Universitext

Editorial Board
(North America):

J.H. Ewing
F.W. Gehring
P.R. Halmos
Universitext
Editors (North America): J.H. Ewing, F.W. Gehring, and P.R. Halmos

Aksoy/Khamsi: Nonstandard Methods in Fixed Point Theory

Aupetit: A Primer on Spectral Theory
Bachumikern: Linear Programming Duality
Benedetti/Petronio: Lectures on Hyperbolic Geometry
Berger: Geometry I, II (two volumes)
Bliedtner/Hansen: Potential Theory
BoossIBleecker: Topology and Analysis
CarlesonlGamelin: Complex Dynamics
Cecil: Lie Sphere Geometry: With Applications to Submanifolds
Chandrasekharan: Classical Fourier Transforms
Charlap: Bieberbach Groups and Flat Manifolds
Chern: Complex Manifolds Without Potential Theory
Cohn: A Classical Invitation to Algebraic Numbers and Class Fields
Curtis: Abstract Linear Algebra
Curtis: Matrix Groups
van Dalen: Logic and Structure
Das: The Special Theory of Relativity: A Mathematical Exposition
DiBenedetto: Degenerate Parabolic Equations
Dimca: Singularities and Topology of Hypersurfaces
Edwards: A Formal Background to Mathematics I alb
Edwards: A Formal Background to Mathematics II alb
Emery: Stochastic Calculus
Foulds: Graph Theory Applications
Frauenthal: Mathematical Modeling in Epidemiology
Fukhs/Rokhlin: Beginner's Course in Topology
GallotIHulinlLafontaine: Riemannian Geometry
Gardiner: A First Course in Group Theory
GArding/Tambour: Algebra for Computer Science
Godbillon: Dynamical Systems on Surfaces
Goldblatt: Orthogonality and Spacetime Geometry
Hahn: Quadratic Algebras, Clifford Algebras, and Arithmetic Witt Groups
Hiawka/Schoissengeier/Taschner: Geometric and Analytic Number Theory
Holmgren: A First Course in Discrete Dynamical Systems
HowelTan: Non-Abelian Harmonic Analysis: Applications of SL(2,R)
Humi/MiIler: Second Course in Ordinary Differential Equations
Hurwitz/Kritikos: Lectures on Number Theory
Iversen: Cohomology of Sheaves
Jones/Morris/Pearson: Abstract Algebra and Famous Impossibilities
Kelly/Matthews: The Non-Euclidean Hyperbolic Plane
Kempf: Complex Abelian Varieties and Theta Functions
Kostrikin: Introduction to Algebra
Krasnoselskii/Pekrovskii: Systems with Hysteresis
Luecking/Rubel: Complex Analysis: A Functional Analysis Approach
MacLane/Moerdijk: Sheaves in Geometry and Logic
Marcus: Number Fields
McCarthy: Introduction to Arithmetical Functions
Meyer: Essential Mathematics for Applied Fields

(continued after List of Symbols)

Richard A. Holmgren

A First Course in
Discrete Dynamical
Systems

With 55 Figures

Springer-Verlag
New York Berlin Heidelberg London Paris
Tokyo Hong Kong Barcelona Budapest
Richard A. Holmgren
Department of Mathematics
Allegheny College
Meadville, PA 16335-3902
USA

Editorial Board
(North America):
J.H. Ewing F. W. Gehring
Department of Mathematics Department of Mathematics
Indiana University University of Michigan
Bloomington, IN 47405 Ann Arbor, MI 48109
USA USA

P.R. Halmos
Department of Mathematics
Santa Clara University
Santa Clara, CA 95053
USA

Mathematics Subject Classifications (1991): 39A12, 58F13, 58F20

Library of Congress Cataloging-in-Publication Data

Holmgren, Richard A.
A first course in discrete dynamical systems/Richard A.
Holmgren.
p. cm.
Includes bibliographical references and index.

1. Differentiable dynamical systems. I. Title.

QA614.8.H65 1994
514' .74-dc20 93-43894

Printed on acid-free paper.

© 1994 Springer-Verlag New York, Inc.

All rights reserved. This work may not be translated or copied in whole or in part without the
written permission of the publisher (Springer-Verlag New York, Inc., 175 Fifth Avenue, New
York, NY 10010, USA), except for brief excerpts in connection with reviews or scholarly
analysis. Use in connection with any form of information storage and retrieval, electronic
adaptation, computer software, or by similar or dissimilar methodology now known or hereaf-
ter developed is forbidden.
The use of general descriptive names, trade names, trademarks, etc., in this publication, even
if the former are not especially identified, is not to be taken as a sign that such names, as
understood by the Trade Marks and Merchandise Marks Act, may accordingly be used freely
by anyone.

Production managed by Francine McNeill; manufacturing supervised by Genieve Shaw.

Photocomposed copy prepared using the author's AMS-LaTeX fIles.
Printed and bound by R.R. Donnelley and Sons, Harrisonburg, VA.

987654321

ISBN-13: 978-0-387-94208-7 e-ISBN-13: 978-1-4684-0222-3

DOl: 10.1007/978-1-4684-0222-3
Preface

An increasing number of colleges and universities are offering undergradu-

ate courses in discrete dynamical systems. This growth is due in part to the
proliferation of inexpensive and powerful computers, which have provided
access to the interesting and complex phenomena that are at the heart of
dynamics. A second reason for introducing dynamics into the undergradu-
ate curriculum is that it serves as a bridge from concrete, often algorithmic
calculus courses, to the more abstract concepts of analysis and topology.
Discrete dynamical systems are essentially iterated functions, and if
there is one thing computers do well, it is iteration. It is now possible
for anyone with access to a personal computer to generate beautiful images
whose roots lie in discrete dynamical systems. The mathematics behind
the pictures are beautiful in their own right and are the subject of this
text. Every effort has been made to exploit this opportunity to illustrate
the beauty and power of mathematics in an interesting and engaging way.
This work is first and foremost a mathematics book. Individuals who read
it and do the exercises will gain not only an understanding of dynamical
systems, but an increased understanding of the related areas in analysis as
well.

How to use this book. This text is suitable for a one-semester course
on discrete dynamical systems. It is based on notes from undergraduate
courses that I have taught over the last few years. The material is intended
for use by undergraduate students with a year or more of college calculus
vi Preface

behind them. Students in my courses have come from numerous disciplines;

most have been majors in other disciplines who are taking mathematics
courses because they have a general interest in the subject. Concepts from
calculus are reviewed as necessary. In particular, Chapters 2 and 3 are
devoted to a review of functions and the real numbers. My students have
found the material in these chapters to be extremely useful as background
for the subsequent chapters. Other concepts are reviewed or introduced in
later chapters.
The interdependence of Chapters 1 to 11 is fairly deep, and these should
be covered sequentially. Students with a good background in real anal-
ysis can skip Chapters 2 and 3. On the other hand, students with only
a year of calculus, and little or no experience reading and writing math-
ematical proofs, are especially encouraged to read these chapters and do
the exercises. They are intended to provide the mathematical sophistica-
tion necessary to handle the remaining portions of the book. Chapters 14
and 15 can be covered immediately after Chapter 11 with the exception of
Section 14.5, which has Chapter 12 as a prerequisite.
Since the heart of any good mathematics textbook is the exercises, I have
provided a liberal quantity of interesting ones. The exercises range from
computational to those requiring a proof. A large number of the exercises
involving the theory can, at the instructor's discretion, be answered with
descriptive paragraphs or drawings rather than formal proofs. Some of the
exercises will be assumed later in the text. These are marked by a black dot
(.) and should not be skipped. Particularly difficult exercises are marked
by a star (*) or, in some cases, even a double star (**).
All students are encouraged to tackle the star problems. Trying to solve
them will deepen one's understanding of the material, even if the particular
exercise is never completed. Indeed, there are one or two for which I do
not have a complete solution, but which are very interesting (and fun) to
work on.
In some cases, it is nearly impossible to complete an exercise without
assistance from a computer. Mathematic6f'Y or a similar package is an ex-
cellent resource for doing most of them. The relevant Mathematica code
is provided in the appendix. It is also very easy to write simple programs
that will assist with the exercises. Details and references are provided in
the bibliography and the appendix.

Acknowledgments. Numerous individuals have assisted in the devel-

opment of this text. First and foremost I would like to thank my students
whose interest and enthusiastic responses encouraged me to write it all
down. I am particularly grateful to Crista Coles, Joe Cary, and James
Gill, who worked most of the exercises in the first 10 chapters and pro-
vided invaluable feedback on the wording and presentation.
 Preface vii

Mark Snavelly and his students at Carthage College used a preliminary

version of the text and provided additional suggestions and encourage-
ment. The idea for the chapter on bifurcations was entirely Mark's. Later
refinements in this chapter were suggested by Roger Kraft. All errors and
misrepresentations were added by the author. Roger Kraft and George
Day deserve special recognition for their careful reading and correction of
the entire manuscript. Finally, I would like to express my gratitude for the
support of my friends and family, without which this work would not be
possible.
The graphics in this book were created on a NeXT™ computer using
Mathematica and the draw program, which comes bundled with the NeXT
operating system. Typesetting was accomplished using AMS-l1\1E;X and
the AMSFonts.

Richard A. Holmgren
Meadville, Pennsylvania
Contents

Preface v
How to use this book . v
Acknowledgments . vi

1 Introduction 1
1.1 Phase Portraits 4
Exercise Set 1 . . . . 6

2 A Quick Look at Functions 9

Exercise Set 2 . . . . . . . . . . 16
3 The Topology of the Real Numbers 21
Exercise Set 3 . . . . . . . . . . . . . 27

4 Periodic Points and Stable Sets 31

4.1 Graphical Analysis 36
Exercise Set 4 . . . . . . . 38

5 Sarkovskii's Theorem 41
Exercise Set 5 . . . . . . . 45

6 Differentiability and Its Implications 47

Exercise Set 6 . . . . . . . . . . . . . . . . 54
x Contents

7 Parametrized Families of Functions and Bifurcations 57

Exercise Set 7 . . . . . . . . . . . . 65

8 The Logistic Function, Part I 67

Exercise Set 8 . . . . . . . . . . . . 75

9 Symbolic Dynamics and Chaos 79

Exercise Set 9 . . . . . . . . . . . . . 90

10 The Logistic Function, Part II: Topological Conjugacy 95

Exercise Set 10 . . . . . . . . . . . . . . . . . . . . . . . . . .. 102

11 The Logistic Function, Part III 105

Exercise Set 11 . . . . . . . . . . . . 111

12 Newton's Method 115

12.1 Newton's Method for Quadratic Functions. 121
12.2 Newton's Method for Cubic Functions 126
12.3 Intervals and Rates of Convergence . 134
Exercise Set 12 . . . . . . . . . . . . . . . . 135

13 Numerical Solutions of Differential Equations 141

Exercise Set 13 . . . . . . . . . . . . . . . . . . . . . 152

14 The Dynamics of Complex Functions 155

14.1 The Complex Numbers. . . . . . . . . 155
14.2 Complex Functions . . . . . . . . . . . 158
14.3 The Dynamics of Complex Functions . 162
14.4 The Riemann Sphere. . . . . . . . . . 166
14.5 Newton's Method in the Complex Plane 170
Exercise Set 14 . . . . . . . . . . . . . . . . . 176

15 The Quadratic Family and the Mandelbrot Set 181

Generating Julia and Mandelbrot Sets on a Computer 185
Exercise Set 15 . . . . . . . . . . . . . . . . . . . . . . . . . 187

AppendixA.
Computer Algorithms 191
A.l Iterating Functions . . . . . . . . . . . . . . . 191
Finding the Value of a Point Under Iteration 191
Tables of Iterates . . . . . . . . . . . . . . . . 192
Controlling the Precision of the Computations. 193
Graphing Iterated Functions. 193
A.2 Graphical Analysis . . 194
A.3 Bifurcation Diagrams. 196
A.4 Julia Sets . . . . . . . 197
 Contents xi

A.5 The Mandelbrot Set . . . . . . . 199

A.6 Stable Sets of Newton's Method. 200

References 203
Dynamical Systems. . . . . . . . . . 203
Topics in Mathematics . . . . . . . . 205
General Interest Books on Dynamics 206
Computer Programs and Algorithms 206

Index 209

List of Symbols 213

1
Introduction

A discrete dynamical system can be thought of as a function which is com-

posed with itself over and over again. For example, consider the function
f(x) = -x3. If we compose f with itself, then we get
f2(X) = (f 0 f)(x) = -( _x 3)3 = x 9.
Iterating the process we get

f3(X) = (f 0 f 0 f)(x) = (f 0 p)(x) = _(x 9 )3 = _x 27 ,

f4(x) = (f 0 f 0 f 0 f)(x) = (f 0 f3)(X) = -( _x 27 )3 = x 8 1,

where n is a natural number.

We would like to answer the question, "Given a real number x, what is
lim fn(x)?" More generally we ask, "What properties does the sequence
n-->oo
X, f(x), P(x), J3(x), ... have?" By the notation r(x) we mean f com-
posed with itself n - 1 times, not the nth power of f nor the nth derivative
of f. By fO(x) we mean the identity function, that is, fO(x) = x. We call
the behavior of points under iteration of the function the dynamics of the
function.
Let us visualize these questions another way. Suppose each of us lives
someplace on the real line and our address is given by the real number on
2 1. Introduction

which our apartment is set. For instance my address might be 2. Each year
the government has decreed that I must move to a new apartment whose
address can be found by cubing my present address and then finding the
additive inverse. That is, I apply the function f(x) = -x3 to my present
address to find my new address. So, I will be living at f(2) = _2 3 = -8
next year. The year after I will be living at f( -8) = j2(2) = _(-8)3 = 512
and after n years I will be living at r(2) = (_1)n23 n • Even ifllive to be
a very old man it is clear that I will never live in the same place twice. In
fact, each year the absolute value of my address will quickly become larger,
and I will move from one side of zero to the other.
Hence, if we start at the point 2, then fn(2) grows without bound in
absolute value and oscillates from one side of 0 to the other. What happens
if we start someplace else? Suppose we begin at the point ~. Then the
next year we will be at fG) = -~. The following year we will be at
f(-~) = f2(~) = 5~2 and after n years we will be at r(~) = (_1)n~.
So each year we will move from one side of 0 to the other, but over time
the move will not be very far.
In general, our address after n years will be r(xo) = (_1)n(xo)3 n ,
where Xo is our first address. If Ixol > 1 we see that
lim
n--+(X)
Ir(xo)1 = lim
n--+ 00
Ixol3n = 00.

r
In absolute value (xo) will increase without bound and the factor (-1) n
will cause r(xo) to oscillate from one side of 0 to the other. On the other
hand, if 0 < Ixol < 1, then lim r(xo) = O. So while r(xo) will still move
n->oo
from one side of 0 to another, its value will approach 0 as n gets larger.
In terms of choosing apartments, an apartment whose address is less than
one in absolute value would be preferable to an apartment whose address
is larger than one since, as time passes, we won't have to move so far each
year.
It remains to see what happens to -1, 0, and 1. Note that f(-1) = 1
and f(1) = -1. Hence, -1 and 1 form a periodic cycle. Thus, if we were
living at 1 and could work out a deal with the person living at -1, then
we could probably leave some of our belongings at each address since we'll
be back every other year, a luxury indeed. Of course, if we lived at 0, then
we'd not have to move at all, since 0 is fixed by f: a stable and perhaps
boring existence.
To recap, a discrete dynamical system consists of a function and its
iterates. If a dynamical system is given, we would like to know where each
point goes as we iterate the function, and what route it takes to get there.
In our example, f(x) = -x3, we have seen a number of possibilities. There
are points which head off to infinity, others which get close to 0, a pair
of points which oscillate back and forth periodically, and a point which
 1. Introduction 3

doesn't go anywhere at all. As we continue our investigations the list of

possibilities will grow.
Before we go on however, it will be useful to look at the motivation
behind the study of discrete dynamical systems. In the whimsical example
above we used the function f(x) = -x3 to determine the location of our
next apartment. By iterating the function we were able to find out where
we would be in any future year. Functions and their iterates can be used
to model more practical problems as well.

EXAMPLE 1.1.
Suppose we have a population of rabbits which is growing by 10% an-
nually. Let's assume we start with Xo rabbits and denote the population
in the nth year by X n . How many rabbits will we have in n years, or
equivalently, what is the value of xn?
Clearly,

Xl = Xo + .1xo = 1.1xo
X2 = Xl + .1XI = 1.1xI

Thus, Xi+l = P(Xi) where p(x)1.1x. So,

Xl = p(xo)
X2 = p(xd = (p 0 p)(xo) = p2(xO)
X3 = p(X2) = (p 0 p2)(xO) = p3(xO)

Therefore, the population of rabbits after n years is determined by ap-

plying the function p(x) = 1.1x to Xo n times. A simple calculation shows
that pn(x) = (1.1)nx. Thus, if we start with more than one rabbit, the
population will grow and keep growing forever. For example, a population
of 10 rabbits will swell to a population of about 26 rabbits in 10 years.
However, according to this model, those same 10 rabbits will swell to a
population of approximately 67 in 20 years, 1174 in 50 years, and 137,806
in 100 years. This illustrates the principal flaw in exponential models of
populations: unlimited growth. 0
4 1. Introduction

In general, models which iterate a function of the form g(x) = rx

are called exponential models. As we demonstrated above, in popula-
tion problems exponential models have limited predictive power when the
value of n becomes large since the predicted population becomes so large
that it is no longer realistic. A more sophisticated model for population,
which takes into account the limits on growth, uses the logistic function,
h(x) = rx(l - x).
Models using the logistic function assume there is an absolute limit on
the size of the population and designate the size of the population as a
fraction of the limit. Hence, the size of any population will be denoted by a
humber in the interval [0,1]. For example, .25 indicates that the population
is 25% of the limit population. If Xo is the population in the first time
period, then the population in the next time period is h( xo) = rxo (1 - xo).
The factor (1 - x) distinguishes models using the logistic function from
exponential models. As x approaches 1 this factor approaches O. Hence,
as x becomes larger the population grows at a slower rate. If x is large
enough, then the population will decline.
We will see in the following chapters that the dynamics of models based
on the logistic function are much harder to understand than exponential
models. In fact, much of this book is devoted to understanding this one
example.

1.1. Phase Portraits

A phase portrait is a graphical representation of the dynamics of a sys-
tem. It consists of a diagram representing possible beginning positions in
the system and arrows which indicate the change in these positions un-
der iteration of the function. It is best understood by looking at a few
examples.

EXAMPLE 1.2.
Let 1(x) = x 2 • The dynamical system we are considering consists of the
domain of 1 and the function itself. The domain is the set of real numbers
which we represent by a line. Note that 0 and 1 are fixed, that is, 1(0) = 0
and 1(1) = 1. We indicate this in our phase portrait by dots at 0 and 1.
We notice that if 0 < x < 1, then In(x) tends towards 0 as n becomes
larger and if x> 1, then In(x) tends towards infinity as n becomes larger.
To represent this, we draw an arrow from 1 towards 0 and another from
one towards positive infinity. Now what happens if x is less than O? The
point -1 becomes fixed at 1 since 1(-1) = 1. This implies r( -1) = 1 for
all n larger than or equal to 1. This is shown by the arrow from -1 to 1.
Points which lie between -1 and 0 are mapped into the interval (0,1) and
 1.1. Phase Portraits 5

FIGURE 1.1. The phase portrait of f(x) = x 2 .

FIGURE 1.2. The phase portrait of f(x) = -x 3 .

then they move towards 0 under iteration of the function. Similarly, points
which are to the left of -1 are mapped to the right of 1 and then move
towards infinity under iteration of the map. We represent this by arrows
from the negative portion of the line to the positive portion. All of this
information is encoded in the phase portrait shown in Figure 1.1. 0

EXAMPLE 1.3.
Consider the function f(x) = -x 3 . 0 is fixed and the point 1 goes to
-1 and then returns to 1 on the second iteration. Points which are greater
than 1 in absolute value oscillate from side to side of zero under iteration
of f and grow ever larger in absolute value. Points which are less than 1 in
absolute value oscillate from side to side of zero under iteration of f and
grow ever smaller in absolute value. All of this is encoded in the phase
portrait shown in Figure 1.2. 0
The reader will be given an opportunity to explore and develop more
phase portraits in the exercises.
6 1. Introduction

Exercise Set 1

1.1 Suppose h(x) = rx(l - x).

a) Let r = 2 and calculate hn(.25) for n = 0,1,2,3, and 4 and
n = 101, 102, 103, 104. Can you determine limn-+oo hn(.25)?
b) Let r = 2 and calculate hn(.2345) for n = 0,1,2,3, and 4 and
n = 101, 102, 103, 104. Can you determine limn-+ oo hn(.2345)?
c) Let r = 3 and calculate hn(.25) for n = 0,1,2,3, and 4 and
n = 101, 102, 103, 104. Can you determine limn-+oo hn(.25)?
d) Let r = 3 and calculate hn(.2345) for n = 0,1,2,3, and 4 and
n = 101, 102, 103, 104. Can you determine limn-+ oo hn(.2345)?
e) Let r = 4 and calculate hn(.25) for n = 0,1,2,3, and 4 and
n = 101, 102, 103, 104. Can you determine limn-+oo hn(.25)?
f) Let r = 4 and calculate hn(.2345) for n = 0,1,2,3, and 4 and
n = 101, 102, 103, 104. Can you determine limn-+ oo hn(.2345)?

1.2 Look up Newton's method for approximating the zeros of a func-

tion in your calculus textbook and describe it as a dynamical sys-
tem. What is the significance of limn-+oo r(x) in this case?

1.3 Draw the Phase Portraits for the functions below.

a) f(x) = x 1 / 3
b) g(x) = arctan x
c) r(x) = ~
d) C(x) = cos(x)

1.4 On page 4 we were introduced to the logistic function as a model

for population growth. Recall that the logistic function is defined
as h(x) = rx(l - x). We stated that if x is large enough, then
the population will decline. Suppose r = 2.5. Find a point Xo in
[0,1] so that h(x) :::; x if and only if Xo :::; x. What happens to the
size of the population over time if the initial population density is
xo? What if the initial density is a little larger than xo? a little
smaller? Answer the same questions for r = 3.4 and r = 1.

1.5 AN INVESTIGATION: Let h(x) = rx(l - x). Investigate the chang-

ing behavior of the finite sequence h 500 (.1), h 501 (.1), ... , h 535 (.1)
 Exercise Set 1 7

as r varies from 3 to 3.6. You might begin by looking at the r

values 3, 3.1, 3.2, 3.3, 3.4, 3.5, and 3.6. Between which values
do you see a qualitative change in behavior? If you see a change
between r = 3.4 and r = 3.5, then look at the behavior at the mid-
point r = 3.45. Continue to focus on smaller and smaller intervals
around changes in behavior. You should find that a change in the
r value as small as .005 can be significant. Describe these changes
and the places they occur to the best of your abilities.

1.6 AN INVESTIGATION: If you have been using a computer to inves-

tigate the behavior of a function under iteration, then you should
be aware that the program you are using rounds off to a fixed
number of significant digits after each iteration. The question is:
does this rounding matter? Write a short program in which the
number of significant digits used in calculations can be chosen by
the user. Such a program for Mathematica can be found in the
appendix. Investigate the behavior of points in [0,1] under itera-
tion of h(x) = rx(l - x) when the computations are done with a
precision of 8 significant digits and again when they are done with
99 significant digits using various values of r between 3 and 4. Do
the number of significant digits matter? For which values of r?
2
A Quick Look at Functions

Before we begin our discussion of dynamics, it is necessary to review a

few definitions. We review the terms function, one-to-one, onto, continu-
ous, and inverse function. A good, working knowledge of these terms is
fundamental to understanding the following material. We also introduce
homeomorphisms and complete the section by reviewing the Intermediate
Value Theorem.

DEFINITION 2.1. A function is a rule which assigns each element of one

set to a unique element of a second set. The first set is called the domain of
the function and the second set is called the codomain. The set of elements
in the codomain which have an element of the domain assigned to them is
called the range of the function. We use the notation f : D -7 C to indicate
a function f with domain D and codomain C. The notation f : D -7 D
indicates that the domain and codomain of the function are the same set.

We often refer to functions as maps or mappings. Note that a function

isn't necessarily a map from one set of real numbers to another. While
functions of the real numbers are certainly important, they are by no means
the only examples. There are many important functions which are not
functions of the real numbers. When we do wish to refer to the set of real
numbers we use the symbol R Thus, the notation g : lR. -7 lR. indicates a
function 9 whose domain and codomain are both the set of real numbers.
10 2. A Quick Look at Functions

If f : D - C and A c D, then f(A) is defined to be the subset of C

containing all the elements of the form f(a) where a is in A. That is,
f(A) = {c in C I there is a in A satisfying f(a) = c}.

This set is often called the image of A under f. Notice that if f : D - C,

then f(D) is the range of f. If Bee, then f-1(B) is called the inverse
image or preimage of B and consists of all elements of D whose image is
contained in B. That is,
f-l(B) = {x in D I f(x) is in B}.

EXAMPLE 2.2.
Let f : lR -lR be defined by f(x) = x 2. Then
f([-1,2]) = [0,4] and f-1((1,4]) = [-2, -1) U (1,2].
Note that
f(1-1([1,4])) = [1,4], but f-1(1([1,2])) = [-2,-1] U [1,2].
Note also that the use of the notation f- 1 does not necessarily imply that
f is an invertible function. 0
DEFINITION 2.3. A function is one-to-one if there is exactly one element
of its domain assigned to each element of its range.

PROPOSITION 2.4. The function f is one-to one if and only if f(x) =

f(y) implies that x = y.

The "if and only if" in this proposition means that the two statements,
"f : D - C is one-to-one" and "f(x) = f(y) implies x = y" are equivalent.
In other words, if one of the statements is true, then the other is true; if one
of the statements is false, then the other is false. The proof of Proposition
2.4 is not hard and is left as an exercise.
Proposition 2.4 provides us an easy way to test whether or not a function
is one-to-one. For example, consider the function f : lR - lR defined by
f(x) = x 2. Clearly, f is not one-to-one since f(l) = 1 = f(-I). On the

°
other hand, the function g: [0,00) -lR defined by g(x) = x 2 is one-to-one
since if g(x) = g(y), then x 2 = y2, X ~ 0, and y ~ imply x = y. Note that
even though the functions f and 9 have the same rule, they have different
domains and, hence, different properties.
A graphical test for determining whether functions of the real numbers
are one-to-one is the horizontal line test. A function is one-to-one if and
only if every horizontal line crosses the graph of the function at most once.
By looking at the graphs of f(x) = x 2 and p(x) = X 1/ 3 we see that f is not
 2. A Quick Look at Functions 11

-2
-I -2

FIGURE 2.1. The graphs of f(x) = x2 and p(x) = xl / 3.

one-to-one and p is. The graphs of f and p are shown in Figure 2.1. In the
exercises we ask the reader to explain why the horizontal line test works.
DEFINITION 2.5. A function is onto if each element of the codomain has
at least one element of the domain assigned to it. In other words, a function
is onto if the range equals the codomain.
When determining whether or not a function is onto we must know
what the codomain is. For example, the function f : lR ---+ lR defined by
f(x) = eX is not onto. There is no point in the domain which is mapped
to the numbers in the interval (-00,0]. On the other hand, the furiction
g : lR ---+ (0,00) defined by g(x) = eX is onto. Of course, every function can
be made into an onto function by restricting its codomain to its range.
What can we say about the graph of an onto function whose domain and
codomain are subsets of the real numbers? If we consider the codomain
as a subset of the y-axis, then it is clear that every horizontal line which
passes through an element of the codomain must also pass through the
graph of an onto function. This geometric description is useful when we
are trying to construct examples of functions with specific properties.

EXAMPLE 2.6.
Construct an example of a function f : [1,3] ---+ [2,4] which is onto but
not one-to-one.
Let us first determine what the graph of f must look like. Since all the
values of f must lie in the codomain, the graph lies between the lines y = 2
and y = 4. As the domain of f is [1,3]' a vertical line will intersect the
graph if and only if it passes through a point on the x-axis between and
including 1 and 3. So the graph of the function lies in the box with corners
(1 , 2), (1 , 4), (3,4), and (3 , 2) as shown in Figure 2.2.
Since f is onto, we know that every horizontal line which passes through
the codomain (thought of as a subset of the y-axis) must intersect the graph
of the function. Also, since f is not one-to-one there must be at least one
12 2. A Quick Look at Functions

(1.4) (3,4)
- -+t--

(3,2)

2 4

FIGURE 2.2. The graph of a function, J : [1,3] --> [2,4],

which is onto, but not one-to-one.

horizontal line through the codomain which strikes the graph two or more
times. Obviously, there are many graphs which satisfy these criteria. One
example is shown in Figure 2.2. 0
While it is often sufficient to think of continuous functions as functions
whose graphs are unbroken lines, it is essential to have a more precise
definition. In particular, if a function's domain or codomain is not an
intervai of the real line, then the graph of the function is not always so easy
to draw and not always informative when it can be drawn. For example, if
Z is the set of integers, then the function J : Z --> Z defined by J(z) = 2z
is continuous, but its graph is a set of discrete points.
DEFINITION 2.7. Let J : D --> C and Xo be a point in D. Then J
is continuous at Xo if for every positive number E, there exists a positive
number 8 such that
if x is in D and Ixo - xl < 8, then IJ(xo) - J(x)1 < E.

A Junction is continuous iJ it is continuous at each point oj its domain.

To understand this definition it helps to think of Ix - yl as the distance
from x to y. Then Ix - yl < E simply means that the distance from x to
y on the real line is smaller than E. When contemplating the definition of
 2. A Quick Look at Functions 13

x I(x) E
• • • j(xo )
D c

FIGURE 2.3. An illustration of continuity at the point Xo.

continuity, visualize Xo and f(xo) as points in the domain and codomain of

f, as shown in Figure 2.3. Now draw a circle around f(xo) with a radius
of f and its center at f(xo). If f is continuous at xo, then we can draw a
circle around Xo whose radius 8 is so small that whenever x is in the circle
around Xo and in the domain of f, then f(x) is in the circle around f(xo).
Note that x is in the circle ofradius 8 around Xo if and only if Ix - xol < o.
That is, x is in the circle of radius 8 around Xo if and only if the distance
from x to Xo is less than o. Similarly, f(x) is in the circle around f(xo) of
radius f if and only if If(x) - f(xo)1 < f.

EXAMPLE 2.8.
Let Z be the set of integers. We use the definition to show that the
function f : Z -+ Z defined by f (z) = 2z is continuous.
Suppose a is any integer and f is any number greater than o. We must
find a number 8 which is greater than 0 and such that If(z) - f(a)1 < f
whenever z is an integer satisfying Iz - al < 8. Let's try 8 = If z is an !.
integer and Iz - al < !, then z must be equal a. Hence, If(z) - f(a)1 is
equal to 0, which is less than f. 0
Lest we think it is always easy to find a 8 which works , we consider
another example. However, in oder to do that we need to introduce the
Triangle Inequality.
The Triangle Inequality needn't be mysterious. As most of us realize,
it is further to go from New York to Chicago by way of Los Angeles than
it is by traveling directly; this is not surprising. Essentially the Triangle
Inequality tells us that the distance from the point a to the point b is less
than or equal to the distance from a to the point e plus the distance from
e to b. We can express the Triangle Inequality mathematically by

la - bl :::; la - el + Ie - bl (2.1)
where a, band e represent real numbers. Recall that the value la - bl
represents the distance from the point a to the point b. If we let x = a - e
and y = e - b in formula 2.1, then we get

Ix + yl :::; Ixl + IYI·

14 2. A Quick Look at Functions

This is th€ form of the Triangle Inequality which is most useful to us; we
use it in virtually every chapter of this text. The reader is well advised to
take the time to prove it and to learn how to use it. Exercise 2.8 at the end
of this chapter provides hints for proving the inequality and practice using
it. Our first use of the the Triangle Inequality is in the following example.

EXAMPLE 2.9.
Show that the function g(x) = x 2 is continuous on the real numbers.
Let a be any real number. We must show g that is continuous at a.
We consider the case where a is not O. Let 10 be any positive real number.
We must find a positive number 8 such that whenever x is a real number
satisfying Ix - al < 8, then If(x) - f(a)1 = Ix2 - a2 1< E. To accomplish
this, we choose 8 so that 8 < lal, 8 < 13:1' and 8 > O. This is possible since
both 10 and lal are positive numbers. If Ix - al < 8, then using the Triangle
Inequality we find that

Ixl = Ix - a + al ~ Ix - al + lal < lal + 8.

It follows that when Ix - al < 8, then

If(x) - f(a)1 = Ix2 - a2 1= Ix - allx + al

< 8lx+al
~ 8(lxl + lal)
< 8(lal + 8+ luI)
< 813al
10
< 13al13al = E.
Hence, If(x) - f(a)1 < 10 when Ix - al < 8 as desired. Thus, g is continuous
at each nonzero real number. Note that the Triangle Inequality was used
twice in the preceding sequence of inequalities; once in going from the
second line to the third and a second time in going from the third line to
the fourth. We leave the proof that g is continuous at 0 as an exercise. 0
Note that in proving that g was continuous at a in the previous example,
our choice of 8 depended on the value of both a and E. This dependence of
10 on the point in question is the rule rather than the exception.
The astute reader may have noticed that our definition of continuity only
makes sense if the notion of a distance is defined on the sets D and C. Given
our current knowledge, this means that D and C must be subsets of the
real numbers. When we look at metric spaces in Chapter 9, we will extend
the definition of continuity to functions whose domain and codomain may
be more general sets.
 2. A Quick Look at FUnctions 15

The Intermediate Value Theorem, which we introduce next, is a useful

property of continuous functions. An application of the Intermediate Value
Theorem is demonstrated in exercise 2.16. Further applications are found
in subsequent chapters.
THEOREM 2.10. THE INTERMEDIATE VALUE THEOREM. Let f be a
continuous function defined on the interval [a, b]. If p is a number between
f(a) and f(b), that is, f(a) ::; p ::; f(b) or f(b) ::; p ::; f(a), then there
exists c in [a, b] such that f(c) = p.
A proof ofthe Intermediate Value Theorem can be found in introductory
analysis textbooks such as those listed in the references. We note that the
condition that f be continuous is necessary. We demonstrate this in the
following example.

EXAMPLE 2.11.
I, for x < 1
Let f : [0,2] --+ lR be defined by f(x) = { . Then f(O) = 1
3, for x 2: 1
and f(2) = 3. Even though 2 is between f(O) and f(2) there is no c in
[0,3] such that f(c) = 2. D
DEFINITION 2.12. Let f : D --+ C be a function with range f(D). The
function 9 : f(D) --+ D is an inverse of f if (g 0 f)(x) = x for all x in D.
If f has an inverse, then it is usually denoted f- 1 and we say that f is
invertible.
The following proposition is well-known and its proof is left as an exer-
cise:
PROPOSITION 2.13. A function is invertible if and only if it is one-to-
one.
DEFINITION 2.14. If a function is one-to-one, onto, and continuous, and
its inverse is continuous, then it is a homeomorphism. In this case, we say
the domain and codomain are homeomorphic to one another.
Homeomorphisms play an important role in mathematics in general,
and in dynamical systems, in particular. When two mathematical spaces
or structures are homeomorphic they are, in some sense, the same. Math-
ematically, we describe this by saying that they have the same topolog-
ical properties. The intervals [1,2] and [3,5] are homeomorphic since
9 : [1,2] --+ [3,5] defined by g(x) = 2x + 1 is a homeomorphism. (The
reader should check that 9 is in fact one-to-one, onto, continuous and has a
continuous inverse.) However, the intervals [1,2] and (3,5) are not homeo-
morphic. That is, there is no function f : [1,2] --+ (3,5) which is one-to-one,
16 2. A Quick Look at Functions

onto, continuous, and has a continuous inverse. This is demonstrated in

exercise 2.16. In Chapter 3 we will see that the topologies of [1,2) and
(3,5) are different as well. Specifically, [1,2) contains all of its limit points
while (3,5) does not.
Note that in all of the examples in this book, if a function is one-to-one,
onto, and continuous, then it is a homeomorphism. This is not true in
general. There are functions that are one-to-one, onto, and continuous and
which are not homeomorphisms. The interested reader may wish to consult
General Topology by S. Willard or a similar reference for further details.

Exercise Set 2

2.1 Let f(x) = x 2 and 8(x) = sinx.

a) Find f([O, 1)).
b) Find f- 1(1([0, 1))) and f(l-1([0, 1))). Why aren't these two
sets the same?
c) Find 8([0,~)) and 8- 1 (8([0, ~))).

2.2 Let 9 : lR -t Z be the step function defined such that g(x) is

the largest integer less than or equal to x. (Z denotes the set of
integers.) Calculate g([!, 10.1)) and g-1({2,4}).

2.3 a) Let A be a subset of the domain of the function f. Prove

f-1(1(A)) :J A and show by example that it may happen that
f-1(1(A)) #- A.
b) Let B be a subset of the codomain of the function f. Determine
which of the following must be true and show that your answer is
correct.

2.4 Suppose f : [-1, l)-t [-1,1) is defined so that f(x) = !x.

a) Find the range of f.
b) Find the range of f 0 f.
c) Find the range of fn for all natural numbers n.
 Exercise Set 2 17

2.5 a) Find a function f : [0,2J ---+ [2,4J so that f is one-to-one and

continuous, but not onto.

b) Find a function f : [0,2J ---+ [2,4J so that f is onto and continu-

ous, but not one-to-one.

c) Find a function f : [0,2J ---+ [2,4J so that f is one-to-one and

onto, but not continuous.

d) Find a function f : [0,2J ---+ [2,4J so that f is a homeomorphism.

That is, f is one-to-one, onto, continuous, and has a continuous
inverse.

2.6 Prove Proposition 2.4.

2.7 Let f be a function whose domain and codomain are subsets of the
real numbers. Prove f is one-to-one if and only if every horizontal
line intersects the graph of f at most once.

2.8 • THE TRIANGLE INEQUALITY.

a) Let a and b be real numbers. Prove: la+ bl ~ lal + Ibl. You may
use the fact that -Ial ~ a ~ lal and -Ibl ~ b ~ Ibl or you may
wish to show that la + W
~ (Ial + Ib1)2.

b) Use part (a) to prove that Ix - yl ~ Ix - zl + Iy - zI.

c) Prove: Ix - yl ~ Ixl + Iyl·
d) Prove: Ixl :::: Ix - yl - Iyl·
e) Prove: Ixl-Iyl ~ Ix - yl·

2.9 a) Justify each of the steps in the second displayed inequality of

Example 2.9.

b) Prove that g(x) = x2 is continuous at 0.

2.10 Prove Proposition 2.13.

2.11 Show that f (x) = ~ is continuous.

18 2. A Quick Look at Functions

-1 for x < 0, . .
2.12 a) Show that f(x) = { 1
for x > ° IS contmuous.

-1 for x S 0, . .
b) Show that f(x) = { 1 IS not contmuous.
forx>O

2.13 Show that every function whose domain is the integers and whose
codomain is the real numbers is continuous.

2.14 • a) Show that all nonconstant linear functions are continuous.

Note: A nonconstant linear function has the form f(x) = mx+b
where m i=- 0 .
• b) Show that all nonconstant linear functions have continuous
inverses.
• c) Show that all nonconstant linear functions are homeomor-
phisms.

2.15 Prove that if a and b are real numbers satisfying a < b, then the
interval [a, b] is homeomorphic to [0,1].
Hint: Find a linear function f : [a, b]-t [0,1].

2.16 a) Let f : [1,2] - t (3,5) be one-to-one and continuous. Show that

f is not onto.
Hint: First show f is strictly increasing or strictly decreasing
by using the Intermediate Value Theorem.
Note: This exercise implies [1,2] and (3,5) are not homeomor-
phic.
b) Let 9 : (0,1) - t [1,3] be onto and continuous. Show that 9 is
not one-to-one.
Hint: First show there is Xo in (0,1) such that g(xo) = 3. What
is the nature of 9 near xo? You may wish to use the Intermediate
Value Theorem as well.
Note: This exercise implies (0,1) and [1,3] are not homeomor-
phic.
c) Let a, b, c, and d be real numbers satisfying a < band c < d.
Prove that the intervals (a, b) and [c, d] are not homeomorphic.

2.17 • Let f : D - t C and 9 : C -t E. Prove that the following

statements are true.
a) If f and 9 are onto, then go f is onto.
 Exercise Set 2 19

b) If f and 9 are one-to-one, then go f is one-to-one.

c) If f and 9 are continuous, then go f is continuous.
d) If f and 9 are homeomorphisms then go f is a homeomorphism.

2.18 • Let f : D ---7 C and 9 : D ---7 C be continuous functions. Prove

that f + 9 and f - 9 are continuous functions.

2.19 * Prove or disprove: If f : ~ ---7 ~ is one-to-one, onto, and con-

tinuous, then f is a homeomorphism. (Prove or disprove means
that you should either prove the statement or find an example of
a function, f : ~ ---7 ~, which is one-to-one, onto, and continuous,
but not a homeomorphism.)
3
The Topology of the
Real Numbers

The topology of a mathematical space is its structure or the characteristics

it exhibits. In calculus we were introduced to a few topological ideas and
we will need a few more in our study of dynamics.
One of the fundamental questions of dynamics concerns the properties
of the sequence x, f(x), P(x), P(x), .... To discuss these properties
intelligently we need to understand convergence, accumulation points, open
sets, closed sets, and dense subsets. In this section we limit our discussion
to subsets of the real numbers; we will revisit the definitions when we
introduce metric spaces in Chapter 9.

DEFINITION 3.1. Suppose U is a subset of the real numbers. Then U is

open if for each x in U there is a positive number E so that Ix - yl < E
implies y is in U.

As with the definition of continuity, it is useful to think of the E in this

definition as a distance. Picture U as a subset of the real line as shown
in Figure 3.1 and let x be a point in U. If U is open, then we can find an
E > 0 such that when we draw a circle around x with radius E, all the real
numbers contained in the circle are contained in U.
This disk with radius E and center x is usually called an E-neighborhood
or simply neighborhood of x. We can think of a neighborhood of x as all
the points which "live near by" x. We make this definition precise by
22 3. The Topology of the Real Numbers

FIGURE 3.1. An open set.

DEFINITION 3.2. Let E > O. The set NE(x) = {y in lR I Ix - yl < E} is

an E-neighborhood, or simply neighborhood of x.
Note that Ix - yl < E if and only if y is in the open interval (x - E, X + E).
Consequently, N€ (x) = (x - E, X + E) and we can think of neighborhoods of
x as open intervals which are centered at x.
The following two propositions follow from a careful examination of the
definitions of open sets and neighborhoods and should be proven by the
reader.
PROPOSITION 3.3. The set U is open if and only if for each x in U there
is a neighborhood of x which is completely contained in U .
PROPOSITION 3.4. Every neighborhood of a real number is an open set.

It is important to notice that when demonstrating a set is open, we only

need to find one positive E which works for each x. On the other hand,
in the case of continuity at a point, there must be an appropriate 8 for
every positive E. In the following example, we use the definition of open to
demonstrate that "open intervals" are indeed open and note that not every
set need be open.

EXAMPLE 3.5.
a) The interval (a , b) is open.
Let x be a point in (a, b). To show that (a, b) is open, we must find E > 0
so that Iy - xl < E implies y is in (a, b) . Suppose E is the smaller of Ix - al
and Ix - bl. In other words, suppose E is the distance from x to the nearest
endpoint of (a, b). Since a < x < b, E is positive. If Ix - yl < E, then
Ix - yl < E :::; Ix - al·
Since x > a, this implies
-x + a < x - y < x-a.
Solving this for y we find that
a < y < 2x - a.
 3. The Topology of the Real Numbers 23

Hence, a < y. To complete the proof that y is in (a, b), we need to show
that y < b. The proof is similar to the one above. The details are left as
an exercise.

b) The set of rational numbers is not an open set.

If x is a rational number, then no matter how small the positive number
E is, there is some irrational number whose distance from x is less than E.
We prove this formally in the exercises. 0

Since there are open sets in mathematics, it's not surprising there are
closed sets as well. To define closed sets we use the concepts of convergent
sequences and accumulation points.

DEFINITION 3.6. Let Xl, X2, X3,'" be a sequence of real numbers. The
sequence converges to X if for each E > 0, there exists N such that if k ;::: N,
then Ix-xkl < E. The sequence grows without bound or converges to infinity
if for each M, there exists N such that if k ;::: N, then Xk > M.

Again, we can think of E as a a distance. If the sequence Xl, X2, X3, ...
converges to X, then when k is large, Xk is close to x, that is, IXk - xl < E.
Notice that in the case of convergent sequences we must be able to find an
N which works for each E. The N may change as E changes and it is not
sufficient to find an N which works for one particular E.

EXAMPLE 3.7.
a) The sequence Xl = .9, X2 = .99, X3 = .999, ... converges to 1.
To see this, let E > O. Then there exists an integer N such that l~N < E.
Thus, 11 - Xk I = 16k < E for all k ;::: N as desired.
b) The sequence Xl = 1, X2 = 2, X3 = 1, X4 = 2, X5 = 1, ... does not
converge to any real number.
To demonstrate this, it is sufficient to find a single E for which there is
no N. Choose E = [Link] claim that if Xk is such that IXk - xl < then i,
IXk+! - xl > i·To see this suppose that Xk = 1; it has to be either 1 or 2.
Then Xk+l = 2 and by the Triangle Inequality,

Ixx+! - xl = 12 - xl
= 11 - (x - 1)1
3
;::: 111 -Ix - 11 > 4'
A similar result holds when .Tk = 2. Hence, when E = i, there does not
exist x and N such that IXk - xl < E for all k ;::: N. 0
24 3. The Topology of the Real Numbers

A number is an accumulation point of a set if we can find points of the

set which are as close to the number as we want. We define this more
precisely.
DEFINITION 3.8. Let S be a subset of the real numbers. Then the real
number x is an accumulation point (or a limit point) of S if every neigh-
borhood of x contains an element of S other than x.
Note that x need not be in S to be an accumulation point of S. This
is demonstrated in Example [Link]. The following proposition gives us two
alternative definitions of accumulation point and is an easy consequence of
the preceding definitions.
PROPOSITION 3.9. Suppose S is a set. Then the following statements
are equivalent.
a) The point x is an accumulation point of the set S. That is, every
neighborhood of x contains at least one element of S other than x.
b) For each E > 0 there exists y in S such that 0 < Ix - yl < E.
c) There is a sequence of points, all different from x and contained in
S, which converges to x.
The reader should prove the equivalence of the three statements. Propo-
sition 3.9 allows us to use anyone of the statements to prove statements
about accumulation points.

EXAMPLE 3.10.
a) The point 0 is an accumulation point of (0,1]. If E is any positive number,
then either 1 < E, or ~ is an element of (0, 1] and 0 < I~I < E. Hence, there
is a point in (0,1] whose distance from 0 is positive and less than Eo So 0
is an accumulation point of (0,1) by part (b) of Proposition 3.9.
b) The irrational number x = .123456789101112131415 ... is an accumu-
lation point of the set of rational numbers.
To prove this we define a sequence of rational numbers by
Xl =.1
X2 = .12
X3 = .123
X4 = .1234

It is easy to prove that this sequence converges to x. Thus, by part (c) of

Proposition 3.9, x is an accumulation point of the rational numbers. 0
 3. The Topology of the Real Numbers 25

DEFINITION 3.11. A set is closed if it contains all of its accumulation

points.

EXAMPLE 3.12.
a) The interval 1= [2,4] is closed.
We prove this by showing that if x is not in I, then x is not an accumu-
lation point of I. This implies that if x is an accumulation point of I, then
x is in I.
Assume x is not in I. Now either x < 2 or x > 4. If x < 2, then
2 - x is a positive number and the neighborhood of x with radius 2 - x is
(x - (2 - x), x + (2 - x)) or (2x - 2,2). Since this neighborhood doesn't
contain any point of [2,4], x is not an accumulation point of [2,4]. A similar
argument will show that if y > 4, then y is not an accumulation point of
[2,4]. Consequently, [2,4] must contain its accumulation points.
b) The set of rational numbers is not closed in the real numbers since
we demonstrated a sequence of rational numbers which converged to an
irrational number in Example 3.10b. 0

We have seen that the rational numbers are neither open nor closed in
the real numbers so it is not true that all sets must be either open or closed.
In the exercises we ask the reader to show that the interval [1,2) is neither
open nor closed.
If A is a set of real numbers, then the complement of A is the set of
all numbers which are not contained in A. For example, the complement
of [1,2) is (-00,1) U [2,00). Often we can discover properties of a set by
examining its complement. The following proposition employs this tech-
nique.

PROPOSITION 3.13. A set is open if and only if its complement is closed.

PROOF. Let A be a subset of the real numbers and let B be the com-
plement of A. We begin by showing that if A is open, then B is closed.
Suppose A is open. If x is an accumulation point of B, then every
neighborhood of x contains an element of B. Consequently, there is no
neighborhood of x which is completely contained in A. As A is open,
Proposition 3.3 implies x is not in A. It follows that x is in B and so
B contains all of its accumulation points. Hence, if A is open, then B is
closed.
It remains to show that B is open when A is closed.
Suppose A is closed. Let x be a real number such that for all E > 0 there
exists y in A satisfying Ix - yl < E. Then x must be in A and it follows
that B is open. Why? 0
26 3. The Topology of the Real Numbers

A number of observations follow from Proposition 3.13. For example we

now know the set of irrational numbers is neither open nor closed. This is
a consequence of Proposition 3.13 and our earlier observation that the set
of rational numbers is neither open nor closed. Additional implications of
Proposition 3.13 appear in the exercises and later in the text.
We occasionally find it useful to speak of a collection of sets such as
{AI, A 2, A 3, ... }, which we will denote as {An}. We will denote the union
of this collection, Al U A2 U A3 U ... , as U An and the intersection, Al n
n
A2 n A3 n ... , as An. The following proposition outlines a few of the
properties of collections of open and closed sets.

PROPOSITION 3.14. Let {An} be a collection of open sets. ThenUA n is

n
open. If the collection contains only a finite number of sets, then An is
open.
Let {Bn} be a collection of closed sets. Then n
Bn is closed. If the
collection contains only a finite number of sets, then UBn is closed.

It is important to note that the intersection of an infinite number of

open sets might not be open. Likewise, the union of an infinite number
of closed sets might not be closed. For example, if we define the infinite
collection of open sets {An} by An = (- 1., 1.),
then nAn = {O}, which is
not open! Similarly, if we define the infinite collection of closed {Bn} by
1.,
Bn = [-1 + 1- 1.l, U
then the union Bn is equal to (-1,1), an open set.
This illustrates that the distinction between infinite and finite collections
is not made lightly. A proof of Proposition 3.14 can be worked out by the
reader or found in W. Rudin's text, Principles of Mathematical Analysis.
The last topological concept we introduce in this chapter is the concept
of a dense subset.

DEFINITION 3.15. Let A be a subset of B. A is dense in B if every point

in B is an accumulation point of A, a point of A, or both.

We can think of a dense subset of B as having parts in every nook and

cranny of B. Another way of defining a dense subset is to say A is dense
in B if, for all points in B, and for all circles around that point, no matter
how small the radius, there is a point of A contained in the circle. This
idea is made precise in the following proposition.

PROPOSITION 3.16. Let A be a subset of B. Then, the following state-

ments are equivalent.
a) A is dense in B.
b) For each point x in B and each E > 0, there exists y in A such that
Ix - yl < E.
 Exercise Set 3 27

c) For every point x in B, there exists a sequence of points contained

in A that converges to x.

The proof of this proposition is not hard and is left as an exercise.

EXAMPLE 3.17.
The rationals are dense in ~.
It suffices to show that every irrational number is an accumulation point
of the rationals.
Let t = * * .tlt2t3t4 ... be an irrational number expressed in decimal
notation. Define a sequence of rational numbers by

Xl = * * .tl
X2 = * * .tlt2
X3 = * * .tlt2t3
X4 = * * .tlt2t3t4

We can easily prove that this sequence converges to t. Hence, the rationals
are dense in the real numbers. 0
It is interesting to note that the irrational numbers are also dense in the
real numbers. This demonstrates that even though a set is large enough to
be dense, it doesn't necessarily have to be so large that there isn't enough
room left over for other large (that is, dense) sets. In fact, we can find
an infinite collection of subsets of the real numbers, each of which is dense
in the real numbers, such that no two of the sets have points in common.
Next time you are waiting patiently in a doctor's office or for a lecture to
end you might try to do just that.

Exercise Set 3

3.1 • Show that Ix - yl < f if and only if y is in the open interval

(x - f,X + f). Thus, the f neighborhood of x or N€(x) is the
interval (x - f, X + f).

3.2 Prove that the interval (a, b) is open. (Note that most of the proof
is contained in Example 3.5.)

3.3 Prove that the interval [a, b] is closed.

28 3. The Topology ofthe Real Numbers

3.4 Show that the interval [1,2) is neither open nor closed.

3.5 Prove that the rationals are not an open set.

3.6 Indicate whether the following sets are open, closed, neither open
nor closed, or both open and closed. Justify your answer.
a) [-1,0) U (O,lj
b) (-l,OjU[O,l)
c) (-l,O)n[O,l)
d) The set of natural numbers, {I, 2, 3, ... }.

e) U(-l,~)
n=l n n

3.7 Prove Propositions 3.3, 3.4, and 3.9.

3.8 Prove that the sequence x, x, x, ... converges to x.

3.9 Let x = .123456789101112 ... and define a sequence by

Xl =.1
X2 = .12
X3 = .123
X4 = .1234

Show that the sequence converges to x.

3.10 Explain the second half of the proof of Proposition 3.13.

3.11 Prove that lR is both open and closed. Use Proposition 3.13 to
show the empty set 0 is both open and closed. Are there any other
subsets of lR which are both open and closed?

3.12 Show that the set g, 1,~, i, ... ,O} is closed.

3.13 Let A be a closed and dense subset of B. Prove that A = B.
 Exercise Set 3 29

3.14 Show that the set of irrational numbers is dense in JR.

3.15 * Prove Proposition 3.14.

3.16 Prove Proposition 3.16.
4
Periodic Points
and Stable Sets

We are now ready to examine the dynamics of discrete systems. We begin

by defining and categorizing the simplest types of behavior. In most of
what follows we assume that the range of the function in question is a
subset of the domain. Exceptions to this practiCe will be noted.

DEFINITION 4.1. If f is a function and f(c) = c, then c is a fixed point

of f·
A function of the real numbers has a fixed point at c if and only if the
point (c, c) is on its graph. Thus, a function has a fixed point at c if and
only if its graph intersects the line y = x at the point (c, c).
Fixed points are important in dynamics. In Chapter 1 we observed that
the function f(x) = -x 3 has a fixed point at 0 which attracts all of the
points in the interval (-1,1) to it under iteration of the function. We will
see that in many dynamical systems fixed points playa similar role.
The following theorem will help us to locate fixed points:

THEOREM 4.2. Let I = [a, b] be a closed interval and f : I -+ I be a

continuous function. Then f has a fixed point in I.
PROOF. If f(a) = a or f(b) = b, then either a or b is fixed and we
are done. Suppose f(a) =F a and f(b) =F b. Let g(x) = f(x) - x. Since
g(x) is the difference of continuous functions, it is a continuous function.
As f(a) =F a and f(a) is in [a, b], f(a) > a. Likewise, f(b) < b. Hence,
32 4. Periodic Points and Stable Sets

FIGURE 4.1. An illustration of the situation in Theorem

4.2. Recall that f has a fixed point at c if and only if the
graph of f intersects the line y = x at c. We see that such
an intersection is required by the conditions on f.

°
g(a) = f(a) - a > and g(b) = f(b) - b < 0. Since g is continuous, the
Intermediate Value Theorem implies there is c in [a, b] such that g(c) = 0.
°
But g(c) = f(c) - c = so f(c) = c and we are done. Figure 4.1 illustrates
this proof. 0

EXAMPLE 4.3.
The function f(x) = 1 - x 2 has a fixed point in the interval [0,1].
To see this we might note that f is a continuous function and the range
of f over [0,1] is contained in [0,1]. Thus, by theorem 4.2, f has a fixed
point in [0,1].
Alternatively, we look at the graph of f shown in Figure 4.2 and note
that it crosses the line y = x in the interval [0,1]. Hence, f has a fixed
point in [0,1].
Finally, we can find the fixed point algebraically by solving the equation
f(x) = x or 1 - x 2 = X . By doing this we find that the fixed points of f
are -~ ± 1 and note that -~ + 1
is in [0,1]. 0

In Theorem 4.2, we proved that if the image of a closed interval under a

continuous map is contained in the interval, then the map has a fixed point
 4. Periodic Points and Stable Sets 33

FIGURE 4.2. The graph of f(x) = 1- x 2 •

in the interval. In the following theorem we show that there is a fixed point
in the interval when the containment is reversed.

THEOREM 4.4. Let I be a closed interval and f : I -+ lR be a continuous

function. If f (I) :J I, then f has a fixed point in I.

PROOF. Let I = [a, b]. Since f(I) :J I there are c and d in I such that
fCc) = a and fed) = b. If c = a or d = b we are done. If not, then a < c < b
and a < d < b. If we define g(x) = f(x)-x, then g(c) = f(c) -c = a-c < 0
and g(d) = fed) - d = b - d > O. Since g(c) < 0 and g(d) > 0, and g is
continuous, the Intermediate Value Theorem implies there is e between c
and d (and hence in I) satisfying gee) = 0 and fee) = e. The reader is
encouraged to draw a figure which illustrates this proof. 0

We noted earlier that the function f(x) = _x 3 examined in Chapter

1 has a fixed point at O. The behavior of the points 1 and -1 under
iteration of f is also worth noting. Recall, f maps 1 to -1 and -1 back
to 1. Consequently, we call 1 and -1 periodic points and the set {-I, I} a
periodic orbit.

DEFINITION 4.5. The point x is a periodic point of f with period k if

fk(x) = x. In other words, x is a periodic point of f with period k if x
is a fixed point of fk. The point x has prime period ko if fko(x) = x and
r (x) =F x whenever 0 < n < k o . That is, x has prime period ko if x returns
back to its starting place for the first time after exactly ko iterations of f.
34 4. Periodic Points and Stable Sets

The set of all iterates of a point x is called the orbit of x and if x is

a periodic point, then it and its iterates are called a periodic orbit or a
periodic cycle.
To illustrate Definition 4.5 we consider f(x) = _x 3 . The points 1 and
-1 form a periodic orbit with prime period 2. The set of points with period
2 is {-I, 1, O} while the set of points with prime period 2 is {-I, I}. The
only fixed point of f is O.
A function may have lots of fixed or periodic points. f(x) = x fixes every
point while every point except 0 is a periodic point with prime period 2
for g(x) = -x. Later we shall see that, for some values of r, the function
h(x) = rx(l - x) has periodic points with all prime periods.
Two other types of points are eventually fixed and eventually periodic
points. For example, the function h(x) = 4x(1 - x) has a fixed point at
O. The point 1 is not fixed, but h(l) = 0 so after one iteration 1 is fixed
at O. Also, h(!) = 1 and h2 (!) = 0 so after two iterations ! is fixed at O.
In the exercises we ask the reader to find two points which are fixed at 0
after exactly three iterations of h. Given enough time it is possible to find
points which are fixed by h after any number of iterations.
Before we look at an example of a point which is eventually periodic but
not eventually fixed we introduce a precise definition of these terms.
DEFINITION 4.6. The point x is an eventually fixed point of the function
f, if there exists N such that r+ 1 (x) = r (x) whenever n ;::: N. The
point x is eventually periodic with period k, if there exists N such that
r+k(x) = r(x) whenever n;::: N.

EXAMPLE 4.7.
Let g(x) = Ix -11- Then 0 and 1 form a periodic cycle. f(2) = 1 so 2 is
eventually periodic. In fact, every integer is eventually periodic. 0
The function f(x) = -x3 has a fixed point at 0 and a periodic cycle
consisting of 1 and -1. Examination reveals that there are no eventually
periodic points which are not periodic. How should we characterize the rest
of the points? For example, if we start at ! and iterate with f, then we get
the sequence, "2'
1 - 23'
1 2§"'
1 - 227'
1 . . .. S0"21·IS not perlO
. d'lC and never reach es
0, though it does get closer and closer to it. That is, fn (!) converges to 0
as n goes to 00. We say! is forward asymptotic to O.
DEFINITION 4.8. Let f be a function and p be a periodic point of f with
prime period k. Then x is forward asymptotic to p if the sequence x, fk (x),
Pk(x), Pk(X), ... converges to p. In other words, limn_co rk(x) = p.
The stable set of p, denoted by WS(p), consists of all points which are
forward asymptotic to p.
 4. Periodic Points and Stable Sets 35

If the sequence lxi, If(x)l, IP(x)l, IP(x)l, ... grows without bound, then
x is forward asymptotic to 00. The stable set of 00, denoted by WS(oo),
consists of all points which are forward asymptotic to 00.
Notice that when we are searching for points in the stable set of a point
with prime period k we must consider the sequence
x, fk(x), f2k(x), f3k(x), ...
not the sequence

EXAMPLE 4.9.
a) Let f(x) = -x3. Then the stable set of 0 consists of all points in the
interval (-1, 1), the stable set of 00 consists of all points in the intervals
(-00, -1) and (1,00), the stable set of -1 contains only the point -1
and the stable set of 1 contains only the point 1. Symbolically we write
WS(O) = (-1,1), WS(oo) = (-00,-1) U (1,00), WS(-l) = {-I}, and
W S(l) = {I}. Notice that it doesn't make sense to discuss the stable set
of any other points since f has no other periodic points.
b) In Example 4.7 we examined the function g(x) = Ix -11- Returning to
that example we see WS(O) is the set of all even integers and W S (l)is the
set of all odd integers. 0
In the preceding example we tacitly assumed that if a point is in the
stable set of one periodic point, then it cannot be in the stable set of a
different periodic point. This idea is important and we prove it in the
following theorem.
THEOREM 4.10. The stable sets of distinct periodic points do not in-
tersect. In other words, if p and q are periodic points and p =I- q, then
WS(p) n WS(q) = 0.
PROOF. Let PI and P2 be periodic points of period kI and k 2 , respec-
tively. We'll show that if WS(PI) n WS(P2) =I- 0, then PI = P2.
Let x be in WS(PI) n W S (P2). Then for each E > 0 there exist NI and
N2 such that n ::::: NI implies IPI - rk1(x)1 < ~ and n ::::: N2 implies
Ip2 - rk2(x)1 < ~. If M is the larger of NI and N 2 , then n ::::: M implies
both IPI - r k1 (x) I < ~ and Ip2 - r k2 (x) I < ~. By the Triangle Inequality
we have
IPI - P21 = IPI - rklk2(x) + rklk2(x) - P21
~ IPI - rklk2(x)1 + Irklk2(x) - P21
E E
< 2" + 2" = E.
36 4. Periodic Points and Stable Sets

Since we have shown the distance between PI and P2 is less than E for each
E > 0, it must be that PI = P2. D

Now that we have defined periodic points, eventually periodic points,

and stable sets, we are able to classify most points in a simple dynamical
system. In our first example, f(x) = -x3, we have three types of points:

°
the fixed point 0, the periodic points 1 and -1, and points in the stable
sets of or 00. As we continue our investigations we shall find that not all
systems are quite so easy to characterize. We will need new tools and ideas
to classify the dynamics of even a simple function like h(x) = 4x(1 - x).
One of the simplest tools we have at our disposal is graphical analysis.

4.1. Graphical Analysis

As the name implies, graphical analysis uses the graph of a function to
analyze its dynamics. It is best understood by studying a few examples.

EXAMPLE 4.11.
We begin our examination of the dynamics of the function f(x) = x 3
by graphing f and the line y = x on the same set of coordinate axes as in
Figure 4.3, and try to determine the itinerary of the point a which lies in
the interval (0,1). Rather than follow the movement of the point on the
x-axis we keep track of its whereabouts on the line y = x. It starts at a
(or, if you prefer, at the point (a, a)). From that point we travel vertically,
until we strike the graph of f. Since we moved vertically, the x value of
the point of intersection must be a and the y value must be f(a). We
now travel horizontally back to the line y = x to arrive at the point f(a).
Repeating the process we travel vertically to the graph of f to arrive at
P
°
the point (f (a), (a)) and then horizontally back to f2 (a) on the y = x
line. Continuing, we see that r(a) approaches as n approaches 00. After

°
trying more points in the interval (0,1), we conclude that every point in
(0,1) will approach under iteration of f(x) = x 3 . Further examination
suggests that W S (0) = (-1,1) and W S (00) = (-00, -1) U (1,00). The
points -1, 0, and 1 are fixed points of f. D

EXAMPLE 4.12.
We now consider the function g( x) = _xl/3. The graph of 9 and the line
y = x are shown in Figure 4.4. We start with a point a which is near and °
°
in the interval (0,1). Note that as n grows the value of gn(a) approaches 1
in absolute value and oscillates from one side of to the other. Note also
 4.1. Graphical Analysis 37

FIGURE 4.3. Graphical analysis of the orbit of the point

a under iteration of f (x) = x 3 .

(/2(a),f3(a»

~=====:;t:====~:JL!(/t4(a),f5ca»

FIGURE 4.4. Graphical analysis of the orbit of a under

iteration of g(x) = _x l / 3 .
38 4. Periodic Points and Stable Sets

that -1 and 1 form a periodic cycle with prime period 2. Further, gn (a)
is always positive when n is even. Thus, we conclude that a is in W S (1).
Further analysis suggests that WS(1) = (0,00), WS(-1) = (-00,0), is
fixed, and WS(oo) = 0. D
°
The reader is asked to develop additional examples of graphical analysis
in the following exercises.

Exercise Set 4

4.1 Let g( x) = Ix -11. Find all fixed points and eventually fixed points
of g. How many periodic points does g have? What are they? How
many eventually periodic points does g have? What are they?

4.2 From exercise 4.1 we know that ~ is fixed by g(x) = Ix - 11. Find
WS(~).

4.3 Find two points in [0, 1] that are eventually fixed at °after exactly
3 iterations of h(x) = 4x(1 - x).

4.4 a) Can a function have a point which is eventually fixed after ex-
actly 2 iterations, if it has no points which are eventually fixed
after exactly 1 iteration? If so, give an example; if not, explain
why.
b) Suppose n is a natural number. Can a function have a point
which is eventually fixed after exactly n + 1 iterations, if it has no
points which are eventually fixed after exactly n iterations? If so,
give an example; if not, explain why.

4.5 In the text we stated that f(x) = -x 3 has no eventually fixed

points (page 34). Explain why this is true.

4.6 For each of the following functions, find all periodic points and
describe the stable sets of each. You may wish to use graphical
analysis.
a) q(x) = x2
b) E (x) = eX - 1
c) p( x) = x 2 - .r
d) r(x) = -~ arctan x
 Exercise Set 4 39

e)q(x)=x2+i
* f) h(x) = 3.5x(1 - x)

4.7 Show that all periodic points of h(x) = 3.2x(1-x) lie in the interval
[0,1].

4.8 Find a function 1 : [0,1] -+ [0,1] with no fixed points and draw its
graph.

4.9 Find a function 1 : [0,1] -+ [0,1] with exactly three fixed points
and draw its graph.

4.10 Find a continuous function 1 : (1,3) -+ (1,3) that doesn't have a

fixed point. Explain why this doesn't contradict Theorem 4.2.

4.11 Draw a figure that illustrates the proof of Theorem 4.4. Use your
figure to explain the proof.

4.12 A function 1 has a fixed point at c if (c, c) is on the graph of 1.

Suppose (a, b) is on the graph of 1 and a is a period 2 point of f.
What other point must be on the graph of I? Can you generalize
this to a period n orbit?

4.13 Find a continuous function of the real numbers which has a periodic
point of period three and indicate the periodic point.
* What other periodic points does the function have?
Hint: Choose the periodic orbit first. Use the fact that (a, b)
is on the graph of 1 if and only if 1(a) = b and find a continuous
function whose graph goes through the appropriate points.
5
Sarkovskii's Theorem

Consider the function p( x) = -,} x 2 + ~ X + 1. It is easy to see that p( 0) = 1,

p(l) = 2, and p(2) = O. So {O, 1, 2} form an orbit with period three. It is
reasonable to ask how many other periodic points p(x) has and what prime
periods are represented. These questions are answered, at least in part, by
the following remarkable theorem:

THEOREM 5.1. If a continuous function of the real numbers has a peri-

odic point with prime period three, then it has periodic points of all prime
periods.

PROOF. Let {a, b, c} be a period three orbit of the continuous function

f. Without loss of generality, we assume a < b < c. There are two cases,
f(a) = b or f(a) = c. We suppose f(a) = b. This implies f(b) = c and
f(c) = a. The proof of the case f(a) = c is similar.
Let 10 = [a, b] and let II = [b, c]. The Intermediate Value Theorem
implies that f(Io) ::J II, f(Id ::J II, and f(h) ::J 10 , Since, f(h) ::J h,
Theorem 4.4 implies that f has a fixed point in II.
Now let n be a natural number larger than 1. Suppose that there is a
nested sequence of closed intervals, h = Ao ::J Al ::J A2 ::J ••• ::J An, with
the following properties.
(1) Ao = h
(2) f(Ak) = Ak-1 for k = 1,2, ... ,n - 2
(3) fk(Ak) = h for k = 1,2, ... ,n - 2
42 5. Sarkovskii's Theorem

(4) r-l(An _ l ) = 10
(5) r(An) = h
Since An C h, (5) and Theorem 4.4 imply that r has a fixed point in
An. Of course, this is identical to saying that f has a periodic point p
with period n in An. In the next two paragraphs we use the other four
conditions to show that p has prime period n.
If p is a periodic point of period n in An, then (1) and (3) imply that
p, !(p), P(p), ... , r - 2(p) are in II = [b,cl and (4) implies r-l(p) is in
10 = [a, bl. If p = c, then f(P) = a which is in not in h. Since the only
one of the first n iterates of p which doesn't have to be in h is fn-l(P),
it must be that n = 2. However, this contradicts the fact that the prime
period of c is three. Hence, p must be in the half-open interval [b, c).
If p = b, then n = 3 since p(P) = a which is not in II and the only
one of the first n iterates of p which is not in h is r-l(p). Assume that
n is not three. It follows that p must be in the open interval (b, c). Since
r-l(p) is in 10 = [a, bl, r-l(p) is not equal to p and so p can't have prime
period n - 1. If the prime period of p were less than n - 1, then (3) and
the fact that p is not b or c imply that the orbit of p is contained entirely
in (b, c) and this would contradict (4). Since r- l (p) is not an element of

(b, c), it must be that p has prime period n.

To complete the proof we demonstrate the existence of such a sequence
of closed sets for each natural number larger than 1. Let n be a natural
number which is larger than 1. We establish each of the requirements in
turn. Obviously we can choose Ao so that Ao = hand (1) is satisfied. The
other properties use the fact that if f is a continuous function and J and
K are closed intervals such that f(J) ::> K, then there exists an interval Jo
such that Jo C J and f(Jo) = K. This is an intuitively clear result whose
proof depends on the continuity of f. The proof is outlined in exercise 5.7
at the end of this chapter.
Now, since Ao = hand f (Id ::> h, we have f (Ao) ::> Ao and so there
is Al contained in Ao such that !(A l ) = Ao. Since Al C Ao this implies
!(A 1 ) ::> AI. Consequently, there is A2 C AI, such that !(A 2) = AI. We
continue in this manner and define Ak for k = 1,2, ... ,n - 2. In each case,
we find Ak contained in Ak-l so that f(Ak) = A k- l for k = 1,2, ... , n - 2
as required by (2). Further, f(Ak) ::> Ak for each k so the process of
defining the intervals Ak can continue indefinitely. Then f(Ak) = Ak-l
implies

P(A k ) = f(f(Ak)) = f(A k- l ) = A k- 2

f3(Ak) = f(f2(Ak)) = f(A k- 2) = A k- 3
 5. Sarkovskii's Theorem 43

fk(Ak) = Ao = II
for each k = 1,2, ... ,n - 2 as required by (3). To define A n - 1 consistent
with (4) we note that
r - 1(A n_ 2 ) = f(r- 2 (A n _ 2 )) = f(11) ::) 10 .
Hence, there is A n - 1 C A n- 2 such that f n - 1(A n_ 1) = 10 , Finally,
r(An-d = f(r- 1(A n _ 1)) = f(Io) ::) II
implies there is An C A n- 1 such that r(An) = II as required by (5) and
the proof is complete. 0
As surprising as Theorem 5.1 might be, it is only a special case of a more
general theorem proven by A. N. Sarkovskii in 1964. This interesting and
beautiful result depends only on the continuity of the function in question
and should give pause to anyone who might think that all of the really
good theorems which don't rely on the mathematical equivalent of a B-1
bomber were proven before the turn of the century.
In his theorem, Sarkovskii places an order on the integers which we
define below. The statement of the theorem follows the definition.
DEFINITION 5.2. SARKOYSKII'S ORDERING. Sarkovskii's ordering of
the natural numbers is
3 >- 5 >- 7 >- ... >- 2 . 3 >- 2 . 5 >- 2 . 7 >- ...
. . . >- 22 . 3 >- 22 . 5 >->- 22 . 7 >- . . . >- 2n . 3 >- 2n . 5 >- 2n . 7 >- ...
. . . >- 23 >- 22 >- 2 >- 1.

a >- b indicates a precedes b in the order. When writing the order, all odd
numbers except one are listed in ascending order, then two times every odd,
then four times every odd, and so on. The order is completed by listing the
powers of two in descending order. Every natural number can be found
exactly once in Sarkovskii's ordering.
THEOREM 5.3. [SARKOYSKII, 1964] Suppose that f : lR --+ lR is contin-
uous and that f has a periodic point with prime period n. If n >- m in
Sarkovskii's ordering, then f also has a periodic point with prime period
m.
While the proof of Sarkovskii's Theorem uses no tools which are not
used in the proof of Theorem 5.1, it is somewhat longer. However, in
spirit it is much the same. Essentially we show that if n >- m, then the
existence of a prime period n point implies the existence of a sequence of
closed intervals h, 12 , •.• ,1m , such that f(1i) ::) IH1 for all i < m and
f(1m) ::) h· Then Theorem 4.4 implies there is a period m point in II.
44 5. Sarkovskii's Theorem

FIGURE 5.1. The graphs of h(x) = 3.2x(1-x), h2 (x), and

h4 (x).

If one of the intervals is disjoint from all of the other intervals, then there
is a point with prime period m. The key then is to find these intervals h
Rather than write the proof out in detail here, we refer the reader to the
proof by Huang on pages 91-102 of the April 1992 issue of Mathematics
Magazine. Huang's exposition is clear and complete. A similar proof can be
found in Devaney's text, Introduction to Chaotic Dynamics, which is listed
in the references. Both authors demonstrate that Sarkovskii's Theorem
is sharp by constructing continuous functions which have period 5 points,
but no period 3 points. The technique used allows for the construction of
a continuous function which has a periodic point with prime period n, and
no periodic point with prime period which precedes n in the Sarkovskii
ordering.
In addition to implying the existence of orbits with prime periods lower
than that of a known orbit, Sarkovskii's Theorem can be used to show that
points with certain periods do not exist. For example, in Figure 5.1 the
graphs of h(x) = 3.2x(1 - x), h2 (x) and h4 (x) are shown.
In the first graph we see that h has fixed points at 0 and near .7. In the
second graph, we see the fixed points again, and a period 2 orbit near .5
and .8. In the third graph, we see that the only period 4 points are at 0
and near .7, .5, and .8. That is, h has no prime period 4 orbits. Therefore,
we can conclude that h has no orbits with prime period other than 2 and
1, since those are the only two orbits less than 4 in the Sarkovskii ordering.
We should note that it is safe to assume that all of the periodic points of
h lie in the interval [0, 1], so we do not need to graph any other portion of
the function. The reader should verify that this assumption is valid.
 Exercise Set 5 45

Exercise Set 5

5.1 Let hand 10 be as described in the proof of Theorem 5.1. Use the
Intermediate Value Theorem to show that lUo) ~ h, IUd ~ h,
and IUd ~ 10 .

5.2 Demonstrate the existence of an orbit of h(x) = 4x(1 - x) with

prime period three. What does Sarkovskii's Theorem tell us about
the periodic points of h(x)? •
Hint: Use a computer graphics package to look at the graphs
of hand h 3 .

5.3 Show that l(x) = """"-rr4 arctan(x) does not have a point with prime
period 32.
Hint: Use a computer graphics package to look at the graphs
of the iterates of f.

5.4 Suppose that A o, AI, A 2, ... , An are closed intervals and 1 is a

continuous function such that I(Ak) ~ Ak+! for 0 ::::; k < n.
a) Show that there is a point Xo in Ao such that li(xo) is in Ai for
[Link]; i::::; n.
b) Prove that if 1(An) ~ A o, then 1 has a periodic point in Ao
with period n.
c) Prove that if Ao is disjoint from the other intervals, then 1 has
a point with prime period n in Ao.

5.5 Let 1 : lR --t lR be continuous, n > 3, and Xl, X2, ... ,Xn be points
such that Xl < X2 < ... < Xn . Show that if l(xi) = Xi+! for
i = 1,2, ... , n - 1 and l(xn) = Xl, then 1 has points with all
prime periods.

5.6 a) Let 1 : 1 --t 1 be a continuous function of the interval 1 and

suppose 1 has a periodic point with prime period 3 in I. Show
that 1 has periodic points of all orders in 1.
Hint: Extend 1 to a continuous function of the real line.
b) Let I be an interval. State and prove an analog of Sarkovskii's
Theorem for 1 : I --t 1. You may assume Sarkovskii's Theorem.
46 5. Sarkovskii's Theorem

5.7 * Let J = [a, b] and [ = [c, d] be closed intervals and let f be a

continuous function satisfying f (J) ::) [. Show that there exists an
interval J o such that Jo c J and f(Jo) = [.
Hint: Let Xo be a point in [a, b] such that f(xo) = c. Without
loss of generality we can assume there is y in [xo, b] such that
f(y) = d. (You should explain why we can do this.) Let X2 be the
greatest lower bound of the set {y in [xo, b]1 f(y) = d} and Xl be
the least upper bound of the set {x in [xo, X2] I f (x) = c}. Show
that f([Xl,X2]) = [c,d].
6
Differentiability and
Its Implications

In many cases the functions we consider are differentiable. In this chapter

we examine the dynamical information contained in the derivative of the
function.
DEFINITION 6.1. Let I be an interval, f : 1-* 1R, and a be a point in I.
The function is differentiable at a if the limit
. f(x) - f(a)
11m .:........:...-=-----'-~

x-a X - a

exists. In this case we say f is differentiable at a and call the limit f' (a)
or the derivative of f at a. A function is differentiable if it is differentiable
at each point in its domain.
Recall that a function which is differentiable at a point is also continuous
at that point. An important property of differentiable functions which we
will often exploit is the Mean Value Theorem.
THEOREM 6.2. MEAN VALUE THEOREM. Let f : [a, b] -* IR be differ-
entiable on [a, b]. Then there exists c in [a, b] such that
f(b) - f(a) = f'(c)(b - a).
After examining a few graphs the reader will realize that the Mean Value
Theorem is an intuitively clear result. A proof can be found in any good
calculus text. We use the Mean Value Theorem to get another result about
48 6. Differentiability and Its Implications

fixed points. Theorem 4.2 states that every continuous function which maps
a closed interval back into itself has at least one fixed point. If the function
is differentiable on the closed interval and the value of the derivative is less
than one in absolute value, we get a much stronger result.
THEOREM 6.3. Let I be a closed interval and f : I --+ I be a differen-
tiable function satisfying 1f'(x)1 < 1 for all x in I. Then f has a unique
fixed point in I. Moreover, if x and yare any two points in I and x I- y,
then If(x) - f(y)1 < Ix - yl·
In the hypothesis of Theorem 6.3 we assume f is differentiable on I.
Consequently, f must also be continuous on I and Theorem 4.2 guarantees
that f has a fixed point. The additional differentiability condition guaran-
tees that f has only one fixed point and points in I become closer together
as we iterate the function. In fact, it can be shown that I is contained
in the stable set of the unique fixed point. If we assume f' is continuous,
then this is a corollary of the proof of Theorem 6.5. If f' is not contin-
uous, then the proof requires more sophisticated tools from mathematical
analysis than we have at our disposal. In particular, an understanding of
least upper bounds is useful. Finally, we note that, as with Theorem 4.2,
Theorem 6.3 does not necessarily hold when I is an open interval. (See
exercise 4.10.)
The proof of Theorem 6.3 follows from the Mean Value Theorem.
PROOF. We begin by proving the second part of the theorem. Let x
and y be any two points in I satisfying x I- y. Without loss of generality
we assume x < y. Then our hypotheses assert that f is differentiable on
[x, yj. By the Mean Value Theorem, there exists c in [x, yj such that

If(x) - f(y)1 = 1!,(c)llx - yl or If'(c)1 = If(x) - f(Y)I. (6.1)

Ix-yl
Since c is contained in [x, yj and [x, yj is contained in I, we know that
1f'(c)1 < 1. Hence, equation (6.1) implies If(x) - f(y)1 < Ix-yl as desired.
It remains to show that f has a unique fixed point. As we noted above,
Theorem 4.2 implies that f must have at least one fixed point in I. Let p
be a fixed point and let x be any other point in I. We have already shown
that
Ip - f(x)1 = If(p) - f(x)1 < Ip - xl·
It follows that f(x) I- x and x is not a fixed point of f. Therefore, p is the
only fixed point of f in I. 0
Notice that by assuming the function was differentiable we were able to
make a much more specific statement about the dynamics of the function.
If in addition to differentiable the derivative is continuous, we gain further
 6. Differentiability and Its Implications 49

FIGURE 6.1. Graphical analysis of f(x) Ix

2
+ ~
2
and
r(x) = -~x + ~.

information about the dynamics. To illustrate this, we consider a few

simple examples.

EXAMPLE 6.4.
a) Let f(x) = ~x +~. Solving ~x + ~ = x we find that 3 is the only fixed
point of f. Notice also that fl (x) = ~ for all x. Hence, if x is any real
number other than 3, then the Mean Value Theorem implies there exist
real numbers, Cl, C2, and C3 such that

If(x) - f(3)1 = I!,(cdllx - 31 = ~Ix - 31

If2(x) - f2(3)1 = 1!,(C2)llf(x) - f(3)1 = ~Ix - 31

and

If 3(x) - f3(3) 1= 1!,(C3) llf2(X) - f2(3)1 = 213 Ix - 31.

Generalizing we find that
Ir(x) - r(3)1 = 2~ Ix - 31·
Since r(3) = 3, we conclude that r(x) converges to 3 and x is in W S (3) .
Consequently, W S (3) = R
This is illustrated using graphical analysis in Figure 6.l.
b) Now let r(x) = -~x +~. Once again 3 is the unique fixed point, but
this time rl(x) = -~. Using the Mean Value Theorem as in part (a) , we
can show that WS(3) = R In Figure 6.1 we illustrate this using graphical
analysis. Notice the relationship between the sign of the derivative and the
50 6. Differentiability and Its Implications

behavior of x under iteration of the function. In I the derivative is positive

and x approaches 3 from one side. In r the derivative is negative and the
iterates of x alternate from one side of 3 to the other. This is a pattern we
will see again.
c) Let g(x) = 2x - 3. Again 3 is a fixed point, but the derivative is
positive and greater than one in absolute value. Before we go on stop and
consider: What implications does the derivative at the fixed point have
for the behavior of the rest of the points? How would one prove that our
hunch is correct?
Since g'(x) = 2 for all x, the Mean Value Theorem implies there exist
Cl, C2, . •• so that

Ig(x) - g(3)1 = Ig'(cdllx - 31 = 21x - 31

Ig2(x) - g2(3)1 = 19'(C2)llg(x) - g(3)1 = 221x - 31

As gn(3) = 3, the distance from gn(x) to 3 grows without bound as n goes

to 00. Consequently, x is in WS(oo) and WS(oo) = (-00,3) U (3,00).
d) That leaves us one more case, k(x) = -2x + 9. Once again 3 is a
fixed point. This time the derivative is -2, which is both negative and
greater than one in absolute value. As the reader should be able to see,
WS (00) = (-00,3) U (3, (0) and the iterates of x alternate from one side to
the other of 3 as they grow without bound. 0

Example 6.4 suggests the following theorem:

THEOREM 6.5. Let I be a differentiable function, let p be a fixed point

of I, and suppose f' is continuous. If II' (p) I < 1, then there exists an open
interval U containing p such that whenever x is in U, then fn(x) converges
to p. That is, U c WS(p). If If'(p) I > 1, then there exists an open interval
containing p such that all points in the interval which are not equal to p
must leave the interval under iteration of I.
Before we prove Theorem 6.5 we make a few observations. First, we
will use the fact that I' is continuous to prove the theorem. However,
this is only for our convenience; the theorem is still true even if f' isn't
continuous. (See the proof of Theorem 4.11 for an outline of a proof of this
fact.) It is difficult to imagine a differentiable function whose derivative is
not continuous, but such functions do exist. This is a point worth pondering
that should be addressed in any good class on Real Analysis. A function
whose derivative exists and is continuous is called 0 1 •
 6. Differentiability and Its Implications 51

Second, in the case where If'(p) I > 1, it is not generally true that once
a point leaves the open interval containing p it is gone for good. When
we examine the logistic function more carefully we will see cases where a
point's iterates return over and over again. As we demonstrated in Chapter
3, we may assume the open set containing p is a neighborhood of p.
Finally, as we noted in Example 6.4, it is true that when the derivative
is negative on the open interval U, then x and f(x) are on opposite sides
of p and when the derivative is positive on U, then x and f(x) are on the
same side of p. The reader is asked to prove this in the exercises.
To sum up, we restate Theorem 6.5.
THEOREM 6.5 (RESTATED). Let f be a 0 1 function and p be a fixed
point of f. Then If'(p) I < 1 implies that there is a neighborhood ofp which
is contained in W'(p) and If'(p)1 > 1 implies that there is a neighborhood
of p all of whose points must leave the neighborhood under iteration of f·
PROOF. Let f be differentiable with a continuous derivative, assume pis
a fixed point of f, and suppose 1f'(p)1 < 1. We wish to find a neighborhood
of p that is contained in WS (p). Our strategy is to find 8 > 0 and E > 0
such that 1f'(x)1 < 1- E for all x in (p- 8,p+ 8). We then apply the Mean
Value Theorem to show that (p - 8,p + 8) C WS(p).
Let E = ~(1 - If'(p)l). Since 1f'(p)1 < 1, E > O. To find 8 we use
the continuity of f'. Since f' is continuous there exists 8 > 0 such that
If'(x) - f'(p) I < E whenever Ip - xl < 8. (The reader should be able to
explain why this is so.) Equivalently, If'(x) - f'(p) I < E whenever x is in
(p - 8,p + 8). As E = ~(l-If'(p)I), this implies that if x is in (p - 8,p+ 8),
then

1!,(x)1 = + !,(p) I
If'(x) - f'(p)
~ If'(x) - f'(p) I + If'(p) I (by the Triangle Inequality)
< E + If'(p) I
= ~(l-If'(p)1) + 1f'(p)1
= ~ + ~1!,(p)1
= 1-(~(1-I!,(p)l))
=1-E.

So, if x is in (p - 8,p + 8), then 1f'(x)1 < 1 - E. It remains to show that

r(x) converges to p whenever x is in (p - 8,p + 8).
Let x be in (p-8,p+8) and not equal to p. By the Mean Value Theorem
there is c between x and p such that

If(x) - pi = If(x) - f(P)1 = If'(c)llx - pl·

52 6. Differentiability and Its Implications

Since c must be in the interval (p - 8, p + 8) this implies

If(x) - f(p)1 < (1 - €)Ix - pI·
Iterating the process we find that
If2(X) - pi < (1 - €)2Ix - pi,
and by induction,
Ir(x) - pi < (1- €)nlx - pl.
Now, since (1 - €) < 1, it follows that (1 - €)n converges to 0 as n goes
to 00. Thus, we can conclude that Ir(x) - pi approaches 0 and r(x)
converges to pas n goes to 00.
The proof ofthe case If'(P) I > 1 is similar and is left as an exercise. 0
Since U is an open set containing p in Theorem 6.5, p is on the "inside"
of the set. In other words, all points near p demonstrate the correct char-
acteristics. If U were a closed interval or set, then p might have been an
endpoint of the interval or on the "edge" of the set. In that case points on
one side of p may not have had the proper characteristics no matter how
close they were.
Fixed points whose derivatives are not equal to one in absolute value
are important enough to have their own name. They are called hyperbolic
fixed points. For obvious reasons fixed points whose derivatives are less
than one in absolute value are said to be attracting, and fixed points whose
derivatives are greater than one in absolute value are said to be repelling.
In Theorem 6.5 we saw that we can make some predictions about the
behavior of points in the neighborhood of a hyperbolic fixed point. In the
following examples we see that life is not quite so predictable near fixed
points whose derivatives are equal to 1 or -1.

EXAMPLE 6.6.
a) Let g( x) = x - x 3 . Then 0 is a fixed point of 9 and g' (0) = 1. Graphical
analysis easily demonstrates that (-1,1) C WS(O).
b) Let p(x) = x + x 3 . Then 0 is again a fixed point of p and p'(O) = 1.
However, in this case WS(O) = 0 and WS(oo) contains all real numbers
except O.
c) Let k(x) = eX -1. In this case 0 is a fixed point of k and k'(O) = 1, but
WS(O) = (-00,0] and WS(oo) = (0,00). That is, all points to the left of 0
approach 0 under iteration of k, but all other points grow without bound
under iteration of k.
The graphs of each of the functions in this example are shown in Fig-
ure 6.2. 0
 6. Differentiability and Its Implications 53

y=x y=x
y=x

k(x)

p(x)

FIGURE 6.2. The graphs ofg(x) =x-x 3 ,p(x) =x+x 3 ,

and k(x) = eX - 1.

Example 6.6 demonstrates that nothing definitive can be said about the
behavior of points near a fixed point whose derivative is equal to one in
absolute value.
Recall that p is a periodic point with period n of the function f if p is a
fixed point ofr. In light of this fact, the following definition of a hyperbolic
periodic point is an obvious extension of the definition of hyperbolic fixed
points.
DEFINITION 6.7. Let p be a periodic point of the differentiable function f
with prime period k. Then p is a hyperbolic periodic point if l(fk)'(p)1 =I- l.
If l(fk)'(p)1 = 1, then p is a nonhyperbolic periodic point.
Notice that if p is a periodic point of f with period k, then it is the
derivative of fk at p that is important in determining whether or not p is
hyperbolic rather than the derivative of f at p. In exercise 6.3 we show
that if the prime period of p is larger than 1, then the derivative of f at p
by itself is not sufficient to determine the behavior of points near p under
iteration of the function.
As a corollary to Theorem 6.5 we get
THEOREM 6.S. Let f be a C 1 function and p be a per-iodic point of f
with prime period k. If l(fk)'(p)1 < 1, then there exists an open interval
containing p that is contained in the stable set of p. If I(fk)'(p)1 > 1, then
there exists an open interval containing p such that all points in the inter-val
except p must leave the inter-val under itemtion of fk.
Theorem 6.S is an easy extension of Theorem 6.5 and its proof is left as
an exercise.
We have observed that whenever p is a periodic point of f with prime
period k and the derivative of fk at p is less than one in absolute value,
then points near p are attracted to p under iteration of fk. On the other
hand, if p is a periodic point of f with prime period k and the derivative
54 6. Differentiability and Its Implications

of fk at p is greater than one in absolute value, then points near p are

moved away from p under iteration of fk. Hence, we make the following
definition.

DEFINITION 6.9. Let p be a periodic point of f with prime period k. If

IUk)'(P) I < 1,then p is an attracting periodic point of f· If IUk)'(P)1 > 1,
then p is a repelling periodic point of f.

We emphasize again that it is the derivative of fk at p which is important

in determining whether or not p is attracting or repelling. The derivative
of f at p by itself provides no information about the nature of the periodic
point. (See exercises 6.2 and 6.3.)
It is true that the fixed point at 0 of the function g (x) = x - x 3 attracts
nearby points (Example 6.6a). However, g'(O) = 1 so we call 0 a weakly
attracting fixed point of g. Similarly, 0 is a fixed point of p( x) = x + x 3
which repels nearby points, but p'(x) = 1 (Example 6.6b). Consequently,
we call 0 a weakly repelling fixed point of p.
In Chapter 7 we will begin to look at how periodic points in particular,
and the dynamics of functions in general, change as we slowly change the
functions. We shall see that hyperbolic periodic points are relatively in-
sensitive to small changes while nonhyperbolic periodic points can change
quickly. The relative stability of hyperbolic periodic points accounts for
their importance.

Exercise Set 6

6.1 Find the periodic points of each of the following functions. Indicate
whether the periodic points are repelling, attracting or neither and
find their stable sets.

a) f(x) = -x 3
b) p(x) =x3 -x

c) k(x) = -x3 - x

d) g(x) = e x - 1

e) E(x) = e- X

f) S(x) = sin x
 Exercise Set 6 55

g) r(x) = _x 1 / 3

h) A(x) = ~4 arctan x

6.2 • a) Let 1 : I -+ I be a differentiable function, let x be a point in

I, and k be a natural number. Prove
(fk)'(X) = !,(x)· I'(f(x)) . 1'(f2(X)) ..... 1'(fk-l(X)).
Hint: Use the chain rule and Mathematical Induction.
b) Let {P -1,p2' ... ,Pn} be the orbit of a periodic point with pe-
riod n. Use part (a) to prove P is an attracting hyperbolic periodic
point if and only if I!'(pd· 1'(P2)· ... · !'(Pn) I < 1. State and prove
a similar statement for repelling hyperbolic periodic points.

6.3 Find a C 1 function 1 : lR. -+ lR. which has 0 as an attracting hy-

perbolic periodic point with period 2, and such that 1'(0) = 2.
Does this contradict the definition of attracting hyperbolic peri-
odic points? Why or why not?
Hint: Use exercise 6.2 and draw a graph of I.

6.4 Find a fixed point of I(x) = -!x + 3 and use the Mean Value
Theorem to prove that every real number is in the stable set of
this fixed point.

6.5 Prove the second half of Theorem 6.5 using the steps outlined be-
low. The proof depends on finding an € > 0 and an open interval
(p - 8,p + 8) so that 1!,(x)1 > 1 + € whenever x is in (p - 8,p+ 8).
!
a) Let € = (I!' (p) I - 1). Choose an appropriate 8 and use the
Triangle Inequality to show that 1/'(x)1 > 1f'(p)l- € whenever x
is in (p - 8,p + 8).
b) Prove 1f'(p)l- € = 1 + € and conclude 1/'(x)1 > 1 + € whenever
x is in (p - 8,p + 8).
c) Show that if y =1= p and Ik(y) is in (p - 8,p + 8) for all k < n,
then Ir(y) - pi> (1 + €)nly - pI.
d) Use (c) to prove the result.

6.6 * Let I be a closed interval and 1 : I -+ I be a C 1 function

satisfying 1f'(x)1 < 1 for all x in I. Theorem 6.3 states that 1 has
a unique fixed point in I. Prove th;1t I is contained in the stable
set of this fixed point.
56 6. Differentiability and Its Implications

6.7 * Let 1 be a C l function and suppose that p is a fixed point of

1 satisfying f'(p) > 1. Does Theorem 6.5 imply that there is a
neighborhood U of p such that Un WS(P) = p? Prove that your
answer is correct.

6.8 Let 1 : I - t lR be a C l function of the interval I and suppose that

p is a fixed point of I.

a) Prove that if I'(p) < 0, then there is an open interval U con-

taining p such that whenever x is in U, I(x) and x are on opposite
sides of p.
b) Prove that if f'(P) > 0, then there is an open interval U con-
taining p such that whenever x is in U, I(x) and x are on the same
side of p.
Hint: First define mathematically what we mean by "being on
the same side of p" , then use the Mean Value Theorem.

6.9 Let Xo and a be positive real numbers and consider the sequence
defined by
Xn+l = ~2 (xn+~).
Xn
a) Find a function 1 such that Xn = r(xo).
b) Describe the periodic points of I. Are they hyperbolic? attract-
ing or repelling?
c) What is the limit of the sequence xo, Xl, X2, •.• ?
Note: 1 defines a very useful algorithm for computing square
roots as it has a strongly attracting fixed point. That is, the se-
quence converges quickly to the desired square root. For a discus-
sion of the error estimate see exercise 16 in Chapter 3 of the third
edition of W. Rudin's text Principles 01 Mathematical Analysis.
7
Parametrized Families of
Functions and Bifurcations

By a family of functions we mean a collection of functions which are of

the same general type. For instance we might consider the family of linear
functions of the form fm(x) = mx where m is allowed to vary over the real
numbers. The variable m is called a parameter and the family represented
by fm(x) = mx is called a parametrized family. The parameter is usually
written as a subscript of the function name. In those cases where the
parameter is clear, we will suppress the subscript to simplify notation.
We are interested in how the dynamics of a function changes as we
change the parameter. To illustrate this, let's look at an example.

EXAMPLE 7.l.
Consider the parametrized family fm(x) = mx. For all values of m
except 1, it is clear that 0 is the only fixed point of fm.
If m < -1, then 0 is a hyperbolic repelling fixed point and all other
points are in the stable set of infinity.
If m = -1, then all points except 0 are periodic points with prime
period 2.
When -1 < m < 1, 0 is an attracting fixed point and WS(O) = R
If m = 1, then all points are fixed points.
Finally, when m > 1, 0 is a repelling fixed point and all other points are
in the stable set of 00. D
58 7. Parametrized Families of Functions and Bifurcations

We note that the dynamics of the family fm(x) = mx are unchanged for
large intervals of parameter values. Then at a particular parameter value
(Le., -1 and 1), the dynamics change suddenly, after which they again
remain constant for a prolonged interval. We call these sudden changes in
dynamics bifurcations. Notice also 0 is the fixed point and that If'(O)1 =
1 when m = 1 or -1. The presence of nonhyperbolic fixed points for
parameter values at which there is a bifurcation is typical.
DEFINITION 7.2. Let fc(x) be a parametrized family of functions. Then
there is a bifurcation at Co if there exists E > 0 such that whenever a and b
satisfy Co - E < a < Co and Co < b < Co + E, then the dynamics of fa (x) are
different from the dynamics of fb(x). In other words, the dynamics of the
function changes when the parameter value crosses through the point Co.
Readers who are used to a more rigorous approach may wonder what
we mean by saying "the dynamics are the same" or the "dynamics are
different". We will rely on an intuitive idea of sameness until we discuss
topological conjugacy in Chapter 10. Then we will say two functions have
the same dynamics if and only if they are topologically conjugate. Inter-
ested readers are encouraged to look ahead. For now, our intuition will be
able to deal well with the functions we examine.
Let's look at a few more simple examples.

EXAMPLE 7.3.
Consider the family of functions Ec(x) = eX +c. We examine the dy-
namics of this family by using graphical analysis. We are all familiar with
the general shape of the exponential function. We should notice that as
the parameter increases in value the shape of the graph remains the same,
but it moves to the left. A careful examination reveals that we have three
cases: the graphs of E(x) and y = x intersect in exactly two points, the
graphs of E(x) and y = x have a single point of tangency, and the graphs
of E(x) and y = x do not intersect. The graphs of typical examples are
shown in Figure 7.1.
If c < -1, then the graphs of E(x) and y = x intersect in two places
as shown in Figure 7.1. We call the two points of intersection a and b,
with a < b. Note that both a and b are fixed points, 0 < E'(a) < 1,
and E'(b) > 1. Hence, a is an attracting hyperbolic fixed point and b is
a repelling hyperbolic fixed point. Graphical analysis demonstrates that
WS(a) = (-00, b) and WS(oo) = (b,oo).
When c = -1, the two fixed points fuse into one fixed point at x = 1.
This fixed point is not hyperbolic since E~ 1 (1) = 1. Graphical analysis
demonstrates that WS(l) = (-00,1] and WS(oo) = (1,00). This case is
illustrated in Figure 7.1.
 7. Parametrized Families of Functions and Bifurcations 59

4
The repelling [Link] poin! b

The anracling Ii cd point 0

2 3

·1

·2

·2 4

-1

·2

FIGURE 7.1. Graphical analysis of Ec for the three cases,

c < -1, c = -1, and c > -1.
60 7. Parametrized Families of Functions and Bifurcations

a) b)

0.5

-\ -\

-\ -\

FIGURE 7.2. The graph of Ak(X) = karctan(x) along with

the line y = x for the parameters values a) k = .8, b) k = 1,
and c) k = 1.2. Note that A~(O) = k when k = 1, so the
fixed point is not hyperbolic.

Finally, if c > -1, then the graphs of E(x) and y = x do not in-
tersect and consequently, E(x) has no fixed point. Since E(x) is con-
tinuous, Sarkovskii's Theorem implies that E(x) can have no periodic
points when c> -1. Using graphical analysis we show in Figure 7_1 that
WS(oo) =~. 0

We make a special note of a few characteristics of this example. First,

as the parameter c grows and approaches -1, the two fixed points a and b
gradually approach one another until, when c = -1, they join to become
one fixed point. Immediately thereafter, they disappear all together. This
type of bifurcation is called a saddle-node bifurcation. The merging and
annihilation (or creation and splitting) of periodic points as we vary a
parameter is a common theme in dynamics. Further, notice that at the
moment of bifurcation, the fixed point is not hyperbolic. This is another
common characteristic of bifurcations.

EXAMPLE 7.4.
Consider the family of functions Ak(X) = karctan(x) for parameter
values near 1. When 0 < k < 1, the function has one attracting fixed point
near 0 and all real numbers are in the stable set of 0_ If k = 1, then 0
is still a fixed point but Ai (0) = 1. Graphical analysis indicates that 0
is weakly attracting and WS(O)is still ~ even though 0 is not hyperbolic.
Finally, when k > 1, we find a repelling fixed point at 0 (A~(O) = k > 1)
and two other fixed points have been formed, both of which are attracting.
If the two fixed points are labeled a and b and a < 0 < b, then we see that
WS(a) = (-00,0) and WS(b) = (0,00)_ The graph of each case is shown in
Figure 7.2 along with the line y = x. 0
 7. Parametrized Families of Functions and Bifurcations 61

The type of bifurcation seen in Ak(X) = karctan(x) when k = 1 is

called a pitchfork bifurcation. In general, pitchfork bifurcations occur when
either an attracting periodic point splits into a repelling periodic point
with an attracting periodic point of the same period as the original point
on each side of it, or alternatively, a repelling periodic point splits into an
attracting periodic point surrounded by two repelling periodic points of the
same period. The rationale for this odd name becomes apparent when we
consider the bifurcation diagram.
A bifurcation diagram is a graph of the periodic points of a function
plotted as a function of the parameter. Typically, the periodic values are
plotted on the vertical axis and the parameter values are plotted on the
horizontal axis. We denote fixed points by solid lines and non-fixed periodic
points by dotted lines. Vertical arrows are added to indicate whether or not
a point is attracting or repelling. Bifurcation diagrams for Ec(x) = e X +c
and Ak(X) = karctan(x) are shown in Figure 7.3. A quick look at the
bifurcation diagram for Ak attests to the origin of the name "pitchfork
bifurcation" .

EXAMPLE 7.5.
The function hr (x) = rx( 1- x) exhibits a third type of bifurcation when
r = 1: a trans critical bifurcation. When 0 < r < 1, h has two fixed points,
one of which is less than one and repelling, and another at 0, which is
attracting. When r = 1, these two points merge .to form one fixed point at
o which attracts points which are greater than 0 and repels points which
are less than O. Finally, if 1 < r < 3, we see that 0 becomes a repelling
fixed point and there is an attracting fixed point which is greater than O.
The graphs of these three situations are shown in Figure 7.4. 0

EXAMPLE 7.6.
Continuing the investigation of hr(x) = rx(l - x) that we started in
Example 7.5, we find that hr undergoes another bifurcation when r = 3.
We have seen that hr has a repelling fixed point at 0 and an attracting
fixed point greater than 0 when 1 < r < 3. A quick look at the graph
of h; demonstrates that hr has no other periodic points when r is in this
range. At r = 3 the larger fixed point is at .75 and is weakly attracting:
it is no longer hyperbolic since h' (.75) = -1. When r > 3 the fixed point
is repelling. (The proof of this is left as an exercise.) Graphical analysis
when r = 3.1 indicates that most points in (0,1) are attracted to a period
2 orbit which straddles the larger fixed point. What has happened?
To understand this, we look at the graph of h;
in the vicinity of the
fixed point Pr = r-;l for parameter values slightly less than 3, equal to 3,
62 7. Parametrized Families of FUnctions and Bifurcations

j 1+
a) b)

rr
2
4

t I
I
2 0

ttl
0
-1

-2
1i~ 1t
-3 -2.5 -2 -1.5 -1 -0.5 0 0.5 1.5 2

FIGURE 7.3_ The bifurcation diagrams of a) Ec(x) = eX +c

and b) Ak (x) = k arctan( x). Note that the value of the
parameter is shown on the x-axis and the domain of the
function is represented on the y-axis. Solid lines represent
fixed points. From graph (a) we see that when c = -2,
Ec(x) has fixed points at about .16 and about 3_2. The
arrows indicate the direction points move under iteration.
Consequently, when c = -2 the arrows indicate the fixed
point at .16 is attracting and the fixed point at 3.2 is re-
pelling. Recall that bifurcations whose diagrams are sim-
ilar to (a) are called saddle node bifurcations and those
whose diagrams are similar to (b) are called pitchfork bi-
furcations. The inspiration for the latter name should be
apparent from the diagram.
 7. Parametrized Families of Functions and Bifurcations 63

a) b) c)

-OJ 0.3

FIGURE 7.4. The graph of hr(x) = rx(1 - x) for three

parameter values near 1. The graph of the line y = x is
also shown for comparison. In (a) r = .78, in (b) r = 1,
and in (c) r = 1.3. Note that h~ (0) = 1 when r = I, so
the fixed point is not hyperbolic when the bifurcation oc-
curs.

~ t t~~{

i i i'f)
0.5

O r-----------~~----------------------------------

·0.5

.1~-----L------~------------~2------------~-------

FIGURE 7.5. The bifurcation diagram for the logistic func-

tion hr(x) = rx(1 - x). The dashed lined represents a
period 2 attracting orbit.
64 7. Parametrized Families of Functions and Bifurcations

/
a) b) /

/
0.15 0.15

0.1 0.1

'''V
0.65

0.6 0.6

0.55 / 0.55
0.55 0.6 0.65 0.1 0.15 •
0.55 0.6 0.65 0.1 0.15

c)
0.15 Q2

\
0.1

0.6S

-----
new allracLing period 2
0.6

0.55 Q.
0.55 0.6 0.65 0.1 0.15

FIGURE 7.6. The graph of h;(x) where hr(x) = rx(l- x)

for three parameter values. In (a) r = 2.9, in (b) r = 3,
and in (c) r = 3.1. Note that the attracting fixed point
splits into a period two attracting orbit with a repelling
fixed point between the two points in the orbit. The period
two orbit is indicated by a dashed line in the bifurcation
diagram shown in Figure 7.5 on the previous page.
 Exercise Set 7 65

and slightly larger than 3. Some examples are shown in Figure 7.6. Notice
that as the parameter value passes 3, the value of (h~)'(Pr) changes from
being less than 1 to being greater than 1. Hence, the fixed point changes
from attracting to repelling. In addition, the continuity of hr requires that
as this happens a period 2 attracting orbit must be added. (The details
of the proof of this fact are also left as an exercise.) This orbit is labeled
by the points al and a2 in Figure 7.6. This type of bifurcation is called a
period doubling bifurcation since a periodic orbit of twice the period of the
original periodic point is added. A bifurcation diagram for the family hr
is shown in Figure 7.5. 0

In closing, we note that there are several ways of drawing bifurcation

diagrams. The choice of diagram depends on the information we are trying
to code. We will see a another type of bifurcation diagram in Chapter
11. We also note that of the four types of bifurcations described here, the
most important are the saddle-node and the period doubling bifurcations.
They arise in many settings and can't be eradicated by small changes in
the parametrized family. 'Iranscritical and pitchfork bifurcations are inter-
esting, but are only observed in certain types of parametrized families and
can be eradicated by small changes.

Exercise Set 7

7.1 Discuss the bifurcations of the following families of functions at the

given parameter values and draw bifurcation diagrams for each.

a) qc(x) = x 2 + cat c = t and c = -£.

b) Fr(x) = rx - x 3 at r = 1.
c) Ec(x) = eX + cat c = -1.
d) gm(X) = X3 + mx at m = 1 and m = -1.
e) Ek(X) = ek(x-l) at k = 1.

7.2 Let hr(x) = rx(l - x).

a) Show that the only fixed points of hr are 0 and r;l.
b) Use the derivative of hr to determine for which parameter values
the fixed points of hr are repelling and for which values they are
attracting.
66 7. Parametrized Families of Functions and Bifurcations

7.3 In Example 7.6 we stated that a period two orbit must be added
as the derivative of h~ at Pr = r-;:l changed from being less than
one to being greater than one. Prove that this is true.
Hint: First show that (h~(Pr))' > 1 implies there exists Xo > Pr
so that h~(xo) > Xo. Then use the fact that h~(l) = 0 and the
Intermediate Value Theorem to show that there must exist a fixed
point of h~ in the interval (Pr, 1). Using a similar technique show
there exists a fixed point of h~ which is in (!, Pr). How do we know
these points are prime period 2 points of h r ? Can you prove that
they are in the same orbit?

7.4 AN INVESTIGATION: Examine the dynamics of hr for the param-

eter values between 0 and 4. Find as many of the bifurcations as
you can and describe each one.
8
The Logistic Function,
Part I

We return to the family offunctions, hr(x) = rx(l-x) where r > O. As we

saw in Chapter 1, this family is a reasonable model of population growth.
We note that the graph of hr is a parabola facing down with x intercepts
at 0 and 1, and with a vertex at the point (~, i).
The following facts have already been discussed or are easily verified:
(1) Solving the equation rx(l- x) = x demonstrates that hr has fixed
points at 0 and at Pr = r~l. Also, 1 and ~ are eventually fixed
points since h r (1) = 0 and hr(~) = Pro
(2) If 0 < r < 1, then Pr < 0, Pr is a repelling fixed point, and 0 is an
attracting fixed point. The stable set of 0 is (Pr" ~), the stable set
of Pr is {Pn ~}, and the stable set of infinity is (-0(1, Pr) U (~, 00).
(3) vVhen r = 1 a transcritical bifurcation occurs, PI = 0 and is not
hyperbolic, the stable set of 0 is [0,1], and the stable set of infinity
is (-00,0) U (1,00). In general, if r :::0: 1, then the stable set of
infinity includes the intervals (-00,0) and (1,00).
(4) If 1 < r < 3, then 0 is a repelling fixed point, Pr is an attracting
fixed point, the stable set of 0 is {O, I}, and the stable set of Pr is
(0,1).
(5) When r = 3 a period doubling bifurcation occurs. 0 is still repelling
and Pr is only weakly attracting. The stable set of Pr is (0,1).
(6) If 3 < r < 3.4, then both 0 and Pr are repelling fixed points.
68 8. The Logistic Function, Part I

°
Note that the stable set of is still {O, I}, but the stable set of
Pr contains an infinite number of points. However, there is now
an attracting period 2 orbit and "most" of the points in (0,1) are
forward asymptotic to one of the two points in the orbit.
(7) When r ~ 3.45 another period doubling bifurcation occurs and the
period 2 attracting orbit splits into a period 4 attracting orbit and
a period 2 repelling orbit.
(8) In the interval 3.44 < r ~ 4 changes occur rapidly. There is a pe-
riod 3 orbit whenever r > 3.6, so by Sarkovskii's Theorem we know
there are periodic points of all orders. Because of the complexity
of the dynamics when r is in this range we will first discuss what
happens when r > 4.
Suppose r > 4. Then h( ~) > 1 and ~ is forward asymptotic to infinity.
Since h(~) > 1, h(l) = 0, and h(O) = 0, the Intermediate Value Theorem

Consequently, h 2(xo) = h 2(Xl) = °

implies there exist Xo in [O,~] and Xl in [~, 1] such that h(xo) = h(Xl) = 1.
and both Xo and Xl are eventually
fixed. (See Figure 8.1 for an illustration of this arrangement.) We leave

°
as an exercise the proof that there are infinitely many points which are
eventually fixed at and that there are infinitely many points which are
eventually fixed at Pro We would like to know what other points of [0,1]
are forward asymptotic to something in [0,1].
For each natural number n we define

An = {X I hn(x) is in [0, I]}. (8.1)

Our goal is to describe the set A = n

(Xl

n=l
An: the set of points that remain in

[0,1] forever under iteration of h. To do this we need to know more about

the sets An. Throughout the remainder of this book, the symbol An will
be used to indicate the set of points that remain in [0,1] under iteration
of h. The symbol A will be consistently used to indicate the intersection
of the sets An. The following proposition is a first step in developing a
description of A.

PROPOSITION 8.1. If h(x) = rx(l - x) and r > 4, then the following

statements are true.
a) The set of real numbers X in [0,1] satisfying the condition that h(x)
is not in [0,1] is the interval

( ~2 _ vr2 - 4r ~
2r' 2 +
vr2 - 4r)
2r
 8. The Logistic Function, Part I 69

and
A= [0 ~2 - vr22r- 4r] U[~2 + vr22r- 4r ' 1].
1 ,

b) The set An consists of 2n disjoint closed intervals for all natural

numbers n.
c) If I is one of the closed intervals in An, then hn : I ----+ [0,1] is
one-to-one and onto.
PROOF. We first note that if h(x) is in [0,1], then x must be in [0,1].
Hence, An is contained in [0,1] for all n. We now prove part (a) of the
proposition.
a) The graph of h(x) is shown in Figure 8.1. We see that the set of
points in [0,1] for which h(x) is not in [0,1] is exactly those points that
satisfy h(x) > 1. But these are the points which lie between the roots of
h(x) = 1. Using the quadratic formula we find that h(x) = rx(1 - x) = 1
when x = ~ ± v'r~;4r and (a) follows.
We prove (b) and (c) by induction. It is clear from part (a) that Al
consists of 21 disjoint closed intervals. Also, we see that h(O) = h(l) = °
°
and h( ~ ± v'r~;4r) = 1. Thus, the endpoints of the two intervals are
mapped to and 1. By continuity and the Intermediate Value Theorem
we can conclude that h : I ----+ [0,1] is onto. To verify that h is one-to-

° °
one on the intervals of AI, we note that it is strictly monotone on these
intervals since h'(x) = r(l - 2x) > when x < ~ and h'(x) < when
x > ~. Therefore, if I is one of the two intervals in AI, then h : I ----+ [0,1]
is one-to-one and onto.
To continue the induction argument, we suppose that Ak consist of 2k
disjoint closed intervals. Further, we suppose that if I is one of the closed
intervals in Ak , then hk : I ----+ [0,1] is one-to-one and onto and that either
° °
(hk)'(x) < for all x in I or (hk)'(x) > for all x in [a, b].
Now consider Ak+1' It is easily seen that A k+1 C Ak; the reader should

°
be able to explain why this is so. Assume [a, b] is one of the intervals in Ak
and (hk)'(x) > for all x in [a,b]. Since hk is strictly increasing on [a,b],
hk is continuous, and hk([a, b]) = [0,1], the Intermediate Value Theorem
implies there exist points X2 and X3 such that
(1) a < X2 < X3 < b,
(2) h k ([a,x2]) = [0, ~ _ v'r~;4r],
(3) hk((x 2, x))
3
-- (1.2 _ v'r 2
2 -4r 1.
r' 2
+ v'r2
2-4r) and
r'
(4) hk([X3, b]) = [~ + v'r~;4r, 1].
The first condition implies that the intervals [a, X2] and [X3, b] are disjoint.
The last three conditions imply that hk+l ([a,x2]) = [0,1]' h k+1(x) > 1
70 8. The Logistic Function, Part I

___ 1 __________ _

FIGURE 8.1. The graph of h(x) = rx(l - x) when r > 4.

The points where h(x) = 1 are Xo = ~ - v'r~;4r and
_1 v'r 2-4r Thevert
Xl - 2 + - 2 - r - ' ex'IS a t(l2' 4r) .

for all x in (X2, X3), and hk+1([X3, b]) = [0,1]. Thus, the set of points in
[a, b] which are also in A k + 1 consists of the two disjoint closed intervals
[a,x2] and [x3,b]. If x is in [a,x2]' then hk(x) is in [0, ~ - v'r~;4T] and
h'(hk(x)) > O. By assumption (hk)'(x) > 0, so for all x is in [a, X2],
(h k+1)'(x) = h'(hk(x))· (hk)'(x) > O.
A similar argument demonstrates (h k+1)'(x) < 0 for all x in [x3,b].
If we begin the argument in the preceding paragraph by assuming that
(hk)'(x) < 0 for all x in [a, b], then we can use an analogous argument to
show that the points of A k + 1 which are in [a, b] are contained in two disjoint
intervals [a,x2] and [.T3,b], hk+1([a,x2]) = [0,1], hk+1([X3,b]) = [0,1], and
that (h k + 1 )' is either strictly positive or strictly negative on each of the
intervals [a, X2] and [X3, b]. Since [a, b] is an arbitrary interval in Ak it
follows that there are twice as many intervals in A k + 1 than there are in A k .
That is, Ak+1 contains 2(2k) = 2k+1 disjoint closed intervals. Further, we
have shown that if J is one of the intervals in A k+1, then hk+1 : J -'> [0,1]
is onto (by conditions (2) and (4) listed above). Also, hk+1 is one-to-one
on J and we can make the next step in the induction since (h k + 1 ), is either
strictly positive or strictly negative on J. D

As we stated above, our goal is to describe the set A = n

00

n=l
An. In the

proof of Proposition 8.1 we saw that Ak+1 was formed by removing an open
interval from the middle of each interval in Ak . The set of points which
are left after this has been done for each n is A. Students of topology
may recognize that this construction is similar to that of a Cantor Middle
Thirds Set.
 8. The Logistic Function, Part I 71

DEFINITION 8.2. A set r c R is called a Cantor set if

a) r
is closed and bounded. (Sets of real numbers with these charac-
teristics are called compact sets.)
b) r contains no intervals. (Sets of this nature are called totally dis-
connected sets.)
c) Every point in r is an accumulation point of r. (Such sets are
called perfect sets.)
While most of the terms used in Definition 8.2 should be familiar, we
have not yet defined bounded sets. A set of real numbers is bounded if
there exists a positive number which is larger than every element of the
set. This is equivalent to saying that there exists a closed interval [a, b]
which contains the set.

EXAMPLE 8.3.
CONSTRUCTING CANTOR SETS.
Cantor sets are usually constructed by iterative processes. To construct
the Cantor Middle Thirds Set we begin with the interval [0,1] and remove
the open set (!, ~). In the second step we remove the middle third of each
remaining interval. That is, we remove the intervals (!,~) and (~, ~). We
continue the process by removing the middle third of each of the remaining
intervals at each step. The set of points remaining is called the Cantor
Middle Thirds Set. As we will see shortly, this is in fact·a Cantor Set.
H 0 < a < 1, then we can construct similar sets called Cantor Middle-a
sets by removing an interval from the center of each remaining interval
whose length is a times the length of the remaining interval. For example,
in the Middle-i set we remove the middle fifth of each interval at each step.
To be more precise, Let ro = [0,1]. Let r l be the two closed inter-
vals of equal length left when the open interval of length a is removed
from the middle of r o. Define r 2 to be the set of 4 closed intervals
of equal length obtained by removing an open interval whose length is
a x (the length of an interval in r d from the middle of each interval in
r l . We continue to define r n inductively as the set of closed intervals
of equal length formed by removing an open interval whose length is a x
(the length of an interval in r n-I) from the center of the closed intervals
in r n-I. The first few iterations of this construction for a = 3 are shown
in Figure 8.2. The Cantor Middle-a Set is the set r = n~o r n.
Clearly, r is not empty since the endpoints of the open intervals removed
are in r. We shall see in Proposition 8.5 that r is perfect. Students who
have studied count ability will be interested to know that nonempty perfect
subsets in lR must be uncountable. This is proven in the analysis text by
Rudin which is listed in the references. It is also interesting to note that
72 8. The Logistic Function, Part I

113 2/3

FIGURE 8.2. The first five sets formed in the construction

of the Cantor Middle Thirds Set.

only count ably many of the points in r are endpoints of the intervals that
are removed, so there are uncountably many points in r which are not
endpoints. 0

LEMMA 8.4. If r n is as defined in Example 8.3, then there are 2n closed

intervals in r n and the length of each closed interval is e2<> r.
Also, the
combined length of the intervals in r n is (1- a)n which approaches 0 as n
approaches infinity.
PROOF. We start with an interval of length 1 and proceed by mathe-
matical induction.
In the first step we remove a gap of length a and are left with 2 = 21
closed intervals with a total length of (1 - a) so each interval has length
1 2<> and the total length of the intervals is 1 - a.
In general, suppose there are 2k intervals left in r k , each with length
e 2<» k for a total length of (1 - a) k. We will show that there are 2k+ 1
intervals left in r k+ b each with length e2<> ) k+l for a total length of
(1 - a)k+l. Note that each time we remove the middle-a portion of a
closed interval, we split the interval into two closed intervals. So in go-
ing from r k to rk+l we double the number of intervals and there are
2(2k) = 2k+l intervals in fk+l. After the kth step each interval has length
e 2<» k. We remove the middle-a portion of each interval so the amount
of each interval from r k left in rk+l is (1 2<» k-a (1 2<» k
= (l-~t+1).
Since this length is left in two intervals, the intervals each have length
! «l-~t+l) = e2<»k+l. Finally, there are 2k+l intervals so their total
length is (2k+l) e2<»k+l = (1- a)k+l.
Since 0 < a < 1, (1- a)n converges to 0 as n grows without bound and
it follows that the total length of the intervals in r n approaches 0 as n goes
to 00. 0
We are now ready to show that Cantor Middle-a Sets are appropriately
named.
 8. The Logistic Function, Part I 73

PROPOSITION 8.5. The Cantor Middle-a Set is a Cantor set.

PROOF. Let r be a Cantor Middle-a set. We must show that a) r is

closed and bounded, b) r contains no intervals, and c) every point of r is
an accumulation point of r.
a) Since r is the intersection of nested closed intervals, Proposition 3.14
implies that it is closed. As r is contained in [0,1], it is also bounded.
b) If r contains an open interval (x,y) with length Iy - xl, then at
each stage in the construction of r, (x, y) must be contained in one of the
remaining closed intervals. However, Lemma 8.4 implies that after n steps
the length of one of these intervals is C2 t
a and we can find an no such
that C2 ato < Iy - xl· That is, the length of each of the closed intervals
in r no is less than the length of (x, y). Hence, the entire interval (x, y)
cannot be contained in r no and r contains no intervals.
c) Suppose that x is a point in r and let Ne(x) = (x - E,X + E) be a
neighborhood of x. We must show that there exists a point in r that is
contained in Ne(x) and is not equal to x. Notice that if a is an endpoint of
one of the intervals which is removed, then a is in r. Now at each stage in
the construction of the Cantor Set, x must be in one of the remaining closed
intervals. That is, for each n there is an interval in r n which contains x.
Choose n large enough so that C2 a t < E. Then there is a closed interval
I with length C2 t
a in r n such that x is in I. Since C2 a t < E, it must
be that the endpoints of I are in N,(x). As there are two endpoints and x
can be equal to, at most, one of them, we are done. 0

Note that the proof of Proposition 8.5 does not use the fact that we
removed the a-interval from the exact middle of the closed interval. The
fact that the total length of all the intervals approaches zero follows from
the observation that if we remove a fraction a at each stage, then the
length is reduced by a factor of 1 - a so the total length of the remaining
intervals is (1- a)n after n steps. In fact, if at each step we remove an open
interval whose length is a times the length of each closed interval and we
remove this interval in such a way that the endpoints of the open interval
do not coincide with either of the endpoints of the closed interval, then we
obtain a Cantor set. It can also be shown that if a subset of the real line
is closed and bounded, contains no open intervals, and is perfect, then it is
homeomorphic to the Cantor Middle Thirds Set.

We complete this chapter by showing that the set of points which remain
in [0,1] under iteration of hr is a Cantor set whenever r > 2 + J5. This is
true for all r > 4, but it is a bit trickier to prove for r < 4 ::::; 2 + J5. To
do this we use the following lemma:
74 8. The Logistic Function, Part I

LEMMA 8.6. Let r > 2 + v'5 and h(x) = rx(l- x). Then there is € > 0
such that Ih'(x)1 > 1 + € whenever x is in A 1 . Further, the length of each
interval in An is less than (l;€)n.

Note that the value of € found in the lemma depends on the value of r.
As r gets smaller € must get smaller as well.
The proof of the first part of the lemma follows from the fact that if
x is in Al. then Ih~(x)1 is greater than or equal to the absolute value of
the derivative of hr at (~ ± v'r~;:-4r). We prove the second portion of the
lemma by using the first part and the Mean Value Theorem. The proof is
outlined in exercise 8.3.

THEOREM 8.7. The set A = n00

n=l
An is a Cantor set.

PROOF. We need to show that a) A is closed and bounded, b) A contains

no intervals, and c) every point in A is an accumulation point of A.
a) Since A is the intersection of nested closed sets, Proposition 3.14
implies that it is closed. As A is contained in [0,1), it is also bounded.
b) If A contains the open interval (x, y) with length Ix-yl, then for each
n, (x, y) must be contained in one of the intervals of An. However, Lemma
8.6 implies that there is an € > 0 such that the length of an interval in An
is less than ( 1) . Since we can find no so that Ix - yl > (1 1) ,
1+€n +€~
the interval (x,y) can not possibly fit in to an interval in Ano' Hence, A
contains no open intervals.
c) Finally, suppose x is a point in A and let N6(X) = (x - 8, x + 8) be a
neighborhood of x. We must show that there is a point in A other than x
which is contained in N 6 (x). Notice that if a is an endpoint of one of the
intervals in An, then a is in A since hn+1(a) = O. Now for each n, x must
be contained in one of the intervals of An. We let € be as in Lemma 8.6
1
and choose n large enough so that (1 + €)n < 8. Then the entire interval
of An must be in N6(X) since the length of each interval in An is less than
( 1) . Since both of the endpoints of the interval are in N6(X) and at
1+€ n
least one of them is not x, we are done. 0
DEFINITION 8.8. The set 0 is a hyperbolic repelling set of the function
f if 0 is closed and bounded, f(O) = 0, and there is N > 0 such that
I(r)'(x) I > 1 for all x in 0 and all n 2:: N. Similarly, the set 0 is
a hyperbolic attracting set of the function f if 0 is closed and bounded,
f(O) = 0, and there is N > 0 such that l(r),(x)1 < 1 for all x in 0 and
all n 2:: N.
 Exercise Set 8 75

In the exercises we ask the reader to show that A is a hyperbolic repelling

set of hr(x) = rx(l - x) when r > 2 + v's.
Cantor sets were discovered by the German mathematician Georg Can-
tor in the last part of the nineteenth century. At the time of their discovery,
they created a considerable stir in the mathematical community since their
properties are counter intuitive. For example, it is hard to imagine points
other than the endpoints of the intervals in each r n which are in r, yet we
know that there are more of them than there are endpoints of the intervals.
Throughout much of the twentieth century Cantor sets were considered to
be little more than a mathematical curiosity. However, the work of Stephen
Smale in the 1960's demonstrated that Cantor sets play an important
role in dynamical systems. We saw one indication of their importance in
Theorem 8.7.
A Cantor set is also an example of a fractal. One of the commonly cited
properties of fractals is that they are self-similar under magnification. In
the exercises we ask the reader to show that the Cantor Middle Thirds Set
is homeomorphic to a subset of itself. More generally, we ask the reader
to show that if I is an open interval whose intersection with the Cantor
Middle Thirds Set is not empty, then the intersection contains a subset
which is homeomorphic to the Cantor Middle Thirds Set.

Exercise Set 8

8.1 Show that h(x) = rx(l - x) has infinitely many eventually fixed
points when r ?: 4.

8.2 a) Prove that An is contained in [0,1] for all n.

b) Explain why Ak+1 c Ak.

8.3 • Prove Lemma 8.6 with the steps outlined below.

2-4r) > 1 and h'(! + v'rL4r) < -1 when
a) Prove that h'(!2 - v'r2r 2 2r
r > 2 + v's. Show that h"(x) < 0 for all x and conclude that h'(x)
is decreasing. Explain why these two statements guarantee that
there is E > 0 so that Ih'(x)1 > 1 + E when x is in AI.
b) Use part (a) and exercise 6.2 to show that if c is any point in
An, then (hn)'(c) > (1 + E)n.
c) Show that the Mean Value Theorem and part (b) imply that
if x and y are the endpoints of one of the intervals in An, then
76 8. The Logistic Function, Part I

Ihn(x) - hn(y)1 > (1 + €)nlx - YI.

Explain how we know that
1
Ihn(x) - hn(y)1 < 1 and conclude that Ix - yl < (1 + €)n·

8.4 a) Use Proposition 8.1 and Theorem 4.4 to show that if r > 2+V5,
then hr(x) = rx(l - x) has 2n periodic points with period n.
b) Use Lemma 8.6 and exercise 6.2 to show that the periodic points
of from part (a) are all repelling.
c) Show that if r > 2 + V5, then the periodic points of hr are dense
in A.

3X, for x < !

8.5 a) Let 9 : lR -lR be defined by g(x) = { - ~ Show
3 - 3x, for x 2: 2.
that the set
r = {x in [0,1]1 gn(x) is in [0,1] for all n}
is the Cantor Middle Thirds Set.
b) Let r be the Cantor Middle Thirds Set. Use part (a) to show the
map f : r - [O,~] n r defined by f(x) = ~x is a homeomorphism
of r to a subset of itself.
c) Show that if I is one of the intervals in r n used in the construc-
tion of the Cantor Middle Thirds Set, then r is homeomorphic to
rnI.
d) Prove that if I is an open interval whose intersection with the
Cantor Middle Thirds Set is not empty, then the intersection of
I and the Cantor Middle Thirds Set contains a subset which is
homeomorphic to the Cantor Middle Thirds Set.

8.6 Show that A is a hyperbolic repelling set of h(x) = rx(l-x) when

r > 2 + V5.

8.7 • THE TENT MAP: Let T: [0,1]- [0,1] be the tent map defined
by
T(x) = {2X' for x ~n [~,~]
2 - 2x, for x 10 [2,1].

a) Graph T(x), T2(x), T3(x), and T 4 (x). (You may wish to use
a computer graphing tool. See the appendix for suggestions as to
how one can define T for the computer.)
 Exercise Set 8 77

b) Show that T has 2n repelling periodic points with period n.

c) Show that the periodic points of T are dense in [0, 1].
9
Symbolic Dynamics
and Chaos

We begin this chapter with a discussion of symbolic dynamics. As we

proceed through the discussion, it may seem that the dynamics of the shift
map on symbol space is an odd place to look for a deeper understanding
of the logistic map, but we shall see that it is precisely the tool we need.

DEFINITION 9.1. Let 2::2 be the set of all infinite sequences of O's and
1 'So This set is called the sequence space of 0 and 1 or the symbol space of
o and 1. More precisely, 2::2 = {(SOS1S2 ... ) I Si = 0 or I}. We will often
refer to elements of 2::2 as points in 2:: 2 .

The concept of distance has proven to be useful when working on the

real line and we use it again in symbol space.

DEFINITION 9.2. Let s = SOS18283 ... and t = tOt1t2t3 ... be points in

2:: 2 • We denote the distance between sand t as drs, t] and define it by

drs, t] = f
i=O
1
8i ; til.

Since lSi - ti I is either 0 or 1,

1
o ::; drs, t] ::; L
00

2i =2
i=O
80 9. Symbolic Dynamics and Chaos

Hence, drs, t] is a number between 0 and 2. Also, it is clear that drs, t] = 0

if and only if s = t. The distance function which we have described is an
example of a metric. In general, we define a metric by

DEFINITION 9.3. Let X be a set and let d be a function from the set of
all ordered pairs of elements of X into the real numbers. If the following
conditions hold for all x, y, and z in X, then d is a metric on X.
a) d[x, y] ~ 0 and d[x, y] = 0 if and only if x = y.
b) d[x,yj = d[y,x].
c) d[x,yj ~ d[x,zj +d[z,yj.

We are already familiar with one metric: the distance between two points
in the real numbers is the absolute value of their difference. For example
the distance between the numbers 2 and 3 is \2 - 3\ = 1 and the distance
between 2 and -5 is \2 - (-5)\ = 7. It should be clear that this distance
function satisfies conditions (a) and (b) of the definition. The reader may
recognize condition (c) in Definition 9.3 as the Thiangle Inequality, which
was proven for the metric on the real numbers in exercise 2.8. In the
exercises at the end of this chapter we show that the distance function in
symbol space is a metric and evaluate several other functions as possible
metrics.
The obvious question now is "Why do we care?". If we refer back to
Chapter 3 we see that all of our topological definitions and the definition
of continuity are based on the idea of distance. So once we have defined a
metric on a set we can define a topology for the set. That is, we can decide
which subsets are open, which are closed, what a convergent sequence is
in the set, and which functions of the set are continuous. All of these
definitions are brought together below.

DEFINITION 9.4. Let X be a set and d be a metric on the set.

a) A subset U of X is open if for each x in U there exists E > 0 such
that d[x,yj < E implies y is in U.
b) Let E > 0 and x be in X. The set NE(x) = {y in X \ d[x,y] < E}
is called a neighborhood of x.
c) Let Xl, X2, X3, . .. be a sequence of elements of X. The sequence
converges to x if for each E > 0, there exists an integer N such
that if k ~ N, then d[x, Xk] < E-
d) Let S be a subset of X. Then the point x in X is an accumulation
point (or a limit point) of S, if every neighborhood of x contains
an element of S which is distinct from x.
e) A subset of X is closed if it contains all of its accumulation points.
 9. Symbolic Dynamics and Chaos 81

f) Let A be a subset of B. Then A is dense in B if every point of

B is an accumulation point of A, a point of A, or both. In other
words, if A is dense in B and x is in B, then every neighborhood
of x contains an element of A.
g) If Y is a set and d2 is a metric on Y, then the function f : X --+ Y
is continuous at the point Xo in X if for every E > 0 there exists
8 > 0 such that
if x is in X and d[xo, x] < 8, then d2 [f(xo), f(x)] < E.

A function is continuous if it is continuous at each point of its

domain. (Note that f(xo) and f(x) are elements ofY so we must
use the metric d2 to measure the distance between them.)
Notice the importance of distance in the previous definitions. The set U
is open if the points near each point (within the distance E of each point)
are in U. The sequence {x n } converges to x if the point Xn is close to x
(within the distance E of x) when n is sufficiently large. The point x is
an accumulation point of S if we can find a point of S as close to x as we
want. A is dense in B if there is a point in A close (within the distance E)
to each point of B. Finally, the function f is continuous at Xo if we can
guarantee f(x) is close to f(xo) whenever x is close enough to Xo.
All of the topological facts that we have shown to be true for the real
numbers are also true for more general metric spaces. We gather together
some of these facts in the following propositions. The proofs of all of these
propositions follow directly from the definitions and are left as exercises.
Students who wish to master the concepts of metric spaces and their topolo-
gies are urged to complete the proofs of all the propositions and study the
examples that follow.
PROPOSITION 9.5. The set U is open if and only if for each x in U there
is a neighborhood of x which is completely contained in U.
PROPOSITION 9.6. If X is a metric space with metric d, x is in X and
E> 0, then the neighborhood N€(x) is open.
PROPOSITION 9.7. Let X be a metric space with metric d. If x is in X
and S is a subset of X, then the following statements are equivalent:
a) x is an accumulation point of S.
b) For each E > 0 there exists y in S such that 0 < Ix - yl < Eo

c) If U is an open set containing x, then U nS contains at least one

point other than x.
d) There is a sequence of points different from x and contained in S
which converges to x.
82 9. Symbolic Dynamics and Chaos

PROPOSITION 9.8. The complement of an open set is closed. Conversely,

the complement of a closed set is open.
PROPOSITION 9.9. Let X be a metric space with metric d and A be a
subset of X. Then the following statements are equivalent:
a) A is dense in X.
b) For all points x in X and all E > 0, there exists a in A such that
d[x,a) < E.
c) For each x in X, there is a sequence of points al,a2,a3, ... con-
tained in A that converges to x.
d) Every open subset of X contains an element of A.
PROPOSITION 9.10. Let X and Y be metric spaces and f: X -+ Y be a
continuous function. If U is an open set in Y, then f- 1 (U) is an open set
in X.
Before moving on to a few examples, we pause to prove a useful lemma
about ~2.
LEMMA 9.11. Let sand t be elements of ~2. If the first n + 1 digits in
sand t are identical, then drs, t) ~ 2~. On the other hand, if drs, t) ~ 2~'
then the first n digits in sand t are identical.
PROOF. Let s = SOSIS2 ... and t = tOtlt2 ... be sequences in ~2. We
note first that the first n + 1 digits of s are so, Sl, ... , Sn. So, sand t
agree on the first n + 1 digits if and only if Si = ti for i ~ n.
Now suppose Si = ti for all i ~ n. Then

drs, t) = f
i=O
lSi; til
n 0 00

=L 2i+ L
i=O i=n+l

On the other hand, if there is j < n such that Sj ¥- tj, then

 9. Symbolic Dynamics and Chaos 83

EXAMPLE 9.12.
a) The distance between 8 = 000000 ... and t = 010101010 ... is

b) If S is the set of all elements of ~2 which begin with the sequence 011,
then S is closed.
To demonstrate this we show that all of the accumulation points of S are
contained in S. If 8 is an accumulation point of S, then every neighborhood
of 8 contains an element of S other than 8. Consequently, there is 8* in
N 1 (8) n S satisfying 8* =I- 8. But then the definition of neighborhood
~
implies d[8*, 8] < d-:> and Lemma 9.11 implies that the first 3 digits of 8 and
8* must agree. Since 8* is in S, the first three digits of 8* are 011. Hence,
the first three digits of 8 are 011 and 8 is in S.
c) The set of all sequences in ~2 which end with an infinite string of O's is
dense in symbol space. That is, if

A= {8081 82 . .. in ~2 I there is N satisfying 8i = 0 for all i ~ N},

then A is dense in ~2'

To show this, let t be in ~2 and N€(t) be a neighborhood of t. We must
prove that there is 8 in An N€ (t). That is, we must show that there is 8 in
A such that d[8, t] < Eo Choose n so that 2~ < E. Let 8 be the element of
~2 whose first n + 1 digits are the same as t's first n + 1 digits and whose
remaining digits are all O's. Then 8 is in A by definition and Lemma 9.11
implies d[8, t] :S 2~ < Eo Hence, N€(t) contains an element of A and we're
done. 0

Now that we have a basic understanding of ~2 we are ready to examine

the dynamics of the shift map on ~2'

DEFINITION 9.13. The 8hift map (J' : ~2 -t ~2 i8 defined by

(J'(808182 .. ·) = 81 8 2 8 3 ....

In other words, the shift map "forgets" the first digit of the sequence. For
example, (J'(01110101 ... ) = 1110101 ....
84 9. Symbolic Dynamics and Chaos

There are several immediate observations we can make about the shift
map.
PROPOSITION 9.14. The shift map is continuous.
PROOF. Let s be an element of E2 and f > O. We must show there is
8 > 0 such that whenever drs, t] < 8, then d[a(s),a(t)] < f. Choose n so
that 2~ < f and let 8 = 2n\1' If drs, t] < 8, then we know from Lemma
9.11 that sand t agree on the first n + 2 digits. So a(s) and a(t) agree on
the first n + 1 digits and d[a(s), a(t)] S 2~ < f as desired. 0
It should be clear by now that two points are close together if and only
if their initial digits agree. The more digits they agree on before they differ
the closer together they are. Of course, this is exactly what Lemma 9.11
claims. Example 9.12 and the proof of Proposition 9.14 illustrate the use
of this idea. We will use this idea once again in the proof of part (a) of
Proposition 9.15 and the reader will be given an opportunity to use it in
proving parts (c) and (d).
PROPOSITION 9.15. The shift map has the following properties:
a) The set of periodic points of the shift map is dense in E 2 •
b) The shift map has 2n periodic points of period n.
c) The set of eventually periodic points of the shift map which are not
periodic is dense in E 2 •
d) There is an element of E2 whose orbit is dense in E 2. That is,
there is s* in E2 such that the set {s*, a(s*), a 2(s*), a 3 (s*), ... } is
dense in E 2 .
e) The set of points which is neither periodic nor eventually periodic
is dense in E 2 •
PROOF. a) Suppose 8 = 808182 ... is a periodic point of a with period
k. Then an(ak(s)) = a n(8). Since an(s) "forgets" the first n digits of 8 we
see that

a n (a k (8081S2"')) = 8n+k Sn+k+1 8n+k+2···

= 8n8n+18n+2 = an (80 8 18 2 ... )
and Sn+k = 8n for all n. This implies that s is a periodic point with period
k if and only if 8 is a sequence formed by repeating the k digits 8081 ... Sk-l
infinitely often.
To prove that the periodic points of a are dense in E 2 , we must show
that for all points t in E2 and all f > 0, there is a periodic point of a
contained in No (t). From the preceding discussion we see that this means
we need to find a point in No (t) which is a sequence formed by repeating
 9. Symbolic Dynamics and Chaos 85

the initial k digits infinitely often. But, if t = totlt2t3 ... and we choose n
so that in < €, then we can let s = totl ... tntotl ... tntotl . . .. As t and s
agree on the first n + 1 digits, Lemma 9.11 implies drs, tj :::; 2~ < € or s is
in NE (t) and s is a periodic point by construction.
The proofs of parts (b) and (c) are left as exercises.
d) The sequence which begins with 0 1 00 01 10 11 and then includes
all possible blocks of 0 and 1 with three digits (there are eight) followed
by all possible blocks of 0 and 1 with four digits (there are sixteen) and so
forth is called the Morse Sequence. We leave as an exercise the proof that
the orbit of this sequence is dense in ~2.
e) Since the set of non-periodic points includes as a subset the orbit of
the Morse Sequence defined above, the truth of part (e) of this proposition
follows from part (d). D
The shift map has several characteristics worth highlighting. First we
note that the periodic points of a are dense in ~2' This was proven in
the previous proposition. This implies that in every neighborhood of every
point in ~2 there is a periodic point of a. Further, by Proposition 9.9, if
s is in ~2' then there is a sequence of periodic points Sl, S2, S3, •.. which
converges to s. We can think of this property as guaranteeing a certain
amount of regularity in the dynamics of the shift map. Periodic points of
a are everywhere in the domain.
Second, we note that ~2 is well mixed by a. We define this mathemati-
cally as topological transitivity.
DEFINITION 9.16. The function f : D ---> D is topologically transitive if
for all open sets U and V in D, there is x in U and a natural number n
such that r(x) is in V.
We sometimes describe the action of a function which is topologically
transitive as mixing the domain. If U is any open set in the domain of the
function, then some point of U will eventually land in every neighborhood
of every point in the domain under iteration of the function. Proposition
9.15d in conjunction with the following proposition guarantees that the
shift map is topologically transitive.
PROPOSITION 9.17. Let f be a function and suppose there is a point
whose orbit under iteration of f is dense in the domain of f. Then f is
topologically transitive.
PROOF. Let f : D ---> D and let U and V be open sets in D. We
must find x in U and a natural number n such that fn(x) is in V. By
hypothesis, there is a point in D, whose orbit is dense in D. If Xo is such
a point, then we know there exists k such that fk(xO) is in U. If we can
86 9. Symbolic Dynamics and Chaos

show that there is a natural number n such that fk+n(xo) is in V, then

we are done since we can let x = fk(xo). In this case, x is in U and
r(x) = rUk(xo)) = r+k(xo) is in V.
Since the orbit of Xo is dense, we know that V contains at least one
iterate of Xo. Suppose there are only finitely many iterates of Xo in V. If
we let v be any element of V which is not an iterate of Xo, then the number

E = min{lv - r(xo)l}

°
is greater than and the neighborhood N€/2 (v) doesn't contain any it-
erates of Xo. Thus, the orbit of Xo is not dense in D, a contradiction.
Consequently, V must contain infinitely many iterates of Xo. Since there
are only finitely many natural numbers less than k, there exists m > k such
that fm(xo) is in V. If we let n = m - k, the proof is complete. 0

Finally, we note that for all s in E2 and E > 0, there is t in E2 and a

natural number n such that d[s,t] < E, but d[an(t),an(s)] = 2. This is
demonstrated by choosing n so that 2~ < E and choosing t so that the first
n digits of t are the same as the first n digits of s and the other digits of t
are all different from the corresponding digits of s. That is, Si = ti if and
only if i < n. In this case, drs, t] = 21n < E. As an(s) = SnSn+1Sn+2'" and
an(t) = tntn+ltn+2 ...

d[an(t),an(s)] = f
i=O
ISHn; tHnl = f
i=O
~ = 2.

In fact, we see that d[ak(s),ak(t)] = 2 for all k 2: n. So, after a finite

number of iterations sand t are as far apart as two sequences can be in
E2 even though we can initially choose them to be as close together as
we want. This is an extreme example of sensitive dependence on initial
conditions.

DEFINITION 9.18. Let D be a metric space with metric d. The function

f : D ---. D exhibits sensitive dependence on initial conditions if there exists
°
a 8 > such that for all x in D and all E > 0, there is a y in D and a
natural number n such that d[x, y] < E and d[r(x), r(y)] > 8.

Practically speaking, sensitive dependence implies that if we are using an

iterated function to model long-term behavior (such as population growth,
the weather, or economic performance) and the function exhibits sensitive
dependence, then any error in measurement of the initial conditions may
result in large differences between the predicted behavior and the actual
behavior of the system we are modeling. Since all physical measurements
include error, this condition severely limits the utility of our model.
 9. Symbolic Dynamics and Chaos 87

Functions which exhibit the three characteristics just outlined for a are
said to be chaotic.
DEFINITION 9.19. (DEVANEY!) Let D be a metric space. Then the
function f : D -+ D is chaotic if
a) the periodic points of f are dense in D,
b) f is topologically transitive, and
c) f exhibits sensitive dependence on initial conditions.
Restated, this definition implies that a chaotic function observes a cer-
tain amount of regularity and mixes the domain well. Further, even the
smallest changes in initial position may result in dramatically different re-
sults in values under iteration of the function. The regularity is evidenced
by the fact that we can always find a periodic orbit in every neighborhood,
no matter how small, of every point in the domain of the function. On the
other hand, the domain is well mixed by the function since if we choose
any open set, then we can find a point in every other open set that will
eventually end up in the first set under iteration of the function. As we
have seen, a : ~2 -+ ~2 exhibits all of these behaviors and is chaotic.
We complete this section with a new and surprising result which ap-
peared in the April 1992 issue of the American Mathematical Monthly
(pages 332-334) and which was proven by Banks, Brooks, Cairns, Davis,
and Stacey.
THEOREM 9.20. Let f : X -+ X be topologically transitive and suppose
the periodic points of f are dense in X. If X contains an infinite number
of elements, then f exhibits sensitive dependence on initial conditions.
This result implies that when the periodic points of a function are dense
in its domain and the function is topologically transitive, then the function
is chaotic. It is not necessary to check for sensitive dependence on initial
conditions. The assumption that X contain an infinite number of elements
is added to eliminate trivial cases. A function of a set containing only one
element will satisfy the hypotheses ofthe theorem, but the conclusion can't
hold. Counter examples exist for larger finite sets as well.
PROOF. Let f : X -+ X be topologically transitive and suppose the
periodic points of f are dense in X. Let d be the metric on X. We begin
by showing there is 80 > 0 such that for all x in X there is a periodic point
q such that the distance from x to r(q) is greater than or equal to 80 for
all n. In other words, every x in X is at least 80 away from a periodic
IThere are several definitions of chaotic. This definition was introduced by R. De-
vaney in his text, Introduction to Chaotic Dynamical Systems, which was published in
1989 and is listed in the references.
88 9. Symbolic Dynamics and Chaos

orbit. We define Do by choosing two periodic points p and q with different

orbits and setting

Do = ~ min{d[r(p), fm(q)ll n and m are natural numbers}.

Since there are only finitely many points in the orbits of p and q, Do > O.
By definition, all of the points in the orbit of p are at least 2Do units away
from all of the points in the orbit of q. Then by the triangle inequality,

2Do ::; d[r(P), fm(q)]

::; d[r(p), xl + d[fm(q), x]
for all natural numbers n and m. It follows that when d[r (p), x] ::; Do for
some n, then d[fm(q), x] ~ Do for all m, and conversely. Hence, every point
x is at least Do units away from either every point in the orbit of q or every
point in the orbit of p.
Now let D = ~Do. We will show that for all x in X and all f > 0, there
is a y such that d[x, y] < f and d[r(x), r(y)) > D for some n. In other
words, we will show that D satisfies the criteria for demonstrating f has
sensitive dependence on initial conditions.
Let x be any element of X and choose f > O. Since the condition is more
restrictive as f becomes smaller we may assume without loss of generality
that f < D. As the periodic points of f are dense in X, we can find a
periodic point p with prime period k such that d[x,p) < f. Also, there
exists a periodic point q such that every point in its orbit is at least a
distance 4D = Do from x. We demonstrated above that such an orbit exists.
Finally, define

V = nk

i=O
f-i(No(fi(q)))

= {x I d[t(X),fi(q)) < D for 0 ::; i ::; k}.

Recall that N 6 (fi(q)) is the set of points within a distance D of fi(q).

The fact that q is in V is obvious from the definition of V. (If it is not
obvious, then a little time used now trying to understand the definition of
V would be time well spent.) We claim V is open as well. Proposition 9.6
implies that N 6 (fi(q)) is open for all i. As fi is continuous, it follows from
Proposition 9.10 that f-i(No(fi(q))) is open for each i. Finally, we stated
in Proposition 3.14 that the intersection of a finite number of open sets is
open, so

nf-
k
V = i (N6 (t(q)))
i=O
 9. Symbolic Dynamics and Chaos 89

.\ .

FIGURE 9.1. An illustration ofthe location of objects used

in the proof of Theorem 9.20.

is open. (Proving Proposition 3.14 from the definition of open sets in a

metric space is a good exercise in topology.)
Now since V and N€(x) are open, the topological transitivity of J implies
there exist y and m such that y is in N€(x) and Jm(y) is in V. Let j be
the integer satisfying T T
< j :::; + 1, that is, 0 < kj - m :::; k.
We are now ready to complete the proof. Before we do, let 's make
sure we have all of our facts straight. We have the arbitrary point x and
o < f < 8. Inside N€(x) we find three points: x; p, a periodic point with
period k; and y. We also have another open set V which doesn't intersect
N€(x) (why is this true?) and such that (1) the periodic point q is in V ,
(2) if z is in V and i :::; k, then the distance between Ji(Z) and Ji(q) is less
than 8, and (3) Jm(y) is in V. This is illustrated in Figure 9.1.
To complete the proof, we will show that either d[Jkj (p) ,Jkj (x)] > 8 or
d[Jkj (x), Jk j (y)] > 8. Since the distance from x to either of the points p
or y is less than f, this will prove that J exhibits sensitive dependence on
initial conditions.
We note that the distance from Jkj(y) = Jkj-m(fm(y)) to Jkj-m(q) is
less than 8 since we chose y so that Jm(y) is in V , the distance between
Ji(z) and Ji(q) is less than 8 whenever z is in V and i :::; k, and since we
chose j so that kj - m :::; k. Also, the Triangle Inequality implies

d[x , Jkj-m(q)] :::; d[x ,p] + d[p, Jk j (y)] + d[Jkj (y), Jkj-m(q)]. (9.1)

Since we have assumed that d[x,p] < f < 8 and we have shown that
d[Jkj(y),Jkj-m(q)] < 8, inequality (9.1) implies

(9.2)

Recall that we chose q to be a periodic point with the property that

90 9. Symbolic Dynamics and Chaos

d[x, fkj-m(q)] > 48 = 80 . Thus, it follows from inequality (9.2) that

48 < dfp, fk j (y)] + 28 or 28 < dfp, fk j (y)]. (9.3)
Since p is periodic with period k, we see that fkj(p) = p and we conclude
from inequality (9.3) that
28 < d[fk j (p), fk j (y)].
Finally, by the Triangle Inequality we have
28 < d[fk j (P), fk j (y)] ~ d[fk j (p), fk j (x)] + d[fk j (x), fk j (y)].
Therefore, either d[fkj(p),Jkj(x)] > 8 or d[fkj(x),fkj(y)] > 8 and our
proof is complete. 0

Exercise Set 9

9.1 a) Show that the distance function defined on ~2 in Definition 9.2

is a metric.
b) Let z = 00000000 ... , r = 11111 ... , and s = 001001001001 ....
Calculate d[z, r], d[r, s], and d[z, s].
c) Show that the largest distance two points can be apart in ~2
is 2.

9.2 Let C be the set of all ordered pairs ofreal numbers. (Le., C is the
Cartesian plane.)
• a) Show that dl [(XI,YI),(X2,Y2)] = V(XI-X2)2 + (YI-Y2)2 is
a metric.
Let y = (1, 2) be a point in C and graph the set
NI(y) = {x in C I dl(x,y) < I}.

b) Show that d2[(XI, YI), (X2' Y2)] = IXI - x21 + IYI - Y21 is a metric.
Let y = (1, 2) be a point in C and graph the set
NI(y) = {x in C I d 2 (x,y) < I}.

9.3 Is the function d3[x, y] = (x - y)2 a metric on the real numbers?

How about d4[X,y] = 1~~~~I?

9.4 Prove Propositions 9.5, 9.6, 9.7, 9.8, 9.9, and 9.10.
Hint: Use the Triangle Inequality for Proposition 9.6.
 Exercise Set 9 91

9.5 Let 8' be the set of all points which don't begin with OIl. Use
Definition 9.4a to prove that 8' is open. Note that we know this
is true since 8' is the complement of the closed set 8 in Example
9.12b.

9.6 Let r = 101010 ... and N~(r) = {s I d[s,r] < n. Describe the set
of points which are in Nl (r).
2

9.7 Let 0 1 , O2 , 0 3 , ... , On be a finite collection of open sets. Prove

that nOk
n

k=l
is an open set.

9.8 Complete the proof of Proposition 9.15.

9.9 Find all the points of period three in I: 2. Which of these points
have prime period three? Which are in the same orbit? Does
Sarkovskii's Theorem apply here? Why or why not?

9.10 a) If s = 101010 ... , then s is a period 2 point of (L Determine the

stable set of s.
b) Let t be any element of I: 2. Find all points r in I:2 that satisfy
the condition that d[a-n(r), a-n(t)] approaches 0 as n goes to infinity.

9.11 Suppose m > 1 and let f : JR. -+ JR. be defined by f(x) = mx.
Show that f exhibits sensitive dependence on initial conditions. Is
f chaotic? Explain.

9.12 • THE DOUBLING MAP ON THE CIRCLE: Let 8 1 be the unit circle
and identify each point on the circle by the radian measure of
the angle between the positive x-axis and the ray beginning at
the origin and passing through the point. We will always measure
angles in a counterclockwise direction. The reader should recognize
this as the usual representation of the unit circle and angles on it.
For example the point (1,0) is labeled 0, and the point (0,1) is
labeled ~. We will also find it useful to identify the point a with
the point oo+2mf where n is any integer. For example, the points 0,
21f, -21f, 41f, ... are all identical. Also, i, 737T , and - 5; are all the
same point on the circle. Again, this is the standard representation
of angles on the unit circle.
We define a metric on 8 1 by letting d[oo,,6] be the length of the
shortest arc on the circle from a to,6. More precisely, if a and ,6
92 9. Symbolic Dynamics and Chaos

are in the interval [0,211"), then

d[a, 11] = {Ia - 111, if la - 111 ::; 11"

la - 111- 11", if la - 111> 11".
Define the doubling function D : 8 1 ---t 8 1 by D(O) = 20.
a) Show that d is a metric for 8 1 .

b) Show that if a = 211"2k where k and n are natural numbers,

n-1
then a is periodic with period n. Conclude that periodic points of
D are dense in 8 1 .
c) Let (a, 11) = {O in 8 1 1 a < 0 < 11} be any interval on 8 1 . Prove
that there exists a natural number n such that D n ((a, 11)) = 8 1 .
Conclude that D is topologically transitive on 8 1 .
d) Use Theorem 9.20 to conclude that D is chaotic on 8 1 .
e) Show that in every open interval on 8 1 there exist points Xl and
X2 and an integer n such that d[D n (X1), D n (X2)] = 11".
Hint: Let Xl be of the form ~:.
f) Use part (e) to prove that D exhibits sensitive dependence on
initial conditions. In particular, demonstrate that for every point
X on 8 1 , there exists a point y in 8 1 and natural number n so that
d[Dn(x),Dn(y)] ~ ~.

9.13 • THE TENT MAP REVISITED: Determine whether or not the tent
map defined by

T(x) = {2X' for Xin [O,~]

2 - 2x, for X in [~, 1]

is chaotic.
Hint: It may help to review the work from exercise 8.7.

9.14 AN INVESTIGATION: Let T be the set of all sequences in E2 which

contain no consecutive 1'so
a) Show a(T) = T.
b) How many periodic points does a have in T?
c) Are the periodic points of a dense in T?
d) Is a : T ---t T chaotic?
 Exercise Set 9 93

9.15 A COMPUTER INVESTIGATION: Use a computer program that al-

lows the user to control the the amount of rounding in its calcu-
lations to empirically test whether or not h: [0,1] --t [0,1] defined
by h(x) = 4x(1 - x) exhibits sensitive dependence on initial con-
ditions. An example of such a Mathematica program can be found
in the appendix.
10
The Logistic Function,
Part II:
Topological Conjugacy

We return now to the analysis of the logistic function. Our goal is to prove
that if r > 2 + yI5, then hr(x) = rx(1- x) is chaotic on A. We recall that
A is the set of all numbers in [0,1] which remain in [0,1] under iteration
of h. That is, A = {x I hn(x) is in [0,1] for all n}. By Theorem 9.20, it is
sufficient to show that the periodic points of h are dense in A and that h
is topologically transitive on A. We have already shown that the periodic
points of h are dense on A in exercise 8.4. Unfortunately, proving that h
is topologically transitive on A directly from the definition is a relatively
difficult task. Consequently, we will show instead that the dynamics of h
on A are the same as the dynamics of a on L: 2 • Mathematically speaking,
we say that h on A is topologically conjugate to a on L: 2 .

DEFINITION 10.1. Let f : D ---+ D and g : E ---+ E be functions. Then

f is topologically conjugate to g if there is a homeomorphism r : D ---+ E
such that r 0 f = g 0 r. In this case, r is called a topological conjugacy.

We represent this relationship by the commutative diagrams

f
x ~ f(x)

or TilT
E~E r(x) ~ g(r(x)).
g g
96 10. The Logistic Function, Part II: Topological Conjugacy

The diagrams imply that if x is an element of D, then T(f(X)) = g(T(X)).

To use a diagram let x be an element of D. If we follow the arrows from the
upper left corner over and down, then we get T(f (x)). Following them down
and over we get g(T(X)) which is the same point we arrived by following
the arrows on the other route since the diagrams are commutative. Hence,
g(T(X)) = T(f(X)). Also, since T is a homeomorphism, we know that T is
one-to-one and T- l is defined and continuous. Thus, T(Y) = g(T(X)) if and
only if y = T-l(g(T(X))). Setting y = f(x) we see that T(f(X)) = g(T(X))
implies T-l(g(T(X))) = f(x). We represent this by

x y

or
E ----7 E T(X) ----7 g(T(X)).
9 9

Reading the diagrams we see that y = f(x) = T-l(g(T(X))). Similarly, if a

is an element of E, then T(f(T-l(a))) = g(a) or

f T-l(a) f f(T-l(a))
D ----7 D ----7

r-11 Ir or r-11 Ir
E ----7 E a ----7 g(a).
9 9

Returning now to the definition of a homeomorphism we recall that

c.p: D---> E is a homeomorphism if and only if c.p is continuous, one-to-one,
onto, and has a continuous inverse. Further, we note that if D and E are
homeomorphic, (Le., there exists a homeomorphism from D to E), then the
topologies of D and E are the same. We make this precise in the following
proposition.
PROPOSITION 10.2. Let D and E be metric spaces and c.p : D ---> E be a
homeomorphism. Then
a) the set U in D is open if and only if the set c.p(D) is open in E,
b) the sequence Xl, X2, X3, .. , in D converges to x in D if and only if
the sequence c.p(Xl), c.p(X2), c.p(X3) , ... converges to c.p(x) in E,
c) the set F is closed in D if and only if the set c.p(F) is closed in E,
d) the set A is dense in D if and only if the set c.p(A) is dense in E.
The proof of this proposition follows immediately from the relevant def-
initions and should be completed by the reader. It may be useful to use
some of Propositions 9.5 through 9.10.
 10. The Logistic Function, Part II: Topological Conjugacy 97

When functions of D and E are topologically conjugate, the existence

of the homeomorphism from D to E guarantees that the topologies of the
two spaces are identical. The condition that T 0 f = gOT guarantees that
the dynamics are the same. This is demonstrated in Theorem 10.3.

THEOREM 10.3. Let D and E be metric spaces, f : D -; D, g : E -; E,

and T : D -; E be a topological conjugacy of f and g. Then,
a) T- 1 : E -; D is a topological conjugacy,
b) T 0 fn = gn 0 T for all natural numbers n,
c) p is a periodic point of f if and only if T(p) is a periodic point of
g. Further, the prime periods of p and T(p) are identical.
d) If p is a periodic point of f with stable set WS(p), then the stable
set ofT(p) is T(WS(p)).
e) The periodic points of f are dense in D if and only if the periodic
points of g are dense in E.
f) f is topologically transitive on D if and only if g is topologically
transitive on E.
g) f is chaotic on D if and only if g is chaotic on E.
Before we begin the proof of Theorem 10.3 it is worth noting that many
students of mathematics (both professional and those with amateur status)
find it useful to use commutative diagrams to keep track of arguments
involving topological conjugacies.
Let us consider for example the proof of part (f) of this theorem. Suppose
f is topologically transitive on D. We wish to show that g is topologically
transitive on E. Given an open set U in E (lower left corner of the diagram),
we want to find an x in U so that for some natural number n, gn(x) lands
in the open set V in E (lower right corner).

UeE VeE

Now we consider T-1(U) (upper-left corner) and T- 1(V) (upper-right cor-

ner). By Proposition 9.10 both of these sets are open in D. Hence, since
f is topologically transitive on U, there is y in T-1(U) and a natural
number n such that r(y) is in T-1(V). We let x = T(y) (in lower-left
corner). Then following around the diagram, up, across, down, we see
that gn(x) = gn(T(Y)) = T(r(y)) by part (b) of Theorem 10.3. Thus,
gn(x) = T(r(y)) which is in V since r(y) is in T-1(V).
98 10. The Logistic Function, Part II: Topological Conjugacy

With a little practice this diagram chasing becomes second nature and an
easy way to complete proofs involving commuting functions like topological
conjugacies.
PROOF OF THEOREM 10.3:
a) It suffices to show r- 1 0 9 = for-I. The details are left to the reader.
b) To prove that r 0 fn = gn 0 r we need only consider the following
commutative diagram consisting of n squares. As each of the squares com-
mutes, the whole diagram commutes. Thus, we get the same result by
going around the outside of the diagram in either direction and the truth
of the assertion follows.

D~D~D~ ~D

E----t E----tE----t ... ----tE

9 9 9 9

Alternatively, for those with a more analytical bent, we may use the
Principle of Mathematical Induction. The statement r 0 fn = gn 0 r is true
when n = 1 since r is a topological conjugacy. Now suppose it is true for
n, that is, r 0 fn = gn 0 r. Then it is true for n + 1 since
r 0 r+ 1 = (r 0 r) 0 f = (gn 0 r) 0 f = gn 0 (r 0 f) = gn 0 (g 0 r) = gn+1 0 r.
c) Let p be a periodic point of f with prime period k. We first show
gk(r(p)) = r(p). Since fk(p) = p, we get the commutative diagram

p p

Therefore, gk(r(p)) = r(p) and r(p) is a period k point of g. If 0 < n < k,

then we see that gn(r(p)) = r(r(p)) i= r(p) since r is one-to-one and p
has prime period k under f. Hence, r(p) has prime period k under g.
Since r- 1 is a topological conjugacy the same argument demonstrates
that if q is a periodic point of 9 with prime period k, then r- 1 (q) is a
periodic point of f with prime period k.
d) Let p be a periodic point of f with prime period k and x be an element
of W S (p). (It may help to keep track of the argument using a commutative
diagram.) Then by definition of WS(p), for each 0 > 0 there is N so that
if n 2: Nand dD is the metric on D, then dD[fkn(x),p] < o. We must
 10. The Logistic Function, Part II: Topological Conjugacy 99

show that for each E > 0, there is M such that whenever n ~ M, then
dE [gkn(r(x)), r(p)] < E.
Suppose E > O. Since r is continuous, there exists 8 > 0 so that if
dD[Y,p] < 8, then dE[r(y), r(p)] < E. Choose M so that if n ~ M, then
dD[fkn(x),p] < 8. Then by continuity,

dE[r(jkn(x)),r(p)] = dE[gkn(r(x)),r(p)] < E

when n ~ M.
e) The preservation of density of periodic points by a topological conju-
gacy follows easily from part (c) of this proposition and part (d) of Propo-
sition 10.2.
f) The argument for this portion is outlined above.
g) This portion follows from parts (e) and (f) and Theorem 9.20. 0
We complete this chapter by constructing a topological conjugacy from
hr : A -+ A to a : E2 -+ E2 when r > 2 +.J5. As a is chaotic on E 2 ,
Theorem 10.3 then implies that h is chaotic on A.
We begin by defining the function 'I/J : A -+ E 2 • Set 10 = [0, ~ - v'r~;4r]
and h = [~ + jr~;4r, 1]. Recall from Proposition 8.1 that

A1 = {x I h(x) is in [0, I]}

= IoUh·

Since A is a subset of A1 , A is also included in IoUh. For each x in A, define

the sequence 'I/J(x) = SOSlS2'" in E2 so that hn(x) is in Isn for each n. For
example , if r -
- ,5 then h O(l-v'5) -- 1-v'5 is in L0, h 1(1-v'5) = 1 is in I 1,
e-
2 2 2
and h k 2v'5) = 0 is in 10 for all k ~ 2. So 'l/Je-
2v'5) = 010000 .... We can
think of 'I/J(x) as the itinerary of x. The digit Sn in 'I/J(x) = SOSlS2 ... is 0 if
and only if hn(x) is in 10 and Sn = 1 if and only if hn(x) is in h. Clearly,
'I/J is well-defined. It remains to show that 'I/J is a topological conjugacy.

THEOREM 10.4. The function 'I/J : A -+ E2 defined above is a topological

conjugacy. That is,
a) 'I/J is one to one and onto,
b) 'I/J is continuous,
c) 'I/J-1 is continuous, and
d) 'I/J 0 h= a 0 'I/J.

PROOF. We begin by defining subsets of An which we call ISQsls2 ... Sn'

We then prove each of the parts of Theorem 10.4 in turn.
100 10. The Logistic Function, Part II: Topological Conjugacy

Let h, 10, and h be defined as above and let SOS2S2." be an element

of E 2 . We claim that the set
Isos182 ... Sn = {x I hk(x) is in 18k for all k ~ n}
is one of the intervals in An+! where
An +l = {x I hn+!(x) is in [0, I]}.
We note that this definition of An is the same as the one we used in Chapter
8. We prove the claim by induction.
By Proposition 8.1, Iso is either 10 or II' Since Al = 10 UIl the claim is
true in this case. Suppose Isosl ... Sn_l is one of the intervals in An. We will
show this implies Is081",Sn is one of the intervals in An+!' Note first that
Isosl",Sn is a subset of Isosl",Sn_l so it suffices to determine which portions
of Isosl ... 8n_l belong to Isosl",Sn'
Let Isosl",Sn_l = [a, b]. Then by Proposition 8.1 hn([a, b]) = [0,1] and
hn is monotone on [a, b]. We assume hn is increasing on [a, b]; the proof
is similar when h n is decreasing. By the Intermediate Value Theorem,
there exist Cl and C2 so that a < Cl < C2 < b, hn(Cl) = ~ - v'r~;4r, and
h n (C2) = ~ + v'r~;4r. Consequently,

and

So, if Sn = 0, then Isosl ... Sn = [a, Cl] and if Sn = 1, then Isosl",Sn = [C2, b].
As this is the same way we derived the intervals of An+! in the proof of
Proposition 8.1, we see that Isosl ... Sn is one of the intervals comprising
A n +l .
a) Let S = SOSlS2S3 ... be an element of E 2 . To show that 'lj; is one to
one and onto we must show that 'lj;-l(S) contains exactly one point. But if
x is in 'lj;-l(s), then x is in Isosl ... Sn for all n. Thus,

'lj;-l(s) = n
00

n=O
Isosl",Sn'

We need to show this intersection is not empty and contains exactly one
point. Let Isosl",Sn = [an, bn]. Suppose 'lj;-l(s) contains two points x and
y. Then Ix-yl ~ Ibn -ani for all n since x and yare in each of the intervals
 10. The Logistic Function, Part II: Topological Conjugacy 101

[an, bn ]. However, Lemma 8.6 implies that Ian - bnl approaches 0 as n goes
to infinity. So it must be that Ix - yl = 0 and x = y. Consequently, we see
that 'lj;-I(S) contains no more than one point.
It remains to show that 'lj;-I(s) contains at least one point. This follows
immediately from the well known Nested Interval Theorem of mathematical
analysis which states that the intersection of a nested set of closed intervals
is not empty. (See Theorem 2.38 of W. Rudin's Principles of Mathematical
Analysis, 3rd edition.) Recall that the intervals 1SQS1, .. Sn are denoted by
[an, bn ]. The proof ofthe Nested Interval Theorem depends on the principle
that there exists a unique real number which is larger than or equal to
every an and smaller than any other number which is also greater than
all the an. Such a number is called the least upper bound of the set
{al,a2, ... ,an , ... }. Now since an ::; am ::; bm whenever m ~ nand
an ::; bn ::; bm whenever m ::; n, we see that each bm is greater than
every an. So the least upper bound, a, is less than or equal to every bm .
Therefore, an ::; a ::; bn for each n and a is in every [an, bn] = 1SQS1 ... Sn' It
follows that a is in 'lj;-I(s) and we have completed the proof of part (a).
b) Let E > 0 and x be in A. To show 'lj; is continuous at x we must find
8> 0 so that if Ix - yl < 8, then d['lj;(x), 'lj;(y)] < E. If we choose n so that
2~ < E, then Lemma 9.11 implies that it is sufficient to show there exists
a 8 > 0 so that the sequences 'lj;(x) and 'lj;(y) agree on the first n + 1 digits
whenever Ix - yl < 8. In other words, if'lj;(x) = SOSIS2 .. " then x is in
1SQ81S2",Sn and we need to show that y is in1sQs1s2",sn whenever y is in A
and Ix - yl < 8.
Let [a b b1], [a2,b 2], ... , [a2n'H,b2n+1] be the 2n+1 intervals in An+l
!
indexed so that bi - 1 < ai for all i. Set 8 = min{lai - bi - 1 1}. Since the
intervals are not overlapping and finite in number, 8 is positive. Also, if
x and yare in An+l and Ix - yl < 8, then it follows that x and yare in
the same interval of An+1. As 1SQS182 ... 8n is an interval in An+1 and x is in
18Q8182 ... 8n' we see that y must be in 18Q8182 ... 8n and the proof of part (b) is
complete.
c) The proof that 'lj;-1 is continuous is,1eft as an exercise.
d) It remains to show 'lj;oh = ao'lj;. Let x be in A and 'lj;(x) = SOSIS2 ....
As we demonstrated in the proof of part (a), x is the unique point in

18Q8182'" = n00

n=O
18Q8182 ... 8n =
n=O
n
00

{x I hk(x) is in 18k for all k ::; n}.

It follows that hex) is the unique point in

n n
00 00

1818283 .• ' = 18182 ... 8n = {x I hk+1(x) is in 18k for all k ::; n}.
n=1 n=O
102 10. The Logistic Function, Part II: Topological Conjugacy

So 'ljJ(h(x)) = SlS2S3··· = (J('ljJ(x)) and we are done. D

COROLLARY 10.5. Let r > 2 + V5, h(x) = rx(l - x), and

A = {x I hn(x) is in [0,1] for all n}.

Then h is chaotic on A.

The proof follows from Theorems 10.3 and 10.4 and Proposition 9.15.

Exercise Set 10

10.1 Prove Proposition 10.2.

10.2 Complete the details of the proof of Theorem 10.3.

10.3 Let 7 : D ~ E be a topological conjugacy of the maps f : D ~ D

and g : E ~ E. Show that 7- 1 : E ~ D is a topological conjugacy
of g to f.

10.4 • THE QUADRATIC MAP qc(x) = x 2 + c:

a) Let f(x) = Ax 2 + Ex + C be a quadratic function. Show that
there exists a linear function 7(X) = mx + d and a parameter c so
that 7 is a topological conjugacy between f and qc(x) = x 2 + c.
Hint: Simplify the compositions 7 0 f and qc 0 7 and equate
coefficients of like terms.
b) Show that if c ::::: 1/4, then qc (x) = x 2 + c is topologically
conjugate to h(x) = rx(l - x) for some r. Use the topological
conjugacy to prove that qc has a single attracting fixed point when
-3/4 < c < 1/4. What happens when c < -3/4? What are the
dynamics of qc when c > 1/4?
Determine for which parameter values qc is chaotic on a Cantor
set. You may assume hr is chaotic on a Cantor set whenever r > 4.

10.5 • THE TENT MAP: THE ADVENTURE CONTINUES. Define the tent
map by
T(x) = {2Xl for x in [O,~]
2 - 2x, for x in [~, 1].
a) Show that the map 7(X) = ~ cos(7r(l-x)) + ~ = cos2(~(1-x))
is a topological conjugacy between the map h(x) = 4x(1 - x) on
[0, 1] and the tent map.
 Exercise Set 10 103

b) Use part (a) of this exercise, exercise 9.13, and exercise 8.7 to
show that h: [0,1]-+ [0,1] is chaotic.

10.6 Show that q(x) = x 2 - 2 is chaotic on the interval [0,2].

Hint: You may wish to use the two previous exercises.

10.7 Prove that 'ljJ-l as defined in Theorem 10.4 is continuous.

10.8 Determine the arrangement of the eight intervals ISQsls2 in [0,1].

10.9 ** Prove h(x) = rx(l - x) is chaotic on A when 4 < r ~ 2 + J5.

11
The Logistic Function,
Part III

In Chapter 7 we characterized the behavior of the parametrized family of

functions hr(x) = rx(l- x) for 0 < r S 3. By this time, the reader should
find it easy to verify that h has, at most, two periodic points for these
parameter values, both of which are fixed. All of the other points in lR.
are in the stable set of one of these points, or the stable set of infinity. In
Chapter 10 we demonstrated that there is a Cantor set in [0,1) on which h
is chaotic when r > 2 + /5. All points of lR. not in the Cantor set are in
the stable set of infinity. We stated without proof that this property also
holds for 4 < r S 2 + /5. Finally, we asked the reader to show that h is
chaotic on [0,1] when r = 4 in exercise 10.5. The analysis of the behavior
of h for parameter values between 3 and 4 remains.
In Example 7.5 we demonstrated that hr undergoes a period doubling
bifurcation when r = 3. The fixed point Pr = r~l changes from being an
attracting fixed point when r < 3 to a repelling fixed point when r > 3.
The bifurcation is called period doubling because when the change occurs
an attracting period 2 orbit is added which straddles Pr. We can visualize
the attracting fixed point as splitting into a period 2 attracting orbit with
a repelling fixed point lodged between the two points in the orbit.
To get an idea of what happens as we continue to increase r we use a
theorem of Pierre Fatou, which was discovered shortly after the turn of the
century.
106 11. The Logistic Function, Part III

THEOREM 11.1. [FATOU] If the quadratic polynomial f(x) = ax 2+bx+c

has an attractive periodic orbit, then the critical point '2:
is in the stable
set of one of the points in the orbit.

We recall from calculus that Xo is a critical point of the differentiable

function f if f'(xo) = O. Note that the theorem does not state that there
is a critical point in the stable set of each attracting periodic point. As
there is only one critical point, that would be impossible. What it does say
is that there is a critical point in the stable set of one of the iterates of an
attracting periodic point. As an easy corollary to Theorem 11.1 we see that
a quadratic polynomial can have at most one attracting periodic orbit. We
will not prove Theorem 11.1 since the proof requires a substantial amount of
material from complex analysis. A proof can be found in the references. In
particular, the reader might wish to consult the second edition of Devaney's
text, Introduction to Chaotic Dynamical Systems.
We note that the critical point of hr(x) = rx(l- x) is !.Consequently,
if hr has an attracting orbit for some r, then we know that ! is in the
stable set of the orbit. We exploit this information to draw a bifurcation
diagram for hr. In doing so, the parameter r is graphed on the horizontal
axis, and the value of h~ (!) is plotted on the vertical axis for values of n
between 100 and 500. For example, when r = 3.8, h~OO(!) = .83668 and
h~Ol(!) = .51924, so the points (3.8, .83668) and (3.8, .51924) are plotted.
We continue by plotting (3.8, h~02(!)), (3.8, h~03(!)), ... , (3.8, h~OO(!)).
This is done for a large number of r values between 2 and 4 and the resulting
graph is shown in Figure 11.1. Since an attracting periodic orbit attracts
the critical point, the diagram should give us a picture of the location of
the attracting orbits.
Note that we get the behavior we expect for 2 < r < 3. After 100
iterations the points have more or less converged to the attracting fixed
point Pr = r;l. When r = 3 there is a period doubling bifurcation and
the orbit of the critical point is attracted to a period 2 orbit when r is a
little larger than 3. It appears that another period doubling bifurcation
occurs at a point somewhat less than r = 3.5. At that point, the period
2 orbit splits into an attracting period 4 orbit, followed by another split
into period 8, another into period 16, and so on. While an analytic proof
that this is what happens is beyond the scope of this book, a convincing
geometric argument is readily accessible.
Consider again the graphs shown in Figure 7.6 on page 64. Note that
as the parameter value passes three, the derivative of h: at the fixed point
passes from being less than one, to being greater than one. Recall that the
fixed point is Pr = r; 1 . When r > 3, the fact that (h: )(Pr) > 1 implies
there is a point Xo to the right of the fixed point satisfying h~(xo) > Xo.
Since h:(l) = 0, the Intermediate Value Theorem implies that there is a
 11. The Logistic Function, Part III 107

o·
2
.
3 4
r-values

FIGURE 11.1. A bifurcation diagram ofhr(x) =rx(l-x)

generated by graphing the 100th through 500th iterates
of ~. The parameter value r is plotted on the horizontal
axis and the x value is plotted on the vertical axis. The
parameter value is allowed to vary from 2 to 4. The range
of x values shown is from 0 to 1.

0.8

FIGURE 11.2. The graph of h~.4 and h 2 .646 . Notice the

similarity of the graphs in the boxed regions.
108 11. The Logistic Function, Part III

fixed point of h~ in the interval (xo, 1). This point is a prime period 2
point. This was proven in exercise 7.2.
Now consider the graph of h~(x) for r = 3.4 shown in Figure 11.2. In
particular, notice the boxed region. If we rotate the graph of h~ on the
interval over which the box sits, then it closely resembles the graph of hs
in the unit square when s = 2.646. By the unit square we mean the box
with corners at (0,0), (0,1), (1,1), and (1,0).
We can draw a square box like the one on the graph of h~ in Figure 11.2
for all parameter values r satisfying 2 < r :s 4. In each case the box has
corners at (!r'r!) ' (!r' r-l)
r '
(r-l r-l)
r' r '
and (r-l
r 'r
!). Note that when r = 2
the upper-right corner of the box coincides with the lower-left corner and
the box is no longer defined. If h~(!) = 2 r (i;;r)is greater than ~, then h2
maps the interval Ir = [~, r;l] to itself. Solving numerically, we see this
occurs when r is less than approximately 3.678.
In Figure 11.3 we consider the graphs of h~(x) for r = 3.4, r = 3.45,
and r = 3.5. In each case the graph is shown on the interval I r . Note that
the interval changes as we change r, but the behavior is identical to that
which we observed when we looked at the graph of h~(x) for r near 3 in
Example 7.6; a period doubling bifurcation occurs. If p is the period two
point, then (h~)'(P) < 1 when r < 3.45. If r = 3.45, then (h~)'(P) ~ 1 and
when r > 3.45, a new attracting period 4 point has appeared.
Looking at the graph of h~ for r = 3.5, we again notice square boxes
in which the graph of h~ is similar to the graph of hs in the unit square.
These boxes are identified in Figure 11.4. This process can be continued
indefinitely. If hr has a periodic point with prime period n, then we can
find regions on the graph of h~ around which we can draw a square box so
that the graph of h~ in the box is similar to the graph of hs on the unit
square. As r changes we will find all of the same dynamics for h~ within
these boxes that we find for hs on the unit square. In particular, since
hs has a period doubling bifurcation when s = 3, h~ will have a period
doubling bifurcation for some value of r.
Recall that h~ : Ir -+ Ir where Ir = [~, r;l] whenever r < 3.678. When
r > 3.679, the graph of h r "sticks out" of the box. That is, there are points
in Ir which are not mapped back into I r . This situation is illustrated in
Figure 11.5. Note that the graph of h~ on Ir for r > 3.679 is similar to the
graph of hs on [0,1] for s > 4 with one significant difference; points which
leave the interval Ir under iteration of h~ when3.679 < r < 4 may return
to the interval, whereas points which leave [0,1] under iteration of hs are
asymptotic to infinity. In Figure 11.5 we show via graphical analysis of h~
that though! leaves Ir when r = 3.8, it returns within a few iterations.
Recall from Chapter 8 that if s > 4, then hs is chaotic on a Can-
tor set which is contained in [0,1]. A similar result holds for h~ on Ir
 11. The Logistic Function, Part III 109

4----.....>..,~--~ period 4 points

0.3
~0~.3-----0-.4----~0.-5-----0.~6----0~.7-

FIGURE 11.3. Graphs of h;

for the parameter values 3.4,
3.45, and 3.5. The point p is a prime period 2 point of
h which undergoes a period doubling bifurcation when r
is approximately 3.45. In each, case the bounds of the
graphed region are ~ and r~ 1 .
110 11. The Logistic Function, Part III

0.3
0.3 0.4 0.5 .6 0.7

FIGURE 11.4. The graph of h;

for the parameter value
r =
3.5 shown over the interval [~, r~l] . Note the boxes
in which the graph is similar to that of the logistic function
on [0,1].

O. 0.8

0.6

0.4

0.2

0 0.2 0.4 0.6 O. 0 0.2 0.4 0.6 O.

FIGURE 11.5. The graph of h;

when r = 3.8. Note that
points in Ir may leave the interval under iteration of h; and
then return. This is illustrated using graphical analysis
beginning at the point ~.
 Exercise Set 11 111

when r > 3.679. There is a Cantor set in Ir on which h~ is chaotic. As

a result we expect that the dynamics for parameter values in the interval
[3.679,4] to be extremely complicated and they are. In fact, the dynamics
of hr for parameters in this interval are not fully understood. A bifurca-
tion diagram for this interval is shown in Figure 11.6. Note the apparent
period 5 and period 3 orbits. If we investigate a region around any of these
orbits we discover another bifurcation diagram apparently identical to the
larger one we first examined in Figure 11.1. An enlargement of a region
surrounding a period 3 point is shown in Figure 11.6.
The reader is encouraged to use a computer to explore other regions of
the bifurcation diagram. With a little effort it should be possible to find a
series of bifurcation diagrams, one contained within the other, all of which
resemble Figure 11.1.
We end this chapter by noting that the dynamics of hr are already quite
complicated by the time r = 3.678. While Figure 11.6 makes it clear that
the dynamics are complex by the time r > 3.679, the sequence of period
doubling bifurcations has deteriorated into chaos long before then. In fact,
the entire period doubling cascade will have finished by the time r = 3.57
and there will be a periodic point with prime period 2n for n whenever
r ~ 3.57. Two techniques for demonstrating this are indicated in the
exercises.

Exercise Set 11

11.1 Let hr = rx(1 - x) and Ir = [~, r-;:l]. Show h~ : Ir ---. Ir when

2 < r < 3.67.

11.2 Use a computer graphics package to demonstrate geometrically

that there is another period doubling bifurcation from period 4
to period 8 inside the boxes shown in Figure 11.3. Explain why
your graphs imply there is a such a bifurcation. At what parameter
value does the bifurcation occur? Identify boxes on the graph of
h 8 within which you would expect another period doubling bifur-
cation.

11.3 Use a graphics package to investigate the properties of h~. Demon-

strate graphically an attracting prime period 3 point of hr. For
which parameter values is there an attracting prime period 3 point?
Find a box containing an attracting prime period 3 point in which
the graph of h~ is similar to the graph of hs on the unit square for
some s. Does a period doubling bifurcation occur in this box? If
112 11. The Logistic Function, Part III

r-values 4

.6

FIGURE 11.6. The top diagram is a bifurcation diagram

for hr in which r varies from 3.679 to 4. Note the verti-
cal strips of white space. These strips correspond to at-
tracting periodic points. Two of the most obvious occur
at parameter values for which there are period 5 and pe-
riod 3 attracting points. A box is drawn about a portion
of the diagram in which there is a period three periodic
point. The bottom diagram is an enlargement of this re-
gion. Note the similarity of this region to the bifurcation
diagram shown in Figure 11.1. Given our discussion, the
reader should be able to explain why this similarity is to
be expected.
 Exercise Set 11 113

so, what is the period of the new attracting fixed point? At what
parameter value does it appear? Support your statements with
graphs and explanations.

11.4 a) Draw a graph of h; for a variety of parameter values in the

range 2 ::; r ::; 4. How does the graph of h;change as r increases?
Explain why your answer is correct. Use the results to argue that
if hr has a periodic point of prime period two when r = ro, then
it has a periodic point with prime period two whenever r :;::: roo

b) Repeat part (a) for h;'.

c) Generalize the statements in part (a) to h~. Explain why the
result holds.

d) Use part (c), Sarkovskii's Theorem, and a bifurcation diagram

to demonstrate that the cascade of period doubling bifurcations
must have been completed before r = 3.575. That is, show there
are orbits with prime period 2n for all n whenever r :;::: 3.575. Can
you do the same for r :;::: 3.57?

11.5 • THE QUADRATIC MAP REVISITED: Let qc(x) = x 2 + c. Use

exercise 10.4 to show that the only periodic points of q are a single
attracting fixed point and a repelling fixed point when -3/4 < c <
1/4. However, when C < -3/4 we see a cascade of period doubling
bifurcations followed by chaos.

11.6 FEIGENBAUM'S CONSTANT: Let qc(x) = x 2 + c.

a) Let Cn be the parameter value at which the period 2n attracting
periodic point of qc splits into a period 2n +1 attracting periodic
point. Estimate en for n ::; 6.

b) Let dn = Cn - Cn-l and use the estimates of Cn found in part (a)

Cn+l - Cn
to approximate dn for n ::; 5. If you did the calculations carefully
you should find that the sequence of dn's appears to converge.
(Mitchell Feigenbaum showed that it converges to approximately
4.699 in 1978. This value is now called Feigenbaum's constant.)

c) Use Feigenbaum's constant, the values Cn, and the fact that
00 1
"'" xn = - - to estimate the value of r at which the period dou-
L I-x
n=O
bling cascade will have been completed.
114 11. The Logistic Function, Part III

11.7 The amazing aspect of Feigenbaum's constant is that it is universal.

By this we mean that if Sc{x) is a parametrized family of functions
which undergoes a period doubling cascade in the same manner as
the logistic function, Cn is the parameter value at which the bifur-
cation from period 2n to period 2n +! occurs, and dn = Cn - Cn-l,
Cn+! - Cn
then lim d n = Feigenbaum's constant.
n->oo
Show numerically that this is true for the family of functions
Sc{x) = csinx on the interval [0,11']. In creating the bifurcation
diagram assume that if Sc{x) has an attracting periodic point,
then a critical value of Sc{x) must be in the stable set of one of its
iterates. Recall that the critical value of a function is a point at
which the derivative is zero.

11.8 Use a computer to find a sequence of 10 bifurcation diagrams for

hr, one contained inside the other, each of which is similar to the
diagram shown in Figure 11.1.

11.9 What causes the shadow curves seen in the bifurcation diagrams
of Figure 11.6?
Note: An interesting discussion of the shadow curves can be
found in section 11.4 of the book Fractals and Chaos by Peitgen,
Jurgens, and Saupe, which is listed in the references.
12
Newton's Method

The search for solutions of the equation f (x) = 0 is ancient and methods
that can solve equations of the form ax2 + bx + c = 0 are several thousand
years old. In the 16th century, Italian mathematicians discovered methods
for solving third and fourth degree polynomials. However, it was shown in
the early part of the nineteenth century that there is no general method
for solving polynomials of degree five or higher. Consequently, methods for
estimating solutions of equations as simple as polynomials are necessary.
Isaac Newton developed such a method, which was later refined by Joseph
Raphson, and which we now know as Newton's method or the Newton-
Raphson Method. Newton's method is easy to use and is often taught in
first semester calculus since it only requires knowledge of the derivative.
To illustrate Newton's method, suppose f (x) is differentiable and that Xo
is a reasonable approximation of a solution to the equation f(x) = O. How
can we find a better approximation? We begin by drawing the line tangent
to the graph of f (x) at the point (xo, f (xo) ). The slope of this line is f' (xo).
We label the point of intersection of the tangent line and the x-axis (Xl, 0)
and take Xl as our next approximation. Since f is differentiable, its graph
is reasonably straight on small intervals and if Xo is close to the root, then
Xl is a better approximation than Xo. This is illustrated in Figure 12.l.

In Figure 12.1 we see that our new estimate Xl is a significant improve-

ment over our initial estimate of the solution, Xo. As long as the function
doesn't curve too much between the first estimate and the solution this
116 12. Newton's Method

FIGURE 12.1. An illustration of Newton's method.

FIGURE 12.2. Two iterations of Newton's method.

is the case. If we apply Newton's method again we get an even better

approximation as shown in Figure 12.2. In fact, when Newton's method
converges to a root, it does so at a rate which is proportional to the square
of the error. The number of correct digits to the right of the decimal point
approximately doubles at each iteration.
To develop an analytic understanding of Newton's method, suppose that
we have a function f(x), know its derivative f'(x), and have Xo as an initial
estimate of a root. To find a second estimate Xl in terms of Xo we find an
equation of the tangent line and solve for X when y = O. As (xo, f(xo)) is
a point on the line and the slope is f' (xo), the equation of the line is

(y - f(xo)) = f'(xo)(x - xo).

 12. Newton's Method 117

Setting y = 0 and x = Xl and solving for Xl we find that

J(xo)
Xl = Xo - J/(xo)·

Now if we let Xl be our estimate and apply the process again to find another
estimate X2, then we see that

We can find another estimate and then another by repeatedly applying

Newton's method. How can we represent this as an iterated function?
J(x)
Define Nf(x) by Nf(x) = X - J/(x). If Xo is our initial estimate for
Newton's method, then

Xl = Nf(xo)
X2 = Nf(XI) = Nf(Nf(xo)) = NJ(xo)

and, in general, if Xn is the nth estimate, then

Xn = N7(xo).
To simplify notation when the function in question is clear, we will suppress
the subscript and write Nf(x) as N(x). We call Nf(x) Newton's function
for J.

EXAMPLE 12.1.
Let p(x) = x3 - x. The roots of pare -1, 0, and 1. Let's see how fast
we can find them using Newton's method. We begin by finding N(x).

N( ) _ _ p(x) _ _ x 3 - X _ 2x 3
X - X p/ (x) - X 3x2 _ 1 - 3x2 _ 1

If we estimate the root as .25 we get the sequence of estimates

Xo = .25
Xl = -.0385
X2 = .000114
X3 = -.00000000000299
118 12. Newton's Method

The sequence converges quickly to the nearest root. If our initial value is
.75 we get

Xo = .75
Xl = 1.227
X2 = 1.051
X3 = 1.003

Again the sequences converge quickly to the nearest root.

Now suppose we try an initial value of .45. Iteration of Newton's method
gives us the following sequence of values.

Xo = .45
Xl = -.464
X2 = .567
X3 = -10.156
X4 = -6.793
X5 = -4.562
X6 = -3.091
X7 = -2.135
Xs = -1.536
Xg = -1.192
XlO = -1.038
Xn = -1.002

It converges, but the convergence is not as quick and it is to the root

which is furthest away. This illustrates one of the problems with Newton's
method. While it often finds a root, it may not find the closest one.
It is also worth noting that the values given for the previous calculations
were rounded after the calculation. The actual calculations were done with
16 significant digits. Given our previous work we should not be too sur-
prised to find out that one needs to be careful when rounding as small
errors can make a difference. For example, if we had started the last cal-
culation at .46 rather than .45, then the sequence would have converged to
1. Similarly, were we to have rounded Xl up to -.47 the sequence would
again converge to 1. To get some idea of the fragility of these calculation
the reader is encouraged to try other initial values for this function. Values
between .4472 and .46 should prove to be especially interesting.
Finally, consider the initial value ~. Then Xl = N(~) = 2/~..J3,
which, of course, is undefined. What happened? If we look at the graph of
 12. Newton's Method 119

0.5

-1

FIGURE 12.3. Newton's method applied to p(X) = X3 - x

with the initial value Xo = )s. Note that p' ( )s) is zero,
so the tangent line is horizontal and we can't find Xl.

p(X) = X3 - X (Figure 12.3) we see that the tangent line at ()s,p()s)) is

horizontal. Consequently, the tangent line never intersects the x-axis and
we can't find Xl. This manifests itself in Newton's method as a point for
which N(x) = X - ;~~) is undefined since p'(x) is zero there. We shall see
later that this difficulty can be avoided for polynomials. 0
The unpredictable behavior of Newton's method in the preceding exam-
ple for initial values near .45 leads one to ask what else might go wrong.
The answer is plenty. We will see that even cubic polynomials can have
regions which not only take a long time to converge, but which never con-
verge to a root at all. However, we will also outline in Theorem 12.15 a
procedure for determining intervals where the convergence is quick and to
the nearest root and we show in Theorem 12.2 that a number is an attract-
ing fixed point of Np(x) for the polynomial p if and only if it is a root of
the polynomial.
THEOREM 12.2. Suppose p(x) is a polynomial. If we allow cancellation
of common factors in the expression of Np(x), then Np(x) is always defined
at roots of p(x), a number is a fixed point of Np(x) if and only if it is a
root of the polynomial, and all fixed points of Np(x) are attracting.

The problem we wish to avoid with the first conclusion in Theorem 12.2
is illustrated by the function, f(x) = x 2 - 2x + 1. The only root of f is
1 and f'(l) = O. As Nf(x) =x - J/~~), Nf(l) is not defined since there
is a zero in the denominator. More generally, if r is a root of p( x) and
120 12. Newton's Method

p'er) = 0, then Np(r) = r - ~

which is undefined. However, we see in the
proof of the theorem that when the derivative at a root is 0, the numerator
and denominator of the fraction ;~;) have common factors, which can be
cancelled. The function Np(x) is defined at the root after cancellation of
these factors.

PROOF. Let p(x) be a polynomial and suppose r is a root of p, that is,

per)= 0. We recall from algebra that there is a natural number n and a
polynomial q such that p(x) = (x - r)nq(x) and q(r) =1= 0. Hence,

p(x)
Np(x) = x - pl(X)
(x - r)nq(x)
= x-
(x - r)nql(x) + n(x - r)n-lq(x)
-;----:----:7-:---"-..,.:...:.~-:---:-....,......,._

(x - r)nq(x)
= x - -;----:---:;-'-;-~---'-,....::-:.~-____;~
(x - r)n-l((x - r)ql(x) + nq(x))
(x - r)q(x)
= x - -----'--,--....,.....:,-=-:......:..--,---,-
(x - r)ql(x) + nq(x)
Thus, after cancelling the term (x - r)n-l we find that Np(r) is defined
and
N(r)=r-
p
(r-r)q(r)
(r - r)q'(r) + nq(r)
=r-
0+ nq(r)
=r °
°
since q(r) =1= by assumption. Therefore, Np(x) is defined at the root and
has a fixed point there.
On the other hand, if Np(r) = r, then

per)
r - p'er) = r

where :,~~) is assumed to be in reduced form. This implies ;~;) = °

or
per)
= 0. Hence, fixed points of Np are roots of p.
Further,

I (p' (x))2 - p(x )p" (x) p(x )p" (x)

Np(x) = 1 - (p' (x))2 = (pl(x))2 . (12.1)

Thus, when r is a root of p and p'er) =1= 0,

and INp(r) I =
I Ip(r)p"(r) I
((p' (r))2 = °< 1.
 12.1. Newton's Method for Quadratic Functions 121

Therefore, r is an attracting fixed point of p. A similar argument holds

when r is a root of p and p'(r) = O. 0
We should note that in the proof of Theorem 12.2 we only used the fact
that the function was a polynomial to guarantee that Np(x) was defined
at the root. The same proof demonstrates that when r is a root of the
function f and Nf(r) is defined, then Nf(r) = rand Nj(r) = O. Fixed
points whose derivatives are zero are often called super attracting fixed
points. We investigate the significance of this designation in the exercises.
As a corollary to the proof of Theorem 12.2 we get
COROLLARY 12.3. If f is a differentiable function, then a fixed point of
Nf(x) must be an attracting fixed point of Nf and a root of f. On the
other hand, if a point is a root of f and Nf is defined at that point, then
the point is an attracting fixed point of Nf.

12.1. Newton's Method for Quadratic Functions

Before we look at some of the strange behaviors associated with Newton's
method, let's look at a case where things go right. Let f(x) = ax 2 +bx+c
where a =f. O. Then

N (x) =x _ ax 2 + bx + c = ax2 - c. (12.2)

f 2ax + b 2ax + b
In the next proposition we show that if we are interested in analyzing the
dynamics of Newton's method, then we may assume a = 1 and b = O.
PROPOSITION 12.4. Let f(x) = ax 2 + bx + c and q(x) = x 2 - A where
2
A = (b -4ac). Then r(x) = 2ax+b is a topological conjugacy from Nf(x)
to Nq(x).
PROOF. In exercise 2.17, we proved that all nonconstant linear func-
tions are homeomorphisms. Thus, we may assume that r(x) = 2ax + b
is a homeomorphism. It remains to show that r 0 N f = N q 0 r. From
ax 2 - c x2 + A
equation (12.2) we have Nf(x) = band Nq(x) = - 2 - ' Then
2ax+ x
roN (x)
f
= 2a (ax2 -
2ax + b
c)
+ b = 2a2x 2 + 2abx + b2 -
2ax + b
2ac

and
Nor = (2ax + b)2 + A = 4a 2x 2 + 4abx + b2 + A
q 2(2ax+b) 2(2ax+b)
4a x + 4abx + b + b - 4ac
2 2 2 2 2a 2x 2 + 2abx + b2 - 2ac
2(2ax+b) 2ax+b
122 12. Newton's Method

FIGURE 12.4. Graphical analysis for Newton's method

applied to p(x) = X 2 - 1 with .2 and -.2 used as starting
points.

Thus, TO Nf = Ng 0 T and the proof is complete. 0

As a consequence of Proposition 12.4 we may restrict our analysis of

Newton's method for quadratic polynomials to polynomials of the form
q(x) = x 2 - c. Obviously, when c > 0 the solutions of q(x) = 0 are ±VC,
when c = 0 the only solution is 0 and when c < 0 there are no real solutions.
We begin our analysis of the case c > 0 with an example.

EXAMPLE 12.5.
Let p(x) = x 2 - 1. From Theorem 12.2 we know that -1 and 1 are the
only fixed points of N(x). Since p'(O) = 0, N(O) is undefined. We will show
that WS(1) = (0, (0) and W S( -1) = (-00,0). Graphical analysis of N(x)
justifies this claim though of course it doesn't prove it. The graph of N (x)
with graphical analysis done for two initial points is shown in Figure 12.4.
We begin by proving the claim for WS(1) . We note that if Xo is in
x2 + 1
(0 , 00) , then N(xo) = _0_ _ is in [1,(0) since this is equivalent to saying
2xo
that x6 - 2xo + 1 2:: O. It remains to show that lim Nn(xd
n-+oo
= 1 whenever
 12.1. Newton's Method for Quadratic Functions 123

, q(x)q"(x) X2 - 1 . ..
From equation (12.1), N (x) = (q'(X))2 =~. DIfferentIatmg
1
again we find that N"(x) = 3' Since N"(x) > 0 when x is in the interval
x
(1, (0) we know that N' (x) is increasing on this interval. As N' (1) = 0 and

x 2 -1 1
lim N'(x) = x->oo
lim - - 2 -
x->oo 2x 2'
this implies that 0 < N'(x) < ~ for all x in (1, (0). The Mean Value
Theorem implies that there is CI between 1 and Xl so that

IN(xd - 11 = IN(xI) - N(I)1 = IN'(cdllxI - 11 < ~lxI -11-

By induction we can find Cn between Nn-I(Xl) and 1 so that

INn(xl) - 11 =IN(Nn-1(xd) - N(N n- I (I))1

= If'(cn)IINn-l(xd - N n- I (I)1
< (~) (~r-l IXI - 11 = (~r IXI -11-
As the last term in the previous inequality approaches 0 as n tends to
infinity, it follows that if Xl > 1, then Nn(xd converges to 1 as n tends to
infinity. Therefore, if Xo > 0, then Nn(xo) = Nn-l(xd converges to 1 as
n tends to infinity as well. Thus, we have shown that (0,00) C WS(I).
The proof that (-00,0) C W S ( -1) is similar. As these two relationships
account for all the points in the domain of N, the result WS(I) = (0, (0)
and WS( -1) = (-00,0) follows. D

The argument in the preceding example suggests the following propo-

sition. The proof of the proposition is left as an exercise. Two possible
approaches to the proof are suggested in exercise 12.18.

PROPOSITION 12.6. If q(x) x 2 - c2 and c > 0, then Nq(x) has fixed

=
points at ±c, both of which are attracting. Further, the stable set of 0 is
the interval (0, (0) and the stable set of -c is the interval (-00,0).

The case c = 0 is easier. In this case, our function is q(x) = x 2 and

N(x) = x - ~: = ~x. Demonstrating that 0 is the only fixed point of N(x)
and that the stable set of 0 is the set of all real numbers is a straightforward
exercise.
Now we turn to the more interesting case, Newton's method for the
function q(x) = x 2 + c when c > O. Of course q(x) doesn't have any real
roots and, as the only points which can be attracting fixed points of N are
roots of q, we expect something different to happen. It is easiest to first
consider the special case c = 1.
124 12. Newton's Method

EXAMPLE 12.7.
x 2 -1
Let p(x) = x 2 + 1. Then N(x) = - - . To get some idea of what
2x
might happen for this function we look at graphical analysis for N(x) with
initial point .1 in Figure 12.5.
Obviously this isn't as simple as the cases we have examined already.
Looking at Figure 12.5 we are led to wonder if the action of this function
might, in fact, be chaotic. We prove that this is the case by showing that
N is related to the doubling map of the circle.
Recall from exercise 9.12 that the circle is denoted 8 1 and is usually
depicted as the unit circle. Angles or points on the circle are measured
in radians in the counterclockwise direction with the initial point on the
positive x-axis. The doubling map, D : 8 1 ~ 81, is defined by D(()) = 2().
We demonstrated in exercise 9.12 that D was chaotic on 8 1 .
We define a map from the circle with the point 0 deleted to the real
numbers by ¢(x) = cot(~). This map satisfies No ¢ = ¢ 0 D, since by the
half angle identities,

cot 2 (:£) -1 cos 2 (:£) - sin2 (:£)

(No¢)(x)- 2 _ 2 2
- 2 cot (~) - 2 cos (~) sin (~)
cos(x)
=-.-(-) = cot (x) = (¢oD)(x).
sm x

Note that ¢ is not defined at 0 and so is not a homeomorphism of 8 1

and~. Hence, it is not a topological conjugacy. We leave as an exercise
the verification of the fact that cot ( ~) is continuous, one-to-one, and onto
when consider as a map from the circle with the point 0 deleted to the
real numbers. We also note that (¢ 0 D)(7f) = ¢(O), which is not defined.
However, (N 0 ¢)(7f) = N(O), which is not defined either since 1'(0) = o.
Now working backwards we see that D2 (7f /2) = D2 (37f /2) = 0 and ¢ 0 D2
is not defined at 7f/2 or 37f/2. On the other hand, both (N 2 0 ¢)(7f/2) and
(N 2 0 ¢)(37f/2) reduce to N(O), which is not defined either. Iterating this
procedure we can show that rn is a point on 8 1 for which Dn(rn) = 0 and
(¢ 0 Dn)(rn) is not defined if and only if Nn(¢(rn)) is not defined. So, ¢
always maps points in 8 1 that are not mapped to 0 by some iterate of D
to points in ~ for which Nn is defined for all natural numbers n. Also,
if "f is a point in 8 1 that is not mapped to 0 by an iterate of D, then
(¢ 0 Dn)("() = (N n 0 ¢)("() for all natural numbers n.
By Theorem 9.20, it suffices to show that periodic points of N are dense
in ~ and that N is topologically transitive on R to demonstrate that N is
chaotic on R Let (a, b) be an interval in R. Then ¢-1((a,b)) = (0'.,{3) is
an interval in 8 1 . By exercise 9.12b, there is "f in (0'., {3) and k such that
Dk ("() = "f. Without loss of generality, we can assume that Dn ("() =f. 0 for
 12.1. Newton's Method for Quadratic Functions 125

L
/
/
./
./

/ -'"

3 -2 2 3
./ V
./
/'

~V
-~

[7
-3
FIGURE 12.5. Graphical analysis for 100 iterations of
Newton's function for p(x) = x 2 + 1 with .1 used as a
starting point.

any n. (Why?) Hence, if c = ¢b), then c is in (a , b) and

Nk(c) = Nk(¢b)) = ¢(Dkb)) = ¢b) = c.

Thus, c is a periodic point of N in (a, b) and the periodic points of N are
dense in R Also, we demonstrated in exercise 9.12c that there exists m
such that D m (( cd3)) = 8 1 . It follows that Nm (( a, b)) = R Therefore, if
(d, e) is any other interval in JR, then there is Xo in (a, b) such that Nm(xo)
is in (d, e) and N is topologically transitive.
A slightly different proof that N is chaotic on JR is outlined in exercise
12.16, and a proof using complex analysis is outlined in Example 14.21. D
Now suppose that q(x) = x 2 +c2 , C -=I- 0, and p(x) = x 2 + 1. By showing
that Nq and Np are topologically conjugate we are able to prove

PROPOSITION 12.8. If q(x) = x2 + c2 , then N q is chaotic on the set

(-00, 0) U (0, (0).

The proof of Proposition 12.8 is left as an exercise. In Chapter 14 we

investigate the dynamics of Newton's method when applied to complex
quadratic functions.
126 12. Newton's Method

12.2. Newton's Method for Cubic Functions

In this section we will find that the behavior of Newton's method for
a cubic polynomial is much richer than its quadratic cousin. As with the
quadratic we can simplify the analysis by reducing to cases where the be-
havior of N varies with a single parameter.

PROPOSITION 12.9. If f(x) = ax 3 +bx 2 +ex+d andg(x) = x 3 +Ax+B

where A = 9ac - 3b2 and B = 27a 2 d + 2b3 - 9abc, then the function
T(X) = 3ax + b is a topological conjugacy from Nf(x) to Ng(x). That is,
ToNf = N g OT.

The proof of Proposition 12.9 is an exercise in algebra and is left to

the reader. More interesting questions are: Why should we expect there
to be a linear function that conjugates N f and N g ? What motivates the
choice of T( x) = 3ax + b as the conjugacy? The answer to the second
question is that T maps the inflection point of f to the inflection point
of g. If we can understand why that is the proper point to consider, we
will understand the answer to the first question. It is really a matter of
translation and scaling. The placement of the y-axis relative to the graph
of the function is irrelevant to the behavior of Newton's method for the
function. Hence, we can translate horizontally. Furthermore, if we divide
f(x) = ax 3 + bx2 + ex + d by a to get p(x) = x 3 + ~X2 + ~x + ~ we find
that N f = N p • In other words, we can rescale f by a. We leave the proof
of this fact as an exercise.

PROPOSITION 12.10. Let f(x) = x 3 + ax + b3 where b =I- 0 and define

g(x) = x 3 + + 1 where c = {is. Then N f and N g are topologically
cx
conjugate via the homeomorphism 7(X) = ix.

Proposition 12.10 coupled with Proposition 12.9 implies that either the
dynamics of Newton's method for a cubic polynomial are the same as the
dynamics of Newton's method for a polynomial in the parametrized family
fc(x) = x 3 + ex + 1, or the dynamics are the same as those of Newton's
method for a polynomial of the form ga(x) = x 3 +ax. The following propo-
sition implies that Newton's method for functions of the latter form must be
topologically conjugate to Newton's method for one of three polynomials:

or

PROPOSITION 12.11. If 7(X) = ~x where a =I- 0, g(x) = x 3 + a2x, and

p+(x) = +x, then N g and Np+ are topologically conjugate via 7. On the
x3
other hand, if f(x) = x 3 - a 2 x and p_(x) = x 3 - x, then 7 is a topological
conjugacy between N g and Np_ .
 12.2. Newton's Method for Cubic Functions 127

Using this proposition and the preceding discussion we conclude that

for all cubic polynomials p(x), Np is topologically conjugate to Nf where
f(x) = x 3 + cx + 1 or Np is topologically conjugate to Newton's method
for one of the three polynomials, p+(x) = x 3 + X, p_(x) = x 3 - x, or
Po (x) = x 3 . The analysis of Np+ and Npo is easy and is left as an exercise.
The results are summarized in the following proposition.

PROPOSITION 12.12. Let p+(x) = x 3 + x and Po(x) = x 3 • Then 0 is the

only fixed point of both Np+ and N po ' and the stable set of 0 is lR for both
functions.

The behavior of Newton's method for p_(x) = x 3 - x is more com-

plicated. We considered some of the issues in Example 12.1. Additional
interesting behavior is examined in exercise 12.5 at the end of this chapter.
We turn now to Newton's method for f(x) = x 3 + cx + 1. In this case
2x 3 -1
N(x) =..,....--,,--
3x 2 + c·
When c > 0 it is easy to verify that N has exactly one root and that the
stable set of the root contains all of R When c = 0 there is still one real
root and the stable set contains all real numbers except O.
Now consider the case c < O. We note that the graph of f is symmetric
about the point (0,1), has a relative minimum at x = J -c13 and a relative
maximum at x = - J -cI3. We label the x value at the minimum point
x m . The graph of f is shown in Figure 12.6 along with the graph of N for
two values of c.
We note that as c decreases without bound, the relative minimum of
f decreases and the relative maximum increases. When c = _~(21/3),
or approximately -1.8898, the relative minimum is 0 and f has two real
roots, one at 0 and the other at x = x m . We denote the point _~(21/3)
by co. Note that when c is less than co, f has three real roots. When c
is greater than co, f has one real root. Looking at Figure 12.6, we are
led to ask what happens to Newton's method when c is between 0 and co,
say for example c = -1. If we start Newton's method at 0, then the next
value is larger than x m . The value after that seems to be smaller than Xm
and then it becomes hard to discern. Iterating on a computer we find that
when c = -1 we get the sequence

Xo = 0, Xl = 1, X2 = .5, X3= 3, X4 = 2.038, X5 = 1.39,

X6 = .912, X7 = .345, Xs = 1.428, Xg = .942, XlO = .405.
It doesn't look like this sequence will converge to the root anytime soon.
In Figure 12.7 we look at the first 100 orbits of 0 for a variety of parameter
values between -2 and o.
128 12. Newton 's Method

y=x

.\~

FIGURE 12.6. The graphs of f(x) = x 3 + ex + 1 and N f

are shown in the first row for the parameter value e = -l.
In the second row we again see f and N f' this time for
the parameter value e = -3.2. The vertical lines in the
graph of N f are the asymptotes created at the points where
f'(x) = o.
 12.2. Newton's Method for Cubic Functions 129

. . . ,. :'fi"..
2
....~'. .. ... :.l;... .'
.. ."' ......;.
1.5 .' 0(' - '"!.
~·I·.;~
:;:.,...~ -:. ~~:
'"

0.5

0
- :C
'.
I :..
.
-0.5
..
~

, '.
" .1

-1

-1. 5

-2
-1.75-1.5-1.25 -1 -0.75-0.5-0.25 0

FIGURE 12.7. The points (c, N~(O)) are plotted where Nc

is Newton's function for fc(x) = x 3 + CX + 1. We use 250
parameter values of c between -2 and 0, and n varies from
o to 100 for each c.
Note that 0 seems to converge to a positive root when such a root exists,
that is, when c < Co ~ -1.89. The dark line running slightly upward across
the bottom of the diagram when c is larger than Co represents the negative
root so we see that on a fairly regular basis if c > Co , then 0 is attracted to
the negative root within 100 iterations of N c . If 0 is not attracted quickly
to one of the roots, then there does not seem to be any apparent pattern to
its behavior under iteration of N c . If we look at the 500th through 600th
iterates of 0 we get a clearer picture, since by that time the iterates should
be near an attracting orbit. The results of such a computation are shown
in Figure 12.8. Again , there is no obvious pattern in the behavior of Nc at
those parameter values for which 0 does not converge to a root , though a
large percentage of the orbits do seem to be converging to a root.
Looking at Figure 12.8 we are led to wonder what is happening for
those parameter values where Newton's method isn't converging to the
root quickly. We begin searching for an answer to this by investigating N
for c = -1.265. This value is chosen since the orbit of 0 tends to clump at
two places, one near 0 and the other near .8.
130 12. Newton's Method

1.5
. ...
1
. .. .. • I •

0.5

o . I t • I •

-------
-0.5

-1

-1.5
-
-2~---1~.~7-5---1~.-5---1~.-2~5---~1----0~.~7~5---0~.5----0~.2~5--~0

FIGURE 12.8. The 500th through 600th points in the or-

bit of zero for parameter values between -2 and o. The
parameter is graphed on the horizontal axis and 500 val-
ues are used between -2 and o.
 12.2. Newton's Method for Cubic Functions 131

0.025

-0.05

FIGURE 12.9. The graph of NJ(x) where f(x) = x 3 -

1.265x + 1. Note that the graph of NJ crosses the line
y = x near the point (-.016, -.016). Note also that the
slope of Nt at the intersection is negative and greater than
-1. A light line with slope -1 is drawn through the point
of intersection.

EXAMPLE 12.13.
2x 3 - 1
Let f(x) = x 3 -1.265x + 1. Then N(x) = 3x 2 - 26' When iterating
1. 5
N starting at 0, we get the following sequence:

Xo = 0, Xl = .79051, X2 = -.019675, X3 = .79125,

X4 = -.015043, X5 = .79094, X6 = -.016973, X7 = .79106,
Xs = -.016232, Xg = .79101, XlO = -.016527.

We seem to be converging to an attracting periodic point with prime period

two!
To investigate further we look at the graph of N 2 (x) in a neighborhood
around O. If there is an attracting periodic point with period 2 in this
neighborhood, then the graph of N 2 (x) should intersect the line y = x and
the slope of the graph of N 2 (x) should be less than one in absolute value
at the point of intersection.
While Figure 12.9 presents powerful evidence that there is an attracting
period two point of N near - .016, it is a fact that should still be proven.
To do this, we look for an interval J such that the intersection of N (J)
132 12. Newton's Method

and I is empty, N 2 : I _ I, and I(N 2)'(x)1 < 1 for all x in I. The first
condition ensures that I contains no fixed point of N, and Theorem 6.3,
along with the last two conditions, guarantees that N has an attracting
periodic point with prime period 2 in I. Figure 12.9 suggests that the
interval we are looking for might be [-.03, .03]. Our goal then is to show
that 1= [-.03, .03] satisfies the following three properties:
(1) N(I)nI=0
(2) N2 : I _ I
(3) I(N 2 )'(x) I < 1 for all x in I
. , f(x)f" (x)
In equatIOn (12.1) we calculated that Nf(x) = (f'(x))2 . Hence, for
_ 3 , _ (x 3 - 1.265x + 1)(6x) .
f(x) - x - 1.265x + 1 we have N (x) - (3x 2 _ 1.265)2 . DIfferen-
tiating again and simplifying we find that

Nil ( ) = 6(2(1.265)x3 - 9x 2 + 2(1.265)2X - 1.265)

X (3X2 _ 1.265)3
6x 2[2.53x - 9] + 6(1.265)[2.53x - 1]
(12.3)
(3X2 - 1.265)3
For x in the interval I = [-.03, .03] we note that both the numerator and
the denominator in N"(x) are negative. Consequently, N"(x) is positive
and N' is increasing on the interval. That is,

- .12 < N'( -.03) :::; N'(x) :::; N'(O) = 0 (12.4)

for x in [-.03,0] and

0= N'(O) :::; N'(x) :::; N'(.03) < .11 (12.5)

for x in [0, .03]. Hence, N is decreasing on [-.03,0] and increasing on
[0, .03]. Since N(-.03) ~ .7922, N(O) ~ .7905, and N(.03) ~ .7922 this
implies that N(I) is contained in the interval [.79, .793]. Clearly, N(I)nI =
o and we have demonstrated condition 1.
Using equation (12.3) again we find that N"(x) < 0 on [.79, .793]. Thus
N' is decreasing on the interval [.79, .793], and if x is in [.79, .793], then

6.4 > N'(.79) ~ N'(x) ~ N'(.793) > 6.1. (12.6)

This implies that N is increasing on [.79, .793]. Since N(.79) ~ -.0229
and N(.793) ~ -.00426, we conclude that N([.79, .793]) C [-.03, .03] and
N 2 : I _ I as required by condition 2.
Finally, we wish to estimate (N 2 )'(x) for x in I. From exercise 6.2 we
know that (N 2)'(x) = N'(x)N'(N(x)). Since inequalities (12.4) and (12.5)
 12.2. Newton's Method for Cubic Functions 133

0.2

0.1

-0 . 1

-0 . 2

-1.29 -1.28 -1.27 -1.26 -1.25

FIGURE 12.10. The 500th through 650th points in the

orbit of zero for parameter values between -1.3 and -1.25.
The parameter is graphed on the horizontal axis and 500
values are used.

imply IN'(x)1 < .12 when x in I and inequality (12.6) implies IN'(x)1 < 6.4
when x is in [.79, .793], we have
I(N 2 )'(x)1 = IN'(x)IIN'(N(x))1 < (.12)(6.4) = .768 < 1
as required by condition 3.
We have shown that N has an attracting periodic point with prime
period 2 in [-.03, .03]. We leave as an exercise the proof that [-.03, .03] is
contained in the stable set of this periodic point. D

In the preceding example we found that when p(x) = x 3 - 1.265x + 1,

there are intervals on which N not only doesn't converge to a root , but
converges to an attracting periodic point instead. Given the size of the
clumps in Figure 12.8 for parameter values near -1.265, we are led to
wonder if we won't find something more interesting than a period 2 point.
In Figure 12.10 we investigate by plotting an enlargement of the lower of
the two clumps.
Amazingly we find the familiar bifurcation diagram of Figure 11.1. In
the exercises we will search for an explanation of this phenomenon. The
134 12. Newton's Method

diagram indicates that N 2 goes through a cascade of period doubling bi-

furcations as the parameter decreases from -1.25 to -1.3. It also indicates
that there is something more going on. Recall that in the case of the logis-
tic equation, the cascade of period doubling bifurcations was followed by
chaos. So, for the cubic case, we are led to the conjecture,

CONJECTURE 12.14. If f(x) = x 3 -cx+1, then there exists c > 0 and a

set of real numbers on which Nt is chaotic. Further, this set is attracting.

We explore this conjecture in the exercises.

12.3. Intervals and Rates of Convergence

The results of the previous section on Newton's method for cubic polyno-
mials should have us a bit concerned by now. After all we have shown that
if we do not choose our initial guess wisely, then not only might the method
not converge, but it might wander periodically, or perhaps even chaotically,
forever. Fortunately, there are tests to determine whether or not our initial
guess will converge to the closest root. Theorem 12.15 outlines one such
criterion. One might also wonder why we would choose to use a method
which is subject to problems like those demonstrated above. The answer is
speed. Newton's method exhibits quadratic convergence. Essentially this
means that the number of correct digits to the right of the decimal place
in our estimate doubles with each iteration of Newton's method. Conse-
quently, if we can find an approximation which is reasonably close to the
root, then we can get a very good approximation of the root with only
a few steps. By contrast, the bisection method which is also commonly
taught in calculus or precalculus classes, converges geometrically. That is,
it takes approximately 3 iterations to add one more correct digit to the
approximation. The rate of convergence of Newton's method is made more
precise in Theorem 12.15.

THEOREM 12.15. Suppose that f(x) is a twice differentiable function

and f' (xo) is not equal to o. Let

E =I f,~:~)) I ' and

If 1f"(x)1 < 1f'(xo)12 fOT all x in I, then Nj(xo) converges to a TOot of f

2If(xo)1
in I and
 Exercise Set 12 135

As one might guess, Theorem 12.15 is proven by induction on the number

of iterations. While the proof requires no mathematical concepts other
than induction and first-year calculus, a proper treatment of it requires
several pages. As inclusion of the proof here will not serve to increase our
understanding of dynamical systems, we refer the interested reader to the
text by Ostrowskii which is listed in the references.

Exercise Set 12

12.1 For each of the following functions use Newton's method to find as
many roots of the equation as you can. Describe the stable set of
each root. Indicate any points you find which are not in the stable
set of any root and explain why they aren't in the stable set of a
root.
a) f(x) = x2 - 3x + 2
b) g(x) = log x
c) s(x) = sec x
d) E(x) = eX
1
e) k(x) = - - 1
x
f) p( x) = X4 + x2
g) r(x) = ..;x
x
h) F(x) = ~1
x +
a
12.2 Draw the graph of function for which Newton's method converges
to infinity whenever the initial value is greater than 2 and for which
it converges to 0 whenever the initial value is less than 2. Find a
formula for such a function. Prove that your formula works; that
is, show that the stable sets of zero and infinity are as desired.

12.3 Investigate Newton's method for f(x) = x 2 • Find the unique fixed
point and prove that its stable set is R

12.4 If p is a fixed point of the function f, then p is said to be super

attracting if f'(p) = O. Prove that if p is a super attracting fixed
point of f, then for each E > 0 there is 8 > 0 so that if Ip - yl < 8,
136 12. Newton's Method

then If(y) - pi < €Ip - yl· Explain why this property suggests the
name super attracting.

12.5 NEWTON'S METHOD FOR p(x) = x 3 x: Suppose that p(x) =

-
2x 3
x 3 - x. In Example 12.1 we demonstrated that N(x) = 2 •
3x -1
As we noted in that example, N (x) is undefined for x = ± ~.

a) Show that WS(O) = (- ]g, Jg). What happens to ]g?

b) Find a repelling periodic point of N with period 2.
c) Show that there exists a sequence of points ao, a1, a2, ... such
that ao = ~, ai > aHl for all natural numbers i, Ni(ai) = ~ if
i is even and Ni(ai) = - ~ if i is odd.
. 1
Show hm ai = ~.
i-+co v5
What happens to Newton's method if our initial point is one of
the numbers ai?
If we numerically solve for a2, and use a computer to apply
Newton's method to this numerical solution, then it converges to
one of the roots -lor 1. Why don't we see the behavior that our
analysis predicts?
Demonstrate the existence of a similar sequence bo, b1 , b2 , . . • ,
which begins at - ~ and converges to - ]g.
d) Show that (aH 1, ai) is in WS (1) whenever i is odd and in
W S (-l) whenever i is even. State and prove a similar fact for
the intervals (b i , bi +1).
e) Summarize by classifying all of the points in IR as to their be-
havior under iteration of N.

h
12.6 A N I NVESTIGATION: ConSl·der t h e sequence di = ai-l - ai were
ai - aHl
the ai are as defined in exercise 12.5. Use a computer to determine
numerically that d i converges. To what number does it converge?
Why is this so? Can you prove that your answer is correct?

12.7 Let p+(x) = x 3 + x and Po(x) = x 3 . Show the only fixed point
of Newton's method for both functions is 0 and that WS(O) is R
(This is Proposition 12.12.)

12.8 Let f(x) = x 3 + ex + 1 where e > O. Show that Nf has exactly one
fixed point and that the stable set of the fixed point contains all
 Exercise Set 12 137

real numbers. Show that when e = 0 there is still only one fixed
point and its stable set contains all real numbers except O.

12.9 Let f(x) = x 3 -1.265x + 1. Show that the interval [-.03, .03) is in
the stable set of an attracting periodic point of N f .
Hint: Use the results of Example 12.13.

12.10 In Example 12.13 we stated that if N(x) has an attracting periodic

point of period two in an interval, then the graph of N (x) will cross
the line y = x in the interval and that IN'(x)1 < 1 at the point of
intersection. Explain why this is so.

12.11 Let f(x) = x 3 + ex + 1. Find a value of the parameter e for which

Nf has an attracting periodic point with prime period 4.
Hint: Use the bifurcation diagram in Figure 12.10.

12.12 a) Figure 12.10 indicates that Newton's method for the function
f(x) = x 3 + ex + 1 goes through a cascade of period doubling
bifurcations as e varies from -1.25 to -1.3. Use a sequence of
graphs of N 2 for parameter values in this range to explain why
this is to be expected. Why do we use the graph of N 2 instead of
N?
b) Using the information from part (a) and our experience with
period doubling cascades in Chapter 11, predict a parameter value
and an interval on which we would expect N 2 to be chaotic. Ex-
plain why this is a good choice. Can you find numerical or graphical
evidence to support your guess? Can you prove that your guess is
correct?
c) If N 2 is chaotic on I, does it follow that N is chaotic on IUN(I)?
Explain.

12.13 Let f(x) = x 3 + ex + 1. Find a value of e for which N f has an

attracting periodic point with prime period 3.
Hint: Examination of Figure 12.7 should yield some likely can-
didates. Note that in Example 12.13 we found a period two point
for e = -1.265 at which value we see two "blobs" in the bifurcation
diagram.

12.14 * Let f(x) = x 3 + ex + 1. Find a value of e for which there is an

interval I in which Nf has periodic points of all prime periods.
138 12. Newton's Method

12.15 Let ¢ : 8 1 \ {O} -+ lR be as defined in Example 12.7.

a) Show that ¢ is one-to-one, onto, and continuous.
b) Show that if "( is a point in 8 1 that is not mapped to 0 by an
iterate of D, then (¢oDn)("() = (Nno¢)("() for all natural numbers
n.
c) Let (a, b) be an interval in lR and (a, f3) be ¢-l((a, b)). Suppose
Dm((a, f3)) = 8 1 . Prove that Nm((a, b)) = lR.

12.16 Let 8* be the set points of 8 1 which are not eventually fixed at 0
by the doubling map D((}) = 2(}. That is,

8* = {p in 8 1 I Dn(p) i= 0 for any n}.

a) Show that D : 8* -+ 8* and that D is chaotic on 8*.

x 2 -1
b) Let N(x) = -- be Newton's method for f(x) = x 2 + 1 and
2x
let
lR* = {x in lR I Nn(x) is defined for all n}.

That is, lR* is the set of points which are never mapped to 0 by N.
Show ¢ : 8* -+ lR * defined by ¢ (x) = cot ( ~) is a homeomorphism.
Hint: It may help to use the previous exercise and the discussion
in Example 12.7.
c) Show that D : 8* -+ 8* and N : lR* -+ lR* are topologically
conjugate. Conclude that N is chaotic on lR*.
d) Use part (c) to show that N is chaotic on lR.

12.17 Show that Theorem 12.15 implies that the number of correct digits
to the right of the decimal place approximately doubles with each
iteration of Newton's method.

12.18 a) Prove Proposition 12.6 by revising the argument in Example

12.5.
b) Prove Proposition 12.6 by showing that Newton's function for
f(x) = x 2 - 1 is topologically conjugate to Newton's function for
q(x) = x2 - c2 •
Hint: Use a linear function that maps the fixed points of Nf to
the fixed points ofNg as the topological conjugacy.
 Exercise Set 12 139

12.19 Prove Proposition 12.8 by demonstrating that Newton's function

for f(x) = x 2 + 1 is topologically conjugate to Newton's function
for q(x) = x 2 + c2 •

12.20 Prove that Newton's methods for f(x) = ax 3 + bx 2 + ex + d and

p( x) = x 3 + ~ x 2 + ~ X + ~ are topologically conjugate. Explain why
we should not be surprised by this.

12.21 Let P(x) be a polynomial and Q(x) = P(x + a) where a is a

real number. Explain why we should expect N p and NQ to be
topologically conjugate without explicitly calculating N p and N Q.
You may find it useful to draw representative graphs of P and Q.
What should the conjugacy be?

12.22 AN OPEN-ENDED PROBLEM: Let q(x) = ax 4 +bx3 +cx2 +dx+e.

Find a fourth degree polynomial with fewer parameters for which
Newton's method is conjugate to N q • What is the fewest number
of parameters necessary? Let p(x) = (x 2 + A)(x2 + B). For which
values of a, b, c, d, and e can we find A and B so that N q and Np
are topologically conjugate? Investigate the behavior of N p • What
types of behavior do you find?
13
Numerical Solutions of
Differential Equations

The reader is probably familiar with the differential equation y' = ky from
calculus. Often it is introduced as a simple model for population growth
or accumulating interest. For example, if p(x) represents the population of
bacteria in a laboratory sample at time x and the population increases by
fixed percentage during each time interval, then it is reasonable to assume
that there is k > 0 so that p' = kp: the rate of change of the size of the
population, as reflected by the derivative, is proportional to the population.
The larger the population, the more individuals that are added to it during
each time interval. If we wish a more sophisticated model that takes into
account the fact that the crowding of a large population may effect growth
negatively, then we might reflect this by adding a quadratic term to get
w' = kw - aw 2 where both k and a are positive. When the population is
small, kw is greater than aw 2 , the derivative is positive, and the population
is growing. As w increases the difference between kw and aw 2 decreases
until we reach a point where aw 2 is larger than kw, the derivative is negative
and the population is shrinking.
Both of these equations are solvable by techniques we learned in calculus.
If ~~ = kp and p(O) = Po, then we can separate the variables to find that
~ dp = k dx. Now we integrate both sides and solve for p.
p
142 13. Numerical Solutions of Differential Equations

log(p) = kx +c
P = Ge kx (13.1)
where G = Po.

. J
Similarly, if we assume that w(O) 2': 0, w(O) =1= ~, separate the variables,
and mtegrate k 2 1 dw, then we can show that
w -aw
kG
w (x) - ----:c---,.- (13.2)
- aG + e- kx

is a solution of Wi = kw - aw 2 where G = k-:~~O). If w(O) = kia, then

the constant function p(x) = k/a is a solution.
However, not all differential equations are as easy to solve as the two
discussed above. It is not hard to write down differential equations that
are difficult or even impossible to solve. This is not surprising given the
difficulties involved with integrating functions, and the fact that not all
differential equations can be solved with integration alone. A variety of
techniques have been developed to estimate the solutions of differential
equations that cannot be solved by other means. The easiest of these is
Euler's method, which we illustrate with the following example.

EXAMPLE 13.1.
Let pi = .5p and suppose p(O) = 1. Without solving the differential
equation, estimate the value of p(5).
If there is a solution to the differential equation, then we know the slope
of the tangent line at each point on its graph. For instance, we know
that (0,1) is on the graph since p(O) = 1 and we know that the slope
of the tangent line at (0,1) is .5(p(O)) = .5. Also, since the function is
differentiable its graph is close to the tangent line for a small interval around
O. The y-coordinate of the point on the tangent line with x-coordinate 1
can be found by adding the product of the slope and the change in x to
p(O). Doing so we find that when x = 1, the y-coordinate on the tangent
line is .5(1) + 1 = 1.5. Thus, we estimate that p(l) is 1.5. We illustrate
this in Figure 13.1.
Now the slope of the tangent to the solution at (1, p(l)) is .5(p(1)) or
approximately .5(1.5) = .75. Using the same procedure as before we obtain
a new x value by adding 1 to the current x value and estimate a new y
value by adding the product of the slope and the change in x to the current
y value. That is, we add (.75)(1) = .75 to 1.5 to find that p(2) ~ 2.25. This
is also illustrated in Figure 13.1.
Iterating the procedure, we estimate the slope of the solution at (2, p(2))
as .5(2.25) = 1.125 and add (1.125)(1) = 1.125 to 2.25 to find p(3) ~ 3.375.
 13. Numerical Solutions of Differential Equations 143

2.5

(2,2.25)

~~"_bY_'75XlIoP(I)
2

1.5
(1,1.5), slope of this line isp'(1) or .75

~ y coordinate is obtained by adding .5 x I to p(O)

slope of this line is p' (0) or .5

0.5

0.5 1.5 2 2.5

FIGURE 13.1. An illustration of the derivation of the first

few points in Euler's approximation of p' = .5p with initial
condition p(O) = 1.

Continuing in this way we find that p(5) ~ 7.59375. A diagram of the

completed process is shown in Figure 13.2. The reader is encouraged to
verify the last two estimates before reading on.
In Figure 13.2, we see that we missed the actual value of p(5) by quite
a wide margin. In fact, p(5) = e2 . 5 ~ 12.18 and our estimate is about 7.59.
It should be apparent that part of the problem is the size of the jumps in
x value. In most cases it is not a good idea to assume that the graph of
the solution is more or less the same as the tangent line over an interval of
length 1. If we add only .1 to the x-value each time, then we must make
50 steps to reach 5 from 0, but we find that our estimate is p(5) ~ 11.47,
which is closer. By decreasing the size of the steps again to .01 we increase
the number of steps to 500, but we find the estimate is now p(5) ~ 12.17,
a much better approximation. 0

Using Example 13.1 as a reference, we describe Euler's method for ap-

proximating the solution of the differential equation p' = f(x,p) where x
is the independent variable and f : JR2 -+ JR. Note that p is a function
of x and f(x,p(x)) is the slope of the line tangent to the solution passing
through the point (x,p(x)). Given the initial point (xo,p(xo)) we develop
an algorithm to estimate the value of p at some point x. For the sake of
convenience, we assume x > xo, though the method works when x < Xo.
144 13. Numerical Solutions of Differential Equations

10

(5.7.59315)

(4.5.0625)

(0. 1)

o 2 3 5

FIGURE 13.2. Estimating a solution of p' = .5p with ini-

tial condition p(O) = 1 using Euler's method with step size
1. The actual solution is shown in gray for comparison.

We divide the interval [xo , xl into n subintervals each of length h = x - Xo

n
and designate the end points of the subintervals in ascending order by xo,
Xl , X2, ... Xn = X. Note that Xi+l = Xi + h for all i between 0 and n .
The number h is called the step size, and moving from Xk to Xk+l in the
approximation process is called a step. Of course, it takes n steps to reach
x from Xo.
Given the value of p(xo) we estimate the value of p(xr) by multiplying
the slope of the tangent line at Xo times h and adding the product to p( xo).
Hence, p(xr) is approximately f(xo,p(xo)) . h + p(xo). This is the same
procedure we used in Example 13.1. We repeat the process to approximate
p(X2) , arriving at P(X2) ~ f(XI ,P(Xl)) . h + P(XI)' In general , we have

or, letting Pk be the approximation of p(Xk) ,

Pk = f(Xk-l ,Pk-l) . h + Pk-l· (13.3)

We continue with this process until we have reached Pn ~ p(x). The first
few steps in the process are illustrated in Figure 13.3.
If we wish to apply the methods of the previous chapters we need to find
a function 9 so that Pn = gn (Po) . Suppose that f is a function of P alone, for
 13. Numerical Solutions of Differential Equations 145

f(x,.p,) ·h
f(xflPo) ·!J

" h II

FIGURE 13.3. An illustration of Euler's method.

example, f(x,p) = .5p, as in Example 13.1. Then setting g(p) = f(p) ·h+p
we can write equation (13.3) as

Thus, in the case where f depends only on p, Euler's method is the itera-
tion of a function of the real numbers. If f depends on both x and p, then
Euler's method can be modeled as the iteration of a function from ]R2 to
]R2. Unfortunately, we haven't the tools necessary to completely analyze
iterated functions of more than one variable. However, we will see in Ex-
ample 13.5 that even in the latter case we can still get some information
from our understanding of iterated maps of the real line.

EXAMPLE 13.2.
Suppose, as in Example 13.1, that p' = .5p, p(O) = Po, and we wish to
approximate p(x) = Pn by using n steps of size h = ~(x). Then

where g(p) = (1 + .5h)p.

To see this we note that from equation (13.3),

Pk = f(Pk-d . h + Pk-l = .5Pk-l h + Pk-l = (1 + .5h)Pk-l.

146 13. Numerical Solutions of Differential Equations

So, if we let g(p) = (1 + .5h)p, then

Pn = (1 + .5h)Pn-l = g(Pn-l)
= g ((1 + .5h)Pn-2) = l(Pn-2)

Recall from calculus that lim (1 + ~)n =e T • So, since h = ~,

n~oo n n

lim gn(po) = lim (1 + .5ht Po = lim (1

h-+O h-+O n-+oo
+ .5~)n
n
Po = e· 5x po.

Therefore, the estimate we get from Euler's method converges to the solu-
tion as the step size approaches 0, and the number of steps goes to infin-
ity. 0

We demonstrated in Example 13.2 that the Euler approximation of the

value of the solution of p' = kp at any point converges to the value of
the actual solution at that point, as the step size converged to zero and
the number of steps went to infinity. Usually, we see a similar behavior:
the Euler approximation converges to the actual value at a fixed point in
the domain as step size decreases. However, we should caution that there
are limits beyond which smaller step sizes are no longer practical. This
is due to the fact that approximations with large numbers of steps must
be mechanized and computers have finite accuracy. To a computer there
are only finitely many numbers between 0 and 1. This limitation intro-
duces possibly fatal errors as the step size becomes smaller. The interested
reader is referred to the text Differential Equations: A Dynamical Systems
Approach by J. Hubbard and B. West for details.
We might also ask whether the Euler approximation for a fixed step
size has the same qualitative behavior as the solution of the differential
equation. By qualitative behavior, we mean the general shape of the so-
lution curve. For example, if the solution is p(x) = Ce- x , then we know
that all solutions tend to 0 as x tends to infinity. Or if the solution is
p(x) = y'X + ~, then we know that the solution tends to y'X as x tends
to infinity. This kind of information can sometimes be more useful than
knowledge of the value of the solution at a specific point. For example,
suppose we know that the population of a given mammal in an ecosystem
is modeled by a differential equation whose general solutions have the form
p(x) = 150sin(ix) + 450~+C where x is in years and p(x) is the number of
individuals in year x. Then we won't be unduly alarmed if the population
drops down to 325 individuals from a high of 575 in a three-year time span
 13. Numerical Solutions of Differential Equations 147

since we know that the qualitative behavior of the solutions is to oscillate

between highs of 600 and and lows of 300 with a period of 8 years.
In the remainder of this chapter we will focus on the qualitative behavior
of Euler approximations.

EXAMPLE 13.3.
Consider pi = kp where k is any real number and p(O) = po. As we
have seen, the actual solution is p(x) = poe kx . Suppose we approximate
the solution by using Euler's method with fixed step size h > O. We
denote successive approximations, Po = p(O), PI ~ p(h), P2 ~ p(2h), ....
Using an analysis similar to that in the preceding example, we find that
Pn = kpn-I h + Pn-I· That is, Pn = (1 + kh)npo or Pn = gn(po) where
g(p) = (1 + kh)p.
We note that the only fixed point of 9 is 0 and that Ig' (p) I = 11 + khl
for all t. If k > 0 we see that 19'(p) I > 1 for all p and that gn(po) tends to
(Xl whenever Po i= 0 and remains constant at 0 when Po = o. Qualitatively

this is the same behavior we see in the solutions of the differential equation;
when k > 0, the solution p(x) = poe kx tends to infinity as x increases and
is identically zero when Po = O.
Now consider k < o. If 0 < h < -~, then g'(p) = 11 + khl < 1. Hence,
when h is in this range, then gn(po) tends 0 for all Po since WS(O) = lEt If
h = - ~, then 19' (p) I = 1 and gn (Po) = Po for all n. Finally, if h > - ~, then
gn(po) tends to infinity. Which behavior is correct? Checking the solution
we see that when k < 0, p(x) = poe kx tends to 0 as x goes to infinity so
Euler's approximation has the same qualitative behavior as the solution if
and only if h < - ~. How small the step size h needs to be in order to
guarantee the correct qualitative behavior depends on the absolute value
of k. No single h will work for all k. The larger k is in absolute value the
smaller h must be to exhibit the correct behavior. D

In Examples 13.2 and 13.3, Euler's approximation reduced to iteration

of a linear function. If we investigate more complicated differential equa-
tions, we expect to see more complicated dynamics and indeed this is what
happens.

EXAMPLE 13.4.
Let pi = P - p2, p(O) = Po, and h represent a positive step size for an
Euler approximation. Again, we denote the estimated value of p( nh) by
Pn. If g(p) = (p - p2)h + p, then

Pn = (Pn-I - P;-I) h + Pn-I = g(Pn-d = gn(po).

148 13. Numerical Solutions of Differential Equations

In exercise 10.4 we saw that all quadratic maps are topologically conjugate
to maps of the form qc(P) = p2 + c via a linear conjugacy. In particular,
there exists T(p) = mp+b so that Toqc = gOT. A few computations reveals
that g(p) = (p - p2)h + p is conjugate to qc(P) = p2 + c where c = 1 ~ h .
2
Since h > 0, there is a positive value of h for which the qualitative behavior
of g(p) is the same as the qualitative behavior of qc(p) = p2 + c for each
c < ~. Recall from exercise 11.5 that when -3/4 < c < 1/4, the only
periodic points of q are a single attracting fixed point and a repelling fixed
point. However, when c < -3/4 we see a cascade of period doubling
bifurcations followed by chaos. Consequently, the qualitative behavior of
the Euler approximation depends on the choice of h and hence c. We now
ask how this corresponds to the qualitative behavior of the solutions of the
differential equation.
From equation (13.2) we know that the solution of p' = p - p2 has the
form p(x) = C C
+ e- X where C = 1-p(O~)'
pO
Therefore, if p(O) is not 1 or 0,
then p(x) tends to 1 as x tends to infinity. If p(O) = 1, then the constant
function p(x) = 1 is a solution which attracts all other solutions and if
p(O) = 0, then the solution is the constant function p = 0 which repels all
other solutions. Returning to the Euler approximation, we are led to ask
if the solutions p = 1 and p = 0 correspond to the attracting and repelling
fixed points we see when -3/4 < c < 1/4.
We note that I-fk is an attracting fixed point of qc and l+fk is
a repelling fixed point. In the exercises we ask the reader to show that
the topological conjugacy maps I-fk to 1 and 1+f4Cto O. So when
-3/4 < c < 1/4, then 1 is an attracting fixed point of Euler's approxi-
mation and 0 is a repelling fixed point. Hence, the qualitative behavior of
the approximation is the same as the qualitative behavior of the solution.
Since c = I-t, -3/4 < c < 1/4 implies 0 < h < 2 and we know that
Euler's approximation displays the correct behavior when h < 2. When
h> 2 the approximation still has fixed points at 0 and 1, but they are both
repelling. As h tends towards infinity, c tends towards negative infinity and
the approximation goes through a series of period doubling bifurcations.
When h > 3, c < -2 and by exercise 10.4, the Euler approximation is
chaotic on a Cantor set with the remaining orbits tending to infinity. A
few typical behaviors are illustrated in Figure 13.4.
As a concluding comment, we note that if Po = 3 and h = 1, then the
Euler approximation does not converge to 1, even though the value of h is
within the range for which 1 is an attracting fixed point. We explore this
apparent contradiction in exercise 13.2. 0
 13. Numerical Solutions of Differential Equations 149

1.4

1.2

0.8

0.6

0.4

0.2

0 10 20 30 40

FIGURE 13.4. The light curve is the graph of the solution

of P = P - p2 with initial condition p(O) = 1.4. The lighter
jagged line is the Euler approximation for h = 1.9 and the
dark jagged line is the Euler approximation for h = 2.1.
Note that the approximation when h = 2.1 appears to
converge to a period 2 cycle whereas the approximation
when h = 1.9 tends to 1, as does the actual solution.

In the previous examples the derivative of the solution of the differential

equation at any point depended only on the value of the solution. We
saw that for a small enough value of h the Euler approximation always
exhibited the correct qualitative behavior. In the following example we see
that it is not always possible to find a value of h small enough to guarantee
the correct qualitative behavior of the approximation.

EXAMPLE 13.5.
Now consider the differential equation pi = - xp. By using techniques
similar to those we used earlier, we find that the solution of this equation
is p(x) = Ce- x2 / 2 where C = p(O). However, computing the Euler ap-
proximation in this case is a bit more complicated since the value of the
derivative is dependent on both the value of x and the value of p. We
extend our notation to include this fact by labeling the initial point in the
approximation (O,Po). Then the first iterate of Euler's approximation with
step size h yields (h,pd, the second yields (2h,P2), and so forth, with the
nth iterate being (nh , Pn). Note that as before we increase the first coor-
dinate by the step size h with each iterate. Also, the slope of the tangent
line at the point (nh,Pn) is -(nh)Pn. To determine Pn+l we use equation
(13.3). We have
150 13. Numerical Solutions of Differential Equations

To write the Euler approximation with step size h as the iteration of a

function as we have in the previous two examples, we must employ the
function G : JR2 --t JR2 defined by

G(x,p) = (x + h, y - xph).

Then (nh,Pn) = Gn(o,po). Even though G is a function of JR2, we will

use the tools we have developed for investigating iterated functions of JR
to determine whether or not the Euler approximation exhibits the correct
qualitative behavior.
We note that the actual solution, p(x) = Ce- x2 / 2 , tends to 0 as x tends
to infinity regardless of the initial value chosen. Setting the step size at .2
and the initial value at p(O) = 1 we use a computer to test the asymptotic
behavior of Euler's method for the first 200 iterates or so. We find

Po = 1 P60 >:::: _10- 35

P10>:::: 10- 1 P70 >:::: _10- 33
P20>:::: 10- 5 , PSO >:::: _10- 30
P30>:::: _10- 30 P90 >:::: 10- 26
P40>:::: _10- 34 P100 >:::: 10- 22 •
P50 >:::: _10- 35

This is more or less the behavior we expected, though the rise in the value
of P for the last few iterations is worrisome. Looking ahead we see things
aren't getting any better: in fact they are substantially worse.

and

Finally,

Clearly, things are getting out of hand. The iterations are swinging wildly
from one side of 0 to the other and the magnitude of the swings is growing.
Perhaps it will settle down again, but it doesn't look good. Taking a
hint from our earlier examples we try to avoid this unpleasant behavior
by using a smaller step size. In the past a small enough step size was
all that we needed to guarantee that the qualitative behavior of the Euler
approximation was the same as that of the actual solution. Letting h = .1
 13. Numerical Solutions of Differential Equations 151

we try again. We find

PlOO ~ 10- 43
P200 ~ _10- 99
P300 ~ _10- 83
P500 ~ _1011
PlOOO ~ _10 412 .

Things are definitely not getting better.

Based on the limited evidence we have acquired so far, one might spec-
ulate that over the long term Euler's approximation does not behave in
the same way as the actual solution. Closer examination of the iter-
ated function bears this out. Recall that (hn,Pn) = Gn(o,po) where
G(x,p) = (x + h, (1 - xh)p). If we write the coordinate functions sepa-
rately, we have G(x,p) = (gl(X,p),g2(X,p)) where gl(X,p) = x + h and
g2(X,p) = (1 - xh)p. As we are principally concerned with the fate of Pn
as n tends to infinity, we focus our attention on the behavior of g2 under
iteration.
If Po is an initial value which remains unchanged under iteration of G,
then g2(X,PO) = Po· Thus, Po = (1 - xh)po and (-xh)po = O. Since we
have assumed h is not 0 and x can not be 0 for more than one iteration,
this implies Po = 0 is the only initial condition which yields a stable state.
This corresponds to the behavior of the solution of the differential equation
in which the constant equation p(x) = 0 is a solution. However, in the
differential equation all other solutions tend to this constant solution over
time; the value of the solution, p(x) = Ce- x2 / 2 , to the differential equation
pi = -xp, tends to 0 as x tends to infinity. On the other hand, for the
two values of h we have checked the Euler approximation tended to 0 for
a time and then began to grow.
To determine if the growth of the Euler approximation continues in an
unbounded fashion for all values of h we again consider g2(X,p) = (l-xh)p.
Note that as x increases, (l-xh) decreases. If x> 2/h, then (l-xh) < -1
and since x continues to grow as we iterate G, the behavior of g2 is the
similar to that of £(p) = mp where m < -1: all points but 0 are in the
stable set of infinity and oscillate from side to side of zero. Hence, no matter
how small we choose h, once x > 2/h, that is once we have completed
2/h 2 iterations, the Euler approximation begins to grow without bound.
We make this statement precise in exercise 13.5. Note that this analysis
predicts that the approximation will begin to grow after 50 iterations when
h = .2 and after 200 iterations when h = .1. This correlates well with the
experimental data shown above. 0
152 13. Numerical Solutions of Differential Equations

Exercise Set 13

13.1 a) Complete the steps in Example 13.1 and verify the estimate for
p(5).
b) Let p' = .5p and p(O) = 1. Using Euler's method with 5 steps
each of length .2, estimate p(l) without using a computer.

13.2 In Example 13.4 we claimed that there is a linear map which is

a topological conjugacy between the Euler approximation of the
differential equation p' = p - p2 with step size h > 0 and the
quadratic map q(x) = x 2 + c where c = 1 4h 2 •
a) Find the topological conjugacy.
b) Show that 0 and 1 are always fixed points of the Euler ap-
proximation and that 1 is an attracting fixed point if and only if
0< h < 2.
c) If h = 1 and Po = 3, the Euler approximation grows without
bound. Why is this?
When h = 1, what is W S (l)?
d) Determine the stable set of 1 for each h between 0 and 2. Given
Po > 1, find a formula which will determine for which h the Euler
approximation with initial value Po will be converge to 1.
e) Determine the values of h for which the Euler approximation has
an attracting periodic orbit with prime period 2. Choose a step
size in the range you have found and use a computer program to
demonstrate several initial values which are attracted to a period
2 orbit. What are the coordinates of the period 2 orbit?
f) Explain why there exists ho so that the Euler approximation has
periodic points of all prime periods if and only if h > h o. Estimate
ho to the best of your ability.

13.3 For each of the following differential equations, represent the Euler
approximation as the iteration of a function of the real numbers.
Determine for which step sizes the qualitative behavior of the ap-
proximation is the same as the qualitative behavior of the solutions.
a) p' = -!p
b) p' = p - 2p2
c) p' = _p2
 Exercise Set 13 153

Note: For part (c), if p(O) f:- 0, then the solution is p(x) = x.!.c
where c = ptO) and if p(O) = 0, then the solution is the constant
function p = O.

13.4 In Example 13.3 we claimed that when p' = kp and k < 0, the
larger k is in absolute value, the smaller the step size must be in
order for the Euler approximation to exhibit the correct qualitative
behavior. Use derivatives to explain why this is so. For which
values of k must we use a step size which is smaller than .01 in
order to guarantee the correct qualitative behavior?

13.5 In Example 13.5 we claim that if x > 2/h, then the growth of the
second coordinate of the function G(x,p) = (x+h, (l-xh)p) under
iteration of G behaves qualitatively like iteration of the function
£(p) = -mp where m > 1. In particular, as n tends to infinity, the
absolute value of the second coordinate of Gn (x, p) tends to infinity
as well. To make this precise, fix h > 0, let (xo,Po) be an initial
point satisfying xo > 2/h, and define £(p) = (l-xoh)p. Prove that
the absolute value of the second coordinate of Gn (xo, Po) is greater
than or equal to l£n(po)1 for all natural numbers n.
Hint: Let g2(X,P) = (1 - xh)p and prove that
Ig2(G n (XO,Po))1 ~ £n+1(po) = 11 - xohln+lpo.

13.6 The solution of the differential equation p' = p - x is

p(x) = x + 1 + Ce x where C = p(O) - 1.
Determine whether or not the qualitative behavior of the Euler
approximation is the same as that of the solutions of the equation.
If not, why does it go awry?
14
The Dynamics of
Complex Functions

Many interesting and beautiful results in dynamics are seen in the realm
of complex functions. In the last two chapters of this text we examine
Newton's method for complex functions and the dynamics of the quadratic
map qc(z) = z2 + e when z is complex. Both of these subjects yield in-
teresting mathematics, as well as beautiful illustrations. In particular, the
well-known Mandelbrot set is intimately connected with the dynamics of
the quadratic map.

14.1. The Complex Numbers

The set of complex numbers consists of all numbers of the form a + bi
where a and b are real numbers and i satisfies i 2 = -1. We denote this set
by the symbol C. If a + bi and e + di are two complex numbers, then we
define addition, multiplication, subtraction, and division as follows:
(1) (a+bi)+(e+di)=(a+e)+(b+d)i
(2) (a + bi) - (e + di) = (a - e) + (b - d)i
(3) (a + bi)(e + di) = ae + bei + adi + bdi 2 = (ae - bd) + (ad + be)i
(4) (a+bi)+(e+di) = a+db~
e+ 2
(a+bi)(e-di) (ae+bd) (be-ad).
= (e + di)(e - di) = e2 - d2 + e2 _ d2 2
156 14. The Dynamics of Complex Functions

Note the use of a technique reminiscent of "rationalizing the denominator"

when simplifying the quotients of complex numbers.
While the algebraic properties of complex numbers were known as early
as the 16th century, their importance and utility was not realized until
their geometric properties were developed by Gauss in the 19th century.
We represent the complex number a+bi in the cartesian coordinate system
by the point (a, b). For example, the number 2 + 3i corresponds to the
point (2,3) in the plane. The set of real numbers considered as a subset
of the complex plane lies on the x-axis and the set of pure imaginary
numbers, which is all complex numbers of the form bi, falls on the y-axis.
Consequently, in the complex plane we often call the horizontal axis the
real axis and the vertical axis the imaginary axis.
A complex number plotted as a point in the plane can profitably be
visualized as a vector extending from the origin to this point. Adding two
complex numbers is analogous to adding the corresponding vectors. This
interpretation also leads to a natural definition of length. The complex
number a + bi has a length or modulus which is the distance from the origin
to the point (a, b) and is computed as J a 2 + b2 . This is the same notion of
length typically employed for vectors. We denote the modulus of a + bi as
la+bil. It is no accident that the notation used is the same as the absolute
value notation in the real numbers. The absolute value of a real number
can also be thought of as the distance from that number to zero. We define
a metric on the complex numbers in the same way that we did for real
numbers. Let z and w be complex numbers. Then the distance from z to
w is Iz - wi. Or, using our earlier notation, if z = a + bi and w = c + di,
then the distance from z to w is

Iz -wi = I(a+bi) - (c+di)1

= I(a - c) + (b - d)il = J(a - c)2 + (b - d)2.
We leave the proof that this definition of distance does in fact form a metric
on the complex numbers as an exercise.
Now that we have a metric on C we can describe its topology using
Definition 9.4. In particular we note that the neighborhood
Ne(z) = {w in C liz - wi < E}
is a disk in the plane with center z and radius E. Hence, an open set in the
complex plane must contain a disk around each point in the set. Readers
for whom this is a first introduction to complex analysis are encouraged to
review Definition 9.4 and see that they understand it as it applies to the
complex plane.
Points in the plane may be represented by polar coordinates as well as the
system of rectangular coordinates used in the cartesian coordinate system.
 14.1. The Complex Numbers 157

FIGURE 14.1. The complex number (a, b) is shown with

its argument indicated. The argument is the angle from
the positive x-axis to the ray through (a , b). Two such
angles are shown here.

If a + bi is a point in the plane, then its position is completely determined

by its distance from the origin and the angle from the positive x-axis to
the ray extending from the origin through a + bi. To be more specific, the
location of the point (a, b) is determined by its modulus, va 2 + b2 , and an
angle a satisfying

and (14.1 )

This angle is generally called the argument of a+bi and denoted arg(a+bi).
Of course, there are a few problems with the definition of the argument.
First, there are infinitely many angles which satisfy equation (14.1). Two of
these are demonstrated in Figure 14.1. Second, the argument isn't defined
when (a, b) = (0,0). The first of these problems is easy to deal with; we just
need to remember that if two complex numbers have the same modulus and
their arguments differ by an integer multiple of 27f, then they are the same
number. For example, V3 + i has modulus 2 and argument ~ + 2n7f where
n can be any integer. We acknowledge the second problem by not defining
an argument for zero. (There appear to be several good candidates for the
argument of zero, but all of them lead to contradictions in complex function
theory. Hence, we must resign ourselves to not being able to express zero
in polar coordinates. As we will see, this is not an undue hardship.)
Now let z be a nonzero complex number, r be the modulus of z , and a
be the argument of z. Then it should be clear that

z = r(cos(a) + isin(a)). (14.2)

158 14. The Dynamics of Complex Functions

If we extend the exponential function to the complex numbers so that it is

differentiable at all points in C and eUe v = eU+v for all complex numbers
u and v, then it can be shown that
cos (a) +isin(a) = eCY.i. (14.3)
This equation is often called Euler's formula. Combining equations (14.2)
and (14.3) we arrive at

where r = Izl and a = arg(z). We call this representation of z the expo-

nential or polar form.
When expressed in exponential form, a geometric interpretation of mul-
tiplication and division is readily apparent. If z = reCY.i and w = sef3i, then
. n~.
zw = (rs)e CCY.+f3l'. On the other hand, z -;- w = -f3' = !:.e CCY.-f3l'. Thus,
se ' S
the modulus of the product of two complex numbers is the product of the
moduli of the factors and the argument of the product is the sum of the
arguments of the factors. A similar statement holds for the quotient of
complex numbers. We can now find the square roots of complex numbers.

EXAMPLE 14.1.
Find the square roots of z = 4e i i .
If we indicate the root by w, then Iwl 2 = 4 and 2 arg(w) = ~. Obviously,
Iwl = 2 and arg(w) = ~ suffice. How do we find the second root?
Recall that we may write arg(z) = ~ + 2mf where n is any integer. So
the equation we really need to solve is 2 arg( w) = ~ + 2mf. Dividing by 2
we arrive at arg( w) = ~ + mr. Writing this as a real number plus integer
powers of 21f we see that arg(w) = ~ + 2n1f or arg(w) =
'IT' 7r' 97l'" .
9;+ 2n1f. So the
roots of z = 4e'42 are WI = 2e s ' and W2 = 2eT'. The complex number z
and its roots are shown in Figure 14.2. 0

14.2. Complex Functions

To aid our investigation of the dynamics of complex functions we restate
a few of the definitions and theorems that we developed for functions of
a real variable. In what follows the domain and range of all functions
are assumed to be subsets of the set of complex numbers unless stated
otherwise.
DEFINITION 14.2. The function f : D -7 C is continuous at the point Zo
in D if for each E > 0 there exists 8> 0 such that If(zo) - f(z)1 < E for all
z in D satisfying Izo - zl < 8. A function is continuous if it is continuous
at each point in its domain.
 14.2. Complex Functions 159

FIGURE 14.2. The geometric representation of the num-

ber z = 4e%i and its square roots. Note that if we double
the argument of either one of the roots we get the argu-
ment of z.

Note that this definition is the same as the definition used for func-
tions of the real numbers (Definition 2.7) and functions on metric spaces
(Definition g.4g). In particular, the numbers E and 8 are still positive real
numbers representing distances. When we visualize continuous functions of
the complex plane, we can use the same imagery as we used when consider-
ing continuous functions of the real numbers. The definition implies that if
f is continuous at zo, then for every neighborhood of f(zo), there is a 8 > 0
so that any point in the neighborhood of Zo with radius 8 is mapped inside
the neighborhood of f(zo). That is, f(N6 (zo)) is contained in N€(f(zo)).
We represent this in the complex plane by drawing the neighborhoods as
disks and visualize the neighborhood of Zo being mapped inside the neigh-
borhood of f(zo). This is illustrated in Figure 14.3.
The definition of the derivative for complex functions is also the same
as that of the derivative for real functions.

DEFINITION 14.3. The function f is differentiable at the point Zo if f is

defined at all points in a neighborhood containing Zo and the limit

. f(z) - f(zo)
11m
Z->Zo Z - Zo

exists. In this case we say f is differentiable at Zo and and call the limit
f' (zo) or the derivative of f at Zo. A function is differentiable if it is
differentiable at each point in its domain.
160 14. The Dynamics of Complex FUnctions

FIGURE 14.3. The behavior of a function which is contin-

uous at Zo0 If J is continuous at Zo, then for each E > 0,
there is 0 > 0 so that J(N8 (zo)) c N,(J(zo)).

Of course, the problem with this definition is that we must define what
. J(z) - J(zo) .
we mean by hm for complex valued functIOns. We recall
Z->Zo z - Zo
from calculus that lim g(x) = L if the value of g(x) is close to L when x
X--+XQ

is close to a. We make this precise by using the metric.

DEFINITION 14.4. Let J be a Junction which is defined on a neighborhood
oj Z00 Then
lim J(z) = L
Z--+ZO

iJ Jor every E > 0, there exists 0 > 0 so that iJ 0 < Iz - zol < 0, then
IJ(z) - LI < E.
If 0 and E are small, then the expression 0 < Iz - zol < 0 implies that
z is close to but not equal to Zo and IJ(z) - LI < E guarantees that J(z)
is close to L. Drawing a diagram in the complex plane which illustrates
Definition 14.4 can be an instructive exercise and the reader is encouraged
to do so. Notice that we allow z to approach Zo from any direction. This
is a significant difference from the real case in which we can approach
numbers only from the left and the right. As a result of this difference the
properties of differentiable complex functions are significantly different from
those of differentiable real functions. All differentiable complex functions
are analytic which means that their Taylor series expansion is defined and
converges on a neighborhood of any point in the domain. However, the
properties of complex polynomials and rational functions are analogous to
those of real polynomials and rational functions. As these are the only types
 14.2. Complex Functions 161

of complex functions we will encounter in this text, we will not develop

the general theory of complex functions. Readers who are interested in
learning more about the truly fascinating properties of complex functions
are encouraged to consult one of the texts listed in the references.
When we discuss the Riemann sphere in Section 14.4, we will find it
useful to have a definition of limits which doesn't explicitly refer to a metric.
We restate Definition 14.4 below using neighborhoods. The reader should
verify that the two statements of the definition are in fact equivalent.

DEFINITION 14.4. (RESTATED) Let 1 be a function which is defined on

a neighborhood 01 z00 Then

lim I(z)
Z-+Zo
=L
if for every neighborhood N of L, there exists a neighborhood N' of Zo so
that if z is in N' and not equal to Zo, then f(z) is in N.

In the following theorems we collect a few elementary facts about differ-

entiable functions, all of which are analogous to the real case:

THEOREM 14.5. Differentiable functions are continuous.

THEOREM 14.6. If f and g are complex functions which are differen-

tiable at Zo, then the following statements hold:
a) 1 + g is differentiable at Zo and (f + g)'(zo) = I'(zo) + g'(zo).
b) If a is a complex number, then (al)'(zo) = al'(zo).
c) The product f 9 is differentiable at Zo and
(fg)'(zo) = I(zo)g'(zo) + I'(zo)g(zo).

d) If g(zo) =1= 0, then £ is differentiable at Zo and

9

(£)'
9
(zo) = g(zo)I'(zo) - I(zo)g'(zo).
(g(ZO))2
THEOREM 14.7. THE CHAIN RULE. If 1 is differentiable at Zo and 9
is differentiable at 1(zo), then the composition (g 0 I) is differentiable at Zo
and (g 0 I)'(zo) = g'(f(zo))l'(zo).

THEOREM 14.8. The formula

_d zn = nzn-l
dz
holds for all natural numbers n.
162 14. The Dynamics of Complex Functions

We do not prove any of these theorems, but we do illustrate the proof of

Theorem 14.8 in the following example. Note that Theorems 14.6 and 14.8
imply that all complex polynomials and rational functions are differentiable
throughout their domains. In addition, the formulas for differentiating
these functions are analogous to those for real polynomials and rational
functions.

EXAMPLE 14.9.
If J(z) = z2, then J'(zo) = 2zo for all Zo since

Z2_ Z2
J' (zo) = lim
Z-+Zo z - Zo
0 = lim (z
Z-+Zo
+ zo) (z-zo)
- - = 2zo.
Z - Zo
o

14.3. The Dynamics of Complex Functions

We begin our investigation of the dynamics of complex functions with
functions of the form J(z) = az where a is a complex number. We leave
investigation of the slightly more general function g(z) = az + b as an
exercise.

EXAMPLE 14.10.
Let J(z) = az where a is a complex number whose modulus is not 1.
We note that 0 is the only fixed point of J and investigate the orbit of Zo
when Zo -# o.
Clearly, P(zo) = J(azo) = a2zo and in general r(zo) = anzo. If we
write a as lale Oi and recall that an = lalne nOi , then we can write r(zo) as
lalnenOizo or in exponential form

Therefore, Ir(zo)1 = lalnlzol, and arg(r(zo)) = arg(zo) + nO. If lal < 1,

then lalnlzol converges to 0 as n increases. Thus, WS(O) = C when lal < 1.
On the other hand, lalnlzol tends to infinity when lal > 1, and so WS(O) = 0
and WS (00 )is all of C except 0 in this case. The argument of a does not
affect the convergence or nonconvergence of In(zo) to zero, but it does
affect the orbit. Each time the function is iterated, arg( a) is added to the
argument of the input. Thus, when the argument of a is not zero the orbit
of any nonzero point winds around the fixed point. This is illustrated in
Figure 14.4. 0
 14.3. The Dynamics of Complex Functions 163

•
/(1)
•
•
/(1)
• •

•
•

FIGURE 14.4. The orbit of 1 under iteration of the linear

function J(z) = az. In the first figure, a = ~eii and in the
second one a = teii. Note that in the first diagram the
modulus is increased by a factor of 1.25 with each iteration
and in the second diagram the modulus of each iterate is .8
times the modulus of the previous iterate. In both cases,
i is added to the argument with each iteration.

In Example 14.10, we saw that 0 is a fixed point of J(z) = az and

that WS(O) = C when 11'(0)1 = lal < 1. On the other hand, WS(O) = 0
when IJ'(O)I = lal > 1. This leads us to conjecture that the derivative
of a fixed point indicates the repelling or attracting nature of fixed points
for complex functions as it does in the real case. We formalize this in the
following theorem.

THEOREM 14.11. Let f be a differentiable complex function and p be

a fixed point of f. If If'(p)1 < 1, then the stable set of p contains a
neighborhood of p. If 11'(p)1 > 1, then there is a neighborhood of p all of
whose points must leave the neighborhood under iteration of f.

PROOF. We prove the theorem for the case 11'(p)1 < 1. The proof of
the case 11'(p)1 > 1 is left as an exercise.
Assume f(p) = p and 11'(p)1 < 1. Since ~(l-IJ'(p)l) > 0, the definitions
of the derivative and limits imply that there exists 8 > 0 such that if
0< Iz - pi < 8, then

- f(p) - 1'(p) I < ~(1 -If'(p)l)·

I f(z)z-p 2
(14.4)
164 14. The Dynamics of Complex Functions

Also, by the triangle inequality

- f(p) I : :; If(z) - f(P) I+ I!'(P)I·

If(z)z-p z-p
- !,(p) (14.5)

Inequalities (14.4) and (14.5) together imply that if 0 < Iz - pi < 8, then

- f(p) I : :; ~(I-I!,(p)l) + I!'(p) I = ~(1 + If'(p)l)·

If(z)z-p 2 2
(14.6)

We note that !(1 + If'(p) I) < 1 since 1f'(P) I < 1 and set A = !(1 + I!,(p)l)·
Multiplying both sides of inequality (14.6) by Iz - pi and substituting A for
!(1 + I!,(p)l) we see that when Iz - pi < 8, then
If(z) - pi = If(z) - f(p) I :::; Alz - pI·
By induction we can show that if Iz - pi < 8 and n is any natural number,
then
Ir(z) - pi:::; Anlz - pI·
Since A < 1, this implies that r(z) tends to p as n tends to infinity
whenever z is such that Iz - pi < 8. Thus, the neighborhood of p with
radius 8 is in the stable set of p. 0
Of course, we can generalize Theorem 14.11 to periodic points in the
same way that we generalized the analogous theorem for fixed points of the
real numbers.
THEOREM 14.12. Let f be a differentiable complex function and p be
a periodic point of f with period k. If IUk),(p)1 < 1, then there exists a
neighborhood ofp which is contained in the stable set ofp. If IUk)'(p) I > 1,
then there exists a neighborhood of p such that all points must leave the
neighborhood under iteration of fk .

Theorem 14.12 is an easy extension of Theorem 14.11 and its proof is

left as an exercise. These theorems motivate the definition of hyperbolic
repelling and attracting points.
DEFINITION 14.13. Let f be a differentiable complex function and p be
a periodic point with period k. Then p is a hyperbolic periodic point if
IUk),(p)1 #1. If IUk),(P)1 < 1, then p is an attracting periodic point. If
IUk)'(p) I > 1, then p is a repelling periodic point.

As with the real case, periodic points which are not hyperbolic do not
have predictable behavior in a neighborhood of the point. In the next
example we examine the behavior of points in C when we iterate a function
of the form f(z) = az where lal = 1. Note that If'(O)1 = 1 and 0 is the
only fixed point of f unless a = 1.
 14.3. The Dynamics of Complex FUnctions 165

EXAMPLE 14.14.
Suppose j(z) = e9i z and Zo is a nonzero complex number. Then the
orbit of Zo under iteration of j lies on the circle which is centered at the
origin and whose radius is Izol. If 0 is a rational multiple of 'Jr, then Zo is
a periodic point of j. If the argument of 0 is not a rational multiple of 'Jr,
then Zo is not a periodic point and its orbit is dense on the circle.
To see this, note that jn(zo) = en9izo. Consequently,

Ir(zo)1 = len9i llzol = Izol

for all n and all iterates of Zo must be on the circle with radius Izol and
center at 0. Also, arg(r(zo)) = arg(zo) + nO so Zo is periodic if and only
if there exists a natural number n and an integer k so that
arg(zo) + nO = arg(zo) + 2k'Jr.
Solving this equation for () we find () = 2~71". Hence, Zo is periodic if and
only if () is a rational multiple of 'Jr.
Now suppose () is not a rational multiple of 'Jr. We claim that the orbit
of Zo is dense on the circle with radius Izol and center at the origin. To
prove this we show that every neighborhood of every point on the circle
with radius Izol and center at the origin must contain an iterate of zoo
Let E > 0, and r = Izol. Then the circumference of the circle with radius
Izol and center at the origin is 2'Jrr. As () is not a rational multiple of'Jr we
know that jm(zo) #- Zo for any natural number m. In particular, if m is
larger than 2'Jrr / E, then none of the first m iterates of Zo are the same and
there must be two between which the distance is less than E. (The reader
is encouraged to explain why this is so.)
Suppose that Ijm2(zo) - jml(zo)1 < E and m2 > mI. Since we know
that jml (zo) = re(m 1 9+arg(zo))i and jm2 (zo) = re(m 2 9+arg(zo))i, this implies
that
Ire(m 2 9+arg(zo))i _ re(ml9+arg(zo))il = le(ml9+arg(zo))i(re(m2 -mt}9i _ r)1
= Ire(m 2 -m tl 9i - rl < E.
Thus,
Ijm 2 -m 1 (zo) - zol = Ire(m 2 -m 1 )9i+ arg(zo)i _ rearg(zo)il
= Ire(m 2 -mt}9i - rl < E. (14.7)
As we noted above, jm 2-m 1 rotates points in C by the nonzero angle
(m2 - mI)B and inequality (14.7) indicates that the distance between suc-
cessive iterates of jm 2-m 1 is less than E. Hence, the sequence of points zo,
jm 2 -m 1 (zo), p(m 2-m tl (zo), j3(m 2-m tl (zo), ... covers the circle of radius
Izol by points which are separated by a distance which is less than E. Thus,
166 14. The Dynamics of Complex Functions

if w is any point on the circle of radius Izol, then the neighborhood of w

with radius € must contain an iterate of zoo As € is arbitrary, this implies
the orbit of Zo is dense on the circle with center at the origin and radius
Izol as desired. D
In the last example, 0 is a fixed point of the function f(z) = efJiz,
11'(0)1 = lefJil = 1, and 0 is neutral in the sense that points near 0 do not
get closer or further away from 0 as the function is iterated. The reader
is encouraged to try to find functions with nonhyperbolic fixed points that
either attract or repel nearby points.
We complete this section with an investigation of the dynamics of the
function q(z) = Z2. We will investigate the more general quadratic family
qc(z) = Z2 + c in Chapter 15.

EXAMPLE 14.15.
Let q( z) = z2. The only fixed points of q are 0 and 1. Since q' (0) = 0
and q' (1) = 2, 0 is an attracting fixed point and 1 is repelling. If z = refJi,
then q(z) = (refJi)(refJi) = r 2e2fJi . Iterating, we find q2(Z) = r 4e4fJi and in
general qn(z) = r2ne2nfJi. Since Iqn(z)1 = r 2n , we see that z is in WS(O) if
and only if Izl < 1 and z is in WS(oo) if and only if Izl > 1.
If Izl = 1, the dynamics are more interesting. Notice that Izl = 1 implies
that Iqn(z)1 = 1 for all n. Thus, if we denote the circle of points whose
modulus is one by 8 1 , then we can write q : 8 1 -+ 8 1 . (The notation 8 1 is
typically used to denote the unit circle in the plane.) Now if arg(z) = (),
then arg(q(z)) = 2() and we recognize our friend from exercise 9.12: the
doubling map on the circle. Each time we apply the function q(z) = z2 to a
point in 8 1 the modulus of the point remains unchanged and the argument
is doubled. We demonstrated in exercise 9.12 that the doubling map on 8 1
is chaotic so it follows that q(z) = z2 is chaotic on 8 1 as well. Orbits of
typical points in C under iteration of q are shown in Figure 14.5. D

14.4. The Riemann Sphere

In this section we introduce the geometric description of the complex
numbers as points on a sphere. While this may seem to be an unnecessary
abstraction, we shall see that it is just the notion we need to simplify
the proof of several interesting facts in dynamics. For example, in the next
section we use the spherical representation of the complex numbers to show
that Newton's method for f(x) = x 2 + 1 is chaotic on the real line. While
we proved this result already in Example 12.7, we will see that the proof
using the Riemann Sphere is much simpler. The principal advantage of the
spherical representation is that it includes the point at infinity. Once we
 14.4. The Riemann Sphere 167

initial pointS

FIGURE 14.5. The orbit of the points .98ei i , ei i , l.02e ii

under iteration of q(z) = z2. Notice that the orbits con-
verge to zero or infinity if the modulus of the initial point
is not equal to one. If the modulus of the initial point
is equal to one, then its orbit remains on the unit circle.
The unit circle is shown in black, the spirals indicating the
orbits are shown in gray.

_ - - - - (0.0.1)

FIGURE 14.6. Stereographic projection of the unit sphere

on to the xy-plane. A complex number z and its image z'
on the Riemann Sphere are shown.
168 14. The Dynamics of Complex Functions

have defined this new representation we will treat 00 as a point and include
it in the domain and range of functions.
To define the Riemann Sphere we use stereographic projection, the same
technique which is sometimes used to create maps of the world. A good way
to visualize it is to picture a sphere with a radius of one and center at the
origin in ]R3. We denote this sphere l by S2. The xy-plane represents the
complex plane and intersects the sphere at the equator. This arrangement
is pictured in Figure 14.6 (p. 167).
At the top of the sphere is the point (0,0,1). If we draw the line through
the point at the top and any point in the complex plane, then it intersects
the sphere at exactly one point other than (0,0,1). If we draw a line from
(0,0,1) to a point which is greater than one in modulus, then it intersects
the top half of the sphere. If we draw one through a point which is on the
unit circle, then it intersects at the equator, and if we draw a line from
(0,0,1) through a point which is less than one in modulus, it intersects the
sphere at a point on the bottom half of the sphere. To get a correspondence
between points on the sphere and the points in the complex plane we assign
each complex number to the point on the sphere which is on the line through
(0,0,1) and the number. It should be geometrically obvious that this is
a one-to-one, onto correspondence between all of the points on the sphere
except (0,0,1) and the complex plane. If in addition we assign (0,0,1)
to the point 00, then we have a one-to-one, onto correspondence between
S2 and C U {oo}. We leave as an exercise the verification of the fact
that (0,0,-1) corresponds to 0, (0,1,0) corresponds to i, and (~,~, 0)
corresponds to (~V3 + 2.;2)(1 + i).
Now consider a sequence of points 0, Zl, Z2, Z3, ... which satisfy IZkl = k.
We would like limk--+oo Zk = 00 since it is clear that the points march
towards (0,0,1) on S2. To define this limit we need to define a set of
neighborhoods for each point on S2. Once we have defined neighborhoods
we can also define open sets, convergence of sequences, and continuous
functions without reference to a metric.

DEFINITION 14.16. Let z' be a point on the Riemann Sphere which cor-
responds to Z in the complex plane. The set U' on the Riemann Sphere

°
is a neighborhood of z' if the corresponding set U is a neighborhood of Z
in the complex plane. Let E > and define NE (00) = {z in C I Izi > ~}.

1 We have already used the symbol 8 1 to indicate a sphere (circle) with radius one
and center at the origin in IR2. The symbol 8 2 is used here to indicate a sphere with
radius one and center at the origin in IR3. In general, 8 n indicates a sphere with radius
one and center at the origin in IRn+1. Note that position can be defined in terms of one
variable (the angle) on 8 1 , it takes two variables to describe position on 8 2 , and it takes
n variables to describe a position on 8 n , hence the superscript.
 14.4. The Riemann Sphere 169

Then the corresponding set N.( (0)' on S2 is a neighborhood of (0,0,1) on

the Riemann Sphere.

Neighborhoods in S2 correspond directly to the usual definitions of

neighborhoods in C with the addition that neighborhoods of 00 are de-
fined as sets of the form N.(oo) = {z in C Ilzl > ~} where E > O. It should
be clear that as E decreases, the neighborhoods of 00 decrease as well. Also,
if we adjoin the point 00 to C, then it is natural to include 00 in N. (00 ).
We caution the reader that it is not a wise idea to select neighborhoods
arbitrarily. The foundations of general topology rest on the properties of
neighborhoods and open sets, and care was taken when writing Definition
14.16 to see that if Nand N' are neighborhoods of x, then N n N' is a
neighborhood of x and x is in both Nand N'. Readers who are intrigued
by the necessity of these seemingly arcane requirements are encouraged to
consult a text on general topology. It is a truly fascinating subject.
Since we are defining all of our neighborhoods on S2 in terms of sets in
cui oo}, we will often refer to the Riemann Sphere as the extended complex
plane. We define functions on the Riemann Sphere using the usual algebra
of C with two additional properties. If f is a function defined on a subset
of the complex plane, then f ((0) = limz->oo f (z) if that limit exists. Also,
if limz->a f(z) = 00, then f(a) = 00.
Now that neighborhoods are defined, we restate the definitions of open
set, convergent sequence, and continuous function. Note that we are not
changing the definitions of these concepts, but reformulating them without
specific mention of a metric. The definitions of accumulation point, closed
set, and dense subset already are defined in terms of neighborhoods.

DEFINITION 14.17. A set is open if every element of the set has a neigh-
borhood that is completely contained in the set.

DEFINITION 14.18. A sequence of points converges to a point if every

neighborhood of that point contains an element of the sequence.

DEFINITION 14.19. Let f : D --t C and Zo be an element of D. Suppose

that for every neighborhood V of f(zo) we can find a neighborhood U oj Zo
so that J(U) c V. Then f is continuous at zoo

Definition 14.17 is a restatement of Proposition 9.5, Definition 14.18 can

easily be shown to be equivalent to Definition 9.4c, and Definition 14.19 is
clearly equivalent to Definition 14.2. Notice in particular that Figure 14.3
on page 160 can be used to illustrate Definition 14.19 as well as Definition
14.2.
In Section 14.5 and Chapter 15 we investigate the dynamics of functions
of the extended complex plane. While doing so, we will use a special class
170 14. The Dynamics of Complex Functions

of functions called Mobius transformations or linear fractional transforma-

tions. These functions have the form
T(z) = az + b
cz+d
where a, b, c, and d are complex numbers that satisfy the condition that
ad - bc ¥= O. To extend T to the complex plane we let T(oo) = %. Also, if
cZo + d = 0, then T(zo) = 00. Mobius transformations have many interest-
ing properties. First, they always map circles on the Riemann Sphere to
circles. Since a line in the complex plane becomes a circle through 00 in the
extended complex plane, this implies that circles and lines in the complex
plane are mapped to circles and lines in the complex plane. This is most
easily shown by demonstrating that every Mobius transformation can be
written as a composition of functions of the form fez) = az, g(z) = z + b,
and k(z) = ~ and then proving that these functions have the requisite prop-
erty. Details can be worked out by the reader or found in a text on complex
analysis. As an example of this property, we note that the transformation
T(z) = ~1 + ~~~z + 1 maps the real axis in the complex plane to the unit
1- 2 Z + 1
circle. The reader may find it instructive to compute the image of the imag-
inary axis as well. The other attribute of Mobius transformations which
we will use is that all Mobius transformations are homeomorphisms of the
extended complex plane. We leave the proof of this fact as an exercise.

14.5. Newton's Method in the Complex Plane

In Chapter 12 we considered Newton's method on the real line. In this
section we extend that discussion to the complex plane. While the geo-
metric interpretation of Newton's method as an approximation by tangent
line is not valid in the complex case, the basic formula is still the same
and the theorems we proved in Chapter 12 still hold. In particular, New-
ton's function for the complex polynomial fez) is defined by the equation
fez)
Nf(z) = z - f'(z). Further, we have

THEOREM 14.20. If p(x) is a complex polynomial and we allow cancel-

lation of common factors in the expression of Np(x), then Np(x) is always
defined at roots of p(x), a number is a fixed point of Np(x) if and only if it
is a root of the polynomial, and all fixed points of Np(x) are attracting.

The proof of Theorem 14.20 is identical to that of Theorem 12.2.

We begin our investigation of Newton's method for complex quadratic
functions with an example.
 14.5. Newton's Method in the Complex Plane 171

EXAMPLE 14.21.
Let J(z) = Z2 - 1. A quick calculation reveals that N(z) = z22: 1.
From Theorem 14.20 we know that 1 and -1 are attracting fixed points of
N. We claim that if a+bi is an element ofC and a> 0, then a+bi is in the
stable set of 1. On the other hand, if a + bi is such that a < 0, then it is in
the stable set of -1. Finally, we claim that N is chaotic on the imaginary
axis. We prove this claim by showing that N(z) is topologically conjugate
to q(z) = z2 on the extended plane via a Mobius transformation.
If such a Mobius transformation exists, we should be able to find it by
searching for one that has the required properties. Since any three points
determine a circle and Mobius transformations must map circles to circles,
we look for three significant points of N and the corresponding points of
q to which they must be mapped. The attracting fixed points of N are 1
and -1, so we map them to 0 and 00, the attracting fixed points of q. The
only other fixed point of N is 00 and so it must be mapped to the repelling
fixed point 1 of q. If we are lucky, the coefficient of z in the numerator
of the Mobius transformation we are looking for is not zero, and we can
divide the numerator and denominator of the transformation by it. In this
case, the requisite transformation has the form

T(z) = z + b .
cz+d
Setting T(l) = 0, T( -1) = 00, and T(oo) = 1 and solving for b, c, and d
we find that
T(z) = z -1.
z+l
Verifying that ToN = q 0 T is an easy exercise in algebra. As T is a
homeomorphism, q and N are topologically conjugate.
Since T(O) = -1, T(i) = i, and T(oo) = 1, the image of the imaginary
axis must be the unit circle. (The unit circle is the only circle passing
through -1, i, and 1, and T maps circles and lines to circles and lines.)
Since we know from Example 14.15 that q is chaotic on the unit circle
and that T is a topological conjugacy, it is tempting to conclude that N
is chaotic on the imaginary axis. This is in fact true, but we have to be
a little bit careful in proving it. Theorem 10.3 states that if a function of
a metric space is topologically conjugate to a second function of a metric
space which is chaotic, then the first function is chaotic as well. But T is a
topological conjugacy between functions on the Riemann Sphere, which we
have not defined as a metric space. As the definition of sensitive dependence
on initial conditions uses a metric, the notion of chaotic as we have defined
it doesn't even make sense for functions of the Riemann Sphere, unless we
define a metric there.
172 14. The Dynamics of Complex Functions

We have two alternatives for dealing with this problem. First, we can
define the distance between two points on the Riemann Sphere by
d[(Xl' X2, X3), (Yl, Y2, Y3)] = J (Yl - xd 2 + (Y2 - X2)2 + (Y3 - X3)2.
This is the usual euclidean metric used in m. 3 . It is easy to show that the
subset U of 8 2 is open (as defined in Definitions 14.16 and 14.17) if and
only if for each x in U there exists a positive real number E so that if Y is
in 8 2 and d[x, y] < E, then Y is in U. Thus, we can use this metric to define
chaotic behavior on the Riemann Sphere and Theorem 10.3 holds.
On the other hand, we can avoid these problems by noting that the
periodic points of q are dense on the unit circle and that q is topologically
transitive on the unit circle. These properties are defined only in terms of
open sets and are preserved by topological conjugacy. Thus, the periodic
points of N are dense on the imaginary axis and N is topologically transitive
there as well. Now, since sets that are open in the complex plane are also
open in the extended complex plane, we may conclude that the periodic
points of N are dense and N is topologically transitive on the imaginary
axis, considered as a subset of the complex plane. It follows from Theorem
9.20 that N is chaotic on the imaginary axis.
It remains to show that a + bi is in the stable set of 1 if and only if a > 0
and it is in the stable set of -1 if an only if a < o. But, T(l) = 0 and
T({a+bi I a> O}) = {z Ilzl < I},
which is the stable set of 0 under iteration of q. Since T preserves stable
sets the first assertion holds. A similar observation demonstrates the second
statement. The reader may wish to verify that the proof that topological
conjugacies preserve stable sets does not depend on the existence of a metric
in the spaces in question. 0
When reflecting on the previous example, we note that the imaginary
axis is the perpendicular bisector of the line segment between the roots of
f(z) = Z2 - 1. Also, recall from our work in Chapter 12 that Newton's
method had similar dynamics for large classes of quadratic polynomials.
We are led to ask for which complex quadratic polynomials the perpen-
dicular bisector of the line segment between the roots divides the stable
sets of the roots, and in which of those is Newton's method chaotic on the
bisector. In 1879, the English mathematician Arthur Cayley proved that
the perpendicular bisector of the line segment joining the roots divides the
stable sets of Newton's function for all complex quadratic polynomials with
distinct roots. We can now add that it is chaotic on the bisector as well.
THEOREM 14.22. [CAYLEY, 1879] If the complex quadratic polynomial
q(x) has two distinct roots, then Nq(x) is chaotic on the perpendicular
bisector of the line segment joining the two roots. Further, the stable set
 14.5. Newton's Method in the Complex Plane 173

FIGURE 14.7. An illustration of Cayley's Theorem. The

roots of the complex quadratic polynomial are indicated
by rl and r2. The stable set of rl under iteration of Nq(x)
is indicated in gray. Nq(x) is chaotic on the perpendicular
bisector labeled as L. The remaining points are in the
stable set of r2.

oj each root under iteration oj Nq(x) is the set oj points which lies on the
same side oj the bisector as the root.

An illustration of Theorem 14.22 is shown in Figure 14.7. There are two

convenient proofs of the Theorem. In one we show that Newton's method
for any complex quadratic polynomial with distinct roots is topologically
conjugate to Newton's method for J(z) = Z2 - 1 via a linear map. In the
other we show that Newton's method for any complex quadratic polynomial
with distinct roots is topologically conjugate to the function J(z) = Z2, and
proceed as in Example 14.21. The details are left to the reader.
We complete the description of Newton's method for complex quadratic
polynomials with the following theorem:

THEOREM 14.23. IJ the complex quadratic polynomial q has exactly one

root, then all points in C are in the stable set oj the root under iteration oj
Nq(x).

Theorem 14.23 is easily proven by showing that it holds for q(z) = Z2 and
then proving that Newton's function for a complex quadratic polynomial
with a single repeated root is topologically conjugate to N q • Theorem
14.23 also generalizes to complex polynomials of higher degree with a single
repeated root.
174 14. The Dynamics of Complex Functions

Given our success with understanding Newton's method for complex

quadratic polynomials we are led to hope that an analysis of Newton's
method for complex cubics will prove to be as simple. However, our en-
thusiasm should be tempered by our experience with Example 12.13 in
which we found an attracting periodic point with prime period two for
Newton's method applied to a real cubic equation. We begin cautiously
and investigate Newton's method for J(z) = z3 - 1.

EXAMPLE 14.24.
Let J(z) = z3 -1. Then N(z) = 2z 3 ~ 1 and has attracting fixed points
h· k·
3z .
at 1, ea ', and ea'. To get some idea of what we mIght expect to find we
conduct a computer experiment. We select a grid of points covering the
square region whose corners are at 2 + 2i, 2 - 2i, -2 - 2i, and -2 + 2i.
We calculate the value of the 100th iterate of each of these points under
iteration of N. If the distance from the 100th iterate to 1 is less than ~,
then we assume the point is in the stable set of 1 and color it light gray. If
the distance from the 100th iterate to e 23" i is less than ~, then we assume
the point is in the stable set of e 2; i and color it dark gray. Finally, if the
distance from the 100th iterate to e 4; i is less than ~, then we assume the
point is in the stable set of e 4; i and color it black. All points which are not
within ~ of one of the roots after 100 iterations of N are left white. The
result of this experiment using a 250 by 250 grid is shown in Figure 14.8.
We can see that the dynamics of Newton's method for this cubic are
much more complicated than those of Newton's method for quadratic func-
tions. In fact, the dynamics of Newton's method for most polynomials of
degree higher than two are not completely understood. The reader is en-
couraged to use experiments similar to the one above to try to understand
the dynamics of Newton's method for cubics. In particular, we would like
to know what happens on the boundary of the stable sets. Specific sugges-
tions for projects can be found in the exercises. 0
 14.5. Newton 's Method in the Complex Plane 175

FIGURE 14.8. An analysis of the dynamics of Newton's

method for J(z) = z3 -1. There are three attracting fixed
points for this function, 1, e 2; i, and e 43" i. Points which
are known to be in the stable set of 1 are shown in light
gray, points which are known to be in the stable set of
e¥i are shown in dark gray, and points which are known
to be in the stable set of e 4; i are shown in black. Points
for which the eventual destination is not known are shown
in white. The corners of the region shown are at -2 - 2i,
-2 + 2i, 2+ 2i, and 2 - 2i.
176 14. The Dynamics of Complex Functions

Exercise Set 14

14.1 Simplify the following expressions:

a) (2 - 3i) + (-1- 4i)
b) !i - (1 + i)
c) (2 + i)(i)
d) (2 + i) -7 i
e) (1 - 4i)(2 + 3i)
f) (1 - 4iH- (2 + 3i)

14.2 Find the modulus and argument of each of the following complex
numbers:
a) 1- 4i

b) ! + £i
c) 2

14.3 Draw a sketch of the neighborhood of 3 + i with radius !. Do the

same for the neighborhood of -i with radius 2.

14.4 Find all of the solutions of the following equations and plot them
on the complex plane. What do you notice about the geometry of
the solutions?
a) z2 = 8.
b)z3-8=0.
c) zn - 1 = 0, for n = 2, 3, 4, 5, and 6.

14.5 Let Zl = reQl i and Z2 = re Q2i • Suppose there exists an integer k

so that Ir(al + 2k7r - (2)1 < f. Prove IZI - z21 < f.
Hint: Draw a representative diagram and consider the geome-
try.

14.6 Verify that the function d[a + bi, c + di] = y'(c - a)2 + (d - b)2 is
a metric on Co
 Exercise Set 14 177

14.7 Let J be a differentiable complex function. Show that if r(p) is

defined for all n :::; k, then
(fk)'(p) = J'(p). !,(f(p)) . !,(f2(p)) ..... J'(fk-l(p)).

14.8 Prove that ! zn = nzn-l for all integers n.

14.9 Prove that J is continuous at Zo if it is differentiable there.

14.10 a) Complete the details of the induction argument used in the proof
of the first half of Theorem 14.1l.
b) Let J be a differentiable complex function and p be a fixed point
of J. Prove that if If'(p) I > 1, then there is a neighborhood of p
all of whose points must leave the neighborhood under iteration of
J.
14.11 Let g(z) = az + b where a and b are complex numbers.
a) Find the fixed point of g if it exists. In what case(s) will there
not be a fixed point?
b) Show that if a of. 1, then g(z) = az+b is topologically conjugate
to a function of the form J(z) = cz.
c) In those cases where g has a fixed point, describe its stable set.
d) In those cases where g does not have a fixed point, describe the
dynamics of g.

14.12 Plot the orbit of a typical point under iteration of gl(Z) = !z + l.

Do the same for g2(Z) = 2z - 2.

14.13 The orbit of 1 under iteration of J(z) = az for a = ~e.g-i and

a = ~e.g-i is shown in Figure 14.4 on page 163. The spiral is drawn
so that if u and v are any two points on the spiral which lie on
the same ray extending from the origin, then there exists a natural
number n so that either r(u) = v or r(v) = u.
a) Find the equation of the spiral when a = ~e.g-i and show that
your equation has the required property. Do the same for a = ~e.g-i.
b) For which values of a is it possible to draw such a spiral?
c) Prove that the spiral through the iterates of q( z) = Z2 shown in
Figure 14.5 can not have this property.
178 14. The Dynamics of Complex Functions

14.14 * Suppose p is a fixed point of the differentiable function J and

If'(p) I = 1. If there exists a neighborhood of p which is contained
in the stable set of p, then p is weakly attracting. If there exists a
neighborhood U of p such that for each point x in U except p there
is a positive integer n so that r(x) is not in U, then p is weakly
repelling. Find an example of a function of the complex numbers
with a weakly attracting or a weakly repelling fixed point.

14.15 Let q(z) = Z2. Show that 1 is a repelling fixed point of q and that
W8(1) is dense on the unit circle in Co Why doesn't this contradict
Theorem 14.11?

14.16 Let J(z) = z3.

a) Find the fixed points of J and determine whether they are at-
tracting, repelling, or nonhyperbolic.
b) Describe the periodic points of J. Determine whether they are
attracting, repelling, or nonhyperbolic.
c) Describe the orbit of all the points in C to the best of your
ability.

14.17 Verify that in the description of the Riemann Sphere outlined in

Section 14.4, (0,0, -1) corresponds to 0, (0,1,0) corresponds to i,
and (!,~, ~) corresponds to (~J3 + 2V2)(1 + i).

14.18 Show that Definition 14.18 and Definition 9.4c are equivalent. To
do this, let Xl, X2, X3, . •• be a sequence of elements in a metric
space, then show that for every E > 0, there exists an integer
N such that d[x, Xk] < E whenever k ~ N if and only if every
neighborhood of X contains an element of the sequence.

14.19 Show that Definition 14.19 and Definition 14.2 are equivalent.

14.20 Let T(z) = az +db be a Mobius transformation. In particular,

ez+
ad - be =f. 0.
a) Show that T is defined and continuous on the extended complex
plane.
Note: T(oo) = ~ and T(zo) = 00 if an only if ezo + d = O.
b) Show that T : C U {oo} -+ C U {oo} is one-to-one and onto.
c) Show that T is a homeomorphism of the extended complex plane.
 Exercise Set 14 179

14.21 Prove Theorem 14.20.

Hint: Mimic the proof of Theorem 12.2.

14.22 Prove that Newton's method for J(z) = z2 - 1 is chaotic on the

imaginary axis by completing the details of the argument in Ex-
ample 14.21.

14.23 Let T be a topological conjugacy between functions which are de-

fined on the extended complex plane. Show that T maps stable
sets of fixed points to stable sets of fixed points without using a
metric on 8 2 .

14.24 Prove Cayley's Theorem (Theorem 14.22) which states that if the
complex quadratic polynomial q(x) has two distinct roots, then
N q (x) is chaotic on the perpendicular bisector of the line segment
joining the two roots. Further, the stable set of each root under
iteration of Nq(x) is the set of points which lies on the same side
of the bisector as the root.
Hint: Use one of the suggestions given after the statement of
the theorem in the text.

14.25 a) Prove that if a complex quadratic polynomial has exactly one

root, then all points in C converge to the root under iteration of
Newton's function for the polynomial. (This is Theorem 14.23.)
b) Describe the dynamics of Newton's function for J(z) = z3 and
show that your description is correct. You should be able to com-
pletely describe the dynamics on C.
c) Characterize the dynamics of Newton's function for any cubic
polynomial with exactly one root of multiplicity three and show
that your characterization is correct.
d) Generalize the statement of part (a) to higher degree polynomi-
als and prove that your generalization is correct.

14.26 AN INVESTIGATION: The polynomial J(z) = z3 - Z2 has exactly

two roots: 0 is a root of multiplicity two and 1 is a root of multiplic-
ityone. Describe the dynamics of Nf to the best of your ability. If
possible, generalize your results to the set of all polynomials with
one root of multiplicity two and one root of multiplicity one.

14.27 AN INVESTIGATION: In Example 12.13 we found that the interval

of the real line [-.03, .03] is attracted to a period two point of
180 14. The Dynamics of Complex Functions

Newton's function for J(z) = Z3 - 1.265z + 1. What other points

of C are attracted to this period two point? Can you prove that
there is a disk in C that is attracted to this period two point? This
would imply that the set of points which is attracted to a root of
J under iteration of Nf is not dense in C. Why? Why is this
important?
Can you find a parameter value c for which Newton's function
for J(z) = z3 - cz + 1 has an attracting prime period four point?
Can you find an open set of parameter values for which this is true?
**Is there a parameter value c for which Newton's method ap-
plied to J(z) = z3 - cz + 1 is chaotic on a set which contains an
open subset of C? Note: It suffices to show that an iterate of Nf is
conjugate to q( z) = z2 - 2 for some c. You should also explain why
this is sufficient as well. Is there an open set of parameter values
for which Nf is chaotic on a subset of C?
15
The Quadratic Family
and the Mandelbrot Set

We return once again to the study of the dynamics of quadratic func-

tions. In this chapter we consider the quadratic family qc(z) = z2 + c. We
demonstrated in exercise 10.4 that all real quadratic functions are topolog-
ically conjugate to a real polynomial of the form qc(.T) = x 2 + c for some
c. This fact extends to the complex quadratic polynomials; all complex
quadratic polynomials are topologically conjugate to a polynomial of the
form qc(z) = z2 + c. We will take direction for our study of the quadratic
family from our previous work with the logistic map hr(x) = rx(1 - x).
Recall that the interesting points in the dynamics of hr were those points
which remained in the interval [0, 1] for all time. Points which left this
interval after a finite number of iterations were in the stable set of infinity.
Given this, it seems reasonable to suppose that the interesting points in the
dynamics of qc are those which have bounded orbits. An orbit is bounded
if there exists a positive real number such that the modulus of every point
in the orbit is less than this number. The following proposition shows that
the orbits of qc do in fact divide up into two sets: those which are bounded
and those which converge to infinity.

PROPOSITION 15.1. Either the orbit of a complex number under itera-

tion of a complex quadratic polynomial is bounded or the number is in the
stable set of infinity.
182 15. The Quadratic Family and the Mandelbrot Set

PROOF. Since all complex quadratic polynomials are topologically con-

jugate to a function of the form qc(z) = z2 + e it suffices to prove the
proposition for this family of functions. (In making this assumption we
are using the fact that if ¢ : C --+ C is a homeomorphism and K c C is
bounded, then ¢(K) is bounded. We haven't proven this and interested
readers are encouraged to do so. A proof can be found in the analysis text
by W. Rudin, which is listed in the references.)
We have already shown the proposition holds for q(z) = z2 (Example
14.15). Let e be a nonzero complex number and q(z) = z2 + e. We begin
by showing that if w is a complex number satisfying Iwl > lei + 1, then w
is in the stable set of infinity.
Since q(w) = w 2 + e,

Iq(w)1 = Iw2+ el ;::: Iw21-lel

;::: (lei + 1)2 - lei = 1c1 2+ lei + 1.
Then,

Iq2(W)1 = Iq(q(w))1
= I(q(W))2 + el
;::: Iq(w)12 - lei
;::: (1c1 2+ lei + 1)2 - lei
= lel 4 + 21el 3 + 31el 2+ lei + 1
;::: 31el 2+ lei + 1.
Continuing by induction we show that

Since the right-hand side of this inequality tends to infinity as n gets larger,
our assertion that w is in the stable set of infinity whenever Iwl > lei + 1
must hold. It follows that either the orbit of a complex number under
iteration of q(z) = z2 + e is bounded by lei + 1 or the number is in the
stable set of infinity. 0
The set of points with bounded orbits is interesting enough to have
earned its own name.

DEFINITION 15.2. The set of points whose orbits are bounded under it-
eration of q(z) = Z2 + e is called the filled Julia set of q and denoted by Kc.
The boundary of a filled Julia set is called the Julia set. The boundary of
a set is the collection of points for which every neighborhood eontains an
element of the set as well an element which is not in the set.
 15. The Quadratic Family and the Mandelbrot Set 183

K ..719 •.251

FIGURE 15.1. The filled Julia sets for c 0, c -1,

c = i, and c = -.719 - .25i.
Julia sets are named after the French mathematician Gaston Julia, who
along with his countryman Pierre Fatou, did much of the fundamental work
on the dynamics of complex functions early in this century. Their results
are all the more remarkable given that they did not have the power of
computers to help them visualize the sets they were considering. The filled
Julia sets for four parameter values are shown in Figure 15.l.
In Chapter 8, we defined Cantor sets as subsets of the real numbers
which are closed and bounded, contain no intervals, and for which every
point is an accumulation point of the set. We extend this definition to
subsets of the complex plane.

DEFINITION 15.3. A set of complex numbers is a Cantor set if it is closed

and bounded, totally disconnected, and if every point is an accumulation
point of the the set. A set of complex numbers is totally disconnected if there
does not exist any path between distinct poi-n ts in the set that is completely
contained in the set.
184 15. The Quadratic Family and the Mandelbrot Set

Fatou and Julia proved many interesting theorems. We stated a special

case of one of their results in Theorem [Link] general statement is given
in the following theorem.

THEOREM 15.4. If a complex polynomial has an attracting periodic or-

bit, then there must be a critical point of the polynomial in the stable set of
one of the points in the orbit.

We recall that Zo is a critical point of the polynomial p if p'(zo) = o. As

the only critical point of qc(z) = Z2 + cis 0, this implies that qc can have
at most one attracting periodic orbit and that 0 must be attracted to it.
A proof of Theorem 15.4 can be found in the second edition of Devaney's
text, Introduction to Chaotic Dynamical Systems, which is listed in the
references. The role of the critical point is again highlighted in the following
theorem of Fatou and Julia.

THEOREM 15.5. If 0 is in the filled Julia set of qc(z), then the filled
Julia set is connected. That is, given any two points in the filled Julia set,
there is a path contained in the set which connects those two points. On the
other hand, if 0 is not in the filled Julia set of qc, then the set is a Cantor
set.

Proposition 15.1 and Theorem 15.5 naturally divide the complex num-
bers into two sets: those parameter values c for which the filled Julia set is
connected and the orbit of 0 is bounded under iteration of qc (z) = z2 + c;
and those for which the filled Julia set is totally disconnected and 0 is in
the stable set of infinity. The set we investigate here is the former. More
specifically, we are interested in

M = {c I q~(O) is bounded}.
While we will not prove Theorem 15.5, in exercises 15.4 and 15.5 the
reader is invited to prove that the filled Julia set is a Cantor set and that
q~(O) tends to 00 when Icl > 2. This implies that the set M is contained
inside a disk of radius 2. In exercise 15.3 we show that the set of param-
eter values for which qc has an attracting fixed point is the interior of the
cardioid parametrized by p(O) = ~eei - ~e2ei. By Theorem 15.4, it follows
that this cardioid is contained in M since 0 must be in the stable set of an
attracting fixed point. In fact, if we define

Mk = {c I qc(z) has an attracting periodic point with prime period k},

then
00
 15. The Quadratic Family and the Mandelbrot Set 185

In exercise 15.3 we also show that M2 is the disk with radius ~ and center
at -1. It is interesting to try to find the other sets Mk.
So far we have stated that M is contained inside the circle of radius
2, M contains the interior of the cardioid p(O) = ~el1i - ~e2I1i, and M
contains the interior of the disk with center at -1 and radius ~. To get a
better picture of M we conduct another computer experiment. We know
that M consists of the parameter values for which the orbit of 0 is bounded
and that if Iq~(O)1 > 2 for any n, then c is not in M. Thus, we cover the
disk of radius 2 by a grid, use each point in the grid as a parameter c, and
calculate q~OO(O). If Iq~OO(O)1 ::::: 2, we assume that c is in M and color the
corresponding grid point black. If Iq~OO(O)1 > 2, we leave the grid point
white. The results of such a program are shown in Figure 15.2.
The reader may recognize the sketch of M in Figure 15.2 as the famous
Mandelbrot set, the first images of which were published by Benoit Man-
delbrot in 1980. The Mandelbrot set has been studied intensely since its
discovery and much is known about it, though there is still more to be
learned. It is known that it is connected, and the manner in which the
various bulbs are connected is also understood. The details of this connec-
tion are well beyond the scope of this book. An interesting and accessible
discussion of the properties of the Mandelbrot set can be found in Chapter
14 of the book Chaos and Fractals by Peitgen, Jurgens, and Saupe. It is
the author's understanding that a more complete account of the properties
of the Mandelbrot set will be included in the forthcoming book on complex
dynamics by J. Milnor. Both books are listed in the references.

Generating Julia and Mandelbrot Sets on a Computer. There

are a number of widely available computer programs which generate pic-
tures of Julia sets and the Mandelbrot set. A few of these are listed in the
references. Individuals with a minimal amount of programming knowledge
will find it easy and instructive to write their own programs. An outline
of the Mathematica programs the author used to create the illustrations
shown in this text are listed in the appendix as models of how such pro-
grams might be structured. M athematica is not recommended for creating
images of the Julia and Mandelbrot sets as it is not very fast for this kind
of computation. For example, several of the images created for this text
took more than four hours to run in M athematica. Similar programs writ-
ten in C or even Basic run in a small fraction of the time. As indicated by
the preceding statements, the most important factor to consider in writing
a program is speed. We outline some of the considerations for efficiently
computing the Mandelbrot set here and note that there are similar consid-
erations for Julia sets.
If we wish to create a picture of the Mandelbrot set with a resolution
of 256 by 256, and wish to do so by determining if the 50th iterate of qc is
186 15. The Quadratic Family and the Mandelbrot Set

2r---~----------~----~--~----~--------~

1.5

0.5

-0.5

-1

-1.5

-2~--~----~----~----~--~----~----~--~
-1.5 -1 -0.5 0 0.5 1 1.5 2

FIGURE 15.2. The set of parameter values c for which

Iq~OO(O)[Link]; 2. This set is a good approximation of M.
The image contains the cardioid p(O) = ~elJi - ie20i and
i
the disk with radius and center at -1 as predicted by
our calculations. Note that while M appears not to be
connected in this figure, it has been proven that M is con-
nected. This was done by A. Douady and J. H. Hubbard
in the mid 1980's.
 Exercise Set 15 187

greater than 2, then we would be foolish indeed to do so by iterating each

of the 65,536 points in the grid 50 times. That would involve more than
3.2 million iterations. It is quicker to check after each iteration to see if the
modulus of that iterate is larger than 2. If it is, then the point is not in the
Mandelbrot set and we can stop. We can use this extra check to increase
the beauty of our picture by coloring points which leave before 5 iterations
one color, points which leave after 5 but before 10 iterations another color,
and so on. A further increase in speed can be gained by automatically
including any point which lies inside the cardioid z = ~elii - ie2Bi or the
disk with radius i and center at -1. Finally, there is no substitute for
clever programming. For example, if we wish to know whether la + bil < 2,
then it is computationally more efficient to check whether a 2 +b2 < 4 rather
than va 2 + b2 < 2. With these long computations one is well advised to
think before writing.
There are algorithms for computing Mandelbrot and Julia sets other
than those described here. Several of the alternatives are discussed in the
text by Peitgen, Jurgens, and Saupe, which is listed in the references.
Happy programming and enjoy the exploration of complex analytic dy-
namics!

Exercise Set 15

15.1 Show that all complex quadratic polynomials are topologically con-
jugate to a polynomial of the form qc(z) = Z2 + c.
Hint: See exercise 10.4.

15.2 Complete the details of the induction argument in the proof of

Proposition 15.1.

15.3 a) Show that the set of parameter values c for which qc(z) = Z2 +c
has an attracting fixed point is the interior of the cardioid
1 t 1 2n"
p(B) = _eofl"
- -e ot.
2 4
Hint: We wish to find c so that there exists Zo = re lii such that
z5 + c = Zo and Iq~ (zo) I < 1. Substitute for Zo and solve for c.
b) Show that the set of parameter values for which qc(z) has an
attracting periodic point with prime period two is the interior of
i
the disk with radius and center at -1.
**c) Describe the set of parameter values for which qc(z) = Z2 + c
has an attracting periodic point with prime period k for as many
188 15. The Quadratic Family and the Mandelbrot Set

other k as you can. Be forewarned that this is very difficult and

perhaps impossible to do analytically. Use a computer and experi-
ment. Remember that we have already studied the dynamics of the
real function qc(x) = x 2 +e (in exercise 11.5 and that those real pa-
rameter values for which we have attracting period k points of the
real function are also parameter values for the complex function.
Good luck!!
An answer to this question can be found in Chapter 14 of the
book by Peitgen, Jurgens, and Saupe, which is listed in the refer-
ences.

The following two exercises demonstrate that if Icl > 2, then 0 is

not in the filled Julia set and the filled Julia set is a Cantor set.
This lends support to the statement of Theorem 15.5.

15.4 Let qc(z) = Z2 + c.

a) Prove that if Izl :2: Icl
and Izl > 2, then Iqc(z)1 > Izl and z is in
the stable set of infinity.
b) Use part (a) to show that if Icl > 2, then q~(O) tends to 00.

15.5 LOOKING FOR CANTOR DUST: It has been suggested that Cantor
sets in the complex plane be called Cantor dust since, like a cloud
of dust, no two points are connected yet there is another point
close to every point in the cloud. In this exercise we show that the
filled Julia set of the function qc(z) = Z2 + c forms such a Cantor
cloud when lei> 2.
Let Icl > 2 and define q(z) = Z2 + c. Let D be the circle with
radius Icl and center at the origin.
a) Use exercise 15.4a to show that the filled Julia set is contained
inside D. In fact, Kc lies inside q-n(D) for all n.
b) Show that q-l(D) is the figure eight with crossing at the origin
shown in Figure 15.3.
c) Prove that q-2(D) must be inside q-l(D) and that q-2(D)
consists of two figure eights as demonstrated in Figure 15.3.
d) Prove that q-n(D) consists of 2n figure eights situated as in
Figure 15.3. Further, show that as n goes to infinity the diameter
of the figure eights in q-n(D) converges to O.
 Exercise Set 15 189

FIGURE 15.3. The inverse images of the disk with radius

2.5 under the map Q2.5. The filled Julia set is the inter-
section of all the inverse images of the disk. Successive
inverse images are drawn lighter. The nth inverse image
consists of 2n figure eights with one in each of the lobes of
the 2n - 1 figure eights in the (n - l)st inverse image.
Appendix A
Computer Algorithms

Most of the computations shown in this text were done by Mathematica

on a NeXT computer. Mathematica is very good at computing tables of
values of fn(x), graphing functions, and graphical analysis. Unfortunately,
it is not a particularly good language to use for graphics intensive projects
like computing bifurcation diagrams, the Mandelbrot set, or Julia sets.
The most efficient way to generate these types of graphics is to plot the
points as you go. With Mathematica one saves the graphic information and
prints it all at the end. As some of the lists generated contain hundreds of
thousands of elements, this is extremely inefficient. Since the Mathematica
code used to create the images in the text can easily be converted to another
language, they are included here for reference.

A.I. Iterating Functions

Finding the Value of a Point Under Iteration. The easiest way
to iterate functions with Mathematica is to use the Nest command. The
format is
Nest [function, start value, number of iterations].

Below we define f(x) = x 3 and then iterate f three times starting at 1.2.
Clear[f] ;
f [x_] ;= x-3;
192 A. Computer Algorithms

Nest [f, 1. 2 ,3]

The output is 137.371.
NestList [ ... ] will show the value of the first n iterates. The format
is similar to that of Nest. The following commands compute the first three
iterates of 1.2 under f(x) = x 3 :
Clear[f] ;
f [x_] := x-3;
NestList[f,1.2,3]
The output is {1.2, 1.728, 5.15978, 137.371}.

Tables of Iterates. We use the next set of commands to generate a

table of iterates. In the first two lines we define the function. Next we tell
the computer what the starting value is. The second and third variables tell
the computer which iterations to print. In the example which follows we are
telling the computer to iterate the function h(x) = 3.5x(1 - x) beginning
at .1 and to print the values of the 10th through the 20th iterates.
Clear[h];
h[x_] := 3.5x (1 - x);

StartingValue = .1;
Firstlteration = 10;
Lastlteration = 20;

i=O;
Y = N[StartingValue];
While[i <= Lastlteration,
If[i >= Firstlteration,Print[i," ",N[y,8]]];
y = h[y] ;
i = i+l]
The output of this program is as follows:
10 0.86414351
11 0.41089828
12 0.84721309
13 0.45305075
14 0.86728519
15 0.40285557
16 0.84197036
17 0.46569696
18 0.87088155
19 0.39356405
20 0.83534986
 A.I. Iterating Functions 193

Controlling the Precision of the Computations. For some of the

exercises in this text we need be able to control the precision of the com-
putations. We use the Mathematica command SetPrecision which writes
an expression to n significant digits, adding zeros or rounding as necessary.
The format is
SetPrecision[expression, n]
Zeros which are added are in binary so the base ten representation of the
new number will not necessarily end in zeros. In the commands that follow,
the variable SigDigi ts controls the amount of rounding.
Clear[h] ;
h[x_]:= 3.5x(1-x);

StartingValue = .1;
Firstlteration = 10;
Lastlteration = 20;
SigDigits = 64;

i=O;
y = SetPrecision[StartingValue,SigDigits];
While[i <= Lastlteration,
If [i >= FirstIteration, Print [i, " " ,N [y, 8]]] ;
y = SetPrecision[h[y] ,SigDigits];
i = i +1]
As the function h(x) = 3.5x(1 - x) is not sensitive to small changes in
initial conditions, the output of this program is the same as that of the
preceding program.

Graphing Iterated Functions. For some of the exercises we need to

graph the iterate of a function. For example in exercise 8.7, we are asked
to graph T2 where

for x in [0, ~]
T(x) = {22x,
2x, for x in [~, 1].

We accomplish this with the following program:

Clear[T] ;
T[x_]:= If[x<=.5,2x,2 - 2x];

xmin O',
xmax 1;

NumberOflterations 2',
194 A. Computer Algorithms

Plot [{Nest [T,x,NumberOflterations] ,x},

{x, xmin, xmax} ,
PlotRange->{xmin,xmax},
AspectRatio->1]
Note the use of an If [ ... ] statement to define T(x). The variables xmin
and xmax denote the endpoints of the domain. NumberOfIterations = 2
tells the program to compute the graph of T2. If we replace the 2 by 3 it
will graph T3. We use Plot and Nest to draw the actual graph. Since we
are lookin g for fixed points of T2 we draw in the line y = x as well. In the
second line of the Plot command we define the domain of the function to be
graphed. PlotRange->{xmin,xmax} tells the computer that the codomain
shown should be equal to the domain. Finally, AspectRatio->1 makes the
graph square. The output is shown in Figure A.I.

FIGURE A.I. The graph of T2 and y = x as generated by

the program listed above.

A.2. Graphical Analysis

One of the most powerful tools available for investigating the dynamics of
functions of the real numbers is graphical analysis. The following program
was used to produce Figure A.2a. We use graphical analysis to analyze the
itinerary of .1 under iteration of h(x) = 3.5x(1 - x). The first 20 iterates
are shown. Lines enclosed by (* ... *) are comments.
 A.2. Graphical Analysis 195

Clear[h];
h[x_] := 3.5x (1-x)

StartingValue = .1;
FirstIt = 0;
LastIt = 20;

xmin = 0;
xmax = 1;

(* Finding the Points *)

i 0;
y = N[StartingValue] ;

While[i < FirstIt, y = h[y]; i = i + 1];

DataTable = {{y,y},{y,h[y]}};

While[i < LastIt,

y = h[y] ;
AppendTo[DataTable,{y,y}] ;
AppendTo[DataTable,{y,h[y]}] ;
i =i + 1];

AppendTo[DataTable,{h[y] ,h[y]}] ;

(* Drawing the Cobweb *)

Cobweb = ListPlot[DataTable,PlotJoined -> True,

PlotRange - > {{xmin, xmax} , {xmin, xmaxH ,
AspectRatio -> 1,PlotStyle->GrayLevel[.3],
DisplayFunction -> Identity];

(* Drawing the graphs of h and y=x *)

Graph = Plot [{h[t] ,t},{t,xmin,xmax},

PlotRange -> {xmin,xmax},AspectRatio -> 1,
DisplayFunction -> Identity];

(* Displaying the result *)

Show[Cobweb,Graph,DisplayFunction -> $DisplayFunction]

196 A. Computer Algorithms

a) b)

FIGURE A.2. Graphical analysis of the itinerary of .1 un-

der iteration of h(x) = 3.5x(1 - x).

To determine the long-term itinerary of .1 under iteration of h we change

the values of FirstIt and LastIt in the above program to 100 and 120,
respectively. Then the first 100 iterates are calculated and 20 iterates of the
cobweb are drawn beginning with h lO0 (.1). The result is shown in Figure
A.2b; apparently, .1 is attracted to a period 4 orbit.

A.3. Bifurcation Diagrams

The following "pseudocode" will generate bifurcation diagrams like the
one shown in Figure 12.10. With minor changes it can be implemented
on most computer systems. The programming conventions used are most
similar to those of M athematica. In particular, the roles of semicolons
and commas in While statements are the re verse of those used in C. A
M athematica program similar to the one below was used by the author
to create the bifurcation diagrams in Chapter 12. As the points had to
be saved and then plotted all at the same time using ListPlot, it took
several hours to create one figure. Plotting the points as you go will save
memory and a lot of time. This program should run in a few minutes
if it is implemented well. If the author hadn't had a large network of
NeXT machines at his disposal he would have written something in a faster
language. In general, it is wise to keep the number of iterates and parameter
steps small until one gets some idea of the length of time it will take the p
rogram to run. Comments in the code are preceded by a %.
 A.4. Julia Sets 197

%We define the function with one parameter, r.

%We define the variables.

% Variable names are similar to ones we've used before.
% rmin and rmax are parameter bounds.
% Norsteps is number of parameter values used.

First1t = 100
LastIt = 200
Seed = .5
rmin = 2
rmax = 4
Norsteps 250
xmin = 0
xmax = 1

%Now we do the computations.

%Points are plotted by the third While[ ... ] statement.
rstep = (cmax - cmin)/Norsteps
p = rmin;
While [p <= rmax,
i = 0;
y = N[seed];
While[i < First1t, y = h[y,p]; i i + 1];
While[i <= Last1t,
Plot{p,y};
1f[y == h[y,p],Break];
i = i + 1];
p = p + rstep]

A.4. Julia Sets

A "pseudocode" for drawing Julia sets of qc{z) = z2+ e is outlined below.
Recall that for a fixed e, the Julia set of qc is the set of points that are not
in the stable set of infinity. If Iq;,"(z) I is larger than both 2 and lei for some
n, then z is in the stable set of infinity. Comment lines in the program are
preceded by a %.
198 A. Computer Algorithms

% The variables realC and imaginC are the real and

% imaginary parts of the parameter. Modsq is the modulus
% of the parameter squared. xmin and xmax are the bounds
% of the real parts of the initial points. xgrid is the
% number of steps used to get from xmin to xmax.

realC = 0
imaginC = 1
modsq = (realC)-2 + (imaginC)-2

xmin = -2
xmax = 2
ymin = -2
ymax = 2
xgrid 250
ygrid = 250

iterations = 25;

% The computations start here.

% Plot[Rectangle[[{x,y},{x+xstep,y+ystep}]]] is an
% instruction to color the rectangle with lower left corner
% (x,y) and upper right corner (x+xstep,y+ystep) black.

xstep (xmax - xmin)/xgrid

ystep (ymax - ymin)/ygrid

For [x xmin, x < xmax, x = x+xstep,

For[y ymin, y < ymax, y y+ystep,
u = (x + xstep/2);
v = (y + ystep/2);
For [i=O, i<=iterations, i++,
If[i == iterations,
Plot[Rectangle[[{x,y},{x+xstep,y+ystep}]]]];
If[(u-2 + v-2) > Max [4,modsq] ,Break[]];
utemp = u-2 - v-2 + realC;
v = 2*u*v + imaginC;
u = utemp
]]J

There are other methods for computing images of Julia sets. One utilizes
the fact that an equivalent definition of a Julia set for the parameter c is
the closure of the set of repelling periodic points of qc (z) = z2 + c. Another
 A.S. The Mandelbrot Set 199

uses the fact that for any positive €, we know the circle with radius 1 + Icl +€
is not in the filled Julia set Kc. Successive inverse images of this circle are
computed. As these inverse images must converge to Kc as the number of
iterations goes to infinity, we can often get a good picture quickly. Both of
these methods are discussed in the book by Peitgen, Jurgens, and Saupe,
which is list ed in the references.

A.5. The Mandelbrot Set

The following Mathematica program was used to create Figure 15.2. It
is included for the readers information. This version took approximately
five hours to run on a NeXT workstation. Again, the major slowdown
was creating and storing the huge list of graphics primitives which were all
invoked at the end:
Clear[q] ;
q[z_,p_] := z~2 + p;

xmin = -2;
xmax 2;
ymin -2;
ymax 2;
xgridpoints = 400;
ygridpoints = 400;
iterations = 75;

xstep N[(xmax - xmin)/xgridpoints];

ystep N[(ymax - ymin)/ygridpoints];

mandelbrot = {};

For[x = N[xmin], x < xmax, x = x+xstep,

For[y = N[ymin], y < ymax, y = y+ystep,
c = (x + xstep/2) + (y + ystep/2) I;
While[(Re[c] + 1)~2 + Im[c]~2 < .0625,
AppendTo[mandelbrot,
Rectangle [{x,y},{x+xstep,y+ystep}]] ;
y = y + ystep;
c = (x + xstep/2) + (y + ystep/2) I];
r = Abs[c];
While[r < Abs[(.5/r) c - (.25/r~2) c~2],
AppendTo[mandel~rot,
Rectangle[{x,y},{x+xstep,y+ystep}]];
200 A. Computer Algorithms

y y + ystep
c = (x + xstep/2) + (y + ystep/2) I;
r = Abs [c]
w = c;
For [i=O, i<=iterations, i++,
If[i == iterations,
AppendTo[mandelbrot,
Rectangle [{x,y},{x+xstep,y+ystep}]]] ;
If[Re[w]-2 + Im[w]-2 > 2,Break[]];
w = q[w,c]
]]]

Show [Graphics [mandelbrot] ,AspectRatio->1,

PlotRange->{{xmin,xmax},{ymin,ymax}}]
An alternative method for computing the Mandelbrot set is discussed in
the book by Peitgen, Jurgens, and Saupe, which is listed in the references.
It uses an inverse iteration scheme similar in concept to the one described
above for creating Julia sets.

A.6. Stable Sets of Newton's Method

The following Mathematica program was used to create Figure 14.8. It
is included here for the reader's information:
Clear[f] ;
f[z_] := (2z-3 + 1)/(3z-2);

xmin -2;
xmax 2;
ymin -2;
ymax 2;
xgridpoints = 300;
ygridpoints = 300;
iterations = 50;

xstep N[(xmax - xmin)/xgridpoints];

ystep N[(ymax - ymin)/ygridpoints];

stable1 {GrayLevel[.8]};
stable2 {GrayLevel [.4] };
stable3 {GrayLevel[O]};
 A.6. Stable Sets of Newton's Method 201

For[x = N[xmin], x < xmax, x = x+xstep,

For[y = N[ymin], y < ymax, y = y+ystep,
w = (x + xstep/2) + (y + ystep/2) I;
For [i=O, i<=iterations, i++
If[(Re[w]- 1)-2+Im[w]-2<= .0625,
AppendTo[stable1,
Rectangle[{x,y},{x + xstep,y + ystep}]];
Break[]] ;
If[(Re[w]+.5)-2 + (Im[w]-.866)-2<=.0625,
AppendTo[stable2,
Rectangle[{x,y},{x + xstep,y + ystep}]];
Break[]] ;
If[(Re[w]+.5)-2 + (Im[w] + .866)-2<=.0625,
AppendTo[stable3,
Rectangle[{x,y},{x + xstep,y + ystep}]];
Break[]] ;
If[w==O.O + [Link],Break[]];
w = fEw]
]]]

Show[Graphics[{stable1,stable2,stable3}],AspectRatio->1]

The boundaries ofthe stable sets of the fixed points of Newton's function
for a cubic polynomial are the closure of the set of repelling periodic points.
By exploiting this fact, one can write algorithms which will compute these
boundaries efficiently. We again refer the interested programmer to the
book by Peitgen, Jurgens, and Saupe for suggestions as to how this might
be done.
References

The items in this bibliography are loosely arranged by topic. In general,

only works with which I have had substantial experience are included; this
list is by no means exhaustive. My personal reactions to the texts are
also included where appropriate. In addition to sources of information
on dynamical systems, I have included references on other mathematical
topics mentioned in the text, general interest titles, and a note on computer
resources.

Dynamical Systems. New books and articles on dynamical systems

appear regularly. Good books on the topic include the following:

DEVANEY, R. An Introduction to Chaotic Dynamical Systems, 2nd edition,

Addison-Wesley, 1989
My text arose out of the desire to provide my students the mathematical
background necessary to understand this text by Devaney. His book is
good, but demands sophistication on the part of the reader. In general,
the notation and terminology of my book are consistent with his and
the reader should find his book to be a reasonable continu ation of the
ideas contained in mine.
204 References

DEVANEY, R. A First Course in Chaotic Dynamical Systems: Theory and

Experiment, Addison-Wesley, 1992
This text by Devaney covers roughly the same ground as mine, but is
written from a different point of view. Individuals interested in writing
programs to do computations in dynamics may wish to consult Appendix
B of Devaney's book which contains True Basic programs that will do
all of the computations necessary to complete the exercises in my text
as well as his.

HIRSCH, M. W. AND SMALE, S. Differential Equations, Dynamical Sys-

tems, and Linear Algebra, Academic Press, 1974
Continuous dynamical systems arise out of the qualitative study of dif-
ferential equations. This text is an excellent treatment of the qualitative
analysis of differential equations and introduces concepts in continuous
dynamical systems which are analogous to the concepts we have seen
in my book. Hirsch and Smale's book does require sophisticat ion on
the part of the reader, but it should be accessible to an advanced under-
graduate and is worth the effort. The intertwining of dynamical systems
theory and linear algebra with the theory of differential equations, mo-
tivates their study. Recommended reading for any student interested in
pursuing the study of dynamical systems.

HUBBARD, J. AND WEST, B. Differential Equations: A Dynamical Sys-

tems Approach, Springer-Verlag, 1991
This is a three-volume text which explores the qualitative behavior of
differential equations. The fifth chapter (in volume 1) does a good job
of introducing numerical methods of solving differential equations as
discrete dynamical systems. Limitations of computer implementations
of numerical approximations to solutions of differential equations are
discussed in Chapter 3. There is a proof of Theorem 12.15 from my text
in Section 5.3.

MILNOR, J. Dynamics in one complex variable: Introductory Lectures,

Stony Brook Institute for Mathematical Sciences Preprint #1990/5
I haven't seen these notes, but understand that they are a fairly complete
set of notes suitable for a first-year graduate level course on complex
dynamics.
 References 205

PEITGEN, H. AND SAUPE, D., EDS. The Science of Fractal Images,

Springer-Verlag, 1988
This book should be informative and interesting to anyone with the
mathematical sophistication necessary to read my text. It contains nu-
merous algorithms for computing fractal images. In particular there are
several different algorithms for computing Julia sets and the Mandelbrot
set.

PEITGEN, H. JURGENS, H. AND SAUPE D. Chaos and Fractals: new fron-

tiers in science, Springer-Verlag, 1992
An encyclopedic account of many current topics in chaotic dynamics
and fractals. I haven't had a lot of time to look at this book, but have
enjoyed those portions which I have read. The last five chapters extend
the material covered in my text. In particular, the chapters on Julia
sets and the Mandelbrot set present many of the recently discovered
properties of these sets. They also include good, efficient programs in
Basic for computing images of these sets. Recommended reading for
individuals interested in learning more about fractals, Julia sets, and
the Mandelbrot set. The graphics are beautiful and it is a fun book to
peruse.

The College Mathematics Journal, Vol. 22, No.1, January 1991,

This volume of CMJ was dedicated to discrete dynamical systems and
contains several good articles. Software and computer exercises using
ideas from dynamics are also discussed.

Topics in Mathematics.
BAK, J. AND NEWMAN, D. J. Complex Analysis, Springer-Verlag, 1982
BELDING, D. AND MITCHELL, K. Foundations of Analysis, Prentice-Hall,
1991
CHURCHILL, R. V. AND BROWN, J. W. Complex Variables and Applica-
tions, 5th edition, McGraw-Hill, 1990
OSTROWSKII, A. M. Solutions of Equations and Systems of Equations, 2nd
edition, Academic Press, 1966
RUDIN, W. Principles of Mathematical Analysis, 3rd edition, McGraw-Hill,
1976
WILLARD, S. General Topology, Addison-Wesley, 1970
206 References

The text by Rudin is an excellent comprehensive introduction to math-

ematical analysis in ]Rn and demands a relatively high degree of math-
ematical sophistication on the part of the reader. It is definitely on my
top ten favorite mathematics textbooks list. The text by Belding and
Mitchell is not as comprehensive, but it is a good, reader-friendly intro-
duction to real analysis. Student consensus is that it is a great book for
the first time student of analysis.
The complex analysis texts are both good introductions to complex anal-
ysis and are accessible to advanced undergraduates.
Willard's textbook on topology is intended for beginning graduate stu-
dents. Most of the topological ideas covered in my book are adequately
covered in the analysis books already mentioned, but individuals inter-
ested in all of the details will find Willard's text useful. Personally, I
think general topology is fascinating.
Ostrowskii's text is an extensive treatment of numerical methods for
solving equations. It is included here since it contains a proof of Theorem
12.15.
General Interest Books on Dynamics.
Gleick, J. Chaos: making a new science, Viking, 1987
This book has had a wide readership and students in my courses on
dynamics have found it to be good companion reading for a course in
dynamics. I prefer the book by Stewart, which is listed below, but there
is room for debate. The review of Gleick's book by John Franks and
the subsequent responses in The Mathematical Intelligencer are worth
reading. They help put Gleick's book in perspective and provoke thought
about the nature of mathematics. The review and a response are in Vol
11, number 1 (winter 1989), pages 65-71. Discussion continues in Vol
11, number 3 (summer 1989), pages 3-13.

STEWART, I. Does God Play Dice?: the mathematics of chaos, Blackwell,

1989
This is a very good and readable discussion of the implications of mod-
ern dynamical systems and chaos theory on the development of science.
Highly recommended as companion reading for a course in dynamical
systems or differential equations.
Computer Programs and Algorithms. Most of the exercises in this
text can be done using Mathematica. Mathematica is a computer algebra
system developed by Wolfram Research of Champaign, Illinois. The pro-
grams and commands necessary to use Mathematica successfully with this
textbook are listed in the appendix.
 References 207

There are a multitude of other options one can pursue. Probably the
easiest choice for those with some computer experience is to write their
own programs. None of the programs for doing the dynamics experiments
mentioned in this text need be more than 20 or 30 lines long. Many will
have less than 10. To write the programs one needs to be able to input
initial data to a program, construct loops to do the iteration, and output
graphics information to the screen; that's it. Several of the books listed
above in the section on dynamics include computer programs. True Basic
programs which can easily be converted into other languages are included
in appendix B of the text by Devaney. Numerous programs are included
in the book edited by Peitgen and Saupe. Efficient programs in Basic for
computing Mandelbrot and Julia sets are found in the book by Peitgen,
Jurgens, and Saupe.
For those who want a canned package the alternatives are almost endless.
There are probably hundreds of public domain packages which will do the
job. Check local bulletin boards or use Archie and search the internet. By
the time this book appears in print, 1 expect to have located or written
programs for the NeXTstep environment which will do the computations
necessary for the exercises in my text. The software will be compatible with
both the original NeXT hardware and the intel version. Those interested
in getting a copy (free or close to it) may contact me at the Mathematics
Department, Allegheny College, Meadville PA 16335. My internet address
is: rholmgre@[Link]. I accept NeXT mail. Anyone who has already
written such programs for the NeXT or knows of such programs is especially
encouraged to contact me. (I know that there is a program to compute the
Mandelbrot set in the NextDeveloperjDemos folder, thanks. If you have a
NeXT and haven't seen it, have a look; it's a nice program.)
Below 1 have included a reference for a workbook and software which
is intended to serve as a supplement to Devaney's text by the same name
(listed above). While I have never seen it, I believe it contains programs
capable of completing all the computations necessary for completing the
exercises in my text. It requires a MacIntosh computer running system
6.0.5 or higher. To get adequate speed one should have a Mac II and
a math coprocessor. For the exact system requirements, please contact
Addison-Wesley.
GEORGES, J. JOHNSON, D. AND DEVANEY, R. L. A First Course in
Chaotic Dynamical Systems: Laboratory Addison-Wesley, 1992
Index

accumulation point 24, 80 circle, doubling map on 91, 124, 138,

equivalent conditions for 24, 81 166
argument of a complex number 157 closed set 25, 80
attracting periodic point 54, 164 equivalent conditions for 25, 82
super attracting 121, 135 codomain 9
weakly attracting 54, 178 commutative diagram 95
compact set 71
bifurcation 57, 58 complement 25
diagram 61, 65, 106 complex functions 158
period doubling 65, 106 complex numbers 155
pitchfork 61 argument of 157
saddle-node 60
arithmetic in 155
transcritical 61
exponential form 158
bifurcation diagram 61, 65, 106
modulus of 156
bounded set 71
polar form 158
in the complex plane 181
on Riemann sphere 166
square roots of 158
c1 functions 50
complex plane 156
Cantor, Georg 75
extended 169
Cantor dust 188
Cantor set 71, 75 continuous function 12, 81
in complex plane 183, 188 of complex numbers 158, 169
Cantor Middle Thirds Set 71, 76 visualizing 13
Cayley, Arthur 172 convergent sequence 23, 80, 169
chaotic 87 converges to infinity 23, 168
conditions for 87 critical point 106, 184
chain rule 55, 161, 176 role in dynamics 106, 184
210 Index

dense subset 26, 81 homeomorphism 15, 96

equivalent conditions for 26, 82 linear functions as 18
derivative of a function 47, 159 hyperbolic attracting set 74
Devaney, R. 87 hyperbolic fixed point 52
differentiable 47, 159 hyperbolic periodic point 53, 164
distance 12, 80 hyperbolic repelling set 74
in symbol space 79
on the Riemann sphere 172
image 10
domain 9
Intermediate Value Theorem 15
doubling map 91, 124, 138, 166
inverse image 10
of an open set is open 82
€-neighborhood 21 invertible function 15
Euler's formula 158 conditions for 15
Euler's method 142
eventually fixed point 34
eventually periodic point 34 Julia, Gaston 183, 184
exponential form of a complex number Julia set 182, 198
158 algorithms for 185, 197, 198
exponential models 4
extended complex plane 169 least upper bound 101
limit 160
family of functions 57 limit point 24
parametrized 57 linear fractional transformation 170
Fatou, Pierre 105, 183, 184 logistic function 4, 6
Feigenbaum, Mitchell 113
Feigenbaum's constant 113
Mandelbrot, Benoit 185
filled Julia set 182, 198
Mandelbrot set 155, 185
fixed point 31
algorithms for 185, 199, 200
eventually fixed 34 map or mapping 9
finding fixed points 31, 33, 48, 50
Mean Value Theorem 47
hyperbolic 52
metric 80
forward asymptotic 34
on Riemann sphere 172
fractal 75
Mobius transformation 170
function 9
modulus of a complex number 156
C 1 50
Morse sequence 85
codomain 9
continuous 12, 158, 169
differentiable 47, 159 neighborhood 21, 27, 80
domain 9 in complex plane 156
family of 57 in extended complex plane 168
inverse of 15 Nested Interval Theorem 101
invertible 15 Newton, Isaac 115
linear 18 Newton's function 117, 170
one-to-one 10 Newton's method 115
onto 11 algorithms for 200
parametrized family of 57 chaotic behavior in 134, 179
range 9 in the complex plane 170
for cubic polynomials 126, 174, 179
graphical analysis 36 for polynomials 119, 170
algorithms for 194 for quadratic polynomials 121, 173
grows without bound 23 Newton-Raphson method 115
 Index 211

nonhyperbolic periodic point 53 range of a function 9

importance of 58 Raphson, Joseph 115
repelling periodic points 54, 164
onto 11 weakly 54, 178
equivalent conditions for 11 Riemann sphere 166
one-to-one 10
equivalent conditions for 10, 15 saddle-node 60
open set 21, 80, 169 Sarkovskii, A. N. 43
equivalent conditions for 22, 25, 81, Sarkovskii's ordering 43
82 Sarkovskii's Theorem 41,43
on Riemann sphere 172 on an interval 45
orbit 33 sensitive dependence on initial
bounded 181 conditions 86
sequence space 79
parameter 57 distance in 79
parametrized family of functions 57 shift map 83
perfect set 71 Smale, Stephen 75
period 33 stable set 34
prime period 33 finding stable sets 50, 53, 163
period doubling bifurcation 65, 106 stereographic projection 168
cascade of 106 super attracting 121, 135
periodic cycle 33 symbol space 79
periodic orbit 33 distance in 79
periodic point 33
attracting 54, 164 tent map 76, 92, 102
eventually 34 topological conjugacy 95, 97
hyperbolic 53, 164 topologically transitive 85
repelling 54, 164 conditions for 85
weakly attracting 54, 178 topology 21
weakly repelling 54, 178 totally disconnected 71, 183
phase portraits 4 transcritical bifurcation 61
pitchfork bifurcation 61 Triangle Inequality 13, 17, 80
polar form of a complex number 158
preimage 10 weakly attracting periodic point 54, 178
of an open set is open 82 weakly repelling periodic point 54, 178
prime period 33 well mixed 85

quadratic map 102, 113, 148

of the complex plane 166, 181
List of Symbols

lR. the set of real numbers, p. 9

r the nth iterate of the function f, p. 1
f(A) the image of the set A, p. 10
f-1(A) the inverse image of the set A, p. 10
N,(x) the neighborhood of x with radius E, p. 21,80
WS(p) the stable set of the periodic point p, p. 34
WS(oo) the stable set of infinity, p. 34
An the set of points which are in [0,1] after n
iterates of h(x) = rx(l - x), p.68
A the set of points which remain in [0,1] under
iteration of h(x) = rx(l- x), p.68
r a Cantor set, p. 71
E2 the sequence space of O's and 1's, p. 79
(J" the shift map, p. 83
d[x,y] the distance between the points x and y, p.80
S1 a circle, p. 91
Nf(x) Newton's function for the function f, p.117
214 List of Symbols

C the set of complex numbers, p. 155

82 the unit sphere in ~3, p. 168
arg(z) the argument of the complex number z, p. 157
Kc the filled julia set, p. 182
min{xl,x2, ... ,xn } the smallest number in the
finite set {Xl,X2, ... ,xn }
Universitext (continued)

[Link]: Banach Lattices

MineslRichmaD/Ruitenburg: A Course in Constructive Algebra
Moise: Introductory Problem Course in Analysis and Topology
Montesinos: Classical Tessellations and Three Manifolds
Nikulln/Shafarevich: Geometries and Groups
0ksendal: Stochastic Differential Equations
Porter/Woods: Extensions and Absolutes of Hausdorff Spaces
Rees: Notes on Geometry
Relsel: Elementary Theory of Metric Spaces
Rey: Introduction to Robust and Quasi-Robust Statistical Methods
Rickart: Natural Function Algebras
Rotman: Galois Theory
Ryhakowski: The Homotopy Index and Partial Differential Equations
Samelson: Notes on Lie Algebras
Schiff: Normal Families of Analytic and Meromorphic Functions
Shapiro: Composition Operators and Classical Function Theory
Smith: Power Series From a Computational Point of View
Smorynski: Logical Number Theory I: An Introduction
SmoryDski: Self-Reference and Modal Logic
Stanilif: The Mathematical Theory of Turbulence
Sdllwell: Geometry of Surfaces
Strooc:k: An Introduction to the Theory of Large Deviations
Sunder: An Invitation to von Neumann Algebras
Tondeur: Foliations on Riemannian Manifolds
Verhulst: Nonlinear Differential Equations and Dynamical Systems
Zaanen: Continuity, Integration and Fourier Theory