Uwe Schäfer

from sperner‘s lemma to
differential equations
in banach spaces
An Introduction to Fixed Point
Theorems and their Applications
Uwe Schäfer

From Sperner‘s Lemma to Differential

Equations in Banach Spaces
An Introduction to Fixed Point Theorems and their Applications
From Sperner‘s Lemma to Differential
Equations in Banach Spaces
An Introduction to Fixed Point Theorems
and their Applications

by
Uwe Schäfer
Priv.-Doz. Dr. Uwe Schäfer
Karlsruher Institut für Technologie (KIT)
Institut für Angewandte und Numerische Mathematik
[email protected]

Impressum

Karlsruher Institut für Technologie (KIT)

KIT Scientific Publishing
Straße am Forum 2
D-76131 Karlsruhe
KIT Scientific Publishing is a registered trademark of Karlsruhe
Institute of Technology. Reprint using the book cover is not allowed.
www.ksp.kit.edu

This document – excluding the cover – is licensed under the

Creative Commons Attribution-Share Alike 3.0 DE License
(CC BY-SA 3.0 DE): http://creativecommons.org/licenses/by-sa/3.0/de/
The cover page is licensed under the Creative Commons
Attribution-No Derivatives 3.0 DE License (CC BY-ND 3.0 DE):
http://creativecommons.org/licenses/by-nd/3.0/de/

Print on Demand 2014

ISBN 978-3-7315-0260-9
DOI 10.5445/KSP/1000042944
to Gabriele Schäfer
Preface

In this book we present several classical fixed point theorems and show how they
have been applied. We further indicate where they are still applied to this very
day.

We start with Brouwer’s fixed point theorem. A well-known application by J.

Nash, proving the existence of an equilibrium in any two-person game is pre-
sented in some detail.

Next, we present the Poincaré-Miranda theorem (by applying the Brouwer fixed
point theorem) and then apply this theorem to prove the correctness of an algo-
rithm used to solve nonlinear complementary questions, which arise, for example,
in free boundary problems.

In the next chapter we deal with so-called verification methods. The idea is to
use Brouwer’s fixed point theorem as well as the Poincaré-Miranda theorem to
verify, by use of a computer, that within a neighborhood of an approximate solu-
tion there really exists a solution of the problem f (x) = 0, where f : Rn → Rn .

Chapter 7 deals with Banach’s Contraction Mapping Principle. The classical

uniqueness result for Cauchy’s problem is derived from this famous principle.

Next, we present Schauder’s fixed point theorem which can be viewed as a nat-
ural generalization of Brouwer’s fixed point theorem to an infinite-dimensional
setting. The classical Peano existence theorem is presented. We also deal with
verification methods in relation to the existence of a solution of a semilinear el-
liptic partial differential equation in a neighborhood of an approximate solution.
These methods were developed in the 1990s and exploit interval arithmetic on a
computer. They still find use today in scientific research helping prove existence
via the computer.

We close with Tarski’s fixed point theorem. Though often overlooked, it is very
elegant and we show how it helps prove the existence of a solution to a parabolic
partial differential equation in a Banach space.

The book is based on a series of lectures on fixed point theorems and their appli-
cations that were given at the KIT (Karlsruhe Institute of Technology, formerly

i
Preface

known as Univerität Fridericiana Karlsruhe (TH)). The idea for this series of lec-
tures arose from the fact that fixed point theorems are used in the verification
methods mentioned above. It seems that this application is not known, even to
experts in fixed point theory.

Most authors of books about fixed point theory are functional analysts, alge-
braic topologists or mathematical economists. They do not consider numerical
methods and so the problem of roundoff errors that occur during calculations
goes unmentioned. As a consequence, no mention is made of verified results
implemented on a computer using interval arithmetic via either the Brouwer or
the Schauder fixed point theorems.

Rather than presenting the many beautiful proofs of Brouwer’s fixed point the-
orem, we opt to expose many nice applications in some detail. It is amazing
that fixed point theorems can be used in such different fields of mathematical
endeavor and we wish to share how this is so with other mathematicians.

We introduce Sperner’s lemma via a board game for children and derive Brouwer’s
fixed point theorem from it. Once that is done the following chapters can be
treated more or less independently.

I thank Prof. Dr. Joseph Diestel from the Kent State University, Ohio for his
interest, advice, and extremely valuable comments. His assiduous proofreading
turned up many infelicities of grammar, logic, and style.

My thanks are also directed to Regine Tobias, to Brigitte Maier, and to Nadine
Klumpp, of KIT Scientific Publishing, for supporting this project.

Karlsruhe, October 2014 Uwe Schäfer

ii
Contents

1 Sperner’s Lemma 1
1.1 Simplexes and Barycentric Coordinates . . . . . . . . . . . . . . 8
1.2 The Proof of Sperner’s Lemma . . . . . . . . . . . . . . . . . . . 11

2 Brouwer’s Fixed Point Theorem 13

2.1 The Proof for Simplexes . . . . . . . . . . . . . . . . . . . . . . . 19
2.2 The Proof for Convex Sets . . . . . . . . . . . . . . . . . . . . . . 21

3 Nash’s Equilibrium 25
3.1 Two-Person Games . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.2 Nash’s Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.3 Nash’s Proof of Existence . . . . . . . . . . . . . . . . . . . . . . 29

4 The Poincaré-Miranda Theorem 33

4.1 The Theorem and its Corollaries . . . . . . . . . . . . . . . . . . 33
4.2 A Fixed Point Version . . . . . . . . . . . . . . . . . . . . . . . . 39

5 The Nonlinear Complementarity Problem 43

5.1 Free Boundary Problems . . . . . . . . . . . . . . . . . . . . . . . 44
5.2 Least Element Theory . . . . . . . . . . . . . . . . . . . . . . . . 46
5.3 Tamir’s Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 48
5.4 The Correctness of Tamir’s Algorithm . . . . . . . . . . . . . . . 51

6 Verification Methods 55
6.1 Interval Computations . . . . . . . . . . . . . . . . . . . . . . . . 56
6.2 The Krawczyk Operator . . . . . . . . . . . . . . . . . . . . . . . 62
6.3 The Moore Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
6.4 The Moore-Kioustelidis Test . . . . . . . . . . . . . . . . . . . . . 65
6.5 The Frommer-Lang-Schnurr Test . . . . . . . . . . . . . . . . . . 68
6.6 A Comparison of the Tests . . . . . . . . . . . . . . . . . . . . . 69
6.7 Final Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73

7 Banach’s Fixed Point Theorem 75

7.1 Banach Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
7.2 Ordinary Differential Equations . . . . . . . . . . . . . . . . . . . 81

iii
Contents

8 Schauder’s Fixed Point Theorem 85

8.1 Compactness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
8.2 Peano’s Existence Theorem . . . . . . . . . . . . . . . . . . . . . 97
8.3 The Poincaré-Miranda Theorem in l2 . . . . . . . . . . . . . . . . 102
8.4 Verification Methods . . . . . . . . . . . . . . . . . . . . . . . . . 104

9 Tarski’s Fixed Point Theorem 109

9.1 Differential Equations in Banach Spaces . . . . . . . . . . . . . . 114
9.2 Applying Tarski’s Fixed Point Theorem . . . . . . . . . . . . . . 116

10 Further Reading 123

Bibliography 125

Index 133

iv
1 Sperner’s Lemma
Suppose two children are playing a board game. The board is a big triangle that
is subdivided into smaller triangles as depicted in Figure 1.1. At the starting
point of the game a red chip lies at the left vertex, a green chip lies at the right
vertex, and a blue chip is lying at the vertex at the top.

x x

Figure 1.1: Starting point of the game

1
1 Sperner’s Lemma

x x

x x

x x

x x

x x

x x x x x x x

Figure 1.2: The rule for the border edges

The rules of the game are as follows. Both children are placing alternatively
chips at the remaining vertices. (It is assumed that enough blue, red, and green
chips are available). The children can place their chips at vertices whereever
they want, but at the border edges of the big triangle there is a certain rule.
At the vertices lying on the left border edge of the big triangle it is only allowed
to place red or blue chips. See Figure 1.2. On the other hand, at the vertices
lying on the right border edge of the big triangle it is only allowed to place green
or blue chips. Finally, at the vertices lying on the border edge at the bottom of
the big triangle it is only allowed to place red or green chips.
Now, the aim of the game is to avoid a complete colored small triangle as long
as possible, where a complete colored small triangle is a small triangle with a

2
 1 Sperner’s Lemma

x x

x x

x x x x

x x x x x

x x x x x x

x x x x x x x

Figure 1.3: Only one vertex is left

blue, a red, and a green chip. In other words, the child who places a chip so
that a complete colored small triangle arises is the loser.
Let us consider Figure 1.3. Up to this point, both children have played very
well, because no complete colored small triangle has arisen, yet. However in this
situation the child who has to place the last chip will lose. For example, if he
(or she) places a green chip at the last vertex, then a complete colored small
triangle arises as depicted in Figure 1.4. Also a red or a blue chip will lead to a
complete colored small triangle. The reader can check this on his own.
The question arises: is there always a loser? In other words, will there arise a
complete colored small triangle during the game? This is true, indeed. In fact,
it is a consequence of Sperner’s lemma in the two-dimensional case.

3
1 Sperner’s Lemma

x x

x x x

x x x x

x x x x x

x x x x x x

x x x x x x x

Figure 1.4: A complete colored small triangle has arisen

How can we prove it? Let us consider Figure 1.4. Assume, we start a path from
the bottom of the big triangle, where it is only allowed to cross an edge of a
small triangle for the case that the two vertices of this edge are red and green.
We will show, that such a path either will end again at the bottom or it will
reach a complete colored small triangle. See Figure 1.5.
Since the vertices lying on the left border edge of the big triangle are blue and
red, and since the vertices lying on the right border edge of the big triangle are
blue and green, such a path can leave the big triangle neither on the left border
edge nor on the right border edge.

4
 1 Sperner’s Lemma

x x

x x x

x x x x

x x x x x

x x x x x x

x x x x x x x

Figure 1.5: The path that reaches a complete colored small triangle

Furthermore, such a path cannot cross three edges of the same small triangle,
because two colors of the three vertices must be the same, if the path crosses
twice the same small triangle. As a consequence, such a path cannot have a
loop. So, such a path either will end again at the bottom or it will reach a
complete colored small triangle. See again Figure 1.5.

The remaining thing that we have to show is that there exists at least one path
that will not end again at the bottom of the big triangle.

To verify this, consider the border edge at the bottom of the big triangle.

x x x x x x x

5
1 Sperner’s Lemma

Going from left to right, since at the beginning of the game at the left vertex
there was a red chip and since at the right vertex there was a green chip, there
must be a small edge with a red chip at the left and a green chip at the right. If
afterwards there is also a small edge with a green chip at the left and a red chip
at the right, then there must be a corresponding small edge with a red chip at
the left and a green chip at the right.
Otherwise the border edge at the bottom will not end at a green vertex.
As a consequence, at the bottom of the big triangle the number of small edges
consisting of a red chip and a green chip must be odd. (Concerning this number
it does not matter if the red chip is on the left and the green chip is on the right
or vice versa.)
See Figure 1.5. At the bottom of the big triangle the number of small edges
consisting of a red and a green chip is 5.
So, at least one path that starts at an edge with a red chip and a green chip
at the bottom of the big triangle will not come back to the bottom. This path
must end in a complete colored small triangle.
This means that the board game indeed ends with a loser. Actually, Sperner’s
lemma says more: Assume the children’s game is not finished when the first
complete colored small triangle occurs but when the last chip is set.
A variant of the children’s game could be that this child who created more
complete colored small triangles than the other one has lost the game.
Also in this variant of the game there is always a loser because - and this is
Sperner’s lemma in the two-dimensional case - the number of complete colored
small triangles is odd. See Figure 1.6.
To prove this we just continue the proof for the fact that there is at least one
complete colored small triangle. Up to this point we have proved the following
two statements:
(S1) The number of edges with a red chip and a green chip at the bottom of
the big triangle is odd.
(S2) Each path from the bottom of the big triangle, where it is only allowed to
cross an edge of a small triangle for the case that the two chips are green
and red, either will end again at the bottom of the big triangle or it will
reach a complete colored small triangle.
With these two statements we can conclude that the number of complete colored
small triangles that were reached by a path started from the bottom of the big
triangle is odd.

6
 1 Sperner’s Lemma

x x

x x x

x x x x

x x x x x

x x x x x x

x x x x x x x

Figure 1.6: Seven complete colored small triangles

Now, if there is another complete colored small triangle that was not reached by
a path that started from the bottom of the big triangle, then a path that starts
from this complete colored small triangle (where it is only allowed to cross an
edge of a small triangle for the case that the two chips are green and red) will
end in a further complete colored small triangle that was also not reached by a
path that started from the bottom of the big triangle. See Figure 1.7.
This means that complete colored small triangles that are not reached by a path
started from the bottom of the big triangle arise pairwise, and so we have proved
a third statement:
(S3) The number of complete colored small triangles that are not reached by a
path started from the bottom of the big triangle is even.

7
1 Sperner’s Lemma

x x

x x x

x x x x

x x x x x

x x x x x x

x x x x x x x

Figure 1.7: Even further complete colored triangles

Together with the two statements (S1) and (S2) it follows that the number of
complete colored small triangles is odd.

1.1 Simplexes and Barycentric Coordinates

In geometry, a simplex (plural simplexes or simplices) is a generalization of

the notion of a triangle or tetrahedron to arbitrary dimensions. Specifically,
a k-simplex is a k-dimensional polytope which is the convex hull of its k + 1
vertices. More formally, in 2 dimensions a simplex is a triangle; in 1 dimension
a simplex is a line segment; in N dimensions a simplex is the set of all points P

8
 1.1 Simplexes and Barycentric Coordinates

−−→
V0 V2

−−→
V0 V1

Figure 1.8: The two vectors spanning the simplex

with
−−→ −−→ −−→ −−−→
OP = x0 · OV0 + x1 · OV1 + ... + xN · OVN
so that

x0 ≥ 0, x1 ≥ 0, ..., xN ≥ 0 and x0 + x1 + ... + xN = 1.

−−→ −−→ −−−→
Here, V0 , ..., VN are the vertices of the simplex, and OV0 , OV1 , ..., OVN are the
associated position vectors. The simplex is called nondegenerate, if the N vec-
tors
−−→ −−→ −−−→
V0 V1 , V0 V2 , ... , V0 VN
are linearly independent. The numbers x0 , x1 , ..., xN are the barycentric coor-
dinates of the point P .

9
 1 Sperner’s Lemma

x
V2

x x

P
x x

x x x x

x x x x x

x x x x x x

V0 V1
x x x x x x x

Figure 1.9: Figure 1.3 combined with the vector addition (1.1)

Example 1.1. Consider Figure 1.1. It is a simplex in 2 dimensions with the

6th subdivision. Let the red vertex be V0 , let the green vertex be V1 , and let
−−→ −−→
the blue vertex be V2 . Since the vectors V0 V1 and V0 V2 are obviously linearly
independent, the simplex is nondegenerate. See also Figure 1.8.

−−→
Each point within the simplex can be reached via a linear combination of V0 V1
−−→
and V0 V2 . For example, let in Figure 1.3 the point P be the vertex without a
color. Then, (see also Figure 1.9)

−−→ 1 −−→ 4 −−→

V0 P = · V0 V1 + · V0 V2 . (1.1)
6 6

10
 1.2 The Proof of Sperner’s Lemma

Via
−−→ −−→ −−→ −−→ −−→ −−→ −−→ −−→ −−→
V0 V1 = OV1 − OV0 , V0 V2 = OV2 − OV0 , and V0 P = OP − OV0

we get from (1.1)

−−→ 1 −−→ 1 −−→ 4 −−→
OP = · OV0 + · OV1 + · OV2 .
6 6 6
This means, that the vertex P has the barycentric coordinates
 
1 1 4
, , .
6 6 6

In general, for n = 2, 3, ... we form the nth barycentric subdivision of the simplex
as follows. The new vertices are the points with barycentric coordinates
 
k0 k1 kN
, , ..., ,
n n n

where the ki are integers satisfying

N
X
all ki ≥ 0 and ki = n.
i=0

The original simplex is the whole body; the subsimplexes are its cells. In general,
in the nth subdivision of a simplex in N dimensions, the number of cells tends
to infinity as n → ∞; but the diameter of each cell tends to zero. If ∆ is the
diameter of the body, then the diameter of each cell is ∆
n.

1.2 The Proof of Sperner’s Lemma

Sperner’s lemma concerning simplexes reads as follows.

Theorem 1.1. Form the nth barycentric subdivision of a nondegenerate simplex

in N dimensions. Label each vertex with an index i ∈ {0, 1, ..., N } that satisfies
the following restriction on the boundary.
Each vertex from the vertices Vp , Vq , ..., Vs , that define a boundary
element, is labeled with an index i ∈ {p, q, ..., s}.
Then, the number of cells in the subdivision, that have vertices with the complete
set of labels: 0, 1, ..., N , is odd.

11
1 Sperner’s Lemma

Proof: The proof is by induction, where the case N = 1 is trivial and the case
N = 2 was done via the board game for children.
For the general proof we use some notation. If the vertices of a cell have the labels
m0 , m1 , ..., mk , we say the cell has type (m0 , m1 , ..., mk ). Here permutations do
not matter, but multiplicities do matter.
A cell in N dimensions has boundary elements of dimensions 0, 1, ..., N − 1.
Those of dimension 0 are vertices; those of dimensions N − 1 we will call faces.
We will call the whole cell an element of dimension N .
By F (a, b, ..., q) we mean the number of elements in the body of type (a, b, ..., q).
Let us look at the faces labeled (0, 1, ..., N − 1). Two such faces occur in each
cell of type
(0, 1, ..., N − 1, m) if m ∈ {0, 1, ..., N − 1}.
On the other hand, one face labeled (0, 1, ..., N − 1) occurs in each cell of type

(0, 1, ..., N − 1, N ).

Now look at the sum

N
X −1
2· F (0, 1, ..., N − 1, m) + F (0, 1, ..., N − 1, N ).
m=0

This sum, abbreviated by S, counts every interior face (0, 1, ..., N − 1) twice,
since every interior face is shared by two cells; but the sum counts every face
(0, 1, ..., N − 1) on the boundary of the body only once, because each face on the
boundary belongs to only one cell. Therefore, the sum S satisfies

S = 2 · Fi (0, 1, ..., N − 1) + Fb (0, 1, ..., N − 1).

Sperner’s lemma applies with dimension N −1, and so Fb (0, 1, ..., N −1) is odd by
induction. Therefore, S is odd, which implies that F (0, 1, ..., N − 1, N ) is odd. 

Remark 1.1. Emanuel Sperner (9 December 1905 - 31 January 1980) was a

German mathematician. The idea to present and to prove his lemma via a board
game for children is from the DVD: Das Spernersche Lemma by W. Dröge, H.
Göttlich, and F. Wille published 1983 by the IWF Göttingen. 

12
2 Brouwer’s Fixed Point Theorem
The Brouwer fixed point theorem was published by him in 1912. Three years
before, Brouwer could only prove his theorem for n = 3, and it was Hadamard
who in 1910 gave the first proof for arbitrary n.1
Brouwer’s proof differs from that of Hadamard and meanwhile, there are many
other proofs. See, for example, any of the books J. Franklin (32), J. Appell and
M. Väth (9) or H. Heuser (46). Brouwer’s fixed point theorem is a topological
result, and most proofs are topological. Here, we decided to present the proof of
Knaster-Kuratowski-Mazurkiewicz given in 1929 using the lemma of Sperner.
This chapter is organized as follows. First, we state the fixed point theorem and
we present some examples to illustrate it. Then, we give some examples to show
that no assumption can be dropped. In following sections we give the proof of
the theorem in two steps. First, we give the proof for simplexes and then, based
on this, we give the proof for convex sets.
Formally, the theorem reads as follows.

Theorem 2.1. (Brouwer’s fixed point theorem) Let K be a nonempty

closed bounded convex subset of Rn and suppose f is a continuous function from
K to itself. Then, there exists some x∗ ∈ K so that f (x∗ ) = x∗ .

Example 2.1. Consider Figure 2.1. There you can see the graph of a function
f : [a, b] → [a, b]. Being a self-mapping means, that the graph must stay within
the square [a, b] × [a, b]. Being continuous means, that the graph has to cross
the function y = x. So, Brouwer’s fixed point theorem is more or less obvious
for the case n = 1. 

Of course, there can be more than one fixed point. See Figure 2.2.
It is not that easy, to verify Brouwer’s fixed point theorem in higher dimensions.
For example, consider a cup of coffee. Imagine, you pour carefully the coffee
from one cup to anoter cup in one go, where both cups have equal size and
equal form. Then, Brouwer’s fixed point theorem says that at least one particle
of the coffee will be at the same position. It is not an easy task to verify this.
1 See (55).

13
2 Brouwer’s Fixed Point Theorem

a b x
Figure 2.1: One fixed point

At least it is not as easy as using Figure 2.1 in order to verify Brouwer’s fixed
point theorem for n = 1.
Before we give a proof of Brouwer’s fixed point theorem, we give some examples
to show that Brouwer’s fixed point theorem is no longer true if some assumptions
are dropped.
Before, let us recall the definition of a norm in Rn . A function k · k : Rn → R is
called a norm in Rn if it fulfills the following three conditions:
kxk ≥ 0 and kxk = 0 ⇔ x = 0,
kλ · xk = |λ| · kxk,
kx + yk ≤ kxk + kyk.
Here, x, y ∈ Rn and λ ∈ R. Note, that we use 0 for the number zero as well as
for the zero vector.

14
 2 Brouwer’s Fixed Point Theorem

a b x
Figure 2.2: More fixed points

Perhaps the most popular norm in Rn is the Euclidean norm which is defined
as follows:  
v x1
uX n
x2j with x =  ...  ∈ Rn .
u  
kxke := t
j=1 xn

Theorem 2.2. All norms in Rn are equivalent; i.e., for any two different norms
k · k1 and k · k2 in Rn there exist two positive numbers α, β so that

α · kxk1 ≤ kxk2 ≤ β · kxk1 for all x ∈ Rn .

Proof: It suffices to show that any norm k · k in Rn is equivalent to the Euclidean

norm. So, let k · k be an arbitrary norm in Rn and let ej denote the vector in

15
2 Brouwer’s Fixed Point Theorem

Rn with a 1 in the jth entry and 0’s elsewhere. Then using

v
uX
n
q
2 n u n 2
max |xj | = max xj ≤ t xj
j=1 j=1
j=1

we get
n
X n
X
kxk = k xj · ej k ≤ |xj | · kej k ≤ β · kxke
j=1 j=1

with
n
X
β= kej k.
j=1

Assume there is no α > 0 so that

α · kxke ≤ kxk

is valid for all x ∈ Rn . Then, for each k > 0 there is some x(k) ∈ Rn with

1
· kx(k) ke > kx(k) k.
k

Dividing by kx(k) ke and setting

1
y (k) := · x(k)
kx(k) k e

we get a sequence {y (k) }∞

k=1 satisfying

1
· ky (k) ke > ky (k) k and ky (k) ke = 1 for all k ∈ N. (2.1)
k
Due to the theorem of Bolzano-Weierstraß we can conclude that there is a con-
vergent subsequence {y (kl ) }∞ ∗
l=1 concerning k · ke ; i.e., there exists some y ∈ R
n

with
lim ky (kl ) − y ∗ ke = 0.
l→∞

Since
ky (kl ) − y ∗ k ≤ β · ky (kl ) − y ∗ ke
this subsequence is also convergent to y ∗ concerning the norm k · k. On the
one hand, (2.1) implies ky ∗ ke = 1, whence y ∗ 6= 0. On the other hand, this
contradicts to (2.1) because of k1l > ky (kl ) k. 

16
 2 Brouwer’s Fixed Point Theorem

Example 2.2. (concerning boundedness) A set K in Rn is bounded if there

is a constant M so that

kxk ≤ M for all x ∈ K

for some norm k · k in Rn . Due to Theorem 2.2 it is irrelevant concerning which

norm this is valid.
Let K = [0, ∞) and f : K → K defined by f (x) = x + 1. Then f has no fixed
point at all. 
Example 2.3. (concerning closedness) A set K in Rn is closed if every
convergent sequence in K has its limit in K.
Let K = (0, 1) and f : K → K defined by f (x) = x2 . Then due to

f (x) = x ⇔ x · (x − 1) = 0

f has two fixed points x∗ = 0 and x∗∗ = 1, but neither belongs to K. 

Example 2.4. (concerning convexity) A set K in Rn is convex if

x, y ∈ K ⇒ λ · x + (1 − λ) · y ∈ K

for all λ ∈ [0, 1] and for all x, y ∈ K. Let

  q
x1 1
K = {x = ∈ R2 : ≤ x21 + x22 ≤ 1}.
x2 2
K is just a disc with a hole and therefore not convex:
3 ! 3 !  
− 0 1 1
x= 4 ,y = 4 ∈ K but K 3 6 = · x + (1 − ) · y.
0 0 0 2 2

Let f : K → K be defined by rotating K by 45 degree to the left. Then the zero

vector is the only fixed point of f . It does not belong to K. 

The next example shows that the phrase ’f is continuous’ actually has to read
’f is continuous subject to some norm’, since the theorem was wrong if the
assumption would be changed to ’f is continuous subject to some metric’.
Example 2.5. (concerning continuity subject to some norm) A function
f : Rn → Rn is continuous subject to some norm k · k if for any x∗ ∈ Rn

lim kf (x(k) ) − f (x∗ )k = 0 if lim kx(k) − x∗ k = 0. (2.2)

k→∞ k→∞

Due to Theorem 2.2 it is irrelevant which norm is used in (2.2).

17
2 Brouwer’s Fixed Point Theorem

Let Rn be equipped with the so-called discrete metric which is defined as follows.
(
∗
1 if x∗ 6= x̂,
d(x , x̂) :=
0 if x∗ = x̂.

Continuity subject to this metric concerning any function f means, that from

lim d(x(n) , x∗ ) = 0
n→∞

it follows
lim d(f (x(n) ), f (x∗ )) = 0.
n→∞

However, according to this property any function f is continuous as we will see.

Let
lim d(x(n) , x∗ ) = 0.
n→∞

This means, that there exists some n0 ∈ N so that

x(n) = x∗ for all n ≥ n0 .

As a consequence,
f (x(n) ) = f (x∗ ) for all n ≥ n0 .
So,
lim d(f (x(n) ), f (x∗ )) = 0.
n→∞

Now, let x0 , y0 ∈ Rn , x0 6= y0 and let

K := {x ∈ Rn : x = x0 + λ · (y0 − x0 ) for some λ ∈ [0, 1]}.

Then, K is nonempty, bounded, closed, convex, and the function f : K → K

defined by (
x0 if x 6= x0
f (x) :=
y0 if x = x0
is continuous subject to the discrete metric, but there is no fixed point. 

18
 2.1 The Proof for Simplexes

2.1 The Proof for Simplexes

We assume that the reader is familiar with Section 1.1.

Lemma 2.1. Let S be a simplex, and let x, y ∈ S be given in barycentric

coordinates. Then,
x ≥ y ⇒ x = y.

Proof: Assume, that x 6= y, although x ≥ y. Then, there exists some i so that

xi > yi . Hence, due to X X
1= xi > yi = 1,
we get a contradiction. 

Theorem 2.3. (Brouwer’s fixed point theorem for simplexes) Let S ⊆

RN be a nondegenerate simplex in N dimensions, and let f : S → S be contin-
uous. Then, there exists some x∗ ∈ S so that x∗ = f (x∗ ).

Proof: For each x ∈ S we have f (x) ∈ S. Therefore,

N
X N
X
xi = 1 = fi (x). (2.3)
i=0 i=0

Assume, for all x ∈ S it holds

f (x) 6= x.
Then, for each x ∈ S there must be some i satisfying xi 6= fi (x), and, due to
(2.3), then there must be some j satisfying

xj > fj (x).

So, for each x ∈ S we define

m(x) := min{j : 0 ≤ j ≤ N, xj > fj (x)}.

Let n be a large positive integer. Forming the nth barycentric subdivision of S,

we see that m(x) has the following property. If x belongs to a boundary element
SF of S, then those coefficients of
−−→ −−→ −−−→
x = x0 · OV0 + x1 · OV1 + ... + xN · OVN

that belongs to vertices not in SF are zero. That means, if

Vp , Vq , ..., Vs ∈ SF , and for the remaining vertices Vi , Vi 6∈ SF ,

19
2 Brouwer’s Fixed Point Theorem

then
−−→ −−→ −−→
x = xp · OVp + xq · OVq + ... + xs · OVs .
Because xj > fj (x) ≥ 0 it follows that
m(x) ∈ {p, q, ..., s} for all x ∈ SF .
Now, Sperner’s lemma says, that there is some cell with N + 1 vertices
X0 (n), X1 (n), ..., XN (n)
satisfying
m = 0 at the vertex X0 (n),
m = 1 at the vertex X1 (n),
..
.
m = N at the vertex XN (n).
Then, by definition of m(x) we have
−−−−−→ 
x0 > f0 (x) for x = OX0 (n), 

−−−−−→ 

x1 > f1 (x) for x = OX1 (n), 
.. (2.4)
. 


−−−−−→  
xN > fN (x) for x = OXN (n).
We can do this for any n that’s large. Since S is bounded and closed, there is a
subsequence nl so that there is X ∗ ∈ S and
−−−−−→ −−−→
lim OX0 (nl ) = OX ∗ .
l→∞

Furthermore, all the vertices of the cell are close to each other if n is large. So,
−−−−−→ −−−→
lim OXj (nl ) = OX ∗ for all j ∈ {0, 1, ..., N }.
l→∞

Now, the continuity of f implies

−−−−−→ −−−→
lim f (OXj (nl )) = f (OX ∗ ) for all j ∈ {0, 1, ..., N }.
l→∞

So, we get in (2.4)

x∗0 ≥ f0 (x∗ ),
x∗1 ≥ f1 (x∗ ),
..
.
x∗N ≥ fN (x∗ ).
−−−→
Finally, Lemma 2.1 leads to x∗ = f (x∗ ) with x∗ = OX ∗ . 

20
 2.2 The Proof for Convex Sets

2.2 The Proof for Convex Sets

The step from a simplex to an arbitrary convex set is done using the following
definition.

Definition 2.1. Two sets A, B are called topologically equivalent if there is a

continuous function Ψ : A → B, with a continuous inverse Ψ−1 : B → A.

We present an example in the following theorem.

Theorem 2.4. Let K ⊆ RN be a closed bounded convex set with nonempty

interior. Then, K is topologically equivalent to a closed ball in RN .

Proof: Let B be an interior point of K and let u be any unit vector in RN . Then,
−−→
K contains a boundary point OB + λ · u with λ > 0. For each u there is only one
such boundary point. Why? Suppose there were two with λ1 < λ2 . Remember
that K contains some small ball Q centered at B. But then the point P1 with
−−→ −−→
OP1 = OB + λ1 · u lies interior to the convex hull of the ball Q and the point
−−→ −−→
P2 with OP2 = OB + λ2 · u. Therefore, P1 is interior to K, which is impossible
for a boundary point.

Thus, for each u in RN , there is a unique

λ = λ(u) > 0
−−→ −−→
so that P with OP = OB + λ · u lies on the boundary of K. The function λ(u)
is called the radial function.

We assert that the radial function is continuous as a function of the unit vector
u. If u(0) were a point of discontinuity, then the boundary of K would contain
points Pk with
−−→ −−→
OPk = OB + λk · u(k) , k = 1, 2, 3, ...
with u(k) → u(0) but with positive λk not converging to λ(u(0) ). Since λ1 , λ2 , ...
is a bounded sequence, there is a subsequence with a limit λ∗ 6= λ(u(0) ). But
−−→
then OB + λ∗ u(0) must lie on the boundary of K, since a limit of boundary
points must be a boundary point, too. Now we would have two boundary points
P∗ and P0 with
−−→ −−→ −−→ −−→
OP∗ = OB + λ∗ · u(0) and OP0 = OB + λ(u(0) ) · u(0) .

But this is impossible, since the boundary point in the direction u(0) is unique.

21
2 Brouwer’s Fixed Point Theorem

Now, we define the function Ψ : K → {y ∈ RN : kyke ≤ 1} as follows:

1 −−→
Ψ(C) :=  −−→  · BC, if C 6= B
BC
λ −
−→
kBCke

and
Ψ(B) := 0.
On the other hand let y ∈ RN with kyke ≤ 1. Then,
−−→ y
Ψ−1 (y) := OB + λ( ) · y, if y 6= 0
kyke

and
Ψ−1 (0) := B.
By definition of the radial function λ(·) we have
−−→ !
−−→ BC
kBCke ≤ λ −−→ if C 6= B.
kBCke

Therefore, if C 6= B
−−→
kBCk
kΨ(C)ke =  −−→ e  ≤ 1.
BC
λ −
−→
kBCke

On the other hand,

 
−1 −−→ Ψ(C)
Ψ (Ψ(C)) = OB + λ · Ψ(C)
kΨ(C)ke
 
−
−→
  BC  
 λ
−
−→
BC  −−→
−−→  kBCke
−−
→
 ·  BC
 −−→ −−→ −−→
= OB + λ 
  −
−→  = OB + BC = OC
 λ −−→ BC
 −−
→
  BC   kBCke
λ − −→
BC
k− −
→
BCke

e

if C 6= B. Since we have already shown that the radial function is continuous,

the only thing that remains to be shown is

lim kΨ(C (k) )ke = 0, if lim C (k) = B

k→∞ k→∞

and
−−→
lim Ψ−1 (y (k) ) = OB, if lim y (k) = 0.
k→∞ k→∞

22
 2.2 The Proof for Convex Sets

However, due to
 −−−−→ 
BC (k)
λ  −−−−→  ≥ dist(B, boundary of K) > 0
kBC (k) ke

and  
y (k)
λ ≤ diameter of K
ky (k) ke
this is obvious. 

Now, we are able to prove Brouwer’s fixed point theorem. For completeness, we
state it once more.
Theorem 2.5. (Brouwer’s fixed point theorem) Let K 6= ∅, K ⊆ Rn be
bounded, convex, and closed. Furthermore, let f : K → K be continuous. Then,
there exists some x∗ ∈ K so that f (x∗ ) = x∗ .

Proof: Case 1. We assume that K has dimension n. Then, by Theorem 2.4 K

is topologically equivalent to a closed ball in Rn ; i.e., there exists a continuous
function
Φ : K → {y ∈ Rn : kyke ≤ 1}
with a continuous inverse Φ−1 . Again, by Theorem 2.4 the set {y ∈ Rn : kyke ≤
1} is topologically equivalent to any nondegenerate n-simplex S with S ⊆ Rn .
Therefore, there exists a continuous function

ϕ : {y ∈ Rn : kyke ≤ 1} → S

with a continuous inverse ϕ−1 . Setting

Ψ := ϕ ◦ Φ

we have that K is topologically equivalent to S with Ψ : K → S. Now, we

define
g := Ψ ◦ f ◦ Ψ−1 . (2.5)
Then, g : S → S is continuous. By Brouwer’s fixed point theorem for simplexes
(see Theorem 2.3) g has a fixed point z ∗ ∈ S; i.e., z ∗ = g(z ∗ ). Via (2.5) we
get
z ∗ = Ψ(f (Ψ−1 (z ∗ ))).
Hence,
Ψ−1 (z ∗ ) = f (Ψ−1 (z ∗ )).
This means, that x∗ := Ψ−1 (z ∗ ) is a fixed point of f satisfying x∗ ∈ K.

23
2 Brouwer’s Fixed Point Theorem

Case 2. We assume that K has dimension m with m < n. Then, there exists
m orthogonal unit vectors u(1) , ..., u(m) so that for all x ∈ K there are uniquely
defined numbers y1 , ..., ym so that

x = y1 · u(1) + ... + ym · u(m) .

With  
  y1
. .  . 
U := u(1) .. ... .. u(m) ∈ Rn×m  ..  ∈ R
and y :=   m

ym
we have x = U · y and U T · x = y. Setting

K̃ := {y ∈ Rm : y = U T x with x ∈ K}

we have that K̃ is bounded, convex, and closed with dimension m. For any
y ∈ K̃ we define
f˜(y) := U T · f (U · y).
Then, f˜ : K̃ → K̃ is continuous. Applying Case 1 with n replaced by m the
function f˜ has a fixed point y ∗ ∈ K̃; i.e.,

y ∗ = f˜(y ∗ ).

It follows
U · y ∗ = f (U · y ∗ ).
This means, that x∗ := U · y ∗ is a fixed point of f satisfying x∗ ∈ K. 

24
3 Nash’s Equilibrium
Although the movie “A Beautiful Mind” from 2001 starts with the claim that
’mathematicians won World War II’, the main focus of the movie is not on
mathematicians but on the life of one particular mathematician, John Forbes
Nash, Jr., whose contributions were post-World War II.
If you talk about this movie with friends and colleagues the question will arise
about what John Nash did. Sooner or later you might have to explain the
meaning of ’Nash equilibrium’. This chapter is aimed at helping you with such
an explanation.

3.1 Two-Person Games

Concerning a two-person game two people (player 1 and player 2) are playing
against each other. Both players have the ability to make their bets. We assume
that player 1 can choose from m possibilities, whereas player 2 can choose from
n (in general other) possibilities. To simplify, we assume that there are given
two m × n matrices
   
a11 a12 · · · a1n b11 b12 · · · b1n
.. ..
 a21 a22 . . .  b21 b22 . . .
   
.  . 
A= . 
, B =  .
  
 .. . .. . .. a   .. . .. . .. b


m−1 n m−1 n
am1 am2 · · · amn bm1 bm2 · · · bmn

with the following meaning. For the case that player 1 places his bet on the
ith possibility and player 2 places his bet on the jth possibility, then player 1’s
costs are defined by
aij = eTi Aej ,

and player 2’s costs are given by

bij = eT
i Bej .

Sometimes both players are allowed to split up their bets into different possibili-
ties. We give an example. Player 1 stakes half of his money on the first possibility

25
3 Nash’s Equilibrium

and bets the other half of his money on the second possibility. Player 2 stakes a
quarter of his money on the first possibility and he stakes the remaining money
on the last possibility. Then, player 1’s costs are defined by
 1 
4
 
 0 
1 1 
..

( 0 · · · 0)A  ,
2 2 
 .

 0 
3
4
and player 2’s costs are given by
 1

4
 
 0 
1 1 
..

( 0 · · · 0)B  .
2 2 
 .

 0 
3
4
This leads us to the definition of a mixed strategy.
Definition 3.1. For m, n ∈ N we define
m
P
Sm := {x = (xi ) ∈ Rm : xi ≥ 0, i = 1, ..., m, xi = 1},
i=1
n
P
Sn := {y = (yj ) ∈ Rn : yj ≥ 0, j = 1, ..., n, yj = 1}.
j=1

Any x ∈ S m is called mixed strategy concerning player 1, and any y ∈ S n is

called mixed strategy concerning player 2. Especially, any x ∈ S m with xi = 1 for
some i is called pure strategy concerning player 1. Analogously, pure strategies
concerning player 2 are defined.
Example 3.1. Suppose we have two software companies working together. One
is called A from Argentina and the other one is called B from Brasil. One part
of their cooperation is given by some data transmission from A to B and vice
versa. Suppose in Argentina there are m different network providers, whereas
in Brasil there are n different network providers. Furthermore, suppose, that
the companies have agreed to the fact that company A will meet the costs in
Argentina and company B will meet the costs in Brasil.
Since the prices of the network providers are given, there are two cost matrices
A, B ∈ Rm×n representing the costs as follows. If company A chooses the ith
network provider for all data and if company B chooses the jth network provider
for all data, then company A has to pay aij and company B has to pay bij .
Now the question arises how the data should be split up into the network
providers so that the costs are as small as possible for both companies. 

26
 3.2 Nash’s Equilibrium

3.2 Nash’s Equilibrium

Concerning a two-person game the question naturally arises if there exists a pair
of strategies so that the costs are as small as possible for both players. That is,
one asks if x̂ ∈ S m and ŷ ∈ S n exist with

x̂T Aŷ ≤ xT Ay for all x ∈ S m , y ∈ S n , (3.1)

and
x̂T B ŷ ≤ xT By for all x ∈ S m , y ∈ S n . (3.2)
The following example will show that such a pair of strategies does not exist in
general.
Example 3.2. Let
   
5 0 5 10
A= and B = .
10 1 0 1

Suppose, there are x̂ ∈ S 2 and ŷ ∈ S 2 with

x̂T Aŷ ≤ xT Ay for all x ∈ S 2 , y ∈ S 2 (3.3)

and
x̂T B ŷ ≤ xT By for all x ∈ S 2 , y ∈ S 2 . (3.4)
We will show that from (3.3) it follows that
   
1 0
x̂ = , ŷ = . (3.5)
0 1

Firstly, for any x ∈ S 2 and for any y ∈ S 2 we have xT Ay ≥ 0 because of A ≥ O.

Since the vectors from (3.5) fulfill x̂T Aŷ = 0, it is sufficient to show
   
1 0
x̂T Aŷ = 0 ⇒ x̂ = , ŷ = .
0 1

So, let x̂T Aŷ = 0. Then, from

0 = x̂T Aŷ = 5x̂1 ŷ1 + 10x̂2 ŷ1 + x̂2 ŷ2

we get the following three conditions

5x̂1 ŷ1 = 0, 10x̂2 ŷ1 = 0, x̂2 ŷ2 = 0. (3.6)

In particular, x̂2 ŷ1 = 0. Since x̂, ŷ ∈ S 2 , by (3.6) we get

ŷ1 = 0 ⇒ ŷ2 = 1 ⇒ x̂2 = 0

27
3 Nash’s Equilibrium

and
x̂2 = 0 ⇒ x̂1 = 1 ⇒ ŷ1 = 0.
Therefore, we have ŷ1 = 0 and x̂2 = 0, which implies (3.5). Secondly, if we
substitute (3.5) in (3.4) we get a contradiction because
 
1
x̂T B ŷ = 10 > 0 = (0 1)B .
0


Example 3.2 shows that, in general, it is possible that no optimal pair of strate-
gies exist in a two-person game. So, it is no surprise that sometimes two compa-
nies have to abandon negotiations, since no agreement could be achieved, even
if dealing was allowed.
A third person who wants to arbitrate can take advantage of Nash’s equili-
brium.
Definition 3.2. Concerning a two-person game defined by A, B ∈ Rm×n , a
pair of strategies (x̂, ŷ), x̂ ∈ S m , ŷ ∈ S n is called a Nash equilibrium if
x̂T Aŷ ≤ xT Aŷ for all x ∈ S m ,
x̂T B ŷ ≤ x̂T By for all y ∈ S n
is valid.

The knowledge of some Nash equilibrium can be exploited by some arbitrator

as follows. The arbitrator is going to player 1 saying:
“I know that player 2 is playing strategy ŷ. With this assumption
it can be guaranteed mathematically that your costs are as small as
possible if you play strategy x̂.”
Afterwards, the arbitrator is going to player 2 saying:
“I know that player 1 is playing strategy x̂. With this assumption
it can be guaranteed mathematically that your costs are as small as
possible if you play strategy ŷ.”
Player 1 and player 2 are not talking to each other (in fact, they quarrel with
each other). So, no player can say if he was contacted by the arbritrator first or
not.
We want to emphasize that it is possible that there exists some pair of strategies
(x, y) satisfying
xT Ay < x̂T Aŷ and xT By < x̂T B ŷ.
This means, that the main advantage of the knowledge of some Nash equilibrium
is not taken by the players but is taken by the arbitrator!

28
 3.3 Nash’s Proof of Existence

Example 3.3. (Prisoners’ dilemma) After a robbery two suspicious persons

were arrested. The custodial judge, who wants these two persons to confess to
this crime, is aware of some small piece of evidence. So, he tries to play one
suspect off against the other one by offering a deal:
If both suspects confess, then both suspects will be arrested for five
years in prison. In case that exactly one suspect confesses, then
the other suspect will be arrested for ten years in prison and the
confessing suspect will be free. If both suspects deny, then both
suspects will be arrested for one year in prison due to illicit possession
of a firearm.
The custodial judge sends the two suspects in different cells, so they have to
make their decisions independently from each other on how to testify. Now, the
suspects are playing a two-person game.
In this game a Nash equilibrium is the situation that both suspects confess. The
custodial judge exploits the knowledge of this Nash equilibrium by visiting the
suspects consecutively by saying:
“If I was you, I would confess, since your fellow has already con-
fessed.”
Neither suspect knows whether he was visited by the custodial judge first. The
prisoners’ dilemma is the fact that the custodial judge benefits from the knowl-
edge of the Nash equilibrium but neither of the prisoners do; if both suspects
claim innocence, both will fare better. 

3.3 Nash’s Proof of Existence

In contrast to Example 3.2 where we have seen that some two-person games
do not have an optimal solution for both players, any two-person game can be
arbitrated by a Nash equilibrium because John Forbes Nash Jr. has shown that
there exists a Nash equilibrium in any two-person game. We give the proof in
the following theorem.
Theorem 3.1. Given a two-person game defined by A, B ∈ Rm×n there exists
a Nash equilibrium.

Proof: Let Ai· denote the ith row of A and let B·j denote the jth column of B.
Then, for (x, y) ∈ S m × S n and via

ci (x, y) := max{xT Ay − Ai· y, 0}, i = 1, ..., m,

dj (x, y) := max{xT By − xT B·j , 0}, j = 1, ..., n,

29
3 Nash’s Equilibrium

we define the continuous mapping T : S m × S n → Rm × Rn by T (x, y) := (x0 , y 0 )

with
xi + ci (x, y) yj + dj (x, y)
x0i := m
P , i = 1, ..., m, yj0 := Pn , j = 1, ..., n.
1+ ck (x, y) 1+ dk (x, y)
k=1 k=1

m
P
For i = 1, ..., m, ci (x, y) ≥ 0 and so ck (x, y) ≥ 0. From x ∈ S m then it
k=1
follows that x0 ∈ S m . Analogously, one can show y 0 ∈ S n . Therefore,

T (x, y) ∈ S m × S n , if (x, y) ∈ S m × S n . (3.7)

Now, we show

T (x, y) = (x, y) ⇔ (x, y) is a Nash equilibrium. (3.8)

”⇐”: Let (x̂, ŷ) be a Nash equilibrium. Then, for all i = 1, ..., m it holds

x̂T Aŷ ≤ eT
i Aŷ = Ai· ŷ.

So, we have ci (x̂, ŷ) = 0 for i = 1, ..., m. By the same argument we get dj (x̂, ŷ) =
0 for j = 1, ..., n. Therefore, we get T (x̂, ŷ) = (x̂, ŷ).
”⇒”: We assume that T (x̂, ŷ) = (x̂, ŷ), although (x̂, ŷ) is not a Nash equilibrium.
Then, either there exists some x ∈ S m with xT Aŷ < x̂T Aŷ or there exists some
y ∈ S n with x̂T By < x̂T B ŷ. Since both cases lead to a contradiction in the
same way, we assume without loss of generality, that for some x ∈ S m it holds

xT Aŷ < x̂T Aŷ. (3.9)

First, we will show that due to (3.9) it holds

Ai· ŷ < x̂T Aŷ for at least one i ∈ {1, ..., m}. (3.10)

Assume, that (3.10) is not true. Then, it follows that min Ai· ŷ ≥ x̂T Aŷ.
1≤i≤m
Therefore,
m
  m
P P
xT Aŷ = xi Ai· ŷ ≥ min Ai· ŷ · xi = min Ai· ŷ ≥ x̂T Aŷ
i=1 1≤i≤m i=1 1≤i≤m

contradicting (3.9). So, (3.10) is shown, and it follows ci (x̂, ŷ) > 0 for at least
one i ∈ {1, ..., m}. This implies
m
X
ck (x̂, ŷ) > 0. (3.11)
k=1

30
 3.3 Nash’s Proof of Existence

Second, we will show:

There exists some i ∈ {1, ..., m} with x̂i > 0 and ci (x̂, ŷ) = 0. (3.12)

Suppose, (3.12) is not true. Then, for every i ∈ {1, ..., m} it holds

x̂i = 0 or Ai· ŷ < x̂T Aŷ. (3.13)

So, we get via

m
X m
X
T
x̂ Aŷ = x̂i Ai· ŷ < x̂i (x̂T Aŷ) = x̂T Aŷ
i=1 i=1

a contradiction. Therefore, (3.12) is true. Concerning the index i from (3.12)

we get togethter with (3.11)

x̂i
x0i = m
P < x̂i .
1+ ck (x̂, ŷ)
k=1

This means that T (x̂, ŷ) 6= (x̂, ŷ). As a consequence, (3.8) is shown.
We’ve shown the equivalence of the existence of a Nash equilibrium and the
existence of a fixed point for the map T . Now the set S m × S n is a nonempty,
closed bounded convex set and the function T is a continuous self-map on S m ×
S n ; so Brouwer’s fixed point theorem tells us T has a fixed point and, with that,
we also have a Nash equilibrium. 

Theorem 3.1 is a pure existence theorem. It presents no algorithm to calculate

a Nash equilibrium. The calculation of a Nash equilibrium can be done via a
so-called linear complementarity problem. This was done by C. E. Lemke and
J. T. Howson in 1964.
Remark 3.1. If you watch the movie “A Beautiful Mind” once more, you will
see that the meaning of a Nash equilibrium was not declared correctly. This was
already noticed by S. Robinson in 2002. 
Remark 3.2. For further contributions of J. Nash to mathematics we refer to
(59). 

31
4 The Poincaré-Miranda Theorem

The well-known intermediate-value theorem which says that a continuous func-

tion f : [a, b] → R satisfying
f (a) · f (b) ≤ 0 (4.1)

equals zero at some point of the interval [a, b] was proved by Bernard Bolzano
(1781-1848). In 1883-1884, Henri Poincaré announced the following generaliza-
tion without proof:

“Let f1 , ..., fn be n continuous functions of n variables x1 , ..., xn : the

variable xi is subjected to vary between the limits ai and −ai . Let
us suppose that for xi = ai , fi is constantly positive, and that for
xi = −ai , fi is constantly negative; I say there will exist a system of
values of x where the f ’s vanish.”

In 1886 Poincaré published his argument on the homotopy invariance of the index
which is a basis for the proof. The result obtained by Poincaré has come to be
known as the theorem of Miranda, who in 1940 showed that it was equivalent
to Brouwer’s fixed point theorem. See (55).

4.1 The Theorem and its Corollaries

We will prove the Poincaré-Miranda theorem by using Brouwer’s fixed point the-
orem as it was done by Miranda in 1940. The following corollaries will generalize
the Poincaré-Miranda theorem.

Theorem 4.1. (The Poincaré-Miranda theorem) Let Ω = {x ∈ Rn : |xi | ≤

Li , i = 1, ..., n} and f : Ω → Rn be continuous satisfying

fi (x1 , x2 , ..., xi−1 , −Li , xi+1 , ..., xn ) ≥ 0 for 1 ≤ i ≤ n,

(4.2)
fi (x1 , x2 , ..., xi−1 , +Li , xi+1 , ..., xn ) ≤ 0 for 1 ≤ i ≤ n.

Then, f (x) = 0 has a solution in Ω.

33
4 The Poincaré-Miranda Theorem

Proof: Case 1: L1 > 0, ..., Ln > 0.

Case 1.a: Suppose that
)
fi (x1 , .., xi−1 , −Li, xi+1 , .., xn ) > 0
i = 1, ..., n. (4.3)
fi (x1 , .., xi−1 , +Li , xi+1 , .., xn ) < 0

Since Ω is compact and since f is continuous, we can define for i = 1, ..., n

mi := min{fi (x) : x ∈ Ω} and Mi := max{fi (x) : x ∈ Ω}.

By (4.3), we have mi < 0 and Mi > 0. Now, we define

δi+ := min{|xi + Li | : x ∈ Ω, fi (x) < 0}

and
δi− := min{|xi − Li | : x ∈ Ω, fi (x) > 0}.
Thanks to (4.3), we have δi+ > 0 and δi− > 0 for all i = 1, ..., n. Because mi < 0
and Mi > 0 for all i = 1, ..., n, there exist εi satisfying

δi+ δi−
0 < εi < min{− , }, i = 1, ..., n.
mi M i

So, we can define f˜ : Ω → Rn by

f˜i (x) := xi + εi · fi (x), i = 1, ..., n.

We claim that
f˜(Ω) ⊆ Ω. (4.4)
Let i ∈ {1, .., n} be arbitrary but fixed. With x ∈ Ω we have −Li ≤ xi ≤ Li .
Firstly, if fi (x) = 0, then

f˜i (x) = xi ∈ [−Li , Li ] .

Secondly, if fi (x) > 0, then from

0 < fi (x) ≤ Mi and εi > 0

it follows
0 < εi · fi (x) ≤ εi · Mi
and together with Li − xi ≥ δi− one can conclude that

δ−
−Li ≤ xi < xi + εi · fi (x) ≤ Li − δi− + εi · Mi < Li − δi− + i Mi = Li .
| {z } Mi
=f˜i (x)

34
 4.1 The Theorem and its Corollaries

Thirdly, if fi (x) < 0, then from

0 > fi (x) ≥ mi and εi > 0

it follows
0 > εi · fi (x) ≥ εi · mi ,
and together with Li + xi ≥ δi+ one can conclude

δ+
Li ≥ xi > xi + εi · fi (x) ≥ −Li + δi+ + εi · mi > −Li + δi+ − i mi = −Li .
| {z } mi
=f˜i (x)

In other words, f˜ is a continuous self-mapping of Ω to Ω. Since Ω is nonempty,

closed, bounded, and convex, we can apply Brouwer’s fixed point theorem. This
means that there exists some x∗ ∈ Ω satisfying f˜(x∗ ) = x∗ which leads to
f (x∗ ) = 0 due to the definition of f˜.

Case 1.b: Let (4.2) be true. Then, we define

 T
(k)
f (k) (x) = f1 (x), ..., fn(k) (x)

with
(k) xi
fi (x) := fi (x) − , i = 1, ..., n, k = 1, 2, 3, ... .
k
Since
(k) Li
fi (x1 , .., xi−1 , −Li , xi+1 , .., xn ) = fi (x1 , .., xi−1 , −Li , xi+1 , .., xn ) + >0
k
and
(k) Li
fi (x1 , .., xi−1 , Li , xi+1 , .., xn ) = fi (x1 , .., xi−1 , Li , xi+1 , .., xn ) − < 0,
k
Case 1.a guarantees the existence of some x(k) ∈ Ω satisfying f (k) (x(k) ) = 0.
Since Ω is compact, there exists a convergent subsequence x(kl ) ; i.e.,

lim x(kl ) = x∗ and x∗ ∈ Ω.

l→∞

Due to the fact that f is continuous, we have

!
(k )
(k ) x l
0 = lim fi l (x(kl ) ) = lim fi (x(kl )
)− i = fi (x∗ ) for all i = 1, .., n.
l→∞ l→∞ kl

35
4 The Poincaré-Miranda Theorem

Case 2: L1 ≥ 0, ..., Ln ≥ 0.
Case 2.a: L1 = 0, ..., Ln = 0. Then, Ω = {0}, and for all i = 1, ..., n we have
fi (0) ≥ 0 and fi (0) ≤ 0, which leads to f (0) = 0.
Case 2.b: Without loss of generality, we assume that

L1 > 0, ..., Lk > 0, Lk+1 = 0, ..., Ln = 0.

Following (4.2), we then have for i = k + 1, ..., n

fi (x1 , ..., xk , 0, ..., 0) ≥ 0 and fi (x1 , ..., xk , 0, ..., 0) ≤ 0.

Hence, fi (x1 , ..., xk , 0, ..., 0) = 0 for i = k + 1, ..., n. Defining Ω̃ := {x ∈ Rk :

|xi | ≤ Li , i = 1, ..., k} and f˜ : Ω̃ → Rk by

f˜i (x) = fi (x1 , ..., xk , 0, ..., 0), i = 1, ..., k,

there exists  
x̃1
 . 
 .  ∈ Ω̃
 . 
x̃k

satisfying f˜(x̃∗ ) = 0 according to Case 1. Finally, we have

f (x̃∗1 , ..., x̃∗k , 0, ..., 0) = 0

with (x̃∗1 , ..., x̃∗k , 0, ..., 0)T ∈ Ω. 

The following corollary will show that it is not necessary that the center of the
domain is the zero vector.

Corollary 4.1. Let xc ∈ Rn , L ∈ Rn , Li ≥ 0, i = 1, ..., n. Furthermore, let

G = {x = (xi ) ∈ Rn : |xci − xi | ≤ Li , i = 1, .., n},

)
G+ c
i = {x ∈ G : xi = xi + Li }
i = 1, .., n.
G− c
i = {x ∈ G : xi = xi − Li }

If g : G → R is continuous satisfying
)
gi (x) ≤ 0 for all x ∈ G−
i
i = 1, ..., n,
gi (x) ≥ 0 for all x ∈ G+
i

then there exists x∗ ∈ G satisfying g(x∗ ) = 0.

36
 4.1 The Theorem and its Corollaries

Proof: Considering Ω := {x̃ ∈ Rn : |x̃i | ≤ Li , i = 1, .., n} and f (x̃) =

−g(x + x̃) we have that f : Ω → Rn is continuous and for x̃ ∈ Ω we get
c

fi (x̃1 , .., x̃i−1 , −Li , x̃i+1 , .., x̃n ) =

−gi (xc1 + x̃1 , .., xci−1 + x̃i−1 , xci − Li , xci+1 + x̃i+1 , .., xcn + x̃n ) ≥ 0
| {z }
∈ G−
i

and fi (x̃1 , .., x̃i−1 , Li , x̃i+1 , .., x̃n ) =

−gi (xc1 + x̃1 , .., xci−1 + x̃i−1 , xci + Li , xci+1 + x̃i+1 , .., xcn + x̃n ) ≤ 0.
| {z }
∈ G+
i

According to Theorem 4.1 there exists x̃∗ ∈ Ω with f (x̃∗ ) = 0. Defining x∗ :=

xc + x̃∗ we get x∗ ∈ G and g(x∗ ) = g(xc + x̃∗ ) = −f (x̃∗ ) = 0. 

The following corollary is motivated by (4.1).

Corollary 4.2. Let xc ∈ Rn , L ∈ Rn , Li ≥ 0, i = 1, ..., n and G = {x = (xi ) ∈
Rn : |xci − xi | ≤ Li , i = 1, .., n}. In addition, let
)
G+ c
i = {x ∈ G : xi = xi + Li }
i = 1, .., n.
G− c
i = {x ∈ G : xi = xi − Li }

If g : G → Rn is continous satisfying

gi (x) · gi (y) ≤ 0 for all x ∈ G+ −

i and for all y ∈ Gi , i = 1, .., n, (4.5)

then there exists some x∗ ∈ G with g(x∗ ) = 0.

Proof: Case 1: For all i ∈ {1, ..., n} either we have

gi (x) ≤ 0 for all x ∈ G− +

i and gi (y) ≥ 0 for all y ∈ Gi (4.6)

or we have

gi (x) ≥ 0 for all x ∈ G− +

i and gi (y) ≤ 0 for all y ∈ Gi . (4.7)

In this case, we define g̃ : G → Rn by

(
gi (x) if i satisfies (4.6)
g̃i (x) =
−gi (x) if i satisfies (4.7).

The set G and the function g̃(x) fulfill the assumptions of Corollary 4.1. There-
fore, there exists x∗ ∈ G with g̃(x∗ ) = 0, whence g(x∗ ) = 0.

37
4 The Poincaré-Miranda Theorem

Case 2: The set {1, ..., n} can be subdivided in two sets of indices, say I
and J, so that for all i ∈ I we have either (4.6) or (4.7), whereas for all
i ∈ J there exist either x(1,i) , x(2,i) ∈ G+
i or x
(1,i)
, x(2,i) ∈ G−
i satisfying
(1,i) (2,i)
gi (x ) < 0, gi (x ) > 0.
Without loss of generality, for all i ∈ J let x(1,i) , x(2,i) ∈ G+
i . Then, we can
conclude from (4.5) that

for all i ∈ J we have gi (y) = 0 for all y ∈ G−

i . (4.8)

Again without loss of generality, we assume that

I = {1, ..., n0 } and J = {n0 + 1, ..., n}.

Then, we define G̃ := {x̃ ∈ Rn0 : |xci − x̃i | ≤ Li , i = 1, ..., n0 } and the function
g̃ : G̃ → Rn0 by
 
g1 (x̃1 , ..., x̃n0 , xcn0 +1 − Ln0 +1 , ..., xcn − Ln )

g̃(x̃) :=  .. 
. .
gn0 (x̃1 , ..., x̃n0 , xcn0 +1 − Ln0 +1 , ..., xcn − Ln )

Due to Case 1, we can conclude that there exists some x̃∗ ∈ G̃ with g̃(x̃∗ ) = 0.
Setting  
x̃∗1
 .. 

 . 

∗
 x̃n

x∗ :=  c
 0 

 xn0 +1 − Ln0 +1 
 
 .. 
 . 
xcn − Ln
we have x∗ ∈ G and g(x∗ ) = 0 by (4.8). 

38
 4.2 A Fixed Point Version

4.2 A Fixed Point Version

Considering the original statement of Poincaré and considering the original paper
of Miranda, one can see, that there is a tiny difference. It would be exactly the
same, if (4.2) read

fi (x1 , x2 , ..., xi−1 , −Li , xi+1 , ..., xn ) < 0 for 1 ≤ i ≤ n,

fi (x1 , x2 , ..., xi−1 , +Li , xi+1 , ..., xn ) > 0 for 1 ≤ i ≤ n,

and if Li was replaced by ai . Of course, the ≤-sign is more general than the
<-sign, but it is not clear, why Miranda changed
“... for xi = ai , fi is constantly positive ...”
into
fi (x1 , x2 , ..., xi−1 , +Li , xi+1 , ..., xn ) ≤ 0.
Of course, concerning the existence of a zero, this is irrelevant thanks Corollary
4.2. However, it is worth mentioning, that one also gets the existence of a fixed
point, when using the original notation of Miranda. This will be shown in the
following theorem.
Theorem 4.2. Let Ω = {x ∈ Rn : |xi | ≤ Li , i = 1, ..., n}. Furthermore, let
g : Ω → Rn be continuous satisfying
)
gi (x1 , ..., xi−1 , −Li , xi+1 , ..., xn ) ≥ 0
for i = 1, ..., n.
gi (x1 , ..., xi−1 , +Li , xi+1 , ..., xn ) ≤ 0

Then, there exists some x∗ ∈ Ω satisfying g(x∗ ) = x∗ .

Proof: Considering f (x) := g(x) − x, x ∈ Ω we have for all i ∈ {1, ..., n}

fi (x1 , , .., xi−1 , −Li , xi+1 , .., xn ) = gi (x1 , .., xi−1 , −Li , xi+1 , .., xn ) + Li ≥ 0,
fi (x1 , , .., xi−1 , +Li , xi+1 , .., xn ) = gi (x1 , .., xi−1 , +Li , xi+1 , .., xn ) − Li ≤ 0.

Then, from Theorem 4.1, we can conclude that there exists x∗ ∈ Ω satisfying
f (x∗ ) = 0. Therefore, g(x∗ ) = x∗ . 

Analogous corollaries for this theorem (such as Corollary 4.1 and Corollary 4.2
for Theorem 4.1) are not true as we can see in Figure 4.1 and Figure 4.2, re-
spectively.

39
4 The Poincaré-Miranda Theorem

-L L x

Figure 4.1: No fixed point in [−L, L]

In Figure 4.1 the graph of a function y = g(x), x ∈ [−L, L] is given satisfying

g(−L) < 0, g(L) > 0. According to Corollary 4.2, g(x) has a zero in [−L, L].
However, g(x) has no fixed point in [−L, L], which is no contradiction to Theo-
rem 4.2, since g(−L) ≥ 0, g(L) ≤ 0 is not valid, here.

40
 4.2 A Fixed Point Version

xc -L xc +L x

Figure 4.2: No fixed point in [xc − L, xc + L].

In Figure 4.2 the graph of a function y = g(x), x ∈ [xc − L, xc + L] is given

satisfying g(xc − L) > 0, g(xc + L) < 0. According to Corollary 4.1 g(x) has a
zero in [xc − L, xc + L]. However, g(x) has no fixed point in [xc − L, xc + L].
Remark 4.1. Note that the function g in Theorem 4.2 is not assumed to be a
self-mapping as is assumed in many other fixed point theorems. 

41
5 The Nonlinear Complementarity
Problem

Given a vector f = (f1 , ..., fn )T of n real, nonlinear functions of a real vector

z = (z1 , ..., zn )T , the nonlinear complementarity problem, abbreviated by NCP,
is to find a vector z so that

z ≥ 0, f (z) ≥ 0, z T · f (z) = 0,

or to show that no such vector exists. Here, the inequality sign is meant com-
ponentwise; i.e.,

z ≥ 0 ⇔ z1 ≥ 0, ..., zn ≥ 0
f (z) ≥ 0 ⇔ f1 (z) ≥ 0, ..., fn (z) ≥ 0.

Especially, the condition z T · f (z) = 0 then implies

zi = 0 or fi (z) = 0 for all i = 1, ..., n.

In 1974, Tamir published an algorithm for solving the NCP for the case that f is
a so-called Z-function. Tamir’s algorithm is a generalization of Chandrasekaran’s
algorithm which solves the linear complementarity problem for the case that the
given matrix M is a so-called Z-matrix.

Proving his results, Tamir used the iterative processes of Gauss-Seidel and of
Jacobi. In Section 5.4 we present a different proof where these iterative processes
are not used. Instead, we use the least element theory and the Poincaré-Miranda
theorem.

Before we are going to show how the Poincaré-Miranda theorem can be used to
prove the correctness of Tamir’s algorithm, we give an application of the NCP
where Tamir’s algorithm can be used.

43
5 The Nonlinear Complementarity Problem

5.1 Free Boundary Problems

In contrast to a boundary problem where a differential equation has to be solved

in a given domain satisfying given initial- and fixed boundary conditions, a
free boundary problem is defined by a differential equation together with the
unknown boundary of the domain.
A simple example is melting ice. If a piece of ice swims in a glass of water,
the content of the glass is separated in two parts: the water and the ice. To
determine the temperature of the water as a function dependent on time and
space, one has to solve a parabolic differential equation (see (18)). The boundary
of the water is moving since the ice is melting. Therefore, the boundary is not
known from the start and it has to be determined as part of the solution.
For more examples concerning free boundrary problems we refer to the book
of Crank. In the following example we show how the NCP arises from free
boundary problems.

Example 5.1. We consider the ordinary free boundary problem:


Find s > 0 and z(x) : [0, ∞) → R such that 
p 

z 00 (x) = 1 + (z(x))2 , for x ∈ [0, s], 
(5.1)
z(0) = 1, z 0 (s) = 0, 




z(x) = 0, for x ∈ [s, ∞).

One can show that (5.1) has a unique solution, say {ŝ, ẑ(x)}, and that
√
ŝ ≤ 2,
ẑ(x) > 0, x ∈ [0, ŝ).

See (83), (99). However, it is not possible to solve (5.1) explicitly. So, one is
satisfied to get approximations for ẑ(x) at discrete points and to get an approx-
imation for ŝ.
√
To show this, we choose n ∈ N and subdivide the interval [0, 2] equidistantly;
i.e., we define
√
2
h := , xi := i · h, i = 0, 1, ..., n + 1.
n+1
Then, we have )
ẑ(xi ) > 0, if xi ∈ [0, ŝ),
√ (5.2)
ẑ(xi ) = 0, if xi ∈ [ŝ, 2].

44
 5.1 Free Boundary Problems

For the case that xi ∈ [0, ŝ],

p
−ẑ 00 (xi ) + 1 + ẑ(xi )2 = 0.
√
On the other hand, if xi ∈ (ŝ, 2], we get
p p
−ẑ 00 (xi ) + 1 + ẑ(xi )2 = −0 + 1 + 02 = 1 > 0.
As a consequence, for all i = 0, ..., n + 1 we have

ẑ(xi ) ≥ 0, 
p 

−ẑ 00 (xi ) + 1 + ẑ(xi )2 ≥ 0, (5.3)
 p  

ẑ(xi ) · − ẑ 00 (xi ) + 1 + ẑ(xi )2 = 0. 

Due to Taylor’s formula with remainder, we get for i = 1, ..., n

1 1 1
ẑ(xi + h) = ẑ(xi ) + h · ẑ 0 (xi ) + h2 · ẑ 00 (xi ) + h3 · ẑ 000 (xi ) + h4 · ẑ 0000 (ξi )
2 6 24
and
1 1 1
ẑ(xi − h) = ẑ(xi ) − h · ẑ 0 (xi ) + h2 · ẑ 00 (xi ) − h3 · ẑ 000 (xi ) + h4 · ẑ 0000 (νi )
2 6 24
with some νi ∈ (xi − h, xi ) and some ξi ∈ (xi , xi + h). Adding both equations
we get
−ẑ(xi − h) + 2 · ẑ(xi ) − ẑ(xi + h) 1
= −ẑ 00 (xi ) − h2 (ẑ 0000 (ξi ) + ẑ 0000 (νi )) .
h2 24
So, for small h we have
−ẑ(xi − h) + 2 · ẑ(xi ) − ẑ(xi + h)
≈ −ẑ 00 (xi ).
h2
√
Defining z0 := ẑ(0) = 1 and zn+1 := ẑ( 2) = 0 we substitute
zi for ẑ(xi )
and
−zi−1 + 2 · zi − zi+1
for − ẑ 00 (xi )
h2
in (5.3); i.e., for i = 1, ..., n we get

zi ≥ 0, 



q 

−zi−1 + 2 · zi − zi+1 
+ 1 + zi2 ≥ 0, (5.4)
h2 

 −z q 

i−1 + 2 · zi − zi+1
 

zi · 2
+ 1 + zi2 = 0. 
h

45
5 The Nonlinear Complementarity Problem

Defining the function f : Rn → Rn by

    p  
2 −1 0 ··· 0 z1 1 + z12 − 12

 .. ..   ..   ..   0h
 −1 2 −1 . . 
 .  
   .  
  .


1 
 . .. ..   ..   ..   . 
f (z) = 2  0 .. . . 0  + + . 
 .   .

h    . 

 . .  .   ..   .
 .. .. −1 2 −1 .
. . 
p .
    
0 ··· 0 −1 2 zn 1 + zn2 0
(5.5)
we see that (5.4) is an NCP.

This example will be continued at the end of this chapter. 

5.2 Least Element Theory

Tamir’s algorithm presented in the next section was developed for the NCP
defined by a so-called Z-function. So, we start with its definition.

Definition 5.1. A function f : Rn → Rn is called a Z-function if for any z ∈ Rn

the functions ϕij (t) := fi (z + tej ), i 6= j, i, j = 1, ..., n where ej denotes the jth
standard unit vector satisfy

t ≤ t̃ ⇒ ϕij (t) ≥ ϕij (t̃).

The function given in (5.5) is a Z-function. We’ll see that Tamir’s algorithm can
be used to solve the discretized free boundary problem.

Moreover, Z-functions have been studied independently from free boundary

problems. They have nice properties. One of them will be described in the
following lemma.

Lemma 5.1. Let Rk+ denote the first orthant of Rk ; i.e.,

Rk+ = {x ∈ Rk : xj ≥ 0, j = 1, ..., k}.

Furthermore, let g : Rk+ → Rk be a Z-function and let

S := {t ∈ Rk+ : g(t) ≥ 0}.

If S 6= ∅, then S has a least element.

46
 5.3 Tamir’s Algorithm

Proof: Let x, y ∈ S and z := inf{x, y}, where

 
min{x1 , y1 }

inf{x, y} =  .. 
. .
min{xk , yk }

Then, since g is a Z-function and since x, y ∈ S we have

xi ≤ yi ⇒ gi (z) ≥ gi (x) ≥ 0,
yi < xi ⇒ gi (z) ≥ gi (y) ≥ 0.

Hence,
inf{x, y} ∈ S. (5.6)
Now, let i ∈ {1, ..., m} be fixed but arbitrary. We define

Mi := {s ∈ R : there exists x ∈ S with xi = s}.

Since S is bounded from below, Mi is bounded from below, too. Furthermore

Mi is closed and is a nonempty subset of R. Therefore, Mi has a least element,
say mi . So, we define
 
m1
x̃ =  ... 
 

mn

and let xi ∈ S satisfy xii = mi . Then, we have

x̃ = inf{x1 , ...xn } = inf{x1 , inf{x2 , ... inf{xn−1 , xn }...}}.

So we finally have x̃ ∈ S due to (5.6). 

Note, that not every bounded nonempty set S ⊂ Rn has a least element. See
Figure 5.1.

We end this section with the definition of the infimum of a set.

Definition 5.2. Let S ⊆ Rn be a nonempty set. Then, the greatest lower bound
of S is denoted by inf S. If z := inf S, then it holds

z≤t for all t ∈ S

and
if v ≤ t for all t ∈ S then v ≤ z.

47
5 The Nonlinear Complementarity Problem

x2 6

@
@
@
@ S
@
r @
z -
x1

Figure 5.1: z = inf S 6∈ S

5.3 Tamir’s Algorithm

Tamir’s algorithm is given in Table 5.1, where Rk+ denotes the first orthant of
Rk ; i.e., Rk+ = {x ∈ Rk : xj ≥ 0, j = 1, ..., k}. We want to remark that the
pseudocode in Table 5.1 is not the original pseudocode presented by Tamir. We
have removed the modified Jacobi process. Instead, we use the lines 5-7.

Example 5.2. (Wilhelm, 2008) Let

    
17 −3 −5 −1 z1 22 + z13
 −8 11 −1 −6   z2   −12 + z23 
f (z) =  
 −3 −5 17 −1   z3
+
  15 + z33  .


−2 −2 −9 9 z4 −50 + z43

We will use Tamir’s algorithm to solve the NCP defined by this Z-function. At
the beginning, we have
 
22
 −12 
k = 0, z (0) = 0, f (z (0) ) = 
 15  ≥
 6 0.

−50

We choose i1 = 2 and get J = {2}. Now, Tamir’s algorithm considers the

function
g (1) : R → R
g (1) (t) := f2 (0, t, 0, 0) = t3 + 11t − 12.

48
 5.3 Tamir’s Algorithm

begin
k := 0 ; z := 0 ; J := ∅ ;
if f (z) ≥ 0 then goto 10
else repeat k := k + 1;
choose ik ∈ {1, ..., n} with fik (z) < 0 ;
J := J ∪ {ik } ;
..., ik } and g (k) : 
let J = {i1 ,  Rk+ → Rk be defined as
k
P
   fi1 ( tj eij ) 
t1  j=1 
 ..   .. 
 .  7→   .
;

 k 
tk  P 
fik ( tj eij )
j=1
5: let M (k) := {t ∈ Rk+ : g (k) (t) = 0, tj ≥ zij , j = 1, ..., k − 1};
6: if M (k) 6= ∅ then
k
P (k)
7: begin t(k) := inf M (k) ; z := tj eij end
j=1
else begin write(’NCP(f ) has no solution’); goto 20 end;
until f (z) ≥ 0 ;
10: write(’The solution is ’,z);
20: end.

Table 5.1: Tamir’s algorithm

Note that g (1) (1) = 0 and the function g (1) (t) is strictly increasing, since
d (1)
g (t) = 3t2 + 11 > 0.
dt
This implies M (1) = {1}. Going on with Tamir’s algorithm, we get
   
0 19
 1   0 
t(1) = 1, z (1) =   (1)
 0  , but f (z ) =  10  6≥ 0.
 

0 −52
This leads to i2 = 4 and J = {2, 4}. Now, Tamir’s algorithm considers the
function
g (2) : R2 → R2    
(2) 11 −6 t1 −12 + t31
g (t1 , t2 ) := + .
−2 9 t2 −50 + t32

49
5 The Nonlinear Complementarity Problem

Because of g (2) (2, 3) = 0, we have

 
2
∈ {t ∈ R2+ : g (2) (t) = 0, t1 ≥ 1} = M (2) .
3

Now, we want to show that M (2) is a singleton. So, assume that g (2) ((a, b)) = 0
with a ≥ 1. Then,
a3 + 11a − 6b − 12 = 0,
b3 + 9b − 2a − 50 = 0.

Substituting
a3 + 11a − 12
b= (5.7)
6
in the second equation, we get
 3
a3 + 11a − 12 9
h(a) := + (a3 + 11a − 12) − 2a − 50 = 0.
6 6

Since
 2
a3 + 11a − 12 3a2 + 11 9 2 87
h0 (a) = 3 · · + a + > 0,
6 6 2 6
the function h(a) is strictly increasing. So, there can be only one a with h(a) = 0.
(2)
Then, via(5.7),
 also b is unique. As a result, M is a singleton and we have
(2) 2
M ={ }. So, we define
3
 
  0
2  2 
t(2) = and z (2) =
 0 .

3
3

Since  
13
 0 
f (z (2) ) = 
 2  ≥ 0,


0
 
0
 2 
the algorithm stops with a solution z =   0  of the NCP.
 
3

50
 5.4 The Correctness of Tamir’s Algorithm

5.4 The Correctness of Tamir’s Algorithm

Before we prove the correctness of Tamir’s algorthm, we need another lemma.
Lemma 5.2. Let f : Rn+ → Rn be a continuous Z-function and M (k) be defined
as in Tamir’s algorithm. If M (k) 6= ∅, then inf M (k) ∈ M (k) .

Proof: We consider the sets S (k) := {t ∈ Rk+ : g (k) (t) ≥ 0}. Since f is a
Z-function, all g (k) are Z-functions, too. Since M (k) ⊆ S (k) , the condition
M (k) 6= ∅ implies S (k) 6= ∅, and it follows by Lemma 5.1 that every S (k) has a
least element, say t(k) . First, we will show by induction, that g (k) (t(k) ) = 0.
k = 1 : With fi1 (0) < 0 and ∅ 6= M (1) the assertion follows by the intermediate-
value theorem and the continuity of f .
k−1 k: Let t(k−1) be the least element of S (k−1) , let
 ! 
k−1
P (k−1)
 fi1 tj eij 
 j=1 
 .. 
o = g (k−1) (t(k−1) ) = 
 
. , (5.8)
 ! 
 k−1

 P (k−1) 
fik−1 tj eij
j=1

and let  
 (k−1)  k−1
(k) t X (k−1)
0 > gk ( ) = fik  tj eij  . (5.9)
0
j=1

By Lemma 5.1, S (k) has a least element t(k) . Since g (k−1) is a Z-function, for all
j ∈ {1, ..., k − 1}
   
Xk k−1
X (k)
(k)
0 ≤ fij  tj eij  ≤ fij  tj eij  .
j=1 j=1

(k) (k)
Therefore, (t1 , ..., tk−1 )T ∈ S (k−1) , whence
   
(k) (k−1)
t1 t1
 .   .. 
 .  ≥ inf S (k−1) =  . (5.10)
 .   . 
(k) (k−1)
tk−1 tk−1

Since t(k) ∈ S (k) , g (k) (t(k) ) ≥ 0. In order to prove g (k) (t(k) ) = 0, we assume that
(k)
there exists some j ∈ {1, ..., k} with gj (t(k) ) > 0. By (5.10), (5.8), and (5.9)

51
5 The Nonlinear Complementarity Problem

we then have
   
 (k−1)
 k k−1
(k) t X (k)
X (k−1)
tj = ⇒ 0 < fij  tj eij  ≤ fij  tj eij  ≤ 0;
0 j j=1 j=1

 
(k) t(k−1)
it follows that tj > . So, we choose δ > 0 so that
0 j
 
(k) t(k−1) (k)
tj −δ > and gj (t(k) − δ · ej ) > 0.
0 j

This is possible, since g (k) is continuous. Then, we have

(k) (k) (k)
gj (t(k) − δ · ej ) > 0, gi (t(k) − δ · ej ) ≥ gi (tk) ) ≥ 0, i 6= j.

Hence, t(k) − δ · ej ∈ S (k) , which is a contradiction. Therefore, g (k) (t(k) ) = 0.

Together with (5.10) it follows t(k) ∈ M (k) .
Finally, we have to show t(k) = inf M (k) . So, let z := inf M (k) . Since t(k) ∈ M (k) ,
it simply follows that z ≤ t(k) .
Next, since t(k) = inf S (k) , we have

t(k) ≤ t for all t ∈ S (k) .

Because of M (k) ⊆ S (k) , we have

t(k) ≤ t for all t ∈ M (k) .

By the definition of the infimum, it holds

t(k) ≤ z.

So, we finally get t(k) = z = inf M (k) . 

We will again state Corollary 4.1 because its notations can be used better within
the proof in the following theorem.
Corollary 5.1. Let a, b ∈ Rn , a ≤ b, Ω be the rectangle Ω := {x ∈ Rn : xi ∈
[ai , bi ], i = 1, ..., n}, and f : Ω → Rn be a continuous function on Ω. If

fi (x1 , x2 , ..., xi−1 , ai , xi+1 , ..., xn ) ≤ 0 for 1 ≤ i ≤ n,

fi (x1 , x2 , ..., xi−1 , bi , xi+1 , ..., xn ) ≥ 0 for 1 ≤ i ≤ n,

then f (x) = 0 has a solution in Ω.

52
 5.4 The Correctness of Tamir’s Algorithm

Theorem 5.1. Let f : Rn → Rn be a continuous Z-function. Then, Tamir’s

algorithm solves the NCP.

Proof: We show by induction: If M (k) = ∅, then the NCP has no solution.

k = 1: Let f (0) 6≥ 0, fi1 (0) < 0, and M (1) = ∅. Suppose, the NCP has a
solution, say z ∗ . Then, we consider g (1) defined on the interval [0, zi∗1 ]. On the
one hand, we have g (1) (0) = fi1 (0) < 0. On the other hand,
 
n
X
0 ≤ fi1 (z ∗ ) = fi1  zj∗ ej  ≤ fi1 (zi∗1 ei1 ) = g (1) (zi∗1 ).
j=1

Since f is continuous, g (1) is continuous, too. So, by the intermediate-value

theorem there must be some ξ ∈ [0, zi∗1 ] satisfying g (1) (ξ) = 0. This is a contra-
diction to the assumption that M (1) = ∅ holds.
k−1 k: Let t(k−1) be the least element of S (k−1) satisfying (5.8) and (5.9),
and let M (k) = ∅. Suppose, the NCP has a solution, say z ∗ . Then,
   
n
X k−1
X
0 ≤ fij (z ∗ ) = fij  zj∗ ej  ≤ fij  zi∗j eij  , j = 1, ..., k − 1. (5.11)
j=1 j=1

Hence, (zi∗1 , ..., zi∗k−1 )T ∈ S (k−1) and it follows that (zi∗1 , ..., zi∗k−1 )T ≥ t(k−1) .
Now, we consider the function g (k) on the product of intervals
 (k−1) ∗

[t1 , zi1 ]

 .. 

 . 
Ω= .
 [t(k−1) , z ∗ ] 
 k−1 ik−1 
[0, zi∗k ]
(k−1)
Let t ∈ Ω; i.e., tj ∈ [tj , zi∗j ] , j = 1, ..., k − 1 and tk ∈ [0, zi∗k ]. Then, for all
l ∈ {1, ..., k − 1} we have
   
k−1
X k−1
X
(k−1) (k−1)
fil  tj eij + tl eil + tk eik  ≤ fil  tj eij  = 0
j=1, j6=l j=1

by (5.8) and
   
k−1
X k
X
fil  tj eij + zi∗l eil + tk eik  ≥ fil  zi∗j eij  ≥ fil (z ∗ ) ≥ 0.
j=1, j6=l j=1

53
5 The Nonlinear Complementarity Problem

Finally,    
k−1
X k−1
X (k−1)
fik  tj eij + 0 · eik  ≤ fik  tj eij  < 0
j=1 j=1

by (5.9) and  
k−1
X
fik  tj eij + zi∗k · eik  ≥ fik (z ∗ ) ≥ 0.
j=1

Since f is continuous, g (k) is continuous, too. So, by Corollary 5.1 applied to g (k)
and Ω there must be some ξ ∈ Ω satisfying g (k) (ξ) = 0. This is a contradiction
to the assumption that M (k) = ∅ holds. 
Example 5.3. We continue Example 5.1. F. Wilhelm has shown that f (z),
z ≥ 0 is injective. As a result, applying Tamir’s algorithm for solving the NCP,
all sets M (k) are either empty or a singleton. In contrast to the original paper
of Tamir, the method for calculating a zero of g (k) is not fixed in Table 5.1. So,
it is left to the programmer which method for calculating a zero is chosen.
The results presented in Table 5.2 are based on an implementation of D. Ham-
mer. The input data are n as well as the tolerance ε > 0. As the method for
calculating a zero of g (k) , Newton’s method was chosen, where


 0 if k = 1
 (k−1) 
tstart := t

 if k > 1
0

was taken as the starting point, respectively. If zi = 0 and zi+1 > 0, then
s̃ := 12 (xi + xi+1 ) was taken as an approximation for ŝ. See Table 5.2 for some
examples. Note, that the exact value of ŝ satisfies ŝ ∈ [1.393206, 1.397715]; see
(84). 

n s̃ running time n s̃ running time

50 1.372619 0.017 s 50 1.372619 0.028 s
100 1.379208 0.114 s 100 1.393210 0.201 s
150 1.390799 0.720 s 150 1.390799 0.831 s
200 1.389587 1.507 s 200 1.389587 2.192 s
250 1.388859 3.962 s 250 1.388859 4.577 s
500 1.387397 20.478 s 500 1.393042 29.514 s

Table 5.2: ε = 10−5 ε = 10−11

54
6 Verification Methods
Many numerical methods for calculating a zero of a function f : Rn → Rn
involve calculating an approximate solution, say x̃. These methods often consist
of iterative algorithms using the stopping criterion
kf (x̃)k∞ < ε, (6.1)
where, here, kf (x̃)k∞ := max{|f1 (x̃)|, ..., |fn (x̃)|} and where ε > 0 is a given
tolerance. However, the criterion (6.1) is not sufficient in order to conclude that
there is a solution of f (x) = 0.
Example 6.1. Let f : R2 → R2 defined by
! !
f1 (x, y) min{x, − 21 x + 12 y − ε}
f (x, y) = = .
f2 (x, y) min{y, − 25 x + 25 y − 3ε}
Assume, an approximate solution is given by
(x̃, ỹ) = (0, ε).
1
Then, kf (x̃, ỹ)k∞ = 2ε < ε. However, a zero of f does not exist. Assume,
f (x∗ , y ∗ ) = (0, 0). Then, the first component gives
1 1
x∗ = 0 or − x∗ + y ∗ − ε = 0.
2 2
Case 1: Let x∗ = 0. Then, y ∗ ≥ 2ε > 0. Substituting this result in f2 (x∗ , y ∗ ) =
0, we get a contradiction; after all,
5 5 5
0 = f2 (x∗ , y ∗ ) = min{y ∗ , − · 0 + · y ∗ − 3ε} ≥ min{2ε, · 2ε − 3ε} = 2ε > 0.
2 2 2
Case 2: Let − 21 x∗ + 12 y ∗ − ε = 0 (and x∗ > 0). It follows
y ∗ = x∗ + 2ε > 0.
Substituting this result within f2 (x∗ , y ∗ ) = 0, again we get a contradiction
0 = f2 (x∗ , y ∗ ) = min{y ∗ , − 25 x∗ + 25 y ∗ − 3ε}

= min{x∗ + 2ε, − 25 x∗ + 5
2 · (x∗ + 2ε) − 3ε} = 2ε > 0. 

55
6 Verification Methods

Considering (6.1), we only can guess that in the so-called interval vector
 
[x̃1 − ε, x̃1 + ε]

 .. 

 . 
[x̃n − ε, x̃n + ε]

there exists an x∗ satisfying f (x∗ ) = 0. A method done by use of a computer

that verifies (i.e. that proves) that this guess is really true is called a verification
method.

How can this be done?

The basic idea is very simple: Check by use of a computer that (4.5) is valid
for  
[x̃1 − ε, x̃1 + ε]
Ω=
 .. 
.
 . 
[x̃n − ε, x̃n + ε]
Then, the corollaries of the Poincaré-Miranda theorem prove that there is some
x∗ ∈ Ω with f (x∗ ) = 0.

Another idea is to transform the problem f (x) = 0 into a fixed point problem
and to verify by use of a computer the assumptions of Brouwer’s fixed point
theorem.

We present three verification methods. One of them is based on Brouwer’s fixed

point theorem and two of them are based on the Poincaré-Miranda theorem. All
of them use interval computations.

6.1 Interval Computations

Let [a] denote a real compact interval; i.e.,

[a] = [a, a] = {x ∈ R : a ≤ x ≤ a}.

Sometimes we also write

sup[a] instead of a and inf[a] instead of a.

Especially we do this in the case that the interval is described by a long term,
see Section 6.6.

56
 6.1 Interval Computations

Furthermore, let IR denote the set of all real compact intervals. Then, we can
define +, −, ·, / in the set theoretic sense; i.e.,
[a] + [b] = {a + b : a ∈ [a], b ∈ [b]},
[a] − [b] = {a − b : a ∈ [a], b ∈ [b]},
[a] · [b] = {a · b : a ∈ [a], b ∈ [b]},
[a] a
= { : a ∈ [a], b ∈ [b]} if 0 6∈ [b].
[b] b
Example 6.2. Let [a] = [1, 3] and [b] = [2, 4]. Then,
[a] + [b] = {a + b : a ∈ [1, 3], b ∈ [2, 4]} = [3, 7],
[a] − [b] = {a − b : a ∈ [1, 3], b ∈ [2, 4]} = [−3, 1],
[a] · [b] = {a · b : a ∈ [1, 3], b ∈ [2, 4]} = [2, 12],
[a] a 1 3
= { : a ∈ [1, 3], b ∈ [2, 4]} = [ , ]. 
[b] b 4 2

Following Example 6.2, it is easy to see that the results of [a] ∗ [b], where ∗ ∈
{+, −, ·, /} can be expressed by the interval boundaries of [a] and [b]. More
precisely,

[a] + [b] = [a + b, a + b], 



[a] − [b] = [a − b, a − b], 



[a] · [b] = [min{a · b, a · b, a · b, a · b}, max{a · b, a · b, a · b, a · b}] (6.2)




[a] 1 1 

= [a] · [ , ] if 0 6∈ [b]. 

[b] b b
So, for any a ∈ [a] and any b ∈ [b] we have
a ∗ b ∈ [a] ∗ [b], ∗ ∈ {+, −, ·, /}.
(For the case ∗ = / we have to assume, that 0 6∈ [b]). This means, that any term
a ∗ b is included within the interval [a] ∗ [b]. For example,
1
+ π ∈ [0.333, 0.334] + [3.141, 3.142] = [3.474, 3.476].
3
The smaller the interval will be, the better an approximation of the exact value
can be propagated. But interval arithmetic is not as easy as it seems. Some
rules, that you are used to know in R, are not valid in IR. For example, the
distributivity law, which reads
a · (b + c) = a · b + a · c,

57
6 Verification Methods

is not valid in IR. Let [a] = [1, 2], [b] = [−1, 1], and [c] = [2, 4]. Then,

[a] · ([b] + [c]) = [1, 2] · [1, 5] = [1, 10],

whereas
[a] · [b] + [a] · [c] = [−2, 2] + [2, 8] = [0, 10].
One can show, that only the so-called sub-distributivity law

[a] · ([b] + [c]) ⊆ [a] · [b] + [a] · [c]

is valid in IR. We refer to the book of G. Alefeld and J. Herzberger for a detailed
introduction into interval computations.
The main thing, that we need in the following sections, is the meaning of a
so-called interval evaluation.

Example 6.3. Let f (x) = x3 + 2x2 + x + 2 and let [x] = [1, 2]. Then, an interval
evaluation is defined via the basic rules (6.2) as follows:

f ([x]) = f ([1, 2]) = [1, 2] · [1, 2] · [1, 2] + 2 · [1, 2] · [1, 2] + [1, 2] + 2

= [1, 8] + [2, 8] + [3, 4] = [6, 20].

Clearly, if there is some unknown number, say η, that we can only include in an
interval, say η ∈ [1, 2], then we have f (η) ∈ f ([1, 2]). 

To get a feeling where we can really benefit from an interval evaluation, we

consider the following detailed example.

Example 6.4. We consider the boundary value problem:


−y 00 (x) = f (x, y(x), y 0 (x)), x ∈ [a, b] 

y(a) = α, (6.3)


y(b) = β.

To present the main ideas, we just consider a special example.


−y 00 (x) = ex , x ∈ [0, 1]  
y(0) = 0, (6.4)


y(1) = 0.

We choose n ∈ N and subdivide the interval [0, 1] equidistantly; i.e., we define

1
h := , xi := i · h, i = 0, 1, ..., n + 1.
n+1

58
 6.1 Interval Computations

Let ŷ(·) denote the solution, then Taylor’s formula with remainder term leads
to
1 1 1
ŷ(xi + h) = ŷ(xi ) + h · ŷ 0 (xi ) + h2 · ŷ 00 (xi ) + h3 · ŷ 000 (xi ) + h4 · ŷ 0000 (ξi )
2 6 24
and
1 1 1
ŷ(xi − h) = ŷ(xi ) − h · ŷ 0 (xi ) + h2 · ŷ 00 (xi ) − h3 · ŷ 000 (xi ) + h4 · ŷ 0000 (νi )
2 6 24
with some νi ∈ (xi − h, xi ) and some ξi ∈ (xi , xi + h). Adding both equations
we get

−ŷ(xi − h) + 2 · ŷ(xi ) − ŷ(xi + h) h2 0000

2
= −ŷ 00 (xi ) − (ŷ (ξi ) + ŷ 0000 (νi )) (6.5)
h 24
for i = 1, ..., n. Keeping in mind, that −ŷ 00 (x) = ex , we know ŷ 0000 (x) = −ex ,
too. For example, let n = 1. Then, h = 21 , x0 = 0, x1 = 21 , x2 = 1 and

1 1
ν1 ∈ [x1 − h, x1 ] = [0, ], ξ1 ∈ [x1 , x1 + h] = [ , 1].
2 2
Formula (6.5) then reads
 2
−ŷ(0) + 2 · ŷ( 12 ) − ŷ(1) 1 1 1


= e2 + eξ1 + eν1 .
1 2 24 2
2

Using ŷ(0) = 0, ŷ(1) = 0, ν1 ∈ [0, 21 ], and ξ1 ∈ [ 12 , 1], the interval evaluation

gives
1 1 1 1  [0, 1 ] 1

ŷ( 21 ) ∈ e2 + · e 2 + e[ 2 ,1]
8 8 96
1 1 1 1 1 1
= e2 + · · [1 + e 2 , e 2 + e] ⊆ [0.209539, 0.211776].
8 8 96
For n > 1 formula (6.5) leads to
     1 
2 −1 0 ··· 0 ŷ(x1 ) ex1 + 24 · h2 (eξ1 + eν1 )
 . . ..   ..   .. 
 −1
 2 −1 . . 


 .  
  . 

1  . . .   ..   .. 
.. .. .. · =
h2  0 0  . .
 
    
 . .   ..   .. 
 . . . . −1 2 −1   .   . 
1
0 ··· 0 −1 2 ŷ(xn ) exn + 24 · h2 (eξn + eνn )

with some νi ∈ (xi − h, xi ) and some ξi ∈ (xi , xi + h), i = 1, ..., n.

59
6 Verification Methods

Applying the interval Cholesky method (see (5)) we get for n = 8:

 
1
ŷ ∈ [0.073376707, 0.073463768]
9
 
2
ŷ ∈ [0.132943708, 0.133096479]
9
 
3
ŷ ∈ [0.177080357, 0.177276777]
9
 
4
ŷ ∈ [0.203974522, 0.204192766]
9
 
5
ŷ ∈ [0.211601374, 0.211819618]
9
 
6
ŷ ∈ [0.197698130, 0.197894550]
9
 
7
ŷ ∈ [0.159736085, 0.159888857]
9
 
8
ŷ ∈ [0.094889365, 0.094976663].
9
The problem (6.4) was chosen, so that the solution can be expressed explicitly
in order to compare the results. The solution of (6.4) reads
ŷ(x) = −ex + (e − 1) · x + 1.
1


So, we have ŷ 2 = 0.21042... and
 
1
ŷ = 0.073401...
9
 
2
ŷ = 0.132992...
9
 
3
ŷ = 0.177148...
9
 
4
ŷ = 0.204057...
9
 
5
ŷ = 0.211692...
9
 
6
ŷ = 0.197787...
9
 
7
ŷ = 0.159811...
9
 
8
ŷ = 0.094936... .
9

60
 6.1 Interval Computations

The idea to use Taylor’s formula with remainder term and to use interval com-
putations in order to get inclusions of the exact solution at discrete points, was
presented by E. Hansen and R. Moore. It was applied to problem (6.3), where
the existence of a unique solution is given, but where this solution cannot be
stated explicitly as in our simple example (6.4). For the technical details we
refer to (43), for example. 

Interval computations can easily be extended to the n-dimensional case. An

interval matrix, for instance, is a matrix with an interval in each entry; i.e.,
 
[1, 2] [2, 3]
[A] = .
[−1, 0] [1, 3]

By IRm×n , we denote the set of all interval matrices with m columns and n
rows. Then, with [A], [B] ∈ IRn×n , we define

[A] ± [B] = ([aij ]) ± ([bij ]) = ([aij ] ± [bij ])

and
n
X
[A] · [B] = ( [aik ] · [bkj ]),
k=1

where the computations of intervals are done as defined in (6.2).

An interval evaluation can also be defined as in the 1-dimensional case.

Example 6.5. Let

f (x, y) = x · y 2 , x ∈ [−2, 1], y ∈ [1, 2].

Then, f ([x], [y]) = [−2, 1] · [1, 2]2 = [−2, 1] · [1, 4] = [−8, 4]. 

Finally, we want to note that not every function has an interval evaluation.

Example 6.6. Consider the function


 sin x , x 6= 0,
f (x) = x

1, x = 0.

Then, f ([−1, 1]) is not defined. 

61
6 Verification Methods

6.2 The Krawczyk Operator

To motivate the definition of the Krawczyk operator, we consider a differentiable
function f : R → R. The mean-value theorem gives

f (y) − f (x(0) ) = f 0 (ξ) · (y − x(0) ) for some ξ between x(0) and y. (6.6)

If f (y) = 0, then
0 = f (x(0) ) + f 0 (ξ) · (y − x(0) ). (6.7)
Multiplying (6.7) by some r 6= 0, we get

0 = r · f (x(0) ) + r · f 0 (ξ) · (y − x(0) ). (6.8)

In order to transform (6.8) into a fixed point equation, we add and subtract the
term y − x(0) , getting

0 = r · f (x(0) ) + (r · f 0 (ξ) − 1) · (y − x(0) ) + y − x(0) .

Thus, y = x(0) − r · f (x(0) ) + (1 − r · f 0 (ξ)) · (y − x(0) ). Now, let [x] ∈ IR be

an interval with y ∈ [x] and x ∈ [x]. Then, also ξ ∈ [x]. Suppose the interval
evaluation of f 0 ([x]) is possible as defined in Section 6.1. Then, f 0 (ξ) ∈ f 0 ([x]).
The Krawczyk operator is then defined as

K(x(0) , r, [x]) = x(0) − r · f (x(0) ) + (1 − r · f 0 ([x])) · ([x] − x(0) ),

where the interval computations are defined as in Section 6.1.

It is often useful to have a multi-dimensional generalization of the Krawczyk
operator along with a corresponding generalized method of interval computation.
The starting point above was the mean-value theorem. This theorem is only valid
in the 1-dimensional case. However, the Krawczyk operator can also be defined
for functions f : Rn → Rn as follows.
Again, we assume that f is differentiable. Then,

f (y) − f (x(0) ) = B(x(0) , y) · (y − x(0) ), (6.9)

where
 ∂f1


∂f1



∂x1 x(0) + λ1 · y − x(0) ··· ∂xn x(0) + λ1 · y − x(0)
B(x(0) , y) = 
 .. .. 
.

.



∂fn ∂fn
∂x1 x(0) + λn · y − x(0) ··· ∂xn
(0)
x + λn · y − x (0)

1
with some λ1 , ..., λn ∈ (0, 1). Note that, in general, the λi will all be distinct.
1 See Section 3.2 in (68).

62
 6.2 The Krawczyk Operator

Now, let R ∈ Rn×n be a nonsingular matrix, let I denote the unit matrix, let
[x] ∈ IRn×n be an interval vector, let y ∈ [x], and let x(0) ∈ [x] be arbitrary but
∂fi
fixed. If ∂x j
([x]) can be evaluated for i, j = 1, ..., n as described in Section 6.1,
then we have
B(x(0) , y) ∈ f 0 ([x]),
where  
∂f1 ∂f1
∂x1 ([x]) ··· ∂xn ([x])
f 0 ([x]) = 
 .. .. 
. . ,
∂fn ∂fn
···
∂x1 ([x]) ∂xn ([x])


because x(0) + λi · y − x(0) ∈ [x] for i = 1, ..., n, since [x] is convex. As a result,
the Krawczyk operator can be defined as

K(x(0) , R, [x]) = x(0) − R · f (x(0) ) + (I − R · f 0 ([x])) · ([x] − x(0) ).

Another technique for finding a matrix B(x(0) , y) satiyfying (6.9) goes back to
E. Hansen. For this, it is sufficient that the function f is continuous. Because
of  
(0)
fi (y1 , . . . , yn ) − fi x1 , . . . , x(0) n =
   
(0) (0)
fi (y1 , . . . , yn ) − fi x1 , y2, . . . , yn + fi x1 , y2, . . . , yn
   
(0) (0) (0) (0)
−fi x1 , x2 , y3 , . . . , yn + fi x1 , x2 , y3 , . . . , yn
     
(0) (0) (0) (0) (0)
± · · · − fi x1 , . . . , xn−1 , yn + fi x1 , . . . , xn−1 , yn − fi x1 , . . . , x(0) n =
   
n fi x(0) , . . . , x(0) , yj , . . . , yn − fi x(0) , . . . , x(0) , yj+1 , . . . , yn  
X 1 j−1 1 j (0)
(0)
· y j − x j
j=1 yj − xj

the matrix H(x(0) , y) with H(x(0) , y)ij :=

    
(0) (0) (0) (0)
 fi x1 , . . . , xj−1 , yj , . . . , yn − fi x1 , . . . , xj , yj+1 , . . . , yn

 (0)
(0)
, yj 6= xj
 yj − xj

 (0)
cij , yj = xj
0
fulfills f (y) − f (x(0) ) = H(x(0) , y) · (y− x(0) ) for any choice of cij ∈ R. If f (x(0) )
∂fi (0) (0)
exists, it is natural to set cij := ∂xj x1 , . . . , xj , yj+1 , . . . , yn . If there exists
an interval extension H(x(0) , [x]) as described in Section 6.1, then the Krawczyk
operator can be defined as

K(x(0) , R, [x]) = x(0) − R · f (x(0) ) + (I − R · H(x(0) , [x])) · ([x] − x(0) ).

63
6 Verification Methods

The difference between f 0 ([x]) and H(x(0) , [x]) will be presented in the following
example.
Example 6.7. Let f (x) = x2 , x ∈ [x] ⊆ R. Then,

f (y) − f (x(0) ) = y 2 − (x(0) )2 = (y + x(0) ) · (y − x(0) )

and Hansen’s method yields

H(x(0) , [x]) = [x] + x(0) .

Considering simple differentiation one gets

f 0 (x) = 2 · x and f 0 ([x]) = 2 · [x].

For example, if [x] = [3, 5] and x(0) = 4, then Hansen’s method yields

H(x(0) , [x]) = [3, 5] + 4 = [7, 9],

whereas simple differentiation leads to

f 0 ([x]) = 2 · [3, 5] = [6, 10],

a result less precise than that of Hansen’s result. 

6.3 The Moore Test

Let f : D → Rn , D ⊆ Rn be continuous. In 1977, R. Moore published a method
for testing/proving the guess, if within a given interval vector [x] ∈ IRn , [x] ⊆ D,
there exists a zero of f . It is based on the Krawczyk operator and on Brouwer’s
fixed point theorem.
Assume that we know an interval (slope) matrix Y (x0 , [x]), which fulfills the
following: let x(0) ∈ [x] ⊆ D be fixed. Then, for any y ∈ [x] there exists a
matrix Y (x(0) , y) ∈ Y (x(0) , [x]) so that

f (y) − f (x(0) ) = Y (x(0) , y) · (y − x(0) ). (6.10)

In Section 6.2 we have presented two possibilities for finding such an interval
matrix. Independently of the choice of Y (x(0) , [x]), the Krawczyk operator is
defined as

K(x̂, R, [x]) = x̂ − R · f (x̂) + (I − R · Y (x̂, [x])) · ([x] − x̂)

where I denotes the unit matrix, R ∈ Rn×n is a nonsingular matrix, and where
x̂ ∈ [x] is arbitrary, but fixed.

64
 6.4 The Moore-Kioustelidis Test

Theorem 6.1. (Moore, 1977) If

K(x̂, R, [x]) ⊆ [x], (6.11)

then there exists some x∗ ∈ [x] with f (x∗ ) = 0.

Proof: Multiplying (6.10) with R, we get after some calculations

−R · f (y) = −R · f (x̂) − R · Y (x̂, y) · (y − x̂),

whence
y − R · f (y) = x̂ − R · f (x̂) + (I − R · Y (x̂, y)) · (y − x̂).
Considering the continuous function

r(y) := y − R · f (y), y ∈ Rn ,

for any y ∈ [x] we have

r(y) = x̂ − R · f (x̂) + (I − R · Y (x̂, y)) · (y − x̂)

∈ x̂ − R · f (x̂) + (I − R · Y (x̂, [x])) · ([x] − x̂)
= K(x̂, R, [x]) ⊆ [x].

According to Brouwer’s fixed point theorem, there exists some x∗ ∈ [x] with
r(x∗ ) = x∗ . Since R is nonsingular, we have f (x∗ ) = 0. 

The check for the validity of (6.11) is called Moore test. A way to find an interval
vector [x], for which it is likely to verify (6.11), has been published by G. Alefeld,
A. Gienger, and F. Potra in 1994.

6.4 The Moore-Kioustelidis Test

If Brouwer’s fixed point theorem can be used for verification methods, then so,
too, can the Poincaré-Miranda theorem. This was remarked by J. B. Kioustelidis
in 1978. Basically the idea is as follows.
Let [x] ∈ IRn , f : [x] → Rn be continuous, and let the opposite faces [x]i,± be
defined by
 T 
[x]i,+ := [x1 ], · · · , [xi−1 ], xi , [xi+1 ], · · · , [xn ] , 

 T i = 1, ..., n. (6.12)

[x]i,− := [x1 ], · · · , [xi−1 ], xi , [xi+1 ], · · · , [xn ] , 

65
6 Verification Methods

Applying the Poincaré-Miranda theorem, one could try to verify for all i =
1, ..., n
sup fi ([x]i,+ ) ≤ 0 ≤ inf fi ([x]i,− ) (6.13)
or
sup fi ([x]i,− ) ≤ 0 ≤ inf fi ([x]i,+ ), (6.14)
using the interval extensions explained in Section 6.1. But it is very unlikely
that this works as we will see in the following example.

Example 6.8. Let f : R2 → R2 be defined as

   
f1 (x, y) x·y
f (x, y) = = .
f2 (x, y) x+y−1

A zero is given by (x∗ , y ∗ )T = (1, 0)T . However, for any interval vector [x] ∈ IR2
with (x∗ , y ∗ )T as its center, neither (6.13) nor (6.14) can be valid. To show this,
let  
[1 − ε, 1 + ε]
[x] = with arbitrarily small ε > 0, δ > 0.
[−δ, δ]
Then, for any
   
x 1,+ 1+ε
∈ [x] =
y [−δ, δ]
we get
f1 (x, y) = f1 (1 + ε, y) = (1 + ε) · y, y ∈ [−δ, δ].
On the other hand, for any
   
x 1−ε
∈ [x]1,− =
y [−δ, δ]

we get
f1 (x, y) = f1 (1 − ε, y) = (1 − ε) · y, y ∈ [−δ, δ].
Thus,
δ δ
f1 (1 + ε, ) · f1 (1 − ε, ) > 0.
2 2
Therefore, the Poincaré-Miranda theorem cannot be applied. 

The idea of Kioustelidis was to apply the Poincaré-Miranda theorem not to f ,

but to some auxiliary function g. If f is differentiable, then

f (x) ≈ f (xc ) + f 0 (xc ) · (x − xc ).

66
 6.4 The Moore-Kioustelidis Test

Furthermore, if f (xc ) ≈ 0 and if f 0 (xc ) is nonsingular, then

 −1
x − xc ≈ f 0 (xc ) · f (x).
 −1
Setting g(x) := f 0 (xc ) · f (x) we get
)
gi (x) ≈ (x − xc )i = xi − xci ≤ 0 if x ∈ [x]i,−
i = 1, ..., n.
gi (y) ≈ (y − xc )i = xi − xci ≥ 0 if y ∈ [x]i,+
This is exactly the behaviour that we need in order to apply one of the corollaries
of the Poincaré-Miranda theorem.
Using a variation of this idea, R. Moore and J. B. Kioustelidis established the
following.
Theorem 6.2. (Moore and Kioustelidis, 1980) Let x̂, s, t ∈ Rn , s ≥ 0,
t ≥ 0, and [x] = [x̂ − s, x̂ + t] ∈ IRn . For i = 1, ..., n the interval vectors [x]i,+
and [x]i,− are defined as in (6.12). Let f : D → Rn , D ⊆ Rn be continuous with
[x] ⊆ D. Assume, that an interval (slope) matrix Y (x0 , [x]) is known, which
fulfills the following. Let x(0) ∈ [x] be fixed. Then, for any y ∈ [x] there exists a
matrix Y (x(0) , y) ∈ Y (x(0) , [x]) so that (6.10) holds. Let ei denote the ith unit
vector, R ∈ Rn×n be nonsingular, g(x) := R · f (x), and
Pn  
[l]i,+ := gi (x̂ + ti ei ) + R · Y (x̂ + ti ei , [x]i,+ ) · [−sj , tj ],
j=1, j6=i ij
Pn  
[l]i,− := gi (x̂ − si ei ) + R · Y (x̂ − si ei , [x]i,− ) · [−sj , tj ].
j=1, j6=i ij

If for i = 1, ..., n it holds that

sup[l]i,− ≤ 0 ≤ inf[l]i,+ (6.15)
or
sup[l]i,+ ≤ 0 ≤ inf[l]i,− , (6.16)
∗ ∗
then there exists some x ∈ [x] with f (x ) = 0.

Proof: Let i ∈ {1, ..., n} be arbitrary, but fixed. Setting x(0) = x̂ + ti ei in (6.10),
then for every x ∈ [x]i,+ it holds
f (x) = f (x̂ + ti ei ) + Y (x̂ + ti ei , x) · (x − x̂ − ti ei ).
After multiplying by R, we get
g(x) = g(x̂ + ti ei ) + (R · Y (x̂ + ti ei , x)) · (x − x̂ − ti ei )

∈ g(x̂ + ti ei ) + (R · Y (x̂ + ti ei , [x]i,+ )) · ([x]i,+ − x̂ − ti ei ).

67
6 Verification Methods

Since  
[−s1 , t1 ]
 .. 

 . 

 
 [−si−1 , ti−1 ] 
i,+
 
[x] − x̂ − ti ei = 
 0 ,

 [−si+1 , ti+1 ] 
 
 
 .. 
 . 
[−sn , tn ]
it follows that gi (x) ∈ [l]i,+ for all x ∈ [x]i,+ . Analogously, we get gi (y) ∈ [l]i,−
for all y ∈ [x]i,− . But (6.15) and (6.16) assure us that

gi (x) · gi (y) ≤ 0 for all x ∈ [x]i,+ and for all y ∈ [x]i,− .

By Corollary 4.2 it follows that there exists some x∗ ∈ [x] with g(x∗ ) = 0. Since
R is nonsingular, f (x∗ ) = 0 as well. 

The check for validity of (6.15) and (6.16) is called Moore-Kioustelidis test. A
comparison to the Moore test will be given in Section 6.6. Beforehand we present
another verification test also based on the Poincaré-Miranda theorem.

6.5 The Frommer-Lang-Schnurr Test

Theorem 6.3. (Frommer, Lang, and Schnurr, 2004) Let [x] ∈ IRn and
for i = 1, ..., n let the interval vectors [x]i,+ and [x]i,− be defined as in (6.12).
Let f : D → Rn , D ⊆ Rn be continuous with [x] ⊆ D. Assume, that an interval
(slope) matrix Y (x(0) , [x]) satisfies: Let x(0) ∈ [x] be fixed. Then, for any y ∈ [x]
there exists a matrix Y (x(0) , y) ∈ Y (x(0) , [x]) so that (6.10) holds. Let x̂ ∈ [x],
R ∈ Rn×n be nonsingular, g(x) := R · f (x), and
n 
P 
[m(x̂)]i,+ := gi (x̂) + R · Y (x̂, [x]) · ([xj ]i,+ − x̂j ),
j=1 ij
Pn  
[m(x̂)]i,− := gi (x̂) + R · Y (x̂, [x]) · ([xj ]i,− − x̂j ).
j=1 ij

If for i = 1, ..., n
sup[m(x̂)]i,− ≤ 0 ≤ inf[m(x̂)]i,+ (6.17)
or
sup[m(x̂)]i,+ ≤ 0 ≤ inf[m(x̂)]i,− , (6.18)
∗ ∗
then there exists some x ∈ [x] with f (x ) = 0.

68
 6.6 A Comparison of the Tests

Proof: Let i ∈ {1, ..., n}. According to (6.10) we get

gi (x) ∈ [m(x̂)]i,+ for all x ∈ [x]i,+

and
gi (y) ∈ [m(x̂)]i,− for all y ∈ [x]i,− .
Calling on Corollary 4.2, (6.17), and (6.18) we conclude that there exists some
x∗ ∈ [x] with g(x∗ ) = 0. Since R is nonsingular, it follows that f (x∗ ) = 0.

The check for the validity of (6.17) and (6.18) is called the Frommer-Lang-
Schnurr test.

6.6 A Comparison of the Tests

Theorem 6.4. The Frommer-Lang-Schnurr test is more powerful than the Moore
test, meaning that (6.11) implies (6.17).

Proof: Suppose that the Moore test is successful. Then, considering the ith
component one obtains
 
xi ≤ inf K(x̂, R, [x])i = inf x̂ − R · f (x̂) + (I − R · Y (x̂, [x])) · ([x] − x̂) .
i

Since g(x) = R · f (x) we have

n
X
K(x̂, R, [x])i = x̂i − gi (x̂) + (I − R · Y (x̂, [x]))ij · ([xj ] − x̂j ),
j=1

and since for j 6= i we have [xj ]i,− = [xj ], we can conclude that K(x̂, R, [x])i =
n
X
x̂i − gi (x̂) + (I − R · Y (x̂, [x]))ii · ([xi ] − x̂i ) − (R · Y (x̂, [x]))ij · ([xj ]i,− − x̂j ).
j=1,j6=i

Furthermore, due to

(I − R · Y (x̂, [x]))ii · ([xi ] − x̂i ) ⊇ (I − R · Y (x̂, [x]))ii · (xi − x̂i )

we have
inf K(x̂, R, [x])i ≤
 n
X 
inf x̂i −gi (x̂)+(I −R·Y (x̂, [x]))ii ·(xi − x̂i )− (R·Y (x̂, [x]))ij ·([xj ]i,− − x̂j ) .
j=1,j6=i

69
6 Verification Methods

From the distributive law [a] · γ + [b] · γ = ([a] + [b]) · γ holding for the scalar
γ = xi − x̂i and the fact that xi = [xi ]i,− it follows that
 
xi ≤ inf K(x̂, R, [x])i ≤ inf xi − [m(x̂)]i,− = xi − sup[m(x̂)]i,− .

Thus,
sup[m(x̂)]i,− ≤ 0.
In the same way, one can show that

sup K(x̂, R, [x])i ≤ xi

implies 0 ≤ inf[m(x̂)]i,+ . 

The Moore test and the Frommer-Lang-Schnurr test require roughly the same
computational work. Therefore, the Frommer-Lang-Schnurr test is more power-
ful than the Moore test by Theorem 6.4.
Comparing the Moore-Kioustelidis test with the Frommer-Lang-Schnurr test,
it is obvious that the latter requires less computational work, since there the
function g(·) and the interval (slope) matrix Y (·, ·) are only evaluated once.
However, the following theorem will present a special situation, where the Moore-
Kioustelidis test is more powerful than the Frommer-Lang-Schnurr test.
Theorem 6.5. Let f : D → Rn , D ⊆ Rn be differentiable and let [x] ⊆ D.
Furthermore, let
Y (x̂, [x]) = f 0 ([x]).
Then, concerning Theorems 6.3 and 6.2, we have

[l]i,+ ⊆ [m(x̂)]i,+ and [l]i,− ⊆ [m(x̂)]i,− . (6.19)

This means, if (6.17) from Theorem 6.3 is valid, then the condition (6.15) from
Theorem 6.2 is fulfilled; analogously, if (6.18) from Theorem 6.3 is valid, then
the condition (6.16) from Theorem 6.2 is fulfilled.

Proof: Let i ∈ {1, ..., n}. According to (6.10) we have

f (x̂ + ti ei ) − f (x̂) = Y (x̂, x̂ + ti ei ) · (x̂ + ti ei − x̂).

After multiplying by R, we get

g(x̂ + ti ei ) = g(x̂) + (R · Y (x̂, x̂ + ti ei )) · ti ei ,

whence
   
gi (x̂ + ti ei ) ∈ gi (x̂) + ti · R · Y (x̂, [x]) = gi (x̂) + ti · R · f 0 ([x]) . (6.20)
ii ii

70
 6.6 A Comparison of the Tests

Since, in general, we have for [a], [b] ∈ IR and any γ ∈ R, that γ ∈ [a] implies
γ + [b] ⊆ [a] + [b], we get from (6.20)
n
X  
gi (x̂ + ti ei ) + R · f 0 ([x]) · [−sj , tj ] ⊆ [m(x̂)]i,+ . (6.21)
ij
j=1, j6=i

Using f 0 ([x]i,+ ) ⊆ f 0 ([x]) we finally get [l]i,+ ⊆ [m(x̂)]i,+ . Analogously, one can
show that [l]i,− ⊆ [m(x̂)]i,− .
The following figure gives a summary for the three verification methods pre-
sented, where A ⇒ B means that the success of test A implies the success of
test B.

F rommer f’([x]) M oore

M oore =⇒ Lang =⇒ Kioustelidis
Schnurr

The following example will present a counterexample concerning the second

implication when Hansen’s slope matrix H(x̂, [x]) is used as Y (x̂, [x]) instead of
f 0 ([x]).

Example 6.9. (Schnurr, 2005) We consider the function

! !
f1 (x1 , x2 ) (4x1 + 0.1)x22 − (x1 + 0.025)x21 + 0.0525x1
f (x1 , x2 ) = =
f2 (x1 , x2 ) x2 + 0.01

and a point x̂ = (x̂1 x̂2 )T . The Jacobian matrix is

!
0 4x22 − 3x21 − 0.05x1 + 0.0525 2x2 (4x1 + 0.1)
f (x1 , x2 ) = ,
0 1

whereas the slope matrix H(x̂, y) of Hansen’s method is given by

f1 (y1 , y2 ) − f1 (x̂1 , y2 )
=
y1 − x̂1

4y22 − y12 − y1 x̂1 − x̂21 − 0.025 · (y1 + x̂1 ) + 0.0525,

f1 (x̂1 , y2 ) − f1 (x̂1 , x̂2 )
= (4x̂1 + 0.1) (y2 + x̂2 ) ,
y2 − x̂2
f2 (y1 , y2 ) − f2 (x̂1 , y2 ) f2 (x̂1 , y2 ) − f2 (x̂1 , x̂2 )
= 0, = 1.
y1 − x̂1 y2 − x̂2

71
6 Verification Methods

Concerning [x] = ([−0.1, 0.1] [−0.1, 0.1])T and x̂ = 0, we compare the Moore
test with the Moore-Kioustelidis test. In both tests we use
 0 −1  1
0

R = f (0) = 0.0525 .
0 1
Then,
f1 (y1 , y2 ) − f1 (0, y2 ) 2
= 4y22 − (y1 + 0.0125) + 0.0525 + 0.01252.
y1 − 0
Therefore, we have
   
0.04 , 0.0925 + 0.01252 [−0.01, 0.01]
Y (0, [x]) =
0 1
and
h i  0.01 0.01  !
0.0925+0.01252 0.04
1− 0.0525 ,1 − 0.0525 − 0.0525 , 0.0525
I − R · Y (0, [x]) =
0 0
 
[−0.77 , 0.24] [−0.2, 0.2]
⊆ .
0 0
Hence,
 
[−0.1, 0.1]
K (0, R, [x]) = 0 − R · f (0) + (I − R · Y (0, [x])) ·
  [−0.1, 0.1]
[−0.077, 0.077] + [−0.02, 0.02]
⊆ ⊆ [x] .
−0.01
Therefore, the Moore test is successful, and due to Theorem 6.4 also the Frommer-
Lang-Schnurr test is successsful to verify that a zero x∗ of f is within [x]. On
the other hand, Theorem 6.2 says that
      
1,+ 0.1 0.1
[l] = R · f (0.1, 0) + R · Y ( , ) · [−0.1, 0.1]
1 0 [−0.1, 0.1] 12
 1   1  1
= · 4 · 10−3 + · 0.5 · [−0.1, 0.1] · [−0.1, 0.1] = [−2, 18].
0.0525 0.0525 105
Thus, the Moore-Kioustelidis test fails. 
Remark 6.1. Being aware of the fact that Brouwer’s fixed point theorem as
well as the Poincaré-Miranda theorem can be proved (very briefly) by using the
so-called degree of a mapping (see the book of J. Cronin and the paper of M.
N. Vrahatis, respectively), one might get the idea to verify the existence of a
zero directly by using the degree of a mapping and interval arithmetic. This was
done by A. Frommer, F. Hoxha, and B. Lang.

72
 6.7 Final Remarks

6.7 Final Remarks

To see the full power of interval computations, one has to use a computer.
There are several programming languages, that support interval arithmetic. For
example, PASCAL-XSC, C-XSC, and INTLAB. See, for example, the book of R.
Hammer, M. Hocks, U. Kulisch, and D. Ratz, the book of R. Klatte, U. Kulisch,
A. Wiethoff, C. Lawo, and M. Rauch, and the paper of S. Rump, respectively.
These programming languages are extensions of PASCAL, C, and Matlab, re-
spectively, by introducing another data type called interval. If [a], [b], and [c]
are declared as an interval, then the operations (6.2) are predefined and the
boundaries of [c] = [a] ∗ [b] with ∗ ∈ {+, −, ·, /} are automatically rounded in
the corresponding direction, for the case, that c and c have to be rounded. For
example, a + b is rounded downwardly and a + b is rounded upwardly to the
next machine number. This means, if a, a, b, and b are machine numbers, then
a + b ∈ [a] + [b] for all real a ∈ [a, a] and all real b ∈ [b, b].
Using interval arithmetic by a computer, one has to extend the ideas of the
preceding sections a little bit.
Example 6.10. Consider the function
√
g(x) = 3x − 2, x ∈ R.
We want to verify, that within the interval
√ [x] = [0.47, 0.48] there is a zero of g.
Since within a computer the number 2 cannot be represented, the test
g(0.47) · g(0.48) ≤ 0
√
cannot be done exactly by a computer. Therefore, the number 2 is included
in an interval; i.e., √
2 ∈ [1.4142, 1.4143],
and the function is extended to the function g(x; a) with a ∈ [1.4142, 1.4143],
x ∈ R. If it is possible to show that for all a ∈ [1.4142, 1.4143] it holds that
g(0.47; a) · g(0.48; a) ≤ 0,
then one can conclude
√ that for all a ∈ [1.4142, 1.4143] there is a zero in [x];
especially for a = 2. 

This means, actually we have to consider functions f (x1 , ..., xn ; a1 , ..., am ) with
(x1 , ..., xn )T ∈ D, where D is the domain of the function and where a1 ,...,am
are parameters, that can vary in given intervals. Indeed, it is easy to extend
the corresponding theorems to this generalized case. However, the hierarchy
presented in the preceding section is no longer true as we will see in the following
example.

73
6 Verification Methods

Example 6.11. We consider the continuous function

( )
2x − 1, x≥1
f (x; a) = a ∈ [a] = [2, 4].
a · x − a + 1, x < 1

Taking (
2x − 1, x≥1
f (x; [a]) =
[a] · x − [a] + 1, x < 1
and
Y ([x]; [a]) = convex hull of {2, [a]} = [2, 4],
we want to verify that within the interval [x] = [ 12 , 2] there is a zero of f .
Choosing R 6= 0 and x̂ = 1 we have by Theorem 6.3

[m(1)]+ = R · 1 + R · [2, 4] · (2 − 1) = R · [3, 5]

and
1
[m(1)]− = R · 1 + R · [2, 4] · ( − 1) = R · [−1, 0],
2
whereas by Theorem 6.2 we have
1
[l]− = R · ([a] · − [a] + 1) = R · [−2, 1].
2
This means, that the Frommer-Lang-Schnurr test is successful, whereas the
Moore-Kioustelidis test fails. 

For a hierachy in the general case we refer to (82). See also (80) for some
numerical examples.
For more examples where interval arithmetic is used to prove conjectures via
the computer we refer to (37).

74
7 Banach’s Fixed Point Theorem
Using Brouwer’s fixed point theorem one can attack the problem of showing the
existence of a fixed point. As already mentioned in Chapter 2 a fixed point
verified by Brouwer’s theorem is not necessarily unique. See Figure 2.2. The
question naturally arises how the assumptions can be strengthened so that a
verified fixed point is unique.
To get the idea, let x∗ be a fixed point of f : [a, b] → [a, b]. In order to guarantee
that x∗ is the only fixed point, the problem f (x) = x must not have a second
solution. If there was a second solution, say x̃, x̃ 6= x∗ it would follow that
x̃ − x∗ f (x̃) − f (x∗ )
1= ∗
= .
x̃ − x x̃ − x∗
If we assume f to be differentiable, the mean-value theorem says that there
exists ξ between x̃ and x∗ so that
f (x̃) − f (x∗ )
= f 0 (ξ).
x̃ − x∗
So, if f : [a, b] → [a, b] is differentiable satisfying

f 0 (x) < 1 for all x ∈ [a, b], (7.1)

then f has exactly one fixed point in [a, b].

How can this idea be extended to the n-dimensional case?
Since, for x̃ > x∗ ,
f (x̃) − f (x∗ )
= f 0 (ξ) < 1
x̃ − x∗
is equivalent to

f (x̃) − f (x∗ ) = f 0 (ξ) · (x̃ − x∗ ) < (x̃ − x∗ ),

a (more or less obvious) generalization to the n-dimensional case is

kf (x̃) − f (x∗ )k < kx̃ − x∗ k if x̃ 6= x∗ (7.2)

for some norm in Rn . (Note, however, that in (7.1) it reads f 0 (x) < 1 and not
|f 0 (x)| < 1.) So, trivially, we have the following theorem.

75
7 Banach’s Fixed Point Theorem

Theorem 7.1. Let K ⊆ Rn be nonenpty, convex, bounded, and closed. Fur-

thermore, let f : K → K satisfy

kf (x) − f (y)k < kx − yk (7.3)

for all x, y ∈ K, x 6= y for some norm in Rn . Then, f has exactly one fixed
point in K.

Proof: By (7.3) the function f is continuous. Therefore, Brouwer’s fixed point

theorem guarantees the existence of a fixed point, say x∗ . If there was another
fixed point, say x̃, with x̃ 6= x∗ , then it would follow due to (7.3)

kx̃ − x∗ k = kf (x̃) − f (x∗ )k < kx̃ − x∗ k

which is a contradiction. 

Of course, if the proof is so simple, there is no reason to name this result after
a mathematician. So, there must be more!
Indeed, the astonishing thing is (beyond the fact that the fixed point can be
approximated iteratively, see Theorem 7.2) the fact that the result is also true
in some vector spaces that can be infinite-dimensional. These vector spaces are
called Banach spaces. We will introduce them in the following section. In order
to emphasize this result, we mention that Brouwer’s fixed point theorem is not
valid in infinite-dimensional vector spaces.

7.1 Banach Spaces

Before we introduce some Banach spaces, we briefly remember the definition
of vector spaces and normed spaces. Already in school R2 was considered as a
vector space. For example,  
4
∈ R2 .
2


Then, 2 · 42 points in the same direction but with doubled length. Also the
vector addition is well known from school. For example
     
4 2 6
+ = ∈ R2 .
2 6 8

See also Figure 7.1.

76
 7.1 Banach Spaces

4 
6
 
2

8 6
3

1 
4

2
0
0 1 2 3 4 5 6 7
Figure 7.1: Vector addition visualized in R2

Let E = R2 , the basic observation for a generalization is the fact that

x ∈ E, y ∈ E ⇒ x + y ∈ E,
(7.4)
λ ∈ R, x ∈ E ⇒ λ · x ∈ E.

Trivially, (7.4) can be verified for E = Rn for all integers n ≥ 2. More important,
however, is the fact, that we can also consider the set of all real-valued continuous
functions with domain [a, b], denoted by C([a, b]), as E. If f, g ∈ C([a, b]) and
λ ∈ R, then also

f + g ∈ C([a, b]) and λ · f ∈ C([a, b])

by defining
(f + g)(x) := f (x) + g(x) for all x ∈ [a, b]
and
(λ · f )(x) := λ · f (x) for all x ∈ [a, b].
Satisfying (7.4) is not the only property that the sets Rn and C([a, b]) have in
common. The idea is to find as many common properties as possible and to
prove theorems only based on these properties. Formally, a real vector space

77
7 Banach’s Fixed Point Theorem

(sometimes also called a real linear space) is a nonempty set E fulfilling (7.4)
and
x + (y + z) = (x + y) + z for all x, y, z ∈ E,
x + y = y + x for all x, y ∈ E,
there exists 0 ∈ E so that x + 0 = x for all x ∈ E,
for all x ∈ E there exists −x ∈ E so that x + (−x) = 0,
λ · (x + y) = λ · x + λ · y for all x, y ∈ E and all λ ∈ R,
(λ + ν) · x = λ · x + ν · x for all x ∈ E and for all λ, ν ∈ R,
(λ · ν) · x = λ · (ν · x) for all x ∈ E and for all λ, ν ∈ R,
1 · x = x for all x ∈ E.
A real vector space E is called a normed vector space, if there is a function
k · k : E → R satisfying

kxk ≥ 0 and kxk = 0 ⇔ x = 0,

kλ · xk = |λ| · kxk
kx + yk ≤ kxk + kyk.

Now, we can give the definition of a Banach space.

Definition 7.1. A normed vector space (E, k·k) is called a Banach space, if any
so-called Cauchy sequence in E (see below) is convergent to a limit that belongs
to E; where a sequence {x(n) }∞n=1 is called a Cauchy sequence if the following
holds: For any ε > 0 there exists n0 (ε) so that

kx(m) − x(n) k < ε if m, n ≥ n0 (ε).

This definition is enough to extend Theorem 7.1.

Theorem 7.2. (Banach’s fixed point theorem) Let (E, k · k) be a Banach
space, and let D ⊆ E be nonempty and closed. If the function T : D → D
satisfies
kT (x) − T (y)k ≤ q · kx − yk for all x, y ∈ D (7.5)
with some q < 1, then within D there exists a unique fixed point x∗ of T .
Furthermore, for any x(0) ∈ D the so-called fixed point iteration

x(n+1) := T (x(n) ), n = 0, 1, 2, ...

converges to x∗ and
1 qn
kx(n) − x∗ k ≤ · kx(n+1) − x(n) k ≤ · kx(1) − x(0) k.
1−q 1−q

78
 7.1 Banach Spaces

Proof: First, we note that from (7.5) it follows that T is continuous. Then, we
choose arbitrarily x(0) ∈ D and consider the sequence

x(n+1) := T (x(n) ), n = 0, 1, 2, 3, ... .

This sequence is well-defined, since T is a self-mapping. From

kx(n+1) − x(n) k = kT (x(n) ) − T (x(n−1) )k ≤ q · kx(n) − x(n−1) k

we get by induction

kx(n+1) − x(n) k ≤ q n · kx(1) − x(0) k. (7.6)

Now, for any p > 0 we have kx(n+p) − x(n) k ≤

kx(n+p) − T (x(n+p) )k + kT (x(n+p) ) − T (x(n) )k + kT (x(n) ) − x(n) k

≤ kx(n+p) − T (x(n+p) )k + q · kx(n+p) − x(n) k + kT (x(n) ) − x(n) k
= kx(n+p) − x(n+p+1) k + q · kx(n+p) − x(n) k + kx(n+1) − x(n) k.

So, by (7.6) it follows

1 qp + 1
kx(n+p) −x(n) k ≤ (q n+p +q n )·kx(1) −x(0) k = ·kx(1) −x(0) k·q n ≤ C·q n
1−q 1−q
with
2
C= · kx(1) − x(0) k.
1−q
Hence, {x(n) }∞ ∗
n=0 is a Cauchy sequence. The latter has a limit, say x , since E
∗
is a Banach space, and x ∈ D, since D is closed. Due to the continuity of T we
have
T (x∗ ) = T ( lim x(n) ) = lim T (x(n) ) = lim x(n+1) = x∗ .
n→∞ n→∞ n→∞

This means, that x∗ ∈ D is a fixed point of T .

To show uniqueness, we assume that T has two fixed points, say x∗ and y ∗ .
Then, with (7.5) it follows

kx∗ − y ∗ k = kT (x∗ ) − T (y ∗ )k ≤ q · kx∗ − y ∗ k.

However, this can only be true if x∗ = y ∗ , since q < 1.

To get the related error estimate we consider

kx(n) − x∗ k ≤ kx(n) − x(n+1) k + kx(n+1) − x∗ k

= kx(n+1) − x(n) k + kT (x(n) ) − T (x∗ )k
≤ kx(n+1) − x(n) k + q · kx(n) − x∗ k.

79
7 Banach’s Fixed Point Theorem

Hence,
1
kx(n) − x∗ k ≤ · kx(n+1) − x(n) k,
1−q
and (7.6) finishes the proof. 

Banach’s fixed point theorem gets its full power, if one is aware of some Banach
spaces. We name just a few. (For a proof of the fact that these vector spaces
are indeed Banach spaces, we refer to any book on functional analysis).
• (Rn , k · k), where k · k can be any norm in Rn .
• (l2 , k · k), where l2 denotes the vector space of all square summable se-
quences of real numbers; i.e.,
∞
X
x= {xi }∞
i=1
2
∈ l :⇔ x2i < ∞,
i=1

and where v
u∞
uX
kxk := t x2 . i
i=1

• (l∞ , k · k), where l∞ denotes the vector space of all bounded sequences of
real numbers, and where
∞
kxk := sup |xi |.
i=1

• (c0 , k·k), where c0 denotes the vector space of all real sequences converging
to zero, and where
∞
kxk := sup |xi |.
i=1

• (C([a, b]), k·k∞ ), where C([a, b]) denotes the vector space of all real-valued,
continuous functions with domain [a, b], and where

kxk∞ := max{|x(t)| : t ∈ [a, b]}.

80
 7.2 Ordinary Differential Equations

7.2 Ordinary Differential Equations

The breakthrough of Banach’s fixed point theorem is in the field of ordinary
differential equations. This will be presented in the following example.
Example 7.1. Given ξ, η, a ∈ R and f : [ξ, ξ + a] × R → R. Then, the initial
value problem is to find a function y : [ξ, ξ + a] → R satisfying

y 0 (x) = f (x, y(x)), x ∈ [ξ, ξ + a]

y(ξ) = η.

A famous result by Picard-Lindelöf says that this problem has a unique solution,
if f (x, y) is continuous and if there exists L ≥ 0 so that

|f (x, y) − f (x, y)| ≤ L · |y − y|

is valid for all x ∈ [ξ, ξ + a] and for all y, y ∈ R. Its proof can be done by
using Banach’s fixed point theorem as follows. We consider the Banach space
C([ξ, ξ + a]) with the weighted norm

kykw := max{|y(x)| · e−2L·x , x ∈ [ξ, ξ + a]}.

Since f is continuous, y(x) is a solution of the initial value problem, if and only
if
Zx
y(x) = η + f (t, y(t))dt.
ξ

So, we define the function

T : C([ξ, ξ + a]) → C([ξ, ξ + a])

by
Zx
T (y)(x) := η + f (t, y(t))dt.
ξ

Then, the initial value problem has a unique solution, if and only if T has a
unique fixed point.

81
7 Banach’s Fixed Point Theorem

To verify the existence of a unique fixed point of T , let y, z ∈ C([ξ, ξ + a]). Then,
 
Rx   
 
|T (y)(x) − T (z)(x)| =  f (t, y(t)) − f (t, z(t)) dt
ξ 
Rx
≤ |f (t, y(t)) − f (t, z(t))| dt
ξ
Rx
≤ L · |y(t) − z(t)| dt
ξ
Rx
= L· |y(t) − z(t)| · e−2L·t · e2L·t dt
ξ
Rx
≤ L · ky − zkw · e2L·t dt
ξ
1
≤ L · ky − zkw · 2L · e2L·x .

It follows
L
|T (y)(x) − T (z)(x)| · e−2L·x ≤ · ky − zkw for all x ∈ [ξ, ξ + a].
2L
Hence,
1
kT (y) − T (z)kw ≤
· ky − zkw .
2
Now, Banach’s fixed point theorem can be applied, and therefore T has a unique
fixed point, which is then the unique solution of the initial value problem.
The iteration y (0) (x) = η and
Zx
(n+1) (n)
y (x) := T (y )(x) = η + f (t, y (n) (t))dt, n = 0, 1, 2, 3, ...
ξ

is called Picard iteration. It converges to the unique solution. 

Remark 7.1. Banach’s fixed point theorem is also called the Contraction-
Mapping Theorem, especially when E = Rn . It can also be used to prove
the Inverse Function Theorem and the Implicit Function Theorem. See (68), for
example. 

Remark 7.2. In (101) splitting methods were considered for solving linear
systems of equations, say Ax = b, where A ∈ Rn×n is nonsingular and b ∈ Rn .
Expressing the matrix A in the form A = M − N with M being nonsingular is
called a splitting of the matrix A.

82
 7.2 Ordinary Differential Equations

Let x∗ denote the unique solution of Ax = b, then we have

(M − N )x∗ = b,

which can be written equivalently as

x∗ = M −1 N x∗ + M −1 b.

Choosing x(0) ∈ Rn arbitrarily we can start the fixed point iteration

x(n+1) = M −1 N · x(n) + M −1 b, n = 0, 1, 2, ... . (7.7)

Special choices of M and N lead to the well-known Jacobi, Gauss-Seidel, and the
successive overrelaxation methods. Note that no fixed point theorem is needed,
because the unique solution is guaranteed by the nonsingularity of A and we
have

x(n+1) − x∗ = M −1 N · (x(n) − x∗ ) = ... = (M −1 N )n · (x(0) − x∗ ).

It can be shown (without any use of fixed point theory) that

lim (M −1 N )n = O, (where O denotes the zero matrix),

n→∞

if and only if the so-called spectral radius of M −1 N is smaller than 1. So,

the proof of convergence for these splitting methods reduces to the calculation
of a spectral radius. Therefore, Theorem 7.2 is associated with these splitting
methods only due the fixed point iteration (7.7), but not due to the proof of
existence and uniqueness of a fixed point.
However, we refer to (7) where a generalization of these splitting methods to
interval computations is done. There, the context to Theorem 7.2 is obvious
and necessary. See also (8), (21), and Chapter 13 in (68). 

83
8 Schauder’s Fixed Point Theorem
In this chapter we deal with the question if the Brouwer fixed point theorem
can be extended from Rn to other normed linear spaces. First we consider
n-dimensional normed linear spaces, second we consider infinite-dimensional
normed linear spaces.
An n-dimensional normed linear space (different from Rn ) is presented in the
following example.
Example 8.1. Let f : [0, 1] → R defined by f (x) = 1, let g : [0, 1] → R defined
by g(x) = x, and let h : [0, 1] → R defined by f (x) = x2 . Then every quadratic
function from [0, 1] to R can be expressed as a linear combination of f, g, and
h. We may write
X = span{f, g, h}.
So, using the maximum norm k · k∞ defined in Chapter 7.1 (X, k · k∞ ) is a three-
dimensional normed linear space. For instance, any solution of the ordinary
differential equation
y 000 (x) = 0, x ∈ [0, 1]
belongs to X. 

As we will see in the following theorem any n-dimensional normed linear space
and Rn are more or less the same.
Theorem 8.1. Let (X, k · kX ) be an n-dimensional normed linear space. Then,
X and Rn are isomorphic; i.e., there exists a linear bijection T : X → Rn so
that both T and T −1 are continuous.

Proof: Let f (1) , f (2) , ..., f (n) be a basis of X. Then, for every f ∈ X there exist
uniquely defined x1 , x2 , ..., xn ∈ R so that
n
X
f= xj · f (j) .
j=1
n
Defining T : X → R by  
x1
T (f ) =  ... 
 

xn

85
8 Schauder’s Fixed Point Theorem

we have a linear bijection with T −1 : Rn → X defined by

 
x1 n
−1  ..  X
T  . = xj · f (j) .
xn j=1

To show continuity of T and T −1 we consider the norm k · k on Rn defined by

 
x1
kxk := kT −1 (x)kX where x =  ...  .
 

xn
By Theorem 2.2 this norm is equivalent to the Euclidean norm on Rn . It follows
that there exist positive numbers α and β so that
α · kxke ≤ kxk ≤ β · kxke for all x ∈ Rn .
In other words, for x, y ∈ Rn and f, g ∈ X we have
kT −1 (x) − T −1 (y)kX = kT −1 (x − y)kX = kx − yk ≤ β · kx − yke
and
1 1
kT (f ) − T (g)ke = kT (f − g)ke ≤ · kT (f − g)k = · kf − gkX .
α α
So, T as well as T −1 are (even Lipschitz) continuous. 

Concerning Brouwer’s fixed point theorem we get the following consequence.

Corollary 8.1. Let (X, k · kX ) be an n-dimensional normed linear space and let
K ⊆ X be nonempty, bounded, closed, and convex. Let f : K → X be continuous
satisfying f (K) ⊆ K, then there exists x∗ ∈ K with f (x∗ ) = x∗ .

Proof: By Theorem 8.1 X and Rn are isomorphic; i.e., there exists a linear
bijection T : X → Rn so that both T and T −1 are continuous. Therefore,
K̃ := T (K) is nonempty, bounded, closed (due to the continuity of T ), and
convex (due to the linearity of T ) in Rn . Defining
g := T ◦ f ◦ T −1
we get g : K̃ → K̃ and we get that g is continuous. By Brouwer’s fixed point
theorem there exists y ∗ ∈ K̃ satisfying g(y ∗ ) = y ∗ . Then, x∗ := T −1 (y ∗ ) ∈ K is
a fixed point of f . 

Brouwer’s fixed point theorem is not true, if one just substitutes an infinite-
dimensional normed linear space (even a very regular one) for Rn . This will be
shown in the following example.

86
 8 Schauder’s Fixed Point Theorem

Example 8.2. (Kakutani, 1943) Consider the space X = l2 (Z) of all doubly-
infinite real-valued sequences {xn }∞
n=−∞ (Z denotes the set of all integers) so
that
X∞
x2n < ∞;
n=−∞

equipped with the norm v

u ∞
u X
kxk2 := t x2n .
n=−∞

(X, k · k2 ) is a Banach space. Furthermore, the closed unit ball

B = {x ∈ X : kxk2 ≤ 1}
is a nonempty, closed, bounded, and convex set in X. Let S(x) denote the
shifted sequence of x, that is

(S(x))n = xn+1 for n ∈ Z;

so, for instance, if x = e0 is the sequence with 0’s in each coordinate except the
0th coordinate where it has 1 for an entry, then S(x) = e1 , the sequence with a
1 in the 1th entry and 0’s elsewhere.
It is plain and easy to see that S takes X onto itself in a linear, isometric manner:
v v
u X u X
u ∞ u ∞
kS(x)k2 = t 2
xn+1 = t x2n = kxk2 .
n=−∞ n=−∞

With Kakutani, define Φ : X → X by

1
Φ(x) = S(x) + (1 − kxk2 ) · e0 .
2
Φ takes B into B: after all, if kxk2 ≤ 1, then
kΦ(x)k2 ≤ kS(x)k2 + k 21 (1 − kxk2 ) · e0 k2

= kxk2 + 21 (1 − kxk2 ) = 1
2 + 1
2 · kxk2 ≤ 1.

However, Φ is fixed point free in B. Indeed, imagine x∗ ∈ X satisfies Φ(x∗ ) = x∗ .

Then
1
x∗ − S(x∗ ) = (1 − kx∗ k2 ) · e0 . (8.1)
2
Can x∗ = 0, where 0 denotes the zero sequence? If so, then (8.1) says
1 1
0= (1 − k0k2 ) · e0 = · e0 ,
2 2

87
8 Schauder’s Fixed Point Theorem

which is not so. So x∗ 6= 0. But this says that

1
x∗ − S(x∗ ) = (1 − kx∗ k2 ) · e0
2
and by the very definition of S(x∗ ) we have

x∗n = x∗n+1 for any n 6= 0.

This means that

... = x∗−2 = x∗−1 = x∗0 , x∗1 = x∗2 = x∗3 = ...

Since x∗ 6= 0, at least one of the numbers x∗0 and x∗1 , respectively, is not 0, in
∞
P
plain contradiction to (x∗n )2 < ∞. 
n=−∞

To get a fixed point theorem in infinite-dimensional spaces one has to assume

more. This was done by Juliusz Schauder and it will be presented in the follow-
ing.
Before, we consider some properties of sets. Some of them depend on the di-
mension, some of them do not. For example, a set K is called to be convex, if
it holds that

x, y ∈ K ⇒ λ · x + (1 − λ) · y ∈ K for all λ ∈ (0, 1).

This definition does not depend on the dimension of K. Next, we consider open
sets. Let (E, k · k) be a normed vector space and let Q ⊆ E. Q is said to be
open, if for any x ∈ Q there exists some ε > 0 so that

U (x; ε) := {y ∈ E : kx − yk < ε}

is a subset of Q. This definition is also independent from the dimension of E.

So, let’s start with some examples, where differences occur.
Example 8.3. Let L = {Li }∞ 2
i=1 ∈ l and let

Ω = {x ∈ l2 : |xi | < Li , i = 1, 2, 3, ...}.

On first sight one may think that Ω is an open set. But this is not the case as
we will show now.
Let ε > 0 be arbitrary but fixed. Since L ∈ l2 , the sequence {Li }∞
i=1 has to be
convergent to 0; i.e., there exists i0 ∈ N so that
ε
Li < for all i ≥ i0 .
4

88
 8 Schauder’s Fixed Point Theorem

Let x ∈ Ω. Then, we define ξ = {ξi }∞

i=1 by


 xi , i 6= i0
ξi := 3·ε

 + xi0 , i = i0 .
4
Considering v
u∞
uX 3·ε
kx − ξk = t (xi − ξi )2 = < ε,
i=1
4

we can conclude that

ξ ∈ U (x; ε) := {y ∈ l2 : kx − yk < ε}.

However, it holds
 
3 · ε  3·ε 3·ε ε ε
|ξi0 | =  + xi0  ≥ − |xi0 | ≥ − = > Li0 ,
4 4 4 4 2

which means that ξ 6∈ Ω. Therefore, regardless of ε > 0

U (x; ε) 6⊂ Ω;

i.e., Ω is not an open set. 

A famous theorem by Bolzano-Weierstraß says that any sequence in a bounded

subset of Rn has a convergent subsequence. This conclusion is not true in
infinite-dimensional normed linear spaces as we will see in the following ex-
ample.

Example 8.4. Let (X, k · k) be an infinite-dimensional normed linear space. We

consider the unit ball

U (x0 ; 1) := {x ∈ X : kx − x0 k ≤ 1}.

We will show that there exists a sequence {x(n) }∞ n=1 with kx

(n)
k = 1, n =
(m)
1, 2, 3, ... satisfying kx − x k ≥ 1, if m 6= k. To achieve this, choose x(1) ∈ X
(k)

so that kx(1) k = 1 and define

X1 := {a · x(1) : a ∈ R}.

Since X is infinite-dimensional, there exists some y ∈ X, y 6∈ X1 with

ky − xk > 0 for all x ∈ X1 . (8.2)

89
8 Schauder’s Fixed Point Theorem

Defining the continuous function ϕ : R → R by

ϕ(a) = ky − a · x(1) k,

we have
lim ϕ(a) = ∞
|a|→∞

since ky − a · x(1) k ≥ |a| · kx(1) k − kyk. Therefore, there exists a1 ∈ R with

ϕ(a1 ) = min ϕ(a) = min ky − a · x(1) k. (8.3)

a∈R a∈R

In addition, ky − a1 · x(1) k > 0 due to (8.2). Setting

1
x(2) := · (y − a1 · x(1) )
ky − a1 · x(1) k

we get kx(2) k = 1 and

(2) (1)
1 (1)

(1)
kx −x k = ky − a · x(1) k · (y − a1 · x ) − x

1

1
(1) (1) (1)
= · (y − a 1 · x ) − ky − a 1 · x k · x
ky − a1 · x(1) k
1
(1) (1)
= · y − (a 1 + ky − a 1 · x k) · x ≥1
ky − a1 · x(1) k

thanks to (8.3). Now, suppose x(1) , ..., x(n) ∈ X have been chosen so that

kx(1) k = kx(2) k = ... = kx(n) k = 1

and
kx(i) − x(j) k ≥ 1 for i 6= j, i = 1, 2, ..., n, j = 1, 2, ..., n.
Then, we define
n
X
Xn := {x = tj · x(j) : t1 , ..., tn ∈ R}.
j=1

Since X is infinite-dimensional, there exists z ∈ X, but z 6∈ Xn with

kz − xk > 0 for all x ∈ Xn . (8.4)

Again, we define a continuous function ϕ : Rn → R by

n
X
ϕ(t1 , ..., tn ) := kz − tj · x(j) k.
j=1

90
 8 Schauder’s Fixed Point Theorem

Then,
n
X n
X n
X
kz − tj · x(j) k = k tj · x(j) − zk = kti · x(i) − (z − tj · x(j) )k.
j=1 j=1 j=1, j6=i

Hence,
n
X
(i)
ϕ(t1 , ..., tn ) ≥ |ti | · kx k − kz − tj · x(j) k
j=1, j6=i
which implies that
lim ϕ(t1 , ..., tn ) = ∞ for all i ∈ {1, ..., n}. (8.5)
|ti |→∞

Let M > 0 be arbitrary but fixed and let

K := {t ∈ Rn : ktke ≤ M }.
Then the function ϕ : K → R has a minimum and a maximum, since K is
compact and since ϕ is continuous. Together with (8.5) it follows that the
function ϕ : Rn → R has a minimum, say
n
X n
X
ϕ(t∗1 , ..., t∗n ) = kz − t∗j · x(j) k = min kz − tj · x(j) k (8.6)
(t1 ,...,tn )T ∈Rn
j=1 j=1

and we have ϕ(t∗1 , ..., t∗n ) > 0 by (8.4). Setting

n
1 X
x(n+1) := n
P · (z − t∗j · x(j) )
kz − t∗j · x(j) k j=1
j=1

we get kx(n+1) k = 1 and for any i ∈ {1, ..., n} we have

n
(n+1) (i)
1 X
∗ (j)

(i)
kx −x k=
Pn · (z − tj · x ) − x =
kz −
t∗j · x(j) k j=1

j=1



n  n 
1 X
∗ (j) ∗
X
∗ (j) (i)
Pn · z − tj · x − ti + kz − tj · x k · x ≥ 1
∗ (j)
kz − tj · x k j = 1, j=1
j=1 j 6= i
thanks to (8.6). So, by induction, we have created a sequence {x(n) }∞
n=1 with
kx(n) k = 1, n = 1, 2, 3, ... satisfying
kx(m) − x(k) k ≥ 1, if m 6= k. (8.7)
This sequence cannot have a convergent subsequence. 

91
8 Schauder’s Fixed Point Theorem

8.1 Compactness
Definition 8.1. Let X be a normed vector space. A nonempty set K ⊆ X is
called compact, if every sequence in K has a subsequence that converges to a
limit that belongs to K.

If X = Rn , then the compact sets are precisely those that are closed and
bounded; this is the celebrated theorem of Bolzano-Weierstraß. But, as we saw
in Example 8.4, this is not so for the closed unit ball of an infinite-dimensional
normed linear space, a closed, bounded set to be sure. Compactness is more spe-
cial than being just closed and bounded; indeed, it is a key ingredient exploited
by Schauder in his fixed point theorem.
We start by establishing a fundamental feature of compact sets in normed vector
spaces - the property formulated is often referred to as ’total boundedness’.
Theorem 8.2. Let X be a normed vector space and let K ⊆ X be com-
pact. Then, for every ε > 0 there exists a finite set of elements of K, say
{v (1) , ..., v (p) }, so that for any y ∈ K there exists v (i) ∈ {v (1) , ..., v (p) } with
kv (i) − yk < ε.

Proof: Assume that K fails the conclusion. That is, assume, that there is some
ε0 > 0 so that for every finite set of elements of K, say {v (1) , ..., v (n) }, there
exists some y ∈ K with

kv (j) − yk ≥ ε0 for all v (j) ∈ {v (1) , ..., v (n) }.

Let x(1) ∈ K. Then, by assumption there is some x(2) ∈ K with

kx(1) − x(2) k ≥ ε0 . (8.8)

Considering the set {x(1) , x(2) } by assumption there is some x(3) ∈ K with

kx(j) − x(3) k ≥ ε0 for j = 1, 2.

Together with (8.8) we get

kx(j) − x(i) k ≥ ε0 for j 6= i, j = 1, 2, 3, i = 1, 2, 3.

By induction, we get a sequence {x(n) }∞

n=1 ⊆ K satisfying

kx(j) − x(i) k ≥ ε0 for j 6= i, j = 1, 2, 3, ......, i = 1, 2, 3, .... .

Such a sequence cannot have a convergent subsequence. As a result, K is not

compact. 

92
 8.1 Compactness

Before we present Schauder’s fixed point theorem in the next section, we want to
give an important, classical example of a compact set in an infinite-dimensional
normed linear space, the so-called Hilbert cube.
Example 8.5. Let X = l2 and let L = {Li }∞ 2
i=1 ∈ l . Then, the set

Ω = {x ∈ l2 : |xi | ≤ Li , i = 1, 2, 3, ...}

is compact. To show this, let

n o∞
(n) ∞
{x(n) }∞
n=1 = {xi }i=1
n=1

be a sequence in Ω.
(n)
i = 1 : {x1 }∞
n=1 ⊆ [−L1 , L1 ]. Due to the theorem of Bolzano-Weierstraß there
(ϕ (n)) ∞
is a subsequence {x1 1 }n=1 and some x∗1 ∈ [−L1 , L1 ] with
(ϕ1 (n))
lim x1 = x∗1 .
n→∞

i=2: ( !)∞  
(ϕ (n))
x1 1 [−L1 , L1 ]
(ϕ (n)) ⊆ .
x2 1 [−L2 , L2 ]
n=1

Again, due to the theorem of Bolzano-Weierstraß there is a subsequence ϕ2 (n)

of ϕ1 (n) and some x∗2 ∈ [−L2 , L2 ] with
(ϕ2 (n))
lim x2 = x∗2 .
n→∞

So, !
(ϕ (n))  
x1 2 x∗1
lim = .
n→∞ (ϕ (n))
x2 2 x∗2

Proceeding in this way we define some

x∗ = {x∗i }∞
i=1 ∈ Ω (8.9)

and for any fixed k we have

  
(ϕk (n))
x1 x∗1

lim  ..   . 
 =  . , (8.10)
n→∞  .  .
(ϕ (n))
xk k x∗k

where
ϕk (n) ⊆ ϕk−1 (n) ⊆ ... ⊆ ϕ2 (n) ⊆ ϕ1 (n) ⊆ N. (8.11)

93
8 Schauder’s Fixed Point Theorem

Defining the diagonal subsequence {x(ϕ(n)) }∞

n=1 of {x
(n) ∞
}n=1 by

ϕ(n) = {ϕ1 (1), ϕ2 (2), ϕ3 (3), ϕ4 (4), ...}

we get for any fixed k due to (8.11)

   
(ϕ(n)) (ϕ (n))
x1 x1 k

 .. 
⊆
 .. 
 for all n ≥ k. (8.12)
 .   . 
(ϕ(n)) (ϕ (n))
xk xk k

Now, we will show that the sequence {x(ϕ(n)) }∞ ∗

n=1 converges to x from (8.9).

Let ε > 0 be arbitrary, but fixed. Since L = {Li }∞ 2

i=1 ∈ l , there exists some
k ∈ N satisfying v
u ∞
u X
ε
t (2 · Lj )2 < .
2
j=k+1

Concerning this k, due to (8.10) and (8.12), there exists some n0 ∈ N so that
v
u k
uX (ϕ(n)) ε
t (x
j − x∗j )2 < for all n ≥ max{n0 , k}.
j=1
2

Then for all n ≥ max{n0 , k} we get

v
uX ∞
u k (ϕ(n)) X (ϕ(n))
kx(ϕ(n)) ∗
− x k = t (xj − x∗j )2 + (xj − x∗j )2
j=1 j=k+1
v v
u k
u ∞
u X
uX (ϕ(n)) ∗ 2 ε ε
≤ t (x
j − xj ) + t (2 · Lj )2 < + = ε.
j=1
2 2
j=k+1

That means, that {x(n) }∞

n=1 has a convergent subsequence and its limit belongs
to Ω by (8.9). Therefore, Ω is compact. 

Since continuity of a function f : K → K is not enough to ensure a fixed

point in infinite-dimensional vector spaces, one has to modify the assumptions
concerning f . The following definition will be sufficient.

94
 8.1 Compactness

Definition 8.2. Let (X, k · k) be a normed vector space and E ⊆ X. A function

(which is called an operator sometimes in infinite-dimensional spaces) T : E →
X is called compact, if the following two conditions are fulfilled.
1. T is continuous (with respect to the given norm);
2. For any bounded M ⊆ E the set T (M ) is compact according to Definition
8.1, where T (M ) denotes the closure of T (M ).

Theorem 8.3. (Schauder’s fixed point theorem) Let X be a normed vector

space and K ⊆ X be a nonempty, convex, bounded, and closed set. Furthermore,
let T : K → K be compact. Then, there exists some x∗ ∈ K satisfying x∗ =
T (x∗ ).

Proof: By Definition 8.2 the set T (K) is a compact set. Furthermore, since K
is closed and since T (K) ⊆ K we have

T (K) ⊆ K.

Now, let ε > 0 be arbitrary but fixed. Then, because of Theorem 8.2 there exist
v (1) , ..., v (p) ∈ T (K) so that for every k ∈ K there is some v (i) ∈ {v (1) , ..., v (p) }
satisfying kv (i) − T (k)k < ε. Therefore, considering the functions
( )
ε − kv (i) − T (k)k if kv (i) − T (k)k < ε,
mi (T (k)) := i = 1, ..., p,
0 otherwise

mi (T (k)) > 0 for some i, 1 ≤ i ≤ p, for each k ∈ K. So, the function Tε : K → X

defined by
p
1 X
Tε (k) := p mi (T (k)) · v (i)
P
mi (T (k)) i=1
i=1

is well defined. Since each v (i) ∈ T (K) ⊆ K and since K is convex, it follows
that Tε (K) ⊆ K. Moreover, for every k ∈ K it holds
p
P
mi (T (k)) · kT (k) − v (i) k
i=1
kT (k) − Tε (k)k ≤ p < ε. (8.13)
P
mi (T (k))
i=1

Now, let {ε(n) }∞

n=1 be a sequence in R satisfying

ε(n) > 0 and lim ε(n) = 0.

n→∞

95
8 Schauder’s Fixed Point Theorem

According to ε(n) we define Xn := span{v (1) , ..., v (pn ) } and Kn := K ∩ Xn .

Then, by Theorem 8.1 it follows that Xn is isomorphic to Rn and obviously Kn
is nonempty, convex, bounded, and closed. Furthermore, we have shown above
that
Tεn (Kn ) ⊆ Kn .
So, by Corollary 8.1 there exists x(n) ∈ Kn so that

Tεn (x(n) ) = x(n) .

Since T (x(n) ) ∈ T (Kn ) ⊆ T (K) and since the set T (K) is compact, the sequence
{T (x(n) )}∞
n=1 must have a convergent subsequence, say

lim T (x(nl ) ) = x∗ ∈ T (K) ⊆ K. (8.14)

l→∞

So, for every ε > 0 there is some l0 ∈ N so that

kx∗ − T (x(nl ) )k < ε for all l > l0 .

According to (8.13), it holds

kT (x(nl ) ) − Tεnl (x(nl ) )k < εnl for all l.

This results in
kx∗ − x(nl ) k = kx∗ − Tεnl (x(nl ) )k

≤ kx∗ − T (x(nl ) )k + kT (x(nl ) ) − Tεnl (x(nl ) )k

< ε + εnl

for all l > l0 . In other words, we have

lim x(nl ) = x∗ .
l→∞

Since T is continuous, we have

lim T (x(nl ) ) = T (x∗ ).

l→∞

So, together with (8.14), we get x∗ = T (x∗ ). 

96
 8.2 Peano’s Existence Theorem

8.2 Peano’s Existence Theorem

In 1890, G. Peano published the following result.

Theorem 8.4. Let f : [ξ, ξ + a] × [η − b, η + b] → R be continuous. Then, by

setting

b
M := max{|f (x, y)| : (x, y) ∈ [ξ, ξ + a] × [η − b, η + b]} and α := min{a, }
M
there exists a continuously differentiable function y(x) satisfying

y 0 (x) = f (x, y(x)), x ∈ [ξ, ξ + α]

y(ξ) = η.

The proof will be done via Schauder’s fixed point theorem. Another auxiliary
result will be the theorem of Ascoli-Arzela which is based on so-called equicon-
tinuous sets of real-valued functions. So, we start with the definition of equicon-
tinuity.

Definition 8.3. Let F = {f (1) , f (2) , ...} be a set of functions where f (i) : [c, d] →
R are continuous for all f (i) ∈ F . The set F is said to be equicontinuous, if for
ε > 0 there exists some δ = δ(ε), so that for all f (i) ∈ F it holds that

|f (i) (x) − f (i) (x)| < ε if |x − x| < δ.

Note, that for all f (i) ∈ F the same δ can be chosen. Obviously, if F = {f };
i.e., if F is a singleton, then F is equicontinuous if f is continuous. Consider the
case that F = {f (1) , f (2) } where both functions are continuous. So, for ε > 0
there exists some δ1 so that

|f (1) (x) − f (1) (x)| < ε if |x − x| < δ1 .

and there exists some δ2 so that

|f (2) (x) − f (2) (x)| < ε if |x − x| < δ2 .

Setting δ := min{δ1 , δ2 } we have shown that F is equicontinuous. In the same

way one can show that any finite set of continuous functions with the same
domain is equicontinuous. Therefore, the question arises, if there are sets of
continuous functions with the same domain that are not equicontinuous. This
question will be answered in the following example.

97
8 Schauder’s Fixed Point Theorem

Example 8.6. Let F be the set of all continuous functions with domain D =
[0, 1] and let ε = 21 . Then, for every δ > 0 we can define two different values
x1 ∈ [0, 1] and x2 ∈ [0, 1] satisfying

0 < x2 − x1 < δ

and based on these two values we can define a continuous function f satisfying
f (x1 ) = 0 and f (x2 ) = 1. For example, f could be chosen as a piecewise linear
function; i.e., 

 0 if x ∈ [0, x1 ],

 x−x
1
f (x) = if x ∈ (x1 , x2 ],
 x2 − x1


 1 if x ∈ (x2 , 1].
Then, we have

1
|x2 − x1 | < δ, but |f (x2 ) − f (x1 )| = |1 − 0| = 1 6< = ε.
2
Hence, the set F is not equicontinuous. 

Equicontinuity plays an important role in the following theorem.

Theorem 8.5. (Theorem of Ascoli-Arzela) Let F = {f (1) , f (2) , ...} be equicon-

tinuous, where f (i) : [c, d] → R for all f (i) ∈ F . If there exists some P ≥ 0 so
that
|f (i) (x)| ≤ P for all x ∈ [c, d], and for all f (i) ∈ F,
then F has a uniformly convergent subsequence and its limit function is contin-
uous.

Proof: Let A = {x1 , x2 , ...} be the set of rational numbers in [c, d]. The sequence
of real numbers {an }∞n=1 with

an := f (n) (x1 )

is bounded by P . The Bolzano-Weierstraß theorem tells us this sequence has a

convergent subsequence, say

f (p1 ) (x1 ), f (p2 ) (x1 ), f (p3 ) (x1 ) ... .

The sequence of real numbers {bn }∞

n=1 with

bn := f (pn ) (x2 )

98
 8.2 Peano’s Existence Theorem

is bounded by P , too. Again, by the theorem of Bolzano-Weierstraß, this se-

quence has a convergent subsequence, say

f (q1 ) (x2 ), f (q2 ) (x2 ), f (q3 ) (x2 ) ... .

Here, {qn }∞ ∞ ∞
n=1 is a subsequence of {pn }n=1 . The third sequence {cn }n=1 with

cn := f (qn ) (x3 )

has a convergent subsequence, too, say

f (r1 ) (x3 ), f (r2 ) (x3 ), f (r3 ) (x3 ) ... .

Continuing this process we get several sequences

f (p1 ) , f (p2 ) , f (p3 ) , f (p4 ) , ... convergent for x = x1 ,

(q1 ) (q2 ) (q3 ) (q4 )
f ,f ,f ,f , ... convergent for x = x1 , x2
(r1 ) (r2 ) (r3 ) (r4 )
f ,f ,f ,f , ... convergent for x = x1 , x2 , x3 ,
...................................

The kth row is a subsequence of the (k − 1)th row being convergent for x =
x1 , x2 , ..., xk . As a result, the diagonal sequence
)
{f (dn) (x)}∞
n=1 := f
(p1 )
(x), f (q2 ) (x), f (r3 ) (x), ...
(8.15)
is convergent for all x ∈ A.

Now, let ε > 0. Since {f (dn) (x)}∞

n=1 is equicontinuous, there exists δ(ε) so that
for all dn ≥ 1 it holds that
ε
|f (dn ) (x) − f (dn) (x)| < if |x − x| < δ(ε). (8.16)
3
Then, we subdivide the interval [c, d] in subintervals J1 , ...,Jp so that

[c, d] = J1 ∪ ... ∪ Jp with diameter of Ji ≤ δ(ε), i = 1, ..., p.

For every Ji there exists xi with

xi ∈ Ji ∩ A.

Furthermore, due to (8.15), there exists n0 (ε) so that

ε
|f (dm ) (xi ) − f (dn ) (xi )| < for all i = 1, ..., p if m, n ≥ no (ε). (8.17)
3

99
8 Schauder’s Fixed Point Theorem

As a result, if x ∈ [c, d], there is some Jk with x ∈ Jk and we get |x − xk | < δ(ε),
whence via (8.16) and (8.17) it follows

|f (dm ) (x) − f (dn) (x)| ≤

|f (dm ) (x) − f (dm ) (xk )| + |f (dm ) (xk ) − f (dn ) (xk )| + |f (dn ) (xk ) − f (dn ) (x)| < ε
if m, n ≥ n0 (ε). Since x ∈ [c, d] was arbitrary, then for m, n ≥ n0 (ε) we can
conclude that

ε > max{|f (dm ) (x) − f (dn ) (x)| : x ∈ [c, d]} = kf (dm ) − f (dn) k∞ .

This means, that {f (dn) }∞n=1 is a Cauchy sequence subject to the maximum
norm. But (C([c, d]), k · k∞ ) is a Banach space, so {f (dn) }∞
n=1 is convergent in
this norm and the limit function is continuous. 

Now, we are able to prove Peano’s existence theorem. Note that ξ, η, b, a, α,

M , and the function f are given, there.
Proof of Theorem 8.4: We define the set of functions

K := {w ∈ C([ξ, ξ + α]) : w(x) = η + v(x), kvk∞ ≤ b}.

Then, we define an operator T on K: for w ∈ K we set

Zx
T (w)(x) := η + f (t, w(t))dt.
ξ

Since f : [ξ, ξ + a] × [η − b, η + b] → R is continuous, T (w)(x) is continuous, too,

and  
Zx 
  b
 
 f (t, w(t))dt ≤ (x − ξ) · M ≤ α · M ≤ ·M =b
  M
ξ 

for all x ∈ [ξ, ξ + α] ⊆ [ξ, ξ + a]. So, T is a self-mapping; i.e.,

T : K → K.

Obviously, K is nonempty, convex, bounded, and closed. In order to apply

Schauder’s fixed point theorem, it remains to be shown that T is compact.
Firstly, we show, that T is continuous. Let

lim kw(n) − w∗ k∞ = 0.
n→∞

100
 8.2 Peano’s Existence Theorem

Then, since f is continuous,

ξ+α
Z
(n) ∗
kT (w ) − T (w )k∞ ≤ kf (·, w(n) (·)) − f (·, w∗ (·))k∞ dt < ε
ξ

if n is sufficiently large. Secondly, let {w(n) }∞

n=1 be a sequence in K. Since T is
a self-mapping and since K is a bounded set, the set {T (w(n) )}∞ n=1 is bounded,
and it only remains to show that {T (w(n) )}∞ n=1 is equicontinuous in order to
apply the theorem of Ascoli-Arzela.
Let ε > 0 be given. Then, we choose
ε
δ := .
M
Hence, for every n ≥ 1 and for every x, x ∈ [ξ, ξ + α] with |x − x| < δ we get
 x 
Z 
(n) (n)
 (n)

|T (w )(x) − T (w )(x)| ≤  f (t, w (t))dt ≤ |x − x| · M < ε.

 
x

Now, the theorem of Ascoli-Arzela can be applied, which means that T is com-
pact. Then Schauder’s fixed point theorem can be applied, which means that
there exists some w̃ ∈ K satisfying w̃ = T (w̃). It follows
Zx
w̃(x) = η + f (t, w̃(t))dt
ξ

for x ∈ [ξ, ξ + α]. Since this is equivalent to

w̃0 (x) = f (x, w̃(x)), x ∈ [ξ, ξ + α]

w̃(ξ) = η

Peano’s existence theorem is proved. 

101
8 Schauder’s Fixed Point Theorem

8.3 The Poincaré-Miranda Theorem in l2

The following result is presented in this chapter, because its proof is similar to
the proof of Schauder’s fixed point theorem.
Theorem 8.6. Let x̂ = {x̂i }∞ 2 ∞ 2
i=1 ∈ l , L = {li }i=1 ∈ l , li ≥ 0 for all i ∈ N,
Ω := {x ∈ l : |xi − x̂i | ≤ li for all i ∈ N} and f : Ω → l2 be a continuous
2

function on Ω. Also let

Fi+ := {x ∈ Ω : xi = x̂i + li }, Fi− := {x ∈ Ω : xi = x̂i − li } for all i ∈ N.
If for all i ∈ N it holds that
fi (x) · fi (y) ≤ 0 for all x ∈ Fi+ and for all y ∈ Fi− , (8.18)
∗ ∗
then there exists some x ∈ Ω satisfying f (x ) = 0, where 0 denotes the number
zero as well as the zero sequence.

Proof: For fixed n ∈ N we consider the function h̃(n) : Ω → l2 defined by

 
f1 (x1 , x2 , ..., xn−1 , xn , xn+1 , ...)
 .. 

 . 

(n)
f (x
h̃ (x) :=  n 1 2
 , x , ..., xn−1 , xn , xn+1 , ...) .


 0 

..
.
As in Example 8.5 we can show that Ω is compact. Since f is continuous, the set
f (Ω) is compact, too. Therefore, for given ε > 0 there is a finite set of elements
v (1) , ..., v (p) ∈ f (Ω) so that if f (x) ∈ f (Ω), then there is a v ∈ {v (1) , ..., v (p) } so
that
kf (x) − vk ≤ ε
and there exists n1 = n1 (ε) ∈ N so that for all n > n1 it holds that
v
u ∞
u X
t (vj )2 ≤ ε for all v ∈ {v (1) , ..., v (p) }.
j=n+1

So, if n > n1 is valid, then for all f (x) ∈ f (Ω) we have some v ∈ {v (1) , ..., v (p) }
so that for all x ∈ Ω
   
0 0
 ..   .. 

 . 

 . 
 
 0  k ≤ kf (x) − vk + k  0  k ≤ 2ε.
  
kf (x) − h̃(n) (x)k = k 
 fn+1 (x)   vn+1 
   
 fn+2 (x)   vn+2 
   
.. ..
. .

102
 8.3 The Poincaré-Miranda Theorem in l2

Now, for fixed n ∈ N we define

 
[x̂1 − l1 , x̂1 + l1 ]

Ωn :=  .. 
. 
[x̂n − ln , x̂n + ln ]

and h(n) : Ωn → Rn by
 
f1 (x1 , x2 , ..., xn−1 , xn , x̂n+1 , x̂n+2 , ...)
h(n) (x) := 
 .. 
. .
fn (x1 , x2 , ..., xn−1 , xn , x̂n+1 , x̂n+2 , ...)

By (8.18) and Corollary 4.2 there exists x(n) ∈ Ωn with

h(n) (x(n) ) = 0,

where 0 denotes the zero vector in Rn . Setting

 
x(n)
 x̂n+1 
x̃(n) := 
 
 x̂n+2 

..
.

it holds that
x̃(n) ∈ Ω and h̃(n) (x̃(n) ) = 0,

where 0 denotes the zero sequence. Now, let n > n1 . Then,

kf (x̃(n) )k = kf (x̃(n) ) − h̃(n) (x̃(n) )k ≤ 2ε.

Hence, lim f (x̃(n) ) = 0, where 0 denotes the zero sequence. Since Ω is compact,
n→∞
the sequence x̃(n) has a convergent subsequence with limit x∗ ∈ Ω. W.l.o.g. we
assume that lim x̃(n) = x∗ holds. On the one hand, it follows that lim f (x̃(n) ) =
n→∞ n→∞
f (x∗ ), since f is continuous. On the other hand, it follows that f (x∗ ) = 0, since
the limit is unique. 
The infinite-dimensional Poincaré-Miranda theorem can be applied to difference
equations that, in fact, are equivalent to some special infinite (denumerable)
systems of equations, and to the controllability of nonlinear repetitive processes.
See (47).

103
8 Schauder’s Fixed Point Theorem

8.4 Verification Methods

In Chapter 6 verification methods were presented based on Brouwer’s fixed point

theorem. The problems, whose solutions have been verified there, were finite-
dimensional. So, the idea is to apply Schauder’s fixed point theorem to problems
that are infinite-dimensional.

But how can a computer, that can only represent finite many num-
bers, be used to verify solutions solving an infinite-dimensional prob-
lem?

We present the main ideas due to M. Plum by an illustrative example.

Example 8.7. Many boundary value problems for semilinear elliptic partial
differential equations allow very stable numerical computations of approximate
solutions, but are still lacking analytical existence proofs. Plum proposes a
method which exploits the knowledge of a good numerical approximate solution,
in order to provide a rigorous proof of existence of an exact solution close to
the approximate one. This goal is achieved by Schauder’s fixed point theorem
as follows.

Consider the problem

)
−(4u)(x) + f (x, u(x)) = 0, x ∈ Ω,
(8.19)
u(x) = 0, x ∈ ∂Ω,

where Ω ⊂ Rn , with n ≤ 3, is a bounded domain with boundary ∂Ω, f : Ω×R →

R is a given nonlinearity with f and ∂f
∂y being continuous, and 4 is the usual
n
P ∂2
Laplacian; i.e., 4 = ∂x2
.
i
i=1

Let ũ(x) be an approximate solution of (8.19) calculated by some classical

method (such as the finite element method). In order to verify, that at least
in a neighbourhood of ũ(x) there really exists a solution of (8.19), we consider
the function set )
U = ũ + V, where
(8.20)
V = {v ∈ C(Ω) : kvk∞ ≤ α}

with α > 0 to be determined. Here, C(Ω) denotes the set of all real-valued
continuous functions with domain Ω, and kvk∞ = max{|v(x)| : x ∈ Ω}.

But how can a solution exist in U , although the Laplacian occurs in

(8.19)?

104
 8.4 Verification Methods

The idea is to find a so-called ’weak solution’. To explain the concept of a weak
solution, we need some definitions. We define
Z
L (Ω) = {u : Ω → R : u measurable , u2 (x)dx < ∞}
2

and H := H 2 (Ω)∩ H̊ 1 (Ω), consisting of all functions u ∈ L2 (Ω) which have weak
2
∂u
derivatives ∂x i
, ∂x∂i ∂x
u
j
in L2 (Ω) (i.e., functions u for which functions

∂u ∂2u
vi =: ∈ L2 (Ω), wij =: ∈ L2 (Ω) i, j = 1, ..., n
∂xi ∂xi ∂xj
exist so that Z Z
∂ϕ
u(x) · (x)dx = − vi (x) · ϕ(x)dx (8.21)
∂xi
Ω Ω

and Z Z
∂2ϕ
u(x) · (x)dx = wij (x) · ϕ(x)dx (8.22)
∂xi ∂xj
Ω Ω

for all ϕ ∈ C0∞ (Ω))1 , and which vanish on ∂Ω in the sense that

∂u ∂ϕ(n)
ku − ϕ(n )kL2 (Ω) → 0, k − kL2 (Ω) → 0 as n → ∞ (8.23)
∂xi ∂xi

for some sequence {ϕ(n) }∞ ∞

n=1 in C0 (Ω). Due to the embedding theorem of
Sobolev-Kondrachev-Rellich (see (1)), we have indeed H ⊂ C(Ω)) and the em-
bedding H ,→ C(Ω)) is compact (i.e., each k · kH -bounded sequence has a k · k∞ -
convergent subsequence).
Now, we go back to our problem (8.19). Assume, the operator

L : H → L2 (Ω)

defined by
∂f
L(u)(x) := −(4u)(x) + (x, ũ(x)) · u(x) (8.24)
∂y
has an inverse
L−1 : L2 (Ω) → H
that is bounded. Then, we define the operator

T : C(Ω) → C(Ω)
1 C ∞ (Ω)
denotes the vector space of all infinitely differentiable functions on Ω with compact
0
support in Ω.

105
8 Schauder’s Fixed Point Theorem

by
 ∂f 
T (u)(x) := L−1 (x, ũ(x)) · u(x) − f (x, u(x)) , x ∈ Ω.
∂y
Note, that H ⊂ C(Ω). If T had a fixed point, say u∗ , then

u∗ (x) = T (u∗ )(x), x∈Ω

would lead to
∂f
L(u∗ )(x) = (x, ũ(x)) · u∗ (x) − f (x, u∗ (x)), x ∈ Ω, (8.25)
∂y
which, due to (8.24), is equivalent to
∂f ∂f
− (4u∗ )(x) + (x, ũ(x)) · u∗ (x) = (x, ũ(x)) · u∗ (x) − f (x, u∗ (x)), (8.26)
∂y ∂y
x ∈ Ω; i.e.,
− (4u∗ )(x) + f (x, u∗ (x)) = 0, x ∈ Ω. (8.27)
Since T (u∗ ) ∈ H, we have that u∗ ∈ H; i.e., u∗ is a weak solution of (8.19).
(The equalities of (8.25)-(8.27), and of u∗ (x) = 0 on ∂Ω are only valid in the
sense of (8.21), (8.22), and (8.23).)
To verify a fixed point of T , Schauder’s fixed point theorem is applied based on
the observation that T is compact, since the embedding H ,→ C(Ω) is compact.
We have
 
T (u)(x) = L−1 ∂f ∂y (x, ũ(x)) · u(x) − f (x, u(x))
 
= ũ(x) − L−1 (L(ũ))(x) + L−1 ∂f ∂y (x, ũ(x)) · u(x) − f (x, u(x))

(8.24)
 ∂f 
= ũ(x) − L−1 − (4ũ)(x) + (x, ũ(x)) · (ũ(x) − u(x)) + f (x, u(x)) .
∂y
Setting
∂f
g(x, y) := f (x, y) − f (x, ũ(x)) − (x, ũ(x)) · (y − ũ(x))
∂y
we get by Taylor’s formula
g(x, y)
→0 as y → ũ(x) (8.28)
|y − ũ(x)|
and
∂f
f (x, ũ(x)) + g(x, u(x)) = f (x, u(x)) + (x, ũ(x)) · (ũ(x) − u(x)).
∂y

106
 8.4 Verification Methods

So, in order to apply Schauder’s fixed point theorem we have to find α that
satisfies  
k − L−1 − 4ũ + f (·, ũ) + g(·, ũ + v) k∞ ≤ α (8.29)

for all v ∈ C(Ω) with kvk∞ ≤ α. Then, we have (see (8.20))

T (U ) ⊆ U, (8.30)

and we can conclude that in U there is a weak solution of (8.19).

With (8.28) in hand it is not unlikely that (8.29) can be verified, if ũ is a good
approximation (and if there really exists a solution).
In order to verify (8.29), operator norm bounds for L−1 , −4ũ + f (·, ũ), and g
are needed. Plum calculated eigenvalue bounds for L of (8.24) and bounds for
an integral with known integrand. Interval arithmetic was used in all numerical
evaluations (but not during the computation of the approximate solution ũ!), in
order to take rounding errors into account on a computer.
We do not go into further details, but we present a numerical example published


by Breuer, McKenna, and Plum in 2003. Let Ω = (0, 1) × (0, 1) and x = xx12 .
They considered the problem
)
−(4u)(x) − (u(x))2 = −s · sin(π · x1 ) · sin(π · x2 ), x ∈ Ω,
(8.31)
u(x) = 0, x ∈ ∂Ω.

It had been conjectured in the PDE community since the 1980’s that problem
(8.31) has at least 4 solutions for s > 0 sufficiently large. For s = 800, Breuer,
McKenna, and Plum were able to compute 4 essentially different approximate
solutions by the numerical mountain pass algorithm developed in (20), where
”essentially different” means that none of them is an elementary symmetry trans-
form of another one. With Plum’s method they could verify, by use of interval
arithmetic exploited on a computer, that in a neighborhood of each of these 4
approximate solutions, indeed there really exists a solution.
Two years after publication, Dancer and Yan gave a more general analytical
proof. They even proved that the number of solutions of problem (8.31) becomes
unbounded as s → ∞. Nevertheless, Breuer, McKenna, and Plum were first. 

Other people have also considered so-called enclosure methods for the problem
(8.19) using Schauder’s fixed point theorem. We want to mention M. T. Nakao
and his co-workers. Their method avoids the computation of a bound for L−1 .
On the other hand, it requires the verified solution of large nonlinear and linear
systems, where the latter moreover have an interval right-hand side, and it
requires explicit knowledge of projection error bounds. We refer to (65).

107
9 Tarski’s Fixed Point Theorem

Continuity was the starting point to guarantee the existence of a fixed point:
) (
f : [a, b] → [a, b], there exists x∗ ∈ [a, b]
⇒
f is continuous with x∗ = f (x∗ ).

We have presented several extensions. For example, f : K → K, where K ⊆ Rn

is convex, bounded, and closed. Then, we have weakened the assumption that f
has to be a self-mapping, and Schauder’s fixed point theorem considers compact
mappings in infinite-dimensional vector spaces.

All these results exploit the fact that f is continuous. So, on first sight, it
seems implausible to get a fixed point theorem that does not assume continuity.
However, let us consider the following situation:

f : [a, b] → [a, b], f is increasing;

i.e., if x ≤ y, then f (x) ≤ f (y). Figure 9.1 shows that such a function must
have a fixed point, although the function is not continuous. It was A. Tarski,
who followed this idea. The starting point is no longer a vector space, but it is
an ordered set (in order to generalize what is meant by increasing).

Definition 9.1. Let Ω be a nonempty set. Then, (Ω, ≤) is called an ordered

set, if for all x, y, z ∈ Ω

x ≤ x,
x ≤ y, y ≤ x ⇒ x = y
x ≤ y, y ≤ z ⇒ x ≤ z.

Example 9.1. We give some examples of ordered sets.

a) (Rn , ≤) with
   
x1 y1
 ..   .. 
 . ≤ .  :⇔ xi ≤ yi for i = 1, ..., n.
xn yn

109
9 Tarski’s Fixed Point Theorem

b) (l2 , ≤), (c0 , ≤), and (l∞ , ≤) with

{xn }∞ ∞
n=1 ≤ {yn }n=1 :⇔ xn ≤ yn for n = 1, 2, 3, ... .

c) (C([a, b]), ≤) with

f ≤g :⇔ f (x) ≤ g(x) for all x ∈ [a, b].

To formulate a fixed point theorem, we need only one more definition.

Definition 9.2. An ordered set (Ω, ≤) is said to be a complete lattice, if for any
nonempty Λ ⊆ Ω
inf Λ as well as sup Λ
exist and
inf Λ ∈ Ω as well as sup Λ ∈ Ω.

Now, we can present Tarski’s fixed point theorem.

Theorem 9.1. (Tarski’s fixed point theorem, 1955) Let (Ω, ≤) be a com-
plete lattice and let Φ : Ω → Ω be a function satisfying

ω≤v ⇒ Φ(ω) ≤ Φ(v), for all ω, v ∈ Ω.

Then, there exists some ω ∗ ∈ Ω satisfying ω ∗ = Φ(ω ∗ ).

Proof: Since (Ω, ≤) is a complete lattice, we have that inf Ω ∈ Ω as well as

sup Ω ∈ Ω exist. So, sup Ω = max Ω and inf Ω = min Ω.
Moreover, since Φ is an increasing self-mapping, we have

min Ω ≤ Φ(min Ω) ≤ Φ(max Ω) ≤ max Ω.

Motivated by Figure 9.1 we define

Λ := {ω ∈ Ω : ω ≤ Φ(ω)}.

We know that Λ 6= ∅, because min Ω ∈ Λ. Again, since (Ω, ≤) is a complete

lattice, the existence of ω0 := sup Λ is guaranteed. This means, that

ω ≤ ω0 for all ω ∈ Λ.

Since Φ is increasing by the definition of Λ we get

ω ≤ Φ(ω) ≤ Φ(ω0 ) for all ω ∈ Λ.

110
 9 Tarski’s Fixed Point Theorem

a b x
Figure 9.1: There must be a fixed point, only due to the fact, that f is
an increasing self-mapping

So, Φ(ω0 ) is an upper bound of Λ. According to the definition of the supremum

of a set we get
ω0 ≤ Φ(ω0 ), (9.1)

whence it follows that Φ(ω0 ) ≤ Φ(Φ(ω0 )) by the monotonicity of Φ. One can

conclude that Φ(ω0 ) ∈ Λ and by the very definition of the supremum of a set we
get
Φ(ω0 ) ≤ ω0 .

Together with (9.1) we finally get Φ(ω0 ) = ω0 . 

As the reader might have noticed, the proof is very short and elementary. How
can it be applied? First, we need to get a hint at what we’re dealing with - what
are some pertinent examples of complete lattices?

111
9 Tarski’s Fixed Point Theorem

Remembering Figure 5.1 at page 48, we know that Rn with n ≥ 2 is not a

complete lattice. However, the next example will present a complete lattice.
On first sight, this example looks strange, but it will be used in the following
sections.
Example 9.2. Let T > 0 be given. Furthermore, let Γ ≥ 0, L̃ ≥ 0, and L∗ ≥ 0
be given. Then, we consider Ω = Ω(Γ, L̃, L∗ ) with
n
Ω := ω = {ωn }∞ ∞
n=1 : R × [0, T ] → l ,

for all n ∈ N it holds |ωn (x, t)| ≤ Γ for all (x, t) ∈ R × [0, T ],

and for all (x, t), (y, s) ∈ R × [0, T ] it holds

o
kω(x, t) − ω(y, s)k ≤ L̃ · |x − y| + L∗ · |t − s| .

I.e., Ω is a set of l∞ -valued, bounded, Lipschitz continuous functions, and we

want to show that (Ω, ≤) is a complete lattice, where for ω, v ∈ Ω it is
(
for all n = 1, 2, 3, ... it holds that ωn (x, t) ≤ vn (x, t)
ω ≤ v :⇔
for all (x, t) ∈ R × [0, T ].

Of course, Ω is nonempty, since the zero element of l∞ , interpreted as the con-

stant sequence of the zero function, belongs to Ω. Now, let Λ ⊆ Ω and let Λ be
nonempty. Then, we have to show, that inf Λ as well as sup Λ exist and that
inf Λ ∈ Ω as well as sup Λ ∈ Ω is fulfilled. By the definition of Ω, we have for all
ω ∈ Ω (and therefore, for all ω ∈ Λ)

|ωn (x, t)| ≤ Γ for all n = 1, 2, 3, ... and for all (x, t) ∈ R × [0, T ]. (9.2)

So, for any n ∈ N and for any (x, t) ∈ R × [0, T ] we define

ω̃n (x, t) := sup{ωn (x, t) : ω ∈ Λ}. (9.3)

By definition (see (9.2)) |ω̃n (x, t)| ≤ Γ for all n ∈ N and for all (x, t) ∈ R × [0, T ].
It remains to show that

kω̃(x, t) − ω̃(y, s)k ≤ L̃ · |x − y| + L∗ · |t − s| for all (x, t), (y, s) ∈ R × [0, T ].

So, let (x, t), (y, s) ∈ R × [0, T ], and let n ∈ N be arbitrary but fixed. Then for
all ω ∈ Λ

ωn (x, t) − ωn (y, s) ≤ |ωn (x, t) − ωn (y, s)| ≤ L̃ · |x − y| + L∗ · |t − s|

and it follows that

112
 9 Tarski’s Fixed Point Theorem

a b x
Figure 9.2: No fixed point

ωn (x, t) ≤ L̃ · |x − y| + L∗ · |t − s| + ωn (y, s) ≤ L̃ · |x − y| + L∗ · |t − s| + ω̃n (y, s).

In other words,
L̃ · |x − y| + L∗ · |t − s| + ω̃n (y, s)
is an upper bound for ωn (x, t), ω ∈ Λ. By definition of the supremum we get
ω̃n (x, t) ≤ L̃ · |x − y| + L∗ · |t − s| + ω̃n (y, s).
As a result
ω̃n (x, t) − ω̃n (y, s) ≤ L̃ · |x − y| + L∗ · |t − s|.
Exchanging (x, t) with (y, s) one can see that
ω̃n (y, s) − ω̃n (x, t) ≤ L̃ · |y − x| + L∗ · |s − t|.
That is
|ω̃n (x, t) − ω̃n (y, s)| ≤ L̃ · |x − y| + L∗ · |t − s|.

113
9 Tarski’s Fixed Point Theorem

Since the right-hand side does not depend on n, we have

kω̃(x, t) − ω̃(y, s)k ≤ L̃ · |x − y| + L∗ · |t − s|.
It follows that ω̃ ∈ Ω, and it is clear from (9.3) that ω̃ = sup Λ. In the same
way, one can show that inf Λ exists and that inf Λ ∈ Ω. 

Finally, we want to note that Tarski’s fixed point theorem is no longer true, if
one substitutes
ω ≤ v ⇒ Φ(ω) ≤ Φ(v)
by
ω≤v ⇒ Φ(ω) ≥ Φ(v).
See Figure 9.2.

9.1 Differential Equations in Banach Spaces

In 1950, Dieudonné published an example that shows that Peano’s existence
theorem is no longer true, if the considered function f is not real-valued, but
has its values in a real Banach space. More precisely, let E be a real Banach
space, let a ∈ E, and let T > 0. Then, we consider the so-called Cauchy
problem )
u0 (t) = f (t, u(t)), t ∈ [0, T ],
(9.4)
u(0) = a,
where
f : [0, T ] × E → E
is continuous and bounded. The following example shows, that, in general, this
problem has no solution.
Example 9.3. (Dieudonné, 1950) Let E = c0 , let 0 < T ≤ 2, and let
 
1 1 1
a = 1, , , , ... ∈ c0 .
2 3 4
Furthermore, let f : [0, T ] × c0 → c0 be defined by
  n  o∞
f t, {xn }∞ n=1 = f m t, {xn } ∞
n=1
m=1

with 
 √0 if xm ≤ 0,
  
∞
fm t, {xn }n=1 = xm if 0 ≤ xm ≤ 4,


2 if 4 ≤ xm .

114
 9.1 Differential Equations in Banach Spaces

I.e., the functions fm are independent of t and we have

 
lim fm t, {xn }∞
n=1 = 0 if lim xm = 0.
m→∞ m→∞

Therefore, we have f : [0, T ] × c0 → c0 . If (9.4) had a solution u : [0, T ] → c0 ;

i.e.,
u(t) = {u1 (t), u2 (t), u3 (t), ...},
it would follow
 
 0 if um (t) ≤ 0, 
 p  1
0
um (t) = um (t) if 0 ≤ um (t) ≤ 4, t ∈ [0, T ], um (0) =

 
 m
2 if 4 ≤ um (t),

for all m ∈ N. For fixed m ∈ N we get the (unique) solution

 2
t 1
um (t) = +√ , t ∈ [0, T ].
2 m

(Note, that T ≤ 2.) However, due to

t2
lim um (t) = >0 for all t ∈ (0, T ],
m→∞ 4
we would have
u(t) = {u1 (t), u2 (t), u3 (t), ...} 6∈ c0 .
Therefore, the problem (9.4) has no solution in this situation. 

Godunov (1974) showed that whenever E is an infinite-dimensional Banach space

there exist an a in E, a T > 0, and a continuous f so that the problem (9.4) has
no solution. So, a new field in mathematics was born: Differential Equations in
Banach Spaces, dealing with the question what conditions have to be fulfilled so
that a solution of (9.4) can be guaranteed. We refer to the book of K. Deimling
and the paper of P. Volkmann for a detailed introduction.
In the following section, we will present an existence theorem for a parabolic
differential equation in l∞ based on Tarski’s fixed point theorem. Its advantage is
that there is no need to consider compact functions as it is the case in Schauder’s
fixed point theorem, so we are able to get an existence theorem in l∞ .

115
9 Tarski’s Fixed Point Theorem

9.2 Applying Tarski’s Fixed Point Theorem

We consider the parabolic differential equation

ut (x, t) − uxx (x, t) = f (x, t, u(x, t)), (x, t) ∈ R × (0, T ),
(9.5)
u(x, 0) = 0, x ∈ R,

with
f : R × (0, T ) × l∞ → l∞ .
I.e., the values of a solution u(x, t) have to be in l∞ . Before we attack this
problem, we need to explain what is meant by ut (x, t) and uxx (x, t).
Definition 9.3. Let E be a Banach space. Then, a function u : R × (0, T ) → E
is partially differentiable with respect to t in (x0 , t0 ) ∈ R × (0, T ), if there exists

ut (x0 , t0 ) ∈ E

so that  
1
lim · u(x0 , t) − u(x ,
0 0t ) − u t 0 0 =0
(x , t )
t→t0 t − t0

holds.

Analogously, one can define ux (x, t) and uxx (x, t).

If the function f (x, t, z) in (9.5) does not depend on z, the problem (9.5) can be
solved by the following theorem.
Theorem 9.2. Let E be any real Banach space and let g(x, t) : R×[0, T ] → E be
bounded, continuous, and Lipschitz continuous in x ∈ R, uniformly with respect
to t; i.e., there exists a constant L so that

kg(x, t) − g(y, t)k ≤ L · |x − y|

for all (x, t), (y, t) ∈ R × [0, T ]. Then, the function

Zt Z∞ (x − ξ)2
1 −
u(x, t) := p · e 4(t − τ ) · g(ξ, τ ) dξ dτ
4π(t − τ )
0 −∞

is a solution of the Cauchy problem

ut (x, t) − uxx (x, t) = g(x, t), (x, t) ∈ R × (0, T )

u(x, 0) = 0, x ∈ R.

In addition, u(x, t) is continuous on R × [0, T ].

116
 9.2 Applying Tarski’s Fixed Point Theorem

For a proof we refer to (33), pp. 1-25. There, the theorem is only given for
real-valued functions, but the theorem remains true if g(x, t) has got its values
in any real Banach space E, because one has just to substitute the norm for the
absolute value. Also, if h : [a, b] → E is continuous, then

Zb Z∞
h(x)dx as well as h(x)dx
a −∞
P
have to be understood as limits of Riemann sums h(ξk ) · (xk − xk−1 ). Now,
we can attack problem (9.5).
Theorem 9.3. Let f : R × [0, T ] × l∞ → l∞ be a function with the following
properties:
i) f is continuous.
ii) There exists a constant L1 so that for all (x, t, z), (y, t, z) ∈ R × [0, T ] × l∞
it holds
kf (x, t, z) − f (y, t, z)k ≤ L1 · |x − y|.

iii) There exists a constant L2 so that for all (x, t, z), (x, t, z̃) ∈ R × [0, T ] × l∞
it holds
kf (x, t, z) − f (x, t, z̃)k ≤ L2 · kz − z̃k.

iv) There exists a constant M so that for all (x, t, z) ∈ R × [0, T ] × l∞ it holds

kf (x, t, z)k ≤ M.

v) For all (x, t, z), (x, t, z̃) ∈ R × [0, T ] × l∞ it holds

z ≤ z̃ ⇒ f (x, t, z) ≤ f (x, t, z̃).

Then, there exists a continuous function u : R × [0, T ] → l∞ satisfying (9.5).

Proof: We consider the following set of functions

n
Ω := ω : R × [0, T ] → l∞ , −Γ ≤ ω ≤ Γ

kω(x, t) − ω(y, s)k ≤ L3 · |x − y| + L4 · |t − s|,

o
for all (x, t), (y, s) ∈ R × [0, T ] ,

with
Γ = {Γn }∞
n=1 and Γn = T · M,

117
9 Tarski’s Fixed Point Theorem

√ q
T
L3 := 2 · M · T and L4 := M + 2(L1 + L2 · L3 ) · π. Moreover, for ω ∈ Ω we
define
Zt Z∞ (x − ξ)2
1 −
Φ(ω)(x, t) := p · e 4(t − τ ) · f (ξ, τ, ω(ξ, τ )) dξ dτ.
4π(t − τ )
0 −∞

Then, for g(x, t) := f (x, t, ω(x, t)) we have

kg(x, t) − g(y, t)k = kf (x, t, ω(x, t)) − f (y, t, ω(y, t))k
≤ kf (x, t, ω(x, t)) − f (y, t, ω(x, t))k+
kf (y, t, ω(x, t)) − f (y, t, ω(y, t))k
≤ L1 · |x − y| + L2 · kω(x, t) − ω(y, t)k
≤ L1 · |x − y| + L2 · L3 · |x − y| = (L1 + L2 · L3 ) · |x − y|
and therefore, by Theorem 9.2,
Φ(ω)t (x, t) − Φ(ω)xx (x, t) = f (x, t, ω(x, t)), (x, t) ∈ R × (0, T ),
Φ(ω)(x, 0) = 0, x ∈ R.
In particular, the function Φ(ω)(x, t) is continuous on R × [0, T ]. This means
that (9.5) has a solution if Φ possesses a fixed point. In order to apply Tarski’s
fixed point theorem, we have to verify three things.
(A1) (Ω, ≤) is a complete lattice.
(A2) Φ : Ω → Ω; i.e., we have to show that Φ is a self-mapping of Ω.
(A3) ω ≤ v ⇒ Φ(ω) ≤ Φ(v) for all ω, v ∈ Ω.
Already in Example 9.2 we have shown the first point. Concerning (A2): We
have to show Φ(ω) ∈ Ω, if ω ∈ Ω; i.e., we have to show three things:
(B1) −T M ≤ Φn (ω)(x, t) ≤ T M for all n ∈ N and for all (x, t) ∈ R × [0, T ].
(B2) kΦ(ω)(x, t) − Φ(ω)(y, t)k ≤ L3 · |x − y| for all x, y ∈ R and for all t ∈ [0, T ].
(B3) kΦ(ω)(x, t) − Φ(ω)(x, s)k ≤ L4 · |t − s| for all x ∈ R and for all s, t ∈ [0, T ].
Concerning
p (B1): Since f (x, t) is bounded by M we get after substituting ξ =
x − 4(t − τ ) α that
(x − ξ)2
Rt R∞ −
kΦ(ω)(x, t)k ≤ M· √ 1
· e 4(t − τ ) dξ dτ
0 −∞ 4π(t−τ )

Rt R∞ 2
= M· √1 · e−α dα dτ.
π
0 −∞

118
 9.2 Applying Tarski’s Fixed Point Theorem

R∞ 2
Using √1 · e−α dα = 1 we get
π
−∞

Zt
kΦ(ω)(x, t)k ≤ M · 1dτ ≤ T · M.
0

Concerning (B2): For fixed n ∈ N we have

Zt Z∞ (x − ξ)2
1 − ξ−x
Φn (ω)x (x, t) = p e 4(t − τ ) fn (ξ, τ, ω(ξ, τ )) dξ dτ
4π(t − τ ) 2(t − τ )
0 −∞

M
and we get |Φn (ω)x (x, t)| ≤ √ (I
π 1
+ I2 ) with

Zt Zx (x − ξ)2
1 − |ξ − x|
I1 := p · e 4(t − τ ) · dξ dτ
4(t − τ ) 2(t − τ )
0 −∞
Zt Zx (x − ξ)2
1 − x−ξ
= p · e 4(t − τ ) · dξ dτ
4(t − τ ) 2(t − τ )
0 −∞

and
Zt Z∞ (x − ξ)2
1 − |ξ − x|
I2 := p · e 4(t − τ ) · dξ dτ
4(t − τ ) 2(t − τ )
0 x
Zt Z∞ (x − ξ)2
1 − ξ−x
= p · e 4(t − τ ) · dξ dτ .
4(t − τ ) 2(t − τ )
0 x
2
Substituting α := − (x−ξ)
4(t−τ ) , −∞ < ξ ≤ x it follows (using dα =
x−ξ
2(t−τ ) dξ) that
 
Zt Z0 Zt Z0
1 1
I1 = p · eα dα dτ = p · eα dα dτ
4(t − τ ) 4(t − τ )
0 −∞ 0 −∞
Zt √
1 √ √
= p dτ = [− t − τ ]|t0 = t ≤ T .
4(t − τ )
0

2
In the same way, substituting α := − (x−ξ)
4(t−τ ) , x ≤ ξ < ∞ it follows using the

119
9 Tarski’s Fixed Point Theorem

same change of variables that

−∞
 0 
Zt Z Zt Z
1 1
I2 = p · eα · (−1) dα dτ = p · eα dα dτ
4(t − τ ) 4(t − τ )
0 0 0 −∞
Zt √
1 √ √
= p dτ = [− t − τ ]|t0 = t ≤ T .
4(t − τ )
0

All together,
M M √ √
|Φn (ω)x (x, t)| ≤ √ (I1 + I2 ) ≤ √ · 2 · T < 2 · M · T = L3 .
π π

Using the mean-value theorem, there exists η between x and y satisfying

|Φn (ω)(x, t) − Φn (ω)(y, t)| = |Φn (ω)x (η, t) · (x − y)| ≤ L3 · |x − y|.

Since the right-hand side does not depend on n, we can conclude

kΦ(ω)(x, t) − Φ(ω)(y, t)k ≤ L3 · |x − y|.

Concerning (B3): Without loss of generality we assume s < t. Then, for D :=

kΦ(ω)(x, t) − Φ(ω)(x, s)k we get



Zs Z∞ (x − ξ)2
1 −
· e 4(t − τ ) · f (ξ, τ, ω(ξ, τ )) dξ dτ

D ≤ p

0 −∞ 4π(t − τ )
| {z }

=:J11

2
s ∞ (x − ξ)
Z Z
1 −
− p ·e 4(s − τ )
· f (ξ, τ, ω(ξ, τ )) dξ dτ
4π(s − τ )

0 −∞
| {z }

=:J12


Z Z (x − ξ)2
t ∞ 1 −
· e 4(t − τ ) · f (ξ, τ, ω(ξ, τ )) dξ dτ .

+ p
s −∞ 4π(t − τ )
| {z }

=:J2

As in Case (B1) one can show kJ2 k ≤ M · (t − s) = M · |t − s|. So, the hard
work is kept in estimating J11 and J12 , respectively. Substituting α = √ x−ξ
4(t−τ )

120
 9.2 Applying Tarski’s Fixed Point Theorem

and α = √ x−ξ , respectively, we get

4(s−τ )

Zs Z∞
1 2
√ √
√ · e−α · f (x − 2α t − τ , τ, ω(x − 2α t − τ , τ ))

kJ11 − J12 k ≤
π
0 −∞

√ √

−f (x − 2α s − τ , τ, ω(x − 2α s − τ , τ )) dα dτ
Zs Z∞
1 2
 √ √ 
√ · e−α · (L1 + L2 · L3 ) · x − 2α t − τ − x + 2α s − τ  dα dτ
 
≤
π
0 −∞
Zs Z∞
1 √ √ 2
= √ · (L1 + L2 · L3 ) · ( t − τ − s − τ ) dτ · e−α 2|α| dα .
π
0 −∞
| {z } | {z }
=:F =2
Concerning F we can conclude
s
2 3 3  2 3 3 3 2 3 3
F = [−(t − τ ) 2 + (s − τ ) 2 ] = (t 2 − s 2 − (t − s) 2 ) < (t 2 − s 2 ) =: P.
3 0 3 3
3 √
Considering the function b(x) := 23 x 2 , 0 ≤ x ≤ T with b0 (x) = x by using the
mean-value theorem we get
2 3 3 √ √
P = (t 2 − s 2 ) = b(t) − b(s) = b0 (η) · (t − s) ≤ T · (t − s) = T · |t − s|
3
for some η ∈ (s, t). All together it holds

D ≤ kJ11 − J12 k + kJ2 k

√
≤ √1 · (L1 + L2 · L3 ) · 2 · T · |t − s| + M · |t − s|
π

= L4 · |t − s|.

Concerning (A3): Let ω ≤ v. Since

(x − ξ)2
1 −
p · e 4(t − τ ) > 0
4π(t − τ )
it follows by the assumption on f for all n ∈ N

Zt Z∞ (x − ξ)2
1 −
Φn (ω)(x, t) = p · e 4(t − τ ) · fn (ξ, τ, ω(ξ, τ )) dξ dτ
4π(t − τ )
0 −∞

121
9 Tarski’s Fixed Point Theorem

Zt Z∞
1 (x−ξ)2
≤ p · e− 4(t−τ ) · fn (ξ, τ, v(ξ, τ )) dξ dτ = Φn (v)(x, t).
4π(t − τ )
0 −∞

So, all assumptions of Tarski’s fixed point theorem are fulfilled, and we can apply
it. 

Finally, we want to note, that is not easy to see, how Theorem 9.3 can be proved
using Schauder’s fixed point theorem.

122
10 Further Reading
To emphasize that mathematical research never ends we add some more fixed
point theorems for further reading. To avoid any kind of rate the following list
is given in alphabetical order:
• The fixed point theorem of Browder, see (17).
• The fixed point theorem of Darbo, see (26).
• The fixed point theorem of Herzog and Kunstmann, see (45).
• The fixed point theorem of Kakutani, see (50).
• The fixed point theorem of Kneser, see (53).
• The fixed point theorem of Lagler and Volkmann, see (56).
• The fixed point theorem of Lefschetz and Hopf, see (30).
• The fixed point theorem of Lemmert, see (58).
• The fixed point theorem of Sadovskii, see (77).
• The fixed point theorem of Tychonoff, see (100).
• The fixed point theorem of Weissinger, see (108).
Some of them are (just) generalizations of others, some of them have also very
nice applications. The reader might search for more fixed point theorems or the
reader might extend the presented applications by his own or, even better, the
reader will apply some fixed point theorem in a new way. Anyway, it’s up to
you now.

123
Bibliography
[1] R. A. Adams, Sobolev spaces, Academic Press, New York, 1975.

[2] M. Aigner, G. M. Ziegler, Proofs from THE BOOK, Springer-Verlag,

2010.

[3] G. E. Alefeld, X. Chen, F. A. Potra, Numerical validation of solu-

tions of complementarity problems: the nonlinear case, Numer. Math., 92
(2002), pp. 1-16.

[4] G. Alefeld, A. Gienger, F. Potra, Efficient numerical validation

of solutions of nonlinear systems, SIAM J. Numer. Anal., 31 (1994), pp.
252-260.

[5] G. Alefeld, G. Mayer, The Cholesky method for interval data, Linear
Algebra Appl., 194 (1993), pp. 161-182.

[6] G. Alefeld, P. Volkmann, Regular splittings and monotone iteration

functions, Numer. Math., 46 (1985), pp. 213-228.

[7] G. Alefeld, J. Herzberger, Introduction to interval computations,

Academic Press, New York, 1983.

[8] G. Alefeld, Anwendungen des Fixpunktsatzes für pseudometrische

Räume in der Intervallrechnung, Numer. Math., 17 (1971), pp. 33-39.

[9] J. Appell, M. Väth, Elemente der Funktionalanalysis, Vieweg, Wies-

baden, 2005.

[10] S. Banach, Sur les opérations dans les ensembles abstraits et leur appli-
cation aux équations intégrales, Fund. math., 3 (1922), pp. 133-181.

[11] W. Benz, H. Karzel, A. Kreuzer, Emanuel Sperner, Gesammelte

Werke, Heldermann Verlag, 2005.

[12] P. Bod, On closed sets having a least element, Optimization and opera-
tions research. Lecture Notes in Econom. Math. Systems, 117 (1976), pp.
23-34.

125
Bibliography

[13] P. Bod, Minimality and complementarity properties of s.c. Z-functions

and a forgotten theorem due to Georg Wintgen, Math. Operationsforsch.
Statist., 6 (1975), pp. 867-872.

[14] K. C. Border, Fixed point theorems with applications to economics and

game theory, University Press, Cambridge, 1985.

[15] B. Breuer, P. J. McKenna, M. Plum, Multiple solutions for a semi-

linear boundary value problem: a computational multiplicity proof, J. Dif-
ferential Equations, 195 (2003), pp. 243-269.

[16] L. E. J. Brouwer, Über Abbildungen von Mannigfaltigkeiten, Math. An-

nalen, 71 (1912), pp. 97-115.

[17] F. Browder, On a generalization of the Schauder fixed point theorem,

Duke Math. J., 26 (1959), pp. 291-303.

[18] J. R. Cannon, Multiphase parabolic free boundary value problems, in: D.

G. Wilson, A. D. Solomon, P. T. Boggs (eds.), Moving boundary
problems, Academic Press 1978, pp. 3-24.

[19] R. Chandrasekaran, A special case of the complementary pivot prob-

lem, Opsearch, 7 (1970), pp. 263-268.

[20] Y. S. Choi, P. J. McKenna, A mountain pass method for the numerical

solutions of semilinear elliptic problems, Nonlinear Anal. Theory Methods
Appl., 20 (1993), pp. 417-437.

[21] L. Collatz, Funktionalanalysis und Numerische Mathematik, Springer-

Verlag, Heidelberg, 1964.

[22] R. W. Cottle, J.-S. Pang, R. E. Stone, The linear complementarity

problem, Academic Press, San Diego, 1992.

[23] J. Crank, Free and moving boundary problems, Clarendon Press 1984.

[24] J. Cronin, Fixed points and topological degree in nonlinear analysis,

Amer. Math. Soc., Mathematical Surveys and Monographs, 1964.

[25] E. N. Dancer, S. S. Yan, On the superlinear Lazer-McKenna conjecture,

J. Differential Equations, 210 (2005), pp. 317-351.

[26] G. Darbo, Punti uniti in trasformazioni a codominio non compatto,

Rend. Sem. Mat. Univ. Padova, 24 (1955), pp. 84-92.

[27] K. Deimling, Ordinary differential equations in Banach spaces, Springer-

Verlag Berlin 1977.

126
 Bibliography

[28] K. Deimling, Nichtlineare Gleichungen und Abbildungsgrade, Springer-

Verlag, Berlin Heidelberg 1974.
[29] J. Dieudonné, Deux exemples singuliers d’équationes différentielles, Acta
Sci. Math., 12B (1950), pp. 38-40.
[30] J. Dugundji, A. Granas, Fixed point theory, Springer-Verlag, 2003.
[31] F. Facchinei, J.-S. Pang, Finite-dimensional variational inequality and
complementarity problems, Volume I + II, Springer-Verlag, 2003.
[32] J. Franklin, Methods of mathematical economics. Linear and nonlinear
programming, fixed-point theorems, SIAM, 2002.
[33] A. Friedman, Partial differential equations of parabolic type, Prentice-
Hall, INC 1964.
[34] A. Frommer, F. Hoxha, B. Lang, Proving the existence of zeros using
the topological degree and interval arithmetic, J. Comput. Appl. Math.,
199 (2007), pp. 397-402.
[35] A. Frommer, B. Lang, Existence tests for solutions of nonlinear equa-
tions using Borsuk’s theorem, SIAM J. Numer. Anal., 43 (2005), pp. 1348-
1361.
[36] A. Frommer, B. Lang, M. Schnurr, A comparison of the Moore and
Miranda existence tests, Computing, 72 (2004), pp. 349-354.
[37] A. Frommer, Proving conjectures by use of interval arithmetic, in: Per-
spectives on enclosure methods, Springer-Verlag, 2001, pp. 1-13.
[38] A. N. Godunov, O teoreme Peano v banahovyh prostranstvah,
Funkcional’. Analiz Prilozén, 9 (1974), pp. 59-60.
[39] D. Hammer, Eine Verallgemeinerung des Algorithmus von Chandraseka-
ran zur Lösung nichtlinearer Komplementaritätsprobleme, Diplomarbeit,
Universität Karlsruhe, 2006.
[40] R. Hammer, M. Hocks, U. Kulisch, D. Ratz, Numerical toolbox for
verified computing. Volume I: Basic numerical problems. Theory, algo-
rithms, and Pascal-XSC programs, Springer-Verlag, 1993.
[41] E. Hansen, G. W. Walster, Global optimization using interval analysis.
Second edition, revised and expanded. With a foreword by Ramon Moore,
Marcel Dekker, Inc., New York, 2004.
[42] E. R. Hansen, Interval forms of Newton’s method, Computing, 20 (1978),
pp. 153-163.

127
Bibliography

[43] E. R. Hansen, On solving two-point boundary-value problems using in-

terval arithmetic. In: Topics in interval analysis, E. Hansen (Editor),
Clarendon Press, Oxford, (1969), pp. 74-90.
[44] P. T. Harker, J.-S. Pang, Finite-dimensional variational inequality and
nonlinear complementarity problems: A survey of theory, algorithms and
applications, Math. Programming, 48 (1990), pp. 161-220.
[45] G. Herzog, P. Kunstmann, A fixed point theorem for decreasing func-
tions, Numer. Funct. Anal. Optim., 34 (2013), pp. 530-538.
[46] H. Heuser, Lehrbuch der Analysis, Teil 2, Teubner, Stuttgart, 1983.
[47] D. Idczak, M. Majewski, A generalization of the Poincaré-Miranda
theorem with an application to the controllability of nonlinear repetitive
processes, Asian Journal of Control, 12 (2010), pp. 168-176.
[48] I. James, Remarkable Mathematicians, from Euler to von Neumann, Uni-
versity Press, Cambridge, 2002.
[49] S. Kakutani, Topological properties of the unit sphere in Hilbert space,
Proc. Imp. Acad. Tokyo, 19 (1943), pp. 269-271.
[50] S. Kakutani, A generalization of Brouwer’s fixed point theorem, Duke
Math. J., 8 (1941), pp. 457-459.
[51] J. B. Kioustelidis, Algorithmic error estimation for approximate solu-
tions of nonlinear systems of equations, Computing, 19 (1978), pp. 313-
320.
[52] R. Klatte, U. Kulisch, A. Wiethoff, C. Lawo, M. Rauch, C-XSC.
A C++ class library for extended scientific computing, Springer-Verlag,
1993.
[53] H. Kneser, Eine direkte Ableitung des Zornschen Lemmas aus dem
Auswahlaxiom, Math. Z., 53 (1950), pp. 110-113.
[54] R. Krawczyk, Newton-Algorithmen zur Bestimmung von Nullstellen mit
Fehlerschranken, Computing, 4 (1969), pp 187-201.
[55] W. Kulpa, The Poincaré-Miranda theorem, Amer. Math. Monthly, 104
(1997), pp. 545-550.
[56] M. Lagler, P. Volkmann Über Fixpunktsätze in geordneten Mengen,
Math. Nachr., 185 (1997), pp. 111-114.
[57] C. E. Lemke, J. T. Howson, Equilibrium points of bimatrix games, J.
Soc. Indust. Appl. Math., 12 (1964), pp. 413-423.

128
 Bibliography

[58] R. Lemmert, Existenzsätze für gewöhnliche Differentialgleichungen in

geordneten Banachräumen, Funkcial. Ekvac., Ser. Internac., 32 (1989),
pp. 243-249.

[59] J. Milnor, John Nash and “A beautiful mind”, Notices Amer. Soc., 45
(1998), pp. 1329-1332.

[60] C. Miranda, Un osservatione su un theorema di Brouwer, Belletino

Unione Math. Ital. Ser. II, (1940), pp. 5-7.

[61] R. E. Moore, J. B. Kioustelidis, A simple test for accuracy of approx-

imate solutions to nonlinear (or linear) systems, SIAM J. Numer. Anal.,
17 (1980), pp. 521-529.

[62] R. E. Moore, A test for existence of solutions to nonlinear systems,

SIAM J. Numer. Anal., 14 (1977), pp. 611-615.

[63] J. J. Moré, Classes of functions and feasibility conditions in nonlinear

complementarity problems, Math. Programming, 6 (1974), pp. 327-338.

[64] S. A. Morris, An elementary proof that the Hilbert cube is compact,

Amer. Math. Monthly, 91 (1984), pp. 563-564.

[65] M. T. Nakao, Solving nonlinear elliptic problems with result verification

using an H −1 type residual iteration, Computing (Suppl. 9) (1993), pp.
161-173.

[66] J. F. Nash, Noncooperative games, Ann. of Math., 54 (1951), pp. 286-295.

[67] M. Neher, The mean value form for complex analytic functions, Com-
puting, 67 (2001), pp. 255-268.

[68] J. M. Ortega, W. C. Rheinboldt, Iterative solutions of nonlinear

equations in several variables, Academic Press, New York, 1970.

[69] G. Owen, Game theory, W. B. Saunders Company, London, 1968.

[70] G. Peano, Démonstration de l’integrabilité des équations différentielles

ordinaires, Math. Ann., 37 (1890), pp. 182-228.

[71] M. Plum, Existence and multiplicity proofs for semilinear elliptic bound-
ary value problems by computer assistance, Jahresber. Deutsch. Math.-
Verein., 110 (2008), pp. 19-54.

[72] M. Plum, Computer-assisted proofs for partial differential equations,

spring tutorial seminar on computer assisted proofs - numeric and symbolic
approaches, 21st Century COE Program, Kyushu University, 2005.

129
Bibliography

[73] W. Poundstone, Prisoners’ dilemma, University Press, Oxford 1993.

[74] W. C. Rheinboldt, On M-functions and their applications to nonlinear

Gauss-Seidel iterations and to network flows, J. Math. Anal. Appl., 32
(1970), pp. 274-307.

[75] S. Robinson, The problem with blondes, SIAM News, 35 (2002), p. 20.

[76] S. Rump, INTLAB - INTerval LABoratory. In: Csendes, Tibor (ed.),

Developments in reliable computing, (1999), pp. 77-104.

[77] B. N. Sadovskii, On a fixed point principle, Funkcional. Anal. i Prilozen,

1 (1967), pp. 74-76.

[78] U. Schäfer, A fixed point theorem based on Miranda, Fixed Point Theory
Appl., Vol. 2007, Article ID 78706 (2007), 5 pages.

[79] U. Schäfer, On Tamir’s algorithm for solving the nonlinear complemen-

tarity problem, PAMM, Proc. Appl. Math. Mech., 7 (2007), pp. 2060057-
2060058.

[80] U. Schäfer, On computer-assisted proofs for solutions of linear comple-

mentarity problems, J. Comput. Appl. Math., 199 (2007), pp. 257-262.

[81] U. Schäfer, Wie erklärt man ein Nash-Gleichgewicht, Elem. Math., 62

(2007), pp. 1-7.

[82] U. Schäfer, M. Schnurr, A comparison of simple tests for accuracy

of approximate solutions to nonlinear systems with uncertain data, J. Ind.
Manag. Optim., Vol. 2, No. 4 (2006), pp. 425-434.

[83] U. Schäfer, Unique solvability of an ordinary free boundary problem,

Rocky Mountain J. Math., 34 (2004), pp. 341-346.

[84] U. Schäfer, Accelerated enclosure methods for ordinary free boundary

problems, Reliab. Comput., 9 (2003), pp. 391-403.

[85] U. Schäfer, Aspects for a block version of the interval Cholesky algo-
rithm, J. Comput. Appl. Math., 152 (2003), pp. 481-491.

[86] U. Schäfer, Two ways to extend the Cholesky decomposition to block

matrices with interval entries, Reliab. Comput., 8 (2002), pp. 1-20.

[87] U. Schäfer, An existence theorem for a parabolic differential equation in

l∞ (A) based on the Tarski fixed point theorem, Demonstratio Math., 30
(1997), pp. 461-464.

130
 Bibliography

[88] J. Schauder, Der Fixpunktsatz in Funktionalräumen, Studia Math., 2

(1930), pp. 171-180.
[89] M. Schnurr, Computing slope enclosures by exploiting a unique point of
inflection, Appl. Math. Comput., 204 (2008), pp. 249-256.
[90] M. Schnurr, D. Ratz, Slope enclosures for functions given by two or
more branches, BIT, 48 (2008), pp. 783-797.
[91] M. Schnurr, On the proofs of some statements concerning the theorems
of Kantorovich, Moore, and Miranda, Reliab. Comput., 11 (2005), pp.
77-85.
[92] Y. A. Shashkin, Fixed points, American Mathematical Society, 1991.
[93] A. Simon, I. Ould-Ahmed-Izid-Bih, I. Moutoussamy, P. Volk-
mann, Structures ordonnées et équations elliptiques semi-linéaires, Rend.
Circ. Mat. Palermo, 41 (1992), pp. 315-324.
[94] A. Simon, P. Volkmann, Équations elliptiques dans les espaces de Ba-
nach ordonnées, C. R. Acad. Sci. Paris, t.315, Serie I, (1992), pp. 1245-
1248.
[95] D. R. Smart, Fixed point theorems, University Press, Cambridge, 1974.
[96] E. Sperner, Neuer Beweis für die Invarianz der Dimensionszahl und des
Gebietes, Abh. Math. Sem. Hamburg, 6 (1928), pp. 265-272.
[97] A. Tamir, Minimality and complementarity properties associated with Z-
functions and M-functions, Math. Programming, 7 (1974), pp. 17-31.
[98] A. Tarski, A lattice-theoretical fixpoint theorem and its applications, Pa-
cific J. Math., 5 (1955), pp. 285-309.
[99] R. C. Thompson, A note on monotonicity properties of a free boundary
problem for an ordinary differential equation, Rocky Mountain. J. Math.,
12 (1982), pp. 735-739.
[100] A. Tychonoff, Ein Fixpunktsatz, Math. Ann., 111 (1935), pp. 767-776.
[101] R. S. Varga, Matrix iterative analysis, Prentice-Hall, 1962.
[102] P. Volkmann, Cinq cours sur les équations différentielles dans les espaces
de Banach, In: A. Granas, M. Frigon (eds.), Topological methods in
differential equations and inclusions, Kluwer Dordrecht 1995, pp. 501-520.
[103] P. Volkmann, Existenzsätze für gewöhnliche Differentialgleichungen in
Banachräumen, Tech. Univ. Berlin, Berlin 1985, pp. 271-287.

131
Bibliography

[104] M. N. Vrahatis, A short proof and a generalization of Miranda’s exis-

tence theorem, Proc. Am. Math. Soc., 107 (1989), pp. 701-703.
[105] W. Walter, Analysis II, Springer-Verlag, 1990.
[106] W. Walter, Gewöhnliche Differentialgleichungen, Springer-Verlag, 1990.
[107] W. Walter, There is an elementary proof of Peano’s existence theorem,
Am. Math. Monthly, 78 (1971), pp. 170-173.
[108] J. Weissinger, Zur Theorie und Anwendung des Iterationsverfahrens,
Math. Nachr., 8 (1952), pp. 193-212.
[109] F. Wilhelm, Der Satz von Miranda und der Algorithmus von Tamir,
Diplomarbeit, Universität Karlsruhe, 2008.
[110] N. Yamamoto, M. T. Nakao, Numerical verifications of solutions for el-
liptic equations in nonconvex polygonal domains, Numer. Math., 65 (1993),
pp. 503-521.

132
Index

C([a, b]), 80 continuous, 17

C0∞ (Ω), 105 convex, 17, 88
IRm×n , 61 custodial judge, 29
Rk+ , 46
inf[a], 56 degree of a mapping, 72
sup[a], 56 discrete metric, 18
c0 , 80 distributivity law, 57
l2 , 80
l2 (Z), 87 edge, 2
l∞ , 80 equicontinuous, 97

A beautiful mind, 25, 31, 129 free boundary problem, 44

arbitrator, 28
guess, 56
Ascoli-Arzela, 98
Hansen’s slope matrix, 71
Banach space, 78
Hilbert cube, 93, 129
Banach’s fixed point theorem, 78
barycentric coordinates, 9 increasing function, 109
barycentric subdivision, 11 infimum of a set, 47
bets, 25 initial value problem, 81
Bolzano-Weierstraß, 89 intermediate-value theorem, 51
boundary value problem, 58 interval evaluation, 58
bounded, 17 interval matrix, 61
Brouwer’s fixed point theorem, 23 INTLAB, 73
isomorphic, 85
C-XSC, 73
Cauchy sequence, 78 Krawczyk operator, 63
chip, 1
Cholesky method, 60 least element, 46
closed, 17 linear complementarity problem, 31,
compact embedding, 105 43
compact function, 95 linear space, 78
compact set, 92
complete lattice, 110 machine number, 73

133
Index

mean-value theorem, 62, 75, 120, 121 vector space, 77

melting ice, 44 vertex, 1

Nash’s equilibrium, 28 weak derivatives, 105

Nash, John, 25
network providers, 26 Z-function, 46
nondegenerate, 10
nonlinear complementarity problem,
43
norm, 14
norm equivalence in Rn , 15
normed vector space, 78

open set, 88
operator, 95

Pascal-XSC, 73
Peano’s existence theorem, 97
Picard-Lindelöf, 81
Poincaré-Miranda theorem, 33
polytope, 8
prisoners’ dilemma, 29

radial function, 21

Schauder’s fixed point theorem, 95

simplex, 8
splitting of the matrix A, 82
strategy
mixed strategy, 26
optimal pair of strategies, 28
pair of strategies, 27
pure strategy, 26
sub-distributivity law, 58
suspect, 29

Tamir’s algorithm, 49
Tarski’s fixed point theorem, 110
Taylor’s formula, 45, 59, 106
tolerance, 54
topologically equivalent, 21
two-person game, 25

vector addition, 76

134
from sperner‘s lemma to
differential equations
in banach spaces
An Introduction to Fixed Point
Theorems and their Applications
ISBN 978-3-7315-0260-9
9 783731 502609