Mathematical Explorations with MATLAB

This book is about the kind of mathematics usually encountered in

first year university courses. A key feature of the book is that this
mathematics is explored in depth using the popular and powerful
package MATLAB. The emphasis is on understanding and investigating
the mathematics, and putting it into practice in a wide variety of
modelling situations. In the process, the reader will gain some fluency
with MATLAB, no starting knowledge of the package being assumed.
The range of material is wide: matrices, whole numbers, complex
numbers, geometry of curves and families of lines, data analysis,
random numbers and simulations, and differential equations form the
basic mathematics. This is applied to a large number of investigations
and modelling problems, from sequences of real numbers to cafeteria
queues, from card shuffling to models of fish growth. All extras to the
standard MATLAB package are supplied on the World Wide Web.

All three authors hold positions at the University of Liverpool.

Ke Chen is Lecturer in Mathematical Sciences, Peter Giblin is Reader
in Mathematics, and Alan Irving is Reader in Theoretical Physics.
Mathematical Explorations with MATLAB

KE CHEN, PETER GIBLIN, ALAN IRVING

AMBRIDGE
UNIVERSITY PRESS
 CAMBRIDGE UNIVERSITY PRESS
Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore, Sao Paulo

Cambridge University Press

The Edinburgh Building, Cambridge CB2 8RU, UK

Published in the United States of America by Cambridge University Press, New York

www.cambridge.org
Information on this title: www.cambridge.org/9780521630788

Cambridge University Press 1999

This publication is in copyright. Subject to statutory exception

and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without the written
permission of Cambridge University Press.

First published 1999

Reprinted 2000

A catalogue record for this publication is available from the British Library

ISBN 978-0-521-63078-8 hardback

ISBN 978-0-521-63920-0 paperback

Transferred to digital printing 2008

Cambridge University Press has no responsibility for the persistence or

accuracy of URLs for external or third-party Internet websites referred to in
this publication, and does not guarantee that any content on such websites is,
or will remain, accurate or appropriate.
 Contents

Preface page xi

Part one: Foundations 1

1 Introduction 3
1.1 First steps with MATLAB 3
1.2 Vectors and plots 6
1.3 Creating and editing script files 9
1.4 Getting hardcopy of things 15
Exercises 16

2 Matrices and Complex Numbers 18

2.1 Vectors and matrices 18
2.2 Complex numbers 23
2.3 Population dynamics: the Leslie matrix 24
Exercises 27

3 Whole Numbers 31
3.1 A loop to calculate Fibonacci numbers 31
3.2 A loop with conditionals: the 3n + 1 or hailstone problem 33
3.3 The euclidean algorithm for greatest common divisors 34
3.4 Fermat's theorem and the power algorithm 36
Exercises 39
3.5 Appendix 41

4 Graphs and Curves 44

4.1 Polynomials 44
4.2 Initial examples of drawing curves 45

v
vi Contents

4.3 Taylor polynomials 47

4.4 Approximations using the function polyfit 49
4.5 The goat problem 50
4.6 Envelopes of lines 51
Exercises 54
4.7 Appendix 58

5 Representation of Data 60
5.1 Data analysis 60
5.2 Least squares fitting 63
Exercises 70
5.3 Appendix 71

6 Probability and Random Numbers 75

6.1 Generating random numbers 75
6.2 Random integers 76
6.3 Simulating uniform distributions 78
6.4 Simulating normal distributions 78
6.5 Simulating negative exponential distributions 79
Exercises 82
6.6 Appendix 84

7 Differential and Difference Equations 88

7.1 Ordinary differential equations (ODEs) 88
7.2 Systems of differential equations 91
7.3 Difference equations 94
Exercises 96

Part two: Investigations 99

8 Magic Squares 101

8.1 Introduction 102
8.2 Magic squares size 3 x 3 102
8.3 Magic squares size 4 x 4 105
8.4 Magic squares size 5 x 5 (optional) 107

9 GCDs, Pseudoprimes and Miller's Test 108

A GCDs of random pairs and triples of numbers 108
B Pseudoprimes and Miller's test 111
 Contents vii

10 Graphics: Curves and Envelopes 116

A Rose curves and epicycloids 116
B Envelopes 119
C Curves of constant width 122

11 Zigzags and Fast Curves 128

A Spirographs and zigzags 128
B Fast curves 135

12 Sequences of Real Numbers 146

12.1 Mobius sequences 146
12.2 Cobweb diagrams 148
12.3 Mobius functions and powers of matrices 149
A Investigation on Mobius sequences 153
B Attracting cycles 156
C Quadratic and exponential sequences; fixed points 159

13 Newton-Raphson Iteration and Fractals 164

13.1 Introduction 164
13.2 The equation z2 + 1 = 0 165
13.3 General quadratic equations 166
13.4 The cubic equation z3 - z = 0 168

14 Permutations 171
A Cycle decompositions 171
B Card shuffling 177
14.1 Introduction 177
14.2 Ins and outs 179
14.3 Cycles 185
14.4 Rough riffles (ruffles) 185
14.5 Appendix 188

15 Iterations for Nonlinear Equations 189

15.1 1D: Method 1 - Newton-Raphson 190
15.2 1D: Method 2 - Gauss-Jacobi 191
15.3 1D: Convergence analysis 191
15.4 2D: Iterations for nonlinear systems 196
15.5 2D: Contour plot and convergence history 199
Exercises 203
viii Contents
16 Matrices and Solution of Linear Systems 207
16.1 Operation counts 208
16.2 Dense linear systems 209
16.3 The iterative refinement algorithm 212
16.4 A perturbation analysis for Ax = b 213
16.5 Sparse matrices, graph ordering and permutations 214
Exercises 217

17 Function Interpolations and Approximation 221

17.1 1D: Introduction 222
17.2 The 1D example M-file intdemo 1. m 223
17.3 1D data fitting 224
17.4 How accurate is my approximation? 227
17.5 Introduction to multi-variable approximation 228
17.6 The 2D M-file intdemo2.m 229
17.7 Contour plots, 3D plots and slicing 229
17.8 The `\' global method 233
17.9 The piecewise method 234
17.10 Comparison of approximations 234
Exercises 236

18 Ordinary Differential Equations 239

18.1 Strategy 239
Exercises 240

Part three: Modelling 243

19 Checkout Queues: Long or Short 245

19.1 Simulating queues 246
19.2 The motorway filling station 250
19.3 The Leo's cafeteria 251
Exercises 253

20 Fish Farming 257

20.1 Preliminary look at the problem 258
20.2 Models of fish growth 258
20.3 Designing the Leslie matrix 261
20.4 Fishing strategy 262
Exercises 263
 Contents ix

21 Epidemics 265
21.1 Preliminary look at some data 266
21.2 The SIR model for the dynamics of an epidemic 267
21.3 Studying the behaviour analytically 268
21.4 Analysing the data 269
Exercises 271

22 Dynamics of Snowboating 273

22.1 Preliminary look at the problem 274
22.2 The equations of motion 275
22.3 Exploring the operating parameters 277
Exercises 278

23 Tides 281
23.1 Preliminary look at the tidal data 282
23.2 Fourier series and methods 282
23.3 Analysis of an electrical signal 284
23.4 Fourier analysis of the tidal data 284
Exercises 285

Appendix 1 MATLAB Command Summary 286

Appendix 2 Symbolic Calculations within MATLAB 290

Appendix 3 List of All M-files Supplied 292

Appendix 4 How to Get Solution M-files 296

Appendix 5 Selected MATLAB Resources on the Internet 297

References 299
Index 301
 Preface

Mathematics and its practitioners have come a long way since the days of
drawing polygons in the sand with a stick. Although this cannot be said
of all our degree courses, there is nevertheless an increasing realisation
in higher mathematics education that current computing technology can
open new doors for students and tutors alike. This book arose out of
a largely successful attempt to complement traditional mathematical
courses with one which took this opportunity seriously.
First year students at a UK university are expected to acquire a wide
range of mathematical skills-the ability to argue logically, absorb new
concepts, calculate accurately, translate everyday problems into appro-
priate mathematical language, construct mathematical models and to
assess the approximations made. We chose to use the popular and pow-
erful computer package MATLAB to help promote some of these skills.
It provided a convenient way to help students understand things graph-
ically, to see the wood rather than the trees in complex problems and to
give access to more realistic modelling situations.
We chose MATLAB rather than one of the increasingly sophisticated
and algebraically based packages because of the very gently sloped learn-
ing curve involved. MATLAB allows the student to graduate smoothly
from the functionality of a hand calculator, through increasing use of
powerful numerical and graphical facilities towards a high level pro-
gramming capability. The latter point was considered a bonus in that
it provided a possible access route to programming for students with no
prior computer background. At the very least, students with no key-
board skills at all can acquire a degree of familiarity with an essential
modern tool, the computer.
The course, and this book, were designed for students coming to grips
with a typical first year honours mathematics course at a UK university.

xi
xii Preface
In our case, students had already completed the first semester of core
units and so already had a basic knowledge of calculus, complex num-
bers, vectors and matrices. In the book, we assume that the reader has a
reasonable level of skill in calculus but only limited familiarity with the
other topics. The typical student will be in the process of extending this
base to include some selection of topics such as elementary statistics,
mechanics, linear algebra, number theory, differential equations, Fourier
series and so on. The book is thus intended to help motivate new topics
and to build on old ones.
Like Gaul, the book is divided into three parts. Part one comprises
a very elementary `hands-on' introduction to the features of MATLAB
followed by a series of methods chapters. In these the reader is taken
through a range of mathematical ideas and given `on the job' training in
those MATLAB techniques which are expected to be of particular value
in the ensuing project chapters. Thus all the standard programming
structures and MATLAB commands are introduced through work on:
matrices; whole numbers and elementary number theory; graphing plane
curves; data fitting and approximations to functions using least squares
techniques; simulation of random distributions; and ordinary differential
equations. In this way, the student learning how to use MATLAB is
taken through mathematics which is (or should be!) interesting for its
own sake.
Part two contains a variety of projects, termed `Investigations', which
build on the earlier ideas. Matrices are applied in the context of magic
squares, permutations and the solution of linear systems; manipulation
of whole numbers is applied to greatest common divisors of random sets
of numbers, primality testing and card shuffling; approximation tech-
niques are applied to solution of nonlinear equations and interpolations;
and so on. In each case, an exploratory attitude is encouraged, backed
up with plenty of explicit exercises, both purely computational and more
mathematical in nature.
Finally, Part three contains a number of `Modelling Projects' in which
the reader is invited to employ some of the skills developed in Part one.
By its nature, mathematical modelling is a rather open-ended process
and requires a certain degree of mathematical maturity that a first year
student may not yet have attained. Nevertheless, we feel that the avail-
ability of techniques to which MATLAB gives access, and the very great
importance of modelling as an applied mathematical skill, mean that
this is an opportunity not to be missed. In practice, we have found that
students cope well with these challenges.
 Preface xiii

At Liverpool, we required students to work through the preliminary

material (Part one, taking six weeks), and allowed them to choose a total
of three projects from Parts two and three, with at least one from each
part. Two weeks seemed to be a good time to allow for the completion
of one project, so that the whole course was twelve weeks long.
We have of course striven for uniformity in important matters through-
out the book. But a discerning reader will detect three different styles in
the project work of Parts two and three, providing a measure of variety
which we feel is entirely healthy.
We have provided appendices which list MATLAB commands, give
some information on symbolic calculations (not used explicitly in the
material of the book) and MATLAB resources, and list the available
M-files chapter by chapter.
Using the book
The book will prove useful in a number of contexts. Firstly it can be
used, as it stands, to deliver a complete course unit. Secondly, the book
should prove useful to course designers with slightly differing require-
ments. In this case the various examples of project work will provide
a convenient source of material and stimulate the creation of further
material tailored to the local need. Thirdly, the book can serve as a
self-contained tutor for the enthusiastic individual who is not following
any formal course structure.
In every case, the reader is intended to work through the book while
sitting at the computer keyboard, although there are also mathematical
exercises to be done off-line. A preliminary skim through this preface
and Chapter 1 will help orient the newcomer before plunging in.
Readers who already have some experience with MATLAB might well
wish to jump straight into Chapter 2. If in doubt, readers can quickly
brush up their skills with the exercises at the end of Chapter 1.
Copies of all the `M-files' to which the text refers are freely available.
Details of how to obtain these are given in the Appendix. Partial so-
lutions and hints are available to course tutors in electronic form on
request from the publisher.
MATLAB is available on a wide variety of platforms. For definiteness,
the book assumes the reader has access to MATLAB within a Microsoft
Windows environment. Should this not be the case, readers may ex-
perience some small inconvenience in the early stages while adapting
file-handling and editing instructions to suit their own installation, but
the M-files should all run correctly and the material of the book itself is
xiv Preface
platform independent. Course providers might wish to make available a
brief summary of key points where the local reader might otherwise go
astray.

Acknowledgments

We are very grateful to colleagues at Liverpool who have helped to

set up this course and have provided input to the material. These are
Nigel Backhouse, Eric Edmond, Toby Hall, Neil Kirk, Dick Wait, Neville
Waters. We are especially grateful to the students who have taken the
course over several years; their efforts have detected possible ambiguities
in the project work and helped to make the material more user-friendly
- and, we hope, correct - than it was at the beginning. Any remaining
faults are of course our responsibility. PG is also grateful to Brown
University in Providence for its generous hospitality during part of the
writing period of this book, and to the Fuibright Commission for a travel
grant.
Liverpool, October 1998 Ke Chen
Peter Giblin
Alan Irving

MATLAB is a registered trademark of The MathWorks, Inc.

Part one
Foundations
 1

Introduction

1.1 First steps with MATLAB

If you haven't already done so, you should start MATLAB by double-
clicking the relevant icon with your mouse or by asking a friend sitting
next to you to show you how. Asking a friend is often the quickest
way to obtain help and, in what follows, we will encourage you to take
this route when all else fails. If `clicking' and `icons' mean nothing to
you, you may need some extra help in getting started with Windows. It
might also be that your system doesn't use Microsoft Windows and that
simply typing matlab will do the trick. For example, if you are using a
Unix system of some kind this may be the case.t If all goes well you will
see a MATLAB prompt

inviting you to initiate a calculation. In what follows, any line beginning

with >> indicates typed input to MATLAB. You are expected to type
what follows but not the >> prompt itself. MATLAB supplies that
automatically.

1.1.1 Arithmetic with MATLAB

MATLAB understands the basic arithmetic operations: add is +, sub-
tract is -, multiply is * and divide is /. Powers are indicated with ,
thus typing
>> 5*5+12"2

results in
t We will make some further remarks, where appropriate, about editing and file
handling in a non-Windows environment.

3
4 Introduction
ans =

169

If you typed the above line and nothing happened, perhaps you omitted
to press the <Enter> key at the end of the line. The laws of precedence
are built in but, if in doubt, you should put in the brackets. For example

>> 8*(1/(5-3)-1/(5+3))

ans =

The sort of elementary functions familiar on hand calculators are also

available. Try

>> sgrt(5"2+12"2)

and

>> exp(log(1.7))

What do think sin(pi/2) will produce? Try it.

In fact MATLAB has the value of 7r = 3.1415926... built in. Simply
type in pi whenever you need it. Try the following:

pi
format long
pi
format short

MATLAB retains considerably more significant figures of accuracy than

suggested by the default setting which is given by format short.

1.1.2 Using variables

You can assign numerical values to `variables' for use in subsequent

calculations. Typing

>> x=3

produces
 1.1 First steps with MATLAB 5

X =

or you might want something more useful like

>> rad=2; ht=3;
>> vol=pi*ht*rad"2

vol =

75.3982

Note that the first line had two `commands' on it, neither of which
seemed to produce a result! When MATLAB encounters an instruction
followed by a semi-colon ; it suppresses any visual confirmation. It
really does obey the instruction but keeps quiet, as you can check with
>> rad=4;
>> rad

rad =

This is useful if you want to avoid cluttering up the screen with inter-
mediate results. Watch out for semi-colons in what follows.
Remember that each variable must somehow be assigned a value be-
fore you can make use of it in further calculations. For example if you
have followed the above examples and now type
>> f = x"2 + 2*x*y + y"2

you should get a result something like

??? Undefined function or variable y

This is self-explanatory. If you now type y=4; and then repeat the
calculation of f you should have more success.
Incidentally, a quick way of repeating a previous MATLAB instruction
is to press the `up-arrow' key (1) until you recover the command you
want. Try it now. If the original instruction was not quite correct, or
if you want to develop a new instruction from a complex but similar
previous one, you can use the same trick. Recover the command which
6 Introduction
you want then use the sideways arrows (- and ->) together with the
delete key to edit the old command suitably. As an exercise, try using
your previous work to calculate the volume of a right-circular cylinder
with radius 2 and height 1/4. Did you get it?
If you have difficulty remembering the names of variables which you
have assigned, you can try typing who or whos. Try them both now. Do
you recognise the variables listed?

1.2 Vectors and plots

One of the pleasures in learning to use MATLAB is discovering the
simplicity of plotting things. The basic principle is:
(i) select a sequence of x-values that is, a vector of values;
(ii) evaluate y = f (x), that is, obtain a corresponding vector of y-
values
(iii) plot y vs X.
Before doing this, it is worth spending a moment learning something
about how MATLAB deals with vectors.

1.2.1 Vectors
Type in the following examples which all result in vector-valued vari-
ables. Pause to think about the result in each case.
u=[2,2,3]
u= [2 2 3]
v=[1,0,-1]
w=u-2*v
range=1:13
odd=1:2:13
down=20:-0.5:0
even=odd+1
xgrid=0:.05:1; x=xgrid*pi
y=sin(x)

The first two lines demonstrate that elements of a vector can be sep-
arated by spaces or commas. If you are worried about inserting blank
spaces by accident you can stick to the comma notation. Thus [1+1 2
3] is the same as [2, 2, 31, whereas [1 +1 2 3] is the same as [1, 1, 2, 31 !
 1.2 Vectors and plots 7

Note that vectors can be of any length. They can be row vectors as
here, or column vectors like
>> w'
ans =

0
2
5

where the apostrophe denotes transpose (T). In MATLAB, vectors are

treated simply as a special case of matrices which you will learn much
more about in the next chapter.
Notice what happens when the displayed vector is too long to fit on a
line. MATLAB just displays as many elements as it can and then puts
the rest on the following lines. The elements in a row vector are treated
as `columns'.
An elementary function of a vector x, such as sin(x), is also a vector
of the same kind. We can use this fact to create plots of functions as
shown in the next section.
MATLAB knows how to multiply matrices of compatible size. This
will be discussed in greater detail in the next chapter. For now, try
typing

Can you make sense of the results? Why did the last one not work?
Now suppose you want a set of values z given by z = y2, where y is
the vector of values already assigned. From the above experiments you
will realise that
>> z=y*y

is not understood by MATLAB.

>> z=y*y'
is understood but is evaluated as the scalar product y y! What you
need to do, to force MATLAB to multiply things element-by-element, is
to type
8 Introduction
>>

where the . inserted before the * symbol is the key feature forcing
-
element-by-element operation. Similarly, u. /v and y. 2 are understood
as element-by-element operations with vectors of the same size.

1.2.2 Plotting things

Type whos at this point to verify that you have vectors x and y defined
as above. They should both be 1 x 21 matrices (that is, row vectors).
Plotting is easy. Just type
>> plot(x,y)

and sit back. A nice simple graph of y = sin x vs x will magically appear.
The axes are chosen automatically to suit the range of variables used.
This is the simplest possible case. In later work you will want to do
more elaborate things. For now, try the following:

>> title('Graph of y=sin(x)')

>> xlabel('x')
>> ylabel('y')
>> yl=2*x;
>> hold on
>> plot(x,y1,'r')

You can probably figure out the significance of each of these commands.
For example, yl=2*x defined new function values y = 2x, hold on told
MATLAB to keep the same graph and plot (x, y 1, ' r') plotted the next
graph on top. Note that the axes were adjustedt and the second curve
was plotted in red.
In the above examples of plot, MATLAB joined up the 21 points
with straight-line segments. Should you not want this, you can specify
plotting points using a choice of symbols as follows. Try this

>> hold off

>> plot(x,y,'+')
>> plot(x,y,'g*')
>> plot(x,y,'w.')

t Assuming you are using MATLAB version 4. There are a number of small differ-
ences between this and earlier versions, particularly with graphics commands.
 1.3 Creating and editing script files 9

Did MATLAB do what you expected? Did you remember to use the t
key to reuse previous commands? If you need some help with how to
use any MATLAB instruction you can type for example
>> help plot
>> help hold
>> help sin

and so on.

1.3 Creating and editing script files

Once you get going, you may find it tiresome to keep reentering the same,
or similar, sequences of commands. Fortunately, there is a simple way
round this: you simply store any frequently used sequence of commands
in a file called a `script' or `M-file'. You can then invoke this list of
commands as often as needed.
For example, in a particular session you might want to find the dis-
tance between two points A and B whose position vectors are given by
a = (1, 0, -2) and b = (2, 3, 1) respectively. Knowing that the vector
displacement between them is
d=b-a
and that
Jd12=d.d,
you might use the following sequence of MATLAB instructions:
>> a=[1,0,-21;
>> b=[2,3,1];
>> d=b-a;
>> dd=d*d';
>> dist=sqrt(dd)

to solve that particular problem. This is fine, but suppose you have a
set of five points and want to check which pair is the closest together?
You would obviously want to store as many as possible of these steps in
a `script' (file) to reuse as required.

1.3.1 Editing and saving a text file

We first need to review file-handling and editing.
10 Introduction
Non-Windows users
If you are not using Microsoft Windows, you will at this point need to
make some slight alteration to procedures. However, no matter how
MATLAB has been installed on your system, there will undoubtedly be
a text editor of some sort. Assuming it is called edit, the easiest way to
invoke it from MATLAB is probably to type
>> !edit fname

where fname is the name of a text file which either exists or will exist,
by the time you have finished! If that doesn't work, consult someone
knowledgeable about your system setup or ask a patient friend.
Windows users
Windows comes with its own basic text file editor called Notepad, whose
icon is usually found within the Accessories Group. A typical MATLAB
setup within Windows makes direct use of this accessory so we will con-
centrate on this method. To create and edit a new file called myf ile. m,
from within MATLAB, proceed as follows:
(i) In the MATLAB Command Window menu, click on File.
(ii) Click New then M-file.
(iii) Within Notepad which you have now started, you can type some
lines, for example
% myfile.m
It doesn't do very much, just identifies itself.
% These 3 lines are comment lines which MATLAB ignores.
disp(' I am an M-file')

(iv) Click File then Save As.

(v) Within the box File Name which is open waiting, type in myf ile. m.
(vi) Click OK.

You have now created a file which MATLAB can find and use. You
hope!
Back in the MATLAB Command Window you can now ask MATLAB
whether it can find the file. Type
>> type myf ile
and you should see a list of the lines which you typed in. If not, go
back to step (i) and start Notepad again by clicking File in the MATLAB
Command Window followed this time by Open M-file. You should see an
 1.3 Creating and editing script files 11

entry corresponding to myf ile. m. If not, you should go right back to

the beginning of this section, perhaps with someone else watching your
steps this time.

1.3.2 Script files

If all is well, you now have your first example of a script file. To use it
you simply type
>> myfile
whereupon you should see something like
I am an M-file
Now for something more useful. We will set up an M-file to repeat
the earlier instructions for getting the distance between two points. Pro-
ceed as above to create a new M-file called distab.m containing a few
comment lines, such as
% distab.m
% Calculates the distance between two vectors a and b
% . . .

including any more comment lines (those starting with %) which you
may need to remind yourself how it works. Then the business lines
d=b-a;
dd=d*d';
dist=sqrt(dd)

Remember to save the file by clicking File then Save As as you did
when creating mf ile. m. Review that example if necessary. When you
have finished editing the file, you might as well close Notepad by clicking
File then Exit. You can leave Notepad open if you want, but it may
get confusing if you leave too many Notepad windows open at once.
When exiting, Notepad always reminds you if there are recent changes
to be saved.
Now pick values for the vectors a and b, if you haven't already done
so

>> a=[1,0,-21;
>> b=[2,3,1];
Then find the distance by simply typing
12 Introduction

>> distab

Did it work? If not, go back into Notepad and keep trying.

Try it for a different choice of points A and B, perhaps for a pair
whose distance you can trivially check, for example A - (1, 2, 3) and
B - (1, 1, 3).
>> a=[1,2,3] ;
>> b=[1,1,3];
>> distab
To find out what M-files you have created, or which others MATLAB
already contains, use the command what. If you now want to check
the purpose of an M-file you can use help (just as with any MATLAB
instruction)
>> help myfile
>> help sqrt
>> help sin

The command help displays the initial comment lines at the top of an
M-file. This is why it is always good practice to include comment lines
(starting with %) at the top of an M-file. It is also a good idea to include
the name or title in the first comment line.
Note that myfile is the name of the MATLAB instruction (what you
type to use it), whereas myfile . m is the name of the file containing its
definition.

1.3.3 Function files

It was tedious to have to assign the two vectors each time before using
the above `script'. You can combine the assignment of input values with
the actual instruction which invokes the M-file by using a function M-
file. Not only that, but you can at the same time assign the answer to a
new variable, that is, you can design a function M-file distfn such that
typing
>> dab=distfn([1,2,3],[1,1,3]);

or

>> a=[1,2,3]; b=[1,1,3]);

>> dab=distfn(a,b);
 1.3 Creating and editing script files 13

assigns the correct distance to the variable ab without any further ado.
Here is how to modify the script M-file distab.m to become the func-
tion M-file distfn.m. We will assume that you didn't leave Notepad
open and so will have to start editing from scratch.
(i) In the MATLAB Command Window menu, click on File.
(ii) Click pen M-file to start Notepad and look for the list of avail-
able M-files, that is, ones ending with . in. (If necessary, change
the *. txt ending showing in the File Name box to *. m).
(iii) Select distab.m.
(iv) Now you can make changes. First change the comment lines
to reflect the new name and purpose and then the modify the
MATLAB instructions so that your file looks like this:
distfn.m
Calculates the distance between two vectors a and b
Usage:-
%
input: a,b (position vectors)
% output: distfn is the distance between them
function dist=distfn(a,b)
d=b-a;
dd=d*d';
dist=sqrt(dd);

(v) Click Save As.

(vi) Within the box File Name which is opened waiting, type in
distfn.m.
(vii) Click OK.
Back in the MATLAB Command Window, you can now type help
distfn to remind yourself how to use it. You type
>> dist=distfn([1,1,1], [2,2,2] )
or

>> dist=distfn(a,b)

to assign the required distance to the variable dist. If you got more than
one number flashing up, perhaps you forgot some of the semi-colons with
the function. If so have another look at it using Notepad.
Function M-files (or `M-functions'), that is, M-files whose first non-
comment line starts function . . ., have a very important feature. Aside
14 Introduction

from the name itself (distfn in this case), all the other variables (a, b,
dd etc.) are purely internal to the function. This can help reduce con-
fusion with other calculations and variables which you may have used.
Check this by typing

>> who
>> clear
>> who
>> dist=distfn([1,1,1],[2,2,2])
>> who

The command clear clears out all previously defined variables. After
executing the function with dist=distfn(. .), no trace of the internal
variables remains in your MATLAB session.

1.3.4 Diary files and saving things

You will sometimes find it handy to keep a copy of what you produce
on the screen. You might want to print out bits of it later (see the
next section). This is done very easily. To see how it works, type the
following

diary sectl.txt
% Beginning of section 1
now some commands
myfile
dist=distfn([1,1,1],[2,2,2])
diary sect2.txt
>> % I want this stored somewhere else
>> x=0:.1:1;
>> y=x.*x
>> plot(x,y)
>> diary off

The session will proceed fairly normally. The command diary faame
tells MATLAB to put a copy of the text output (numbers and letters)
in the file f name. Typing diary off switches it off. The above example
puts some output into one file and some into another. You can now use
Notepad to have a look at what was generated and, if needed, edit it.
Invoke Notepad as usual by clicking File and try to pen files with the
ending * . txt. You should see sect 1. txt and sect2. txt in the list.
 1.4 Getting hardcopy of things 15

When you look in sect2.txt you may be disappointed to find no plot

of x2 there! This is because plots and other graphical images can't easily
be represented as text files. The next section shows how to get round
this.
When you end a MATLAB session all current variables and their
values are lost. Usually this doesn't matter. The effect of stopping and
restarting a session is the same as typing
>> clear

that is, you start with a clean sheet. Should you actually wish to save
what you are doing for another time you can type, for example,
>> save monday

or

>> save monday x,y,mymatrix

where the second version saves only the explicitly mentioned variables
in the specified file monday. You can then reload things the next day or
whenever with
>> load monday

Remember that only the variables themselves that is, their current val-
ues, are saved using these techniques. Any formulae which you have
typed in will be lost unless you have entered and saved them in an
M-file.

1.4 Getting hardcopy of things

1.4.1 Windows users
An advantage of using a Windows environment is that printing text files,
such as diary files and M-files, is performed in a completely standard way
in all applications. MATLAB is normally set up to take full advantage
of this. The same is true for the plot images which MATLAB produces.
To get a copy of an M-file or other text file just open it up using
Notepad in the usual way (see previous section). Click on File and
then Print. That's it! If it isn't, then perhaps you need to check Print
Setup on the same menu to see where the output is being sent. If
necessary, you may need to check with your local expert. Usually things
have already been set up so all output goes to the most convenient local
printer.
16 Introduction

Table 1.1. Summary of basic commands introduced so far.

plot (X, Y) plot(x,y,'*') plot(x,y,'+g')
title('The title') xlabel('x label') xlabel('x label')
sqrt(x) sin(x) exp(x)
hold on hold off
x=-1:.2:1 y=x.*x dotprod=x*y' a
format long format short
help sqrt help myfile save fname
diary filel.txt diary off load fname

a (' - transpose)

For plots, you go about getting a hardcopy in the same way but you
use the File and Print buttons on the Window containing the figure
(usually labelled Figure No. 1).

1.4.2 Non-Windows users

If you don't have a Windows system or want to by-pass it, you can
usually get away with something like

>> !print fname.txt

or use the appropriate print command of the underlying operating sys-

tem (for example lp or lpr for Unix).
For plots, the MATLAB command print is usually set up to print a
copy of the current figure on the default printer. If this is not so, you
may make some progress by following the advice in help print. Failing
that, try asking your patient friend.

Exercises
Before proceeding, check your level of skill by completing all of the
following short exercises. If you can't remember some of the commands
have a look at the list in Table 1.1. If necessary, go back and reread the
relevant section.

1.1 Find the sum of the first four terms in the sequence
1 2 3
2x3' 3x4' 4x5'
 Exercises 17

1.2 Define a vector t with values evenly spaced by 0.2 between 0

and 6 inclusive. Now use this to obtain plots of
f (t) = sin(7rt)
and
g(t) = exp(-t) sin(irt)
on the same graph with the first in green and the second in
yellow. If you are unsure of the MATLAB functions needed,
use help exp etc. Enhance the graph by adding a white line
corresponding to y = 0.
1.3 Use the editor to create an M-file which deduces the length of
each side of a triangle ABC whose vertices have position vectors
a = [1, 2, 3], b = [2, 3, 4] and c = [3, 4, 5].
 2

Matrices and Complex Numbers

2.1 Vectors and matrices

2.1.1 Vectors
We recall briefly how to enter vectors. Let
a=(-1,2,4) and b=(1.5,2,-I).
To assign these vectors to variables a and b type either
>> a [-124]
>> b = [1.5 2 -1]
or

>> a = [-1,2,4]
>> b = [1.5,2,-1]

Thus spaces or commas can be used.

One way of finding the dot or scalar product, of two vectors, say a b,
was introduced in 1.2. Here is another which uses the important
idea of element by element multiplication also introduced in Chapter 1.
Typing
>> c=a.*b

where there is a dot before the multiplication sign *, multiplies the vec-
tors a and b element-by-element: in this case, c = (-1.5,4, -4). The
dot product is then obtained by summing the elements of c:
>> sum(c)

gives a b = -1.5. Similarly

>> sqrt(sum(a.*a))

18
 2.1 Vectors and matrices 19

gives the magnitude of a. In fact, the MATLAB command, norm, will

directly find the magnitude of a vector. To find the angle 0 between a
and b, we can use 0 = cos-1 a- b/lal Ibi. In MATLAB, cos-1 is acos so
the the complete calculation, performed this way, would be written
>> theta = acos( sum(a.*b) / sqrt(sum(a.*a)*sum(b.*b)) )

and gives 0 = 1.693 radians approximately.

2.1.2 Matrices
The matrix
7-1 1 2
A= 3 -1 1
-1 3 4

is entered in MATLAB as
>> A=[-1 1 2;3 -1 1;-1 3 4]

with semicolons separating the row vectors, or as

>> A= [-1 1 2
3 -1 1
-1 3 4]
with different rows separated by <Enter>, therefore appearing on dif-
ferent lines. The MATLAB prompt >> will not reappear until you have
finished defining the matrix by closing the bracket]. If you accidentally
make one row have more entries in it than another row then you will
get an error message. Note that, if you wish, you can insert commas
between the entries which lie in the same row:
>> A=[-1,1,2;3,-1,1;-1,3,4]

If you find the line spacing in MATLAB output is over-generous for

your taste you can suppress extra spacing with the command
>> format compact

At some stage, have a look at help format to see what else you can do
to control the look of MATLAB output.
The equations
-XI + X2 + 2x3 10
3x1 - x2 + X3 -20 (2.1)
-xi + 3x2 + 4x3 40
20 Matrices and Complex Numbers
can be written as
Ax = b (2.2)

say, where x is the column vector (xl, x2i x3)T and b is the column
vector (10, -20, 40)T of right-hand sides of the equations. In MATLAB,
we can write b as the transpose of a row vector by using
>> b=[10 -20 401'
In order to solve linear equations Ax = b, where the determinant of
A is nonzero, we can use the inverse A-l of A, that is, the matrix such
that AA-' = A-'A = I, the 3 x 3 identity matrix. To discover whether
the determinant is zero, type
>> det(A)

which gives the answer 10. The solution for x is then obtained by
x = A-'Ax = A-'b,
which in MATLAB is
>> x=inv(A)*b

displaying the result as a column vector x (in this case (1, 19, -4)T).
Thus matrix products are obtained by using the multiplication symbol
* but the matrices must be of compatible sizes or MATLAB produces
an error. For example b*A is meaningless since the number of columns
of b, namely 1, does not equal the number of rows of A, namely 3. You
can test matrix multiplication by
C=A*A
det(C)
D=A-3
det (D)
Thus C = A2, D = A3 and, by the well-known rule for square matrices
of the same size det(P)det(Q) = det(PQ), the determinants are 100,
1000 respectively.
Actually, one does not usually solve linear equations such as equa-
tions (2.1) or (2.2) by first finding the full inverse of the corresponding
matrix. One applies linear operations to the augmented matrix formed
from A and b. MATLAB can also perform the calculation in this way.t
You simply type
t See Chapter 16 for a more detailed discussion of the solution of linear systems.
 2.1 Vectors and matrices 21

>> x = A\b

Again, the matrices must have compatible dimensions. Try it and check
your answer by forming the product Ax or by finding the `residual'
r = b - Ax.
You can also add two matrices of the same size. MATLAB doesn't
even complain if you add a scalar to a matrix. In fact
>> E=A"2+2*A+1

produces the matrix

1 1 1
E=A2+2A+ 1 1 1
1 1 1

So don't assume that 1 means the identity matrix.t The identity matrix
and the matrix all of whose elements are 1 can be called up by typing
>> diag([1 1 1] )
>> ones(3,3)
so that diag makes a diagonal matrix with the given diagonal entries
and ones makes a matrix of is of the given size. What does
>> diag(ones(1,3))

give? Note the brackets when using the function diag. The identity
matrix is actually built into MATLAB as (e.g. for 3 x 3)
>> eye (3)
As already noted, one may wish to solve the equations Ax = b, by
forming the corresponding augmented matrix
>> F = [A b]
In fact, this method is applicable even when the determinant is zero or
when the matrix A of the equations is not square. The procedure is to
reduce the augmented matrix to `row reduced echelon form', that is, to
use row operations to produce a matrix in which there is a leading 1 in
each row, the entries before, above and below which are all zeros. We
shall not use this in any important way here, so if you have not met
the idea just skip over the remainder of this section. The row reduced
echelon form is used in the project of Chapter 8 on magic squares.
t Unfortunately, typing A*1 produces A again, suggesting that in multiplication
MATLAB thinks 1 means the identity matrix!
22 Matrices and Complex Numbers
Typing
>> G=rref(F)

produces the row reduced echelon form of F. In this case it is

1 0 0 1

G= 0 1 0 19
0 0 1 -4
This means that the original equations are equivalent to the equations
whose augmented matrix is the matrix G, and these equations are simply
x = 1, y = 19, z = -4 so the solutions can be read off.
Don't forget the round brackets when using a MATLAB function such
as rref. Thus if you type in a matrix directly you must still include
them, as in
>> rref ([1 2 3; 4 5 6; 7 8 9] )
Note that this produces a single row of zeros (the last row), indicating
that the rows of the original matrix are linearly dependent. You can
check this by finding the determinant, which is zero.
You may also care to try typing
>> rrefmovie(F)

which gives an animated demonstration of the row reduction process.

2.1.3 Eigenvalues and eigenvectors

If A is an n x n matrix, A is a number and x is a nonzero n x 1 (column)
vector, such that
Ax = Ax,
then x is called an eigenvector of A and A is the corresponding eigen-
value.t The eigenvalues can also be thought of as the numbers A such
that det(A - AI) = 0, where I is the n x n identity matrix.
For example, try typing
>> A=diag([1 2 3])
>> eig(A)
>> P=[1 2 3;4 5 6; 578]
>> det(P)
t Geometrically, this means that the linear map associated with A sends x to another
vector along the same line through the origin as x.
 2.2 Complex numbers 23

B=inv(P)*A*P
eig(B)
[Y,D]=eig(B)

The eigenvalues of A should be simply the diagonal entries 1, 2 and 3.

Since the matrix P is nonsingular (det(P) = 3), the matrices A and
B = P-'AP are `similar' and so have the same eigenvalues. Note that
MATLAB may order them differently. The last line finds the eigenvec-
tors of B as well: the matrix Y has columns which are eigenvectors and
the diagonal matrix D has diagonal entries which are eigenvalues. The
eigenvector corresponding to the first diagonal element of D is the first
column of Y, and so on. In this case, all eigenvalues and eigenvectors
are real. For an example where this does not happen, try the matrix A
of 2.1.2.

2.2 Complex numbers

MATLAB has the symbol i () built in together with the rules of
complex arithmetic. MATLAB also reserves the symbol j for. Try
typing

and see what happens.

Warning: when dealing with complex numbers, there is a risk that
MATLAB will misunderstand the symbol i for the square root of -1.
This can happen if i has been used recently for an indexing variable,
usually an integer (we shall meet indexing in for loops later; see 3.1).
To safeguard against this, just type
>> clear i,j

before starting work on complex numbers, and avoid using i for anything
else!
You will already have seen some complex numbers if you found the
eigenvalues and eigenvectors in 2.1. MATLAB can handle complex
numbers easily. For example,
>> a=1+i; b=2-3i;
>> c=a*b
24 Matrices and Complex Numbers
>> d=sqrt(a)
produces the answers c = 5 - i, d = 1.0987 + 0.4551i. Note that just one
square root is shown. Likewise
(-1)"(1/2)
(-2+21)"(1/3)

produce respectively the answers i, 1 + i. The other square or cube

roots must be obtained by multiplying by -1 or by cube roots of unity,
respectively.
The modulus, argument (> -ir and < it radians) and real parts of an
already specified complex number a are obtained by
>> abs(a)
>> angle(a)
>> real(a)

respectively. The command

>> imag(a)

produces the imaginary part without the i attached. For example with
a = 1 + i the answer is 1.
>> conj (a)
produces the complex conjugate a of a. (If a = x + iy where x and y are
real, then a = x - iy.) Thus
>> a*conj(a)-abs(a)"2

should always give answer 0 (but might give a very small answer such
as 1.7764e-015, that is, 1.7764 x 10-15)
Try also
>> exp(i*pi)
which produces eivr = -1.

2.3 Population dynamics: the Leslie matrix

There are many uses of matrices and their eigenvalue properties in eco-
nomics, life-sciences and probability theory as well as in the physical
sciences. A very simple example is in the following discrete time model
 2.3 Population dynamics: the Leslie matrix 25

for the age structure of the population of a country or other large commu-
nity. The basic idea is to take a vector representing the age distribution
in one year, construct a matrix of transition probabilities from one year
to the next and then, by matrix multiplication, predict the probable
age distribution for the next year. Predictions for subsequent years are
obtained by further matrix multiplication. The hard bit is to make a
model for the probability matrix! Here is an example.
In some given year, we count the number of people in age-bands 0-
5, 6-19, 20-59 and 60-69. Further subdivisions are of course possible
with a little more work. We then make the following, fairly drastic,
simplifying assumptions

Within one age-band, the age distribution is constant, that is, there
are the same number of people in each year group.
We do not consider anyone who lives beyond 69!
Deaths only occur in the 60-69 age-band at the rate of dA% per annum
and in the 0-5 band at dI% per annum.
Births are due to people in the 20-59 age-band at the rate of b% per
annum.
With these assumptions, we want to relate the column vector
N(2) = (ni(2), n2(2), n3(2), n4(2))T
of populations in the four age-bands in year 2 to the column vector

N(1) = (ni(1),n2(1),n3(1),n4(1))T
of populations in year 1. Recall that, as above, T stands for transpose
of a vector or matrix.
To see how to obtain this relationship, consider Table 2.1.
The various columns contain: the age range within the band; the
symbol nz for the current number of persons in that band; the number
of year groups (for example, ages 0, 1, 2, 3, 4, 5 make six year groups);
the number that, according to our rules, die in that year; the consequent
number (the rest) who survive the year; the number of these survivors
who leave the band (graduating to the next one or, in the case of the
age-band 60-69, passing out of our calculations); and the number who
enter the band, either through being born into it or through graduating
from the one below.
Of course, each band's leavers become the next band's entrants. Ac-
cording to the simplifying assumptions, the number in each individual
26 Matrices and Complex Numbers

Table 2.1. Construction of a Leslie matrix.

Age- No. in Year Die Survive Leave Enter

band band grps. band band

0-5 n1 6 00nl (1 - 00 n1 (1 - 0) 6 100 n3

6-19 n2 14 0 n2 144 (1 -0s

20-59 n3 40 0 n3 El n2
40 14

60-69 n4 10 100
n4 (1 - 100) 14 (1 - loo) 10 40

year within an age-band is the same, so the `leavers' are just the 'sur-
vivors' divided by the number of years covered by the age-band. In prac-
tice much narrower age-bands might be taken, to make this assumption
more plausible.
It is now easy to see that the vector N(2) of populations of the various
age-bands in year 2 is related to the corresponding vector N(1) for year
1 by a matrix equation
N(2) = LN(1),
where L is the so-called Leslie matrix
1- 0)
b 0
s 0 100
/
01

13
L s11-100 1 0
39
0
0 14 40
0
0 0
1 9 // _ dA )
40 10 ll 1001 1

Similarly,
N(3) = LN(2) = L2N(1),
and so on. In general, we have:
N(t + 1) = LN(t).
Given data for t = 0, say, and some estimates of b, dI and dA, one can
then predict the likely age structure of the population for a number of
years ahead. This is very useful if you have to plan pension schemes,
university sizes or day-care provision.
An M-file leslie.m is provided to allow you to explore these ideas.
Type
 Exercises 27

>> leslie

to obtain a 4 x 4 matrix L describing the evolution of the above model

population. You are prompted to supply the birth rate (b) as a percent-
age per annum, for example 2.5, the infant mortality rate (for example
dI = 1) and the age-related death rate (for example dA = 10) etc. Now
set up an initial population, say
>> N = [ 3.6 11.4 29.6 10.6]'

which is a very rough approximation to the 1996 UK population distri-

bution in millions.
You should try the following:
Multiply N by L 4 times (N1=L*N, N2=L*N1.... ) to predict the UK
population distribution in the year 2000 according to this model. What
significant changes are there if any? Note that you can also type things
like

>> L"4
>> N50=L"50*N

Type

>> [Y D] = eig(L)

to obtain the eigenvectors and eigenvalues of the Leslie matrix L. Find

the largest eigenvalue and its corresponding eigenvector E. You can com-
pare the population vector after say 50 years, L50N, with this eigenvector
E by
>> N50=L"50*N
>> N50./E

The second command divides the three elements of N50 by the three
corresponding elements of E. The result should be three approximately
equal numbers, showing that the population after 50 years is roughly
proportional to the eigenvector corresponding to the largest eigenvalue.
What happens if the initial population distribution is proportional to
this eigenvector?

Exercises
2.1 As a simple example of manipulation of complex numbers, con-
sider the following, which will be covered in more general terms
28 Matrices and Complex Numbers
in the project of Chapter 13. (Note the warning given above, in
2.2; for safety you can type `clear i' before starting.) Enter
any complex number into MATLAB, for example
>> z=3+4i

Now type
>> z=(z"2-1)/(2*z)

Using the T key, repeat the last command several times. Even-
tually the answers settle down-converge--to either i or -i; for
the starting value 3 + 4i they settle down to i.
Can you discover the rule which determines, from the starting
z, whether the numbers will converge to i or -i? (You are not
expected to prove the rule. That is covered in Chapter 13.)
If you start with z = 0, MATLAB produces in turn -oo, 0,
oo, 0, which doesn't make much sense since z = too, if it
gives anything at all, should give z = oo again. But there
are other starting values such as z = 2 which produce more
reasonable nonconvergence. Which starting values for z pro-
duce nonconvergence? (Again only an experimental answer is
expected.)
2.2 This exercise combines matrix work with editing files and using
diaries. Make an M-file which contains the following (remember
to leave spaces between the entries of a single row, or to insert
commas there):
A=[1 1/2 1/3
1/2 1/3 1/4
1/3 1/4 1/5]
b=[1 0 0]'
det (A)
X=inv(A)*b

You could call this by some convenient reference name like

ch2g2.m.
This M-file solves the equations with A as matrix of coeffi-
cients and the column vector b as the `right-hand sides' of the
equations. Note that the determinant of A is calculated first.
Even if det A 0 0 the numerical solution may be very sensitive
to small inaccuracies or changes in the parameters. You can
 Exercises 29

expect this if det A is small. This is just one example of an

ill-conditioned system. f
(a) Make a copy of ch2g2.m, say ch2g2a.m age-band edit the
file so that it solves the equations in which the elements 1/3 of
A are replaced by 0.333.
(b) Do the same replacing 1/3 by 0.33; you could call the M-file
ch2q2b.m.
In MATLAB, run your M-files and save the results in a diary.
Thus you type (in MATLAB) diary ch2g2. txt then ch2g2 to
run this M-file, then ch2q2a to run the next M-file, and finally
ch2q2b to run the third one. When you have run them all type
diary off. Return to the editor and look at the file ch2q2.txt
to see what is in it. You should find that the results of solving
these similar sets of equations are very different. You could of
course add your own comments on this by typing them alongside
the results and print out the diary so that you have a `hardcopy'
record.
2.3 Copy the M-file leslie.m and edit it as follows. The new M-file
is to be suitable for studying a population of cats described by
age bands 0-1, 2-5, 6-10 and 11-15 and obeying the following
modelling assumptions:
No cats beyond age 15 are considered.
Births only arise from the 2-5 band (b% per annum).
Kitten deaths occur at the rate dK% per annum in the 0-1
band.
All survive in the 2-5 band.
Deaths from being run over occur in the 6-10 band at a rate
dR% per annum and from old age in the 11-15 band at dA%
per annum.
An appropriate name for the resulting M-file is cats.m.
(a) Find the matrix generated when b = 15, dK = 5, dR = 2 and
dA = 30. (Thus type, for example, diary ch2g3. txt, then run
the M-file by typing cats and entering the appropriate data,
then type diary off.)
(b) Assuming an initial population distribution [20, 20, 20, 20],
does the overall population increase or decrease over a five-year
t See Chapter 16 for a discussion of ill-conditioned systems.
30 Matrices and Complex Numbers
period? You can use sum to get the total of a vector. You can
use diary ch2g2g3.txt again to include the result of finding
the population during this period, and type in whether it is in-
creasing or decreasing. Note that using the same diary name
appends the new material to the old. It does not overwrite the
original diary material. Don't forget to type diary off every
time you have temporarily stopped writing to this diary.
(c) Find the largest (in magnitude) eigenvalue and correspond-
ing eigenvector of the Leslie matrix C for the cat population
evolution. Note that
>> V=C(:,1)

picks out the first column V of a matrix C. This is very useful

if you want to do something further with this column, such as
compare it with another vector.
(d) Using the initial population in (b), show that the population
vector (N) after 50 years is approximately proportional to the
eigenvector found in (c). Repeat for another initial population
of your own choice. Recall that
>> V1./V2

gives a vector containing the ratios of the elements of V1 and

V2.
 3
Whole Numbers

In this chapter, we shall introduce further MATLAB structures in the

context of some properties of whole numbers.

3.1 A loop to calculate Fibonacci numbers

Type in succession
f = [1 1]
f (3) = f (1) + f (2)
f
f (4) = f (2) + f (3) ;
f
The last command produces the vector [1 1 2 3]. Thus f (1) refers to
the first entry of the vector f, f (2) to the second, etc.
We can do this over and over in a loop:
>> f = [1 1] ;
for k = 1:15
f (k+2) = f (k+1) + f (k) ;
end
>> f

Note that MATLAB suppresses the >> prompt until the loop is com-
pleted by end. What the for... end loop does is to take in succession
the values 1, 2, ... ,15 for the variable k and to augment the vector f by
a new entry f(k + 2) each time round. Thus k = 1 makes f emerge as
[1 1 2]; k = 2 makes it emerge as [1 1 2 3], and so on. The semi-colon
after the equation for f(k+2) suppresses output during the 15 times the
loop is executed.

31
32 Whole Numbers
Typing
>> plot(f,'*')

gives a plot of the values, placing an asterisk at each point (i, f (i)).
Just plot (f) joins these points up with straight segments, producing
a steeply sloping curve. The entries of the vector f are the Fibonacci
numbers, 1, 1, 2, 3, 5, 8, 13, 21, 34. .... The rule for forming them is, as
above, f(1) = f(2) = 1;f(k + 2) = f(k + 1) + f(k) fork > 1.
The above commands can be put into an M-file. As a variant on the
above we could use a `while loop', as follows:
f=[1 11;
k=1;
while f(k) < 1000
f (k+2)=f (k+1)+f (k) ;
k=k+1;
end
f
plot(f)
This M-file is stored under the name f ibno. m and is executed by
typing f ibno. The indentation used is intended to display the structure
of the M-file and has no effect on the running or output. This time the
two commands between while and end are executed while the condition,
f(k) < 1000, of the while loop is valid. Since f(16) = 987 and f(17) =
1597 the last value of k which allows execution of the while loop is
k = 16. The value of k at the end of the whole M-file is 17 since k
is incremented by 1 during the final execution of the loop. Notice an
important difference between for and while loops:
In a for loop the loop variable (k in the above example) is automat-
ically incremented by 1 for each pass through the loop. If you want
the variable incremented by d each time and starting at a while not
exceeding b you use
for k=a:d:b

end

In a while loop the variable (k above) is not incremented automat-

ically. You have to increment it explicitly during the loop, as above
where this is done with k=k+1.
 3.2 A loop with conditionals: the 3n + 1 or hailstone problem 33

3.2 A loop with conditionals: the 3n + 1 or hailstone problem

Let n be a positive integer. We iterate the following process: if n is even
we divide it by 2; if n is odd we replace it by 3n + 1. Thus, starting with
n = 10 we get successively the numbers 5,16,8,4,2,1. Here is an M-file
which performs this automatically on any given (input) n. It is called
hail.m. Note that in MATLAB rem(a,b) is the remainder on dividing
a by b.

n=input('Enter n')
f=[n] ;
k=1;
while n>1
k=k+1;
if rem(n,2)==O
n=n/2;
else
n=3*n+1;
end
f (k)=n;
end
bar(f)
To use this we type hail and then input say 25 <Enter>. This
version produces a bar-chart rather than a graph.
The resulting bar-chart for n = 1000 is shown in Figure 3.1.
Typing k after running the above M-file produces an answer which is
the exact number of steps taken to reach 1, counting the original number
as Step 1. (Thus for example n = 8 produces 8, 4, 2, 1 which counts as
4 steps.) For n = 1000 it takes 112 steps, as you can roughly see from
the bar chart of Figure 3.1. Typing max (f) after running the M-file
produces the maximum value attained by n during the iteration: the
highest point on the bar-chart. This is 9232 for a starting value of 1000.
Here are some notes on the way that hail.m works.

(i) Note the incrementing of k in the while loop. This could equally
well be done later in the loop, after the f (k) = n line.
(ii) Note the if ... else ... end construction. All conditionals must
be closed with an end statement. You can have if ... end without
an alternative `else'.
(iii) The == symbol is used for comparison. The quickest way to
34 Whole Numbers
10000

9000

8000

7000

6000

5000

4000

3000

2000

1000

20
A 40
Ih N lll1110llllll111L
60 80 100 120

Fig. 3.1. The `hailstone' iteration starting at n = 1000,

check whether something is even is to compute the remainder on

division by 2, which is done by rem (n , 2) .
(iv) Semi-colons after the if statement and the else statement are
not needed (but will not do harm).
It is not known whether the above process always terminates in a 1 for
any input n. No number has yet been found which does not terminate
in 1 but no general proof exists either. You can read more in [13].

3.3 The euclidean algorithm for greatest common divisors

To find the greatest common divisor (gcd) of two whole numbers a and
b, where b > 0, we first divide b into a:

a=bq+r,where0<r< b.
Here q is the quotient and r is the remainder, which is calculated by

r=rem(a,b)
(If needed, the quotient q can be calculated by q=f loor (a/b).) Then
we always have

gcd(a, b) = gcd(a - bq, b) = gcd(r, b) = gcd(b, r). (3.1)

 3.3 The euclidean algorithm for greatest common divisors 35

Table 3.1. Euclid's algorithm.

Step Current a Current b Current q Current r
1 69 15 4 9
2 15 9 1 6
3 9 6 1 3
4 6 3 2 0

In the special case where r is an exact factor of b (denoted r1b and

pronounced `r divides b'), the last term is simply r.
Sketch proof of (3.1) The first = in equations (3.1) follows because the
pair a, b have exactly the same divisors as the pair a - bq, b, hence certainly the
same gcd. To see this, let d1a and ddb. Then a = aid, b = bid say for integers
al, bi. Then a - bq = d(ai - biq), so d is also a factor of a - bq. Conversely,
the same kind of argument shows that a common factor of a - bq and b is
also a factor of a. The second = in equations (3.1) is from the definition of r
and the third = just reverses the order of the numbers without affecting their
gcd.

The idea of Euclid's algorithm is to replace a by b and b by r and

repeat. Let us take a = 69, b = 15. The calculation can be organised
as shown in Table 3.1. Then, gcd(69,15) = gcd(15, 9) = gcd(9, 6) =
gcd(6, 3) = 3 since, at the end, 3 goes exactly into 6, and hence the gcd
is 3.
To realise this in MATLAB, we do the division and then replace a by
b, b by q. The calculation is continued while the remainder r calculated
is > 0, but there is some advantage in using the condition b > 0 since, at
the end of the loop, they are equal anyway and r does not have a value
until the loop has already been entered. It is much easier to understand
the final M-file than to describe it! Can you see why the gcd is the final
value of a calculated by the procedure, rather than the final value of b
or r?
Function for calculating the gcd of a and b.
To use, type e.g. gcd(69,15)
For convenience we make sure a,b nonnegative to start with

function h=gcdiv(a,b)
if a<O a=-a;
end
36 Whole Numbers
if b<0 b=-b;
end
while b > 0
r=rem(a,b);
a=b;
b=r;
end
h=a;

This M-file, which is available as gcdiv.m, is written as a function.

Note that the function value is h which is the outcome of the euclidean
algorithm procedure. The M-file is used by a command of the form
gcdiv(a,b) where a and b are either known to MATLAB at that mo-
ment, or are replaced by explicit whole numbers such as 69 and 15.
What happens if you type gcdiv(0,4) ? Do you think the answer is
reasonable? What about gcdiv(4,0) ?
There is an exercise to extend this to calculating the gcd h of three
numbers, at the end of this chapter (Exercise 3.7), given the definition
gcd(a, b, c) = gcd(gcd(a, b), c). That is, you work out the gcd x of a and
b and then h = gcd(x, c). There is also an exercise to get the computer
to generate a pair of random numbers for the calculation of the gcd. In
the Investigations (Chapter 9) there is a project involving the gcds of
large numbers of randomly chosen numbers. For this it is necessary to
write an M-file which repeatedly uses the function gcdiv above.

3.4 Fermat's theorem and the power algorithm

The most interesting examples from elementary number theory use a
wonderful theorem first stated in general by Pierre de Fermat in 1640,
though the first full published proof is due to Leonhard Euler in 1736.
A special case was known to Chinese mathematicians hundreds of years
before Fermat. The theorem is proved in any book of elementary number
theory, for example [7]. Here, it will merely be stated and some of its
consequences explored. Recall that a prime number is a whole number
p > 1 which has no factor besides 1 and p itself. The theorem is this:
Fermat's theorem Let p be prime and suppose a is not divisible by p.
Then aP-1 - 1 is divisible by p.
The conclusion can also be written rem(0-1, p) = 1, meaning `the
remainder on dividing aP-1 by p is 1', or aP-1 - 1 mod p. We shall not
 3.4 Fermat's theorem and the power algorithm 37

make use of the notation - here; a - b mod m, where a, b, m are whole

numbers and m j4 0, means simply that m is a factor of a - b.
We shall use Fermat's theorem shortly to give a method for showing
when numbers are not prime, but for the moment we concentrate on the
following:

Problem If n and m are reasonably large numbers, how can we calculate

the remainder on dividing an by m?
If say n = 1000000, then even with a = 2 it is completely infea-
sible to work out an and then get the remainder from that, because
21000000 has about 300000 digits! MATLAB fails at much smaller num-
bers, for example it correctly says that rem(7"2,10) = 9 but it claims
that rem(7"20,10) = 0 which is obviously false!

3.4.1 The power algorithm

This is a cunning method for working out rem(an, m) when the numbers
are large. The mathematical details are given in 3.5.1 for your interest.
The main thing to remember is that the power algorithm does not work
out an and only needs to handle numbers about as big as m2 during the
calculation. To use the power algorithm for working out rem(an, m),
simply type
>> pow(a,n,m)

Remember that you can get information on the M-file pow. m by typing
help pow.
Using the power algorithm it is easy to verify many examples of Fer-
mat's theorem such as 161998 - 1 mod 1999, which is so because 1999 is
prime.
Examples
(i) pow(3,118,119) gives 32. Hence, by Fermat's theorem, 119 is
not prime (for, if it were, then the answer would have been 1).
Of course, 119 = 7 x 17.

(ii) pow (2, 1993002, 1993003) gives 1121689, so 1993003 is not prime
(in fact, it is 997 x 1999).

(iii) pow(2,10"'6+2,10"6+3) gives 1, so this is consistent with 106 +

3 = 1000003 being prime, but doesn't prove it. Also
38 Whole Numbers
pow(3,10"6+2,10"6+3)
pow(5,10"6+2,106+3)
both give 1. All these are in some sense `evidence' for 106 + 3
being prime, but not a proof. (In fact, 106 + 3 is prime.) See
3.4.2 for some nasty examples where the `evidence' for primality
is in fact false evidence!

Eventually the arithmetic in MATLAB starts to go wrong with round-

ing errors. For example, 109 + 7 happens to be prime, but
>> pow(2,10"9+6,10"9+7)

does not give the correct answer 1. This is because ten figure numbers
cannot be reliably manipulated by MATLAB. You can rely on pow for
seven or eight figure numbers.
More powerful packages can handle hundreds of figures. These pack-
ages are routinely used to find very large (at least 100-digit) numbers
which are `probably prime', using Fermat's theorem and a more subtle
variant called Miller's test. These `probable primes' are used in cryptog-
raphy, that is in sending messages by virtually unbreakable codes. You
can read about such codes in [12]. Miller's test for more modestly sized
numbers appears in Chapter 9.

3.4.2 Pseudoprimes
There are numbers which satisfy the conclusion of Fermat's theorem
without satisfying the hypothesis that p is prime. That is, we can some-
times have rem(a'm-1, m) = 1, for some a > 1, without having m prime.
When this happens we call m a pseudoprime to base a.
For example, rem(724, 25) = 1, rem(2340 341) = 1, so that 25 = 52 is
a pseudoprime to base 7, and 341 = 11 x 31 is a pseudoprime to base
2. Nevertheless if, for a fixed in, lots of values of a are found for which
rem(am_l, m) = 1, then this can be regarded as accumulating evidence
for the primality of m. For very large m this is by far the most efficient
way of selecting `probable' primes, though in practice a variant known
as Miller's test is used. This is the subject of one of the Investigations.
See Chapter 9.
The number 561 (= 3 x 11 x 17) has the rather unpleasant property
that rem(a560 561) = 1 for any a not divisible by 3, 11 or 17. It is called
a Carmichael number. Numbers such as these masquerade as primes
under a large variety of tests, but can usually be unmasked with a little
 Exercises 39

patience. It was proved only in 1992 that there are infinitely many
Carmichael numbers.

Exercises
3.1 Run the M-file f ibno. m, i.e. type f ibno in MATLAB. The
length of the vector f is obtained by length(f). What is the
length of f in this case? What is the largest Fibonacci number
< 1000 and the smallest one > 1000?
3.2 Use the hail command to find the number of steps which 27
takes to reach 1, and the largest value reached during the iter-
ation.
Modify hail.m so that it prints out the number of iterations
and the largest value reached, but does not plot the values at
the end. Call the resulting M-file hail2.m.
Modify hail2.m a s follows, calling the result hail3.m. Re-
move the line n = input ( ' Enter nI); , insert the lines

for k=50:100
n=k

at the beginning, and

end % of for k=1:50 loop

at the end. What does this M-file do? Check this by running
it, i.e. typing hail3 in the MATLAB window.
Using a diary (1.3.4) for storing the output from hail3.m
read off all the values of n between 50 and 100 which have the
same largest value during the hailstone iteration as 27 has. For
each one, state the number of iterations each takes to reach 1.
3.3 Type format short <Enter>. Let 6,,, and 7n denote the whole
numbers written in ordinary decimal notation as strings of n 6s
and n 7s, respectively. So 64 = 6666, etc. Use the command
gcdiv to find the gcd of 6n and 7n for n = 1, 2, 3. .... When is
the displayed answer wrong? How do you know it is wrong?
Try typing format long and repeating the exercise. What is
the new value of n? Again, how you can be sure that, for this
n, the answer is wrong?
40 Whole Numbers

3.4 Typing pow(17,1000,100) should give the answer 1, so that

171000 is 1 + a multiple of 100, i.e. 171000 - 1 is a multiple of
100. What are the last two digits of 171000? Show in a similar
way that 171000 - 1 is a multiple of 99 and 101. Why can you
deduce now that 171000 -1 is a multiple of 99 x 100 x 101? (This
uses the fact that no two of 99, 100 and 101 have a common
factor.)
Use pow to find the last three digits of 192000.
Find the remainders on dividing 173313 and 331317 by 112643.
Note: The MATLAB command rem is of no use here: the num-
bers are far too large. You must use the power algorithm M-file
pow. Writing 17 = 1 + 16, why does it follow that 173313 - 1 is
divisible by 16? (That's a mathematical question!) Can you see
why, similarly, 331317 - 1 is divisible by 16? What can you say
about gcd(173313 -1,3313 17 - 1)?
3.5 Use pow to find rem(213946761394677). Why can you deduce
that 1394677 is not prime?
3.6 Type in an M-file as follows, calling it mypow. m

m=10"'7+1;
while m<10"7+100
m % Note no semicolon!
pow(2,m-1,m) % Note no semicolon!
m=m+2;
end;

What does this do? Use it (and a diary if necessary, though

the pause button can be useful too) to find the only two numbers
between 107 and 107 + 100 which could possibly be prime.
3.7 Create a function h=gcdiv3(a,b, c) which calculates the gcd of
three numbers a, b, c. Here is one way to do this. Your M-file
should start with the line
function h=gcdiv3(a,b,c)
and it should simply work out x = gcd(a, b) and then h =
gcd(x, c), in each case by calling on the function gcdiv which
already exists. It is a general fact that
gcd(a,b,c) = gcd(gcd(a,b),c)

so your M-file should be very short! Call the completed M-file

gcdiv3. m to indicate that it is an M-file.
 3.5 Appendix 41

Test the function gcdiv3 on some triples of small numbers to

see whether it gives the right answer. Then calculate
gcd(414304,56496,351824) and gcd(196054,175131,133407).
Create an M-file gcdrand.m with the following lines (which
are one way of generating `random' numbers, in this case num-
bers between 1 and 106)
t = fix(clock)
n = t(6)
rand('seed',n)
a = round(rand*10"6)
b = round(rand*10"6)
c = round(rand*10"6)
gcdiv3(a,b,c)
(We will be discussing `random' numbers in more detail in Chap-
ter 5.) Run this M-file 20 times and write down the gcds ob-
tained (don't write down the numbers a, b, c). What percentage
of trials give gcd equal to 1? There is an investigation later
(Chapter 9) which finds the `probability' that three randomly
chosen numbers have gcd equal to 1, both theoretically and ex-
perimentally.

3.5 Appendix
3.5.1 Proof of the power algorithm
The key fact is that we can take remainders at any time during a
calculation and always arrive at the same answer. Thus if we want
x = rem(ab, m) we can work out successively
p = rem(a, m), q = rem(b, m), x = rem(pq,m). (3.2)

For example rem(973 x 58,10) = rem(3 x 8, 10) = rem(24,10) = 4.

This is of course just the units digit of 973 x 58, since we are taking
remainders after division by 10.
There is an extremely fast algorithm for working out remainders of
large powers. Here is an example. Suppose we want to work out x =
rem(750, 11). First, we work out the binary representation of the power,
50. This is
50=(0x1)+(1x2)+(0x22)+(0x23)+(1x24)+(1x25).
42 Whole Numbers

Table 3.2. The power algorithm.

b rem(b,11) d=bit x so far
7 7 0 1
72 5 1 5
74 3 0 5
78 9 0 5
716
4 1 9
7 32
5 1 1

Thus
750 = 72 x 716 x 732 .

We'll see shortly how to do this quickly. Then in succession we work

out the remainders on division by 11 of 72, 74 7a 716 732 This is done
by repeated squaring and using equations (3.2) successively. Thus we
calculate rem(72,11) explicitly, then square the answer (5) and take the
remainder again. At that point we just get 3 and the square of that is
9-no need to take a remainder. The full sequence is:
rem(72,11) = 5, rem(52,11) = 3, 9, rem(92, 11) = 4, rem(42,11) = 5.
Now using (3.3), we only need the first, fourth and fifth of these. It is
best to set everything out as in Table 3.2. Note that the only time the `x
so far' entry changes is when there is a `1' in the `bit' column, indicating
that that power of 2 does indeed occur in the expression (3.3).
The way that the bits are calculated is as follows. We start with
d = rem(n, 2) as the `units bit' for n, that is, the right-most bit in
the binary expansion of n. At every step after the first, n is replaced
by (n - d)/2; this brings the next bit in the binary expansion to the
right-most position. Thus for 50 we get successively:
d=rem(50,2)=0; n = (5 0 - 0)/2 = 25;
d=rem(25,2)=1; n = (2 5 - 1)/2 = 12;
d=rem(12,2)=0; n = (1 2 - 0)/2 = 6;
d=rem(6,2)=0; n = (6 - 0)/2 = 3;
d=rem(3,2)=1; n = (3 - 1)/2 = 1;
d=rem(1,2)=1.
The M-file pow. m to implement this is as follows:

function x=pow(a,n,m)
 3.5 Appendix 43

b=a;
x = 1;
while n>O
d = rem(n,2);
if d==1
x = rem(x*b,m);
end
b = rem(b * b,m);
n = (n-d)/2;
end
 4
Graphs and Curves

In this chapter we shall use MATLAB to draw graphs of functions y =

f (x), to approximate functions in two different ways as polynomials,
to solve equations f (x) = 0 and to draw systems of lines in the plane
forming an `envelope'. (For graphs z = f (x, y), see Chapter 17.) The
technique of approximation is particularly important in applications,
where it is desirable to replace a relatively complicated function (or
indeed data set; compare Chapter 5) by a simple polynomial which can
be easily handled. The choice here is between a very good approximation
of f (x) close to some given value xo of x, which gets steadily worse as x
moves away (Taylor polynomials), and a reasonably good approximation
over a larger range ('polyfit' approximations). Each has its uses.

4.1 Polynomials
A polynomial such as p(x) = x4 + 2x3 - 3x2 + 4x + 5 is entered by means
of its coefficients:

>> p= [1 2 -3 4 5]

Note the spaces between the numbers; commas are also allowed for sep-
aration. In other words, a polynomial is just entered as a vector con-
taining the coefficients, starting with the `leading' coefficient in front of
the highest power of the variable x.
The roots of p (i.e., the solutions of p(x) = 0) are obtained by typing
>> roots(p)

Notice that this finds the complex roots as well as the real ones. We can
increase the number of figures displayed by typing

44
 4.2 Initial examples of drawing curves 45

>> format long

We can draw the graph of p by typing
>> x=-4:.05:2;
>> y=polyval(p,x);
>> plot(x,y)

The first line creates a vector x whose entries are the numbers from -4
to 2 at 0.05 intervals. Thus
x = [-4.0 - 3.95 - 3.9 ... 1.95 2.0].
As usual, the semi-colon merely suppresses output to the screen.
The second line evaluates the polynomial p at all these x, producing
a vector y of the same length as x containing p(-4),.. ., p(2).
The third line plots values of y against values of x, taking them in
corresponding pairs and joining up with (very short) straight segments.
By adding
>> hold on
>> plot([-4,2],[0,0])

the x-axis is also plotted. The plot command takes the first entry from
each square bracket, making the point (-4, 0), and joins it to the point
(2, 0) obtained from the second entry from each bracket, thus drawing
the x-axis across the screen. From this it is quite clear that p has two
real roots in the interval from -4 to 2, and their positions agree with
the calculation of roots as -3.18... and -0.728.... See Figure 4.1.

4.2 Initial examples of drawing curves

First recall the basic steps involved in plotting (1.2.2). For example, to
draw the curve y = sin x for x between 0 and 2ir we type
>> x=0:.05:2*pi;
>> y=sin(x);
>> plot(x,y)

The command plot works with two (identical length) vectors of values
specifying pairs of coordinates. In this case, if you type length(x) or
length(y) you should get the response 126. Although axes are scaled
automatically by the maximum values of x and y, they can be manually
changed if necessary. For example, the line
46 Graphs and Curves

50

40

20

10

-10

-20
-4 -3 -2 2

Fig. 4.1. Graph of y = x4 + 2x3 - 3x3 + 4x + 5, with the x-axis also drawn.

>> axis([0 10 -2 2])

makes the screen cover the region 0 < x < 10, -2 < y < 2. Try this to
see the effect. (In versions of MATLAB before Version 4, you may need
to reissue the plot(x,y) command after axis.)
Similarly, to plot the graph of y = x sin x, say, we could use
>> x=0:.05:2*pi;
>> y=x.*sin(x);
>> plot(x,y)

Note the . * operation (1.2.1), which multiplies corresponding entries of

the vectors x and sin(x), and does not attempt the impossible task of
multiplying x and sin(x) together as matrices (they're the wrong size).
Now try typing
>> x=0.01:.01:1;
>> y=1/x;
>> y=1./x;
>> plot(x,y)

The first attempt to define y = 1/x will not work; the second should
produce the graph y = 1/x, for -1 < x < 1. Recall (1.2.1) that
the symbol . / is used for elementwise division: divide 1 by each of the
elements of the vector x in turn and call the resulting vector y. The
 4.3 Taylor polynomials 47

instruction 1/x makes no sense since you can't divide 1 by a vector. (In
versions of MATLAB before Version 4, y=1. /x was not acceptable either
and a space had to be left after the 1 so that the dot was not confused
with a decimal point.)
Typing
>> t=0:.05:2*pi;
>> x=cos(t);
>> y=sin(t);
>> plot(x,y)

produces the parametric curve { (cos t, sin t) }, which is of course a circle

of radius 1 centred at (0,0), although it won't look very circular because
of the automatic scaling. Try
>> axis('square')

and note how the plot changes to being circular. You can also use
>> axis('equal')
which makes the scales on the two axes equal no matter what the shape
of the figure. Thus
>> t=0:.05:2*pi;
>> x=2*cos(t);
>> y=sin(t);
>> plot(x,y)
>> axis('equal')

produces a picture of an ellipse where the major and minor axes have
their true ratio of 2:1. Replacing axis (' equal') with axis (' square' )
does not have the same effect. Try it!
There are parametric curves in the exercises and investigations which
are more exciting than this!

4.3 Taylor polynomials

A good illustration of the power of graphics is to compare the graph of
a function such as sin x and the graph of its Taylor polynomial approx-
imation
x3 x5 x2k-1
x - 3!- + -5! - ... + (-1)k-1

(2k - 1)!
48 Graphs and Curves

Fig. 4.2. Sine curve (solid) and Taylor approximation of degree 9 (dotted).

The M-file tsine . m calls for the value of k (this is the number of terms in
the polynomial), then plots the `true' curve in green and, once <Enter>
is pressed, plots both this and the approximation in red. MATLAB uses
first the scale appropriate to the sine curve alone and then the scale
appropriate to the two curves together. Remember that to run this M-
file you just type tsine followed by <Enter>. You never type the m
when running M-files.
In Figure 4.2 the sine curve is the solid line and the Taylor approxi-
mation with k = 5 (that is, degree 2k - 1 = 9) is the dotted line.
The M-file tsine2.m on the other hand plots the actual sine curve (in
green) and then freezes the scaling while the approximation is plotted
(in red). This is achieved by the command axis (axis) after the first
plot command in the M-file. Notice the difference with say k = 4.
The approximation hugs the true curve more and more closely from
x = 0 onwards, as k increases. By the time k reaches about 10, the
curves are more or less indistinguishable over the range from 0 to 27r.
Here is the text of the M-file tsine. Even if you don't follow all the
details, you should note how easy it is to do this relatively complicated
thing in MATLAB.
Draws the Taylor approximation
to a sine curve of degree 2k-1
k=input('Type the number of terms
 4.4 Approximations using the function polyfit 49

of the Taylor series for sine: ');

x=0:.05:2*pi;
z=sin(x);
plot(x,y,'g') %Actual sine curve is drawn in green
hold on
pause % This holds up execution until Enter is pressed
w=x;
y=x;
s=-1;
for j=1:k-1
w=w.*x.*x/(2*j*(2*j+1));
Y=Y+s*w;
s=-s;
end
plot(x,y,'r')
%Approximation is drawn in red
hold off
Note the use of 'g' to draw in green (see 1.2.2). A dashed curve is
drawn by plot (x, y, ' - - '); when needed, the symbol - denotes a solid
curve. We have also already met the hold on statement which holds the
graph and puts the second one on top of it. It is necessary to `release'
the graphs by a call to hold off at the end. Notice again the use of
elementwise multiplication . * in this file.

4.4 Approximations using the function polyfit

An entirely different approach to approximating a function by a polyno-
mial is taken by the available function polyf it, which fits a polynomial
of a given degree to a curve by a method which in a sense minimises the
distance of the polynomial graph from the true curve graph, averaged
along their whole lengths. (See also Chapter 5.) For example (this M-file
is called polyex.m):
a=0:.05:2*pi;
b=sin(a);
c=polyfit(a,b,5);
d=polyval(c,a);
plot(a,d,'r')
hold on
pause
50 Graphs and Curves

Fig. 4.3. Sine curve (solid) and `polyfit' approximation of degree 3 (dotted).

plot(a,b,'g')
hold off
This finds a polynomial approximation of degree 5 to the sine curve over
the range 0 - 6.3, using a for a change as the independent variable. The
polynomial approximation c is simply a list of coefficients;, to evaluate it
at the points of a we use the function polyval. The first `plot' command
draws the approximation in red and the second `plot' command draws
the true sine curve in green. Notice how completely different this ap-
proximation is from the Taylor approximation of the same degree (given
by k = 3 in the M-file tsine.m).
Figure 4.3 shows the result of running polyex. m with degree 3.

4.5 The goat problem

Here is a well-known problem which leads to a frightful equation. We
can solve it graphically, or numerically.
Problem A goat is tethered to a point on the circumference of a circular
field of radius 10 metres. How long should the rope be so that the goat
can graze over exactly half the area of the field?
Let x be the length of the rope in metres. (That was the easy bit!)
Some messing around with areas of sectors and areas of triangles gives
 4.6 Envelopes of lines 51

the following equation for x:

2(x2 - 200) cos-1 (20) - x 400 - x2 + 100,7r = 0. (4.1)

See Figure 4.4 and 4.7.1. Note that cos-1 is the inverse cosine, which in
MATLAB is acos. It does not seem very feasible to solve this equation
explicitly! For a graphical solution the left-hand side of (4.1) can easily
be plotted-the M-file goatgr.m does this. It should be clear from the
plot (also perhaps from looking at Figure 4.4) that there is a solution
somewhere around x = 10. One way of getting MATLAB to calculate an
accurate value of the solution is to use the built-in function fzero. For
this we need an M-file containing a definition of the function in question.
In this case it is goatfn.m and is as follows (this function is available to
you; no need to type it in!):
function y=goatfn(x)
y=2*(x.*x-200).*acos(x/20)-x.*sgrt(400-x.*x) + 100*pi;
end;

Note that the MATLAB command goatfn will produce an error, but
say goatfn(10) will give the value of the function when x = 10. What
happens when you type goatfn(25) ? What in fact is the `domain' of
the goat function, that is, the range of values of x for which y is real?
We now type
a=fzero('goatfn',10)

meaning `look for the solution a of goatfn(x)=0 near to x = 10'. By

using format long a very accurate estimate is found.

4.6 Envelopes of lines

There is an attractive way of using the graphics in MATLAB to draw
systems of lines in the plane and their envelopes. For instance the M-file
parnorm.m draws the normals to the parabola y = x2, parametrised by
x = t, y = P. The normal to the parabola at (t, t2) has the equation

y - t2 = - 2t (x - t), that is, x + 2ty - t - 2t3 = 0.

This is p(t)x + q(t)y + r(t) = 0, where p(t) = 1, q(t) = 2t and r(t) _
-t - 2t3. So the functions p, q, r, written in MATLAB, are
p=ones(size(t));
52 Graphs and Curves

Fig. 4.4. A goat is tethered at G. We want to discover the length x of rope

which will allow the goat to graze over half the field. The resulting equation
for x is (4.1), and this is proved in 4.7.1. The shading is for that proof.

q=2*t;
r=-t-2*t.*t.*t;
Note the use of ones(size(t)) when the function is simply a con-
stant equal to 1. If it had been p(t) = 5 say, then we would use
p=5*ones(size(t)). Note also the use of .* for cubing each value
of t. The form r=-t-2*t. "3 also works.
Other possibilities for p, q, ,r are considered below.
The M-file parnorm.m requires input of

(i) the bounds on x (use xl = -2, xu = 2, that is we are restricting

the x-coordinate to lie between -2 and 2);
(ii) the lower limit of y (use yl = -1). The upper limit of y is then
determined if we require that the range of x, namely 4, equals
the range of y;
(iii) the bounds on the parameter t (use tl = -2, to = 2, so that t
goes from -2 to 2, like x).
When you run parnorm.m (type parnorm and <Enter>) the effect is to
draw all the lines on the screen and the eye irresistibly picks out another
curve with a downwards cusp, which is tangent to all the lines. This
cusped curve is called the envelope of the lines. Using <Enter> after
 4.6 Envelopes of lines 53

2.5

1.5

0.5
NM
0

-0.5

-1
-2 -1.5 -0.5 0 0.5 1 1.5 2

Fig. 4.5. The normals to the parabola y = x2 and their `envelope', a downward
cusped curve.

the lines are drawn actually emphasises this cusped curve by drawing it
in white. See Figure 4.5.
The M-file parnorm. m is a special case of the M-file linen . m, which
draws a family of lines
p(t)x + q(t)y + r(t) = 0,
where p, q, r are any functions of a parameter t, over some pre-defined
range of values of t. Thus we obtain the M-file parnorm.m from linenv.m
by specifying the functions p, q, r in the correct place (see below). It has
been adapted to draw the parabola in red and (after pressing <Enter>)
the envelope curve itself in white, by adding a few lines to the file. See
Exercise 4.7 for a method of calculating the envelope curve. You can
read more about envelopes in Chapter 10.
The general M-file linenv.m is incomplete: before it will run you
need to specify the functions p, q, r, in the following place:

p= % These three lines are where the

q= % equation of the line, px+qy+r=0,
54 Graphs and Curves
r= % is placed. p,q,r are functions of t.

BEWARE!! that say q=4 will not produce a

vector but a constant
number. So if any of p,q,r is
constant, say k, then you need
to write it as k*ones(size(t)).
Here, ones((size(t)) is a vector of 1's of
the same length as the vector t.

In the particular case of the normals to a parabola (as in parnorm.m)

we need the equation of the normal to the curve y = x2 at the point
(t, t2), which is calculated at the beginning of this section.

Exercises
4.1 Try

>> a=1;
>> x=-5:.05:5;
>> y=x.*x.*x - a*x + 1;
>> plot(x,y)

Despite the fact that 1 is a number, MATLAB is sufficiently

intelligent to add it to the vector x. *x. *x whose entries are the
values of x3 for x = -5, x = -4.95, x = -4.9, ... , x = 5. Thus
1 is added to each entry in this vector.
Can you tell from the graph how many real roots the equation
x3-x+1=Ohas? Try
axis([-5 5 -3 3] )
This should help: it forces MATLAB to adopt -5 < x < 5 and
-3 < y < 3 on the axes.
The M-file cubics.m does the above, allowing you to choose
your range of values of x and y, and also the value of a. So
cubics <Enter> followed by, in succession, the five numbers
-5,5, -3,3, 1, each followed by <Enter>, plots y = x3 - ax + 1
with -5 < x < 5, -3 < y < 3, and putting a = 1. It also draws
the x-axis in white. Look at the M-file if you want to see how
it does this.
 Exercises 55

Try running cubics with other values of a and find approx-

imately the value of a at which the graph is tangent to the
x-axis, so that there is a double root. Once you have a fairly
good idea of the value of a, run cubics with 0.5 < x < 1 and
-0.05 < y < 0.05. You'll find you can tell very accurately
whether the graph is tangent to the axis. Find the best approx-
imation you can for a.
4.2 In the M-file t s ine . m, change
plot(x,z,'g') to plot (x,z.*z,'g'),and
plot (x,y,'r') to plot (x,y.*y,'r').
Call the resulting M-file say tsine3.m. Which functions are now
being plotted in red and in green? Give formulae for them. Run
the M-file to discover the least number of terms (k) you need in
order to get a good polynomial approximation for 0 < x < 4.
What is the degree of the resulting polynomial approximation?
(Hint: It's not k.)
4.3 For the `goat' problem (4.5), find the value of x close to 10
using f zero. Is there just one possible answer to the original
problem? (Look at the graph of the goat function.)
Try also the following, which approximates the goat function
y by a polynomial z of degree 3 (you could do this by direct
input to MATLAB or via an M-file):
X=-19:.05:19;
y=goatfn(x);
z=polyfit(x,y,3);
v=polyval(z,x);
plot(x,y,'g',x,v,'r')
This plots the true goat function in green and the approximation
in red. You should find that the two curves are very close over
the range from -19 to 19.
Type z to find out what the polynomial z is. Use roots (z)
to find its roots. Why is there an `extra' root for z which is not
related to a zero of the goat function?
4.4 Type x=.05:.01:2; and then type in the necessary lines to
draw the graph of y = xsin !, over the range 0.05 < x < 2.
(Beware of that innocent-looking 1. Compare 4.2 above.) Use
a sequence of commands similar to that in Exercise 4.3 to plot
a polynomial approximation of degree 7 on the same diagram
as the original curve. Why do you think the fit is so bad?
56 Graphs and Curves

Fig. 4.6. Generation of a hypocycloid by a circle radius b rolling inside a

circle radius a > b. The hypocycloid is generated by a point P fixed to a
radius CQ of the moving circle, where the length of CP is d. In the figure,
a = 3, b = 1, d = 0.7.

4.5 The M-file hypocy.m draws so-called hypocycloid curves (also

known as spirographs), which are obtained by rolling a circle of
radius b inside a circle of radius a > b. The curve is traced out
by a point P rigidly attached to the rolling circle, at a distance
d from its centre.
The parametrisation of the hypocycloid is

x = (a-b) cos (a bt b) +d cost, y = (a-b) sin (a bt b l -d sin t.

\ /J \ JJ
(4.2)
See the 4.7.2 and Figure 4.6. The M-file calls for the values of
a, b, d and for the upper limit of t, that is, 0 < t < nir, where n is
input by you. Draw the curve using hypocy.m for b = d = 1 and
a = 2. Explain what you see, using the above parametrisation.
Take a = 3, 4, 5, 6, 7, with b = d = 1 and find in each case: (i)
the smallest integral value of n which makes the curve close up
and start to repeat; (ii) the number of `cusps' (sharp points) on
the curve.
 Exercises 57

Can you formulate the general results here, for any integer
a>2?
4.6 Show that the perpendicular bisector of the line joining (a, b) to
(c, d) has equation
2x(a - c) + 2y(b - d) -a2 -b2+c2+d2 =0.
(Hint: Use the equal distance property: (x, y) is on the perpen-
dicular bisector if and only if its distances from (a, b) and (c, d)
are equal. This gives
(x - a)2 + (y - b)2 = (x - c)2 + (y -
Deduce that the perpendicular bisector of the line joining
(0, 1) to (t, 0) has equation
2xt-2y+1-t2=0. (4.3)

Complete the M-file linenv.m to plot the envelope of these

perpendicular bisectors for -2 < t < 2.
4.7 Suppose that the equations
px+qy+r=0 and p'x+q'y+r'=0 (4.4)

(where p, q, r are functions of t as in the above examples, and '

denotes differentiation with respect to t) are solved for x and y:

x pq -pqq) y pq, - pq. (4.5)

(This is just solving two simultaneous linear equations.) Thus

x and y are functions of the parameter t; as t varies the point
(x, y) traces out a curve.
A straightforward but tedious calculation shows that:
The tangent to this curve (x, y) at the point with parameter t is pre-
cisely the original line p(t)x + q(t)y + r(t) = 0.
Thus (4.5) gives the envelope curve, which is tangent to all the
lines.
In practice it is usually simpler to solve directly rather than
use (4.5). For example, with the normals to a parabola in 4.6,
we get the envelope from
x + 2ty - t - 2t3 =0 and 2y-1-6t2=0,

where the second equation is obtained by differentiating the first

58 Graphs and Curves
with respect to t. The second equation gives y directly, and then
we get x from the first equation:

X = -4t3, y = 2 (1 + 6t2).
This is the downward cusped curve in Figure 4.5. We can find
the equation of the curve in x, y coordinates by eliminating t.
Thus 6t2 = 2y - 1 so 216t6 = (2y - 1)3. Also 4t3 = -x so
16t6 = x2. Hence 216x2 = 16(2y - 1)3, that is,
27x2 = 2(2y - 1)3

is the equation of the envelope curve in this case.

Find the equation of the envelope in the case of the perpen-
dicular bisectors example (4.3). How does this relate to the
shape of the curve which appears when all these perpendicular
bisectors are drawn on the screen?

4.7 Appendix
4.7.1 Proof of the goat equation
Referring to Figure 4.4, the region the goat, tethered at G, can graze
is enclosed by a circular arc from G along the boundary of the field to
A, then along a circular arc centred at G to B and finally along the
boundary of the field back to G. The area of this region is the sum of:
(i) the area of a circular sector of the circle centre G, radius x,
bounded by the straight lines BG, GA and the arc centred at
G from A to B;
(ii) twice the area of the piece of the field below the chord BG.
The area of the latter piece is the area of a circular sector of the field
from C along the radius to B, along the circumference to G and along
the radius to C, minus the area of the shaded triangle.
Let 0 be the angle CGA (equal to the angle CGB). The area (i) is
2x2(20) and 0 = cos' Zo, so the area is x2 COs-1 20. The area of the
piece in (ii) is

x102x2(-0)-x 100- ()2.

Adding (i) to twice (ii) and putting this equal to half the area of the
circular field, 507r, gives equation (4.1).
 4.7 Appendix 59

4.7.2 Proof of the parametrisation (4.2)

Using the notation of Figure 4.6, the centre C of the rolling circle has
reached an angle u anticlockwise from the horizontal, so that
C = ((a - b) cos u, (a - b) sin u).
At the same moment, the line joining C to the moving point P has
rotated an angle t clockwise from its initial horizontal position. Here,
we think of the small circle starting off with C on the positive x-axis, P
starting at PO. The rolling condition is that the arc of the large circle
from Po to A equals the arc of the small circle from A to Q, where CPQ
is a straight line as in the figure. Hence
bt
b(u + t) = au, so that u =
a-b*
The formula (4.2) follows because the point P is
((a - b) cos u, (a - b) sin u) + d(cos t, - sin t);
the last - sign occurs because t is measured clockwise.
 5
Representation of Data

As technologies advance, the amount of data becoming available in any

practical application area is increasing rapidly. The purpose of repre-
senting data in another form is to help make use of this mass of data
and to extract further information from it in a useful way.
In this chapter we present an introduction to data analysis using some
of the elementary statistical tools within MATLAB. We consider char-
acteristics within a data set, relationships between data sets and the
extraction of particular data from data sets. Thus the topics we cover
here coincide with the elementary functions of many modern statistical,
spreadsheet and database packages.

5.1 Data analysis

Given a set of data D = [dl, d2i ... , dn], we can compute its simple
statistics such as maximum, minimum, mean and median (the value
with 50% of the data above it and 50% below it). The corresponding
MATLAB commands are max, min, mean, median. Here D can also
be a column vector or a matrix.
To introduce some more commonly used commands, let us take
>> D = [21347]
as an example. Then Table 5.1 shows how the statisticst of this data set
f If we denote the mean of D by d, then a measure of the dispersion or variability
is given by the standard deviations with

(d_n22)
s2 (dz - d)2 =
nI
I

60
 5.1 Data analysis 61

Table 5.1. Computing the statistics of a data set (vector D).

Description MATLAB input line Result
Minimum m= min(D) 1
Maximum M= max(D) 7
Mean p = mean(D) 3.4
Median P = median(D) 3
Product of elements in D pr = prod(D) 168
Sum of elements in D dsum = sum(D) 17
Cumulative Sum csum = cumsum(D) [2 3 6 10 17]
Standard deviation s = std(D) 2.3022
Sorting D so = sort(D) [1 2 3 4 7]

can be calculated with MATLAB. Here note that sorting is only done in
the ascending order; so what result would the following produce?

>> D1 = - sort(-D)

A useful aid for displaying data is the histogram. The MATLAB com-
mand is simply hist; for example, hist(D) will display the histogram
of the data set D. By default, the interval min(D) < x < max(D) is
partitioned into ten equally spaced subintervals and the height of the
bars is the number of data in the subinterval. We can vary the number
of bars; to have three subintervals, type

>> hist(D, 3)

So, here, the intervals are 1 < x < 3, 3 < x < 5, 5 < x < 7 and the bars
are centred on 2, 4, 6 respectively. To find the sizes (y) of the groups,
type

>> y = hist(D, 3)

and to find the sizes (y) of the groups and mid-points (x) of the intervals,
type

>> [y x] = hist(D, 3)
Here the results are y = [3 1 1] and x = [2 4 6].
There is a link with bar charts: type help bar to find details on
using bar and you will see that bar (x, y) produces an identical plot to
hist(D, 3).
62 Representation of Data
5.1.1 Sorting
Often it is necessary to sort a number of data sets so that one of them
is in increasing or decreasing order. For instance, when there are two
related data sets, the second set must be reordered consistently with the
first set.
An example of this is in the M-file tomato.m, which compares the
amount x of fertiliser applied to a tomato plant with the weight y of
tomatoes produced.
Typing tomato produces two vectors x and y but the entries in x
are not in increasing order. (You can see the vector x by typing x, of
course.) Typing plot (x, y, ' g*') produces a scatter plot of the data, x
against y, the symbol used being green `*'.
Typing plot (x, y) produces a line plot, but because the data are not
ordered it appears as a scribble rather than a graph. One way of making
a proper plot is by typing
[sx k] = sort(x);
sy = y(k);
plot(sx, sy)

The first line sorts x into increasing order, noting the positions of their
components in the original order in the vector k. The sorted x are placed
in a vector sx. The second line reorders y so that their components
correspond to those of x correctly. Type [x y sx sy] to see the four
columns of figures at once.
You will find that the line plot now looks more like a graph. Finally,
as we mentioned earlier in this chapter, it is easy to sort data into a
decreasing order with sort by taking the negative of the values.

5.1.2 Querying
Querying or extracting data lies at the heart of databases, now increas-
ingly used in various management information systems. Here we shall
take two simple examples to illustrate. The essential quantity to specify
in using query is a condition often called the criterion. In MATLAB,
this condition can be represented by a vector (or a matrix) with Os and
is, meaning `not satisfied' and `satisfied' respectively.
As a first example, consider those tomato plants which give more fruit.
The task is to show a line plot of fertiliser versus weight for those with
over 5 kg of fruit. To identify such cases out of the entire data set, we
 5.2 Least squares fitting 63

use a vector d with is and Os, count the total number t of the Is, and
extract the satisfied cases in new vectors x1 and yl (both sized t):
>> tomato;
>> d = (y>=5); % Query Condition !!!!
>> t = sum(d) % Number of `satisfied'
>> [v k]=sort(-d); % Bring `satisfied entries' to start
>> k = k(1:t) % Shorten the index to `t' (Query)
>> xl=x(k); yl=y(k); % Pick up query results from x,y

If we now wish to plot xl versus yl, a sorting of both according to xl

has to be done first, as we discussed before. Here t = 7.
The second example is a simple geometric task. Given a set of 42
points (x, y) on a unit circle, we are to identify those points which lie in
the second quadrant, and plot them:
format compact; t=0: 0.15: 2*pi; n_count=length(t)
x=cos(t); y=sin(t); % ------ the given data set---
d = (y>=0) & (x<=0); % Query Conditions !

t = sum(d) % Number `satisfied'

[v i] =sort (-d) ; % sort `satisfied entries'
i = i(1:t) Shorten index to `t' (Query)
xl=x(i); yl=y(i); Pick up query results.
plot(xl,yl) 0 ------ plot them -----------

Try it. Did you find t = 10?

5.2 Least squares fitting

Given a pair of data sets (say x and y in the `tomato' example above)
we may infer, by inspecting their scatter plot, a relationship or associ-
ation between them. We want to find a function which approximately
describes such a relationship. Of course, different choices of functions
can be tried for such a purpose depending on what the scatter plot
looks like. Suppose that such a pair of data sets is given by (xi, y2) for
i = 1, 2, ... , n. We can conveniently write these as two vectors:
X = Ixl, x2, ... , xnj T ,

Y = [Y1, Y2, ... , ynIT

Here, we have taken them to be column vectors.

64 Representation of Data
5.2.1 Fitting a straight line (the least squares line)
Perhaps the easiest function to try is the linear function
Y=a+/3x,
that is, we try to find two scalars a, /3 such that the linear relationship
Yl = a +/3x1, Y, 1 x1
Y2 = a +/3x2, Y2 1 x2
that is, _
Yn = a + /3xn, Yn 1 xn

produces Y = yi for all i. We denote this linear system by Ac = Y with

cT = (a /3). Do such scalars exist? This is an over-determined system
of linear equations for unknown quantities a and /3 which means that
the system cannot be satisfied exactly,t in general, for any a and /3.
An arbitrary choice of a and /3 would give an error of size l yi - Yi
for each i, and trying to satisfy the system approximately implies a
compromise with such a residual error. The Gaussian method of least
squares says that$ we should find that unique solution a and /3 which
minimizes the sum of residual squares
n
R2 1:(Y, - Y)2>
=
i=1

where for all i,

Yi=a+/3xi.
5.2.1.1 First method: using \
In MATLAB, given data sets x and y, we can find the best solution a
and /3 easily by doing the following
>> A=[ones(size(x)) x]
>> c=A\y

Here, A is the matrix appearing above, with is down its first column
and x down its second column. The second line instructs MATLAB to
find the best approximation c to a solution of the equations Ac = y. As
defined before, c(1) = a and c(2) = /3.
t Mathematically the solvability depends on whether vector Y lies in the span of
the two column vectors of A! As rank(A) < 2 and often in practice n > 2, the
system is more likely to be inconsistent, that is, have no solution!
t Further details are given in an appendix to this chapter.
 5.2 Least squares fitting 65

Once a and 0 are available, we may use the relationship

Y=a+,6x
to find any points on the best-fitting line; for example, the following

>> Y = c(1) + c(2)*x;

>> plot(x,Y)

plots the best-fitting line (the least squares line).

To find the residual R2 of the least squares fit, type (noting that y,
Y, r are column vectors)

>> Y=A*c ;
>> r = y - Y
>> R2 = r'*r

5.2.1.2 Second method: using polyf it

Given data as above, type

>> C = polyfit(x, y, 1)

to find the best-fitting curve of degree 1 (that is, a line). This gives
C(2) = a and C(1) = 03. Notice that c was a column vector while C is
a row vector! To find the residual of the least squares fit using the row
vector C, type
>> Y = polyval(C,x);
>> r=y - Y;
>> R2 = r'*r

Remember from 4.1 that polyval evaluates the polynomial C at the

values of x in the vector x. That is, it finds the points on the best-fitting
line. Since x, y are column vectors, so are Y, r.

5.2.1.3 Example data: tomatoes again

The data have been described in 5.1.1. Some information can also be
obtained by typing help tomato. You should find the least squares
solution is a = 4.3985 and 0 = 0.0966, and the residual R2 = 0.388.
You can fit and then plot a straight line without sorting the data, but
if you want to fit and then plot any other curve (see 5.2.2) then you do
need to sort the data.
66 Representation of Data
5.2.2 Fitting a polynomial function
If the relationship between data sets x and y is not really linear, we may
try nonlinear functions. The simplest example is the use of a quadratic
function
Y=,32x2 + i31x+a.
As before, we want to find coefficients ,32i /31 and a such that the fol-
lowing residual is minimised
n
R2
= E(yi - Y)2,
i=1
where for all i
Yi = /32x? + (31xi + a.

With MATLAB simply type

>> C = polyfit(x, y, 2)

to obtain the coefficients C(1) = ,32i C(2) = /31 and C(3) = a. To find
the residual of the least squares fit using the row vector C, type
>> Y = polyval(C,x);
>> r = y - Y ;

>> R2 = r'*r

Typing help polyf it, you will find out that polyf it can fit a poly-
nomial function of any degree k to data sets x and y. The general
function will be of the form
Y=/3kxk+...+/32x2 + ,31x+a

and so the MATLAB command C = polyfit(x, y, k) gives the re-

quired coefficients in a row vector form
C = [/3k, ... , 132, 01, 011
Here the estimated values at all xis can be found by
>> Y = polyval(C,x);

So with polyf it, higher order polynomials can be considered with ease.
But it cannot deal with the multiple variable case as discussed in 5.2.3.
For those who enjoy the beauty of the \ method, you will be pleased
to know that this can be done. With the following hint on how to set
up A, you should obtain the same solution as with polyf it
 5.2 Least squares fitting 67

>> A=[ones(size(x)) x x."21; % for component-wise

Remark: The nonlinear relationship between data sets x and y need not
be polynomial; they may be other functions, for example, we can try the
following
Y=132e-x+/3ieX+a

or

Y = /32 sin x+/31 cos x + a.

5.2.2.1 Example data: same tomatoes

Using the data sets in tomato.m, we can use polyfit(x, y, 2) to fit a
quadratic function. The coefficients will be
[/32 /31 a] = [-0.0045 0.1776 4.1826]
and the residual R2 =0.217. Finally use polyfit(x, y, 3) to fit a
cubic function. The coefficients will be
[/33 /32 /31 a] = [0.0004 - 0.0144 0.2451 4.1088]
and the residual R2 =0.1905.
Typing toms plots a linear fit, then a quadratic fit without sorting and
finally a quadratic fit with sorting. (Hit <Enter> between plots to clear
any pause.) Clearly sorting is vital for plotting quadratic fits! Look at
the M-file toms.m to see how this is done, using sorting as in 5.1.1.
Here is the relevantt part:
[sx o] = sort(x);
sy = y(o);
C = polyfit(sx,sy,2);
Y = polyval(C,sx);
plot(sx,sy,'*w',sx,Y,'w')

5.2.3 Data fitting for multiple variables

The above methods assume that we have an `output' dependent variable
y which is a function of a single `input' independent variable x. In the
tomato example, the weight of tomatoes y is assumed to depend solely
on the weight of fertiliser x; we have ignored sunlight and moisture
t Refer to Chapter 17 for plotting curved lines with sufficient plot points.
68 Representation of Data
among other possible factors! It is often more realistic to assume that
the `output' data y is a function of several independent `input' variables,
and here we show how to handle this case on the assumption that the
relationship is linear in all input variables. This assumption is only made
to simplify the presentation; higher order polynomial relationships can
be considered similarly. Thus, calling the `input' variables x(1), ... , x(P),
we are trying to find a relationship of the form

y=a+131x(1)+...+0Px(P)
in which x(i) occurs only to degree 1. In the example below, the response
y to treatment of diabetics is assumed to depend on three factors, x(1) =
age, x(2) = weight and x(3) = diet (which is converted into a numerical
value according to some scheme).
We are often dealing with discrete data sets and so we assume p vari-
ables (or `input' factors) and n values ('output' observations) of each
variable which are put into a vector:

Y [Y1 Y2
... yn1T

X(1) [x(1) x(1) ... x(1)]T

X(2) . [x(2) x(2) ... x(n2)]T

X(P)X(P)
[ 1 2
... x(P) T
j

As above with p = 1, we could assume a relationship between y and the

independent variables x(1), x(2), ... , x(P) of the form

Y = a + Olx(1) + 02x(2) + ... + OPx(P).

Consider the case of p = 3 (or 3 `input' factors). The problem is to

find

Y = a + 131x(1) + 132x(2) + 03x(3)

(that is, a particular set of values for coefficient vector c) that minimises
n
E[y2 - Y12,
R2 = IIY - YI 2 =
i=1
 5.2 Least squares fitting 69

where c = (a /31 02 /33)T,

Y1 Yi 1 x11 x12 x1
Y2 Y2 1 x21 x22 x23
Y= Y= . A=
yn Y. 1 xn1) xn2) xn3)

This least squares problem can be written in an equivalent form (suitable

for MATLAB)
Ac = y.
You must have realised that this is the \ method! Yes, and the same
idea can be applied similarly to the higher order polynomial fitting (or
other nonlinear function fitting) for this multiple variable case! As pre-
viously remarked, the polyf it method can no longer help. As it stands,
polyf it only deals with the single variable case. In the project of Chap-
ter 17, polyfit2.m will be developed for treating the p = 2 case.
To find the coefficient column vector c = [a i31 /32 /33]T and the
residual R2 = R2, the MATLAB commands will be (using xl for the
vector x(i), x2 for x(2) and so on)
>> A=[ones(size(xl)) xl x2 x3];
>> c=A\y; Y=A*c; r=y-Y;
>> R2 = r'*r

5.2.3.1 Example data: diabetic.m

The M-file diabetic.m contains the data and a brief explanation of it
so type

>> help diabetic

>> diabetic

The data sets will be kept as y, xl, x2, x3. You should find that
the coefficients are cT = [36.9601 - 0.1137 - 0.2280 1.9577] and
R2 = R2 = 567.6629.
Then one may infer that the following function (that is, our least
squares solution) describes the response to treatment
Y = Y(x(1), x(2), x(3)) = 36.9601 - 0.1137x(1) - 0.2280x(2) + 1.9577x(3),

where x(1) = age, x(2) = weight and x(3) = diet.

70 Representation of Data
Exercises
5.1 The M-file marks. m contains the marks obtained by a class of
engineering students, on a differential equations course, in a
class test and in the homework prior to test. If you run the
M-file by typing marks, you will generate a copy of the marks
in matrix mkdata. Typing
>> x=mkdata(:,1);
>> y=mkdata(:,2);

pulls out the first and second columns of data and names them
as x (test marks) and y (homework) respectively. Now try the
following:
(a) Compute the least squares line (see 5.2)
Y=a+Ox
and the least squares line
X =ry+Sy.
In this problem it is not possible to state that one variable
is dependent and the other independent, but it is reason-
able to assume some relationship. If the relationship was
exact (that is, if the residual of the least squares fit was
zero), then the two lines would be coincident. Find the
equations of the two lines and describe briefly the method
you used to obtain your answer.
(b) Plot the data as a scatter plot, and superimpose on it
a plot of the two lines. Calculate the value of the angle
between the lines.t Again, describe the method you used
to obtain your answer.
Hint: you may use this formula
tan 01 - tan 02
tan(Bl - 02) =
1 + tan 01 tan 02
5.2 The M-file mannheim.m gives some production figures for car
power steering units together with the cost of production. After
running mannheim, the data will be in a matrix called mdata
of size 22 x 3. We shall name the first column of mdata as x
and second column as y. Now try the following:
t The relationship between data sets is known as correlation. Here the angle is an
indicator of how the data sets are correlated.
 5.3 Appendix 71

(a) Begin by sorting the data according to increasing x, nam-

ing the sorted data as sx, sy.
(b) Assume that y is the cost of production and x is the
number of units produced. Fit a linear relation y = a +
/3x to the given data set.
(c) Define two new independent variables x(1) and x(2) as
follows; we omit interpretation of their practical mean-
ings

X,
x(1) = j x < ry x(2) = rll 0, x < ry
tt ', x?'Y, x - 'Y, x ?'Y

with ry = 3000.
Firstly, work out the new data sets xl = x(1), x2 = x(2).
Secondly, use the method of 5.2.3 with the sorted data
and p = 2 to compute a least squares fit of the form
y = a +,1x(1) + 02x(2)

and find the residual of this fit. Finally, plot the non-
linear fit using plot (sx,sy,'*',sx,Y), where Y is, as
usual, the points of the nonlinear fit, as in 5.2.3. Does
the graph appear to fit the data substantially better than
the linear fit? Suggest a possible (real life) reason why
the fit using two straight lines is so much better than the
fit using one.
Hint What might happen once the number of units be-
ing produced exceeds a certain number? Try also help
mannheim. Note that x(1) = min(-y, x) and x = x(1) +x(2)
for all values of x.
EXTRAS Consider a further fitting
y = a + 011x(1) + 021x(2) (X(l))2
+012 + 022(x(2))2
or alternatively
(1) 2)
y=a+/31e D +/.32e

5.3 Appendix
Here we consider the minimisation of R2 by selecting a and 0; see 5.2.1.
To proceed, we first prove the famous Cauchy-Schwarz inequality.
72 Representation of Data
Theorem 5.1 (Cauchy-Schwarz inequality) For any x, y E Rn we
have
n
EIxjyjI <- IIXII2IIYII2.
j=1

Proof (i) We only need to consider x 0 and y 0, as otherwise the

result is trivially 0 = 0;
(ii) We claim that it is sufficient to prove:
n
Exjvj < IIXII2IIVII2 (5.2)
j=1
for any nonzero vectors x, v E R. For if this is true, we may construct
v by vj = sign (xj)sign(yj )yj for any y E Rn and j = 1, ... , n so that
(5.1) is valid;
(iii) To prove (5.2), note that for any A E R
0 < Ilx + Av112 = (x + AV)T(X + Av),
so
n
2A IxjvjI < IIXII2+A2IIvII2
j=1
Now select a specific A = IIxIl2/IIv112, simplifying the above inequality as
n
2 E Ixjvj I < 2IIXII2IIvII2,
j=1

which proves inequality (5.2). Thus inequality (5.1) is proved.

5.3.1 Derivation of the least squares method

As R2 is quadratic in a, the minimum values may be found by solving

R and a R2 = 0.
The least squares equations can be written out as
a
E(yi - a - /3x2)2 = (-2) J:(yi - a - /3x2) = 0,
as i i
C a - ,3x2)2 = (-2) E xi(yi - a - /3xi) = 0,
i i
 5.3 Appendix 73

that is,

na +
li
(Xi)=Yi
Zl

xi l
a+ X2) xiyi.

Using the usual matrix notation, cT = (a 0),

x1
r11 x2
A= x=
\ 1 xn yn

we can write (and define)

4)(c) = R2(c) = (Y - Ac)T(y - Ac).

Then the above least squares equations reduce to

ATAc=ATy.

Mathematically speaking, to solve this 2x2 linear system, we first

need to ensure

det(ATA) = n E x2
i i

which is true if at least two xis are distinct! Then we can easily find the
unique solution for c = (a 3)T

where x and y are the means of data sets x and y respectively.

To convince ourselves that the above solution is a minimum instead
of a saddle point, we need to work out all the second derivatives (which
74 Representation of Data
are constants in this quadratic case). We may verify that
a2 4) a2.D 82.D
2n, I
aa2 aaa)3 a,Qaa
a2.D
02 = 2 E x?.

Then from the Cauchy-Schwartz inequality, the discriminant is negative

since
2
2 ) (92,D
-nEx? <0
C J- as aa2J
4
CExz/f
fl 7
Z

whenever det(ATA) 0 0, and both as > 0 and aa2 > 0 as at least

two xis are distinct (that is, they have no chance to be zero at the same
time). This proves that the extreme point is indeed a minimum.
An alternative and simpler proof is as follows. Consider the following
functional for all w = c + h E R2 (space)
4) (w) = (y - Aw)T(y - Aw) = wTATAw - 2yTAw+yTy.
We can first verify that
(D(c + h) = 4)(c) + (Ah)TAh + 2hT (AT Ac - AT y).
Assuming AT Ac = AT y, we then have
R2(c + h) = 4(c + h) = 4)(c) + IIAhII2 > (D(c) = R2(c).
Therefore the solution to equation (5.3) is the unique minimum of the
least squares problem.
 6
Probability and Random Numbers

The study of probability and random, or stochastic, processes remains

an important subject, simply because most real world problems exhibit
random variations. Such variations can give rise to nondeterministic
factors that definitive mathematics, even with approximations, cannot
immediately describe. The probability of an event in classical probabil-
ity theory is defined as the ratio of the frequency of the event to the
total number of all possible outcomes. Often it is more convenient to
model probability by a density distribution for both discrete and contin-
uous random systems. Fortunately for many practical problems, suitable
probability density functions, representing probability distributions, are
known.
In this chapter, we introduce the elementary probability distributions
and simulate them by generating appropriate random numbers with
MATLAB.

6.1 Generating random numbers

The MATLAB function rand generates pseudorandom numbers on the
interval (0, 1). These numbers are pseudorandom because they appear
to be random sequences but there is a method to duplicate them! The
sequences are generated by a deterministic algorithm but they can be
`seeded' to yield a particular sequence.t
To see how this command works, try the following
>> rand % Generates a random number in (0,1)
>> rand('seed',13) % Set the `seed' of algorithm to 13

t The algorithm is based on a multiplicative congruential method. See the appendix

to this chapter for details.

75
76 Probability and Random Numbers
bi = rand(25,1) % A random column vector in (0,1)
rand('seed',O) % Reset the `seed' to DEFAULT.
A = rand(3,4) % 3x4 random matrix in (0,1)
rand('seed',13) % Reset the `seed' to 13
b2 = rand(25,1) % Duplicate bi !

It is possible to `increase' the randomness by starting the sequence of

numbers at a `random' term, for example, by linking the seed to the
internal clock of the computer. The current clock time is
>> time=round(clock) % round can be fix, ceil or floor

and this vector represents time=[Year Month Day Hour Min Sec] so
we could use
>> rand('seed',time(6)) %or rand('seed',time(5)*time(6))

to link to the nearest second, though this doesn't always seem to work as
well as setting the seed yourself, preferably to an odd number. Numbers
in the range 10000-20000 (or actually the default 0) are good choices.
The simple MATLAB function rand generates a column of n random
numbers uniformly distributed in (0, 1) by x=rand (n, 1), or a row by
x=rand (1, n) . Similarly, rand (n) generates a n x n random matrix
while rand(m,n) produces a m x n matrix with entries in (0, 1).
To generate random numbers in any interval (a, b), use a linear trans-
formation (x --* y), for example
>> x = rand(1,30); % 30 random numbers in (0,1)
>> a = 12; b = 99; % set up the interval
>> y = a + (b-a) * x; % 30 random numbers in (a,b)

Notice that the elements of vector y are now within (a, b).

6.2 Random integers

MATLAB functions ceil, fix, floor and round, combining with rand,
generate random integers. Here ceil (x) is the smallest integer > x,
floor(x) is the largest integer < x and round(x) is x rounded to the
nearest integer. The command fix mixes ceil and floor depending
on the sign of a number, because it rounds numbers towards zero to the
nearest integer. For instance, test the followings out
ceil(1.2)=2, ceil(3)=3, ceil(1.9)=2, ceil(1.5)=2,
fix(1.9)=1, fix(2)=2, fix(-1.6)=-1, fix(-2)=-2,
 6.2 Random integers 77

floor(1.2)=1, floor(3)=3, floor(1.9)=1, floor(1.5)=1,

round(1.2)=1, round(3)=3, round(1.9)=2,round(-1.5)=-2.

We can get a column of n numbers taking integer values 0, 1, ... , k

(that is, integer values in the closed interval [0, k]) with equal probability
using

>> x = rand(n,1); floor((k+1)*x)

We can get a column of n numbers taking integer values 1, . . . , k (that is,

integer values in the closed interval [1, k]) with equal probability using
>> ceil(k*x)

The use of round is a little different. For example, the following

>> round(10*x)
generates n numbers which do not take the values 0, 1, . . . , 10 with equal
probability. (Can you see why? Compare the x-values rounded to 0 with
those rounded to 1, and consider also round( 9.99 x + 0.5 ).)
If we simply want to generate a random permutation of k integers
1, 2, ... , k, use the command randperm; for example, we may permute
the rows of a matrix in a random manner
>> r = randperm(9)
>> A = ceil( rand(9)*5 ) Generate a matrix in (1,5)
>> B = A(r,:) Permute rows via "r"

Here is a cunning way to produce a column vector of Os and 1s, where

the Os occur with `probability 3'. That is, for a very large vector, one
third of them are likely to be 0 and the rest 1. Type
>> rand('seed',O); x=rand(100,1); % Note x in (0,1)
c=(x>1/3)

The second line here instructs the computer to print 0 if x `fails' the test
x > s and 1 if x `passes'. Assuming that x has a uniform distribution
in [0, 1] you should get Os in approximately s of the cases. You can test
this by typing sum(c)/100 and seeing whether the answer is close to 3
However, a more challenging method than c=(x>1/3) is the following

>> rand('seed',O); x=rand(100,1); % Note x in (0,1)

>> c=round(x + 1/6)
78 Probability and Random Numbers
Here the trick is that, by a shift of 6, only numbers in x less than
2 - = a are rounded to zero! Can you draw a picture to illustrate this
idea?s In a similar way, you can experiment with any probability p. Can
you suggest a formula for the shift?
Example Starting from seed=121, to generate two random integer ma-
trices of size 2x6 with elements in [1, 4], and [0,3] respectively, we do
the following

>> rand('seed',121); x = rand(2,6)

>> y = ceil( 4*x ) % Case (1) with [1,4]
>> z =floor( 4*x ) % Case (2) with [0,3]

Here are the results:

y= 1 2 2 1 3 2
3 3 4 2 1 1

z= 0 1 1 0 2 1

2 2 3 1 0 0

6.3 Simulating uniform distributions

We have discussed that the MATLAB function rand simulates numbers
from a uniform distribution U(0,1).t As mentioned, we can use random
numbers to simulate uniform distributions in any interval when we com-
bine U(0,1) with a linear transformation; for your convenience we have
developed an M-file called unirand.m.
Example To simulate 5000 rolls of three dice, we try
>> rand('seed',19) % Set 'seed' to 19
>> z=unirand(0.5,6.499,5000,3);% Random in (0.5, 6.499)
>> r = round(z); % Random numbers in [1,6]
>> [x y]=hist(r,6) % Count faces in x

6.4 Simulating normal distributions

To obtain a column of 1000 random numbers with normal distribution
having mean 0 and standard deviation 1, type
>> randn('seed',12); x=randn(1000);

t See the appendix to this chapter for a description of uniform distributions.

 6.5 Simulating negative exponential distributions 79

where seed is similar to that used for rand, but is specific to randn!
Using again a linear transformation, to obtain numbers with mean m
and standard deviation s, type
>> y=m+s*x;
For your convenience we have developed an M-file called normrand.m.
Example To simulate heights of people in two different regions, both
with mean m = 3 and standard deviation s = 5, we take a sample of
2000 people from each region and do the following
>> randn('seed',11) % Set 'seed' to 11
>> y = normrand(3,5,2000,2);% N(3,5) with mean=3 & Std=5
>> hist(y) % Plot the simulated data y

or without normrand.m
>> randn('seed',11) % Set 'seed' to 11
>> x = randn(2000,2); % N(0,1) of mean=0 & Std=1
>> y = 3 + 5*x; % Transformation
>> hist(y) % Plot the simulated data y
and you should find that mean(x) = [0.0318 - 0.0346], mean(y) _
[3.1591 2.8270] and std(y) = [4.9989 5.0467].
Again as with rand, to obtain random integers, combine randn with
any of the functions ceil, fix, floor and round. To get, for example,
seven columns of nine whole numbers with normal distribution of mean
m = 3 and standard deviation s = 4, type
>> n1 = 9; n2 = 7; x = randn(nl,n2); m=3; s=4;
>> y = round(m + s*x);

Remark: The two kinds of distributions are distinguished by rand and

randn. By default, rand produces a uniform distribution. However,
commands rand('normal') and rand('uniform') can force rand to
toggle between the two distributions! We discourage this practice, as
one may lose track of this toggling unless this is set each time rand is
used!

6.5 Simulating negative exponential distributions

As the negative exponential distribution is related to the uniform distri-
bution,t its simulation can be done easily using the MATLAB random
t See the appendix to this chapter.
80 Probability and Random Numbers
number generator rand. To simulate a negative exponential distribu-
tion, with both mean and standard deviation being , we can use the
following relation
z = -f log(1 - x),
assuming that x is from a uniform distribution U(0,1). For your conve-
nience we have developed an M-file exprand. m, generating this distribu-
tion.
To illustrate, we can simulate and compare samples of n = 20000
numbers of
a uniform distribution in (0, 2.4) with mean m = 1.2;
a normal distribution with mean m = 1.2 and standard deviation
s = m = 1.2;
an exponential distribution with mean f = 1.2:
>> rand('seed',O); n=20000; % Reset the 'seed' value
>> m=1.2; s=1.2; mean and std
x = unirand(0,2.4,n,1); Uniform in (0,2.4)
randn('seed',0); Reset the 'seed' value
y = normrand(m,s,n,1); Normal (m,s)=(1.2,1.2)
rand('seed',0); Reset the 'seed' value
o10
z = exprand(m, n,l); Exponential (mean m)
subplot(131); hist(x,90) IO Plot x
subplot(132); hist(y,20) Plot y
subplot(133); hist(z,40) Plot z

This example is available in the M-file c6exp. m. The results can be seen
in Figure 6.1, where one graph shows an exponential distribution! The
mean and standard deviation of a negative exponential distribution are
both f = m and so you may use MATLAB commands mean and std to
verify that, for z, both quantities are close.
As discussed in 6.6, in practice, inter-arrival times satisfy the negative
exponential distribution. For instance, take the example of five arrival
times (occurrences) of an event as arr = [1.3 2.4 5.1 6 8.3]. Then inter-
arrival times are the time differences between arrival times (occurrences),
that is, int_arr = [1.3 1.1 2.5 0.9 2.3] counting from t = 0.
So we can simulate inter-arrival times by exprand. m but how do we
work out arrival times then? Each arrival time is a cumulative sum of
inter-arrival times; for the above example, given int_arr, we work out
arr by
 6.5 Simulating negative exponential distributions 81

200

100

0
0 0.5 1 1.5 2
4000

2000

0
-4 -3
-0
-2 0 1 2 3 4 5 6

6000

4000

2000

0
0 1 2 3 4 5 6 7 8 9 10

Fig. 6.1. Illustration of three probability distributions.

arr(1) = int_arr(1) = 1.3;

arr(2) = int_arr(1) + int_arr(2) = 2.4;
arr(3) = int_arr(1) + int_arr(2) + int_arr(3) = 5.1;
arr(4) = sum of `int_arr(1) +...+ int_arr(4)' = 6;
arr(5) = sum of `int_arr(1) +...+ int_arr(5)' = 8.3.

In MATLAB, the command for cumulative sums is simply

>> t = cumsum(z) ;

For the above example in Figure 6.1, the cumulative sum for z can be
written as
k
tk=Ez2, k=1,2,...,20000.
=1
Here t(1) = 0.2966 and t(20000) = 24036.2043.
Example There is a concern about the long queues at a bank cash
machine and we have been asked to simulate the arrivals during peak
times of the day. It is known from computer records that (on average)
there are 131 customers between 9 am and noon. In order to simulate
the arrival of the first 50 customers, first work out the mean f = isi
since the average inter-arrival time is 13 minutes. Hence type
82 Probability and Random Numbers
m = 180 / 100; rand('seed',O);
z = exprand(m,50,1); hist(z)

As z is the inter-arrival time, the arrival time t is simply

>> t=cumsum(z)
with t(1) = 0.3396 and t(50) = 69.3055 minutes.

Exercises
6.1 An electrical retailer estimates that, in the weeks before Christ-
mas, there are, on average, 140 customers per day for electronic
games, with inter-arrival times satisfying a negative exponential
distribution. (A shop day is 9 am-5 pm.) On average 3 pay cash
and the rest use a credit card, 50% of the sales are for Nintendo
(Super Mario) and 50% are for Sega (Sonic the Hedgehog). You
are asked to simulate (the first) 30 sales from the opening time,
assuming that each customer buys a product.
(a) Simulate the inter-arrival and arrival times for the first
30 customers of the day. Find the mean and standard
deviation for the inter-arrival times.
(b) Print out a table simulating the first 30 sales in the form
of Arrival Time, in minutes with 1 decimal place, Pay-
ment Method (Cash as +1 and Credit as -1) and Product
Sold (Nintendo as +1 and Sega as -1)
Hints:
Your answer to question (b) will be three columns (all
of size 30):
one set of arrival times,
one set to select cash or credit card and
one set to select Nintendo or Sega.
Consult 6.2 to see how to obtain numbers which are
0, 1 with probability 13 , 3, or both with probability 21.

How do you then convert such a vector of Os and is

into -Is and +ls ?
MATLAB cannot print different columns of the same
matrix in different formats; they all need to be integers
or all decimals. To change this you need to use the
command fprintf or sprintf. For example, try
 Exercises 83

>> a=pi, b=35, c= 15

>> c_1= [abc]
fprintf('c_2 = %5.3f %5d %d\n',a,b,c)
c_3 = sprintf('%5.3f %5d %d\n',a,b,c)

where `%5.3f' means printing a real number with three

decimals in a total length of five, `%5d' means printing
an integer with five digits while `%d' prints a default
number of digits. To print the Arrival Time, you need
`%6.lf'.
6.2 It is estimated that, on average, 15 trains per hour arrive at
Liverpool Lime Street station, with inter-arrival times satisfying
a negative exponential distribution. In addition it is known
that 75% of the trains are small local diesel units on which
the number of passengers (per train) is uniformly distributed
on [0, 80]. The remaining 25% are inter-city trains on which
the number of passengers is assumed normally distributed with
mean 90 and standard deviation 20.
(a) Make a 15 x 5 matrix simulating the train types (0 =
inter-city, 1 = diesel) for the first 15 trains on 5 days.
(See 6.2 again for obtaining a column of numbers which
take the value 0 with probability 4 and 1 with probability
3
4')
(b) Use the given distributions to simulate the numbers of
passengers on 15 diesel trains and 15 inter-city trains for
5 days. (So your answer will be two matrices, each 15 x 5,
one matrix for each train type.)
(c) Combine (a) and (b) to simulate the total number of
passengers on the first 15 trains arriving at Lime Street,
for 5 days.
(d) Simulate the arrival times of the first 15 trains on 5 days,
and use this information and (c) to estimate the total
number of passengers arriving in the first half-hour of
each day.
(e) Make a plot of the cumulative passenger totals against
arrival times for the five days on the same graph.
Hint If the arrival times are in a 15 x 5 matrix arr and
the cumulative passenger totals are in a 15 x 5 matrix
cumpass then plot (arr, cumpass) plots all five graphs at
84 Probability and Random Numbers
once. From help plot or arr (15 , :) and cumpas s (15 , :) ,
can you distinguish the curves?

6.6 Appendix
6.6.1 Generation of random numbers
The popular approach of generating random numbers is by the so-called
multiplicative congruential method. It is implemented in the MATLAB
command rand (' seed' , j) where j is an integer as discussed before
and is called the MATLAB 4 generator. Another approach of gener-
ating random numbers is based on scaled Fibonacci numbers with a
shift (see Chapter 3). This is implemented in the MATLAB command
rand('state', j) where j is an integer and is called the MATLAB 5
generator. For example with the default j = 0,
>> rand('state', 0); rand '/,I, gives rand = 0.9501
>> rand('seed', 0); rand %% gives rand = 0.2190

We illustrate here how the congruential method works. It uses

ui+i = Kui(Mod M), i = 1, 2,3,...,
where `Mod M means that we subtract as many multiples of M as pos-
sible, and take the remainder as our answer. For any i, ui is a random
number such that 0 < ui < M. Therefore ui corresponds to a random
number ri = ui/M E [0, 1). To use this method, once suitable constants
K, M (usually integers) are chosen, we only need to supply an initial
value ul which is usually called a `seed'! Usually K is of about the same
magnitude as the square root of M but M, although large, must not be
a multiple of K.
With MATLAB, the method for one step is simply implemented as
>> unew = rem ( K * uold, M ), rand_num = unew / M

Taking ul = seed which is supplied by the user, we can represent n

steps of a congruential method as follows
>> for j = 1:n
>> seed = rem( K*seed + shift, M), rand_num = seed / M
>> end

where shift is zero in the standard method but can be nonzero!

For the benefit of interested readers, we have developed an M-file
randme. m for implementing this method. Type randme to see what it
 6.6 Appendix 85

does! For example we have used a=randme (4) with seed=2 and shift=0,
taking K = 32 and M = 5, to produce the following `home-made' four
random numbers: rand = 0.8 , 0.6, 0.2, 0.4.
To find out what the current seed is, type s = rand('seed ' ), while
rand('seed ' ,45) is to change the seed to 45. Therefore, to observe a
sequence of seed changes in MATLAB type

>> rand('seed',O); % Fix the seed to a known value

>> s = rand('seed') % Find out the default seed!
>> rand('seed',1); % Fix the seed to a known value
>> s = rand('seed'); % Check that it is indeed there
>> fprintf('MATLAB Seed = %12.Of (initially set)\n',s)
>> for j = 1:5, r=rand; s=rand('seed');
>> fprintf('MATLAB Seed = %12.Of rand = %d',s,r);end
We believe that MATLAB 4 generator uses K = 75 and M = 231 - 1
but, experimenting with randme. m and these two parameters, one may
conclude that there must be an unpublished integer shift involved! For
our purpose, we only need to know the basic algorithm
seed = (Kseed + shift) (Mod M).

6.6.2 Distribution functions

The distribution of the pseudorandom numbers is continuous (up to ma-
chine precision) and uniform (equally likely in the range (0, 1)). For any
distribution of random numbers, the probability of sampling a particular
value is governed by the probability density function (pdf) f so that if
we denote by X the random variable, then for any values x1 and x2
such that x1 < x2, it follows that
x2
P(xl <
-X <- x2) = f (t)dt.
Jxl
Alternatively a distribution can be defined in terms of the cumulative
distribution function (cdf) F, where
F(x) = P(X < x);
the relation between the two functions is

F(x) = x f (t)dt with L:f(t)dt=1 .

J 00
86 Probability and Random Numbers
Uniform distributions For a uniform distribution on the interval [0, 1],
known as the normalised uniform distribution U(0,1), the pdf is
(1, X E [0, 1],
f(x) =
1 0, otherwise,
and the cdf
0, x < 0,
F(x) = x, x E [0, 1),
1, X>1.
Normal distributions The distribution of a normal random variable
X is governed by the pdf

exp (x
f (x) _a 2a2)
or equivalently the cdf
/ ( )2 l
J- exp { (t } dt.
F(x) v 1 2- l
2Q2 JJJ

This distribution has the mean p and standard deviation a, denoted by

N(, a2).
When p = 0 and a = 1, N(0,1) is the so-called standard normal
distribution which is simulated by randn in MATLAB. The distribution
N(, a2) can be obtained by the transformation z = ax + p, if x is the
standard normal variable N(0,1).
Negative exponential distributions A negative exponential distribu-
tion describes a Poisson process. Recall that a series of events (denoted
by random variable X) is a Poisson process if
the number of outcomes in each period is independent;
the probability of more than one event in such a small interval is
negligible;
the probability of a single event during a (very short) interval is pro-
portional to the length of the interval, that is,

lo P {an event occurs in the time interval(t, t + h)} - A ,

where P denotes the probability and A is fixed.

The distribution is governed by the pdf
Ae-az z > 0,
P z) _ 0, otherwise,
 6.6 Appendix 87

or equivalently the cdf

e-atdt = 1 - e-az z>0,
F(z) = f
0, otherwise.
As the cdf is in a simple and closed form, we can link it with the uni-
form distribution (another simple cdf) and thus do not need a separate
random number generator. If x is selected from a uniform random dis-
tribution, to generate an exponential distribution z, let
1-e-Az=x

giving (P = A-')
z = -flog(1 - x).
This is actually what M-file exprand.m does. For the bank example in
6.5, where there are 131 customers (arrivals) in 180 minutes, the average
inter-arrival time is = ho = 13 minutes. Therefore, the probability of
an arrival (event) in any short interval (t, t + ho) is P = 1 so

lim
P{(t,t+h)} N P{(t,t+ho)} _ 1/ho = 131/180-A.
n-0 h ho
Of course, = 1/A as expected.
 7
Differential and Difference Equations

In this chapter we show how MATLAB can be used to explore, and to

solve numerically, various simple types of differential equation.

7.1 Ordinary differential equations (ODEs)

You will probably have come across the differential equation
dx = -Ax
dt
which is easily solved by writing it as
dx = -Adt
x
and integrating both sides. This gives
lnx= -At+c
that is,
xoe-at
x(t) = , where xo = x(0) . (7.2)

This describes, for example, radioactive decay (positive A) or exponential

growth (negative A).
The `dot' notation for differentiation is often used when the indepen-
dent variable is t, as in this case. Thus equation (7.1) is often written,
in compact notation as

Another interesting example, but a little harder to solve, is the logistic

equation
x=rx 1-Kx
88
 7.1 Ordinary differential equations (ODES) 89

0.9

0.8

0.7

0.6 N\\\\\\\\\\
x 0.5
\\\\ \ \ \ \ \ \
0.4
N\\ \\\\\\\\
0.3

0.2

0.1

0L_

0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2

Fig. 7.1. Grain plot for the ODE dx/dt = -x. The curve is a particular
numerical solution.

describing, for example, the rate of infection in a population. The solu-

tion is
K
x(t) _
1 - (1 - K0 )e-rt
MATLAB can be used to solve these and other more complicated ODEs
by a numerical integration procedure. Its graphical capabilities can
also be used to study visually the behaviour of solutions. An M-file
fodesol.m is provided to demonstrate this. As usual, typing
>> help fodesol
gives some information on how it is used.
The first order differential equation dx/dt = f (x, t) can be thought of
as defining a `grain' on the t, x plane like the grain of a piece of wood, or
the magnetic lines due to a bar magnet. See Figure 7.1 for an example.
The direction of the grain at the point (t, x) is defined by f (x, t),
which, being equal to dx/dt, gives the tangent of the angle the grain
90 Differential and Difference Equations
makes with the t-axis, that is, the slope. The differential equation tells
us that a solution passing through the point (t, x) travels in the direction
defined by f (x, t). The solution curves follow the direction of the grain.
Solving the differential equation is the process of drawing lines following
this grain at each point on their path. Each choice of initial condition
(x = xo when t = to) gives a different path (that is, solution). You can
explore the above examples (7.1) and (7.3) using fodesol.
To begin type
>> fodesol('fnxt',0,2.5,-3,3)
where fnxt . m is a simple example of an M-file containing the right-hand
side of the ODE to be studied (already provided for you). In this case it
supplies f (x, t) = -x, that is, the right-hand side required to study (7.1)
when A = 1. The four numbers supplied are, respectively, train, tmax,
Xmin and xmax, the limits of the plotting region. fodesol first displays
the relevant grain plot, or `slope field', from which you can visualise how
the solutions will behave. It then draws a solution curve starting from
a point (to, xo) which you must specify. It will prompt you for these in
turn. If you want another curve through a different point, respond with
y and give it another starting point.
When you have got the hang of this, create another version of fnxt . m,
say myfnxt . in, which contains the right-hand side of the logistic equation
(7.3) where, say,

K=r=1 .

Of course you can use fodesol to study any first order equation irre-
spective of the original choice of variable name used for the independent
and dependent variables (here t and x respectively).
Explore solutions x(t) which have xo = x(to) < 1 and also those with
xo > 1.
Another interesting ODE to look at is

dx
=x -x'
dt t

for example, for

-3<t<4, -3<x<4 .

You should create another version of fnxt . m and look at its solutions.
 7.2 Systems of differential equations 91

7.2 Systems of differential equations

MATLAB can also numerically solve systems of coupled first order dif-
ferential equations. For example, the equations
dx/dt = f (x, y), dy/dt = g(x, y) (7.6)

are first order and coupled since the derivative of x, say, depends on
y as well as on x. They are also a bit special since t is not explicitly
involved on the right-hand side. They are called autonomous differen-
tial equations for this reason. Equations like these are widely used to
describe the interaction of competing species as shown in the example
below equations (7.7). First some theory.

7.2.1 The phase plane

Equations (7.6) describe the motion of a point with coordinates (x, y)
in the x, y plane, called here the phase plane. See Figure 7.2 for an
example. As time changes, the point (x, y) will move unless of course
both f (x, y) = 0 and g(x, y) = 0. If this is the case then the point (x, y)
will remain stationary. Such stationary points are called fixed points
of the system of differential equations. Thus the fixed points can be
found from solving f (x, y) = 0 and g = 0. If this is not easy to do,
experimenting with different starting values for x and y may indicate
where possible fixed points may lie.
Near a fixed point there are a number of possibilities:
(i) The path followed by the solution may always move closer to the
fixed point. The point is then said to be stable.
(ii) The path may approach the point and then move away. The point
is then called a saddle point and is unstable.
(iii) The path may always move away from the point. Such points are
also called unstable.
(iv) The path may just move around the point in some sort of orbit.
You should now see if you can identify these behaviours in the following
example.

7.2.2 Competing species

The M-files species.m models the behaviour of two different creatures
which compete for the same food source. This is modelled by the equa-
92 Differential and Difference Equations

Fig. 7.2. The phase plane for the system of equations 7.7 for some choice of
parameters. The curves are particular numerical solutions.

tions
x(a - bx - cy), y = y(d - ex - fy) (7.7)

where x and y are the numbers of the two different species.

The parameters a, b, c and d determine the behaviour of the system.
For some values you will find that one species always wipes out the other.
For different values the species can coexist, while for a further set the
winning species depends on the initial populations, and so on.
The four types of fixed point behaviour described above are illustrated
by the sets of parameters given below. Investigate each of these cases.
To begin just type species.
(i) a=24 b=6 c=8 d=24 e=8 f =6
(ii) a=24 b=4 c=6 d=24 e=6 f =8
 7.2 Systems of differential equations 93

(iii) a=6 b=0 c=1 d=-4 e=-1 f =0

See if you can identify an example of each type of fixed point. An im-
portant first step is to identify the fixed points that is, where, according
to equation (7.7),
x=y=0.
This means you need to solve
x(a - bx - cy) = 0,
y(d - ex - fy) = 0

for the fixed points (x, y) in terms of a, b, c, . . . , f . You then just need
to substitute values for a, b etc to get the fixed points (at most 4 of
these) for each case. This helps you start your exploration. In each
case, how many different solutions (fixed points) did you find which had
x > 0 and y > 0? You are, of course, only interested in solutions with
non-negative numbers of each species! The differential equations may
well have solutions which are not of relevance to the physical application
at hand.
Use species to explore the neighbourhood of each of the fixed points
noting the behaviour close to them. The stable points are easy to find
but the unstable points are more difficult. Notice that as you get nearer
to a fixed point the motion of the point slows down and when you go
further away it speeds up. As well as a graph, the M-file displays the
coordinates at the end of each time-step.

7.2.3 Higher order equations

High order equations can be written as a system of first order ODEs, by
a simple trick using derivatives. Then the MATLAB/numerical ODE
solver can be used to give numerical solutions. Consider the following
second order differential equation in t

d 2 - k(1 - w2) dt + w = 0, (7.8)

called the Van der Pol equation (k is some positive constant). It de-
scribes oscillations in some electrical component. It can be converted
into an autonomous system with variables v and w by setting v = dwldt.
Thus you get the pair of first order equations
dw
=
dt v'
94 Differential and Difference Equations
dv
dt = k(1 - w2)v - W. (7.9)

The M-file vderpol.m illustrates the solution of this equation. Run

it for different values of the parameter k (just type vderpol and follow
the instructions). For very small k, say about 0.1, you will find a limit
cycle which is almost circular. In a limit cycle: the solution path settles
down into a closed loop of some shape, irrespective of its initial starting
point. Note that this is not the same as the orbit behaviour described in
7.2.1, where the solution curve always closes on itself. For larger values
of k the shape of the limit cycle is very different.
Have a look at the M-files vderpol.m and vdplfn.m. The latter pro-
vides the right-hand side of the coupled equations (7.9). You can adapt
these to solve your own higher order differential equations as needed, for
example the equation of a damped harmonic oscillator or a projectile
in a resistive medium. One of the projects in Part three of the book
(Chapter 22) involves doing this amongst other things. When making
modifications of the files vderpol.m and vdplfn.m for solving your own
systems of equations make your own copies and rename them: for ex-
ample mysolver.m and myslvfn.m. This should help to avoid confusion
with the original M-files and to ensure that you are solving the equation
you mean to solve!

7.3 Difference equations

The radioactive decay equation (7.1) expresses the fact that, in unit
time, a constant fraction A of the number of atoms present (x) is likely
to decay. Calculus allows us to make this statement precise over an
arbitrarily small time interval. In many applications, it is convenient to
stick to a particular time interval and discuss differences x(t + 1) - x(t)
rather than derivatives which describe instantaneous rates of change.
In the case of radioactive decay we could write an (almost) equivalent
difference equation
xn+1 - xn = -AX". (7.10)

Here, n is a positive integer labelling the time measured in units of, say,
1 second, that is,
x(t) = xn, t= n x 1s.
It is not hard to verify that
xn = xo(l - ))n (7.11)
 7.3 Difference equations 95

is a solution of the difference equation (7.10) since the latter can be

written as
xn+1
xn
--1-A.
For small values of A you can also check that this solution is equivalent
to the differential equation solution (7.2) since, in this case,
e-naN1-nAti(1-A)n.
Difference equations may, of course, be studied as mathematical tools
in their own right. In many applications, they arise more naturally out
of the problem than do differential equations. In situations involving
discrete variables, for example, one may wish to solve the difference
equation directly rather than make some continuous approximation.
The classification and nomenclature of difference equations are simi-
lar to those used for differential equations. In particular, the treatment
of inhomogeneous difference equations is the same. One adds a partic-
ular solution ('particular integral') to the solution of the homogeneous
equation.
Here is a typical example of a homogeneous, linear difference equation
that can be solved easily by a trial method similar to that used to solve
the corresponding constant coefficient ODE:
xn+2 - 5xn+1 + 6xn = 0 . (7.12)

This is a second order equation since two differences Xn+2 - xn+1 and
xn+1 - xn are involved. (Try rewriting (7.12) in terms of differences.)
We look for a solution similar to (7.11)
xn = zn . (7.13)

Substituting we get
zn+2 - 5Zn+l + 6Zn = 0

or

zn (z2 - 5z + 6) = 0.
Ignoring the trivial solution z = 0, we solve the quadratic factor to find
z=2 or z=3.
So we get two solutions of the form (7.13) which must be added in the
usual way to give
xn = A2n + B3n
96 Differential and Difference Equations
as the general solution. Notice there are two arbitrary constants since
this is a second order difference equation. Just as for differential equa-
tions, analytic solutions like this one are not always easy to obtain.
However, it is usually simple to get the computer (for example using
MATLAB) to perform the iterations implied by such equations directly.
The M-file diffegn.m is set up to solve the above example in this way.
To begin type
>> diffeqn
and follow the instructions. You should look inside this M-file and
dfeqfn.m so you can understand how to use or modify them to solve
other similar homogeneous equations, or indeed general inhomogeneous
equations.
Try various different initial conditions including [xo, xl] = [1.0, 2.0],
[1.0,1.8] etc.

Exercises

7.1 Use fodesol.m, and a modified version of myfunxt.m, to study

solutions of the following equation
dx
dt = x(1 - x2) . (7.14)

In particular:
(a) Find y(2) given the initial condition y(O) = 1.5.
(b) Find y(3) given the initial condition y(O) = 0.055.
(c) Sketch the general behaviour of all the different types
of solutions which you expect, having studied the `slope
field' or `grain plot'.
7.2 From the competing species examples in 7.2.2, identify an ex-
ample of each of
(a) fixed point surrounded by an orbit;
(b) stable fixed point;
(c) saddle point.
In each case make a rough sketch of the behaviour of the trajec-
tories (solutions) in the neighbourhood of the fixed point, that
is, showing the direction of flow of the solutions.
 Exercises 97

7.3 Modify the M-files vderpol.m and vdplfn.m to study the (non-
autonomous) second order differential equation
dew 2
dt2 - a(w - t) = 0 ,

where a is some parameter. (This is 'non-autonomous' since it

has a term which is an explicit function of t, but it can be solved
in the same way.) For the case a = 3, find the value of w(4), and
give a rough sketch of w(t), for each of the solutions satisfying:
(a) w(0) = 0, w(0) = 1;
(b) w(0) _ -1, tb(0) = 1;
(c) w(0) = 0, rb(0) = -1.

Suggest initial conditions which give a solution w(t) which is

monotonically increasing.
7.4 Use diffeqn.m to solve the difference equation
6xn+2 - xn+1 + 2xr,, = n2 + 3n + 2
subject to x0 = 1 and x1 = 2 and, in particular, find x10. Verify
your solution by comparing the left- and right-hand sides of the
equation for n = 8.
 Part two
Investigations
 8
Magic Squares

Aims of the project

Magic squares have been known for centuries. This project explores
their properties from the perspective of matrix algebra, that is, using
addition and multiplication of matrices. The project is not concerned
with the number-theoretic problem of finding magic squares containing
consecutive integers. The project is self-contained, but it may be of
interest to know that several of the mathematical results come from the
article [16]. This article also contains other results on the same subject.

Mathematical ideas used

Matrix multiplication, row reduced echelon form and solution of linear
equations are used. Also, for example, 3 x 3 matrices are regarded as
lying in nine-dimensional space R9, and subspaces of R9 are considered.
(There is no requirement to know the definition of an abstract vector
space: all spaces are contained in some Rn.) The ideas of linear inde-
pendence and basis are used. It is necessary to know that, in a subspace
X of dimension r in Rn, a set of r vectors in X which is linearly inde-
pendent automatically spans X and so forms a basis. It is necessary to
know the definitions of eigenvalues and eigenvectors, and to use these in
a simple argument involving powers of matrices.

MATLAB techniques used

The project is about matrices, so you will need the techniques described
in Chapter 2. At one point there is an M-file with several `for' loops,
and `if' statements, so you will need to understand these ideas. See
Chapter 3. Note that the project is somewhat `open-ended': students
who work quickly might like to go on to the final section, on 5 x 5 magic
squares, which could be regarded as optional.

101
102 Magic Squares
8.1 Introduction
This project is about magic squares. An n x n magic square is an n x n
matrix of real numbers with the following property:

All rows, all columns and the two `main' diagonals of the matrix add up to the
same number, r say, called the magic constant.

For example,
3 1 2
1 2 3
2 3 1

has this property for r = 6. The two main diagonals in this case are
3 + 2 + 1 (top left to bottom right) and 2 + 2 + 2 (top right to bottom
left).
There are many algorithms for producing magic squares with the ad-
ditional property that the entries are the integers 1, 2, ... , n2 in some
order. For example, with n = 3,

8 1 6
3 5 7
4 9 2

The MATLAB function magic does this (try typing magic (3) ). In this
project we shall not go into these algorithms but instead investigate the
algebra underlying magic squares, using your knowledge of matrices and
solution of linear equations.

8.2 Magic squares size 3 x 3

(i) Consider a general 3 x 3 matrix

all a12 a13

A = (aij) = a21 a22 a23
a31 a32 a33

Thus there are nine entries all, ... , a33. Explain why, writing down the
conditions for all rows, all columns and the two main diagonals to add
to the same number r, we get the condition My = 0, where M is the
 8.2 Magic squares size 3 x 3 103

8 x 10 matrix
-1 1 1 1 0 0 0 0 0 0
-1 0 0 0 1 1 1 0 0 0
-1 0 0 0 0 0 0 1 1 1

-1 1 0 0 1 0 0 1 0 0
-1 0 1 0 0 1 0 0 1 0
-1 0 0 1 0 0 1 0 0 1

-1 1 0 0 0 1 0 0 0 1

-1 0 0 1 0 1 0 1 0 0

and v is the column vector

)T
(r all a12 a13 a21 a22 a23 a31 a32 a33

(This really is just a matter of blindly applying the definition!)

(ii) Use the MATLAB function rref to find the reduced row echelon
form of the matrix M. Why does it follow that, if A is to be magic, then
three of the entries in the matrix A, say those in the bottom row, can be
chosen arbitrarily and the rest are then determined? (Note that, after
row reduction of Al, the top row of the reduced matrix can be ignored
since it just gives the equation
r = sum of elements in bottom row of A,
and this is now the only equation involving r.)
Another way of saying this is that, writing the rows of a 3 x 3 matrix
in succession to give a vector
(all a12 a13 a21 a22 a23 a31 a32 a33) E R9,

the subspace of magic squares has dimension 3.

(iii) Verify that

0 1 -1 1 -1 0
El = -1 0 1 , E2 = -1 0 1
1 -1 0 0 1 -1
1 1 1
E3 = 1 1 1
1 1 1

are all magic squares. Why does it follow from (ii) that every 3 x 3 magic
square A can be written uniquely in the form of a linear combination
A =,\1E1 + \2E2 +,\3E3
104 Magic Squares
for Al, A2, A3 E R? (Hint: You are being asked to show that El, E2, E3
form a basis for the three-dimensional subspace of magic squares. Since
there are three matrices Ei it is enough to show that they are linearly
independent, considered as vectors in R9.) What are the Ai when A is
given by (8.1) above?
(iv) A 3 x 3 matrix has, in addition to its two `main' diagonals, four
other `broken' diagonals:

all + a23 + a32, a12 + a21 + a33, a13 + a21 + a32, a12 + a23 + a31.

If, for a magic square A, we require in addition all the broken diago-
nals to add up to the same magic constant r, then the square is called
pandiagonal.
In general, an n x n matrix has two main diagonals and 2n - 2 broken
diagonals, and if all these and all the rows and all the columns add up
to the same r, then the matrix is a pandiagonal magic square.
In the 3 x 3 case, expand your matrix M by the addition of four rows
corresponding to the four broken diagonals and use the new matrix (M1
say) to show that the only pandiagonal magic squares are AE3 for A E R.
(Thus you should find from the row reduced echelon form of M1 that
only one entry of A is now arbitrary, and the others are all equal to this
one.)
(v) We now look at products of 3 x 3 magic squares. By (iii) the product
of two such squares has the form
(A1E1 + A2E2 + A3E3)(liiEl + 2E2 + /-i3E3)
Let
0 0 1
P= 0 1 0
1 0 0

Verify that all products EiEj (including i = j) can be written as linear

combinations of I, P, E3 (I being the 3 x 3 identity matrix), and hence
that (8.2) is such a linear combination.
Deduce that the product of an even number of 3 x 3 magic squares is
a linear combination of I, P, E3. Finally deduce that the product of an
odd number of 3 x 3 magic squares is also magic. (For the `even' case,
using the result just proved, you need to show that multiplying together
linear combinations of I, P, E3 gives again a linear combination of these
three matrices. This can be expressed by saying `the subspace of R9
 8.3 Magic squares size 4 x 4 105

spanned by I, P, E3 is closed under matrix multiplication'. Note this is

not true of the subspace of magic squares itself. For the product of an
odd number of magic squares you now know that this can be written as
the product of a linear combination of I, P, E3 and a linear combination
of El, E2, E3.)
(vi) Explain why the condition that all the rows of a 3 x 3 matrix A add
up to r is equivalent to the condition that (1, 1, 1) is an eigenvector of
A with corresponding eigenvalue r. Explain also why this last condition
implies that, for any integer k > 0, (1, 1, 1) is an eigenvector of Ak with
eigenvalue rk. (This is a standard property of eigenvectors and follows
from the definition.) Now deduce from (v) that if A is a 3 x 3 magic
square with magic constant r, and k is odd, then Ak is a magic square
with magic constant rk.

8.3 Magic squares size 4 x 4

(i) Set up a matrix corresponding to M above in the 4 x 4 case. It will be
a 10 x 17 matrix. Use MATLAB to show that the rank is 9 and deduce
that in R16 the 4 x 4 magic squares form a subspace of dimension 8.
(ii) Let p = [p(l) p(2) p(3) p(4)] be a permutation of 1, 2, 3, 4, i.e. the
numbers 1, 2, 3, 4 in some order. The permutation matrix A correspond-
ing to p is the 4 x 4 matrix which is all zeros except for A(i,p(i)) = 1
for i = 1, 2, 3, 4. So, for example, if p = [3 1 2 4] then
0 0 1 0
_ 1 0 0 0
4 0 1 0 0
0 0 0 1

Note that this is in fact magic.

There are seven other permutations of 1, 2, 3, 4 which give magic
squares in this way. Find them. As a hint, hcrc is one fairly brute force
way to enumerate the permutations of 1, 2, 3, 4 in MATLAB.
106 Magic Squares
for a=1:4
for b=1:4
if b"=a
for c=1:4
if c""=a & c"=b
for d=1:4
if d"'=a & d"=b & d"=c
p= [a b c d] ;

end;
end;
end;
end;
end;
end;
end;

After the permutation p is found, you will want to calculate the permu-
tation matrix, then test whether it is magic. The quickest way to do this
is to use the criterion My = 0 as in Question (i) in 8.2. Now, however,
M will be the 10 x 17 matrix you have just found, and v will be a 17 x 1
column vector of the form (-1 all ... a44)T .
(iii) Add more rows to your matrix M to test for pandiagonal 4 x 4 magic
squares (see Question (iv) in 8.2 for the definition). Call the resulting
matrix M1. What is the dimension of the space now? Are any of these
given by permutation matrices?
(iv) By finding the reduced row echelon form of your matrix Ml, show
that every 4 x 4 pandiagonal magic square has the form
a-b-c+d+e -a+b+c+d+e a+b-c-d+e -a+b+3c+d-e
b+c-d+e b+c+d-e -b+c+d+e b-c+d+e
-a+2c+2d a-2c+2e -a+2b+2c a
2b 2c 2d 2e

where a, b, c, d, e are arbitrary real numbers. Write an M-file to generate

such matrices and use MATLAB to find the eigenvalues of several exam-
ples. Do you have any conjectures about the general result here? Can
you explain one of the eigenvalues in the same way as Question (vi) in
8.2? (One relation between eigenvalues follows from the fact that the
sum of the eigenvalues of any square matrix is equal to the trace of the
 8.4 Magic squares size 5 x 5 (optional) 107

matrix, that is, the sum of the entries in the leading diagonal-top left
to bottom right.)

8.4 Magic squares size 5 x 5 (optional)

Find out what you can about the dimension of the space of 5 x 5 magic
squares, and of the pandiagonal ones. Are the pandiagonal ones spanned
in this case by permutation matrices? If so, find a basis consisting of
permutation matrices. (You can find details in the article [16].)
 9
GCDs, Pseudoprimes and Miller's Test

This chapter contains two investigations of a number-theoretic nature,

building on the ideas of Chapter 3. Investigation A is on greatest com-
mon divisors (gcds), and B is on pseudoprimes and Miller's test for
primality.
A GCDs of random pairs and triples of numbers
Aims of the project
The idea is to find, by theoretical and experimental means, estimates
for the probability that a pair of randomly chosen positive integers have
no common factor. This is also extended to triples of numbers.
Mathematical ideas used
Elementary ideas of probability are used (e.g., probability of two inde-
pendent events is the product of their individual probabilities). There
are several mathematical arguments given which need to be `filled in'
with some details-in fact, this project is more mathematical than com-
putational. The idea is to use mathematical arguments and experiment
to determine certain probabilities, such as: given three random integers,
what is the probability that each of the three resulting pairs of integers
are coprime-that is, have gcd 1 ?
MATLAB techniques used
The project uses the M-file gcdiv.m from Chapter 3, which calculates
the gcd of two integers. It also requires the writing of some M-files which
use loops and conditionals.
The M-file gcdran.m takes n pairs of `random' numbers, each < 1000,
and finds the gcd of each pair, then prints out the percentage of pairs
with gcd equal to 1.

108
 GCDs, Pseudoprimes and Miller's Test 109

(i) Try running the M-file several times for say n = 500 and find the
average percentage with gcd equal to 1.
(ii) Adapt the M-file gcdran to find the percentage which have gcd 2,
and also the percentage with gcd 1 or 2 or 3. Thus for the latter case,
for example (4, 10), (6, 9), (12, 21) and (6, 35) count towards those with
gcd 1 or 2 or 3, whereas (5, 10), (7, 35), (30, 50) do not.
(iii) By a bar-chart, histogram or other means, display (for a single run
of say 1000 random pairs) the numbers of pairs which have gcds from 1
to 20. Hint: The best thing is to use a vector, say v, of length 20, whose
kth entry records the number of pairs with gcd equal to k. So you start
by initialising
v=zeros(1,20)
and then, if a gcd is calculated to be k < 21 you increment
v(x) = v(x) + 1;
At the end, bar (v) will display the bar chart.
(iv) Consider the following `argument', writing it out for yourself with
any explanations which you feel able to offer.
Let x be the probability that a randomly chosen pair of numbers has
gcd equal to 1, that is, the pair is coprime. Recall that for integers a, b,
the symbol a I b means `a is a factor of b', that is, b/a is a whole number.
Now

gcd(a, b) = h should be (1/h)(1/h)x. (Informally,

the probability that h divides a random integer a is 1/h.) Since every
pair has some gcd we have
00
X _
T2
=1.

h=1

Now Eh1(1/h2) = 7r2/6. (This standard and remarkable result was

proved by L. Euler around 1750. Proofs can be found in most books on
complex analysis or Fourier series. Here, we simply take it as known.)
What does this give for x? How well does this accord with your
experimental results? According to this, roughly what shape should
your bar-chart/histogram for gcds up to 20 take?
(v) Extend the above argument (writing the new argument out for
yourself) to suggest the probability y that three randomly chosen whole
numbers a, b, c have gcd(a, b, c) = 1. Adapt the M-file above (or your
110 GCDs, Pseudoprimes and Miller's Test
gcdiv3. m from Chapter 3) to find the number of random triples with
gcd 1 and compare with the theoretical answer. If you need to know
an approximate value for Eh_1(1/h3), then why not use MATLAB to
calculate this sum to say 50 terms?
(vi) Here is an alternative `argument' to that given in Question (iv)
above. Write it out, filling in what details you can. Let p be a prime.
The probability that a given integer a is divisible by p is 1/p. So given
two integers a, b the probability that they are not both divisible by p is
1- - . Assuming these events, for different primes p, are all independent,
the probability that there is no prime dividing both a and b is the product

ri (1- 1p2
P
,

taken over all primes p = 2,3,5,7, 11 ..... This is the probability that a
and b are coprime.
An M-file called primes. m is available to you which produces a vector
p containing the primes < 5000. (So p(l) = 2, p(2) = 3, etc.) Run this
as usual by typing primes . How many such primes are there? Find the
above product over the primes which are < 5000. (This will be a good
approximation to the product over all primes.)
Does this agree with the probability x obtained in Question (iv)?
(vii) Take a random sample of triples of numbers a, b, c and find what
percentage turn out to be pairwise coprime. Here, a, b, c are pairwise
coprime if the gcds of a, b, of b, c and of a, c are all 1. Simply cubing the
probability x that two numbers are coprime does not (you should find)
give the probability that these three pairs are all coprime. Why do you
think this is? Warning: To express `x = y = z = 1' in MATLAB do not
write
x==y==z==1
since this statement will be `true', that is, take the value 1, if, for exam-
ple, x = y and z = 1. Instead, write
x==1 & y==1 & z==1.

(viii) Here is a sketch of an `argument' which suggests the probability of

three numbers being pairwise coprime as above. Fill in what details you
can. Let p be prime. The probability that none of the three numbers
a, b, c is divisible by p is
3
1

Cl p)
 GCDs, Pseudoprimes and Miller's Test 111

The probability that exactly one of a, b, c is divisible by p and the other

two are not is

C1-p 12.
P
Hence the probability that at most one of the three is divisible by p is
the sum of these two which after rearrangement becomes

(1_)2(1_(Pl)2).
Now a, b and c are coprime if and only if, for all primes p, at most one
of the numbers is divisible by p. Assuming these are all independent
events the probability z that the three numbers are pairwise coprime
is the product of the expressions (9.1) for all primes p. You should
recognise the squared factor in front from Question (vi) above. Work
out the value of the other factor using the same vector p of primes you
used above. Now deduce the theoretical value of z.
Does this agree with the experimental value you found?
(ix) Consider the set of all pairs of numbers a, b where a and b are > 2
and < m. For various m, what percentage of these are coprime? How
close is this to the probability of random pairs being coprime?

B Pseudoprimes and Miller's test

Aims of the project
This project is based on the idea, introduced in Chapter 3, of pseudo-
prime. Certain numbers `masquerade as primes' and we shall explore
one method commonly used to `unmask' such numbers-to show that,
in fact, they are not prime.
Mathematical ideas used
Fermat's theorem, given in 3.4, is used, and also the ability to work
out remainders rem(a', m), where n and m may be fairly large num-
bers. This is done automatically with the MATLAB function pow. A
relatively sophisticated primality test, Miller's test, is also introduced
here. This can be used to `unmask' some pseudoprimes, proving that
they are composite without actually trying to factorise them.

MATLAB techniques used

The M-files for this project have been written for you, but you need
112 GCDs, Pseudoprimes and Miller's Test
to amend one of them slightly to perform different tasks. The M-file
gcdiv. m for calculating gcds is also used.
(i) An M-file called primes. m is available to you. This takes the numbers
up to 5000 and decides for each one whether it is prime, putting the
primes in a vector p = [p(1), ... , p(k)], where k is the number of primes
involved. Thus p(l) = 2, p(2) = 3, p(3) = 5, p(4) = 7, p(5) = 11, etc.
So begin by typing primes. Once this has run, typing p will produce a
long vector containing all these primes; say p(5) will produce the fifth
prime, which is 11.
(ii) Recall from Chapter 3 that a pseudoprime to base a is a number
n which is not prime, but which nevertheless satisfies rem(an-1, n) = 1.
The significance of this is that every prime n automatically satisfies this
relationship (so long as a is not a multiple of n), by Fermat's theo-
rem (see Section 3.4). However, there is a very small minority of non-
primes (composites) which `masquerade as primes' to the extent that
they, too, satisfy this relationship. One way of checking that a num-
ber n is not prime, without actually trying to factorise it-a hopelessly
lengthy procedure for very large numbers-is to show that it fails to
satisfy rem(an-1, n) = 1, for some a which is not a multiple of n. Pseu-
doprimes resist this attempt to show that they are not prime. Note
that in this book we can only illustrate these ideas with relatively small
numbers!
The M-file psp2.m finds pseudoprimes to base 2 which are < m for
a value of m input by the user. Note that we require the even number
2n-1 to leave remainder 1 when divided by n, say 2n-1 - kn = 1. If n
were even the left hand side of this equation would be even, which is a
contradiction, so it is necessarily true that n is odd. To save some time
in execution, the M-file checks only odd numbers n, starting at n = 3.
This M-file has comments to explain how it works. Take say m = 2000
and check that each number n output is indeed composite and that
2n-1 - 1 mod n, that is, 2n-1 leaves remainder 1 when it is divided by
n. Of course you need to use
>> pow(2,n-1,n)

to do this.
Make amendments to psp2. m as necessary to find all numbers < 5000
which are:

(1) pseudoprimes to base 2,

 GCDs, Pseudoprimes and Miller's Test 113

(2) pseudoprimes to base 2 and 3,

(3) pseudoprimes to base 2, 3 and 5.
You should find that your list becomes steadily shorter: fewer and fewer
numbers can masquerade as primes when several bases are used. For
each pseudoprime, work out its factorisation into primes.
How many primes are there < 5000? (Use Question (i) above.)
(iii)
Suppose you are presented with an odd number n, where 1 < n < 5000,
and
rem(a'-1, n) = 1 for a = 2, 3 and 5.
What is the probability that n is not a prime? The moral here is that
n is `very likely' to be prime.
(iv) If you want to find the numbers n which are pseudoprimes just to
base 3, rather than simultaneously to bases 2 and 3, then you should
also consider even numbers n. The same holds for base 5. Are there in
fact any even pseudoprimes n to base 3 or base 5 which are less than
10000?

(v) Miller's test If a number n has rem(an-1, n) 1, (that is, an-1

1 mod n) for some a with n not dividing a, then we know from Fermat's
theorem that n is not prime. There is a more sophisticated method for
flushing out composite (that is, not primet) numbers, called Miller's
test, which was published in 1976. For more details of the theory of
Miller's test see, for example, [7], Ch.5.
Here is the test. Start with an odd n > 1, and a number b (the base),
which is coprime to n, that is, gcd(b, n) = 1. Then carry through the
following steps:
Step 1 Let k = n - 1, r = rem(bk, n). If r 54 1 then Miller's test to base b has
been failed, otherwise continue. (Note that if n passes Step 1, then n is either
prime or a pseudoprime to base b.)

While k is even and r = 1, repeat the following

Step 2 Replace k by k/2, then replace r by the new value of rem(bk, n).

When k becomes odd or r $ 1:

if r = 1 or r = n - 1 then n has passed Miller's test to base b;
if r 1 and r 54 n - 1 then n has failed Miller's test to base b.
t For technical reasons, the number 1 is regarded as neither prime nor composite.
This need not worry us in the slightest!
114 GCDs, Pseudoprimes and Miller's Test
The important result is that if n is prime then n always passes the
test. Thus if n fails the test then it is composite. Note that failing Step 1
amounts to failing to satisfy Fermat's theorem.
The M-file miller. m implements this test. See if you can understand
how it does this. Use the M-file to do the following:
(vi) Test miller.m by using some primes, which must pass the test.
Write out the residues obtained. You can get plenty of primes to try
from the result of Question (i) above.
(vii) Check that n = 1373653 passes the test for b = 2 and b = 3 but fails
it for b = 5; write down the results of doing the test for these b. Is this
n prime or composite? Let c = rem(5n-1, n). (Note that this cannot be
evaluated by the MATLAB function rem since the numbers are far too
big. But the M-file miller.m tells you what c is !) Find gcd(c - 1, n)
and check that this is a proper factor of n, that is, a factor which isn't
1 or n.
(viii) Check that n = 25326001 passes Miller's test to bases 2, 3 and 5
but fails it for base 7 (write out the residues occurring in the tests). Let
c = rem(71-1, n). Find gcd(c -1, n) and check that it is a proper factor
of n.
(ix) Find the first number > 5 x 107 which could possibly be prime.
Write down exactly what you do to find this number. (So the fewer the
times you need to use Miller's test the better. For example, do not test
50000002 for primality by Miller's test!)
(x) Which of the pseudoprimes found above (Question (ii)) are un-
masked as composites by using Miller's test to base 2?
(xi) There is a composite number between 2000 and 3000 which passes
Miller's test to base 2. Find it. Does it pass Miller's test to base 3?
(xii) A tiny bit of theory. The notation x - y mod n ('x is congruent to
y mod n') means that n I (x - y) (that is, n divides exactly into x - y).
Thus, if r = rem(a, n), then a - r mod n. You may assume the standard
properties (which are easy to prove): If x - y and u - v mod n then
x u - y v and xu - yv mod n. In particular, x2 - y2 mod n.
Now suppose that n is odd and 2'+1 (n - 1), where r is an integer
1

> 0, and that gcd(b, n) = 1. Suppose that

bin-' '2' _ 1 mod n, (9.2)

c = b(n-1)/2-+' # 1 mod n. (9.3)

 GCDs, Pseudoprimes and Miller's Test 115

Use (9.2) and successive squaring to show that n is a pseudoprime to

base b. Show also that n fails Miller's test to base b, and that n I (c2 -1),
using (9.3).
Let h = gcd(c - 1, n). The claim is that h 0 1 and h n, that is,
that h is a proper factor of n. Show this as follows. (a) If h = 1, deduce
n I (c+ 1). (You may use the standard result x I yz, gcd(x, y) = 1, = x
z.) Why is this a contradiction? (b) If k = n then deduce n I (c - 1).
Why is this a contradiction?
Check that 10004681 is a pseudoprime to base 2 which fails Miller's
test to base 2. Factorise it using the above method. (Ideally you should
end up with the prime factors of 10004681. Remember that the vector
p of primes is available to you.)
For the pseudoprimes to base 2 which you found in (a) above, check
which ones can be factorised by the above method. Write down the
steps you go through in testing the numbers, and list those for which
the method does not work.
 10
Graphics: Curves and Envelopes

This chapter contains three graphics investigations. Investigation A is

about certain special cases of curves obtained by rolling one circle on
another circle; B is a further exploration of envelopes of families of lines,
introduced in Chapter 4, and C is on curves which have `constant width'
in the sense that all pairs of parallel tangents are the same distance apart.

A Rose curves and epicycloids

Aims of the project
Rose curves are curves with 'petals'-they don't particularly resem-
ble roses!-and we shall explore these theoretically and experimentally.
(Some more ideas for project work can be found in [8].) Then we shall
take a certain envelope defined by chords of a circle and relate it to a
curve obtained by rolling one circle on another.
Mathematical ideas used
The project is about parametrised plane curves, including polar coordi-
nates, and envelopes of lines (4.6). Trigonometrical formulae come into
some of the calculations.
MATLAB techniques used
You will need to use some M-files from Chapter 4 to draw curves and
envelopes of lines.
Rose curves These are the special case of hypocycloids (see Exer-
cise 4.5), where a = 2m/(m + 1), b = 1 and d = a - 1. Here, m is an
integer > 0. Figure 10.1 shows the example m = 4.
Make an investigation of these curves. Use the M-file hypocy. m . In
particular investigate experimentally the following numbers, expressed
in terms of m: (i) the smallest value of t which makes the curve close

116
 Graphics: Curves and Envelopes 117

0.5

-0.5

-0.5 0.5

Fig. 10.1. A rose curve with m = 4 and eight `petals'.

up and start to repeat itself; (ii) the number of times the curve passes
through the origin before it closes up and starts to repeat itself; (iii) the
number of `petals' the rose has.
Then do a mathematical investigation as follows. Put t = (m - 1)0
in the parametrisation of the hypocycloid given in equation (4.2). Then
use the famous formulae

cos a+cos,6 = 2cos (+/3\ ) cos I a

2 0) ,

sin a - sin,3 = 2 cos a '6) sin l a

2 2 )
to show that the rose curve can be written as r = 2 n+i cos(m0) in polar
coordinates r, 0. (This just means that x = r cos 0, y = r sin 0.) You can
assume that a polar curve r = f (0), where f is some function, will close
up and start to repeat (a) when 0 = it if f (0 + 7r) = -f (0) for all 0, and
(b) when 0 = 27r if f (0 + 7r) -f (0) but f (0 + 27) = f (0) for all 0. Use
this to find the answers to (i), (ii), (iii) above, and compare with your
experimental answers.
Epicycloids
(i) Show that the equation of the chord of the circle x2 + y2 = 1 joining
(cost, sin t) to (cos(mt), sin(mt)) is
x(sint - sinmt) - y(cost - cosmt) + sin(m - 1)t = 0. (10.1)

(You will need to remember the trigonometrical formula for sin(a - b).)
118 Graphics: Curves and Envelopes

-2 6

Fig. 10.2. An epicycloid obtained by rolling a circle radius 1 outside a circle

radius 5.

Use linenv.m to draw the envelope of these chords for some values of
m, say 2, 3 and 4. Print one of these out.
(ii) Verify that the point
in cost + cos(mt) _ m sin t + sin(mt)
X = y(t)
m+1 m+1
satisfies both equation (10.1) and the equation obtained by differenti-
ating equation (10.1) with respect to t. This is a special case of the
situation of Exercise 4.7, and shows that the above point Wt), y(t))
gives a parametrisation of the envelope of the chords. That is, the curve
which your eye picks out from the drawing of all the chords, where the
lines appear to cluster, is parametrised as above.
(iii) An epicycloid is similar to a hypocycloid (Exercise 4.5) but the
rolling circle rolls outside the fixed one. The parametrisation is (using
u for a reason which will become clear)

x=(a+b)cosCa+b)+dcosu, y=(a+b)sin( a+b)+dsinu.

Figure 10.2 shows` the example a = 5, b = 1, d = 1, with the fixed circle

also drawn and the rolling circle in its position for u = 0.
Show (mathematically!) that by putting u = mt and taking suitable
values of a, b, d in terms of m, this epicycloid can be made to coincide
with the above parametrization of the envelope of chords. You need, for
example, m/(m + 1) = a + b, 1/(m + 1) = d.
 Graphics: Curves and Envelopes 119

(iv) Amend hypocy. m so that it draws epicycloids. Note: You will need
to make one subtle amendment, namely `upper' should be redefined as
abs(a+b)+abs(d). If you name the parameter as u then you also need
to use u=tl: tstep: to instead of t=tl : tstep: tu. There is no need to
rename tl and tu. The rest of the changes are simply in the x and y
lines. Draw some of these epicycloids for small (integral) values of m
and compare with the envelope of chords.
(v) Prove the formula for the parametrisation of an epicycloid (in Ques-
tion (iii) above).

B Envelopes
Aims of the project
The idea is to look at two situations where a family. of lines creates an
envelope. One is a family of perpendicular bisectors similar to that en-
countered in Chapter 4 (Exercise 4.6), and the other comes from a sliding
ladder. The investigation is both mathematical and experimental.
Mathematical ideas used
These include parametric equations of curves, elimination of a variable
between two equations, tangents to curves and simple trigonometry. The
ideas on envelopes come from Chapter 4.
MATLAB techniques used
The first part of the investigation uses the M-file linenv. m to draw en-
velopes of perpendicular bisectors. You also need to draw an additional
curve on the same figure. The second part involves the solution of poly-
nomial equations, drawing parametric curves and drawing envelopes of
lines.
Perpendicular bisectors
Investigate the envelope of perpendicular bisectors of lines joining the
point (a, 0) (where you can assume a > 0) to the points of the unit circle
x2 + y2 = 1, which is parametrised by (cost, sin t). Thus to begin with,
show that the equation of the perpendicular bisector is
2x(a - cost) - 2y sin t + 1 - a2 = 0 (10.2)

(see Exercise 4.6). Then use linenv.m to draw the envelope for various
values of a. The M-file should have a couple of lines added to it so that
it draws this circle as well, in red. Print out one example for a > 1 and
one for a<1.
120 Graphics: Curves and Envelopes
The rest of this part is a mathematical investigation to explain the
pictures just obtained. For the perpendicular bisectors given by equation
(10.2) use the method of Exercise 4.7, to show that the equation of the
envelope of perpendicular bisectors is, for a 1,
z
4x2 - 4ax + 14ya2 = 1-a2.
So you have to eliminate t between two equations, namely equation
(10.2) and the derivative of this equation with respect to t. One way
to eliminate t is to arrange the two equations as equations for sin t and
cost, solve for these two, and then use sin2 t + cost t = 1.
What is the envelope if a = 1? Is there a simple geometrical explana-
tion for this?
Assume a # 1. Make the substitution X = x - 2 to reduce the
equation of the envelope to the form
2

4X2 + 4y = 1.
1-a2
What kind of curve is this? Explain with examples how it fits with the
pictures.
Can you find any significance in the particular perpendicular bisectors
which arise when the line joining (a, 0) to the circle is tangent to the
circle? Of course this requires a > 1. (Hint: Show that these correspond
with cost = 1/a and have equations X,/a2 - 1 = y.)
Sliding ladders
A ladder of length 1 rests on the ground and on a vertical wall, just also
resting on a rectangular box of sides a and b, as in Figure 10.3, left.
Writing x for the length shown, show that f (x) = 0, where
P X) = x4 - 2ax3 + (a2 +b 2 -12)x2 + 2a12x - a212.
Take 1 = 10 and a = 2. Verify that b = 5 gives a physically possible
solution (that is, the equation has a root x which is physically possible),
but b = 6 does not. Illustrate with the graph of f but also use roots
to find the roots numerically. The M-file paramc. m draws parametrised
curves (x, y) _ (x(t), y(t)). In the special case of a function graph y =
f (x) you can take x = t, y = f (t), so you have to insert these in the
correct place in paramc. m. (Do not confuse the y here with the y in
either of the diagrams! The M-file is a general purpose parametric curve
plotter, so x, y are the most natural variables to use in it.)
Take a = 2 and b = 5 and find all solutions when 1 = 20. Try also
 Graphics: Curves and Envelopes 121

Fig. 10.3. Left: a ladder of length l leaning against a vertical wall and resting
on a box of dimensions a and b. Right: finding the equation of the ladder.
Note that x, y have different meanings in the two diagrams!

a = 2, b = 5, 1 = 9.582299. What do you notice here? What happens if

1 is decreased or increased slightly?
We now take the box away and imagine the ladder sliding down the
wall, keeping in a vertical plane (a frightening possibility!) and find the
envelope of all the ladder positions.
Show that the equation of the straight line along the ladder, given the
angle t in the right-hand diagram of Figure 10.3, is
x
cost sin t
+=1.
Show from the method of Exercise 4.7 that the envelope of the ladder
lines is given by x = 1 cos' t, y = 1 sin3 t. Now eliminate t to show that
the equation of the envelope is
x2/3 + y2/3 = 12/3.

Also use 1 inenv . m to draw the envelope of the ladder lines.

Show that the mid-point of the ladder describes part of a circle as the
ladder slides. Can you include this on the picture of the envelope? Can
you see a connection with up-and-over garage doors?
Why can you deduce from the equation of the envelope that the con-
dition for a physically possible solution of the original ladder and box
problem to exist is
a2/3 + b2/3 < 12/3?

Check this with the examples above.

122 Graphics: Curves and Envelopes

Fig. 10.4. All widths, measured between parallel tangents, are equal to 2.

C Curves of constant width

Aims of the project
The circle is by no means the only curve which has `constant width' in the
sense that all pairs of parallel tangents are the same distance apart. This
project investigates properties of curves of constant width, including
how to make them, and how they are related to other geometrical ideas
such as envelopes of lines. (There are some interesting observations on
curves and surfaces-of constant width in [5].)
Mathematical ideas used
Equations of lines and tangents, linear equations and 3 x 3 determinants;
envelopes of lines (Chapter 4) come in too.
MATLAB techniques used
Drawing parametric curves and envelopes of lines, using given M-files.
This project is about curves of constant width. The width of a curve is
measured between a pair of parallel tangents and constant width means
that all such widths are equal. Surprisingly, the circle is not the only
curve with this property. See Figure 10.4. Curves of constant width
have been applied in very practical situations, for example the W inkel
car engine and bits for drilling (nearly) square holes. Coins have to be
curves of constant width to work in coin-operated machines; the British
20p and 50p coins are noncircular curves of constant width.
We shall generally only consider curves that are `convex'; this implies
 Graphics: Curves and Envelopes 123

Fig. 10.5. `Pedal coordinates' of a line.

that there are exactly two tangents parallel to any given direction. We
shall meet some more `singular' examples later.
See 4.6 and Exercise 4.7 for details of envelopes.

The `pedal' construction

Let h be a function of the angle t. In Figure 10.5, l(t) is the line through
the point (h(t) cost, h(t) sin t), perpendicular to the direction from this
point to the origin.
(i) Show that the equation of l(t) is
x cos t + y sin t = h(t). (10.3)

(ii) Let h(t) = 1+1 cos(3t). Amend the M-file linenv.m so that it draws
the envelope E of the lines l(t) for 0 < t < 27r. Print this envelope out,
taking care to adjust the limits of x and y so that the whole envelope E
is clearly visible on the picture.
(iii) The lines l (t) are of course all tangent to the envelope E drawn
in (ii). Check that h(t) + h(t + ir) = 2 for all t with the function h
given there. Why does this imply that the pairs of parallel tangents to
E are all a distance 2 apart, and hence that E is a curve of constant
width 2?
(iv) Recall from Exercise 4.7 that to find equations for the envelope we
take the equation (10.3), together with the equation
-x sin t + y cos t = h'(t) (10.4)
124 Graphics: Curves and Envelopes
obtained by differentiating the equation (10.3) with respect to t. (Here
and below,' stands for differentiation, d/dt.) Solve the equations (10.3)
and (10.4) to obtain
x (t) = h (t) cos t - h'(t) sin t y (t) = h (t) sin t + h'(t) cos t, (10.5)

which are therefore the parametric equations of E.

Amend the M-file paramc. m so that it draws the parametrised curve
given by equations (10.5), with h as in (ii), calling the amended M-file
constwl. in. Adjust the limits of x and y so that E just fits in the square
on the screen (of course, being constant width, it will just fit inside a
square!), and print out this picture. The curve E should match up with
the one you printed out in (ii), apart from possibly different scaling. Join
(by hand) the points where your curve touches the top and bottom of
the square, and the points where it touches the two sides of the square.
Do you notice anything special about these two lines?
(v) Explain why, changing h(t) (as in (ii)) by adding a constant a, an-
other curve of constant width is obtained. What is the width of this
curve? Amend your M-file constwl.m so that it draws this curve for
each of the eleven values a = 0,0.1,0.2,... ,1.0 on the same picture,
calling the resulting M-file constw2.m. Remember that using hold on
between the drawing of the various curves keeps them all on the same
picture. Use hold off at the end of the M-file. Adjust the limits of x
and y so that the largest of the eleven curves fits exactly on the square
of the screen and print out this picture.
(vi) What happens if you take a = -0.5 in (v)? Draw the resulting
`curve' on the screen and make a hand sketch of it. Do you think it has
constant width?
(vii) As already noted, the function h in (ii) satisfies
h(t) + h(t + 7r) = k, (10.6)

where k is constant (actually k = 2 in this example). Another way of

satisfying the equation (10.6) is to make h itself constant. What curve
E is produced by the equation (10.5) in that case? (You can answer that
without using the computer!) Choose a nonconstant function h which
satisfies the equation (10.6) and produces a curve E different from that
studied above. Print out the curve E as given by your h and the equation
(10.5), choosing scales so that it fits exactly on the square of the picture.
(viii) Let h be any function satisfying the equation (10.6), and let E be
 Graphics: Curves and Envelopes 125

Fig. 10.6. The line joining the two points of contact of parallel tangents is in
fact normal to the curve at both contact points.

the envelope parametrised as before by equations (10.5). Show that

x(t + ir) = x(t) - k cost, y(t + 7r) = y(t) - k sin t.
Deduce from this that the line joining the points of the envelope E at
t and t + it is always perpendicular to the lines l(t) and l(t + 7r) (see
Figure 10.6).
Completing a given 'half-curve'
Here, we start with a curve which has parallel tangents at its endpoints
but nowhere else. We shall take the example of the half-ellipse in Fig-
ure 10.7. Here, the two parallel tangents are a distance 2 apart and we
construct a curve of constant width 2. The half-ellipse is parametrised
by
7r 327r
x = a cos t, y = sin t, for <t<
2
For each tangent to the half-ellipse, say at (a cost, sin t), we take the line
m(t) parallel to this tangent, at distance 2 (see Figure 10.7).
(ix) Show that the line m(t) has equation
x cos t + ya sin t = a - 2 a2 sin2 t + cost t. (10.7)

(x) The equation (10.7) is not in quite the same form as equation (10.3),
because of the presence of a in the coefficient of y. However, the method
of solving for the points of the envelope E of the lines m(t) is still the
126 Graphics: Curves and Envelopes

Fig. 10.7. A curve of constant width based on half an ellipse.

same as before. Writing r(t) for the right-hand side of equation (10.7),
show that the points of the envelope are

x(t) = r(t) cost - r(t) sint, y(t) = 1(r(t) sint + r(t) cost). (10.8)
a
Why does the envelope E, together with the half-ellipse, make a curve
of constant width 2?
(xi) Amend your M-file constwl.m, or write another M-file, to draw the
envelope curve E given by equation (10.8) for it/2 < t < 37r/2, and also
to draw the original ellipse for the same range of values of t. Call the
result constw3.m.
(xii) Draw the following three examples of E:
a = 0.75, xl = -1, xu = 2, yl = -1.5;
a = 1.25, xl = -1.25, xu = 1.25, yl = -1.25;
a = 3, xl = -3, xu = 1, yl = -2.
As usual xl, xu, yl stand for the lower and upper limits of x and the
lower limit of y. In each case make a hand sketch of the half-ellipse and
envelope. You should find that the third curve is singular, that is, has
cusps or sharp points.
(xiii) By taking various values of a (and suitable limits for x and y)
determine as accurately as you can the range of values of a for which
the envelope E above is free from cusps. You need only consider a > 0.
(xiv) The remaining work here uses the theory of envelopes to find
 Graphics: Curves and Envelopes 127

the range of values of a > 0 for which there are no cusps on the above
envelope. Writing equation (10.7) as
px+qy=r,
it can be shown that the condition for a singular point (cusp) to exist is
that (using ' for differentiation as before)
p q r
p' q' r' = 0
p"r q// r"
has a solution t between it/2 and 3ir/2. Using p = cos t, q = a sin t
show that adding the top row to the bottom row and evaluating the
determinant by the bottom row reduces this condition to simply r+r" =
0 (remember that a is nonzero). Hence show that the condition is v = 0,
where

v = -2u+a+ 2(a2 - 1)2 u3

sin 2 tcos2 t
- 2(a2 - 1)(cos2
u
t - sine t)

and u = a2 sin2 t + cost t.

Write a (short) M-file to draw the graph of v:
t=-pi/2:0.01:3*pi/2;
v= above expression;
plot(t,v)
hold on
plot([pi/2,3*pi/2] , [0,0] )
hold off
What does the second plot statement do?
Draw a hand sketch of the resulting graph of v for the three values
a = 0.75,1.25, 3 and verify that in these cases the curve in (xi) above is
free from cusps precisely when the graph of v fails to cross the axis.
What is the range of values of a, according to this method, for which
the envelope E is free from cusps?
 11
Zigzags and Fast Curves

This chapter contains two investigations which use graphics: A, on

spirographs and zigzags, and B, on the problem of determining the shape
of a wire which gives the fastest time of descent for a bead sliding down
the wire.

A Spirographs and zigzags

Aims of the project
The idea here is to draw a zigzag line determined by a simple rule. The
interest lies in determining how many times the line must zig and zag
before it closes up, and in working out how large the resulting picture
is, so that it can be drawn fitting neatly on the screen. We shall also
investigate the connection between the zigzags and certain spirograph
curves (epicycloids). The basic idea for this project comes from the book
[1].

Mathematical ideas used

Vectors in the plane, linear equations and 2 x 2 matrices, regular poly-
gons, gcds, trigonometric formulae and ellipses all come into this project.
MATLAB techniques used
This is a project about drawing sequences of lines. You need to amend
a given program to make it do successively more tasks automatically.
One of these involves calculating gcds, and another involves setting the
screen size to fit the zigzag.
Construction of the zigzag
This project involves drawing zigzag lines according to a simple rule, and
examining the mathematics behind these constructions. There is an in-
timate connection between the zigzags and certain epicycloids, which are

128
 Zigzags and Fast Curves 129

Fig. 11.1. The basic zigzag, defined by lengths of 100 and 1, and angles 01, 02.
In the right-hand figure, pi and p2 are rotations through 01 and 02 respectively.

obtained by rolling one circle on another and drawing the path traced
out by a point rigidly attached to the rolling circle. (Compare Investi-
gation A of Chapter 10.)
The basic idea is illustrated in Figure 11.1. A straight horizontal line
is drawn to the right from the origin, of length 100. At the end of this
another straight horizontal line is drawn of length 1. If 1 < 0 then the
line is drawn to the left and otherwise to the right; in either case it
terminates at (100 + 1, 0). At this stage we say that one step has been
completed.
Now the zigzagging begins. We have two angles 01, 02 given to us
(usually they will be whole numbers of degrees). We draw a straight line
of length 100 from (100+1, 0), at an angle 01 with the positive x-axis (so
this angle is measured anticlockwise from this axis). The termination of
this line is at (100 + 1 + 100 cos 01 i 100 sin 01). From this point we draw
a line of length 1 at an angle 02 with the horizontal, thereby arriving at
the point
(100+1+ 100 cos 01 +lcos02i100sin01 +lsin02).
At this stage, two steps have been completed.
The lengths of the added lines are always alternately 100 and 1, but
the angles between the added lines and the horizontal go up by 01 and
02 respectively at every step. Thus the third step consists of two lines at
angles of 201i 202 to the horizontal, the fourth step of two lines at angles
of 301i 302 to the horizontal, etc.
130 Zigzags and Fast Curves

The mathematical problems underlying this construction are:

Can we find a reasonable `closed' formula for the point reached at the
end of k steps? ('Closed' means an explicit formula without any ...s.)
Does the zigzag close up and if so how many steps are needed?
How big is the resulting figure (we need to know this so we can draw
it without pieces whizzing off the screen).
There are other problems too which will be mentioned later: for example
it is possible to define a `spirograph' curve through the points arrived
at after 1, 2, 3, ... , k steps which turns out to be an epicycloid. This
epicycloid in some sense approximates the zigzag itself, and there are
examples where the epicycloid is a simple curve like an ellipse.
The M-file zz 1. m draws the zigzag but it requires you to give a lot of
help, by specifying the number of steps to be executed and the size of
the graphics window (xl < x < xu, yl < y < yu). In due course you will
have amended this so that the number of steps to closure is calculated
in advance, and the size of the window set automatically. The one thing
zz l . m does do for you is to make the graphics window square, that is,
xu - xl = yu - yl. In fact yu is put equal to yl + xu - xl.
As a start, try running zzl.m with the inputs 1 = 40, 01 = 45, 02 =
9, steps = 40, xl = -320, xu = 460, yl = -15. Try also 1 = 50, 01 =
175, 02 = 185, steps = 72, xl = 0, xu = 150, yl = -74.
A vector formula for the point reached after k steps
For consistency of notation we shall replace 100 by 11 and 1 by 12 in what
follows. Let v1i v2 be horizontal vectors of lengths 11, 12 respectively,
and let p1, p2 denote rotations anticlockwise through 01 i 02 respectively.
Thus for example if v is the vector (a, b) then the effect of p1 on v is to
turn it into the vector
cos 01 - sin 01 a _ a cos 01 - b sin 01
sin 01 cos 01 b a sin 01 + b cos 01

We can write pi for the rotation through k01; the matrix for this is
obtained by replacing 01 by k01. For P2, we replace 01 by 02. Using this
notation we can write the position of the point at the end of k steps of
the zigzag as (see the right-hand part of Figure 11.1):
p2_lv2
vl + V2 + plvl + p2v2 + pivl + p2v2 + ... pi-lvl +
(v1 + plvl + plv1 + ... pi-lv1)
p2-l
+ (v2 + p2v2 + P2v2 + ... V2)- (11.2)
 Zigzags and Fast Curves 131

Fig. 11.2. Part of a `regular polygon' with central angle B1 and centre cl.

The two brackets in equation (11.2) can be rewritten in a much more

convenient form. In fact the vectors occurring in the first bracket are
simply the vectors along the sides of a regular polygon with side equal to
the length 11 of v1 as shown in Figure 11.2. The point c1 is the centre of
the polygon and the distances from c1 to the various vertices (starting
at 0) are all the same. Explain the labels on the angles and sides in the
figure and the following two equations.
vl = cl - plc1, (11.3)

vl + plvl + plvl + ... + pi lvl = cl - pi c1. (11.4)

Hence the point of the zigzag reached after k steps is, using equations
(11.2), (11.4) and the corresponding result about v2i
cl - picl + c2 - p2c2. (11.5)

So to find a good expression for the point that the zigzag has reached
we just need to find cl and c2. This is easy from equation (11.3) and
its counterpart for v2, but first we look at the number of steps.
The number of steps to closure
It is not obvious that the zigzag ever closes, of course. But so long as
132 Zigzags and Fast Curves
we stick to simple angles (for example, whole numbers of degrees) it will
close, and we can calculate the number of steps needed.
From now on we assume that 91, 02 are whole numbers of degrees.
From equation (11.5) it follows that the zigzag will certainly close if
both c1 = pic1 and c2 = p2c2. (It is conceivable that the zigzag closes
before both these happen. You might like to investigate that possibility!
But we shall say no more about it here.) Now if a nonzero vector cl is
returned to its original state by k rotations pi through 01 degrees, then
it must be that k01 is a multiple of 3600. Explain why this is the same
as
360
k = multiple of gcd(360, 01)

So the zigzag will close provided k is a multiple of both

360 360
and
gcd(360,01) gcd(360,02)'

Explain why this is the same as saying that the zigzag will definitely
close when k = s, where
360
s= (11.6)
gcd(360, 01, 02)'
(Thus you need to show s is a common multiple of the two numbers
given above. In fact it is the least common multiple. .)
Use equation (11.6) to calculate the quantity `steps' that is, the num-
ber of steps to closure-in zz 1. m, using the available M-file gcdiv. m for
calculating gcds. Call the resulting M-file zz2.m. Check several exam-
ples to see that the zigzag does close. Write out the data which you
input to the M-file, and the calculated number of steps to closure.

Size of the zigzag

We want to calculate in advance the size of the zigzag so that the graphics
window can be set accordingly. In terms of equation (11.5) we shall take
c = C1 + c2 at the centre of the screen, i.e. at (2 (xl + xu), 2 (yl + yu)).
So we presumably need to find c and also to find the other terms of the
equation (11.5).
Finding the `centre' c
Show from equation (11.3), the matrix for the rotation p1 given in equa-
tion (11.1), and the fact that v1 = (l1i 0)T (as a column vector) that
-1
1 - cos01 sin 01
Cl v,
- sin e1 1- cose1
 Zigzags and Fast Curves 133

11
sin 10,
2 , cos 2 Bl
2 sin 101
2 ( T
Hint: You will need to use the 'half-angle formulae'

1 - cos91 = 2sin2 201isin91 = 2sin 201 cos 201.

Of course a similar formula holds for c2i all suffix is being replaced
by suffix 2s.
Now use the matrix for the rotation pi through k01, namely
( cos k01 - sin k01
sin k01 cos kGl J'
a similar matrix for p2, and equation (11.5) to show that the end of the
zigzag after k steps is at the point with coordinates

c+ (sin (k - 2 1 91, -cos (k - 1)01)

2 sin 191
2 /
+ 2 sin 192
(sin k- 2 02,- cos k- / 02) . (11.7)
2L

2
(This time you will need the well-known formulae for sin(a + b) and
cos(a + b).) Of course it is possible to incorporate the formulae obtained
above for c = cl + c2 but for most purposes it is better to use equation
(11.7) which gives the position relative to the `centre' c.

The size formula

We need to know how far away the point given by equation (11.7) can
get from the centre c. Placing c at the centre of the screen we can then
arrange to get all of the zigzag on the screen. Show that, removing the
initial term c from the expression (11.7), the remaining vector has length
at most
ll 12
d (11.8)
2 sin 01 + 2 sin 102
2
Write the vector c as (a, b). Thus with c at the centre of the screen we
want to make the graphics area a square of side 2d. This amounts to
xl=a-d, xu=a+d, yl=b-d, yu=b+d.
Incorporate these into zz2. m to give zz3. m: zigzag program mark 3,
which will automatically scale the picture to fit neatly into the screen.
Test with several examples, giving details of the ones you test and how
well they fit. Remember that in the original setup, 11 = 100.
134 Zigzags and Fast Curves

700

600

500

400

300

200

100

0
-300 -200 -100 0 100 200 300 400

Fig. 11.3. A zigzag (with l1 = 100, 12 = 40, 01 = 45, 02 = 9), and the
associated spirograph which passes through the ends of all the segments of
length 12.

The spirograph
The object here is to draw a `spirograph' curve which goes through
all the positions which the zigzag has reached after k steps, for k =
1,2,3 ..... In fact the curve is given precisely by equation (11.7), except
that instead of k taking only integer values 1, 2, ... , s, where s is the
number of steps (given by equation (11.6)), it takes on all real values
from 0 to s. Thus dividing the interval [0, s] into say 1000 parts we will
set up an array of values of k:
n=1:1000;
k=s*n/1000;
Thus k goes in 1000 steps from s/1000 to s. Plot the curve given by
equation (11.7) in white after the zigzag has been drawn, calling the
resulting M-file =4.m. Give examples. Figure 11.3 shows the case 11 =
100, 12 = 40, 01 = 45, 02 = 9. Can the spirograph close up before the
zigzag does?

The special case 01 = -02

Show that, apart from the initial c term in equation (11.7), the remaining
terms become, when 01 = -02 = 0say,
/
((li + 12) sin Ck - 2 0, (-1k + l2) cos I k - 1)0) .
2 sin i 0

Writing this as (X (k), Y(k)), where k, as above, is now a real number

 Zigzags and Fast Curves 135

between 1 and s, show that (X, Y) describes an ellipse with horizontal

and vertical axis lengths
11 + 12 11 - 12
sin 20 sin B
2
respectively. (Note that this ellipse has centre at c, not at the origin,
since we ignored the c in equation (11.7).) Test this out with some values
of 0, drawing both the zigzag and the spirograph. How do you make an
ellipse with horizontal axis shorter than its vertical axis? When is the
ellipse a circle?

Spirographs with cusps

The form of equation (11.7) shows that the spirograph curve is actu-
ally an epicycloid (compare Chapter 10, Investigation A) obtained by
rolling a circle on another circle and tracing the path of a point P rigidly
attached to the latter circle. In fact it can be shown that the fixed cir-
cle has radius a, the rolling circle has radius b and the point P is at a
distance d from the centre of the rolling circle (see Figure 11.4 and note
that a, b, d have nothing to do with their previous uses in this project!),
where
011, , 12
a = (01 - 02)l1 b = d =
202 sin 01 ' 202 sin 01 2 sin 202

2 2
(You need not check these.) A case of special interest is b = d which
means that the point P is on the circumference of the rolling circle. This
always produces cusps on the spirograph. Show that this amounts to
1101 sin 202
2
02 sin 201

Amend zz4. m to zz5. m which chooses this automatically for 12 once

01i 02 are given (remember 11 = 100).
Investigate cases where there are cusps on the spirograph. For exam-
ple, what determines the number of cusps?

B Fast curves
Aims of the project
We shall examine various candidates for a `fast curve': taking a smooth
wire in the shape of such a curve, we want to minimise the time taken
by a bead to slide down the wire. The `fastest curve' is known to have
the shape of a cycloid (see, for example, [17]) but here we shall also use
136 Zigzags and Fast Curves

Fig. 11.4. Generation of an epicycloid (spirograph) by a circle rolling outside

another circle.

numerical integration techniques to compare the `speed' of various simple

curves such as parabolas and broken lines. There is more information
on the same topic in [9].
Mathematical ideas used
This project is about minimisation: finding shortest times of descent.
These are expressed by integrals. Some differentiation is involved, though
most of the minimisation is done numerically. L'Hopital's rule is used
to evaluate a limit. Various curves such as parabolas and cycloids are
used as examples.
MATLAB techniques used
Integration via the quad8 numerical integration package is used, min-
imisation of a function of two variables using fmins and solution of
equations using f solve. There is some use of global variables in the
given M-files.

Introduction and formulae for reference (no friction case)

This project is about a bead sliding down a smooth wire, lying in a
vertical plane, between two points not in the same vertical line. The
object is to investigate the time taken and to compare the times for
different shaped curves. There is a famous result, dating back to the
 Zigzags and Fast Curves 137

seventeenth century, which says that the shortest time of descent is

achieved by a curve called a `cycloid'. We shall meet the cycloid
curve later. The proof that this is the `curve of shortest descent time' or
`brachistochrone' will not be given here; it is not hard, but uses the first
steps in a subject called the calculus of variations. See, for example, [17].
The proof assumes that the wire is smooth: no friction forces act. When
there is friction a slightly different result holds; there is a little work on
this case at the end of the project.
Consider a curve joining the points (0, c) and (d, 0) in the plane, as
in Figure 11.5, where c = 2, d = 4. If a bead slides under gravity but
without friction or air resistance down this curve, starting at rest at
the point (0, c), then energy is conserved. This implies that the kinetic
energy at any moment is equal to the potential energy lost up to that
moment, so that if v is the speed at time t and y is the y-coordinate at
time t, then
1
2mv2 = mg(c - y), i.e. v = 2g(c - y), (11.9)

where g is the acceleration due to gravity and m is the mass of the bead.
Now v = ds/dt, where s is the arc-length along the curve, so the time
t can be found by integration. Of course the value of g will depend on
what units are being used. In order to make the times of descent of a
reasonable size (round about 10), the value of g used in the M-files is
set at 0.1.
There are various forms of the resulting formula for the total time of
descent from (0, c) to (d, 0), depending on whether we integrate with
respect to x or whether both x and y are functions of another variable
0 say, as will be the case for the cycloid. We state these forms together
here; you can assume them during the project. Note that most of the
integrations are performed numerically using a MATLAB routine called
quad8. At the end of this project there is a note about evaluating the
integrals numerically.
The time of travel from (0, c) to (d, 0), where it is assumed that c >
0, d > 0 and that x increases steadily along the curve, is

1
d
1 + y/2 1 f91 x/2 + y/2
dx, (ii) d0 (11.10)
(1) 2g J0 c-y 0
c-y
where, in (i), y is a function of x and ' means dl dx, while, in (ii), x, y
are both functions of 0, ' means d/d0 and 00, 01 are the values of 0 at
the points (0, c) and (d, 0) respectively.
138 Zigzags and Fast Curves

Fig. 11.5. A curve joining the points (0, 2) and (4, 0).

Note that there is a small problem in evaluating the integrals above,

for, at x = 0 in (i) or at 0 = Bo in (ii), we have y = c so that the integrand
is infinite. Nevertheless, so long as the curve does not go above the line
y = c, the integrals do converge (the time of descent is finite!).
MATLAB has a numerical integration package called quad8, but it
requires that the integrand is finite over the range of integration. We
get round this problem by a completely naive approacht. We use quad8
to integrate over an interval which starts just after x = 0 (or 0 = Bo)
and use a separate estimate for the remaining tiny piece of the original
range of integration. This seems to work reasonably well in practice.
At the end of this project we shall go into some more details on the
integrals. For the time being, you can just use the M-files as directed
and take it on trust that they give reasonably accurate results!

t There are techniques called singularity subtraction techniques which are less naive.
 Zigzags and Fast Curves 139

The fastest parabola

Here, we shall take various parabolas between the points (0, c) and (d, 0)
and find the times of travel, finding in the process the fastest parabola.
(i) Consider a parabola y = axe + Ox + c which passes through the
point (0, c). We shall insist that a > 0 so that the parabola is `convex
upwards'. Show that if m is the x-coordinate of the minimum point on
the parabola, and if the parabola passes through (d, 0), then
a= c
d(2m - d)
-tam.
Note that in order to fulfil our requirement that a > 0 we need m > a d.
We are only interested in that part of the parabola between x = 0 and
x = d. If d < m < d then the bead goes below the x-axis before rising
2 (d, 0) while if m > d then it is downhill all the way for the
to the point
bead.
(ii) Take c = 2, d = 4. Sketch the parabolas for m = 2.5, 3, 4 and
5. (The M-file paramc. m, which plots parametric curves, is available if
you really need it, but it would be far quicker here to sketch by hand,
knowing the general shape of a parabola.) Make a guess as to which
parabola you think might give the fastest ride for a bead sliding from
(0, 2) to (4, 0).
(iii) The M-file slide 1. m uses the MATLAB package quad8, plus a
separate approximation for 0 < x < 0.01 (see the Introduction above
and the note at the end of this project) to evaluate the integral (i) in
equation (11.10) in the case of the parabola. It prints out these two
times and then the total time of descent is given.
The M-file requires another function file containing the function to
be integrated, and this is called slidelfn.m. Look carefully at both
these M-files since you will need to amend them for other curves later.
In particular look at the (rather tiresome) need for global variables in
these two M-files. Both the function y and its derivative, called yl in
the M-files, are used.
Keeping to c = 2, d = 4, find the times of descent for m = 2.5, 3, 4
and 5.
Now amend the M-file (calling it slide2.m) so that it takes values of
m from say d + 0.1 to din steps of 0.1 and finds the time of descent for
a
each one, storing it in a vector called result, finally plotting the graph
of m against the time. So you will want something like
m=d/2+.1:.1:d;
140 Zigzags and Fast Curves

(d,0)

Fig. 11.6. A broken line joining (0, c) and (d, 0), the break point being (a, b).

result=zeros(length(m),1);
for i=1:length(m)

result(i)=time;
end % of i=1:length(m) loop
plot(m,result)
Note that the global statement in slide1.m MUST remain outside
the for loop! There is no need to change slidelfn.m.
For c = 2, d = 4 print out the graph of m against time of descent
and estimate from it the value of m which gives the minimum time of
descent. Draw a hand sketch of the corresponding parabola. (Was your
guess right?!) For c = 2, d = 3 use the same method to find the `fastest
parabola' and sketch it, but do not print out the graph of m against the
time.

Broken lines
Suppose we join (0, c) and (d, 0) by two straight segments, joining at the
`break point' (a, b) say. See Figure 11.6. We shall look for the fastest
broken line. So here c and d are given, and a and b can be changed.
(iv) Show that the equations of the two segments are
ac)x x
y=c+ (x < a), y = b(a (x > a).
(b
- d)
(v) For a straight line descent it is actually possible to calculate the time
exactly, by evaluating the first integral in equation (11.10). Consider a
line of slope m, between (xo, yo) and (xl, yi); note that m will generally
 Zigzags and Fast Curves 141

be negative here. Replacing dx by dy/m in the integral, show that the

time of descent from height yo to height yl is
1+m2 yl 2 1+m2
mv29 f,,,
1

c-
dy_

mVg
2g
c-yo- c-yl). (11.11)

Taking c = 2, d = 4 as before (and 0 < a < d, 0 < b < c), show that
the time of descent along the broken line is
2 ( Va2 2(2 b b)2 + (f -)b (4 - a)2 + b2
(11.12)

You can use the built-in MATLAB function fmins to find the minimum
of the function of a and b defined by equation (11.12). You have to make
an M-file, say slide3fn.m, containing the function to be minimised.
This takes the form
function realtime=slide3fn(p)
g=.1; c=2; d=4;
a=p(1); b=p(2);
realtime = ;% Formula for the real time of descent,
as above

You then call fmins by fmins ('slide3fn' , [*, *] ' ), where the asterisks
are replaced by a sensible guess as to the values of a, b which give the
minimum. (Note the ' after the vector, indicating transpose.) Make a
sensible guess and find the values giving the minimum time of descent.
(Note that if you make a guess of say a = 1, b = 2.1 > c then, as you
might expect, fmins does not give a reasonable answer.) Draw a sketch
of the broken line which achieves the minimum time. Also find the
corresponding minimum time. How does this compare with the fastest
parabola?
(vi) An interesting case of equation (11.12) is when b = 0, which makes
the second term now of the form 0/0. Use L'Hopital's rule on the ex-
pression

to show that when b = 0 the time of descent is (still with c = 2, d =

4, a < d),
a2+4+4-a).
142 Zigzags and Fast Curves

rO

Fig. 11.7. One complete arch of the cycloid starting at (0,2), generated by
rolling a circle of radius r = 1.5 underneath the line y = 2. The cycloid is
parametrised by 0.

Show, using calculus, that the value of a between 0 and 4 which makes
this a minimum is 2/v'-3-.

The cycloid
A cycloid curve is obtained by rolling a circle along a straight line and
tracing the path of a point attached to the circumference of the circle.
In our case we roll a circle of radius r `under' the straight line y = c. See
Figure 11.7. The cycloid is parametrised by the angle 8 in the figure;
the coordinates of the moving point P are
x=r(8-sine), y=c-r+rcos0. (11.13)

Figure 11.7 shows a complete `arch' of a cycloid, for 0 < 8 < 21r. Note
that 0 = 0 gives (x, y) = (0, c) = (0, 2) in the figure.
As before we take c and d as given. We want the cycloid to pass
through the point (d, 0); in order for this to happen we need, for some
r>0and0<8<21r,
rsinO-rO+d=O, rcos8-r+c=0. (11.14)

Since these are two equations for two unknowns r, 8 it is reasonable

that they might have a solution, maybe even a unique solution. Since
they are rather unpleasant equations, we have to solve them using one of
MATLAB's equation solvers, e.g. f solve. (If by any chance your version
 Zigzags and Fast Curves 143

of MATLAB lacks f solve then you could try using one of the custom
written M-files from Chapter 15 called full_new.m and gauss_ja.m.)
(vii) Taking c = 2, d = 4, find the solution of equation (11.14) as follows.
Create a function file say cycfun.m of the following form
function q=cycfun(p)
c=2; d=4;
r=p(1); theta=p(2);
q=zeros(2,1);
q(1)=r*sin(theta)-r*theta+d;
q(2)=r*cos(theta)-r+c;

You can then solve the equations (11.14) with c = 2, d = 4 by

fsolve('cycfun',[r0, theta0]')
where rO and theta0 are a first guess as to the solution for r and 0. Don't
forget the ' after the vector, indicating transpose. Make a reasonable
guess and find the resulting solution. (Hint: Try r = 1 for the radius.)
Denote the solution for 0 by 01 (referred to as thetas in the M-files).
Sketch the cycloid over the range from 0 = 0 to 0 = 01, i.e. between the
points (0, 2) and (4, 0).
The M-file slide4.m uses the second formula of (11.10) to calculate
the time of descent when a curve is parametrised by 0. The range of
values of 0 is from 0 to 01. This M-file uses a function slide4fn.m in
the same way that slidel.m uses the function slidelfn.m.
What is the time of descent for the cycloid? There is a famous
theorem, first proved in the seventeenth century by Newton, Leibniz,
Bernoulli and others, that the cycloid is the curve which gives the fastest
descent. How much faster is it (as a percentage) than the fastest broken
line?
(viii) Notice that (in the present friction-free case) the cycloid is still
best when c = 0, i.e. when the start and end are on the same horizontal
level. In this case the bead, on arriving at (d, 0), has zero speed. When
c = 0, d = 4 what is the time taken along the cycloid? Beware that you
might need to take 01 just a little less than the true value to avoid an
infinite time being produced by slide4.m. Alternatively, with a little
cunning you can find the time it takes to travel half-way along the arch
of this cycloid and double the answer.
144 Zigzags and Fast Curves
The square root curve
Find the value of a so that the curve
y=aV1X- +c,
which passes through (0, c), also passes through (d, 0). Find the time of
travel between these points along this curve in the case c = 2, d = 4 and
also sketch the curve in this case. You can adapt slide1.m here, calling
the result slide5.m, remembering to produce also an M-file slide5fn.m
which contains the function y and its derivative yl, as in slide if n.m.
How does this time compare with that for the cycloid?

The integrals (optional reading)

Using the formula (11.9) in the form
ds =
2g(c - y),
dt

and the standard formulae

ds ds dx _ 1 + y/2 dx
dt dx dt dt
(where y' = dy/dx), prove formula (11.10)(i).
As was pointed out in the introduction to this chapter, MATLAB's
numerical integrator assumes that the function being integrated is finite
over the range of integration, so in this project the integrator is used in
equation (11.10)(i) over the range 0.01 < x < d and a separate estimate
is obtained over the initial small interval 0 < x < 0.01. In fact, to
obtain the estimate, we shall assume that y' = dy/dx is constant for
0 < x < 0.01. It can be shown that the integral between these limits is
then given by the formula of (11.11), namely
2(1 + yi2)(c - y)
7

(note that y' will be negative).

In the approximations used in this project, the values of y and y' are
taken at the endpoint x = 0.01.

Friction (optional extra)

It is interesting to make the above considerations more `realistic' by
introducing friction. This is surprisingly easy to do, since the only effect
is to replace the formula (11.9) by the slightly more complicated formula
v = f2g(c - y - x),
 Zigzags and Fast Curves 145

where p is the coefficient of (kinetic) friction. This means that the bead
is acted on by gravity downwards, the normal reaction of the wire, and
friction, which is p times the normal reaction. The equation for v then
follows from Newton's second law of motion (you need not check this
unless you want to!).
As a result, the formula (11.10)(i) for the time of travel becomes

1 0d 1 + yrz
dx.
VI-2-g c - y -x
Adapt slidei.m and slidelfn.m to find the fastest parabola when =
0.3. Call the M-files slide6.m and slide6fn.m . Beware that taking
values of p larger than about 0.5 may result in `infinite' times of travel
along some parabolas. Why is this?
 12
Sequences of Real Numbers

This chapter is about sequences of real numbers, and we begin with an

introduction to special classes of sequences. There follow three investi-
gations which explore some more properties. The first (A) is on Mobius
sequences, which are described below, while the second (B) is entirely
on quadratic sequences. The third one (C) is mostly on quadratic and
exponential sequences but there is some mention of Mobius sequences
again. Check with each investigation to see which parts of the prelimi-
nary material (12.1-12.3) are required.

12.1 Mobius sequences

We shall consider several sequences of the form
axn + b
xn+1 = Cxn + d' n = 0, 1, 2,37 .. .

where a, b, c, d are real numbers. Once x0, also real, is given, the whole
sequence is determined and consists of real numbers. It is called a
Mobius sequence. There are two M-files available. The first is mob ius . m
which calls for the numbers a, b, c, d, x0 and the number of iterations
to be performed, that is, how many terms of the sequence are to be
calculated. It then plots a bar-chart of the values.
The second M-file is mobius i . m which dispays the values of the suc-
cessive xi as a column vector, using `long format' for greater accuracy.
(i) (a, b, c, d) = (1, 2, 1, 1), x0 = 3. It is fairly clear from mobius.m and
mobius1.m (use 20 iterations) that this sequence is convergent. The
limit also looks suspiciously familiar. Running with other values of x0
produces the same result. Note that some values of x0 produce 00 on
the way, as with x0 = -1.5 for example. (Here, x2 = 00.) However,

146
 12.1 Mbbius sequences 147

this is not a serious problem since the formula suggests that if xn = 00

then x,,,+1 should be simply a/c. In fact the M-files take account of this
and do not crash if an infinite value is obtained. (In practice it is more
usually a very large value, positive or negative.)
The M-file mob ius . m is reproduced at the end of this section to show
you how the `infinity' problem is overcome.
(ii) (a, b, c, d) = (1, 1, -2,1), xo = 3. Using mobius. m and 500 iterations
seems not to produce a limit. Using mobius i . m and 20 iterations it
certainly seems that the numbers are jumping around a lot. Such a
sequence could be called `wildly divergent' or `oscillatory' or `chaotic'.
(iii) (a, b, c, d) _ (1, 1, -1, 1), xo = 3. This gives a repeating cycle of four
values 3, -2, - 32, which can be seen by either of the two M-files. Such
a sequence is said to be periodic with period four.
Clearly a lot of different behaviours are possible. Note that, in the
last two above, we can argue that if x,,, -+ 1, then xn+1 -* 1 too, so that,
respectively,
1
-21+1' thus l2=-2,
11+11;
l= thusl2=-1.
In both cases this is a contradiction (remember we are working with
real numbers). Hence it is certain that neither of these two sequences
can have a single limit 1, but the modes of divergence are very different.
Note that an apparently wildly divergent sequence might in fact be pe-
riodic with a very large period, and this might be obscured by numerical
inaccuracies in computation. There is no easy answer to this-and we
shall not look for an answer here!
Applying the above method to the first example shows that if the
sequence has a limit then that limit must be fvf2-. It is not quite so
easy to prove that it does have a limit, but this can be done. The `cobweb
diagram' we shall meet shortly makes this very plausible indeed-in fact
it can be made into a proof. It is slightly mysterious that 1 = -\ is
never the limit unless you actually start with xo = -v"2-.
Here is the M-file mobius. m. Note that x,,, can be `oo'-this happens
when xn_1 = -d/c. But it is still possible to calculate Xn+1, namely
xn+1 = a/c. Note that `infinity' here just means a very large number,
which in MATLAB is referred to as Inf.
148 Sequences of Real Numbers

M-file to plot Mobius sequences as bar chart.

If x is the current term then the next one
is (ax+b)/(cx+d).
The constants a,b,c,d are input by the user.
You are also asked for the number of iterations
to be performed and the starting value x0.

a=input('Type a
b=input('Type b ');
c=input('Type c ');
d=input('Type d ');

xO=input('Type xO ');
n=input('Type the number of times to be iterated ');

X= [x0] ;

for j=1:n
if x(j)==Inf I x(j)==-Inf
x(j+1)=a/c;
else
x(j+1)=(a*x(j)+b)/(c*x(j)+d);
end
end;

bar (x)

12.2 Cobweb diagrams

These are a diagrammatic way of following sequences under iteration.
Given a function y = f (x), we start with x0 and find in succession
x1 = f(xo), x2 = f(xi), etc. Start with the graph y = f(x) and
add the line y = x; then proceed as follows. From the point (xo, 0) on
the x-axis a vertical line is drawn to meet the graph at (x0, xi), then a
horizontal line to meet the line y = x at (x1, x1). The process is then
repeated: a vertical line to meet the graph at (x1i x2), then a horizontal
line to meet y = x at (X2, x2), and so on. The successive vertical lines
have x-coordinates x1iX2..... Thus the progress of the sequence can be
followed.
 12.3 Mobius functions and powers of matrices 149

Figure 12.1 shows two examples of cobweb diagrams, one in fact for a
Mobius sequence and one for a quadratic sequence (these are discussed
in Project B below). Possibly the second one gives more of an idea of
why the name `cobweb' is used!
Mobius sequence cobwebs
The M-file cobm.m plots cobweb diagrams for Mobius sequences, which
use the function f (x) = (ax+b)/(cx+d). It requires input of a, b, c, d, x0
and the upper and lower limits of x and y on the diagram. Iteration is
done 30 times, or until it seems that convergence is certain. The sequence
of values of x is printed out at the end.
Note that, when there is a vertical asymptote to the graph of the
function f, this will be replaced by a vertical line joining two points of
the graph. Here are some suitable values:
1 In Example (i) above, lower limits of x and y equal to -3 and
upper limits of x and y equal to 3, x0 = 2 or - 2. If you want to
see the action in close-up then you could take say 1.3 < x < 1.5
and the same for y.
2 In Example (ii) above, lower limits of x and y equal to -10 and
upper limits of x and y equal to 10, xo = 2.
3 In Example (iii) above, lower limits of x and y equal to -5 and
upper limits of x and y equal to 5, x0 = 3.
It is fairly clear in each case what is happening.
The points where the graph of f meets the line y = x are fixed points
of the function y, that is, writing y = f (x), they satisfy f (x) = x. It is
these points, if any, which are limits of the sequence.

12.3 Mobius functions and powers of matrices

Let
a b axn + b
A= c d ' xn+1 = cxn + d

Clearly there is going to be a close connection between the matrix A

and the Mobius sequence. Here is one of many ways of expressing this
connection.
Theorem 1 (i) There is a real number a (depending on n) such that

AI I =al x11 I

1
150 Sequences of Real Numbers

1.6

1.4

1.2

0.8

0.6

0.8 1 1.2 1.4 1.6 1.8

0.9

0.85

0.8

0.75

0.7

0.65

0.6

0.55

0.5

0.45

0.4
0.4 0.45 0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9

Fig. 12.1. Two cobweb diagrams, the upper one corresponding to a Mobius
sequence and the lower one to a quadratic sequence.

In other words,

if A I )=(wl2 ) then xn+l = w2.

1 W
 12.3 Mobius functions and powers of matrices 151

(ii) For any integer k > 1 there is a real number a (depending on k)

such that

Ak (
In other words,
/ x0 ) wl
if A c then xk =
1 W2

Example 1
Let

A= 1 1 1 ), x0=3.

Then with v = (3,1)T, we have Av = (5, 4)T and the a in (i) of Theorem
1 is 4, while xl = 4. Taking k = 4 in (ii), A4v = (75,53)T and x4 =
53 = 1.4151; the value of 0 is 53.
Example 2
If you type
format long
A=[1 2; 1 11;
v=[2 1] ' ;

w=A- 20*v ;
w(1)/w(2)
then you should get the answer 1.41421356237310, which is the same
answer as you get by typing sqrt (2) . By (ii) of the Theorem 1, the
number obtained is x20 when x0 = 2.
The proof of Theorem 1 is straightforward: (i) is proved by multiplying
out the two sides and taking a = cx,,,+d. Then (ii) is proved by induction
on k. The case k = 1 is just (i) with n = 0. Assuming (ii) for k, we
proceed as follows:

Ak+1 ( i ) = A/3 1 /3A 1 ) = ,3a (x )

1 1 1
The first = uses the induction hypothesis, the second = uses the fact
that /3 is just a number and the third = uses (i) with n replaced by k.
This completes the proof of (ii) by induction.
If Ak is a scalar matrix-that is, of the form AI, where I is the identity
152 Sequences of Real Numbers
2 x 2 matrix-then we have

Ak( )=AI I,
1 1
so that by (ii) of the theorem xk = x0i i.e. the values of the Mobius
sequence have recycled back to xo. So periodic sequences can be detected
by calculating powers of the corresponding matrix.
We can also work `backwards': for example, let us work out the start-
ing value x0 which, for the sequence
3xn, + 5
xni-1 =
xn+2
1

gives x6 = -2. Note that when x0 is chosen in this way, the next term
x7 will be `infinite', since the denominator x6 + 2 vanishes. We want
x
A6 ( 1 12 /
for some ,Q. This gives

-6 1-22658
\ 1 ) A 12)-( 2649
where you can verify the last equality by typing A" (-6) * [-2 1] ' hav-
ing first input the matrix

A=(13 5
2).

(You may find that -22658 actually appears as something like -22657.999
of course!) Thus x0 = -22658/12649 = -1.7913. If you use mobiusl.m
to find x7 for this sequence (input x0 in the above fractional form) you
will find that it gives a very large value. Nevertheless x8 is given cor-
rectly as 3.
There is a highly significant fact about the direction of the vector
w = Akv for large values of k. Going back to the matrix A of Example
2 above, try typing
[X D]=eig(A)

You will find that one of the eigenvectors of A is parallel to the vector
w obtained in Example 2. For example, you can type
xl=X(:,1)
y=x1./w
 12.3 Mobius functions and powers of matrices 153

to pull out the first eigenvector of A and then compare it with w in direc-
tion. The two entries in the vector y should be approximately the same.
Note that the eigenvector used here is in fact the one corresponding to
the larger eigenvalue.
With matrices we can equally well consider 3 x 3, but note that this
does not have an obvious analogue in terms of sequences. Try
A= [1 2 3;-1 2 -3;1 -3]
-2
and (almost) any starting vector v. This matrix has one real eigenvector,
(-0.5433, 0.2849, 0.7897), as you can discover by typing [X, D] = eig (A).
This has the same direction as Anv for large values of n.
Note on two M-files
It can happen that powers of a matrix A have entries which become very
large indeed, and working out Anv for a large n can be hazardous. There
are two M-files available which overcome this problem when we only
want to know the direction of Anv. They work out Av, A2v, A3v,.. .
but after each step rescale the answer to be a unit vector. They are
called matit2.m for 2 x 2 matrices and matit3.m for 3 x 3 matrices.
Both require previous input of the matrix A but not of the vector v.
Look at them if you are curious.
For those who like to know the theory, we have the following.
Theorem 2 If A n v does have a limiting direction, then this limit must
be an eigenvector of the matrix A.
Proof If we suppose Anv/IIAnvIJ -+ w for some w 0 as n -+ oc, then
A(Anv/II AnvII) -p Aw, so
An+1V
IIAn+1vlI
IIAn+1vII
IAnvII -' Aw.
But, as n oo the first fraction here tends to w and the second to
some real number a, so we obtain Aw = aw, which means that w is
an eigenvector of A. Of course this does not prove that Anv does have
a limiting direction equal to an eigenvector, only that eigenvectors are
the only possible limiting directions.

A Investigation on Mobius sequences

Aims of the project
We shall explore Mobius sequences both experimentally and theoreti-
cally. You should read through 12.1-12.3 first.
154 Sequences of Real Numbers

Table 12.1. Values to try for the Mobius sequence.

a b c d

2 1 1 -2
3 2 4 3
3 2 2

-/ /
1
1 3 -2 1

1 1
-1 1

Mathematical ideas used

Matrices, sequences, mathematical induction and limits of sequences.
There is some calculation of derivatives in the theory part of the project.
You should read through 12.1-12.3 first.
MATLAB techniques used
Only pre-written M-files are used here to iterate Mobius sequences, to
find powers of matrices and to plot cobweb diagrams.

Mobius functions and matrices

(i) For the values of a, b, c, d given by the Table 12.1 decide using the
M-files mobius.m and/or mobius1.m whether the corresponding Mobius
sequence is convergent or divergent, and in the latter case whether it
appears to be `wildly divergent' ('chaotic') or appears to be periodic
(cycling round in a loop). Write down your results, specifying your
initial value(s) xo and the numbers of iterations you used.
Also use the M-file cobm.m to plot cobweb diagrams. To iterate you
use <Enter>. Make rough sketches of these on paper in the cases where
you assert the sequence is convergent or periodic, specifying your range
of values of x and y as well as starting value(s) xo.
Note: Use the expression sqrt (3) when prompted for an input. MAT-
LAB will substitute the numerical value for you.
(ii) Take the same examples as in the previous exercise, associating
(a, b, c, d) with the matrix

Use the matrix multiplication of MATLAB to verify that, for the

periodic sequences, the nth power of the matrix (n = the length of the
period) is a scalar matrix. This means a matrix of the form
 12.3 Mobius functions and powers of matrices 155

aI = I A 0
),
where I is the 2 x 2 identity matrix. Write down the value of A in each
case. Also verify that, for (a, b, c, d) = (1, 1, -1, 1), the fourth power of
the matrix is a scalar matrix, and write down the value of A.
(iii) What is the connection between A4 = AI in the last part of the
previous exercise and the fact that the corresponding Mobius sequence
is periodic with period 4? (See 12.3.)
(iv) For which of the examples in Exercise (i) does the M-file matit2.m
give a limiting direction for the iterated product Anv? State the value(s)
of the vector v which you used and the limiting direction obtained.
Verify that this is an eigenvector of A by calculating the eigenvectors.
State both eigenvectors in these cases.
(v) For (a, b, c, d) = (1, 2, 1, 1), find the inverse of the corresponding
matrix (e.g. by using inv(A)). Hence find (exactly) a starting value
x0 for which x5 = oo. Is there any possibility of finding xo such that
X50 = oo?
(vi) This part has nothing to do with Mobius functions but involves pow-
ers of 3 x 3 matrices. Compare 12.3. A population of bugs is of three
kinds: black, red and green. At time t = 0 there are 10 black ones but
no red or green ones. At time t = 1, the following three things happen
simultaneously:
(a) each black bug splits into a black bug, 2 red bugs and a green bug;
(b) each red bug splits into 1 black bug and 2 green bugs;
(c) each green bug splits into 3 red bugs.
The same happens at times t = 2, t = 3, etc. The bugs in this problem
are immortal (this is obviously pure mathematics).
How can you use matrix iteration to discover, for large values of t,
what percentages of the population are black, red and green? Does the
initial population (at t = 0) make any difference? Is there a connection
with one of the eigenvectors of a matrix? Hint: Consider the matrix
1 1 0
12 0 3
1 2 0
156 Sequences of Real Numbers
Theory of Mobius sequences
Here you have the opportunity to work out some of the theory of these
sequences. Illustrate the theory with suitable examples, using the M-
files.

(vii) Let f (x) = (ax + b)/(cx + d), where c 0 and ad - be 0.

The fixed points of f are numbers x such that f (x) = x. Show that if
(a - d)2 + 4bc > 0 then there are two such real fixed points a and /3, say.
(viii) Remembering xn+l = f (xn), f (a) = a, f (/3) = /3, show that
xn+1 - a c/3+d xn - a (12.1)
xn+1-/3 ca+d xn - a
Hint: First show, using a = (aa + b)/(ca + d), that
(ad - bc)(xn - a)
xn+1 - a = (cxn + d)(ca + d)'
and a similar formula for xn+1 - /3.
Now deduce, by repeated application of equation (12.1) for n - 1, n -
2,..., that

xn - a (ca+d) n (xo-/3
(ix) Show that f'(a) = (c/3+d)/(ca+d), and f'(/3) = (ca+d)/(c/3+d).
Hint: It is easy to check that f'(x) = (ad - bc)/(cx+d)2. Remembering
that a and /3 are the roots of cx2 - (a - d)x - b = 0, deduce that
(ca + d) (c/3 + d) = ad - bc.
(x) Deduce that if I f(a) 1< 1, then xn -+ a as n -> oo.
(xi) Show that if (a - d)2 + 4bc < 0, then the fixed points are not real.
Why does this imply that the sequence {xn} cannot have a limit?

B Attracting cycles
Aims of the project
This project is about sequences formed by iteration of a quadratic (de-
gree 2) function. These sequences do not generally have a unique limit,
but have `many limits' in the sense that successive values of the sequence
come close to a collection of different numbers. We shall investigate these
`attracting cycles' both mathematically and experimentally. More de-
tails of the mathematics involved here can be found in [2]. There is
 12.3 Mobius functions and powers of matrices 157

no need to read the material on Mobius sequences in order to do this

project.
Mathematical ideas used
We use sequences of real numbers defined by a quadratic formula. There
is some manipulation of polynomials and curve sketching involved in the
calculations.
MATLAB techniques used
The project uses existing M-files for analysing quadratic sequences, and
also to find the roots of a polynomial equation.
We shall not go much into the theory of attracting cycles; the point of
this investigation is to find some examples. We write f (x) = Ax(1 - x)
and xn+1 = f (xn), n = 0,1,2,..., to define the quadratic sequence
xo, X1, x2, .... Once A is given, this sequence is determined by the first
term x0.
Write f P for the pth iterate of f, that is, f2(X) = f (f (x)), f 3 (x) _
f(f(f(x))), etc. Thus xP = fP(xo). For a given value of p, we are
particularly interested in finding values of A such that

(12.2)

This is because of the following result: If A is such a value, then, for

some q which is a factor of p, the numbers 2, f '(1), for i = 1, ... , q - 1
form an `attracting q-cycle'. This means that if you start from x0 in
the interval (0, 1) and iterate f, then the values fP(xo), p = 1, 2, 3, ... ,
come close in succession to these q numbers.
(i) For p = 2 show that f2(x) = A2x(1 - x)(1 - Ax + Axe) and hence
show that equation (12.2) becomes A3-4A2+8 = 0. Put the coefficients
into a vector v (starting with the leading term) and use roots (v) to find
the roots of this equation. Ignore the negative root. One positive root
is 2; for this value check that in fact f(2) = 2. For the other root,
use cobq.m to confirm that, starting with various x0, the values of the
sequence approach two numbers. Write down these numbers as given by
the M-file.
(ii) The M-file quadn.m plots the graph of any iterate of f (for any given
A), and plots the lines x = 2, y = 2 in green on the same diagram. It
allows you to change A without changing the other settings. Instructions
appear on the screen. Note that after a graph is drawn, you will need to
158 Sequences of Real Numbers
press <Enter> to return control to the main MATLAB window. Then
you will be prompted to press 0 to stop, or 1 to choose another A.
By homing in on the point (2 , 2) you can get an extremely accurate
value of A such that equation (12.2) holds. (Adjust A until the graph goes
through the intersection of the green lines.) Use it for the third iterate
f3, and find the value of A near to 3.83 such that equation (12.2) holds.
You should be able to manage a couple more decimal places at least,
with judicious use of the M-file. Having found A, use cobq.m to check
that, starting with various x0, the sequence approaches the `attracting
3-cycle' as above. Write down the numbers occurring in this 3-cycle.
For each x0 that you use, state how many iterations were needed before
the sequence approached these numbers to four decimal places.
(iii) For p = 6 there is a solution of equation (12.2) near to 3.63. Find
this solution as nearly as you can, using quadn.m, and use cobq.m to
find the set of six numbers (the `attracting 6-cycle') which the sequence
approaches. For p = 16 there is a solution near to A = 3.55. Does this
give an attracting 16-cycle? Give details (using the M-files, of course).
(iv) Show that the (12.2) for p = 3 is
A 7 - 8A6 + 16A 5 + 16A 4 - 64A 3 + 128 = 0.

Put the coefficients into a vector p (starting with the leading coefficient
1) and find the solutions with roots (p). Ignore negative real roots and
complex roots. One root should be 2; check that for this A, f(i2) = i2
For the other root, use cobq. m to verify that there really is an attracting
3-cycle, and write down the numbers in it (you should get the same
limiting numbers whatever x0 in (0, 1) you start from).
(v) For p = 4 there is a solution near to A = 3.5. Find this solution
to several more decimal places, using quadn.m, and use it in cobq.m to
find the numbers in the resulting attracting 4-cycle.
(vi) Show that the graph y = f (x) crosses the line y = x when x = 0 or
x = 1 - 1. Deduce that, for 0 < A < 1, the graph of y = f (x) is entirely
under the graph of y = x for 0 < x < 1. Explain with a (hand-drawn)
cobweb diagram how this makes it very plausible that, for any x0 with
0 < xo < 1, the resulting quadratic sequence is convergent with limit 0.
For 1 < A < 2 show that the graph y = f (x) crosses the line y = x
just once for 0 < x < 1, namely for x = 1- a . Calculate the slope of the
graph y = f (x) where this happens, showing that it is positive. Again
using a hand-drawn cobweb diagram explain why this makes it plausible
 12.3 Mobius functions and powers of matrices 159

Fig. 12.2. The `period doubling' diagram. For each value of A on the horizontal
axis, the points above it are the values of x in the resulting attracting cycle.
So from A around 3.5 there are longer and longer attracting cycles.

that, for 0 < x0 < 1, the resulting sequence is convergent. What is the
limit this time? Confirm your assertion by means of some examples,
using the M-files.
(vii) The M-file perdoub. m (for `period doubling'!) takes 100 values
of A between two chosen limits 11 and 12. For each one the numbers
xl, , x1oo are calculated (with xo = 1) and just the terms from the
75th on are plotted on the vertical line above the value A. So the picture
has axes A, x, where 11 < A < 12 and 0 < x < 1. Figure 12.2 shows the
picture for 1 < l < 4.
Try running it for say 11 = 0, 12 = 2. How does the resulting picture
confirm what is shown in (vi) above?
Try running it for 11 = 2,12 = 3.5. What do you observe?
(viii) (Optional extra) Between A = 3.9 and A = 4 there are several
more attracting cycles. Find some of them, using the M-files to the best
advantage that you can. (It is possible to be reasonably systematic in
this.)

C Quadratic and exponential sequences; fixed points

Aims of the project
This is again about sequences, their limits and their attracting cycles-
160 Sequences of Real Numbers
sets of numbers which the terms of the sequence approach. We shall
study two types of sequence, quadratic sequences (as in B above) and
exponential sequences, for example x, 2x, 2(2') and so on. (Exponential
sequences are treated in detail in [2], pp.406ff.) In particular we shall
study `fixed points' of sequences and relate these to their convergence
behaviour. There is also a very little on Mobius sequences in this in-
vestigation; see 12.1 for the definition. You might also want to look at
12.2 which describes `cobweb diagrams'.
Mathematical ideas used
The main mathematical idea is that of a sequence of real numbers. The
theoretical part of the investigation involves calculus and manipulation
of functions.
MATLAB techniques used
Existing M-files are used. You need to amend the M-file which plots
cobweb diagrams of quadratic sequences so that it covers the case of
`exponential sequences'.

Quadratic sequence cobwebs

For a general discussion of cobweb diagrams, see 12.2.
The M-file cobq. m plots a cobweb diagram for the sequence
xn+1 = Axn(1 - xn),
where A is a fixed real number which we take to be > 0. That is,
the function being used is y = f (x) = Ax(1 - x). Try, for example,
A = 3.9221934, xl = 0, xu = 1, yl = 0, yu = 1,x0 = 0.5. (As usual, xl
denotes the lower limit of x, xu the upper limit, and yl, yu are the lower
and upper limits of y.) Try 20 iterations. This produces a 7-cycle of
values, a so-called superattracting 7-cycle.

Quadratic and exponential sequences

(i) Use the M-file cobq.m, in the cases given by Table 12.2. Once
the cobweb diagram is plotted the values of the xi are printed out, so
looking at the last ones can usually confirm your visual impression of a
limit/periodic sequence/chaos (wild divergence). Table 12.2 shows the
numbers to input at the prompts; n is the number of iterations.
In each case state what you consider the behaviour to be; in the case
of convergence give the limit and in the case of periodic sequences, give
the length of the period and the various values of x occurring in the
period. Note that here a `convergent' sequence, means one whose values
 12.3 Mobius functions and powers of matrices 161

Table 12.2. Values to try in the quadratic sequence.

A xo xl xu yl yu n
2.5 0.5 0.5 0.7 0.5 0.7 20
3.2 0.5 0.4 0.9 0.4 0.9 20
3.4 0.5 0.4 0.9 0.4 0.9 40
3.5 0.5 0.3 0.9 0.3 0.9 20
3.8 0.5 0 1 0 1 50
3.83 0.5 0 1 0 1 20

approach a single limiting value. On the other hand a `periodic' sequence

of period k means one f o r which there are k numbers 1k, ... , lk and the
values of the sequence come close in succession to these k numbers. (If
you start with xo = ll say, then the sequence actually cycles round the
numbers 12 i ... and back to 11 again. As in Investigation B, the numbers
11, ... , lk are said to form an attracting k-cycle.) So you look at the
numbers produced by the sequence and see if these `settle down' to a
single number or to a succession of the same numbers repeated.
(ii) Amend the M-file cobq.m to plot the graph of y = kx, where k
is a fixed real number, and iterate this function, producing a cobweb
diagram as before. (Call the resulting M-file cobexp.m.) This is just a
matter of replacing the quadratic function with the function y = kx in
the M-file cobq.m. Note that the sequence being considered here, for a
fixed value of k, is x, k1, k(k') , and so on. It is called an exponential
sequence. Remember that since in the M-file x is a vector of values, you
will have to use k. "x
Investigate whether sequences starting at say xo = 0 are convergent
or divergent, and, if convergent, what the limit is. If it depends on the
value of k, then attempt to find the value(s) of k at which behaviour
changes.
Here are some values to put in cobexp. m to get you started:

k = 0.05, xl = -0.1, xu = 1.1, yl = -0.1, yu = 1.1;

k = 0.5, xl = -0.1, xu = 1.1, yl = -0.1, yu = 1.1;
k = 1.4, xl = 0, xu = 5, yl = 0, yu = 5;
k = 1.5, xl = 0, xu=10, yl = 0, yu = 10.
As an aid to thinking about what happens between k = 1.4 and
k = 1.5, consider where the line y = x meets the curve y = F. Show
162 Sequences of Real Numbers
that this is where k = xl/x. Sketch the graph of y = x1Ix. Hint: Use
x11x = eln(x)/x and show that the derivative of xl/x is xl/x(1-ln(x))/x2.
What is the largest value which this y takes? (Answer this question
by using calculus to find the maximum of y.) How does this relate to
the behaviour of the exponential sequence?

Fixed points
A fixed point of a function f is a value of x such that f (x) = x. Consider
the functions f (x) = (ax+b)/(cx+d), where c 0, and f (x) = Ax(1-x),
where .A > 0. What are their real fixed points (if any)? Let a be a real
fixed point. It is called

attracting if I f' (a) I< 1;

repelling if I f'(a) I> 1;
indifferent if I f'(a) I= 1.

The general idea is that if x0 is close to an attracting fixed point a,

then the iterates xn will be attracted to a, that is x,, - * a as n - oc.
Similarly if x0 is close to a repelling fixed point a, then x1 will be further
away from a. For an indifferent fixed point various things can happen.
Except in (vii) below, you can take your starting value x0 to satisfy
0<x0<1.
(iii) For the case f (x) = Ax(1 - x) show (using mathematics, not the
computer!) that the only values of A > 0 giving indifferent fixed points
are A = 1 and 3. Use the M-files to find out what happens to sequences
x,,, in these cases.
(iv) Show that for f (x) = (-3x - 2)/(4x + 3), there are two indifferent
fixed points. Use the M-files to investigate what happens to sequences
in this case. More generally, show that f (x) = (ax + b)/(cx + d) always
has indifferent fixed points when a + d = 0 and ad - be = -1.
(v) Investigate the fixed points and behaviour of the function given by
f(x) = (7x - 2)/(2x + 3).
(vi) Find one example of a Mobius sequence (f(x) = (ax+b)/(cx+ d))
and one of a quadratic sequence (f (x) = Ax(1 - x)) with an attractive
fixed point, and investigate with the M-files whether sequences really do
have that as limit. (For the Mobius sequence, you are probably better
off guessing than trying to find the condition for an attractive fixed
point.)
 12.3 Mobius functions and powers of matrices 163

(vii) Investigate the indifferent fixed points of f (x) = -x and f (x) _

x2 + X. (In these cases you might want to take xo outside the range
0<xo<1.)
 13
Newton-Raphson Iteration and Fractals

Aims of the project

We shall investigate sequences of points in the complex plane which
have limits equal to roots of certain simple equations. The sequences
are generated by the complex number version of the well-known Newton-
Raphson formula. We shall in particular contrast the case of a quadratic
equation with that of a cubic equation (the contrast could not be greater!).
For a detailed discussion of the Newton-Raphson method for real equa-
tions, see Chapter 15. There are many semi-popular books on fractals,
for example [14].
Mathematical ideas used
Complex numbers, including solution of quadratic equations with com-
plex coefficients are used. There is also mention of a special cubic equa-
tion. Some plane geometry is used, such as perpendicular bisectors of
segments.
MATLAB techniques used
This project uses the ability of MATLAB to draw points and lines in the
plane. The points here will represent complex numbers in the usual Ar-
gand diagram. Basic M-files are written for you, and need some amend-
ments to apply to other examples.

13.1 Introduction
You may have met the Newton-Raphson method for approximating the
roots of an equation f (x) = 0. Briefly, this consists in making a guess
x0 at a root and then refining it by x1, x2, x3, ..., where
f(xk)
Xk+1 = Xk - fT(xk)> k = 0 1 2 .... (13.1)

164
 13.2 The equation z2 + 1 = 0 165

In this project we shall apply this method to the solution of certain

complex equations.

13.2 The equation z2 + 1 = 0

As you know the above equation has solutions i. (Since the variable
is complex we shall call it z.) For this equation, the N-R formula above
becomes
zk+1 zk-1
zk+1 = zk - 2zk 24
In fact, let us write N(z) = Z2z 1, so Zk+1 = N(Zk). Starting with a
guess z (which had better not be 0!), the successive approximations to
a root are N(z),N2(z) = N(N(z)),N3(z) = N(N2(z)), etc.
Note that we have to be careful in this project about distinguishing
iteration, that is repeating the same function, as with N above, and
raising to a power. If we want to write down say the fourth power of
N(z) then we shall write (N(z))4, using brackets and the position of the
4 to remove ambiguity.
To begin with, here is some theory. You should write it out, filling in
the details. We want to determine which root, if any, is found by taking
the limit of numbers Nk(z) as k - oo.
(i) Define T(z) by T(z) = z+? (this is of course connected with the fact
that the roots of z2 + 1 = 0 are i and -i). Verify by substituting N(z)
into T that T(N(z)) = (T(z))2; repeated application of this shows that
T(N2(z)) = (T(N(z)))2 = (T(z))4.
What will be the general result here? T(Nk(z)) will be what power of
T(z)? (Hint: it is not the power 2k. Try working out T(N3(z)) for a
start.)
(ii) Why is IT(z) J = 1 if and only if z is the same distance from i
and from -i ? (Recall that Iz - al is the distance between the complex
numbers z and a.) Let L be the set of points z with this property. What
is L? Draw it on an Argand diagram of the complex numbers. It follows
that points on one side of L satisfy T(z)J > 1 and points on the other
side satisfy JT(z) I < 1. Which side of L is which?
(iii) Assume that IT(z)I < 1. Why does it follow that (T(z))r - 0 as
r - oo? (Hint: I(T(z))rl = T(z)1r.) Deduce from the above that, as
k , oo, we have T(Nk(z)) - 0. Why does it follow that Nk(z) -+ i?
166 Newton-Raphson Iteration and Fractals
What is the corresponding result when JT(z)I > 1 ? (Hint: Let
1/T(z) = U(z) so that IU(z)I < 1, then apply the same argument (check
U(N(z)) = (U(z))2), noting that U-1(0) = -i.)
(iv) Now consider the two regions of the plane:
R-={z:Nk(z)-*-Tas k-moo}.
Draw a diagram to illustrate these regions, the line L and the roots i and
-i. We call R+ the basin of attraction for the root +i, and similarly
R_ is the basin of attraction for the root -i.
Show that if z is on the set L (the common boundary of the two
regions R+ and R_), then Nk(z) stays on L for all values of k. (This
is easy once you identify what L is.) So in this case iteration does not
produce a root at all.

13.3 General quadratic equations

(i) The M-file cnrl. m (cur is complex Newton-Raphson) takes any two
complex numbers p, q and finds the quadratic equation
z2+az+b=0 (13.2)

with these as roots. Thus, as is well known, a = -p - q, b = pq. The

Newton-Raphson formula (13.1) for this case replaces the current `guess'
z by
=z-z2+az+b_z2-b
Nz
O 2x+a 2z+a*
(13.3)

The M-file chooses a random starting point z and joins up the successive
complex numbers N(z), N2 (z), N3 (Z).... until they come very close to
one of the roots p, q (marked with large crosses, one red and one green).
So the zig-zag line produced shows how the method gradually approx-
imates to a root of the equation. To choose another random starting
point, press <Enter>. The M-file allows ten such starting points.
Try running the M-file and inputting i, -i for the roots. You need to
specify the square region of the Argand diagram shown on the screen,
namely xl < x < xu, yl < y < yu, by giving in succession xl, xu, yl (yu
is then calculated). Observe which iterations tend to i and which to -i;
this should agree with the theory above!
Try running the M-file with, for example, 1 + 2i, 3 + 4i for the roots,
and a suitable region, for example, 0 < x < 5, 0 < y < 5. Do you have
 13.3 General quadratic equations 167

any conjecture about which starting points end up at 1 + 2i and which

at 3 + 40
(ii) The M-file cnr2.m takes 500 random starting points and works
out what happens to them under repeated application of the Newton-
Raphson formula applied to the same quadratic equation z2 + az + b = 0
with roots p, q. The roots are marked with a red cross for p and a
green cross for q. Those starting points which tend to the first root
(p) are plotted as small red circles and those which tend to the second
root (q) are plotted as small green circles. Try the same examples as in
Question (i) above, and some others. Write down the values of p and
q which you use, and the limits of x and y on the screen, and sketch
the line L separating the two `basins of attraction' for each example
you consider. (The basin of attraction for p is the area of the plane
containing all those starting points z which give approximations tending
to p.)
Can you guess what the line L separating the two `basins of attraction'
is for a general quadratic equation? Does this support your guess in
Question (i)?
(iii) Guided by the following hints, find the dividing line between basins
of attraction for a general quadratic equation with distinct roots p, q by
a theoretical argument.
The Newton-Raphson map is still defined by equation (13.3). Let T
be the transformation
z P,
T(z) =

where as before p and q are the roots of the quadratic equation (13.2).
Let L be the set of points z in the plane for which JT(z) J = 1, i.e.
Iz -p1 = Iz - qJ. Draw a diagram to illustrate how L is related top and
q.

Verify that T(N(z)) = (T(z))2, just as in 13.2. It follows, as there,

that T(Nc(z)) = what power of T(z)? The argument now splits into
two cases according to the side of L on which the starting point z lies.
These two sides are distinguished by T(z)J < 1 and T(z)J > 1. Use the
same idea as in 13.2 to show that if IT(z)I < 1, then T(N'(z)) ---+ 0 as
k -> oo and deduce that Nk(z) __+ p. Similarly if JT(z)J > 1, then (for
example, by using U = 1/T as in 13.2) Nk(z) --> q. So L is the dividing
line.
Can you show that, if z lies on L, then all Nk(z) also lie on L?
What happens if p = q? First run curl. m and cnr2. m for a few
168 Newton-Raphson Iteration and Fractals
choices of equal roots and observe what happens. Formulate a general
result and see if you can prove it. (Hint: Try T(z) = z - p.)
(iv) Returning to the case p q, what happens when we take our initial
guess z on the dividing line L? The case z2 + 1 = 0 is typical, and here
L is, as you should realize by now, the x-axis ! In Question (iv) of 13.2
you showed that if z lies on L for this case, then N(z) does too: the
iterates N(z), N2(z), N3(z), ... all lie on L.
Illustrate this by amending cnr i . m, calling the result cnr3. m, so that
p = i, q = -i, the screen depicts -10 < x < 10, -10 < y < 10,
and a single random starting point z on the line L is chosen. A white
circle should be placed at this point and at each of the 100 subsequently
calculated points by
plot(real(z),imag(z),'wo')
You should also store the values z (which will actually be real in this
example as they are all on L) in a vector xvalue. You do this by
initialising
xvalue=[];
and then after each new z is calculated using
xvalue=[xvalue, real(z)];
After the end of the calculation, plot these values against the count
number by
plot([0:100],xvalue)
Print out this plot. Does is suggest to you that the values produced
by the Newton-Raphson method do not converge to anything when the
starting value is on L? Give a reason for your answer.

13.4 The cubic equation z3 - z = 0

Now we shall look at the cubic equation z3 - z = 0, which has roots
p = -1, q = 0 and r = 1. Verify that the Newton-Raphson formula is
now
N(z) = 2z3
3z2-1'
The idea is to find the basins of attraction, that is, the regions of the
plane defined by
R_1={z:N'(z)->-1ask--+ oo}
and similarly for Ro and R1 . You will see that this is rather a tall order.
 13.4 The cubic equation z3 - z = 0 169

Fig. 13.1. A close up view of part of the basin of attraction for the root r = 1
of the cubic equation z3 - z = 0.

(i) Amend cnr2. m so that it applies to this example, plotting circles

in three colours: a circle of a particular colour is plotted at the point
(x, y) if with starting value z = x + iy the N-R iteration converges to
a particular root. You should take 1000 random points as the starting
points. Call the resulting M-file cnr4. m.

(ii) Run the program with xl = -2, xu = 2, yl = -2 (so x and y go

from -2 to 2). Make a rough sketch of the three basins of attraction.
Now run it for xl = -0.6, xu = -0.4, yl = -0.1, that is concentrating
on a small portion of the earlier picture. (Mark this portion roughly on
your first sketch.) Make a rough sketch of the three basins of attraction
again.
The basins of attraction in the case of this very innocent-looking cubic
equation are in fact infinitely complicated, in the sense that no matter
how much you magnify them they still look complicated (like the Man-
delbrot set). Figure 13.1 is obtained from a program just like cnr4.m
except that it scans the screen pixel-by-pixel instead of taking a large
number of randomly scattered points. It shows the basin of attraction
for the root r = 1, with xl = -0.6, xu = -0.4, yl = -0.1. Each of the
170 Newton-Raphson Iteration and Fractals
little excrescences round the main blob is infinitely complicated. The
basin of attraction is a fractal set: however much you magnify a small
portion, it never looks any simpler.
As you can see, cubics are very different from quadratics!
 14
Permutations

In this chapter there are two projects. The first one (A) is about ran-
dom permutations of a finite set, cycles and permutation matrices. The
second (B) is an investigation of card shuffling, introducing many of the
standard ideas including perfect and approximate riffle shuffles.

A Cycle decompositions
Aims of the project
We shall use MATLAB to investigate `random' permutations, especially
their disjoint cycle decompositions. There is a theoretical and experi-
mental investigation of the [5 [5 average number of disjoint cycles occur-
ring in a random permutation. The basic material on permutations is
generally covered in a first course on abstract algebra; see, for example,
[11].

Mathematical ideas used

This investigation studies permutations of a finite set, decompositions
into disjoint cycles, the order of a permutation and permutation matri-
ces. The order of a permutation involves the idea of the least common
multiple (1cm) of a set of integers. Also the average number of disjoint
cycles occurring in permutations of a given finite set is investigated both
experimentally and theoretically. Note: We always write composition of
permutations from right to left: the notation 7r2ir1 means `do 7r1 first
and then do 7r2'.
MATLAB techniques used
A given M-file produces `random' permutations of the consecutive num-
bers 1, 2, ... , n, and another breaks a permutation up into disjoint cycles.
You need to amend these to calculate the lengths of cycles and the least

171
172 Permutations
common multiple of the lengths. Also you need to count the total num-
ber of disjoint cycles in a permutation and to average this over a large
number of trials.
A permutation of 1, 2, ... , n is a rearrangement of these numbers in
some order. For example, when n = 8, the permutation
'r _ ( 1 2 3 4 5 6 7 8 ) (14.1)
7 3 8 6 5 4 1 2)
takes 1 to 7, 2 to 3, 3 to 8 and so on. The rearrangement is simply
written in the bottom row, which we shall call the vector p. Any per-
mutation can be written in disjoint cycle notation, . For it this gives
(17)(238)(46)(5), which means that it can be effected by leaving 5 alone;
taking 4 to 6 and 6 to 4; taking 2 to 3, 3 to 8 and 8 to 2; and taking 1 to
7 and 7 to 1. The cycle (46) is called a 2-cycle or transposition, (238) is
a 3-cycle, etc. Often the `1-cycle' (5) is omitted from the notation, but
here we shall include it. We shall be interested in (among other things)
the total number of disjoint cycles in the expression of a permutation.
In the case of it, this number is 4. (It is reasonably clear and true!-
that the disjoint cycle decomposition of a permutation 7r is completely
determined by it, except that the cycles may be written in a different
order and each cycle may itself be cyclically permuted. For example,
the above it is also equal to (5)(382)(71)(46).)
We denote by S(n) the set of all permutations of 1, 2, ... , n. There are
n! of these permutations, since there are n choices for what 1 is taken
to, then n - 1 choices for what 2 is taken to, etc., making
n(n - 1)(n - 2)...1 = n!

choices altogether.
(i) The available MATLAB function randperm(n) produces a 'random'
permutation of the numbers 1, 2, ... , n, that is, a random element of
S(n). Thus you type randperm(10) to get a random permutation of
1, 2, ... ,10. Take a look at this M-file. Use your knowledge of the sort
command to write down how randperm.m works. (Remember you can
type help sort if you've forgotten about sort !)
Write a short M-file which creates 1000 random permutations and
writes them all as the rows of a matrix bigp of size 1000 x n.
Hints: You will want to start with
n=input('Type the value of n')
 Permutations 173

so that you can use different values of n in different examples. Each

random permutation is then generated by
p=randperm(n);
You will need a loop,
for j = 1:1000
starting after the input line for n. If you write, within the loop,
bigp(j,:)=p;
this makes p the jth row of the matrix bigp. If you do several examples
in succession, it might be a good idea to type (in MATLAB)
>> clear bigp
between the examples, or add this at the beginning of your M-file.
Now take n = 5 and draw histograms of the columns of bigp. (Note
that
hist(v,1:5)
draws the histogram of a vector v with the numbers 1,2,3,4,5 as the cen-
tres of the `bins'.) Print out one of these histograms and state whether
they are all reasonably level, which would suggest that the numbers
1,2,3,4,5 are equally likely to appear in any position in the permutation
(and hence that randperm.m is a reasonably successful random permu-
tation generator).
(ii) The M-file cycles.m generates a random permutation and then
breaks it into disjoint cycles. It does this in exactly the same way that
you would, by following numbers through the permutation until a cycle
is formed and then going on to a new number to start another cycle.
To help the computer realise when all the numbers from 1 to n have
been `used up' in cycles, the entries in p are changed to zero as they are
used in a cycle. Try running the M-file a few times to see what output
it produces. For one example with n = 10, break the permutation into
cycles by hand, checking that you get the same answer as the computer
does.
(iii) Amend cycles. m so that it counts the orders (lengths) of the cycles
and then at the end finds the 1cm of the lengths. It is best if the lengths
are stored in a vector, say clength, which starts out empty:
clength = []
and is then augmented each time a new length is found:
clength = [clength newlength]
174 Permutations
where `newlength' stands for the length just found. To find the lcm,
proceed as follows: you can assume that 1 = lcm(xl, x2, ... , xk) can be
obtained by the sequence
l1 = lcm(xl, x2), 12 = lcm(ll, x3), ... , lk_1 = lcm(lk-2, Xk),
and the last number lk-1 calculated is the lcm 1 required. You can also
assume that for two numbers a, b
ab
lcm(a b) =
gcd(a,b)*
You can use the existing function gcdiv. m to find the gcd.
So the MATLAB commands needed to calculate the lcm of the lengths
are

ord=clength(1);
for i=2: length(clength)
ord=(ord*clength(i))/gcdiv(ord,clength(i));
end

and then of course you will want to display the order, here called ord.
Give a few examples of results from this M-file.
(iv) Amend randperm.m so that it also calculates the corresponding
permutation matrix A of the permutation 7r E S(n). This is defined as
the matrix which starts off all zeros and then adds a 1 in the ith row at
position 7r(i), for i = 1, ... , n. So you first define A = zeros(n,n) and
then have a loop
for i=1:n
A(i,p(i))=1;
end;
(Recall that p is the name of the vector giving the rearrangement, the
bottom row of equation (14.1) above.)
Why is it true that when you multiply A by the column vector u =
(1, 2, ... , n)T to form Au, you get the permutation vector p, written as
a column? Note that in MATLAB, u could be written u= [ 1 : n] ' . It fol-
lows that powers of the permutation can be calculated by A2u, A3u,....
Use your examples from Question (iii) above to illustrate the fact that
the lcm of the lengths of the cycles equals the order of a permutation
(that is the smallest power k > 1 for which Aku = u). Note: You
may find it better to produce a new version of the M-file you wrote for
the earlier part of this question, in which you are able to specify the
permutation instead of having it chosen randomly for you.
 Permutations 175

(v) Amend cycles.m to calculate also the total number of cycles (in-
cluding 1-cycles) in the disjoint cycle expression for a permutation. Then
amend it again so that it takes a given number of random permutations
(which might be 1000 in practice) and calculates the average number of
cycles which are in the disjoint cycle representations of these permuta-
tions. Run this several times with different values of n to get experi-
mental estimates for these averages for the different n.
(vi) This is a theoretical argument which predicts the average found
experimentally in Question (v) above. It involves no computing. You
should fill in the details of the arguments and answer the questions.
We calculate the average E of disjoint cycles calculated over all n!
permutations in S(n). Thus let Pk(n) be the total number of k-cycles
occurring among the disjoint cycle representations of all elements of
S(n). We need
P1(n)+P2(n)+...+Pn(n)
E
n!
For example, let n = 3. The disjoint cycle representations of the six
elements of S(3) are

(1) (2) (3), (1) (23), (2) (31), (3) (12), (123), (132).

Thus P1(3) = 6, P2(3) = 3, P3(3) = 2, and E = s is the average

number of disjoint cycles for permutations of three objects.
How many permutations in S(n) contain the particular 1-cycle (1)?
The remaining n -1 numbers can be permuted to anything, so there are
(n-1)! such permutations. Similarly f o r the 1-cycles (2), ... , (n). So for
each of the n possible 1-cycles there are exactly (n - 1)! permutations
containing that 1-cycle. So there are n(n - 1)! = n! 1-cycles occur-
ring somewhere among the complete list of permutations, expressed in
disjoint cycle form, in S(n). Thus Pi(n) = n!.
Now consider 2-cycles. Any particular 2-cycle, such as (12), occurs
in exactly (n - 2)! permutations in S(n), and there are n(2 1) possible
2 -cycles (why?). So P2(n) = z .
Similarly show that P3 (n) = 3 . (To count the number of possible
3-cycles, first count the number of ordered triples of distinct numbers,
(a, b, c), all between 1 and n. Then use the fact that any particular
3-cycle comes from three of these, say (a, b, c), (b, c, a), (c, a, b), so divide
by 3.)
What is the general result for Pk(n)? What is the value of E? Com-
176 Permutations
pare this average with those calculated experimentally in Question (v)
above.
(vii) Finally, consider the following slightly bizarre calculation. We are
given n and a second number m, with 1 < m < n. The disjoint cycles
of any particular permutation 7r which contain the numbers 1, 2, ... , m
might or might not actually contain all of 1, 2, ... , n. For example, with
n = 4, m = 2 consider the permutations
7F1 _ 1 2 3 4 _ 1 2 3 4
2 1 4 3 ' 72 = (1 3 4 2

For 71 = (12)(34), the cycles containing 1,2 do not also contain 3,4. For
r2 = (1)(234), they do.
Amend cycles. m to calculate the proportion of a random selection
of say 1000 permutations which do have the above property, for a given
m (input at the beginning, with n). Hint: This is a great deal easier
than it sounds. You need to input m and the number of permutations to
test, say testnum permutations. You will also need a counting variable,
say count, to count the number of permutations which do have the
property above. The only other essential change that needs to be made
is to change
while i<n
to

while i<m
Once the cycle decomposition has been found, you need to check whether
it has the required property; this is done by
if p == zeros(size(p))
count=count+1;
end
Can you see why this does the trick?
Taking say n = 8 and various in, conjecture a formula for this pro-
portion in terms of n and m.
(viii) The calculation in Question (vii) can be interpreted, as follows.
We have n boxes, each with its unique key which fits no other box, and
each with a small slot in it through which you can push a key but can't
get it back! The boxes are locked, and the keys are dropped at random
into the boxes. Now boxes 1, 2, ... , m are broken open. What is the
probability that every box can now be unlocked? If you can explain the
connection of this with the calculation above, please do so.
 14.1 Introduction 177

B Card shuffling
Aims of the project
Card shuffling is a situation in which permutations occur in (some peo-
ple's) everyday life. In this project we shall study both `perfect' shuffles
in which there is an extreme degree of regularity, and also `rough' shuf-
fles where there is a certain amount of chance, modelling an ordinary
person's (as opposed to a card magician's) riffle shuffle of a pack of 52
cards. We shall find, surprisingly, that three (rough or perfect) riffle
shuffles of a pack are by no means enough to randomise its ordering. In
fact if three such shuffles are performed and then a single card is moved
to another place in the pack, it is usually possible to discover the identity
of that card by a simple technique of laying out the cards in columns.
There is a discussion of card shuffling in [6].
Mathematical ideas used
This project involves permutations of a finite set, thought of as a pack of
cards numbered 1, 2, ... , n. So it is necessary to multiply permutations
(that is, do one after the other). We use the `congruence' notation:
a - b mod m, where a, b, m are integers and m 54 0, means that a - b =
Am for some integer A, which can be > 0, < 0 or zero. The order of
a permutation obtained by shuffling a pack of cards is investigated-
this means the number of times a particular shuffle has to be exactly
repeated so as to return the cards to their original ordering. (There is
an unfortunate clash of terminology here between `order' and 'ordering'!
The latter refers to the sequence of card values top to bottom of the
pack.) There is use of the disjoint cycle representation of a permutation.
Note: We always write composition of permutations from right to left:
the notation 7r2ir1 means `do 7r1 first and then do 1r2'
MATLAB techniques used
Some amendments to existing M-files are required, adapting from odd
numbers of cards to even numbers for example. The MATLAB command
sort will come in handy.

14.1 Introduction
14.1.1 Position permutations (pps)
Consider a pile of six cards, with face values 1, 2, 3, 4, 5, 6 top to bottom.
Suppose these are rearranged (shuffled) so that the face values become
178 Permutations
4, 1, 5, 2, 6, 3 top to bottom. The corresponding position permutation
(pp) is
2 3 4 5 6
Ir= C1

2 4 6 1 3 5 '

where 7r(x) = y (indicated by x in the top row and y directly underneath)

means that the card initially in position x from the top of the pack moves
to position y from the top of the pack.
Most shuffles are best described by their pps. For example the above
is a `riffle shuffle' where the cards are first divided into two equal piles
which have face values 4, 5, 6 top to bottom and 1, 2, 3 top to bottom.
These two piles are then interleaved. Doing the same riffle shuffle again
the piles have face values 2, 6, 3 and 4, 1, 5 and the final ordering of
cards has face values 2, 4, 6, 1, 3, 5 top to bottom. You can check for
yourself that the pp corresponding to this double shuffle is simply the
square of 7r (do 7r and then do 7r again):
_ 1 2 3 4 5 6
'r2 4 1 5 2 6 3

We could now `cut' the pack by making it into two piles 2, 4, 6 and
1, 3, 5 and reassembling so that the ordering now has face values 1, 3,
5, 2, 4, 6. The whole sequence of three shuffles (two riffles and a cut)
therefore has pp equal to
1 2 3 4 5 6
(14.2)
1 4 2 5 3 6

Note that the cut itself has pp equal to

_ 1 2 3 4 5 6
4 5 6 1 2 3

(card in position 1 -+ position 4, etc.). You can easily verify that the
permutation in equation (14.2) is precisely air2. Thus we can very con-
veniently work out the pp of the combined effect of the three shuffles by
multiplying together the pps of the individual shuffles: it then it then or,
written o,7r2.
Generally:
suppose we do a sequence of shuffles, with pps 7r1i then 7r2, and so on
up to 7rr. Then the pp of the total shuffle is the product permutation
irr7rr_1 ...7f1. In particular, repeating a shuffle with pp it a total of r
times gives a shuffle with pp equal to 7rr.
 14.2 Ins and outs 179

Note that to save excessive writing we shall from now on indicate pps
by their bottom line only. The top line is always 1, 2, ... , n. Sometimes
(14.3) we shall be interested in the disjoint cycles representation of a
PP.
Of course, ordinary playing cards have suits. We shall label them with
face values 1, 2, ... , 52 instead. This is the same as ordering the suits
and ordering the cards A,2,..., 10, J,Q,K within each suit.

14.1.2 Using the MATLAB `sort' command

MATLAB has an automatic way of converting from position permuta-
tions to face values, namely the sort command. Suppose we start with
six cards in the natural ordering 1, 2, 3, 4, 5, 6. Try the following in
MATLAB.
p=[2 6 4 1 5 3]; % Or any other list: this is the pp
[v q]=sort(p);
q
You will find that the vector q gives the face values of the cards left to
right (which we often think of as top to bottom in a pack). Note that
v is not used here, but it needs to be included in the sort command.
Note that, conversely, regarding q as the position permutation, p is the
vector of face values of the cards left to right: it works both ways round.
So the crucial advantage of the pp method is that by multiplying the
pps of two shuffles in the ordinary way we get the pp of the combined
shuffle of the cards. We can always convert to face-value orderings by
using sort, and this is done for you in the M-files.

14.2 Ins and outs

Take a pack of n cards labelled 1, 2, ... , n and divide it into two parts,
as nearly equal as possible. A perfect riffle shuffle is obtained by alter-
nating cards from the two half-packs. There are four cases, according to
whether n is even or odd, and according to which card goes on top.

14.2.1 Even in-shuffles

If n is even, say n = 2k, then the two half-packs contain 1, 2, ... , k and
k + 1, k + 2, ... , 2k. If card k + 1 goes on top this is called an in-shuffle
180 Permutations
4 1
4
5 2 2 5
6 3 3
6
Even in-shuffle Even out-shuffle

3 - 4
1
1
4 2
2 5
5 3

Odd in-shuffle Odd out-shuffle

Fig. 14.1. In- and out-shuffles for an even and odd number of cards, which
start out top to bottom in the order 1, 2, 3, ....

and the face-values of the cards are now

k+ 1,1,k+2,2,k+3,3,...,2k,k.
See Figure 14.1 for the case k = 3.
Write down the corresponding position permutation, 7r say, and verify
(e.g. by a couple of examples like n = 6, n = 10) that it can be written
7r(x) m 2x (mod 2k + 1),
where recall that this means ir(x) - 2x is an exact multiple of 2k + 1.j
In this case ir(x) will either equal 2x or 2x - (2k + 1).
Note that since this is the position permutation we know immediately
that doing the same shuffle again results in a pp obtained by doing 7
twice:
7'(x) = ir(ir(x)) - ir(2x) - 4x (mod 2k + 1). (14.3)

Thus here the card which starts in position x before any shuffling is
done ends up, after two shuffles, in position 4x (mod 2k + 1). Verify this
directly in the case n = 6.
What is the smallest number of times you need to repeat this shuffle,
for n = 6, to get the cards back in their original ordering?
t The only properties of = mod n we need are: if a - b and c - d, then (i)
a c = b d and (ii) ac bd. In particular if a = b, then a' = br for any integer
r>0.
 14.2 Ins and outs 181

To illustrate how to combine permutations in an M-file, the M-file

riffle1.m repeats the in-shuffle, for n even, any number of times. Study
this M-file: you will neeu to amend it later. Note in particular that it
uses sort to print out the final face-value ordering of the cards, since
that is what we want to know. Use it to find the smallest number of in-
shuffles of 14 cards which are needed to get them back into the original
ordering.
Now look at the M-file rifflela.m . This keeps multiplying 2 by
itself and reducing the result mod 2k + 1, that is, taking the remainder
on dividing the result by 2k + 1. Try k = 7; the value of r which it
produces should be the same as the number of in-shuffles of 14 cards
above.
The explanation for this is as follows. By repeating the argument
leading to equation (14.3) we find that the position permutation for r
repetitions of the shuffle is the permutation 7r' which has the property
that for each position x,

,7r''(x) - 2'x (mod 2k + 1).

It follows that if 2'' - 1, then every card is back in its original position!
Conversely if every card is back in its original position then in particular
the card in position 1 is back to position 1, so 2' =_ 1. So the smallest
number of repetitions of the shuffle which return all cards to their orig-
inal positions is the smallest r with 2' - 1 (mod 2k + 1). This smallest
number is called the order of the shuffle, or the order of the permuta-
tion giving the shuffle. (The terminology is of course very unfortunate!
Orderings of cards say what the face values are top to bottom of the
pack.)
Use riffle la. m to find this smallest r for all even numbers from 40
to 52 inclusive.

Your task in what follows is to imitate the above analysis in three other
cases. Because these are slightly different, some hints will be given.

14.2.2 Even out-shuffles

Here there are n = 2k cards but 1 goes on top, so the face-value ordering
of the cards after one shuffle is

1,k+1,2,k+2,3,k+3,..., k, 2k
182 Permutations
See Figure 14.1 for the case k = 3. Verify by looking at say n = 6, n = 10
that the position permutation in this case is ir, where
7r(x) =-2x-1 (mod 2k-1) for 1 <x <2k- 1,
while,7r(2k) = 2k. There is one additional complication: for x = k, the
position is 2k - 1, not 0. So in working out 7r(x) we want to take the
remainder of 2x - 1 on division by 2k - 1, but, since there is no position
0, remainder 0 means position 2k - 1.
To make this easier, a special function called remm.m has been created
for you. This equals the ordinary rem unless the remainder is 0. Thus
remm (19 ,11) gives 8,
but
remm(22,11) and remm(11,11)
give 11.
First, use the function remm.m to modify riffles .m to handle the case
of even out-shuffles. You will need to change the i-loop to go only to
2k - 1 and add store (2*k) =2*k; after the end of the i-loop but before
position-perm=store. Call your modified M-file riffle2.m and use it
to verify that for 52 cards you return to the original ordering after only
eight out-shuffles. How does this compare with the number of in-shuffles
for 52 cards which you found using rifflela.m above?
Next, fill in the details of the following sketch to show that the number
of in-shuffles required to return 2k cards to their original positions is the
smallest r for which
2r = 1 (mod 2k - 1).
Here is the sketch of an argument. Note that the card starting in po-
sition 2k stays there after any number of out-shuffles, so we exclude
x = 2k in what follows. As before, combining shuffles just amounts
to multiplying position permutations in the usual way, so two shuffles
produces a position permutation ir2, where
7r2(x)-2(2x-1)-1=4x-3 (mod 2k- 1).
By induction it follows that for r shuffles we get pp equal to 7rr, where
ir''(x) - 2rx - (2r - 1) (mod 2k - 1).
Hence, if 21 - 1, then 7rr returns every card to its original position, since
7rr(x) - x for all x. Conversely, if every card is returned to its original
position, then the card in position 2 is returned, so that 2r+' -2r+1 - 2,
 14.2 Ins and outs 183

i.e. 2' - 1. So the smallest number of in-shuffles returning every card

to its original position is the smallest r such that 2' - 1 (mod 2k - 1).
Use this and a suitable amendment to rifflela.m (call it riffle2a.m)
to find the smallest number of in-shuffles returning every card to its orig-
inal position, for every even number n = 2k from 40 to 52 inclusive.

14.2.3 Odd packs of cards

If the number of cards is odd, say 2k + 1, then we split the pack into
two, having k and k + 1 cards in the two parts. The two parts are riffled
together in such a way that the top and bottom cards of the k + 1 part
become the top and bottom cards of the combined pack (the k + 1 cards
`straddle' the k cards). There are two choices, according as the part
with k cards consists of those with face values 1, ... , k (in-shuffle) or
k + 2, ... , 2k + 1 (out-shuffle). See Figure 14.1 for the case k = 2. Verify
that the corresponding position permutations are given by respectively
ir(x) - 2x (mod 2k + 1) and 7r(x) - 2x - 1 (mod 2k + 1). (14.4)

Adapt riffle 1.m to these two cases (calling the M-files riffle3.m and
riff1e4.m) and find the smallest number of in- and out-shuffles which
return 15 cards to their original ordering. Note that you will want to
use remm.m rather than rem to avoid having a zero remainder.
Show that for both in- and out-shuffles, the smallest number of shuffles
to return the cards to their original ordering is the smallest value of r
such that 2' - 1 (mod 2k + 1). Use rifflela.m to find the smallest
number r for odd packs of cards with n = 2k + 1 from 39 to 51 inclusive.

14.2.4 Riffles and cuts for odd packs

There is a very interesting phenomenon which holds for odd packs and
which we shall investigate theoretically here. No computing is required,
but of course you are at liberty to write some illustrative M-files if you
wish!
A cut of a pack is a cyclic permutation, that is, it preserves the `cyclic'
ordering of the cards, which is what you get by writing the face-values
round a circle. For example if the pack 1, 2, 3, 4, 5 is cut below 3
the new face-value ordering is 4, 5, 1, 2, 3. If these numbers are written
round a circle then the ordering is the same as it was at the beginning. In
general, for n cards, explain by giving some examples why cutting below
184 Permutations
the card in position c below the top produces a position permutation
n-c+1,n-c+2,...,n-1,n,1,2,...,n-c-1,n-c.
Let us call this permutation 6c, so that a, (x) = x - c (mod n).
Now assume n is odd and equal to 2k + 1. According to equation
(14.4), a perfect riffle of n cards is given by the position permutation it
where ir(x) __ 2x or 2x - 1 (mod n). Show that, in both cases,
7rUc = Q2c7r.

Here we are just multiplying permutations in the usual way, so the equa-
tion states that
7r(Uc(x)) Q2c(7r(x)) (modn) for all x.

Show that both sides are congruent to 2x - 2c mod n. Explain why this
implies that when riffles are interspersed with cuts, the same effect can
be achieved by doing a sequence of riffles first and then a sequence of
cuts. Furthermore, it is pretty clear that doing any sequence of cuts one
after the other is the same as doing a single cut. Thus a sequence of
riffles and cuts has pp of the form

QIrr7rr-1 ... 71,

where o, is a cut and the 7ri are (in- or out-)shuffles.

Now think back to the smallest number of times, r that (say) an in-
shuffle of the n = 2k + 1 cards must be repeated to return the cards
to their original ordering (14.2.3). Suppose that r in-shuffles (for this
r) are interspersed with any number of cuts. It follows from the above
that the overall position permutation can be written as 0,7rr, where a is
some cut. Why does it follow that the cards are now in the same cyclic
ordering as they started?
Show that the same result holds for out-shuffles.
It is slightly harder to deduce that the same holds for any mixture of
in- and out-shuffles, so long as there are r of them. (A sketch is given in
14.5 below.) Thus riffle shuffles and cuts are a very ineffective way of
randomising an odd pack of cards: they always return to their original
cyclic ordering after a relatively small number of shuffles. The situation
for even packs, incidentally, is completely different. It can be shown
that a suitable sequence of riffles and cuts can produce any of the n!
permutations of n = 2k cards, so riffles and cuts are an effective way of
randomising even packs.
 14.4 Rough riffles (ruffles) 185

14.3 Cycles
Like any other permutation, the position permutation corresponding to
a riffle shuffle can be broken into disjoint cycles. An M-file which does
this for even in-shuffles has been written for you: it is called riffle 1 c . m.
It displays the cycles and the cycle lengths. For even numbers from 40 to
52 use riffle1c.m and the known fact that the order of a permutation
is the lcm of the lengths of its disjoint cycles to verify your results in
14.2.1. For each even number, write down the lengths of the cycles and
the lcm. Note that 52 cards have a special property: the even in-shuffle
is a single cycle. What is the next even number with this property?

14.4 Rough riffles (ruffles)

14.4.1 One ruffle
Unless you are extremely skilled, a perfect riffle shuffle is beyond your
powers. So here we consider how we might model an ordinary person's
attempt at a riffle, which we shall call a rough riffle, or ruffle for short.
See Figure 14.2. To fix ideas we shall consider an even number of cards,
n = 2k, which is initially divided into two parts, with m and 2k - m
cards, where we shall assume that k < m < k + 3: an almost perfect cut
of the cards.
The m cards at the top of the pack (when we start with a new pack
these are labelled 1, 2, ... , m) are now ruffled into the the remaining
cards. There are 2k - m + 1 `slots' into which a card from the top m
cards can go, namely the slot above card m+1 and the slots below cards
m + 1, m + 2, ... , 2k. We shall assume that the ruffle puts 0, 1, 2 or 3
cards into each one of these slots.
We try now to make a reasonable guess for the probabilities of 0, 1, 2
or 3 cards being put into a particular slot. Let pi be the probability of
i cards going into a slot, for i = 1, 2, 3, 4. Then we certainly want

Po+Pi+p2+p3=1.
The `expected' number of cards to go into a slot is pl + 2P2 + 3p3 and
there are 2k + m + 1 slots, so the `expected' number of cards used up is
(2k - m + 1)(pl + 2P2 + 3p3). Perhaps we should equate this with the
number of cards available to go into the slots, which is in. Thus we shall
try
(2k - m + 1)(pi + 2p2 + 3p3) = M.
186 Permutations

Fig. 14.2. A perfect riffle shuffle of 22 cards (left) and a rough riffle shuffle
(ruffle) of the cards (right).

Thus if P2 and p3 are given (as well as k and m) then po, pl are given by
In
pi = 2k - m + 1 - 2p2 - 3p3, po = l -PI -P2 -P3- (14.5)

Of course, we need to choose p2, p3 so that po, pi are positive and < 1.
Replacing the fraction in equations (14.5) by 1 (which is a reasonable
approximation when k is large) show that we need only choose P2,P3
with 2P2 + 3p3 < 1. For large k this will be a guide to choosing p2, p3.
What do you think are reasonable values for P2, P3? For n = 2k = 52,
an ordinary pack, what does this give for po, pl, taking the four possible
values of m in turn?
An M-file has been written for you, called ruffle 1. m, which imple-
ments the above model of a ruffle. It is rather complicated since it has
to take care of various possibilities such as the cards 1, 2, ... , m running
out before the slots, or vice versa. If a negative probability po or pl is
detected, then the M-file simple ends.
Run the M-file ruffle 1. m several times, using what you consider rea-
sonable probabilities, and taking k = 26 (an ordinary pack). Perhaps
you can take one of these and show how the cards from one (rough) half
have been slotted into the gaps between cards of the other half, using a
suitable diagram.
 14.4 Rough riffles (ruffles) 187

The M-file ruffle 1 c . m finds the cycles for a ruffle, as in 14.3 above.
Run this three times and for each one calculate the order of the ruffle,
which is the lcm of the lengths of the cycles. This equals the number of
times this same ruffle would have to be repeated to get the cards back
to their original ordering. The only point here is that you might find
the order very different from that of a perfect riffle of 52 cards.

14..4.2 Two or more ruffles

The M-file ruffle2.m is more elaborate in that it does a sequence of
ruffles. In your experiments you will use up to four ruffles. The idea is
to discover whether 2, 3 or 4 ruffles have the effect of `randomising' the
pack. Having performed the ruffles, you are invited to select a single
card, the ith from the left after the ruffle (in the pack this would be the
ith card from the top), and move it to the jth position from the left,
where i and j should not be too close together. (Suggestion: to simplify
your experiments, stick to i = 1: the first card is moved somewhere
else.) Note down the face value of the card which you have moved. We
shall see whether it is possible to detect which card was moved; if so,
this is a sure sign that the pack has not been randomised.
After moving one card, the cards are laid out in columns in a certain
way. By running the M-file with some fairly small values for k (you might
have to be very careful about your values for P2, P3 if k is small), see if
you can discover the method used for laying out the cards in columns.
Ignore all -1s in the columns, they are merely place fillers.
With luck, the moved card will appear in a column by itself (apart
from -1's below it). Experiment with two, three or four ruffles and
report your results on whether the moved card can be detected. Try it
out with a real pack!
Here is the output from a typical run, using p2 = 0.2, p3 = 0.05 and
three ruffles. The card numbered 28 was moved to the tenth position (so
Ci, j] is [1 101), and, as you can see, it stands out like a sore thumb!
ans =

1 22 39 29 35 45 10 28 17
2 23 40 30 36 46 11 -1 18
3 24 41 31 37 47 12 -1 19
4 25 42 32 38 48 13 -1 20
5 26 43 33 -1 49 14 -1 21
188 Permutations
6 27 44 34 -1 50 15 -1 -1
7 -1 -1 -1 -1 51 16 -1 -1
8 -1 -1 -1 -1 52 -1 -1 -1
9 -1 -1 -1 -1 -1 -1 -1 -1

14.5 Appendix
This is a sketch proof of a result in 14.2.4. All permutations 7ri here
are in- or out-shuffles for n odd. We perform any sequence of in- or
out-shuffles, interspersed with cuts. As in 14.2.4, this is equivalent to
doing all the shuffles first and then a cut o,.
The idea is to prove that if we choose r with 2' - 1 mod n, then
any sequence of r in- or out-shuffles, interspersed with cuts, produces
merely a cut, that is, a cyclic permutation of the n cards. We also prove
conversely that for this to happen after r in- or out-shuffles interspersed
with cuts, we must have 2' - 1 mod n.
Now 7r1 (x) = 2x or 2x - 1, 7r27rl(x) = 4x or 2(2x - 1) or 2(2x) - 1 or
2(2x - 1) - 1, that is, 4x or 4x - 1 or 4x - 2 or 4x - 3. In general we
have
7r = 7rr7rr-1 ... 7r1(x) - 2rx + 1,

where 1 can be any of 0, 1, 2, ... , 2' - 1. Here and below, - means

mod n.
Suppose 2' = 1. Then 7r(x) - x + 1 for all x and a fixed 1, and this is
a cyclic permutation, that is, a cut. Hence o,7r is also a cut.
Conversely suppose that our is cyclic, so that 7r is cyclic. Then
2'x+1-x+1'
for all x and some fixed 1,1'. Write 1' - 1 = m, so that 2rx x + m.
For x = 1 this gives 2' - 1 + m and for r = 2 it gives 2r+1 2 + m.
Subtracting these gives 2r - 1.
 15
Iterations for Nonlinear Equations

One can find a lot of mathematical examples involving iterations and

the solution of nonlinear equations. Such equations arise from solving
nonlinear problems - either differential or statistical. Real world prob-
lems are often nonlinear and may involve more than one independent
variable, although techniques may resemble or reduce to the one dimen-
sional case. The main reason for iterations is that direct (i.e. analytical)
solutions of nonlinear equations are in general difficult to find, and nu-
merical solutions need more than one step to converge. This chapter
mainly considers equations with real coefficients; see Chapter 13 for the
complex case.

Aims of the project

The purpose of this investigation is to study if, when and how numerical
methods work in the context of solving nonlinear equations. In particu-
lar, the important issues of accuracy and convergence speed of iterative
methods are considered.

Mathematical ideas used

This project involves vectors and matrices. The method is based on
iterations and linearisations of nonlinear equations. We first consider
one equation in one unknown (prefixed by 1D for simplicity), and then
consider systems of equations in multiple unknowns (prefixed by 2D).
Two M-files full-new.m and gauss-ja.m together with fsate.m are
used throughout this chapter.
MATLAB techniques used
This project is about solving equations. Two M-files full-new.m and
gauss-ja.m will be used for such a purpose. To use them, you need
to supply the equation information by a simple M-file. To analyse the

189
190 Iterations for Nonlinear Equations
convergence speed, use the M-file f -rate. m and a visualisation M-file
cont4.m is given as an example.
Throughout this chapter, superscripts are used for sequences while sub-
scripts are reserved for coordinates in the case of multiple variables.

15.1 1D: Method 1 - Newton-Raphson

Write a single equation as
F(x) = 0.
The Newton-Raphson method takes the form:
x(n+1) = x(") - F(x(n))/F'(x(n)),
where F' is the first derivative of function F. Given an initial guess
x = x() for the root, we hope that the sequence of iterates is getting
closer and closer to the true solution, that is, x(n) - x = x* as n -p oo.
For your convenience, we have developed an M-file full-new.m imple-
menting the Newton method. To illustrate, consider the solution of this
specific example:
Example 1
F(x) = x3 - 1Ox2 + 27x - 18 = 0, (15.1)

given that the root is approximately near x() = 7.4 (note: the exact
solution is x* = 6). As the first derivative F'(x) = 3x2 - 20x + 27, we
can prepare a file (say) f_exln.m which contains the followingt

function [F,J] = fun-name ( P ; % Example 1

x = P(1) ; % (Newton-Raphson)
F = [x"3-10*x"2+27*x-18]; % The equation
J = [3*x-2 - 20*x + 27 ]; % The Jacobian

Then the MATLAB command is simply

>> Root = full_new('f_exln', 7.4) % Form A

>> Root = full_new('f_exln', 7.4, tni) % Form B
>> Root = full new('f_exln', 7.4, tni, tol) % Form C
>> Root = full_new('f_exln', 7.4, tni, tol, hi)% Form D

t For users of MATLAB V3.5, it should be noted that the second line x = P(1)
in the M-file f_exln.m must be modified to x = P, and the same modification is
needed in f_exlg.m later.
 15.3 1D: Convergence analysis 191

where tni is the total number of iterations requested (say 12) starting
from x() = 7.4, tol is the stopping tolerance for IF(x)l < tol (say
1.0E-4) and hi specifies if intermediate iterates are shown (hi = 1 for
`yes', hi = 0 for `no'). With Forms A-C, only the final iterate is displayed
(i.e. hi = 0) and variables tni and tol, if not specified, take the default
values of 20 and 1.0E-3 respectively.

15.2 1D: Method 2 - Gauss-Jacobi

The Gauss-Jacobi method rearranges a given equation F(x) = 0 as
x = G(x) and then forms the iteration
x(n+1) = G(x( ) ) , for n = 0,1, 2, ... .
This method has been coded into the M-file gauss-ja.m. Its usage is
similar to full-mew.m differing only in preparation of a function M-file.
Thus Example 1 above can be solved by (similar to full-new.m)
>> Root = gauss_ja('f_exlg', 7.4) % Form A
>> Root = gauss_ja('f_exlg', 7.4, tni, tol, hi) % Form D

where a shorter function M-file f_exlg.m contains

function G = fun-name( P ); % Example 1
x = P(1); % (Gauss-Jacobi)
R = 18 - x"3 + 10*x-2;
G = R / 27; % The following is also ok
R = 10*x"2-27*x+18; % Intermediate quantity
% G = sign(R)*abs(R)"(1/3); % use SIGN to avoid complex
where line one contains no [ ]. In fact, you should find that the above
choice of G gives rise to poor convergence.
Note G(x) is not unique. We can even view the Newton method as a
Gauss-Jacobi method with the particular choice G(x) = x-F(x)/F'(x).
For the above Example 1, F(x) = x3 - 10x2 + 27x - 18 = 0 has been
rewritten as x = G(x) with G(x) _ (18 + 10x2 - x3)/27, but this G(x)
is not unique as G(x) = (10x2 - 27x + 18) 13 is another possibility.

15.3 1D: Convergence analysis

Neither method 1 nor 2 works well all the time. When a method does
not work, there are two possible reasons: (1) the method is no good and
in this case we have to try something else; (2) the initial guess is too far
192 Iterations for Nonlinear Equations
away from a true root and this problem can be cured to some extent.
When a method does not work well, it usually means slow convergence
- a situation similar to (1) above - and is a sign of an inferior method.
Below we hope to demonstrate these problems with some analysis.

15.3.1 Search region

By `search region', we mean the region that contains the required root(s).
R = (-oo, +oo) can be a rearch region but we hope to do better than
being content with this large region. As most equations come from prac-
tical problems, we may choose a `search region' based on a priori knowl-
edge; for example [1, 10] can be a good search region when x represents
the height of people in feet.
In the present one-variable case, i.e. F(x) = 0, we may get some
idea of the roots of F(x) in an interval [a, b] by using the MATLAB
command plot. Consider
Example 2
F(x) = (2 - sin x)(x3 - 10x2 + 27x - 18) = 0. (15.2)

For this example, try

>> x = 0.5 : 0.01 : 6.5;
>> F = (2-sin(x)) .* (x.'3 - 10*x.-2 + 27*x - 18)
>> plot(x,F, x,zeros(size(x)) ) ; grid

to get a graph like Figure 15.1. Then it can be observed that all roots
are in (say) [0, 7] and so [0, 7] is a good search region for the initial guess
x=x(o).

15.3.2 Convergence region

The convergence region for a method is defined to be the collection of
all initial points of a search region from which a convergent solution can
be obtained. Obviously a convergence region is embedded in the search
region. To find such a convergence region, we have to start from `all'
points of a search region and record those successful initial points. In
practice, consider only those points in some step-length.
For instance, with the search region [0, 8] for Example 1, we may start
in turn at x() = 0, 1/8, 2/4, ..., 8 - 1/8, 8; here a step-length
of Ax = 1/8 = 0.125 is used. In MATLAB notation, this means that
 15.3 1D: Convergence analysis 193

Fig. 15.1. Decide on a search region.

6 0 0

1 0000000000000000 000

1 2 3 4 5 0 7 8

Fig. 15.2. Find a convergence region of three intervals.

xO = 0 : 0. 125 : 8. Furthermore, we can prepare an M-file as follows

to generate a graph like Figure 15.2:

tni = 30 ; X = [] ; R = [] ;
for xO = 0 : 1/8
8 %%%%% f_exln.m for Example 1 %%%%%
:

root = full_new('f_exin', x0, tni) ;

194 Iterations for Nonlinear Equations

1 2 3 4 5 6 7 8 9
[0 1.875] for 1 [2.215 4.25] for 3 [4.875 8] for 6

Fig. 15.3. Example 2: convergence region for roots x = 1, 3, 6.

if 0 <= root <= 8, X = [X; x0] ; R = [R; root] ; end

end %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
Convergence-region = [X R], plot(X, R, 'o')
axis([O 8 0 7])

Apart from a few `odd' points, there are essentially three intervals which
make up a convergence region. So Figure 15.2 can be presented better
in the form of Figure 15.3 which is produced by
x = [0 0 1.875 1.8751; y =[O 1 1 01;
plot(x,y); hold on %% plot(x,y,'o'); % Optional
plot([2.215 2.215 4.25 4.25], [0 3 3 0]);
plot([4.875 4.875 8 8], [0 6 6 0]);
axis ([-1 9 -1 81); hold off

15.3.3 Convergence order

Different iterative schemes converge to the solution at different rates or
speeds, assuming that they converge. These rates provide a measure of
the performance of each method. The rate of convergence of a method
is characterised by its convergence order. If a sequence x('), x(i), .. .
x(n),... satisfies Ix(n+l) - xI a IX(n) - xl k, that is,
I x(n+1) - xI = CI x(n) - xI k, n = 0, 1, 2, .. .
 15.3 1D: Convergence analysis 195

that is,
Ix(l) - xl CIx() -xlk
x(2) _X1 CIx(l) -Xlk
Ix(n+l) - x1 = Clx(n) - xlk

for some constants C and k (independent of n), then the method that
produces this sequence of iterates is of convergence order k. Here x is
a solution of F(x) = 0, a means `proportional' and I denotes the
I

absolute value (modulus).

The higher the value of k, the faster is the convergence and the better
is the numerical method being used. To estimate constant C and order
k, we need at least three iterates and the exact root, or four iterates
with the last one used as the `exact' root.

The M-file: f_rate.m

If the complete sequence is more than four iterates, we may use the
M-file f_rate.m to calculate estimates of its convergence order k. As
every three iterates plus the very last one produce an estimate, the M-
file f_rate.m offers a choice of printing all estimates or just an average.
The syntax is simply:
>> k = f_rate (root, hi)

where root is the sequence of iterates that are obtained from full-new. m
or gauss_ga.m, and hi = 1 asks for all estimates to be recorded (while
hi = 0 only requires the averaged and final k estimate to be recorded).
For Example 1, we can call f -rate. m as follows:
>> root = full_new('f_exln', 7.4, 20, 0.0001, 1)
>> k = f_rate(root, 1)
>> mean(k)

where mean finds an average value. As a final example, one may explicitly
specify root
>> root = [ 3.46 3.5 3.61 3.735 3.755]'
>> K = f_rate(root, 1)
K = 3.8743 3.5091 %% Order estimates
>> A = mean(K)
A = 3.6917 %% Average
196 Iterations for Nonlinear Equations
15.4 2D: Iterations for nonlinear systems
In the remaining part of this chapter, we study iterative methods for
systems of equations. The two M-files fullslew.m and gauss-ja.m to-
gether with cont4.m and f_rate.m will be used.
Denote a system of m nonlinear equations (in vector notation) as
F(x) = 0, where
F = (F1iF2,...,Fm)T and X= (xl,x2,...,x,m,)T.
Denote the Jacobian matrix of F by J, that is, the matrix of all first
derivatives with J2j = a . The essential tool we use is the Taylor
theorem that is usually covered in a calculus course.
Given a vector function F(x) with m = 2, i.e. x E R2, each compo-
nent of F(x) can be expanded in the Taylor series, that is,

Fi(xl X2) = Fl(xi), x2l) + (XI - xii)OFl + (x2 - xzOFl + hot

ax 8x
F2(xl, X2) = F2(xii, xzl) + (x1 - xii) + (x2 - xzi + hot,

that is,
F(x(1)) = F(x()) + J J. [x(1) - x(0)] + hot,
where hot stands for `higher order terms' and J is the Jacobian matrix
evaluated at x = x()
dF1 dF1
a21 dx2
J=
8F2 9F2
8x1 a22
Further the famous Newton-Raphson method can be derived.

15.4.1 2D: Method 1 - Newton-Raphson

In the general case of m dimensions, i.e. x E R', the Taylor theorem
gives a linear approximation to F(x) in the vicinity of any point x().
If this point x() is an initial guess close to the true solution x = x*,
hoping that the linear approximation is close to F(x), we may ignore the
higher order terms and solve for the `root' of this linear equation giving
x(l) = x() -J 1(x())F(x()),
where the Jacobian matrix J is evaluated at x = x(). Repeating the
above process, we obtain the Newton-Raphson method
x(n+1) = X(n) - J 1(x(n))F(x(" )) , n = 0,1, 2, .. .
 15.4 2D: Iterations for nonlinear systems 197

where the Jacobian matrix J is evaluated at x = x(n).

As with 1D, we hope the sequence of iterates is getting closer and
closer to the true solution as n - oo. Here don't be put off by the
matrix J, as it only plays the role of F'(x(n)) in the single variable case;
so J-i(x(n)) corresponds to 1/F(x(n)).
Now consider the solution of the following two equations, that is,
F(x) = (Fl, F2 )T = 0 with the unknown x = (x, y)T
Example 3
Fi(x,y) = sin x - y3 - 8 = 0,
(15.3)
{ F2(x,y)=x+y-3=0.
The Jacobian matrix is
QFl QEl \ cosx -3 * y 2
a /
/1I

J=(Jj)= ( aF2
OX
ax
aF2
ay
= 1 1

Again given an initial guess for the root(s), say x() = ( 4.7, -1.9 )T,
we can use the Newton-Raphson iterations to find the exact solution
x = (5.075, -2.075)T.
The iteration formula can be written as x(n+l) = x(n) - d(n), where
d(n) is the solution of the linear equations J(x(n))d(n) = F(x(n)). The
rewritten iteration formula is more efficient as it avoids using, or calcu-
lating, the inverse of matrix J (see Chapter 16). Of course for the single
variable case, there is no difference; refer to 15.1. As we know, the
MATLAB command for finding d(n) is simply d = J \ F. This is how we
have implemented the Newton method in the following M-file.

The M-file: fullnew.m

To solve Example 3, prepare the following function M-file f_ex2n.m
function [ F, J ] = anyname (P ); % Ex 3 - multiple
x = P(1); y = P(2); equations
F = [ sin(x)-y"3-8; Equation-1
x+y-3 1; Equation-2
J = [ cos(x), -3*y"2; Jacobian 2 x 2
1, 11; matrix

(Note To solve three or more equations, the function M-file can be pre-
pared similarly.) The MATLAB command is simply
>> Root=full_new('f_ex2n', [4.7 -1.9]') % A
198 Iterations for Nonlinear Equations
>> Root=full_new('f_ex2n', [4.7 -1.9]', tni) % B
>> Root=full_new('f_ex2n', [4.7 -1.9]', tni,tol) % C
>> Root=full_new('f_ex2n', [4.7 -1.9]', tni,tol, hi) % D

where tni is the total number of iterations requested (say 12) starting
from x() = [4.7 - 1.9]T, tol is the stopping tolerance for IF1 < tol
(say 1.0E-4) and hi specifies if intermediate iterates are shown (hi = 1
for `yes', hi = 0 for `no'). One can see that the usage is identical to the
1D case.

15.4.2 2D: Method 2 - Gauss-Jacobi

The Gauss-Jacobi iteration involves rearranging F(x) = 0 as x = G(x)
and forming the iteration
x(n+i) for n = 0,1, 2. ... .
= G(x(n))
This method, taking a variety of names in different contexts, is intu-
itively simple. However, as with 1D, the Newton-Raphson method can
be considered as a special case of Gauss-Jacobi if we choose
G(x) = x - J-'(x)F(x).
For the above Example 3, F(x) = 0 may be rewritten as x = G(x)
with

G(x) = 3-y
(sin x - 8)13

Here the two equations in (15.3) have been swapped; this does not change
the solution.

The M-file : gauss-ja.m

The M-file gauss_ja.m, implementing the Gauss-Jacobi method, can
be used similarly to full-new.m, but the main difference lies in the
preparation of function M-files. As presence of the J matrix is not
required, it is somewhat simpler.
Thus Example 3 above can be solved by (similar to using full-new.m)
>> Root = gauss_ja('f_ex2g', [4.7 -1.9]') % A
>> Root = gauss_ja('f_ex2g', [4.7 -1.9]',tni,tol,hi) % D

where the shorter function M-file f _ex2g. m contains

 15.5 2D: Contour plot and convergence history 199

52

Fig. 15.4. Contour plot of F1 (x, y) of two variables.

function G = fun-name ( P ); % Ex 3 - (2 variables)

x = P(1); y = P(2); % (Gauss-Jacobi)
G = [ 3-y; % Here G is of size 2
-(8-sin(x))"(1/3) ]; % Avoid complex roots

15.5 2D: Contour plot and convergence history

If we plot function values of F, the roots of F(x) are intersections of
all zero contour lines. This gives us some rough idea about where the
roots are and where we should start iterations from; refer to the M-file
cont4.m.
To produce a contour plot, the main preparation involves the setting
up of a uniform grid covering a suitable rectangle region. This is done
via the commandf meshgrid. Suppose that we want to draw a contour
plot of function Fl (x, y) = (x6 - y3 - 0.5) exp(-x2 - y2) in say S2 =
[-2, 2] x [-2, 2]. The procedure is the following (see Figure 15.4)
xO = -2 : 0.4 2;
yO = -2 : 0.25 2; % Two 1D vectors

t For users of MATLAB V3.5, the command is meshdom.

200 Iterations for Nonlinear Equations
Green stars for F1=O and blue circles for F2=0

Fig. 15.5. Intersecting zero contours of functions F1(x, V) and F2 (x, y).

[x y] = meshgrid (xO, yO) ; % Two matrices

F1=(x."6-y."3-0.5) .* exp(-x."2-y.'2); % F1 = matrix
M=max(F1); m=min(F1); % Row vectors
V=[m 0 M]; % Some values of F1
C1 = contour(x,y,F1, V, '-r'); % MAIN step
clabel(C1); % Optional labels
grid ; %, Optional grid

If M=max(max(F1)) and m=min(min(F1)) are used instead, M and m

are scalars and so V is of size 4. In fact, V may be set to any desirable
values within the range of Fl.
If the zero contour lines of a second function (say) F2 (x, y) = 20 *
(x2 - y2) - 5 = 0 are superimposed on those of Fl (x, y) above, using
different plotting symbols, the intersections must be the solution. This
is also done in the (second part of) M-file cont4.m; see the results in
Figure 15.5 where we can see rough locations of four roots.
Obviously a converged root (one of the four roots) of the system
Fl (x, y) = 0 and F2 (x, y) = 0 should be close to a zero contour line
(or an intersection of lines). Further if we plot all iterates on the con-
tour plot, then we should see a zigzag curve converging to the root. Try
the following for example; you may modify cont4.m
 15.5 2D: Contour plot and convergence history 201

_2L
-2 -1.5 -0.5 0 0.5 1 1.5 2

Fig. 15.6. Contour plot of F1 (x, y) = 0 and convergence history.

>> Root = [ 0.5 -1.5; 0.25 -1.7; 0.55 -0.25;

>> -0.45 -0.1; 0.1 0.55; -1.0 0.8];
>> xO=-2:0.4:2; yO=-2:0.25:2; R=Root;
>> [x y] = meshgrid (x0, y0) ;

>> F1=(x."6-y."3-0.5) .* exp(-x."2-y."2) ; % F1=matrix

>> V=[-eps 0 eps]; % Vector of almost 0's
>> C1 = contour(x,y,F1,V, '-r'); % MAIN step
>> clabel(C1); grid hold on;
;

>> plot(R(:,1),R(:,2),'-g', R(:,1),R(:,2),'ow');

>> hold off

The plot is shown in Figure 15.6, where we assume Root contains these
iterates

Root = 0.5 -1.5 % x"(0) - starting vector

0.25 -1.7
0.55 -0.25
-0.45 -0.1
0.1 0.55
-1.0 0.80 % Last iterate -- as root
202 Iterations for Nonlinear Equations
(Of course such a quantity Root would come out straightaway from
full-new.m and gauss_ja.m.)

15.5.1 Convergence order

Different iterative schemes converge to the solution at different rates or
speeds. As with 1D, we now consider the convergence order. If a vector
sequence x(), x(1), ..., x('n), ... satisfies IIx(n+1) -xli oc IIx(n) -x1Ik,
that is,
CIIx(n)
IIx(n+1) _ xii = _ xlIk,
for some constants C and k (independent of n), the method that pro-
duces this sequence is of convergence order k. Here x is a solution of
F(x) = 0, and II' I I denotes some norm (say the 2-norm; type help norm
and refer also to Chapter 16 for more details).
The higher the value of k, the faster is the convergence and the better
is the underlying numerical method. As with 1D, to estimate constant
C and order k, we use the following M-file.

The M-file : f-rate.m

If the complete sequence is more than four iterates, we may use the
M-file f -rate. m to calculate estimates of k. If root is the sequence of
iterates that are obtained from full_new.m or gauss_ga.m, type
>> k = f_rate (root, hi)

where hi = 1 asks for all estimates of k to be recorded (while hi = 0

only requires the averaged and final k estimate to be recorded in k).
For Example 3, we may call f -rate. m as follows:
>> Root=full_new('f_ex2n',[4.7 -1.9]', 30,1.0E-4, 1) % D
>> K = f_rate(Root, 1)
>> mean(K)

where mean finds an average value. As a final example, specify

>> root = [ 3.46 2.5;
3.5 2.59;
3.61 2.65;
3.735 2.69;
3.755 2.695]
 Exercises 203

and call f -rate. m as follows

>> KO = f_rate(root, 0) %% or
>> K1 = f_rate(root, 1)

to obtain the results (average = 2.8728)

KO = 2.8728 3.3453
K1 = 2.4003 3.3453

Exercises
15.1 Using full new.m and x() = 1.5, find a real root of
F(x) = ex + e-x + 2 cosx - 6.
15.2 Apply gauss-ja.m and re-solve Exercise 15.1 by selecting two
different G(x) functions to find the root in [1, 2]; e.g. x =
ln(6 - ex - 2cosx) may be a choice.
15.3 Show that ifIx(1)-xI = CIx(0)-xlk, Ix(2)-xI = CIx(1)-xIk
and all terms are assumed to be nonzero, then we have formulae
- xI
Ix(2) Ix(1) - xl.
(a) k = In Ix(1) In
- xI WO) - X1,
(b) C = Ix(2) -xl/lx(1) -xlk.
15.4 A small object, dropped from rest, encounters air resistance as
it falls. Using Newton's second law to model this situation we
obtain the following equation for the height H in metres that it
falls in t seconds:
H(t) = 12.25 [t + 1.25(e-0-8t - 1)] .

To determine the time taken to fall 6.125 m, we must solve the

equation H(t) = 6.125, that is,
F(t) = 1.25e -0-8t + t - 1.75 = 0.
Using the M-file gauss-ja.m for each of the following methods:
(a) Newton-Raphson: G(t) = t - F(t)/F'(t) (as in 15.2);
(b) Gauss-Jacobi: use G1(t) = 1.75 - 1.25e-0.8t;

(c) G auss-J acobi : use G2( t) = t - F(t)

1-(0.8t-F0.6) exp(-0.8t)
and with t() = 0.5,
(i) determine a solution correct to four decimal places;
204 Iterations for Nonlinear Equations
(ii) estimate the convergence order in (i) if possible;
(iii) find and plot the convergence region.
Hint Use the search region [-5, 5] and axis ([-5 5 -4 4]) after
plot.
15.5 Given the nonlinear function F(x) = 0.6 - (3 + In x)x + ex:
(a) Generate four random numbers x1, x2i x3 and x4 in inter-
vals (1,1.3), (1.4,1.5), (1.5,1.6) and (1.7, 2) respectively,
and evaluate F(x) at x = x1ix2,x3,X4. For each root of
F(x) in [1, 2], find an interval that contains it. Mark all
four points on a graph of the function F(x) over some
search region and verify that it has exactly two roots a
and 0 in [1, 2].
Hint Use the intermediate value theorem on sign changes
of a function.
(b) Try to use the Gauss-Jacobi method (tol = 10-5) with
G(x) = 0.6 - (2 + In x)x + ex, to compute roots a and 0
and estimate the convergence order for each case; if this
G(x) is not suitable for 0, you need to find something
that is suitable.
15.6 Using the Taylor theorem in one variable, deduce that, for a
single function H = H(x) = H(xi, x2) in two variables (m = 2),
H(x) = H(x()) + (x(1) - x()) V/H + hot,
where V/H = gradH = (ate , aL ). Use this formula to find a
linear approximation to
H(x) = ln(xlx2) at x() = (1, 1)T.
EXTRAS Can you generalise this result to the case of three
variables (m = 3) and give an example?
15.7 Consider the system F(x) = 0, given by
F1(xl,X2) = exl-x2 - sin(xi + x2) = 0,

1 F2(xi,x2) = X2 1X2 - cos(xi + x2) = 0-

(a) Using full-new.m and x(O) = (-2,1)T, find a root up to

four digits of accuracy.
(b) Using gauss_ja.m and x() _ (0.5,0)T, find a root with
no more than 40 iterations using the formula G = x -
wJ\F. Take w = 0.5, and w = 1.5.
 Exercises 205

Hint When w = 1 the method is identical to the Newton

method. Can you find a different G(x) formula that
works?
15.8 Show that if 11x(1)_xjj /jjx()-xjj' = llx(2)-xjj /11x(')-xll' = C,
and all terms are assumed to be nonzero, we have
(a) k = In 11x(1)
xll / In 11x(1)
- xll
iIxO - x 11x() - xli
11x(2)
(b) C = _ xli/llx(1) - x11k.
15.9 Following Exercise 15.7,
(a) using a grid spacing of 0.2 on the region
R={x:-2<xl <1.6, 0<x2<2}
deduce, from the contour plots of F1 and F2 on the same
graph, that there are two roots in R;
(b) for the root nearest to the point (0.5, 0.0)T, plot its New-
ton iterates on the contour plot and find the averaged
convergence rate of the iterates. Further plot on the same
graph Newton iterates starting from (-2,1)T.
15.10 Consider the system of equations F(x) = 0 given by
Fl (x) = 12x1 - 3x2 - 4x3 - 7.17 = 0,
F2(x) = x1 + 10x2 - x3 - 11.54 = 0,
F3 (x) = x2 + 7x3 - 7.631 = 0.
(a) Write down the Jacobian matrix for F and use the Newton-
Raphson method with x() = (0, 0, 0) T to obtain the
solution correct to five decimal places.
(b) Use the third equation to eliminate x3 and use Newton-
Raphson to solve the reduced system of two equations
for x1 and x2 correct to five decimal places. Evaluate x3.
(c) EXTRAS Using a grid spacing of 0.1 on the region
R={x:-5<xl <25,-15<x2<5},
obtain the contour plots of Fl and F2, based on (b), and
further decide on the total number of solutions to the
original system. Find any remaining solutions correct
to five decimal places using Newton-Raphson. Write
down a version of Gauss-Jacobi iteration formulae for
the original system and investigate the convergence of
this method to each root.
206 Iterations for Nonlinear Equations
(d) EXTRAS Other methods exist for the solution of non-
linear equations but we leave you to find out about the
MATLAB command f solve whicht is based on a method
of nonlinear least-squares. Compare its performance with
full-new. m.

t This command is usually available in the MATLAB optimization toolbox. How-

ever, if your MATLAB does not have it, get a copy from the web address given in
Appendix 3.
 16
Matrices and Solution of Linear Systems

Linear systems of algebraic equations are one of the most important sub-
jects in mathematics, since most other subjects, methods and problems
involve or reduce to this subject.

Aims of the project

As numerical methods involve truncations and finite precisions, we in-
vestigate their effect on solution accuracy. Most of us know some theory
about linear systems but may not be aware of the good or bad choice
of solution methods on computers, what determines the accuracy of the
numerical solution and whether an obtained solution can be improved.
Large scale problems arising from practice often involve sparse matrices
and special techniques can be developed. This project addresses all such
issues.

Mathematical ideas used

You have learnt that to solve Ax = b you just type x = A\b. To
investigate the sensitivity of the solution x with regard to the matrix
A (or its condition number) and the right-hand side vector b, we use a
controllable number of digits in our calculations. The use of iterative
refinements is illustrated. Finally we discuss how a sparse matrix may
be condensed towards band forms by use of permutation matrices.

MATLAB techniques used

Seven M-files lin_soly. m, chop. m, lu2. m, lu3. m, lu4 . m, solv6. m
and spar_ex.m are used to assist this project. The last two M-files are
listed in the chapter. Here chop. m, used to fix digits, is used by lu2. m,
lu3. m and lu4 . m. You will get useful experience of MATLAB's easy
and simple commands for both dense and sparse matrices.

207
208 Matrices and Solution of Linear Systems
16.1 Operation counts
Operation counts give an indication of how many floating point oper-
ations (flops) are needed for a given mathematical problem. Any re-
duction in operation counts leads to a faster solution. This project will
highlight three basic ideas to this end. The first idea, discussed below,
is to implement an existing method in a more efficient way in order to
achieve speed-up. The second idea, discussed in 16.2 and 16.3, is to
seek new and alternative methods for solving the same problem. The
third idea, discussed in 16.5, is to reformulate the given problem so
that existing methods can be efficient.
To illustrate the first idea, we consider the simple task of calculating
z = ABCb, where A, B, C are n x n matrices, and b is a vector of size
n. A naive way to do this is to work out a matrix D = ABC first before
doing z = Cb; this costs roughly 4n3 floating point operations (flops).
A better method is to compute three matrix-vector products, that is,
zl = Cb, Z2 = Bzl, z = Az2; this costs only 6n2 flops!
The efficiency of different methods is therefore crucially determined
by operation counts. We may get estimates of operation counts theoret-
ically from the mathematical formulae but such a job can be done easily
by the MATLAB command flops; to demonstrate its usage, consider
the second method in the above example
>> flops(0); n=2000; % Set flops counter to zero
>> A=rand(n); B=rand(n); % Setup matrices
>> C=rand(n);b=rand(n,1);% Setup matrices
>> z=C*b; z=B*z; z=A*z; % Operations done
>> w=flops, ratio=w/n"2 % Work out flops

where you should find w = 24000000 and ratio = 6. Note that rand
does not contribute to f lops.
Apart from flops, three other commands can be used to monitor the
time spent by MATLAB to carry out a particular calculation; these are
shown in Table 16.1, where `Method' refers to operations of a numeri-
cal method and row `E' shows results from doing the simple operation
rand(9) *5, that is, replacing `Method' by rand(9) *5. Note etime checks
the difference in clock and converts it into seconds; type help etime
to be aware of a possible problem.
Remark Broadly speaking many classical mathematical methods, al-
though elegant in theory and useful in proving the existence of solu-
tions, are not suitable for computer use. For they may either take too
 16.2 Dense linear systems 209

Table 16.1. Monitoring the efficiency of a numerical method.

flops tic/toc cputime etime

>> flops(0) >> tic >> t0=cputime; >> t0=clock;

>> `Method' >> `Method' >> `Method' >> `Method'

>> t=flops >> t=toc >> t=cputime; >> t=clock;

>> t=t-t0 >> t=etime(t,tO)
E t = 81 t = 6.09 t = 5.27 t = 3.13

long (even years) or lead to large errors; refer to 16.2.3. For example,
A-' = adj(A)/det(A) is good for a theory but not suitable for numeri-
cal computations because the number of operations required is too large!
Here adj(A) denotes the adjoint of matrix A - a matrix of cofactors.
Therefore the analysis, selection (validation) and development of prac-
tical and useful numerical methods are the main aims of computational
mathematics.

16.2 Dense linear systems

In what follows, you may issue the MATLAB command format compact
first to control output.

16.2.1 MATLAB solution of Ax = b and pivoting

For square matrices, the simple instruction x = A\b is not the same as
x = inv(A) * b because each represents a different method for solv-
ing the same problem. Convince yourself by doing the following MAT-
LAB experiment before proceeding. First generate a random matrix A
of size 8 x 8 using A=rand(8) and a column vector b of size 8 using
b=rand(8,1). Then type
>> flops(O) ; % Set flops counter to zero
>> x=A\b %Ais8x8 andbis8x 1.
>> fl = flops % Give number of operations used
>> flops(0) ; % Set counter to zero again
>> x = inv(A)*b
>> f2 = flops % Should be larger
210 Matrices and Solution of Linear Systems
Here and from now on where flops is used, it is always possible to use
the other MATLAB commands listed in Table 16.1.
The actual MATLAB method for solving Ax = b is based on the
Gauss elimination method and may be described as a three-stage ap-
proach:

(1) Compute a lower triangular matrix L, an upper triangular matrix

U and a permutation matrix P such that PA = LU. Here P is
simply the identity matrix I with some of its rows swapped.
This swapping ensures that in reducing the new matrix Al = PA
to its row echelon form, all multipliers in the elimination process
are less than or equal to one. The use of P implies that we solve
PAx = Pb or rather LUx = Pb instead of Ax = b. Such a
process is known as partial pivoting and helps to maintain the
accuracy of computations.
(2) Solve Ly = Pb, known as forward substitution.
(3) Solve Ux = y, known as backward substitution.
The first stage known as triangular factorisation (or LU decomposition),
related to the row echelon form reduction,] is the most important step. It
can be done via the command lu. Therefore the followings is equivalent
to using x = A\b
>> [L,U,P] = lu(A); % Find L, U and P matrices (A1=PA=LU)
>> B = P*b ; % A x = b becomes Al x=B or LU x=B
>> y=L\B ; %Solve Ly=B as L (Ux) =B
>> x=U\y %Solve Ux=y

16.2.2 L and U are reusable

Suppose that we hope to solve Ax = b and Ax = e (or for that matter a
number of linear systems with the same matrix but different right-hand
sides). Using the \ approach twice is not a good idea, because each time
MATLAB would do one LU factorisation of PA.
A better procedure is the following:
t The MATLAB command rref does this job but one cannot see much output
from it. Another command rrefmovie is more useful. Fortunately one is allowed
to view the sources of rref.m and rrefmovie.m. Try A=round(rand(6)*7); [R
J]=rref(A), rrefmovie(A).
t Note that y = L\B does not take as many operations as A\B because MATLAB is
clever enough to recognize the triangular nature of L! The same is true for x = U\y.
Test this out using flops.
 16.2 Dense linear systems 211

>> [LAP] = lu(A); % Find L,U,P matrices (PA=LU)

>> B = P*b; % Solve PAx=Pb (eqn 1)
>> z = L \ B; % Solve PAx=Pb
>> x = U \ z % Solve PAx=Pb
>> E = P*e; % (eqn 2) Solve PAx=Pe
>> z = L \ E; % Solve PAx=Pe
>> x = U \ z % Solve PAx=Pe

16.2.3 Significant decimal digits

In MATLAB and most other computer packages or languages, only
a finite number of digits are actually used in calculations, and so most
numbers must be rounded. This gives rise to computing errors. The
more computing steps we have, the larger the final error. This partially
explains why we should always abandon a method like x = inv(A) * b
on computers, as it not only takes much longer for a start but also may
accumulate larger errors.
All MATLAB calculations by default use 16 decimal digits. As we
want to investigate the effect of finite precision arithmetic on numerical
accuracy, we shall mostly work with t-digit arithmetic in this project
with 1 < t < 16. The M-file chop. m provides us with an easy way to do
this. Try help chop, chop(0.1364, 2) * 10 or chop(1.223345, 6)
and chop (1.223345 , 5) to see how it works.

16.2.4 The M-files: lin-solv.m, lu2.m and lu3.m

We shall explore the importance of pivoting and significant digits in the
context of solving Ax = b. The main M-file lin_solv.m is used as
follows
>> x = lin_soly( 'lu_name', A, b, t) ;

where lu name is the name of the factorisation function M-file which may
be replaced by lu2 or lu3 or lu4. Here t indicates that all calculations
are performed to the fixed precision of t-digits only.
The function M-file lu2. m is mathematically the same as the built-in
system file lu.m except that now we only perform t-digit arithmetic.
So if you input t = 16 with lu2.m, then lin_solv.m should produce
identical results to x = A\b.
The function M-file 13. m does not allow partial pivoting and is only
to be used as a comparison, not to compute accurate solutions as no
212 Matrices and Solution of Linear Systems
permutation matrix P is involved and multipliers larger than one are
permitted. The M-file lu4. m always produces the same results as lu3. m
but uses fewer flops for sparse matrices; see 16.5.
The two M-files 1u2. m and 1u3. m can be used individually for factori-
sation (similar to 1u)
>> [L,U,P]=1u2(A,t); %Factorise A (t digits & P)
>> [L,U] =1u3(A,t); %Factorise A (t digits & no P)

So if you solve Ax = b by
>> x = lin_soly( '1u2', A, b, t) ;

we may say the method is Gaussian elimination with partial pivoting.

Similarly the following is called Gaussian elimination without partial
pivoting
>> x = lin_soly( 'lu3', A, b, t) ;

16.3 The iterative refinement algorithm

If y is an approximate solution of Ax = b, define its residual as r =
b - Ay which may not be zero (or small). If we solve Ae = r, then we
can usually expect x1 = y + e to be a better approximation to the true
solution x. The procedure may be repeated if the residual for x1 is not
`small' enough and we have an iterative process. This is the basic idea
of the iterative refinement algorithmt used to improve solutions:
Step 0. Use Gauss elimination with a fixed precision t < 16 to solve
Ax = b and keep L, U, P matrices. Denote the approximate solution
by y. (Note P = I when no partial pivoting is used.)
Step 1. Compute the residual r = b - Ay (if r = 0 or rather if its
norms is small enough, we stop here and accept x = y).
Step 2. Use the same L, U, P to solve LUe = Pr.
Step 3. Update the solution y - y + e.
Step 4. Go back to Step 1 and repeat.
t We remark that iterative refinement is closely related to another idea called resid-
ual correction, where A may represent other operators (e.g. differential, integral,
nonlinear). This idea forms the basis of a modern and powerful numerical tech-
nique - the multi-level method (or multi-grid method).
t With MATLAB, if r E Rn, then its p-norm is jjrjjp = norm(r, p) with p being either
any positive integer or `inf' (i.e. for oo-norm). However, for matrices, MATLAB
command norm can only compute four norms, namely, IIAII = norm(A, p), where p
can be 1, 2, `inf' or `fro' that denotes, respectively, 1-norm, 2-norm, oo-norm
and F-norm (Frobenius). The default norm(A) gives the 2-norm.
 16.4 A perturbation analysis for Ax = b 213

The values in both r and e give indications of the accuracy of y. They

can be measured by the MATLAB command norm.

16.4 A perturbation analysis for Ax = b

In general, the accuracy of x = A\b depends on the property of matrix
A (the so-called conditioning). A measure of conditioning is by its con-
dition number which is defined by cond(A) = IIA-1II IJAII, where II' II can
be any matrix norm. However, for the condition number, MATLAB has
a simple commandt cond in the 2-norm; just type cond(A). For `good'
matrices, the digit number t does not affect accuracy much; but for `bad'
matricest (or ill conditioned ones), t is significant!
Now consider the solution accuracy when t < 16. Suppose Ax =
b denotes the true linear system involving no rounding in A and b.
However, in reality, due to rounding and possible data collection error,
our numerical solution y actually only satisfies a so-called perturbed
system (nearby)
(A + AA)y = b + Ob.
Here (A + AA) is viewed as the matrix A in t precision and similarly
for b + Ab. The error terms AA (matrix) and Ab (vector) are usually
known. For example, when we use t-digit arithmetic,
IIobII = norm(A - Al) and IIobII = norm(b - bl),
where Al = chop(A, t) and bl = chop(b, t).
The solution accuracy is a measure of how close the numerical solution
y is to the true solution x. A linear system is well conditioned if small
values of IIDAII/IIAII and IIobII/IIbII lead to a small value of IIx-yll/IIxII.
Otherwise, it is ill conditioned. Obviously we want to know how these
quantities are related. We have the following result:

Theorem 16.1 Suppose that the exact equation is Ax = b and we have

actually solved
(A+AA)y=b+Ob.
The relative error will be
t To compute cond(A) in other norms, we may use the definition involving A-' but
this can be expensive in terms of flops. For 1-norm, try condest (A).
t For a `bad' matrix, to achieve a good accuracy, the solution strategy is to convert
it into a `good' one rather than to demand an unlimited increase of t. Such a
method is called pre-conditioning.
214 Matrices and Solution of Linear Systems

lix - YII (11o&

IIxII -C IJAII + libJJ '

where the constant is given by

cond(A)
C= oA
IIIIAIIIII.

1-condA
( )

Therefore we say a problem Ax = b is well conditioned if C is not too

large and ill conditioned if C is large. For the purpose of this project,
let us call a problem ill conditioned if C > 50.
Example Solve a random 11 x 11 linear system Ax = b with t = 2
digit arithmetic; see Chapter 6 for using rand. We want to know if the
problem is well conditioned. The following M-file is also available as
solv6.m
rand('seed',1998); n=11; % To fix 'seed'
A = 10*rand(n,n) ;

b = rand(n,1); % Generate the system

t = 2 ; % Set the precision
x = A \ b ; % `Exact' solution
y = lin_soly( '1u3', A, b, t) % Numerical solution
;

A_rel=norm(A-chop(A,t))/norm(A);% Relative error in A

b_rel=norm(b-chop(b,t))/norm(b);% Relative error in b
k = cond(A) ; % Condition number
C=k/ ( 1-k*A_rel) ;
if C >= 50, disp('Problem is ill conditioned ...'),
else, disp('Problem is well conditioned.'), end
error-theory = C*(A_rel+b_rel) % Formula in Theorem
error-found = norm(x-y)/norm(x)% Exact x, numerical y

Here you should find that C = 97.561 so the problem is ill conditioned,
error-theory = 0.7618 is the predicted error by theory (larger) and
error-found=0.3381 is the actual error observed (smaller).

16.5 Sparse matrices, graph ordering and permutations

Matrices that are generated from applied mathematical problems are
often sparse; by `sparse' we mean many entries (say 50%) are actually
zeros. Therefore avoiding any operations with these zero positions can
lead to the speeding up of practical algorithms. One general approach
 16.5 Sparse matrices, graph ordering and permutations 215

is to use permutation matrices to gather all nonzeros together to form

a desirable pattern (e.g. band matrices that have zeros in all entries
except on the main diagonal and a small number of subdiagonals).
To visualise sparse matrices, we may use the MATLAB command spy,
for example, try the following (refer to M-file spar _ex . m)
>> A = diag(0:9); % Generate a matrix of size 10
>> spy(A); grid % See what it looks like
>> a = eye(10,1)*ones(1,10)*2; % Nonzero Row 1
>> b = ones(10,1)*eye(1,10)*3; % Nonzero Column 1
>> A = A + a + b; % Formed an arrow head matrix
>> spy(A); grid % See what it looks like now

To appreciate the techniques used in sparse matrices, consider the

following
>> Al = A; % Note this A follows the above
>> flops(0) % Counter reset (try also tic/toc)
>> [L1 U1]=lu3(Al);% Factorise A as a full matrix
>> Work-1 = flops % Operations performed
>> al = symrcm(Al) % Find new ordering for matrix Al
>> A2 = A1(al,al) % Reorder A2=P*Al*P'
>> flops(0) % Counter reset (try also tic/toc)
>> [L2 U2]=lu4(A2);% Factorise A as a sparse matrix
>> Work-2 = flops % Operations performed (new)

You should find that Work_1 = 8434 and Work_2 = 1506. Here the trick
is that we have found a new ordering for matrix Al; to see and use the
permutation matrix, try
>> P = eye(10); % Generate a 10 x 10 identity matrix
>> P = P(al,:) % Reorder its rows
>> A3 = P*Al*P' % Here A3 = A2 !

As we know, P is orthogonalt and thus the eigenvalues of A, A2 and

A3 are the same. Evidently it is more advantageous to work with A3 to
find eigenvalues.
Finally, we show how to represent sparse matrices by graphs! A sparse
matrix A,, can be viewed as a representation of zeros and nonzeros,
and each nonzero represents a relationship (or an interaction) between
t Check this out by verifying pTp = I. Note: Advanced users of MATLAB may
use speye, instead of eye, to form an identity matrix; try also help sparse to
find out more about building sparse matrices.
216 Matrices and Solution of Linear Systems
two numbers (i.e. a row and a column index, both belonging to the same
index set 1, 2, ..., n). In graph theory, a graph G with n nodes (not
all nodes connecting each other) may have a variable number of edges,
each edge representing a connection of two nodes.
Therefore, a sparse matrix A can be naturally linked with a graph
G(A) if we `identify' an index from the former with a node from the
latter, and also a nonzero from the former with an edge with the latter.
Obviously zero entries of A correspond to nonexistent edges of G(A)!
Thus the connectivities (edges) of n nodes of a graph can represent a
sparse matrix. More precisely, edges of a graph G(A) represent nonzero
elements of A, that is to say, node i connects node j in the graph if and
only if entry A2j 0 in the matrix.
For example, a tridiagonal matrix can represent n nodes forming an
open chain (connectivity), with each middle node only connecting its
two immediate neighbours, that is, the first end node 1 only connecting
to node 2, the middle node k only connecting to nodes k-1 and k + 1 and
the last end node n only connecting to node n - 1. To put it the other
way round, with such a G(A), node k only connects to nodes j = k - 1
and j = k + 1, and so for the underlying matrix, entry Ak : 0 only for
k - jI < 1. This implies that A must be tridiagonal!
In the above example of a 10 x 10 matrix, the original matrix A
represents the connectivities of 10 nodes in the natural ordering a =
1 : 10 and A2 represents the connectivities in a new ordering al =
[10 9 8 4 6 5 2 7 1 3].
For a symmetric matrix A, its graph G(A) can be plotted by the
MATLAB command gplot. For the above example with A10,<10, try
the following (available in the M-file spar_ex. m)

>> xy = [ 0 0
1.5 1.3
0.35 2.0
-1.0 1.7
-1.9 0.68
-1.9 -0.68
-1.0 -1.7
0.35 -2.0
1.5 -1.3
2.0 0.0 ], figure(2)
>> gplot(A,xy); hold on % Plot all edges
>> gplot(A,xy,'o') % Plot all nodes as 'o'
 Exercises 217

>> for k=1:10,

tt = sprintf('%d\n',k);
text(xy(k,1),xy(k,2),tt) % Print numbers
end; axis off
>> title('A with the original ordering a')
>> hold off; figure(3) % Now plot G(A2) below
>> xy2 = xy(al,:); % Reorder nodes by al
>> gplot(A2,xy2); hold on % Plot all edges
>> gplot(A2,xy2,'o') % Plot all nodes as 'o'
>> for k=1:10,
tt = sprintf('%d\n',k);
text(xy2(k,1),xy2(k,2),tt) % Print numbers
end; hold off; axis off
>> title('Al with the new ordering al')

The study of sparse matrix techniques is of importance in applied sci-

ences. Interested readers may consult two books on sparse matrices, [4]
and [15].

Exercises
16.1 Generate at least three random matrices A,,,,,, of random orders
n = 50 to n = 170 and random column vectors b of size n. Then
use flops to show the following results hold approximately and
estimate constant C in each case:
x = A\b takes Cn3 operations;
y = det(A) takes Cn3 operations;
z = inv(A) takes Cn3 operations;
t = b * b' * b takes Cn2 operations;
u = b * (b' * b) takes Cn operations;
v = b * 9 + 6 takes Cn operations;
w = A * b takes Cn2 operations.
16.2 Solve the following system Ax = b using both A\b and two
separate steps of the LU approach (are the two solutions iden-
tical?)
1 1 0 3 xl 4
2 1 -1 1 x2 _ 1

3 -1 -1 2 x3 -3
-1 2 3 -1 x4 4
218 Matrices and Solution of Linear Systems
16.3 Try the approach of 16.2.2 to solve new systems Ay = e and
Ax = f, immediately after solving Ax = b in Exercise 16.2,
with e = (4 2 - 2 1)T and f = (34 10 1 36)T.
16.4 Based on [L U P] =lu(A) for Exercise 16.2, calculate the follow-
ing relative errors
eL= IIL-L211
andeu= IIU-U211

IILII IIUII

where L2, U2 are chopped from L, U with t = 1, and II ' II here

refers to a norm computed by the MATLAB command norm,
for example, IILII computed by norm (L) .
16.5 Using l in-soly . m with lu2. m and lu3. m, solve Exercise 16.2
again by t = 1 digit arithmetic. Compare the two solutions.
Which one is more accurate? Use command norm as in Exercise
16.4.

16.6 Consider the system of linear equations Ax = b, where

1.1756 4.0231 -2.14170 5.1967
_ -4.0231 1.0002 4.5005 1.1973
A
-10.179 -5.2107 1.1022 0.10034
886.19 7.0005 -6.6932 -4.1561

and b = ( 15.721 19.392 2.9507 -38.089 )T. Take xe =

A\b as the true solution.
(a) Find the approximate solution y to Ax = b with t = 4
decimal digit arithmetic by means of Gauss elimination
without partial pivoting. Calculate the relative error

er= Ilxe - yll

IIXell

where II ' II denotes the 2-norm.

(b) Perform one iterative refinement based on (a), and cal-
culate the new relative error.
(c) Following (b) and using t = 4 again, how many more
steps of iterative refinements do we need to reach the
limit of iterative refinement, that is, to obtain IIell < 10-9
in y - y + e of the iterative refinement algorithm? Show
all intermediate residuals, their norms and the relative
error of the final solution.
Hint use the command while as in
 Exercises 219

y = lin_soly( ... );
count = 0
while ( norm(e) > 1.0E-9 )
r = b - A * y
e = . . . . . .

y = y + e ;

count = count + 1
norm_e_is = norm(e)
end

16.7 Following Exercise 16.6, solve the two modified problems below
by Gauss elimination with partial pivoting using t = 5 digit
arithmetic. Decide on the conditioning of each case.
(a) Keep the original A as above but modify b by values that
are randomly distributed on the interval [-0.001, 0.001].
Hint b - b + 0.002 * rand(4, 1) - 0.001 * ones(4, 1).
(b) Keep the original b but modify A by values that are
randomly distributed on the interval [-0.002,0.002].
16.8 EXTRAS Find out about how to generate a Hilbert matrix by
help hilb. Let A be a Hilbert matrix of order 6 and bT =
2 : 1 : 7. Find the actual relative error E for the numerical
solution y by solving Ax = b by Gauss elimination without
partial pivoting and using t = 8 digit arithmetic.
Modify your M-file so that it runs for t = 1 : 12, and for each
case print out a relative error for y.

16.9 For the following 9 x 9 sparse matrix

1 0 1 1 0 0 0 1 1

0 1 1 0 0 0 0 0 0
1 1 1 0 0 0 0 0 0
1 0 0 1 1 0 0 0 0
A= 0 0 0 1 1 0 0 0 0
0 0 0 0 0 1 0 0 1

0 0 0 0 0 0 1 1 0
1 0 0 0 0 0 1 1 0
1 0 0 0 0 1 0 0 1

use symrcm to obtain a new ordering a and a new permuted

matrix Al = A(a, a). Plot both graphs G(A) and G(A1) using
220 Matrices and Solution of Linear Systems
gplot, assuming that graph G(A) has these nodal positions
0 0
0 1

0 0.5
0 -0.5
xy= 0 -1
-1 0
1 0
0.5 0
-0.5 0

16.10 For the fo llowin g 1 1 x 11 matrix A , decom pose it by lu 3. m and

lu4. m and find the numbe r of flops required in each case
p. -2 0 30 3 3 0 4 0 3 -2
0 p 4 4 0 3 -2 0 0 3 0
0 4 It 3 0 0 -2 0 0 0 0
0 0 2 p 0 0 6 0 0 0 0
-2 6 0 0 p -2 4 -2 0 4 0
3 3 6 0 4 y -2 0 0 -2 0
0 2 4 4 0 0 p 0 0 4 0
-2 0 0 0 3 3 0 4 -2 3
2 0 0 0 3 0 0 -2 0 6
0 -2 3 4 3 -2 4 0 0 0
4 0 0 0 -2 2 0 -2 -2 0
where p = 999. Further use the MATLAB reorderingt com-
mand symrcm to obtain a new matrix Al. Compare the f lops
needed to decompose matrix Al using lu4. m with the previous
two cases and work out percentages of saving. Plot all nonzero
positions of A and Al using spy. Can you suggest an ordering
that is better than symrcm?

f Note that this matrix is not really symmetric but symrcm can still be useful.
 17
Function Interpolations and Approximation

Function approximation lies at the very heart of computational mathe-

matics where unknown or complicated functions are hoped to be repre-
sentable by simple ones. Interpolation is a simple and convenient way of
approximation. Polynomials are among the most useful and well known
as well as the simplest class of functions.

Aims of the project

The purpose of this investigation is to approximate general functions by
the polynomials. We consider the interpolation technique and the least
squares fitting. We shall consider the one dimensional case first (1D)
and then the multiple dimensions (mainly 2D). The two cases may be
considered as two separate projects.

Mathematical ideas used

Polynomials are used to represent discrete data or approximate other
functions. We consider the methods and criteria of constructing such
polynomials. Both the interpolation and least squares processes essen-
tially form matrix equations (linear system) for polynomial coefficients.
We address two methods of comparing such functions.

MATLAB techniques used

This project is about functions and their approximations. The MAT-
LAB commands interpl, interp2, polyf it and polyval, and M-files
polyf it2. m and polyval2. m are used. There will be an extensive use
of MATLAB's graphics capabilities to visualise different functions, all
illustrated by two M-files intdemol.m and intdemo2.m.

221
222 Function Interpolations and Approximation
17.1 1D: Introduction
In one space variable, a degree m polynomial may be written as
P,,,, (x) = C(1)xm + C(2)x"c-1 + + C(m)x + C(m + 1),
where C = [C(1) C(2) . . . C(m) C(m + 1)] is the coefficient vector.
In MATLAB, such a polynomial is evaluated by the command polyval:
Example 1
P4(x) = 2x4 - 5x3 + 3x2 + 8x - 1,
that is, m = 4, C = [ 2 -5 3 8 -1 ]. We can evaluate P4 at
points x = -9.0 : 4.5 : 10 and have a dotted as well as a line plot by
>> C = [ 2 -5 3 8 -1 ] ; % Type in the coefficient
>> x = -9.0 : 4.5 : 10.0 ; % A vector in [-9,10]
>> y = polyval(C, x) ; % Evaluate at x
>> plot (x,y,'ow');hold on; % [0] Scattered points
>> plot (x,y,'-w') ; % [1] Plot in solid lines
>> xp = -9.0 0.01 :10.0 : ; % A large vector in [-9, 10]
>> yp = polyval(C, xp) ; % Evaluate at xp for plots
>> plot ( xp, yp, ':g'); % [2] Dotted lines (better)
>> axis([-9 10 -2000 17500]);% Fix axis display
>> CQ = polyfit(x,y,2) % Quadratic fitting at (x,y)
;

>> y2 = polyval(CQ, xp) % Values of approximation

;

>> plot(xp, y2, '--r'); % [3] Dashed line plot

>> title('Example 1 -- plotting curves'); hold off

Here the solid line plott is not representative of the polynomial because
it is not a curve at all and the polynomial is far from being a straight
line. We have to use more points to show it as a proper curve (the dotted
line). This is the first important point on plots to be noted in doing this
project; see Figure 17.1.
Notation and Convention All approximations are based on a knowl-
edge of given data points, but test points are artificially chosen for anal-
ysis. To avoid possible confusion, we use x, y (size n) for data points, xi,
yi (size T) for test points and xp, yp (size L) for plot points. Obviously
L should be sufficiently large (say L >> n) but T can be any number.
For the Example 1 above, we have used n = 5 and L = 1901.
t Here we assume that the background colour is black. On some systems where this
is not the case, type set(gca,'Color ','k') to reset or replace the `white' (`ow'
and `-w') by `black' ('ok' and '-k').
 17.2 The 1D example M-file intdemol.m 223

-2000
-8 -6 -4 -2 2 10

Fig. 17.1. Example 1 - plotting curves (solid lines - polynomial C with few
plot points, dotted - with more plot points, dashed - a quadratic fitting).

In practice, we can identify two situations where polynomials are

deemed to be useful: (i) given a complicated function, we wish to use
a simple polynomial of a certain degree to represent it reasonably; (ii)
given a discrete set of points say (xi, yi), (x2, y2), , (x,,,, y,,), we wish
to find a polynomial function to represent the data.
In Chapter 4, we used polyf it for fitting a polynomial to case (ii).
This method using a single polynomial for approximation is called a
global method; an example can be seen in Figure 17.1 where a quadratic
polynomial approximation is shown. We shall compare such global meth-
ods with the so-called piecewise approximation methods where several
piecewise interpolating polynomials are put together. Here the former
uses polyf it and the latter uses interpl.

17.2 The 1D example M-file intdemol.m

To illustrate intdemol.m, we take y = cos(2x) in [-2,4] with n = 13
Example 2
f x=-2 : 0.5 : 4
y = cos(2 * x).
All approximations to this function are compared at 16 points xi = -2 :
0.4 4 but plotted at 151 points xp = -2 : 0.04 4. We suggest you
: :
224 Function Interpolations and Approximation
print out and study the demo file intdemo1.m which is based on this 1D
example; see Figures 17.2 and 17.3.

17.3 1D data fitting

Assume that a set of data points (xi, y1), (x2, y2), ... , (xn, yn) is given.
We hope to find a polynomial for fitting the data set. This problem has
been considered in Chapter 5.

17.3.1 Global methods (least squares method)

If the polynomial under consideration is of degree m, i.e.
Pm,,,(x) = C(1)x' + C(2)xm'-1 + + C(m)x + C(m + 1),
then the coefficients (vector C) are not known and will be determined
by the way we minimize P,,,, (xi) - yi. The least squares method proposes
to minimise the squares error with regard to C(i)s

E=E(C)= Y)2
i=1

where Y = P,(xi). When m = 1, we have the linear approximation

P1 (x) = C(1)x + C(2) as discussed in Chapter 5.
The MATLAB commands implementing such a method, for Example
2 with a fifth degree polynomial (m = 5), are simply
x = -2 : 0.5 : 4 % Set up given vector x
y = cos(2*x) Set up given vector y
C = polyfit(x,y,5); Find out C_i's in P_5(x)
xp = -2 : 0.04: 4; Points set up for plot
fp = polyval(C, xp) Evaluate P_m for plot
plot(xp, fp) Plot the approximating curve

17.3.2 Piecewise approximations

You must have realised from Example 1 (and Exercise 17.1) that polyno-
mials or rather global polynomials can be oscillatory or unsatisfactory.
We look for alternative approximation methods.
Interpolation is a similar but different approach to least squares fitting.
Here we want a polynomial to pass all n data points, i.e. to commit no
errors. But the degree m of such polynomials has to be high, namely
 17.3 1D data fitting 225

-1 0 1 2 3 4
File intdemol.m - example 1 D (dotted lines - true curve)

Fig. 17.2. 1D global approximation - linear(o), cubic(*), order 6(x).

-1 0 1 2 3 4
File intdemol.m - example 1 D (dotted lines - true curve)

Fig. 17.3. 1D piecewise approximation - linear(o), cubic(*), spline(x).

m = n - 1 (i.e. equal to one less than the number of data points). As

global and high order polynomials can be unsatisfactory, interpolation
appears to be no better than the least squares method except that the
former gives the advantage of exact data representation at data points.
It turns out that polynomials can give good approximations locally
(i.e. nonglobally and over a short range), and different or even same
226 Function Interpolations and Approximation
order polynomials can be effective if we can divide the data set into small
and nonoverlapping sets. This leads to the piecewise approximation
approach.f
Therefore, instead of constructing degree m = n - 1 polynomialst
as in Chapter 5, we consider the piecewise approach in the context of
interpolation.

17.3.2.1 Piecewise linear and cubic interpolations

The piecewise approach is to use several low order polynomials of in-
terpolation and piece them together (hence the name `piecewise'). The
overall approximation is in general not differentiable, though continu-
ous, at the joints, but it is efficient and accurate. Piecewise linear and
cubic interpolations are two examples.
The MATLAB command interpi can produce both piecewise linear
and cubic interpolations, as used below for Example 2:
>> x = -2 : 0.5 : 4; % Set up given vector x
>> y = cos(2*x); % Set up given vector y
>> xx = -2 : 0.04 : 4; % For interpolations and plots
>> FL = interpl(x,y, xx,'linear'); % Used for plot
>> FC = interpl(x,y, xx,'cubic'); % Used for plot
>> plot(xx,FL,'-r', xx,FC,'-g')

Here xx is chosen with a small spacing to plot the two interpolation

curves but it can be of any size, e.g. xx = -1 : 0.3 : 1.

17.3.2.2 Continuously differentiable cubic interpolations splines

The shortcoming of piecewise interpolations or the lack of differentia-
bility is overcome by insisting that the approximation have continuous
derivatives (say first and second order) at all interpolation points. For
cubic approximations, this gives rise to the popular spline interpolation
widely used in solving modern engineering problems.
The MATLAB command interpi can generate a cubic spline inter-
polation for Example 2 by
>> x = -2 : 0.5 : 4; % Set up given vector x
>> y = cos(2*x); % Set up given vector y

f The modern finite element method is one of the best application examples of a
piecewise approximation approach.
t Interested readers can consult any numerical analysis textbook for details and
formulae. Test this out using polyfit!
If one insists, the overall approximation can be made differentiable; see 17.3.2.2.
 17.4 How accurate is my approximation? 227

>> xp = -2 : 0.04 : 4; % Points set up for plot

>> FS = interpi(x,y, xp,'spline'); % Values for plot
>> plot(xp,FS,'-b')

17.4 How accurate is my approximation?

If an approximation, when plotted against the exact function,] almost
overlaps it, then this is a sign of a good fit. Here we discuss an analytical
way of measuring approximation. We shall use and compare it at some
test points.
Suppose that xi, x2, ... , xT (in vector xi) are our test points with the
known values 1, y2, ... , 9T (in vector yi). Then if f;s (in vector fl) are
the values of our approximation at xi, we may check how the method
performs by calculating (the so-called root mean square error)

Error = JIIfi_yiII 2 = I 1:
T
(fi _ yi)2.
rp
V i=1

If this error is small enough, we say the underlying approximation is

good. In theory, we could have xi = xi for all i (i.e. test points and data
points coincide with T = n), but it may not make much sense as with
interpolation Error = 0. (Think about the definition of `interpolation'.)
For Example 2, with a least squares approximation by a fifth degree
polynomial, we may proceed as follows to estimate the approximation
error in root mean square norm.

>> x = -2 : 0.5 : 4; % Set up given data point x

>> y = cos(2*x) ; % Set up given data point y
>> C = polyfit(x,y,5); % Find out C_i's in P_5(x)
>> xi = -2 0.4 4;
: : % Set up test points
>> fi = polyval(C, xi); % Approximations at test points
>> yi = cos(2*xi); % Exact values at test points
>> T = length(xi); % Get the size of "xi"
>> Error = sqrt( norm(fi - yi)-2 / T) % Root mean square

Here recall that norm is a MATLAB command. We remark that in the

last line, if f i is replaced by a vector generated from interpi (e.g. FL),
t In real life where the exact function or solution is not available, one usually com-
pares with a known and accurate approximation (which may be expensive to obtain
in terms of time).
228 Function Interpolations and Approximation
we have to check if we need to transpose it before taking the differencet
fl - yi.

17.5 Introduction to multi-variable approximation

In 17.1 with 1D, we considered various ways to find a polynomial Pm(x)
(defined by its coefficient vector C). However, real world problems often
involve more than one space variable; a general mth degree polynomial
in n variables (Rn) may be represented by
P. = Pm(xi, x2, ... xn) = Xn

1 + ... +
rnO...Qxm

00...0)

where ik = 0, 1, ... , m for k = 1, 2, ... , n. For simplicity, we shall con-

centrate on 2D problems with n = 2 and on two special polynomials
with m = 1 (linear) and 3 (cubic).
In the case of m = 1, we consider the so-called bilinear function of the
form
Z(x,y) =a+bx+cy+dxy,
while in case of m = 3, we approximate functions by the so-called bicubic
function
Z(x, y) = a+bx+cy+dxy+dixy2+d2x2y+elx2+e2x3+fly2+f2y3.
In practice, similarly to the 1D case, we can identify two situations
where polynomials are deemed to be useful: (i) given a complicated
function z = z(x, y), we wish to use a simple polynomial Z = Z(x, y)
of certain degree (m = 1 or 3) to represent it reasonably (based on N
points); (ii) given a set of discrete data points (xl, yl, zi), (x2, y2, z2),
... , (xN, yN, zN), we wish to find a polynomial function Z = Z(x, y)
to represent the data.
We shall first investigate ways of finding such a polynomial and then
study methods of measuring how good each fit is.
You will need two main M-files specially developed for the remaining
part of this project:
polyfit2.m - Least squares fit in 2D (Linear and Cubic)
polyval2.m - Evaluation of "polyfit2.m" (Linear & Cubic)

t More precisely, with MATLAB V.4, FS = interpl(x,y, xp, 'spline') always

produces a column vector even when xp is a row vector! This is not so with
MATLAB V.5.
 17.7 Contour plots, 3D plots and slicing 229

together with example M-files intdemo2. m, cont4. m and cont7. m.

Notation and convention The careful reader may have noticed that
m the degree of polynomials;
n the number of independent variables;
N the number of data points.

17.6 The 2D M-file intdemo2.m

To illustrate the above M-files, a demonstration M-file has been prepared
for your reference, where the following example is used (N = 25):
Example 3
x=00.25: 1
y=00.251
z = r ln(r),
where r = (x - 1/2)2 + (y - 1/4)2. As before, you should run and
print out this M-file intdemo2. m before proceeding.

17.7 Contour plots, 3D plots and slicing

As we are to work with functions of n = 2 variables, we should investigate
how to visualise such functions.
In MATLAB, we can draw contour plots and 3D plots, both involving
similar preparations. Suppose that we want to draw the function
z = z(x,y) _ (x6 - y3 - 0.5)exp(-x2 - y2)

in S2 = [-2, 2] x [-2, 2]. The command meshgrid is first used to set up

a grid, followed by calculations of values (if they are not yet known):
>> xO = -2 . 0.4 . 2;
>> yO = -2 : 0.25 : 2; %% x0, y0: 2 1D vectors
>> [x y] = meshgrid (xO,yO); U To get two matrices
>> ZZ=(x."6-y."3-0.5) .* exp(-x.-2-y."2); % ZZ = matrix

3D plots Two commands mesh and surf (and their variantst) can be
used to visualise ZZ; (see Figures 17.4 and 17.5 for actual plots).
>> mesh(ZZ); title('A mesh plot');
>> disp('Pause ...'); pause
>> surf(ZZ); title('A surf plot')
t Type help mesh to see how many other commands are referred to. Some com-
mands such as surf and plot3 are only available from MATLAB V.4.
230 Function Interpolations and Approximation

Fig. 17.4. A mesh plot.

To rotate a graph, try the command view, e.g. view(20, 70).

Contour plot The command contour may be used to draw contour
plots but we need to specify values where contourt lines are requested
(refer also to the M-file cont4.m)
>> M=max(ZZ); mi=min(ZZ); % Row vectors
>> V=[mi mi/2 0 M/2 M]; % Some heights of ZZ
>> C1 = contour(x,y,ZZ, V,'-r'); % MAIN step
>> clabel(C1); % Labels (numbers)
>> title('Contour plot of z(x,y)')
>> C2 = contour(x,y,ZZ, 7,'-g'); % Simpler 7 levels
Figure 17.6 shows the result of the first four lines of commands (i.e. use
the height vector V).
Note: You must have noticed that, with x0 = -2 : 0.4 : 2 and y0 =
-2 : 0.25: 2, both surface and contour plots appear very nonsmooth.
To increase the smoothness, i.e. to view the true curve surfaces, we need
to reduce the step-lengths, e.g. use x0 = -2 : 0.01 : 2 and y0 = -2
0.02 : 2 to generate more grid points.
t We remark that in the older versions of MATLAB (say Versions 4-), vari-
ations of the input method shown here are also acceptable, e.g. Cl
= contour(ZZ,7, xO,yO, '-r') or C1 = contour(ZZ,V, xO,yO, '-r') or C1 =
contour (x0,yO,ZZ,7, -r'). Only the very last variation is valid with MATLAB
V.S.
 Contour plots, 3D plots and slicing 231

12

Fig. 17.5. A surf plot.

Fig. 17.6. A contour plot.

Sliced plots Sliced plots t are line plots that are sliced off 3D plots in
one axis direction. For Example 3, we may slice at mid-intervals in each
direction and then plot each slice separately (see Figure 17.7):

t MATLAB V.5 introduces a new and convenient command called slice for the
same purpose. Type help slice to find out more details.
232 Function Interpolations and Approximation

0.4

0.2
0
X
0
c0

c2 -0.2
.2
-0.4

-0.61
1 1 1 1 1 1

-2 -1.5 -1 -0.5 0 0.5 1 1.5 2

Fig. 17.7. Sliced plots.

>> x = 0; y = -2 : 0.25 : 2; %% x=value, y=vector

>> Zx=(x."6-y.'3-0.5) .* exp(-x."2-y."'2); % Zx=vector
>> subplot(211);plot(y, Zx,'-r'); title('Sliced plots')
>> ylabel('Sliced at x=0')
>> y = 0; x = -2 : 0.25 : 2; % Fix y
>> Zy=(x."6-y."3-0.5) .* exp(-x."2-y."2); % Zy=vector
>> subplot(212); plot(x, Zy, '-g');
>> ylabel('Sliced at y=0')

A different (and maybe better) method of plotting sliced plots is by use

of the command plot3; we repeat the above example as follows:
>> subplot(111)
>> x = 0; y = -2 : 0.25 : 2;
>> Zx=(x."6-y."3-0.5) .* exp(-x."2-y."2)
 17.8 The `\' global method 233

hl=plot3(zeros(size(y)), y, Zx, -r');

hold on %%% Necessary
y = 0; x = -2 : 0.25 : 2;
Zy=(x."6-y."3-0.5) .* exp(-x."2-y."2)
h2=plot3(x, zeros(size(x)), Zy, '-g');
zlabel('Slices at planes x=0 and y=0')
grid on; hold off % Change viewpoint below
view(10,50) % Try view(80,50) or view(175,50)
set(hl,'Linewidth', 3); set(h2,'Linewidth', 2)
set(gca,'Ytick', -2:1:2)

Here the last two lines annotate the graph in an advanced way; type
get (gca) and help set to find out which other attributes of a graph can
be modified. As sliced plots do not represent the full graph, they should
be used with caution but they can be useful when comparing several
approximations. Of course, slicing can be in any suitable directions, e.g.
along x + y = 1.

17.8 The '\' global method

Suppose Z(x, y) is our approximation to the true function z(x, y) and
zi = z(xi, yi); it may be possible that z(x, y) is not known but that its
discrete values are known.
With the least squares approach, we hope to minimise the squares
error (assume Zi = Z(xi, yi) for all i)
n
E2=E(zi-Zi)2.
i=1
The minimisation is with respect to the underlying parameters; e.g.
a, b, c, d for the bilinear case of 17.5. The method has been imple-
mented in two M-files where the `\' approach has been used (see the
appendix to Chapter 5 for a theory). The syntax to use polyf it2.m
and polyval2.m for Example 3 is as follows (see cont7.m):
x0 = 0:0.25:1; y0 = x0; %% Interpolation nodes
[x y] = meshgrid(xO, y0);
R = sqrt ( (x-1/2)."2 + (y-1/4)."2 );
z = R.*log(R+eps) %% True and given values
;

C1 = polyfit2( x,y,z, 'linear');

C2 = polyfit2( x,y,z, 'cubic' );
%%%%%%%%%%%%%%%%%%%%%% Evaluations for plots only %%%%%%%
234 Function Interpolations and Approximation
x0 = 0:0.02:1; y0 = xO ; % set nodes
[xi yi] = meshgrid(xO, y0);
Z1 = polyval2( C1, xi,yi, 'linear'); % Bilinear
Z2 = polyval2( C2, xi,yi, 'cubic' ); % Bicubic
%%%%%%%%%%%%%%%%%%%%% Contour Plots %%%%%%%%%%%%%%%%%%%%%
contour(Z1,9, xO,yO, '-g'); hold on; % 9 levels
contour(Z2,9, xO,yO, ':r'); hold off

17.9 The piecewise method

The global method tends to fail for certain difficult functions because
high order polynomials are oscillatory. One alternative, as in 1D, is to
use piecewise methods that have the advantage of smoothness of low
order polynomials and reasonable local accuracy.
Here we shall consider the interpolation method using piecewise low
degree polynomials. The MATLAB command interp2 can be used for
both bilinear and bicubic interpolations. As we know, the intersection
of two pieces of surfaces is a curve that has an uncountable number of
points. It is not possible to have continuity and differentiabilityt over
this curve, because the number of freedoms of any degree polynomial is
limited. Hence there are no `splines' in other than one dimension.
We now show how to use interp2 for Example 3 (see cont7.m)
>> x=0 :0.25: 1; y=x; [x y]=meshgrid(x, y);
>> R=sqrt ( (x-1/2)."2 + (y-1/4)."2 ); z = R.*log(R+eps);
>> xa=0 :0.05: 1 ya=xa ;
; % New 1D vectors
>> [xi yi] = meshgrid(xa, ya) % New 2D matrices
;

>> zL = interp2(x,y,z, xi,yi, 'linear');

>> zC = interp2(x,y,z, xi,yi, 'cubic');
>> contour(zL,9, xa,ya, '-w'); hold on;
>> contour(zC,9, xa,ya, '-b'); hold off

17.10 Comparison of approximations

For a function z = z(x, y), we have discussed four kinds of approxima-
tions Z = Z(x, y). We hope to decide which one is the closest to the
original function by comparing them at some selected points.
f However, in such cases, one may be content with incomplete differentiability, i.e.
differentiability at some points (not all points of a curve). Tensor products can
achieve such a result.
 17.10 Comparison of approximations 235

These selected points, called test points as in 1D, can be those used
to generate various plots but should not be the starting data points x0
and y0.
Suppose that an approximation Z and its corresponding true function
values z are in the form of matrices (say of size N1 x N2). In the example
of the last two sections, such an approximation Z may be Zl, Z2, zL, zC
and z is calculated using the original function z(x, y). Then the following
two methods can be considered.
Geometric comparison Using the hold on option, we may overlap an
approximation Z on the contour plot of the true values z (using different
colours of course). Then large discrepancies would indicate that the
underlying approximation is not very good and vice versa. Note it is not
easy to overlap plots with mesh and surf.
One may also generate a contour plot and decide where the main fea-
tures of a function are, before producing suitable sliced plots for compar-
isons. Conclusions drawn from a geometric comparison can be subjective
but it is effective to identify the poorest approximation method.
Algebraic comparison A very useful but boring method for measur-
ing discrepancies between two matrices is by use of norms, e.g. E _
norm(Z - z) as an error indicatort - a small E indicates good approx-
imation and vice versa.
In engineering applications, as in 1D, it is more common to use the
so-called root mean squares norm defined by
Ni N2
E= NlN2 1
(ZZj - zi.7)2.

i=1 j=1
The smaller such an error is, the better the method is.
Luckily the above formula resembles the Frobenius norm of matrix
(Z - z) that can be computed via the MATLAB command norm(Z-z,
' fro') . Thus the quantity E can be computed as follows
>> [m n] = size(Z); % Get the dimension of matrix
>> mn = m * n ; % Total number of nodes
>> format compact; % Don't leave extra space lines
>> disp('Root Mean Squares error is :');
>> E = sqrt( norm( Z-z, 'fro')-2 / mn )

Finally, we remark that data fitting and function approximations play

f See Chapter 16 for more examples of norm.
236 Function Interpolations and Approximation
important roles in scientific computing. As this project shows, an effi-
cient method may not work well for all problems - a fact that typifies
the difficulty and challenge for mathematicians doing real modelling and
simulations. There are many other choices of global functions that may
be substitutes for polynomials such as trigonometric functions, radial
basis functions and, more recently, wavelet functions. Again for each
case, one may look at some kind of piecewise versions - functions with
relatively small compact support. Overall, the design of a robust method
involves careful experiments and observations, and above all, mathemat-
ical analysis.

Exercises
17.1 Given the function f (x) = 1/(1 + 25x2) in [-1, 1] and nine data
points x = -1 : 0.25: 1, plot the function against its polynomial
approximations of degrees 3 and 7.
17.2 Following Exercise 17.1, plot the original function against its
piecewise linear, cubic and spline approximations.
17.3 For the following given functions
(a) f (x) = -(x + a)R + [aR(1 - x) + (1 + a)Rx] with a = 10-2
and R = -1 in [0,1],
(b) f (x) = (1 - x){tan-1 [b(x - xo)] + tan-1 [bxo] } with b= 100
and x0 = 1/4 in [0,1],
(c) f (x) = tanh[20(x - 1/2)] in [0, 1],
(d) EXTRAS f(x) = 10000(f1(x)+f2(x))I(x-1/4)(x-3/4)3
with fi(x) = -(x + a)R + [aR(1 - x) + (1 + a)Rx] and
f2(x) = -(1 - x + a)R + [aRx + (1 + a)R(1 - x)] where
a = 0.3 and R = -1/4, in [0,1],
(i) Compare the performance of four approximations: least
squares method using for m = 1 20, piecewise lin-
:

ear method, piecewise cubic method and piecewise cu-

bic spline method by using n = 10 uniformly distributed
data points (x = 0 : 1/9 : 1);
(ii) Try to improve the performance of at least one of the two ap-
proximations: least squares method using for m = 1:20
and piecewise linear method, by selecting n = 10 manu-
ally placed data points that are nonuniform (say x = [0 :
0.1:0.4 0.6:0.1:1] or x = [0 0.26: 0.15: 0.44 0.47 :
 Exercises 237

0.035 : 0.65 1]). Plot the two approximations on the

same graph.
Notes In each case, to simplify the problem, you should
(1) test the root mean squares error at T = 10n+1 uniformly
distributed test points (i.e. about 1/10 spacing of data
points);
(2) rank the methods from the best to the worst in a perfor-
mance table (according to error indicators at test points);
(3) plot all competitors on the same graph.
Hints:
(I) For the least squares method, you should try up to the
degree of m = 20 polynomial but only quote the best
result for comparison.
(II) For the case of using `manually placed data points', if you
are not sure about how to find these points, the following
may help
plot a given function,
plot the first derivative of a given function,
then place more points near large variations of the func-
tion or near large values of its derivative.
dtanh(x) d tan-'(x) _ 1
(III) _ 1 - tanh2 (x) and
dx dx l + x2
(IV) On completion, you should obtain
(1) 6 graphs;
(2) 3 x 4 sets of root mean squares errors and 3 orderings
from using uniform points;
(3) 3 x 2 sets of root mean squares errors and 3 orderings
from using nonuniform points.
17.4 Plot the following two functions and also show their sliced plots
at some suitable positions (only one 3D and one slice plot are
required for each case):
gl (x, y) = tanh[20(x + y - 1)] in [0,1] x [0,1];
g2(x, y) = tanh[20(x2 + y2 - 1/2)] in [0, 1] x [0, 1].
17.5 For the two functions in Exercise 17.4, show contour plots (with
five levels) of their linear and cubic approximations; use N = 11
points in each coordinate direction, i.e. x0 = y0 = 0 : 0.1 : 1.
238 Function Interpolations and Approximation
17.6 For the two functions in Exercise 17.4, show contour plots (with
five levels) of their piecewise linear and cubic approximations;
use N = 9 points in each coordinate direction, i.e. xO = yO =
0 : 1/8 : 1.
17.7 For one of the two functions in Exercise 17.4, compare its global
linear and cubic approximations based on Nl = N2 = 13 points
(see Exercise 17.5 with xO = yO = 0 : 1/12 : 1) against its
piecewise linear and cubic approximations based on Nl = N2 =
8 points (see Exercise 17.6 with xO = yO = 0 : 1/7 : 1):
(a) First carry out algebraic comparisons and rank the four
methods in each case from the best to the worst. The
number of test points should be more than 4 * Nl * N2.
(b) Then present suitable graphs supporting your conclu-
sions. (Definitely sliced plots are effective but try to get
at least one contour or a 3D plot as well.)
 18
Ordinary Differential Equations

Aims of the project

You are invited to study the solutions of a list of ordinary differential
equations using whatever methods you have at your disposal.
Mathematical ideas used
Some of the equations are capable of analytic solution using such math-
ematical tools as: separation of variables, integrating factors or series
solutions. Others are examples of homogeneous, or constant coefficient
differential equations. What you can bring to bear will very much de-
pend on your mathematical background at this point.
MATLAB techniques used
All the numerical techniques required have been introduced in Chapter 7.
For example, grain (or phase) plot analysis and the numerical solution of
coupled first order equations. You may find helpful the M-files associated
with that work (fodesol.m, species.m, vderpol.m,...). You can use
these directly or copy and modify them as required.

18.1 Strategy
Your aim, for each equation in the list of exercises, is to provide the
following information as appropriate:

(a) For first order equations, a grain plot with typical solutions su-
perimposed. For second order, or coupled first order equations,
sketch of a typical phase plot.
(b) An analytic general solution if you can find one.
(c) The particular solution for the specified initial conditions.

239
240 Ordinary Differential Equations
(d) Any other comments you wish to make on the nature of the so-
lutions, their stability etc.

In each case, start by classifying the type of the differential equation.

Is it linear? What order is it? Has it got constant coefficients? Is it
homogeneous? If it is of a type which you recognise, then try to solve it
`analytically', that is, by paper and pencil. If you can't, then see if you
can use one of the techniques discussed in Chapter 7. If you have been
able to solve the equation exactly, you can compare the solution with
the numerical or graphical one got by using MATLAB. In each exercise,
you are given a particular solution to find. You should try your best to
find out as much as you can about all possible types of solution.

Exercises
18.1 Find y at x = 1 for the particular solution of

dy = -2xy
dx
satisfying the initial condition y(O) = 1.
8.2 Find y at x = 1 for the particular solution of
dy = e_2xy
dx
satisfying the initial condition y(0) = 1.
8.3 Find y at x = 5 for the particular solution of

(x-2 xy)dy-y=0, (x>0,y>0)

satisfying the initial condition y(l) = 10.
18.4 Find y at x = 1 for the particular solution of
2
dy
+ 2 y = ex
dx +32

satisfy the initial condition y(O) = 1, y'(0) = 0.

18.5 Find y at x = 4 for the particular solution of
d2y
dx2
+y 2 -x 2 =0
satisfying the initial condition y(0) = 1, y'(0) = 0.
 Exercises 241

18.6 Find y at x = 4 for the particular solution of

d2y
= cos(2x)y
dx2
satisfying the initial condition y(O) = 1, y'(0) = 0.
18.7 Find y at x = 4 for the particular solution of
2

dx2 dx
+x2y=0
satisfying the initial condition y(O) = 1, y'(0) = 0. (Hint: You
may have difficulty finding the numerical solution when x is
near 0 (called a `regular singular point'). By considering a series
(Maclaurin) expansion such as y(x) .: a+bx+cx2 you can work
out how y and its derivative must behave for very small x-values
should your M-file need to know this.)
18.8 Find x at t = 1 for the particular solution of
d
dt = x(2x + y - 3), t = y(x + 2y - 3)
satisfying the initial condition x(0) = 5, y(0) = 1. (Hint: Think
also about fixed points and stability as studied in Chapter 7.)
18.9 Find y at x = 1 for the particular solution of
z
(1-x2)ddx2y -2xdy+12y=0
dx
satisfying the initial condition y(O) = 0, y'(0) _ -2. (Hint: See
the advice given for exercise 18.7, if you experience difficulty in
finding solutions at particular x-values.)
Part three
Modelling
 19
Checkout Queues: Long or Short

In daily life, we meet many examples of first-in-first-out or fifo queues

that are usually simple queueing systems. A more complex multiqueue-
ing system might on one hand involve a fan-in structure in which several
separate queues have to merge at a later facility, on the other hand, in-
volve a fan-out structure where a single queue branches to form distinct
subqueues. The people, objects, customers that flow through the sys-
tem, known as entities, can have attributes, such as requiring leaded or
unleaded petrol, paying by cash or cheque, that they carry with them
through the system.
Aims of the project
The purpose of modelling queues is two-fold. Firstly, as customers, we
tend to prefer short queues in order to save time. Secondly, as busi-
ness managers, we hope the service utility (i.e. the ratio of actual time
when a service is utilised over the maximum available service time) is ap-
proaching 100% to maximise profits but still want to avoid long queues
for the sake of customer loyalties. Therefore it is of practical interests
to predict the peak times of long queues before deciding how to improve
the service.

Mathematical ideas used

Statistical distributions and probabilities used here have been discussed
in Chapter 6. Sparse vectors of Os and Is, for data extraction, are gen-
erated by various conditions. Flexible vector operations will be needed.
MATLAB techniques used
The supplied M-file queue. m is mainly used, assisted by three other M-
files exprand.m, normrand.m, unirand.m. There will be an extensive
use of MATLAB commands plot, bar, sort, rand.

245
246 Checkout Queues: Long or Short
This project is divided into three parts. 19.1 deals with some tasks
of modelling simple queues and introduces the main M-file queue. m.
19.2 models queues at a busy motorway petrol filling station while
19.3 considers queues at the Leo's cafe in a retail park. Both problems,
each representing a separate project, involve making decisions on staffing
levels, and improving profit margins and promotion issues.

19.1 Simulating queues

A simulation is performed as a sequence of events which are of two types

Arrival: An entity arrives at a service and joins the queue. If the

queue was previously empty and the service idle then the entity service
starts immediately, otherwise the entity joins the back of the queue
and waits in line.
End-of-service: An entity moves on to next stage. If there are no
entities waiting in the queue then the next service starts immediately
and is not noted as a separate event, as far as the current entity is
concerned. Of course, if the current queue is empty then the service
is unoccupied until the next arrival.

If there is a further stage, an entity moves directly to the next queue

and this arrival at the new queue is not a separate event. If there are no
further stages, the entity leaves the system. In the examples of the next
section, `vehicle' = entity and `petrol pump' = service while in 19.3,
`people' = entity and `travel from stores to cafe' = service.
All entities have to be separated into two distinct queues at some
stage. In the case of a single queue and a single server, we typically
need

vector 1 - a sequence of inter-arrival times;

vector 2 - a sequence of service times (attributes).
Note that with cumsum inter-arrival times are converted into arrival
times. As we said one cares both about an entity (or entities) and about
the overall utility of a service, so as to make the underlying business
convenient and efficient as well as cost effective. The above two vectors
can be used to work out the required information - we have developed
a M-file queue. m for such a purpose. Firstly, recall how we model the
inter-arrival times and service times.
 19.1 Simulating queues 247

19.1.1 The statistical theory

The inter-arrival time It may be described by a negative exponential
distribution. As in Chapter 6, this can be denoted by
It = -L ln(1 - r),
where r is the uniform random variable, L is the mean value for inter-
arrivals, and It is the inter-arrival time (simulated).
Service times are usually based on simple discrete attributes (such as
40% turn left and 60% turn right) and can be simulated by MATLAB's
random numbers as in Chapter 6.

19.1.2 The M-file queue. m and some associated M-files

The supplied M-file queue. m is designed to work out the waiting time
and end-of-service time, given an arrival time and a service time. The
syntax is as follows
>> [Mean_q Serv_u Wait_t Stop-t] queue(Arr_t, Ser_t)

where the two input vectors are

Arr_t the vector of arrival times;
Ser_t the vector of service times;
and the four output parameters are
Mean_q a scalar showing the mean queue length for the period of the
simulation;
Serv_u a scalar showing the fraction of total time that the service was
in use;
Wait_t the vector of waiting times (queueing to be served), excluding
the service times;
Stop_t the vector of end-of-service times.
For example, with two arrivals Arr = [2 8] and corresponding service
time Ser = [7 4] (in minutes), we have
>> [M S Wait Stop] = queue(Arr, Ser)
M = 0.9231
S = 0.8462
Wait = 0 1 % No waiting for 1st but only for 2nd
Stop = 9 13 % The time when service is finished
248 Checkout Queues: Long or Short

Table 19.1. Simulation data of a supermarket checkout.

Inter-arrival time IT Number of items N Service time ST = S * N
47 9 90
94 7 70
58 7 70
103 3 30
3 6 60

The M-file queue. in, as used above, has been devised to simulate a simple
queue. For more complicated examples it is necessary to write a simple
program that calls queue. m more than once. f
To simplify the generation of random numbers, you may use the fol-
lowing three M-files: exprand.m, normrand.m and unirand.m; these
M-files should be easy to understand as illustrated in Chapter 6 but
type help to see the usage if necessary. To compute a matrix of n x k,
exponentially distributed with mean .t = mu, type
>> matrix = exprand(mu,n,k)

To compute a matrix of size n x k, normally distributed with mean

p = mu and standard deviation o, = sigma, type
>> matrix = normrand(mu,sigma,n,k)

To compute a matrix of size n x k, uniformly distributed between a and

b, type
>> matrix = unirand(a,b, n,k)

19.1.3 The best time to go shopping

Customers arrive at a supermarket express checkout according to a neg-
ative exponential distribution of inter-arrival times with mean L = 50
seconds. They have I = 10 items or less and the checkout time is S = 10
seconds per item. Here the inter-arrival time of customers can be simu-
lated by IT = exprand(5, L) for five arrivals, and the number of their
purchases is simulated by N = ceil(unirand(5, 0, 10)). Some sample
data are shown in Table 19.1.
t We remark that from an efficiency point of view, service utility should reach 1.0
(or 100%) to minimise cost, but this inevitably means long queues that should be
avoided to maximise customer satisfaction and loyalty.
 19.1 Simulating queues 249

The arrival time AT can be worked out using cumsum based on the
inter-arrival time IT. Further we can work out when a person leaves the
supermarket till, how long each person needs to wait in the queue, the
average waiting time and the percentage of time when the till is in full
use. All this information can be obtained by a single call to the M-file
queue. m as follows

>> IT = [47 94 58 103 31;

>> AT = cumsum( IT )
>> ST = [90 70 70 30 60]; % Service time
[Mean_q Ser_u Wait_t Stop-t] = queue(AT,ST)

The solution from MATLAB will be

AT = 47 141 199 302 305

Mean_q = 0.9158
Ser u = 0.8163
Wait -t = 0 0 12 0 27
Stop_t = 137 211 281 332 392

This means that on average a customer finds a queue of length Mean_q =

0.9158: 1 person waiting for service (i.e. the till service is fast). In
particular of the five arrivals, the first, second and fourth person need
not to wait but the third and the fifth person are a bit unfortunate
having to wait in the queue for one person ahead.

19.1.4 An event table

Once a simulation is performed, we can construct an event table based
on the information obtained. An event table can show the full details
of each entity: arrival time, details of attributes, queue length and end-
of-service time. Thus we may keep a track record of events to facilitate
the efficient running and management of a service.
For the above shopping example, most queueing observations can be
reflected in an event table (see Table 19.2) which is built up from arrival
time, waiting time, and stopping time etc, where N denotes the number
of items and EoS the end-of-service time while QL represents the queue
length that is calculated by hand (note: the mean of these QLs has been
given by queue. m and the last column EoS is equal to the stopping time
as given by queue.m).
250 Checkout Queues: Long or Short

Table 19.2. Simulation results of a supermarket checkout.

Event Clock (seconds) N ST QL Next arrival EoS
start 0 0 47 free
arrival(l) 47 9 90 1 141 137
arrival(2) 141 7 70 1 199 211
arrival(3) 199 7 70 2 302 281
arrival(4) 302 3 30 1 305 332
arrival(5) 305 6 60 2 343 392

Main pump

Kiosk at shop

Side pump

Fig. 19.1. Layout of the filling station (Main = motorway).

19.2 The motorway filling station

The motorway filling station (see Figure 19.1) stands at a road junction
with one entrance from the motorway and one from a side road. There is
one petrol pump for each entrance with cars from the motorway queueing
at one pump and cars from the side road queueing at the other.
All customers queue to pay at a single cash point after serving them-
selves with petrol, where 20% pay by cash, 30% by credit card and 50%
by cheque (think how you could distinguish these). The inter-arrival
times of cars at each entrance are distributed according to a negative
exponential distribution, the mean inter-arrival times being 100 seconds
on the motorway and 150 seconds on the side road. The service times
at each pump are distributed uniformly, ranging from 10 seconds to 100
seconds at the motorway pump and from 30 seconds to 150 seconds at
the side road pump. The time taken at the cash kiosk is 30 seconds for
payment by cash, 40 seconds for payment by credit card and 60 seconds
for payment by cheque.
Based on a small sample of the first few arrivals, we can determine
their arrival times, service times and cash point times (to the nearest
second) by using pseudorandom numbers as shown in Table 19.3, where
main refers to the motorway pump and the random vector r is used to
 19.3 The Leo's cafeteria 251

Table 19.3. Sample data for the motorway filling station.

Arrival AT Service ST Payment method Time taken
main side main side random r symbol s at cash point
2 17 56 46 0.96 +1 60
246 72 52 129 0.96 +1 60
344 98 94 74 0.07 -1 30
400 356 33 19 0.22 0 40
405 0.11 -1 30

work out the payment method (vector s). All times, shown in tables
of this section, are converted into seconds. Further in Table 19.4, we
can construct the corresponding event table to display the results (note:
EoS = end-of-service corresponds to the event in column 1). Again we
may use queue. m to verify the correctness of this EoS column as follows:
>> Arr = [ 58 63 201 275 298]; %% Arrival at kiosk
>> Ser = [ 60 60 30 40 301;
>> [mq su wait_t EoS] = queue(Arr, Ser)

where Arr and Ser are collected from the `*.k' rows (i.e. the actual
combined arrivals at the cash kiosk) of Table 19.4. The results (for EoS)
match those corresponding entries in the last column of Table 19.4, as
expected:
[mq su] = 0.8464 0.6377
wait_t = 0 55 0 0 17
EoS = 118 178 231 315 345 % Good %

Note: For the cash point times, if r denotes a vector of uniformly dis-
tributed numbers in [0, 1], let customers pay by cash if 0 < r < 0.2, by
credit card if 0.2 < r < 0.5 and by cheque if 0.5 < r < 1. A simple
method of generating is for those r < 0.2 and Os for those r > 0.2 is
the following k1 = (r < 0.2).

19.3 The Leo's cafeteria

The Leo's cafeteria is situated on the exit route in a busy retail park,
in a city centre, where there are three main businesses: a superstore, a
DIY store and a garden centre. Customers visit these businesses first
before dropping into the cafe for some food and drink. As illustrated in
Figure 19.2, customers arrive at one of three businesses to buy various
252 Checkout Queues: Long or Short

Table 19.4. A sample event table for the motorway example

('ml.p' means the first arrival `ml' at the motorway pump (referred to
as main), 'ml.k' means the arrival of `ml' at the kiosk, and similarly
`sl.p' and `sl.k' are for the side road).

Event Clock Pay Q ueue length (QL) Next arrival EoS

code (s) time ma in side kiosk main side (exit)
start 0 0 0 0 2 17
ml.p 2 1 0 0 24 6 17 58
sl.p 17 1 1 0 24 6 72 63
ml.k 58 60 0 1 1 24 6 72 118
sl.k 63 60 0 0 2 24 6 72 178
s2.p 72 0 1 2 24 6 98 201
s3.p 98 0 2 2 24 6 356 275
s2.k 201 30 0 1 1 24 6 356 231
m2.p 246 1 1 0 34 4 356 298
s3.k 275 40 1 0 1 34 4 356 315
m2.k 298 30 0 0 2 34 4 356 345
m3.p 344 1 0 1 40 0 405 438
s4.p 356 1 1 0 40 0 405 375

Superstore 40%, dl

50%, d2
DIY store Leo's cafeteria

Garden centre 60%, d3

Fig. 19.2. Location of the Leo's cafe.

items and after some time most of customers arrive at the last stop
the cafe (via the front entrance). The cafe has two entrance doors: a
front (main) one and a rear one, but only one cash kiosk. Previous data
show that:

(a) The inter-arrival times of customers at superstore, DIY store and

 Exercises 253

garden centre all satisfy the negative exponential distribution,

with means of ml, m2 and m3 seconds respectively.
Customers arriving at the front door of the cafe are from one of
these three businesses but a smaller number of customers arriving
at the rear door of the cafe are mostly from a nearby city centre
car park; the inter-arrival time for these people alone (at the rear
door) satisfies the negative exponential distribution, with a large
mean of m4 seconds.
(b) 40% of customers at the superstore will visit the cafe next.
50% of customers at DIY store will visit the cafe next.
60% of customers at garden centre will visit the cafe next.
(c) It takes customers from the superstore, DIY store and garden
centre (on average) dl, d2 and d3 seconds respectively from the
moment they first enter a store to the moment they arrive the
cafe.
(d) Of all customers arriving at the cafe, the time taken at the cash
kiosk is 20 seconds for payment by cash but 40 seconds for pay-
ment by credit card. Actually 90% of people pay by cash and so
only 10% by credit card.

Exercises
19.1 Assume the mean value L for inter-arrivals of fishing boats at a
sea port is 3 minutes. Consider three kinds of distributions for
the inter-arrival time It:
(a) uniform distribution in [0, 6];
(b) normal distribution with standard deviation or = 1;
(c) negative exponential distribution.
Take the first 100 arrivals in each distribution and, using queue. m
with the service time of 2 minutes for each boat, plot all three
arrivals versus respective waiting times on the same graph. Find
the peak waiting time for each distribution and mark it on the
graph.

19.2 For the supermarket example in 19.1.4, taking L = 45, I =

12, S = 8, use queue. m to simulate nine arrivals and complete
a similar event table.

19.3 For the motorway filling station of 19.2:

254 Checkout Queues: Long or Short
(a) Generate 145 arrivals on the main road (motorway) and
96 arrivals on the side road, and perform a simulation
before working out all three waiting times (at two pumps
plus cash kiosk) with the help of queue. m. Further:
obtain bar charts (or other forms of suitable plots) to
present the information on arrival times against the
waiting times (for each of the pumps and for the cash
kiosk);
work out a peak time (in terms of the longest waiting
time) for each pump and also for the cash kiosk. Mark
the peak time on the respective plots.
Hints The results will be in three vectors of waiting time.
The event table is not required here but do state the
methods used. The arrival time at the cash kiosk is the
combination of the two EoS times (stop-times) from the
two pumps. This arrival time must be sorted. In MAT-
LAB, an easy way to combine two row vectors a and b
into c is by c = [a b].
(b) Generate random numbers under each heading and com-
plete (by hand again) an event table similar to Table 19.4
until a clock time of at least 500 seconds.
Hint Do a large simulation and then find out how many
are actually needed for a clock time close to and above
500 seconds.
(c) Assume that in the motorway problem, each customer
pays between 2 and 18 in a uniform distribution. Cal-
culate the total income of first half-hour. Suppose that
a promotion costing 70 would change the mean inter-
arrival time to 80 seconds and 130 seconds for the mo-
torway and side road pumps respectively.
Decide if the promotion is worthwhile.
Run your M-file n times (where n > 10) and work
out the probability P that your decision above is correct;
(P ~ k/n if the same decision occurs k times in n trial
runs).
Hint For this final question, you should use a variable
seed for generator rand. The following may be used
>> use-seed = sum(100*clock);
>> rand('seed', use-seed)
before each run of your M-file.
 Exercises 255

19.4 For the Leo's cafeteria example of 19.3 using ml = 44, m2 = 55,
m3 = 36, m4 = 100 and dl = 145, d2 = 108, d3 = 69:
(a) Generate 100 arrivals for each of the three main businesses
and 50 arrivals at the rear door of cafe. Then work out
how many customers out of each 100 actually arrive at
the cafe. Show all four kinds of arrivals (in minutes) at
the cafe on the same graph (using different colours and
line patterns).
(b) In (a), what is the total number of customers arriving at
the cafe. Plot arrival time (in minutes) versus service
time (in seconds) at the cafe. Further show arrival time
(in minutes) against waiting time (in seconds) on another
graph after calling queue. m. Find the maximum waiting
time (or the peak of graph) and mark this point by the
symbol `o' on this second graph.
(c) Following (a) and (b), how many customers of each kind
of arrival have arrived at the cafe within the first 40
minutes. What is the total number of customers in the
first 40 minutes? Suppose that the amounts of money
customers spend in the cafe satisfy a normal distribu-
tion with a mean of 1.49 and a standard deviation of
30 pence. Calculate the takings in the first 40 minutes.
What is the profit margin if the running cost of the cafe
is 90?
Hint 1 For the cash point times, as in 19.2, let customers pay by
cash ifs = (0<=r & r>=0.9), and by credit card ifs =
(0.9 < r), where r is the normalised random vector.
Hint 2 The MATLAB command nnz counts the number of nonze-
ros in a vector. Also, the command sort may be used to
pick up all nonzeros in a vector as in
>> arr = [ 2 0 1 5 0 0 4 0 9]
>> arr_1 = -sort(-arr), number = nnz(arr)
>> arrival-2 = sort( arr_1(1:number) )

giving number = 5 and

arr 1= 9 5 4 2 1 0 0 0 0
arrival-2 = 1 2 4 5 9
 20
Fish Farming

Aims of the project

You are given data on the current distribution in size and age of a
particular species of fish on some fishing banks. You are invited to help
establish a sensible fishing strategy (size of net mesh and frequency of
harvest) so that the stocks of fish can be maintained at a viable level.
As part of this, you are required to develop a model for the growth rate
of fish.

Mathematical ideas used

Leslie matrices (Chapter 2) are used to describe the evolution of the
age distribution and differential equations (Chapter 7) to describe the
growth of the fish. Least squares fitting (Chapter 5) is a further optional
technique.

MATLAB techniques used

Much of the above analysis can be accomplished with the M-files listed
below. You may have to modify some of these in the course of your
study. In each case typing help will give information on the pur-
pose and usage. You will also make use of standard MATLAB com-
mands such as eig (eigenvalues), plot and bar (bar graphs).
fishy . m - loads fish data
leslie.m - example of population evolution (see Chapt 2)
lmf i sh. m - basic Leslie matrix for fish population
fodesol.m - graphical and numerical solution of first order ODE
mparft. m - multi-parameters least squares fit
mparst4.m - example set up for the above
resid4.m - residuals function for the above
f ishdat . m - uses fishy.m to load data for above fit

257
258 Fish Farming
20.1 Preliminary look at the problem
Use the command fishy to load some data on the fish. Typing help
fishy will show you what is there. You could use the usual plot com-
mand to have a look at how the weight of a fish and its fertility depends
on its age. Similarly you can make a plot or a bar graph of the current
numbers of fish versus age. (Type help bar.)
A first attempt at a Leslie matrix (cf. Chapter 2) has been provided.
Type help lmfish to see what it is supposed to do. To see the whole
M-file in more detail use Notepad. You will find that several additions
and corrections are required, for example to take account of the fact
that the egg-production (fertility) rate refers only to female fish and
to model more realistically the removal rate due to fishing. The latter
clearly depends on the mesh size of the fishing nets, the size distribution
of the particular fish species and the frequency of harvesting. Part of
this procedure will require you to establish a mathematical model which
fits the weight vs age data and also predicts how the cross-section of a
fish depends on its age.
We will return to each of these tasks in due course. In the meantime,
it is worth trying the first two exercises at the end of the chapter.

20.2 Models of fish growth

There are two simple models which you are invited to consider. Further
details of these can be found in [3]. The predictions of each should be
compared with the given data on fish weight as a function of age. First,
we define a few variables and parameters which will come in handy:
m(t) - mass of a fish at time t
V (t) - volume of a fish at time t
t - age of the fish in years
p - density of a fish in g/cm3 (pure water is 1 g/cm3)
h, w, 1 - height, width and length of a fish (cm)
(Note: in colloquial usage, the terms mass and weight are often used in-
terchangeably. Actually the weight and mass are numerically the same
only if the former is expressed in units of g.)

You should be able to estimate a value for p simply from the fact
that the fish swim in water with no apparent difficulty. Specifying the
dimensions h, w and 1 pre-supposes a very simple rectangular box model
 20.2 Models of fish growth 259

of a fish. You may be able to do better. The box model implies that the
volume V = hlw. In any case, m = pV. We will return to these points.

20.2.1 Gompertz growth model

In this model, the rate of growth (increase in mass) is proportional to the
amount of tissue doing the growing but with the factor of proportionality
decreasing exponentially with time. This models an aging process and
gives
d
t = re-Atm, (20.1)

where the parameters r and A (both > 0) have to be determined to fit the
circumstances. Clearly r sets the overall rate of growth and A determines
the time-scale over which the period of rapid growth takes place. You can
estimate 1/A for our fish very roughly by looking at your plot of weight
gain to see at what age the maximum rate of growth occurs. Likewise
you can estimate r very roughly by applying the formula (20.1) at a
couple of t-values, where m and its slope can be estimated. Make a
rough guess of each in this way and use fodesol to see how the model
behaves. Note that you have to supply an initial condition. Choose a
t-value for which you know the corresponding m-value (t = 1?). You
can adjust the parameters in the light of what you see.
However, you might well find you can solve the ODE (20.1) by inte-
gration to obtain an analytic solution for m(t). A good plan would be
to do this and then use fodesol with suitably modified fnxt to obtain
a plot and numerical solution as a cross check. In this way you can be
sure of correct answers. If unsure of fodesol, you can have another look
at Chapter 7.

20.2.2 Von Bertalanffy growth model

This more sophisticated model, here written for the increase in volume
rather than mass, gives

dt =aV3 -/3V, (20.2)

where the parameters a and 0 are positive. The idea behind it is one of
energetic balance. The first positive term represents energetic (nutrient)
input and is proportional to surface area (hence V23), while the negative
260 Fish Farming
second term represents the drain on resources, that is, food for body
tissue, and is proportional to the volume itself. Again, you should try to
solve this both graphically/numerically with fodesol and analytically
where you may find the change of variable L = V 3 helpful. Here, L
obviously has the dimensions of length. To estimate the parameters in
this case think about the maximum size achieved when dV/dt = 0. You
can also look at spot values of the slope of the data if this helps. (Use
two neighbouring points.)

20.2.3 Choosing the best model and parameters

Using the above methods you should be able to arrive at a reason-
able set of parameters to describe the fish growth data [fish (: , 2) vs
fish (: , 1)]. If you actually manage to obtain correct analytic solutions,
you can go much further and try a `least squares' best fit of the data.
This will show you which growth model is better and also determine the
best fit parameters for it.
The idea of least squares was discussed in Chapter 5 within the con-
text of linear least squares. The same idea of minimising the residual
function R2 with respect to the choice of parameters can be applied to
any parametrisation provided you have a means of minimising. MAT-
LAB is able to do this (of course) and appropriate M-files to perform
a multi-parameter fit mparst4. m and mparft . m have been set up to do
this for you. Type mparst4 to set up the example needed here, followed
by mparft. This particular M-file sets up a fit of the mathematical
model
m = aebVt + c (20.3)

to the growth data. When prompted, try giving [1 1 1] as suggested

starting values for the parameters [a b c]. As you will see from the fi-
nal plot and quoted residual R2, the fit is rather poor! You could now
modify the fitted function in resid4.m to reproduce, in turn, your two
model parametrisations (solutions to equations (20.1) and (20.2)).
The best parametrisation can be identified either by eye, if the dif-
ference is marked, or by looking for that with the smallest residual R2
which measures the sum of the squares of the differences between the
model and the data.
Now try the exercises concerning the growth models and their param-
eters.
 20.3 Designing the Leslie matrix 261

20.3 Designing the Leslie matrix

You may need to review the structure and use of Leslie matrices as
illustrated in Chapter 2. The M-file leslie.m is a simple example. The
file lmfish.m is provided as a starting point for further development
according to your needs. It already incorporates the following features:

(i) It uses the egg-production data in fishy.m and the additional

information that only 1 in 5000 eggs survive to give a one year
old fish. However it needs altering to take account of the fact
that the egg-production rate supplied by fishy.m is per female
fish. You should correct that point now.
(ii) It also incorporates a non-survival probability, different for each
age group. The probabilities in here are just more or less arbitrary
numbers (less than one of course) and increasing since we expect
older/larger fish to be more likely to be caught by a fishing net.
(iii) It has the key feature of a Leslie matrix that the survivors from
one age band grow older and so populate the next age band.
(iv) No fish survives beyond 10 years of age!

These features are annotated in the M-file. Your next task should be
to improve the survival probability model to incorporate your knowledge
of the size of a fish of a given age and to take account of different mesh
sizes.

20.3.1 Cross-section of a fish

At this point you may have a good model which tells you the mass and
hence volume V at any arbitrary time t. If you didn't succeed in getting
a very reliable model for that, you can still proceed by using the data
points fish (: , 2) themselves (modified by the density if necessary) since
you will normally only need V (t) at integer values of t anyway.
To estimate how easily the fish can go through a net of mesh size d
cm, say, you need to relate V to the transverse dimensions of the fish.
The simplest possible model is to approximate the fish by a rectangular
cuboid. You then need to make some assumption about the shape, that
is, ratio of length to height (l/h) and width to height (w/h). From this
you can deduce h, say, from V. Clearly, if h > d the fish is caught. You
may have a better model for the fish shape. If so, try to establish a
formula for its maximum `height' in terms of its volume as shown above.
262 Fish Farming
20.3.2 Probability of getting caught
Even if h < d, the fish can become snarled up in a moving net. The best
idea is to dream up a formula for the probability that a fish of height
h will get caught in a net of mesh size d. For example, the probabilty
function
Pc = min{1, (h/d)2} (20.4)

means certain capture for fish bigger than the mesh size and a decreasing
likelihood for smaller fish. One might justify using (h/d)2 rather than
some other function by claiming that the cross-sectional area rather
than a linear dimension is significant here. There is a further modelling
feature which you should include and that is the efficiency of trawling
(harvesting) in any one year. Even if a very fine mesh is used, the
total catch depends on how often and how completely this is done. You
could combine these factors into one `efficiency' factor fe f f with which
to multiply Pc before going on to calculating the probability 1 - P of
not getting caught in a given year.
You should adapt lmfish.m to incorporate as many of the above fea-
tures that you can. The appropriate places to insert these additions
are annotated in the M-file. Add your own comments (lines beginning
with %) as you add more features. This helps to remind you what you
have done. Since you will want to use this code repeatedly to obtain
and study the Leslie matrix H, it might be as well to get the M-file to
prompt you for the necessary values of d and feff. You can do this with
a line such as
d=input(' Give me the mesh size d ');

which will do just that and make the input value of d (10 say) available
to subsequent lines in 1mf i sh. m. In this way you can keep reusing the
M-file without having to edit it every time you need a new value!

20.4 Fishing strategy

If you understood the Chapter 2 material on Leslie matrices very well,
you will recall that the eigenvalues of the matrix determine the long term
fate of a population. If the largest is greater than 1, it will prosper. If
it is less than 1, the population will eventually die out. This gives you
a quick way of determining whether harvesting each year (multiplying a
population vector by H) will eventually wipe out the stocks of fish. You
 Exercises 263

can check explicitly what happens in the first year by taking the current
population, for example
>> p0=fish(:,3)

and performing one year's fishing by forming

>> pl=H*pO
and so on forming [p0 p1 p2 p3...]. You might like to display these
on a single plot to get an idea of what is really happening. You will find
that the fate of the fish stocks depends crucially on the net size d and the
frequency of fishing in the year fef f. There are several possibilities for
establishing a sensible fishing policy. You can keep fishing at a constant
low level with a sensibly chosen net size so that the younger fish have a
chance to grow. Alternatively you could conduct heavy fishing for a year
or so then give the stocks a rest. You could model this by having two
different Leslie matrices Hl and and Ho which are multiplied in different
sequences according to your strategy. Again, you can check the overall
fate of the population by looking at the eigenvalues of the products.
There are many things to experiment with. The fishing industry is
not directly interested in the numbers of fish swimming about, but in
the maximum biomass (total mass of fish) which can be removed over
an extended period. Ideally, your strategy should be to maximise this
yield. You should work out how to extract the numbers caught each
year in each age group. Perhaps the simplest way is to adapt a version
of lmfish.m to produce a diagonal matrix which, when applied to the
population of a given year, gives a vector of caught fish. Multiplying
(.*) by the vector of fish weights gives the vector of biomass caught.
There are also many imperfections in such a model which you should
think about. For example, if you set d = 100 and fe f f = 0 the model
suggests an exponentially growing population of fish. Of course, this
does not actually happen since the food supply is finite and disease can
set in. If the population is kept under control, however, one may be able
to ignore these limitations.

Exercises
20.1 Use the data loaded in with fishy to construct a bar graph of
the total mass of fish (the biomass) as a function of age.
20.2 Find the total biomass (in some sensible units) of fish of all ages
currently swimming about in one square kilometre of sea.
264 Fish Farming
20.3 Estimate the fish density p (in g/cm3) and hence obtain a plot
of the height, width and length of the fish as a function of age
according to the rectangular `box' model of a fish.
20.4 Solve equations (20.1) and (20.2) by the method of separation
of variables.
20.5 Use fodesol to obtain numerical solutions of equations (20.1)
and (20.2) with guessed values of the parameters. By comparing
with the data for fish mass (or volume), try to find better values
of the parameters.
20.6 If you managed to obtain (by integration) closed expressions for
V(t) for either of the models, use mparft to obtain `best fit'
parameters and compare these with those which you estimated
with the help of fodesol.
20.7 Give your estimate of the parameters [mo, r, A] or [mo, a,,3] of
the model (20.1) or (20.2) which, in your opinion, best describes
the growth data for fish. Also give an estimate of the maximum
weight and length which one of your model fish could achieve.
20.8 Estimate how the cross-section of a fish varies with age.
20.9 Adapt the Leslie matrix lmf ish . m to incorporate the effects of
net size, fish size and fishing efficiency as discussed above.
20.10 Give an example of a viable fishing strategy (net size and effi-
ciency factor which will yield a good harvest but conserve stocks
after five consecutive years fishing. What annual biomass does
your harvest yield from each square kilometre?
20.11 Comment on alternative strategies in which years of heavy fish-
ing are interspersed with recovery years. Obtain some numbers
or plots to support these comments.
 21
Epidemics

Aims of the project

You are given some data describing the development of epidemics which
occurred in various communities and are provided with a basic model
describing the dynamics of such a system. You are then invited to
analyse the data as best you can to discover the underlying behaviour
of the disease and the response of the community to it. The model is
based on a set of coupled first order differential equations. You must
obtain approximate analytic solutions and full numerical solutions using
the routines provided.

Mathematical ideas used

You will need to: work with coupled first order differential equations
(Chapter 7). make linear approximations; know how to integrate sim-
ple linear first order equations; understand the least squares fit idea
(Chapter 5);

MATLAB techniques used

The numerical techniques for differential equations are those first intro-
duced in Chapter 7. The multi-parameter least squares fit package is
that first described in Chapter 20.
For convenience, here is a list of relevant M-files, both standard ones
and special ones provided for this particular project. You may have to
modify some of these in the course of your study. In each case typing
help will give information on the purpose and usage.

265
266 Epidemics
f ludat. m - data for school flu epidemic
plagdat. m - data for Bombay plague
colddat. m - data for island cold epidemic
sirepi. m - SIR epidemic model integrator
sirfn.m - derivative function for the above
mparft. m - multi-parameter least squares fit
mparst3. m - example set up for the above
resid3.m - residuals function for the above
lagsum.m - cumulative sum

21.1 Preliminary look at some data

Use the command f ludat to load some sample data, in this case the
number of boys infected with flu as a function of time in days during an
outbreak of the illness in a public school. You can regard the school as
a closed community of N (= 763) individuals with initially perhaps just
one infected pupil (I = 1) and S = N - 1 = 762 potentially `susceptible'
individuals. As time progresses, the number I of infected pupils rises but
not indefinitely because there are fewer left still susceptible to infection.
The ones who have successfully recovered have now got some, at least
temporary, immunity and so can be thought of as 'removed' from the
system (R). This classification of the N pupils into these three classes
of individuals

N = S(t) + I(t) + R(t) (21.1)

is used in the so-called SIR model described in the next section.

The M-file plagdat. m also loads some data, this time for the number
of deaths per week report during a plague epidemic in Bombay in 1905
and 1906. Unlike the above case, the class of `removed' individuals R(t)
is unfortunately due to death rather than acquired immunity. The net
result from the point of view of model building is, however, the same:
they take no further part in receiving or promoting infection. The datum
loaded is, roughly speaking, the rate of removal (death), that is dR/dt.
The third data sample colddat is similar to the first two data samples
and concerns the spread of a common cold outbreak within the popula-
tion of a remote Atlantic island following one of the infrequent visits of
a supply ship. This time the numbers of new cold cases each day were
recorded. The total island population was 280.
 21.2 The SIR model for the dynamics of an epidemic 267

21.2 The SIR model for the dynamics of an epidemic

This model, due originally to Kermack & McKendrick in the 1930s, uses
first order coupled differential equations to describe the evolution of the
populations S, I and R identified as above, hence the model's name.
For an introduction to such models of epidemics see, for example, [10]
or [3]. Since the three variables sum to a constant (see equation (21.1)),
only two of the following equations are actually independent:

dS
-rSI , (21.2)
dt
dI
dt
rSI - aI , (21.3)
dR
aI. (21.4)
dt

To recap:

S(t) is the number of `susceptibles', i.e. the number in the community

of N ready and waiting to be infected with the disease at time t.
I(t) is the number who actually have the disease and are capable
of infecting others (the susceptibles) and are no longer susceptible
themselves.
R(t) is the number `removed' from the system because they are nei-
ther infected nor susceptible since they have either recovered, and so
acquired immunity, or, in severe cases, are dead.
The origin of the simple model represented by the right-hand side of
these equations is very simple to understand in the light of the above
examples. Look at equation (21.4). This says that the rate at which
people are removed (die or recover) is proportional to the number of
cases I. Here
a = removal rate of infectives (a > 0) . (21.5)

The second term in equation (21.3) reflects the corollary to this state-
ment since these same individuals are being lost from class I. The first
term in this equation is the statement that the number of new cases
per unit time is proportional both to the number of potential victims
(susceptibles) S and to the number already having the disease I and so
capable of infecting others. The constant of proportionality is
r = infection rate (r > 0). (21.6)
268 Epidemics
Finally, equation (21.2) represents the corollary to this statement since
these same individuals are being lost to class S.
Any particular model is now fully specified by the infection and re-
moval rates a and r and by a set of intial conditions. Usually, one takes
R(0) = 0 since no one has yet had the chance to recover or die. Thus,
a choice of 1(0) = Io is enough to define the system at t = 0 since then
So = N - 10. Typically, one takes t = 0 to be the moment at which
I = Io = 1 if the epidemic is triggered by a single infected individual.

21.3 Studying the behaviour analytically

The first thing to notice is that dS/dt < 0 at all times so that S(t) is
monotonically decreasing whereas
dI/dt = 0 when S = p (21.7)
and so I(t) can have a stationary point at some time t,,,ax. This is easily
seen to be a maximum since I(t) is usually rather small initially. Here
p = a/r (21.8)
is called the relative removal rate. Thus, if So < p, condition (21.7) can
never be satisfied and the infection dies out. There is no epidemic. The
size of p is therefore a key parameter in the model. The definition of an
epidemic is just that 1(t) > Io at some later time t.
One can partially solve the system by eliminating t as an explicit
variable and solving the resulting differential equation. For example,
from equations (21.2) and (21.3)
dI
1+S, (21.9)
dS
which you should be able to show gives the solution

R=ln(So) (21.10)

In general, you still need to solve the system numerically to get the t
dependence. However, in certain circumstances, you can make approx-
imations to progress further analytically. For example, if you put the
exact result (21.10) back into the differential equation for R(t) (21.4),
you get
dR =a(N - R - Soe P ). (21.11)

As it stands, this requires numerical solution. For example, we could use

 21.4 Analysing the data 269

fodesol.m and the techniques of Chapter 7. However, when R/p < < 1
the exponential can be expanded to, say, second order in R giving a
separable variable equation which is tractable. To do the integration,
one can use `completing the square' to get something of the form
dR
- adt . (21.12)
A2 - (R - R)2
You should try this for yourself (see the exercises). The result is an
explicit function R(t) which is then an approximate solution to equation
(21.11).
If you then calculate dR/dt you get an expression which could, for
example, be used to compare directly with the plague data loaded with
plagdat. m. The result is of the form
dR
= Aa sech2 (at +,3). (21.13)
dt
Equation (21.10) can also be used to study the final outcome of the
epidemic, when I becomes effectively zero as t - oo. In that case,
equation (21.10) can be rewritten

S(oo)/So = exp - (N - S(oo))

p
(21.14)

which could be solved numerically, for example using the MATLAB func-
tion f zero. If we happen to know the final number of susceptibles and
the total community involved, we can use this to estimate p. Equation
(21.10) can also be combined with equation (21.7) to get further rela-
tions between the parameters and the height of the maximum epidemic
Ima,x. There are lots of different ways of using such relations depending
on the type of data available.

21.4 Analysing the data

In the case of the plague data, you don't know much about the model
parameters - even the size of the relevant Bombay community. How-
ever, assuming that R/p is not too big, you can test the model via its
prediction (21.13) by trying to find parameters A, a and 0 which fit the
data in the sense of a least squares fit. The least squares fit described
in Chapter 5 involved a straight line fit. Here, and in Chapter 20, the
parametrisation is more complicated but the same principle applies: the
270 Epidemics
best fit is one which minimises the sum of the squared `residuals' be-
tween theory and data. The M-file mparft . m does this. You should
type mparst3 to set up the example provided which attempts to fit a
parametrisation

dR (t t )2
Ke 26 (21.15)
dt

to the plague data. You then type mparft to start the least squares fit
and follow the prompts. The `residuals' are calculated in an M-file called
resid3.m. When you have tried this successfully, you should go on to
modify a copy of resid3.m to implement the model expression (21.13).
A full numerical solution of the SIR model is accomplished using the
M-file sirepi. m. You should try to find parameters a and r which give
a reasonable description of the school flu data. In order to get rough
estimates of the sort of values to try, you can study the small t and
large t data separately. At small t, before R becomes large you can
approximate S by N - I and equation (21.3) by

dI
dt = rNI. (21.16)

By solving this and comparing the result with the small t data, you can
get an estimate of r. Similarly, at large t when the epidemic has run its
course, S and I are both small so that dI/dt is dominated by the last
term. Again, you can integrate this equation and get an estimate of a
this time. You can use these rough values to get started with sirepi.
In the case of the Atlantic island data, colddat.m provides you with
the number of new cases each day which is clearly something to do with
the infection rate. To get the actual number of infectives I (t), you need
to know something about how long the infection lasts. This will vary
but a reasonable model might be to take it to be, say, 7 days. From this
you can deduce, not only I(t), but R(t) and hence S(t) as well. You
can manipulate the data by hand or you can use MATLAB's facilities
to help you. An M-file lagsum.m is provided which may be of some use.
Given a complete breakdown of the population into the categories S, I
and R, you can test the SIR model in some detail. By comparing with
the SIR model predictions you can decide whether your cold infection
period of 7 days was too long or too short and so on.
 Exercises 271

Exercises
21.1 Check the nature of any turning value given by equation (21.7)
using I (tmax)
21.2 Derive equation (21.10) from equation (21.9) using the initial
conditions discussed at the end of 21.2.
21.3 Integrate equation (21.12) using the standard integral
1 _ 1

f a2 - x2 a
tanh-1(x)
a
(21.17)

to obtain equation (21.13) and try to find expressions for A, a

and 0 in terms of the model parameters a, r, N and So.
21.4 Find parameters which give a reasonable description of the school
flu epidemic data assuming that I = 1 at t = 0 and the total
number of boys in the school is N = 763.
21.5 Estimate R/p at the height of the flu epidemic?
21.6 How many boys have escaped the flu by the fifteenth day?
21.7 Find the parameters A, a and /3 which best describe the plague
data using the approximate expression (21.13).
21.8 Adapt a copy of sirepi.m to include a plot dR/dt and so try
to find a complete model (N, r and a) for the plague data. You
will have to make some choice of the total community involved
in the Bombay plague.
21.9 Compare the parameters found in this model with those from
the best fit expression found previously.
21.10 Convert colddat. m to supply a table and plot of estimates of
S(t), I(t) and R(t) for the Atlantic island cold data. Estimate
R(oo).
21.11 Find reasonable SIR parameters to describe the Atlantic island
cold data. Comment on the success or otherwise of your mod-
elling attempt.
 22
Dynamics of Snowboating

Aims of the project

You are given a mathematical description of a winter sports area and
invited to help design a run for the new winter sport of `snowboating'
- sliding down mountainsides in a rubber boat, more or less out of
control. You will have to obtain reasonable model parameters to define
safe operating conditions.
Mathematical ideas used
You are given a fairly realistic dynamical model based on Newton's laws.
The equations of motion are coupled differential equations.
MATLAB techniques used
You can solve the differential equations using the numerical methods and
MATLAB programmes of (Chapter 7). For convenience, there follows
a list of relevant M-files, both standard ones and special ones provided
for this particular project. You will have to modify some of these in the
course of your study. In each case typing help gives information on the
purpose and usage.

topog - 3D and contour plots of slopes

snowsl - definition of the topographical surface
snowboat - solves the equations of motion for the snowboat
snbtfn - derivative function for the above
fsnowO - surface function f (scalar args.)
fsnowl - ff and fy
fsnow2 - f.,y and fyy
snowmn - minimisation function for the surface
snowmx - maximisation function for the surface

273
274 Dynamics of Snowboating
22.1 Preliminary look at the problem
Use the command topog to show a mesh plot followed by a contour plot
of the snow hills where the action will take place. It is defined by
z = f(x,y), (22.1)

where f is some smooth function which can be examined in either

snows 1. m or f snow0 . m. The M-file topog. m also allows you to find
out where the hill tops (maxima) and bottoms of the valleys (minima)
are. It uses the standard M-file fmins to minimise -f and f to find
local turning points. Actually, you are provided with the first and sec-
ond order partial derivatives (see f snows . m and f snow2. m) so you could
use these to find the turning points analytically. However, if you look
inside them, the derivatives are a bit long and messy and so the general
numerical minimisation routine fmins is more convenient.
Your first task should be to find and note down some of the geo-
graphic features. In the absence of any frictional forces, a snowboat (an
unsteered toboggan) set going from the highest point will head down
the path of steepest descent gathering speed continually. Since in that
case the total energy is conserved, the speed v at any moment can be
calculated from the loss in potential energy:

E = Ep + EK = constant, (22.2)

where
1
Ep = mgz, EK = 2 mv2 (22.3)

You can work out what speed the snowboat will be doing when it
drops to its lowest point - a totally unrealistic and dangerous speed. Of
course, in real life, there are resistive forces:

friction between boat and snow, which will be more or less constant;
air resistance which rises with speed.

Part of your task will be to obtain reasonable values for the parameters
describing these forces. If necessary, you can increase the friction by
roughening the underside of the boat and so obtain a slower ride. The
other major task will be to find a safe starting point and initial velocity
to give an exhilarating, but not life-threatening, ride down the mountain.
Before continuing, try the first three exercises at the end of the chap-
ter. These will help you visualise the problem.
 22.2 The equations of motion 275

22.2 The equations of motion

First, we define some key constants and variables.
m - mass of the snowboat: take as 150 kg including one passenger;
A - cross-sectional area: say 1.5 x 0.5 m2;
g - acceleration due to gravity: 9.81 m/s;
- coefficient of sliding friction between boat and snow;
k - coefficient of air resistance (Fk = kv2);
r - position vector of snowboat;
F - total external force on the snowboat;
F9 - gravitational force;
F - resistive force due to sliding friction;
Fk - force due to air resistance;
N - normal reaction from the snow surface.
According to Newton's second law,

mr=F=F9+F,,,+Fk+N, (22.4)

where the forces acting are shown schematically in Figure 22.1.

F9

Fig. 22.1. Forces acting on the snowboat (not to scale).

While the snowboat remains on the slope, the motion is described by

equation (22.4) together with equation (22.1). If it goes over a bump or
cliff and leaves the ground, it will behave as a projectile in a resisting
medium subject only to equation (22.4) with F. = N = 0. For sliding,
276 Dynamics of Snowboating
the following is a reasonable model of friction

FN, -iri INS. (22.5)

Notice that friction always acts in the opposite direction to the sliding.
This simple model therefore breaks down whenr = 0 where the force
changes discontinuously! The numerical solution method described in
Chapter 7 could have problems with this. Fortunately, we are not inter-
ested in following the snowboat after it has come to rest (r = 0)! It will
be left to you to determine a realistic value for p. (See the exercises.)
For the air resistance, it is found that cars, people on bicycles, free-fall
parachutists etc. experience a force like

Fk = -krli I . (22.6)

Again, this always opposes motion but has reasonable behaviour as the
velocity goes to zero. You should be able to figure out a sensible value for
k using the experimental information that a free-fall parachutist weigh-
ing the equivalent of 70 kg, in a spreadeagled pose, reaches a terminal
velocity of around 54 m/s. To convert his/her k-value to that for a
snowboat, you will need to estimate the frontal area of each and use
the fact that the force is proportional to the area presented. To get the
terminal velocity equation, just use the free-fall (projectile) equation de-
scribed above and study it whenr = 0 since the parachutist is no longer
accelerating when terminal velocity is reached (by definition!)
The gravitational force is easy to model:

F9 = mg(0, 0, -1) (22.7)

using the same coordinate choice used to describe the snowfield.

Note that N r = 0 since N is normal to the surface and r, the velocity
vector, is tangential to the surface. If you have met some elementary
vector calculus, you may know that the normal to the surface (22.1) is
parallel to the grad of the function 0 = z - f (x, y). This gives an easy
way of finding it numerically.
Newton's equations (22.4) therefore represent a set of three coupled
second order differential equations where only two coordinates (x and y
say) are independent because of (22.1). If we eliminate z, z and z using
this, we get two second order coupled equations in x and y. Using the
trick learned in Chapter 7, these can be reexpressed in terms of four first
 22.3 Exploring the operating parameters 277

order equations in x, y, u and v, say, where

u=x, v=y.
The first order equations are
u _ q-1 ( -- Irlu+qxg l (22.8)
C l-\v C v/ \ -mIrly+qy9 /
where
A
_ 1 - gxfx -gxfy )
-gyfx 1 - gyfy
Here,
k
g=g+fxxu2+2fxyuv+fyyv2+ -ri(fxu+fyv),
m

qx=px-,p z
z
px=nx - Ax

and (nx, ny, nz) is parallel to grad 0, the surface normal.

You may wish to verify these equations but be ready to do quite a bit
of calculus and algebra.
The M-file snowboat. m solves these equations numerically in the same
manner as the examples studied in Chapter 7. The function snbtfn.m
simply contains the derivatives as expressed in equation (22.8).

22.3 Exploring the operating parameters

Having found out where the hill tops are, made a reasonable estimate of
the air resistance parameter k and guessed a trial value for (the coef-
ficient of sliding friction.), you can use snowboat to check out a sensible
run. Notice that if you start on a relatively flat area you will need quite
a large downhill push (large VI-U-0 + v0) to get going. Remember that
the friction model is not very reliable at low speeds on gently sloping
ground where the snowboat is liable to grind to a halt.
A useful way to proceed is to choose the surface function f in equation
(22.1) to be that for a simple sloping plane. In this case, you can solve
the equations analytically and check out the performance of the numer-
ical procedure. To do this, make new copies of snowsl. m, f snowsO. m,
f snows 1. in, and f snows2. m corresponding to the choice of plane, say
z = f(x,y) = y/2. (22.9)

The partial derivatives are not hard to work out! The problem becomes
278 Dynamics of Snowboating
one dimensional since everything can be written in terms of, say, y.
To solve the single second order ODE, the best method is to use v
and y rather than y and t as the dependent and independent variables
respectively, using

y = vv . (22.10)
y
In this way, you can check the validity of much of the formalism and see
how well the numerical procedure is performing.
The M-file snowboat. m prompts you for the parameters y and k, the
starting velocity (horizontal components) uo and v0 and a time interval
6t over which to integrate numerically. Note that you control the start-
ing direction with u0 and v0 making them negative if necessary. The
subsequent direction of motion is determined by the topographical fea-
tures of the snow slopes via the equations of motion (22.8). Ideally, you
want the run to start near the top of a hill and grind to a halt near the
lowest point within the area. Excursions into neighbouring valleys are
expensive.

Exercises
22.1 Print out, or make rough sketch copies of the 3D-view plot and
of the contour plot provided by topog.
22.2 Give the locations of all hill tops and bottoms of valleys within
the area which is 2 km x 2 km. Give the altitudes as well.
22.3 Sketch on the plots a couple of paths suitable for snowboating.
22.4 Quote your estimate of a reasonable value for k based on the
free-fall parachutist data and your estimate of the relative frontal
areas involved.
22.5 Modify the surface function routines to describe the sloping
plane (22.9) as discussed above. Use snowboat to describe mo-
tion down the plane from the point (0, 2000) with initial velocity
(0, -2) m/s and compare your results with the exact answers got
by integrating the equation analytically.
22.6 If, instead, the sloping plane had the equation z = y/10 what
value of a, the coefficient of friction between snowboat and snow,
would lead to zero acceleration down the hill?
22.7 Find a reasonable coefficient of friction , and initial velocity
which allows the snowboat to reach the bottom of the valley
reasonably smoothly starting from the very highest point in the
 Exercises 279

snow hills (use the original versions of snowsl.m, etc.). Give the
initial parameters, the approximate time taken and the approx-
imate top speed reached during the descent.
22.8 Try to find other similar runs from the remaining hill tops. Com-
ment on how the initial velocity might be achieved in each case.
22.9 Modify snowboat. m to display the change in potential energy
and the change in kinetic energy from the initial and final po-
sitions and give the results for the trip down from the highest
hill.
22.10 Give the energy differences for the same run but when the fric-
tion and air resistance are switched off. Comment on the two
sets of results.
 23
Tides

Aims of the project

You are given data consisting of a series of tidal measurements taken
over some period of time and are required to analyse these, draw some
conclusions and make some predictions. Specifically, the data are the
measured heights (in metres) of the sea level taken at the given equally
spaced times (in hours) at a fixed place. You are asked to find such
things as the mean sea level, mean tidal range, mean period between
high tides, estimated sea level at particular time on a particular day etc.
The times given, are with respect to 0 : 00 hours on 1 January 1992.

Mathematical ideas used

Since tidal effects are clearly periodic in nature, you are given some
assistance in learning about Fourier series and their application.

MATLAB techniques used

There are a number of built-in MATLAB routines which can be used to
help with Fourier analysis and manipulating vectors of data:
fourier - demo of Fourier series for square wave;
fftdemo - demo of FFT for signal analysis;
fft - Fourier transform;
ifft - inverse Fourier transform;
spline - interpolation of data points;
mean - mean of data;
max - maximum of data.
In addition to these standard procedures, the following M-files are pro-
vided for this project. You may have to modify some of these in the
course of your study. In each case typing help gives some information
on the purpose and usage.

281
282 Tides
tides.m - load tidal data tm(1:nt), hm(l:nt);
f ourdat . m - load sample signal data;
f ouran. m - Fourier analysis of data;
foursig.m - for checking purposes only;
locmax.m - local maxima of data;
times.mat - vector of tide measurement times;
heights.mat - vector of tide heights;
noisyt.mat - vector of sample signal times;
noisys.mat - vector of sample signal data.

23.1 Preliminary look at the tidal data

Use the command tides to load the tidal information from the MAT-
LAB data files times.mat and heights.mat. Now have a look at the
data. The times (tm) are in hours and the heights (hm) are in metres.
The first few exercises at the end of the chapter should help you to get
a general picture of what is going on.
Having explored these and other features, you will see that the tidal
heights very roughly satisfy something like
h(t) = ho + Asin(wt + 0) . (23.1)

Estimate ho, A, w and 0 as best you can and comment on the reliability
of each of these estimates. To do this, start by estimating A by noting the
`amplitude' of the main oscillation which is half the difference between
maximum and minimum. Then establish 0 by noting where each the
oscillation period starts. You can estimate ho from the water level at
mid-tide.

23.2 Fourier series and methods

If the tides exhibited an exact sinusoidal behaviour as in equation (23.1),
there would be no need for tide tables. In fact, the amplitude A and the
frequency w defined in this way vary with time. Another, more precise,
way to describe this is to recognise that more than one frequency and
corresponding amplitude are present. In general, any periodic distur-
bance such as a sound will contain many modes of oscillation. For an
orchestra this is obvious. However, even the purest note of a top class
singer contains many frequencies besides the `fundamental'. You may
well have seen demonstrations of this using an oscilloscope to analyse
and display the frequency content of a human voice or other sound.
 23.2 Fourier series and methods 283

Mathematically, the total sound or other signal f (t) is decomposed as

a Fourier sum (discrete time sample version) or Fourier integral (contin-
uous time version) over its constituent modes. The Fourier sum looks
like

CO

f (t) _ Ane T (23.2)

n=-oo
where T is the total time interval over which measurements are made
and n runs over all integers. You sometimes see this written in terms of
sin and cos since these are related to the complex exponential:

e ti2nnt
T= cos ( t) f i sin ( Tt) . (23.3)
T
Since sin(2irnt/T) and cos(2irnt/T) have period T/n as a function of
t, each mode labelled by n corresponds to frequency n/T. Conversely
you can write the frequency coefficient An as an integral over t

An = T
fT
f (t)edt . (23.4)

Where the signal f (t) is only known at discrete (but regularly spaced)
times, this integral can be replaced by a sum. This is essentially what
the MATLAB routines fft (fast fourier transform) and its inverse ifft
do. There are two useful MATLAB demos to help you understand how
Fourier sums and transforms work.

fourier: this shows how a periodic `square wave' can be reproduced

by a (infinite) sum of sinusoidal frequencies. Because the shape of the
periodic signal is far from sinusoidal here, you need many sine waves
to reproduce the original signal well. Actually, when discontinuities
are present, the Fourier sum inevitably gives the wrong limit at the
point of discontinuity - the so-called 'Gibbs phenomenon'.
fftdemo: this demo starts by making up a signal which is basically
the sum of two sinusoidal ones (that is, two frequencies) but adds
some random noise to make it look like the sort of thing you might
encounter playing music over a noisy telephone line. The data look
very random when plotted against time (much more so than the tide
data!). However, when you look at the frequency coefficients (Fourier
transform of the data) you clearly see the original sinusoidal compo-
nents. Note that the frequency unit 1 Hertz is just 1 s-1.
284 Tides

It is clear that a similar Fourier analysis will help one understand

what influences the tidal data. Which periods or frequencies dominate?

23.3 Analysis of an electrical signal

As a prelude to analysing the tidal information, some data similar to
that in the fftdemo have been provided (fourdat) together with a basic
analysis M-file (fouran). As usual, typing help f name will give you
some help in getting started. You should identify which frequencies are
present in the original signal data and estimate the relative importance
of each frequency component. To get this, look at the square of the
coefficient of the mode as given by the Fourier transform. You may wish
to play around with a copy of fouran.m to get it to display or plot what
you particularly want to see. You can also arrange to print out hardcopy
of interesting plots.
Estimate which frequencies are present in the original signal. After
you have done this, look in the M-file foursig.m to check your answer
and see how the signal was actually created.

23.4 Fourier analysis of the tidal data

Having understood the basic operation of the M-file (fouran) you can
now adapt it to cope with the tidal data. Edit the load statement to use
the tidal data (times.mat and heights.mat). Then make corresponding
alterations to the rest of the code. Note that the sensible units for
the frequencies and periods in this case will involve hours rather than
seconds. Have you identified which variable holds the `signal' and which
holds the time for the tidal data? In the original version of fouran.m
the signal was in s and and the time in t. To look at the tidal data just
type, for example,
>> clear
>> load times.mat
>> load heights.mat
>> whos
>> plot( ....) etc.

The exercises can be accomplished using a combination of techniques.

These include:

making plots of the data over a variety of ranges if necessary;

 Exercises 285

studying the Fourier analysis plots and the spectrum results them-
selves;
making simple parametrisations of the dominant sinusoidal behaviour
as in equation (23.1);
using spline to interpolate subsets of data.

Exercises
23.1 Check out the time interval between measurements and the total
time covered by the tidal measurements which are supplied.
23.2 Look at the maximum, minimum and mean of the data.
23.3 Plot the height vs time over different time scales - a day, a week
etc.
23.4 Use the function locmax to estimate the heights of the high and
low tides. Are these really the actual high tide values?
23.5 Estimate as best you can, the typical time between consecutive
high tides.
23.6 Find on which days the tidal variation is strongest/weakest.
23.7 Find the dates covered by the given tidal data and the mean
measured tidal height in this period.
23.8 Find the absolute maximum and minimum measured heights
and when they were recorded.
23.9 Estimate the highest tide (whether measured directly or not)
and when it occurred. Do the same for the lowest tide.
23.10 Find the dominant frequency components in descending order
of importance. Comment on the numerical values of the corre-
sponding periods.
23.11 Estimate the average time between consecutive low or high tides.
23.12 Estimate the sea level at 3.10 pm on 13 January and at at 6pm
on 4 March in the year 1992.
23.13 Construct a plot of the rate in metres/second at which the tide
is going out during the afternoon of 22nd January 1992 over
a beach which has a gradient of 1 in 200 (assuming the tidal
variation is independent of the local geography). What is the
maximum speed achieved across the sand and at what time is
it achieved? You should use the parametrisation (23.1) to help
you do this.
 Appendix 1
MATLAB Command Summary

Most commands can be found by using help but here is a quick and
compact summary. For further details, type: help command.

Basic operations

help fname displays help comments at the top of M-file fname

quit quit MATLAB
exit same as quit
type fname lists contents of M-file fname
who lists currently used variables
whos as above but with more detail
clear clear variables from memory
what directory listing of all M-files on disk
which locate M-files
format change display format of results
demo runs some MATLAB demos

Built in values
pi 7r i,j
inf oo ans current answer
flops floating point count clock wall clock time

286
 MATLAB Command Summary 287

Arithmetic and matrix operations

+ add numbers (scalars), vectors or matrices
as for add
multiplication of numbers or compatible matrices, vectors
element-by-element multiplication of same size vectors, or matrices
division of numbers, right division of compatible matrices
element-by-element division of same size vectors, or matrices
left division of compatible matrices
power of number or square matrix
element-by-element power of vector or matrix
transpose
size size of matrix or vector
length length of a vector
sum sum of elements of vector
norm magnitude of a vector

Common maths functions

sin, cos, tan usual trigonometric functions
acos, asin inverse of these
exp, log exponential and natural logarithm
sqrt square root
rand random numbers in [0, 1)
round round to nearest integer
fix round towards zero
abs absolute magnitude of real or complex number
angle phase or argument
real, imag real and imaginary parts
conj complex conjugate

Matrix functions
det determinant of square matrix
eig eigenvalue and eigenvectors of a square matrix
inv inverse
rref reduced row echelon form
rank rank of a matrix
288 MATLAB Command Summary
Graph operations

plot plot vector of points (linear X, Y-plot)

hold on/off superimpose plots
clg clear graphics screen
mesh 3D mesh surface
meshdom domain for mesh plots
contour contour plot
bar bar graph
title add a title
xlabel, ylabel add labels
axis axis scaling
text annotate with text
print print current graph

Function operations

fmin minimum of function of one variable

(to be replaced by fminbnd from MATLAB 5.3)
fmins minimum of function of several variables
(to be replaced by fminsearch from MATLAB 5.3)
f zero find zero of function of one variable
syms declare a symbolic variable (MATLAB 5 symbolic toolbox)
spline find spline function for data vector
quad numerical integration
ode23 2nd/3rd order Runge-Kutta solution of ODEs

Statistics operations

mean mean of data vector

std standard deviation
cumsum cumulative sum
cov correlation between data vectors
min, max minimum and maximum
 MATLAB Command Summary 289

Load/save operations
save save variables to file
load load variables from file
diary fname log all MATLAB operations to file fname
chdir change file directory
dir list contents of directory
 Appendix 2
Symbolic Calculations within
MATLAB

MATLAB provides an interface calling directly Maplet commands for

symbolic calculations. As this topic can be useful for readers, but is not
essential for this book, we here use the same examples that appeared
in this book for illustrations. Note that, within maple, the string is in
Maple syntax. However, MATLAB 5 introduces the command syms to
make it much easier to do symbolic calculations in the MATLAB syntax
(as seen from one example below).

Differentiation and nonlinear solution - Chapter 4

To find g = at for f = x + 2ty - t - 2t3 and solve g = 0 for y, do:
>> maple(' f x + 2*t*y - t - 2*t"3 ;' )

>> maple(' g diff(f,t) ') %''. giving g 2*y-1-6*t"2

>> maple(' h := solve(g,y) ') %% giving h := 1/2+3*t-2

Here the semicolon is only optional as MATLAB will add one to meet
the Maple syntax. With syms, the familiar MATLAB syntax applies:
.............
>> syms x t y %%%'/'/%%%'/%%%'/ Declare 3 symbolic variables
>> f = x+2*t*y-t-2*t"3
>> g = diff(f,t), h = solve(g,y)

Solution of differential equations - Chapter 7

To solve the ODE, a = i - x, try the following
t Maple toolbox for MATLAB is developed and distributed by Waterloo
Maple Software, Inc. See the web site: http://www.maplesoft.on.ca/ or
http://www.maplesoft.com for details of Waterloo Maple.

290
 Symbolic Calculations within MATLAB 291

>> maple(' a diff(x(t),t)=x(t)/t - x(t) ')

>> maple(' b := dsolve(a,x(t)) ')

The output will be (_Cl denotes a generic constant)

b := x(t) = t*exp(-t)*_C1

Integration - Chapter 11
To verify that f 1 , y dy = 2 ( c --yo - c - yl for c = 2, try
>> maple( 'c := 2; a := int(1/sqrt(c-y), y=y_O..y_l)' )

The result will be

a :_ -2*(2-y_1)"(1/2) + 2*(2-y_O)"(1/2)

Solution of linear systems - Chapter 16

To check that the solution of
1 1 0 3 fxl 4
2 1 -1 1 x2 1

3 -1 -1 2 x3 -3
-1 2 3 -1 x4 4

is x = [-1 2 0 1]T, try

>> maple(' A := matrix([[1,1,0,3],[2,1,-1,1],
[3,-1,-1,2],[-1,2,3,-1]])')
>> maple(' print(A) ')
>> maple(' x ''x'' ')
>> maple(' b array(1..4,[4, 1,-3,41)')
>> maple(' x := linsolve(A,b)' )

Note that the first two lines for defining A should be typed in as one line
- here the `breaking' is due to text processing only. The final result
will be
x := VECTOR([-1, 2, 0, 1])
To get more help on Maple, type
>> mhelp % Also maple('help')
>> mhelp solve % Here for help on "solve"
 Appendix 3
List of All M-files Supplied

The following lists all M-files which we provide for the reader via the
internet from
http://www.cup.cam.ac. uk/Scripts/webbook.asp?isbn=0521639204
or
http://www. cup. cam. ac. uk/Scripts/webbook. asp?isbn=0521630789
These are listed under the chapters in which they are first used. Several
are of course reused in subsequent chapters.
To locate the exact page number where an M-file or a MATLAB com-
mand is used in the book, check the index pages under headings `M-files'
and `MATLAB commands'.

Part one: INTRODUCTION

Chapter 1 Chapter 4
none tsine.m
tsine2.m
polyex.m
Chapter 2 goatgr.m
leslie.m goatfn.m
parnorm.m
linenv.m
Chapter 3
cubics.m
fibno.m hypocyc.m
hail.m
gcdiv.m
pow.m

292
 List of All M-files Supplied 293

Chapter 5 Chapter 7
tomato.m fodesol.m
toms.m fnxt.m
diabetic.m odegr.m
marks.m mode23.m
mannheim.m species.m
specfn.m
ode23k.m
Chapter 6 vderpol.m
c6exp.m vdplfn.m
exprand.m diffeqn.m
normrand.m dfeqfn.m
randme.m
unirand.m

Part two: INVESTIGATIONS

Chapter 8 Chapter 11
none zzl.m
slidel.m
slidelfn.m
slide4.m
slide4fn.m

Chapter 9 Chapter 12
gcdran.m mobius.m
primes.m mobiusl.m
psp2.m cobm.m
miller.m cobq.m
matit2.m
matit3.m
Chapter 10
quadn.m
hypocy.m perdoub.m
linenv.m
paramc.m
294 List of All M-files Supplied

Chapter 13 Chapter 16
cnrl.m linsolv.m
cnr2.m chop.m
lu2.m
lu3.m
Chapter 14 lu4.m
randperm.m solv6.m
cycles.m spar_ex.m
rifflel.m
rifflela.m
Chapter 17
remm.m
rifflelc.m intdemol.m
rufflel.m intdemo2.m
rufflelc.m cont4.m
ruffle2.m cont7.m
polyfit2.m
polyval2.m
Chapter 15
full_new.m
Chapter 18
gauss ja.m
f_rate.m See Chapter 7
cont4.m

Part three: MODELLING

Chapter 19 Chapter 20
queue.m fishy.m
exprand.m lmfish.m
normrand.m mparft.m
unirand.m mparst4.m
resid4.m
fishdat.m
 List of All M-files Supplied 295

Chapter 21 Chapter 23
fludat.m tides.m
plagdat.m fourdat.m
colddat.m foursig.m
sirepi.m locmax.m
sirfn.m times.mat
mparst3.m heights.mat
resid3.m noisyt.mat
lagsum.m noisys.mat

Chapter 22
topog.m
snowsl.m
snowboat.m
snbtfn.m
fsnowO.m
fsnowl.m
fsnow2.m
snowmn.m
snowmx.m
 Appendix 4
How to Get Solution M-files

Solution M-files and sketches of some sample solutions are available

through the internet via a password made available to those using the
book as a course text. Course leaders should contact the publisher -
CUP:
for paperback
http://www.cup.cam.ac.uk/Scripts/webbook.asp?isbn=0521639204
or for hardback
http://www. cup. cam. ac. uk/Scripts/web book. asp?isbn=0521630789

296
 Appendix 5
Selected MATLAB Resources on the
Internet

The name MATLAB stands for matrix laboratory. MATLAB was orig-
inally written to provide easy access to matrix software developed by the
UNPACK and EISPACK projects, which together represent the state
of the art in software for matrix computation. MATLAB converts user
commands (including M-files) into C or C++ codes before calling other
program modules (subroutines). Today MATLAB has been developed
to cover many application areas. You may search the internet with the
key word MATLAB to see some evidence.
To learn more about the product developers, check the web page
http://www.mathworks.com/
If you hope to find out more of, or wish to use directly, the LAPACK
and EISPACK programs, check the following
http://www.netlib.org/lapack/
For users who would like a gentle and easy introduction to MATLAB,
try the following
http://www.ius.cs.cmu.edu/help/Math/vasc-help-matlab.html
http://www.unm.edu/cirt/info/software/apps/matlab.html
For novice users who prefer short guides to MATLAB with concrete
and simple examples, try
http://www.liv.ac.uk/CSD/acuk-html/486.dir/486.html
http://www-math.cc.utexas.edu/math/Matlab/Matlab.html
http://www.indiana.edu/-statmath/math/matlab/index.html
Advanced users of MATLAB may check the following pages for more
comprehensive guides to MATLAB

297
298 Selected MATLAB Resources on the Internet
http://www.rrz.uni-hamburg.de/RRZ/software/Matlab/
http://www.engin.umich.edu/group/ctm/
ftp://ftp.math.ufl.edu/pub/matlab/
Even more sophisticated examples may be available on the internet; e.g.
try the following for more M-files
ftp://ftp.cc.tut.fi/pub/math/piche/numanal/
http://www.ma.man.ac.uk/-higham/testmat.html
http://www.tc.cornell.edu/-anne/projects/MM.html
http://users.comlab.ox.ac.uk/nick.trefethen/multimatlab.html
ftp://ftp.mathworks.com/pub/mathworks/toolbox/matlab/sparfun/
The following sites contain technical information on MATLAB and
FAQ (frequently asked questions) pages
http://www.mathworks.com/
http://www-europe.mathworks.com/
http://www.math.ufl.edu/help/matlab-faq.html
 References

[1] H.Abelson and A.diSessa, Turtle Geometry, Cambridge, Mass: MIT

Press, 1981.
[2] J.W.Bruce, P.J.Giblin and P.J.Rippon, Microcomputers and
Mathematics, Cambridge: Cambridge University Press, 1990.
[3] P.Doucet and P.B.Sloep, Mathematical Modelling in the Life Sciences,
Chichester: Ellis Horwood, 1992.
[4] I.S.Duff, A.M.Erisman and J.K.Reid, Direct Methods for Sparse
Matrices, Oxford: Clarendon Press, 1986.
[5] G.Fischer, Mathematical Models, Braunschweig: Friedrich Vieweg Sohn,
1986 (Commentary volume, p.53.)
[6] M.Gardner, Mathematical Carnival, Washington DC: Mathematical
Association of America, 1989.
[7] P.J.Giblin, Primes and Programming, Cambridge: Cambridge
University Press, 1993.
[8] L.M.Hall, `Trochoids, roses, and thorns-beyond the spirograph',
College Mathematics Journal 23 (1992), 20-35.
[9] L.Haws and T.Kiser, `Exploring the brachistochrone problem',
American Mathematical Monthly 102 (1995), 328-336.
[10] F.C.Hoppensteadt, Mathematical Theories of Populations:
Demographics, Genetics and Epidemics, Philadelphia: Society for
Industrial and Applied Mathematics, 1975.
[11] J.F.Humphreys and M.Y.Prest, Numbers, Groups and Codes,
Cambridge: Cambridge University Press, 1989.
[12] N.Koblitz, A Course in Number Theory and Cryptography, New York:
Springer-Verlag, 1987.
[13] J.C.Lagarias, `The 3x + 1 problem and its generalizations', American
Mathematical Monthly 92 (1985), 3-23.
[14] H.-O.Peitgen and P.H.Richter, The Beauty of Fractals : Images of
Complex Dynamical Systems Berlin: Springer-Verlag, 1986.
[15] Y.Saad, Iterative Solution for Sparse Linear Systems, Boston: PWS
Int. Thompson Pub. (ITP), 1996.
[16] A.C.Thompson, `Odd magic powers', American Mathematical Monthly
101 (1994), 339-342.

299
300 References
[17]V.M.Tikhomirov, Stories About Maxima and Minima, Washington DC:
Mathematical Association of America, 1990.
 Index

approximation, 221 parametric, 47

error, 227 plotting, 45
arithmetic, 3 rose, 116
basic operations, 3 sine, 48, 50
arrival time, 80, 247 spirograph, 56, 134
square root, 144
basin of attraction, 166-168 cusp, 56, 126, 135
bits cycloid, 137, 142
calculation of, 42
boxes, unlocking, 176 data analysis, 60
brachistochrone, 137 data set, 60
database, 62
Carmichael number, 38 derivative
Cauchy-Schwarz inequality, 71 partial, 276, 278
proof, 71 determinant, 20
cobweb diagram, 148, 149, 154, 158, difference equations, 94
160, 161 first order, 94
commands second order, 96
repeating, 5 differential equations, 88
complex numbers, 23, 165 autonomous, 91
argument, 24 coupled, 273, 277
conjugate, 24 coupled first order, 265
distance, 165 first order, 90, 91
imaginary part, 24 higher order, 93
modulus, 24 ordinary, 239
real part, 24 digits
congruence, 114, 180 significant, 211
convergence order, 191, 202 using power algorithm, 40
convergence region, 192 disjoint cycle notation, 172
coprime, 108-110 distribution
cubic equation, 54, 168 exponential, 79, 247
double root, 55 normal, 78
curve, 44 random, 247
constant width, 122 uniform, 76, 247
cycloid, 137, 142 distribution function, 85
ellipse, 47, 135 division
epicycloid, 117, 135 element-by-element, 27
hypocycloid, 56, 59, 116
normal, 51 echelon form, 103, 210
parabola, 51, 139 edit a file, 9

301
302 Index
eigenvalue, 22, 105, 106, 153, 155, 215 2D, 228
eigenvector, 22, 105, 152, 155 linear and cubic, 226
envelope, 51, 57, 118, 119, 121, 122 spline, 226
epicycloid, 135 iterations, 189, 196
epidemics, 265 iterative refinement, 212
SIR model, 266
equations of motion, 275 kinetic energy, 137
errors
undefined variable, 5 ladder, 120
Euclid's algorithm, 34 least common multiple, 132, 173
event table, 249 least squares fitting, 63, 260, 270
formulae derivations, 71
Fermat's (little) theorem, 36 nonlinear, 67
Fermat's theorem, 112 least squares method, 224
Fibonacci numbers, 31, 39 1D, 224
files 2D, 233
edit, 9 proof of, 72
save, 9 line
fish, 258 broken, 140
catching, 262 LU decomposition, 210
growth, 258
fixed point, 156 M-file, 9
floating point operations (flops), 208 M-files
Fourier analysis, 284 c6exp, 80
Fourier integral, 282 chop, 211
Fourier methods, 281, 282 curl, 166
Fourier sum, 282 cnr2, 167
fractal set, 170 cobra, 149
friction, 144, 276 cobq, 157, 160
function M-file, 12, 13 colddat, 266
functional, 74 cont4, 199, 230
cont7, 233, 234
garage door, 121 cubics, 54
Gauss elimination method, 210 cycles, 173
Gauss-Jacobi, 191 dfeqfn, 96
global variable, 139 diabetic, 69
goat problem, 50, 55, 58 diffeqn, 96
grad, 276 exprand, 80, 248
grain plot, 90 f_rate, 195, 202
graph, 216 fftdemo, 281
annotation, 233 fibno, 32, 39
connectivities, 216 fishdat, 257
edge, 216 fishy, 257
node, 216 fludat, 266
gravity, 137 fnxt, 90
greatest common divisor, 34, 108, 114 fodesol, 89, 257
three numbers, 40, 110 fouran, 282
fourdat, 282
hailstone function, 33, 39 fourier, 281
hole, square, 122 foursig, 282
hypocycloid, 56, 59 fsnow0, 273
fsnowl, 273
integral, 137 fsnow2, 273
improper, 138 full-new, 190, 197
inter-arrival time, 80, 247 gauss ja, 191, 198
interpolation gcdiv, 36
1D, 221 gcdran, 108
 Index 303

goatfn, 51 solv6, 214

goatgr, 51 spar-ex, 216
hail, 33, 39 species, 91
hypocy, 56, 116 tides, 282
intdemol, 224 tomato, 62, 67
intdemo2, 229 toms, 67
lagsum, 266 topog, 273
leslie, 26, 257 tsine, 48
lin_soly, 211 tsine2, 48
linenv, 53, 119 unirand, 78, 248
lmfish, 257, 261 vderpol, 94
locmax, 282 vdplfn, 94
1u2, 211 zzl, 130
1u3, 211 magic square, 102
1u4, 211, 220 pandiagonal, 104, 106
mannheim, 70 product of, 104
marks, 70 Mandelbrot set, 169
matit2, 153, 155 MATLAB commands
matit3, 153 \, 197
miller, 114 abs, 24
mobius, 146, 147, 154 acos, 19, 51
mobiusl, 146, 154 angle, 24
mparft, 257, 260, 266 arithmetic operations, 287
mparst3, 266 axis, 46, 54
mparst4, 257, 260 axis('equal'), 47
normrand, 79, 248 axis('square'), 47
paramc, 120 axis(axis), 48
parnorm, 51 bar, 33, 61, 257
perdoub, 159 basic operations, 286
plagdat, 266 ceil, 76
polyex, 49 clear, 14
polyfit2, 233 common functions, 287
polyval2, 233 comparison (==), 33
pow, 37, 40, 42 cond, 213
primes, 110, 112 condest, 213
psp2, 112 conj, 24
quadn, 157 contour, 273
queue, 247, 254, 255 cumsum, 61
randme, 84 det, 20
remm, 182 diag, 21
resid3, 266 diary, 14
resid4, 257 eig, 22, 257
rifflel, 181 exp, 17
rifflela, 181 fft, 281
rifflelc, 185 fix, 76
rufflel, 186 floor, 34, 76
rufflelc, 187 flops, 209
ruflie2, 187 fmins, 141, 273
sirepi, 266 for, 31, 32
sirfn, 266 format, 4, 19, 235
slidel, 139 fprintf, 82
slidelfn, 139 fsolve, 142, 206
snowboat, 273 function operations, 288
snowbtfn, 273 fzero, 51, 269
snowmn, 273 get, 232
snowmx, 273 gplot, 216
snowsl, 273 graph operations, 288
304 Index
help, 9, 12 std, 61
hist, 61, 173 sum, 18, 61
hold, 8, 9, 45, 49, 235 surf, 229
i, 23 symrcm, 215
if...else, 33 tic, 209
ifft, 281 title, 8
imag, 24 toc, 209
Inf, 147 type, 10
input, 262 view, 230, 233
interpl, 226 what, 12
interp2, 234 while, 32, 36, 218
inv, 20 who, 6
length, 45 whos, 6
load, 15 xlabel, 8
load and save operations, 289 ylabel, 8
magic, 102 matrix, 19
matrix functions, 287 augmented, 21
matrix operations, 287 determinant, 20
max, 33, 61, 281 diagonal, 21
mean, 61, 281 echelon form, 21
median, 61 identity, 20
mesh, 229, 273 inverse, 20
meshdom, 199 Leslie, 24, 258, 261
meshgrid, 199, 229 magic square, 102
min, 61 multiplication, 20
nnz, 255 nonsingular, 23
norm, 19, 228, 235 permutation, 105, 215
ones, 21 reduced echelon form, 103
pi, 4 rotation, 130
plot, 8, 45, 62, 192 sparse, 215
plot3, 232 matrix powers, 149
polyfit, 49, 55, 65 Miller's test, 38, 113
polyval, 45, 50, 55, 66, 222 minimisation, 71
prod, 61 minimum, 71
quad8, 137, 139 minimum of a function, 141
rand, 41, 75, 76, 78, 79, 85, 214 multiple variables, 67
randn, 79, 86 multiplication
randperm, 77, 172, 174 element-by-element, 7, 18
real, 24 scalar product, 18
rem, 33, 34, 37, 41, 182
roots, 44, 55, 157, 158 Newton-Raphson method, 164, 190
round, 76 1D, 190
rref, 21, 103, 210 2D, 196
rrefmovie, 210 complex, 165
save, 15 divergent, 168
semicolons, 34 nonlinear equations, 189
set, 222, 232 nonlinear systems, 196
sin, 4 normal to a curve, 51
slice, 231 notepad, 10
sort, 61, 172, 179, 255
sparse, 215 order, 181
speye, 215
spline, 281 parabola, 139
sprintf, 217 partial pivoting, 210
spy, 215 pedal construction, 122
sqrt, 18, 23, 235 period doubling, 159
statistics operations, 288 permutation, 77, 105, 172, 177
 Index 305

cyclic, 183 ruffle, 185

disjoint cycles, 172, 185
number of cycles, 175 saving work, 14
order of, 181 scale
position, 178, 179 change by axis command, 47
transposition, 172 freezing, 48
permutation matrix, 215 with two curves, 48
perpendicular bisector, 57, 119 script, 9, 11
perturbation analysis, 213 sequence
phase plane, 91 attracting cycle, 157, 161
fixed point, 91 basin of attraction, 166-168
orbit, 91 cobweb diagram, 148, 149, 154, 158,
saddle point, 91 160, 161
phase plane:limit cycle, 94 convergent, 154, 158, 160
plot, 8 exponential, 161
3D line, 231 fixed point
contour, 230 attracting, 162
sliced, 231 attractive, 162
subplot, 231 indifferent, 162
surface, 229 repelling, 162
plotting, 8 Mobius, 146, 149, 154, 162
Poisson process, 86 theory, 156
polygon Newton Raphson, 165
regular, 131 period doubling, 159
polynomial, 44, 222 periodic, 147, 154, 160, 161
in n dimensions, 228 quadratic, 149, 157, 162
roots, 44, 54 superattracting cycle, 160
Taylor, 47 wildly divergent, 147, 160
polynomial function fitting, 66 shell commands, 16
population, 155 shuffle, 177
potential energy, 137 cut, 178, 183, 188
power algorithm, 37, 40, 41, 112 for odd pack, 183
prime, 110, 112 in, 179, 183, 184, 188
print, 15 order of, 181
printing, 15 out, 181, 183, 184, 188
probability, 75, 108 riffle, 178
pseudoprime, 38, 112 ruffle, 185
sliding bead, 137
quadratic equation, 166 sorting, 62
equal roots, 168 sparse matrix, 215
quadratic functional, 74 graph, 216
querying, 62 spirograph, 56, 134
queue, 245 statistics, 61

random tangents
distributions, 80 parallel, 122
integers, 76 Taylor polynomial, 47, 196
numbers, 75 tides, 281
pairs, 108 data, 282
triples, 109 trace, 106
variable, 85 transpose, 7
remainder, 33 transposition, 172
when numbers are large, 37, 40
residual correction, 212 vector
resistance, 276 representing polynomial, 44
root mean square error, 227, 235 vectors, 6, 18, 31
rotation, 130 column, 7
306 Index

row, 7

W inkel engine, 122

waiting time, 247
width, 122
windows, 3, 10

zero vector, 233

zigzag, 128