Chapter 1

Fixed point theorems

One of the most important instrument to treat (nonlinear) problems with the aid of
functional analytic methods is the fixed point approach. This approach is an important
part of nonlinear (functional-)analysis and is deeply connected to geometric methods of
topology. We consider in this chapter the famous theorems of Banach, Brouwer and
Schauder. A more detailed description of the fixed point theory can be found for instance
in [1, 2, 6, 8, 9] dar.

1.1 The fixed point approach – preliminaries

Consider the following fixed point equation

x = F(x) (1.1)

Here, F is a mapping from X −→ X . We assume that X is endowed with the metric d .

A point z ∈ X which satisfies z = F(z) is called a fixed point of F. Fixed point theo-
rems guarantee the existence and/or uniqueness when F and X satisfy certain additional
conditions. A simple example of a mapping F which doesn’t posses a fixed point is the
translation in a vector space X :

F : X � x �−→ x + x0 ∈ X where x0 �= θ .

Many problems of applied mathematics may be formulated as a fixed point equation or

reformulated in this way: Linear equations, differential equations, zeroes of gradients, . . . .
The following examples show how the treatment of nonlinear problems may be translated
into fixed point problems:
• Determination of zeros of nonlinear functions g : R −→ R:

g(x) = 0

There are different possibilities to translate this equation into a fixed point problem:

F(x) := x − g(x) simplest method

F(x) := x − ωg(x) linear relaxation with relaxation parameter ω > 0
F(x) := x − (g � (x))−1 g(x) Newton method

1
 • Consider the determination of a zero z of a function f : Rn −→ Rn . Clearly, such a
zero is a fixed point of the mapping F : Rn � x �−→ x + f(x) ∈ Rn . The formulation
of equations as a fixed point equation is possible in various ways. Consider

f : R � x �−→ x3 − 13x + 18 ∈ R .

z is a zero of f iff z is a fixed point of gi , i = 1, 2, 3, 4, where

3
g1 (x) := x3 − 12x + 18 g2 (x) := x + 18
13
g4 (x) := 13x − 18
1
g3 (x) := (13x − 18) 3
x2

• Ordinary differential equations

y � = f(t, y) , y(t0 ) = y0 (1.2)

If the righthandside f is continuous on an appropriate domain Q ⊂ R × Rn then this

initial value problem is equivalent to the following integral equation of fixed point
character: �t
0
y(t) = F(y)(t) := y + f(s, y(s))ds
y0

• (Nonlinear) partial differential equations:

−Δu = f(u), in Ω , u = θ in ∂Ω

f is given and the solution u has to be found. Under certain conditions one can
define the inverse (−Δ)−1 of Δ on spaces of functions u which vanish on ∂Ω . Then
we obtain the following fixed point equation:

u = F(u) := (−Δ)−1 (f(u))

Suppose we have a mapping F : X −→ X as given above. We want to consider

methods to compute such a fixed point. Here are some preliminary remarks. We define

F0 := id , Fn+1 := F ◦ Fn , n ∈ N0 .

Choosing x ∈ X, the sequence (Fn (x))n∈N is called the orbit of x under F . This orbit is
the result of the iteration

xn+1 := F(xn ) , n ∈ N0 ; x0 := x . (1.3)

This iteration process is in the focus of the following chapters. The main question which
we will analyze is under which assumptions (concerning F and X) the orbit converges.

2
1.2 Fixed point theorem of Banach
Consider the fixed point equation (1.1) again. The essential step to prove the existence
of a solution of this equation via the iteration
xn+1 := F(xn ) , n ∈ N0 ; x0 := x , (1.4)
is to prove that the orbit (Fn (x))n∈N is a Cauchy sequence (for every x ∈ X). The
assumption which guarantees this is the assumption that F is Lipschitz continuous with
Lipschitz constant c satisfying c < 1 . This means
There exists c ∈ [0, 1) such that d(F(x), F(y)) ≤ cd(x, y) for all x, y ∈ X . (1.5)
If (1.5) is satisfyied, F is called a contraction. Notice that cn is the Lipschitz constant
of Fn .
Lemma 1.1. Let (X, d) be a metric space and let F : X −→ X be a contraction. If
c ∈ [0, 1) is the Lipschitz constant of F, then we have
1
d(x, y) ≤ (d(x, F(x)) + d(y, F(y))) , x, y ∈ X . (1.6)
1−c
Proof:
Let x, y ∈ X . Using the triangle inequality
d(x, y) ≤ d(x, F(x)) + d(F(x), F(y)) + d(F(y), y)
and using d(F(x), F(y)) ≤ cd(x, y) the result follows. �

(1.6) is a very strange inequality: it says that we can estimate how far apart any two
points x, y are just from knowing how far x is from its image F(x) and how far y is from
its image F(y) . A first application is
Corollary 1.2. Let (X, d) be a metric space and let F : X −→ X be a contraction. Then
F has at most one fixed point.
Proof:
Suppose that x, y are fixed poits of F . Then d(x, F(x)), d(y, F(y)) are zero, so by the
inequality (1.6) d(x, y) is zero too. �
Lemma 1.3. Let (X, d) be a metric space and let F : X −→ X be a contraction. Then
the orbit (Fn (x))n∈N0 is a Cauchy sequence.
Proof:
Let c ∈ [0, 1) be the Lipschitz constant of F . From the inequality (1.6) we read off
1
d(Fn (x), Fm (x)) ≤ (d(Fn (x), Fn+1 (x)) + d(Fm (x), Fm+1 (x))) , m, n ∈ N .
1−c
Recalling that ck is the Lipschitz constant of Fk we get
cn + c m
d(Fn (x), Fm (x)) ≤ d(x, F(x)) , m, n ∈ N .
1−c
Since limk ck = 0 we obtain limm,n→∞ d(Fn (x), Fm (x)) = 0 . �

3
Theorem 1.4 (Contraction Theorem, Banach 1922). Let (X, d) be a complete metric
space and let F : X −→ X be a contraction. If c ∈ [0, 1) is the Lipschitz constant of F,
then the following assertions hold:

(a) There exists a uniquely determined fixed point z of F .

(b) The iteration

xn+1 := F(xn ) , n ∈ N0 , x0 := x , (1.7)
defines a sequence (xn )n∈N0 which converges to z for all x ∈ X .
n
(c) d(xn , z) ≤ c d(F(x), x) , n ∈ N0 .
1−c
(d) d(xn , z) ≤ c d(xn , xn−1 ) , n ∈ N .
1−c
(e) d(xn , z) ≤ c d(xn−1 , z) , n ∈ N .

Proof:
Ad (a), (b)
The uniqueness of a fixed point follows from Corollary 1.2. Let x ∈ X . From Lemma (1.3)
we obtain that the orbit (Fn (x))n∈N0 converges to a point z since (X, d) is complete. Since
F is continuous due the fact that F is a contraction – see (1.5) – z is a fixed point.
Ad (c)
Let x ∈ X . We may prove with the help of (1.5) inductively

d(Fn+1 (x), Fn (x)) ≤ cn d(F(x), x) . (1.8)

Then with (1.8)

m
� m
�
m n j j−1
d(F (x), F (x)) ≤ d(F (x), F (x)) ≤ cj−1 d(F(x), x) (1.9)
j=n+1 j=n+1
m−n
1−c
≤ cn d(F(x), x) , m ≥ n . (1.10)
1−c
Since c ∈ [0, 1) we obtain the assertion (c) by going to the limit m → ∞ in (1.9),(1.10) .
We obtain from (1.8)

1 − cm−1
m
d(F (x), F(x)) ≤ c d(Fn (x), Fn−1 (x)) , m ∈ N . (1.11)
1−c
(d), (e) are simple consequences of these inequalities. �

Theorem 1.5 (Banach’s fixed point theorem). Let (X, � · �) be a Banach space, let V be
a closed subset of X and let F : V −→ V be a contraction, that is

∃ c ∈ [0, 1) ∀ x, y ∈ V (d(F(x), F(y)) ≤ cd(x, y)) . (1.12)

Then the following statements hold:

a) There exists a uniquely determined fixed point z of F in V .

4
 b) The successive iteration

xn+1 := F(xn ) = Fn+1 (x) , n ∈ N0 , x0 = x , (1.13)

defines a sequence (xn )n∈N0 which converges to z for all x ∈ V .

n
c) �xn − z� ≤ c �F(x) − x� , n ∈ N .
1−c
(d) �xn − z� ≤ c �xn − xn−1 � , n ∈ N .
1−c
(e) �xn − z� ≤ c �xn−1 − z� , n ∈ N .
Proof:
Consider the metric space (V, d) with d(x, y) := �x−y� , x, y ∈ V . This space is complete
due to the closedness of V . Now apply Theorem 1.4. �

Banach’s fixed point theorem is a perfect theorem“. It formulates an easy to check

”
assumption under which we have existence and uniqueness and computatibilty of a fixed
point is guarateed. Moreover, it presents the order of the convergence of the approximating
sequence and gives hints to stop the iteration in finite steps; see below.
Remark 1.6. The application of the fixed point theorem of Banach to an initial value
problem for ordinary differential equations leads to an existence and uniqueness result,
the so called Picard–Lindelöff theorem; see [3]. The assumption we need to apply
Theorem 1.5 is that the righthand side f in (1.2) is continuous and Lipschitz continuous
with respect to y uniformly in t . �

In many applications, the mapping F in a fixed point equation is dependent on a

parameter p ∈ P :
x = Fp (x) , x ∈ X, p ∈ P . (1.14)
In this situation we have
Corollary 1.7. Let (X, d) be a complete metric space, let (P, d � ) be a metric space and
let Fp : X −→ X be a contraction for each p ∈ P . Let cp ∈ [0, 1) be the Lipschitz constant
of Fp . We assume cp ≤ c < 1, p ∈ P . Moreover we assume, that there exists p0 ∈ P with
limp→p0 Fp (x) = Fp0 (x) for all x ∈ X . Then the equation (1.14) posesses for each p ∈ P a
uniquely determined solution xp ∈ X and we have limp→p0 xp = xp0 .
Proof:
Let xp ∈ P be the uniquely determined solution of (1.14). This solution exists due to
Theorem 1.4. Then

d(xp , xp0 ) = d(Fp (xp ), Fp0 (xp0 ))

≤ d(Fp (xp ), Fp (xp0 )) + d(Fp (xp0 ), Fp0 (xp0 ))
≤ cd(xp , xp0 ) + d(Fp (xp0 ), Fp0 (xp0 ))

and we conclude
1
d(xp , xp0 ) ≤d(Fp (xp0 ), Fp0 (xp0 )) .
1−c
By taking the limit p → p0 we obtain limp→p0 xp = xp0 . �

5
 1
Example 1.8. X = R , � · � := | · | , M := [0, ∞), F : M � t �−→ t + ∈ M . We
1+t
have
|F(t) − F(s)| < |t − s| .
F posesses no fixed point! Notice that

|F(t) − F(s)| ≤ c|t − s| , t, s ∈ M ,

for no c ∈ [0, 1) is possible. �

Corollary 1.9. Let X be a Banach space and let F : Br −→ X be a contraction, i. e.

∃ L ∈ [0, 1) ∀ x, y ∈ Br (�F(x) − F(y)� ≤ L�x − y�) . (1.15)

If F(∂Br ) ⊂ Br then F has a uniquely determined fixed point in Br .

Proof:
Define
1
G : Br � x �−→ (x + F(x)) ∈ X .
2
Let x ∈ Br , x �= θ , and consider u := rx�x�−1 ∈ ∂Br ⊂ Br . Then we obtain

�F(x) − F(u)� ≤ L�x − u� = L(r − �x�)

and hence

�F(x)� ≤ �F(u)� + �F(x) − F(u)� ≤ r + L(r − �x�) ≤ 2r − �x�

and
1 1
�G(x)� = �x + F(x)� ≤ (�x� + �F(x)�) ≤ r .
2 2
Due to the continuity of G we obtain �G(θ)� ≤ r .
This shows that G : Br −→ Br and G is a contraction too as we conclude from
1 1
�G(x) − G(y)� ≤ (�x − y� + L�x − y�) = (1 + L)�x − y� , x, y ∈ Br .
2 2
Theorem 1.5 implies that G has a fixed point x ∈ Br . This implies x = F(x) . Due to the
fact that F is a contraction this fixed point is uniquely determined. �

Remark 1.10. What is a kind of the inverse of the contraction fixed point theorem? The
most elegant result in this direction is due to Bessaga [5].
Suppose that X is an arbitrary nonempty set and suppose that F : X −→ X has
the property that F and each of its iterates T n has a unique fixed point. Then for each
c ∈ (0, 1) there exists a metric dc on each X such that (X, dc ) is complete and for which

dc (F(x), F(y)) ≤ cdc (x, y) for all x, y ∈ X .

6
Theorem 1.11 (Edelstein, 1961). Let X be a Banach space and let V ⊂ X . Suppose that
F : V −→ V is contractive, i. e.

∀ x, y ∈ V, x �= y, (�F(x) − F(y)� < �x − y�) . (1.16)

Then we have

a) F has at most a fixed point in V .

b) If V is compact then F posesses exactly one fixed point.

Proof:
Ad a) Obvious.
Ad b) Consider the mapping V � v �−→ �v − F(v)� ∈ R . Since V is compact and since
the norm is continuous there exists x ∈ V with �x − F(x)� = inf v∈V �v − F(v)� . Now we
must have x = F(x) because otherwise we would have

�F(x) − (F ◦ F)(x)� < �x − F(x)� .

Remark 1.12. The completeness that we assume in Theorem 1.5 and in the following
results is more or less necessary for the existence of fixed points; see for instance [7, 11,
10, 14, 15]. �

Remark 1.13. An application of the contraction theorem of Banach is the theorem of the
inverse mapping; see for instance [4]. �

1.3 Successive approximation for a contraction

Consider
xn+1 := F(xn ) , n ∈ N0 ; x0 := x , (1.17)
with a contraction F in a metric space (X, d) . From a practical programming point of
view, we need a stopping rule for the successive approximation in order to compute the
fixed point in a finite number of iteration steps; see above. The goal of a stopping rule is
to guarantee for a given accuracy quantity ε > 0 an approximation xN ∈ X with

d(xN , z) ≤ ε

where (Fn (x))n∈N ist the orbit generated by the iteration (1.17). To implement such a
stopping rule one may pursue two different concepts.

A priori concept
This results from the estimation (c) in Theorem 1.4:

cn
n
d(x , z) ≤ d(F(x), x) , n ∈ N .
1−c
A priori stopping rule

7
If δ := d(F(x), x) and N > (log(ε) + log(1 − c) − log(δ))/ log(c) then d(xN , z) ≤ ε .
Notice that such a number N exists due to the fact that 0 ≤ c < 1 . This number can be
computed before we compute the orbit (Fn (x))n∈N0 .
Suppose we take ε := 10−m in our stopping rule inequality. In order to obtain one more
decimal digit of precision we have to do (roughly) 1/| log(c)| more iterations steps. Stated
a little differently, if we need N iterative steps to get m decimal digits of precision, then
we need another N to double the precision to 2m digits. This kind of error behavior is
called linear convergence.

A posteriori stopping concept

The basis for this conception is the estimation in (d) of Theorem 1.4.

A posteriori stopping rule

If d(xN , xN−1 ) ≤ ε(1 − c)/c then d(xN , z) ≤ ε .

Algorithm 1.1 Successive approximation – A posteriori stopping rule

Given a fixed point equation x = F(x) and an accuracy parameter ε > 0 . This algorithm
computes a sequence x0 , x1 , , . . . , xN such that d(xN , z) ≤ ε . Here x0 is a given starting
value.
Assumptions: see Theorem 1.4

(1) k := 0

(2) u := xk

(3) If u = F(u) set N := 0 and go to line (7)

(4) xk+1 := F(u)

(5) If d(xk+1 , xk ) ≤ ε(1 − c)/c set N := k + 1 and go to line (7).

(6) Set k := k + 1 and go to line (2).

(7) STOP with an approximation xN for the fixed point z which satisfies d(xN , z) ≤ ε

For the realization of the stopping rules above we need the Lipschitz constant c . It is
obvious that an approximation c̃ with c̃ ∈ [c, 1) is sufficient.

Now, we consider the successive approximation method under errors. We follow mainly
[12]. We consider in the complete metric space (X, d) the fixed point equation (1.1)

x = F(x) (1.18)

and the associated successive approximation

xk+1 := F(xk ) , k ∈ N0 x0 := x, (1.19)

8
where F : X −→ X is a given mapping and x0 is a given starting value of the iteration.
This defines an approximating family (xk )k∈N . We assume that F is a contraction with
contraction constant c ∈ [0, 1) . Then – by the contraction theorem – (xk )k∈N converges
to the unique fixed point z of the equation (1.18). Now, instead of (xk )k∈N0 we have at
hand an approximating sequence (x̃k )k∈N0 only. This sequence may be constructed by an
iteration of a Lipschitz continuous mapping F̃ : X −→ X as follows:
x̃k+1 := F̃(x̃k ) , k ∈ N0 , x̃0 := x . (1.20)
In general, we do not have limn∈N x̃k = z when F̃ is different from F .
Theorem 1.14. Let (X, d) be a complete metric space and let F : X −→ X be a
contraction with Lipschitz constant c ∈ [0, 1) and a fixed point z . Let δ > 0 and suppose
that (xk )k∈N0 , (x̃k )k∈N0 = (x̃(δ)k )k∈N0 are sequences such that the following conditions are
satisfied:
xn+1 := F(xn ) , n ∈ N0 , x0 = x . (1.21)
d(x̃n+1 , F(x̃n )) ≤ δ , n ∈ N0 . (1.22)
d(x̃0 , x0 ) ≤ δ . (1.23)
Stopping rule: Suppose there exists N(δ) ∈ N with
d(x̃n+1 , x̃n ) ≤ d(x̃n+1 , x̃n ), n = 0, . . . , N(δ) , d(x̃N(δ)+1 , x̃N(δ) ) > d(x̃N(δ) , x̃N(δ)−1 ) (1.24)
Then
lim x̃N(δ) = z (1.25)
δ↓0

Proof:
Let m ≤ N(δ) . Then
d(x̃m , xm ) ≤ d(x̃m , F(xm−1 ) ≤ δ + cd(x̃m−1 , xm−1
..
.
≤ δ + cδ + · · · + cm−1 δ + cm d(x̃0 , x0 )
δ
≤ + d(x̃0 , x0 ) .
1−c
This implies
lim d(x̃m , xm ) = 0 , m = 1, . . . , N(δ) , lim d(x̃N(δ) , xN(δ) ) = 0 . (1.26)
δ↓0 δ↓0

Since
d(x̃m , z) ≤ d(x̃m , xm ) + d(xm , z) , m = 0, . . . , N(δ), (1.27)
we have to show limδ↓0 d(xN(δ) , z) = 0 .
Case 1: limδ↓0 N(δ) = ∞ .
Since limn d(xn , z) = 0, nothing has to be proved.
Case 2: If limδ↓0 N(δ) = ∞ does not hold then there exist a sequence (δl )l∈N and q ∈ N
with Nl := N(δl ) ≤ q, l ∈ N .
Let m ≤ q . Then
d(xq , xq−1 ) ≤ cd(xq−1 , xq−2 ) ≤ · · · ≤ cq−m d(xm , xm−1 ) . (1.28)

9
Moreover, due to cq−m ≤ 1 we have

d(xq , xq−1 ) ≤ d(xNl , x̃Nl ) + d(x̃Nl , x̃Nl −1 ) + d(x̃Nl −1 , xNl −1 ) . (1.29)

For l → ∞ the first and the second term in (1.29) converge to zero due to 1.26. From the
stopping rule we obtain
d(xNl , xNl −1 ) ≤ d(xNl +1 , xNl )
and hence

d(x̃Nl , x̃Nl −1 ) ≤ d(xNl +1 , F(x̃Nl )) + d(F(x̃Nl ), F(x̃Nl −1 )) + d(F(x̃Nl −1 ), x̃Nl ) .

This implies
2δ
d(x̃Nl , x̃Nl −1 ) ≤
. (1.30)
1−c
Therefore, the seond term in (1.29) converges to zero too if l goes to infinity. From this
we conclude that we must have

d(xq , xq−1 ) = d(F(xq−1 ), xq−1 ) = 0 .

This implies xq−1 = z and

d(xn , z) = 0 for n ≥ q − 1 . (1.31)
For n < q − 1 we have

d(xn , z) = d(xn , xq−1 ) ≤ d(xn , xn+1 ) + · · · d(xq , xq−1 ) .

Each term in the sum on the righthand side can be extimated analogous to (1.29) and
converges to zero as l goes to infinity. This shows

lim d(xN(δ) , z) = 0 . (1.32)

δ↓0

From (1.26) and (1.32) we conclude (1.25). �

Remark 1.15. The stopping rule may be formulated a little bit more general with a
general distance function d̃ which satisfies

|d(x, y) − d̃(x, y)| ≤ ε(δ), x, y ∈ X,

with limδ↓0 ε(δ) = 0 . �

1.4 Brouwer’s fixed point theorem

Probably, the most simplest existence result for a fixed point is the following one:
Every continuous function f : [a, b] −→ [a, b] has a fixed point.
To prove this, let g(t) := f(t) − t, t ∈ [a, b] . Then, since f is a mapping from [a, b]
into [a, b], we must have g(a) ≥ a, g(b) ≤ b . By the mean value theorem there exists
z ∈ [a, b] with g(z) = 0 . Clearly, z is a fixed point of f .
This result holds in particular for the interval [−1, 1] . Now, the classical formulation
of Brouwer’s fixed point theorem is an analogous“ result in a finite-dimensional normed
”
vector space. For the proof we refer to the literature; see for instance [6, 8, 16].

10
Theorem 1.16 (Fixed point theorem of Brouwer, 1912; Poincaré, 1883). Let B1 be the
closed unit ball in the euclidean space Rn and let F : B1 −→ B1 be a continuous mapping.
Then F has a fixed point.

Let C be a closed convex subset of Rn with nonempty interior. We may assume that
θ ∈ C . Then C is homeomorph to the closed unit ball B1 in Rn . This can easily seen
using the Minkovski functional with respect to the ball which is contained in the interior
of C . If C is an arbitrary nonempty closed convex subset in Rn then C has a nonempty
interior as a subset of the affine hull aff(C) of C . Therefore, C is homeomorph to the unit
ball B1 in Rk for some k ≤ n . Let T : B1 −→ C be such a homoemorphism. We set
G := T −1 ◦ F ◦ T : B1 −→ B1 . Then G is continuous and has a fixed point z . Then T (z)
is a fixed point of F . This consideration leads to a more general formulation of Brouwer’s
fixed point theorem.
Theorem 1.17. Let C be a nonempty convex compact subset of a finite-dimensional
normed space and let F : C −→ C be a continuous mapping. Then F has a fixed point.
Proof:
We present a proof different of the considerations above. It is more in the spirit of the
main topic in this monograph.
Without loss of generality we may assume that C is a subset of the euclidean space Rk
for some k . Since C is compact it is bounded. Suppose C ⊂ Br . Consider the mapping
PC : Rk −→ C defined by the property �x − PC (x)� = inf �x − u� .
u∈C

Such a map is well defined since the function f : Rd � x �−→ �x − u� ∈ R is continuous

and the set C is compact; the uniqueness of the minimum of f over C is obvious. We will
study this in the next chapter much more detailed. Notice that PC is the identity on C .
The mapping PC is continuous. To prove this, let x ∈ Rk and let (xn )n∈N be a sequence
with x = limn xn . Since the points PC (xn ) belong to the compact set C for all n ∈ N
(PC (xn ))n∈N has a cluster point y which is contained in C ; y := liml PC (xnl )) . From
�x − y� = lim �xnl − PC (xnl )� = lim dist(xnl , C) = dist(x, C)
l l

we conclude PC (x) = y since PC (x) is uniquely determined. For the fact liml dist(xnl , C) =
dist(x, C) we use that �dist(·, C) is Lipschitz-continuous; see Lemma 2.1.
Now consider R := PC �Br . Then G := F ◦ R is a continuous mapping from Br into Br and
has, due to Brouwer’s fixed point theorem, a fixed point x . Since G(x) ∈ C x is in C too.
Hence R(x) = x and F(x) = x . �
Remark 1.18. Deeply connected to the Brouwer fixed point theorem are the following
theorems:
• Hairy ball therorem
This theorem states that there is no nonvanishing continuous tangent vector field on
evendimensional n-spheres.
• Brötchensatz“
”
This ham sandwich theorem states that given n (measurable) objects“ in n-dimensional
”
space can be divided in half (with respect to their measure) with a single (n − 1)-
dimensional hyperplane.

11
 • Theorem of Borsuk-Ulam
This theorem states that every continuous function from an n-sphere into an eu-
clidean n-space maps some pair of antipodal points to the same point.
The connection between these results is the following:
Theorem of Borsuk-Ulam =⇒ Brötchensatz“
”
Theorem of Borsuk-Ulam =⇒ Brouwers fixed point Theorem
The hairy ball Theorem =⇒ Brouwers fixed point Theorem
�
Example 1.19. The following example shows that the result does not hold in infinite-
dimensional Banach spaces.
Let H be the space of the square-summable sequences,i.e.: H = l2 . This space is a Hilbert
space endowed with the inner product
�
�(xn )n∈N |(yn )n∈N � := xn y n
n∈N

and norm � 1
�(xn )n∈N � := ( x2n ) 2 .
n∈N
We define a mapping F on l2 by
�
F((xn )n∈N ) := ( 1 − �(xn )n∈N �2 , x1 , x2 , . . . ) .
Clearly, �F((xn )n∈N )� = 1 for all (xn )n∈N ∈ l2 and hence F : B1 −→ B1 . F is continuous
since limk xk = x in l2 implies
� �
lim �F(xk ) − F(x)�2 = lim(( 1 − �xk �2 − 1 − �x�2 )2 + �xk − x�2 ) = 0 .
k k

Assume F has a fixed point x = (xn )n∈N ∈ B1 then �x� = 1 (see above) and hence
x1 = 0, x2 = x1 = 0, x3 = 0, . . . , . Thus, �x� = 0, contradicting the fact that �x� = 1 . �

In Example 1.19 we have seen that Brouwer’s fixed point theorem in a infinite-dimen-
sional Hilbert space does not hold in general. The following fixed point theorem of Kaku-
tani shows that the generalization of Brouwer’s fixed point theorem without additional
assumptions is not possible; we omit the proof.
Theorem 1.20. Let H be an infinite-dimensional Hilbert space. Then there is a contin-
uous mapping F : H −→ H which maps B1 into B1 and do not have a fixed point.

1.5 The fixed point property

A topological space X has the fixed point property, if every continuous mapping f :
X −→ X has a fixed point. Theorem 1.16 and 1.17 say that the unit ball in the euclidean
space Rn and every nonempty compact convex subset in a finite-dimensional normed space
has the fixed point property. Thus, we see that the unit ball in Rn , independently of the
norm chosen, has the fixed point property.
In Chapter 7 we shall prove the following fixed point theorem.
1
Theorem 1.21 (Browder, Göhde, Kirk, 1965). Let X be a uniformly convex Banach
1
This theorem has been proved independently by Browder, Göhde and Kirk in 1965

12
space and let C := B1 be the closed unit ball in X . If F : C −→ C is a nonexpansive
mapping then F has a fixed point.

Let us consider Theorem 1.21 in a Hilbert space H. Such a space is is uniformly

convex; see Section 4.1. Then the result can be proved by considering the contraction
Ft , Ft (x) := tu + (1 − t)F(x) for some u ∈ C and for t ∈ [0, 1) .

Theorem 1.22. Let H be a Hilbert space, let C be a bounded closed convex subset of H
and let F : C −→ C be a nonexpansive mapping. Then F has a fixed point.

Proof:
Consider for t ∈ (0, 1] the map Ft : C � x �−→ tu + (1 − t)F(x) ∈ H . Since C is convex
the range of Ft is contained in C . We have

�Ft (x) − Ft (y)� ≤ (1 − t)�x − y� , x, y ∈ C ,

and we conclude from Banach’s contraction theorem that Ft has a uniquely determined
fixed point xt ∈ C: xt = tu+(1−t)F(xt ), t ∈ (0, 1] . Since C is a bounded set the diameter
diam(C) is finite. We obtain

�xt − F(xt )� = t�u − F(xt )� ≤ t diam(C) .

This shows limt↓0 �xt − F(xt )� = 0 and we say that F has approximate fixed points.
Consider Fixε (F) := {z ∈ C : �F(z) − z� ≤ ε} . We conclude from the fact that limt↓0 �xt −
F(xt )� = 0 that Fixε (F) �= ∅ for all ε > 0 . Clearly, Fixε (F) is closed for all ε > 0 since F
is continuous. Moreover, Fixε (F) is convex for all ε > 0 since C is a subset of a Hilbert
space. We shall prove this in Chapter 7 in a more general context; see 7.16. Thus we know
that Fixε (F) is weakly closed for all ε > 0 and therefore weakly compact as a subset of the
weakly compact subset C . This implies ∩ε>0 Fixε (F) �= ∅ . Clearly, every x ∈ ∩ε>0 Fixε (F)
is a fixed point of F . �

The approach which is used in the proof of Theorem 1.22 is a part of a huge theory
on the iterative computation of fixed points of nonexpansive mappings. For example,
Halpern’s iteration
xn+1 := αn xn + (1 − αn )F(xn ) , n ∈ N0 , (1.33)
(see Chapter 7) is used to compute a fixed point of F when we choose the sequence (αn )n∈N0
in an appropriate way. Surely, limn αn = 0 should be satisfied. We caome back to this
iteration method in Chapter 7.

1.6 Fixed point theorem of Schauder

Definition 1.23. Let X be a Banach space, let M ⊂ X and let F : M −→ X be a
mapping.

(a) F is called bounded if F(M ∩ B) is bounded for every bounded subset B of X .

(b) F is called completely continuous if F maps every weakly convergent sequence (in
M) into a convergent sequence.

13
(c) F is called compact if F(M ∩ B) is compact for every bounded subset B of X .
�

Now, we present three versions of Schauder’s fixed point theorem. It is sufficient to

proof only one version since they are equivalent. We shall proof the second version in
Section 5.8 as a consequence of the considerations concerning the metric projection.
Theorem 1.24 (Fixed point Theorem of Schauder, 1930/First Version). Let X be a
Banach space and let M be a nonempty bounded closed convex subset of X . Then every
completely continuous mapping F : M −→ M posesses a fixed point.
Theorem 1.25 (Fixed point Theorem of Schauder, 1930/Second Version). Let X be a
Banach space and let M be a nonempty compact convex subset of X . Let F : M −→ M
be continuous. Then F posesses a fixed point.
Proof:
See the proof of Theorem 5.14. �
Theorem 1.26 (Fixed point Theorem of Schauder, 1930/Third Version). Let X be a
Banach space and let M be a nonempty closed convex subset of X . Let F : M −→ M be
a continuous mapping and let F(M) be compact. Then F posesses a fixed point.
Definition 1.27. A metric space (X, d) is totally bounded if and only if for every r > 0
there exists a finite collection of open balls in X of radius r whose union contains X . �
Theorem 1.28. Let X , Y be Banach spaces, let U ⊂ X open and let F : U −→ Y . Let
A ⊂ U bounded. Then the following conditions are equivalent:
(a) F : A −→ Y is completely continuous.
(b) For every ε > 0 there exists a finite-dimensional subspace Yε of Y and a continuous
bounded mapping Fε : A −→ Yε such that
sup �F(x) − Fε (x)� ≤ ε .
x∈A

Fε can be chosen such that Fε (A) ⊂ co(F(A)) holds.

Proof:
Ad (a) =⇒ (b)
Since F is compact and A is bounded F(A) is compact. Let ε > 0 . Then there exist
y1 , . . . , ym ∈ F(A) with
F(A) ⊂ ∪m i
i=1 Bε (y ) .

We define φj : A −→ [0, 1] by
dist(F(x), Y\Bε (yj ))
φj (x) := �m ,x∈X.
dist(f(x), Y\Bε (yi ))
i=1

We have φ1 (x) + · · · + φm (x) = 1, x ∈ A . Set

m
�
Fε := φj yj , Yε := span(y1 , . . . , ym ) ⊂ Y .
j=1

14
Now we have

sup �F(x) − Fε (x)� ≤ ε , Fε (A) ⊂ co(y1 , . . . , ym ) ⊂ co(F(A)),

x∈A

and since co(F(A)) is bounded Fε (A) is bounded too.

Since the function dist(·, M) is Lipschitz continuous in every metric space M (see Lemma
2.1) we conclude due to the continuity of F that every φj is continuous. Hence Fε is
continuous.
Ad (b) =⇒ (a)
Since F is the uniformly limit of a family of continuos mappings F is continuous. Since
in a Banach space each point possesses a contable basis of neighborhoods it is sufficient
to show that F(A) is sequentially compact. Hence it is sufficient to show for a sequence
(F(xk ))k∈N in F(A): there exists a subsequence (F(xkl )l∈N which is convergent to a point
y := liml F(xkl ) in F(A) .
(Fn (xk ))k∈N is for each n ∈ N a bounded sequence in a finite-dimensional space. Hence
there exist subsequences of (xk )k∈N und yn ∈ Y with yn = lim Fn (xk ), n ∈ N . The
sequence (yn )n∈N is a Cauchy sequence due to

�yl − ym � ≤ �yl − Fl (xk )� + �Fl (xk ) − Fm (xk )� + �Fm (xk ) − ym �

if one chooses k = k(l, m) appropriate. As a consequence the sequence (yn )n∈N ; congerges.
y := limn yn . From

�F(xk ) − y� ≤ �F(xk ) − Fn (xk )� + �Fn (xk ) − yn � + �yn − y�

we conclude y = limk F(xk ) . Obviously, y ∈ F(A) .

�

1.7 Schauder’s fixed point theorem and successive

approximation
Now, we want to consider the successive approximation of a fixed point under the as-
sumption of the fixed point theorem of Schauder. The idea is to consider the successive
approximation in a partial ordered Banach space. This has been done in a rigorous way
for the first time by J. Schröder [13].

Definition 1.29. Let X be Banach space endowed with a partial order � .

We say that a mapping T : DT −→ X , DT ⊂ X , is syntone iff v � w implies Tv � Tw .
We say that a mapping T : DT −→ X , DT ⊂ X , is antitone iff v � w implies Tw � Tv .
�

Consider for a continuous map T : DT −→ X , DT ⊂ X convex, an equation

x = F(x) := T (x) + y for x ∈ X , (1.34)

and suppose that T can be decomposed as follows:

T = T1 + T2 where T1 is syntone and T2 is antitone . (1.35)

15
Then we may consider the iteration

vn+1 := T1 (vn ) + T2 (wn ) + y , wn+1 := T1 (wn ) + T2 (vn ) + y (1.36)

with starting values v0 , w0 ∈ DT . If we can verify that

v0 � v1 � w1 � w0 (1.37)

then the iteration (1.36) can be continued and we obtain sequences (vn )n∈N0 , (wn )n∈N0
with
v0 � v1 � v2 � · · · � vn � wn � · · · � w2 � w1 � w0 , n ∈ N . (1.38)
Under assumptions which imply that the interval“ {u ∈ X : v0 � u � w0 } is mapped
”
into a relative compact subset of X there exists a fixed point x ∈ X with

vn � x � wn for all n ∈ N0 (1.39)

by Schauder’s fixed point theorem 1.26. Thus, we may construct approximations vn , wn

for the fixed point x of F which allow the inclusion

vn � x � wn for all n ∈ N0 , (1.40)

if we have the additional property that θ � u � � u � u �� implies �u �� − u� ≤ �u �� − u � � .

This approach may be used in various situations. Notice that a matrix A ∈ Rn,n may
be decomposed into A = A1 + A2 where both A1 , −A2 have nonnegative entries. Then
the mapping u �−→ A1 u is syntone and u �−→ A2 u is antitone with respect to the
usual partial order in Rn .
Another area of application is the theory of integral equations x = F(x) = Tx + y when T
is an integral operator of (Hammerstein-)type:
�
Tu(t) := κ(t, s)φ(u(s))ds , t ∈ B ,
B

We may decompose the kernel κ and/or φ such that the method above may be applied.
Example 1.30. Consider the equation
�1
1
x(t) − 1 = T (x)(t) := |t − s|(x(s) − x(s)2 )ds , t ∈ [0, 1],
0 2
in the space C[0, 1] endowed with the supremum-norm. The partial order in C[0, 1] is
given by f � g ⇐⇒ f(t) ≤ g(t) for all t ∈ [0, 1] . T may be decomposed as follows:
�1 �
1 1
T1 (u)(t) := |t − s|u(s)ds , T2 (u)(t) := − |t − s|u(s)2 ds , t ∈ [0, 1] .
0 2 0

Clearly, T1 is syntone, T2 is antitone. With v0 ≡ 0, w0 ≡ 2 we obtain v1 (t) := 2(t −

t2 ), w1 (t) = 2(1 − t + t2 ) , t ∈ [0, 1] . We can verfify v0 � v1 � w1 � w0 . Then we obtain
a solution x of the equation above with

v1 (t) ≤ x(t) ≤ w1 (t) , t ∈ [0, 1] .

16
1.8 Appendix: Variational principle of Ekeland and
applications
Wir beweisen das Variationslemma von Ekeland in einer endlichdimensionalen Version.
Der Beweis ist hier elementar im Gegensatz zur Situation in metrischen oder topologischen
Räumen. Sei nun in diesem und dem nächsten Abschnitt | · | stets die euklidische Norm
in Rn , assoziert zum euklidischen Skalarprodukt �·|·� .
Theorem 1.31 (Variationslemma von Ekeland, 1972). Sei f : Rn −→ R unterhalbstetig,
nach unten beschränkt. Sei ε > 0, x∗ ∈ Rn mit

f(x∗ ) ≤ infn f(x) + ε . (1.41)

x∈R

Dann gibt es zu jedem γ > 0 ein x ∈ Rn mit

(a) f(x) ≤ f(x∗ );
ε;
(b) |x∗ − x| ≤ γ
(c) f(x) ≥ f(x) − γ|x − x| für alle x ∈ Rn .
Proof:
O. E. inf{f(x)|x ∈ Rn } = 0 , also f(x) ≥ 0 für alle x ∈ Rn . Betrachte die Abbildung

F : Rn � x �−→ f(x) + γ|x − x∗ | ∈ R .

Da F nach unten beschränkt und unterhalbstetig ist, und da die Niveaumengen Nr (F) :=
�x ∈ Rn |F(x) ≤ r} beschränkt und abgeschlossen sind, also kompakt sind, gibt es ein x,
das ein Minimum von F darstellt (siehe etwa [4], Satz 3.12) Nun haben wir

f(x) + γ|x − x∗ | ≤ f(x) + γ|x − x∗ | für alle x ∈ Rn , ε ≥ f(x∗ ) ≥ f(x) + γ|x − x∗ | .

Daraus folgt (b), (c) sofort. �

Corollary 1.32. Sei f : Rn −→ R Fréchet-differenzierbar und nach unten beschränkt.
Sei ε > 0, x∗ ∈ Rn mit
f(x∗ ) ≤ infn f(x) + ε . (1.42)
x∈R

Dann gibt es ein x ∈ R mitn

(a) f(x) ≤ f(x∗ );

√
(b) |x∗ − x| ≤ ε;
√
(c) |∇f(x)| ≤ ε .
Proof: √
Wende Satz 10.16 an mit γ := ε . Dann gibt es also x ∈ Rn mit
√ √
f(x) ≤ f(x∗ ), |x∗ − x| ≤ ε, ε|x − x| ≥ (f(x) − f(x)) für alle x ∈ Rn .

Aus der dritten Ungleichung folgt

√
f(x + tw) − f(x)) ≥ −t ε|w| für alle t > 0, w ∈ Rn .

17
Daraus folgt √
�∇f(x), w� ≥ − ε|w| für alle w ∈ Rn ,
was nun (c) impliziert. �

Für den Beweis des Variationslemmas im Kontext metrische Räume/unendlichdimen-

”
sionale Räume“ stellen wir ein allgemeines Prinzip vor.
Erinnerung: Sei X eine Menge. Eine Relation ≤ “heisst Halbordnung auf X genau
”
dann, wenn gilt:

i) x ≤ x (Reflexivität)

ii) x ≤ y, y ≤ x =⇒ y = x (Symmetrie)

iii) x ≤ y, y ≤ z =⇒ x ≤ z ∀x, y, z ∈ X (Transitivität)

Wir schreiben dann (X, ≤), vereinbaren die Schreibweise

S(x) := {y ∈ X|y ≥ x} ,

und können für eine Folge (xn )n∈N in X definieren:

(xn )n∈N monoton wachsend, falls xn ≤ xn+1 für alle n ∈ N .

Theorem 1.33 (Brezis–Browder, 1976). Sei (X, ≤) eine halbgeordnete Menge und ϕ :
X −→ R monoton wachsend, d.h. ϕ(x) ≤ ϕ(y), falls x ≤ y. Es gelte:

i) Für jede monoton wachsende Folge (xn )n∈N mit sup{ϕ(xn )|n ∈ N} < ∞ existiert
y ∈ X mit xn ≤ y für alle n ∈ N .

ii) Für alle x ∈ X gibt es u ∈ X mit x ≤ u, ϕ(x) < ϕ(u) .

Dann gilt: sup{ϕ(y)|y ∈ S(x)} = ∞ für alle x ∈ X.

Proof:
Für u ∈ X setze µ(u) := sup{ϕ(y)|y ∈ S(u)} .
Annahme: Es gibt x ∈ X mit µ(x) < ∞.
Definiere eine Folge (xn )n∈N induktiv in folgender Weise: x1 := x; sind x1 , . . . , xn bestimmt,
wähle xn+1 ∈ S(xn ) mit µ(xn ) − 1/n ≤ ϕ(xn+1 ) .
Wegen µ(xn ) ≤ µ(x) < ∞ existiert xn+1 und wegen xn+1 ≥ xn gilt µ(xn+1 ) ≤ µ(xn ) . Nun
gilt also

xn ≥ x, d. h. xn ∈ S(x), n ∈ N; (xn )n∈N ist monoton wachsend; ϕ(xn ) ≤ µ(x), n ∈ N.

Wegen i) gibt es y ∈ X mit xn ≤ y für alle n ∈ N . Wähle nach ii) u ∈ X mit u ∈ S(y)
und ϕ(y) < ϕ(u) . Nun gilt xn ≤ u, ϕ(u) ≤ µ(xn ), n ∈ N. Also

1
ϕ(u) ≤ µ(xn ) ≤ ϕ(xn+1 ) + 1/n ≤ ϕ(y) + , n ∈ N,
n
d. h. ϕ(u) ≤ ϕ(y) < ϕ(u), was ein Widerspruch ist. �

18
Corollary 1.34. Sei (X, ≤) eine halbgeordnete Menge, ψ : X −→ R nach oben
beschränkt und es gelte:

Für jede monoton wachsende Folge (xn )n∈N in X existiert y ∈ X mit xn ≤ y , n ∈ N .

(1.43)
Dann gilt:
(a) Ist ψ monoton wachsend, so folgt: ∀u ∈ X ∃v ∈ S(u) ∀w ∈ S(v) (ψ(v) = ψ(w)) .
(b) Ist ψ streng monoton wachsend, so folgt: ∀u ∈ X ∃v ∈ S(u) ∀w ∈ X (w ≥ v =⇒
w = v) .
Proof:
Zu (a) .
Sei u ∈ X. Betrachte (S(u), ≤), ϕ := ψ|S(u) . Dann ist ϕ monoton wachsend und nach oben
beschränkt.
Annahme: ∀ v ∈ S(u) ∃ w ≥ v (ϕ(v) < ϕ(w))
Nun sind die Voraussetzungen des Satzes 10.26 für (S(u), ≤), ϕ erfüllt. Es folgt:

∀ w ∈ S(u) (sup{ϕ(y)|y ∈ S(w) ∩ S(u)} = ∞) .

Dies ist ein Widerspruch zur Beschränktheit von ϕ .

Zu (b) .
Sei u ∈ X. Nach (a) gibt es v ∈ S(u) mit ϕ(w) = ϕ(v) für alle w ∈ X, w ≥ v . Da ϕ
streng monoton wachsend ist, gilt also:

w ∈ X, w ≥ v =⇒ ϕ(w) = ϕ(v) =⇒ w = v .

�
Corollary 1.35. Sei (X, d) ein metrischer Raum und sei (X, ≤) halbgeordnet. Sei η :
X −→ R streng monoton fallend und nach unten beschränkt. Es gelte:
(a) Für alle x ∈ X ist S(x) abgeschlossen.
(b) Für jede monoton wachsende Folge (xn )n∈N in X ist {xn |n ∈ N} kompakt.
Dann gilt:
∀ u ∈ X ∃ v ∈ S(u) ∀ w ∈ X (w ≥ v =⇒ w = v) .
Proof:
Sei ψ := −η. Wir weisen (10.19) nach. Sei dazu (xn )n∈N eine monoton wachsende Folge
in X . Dann existiert wegen (b) ein y ∈ X und eine Teilfolge (xnk )k∈N mit y = limk xnk .
Da xnk ∈ S(xnj ) für alle k ≥ j gilt, folgt wegen (a) y ∈ S(xnj ) ⊂ S(xn ), nj ≥ n, n ∈ N.
Also y ≤ xn für alle n ∈ N. Nun können wir Folgerung 10.27 (b) anwenden. �

Theorem 1.36. Sei (X, d) ein vollständiger metrischer Raum und sei f : X −→ (−∞, ∞]
unterhalbstetig, nach unten beschränkt und eigentlich, d. h. es gibt z ∈ X mit f(z) < ∞ .
Sei ε > 0, x∗ ∈ X mit
f(x∗ ) ≤ inf f(v) + ε . (1.44)
v∈X

Dann gibt es x ∈ X mit

19
 1. f(x) ≤ f(x∗ ) ;

2. d(x, x∗ ) ≤ 1 ;

3. f(y) + εd(y, x) > f(x) für alle y ∈ X\{x} .

Proof:
f(x∗ ) ist endlich wegen (10.20). X � := {z ∈ X|f(z) ≤ f(x∗ )} ist abgeschlossen, da f unter-
� d) ein vollständiger metrischer Raum mit x∗ ∈ X
halbstetig ist. Also ist (X, � . Haben wir
das Resultat in (X,� d) für f� := f � bewiesen, dann ist das Resultat in (X, d) bewiesen,
|X
denn die Bedingung 3. folgt dann mit 1. aus der Zeile
�.
f(y) + εd(y, x) > f(x∗ ) + εd(y, y) ≥ f(x∗ ) ≥ f(x) , y ∈ X\X

O. E. können wir daher nun X� = X annehmen.

Wir definieren
u ≤ v : ⇐⇒ f(v) − f(u) ≤ −εd(u, v) , u, v ∈ X ,
und zeigen, dass dadurch auf X eine Halbordnung erklärt ist.
Die Reflexivität ist klar. Seien u ≤ v, v ≤ u , d. h.

f(v) − f(u) ≤ −εd(u, v) , f(u) − f(v) ≤ −εd(u, v) .

Daraus folgt d(u, v) = 0 und daher ist die Symmetrie gezeigt. Seien u, v, w ∈ X mit
u ≤ v, v ≤ w, d. h.

f(v) − f(u) ≤ −εd(u, v) , f(w) − f(v) ≤ −εd(v, w) .

Mit der Dreiecksungleichung folgt

f(w) − f(u) ≤ −εd(u, w) , d. h. u ≤ w .

Damit ist auch die Transitivität klar.

Wir zeigen nun, dass f streng monoton fallend ist ist bezüglich der Halbordnung ≤ . Seien
dazu u, v ∈ X, u ≤ v, u �= v ; also

f(v) − f(u) ≤ −εd(u, v) < 0 , d. h. f(v) < f(u) .

Wir zeigen (a) in Folgerung 10.28. Sei dazu u ∈ X.

S(u) = {y ∈ X|y ≥ u} = {y ∈ X|f(y) + εd(y, u) ≤ f(u)}

ist abgeschlossen, da f und d(u, ·) unterhalbstetig sind.

Wir zeigen (b) in Folgerung 10.28. Sei dazu (xn )n∈N eine monoton wachsende Folge in X .
Dann konvergiert die Folge (f(xn ))n∈N , da sie monoton fallend und nach unten beschränkt
ist. Wir lesen aus
n−1
� n−1
�
εd(xn , xm ) ≤ εd(xi , xi+1 ) ≤ (f(xi+1 ) − f(xi )) = f(xn ) − f(xm ) , n, m ∈ N, n > m,
i=m i=m

20
ab, dass (xn )n∈N eine Cauchyfolge und daher konvergent ist.
Wir können nun Folgerung 10.28 anwenden: Es gibt x ∈ S(x∗ ) mit der Eigenschaft

u ≥ x =⇒ u = x . (1.45)

Wegen x ∈ S(x∗ ) folgt

0 ≤ εd(x∗ , x) ≤ f(x∗ ) − f(x) ≤ inf f(v) + ε − f(x) ≤ ε .

v∈X

Damit ist 1., 2. gezeigt. Sei y ∈ X\{x}. Wegen (10.21) gilt nicht y ≥ x, d. h. f(y) − f(x) >
−εd(y, x) . �

Corollary 1.37. Sei (X, d) ein vollständiger metrischer Raum und sei f : X −→
(−∞, ∞] unterhalbstetig, nach unten beschränkt und eigentlich. Sei γ > 0, x∗ ∈ X .
Dann gibt es x ∈ X mit

i) f(x) < f(x) + γd(x, x) für alle x ∈ X\{x};

ii) f(x) ≤ f(x∗ ) − γd(x∗ , x)) ;

Proof:
Sei Y := {x ∈ X|f(x) + γd(x, x∗ ) ≤ f(x∗ )} , g := γ1 f|Y . Y ist abgeschlossen, da f unterhalb-
stetig ist. Also ist (Y, d) ein vollständiger metrischer Raum. g ist unterhalbstetig, nach
unten beschränkt und eigentlich. Sei α := inf v∈Y g(v). Es gibt x∗∗ ∈ Y mit g(x∗∗ ) ≤ α + 1.
Also ist Satz 10.29 anwendbar mit ε = 1 :
1 1
∃ x ∈ Y ∀x ∈ Y\{x} ( f(x) < f(x) + d(x, x)) .
γ γ
Also
f(x) < f(x) + γd(x, x)) für alle x ∈ Y\{x}) , f(x) ≤ f(x∗ ) − γd(x∗ , x) .
ii) ist damit schon bewiesen.
Ist x ∈ X\Y, so gilt f(x) + γd(x, x∗ ) > f(x∗ ) und mit ii) folgt

f(x) ≤ f(x∗ ) − γd(x∗ , x) < f(x) + γ(d(x, x∗ ) − d(x∗ , x)) ≤ f(x) + γd(x, x) .

Damit ist auch i) bewiesen. �

Theorem 1.38 (Variationslemma von Ekeland, 1972). Sei (X, d) ein vollständiger metrischer
Raum und sei f : X −→ (−∞, ∞] unterhalbstetig, nach unten beschränkt und eigentlich.
Sei ε > 0, x∗ ∈ X mit
f(x∗ ) ≤ inf f(v) + ε . (1.46)
v∈X

Dann gibt es zu jedem γ > 0 ein x ∈ X mit

(a) f(x) ≤ f(x∗ );

(b) d(x∗ , x) ≤ γ;
(c) f(x) + γ ε d(x, x) > f(x) für alle x ∈ X\{x}.

21
Proof:
1d.
Sei γ > 0. Wende Satz 10.29 an unter Verwendung der zu d äquivalenten Metrik d̃ := γ
�
√
Beachte: Es liegt folgender Kompromiss für die Wahl von γ nahe: γ := ε; siehe
Folgerung 10.17.

Das folgende Resultat stellt eine Existenzaussage im Kontext der Voraussetzungen von
Satz 10.29 bereit. Ob es als Existenzsatz Verwendung finden kann, ist zweifelhaft, wir
werden ihn aber verwenden können.
Theorem 1.39 (Takahashi, 1989). Sei (X, d) ein vollständiger metrischer Raum und sei
f : X −→ (−∞, ∞] unterhalbstetig, eigentlich und nach unten beschränkt. Es gelte:

Für alle u ∈ X mit f(u) > inf f(x) gibt es v ∈ X mit v �= u, f(v) + d(u, v) ≤ f(u) .
x∈X

Dann gibt es x ∈ X mit f(x) = inf x∈X f(x).

Proof:
Annahme: f(u) > µ := inf x∈X f(x) für alle u ∈ X .
Nach Satz 10.31 gibt es x ∈ X mit

f(y) + d(y, x) > f(x) für alle y ∈ X\{x} .

Wegen f(x) > µ, gibt es v ∈ X, v �= x, mit f(v) + d(v, x) ≤ f(x) . Dies ist ein Widerspruch.
�

Die Vollständigkeit der metrischen Räume in den obigen Resultaten war wesentlich
für die Beweise. Sie ist auch wesentlich für die Gültigkeit der Resultate, denn man kann
zeigen, dass die Vollständigkeit in einem metrischen Raum schon durch die Tatsache
charakterisiert wird, dass die Aussage von Satz 10.31 für jedes f mit den dortigen Eigen-
schaften gilt.
Theorem 1.40. Sei (X, d) ein metrischer Raum. Dann sind äquivalent:
(a) (X, d) ist vollständig.
(b) Für jede unterhalbstetige Abbildung f : X −→ R mit inf v∈X f(v) ≥ 0 gilt: Ist
ε > 0, x∗ ∈ X mit f(x∗ ) ≤ inf v∈X f(v) + ε, so gibt es x ∈ X mit
1. f(x) ≤ f(x∗ ) ;
2. d(x, x∗ ) ≤ 1 ;
3. f(y) + εd(y, x) ≥ f(x) für alle y ∈ X .
Proof:
(a) =⇒ (b). Satz 10.29.
(b) =⇒ (a).
Sei (xn )n∈N eine Cauchyfolge in X . Dann existiert offenbar limn d(xn , y) für jedes y ∈ X .
Definiere f : X −→ R durch

f(y) := lim d(xn , y), y ∈ X .

n

22
f ist unterhalbstetig, da d(y, ·) stetig ist für alle y ∈ X, und es gilt inf v∈X f(v) = 0 . Sei
ε ∈ (0, 1), x∗ ∈ X mit f(x∗ ) ≤ ε . Dann gibt es nach Satz 10.29 x ∈ X mit

f(x) ≤ f(x∗ ) ≤ ε , f(y) + εd(y, x) ≥ f(x) für alle y ∈ X .

Wir zeigen: f(x) ≤ εl für alle l ∈ N .

Der Induktionsanfang l = 1 ist klar.
Sei η > 0 beliebig. Da f(x) ≤ εl ist nach Induktionsvoraussetzung, gibt es xm mit
d(xm , x) < εl + η, f(xm ) < η . Mit 3. folgt

η + ε(εl + η) > f(xm ) + εd(xm , x) ≥ f(x) .

Durch Grenzübergang η → 0 folgt f(x) ≤ εl+1 .

Da ε ∈ (0, 1) ist, gilt also f(x) = 0 , d. h. x = limn xn . �
Theorem 1.41. Sei (X, � · �) ein Banachraum und sei f : X −→ (−∞, ∞] eigentlich,
unterhalbstetig, Gateaux–differenzierbar in dom(f) := {x ∈ X|f(x) < ∞} und nach unten
beschränkt Dann gibt es zu jedem ε > 0, γ > 0 und x∗ ∈ X mit

f(x∗ ) ≤ inf f(v) + ε

v∈X

ein x ∈ X mit
1. f(x) ≤ f(x∗ ) .
2. �x − x∗ � ≤ γ.
ε.
3. �DG f(x)� ≤ γ

Proof:
Wähle x ∈ X gemä”s Satz 10.31. Dann gilt
ε
f(y) ≥ f(x) − �y − x�, y ∈ X . (1.47)
γ
Also
ε ε
f(x + tw) ≥ f(x) − �w�t für alle t > 0, d. h. DG f(x)(w) ≥ − �w� .
γ γ
Dies zeigt
ε
�DG f(x)� ≤ .
γ
�
Remark 1.42. Satz 10.34 zeigt, dass für ein Optimierungsproblem

Minimiere f(x), x ∈ X,

unter den dortigen Voraussetzungen Punkte x ∈ X existieren, die f fast minimieren“ und
”
die die notwendige Bedingung DG f(x) = θ fast erfüllen“:
”
ε
f(x) ≤ inf f(v) + ε , �DG f(x)� ≤ .
v∈X γ
�

23
Theorem 1.43 (Kontraktionssatz/Fixpunktsatz von Banach). Sei (X, d) ein vollständi-
ger metrischer Raum und sei F : X −→ X eine Kontraktion, d. h.

d(F(x), F(y)) ≤ L d(x, y) für alle x, y ∈ X

mit L ∈ [0, 1) . Dann besitzt F einen eindeutigen Fixpunkt.

Proof:
Betrachte
f : X � x �−→ d(x, F(x)) ∈ R .
Offenbar ist f stetig, nach unten beschränkt und eigentlich. Sei x∗ ∈ X, ε := d(x∗ , F(x∗ )), γ :=
1 − L . Wende damit Satz 10.29 an: Es gibt x ∈ X mit

d(x, F(x)) < d(x, F(x)) + (1 − L)d(x, x) für alle x �= x .

Dann muss x ein Fixpunkt sein, denn anderenfalls würde x := F(x) einen Widerspruch
ergeben gemä”s

d(x, F(x)) < d(F(x), F(F(x))) + (1 − L)d(x, F(x)) ≤ Ld(x, F(x)) − Ld(x, F(x)) + d(x, F(x)) .

Die Eindeutigkeit liest man mit mit zwei Fixpunkten u, v ∈ X aus

d(u, v) = d(F(u), F(v)) ≤ L d(u, v)

sofort ab. �

Theorem 1.44 (Fixpunktsatz von Caristi–Kirk, 1976). Sei (X, d) vollständiger metrischer
Raum, sei f : X −→ R nach unten beschränkt und unterhalbstetig. Sei F : X −→ X; es
gelte:
d(x, F(x)) ≤ f(x) − f(F(x)) für alle x ∈ X.
Dann besitzt F einen Fixpunkt, d. h. es gibt z ∈ X mit F(z) = z.

Proof:
Annahme: Es gilt F(z) �= z für alle z ∈ X.
Nach Satz 10.29 gibt es x ∈ X mit f(z) − f(x) > −d(z, x) für alle z ∈ X\{x}) . Also

f(x) − f(F(x)) < d(x, F(x)) ≤ f(x) − f(F(x)) ,

was ein Widerspruch ist. �

Remark 1.45. Bemerkenswert an Satz 10.37 ist, dass von der Abbildung F nicht einmal
Stetigkeitseigenschaften verlangt werden.
Unter den Voraussetzungen von Satz 10.37 liegt i. a. keine Eindeutigkeit vor, wie folgendes
Beispiel zeigt:
X := [0, 1], d(x, y) := |x − y|, f(x) := |x|, F(x) := x
Beachte: Man kann den Satz 10.29 auch aus dem Fixpunktsatz von Caristi–Kirk (Satz
10.37) herleiten. �

Betrachte folgende Sätze:

24
Theorem 1.46. Sei (X, d) ein vollständiger metrischer Raum und sei f : X −→ (−∞, ∞]
unterhalbstetig, eigentlich und nach unten beschränkt. Sei T : X −→ POT(X)\{∅}. Es
gelte:
∀ x ∈ X, x ∈
/ T (x) ∃ y �= x (f(y) + d(y, x) ≤ f(x)) .
Dann gibt es z ∈ X mit z ∈ T (z).

Theorem 1.47. Sei (X, d) ein vollständiger metrischer Raum und sei f : X −→ (−∞, ∞]
unterhalbstetig, eigentlich und nach unten beschränkt. Sei T : X −→ POT(X)\{∅}. Es
gelte für alle x ∈ X :

f(y) + d(y, x) ≤ f(x) für alle y ∈ T (x) .

Dann gibt es z ∈ X mit {z} = T (z).

Theorem 1.48. Sei (X, d) ein vollständiger metrischer Raum und sei f : X −→ (−∞, ∞]
unterhalbstetig, eigentlich und nach unten beschränkt. Sei T : X −→ POT(X)\{∅} . Es
gelte für alle x ∈ X
f(y) + d(y, x) ≤ f(x) für alle y ∈ T (x) .
Dann gibt es z ∈ X mit z ∈ T (z).

Theorem 1.49. Sei (X, d) ein vollständiger metrischer Raumund sei f : X −→ (−∞, ∞]
unterhalbstetig, eigentlich und nach unten beschränkt. Sei T : X −→ POT(X)\{∅}
abgeschlossen. Es gelte:

d(x, T (x)) ≤ f(x) − sup f(y) für alle x ∈ X .

y∈T (x)

Dann gibt es z ∈ X mit T (z) = {z}.

Lemma 1.50. Die Sätze 10.32, 10.39, 10.40, 10.41, 10.42 sind äquivalent.

Proof:
Aus Satz 10.32 folgt Satz 10.39, denn:
Annahme: T hat keinen Fixpunkt, d. h. x ∈ / T (x) für alle x ∈ X ..
Sei u ∈ X mit f(u) > inf x∈X f(x). Da T (u) �= ∅ und da u ∈ / T (u) gibt es v ∈ T (u), v �= u,
mit f(v) + d(v, u) ≤ f(u) . Damit liefert nun Satz 10.32 die Existenz von x mit f(x) =
inf x∈X f(x) . Sei y ∈ T (x). Aus

0 < d(x, y) ≤ f(x) − f(y) ≤ f(y) − f(y) = 0

lesen wir einen Widerspruch ab.

Aus Satz 10.39 folgt Satz 10.40, denn:
Annahme: Es gibt kein u ∈ X mit {u} = T (u) .
Definiere g : X � x �−→ T (x)\{x} ∈ POT(X) . Es gilt

f(g(x)) + d(g(x), x) ≤ f(x) , x ∈ X .

Aus Satz 10.39 folgt die Existenz von z ∈ X mit z ∈ T (z) . Dies ist auch ein Fixpunkt von
g nach Satz 10.37.
Aus Satz 10.40 folgt Satz 10.32, denn:

25
Definiere T : X � x �−→ {y ∈ X|f(y) + d(x, y) ≤ f(x)} ∈ POT(X) . Auf Grund der
Voraussetzung in Satz 10.32 gilt T (x) �= ∅ für alle x ∈ X .
Annahme: Es gibt kein x ∈ X mit f(x) = inf v∈X f(v) . Nun gilt

f(y) + d(x, y) ≤ f(x) , y ∈ T (x), x ∈ X .

Aus Satz 10.40 folgt die Existenz von z ∈ X mit {z} = T (z) . Dies zeigt, dass kein v ∈ X
esistiert mit f(y) + d(y, z) ≤ f(z), was ein Widerspruch ist.
Aus Satz 10.40 folgt Satz 10.41, denn:
Dies ist offensichtlich. Aus Satz 10.41 folgt Satz 10.42, denn:
Setze g(x) := 21 f(x), x ∈ X . Ist d(x, T (x)) = 0, dann ist x ∈ T (x), da T (x) abgeschlossen
ist. Hat also T keinen Fixpunkt, dann ist d(x, T (x)) > 0 für alle x ∈ X . Sei y ∈ T (x) so,
dass d(x, y) < 12 d(x, T (x)) . Dann haben wir

1 1
d(x, y) ≤ d(x, T (x)) ≤ (f(x) − sup f(y)) ≤ g(x) − g(y) .
2 2 y∈T (x)

Da g eigentlich, unterhalbstetig und nach unten beschränkt ist, hat T einen Fixpunkt
nach Satz 10.41.
Aus Satz 10.42 folgt Satz 10.40, denn Satz 10.40 ist ein Spezialfall von Satz 10.42. �

1.9 Conclusion and comments

Nonlinear functional analysis is the study of operators lacking the property and equations
which are governed by such operators. The fixed point principle is the bunch of methods
to translate or reformulate (analytic) problems in the applied sciences to equations of
fixed point type, mostly governed by nonlinear mappings.
The most earliest observation that a fixed point principle is helpful is the method of
computing the square root of a nonnegative number known as Babylonian method which
is nowadays considered as a Newton-type method. Another method which shows that the
fixed point principle is very powerful is the proof that a matrix with nonnegative entries
has an eigenvalue which is nonnegative (see exercises). Other examples we have seen in
Section 1.1.
In the sections of this chapter we have presented the most important fixed point the-
orems in the analytic study of problems in the applied science: Banach’s and Schauder’s
fixed point theorem. In our context, Brouwer’s fixed point may be considered as a result
which enables us to proof the Schauder fixed point theorem. The fixed point theorem of
Darbo is the central result of a unified approach for Banach’s and Schauder’s theorem.
A first impression for fixed point theorems for nonexpansive mappings is the proof
of a version of the fixed point theorem of Browder, Göhde, Kirk. As we will see, fixed
point theory for nonexpansive mappings benefit mostly of geometric properties of Banach
spaces.
There is a lot of fixed point theorems mostly located in other topics of mathematics: the
fixed point theorems of Borsuk, Krasnoselski, Caristi (see Section 10.6), Knaster-Tarski,
Lefschetz, Kakutani; see [2, 8, 9].

26
1.10 Exercises
1.) Let a < b let F : [a, b] −→ [a, b] be continuous and let F be differentiable in
(a, b) . We assume: There exists q ∈ [0, 1) with

|F � (ξ)| ≤ q for all ξ ∈ (a, b) . (1.48)

Show: F posesses a uniquely determined fixed point.

2.) Let X := R , � · � := | · | , M := (0, ∞) , F : M � t �−→ qt ∈ M with q ∈ [0, 1) .
Show that no z ∈ M exists with z = F(z) .
3.) Consider in the inner product space (Rn , �·�) and a continuous mapping g : Br −→
Rn with
�g(x)|x� ≥ 0 , x ∈ Rn with �x� = r .
Show: The equation g(x) = θ has a solution in Br .
4.) Let X be a Banach space, let C ⊂ X open and convex, let θ ∈ C .
Show: For all x ∈ X \C there exists a uniquely determined σ(x) ∈ (0, 1] with
σ(x)x ∈ ∂C .
5.) Let X be a Banach space and let A ⊂ X be compact. Consider continuous map-
pings Fi : A −→ X, i ∈ N, and a mapping F : A −→ X . Suppose that
limi supx∈A �Fi (x) − F(x)� = 0 . Show:
(a) F is continuous.
(b) If each Fi has a fixed point then F has a fixed point.
6.) Consider in C[0, 1], endowed with the supremum norm, the (nonlinear) mapping
�1
G(x)(t) := κ(t, s)f(s, x(s))ds , t ∈ [0, 1], x ∈ C[0, 1] .
0

Here, κ ∈ C([0, 1] × [0, 1]), f ∈ C([0, 1] × R) . Show: If

�1
|f(s, u) − f(s, v)| ≤ L|u − v| , s ∈ [0, 1], u, v ∈ R, L · max |κ(t, s)|ds < 1,
t∈[0,1] 0

then G : C[0, 1] −→ C[0, 1] is a contraction.

7.) Let X be a Banach space, let V be a compact subset of X and let F : V −→ V be
contractive, i. e.

∀ x, y ∈ V, x �= y, (�F(x) − F(y)� < �x − y�) . (1.49)

Show:
(a) F has a uniquely determined fixed point x .
(b) The iteration xn+1 := F(xn ), x0 = x, converges for each x ∈ X to x .
Hint for (b): One can show that the function x �−→ �x − F(x)� along the iteration
is monotone decreasing.
8.) Let X be a compact metric space. Show: X is separable.

27
9.) Let C[a, b] be the space of continuous functions on [a, b] endowed with the maximum-
norm. Show that C[a, b] is separable but not compact.
10.) Consider the following system of nonlinear equations:

2x21 − x22 − 8x1 = 0 , x21 + x1 x2 − 4x2 + 1 = 0

Find a reformulation as a fixed point equation in R2 and solve this equation (by
Brouwer’s fixed point theorem).
11.) Consider
1 2
F : R2 � (x1 , x2 ) �−→ (x1 e−x2 + x1 x2 + 3, ln(1 + x21 + x22 ) − 1) ∈ R2 .
6
Let R2 be endowed with the maximum norm � · �∞ .
(a) Show that F is Lipschitz-continuous in [0, 1] × [0, 1] with Lipschitz-constant
c = 56 .
(b) Show that there exists a uniquely determined fixed point of F .
(c) How many steps k do we need to obtain an accuracy �xk − z�∞ ≤ 10−3 using
the fixed point iteration xk+1 := F(xk ) starting with x0 := (0, 0) .
12.) Show that
x + 1/2
f : [0, ∞) � x �−→ ∈ [0, ∞)
x+1
is a contraction. Compute the fixed point of f and an approximation x1 using the
starting point x0 = 1 for the fixed point iteration.
13.) Show that the following system of nonlinear equations

sin(x1 + x2 ) − x2 = 0 , cos(x1 + x2 ) − x1 = 0

is solvable in R2 .
14.) Let A = (aij ) ∈ Rn,n be a matrix with entries aij ≥ 0 . Show that A has a
nonnegative eigenvalue λ with associated eigenvector x = (xi ) with entries xi ≥ 0 .
15.) Let F : Rn −→ Rn be a continuous mapping with

�F(x) − F(y)� ≥ c�x − y� for all x, y ∈ Rn

with c > 1 . Show: If F is surjective then F has a uniquely determined fixed point.
16.) Let C[0, 1] endowed with the maximum-norm � · �∞ . Let C := {x ∈ C[0, 1] : 0 =
x(0) ≤ x(s) ≤ x(1) = 1} and consider

T : C[0, 1] � x �−→ T (x) ∈ C[0, 1] with T (x)(s) := sx(s), s ∈ [0, 1] .

Show:
(1) T is a linear nonexpansive mapping with �T � = 1 .
(2) T (x) ∈ C if x ∈ C .
(3) C is a bounded closed convex set.

28
 (4) T posesses a uniquely determined fixed point belonging to C[0, 1]\C .
(5) There exists xt ∈ C with xt = tu + (1 − t)T (xt ) where u ∈ C, u(s) := s .
(6) inf x∈C �x − T (x)� = 0 .
(7) limn �x0 − T n (x0 )�∞ = diam(C) .
17.) Let X be a Banach space, let C ⊂ X be compact and let F : X −→ X be
continuous. Then the following statements are equivalent:
(a) inf x∈X �x − F(x)� = 0 .
(b) F posesses a fixed point.
18.) Let Xs be a Banach space, let C ⊂ X be compact and convex and let F : X −→ X
be nonexpansive. Then F posesses a fixed point.
19.) Let H be the euclidean space R2 and let C0 := {(x, y) ∈ R2 : x2 + y2 ≤ 1},
C1 := {(x, y) ∈ R2 : x ≥ 0, |y| ≤ x}, C2 := {(x, y) ∈ R2 : x ≤ 0, |y| ≤ |x|} und
C := (C0 ∩ C1 ) ∪ (C ∩ C2 ) . Show: C ⊂ co({PC (x, y) : (x, y) ∈ R2 \C}) .
20.) Let X be a Banach space and let U be a closed subspace of X . Consider the
quotient space

X /U := {[x] : x ∈ X } where [x] := {x + u : u ∈ U} .

Show that �[x]� := dist(x, U), x ∈ X , defines a norm in X /U .

21.) Let X be a Banach space nd let U be a closed subspace of X , different from X .
Show that for every ε >= there exists x ∈ B1 with dist(x, U) ≥ 1 − ε .

29
Bibliography

[1] R.P. Agarwal, M. Meehan, and D. O’Regan. Fixed point theory and applications.
Cambridge University Press, 2004.
[2] J. Appell and M. Väth. Elemente der Funktionalanalysis. Vieweg, Braunschweig,
2005.
[3] J. Baumeister. Differentialgleichungen. Skriptum 2006, Goethe–Universität Frank-
furt/Main.

[4] J. Baumeister. Konvexe Analysis, 2014. Skriptum WiSe 2014/15, Goethe–Universität

Frankfurt/Main.

[5] C. Bessaga. On the converse of the banach fixed-point principle“. Colloquio Math., 7:41–
”
43, 1959.

[6] J. Cronin. Fixed points and topological degree in nonlinear analysis. Amer. Math. Soc.,
Providence, 1964.

[7] M. Edelstein. An extension of Banach’s contraction principle. Proc. Amer. Math Soc.,
12:7–10, 1961.

[8] A. Granas and J. Dugundij. Fixed point theory. Springer, Berlin, 2003.

[9] V.I. Istratescu. Fixed point theorems: an introduction. Reidel, Dordrecht, 1988.

[10] O. Kada, T. Suzuki, and W. Takahashi. Non convex minimization theorems and fixed point
theorems in complete metric spaces. Math. Japon., 44:381–391, 1996.

[11] E. Karapinar. Edelstein type fixed point theorems. Ann. Funct. Anal., 2:51–58, 2011.

[12] R. Krawczyk. Abbrechkriterium für Iterationsverfahren. Z. Angewandte Math. Mech.,

52:227–232, 1972.

[13] J. Schröder. Anwendung von Fixpunktsätzen bei der numerischen Behandlung nichtlinearer
Gleichungen in halbgeordneten Räumen. Arch. Rational Mech. Anal., 4:177–192, 1959.

[14] S.L. Singh and S.N. Mishra. Remarks on recent fixed point theorems. Fixed point theory
and applications, pages 1–18, 2010.

[15] T. Suzuki and W. Takahashi. Fixed point theorems and characterizations of metric com-
pleteness. Topological methods in nonlinear Analysis, 8:371–382, 1996.

[16] P. Wittbold. Funktionalanalysis I. https://fsr6.files.wordpress.com/2014/07/funktional-

ana− skript.pdf, 2006. Humboldt Universität Berlin.

30