Calculus on Rn

Seminar 9 2020

1 Conditioned Extremum Points

Let us consider the following framework


 n<m



A ⊆ Rn








f :A→R




F : A → Rm








C = {x ∈ A : F (x) = 0m }


A point c ∈ C is said to be a conditioned extremum point of f with

respect to the set C , if it is an extremum point for the restriction of f to
C.

The Lagrange multipliers rule

In addition to the framework from the beginning, consider that

A is nonempty and open

and
F = (F1 , ..., Fm ) : A → Rm is a C 1 function .
If c is a condition extremum of f with respect to C, and

rankJ(F )(c) = m,

then, there exists λ0 ∈ Rm such that (c, λ0 ) ∈ A × Rm is an extremum point of

the function
L : A × Rm → R
L(x, λ) = f (x) + hλ, F (x)i = f (x) + λ1 F1 (x) + ... + λm Fm (x).

Here L is called the Lagrange function associated to f , and λ01 , ...λ0m are the
Lagrange multipliers.
In practice, we are interested in point 2, from Theorem 2.21.5 from the lecture notes,
which says that: If c ∈ C and there exists λ0 ∈ Rm such that

dL(c, λ0 ) = 0 and d2 L0 (c, h) > 0 (respectively < 0)

1
 for all h ∈ Rn \{0n } such that dF (c)(h) = 0M , where

L0 (x) = f (x) + hλ0 , F (x)i,

then
c is a conditioned local minimum (respectively maximum)
point o f w.r.t. C.

Homework 1: Solve the Exercise 7,9 form Seminar8-ttrif

2 Local Extremum Points for Twice Differentiable

Functions
Algorithm for determining the local extremum points of a C 2
function
6 A ⊆ Rn an open set, and let f : A → R be a C 2 function on A.
Let ∅ =

Step 1 Determine the first order partial derivatives of f

Step 2 Determine the critical/stationary points of f , i.e all the points a ∈ A s.t.
∂f
(a) = 0, ∀j ∈ {1, ..., n}.
∂xj

More precisely, determine the set

C = {c ∈ A : ∇f (c) = 0n }
If f has no critical points, (i.e C = ∅) then f has no local extremum points.
If f has some critical points, (i.e. C 6= ∅) GO TO Step 3

Step 3 Determine all the second-order partial derivatives of f at a random point in

a ∈ A. Construct the Hessian matrix attached to the function f at the point a,
denoted by Hf (a), and defined by:

∂2f ∂2f ∂2f

 
∂x21
(a) ∂x2 ∂x1 (a) ... ∂xn ∂x1 (a)
 
 
2 2 2
∂ f ∂ f ∂ f
 

∂x1 ∂x2 (a) ∂x22
(a) ... ∂xn ∂x2 (a) 
Hf (a) = 



 

 ... ... ... ... 

∂2f ∂2f ∂2f
∂x1 ∂xn (a) ∂x2 ∂xn (a) ... ∂x2 (a) n

Step 4 For every critical point c ∈ C, analyze the Hessian matrix, hence, analyze Hf (c)

2
 – If Hf (c) is POSITIVE DEFINITE, then c is a local minimum point of
f.
– If Hf (c) is NEGATIVE DEFINITE, then c is a local maximum point
of f .
– If Hf (c) is INDEFINITE, then c is not a local extremum point of f , case
in which it is called a a saddle point.

• Recall to repeat Step 4 for each critical point c ∈ C

Now, the natural question arising is concerning the nature of the Hessian matrix. We
use Sylvester’s Theorem applied to the quadratic matrix Hf (c). In order to formulate
this theorem, we make the following notations for the determinants for the minors (on
the first diagonal) of this matrix. Thus, for

∂2f ∂2f ∂2f

 
∂x21
(a) ∂x2 ∂x1 (a) ... ∂xn ∂x1 (a)
 
 
∂2f ∂2f ∂2f
 

∂x1 ∂x2 (a) ∂x2 2 (a) ... ∂xn ∂x2 (a) 
Hf (a) = 



 

 ... ... ... ... 

∂2f ∂2f ∂2f
∂x1 ∂xn (a) ∂x2 ∂xn (a) ... ∂x2 (a) n

 ∂2f ∂2f

2 (a) (a)
 2 
∂ f  
∂x2 ∂x1

∆1 =  2 (a) , ∆2 =  ∂∂x2 f1  ...
 
∂2f
∂x1  ∂x ∂x (a)
1 2 ∂x22
(a) 
 2 2

 ∂ f (a) ∂ f ∂2f
 ∂x21 ∂x2 ∂x1
(a) ... ∂xn ∂x1
(a) 

 
 
 ∂2f ∂2f ∂2f

 (a) ∂x22
(a) ... (a) 
∆n =  ∂x1 ∂x2 ∂xn ∂x2 

 

 ... ... ... ... 

 ∂2f ∂2f 2
∂ f
 ∂x1 ∂xn (a) ∂x2 ∂xn (a) ... (a)

∂x2

n

Sylvester’s Theorem:
A quadratic matrix of the form Hf (a) is

• POSITIVE DEFINITE if ∆k > 0, for all k ∈ {1, ..., n}

• NEGATIVE DEFINITE if (−1)K ∆k > 0, for all k ∈ {1, ..., n}

• INDEFINITE if we are not in the two cases above, and still ∆k > 0 for
all k ∈ {1, ..., n}

3
 This theorem can be used in most of the cases, but it cannot be applies when some
of the determinants are 0.
For the case when there exists k ∈ {1, ..., n} such that ∆k = 0, we have to consider
the second order differential function attached to f at the point a, which is

d2 f (a) : Rn → R

given by
d2 f (a)(h) = hT · Hf (a) · h, ∀h ∈ Rn ,
which, if you do the matrix multiplication turns into
n X n
X ∂ 2f
2
d f (a)(h) = (a) · hi · hj , ∀h = (h1 , ..., hn ) ∈ Rn .
i=1 j=1
∂x i ∂x j

If we may emphasize two particular points u, v ∈ Rn such that

d2 f (a)(u) < 0 < d2 f (a)(v)

this means that the Hf (a) is indefinite, therefore, at Step 4 of the algorithm, the
conclusion is that
a is a saddle point,
hence it is not a local extremum point.
Hf (a) is said to be semidefinite if

d2 f (a)(h) ≤ 0

or ≥ 0 for all h ∈ Rn (it is not stict). In this case the nature of the point a has to be
analyzed by going back to the original form of the function (see some of the following
examples )

Homework 2: Solve all of the points of Exercise 10 form Seminar8-

ttrif
Moreover, consider the following briefly solved examples

Example 1:Determine all local extrema, their type (minima or

maxima) and the corresponding extreme values of the following
functions:
a) f : R3 → R, f (x, y, z) = 2x2 −xy+2xz−y+y 3 +z 2 ; b) f : R2 → R,
f (x, y) = x2 + y 2 (1 − x)3 .
Solution
a) f : R3 → R, f (x, y, z) = 2x2 − xy + 2xz − y + y 3 + z 2 .

4
 For all (x, y, z) ∈ R3 it holds ∂f
∂x
(x, y, z) = 4x − y + 2z, ∂f
∂y
(x, y, z) = −x − 1 + 3y 2 and
∂f
∂z
(x, y, z) = 2x + 2z. The stationary points of f are the solutions of the system

 4x − y + 2z = 0
−x − 1 + 3y 2 = 0
2x + 2z = 0.


From the last equation it follows that x = −z. By replacing this in the first equaiton
we get y = 2x, equality which, together with the second equation, provides us with
12x2 − x − 1 = 0. Thus x ∈ {− 41 , 13 }. It follows that (− 14 , − 21 , 14 ) and ( 13 , 23 , − 31 ) are the
stationary points of the function.
For all (x, y, z) ∈ R3 it holds

∂ 2f ∂ 2f ∂ 2f
(x, y, z) = 4, (x, y, z) = 6y, (x, y, z) = 2,
∂x2 ∂y 2 ∂z 2

∂ 2f ∂ 2f ∂ 2f
(x, y, z) = −1, (x, y, z) = 2, (x, y, z) = 0,
∂x∂y ∂x∂z ∂y∂z
thus
   
  4 −1 2   4 −1 2
1 1 1 1 2 1
H(f ) − , − , =  −1 −3 0  , H(f ) , ,− =  −1 4 0 .
4 2 4 3 3 3
2 0 2 2 0 2

The matrix H(f ) 13 , 23 , − 13 is positive definite, thus ( 13 , 32 , − 13 ) is a local minimum point.

Moreover
1 2 1 −13
f( , , − ) = .
3 3 3 27
Since Sylvester’s criterion cannot be applied to the matrix H(f ) − 41 , − 12 , 14 , we will

study its nature by other means. The quadratic from associated to this matrix is

Φ(h1 , h2 , h3 ) = 4(h1 )2 − 2h1 h2 + 4h1 h3 − 3(h2 )2 + 2(h3 )2 .

Since Φ(0, 1, 0) = −3 < 0 and Φ(0, 0, 1) = 2 > 0, it follos that this quadratic form (and
its corresponding matrix) is indefinite. Therefore, (− 41 , − 21 , 14 ) is not a local extremum, it
is a saddle point.

b) f : R2 → R, f (x, y) = x2 + y 2 (1 − x)3 .
For all (x, y) ∈ R2 it holds ∂f
∂x
(x, y) = 2x − 3y 2 (1 − x)2 and ∂f
∂y
(x, y) = 2y(1 − x)3 . The
stationary points of f are the solutions of the system

2x − 3y 2 (1 − x)2 = 0
2y(1 − x)3 = 0.

From the last equation we get y = 0 or x = 1. If y = 0, then from the first equation of the
system we get x = 0. We notice that x = 1 does not obey the first equation of the system,
thus it is excluded. In conclusion (0, 0) is the only stationary point of the function.

5
 For all (x, y) ∈ R2 it holds

∂ 2f ∂ 2f ∂ 2f
(x, y) = 2 + 6y 2 (1 − x), (x, y) = −6y(1 − x)2 , (x, y) = 2(1 − x)3 ,
∂x2 ∂x∂y ∂y 2
thus  
2 0
H(f )(0, 0) = .
0 2
Since 2 > 0 and 4 > 0, it follows that H(f )(0, 0) is positive definite, thus (0, 0) is a local
minimum point. Moreover
f (0, 0) = 0.

Example 2:
Determine all local extrema, their type (minima or maxima) and
the corresponding extreme values of the following functions:

a) f : R3 → R, f (x, y, z) = x3 − 3x + y 2 + z 2 ;
b) f : R2 → R f (x, y) = x4 + y 4 − 4(x − y)2 ;
c) f : R3 → R f (x, y, z) = z 2 (1 + xy) + xy;
d) f : R2 → R f (x, y) = x3 + 3xy 2 − 15x − 12y.

Solution

a) f : R3 → R, f (x, y, z) = x3 − 3x + y 2 + z 2
Step 1: The critical/stationary points For all (x, y, z) ∈ R3 it holds ∂f ∂x
(x, y, z) =
2 ∂f ∂f
3x −3, ∂y (x, y, z) = 2y and ∂z (x, y, z) = 2z. The critical/stationary points of the function
f are the solutions of the system
 2
 3x − 3 = 0
2y = 0
2z = 0.


Therefore (−1, 0, 0) and (1, 0, 0) are the only stationary points of the function. The set
of the critical points is
C = {(−1, 0, 0), (1, 0, 0)}
For all (x, y, z) ∈ R3 it holds

∂ 2f ∂ 2f ∂ 2f
(x, y, z) = 6x, (x, y, z) = 2, (x, y, z) = 2,
∂x2 ∂y 2 ∂z 2

∂ 2f ∂ 2f ∂ 2f
(x, y, z) = 0, (x, y, z) = 0, (x, y, z) = 0,
∂x∂y ∂x∂z ∂y∂z

6
thus    
−6 0 0 6 0 0
H(f )(−1, 0, 0) =  0 2 0  , H(f )(1, 0, 0) =  0 2 0  .
0 0 2 0 0 2
The matrix H(f )(1, 0, 0) positive definite, thus (1, 0, 0) is a local minimum point of the
function.
Since Sylvester’s criterion cannot be applied to the matrix H(f )(−1, 0, 0), we will
study its nation by other means. The quadratic form associated to this matrix is

Φ(h1 , h2 , h3 ) = −6(h1 )2 + 2(h2 )2 + 2(h3 )2 .

Since Φ(1, 0, 0) = −6 < 0 and Φ(0, 1, 0) = 2 > 0, it follows that this quadratic form (and
therefore its associated matrix) is indefinite. Thus (−1, 0, 0) is not a local extrema, it is
a saddle point.

b) f : R2 → R, f (x, y) = x4 + y 4 − 4(x − y)2

For all (x, y) ∈ R2 it holds ∂f
∂x
(x, y) = 4x3 − 8(x − y) and ∂f
∂y
(x, y) = 4y 3 + 8(x − y).
The stationary points of the function f are the solutions of the system
 3
4x − 8(x − y) = 0
4y 3 + 8(x − y) = 0.

By adding the two equations we get x3 = −y 3 , thus x = −y. By replacing this in the first
equation e get x3 − 4x = 0, thus x ∈ {−2, 0, 2}. Therefore the stationary points of the
function are (−2, 2), (0, 0) and (2, −2).
For all (x, y) ∈ R2 it holds

∂ 2f ∂ 2f ∂ 2f
2
(x, y) = 12x2 − 8, (x, y) = 8, 2
(x, y) = 12y 2 − 8.
∂x ∂x∂y ∂y
For the three stationary points we get
   
40 8 −8 8
H(f )(−2, 2) = H(f )(2, −2) = and H(f )(0, 0) = .
8 40 8 −8

Since the matrix H(f )(−2, 2) = H(f )(2, −2) is positive definite the points (−2, 2) and
(2, −2) are local minimum points. Then

f (−2, 2) = −32 and f (2, −2) = −32.

Due to the fact that the determinant of the matrix H(f )(0, 0) is zero, the algorithm
for determining the stationary points provides us with no useful information connected
to the stationary point (0, 0). Nevertheless we notice that f (0, 0) = 0, f (x, x) = 2x4 ,
f (x, 0) = x2 (x2 − 4). It follows that in each neighnborhood of (0, 0) the function f takes
both positive and negative values (for example, for all n ∈ N it holds f ( n1 , n1 ) > 0 and
f ( n1 , 0) < 0). Therefore (0, 0) is not a local extremum, it is a saddle point.

c) f : R3 → R, f (x, y, z) = z 2 (1 + xy) + xy;

7
 For all (x, y, z) ∈ R3 it holds ∂f∂x
(x, y, z) = yz 2 + y, ∂f
∂y
(x, y, z) = xz 2 + x and
∂f
∂z
(x, y, z) = 2z(1 + xy). The stationary points of f are the solutions of the system

 y(z 2 + 1) = 0
x(z 2 + 1) = 0
2z(1 + xy) = 0.


From the first two equations it follows that y = 0 andx = 0, respectively. By replacing
these values in the last equation we reach the conclusion that z = 0. Hence (0, 0, 0) is the
only stationary point of the function.
For all (x, y, z) ∈ R3 it holds
∂ 2f ∂ 2f ∂ 2f
(x, y, z) = 0, (x, y, z) = 0, (x, y, z) = 2(1 + xy),
∂x2 ∂y 2 ∂z 2
∂ 2f ∂ 2f ∂ 2f
(x, y, z) = z 2 + 1, (x, y, z) = 2yz, (x, y, z) = 2xz,
∂x∂y ∂x∂z ∂y∂z
thus  
0 1 0
H(f )(0, 0, 0) =  1 0 0  .
0 0 2
Since Sylvester’s criterion cannot be applied to the matrix H(f )(0, 0, 0), we study its
nature by other means. Its associated quadratic form is

Φ(h1 , h2 , h3 ) = 2h1 h2 + 2(h3 )2 .

Since Φ(−1, 1, 0) = −2 < 0 and Φ(0, 0, 1) = 2 > 0, it follows that this quadratic form
(therefore its corresponding matrix) is indefinite. Thus (0, 0, 0) is not a local extremum
point of the function, it is a saddle point.

d) f : R2 → R, f (x, y) = x3 + 3xy 2 − 15x − 12y.

For all (x, y) ∈ R2 it holds ∂f
∂x
(x, y) = 3x2 + 3y 2 − 15 and ∂f
∂y
(x, y) = 6xy − 12. The
stationary points of f are the solutions of the system
 2
3x + 3y 2 − 15 = 0
6xy − 12 = 0.

It follows that x2 + y 2 = 5 and xy = 2. As x2 + y 2 = (x + y)2 − 2xy it follows that

(x + y)2 = 9, which means that x + y ∈ {−3, 3}. Therefore x and y are the solutions of
one of the equations t2 + 3t + 2 = 0 or t2 − 3t + 2 = 0. Thus the stationary point of the
functions are (−2, −1), (−1, −2), (1, 2) and (2, 1).
For all (x, y) ∈ R2 it holds
∂ 2f ∂ 2f ∂ 2f
(x, y) = 6x, (x, y) = 6y, (x, y) = 6x.
∂x2 ∂x∂y ∂y 2
For the four stationary points of the function we get
   
−2 −1 −1 −2
H(f )(−2, −1) = 6 , H(f )(−1, −2) = 6 ,
−1 −2 −2 −1

8
    
1 2 2 1
H(f )(1, 2) = 6 , H(f )(2, 1) = 6 .
2 1 1 2
The matrix H(f )(−2, −1) is negative definite, H(f )(2, 1) is positive definite while
(f )(−1, −2) ane H(f )(1, 2) are indefinite. Thus (−2, −1) is a local maximum point of f ,
and (2, 1) is a local minimum, while (−1, −2) and (1, 2) are not local extremum point,
they are saddle points. Moreover

f (−2, −1) = 18 and f (2, 1) = −18.