Sie sind auf Seite 1von 102

Advanced Microeconomics

+
Konstantinos Serfes
September 24, 2004

These notes have borrowed content and notation from the following two textbooks: i) Microeconomic Theory by
A. Mas-Colell, M.D. Whinston and J. Green and ii) Advanced Microeconomic Theory by G.A. Jehle and P.J. Reny.
You should study these notes in conjunction with the above two textbooks. These lectures notes are likely to contain
a number of errors.
1 CONSUMER THEORY
Formulate the consumers utility
maximization problem subject to a budget
constraint
Solve the consumers problem to
derive the demand functions
Study the properties of the demand
functions
A QUICK OVERVIEW OF CONSUMER
THEORY
There are four building blocks in any model of consumer choice: i) the consumption set, ii) the
preference relation, iii) the feasible set, and iv) the behavioral assumption.
The consumption set depicts all the options that are available to the consumer (whether these
options are aordable of not). The preference relation puts some structure on how the consumer
ranks the dierent alternatives available to him. The feasible set includes all the options from
the consumption set that are also aordable (in a sense that will be made precise later). Finally,
the behavioral assumption is about how the consumer chooses among his alternatives, given his
preferences. In particular, we will assume that the consumer chooses the best alternative.
We will study each one of the four building blocks separately and at the end we will put them
together to pose the consumers constraint maximization problem.
The consumption set A
We assume that there are 1 commodities. Each commodity / is measured in some innitely
divisible unit. Let r

1
+
(non-negative) represent the number of units of good /. We denote
1
a vector of the 1 goods (consumption bundle) by x = (r
1
, r
2
, ..., r
1
) 1
1
+
.
1
We make the
following assumptions about A.
1. O , = A _ 1
1
+
.
2. A is closed.
2
3. A is convex.
3
4. 0 A.
First, we assume that the consumption set is a non-empty (i.e., it contains at least one element)
subset of an 1dimensional Euclidean space. Assumption 2 is needed for technical reasons that
will become clear later on (1
1
+
for instance is closed). The third assumption ensures that A has no
holes in it and no missing parts (i.e., averages of any two bundles in the set also belong in the set).
For example, if the goods were not perfectly divisible (say each good is measured only in integers
0, 1, 2, see gure 1), then A would not be convex. Finally, assumption 4 allows for the possibility
of consuming nothing.
x
1
x
2
0 1 2
1
2
Figure 1: A non-convex consumption set
1
We will use bold letters to denote vectors.
2
For a denition of a closed set go to pages 943-944 of Mas-Colell.
3
For a denition go to page 946 of Mas-Colell.
2
Preferences
A preference relation is denoted by %. This is a binary relation on A. If x and y are in A and
x % y, then we say that x is at least as good as y. Next, we list the basic axioms that a preference
relation must satisfy.
AXIOM 1: (Completeness). For all x and y in A, either x % y, or y % x.
This axiom basically says that the consumer, when he is confronted with two dierent consump-
tion bundles, he must be able to say which one he prefers more, or that he is indierent between the
two. What is ruled out by this axiom is a situation where the consumer cannot rank two bundles. A
deeper inspection should convince you that this axiom is absolutely necessary for consumer theory.
After all, we want the consumer to choose the best bundle for him. But to choose the best element
of A, he must be able to make pairwise comparisons.
AXIOM 2: (Transitivity). For any three elements x, y and z in A, x % y and y % z, then
x % z.
Without the transitivity axiom there may not exist a best consumption bundle in A, even when
pairwise comparisons are allowed for. For example, suppose that A = x, y, z, i.e., it consists of
only three consumption bundles. Further assume that axiom 2 is violated. More specically, x % y
and y % z, but z ~ x. You can easily check that there is no best element in A. If the consumer
has to choose between x and y he will choose x. Between y and z he will choose y. But between x
and z he will choose z! We made a full cycle. Without being able to identify a set of best elements,
consumer theory does not have much predictive power.
Denition 1: The preference relation % on the consumption set A is rational, if it satises
axioms 1 and 2.
Denition 2: (Strict preference relation). The binary relation ~ on A is dened as follows:
x ~ y if and only if x % y and y x.
The phrase x ~ y is read, x is strictly preferred to y.
Notice that ~ is not complete as it cannot compare indierent bundles.
3
Denition 3: (Indierence relation). The binary relation ~ on A is dened as follows:
x ~ y if and only if x % y and y % x.
The phrase x ~ y is read, x is indierent to y.
Denition 4: (Sets in X derived from the preference relation). Let x be a consumption bundle
in A. Relative to this point we can dene the following subsets of A:
1. % (x) = y A : y % x , called the at least as good as set.
2. - (x) = y A : y - x , called the no better than set.
3. - (x) = y A : y - x , called the worse than set.
4. ~ (x) = y A : y ~ x , called the preferred to set.
5. ~ (x) = y A : y ~ x , called the indierence set.
Axioms 1 and 2 have put very few restrictions on the preferences. In particular, the induced
from % sets (i.e., 1-5 above) can practically admit a wide range of shapes (see gure 2 where
A = 1
2
+
). For instance, the indierence set may be thick, the at least as good set may not be
closed or it may not be convex.
x
>(x) ~(x)
<(x)
x
1
x
2
Figure 2
Before we attempt to put more structure on the preferences, by imposing more axioms, we
introduce the concept of the utility function. Later on more axioms will be assumed.
4
Utility Functions
Denition 5: (Utility function) A function n : A 1 is a utility function representing
preference relation %, if for all x, y A,
x % y if and only if n(x) _ n(y).
We want to construct a function which, given any two consumption bundles, will assign a greater
number to the bundle that the consumer prefers more. Dealing with utility functions is much more
convenient (because functions can be dierentiated, integrated etc.) than dealing with preferences.
However, we have to demonstrate that such a transition (i.e., from preferences to utility functions
that are consistent with these preferences) is possible and without any loss of generality. This is
what we will eventually prove.
The utility function n is not unique. For any strictly increasing function ) : 1 1, )(n(x)) is a
new utility function representing the same preferences [see exercise 1.B.3]. For example assume that
there are only two goods and each good is measured only in integers 0, 1, 2 . Moreover, assume
that the preferences are dened as follows:
x ~ y if and only if r
1
+r
2
j
1
+j
2
and
x ~ y if and only if r
1
+r
2
= j
1
+j
2
.
Observe that % as dened above are complete and transitive. One utility representation is:
n(0, 0) = 0, n(0, 1) = n(1, 0) = 1, n(2, 0) = n(0, 2) = n(1, 1) = 2
n(2, 1) = n(1, 2) = 3 and n(2, 2) = 4.
The above function is indeed a utility function. But it is not unique. Any monotonic transfor-
mation of the utility function preserves the rankings. If you take, for example, the logarithm (or
the square root) then you obtain another utility function which represents the same preferences.
What matters is the ranking of the consumption bundles not the actual number that the function
assigns to a bundle.
Properties of the utility function that are invariant to a strictly increasing transformation are
called ordinal.
5
Cardinal, on the other hand, are not preserved under all such transformations.
The preference relation induced by n, as we demonstrated above, is an ordinal property. The
numerical values assigned by n are cardinal. I will list more ordinal and cardinal properties later
in the course.
An important issue in consumer theory is to identify the axioms that would guarantee that for
any preference relation % which satises these axioms there exists a utility function representing
these preferences. What the next proposition says is that completeness and transitivity are the
minimum set of axioms that must be imposed on %. Without these two axioms it is impossible to
construct a utility function.
Proposition 1: A preference relation % can be represented by a utility function only if it is
rational.
Proof: We want to prove the following. Suppose that there exists n : A 1 representing %.
Then % must be complete and transitive.
Suppose by way of contradiction that this is not true. That is, there exists such a n but
preferences are not complete and transitive. So, they may be either incomplete, or intransitive, or
both.
Completeness: Take any x and y in A such that the consumer cannot compare. Then, n(x)
cannot be compared with n(y). But n is a real-valued function and n(x) and n(y) are real numbers,
which can be ranked. Contradiction.
Transitivity: Take any x, y and z in A such that x % y and y % z but z ~ x. This implies that
n(x) _ n(y) and n(y) _ n(z). But since n is a real-valued function, it must be that n(x) _ n(z).
Contradiction.
In other words, rational preferences is a necessary condition for the existence of a utility function
representing these preferences. However, rationality of preferences in general is not sucient for
them to be represented by a utility function. As we will shortly see we need more axioms to
guarantee that. Nonetheless, we can ask the following question.
Question: Under what conditions rationality of preferences is also a sucient condition for
the existence of a utility function?
6
Exercise 1.B.5.: Show that if A is nite and % is a rational preference, then there exists
n : A 1 that represents %.
You can easily see this and a proof is not needed. The point is that niteness of the consumption
set and rationality of preferences are sucient conditions for the existence of a utility function. As
it will become evident later, when A is not nite, this assertion is not necessarily true (when for
example 1 = 2 and A = [0, 1] [0, 1]).
We will return to preferences again to impose more axioms.
The feasible set
Let p = (j
1
, j
2
, ..., j
1
) 1
1
++
denote a vector of prices and n 1
++
denote the consumers
income. We assume that p is a row vector and x a column vector. A competitive budget is dened
as,
1
p,&
= x A : p x =j
1
r
1
+j
2
r
2
+ +j
1
r
1
_ n .
The set,
1
p,&
= x A : p x = n
is called the budget hyperplane. In two dimensions (i.e., when 1 = 2) the feasible set is the shaded
area in gure 3 below.
x
1
x
2
w/p
1
w/p
2
Slope = - (p
1/
p
2
)
B
p,w
{x in R
L
+
: px = w}
Figure 3
7
Claim: The competitive budget 1
p,&
is a convex set.
The above assertion can be proved as follows: Take y and z in 1
p,&
. We have to show that
x
0
= cy + (1 c)z 1
p,&
,
where c [0, 1] . More precisely, we must show two things: i) x
0
A and ii) p x
0
_ n.
Clearly x
0
A since by assumption A is convex. Also we have,
p y _ n and p z _ n,
which implies,
cp y _ cn and (1 c)p z _ (1 c)n.
By adding the two inequalities we obtain,
cp y + (1 c)p z = p cy +p (1 c)z = p (cy+(1 c)z) =
p x
0
_ cn + (1 c)n = n,
which completes the proof.
Behavioral assumption
The consumer chooses the bundle that is most preferred among the ones that are in the feasible
(budget) set.
We continue by imposing more axioms on the preferences.
AXIOM 3 (Monotonicity):
Strong monotonicity: If x and y A such that y _ x and y ,= x, then y ~ x. (y _ x means
that every coordinate of y is weakly greater than the corresponding coordinate of x, but the
assumption y ,= x prevents all of them from being equal).
Monotonicity: If x and y A such that y x then y ~ x (y x means that every coordinate
of y is strictly greater than the corresponding coordinate of x).
8
Locally non-satiated: For all x A and for all - 0 there exists a y A such that,
|y x| _ - and y ~ x,
where |y x| is the Euclidean norm, i.e., |y x| =
_

1
=1
(r

)
2
.
To see the dierence between the three monotonicity concepts consider rst the preferences
which are represented by a Cobb-Douglas utility function: n = r
1
r
2
. These preferences are strongly
monotone, monotone and locally non-satiated. A bundle that contains more of only one good is
strictly preferred to the initial bundle.
Now consider preferences which are represented by a Leontief utility function: n = minr
1
, r
2
.
These preferences are monotone and locally non-satiated, but not strongly monotone. A new bundle
must contain more of both goods to be strictly preferred.
The following is true (PROVE IT AS AN EXERCISE):
Strong monotonicity =Monotonicity =Local non-satiation.
Therefore, strong monotonicity is the strongest assumption and local non-satiation is the weak-
est. For most of the subsequent analysis local non-satiation is enough.
Question: What does each one of the monotonicity assumptions imply about the shape of the
indierence curves?
With local non-satiation we eliminate thick indierence curves (gure 4). With monotonicity we
can guarantee that the indierence curves are not thick, do not have any increasing parts (at part
are allowed; gure 5)) and the preferred to set is above the indierence curve and the worse
than set below. Finally, with strong monotonicity we obtain indierence curves which are not
thick, are always (strictly) downward sloping (gure 6) and the preferred to set is above the
9
indierence curve and the worse than set below. [Show why the above claims are true]
x
x
1
x
2
Figure 4
x
x
1
x
2
Figure 5
x
x
1
x
2
Figure 6
AXIOM 4 (Continuity):
For all x A, the set % (x) (i.e., the at least as good as set, or the upper contour set) and
- (x) (i.e., the no better than set, or the lower contour set) are closed in A.
A set is closed in a particular domain if its complement is open in that domain. So, if % (x)
is closed in A then its complement - (x) is open in A. The continuity axiom guarantees that
sudden preference reversals do not occur. The continuity axiom can equivalently be expressed
using sequences of bundles, rather than open and closed sets. We can then say that % (x) is closed
if for any sequence y
a
% (x) such that y
a
y, we have y % (x).
Notice that if both % (x) and - (x) are closed, then so is ~ (x) since ~ (x) =% (x) - (x)
and an intersection of closed sets is closed. This implies that the continuity axiom rules out open
areas in the indierence set like the one in gures 4,5 and 6.
We need both the upper and the lower contour sets to be closed. Take a look at the preferences
as given in gure 7 (no monotonicity assumption is satised in this example). The indierence set
is closed (the dark grey ball). The strictly lower contour set (the light grey ball) does not contain
its boundary points. Everything else is the strictly upper contour set. Notice that the set ~ (x) is
closed. The set - (x), however, is not closed. Notice that - (x) =~ (x)' - (x) which is the union
of a closed with an open set which does not have to be closed. Nevertheless, the upper contour set
% (x) is closed.
10
I can nd a converging sequence y
a
- (x) such that y
a
y and y is located on the dotted
circumference of the light grey ball. In this case the limit point of y
a
which is y belongs in ~ (x)
(sudden preference reversal).
x
1
x
2
x
~(x)
<(x)
>(x)
y y
n
Figure 7
AXIOM 5 (Convexity):
Weak convexity: Preferences % on A are convex if for all x, the set % (x) is a convex set, that is,
if y % (x) and z % (x), then w =cy + (1 c)z % (x), for all c [0, 1] .
Strict convexity: Preferences % on A are strictly convex if for all x, the set % (x) is a strictly
convex set, that is, if y % (x) and z % (x) and y ,= z then w = cy + (1 c)z ~ (x), for
all c (0, 1) .
Convexity of preferences implies that the consumer prefers average bundles to extreme ones.
Another interpretation, is that the consumer has a taste in favor of diversication (see Chapter 6
of the main text).
When 1 = 2 the absolute value of the slope of an indierence curve is called the marginal rate
of substitution (MRS), which measures how many units of good 2 the consumer is willing to give
up for a marginal unit of good 1, such that he remains indierent after the exchange. If preferences
are monotonic (which implies that the upper contour set is above and the lower contour set below)
11
convexity implies that the amount of good 2 that the consumer is willing to give up for an extra
(marginal) unit of good 1 decreases as the amount of good 1 increases (see gure 8). In particular
we have,
Convexity = Diminishing (MRS)
Strict convexity = Strictly diminishing (MRS).
x
1
x
2
x
y
z
w
Figure 8: Strictly convex preferences and diminishing MRS
EXISTENCE OF A UTILITY FUNCTION
A central question in consumer theory is: what is the minimum set of axioms that we should
impose on preferences to guarantee the existence of a utility function representing these prefer-
ences? As we have seen in the previous lecture rationality of preferences is an absolutely necessary
assumption. Without it we can guarantee the non-existence of a utility function. We also saw that
if A is nite, rationality of preferences guarantees the existence of n.
First, we oer an example where A is not nite, preferences are rational but not continuous,
and it is impossible to construct a utility function.
Assume that 1 = 2. Dene % as follows,
x % y if either: i) r
1
j
1
or ii) r
1
= j
1
and r
2
_ j
2
12
This is known as lexicographic preference relation. (Draw an indierence curve).
Claim: These preferences do not satisfy the axiom of continuity. They are however rational,
strongly monotone and strictly convex [see exercise 3.C.1].
I will show that the lower contour set is not closed. Let y = (2, 3) and y
a
= (2 1,:, 3) . Also
let x = (2, 2) . Clearly y
a
y as : and y
a
- (x) for all :. But y , - (x), since y ~ x. A
similar argument can show that the upper contour set is not closed either.
Question: Can we nd a n : A 1 that represents lexicographic preferences?
Without any loss of generality, assume that A = [0, 1] [0, 1] . Suppose that there exists a
n : A 1 which represents the lexicographic preferences.
x
1
x
2
x
1
1
1
y
y

z
z

x
x

1
x

Figure 9
The proof follows gure 9 closely. It can be easily seen that n(y) n(y
0
). Since rationals are
dense we can always nd a rational between two real numbers. Hence, we can nd a bundle x
such that n(x) is a rational number [denoted by n(x) = r(x)] and n(y) r(x) n(y
0
). We also
have n(z) n(z
0
) and by the same logic there exists another rational, denoted by r(x
0
) such that,
n(z) r(x
0
) n(z
0
). Moreover, n(z
0
) n(y) [since bundle z
0
contains more of good 1 than bundle
y] and therefore,
n(z) r(x
0
) n(z
0
) n(y) r(x) n(y
0
).
13
Notice that r
0
1
r
1
and r(x
0
) r(x). I have constructed a one-to-one function from the set of
reals (the rs) to the set of rationals (the r(x)s). But this is a mathematical impossibility since
the set of reals is an uncountable set while the set of rationals is countable (the rationals can be
put into a one-to-one correspondence with the set of natural numbers which is also countable; the
reals cannot). They do not have the same cardinality and therefore it is not plausible to construct
a one-to-one mapping. Therefore, a utility function does not exist.
Theorem 1 (Existence of a continuous utility function): If % are complete, transi-
tive, continuous and strongly monotone, then there exists a continuous real-valued function (utility
function) n : A 1 representing %.
Idea of the proof: The intuition behind the proof can be grasped by studying gure 10.
x
1
x
2
x
u(x)e
~(x)
u(x)
u(x)
1
1
e
45
o
~(y)
y
u(y)e
Figure 10
Fix a bundle x. Find all the indierent to x bundles (indierence curve). Based on the axioms
this set exists and it is not thick. Assign a number to all these bundles. A dierent bundle y that
is strictly preferred to x must lie on a higher indierence curve. Assign a higher number to y (and
to the bundles indierent to y). This is how we construct a utility function which represents these
preferences. The function is also continuous, because there are no jumps in the preferences and
therefore no jumps in the numerical assignments.
Rigorous proof: We will prove that at least one such utility function exists. Let
e = (1, 1, ..., 1) 1
1
++
14
and consider the mapping n : A 1 dened so that the following is satised,
n(x) e ~ x. (*)
Question 1: Does there always exist a number n(x) satisfying (*) ?
Question 2: Is it uniquely determined so that n(x) is a well-dened function ?
We attempt to answer the rst question. Fix x A and consider the following two subsets of
the real numbers:
= t _ 0 : t e % x and 1 = t _ 0 : t e - x .
If t
+
1, then t
+
e ~ x, so setting n(x) = t
+
would satisfy (*). Thus, we have to show
that 1 ,= O.
Continuity of %: This implies that and 1 are closed sets
Strong monotonicity of %: t =t
0
for all t
0
_ t. In other words, must be a closed
interval of the form [t

, ). Similarly, 1 must be a closed interval of the form [0,

t].
Completeness of %: This implies that either t e % x or t e - x, that is t '1. But this
means that 1
+
= ' 1 = [t

, ) ' [0,

t].
Therefore, t

t so that 1 ,= O.
Now lets answer the second question. We must show that there exists only one number t _ 0
such that t e ~ x.
If t
1
e ~ x and t
2
e ~ x then by transitivity t
1
e ~t
2
e. By strong monotonicity it must be
that t
1
= t
2
. So for every x A there exists exactly one number n(x) such that
n(x) e ~ x.
Next, we have to show that n (as we constructed it) represents %.
Consider two bundles x and y and n(x), n(y) which by denition must satisfy
n(x) e ~ x and n(y) e ~ y.
15
We have
x % y ==n(x) e ~ x % n(y) e ~ y
transitivity ==n(x) e %n(y) e
strong monotonicity ==n(x) _ n(y).
Now it remains to be shown that n is continuous. We must show that the inverse image under
n of every open ball in 1 is open in A.
n
1
((a, /)) = x A : a < n(x) < /
monotonicity = x A : a e - n(x) e - / e
n(x) e ~ x and transitivity = x A : a e - x - / e
= ~ (a e) - (/ e).
% and - are closed sets and therefore - and ~ are open. n
1
((a, /)) is the intersection of two
open sets and hence it is also open.
Exercise 3.C.4. Exhibit an example of a preference relation that is not continuous but is
representable by a utility function.
Let A = 1
+
and dene n() : A 1 by,
n =
_
_
_
0, if r < 1
1, if r 1
any in [0, 1], if r = 1.
Denote by % the preference relation represented by n(). We will show that % is not continuous,
by showing that either the upper or the lower contour set is not closed. First, suppose that n(1) 0.
Let
r
a
= 1
1
:
~ 0.
r
a
- (0) and it converges to r = 1 , - (0) since 1 ~ 0. Thus, the lower contour set is not
closed.
Second suppose that n(1) < 0. Let
r
a
= 1 +
1
:
~ 2.
r
a
% (2) and it converges to r = 1 , % (2) since 2 ~ 1. Thus, the upper contour set is not
closed.
16
1.1 Properties of preferences and utility functions
Let % be represented by n : A 1. Then,
1. n(x) is strictly increasing if and only if % is strongly monotonic.
2. n(x) is quasi-concave if and only if % is convex.
3. n(x) is strictly quasi-concave if and only if % is strictly convex.
Homotheticity: A continuous % is homothetic if and only if it admits a utility function that is
homogeneous of degree one.
A function ) : 1
a
1 is homogeneous of degree one if )(c x) = c)(x), for any
c 1
++
.
A function / : 1
a
1 is homothetic if it can be written as q()(x)) where q : 1 1
is strictly increasing and ) : 1
a
1 is homogeneous of degree 1.
If preferences are homothetic then the slopes of the level sets of n (Level set = x A : n(x) = /)
are unchanged along any ray through the origin (see gure 11). As it will become evident later on,
this implies that if preferences are homothetic, the goods are not inferior goods.
17
x
1
x
2
x
2x
y
2y
Figure 11
Example: The preferences represented by a Cobb-Douglas utility function, n = r
o
1
r
b
2
, a 0 and
/ 0 are strongly monotonic, strictly convex and homothetic. The Cobb-Douglas utility function
is strictly quasi-concave, strictly increasing and homothetic.
The fact that n is strictly increasing is easy to check. To verify that it is strictly quasi-concave,
the bordered Hessian matrix must be negative denite [see Mas-Colell p.938], i.e.,

n
11
n
12
n
1
n
21
n
22
n
2
n
1
n
2
0

0.
The latter is true if and only if
2n
1
n
2
n
12
[n
1
]
2
n
22
[n
2
]
2
n
11
0.
Now you can check that the following is true,
2n
1
n
2
n
12
[n
1
]
2
n
22
[n
2
]
2
n
11
= a/(a +/)r
(3o2)
1
r
(3b2)
2
0.
Put it dierently, we require n to be strictly concave (i.e., to have a negative denite Hessian
matrix 1
2
n(x)) not everywhere but in the subspace z A : \n(x) z = 0, where,
\n(x) =
_
0n
0r
1
,
0n
0r
2
_
,
is the gradient vector (see gure 12).
18
x
1
x
2
Gradient vector
z
x
Figure 12
So the condition says that n must be strictly concave when we move in such a way that the
level of the function does not change.
Another way to prove the strict quasi-concavity of n is to show that the upper contour set is
strictly convex. The upper contour set is,
_
x A : r
o
1
r
b
2
_ /
_
.
The indierence curve is
r
2
=
_
/
r
o
1
_
1b
.
Dierentiate the above function twice and you will see that the outcome is positive which
indicates that the indierence curve is strictly convex and the upper contour set as a consequence
is also convex.
To prove that preferences are homothetic you have to show that preferences admit a utility
function which is homogeneous of degree 1. Right now, it is not necessarily true that n = r
o
1
r
b
2
is
homogeneous of degree 1. But all we have to show is that % admit a homogeneous of degree 1 utility
function. Now remember that any strictly monotonic transformation of n leaves the preferences
unchanged. Transform n as follows,
~ n =
_
r
o
1
r
b
2
_
1(o+b)
.
Finally verify that ~ n is homogeneous of degree 1,
~ n(t x) =
_
(tr
1
)
o
(tr
2
)
b
_
1(o+b)
= t
o(o+b)
t
b(o+b)
_
r
o
1
r
b
2
_
1(o+b)
= t
(o+b)(o+b)
~ n = t~ n.
19
Exercise: Consider %which are represented by the following utility function: n =
_
r
2
+
_
r
1
r
2
.
Are these preferences: i) monotonic, ii) convex and iii) homothetic?
Solution: Preferences are clearly strongly monotonic. They are also strictly convex. To see this
compute the determinant of the bordered Hessian matrix, which yields (verify the result),
.0625
5
_
r
1
r
2
+ 4r
1
_
r
2
+
_
r
2
r
1
_
r
2
_
r
1
r
2
0.
To prove that preferences are homothetic, we must show that the slope of the indierence curves
is constant along a straight line emanating from the origin. In other words, the following set of
bundles
_
(r
1
, r
2
) A :
'l
1
'l
2
= /
_
forms a straight line. 'l
i
= 0n,0r
i
(marginal utility) and the ratio of the marginal utilities is the
marginal rate of substitution. This would imply that the locus where the slope of the indierence
curves is constant is a straight line.
'l
1
'l
2
= / =
1
2
(r
1
r
2
)
12
r
2
1
2
r
12
2
+
1
2
(r
1
r
2
)
12
r
1
= / =
1
2
r
12
2
+
1
2
(r
1
r
2
)
12
r
1
1
2
(r
1
r
2
)
12
r
2
=
1
/
=
_
r
1
+r
1
r
2
=
1
/
=
r
2
= /
_
_
r
1
+r
1
_
Preferences are not homothetic since the slope of the indierence curves is not constant on a
straight line starting from the origin.
In the remaining of the course we will assume (without proving it) that the utility function is
dierentiable (that is, the right and left derivatives at any point are equal). Notice that continuity
does not guarantee dierentiability. A standard example is the function j = [r[ which is clearly
continuous, but at zero is not dierentiable. The left derivative is 1 and the right +1. To ensure
the dierentiability of the utility function we have to assume that preferences are smooth (in some
sense that will not be made precise). Basically, one thing that is not allowed is indierence curves
with kinks. For a rigorous treatment see G. Debreu, Smooth preferences, Econometrica (1972).
Given that we can dierentiate n, we can formally introduce the following concepts.
Marginal utility of good /:
'l

(x) =
0n(x)
0r

.
20
Marginal rate of substitution between goods i and ,:
'1o
i)
(x) =
0n(x),0r
i
0n(x),0r
)
.
Now we are ready to state and solve the consumers utility maximization problem.
1.2 The utility maximization problem (UMP)
The consumer takes the market prices p and his income n as given and chooses the consumption
bundle x which gives him the highest utility provided that the chosen bundle is aordable. Formally,
max
xA
n(x)
subject to : p x _ n.
The solution to this problem is known as Marshallian (Mas-Colell calls them Walrasian)
demand functions and are represented as,
x(p, n) =
_
_
_
_
_
_
r
1
(p, n)

r
1
(p, n)
_
_
_
_
_
_
1
1
+
.
The vector x(p, n) shows how many units of each good the consumer will choose so that he
maximizes his utility given the vector of prices p and his income n.
First, we will prove that a solution to the UMP exists.
Proposition: If j 0 and n is continuous, then the UMP has a solution.
Proof: If j 0 then the budget set 1
p,&
= x A : p x _ n is both closed and bounded.
Bounded: 0 _ r

_ n,j

. There exists a clear upper and lower bound to the consumption of


each good.
Closed: Take any sequence x
a
1
p,&
such that x
a
x. We must show that x 1
p,&
. Suppose
by way of contradiction that this is not the case, i.e., x , 1
p,&
. This means that p x n. Since
21
the inner product p x is continuous in x, I can nd an ~ : (large enough) and an associated bundle
x
~ a
very close to x, such that p x
~ a
n. But this is a contradiction, since we have assumed that
x
a
1
p,&
for any :.
By the Heine-Borel theorem (which states that a subset of 1
1
is compact if and only if it is
closed and bounded) 1
p,&
is a compact set.
Now we can apply the Maximum Theorem, which states that a real-valued continuous function
on a compact set attains its maximum and minimum.
In our case the function is the utility function (which is continuous if axioms 1-4 hold) and is
dened on the budget set which is compact. The maximum is attained by the solution to the UMP
which is the demand functions.
1.2.1 Demand functions and comparative statics
Denition: x(p, n) is homogeneous of degree zero in (p, n) if, x(cp, cn) = x(p, n), for all
p, n and c 0.
This says that if all prices and income increase (or decrease) by the same proportion, then the
optimal demand should not change. This is like a synchronized ination which raises all prices and
incomes by the same percentage.
Denition: x(p, n) satises the Walras Law if, p x(p, n) = n, for all p, n.
This says that the consumer will spend all his income. As we will see below, this is clearly true
when preferences are locally non-satiated.
Denition (Income eects): A commodity / is normal at (p, n) if,
0r

(p, n)
0n
_ 0,
22
and inferior if,
0r

(p, n)
0n
_ 0.
If every commodity is normal at all (p, n), then we say that the demand is normal.
Below, I present the income (wealth) eects in vector representation,
1
&
x(p, n) =
_
_
_
_
_
_
0r
1
(p, n),0n

0r
1
(p, n),0n
_
_
_
_
_
_
1
1
.
Denition (Price eects): A commodity / is ordinary at (p, n) if,
0r

(p, n)
0j

_ 0,
and Gien if,
0r

(p, n)
0j

_ 0.
If a goods demand curve is everywhere downward sloping in the goods own price, then the
good is ordinary. Otherwise, it is Gien.
Below, I present the price eects in matrix representation,
1
j
x(p, n) =
_
_
_
_
_
_
_
_
0r
1
(p, n),0j
1
0r
1
(p, n),0j
1



0r
1
(p, n),0j
1
0r
1
(p, n),0j
1
. .
_
_
_
_
_
_
_
_
11 matrix
.
Denition (Price elasticities): Good /s elasticity with respect to good /s price is dened
as,
-
I
(p, n) =
0r

(p, n)
0j
I
j
I
r

(p, n)
.
It measures good /s percentage change following a marginal percentage change in good /s
price. If / = /, then we call it own price elasticity, while if / ,= / it is called cross price elasticity.
23
Denition (Income elasticities): Good /s elasticity with respect to income n is dened as,
-
&
(p, n) =
0r

(p, n)
0n
n
r

(p, n)
.
It measures good /s percentage responsiveness to a small percentage change of income.
Elasticities do not depend on the units chosen for measuring commodities and for this reason
they provide a unit-free way of capturing demand responsiveness. They are used extensively in
applied work.
Proposition: Suppose that n is a continuous and strictly quasi-concave utility function rep-
resenting locally non-satiated preferences. Then the Marshallian demand functions x(p,n) satisfy
the following properties:
1. They are homogeneous of degree zero in (p, w).
2. They satisfy the Walras Law.
3. x(p, n) consists of a single element, i.e., it is a function rather than a correspondence (multi-
valued function).
Proof:
1. We have to show that x(cp, cn) = x(p, n), for all p, n and c 0. Multiply each price and
income by a scalar c. First notice that the utility is not aected by c. Simply c does not
enter n. Second observe that
1
cp,c&
= x A : cp x _ cn = 1
p,&
= x A : p x _ n .
The budget set remains unchanged. Therefore, c does not change the maximization problem
in any way and consequently the solution must be the same, regardless of the c.
2. Suppose by way of contradiction that p x(p, n) < n. Then there must be another bundle y
very close to x such that: i) (by continuity of the inner product) p y < n and ii) (by local
non-satiation) y ~ x. So I found an aordable bundle which is strictly preferred to x. But
this contradicts the optimality of x.
3. I must show that x(p, n) is single-valued. In other words, for any (p, n), there is only one
value of each good which solves the consumers maximization problem. Suppose otherwise,
24
i.e., there are two bundles q and y with q ,= y such that both belong to x(p, n). This means
two things: i) both bundles are aordable and ii) there is no other aordable bundle that is
strictly preferred to these two. Let
z = cq + (1 c)y, c (0, 1) .
Since the utility function is strictly quasi-concave, preferences are strictly convex and a convex
combination of q and y (which is the new bundle z) must be strictly preferred, i.e., z ~ q
and z ~ y. But z is also aordable, i.e.,
p z = p (cq + (1 c)y) _ cn + (1 c)n = n.
Hence, q and y cannot be optimal at the same time. This contradiction leads us to conclude
that x(p, n) is single-valued.
1.2.2 Implications of homogeneity and Walras Law for price and wealth eects
N The homogeneity of degree zero implies that for all (p, n),
1

I=1
-
I
(p, n) +-
&
(p, n) = 0, for / = 1, ..., 1. (*)
To see this x one good, say good /. The above equation says that the sum of all the price (own
and cross) and income elasticities is zero. If all prices and income increase by 1%, the demand of
good / will remain unaected. The same holds for all goods.
Proof : Homogeneity of degree zero means,
x(cp, cn) x(p, n) = 0, c 0.
Dierentiate the above expression with respect to c,
1

I=1
0r

(cp, cn)
0 (cj
I
)
j
I
+
0r

(cp, cn)
0 (cn)
n = 0, / = 1, ..., 1,
and set c = 1 to obtain (the above equation holds for any c 0 and it will certainly hold for
c = 1),
1

I=1
0r

(p, n)
0 (j
I
)
j
I
+
0r

(p, n)
0 (n)
n = 0, / = 1, ..., 1.
25
If we divide both sides of the above equation by r

(p, n) and using the elasticity formulas we


obtain (*).
N The Walras Law implies that for all (p, n),
(Cournot aggregation)
1

I=1
/
I
(p, n)-
I
(p, n) +/

(p, n) = 0, for / = 1, ..., 1, (**)


which says that total expenditure cannot change in response to a change in prices, and
(Engel aggregation)
1

I=1
/
I
(p, n)-
&
(p, n) = 1, for / = 1, ..., 1, (***)
which says that total expenditure must change by an amount equal to any wealth change. /
I
(p, n) =
j
I
r
I
(p, n),n is the budget share of the consumers expenditure on good / given the prices and
income.
Proof : From Walras Law we have,
p x(p, n) = 0.
Dierentiating the above with respect to prices we obtain,
1

I=1
j
I
0r
I
(p, n)
0j

+r

(p, n) = 0, for / = 1, ..., 1.


By multiplying both sides by j

,n and multiplying and dividing all the terms except the last
one by r
I
(p, n), we obtain,
1

I=1
j
I
r
I
(p, n)
r
I
(p, n)
0r
I
(p, n)
0j

n
+
r

(p, n)j

n
= 0, for / = 1, ..., 1,
which yields (**).
A similar argument can be used to prove (***).
1.2.3 Solving the UMP
The UMP,
max
xA
n(x)
s.t. : p x _ n
26
is a nonlinear programming problem with one inequality constraint. From now on and due to a
monotonicity assumption on the preferences we will assume that the budget constraint is satised
with equality. The Lagrangian of the above problem is,
1(x) = n(x) +`(n p x) .
The Kuhn-Tucker necessary conditions say that if x
+
(p, n) is a solution to the UMP, then there
exists a Lagrange multiplier ` _ 0 such that,
01(x
+
)
0r

=
0n(x
+
)
0r

`j

_ 0, and r
+

01(x
+
)
0r

= 0, / = 1, ..., 1, (1)
and
01(x
+
)
0`
= (n p x) _ 0, and `
01(x
+
)
0`
= 0. (2)
(2) will always be satised with equality since preferences are monotonic, i.e.,
01(x
+
)
0`
= (n p x) = 0. ((2
t
))
If the solution is interior, i.e., x
+
(p, n) 0, then (1) is also satised with equality (see gure
13),
01(x
+
)
0r

=
0n(x
+
)
0r

`j

= 0. ((1
t
))
(1
t
) gives the famous condition that at the optimum '1o
i)
= j
i
,j
)
. In this case the gradient
vector is proportional to the price vector.
x
1
x
2
w/p
1
w/p
2
Slope = - (p
1/
p
2
) = -MRS
12
(x
*
)
x
*
p

u(x
*
)
Figure 13: Interior solution
27
Corner solutions, however, where at least one good is zero cannot be ruled out a priori. In such
a case, (1) becomes 01(x
+
),0r

_ 0 (which means that by lowering r

the objective increases, but


the decrease of r

has to stop at zero) and r


+

(01(x
+
),0r

) = 0 since r
+

= 0 (see gure 14).


x
1
x
2
w/p
1
w/p
2
Slope = -MRS
12
(x
*
)
x
*

u(x
*
)
p
p
Figure 14: Corner (boundary) solution
If the utility function is quasi-concave, then the Kuhn-Tucker rst order conditions are not only
necessary but also sucient.
EXAMPLES
1. Cobb-Douglas. n = r
o
1
r
b
2
. The solution to the UMP
max
(a
1
,a
2
)
n = r
o
1
r
b
2
s.t. : j
1
r
1
+j
2
r
2
= n
can take one of the following four forms: i) both goods are strictly positive (interior solution), ii)
and iii) only one is strictly positive and iv) both goods are zero (boundary solutions). The Lagrange
function is,
1 = r
o
1
r
b
2
+`(n j
1
r
1
j
2
r
2
) .
Since the Cobb-Douglas function is strictly quasi-concave the rst order K-T conditions are
necessary and sucient.
28
Case 1 . r
1
0 and r
2
0. The K-T conditions must be satised with equality,
01
0r
1
= 0 =ar
o1
1
r
b
2
`j
1
= 0, (1)
01
0r
2
= 0 =/r
o
1
r
b1
2
`j
2
= 0, (2)
01
0`
= 0 =n j
1
r
1
j
2
r
2
= 0. (3)
Divide (1) by (2) to obtain,
ar
2
/r
1
=
j
1
j
2
. (4)
Solve (4) with respect to r
2
and plug it into (3) and solve for r
1
. This yields,
r
1
=
an
(a +/) j
1
. (5)
Plug (5) back into (4) to obtain,
r
2
=
/n
(a +/) j
2
. (6)
Case 2 . r
1
0 and r
2
= 0. This is impossible to be optimal since the utility even when one
good is zero is zero and the consumer can clearly do better than that. For the same reason the
remaining cases will not be optimal.
(5) and (6) are the Marshallian demand functions.
2. Perfect substitutes. n = ar
1
+/r
2
. As before, the solution to the UMP
max
(a
1
,a
2
)
n = ar
1
+/r
2
s.t. : j
1
r
1
+j
2
r
2
= n
can take one of the following four forms: i) both goods are strictly positive (interior solution), ii)
and iii) only one is strictly positive and iv) both goods are zero (boundary solution). The Lagrange
function is,
1 = ar
1
+/r
2
+`(n j
1
r
1
j
2
r
2
) .
Since the utility function is quasi-concave the rst order K-T conditions are necessary and
sucient.
Case 1 . r
1
0 and r
2
0. The K-T conditions must be satised with equality,
01
0r
1
= 0 =a `j
1
= 0, (1)
29
01
0r
2
= 0 =/ `j
2
= 0, (2)
01
0`
= 0 =n j
1
r
1
j
2
r
2
= 0. (3)
By looking at the ratio (1)/(2) which is a,/ = j
1
,j
2
we can say that case 1 is valid only when
a
/
=
j
1
j
2
.
In any other case, the K-T conditions are not satised with equality. In this case the optimal
bundle must satisfy (3) and it is not unique.
Case 2 . r
1
0 and r
2
= 0. The K-T conditions are,
01
0r
1
= 0 =a `j
1
= 0, (4)
01
0r
2
< 0 =/ `j
2
< 0. (5)
From (4) ` = a,j
1
which if it is plugged into (5) yields the condition under which case 2 is
valid,
/
a
j
1
j
2
< 0 =
j
2
j
1

/
a
. (6)
The optimal solution in this case is,
r
1
=
n
j
1
and r
2
= 0.
Case 3 . r
1
= 0 and r
2
0. The K-T conditions are,
01
0r
1
< 0 =a `j
1
< 0, (7)
01
0r
2
= 0 =/ `j
2
= 0. (8)
From (8) ` = /,j
2
which if it is plugged into (7) yields the condition under which case 3 is
valid,
a
/
j
2
j
1
< 0 =
j
2
j
1
<
/
a
. (9)
The optimal solution in this case is,
r
1
= 0 and r
2
=
n
j
2
.
30
Obviously, case 4 will not be optimal.
The demand correspondence (notice that the solution is not a function) is,
r
1
=
_

_
&
j
1
, if
j
2
j
1

b
o
[0, n,j
1
] , if
j
2
j
1
=
b
o
0, if
j
2
j
1
<
b
o
and r
2
=
_

_
&
j
2
, if
j
2
j
1
<
b
o
[0, n,j
2
] , if
j
2
j
1
=
b
o
0, if
j
2
j
1

b
o
3. Concave preferences. n = ar
2
1
+/r
2
2
. The Lagrange function is,
1 = ar
2
1
+/r
2
2
+`(n j
1
r
1
j
2
r
2
) .
In this case, the utility function is not quasi-concave and the K-T conditions are only necessary.
We can guess the solution and then verify the K-T conditions.
Case 1. r
1
0 and r
2
0. This case will yield a minimum, not a maximum, since the worse
than set is now convex (see g. 15).
x
1
x
2
(c/a)
.5
(c/b)
.5
Better than
set
Figure 16: Non-convex preferences
Hence the solution will be a boundary solution.
The solution is,
r
1
=
_

_
&
j
1
, if
j
1
j
2
<
_
o
b
0 or
&
j
1
, if
j
1
j
2
=
_
o
b
0, if
j
1
j
2

_
o
b
and r
2
=
_

_
&
j
2
, if
j
1
j
2

_
o
b
0 or
&
j
2
, if
j
1
j
2
=
_
o
b
0, if
j
1
j
2
<
_
o
b
31
Now you can easily verify that the above solution satises the K-T conditions (provided that
we also choose the value of `).
Lets see how. It will helpful if we explicitly introduce all the constraints,
j
1
r
1
+j
2
r
2
= n ((`
1
))
r
1
_ 0 ((`
2
))
r
2
_ 0 ((`
3
))
where the `s are the respective multipliers. Now suppose that
j
1
j
2
<
_
o
b
and the optimal point
is,
x
+
=
_
n
j
1
, 0
_
.
The gradient of n is,
On(x
+
) =
_
2an
j
1
, 0
_
and the gradients of the constraints are,
j = (j
1
, j
2
), (1, 0) and (0, 1).
The second constraint is not binding (i.e., r
1
0) and therefore `
2
= 0. The K-T conditions
now say that there exist (`
1
, `
3
) _ 0 such that,
_
2an
j
1
, 0
_
= `
1
(j
1
, j
2
) +`
3
(0, 1) ,
or,
2an
j
1
= `
1
j
1
0 = `
1
j
2
`
3
.
The solution is,
`
1
=
2an
j
2
1
and `
3
=
2anj
2
j
2
1
32
x
1
x
2
x
*

u(x
*
)=(2a/p
1
,0)
p
(0,-1)
Figure 17
Figure 17 depicts graphically the solution. The K-T conditions say that: rst ignore the slack
constraint. Then the gradient of the utility must be below p and above the gradient of the binding
constraint (0, 1). In this case, there exists positive weights on p and (0, 1) that give On(x
+
) as
a linear combination.
Now notice that if we had written the Lagrange function without explicitly accounting for the
non-negativity constraints (as I did in a previous lecture), i.e.,
1 = ar
2
1
+/r
2
2
+`
1
(n j
1
r
1
j
2
r
2
),
then at the optimum we would have,
01
0r
1
= 2ar
1
`
1
j
1
= 0 (1
t
)
01
0r
2
= 2/r
2
`
1
j
2
< 0 (2
t
)
since `
2
= 0 and `
3
0. Hence, if the non-negativity constraints are not explicitly included then at
the optimum if a variable is zero the partial of the Lagrange function with respect to this variable
must be negative.
Claim: The Lagrange multiplier ` (on the budget constraint) gives the marginal value of relax-
ing the constraint (by increasing the consumers income).
Proof: Let x(p, n) 0. The change in the utility from a marginal increase in n is given by,
On(x(p, n)) 1
&
x(p, n) (*)
33
But from the K-T rst order conditions (for interior maximum) we have,
On(x(p, n)) = ` p (**)
Plug (**) into (*),
On(x(p, n)) 1
&
x(p, n) = ` p1
&
x(p, n)
which after using the Engel aggregation condition (p 1
&
x(p, n) = 1) becomes,
On(x(p, n)) 1
&
x(p, n) = `.

1.2.4 Continuity and dierentiability of the Marshallian demand correspondences


(MDC)
Continuity Recall that when preferences are strictly convex then the solution is a function not a
correspondence (multi-valued function). But when preferences are not strictly convex (e.g. perfect
substitutes) then the solution is in general a correspondence.
First, we will introduce a new notion of continuity (called upper hemicontinuity) that is used
whenever we have a correspondence and it coincides with the standard notion of continuity for
functions when the correspondence becomes a function.
Denition (upper hemicontinuity): A MDC x(p, n) is upper hemicontinuous (uhc) at
( p, n) if whenever (p
a
, n
a
) ( p, n) as : , x
a
x(p
a
, n
a
) for all : and x = limx
a
, we have,
x x( p, n).
34
p
x(p)
p
bar
p
n
x
bar
x
n
Figure 18: A correspondence that is NOT uhc
In gure 18 we depict a correspondence that is not uch. The sequence x
a
is in the correspondence
for all : but the limit x is not.
Proposition (uhc of the MDC): Suppose that n is a continuous utility function representing
locally non-satiated preferences. Then the solution to the UMP x(p, n) is an uhc correspondence
at all (p, n) 0.
Proof : By way of contradiction suppose that the demand correspondence x(p, n) is not uhc.
Take a sequence (p
a
, n
a
) ( p, n) and a sequence x
a
such that x
a
x(p
a
, n
a
) for all : but
limx
a
= ~ x , x( p, n). This implies the following about ~ x: i) either it is not aordable, or ii) it is
aordable but there exists another aordable bundle with a strictly higher utility. We will rst
show that it is indeed aordable.
Observe that p
a
x
a
_ n
a
for all : (since x
a
x(p
a
, n
a
) and therefore it must be aordable).
By taking limits as :
p
a
x
a
_ n
a
| | |
p ~ x _ n
Hence, ~ x 1
p, &
, i.e., it is aordable.
It must be then that there exists at least one other aordable bundle y ( p y _ n) such that
35
n(y) n(~ x). By the continuity of n there exists a bundle z close to y such that p z < n and
n(z) n(~ x). (*)
If now : is large enough p
a
z _ n
a
. Hence, z 1
p
n
,&
n and we must have n(x
a
) _ n(z) because
x
a
x(p
a
, n
a
). Taking limits as : the continuity of n implies that n(~ x) _n(z) contradicting
(*).
Remark: If x(p, n) is a function, then the above proposition says that the Marshallian demand
functions are continuous.
Dierentiability We assume that n is twice continuously dierentiable and \n(x) ,= 0 for all x.
Suppose that x(p, n) 0. This is the solution to the system of 1+1 equations in 1+1 unknowns,
\n(x) ` p = 0
n p x = 0.
By the implicit function theorem (IFT) (Theorem M.E.1 p.941) we know that the dierentiabil-
ity of x(p, n) depends on the Jacobian matrix of the system having a non-zero determinant. The
Jacobian of the system is,
_
1
2
n(x) p
p
T
0
_
Example: Suppose 1 = 2. The Lagrangian is,
1 = n(r
1
, r
2
) +`(n j
1
r
1
j
2
r
2
).
The system of the 3 equations is,
0n
0r
1
`j
1
= 0
0n
0r
2
`j
2
= 0
n j
1
r
1
j
2
r
2
= 0.
The Jacobian of this system is,
_
_
_
0
2
&
0a
2
1
0
2
&
0a
1
0a
2
j
1
0
2
&
0a
2
0a
1
0
2
&
0a
2
2
j
2
j
1
j
2
0
_
_
_
36
Since the F.O.C.s hold (i.e., 0n,0r

= `j

) the above matrix has a non-zero determinant if


and only if the determinant of the bordered Hessian

0
2
&
0a
2
1
0
2
&
0a
1
0a
2
0&
0a
1
0
2
&
0a
2
0a
1
0
2
&
0a
2
2
0&
0a
2
0&
0a
1
0&
0a
2
0

,= 0
The Cobb-Douglas utility function, for example, has a bordered Hessian with a strictly positive
determinant and we can conclude that the demand functions must be dierentiable.
In general, and following the same logic as in the example above, the Jacobian is non-zero i
the bordered Hessian

1
2
n(x) \n(x)
\n(x) 0

,= 0.
This condition means that the indierence set through x has a non-zero curvature at x (it is
not even innitesimally at). This condition is a slight strengthening of quasi-concavity.
1.2.5 The indirect utility function
The indirect utility function measures the highest utility the consumer can attain for any given
prices and income, i.e.,
(p, n) = n(x(p, n))
where x(p, n) are the Marshallian demand functions. This is the value function of the UMP.
Proposition (properties of the indirect utility function): The IUF is,
1. Homogeneous of degree zero in (j, n).
2. Strictly increasing in n and non-increasing in j

for any /.
3. Quasi-convex, that is, the set (p, n) : (p, n) _ is convex for any .
4. Continuous in (j, n).
Proof:
37
1. Since the demand functions are homogeneous of degree zero it follows that,
(cp, cn) = n(x(cp, cn)) = n(x(p, n)) = (p, n).
2. Easy to see.
3. Consider (p, n) and (p
t
, n
t
) such that (p, n) _ and (p
t
, n
t
) _ . Consider
(p
tt
, n
tt
) =
_
cp + (1 c)p
t
, cn + (1 c)n
t
_
for c [0, 1] .
We need to show that (p
tt
, n
tt
) _ . We have to show that for any x such that p
tt
x _ n
tt
we
must have n(x) _ . Notice rst that if, p
tt
x _ n
tt
then,
cp x + (1 c)p
t
x _ cn + (1 c)n
t
.
Thus, either px _ n or p
t
x _ n
t
or both. If the rst inequality holds, then n(x) _ (p, n) _ .
If the second inequality holds then n(x) _ (p
t
, n
t
) _ .
4. When preferences are strictly convex, we proved that the demand functions are continuous.
Hence (p, n) = n(x(p, n)) is a composition of two continuous functions and it should be contin-
uous. Note that (p, n) is continuous even when preferences are not strictly convex (Maximum
Theorem). We will not present a more general proof.
1.3 The expenditure minimization problem
The expenditure minimization problem (EMP) can be stated as follows,
min
x
p x
subject to : n(x) _ n.
So, in this problem the consumer chooses the cheapest consumption bundle provided that he
obtains a level of utility at least n. We denote the solution to the EMP by h(p, n). This is an 11
38
vector and is called the Hicksian (or compensated) demand functions. The minimized expenditure
is denoted by c(p, n) = p h(p, n) and is called the expenditure function.
Proposition (comparison between the UMP and the EMP):
i) If x
+
is optimal in the UMP when wealth is n 0, then x
+
is optimal in the EMP when the
required utility level is n(x
+
). Moreover, the minimized expenditure level in this EMP is exactly n.
ii) If h
+
is optimal in the EMP when utility is n n(0), then h
+
is optimal in the UMP when
the wealth is p h
+
. Moreover, the maximized utility level in this UMP is exactly n.
Proof: Straightforward.
What is implied by the above proposition is the following relationship between indirect utility
and expenditure functions,
c(p, (p, n)) = n and (p, c(p, n)) = n.
x
1
x
2
w/p
1
w/p
2
u
1
h = x
w/p
1

u
2
Figure 19
price of good 1
h, x
p
1
x(p,w)
h(p,u
1
)
p
1

h(p,u
2
)
Figure 20
Figures 19 & 20 depict the relationship between the UMP and the EMP graphically. Fix (p, n)
and solve both problems. The solutions coincide (h = x). If j
1
decreases the bundle that solves
the EMP must lie on the same indierence curve. The bundle that solves the UMP in general
39
lies on a higher indierence curve. The Marshallian (x(p, n)) and Hicksian (h(p, n)) demands
intersect at the point where (p, n) = n. That is, at the point where the maximum utility that
the consumer can attain given the prices and income is equal to the least utility that the consumer
must attain when he minimizes his expenditure. At this point the same consumption bundle solves
both problems. At any other point the two solutions diverge (as gure 20 clearly illustrates). Now
if n is not xed but it varies with (p, n) such that (p, n) = n (see gure 20) then we obtain the
following duality result between the Marshallian and the Hicksian demand functions,
x(p, n) = h(p, (p, n)).
Similarly, if n is not xed but it varies with (p, n) such that c(p, n) = n, then we obtain the
second duality between the Marshallian and the Hicksian demand functions,
h(p, n) = x(p, c(p, n)).
Example
Let n = r
o
1
r
1o
2
. The Lagrange function is,
1 = j
1
r
1
+j
2
r
2
+`( n r
o
1
r
1o
2
).
Since the constraint set is strictly convex the rst order conditions are also sucient and the
solution will be in the interior.
01
0r
1
= 0 =j
1
= `ar
o1
1
r
1o
2
,
01
0r
2
= 0 =j
2
= `(1 a)r
o
1
r
o
2
,
01
0`
= 0 = n = r
o
1
r
1o
2
.
Solving the above system of equations we obtain the Hicksian demand functions,
/
1
= n
_
aj
2
(1 a)j
1
_
(1o)
and /
2
= n
_
(1 a)j
1
aj
2
_
o
. (1)
The expenditure function is,
c(p, n) = j
1
/
1
+j
2
/
2
=
_
a
o
(1 a)
o1

j
o
1
j
1o
2
n. (2)
Checking for the duality
40
Recall that the Marshallian demand functions are,
r
1
=
an
j
1
and
(1 a)n
j
2
. (3)
Therefore, the Cobb-Douglas indirect utility function is,
(p, n) =
_
an
j
1
_
o
_
(1 a)n
j
2
_
1o
. (4)
Now it must be the case (verify it) that if we plug (4) into (1) (in place of n) we obtain (3).
And if we plug (2) in (3) (in place of n) we obtain (1).
Also, if we set (2) equal to n and solve with respect n we will obtain the indirect utility function.
Conversely, if we set (4) equal to n and solve with respect n we will obtain the expenditure function.
Proposition (properties of the expenditure function):
i) Homogeneous of degree 1 in j.
ii) Strictly increasing in n and non-decreasing in j

for all /.
iii) Concave in j.
iv) Continuous in j and n.
Proof:
i) We need to show that c(cp, n) = cc(p, n). Let h
+
be the bundle that solves the EMP at
(p, n). Now multiply all prices by c. First note that the constraint set x A : n(x) _ n is not
aected. The objective function p x gets only re-scaled and hence whatever was optimal under
p x must be optimal under (cp) x. Thus,
c(cp, n) = (cp) h
+
= cc(p, n).
ii) Proof straightforward.
41
iii) By the denition of a concave function, we have to show the following. Fix p and p
t
.
c(cp+(1 c)p
t
, n) _ cc(p, n) + (1 c)c(p
t
, n), for any c [0, 1] .
Let
p
tt
= cp+(1 c)p
t
,
and assume that h
tt
is optimal at p
tt
. That is,
c(p
tt
, n) = p
tt
h
tt
= cp h
tt
+ (1 c)p
t
h
tt
_ cc(p, n) + (1 c) c(p
t
, n).
The last inequality follows from the fact that h
tt
is not the expenditure minimizing bundle at
p and p
tt
.
iv) It follows from Berges maximum theorem. The idea is similar to the continuity of the
indirect utility function.
Proposition (properties of the expenditure function):
i) Homogeneous of degree 1 in j.
ii) Strictly increasing in n and non-decreasing in j

for all /.
iii) Concave in j.
iv) Continuous in j and n.
Proof:
i) We need to show that c(cp, n) = cc(p, n). Let h
+
be the bundle that solves the EMP at
(p, n). Now multiply all prices by c. First note that the constraint set x A : n(x) _ n is not
42
aected. The objective function p x gets only re-scaled and hence whatever was optimal under
p x must be optimal under (cp) x. Thus,
c(cp, n) = (cp) h
+
= cc(p, n).
ii) Proof straightforward.
iii) By the denition of a concave function, we have to show the following. Fix p and p
t
.
c(cp+(1 c)p
t
, n) _ cc(p, n) + (1 c)c(p
t
, n), for any c [0, 1] .
Let
p
tt
= cp+(1 c)p
t
,
and assume that h
tt
is optimal at p
tt
. That is,
c(p
tt
, n) = p
tt
h
tt
= cp h
tt
+ (1 c)p
t
h
tt
_ cc(p, n) + (1 c) c(p
t
, n).
The last inequality follows from the fact that h
tt
is not the expenditure minimizing bundle at
p and p
t
.
iv) It follows from Berges maximum theorem. The idea is similar to the continuity of the
indirect utility function.
Proposition (properties of the Hicksian demand functions): For all p 0, h(p, n) has
the following properties:
i) Homogeneous of degree zero in p,
h(cp, n) = h(p, n), for all c 0.
ii) No excess utility: For all h h(p, n), n(h) = n.
43
iii) Convexity/Uniqueness: If % are convex, then h(p, n) is a convex-set (i.e., the solution
might be a correspondence, but for any xed set of parameters the set of solutions is a convex set).
In other words, the correspondence is convex-valued. If % are strictly convex, then h(p, n) is a
singleton (the solution is a function, not a correspondence).
Proof: We will prove only iii). The rst two parts are easy.
First assume that preferences are convex. We want to show that h(p, n) is convex-valued. Fix
(p, n). We need to show that for any two h
t
and h
tt
that belong in h(p, n),
h
ttt
= ch
t
+ (1 c)h
tt
h(p, n), for all c [0, 1] .
Since h
t
and h
tt
h(p, n), it must be that
p h
t
= p h
tt
= n
+
,
and n(h
t
) _ n, n(h
tt
) _ n. That is, both bundles satisfy the constraint and attain the same level
of minimum expenditure. Since preferences are convex, or n is quasi-concave, it must be (by the
denition of a quasi-concave function) that, n(h
ttt
) _ minn(h
t
), n(h
tt
) _ n. Moreover,
p h
ttt
= cp h
t
+ (1 c)p h
tt
= cn
+
+ (1 c)n
+
= n
+
.
Therefore, h
ttt
h(p, n).
Now assume that preferences are strictly convex. We have to show that h(p, n) consists of only
one element. Suppose by way of contradiction that this is not the case and that there are two
bundles that solve the EMP problem, i.e., h
t
h(p, n) and h
tt
h(p, n). It must be that (as we
argued above)
p h
t
= p h
tt
= n
+
,
and n(h
t
) _ n, n(h
tt
) _ n. Since preferences are strictly convex,
n(h
ttt
) minn(h
t
), n(h
tt
) _ n,
where h
ttt
= ch
t
+ (1 c)h
tt
h(p, n), with c (0, 1). Now note that,
p h
ttt
= cp h
t
+ (1 c)p h
tt
= cn
+
+ (1 c)n
+
= n
+
.
Since n(h
ttt
) n and p h
ttt
= n
+
and by the continuity of n we can nd a bundle z in the
neighborhood of h
ttt
such that n(z) _ n and p z < n
+
. This contradicts the initial assumption
that both h
t
and h
tt
are optimal.
44
Proposition (compensated law of demand): For all p
t
, p
tt
the following holds,
_
p
tt
p
t
_

_
h(p
tt
, n) h(p
t
, n)

_ 0. (*)
Proof: From the denition of the expenditure function we have,
p
tt
h(p
tt
, n) _ p
tt
h(p
t
, n),
and
p
t
h(p
tt
, n) _ p
t
h(p
t
, n).
By subtracting the above two inequalities we obtain (*).
For example, when 1 = 2 (*) can be written as follows,
_
_
_ (j
tt
1
j
t
1
) , (j
tt
2
j
t
2
)
. .
(12 vector)
_
_
_
_
_
_
_
_
[/
1
(p
tt
, n) /
1
(p
t
, n)]
[/
2
(p
tt
, n) /
2
(p
t
, n)]
. .
(21 vector)
_
_
_
_
_
=
_
j
tt
1
j
t
1
_ _
/
1
(p
tt
, n) /
1
(p
t
, n)

+
_
j
tt
2
j
t
2
_ _
/
2
(p
tt
, n) /
2
(p
t
, n)

_ 0.
What this proposition says is that the Hicksian demand functions are downward sloping in each
goods own price. If, for instance, only one price changes then the above implies that,
_
j
tt

j
t

_ _
/

(p
tt
, n) /

(p
t
, n)

_ 0.
This is not necessarily true for the Marshallian demand functions, e.g. when a good is Gien
then the Marshallian demand is upward sloping. We will introduce a modied law of demand for
Marshallian demand functions later.
1.3.1 Relationships between demand indirect utility and expenditure functions
Proposition (relationship between Hicksian and expenditure function): For all p and
n the Hicksian demand h(p, n) is the derivative vector of the expenditure function with respect to
prices, i.e.,
h(p, n) = \
p
c(p, n).
45
Proof: Suppose p changes. Using the chain rule the change in expenditure can be written as,
\
p
c(p, n) = \
p
[p h(p, n)] = h(p, n) + [p 1
p
h(p, n)]
T
. (*)
Recall that the rst order condition for an interior solution to the EMP is,
p = `\n(h(p, n)). (**)
By substituting (*) into (**) we obtain,
\
p
c(p, n) = h(p, n) + [`\n(h(p, n)) 1
p
h(p, n)]
T
. (***)
Since the constraint n(h(p, n)) = n holds for all prices in the EMP, we must have \n(h(p, n))
1
p
h(p, n) = 0. Given this (***) becomes,
\
p
c(p, n) = h(p, n).

Proposition (more properties of the Hicksian demand functions): Suppose that %


are strictly convex and locally non-satiated and h(, n) is continuously dierentiable at (p, n) and
denote its derivative 1 1 matrix by 1
j
h(p, n). Then,
i) 1
p
h(p, n) = 1
2
c(p, n).
ii) 1
p
h(p, n) is a negative semi-denite matrix.
iii) 1
p
h(p, n) is a symmetric matrix. The symmetry implies that 0/

,0j
I
= 0/
I
,0j

for any
/, /.
iv) 1
p
h(p, n) p = 0.
Proof :
46
i) From the previous proposition we know that,
\
p
c(p, n) = h(p, n).
If we dierentiate it again with respect to p we get the result.
ii) + iii) We proved in a previous proposition that c(p, n) is a concave function and there-
fore it must have a negative semi-denite and symmetric Hessian matrix, 1
2
c(p, n). but from i)
1
p
h(p, n) = 1
2
c(p, n) and the result follows.
iv) In a previous proposition, we showed that h(p, n) is homogeneous of degree zero in p, i.e.,
h(cp, n) h(p, n) = 0, for all c 0.
By dierentiating the above with respect to c and setting c = 1 we obtain, 1
p
h(p, n) p = 0.
For example, when 1 = 2 the above becomes,
0/
1
0j
1
j
1
+
0/
1
0j
2
j
2
= 0,
0/
2
0j
1
j
1
+
0/
2
0j
2
j
2
= 0.

Denition (substitutes and complements): Two goods / and / are net substitutes at (p, n)
if 0/

,0j
I
_ 0 and net complements if 0/

,0j
I
_ 0. If, on the other hand, 0r

,0j
I
_ 0, then we
say that good / is a gross substitute for good /, and if 0r
I
,0j

_ 0, then we say that good / is a


gross substitute for good /. Similarly, if 0r

,0j
I
_ 0, then we say that good / is a gross complement
for good /, and if 0r
I
,0j

_ 0, then we say that good / is a gross complement for good /.


Remark 1: While the 1
p
h(p, n) matrix is symmetric and therefore 0/

,0j
I
= 0/
I
,0j

for
any /, /, the 1
p
x(p, n) matrix (as it will become clear later) is not symmetric. Hence, it may very
well be the case that 0r

,0j
I
_ 0 while 0r
I
,0j

_ 0.
Remark 2: When there are only two goods the above imply that they must be net substitutes.
With more than two goods the above implies that every good must have at least one net substitute
(WHY?).
47
1.3.2 Income and substitution eects
When the price of a good declines there are at least two conceptually separate reasons why we expect
some change in the quantity demanded. First that good becomes relatively cheaper compared to
other goods. We would expect the consumer to substitute the relatively cheaper good for the now
relatively more expensive ones. This is the substitution eect. At the same time, however, when
the price of a good declines, the consumers real income (purchasing power) increases which allows
him to change his purchases of all goods in an optimal way. This is the income eect. Intuition
tells us that we can decompose the total eect of a price change into these separate conceptual
categories. We will follow the method proposed by Hicks (1939). It starts with the observation
that the consumer achieves some level of utility at the original prices. The substitution eect is the
(hypothetical) change in the consumption that would occur if prices were to change to their new
levels but the maximum utility were kept the same as before the price change. This can eectively
happen if the consumer gets compensated for a price increase or gets de-compensated for a price
decrease. The income eect is the residual eect. This decomposition is captured by the Slutsky
equation.
Proposition (Slutsky equation): For all (p, n) and n = (p, n) we have,
0/

(p, n)
0j
I
=
0r

(p, n)
0j
I
+
0r

(p, n)
0n
r
I
(p, n) , for all / and /.
or equivalently in matrix notation,
1
p
h(p, n) = 1
p
x(p, n) +1
&
x(p, n) x(p, n)
T
.
Proof: Fix ( p, n) and let n = ( p, n). We know that n = c( p, n). From the fact that h( p, n) =
x( p, c( p, n)) we have that
/

( p, n) = r

( p, c( p, n)).
Dierentiate the above expression with respect to j
I
.,
0/

( p, n)
0j
I
=
0r

( p, c( p, n))
0j
I
+
0r

( p, c( p, n))
0n
0c( p, n)
0j
I
.
Using the fact that /

= 0c,0j

, the above becomes,


0/

( p, n)
0j
I
=
0r

( p, c( p, n))
0j
I
+
0r

( p, c( p, n))
0n
/
I
( p, n).
48
Finally, and since n = c( p, n) and /
I
( p, n) = x( p, c( p, n)) = x( p, n) we have,
0/

( p, n)
0j
I
=
0r

( p, n)
0j
I
+
0r

( p, n)
0n
r
I
( p, n).

If we re-arrange the terms of the Slutsky equation as follows,


0r

( p, n)
0j
I
. .
Total eect
=
0/

( p, n)
0j
I
. .
Substitution eect

0r

( p, n)
0n
r
I
( p, n)
. .
Income eect
,
we can obtain the total eect as the sum of the substitution and the income eect.
Consider the own price eect,
0r

( p, n)
0j

. .
Total eect
=
0/

( p, n)
0j

. .
Substitution eect

0r

( p, n)
0n
r

( p, n)
. .
Income eect
.
We have already proved (see the compensated law of demand) that the substitution eect is
negative, i.e., as j

and vice vera. We can deduce from the above the following relationships,
If good / is normal (i.e., 0r

,0n _ 0), then good / must be an ordinary good (i.e., 0r

,0j

_
0). This the law of demand which says that if a good is normal, then the demand curve is
downward sloping. The converse is not necessarily true. That is, if a good is ordinary it is
not necessarily normal, it may be inferior.
If good / is Gien (i.e., 0r

,0j

_ 0), then it must be inferior. The reverse is not necessarily


true. An inferior good might not be Gien.
The Slutsky substitution matrix can be written as,
o(p, n) =
_
_
_
_
0I
1
(p,&)
0j
1

0I
1
(p,&)
0j
L
.
.
.
.
.
.
.
.
.
0I
L
(p,&)
0j
1

0I
L
(p,&)
0j
L
_
_
_
_
=
=
_
_
_
_
0a
1
(p,&)
0j
1
+
0a
1
(p,&)
0&
r
1
(p, n)
0a
1
(p,&)
0j
L
+
0a
1
(p,&)
0&
r
1
(p, n)
.
.
.
.
.
.
.
.
.
0a
L
(p,&)
0j
1
+
0a
L
(p,&)
0&
r
1
(p, n)
0a
L
(p,&)
0j
L
+
0a
L
(p,&)
0&
r
1
(p, n)
_
_
_
_
49
Because o(p, n) = 1
p
h(p, n) (with n = (p, n)) then from Proposition (more properties
of the Hicksian demand functions) above it follows that the substitution matrix must be: i)
negative semi-denite, ii) symmetric and iii) satisfy o(p, n) p = 0.
Proposition (Roys identity): Suppose that the indirect utility is dierentiable at ( p, n)
0. Then,
x( p, n) =
1
\
&
( p, n)
\
p
( p, n) .
That is, for every / = 1, ..., 1
r

( p, n) =
0 ( p, n) ,0j

0 ( p, n) ,0n
.
Proof: Fix n = ( p, n). Since (p, c(p, n)) = n, when we dierentiate it with respect to p and
set p = p we obtain,
\
p
( p, c(p, n)) +
0 ( p, c(p, n))
0n
\
p
c(p, n) = 0.
Since \
p
c(p, n) = h(p, n) we have,
\
p
( p, c(p, n)) +
0 ( p, c(p, n))
0n
h(p, n) = 0.
Since n = c( p, n) and h( p, n) = x( j, n) we can obtain,
\
p
( p, n) +
0 ( p, n)
0n
x( j, n) = 0. (*)
Rearranging (*) we obtain the desired result.
1.3.3 A comprehensive example
Consider the constant elasticity of substitution (CES) utility function: n = (r
v
1
+r
v
2
)
1v
, where
r (, 1]. When r = 1 the indierence curves become linear (i.e., perfect substitutes). When
r 0 then the utility function represents the same preferences as a Cobb-Douglas (n = r
1
r
2
)
utility function. Finally, when r the indierence curves become right angles (i..e, Leontief
preferences, or perfect complements).
1. Solve the UMP.
The Lagrange function is,
50
1 = (r
v
1
+r
v
2
)
1v
+`(n j
1
r
1
j
2
r
2
) .
First observe that the utility function is quasi-concave (strictly for r < 1). To see this set
(r
v
1
+r
v
2
)
1v
= c and solve for r
2
to obtain an indierence curve, i.e.,
r
2
= (c
v
r
v
1
)
1v
0.
Dierentiate r
2
with respect to r
1
twice to obtain,
d
2
r
2
dr
2
1
= (1 r)
_
(c
v
r
v
1
)
(12r)
r
r
2(v1)
1
+ (c
v
r
v
1
)
(1r)
r
r
v2
1
_
_ 0.
Therefore, preferences are convex and the rst order necessary Kuhn-Tucker conditions are also
sucient.
01
0r
1
= 0 =
1
r
(r
v
1
+r
v
2
)
(1v)1
rr
v1
1
`j
1
= 0, (1)
01
0r
2
= 0 =
1
r
(r
v
1
+r
v
2
)
(1v)1
rr
v1
2
`j
2
= 0. (2)
(1)/(2) yields,
r
1
= r
2
_
j
1
j
2
_
1(v1)
. (3)
Plug (3) into the budget constraint and solve for r
2
. After some algebra we obtain,
r
2
=
nj
1(v1)
2
j
v(v1)
1
+j
v(v1)
2
. (4)
Now plug (4) back into (3) to obtain,
r
1
=
nj
1(v1)
1
j
v(v1)
1
+j
v(v1)
2
. (5)
(4) and (5) are the Marshallian demand functions.
To simplify the notation let / = r,(r 1). Then, / 1 = 1,(r 1). The Marshallian functions
can be rewritten as
r
1
=
nj
I1
1
j
I
1
+j
I
2
and r
2
=
nj
I1
2
j
I
1
+j
I
2
. (6)
Note that,
lim
vo
r
1
= lim
vo
r
2
=
n
j
1
+j
2
,
51
which is the solution to the UMP with a Leontief utility function (n = minr
1
, r
2
).
If we now plug (6) back into the utility function we obtain the indirect utility function,
(p, n) = n
__
j
I
1
+j
I
2
__
1I
. (7)
First, note that (7) is in n. Also, it is in j
1
and j
2
, i.e.,
0
0j
1
= n
_
j
I
1
+j
I
2
_
(1+I)I
j
(I1)
1
< 0 and
0
0j
2
= n
_
j
I
1
+j
I
2
_
(1+I)I
j
(I1)
2
< 0.
Moreover, it is homogeneous of degree zero in (p, n), i.e.,
(cp, cn) = cn
__
(cj
1
)
I
+ (cj
2
)
I
__
1I
= cn
_
c
I
_
j
I
1
+j
I
2
__
1I
=
= (cn)
_
c
I
_
1I
__
j
I
1
+j
I
2
__
1I
= cc
1
(p, n) = (p, n).
Finally, we also know that (p, n) is quasi-convex in income and prices but it is tedious to verify
it.
2. Solve the EMP.
The Lagrange function is,
1 = j
1
r
1
+j
2
r
2
+`
_
n (r
v
1
+r
v
2
)
1v
_
.
The F.O.C. are,
01
0r
1
= 0 =j
1
`
1
r
(r
v
1
+r
v
2
)
(1v)1
rr
v1
1
= 0, (8)
01
0r
2
= 0 =j
2
`
1
r
(r
v
1
+r
v
2
)
(1v)1
rr
v1
2
= 0. (9)
(8)/(9) yields,
r
1
= r
2
_
j
1
j
2
_
1(v1)
. (10)
Plug (10) into the utility constraint n = (r
v
1
+r
v
2
)
1v
and solve for r
2
. After some algebra we
obtain,
/
2
= n
_
j
I
1
+j
I
2
_
(1k)
k
j
I1
2
. (11)
Now plug (11) back into (10) to obtain,
/
1
= n
_
j
I
1
+j
I
2
_
(1k)
k
j
I1
1
. (12)
52
(11) and (12) are the Hicksian demand functions. If we plug (11) and (12) back into
j
1
r
1
+j
2
r
2
we obtain (after some simplications) the expenditure function,
c(p, n) = n
_
j
I
1
+j
I
2
_
1I
. (13)
The expenditure function is clearly increasing in n and increasing in prices, i.e.,
0c
0j
1
= n
_
j
I
1
+j
I
2
_
(I1)I
j
(I1)
1
0 and
0c
0j
2
= n
_
j
I
1
+j
I
2
_
(I1)I
j
(I1)
2
0.
It is also homogeneous of degree 1 in prices, i.e.,
c(cp, n) = n
_
(cj
1
)
I
+ (cj
2
)
I
_
1I
= n
_
c
I
_
1I
_
j
I
1
+j
I
2
_
1I
= c n
_
j
I
1
+j
I
2
_
1I
= cc(p, n).
The expenditure function is concave in prices, but it is tedious to verify it with this example.
Numerical example:
The graph below plots r
1
and /
1
against j
1
when n = 10, n = 3, j
2
= 5 and r = .5. Notice
that the two demand functions intersect at j
1
= 10. At this point (p, n) = n. The Marshallian
demand function is above the Hicksian for j
1
< 10.
Figure 21: Hicksian and Marshallian demand functions
3. Verify all the dualities.
53
a. Duality between indirect utility and expenditure functions.
c(p, (p, n)) = n. (14)
and
(p, c(p, n)) = n. (15)
These conditions imply that for a xed vector p, c( p,) and ( p,) are inverses to one another.
Lets see this in our example. From (13) we have
n
_
j
I
1
+j
I
2
_
1I
= n == n = n
__
j
I
1
+j
I
2
__
1I
,
same as (7) the indirect utility function. Similarly, we can obtain the expenditure function from
the indirect utility function. From (7) we have,
n
__
j
I
1
+j
I
2
__
1I
= n ==n = n
_
j
I
1
+j
I
2
_
1I
,
same as (13), the expenditure function.
b. Duality between Marshallian and Hicksian demand functions.
h(p, n) = x(p, c(p, n)),
and
x(p, n) = h(p, (p, n)).
The Marshallian demand functions are,
r
1
=
nj
I1
1
j
I
1
+j
I
2
and r
2
=
nj
I1
2
j
I
1
+j
I
2
,
and the Hicksian are,
/
1
= n
_
j
I
1
+j
I
2
_
(1k)
k
j
I1
1
and /
2
= n
_
j
I
1
+j
I
2
_
(1k)
k
j
I1
2
.
Plug the indirect utility function (p, n) = n
__
j
I
1
+j
I
2
_
1I
into the Hicksian demand func-
tions in place of n. This yields (we only look at good 1),
/
1
= n
__
j
I
1
+j
I
2
__
1I
_
j
I
1
+j
I
2
_
(1k)
k
j
I1
1
=
nj
I1
1
_
j
I
1
+j
I
2
_ = r
1
.
Now plug the expenditure function c(p, n) = n
_
j
I
1
+j
I
2
_
1I
into the Marshallian demand func-
tions in place of n. This yields (again we only look at good 1)
r
1
=
n
_
j
I
1
+j
I
2
_
1I
j
I1
1
j
I
1
+j
I
2
= n
_
j
I
1
+j
I
2
_
(1k)
k
j
I1
1
= /
1
.
54
c. h(p, n) = \c(p, n).
With two goods the above duality can be written as,
_
/
1
/
2
_
=
_
0c
0j
1
0c
0j
2
_
.
0c
0j
1
=
0
_
n
_
j
I
1
+j
I
2
_
1I
_
0j
1
= n
_
j
I
1
+j
I
2
_
(1k)
k
j
I1
1
= /
1
,
and
0c
0j
2
=
0
_
n
_
j
I
1
+j
I
2
_
1I
_
0j
2
= n
_
j
I
1
+j
I
2
_
(1k)
k
j
I1
2
= /
2
.
4. Slutsky substitution matrix
The substitution matrix is,
o(j
1
, j
2
, n) =
_
0I
1
0j
1
0I
1
0j
2
0I
2
0j
1
0I
2
0j
2
_
=
_
0a
1
0j
1
+
0a
1
0&
r
1
,
0a
1
0j
2
+
0a
1
0&
r
2
0a
2
0j
1
+
0a
2
0&
r
1
,
0a
2
0j
2
+
0a
2
(p,&)
0&
r
2
_
=
=
_
_
_
&j
k2
1
j
k
2
(I1)
(j
k
1
+j
k
2
)
2
,
&j
k1
1
j
k1
2
(I1)
(j
k
1
+j
k
2
)
2

&j
k1
1
j
k1
2
(I1)
(j
k
1
+j
k
2
)
2
,
&j
k
1
j
k2
2
(I1)
(j
k
1
+j
k
2
)
2
_
_
_
First note that the diagonal elements are negative, verifying the compensated law of demand.
Moreover o(j
1
, j
2
, n) is symmetric, i.e., 0/
1
,0j
2
= 0/
2
,0j
1
. Also o(j
1
, j
2
, n) is negative semi-
denite, i.e., the diagonal elements are negative and the determinant is zero.
We also know that o(p, n) p = 0, i.e.,
_
_
_
&j
k2
1
j
k
2
(I1)
(j
k
1
+j
k
2
)
2
,
&j
k1
1
j
k1
2
(I1)
(j
k
1
+j
k
2
)
2

&j
k1
1
j
k1
2
(I1)
(j
k
1
+j
k
2
)
2
,
&j
k
1
j
k2
2
(I1)
(j
k
1
+j
k
2
)
2
_
_
_
_
j
1
j
2
_
=
_
0
0
_
.
5. Substitutes and complements
Net. Using the Hicksian demand functions we can see that the two goods are net substitutes,
i.e., 0/
1
,0j
2
= 0/
2
,0j
1
_ 0. This holds regardless of the value of /.
55
Gross. By taking the cross partial derivatives of the Marshallian demand function we obtain,
0r
1
0j
2
=
0r
2
0j
1

nj
(I1)
2
j
(I1)
2
/
_
j
I
1
+j
I
2
_
2
.
Note that it is not necessarily true that
0a
1
0j
2
=
0a
2
0j
1
, although in this example it happens to be the
case. Recall that / = r,(r 1) and r (, 1]. Hence, if r _ 0, / _ 0 and the two goods are gross
substitutes. If, on the other hand, r _ 0, then / _ 0 and the two goods are gross complements.
This conrms what we would expect to obtain, since when r the two goods are perfect
(gross) complements (n = minr
1
, r
2
), while at r = 1 they are perfect (gross) substitutes (n =
r
1
+r
2
).
6. Roys identity
r
1
=
0,0j
1
0,0n
=
n
_
j
I
1
+j
I
2
_
(1+I)I
j
(I1)
1
_
j
I
1
+j
I
2
_
1I
=
nj
I1
1
j
I
1
+j
I
2
.
r
2
=
0,0j
2
0,0n
=
n
_
j
I
1
+j
I
2
_
(1+I)I
j
(I1)
2
_
j
I
1
+j
I
2
_
1I
=
nj
I1
2
j
I
1
+j
I
2
.
END OF EXAMPLE.
Dualities The gure below summarizes all the dualities.
56
UMP EMP
h(p,u)
Slutsky equation
v(p,w)
R
o
y

s

i
d
e
n
t
i
t
y
e(p,u)
h

=

e
e(p,v(p,w)) = w
v(p,e(p,u)) = u
h
(
p
,
u
)

=

x
(
p
,
e
(
p
,
u
)
)
x
(
p
,
w
)

=

h
(
p
,
v
(
p
,
w
)
)
DUAL PROBLEMS
x(p,w)
Figure 22: Dualities
If we start from the UMP, we can solve it to derive the Marshallian demand functions. Then
we can derive the indirect utility function. If we invert it we obtain the expenditure function. By
dierentiating the expenditure function, we obtain the Hicksian demand functions. Conversely, if
we start from the EMP, we can solve it to derive the Hicksian demand functions. Using them we
obtain the expenditure function, whose inverse is the indirect utility function. Using Roys identity
we derive the Marshallian demand functions.
1.3.4 An important conclusion
Based on our results so far we can say that: If the Marshallian demand functions x(p, n) are
generated by preferences which satisfy our AXIOMS 1-5 and are continuously dierentiable then
they must satisfy the following,
1. Homogeneous of degree zero.
2. Satisfy Walras Law (and as a consequence Cournot and Engel aggregation).
3. Substitution matrix o(p, n) which is: i) negative semi-denite, ii) symmetric and iii) satisfy
o(p, n) p = 0.
When researchers try to estimate a system of demand equations they want to know whether it
is: i) consistent with rational behavior and ii) an outcome of a well-dened utility maximization
57
problem. One way is to start from a well-dened utility function derive the Marshallian demands
and then estimate them. This however is restrictive and one may want to begin from a system of
demand functions which t the problem under investigation better, rather than to begin from a
utility function. Is it true that this system is compatible with utility maximization? The answer
is yes, as long as the system satises 1., 2. and 3. above. All one has to do is to check for these
three properties. In other words, 1.,2., and 3. are not only necessary but are also sucient for the
existence of rational generating preferences. This is the converse to the conclusion above and is
called the integrability problem, which we will study next.
1.4 Integrability
Theorem (integrability theorem): A continuously dierentiable system x(p, n) is the sys-
tem of demand functions generated by some strictly increasing, strictly quasi-concave and continu-
ous utility function, if and only if the system satises: Walras Law and the substitution matrix is
symmetric and negative semi-denite.
In other words, Walras Law, symmetry and negative semi-deniteness are not only the con-
sequences of preference-based demand theory, but also are all of its consequences. So, if we start
with a system of equations, we will be certain that it is utility generated if it satises Walras Law,
symmetry and negative semi-deniteness.
Next, I demonstrate, with the aid of an example, how we can go from the demand functions
back to utility.
Consider the following system,
r
1
=
cn
j
1
and r
2
=
(1 c)n
j
2
. (1)
Is it compatible with utility maximization? If so, what is the utility function?
The answer to the rst question is armative, by invoking the integrability theorem. Verify
that (1) satises Walras Law, and the Slutsky substitution matrix is symmetric and negative
semi-denite.
58
Lets now nd the utility which generated this demand system.
N Recovering the expenditure function from the demand system
Fix n and let n = (j
1
, j
2
, n). Given this we know that x(p, n) = h(p, n). Using h(p, n) =
\
p
c(p, n) we obtain,
0c(j
1
, j
2
, n)
0j
1
= /
1
(j
1
, j
2
, n) = r
1
(j
1
, j
2
, n) =
c n
j
1
=
cc(j
1
, j
2
, n)
j
1
, (2)
and
0c(j
1
, j
2
, n)
0j
2
= /
2
(j
1
, j
2
, n) = r
2
(j
1
, j
2
, n) =
(1 c) n
j
2
=
(1 c) c(j
1
, j
2
, n)
j
2
. (3)
We should nd c(j
1
, j
2
, n) that solves (2) and (3). We are looking for a solution to the system
of partial dierential equations. We can re-write it as,
0 ln[c(j
1
, j
2
, n)]
0j
1
=
c
j
1
and
0 ln[c(j
1
, j
2
, n)]
0j
2
=
(1 c)
j
2
.
The above can be written as,
ln[c(j
1
, j
2
, n)] = clnj
1
+c(j
2
, n), (4)
and
ln[c(j
1
, j
2
, n)] = (1 c) lnj
2
+c(j
1
, n). (5)
We must choose c(j
1
, n) and c(j
2
, n) such that (4) and (5) hold simultaneously. This yields,
ln[c(j
1
, j
2
, n)] = clnj
1
+ (1 c) lnj
2
+c( n),
or,
c(j
1
, j
2
, n) = j
c
1
j
(1c)
2
c( n). (6)
We must ensure that c is increasing in n so we can choose c( n) to be any strictly increasing
function in n.
This is the expenditure function when n = r
c
1
r
1c
2
(Cobb-Douglas). Recall that with this
utility,
c(j
1
, j
2
, n) =
_
c
c
(1 c)
(c1)
_
j
c
1
j
(1c)
2
n. (7)
59
Now if we choose c( n) =
_
c
c
(1 c)
(c1)
_
n then (6) is equivalent to (7). Note that it does not
matter what c( n) we end up choosing since we know that implied demand behavior is independent
of strictly increasing transformations.
N Recovering the indirect utility function from the expenditure function
From (7) we have,
_
c
c
(1 c)
(c1)
_
j
c
1
j
(1c)
2
n = n == = nj
c
1
j
(1c)
2
C,
where C is the constant.
N Recovering the utility function from the indirect utility function
Idea: Fix x
0
, and let n(x
0
) denote the utility the consumer attains, if he consumes x
0
. The
following inequality must hold for any p,
(p, p x
0
) _ n(x
0
).
The above inequality says that when the income is p x
0
= n
0
, then the highest utility that the
consumer can attain is at least n(x
0
), since x
0
does not have to be the best bundle at these prices
and income. Moreover, there will be one price vector p
0
such that,
(p
0
, p
0
x
0
) = n(x
0
).
In other words when the (p, p x
0
) is minimized it is equal to the utility at x
0
. This can be
done for any x and therefore we can recover the utility if we solve,
n(x) = min
p
(p, p x).
Since is homogeneous of degree zero in p and n we have: (p, p x) = (cp, c(p x)). Now
if you take c = 1, (p x) we have (p, p x) = (p, (p x) , 1). The above minimization problem
can be equivalently stated as,
n(x) = min
p
(p, 1)
subject to: p x = 1.
60
In our example,
min1j
c
1
j
(1c)
2
s.t. : j
1
r
1
+j
2
r
2
= 1.
Recall that is quasi-convex in prices and income. Hence, this minimization problem can be
solved by solving the rst order necessary and sucient Kuhn-Tucker conditions. The Lagrange
function is,
1 = j
c
1
j
(1c)
2
+`(j
1
r
1
+j
2
r
2
1) .
The F.O.C.s are,
01
0j
1
= cj
c1
1
j
(1c)
2
+`r
1
= 0 (1)
01
0j
2
= (1 c) j
c
1
j
(1c)1
2
+`r
2
= 0. (2)
(1)/(2) yields,
cj
2
(1 c) j
1
=
r
1
r
2
==j
2
=
(1 c) j
1
r
1
cr
2
. (3)
Plug (3) into j
1
r
1
+j
2
r
2
= 1 to obtain,
j
1
r
1
+
(1 c) j
1
r
1
cr
2
r
2
= 1. (4)
Solving (4) with respect to j
1
yields,
j
1
=
c
r
1
. (5)
Plug (5) back into (3) to obtain,
j
2
=
1 c
r
2
. (6)
Finally, plug (5) and (6) into the objective function j
c
1
j
(1c)
2
. This yields,
_
c
r
1
_
c
_
1 c
r
2
_
(1c)
=
r
c
1
r
(1c)
2
c
c
(1 c)
(1c)
= r
c
1
r
(1c)
2
,
where = c
c
(1 c)
(1c)
. This utility function represents Cobb-Douglas preferences.
So, we started from a system of demand functions and we recovered the utility that generated
this demand system.
61
1.5 Welfare analysis
Consider a consumer with preferences % which satisfy axioms 1-5. Suppose that prices are
initially p
0
and they change to p
1
. We would like to measure the change in the consumers welfare.
Think for example a situation where the price change is due to taxation and the authorities would
like to measure the impact of the tax on the consumers.
Let
n
0
= (p
0
, n) and n
1
= (p
1
, n)
be the highest levels utility that the consumer can attain at p
0
and p
1
respectively, given that his
income is n. Also note, from the duality, that
c(p
0
, n
0
) = c(p
1
, n
1
) = n.
That is, regardless of the price level the consumer will spend n. Consider the following measure
which captures the change in the welfare.
Compensating Variation (CV),
C\ (p
0
, p
1
, n) = c(p
1
, n
1
) c(p
1
, n
0
) = n c(p
1
, n
0
). (1)
Lets rst interpret (1). Consider gure 23
x
1
x
2
w/p
1
0
w/p
2
0
w/p
1
1
u
0
u
1
CV
62
Figure 23
Initially the prices are j
0
1
, j
0
2
. At that level of prices and income n the consumers highest level
of utility is n
0
. Now suppose good 1 becomes cheaper, i.e., j
0
1
j
1
1
. The new highest level of utility
is n
1
. Clearly the consumer is better o. You can think of this as a technological improvement
which lowered the cost of producing good 1 and therefore its price. The question which arises is
what is the maximum that the consumer is willing to pay for this technology? If he pays exactly
the C\ then at the new prices (i.e., with the technology) is as well o as he was without the
technology. Hence, the C\ is the maximum that he is willing to pay. If he pays zero his welfare
increase, measured in money, is equal to the C\ . Or think of a tax which increases the price. The
C\ is the amount of money that the consumer must receive in order to be indierent between tax
and no tax.
The CV has an interesting interpretation in terms of the Hicksian demand function. Suppose
that only the price of good 1 changes from j
0
1
to j
1
1
. All the other prices are held xed.
C\ (p
0
, p
1
, n) = n c(p
1
, n
0
) (2)
= c(p
0
, n
0
) c(p
1
, n
0
)
=
_
j
0
1
j
1
1
/
1
(j
1
, j
1
, n
0
)dj
1
.
So the CV is the area under the Hicksian demand from j
0
1
to j
1
1
. If there are no wealth eects
then this will also be the area under the Marshallian. However, this is not true when wealth eects
are present. In this case the Hicksian and the Marshallian demands are not the same. What we
usually use in economics to measure changes in welfare is the area under the Marshallian demand.
This is the standard consumer surplus (CS), but it is just an approximation of the correct
measure which is the area under the Hicksian.
If preferences are quasi-linear then there are no income eects and the CV coincides with the
CS. For example, if n = lnr
1
+ r
2
, then preferences are quasi-linear. [Find the demand functions
and verify that there is no income eect for good 1. The Hicksian coincides with the Marshallian].
Example: The deadweight loss from commodity taxation.
Suppose the government taxes good 1. In particular, it imposes a tax t per unit. Hence the
new per unit price is j
1
1
= j
0
1
+t. The total revenue raised is T = tr
1
(p
1
, n). An alternative to this
63
commodity tax is a lump-sum tax of T directly on the consumers wealth. Under which taxation
scheme is the consumer better o?
We will show that the revenue the government raises (i.e., T = tr
1
(p
1
, n)) is less than what
the government should pay the consumer to keep his utility level equal to that before the tax
imposition. This implies that the government could extract more from the consumer by imposing
a lump-sum tax. This is called the deadweight loss from commodity taxation. Lets see this.
When the government imposes the tax the consumption of good 1 is: r
1
(p
1
, n) and the tax
revenue is: T = tr
1
(p
1
, n). If the government were to compensate the consumer to keep his utility
level the same as before the tax, it would run a decit equal to,
C\ (p
0
, p
1
, n) t/
1
(p
1
, n) =
_
j
0
1
+t
j
0
1
/
1
(j
1
, j
1
, n
0
)dj
1
t/
1
(j
0
1
+t, j
1
, n
0
)
=
_
j
0
1
+t
j
0
1
_
/
1
(j
1
, j
1
, n
0
) /
1
(j
0
1
+t, j
1
, n
0
)

dj
1
0,
since /
1
is strictly decreasing in j
1
. This implies that the government would run a decit.
Numerical example
Consider the CES utility function with r = .5. Also, let j
1
= 1, j
2
= 1 and n = 10. The
quantity demanded (using the Marshallian demands) is: r
1
= r
2
= 5. The indirect utility, using
= n
_
j
I
1
+j
I
2
_
1I
from a previous lecture is: 20. Now suppose that the government imposes a 10% per unit tax on good
1. Its price becomes j
1
= 1.1. The indirect utility is, 19.1 and the consumption of good 1 (using
the Mashallian demand) is, r
1
= 4.33. The governments revenue is: T = tr
1
= 10%4.33 = .433.
The CV (using the expenditure function) is,
c(p
0
, n
0
) c(p
1
, n
0
) = 10 10.476 = .476.
The decit (deadweight loss) is
.433 .476 = .043 < 0.
That is, consumers loose more than what the government gains. The reduction in the consumer
surplus is,
Co =
_
1.1
1
10j
I1
1
(j
I
1
+ 1)
dj
1
= .4652,
which is quite close to .476.
64
2 THE THEORY OF THE FIRM
We assume that a rm produces a single product j using multiple inputs z = (.
1
, ..., .
1
). The
input vector as well as the output must be non-negative. We will describe the rms technology in
terms of a production function j = )(z), which means that j units of output can be produced using
the input vector z and the technology )(). The production function is analogous to the utility
function. Unlike the utility function, the production function is not derived (like we derived the
utility function from preferences) but it is taken as given.
Properties of the production function: The production function ) : 1
1
+
1
+
is continu-
ous strictly increasing and strictly quasi-concave on 1
1
+
, and )(0) = 0.
Marginal product: 0)(z),0.

is the marginal product of input /. It is analogous to marginal


utility.
Isoquant: Q(j) = z _ 0 : )(z) = j. This set consists of all the combinations of inputs that
produce a level of output equal to j. It is analogous to the indierence curve.
Marginal rate of technical substitution (MRTS):
'1To
i)
(z) =
0)(z),0.
i
0)(z),0.
)
.
With two inputs, MRTS is the slope of the isoquant through (.
1
, .
2
). It is analogous to the
MRS.
Elasticity of substitution: For a production function )(z) the elasticity of substitution
between inputs i and , at the point z is dened as,
o
i)
=
d ln(.
)
,.
i
)
d ln()
i
(z),)
)
(z))
=
d (.
)
,.
i
)
.
)
.
i
)
i
(z),)
)
(z)
d ()
i
(z),)
)
(z))
where )
i
and )
)
are the marginal products of inputs i and ,. It measures how a small percentage
change of the isoquant slope changes the input ratio.
Example: Consider the Cobb-Douglas production j = .
o
1
.
b
2
. The MRTS is,
'1To =
)
1
)
2
=
a.
2
/.
1
.
Then,
d ()
1
(z),)
2
(z))
d (.
2
,.
1
)
=
a
/
.
65
The elasticity of substitution is,
o
12
=
/
a
a.
2
,/.
1
.
2
,.
1
= 1.
Returns to scale:
1. Constant returns to scale (CRS) if )(tz) = t)(z) for all t 0 and all z.
2. Increasing returns to scale (IRS) if )(tz) t)(z) for all t 1 and all z.
3. Decreasing returns to scale (DRS) if )(tz) < t)(z) for all t 1 and all z.
Example: In the Cobb-Douglas case, the production function exhibits constant returns to scale
if a +/ = 1, increasing if a +/ 1 and decreasing if a +/ < 1.
CES production function: j = )(z) = [a
1
.
v
1
+a
2
.
v
2
]
bv
.
)(tz) = [a
1
(t.
1
)
v
+a
2
(t.
2
)
v
]
bv
= t
v(bv)
)(z) = t
b
)(z).
If / = 1 then we have CRS. If / 1 (/ < 1) then we have IRS (DRS).
2.1 Prot maximization (Competitive rm)
Let j denote the per-unit output price and w = (n
1
, ..., n
1
) is the input-price vector. The
rm, given the technology and the prices (output and input), chooses the combination of inputs to
maximize the prots (revenue minus cost). The problem is stated as follows,
max
(j,z)0
jj w z
subject to : )(z) _ j.
Since it is not optimal to produce j < )(z), then the constraint is satised by equality and the
problem becomes,
max
z0
j)(z) w z.
66
If , z
+
is optimal then it must satisfy
j
0)(z
+
)
0.

_ n

and with equality if .


+

0.
If the production function is concave, then the rst order conditions are also sucient. For
example, when the inputs are two at an interior solution the MRTS must be equal to the input
price ratio, i.e.,
0)(z),0.
1
0)(z),0.
2
=
n
1
n
2
.
The optimal choice of output j
+
= j(j, w) is called the rms output supply function. The
optimal choice of inputs z
+
= z(j, w) is called the input demand functions. The prot function
is:
(j, w) = jj
+
w z
+
.
Properties of the prot function: The prot function (j, w) satises:
1. Increasing in j.
2. Decreasing in w.
3. Homogeneous of degree 1 in (j, w).
4. Convex in (j, w).
5. Hotellings lemma:
0(j, w)
0j
= j(j, w) and
0(j, w)
0n

= .

(j, w).
Proof:
1. Assume an interior solution so that the rst order conditions are satised with equality,
j
0)(z
+
)
0.

= n

, for all /.
The prot function is,
(j, w) = j)(z
+
) w z
+
.
67
For simplicity suppose that 1 = 2. Dierentiate (j, w) with respect to j and use the rst
order conditions (envelope theorem).
0(.
1
(j, w), .
2
(j, w))
0j
= j +j
0j
0.
1
0.
1
0j
+j
0j
0.
2
0.
2
0j
n
1
0.
1
0j
n
2
0.
2
0j
= j 0.
2. Can be proved similarly by dierentiating the prot function with respect to the input prices.
3. First notice that the input demand functions are homogeneous of degree zero in (j, w). That
is, the optimal choice (the tangency) does not change when the output price and inputs change by
the same proportion.
(tj, tw) = (tj) )(z (tj, tw)) (tw) z (tj, tw) = (tj) )(z (j, w)) (tw) z (j, w)
= t(j, w).
4. Let j and . maximize prots at j and n, and let j
t
and z
t
maximize prots at j
t
and w
t
.
Dene
j
tt
= tj + (1 t)j
t
and
w
tt
= tw+ (1 t)w
t
for 0 _ t _ 1. Let j
tt
and z
tt
maximize prots at j
tt
and w
tt
. Then
(j, w) = jj w z _jj
tt
w z
tt
,
(j
t
, w
t
) = j
t
j
t
w
t
z
t
_j
t
j
tt
w
t
z
tt
.
Therefore for 0 _ t _ 1 we have,
t(j, w) + (1 t)(j
t
, w
t
) _
_
tj + (1 t)j
t
_
j
tt

_
tw+ (1 t)w
t
_
z
tt
= (j
tt
, w
tt
).
5. See the proof of part 1.
Properties of output supply and input demand functions.
1. Homogeneity of degree zero in:
j(tj, tw) = j(j, w), for all t 0,
.

(tj, tw) = .

(tj, tw), for all t 0 and / = 1, ..., 1.


68
2. The (1 + 1) (1 + 1) substitution matrix
_
_
_
_
_
_
0j(j,w)
0j
0j(j,w)
0&
1

0j(j,w)
0&
L

0:
1
(j,w)
0j

0:
1
(j,w)
0&
1

0:
1
(j,w)
0&
L
.
.
.
.
.
.
.
.
.
.
.
.

0:
L
(j,w)
0j

0:
L
(j,w)
0&
1

0:
L
(j,w)
0&
L
_
_
_
_
_
_
is symmetric and positive semi-denite.
Proof:
1. We proved it above.
2. By dierentiating the prot function and using the rst order conditions we can obtain,
j(j, w) =
0(j, w)
0j
and .

(j, w) =
0(j, w)
0n

, / = 1, ..., 1.
It can now be easily veried that the substitution matrix is equal to the Hessian of the prot
function. The latter is convex and therefore the substitution matrix must be positive semi-denite.
In particular the diagonal elements are all positive.
Example. j = .
o
1
.
b
2
where a +/ < 1 DRS.
max
(:
1
,:
2
)
j.
o
1
.
b
2
n
1
.
1
n
2
.
2
.
By assuming an interior solution the F.O.C. reduces to,
aj.
o1
1
.
b
2
n
1
= 0 (1)
/j.
o
1
.
b1
2
n
2
= 0. (2)
Solve (1 & 2) with respect to .
1
and .
2
to obtain the input demand functions,
.
1
= j
1(1ob)
_
/
n
2
_
b(1ob)
_
a
n
1
_
(1b)(1ob)
. (3)
.
2
= j
1(1ob)
_
/
n
2
_
(1o)(1ob)
_
a
n
1
_
o(1ob)
. (4)
69
The output supply function is obtained if we substitute (3 & 4) into the production function.
j = j
(o+b)(1ob)
_
/
n
2
_
b(1ob)
_
a
n
1
_
o(1ob)
. (5)
The prot function is,
= j
1(1ob)
n
o(1ob)
1
n
b(1ob)
2
(/)
b(1ob)
(a)
o(1ob)

j
1(1ob)
n
o(1ob)
1
n
b(1ob)
2
(/)
b(1ob)
(a)
(1b)(1ob)

j
1(1ob)
n
o(1ob)
1
n
b(1ob)
2
(/)
(1o)(1ob)
(a)
o(1ob)
= j
1(1ob)
n
o(1ob)
1
n
b(1ob)
2
_
(/)
b(1ob)
(a)
o(1ob)

(/)
b(1ob)
(a)
(1b)(1ob)
(/)
(1o)(1ob)
(a)
o(1ob)
_
= j
1(1ob)
n
o(1ob)
1
n
b(1ob)
2
(/)
b(1ob)
(a)
o(1ob)
[1 a /] .
2.2 Cost minimization
The rm chooses its level of inputs to minimize the cost of producing a certain level of output.
Formally, the problem is stated as follows,
min
z
w z
subject to : )(z) _ j.
This problem is equivalent to the expenditure minimization problem. The solution to this
problem are the conditional input demand functions, z(w, j) and the value function is the cost
function c(w, j). To solve this problem, quasi-concavity is enough. The necessary (and sucient if
the production function is quasi-concave) condition for an interior solution is,
0)(z),0.
i
0)(z),0.
)
=
n
i
n
)
,
for any two inputs.
Properties of the cost function (similar to the properties of the expenditure func-
tion)
70
1. Zero when j = 0.
2. Continuous on its domain.
3. For all w 0, strictly increasing and unbounded above in j.
4. Increasing in w.
5. Homogeneous of degree 1 in w.
6. Concave in w.
7. Shephards lemma:
\
w
c(w, j) = z(w, j).
Properties of the conditional input demands (similar to the properties of the Hick-
sian demand functions)
1. z(w, j) is homogeneous of degree zero in w.
2. The substitution matrix, dened
_
_
_
_
0:
1
(w,j)
0&
1

0:
1
(w,j)
0&
L
.
.
.
.
.
.
.
.
.
0:
L
(w,j)
0&
1

0:
L
(w,j)
0&
L
_
_
_
_
is symmetric and negative semi-denite.
Example: Cobb-Douglas
minn
1
.
1
+n
2
.
2
s.t. : .
o
1
.
b
2
_ j.
Since the production is strictly increasing the constraint is satised with equality. The Lagrange
function is,
1 = n
1
.
1
+n
2
.
2
+`
_
j .
o
1
.
b
2
_
.
The FOCs are,
01
0.
1
= n
1
`a.
o1
1
.
b
2
= 0,
71
01
0.
2
= n
2
`/.
o
1
.
b1
2
= 0.
Taking the ratio of the two FOCs we obtain,
n
1
n
2
=
a.
2
/.
1
. (1)
Solve (1) with respect to .
2
, plug it back into the constraint and solve for .
1
. This yields,
j = .
o
1
_
/n
1
.
1
an
2
_
b
==.
1
= j
1(o+b)
_
an
2
/n
1
_
b(o+b)
.
This is the conditional demand of input 1. Similarly, the conditional demand of input 2 is,
.
2
= j
1(o+b)
_
/n
1
an
2
_
o(o+b)
.
The cost function is,
c(w, j) = n
1
j
1(o+b)
_
an
2
/n
1
_
b(o+b)
+n
2
j
1(o+b)
_
/n
1
an
2
_
o(o+b)
= j
1(o+b)
n
o(o+b)
1
n
b(o+b)
2
_
_
a
/
_
b(o+b)
+
_
/
a
_
o(o+b)
_
.
Note that if a + / 1, then the cost function is concave in j. In this case (increasing returns
to scale) the average cost function,
C(j) =
c(j)
j
= j
(1ob)(o+b)
n
o(o+b)
1
n
b(o+b)
2
_
_
a
/
_
b(o+b)
+
_
/
a
_
o(o+b)
_
,
is decreasing in j. That is, when the production exhibits increasing returns to scale, the per-unit
cost decreases as the production increases. The marginal cost,
'C(j) =
0c
0j
=
j
(1ob)(o+b)
(a +/)
n
o(o+b)
1
n
b(o+b)
2
_
_
a
/
_
b(o+b)
+
_
/
a
_
o(o+b)
_
is also decreasing.
When a+/ = 1 (constant returns to scale), then the cost function is linear and the average and
marginal cost functions are constant. Finally, when a +/ < 1, then the cost function is convex and
the average and marginal cost functions are increasing.
The input shares, dened as,
:

=
n

(j, w)
c(j, w)
,
72
are, for the Cobb-Douglas problem, equal to
:
1
= a and :
2
= /.
Cost and conditional input demands when production is homothetic
1. When the production function is strictly increasing, strictly quasi-concave and homothetic,
(a) the cost function is multiplicatively separable in input prices and output and can be written
c(j, w) = /(j)c(w, 1), where /(j) is strictly increasing and c(w, 1) is the unit cost function,
or the cost of 1 unit of output;
(b) the conditional input demands are multiplicatively separable in input prices and output and
can be written z(w, j) = /(j)z(w, 1), where /(j) is strictly increasing and z(w, 1) is the
conditional input demand for 1 unit of output.
2. When the production function is homogeneous of degree c 0,
(a) c(j, w) = j
1c
c(w, 1);
(b) z(w, j) = j
1c
z(w, 1).
Proof: We begin by proving part 1a. Let 1 denote the production function. Because it is
homothetic, it can be written as 1(z) = )(q(z)), where ) is strictly increasing and q is homogeneous
of degree 1. Assume that the image of 1 is all of 1
+
. So )
1
(j) 0, for all j 0. For some j 0
let t = )
1
(1),)
1
(j) 0. Now note that,
)(q(z)) _ j ==q(z) _ )
1
(j) ==q(tz) _t)
1
(j) = )
1
(1) ==) (q(tz)) _ 1.
Therefore, we can express the cost function associated with 1 as follows,
c(w, j) = min
z
w z s.t. )(q(z)) _ j
= min
z
w z s.t. ) (q(tz)) _ 1
=
1
t
min
z
n tz s.t. ) (q(tz)) _ 1
=
1
t
min
x
n x s.t. ) (q(x)) _ 1
=
)
1
(j)
)
1
(1)
c(w, 1),
73
where in the second to last inequality we let x = tz.
Because ) is strictly increasing implies that )
1
is as well, the desired result holds for all j 0.
Part 1b is follows from Shephards lemma. Part 2 follows by mimicking the proof of part 1.
Verify the above proposition using the Cobb-Douglas production function.
2.2.1 The short-run cost function
One can argue that in the short-run (say for a time horizon less than a year) some inputs are xed,
due to the rms prior commitments and obligations. I will derive the short-run cost function using
the Cobb-Douglas production function and assuming that the second input is xed. The rms cost
minimization problem becomes,
min
:
1
n
1
.
1
+n
2
.
2
s.t. .
o
1
.
b
2
= j.
From the constraint, .
1
=
_
j, .
b
2
_
1o
. This is the short-run conditional input demand function.
The short-run cost function is,
:c = j
1o
.
bo
2
n
1
+n
2
.
2
.
So, the total cost is decomposed into two parts: the variable cost j
1o
.
bo
2
n
1
and the xed
cost n
2
.
2
.
74
The long-run cost function is the lower envelope of the entire family of short-run cost functions.
In the gure above, I have plotted the long-run cost function and two short-run cost function when
.
2
= 1 and .
2
= 3 respectively. Also a = / = .4 and n
1
= n
2
= 1. The short-run cost functions
are always above the long-run except at one point where they coincide. This is because, in the
short-run the rm has more constraints and therefore its cost must be higher than the long-run
cost.
Recall also from undergraduate Micro that in the short-run the rm will shut-down if the price
is below the minimum average variable cost. If the price is too low that the rm cannot cover its
variable cost, then it is better o to shut down and lose the xed cost. Otherwise, it is protable
(even if it is losing money) in the short-run to stay in business. In the above example, the AVC is,
\ C(j) =
j
1o
.
bo
2
n
1
j
= j
(1o)o
.
bo
2
n
1
.
In this case, the minimum AVC is zero and therefore the rm never shuts down.
Prot maximization using the cost function
We can re-state the rms prot-maximization problem as follows,
max
j
jj c(w, j).
The necessary FOC for j
+
to be prot maximizing is then,
j
0c(w, j
+
)
0j
_ 0, with equality if j
+
0.
In words, at an interior optimum, price equals marginal cost. This condition gives the output
supply function. If c(w, j) is convex in j then the rst order condition is also sucient.
Example: Cobb-Douglas.
max
j
jj j
1(o+b)
n
o(o+b)
1
n
b(o+b)
2
_
_
a
/
_
b(o+b)
+
_
/
a
_
o(o+b)
_
.
If a +/ < 1, then the cost function is convex in j and the FOC is sucient. The FOC is,
j = 'C(j)
j =
j
(1ob)(o+b)
(a +/)
n
o(o+b)
1
n
b(o+b)
2
_
_
a
/
_
b(o+b)
+
_
/
a
_
o(o+b)
_
.
75
The output supply function is,
j = j
(o+b)(1ob)
n
o(1ob)
1
n
b(1ob)
2
_
(a +/),
_
_
a
/
_
b(o+b)
+
_
/
a
_
o(o+b)
__
(o+b)(1ob)
.
Exercise: Consider a rm with the cost function,
c(n
1
, n
2
, j) = j
2
(n
1
+n
2
) .
a) Sketch the rms marginal and average total cost curves and its output supply function.
C(j) = j (n
1
+n
2
) .
'C(j) = 2j (n
1
+n
2
) .
The output supply function is given by the rms prot-maximizing rst order condition,
j = 'C(j) ==j = 2j (n
1
+n
2
) .
b) Sketch the input demand for input .
1
against n
1
.
From Shephards lemma,
0c
0n
1
= .
1
==j
2
.
So, the conditional input demand, depends only on the level of output. The same is true for
the other input.
Recovering the technology from the cost function
This is similar to the integrability problem in consumer theory. By setting,
j
2
(n
1
+n
2
) = :
and solving with respect to j we obtain,
j =
_
:
(n
1
+n
2
)
. (*)
76
The above is the value function of the following problem,
max
z
j = )(z)
s.t. : n
1
.
1
+n
2
.
2
= :.
That is, the rm maximizes the level of production subject to a budget constraint. This is
the same as the utility maximization problem. Although, we are not terribly interested in this
problem, when we analyze rm behavior, (*) can be used (same as in consumer theory) to recover
the production function.
min
(&
1
,&
2
)
_
1
(n
1
+n
2
)
s.t. : n
1
.
1
+n
2
.
2
= 1.
If .
1
.
2
, then n
1
= 0 and n
2
= 1,.
2
. In this case, the objective function is:
_
.
2
.
If .
1
< .
2
, then n
2
= 0 and n
1
= 1,.
1
. In this case, the objective function is:
_
.
1
.
Combining the above two cases, the production function is,
_
min.
1
, .
2
.
This represents a Leontief-type production technology with decreasing returns to scale.
77
3 CHOICE UNDER UNCERTAINTY
3.1 Expected utility theory
Let C denote the set (nite) of possible outcomes, : = 1, ..., .
Denition: A simple lottery 1 is a list 1 = (j
1
, j
2
, ..., j
.
), with j
a
_ 0, for all : and

a
j
a
= 1, where j
a
is the probability of outcome : occurring.
A simple lottery can be represented as a point in the ( 1) dimensional simplex
=
_
p 1
.
+
: j
1
+ +j
.
= 1
_
.
In the two-dimensional simplex, each vertex is a lottery where one outcome is certain. Each
edge represents a lottery where two outcomes are certain and the other is impossible.
For example when C = 1, 2, 3, the two dimensional simplex is an equilateral triangle. We
normalize the altitude of the triangle to 1. Then the sum of the perpendiculars from any point to
the three sides is equal to one. It is common to depict a lottery as a point in the triangle. The
probability of outcome : occurring is the perpendicular distance from the point to the side opposite
the vertex labeled :, see graph below.
1
2
3
p
1
p
3
p
2
78
Denition: Given 1 simple lotteries 1
I
=
_
j
I
1
, j
I
2
, ..., j
I
.
_
, / = 1, ..., 1 and probabilities
c
I
_ 0,

c
I
= 1, the compound lottery (1
1
, ..., 1
1
; c
1
, ..., c
1
) is the risky alternative that yields
the simple lottery 1
I
with probability c
I
, / = 1, ..., 1.
Any compound lottery can be reduced to a simple lottery. The probability of outcome : in the
reduced lottery is,
j
a
= c
1
j
1
a
+ +c
1
j
1
a
.
Therefore the reduced lottery 1 of any compound lottery (1
1
, ..., 1
1
; c
1
, ..., c
1
) can be ob-
tained
1 = c
1
1
1
+ +c
1
1
1
.
Preferences over lotteries
Suppose there are three possible outcomes C = 1, 2, 3 . One can view these outcomes as
money. Now consider the following three lotteries,
1
1
= (1, 0, 0) , 1
2
=
_
1
4
,
3
8
,
3
8
_
and 1
3
=
_
1
4
,
3
8
,
3
8
_
.
The rst lottery yields 1 with probability 100%. The second yields 1 with probability 25%, 2
with probability 3/8 and so on.
Now lets construct a compound lottery. Suppose that each lottery occurs with probability 1/3.
The reduced lottery is,
^
1 =
1
3
1
1
+
1
3
1
2
+
1
3
1
3
=
_
1
2
,
1
4
,
1
4
_
.
Now consider two more lotteries,
1
4
=
_
1
2
,
1
2
, 0
_
and 1
5
=
_
1
2
, 0,
1
2
_
.
Each one of the above lotteries occurs with probability 1/2. The reduced lottery is,
~
1 =
1
2
1
4
+
1
2
1
5
=
_
1
2
,
1
4
,
1
4
_
.
The decision maker views
^
1 and
~
1 the two lotteries as equivalent. We denote by L the set of all
simple lotteries over C. We assume that the decision maker has a rational (complete + transitive)
preferences % on L.
79
Denition: % on 1 is continuous if for any 1, 1
t
and 1
tt
1 the sets,
_
c [0, 1] : c1 + (1 c)1
t
% 1
tt
_
[0, 1]
and
_
c [0, 1] : 1
tt
% c1 + (1 c)1
t
_
[0, 1]
are closed.
Consider the following set of outcomes: C = $1, 000, $10, death. Most would agree that
$1, 000 ~ $10 ~death. Now consider the following lexicographic type of preferences where the
decision maker puts safety rst (i.e., no death) and then money. Consider the following lottery:
1 = (1, 0, 0). This lottery is strictly preferred to 1
t
= (0, 1, 0). Both lotteries guarantee safety
and the rst one yields more money. Now suppose that the rst lottery is mixed with a third one
1
tt
= (0, 0, 1) using a weight c for 1. Now we have,
1
t
% c1 + (1 c)1
tt
, for all c < 1
but
c1 + (1 c)1
tt
~ 1
t
, for c = 1.
Hence the set,
_
c [0, 1] : 1
t
% c1 + (1 c)1
tt
_
[0, 1]
is not closed. In particular the c = 1 is missing from the set. This sudden change of preferences is
avoided with the continuity assumption.
The completeness, transitivity and continuity axioms imply the existence of a utility function
representing %, a function l : L 1 such that
1 % 1
t
==l(1) _ l(1
t
).
The existence of such a function is guaranteed if the above axioms are satised, as we saw in
consumer theory.
A utility function over lotteries, however, is conceptually cumbersome and many people have
trouble articulating such preferences in a consistent way. The way alternatives are framed is very
important.
Furthermore, a lottery does have a special structure-it is a probability distribution over possible
outcomes. The question is can the utility function over lotteries be deduced from a utility function
over outcomes? To be able to answer armatively, we need one more axiom.
80
Denition (Independence axiom): % on 1 satises the independence axiom if for all 1, 1
t
and 1
tt
and c (0, 1) we have
1 % 1
t
==c1 + (1 c) 1
tt
% c1
t
+ (1 c) 1
tt
.
In words, the preference ordering between 1 and 1
t
does not depend on the third lottery 1
tt
.
The independence axiom is not present in consumer theory. There, there is no reason to believe
that the preference ordering over 2 bundles of goods should be independent of the quantities of
other goods that the consumer consumes. In choice under uncertainty, however, it is natural to
assume that the presence of 1
tt
will not aect the preference over the other two, since 1
tt
will never
be consumed together with 1 or 1
t
, but rather either one or the other. In contrast, goods are
consumed together.
Denition (Expected utility form) : l : L 1 has the expected utility form if there
is an assignment of numbers (n
1
, ..., n
.
) to the outcomes such that for every simple lottery
1 = (j
1
, ..., j
.
) 1 we have,
l(1) = n
1
j
1
+ +n
.
j
.
.
A utility function l : L 1 with the expected utility form is called a von Neumann-
Morgenstern expected utility function. l(1) =

a
n
a
j
a
is a linear function in probabilities.
Proposition: A utility function l : L 1 has the expected utility form if and only if it is
linear, that is, if and only if satises the property that
l
_
1

I=1
c
I
1
I
_
=
1

I=1
c
I
l(1
I
),
for any 1 lotteries 1
I
L and probabilities(c
1
, ..., c
1
) _ 0, with

c
I
= 1.
The expected utility form is preserved only by increasing linear transformations (cardinal prop-
erty).
Proposition: l : L 1 is a von Neumann-Morgenstern utility function for % on L. Then
~
l : L 1 is another von Neumann-Morgenstern utility function for % if and only if there exists
, 0 and such that
~
l(1) = ,l(1) +, for all 1 L.
81
Note that if % on L is representable by l() that has the expected utility form, then it is
continuous in the probabilities, since linear functions are continuous.
Exercise 6.B.2: Show that if % on 1 is represented by a utility function l() that has the
expected utility form, then % satisfy the independence axiom.
Assume that preferences are represented by a von Neumann-Morgenstern expected utility func-
tion
l(1) =
.

a=1
j
a
n
a
, for all 1 L.
Let 1, 1
t
and 1
tt
L and c (0, 1). Then 1 % 1
t
if and only if
.

a=1
j
a
n
a
_
.

a=1
j
t
a
n
a
.
This is equivalent to,
c
_
.

a=1
j
a
n
a
_
+ (1 c)
.

a=1
j
tt
a
n
a
_ c
_
.

a=1
j
t
a
n
a
_
+ (1 c)
.

a=1
j
tt
a
n
a
.
The above inequality holds if and only if,
c1 + (1 c) 1
tt
% c1
t
+ (1 c) 1
tt
.
Hence, 1 % 1
t
if and only if
c1 + (1 c) 1
tt
% c1
t
+ (1 c) 1
tt
.
The independence axiom holds.
3.2 Expected utility theorem
Theorem: Suppose that preferences on L are rational and satisfy the continuity and indepen-
dence axioms. Then % admit a utility representation of the expected utility form.
4
4
There is, by now, a large number of theories, called non-expected utility theories, that challenge
some of the axioms imposed by expected utility theory. For a recent survey of these theories see
http://www.nottingham.ac.uk/~lezcs/pdf_les/STARMER_JEL.PDF
82
That is, we can assign a number n
a
to each outcome : = 1, ..., in such a manner that for any
two lotteries 1, 1
t
we have,
1 % 1
t
if and only if
.

a=1
n
a
j
a
_
.

a=1
n
a
j
t
a
.
Remark: One possible interpretation of the alternative outcomes is to view them as consump-
tion bundles. Then, n can be dened over the consumption set A. So in our standard consumer
theory formulation n : A 1. A lottery gives the probability with which each bundle occurs.
Fact 1: Indierence curves are straight lines. This means that for all 1, 1
t
with 1 ~ 1
t
==
c1+(1 c)1
t
~ 1, for all c [0, 1] . Otherwise, the independence axiom would be violated.
To see this, suppose that 1 ~ 1
t
, but .51
t
+ .51 ~ 1. This implies that IC are not straight
lines. This is equivalent to .51
t
+ .51 ~ .51 + .51. Since 1 ~ 1
t
the independence axiom
implies that .51
t
+.51 ~ .51 +.51, a contradiction.
Fact 2: Indierence curves are parallel to each other. Suppose not. Take 1 ~ 1
t
and assume
that
1
3
1 +
2
3
1
tt
%
1
3
1
t
+
2
3
1
tt
,
which implies the two indierence curves are not parallel. But, again, this violates the
independence axiom.
3.3 Money lotteries and risk aversion
Denote amounts of money by a continuous variable r. A monetary lottery is: 1 : 1 [0, 1] .
1(r) is the probability that the realized payo is less than or equal to r. So, 1() is the cumulative
probability function (cdf), or distribution function and )() is the density function. We know that
1(r) =
_
a
o
)(t)dt.
We will work with distribution functions. Let L be the set of all distribution functions over
[0, ). The interval [0, ) is the set of outcomes (money in this case) and each lottery is represented
by a distribution function.
83
The expected utility theorem tells us that there is an assignment of utility values n(r) to non-
negative amounts of money with the property that any 1() can be evaluated by l() of the form,
l(1) =
_
n(r)d1(r).
Note that l() is linear in 1(). In other words, l(a1
1
+ /1
2
) = al(1
1
) + /l(1
2
), for any 1
and a, /.
The strength of the expected utility is that it preserves the useful expectation form while making
the utility of monetary lotteries sensitive not only to the mean but also to higher moments of the
distribution of monetary payments.
We call l() von Neumann-Morgenstern utility function and n() Bernoulli utility function.
3.3.1 Risk aversion and its measurement
Denition (Risk aversion): A decision maker is risk averse if for all 1() the degenerate
lottery that yields
_
rd1(r) with certainty is at least as good as 1() itself.
In other words,
Risk averse ==n
__
rd1(r)
_
_
_
n(r)d1(r).
This is Jensens inequality and it holds as long as n() is a concave function. So, concavity of
the Bernoulli utility function will be used to model a risk averse decision maker.
Denition (Certainty equivalent): Given n(), the certainty equivalent of 1(), denoted by
c(1, n) is the amount of money for which the individual is indierent between the gamble 1() and
the certain amount c(1, n), that is,
n(c(1, n)) =
_
n(r)d1(r).
Consider an example where there are two possible outcomes 1 and 3 which are equally likely.
The certainty equivalent is less than 2 and it represents the certain amount of money that the
decision maker has to receive in order to be indierent between that and the gamble which yields
1 or 3 with 50% likelihood, see gure below.
84
x
1 2 3
u(1)
u(2)
u(3)
c(F,u)
.5(1)+.5u(3)
Proposition: The decision maker is an expected utility maximizer with a Bernoulli utility
function n() dened on amounts of money. Then the following properties are equivalent,
1. Decision maker is risk averse.
2. n() is concave.
3. c(1, n) _
_
rd1(r).
Insurance problem Consider a decision maker with wealth n who runs the risk of losing $1
with probability . He wishes to buy insurance. One unit of insurance costs < 1 and pays 1
dollar if the loss occurs. If c units are bought the wealth will be
n c, if no loss
n c 1 +c, if loss.
The decision makers expected wealth is,
n 1 +c( ) .
85
Following the expected utility theory, the decision makers expected utility is,
(1 ) n(n c) +n(n c 1 +c).
The decision maker chooses c (how much insurance to buy) to maximize his expected utility,
i.e.,
max
c
(1 ) n(n c) +n(n c 1 +c).
If c
+
is optimum it must satisfy,
(1 ) n
t
(n c
+
) + (1 ) n
t
(n 1 +c
+
(1 )) _ 0,
and with equality if c
+
0.
No insurance. c
+
= 0. The rst order condition becomes,
(1 ) n
t
(n) + (1 ) n
t
(n 1) .
Note that if we assume that the decision maker is risk averse, then n must be concave and
n
t
(n 1) n
t
(n).
If (1 ) (1 ), then at c
+
= 0 the rst order condition is strictly positive which implies
that c
+
will be optimally increased. This is true as long as _ . It will also be positive even if
< depending on the concavity of n. But now lets assume that = , that is is actuarially
fair.
The rst order condition becomes,
n
t
(n c
+
) +n
t
(n 1 +c
+
(1 )) .
At c
+
= 1 the rst order condition is satised with equality. Since n is concave the objective
function is also concave and the rst order condition is also sucient. A risk averse decision maker
when the cost of insurance is actuarially fair he will insure completely.
Note that if c
+
= 1 the decision makers income is n 1 regardless of which event occurs.
His income is completely smoothed and the risk has been eliminated.
86
Portfolio choice problem An investor has wealth n 0. There are two assets: a riskless asset
with a return of 1 dollar per dollar invested and a risky asset with a random return . per dollar
invested. The distribution function of . is 1(.). Assume that,
_
.d1(.) 1.
That is, the risky assets expected return is higher than the safe assets return. Denote by c
and , the amounts of wealth invested in the risky and safe asset respectively. For any realization
of . the portfolio pays
c. +,,
where c+, = n. We assume the existence of a von Neumann-Morgenstern expected utility function.
The investors maximization problem is,
max
(c,o)0
_
n(c. +,)d1(.)
subject to : c +, = n.
Alternatively, the maximization problem, using , = n c, can be written as,
max
c0
_
n(n +c(. 1))d1(.)
subject to : 0 _ c _ n.
If c
+
is optimal, the necessary rst order condition is,
c(c
+
) =
_
n
t
(n +c
+
(. 1)) (. 1) d1(.)
_
_
_
_ 0, if c
+
< n
_ 0, if c
+
0.
Suppose c
+
= 0.
c(0) =
_
n
t
(n) (. 1) d1(.) =
_
n
t
(n).d1(.)
_
n
t
(n)d1(.)
= n
t
(n)
__
.d1(.) 1
_
0
since the risky asset has a higher expected return than 1. Therefore, c
+
= 0 cannot satisfy the rst
order condition.
If risk is actuarially favorable (i.e., expected return of the risky asset greater than the riskless
return), as it is the case here, then a risk averse investor will always accept at least a small amount
of it. This is captured by the optimal c being greater than zero. In other words, part of the optimal
87
portfolio consists of the risky asset. Note that this would not be true if the expected return of the
risky asset was equal to that of the riskless. In this case, a risk averse investor would avoid the
risky asset, as it happened in the insurance problem.
A numerical example
Let n() = ()
c
, with 0 < c _ 1. Also let n = 10 and . be distributed uniformly in [0, 2.2].
The riskless assets expected return is 1 per 1 dollar invested. The table below gives the optimal
portfolio for dierent values of c.
c .1 .2 .3 .4 .5 .6 .7 1
c 2.763 3.1 3.53 4.085 4.839 5.9 7.43 10
Measurement of risk aversion
Denition: Given n() the Arrow-Pratt coecient of absolute risk aversion at r is dened as,
r

(r) =
n
tt
(r)
n
t
(r)
.
Risk version is related to the curvature of the Bernoulli utility function n.
Example 1: n = r
c
. The Arrow-Pratt coecient is,
r

=
c(c 1)r
c2
cr
c1
=
(1 c)
r
.
Example 2: n = c
oa
, a 0. The Arrow-Pratt coecient is,
r

= a.
3.3.2 Comparisons across individuals
We can say that n
2
() is more risk averse than n
1
() if,
1. r

(r, n
2
) _ r

(r, n
1
), for all r.
88
2. There exists an increasing concave function c() such that n
2
(r) = c(n
1
(r)). That is, n
2
is a
concave transformation of n
1
.
3. c(1, n
2
) _ c(1, n
1
), for any 1(). The more risk averse decision maker is willing to accept a
lower certain amount of money in order not to enter the gamble.
Portfolio choice problem (continued) We will formally prove the following claim: The more
risk averse investor (investor 2) invests less in the risky asset than the less risk averse investor
(investor 1).
Investor 1s problem is,
max
&c0
_
n
1
(n +c
1
(. 1))d1(.).
Assuming an interior solution,
c
1
(c
+
1
) =
_
n
t
1
(n +c
+
1
(. 1)) (. 1) d1(.) = 0. (1)
For investor 2 the FOC is,
c
2
(c
+
2
) =
_
n
t
2
(n +c
+
2
(. 1)) (. 1) d1(.) = 0. (2)
We would like to show that c
+
2
< c
+
1
.
By dierentiating c
2
with respect to c
2
we obtain,
c
t
2
(c
2
) =
_
n
tt
2
(n +c
2
(. 1)) (. 1)
2
d1(.) < 0,
since n
tt
< 0 due to the concavity assumption. If we show that,
c
2
(c
+
1
) < 0,
this would imply that c
+
2
< c
+
1
. This is because investor 2 would have an incentive to choose an c
2
less than c
1
, since by doing this his objective function increases.
Now use the following,
n
2
(r) = c(n
1
(r)).
89
Then,
n
0
2
(r) = c
t
(n
1
(r))n
t
1
(r).
Using the above two equations (2) can be re-written as follows,
c
2
(c
+
1
) =
_
c
t
(n
1
(n +c
+
1
(. 1)) n
t
1
(n +c
+
1
(. 1)) (. 1) d1(.). (4)
Note that (4), without the c
t
(n
1
(n +c
+
1
(. 1)) term, is zero, see (1) investor 1s FOC. Now
c
t
is a decreasing function since c is concave. Therefore, the high realizations of . (higher than 1)
get a smaller weight compared to the low realizations of . (less than 1). Hence (4) is going to be
negative and the result follows.
Numerical example (continued)
Consider two investors with the following Bernoulli utility functions: n
1
(r) =
_
r and n
2
(r) =
r
.2
. The Arrow-Pratt coecients of absolute risk aversion are,
r

(r, n
1
) =
.5
r
< r

(r, n
2
) =
.8
r
, for all r.
Investor 2 is more risk averse than investor 1. We know that c
+
1
= 4.839 c
+
2
= 3.1, which
conrms the general result we derived above.
3.3.3 Comparison across wealth levels
Denition (Decreasing absolute risk aversion): n() exhibits decreasing absolute risk aver-
sion if r

(r, n) is a decreasing function of r.


As the wealth increases, individuals take more risks. For example, n(r) =
_
r exhibits decreasing
absolute risk aversion, while n(r) = c
oa
, exhibits constant absolute risk aversion.
Denote the increments or decrements to wealth by ..
Proposition: The following properties are equivalent:
90
1. n() exhibits decreasing absolute risk aversion.
2. Whenever r
2
< r
1
, n
2
(.) = n(r
2
+.) is a concave transformation of n
1
(.) = n(r
1
+.).
3. For any risk 1(.), the certainty equivalent of the lottery formed by adding risk . to wealth
level r, given by the amount c
a
at which n(c
a
) =
_
n(r + .)d1(.), is such that (r c
a
) is
decreasing in r. That is, the higher r is, the less the individual is willing to pay to get rid of
the risk.
Exercise 6.C.8. n() exhibits decreasing absolute risk aversion. Show that in the portfolio
choice problem, the optimal allocation between the safe and the risky assets places an increasing
amount of wealth in the risky asset as the investors wealth (n) increases. In other words, the risky
asset is a normal good.
Consider two levels of wealth n
1
n
2
. Dene
n
1
(.) = n(n
1
+.) and n
2
(.) = n(n
2
+.).
Now, we can view the problem as a portfolio choice problem with two investors each having
wealth n, rather than with one investor at two dierent levels of wealth. Given the proposition
above, these two approaches are equivalent. Then n
2
() is a concave transformation of n
1
(), by
the above proposition. We have shown before that the demand for the risky asset under n
1
() is
greater than that under n
2
(). This implies that the demand for the risky asset is greater at higher
levels of wealth.
Relative risk aversion
Absolute risk aversion is when outcomes are absolute gains or losses from current wealth. The
concept of relative risk aversion evaluates percentage gains or losses from current wealth.
Denition (relative risk aversion): Given n(), the coecient of relative risk aversion at r
is
r
1
(r, n) =
rn
tt
(r)
n
t
(r)
.
91
Note that r
1
(r, n) = rr

(r, n). Hence, an individual with a decreasing relative risk aversion


(drra) will also exhibit decreasing absolute risk aversion (dara). The converse is not necessarily
true. Thus, drra is a stronger concept than that of dara.
For example, n(r) =
_
r, exhibits constant relative risk aversion (crra) and dara. That is,
r

(r, n) =
1
2r
and r
1
(r, n) =
1
2
.
On the other hand, n(r) = c
oa
, exhibits increasing relative risk aversion (irra) and constant
absolute risk aversion (cara). That is,
r

(r, n) = a and r
1
(r, n) = ar.
Portfolio insurance problem (continued) Exercise 6.C.11. Show that if r
1
(n, n) is decreasing
in n, then the proportion of wealth invested in the risky asset (n) = c(n),n is decreasing in n.
Similarly, if r
1
(n, n) is decreasing in n, then (n) = c(n),n is increasing in n. Finally, when the
n exhibits crra, then the proportion is constant.
We will rst show that if the coecient of relative risk aversion is increasing in n, then
t
(n) < 0.
The proof when the coecient is decreasing or constant is similar.
The investor chooses to maximize,
max
01
_
n(n(. + (1 ))) d1(.).
The FOC for an interior solution is,
_
n
t
(n(. + (1 ))) (. 1) nd1(.) = 0. (1)
Totally dierentiate (1) with respect to and n. This yields,
__
n
tt
(n(. + (1 )))(. 1)
2
n
2
d1(.)
_
d
+
__
n
tt
(n(. + (1 ))) (. 1) (. + (1 )) nd1(.)
_
dn
==
d
dn
=
_
n
tt
(n(. + (1 ))) (. 1) (. + (1 )) nd1(.)
_
n
tt
(n(. + (1 )))(. 1)
2
n
2
d1(.)
. (2)
92
We need to show that (2) is negative. The denominator is negative since n is a concave function
and therefore all we need to show is that the numerator is positive.
By the denition of the coecient of the relative risk aversion
n
tt
(n(. + (1 ))) (. + (1 )) n
= r
1
(. + (1 )) n
t
(n(. + (1 ))) . (3)
(3) holds for every realization of ..
First suppose that . 1, then (. + (1 )) n n. Since the coecient of relative risk aversion
is increasing, this implies that,
r
1
((. + (1 )) n) r
1
(n).
Therefore,
n
tt
(n(. + (1 ))) (. + (1 )) n
r
1
(n) n
t
(n(. + (1 ))) . (4)
Since . 1 0,
n
tt
(n(. + (1 ))) (. + (1 )) n(. 1)
r
1
(n) n
t
(n(. + (1 ))) (. 1) . (5)
(5) will also hold if . < 1.
Therefore,

_
n
tt
(n(. + (1 ))) (. 1) (. + (1 )) n(. 1) d1(.)

_
r
1
(n) n
t
(n(. + (1 ))) (. 1) d1(.)
= r
1
(n)
_
n
t
(n(. + (1 ))) (. 1) d1(.) = 0,
by the FOC.
93
3.4 Comparison of payo distributions in terms of risk and return
Previous sections compared Bernoulli utility functions across individuals and dierent levels of
wealth. Here we compare payo distributions. In particular, we will be interested in the following
two comparisons,
Comparing the level of returns, i.e., whether a lottery 1() yields a higher return than a
lottery G().
Comparing the level of dispersion, i.e., whether a lottery 1() is riskier than a lottery G().
3.4.1 First order stochastic dominance (fosd)
Denition: The distribution 1() fosd G(), if for all non-decreasing functions n : 1 1 we
have,
_
n(r)d1(r) _
_
n(r)dG(r).
Any decision maker with an increasing Bernoulli utility function would choose 1() over G(),
regardless of his degree of risk aversion. This is because the decision makers expected utility under
1() is greater than that under G().
Proposition: The distribution of monetary payos 1() fosd the distribution G() if and only
if 1(r) _ G(r), for all r. That is,
1Oo1 ==1(r) _ G(r), for all r.
Proof: First we prove ==. Suppose 1() fosd G(). We want to show that 1(r) G(r) < 0,
for all r. Denote,
H(r) = 1(r) G(r).
By way of contradiction, suppose, H( r) 0, for some r. Dene a non-decreasing function as
follows,
n(r) =
_
_
_
1, r r
0, r _ r.
94
This function has the property that
_
n(r)d1(r)
_
n(r)dG(r) =
_
n(r)dH(r)
=
_
a
0
0dH(r) +
_
o
a
1dH(r) = H() H( r)
(since lim
ao
[1(r) G(r)] = 0), = 0 H( r) = H( r) < 0.
Contradiction to the assumption that 1() fosd G().
Now we will prove ==. Let n(r) be an increasing function. Integrating by parts
5
we obtain,
_
n(r)dH(r) = [n(r)H(r)]
o
0

_
n
t
(r)H(r)dr.
Since H(0) = 0 and lim
ao
H(r) = 0, the rst term is zero. It follows that,
_
n(r)dH(r) =
_
n(r)d1(r)
_
n(r)dG(r) _ 0,
if and only if,
_
n
t
(r)H(r)dr _ 0.
Thus, if H(r) _ 0 for all r and n() is increasing, then
_
n
t
(r)H(r)dr _ 0.

Example 1: The are four outcomes C = 1, 2, 3, 4 and two lotteries: G = 1,4, 1,4, 1,4, 1,4
and 1 = 0, 0, 1,2, 1,2.
G(r) 1(r)
Pr o/(r _ 1) = 1,4 0
Pr o/(r _ 2) = 1,2 0
Pr o/(r _ 3) = 3,4 1,2
Pr o/(r _ 4) = 1 1
5
Here is how we derive the formula: dierentiate uv and then integrate,
Z
d (uv) =
Z
udv +
Z
vdu
=)uv =
Z
udv +
Z
vdu
Z
udv = uv
Z
vdu.
In our case, u(x) = u and dH = dv:
95
1 fosd G. Also note that 1(r) _ G(r) for all r.
Example 2: Consider the following two normal densities (both with standard deviation o
1
=
o
2
= 1),
)(r) =
c
(aj
1
)
2
2o
2
1
_
2o
2
1
and q(r) =
c
(aj
2
)
2
2o
2
2
_
2o
2
2
.
The rst distribution has mean j
1
= 10 and the second one j
2
= 5. Moreover, the rst one
fosd the second. The graph below depicts the two distribution functions.
Remark: 1() fosd G() ==mean 1 mean G. To see this, choose n(r) = r. The reverse,
however, is not necessarily true. The entire distribution matters, not just the rst moment.
The next examples, demonstrate that
mean 1 mean G ;1() fosd G().
Example 3: The are four outcomes C = 1, 2, 3, 4 and two lotteries: G = 1,4, 1,4, 1,4, 1,4
and 1 = 1,3, 0, 0, 2,3 .
G(r) 1(r)
Pr o/(r _ 1) = 1,4 1,3
Pr o/(r _ 2) = 1,2 1,3
Pr o/(r _ 3) = 3,4 1,3
Pr o/(r _ 4) = 1 1
96
Note that the mean under 1 is 3 which is greater than the mean under G which is 2.5. Never-
theless, neither distribution dominates the other in the rst order sense.
Example 4: Consider the following two normal densities (both with standard deviation 1),
)(r) =
c
(aj
1
)
2
2o
2
1
_
2o
2
1
and q(r) =
c
(aj
2
)
2
2o
2
2
_
2o
2
2
.
The rst distribution has mean 10 and standard deviation 1 and the second one has mean 5
and standard deviation 5. Neither one fosd the other, although the rst one has a higher mean.
The graph below depicts the two distribution functions.
3.4.2 Second order stochastic dominance (sosd)
Denition: For any two distributions 1() and G(), with the same mean, 1() sosd (or is less
risky than) G() if for every non-decreasing concave function n : 1 1 we have,
_
n(r)d1(r) _
_
n(r)dG(r).
In other words, a risk averse decision maker would prefer 1 to G, since 1 has the same mean
as G and at the same it is less risky.
An alternative way to characterize sosd is in terms of mean preserving spreads (mps).
97
Mean preserving spreads
Consider the following compound lottery. In the rst stage, r is distributed according to 1(r).
In the second stage, we randomize each outcome further, by adding to r a random variable ., (r+.)
with . being distributed according to H
a
(.) and
_
.dH
a
(.) = 0. We basically add uncorrelated
noise to the random variable r.
The mean of r +. is r. Let the resulting lottery be denoted by G(). We say that G is a mps
of 1.
Example 5: Consider the lottery 1 = 1,2, 1,2 on C = 5, 10. Now a mps is constructed as
follows. Let 1,2, 1,2 be a lottery on a new set of outcomes 7 = 1, 1. Add this new lottery
to the existing one to obtain the lottery,
G = 1,4, 1,4, 1,4, 1,4
on
4, 6, 9, 11 .
1 sosd G. Also, G is a mps of 1.
Result: G() is a mps of 1() ==1() sosd G().
Proof : ==. Let n() be a concave Bernoulli utility function. We can conclude that,
_
n(r)dG(r) =
_ __
n(r +.)dH
a
(.)
_
d1(r)
_
_
n
__
(r +.) dH
a
(.)
_
d1(r)
=
_
n(r)d1(r).
The rst equality is by the denition of a mps, the second inequality follows by Jensens in-
equality and the third equality follows from the fact that the noise we add has mean zero.
==. The converse follows if we reverse the steps of the above case.
An elementary increase in risk
98
An elementary increase in risk (eir) is constructed as follows: take all the mass that 1() assigns
to the interval [r
t
, r
tt
] and transfer it to the endpoints r
t
and r
tt
in such a way that the mean does
not change, see gure below.
Below we oer yet another way to capture the sosd idea, based on the eir concept.
Consider two distribution functions 1() and G() with the same mean. Note that
1( r) = G( r) = 1,
for some suciently high r. Integrate by parts.
_
a
0
(1(r) G(r)) dr = r(1(r) G(r))[
a
0

_
a
0
rd (1(r) G(r))
= r 0 0 0 = 0.
This is because the distributions have the same mean. This implies that when the two distrib-
ution functions have the same mean, then the area under each distribution function must be equal
to each other. This, in turn implies that the area under A must equal the area under B, which
further implies,
_
a
0
G(t)dt _
_
a
0
1(t)dt, for all r. (*)
x
1
x x
B
A
F
G
0
99
That is, if 1 sosd G, then the area under G is greater than that under 1 for any r.
Proposition: Suppose 1() and G() have the same mean. Then, the following are equivalent,
i) 1() sosd G().
ii) G() is a mps of 1().
iii) (*) holds.
Example 6: Consider the following two normal densities,
)(r) =
c
(aj
1
)
2
2o
2
1
_
2o
2
1
and q(r) =
c
(aj
2
)
2
2o
2
2
_
2o
2
2
.
Both distributions have mean 10. The rst one has standard deviation 1 and the second 5. The
graph below depicts the two distribution functions. 1() sosd G().
100
101

Das könnte Ihnen auch gefallen