1 Differentiation in Rn

Recall first the usual definition of differentiation of functions.

Definition 1 A function f : R → R is differentiable at x ∈ R if there is a

number f 0 (x) such that
f (x + h) − f (x)
lim = f 0 (x) . (1)
h→0 h
By rewriting 1 we obtain a new geometric interpretation,

f (x + h) − f (x) = f 0 (x) h + r (h) , (2)

where the remainder is small in the sense that

r (h)
lim = 0,
h→0 h

which we will write as r (h) = o (h).

With this definition at hand, we can interpret the derivative in a slightly
different manner. Notice that 2 expresses the behavior of the original function
f around a point x (i.e. at points of the form x + h with h small) as a sum of
a linear function in h, f 0 (x) h and the error term r (h) . In other words, the
operation of taking derivatives is viewed as assigning linear functions that
approximate the original function around a given point to all points in the
domain of f .
The equation 1 certainly makes no sense in the general case of a function
f : Rn → Rm but the view of the derivative as a local linear approximation
to f around a point x certainly does.

Definition 2 A function f : Rn → Rm is differentiable at x ∈ Rn if there is

a linear function λ : Rn → Rm such that
kf (x + h) − f (x) − λhk
lim = 0.
h→0 khk

1
 The linear transformation λ is denoted Df (x) and called the derivative
(or differential or total derivative) of f at x. The matrix of Df (x) : Rn → Rm
is a m × n matrix and is called the Jacobian matrix of f at x. If f : Rn → R,
then the acobian matrix is a row vector.

Proposition 1 If a function f : Rn → Rm is differentiable at x ∈ Rn then

there is a unique linear function λ : Rn → Rm such that
kf (x + h) − f (x) − λhk
lim = 0.
h→0 khk
Proposition 2 (Chain Rule) Suppose E is an open set in Rn , f maps E
into Rm , f is differentiable at x0 ∈ E, g maps an open set containing f (E)
into Rk , and g is differentiable at f (x0 ) . Then the mapping F of E into Rk
defined by
F (x) = g (f (x)) ,
is differentiable at x0 and

DF (x0 ) = Dg (f (x0 )) Df (x0 ) .

Let f be a function, f : Rn → Rm . The components of f are the real

valued functions f1 , ..., fm defined by
m
X
f (x) = fi (x) ei
i=1

or equivalently fi (x) = f (x) ei , where ei is the unit vector in the i − th

direction. For x ∈ Rn , 1 ≤ i ≤ m, 1 ≤ j ≤ n, we define
fi (x + hej ) − fi (x)
Dj fi = lim ,
h→0 h
provided that the limit exists. Writing fi (x1 , ..., xn ) in place of fi (x) we see
that Dj fi is the derivative of fi with respect to xj :
∂fi
Dj fi (x) = ,
∂xj
and is called a partial derivative.

2
Proposition 3 If f : Rn → Rm is differentiable at a, then Dj fi exists for
1 ≤ i ≤ m, 1 ≤ j ≤ n and Df (x) is the m × n matrix Dj fi (x).

Proposition 4 If f : Rn → Rm , then Df (x) exists if all Dj fi (x) exist and

are continuous in an open set containing x.

A function that has continuous partial derivatives in an open set contain-

ing x is called continuously differentiable at x.

1.1 Partial Derivative, Directional Derivative and Gra-

dient
In this section, we consider functions f : Rn → R.

Definition 3 The directional derivative of f at x in direction u ∈ Rn is

denoted by the symbol f 0 (x; u) and is defined by the equation
f (x + hu) − f (x)
f 0 (x; u) = lim ,
h→0 h
whenever the limit on the right exists.

If u = ei , the i − th unit coordinate vector, then f (x, ei ) is called the par-

tial derivative and is denoted by Di f (x). If f (x; u) exists in ecery direction,
then in particular all the partial derivatives D1 f (x) , ..., Dn f (x) exist. The
converse is not true.

Example 1 Consider the real-valued function f : R2 → R given by

(
x + y if x = 0 or y = 0,
f (x, y) =
1 otherwise.

A rather surprising fact is that a function can have a finite directional

derivative f (x; u) for every u but may fail to be even continuous and therefore
at x.

3
Example 2 Let
xy 2
(
(x2 +y 4 )
if x 6= 0,
f (x, y) =
0 if x = 0.

For this reason, directional derivatives as well as partial derivatives are a

somewhat unsatisfactory extension of the one-dimesnional concept of deriva-
tive. The generalization that we defined earlier does not suffer from these
defects..
Consider next a real valued function f : Rn → R Associated with each
x ∈ Rn is a vector, the gradient of f at x, defined by

∇f (x) = Df (x)> .

The geometric interpretation of the gradient of a real-valued function of a

vector variable is worth noting. Recall that

y − f (x̂) = f 0 (x̂) (x − x̂)

is the equation in the hyperplane in Rn+1 , which is the tangent to y = f (x)

where x = x̂.
f (x + hu) − f (x)
lim = ∇f (x) · u, (3)
h→0 h
we define the directional derivative of f at x in the direction of the unit
vector u, and is denoted by (Du f ) (x). If f and x are fixed, but u varies,
then 3 shows that (Du f ) (x) attains its maximum when u is a positive scalar
of (∇f ) (x). In other words, the gradient of a function gives the direction in
which the function increases at its quickest rate.
Let v ∈ Rn be such that ∇f (x) · v = 0. Then the value of the function is
locally constant around x in direction v. The set of points {x ∈ Rn |f (x) = c }
is called the level curve of f at value c. This argument shows that the gradient
is perpendicular to the tangent to the level curve of the function.

4
2 Taylor’s Formula
The basic objective in differentiating a function is to obtain a linear approx-
imation of the function under analysis. This allows for an easier analysis
of the function locally around the point at which it is differentiated. For
some questions, the linear approximation contains all the information that is
really needed. For example, in the study of dynamical systems (which form
an essential part of modern macroeconomics), it is often enough to linearize
the system around a steady state of the system in order to analyze the local
behavior of the system around that steady state. For other questions, the
information contained in the linear approximation is insufficient. The pri-
mary instance of this is optimization theory. In this area, the basic task is
to locate the points in the domain of a function where that function achieves
its maximal (or minimal) value. Linear approximations are used in locating
these points, but in order to tell apart the points where the maximum is
obtained from the points where the function is minimized, more information
is needed.
In this section, a function is approximated around a fixed point x by a
polynomial. This approximation will obviously be at least as accurate as the
linear approximation obtained in the theory of differentiation (since linear
functions are also polynomials). The information contained in the polynomial
approximation of a function is sufficient for the purposes of finding minima
nad maxima.
We start by recalling some facts from eelementary calculus.

Proposition 5 (Rolle). Assume f has a derivative at each point of an open

interval (a, b), and assume that f is continuous at both endpoints a and b. If
f (a) = f (b) there is at least one interior point c at which f 0 (c) = 0.

Proposition 6 (Generalized Mean Value Theorem) Let f and g be two

functions, each having a derivative at each point of an open interval (a, b)

5
and each continuous at the endpoints a and b. Assume also that there is no
interior point x at which both f ’(x) and g’(x) are infinite. Then for some
interior point c we have

f 0 (c) [g (b) − g (a)] = g 0 (c) [f (b) − f (a)] .

These two results can be used to prove the famous approximation result
for continuously differentiable real valued functions on the real line.

Proposition 7 (Taylor’s theorem for R). Let f be a function having

finite n-th derivative f (n) everywhere in an open interval (a, b) and continuous
(n − 1) st derivatives in the closed interval [a, b] . Assume that c ∈ [a, b].
Then, for every x ∈ [a, b] and x 6= c, there exists a point x1 interior to the
interval joining x and c such that:
n−1 (k)
X f (c) f (n) (x1 )
f (x) = f (c) + (x − c)k + (x − c)n .
k=1
k! n!

Taylor’s formula can be extended to real-valued functions f defined on

subsets of Rn . In order to state the general theorem in a form which resembles
the one-dimensional case, we introduce special symbols

f 00 (x; t) , f 000 (x; t) , ..., f (m) (x; t) ,

for certain sums that arise in Taylor’s formula. These play the role of higher
order directional derivatives, and they are defined as follows. If x is a point in
Rn where all second-order partial derivatives of f exist, and if t = (t1 , ..., tn )
is an arbitrary point in Rn , we write
n X
X n
00
f (x; t) = Dij f (x) tj ti .
i=1 j=1

We also define n X
n X
n
X
000
f (x; t) = Dijk f (x) tk tj ti
i=1 j=1 k=1

6
if all third-order partial derivatives exist at x. The symbol f (m) (x; t) is
similarly defined if all m − th order partials exist.

Proposition 8 (Taylor’s theorem for Rn ). Assume that f and all its

partial derivatives of order < m are differentiable at each point of an open
set S in Rn . If a and b are two points of S such that L(a, b) ⊆ S, then there
is a point z on the line segment L(a, b) such that
m−1
X f (k) (a; b − a) f (m) (z; b − a)
f (b) − f (a) = + .
k=1
k! m!

In the lectures on optimization, we will use the properties of f 00 (x; t) to

classify points where f 0 (x) = 0 as either local maxima or local minima of
the function around that point.

3 The Implicit Function Theorem

The functions we analyzed so far were almost all explicit and of the form

y = f (x) ,

where the exogenous variable x appeared on the right hand side and the
endogenous variable appeared on the left hand side. When endogenous and
exogenous variables cannot be separated as in

f (x, y) = c, (4)

then they define an implicit function. The problem is to decided whether

this equation determines y as a function of x. If so, we have

y = g (x) ,

for some function g. We say that g is defined implicitly by (4) . The problem
assumes a more general form when we have a system of several equations

7
involving several variables in terms of the remaining variables. An important
special case is the familiar problem in algebra of solving n linear equations
of the form n
X
aij yj = ti , (i = 1, 2, ...., n) , (5)
j=1

Each equation (5) can be written in the form of (4) by setting

n
X
fi (y, x) = aij yj − yi = 0.
j=1

More generally, consider the linear function f : Rn+k → Rn given by

the n × (n + k) matrix A. Denote the first n components of Rn+k by y =
(y1 , ..., yn ) and the last k components by x = (x1 , ..., xk ) . Suppose that we
are interested in the behavior of the component y in the solutions to the
equation
A (y, x) = 0.

Using the results from linear algebra, we may try to solve explicitly for the
y 0 s in terms of the x0 s. For this purpose, write

A = (Ay , Ax )

where
   
a11 · · · a1n a1(n+1) · · · a1(n+k)
 . ... .. 
  .. .. .. 
Ay =  . .  and Ax =  . .
 .  . . 
an1 · · · ann an(n+1) · · · an(n+k)

Then the linear equation can be written as

Ay y+Ax x = 0,

and thus we can solve for y:

y= −A−1
y Ax x.

8
Clearly, this can only be done if Ay is invertible, i.e. if det (Ay ) 6= 0.
In this section, our goal is to extend this analysis beyond linear func-
tions. Unfortunately, it is in general impossible to find explicit solutions for
n non-linear equations in n unknowns (the y above). Nevertheless, we can
sometimes achieve a more modest goal by looking at the set of equations
locally around a solution to the system. Recall that the basic method of
local analysis was to linearize the function in question around a given point
and analyze the linear approximation. The derivative of the function gives
us the desired approximation, and we can use a similar line of argument to
get local results as we did in thelinear example above.
More specifically, for a given implicit function f (y, x) we want to know
the answers to the following two questions:

1. does f (y 0 , x0 ) determine y 0 as a continuous function of x for x near

x0 ?

2. how does a change in x affect the corresponding y0 s ?

The answer to these questions is provided in the following theorem.

Theorem 9 (Implicit Function Theorem). Let f = (f1 , ..., fn ) be a

vector-valued function defined on an open set S in Rn+k with values in Rn .
Suppose f ∈ C 1 on S. Let (y0 ; x0 ) be a point in S for which f (y 0 ; x0 ) = 0 and
for which the n × n determinant det [Dy f (y 0 ; x0 )] 6= 0. Then there exists a
k-dimensional open set X 0 containing x0 and a unique vector valued function
g defined on X 0 and having values in Rn , such that
(a) g ∈ C 1 on T 0 ,
(b) g (x0 ) = y 0 ,
(c) f [g (x) ; x] = 0, for every x ∈ X 0 .

9
 −1
(d) Dx g (x0 ) = − (Dy f (y 0 , x0 )) Dx f (y 0 , x0 ) , where
   
∂f1 ∂f1 ∂f1 ∂f1
∂y1
· · · ∂yn ∂x1
··· ∂xk
 . .
. . . . ..  and Dx f y 0 , x0 =  .. . . .
  . .. 
Dy f y 0 , x 0 = 



 .   .
. 
∂fn ∂fn ∂fn ∂fn
∂y1
· · · ∂yn ∂x1
··· ∂xk

Another important result from multivariate calculus that uses the local
linear approximation of the original function by its derivative is the following.

Proposition 10
Theorem 11 (Inverse Function Theorem). Assume f = (f1 , ..., fn ) ∈
C 1 on an open set S in Rn , and let T = f (S). If the Jacobian determinant
Jf (a) 6= 0 for some point a ∈ S, then there are two open sets X ⊆ S and
Y ⊆ T and a uniquely determined function g such that:
(a) a ∈ X, f (y) ∈ Y,
(b) Y = f (X) ,
(c) f is one-to-one on X,
(d) g is defined on Y , g (Y ) = X, and g [f (x)] = x, for every x∈X,
(e) g ∈ C 1 on Y.

The Inverse Function Theorem is a powerful result. In general, it is

nearly impossible to check directly that a given map is one-to-one and onto.
By Theorem 11 we need only to show that its derivative is a nonsingular
matrix - a much easier task.
Here are two examples of the use of Implicit Function Theorem that are
typical in Economics.

Example 3 Let f : Rn → R. The level curve of f at value c is defined as

the set {x ∈ Rn |f (x) = c } . We may then ask when one of the endogenous
variables, say x1 is defined implicitly as a function of the other variables
(x2 , ..., xn ) ∈ Rn−1 around a point x∗ on the level curve, i.e. at a point

10
x∗ satisfying f (x∗ ) = c. Implicit Function Theorem shows that whenever
∂f (x∗ )
∂x1
6= 0, there is a function g : Rn−1 → R and an ε > 0 such that
i) x∗1 = g (x∗2 , ..., x∗n ) ,
ii) For all (x2 , ..., xn ) ∈ Bε ((x∗2 , ..., x∗n )) , f (g (x2 , ..., xn ) , (x2 , ..., xn )) = c,
∂f (x∗ )
∂g ((x∗2 ,...,x∗n ))
iii) ∂xi
= − ∂f∂x(xi∗ ) .
∂x1
Observe that the last conclusion follows from ii) by the chain rule. The
interpretation of the result is the following. Suppose that one of the variables,
xi is increased by a small amount dxi . The change in x1 needed in order to
keep the value of the function f constant at c is then given by
∂f (x∗ )
− ∂f∂x i
(x∗ )
.
∂x1

In microeconomics, the problem of utility maximization results in level curves

that are called indifference curves (since the buyer is indifferent between all
those points as the utility index corresponding to them is constant). Expres-
∂f (x∗ )
∂xi
sions of the form ∂f (x∗ ) are called the marginal rates of substitution between
∂xj

i and j.

Example 4 Implicit function theorem is the primary tool of what is called

the comparative statics analysis of economic models. In this analysis, eco-
nomic considerations lead to behavioral equations linking the endogenous
quantities (e.g. price and quantity of oil demanded and supplied) to exoge-
nous variables (weather conditions, outbreak of a war etc.). The main task of
the analysis is to predict how the endogenous variables will change in response
to changes in the exogenous variables.
Let y1 denote the quantity of oil demanded, let y2 denote the quantity
of oil supplied and let y3 denote the price of oil. Let also x1 denote the
average temperature (say that we are talking about the winter months) and
x2 denote a variable recording the military activity in the middle east. From

11
the consumers’ optimal oil purchasing decisions, we can derive the demand
function for oil:
y1 = M (y3 , x1 ) .

From the producers’ optimal production decisions, we can derive the supply
function for oil:
y2 = S (y3 , x2 ) .

Finally, a market is in equilibrium if:

y1 = y2 .

Using the notation from the implicit function theorem, we can write

f1 = M (y3 , x1 ) − y1 ,

f2 = S (y3 , x2 ) − y2 ,

f3 = y1 − y2 .

Then any point y 0 , x0 = (y10 , y20 , y30 , x01 , x02 ) satisfying

f y 0 , x0 = f1 y 0 , x0 , f2 y 0 , x0 , f3 y 0 , x0 = 0

gives an equilibrium of the economic model.

In order to see if implicit function theorem is applicable, we need to eval-
uate Dy f (y 0 , x0 ) . But this is clearly:
∂M (y0 ,x0 )
 
−1 0 ∂y3
∂S (y0 ,x0 )
0 0

 
Dy f y , x =   0 −1
.
∂y3 
1 −1 0

∂M (y0 ,x0 ) ∂S (y0 ,x0 )

Hence det Dy f (y 0 , x0 ) = ∂y3
− ∂y3 . In any reasonable model of de-
∂M (y0 ,x0 ) ∂S (y0 ,x0 )
mand and supply, ∂y3
< 0 and ∂y3
> 0, i.e. demand is downward
sloping and supply is upward sloping. Hence the deteminant is megative and

12
thus nonzero. As a result, the implicit function theorem is applicable. Hence
we may write around the initial point (y 0 , x0 ) y = g (x0 ) , where g : R2 → R3 .
Since
∂M (y0 ,x0 )
 
∂x1
0
∂S (y0 ,x0 )

 
Dx f y 0 , x 0 =  0
,
 ∂x2 
0 0

1
Dx g x0 = − ∂M (y0 ,x0 ) ∂S(y0 ,x0 ) ×

∂y3
− ∂y3
( ) ∂M (y0 ,x0 ) ∂M (y0 ,x0 ) ∂M (y0 ,x0 )
 ∂S y0 ,x0  
− 0
 ∂S ∂y3 ∂y3 ∂y3 ∂x1
 (y0 ,x0 ) ∂M (y0 ,x0 ) ∂S (y0 ,x0 ) ∂S (y0 ,x0 )
 
− ∂y3 0
 
 ∂y3 ∂y3  ∂x2 
1 −1 1 0 0

It makes sense to assume that a decrease in mean temperature increases

∂M (y0 ,x0 )
the demand for oil at all prices (during winter months). Thus ∂x1
<0
and we can calculate the effect of a change in temperature on the oil price
∂M (y0 ,x0 )
∂y3 ∂g3 (y0 ,x0 )
(i.e. ∂x1 ) by evaluating the term ∂x1
= − ∂M (y0 ,x0 )∂x1∂S(y0 ,x0 ) < 0.
∂y3
− ∂y3

4 Exercises
1. A function f : R2 → R2 is given. Verify that the inverse function
theorem is applicable and find the inverse function f −1 .
(i) y1 = x1 , y2 = x21 + x2 ,
(ii) y1 = 2x1 − 3x2 , y2 = x1 + 2x2 .

2. In each of the next two problems a function F and a point P are given.
Verify that the implicit function theorem is applicable. Denoting the
implicitly defined function by g, find the values of all the first partial
derivatives of g at P .

13
 (i) F = (F 1 , F 2 ) , P = (0, 0, 0, 0) , where F 1 = 2x1 −3x2 +y1 −y2 , F 2 =
x1 + 2x2 + y1 − 2y2 ,
(ii) F = (F 1 , F 2 ) , P = (0, 0, 0, 0) , where F 1 = 2x1 −x2 +2y1 −y2 , F 2 =
3x1 + 2x2 + y1 + 2y2 .

3. Consider the following system of equations:

f1 (y1 , y2 ; α, β, q) = y1 + αy22 − q = 0
f2 (y1 , y2 ; α, β, q) = βy12 + y2 − q = 0

i)Can you solve (y1 , y2 ) as functions of (α, β, q) in a neighborhood of

the point (y1 , y2 ) = (1, 1) , (α, β, q) = (1, 1, 2)?
ii) Find the matrix of partial derivatives:
!
∂y1 ∂y1 ∂y1
∂α ∂β ∂q
∂y2 ∂y2 ∂y2
.
∂α ∂β ∂q

4. The directional derivative of f : Rn → R is given by the formula

f (x + hu) − f (x)
lim = ∇f (x) · u.
h→0 h
Show that for all u such that kuk = 1,
∇f (x) · ∇f (x)
∇f (x) · u ≤ ,
k∇f (x)k
thus verifying that the gradient gives the largest directional derivative
for the function.

5. A function f : R → R is said to be of class C ∞ if it has continuous

derivatives of all orders. f is said to be analytic, if it can be expressed
as a limit of sums of power functions, i.e.
∞
X
f (x) = an x n (6)
n=0

14
(i.e. the limit on the right hand side must be well defined). Obviously
all polynomials are analytic functions.
Define the exponential function as follows:
∞
X xn
ex = .
n=0
n!

Show that the series on the right converges for all x and hence the
exponential function is also analytic.
It is a well known theorem for analytic functions that the derivatives
of f can be obtained by differentiating the series of functions term by
term, i.e.
∞
X
0
f (x) = nan xn−1 . (7)
n=0

Use this fact to evaluate the derivative of the exponential function and
show that the exponential function has a well defined inverse called
the logarithmic function. Evaluate the derivative of the logarithmic
function.
Deduce from 6 and 7 Taylor’s theorem for the case where f is analytic:
∞
X f (n) (a)
f (x) = (x − a)n . (8)
n=0
n!

Notice that all analytic functions are in class C ∞ since their derivatives
exist for all orders. Finally, show that there are C ∞ functions that are
not analytic. To achieve this, evaluate the derivatives of the following
function ( 1
e− x2 if x 6= 0,
f (x, y) =
0 if x = 0.
Conclude that 8 fails for a = 0.

15
6. Let
xy 2
(
(x2 +y 4 )
if x 6= 0,
f (x, y) =
0 if x = 0.
Prove that f has well defined finite directional derivatives in all direc-
tions in R2 and prove also that f is not continuous at 0.

16