Extrema of functions of several variables

January 4, 2021

For the beginning we give some notions and results concerning quadratic
forms. Their proofs can be found in the linear algebra courses.

1 Quadratic forms
Definition 1 A function ϕ : Rn → R given by the relation
n
X
ϕ(x) = aij xi xj , ∀x = (x1 , x2 , . . . , xn ) ∈ Rn ,
i,j=1

where aij ∈ R, aij = aji , for every i, j ∈ {1, 2, . . . , n} is called quadratic

form.
The symmetric matrix A = (aij ) ∈ Mn (R) is called the matrix of the
quadratic form ϕ.
Remark that ϕ can be rewritten as
n
X X
ϕ(x) = aii x2i + 2 aij xi xj .
i=1 1≤i<j≤n

Example 1.1 The function ϕ : R3 → R given by the relation

ϕ(x1 , x2 , x3 ) = x21 + 2x22 + 6x23 − 2x1 x2 − 6x2 x3 + 2x3 x1
is a quadratic form with the matrix A,
 
1 −1 1
A = −1 2 −3 .
1 −3 6

1
Definition 2 A quadratic form ϕ : Rn → R is said to be

1. positive definite if ϕ(x) > 0, ∀x ∈ Rn \ {0};

2. positive semidefinite if ϕ(x) ≥ 0, ∀x ∈ Rn \ {0} and there exists

x0 ∈ Rn , x0 6= 0 such that ϕ(x0 ) = 0;

3. negative definite if ϕ(x) < 0, ∀x ∈ Rn \ {0};

4. negative semidefinite if ϕ(x) ≤ 0, ∀x ∈ Rn \ {0} and and there

exists x0 ∈ Rn , x0 6= 0 such that ϕ(x0 ) = 0;

5. indefinite if ϕ(x) < 0, ϕ(y) > 0 for some x, y ∈ Rn .

Example 1.2 Let n = 2. Then

ϕ1 (x1 , x2 ) = x21 + 2x22

is obviously positive definite,

ϕ2 (x1 , x2 ) = x21 − 2x1 x2 + x22

is positive semidefinite since ϕ(x1 , x2 ) ≥ 0 for all (x1 , x2 ) ∈ R2 and ϕ2 (1, 1) =

0. The quadratic form
ϕ3 (x1 , x2 ) = x21 − x22
is indefinite, since ϕ3 (1, 0) = 1 and ϕ3 (0, 1) = −1.

A useful criterion for the evaluation of a quadratic form is given in the

following.
For a quadratic form ϕ : Rn → R of matrix A = (aij ) ∈ Mn (R) let

a11 a12 ··· a1k

a a22 ··· a2k
∆k = 21 , 1 ≤ k ≤ n.
··· ··· ··· ···
ak1 ak2 ··· akk

We denote by λ1 , λ2 , . . . , λn the eigenvalues of A. The eigenvalues of A

are real numbers since A is a symmetric matrix.

2
Theorem 1 (Sylvester) Let ϕ : Rn → R be a quadratic form of matrix
A = (aij ) ∈ Mn (R). Then:

1. ϕ is positive definite ⇔ λ1 > 0, λ2 > 0, . . . , λn > 0 ⇔

∆1 > 0, ∆2 > 0, . . . , ∆n > 0.

2. ϕ is negative definite ⇔ λ1 < 0, λ2 < 0, . . . , λn < 0 ⇔

∆1 < 0, ∆2 > 0, ∆3 < 0, . . . , (−1)n ∆n > 0.

3. ϕ is indefinite ⇔ A has both positive and negative eigenvalues

Remark 1 The quadratic form from Example 1.2 is positive definite since

1 −1 1
1 −1
∆1 = 1, ∆2 = = 1, ∆3 = −1 2 −3 = 1.
−1 2
1 −3 6

One can prove that ϕ is positive definite since ϕ is the sum of three
squares as follows

ϕ(x1 , x2 , x3 ) = (x1 − x2 + x3 )2 + (x2 − 2x3 )2 + x23

Generally a quadratic form can be written as a sum and difference of squares

as follows: first we take all the terms containing x1 and form a square, then
we take all the terms containing x2 and form a square, and so on . . .. At
the end we obtain a sum and difference of squares. This method was given
by Gauss.

Theorem 2 Let ϕ : Rn → R be a positive definite quadratic form. Then

there exists λ > 0 such that

ϕ(x) ≥ λkxk2 , x ∈ Rn .

Proof. For x = 0 the relation is obvious and for x 6= 0 it is equivalent to

 
1 x
ϕ(x) ≥ λ or ϕ ≥ λ.
kxk2 kxk2

Let S(0, 1) = {x ∈ Rn |kxk = 1} be the unit sphere in Rn .

3
 The function ϕ is continuous on S(0, 1), S(0, 1) is a compact subset of
Rn , hence ϕ attains its minimum λ on S(0, 1), i.e., there exists ξ ∈ S(0, 1)
such that ϕ(ξ) = λ.
On the other hand ϕ is positive definite, ξ 6= 0, hence

λ = ϕ(ξ) > 0.
x
Now take x ∈ Rn , x 6= 0. Then kxk ∈ S(0, 1), therefore
 
x
ϕ ≥ λ ⇐⇒ ϕ(x) ≥ kxk2 .
kxk

2 Extrema of functions of several variables

Definition 3 Let f : A → R, A ⊆ Rn be a function of n variables. A point
a ∈ A is called:

1. point of local minimum (maximum) of f if there exists u ∈ V(a)

such that

f (a) ≤ f (x) (f (a) ≥ f (x)), ∀x ∈ U ∩ A.

2. point of global minimum (maximum) of f if

f (a) ≤ f (x) (f (a) ≥ f (x)), ∀x ∈ A.

A point of local minimum (maximum) of f is called point of local ex-

tremum of f . A point of global minimum (maximum) of f is called point
of global extremum of f .

Definition 4 Let A ⊆ Rn be a nonempty set and f : A → R. A point

a ∈ intA with the property that f is differentiable at a and df (a) = 0 is
called a stationary point (critical point).

Theorem 3 (Fermat) Let f : A → R, A ⊆ Rn be differentiable at a ∈

intA. If a is a point of local extremum of f then df (a) = 0.

4
Proof. Suppose that a is a point of local minimum of f . Then there exists
a ball B(a, r) ⊂ A such that f (a) ≤ f (x) for every x ∈ B(a, r).
Let s ∈ Rn , ksk = 1 be fixed. Define the function

g : (−r, r) → R g(t) = f (a + ts), ∀t ∈ (−r, r).

The function g is well defined since

k(a + ts) − ak = ktsk = |t|ksk = |t| < r,

hence a + ts ∈ B(a, r) ⊆ A for every t ∈ (−r, r).

The relation
g(0) ≤ g(t), t ∈ (−r, r)
implies that t = 0 is a point of local minimum for the function of one variable
g. Then g 0 (0) = 0 in view of Fermat theorem for functions of one variable.
We have
g(t) − g(0) f (a + ts) − f (a)
g 0 (0) = lim = lim
t→0 t t→0 t
df
= (a) = df (a)(s) =⇒ df (a)(s) = 0.
ds
Consequently taking s = ek = (0, . . . , 0, 1, 0, . . . , 0) ∈ Rn , 1 ≤ k ≤ n, it
follows fxk (a) = 0, 1 ≤ k ≤ n.

Remark 2 The converse of Fermat’s theorem is not true.

For the function f : R2 → R, f (x, y) = x2 − y 2 , the point (0, 0) is a
stationary point since

df (0, 0) = fx0 (0, 0)dx + fy0 (0, 0)dy = 0

but f (0, 0) = 0, f (0, h) = −h2 ≤ 0, f (h, 0) = h2 ≥ 0 for all h ∈ R, therefore

(0, 0) is not a point of local extremum of f.

Theorem 4 Let A ⊆ Rn be an open set and f ∈ C 2 (A). If a ∈ A is a local

minimum (maximum) of f, then

d2 f (a)(x) ≥ 0 (d2 f (a)(x) ≤ 0), x ∈ A.

5
Proof. Suppose that a is a locl minimum of f. A is an open set, hence there
exists a ball B(a, r) ⊆ A such that f (a) ≤ f (x) for all x ∈ B(a, r). For all
x ∈ B(a, r) there exists ξ ∈ [a, x] such that
1 1
(2.1) f (x) = f (a) + df (a)(x − a) + d2 f (ξ)(x − a)
1! 2!
in view of Taylor’s formula. But df (a) = 0, since f is a local extremum of
f, then the previous relation leads to
1 2
f (x) − f (a) = d f (ξ)(x − a)
2!
1 2 1 2
d f (ξ)(x − a) − d2 f (a)(x − a) ≥ 0


(2.2) = d f (a)(x − a) +
2 2
Denoting x − a = h we obtain
1 1 2
f (a + h) − f (a) = d2 f (a)(h) + d f (ξ)(h) − d2 f (a)(h) ≥ 0


(2.3)
2 2
for all h with khk < r.
Take now x ∈ Rn , x 6= 0 and let t ∈ R such that
r
ktxk < r ⇐⇒ |t| < .
kxk
By the relation (2.3) for h = tx we obtain
1 2 1 2
d f (ξt )(tx) − d2 f (a)(tx) ≥ 0


d f (a)(tx) +
2 2
r
for all t, |t| < kxk and consequently

t2 2 r
d f (a)(x) + d2 f (ξt )(x) − d2 f (a)(x) ≥ 0,



|t| < ,
2 kxk
or
r
d2 f (a)(x) + d2 f (ξt )(x) − d2 f (a)(x) ≥ 0,


|t| < .
kxk
Letting t → 0 in the last relation it follows
d2 f (a)(x) ≥ 0
in view of the continuity of d2 f (a).
In the next theorem we give sufficient conditions for a stationary point
to be a local extremum.

6
Theorem 5 Let A ⊆ Rn be an open set and f ∈ C 2 (A) and a ∈ A be a
stationary point of f . Then:

1. If d2 f (a) is positive definite, then a is a point of local minimum of f .

2. If d2 f (a) is negative definite, then a is a point of local maximum of f .

3. If d2 f (a) is indefinite, then a is not a point of local extremum of f .

Proof.

1. Suppose that d2 f (a) is positive definite. Taking account of Theorem

2 it follows that there exists λ > 0 such that

d2 f (a)(x − a) ≥ λkx − ak2 , ∀x ∈ Rn .

Take B(a, r) ⊆ A. For every x ∈ B(a, r) taking account of Taylor’s

formula, we find ξ ∈ [a, x] such that
1 1
f (x) = f (a) + df (a)(x − a) + d2 f (ξ)(x − a)
1! 2!
1
= f (a) + d2 f (ξ)(x − a).
2
We obtain
1 2 1 2
d f (ξ)(x − a) − d2 f (a)(x − a)


f (x) − f (a) = d f (a)(x − a) +
2 2
n 
1 X ∂ 2 f (ξ) ∂ 2 f (a)

1 2
= d f (a)(x − a) + − (xi − ai )(xj − aj )
2 2 ∂xi ∂xj ∂xi ∂xj
i,j=1
n 
kx − ak2 X ∂ 2 f (ξ) ∂ 2 f (a) (xi − ai )(xj − aj )

1 2
= d f (a)(x − a) + + −
2 2 ∂xi ∂xj ∂xi ∂xj kx − ak2
i,j=1

where a = (a1 , a2 , . . . , an ), x = (x1 , x2 , . . . , xn ).

Denoting
n  2
∂ 2 f (a) (xi − ai )(xj − aj )

X ∂ f (ξ)
α(x) = −
∂xi ∂xj ∂xi ∂xj kx − ak2
i,j=1

7
 for x ∈ B(a, r) \ {a}, we have lim α(x) = 0, since f ∈ C 2 (A) and
x→a

|(xi − ai )(xj − aj )| 1 (xi − ai )2 + (xj − aj )2 1

2
≤ · 2
≤ .
kx − ak 2 kx − ak 2
It follows
1 2 kx − ak2
f (x) − f (a) = d f (a)(x − a) + α(x)
2 2
kx − ak2
≥ α(x)(λ + α(x))
2
for all x ∈ B(a, r) \ {a}. Taking account of lim α(x) = 0 we find a
x→a
neighbourhood V ∈ V(a) such that

λ + α(x) > 0, ∀x ∈ V ∩ B(a, r),

therefore f (x) − f (a) ≥ 0 for all x ∈ V ∩ B(a, r). It follows that a is a

local minimum of f .
Analogously one prove 2. and 3.

Corollary 2.1 Let D ⊆ R2 be an open set and f ∈ C 2 (D) and a = (x0 , y0 )

be a stationary point of f . Denote:

∂ 2 f (a) ∂ 2 f (a) ∂ 2 f (a)

A= , B= , C= .
∂x2 ∂x∂y ∂y 2
Then:
1. If B 2 − AC < 0 and A > 0 =⇒ a is a point of local minimum of f ;

2. If B 2 − AC < 0 and A < 0 =⇒ a is a point of local maximum of f ;

3. If B 2 − AC > 0 =⇒ a is not a point of local extremum of f .

Proof. We have

d2 f (a) = Adx2 + 2Bdxdy + Cdy 2

and we apply Theorem 5

8
Remark 3 If B 2 − AC = 0 the Corollary 2.1 gives no answer. The sign of
f (x, y) − f (x0 , y0 ) depends on the differentials of higher order.

Remark 4 If K ⊆ Rn is compact set and f ∈ C 2 (K), then the global

extrema of f are attained on K, in view of Weierstrass theorem. If they
belong to intK, then there are stationary points of f and can be find using
Theorem 5. If not, they belong to the boundary of K and the theory presented
in this paragraph is not efficient.

Example 2.1 Find the local extrema of the function f : R2 → R,

f (x, y) = f (x, y) = x4 + y 4 − 2x2 + 4xy − 2y 2 .

Solution 1 The stationary points of f are the solutions of the system of

equations  0
fx (x, y) = 4x3 − 4(x − y) = 0
fy0 (x, y) = 4y 3 + 4(x − y) = 0.
Adding the above relations one obtains x3 + y 3 = 0 =⇒ y = −x. √ √
Consequently,
√ √ we get the following stationary points (0, 0), ( 2, − 2)
and (− 2, 2).
The second order partial derivatives of f are

• fx002 (x, y) = 12x2 − 4,

• fy002 (x, y) = 12y 2 − 4,

9
 00 (x, y) = 4,
• fxy

therefore the second differential of f is given by the relation

d2 f (x, y) = (12x2 − 4)dx2 + (12y 2 − 4)dy 2 + 8dxdy.

Next, we √determine
√ the nature of√each √stationary point.
2
For ( 2, − 2) we have √ d √f ( 2, − 2) = 20dx2 + 20dy 2 − 8dxdy, it is
positive definite,√hence√ ( 2, − 2) is a point of local minimum.
Similarly (− 2, 2) is a point of local minimum of f.
For (0, 0) d2 f (0, 0) = −4dx2 + 8dy 2 − 4dxdy, it is negative semidefinite,
therefore we study the sign of f in a neighbourhood of the origin. We have:

• f (0, 0) = 0;

• f (x, x) = 2x4 ≥ 0, ∀x ∈ R, and

√ √
• f (x, 0) = x4 − 2x2 = x2 (x2 − 2) < 0, ∀x ∈ ( 2, − 2), x 6= 0.

=⇒ (0, 0) is not a point of local extremum.

10