23 .
1 + 12
37 1 + 12
37 1 + 12
37 1 + 12
37
Each optimal bundle for consumer 2 satisfies 1 12 +22 = 1, and we can solve consumer 2’s problem by maximizing
³ ´−1 ³ ´−2 ³ ´
1
̂2 (12 ) = 12 + (1−1112 )2 with respect to 12 ∈ (0 11 ). Then ̂02 (12 ) = 2 12 + (1−1112 )2
3 − (1−1 12 )3 ,
12 12 12
12 12
which is positive if 12 12
37
13 ∈ (0 11 ), is negative if 12 is between 12
37
13 and 1
1 . Hence 12 (1 ) =
37 1 +1 37 1 +1
12 13
1
12
37
13 , 22 (1 ) = 12 13 .
37 1 +1 37 1 +1
23 12
1
The market clearing condition for market 1 is 23 + 37
13 − 1 = 0, and the graph of the left hand side is
1 + 1237
12
37 1 +1
54 A few of these themes will be considered in the course Microeconomics 2.
75
0.010
0.005
0.000
1 2 3 4 5
-0.005
The equilibrium allocation and utilities are as follows
27
• If 1 = 64 , then 1 = ( 111 27
175 175 ), 1 =
1367 631
5359 375 ,
64 148
2 = ( 175 175 ), 2 = 3241 792
5359 375 .
• If 1 = 1, then 1 = ( 37 12
49 49 ), 1 =
50 653
117 649 , 2 = ( 12 37
49 49 ), 2 =
50 653
117 649 .
64
• If 1 = 27 , then 1 = ( 148 64
175 175 ), 1 =
3241 792
5359 375 ,
27 111
2 = ( 175 175 ), 2 = 1367 631
5359 375 .
76
You might also like
- A (Micro) Course in Microeconomic Theory For MSC Students100% (3)A (Micro) Course in Microeconomic Theory For MSC Students132 pages
- An Introduction To Game Theory Notes - LatestNo ratings yetAn Introduction To Game Theory Notes - Latest125 pages
- Mailath - Economics703 Microeconomics II Modelling Strategic BehaviorNo ratings yetMailath - Economics703 Microeconomics II Modelling Strategic Behavior264 pages
- Haurie A., An Introduction To Dynamic Games LCTNNo ratings yetHaurie A., An Introduction To Dynamic Games LCTN125 pages
- Introduction To Game Theory Markus Mobius Harvard100% (3)Introduction To Game Theory Markus Mobius Harvard176 pages
- Introduction To Game Theory (Harvard) PDFNo ratings yetIntroduction To Game Theory (Harvard) PDF167 pages
- Lecture I-II: Motivation and Decision Theory - Markus M. MÄobiusNo ratings yetLecture I-II: Motivation and Decision Theory - Markus M. MÄobius9 pages
- Lecture I-II: Motivation and Decision Theory: 1 Motivating Experiment: Guess The AverageNo ratings yetLecture I-II: Motivation and Decision Theory: 1 Motivating Experiment: Guess The Average8 pages
- Economics - Lecture Notes in Microeconomic Theory100% (1)Economics - Lecture Notes in Microeconomic Theory493 pages
- Economics Lecture Notes in Microeconomic TheoryNo ratings yetEconomics Lecture Notes in Microeconomic Theory493 pages
- Ssentials of AME Heory: A Concise Multidisciplinary IntroductionNo ratings yetSsentials of AME Heory: A Concise Multidisciplinary Introduction4 pages
- 2019 Algorithmic Game Theory Lecture NotesNo ratings yet2019 Algorithmic Game Theory Lecture Notes106 pages
- Inferior - Goods - Giffen - Goods - and - Shochu (Kannai 2001)No ratings yetInferior - Goods - Giffen - Goods - and - Shochu (Kannai 2001)10 pages
- Dual Process Theory Evans - Stanovich - PoPS13 PDFNo ratings yetDual Process Theory Evans - Stanovich - PoPS13 PDF19 pages
- Rationality in Economics: Constructivist and Ecological FormsNo ratings yetRationality in Economics: Constructivist and Ecological Forms5 pages
- Why We Should Reject Nudge': Research and AnalysisNo ratings yetWhy We Should Reject Nudge': Research and Analysis8 pages
- Can Policing Disorder Reduce Crime? A Systematic Review and Meta-AnalysisNo ratings yetCan Policing Disorder Reduce Crime? A Systematic Review and Meta-Analysis22 pages
- Optimize, Satisfice, or Choose Without Deliberation? A Simple Minimax-Regret AssessmentNo ratings yetOptimize, Satisfice, or Choose Without Deliberation? A Simple Minimax-Regret Assessment19 pages
- Do Predictive Power of Fibonacci RetraceNo ratings yetDo Predictive Power of Fibonacci Retrace6 pages
- Ashenfelter Predicting Quality Boudeux WineNo ratings yetAshenfelter Predicting Quality Boudeux Wine14 pages
- PWT - 8 NSW Ie'K LWHNQ: XWD R'KX Xog G'LWNo ratings yetPWT - 8 NSW Ie'K LWHNQ: XWD R'KX Xog G'LW2 pages
- DR.V.N Bedekar Institute of Management Presents Enthusia 4.0No ratings yetDR.V.N Bedekar Institute of Management Presents Enthusia 4.015 pages
- Devil Fruit Encyclopedia (D&D 5e System Homebrew) - The HomebreweryNo ratings yetDevil Fruit Encyclopedia (D&D 5e System Homebrew) - The Homebrewery148 pages
- Best Inquisitor Skill Build in Dragon Nest SEA (Guide) - Web Junkies BlogNo ratings yetBest Inquisitor Skill Build in Dragon Nest SEA (Guide) - Web Junkies Blog9 pages
