Econ 509, Introduction to Mathematical Economics I

Professor Ariell Reshef

University of Virginia
Lecture notes based on Chiang and Wainwright, Fundamental Methods of Mathematical Economics.
1 Mathematical economics
Why describe the world with mathematical models, rather than use verbal theory and logic? After
all, this was the state of economics until not too long ago (say, 1950s).
1. Math is a concise, parsimonious language, so we can describe a lot using fewer words.
2. Math contains many tools and theorems that help making general statements.
3. Math forces us to explicitly state all assumptions, and help preventing us from failing to
acknowledge implicit assumptions.
4. Multi dimensionality is easily described.
Math has become a common language for most economists. It facilitates communication between
economists. Warning: despite its usefulness, if math is the only language for economists, then
we are restricting not only communication among us, but more importantly we are restricting our
understanding of the world.
Mathematical models make strong assumptions and use theorems to deliver insightful conclu-
sions. But, remember the A-A C-C Theorem:
Let C be the set of conclusions that follow from the set of assumptions A. Let A be a small
perturbation of A. There exists such A that delivers a set of conclusions C that is disjoint
from C.
Thus, the insightfullness of C depends critically on the plausibility of A.
The plausibility of A depends on empirical validity, which needs to be established, usually
using econometrics. On the other hand, sometimes theory informs us on how to look at existing
data, how to collect new data, and which tools to use in its analysis. Thus, there is a constant
discourse between theory and empirics. Neither can be without the other (see the inductivism vs.
deductivism debate).
Theory is an abstraction of the world. You focus on the most important relationships that
you consider important a priori to understanding some phenomenon. This may yield an economic
model.
1
2 Economic models
Some useful notation: \ for all, exists, ! exists and is unique. If we cross any of these, or prex
by or , then it means "not": e.g., @, and all mean "does not exist".
2.1 Ingredients of mathematical models
1. Equations:
Denitions/Identities : = 1 C
: 1 = C +1 +G+A '
: 1
t+1
= (1 c) 1
t
+1
t
: ' = 11
Behavioral/Optimization :
o
= c ,j
: 'C = '1
: 'C = 1
Equilibrium :
o
=
c
2. Parameters: e.g. c, ,, c from above.
3. Variables: exogenous, endogenous.
Parameters and functions govern relationships between variables. Thus, any complete mathematical
model can be written as
1 (0, 1, A) = 0 ,
where 1 is a set of functions (e.g., demand, supply and market clearing conditions), 0 is a set
of parameters (e.g., elasticities), 1 are endogenous variables (e.g., price and quantity) and A are
exogenous, predetermined variables (e.g., income, weather). Some models will not have explicit
A variables. Moving from a "partial equilibrium" model closer to a "general equilibrium" model
involves treating more and more exogenous variables as endogenous.
Models typically have the following ingredients: a sense of time, model population (who makes
decisions), technology and preferences.
2.2 From chapter 3: equilibrium analysis
One general denition of a models equilibrium is "a constellation of selected, interrelated vari-
ables so adjusted to one another that no inherent tendency to change prevails in the model
2
which they constitute".
Selected: there may be other variables. This implies a choice of what is endogenous and
what is exogenous, but also the overall set of variables that are explicitly considered in the
model. Changing the set of variables that is discussed, and the partition to exogenous and
endogenous will likely change the equilibrium.
Interrelated: The value of each variable must be consistent with the value of all other
variables. Only the relationships within the model determine the equilibrium.
No inherent tendency to change: all variables must be simultaneously in a state of rest,
given the exogenous variables and parameters are all xed.
Since all variables are at rest, an equilibrium is often called a static. Comparing equilibria is called
therefore comparative statics (there is dierent terminology for dynamic models).
An equilibrium can be dened as 1
+
that solves
1 (0, 1, A) = 0 ,
for given 0 and A. This is one example for the usefulness of mathematics for economists: see how
much is described by so little notation.
We are interested in nding an equilibrium for 1 (0, 1, A) = 0. Sometimes, there will be no
solution. Sometimes it will be unique and sometimes there will be multiple equilibria. Each of
these situations is interesting in some context. In most cases, especially when policy is involved,
we want a model to have a unique equilibrium, because it implies a function from (0, A) to 1
(the implicit function theorem). But this does not necessarily mean that reality follows a unique
equilibrium; that is only a feature of a model. Warning: models with a unique equilibrium are
useful for many theoretical purposes, but it takes a leap of faith to go from model to realityas if
the unique equilibrium pertains to reality.
Students should familiarize themselves with the rest of chapter 3 on their own.
2.3 Numbers
Natural, N: 0, 1, 2... or sometimes 1, 2, 3, ...
Integers, Z: ... 2, 1, 0, 1, 2, ...
Rational, Q: :,d where both : and d are integers and d is not zero. : is the numerator and
d is the denominator.
Irrational numbers: cannot be written as rational numbers, e.g., , c,
_
2.
3
Real, R: rational and irrational. The real line: (, ). This is a special set, because it is
dense. There are just as many real numbers between 0 and 1 (or any other two real numbers)
as on the entire real line.
Complex: an extension of the real numbers, where there is an additional dimension in which
we add to the real numbers imaginary numbers: r +ij, where i =
_
1.
2.4 Sets
We already described some sets above (N, Q, R, Z). A set o contains elements c:
o = c
1
, c
2
, c
3
, c
4
,
where c
i
may be numbers or objects (say: car, bus, bike, etc.). We can think of sets in terms of
the number of elements that they contain:
Finite: o = c
1
, c
2
, c
3
, c
4
.
Countable: there is a mapping between the set and N. Trivially, a nite set is countable.
Innite and countable: Q. Despite containing innitely many elements, they are countable.
Uncountable: R, [0, 1].
Membership and relationships between sets:
c o means that the element c is a member of set o.
Subset: o
1
o
2
: \c o
1
, c o
2
. Sometimes denoted as o
1
_ o
2
. Sometimes a strict subset
is dened as \c o
1
, c o
2
and c o
2
, c , o
1
.
Equal: o
1
= o
2
: \c o
1
, c o
2
and \c o
2
, c o
1
.
The null set, ?, is a subset of any set, including itself, because it does not contain any element
that is not in any subset (it is empty).
Cardinality: there are 2
a
subsets of any set of magnitude : = [o[.
Disjoint sets: o
1
and o
2
are disjoint if they do not share common elements, i.e. if @c such
that c o
1
and c o
2
.
Operations on sets:
Union (or): ' 1 = c[c or c 1.
4
Intersection (and): 1 = c[c and c 1.
Complement: dene as the universe set. Then

or
c
= c[c and c , .
Minus: for 1 , 1 = c[c and c , 1. E.g.,

= .
Rules:
Commutative:
' 1 = 1 '
1 = 1
Association:
(' 1) ' C = ' (1 ' C)
( 1) C = (1 C)
Distributive:
' (1 C) = (' 1) (' C)
(1 ' C) = ( 1) ' ( C)
Do Venn diagrams.
2.5 Relations and functions
Ordered pairs: whereas r, j = j, r because they are sets, but not ordered, (r, j) ,= (j, r)
unless r = j (think of the two dimensional plane R
2
). Similarly, one can dene ordered triplets,
quadruples, etc.
Let A and 1 be two sets. The Cartesian product of A and 1 is a set o that is given by
o = A 1 = (r, j) [r A, j 1 .
For example, R
a
is a Cartesian product
R
a
= R R ... R = (r
1
, r
2
, ...r
a
) [r
i
R .
Cartesian products are relations between sets:
\r A, j 1 such that (r, j) A 1 ,
5
so that the set 1 is related to the set A. Any subset of a Cartesian product also has this trait.
Note that each r A may have more than one j 1 related to it (and vice versa). Thus the
relation may assign to any r A a set of values in 1 , o
a
1 . (Analysis of the shape of these sets
in the context of relations will be useful when discussion dynamic programming.)
If
\r A, !j 1 such that (r, j) o A 1 ,
then j is a function of r. We write this in shorthand notation as
j = ) (r)
or
) : A 1 .
The second term is also called mapping, or transformation. Note that although for j to be a
function of r we must have \r A, !j 1 , it is not necessarily true that \j 1, !r A. In
fact, there need not exist any such r at all. For example, j = a +r
2
, a 0.
In j = ) (r), j is the value or dependent variable; r is the argument or independent variable.
The set of all permissible values of r is called domain. For j = ) (r), j is the image of r. The
set of all possible images is called the range, which is a subset of 1 .
2.6 Functional forms
Students should familiarize themselves with polynomials, exponents, logarithms, "rectangular hy-
perbolic" functions (unit elasticity), etc. See Chapter 2.5 in CW.
2.7 Functions of more than one variable
. = ) (r, j) means that
\(r, j) domain A 1, !. 7 such that (r, j, .) o A 1 7 .
This is a function from a plane in R
2
to R or a subset of it. j = ) (r
1
, r
2
, ...r
a
) is a function from
the R
a
hyperplane or hypersurface to R or a subset of it.
3 Equilibrium analysis
Students cover independently. Conceptual points are reported above in Section 2.2.
6
4 Matrix algebra
4.1 Denitions
Matrix:

na
=
_

_
a
11
a
12
. . . a
1a
a
21
a
22
a
2a
.
.
.
.
.
.
a
n1
a
n2
. . . a
na
_

_
= [a
i)
] i = 1, 2, ...:, , = 1, 2, ...: .
Notation: usually matrices are denoted in upper case; : and : are called the dimensions.
Vector:
r
n1
=
_

_
r
1
r
2
.
.
.
r
n
_

_
.
Notation: usually lowercase. Sometimes called a column vector. A row vector is
r
t
=
_
r
1
r
2
r
n

.
4.2 Matrix operations
Equality: = 1 i a
i)
= /
i)
\i,. Clearly, the dimensions of and 1 must be equal.
Addition/subtraction: 1 = C i a
i)
/
i)
= c
i)
\i,.
Scalar multiplication: 1 = c i /
i)
= c a
i)
\i,.
Matrix multiplication: Let
na
and 1
I|
be matrices.
if : = / then the product
na
1
a|
exists and is equal to a matrix C
n|
of dimensions
:|.
if : = | then the product 1
In

na
exists and is equal to a matrix C
Ia
of dimensions
/ :.
7
If product exists, then

na
1
a|
=
_

a
11
a
12
. . . a
1a
a
21
a
22
a
2a
.
.
.
.
.
.
a
n1
a
n2
. . . a
na
_

_
_

_
|
/
11
/
12
. . . /
1|
/
21
/
22
/
2|
.
.
.
.
.
.
/
a1
/
a2
. . . /
a|
_

_
=
_
c
i)
=
a

I=1
a
iI
/
I)
_
i = 1, 2, ...:, , = 1, 2, ...| .
Transpose: Let
na
= [a
i)
]. Then
t
an
= [a
)i
]. Also denoted
T
. Properties:
(
t
)
t
=
(+1)
t
=
t
+1
t
(1)
t
= 1
t

t
Operation rules
Commutative addition: +1 = 1 +.
Distributive addition: (+1) +C = + (1 +C).
NON commutative multiplication: 1 ,= 1, even if both exist.
Distributive multiplication: (1) C = (1C).
Association: premultiplying (1 +C) = 1+C and postmultiplying (+1) C =
C +1C.
4.3 Special matrices
Identity matrix:
1 =
_

_
1 0 . . . 0
0 1 0
.
.
.
.
.
.
.
.
.
0 0 . . . 1
_

_
.
1 = 1 = (of course, dimensions must conform).
Zero matrix: all elements are zero. 0 + = , 0 = 0 = 0 (of course, dimensions must
conform).
Idempotent matrix: = .
I
= , / = 1, 2, ...
Example: the linear regression model is j
a1
= A
aI
,
I1
+ -
a1
. The estimated model
8
by OLS is j = A/ + c, where / = (A
t
A)
1
A
t
j. Therefore we have predicted values
j = A/ = A (A
t
A)
1
A
t
j and residuals c = j j = j A/ = j A (A
t
A)
1
A
t
j =
_
1 A (A
t
A)
1
A
t
_
j. We can dene the projection matrix as 1 = A (A
t
A)
1
A
t
and the
residual generating matrix as 1 = [1 1]. Both 1 and 1 are idempotent. What does it
mean that 1 is idempotent? And that 1 is idempotent? What is the product 11, and what
does that imply?
Singular matrices: even if 1 = 0, this does NOT imply that = 0 or 1 = 0. E.g.,
=
_
2 4
1 2
_
, 1 =
_
2 4
1 2
_
.
Likewise, C1 = C1 does NOT imply 1 = 1. E.g.,
C =
_
2 3
6 9
_
, 1 =
_
1 1
1 2
_
, 1 =
_
2 1
3 2
_
.
This is because , 1 and C are singular: there is one (or more) row or column that is a linear
combination of the other rows or columns, respectively.
Nonsingular matrix: a square matrix that has an inverse.
Diagonal matrix
1 =
_

_
d
11
0 . . . 0
0 d
22
0
.
.
.
.
.
.
.
.
.
0 0 . . . d
aa
_

_
.
Upper triangular matrix. Matrix l is upper triangular if n
i)
= 0 for all i ,, i.e. all
elements below the diagonal are zero. E.g.,
_
_
a / c
0 c )
0 0 i
_
_
.
Lower triangular matrix. Matrix \ is lower triangular if n
i)
= 0 for all i < ,, i.e. all
elements above the diagonal are zero. E.g.,
_
_
a 0 0
d c 0
q / i
_
_
.
Symmetric matrix: =
t
.
Permutation matrix: a matrix of 0s and 1s in which each row and each column contains
9
exactly one 1. E.g.,
_
_
0 0 1
1 0 0
0 1 0
_
_
.
Multiplying a conformable matrix by a permutation matrix changes the order of the rows or
the columns (unless it is the identity matrix). For example,
_
0 1
1 0
_ _
1 2
3 4
_
=
_
3 4
1 2
_
and
_
1 2
3 4
_ _
0 1
1 0
_
=
_
2 1
4 3
_
.
Partitioned matrix: a matrix with elements that are matrices themselves, e.g.,
_

jI
1
ji
C
j)
1
II
1
II
1
II
_
(j+I)(I+i+))
.
Note that the dimensions of the sub matrices must conform.
4.4 Vector products
Scalar multiplication: Let r
n1
be a vector. Then the scalar product cr is
cr
n1
=
_

_
cr
1
cr
2
.
.
.
cr
n
_

_
.
Inner product: Let r
n1
and j
n1
be vectors. Then the inner product is a scalar
r
t
j =
n

i=1
r
i
j
i
.
This is useful for computing correlations.
Outer product: Let r
n1
and j
a1
be vectors. Then the outer product is a matrix
rj
t
=
_

_
r
1
j
1
r
1
j
2
. . . r
1
j
a
r
2
j
1
r
2
j
2
r
2
j
a
.
.
.
.
.
.
r
n
j
1
r
n
j
2
. . . r
n
j
a
_

_
na
.
This is useful for computing the variance/covariance matrix.
10
Geometric interpretations: do in 2 dimensions. All extends to : dimensions.
Scalar multiplication.
Vector addition.
Vector subtraction.
Inner product and orthogonality (rj = 0 means rlj).
4.5 Linear independence
Denition 1: a set of / vectors r
1
, r
2
, ...r
I
are linearly independent i neither one can be expressed
as a linear combination of all or some of the others. Otherwise, they are linearly dependent.
Denition 2: a set of / vectors r
1
, r
2
, ...r
I
are linearly independent i a set of scalars c
1
, c
2
, ...c
I
such that c
i
,= 0 for some or all i and

I
i=1
c
i
r
i
= 0. Otherwise, they are linearly dependent. I.e.,
if such set of scalars exists, then the vectors are linearly dependent.
Consider R
2
:
All vectors that are multiples are linearly dependent. If two vectors cannot be expressed as
multiples then they are linearly independent.
If two vectors are linearly independent, then any third vector can be expressed as a linear
combination of the two.
It follows that any set of / 2 vectors in R
2
must be linearly dependent.
4.6 Vector spaces and metric spaces
The complete set of vectors of : dimensions is a space, a vector space. If all elements of these
vectors are real numbers ( R), then this space is R
a
.
Any set of : linearly independent vectors is a base for R
a
.
A base spans the space to which it pertains. This means that any vector in R
a
can be
expressed as a linear combination of the base (it is spanned by the base).
Bases are not unique.
Bases are minimal: they contain the smallest number of vectors that span the space.
Example: unit vectors. Consider the vector space R
3
. Then
c
1
=
_
_
1
0
0
_
_
, c
2
=
_
_
0
1
0
_
_
, c
3
=
_
_
0
0
1
_
_
11
is a base. Indeed, c
1
, c
3
, c
3
are linearly independent.
Distance metric: Let r, j o, some set. Dene the distance between r and j by a function
d: d = d (r, j), which has the following properties:
d (r, j) _ 0.
d (r, j) = d (j, r).
d (r, j) = 0 = r = j.
d (r, j) 0 = r ,= j.
d (r, j) _ d (r, .) +d (., j) \r, j, . (triangle inequality).
A metric space is given by a vector space + distance metric. The Euclidean space is given
by R
a
+ the following distance function
d (r, j) =

_
a

i=1
(r
i
j
i
)
2
=
_
(r j)
t
(r j) .
But you can imagine other metrics that give rise to other dierent metric spaces.
4.7 Inverse matrix
Denition: if for some square (: :) matrix there exists a matrix 1 such that 1 = 1, then 1
is the inverse of , and is denoted
1
, i.e.
1
= 1.
Properties:
Not all square matrices have an inverse. If
1
does not exist, then is singular. Otherwise,
is nonsingular.
is the inverse of
1
and vice versa.
The inverse is square.
The inverse, if it exists, is unique. Proof: suppose not, i.e. 1 = 1 and 1 ,=
1
. Then

1
1 =
1
1, 11 = 1 =
1
, a contradiction
Operation rules:

1
_
1
= . Proof: suppose not, i.e.
_

1
_
1
= 1 and 1 ,= . Then
1
=
1 =
_

1
_
1
= 1
1
=
_

1
_
1

1
= 1 = 1
1
= 1 = 11 = 1 = , a
contradiction
12
(1)
1
= 1
1

1
. Proof: Let (1)
1
= C. Then (1)
1
(1) = 1 = C (1) =
C1 = C11
1
= C = 11
1
= 1
1
= C
1
= C = 1
1

(
t
)
1
=
_

1
_
t
. Proof: Let (
t
)
1
= 1. Then (
t
)
1

t
= 1 = 1
t
= (1
t
)
t
=
1
t
= 1
t
= 1 =
1
1
t
=
1
1 = 1
t
=
1
= 1 =
_

1
_
t

Conditions for nonsingularity:
Necessary condition: matrix is square.
Given square matrix, a sucient condition is that the rows or columns are linearly inde-
pendent. It does not matter whether we use the row or column criterion because matrix is
square.
is square + linear independence
. .
necessary and sucient conditions
= is nonsingular =
1
How do we nd the inverse matrix? Soon... Why do we care? See next section.
4.8 Solving systems of linear equations
We seek a solution r to the system r = c

aa
r
a1
= c
a1
= r = c
1
,
where is a nonsingular matrix and c is a vector. Each row of gives coecients to the elements
of r:
row 1 :
a

i=1
a
1i
r
i
= c
1
row 2 :
a

i=1
a
2i
r
i
= c
2
Many linear (or linearized) models can be solved this way. We will learn clever ways to compute
the solution to this system. We care about singularity of because (given c) it tells us something
about the solution r.
4.9 Markov chains
We introduce this through an example. Let r denote a vector of employment and unemployment
rates: r
t
=
_
c n

, where c +n = 1 and c, n _ 0. Dene the matrix 1 as a transition matrix that
gives the conditional probabilities for transition from the state today to a state next period,
1 =
_
j
cc
j
c&
j
&c
j
&&
_
,
13
where j
i)
= Pr (state , tomorrow[state i today). Each row of 1 sums to unity: j
cc
+ j
c&
= 1 and
j
&c
+ j
&&
= 1; and since these are probabilities, j
i)
_ 0 \i,. Now add a time dimension to r:
r
t
t
=
_
c
t
n
t

.
We ask: What is the employment and unemployment rates going to be in t+1 given r
t
? Answer:
r
t
t+1
= r
t
t
1 =
_
c
t
n
t

_
j
cc
j
c&
j
&c
j
&&
_
=
_
c
t
j
cc
+n
t
j
&c
c
t
j
c&
+n
t
j
&&

.
What will they be in t + 2? Answer: r
t
t+2
= r
t
t+1
1 = r
t
t
1
2
. More generally, r
t
t
0
+I
= r
t
t
0
1
I
.
A transition matrix, sometimes called stochastic matrix, is dened as a square matrix
whose elements are non negative and all rows sum to 1. This gives you conditional transition
probabilities starting from each state, where each row is a starting state and each column is the
state in the next period.
Steady state: a situation in which the distribution over the states is not changing over time.
How do we nd such a state, if it exists?
Method 1: Start with some initial condition r
0
and iterate forward r
t
I
= r
t
0
1
I
, taking / .
Method 2: dene r as the steady state value. Solve r
t
= r
t
1. Or 1
t
r = r.
14
5 Matrix algebra continued and linear models
5.1 Rank
Denition: The number of linearly independent rows (or, equivalently, columns) of a matrix is
the rank of : r = ra:/ ().
If
na
then ra:/ () _ min:, :.
If a square matrix
aa
has rank :, then we say that is full rank.
Multiplying a matrix by a another matrix 1 that is full rank does not reduce the rank of
the product relative to the rank of .
If ra:/ () = r

and ra:/ (1) = r

1
, then ra:/ (1) = minr

, r
1
.
Finding the rank: the echelon matrix method. First dene elementary operations:
1. Multiply a row by a non zero scalar: c 1
i
, c ,= 0.
2. Adding c times of one row to another: 1
i
+c1
)
.
3. Interchanging rows: 1
i
1
)
.
All these operations alter the matrix, but do not change its rank (in fact, they can all be
expressed by multiplying matrices, which are all full rank).
Dene: echelon matrix.
1. Zero rows appear at the bottom.
2. For non zero rows, the rst element on the left is 1.
3. The rst element of each row on the left (which is 1) appears to the left of the row directly
below it.
The number of non zero rows in the echelon matrix is the rank.
We use the elementary operations in order to change the subject matrix into an echelon matrix,
which has as many zeros as possible. A good way to start the process is to concentrate zeros at the
bottom. Example:
=
_
_
0 11 4
2 6 2
4 1 0
_
_
1
1
1
3
:
_
_
4 1 0
2 6 2
0 11 4
_
_
1
4
1
1
:
_
_
1
1
4
0
2 6 2
0 11 4
_
_
15
1
2
21
1
:
_
_
1
1
4
0
0 5
1
2
2
0 11 4
_
_
1
3
+ 21
2
:
_
_
1
1
4
0
0 5
1
2
2
0 0 0
_
_
2
11
1
2
:
_
_
1
1
4
0
0 1 4,11
0 0 0
_
_
There is a row of zeros: ra:/ () = 2. So is singular.
5.2 Determinants and nonsingularity
Denote the determinant of a square matrix as [
aa
[. This is not absolute value. If the determinant
is zero then the matrix is singular.
1. [
11
[ = a
11
.
2. [
22
[ = a
11
a
22
a
12
a
21
.
3. Determinants for higher order matrices. Let
II
be a square matrix. The i-, minor ['
i)
[
is the determinant of the matrix given by erasing row i and column , from . Example:
=
_
_
a / c
d c )
q / i
_
_
, ['
11
[ =

c )
/ i

.
The Laplace Expansion of row i gives the determinant of :
[
II
[ =
I

)=1
(1)
i+)
a
i)
['
i)
[ =
I

)=1
a
i)
C
i)
,
where C
i)
= (1)
i+)
['
i)
[ is called the cofactor of a
i)
(or the i-,
tI
cofactor). Example:
expansion by row 1

a / c
d c )
q / i

= aC
11
+/C
12
+cC
13
= a ['
11
[ / ['
12
[ +c ['
13
[
= a

c )
/ i

d )
q i

+c

d c
q /

= a (ci )/) / (di )q) +c (d/ cq) .

In doing this, it is useful to choose the expansion with the row that has the most zeros.
Properties of determinants
1. [
t
[ = [[
2. Interchanging rows or columns ips the sign of the determinant.
16
3. Multiplying a row or column by a scalar c multiplies the determinant by c.
4. 1
i
+c1
)
does not change the determinant.
5. If a row or a column are multiples of another row or column, respectively, then the determinant
is zero: linear dependence.
6. Changing the minors in the Laplace expansion by alien minors, i.e. using ['
a)
[ instead of
['
i)
[ for row i ,= :, will give zero:
I

)=1
a
i)
(1)
i+)
['
a)
[ = 0 , i ,= : .
This is like forcing linear dependency by repeating elements.

I
)=1
a
i)
(1)
i+)
['
a)
[ is the
determinant of some matrix. That matrix can be reverse engineered from the last expression.
If you do this, you will nd that that reverse-engineered matrix has linear dependent columns
(try a 3 3 example).
Determinants and singularity: [[ , = 0
= is nonsingular
= columns and rows are linearly independent
=
1
= for r = c , !r =
1
c
= the column (or row) vectors of span the vector space.
5.3 Finding the inverse matrix
Let be a nonsingular matrix,

aa
=
_

_
a
11
a
12
. . . a
1a
a
21
a
22
a
2a
.
.
.
.
.
.
a
a1
a
a2
. . . a
aa
_

_
.
The cofactor matrix of is C

:
C

=
_

_
C
11
C
12
. . . C
1a
C
21
C
22
C
2a
.
.
.
.
.
.
C
a1
C
a2
. . . C
aa
_

_
,
17
where C
i)
= (1)
i+)
['
i)
[. The adjoint matrix of is adj = C
t

:
adj = C
t

=
_

_
C
11
C
21
. . . C
a1
C
12
C
22
C
a2
.
.
.
.
.
.
C
1a
C
2a
. . . C
aa
_

_
.
Consider C
t

:
C
t

=
_

a
)=1
a
1)
C
1)

a
)=1
a
1)
C
2)
. . .

a
)=1
a
1)
C
a)

a
)=1
a
2)
C
1)

a
)=1
a
2)
C
2)

a
)=1
a
2)
C
a)
.
.
.
.
.
.

a
)=1
a
a)
C
1)

a
)=1
a
a)
C
2)
. . .

a
)=1
a
a)
C
a)
_

_
=
_

a
)=1
a
1)
C
1)
0 . . . 0
0

a
)=1
a
2)
C
2)
0
.
.
.
.
.
.
0 0 . . .

a
)=1
a
a)
C
a)
_

_
=
_

_
[[ 0 . . . 0
0 [[ 0
.
.
.
.
.
.
0 0 . . . [[
_

_
= [[ 1 ,
where the o diagonal elements are zero due to alien cofactors. It follows that
C
t

= [[ 1
C
t

1
[[
= 1

1
= C
t

1
[[
=
adj
[[
.
Example:
=
_
1 2
3 4
_
, C

=
_
4 3
2 1
_
, C
t

=
_
4 2
3 1
_
, [[ = 2 ,
1
=
_
2 1
3
2

1
2
_
.
And you can verify this.
5.4 Cramers rule
For the system r = c and nonsingular , we have
r =
1
c =
adj
[[
c .
18
Denote by
)
the matrix with column , replaced by c. Then it turns out that
r
)
=
[
)
[
[[
.
To see why, note that each row of C
t

c is c times row of C
t

, i.e. each row r is

)
C
)v
c
)
, which is
a Laplace Expansion by row r of some matrix. That matrix is
)
and the Laplace expansion gives
the determinant of
)
.
5.5 Homogenous equations: Ax = 0
Let the system of equations be homogenous: r = 0.
If is nonsingular, then only r = 0 is a solution. Recall: if is nonsingular, then its columns
are linearly independent. Denote the columns of by
i
. Then r =

a
i=1
r
i

i
= 0 implies
r
i
= 0 \i by linear independence of the columns.
If is singular, then there are innite solutions, including r = 0.
5.6 Summary of linear equations: Ax = c
For nonsingular A:
1. c ,= 0 = !r ,= 0
2. c = 0 = !r = 0
For singular A:
1. c ,= 0 = r, innite solutions ,= 0.
If there is inconsistency linear dependency in , the elements of c do not follow the
same linear combination there is no solution.
2. c = 0 = r, innite solutions, including 0.
One can think of the system r = c as dening a relation between c and r. If is nonsingular,
then there is a function (mapping/transformation) between c and r. In fact, when is nonsingular,
this transformation is invertible.
19
5.7 Inverse of partitioned matrix (not covered in CW)
Let be a partitioned matrix such that
=
_

11

12

21

22
_
,
Sucient conditions for nonsingularity of are that
11
and
22
are square, nonsingular ma-
trices. In that case

1
=
_
1
11
1
11

12

1
22

1
22

21
1
11

1
22
+
1
22

21
1
11

12

1
22
_
, (1)
where 1
11
=
_

11

12

1
22

21
_
1
, or alternatively

1
=
_

1
11
+
1
11

12
1
22

21

1
11

1
11

12
1
22
1
22

21

1
11
1
22
_
, (2)
where 1
22
=
_

22

21

1
11

12
_
1
. (This is useful for econometrics.)
To prove the above start with 1 = 1 and gure out what the partitions of 1 need to be. To
get (1) you must assume (and use)
22
nonsingular; and to get (2) you must assume (and use)
11
nonsingular.
Note that
11
and
22
being nonsingular are not necessary conditions in general. For example,
=
_
0 1
1 0
_
is nonsingular but does not meet the sucient conditions. However if is positive denite (we will
dene this below; a bordered Hessian is not positive denite), then
11
and
22
being nonsingular
is also a necessary condition.
5.8 Leontief input/output model
We are interested in computing the level of output that is required from each industry in an
economy that is required to satisfy nal demand. This is not a trivial question, because output of all
industries (depending on how narrowly you dene an industry) are inputs for other industries, while
also being consumed in nal demand. These inter-industry relationships constitute input/output
linkages.
Assume
1. Each industry produces one homogenous good.
2. Inputs are used in xed proportions.
20
3. Constant returns to scale.
This gives rise to the Leontief (xed proportions) production function. The second assumption
can be relaxed, depending on the interpretation of the model. If you only want to use the framework
for accounting purposes, then this is not critical.
Dene a
ic
as the unit requirement of inputs from industry i used in the production of
output o. I.e., in order to produce one unit of output o you need a
ic
units of i. If some
industry o does not require its own output for production, then a
cc
= 0.
For : industries
aa
= [a
ci
] is a technology matrix. Each column tells you how much
of each input is required to produce one unit of output of that column. Alternatively, each
row tells you the requirements from the industry of that row if all other industries produced
exactly one unit.
If all industries were used as inputs as well as output, then there would be no primary inputs
(i.e. time, labor, entrepreneurial talent, natural resources, land). To accommodate primary
inputs, we add an open sector. If the a
ic
are denominated in monetary valuesi.e., in
order to produce $1 in industry o you need $a
ic
of input ithen we must have

a
i=1
a
ic
_ 1,
because the revenue from producing output o is $1. And if there is an open sector, then we
must have

a
i=1
a
ic
< 1. This means that the cost of intermediate inputs required to produce
$1 of revenue is less than $1. By CRS and competitive economy, we have the zero prot
condition, which means that all revenue is paid out to inputs. So primary inputs receive
(1

a
i=1
a
ic
) dollars from each industry o.
Equilibrium implies
supply = demand
= demand for intermediate inputs + nal demand .
In matrix notation
r = r +d .
And so
r r = (1 ) r = d .
21
Let
t
c
be the o
tI
row vector of . Then for some output o (row) we have
r
c
=
t
c
r +d
c
=
a

i=1
a
ci
r
i
+d
c
= a
c1
r
1
+a
c2
r
2
+... +a
ca
r
a
. .
intermediate inputs
+ d
c
..
nal
.
For example, a
c2
r
2
is the amount of output o that is required by industry 2, because you need a
c2
units of o to produce each unit of industry 2 and r
2
units of industry 2 are produced. This implies
a
c1
r
1
a
c2
r
2
+... (1 a
cc
) r
c
a
c,c+1
r
c+1
... a
ca
r
a
= d
c
.
In matrix notation
_

_
(1 a
11
) a
12
a
13
a
1a
a
21
(1 a
22
) a
23
a
2a
a
31
a
32
(1 a
33
) a
3a
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
a
a1
a
a2
a
a3
(1 a
aa
)
_

_
_

_
r
1
r
2
r
3
.
.
.
r
a
_

_
=
_

_
d
1
d
2
d
3
.
.
.
d
a
_

_
.
Or
(1 ) r = d .
(1 ) is the Leontief matrix. This implies that you need to produce more than just nal
demand because some r are used as intermediate inputs (loosely speaking, "1 < 1").
r = (1 )
1
d .
You need nonsingular (1 ). But even then the solution to r might not be positive. While in
reality this must be trivially satised in the data, we need to nd theoretical restrictions on the
technology matrix to satisfy a non-negative solution for r.
5.8.1 Existence of non negative solution
Consider
=
_
_
a / c
d c )
q / i
_
_
.
Dene
Principal minor: the determinant of the matrix that arises from deleting the i-th row and
22
i-th column. E.g.
['
11
[ =

c )
/ i

, ['
22
[ =

a c
q i

, ['
33
[ =

a /
d c

.
/-th order principal minor: is a principal minor of dimensions / /. If the dimensions
of the original matrix are : :, then a /-th order principal minor is obtained after deleting
the same : / rows and columns. E.g., the 1-st order principal minors of are
[a[ , [c[ , [i[ .
The 2-nd order principal minors are ['
11
[, ['
22
[ and ['
33
[ given above.
Leading principal minors: these are the 1
ct
, 2
ao
, 3
vo
(etc.) order principal minors, where
we keep the upper most left corner of the original matrix in each one. E.g.
['
1
[ = [a[ , ['
2
[ =

a /
d c

, ['
3
[ =

a / c
d c )
q / i

.
Simon-Hawkins Condition (Theorem): consider the system of equations 1r = d. If
(1) all o-diagonal elements of 1
aa
are non positive, i.e. /
i)
_ 0, \i ,= ,;
(2) all elements of d
a1
are non negative, i.e. d
i
_ 0, \i;
Then r _ 0 such that 1r = d i
(3) all leading principal minors are strictly positive, i.e. ['
i
[ 0, \i.
In our case, 1 = 1 , the Leontief matrix. Conditions (1) and (2) are satised. To illustrate
the economic meaning of SHC, use a 2 2 example:
1 =
_
1 a
11
a
12
a
21
1 a
22
_
.
Condition (3) requires ['
1
[ = [1 a
11
[ = 1 a
11
0, i.e. a
11
< 1. This means that less than the
total output of r
1
is used to produce r
1
, i.e. viability. Next, condition (3) also requires
['
2
[ = [1 [
= (1 a
11
) (1 a
22
) a
12
a
21
= 1 a
11
a
22
+a
11
a
22
a
12
a
21
0
Rearranging terms we have
(1 a
11
) a
22
. .
0
+a
11
+a
12
a
21
< 1
23
and therefore
a
11
..
direct use
+ a
12
a
21
. .
indirect use
< 1
This means that the total amount of r
1
demanded (for production of r
1
and for production of r
2
)
is less than the amount produced (=1), i.e. the resource constraint is kept.
5.8.2 Closed model version
The closed model version treats the primary sector as any industry. Suppose that there is only
one primary input: labor. The interpretation is that each good is consumed in xed proportions
(Leontief preferences). In the case when a
i)
represents value, then the interpretation is that expen-
diture on each good is in xed proportions (these preferences can be represented by a Cobb-Douglas
utility function).
In this model nal demand, as dened above, must equal zero. Since income accrues to primary
inputs (think of labor) and this income is captured in r, then it follows that the d vector must be
equal to zero. Since nal demand equals income, then if nal demand was positive, then we would
have to have an open sector to pay for that demand (from its income). I.e. we have a homogenous
system:
(1 ) r = 0
_
_
(1 a
00
) a
01
a
02
a
10
(1 a
11
) a
12
a
20
a
21
(1 a
22
)
_
_
_
_
r
0
r
1
r
2
_
_
=
_
_
0
0
0
_
_
,
where 0 denotes the primary sector (there could be more than one).
Each column o in the technology matrix must sum to 1, i.e. a
0c
+ a
2c
+ ... + a
ac
= 1, \o,
because all of the revenue is exhausted in payments for inputs (plus consumption). But then each
column in 1 can be expressed as minus the sum of all other columns. It follows that 1
is singular, and therefore r is not unique! It follows that you can scale up or down the economy
with no eect. In fact, this is a general property of CRS economies with no outside sector or
endowment. One way to pin down the economy is to set some r
i
to some level as an endowment
and, accordingly, to set r
ii
= 0 (you dont need land to produce land).
6 Derivatives and limits
Teaching assistant covers. See Chapter 6 in CW.
24
7 Dierentiation and use in comparative statics
7.1 Dierentiation rules
1. If j = ) (r) = c, a constant, then
oj
oa
= 0
2.
o
oa
ar
a
= a:r
a1
3.
o
oa
lnr =
1
a
4.
o
oa
[) (r) q (r)] = )
t
(r) q
t
(r)
5.
o
oa
[) (r) q (r)] = )
t
(r) q (r) +) (r) q
t
(r) = [) (r) q (r)]
)
0
(a)
)(a)
+ [) (r) q (r)]
j
0
(a)
j(a)
6.
o
oa
_
)(a)
j(a)
_
=
)
0
(a)j(a))(a)j
0
(a)
[j(a)]
2
=
)(a)
j(a)
)
0
(a)
)(a)

)(a)
j(a)
j
0
(a)
j(a)
7.
o
oa
) [q (r)] =
o)
oj
oj
oa
(the chain rule)
8. Inverse functions. Let j = ) (r) be strictly monotone (there are no "ats"). Then an inverse
function, r = )
1
(j), exists and
dr
dj
=
d)
1
(j)
dj
=
1
dj,dr
=
1
d) (r) ,dr
,
where r and j map one into the other, i.e. j = ) (r) and r = )
1
(j).
Strictly monotone means that r
1
r
2
=) (r
1
) ) (r
2
) (strictly increasing) or ) (r
1
) <
) (r
2
) (strictly decreasing). It implies that there is an inverse function r = )
1
(j)
because \j Range !r domain (recall: \r domain !j Range denes ) (r)).
7.2 Partial derivatives
Let j = ) (r
1
, r
2
, ...r
a
). Dene the partial derivative of ) with respect to r
i
:
0j
0r
i
= lim
a
i
0
) (r
i
+ r
i
, r
i
) ) (r
i
, r
i
)
r
i
.
Operationally, you derive 0j,0r
i
just as you would derive dj,dr
i
, while treating all other r
i
as
constants.
Example. Consider the following production function
j = . [c/
,
+ (1 c) |
,
]
1,
, , _ 1 .
Dene the elasticity of substitution as the percent change in relative factor intensity (/,|) in re-
sponse to a 1 percent change in the relative factor returns (r,n). What is the elasticity of sub-
stitution? If factors are paid their marginal product (which is a partial derivative in this case),
25
then
j
I
=
1
,
. []
1
'
1
,c/
,1
= r
j
|
=
1
,
. []
1
'
1
,(1 c) |
,1
= n .
Thus
r
n
=
c
1 c
_
/
|
_
,1
and then
/
|
=
_
c
1 c
_ 1
1'
_
r
n
_

1
1'
.
The elasticity of substitution is o =
1
1,
and it is constant. This production function exhibits
constant elasticity of substitution, denoted a CES production function. A 1 percent increase in
r,n decreases /,| by o percent.
7.3 Gradients
j = ) (r
1
, r
2
, ...r
a
)
The gradient is dened as
\) = ()
1
, )
2
, ...)
a
) ,
where
)
i
=
0)
0r
i
.
We can use this in rst order approximations:
)[
a
0
- \) (r
0
) r
) (r) ) (r
0
) - ()
1
, )
2
, ...)
a
)[
a
0
_
_
_
_

_
r
1
.
.
.
r
a
_

_
_

_
r
0
1
.
.
.
r
0
a
_

_
_
_
_ .
Application to open input/output model:
(1 ) r = d
r = (1 )
1
d = \ d
_

_
r
1
.
.
.
r
a
_

_ =
_

11

1a
.
.
.
.
.
.
.
.
.

a1

aa
_

_
_

_
d
1
.
.
.
d
a
_

_ .
26
Think of r as a function of d:
\r
i
=
_

i1

i2

ia
_

i)
=
0r
i
0d
)
.
And more generally,
r = \r d = \ d .
7.4 Jacobian and functional dependence
Let there be two functions
j
1
= ) (r
1
, r
2
)
j
2
= q (r
1
, r
2
) .
The Jacobian determinant is
[J[ =

0j
0r
t

0
_
j
1
j
2
_
0 (r
1
, r
2
)

0j
1
0a
1
0j
1
0a
2
0j
2
0a
1
0j
2
0a
2

.
Theorem (functional dependence): [J[ = 0 \r i the functions are dependent.
Example: j
1
= r
1
r
2
and j
2
= lnr
1
+ lnr
2
.
[J[ =

r
2
r
1
1
a
1
1
a
2

= 0 .
Example: j
1
= r
1
+ 2r
2
2
and j
2
= ln
_
r
1
+ 2r
2
2
_
.
[J[ =

1 4r
2
1
a
1
+2a
2
2
4a
2
a
1
+2a
2
2

= 0 .
Another example: r = \ d,
_

_
r
1
.
.
.
r
a
_

_ =
_

11

13
.
.
.
.
.
.
.
.
.

a1

aa
_

_
_

_
d
1
.
.
.
d
a
_

_ =
_

1i
d
i
.
.
.

ai
d
i
_

_ .
So [J[ = [\ [. It follows that linear dependence is equivalent to functional dependence for a system
of linear equations. If [\ [ = 0 then there are solutions for r and the relationship between d and
r cannot be inverted.
27
8 Total dierential, total derivative and the implicit function the-
orem
8.1 Total derivative
Often we are interested in the total rate of change in some variable in response to a change in some
other variable or some parameter. If there are indirect eects, as well as direct ones, you want to
take this into account. Sometimes the indirect eects are due to general equilibrium constraints
and can be very important.
Example: consider the utility function n(r, j) and the budget constraint j
a
r +j
j
j = 1. Then
the total eect of a small change in r on utility is
dn
dr
=
0n
0r
+
0n
0j

dj
dr
.
More generally: 1 (r
1
, ...r
a
)
d1
dr
i
=
a

)=1
01
0r
)

dr
)
dr
i
,
where we know that dr
i
,dr
i
= 1.
Example: . = ) (r, j, n, ), where r = r(n, ) and j = j (n, ) and = (n).
d.
dn
=
0)
0r
_
0r
0n
+
0r
0
d
dn
_
+
0)
0j
_
0j
0n
+
0j
0
d
dn
_
+
0)
0n
+
0)
0
d
dn
.
If we want to impose that is not directly aected by n, then all terms that involve d,dn are zero:
d.
dn
=
0)
0r
dr
dn
+
0)
0j
dj
dn
+
0)
0n
.
Alternatively, we can impose that is constant; in this case the derivative is denoted as
o:
o&

and
the result is the same as above.
8.2 Total dierential
Now we are interested in the change (not rate of...) in some variable or function if all its arguments
change a bit, i.e. they are all perturbed. For example, if the saving function for the economy is
o = o (j, r), then
do =
0o
0j
dj +
0o
0r
dr .
More generally, j = 1 (r
1
, ...r
a
)
dj =
a

)=1
01
0r
)
dr
)
.
28
One can view the total dierential as a linearization of the function around a specic point, because
01,0r
)
must be evaluated at some point.
The same rules that apply to derivatives apply to dierentials; just simply add dr after each
partial derivative:
1. dc = 0 for constant c.
2. d (cn
a
) = c:n
a1
dn =
0(c&
n
)
0&
dn.
3. d (n ) = dn d =
0(&)
0&
dn +
0(&)
0
d.
d (n n) = dn d dn =
0(&&)
0&
dn +
0(&&)
0
d +
0(&&)
0&
dn.
4. d (n) = dn +nd =
0(&)
0&
dn +
0(&)
0
d = (n)
o&
&
+ (n)
o

.
d (nn) = ndn +nnd +ndn =
0(&&)
0&
dn +
0(&&)
0
d +
0(&&)
0&
dn.
5. d (n,) =
o&&o

2
=
0(&)
0&
dn +
0(&)
0
d =
_
&

_
o&
&

_
&

_
o

.
Example: suppose that you want to know how much utility, n(r, j), changes if r and j are
perturbed. Then
dn =
0n
0r
dr +
0n
0j
dj .
Now, if you imposed that utility is not changing, i.e. you are interested in an isoquant (the
indierence curve), then this implies that dn = 0 and then
dn =
0n
0r
dr +
0n
0j
dj = 0
and hence
dj
dr
=
0n,0r
0n,0j
.
This should not be understood as a derivative, but rather as a ratio of perturbations. Soon we
will characterize conditions under which this is actually a derivative of an implicit function (the
implicit function theorem).
Log linearization. Suppose that you want to log-linearize . = ) (r, j) around some point,
say (r
+
, j
+
, .
+
). This means nding the percent change in . in response to a percent change in r
and j. We have
d. =
0.
0r
dr +
0.
0j
dj .
29
Divide through by .
+
to get
d.
.
+
=
r
+
.
+
0.
0r
_
dr
r
+
_
+
j
+
.
+
0.
0j
_
dj
j
+
_
. =
r
+
.
+
0.
0r
r +
j
+
.
+
0.
0j
j ,
where
. =
d.
.
+
- d ln.
is approximately the percent change.
Another example:
1 = C +1 +G
d1 = dC +d1 +dG
d1
1
=
C
1
dC
C
+
1
1
d1
1
+
G
1
dG
G

1 =
C
1

C +
1
1

1 +
G
1

G .
8.3 The implicit function theorem
This is a useful tool to study the behavior of an equilibrium in response to a change in an exogenous
variable.
Consider
1 (r, j) = 0 .
We are interested in characterizing the implicit function between r and j, if it exists. We already
saw one implicit function when we computed the utility isoquant (indierence curve). In that case,
we had
n(r, j) = n
for some constant level of n. This can be rewritten in the form above as
n(r, j) n = 0 .
From this we derived a dj,dr slope. But this can be more general and constitute a function.
Another example: what is the slope of a tangent line at any point on a circle?
r
2
+j
2
= r
2
r
2
+j
2
r
2
= 0
1 (r, j) = 0
30
Taking the total dierential
1
a
dr +1
j
dj = 2rdr + 2jdj = 0
dj
dr
=
r
j
, j ,= 0 .
For example, the slope at
_
r,
_
2, r,
_
2
_
is 1.
The implicit function theorem: Let the function 1 (r, j) C
1
on some open set and
1 (r, j) = 0. Then there exists a (implicit) function j = ) (r) C
1
that satises 1 (r, ) (r)) = 0,
such that
dj
dr
=
1
a
1
j
on this open set.
More generally, if 1 (j, r
1
, r
2
, ...r
a
) C
1
on some open set and 1 (j, r
1
, r
2
, ...r
a
) = 0, then
there exists a (implicit) function j = ) (r
1
, r
2
, ...r
a
) C
1
that satises 1 () (r) , r) = 0, such that
dj =
a

i=1
)
i
dr
i
.
This gives us the relationship between small perturbations of the rs and perturbation of j.
If we allow only one specic r
i
to be perturbed, then )
i
=
0j
0a
i
= 1
a
i
,1
j
. From1 (j, r
1
, r
2
, ...r
a
) =
0 and j = ) (r
1
, r
2
, ...r
a
) we have
01
0j
dj +
01
0r
1
dr
1
+... +
01
0r
a
dr
a
= 0
dj = )
1
dr
1
+... +)
a
dr
a
so that
01
0j
()
1
dr
1
+... +)
a
dr
a
) +1
a
1
dr
1
+... +1
an
dr
a
= (1
a
1
+1
j
)
1
) dr
1
+... +(1
an
+1
j
)
a
) dr
a
= 0 .
This gives us a relationship between perturbations of the rs. If we only allow r
i
to be perturbed,
dr
i
= 0, then (1
a
i
+1
j
)
i
) = 0 and so )
i
= 1
a
i
,1
j
, as above.
8.4 General version of the implicit function theorem
Implicit Function Theorem: Let 1 (r, j) = 0 be a set of n functions where r
n1
(exogenous)
and j
n1
(endogenous). Note that there are n equations in n unknown endogenous variables. If
1. 1 C
1
and
2. [J[ =

01
0j
0

,= 0 at some point (r
0
, j
0
) (no functional dependence),
31
then j = ) (r), a set of : functions in a neighborhood of (r
0
, j
0
) such that ) C
1
and
1 (r, ) (r)) = 0 in that neighborhood of (r
0
, j
0
).
We further develop this. From 1 (r, j) = 0 we have
_
01
0j
t
_
aa
dj
a1
+
_
01
0r
t
_
an
dr
n1
= 0 =
_
01
0j
t
_
dj =
_
01
0r
t
_
dr . (3)
Since [J[ = [01,0j
t
[ , = 0, then [01,0j
t
]
1
exists and we can write
dj =
_
01
0j
t
_
1
_
01
0r
t
_
dr . (4)
So there is a mapping from dr to dj.
From j = ) (r) we have
dj
a1
=
_
0j
0r
t
_
an
dr
n1
Combining into (3) we get
_
01
0j
t
_
aa
_
0j
0r
t
_
an
dr
n1
=
_
01
0r
t
_
an
dr
n1
.
Now suppose that only r
1
is perturbed, so that dr
t
=
_
dr
1
0 0

. Then we get only the
rst column in the set of equations above:
row 1 :
_
01
1
0j
1
0j
1
0r
1
+
01
1
0j
2
0j
2
0r
1
+... +
01
1
0j
a
0j
a
0r
1
_
dr
1
=
01
1
0r
1
dr
1
.
.
.
row : :
_
01
a
0j
1
0j
1
0r
1
+
01
a
0j
2
0j
2
0r
1
+... +
01
a
0j
a
0j
a
0r
1
_
dr
1
=
01
a
0r
1
dr
1
By eliminating the dr
1
terms we get
row 1 :
_
01
1
0j
1
0j
1
0r
1
+
01
1
0j
2
0j
2
0r
1
+... +
01
1
0j
a
0j
a
0r
1
_
=
01
1
0r
1
.
.
.
row : :
_
01
a
0j
1
0j
1
0r
1
+
01
a
0j
2
0j
2
0r
1
+... +
01
a
0j
a
0j
a
0r
1
_
=
01
a
0r
1
and thus, stacking together
_
01
0j
t
_
aa
_
0j
0r
1
_
a1
=
_
01
0r
1
_
a1
.
32
Since we required [J[ =

01
0j
0

,= 0 it follows that the

_
01
0j
0
_
aa
matrix is nonsingular, and thus
!
_
0j
0a
1
_
a1
, a solution to the system. This can be obtained by Cramers rule:
0j
)
0r
1
=
[J
)
[
[J[
,
where [J
)
[ is obtained by replacing the ,
tI
column in [J
)
[ by
_
01
0a
1
_
. In fact, we could have jumped
directly to here from (4).
Why is this useful? We are often interested in how a model behaves around some point, usually
an equilibrium or a steady state. But models are typically nonlinear and the behavior is hard to
characterize without implicit functions. Think of r as exogenous and j as endogenous. So this
gives us a method for evaluating how several endogenous variables respond to a small change in
one an exogenous variable or policy while holding all other rs constant. This describes a lot of
what we do in economics.
A fuller description of whats going on:
_

_
01
1
0j
1
01
1
0j
2

01
1
0jn
01
2
0j
1
01
2
0j
2

01
2
0jn
.
.
.
.
.
.
.
.
.
.
.
.
01
n
0j
1
01
n
0j
2

01
n
0jn
_

_
_

_
dj
1
dj
2
.
.
.
dj
a
_

_
+
_

_
01
1
0a
1
01
1
0a
2

01
1
0am
01
2
0a
1
01
2
0a
2

01
2
0am
.
.
.
.
.
.
.
.
.
01
n
0a
1
01
n
0a
2

01
n
0am
_

_
_

_
dr
1
dr
2
.
.
.
dr
n
_

_
= 0
_

_
01
1
0j
1
01
1
0j
2

01
1
0jn
01
2
0j
1
01
2
0j
2

01
2
0jn
.
.
.
.
.
.
.
.
.
.
.
.
01
n
0j
1
01
n
0j
2

01
n
0jn
_

_
_

_
dj
1
dj
2
.
.
.
dj
a
_

_
=
_

_
01
1
0a
1
01
1
0a
2

01
1
0am
01
2
0a
1
01
2
0a
2

01
2
0am
.
.
.
.
.
.
.
.
.
01
n
0a
1
01
n
0a
2

01
n
0am
_

_
_

_
dr
1
dr
2
.
.
.
dr
n
_

_
_

_
dj
1
dj
2
.
.
.
dj
a
_

_
=
_

_
0j
1
0a
1
0j
1
0a
2

0j
1
0am
0j
2
0a
1
0j
2
0a
2

0j
2
0am
.
.
.
.
.
.
.
.
.
0j
n
0a
1
0j
n
0a
2

0j
n
0am
_

_
_

_
dr
1
dr
2
.
.
.
dr
n
_

_
= 0
_

_
01
1
0j
1
01
1
0j
2

01
1
0jn
01
2
0j
1
01
2
0j
2

01
2
0jn
.
.
.
.
.
.
.
.
.
.
.
.
01
n
0j
1
01
n
0j
2

01
n
0jn
_

_
_

_
0j
1
0a
1
0j
1
0a
2

0j
1
0am
0j
2
0a
1
0j
2
0a
2

0j
2
0am
.
.
.
.
.
.
.
.
.
0jn
0a
1
0jn
0a
2

0jn
0am
_

_
_

_
dr
1
dr
2
.
.
.
dr
n
_

_
=
_

_
01
1
0a
1
01
1
0a
2

01
1
0am
01
2
0a
1
01
2
0a
2

01
2
0am
.
.
.
.
.
.
.
.
.
01
n
0a
1
01
n
0a
2

01
n
0am
_

_
_

_
dr
1
dr
2
.
.
.
dr
n
_

_
33
8.5 Example: demand-supply system
8.5.1 Using the implicit function theorem
demand :
o
= d(

j,
+
j)
supply :
c
= :(
+
j)
equilibrium :
o
=
c
.
Let d, : C
1
. By eliminating we get
:(
+
j) d(

j,
+
j) = 0 ,
which is an implicit function
1 (j, j) = 0 ,
where j is endogenous and j is exogenous.
We are interested in how the endogenous price responds to income. By the implicit function
theorem j = j (j) such that
dj
dj
=
1
j
1
j
=
d
j
:
j
d
j
=
d
j
:
j
d
j
0
because d
j
< 0. An increase in income unambiguously increases the price.
To nd how quantity changes we apply the total derivative approach to the demand function:
d
dj
=
0d
0j
dj
dj
. .
"substitution eect"<0
+
0d
0j
..
"income eect"0
so the sign here is ambiguous. The income eect is the shift outwards of the demand curve. If
supply did not respond to price (innite elasticity), then that would be it. The substitution eect
is the shift along the (shifted) demand curve that is invoked by the increase in price. But we can
show that d,dj is positive by using the supply side:
d
dj
=
0:
0j
dj
dj
0 .
Draw demand-supply system.
This example is simple, but the technique is very powerful, especially in nonlinear general
equilibrium models.
34
8.5.2 Using the implicit function theorem in a system of two equations
Now consider the system by writing it as a system of two implicit functions:
1 (j, ; j) = 0
1
1
(j, , j) = d (j, j) = 0
1
2
(j, , j) = : (j) = 0 .
Apply the general theorem. Check for functional dependence in the endogenous variables:
[J[ =

01
0 (j, )

d
j
1
:
j
1

= d
j
+:
j
0 .
So there is no functional dependence. Thus j = j (j) and = (j). We now wish to compute
the derivatives with respect to the exogenous argument j. Since d1 = 0 we have
01
1
0j
dj +
01
1
0
d +
01
1
0j
dj = 0
01
2
0j
dj +
01
2
0
d +
01
2
0j
dj = 0
Thus
_
01
1
0j
01
1
0q
01
2
0j
01
2
0q
_
_
dj
d
_
=
_
01
1
0j
dj
01
2
0j
dj
_
Use the following
dj =
0j
0j
dj
d =
0
0j
dj
to get
_
01
1
0j
01
1
0q
01
2
0j
01
2
0q
__
0j
0j
dj
0q
0j
dj
_
=
_
01
1
0j
dj
01
2
0j
dj
_
_
01
1
0j
01
1
0q
01
2
0j
01
2
0q
__
0j
0j
0q
0j
_
=
_
01
1
0j
01
2
0j
_
Using the expressions for 1
1
and 1
2
we get
_
0o
0j
1
0c
0j
1
__
0j
0j
0q
0j
_
=
_

0o
0j
0
_
.
35
We seek a solution for
0j
0j
and
0q
0j
. This is a system of equations, which we solve using Cramers
rule:
0j
0j
=
[J
1
[
[J[
=

0o
0j
1
0 1

[J[
=
0o
0j
[J[
0
and
0
0j
=
[J
2
[
[J[
=

0o
0j

0o
0j
0c
0j
0

[J[
=
0o
0j
0c
0j
[J[
0 .
Try this with three functions for three endogenous variables, i.e. 1
_
j,
c
,
o
; j
_
= 0.
8.5.3 Using the total derivative approach
Now we use the total derivative approach. We have
: (j) d (j, j) = 0 .
Take the total derivative with respect to j:
0:
0j
dj
dj

0d
0j
dj
dj

0d
0j
= 0
Thus
dj
dj
_
0:
0j

0d
0j
_
=
0d
0j
and so
dj
dj
=
0o
0j
0c
0j

0o
0j
0 .
36
9 Optimization with one variable and Taylor expansion
A function may have many local minimum and maximum. A function may have only one global
minimum and maximum, if it exists.
9.1 Local maximum, minimum
First order necessary conditions (FONC): Let ) C
1
on some open convex set (will be
dened properly later) around r
0
. If ) (r
0
)
t
= 0, then r
0
is a critical point, i.e. it could be either
a maximum or minimumor neither.
1. r
0
is a local maximum if )
t
(r
0
) changes from positive to negative as r increases around
r
0
.
2. r
0
is a local minimum if )
t
(r
0
) changes from negative to positive as r increases around r
0
.
3. Otherwise, r
0
is an inection point (not max nor min).
Second order sucient conditions (SOC): Let ) C
2
on some open convex set around r
0
.
If ) (r
0
)
t
= 0 (FONC satised) then:
1. r
0
is a local maximum if )
tt
(r
0
) < 0 around r
0
.
2. r
0
is a local minimum if )
tt
(r
0
) 0 around r
0
.
3. Otherwise ()
tt
(r
0
) = 0) we cannot be sure.
Extrema at the boundaries: if the domain of ) (r) is bounded, then the boundaries may be
extrema without satisfying any of the conditions above.
Draw graphs for all cases.
Example:
j = r
3
12r
2
+ 36r + 8
FONC:
)
t
(r) = 3r
2
24r + 36 = 0
r
2
8r + 12 = 0
r
2
2r 6r + 12 = 0
r(r 2) 6 (r 2) = 0
(r 6) (r 2) = 0
37
r
1
= 6, r
2
= 2 are critical points and both satisfy the FONC.
)
tt
(r) = 6r 24
)
tt
(2) = 12 = maximum
)
tt
(6) = +12 = minimum
9.2 The N
tI
derivative test
If )
t
(r
0
) = 0 and the rst non zero derivative at r
0
is of order :, )
(a)
(r
0
) ,= 0, then
1. If : is even and )
(a)
(r
0
) < 0 then r
0
is a local maximum.
2. If : is even and )
(a)
(r
0
) 0 then r
0
is a local minimum.
3. Otherwise r
0
is an inection point.
Example:
) (r) = (7 r)
4
.
)
t
(r) = 4 (7 r)
3
,
so r = 7 is a critical point (satises the FONC).
)
tt
(r) = 12 (7 r)
2
, )
tt
(7) = 0
)
ttt
(r) = 24 (7 r) , )
ttt
(7) = 0
)
tttt
(r) = 24 0 ,
so r = 7 is a minimum: )
(4)
is the rst non zero derivative. 4 is even. )
(4)
0.
Understanding the
tI
derivative test is based on Maclaurin expansion and Taylor expan-
sion.
9.3 Maclaurin expansion
Terms of art:
Expansion: express a function as a polynomial.
Around r
0
: in a small neighborhood of r
0
.
38
Consider the following polynomial
) (r) = a
0
+a
1
r +a
2
r
2
+a
3
r
3
+... +a
a
r
a
)
(1)
(r) = a
1
+ 2a
2
r + 3a
3
r
2
+... +:a
a
r
a1
)
(2)
(r) = 2a
2
+ 2 3a
3
r +... + (: 1) :a
a
r
a2
.
.
.
)
(a)
(r) = 1 2 ... (: 1) :a
a
.
Evaluate at r = 0:
) (0) = a
0
= 0!a
0
)
(1)
(0) = a
1
= 1!a
1
)
(2)
(0) = 2a
2
= 2!a
2
.
.
.
)
(a)
(0) = 1 2 ... (: 1) :a
a
= :!a
a
.
Therefore
a
a
=
)
(a)
:!
.
Using the last results gives the Maclaurin expansion around 0:
) (r)[
a=0
=
) (0)
0!
+
)
(1)
(0)
1!
r +
)
(2)
(0)
2!
r
2
+
)
(3)
(0)
3!
r +...
)
(a)
(0)
:!
r
a
.
9.4 Taylor expansion
Example: quadratic equation.
) (r) = a
0
+a
1
r +a
2
r
2
.
Dene r = r
0
+c, where we x r
0
as an anchor and allow c to vary. This is essentially relocating
the origin to (r
0
, ) (r
0
)).
q (c) = a
0
+a
1
(r
0
+c) +a
2
(r
0
+c)
2
= ) (r) .
Note that
q (c) = ) (r) and q (0) = ) (r
0
) .
39
Taking derivatives
q
t
(c) = a
1
+ 2a
2
(r
0
+c) = a
1
+ 2a
2
r
0
+ 2a
2
c
q
tt
(c) = 2a
2
.
Use Maclaurins expansion for q (c) around c = 0:
q (c)[
c=0
=
q (0)
0!
+
q
(1)
(0)
1!
c +
q
(2)
(0)
2!
c
2
.
Using c = r r
0
and the fact that r = r
0
when c = 0, we get a Maclaurin expansion for ) (r)
around r = r
0
:
) (r)[
a=a
0
=
) (r
0
)
0!
+
)
(1)
(r
0
)
1!
(r r
0
) +
)
(2)
(r
0
)
2!
(r r
0
)
2
.
More generally, we have the Taylor expansion for an arbitrary C
a
function:
) (r)[
a=a
0
=
) (r
0
)
0!
+
)
(1)
(r
0
)
1!
(r r
0
) +
)
(2)
(r
0
)
2!
(r r
0
)
2
+... +
)
(a)
(r
0
)
:!
(r r
0
)
a
+1
a
= 1
a
+1
a
,
where 1
a
is a remainder (Theorem):
As we choose higher :, then 1
a
will be smaller and in the limit vanish.
As r is farther away from r
0
1
a
may grow.
The Lagrange form of 1
a
: for some point j [r
0
, r] (if r r
0
) or j [r, r
0
] (if r < r
0
) we
have
1
a
=
1
(: + 1)!
)
(a+1)
(j) (r r
0
)
a+1
.
Example: for : = 0 we have
) (r)[
a=a
0
=
) (r
0
)
0!
+1
a
= ) (r
0
) +1
a
= ) (r
0
) +)
t
(j) (r r
0
) .
Rearranging this we get
) (r) ) (r
0
) = )
t
(j) (r r
0
)
for some point j [r
0
, r] (if r r
0
) or j [r, r
0
] (if r < r
0
). This is the Mean Value Theorem:
40
9.5 Taylor expansion and the N-th derivative test
Dene: r
0
is a maximum (minimum) of ) (r) if the change in the function, ) = ) (r) ) (r
0
), is
negative (positive) in a neighborhood of r
0
, both on the right and on the left of r
0
.
The Taylor expansion helps determining this.
) = )
(1)
(r
0
) (r r
0
)+
)
(2)
(r
0
)
2
(r r
0
)
2
+...+
)
(a)
(r
0
)
:!
(r r
0
)
a
+
1
(: + 1)!
)
(a+1)
(j) (r r
0
)
a+1
. .
remainder
.
1. Consider the case that )
t
(r
0
) ,= 0, i.e. the rst non zero derivative at r
0
is of order 1. Choose
: = 0, so that the remainder will be of the same order of the rst non zero derivative and
evaluate
) = )
t
(j) (r r
0
) .
Using the fact that j is very close to r
0
, so close that )
t
(j) ,= 0, we have that ) changes
signs around r
0
, because (r r
0
) changes sign around r
0
.
2. Consider the case of )
t
(r
0
) = 0 and )
tt
(r
0
) ,= 0. Choose : = 1, so that the remainder will
be of the same order of the rst non zero derivative (2) and evaluate
) = )
t
(r
0
) (r r
0
) +
)
tt
(j)
2
(r r
0
)
2
=
1
2
)
tt
(j) (r r
0
)
2
.
Since (r r
0
)
2
0 always and )
tt
(j) ,= 0 we get ) is either positive (minimum) or negative
(maximum) around r
0
.
3. Consider the case of )
t
(r
0
) = 0, )
tt
(r
0
) = 0 and )
ttt
(r
0
) ,= 0. Choose : = 2, so that the
41
remainder will be of the same order of the rst non zero derivative (3) and evaluate
) = )
t
(r
0
) (r r
0
) +
)
tt
(j)
2
(r r
0
)
2
+
)
ttt
(j)
6
(r r
0
)
3
=
1
6
)
ttt
(j) (r r
0
)
3
.
Since (r r
0
)
3
changes signs around r
0
and )
ttt
(j) ,= 0 we get ) is changing signs and
therefore not an extremum.
4. In the general case )
t
(r
0
) = 0, )
tt
(r
0
) = 0, ... )
(a1)
(r
0
) = 0 and )
(a)
(r
0
) ,= 0. Choose
:1, so that the remainder will be of the same order of the rst non zero derivative (:) and
evaluate
) = )
(1)
(r
0
) (r r
0
) +
)
(2)
(r
0
)
2
(r r
0
)
2
+... +
)
(a1)
(r
0
)
(: 1)!
(r r
0
)
a1
+
1
:!
)
(a)
(j) (r r
0
)
a
=
1
:!
)
(a)
(j) (r r
0
)
a
.
In all cases )
(a)
(j) ,= 0.
If : is odd, then (r r
0
)
a
changes signs around r
0
and ) changes signs and therefore not an
extremum.
If : is even, then (r r
0
)
a
0 always and ) is either positive (minimum) or negative
(maximum).
Warning: in all the above we need ) C
a
at r
0
. For example,
) (r) =
_
c

1
2
a
2
r ,= 0
0 r = 0
is not C
1
at 0, and yet r = 0 is the minimum.
42
10 Exponents and logs
These are used a lot in economics due to their useful properties, some of which have economic
interpretations, in particular in dynamic problems that involve time.
10.1 Exponent function
j = ) (t) = /
t
, / 1 .
(the case of 0 < / < 1 can be dealt with similarly.)
) (t) C
o
.
) (t) 0 \t R (since / 1 0).
)
t
(t) 0, )
tt
(t) 0, therefore strictly increasing and so t = )
1
(j) = log
b
j, where
j R
++
.
Any j 0 can be expressed as an exponent of many bases. Make sure you know how to
convert bases:
log
b
j =
log
o
j
log
o
/
.
10.2 The constant e
The expression
j = c
vt
describes constantly growing processes.
d
dt
c
t
= c
t
d
dt
_
c
vt
_
= rc
vt
.
It turns out that
lim
no
_
1 +
1
:
_
n
= lim
a0
(1 +:)
1a
= c = 2.71828...
Think of 1,: = : as time. To see this, use a Taylor expansion of c
a
and evaluate it around zero:
c
a
= c
0
+
1
1!
(c
a
)
t

a=0
(r 0) +
1
2!
(c
a
)
tt

a=0
(r 0)
2
+
1
3!
(c
a
)
ttt

a=0
(r 0)
3
+...
= 1 +r +
1
2!
r
2
+
1
3!
r
3
+...
43
Evaluate this at r = 1:
c
1
= c = 1 + 1 +
1
2!
+
1
3!
+... = 2.71828...
10.3 Examples
10.3.1 Interest compounding
Suppose that you are oered an interest rate r on your savings after a year. Then the return after
one year is 1 +r. If you invested , then at the end of the year you have
(1 +r) .
Now suppose that an interest of
_
v
n
_
is oered for each 1,: of a year. In that case you get a
_
v
n
_
return compounded : times throughout the year. In that case an investment of will be worth
at the end of the year

_
1 +
r
:
_
n
=
_
_
1 +
r
:
_
nv
_
v
.
Now suppose that you get a instant rate of interest r for each instant (a period of length 1,:,
where : ), compounded : times throughout the year. In that case an investment of
will be worth at the end of the year
lim
no

_
1 +
r
:
_
n
= lim
no

_
_
1 +
r
:
_
nv
_
v
=
_
lim
no
_
1 +
r
:
_
nv
_
v
=
_
lim
&=vn0
(1 +n)
1&
_
v
= c
v
.
Thus, r is the instantaneous rate of return.
Suppose that we are interested in an arbitrary period of time, t, where, say t = 1 is a year (but
this is arbitrary). Then the same kind of math will lead us to nd the value of an investment
after t time to be

_
1 +
r
:
_
nt
=
_
_
1 +
r
:
_
nv
_
vt
.
If : is nite, then that is it. if we get continuous compounding (: ), then the value of the
investment after t time will be
c
vt
.
10.3.2 Growth rates
The interest rate example tells you how much the investment is worth when it grows at a constant,
instantaneous rate:
growth rate =
d\,dt
\
=
rc
vt
c
vt
= r per instant (dt).
44
Any discrete growth rate can be described by a continuous growth rate:
(1 +i)
t
= c
vt
,
where
(1 +i) = c
v
.
10.3.3 Discounting
The value today of A t periods in the future is
PV =
A
(1 +i)
t
,
where 1, (1 +i)
t
is the discount factor. This can also be represented by continuous discounting
PV =
A
(1 +i)
t
= Ac
vt
,
where the same discount factor is 1, (1 +i)
t
= (1 +i)
t
= c
vt
.
10.4 Logarithms
Log is the inverse function of the exponent. For / 1, t R, j R + +
j = /
t
= t = log
b
j .
This is very useful, e.g. for regressions analysis.
E.g.,
2
4
= 16 = 4 = log
2
16 .
5
3
= 125 = 3 = log
5
125 .
Also, note that
j = /
log
b
j
.
Convention:
log
c
r = lnr .
Rules:
ln(n) = lnn + ln
ln(n,) = lnn ln
45
ln
_
an
b
_
= lna +/ lnn
log
b
r =
log
a
a
log
a
b
, where a, /, r 0
Corollary: log
b
c =
ln c
ln b
=
1
ln b
Some useful properties of logs:
1. Log dierences approximate growth rates:
lnA
2
lnA
1
= ln
A
2
A
1
= ln
_
A
2
A
1
1 + 1
_
= ln
_
1 +
A
2
A
1
A
1
_
= ln(1 +r) ,
where r is the growth rate of A. Take a rst order Taylor approximation of ln(1 +r) around
ln(1):
ln(1 +r) - ln(1) + (ln(1))
t
(1 +r 1) = r .
So we have
lnA
2
lnA
1
- r .
This approximation is good for small percent changes. Beware: large log dierences give
much larger percent changes (e.g., a log dierence of 1=100% is 2.7=270%).
2. Logs "bend down" their image relative to the argument below the 45 degree line. Exponents
do the opposite.
3. The derivative of log is always positive, but ever diminishing: (log r)
t
0, (log r)
tt
< 0.
4. Nevertheless, lim
ao
log
b
r = . Also, lim
a0
log
b
r = . Therefore the range is R.
5. Suppose that j = c
vt
. Then lnj = ln+rt. Therefore
t =
lnj ln
r
.
This answers the question: how long will it take to grow from to j, if growth is at an
instantaneous rate of r.
6. Converting j = /
ct
into j = c
vt
: /
c
= c
v
, therefore c ln/ = r, therefore j = c
vt
= j =
c
(c ln b)t
.
10.5 Derivatives of exponents and logs
d
dt
lnt =
1
t
d
dt
log
b
t =
d
dt
lnt
ln/
=
1
t ln/
46
d
dt
c
t
= c
t
Let j = c
t
, so that t = lnj:
d
dt
c
t
=
d
dt
j =
1
dt,dj
=
1
1,j
= j = c
t
.
By chain rule:
d
dt
c
&
= c
&
dn
dt
d
dt
lnn =
dn,dt
n
Higher derivatives:
d
a
(dt)
a
c
t
= c
t
d
dt
lnt =
1
t
,
d
2
(dt)
2
lnt =
1
t
2
,
d
3
(dt)
3
lnt =
2
t
3
...
d
dt
/
t
= /
t
ln/ ,
d
2
(dt)
2
/
t
= /
t
(ln/)
2
,
d
3
(dt)
3
/
t
= /
t
(ln/)
3
...
10.6 Application: optimal timing
The value of / bottles of wine is given by
\ (t) = /c
_
t
.
Discounting: 1(t) = c
vt
. The present value of \ (t) today is
1\ = 1(t) \ (t) = c
vt
/c
_
t
= /c
_
tvt
.
Choosing t to maximize 1\ = /c
_
tvt
is equivalent to choosing t to maximize ln1\ = ln/+
_
trt.
FONC:
0.5t
0.5
r = 0
0.5t
0.5
= r
Marginal benet to wait one more instant = marginal cost of waiting one more instant. t
+
=
1,
_
4r
2
_
.
SOC:
0.25t
1.5
< 0
so t
+
is a maximum.
47
10.7 Growth rates again
Denote
d
dt
r = _ r .
So the growth rate at some point in time is
dr,dt
r
=
_ r
r
.
So in the case r = c
vt
, we have
_
\
\
= r .
And since r(0) = c
v0
= , we can write without loss of generality r(t) = r
0
c
vt
.
Growth rates of combinations:
1. For j (t) = n(t) (t) we have
_ j
j
=
_ n
n
+
_

q
j
= q
&
+q

Proof:
lnj (t) = lnn(t) + ln (t)
d
dt
lnj (t) =
d
dt
lnn(t) +
d
dt
ln (t)
1
j (t)
dj
dt
=
1
n(t)
dn
dt
+
1
(t)
d
dt
2. For j (t) = n(t) , (t) we have
_ j
j
=
_ n
n

_

q
j
= q
&
q

Proof: similar to above.

3. For j (t) = n(t) (t) we have
q
j
=
n
n
q
&

n
n
q

48
10.8 Elasticities
An elasticity of j with respect to r is dened as
o
j,a
=
dj,j
dr,r
=
dj
dr
r
j
.
Since
d lnr =
0 lnr
0r
dr =
dr
r
we get
o
j,a
=
d lnj
d lnr
.
49
11 Optimization with more than one choice variable
11.1 The dierential version of optimization with one variable
This helps developing concepts for what follows. Let . = ) (r) C
1
, r R. Then
d. = )
t
(r) dr .
FONC: an extremum may occur when d. = 0, i.e. when )
t
(r) = 0. Think of this condition as
a situation when small arbitrary perturbations of r do not aect the value of the function;
therefore dr ,= 0 in general. No perturbation of the argument (dr = 0) will trivially not
induce perturbation of the image.
SOC:
d
2
. = d [d.] = d
_
)
t
(r) dr

= )
tt
(r) dr
2
.
A maximum occurs when )
tt
(r) < 0 or equivalently when d
2
. < 0.
A minimum occurs when )
tt
(r) 0 or equivalently when d
2
. 0.
11.2 Extrema of a function of two variables
Let . = ) (r, j) C
1
, r, j R. Then
d. = )
a
dr +)
j
dj .
FONC: d. = 0 for arbitrary values of dr and dj, not both equal to zero. A necessary condition
that gives this is
)
a
= 0 and )
j
= 0 .
As before, this is not a sucient condition for an extremum, not only because of inection points,
but also due to saddle points.
Note: in matrix notation
d. =
_
0)
0 (r, j)
_ _
dr
dj
_
= \)dr =
_
)
a
)
j

_
dr
dj
_
= )
a
dr +)
j
dj .
If r R
a
then
d. =
_
0)
0r
t
_
dr = \)dr =
_
)
1
)
a

_

_
dr
1
.
.
.
dr
a
_

_ =
a

i=1
)
i
dr
i
.
50
Dene
)
aa
=
0
2
)
0r
2
)
jj
=
0
2
)
0j
2
)
aj
=
0
2
)
0r0j
)
ja
=
0
2
)
0j0r
Youngs Theorem: If both )
aj
and )
ja
are continuous, then )
aj
= )
ja
.
Now we apply this
d
2
. = d [d.] = d [)
a
dr +)
j
dj] = d [)
a
dr] +d [)
j
dj]
= )
aa
dr
2
+)
ja
drdj +)
aj
djdr +)
jj
dj
2
= )
aa
dr
2
+ 2)
aj
drdj +)
jj
dj
2
.
(The d [dr] and d [dj] terms drop out. The reason is that we are considering dr and dj as variables,
but once they are set they do not change.) In matrix notation
d
2
. =
_
dr dj

_
)
aa
)
aj
)
aj
)
jj
_ _
dr
dj
_
.
And more generally, if r R
a
then
d
2
. = dr
t
_
0
2
)
0r0r
t
_
. .
Hessian
dr .
SONC (second order necessary conditions): for arbitrary values of dr and dj
d
2
. _ 0 gives a maximum.
d
2
. _ 0 gives a minimum.
SOSC (second order sucient conditions): for arbitrary values of dr and dj
d
2
. < 0 gives a maximum. In the two variable case
d
2
. < 0 i )
aa
< 0, )
jj
< 0 and )
aa
)
jj
)
2
aj
.
d
2
. 0 gives a minimum. In the two variable case
d
2
. 0 i )
aa
0, )
jj
0 and )
aa
)
jj
)
2
aj
.
51
Comments:
SONC is necessary but not sucient, while SOSC are not necessary.
If )
aa
)
jj
= )
2
aj
a point can be an extremum nonetheless.
If )
aa
)
jj
< )
2
aj
then this is a saddle point.
If )
aa
)
jj
)
2
aj
0, then )
aa
)
jj
)
2
aj
_ 0 implies sign()
aa
) =sign()
jj
).
11.3 Quadratic form and sign deniteness
This is a tool to help analyze SOCs. Relabel terms for convenience:
. = ) (r
1
, r
2
)
d
2
. = , dr
1
= d
1
, dr
2
= d
2
)
11
= a , )
22
= / , )
12
= /
Then
d
2
. = )
11
dr
2
1
+ 2)
12
dr
1
dr
2
+)
22
dr
2
2
= ad
2
1
+ 2/d
1
d
2
+/d
2
2
=
_
d
2
d
1

_
a /
/ /
_ _
d
1
d
2
_
.
This is the quadratic form.
Note: d
1
and d
2
are variables, not constants, as in the FONC. We require the SOCs to hold
\d
1
, d
2
, and in particular \d
1
, d
2
,= 0.
Denote the Hessian by
H =
_
0
2
)
0r0r
t
_
The quadratic form is
= d
t
Hd
Dene
is
_

_
positive denite
positive semidenite
negative semidenite
negative denite
_

_
if is invariably
_

_
0
_ 0
_ 0
< 0
_

_
,
regardless of values of d. Otherwise, is indenite.
52
Consider the determinant of H, [H[, which we call here the discriminant of H:
is
_
positive denite
negative denite
_
i
_
[a[ 0
[a[ < 0
_
and [H[ 0 .
[a[ is (the determinant of) the rst ordered minor of H. In the simple two variable case, [H[ is (the
determinant of) the second ordered minor of H. In that case
[H[ = a/ /
2
.
If [H[ 0, then a and / must have the same sign, since a/ /
2
0.
11.4 Quadratic form for n variables and sign deniteness
= d
t
Hd =
a

i=1
a

)=1
/
i)
d
i
d
)
.
is positive denite i all (determinants of) the principal minors are positive
[H
1
[ = [/
11
[ 0, [H
2
[ =

/
11
/
12
/
21
/
22

0, ... [H
a
[ = [H[ 0 .
is negative denite i (determinants of) the odd principal minors are negative and the
even ones are positive:
[H
1
[ < 0, [H
2
[ 0, [H
3
[ < 0, ...
11.5 Characteristic roots test for sign deniteness
Consider some :: matrix H
aa
. We look for a characteristic root r (scalar) and characteristic
vector r
a1
(: 1) such that
Hr = rr .
Developing this expression:
Hr = r1r = (H r1) r = 0 .
Dene (H r1) as the characteristic matrix:
(H r1) =
_

_
/
11
r /
12
/
1a
/
21
/
22
r /
2a
.
.
.
.
.
.
.
.
.
.
.
.
/
a1
/
a2
. . . /
aa
r
_

_
If (H r1) r = 0 has a non trivial solution r ,= 0, then (H r1) must be singular, so that
[H r1[ = 0. This is an equation that we can solve for r. The equation [H r1[ = 0 is the
53
characteristic equation, and is an : degree polynomial in r, with : non trivial solutions (some
of the solutions can be equal). Some properties:
If H is symmetric, then we will have r R. This is useful, because many applications in
economics will deal with symmetric matrices, like Hessians and variance-covariance matrices.
For each characteristic root that solves [H r1[ = 0 there are many characteristic vectors r
such that Hr = rr. Therefore we normalize: r
t
r = 1. Denote the normalized characteris-
tic vectors as . Denote the characteristic vectors (eigenvector) of the characteristic root
(eigenvalue) as
i
and r
i
.
The set of eigenvectors is orthonormal, i.e. orthogonal and normalized:
t
i

)
= 0 \i ,= , and

t
i

i
= 1.
11.5.1 Application to quadratic form
Let \ = (
1
,
2
, ...
a
) be the set of eigenvectors of the matrix H. Dene the vector j that solves
d = \ j. We use this in the quadratic form
= d
t
Hd = j
t
\
t
H\ j = j
t
1j ,
where \
t
H\ = 1. It turns out that
1 =
_

_
r
1
0 0
0 r
2
0
.
.
.
.
.
.
0 0 r
a
_

_
Here is why:
\
t
H\ = \
t
_
H
1
H
2
H
a

=
_

t
1

t
2
.
.
.

t
a
_

_
_
r
1

1
r
2

2
r
a

a

=
_

_
r
1

t
1

1
r
1

t
1

2
r
1

t
1

a
r
2

t
2

1
r
2

t
2

2
r
2

t
2

a
.
.
.
.
.
.
.
.
.
.
.
.
r
a

t
a

1
r
a

t
a

2
r
a

t
a

a
_

_
= 1 ,
where the last equality follows from
t
i

)
= 0 \i ,= , and
t
i

i
= 1. It follows that sign() depends
only on the characteristic roots: = j
t
1j =

a
i=1
r
i
j
2
i
.
54
11.5.2 Characteristic roots test for sign deniteness
is
_

_
positive denite
positive semidenite
negative semidenite
negative denite
_

_
i all r
i
_

_
0
_ 0
_ 0
< 0
_

_
,
regardless of values of d. Otherwise, is indenite.
When : is large, nding the roots can be hard, because it involves nding the roots of a
polynomial of degree :. But the computer can do it for us.
11.6 Global extrema, convexity and concavity
We seek conditions for a global maximum or minimum. If a function has a "hill shape" over its
entire domain, then we do not need to worry about boundary conditions and the local extremum
will be a global extremum. Although the global maximum can be found at the boundary of the
domain, this will not be detected by the FONC.
If ) is strictly concave: the global maximum is unique.
If ) is concave, but not strictly: this allows for at regions, so the global maximum may not
be unique.
Let . = ) (r) C
2
, r R
a
.
If d
2
. is
_

_
positive denite
positive semidenite
negative semidenite
negative denite
_

_
\r in the domain, then ) is
_

_
strictly convex
convex
concave
strictly concave
_

_
,
When an objective function is general, then we must assume convexity or concavity. If a specic
functional form is used, we can check whether it is convex or concave.
11.7 Convexity and concavity dened
Denition 1: A function ) is concave i \) (r) , ) (j) graph of ) the line between ) (r) and
) (j) lies on or below the graph.
If \r ,= j the line lies strictly below the graph, then ) is strictly concave.
For convexity replace "below" with "above".
55
Denition 2: A function ) is concave i \r, j domain of ), which is assumed to be a convex
set (see below), and \0 (0, 1) we have
0) (r) + (1 0) ) (j) _ ) [0r + (1 0) j] .
For strict concavity replace "_" with "<" and add \r ,= j.
For convexity replace "_" with "_" and "<" with "".
The term 0r + (1 0) j, 0 (0, 1) is called a convex combination.
Properties:
1. If ) is linear, then ) is both concave and convex, but not strictly.
2. If ) is (strictly) concave, then ) is (strictly) convex.
Proof: ) is concave. Therefore \r, j domain of ) and \0 (0, 1) we have
0) (r) + (1 0) ) (j) _ ) [0r + (1 0) j] , (1)
0 [) (r)] + (1 0) [) (j)] _ ) [0r + (1 0) j]
3. If ) and q are concave functions, then ) + q is also concave. If one of the concave functions
is strictly concave, then ) +q is strictly concave.
Proof: ) and q are concave, therefore
0) (r) + (1 0) ) (j) _ ) [0r + (1 0) j]
0q (r) + (1 0) q (j) _ q [0r + (1 0) j]
0 [) (r) +q (r)] + (1 0) [) (j) +q (j)] _ ) [0r + (1 0) j] +q [0r + (1 0) j]
0 [() +q) (r)] + (1 0) [() +q) (j)] _ () +q) [0r + (1 0) j]
The proof for strict concavity is identical.
11.7.1 Example
Is . = r
2
+j
2
concave or convex? Consider rst the LHS of the denition:
(i) : 0) (r
1
, j
1
) + (1 0) ) (r
2
, j
2
) = 0
_
r
2
1
+j
2
1
_
+ (1 0)
_
r
2
2
+j
2
2
_
.
56
Now consider the RHS of the denition:
(ii) : ) [0r
1
+ (1 0) r
2
, 0j
1
+ (1 0) j
2
] = [0r
1
+ (1 0) r
2
]
2
+ [0j
1
+ (1 0) j
2
]
2
= 0
2
_
r
2
1
+j
2
1
_
+ (1 0)
2
_
r
2
2
+j
2
2
_
+ 20 (1 0) (r
1
r
2
+j
1
j
2
) .
Now subtract (i)(ii):
0 (1 0)
_
r
2
1
+j
2
1
+r
2
2
+j
2
2
_
20 (1 0) (r
1
r
2
+j
1
j
2
) = 0 (1 0)
_
(r
1
r
2
)
2
+ (j
1
j
2
)
2
_
_ 0 .
So this is a convex function. Moreover, it is strictly convex, since \r
1
,= r
2
and \j
1
,= j
2
we have
(i)(ii) 0.
Using similar steps, you can verify that
_
r
2
+j
2
_
is strictly concave.
11.7.2 Example
Is ) (r, j) = (r +j)
2
concave or convex? Use the same procedure from above.
(i) : 0) (r
1
, j
1
) + (1 0) ) (r
2
, j
2
) = 0 (r
1
+j
1
)
2
+ (1 0) (r
2
+j
2
)
2
.
Now consider
(ii) : ) [0r
1
+ (1 0) r
2
, 0j
1
+ (1 0) j
2
] = [0r
1
+ (1 0) r
2
+0j
1
+ (1 0) j
2
]
2
= [0 (r
1
+j
1
) + (1 0) (r
2
+j
2
)]
2
= 0
2
(r
1
+j
1
)
2
+ 20 (1 0) (r
1
+j
1
) (r
2
+j
2
) + (1 0)
2
(r
2
+j
2
)
2
.
Now subtract (i)(ii):
0 (1 0)
_
(r
1
+j
1
)
2
+ (r
2
+j
2
)
2
_
20 (1 0) (r
1
+j
1
) (r
2
+j
2
)
= 0 (1 0) [(r
1
+j
1
) (r
2
+j
2
)]
2
_ 0 .
So convex but not strictly. Why not strict? Because when r + j = 0, i.e. when j = r, we get
) (r, j) = 0. The shape of this function is a hammock, with the bottom at j = r.
11.8 Dierentiable functions, convexity and concavity
Let ) (r) C
1
and r R. Then ) is concave i \r
1
, r
2
domain of )
)
_
r
2
_
)
_
r
1
_
_ )
t
_
r
1
_ _
r
2
r
1
_
.
For convex replace "_" with "_".
When r
2
r
1
and both r
2
, r
1
R we can divide through by
_
r
2
r
1
_
without changing the
57
direction of the inequality to get
)
t
_
r
1
_
_
)
_
r
2
_
)
_
r
1
_
r
2
r
1
, r
2
r
1
.
I.e. the slope from r
1
to r
2
is smaller than the derivative at r
1
. Think of r
1
as the point of reference
and r
2
as a target point. When r
2
< r
1
we can divide through by
_
r
2
r
1
_
but must change the
direction of the inequality to get
)
t
_
r
1
_
_
)
_
r
2
_
)
_
r
1
_
r
2
r
1
=
)
_
r
1
_
)
_
r
2
_
r
1
r
2
, r
2
< r
1
.
I.e. the slope is larger than the derivative at r
1
.
Derivative Condition for Concave Function Derivative Condition for Convex Function
If r R
a
, then ) C
1
is concave i \r
1
, r
2
domain of )
)
_
r
2
_
)
_
r
1
_
_ \)
_
r
1
_ _
r
2
r
1
_
For convex replace "_" with "_".
Let . = ) (r) C
2
and r R
a
. Then ) is concave i \r domain of ) we have d
2
. is negative
semidenite. If d
2
. is negative denite, then ) is strictly concave (but not "only if"). Replace
"negative" with "positive" for convexity.
11.9 Global extrema, convexity and concavity again
Suppose a point r
0
satises the FONC: you have found a critical point of the function ). Then
you examine the SOC: if = d
2
. is negative (positive) denite, then r
0
is at a local maximum
58
(minimum), i.e. r
0
is a local maximizer (minimizer). This implies examining the Hessian at r
0
.
But if you know something about the concavity/convexity properties of ), then you know
something more. If ) is concave (convex), then you know that if r
0
satises the FONC, then r
0
is at a global maximum (minimum), i.e. r
0
is a global maximizer (minimizer). And if ) is strictly
concave (convex), then you know that r
0
is at a unique global maximum (minimum), i.e. r
0
is a
unique global maximizer (minimizer).
Determining concavity/convexity (strict or not) of a function ) implies examining the Hessian
at all points of its domain. As noted above, sign deniteness of d
2
. is determined by the sign
deniteness of the Hessian. Thus
If H is
_

_
positive denite
positive semidenite
negative semidenite
negative denite
_

_
\r in the domain, then ) is
_

_
strictly convex
convex
concave
strictly concave
_

_
.
11.10 Convex sets in R
a
This is related, but distinct from convex and concave functions.
Dene: convex set in R
a
. Let the set o R
a
. If \r, j o and \0 [0, 1] we have
0r + (1 0) j o
then o is a convex set. (This denition holds in other spaces as well.) Essentially, a set is convex
if it has no "holes" (no doughnuts) and the boundary is not "dented" (no bananas).
11.10.1 Relation to convex functions 1
The concavity condition \r, j domain of ) and \0 (0, 1) we have
0) (r) + (1 0) ) (j) _ ) [0r + (1 0) j]
assumes that the domain is convex: \r, j domain of ) and \0 (0, 1)
0r + (1 0) j domain of ) ,
because ) [0r + (1 0) j] must be dened.
11.10.2 Relation to convex functions 2
Necessary condition for convex function: if ) is a convex function, then \/ R the set
o = r : ) (r) _ /
59
Figure 1: Convex set, but function is not convex
is a convex set.
This is NOT a sucient condition, i.e. the causality runs from convexity of ) to convexity of
o, but not vice versa. Convexity of o does not necessarily imply convexity of ).
If ) is a concave function, then the set
o = r : ) (r) _ / , / R
60
is a convex set. This is NOT a sucient condition, i.e. the causality runs from concavity of ) to
convexity of o, but not vice versa. Convexity of o does not necessarily imply concavity of ).
This is why there is an intimate relationship between convex preferences and concave utility
functions.
11.11 Example: input decisions of a rm
= 1 C = j n| r/ .
Let j, n, r be given, i.e. the rm is a price taker in a competitive economy. To simplify, let output,
, be the numeraire, so that j = 1 and everything is then denominated in units of output:
= n| r/ .
Production function with decreasing returns to scale:
= /
c
|
c
, c < 1,2
so that
= /
c
|
c
n| r/ .
Choose /, | to maximize . FONC:
0
0/
= c/
c1
|
c
r = 0
0
0/
= c/
c
|
c1
n = 0 .
61
SOC: check properties of the Hessian
H =
_

_
0
2

0
_
/
|
_
0
_
/ |
_
_

_
=
_
c(c 1) /
c2
|
c
c
2
/
c1
|
c1
c
2
/
c1
|
c1
c(c 1) /
c
|
c2
_
.
[H
1
[ = c(c 1) /
c2
|
c
< 0 \/, | 0. [H
2
[ = [H[ = c
2
(1 2c) /
2(c1)
|
2(c1)
0 \/, |. Therefore
is a strictly concave function and the extremum will be a maximum.
From the FONC:
c/
c1
|
c
= c

/
= r
c/
c
|
c1
= c

|
= n
so that r/ = n| = c. Thus
/ =
c
r
| =
c
n
.
Using this in the production function:
= /
c
|
c
=
_
c
r
_
c
_
c
n
_
c
= c
2c

2c
_
1
rn
_
c
= c
2
12
_
1
rn
_
12
,
so that
/ = c
1
12
_
1
r
_ 1
12
_
1
n
_
12
| = c
1
12
_
1
r
_
12
_
1
n
_ 1
12
.
62
12 Optimization under equality constraints
12.1 Example: the consumer problem
Objective : Choose r, j to maximize n(r, j)
Constraint(s) : s.t. (r, j) 1 = (r, j) : r, j _ 0, rj
a
+jj
j
_ 1
(draw the budget set, 1). Under some conditions, which we will explore soon, we will get the result
that the consumer chooses a point on the budget line, s.t. rj
a
+ jj
j
= 1, and that r, j _ 0 is
trivially satised (nonsatiation and quasi-concavity of n). So we state a simpler problem:
Objective : Choose r, j to maximize n(r, j)
Constraint(s) : s.t. rj
a
+jj
j
= 1 .
The optimum will be denoted (r
+
, j
+
). The value of the problem is n(r
+
, j
+
). Constraints can
only hurt the unconstrained value (although they may not). This will happen when the uncon-
strained optimum point is not in the constraint set. E.g.,
Choose r, j to maximize r r
2
+j j
2
has a maximum at (r
+
, j
+
) = (1,2, 1,2), but this point is not on the line r + j = 2, so applying
this constraint will move us away from the unconstrained optimum and hurt the objective.
12.2 Lagrange method: one constraint, two variables
Let ), q C
1
. Suppose that (r
+
, j
+
) is the solution to
Choose r, j to maximize . = ) (r, j) , s.t. q (r, j) = c
and that (r
+
, j
+
) is not a critical point of q (r, j), i.e. not both q
a
,= 0 and q
j
,= 0 at (r
+
, j
+
).
Then there exists a number `
+
such that (r
+
, j
+
, `
+
) is a critical point of
/ = ) (r, j) +`[c q (r, j)] ,
i.e.
0/
0`
= c q (r, j) = 0
0/
0r
= )
a
`q
a
= 0
0/
0j
= )
j
`q
j
= 0 .
63
From this it follows that at (r
+
, j
+
, `
+
)
q (r
+
, j
+
) = c
`
+
= )
a
,q
a
`
+
= )
j
,q
j
.
The last equations make it clear why we must check the constraint qualications, that not
both q
a
,= 0 and q
j
,= 0 at (r
+
, j
+
), i.e. check that (r
+
, j
+
) is not a critical point of q (r, j).
For linear constraints this will be automatically satised.
Always write +`[c q (r, j)].
If the constraint qualication fails then this means that we cannot freely search for an optimum.
It implies that the theorem does not apply; it does not imply that there is no optimum. Recall that
the gradient \q (r, j) is a vector that tells you in which direction to move in order to increase q as
much as possible at some point (r, j). But if both q
a
= 0 and q
j
= 0 at (r
+
, j
+
), then this means
that we are not free to search in any direction.
Recall that for unconstrained maximum, we must have
d. = )
a
dr +)
j
dj = 0 ,
and thus
dj
dr
=
)
a
)
j
In the constrained problem this still holdsas we will see belowexcept that now dr and dj are
not arbitrary: they must satisfy the constraint, i.e.
q
a
dr +q
j
dj = 0 .
Thus
dj
dr
=
q
a
q
j
.
From both of these we obtain
q
a
q
j
=
)
a
)
j
,
i.e. the objective and the constraint are tangent. This follows from
)
j
q
j
= ` =
)
a
q
a
.
A graphic interpretation. Think of the gradient as a vector that points in a particular
direction. This direction is where to move in order to increase the function the most, and is
64
perpendicular to the isoquant of the function. Notice that we have
\) (r
+
, j
+
) = `
+
\q (r
+
, j
+
)
()
a
, )
j
) = `
+
(q
a
, q
j
) .
This means that the constraint and the isoquant of the objective at the optimal value are parallel.
They may point in the same direction if ` 0 or in opposite directions if ` < 0.
Gradient Condition for Optimization
In the gure above: the upper curve is given by the isoquant ) (r) = ) (r
+
) and the lower curve
is given by q (r) = c.
12.3 is the shadow cost of the constraint
` tells you how much ) would increase if we relax the constraint by one unit, i.e. increase or
decrease c by 1 (for equality constraints, it will be either-or). For example, if the objective is utility
and the constraint is your budget in euros, then ` is in terms of utils/euro. It tells you how many
more utils you would get if you had one more dollar.
Write the system of equations that dene the optimum as identities
1
1
(`, r, j) = c q (r, j) = 0
1
2
(`, r, j) = )
a
`q
a
= 0
1
2
(`, r, j) = )
j
`q
j
= 0 .
This is a system of functions of the form 1 (`, r, j, c) = 0. If all these functions are C
1
and [J[ , = 0
65
at (`
+
, r
+
, j
+
), where
[J[ =

01
0 (` r j)

0 q
a
q
j
q
a
)
aa
`q
aa
)
aj
`q
aj
q
j
)
aj
`q
aj
)
jj
`q
jj

,
then by the implicit function theorem we have `
+
= `(c), r
+
= r(c) and j
+
= j (c) with well
dened derivatives that can be found as we did above. The point is that such functions exist and
that they are dierentiable. It follows that there is a sense in which d`
+
,dc is meaningful.
Now consider the value of the Lagrangian
/
+
= /(`
+
, r
+
, j
+
) = ) (r
+
, j
+
) +`
+
[c q (r
+
, j
+
)] ,
where we remember that (r
+
, j
+
, `
+
) is a critical point. Take the derivative w.r.t. c:
d/
+
dc
= )
a
dr
+
dc
+)
j
dj
+
dc
+
d`
+
dc
[c q (r
+
, j
+
)] +`
+
_
1 q
a
dr
+
dc
q
j
dj
+
dc
_
=
dr
+
dc
[)
a
`
+
q
a
] +
dj
+
dc
[)
j
`
+
q
j
] +
d`
+
dc
[c q (r
+
, j
+
)] +`
+
= `
+
.
Therefore
d/
+
dc
= `
+
=
0/
+
0c
.
This is a manifestation of the envelope theorem (see below). But we also know that at the
optimum we have
c q (r
+
, j
+
) = 0 .
So at the optimum we have
/(r
+
, j
+
, `
+
) = ) (r
+
, j
+
) ,
and therefore
d/
+
dc
=
d)
+
dc
= `
+
.
12.4 The envelope theorem
Let r
+
be a critical point of ) (r, 0). Then
d) (r
+
, 0)
d0
=
0) (r
+
, 0)
00
.
Proof: since at r
+
we have )
a
(r
+
, 0) = 0, we have
d) (r
+
, 0)
d0
=
0) (r
+
, 0)
0r
dr
d0
+
0) (r
+
, 0)
00
=
0) (r
+
, 0)
00

66
Drawing of an "envelope" of functions and optima for ) (r
+
, 0
1
), ) (r
+
, 0
2
), ...
12.5 Lagrange method: one constraint, many variables
Let ) (r) , q (r) C
1
and r R
a
. Suppose that r
+
is the solution to
Choose r to maximize ) (r) , s.t. q (r) = c .
and that r
+
is not a critical point of q (r) = c. Then there exists a number `
+
such that (r
+
, `
+
) is
a critical point of
/ = ) (r) +`[c q (r)] ,
i.e.
0/
0`
= c q (r, j) = 0
0/
0r
i
= )
i
`q
i
= 0 , i = 1, 2, ...: .
The constraint qualication is similar to above:
\q
+
= (q
1
(r
+
) , q
2
(r
+
) , ...q
a
(r
+
)) ,= 0 .
12.6 Lagrange method: many constraints, many variables
Let ) (r) , q
)
(r) C
1
, = 1, 2, ...:, and r R
a
. Suppose that r
+
is the solution to
Choose r to maximize ) (r) , s.t. q
1
(r) = c
1
, q
2
(r) = c
2
, ...q
n
(r) = c
n
.
and that r
+
satises the constraint qualications. Then there exists : numbers `
+
1
, `
+
2
, ...`
+
n
such
that (r
+
, `
+
) is a critical point of
/ = ) (r) +
n

)=1
`
)
_
c
)
q
)
(r)

,
i.e.
0/
0`
)
= c
)
q
)
(r) = 0 , , = 1, 2, ...:
0/
0r
i
= )
i
`q
i
= 0 , i = 1, 2, ...: .
The constraint qualication now requires that
ra:/
_
0q
0r
t
_
na
= : ,
67
which is as large as it can possibly be. This means that we must have : _ :, because
otherwise the maximal rank would be : < :. This constraint qualication, as all the others,
means that there exists a : : dimensional tangent hyperplane (a R
an
vector space).
Loosely speaking, it ensures that we can construct tangencies freely enough.
12.7 Constraint qualications in action
This example shows that when the constraint qualication is not met, the Lagrange method does
not work.
Choose r, j to maximize r, s.t. r
3
+j
2
= 0 .
The constraint set is given by
j
2
= r
3
= j = r
32
for r _ 0 ,
i.e.
C =
_
(r, j) : r _ 0, j = r
32
, j = r
32
_
Notice that (0, 0) is the maximum point. Let us check the constraint qualication:
\q =
_
3r
2
2j
_
\q (0, 0) = (0, 0) .
This violates the constraint qualications, since (0, 0) is a critical point of q (r, j).
68
Now check the Lagrangian
/ = r +`
_
r
3
j
2
_
/
A
= r
3
j
2
= 0
/
a
= 1 `3r
2
= 0 = ` = 1,3r
2
/
j
= `2j = 0 = either ` = 0 or j = 0 .
Suppose r = 0. Then ` = not admissible.
Suppose r ,= 0. Then ` 0 and thus j = 0. But then from the constraint set r = 0 a
contradiction.
12.8 Constraint qualications and Fritz John Theorem
Let ) (r) , q (r) C
1
, r R
a
. Suppose that r
+
is the solution to
Choose r to maximize ) (r) , s.t. q (r) = c
Then there exists two numbers `
+
0
and `
+
1
such that (`
+
1
, r
+
) is a critical point of
/ = `
0
) (r, j) +`
1
[c q (r, j)] ,
i.e.
0/
0`
= c q (r, j) = 0
0/
0r
i
= `
0
)
i
`
1
q
i
= 0 , i = 1, 2, ...:
and
`
+
0
0, 1
`
+
0
, `
+
1
,= (0, 0) .
This generalizes to multi constraint problems.
12.9 Second order conditions
We want to know whether d
2
. is negative or positive denite on the constraint set. Using the
Lagrange method we nd a critical point (r
+
, `
+
) of the problem
/ = ) (r) +`[c q (r)] .
69
But this is not a maximum of the / problem. In fact, (r
+
, `
+
) is a saddle point: perturbations of
r around r
+
will hurt the objective, while perturbations of ` around `
+
will increase the objective.
If (r
+
, `
+
) is a critical point of the / problem, then: holding `
+
constant, r
+
maximizes the value
of the problem; and holding r
+
constant, `
+
minimizes the value of the problem. This makes sense:
lowering the shadow cost of the constraint as much as possible at the point that maximizes the
value.
This complicates characterizing the second order conditions, to distinguish maxima from min-
ima. We want to know whether d
2
. is negative or positive denite on the constraint set. Consider
the two variables case
d. = )
a
dr +)
j
dj
From q (r, j) = c we have
q
a
dr +q
j
dj = 0 = dj =
q
a
q
j
dr ,
i.e. dj is not arbitrary. We can treat dj as a function of r and j when we dierentiate d. the
second time:
d
2
. = d (d.) =
0 (d.)
0r
dr +
0 (d.)
0j
dj
=
0
0r
[)
a
dr +)
j
dj] dr +
0
0j
[)
a
dr +)
j
dj] dj
=
_
)
aa
dr +)
ja
dj +)
j
0 (dj)
0r
_
dr +
_
)
aj
dr +)
jj
dj +)
j
0 (dj)
0j
_
dj
= )
aa
dr
2
+ 2)
aj
drdj +)
jj
dj
2
+)
j
d
2
j ,
where we use
d
2
j = d (dj) =
0 (dj)
0r
dr +
0 (dj)
0j
dj .
This is not a quadratic form, but we use q (r, j) = c again to transform it into one, by eliminating
d
2
j. Dierentiate
dq = q
a
dr +q
j
dj = 0 ,
using dj as a function of r and j again:
d
2
q = d (dq) =
0 (dq)
0r
dr +
0 (dq)
0j
dj
=
0
0r
[q
a
dr +q
j
dj] dr +
0
0j
[q
a
dr +q
j
dj] dj
=
_
q
aa
dr +q
ja
dj +q
j
0 (dj)
0r
_
dr +
_
q
aj
dr +q
jj
dj +q
j
0 (dj)
0j
_
dj
= q
aa
dr
2
+ 2q
aj
drdj +q
jj
dj
2
+q
j
d
2
j = 0
70
Thus
d
2
j =
1
q
j
_
q
aa
dr
2
+ 2q
aj
drdj +q
jj
dj
2

.
Use this in the expression for d
2
. to get
d
2
. =
_
)
aa
)
j
q
aa
q
j
_
dr
2
+ 2
_
)
aj
)
j
q
aj
q
j
_
drdj +
_
)
jj
)
j
q
jj
q
j
_
dj
2
.
From the FONCs we have
` =
)
j
q
j
and by dierentiating the FONCs we get
/
aa
= )
aa
`q
aa
/
jj
= )
jj
`q
jj
/
aj
= )
aj
`q
aj
.
We use all this to get
d
2
. = /
aa
dr
2
+ 2/
aj
drdj +/
jj
dj
2
.
This is a quadratic form, but not a standard one, because, dr and dj are not arbitrary. As
before, we want to know the sign of d
2
., but unlike the unconstrained case, dr and dj must satisfy
dq = q
a
dr +q
j
dj = 0. Thus, we have second order necessary conditions (SONC):
If d
2
. is negative semidenite s.t. dq = 0, then maximum.
If d
2
. is positive semidenite s.t. dq = 0, then minimum.
The second order sucient conditions are (SOSC):
If d
2
. is negative denite s.t. dq = 0, then maximum.
If d
2
. is positive denite s.t. dq = 0, then minimum.
These are less stringent conditions relative to unconstrained optimization, where we required con-
ditions on d
2
. for all values of dr and dj. Here we consider only a subset of those values, so the
requirement is less stringent, although slightly harder to characterize.
12.10 Bordered Hessian and constrained optimization
Using the notations we used before for a Hessian,
H =
_
a /
/ /
_
71
(except that here it will be the Hessian of /, not of )) we can write
d
2
. = /
aa
dr
2
+ 2/
aj
drdj +/
jj
dj
2
as
d
2
. = adr
2
+ 2/drdj +/dj
2
.
We also rewrite
q
a
dr +q
j
dj = 0
as
cdr +,dj = 0 .
The second order conditions involve the sign of
d
2
. = adr
2
+ 2/drdj +/dj
2
s.t. 0 = cdr +,dj .
Eliminate dj using
dj =
c
,
dr
to get
d
2
. =
_
a,
2
2/c, +/c
2

dr
2
,
2
.
The sign of d
2
. depends on the square brackets. For a maximum we need it to be negative. It
turns out that
_
a,
2
2/c, +/c
2

0 c ,
c a /
, / /

.
Notice that H contains the Hessian, and is bordered by the gradient of the constraint. Thus, the
term "bordered Hessian".
The n-dimensional case with one constraint
Let ) (r) , q (r) C
2
, r R
a
. Suppose that r
+
is a critical point of the Lagrangian problem. Let
H
aa
=
_
0
2
/
0r0r
t
_
be the Hessian of / evaluated at r
+
. Let \q be a linear constraint on d
a1
(= dr
a1
), evaluated
at r
+
:
\qd = 0 .
72
We want to know what is the sign of
d
2
. = = d
t
Hd
such that
\qd = 0 .
The sign deniteness of the quadratic form depends on the following bordered Hessian
H
(a+1)(a+1)
=
_
0 \q
1a
\q
t
a1
H
aa
_
.
Recall that sign deniteness of a matrix depends on the signs of the determinants of the leading
principal minors. Therefore
d
2
. is
_
positive denite (min)
negative denite (max)
_
s.t. dq = 0 i
_

H
3

H
4

, ...

H
a

< 0

H
3

0,

H
4

< 0,

H
5

0, ...
_
,
Note that in the text they start from

H
2

, which they dene as the third leading principal

minor and is an abuse of notation. We have one consistent way to dene leading principal
minors of a matrix and we should stick to that.
The general case
Let ) (r) , q
)
(r) C
2
, = 1, 2, ...:, and r R
a
. Suppose that r
+
is a critical point of the
Lagrangian problem. Let
H
aa
=
_
0
2
/
0r0r
t
_
be the Hessian of / evaluated at r
+
. Let

na
=
_
0q
0r
t
_
be the set of linear constraints on d
a1
(= dr
a1
), evaluated at r
+
:
d = 0 .
We want to know the sign of
d
2
. = = d
t
Hd
such that
d = 0 .
73
The sign deniteness of the quadratic form depends on the bordered Hessian
H
(n+a)(n+a)
=
_
0
nn

na

t
an
H
aa
_
.
The sign deniteness of H depends on the signs of the determinants of the leading principal minors.
For a maximum (d
2
. negative denite) we require that

H
2n

H
2n+1

...

H
n+a

alternate
signs, where sign
_

H
2n

_
= (1)
n
(Dixit). Note that we require : < :, so that 2: < :+:.
An alternative formulation for a maximum (d
2
. negative denite) requires that the last ::
leading principal minors alternate signs, where sign
_

H
a+n

_
= (1)
a
(Simon and Blume).
Anyway, the formulation in the text is wrong.
For a minimum...? We know that searching for a minimum of ) is like searching for a maximum
of ). So one could set up the problem that way and just treat it like a maximization problem.
12.11 Quasiconcavity and quasiconvexity
This is a less restrictive condition on the objective function.
Denition: a function ) is quasiconcave i \r
1
, r
2
domain of ), which is assumed to be
a convex set, and \0 (0, 1) we have
)
_
r
2
_
_ )
_
r
1
_
= )
_
0r
1
+ (1 0) r
2

_ )
_
r
1
_
.
For strict quasiconcavity replace the second inequality with a strict inequality, but not the
rst. More simply put
)
_
0r
1
+ (1 0) r
2

_ min
_
)
_
r
2
_
, )
_
r
1
__
.
In words: the image of the convex combination is larger than the lower of the two images.
Denition: a function ) is quasiconvex i \r
1
, r
2
domain of ), which is assumed to be
a convex set, and \0 (0, 1) we have
)
_
r
2
_
_ )
_
r
1
_
= )
_
0r
1
+ (1 0) r
2

_ )
_
r
2
_
.
For strict quasiconvexity replace the second inequality with a strict inequality, but not the
rst. More simply put
)
_
0r
1
+ (1 0) r
2

_ max
_
)
_
r
2
_
, )
_
r
1
__
.
74
In words: the image of the convex combination is smaller than the higher of the two images.
Strict quasiconcavity and strict quasiconvexity rule out at segments.
0 , 0, 1.
Quasiconcave Function, not Strictly Strictly Quasiconvex Function
Due to the at segment, the function on the left is not strictly quasiconcave. Note that neither
of these functions is convex nor concave. Thus, this is a weaker restriction. The following function,
while not convex nor concave, is both quasiconcave and quasiconvex.
Quasiconcave and Quasiconvex
75
Compare denition of quasiconcavity to concavity and then compare graphically.
Properties:
1. If ) is linear, then it is both quasiconcave and quasiconvex.
2. If ) is (strictly) quasiconcave, then ) is (strictly) quasiconvex.
3. If ) is concave (convex), then it is quasiconcave (quasiconvex)but not only if.
Note that unlike concave functions, the sum of two quasiconcave functions is NOT necessarily
quasiconcave. Similarly for quasiconvex functions.
Alternative "set" denitions: Let r R
a
.
) is quasiconcave i \/ R the set
o = r : ) (r) _ / , / R
is a convex set (for concavity it is "only if", not "i").
) is quasiconvex i \/ R the set
o = r : ) (r) _ / , / R
is a convex set (for convexity it is "only if", not "i").
These may be easier to verify. Recall that for concavity and convexity the conditions above were
necessary, but not sucient. Here, these are "set" denitions, so they are necessary and sucient
conditions (i).
Consider a continuously dierentiable function ) (r) C
1
and r R
a
. Then ) is
quasiconcave i \r
1
, r
2
domain of ), which is assumed to be a convex set, we have
)
_
r
2
_
_ )
_
r
1
_
= \)
_
r
1
_ _
r
2
r
1
_
_ 0 .
In words: the function does not change the sign of the slope more than once.
quasiconvex i \r
1
, r
2
domain of ), which is assumed to be a convex set, we have
)
_
r
2
_
_ )
_
r
1
_
= \)
_
r
2
_ _
r
2
r
1
_
_ 0 .
In words: the function does not change the sign of the slope more than once.
76
For strictness, change the second inequality to a strict one, which rules out at regions.
Consider a twice continuously dierentiable function ) (r) C
2
and r R
a
. The Hessian
of ) is denoted H and the gradient as \). Dene
1 =
_
0
11
\)
1a
\)
t
a1
H
aa
_
(a+1)(a+1)
.
Conditions for quasiconcavity and quasiconvexity in the positive orthant, r R
a
+
involve the
leading principal minors of 1.
Necessary condition: ) is quasiconcave on R
a
+
if (but not only if) \r R
a
+
, the leading
principal minors of 1 follow this pattern
[1
2
[ _ 0, [1
3
[ _ 0, [1
4
[ _ 0, ...
Sucient condition: ) is quasiconcave on R
a
+
only if \r R
a
+
, the leading principal minors
of 1 follow this pattern
[1
2
[ < 0, [1
3
[ 0, [1
4
[ < 0, ...
Finally, there are also explicitly quasiconcave functions.
Denition: a function ) is explicitly quasiconcave if \r
1
, r
2
domain of ), which is
assumed to be a convex set, and \0 (0, 1) we have
)
_
r
2
_
)
_
r
1
_
= )
_
0r
1
+ (1 0) r
2

)
_
r
1
_
.
This rules out at segments, except at the top of the hill.
Ranking of concavity, from strongest to weakest:
1. strictly concave
2. concave
3. strictly quasiconcave
)
_
r
2
_
_ )
_
r
1
_
= )
_
0r
1
+ (1 0) r
2

)
_
r
1
_
.
(no at regions)
77
4. explicitly quasiconcave
)
_
r
2
_
)
_
r
1
_
= )
_
0r
1
+ (1 0) r
2

)
_
r
1
_
.
(only one at region allowed, at the top)
5. quasiconcave
)
_
r
2
_
_ )
_
r
1
_
= )
_
0r
1
+ (1 0) r
2

_ )
_
r
1
_
.
12.12 Why is quasiconcavity important? Invariance to positive monotone trans-
formation
Quasiconcavity is important because it allows arbitrary cardinality in the utility function, while
maintaining ordinality. Concavity imposes decreasing marginal utility, which is not necessary for
characterization of convex preferences, convex indierence sets and convex upper contour sets.
Only when dealing with risk do we need to impose concavity. We do not need concavity for global
extrema.
Quasiconcave functions preserve quasi concavity under any positive monotone transformation.
Suppose that some utility function n is quasiconcave, i.e. the set o = r : n(r) _ / is a convex
set \/ R. This means that any r o is at least as good as any r , o. Now consider a positive
monotone transformation of n(r), denoted t (n(r)). Then the set T = r : t (n(r)) _ t (/) is still
convex \/ R. Moreover, T = o, i.e. the same r o that are at least as good as any r , o are
the same r T that are at least as good as any r , T. A corollary is that if I nd a maximizer of
n(r), it is also a maximizer of t (n(r)).
Concave functions DO NOT preserve concavity under any positive monotone transformation.
Proof: By example. r
2
is concave (make sure that you know how to prove this). But c
a
2
a pos-
itive monotone transformation of r
2
is not concave (it has the shape of the normal distribution
function)
Note that since concave functions are also quasiconcave, then any positive monotone transfor-
mation will keep it at least quasiconcave, like in the example above: c
a
2
is indeed quasiconcave.
12.13 Why is quasiconcavity important? Global maximum
Suppose that r
+
is the solution to
Choose r to maximize ) (r) , s.t. q (r) = c .
If
1. the set r : q (r) = c is a convex set (e.g., a hyperplane or a simplex); and
78
2. ) is explicitly quasiconcave,
then ) (r
+
) is a global (constrained) maximum. If in addition ) is strictly quasiconcave, then this
global maximum is unique.
This doesnt apply for a quasi-concave function (not explicitly, not strictly) because then we
can have several at regions, not only at the top. This will not allow distinguishing local
maxima from global maximum based only on the properties of ).
12.14 Application: cost minimization
We like apples (a) and bananas (/), but want to reduce the cost of any (a, /) bundle for a given
level of utility (l(
+
a,
+
/)) (or fruit salad, if we want to use a production metaphor).
Choose a, / to minimize cost C = aj
o
+/j
b
, s.t. l (a, /) = n
Set up the appropriate Lagrangian
/ = aj
o
+/j
b
+`[n l (a, /)] .
Here ` is in units of $/util: it tells you how much an additional util will cost. If l was a production
function for salad, then ` would be in units of $ per unit of salad, i.e. the price of one unit of salad.
FONC:
0/
0`
= n l (a, /) = 0
0/
0a
= j
o
`l
o
= 0 = j
o
,l
o
= ` 0
0/
0/
= j
b
`l
b
= 0 = j
b
,l
b
= ` 0 .
Thus
'1o =
l
o
l
b
=
j
o
j
b
So we have tangency. Let the value of the problem be
C
+
= a
+
j
o
+/
+
j
b
.
Take a total dierential at the optimum to get
dC = j
o
da +j
b
d/ = 0 =
d/
da
=
j
o
j
b
< 0 .
We could also obtain this result from the implicit function theorem, since C (a, /) , l (a, /) C
1
79
and [J[ , = 0. Yet another way to get this is to see that since l (a, /) = n, a constant,
dl (a, /) = l
o
da +l
b
d/ = 0 =
d/
da
=
l
o
l
b
< 0 .
At this stage all we know is that the isoquant for utility slopes downward, and that it is tangent
to the isocost line at the optimum, if the optimum exists.
SOC:
_
H

=
_
_
0 l
o
l
b
l
o
`l
oo
`l
ob
l
b
`l
ob
`l
bb
_
_
.
We need positive deniteness of d
2
C
+
for a minimumwhich requires negative deniteness of Hso
we need

H
2

< 0 and

H
3

< 0.

H
2

0 l
o
l
o
`l
oo

= l
2
o
< 0 ,
so this is good (in fact,

H
2

is always _ 0, just not necessarily < 0). But

H
3

= 0 l
o
[l
o
(`l
bb
) (`l
ob
) l
b
] +l
b
[l
o
(`l
ob
) (`l
oo
) l
b
]
= l
2
o
`l
bb
l
o
`l
ob
l
b
l
b
l
o
`l
ob
+l
2
b
`l
oo
= `
_
l
2
o
l
bb
2l
o
l
ob
l
b
+l
2
b
l
oo
_
Without further conditions on l, we do not know whether the expression in the parentheses is
negative or not (` 0).
The curvature of the utility isoquant is given by
d
da
_
d/
da
_
=
d
2
/
da
2
=
d
da
_

l
o
l
b
_
=
d
da
_
l
o
(a, /)
l
b
(a, /)
_
=
=
_
l
oo
+l
ob
ob
oo
_
l
b
l
o
_
l
bb
ob
oo
+l
ob
_
l
2
b
=
_
l
oo
+l
ob
_

la
l
b
__
l
b
l
o
_
l
bb
_

la
l
b
_
+l
ob
_
l
2
b
=
l
oo
l
b
l
o
l
ob
+l
2
o
l
bb
,l
b
l
o
l
ob
l
2
b
=
l
oo
l
2
b
2l
o
l
ob
l
b
+l
2
o
l
bb
l
3
b
=
1
l
3
b
_
l
oo
l
2
b
2l
o
l
ob
l
b
+l
2
o
l
bb
_
.
This involves the same expression in the parentheses. If the indierence curve is convex, then
80
o
2
b
oo
2
_ 0 and thus the expression in the parentheses must be negative. This coincides with the
positive semi-deniteness of d
2
C
+
. Thus, convex isoquants and existence of a global minimum in
this case come together. This would ensure a global minimum, although not a unique one. If
o
2
b
oo
2
0, then the isoquant is strictly convex and the global minimum is unique, as dC
+
is positive
denite.
If l is strictly quasiconcave, then indeed the isoquant is strictly convex and the global
minimum is unique.
12.15 Related topics
12.15.1 Expansion paths
Consider the problem described above in Section 12.14. Let a
+
(j
o
, j
b
, n) and /
+
(j
o
, j
b
, n) be the
optimal quantities of apples and bananas chosen given prices and a level of promised utility (AKA
"demand"). The expansion path is the function / (a) that is given by changes in n, when prices are
xed.
One way to get this is to notice that the FONCs imply
l
o
(a
+
, /
+
)
l
b
(a
+
, /
+
)
=
j
o
j
b
l (a
+
, /
+
) = n ,
which dene a system of equations, which can be written as
1 (a
+
, /
+
, n, j
o
, j
b
) = 0 .
Fix prices. By applying the implicit function theorem we get a
+
(n) and /
+
(n). Each level of
n denes a unique level of demand. The expansion path / (a) is the function of all the unique
combinations of a
+
(n) and /
+
(n) at all levels of n.
12.15.2 Homotheticity
This is when the expansion path / (a) is a ray (a straight line starting at the origin). Equivalently,
homotheticity implies thatand is implied bythe ratio
b

(&)
o

(&)
=
_
b

_
(n) is constant, not aected
by n.
12.15.3 Elasticity of substitution
An elasticity j is dened as the percent change in j that is invoked by a percent change in r:
j
j,a
=
dj,j
dr,r
=
dj
dr
r
j
=
d lnj
d lnr
.
81
an elasticity of substitution is usually referred to as an elasticity that does not change the level of
some function. For example, if
1 (j, j) = c ,
then the elasticity of substitution is the percent change in j that is invoked by a percent change in
j. By the implicit function theorem
dj
dj
=
1
j
1
j
and
j
j,j
=
1
j
1
j
j
j
.
Elasticities of substitution often arise in the context of optimization. For example, consider the
problem described above in Section 12.14. Let a
+
(j
o
, j
b
, n) and /
+
(j
o
, j
b
, n) be the optimal quan-
tities of apples and bananas chosen given prices and a level of promised utility (AKA "demand").
The elasticity of substitution in demand (between a and /) is given by
d (a
+
,/
+
) , (a
+
,/
+
)
d (j
o
,j
b
) , (j
o
,j
b
)
.
This number tells you how much the relative intensity of consumption of a (relative to /) changes
with the relative price of a (relative to /).
12.15.4 Constant elasticity of substitution and relation to Cobb-Douglas
A general production function that exhibits constant elasticity of substitution (CES) is
= . [c/
,
+ (1 c) |
,
]
1,
. (5)
is the quantity of output and / and | are inputs. c [0, 1] is called the distribution parameter. .
is a level shifter ("productivity" in the output context). The function is CES because a 1 percent
change in the marginal products implies o percent change in the input ratio:
d (/,|) , (/,|)
d ('1
I
,'1
|
) , ('1
I
,'1
|
)
= o =
1
1 ,
. (6)
To see this, note that
'1
I
=
0
0/
=
1
,
. [c/
,
+ (1 c) |
,
]
1,1
,c/
,1
'1
|
=
0
0|
=
1
,
. [c/
,
+ (1 c) |
,
]
1,1
,(1 c) |
,1
82
so that
'1
I
'1
|
=
c
1 c
_
/
|
_
,1
=
c
1 c
_
/
|
_
1o
.
Taking the total dierential we get
d
_
'1
I
'1
|
_
=
1
o
c
1 c
_
/
|
_
1o1
d
_
/
|
_
.
Dividing through by '1
I
,'1
|
and rearranging, we get (6).
This general form can be applied as a utility function as well, where represents a level of
utility and where / and | represent quantities of dierent goods in consumption.
Now, it follows that when , = 0 we have o = 1. But you cannot simply plug , = 0 into (5)
because
= lim
,0
. [c/
,
+ (1 c) |
,
]
1,
= "0
o
" .
In order to nd out the expression for when , = 0 rewrite (5) as
ln(,.) =
ln[c/
,
+ (1 c) |
,
]
,
.
Now take the limit
lim
,0
ln(,.) = lim
,0
ln[c/
,
+ (1 c) |
,
]
,
=
0
0
.
Now apply LHopitals rule:
lim
,0
ln[c/
,
+ (1 c) |
,
]
,
= lim
,0
c/
,
ln/ + (1 c) |
,
ln|
1 [c/
,
+ (1 c) |
1,
]
= cln/ + (1 c) ln| .
So that
lim
,0
ln(,.) = cln/ + (1 c) ln|
or
= ./
c
|
1c
, (7)
which is the familiar Cobb-Douglas production function. It follows that (7) is a particular case of
(5) with o = 1.
Note: to get this result we had to have the distribution parameter c. Without it, you would
not get this result.
83
13 Optimization with inequality constraints
13.1 One inequality constraint
Let ) (r) , q (r) C
1
, r R
a
. The problem is
Choose r to maximize ) (r) , s.t. q (r) _ c .
Write the constraint in a "standard way"
q (r) c _ 0 .
Suppose that r
+
is the solution to
Choose r to maximize ) (r) , s.t. q (r) c _ 0
and that if the q (r) c _ 0 constraint bindsi.e., q (r) = cthen r
+
is not a critical point of
q (r). Write down the Lagrangian function
/ = ) (r) +`[c q (r)] .
Then there exists a number `
+
such that
(1) :
0/
0r
i
= )
i
`
+
q
i
= 0 , i = 1, 2, ...:
(2) : `
+
[c q (r, j)] = 0
(3) : `
+
_ 0
(4) : q (r) _ c .
The standard way: write q (r)c _ 0 and then ip it in the Lagrangian function `[c q (r)].
Conditions 2 and 3 are called complementary slackness conditions. If the constraint is
not binding, then changing c a bit will not aect the value of the problem; in that case ` = 0.
The constraint qualications are that if the constraint binds, i.e. q (r
+
) = c, then
\q (r
+
) ,= 0.
84
Conditions 1-4 in CW are written dierently, although they are an equivalent representation:
(i) :
0/
0r
i
= )
i
`q
i
= 0 , i = 1, 2, ...:
(ii) :
0/
0`
= [c q (r, j)] _ 0
(iii) : ` _ 0
(iv) : `[c q (r, j)] = 0 .
Notice that from (ii) we get q (r) _ c. If q (r) < c, then /
A
0.
13.2 One inequality constraint and one non-negativity constraint
There is really nothing special about this problem, but it is worthwhile setting it up, for practice.
Let ) (r) , q (r) C
1
, r R
a
. The problem is
Choose r to maximize ) (r) , s.t. q (r) _ c and r _ 0 .
Rewrite this as
Choose r to maximize ) (r) , s.t. q (r) c _ 0 and r 0 _ 0 .
Suppose that r
+
is the solution to this problem and that r
+
is not a critical point of the constraint
set (to be dened below). Write down the Lagrangian function
/ = ) (r) +`[c q (r)] +,[r] .
Then there exist two numbers `
+
and ,
+
such that
(1) :
0/
0r
i
= )
i
`q
i
+, = 0 , i = 1, 2, ...:
(2) : `[c q (r, j)] = 0
(3) : ` _ 0
(4) : q (r) _ c
(5) : ,[r] = 0
(6) : , _ 0
(7) : r _ 0 =r _ 0 .
The constraint qualication is that r
+
is not a critical point of the constraints that bind.
If only q (r) = c binds, then we require \q (r
+
) ,= 0. See the general case below.
85
The text gives again a dierentand I argue less intuitiveformulation. The Lagrangian is set
up without explicitly mentioning the non-negativity constraints
? = ) (r) +,[c q (r)] .
In the text the FONCs are written as
(i) :
0?
0r
i
= )
i
,q
i
_ 0
(ii) : r
i
_ 0
(iii) : r
i
0?
0r
i
= 0 , i = 1, 2, ...:
(iv) :
0?
0,
= [c q (r)] _ 0
(v) : , _ 0
(vi) : ,
0?
0,
= 0 .
The unequal treatment of dierent constraints is confusing. My method treats all constraints
consistently. A non-negativity constraint is just like any other.
13.3 The general case
Let ) (r) , q
)
(r) C
1
, r R
a
, , = 1, 2, ...:. The problem is
Choose r to maximize ) (r) , s.t. q
)
(r) _ c
)
, , = 1, 2, ...: .
Write the the problem in the standard way
Choose r to maximize ) (r) , s.t. q
)
(r) c
)
_ 0 , , = 1, 2, ...: .
Write down the Lagrangian function
/ = ) (r) +
n

)=1
`
)
_
c
)
q
)
(r)

.
86
Suppose that r
+
is the solution to the problem above and that r
+
does not violate the constraint
qualications (see below). Then there exists : numbers `
+
)
, , = 1, 2, ...: such that
(1) :
0/
0r
i
= )
i

)=1
`
)
q
)
i
(r) = 0 , i = 1, 2, ...:
(2) : `
)
_
c
)
q
)
(r)

= 0
(3) : `
)
_ 0
(4) : q
)
(r) _ c
)
, , = 1, 2, ...: .
The constraint qualications are as follows. Consider all the binding constraints. Count
them by ,
b
= 1, 2, ...:
b
. Then we must have that the rank of
_
0q
1
(r
+
)
0r
t
_
=
_

_
0j
1
(a

)
0a
0
0j
2
(a

)
0a
0
.
.
.
0j
m
b (a

)
0a
0
_

_
n
b
a
is :
b
, as large as possible.
13.4 Minimization
It is worthwhile to consider minimization separately, although minimization of ) is just like maxi-
mization of ). We compare to maximization.
Let ) (r) , q (r) C
1
, r R
a
. The problem is
Choose r to maximize ) (r) , s.t. q (r) _ c .
Rewrite as
Choose r to maximize ) (r) , s.t. q (r) c _ 0
Write down the Lagrangian function
/ = ) (r) +`[c q (r)] .
FONCs
(1) :
0/
0r
i
= )
i
`q
i
= 0 , i = 1, 2, ...:
(2) : `[c q (r, j)] = 0
(3) : ` _ 0
(4) : q (r) _ c .
87
Compare this to
Choose r to minimize /(r) , s.t. q (r) _ c .
Rewrite as
Choose r to minimize /(r) , s.t. q (r) c 0
Write down the Lagrangian function
/ = /(r) +`[c q (r)] .
FONCs
(1) :
0/
0r
i
= /
i
`q
i
= 0 , i = 1, 2, ...:
(2) : `[c q (r, j)] = 0
(3) : ` _ 0
(4) : q (r) _ c .
Everything is the same. Just pay attention to the inequality setup correctly. This will be equivalent.
Consider the problem
Choose r to maximize /(r) , s.t. q (r) _ c .
Rewrite as
Choose r to maximize /(r) , s.t. c q (r) _ 0 .
and set up the proper Lagrangian function for maximization
/ = /(r) +`[q (r) c] .
This will give the same FONCs as above.
13.5 Example
Choose r, j to maximize minar, /j , s.t. rj
a
+jj
j
_ 1 ,
where a, /, j
a
, j
j
0. Convert this to the following problem
Choose r, j to maximize ar, s.t. ar _ /j, rj
a
+jj
j
1 _ 0 .
This is equivalent, because given a level of j, we will never choose ar /j, nor can the objective
exceed /j by construction.
Choose r, j to maximize ar, s.t. ar /j _ 0, rj
a
+jj
j
1 _ 0 .
88
Set up the Lagrangian
/ = ar +`[1 rj
a
jj
j
] +,[/j ar] .
FONC:
1 : /
a
= a `j
a
a, = 0
2 : /
j
= `j
j
+/, = 0
3 : `[1 rj
a
jj
j
] = 0
4 : ` _ 0
5 : rj
a
+jj
j
_ 1
6 : ,[/j ar] = 0
7 : , _ 0
8 : ar _ /j .
The solution process is a trial and error process. The best way is to start by checking which
constraints do not bind.
1. Suppose , = 0. Then from 2: `j
j
= 0 = ` = 0 = from 1: a a, = 0 = , = 1 0
a contradiction. Therefore ' > 0 must hold. Then from 6: ar = /j = j = ar,/.
2. Suppose ` = 0 (while , 0). Then from 2: /, = 0 = , = 0 a contradiction (even if , = 0,
we would reach another contradiction from 1: a = 0). Therefore > 0. Then rj
a
+jj
j
= 1
= rj
a
+arj
j
,/ = 1 = r(j
a
+aj
j
,/) = 1 = r
+
= 1, (j
a
+aj
j
,/), j
+
= a1, (/j
a
+aj
j
).
Solving for the multipliers (which is an integral part of the solution) involves solving 1 and 2:
`j
a
+a, = a
`j
j
/, = 0 .
This can be written in matrix notation
_
j
a
a
j
j
/
_ _
`
,
_
=
_
a
0
_
.
The solution requires nonsingular matrix:

j
a
a
j
j
/

= /j
a
aj
j
< 0 .
89
Solving by Cramers Rule:
`
+
=

a a
0 /

/j
a
aj
j
=
a/
/j
a
+aj
j
0
,
+
=

j
a
a
j
j
0

/j
a
aj
j
=
aj
j
/j
a
+aj
j
0 .
Try to build economic interpretations for the shadow costs:
`
+
=
a/
/j
a
+aj
j
=
a
j
a
+
o
b
j
j
.
`
+
tells you how many more utils you would get if income (1) increased by one unit. What do you
do with this additional unit of income? You spend it on r and joptimally: for each x you buy,
you also buy a,/ units of j. How much does this cost you? Exactly the denominator of `
+
. So you
get 1,
_
j
a
+
o
b
j
j
_
additional units of r for each unit of income (at the margin). And each one gives
you an additional a utils.
Finally, we check the constraint qualications. Since both constraints bind (`
+
0, ,
+
0),
we must have a rank of two for the matrix
0
_
rj
a
+jj
j
1
ar /j
_
0
_
r j
=
_
j
a
j
j
a /
_
.
In this case we can verify that the rank is two by the determinant, since this is a square 2 2
matrix:

j
a
j
j
a /

= /j
a
aj
j
< 0 .
It is no accident that the determinant is the same as above.
13.6 Another example
Choose r, j to maximize r
2
+r + 4j
2
, s.t. 2r + 2j _ 1 , r, j _ 0
Rewrite as
Choose r, j to maximize r
2
+r + 4j
2
, s.t. 2r + 2j 1 _ 0 , r _ 0 , j _ 0
90
Consider the Jacobian of the constraints
0
_
_
2r + 2j
r
j
_
_
0 [r j]
=
_
_
2 2
1 0
0 1
_
_
.
This has rank 2 for any submatrix \(r, j) R
2
, and since at most two constaraints can bind, the
constraint qualications are never violated. The constraint set is a triangle: all the constraints are
linear and independent, so the constraint qualication will not fail.
Set up the Lagrangian function
/ = r
2
+r + 4j
2
+`[1 2r 2j] +,[r] +, [j]
FONCs
/
a
= 2r + 1 2` +, = 0
/
j
= 8j 2` +, = 0
`[1 2r 2j] = 0 ` _ 0 2r + 2j _ 1
,r = 0 , _ 0 r _ 0
,j = 0 , _ 0 j _ 0
1. From /
a
= 0 we have
2r + 1 +, = 2` 0
with strict inequality, because r _ 0 and , _ 0. Thus > 0 and the constraint
2r + 2j = 1
binds, so that
j = 1,2 r or r = 1,2 j .
2. Suppose , 0. Then r = 0 = j = 1,2 = , = 0 = ` = 2 = , = 3. A candidate
solution is (r
+
, j
+
) = (0, 1,2).
3. Suppose , = 0. Then
2r + 1 = 2` .
From /
j
= 0 we have
8j +, = 2` .
91
Combining the two we get
2r + 1 = 8j +,
2 (1,2 j) + 1 = 8j +,
2 2j = 8j +,
10j +, = 2 .
The last result tells us that we cannot have both , = 0 and j = 0, because we would get
0 = 2 a contradiction (also because then we would get ` = 0 from /
j
= 0). So either , = 0
or j = 0 but not both.
(a) Suppose j = 0. Then r = 1,2 = ` = 1 = , = 2. A candidate solution is
(r
+
, j
+
) = (1,2, 0).
(b) Suppose j 0. Then , = 0 = j = 0.2 = r = 0.3 = ` = 0.8. A candidate solution
is (r
+
, j
+
) = (0.3, 0.2).
Eventually, we need to evaluate the objective function with each candidate to see which is the
global maximizer.
13.7 The Kuhn-Tucker suciency theorem
Let ) (r) , q
)
(r) C
1
, , = 1, 2, ...:. The problem is
Choose r R
a
to maximize ) (r) ,
s.t. r _ 0 and q
)
(r) _ c
)
, , = 1, 2, ...: .
Theorem: if
1. ) is concave on R
a
,
2. q
)
are convex on R
a
,
3. r
+
satises the FONCs of the Lagrangian
then r
+
is a global maximumnot necessarily unique.
We know: if q
)
(r) are convex then
_
r : q
)
(r) _ c
)
_
are convex sets. One can show that the
intersection of convex sets is also a convex set, so that the constraint set is also convex. So
the theorem actually says that trying to maximize a concave function on a convex set give a
global maximum, if it exists. Whether it exists on the border or not, the FONCs will detect
it.
92
Also note that if ) is concave, then the set r : ) (r) _ / is convex. In particular, the set
r : ) (r) _ ) (r
+
) is also convex: the optimum set is convex.
But these are strong conditions on our objective and constraint functions. The next theorem
relaxes things quite a bit.
13.8 The Arrow-Enthoven suciency theorem
Let ) (r) , q
)
(r) C
1
, , = 1, 2, ...:. The problem is
Choose r R
a
to maximize ) (r) ,
s.t. r _ 0 and q
)
(r) _ c
)
, , = 1, 2, ...: .
Theorem: if
1. ) is quasi concave on R
a
+
,
2. q
)
are quasi convex on R
a
+
,
3. r
+
satises the FONCs of the Kuhn-Tucker Lagrangian,
4. Any one of the following conditions on ) holds:
(a) i such that )
i
(r
+
) < 0.
(b) i such that )
i
(r
+
) 0 and r
+
i
0 (i.e. it does not violate the constraints).
(c) \) (r
+
) ,= 0 and ) C
2
around r
+
.
(d) ) (r) is concave.
then r
+
is a global maximum, not necessarily unique.
Arrow-Enthoven constraint qualication test for a maximization problem
If
1. q
)
(r) C
1
are quasi convex,
2. r
0
R
a
+
such that all constraints are slack,
3. Any one of the following holds:
(a) q
)
(r) are convex.
(b) 0q (r) ,0r
t
,= 0 \r in the constraint set.
then the constraint qualication is not violated.
93
13.9 Envelope theorem for constrained optimization
Recall the envelope theorem for unconstrained optimization: if r
+
is a critical point of ) (r, 0).
Then
d) (r
+
, 0)
d0
=
0) (r
+
, 0)
00
.
This was due to
0)(a

,0)
0a
= 0.
Now we face a more complicated problem:
Choose r R
a
to maximize ) (r, 0) , s.t. q (r, 0) = c .
For a problem with inequality constraints we simply use only those constraints that bind. We will
consider small perturbations of 0, so small that they will not aect which constraint binds. Set up
the Lagrangian function
/ = ) (r, 0) +`[c q (r, 0)] .
FONCs
/
A
= c q (r, 0) = 0
/
a
i
= )
i
(r, 0) `q
i
(r, 0) = 0 , i = 1, 2, ...:
We apply the implicit function theorem to this set of equations to get r
+
= r
+
(0) and `
+
= `
+
(0)
for which there well dened derivatives around (`
+
, r
+
). We know that at the optimum we have
that the value of the problem is the value of the Lagrangian function
) (r
+
, 0) = /
+
= ) (r
+
, 0) +`
+
[c q (r
+
, 0)]
= ) (r
+
(0) , 0) +`
+
(0) [c q (r
+
(0) , 0)] .
Dene the value of the problem as
(0) = ) (r
+
, 0) = ) (r
+
(0) , 0) .
Take the derivative with respect to 0 to get
d (0)
d0
=
d/
+
d0
= )
+
a
dr
+
d0
+)
+
0
+
d`
+
d0
[c q (r
+
(0) , 0)] `
+
_
q
+
a
dr
+
d0
+q
+
0
_
= [)
+
a
`
+
q
+
a
]
dr
+
d0
+
d`
+
d0
[c q (r
+
(0) , 0)] +)
+
0
`
+
q
+
0
= )
+
0
`
+
q
+
0
.
94
Of course, we could have just applied this directly using the envelope theorem:
d (0)
d0
=
d/
+
d0
=
0/
+
00
= )
+
0
`q
+
0
.
13.10 Duality
We will demonstrate the duality of utility maximization and cost minimization. But the principles
here are more general than consumer theory.
The primal problem is
Choose r R
a
to maximize n(r) , s.t. j
t
r = 1 .
(this should be stated with _ but we focus on preferences with nonsatiation and strict convexity
and therefore strict concavity of nso the solution lies on the budget line and r 0 is also
satised). The Lagrangian function is
/ = n(r) +`
_
1 j
t
r

FONCs:
/
a
i
= n
i
`j
i
= 0 = ` = n
i
,j
i
, i = 1, ...:
/
A
=
_
1 j
t
r

= 0
Recall: ` tells you how many utils we get for one additional unit of income.
We apply the implicit function theorem to this set of equations to get Marshallian demand
r
n
i
= r
n
i
(j, 1)
and
`
n
i
= `
n
i
(j, 1)
for which there are well dened derivatives around (`
+
, r
+
). We dene the indirect utility func-
tion
(j, 1) = n[r
n
(j, 1)] .
The dual problem is
Choose r R
a
to minimize j
t
r s.t. n(r) = n ,
where n is a level of promised utility (this should be stated with n(r) _ n but we assume that n
is strictly increasing in r, and since the objective is also strictly increasing in r, the solution must
95
lie at n(r) = n). The Lagrangian function is
7 = j
t
r +,[n n(r)] .
FONCs:
7
a
i
= j
i
,n
i
= 0 = , = j
i
,n
i
, i = 1, ...:
7
,
= n (r) = 0
Recall: , tells you how much an additional util will cost.
We apply the implicit function theorem to this set of equations to get Hicksian demand
r
I
i
= r
I
i
(j, n)
and
,
I
i
= ,
I
i
(j, n)
for which there are well dened derivatives around (,
+
, r
+
). We dene the expenditure function
c (j, n) = j
t
r
I
(j, n) .
Duality: all FONCs imply the same thing:
n
i
n
)
=
j
i
j
)
,
Thus, at the optimum
r
n
i
(j, 1) = r
I
i
(j, n)
c (j, n) = 1
(j, 1) = n .
Moreover,
, =
1
`
and this makes sense given the interpretation of , and `.
Duality relies on unique global extrema. We need to have all the preconditions for that.
Make drawing.
96
13.11 Roys identity
(j, 1) = n(r
n
) +`
n
_
1 j
t
r
n
_
.
Taking the derivative with respect to a price,
0
0j
i
=
a

)=1
n
)
0r
n
)
0j
i
+
0`
0j
i
_
1 j
t
r
n
_
`
_
_
a

)=1
j
)
0r
n
)
0j
i
+r
n
i
_
_
=
a

)=1
(n
)
`j
)
)
0r
n
)
0j
i
+
0`
0j
i
_
1 j
t
r
n
_
`r
n
i
= `r
n
i
.
An increase in j
i
will lower demand by r
n
i
, which decreases the value of the problem, as if by
decreasing income by r
n
i
times ` utils/$ per dollar of lost income. In other words, income is now
worth r
n
i
less, and this taxes the objective by `r
n
i
. Taking the derivative with respect to income,
0
01
=
a

)=1
n
)
0r
n
)
01
+
0`
01
_
1 j
t
r
n
_
+`
_
_
1
a

)=1
j
)
0r
n
)
01
_
_
=
a

)=1
(n
)
`j
)
)
0r
n
)
01
+
0`
01
_
1 j
t
r
n
_
+`
= ` .
An increase in income will increase our utility by `, which is the standard result.
In fact, we could get these results applying the envelope theorem directly:
0
0j
i
= `r
n
i
0
01
= ` .
Roys identity is thus

0,0j
i
0,01
= r
n
i
.
Why is this interesting? Because this is the amount of income needed to compensate consumers
for (that will leave them indierent to) an increase in the price of some good r
i
. To see this, rst
consider
(j, 1) = n ,
where n is a level of promised utility (as in the dual problem). By the implicit function theorem
1 = 1 (j
i
) in a neighbourhood of r
n
, which has a well dened derivative d1,dj. This function
97
is dened at the optimal bundle r
n
. Now consider the total dierential of , evaluated at the
optimal bundle r
n
:

j
i
dj
i
+
1
d1 = 0 .
This dierential does not involve other partial derivatives because it is evaluated at the the optimal
bundle r
n
(i.e. the envelope theorem once again). And we set this dierential to zero, because
we are considering keeping the consumer exactly indierent, i.e. her promised utility and optimal
bundle remain unchanged. Then we have
d1
dj
i
=

j
i

1
=
0,0j
i
0,01
= r
n
i
.
This result tells you that if you are to keep the consumer indierent to a small change in the price
of good i, i.e. not changing the optimally chosen bundle, then you must compensate the consumer
by r
n
i
units of income. We will see this again in the dual problem, using Shephards lemma, where
keeping utility xed is explicit. We will see that
0c
0j
i
= r
I
i
= r
n
i
is exactly the change in expenditure
that results from keeping utility xed, while increasing the price of good i.
To see this graphically, consider a level curve of utility. The slope of the curve at (j, 1) (more
generally, the gradient) is r
n
.
Roys Identity
13.12 Shephards lemma
c (j, n) = j
t
r
I
+,
I
_
n n
_
r
I
__
.
98
Taking the derivative with respect to a price,
0c
0j
i
= r
I
i
+
a

)=1
j
)
0r
j
)
0j
i
+
0,
0j
i
_
n n
_
r
I
__
,
a

)=1
n
)
0r
I
)
0j
i
=
a

)=1
(j
)
,n
)
)
0r
I
)
0j
i
+
0,
0j
i
_
n n
_
r
I
__
+r
I
i
= r
I
i
.
An increase in j
i
will increases cost by r
I
i
while keeping utility xed at n (remember that this is a
minimization problem so increasing the value of the problem is "bad"). Note that this is exactly
the result of Roys Identity. Taking the derivative with respect to promised utility,
0c
0n
=
a

)=1
j
)
0r
I
)
0n
+
0,
0n
_
n n
_
r
I
__
+,
_
_
1
a

)=1
n
)
0r
I
)
0n
_
_
=
a

)=1
(j
)
,n
)
)
0r
I
)
0n
+
0,
0n
_
n n
_
r
I
__
+,
= , .
An increase in utility will increase expenditures by ,, which is the standard result.
In fact, we could get these results applying the envelope theorem directly:
0c
0j
i
= r
I
i
0c
0n
= , .
This is used often with cost functions in the context of production. Let c be the lowest cost to
produce n units of output (with n(r) serving as the production function that takes the vector of
inputs r and where j are their prices). Then taking the derivative of the cost function c w.r.t. j
gives you demand for inputs. And taking the derivative of the cost function c w.r.t. the quantity
produced (n) gives you the cost (price) of producing one additional unit of output.
13.13 Mundlak (1968) REStud example
The following is based on ?.
Let j = ) (r) C
2
, r R
a
be the strictly concave production function of a competitive rm.
99
The rm solves the cost minimization problem
choose r to minimize c = j
t
r ,
s.t. ) (r) _ j .
Setting y as numeraire normalizes j
j
= 1, so that the factor prices are interpreted as real prices,
interms of units of output. The Lagrangian can be written as (with appropriate interpretation of
the multiplier `):
/ = j
t
r +
1
`
[j ) (r)]
If (r
+
, `
+
) is a critical point then it satises the FONCs
)
i
(r
+
) = `
+
j
i
, i = 1, 2, ...:
) (r
+
) = j .
The values of r and ` are optimal, so that small perturbations of prices will have no indirect eect
on the objective through them (the envelope theorem). To ease notation I will not carry on the
asterisks in what follows.
We treat the FONCs as identities and dierentiate around the optimal choice

)
)
i)
dr
)
= d`j
i
+`dj
i
=
d`
`
`j
i
+`dj
i
=

`)
i
+`dj
i
for i = 1, 2, ...: ,
where

` = d`,` = d ln`. Rearranging gives the endogenous perturbations of the optimal values
induced by perturbations of the exogenous prices

)
)
i)
dr
)

`)
i
= `dj
i
for i = 1, 2, ...:
and in matrix notation
Hdr \)

` = `[dj] .
I am writing the gradient as a column vector, rather than the conventional row vector of partials.
This can be written as
_
\) H

_

`
dr
_
= `[dj] .
The other dierential is
dj =

i
)
i
dr
i
= \)
t
dr .
100
Dene
1
d. =
1
`
dj ,
which is how tightening the constraint aects the objective (value of the problem). Stacking this
on top of the previous matrices gives
_
0 \)
t
\) H
_ _

`
dr
_
= `
_
d.
dj
_
.
The rst matrix on the LHS is just the bordered Hessian of the problem. Given regularity conditions
(and conditions for a minimum), the bordered Hessian is invertible so we can write
_

`
dr
_
= `
_
0 \)
t
\) H
_
1
_
d.
dj
_
.
Mundlak denes
1 =
_
1
00
1
0)
1
i0
1
i)
_
= `
_
0 \)
t
\) H
_
1
= `
_

_
\)
t
H
1
\)
_
1
_
\)
t
H
1
\)
_
1
)
t
i
)
1
i)
H
1
\)
_
\)
t
H
1
\)
_
1
H
1
H
1
\)
_
\)
t
H
1
\)
_
1
\)
t
H
1
_
,
(see Section 5.7 for inverting partitioned matrices) so that
_

`
dr
_
=
_
1
00
1
0)
1
i0
1
i)
_ _
d.
dj
_
. (8)
Furthermore,
\)
t
1
i0
= `
\)
t
1
i)
= 0
1
i)
\) = 0 .
I.e. 1
i)
is a singular matrix. Finally, it is useful to write (8) without its rst row:
dr = [1
i)
] dj + [1
i0
] d.
_

_
dr
1
dr
2
.
.
.
dr
a
_

_
=
_

_
/
11
/
12
/
1a
/
21
/
22
/
2a
.
.
.
.
.
.
.
.
.
/
a1
/
aa
_

_
_

_
dj
1
dj
2
.
.
.
dj
a
_

_
+
_

_
/
10
/
20
.
.
.
/
a0
_

_
d.
1
In other optimization problems, e.g. constant cost and constant marginal cost, dz will take a dierent form.
101
Along the output isoquant (dj = 0 =d. = 0) we have
0r
i
0j
)
= /
i)
. (9)
Since the objective is c = r
t
j, the envelope theorem gives (on the output isoquant) (Shephards
lemma):
0c
0j
)
= r
)
. (10)
Writing this as an elasticity gives
0 lnc
0 lnj
)
=
0c
0j
)
j
)
c
=
j
)
r
)
c
=
)
, (11)
which is the cost share.
Remark 1 The cost elasticity w.r.t. the price of a particular input is the cost share of that input.
A corollary is that the sum of the elasticities of cost w.r.t. all prices is unity.
Now writing (9) as a price elasticity, we have
j
i)
=
0 lnr
i
0 lnj
)
=
0r
i
0j
)
j
)
r
i
= /
i)
j
)
r
i
= /
i)
c
r
i
r
)
j
)
r
)
c
= o
i)

)
, (12)
where /
i)
is the i-, element of [1
i)
], o
i)
is the (Allen partial) elasticity of substitution and
)
is
dened in (11).
Remark 2 The elasticity of input i w.r.t. the price of input , is equal to the (Allen partial)
elasticity of substitution of input i w.r.t. input , times the cost share of input ,.
Note that for a particular input i

)
j
i)
=

)
/
i)
j
)
r
i
=
1
`r
i

)
`j
)
/
i)
=
1
`r
i

)
)
)
/
i)
= 0 , (13)
since the summation is just one column of \)
t
1
i)
= 0.
Remark 3 The sum of price elasticities (12) for any particular input i is zero.
From (13) we have

)
j
i)
=

)
o
i)

)
= o
ii

i
+

),=i
o
i)

)
= 0
and therefore
o
ii
=
1

),=i
o
i)

)
< 0 .
102
This result holds for any factor i that is actually used in production (
i
0, i.e. no corner
solution for i) under the following conditions: there is at least one factor , ,= i that exhibits some
substitutability with i, and which is actually used in production, i.e. , ,= i s.t. o
i)
0 and
)
0
(
)
_ 0 and o
i)
_ 0 for all , for concave CRS production functions). While the object o
ii
does not
have a clear interpretation (at least not to me), it helps sign the own-price elasticity:
j
ii
= o
ii

i
=

),=i
o
i)

)
< 0 .
Remark 4 The own-price elasticity is negative.
For example, ? shows (page 342) that in the case of two inputs
o
12
=
)
1
)
2
(r
1
)
1
+r
2
)
2
)
r
1
r
2
_
)
11
)
2
2
2)
12
)
1
)
2
+)
22
)
2
1
_ =
)
1
)
2

_
)
11
)
2
2
2)
12
)
1
)
2
+)
22
)
2
1
_
(r
1
)
1
+r
2
)
2
)
r
1
r
2
.
Substituting the FONCs, we have
o
12
=
)
1
)
2

_
)
11
)
2
2
2)
12
)
1
)
2
+)
22
)
2
1
_
(r
1
`j
1
+r
2
`j
2
)
r
1
r
2
= `
)
1
)
2

_
)
11
)
2
2
2)
12
)
1
)
2
+)
22
)
2
1
_
. .
I
12
(r
1
j
1
+r
2
j
2
)
r
1
r
2
= /
12
c
r
1
r
2
.
So that
j
12
= o
12

2
= /
12
c
r
1
r
2
r
2
j
2
c
= /
12
j
2
r
1
as above.
103
14 Integration
14.1 Preliminaries
Consider a continuous dierentiable function
r = r(t)
and its derivative with respect to time
dr
dt
= _ r .
This is how much r changes during a very short period dt. Suppose that you know _ r at any point
in time. We can write down how much r changed from some initial point in time, say t = 0, until
period t as follows:
_
t
0
_ rdt .
This is the sum of all changes in r from period 0 to t. The term of art is integration, i.e. we are
integrating all the increments. But you cannot say what r(t) is, unless you have the value of r at
the initial point. This is the same as saying that you know what the growth rate of GDP is, but
you do not know the level. But given r
0
= r(0) we can tell what r(t) is:
r(t) = r
0
+
_
t
0
_ rdt .
E.g.
_ r = t
2
_
t
0
_ rdt =
_
t
0
n
2
dn =
1
3
t
3
+c .
The constant c is arbitrary and captures the fact that we do not know the level.
Suppose that the instant growth rate of j is a constant r, i.e.
_ j
j
= r .
This can be written as
_ j rj = 0 ,
which is an ordinary dierential equation. We know that j = c
vt
gives _ j,j = r. But so does
j = /c
vt
. So once again, without having additional information about the value of j at some initial
point, we cannot say what j (t) is.
104
14.2 Indenite integrals
Denote
) (r) =
d1 (r)
dr
.
Therefore,
d1 (r) = ) (r) dr .
Summing over all small increments we get
_
d1 (r) =
_
) (r) dr = 1 (r) +c ,
where the constant of integration, c, denotes that the integral is correct up to an indeterminate
constant. This is so because knowing the sum of increments does not tell you the level. Another
way to see this is
d
dr
1 (r) =
d
dr
[1 (r) +c] .
Integration is the opposite operation of dierentiation. Instead of looking at small perturbations,
or increments, we look for the sum of all increments.
Commonly used integrals
1.
_
r
a
dr =
a
n+1
a+1
+c
2.
_
)
t
(r) c
)(a)
dr = c
)(a)
+c ,
_
c
a
dr = c
a
+c ,
_
)
t
(r) /
)(a)
dr =
b
f(x)
ln b
+c
3.
_
)
0
(a)
)(a)
dr = ln[) (r)] +c ,
_
1
a
dr =
_
oa
a
= lnr +c
Operation rules
1. Sum:
_
[) (r) +q (r)] dr =
_
) (r) dr +
_
q (r) dr
2. Scalar multiplication: /
_
) (r) dr =
_
/) (r) dr
3. Substitution/change of variables: Let n = n(r). Then
_
) (n) n
t
dr =
_
) (n)
dn
dr
dr =
_
) (n) dn = 1 (n) +c .
E.g.
_
2r
_
r
2
+ 1
_
dr = 2
_
_
r
3
+r
_
dr = 2
_
r
3
dr + 2
_
rdr =
1
2
r
4
+r
2
+c
105
Alternatively, dene n = r
2
+ 1, hence n
t
= 2r, and so
_
2r
_
r
2
+ 1
_
dr =
_
dn
dr
ndr =
_
ndn =
1
2
n
2
+c
t
=
1
2
_
r
2
+ 1
_
2
+c
t
=
1
2
_
r
4
+ 2r
2
+ 1
_
+c
t
=
1
2
r
4
+r
2
+
1
2
+c
t
=
1
2
r
4
+r
2
+c .
4. Integration by parts: Since
d (n) = nd +dn
we have
_
d (n) = n =
_
nd +
_
dn .
Thus the integration by part formula is
_
nd = n
_
dn .
To reduce confusion denote
\ = \ (r) , (r) = d\ (r) ,dr
l = l (r) , n(r) = dl (r) ,dr
Then we write the formula as
_
l (r) d\ (r) = l (r) \ (r)
_
\ (r) dl (r)
_
l (r) (r) dr = l (r) \ (r)
_
\ (r) n(r) dr .
E.g., let ) (r) = ,c
,a
. Then
_
r,c
,a
dr = rc
,a

_
c
,a
dr
In the notation above, we have
_
r
..
l
,c
,a
. .

dr = r
..
l
c
,a
. .
\

_
1
..
&

_
c
,a
_
. .
\
dr
106
14.3 Denite integrals
The area under the ) curve for a continuous ) on [a, /], i.e. between the ) curve and the horizontal
axis, from a to / is
_
b
o
) (r) dr = 1 (/) 1 (a) .
This is also called the fundamental theorem of calculus. Note that this area may be positive
or negative, depending on whether the area lies more above the horizontal axis or below it.
The Riemann Integral: create : rectangles that lie under the curve, that take the minimum of
the heights: r
i
, i = 1, 2...:. Then create : rectangles with height the maximum of the heights: 1
i
,
i = 1, 2...:. As the number of these rectangles increases, the sums of the rectangles may converge.
If they do, then we say that ) is Reimann-integrable. I.e. if
lim
ao
a

i=1
r
i
= lim
ao
a

i=1
1
i
then
_
b
o
) (r) dr
exists and is well dened.
Properties of dinite intergrals:
1. Minus/switching the integration limits:
_
b
o
) (r) dr =
_
o
b
) (r) dr = 1 (/)1 (a) = [1 (a) 1 (/)]
2. Zero:
_
o
o
) (r) dr = 1 (a) 1 (a) = 0
3. Partition: for all a < / < c
_
c
o
) (r) dr =
_
b
o
) (r) dr +
_
c
b
) (r) dr .
4. Scalar multiplication:
_
b
o
/) (r) dr = /
_
b
o
) (r) dr , \/ R
5. Sum:
_
b
o
[) (r) +q (r)] dr =
_
b
o
) (r) dr +
_
b
o
q (r) dr
6. By parts:
_
b
o
ldr = l\ [
b
o

_
b
o
n\ dr = l (/) \ (/) l (a) \ (/)
_
b
o
n\ dr
7. Substitution/change of variables: Let n = n(r). Then
_
b
o
) (n) n
t
dr =
_
b
o
) (n)
dn
dr
dr =
_
&(b)
&(o)
) (n) dn = 1 (n) +c .
107
Suppose that we wish to integrate a function from some initial point r
0
until some indenite
point r. Then
_
a
a
0
) (t) dt = 1 (r) 1 (r
0
) .
and so
1 (r) = 1 (r
0
) +
_
a
a
0
) (t) dr .
14.4 Leibnitzs Rule
Let ) C
1
(i.e. 1 C
2
). Then
0
00
b(0)
_
o(0)
) (r, 0) dr = ) (/ (0) , 0)
0/ (0)
00
) (a (0) , 0)
0a (0)
00
+
b(0)
_
o(0)
0
00
) (r, 0) dr .
Proof: let ) (r, 0) = d1 (r, 0) ,dr. Then
0
00
b(0)
_
o(0)
) (r, 0) dr =
0
00
[1 (r, 0)[
b(0)
o(0)
=
0
00
[1 (/ (0) , 0) 1 (a (0) , 0)]
= 1
a
(/ (0) , 0)
0/ (0)
00
+1
0
(/ (0) , 0) 1
a
(a (0) , 0)
0a (0)
00
1
0
(a (0) , 0)
= ) (/ (0) , 0)
0/ (0)
00
) (a (0) , 0)
0a (0)
00
+ [1
0
(/ (0) , 0) 1
0
(a (0) , 0)]
= ) (/ (0) , 0)
0/ (0)
00
) (a (0) , 0)
0a (0)
00
+
b(0)
_
o(0)
d
dr
1
0
(r, 0) dr
= ) (/ (0) , 0)
0/ (0)
00
) (a (0) , 0)
0a (0)
00
+
b(0)
_
o(0)
0
00
) (r, 0) dr .
The last line follows from Youngs Theorem: for a continuously dierentiable 1,
0
2
1 (r, j)
0r0j
=
0
0r
01 (r, j)
0j
=
0
0j
01 (r, j)
0r
=
0
2
1 (r, j)
0j0r
.
If the integration limits do not depend on 0, then
0
00
b
_
o
) (r, 0) dr =
b
_
o
0
00
) (r, 0) dr ,
108
and if ) does not depend on 0, then
0
00
b(0)
_
o(0)
) (r) dr = ) (/ (0))
0/ (0)
00
) (a (0))
0a (0)
00
.
14.5 Improper integrals
14.5.1 Innite integration limits
_
o
o
) (r) dr = lim
bo
_
b
o
) (r) dr = lim
bo
1 (/) 1 (a) .
E.g., A ~exp(,): 1 (r) = 1 c
,a
, ) (r) = ,c
,a
for r _ 0.
_
o
0
,c
,a
dr = lim
bo
_
b
0
,c
,a
dr = lim
bo
c
,b
+c
,0
= 1 .
Also
1 (r) =
_
o
0
r) (r) r =
_
o
0
r,c
,a
dr =
_
rc
,a

o
0

_
o
0
c
,a
dr
= " c
,o
" + 0c
,0
+
_

1
,
c
,a

o
0
= 0
1
,
c
,o
+
1
,
c
,0
=
1
,
.
E.g.
_
o
1
1
r
dr = lim
bo
_
b
1
1
r
dr = [ln(r)[
o
1
= ln() ln(1) = 0 = .
14.5.2 Innite integrand
E.g., sometimes the integral is divergent, even though the integration limits are nite:
_
1
0
1
r
dr = lim
b0
_
1
b
1
r
dr = [ln(r)[
1
0
= ln(1) ln(0) = 0 += .
Suppose that for some j (a, /)
lim
aj
) (r) = .
Then the integral from a to / is convergent i the partitions are also convergent:
_
b
o
) (r) dr =
_
j
o
) (r) dr +
_
b
j
) (r) dr .
E.g.
lim
a0
1
r
3
= .
109
Therefore, the integral
_
1
1
1
r
3
dr =
_
0
1
1
r
3
dr +
_
1
0
1
r
3
dr =
_

1
2r
2

0
1
+
_

1
2r
2

1
0
does not exist, because neither integral converges.
14.6 Example: investment and capital formation
In discrete time we have the capital accumulation equation
1
t+1
= (1 c) 1
t
+1
t
,
where 1
t
is gross investment at time t. Rewrite as
1
t+1
1
t
= 1
t
c1
t
.
We want to rewrite this in continuous time. In this context, investment, 1
t
, is instantaneous and
capital depreciates at an instantaneous rate of c. Consider a period of length . The accumulation
equation is
1
t+
1
t
= 1
t
c1
t
.
Divide by to get
1
t+
1
t

= 1
t
c1
t
.
Now take 0 to get
_
1
t
= 1
t
c1
t
,
where it is understood that 1
t
is instantaneous investment at time t, and 1
t
is the amount of capital
available at that time. c1
t
is the amount of capital that vanishes due to depreciation. Write
_
1
t
= 1
a
t
,
where 1
a
t
is net investment. Given a functional form for 1
a
t
we can tell how much capital is around
at time t, given an initial amount at time 0, 1
0
.
Let 1
a
t
= t
o
. then
1
t
1
0
=
_
t
0
_
1dt =
_
t
0
1
a
&
dn =
_
t
0
n
o
dn =
_
n
o+1
a + 1

t
0
=
t
o+1
a + 1
.
14.7 Domars growth model
Domar was interested in answering: what must investment be in order to satisfy the equilibrium
condition at all times.
110
Structure of the model:
1. Fixed saving rate: 1
t
= :1
t
, : (0, 1). Therefore
_
1 = :
_
1 . And so
_
1 =
1
:
_
1 ,
i.e. there is a multiplier eect of investment on output.
2. Potential output is given by a CRS production function

t
= j1
t
,
therefore
_ = j
_
1 = j1 .
3. Long run equilibrium is given when potential output is equal to actual output
= 1 ,
therefore
_ =
_
1 .
We have three equations:
(i) output demand :
_
1 =
1
:
_
1
(ii) potential output : _ = j1
(iii) equilibrium : _ =
_
1 .
Use (iii) in (ii) to get
j1 =
_
1
and then use (i) to get
j1 =
1
:
_
1 ,
which gives
_
1
1
= j: .
Now integrate in order to nd the level of investment at any given time:
_
_
1
1
dt =
_
j:dt
ln1 = j:t +c
1
t
= c
(jc)t+c
= c
(jc)t
c
c
= 1
0
c
(jc)t
.
111
The larger is productivity, j, and the higher the saving rate, :, the more investment is required.
This is the amount of investment needed for output to keep output in check with potential output.
Now suppose that output is not equal to its potential, i.e. ,= 1 . This could happen if the
investment is not growing at the correct rate of j:. Suppose that investment is growing at rate a,
i.e.
1
t
= 1
0
c
ot
.
Dene the utilization rate
n = lim
to
1
t

t
.
Compute what the capital stock is at any moment:
1
t
1
0
=
_
t
0
_
1dt +
_
t
0
1
t
dt =
_
t
0
1
0
c
ot
dt =
1
a
1
0
c
ot
(the constant of integration is absorbed in 1
0
.) Now compute
n = lim
to
1
t

t
= lim
to
1
c
1
t
j1
t
=
1
j:
lim
to
1
t
1
t
=
1
j:
lim
to
1
0
c
ot
1
o
1
0
c
ot
+1
0
=
a
j:
lim
to
1
0
c
ot
1
0
c
ot
+a1
0
=
a
j:
.
The last equality can be derived using LHopitals rule, or by simply noting that
1
0
c
at
1
0
c
at
+o1
0
=
1
1+o1
0
1
0
c
at
1 as t . If a j: then n 1 there is a shortage of capacity, excess demand.
If a < j: then n < 1 there is a excess of capacity, excess supply. Thus, in order to keep output
demand equal to output potential we must have a = j: and thus n = 1.
In fact, this holds at any point in time:
_
1 =
d
dt
1
0
c
ot
= a1
0
c
ot
.
Therefore
_
1 =
1
:
_
1 =
a
:
1
0
c
ot
_ = j1 = j1
0
c
ot
.
So
_
1
_
=
o
c
1
0
c
ot
j1
0
c
ot
=
a
:j
= n .
If the utilization rate is too high n 1, then demand growth outstrips supply,
_
1 _ . If the
utilization rate is too low n < 1, then demand growth lags behind supply,
_
1 < _ .
Thus, the razor edge condition: only a = :j keeps us at a sustainable equilibrium path:
If n 1, i.e. a :j, there is excess demand, investment is too high. Entrepreneurs will try
to invest even more to increase supply, but this implies an even larger gap between the two.
112
If n < 1, i.e. a < :j, there is excess supply, investment is too low. Entrepreneurs will try to
cut investment to lower demand, but this implies an even larger gap between the two.
This model is clearly unrealistic and highly stylized.
113
15 First order dierential equations
We deal with equations that involve _ j. The general form is
_ j +n(t) j (t) = n(t)
The goal is to characterize j (t) in terms of n(t) and n(t). Note that this can be written as
_ j = ) (j, t) .
First order means
oj
ot
, not
o
2
j
ot
2
.
No products: _ j j is not permitted.
In principle, we can have d
a
j,dt
a
, where : is the order of the dierential equation. In the next
chapter we will deal with up to d
2
j,dt
2
.
15.1 Fundamental theorem of dierential equations
Consider solving for j (t) in
_ j = ) (j, t) , (14)
where j (t
0
) = j
0
is known.
Suppose that ) is a continuous function at (t
0
, j
0
). Then there exists a C
1
function j : 1 R
on the open interval 1 = (t
0
a, t
0
+a) such that j (t
0
) = j
0
and _ j (t) = ) (j (t) , t) for all t 1,
i.e. j (t) solves (14). If in addition ) C
1
at (t
0
, j
0
), then the solution j (t) is unique; any two
solutions of (14) must be equal to each other on the intersection of their domains.
Most dierential equations in economic applications will have ) C
1
, so all solutions will be
unique. But in the following case
_ j = 3j
23
there are multiple solutions because 3j
23
is not dierentiable at 0 (the derivative is ). Since 3j
23
is nonetheless continuous at 0, a solution exists but it is not unique. For example, both j (t) = 0
and j (t) = t
3
solve the dierential equation.
15.2 Constant coecients
15.2.1 Homogenous case
_ j +aj = 0
114
This gives rise to
_ j
j
= a
which has solution
j (t) = j
0
c
ot
.
We need an additional condition to pin down j
0
.
15.2.2 Non homogenous case
_ j +aj = / ,
where / ,= 0. The solution method involves splitting the solution into two:
j (t) = j
c
(t) +j
j
(t) ,
where j
j
(t) is a particular solution and j
c
(t) is a complementary function.
j
c
(t) solves the homogenous equation
_ j +aj = 0 ,
so that
j
c
(t) = c
ot
.
j
j
(t) solves the original equation for a stationary solution, i.e. _ j = 0, which implies that j is
constant and thus j = /,a, where a ,= 0. The solution is thus
j = j
c
+j
j
= c
ot
+
/
a
.
Given an initial condition j (0) = j
0
, we have
j
0
= c
o0
+
/
a
= +
/
a
= = j
0

/
a
.
The general solution is
j (t) =
_
j
0

/
a
_
c
ot
+
/
a
= j
0
c
ot
+
/
a
_
1 c
ot
_
.
One way to think of the solution is a linear combination of two points: the initial condition j
0
and
the particular, stationary solution /,a. (If a 0, then for t _ 0 we have 0 _ c
ot
_ 1, which yields
115
a convex combination). Verify this solution:
_ j = a
_
j
0

/
a
_
c
ot
= a
_

_
_
j
0

/
a
_
c
ot
+
/
a
. .
j

/
a
_

_
= aj +/
= _ j +aj = / .
Yet a dierent way to look at the solution is
j (t) =
_
j
0

/
a
_
c
ot
+
/
a
= /c
ot
+
/
a
,
for some arbitrary point /. In this case
_ j = a/c
ot
,
and we have
_ j +aj = a/c
ot
+a
_
/c
ot
+
/
a
_
= / .
When a = 0, we get
_ j = /
so
j = j
0
+/t .
This follows directly from
_
_ jdt =
_
/dt
j = /t +c ,
where c = j
0
. We can also solve this using the same technique as above. j
c
solves _ j = 0, so
that this is a constant j
c
= . j
j
should solve 0 = /, but this does not work unless / = 0. So
try a dierent particular solution, j
j
= /t, which requires / = /, because then _ j
j
= / = /.
So the general solution is
j = j
c
+j
j
= +/t .
Together with a value for j
0
, we get = j
0
.
116
E.g.
_ j + 2j = 6 .
j
c
solves _ j + 2j = 0, so
j
c
= c
2t
.
j
j
solves 2j = 6 ( _ j = 0), so
j
j
= 3 .
Thus
j = j
c
+j
j
= c
2t
+ 3 .
Together with j
0
= 10 we get 10 = c
20
+ 3, so that = 7. This completes the solution:
j = 7c
2t
+ 3 .
Verifying this solution:
_ j = 14c
2t
and
_ j + 2j = 14c
2t
+ 2
_
7c
2t
+ 3
_
= 6 .
15.3 Variable coecients
The general form is
_ j +n(t) j (t) = n(t) .
15.3.1 Homogenous case
n(t) = 0:
_ j +n(t) j (t) = 0 =
_ j
j
= n(t) .
Integrate both sides to get
_
_ j
j
dt =
_
n(t) dt
lnj +c =
_
n(t) dt
j = c
c
R
&(t)ot
= c

R
&(t)ot
,
where = c
c
. Thus, the general solution is
j = c

R
&(t)ot
.
117
Together with a value for j
0
and a functional form for n(t) we can solve explicitly.
E.g.
_ j + 3t
2
j = 0
_ j +
_
3t
2
_
j = 0 .
Thus
_ j
j
= 3t
2
_
_ j
j
dt =
_
3t
2
dt
lnj +c =
_
3t
2
dt
j = c
c
R
3t
2
ot
= c
t
3
.
15.3.2 Non homogenous case
n(t) ,= 0:
_ j +n(t) j (t) = n(t) .
The solution is
j = c

R
&(t)ot
_
+
_
n(t) c
R
&(t)ot
dt
_
.
Obtaining this solution requires some footwork. But rst, see that it works: e.g.,
_ j +t
2
j = t
2
= n(t) = t
2
, n(t) = t
2
.
_
n(t) dt =
_
t
2
dt =
1
3
t
3
_
n(t) c
R
&(t)ot
dt =
_
t
2
c
t
3
3
dt = c
t
3
3
,
since
_
)
t
(j) c
)(j)
dj = c
)(j)
.
Thus
j = c
t
3
3
_
+c
t
3
3
_
= c
t
3
3
+ 1 .
Verifying this solution:
_ j = t
2
c
t
3
3
118
so
_ j +n(t) j (t) = t
2
c
t
3
3
+
_
t
2
_
_
c
t
3
3
+ 1
_
= t
2
c
t
3
3
+t
2
c
t
3
3
+t
2
= t
2
= n(t) .
15.4 Solving exact dierential equations
Suppose that the primitive dierential equation can be written as
1 (j, t) = c
so that
d1 = 1
j
dj +1
t
dt = 0 .
We use the latter total dierential to obtain 1 (j, t), from which we obtain j (t). We set 1 (j, t) = c
to get initial conditions.
Denition: the dierential equation
'dj +dt = 0
is an exact dierential equation i 1 (j, t) such that ' = 1
j
and = 1
t
.
If such a function 1 (j, t) exists, then by Youngs theorem we have
0'
0t
=
0
2
1
0t0j
=
0
0j
.
And this relationship is what we will be checking in practice to verify that a dierential equations
is indeed exact.
E.g., let 1 (j, t) = j
2
t = c. Then
d1 = 1
j
dj +1
t
dt = 2jtdj +j
2
dt = 0 .
Set
' = 2jt , = j
2
.
119
Check:
0
2
1
0t0j
=
0'
0t
= 2j
0
2
1
0t0j
=
0
0j
= 2j
So this is an exact dierential equation.
Solving exact dierential equations:
Before solving, one must always check that the equation is indeed exact.
Step 1: Since
d1 = 1
j
dj +1
t
dt
we could integrate both sides, but this does not lead to right answer. Instead, integrate only
1
j
over j and add a residual function of t alone:
1 (j, t) =
_
1
j
dj +,(t)
=
_
'dj +,(t) ,
where ,(t) is a residual function.
Step 2: Take the derivative of 1 (j, t) from step 1 w.r.t. t, , and equate it to 1
t
from the
original dierential function. This identies ,(t).
Step 3: Solve for j (t), taking into account 1 (j, t) = c.
Example:
2jt
..
A
dj + j
2
..
.
dt = 0 .
Step 1:
1 (j, t) =
_
'dj +,(t) =
_
2jtdj +,(t) = j
2
t +,(t) .
Step 2:
01 (j, t)
0t
=
0
0t
_
j
2
t +,(t)

= j
2
+,
t
(t) .
Since = j
2
we must have ,
t
(t) = 0, i.e. ,(t) is a constant function, ,(t) = /, for some /. Thus
1 (j, t) = j
2
t +/ = c ,
120
so we can ignore the constant / and write
1 (j, t) = j
2
t = c .
Step 3: We can now solve for j (t):
j (t) = c
12
t
12
.
The solution for c will be given by an initial condition.
Example:
(t + 2j) dj +
_
j + 3t
2
_
dt = 0 .
So that
' = (t + 2j)
=
_
j + 3t
2
_
.
Check that this equation is exact:
0'
0t
= 1 =
0
0j
,
so this is indeed an exact dierential equation.
Step 1:
1 (j, t) =
_
'dj +,(t) =
_
(t + 2j) dj +,(t) = tj +j
2
+,(t) .
Step 2:
01 (j, t)
0t
=
0
0t
_
tj +j
2
+,(t)

= j +,
t
(t) = = j + 3t
2
,
so that
,
t
(t) = 3t
2
and
,(t) =
_
,
t
(t) dt =
_
3t
2
dt = t
3
.
Thus
1 (j, t) = tj +j
2
+,(t) = tj +j
2
+t
3
.
Step 3: we cannot solve this analytically for j (t), but using the implicit function theorem, we
can characterize it.
Example:
121
Let T ~ 1 (t) be the time until some event occurs, T _ 0. Dene the hazard rate as
/(t) =
) (t)
1 1 (t)
,
which is the "probability" that the event occurs at time t, given that it has not occurred by time t.
We can write
/(t) =
1
0
(t)
1(t)
,
where 1(t) = 1 1 (t). We know how to solve such dierential equations:
1
0
(t) +/(t) 1(t) = 0 .
1(t) = c

R
t
I(c)oc
.
Since 1(0) = 1 (the probability that the event occurs at all), then we have = 1:
1(t) = c

R
t
I(c)oc
.
It follows that
) (t) = 1
t
(t) = c

R
t
I(c)oc
0
0t
_

_
t
/(:) d:
_
= c

R
t
I(c)oc
[/(t)] = /(t) c

R
t
I(c)oc
.
Suppose that the hazard rate is constant:
/(t) = c .
In that case
) (t) = cc

R
t
coc
= cc
ct
,
which is the p.d.f. of the exponential distribution.
Now suppose that the hazard rate is not constant, but
/(t) = c,t
o1
.
In that case
) (t) = c,t
o1
c

R
t
coc
1
oc
= c,t
o1
c
ct

,
which is the p.d.f. of the Weibull distribution. This is useful if you want to model an increasing
hazard (, 1) or decreasing hazard (, < 1). When , = 1 or we get the exponential distribution.
122
15.5 Integrating factor and the general solution
Sometimes we can turn a non exact dierential equation into an exact one. For example,
2tdj +jdt = 0
is not exact:
' = 2t
= j
and
'
t
= 2 ,=
j
= 1 .
But if we multiply the equation by j, we get an exact equation:
2jtdj +j
2
dt = 0 ,
which we saw above is exact.
15.5.1 Integrating factor
We have the general formulation
_ j +nj = n ,
where all variables are functions of t and we wish to solve for j (t). Write the equation above as
dj
dt
+nj = n
dj +njdt = ndt
dj + (nj n) dt = 0 .
The integrating factor is
c
R
t
&(c)oc
.
If we multiply the equation by this factor we always get an exact equation:
c
R
t
&(c)oc
dj +c
R
t
&(c)oc
(nj n) dt = 0 .
To verify this, write
' = c
R
t
&(c)oc
= c
R
t
&(c)oc
(nj n)
123
and
0'
0t
=
0
0t
c
R
t
&(c)oc
= c
R
t
&(c)oc
n(t)
0
0j
=
0
0j
c
R
t
&(c)oc
(nj n) = c
R
t
&(c)oc
n(t) .
So 0',0t = 0,0j.
This form can be recovered from the method of undetermined coecients. We seek some
such that

..
A
dj +(nj n)
. .
.
dt = 0
and
0'
0t
=
0
0t
=
_

0
0j
=
0
0j
[(nj n)] = n
are equal. This means
_
= n
_
, = n
= c
R
t
&(c)oc
.
15.5.2 The general solution
We have some equation that is written as
_ j +nj = n .
Rewrite as
dj + (nj n) dt = 0 .
Multiply by the integrating factor to get an exact equation
c
R
t
&(c)oc
. .
A
dj +c
R
t
&(c)oc
(nj n)
. .
.
dt = 0 .
Step 1:
1 (j, t) =
_
'dj +,(t) =
_
c
R
t
&(c)oc
dj +,(t) = jc
R
t
&(c)oc
+,(t) .
124
Step 2:
01
0t
=
0
0t
_
jc
R
t
&(c)oc
+,(t)
_
= jc
R
t
&(c)oc
n(t) +,
t
(t) = .
Using from above we get
= jc
R
t
&(c)oc
n(t) +,
t
(t) = c
R
t
&(c)oc
(nj n) ,
so that
,
t
(t) = c
R
t
&(c)oc
n
and so
,(t) =
_
c
R
t
&(c)oc
ndt .
Now we can write
1 (j, t) = jc
R
t
&(c)oc

_
c
R
t
&(c)oc
ndt = c
Step 3, solve for j:
j = c

R
t
&(c)oc
_
c +
_
c
R
t
&(c)oc
ndt
_
.
15.6 First order nonlinear dierential equations of the 1st degree
In general,
_ j = /(j, t)
will yield an equation like this
) (j, t) dj +q (j, t) dt = 0 .
In principle, j and t can appear in any degree.
First order means _ j, not j
(a)
.
First degree means _ j, not ( _ j)
a
.
15.6.1 Exact dierential equations
See above.
15.6.2 Separable variables
) (j) dj +q (t) dt = 0 .
Then just integrate
_
) (j) dj =
_
q (t) dt
125
and solve for j (t).
Example:
3j
2
dj tdt = 0
_
3j
2
dj =
_
tdt
j
3
=
1
2
t
2
+c
j (t) =
_
1
2
t
2
+c
_
13
.
Example:
2tdj jdt = 0
dj
j
=
dt
2t
_
dj
j
=
_
dt
2t
lnj =
1
2
lnt +c
j = c
1
2
ln t+c
= c
ln t
1=2
c
c
= c
c
t
12
.
15.6.3 Reducible equations
Suppose that
_ j = /(j, t)
can be written as
_ j +1j = Tj
n
, (15)
where
1 = 1(t)
T = T (t)
are functions only of t and
: ,= 0, 1 .
When : = 0 we get _ j +1j = T, which we know how to solve.
When : = 1 we get _ j +1j = Tj, and then we solve _ j,j = (T 1).
126
Equation (15) is a Bernoulli equation, which can be reduced to a linear equation by
changing variables and solved as such. Once the solution is found, we can back out the original
function j (t). Heres how:
_ j +1j = Tj
n
1
j
n
_ j +1j
1n
= T
Use a change of variables
. = j
1n
so that
_ . = (1 :) j
n
_ j
_ j
j
n
=
_ .
1 :
.
Plug this in the equation to get
_ .
1 :
+1. = T
d. +
_
_
(1 :) 1
. .
&
. (1 :) T
. .
&
_
_
dt = 0
d. + [n. +n] dt = 0 .
This is something we know how to solve:
. (t) = c

R
t
&(c)oc
_
+
_
c
R
t
&(c)oc
ndt
_
.
from which we get the original
j (t) = . (t)
1
1m
.
Example:
_ j +tj = 3tj
2
In this case
1 = t
T = 3t
: = 2 .
127
Divide by j
2
and rearrange to get
j
2
_ j +tj
1
3t = 0 .
Change variables
. = j
1
_ . = j
2
_ j
so that we get
_ . +t. 3t = 0
d. + (t. + 3t) dt = 0 .
so that we set
n = t
n = 3t .
Using the formula we get
. (t) = c

R
t
&(c)oc
_
+
_
c
R
t
&(c)oc
ndt
_
= c
R
t
coc
_
3
_
c

R
t
coc
tdt
_
= c
t
2
2
_
3
_
c
t
2
2
tdt
_
= c
t
2
2
_
+ 3c
t
2
2
_
= c
t
2
2
+ 3 .
So that
j (t) =
1
.
=
_
c
t
2
2
+ 3
_
1
.
Example:
_ j +j,t = j
3
.
128
In this case
1 = 1,t
T = 1
: = 3 .
Divide by j
3
and rearrange to get
j
3
_ j +t
1
j
2
1 = 0 .
Change variables
. = j
2
_ . = 2j
3
_ j
so that we get

_ .
2
+
.
t
1 = 0
_ . +2
.
t
+ 2 = 0
d. +
_
2
.
t
+ 2
_
dt = 0 .
so that we set
n = 2,t
n = 2 .
Using the formula we get
. (t) = c

R
t
&(c)oc
_
+
_
c
R
t
&(c)oc
ndt
_
= c
2
R
t
t
1
oc
_
2
_
c
2
R
t
t
1
oc
dt
_
= c
2 ln t
_
2
_
c
2 ln t
dt
_
= t
2
_
2t
2

= t
2
2 .
So that
j (t) =
1
.
2
=
_
t
2
2
_
2
.
129
15.7 The qualitative graphic approach
Given
_ j = ) (j)
we can plot _ j as a function of j. This is called a phase diagram. This is an autonomous dierential
equation, since t does not appear explicitly as an argument. We have three cases:
1. _ j 0 : j is growing, so we shift to the right.
2. _ j < 0 : j is decreasing, so we shift to the left.
3. _ j = 0 : j is stationary, an equilibrium.
Stable Dierential Equation Unstable Dierential Equation
System A is dynamically stable: the _ j curve is downward sloping; any movement away from
the stationary point j
+
will bring us back there.
System B is dynamically unstable: the _ j curve is upward sloping; any movement away from
the stationary point j
+
take farther away.
For example, consider
_ j +aj = /
with solution
j (t) =
_
j
0

/
a
_
c
ot
+
/
a
=
_
c
ot
_
j
0
+
_
1 c
ot
_
/
a
.
130
This is a linear combination between the initial point j
0
and /,a.
System A happens when a 0: lim
to
c
ot
0, so that lim
to
j (t) /,a = j
+
.
System B happens when a < 0: lim
to
c
ot
, so that lim
to
j (t) .
15.8 The Solow growth model (no long run growth version)
1. CRS production function
1 = 1 (1, 1)
j = ) (/)
where j = 1,1, / = 1,1. Given 1
1
0 and 1
11
< 0 we have )
t
0 and )
tt
< 0.
2. Constant saving rate: 1 = :1 , so that
_
1 = :1 c1.
3. Constant labor force growth:
_
1,1 = :.
_
1 = :1 (1, 1) c1 = :1) (/) c1
_
1
1
= :) (/) c/ .
Since
_
/ =
d
dt
_
1
1
_
=
_
11 1
_
1
1
2
=
_
1
1

1
1
_
1
1
=
_
1
1
/: ,
we get
_
/ = :) (/) (: +c) / .
This is an autonomous dierential equation in /.
Since )
t
0 and )
tt
< 0 we know that / such that :) (/) < (: +c) /. And given the Inada
conditions ()
t
(0) = and ) (0) = 0), then / such that :) (/) (: +c) /. Therefore,
_
/ 0 for
low levels of /; and
_
/ < 0 for high levels of /. Given the continuity of ) we know that /
+
such
that
_
/ = 0, i.e. the system is stable.
131
Solow Model
132
16 Higher order dierential equations
We will discuss here only second order, since it is very rare to nd higher order dierential equations
in economics. The methods introduced here can be extended to higher order dierential equations.
In fact the fundamental theorem of dierential equations, described above in Section 15.1 ex-
tends to ,
tI
order dierential equations, provided that we specify , initial conditions:
j (t
0
) = j
0
, j
t
(t
0
) = j
1
, j
tt
(t
0
) = j
2
, ... j
[)1]
(t
0
) = j
)1
.
16.1 Second order, constant coecients
j
tt
+a
1
j
t
+a
2
j = / ,
where
j = j (t)
j
t
= dj,dt
j
tt
= d
2
j,dt
2
,
and a
1
, a
2
, and / are constants. The solution will take the form
j = j
j
+j
c
,
where the particular solution, j
j
, characterizes a stable point and the complementary function, j
c
,
characterizes dynamics/transitions.
The particular solution. We start with the simplest solution possible; if this fails, we move up
in the degree of complexity.
If a
2
,= 0, then j
j
= /,a
2
is solution, which implies a stable point.
If a
2
= 0 and a
1
,= 0, then j
j
=
b
o
1
t .
If a
2
= 0 and a
1
= 0, then j
j
=
b
2
t
2
.
In the latter solutions, the "stable point" is moving. Recall that this solution is too restrictive,
because it constrains the dynamics of j. That is why we add the complementary function.
The complementary function solves the homogenous equation
j
tt
+a
1
j
t
+a
2
j = 0 .
133
We "guess"
j = c
vt
which implies
j
t
= rc
vt
j
tt
= r
2
c
vt
and thus
j
tt
+a
1
j
t
+a
2
j =
_
r
2
+a
1
r +a
2
_
c
vt
= 0 .
Unless = 0, we must have
r
2
+a
1
r +a
2
= 0 .
This is sometimes called the characteristic equation. The roots are
r
1,2
=
a
1

_
a
2
1
4a
2
2
.
For each root r
i
there is a potentially dierent
i
coecient. So there are two possible solutions
for the complementary function:
j
1
=
1
c
v
1
t
j
2
=
2
c
v
2
t
.
We cannot just chose one of the two solutions because this would restrict the dynamics of j. Thus,
we have
j
c
=
1
c
v
1
t
+
2
c
v
2
t
.
Given two initial conditions on j and j
t
at some point in time we can pin down
1
and
2
.
There are three options for the composition of the roots:
Two distinct real roots: r
1
, r
2
R and r
1
,= r
2
. This will give us values for
1
and
2
,
given two conditions on j.
j
c
=
1
c
v
1
t
+
2
c
v
2
t
.
Repeated real root: r
1
= r
2
R, r = a
1
,2. It might seem that we can just add up the
solution as before, but this actually not general enough because it restricts the dynamics of
j. Moreover, if we used j
c
= (
1
+
2
) c
vt
, then we cannot separately identify
1
from
2
.
134
We guess again:
j
1
=
1
c
vt
j
2
=
2
t c
vt
.
This turns out to work, because both solve the homogenous equation. You can check this.
Thus for repeated real root the complementary function is
j
c
=
1
c
v
1
t
+
2
tc
v
2
t
.
Complex roots: r
1,2
= r /i, i =
_
1, a
2
1
< 4a
2
. This gives rise to oscillating dynamics
j
c
= c
vt
[
1
cos (/t) +
2
sin(/t)]
We do not discuss in detail here.
Stability: does j
c
0?
r
1
, r
2
R: need both r
1
, r
2
< 0.
r
1
= r
2
= r R: need r < 0.
Complex roots: need r < 0.
Why do we need both
1
c
v
1
t
2
and
2
c
v
2
t
in j
c
=
1
c
v
1
t
+
2
c
v
2
t
when there are two distinct
real roots? Denote j
1
=
1
c
v
1
t
and j
2
=
2
c
v
2
t
. First, note that since both j
1
and j
2
solve the
homogenous equation and because the homogenous equation is linear, then j
1
+j
2
also solves the
homogenous equation and therefore j
c
= j
1
+ j
2
is a solution. To see why the general solution
must have this form, denote j (t
0
) = j
0
and j
t
(t
0
) = j
1
as the initial conditions for the problem
and suppose that w.o.l.g. t
0
= 0. Then
j (0) =
1
c
v
1
0
+
2
c
v
2
0
=
1
+
2
= j
0
j
t
(0) = r
1

1
c
v
1
0
+r
2

2
c
v
2
0
= r
1

1
+r
2

2
= j
1
,
which implies
_
1 1
r
1
r
2
_
. .
1
_

1

2
_
=
_
j
0
j
1
_
and since r
1
,= r
2
, the matrix 1 is nonsingular so that given any values for j
0
and j
1
there is a
unique solution.
135
When there is only one real root the matrix 1 is singular and therefore this general form for
the solution does not work: there would be an innite number of solutions. That is why we use
j
1
=
1
c
vt
and j
2
=
2
tc
vt
. As before, since each one solves the homogenous equation and because
the homogenous equation is linear, then j
1
+j
2
also solves the homogenous equation and therefore
j
c
= j
1
+j
2
is a solution. Following the steps from above,
j (0) =
1
c
v0
+
2
0c
v0
=
1
= j
0
j
t
(0) = r
1
c
v0
+
2
_
c
v0
+r0c
v0

= r
1
+
2
= j
1
,
which implies
_
1 0
r 1
_
. .
1
_

1

2
_
=
_
j
0
j
1
_
.
Now the matrix 1 is nonsingular so that given any values for j
0
and j
1
there is a unique solution
for
1
and
2
.
16.2 Dierential equations with moving constant
j
tt
+a
1
j
t
+a
2
j = / (t) ,
where a
1
and a
2
are constants. We require that / (t) takes a form that combines a nite number of
"elementary functions", e.g. /t
a
, c
It
, etc. We nd j
c
in the same way as above, because we consider
the homogenous equation where / (t) = 0. We nd j
j
by using some educated guess and verify
our guess by using the method of undetermined coecients. There is no general solution
procedure for any type of / (t).
Example: polynomial / (t):
j
tt
+ 5j
t
+ 3j = 6t
2
t 1 .
Guess:
j
j
= ,
2
t
2
+,
1
t +,
0
.
This implies
j
t
j
= 2,
2
t +,
1
j
tt
j
= 2,
2
.
Plug this into j
j
to get
j
tt
+ 5j
t
+ 3j = 2,
2
+ 5 (2,
2
t +,
1
) + 3
_
,
2
t
2
+,
1
t +,
0
_
= 3,
2
t
2
+ (10,
2
+ 3,
1
) t + (2,
2
+ 5,
1
+ 3,
0
) .
136
we need to solve
3,
2
= 6
10,
2
+ 3,
1
= 1
2,
2
+ 5,
1
+ 3,
0
= 1 .
This gives ,
2
= 2, ,
1
= 7, ,
0
= 10. Thus
j
j
= 2t
2
7t + 10 .
But this may not always work. For instance, if
j
tt
+a
1
j
t
+a
2
j = t
1
.
Then no guess of the type j
j
= ,t
1
or j
j
= ,lnt will work.
Example: missing j (t) and polynomial / (t)
j
tt
+ 5j
t
= 6t
2
t 1 .
The former type of guess,
j
j
= ,
2
t
2
+,
1
t +,
0
,
will not work, because ,
0
will never show up in the equation, so cannot be recovered. Instead, try
j
j
= t
_
,
2
t
2
+,
1
t +,
0
_
.
If this fails, try
j
j
= t
2
_
,
2
t
2
+,
1
t +,
0
_
,
and so on.
Example: exponential / (t)
j
tt
+a
1
j
t
+a
2
j = 1c
vt
.
Guess:
j
j
= tc
vt
with the same r and look for solutions for . The guess j
j
= c
vt
will not work. E.g.
j
tt
+ 3j
t
4j = 2c
4t
.
137
Guess:
j
j
= tc
4t
j
t
j
= c
4t
+4tc
4t
= c
4t
(1 4t)
j
tt
j
= 4c
4t
(1 4t) +4c
4t
= c
4t
(8 + 16t) .
Plug in the guess
j
tt
+ 3j
t
4j = c
4t
(8 + 16t) + 3c
4t
(1 4t) +4tc
4t
= c
4t
(8 + 16t + 3 12t 4t)
= 5c
4t
We need to solve
5c
4t
= 2c
4t
so = 0.4 and
j
j
= 0.4tc
4t
.
138
17 First order dierence equations
j
t+1
+aj
t
= c .
As with dierential equations, we wish to trace out a path for some variable j over time, i.e. we
seek j (t). But now time is discrete, which gives rise to some peculiarities. Dene
j
t
= j
t+1
j
t
,
(not the standard notation) which is like
j
t
t
=
j
t+t
j
t
t
,
where t = 1.
17.1 Backward iteration
1. j
t
= j
t+1
j
t
= c.
j
1
= j
0
+c
j
2
= j
1
+c = j
0
+c +c = j
0
+ 2c
j
3
= j
2
+c = j
0
+ 2c +c = j
0
+ 3c
.
.
.
j
t
= j
0
+ct .
2. aj
t+1
/j
t
= 0, a ,= 0. Then j
t+1
= /j
t
, where / = /,a.
j
1
= /j
0
j
2
= /j
1
= /
2
j
0
.
.
.
j
t
= /
t
j
0
.
17.2 General solution
j
t+1
+aj
t
= c ,
where a ,= 0. The solution method involves splitting the solution into two:
j (t) = j
c
(t) +j
j
(t) ,
where j
j
(t) is a particular solution and j
c
(t) is a complementary function.
139
j
c
(t) solves the homogenous equation
j
t+1
+aj
t
= 0 .
Guess
j
t
= /
t
so that
j
t+1
+aj
t
= 0
implies
/
t+1
+a/
t
= 0
/ +a = 0
/ = a .
j
c
(t) = (a)
t
,
where a ,= 0.
a ,= 1. j
j
(t) solves the original equation for a stationary solution, j
t
= /, a constant. This
implies
/ +a/ = c
/ =
c
1 +a
.
So that
j
j
=
c
1 +a
, a ,= 1 .
a = 1. Guess j
j
(t) = /t. This implies
/ (t + 1) /t = c
/ = c .
So that
j
j
= ct , a = 1 .
The general solution is
j
t
= j
c
(t) +j
j
(t) =
_
(a)
t
+
c
1+o
if a ,= 1
+ct if a = 1
.
Given an initial condition j (0) = j
0
, then
140
for a ,= 1
j
0
= +
c
1 +a
= = j
0

c
1 +a
.
for a = 1
j
0
= .
The general solution is
j
t
=
_ _
j
0

c
1+o
_
(a)
t
+
c
1+o
if a ,= 1
j
0
+ct if a = 1
.
For a ,= 1 we have
j
t
= j
0
(a)
t
+
_
1 (a)
t

c
1 +a
,
which is a linear combination of the initial point and the stationary point
c
1+o
. And if a (1, 1),
then this process is stable. Otherwise it is not. For a = 1 and c ,= 1 the process is never stable.
Example:
j
t+1
5j
t
= 1 .
First, notice that a ,= 1 and a ,= 0. j
c
solves
j
t+1
5j
t
= 0 .
Let j
c
(t) = /
t
, so that
/
t+1
5/
t
= 0
/
t
(/ 5) = 0
/ = 5 ,
so that
j
c
(t) = 5
t
.
j
j
= / solves
/ 5/ = 1
/ = 1,4 ,
so that j
j
= 1,4.
j
t
= j
c
(t) +j
j
(t) = 5
t
1,4 .
Given j
0
= 7,4 we have = 2, which completes the solution.
141
17.3 Dynamic stability
Given
j
t
=
_
j
0

c
1 +a
_
/
t
+
c
1 +a
,
the dynamics are governed by / (= a).
1. / < 0 will give rise to oscillating dynamics.
1 < / < 0: oscillations diminish over time. In the limit we converge on the stationary
point
c
1+o
.
/ = 1: constant oscillations.
/ < 1: growing oscillations over time. The process is divergent.
2. / = 0 and / = 1: no oscillations, but this is degenerate.
/ = 0 means a = 0, so j
t
= c.
/ = 1 means a = 1, so j
t
= j
0
+ct.
3. 0 < / < 1 gives convergence to the stationary point
c
1+o
.
4. / 1 gives divergent dynamics.
Only [/[ < 1 gives convergent dynamics.
17.4 Application: cobweb model
This is an early model of agriculture markets. Farmers determined supply last year based on the
prevailing price at that time. Consumers determine demand based on current prices. Thus, three
equations complete the description of this model
supply :
c
t+1
= : (j
t
) = +cj
t
demand :
o
t+1
= d (j
t+1
) = c ,j
t+1
equilibrium :
c
t+1
=
o
t+1
,
where c, ,, , c 0. Imposing equilibrium:
+cj
t
= c ,j
t+1
j
t+1
+
_
c
,
_
. .
o
j
t
=
c +
,
. .
c
.
142
The solution to this dierence equation is
j
t
=
_
j
0

c +
, +c
_ _

c
,
_
t
+
c +
, +c
.
The process is convergent (stable) i [c[ < [,[. Since both are positive, we need c < ,.
Interpretation: what are , and c? These are the slopes of the demand and supply curves,
respectively. If follows that if the slope of the supply curve is lower than that of the demand curve,
then the process if convergent. I.e., as long as the farmers do not "overreact" to current prices next
year, the market will converge on a happy stable equilibrium price and quantity. Conversely, as
long as consumers are not "insensitive" to prices, then...
Stable Cobweb Dynamics Unstable Cobweb Dynamics
17.5 Nonlinear dierence equations
We will use only a qualitative/graphic approach and restrict to autonomous equations, in which t
is not explicit. Let
j
t+1
= ,(j
t
) .
Draw a phase diagram with j
t+1
on the vertical axis and j
t
on the horizontal axis and the 45
degree ray starting from the origin. For simplicity, j 0. A stationary point satises j = ,(j).
But sometimes the stationary point is not stable. If [,
t
(j)[ < 1 at the stationary point, then the
process is stable. More generally, as long as [,
t
(j
t
)[ < 1 the process is stable, i.e. it will converge
to some stationary point. When [,
t
(j
t
)[ _ 1 the process will diverge.
143
Stable Nonlinear Cobweb Dierence Equation Stable Nonlinear Dierence Equation
Unstable Nonlinear Dierence Equation
Example: Galor and Zeira (1993), REStud.
144
145
18 Phase diagrams with two variables (19.5)
We now analyze a system of two autonomous dierential equations:
_ r = 1 (r, j)
_ j = G(r, j) .
First we nd the _ r = 0 and _ j = 0 loci by setting
1 (r, j) = 0
G(r, j) = 0 .
Apply the implicit function theorem separately to the above, which gives rise to two (separate)
functions:
_ r = 0 : j = )
_ a=0
(r)
_ j = 0 : j = q
_ j=0
(r) ,
where
)
t
=
1
a
1
j
q
t
=
G
a
G
j
.
Now suppose that we have enough information about 1 and G to characterize ) and q. And
suppose that ) and q intersect, which is the interesting case. This gives rise to a stationary point
(r
+
, j
+
), in which both r and j are constant:
)
_ a=0
(r
+
) = q
_ j=0
(r
+
) = j
+
.
There are two interesting cases, although you can characterize the other ones, once you do this.
18.1 Case 1: dynamic stability
1
a
< 0, 1
j
0
G
a
0, G
j
< 0 .
Both ) and q are upward sloping and ) is steeper than q at the intersection: )
t
(r
+
, j
+
) q
t
(r
+
, j
+
).
Consider a point on the )
_ a=0
locus. Now suppose that you move slightly above it or slightly
below it. How does this aect the _ r? And similarly for points slightly above or below the _ j locus.
146
By looking at the partial derivatives of 1 and G:
at all points to the right of the )
_ a=0
locus (or above the )
_ a=0
locus ) _ r < 0 and in all points
to the left of the )
_ a=0
locus _ r 0 (1
a
< 0).
at all points above the q
_ j=0
locus _ j < 0 and in all points below the q
_ j=0
locus _ j 0 (G
j
< 0).
Given an intersection, this gives rise to four regions in the (r, j) space:
1. Below )
_ a=0
and above q
_ j=0
: _ r < 0 and _ j < 0.
2. Above )
_ a=0
and above q
_ j=0
: _ r 0 and _ j < 0.
3. Above )
_ a=0
and below q
_ j=0
: _ r 0 and _ j 0.
4. Below )
_ a=0
and below q
_ j=0
: _ r < 0 and _ j 0.
This gives rise to a stable system. From any point in the (r, j) space we converge to (r
+
, j
+
).
Dynamically Stable Phase Diagram
Given the values that _ r and _ j take (given the direction in which the arrows point in the gure),
we can draw trajectories. In this case, all trajectories will eventually arrive at the stationary point
at the intersection of _ r = 0 and _ j = 0.
Notice that at the point in which we cross the _ r = 0 the trajectory is vertical.
Similarly, at the point in which we cross the _ j = 0 the trajectory is horizontal.
This will become important below.
147
18.2 Case 2: saddle point
1
a
0, 1
j
< 0
G
a
< 0, G
j
0 .
Both ) and q are still upward sloping, but now the pattern is dierent, because q
_ j=0
crosses )
_ a=0
at a steeper slope. Notice that
in all points above )
_ a=0
_ r < 0 and in all points below )
_ a=0
_ r 0.
in all points above q
_ j=0
_ j 0 and in all points below q
_ j=0
_ j < 0.
Given an intersection, this gives rise to four regions in the (r, j) space:
1. Below )
_ a=0
and above q
_ j=0
: _ r 0 and _ j 0.
2. Above )
_ a=0
and above q
_ j=0
: _ r < 0 and _ j 0.
3. Above )
_ a=0
and below q
_ j=0
: _ r < 0 and _ j < 0.
4. Below )
_ a=0
and below q
_ j=0
: _ r 0 and _ j < 0.
This gives rise to an unstable system. However, there is a stationary point at the intersection,
(r
+
, j
+
). But in order to converge to (r
+
, j
+
) there are only two trajectories that bring us there,
one from the region above )
_ a=0
and below q
_ j=0
, the other from the region below )
_ a=0
and above
q
_ j=0
. These trajectories are called stable branches. If we are not on those trajectories, then we
are on unstable branches. Note that being in either region does not ensure that we are on a stable
branch, as the gure illustrates.
148
Saddle Point Phase Diagram
149
19 Optimal control
Like in static optimization problems, we want to maximize (or minimize) an objective function.
The dierence is that the objective is the sum of a path of values at any instant in time; therefore,
we must choose an entire path as a maximizer.
2
The problem is generally stated as follows:
Choose n(t) to maximize
_
T
0
1 (j, n, t) dt
s.t.
Law of motion : _ j = q (j, n, t)
Initial condition : j (0) = j
0
transversality condition : j (T) c
vT
_ 0 .
where r is some average discount rate that is relevant to the problem. To this we need to sometimes
add
Terminal condition : j (T) = j
T
Constraints on control : n(t) l
The function j (t) is called the state variable. The function n(t) is called the control variable.
It is useful to think of the state as a stock (like capital) and the control as a ow (like investment or
consumption). Usually we will have 1, q C
1
, but in principle we could do without dierentiability
with respect to n. I.e., we only need that the functions 1 and q are continuously dierentiable with
respect to j and t.
In a nite horizon problem (T < ) c
vT
0 so that the transversality condition immediately
implies that j (T) _ 0, but also something more: If this constraint binds, then j (T) = 0. Either
way, this tells you that the value of j at the end of the problem cannot be negative. This will
become clearer below, when we discuss the Lagrangian approach.
If there is no law of motion for j, then we can solve the problem separately at any instant as
a static problem. The value would just be the sum of those static values.
There is no uncertainty here. To deal with uncertainty, wait for your next course in math.
To ease notation we will omit time subscripts when there is no confusion.
2
The theory behind this relies on "calculus of variations", which was rst developed to compute trajectories of
missiles (to the moon and elsewhere) in the U.S.S.R.
150
Example: the saving/investment problem for individuals.
1. Output: 1 = 1 (1, 1).
2. Investment/consumption: 1 = 1 C = 1 (1, 1) C.
3. Capital accumulation:
_
1 = 1 c1.
We want to maximize the present value of instantaneous utility from now (at t = 0) till we die
(at some distant time T). The problem is stated as
Choose C (t) to maximize
_
T
0
c
jt
l [C (t)] dt
s.t.
_
1 = 1 c1
1 (0) = 1
0
1 (T) = 1
T
.
19.1 Pontryagins maximum principle and the Hamiltonian function
Dene the Hamiltonian function:
H (j, n, t, `) = 1 (j, n, t) +`q (j, n, t) .
The function `(t) is called the co-state function and also has a law of motion. Finding ` is part
of the solution. The FONCs of this problem ensure that we maximize H at every point in time,
and as a whole. If n
+
is a maximizing plan then
(i) : H (j, n
+
, t, `) _ H (j, n, t, `) \n l
or :
0H
0n
= 0 if 1, q C
1
State equation (ii) :
0H
0`
= _ j = _ j = q (j, n, t)
Costate equation (iii) :
0H
0j
=
_
` =
_
` +1
j
+`q
j
= 0
Transversality condition (iv) : `(T) = 0 or other (see below) .
Conditions (ii)+(iii) are a system of rst order dierential equations that can be solved explicitly
if we have functional forms and two conditions: j (0) = j
0
and `(T) = 0. But `(T) = 0 is only
one way to get a transversality/reminal condition.
151
Interpretation of the Hamiltonian: n and j aect the value of the problem directly
through 1. But they also aect the value of the problem indirectly, through their eect on
_ j. This is captured by `q. So in this context ` is the cost/benet of allowing j to grow a bit
faster. So ` has the same interpretation as the Lagrange multiplier: it is the shadow cost of
the constraint at any instant.
We adopt the convention that j (0) = j
0
is always given. There are a few way to introduce
terminal conditions, which gives the following taxonomy
1. When T is xed,
(a) `(T) = 0, j (T) free.
(b) j (T) = j
T
, `(T) free.
(c) j (T) _ j
min
(or j (T) _ j
max
), `(T) free. Add the following complementary slackness
conditions:
j (T) _ j
min
`(T) _ 0
`(T) (j (T) j
min
) = 0
2. T is free and j (T) = j
T
. Add H (T) = 0.
3. T _ T
max
(or T _ T
min
) and j (T) = j
T
. Add the following complementary slackness
conditions:
H (T) _ 0
T _ T
max
H (T) (T
max
T) = 0
19.2 The Lagrangian approach
This is based on Barro and Sala-i-Martin (2001), Economic Growth, MIT Press, third edition.
The problem is
Choose n(t) to maximize
_
T
0
1 (j, n, t) dt
s.t.
152
_ j = q (j, n, t)
j (T) c
vT
_ 0
j (0) = j
0
.
You can think of _ j = q (j, n, t) as an inequality _ j _ q (j, n, t), in which we control both n and j.
We can write this up as a Lagrangian. For this we need Lagrange multipliers for the law of motion
constraint at every point in time, as well as an additional multiplier for the transversality condition:
/ =
_
T
0
1 (j, n, t) dt +
_
T
0
`(t) [q (j, n, t) _ j] dt +0j (T) c
vT
=
_
T
0
[1 (j, n, t) +`(t) q (j, n, t)]
. .
1(j,&,t,A)
dt
_
T
0
`(t) _ j (t) dt +0j (T) c
vT
.
In this context, both n and j are part of the "critical path" (paraphrasing critical point). The
problem here is that we do not know how to take the derivative of _ j w.r.t. j. To avoid this, use
integration by parts to get

_
` _ jdt = `j +
_
_
`jdt
so that
/ =
_
T
0
[1 (j, n, t) +`(t) q (j, n, t)] dt [`(t) j (t)[
T
0
+
_
T
0
_
`(t) j (t) dt +0j (T) c
vT
=
_
T
0
[1 (j, n, t) +`(t) q (j, n, t)] dt `(T) j (T) +`(0) j (0) +
_
T
0
_
`(t) j (t) dt +0j (T) c
vT
.
The FONCs for the Lagrangian are
(i) : /
&
= 1
&
+`q
&
= 0
(ii) : /
j
= 1
j
+`q
j
+
_
` = 0
(iii) : _ j = q .
These are consistent with
(i) : H
&
= 1
&
+`q
&
= 0
(ii) : H
j
= 1
j
+`q
j
=
_
`
(iii) : H
A
= q = _ j ,
which are the FONCs for the Hamiltonian
H (j, n, t, `) = 1 (j, n, t) +`(t) q (j, n, t) .
153
The requirement that j (0) = j
0
can also be captured in the usual way, as well as j (T) = j
T
, if
it is required. The transversality condition is captured by the complementary slackness conditions
j (T) c
vT
_ 0
0 _ 0
0j (T) c
vT
= 0 .
We see here that if j (T) c
vT
0, then its value, 0, must be zero.
19.3 Autonomous problems
In these problems t is not an explicit argument.
Choose n to maximize
_
T
0
1 (j, n) dt s.t. _ j = q (j, n)
plus boundary conditions. The Hamiltonian is thus
H (j, n, `) = 1 (j, n) +`q (j, n) .
These problems are easier to solve and are amenable to analysis by phase diagrams.
19.3.1 Example: the cake eating problem (with no discounting)
Objective: You want to eat your cake in an optimal way, maximizing your satisfaction from eating
it, starting now (t = 0) and nishing before bedtime, at T.
The cake starts at size o
0
.
When you eat cake, the size diminishes by the amount that you ate:
_
o = C.
You like cake, but less so when you eat more: l
t
(C) 0, l
tt
(C) < 0. If you are not eating
cake, you really want some: l
t
(0) = . But if you have too much, you dont enjoy any more
of it: l
t
() = 0.
The problem is
Choose C to maximize
_
T
0
l (C) dt s.t.
_
o = C
o (0) = o
0
o (T) _ 0 .
154
This is an autonomous problem. The Hamiltonian is
H (C, o, `) = l (C) +`[C] .
FONCs:
(i) :
0H
0C
= l
t
(C) ` = 0
(ii) :
0H
0`
= C =
_
o
(iii) :
0H
0o
= 0 =
_
`
(iv) : o (T) _ 0, `(T) _ 0, o (T) `(T) = 0 .
From (iii) it follows that ` is constant. From (i) we have l
t
(C) = ` 0, and since ` is constant,
C is constant too. Then given a constant C we get from (ii) that
o = Ct .
And given o (0) = o
0
we have
o = o
0
Ct .
But we still do not know what is C, except that it is constant. So we solve for the complementary
slackness conditions, i.e., will we leave leftovers?
Suppose ` 0. Then o (T) = 0. Therefore
0 = o
0
CT ,
which gives
C =
o
0
T
.
Suppose ` = 0. Then it is possible to have o (T) 0. But then we get l
t
= 0a contradiction.
The solution is thus
C (t) = o
0
,T
`(t) = l
t
(o
0
,T)
o (t) = o
0
(o
0
,T) t ,
where only o (t) evolves over time and C and ` are constants.
If we allowed a at part in the utility function after some satiation point, then we could have
a solution with leftovers o (T) 0. In that case we would have more than one optimal path: all
would be global because with one at part l is still quasi concave.
155
19.3.2 Anecdote: the value of the Hamiltonian is constant in autonomous problems
We demonstrate that on the optimal path the value of the Hamiltonian function is constant.
H (j, n, t, `) = 1 (j, n, t) +`q (j, n, t) .
The derivative with respect to time is
dH
dt
= H
&
_ n +H
j
_ j +H
A
_
` +H
t
.
The FONCs were
H
&
= 0
H
j
=
_
`
H
A
= _ j .
Plugging this into dH,dt gives
dH
dt
=
0H
0t
.
This is in fact a consequence of the envelope theorem, although not in a straightforward way. If
time is not explicit in the problem, then
01
0t
= 0, which implies the statement above.
19.4 Current value Hamiltonian
Many problems in economics involve discounting, so the problem is not autonomous. However,
usually the only place that time is explicit is in the discount factor,
_
T
0
1 (j, n, t) dt =
_
T
0
c
vt
G(j, n) dt .
You can try to solve those problems "as-is", but an easier way (especially if the costate is of no
particular interest) is to use the current value Hamiltonian:

H = c
vt
H = G(j, n) +,q (j, n) ,
where
, = `c
vt
.
156
A maximizing plan n
+
satises the following FONCs:
(i) :

H (j, n
+
, ,) _

H (j, n, ,) \n l
or :
0

H
0n
= 0 if

H, q C
1
State equation (ii) :
0

H
0,
= _ j = _ j = q (j, n)
Costate equation (iii) :
0

H
0j
= _ , +r, = _ , r, +1
j
+`q
j
= 0
Transversality condition (iv) : ,(T) = 0 or

H (T) = 0 or other.
Since , = `c
vt
we have
_ , =
_
`c
vt
+`rc
vt
=
_
`c
vt
+r, .
Therefore

_
`c
vt
= _ , +r, ,
which is what
0

H
0j
=
0
0j
_
c
vt
H

= c
vt
0H
0j
=
_
`c
vt
.
implies.
19.4.1 Example: the cake eating problem with discounting
We now need to choose a functional form for the instantaneous utility function. The problem is
Choose C to maximize
_
T
0
c
vt
ln(C) dt s.t.
_
o = C
o (0) = o
0
o (T) _ 0 .
We write the present value Hamiltonian

H = lnC +,[C]
157
FONCs:
(i) :
0

H
0C
=
1
C
, = 0
(ii) :
0

H
0,
= C =
_
o
(iii) :
0

H
0o
= 0 = _ , +r,
(iv) : o (T) _ 0, ,(T) _ 0, o (T) ,(T) = 0 .
From (iii) we have
_ ,
,
= r ,
hence
, = 1c
vt
,
for some 1. From (i) we have
C =
1
,
=
1
1
c
vt
.
From (ii) we have
_
o = C
_
t
0
_
od. =
_
t
0
Cd.
o (t) = +
_
t
0
Cd. ,
which, together with o (0) = o
0
implies
o (t) = o
0

_
t
0
Cd. ,
which makes sense. Now, using C = 1
1
c
vt
we get
o (t) = o
0

_
t
0
1
1
c
v:
d.
= o
0
1
1
_

1
r
c
v:

t
0
= o
0
1
1
_

1
r
c
vt
+
1
r
c
v0
_
= o
0

1
r1
_
1 c
vt

Suppose ,(T) = 0. Then 1 = 0 and C (T) = not possible (we could have obtained this
158
result directly from (i)). So ,(T) 0, which implies o (T) = 0. Therefore
0 = o
0

1
r1
_
1 c
vT

1 =
_
1 c
vT

ro
0
Therefore
C =
ro
0
[1 c
vT
]
c
vt
,
which is decreasing, and
, =
_
1 c
vT

ro
0
c
vt
,
which is increasing. And nally
o (t) = o
0
_
1
1 c
vt
1 c
vT
_
.
This completes the characterization of the problem.
19.5 Innite time horizon
When the problems horizon is innite, i.e. never ends, we need to modify the transversality
condition. These are
lim
To
`(T) j (T) = 0
for the present value Hamiltonian, and
lim
To
,(T) c
vT
/ (T) = 0
for the current value Hamiltonian.
19.5.1 Example: The neoclassical growth model
1. Preferences: n(C), n
t
0, n
tt
< 0. Inada conditions: n(0) = 0, n
t
(0) = , n
t
(C) 0 as
C .
2. Aggregate production function: 1 = 1 (1, 1), CRS, 1
i
0, 1
ii
< 0. Given this we can
write the per-worker version j = ) (/), where )
t
0, )
tt
< 0 and j = 1,1, / = 1,1. Inada
conditions: ) (0) = 0, )
t
(0) = , )
t
() = 0.
3. Capital accumulation:
_
1 = 1 c1 = 1 C c1. As we saw in the Solow model, we can
write this in per worker terms
_
/ = ) (/) c (: +c) /, where : is the constant growth rate
of labor force.
159
4. There cannot be negative consumption. In addition, once output is converted into capital,
we cannot eat it. This can be summarized in 0 _ C _ 1 (1, 1). This is an example of a
restriction on the control variable.
5. A social planner chooses a consumption plan to maximize everyones welfare, in equal weights.
The objective function is
\ =
_
o
0
1
0
c
at
c
jt
n(c) dt =
_
o
0
c
vt
n(c) dt ,
where we normalize 1
0
= 1 and we set r = j : 0, which ensures integrability. Notice
that everyone gets the average level of consumption c = C,1.
The problem is
Choose c to maximize \ s.t.
_
/ = ) (/) c (: +c) /
0 _ c _ ) (/)
/ (0) = /
0
Write down the current value Hamiltonian
H = n(c) +,[) (/) c (: +c) /] .
FONCs:
H
c
= n
t
(c) , = 0
H
,
= [) (/) c (: +c) /] =
_
/
H
I
= ,
_
)
t
(/) (: +c)

= r, _ ,
lim
To
,(T) c
vT
/ (T) = 0
Ignore for now 0 _ c _ ) (/). The transversality condition here is a sucient condition for a
maximum, although in general this specic condition is not necessary. If this was a present value
Hamiltonian the same transversality condition would be lim
To
`(T) / (T) = 0, which just means
that the value of an additional unit of capital in the limit is zero.
From H
c
we have n
t
(c) = ,. From H
I
we have
_ ,
,
=
_
)
t
(/) (: +c +r)

.
160
We want to characterize the solution qualitatively using a phase diagram. To do this, we need
two equations: one for the state, /, and one for the control, c. Notice that
_ , = n
tt
(c) _ c ,
so
n
tt
(c) _ c
n
t
(c)
=
_
)
t
(/) (: +c +r)

.
Rearrange to get
_ c
c
=
n
t
(c)
cn
tt
(c)
_
)
t
(/) (: +c +r)

.
Notice that

cn
tt
(c)
n
t
(c)
is the coecient of relative risk aversion. Let
n(c) =
c
1o
1 o
.
This is a class of constant relative relative risk aversion (CRRA) utility functions, with coecient
of RRA = o.
Eventually, our two equations are
_
/ = ) (/) c (: +c) /
_ c
c
=
1
o
_
)
t
(/) (: +c +r)

.
From this we derive
_
/ = 0 : c = ) (/) (: +c) /
_ c = 0 : )
t
(/) = : +c +r .
The _ c = 0 locus is a vertical line in the (/, c) space. Given the Inada conditions and diminishing
returns to capital, we have that the
_
/ = 0 locus is hump shaped. Since r 0, the peak of the hump
is to the right of the vertical _ c = 0 locus.
The phase diagram features a saddle point, with two stable branches. If / is to the right of the
_ c = 0 locus, then _ c < 0 and vice versa for / to the left of the _ c = 0 locus. For c above the
_
/ = 0
locus we have
_
/ < 0 and vice versa for c below the
_
/ = 0 locus. See textbook for gure.
Dene the stationary point as (/
+
, c
+
). Suppose that we start with /
0
< /
+
. Then the optimal
path for consumption must be on the stable branch, i.e. c
0
is on the stable branch, and c (t) will
eventually go to c
+
. The reason is that any other choice is not optimal. Higher consumption will
eventually lead to depletion of the capital stock, which eventually leads to no output and therefore
161
no consumption (U.S.A.). Too little consumption will lead rst to an increase in the capital stock
and an increase in output, but eventually this is not sustainable as the plan requires more and
more consumption forgone to keep up with eective depreciation (: + c) and eventually leads to
zero consumption as well (U.S.S.R.).
One can do more than just analyze the phase diagram. First, given functional forms we can
compute the exact paths for all dynamic variables. Second, we could linearize (a rst order Taylor
expansion) the system of dierential equations around the saddle point to compute dynamics around
that point (or any other point, for that matter).
162