This action might not be possible to undo. Are you sure you want to continue?
Applied Engineering Mathematics
XinShe Yang
University of Cambridge, Cambridge, United Kingdom
CAMBRIDGE INTERNATIONAL SCIENCE PUBLISHING
Published by
Cambridge International Science Publishing
7 l\!Ieadow Walk, Great Abington, Cambridge CBl 6AZ, UK
http:/ fwww .cisppublishing.com
First Published 2007
@Cambridge International Science Publishing
@XinShe Yang
Conditions of Sale
All rights reserved. No part of this publication may be repro
duced or transmitted in any form or by any means, electronic
or mechanical, including photocopy, recording, or any infornla
tion storage and retrieval system, without permission in writing
from the copyright holder.
British Library Cataloguing in Publication Data
A catalogue record for this book is available from the British Library
ISBN 9781904602569
Cover design by Terry Callanan
Printed and bound in the UK by Lightning Source UK Ltd
Preface
Engineering mathematics including numerical methods and
application is the essential part of key problemsolving skills for
engineers and scientists. Modern engineering design and pro
cess modelling require both mathematical analysis and com
puter simulations. Vast literature exists on engineering math
ematics, mathen1atical modelling and numerical methods. The
topics in engineering mathematics are very diverse and the syl
labus of mathematics itself is evolving. Therefore, there is a
decision to select the topics and limit the number of chapters
so that the book remains concise and yet comprehensive enough
to include all the important mathematical methods and popu
lar numerical methods.
This book endeavors to strike a balance between mathemat
ical and numerical coverage of a wide range of mathematical
methods and numerical techniques. It strives to provide an
introduction, especially for undergraduates and graduates, to
engineering mathematics and its applications. Topics include
advanced calculus, ordinary differential equations, partial dif
ferential equations, vector and tensor analysis, calculus of varia
tions, integral equations, the finite difference method, the finite
volume method, the finite element method, reactiondiffusion
system, and probability and statistics. The book also e m p h ~
sizes the application of important mathematical methods with
dozens of worked examples. The applied topics include elastic
ity, harmonic motion, chaos, kinematics, pattern formation and
hypothesis testing. The book can serve as a textbook in en
gineering mathematics, mathematical modelling, and scientific
computing.
XinShe Yang
Cambridge, 2007
Acknowledgements
First and foremost, I would like to thank my mentors, tutors
and colleagues: Prof. A C Fowler and Prof. C J :Mcdiarmid
at Oxford University, Dr J l\II Lees and Dr C T Morley at
Cambridge University, Prof. A C :rvlclntosh, Prof. J Brindley,
Prof. R W Lewis, Prof. D T Gethin, and Prof. Andre Revil
for their help and encouragement. I also thank Dr G. Parks,
Dr T. Love, Dr S. Guest, Dr K. Seffen, and many colleagues
for their inspiration. I thank many of my students, especially
Hugo Whittle and Charles Pearson, at Cambridge University
who have indirectly tried some parts of this book and gave their
valuable suggestions.
I also would like to thank my publisher, Dr Victor Riecan
sky, for his kind help and professional editing.
Last but not least, I thank my wife, Helen, and son, Young,
for their help and support.
XinShe Yang
ii
About the Author
XinShe Yang received his D.Phil in applied n1athematics
from the University of Oxford. He is currently a research fellow
at the University of Cambridge. Dr Yang has published exten
sively in international journals, book chapters, and conference
proceedings. His research interests include asymptotic anal
ysis, bioinspired algorithms, combustion, computational engi
neering, engineering optimization, solar eclipses, scientific pro
gramming and pattern formation. He is also the author of a
book entitled: "An Introduction to Computational Engineering
with Jvlatlab", published in 2006 by Cambridge International
Science Publishing Ltd.
iii
Contents
1 Calculus 1
1.1 Differentiations ....... 1
1.1.1 Definition . . . . . . 1
1.1.2 Differentiation Rules 2
1.1.3 In1plicit Differentiation . 4
1.2 Integrations . . . . . . . . . 5
1.2.1 Definition ........ 5
1.2.2 Integration by Parts .. 6
1.2.3 Taylor Series and Power Series 8
1.3 Partial Differentiation ........ 9
1.3.1 Partial Differentiation .... 9
1.3.2 Differentiation of an Integral 12
1.4 Multiple Integrals ..... 12
1.4.1 l\1ultiple Integrals 12
1.4.2 Jacobian ...... 13
1.5 Some Special Integrals . . 16
1.5.1 Asymptotic Series 17
1.5.2 Gaussian Integrals 18
1.5.3 Error Functions . . 20
1.5.4 Gamma Functions 22
1.5.5 Bessel Functions 24
2 Vector Analysis 27
2.1 Vectors ... . . . . . . . . . . 27
2.1.1 Dot Product and Norm 28
v
CONTENTS
2.1.2 Cross Product
2.1.3 Vector Triple .
2.2 Vector Algebra . . . .
2.2.1 Differentiation of Vectors
2.2.2 Kinematics . . . . . . .
2.2.3 Line Integral . . . . . . .
2.2.4 Three Basic Operators ..
2.2.5 Son1e Important Theorems
2.3 Applications . . . . . . . . . .
2.3.1 Conservation of 1vlass
2.3.2 Saturn's Rings
3 Matrix Algebra
3.1 :rviatrix ....
3.2 Determinant.
3.3 Inverse . . . .
3.4 :rviatrix Exponential.
3.5 Hermitian and Quadratic Forms
3.6 Solution of linear systems
4 Complex Variables
4.1 Complex Numbers and Functions .
4.2 Hyperbolic Functions .
4.3 Analytic Functions
4.4 Complex Integrals . .
5 Ordinary Differential Equations
5.1 Introduction . . . . .
5.2 First Order ODEs ...
5.2.1 Linear ODEs ..
5.2.2 Nonlinear ODEs
5.3 Higher Order ODEs ..
5.3.1 General Solution
5.3.2 Differential Operator .
5.4 Linear System ...... .
5.5 SturmLiouville Equation
vi
CONTENTS
30
31
32
32
33
37
38
40
41
41
42
47
47
49
50
52
53
56
61
61
65
67
70
77
77
78
78
80
81
81
84
85
86
CONTENTS
5.5.1 Bessel Equation ........ .
5.5.2 Euler Buckling . . . . . . . . .
5.5.3 Nonlinear SecondOrder ODEs
6 Recurrence Equations
6.1 Linear Difference Equations
6.2 Chaos and Dynamical Systems
6.2.1 Bifurcations and Chaos
6.2.2 Dynamic Reconstruction.
6.2.3 Lorenz Attractor ..
6.3 Selfsimilarity and Fractals ...
7 Vibration and Harmonic Motion
7.1 Undamped Forced Oscillations
7.2 Damped Forced Oscillations .
7.3 Normal Ivlodes . . . . . . . .
7.4 Small Amplitude Oscillations
8 Integral Transforms
8.1 Fourier Transform
8.1.1 Fourier Series . .
8.1.2 Fourier Integral .
8.1.3 Fourier Transform
8.2 Laplace Transforms.
8.3 ~ a v e l e t . . . . . . . . . .
9 Partial Differential Equations
9.1 First Order PDE
9.2 Classification
9.3 Classic PDEs . .
10 Techniques for Solving PDEs
10.1 Separation of Variables .
10.2 Transform l\1ethods ... .
10.3 Similarity Solution ... .
10.4 Travelling ~ a v e Solution .
vii
CONTENTS
88
90
91
95
95
98
99
. 102
. 103
. 105
109
. 109
. 112
. 116
. 119
125
. 126
. 126
. 128
.. 129
. 131
. 134
137
. 138
. 139
. 139
141
. 141
. 143
. 145
. 147
CONTENTS
10.5 Green's Function
10.6 Hybrid Method .
11 Integral Equations
11.1 Calculus of Variations ..... .
11.1.1 Curvature . . . . . . . . .
11.1.2 EulerLagrange Equation
11.1.3 Variations with Constraints
11.1.4 Variations for l\1ultiple Variables
11.2 Integral Equations . . . . . . . .
11.2.1 Linear Integral Equations
11.3 Solution of Integral Equations .
11.3.1 Separable Kernels . .
11.3.2 Displacement Kernels
11.3.3 Volterra Equation
12 Tensor Analysis
12.1 Notations ..
12.2 Tensors . . .
12.3 Tensor Analysis .
13 Elasticity
13.1 Hooke's Law and Elasticity
13.2 l\1axwell's Reciprocal Theorem
13.3 Equations of l\1otion . . . . .
13.4 Airy Stress Functions ....
13.5 EulerBernoulli Beam Theory
CONTENTS
. 148
. 149
153
. 153
. 153
. 154
. 160
. 165
. 167
. 167
. 169
. 169
. 170
. 170
173
. 173
. 174
. 175
181
. 181
. 185
. 189
. 192
. 196
14 Mathematical Models 201
14.1 Classic l\1odels . . . . . . . . . . . . . . . . 201
14.1.1 Laplace's and Poisson's Equation . . 202
14.1.2 Parabolic Equation . . 202
14.1.3 Wave Equation . . . . . . 203
14.2 Other PDEs . . . . . . . . . . . . 203
14.2.1 Elastic Wave Equation . . 203
14.2.2 l\tlaxwell's Equations . 204
viii
CONTENTS
14.2.3 ReactionDiffusion Equation.
14.2.4 FokkerPlank Equation
14.2.5 BlackScholes Equation .
14.2.6 Schrodinger Equation ..
14.2.7 NavierStokes Equations .
14.2.8 SineGordon Equation
CONTENTS
. 204
. 205
. 205
. 206
. 206
. 207
15 Finite Difference Method 209
15.1 Integration of ODEs . . . 209
15.1.1 Euler Scheme . . . 210
15.1.2 LeapFrog Jviethod . 212
15.1.3 RungeKutta Jviethod . 213
15.2 Hyperbolic Equations . . . . . 213
15.2.1 FirstOrder Hyperbolic Equation . 214
15.2.2 SecondOrder Wave Equation . 215
15.3 Parabolic Equation . . 216
15.4 Elliptical Equation . . 218
16 Finite Volume Method 221
16.1 Introduction . . . . . . 221
16.2 Elliptic Equations . . 222
16.3 Parabolic Equations . 223
16.4 Hyperbolic Equations . 224
17 Finite Element Method 227
17.1 Concept of Elements . . . . . . . . 228
17.1.1 Simple Spring Systems . . . 228
17.1.2 Bar and Beam Elements . . 232
17.2 Finite Element Formulation . 235
17.2.1 Weak Formulation . 235
17.2.2 Galerkin Jviethod . 236
17.2.3 Shape Functions . . . 237
17.3 Elasticity . . . . . . . . . . 239
17.3.1 Plane Stress and Plane Strain . . 239
17.3.2 Implementation . . 242
17.4 Heat Conduction . . 244
ix
CONTENTS
CONTENTS
17.4.1 Basic Formulation . . . . . . . . . . . 244
17 .4.2 ElementbyElement Assembly . . . . 246
17.4.3 Application of Boundary Conditions . 248
17.5 TimeDependent Problems . . . 251
17.5.1 The Time Dimension. . . . . . 251
17.5.2 TimeStepping . . . . . . . . . 253
17.5.3 1D Transient Heat Transfer . . 253
17.5.4 Wave Equation .. . 254
18 Reaction Diffusion System 257
18.1 Heat Conduction Equation . 257
18.1.1 Fundamental Solutions . . 257
18.2 Nonlinear Equations . . . . 259
18.2.1 Travelling Wave . . . 259
18.2.2 Pattern Formation . . 260
18.3 ReactionDiffusion System . . 263
19 Probability and Statistics 267
19.1 Probability . . . . . . . . . . . . . . . 267
19.1.1 Randomness and Probability . 267
19.1.2 Conditional Probability . . . . 275
19.1.3 Random Variables and Ivloments . 277
19.1.4 Binomial and Poisson Distributions. . 281
19.1.5 Gaussian Distribution . . . . . 283
19.1.6 Other Distributions . . . . . . 286
19.1.7 The Central Limit Theorem . . 287
19.2 Statistics . . . . . . . . . . . . . . . 289
19.2.1 Sample Ivlean and Variance . 290
19.2.2 Iviethod of Least Squares . 292
19.2.3 Hypothesis Testing . . 297
A Mathematical Formulas 311
A.1 Differentiations and Integrations . 311
A.2 Vectors and Matrices . . 312
A.3 Asymptotics
. 314
A.4 Special Integrals . . . . 315
X
Chapter 1
Calculus
The prelin1inary requirements for this book are the precalculus
foundation mathematics. We assume that the readers are fa
miliar with these preliminaries, and readers can refer to any
book that is dedicated to these topics. Therefore, we will only
review some of the basic concepts of differentiation and inte
gration.
1.1 Differentiations
1.1.1 Definition
For a known function or a curve y = f ( x) as shown in Figure
1.1, the slope or the gradient of the curve at the point P( x, y)
is defined as
dy = df(x) = !'(x) = lim f(x + ~ x )  f(x), (1.
1
)
dx dx .dx+0 ~ x
on the condition that there exists such a limit at P.
This gradient or limit is the first derivative of the function
f ( x) at P. If the limit does not exist at a point P, then we
say that the function is nondifferentiable at P. By conven
tion, the limit of the infinitesimal change ~ x is denoted as the
differential dx. Thus, the above definition can also be written
1
1.1 Differentiations
Calculus
as
dl(x)
1
dy = dl = ~ d x =I (x)dx,
(1.2)
which can be used to calculate the change indy caused by the
small change of dx. The primed notation
1
and standard nota
tion d ~ can be used interchangeably, and the choice is purely
out of convenience.
Figure 1.1: Gradient of a curve
The second derivative of l(x) is defined as the gradient of
l
1
(x), or
d
2
y =I"( ) = dl(x)
dx2 x dx ·
(1.3)
The higher derivatives can be defined in a similar manner.
Thus,
d
3
y =I"'( ) = dl"(x)
dx
3
 x dx '
lf'ly dl(nl)
... ,  = l(n) = .
dxn dx
(1.4)
1.1. 2 Differentiation Rules
If a more complicated function I ( x) can be written as a prod
uct of two simpler functions u(x) and v(x), we can derive a
differentiation rule using the definition from the first princi
2
Calculus 1.1 Differentiations
pies. Using
f(x +  f(x) u(x + +  u(x)v(x)
=
and subtracting and adding u(x + + u(x +
[= 0] terms, we have
df d[u(x)v(x)]
dx = dx
I
. [ ( A )v(x + v(x) ( )u(x + u(x)]
1m +vx
X
dv du
= u(x) dx + dx v(x),
(1.5)
which can be written in a contract form using primed notations
J'(x) = (uv)' = u'v + uv'. (1.6)
If we differentiate this equation again and again, we can gen
eralize this rule, we finally get the Leibnitz's Theorem for dif
ferentiations
+ ... + uv(n),
(1.7)
where the coefficients are the same as the binomial coefficients
nc = ( n) = nl
r r rl(nr)!'
(1.8)
If a function f(x) [for example, f(x) = exn] can be written
as a function of another function g(x), or f(x) = f[g(x)] [for
example, f(x) = eg(x) and g(x) = xn], then we have
!
'( )  I'
X liD
.6x+O
(1.9)
3
1.1 Differentiations
which leads to the following chain rule
f
1
(x) = df dg,
dgdx
or
{f[g(x)]}
1
= J'[g(x)] · g
1
(x).
In our example, we have f' ( x) = (ex")' = ex" nxn l.
Calculus
(1.10)
(1.11)
If one use 1/v instead of v in the equation (1.6) and (1/v)' =
v
1
fv
2
, we have the following differentiation rule for quotients:
I I
(!)I = u v  uv .
v v
2
(1.12)
0 Example 1.1: The deriva.tive of f(x) = sin(x)ecos(x) can be
obtained using the combination of the above differentiation rules.
f'(x) = [sin(x))'ecos(x) + sin(x)[ecos(x))'
= cos(x)e cos(x) + sin(x)e cos(x)[ cos(x))'
= cos(x )e cos(x) + sin2(x )e cos(x).
The derivatives of various functions are listed in Table 1.1.
1.1.3 Implicit Differentiation
The above differentiation rules still apply in the case when there
is no simple explicit function form y = f ( x) as a function of
x only. For example, y + sin(x) exp(y) = 0. In this case, we
can differentiate the equation tern1 by tern1 with respect to x
so that we can obtain the derivative dyfdx which is in general
a function of both x and y.
0 Example 1. 2: Find the derivative ::; if y
2
+ sin( x )eY = cos( x).
Differentiating term by term with respect to x, we have
2y ~ ~ + cos(x)eY + sin(x)eY: =  sin(x),
dy cos(x)eY + sin(x)
dx  2y + sin(x)eY ·
0
4
Calculus 1.2 Integrations
Table 1.1: First Derivatives
f(x) f'(x)
xn nxn1
ex ex
ax(a > 0) axlna
lnx
!
X
logax
1
xlna
sinx cosx
cosx sinx
tanx sec
2
x
sin
1
x
1
~
cos
1
x
1
 v'1x
2
tan
1
x
1
1+x
2
sinhx coshx
coshx sinhx
1.2 Integrations
1.2.1 Definition
Integration can be viewed as the inverse of differentiation. The
integration F( x) of a function f ( x) satisfies
d ~ ~ x ) = f(x),
(1.13)
or
F(x) = 1x ! ( ~ ) ~ ,
xo
(1.14)
where f ( x) is called the integrand, and the integration starts
from xo (arbitrary) to x. In order to avoid any potential confu
sion, it is conventional to use a dummy variable (say, ~ ) in the
integrand. As we know, the geometrical meaning of the first
derivative is the gradient of the function f ( x) at a point P, the
5
1.2 Integrations
Calculus
geometrical representation of an integral J: f ( ~ ) d ~ (with lower
integration limit a and upper integration limit b) is the area
under the curve f(x) enclosed by xa.xis in the region x E [a, b].
In this case, the integral is called a definite integral as the limits
are given. For the definite integral, we have
fb f(x)dx = 1b f(x)dx 1a f(x)dx = F(b) F(a). (1.15)
la xo xo
The difference F (b)  F (a) is often written in a compact form
F l ~ = F(b) F(a). As F'(x) = f(x), we can also write the
above equation as
l f(x)dx = l F'(x)dx = F(b)  F(a). (1.16)
Since the lower limit x
0
is arbitrary, the change or shift
of the lower limit will lead to an arbitrary constant c. When
the lower limit is not explicitly given, the integral is called an
indefinite integral
I f(x)dx = F(x) + c, (1.17)
where cis the constant of integration.
The integrals of some of the con1mon functions are listed in
Table 1.2.
1.2.2 Integration by Parts
From the differentiation rule ( uv )' = uv' + u' v, we have
uv' = (uv)' u'v.
(1.18)
Integrating both sides, we have
udx = uv  vdx
I
dv ldu
dx dx '
(1.19)
in the indefinite form. It can also be written in the definite
form as
1
b dv lb 1b du
ud dx = [uv] + vd dx.
a X a a X
(1.20)
6
Calculus 1.2 Integrations
Table 1.2: Integrals
f(x) f f(x)
xn(n tf 1)
xn
n+1
1
lnlxl
X
ex ex
sinx cosx
cosx sinx
1
1 tan
1
i!.
a2tx2 a a
1 In a+x
a2x2
2a ax
1
. . ! . . I n ~
x2a2
2a x+a
1
sin
1
~
v'a1x1
a
1
ln(x + v'x2 + a2)
~
[or sinh
1
~ ]
1
ln(x + v'x2  a2)
Jx2a2
(or cosh
1
~ ]
sinhx coshx
coshx sinhx
tanhx lncoshx
The integration by parts is a very powerful method for evalu
ating integrals. 1\.fany complicated integrands can be rewritten
as a product of two simpler functions so that their integrals can
easily obtained using integration by parts.
0 Example 1.9: The integral of I= J x lnx dx can be obtained
by setting v' = x and u = lnx. Hence, v = x
2
2
and u' = ~ · \Ve now
have
I= xlnxdx =  dx
J
x
2
lnx J x
2
1
2 2 X
0
7
1.2 Integrations
Calculus
Other important methods of integration include the sub
stitution and reduction methods. Readers can refer any book
that is dedicated to advanced calculus.
1.2.3 Taylor Series and Power Series
From
l f(x)dx = F(b)  F(a),
and 1:£ = F' = f(x), we have
1
xo+h
/'(x)dx = f(xo +h) f(xo),
xo
which means that
1
xo+h
f(xo +h) = f(xo) + f'(x)dx.
xo
(1.21)
(1.22)
(1.23)
If h is not too large or f' ( x) does not vary dramatically, we can
approximate the integral as
1
xo
f'(x)dx ~ !'(xo)h.
xo
(1.24)
Thus, we have the firstorder approximation to f ( xo + h)
f(xo + h ) ~ f(xo) + hf'(xo). (1.25)
This is equivalent to say, any change from xo to xo+h is approx
imated by a linear term hf'(xo). If we repeat the procedure for
f'(x), we have
J'(xo + h ) ~ !'(xo) + hf"(xo), (1.26)
which is a better approximation than f'(xo +h) ~ f'(x
0
). Fol
lowing the same procedure for higher order derivatives. we can
reach the nth order approximation
f(xo +h)= f(xo) + hf'(xo) + ~ ~ !"(xo) + ; ~ !"'(xo)
8
Calculus 1.3 Partial Differentiation
(1.27)
where Rn+t (h) is the error of this approximation and the no
tation means that the error is about the same order as n + 1th
term in the series. This is the wellknown Taylor theorem and
it has many applications. In deriving this formula, we have im
plicitly assumed that all the derivatives l'(x), l"(x), ... , l(n)(x)
exist. In almost all the applications we meet, this is indeed the
case. For example, sin( x) and ex, all the orders of the deriva
tives exist. If we continue the process to infinity, we then reach
the infinite power series and the error limnoo Rn+ 1 + 0 if the
series converges. The end results are the Maclaurin series. For
example,
and
x3 xs
sin x = x   +   ... , (x E 'R),
3! 5!
x2 x4
cosx = 1+ ... , (x E 'R),
2! 4!
1.3 Partial Differentiation
1.3.1 Partial Differentiation
(1.28)
(1.29)
(1.30)
The differentiation defined above is for function I ( x) which has
only one independent variable x, and the gradient will generally
depend on the location x. For functions I ( x, y) of two variables
x andy, their gradient will depend on both x andy in general.
In addition, the gradient or rate of change will also depend on
the direction (along xaxis or yaxis or any other directions).
For example, the function l(x, y) = xy shown in Figure 1.2
has different gradients at ( 0, 0) along xaxis and yaxis. The
9
1.3 Partial Differentiation Calculus
1 1
Figure 1.2: Variation of f(x, y) = xy.
gradients along the positive x and y directions are called the
partial derivatives respect to x and y, respectively. They are
denoted as U and %, respectively.
The partial derivative of f ( x, y) with respect to x can be
calculated assuming that y =constant. Thus, we have
af(x,y) = f = 8!1
8x  x 8x Y
= lim
f(x + ~ x , y) f(x, y)
~ X
.6.xO,y=const
Similarly, we have
af(x,y) = f = 8!1
ay  yayx
lim
.6.yO ,x=const
f(x, y + ~ y )  f(x, y)
~ y
(1.32)
(1.33)
The notation fxly emphasizes the fact that y is held constant.
The subscript notation fx (or /y) emphasizes the derivative is
carried out with respect to x (or y). :Mathematicians like to
10
Calculus 1.3 Partial Differentiation
use the subscript forms as they are simpler notations and can
be easily generalized. For example,
(1.34)
Since flxfly = flyflx, we have fxy = !yx·
0 Example 1.4: The first partial derivatives of f(x, y) = xy +
sin(x)eY are
8!
fx = ax = y + Cos(x)eY,
The second partial derivative of f(x, y) is
fxx = sin(x)eY,
and
fxy = fyx = 1 cos(x)eY.
0
For any small change Llf = f(x+flx, y+fly) f(x, y) due
to flx and fly, the total infinitesimal change df can be written
as
(1.35)
If x and y are functions of another independent variable ~ , then
the above equation leads to the following chain rule
df af dx af dy
~ = 8 x d ~ + 8 y ~ '
(1.36)
which is very useful in calculating the derivatives in parametric
form or for change of variables. If a complicated function f ( x)
can be written in terms of simpler functions u and v so that
f(x) = g(x, u, v) where u(x) and v(x) are known functions of
x, then we have the generalized chain rule
dg ag ag du ag dv
=++
dx ax au dx av dx.
11
(1.37)
1.4 !v!ultiple Integrals Calculus
The extension to functions of more than two variables is
straightforward. For a function p(x, y, z, t) such as the pressure
in a fluid, we have the total differential as
df = op dt + op dx + op dy +
8
P dz. (1.38)
8t 8x oy 8z
1.3.2 Differentiation of an Integral
When differentiating an integral
<11(x) = l <P(x, y)dy, (1.39)
with fixed integration limits a and b, we have
84>(x) = [b 8</J(x, y) dy.
OX Ja OX
(1.40)
When differentiating the integrals with the limits being func
tions of x,
I(x) = ru(x) 1/J(x, T)dT = w[x, u(x)]  w[x, v(x)],
lv(x)
the following formula is useful:
dl 1u(x) o'l/J du dv
d = 
8
dT + [1/J(x, u(x))d 1/J(x, v(x))d ].
X v(x) X X X
This formula can be derived using the chain rule
dl ol ol du 8 I dv
=++
dx 8x ou dx ov dx'
where ~ = 1/J(x, u(x)) and §f = 1/J(x, v(x)).
1.4 Multiple Integrals
1.4.1 Multiple Integrals
(1.41)
(1.42)
(1.43)
As the integration of a function f ( x) corresponds to the area
enclosed under the function between integration limits, this
12
Calculus 1.4 A1ultiple Integrals
can extend to the double integral and multiple integrals. For a
function f(x, y), the double integral is defined as
F =In f(x, y)dA,
(1.44)
where dA is the infinitesin1al element of the area, and n is the
region for integration. The simplest form of dA is dA = dxdy
in Cartesian coordinates. In order to emphasize the double
integral in this case, the integral is often written as
I = J In f(x, y)dxdy.
(1.45)
0 Example 1. 5: The area moment of inertia of a thin rectangular
plate, with the length 2a and the width 2b, is defined by
I= J In y
2
dS = J In y
2
dxdy.
The plate can be divided into four equal parts, and we have
0
1.4.2 Jacobian
Sometimes it is necessary to change variables when evaluating
an integral. For a simple onedimensional integral, the change
of variables from x to a new variable v (say) leads to x = x(v).
This is relatively sin1ple as dv = ~ ~ d v , and we have
1
xb 1b dv
f(x)dx = f(x(v))d dv,
Xa a X
(1.46)
where the integration limits change so that x(a) Xa and
x(b) = Xb. Here the extra factor dxjdv in the integrand is
referred to as the Jacobian.
13
1.4 !v!ultiple Integrals Calculus
For a double integral, it is more complicated. Assuming
x = 17), y = 17), we have
jj f(x, y)dxdy = j j (1.47)
where J is the Jacobian. That is
J = a(x, y)
I
{)x {)x I I {)x §JL I
 "lJ7i 
 §JL §JL  {)x {JiJ •
"lJ7i
(1.48)
The notation a( x, y) 1J) is just a useful shorthand. This
equivalent to say that the change of the infinitesimal area dA =
dxdy becomes
d d = I a(x, y) 1,/Cd = I ax ay  ax ay 1,/Cd
X y 1J) 1J a1] a1] 1].
(1.49)
0 Example 1. 6: \Vhen transforming from ( x, y) to polar coordi
nates (r, 8), we have the following relationships
x = rcos8, y = rsin8.
Thus, the Jacobian is
J _ a( X, y) _ ax ay ax ay
 a(r,8)  ar a8 ae ar
= cos 8 x r cos()  ( r sin 8) x sin 8 = r[cos
2
8 + sin
2
8) = r.
Thus, an integral in (x, y) will be transformed into
/! (j>(x,y)dxdy = J J 4>(rcos8,rsin8)rdrd().
0
In a similar fashion, the change of variables in triple inte
grals gives
V = jjfn c/J(x,y,z)dxdydz = JL (1.50)
14
Calculus 1.4 !v!ultiple Integrals
and
Ox Oy 0:::
J = {)(X, y, Z) =
~
~
~
Ox 0:::
(1.51)
iJ1] F F
 { ) ( ~ , 1J, ()
Ox 0 ~ 0 ~
0< 0< m:
For cylindrical polar coordinates ( r, (jJ, z) as shown in Figure
1.3, we have
x = rcos(jJ, y = rsin¢,
The Jacobian is therefore
cos 4> sin¢ 0
J = 0( X, y, Z) =
8(r,(jJ,z)
rsin¢ rcos¢ 0
0 0 1
z
Z =Z.
=r.
Figure 1.3: Cylindrical polar coordinates.
(1.52)
(1.53)
For spherical polar coordinates (r, 8, 4>) as shown in Figure
1.4, where 8 is the zenithal angle between the zaxis and the
position vector r, and 4> is the azimuthal angle, we have
x = rsin8cos¢, y = r sin 8 sin¢, z = r cos 8. (1.54)
Therefore, the Jacobian is
sin 8 cos 4> sin 8 sin 4J cos 8
J = r cos 8 cos 4J r cos 8 sin 4> r sin 8 = r
2
sin 8. (1.55)
r sin 8 sin 4J r sin 8 cos 4J 0
15
1.5 Some Special Integrals Calculus
z
Figure 1.4: Spherical polar coordinates.
Thus, the volu1ne element change in the spherical systen1 is
dxdydz = r
2
sin OdrdOdc/>. (1.56)
0 Example 1. 7: The volume of a solid ball with a radius R is
defined as
V= !!In dV.
Since the infinitesimal volume element dV = r
2
sin( O)drdOd¢ in spher
ical coordinates r ~ 0, 0 ~ () ~ 1r and 0 ~ ¢ ~ 21r, the ball can be
divided into two equal parts so that
{R {rr/2 {21r
V = 2 Jo {}
0
sin0[}
0
d¢]dO}dr
{R {rr/2
= 2 Jo {27r Jo sin(O)dO}dr
0
1.5 Some Special Integrals
Some integrals appear so frequently in engineering mathemat
ics that they deserve special attention. l\1ost of these special
16
Calculus 1.5 Some Special Integrals
integrals are also called special functions as they have certain
varying parameters or integral limits. We only discuss four of
the n1ost common integrals here.
1.5.1 Asymptotic Series
Before we discuss any special functions, let us digress first to
introduce the asymptotic series and order notations because
they will be used to study the behaviours of special functions.
Loosely speaking, for two functions f ( x) and g( x), if
f(x) + K
g(x) '
X+ XQ,
where K is a finite, nonzero limit, we write
f = O(g).
(1.57)
(1.58)
The big 0 notation means that f is asymptotically equivalent
to the order of g(x). If the limit is unity or K = 1, we say f(x)
is order of g( x). In this special case, we write
J ~ g ,
(1.59)
which is equivalent to f fg + 1 and gf f+ 1 as x + xo. Ob
viously, x
0
can be any value, including 0 and oo. The notation
~ does not necessarily mean ~ in general, though they might
give the same results, especially in the case when x + 0 [for
example, sin x ~ x and sin x ~ x if x + 0].
When we say f is order of 100 (or f ~ 100), this does not
mean f ~ 100, but it can mean that f is between about 50 to
150. The small o notation is used if the limit tends to 0. That
is
f
+ 0,
g
X+ XQ, (1.60)
or
f = o(g). (1.61)
17
1.5 Some Special Integrals Calculus
If g > 0, f = o(g) is equivalent to f << g. For example, for
Vx E 'R, we have e x ~ 1 + x + O(x
2
) ~ 1 + x + ~ + o(x).
Another classical example is the Stirling's asymptotic series
for factorials
n >> 1. (1.62)
In fact, it can be expanded into more terms
. ~ n n 1 1 139
nl 'V v ..::;1rn(;) (
1
+ 12n + 288n2  51480n3  ... ).
(1.63)
This is a good example of asymptotic series. For standard
power expansions, the error Rk(hk) + 0, but for an asymptotic
series, the error of the truncated series Rk decreases and gets
smaller con1pared with the leading term [here V21m(nfe)n].
However, Rn does not necessarily tend to zero. In fact, R2 =
tin· V21m(nje)n is still very large as R2 + oo if n >> 1. For
example, for n = 100, we have n! = 9.3326 x 10
157
, while the
leading approximation is V21m(nfe)n = 9.3248 x 10
157
. The
difference between these two values is 7. 77 40 x 10
154
, which is
still very large, though three orders smaller than the leading
approximation.
1.5.2 Gaussian Integrals
The Gaussian integral appears in n1any situations in engineer
ing mathematics and statistics. It can be defined by
!
00 2
/(a) = oo ea:x dx. (1.64)
In order to evaluate the integral, let us first evaluate /
2
. Since
the Gaussian integral is a definite integral and must give a
constant value, we can change the dummy variable as we wish.
We have
1
00 2 100 2 100 2
/
2
= [ oo ea:x dxf = oo ea:x dx oo ea:y dy
18
Calculus 1. 5 Some Special Integrals
= i: i: eo(x2+y2)dxdy. (1.65)
Changing into the polar coordinates ( r, 0) and noticing r
2
=
x
2
+ y
2
and dxdy = rdrdO, we have
roo [21r 2
I2 = Jo dr Jo rear dO
lo
oo 1 ar2d( 2) 7r
= 21r e ar = .
o a a
(1.66)
Therefore,
( )
1
00
ax2d ~
I a = e x = .
oo a
(1.67)
Since a is a parameter, we can differentiate both sides of this
equation with respect to a, and we have
!
00
2 ox2d 1 ~
X e X= .
oo 2a a
(1.68)
By differentiating both sides of the Gaussian integral ( equa
tion 1.67) n times with respect to a, and we get the generalized
Gaussian integral In
I
1
00
2n ox
2
n = X e
oo
(1.69)
where n > 0 is an integer.
For a special case when a = ~ and n = 0, the equation
v2u""
(1.67) can be rearranged as
/
00
00
f(x, a)dx = 1, (1.70)
The function f(x, a) is a zeromean Gaussian probability func
tion. As a ~ 0, f(x) ~ 8(x) where 8(x) is the Dirac 8function
which is defined by
8(x) i= 0 (at x = 0), but 8(x) = 0, for xi= 0, (1.71)
19
1.5 Some Special Integrals Calculus
and
/00
00
o(x)dx = 1.
(1.72)
It has an interesting property that
J f(x)o(x (3)dx = !((3), (1.73)
where f(x) is a function.
1.5.3 Error Functions
The error function, which appears frequently in heat conduc
tion and diffusion problems, is defined by
2 rx 2
erf(x) = v'i Jo e
11
d'f/.
Its complementary error function is defined by
2 1
00
2
erfc(x) = 1 erf(x) = y'i x et dt.
The error function is an odd function: erf( x)
Using the results from the Gaussian integral
1
00 2
e 
11
d'TJ = .;:i,
oo
(1.74)
(1.75)
= erf(x).
(1.76)
together with the basic definition, we have erf(O) 0, and
erf( oo) = 1. Both the error function and its complementary
function are shown in Figure 1.5.
The error function cannot be easily evaluated in closed
form. Using Taylor series for the integrand
_112  2 1 4 1 6
e  1  'TJ + 2"'  6"' + ... ,
(1.77)
and integrating term by term, we have
2 x
3
x
5
x
7
erf(x) = yli"[x 3 + 10 42 + ... ],
(1.78)
20
Calculus 1.5 Some Special Integrals
or
2 

1.5
1
.!So.
"t: 0.5
CD
0
0.5
5 0
X
Figure 1.5: Error functions.
 erf(x)
 erfc(x)
5 10
2 oo ( 1) n x2n+ 1
erf(x) =  L 
1
. (1.79)
...fi n=O 2n + 1 n.
The integrals of the complementary function are defined by
ierfc(x) = Loo erfc(TJ)dTJ, (1.80)
and
(1.81)
Using integration by parts, we can prove the following asymp
totic series
ex2
erf(x) "J 1 r;;' (x? oo). (1.82)
Xy7r
On the other hand, if we replace x in the error function by
{3x, we have
lim 
2
1
[1 + erf(/3x)] ? H(x),
[3oo
(1.83)
where H ( x) is a Heaviside function or a unit step function
which is defined by
H(x) = 1 (for x > 0), H(x) = 0 (for x < 0). (1.84)
21
1.5 Some Special Integrals Calculus
At x = 0, it is discontinuous and it is convention to set H(O) =
1/2. Its relationship with the Dirac 8function is that
d ~ ~ x ) = 8(x).
(1.85)
1.5.4 Gamma Functions
The special function is the gamma function which is defined by
r(x) = fooo txletdt = fooo et+(xl)lntdt. (1.86)
Using integral by parts, we have
r(x + 1) = fooo txetdt = txetl: + fooo xtxletdt
= xr(x). (1.87)
When x = 1, we have
The variation of r(x) is shown in Figure 1.6.
If X= n is an integer (n E JV), then r(n + 1) = n!. That is
to say,
n! = r(n + 1) = fooo enlnttdt. (1.89)
The integrand f(n, t) = exp[n ln tt] reaches a tnaximum value
at
af =O
8t '
or t =n. (1.90)
The maximum is !max = exp[n ln n n). Thus, we now can set
t = n + r = n(1 + () so that r = n( varies around n and (
around 0. For n >> 1, we have
n! = /_: e{n lnln(l+()Jn(l+<)} dr
= /_: e{(n In nn)+nlln(l+()<J} dr, (1.91)
22
Calculus 1.5 Some Special Integrals
20
10
10
20
5 0
X
Figure 1.6: Variation of f(x).
5
where we have used ln[n(l + ()] = Inn+ ln(l + (). The inte
gration limits for r = n( (not () are from oo to oo. Using
(2 (3
ln(l + () = ( 2" + "3  ... ,
(1.92)
we have
(1.93)
From the Gaussian integral with a= 1/(2n)
1
00 2 ~
ear dr =  = /2rn,
oo a
(1.94)
we now obtain the Stirling's asymptotic formula
From the basic definition, it can be shown that
r(
1 ) r,;; r ( ~ 2 ) = ..fi27r.
2 = y7r,
(1.96)
The standard gan1ma function can be decomposed into two
incomplete functions: the lower incomplete gamma function
23
1.5 Some Special Integrals Calculus
/'(0:, x) and the upper incomplete gamnla function r(a, x) so
that r(x) = 'Y(O:, x) + r(a, x).
The lower incomplete gamma function is defined by
(1.97)
while the upper incomplete gamma function is defined by
r(a,x) = 1
00
to:letdt.
(1.98)
Obviously, l(o:,x) ~ r(a) as X ~ 00. As r(!) = .Jif, we
have
1 1 2
erf(x) = fi'Y(
2
,x ). (1.99)
Another related function is a beta function
(1.100)
From the definition, we know that the beta function is sym
metric, B(x, y) = B(y, x). The beta function is linked to r
function by
B( )
= r(x)r(y)
x,y r(x +Y).
(1.101)
1.5.5 Bessel Functions
Bessel functions come from the solution of the Bessel's equation
(1.102)
which arises from heat conduction and diffusion problems as
well as wave propagation problems. The solution (see later
chapters in this book) can be expressed as Taylor's series, and
the Bessel function associated with this equation can be defined
by
(1.103)
24
Calculus 1.5 Some Special Integrals
 A.=O
 1
 2
..... 1
0.5

1
o!::=5
X
Figure 1.7: Bessel functions.
where A is a real paran1eter. These are the Bessel functions of
the first kind. It can also be defined by the Bessel integral
1 [27r
J>..(x) =
2
7r Jo cos[ At x sin t]dt.
(1.104)
The Bessel functions of the second kind are related to J >.., and
can be defined by
y _ J>.. cos(A7r) J_>..
>..  sin(A7r) ·
(1.105)
When A = k is an integer, they have the following properites
The Bessel functions of the first kind are plotted in Figure 1. 7.
With these fundamentals of preliminary mathematics, we
are now ready to study a wide range of mathematical methods
in engineering.
25
Chapter 2
Vector Analysis
1\tiany quantities such as force, velocity, and deformation in en
gineering and sciences are vectors which have both a n1agnitude
and a direction. The manipulation of vectors is often associated
with matrices. In this chapter, we will introduce the basics of
vectors and vector analysis.
2.1 Vectors
A vector x is a set of ordered numbers x = (xt, x2, ... , Xn),
where its components x
1
, x
2
, ... , Xn are real numbers. All these
vectors form a ndimensional vector space vn. To add two
vectors x = (xb x2, ... , Xn) andy = (Yt, Y2, ... , Yn), we simply
add their corresponding components,
Z =X+ Y = (xt + Yt,X2 + Y2, ... ,Xn + Yn), (2.1)
and the sum is also a vector. This follows the vector addition
parallelogram as shown in Fig 2.1
The addition of vectors has comn1utability (u + v = v + u)
and associativity [(a+ b) + c = a+ (b +c)]. Zero vector
0 is a special vector that all its components are zeros. The
multiplication of a vector x with a scalar or constant a: is carried
27
2.1 Vectors
out by the multiplication of each component,
ay = (ay1, oy2, ... , ayn)·
Vector Analysis
(2.2)
Thus, y = (  y ~ , y2, ... , Yn)· In addition, (a{3)y = a(f3y)
and (a+ f3)y = ay + {3y.
4 ......•...•... ~ · · · · · · · · · ·
')/
:"
.·
...
. ·
.:
:
:
//
.·
.·
:
:·
Figure 2.1: Vector addition.
Two nonzero vectors a and b are said to be linearly inde
pendent if aa + {3b = 0 implies that a= {3 = 0. If a, {3 are not
all zeros, then these two vectors are linearly dependent. Two
linearly dependent vectors are parallel ( a//b) to each other.
Three linearly dependent vectors a, b, c are in the same plane.
2.1.1 Dot Product and Norm
The dot product or inner product of two vectors x and y is
defined as
n
X· Y = XtYl + X2Y2 + ··· + XnYn = L XiYi, (2.3)
i=l
which is a real number. The length or norm of a vector x is
the root of the dot product of the vector itself,
lxl = llxll = VX:X= ~ ~ x r
(2.4)
28
Vector Analysis 2.1 Vectors
When llxll = 1, then it is a unit vector. It is straightforward
to check that the dot product has the following properties:
X·Y = y ·X, X· (y+z) =x·y+x·z, (2.5)
and
(a:x) · (f3y) = (a:f3)x · y, (2.6)
where a:, {3 are constants.
If() is the angle between two vectors a and b, then the dot
product can also be written
a· b = llall llbll cos(8),
0 ::; () ::; 7r.
(2.7)
If the dot product of these two vectors is zero or cos(8) = 0 (i.e.,
() = 1r /2), then we say that these two vectors are orthogonal.
Rearranging equation (2. 7), we obtain a formula to calcu
late the angle () between two vectors
a·b
cos(()) = II all lib II
(2.8)
Since cos(8) ::; 1, then we get the useful CauchySchwartz in
equality:
(2.9)
Any vector a in a ndimensional vector space vn can be
written as a combination of a set of n independent basis vectors
or orthogonal spanning vectors e1, e2, ... , en, so that
n
a= a:1e1 + a:2e2 + ... + O:nen = L O:iei,
i=l
(2.10)
where the coefficients/scalars 0:1,0:2, ... , O:n are the components
of a relative to the basis e1, e2 ... , en. The most common basis
vectors are the orthogonal unit vectors. In a threedimensional
case, they are i = (1, 0, 0), j = (0, 1, 0, k = (0, 0, 1) for three
x, y, zaxis, and thus x = x1i + x2.i + x3k. The three unit
vectors satisfy i · j = j · k = k · i = 0.
29
2.1 Vectors Vector Analysis
2.1.2 Cross Product
The dot product of two vectors is a scalar or a number. On the
other hand, the cross product or outer product of two vectors
is a new vector
c=axb
which is usually written as
i j k
a X b = Xt X2 X3
Yt Y2 Y3
= I :: :: li +I :: :: lj +I :: :: lk. (
2
.
12
)
The angle between a and b can also be expressed as
. lla x bll
sin f) = llallllbll.
(2.13)
In fact, the norm lla x bll is the area of the parallelogram
formed by a and b. The vector c =ax b is perpendicular to
both a and b, following a righthand rule. It is straightforward
to check that the cross product has the following properties:
xxy = yxx, (x + y)xz = xxz + yxz, (2.14)
and
(ax)x (,By) = (a,B)xxy. (2.15)
A very special case is axa = 0. For unit vectors, we have
ixj = k, jxk = i, kxi = j. (2.16)
0 Example 2.1: For two 3D vectors a = (1, 1, 0) and b =
(2, 1, 0), their dot product is
a· b = 1 X 2 + 1 X ( 1) + 0 = 1.
30
Vector Analysis 2.1 Vectors
As their moduli are
we can calculate the angle(} between the two vectors. We have
a·b 1
cosO= llallllbll = J2J5'
or
(}
l 1 0
= cos 1'17\ ~ 71.56 .
vlO
Their cross product is
V =a X b = (1 X 00 X (1),0 X 11 X 0,1 X (1) 2 X 1)
= (0,0, 3),
which is a vector pointing in the negative zaxis direction. The vector
v is perpendicular to both a and b because
a · V = 1 X 0 + 1 X 0 + 0 X ( 3) = 0,
and
b · V = 2 X 0 + ( 1) X 0 + 0 X ( 3) = 0.
0
2.1.3 Vector Triple
For two vectors, their product can be either a scalar (dot prod
uct) or a vector (cross product). Similarly, the product of triple
vectors a, b, c can be either a scalar
ax ay a::
a· (b x c)= bx by b::
'
(2.17)
Cy Cz
or a vector
ax (b x c)= (a· c)b (a· b)c. (2.18)
31
2.2 Vector Algebra Vector Analysis
As the dot product of two vectors is the area of a parallel
ogram, the scalar triple product is the volume of the paral
lelepiped formed by the three vectors. From the definitions, it
is straightforward to prove that
a· (b x c)= b · (c x a)= c ·(ax b)= a· (c x b), (2.19)
a X (b X c) #(a X b) XC, (2.20)
and
(ax b)· (c x d) =(a· c)(b ·d) (a· d)(b ·c). (2.21)
2.2 Vector Algebra
2.2.1 Differentiation of Vectors
The differentiation of a vector is carried out over each compo
nent and treating each con1ponent as the usual differentiation
of a scalar. Thus, for a position vector
P(t) = x(t)i + y(t)j + z(t)k, (2.22)
we can write its velocity as
v = ~ = ±(t)i + y(t)j + z(t)k, (2.23)
and acceleration as
a= ~ t ~ = X(t)l + ij(t)j + Z(t)k, (2.24)
where () = d()jdt. Conversely, the integral of vis
P = J vdt+c,
where cis a vector constant.
32
(2.25)
Vector Analysis 2.2 Vector Algebra
From the basic definition of differentiation, it is easy to
check that the differentiation of vectors has the following prop
erties:
d(a:a) da d(a ·b) da b db
dt = dt · +a·dt'
(2.26)
and
d(axb) _ da b db
dt  dt x + ax dt ·
(2.27)
2.2.2 Kinematics
As an application of vector algebra, let us study the motion
along a curved path. In there are three coordinate
systems which can be used to describe the motion uniquely.
The first one is the Cartesian coordinates (x, y) with two unit
vectors i (along positive xaxis) and j (along positive yaxis),
and the second one is the polar coordinates (r, 0) with two unit
vectors er and ee as shown in Figure 2.2.
y
Figure 2.2: Polar coordinates, their unit vectors and their re
lationship with Cartesian coordinates.
The position vector r = x(t)i + y(t)j at point P at any
instance t in the Cartesian coordinates can be expressed as
( r, 0). The velocity vector is
(2.28)
33
2.2 Vector Algebra Vector Analysis
and the acceleration is
.2 .. .
a= v = (frO )er + (rO + 2rO)eo. (2.29)
The third coordinate system is the intrinsic coordinate sys
tem (s, 'l/J) where sis the arc length from a reference point (say,
point 0) and 'l/J is the angle of the tangent at the point P (see
Figure 2.3). The two unit vectors for this systems are et along
the tangent direction and en which is the unit normal of the
curve.
Figure 2.3: Intrinsic coordinates and their unit vectors.
In the intrinsic coordinates, the position is uniquely deter
mined by (s, 1/J), and the velocity is always along the tangent.
Naturally, the velocity is simply
(2.30)
The acceleration becomes
.. 82
a= set+ en,
p
(2.31)
where pis the radius of the curvature at point P.
For the circular n1otion such as a moving bicycle wheel as
shown in Figure 2.4, the three coordinate systems are intercon
nected. In a rotating reference frame with an angular velocity
34
Vector Analysis 2.2 Vector Algebra
y
Figure 2.4: Three coordinate systems for a wheel in circular
motion.
w = Ok where k point to the zaxis, the velocity and accelera
tion at any point (say) P can be calculated using another fixed
point A on the rotating body (or wheel). The velocity is
drl
V p = VA + dt A + W X r,
(2.32)
and the acceleration is
d
2
rl dw
ap = aA + dt2 A+ ;Jf X r + acor + acent,
(2.33)
where
drl
acor = 2w X dt A'
(2.34)
is the Coriolis acceleration, and
acent = W X ( W X r), (2.35)
is the centripetal acceleration. It is worth noting that the ve
locity v A and acceleration aA is the velocity and acceleration
in a nonrotating frame or an inertia frame.
In addition, the differentiation of the unit vectors are con
nected by
(2.36)
35
2.2 Vector Algebra Vector Analysis
and
(2.37)
In the intrinsic coordinates, we have s = R</> where R =constant
is the radius of the wheel in circular motion. Thus, 8 =
The velocity for this circular motion is simply
(2.38)
Differentiating it with respect to time and using the relation
ships of unit vectors, we have
• .. '2
a = v = R</>et + R</> en, (2.39)
where the unit vectors are
et =cos </>i +sin </>j, en = sin </>i +cos </>j. (2.40)
0 Example 2.2: A car is travelling rapidly along a curved path
with a speed of 30 mjs at a given instance. The car is fitted with
an accelerometer and it shows that the car is accelerating along the
curved path at 2 The accelerometer also indicates that the
component of the acceleration perpendicular to the travelling direc
tion is 5 mjs2. lVhat is the direction of the total acceleration at this
instance? lVhat is the radius of the curvature? Suppose the car has
a height of 2 meters and a width of 1.6 meters, is there any danger
of toppling over?
Let () be the angle between the acceleration vector and the veloc
ity vector, and let a be the magnitude of the total acceleration. In the
intrinsic coordinates, the velocity is v =set= 30et. The accelera.tion
is given by
·2
a= set+ a(cos8et + sin8en)·
p
Therefore, we have
s
2
30
2
 =  = a sin() = 5,
p p
36
Vector Analysis 2.2 Vector Algebra
or the instantaneous radius of curvature is p = 30
2
/5 = 180m. \Ve
know that the magnitude of the acceleration is a = v2
2
+ 52 = J29.
The angle is
5
(} = tan
1
'2 ~ 68.20°.
In addition, we can assume that the centre of gravity is approx
imately at its geometrical centre. Thus, the centre is 1m above the
road surface and 0.8m from the edges of the outer wheels. If we take
the moment about the axis through the two contact points of the
outer wheels, we have the total moment
v2
1 X 1\t[ 0.8Mg ~ 2.8A1 < 0,
p
where J\tl is the mass of the car. There is no danger of toppling over.
However, if the car speeds up to v = 42 m/s (about 95 miles per
hour), there is a danger of toppling over when the moment of the
weight is just balanced by the moment of the centripetal force. 0
2.2.3 Line Integral
I
ldy
I ________ _.
dx
ds = Jdx'l + dy'l
Figure 2.5: Arc length along a curve.
An important class of integrals in this context is the line
integral which integrates along a curve r(x, y, z) = xi+ yj + zk.
For example, in order to calculate the arc length L of curve r
37
2.2 Vector Algebra Vector Analysis
as shown in Figure 2.5, we have to use the line integral.
L = [s ds = [s J dx2 + dy2 = 1x
} ~ ~ 0 xo
(2.41)
0 Example 2.9: The arc length of the parabola y(x) = tx
2
from
x = 0 to x = 1 is given by
L = 1' v't + y'
2
dx = 1' v't + x2dx
= ~ [ x ~ + ln(x + ~ > {
1
=
2
[V2 ln(V2 1)] ~ 1.14779.
2.2.4 Three Basic Operators
Three important operators commonly used in vector analysis,
especially in fluid dynamics, are the gradient operator (grad or
V), the divergence operator ( div or V ·) and the curl operator
(curl or Vx).
Sometimes, it is useful to calculate the directional derivative
of a function ¢ at the point ( x, y, z) in the direction of n
a¢ a¢ a¢ 8¢
an = n · V¢ = ax cos( a)+ ay cos(f3) +
8
z cos( I), (2.42)
where n = (cosa,cos{3,cos1) is a unit vector and o.,f3,1 are
the directional angles. Generally speaking, the gradient of any
scalar function¢ of x, y, z can be written in a sin1ilar way,
(2.43)
This is equivalent to applying the del operator V to the scalar
function¢
r7 8. 8. 8k
v = 1+J+.
ax ay az
38
(2.44)
Vector Analysis 2.2 Vector Algebra
The direction of the gradient operator on a scalar field gives a
vector field. The gradient operator has the following properties:
\l(o:'l/J+/3¢) = o:\11/J+/3\1¢, \7(1/J¢) = 1/J\14>+¢\11/J, (2.45)
where a, (3 are constants and 1/J, 4> are scalar functions.
For a vector field
u(x, y, z) = ut(x, y, z)i + u2(x, y, z)j + u3(x, y, z)k, (2.46)
the application of the operator \1 can lead to either a scalar field
or a vector field, depending on how the del operator applies to
the vector field. The divergence of a vector field is the dot
product of the del operator \1 and u
d
. t"'7 8ut 8u2 8u3
lVU= v ·U=++
8x oy oz'
(2.47)
and the curl of u is the cross product of the del operator and
the vector field u
i j k
curl u = \1 x u =
1:) 1:) {)
(JX (JY IE
(2.48)
Ut u2 u3
It is straightforward to verify the following useful identities
associated with the \1 operator:
\1 · \1 X U = 0, (2.49)
\1 X \11/J = 0, (2.50)
\1 X (1/Ju) = 1/J\1 Xu+ (\11/J) XU, (2.51)
'V· (1/Ju) = 1/J\1· u + (\11/J) · u, (2.52)
\1 x (\1 x u) = \1(\1 · u)  \7
2
u. (2.53)
One of the most common operators in engineering and sci
ence is the Laplacian operator is
2 8
2
w 8
2
w 8
2
w
\1 w = \1. ('Vw) = 8x2 + 8y2 + 8z2 , (2.54)
39
2.2 Vector Algebra Vector Analysis
for Laplace's equation
(2.55)
In engineering mathematics, it is sometimes necessary to ex
press the Laplace equation in other coordinates. In cylindrical
polar coordinates (r, </>, z), we have
\7. u = ~ 8(rur) + ~ 8ucp + 8u::
r 8r r 8</> 8z ·
(2.56)
The Laplace equation becomes
(2.57)
In spherical polar coordinates (r, 8, ¢), we have
\7. u = 1 8
2
(r
2
ur) + 1 8(sin8ue) + _1_8uct> (
2
.
58
)
r2 8r2 r sin 8 88 r sin 8 8<1> ·
The Laplace equation can be written as
(2.59)
2.2.5 Some Important Theorems
The Green theorem is an important theorem, especially in fluid
dynamics and the finite element analysis. For a vector field
Q = ui + vj in a 2D region n with the boundary r and the unit
outer normal nand unit tangent t. The theorems connect the
integrals of divergence and curl with other integrals. Gauss's
theorem states:
Jjfn(v. Q)dn = Jfs Q. ndS,
(2.60)
which connects the volume integral to the surface integral.
40
Vector Analysis 2.3 Applications
Another important theorem is Stokes's theorem:
Jfs(v x Q). kdS = £ Q. tdr = £ Q. dr, (2.61)
which connects the surface integral to the corresponding line
integral.
In our simple 2D case, this becomes
f
[ [ 8v 8u
(udx + vdy) = J ln (
0
x {)y)dxdy.
(2.62)
For any scalar functions 'l/J and ¢, the useful Green's first
identity can be written as
J 'l/JV</>dr = { (,PV
2
¢ + V'l/1 · V¢)d0., (2.63)
!on ln
where dO. = dxdydz. By using this identity twice, we get
Green's second identity
2.3 Applications
In order to show the wide applications of vector analysis, let
us apply them to study the mechanical and flow problems.
2.3.1 Conservation of Mass
The mass conservation in flow mechanics can be expressed in
either integral form (weak form) or differential form (strong
form). For any enclosed volume n, the total mass which leaves
or enters the surface S is
fspu · dA,
where p(x, y, z, t) and u(x, y, z, t) are the density and the ve
locity of the fluid, respectively. The rate of change of mass in
n is
2.3 Applications Vector Analysis
The n1ass conservation requires that the rate of loss of mass
through the surface S is balanced by the rate of change in n.
Therefore, we have
Using Gauss's theorem for the surface integral, we have
In V · (pu)dV + ! In pdV = 0.
Interchange of the integration and differentiation in the second
term, we have
1
8p
[ + V · (pu)]dV = 0.
n 8t
This is the integral form or weak form of the conservation of
mass. This is true for any volume at any instance, and subse
quently the only way that this is true for all possible choice of
n is
ap
8t+V·(pu)=0,
which is the differential form or strong form of mass conser
vation. The integral form is more useful in numerical meth
ods such as finite volume methods and finite element methods,
while the differential form is more natural for mathmnatical
analysis.
2.3.2 Saturn's Rings
We all know that Saturn's ring system ranks among the most
spectacular phenomena in the solar system. The ring systen1
has a diameter of 270,000 km, yet its thickness does not ex
ceed 100 meters. The sizes of particles in the rings vary from
centimeters to several meters, and this size distribution is con
sistent with the distribution caused by repeated collision. The
ring system has very complicated structures. One natural ques
tion is why the formed structure is a ring system, why not a
42
Vector Analysis 2.3 Applications
spherical shell system? This is a challenging topic which has
not yet fully understood. However, under some reasonable as
sumptions, we can understand why the ring system is so.
When the debris particles surrounding a planet will ulti
mately settle into flat circular rings or disks, which are the
natural consequence of energy dissipation in rotating systems.
The interparticle collisions dissipate energy while conserving
the total angular momentum. Laplace in 1802 showed that
such rings could not be solid because the tensile strength of
the known materials was too small to resist tidal forces from
Saturn. Later, Niaxwell in 1890 showed that a fluid or gaseous
ring was unstable, therefore, the rings must be particulate.
Suppose the whole particulate system consists of N par
ticles (i = 1, 2, ... , N). Its total angular momentum is h. By
choosing a coordinate system so that ( x, y) plane coincides with
the plane of the rings, and the zaxis (along k direction) is nor
mal to this plane. If we now decompose the velocity of each
particle into Vi = (vir, Vie, Viz), the total angular momentum is
then
N
h = k · [L miri x vi]
i=l
N N N
= L mi(ri X Viz)· k + L mi(ri x Vir)· k + L mi(ri x Vie)· k.
i=l i=l i=l
(2.65)
The first two terms disappear because Viz is parallel to k and
axial velocity does not contribute to the angular momentum.
So only the tangential terms are nonzero, and we have
N
h = L mirivie·
i=l
The total mechanical energy is
43
(2.66)
(2.67)
2.3 Applications Vector Analysis
where 1/J(ri) is the potential per unit mass due to Saturn's grav
ity. The interparticle collisions will dissipate the energy, there
fore, the system will evolve towards an energy minimum. From
both expressions for h and E, we can see that the minimization
of E while h is held constant requires that Viz ~ 0 and vie ~ 0.
This means that the collisions dissipate energy while flattening
the system into a disk or rings.
Now let us see why the minimization of the rotational en
ergy will also lead to the same conclusion of ring formation.
Loosely speaking, we can assume that the angular velocity
w = iJ is almost the same for all particles as t ~ oo (or any
reasonable long time) so that collisions are no longer significant
or the rate of energy dissipation is small. If there are differ
ent angular velocities, one particle may move faster and ulti
mately collides with other particles, subsequently redistribut
ing or changing its angular velocity. If we further assume that
the potential energy does not change significantly (this is true
if the particles do not n1ove significantly along the radial di
rection), thus the minimization of total energy leads to the
minimization of the total rotational energy.
This will essentially lead to a quasisteady state. With these
assumptions, we have Vie = TiW. Therefore, the angular mo
mentum becomes
h = :Lmirlw = Iw,
i=l
N
I= :Lmirl,
i=l
where I the moment of inertia of the particulate system. The
total rotational energy is
1 N 2 1 2 1 h2
T= 2 ~ m i r i w = 2Iw = 2T ~ Tmin·
l=l
In order to minimize T, we have to maximize I because h is
constant. For a disk with a radius a, a thickness t <<a and the
total mass m, we have
I= !v r
2
dm = t foR r
2
prdr fo
2
" dO= 21ftp foR r
3
dr = 1ftp ~
4
.
44
Vector Analysis
Using the density p = mj(t1r R
2
), we have
1 2
/disk= 2mR.
2.3 Applications
If all the mass is concentrated at a ring, we have
Similarly, for a solid ball with the same mass and same
radius R, we have
2 2
!ball= 5mR.
For a spherical shell, we have
Therefore, we have
2 2
!sphere = 3mR ·
Iring > /disk > !sphere > !ball·
This means that the total rotational energy is minimized if the
particle system evolves into a ring or at least a disk. This is
probably the main reason why the planetary system and rings
are formed.
45
Chapter 3
Matrix Algebra
3.1 Matrix
1\tiatrices are widely used in almost all engineering subjects. A
matrix is a table or array of numbers or functions arranged in
rows and columns. The elements or entries of a n1atrix A are
often denoted as aii. A matrix A has m rows and n columns,
(3.1)
we say the size of A is m by n, or m x n. A is square if m = n.
For example,
(
1 2 3)
A= 4 5 6 '
B=
(
ex
icosx
sinx)
eiO '
(3.2)
and
(3.3)
where A is a 2 x 3 matrix, B is a 2 x 2 square matrix, and u is
a 3 x 1 column matrix or column vector.
47
3.1 !v!atrix
1\llatrix Algebra
The sum of two matrices A and B is only possible if they
have the same size m x n, and their sum, which is also m x n, is
obtained by adding corresponding entries
C=A+B, (3.4)
where (i = 1, 2, ... , m;j = 1, 2, ... , n). We can multiply a matrix
A by a scalar a: by multiplying each entry by a:. The product
of two matrices is only possible if the number of columns of A
is the same as the number of rows of B. That is to say, if A is
mxn and B is nxr, then the product Cis mxr,
n
Ctj = (AB)ii = L aikbki·
k=l
(3.5)
n
~
If A is a square matrix, then we have An = AA ... A. The
multiplications of matrices are generally not commutive, i.e.,
AB "# BA. However, the multiplication has associativity
A(uv) = (Au)v and A(u + v) =Au+ Av.
The transpose AT of A is obtained by switching the posi
tion of rows and columns, and thus AT will be nxm if A is
mxn, (aT)ij = aii,(i = 1,2, ... ,m;j = 1,2, ... ,n). In general,
we have
(3.6)
The differentiation and integral of a matrix are done on
each member element. For example, for a 2 x 2 matrix
dA . ( dau da12 )
A dt dt
dt   da21 ~ '
(If'"" dt
(3.7)
and
(3.8)
A diagonal matrix A is a square matrix whose every entry
off the main diagonal is zero ( aii = 0 if i "# j). Its diagonal
48
1\t/atrix Algebra 3.2 Determinant
elements or entries may or n1ay not have zeros. For example,
the matrix
(
1 0 0)
I= 0 1 0
0 0 1
(3.9)
is a 3 x 3 identity or unitary matrix. In general, we have
AI=IA=A. (3.10)
A zero or null matrix 0 is a matrix with all of its elements being
zero.
3.2 Determinant
The determinant of a square matrix A is a number or scalar
obtained by the following recursive formula or the cofactor or
Laplace expansion by column or row. For example, expanding
by row k, we have
n
det(A) = IAI = L(1)k+iakjMkj'
j=l
(3.11)
where Mij is the determinant of a minor matrix of A obtained
by deleting row i and column j. For a simple 2 x 2 matrix, its
determinant simply becomes
(3.12)
It is easy to verify that the determinant has the following prop
erties:
laAI =alAI, IABI = IAIIBI, (3.13)
where A and Bare the same size (nxn).
A nxn square matrix is singular if IAI = 0, and is nonsin
gular if and only if IAI =/= 0. The trace of a square matrix tr(A)
49
3.3 Inverse
1\llatrix Algebra
is defined as the sum of the diagonal elements,
n
tr(A) = L ~ i = a11 + a22 + ... + G.nn· (3.14)
i=l
The rank of a matrix A is the number of linearly indepen
dent vectors forming the tnatrix. Generally, the rank of A is
rank( A) ~ min( m, n). For a n x n square matrix A, it is non
singular if rank( A)= n.
From the basic definitions, it is straightforward to prove the
following
(AB ... z)T = zT ... BT AT,
(3.15)
IAB .... ZI = IAIIBI ... IZI, (3.16)
tr(A) = tr(AT), (3.17)
tr(A +B) = tr(A) + tr(B), (3.18)
tr(AB) = tr(BA), (3.19)
1 1
det(A ) = det(Ar
(3.20)
det(AB) = det(A)det(B). (3.21)
3.3 Inverse
The inverse matrix At of a square matrix A is defined as
A 
1
A = AA 
1
= I.
(3.22)
It is worth noting that the unit matrix I has the same size as
A. The inverse of a square matrix exists if and only if A is
nonsingular or det(A) =I= 0. From the basic definitions, it is
straightforward to prove that the inverse of a matrix has the
following properties
(A 1)1 =A,
(3.23)
50
1\t/atrix Algebra 3.3 Inverse
and
(3.24)
A simple useful formula for obtaining the inverse of a 2 x 2
matrix is
(
a b ) l 1 ( d
c d =(adbe) c
(3.25)
0 Example 9.1: For two matrices
A= ( ~ 1 i ~ ) , B = ( ~ ~
1
) ,
3 2 2 1 7
we have
where
V
11
= 1 x 1 + 2 x 2 + 3 x 1 = 8, V
12
= 1 x ( 1) + 2 x 3 + 3 x 7 = 26;
V21 = 1 X 1 + 1 X 2 + 0 X 1 = 1, V22 = 1 X ( 1) + 1 X 3 + 0 X 7 = 4;
V
31
= 3 x 1 + 2 x 2 + 2 x 1 = 9, V
32
= 3 x ( 1) + 2 x 3 + 2 x 7 = 17.
Thus,
(
8 26)
AB=V= 1 4 .
9 17
However, BA does not exist. The transpose matrices of A and B
are
AT= 0 ~ 1 ~ ) .
B
T = ( 1 2 1 )
1 3 7 .
Similarly, we have
51
3.4 !\latrix Exponential Matrix Algebra
The inverse of A is
and the determinant of A is
detiAI = 9.
The trace of A is
tr(A) = A11 + A22 + Aaa = 1 + 1 + 2 = 4.
0
3.4 Matrix Exponential
Sometimes, we need to calculate exp[A], where A is a square
matrix. In this case, we have to deal with matrix exponentials.
The exponential of a square matrix A is defined as
A ~ 1 n 1 2
e =L..,A =I+A+
2
A + ....
n=O nJ
(3.26)
where I is a unity matrix with the same size as A, and A
2
=
AA and so on. This (rather odd) definition in fact provides
a n1ethod to calculate the matrix exponential. The matrix
exponentials are very useful in solving systems of differential
equations.
0 Example 9.2: For a simple matrix
A=U
n.
we have
( e'
eA
 0
~ ) .
For
A=(: : ),
52
1\t/atrix Algebra 3.5 Hermitian and Quadratic Forms
we have
eA  ( ~ ( 1 + e2t) ~ ( e 2 t  1) )
 !(e2t 1) !(1 + e2t) .
For a slightly complicated matrix
A= ( t w )·
w t .
we have
A= ( et cosw et sinw )
e et sinw et cosw ·
0
As you see, it is quite complicated but still straightforward
to calculate the matrix exponentials. Fortunately, it can be eas
ily done using a computer. By using the power expansions and
the basic definition, we can prove the following useful identities
tA ~ 1 n i
2
2
e = L
1
(tA) =I+ tA + 2A + ... ,
n=O n.
(if AB = BA),
!!:_etA = AetA = etA A
dt '
(eA)1 = eA,
det( eA) = etrA.
3.5 Hermitian and Quadratic Forms
(3.27)
(3.28)
(3.29)
(3.30)
(3.31)
(3.32)
The matrices we have discussed so far are real matrices because
all their elements are real. In general, the entries or elements
of a matrix can be complex numbers, and the matrix becomes
a complex matrix. For a matrix A, its complex conjugate A*
53
3.5 Hermitian and Quadratic Forms Matrix Algebra
is obtained by taking the complex conjugate of each of its ele
ments. The Hermitian conjugate At is obtained by taking the
transpose of its complex conjugate matrix. That is to say, for
A = ( : ~
1
1 , : ~ ~ · ::: ) , (3.33)
we have
= ~ ~ :::),
(3.34)
and
(3.35)
A square matrix A is called orthogonal if and only if A 
1
=
AT. If a square matrix A satisfies A* = A, it is said to be an
Hermitian matrix. It is an antiHermitian matrix if A* =A.
If the Hermitian matrix of a square matrix A is equal to the in
verse of the matrix (or At = A 
1
), it is called a unitary matrix.
0 Example 9. 9: For a matrix
a = ( 2 ~ i 3  2i 1 )
et11' 0 1  i1r '
its complex conjugate a• and Hermitian conjugate at are
a· = ( 2 : i 3 + 2i 1 )
e
1
11' 0 1 + i1r '
For the rotation matrix
A = ( cos 9 sin 9 ) .
 sin 9 cos 9 ·
54
1\t/atrix Algebra 3.5 Hermitian and Quadratic Forms
its inverse and transpose are
sin() )
cos() '
and
Since cos
2
() + sin
2
() = 1, we have AT = A l. Therefore, the original
matrix A is orthogonal.
0
A very useful concept in engineering mathematics and com
puting is quadratic forms. For a real vector qT = ( Q1, Q2, q3, ... , qn)
and a real square matrix A, a quadratic form 'l,b(q) is a scalar
function defined by
'f,b(q) = qT Aq
= ( Ql Q2 •••
(
An
Qn ) ~ ~ ~
Ant
which can be written as
n n
1/J(q) = LL QiAijQj·
i=l j=l
) ( l:}
(3.36)
(3.37)
Since 'l/J is a scalar, it should be independent of the coordi
nates. In the case of a square matrix A, 'l,b might be more
easily evaluated in certain intrinsic coordinates Qb Q2, ... Qn.
An important result concerning the quadratic form is that it
can always be written through appropriate transformations as
n
1/J(q) =I: AiQr = AtQI + A 2 Q ~ + ... A n Q ~ .
(3.38)
i=l
The natural extension of quadratic forms is the Hermitian form
that is the quadratic form for complex Hermitian matrix A.
55
3.6 Solution of linear systems !via trix Algebra
Furthermore, the matrix A can be linear operators and func
tionals in addition to numbers.
0 Example 9.4: For a vector q = (q1, q2) and the square matrix
(
1 2)
A= 2 1 '
we have a quadratic form
0
3.6 Solution of linear systems
A linear system of m equations for n unknowns
(3.39)
can be written in the compact form as
(3.40)
or simply
Au=b. (3.41)
In the case of m = n, we multiply both sides by A l (this
is only possible when m = n),
A
1
Au=A
1
b,
(3.42)
56
1\t/atrix Algebra 3.6 Solution of linear systems
we obtain the solution
u=A
1
b.
(3.43)
A special case of the above equation is when b =Au, and
this becomes an eigenvalue problem. An eigenvalue A and cor
responding eigenvector v of a square matrix A satisfy
Av = Av, (3.44)
or
(A Al)v = 0. (3.45)
Any nontrivial solution requires
au A at2
=0, (3.46)
a21 a22 A
... ann A
which is equivalent to
(3.47)
In general, the characteristic equation has n solutions. Eigen
values have the interesting connections with the matrix,
n
tr( A) = L B.ii = A 1 + A2 + ... + A
0
•
i=l
(3.48)
For a symmetric square matrix, the two eigenvectors for two
distinct eigenvalues Ai and Aj are orthogonal vTv = 0.
Some useful identities involving eigenvalues and inverse of
matrices are as follows:
(AB ... z)
1
= z
1
... a
1
A 
1
,
Ai = eig(A),
57
(3.49)
(3.50)
3.6 Solution of linear systems
eig(AB) = eig(BA),
tr(A) = L Aii = L Ai,
0 Example 9.5: For a simple 2 x 2 matrix
its eigenvalues can be determined by
1
1 A 5 I
2 4 A = O,
or
{1  A)(4 A)  2 X 5 = 0,
which is equivalent to
(A+ 1){A 6) = 0.
Matrix Algebra
(3.51)
(3.52)
(3.53)
Thus, the eigenvalues are At = 1 and A
2
= 6. The trace of A is
tr(A) =Au + A22 = 1 + 4 = 5 =At + A2.
In order to obtain the eigenvector for each eigenvalue, we assume
For the eigenvalue At = 1, we plug this into
and we have
or
1{1)
2
IAAllv = 0,
5 I ( VV2t ) = 0,
4{1)
I ; ~ I ( ~ : ) = o,
which is equivalent to
or
58
1\t/atrix Algebra 3.6 Solution of linear systems
This equation has infinite solutions, each corresponds to the vector
parallel to the unit eigenvector. As the eigenvector should be normal
ized so that its modulus is unity, this additional condition requires
+ = 1,
which means
(
5v2 )2 2 1
2 +v2 .
We have VI = v2 = 2/v'29. Thus, we have the first set of
eigenvalue and eigenvector
Vt = ( to ) .
V29
At= 1, (3.54)
Similarly, the second eigenvalue .X
2
= 6 gives
1
1  6 5 I ( Vt ) = 0.
2 46 v2
Using the normalization condition v'f + = 1, the above equation
has the following solution
For a linear system Au= b, the solution u =Alb gen
erally involves the inversion of a large matrix. The direct in
version becomes impractical if the n1atrix is very large (say, if
n > 1000). l\1any efficient algorithms have been developed for
solving such systems. Gauss elimination and L U decomposition
are just two examples.
59
Chapter 4
Complex Variables
Although all the quantities are real variables in the physical
world, however, it is son1etimes easy or even necessary to use
complex variables in mathematics and engineering. In fact,
the techniques based on complex variables are among the most
powerful methods for mathematical analysis and solutions of
mathematical models.
4.1 Complex Numbers and Functions
lviathematically speaking, a complex number z is a generalized
set or the order pair of two real numbers (a, b), written in the
form of
z =a+ ib,
·2 1
z ='
a, bE 'R, (4.1)
which consists of the real part ~ ( z ) =a and the imagery part
SJ<( z) = b. It can also be written as the order pair of real nunl
bers using the notation (a, b). The addition and substraction
of two complex numbers are defined as
(a+ ib) ± (c + id) =(a± c)+ i(b ±d). (4.2)
The multiplication and division of two complex numbers are in
the similar way as expanding polynon1ials
(a+ ib) · (c + id) = (ac bd) + i(ad +be), (4.3)
61
4.1 Complex Numbers and Functions Complex Variables
and
a + ib ac + bd . be  ad
c _+_i_d = c2 + d2 + ·
(4.4)
Two complex numbers are equal a + ib = c + id if and only if
a= c and b =d. The complex conjugate or simply conjugate
z (also z*) of z =a+ ibis defined as
z =a ib. (4.5)
The order pair (a, b), similar to a vector, implies that a
geometrical representation of a complex number a + ib by the
point in an ordinary Euclidean plane with xaxis being the
real axis and yaxis being the imaginary axis (iy). This plane
is called the complex plane. This representation is often called
the Argand diagram (see Figure 4.1). The vector representation
starts from (0, 0) to the point (a, b). The length of the vector
is called the magnitude or modulus or the absolute value of the
complex number
r = lzl = v' a
2
+ b
2
.
(4.6)
Figure 4.1: Polar representation of a complex number.
The angle () that the vector makes with the positive real
axis is called the argument (see Fig 4.1),
() = arg z.
(4.7)
In fact, we may replace () by () + 2n7r ( n E N). The value range
1r < () :::; 1r is called the principal argument of z, and it is
62
Complex Variables 4.1 Complex Numbers and Functions
usually denoted as Argz. In the complex plane, the complex
number can be written as
z = rei() = r cos( 0) + ir sin( 0). (4.8)
This polar form of z and its geometrical representation can
result in the Euler's formula which is very useful in the complex
analysis
ei
6
= cos(O) + isin(O). (4.9)
The Euler formula can be proved using the power series. For
any z E C, we have the power series
z2 zn
ez = 1 + z + 2f + ... + ~ + ... ,
and for a special case z = iO, we have
ilJ 1 . 0 02 i03
e = + z  2! + 31  ... ,
02 03
= (1 2f + ... ) + i(O 3f + ... ).
Using the power series
and
03 os
sin 0 = 0   +   ... ,
3! 51
02 04
cos 0 = 1   +   ... ,
2! 4!
(4.10)
( 4.11)
(4.12)
(4.13)
we get the wellknow Euler's formula or Euler's equation
e ilJ = cos 0 + i sin 0.
(4.14)
For 0 = 1r, this leads to a very interesting formula
(4.15)
If we replace 0 by 0, the Euler's formula becomes
e ilJ = cos( 0) + i sin( 0) = cos 0  i sin 0.
(4.16)
63
4.1 Complex Numbers and Functions Complex Variables
Adding this equation to ( 4.14), we have
ei
0
+ eiO = 2 cos 8,
( 4.17)
or
eiO + eiO
cos8 =
2
(4.18)
Similarly, by deducting (4.16) from (4.14), we get
eiO eiO
sin8 =
2
i
( 4.19)
For two complex numbers z
1
= r
1
ei
01
and z
2
= r2ei
02
, it is
straightforward to show that
ZJZ2 = r1r2ei(o
1
+
02
) = r1r2[cos(a1 + a2) + isin(at + a2)],
( 4.20)
which can easily be extended to get the wellknown de l\1oivre's
formula
[cos( B)+ isin(8)r = cos(n8) + isin(n8). (4.21)
0 Example 4.1: Find z
4
if z = 1 + J3i. lVe can evaluate it by
direct calculation
z
4
= (1 + J3i)
4
= [(1 + J3i)
2
)
2
= [1  3 + 2J3i)
2
= 2
2
( 1 + v'3i)
2
= 4(1  3 2v'3) = 8 8v'3i.
lVe can also use Moivre's formula. The modulus of z is r = lzl =
V1
2
+ J3
2
= 2. The argument()= tan
1
f = 7r/3 = 60°. Thus,
z = 2e7rf
3
. We now have
z
4
= 2
4
e
4
7r/J = 16(cos
4
11" + isin
4
11")
3 3
= 16(  ~  v; i) = 8 8v'3i,
which is exactly the same result as we obtained earlier. The second
method becomes much quicker if you want to evaluate (say) z
100
.
0
64
Complex Variables 4.2 Hyperbolic Functions
5 ~    , ~     ~     ~     ~     ~     ~
\
\
~
\
\
'
'
'
'
'
'
' ...
>C
 ·  ·    ~  ·  ·  ·  

~ 0
c
u;
sinh
 cosh
··
tanh
5
3 2 1 0 1 2 3
X
Figure 4.2: Hyperbolic functions.
4.2 Hyperbolic Functions
Hyperbolic functions occur in many applications and they can
be thought as the complex analogues of trigonometric func
tions. The fundamental definitions are
and
ex ex
sinhx =
2
sinhx
tanhx =   h  ~
COS X
ex+ ex
coshx =
2
,
1
coth x = h.
tan x
(4.22)
(4.23)
Figure 4.2 shows the variation of sinh, cosh, and tanh. If
we replace x by ix and use Euler's formula, then we have
eix _ eix
sinh ix =
2
1 [( . . ) ( . . )] . .
=
2
COSX + ZSlnX  COSX ZSlnX = ZSlllX. (4.24)
65
4.2 Hyperbolic Functions Complex Variables
Similarly, we have
1 °
0
cosh ix = (ezx + ezx)
2
= ~ [ ( c o s x + isinx) + (cosx isinx)] = cosx.
In a similar fashion, we can also prove that
cosix = coshx, sin ix = i sinh x.
Some identities are as follows:
cosh
2
x sinh
2
x = 1,
sinh 2x = 2 sinh x cosh x,
and
cosh 2x = sinh
2
x + cosh
2
x.
(4.25)
(4.26)
(4.27)
(4.28)
(4.29)
0 Example 4.2: Prove that cosh
2
x sinh
2
x = 1. From the
definitions, we have
and
Thus, we have
1
= 4[2 (2)) = 1.
0
The inverses of hyperbolic functions are defined in a similar
way as trigonometric functions. For example, y = cosh x, its
inverse is defined as x = cosh t y. From the basic definitions,
we have
sinhx + coshx =ex. (4.30)
66
Complex Variables 4.3 Analytic Functions
Using sinhx = Vcosh
2
x 1, we have
Vcosh
2
x 1 + coshx =ex, ( 4.31)
or
(4.32)
which gives
4.3 Analytic Functions
Analytic Functions
Any function of real variables can be extended to the function
of complex variables in the same fonn while treating the real
numbers x as x + iO. For example, f ( x) = x
2
, x E 'R becomes
f ( z) = z
2
, z E C. Any complex function f ( z) can be written as
f(z) = f(x + iy) = ~ ( f ( z ) ) + ir;s(J(z))
= u(x, y) + iv(x, t), (4.34)
where u(x, y) and v(x, y) are realvalued functions of two real
variables.
A function f ( z) is called analytic at z
0
if f' ( z) exists for all z
in smne €neighborhood of zo, that is to say, it is differentiable
in some open disk lz z
0
1 < f. If f(z) = u + iv is analytic at
every point in a domain n, then u(x, y) and v(x, y) satisfying
the CauchyRiemann equations
8u
8x
8v
8y'
(4.35)
Conversely, if u and v of f ( z) = u + iv satisfy the Cauchy
Riemann equation at all points in a domain, then the complex
function f(z) is analytic in the same domain. For example, the
67
4.3 Analytic Functions Complex Variables
elementary power function w = zn, ( n > 1) is analytic on the
whole plane, w = pei(j>, z = rei
9
, then
p = rn, 4> = n8.
(4.36)
The logarithm is also an elementary function w = In z
In z =In lzl + i arg(z) =In r + i(8 + w1rk), ( 4.37)
which has infinitely many values, due to the multiple values of
8, with the difference of 21rik(k = 0, ±1, ±2, ... ). If we use the
principal argument Argz, then we have the principal logarithm
function
Ln(z) =In lzl + Argz. (4.38)
If we differentiate the CauchyRiemann equations, we have
fJ2uj8x8y = 8
2
uj8y8x. After some calculations, we can reach
the following theoren1. For given analytic function f(z)
u + iv, then both u and v satisfy the Laplace equations
(4.39)
This is to say, both real and imaginary parts of an analytic
function are harmonic.
A very interesting analytical function is the Riemann zeta
function ((s), which is defined by
00 1
((s) = ~ ,
L ns
n=l
(4.40)
where s is a complex number with its real part more than unity.
That is s E C and ~ ( s ) > 1. This function (infinite series) is
analytic, and it can be extended for all complex numbers s =I= 1.
For example,
1 1 7r
2
((2) = 1 + 22 + 32 + ... = 6'
( 4.41)
but
1 1 1
((1) = 1 + 2 + 3 + 4 + ... = 00.
(4.42)
68
Complex Variables 4.3 Analytic Functions
This ((s) function has trivial zeros s = 2, 4, 6, ... and it
also has nontrivial zeros.
There is a famous unsolved problem, called the Riemann
hypothesis, related to this function. The Riemann hypothesis
conjectured by Bernhard Reimann in 1859 states that all real
parts of any nontrial zero of the Riemann zetafunction ((s)
are !· That is to say, all the nontrivial zeros should lie on a
straight lines= ! + iy. This is amilliondollar open problem
as the Clay Mathematical Institute in 2000 offered a million
dollars to search for a proof, and yet it still remains unsolved.
Laurent Series
For an analytic function p(z), one of important properties is
the singularity such as the pole. If p( z) can be written as
q(z)
p(z) = ( )n'
z zo
(4.43)
where n > 0 is a positive integer while q(z) =/= 0 is analytic
everywhere in the neighbourhood containing z = zo, we say
that p( z) has a pole of order n at z = z
0
. The above definition
is equivalent to say that the following limit is finite
lim [p(z)(z zot] = (,
zzo
11(11 < oo, ( E C. (4.44)
Any analytic function f ( z) can be expanded in terms of the
Taylor series
( 4.45)
This expansion is valid inside the analytic region. However,
if the function f ( z) has a pole of order n at z = zo and it is
analytic everywhere except at the pole, we can then expand the
function p(z) = (zz
0
)n f(z) in the standard Taylor expansion.
69
4.4 Complex Integrals Complex Variables
This means that original function f ( z) can be written as a
power series
!(
)
O_n O_t
z=( ) + ... +( )
zzon zzo
oo(z zo) + ... + ok(z zo)k + ... , (4.46)
which is called a Laurent series, and it is an extension of the
Taylor series. In this series, it is often assumed that On =F 0.
The tern1s with the inverse powers an/(zzo)n+ ... +at/(z
zo) are called the principal part of the series, while the usual
terms ao(z zo) + ... + ok(z zo)k + ... are called the analytic
part.
Furthermore, the most in1portant coefficient is probably
Ot which is called the residue of f(z) at the pole z = zo.
In general, the Laurent series can be written as
00
f(z) = L ok(z zo)k, (4.47)
k=n
where n may be extended to include an infinite number of terms
n+ oo.
4.4 Complex Integrals
Given a function f ( z) that is continuous on a piecewise smooth
curve r, then the integral over r, fr f(z)dz, is called a contour
or line integral of f ( z). This integral has similar properties as
the real integral
frlof(z) + fig(z)]dz = o £ f(z)dz + .8£g(z)dz. (4.48)
If F(z) is analytic and F'(z) = f(z) is continuous along a curve
r, then
l f(z)dz = F[z(b)] F[z(a)]. (4.49)
70
Complex Variables 4.4 Complex Integrals
Cauchy's Integral Theorem
We say a path is simply closed if its end points and initial
points coincide and the curve does not cross itself. For an
analytic function f(z) = u(x, y) + iv(x, y), the integral on a
sin1ply closed path
I= lr f(z)dz = fr (u + iv)(dx + idy)]
= fr (udx vdy) + i fr (vdx + udy).
(4.50)
By using the Green theorem, this becomes
[ 8u av . [ 8u 8v
I = Jn ( ay  ax )dxdy + z Jn (ax  ay )dxdy.
( 4.51)
Fron1 the CauchyRiemann equations, we know that both in
tegrals are zero. Thus, we have Cauchy's Integral Theorem,
which states that the integral of any analytic function f(z) on
a simply closed path r in a simply connected domain n is zero.
That is
fr f(z)dz = 0.
SS(z)
Figure 4.3: Contours for Cauchy integrals.
This theorem is very important as it has interesting conse
quences. If the close path is decomposed into two paths with
71
4.4 Complex Integrals Complex Variables
reverse directions r
1
and r
2
(see Figure 4.3), then r1 and r2
form a close path, which leads to
{ f(z)dz = f f(z)dz.
lr1 lr2
(4.52)
That is to say that the integrals over any curve between two
points are independent of path. This property becomes very
useful for evaluation of integrals. In fact, this can be extended
to the integrals over two closed paths r and y such that y is a
very small circular path inside r. Using a small cut with two
curves C
1
and C
2
so that these two curves combine with r and
y form a closed contour (see Figure 4.3), the Cauchy integral
theorem implies that
fr f(z)dz = L f(z)dz,
(4.53)
since the contribution from the cut is zero.
For an analytic function with a pole, we can make the con
tour y sufficiently small to enclose just around the pole, and
this makes the calculation of the integral much easier in some
cases.
For the integral of p( z) = f ( z) / ( z  zo) over any simply
closed path r enclosing a point zo in the domain n,
I= frp(z)dz, (4.54)
we can use the Laurent series for p( z)
a1 ) ( k
p(z) = ( ) +ao(zzo + ... +ak zzo) + ... , (4.55)
z zo
so that the expansion can be integrated term by term around
a path. The only nonzero contribution over a small circular
contour is the residue a1· We have
I= f p(z)dz = 21ria1 = 21ri Res[p(z)]j ,
k ~
(4.56)
72
Complex Variables 4.4 Complex Integrals
which can be written in terms of f(z) as
1 i f(z)

2
. dz = f(zo).
1rz r z zo
(4.57)
Similarly, this can be extended for higher derivatives, and we
have
j f(z) dz = 21rij(n)(zo).
Jr (z zo)n+l nl
00
00
Figure 4.4: Contour for the integral J(o, (3).
Residue Theorem
For any analytic f ( z) function in a domain n except isolated
singularities at finite points Z1, z2, ... , ZN, the residue theorem
states
N
i f(z)dz = 21fi L Resf(z)lzk'
r k=t
where r is a simple closed path enclosing all these isolated
points. If f(z) has a pole of order Nat zo, the following forn1ula
gives a quick way to calculate the residue
_ 1 . dNt [(z zo)N f(z)]
Resf(z)lzo  (N 1)1 } . ! . . . ~ o dzN1 .
(4.58)
The residue theoren1 serves a powerful tool for calculating
some real integrals and summation of series, especially when
73
4.4 Complex Integrals Complex Variables
the integrand is a function of sin and cos that can be changed
into a complex integral. The real integral f ~ o o 1/J(x)dx becomes
27ri multiplying the sun1 of the residues of 1/J(x) at the poles in
the upper halfspace.
0 Example 4,.9: In order to evaluate the integral
it is necessa1y to construct a contour (see Figure 4.4). As the function
</> = ei
02
< / (/3
4
+ (
2
) has two poles ( = +i/3
2
and i/3
2
and only one
pole ( = +i/3
2
is in the upper half plane, we can construct a contour
to encircle the pole at ( = i{J2 by adding an additional arc at the
infinity (( + oo) on the upper half plane. Combining the arc with
the horizontal line from the integral limits from oo to oo along the
(axis, a contour is closed. Hence, we have
eio2c; /(( + ij32) f(()
4> = (  i(32 = (  ij32 '
where f(() = eio
2
c; /(( + i/3
2
). Using the residue theorem, we have
In a special case when o = 0, we have
f
oe 1 1r
2 j34 d( = /l2.
00 ( + fJ
Another important topic in complex variables is the confor
mal mapping. The essence of a conformal mapping
w = f(z), z,w EC, (4.59)
is that this mapping preserves the angles between curves and
their orientations. One of the widely used mappings is 1\tlobius
linear fractional transformation
az+f3
W= '
"fZ + 8
(4.60)
74
Complex Variables 4.4 Complex Integrals
By choosing the appropriate coefficients o., (3, "/, {J E C, this
mapping can include all major geometrical transformations such
as translations, rotations, inversion, and expansions and con
tractions. Conformal mappings are useful in solving steady
state problems involving harmonic functions by transforming
the problem from a complicated geon1etrical domain to a reg
ular domain such as circles and rectangles, and subsequently
the techniques based on conformal mapping are widely used in
solving Laplace's equation in engineering.
75
Chapter 5
Ordinary Differential
Equations
1viost n1athematical models in engineering are formulated in
terms of differential equations. If the variables or quantities
(such as velocity, temperature, pressure) change with other in
dependent variables such as spatial coordinates and tin1e, their
relationship can in general be written as a differential equation
or even a set of differential equations.
5.1 Introduction
An ordinary differential equation (ODE) is a relationship be
tween a function y( x) of an independent variable x and its
derivatives y', y", ... , y(n). It can be written in a generic form
,y,( I II (n)) _ Q
';l' x,y,y,y , ... ,y  . (5.1)
The solution of the equation is a function y = f ( x), satisfying
the equation for all X in a given domain n.
The order of the differential equation is equal to the order
n of the highest derivative in the equation. Thus, the Riccati
equation:
y' + a(x)y
2
+ b(x)y = c(x), (5.2)
77
5.2 First Order ODEs Ordinary Differential Equations
is a first order ODE, and the following equation of Eulertype
(5.3)
is a second order. The degree of the equation is defined as the
power to which the highest derivative occurs. Therefore, both
Riccati equation and Euler equation are of the first degree. An
equation is called linear if it can be arranged into the form
an(x)y(n) + ... + a1 (x)y' + ao(x)y = ¢(x), (5.4)
where all the coefficients depend on x only, not on y or any
derivatives. If any of the coefficients is a function of y or any
of its derivatives, then the equation is nonlinear. If the right
hand side is zero or ¢(x) = 0, the equation is homogeneous. It
is called nonhomogeneous if ¢( x) =I= 0.
The solution of an ordinary differential equation is not al
ways straightforward, and it is usually very complicated for
nonlinear equations. Even for linear equations, the solutions
can only be obtained for a few simple types. The solution of
a differential equation generally falls into three types: closed
form, series form and integral form. A closed form solution is
the type of solution that can be expressed in terms of elemen
tary functions and some arbitrary constants. Series solutions
are the ones that can be expressed in terms of a series when a
closedform is not possible for certain type of equations. The
integral form of solutions or quadrature is sometimes the only
form of solutions that are possible. If all these forms are not
possible, the alternatives are to use approximate and numerical
solutions.
5.2 First Order ODEs
5.2.1 Linear ODEs
The general form of a first order linear differential equation can
be written as
y' + a(x)y = b(x). (5.5)
78
Ordinary Differential Equations 5.2 First Order ODEs
This equation is always solvable using the integrating factor
and it has a closed form solution.
Multiplying both sides of the equation by explf a(x)dx],
which is often called the integrating factor, we have
y'ei a(x)dx + a(x)yei a(x)dx = b(x)ei a(x)dx, (5.6)
which can be written as
[yei a(x)dx]' = b(x)ei a(x)dx.
(5.7)
By simple integration, we have
yei a(x)dx = J b(x)ei a(x)dxdx +C. (5.8)
So its solution becomes
y(x) = e I a(x)dx j b(x)ei a(x)dxdx + ce I a(x)dx, (5.9)
where C is an integration constant. The integration constant
can be determined if extra requirements are given, and these
extra requirements are usually the initial condition when time
is zero or boundary conditions at some given points which are
at the domain boundary. However, the classification of condi
tions may also depend on the meaning of the independent x. If
xis spatial coordinate, then y(x = 0) =Yo is boundary condi
tion at x = 0. However, if x = t means tilne, then y(t = 0) =Yo
can be thought of as the initial condition at t = 0. Neverthe
less, one integration constant usually requires one condition to
determine it.
0 Example 5.1: lVe now try to solve the ordina1y differential
equation ~ ~ + ty(t) = t with an initial condition y(O) = 0. As
a(t) = t, b(t) = t, its general solution is
y(t) = e I tdt j ( t)ei tdtdt + ce I tdt
79
5.2 First Order ODEs Ordinary Differential Equations
t2 t2 t2 t2
= e"TeT +GeT= 1 +GeT.
From the initial condition y(O) = 0 at t = 0, we have
0 = 1 +C, or C=l.
Thus, the solution becomes
0
5.2.2 Nonlinear ODEs
For some nonlinear first order ordinary differential equations,
sometimes a transform or change of variables can convert it into
the standard first order linear equation (5.5). For example, the
Bernoulli's equation can be written in the generic form
y' + p(x)y = q(x)yn,
(5.10)
In the case of n = 1, it reduces to a standard first order linear
ordinary differential equation. By dividing both sides by yn
and using the change of variables
we have
1
u(x) =,
yn1
1
(1 n)y'
U= '
yn
u' + (1 n)p(x)u = (1 n)q(x),
(5.11)
(5.12)
which is a standard first order linear differential equation whose
general solution is given earlier in this section.
0 Example 5.2: To solve y'(x) + xy = y
20
, we first use u(x) =
1/y
19
, and we u' = 19y' jy
20
. The original equation becomes
u' 19xu = 19,
whose general solution is
u(x) = Ae
19
x + 1.
80
Ordinary Differential Equations 5.3 Higher Order ODEs
Therefore, the solution to the original equation becomes
19 1
y = Ael9x + 1'
or
0
5.3 Higher Order ODEs
Higher order ODEs are more complicated to solve even for the
linear equations. For the special case of higherorder ODEs
where all the coefficients an, ... , at, ao are constants,
any(n) + ... + aty' + aoy = f(x), (5.13)
its general solution y( x) consists of two parts: the complemen
tary function Yc( x) and the particular integral or particular
solution y;(x). We have
y(x) = Yc(x) + y;(x). (5.14)
5.3.1 General Solution
The complementary function is the solution of the linear homo
geneous equation with constant coefficients and can be written
in a generic form
(n) + (n1) + + I +  0
anYc anlYe ··· alYc ao  ·
(5.15)
Assuming y = AeAx, we get the polynomial equation of char
acteristics
(5.16)
which has n roots in general. Then, the solution can be ex
pressed as the summation of various terms Yc(x) = Ek=l CkeAkx
if the polynomial has n distinct zeros At, ... An. For complex
roots, and complex roots always occur in pairs A= r ± iw, the
corresponding linearly independent terms can then be replaced
by erx[Acos(wx) + Bsin(wx)].
81
5.3 Higher Order ODEs Ordinary Differential Equations
The particular solution y;(x) is any y(x) that satisfies the
original inhomogeneous equation (5.13). Depending on the
form of the function f ( x), the particular solutions can take vari
ous forms. For most of the combinations of basic functions such
as sin x, cos x, ekx, and xn, the method of the undetermined c ~
efficients is widely used. For f(x) = sin(ax) or cos(ax), then
we can try y; = A sin ax+ B sin ax. We then substitute it into
the original equation (5.13) so that the coefficients A and B can
be determined. For a polynomial f(x) = xn(n = 0, 1, 2, .... , N),
we then try y; = A+ Bx + Cx
2
+ ... + Qxn (polynomial).
For f(x) = ekxxn, y; = (A+ Bx + Cx
2
+ ... Qxn)ekx. Sim
ilarly, f ( x) = ekx sin ax or f ( x) = ekx cos ax, we can use
y; = ekx (A sin ax + B cos ax). l\!Iore general cases and their
particular solutions can be found in various textbooks.
0 Example 5.9: In order to solve the equation y"'(x) 2y"(x)
y' ( x) + 2y( x) = sin x, we have to find its complementary function
Yc(x) and its particular integral y*(x). we first tzy to solve its com
plementary equation or homogeneous equation
y"'(x) 2y"(x) y'(x) + 2y(x) = 0.
Assuming that y = AeAx, we have the characteristic equation
or
(.X 1)(.\ + 1)(.\ 2) = 0.
Thus, three basic solutions are ex, ex and e
2
x. The general comple
mentazy function becomes
As the function f ( x) = sin x, thus we can assume that the particular
integral takes the form y*(x) = asinx + bcosx. Substituting this
into the original equation, we have
(acosx + bsinx) 2(asinx bcosx)
(acosx bsinx) + 2(asinx + bcosx) = sinx,
82
Ordinary Differential Equations 5.3 Higher Order ODEs
or
(b + 2a + b + 2a 1) sinx + (a+ 2b a+ 2b) cosx = 0.
Thus, we have
4a + 2b = 1,  2a + 4b = 0,
whose solution becomes
1
a= 5'
1
b = 10'
Now the particular integral becomes
y*(x) = ~ sinx +
1
1
0
cosx.
Finally, the general solution
y = ~ s i n x + ..!:..cosx + Aex + Bex + Ce
2
x.
5 10
0
The methods for finding particular integrals work for most
cases. However, there are some problems in the case when the
righthand side of the differential equation has the same form
as part of the complementary function. In this case, the trial
function should include one higher order term obtained by mul
tiplying the standard trial function by the lowest integer power
of x so that the product does not appear in any term of the
complementary function. Let us see an example.
0 Example 5.4: Consider the equation
y"(x) 3y'(x) + 2y(x) =ex.
Usingy(x) = Ke>..x, we have the characteristic equation
Its complementary function is
As the right hand side f ( x) = ex is of the same form as the first term
of Yc, then standard trial function aex cannot be a particular integral
83
5.3 Higher Order ODEs Ordinary Differential Equations
as it automatically satisfies the homogenous equation y"(x)3y'(x)+
2y(x) = 0. We have to tzy y; = (a+ bx)ex first, and we have
Dividing both sides by ex, we have
(a+ 2b + bx) 3(a + b + bx) + 2(a + bx) = 1,
or
b = 1.
As there is no constraint for a, thus we take it to be zero (a= 0 ). In
fact, any nonzero aex can be absorbed into Aex. Thus, the general
solution becomes
0
5.3.2 Differential Operator
A very useful technique is to use the method of differential
operator D. A differential operator D is defined as
d
D=
 dx·
(5.17)
Since we know that De>.x = Ae>.x and Dne>.x = Ane>.x, so they
are equivalent to D ...... A, and nn ...... An. Thus, any polyno
mial P(D) will map to P(A). On the other hand, the integral
operator nt = J dx is just the inverse of the differentiation.
The beauty of using the differential operator form is that one
can factorize it in the same way as for factorizing polynomials,
then solve each factor separately. Thus, differential operators
are very useful in finding out both the complementary func
tions and particular integral.
0 Example 5.5: To find the particular integral for the equation
y""' + 2y = 17e
2
x,
84
Ordinary Differential Equations 5.4 Linear System
we get
or
* 17 2x
y = D5 +2e
Since D
5
..\
5
= 2
5
, we have
17e2x e2x
y* = 25 +2 = 2"
0
This method also works for sin x, cos x, sinh x and others,
and this is because they are related to e>.x via sin() = ( ei
9

eie) and coshx =(ex+ ex)/2.
Higher order differential equations can conveniently be writ
ten as a system of differential equations. In fact, an nthorder
linear equation can always be written as a linear system of n
firstorder differential equations. A linear system of ODEs is
more suitable for mathematical analysis and numerical integra
tion.
5.4 Linear System
For a linear n order equation (5.15), it can be always written
as a linear system
= an1(X)Yn1 + ... + a1(x)y1 + ao(x)y + ¢(x),
(5.18)
which is a system for u = [y Y1 Y2 ... Ynt]T.
For a secondorder differential equation, we can always write
it in the following form
du
dx = f(u, v, x),
dv
dx = g(u, v, x).
(5.19)
If the independent variable x does not appear explicitly in f
and g, then the system is said to be autonomous. Such a system
85
5.5 SturmLiouville Equation Ordinary Differential Equations
has important properties. For simplicity and in keeping with
the convention, we use t = x and u = dufdt in our following
discussion. A general linear system of nth order can be written
as
(
~ ~ l ( au a12 atn l ( Ut l u2 a21 a22 ... a2n u2
.  . . . '
. . . .
. . . .
Un ant an2 ... ann Un
(5.20)
or
ii =Au. (5.21)
If u = vexp(At), then this becomes an eigenvalue problem for
matrix A,
(A Al)v = 0, (5.22)
which will have a nontrivial solution only if
det(A AI) = 0. (5.23)
5.5 SturmLiouville Equation
One of the commonly used secondorder ordinary differential
equation is the SturmLiouville equation in the interval x E
[a, b]
d dy
dx [p(x) dx] + q(x)y + Ar(x)y = 0,
(5.24)
with the boundary conditions
y(a) + ay'(a) = 0, y(b) + (3y'(b) = 0, (5.25)
where the known function p( x) is differentiable, and the known
functions q(x), r(x) are continuous. The para1neter A to be
determined can only take certain values An, called the eigen
values, if the problen1 has solutions. For the obvious reason,
this problem is called the SturmLiouville eigenvalue problem.
86
Ordinary Differential Equations 5.5 SturmLiouville Equation
For each eigenvalue An, there is a corresponding solution
1/J>..n, called eigenfunctions. The SturmLiouville theory states
that for two different eigenvalues Am =/= Am their eigenfunctions
are orthogonal. That is
• = 0.
(5.26)
or more generally
(x)l/1>.. (x)r(x)dx = Omn· (5.27)
It is possible to arrange the eigenvalues in an increasing order
At < A2 < .... < An < ... + oo. (5.28)
Sometimes, it is possible to transform a nonlinear equation
into a standard linear equation. For example, the Riccati equa
tion can be written in the generic form
y' = p(x) + q(x)y + r(x)y
2
, r(x) =/= 0. (5.29)
If r(x) = 0, then it reduces to a first order linear ODE. By
using the transform
or
we have
where
P(x) =
y(x) =
u'(x)
r(x)u(x)'
u(x) = e J r(x)y(x)dx,
u" P(x)u' + Q(x)u = 0,
r'(x) + r(x)q(x)
r(x)
87
Q(x) = r(x)p(x).
(5.30)
(5.31)
(5.32)
(5.33)
5.5 SturmLiouville Equation Ordinary Differential Equations
5.5.1 Bessel Equation
The wellknown Bessel equation
(5.34)
is in fact an eigenvalue problem as it can be written as
l/2
(xy')' + (x )y = 0.
X
(5.35)
Although v can be any real values, but we only focus on the
case when the values of v are integers. In order to solve this
equation, we assume that the solution can be written as a series
expansion in the form
00 00
y(x) = xs I: anxn =I: anxn+s,
ao =/= 0, (5.36)
n=O n=O
where s is a parameter to be determined. If a
0
= 0, we can
always change the value of s, so that the first term of an is not
zero. Thus, we assume in general ao =/= 0. This method is often
called the Frobenius method which is essentially an expansion
in terms of a Laurant series. Thus, we have
dy 00
 = L an(n + s)xn+s1'
dx n=O
(5.37)
(5.38)
Substituting these expression into the Bessel equation, we have
00 00
L(k + s)(k + s l)anxn+s + L(n + s)anxn+s
n=O n=O
00 00
+ L anxn+s+2 l/2 L anxn+s = 0.
(5.39)
n=O n=O
88
Ordinary Differential Equations 5.5 SturmLiouville Equation
Equating the coefficients of the same power xn(n = 0, 1, 2, ... ),
we can get some recurrence relationships. For n = 0, we have
Since a
0
=F 0, we thus have
s = ±v.
From n = 1 terms, we have
or
at(2s + 1) = 0,
1
a1 = 0, (s =F 
2
).
For the rest of terms n = 2, 3, 4, ... , we have
(5.40)
(5.41)
(5.42)
(5.43)
an(n + s)(n + s 1) + an(n + s) + an2 v
2
an = 0, (5.44)
or
an2 an2
(n + s)
2
 v
2
= n(n + 2v)'
(5.45)
Since we now know that a1 = 0, thus aa =as= a7 = ... = a1 =
0. All the even terms contain the factor a
0
, we finally have
(5.46)
where we have used n = 2k = 0, 2, 4, ... so that k = 0, 1, 2, 3, ....
The function J v
oo (l)n
J v =
'
(5.47)
is called the Bessel function of the order v. This is the Bessel
function of the first kind. It has many interesting properties:
(5.48)
89
5.5 SturmLiouville Equation Ordinary Differential Equations
00
I: J, = 1, Jv(x) = ( 1)"' J,(x); (5.49)
v=oo
and
fox uJo(u)du = xJ1 (x). (5.50)
There are other properties as well such as the orthogonality
l xJv(ox)Jv(f3x)dx = 0, (a :/= /3). (5.51)
5.5.2 Euler Buckling
As an example, let us look at the buckling of an Euler column
which is essentially an elastic rod with one pinjointed end and
the applied axial load P at the other end. The column has a
length of L. Its Young's modulus is E and its second moment
of area is I= J y
2
dA =canst (for a given geon1etry). Let u(x)
be the transverse displacement, the Euler bean1 theory gives
the following governing equation
or
2 p
a = EI'
which is an eigenvalue problem. Its general solution is
u = Asinax + Bcosax.
(5.52)
(5.53)
(5.54)
Applying the boundary conditions, we have at the fixed end
(at x = 0), B=O, (5.55)
and at the free end
u=O, (at x = L), Asin(aL) = 0. (5.56)
90
Ordinary Differential Equations 5.5 SturmLiouville Equation
Thus we have two kinds of solutions either A = 0 or sin( o.L) =
0. For A = 0, we have u(x) = 0 which is a trivial solution. So
the nontrivial solution requires that
sin( o.L) = 0, (5.57)
or
o.L = 0 (trivial), 1r, 27r, ... , n1r, ... (5.58)
Therefore, we have
P
 2E/ n27r2EJ
a  £2 '
(n=1,2,3, ... ). (5.59)
The solutions have fixed mode shapes (sine functions) at some
critical values (eigenvalues Pn)· The lowest eigenvalue is
7r2EJ
p* = y;'
which is the Euler buckling load for an elastic rod.
5.5.3 Nonlinear SecondOrder ODEs
(5.60)
For higherorder nonlinear ordinary differential equations, there
is no general solution technique. Even for relatively simple
secondorder ODEs, different equations will usually require dif
ferent n1ethods, and there is no guarantee that you can find the
solution. One of the best methods is the change of variables
so that the nonlinear equation can be transformed into a lin
ear ordinary differential equation or one that can be solved by
other methods. This can be beautifully demonstrated by the
solution process of finding the orbit of a satellite.
As the satellite orbits the Earth, the force is purely radial in
the polar coordinates, therefore, its total angular momentum L
is conserved. L = mr
2
!f/f = const, or r
2
iJ = Lfm = h = const.
The radial acceleration is ar = f riJ. Using Newton's second
law of motion and Newton's law of gravity, we have
m[d
2
r _ r(d8)
2
] = _ GMm
dt
2
dt r2 '
(5.61)
91
5.5 SturmLiouville Equation Ordinary Differential Equations
where M and m are the masses of the Earth and the satellite,
respectively. G is the universal constant of gravitation. Using
the conservation of angular momentum so that
we then have
L
h=,
m
(5.62)
d
2
r h
2
GM
dt2  r3 + ;r = 0, (5.63)
which is a nonlinear equation. By using the change of variables
u = lfr, the conservation of angular momentum becomes
(5.64)
which is equivalent to dt = d(J / ( hu
2
) and this can be used to
eliminate t. Then, we have
dr _ _ _
2
du _ h du
dt  u dt  d(J '
(5.65)
and
d
2
u = dr = h ~ u d(J = h2u2 ~ u
dt
2
dt d8
2
dt d8
2
.
(5.66)
Now the governing equation becomes
(5.67)
or
(5.68)
Since this is a secondorder linear ordinary differential equation,
it is straightforward to write down its solution
u = S + Acos8 + Bcos8 = S[l + ecos(8 + 'l/7)], (5.69)
where A and B are integration constants, which can be con
verted into the eccentricity e and the initial phase 'l/J. The final
solution is
1
r  ~    ~  ~
 S[l+ecos(B+'l/1)]'
(5.70)
92
Ordinary Differential Equations 5.5 SturmLiouville Equation
which corresponds to an ellipse where e is the eccentricity of
the orbit. If we set the polar coordinates in such a way that
'l/J = 0 (say, along the n1ajor axis) and one focus at the origin,
then the equation simply becomes
h2
r= '
GM[l + ecos8]
(5.71)
which is the orbit for satellites and planets.
93
Chapter 6
Recurrence Equations
6.1 Linear Difference Equations
Differential equations always concern the quantities that vary
continuously. However, some problems such as finance are con
cerned with quantities (say, interest rate) that are discrete and
do not vary continuously, and even the independent variables
such as time are not continuously counted or measured (in sec
onds or years). For this type of problem, we need the differ
ence equation or the recurrence equation as the counterpart in
differential equations. In fact, there many similarity between
difference equations and differential equations, especially the
linear ones.
A linear difference equation of Norder can generally be
written as
aoYn + atYn1 + a2Yn2 + ... + aNYnN = f(n), (6.1)
where ai(i = 0, ... , 1V) are coefficients which are not functions of
y. Yn [= y(n)] is the value of the variable y at n = 0, 1, 2, .... If
f(n) = 0, we say that the difference equation is homogeneous.
If all the coefficients ai are constant, the equation is called the
linear equation with constant coefficients (as the counterpart
in the differential equations). In this book, we only focus on
95
6.1 Linear Difference Equa..tions Recurrence Equations
the secondorder linear difference equation. Thus, we have
ayn + bYn1 + CYn2 = f(n),
(6.2)
which can also be written as
(6.3)
where g(n) = f(n 1). If f(n) = 0, we say the equation is
homogeneous. The most fan1ous difference equation is probably
the recurrence relation
Yn = Yn1 + Yn2,
(6.4)
for the Fibonacci sequence (0, 1, 1, 2, 3, 5, 8, ... ). The recurrence
equation is valid for n = 2, 3, ... and the initial conditions are
y(O) = 0, y(1) = 1.
Similar to the solution procedure of linear ordinary differ
ential equations, the general solution of a difference equation
Yn = Un + Pn where Un is the complementary solution to the
homogeneous equation
(6.5)
while Pn is any particular solution of (6.2).
In order to obtain un, we assume that Un = o:An (similar to
Yc = Ae>.x for differential equations). Substituting into (6.5),
we reach a characteristic equation
aA
2
+ bA + c = 0. (6.6)
It has two solutions A1 and A2 in general. Therefore, we have
un = AA} + BA2. (6.7)
0 Example 6.6: Find the solution of
Yn + 3Ynl + 2Yn2 = 0.
96
Recurrence Equations 6.1 Linear Difference Equations
This is a homogeneous equation, we assume Un = O'An, we have
or (A+ 1){A + 2) = 0.
Thus, we have A
1
= 1 and A
2
= 2. Therefore, tl1e general solution
can be written as Yn = A(1t + B(2t. 0
For given initial values, we can determine the constant in
the general solution so that an exact expression can be ob
tained.
0 Example 6. 7:
difference equation
The Fibonacci sequence is governed by the
Yn = Yn1 + Yn2,
with initial conditions Yo = 0, y
1
= 1. This is a homogeneous equa
tion. The characteristic equation is
A
2
 A 1 = 0,
whose solution is A=
1
±
2
v's. The general solution is therefore
In order to determine A and B, we first use Yo = 1, we get
O=A+B.
For n = 1, Y1 = 1 gives
1 = A( 1 +2 y5) + B( 1 2 y5 ).
Now we have A = 1/ v'5, B = 1/ .;5. The general solution becomes
= _1 ( 1 + v'5)n __ 1 ( 1  y5)n
Yn y'5 2 y'5 2 .
0
For finding a particular solution Pn, we can use the similar
technique used in ordinary differential equations. For f ( n) =
k = const, we try Pn =a. For f(n) = kn, we try Pn =a+ {3n
where a and {3 will be determined. For f(n) = k1n, we try
97
6.2 Chaos and Dynamical Systems Recurrence Equations
Pn = O:')'n. Other forms such as polynomials can be done in a
similar fashion.
0 Example 6.8: For the equation
Yn+t Yn 6Ynt = n,
its complementary equation is
Yn+l  Yn  6Yn1 = 0,
and the characteristic equation is
..\
2
 .X 6 = 0, or (.X+ 2)(..\ 3) = 0.
The general complementary solution can now be written as
Finding any pa.rticular solution requires that we t1y Pn = a+ bn. \Ve
now have
a+ b(n + 1) (a+ bn) 6[a + b(n 1)] = n,
or
6a + 7b 6bn = n.
As for any n, this equation is valid, therefore we have a = ;
6
, and
b = k· Finally, the general solution is
y(n) = A(2)n + B3n ~  ;
6
.
6.2 Chaos and Dynamical Systems
As you may have noticed that the above analysis is mainly
about the linear equations, what happens to the nonlinear
equations? The main problem with nonlinear equations is that
there is no general solution techniques available for most cases.
Even for the simplest case, the analysis is not easy. In addition,
the behaviour of nonlinear equations is very complex, even for
the simplest equations. Often, nonlinear equations may have
chaotic behaviour under appropriate conditions.
98
Recurrence Equations 6.2 Chaos and Dynamical Systems
0.8 0.8
0.6
~ u 1
c
>o
0.2
00
20 40 60 80 100 20 40 60 80 100
n n
Figure 6.1: Variations of Yn for v = 2 and v = 3.
6.2.1 Bifurcations and Chaos
In order to show this, we now briefly introduce the concept of
chaos by studying the following nonlinear difference equation
Yn+l = VYn(1 Yn), (6.8)
where v is a constant parameter. This is the wellstudied lo
gistic map, which is essentially an iteration process because all
the values Y2, Y3, ... can be determined for a given parameter v
and an initial condition Yt· This is one of the simplest dynanl
ical systems. This seemingly simple equation is in fact very
complicated. If you try to use the method to solve the linear
difference equations discussed earlier, it does not work.
For a given value v = 2, we can use a computer or a pocket
calculator to do these calculations. If the initial value Yt = 0
or Yt = 1, then the system seems to be trapped in the state
Yn = 0 (n=2,3, ... ). However, if we use a slight difference value
(say) Yt = 0.01, then we have
Yt = 0.01, Y2 = 0.0198, Y3 = 0.0388, Y4 = 0.0746,
Ys = 0.1381, Y6 = 0.2381, Y7 = 0.3628, YB = 0.4623,
y
9
= 0.4972, y
10
= 0.5000, y
11
= 0.5000, .... (6.9)
99
6.2 Chaos and Dynamical Systems Recurrence Equations
0.2
n n
Figure 6.2: Variations of Yn for v = 3.5 and v = 4.
Then, the values are trapped or attracted to a single value or
state Yoo = 0.5000. All the values are then plotted in a graph
as shown in Figure 6.1.
If we use a different parameter v = 3 and run the simula
tions again from the same initial value YI = 0.01, the results
are also plotted on the right in Figure 6.1. Now we have a dif
ference phenomenon. The final values do not settle to a single
value. Instead, they oscillate between two states or two final
values Yoo = 0.6770 and Yoo* = 0.6560. The iteration system
bifurcates into two states as the parameter v increases. If we
do the same simulations again using a different value v = 3.5
(shown in Figure 6.2), there are four final states. For v = 4,
every values seems difference, the system is chaotic and the
values looks like a randmn noise.
Following exactly the same process but using different val
ues of v ranging from v = 0.001 to v = 4.000, we can plot out
the number of states (after 1V = 500 iterations) and then we get
a bifurcation map shown in Figure 6.3. It gives a detailed map
about how the system behaves. From the chaotic map, we see
that for v < 1, the final state is zero (the system is attracted
to a stable state Yoo = 0). For 1 < v < 3, the system settles
(or attracts) to a single state. For 3 < v < 3.45, the system
bifurcates into two states. It seems that the system is attracted
100
Recurrence Equations 6.2 Chaos and Dynamical Systems
Figure 6.3: Bifurcation diagram and chaos .
....
0.2
•,···1 .,
·· .. . : ~
·=
.......
'• . .. ·.
0.75 ··· .. ··
lift:
i
:
0
3 3.2 3.4 3.6 3.8 4
O.J.4
3.5 3.6 3.7 3.8 3.9
v
Figure 6.4: Bifurcation and similarity.
by these discrete states. For this reason, the map is also called
the attractors of the dynamical system. The system becomes
chaotic for v ;::: v* where v* ~ 3.57 is the critical value.
It is a bit surprising for a seemingly determined system
VYn ( 1  Yn) + Yn+ 1 because you may try many times to simu
late the same system at the same initial value Yo = 0.01 (say)
and parameter v. Then, we should get the same set of values
Yt, Y2, .... You are right. So where is the chaos anyway? The
problem is that this system is very sensitive to the small vari
ation in the initial value y
0
. If there is a tiny different, say,
101
6.2 Chaos and Dynamical Systems Recurrence Equations
y
0
= 0.01 ± 0.000000001 or even 10
1000
difference, then the
set of values you get will be completely different, and there is
no way of predicting the final values (say, Ysoo or Yt,ooo,ooo).
Since there is always uncertainty in the real world, even the
computer simulations can only use a finite number of digits,
so the chaos is intrinsic for nonlinear dynamical systems. In
fact, there is a famous 'butterfly effect'. It says that the wing
flip of a butterfly in Africa can cause a tornado in America
or anywhere. Obviously, this is exaggerated too much, but it
does provide some vivid picture for the nature of chaos and
sensitivity to the initial uncertainty in chaotic systems.
If we study Figure 6.4 closely, we can see there is a sim
ilarity between the whole map and its certain part (enclosed
by a box) which is enlarged and plotted on the right in the
same figure. In addition, the ratio between the lengths of the
parameter intervals for two successive bifurcation approaches
the Feigenbaum constant 8F = 4.669 ... , which is universal for
the chaotic systen1s.
This selfsimilarity is one of the typical behaviours of chaotic
systems and it also occurs in other nonlinear systems such as
and
\ . 2
Yn+l = ASlll Yn,
which are plotted in Figure 6.5.
6.2.2 Dynamic Reconstruction
(6.10)
(6.11)
When a nonlinear system becomes chaotic, it seems that it is
very difficult to understand the behaviour. However, there may
be some regularity such as attractors and selfsimilarity as we
have seen earlier. In some case, it is even possible to reconstruct
the system itself.
Suppose we do not know the exact form of the dynami
cal system, but we do know it only depends on one parameter
102
Recurrence Equations 6.2 Chaos and Dynamical Systems
2.5
2
1.5
0.5 Yn+1=A sin2 (yn)
0
2 2.2 2.4 2.6 2.8 3
A.
Figure 6.5: Bifurcation for Yn+1 = A sin Yn and Yn+1 =
.Xsin
2
Yn·
and we can observe the states Y1 , Y2, ... , Yn (to a certain de
gree). From our observations, we can actually reconstruct the
system using the sorted data and plotting Yn versus Yn
1
. If
there are N observations, we have U = Y2, ... , YN as one set
and V = Yb Y2, ... , Yn1 as another set. We plot U versus V,
then the system can be dynamically reconstructed. For exam
ple, from the 100 data for the nonlinear system discussed in
the previous section, the constructed system is a parabola as
plotted in Figure 6.6. The parabola is essentially the original
function y(1 y). The mapping is then y ~    + y(1 y). With a
free parameter v and discrete time n, we obtain the dynamical
system
Yn = VYn1(1 Yn1),
or
Yn+1 = VYn(l  Yn)· (6.12)
We see that even a simple nonlinear equation in 1D can
show the rich complexity of dynan1ical behaviour. Now we
briefly look at a Lorenz system as a classical example.
6.2.3 Lorenz Attractor
The Lorenz attractor was first discovered by Edward Lorenz
when he studied the weather model in 1963. The Lorenz equa
103
6.2 Chaos and Dynamical Systems Recurrence Equations
\
'
\
\
0.2 0.4 0.6 0.8
Figure 6.6: Dynamic reconstruction of the nonlinear function.
Figure 6.7: Lorenz strange attractor.
tions are
x = f3x + yz,
il = ~ ( z  y),
z = ("Y x)y z, (6.13)
where·= ft, and /3, ~ ' "Y are parameters. Specifically, ~ is the
Prandtl number, and "Y is the Rayleigh number.
For certain ranges of parameters, for example f3 = 8/3, ~ =
10, "Y = 28, the system becomes chaotic. The system moves
around in a curious orbit in 3D as shown in Figure 6.7, and the
104
Recurrence Equations 6.3 Selfsimilarity and Fractals
orbit is bounded, but not periodic or convergent. This strange
characteristic gives the name 'strange' to the attractor, and
thus the Lorenz strange at tractor. The chaotic map looks like
a butterfly.
6.3 Selfsimilarity and Fractals
In the bifurcation map of chaotic systems, we have seen the
selfsimilarity. This selfsimilarity is quite universal in many
phenomena, and it is also linked with the concept of fractals. In
fact, selfsimilarity and fractals occur frequently in nature. Two
classical exa1nples are fern leaves and the pattern of lightening.
This observation suggests that even if the system is chaotic,
there is still some pattern or certain regularity in its behaviour,
and thus chaos is very different from the random noise, even
though they sometimes may look the same.
In geometry, we know that a point is zero dimension and a
line is one dimension. Shnilarly, the dimension for a plane is
two and the dimension for a volume is three. All these dimen
sions are integers (0, 1, 2, 3). But there are other possibilities in
nature that the dimension can be a fraction or a real number.
This is where the fractals arise. The basic idea of fractals can
be demonstrated using the generation of the Koch curve (also
called the Koch snowflake curve). The generator or seed of
this curve is the four straight line segments as shown in Figure
6.8. By starting with this generator (left) denoted as 8
1
, and
replacing each straight line segment by the generator itself, we
get the second generation curve (middle, 8
2
). If we continue
this procedure many times, we get the Koch curve. A snapshot
at generation 5 is shown on the right in Figure 6.8. The total
length of the seed 81 is Lt = 4/3 since it has four segments
and each segment has a length of 1/3, for the 82 curve, the
total length is 4
2
/3
2
. This is because there are 4 more seg
ments added, but each segment is only 1/3 of the previous one
in length. For the nth generation, we have the total length
105
6.3 Selfsimilarity and Fractals Recurrence Equations
Figure 6.8: Generation of the Koch curve.
( ~ )n with 4n segments in total.
The fractal dimension can be defined by covering the curve
using small balls (circles in a plane) and counting the number
of balls with radius r = f > 0. For a curve r in a compact
set of a metric space and for each f as the radius of the small
balls, the smallest number N (f) of balls to cover the curve r
will varies with f. The fractal dimension is defined by the limit
d =lim lnN(f).
e+0 In f
(6.14)
For the Koch curve, the minimum radius of the balls is f =
1/3n, and the total number is N(f) = 4n. Hence, we have the
fractal dimension of the Koch curve
. ln4n ln4
d= hm
1 3
= 
1 3
~ 1.2619.
E+0 n n n
(6.15)
Another famous example of fractals, which is related to the
iterative dynamical systems, is the Mandelbrot set z = z
2
+ c,
or
 2
Zn+l Zn + C, (6.16)
where z E C is in the complex plane, and c is a complex number
to be tested. If you start with zo = 0 (say), the iterations
continue until certain criteria are met. In this set, the stopping
criterion is simply lzl 2:: 2. That is to say, for each given c, the
magnitude of z during the iterations changes, it either stays
small ( < 2) or it will eventually surpass two (> 2). If lzl < 2,
we say that c belongs to the Mandelbrot set, otherwise, it is not
106
Recurrence Equations 6.3 Selfsimilarity and Fractals
Figure 6.9: Fractal nature of the l\1andelbrot set: z = z
2
+c.
part of the l\1andelbrot set. The iterations will go on until the
modulus of z reach 2, and the point cis marked on the complex
plane (x, iy) if it is not part of the Mandelbrot set. Then, we
change a different value of c, and follow the same iterative
process again. After many iterations across a region of the
complex plane, the results become the wellknown picture of the
1\!Iandelbrot set (shown in Figure 6.9). It is an usual practice
to mark the points with colours depending on the number of
iterations for each point to reach the modulus lzl = 2. This
simple iterative system can produce beautiful patterns. You
can view this system as a dynamical system, but it is a very
complicated system.
Vast literature exists on this subject, and it is still an active
research area.
107
Chapter 7
Vibration and
Harmonic Motion
After the studies of the complex numbers and ordinary dif
ferential equations, it is time to see how they are applied to
engineering problems. This chapter concerns the vibration and
harn1onic motion of mechanical systems.
7.1 Undamped Forced Oscillations
The sin1ple system with a spring attached with a mass m is
a good example of harmonic motion (see Figure 7.1). If the
spring stiffness constant is k, then the governing equation of
the oscillations is a secondorder ordinary differential equation
for undamped forced harn1onic motion, which can be written
as
y" + wfiy = f(t), (7.1)
where wfi = kjm, and f(t) is the a known function oft. In the
case of f(t) = acoswt, we have
y" + w ~ y = acoswt, (7.2)
where wo is the natural frequency of the system, and a is the
amplitude of external forcing.
109
7.1 Undamped Forced Oscillations Vibration and Harmonic A1otion
Figure 7.1: Undamped forced oscillations.
The general solution y( t) = Yc + Yp consists of a comple
mentary function Yc and a particular integral Yp· The comple
mentary function Yc satisfies the hon1ogeneous equation
y" = 0. (7.3)
Its general solution is
Yc(t) = A sin wot + B cos wot. (7.4)
For the particular integral Yp, we have to consider two different
cases w =I= wo and w = wo because for w = wo the standard
particular a sin wt + b cos wt does not work. It needs some mod
ifications. For w =I= wo, we assume that Yp = asinwt + bcoswt.
We thus obtain
Q'
Yp =
2 2
coswt. (7.5)
w
0
w
Therefore, the general solution
• Q'
y( t) = A sin w
0
t + B cos w
0
t +
2 2
cos wt. (7 .6)
w
0
w
If we further assume that the system is initially at rest when
the force starts to act, we have the initial conditions y(O) = 0
and y' (0) = 0. With these conditions, we have A = 0 and
B =  w
2
) in the general solution. We now have
Q'
y(t) =
2 2
(coswt cosw
0
t). (7.7)
w
0
w
110
Vibration and Harmonic !vlotion 7.1 Undamped Forced Oscillations
Using
we have
where
5 (l)=:(l) =2
0
5 10
time (t)
15 20
Figure 7.2: Variation of an1plitude at w = wo.
C D 2
. (C+D) . (CD)
cos  cos =  sm
2
s1n . . . . ; . . . . _ 
2
 ~ ,
( )
2a . ( w  wo )t . ( w + wo) t
y t =
2 2
sm
2
s1n
2 w
0
w
= A(t) sin (w +
2
wo)t = A(t) sinwt, (7.8)
A( )
_ 2a . (w wo)t
t  2 2 sm 2 .
w
0
w
(7.9)
As lw wol < lw + wol, we can see that the oscillator oscillates
with a major and fast frequency w = (w + wo)/2, while its
amplitude or envelope slow oscillates with a frequent 8w =
( w  wo) /2. This phenomenon is called beats.
For the special case of w = wo, the complementary function
is the same as before, but the particular solution should take
the following form
Yp = t(asinwt + bcoswt), (7.10)
111
7.2 Damped Forced Oscillations Vibration and Harmonic A1otion
which gives
Q .
Yp(t) = 
2
t sm wot.
wo
The general solution is therefore
(7.11)
y(t) = Asinw
0
t + Bcosw
0
t + ~
0
tsinw
0
t. (7.12)
Similarly, the initial solution y(O) = y' (0) = 0 in1plies that
A = B = 0. We now have
y(t) =
2
° tsinwot = A(t)sinwot,
wo
(7.13)
where A(t) = atj(2wo). As the amplitude A(t) increases with
time as shown in Figure 7.5, this phenomenon is called reso
nance, and the external forcing cause the oscillations grow out
of control when the forcing is acted at the natural frequency
wo of the system.
7.2 Damped Forced Oscillations
k
y ~
m I Ty(t)
Figure 7.3: Damped harmonic motion.
As all the real systems have a certain degree of friction,
thus damping should be included. An example of damping is
shown in Figure 7.3. With damping, the equation of forced
oscillations becomes
y"(t) + 2Xy'(t) + w5y(t) = acoswt, (7.14)
112
Vibration and Harmonic 1\lfotion 7.2 Damped Forced Oscillations
where A is the damping coefficient. In principle, one can try
to solve this equation using the standard method, but it may
become a little awkward as it involves complex numbers. In
fact, there is quite an elegant method using the complex vari
ables. In order to do this, we write the companion equation for
equation (7.14) with a different forcing term
rl'(t) + 2Ar,'(t) + w51J(t) = asinwt. (7.15)
Since eiwt = coswt + isinwt, we can multiply (7.15) by i, and
add it to (7.14), and we have
z"(t) + 2Az + w5z = aezwt, (7.16)
where z(t) = y(t) + i17(t). By solving this equation, we essen
tially solve both equations (7.14) and (7.15) at the same time
if we can separate the real and imaginary parts. The comple
mentary function corresponds to the transient part while the
particular function corresponds to the steady state. For the
transient part, the characteristic equation gives
(7.17)
or
Jl = A ± J A
2
 wfi. (7.18)
If A
2
w5, then Jl < 0. If A
2
< w5, then fl.= A+ iJwfi A
2
and < 0. In both cases < 0, thus the solution
Zc oc + 0. In engineering, it is conventional to define
a case of critical damping when { = Ajwo = 1. The quality
factor Q =
is also commonly used. We now have
fl.= wo({ ± FI). (7.19)
For = 0, we have fl. = iwo, which corresponds to the harmonic
oscillations without damping. For { = 1, Jl = wo, it is critical
damping as the imaginary term is zero. The amplitude de
creases exponentially at just the slowest possible manner with
out any oscillations. 1, we get Jl = w
0
{
113
7.2 Damped Forced Oscillations Vibration and Harmonic A1otion
10 15 20 25
time (t)
Figure 7.4: Critical 1), 0.1),
and overdamping = 1.5).
The real part corresponds to the exponential decrease of the
amplitude and the imaginary part corresponds to oscillations.
For this reason, it is called underdamped. Finally, > 1 leads
to 1L = wo( ± < 0. The imaginary part is zero (no
oscillation). As the an1plitude decreases much faster than that
at the critical damping, this case is thus called overdamped.
Figure 7.4 shows the characteristics of these three cases.
If time is long enough ( t >> 1), the transient part Yc will
becomes negligible as t increases. Therefore, we only need to
find the particular solution Zp·
If we try the particular solution in the form z = zoeiwt, we
have
z" + 2Xz' + wfi = P( iw )z, (7.20)
and
P(iw) = (iw)
2
+ 2X(iw) + w5 = (w5 w
2
) + 2Xwi, (7.21)
which is essentially the characteristic polynomial. The general
solution becomes
Q Q . t
z(t) = P(iw) ezwt = [(wfi w2) + 2i.Xw] ezw .
(7.22)
114
Vibration and Harmonic !vlotion 7.2 Damped Forced Oscillations
3
 A.:O.O
 0.25
2.5

0.5
0.75
2
·
1.0
s
c;1.5
ro
0
:2
0.5
1 4 5
Figure 7.5: Variations of amplitude with frequency w and
damping coefficient .X.
It is conventional to call H(iw) = 1/ P(iw) the transfer func
tion. We can always write the general solution z = Aei(wt+<l>),
where A= lzl is the modulus and¢ is the phase shift. There
fore, we have
z = Aei(wt+<l>)'
(7.23)
where
(7.24)
and
_
1
2Xw
¢=tan
2
2
.
w
0
w
(7.25)
As the amplitude of the forcing is a, the gain G(w) of the
oscillation is
A 1
G(w) = = ,
a J (w5  w2)2 + 4.X2w2
(7.26)
which is shown in Figure 7.5.
Finally, the solution of the original equation (7.14) is the
real part. That is
y(t) = Acos(wt + ¢). (7.27)
115
7.3 Normal Modes Vibration and Harmonic A1otion
Some special cases where w+ 0 and w+ oo are very interest
ing. For w << wo, the driving force is at very low frequency, we
have
That is
</>+ 0.
a
y(t) ~ , coswt.
wo
The system is in the same phase with the forcing.
(7.28)
(7.29)
If w >> wo, the forcing is at very high frequency. We have
</>+ 7r. (7.30)
The oscillator is completely out of phase with the forcing.
If w ~ wo, we have
and
a
A+ 2Awo'
y(t) =
2
:
0
sinwot.
(7.31)
(7.32)
At resonance frequency w; = w ~  2A
2
, the amplitude of the
oscillations increases dramatically.
7.3 Normal Modes
In the above harmonic oscillations, we know wo is the natural
frequency of the concerned system, which is relatively simple.
In general, there may be many natural frequencies or modes in
a system, and the natural frequencies are in fact determined
from the eigenvalue problem resulting from the system. Now
let us study a more complicated system with three mass blocks
attached in by two springs as shown in Figure 7.6. This system
can be thought of as a car attached to two caravans on a flat
road.
Let u 1, u2, u3 be the displacement of the three mass blocks
m
1
, m
2
, m
3
, respectively. Then, their accelerations will be ii.1 ,
116
Vibration and Harmonic 1\lfotion 7.3 Normal !v!odes
u2, ih where ii. = d
2
uj dt
2
• From the balance of forces and
Newton's law, we have
mtii.t =kt(u2ul), (7.33)
m2u2 = k2(u3 u2) kt(U2 Ut), (7.34)
ffi3U3 = k2(u3 u2). (7.35)
These equations can be written in a matrix form as
(I
0
0 ) ( Ut )
m2 0 ii.2
0
m3 u3
( kt
kt
~ : 2 ) ( : : ) = ( ~ }
+ ;1
kt + k2
k2
or
Mii+Ku = 0,
where uT = (ut, u2, u3). The mass matrix is
0
and the stiffness matrix is
(7.36)
(7.37)
(7.38)
Equation (7.37) is a secondorder ordinary differential equa
tion in terms of matrices. This homogeneous equation can be
solved by substituting Ui = Ui cos(wt) where Ui(i = 1, 2, 3) are
constants and w
2
can have several values which correspond to
the natural frequencies. Now we have
w[MUi cos(wt) + KUi cos(wt) = 0, (7.40)
117
7.3 Normal Modes Vibration and Harmonic A1otion
~ 1
~ 3
Figure 7.6: Harmonic vibrations.
where i = 1, 2, 3. Dividing both sides by cos(wt), we have
(7.41)
This is essentially an eigenvalue problem because any non
trivial solutions for ui require
(7.42)
Therefore, the eigenvalues of this equation give the natural fre
quencies.
0 Example 7.1: For the simplest case when m
1
= m
2
= m
3
= m
and k
1
= k
2
= k, we have
=0,
or
w
2
(k w
2
m)(3km w
2
m
2
) = 0.
This is a. cubic equation in terms of w
2
, and it has three solutions.
Therefore, the three natural frequencies are
w ~ = 0,
2 3k
w3 = :;;·
For w ~ = 0, we have (U
1
, U
2
: U
3
) = *(1, 1, 1), which is the rigid
body motion. For w
2
= k/m, the eigenvector is determined by
k
k
k
118
Vibration and Harmonic 1\lfotion 7.4 Small Amplitude Oscillations
which leads to U
2
= 0, and U
1
= U
3
. \Vritten in normalized form, it
becomes (U
1
, U
2
, U
3
) = ~ ( 1 , 0 , 1). This means that block 1 moves
in the opposite direction away from block 3, and block 2 remains
stationary. For wi = 3k/m, we have (U
1
, U2, U
3
) = "*(1, 2, 1).
That is to say, block 2 moves in the different direction from block 3
which is at the same pace with block 1. 0
7.4 Small Amplitude Oscillations
For a mechanically conservative system, its total energy E =
T + V is conserved, where Tis its total kinetic energy and V
is its total potential energy. The configuration of the mechan
ical system can be described by its general coordinates q =
( q1, q2, ... , qn). The general coordinates can be distance and an
gles. Thus, the velocities of the system will be q = 41, ti2, ... , tin
If we consider the system consists of many small particles or
even imaginary parts, then the total kinetic energy T is a func
tion of q and sometimes q, but the potential energy V is mainly
a function of q only. As we are only concerned with small am
plitude oscillations near equilibrium Vmin = V(O) = Vo, we can
always take q = 0 at the equilibrium so that we can expand V
in terms of q as a Taylor series
V(q) = Vmin +I: ~ ~ q; +I: I; K;iq;qi + ... ,
l l J
(7.43)
where the stiffness matrix is
K
.. _ ~ 8
2
Vo I
lJ .
2 8qi8qj qi=O,qj=O
(7.44)
Since potential energy is always relative to an arbitrary refer
ence point, we can thus take the potential energy at equilibrium
V min to be zero. In addition, the equilibrium or the minimum
value of V requires ~ = 0 at the equilibrium point qi = 0,
and the force Fi = g ~ shall be zero. This is correct because
the resultant force must be zero at equilibrium, otherwise, the
119
7.4 Small Amplitude Oscillations Vibration and Harmonic A1otion
systen1 will be driven away by the resultant force. The com
ponent of the resultant force along the general coordinate qi
should also be zero. Therefore, the total potential energy is
now simplified as
v = L L qiKijqj = qTKq,
j
which is a quadratic form.
(7.45)
For any small oscillation, the velocity is linear in terms of
Qi, and thus the corresponding kinetic energy is ~ m q l . The
total kinetic energy is the sum of all the components over all
particles or parts, forming a quadratic form. That is to say,
T
"""' """' . . . TM .
= L L ffiijqiqj = q q,
(7.46)
i j
where M = [mij] is the mass matrix.
For a conservative system, the total mechanical energy E =
T +Vis conserved, and thus timeindependent. So we have
d(T + V) = !!:_ [ . T M . TK ] = 0
dt dt q q + q q .
(7.47)
Since M and K are symmetric matrices, this above equation
becomes
Mq+Kq=O. (7.48)
This is a second order ordinary differential equation for matri
ces. Assuming the solution in the form qT = ( q
1
, q
2
, ... , qn) =
(Ut coswt, U2 coswt, ... , Un coswt) and substituting it into the
above equation, we have
(7.49)
which is an eigenvalue problem.
As an application, let us solve the same system of three
mass blocks discussed earlier as shown in Figure 7.6. The total
potential energy T is the sum of each mass block
120
Vibration and Harmonic 1\lfotion 7.4 Small Amplitude Oscillations
0
(7.50)
which can be written as a quadratic form
T
·TM·
=U U, (7.51)
0
(7.52)
We see that M is a symmetric matrix.
For a spring, the force is f = kx, thus the potential energy
stored in a spring is
(7.53)
Therefore, the total potential energy of the twospring system
is
(7.54)
Since interchange of u
1
and u
2
does not change V, it is thus
symmetric in terms of u1, u2 etc, which implies that Kij should
be symmetric as well.
The stiffness matrix K = [Kij] can be calculated using
(7.55)
For example,
(7.56)
and
121
7.4 Small Amplitude Oscillations Vibration and Harmonic A1otion
1 {) kt
= 2 {)ul [kt (u2 ut) + k2(u3 u2)] = 
2
.
(7.57)
Following the similar calculations, we have
(7.58)
which is exactly 1/2 multiplying the stiffness matrix we ob
tained earlier in equation (7.39). Thus, the equation for small
amplitude oscillation is
Mii+Ku= 0. (7.59)
For the special case of mt = m2 = m3 = m and kt = k2 = k,
its eigenvalues are determined by
=0, (7.60)
which is exactly the problem we solved in the previous section
(see example 7.1).
For a simple system such as a pendulum, equation (7.39) is
equivalent to the following simple formula for calculating the
natural frequency
w=
V"(q)
M(q)'
(7.61)
where V" > 0 because the potential energy at equilibrium is
minimum.
0 Example 7.2: A simple pendulum with a mass m is hanged
vertically from a ceiling with a distance L from the fixed point. Let
8 be the small angle from its equilibrium, then the kinetic energy is
T = !mv
2
= !m£
2
(0)
2
• The potential energy is
V = mgL(1 cosO).
122
Vibration and Harmonic 1\lfotion 7.4 Small Amplitude Oscillations
Therefore, the stiffness is K = !V"(8) = !mgLcos8le=O = mgL/2.
The equivalent mass is 1\tf ( 8) = !mL
2
• The governing equation be
comes
or
1 2 .. 1
2
mL 8 +
2
mgL8,
.. L
8+ 8 = 0.
g
The natural frequency for small oscillations is
The period of this pendulum is
T =
2
1r = 21r !I..
w y;
123
0
Chapter 8
Integral Transforms
The mathematical transform is a method of changing one kind
of functions and equations into another kind, often simpler or
solvable one. In general, the transform is essentially a math
ematical operator that produces a new function F( s) by inte
grating the product of an existing function f(t) and a kernel
function K(t, s) between suitable limits
F(s) = J K(t, s)f(t)dt. (8.1)
In the Laplace transform, the kernel is simply exp( st) and
integration limits are from 0 to oo. In the Fourier transform,
the kernel is exp(±ist) with a normalized factor 1/../2ff.
The Fourier transform maps the time domain of a time
dependent series such as a signal into a frequency domain,
which is common practice in signal processing. The Laplace
transform is a very powerful tool in solving differential equa
tions. In this chapter, we will focus on the three major trans
forms: Fourier, Laplace and Wavelet transforms. They are
commonly encountered in engineering and computational sci
ences.
125
8.1 Fourier Transform Integral Transforms
8.1 Fourier Transform
8.1.1 Fourier Series
For a function f ( t) on an interval t E [T, T], the Fourier series
is defined as
where
11T
ao = T T f(t)dt,
(8.3)
and
11T n1rt
an = T T f(t) cos( T )dt,
(8.4)
11T . n1rt
bn = f(T) s1n()dt, (n = 1, 2, ... ).
T T T
(8.5)
Here an and bn are the Fourier coefficients of f ( t) on [T, T].
The function f ( t) can be continuous or piecewise continuous
with a jump discontinuity. For a jump discontinuity at t =to,
if !'(to) and !'(to+) both exist, but !(to) f f(to+ ), then
the Fourier series converges to [!(to)+ !(to+ )]/2. The Fourier
series in general tends to converge slowly.
From the coefficients an and bn, one can easily see that bn =
0 for an even function f( t) = f(t). Similarly, ao = an = 0
for an odd function f (t) =  f ( t). In both cases, only one
side [0, T] of the integration can be used due to the symmetry.
Thus, for an even function f ( t), we have the Fourier cosine
series on [0, T]
(8.6)
For an odd function f ( t), we have the Fourier sine series
00
n1rt
f(t) = L sin( T ).
n=l
(8.7)
126
Integral Transforms 8.1 Fourier Transform
f
t
4 2 0 2 4
Figure 8.1: Triangular wave with a period of 2.
0 Example 8.1: The triangular wave is defined by f(t) = ltl for
t E 1] with a period of2 or f(t+2) = f(t). Using the coefficients
of the Fourier series, we have
11
1
110 11 1
ao =
2
ltldt =
2
[ ( t)dt + tdtJ = 2·
1 1 0
Since both It I and cos(n1rt) are even functions, we have for any n 1,
an = {
1
ltl cos(n7rt)dt = 21
1
t cos(n7rt)dt
}_1 0
= 2tsin(n1rt)l
1
 {
1
sin(n1rt)dt =
2
2
2
[cos(n1r) 1].
n1r
0
n1r }
0
n 7r
Because ltl sin(n1rt) is an odd function, we have
bn = /
1
ltl sin(n1rt)dt = 0.
1
Hence, the Fourier series for the triangular wave can be written as
1 cos(n1r) 1
f(t) = 2 + 2 L n
2
1r
2
cos(n1rt).
n=O
0
The nterm of the Fourier series, that is an cos(n1rtjT) +
bn sin(n1rtjT), is called the nth harmonic. The energy of the
n harmonic is defined by = + The sequence of An
forms the energy or power spectrum of the Fourier series. The
energy spectrum of the triangular wave is shown in Figure 8.2.
127
8.1 Fourier Transform Integral Transforms
0.2
0.15
i;;
! 0.1
0.05
Figure 8.2: Energy spectrum of the triangular wave.
8.1.2 Fourier Integral
For the Fourier coefficients of a function defined on the whole
real axis [oo, oo], we can take the limits
and
a(wn) = /_: f(t) cos(wnt)dt,
n1f'
Wn=T'
as the limits T + oo and Wn + 0. We have ao + 0 if
(8.8)
(8.9)
In this case, the Fourier series becomes the Fourier integral
f(t) = fooo [a(w) cos(wt) + b(w) sin(wt)dw, (8.10)
where
1 !00
a(w) = ; oo f(t) cos(wt)dt,
(8.11)
1 !00
b(w) = f(t) sin(wt)dt.
1f' oo
(8.12)
128
Integral Transforms 8.1 Fourier Transform
Following the similar discussion for even and odd functions, we
know that even functions lead to Fourier cosine integrals and
odd functions lead to Fourier sine integrals.
8.1.3 Fourier Transform
The Fourier transform .1"[/ ( t)] of f ( t) is defined as
F(w) = .1"[/(t)] = . !:: Joo f(t)eiwtdt, (8.13)
V 27r co
and the inverse Fourier transform can be written as
f(t) = .rl [F(w)] = . !:: Joe F(w)eiwtliv, (8.14)
V 27r oo
where exp[iwt] = cos(wt) + i sin(wt). The Fourier transform
is a linear operator, and it has most of the properties of the
differential operator and the integral operator. Hence, it is
straightforward to prove that it has the following properties:
F[f(t) + g(t)] = .1"[/(t)] + .F(g(t)], (8.15)
F[of(t)] = o.1"[f(t)], (8.16)
.1"[( it)n f(t)] =
dnF(w)
(8.17)
dwn '
and
.1"[/(n)(t)] = (iwt F(w), (8.18)
if f(t + ±oo) = f'(t + ±oo) = ... = f(nl)(t + ±oo) + 0.
The transform can have different variations such as the Fourier
sine transforn1 and the Fourier cosine transform. The Fourier
transforms of some common functions are listed in the follow
ing table 8.1.
D Example 8.2: For the triangle function f(t) = 1ltl for (It I ~ 1)
and f ( t) = 0 for It I > 1, its Fourier transform is
\1'21rF(w) = /_: (1 ltl)eiwtdt
129
8.1 Fourier Ti·ansform Integral Transforms
0 1
= f {1 + t)eiwtdt + r (1 t)eiwtdt.
t lo
Integrating by parts and using
. 2( /2) 1  cosw
Sill W =
2
,
we have
rn=F( ) = iw+ 1 eiw _ etW + iw 1 = 2{1 cosw)
v £.1f w 2 2 2 .
w w w
Hence, we have
1 sin
2
(w/2) 1 .
2
F(w) = .j2; (w/
2
)2 = .j2;smc (w/2).
Table 8.1: Fourier Transforms
f(t)
f(t to)
j(t)eiw0 t
8(t)
1
sign(t)
eoltl
e(at)2 (a > 0)
f(at)
1
a2+t2
cos(wot)
sin(wo)
sinxax (a > 0)
F(w) = F[f(t)]
F(w)etwto
F(w wo)
1/v"Fi
v"Fi8(w)
.1.
tw
2o
a2+w2
w2
1 7:'2"
""'7'f!r"e 4a
v:.::o
! F ( ~ )
@ealwl
vt· a
J1[8(w wo) + 8(w + wo)]
i ;[8(w + wo) 8(w wo)]
~ ' (lwl < a); 0, (lwl > a)
0
The most useful Fourier transform for engineering and com
putational science is probably the discrete form, especially in
digital signal processing. The discrete Fourier transform (DFT),
130
Integral Transforms 8.2 Laplace Transforms
for a periodic discrete function or signal x( n) with a period N,
is defined by
N1
X[k] = L
(8.19)
n=O
and the inverse transform, also called the signal reconstruction,
is defined by
1
N1
x[n] = L
N k=O
(8.20)
A periodic signal x( n + N) = x( n) has a periodic spectrum
X[k + N] = X[k]. The discrete Fourier transform consists
of N multiplications and N 1 additions for each X[k], thus
for N values of k, the computational complexity is of O(N
2
).
However, if N =2m (mE JV), many of the DFT calculations
are not necessary. In fact, by rearranging the formula, one can
get the complexity of O(N log
2
N). This type of algorithms
is called Fast Fourier Transform (FFT). Vast literature exists
on the signal processing such as FFT, filter design and signal
reconstruction.
8.2 Laplace Transforms
The Laplace transform .C [f ( t)] of a function f ( t) is defined as
F(8) = .C[f(t)] =loco f(t)estdt, (8.21)
where 8 > 0. The inverse Laplace transform .c
1
(F(8)] is f(t)
or f(t) = .c
1
[F(8)]. The Laplace transforms of most simple
functions can be obtained by direct integration. For simple
functions t and eo:t, we have
lo
co loco 1 [ t l
00
1
.C(t] = testdt = estdt + est = .
o o 8 8 o 8
2
.C(eo:t] = {oo eo:testdt = [1e<so:)tl oo = _I __
lo 8 a o 8 a
131
8.2 Laplace Transforms Integral Transforms
Conversely .c
1
[
1
] = t .c
1
[
1
] = eat. For the Dirac
' ST ' so
<5'function, we have its Laplace transform
.C[d'(t)] = rX) d'(t)estdt = estl = 1. (8.22)
lo t=O
However, the inverse of a Laplace transfonn is usually more
complicated. It often involves the partial fractions of polynomi
als and usage of different rules of Laplace transforn1s. From the
basic definition, it is straightforward to prove that the Laplace
transform has the following properties:
.C[af(t) + fjg(t)] = a.C[f(t)] + /3.C[g(t)],
.C[e
0
t f(t)] = F(s a), s >a,
.C[f(t a)] = eos .C[f(t)],
.C[f' (t)] = s.C[f(t)]  f(O),
.C[ ft f( r)dr] = .!..c[j],
lo s
(8.23)
(8.24)
(8.25)
(8.26)
(8.27)
The Laplace transform pairs of common functions are listed
below in table 8.2.
0 Example 8.9: In order to obtain the Laplace transform of
f(t) = coswt, we shall first write
1 . t . t
f(t) = coswt = '2(e
1
w + ezw ).
Then, we have
= ~ [ [
00
e<s+iw)tdt + foe esiw)tdt)
2 Jo lo
1 1 1 s
= 2[s iw + s + iw] = s2 +w2
0
132
Integral Transforms 8.2 Laplace Transforms
Table 8.2: Laplace Transform
Function f(t)
1
8(t)
tn,n > 0
cos( at)
sin( at)
eat
t1/2
t1/2
tn f(t)
cos( at+ !3)
sinh(ot)
cosh( at)
a2
_1_e:rr
..;:;;
sinoVt
(a> 0)
J3 (eat _ ePt) ( 0 =f= /3)
Laplace Transform F( s)

s
1
n!
sn+T
s
s:l+a:l
a
s:l+a:l
1
sa
1 ( 11" ) 1/2
'2'S3'
(
1)ndnF(s)
dsn
s
s2+a2
a
s1a:l
s
s2a2
leavs
s
Both Fourier and Laplace transforms follow the convolution
theorem. For two functions f and g, their convolution f * g
obeys
(8.28)
and their Laplace transforms follow
.C[f(t) * g(t)] = F(s)G(s), (8.29)
C
1
[F(s)G(s)] = fo' f(t a)g(a)do:. (8.30)
133
8.3 lVa.velet Integral Transforms
The Fourier transform has the similar properties
f(t) * g(t) = /_: f(t)g(t u)du,
F[j(t) * g(t)] = F(w)G(w).
8.3 Wavelet
(8.31)
(8.32)
The Fourier transform is an ideal tool for studying the sta
tionary time signal whose properties are statistically invariant
over time. In the Fourier transform, the stationary signal is
decomposed into linear combinations of sine and cosine waves
. cos(nt), (n = 1, 2, ... ).
(8.33)
The Fourier transform is very useful to analyse stationary
signals where a stationary signal means that the frequencies of
the signal do not change with time. For nonstationary sig
nals whose frequencies f = w vary with time (see Figure
8.3), the Fourier transform does not work well. In addition,
in the Fourier transform there is a tradeoff between frequency
resolution and time resolution,
1
2
, (8.34)
which is similar to the Heisenberg uncertainty principle for spa
tial and velocity intervals. The wavelet transform is an alter
native approach to the Fourier transform to overcome the res
olution problen1 using the Ivlother wavelet 1/J or prototype for
generating the other windows functions, and all the used win
dows are in the form of either dilated/compressed or shifted. As
a result, the wavelet transform is very powerful in dealing with
nonstationary signals because the Fourier transforn1 is not
suitable for such signals. In the wavelet transform, a transient
signal is decomposed into elementary components of wavelets
134
Integral Transforms
> 1
...
as
s 0
=
as
u;
1
0
2 4 6 8
f:ro/21t varies with time (t)
8.3 Wavelet
10
10
Figure 8.3: A stationary signal with two fixed frequencies
(ft = 0.5 Hz and /2 =2Hz) and a nonstationary signal whose
frequency varies with time.
or wavelet packets. There are three major type of wavelets:
Grossmannl\1orlet wavelets, Daubechies wavelets and Gabor
Ivlalvar wavelets.
Wavelets are defined as a realvalued function 'ljJ(t) (t E
'R) in terms of the generator wavelet or n1other wavelet. The
function 'l/J is both well localized, decreasing rapidly as t .. oo
and oscillating in a wavery manner. To generate other wavelets,
'l/J( a, (3, t) is used by translating in time and change of scales.
GrossmannIvlorlet wavelets are of the form
1 t(3
'l/1(), Q > 0,
Q Q
a,b E 'R, (8.35)
where 'l/J a generator wavelet. The Daubechies wavelets have
the form
m,nEZ. (8.36)
135
8.3 lVa.velet Integral Transforms
The GaborlVIalvar wavelets are in the form
1
w(t m) cos[1r(n + '2 )(t m)), mE Z,n EN. (8.37)
The continuous wavelet transform can be defined by
1 J tT
\II J(T, s) = JJST f(t) · 1/J(
8
)dt,
(8.38)
where T is the translation of the location of the window and s
is the scale where s = 1 is for the most compressed wavelet.
Wavelet analysis has vast literature and it is still an active
research area in signal processing. Readers can search the latest
research journals to follow the latest developments.
136
Chapter 9
Partial Differential
Equations
Partial differential equations are much more complicated com
pared with the ordinary differential equations. There is no
universal solution technique for nonlinear equations, even the
numerical simulations are usually not straightforward. Thus,
we will mainly focus on the linear partial differential equations
and the equations of special interests in engineering and com
putational sciences. A partial differential equation (PDE) is a
relationship containing one or more partial derivatives. Simi
lar to the ordinary differential equation, the highest nth partial
derivative is referred to as the order n of the partial differential
equation. The general form of a partial differential equation
can be written as
8u 8u 8
2
u 8
2
u 8
2
u
1/J(x, y, ... , 8x' ay' 8x2' ay2' 8xay' ... ) = O. (
9
·
1
)
where u is the dependent variable and x, y, ... are the indepen
dent variables.
A simple example of partial differential equations is the lin
ear first order partial differential equation, which can be written
as
8u 8u
a(x, y) 8x + b(x, y) ay = f(x, y).
(9.2)
137
9.1 First Order PDE Partial Differential Equations
for two independent variables and one dependent variable u.
If the right hand side is zero or simply f(x, y) = 0, then the
equation is said to be homogeneous. The equation is said to be
linear if a, b and f are functions of x, y only, not u itself.
For simplicity in notations in the studies of partial differ
ential equations, compact subscript forms are often used in the
literature. They are
{)
2
u {]2u
Uyy = {)y2 ' Uxy = 8x8y'
(9.3)
and thus we can write (9.2) as
aux +buy= f. (9.4)
In the rest of the chapters in this book, we will use these nota
tions whenever no confusion occurs.
9.1 First Order PDE
The first order partial differential equation of linear type can
be written as
a(x, y)ux + b(x, y)uy = f(x, y), (9.5)
which can be solved using the method of characteristics
(9.6)
This is equivalent to the following equation in terms of param
eter s
(9.7)
which essentially forms a system of firstorder ordinary differ
ential equations.
138
Partial Differential Equations 9.2 Classification
The simplest example of a first order linear partial differ
ential equation is the first order hyperbolic equation
Ut +cux = 0, (9.8)
where cis a constant. It has a general solution of
u = ,P(x ct), (9.9)
which is a travelling wave along xaxis with a constant speed c.
If the initial shape is u(x,O) = 1/J(x), then u(x, t) = 1/J(xct) at
timet, therefore the shape of the wave does not change though
its position is constantly changing.
9.2 Classification
A linear secondorder partial differential equation can be writ
ten in the generic form in terms of two independent variables
x andy,
auxx + buxy + CUyy + gux + huy + ku = /,
(9.10)
where a, b, c, g, h, k and f are functions of x and y only. If
f ( x, y, u) is also a function of u, then we say that this equation
is quasilinear.
If ~ = b
2
 4ac < 0, the equation is elliptic. One famous
example is the Laplace equation Uxx + Uyy = 0.
I f ~ > 0, it is hyperbolic. One example is the wave equation
Utt = c2Uxx·
If ~ = 0, it is parabolic. Diffusion and heat conduction
equations are of the parabolic type Ut = KUxx·
9.3 Classic PDEs
Iviany physical processes in engineering are governed by three
classic partial differential equations so they are widely used in
a vast range of applications.
139
9.3 Classic PDEs Partial Differential Equations
Laplace's and Poisson's Equation
In heat transfer problems, the steady state of heat conduction
with a source is governed by the Poison equation
(x,y)En, (9.11)
or
Uxx + Uyy = q(x, y, t), (9.12)
for two independent variables x and y. Here k is the thermal
diffusivity and f(x, y, t) is the heat source. If there is no heat
source (q = 0), this becomes the Laplace equation. The so
lution or a function is said to be harmonic if it satisfies the
Laplace equation.
Heat Conduction Equation
Timedependent problems, such as diffusion and transient heat
conduction, are governed by parabolic equations. The heat
conduction equation
Ut = kuxx, (9.13)
is a famous example. For diffusion problem, k is replaced by
the diffusion coefficient D.
Wave Equation
The vibrations of strings and travelling sound waves are gov
erned by the hyperbolic wave equation. The 1D wave equation
in its simplest form is
2
Utt = C Uxx, (9.14)
where cis the speed of the wave.
There are other equations that occur frequently in mathe
matical physics, engineering and computational sciences. We
will give a brief description of some of these equations in later
chapters.
140
Chapter 10
Techniques for Solving
PDEs
Different types of equations usually require different solution
techniques. However, there are some methods that work for
most of the linearly partial differential equations with appro
priate boundary conditions on a regular domain. These meth
ods include separation of variables, series expansions, similarity
solutions, hybrid methods, and integral transform methods.
10.1 Separation of Variables
The separation of variables attempts a solution of the forn1
u = X(x)Y(y)T(t), (10.1)
where X(x), Y(y), T(t) are functions of x, y, t, respectively. In
order to determine these functions, they have to satisfy the
partial differential equation and the required boundary condi
tions. As a result, the partial differential equation is usually
transformed into two or three ordinary differential equations
(ODEs), and these ordinary differential equations often appear
as eigenvalue problems. The final solution is then obtained by
solving these ODEs. As a classic example, we now try to solve
141
10.1 Separation of Variables Techniques for Solving PDEs
the 1D heat conduction equation in the domain x E [0, L] and
t>O
(10.2)
with the initial value and boundary conditions
u(O, t) = u(L, t) = 0, u(x, 0) = 1/J. (10.3)
Letting u(x, t) = X(x)T(t), we have
X"(x) T'(t)
y = kT.
(10.4)
As the left hand side depends only on x and the right hand
side only depends on t, therefore, both sides must be equal
to the same constant, and the constant can taken to be as
A
2
. The negative sign is just for convenience because we
will see below that the finiteness of the solution T(t) requires
that eigenvalues A
2
> 0 or A are real. Hence, we now get two
ordinary differential equations
X"(x) + A
2
X(x) = 0, T'(t) + kA
2
T(t) = 0, (10.5)
where A is the eigenvalue. The solution for T(t) is
(10.6)
The solution for X(x) is in a generic form
X(t) =a cos Ax+ ,BsinAx. (10. 7)
From the boundary condition u(O, t) = 0, we have a = 0. From
u(L, t) = 0, we have
sinAL= 0, (10.8)
which requires that AL = n1r. Please note that n f 0 be
cause the solution is trivial if n = 0. Therefore, A cannot be
continuous, and it only takes an infinite number of discrete val
ues, called eigenvalues. Each eigenvalue A = An = n1r / L, ( n =
142
Techniques for Solving PDEs 10.2 Transform }.!fethods
1, 2, ... )has a corresponding eigenfunction Xn = sin(.Xnx). Sub
stituting into the solution for T(t), we have
(10.9)
By superimposing Un = XnTn and expanding the initial condi
tion into a Fourier series so as to determine the coefficients, we
have
00
u(x, t) = L:: an sin( )e< nz >
2
kt,
n=l
2 [L .
an = L Jo 1/;(x) s1n( y
(10.10)
10.2 Transform Methods
Laplace Transform
The basic idea of the integral transform method is to reduce
the number of the independent variables. For the 1D time
dependent heat conduction, it transforms the partial differen
tial equation into an ordinary differential equation. By solving
the ordinary differential equation, the solution to the original
problem is obtained by inverting back from the Laplace trans
form. As an example, we now solve the 1D heat conduction in
semiinfinite interval [0, oo ),
(10.11)
with the boundary conditions
T(x, 0) = 0, T(O, t) = T
0
. (10.12)
Let T(x, s) = f
0
00
T(x, t)estdt be the Laplace transform of
T(x, t), the equation then becon1es
(10.13)
143
10.2 TI·ansform 1\tlethods Techniques for Solving PDEs
and the boundary condition at x = 0 becomes f'x=O = To/ s.
The general solution to the ordinary differential equation can
be written as
 IBx · 1'Fx
T=Aev"f· +BeV"f·.
The finiteness of the solution as x + oo requires that B = 0,
and the boundary conditions lead to
T
 To  lEx
= e V"k'.
s
By the inversion of the Laplace transform, we have
X
T = Toerfc( . /U),
2vkt
where erfc( x) is the complementary error function.
Fourier Transform
Fourier transform works in the similar manner as the Laplace
transform. The famous example is the classical wave equation
2
Utt = V Uxx'
(10.14)
with the initial conditions u(x, 0) = 1/;(x) = exp[(x  a)
2
],
and Ut(x, 0) = 0. Let u(w, t) = Jk: f ~ o o u(x, t)eiwxdx be the
Fourier transform of u(x, t). This transforms the PDE problem
into an ODE
d
2
u 2 2
dt2 = v w u,
(10.15)
with
__ ,.T.( ) du(w, 0) _
0
UtpW, dt .
(10.16)
The general solution in tenns of the parameter w is
u(w, t) = ijj(w) cos(vwt).
By using the inverse Fourier transform, we finally have
1 !00  .
u(x, t) = . !7C 1/;(w) cos(vwt)ezwxdw
V 27r oo
144
Techniques for Solving PDEs 10.3 Similarity Solution
1
=
2
[1P(x + vt) + 'l/1(x vt)]
_ 1 [ (xa+vt)2 + (xavt)2]

2
e e , (10.17)
which implies two travelling waves: one travels along the xaxis
and the other along the negative xaxis direction.
10.3 Similarity Solution
The essence of similarity solution is to use the socalled sinli
larity variable = xjtf3 so as to reduce the partial differential
equation to an ordinary differential equation. For example, the
diffusion equation
{10.18)
can be solved by using the similarity method by defining a
similar variable
or
X
(=
#"
{10.19)
(10.20)
In general, we can assume that the solution to the equation has
the form
U = {Kt)
0
j[
(10.21)
By substituting it into the diffusion equation, the coefficients
a and f3 can be detern1ined. For most applications, one can
assume a= 0 so that u = f(() (see the following example for
details). In this case, we have
{10.22)
or
f" = (f',B(Kt)
2
/3l,
(10.23)
145
10.3 Similarity Solution Techniques for Solving PDEs
where !' = df I d(. In deriving this equation, we have used the
chain rules of differentiations fx = *: and £ = & so that
au ={3kf'(()x =
8t ,
(10.24)
and
(10.25)
Since the original equation does not have timedependent
terms explicitly, this means that all the exponents for any t
terms must be zero. Therefore, we have 2{3 = 1, or {3 = ! .
Now, the diffusion equation becomes
!"(() =
(10.26)
Using (ln !')' = !"If' and integrating the above equation once,
we get
or (10.27)
Integrating it again and using the (
2
= 4{
2
, we obtain
·2 X
u =A et; = Cerf( . .1'7'::7) + D,
t;o
(10.28)
where C and D are constants that can be determined from
appropriate boundary conditions. For the same problem as
(10.12), the boundary condition as x+ oo implies that C+D =
0, while u(O, t) =To means that D = C =To. Therefore, we
finally have
X X
u = T
0
[1 erf( .;;w)] = T
0
erfc( .;;w)· (10.29)
0 Example 10.1: For the similarity solution of the diffusion
equation Ut = KUxx' we assume u = (Kt)o: !(,.})
13
• Now we want to
know why a = 0. Since
146
Techniques for Solving PDEs 10.4 Travelling lVave Solution
the diffusion equation becomes
The requirement that no explicit (time t) appears in the equation
leads to
Q' = 0, Q'  2,8 + 1 = 0.
Thus,
Q' = 0,
You may think that why not divide both sides of the above
equation by ( Kt )o:, then we do not impose any requirement on
a, thus f3 = ~  This nonzero a indeed appears in some
nonlinear diffusions equations where the diffusion coefficient is
not a constant and K{ u) may depend on u itself.
10.4 Travelling Wave Solution
The travelling wave technique can be demonstrated using the
famous Kortewegde Vries (KdV) equation
(10.30)
or
(10.31)
which is a thirdorder nonlinear partial differential equation.
The interesting feature of this equation is that it has a solitary
wave solution or soliton. The soliton phenomenon was first
observed by John Russell in 1834 when he travelled along the
Union canal in Scotland. Nowadays, telecommunications use
solitons to carry signals in optical fibres.
Now we seek the travelling wave solution in the form
u = <f>(x vt), (10.32)
147
10.5 Green's Function Techniques for Solving PDEs
where v is the speed of the travelling wave. By substituting
into the KdV equation, we have
¢"' + 6c/>c/>
1
 v¢ = 0. (10.33)
Using ( ¢
2
/2)' = c/>c/>
1
and integrating the above equation once,
we have
¢" + 3¢
2
 vel> = A, (10.34)
where A is an integration constant. The requirement of c/>, cl>',
¢" + 0 at far field x + ±oo leads to A = 0. Let '1/J = cl>', we
get
d'lj; 2
'1/J d¢ + 3¢  v¢ = 0.
(10.35)
Integrating with respect to¢, we get
= ¢3 + (10.36)
Integrating it again and substituting back to u, we have the
travelling wave solution
v ]
u =
2
sech T ( x  vt  <5) , (10.37)
where <5 is a constant and v /2 is the amplitude of the wave. We
can see that the speed of the wave depends on the amplitude or
height of the wave. That is to say, big waves travel faster than
smaller waves. For linear wave equations, waves can travel in
both directions, but here it is only possible for the soliton to
travel in one direction, that is along xaxis direction in this
scenario.
10.5 Green's Function
The method of Green's function is very powerful in solving
elliptic equations. A Green's function inside the domain n is
defined as
(10.38)
148
Techniques for Solving PDEs 10.6 Hybrid J.1etbod
where 8(x) is the Dirac delta function. It usually requires that
G = 0 on the boundary surface r. Generally speaking, G =
G(x, y, z; 17, ().
For the hyperbolic equations such as the wave equation, we
can define the Green's function as
Gtt  c
2
= 8(x 17)8(z ()8(t r). (10.39)
The fundamental Green's function for this case is
1 r
= 
4
2
8[(t r) ].
1rc r c
(10.40)
The Green's function method is very complicated, but it can
be very neat in obtaining solutions. Readers can refer to the
literature listed at the end of the book.
10.6 Hybrid Method
Some differential equations can be solved using one of the meth
ods described above, but often a single method simply does
not work. In this case, a hybrid method that combines several
methods is needed. For example, the Crank's diffusion problem
in an infinite cylinder is governed by the following equation:
8u 18 8u
{)t =; 8r [rll: 8r],
(10.41)
where ll: is the diffusion coefficient, and r is the distance in the
polar coordinates. u(r, t) can be concentration or any other
quantity. Now we want to solve this equation with the following
boundary conditions:
u(r, t) = 0, r =a, (10.42)
and
u(r, 0) = 1/J(r), r E (O,a). (10.43)
First, we use the separation of variables, we have
(10.44)
149
10.6 Hybrid Method Techniques for Solving PDEs
then we have the equation for v( r)
1
v" + v' + A
2
v = 0,
r
(10.45)
where v' = dv / dr and A is a parameter to be determined. This
equation is essentially the Bessel equation of the zeroorder
v = 0. Thus, we can now assume that the general solution for
u(r, t) has the following form
00
u(r, t) = L DiJo(Air)e\rKt,
i=l
(10.46)
where Di are undetermined coefficients. The boundary condi
tion [u(r =a) = 0] requires that
(10.47)
which means that paratneter Ai are the roots of the Bessel func
tion J
0
( Aia) = 0. The initial condition gives
00
1/J(r) = L DiJo(Air).
i=l
Using the basic properties of the Bessel functions
and
loa rJo(Air)dr = ;i Jt(aAi),
the general solution can be written as
2 ~ Jo(rAi)   \ ~ K t Ina
u(r, t) =
2
L J
2
( A·) e • r'l/J(r)Jo(rAi)dr.
a i=l 1 a z o
150
(10.48)
(10.49)
(10.50)
(10.51)
(10.52)
Techniques for Solving PDEs 10.6 Hybrid P.1ethod
The solution procedure shows that it requires a combination of
separation of variables, Bessel functions, and power series.
0 Example 10.2: If we now want to solve the same diffusion
equation in the cylindrical coordinates with slight different boundary
conditions:
u(r =a, t) = uo = const, u(r, 0) = ¢(r),
we have to make a transformation u = w + uo. Both u and w satisfy
the same diffusion equation, but now the boundary conditions for w
become
w(r =a, t) = 0, w(r, 0) = ¢(r)  uo = 1/J(r),
which is the problem we have just solved. By substituting u = w + uo
and 1/J(r) = ¢(r)  uo and using the properties of Bessel functions,
we finally obtain
where hi = foa r¢(r)Jo(r>..i)dr. For a very special case when </>(r) =
0, we have
There are other important methods for solving partial dif
ferential equations. These include series methods, asymptotic
methods, approximate methods, perturbation n1ethods and nat
urally the numerical methods.
151
Chapter 11
Integral Equations
The calculus of variations is important in many optin1ization
problems and computational sciences, especially the formula
tion of the finite element methods. On the other hand, integral
equations are a different type of equation and they frequently
occur in applied mathematics and natural sciences. In this
chapter, we will briefly touch these topics.
11.1 Calculus of Variations
The main aim of the calculus of variations is to find a function
that makes the integral stationary, making the value of the
integral a local maximum or minin1um. For example, in me
chanics we may want to find the shape y( x) of a rope or chain
when suspended under its own weight from two fixed points.
In this case, the calculus of variations provides a method for
finding the function y(x) so that the curve y(x) minimizes the
gravitational potential energy of the hanging rope system.
11.1.1 Curvature
Before we proceed to the calculus of variations, let us first dis
cuss an important concept, namely the curvature of a curve.
In general, a curve y(x) can be described in a parametric form
153
11.1 Calculus of Variations Integral Equations
in terms of a vector r( s) with a parameter s which is the arc
length along the curve measured from a fixed point. The cur
vature K of a curve is defined as the rate at which the unit
tangent t changes with respect to s. The change of arc length
is
(11.1)
We have the curvature
dt 1
 =Kn= n.
ds p ·
(11.2)
where p is the radius of the curvature, and n is the principal
normal. As the direction of the tangent is defined by the angle
8 made with the xaxis by t, we have tan() = y'. Hence, the
curvature becomes
d() dO dx
K==
ds dx ds ·
(11.3)
Fron1 () = tan
1
y'(x), we have
dO [ _
1
( ')]' y"
dx = tan y = ( 1 + yl2 r
(11.4)
Using the expression for dsjdx, the curvature can be written
in terms of y(x), and we get
(11.5)
11.1.2 EulerLagrange Equation
Since the calculus of variations is always related to some min
imization or maximization, we can in general assume that the
integrand 1/J of the integral is a function of the shape or curve
y( x) (shown in Figure 11.1), its derivative y' ( x) and the spatial
154
Integral Equations 11.1 Calculus of Variations
coordinate x (or time t, depending on the context). For the
integral
I = l '1/;(x, y, y')dx, (11.6)
where a and bare fixed, the aim is to find the solution of the
curve y( x) such that it makes the value of I stationary. In this
sense, I[y(x)] is a function of the function y(x), and thus it is
referred to as the functional.
y
a
'__,. X
Figure 11.1: Variations in the path y(x).
Here, stationary n1eans that the small change of the first
order in y( x) will only lead to the secondorder changes in val
ues of I[y(x)], and subsequently, the change 8I of I should be
virtually zero due to the small variation in the function y(x).
'Iranslating this into the 1nathematical language, we suppose
that y(x) has a small change of magnitude off so that
y(x) + y(x) + f((x), (11.7)
where ( ( x) is an arbitrary function. The requirement of I to
be stationary means that
or more accurately,
di
df 'E=0 = 0,
8I = 0,
for all ((x).
155
(11.8)
(11.9)
11.1 Calculus of Variations Integral Equations
Thus I becomes
I(y, <) = t ,P(x, y + <(, y
1
+ <(')dx
= [ 1/J(x,y,y')dx+ [[<((: +(':Jdx+0(<
2
). (11.10)
The first derivative of I should be zero, and we have
(11.11)
which is exactly what we mean that the change ~ I (or the first
order variation) in the value of I should be zero. Integrating
this equation by parts, we have
(11.12)
If we require that y(a) and y(b) are known at the fixed points
x = a and x = b, then these requirements naturally lead to
((a) =((b)= 0. This means that the above right hand side of
the equation is zero. That is,
[
81/J] b
( 8y' a= O,
(11.13)
which gives
1
b [81/J ~ 81/J](dx = 0.
a 8y dx 8y'
(11.14)
As this equation holds for all ( ( x), the integrand must be zero.
Therefore, we have the wellknown EulerLagrange equation
(11.15)
It is worth pointing out that this equation is very special in the
sense that 1/J is known and the unknown is y(x). It has many
applications in mathematics, natural sciences and engineering.
156
Integral Equations 11.1 Calculus of Variations
The simplest and classical example is to find the shortest
path on a plane joining two points, say, (0,0) and (1, 1). We
know that the total length along a curve y( x) is
L= l J!+y'2dx.
(11.16)
Since '1/J = J1 + y'
2
does not contain y, 0. Fron1 the
EulerLagrange equation, we have
(11.17)
its integration is
8'1/J y'
ay' = v1 + yn = A.
(11.18)
Rearranging it as
'2 A2
y = 1 A2 '
or
I A
y = v'1 A2'
(11.19)
and integrating again, we have
y = kx +c,
k A
 v'1 A2•
(11.20)
This is a straight line. That is exactly what we expect from
the plane geometry.
D Example 11.1: The EulerLagrange equation is very general
and includes many physical laws if the appropriate form of 1/J is used.
For a point mass m following under the Earth's gravity g, the action
(see below) is defined as
1/J =
mgy =
 mgy,
where y(t) is the path, and now x is replaced by t. v = y is the
velocity. The EulerLagrange equation becomes
81/J d 81/J
8y = dt ( 8v ),
157
11.1 Calculus of Variations Integral Equations
or
d
mg = dt (mv),
which is essentially the Newton's second law F = ma because the
right hand side is the rate of change of the momentum mv, and the
left hand side is the force.
Well, you may say, this is trivial and there is nothing new
about it. This example is indeed too simple. Let us now study
a more complicated case so as to demonstrate the wide appli
cations of the EulerLagrange equation. In mechanics, there is
a Hamilton's principle which states that the configuration of
a mechanical system is such that the action integral I of the
Lagrangian .C = T  V is stationary with respect to the varia
tions in the path. That is to say that the configuration can be
uniquely defined by its coordinates qi and when moving
from one configuration at time t
0
to another time t = t*
I = fo'• .C(t, q;, tj;)dt, i = ... ,N, (11.21)
where T is the total kinetic energy (usually, a function of Qi),
and V is the potential energy (usually, a function of q). Here
Qi means
. 8qi
qi = m·
(11.22)
In analytical mechanics and engineering, the Lagrangian .C
(=Kinetic energy  Potential energy) is often called the action,
thus this principle is also called the principle of least action.
The physical configuration or the path of movement follows
such a path that makes the action integral stationary.
In the special case, x + t, the EulerLagrange equation
becomes
(11.23)
which is the wellknown Lagrange's equation. This seems too
abstract. Now let us look at a classic example.
158
Integral Equations 11.1 Calculus of Variations
Figure 11.2: A simple pendulum.
0 Example 11.2: For a simple pendulum shown in Figure 11.2,
we now try to derive its equation of oscillations. We know the kinetic
energy T and the potential energy V are
V = mgh = mgl(l cos 8).
Using C, = T  V, q = 8 and q = iJ, we have
which becomes
d .
mgl sin 0  dt ( ml
2
0) = 0.
Therefore, we have the pendulum equation
This is a nonlinear equation. If the angle is very small (8 << 1 ),
sin 8 ~ 8, we then have the standard equation for the linear harmonic
motion
0
159
11.1 Calculus of Variations Integral Equations
11.1.3 Variations with Constraints
Although the stationary requirement in the calculus of varia
tions leads to the minimization of the integral itself, there is
no additional constraint. In this sense, the calculus of varia
tion discussed up to now is unconstrained. However, sometimes
these variations have certain additional constraints, for exam
ple, the sliding of a bead on a hanging string. Now we want to
make the integral I stationary under another constraint inte
gral Q that is constant. We have
I= t 1/J(x,y,y')dx, (11.24)
subjected to the constraint
Q = t cf>(x, y, y')dx. (11.25)
As for most optimization problems under additional constraints,
the method of Lagrange multipliers can transform the con
strained problem into an unconstrained one by using a com
bined functional J = I + >..Q or
J = t[,P + >.¢]dx, (11.26)
where ).. is the undetermined Lagrange multiplier. Replacing
1/J by [1/J + >..<P] in the EulerLagrange equation or following the
same procedure of the derivations, we have
(11.27)
Now we can come back to our example of the hanging rope
problem with two fixed points. The total length of the rope is
L, and it hangs frmn two fixed points ( d, 0) and (d, 0). From
the geometric consideration, it requires that 2d < L. In order
to find the shape of the hanging rope under gravity, we now
define its gravitational potential energy EP as
Ep = j_:[pgy(x)ds] = pg j_:yJl + y12dx.
160
(11.28)
Integral Equations 11.1 Calculus of Variations
The additional constraint is that the total length of the rope is
a constant ( L). Thus,
Q = Jd J1 +y
12
dx = L.
d
(11.29)
By using the Lagrange multiplier A, we have J = Ep + ,\Q, or
J = Jd [pgy +X] VI+ y'
2
dx.
d
(11.30)
Since \II = [pgy + X] vi + y'
2
does not contain X explicitly, or
= 0, then the EulerLagrange equation can be reduced into
a simpler form in this special case. Using
d\11 8\11 ow dy ow dy'
=++
dx 8x 8y dx oy' dx
,ow ,ow
= 0 + y oy + y oy'' (11.31)
and the EulerLagrange ), we have
(11.32)
which can again be written as
d ,8\11
dx[w Y 8y'] = o.
(11.33)
The integration of this equation gives
,8\11
\II  y
8
y' = A = const.
(11.34)
Substituting the expression of \II into the above equation, the
stationary values of J requires
(11.35)
161
11.1 Calculus of Variations Integral Equations
Multiplying both sides by vf1 + y''l and using the substitution
A cosh(= pgy +A, we have
(11.36)
whose solution is
h
t[PYY +A]= xpg + K
cos A A .
(11.37)
Using the boundary conditions at x = ±d and the constraint
Q = L, we have K = 0 and in1plicit equation for A
. h(pgd) _ pgL
SlllA2A.
(11.38)
Finally, the curve for the hanging rope becomes the following
catenary
A pgx pgd
y(x) = pg[cosh(A) cosh(y)].
(11.39)
0 Example 11.9: For the hanging rope problem, what happens
if we only fix one end at (a, 0), while allowing the free end of the
hanging rope to slide on a vertical pole? lVell, this forms a variation
problem with variable endpoint(s). We assume that free end is at
(0, y) where y acts like a free parameter to be determined. Now the
boundazy condition at the free end is different. Since the variation
of 81 = 0, we have
l
b 8\J! d 8\J! 8\J!
8J = [ ()](dx + [(]b = 0.
a 8y dx 8y' oy' a
As the variation ( is now nonzero at the free end point, we then have
8\J!
oy' = o.
From J = Ep + >..Q, we have \J! = (pgy + >..) J 1 + y'
2
• Thus, we get
a
8y' [(pgy + >..)vfl + y'
2
] = 0,
162
Integral Equations 11.1 Calculus of Variations
or
y'(pgy + >..)/ Jt + y'
2
= 0, or y' = 0.
In other words, the slope is zero at the free end. 0
Such a boundary condition of y' = 0 has the real physical
meaning because any nonzero gradient at the free end would
have a nonzero vertical component, thus causing the vertical
slip along the rope due to the tension in the rope. The zero
gradient leads to the static equilibrium. Thus, the whole curve
of the hanging rope with one free end forms half the catenary.
0 Example 11.4: Dido's problem concerns the strategy to enclose
a maximum area with a fixed length circumference. Legend says
that Dido was promised a piece of land on the condition that it was
enclosed by an oxl1ide. She had to cover as much as land as possible
using the given oxhide. She cut the oxhide into narrow strips with
ends joined, and a whole region of a hill was enclosed.
Suppose the total length of the oxhide strip is L. The enclosed
area A to be maximized is
1
xb
A= y(x)dx,
X a
where X a and xb are two end points (of course they can be the same
points). \rVe also ha.ve the additional constraint
1
Xb
Jl + y
1
2dx = L = const.
X a
This forms an isoperimetric variation problem. As L is fixed, thus
the maximization of A is equivalent to make I = A + >..L stationary.
That is
I= A+ AL = {' [y + A.jl +ifl)dx.
Using the EulerLagrange equation, we have
or
163
11.1 Calculus of Variations
which becomes
1  >.. .!!._ ( y' ) = 0
dx Jt + y'2 ·
Integrating it once, we get
>..y'
====x+K,
Jt + y'2
Integral Equations
where K is the integration constant. By rearranging, we have
y' = ± X+ K
J>..2 (x + K)2
Integrating this equation again, we get
y(x) = =F)>..
2
 (x + K)2 + B,
where B is another integration constant. This is equivalent to
which is essentially the standard equation for a circle with the centre
at (K, B) and a radius >... Therefore, the most area that can be
enclosed by a fixed length is a circle. 0
An interesting application is the design of the slides in play
grounds. Suppose we want to design a sn1ooth (frictionless)
slide, what is the best curve/shape the slide should take so
that a child can slide down in a quickest way? This problem is
related to the brachistochrone problem, also called the short
est time problem or steepest descent problem, which initiated
the development of the calculus of variations. In 1696, Johann
Bernoulli posed a problem to find the curve that n1inimizes the
time for a bead attached to a wire to slide from a point (0, h)
to a lower point (a, 0). It was believed that Newton solved it
within a few hours after receiving it. From the conservation
of energy, we can determine the speed of the bead from the
equation ~ m v
2
+ mgy = mgh, and we have
v = J2g(h y). (11.40)
164
Integral Equations 11.1 Calculus of Variations
So the total time taken to travel from (0, h) to (a, 0) is
lo
a 1 loa )1 + y'2
t = ds = dx.
0 v 0 J2g(h  y)
(11.41)
Using the simplified EulerLagrange equation (11.34) because
the integrand \ll = )1 + yrl / J2g(h y) does not contain x
explicitly, we have
(1 + y
12
) I 8 (1 + y
1
2)
y [ ] A
2g(h y)  8y' 2g(h y)  .
(11.42)
By differentiation and some rearrangements, we have
'2 Bh+y
y = '
hy
1
B = 2gA2.
(11.43)
By changing of variables TJ = h y = ~ ( 1 cos 0) and inte
grating, we have
x = ~ [0  sin 0] + k, (11.44)
where 0 < 1r and k is an integration constant. As the curve
must pass the point (0, h), we get k = 0. So the parametric
equations for the curve become
x = ~ ( 0  sin 0),
B
y = h "2(1 cosO). (11.45)
This is a cycloid, not a straight line, which seems a bit surpris
ing, or at least it is rather counterintuitive. The bead travels
a longer distance, thus has a higher average velocity and sub
sequently falls quicker than traveling in a straight line.
11.1.4 Variations for Multiple Variables
What we have discussed so far mainly concerns the variations in
2D, and subsequently the variations are in terms y(x) or curves
only. What happens if we want to study a surface in the full 3
D configuration? The principle in the previous sections can be
165
11.1 Calculus of Variations Integral Equations
extended to any dimensions with multiple variables, however,
we will focus on the minimization of a surface here. Suppose
we want to study the shape of a soap bubble, the principle of
least action leads to the minimal surface problem. The surface
integral of a soap bubble should be stationary. Now we assume
that the shape of the bubble is u( x, y), then the total surface
area is
where
W=
{)u au v
1 + ()2 + ()2 = 1 + u2 + u2.
ax {)y X y
(11.47)
In this case, the extended EulerLagrangian equation for two
variables x and y becomes
(11.48)
Substituting w into the above equation and using ~ ! = w u = 0
since w does not contain u explicitly, we get
(11.49)
or
(1 + u ~ ) U x x  2uxUy + (1 + u;)uyy = 0. (11.50)
This is a nonlinear equation and its solution is out of the scope
of this book. This nonlinear equation has been one of the active
research topics for more than a century. It has been proved that
the fundamental solution to this equation is a sphere, and in
fact we know that all bubbles are spherical. For some problems,
we can approximately assume that Ux and uy are small, thus
the above equation becomes Laplace's equation
Uxx + Uyy = 0. (11.51)
166
Integral Equations 11.2 Integral Equations
The calculus of variations has many applications. The other
classical examples include Fermat's principle in optics, Sturm
Liouville problem, surface shape minimization, the action prin
ciple, and of course the finite element analysis.
11.2 Integral Equations
From the calculus of variations, we know that the unknown
y( x) to be optimized is inside the integrand of I. In certain
sense, this is an integral equation. In fact, many physical
processes and laws of conservation are expressed in terms of
integral forms rather than their differentiation counterparts.
Naturally, one of the ways of constructing an integral equa
tion is to integrate from a differential equation. Integral equa
tions are much more complicated compared with the differential
equations. There is no universal solution technique for nonlin
ear equations, even the numerical simulations are usually not
straightforward. Thus, we will mainly focus on the simplest
types of integral equations.
11.2.1 Linear Integral Equations
Fredholm Integral Equations
A linear integral equation for y( x) can be written in the follow
ing generic fonn
u(x) +A l K(x, TJ)Y(TJ)dTJ = v(x)y(x), (11.52)
where K(x, rJ) is referred to as the kernel of the integral equa
tion. The parameter A is a known constant. If the function
u( x) = 0, the equation is then called homogeneous. If u( x) =/= 0,
the equation is inhomogeneous.
If the function v( x) = 0, then the unknown y( x) appears
only once in the integral equation, and it is under the integral
167
11.2 Integral Equations Integral Equations
sign only. This is called the linear integral equation of the first
kind
u(x) +At K(x, 1/)Y('I)d'l = 0.
(11.53)
On the other hand, if v(x) = 1, equation (11.52) becomes the
integral equation of the second kind
u(x) +A l K(x, 1/)Y('I)d'l = y(x). (11.54)
An integral equation with the fixed integration limits a and b,
is called a Fredholm equation. If the upper integration limit
b is not fixed, then the equation becomes a Volterra equation.
The integral equation becomes singular and at least one of its
integration limits approaches infinite.
Volterra Integral Equation
In general, the Volterra integral equation can be written as
u(x) +A lx K(x, TJ)Y(TJ)dTJ = v(x)y(x). (11.55)
The first kind [or v(x) = 0] and second kind [or v(x) = 1] are
defined in the similar manner.
The kernel is said to be separable or degenerate if it can be
written in the finite sum form
N
K(x, TJ) = L, fi(x)gi(TJ), (11.56)
i=l
where fi(x) and gi(TJ) are functions of x and TJ, respectively. A
kernel is called a displacen1ent kernel if it can be written as a
function of the difference ( x  TJ) of its two arguments
K(x, TJ) = K(x TJ). (11.57)
168
Integral Equations 11.3 Solution of Integral Equations
11.3 Solution of Integral Equations
Iviost integral equations do not have closedform solutions. For
linear integral equations, the closedform solutions are only pos
sible for the special cases of separable and displacement kernels.
11.3.1 Separable Kernels
For a Fredholm integral equation of the second kind with sep
arable kernels, we can substitute the kernel (11.56) into the
equation and we have
b N
u(x) +A J. tr J;(x)g;(17)d1J = y(x),
(11.58)
which becomes
N b
u(x) +A tr /;(x) J. g;(17)d17 = y(x).
(11.59)
Because the integration limits are fixed, the integrals over 1J
should be constants that are to be determined. By defining
<>; = t g;(1J)Y(1J)d1J,
we now have the solution in the form
N
y(x) = u(x) +A L aifi(x),
i=l
where the N coefficients Oi are determined by
(11.60)
(11.61)
(11.62)
fori= 1, 2, ... , N. Only for a few special cases, these coefficients
can be written as simple explicit expressions.
169
11.3 Solution of Integral Equations Integral Equations
11.3.2 Displacement Kernels
For a singular integral equation with a displacement kernel, the
equation can be solved by Fourier transforms if both integration
limits of the integral are infinite. In this case, we have
u(x) +A/_: K(x 1J)Y(1J)d1J = Y(1J).
(11.63)
Using the Fourier transforms and the convolution theorem, we
have
U(w) + Av'2?fK(w)Y(w) = Y(w), (11.64)
which is an algebraic equation for Y(w). Its solution is simply
Y(w) = U(w) .
1  Av'2iK(w)
(11.65)
The solution y( x) can be obtained using the inverse Fourier
transform
y(x) = _1_ roo V(w) eiwxdw.
y'2i 1oo [1 Av'2iK(w)]
(11.66)
11.3.3 Volterra Equation
A Volterra equation with separable kernels may be solved by
transforming into a differential equation via direct differentia
tion. In the case of a simple degenerate kernel
K(x, 17) = f(x)g(1J), (11.67)
we have
y(x) = u(x) +A fox f(x)g(1J)Y(1J)d1J, (11.68)
which becomes
y(x) = u(x) + Aj(x) fox g(1J)Y(1J)d1J. (11.69)
170
Integral Equations 11.3 Solution of Integral Equations
If f ( x) =/= 0, it can be written as
y(x) u(x) rx
f(x) = f(x) +A lo g(TJ)Y(TJ)dTJ.
(11.70)
Putting ¢(x) = u(x)/ f(x) and differentiating it, we have
~ ~ ~ : ~ ] ' = 4>'(x) + Ag(x)y(x).
(11.71)
By letting W = y(x)j f(x), we have
w'(x) .Xf(x)g(x)\ll(x) = ¢'(x), (11.72)
which is a firstorder ordinary differential equation for W ( x).
This is equivalent to the standard form
w' + P(x)\ll = Q(x), (11.73)
and
P(x) = .Xf(x)g(x),
Q(x) = ~ ~ ~ : ~ ] ' .
(11.74)
We can use the standard technique by n1ultiplying the integrat
ing factor exp[J P(x)dx] to obtain the solution. We get
y(x) = f(x)[e J P(x)dx]{j [Q(x)ef P(x)dx]dx }.
(11.75)
With appropriate boundary conditions, the exact form of the
solution can be obtained.
0 Example 11.5: Let us tzy to solve the integral equa.tion of
Volterra type
y(x) =ex+ lax ex sin(()y(()d(.
First, we divide both sides by ex, we get
y(x) = 1 + rx sin(()y(()d(,
ex Jo
171
11.3 Solution of Integral Equations Integral Equations
whose differentiation with respect to x leads to
y(x) ' .
[&] = y(x)sm(xL
or
..!..y'(x) y(x)ex = y(x) sin(x).
ex
Divide both sides by y(x) and using [Iny(x)]' = y'(x)fy(x), we have
[In y(x)]' =ex sin x + 1.
By direct integration, we have
lny(x) = x ~ e x cosx + ~ e x sinx.
Thus, we finally obtain
ex
y(x) = exp[x
2
(cosx sinx)].
0
There are other n1ethods and techniques of solving integral
equations such as the operator method, series method and the
Fredholm theory. However, most integral equations do not have
closedform solutions. In this case, numerical methods are the
best alternative.
172
Chapter 12
Tensor Analysis
Ivlany physical quantities such as stresses and strains are ten
sors. Vectors are essentially firstorder tensors. Tensors are the
extension of vectors, and they can have any number of dimen
sions and any orders, though most commonly used tensors are
second order tensors.
12.1 Notations
In tensor analysis, the Einstein sun1mation convention or Ein
stein notations
1
and notations for subscripts are widely used.
Any lowercase subscript that appears exactly twice in any term
of an expression means that sum is over its all possible values
of the subscript. For example, in the threedimensional case,
we have
3
GiXi = L aiXi = GtXt + a2x2 + G3X3. (12.1)
i=l
3
~ j B j k = L AijBjk = AitBlk + Ai2B2k + Ai3B3k· (12.2)
j=l
1
This notation convention was introduced by Albert Einstein in 1916
when formulating the theory of General Relativity.
173
12.2 Tensors Tensor Analysis
OUi = \7 . u = OUt + OU2 + OU3 (12.3)
8xi  OXt 8x2 8x3 ·
The Kronecker delta 8ij which is a unity tensor (like the unity
matrix I in matrix analysis), is defined as
{
1 (if i = j),
8
ij = 0 (if i i= j).
(12.4)
For a tensor with three subscripts sin1ilar to 8ij, the LeviCivita
symbol or tensor is defined as
{
+1 if (i,j, k) is an even permutation of (1, 2, 3),
Eijk = 1 if (i,j, k) is an odd permutation of (1, 2, 3),
0 (otherwise).
(12.5)
The tensors 8ij and Eijk are related by
fijkfkpq = 8ip8jq  8iq8jp·
(12.6)
Using the summation conventions, the matrix equation
Ax=b, (12.7)
can alternatively be written as
(i=1,2, ... ,n). (12.8)
12.2 Tensors
When changing the bases from the standard Cartesian e
1
= i,
e2 = j, e3 = k to a new set of bases ei, e ~ , e ~ , a position vector
x = (x
1
, x2, x
3
) in the old bases is related to the new vector
x' = ( x!, x ~ , x ~ ) in the new bases by a coefficient matrix Sij.
Sij can be the rotation, translation, enlargement or any of their
combinations. For example, the matrix for a simple rotation,
with an angle of () around a fixed axis, Sij becomes
Sij= ( ~ : o ~ : D·
(12.9)
174
Tensor Analysis 12.3 Tensor Analysis
The orthogonality of Sij requires that ssT= sTs =I or
(12.10)
If the components ui of any variable u are transformed to the
components u ~ in the new bases in the same manner as
(12.11)
then ui(i = 1, 2, 3) are said to form a firstorder Cartesian
tensor (or vector in this case). If components of a variable such
as strains aij are transformed as
(12.12)
we say these components form a secondorder tensor.
The order of a tensor is also called its rank. Scalars have
rank zero, vectors have rank 1, and secondorder tensors have
rank 2. In engineering and computing, the rank is associated
with the number of indices to describe a tensor in terms of a
multidimensional array. In this sense, a secondorder tensor is
equivalent to a twodimensional array or a matrix.
In a similar fashion, higherorder tensors can be defined,
and for each order increase, then there is one Sij extra in the
product for transforming, but no subscripts are allowed to ap
pear more than twice
(12.13)
and
(12.14)
12.3 Tensor Analysis
Tensors in Cartesian Coordinates
One of the main advantages of tensors is that a tensor is in
dependent of any chosen frame of reference. Therefore, any
175
12.3 Tensor Analysis Tensor Analysis
physical laws or equations that are formulated in terms of ten
sors should be independent of frame of reference. For example,
the stress strain relation in linear elasticity is independent of
frame of reference (for details see next chapter)
(12.15)
where aij and Cij are the stress tensor and strain tensor, r&
spectively. J.l and A are Lame constants.
In the similar way as n1ultidimensional arrays or matri
ces, two tensors can be added or subtracted componentby
component if and only if they are the tensors of the same order.
For secondorder tensors, a tensor Tij is said to be symmetric
if Tij = Tji, and antisymmetric if Tij = Tji· An interesting
property of a tensor Tij is that it can always be written as a
sum of a symmetric tensor and an antisymmetric tensor
Tij = ~ ( T i j + Tji)[sym.] + ~ ( T i j  Tji)[antisym.].
(12.16)
All the formulas in vector analysis can be rewritten in the
tensor forms using the summation convention and notations
\1
2
1/J = {)21/J = Oij {)1/J
8xi8Xi 8xi8Xj'
8uq
\7 X (\7 X u)i = fijkfkpq
8
.
XjXp
(12.17)
(12.18)
(12.19)
Similarly, the divergence theorem can be rewritten as the fol
lowing form
f
8
8
ui dV = J uinidS.
lv xi !s
(12.20)
The tensor forn1s are sometimes useful to the proof of the com
plex relationship among vectors and tensors. They also become
handy for the implementation of numerical algorithms.
176
Tensor Analysis 12.3 Tensor Analysis
Using the tensor notations, we have the identity oiiAi = Ai·
The cross product Ax B can be expressed as
(12.21)
Another way of denoting the derivatives of tensors a and v
is to use following notations
a a
a i = ~ ,
' UXi
(12.22)
where the index of the spatial component Xi is denoted by a
comma to avoid any potential confusion with other indices.
With these notations, the important operators involving the
V' operator can be written as
and
grad¢ = ¢, i = V' ¢,
Y'
2
¢ = ¢,ii = ~ ¢ ,
8
2
u1 8
2
u2 8
2
u3
V'.v = Vi,i = !i""""2 + '£l"2 + 
8
2 ·
UXl UX2 X3
A very special case is that
(12.23)
(12.24)
(12.25)
(12.26)
0 Example 12.1: Let us now use the tensor notations to prove
ax (b x c) = b(a ·c)  c(a ·b). Fl·om above expressions for cross
products, we know that
a X (b X c) = Eijkaj(Ekpqbpcq) = EijkEkpqajbpCq
= (8ip8jq 8iq8jp)ajbpcq = 8ipbp(8iqajcq) 8iqcq(8iPaibp)·
Using 8ipbp = b and 8iqcq = c, we have
a(b x c) = b(tSjqajcq)  c(tSiPaibp)·
By renaming the indices (j  i and q  j) so that 8iqajcq
8ijaici = a • c and tSiPaibP = a • b, we finally obtain
ax (b x c)= b(a ·c) c(a ·b).
177
0
12.3 Tensor Analysis Tensor Analysis
Tensors in NonCartesian Coordinates
The tensors we have discussed so far are expressed in Carte
sian coordinates. In nonCartesian coordinates, they are more
complicated. Tensor analysis is very important in theoretical
physics and differential geometry where formal mathematical
theory is required. In fact, many books on tensor analysis use
a modern approach in terms of tensor duality, covariance and
contravariance concepts. In the simplest term, a tensor such
as a vector v can be expressed as the sum of its components
multiplying by the basis vectors
(12.27)
where vi(i = 1, 2, 3) are called the covariant components of v
in the contravariant basis vectors ei, while vi(i = 1, 2, 3) are
called the contravariant components of v in the covariant basis
vectors ei. For the curvilinear coordinates ( Ql , Q2, q3) at any
point P on a position vector r(q
1
, q
2
, q
3
), the basic vectors are
given by
(i=1,2,3), (12.28)
where ei and ei are reciprocal systems of vectors, and ei · ei =
8j, where 8j acts in the similar way as 8ii· In many books on
tensor analysis, ei is also written as €i = ei.
Furthermore, the tensor product (also called outer product)
of two tensors is rather complicated. For example, the tensor
product u Q9 v = uvT of two vectors u and vis given by
U@V = ( ~ ~ )@(Vt V2 V3) = ( ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ : ) . (12.29)
U3 U3V1 U3V2 U3V3
Using these basis vectors, we can write a secondorder tensor
in tern1s of covariant components aii and contravariant com
ponents aii as
(12.30)
178
Tensor Analysis 12.3 Tensor Analysis
In a given frame of reference, the fundamental metric tensor is
defined by
(12.31)
This tensor is always symmetric 9ii = 9ji, its determinant g =
l9ij I = det(gij) is related to the Jacobian J = Vff·
The derivatives in nonCartesian coordinates are far more
complicated, and they usually involve the Christoffel coeffi
cients rt = ek · ~ . For example, the divergence of a vector
is defined by the covariant differentiation u1i or V · u = u ~ i =
~ u ~ + r ~ . u i = ui.+r1
1
.ui. It is worth pointing out that ei and ei
vq
1
l] ,] •
become identical and r ~ j = 0 in Cartesian coordinates, and it
is therefore not necessary to distinguish the contravariant and
covariant vectors and components.
Mathematically speaking, the formal approach is preferred.
In engineering mathematics, however, the simple formulation in
terms of multidimensional arrays in Cartesian coordinates is a
more convenient approach, especially from the computational
point of view. That is why we have used an oversimplified
approach here. In the next chapter, we will study the theory
of linear elasticity as an application of tensor analysis.
179
Chapter 13
Elasticity
13.1 Hooke's Law and Elasticity
The basic Hooke's law of elasticity concerns an elastic body
such as a spring, and it states that the extension x is propor
tional to the load F, that is
F=kx, (13.1)
where k the spring constant. However, this equation only works
for 1D deformations. For a bar of uniform crosssection with
a length L and a cross section area A, it is n1ore convenient to
use strain e and stress u. The stress and strain are defined by
F
u = A, e = L' (13.2)
where the extension. The unit of stress is N jm
2
, while the
strain is dimensionless, though it is conventionally expressed in
mjm or% (percentage) in engineering. For the elastic bar, the
stressstrain relationship is
u=Ec, (13.3)
where E is the Young's modulus of elasticity. Written in terms
F and x = we have
EA
F=
181
k= EA
L'
(13.4)
13.1 Hooke's Law and Elasticity Elasticity
where k is the equivalent spring constant for the bar. This
equation is still only valid for any unidirectional compression
or extension. For the 2D and 3D deformation, we need to
generalize Hooke's law. For the general stress tensor (also called
Cauchy stress tensor)
(13.5)
and strain tensor
e = ( ::: ::: ::: ) = ( : ~ ~ : ~ ~ : ~ : ) '
c::;x c::;y c::;::; c31 c32 c33
(13.6)
it can be proved later that these tensors are symmetric, that is
u = uT and e = eT, which leads to
O"xy = O"yx' O"xz = O":;x, O"yz = O"::;y'
(13.7)
and
cxy = cyx, cxz = c::;x, cy::: = c:::y·
(13.8)
Therefore, we only have 6 independent components or unknowns
for stresses and 6 unknown strain cmnponents.
The strain tensor is defined by the displacement u T =
1 8ui 8uj
Cij = 
2
(
8
+ 
8
), (13.9)
Xj Xi
where Xt = x, x2 = y, and x3 = z. Sometimes, it is useful to
write
1
e = 2'(V'u + V'uT). (13.10)
The generalized Hooke's law can be written as
1
Cxx = E[axx v(ayy + O"::;z)],
(13.11)
1
cyy = E[ayy v(axx + O"zz)],
(13.12)
182
Elasticity 13.1 Hooke's Law and Elasticity
1
c::::: = E [u:::: v(uxx + O"yy)],
(13.13)
1+v
cxy = gO"xy,
(13.14)
1+v
ex;:: = gO"x;:;,
(13.15)
(13.16)
where v is the Poisson's ratio, and it measures the tendency
of extension in transverse directions (say, x and y) when the
elastic body is stretched in one direction (say, z). It can be
defined as the ratio of the transverse contract strain (normal
to the applied load) to the axial strain in a stretched cylindri
cal bar in the direction of the applied force. For a perfectly
incompressible material, v = 0.5, and v = 0 rv 0.5 for most
common materials. For example, steels have v = 0.25 rv 0.3.
Some auxetic material such as polymer foams or antirubbers
have a negative Poisson's ratio v < 0.
This generalized Hooke's law can concisely be written as
(13.17)
where we have used the Einstein's summation convention O"kk =
O"xx+O"yy+O"::::· Another related quantity is the pressure, which
is defined by
1
p = O"kk =
3
(13.18)
The negative sign comes from the conventions that a positive
normal stress results in tension, and negative one in compres
sion, while the positive pressure acts in compression. Some
times, it is more convenient to express the stress tensor in terms
of pressure and devitoric stress tensor Sij
(13.19)
183
13.1 Hooke's Law and Elasticity Elasticity
If we want to invert equation (13.17), we have first express
akk in terms of Ckk so that the right hand side of the new
expression does not contain the stress akk· By contraction using
j + i, we have
(13.20)
where we have used 8ii = 811 + 822 + 833 = 1 + 1 + 1 = 3 and
aii = akk· In engineering, the quantity
fPut 8
2
u2 8
2
u3
Ckk = Cxx + Cyy + Cz::: = Bx
2
+ By
2
+ Bz
2
= V · u, (13.21)
means the fractional change in volume, known as the dilation.
This gives that
E
CJii = CJkk =
1
_
2
v Ckk·
(13.22)
Substituting it into equation (13.17), we have
1 +v v E
Cij = gaii  E (
1
_
2
v Ckk)8ij,
(13.23)
or after some rearrangement
1+v v
gaii = Cij +
1
_
2
v Ckk8ij,
(13.24)
which can be written as
(13.25)
where J.L and A are La1ne constants. They are
E
G = J.l = 2(1 + v)'
A= vE .
(1 + v)(1  2v)
(13.26)
This stressstrain relationship can also be written as
u = 2Ge + A(V · u)l5. (13.27)
In engineering, G = J.L is called the shear modulus, while K =
3
(
1
~
2
, ) is called the bulk modulus which is the ratio of pressure
p to the volume change ~ V.
184
Elasticity 13.2 P.tfaxwell's Reciprocal Theorem
13.2 Maxwell's Reciprocal Theorem
For an elastostatic problem, the balance of force leads to
V ·U+ b = 0, (13.28)
where b is the body force or force per unit volume. For a small
cube volun1e element, the total body force dfi along the xiaxis
is
8aii
dfi = dbi = 
0
dV,
x·
J
(13.29)
here we have used the index summation conventions. Similarly,
the total force along the jaxis is
d/j = oaiidV.
OXi
(13.30)
Since there is no relative rotation of the cube element because
the cube element must be at rotational equilibrium, thus the
result moment must be zero. Taking the moment of the two
force components about any point (say, a corner of the cube).
This leads to
(13.31)
which means the stress tensor is symmetric. This is in fact the
compatibility condition for stresses.
Alternatively, consider a cube element at rotational equi
librium with a volume dV = 8x8y8z, and the dimensions of the
elements are 8x, 8y and 8z, respectively. We consider the forces
along the four faces that are parallel to zaxis (shown in Figure
13.1). If we take the moment about a line which is parallel
to zaxis and goes through one corner point, we only have to
consider the two faces that are far from this point because the
two faces through this point do not contribute to the moment.
The forces on the two faces are O'xydAtdY and ayxdA2dx where
the surface areas are dAt = 8x8z and dA2 = 8y8z. The total
moment about point D is
(13.32)
185
13.2 !viaxwell's Reciprocal Theorem Elasticity
Thus, axy = ayx· Similar arguments for other faces along x
and ydirections about any other points and we have aii = Uji·
z
y
8y
Figure 13.1: A cubic element in an elastostatic body.
Using the stressstrain relationship, it is straightforward to
prove that the strain tensor is also symmetric (cij = cji)· In
engineering, this tensor is often written as
(13.33)
so as to emphasize that the nondiagonal elements are for shear
components Txy etc.
From the matrix algebra, we know that a square matrix can
always be expressed in terms of eigenvalues and eigenvectors.
For the stress tensor, the eigenvalue problen1
(u ail)ft = 0, (13.34)
provides the the principal stresses ai(i = 1, 2, 3) (eigenvalues)
and their principal directions ftT = n1, n2, n3 (eigenvectors).
In the coordinate system formed by the three eigenvectors, the
stress is expressed by the three principal stresses along three
principal directions, and there are no shear stress components.
The nontrivial solutions require the determinant of the co
186
Elasticity 13.2 P.tfaxwell's Reciprocal Theorem
efficient matrix must be zero. That is
(Jx (J
Txy Tx:;
Txy ay a Ty:; = 0. (13.35)
Tx::; Ty:;
(J:;  (J
It can be expanded into a cubic equation
a3 (ax+ ay + a::;)a2
+[axay + aya:; + a::;ax (r;y + r;::; + r;:;)]a
[axaya::; + 2TxyTy:;Tx::; (axr;::; + ayr;::; + a::;T;y)] = 0. (13.36)
There are three invariants (It, h and I3) for the secondorder
symmetric stress tensor aii, and these invariants satisfy the
characteristic equation
where
I 1 = tr( a ii) = ax + a y + a::;,
h = axay + aya:: + a:;ax (r;y + r;:; + r;:;;),
and
(13.37)
(13.38)
(13.39)
Under appropriate transformations, this tensor can be trans
formed into a diagonal form
(13.41)
where a1, a2, a3 are principal stresses. Written in terms of prin
cipal stresses, the three invariants become
187
(13.42)
(13.43)
13.2 !viaxwell's Reciprocal Theorem Elasticity
and
(13.44)
Now consider an elastic body when n concentrated load
f
1
, f
2
, ... , fn acted upon the body at n different points. The dis
placements at each point in the direction of the corresponding
force are Ql, Q2, ... , Qn. For a linear elastic body, the principle
of superposition applies and we have
(13.45)
(13.46)
(13.47)
where Cij are the influence coefficients or flexibility matrix.
The total work done due to this set of loads is
(13.48)
There is a very useful theorem concerning these coefficients.
It is called Maxwell's reciprocal theorem or MaxwellBetti the
orem, which states that the influence coefficients (or flexibility)
matrix is symmetric Cii = Cji· That is to say, the displace
ment at point i due to a unit load at another point j is equal
to the displacement at j due to a unit load at point i. This
theorem is essentially equivalent to say the displacements are
pathindependent and independent of the order of the loads
applied upon the elastostatic body.
For two forces f i and f i, the final displacements are the
same where fi is applied first, then fi, or fi is applied first
then f i, or even both are applied at the same time. In other
words, the system has no memory of the load history. In the
case of only two forces, we first apply /i slowly (so as to reduce
the dynan1ical effect) with fi = 0, the displacement of point
i is Cii/ i and the displacement of point j is Ciif i. The work
188
Elasticity 13.3 Equations of !vlotion
done is !Ciifl where Jl = fi · fi. Now with /i kept fixed, we
apply I i slowly, the additional displacement at point i is Cij f i
and the additional displacement at point j is Cjj f j. In this
case, the extra work done is Cijfi · fj + !Cii!J. The total work
done at the final state is
1 2 1 2
W =
2
ciifi +
2
ciifi + Cii(fi · fj).
(13.49)
If follow the same procedure but slowly apply I i first, then
f i, the total work done is now
(13.50)
As the total work done should be independent of the order in
which the loads are applied, this requires W = W, which leads
to
(13.51)
Since fi · fj is a dot product and thus a scalar, we now get
(13.52)
This completes the proof of the :rviaxwell's reciprocal theorem.
This theorem is the important basis for boundary element anal
ysis and virtual work method in computational engineering.
13.3 Equations of Motion
For a general solid where the inertia is not negligible, we have
(13.53)
where p is the density of the elastic body. In some books,
the following form of body force b = pf is used, in this case,
the force f means the force per unit mass. Together with the
generalized Hooke's law and relationship with displacement u,
189
13.3 Equations of Motion Elasticity
we have the following set of equations of motion for an elastic
body.
(13.54)
(13.55)
(13.56)
There are 15 equations (6 for stresses, 6 for strains, 3 for dis
placement) and we have 15 unknowns (6 stress con1ponents, 6
strain components and 3 displacements). Therefore, the elastic
field should be uniquely determined if appropriate boundary
conditions are given. There are other compatibility equations
as well, and we will briefly discuss them later.
If we write the equations of motion using the bold font
notations, we have
{)2u
V·u+b=p8t
2
,,
u = 2Ge + .X(V · u)d,
1
e = 2'(Vu + VuT).
(13.57)
(13.58)
(13.59)
If we substitute the generalized Hooke's law and displacement
into the first equation (13.57), we have
or
{)2u
V · [2Ge + .X(V · u )«5] + b = P 8t
2
,
{)2u
V. [G(Vu + VuT) + .X(V · u)«5] + b = P {)t
2
,
which leads to
[)2u
(G + .X)V(V · u) + GV
2
u + b = P {)t
2
•
190
(13.60)
(13.61)
(13.62)
Elasticity 13.3 Equations of !vlotion
Using G +A= G/(1 2v) and after some rearrangements, we
have
(13.63)
which is the wellknown CauchyNavier equation. This equa
tion supports both longitudinal wave ( P wave) and transverse
wave (S wave). In the simplest 1D case without any (external)
body force b, we can take V · u = 0 for Swave, the equation
is simplified as
(13.64)
thus its wave speed is
v s = ~ ·
(13.65)
For the Pwave in 1D, the displacement field is nonrotational,
i.e., V x (V xu)= 0. From the identity V(V · u) = V x (V x
u) + V
2
u, the 1D CauchyNavier equation becomes
8
2
u1 a
2
ut
P fJt2 = (A + 2G) ax2 .
(13.66)
Then, the speed of Pwave is
Vp = J(.>..+p2G).
(13.67)
Since A+ 2G > G, therefore, Pwaves always travel faster than
Swaves.
Furthermore, from the definitions of the strain components
in terms of displacements uT = (ut, u2, u3) = (u, v, w), we have
au
Cxx =ax'
8v
Cyy = oy'
By assuming the displacements are continuous and differen
tiable functions of positions, we differentiate Cxx with respect
191
13.4 Airy Stress Functions Elasticity
to y twice, we have
(13.69)
Similarly, differentiate cyy with respect to x twice, we have
(13.70)
Now differentiate cxy with respect toy once, and with respect
to x once, we have
(13.71)
where we have used the interchangeability of partial derivatives
fPvj8x8y = 8
2
vj8y8x. This can be rearranged as
8
2
cxx + 8
2
cyy =
2
8
2
cxy
8y2 8x
2
8x8y '
(13.72)
which is the compatibility equation. In the same fashion, we
can derive other compatibility equations
(13. 73)
(13.74)
13.4 Airy Stress Functions
For certain engineering problems, the solutions in a plane is
of concern. In this case, we are dealing with plane strain and
plane stress problems. For a plane stress problem, we assume
that a:::: = 0 (but c:;:; =I= 0), then the plane stress problem
involves no stress components depending on z. That is to say
Uxz = Uyz = Uzz = 0. We have only three independent stress
192
Elasticity 13.4 Airy Stress Functions
components axx' aYY' and axy· The generalized Hooke's law
reduces to
However,
1
Cxx = E(axx vayy),
1
Cyy = E(ayy vaxx),
1+v
Cxy = gaxy·
v
C::;::; = 
1
(Cxx + Cyy),
v
which is not zero in general.
(13.75)
(13.76)
(13.77)
(13.78)
For plane strain problems, it is assumed that C::;::; = 0. Thus,
there are only three independent strain components Cxx' Cyy'
and Cxy, however, the stress a::;::; = v(axx + ayy) is not zero.
The compatibility equation becomes
8
2
Cxx + fPcyy =
2
fPcxy
8y
2
8x
2
8x8y ·
(13.79)
For plane strain problems with no body forces, the equilibrium
equations are automatically satisfied if the stress components
are related to a scalar function q,, called Airy's stress function.
The Airy's stress function is defined by
(13.80)
In this case, the compatibility equation becomes
(13.81)
which is a biharmonic equation and can be written as
(13.82)
In cylindrical polar coordinates (r, 8, z), it becomes
8
2
18 18
2
2
[ 8r2 + ; 8r + r2 882] q,  O.
(13.83)
193
13.4 Airy Stress Functions Elasticity
Figure 13.2: A crack in an elastic plate.
Now consider a semiinfinite crack in an infinite solid as
shown in Figure 13.2, assuming the elastic body deforms in
plane strain. The far field is subjected to bounded stress at
infinity. The surfaces of the crack shall be stress free, which
leads to the following boundary conditions
[)2 cp {) 1 {)cp
U()() = {)r2 = 0, Ur() =  {)r (; {)O) = 0, at 0 = ±7r. (13.84)
Let us try a solution of the form
cp = rn+l f(O),
(13.85)
and substitute it into the governing biharmonic equation, we
get
d? 2 d? 2
[d
02
+ (n + 1) ][d
02
+ (n 1) ]f(O) = 0.
(13.86)
As the secondorder equation y" + >..
2
y = 0 has a general solu
tion y = A sin >..O + B cos >..O, we can here use this method twice,
the general solution takes the following form
f(O) = Acos(n + 1)0 + Bsin(n + 1)0
+C cos(n 1)0 + D sin(n 1)0.
The boundary conditions become
aeo = rn
1
n(n + 1)/(0),
194
(13.87)
(13.88)
Elasticity 13.4 Airy Stress Functions
and
Ure = rnln{(n + 1)[Asin(n + 1)8 Bcos(n + 1)8]
+(n 1)[C sin(n 1)8 D cos(n 1)8]}, (13.89)
at 8 = ±1r. We know n = 0 is trivial. From the first equation,
we have
sin(2n7r) = 0,
1 3
n = ±2, ±1, ±2, ... ,
(13.90)
and rn(n ~ 1) does not converge, therefore, they are not suit
able solutions. The constraint now becomes n ::; 0, but the
solutions has singularity as r+ 0. This is however acceptable
in the crack propagation as the stress concentrations do physi
cally exist. Substituting the general solution into the boundary
conditions with n = 1/2 and 8 = ±1r, we get
3A+C= 0, B D = 0. (13.91)
By defining the stress intensity factor K1 for the crack,
K _ 3A/2i
I 4 '
(13.92)
which is for the opening (model I) of the crack. It is a limit of
stress at 8 = 0
K1 =lim uee(r, 8)1 .
ro 6=0
(13.93)
Finally, the solution of stresses can be written as
}(/ . 2 8 8
Urr = . M=:(1 + Slll )COS,
v 21rr 2 2
(13.94)
K
1
3
8
U()() = v'27fTCOS 2'
(13.95)
}(/ 2 8 . 8
Ure = v'21fT cos 2 s1n 2·
(13.96)
195
13.5 EulerBernoulli Beam Theory Elasticity
Once we have the stress distribution, we can get the strains.
Then the displacements are the integration of the strains, and
we have
(13.97)
and
K1(1  v) fiT . 30 . 0
ue = E V ;r[sin
2
 (5 4v) sm
2
].
(13.98)
13.5 EulerBernoulli Beam Theory
The EulerBernoulli beam theory is a simplified theory for cal
culating the deflection of beams under a distribution of load
force using the linear isotropic theory of elasticity. The basic
assumptions for the beam theory are: 1) the beam is isotropic
and elastic; 2) the beam deformation is dominated by bending,
and distortion and rotation are negligible; 3) the beam is long
and slender with a constant cross section along the axis. Under
these assumptions, we can now derive the governing equations.
Let u(x, t) be the deflection of the beam (shown in Figure
13.3), A be the area of the cross section, and f(x, t) be the force
per unit length. The first assumption implies that the bending
moment M is proportional to the curvature K of the bending.
That is
M=EIK, (13.99)
where E is the Young's modulus and I is the area moment of
the beam's cross section. In mechanics, I is also called the
second moment of area or the area moment of inertia. It is
worth pointing out that the area moment about a horizontal
axis through the centroid is defined by
(13.100)
196
Elasticity 13.5 EulerBernoulli Beam Theory
which has a unit of [m]
4
, and it should not be confused with
the mass moment of inertia J (also often denoted as I, but we
use J here) about an axis, which is defined by
(13.101)
with a unit of [Kg] [m]
2
. Both E and I do not change along
the xaxis. For a cylindrical rod with a radius of R, we have
I= 7rR
4
j4. For a rectangular beam with a base width of band
a depth of h, we have I = bh
3
/12.
X
V(x)/ rI
/v(x+dx)
f(x, t)dx
        ,
_____ _J
Tu(x,t)
Figure 13.3: Beam bending.
The second assumption means that the shear V ( x) is related
to the bending moment
8M
8x = V(x),
(13.102)
and the third assumption means tx << 1. Therefore, we have
8
2
u
M::::::: EI 8x2'
(13.103)
or
V:::::::
8x 8x
2
(13.104)
For a small volume element (also shown in Figure 13.3), the
mass of the element is pAdx where p is the density, and the
197
13.5 EulerBernoulli Beam Theory Elasticity
acceleration is fff. The shear force variation V(x + dx)
V ( x) + ~ d x , and the total force is
8V
V(x) V(x + dx) + f(x, t)dx = [f(x, t) ax ]dx. (13.105)
Using the Newton's second law of motion, we have
8V 8
2
u
f(x, t)  ax = pA at2 .
(13.106)
Substituting the above expression for V, we have
8
2
u 8
2
8
2
u
pA at2 + ax2 [EI ax2] = f(x),
(13.107)
which is the EulerBernoulli equation. If there is no force
f(x, t) = 0, the equation becomes a homogeneous form
{)2u {)2 a2u
pA fJt2 + {)x2 [EI {)x2] = 0,
(13.108)
which is a fourthorder wave equation. It governs the waves
that travel along a beam, a rod or any slender column.
This equation can essentially explain why spaghetti and dry
pasta almost always break into more than two fragments. You
can try in your kitchen to break a slender spaghetti by hold
ing its two ends and gradually form an arc and bend beyond
its curvature limit. When a spaghetti rod snaps, it will gener
ally break three to pieces. This phenomenon is very interesting
and once puzzled the famous physicist Richard Feynman for
quite a while, and it was recently studied by two scientists that
the brittle fragmentation process is virtually governed by this
CauchyN a vier equation. They found that the sudden relax
ation of the curvature by first breaking will lead to a burst of
flexural waves along the spaghetti rod, and these waves locally
increase the curvature in the rod, resulting in more fragmented
pieces.
()2
For the elastostatic problem, 7Jfi ::::::: 0, we have
82 {)2u
ax2 [EI {)x2] = q(x),
(13.109)
198
Elasticity 13.5 EulerBernoulli Beam Theory
where q(x) = f(x) is the applied force per unit length. This
equation will be used to determine the deflection of a bean1.
0 Example 19.1: Let us now use the EulerBernoulli theory to
calculate the shape of a heavy cantilever with a uniform cross section
under its own gravity. For a beam under its own gravity, the force is
constant q(x) = pgA per unit length where g is the acceleration due
to the Earth's gravity. If the length of the cantilever is L, then the
total weight is W = pgAL, thus q = ~ . Therefore, we have
d
4
u W
EI dx4 = q = T
where we have use EI =canst. Integrating it twice, we have
cPu q
2
EI dx
2
=
2
x +Ax+B.
At the free end, the beam cannot support bending moment and/or
shear, which implies that M(L) = 0 and V(L) = q*L+A = 0. These
conditions lead to A = qL and B = qL
2
/2. Integrating the above
equation again, we have
du q qL qL
2
EI = x
3
 x
2
+ x+C
dx 6 2 2 ·
As the beam is fixed at x = 0, we have u = 0 and :: = 0 at x = 0.
Thus we have C = 0 from u = 0. Integrating once again, we have
As Ux = 0 at x = 0, we have D = 0. Therefore, the final deflection
curve becomes
The end deflection 8 = u(x = L) is
0
199
Chapter 14
Mathematical Models
What we have discussed so far in terms of differential equa
tions is very limited. In engineering and natural sciences, there
are so many different kinds of phenomena that require both
mathematical modelling and computer simulations as well as
experimental studies. In most cases, the classical models (us
ing the heat conduction equation and the wave equation and
others) are simply not adequate to described these phenomena.
Therefore, we have to broaden our view to study other kinds of
partial differential equations. In fact, mathematical modelling
per se is a subject with vast literature, and subsequently we
have to focus on the relevant equations and to introduce them
very briefly.
14.1 Classic Models
Before we introduce more complicated partial differential equa
tions, let us first remind us the three types of classic partial
differential equations because they are widely used and occur
in a vast range of applications. To a certain extent, almost
all books or studies on the partial differential equations will
have to deal with these three types of basic partial differential
equations.
201
14.1 Classic Models l.!fathematical l.!fodels
14.1.1 Laplace's and Poisson's Equation
In heat transfer problems, the steady state of heat conduction
with a source is governed by the Poison equation
(x,y) En, (14.1)
or
Uxx + Uyy = q(x, y, t), (14.2)
for two independent variables x and y. Here k is thermal diffu
sivity and I ( x, y, t) is the heat source. If there is no heat source
(q = 0), this becomes the Laplace equation. The solution or a
function is said to be harmonic if it satisfies Laplace's equation.
In order to determine the temperature u completely, the ap
propriate boundary conditions are needed. A simple boundary
condition is to specify the temperature u = uo on the bound
ary an. This type of problem is the Dirichlet problem. On
the other hand, if the temperature is not known, but the gradi
ent 8uf 8n is known on the boundary where n is the outward
pointing unit normal, and this forms the Neumann problem.
FUrthermore, some problems may have a mixed type of bound
ary conditions in the combination of au + (3 g ~ = 1', which
naturally occur as a radiation or cooling boundary condition.
14.1.2 Parabolic Equation
Timedependent problems, such as diffusion and transient heat
conduction, are governed by the parabolic equation
(14.3)
Written in thendimensional case Xt = x, x2 = y, X3 = z, ... , it
can be extended to the reactiondiffusion equation
(14.4)
where I is the reaction rate.
202
1\t/athematical 1\t/odels 14.2 Other PDEs
14.1.3 Wave Equation
The vibration of strings and travelling sound waves are gov
erned by the hyperbolic wave equation. The 1D wave equation
in its simplest form is
 2
Utt C Uxx'
(14.5)
where cis the velocity of the wave. Using a transformation of
the pair of independent variables
~ = x+ct, 'f/ = x ct, (14.6)
fort> 0 and oo < x < oo, the wave equation can be written
as
u ~
1 1
= 0. (14.7)
Integrating twice and substituting back in terms of x and t, we
have
u(x, t) = f(x + ct) + g(x ct), (14.8)
where f and g are arbitrary functions of x + ct and x  ct,
respectively. We can see that there are two directions that the
wave can travel. One wave moves to the right and one travels
to the left at a constant speed c.
14.2 Other PDEs
We have shown examples of the three major equations of second
order linear partial differential equations. There are other equa
tions that occur frequently in mathematical physics, engineer
ing and computational sciences. We will give a brief description
of some of these equations.
14.2.1 Elastic Wave Equation
As we have seen in the linear elasticity, the wave in an elas
tic isotropic homogeneous solid is governed by the following
203
14.2 Other PDEs !Ylathematical A1odels
equation in terms of displacement u,
8
2
u
p 8t
2
= J.L\7
2
u +(A+ J.L)\7(\7 · u) + b,
(14.9)
where p is density. A and J.l are Lame constants. b is body force.
Such an equation can have two types of wave: transverse wave
(Swave) and longitudinal or dilatational wave (Pwave).
14.2.2 Maxwell's Equations
The scientific essence of modern wireless communications is
governed by the :Maxwell's equations for electromagnetic waves
8B
VxE= at'
\7 · B = 0,
(14.10)
(14.11)
(14.12)
1 8E
\7 x B = J.LoJ + c
2
7K' (14.13)
where Pe is the charge density, E is the electric field, B is the
magnetic field and J is the current density. eo and J.lo are the
permittivity and permeability of the free space, respectively.
Finally, c is the speed of light. The first equation is the elec
trostatic equation, the second one is the Faraday's law and the
last equation is the AmpereIviaxwell's law.
14.2.3 ReactionDiffusion Equation
The reactiondiffusion equation is an extension of heat conduc
tion with a source
Ut = Dtl.u + f(x, y, z, u), (14.14)
where D is the diffusion coefficient and f is the reaction rate.
One example is the combustion equation
D
Q
>..ju
Ut = Uxx + ue , (14.15)
204
1\t/athematical 1\t/odels 14.2 Other PDEs
where D, Q and A are constants. The other example is the
FitzHughN agumo equations for transport of a nerve signal
Ut = Uxx + u( 1  U) ( U  Q)  V,
Vt = f(U /V),
(14.16)
(14.17)
where /' > 0, 0 < a < 1 and E << 1. These equations are
sometimes also called the equations in excitable media.
14.2.4 FokkerPlank Equation
The time evolution of the probability density function of posi
tion and velocity of a particle system is described by the Fokker
Plank equation
8p n aD!ll n n 82 n ! ~ l
=[L:1 +L:L: lJ ]p,
8t i=l axi i=l i=l 8xi8xi
which can be generalized as
N
8p ="" "" (
1
)k_!__ _!__{D(kJ . }
at L . L. 8Xii ... 8Xik ZI , •••• 1kp '
k=lt}, .... lk .
(14.18)
(14.19)
where D[k] (x
1
, x
2
, ... , xn) are tensors. In the special case of
N = 2, nl
1
1(xt' ... , Xn) is the drift vector, while Dl
2
1 is the
diffusion tensor.
14.2.5 BlackScholes Equation
In the option pricing model, the value of an option u( 8, t) at
time t is governed by the wellknown Nobelwinning Black
Scholes equation
au au 1 2 2 8
2
u
{)t = ru  r 8 a8  2a 8 882 '
(14.20)
where 8 is the current stock price of the underlying stock and
r is the riskfree interest rate. t is the time until option expi
ration. a is the stock volatility or the standard deviation of
205
14.2 Other PDEs !Ylathematical A1odels
stock returns. The BlackScholes equation is very similar to
the extended version of the diffusion equation. The interesting
feature for a call option with an exercise price E and expiry
time T, the change of variables
T = (T t)a
2
/2,
V = uer(Tt)
(14.21)
and
2r
x = ln(S/E) + (
2
1)r,
(J
(14.22)
can transform it into a standard diffusion equation
(14.23)
Then, we can use the standard methods such as integral trans
forms to solve this equation.
14.2.6 Schrodinger Equation
The famous Schrodinger equation is the revolutionary equation
in quantum mechanics and molecular dynamics
·i:_ aw  tt2 t"'72 ,y, u·T·
ZHv ':l'+ ':l'
8t 2m '
(14.24)
where tt is a Planck constant. This equation can be obtained
from the energy formE=::+ U (where pis the momentum)
using differential operator mapping E + itt£ and p + itt V.
W is the probability wave function and U = V ( r)  E is the en
ergy potential. This is a complex partial differential equation.
14.2. 7 NavierStokes Equations
The NavierStokes equations for an incompressible flow can be
written as
V·u=O,
1 2
Ut + ( u · V)u = ReV u  Vp,
206
(14.25)
(14.26)
1\t/athematical 1\t/odels 14.2 Other PDEs
where Re = pUL/ J.l is the Reynolds number. U is the typical
velocity and L is the length scale. p and J.l are the density
of the fluid and its viscosity, respectively. In computational
fluid dynamics, most simulations are mainly related to these
equations.
In the limit of Re << 1, we have the Stokes flow (slow flow)
governed by
(14.27)
In another limit Re >> 1, we have the inviscous flow
V'. u = 0,
Ut + (u · V')u = V'p. (14.28)
We can see that the equations are still nonlinear even in this
simplified case.
14.2.8 SineGordon Equation
Another important equation that appears in a wide range of
applications in physics and many other fields is the SineGordon
equation
Utt Uxx + sin(u) = 0,
which can generally be written as
Utt = Uxx + asin(wu).
This is a nonlinear hyperbolic equation.
(14.29)
(14.30)
Almost all these equations are very difficult for mathemat
ical analysis, and most of them do not have closedform solu
tions under most common boundary conditions. Therefore, the
numerical methods are a good alternative in this case. In the
next few chapters, we will introduce various numerical methods
in details.
207
Chapter 15
Finite Difference
Method
The finite difference method is one of the n1ost popular meth
ods that are used commonly in computer sin1ulations. It has
the advantage of simplicity and clarity, especially in 1D con
figuration and other cases with regular geon1etry. The finite
difference method essentially transforms an ordinary differen
tial equation into a set of algebraic equations by replacing the
continuous derivatives with finite difference approximations on
a grid of mesh or node points that spans the domain of inter
est based on the Taylor expansions. In general, the boundary
conditions and boundary nodes need special treatment.
15.1 Integration of ODEs
The secondorder or higher order ordinary differential equations
can be written as a firstorder systen1 of ODEs. Since the
technique for solving a system is essentially the same as that
for solving a single equation
dy
dx = f(x,y),
(15.1)
209
15.1 Integration of ODEs Finite Difference Metbod
then we shall focus on the firstorder equation in the rest of this
section. In principle, the solution can be obtained by direct
integration,
y(x) =Yo+ {x f(x, y(x))dx,
lxo
(15.2)
but in practice it is usually impossible to do the integration
analytically as it requires the solution of y( x) to evaluate the
righthand side. Thus, some approximations shall be utilized.
Numerical integration is the most common technique to obtain
approximate solutions. There are various integration schemes
with different orders of accuracy and convergent rates. These
schemes include the simple Euler scheme, RungeKutta method,
Relaxation method, and many others.
15.1.1 Euler Scheme
Using the notations h = ~ x = Xn+l  Xn, Yn = y(xn), Xn =
x
0
+ n ~ x (n = 0, 1, 2, ... , N), and ' = djdx for convenience,
then the explicit Euler scheme can simply be written as
1
Xn+l
Yn+l = Yn + f(x, y)dx ~ Yn + hf(xn, Yn)·
Xn
(15.3)
This is a forward difference method as it is equivalent to the
approximation of the first derivative
1 Yn+l Yn
Yn = ~ X
(15.4)
The order of accuracy can be estimated using the Taylor ex
pansion
h
'I h2 "I
Yn+l = Yn + Y n + TY n + ···
~ Yn + hf(xn, Yn) + O(h
2
). (15.5)
Thus, the Euler method is first order accurate.
For any numerical algorithms, the algorithm must be sta
ble in order to reach convergent solutions. Thus, stability is
an important issue in numerical analysis. Defining 8y as the
210
Finite Difference Method 15.1 Integration of ODEs
discrepancy between the actual numerical solution and the true
solution of the Euler finite difference equation, we have
(15.6)
In order to avoid the discrepancy to grow, it requires the fol
lowing stability condition 1 ~ 1 ::; 1. The stability restricts the
size of interval h, which is usually small. One alternative
that can use larger h is the implicit Euler scheme, and this
scheme approximates the derivative by a backward difference
y ~ = (Yn Ynd/h and the righthand side of equation (15.2)
is evaluated at the new Yn+l location. Now the scheme can be
written as
(15.7)
The stability condition becomes
(15.8)
which is always stable if f'(y) = U ::; 0. This means that
any step size is acceptable. However, the step size cannot be
too large as the accuracy reduces as the step size increases.
Another practical issue is that, for most problems such as non
linear ODEs, the evaluation of y' and f'(y) requires the value of
Yn+l which is unknown. Thus, an iteration procedure is needed
to march to a new value Yn+l, and the iteration starts with a
guess value which is usually taken to be zero for most cases.
The implicit scheme generally gives better stability.
0 Example 15.1: To solve the equation
dy
dx = f(y) = ey y,
we use the explicit Euler scheme, and we have
211
15.1 Integration of ODEs Finite Difference Method
Suppose the discrepancy between real solution and the numerical
Yn is 8yn so that = Yn + 8yn, then the real solution satisfies
= +
Since = f(Yn) + the above equation becomes
Yn+l + 8yn+l = Yn + 8yn + h[f(Yn) + J'(Yn)8yn]·
Together with the Euler scheme, we have
8yn+l = 8yn + J'8Yn·
Suppose that 8yn ex en, then we have
en+l =en+ h/'en' or
e = 1 + hf'.
In order for the scheme to be stable (or en 0), it requires that
or  1 1 + hf' = 1  h(eYn + 1) 1.
The stability condition becomes
  ey, + 1
15.1.2 LeapFrog Method
The Leapfrog scheme is the central difference
1 Yn+1  Yn1
Yn = 2Ax '
which leads to
Yn+1 = Yn1 + 2hf(xn, Yn)·
0
(15.9)
(15.10)
The central difference method is second order accurate. In a
similar way as equation (15.6), the leap frog method becomes
8yn+1 = 8Yn1 + 2hf'(y)8yn, (15.11)
or
8yn+1 =
(15.12)
where
= 1 + This scheme is stable only if ::; 1,
and a special case is = 1 when f'(y) is purely in1aginary.
Therefore, the central scheme is not necessarily a better scheme
than the forward scheme.
212
Finite Difference Method 15.2 Hyperbolic Equations
15.1.3 RungeKutta Method
We have so far seen that stability of the Euler method and the
central difference method is limited. The RungeKutta method
uses a trial step to the midpoint of the interval by central dif
ference and combines with the forward difference at two steps
Yn+l/2 = Yn + .;. f(xn, Yn),
(15.13)
(15.14)
This scheme is second order accurate with higher stability com
pared with previous shnple schemes. One can view this scheme
as a predictorcorrector method. In fact, we can use multisteps
to devise higher order methods if the right combinations are
used to eliminate the error terms order by order. The popular
classical RungeKutta method can be written as
a= hf(xn, Yn),
b = hf(xn + h/2, Yn + a/2),
C = hf(xn + h, Yn + b/2),
d = hf(xn + h,yn +c),
a+ 2(b +c)+ d
Yn+l = Yn +
6
'
(15.15)
which is fourth order accurate. Generally speaking, the higher
order scheme is better than the lower scheme, but not always.
15.2 Hyperbolic Equations
Nun1erical solutions of partial differential equations are more
complicated than that of ODEs because it involves time and
space variables and the geometry of the domain of interest.
Usually, boundary conditions are more complex. In addition,
nonlinear problems are very common in engineering applica
tions. Now we start with the simplest first order equations and
then move onto more complicated cases.
213
15.2 Hyperbolic Equations Finite Difference Metbod
15.2.1 FirstOrder Hyperbolic Equation
For simplicity, we start with the onedimensional scalar equa
tion of hyperbolic type,
(15.16)
where c is a constant or the velocity of advection. By using the
forward Euler scheme for time and centeredspaced scheme, we
have
 u"! un un
J J + c[ i+1; j1] = 0,
(15.17)
where t = n = 0, 1, 2, ... , x = x
0
+ jh,j = 0, 1, 2, ... , and
h = In order to see how this method behaves numerically,
we use the von Neumann stability analysis.
Assun1ing the independent solutions or eigenmodes (also
called Fourier modes) in spatial coordinate x in the form of
uj = and substituting into equation (15.17), we have
ikh(j+1) ihk(j1)
1 ikhj  ikhj = c e  e
e h 2 .
(15.18)
Dividing both sides of the above equation by exp( ikhj) and
using sinx = (eix eix)/2i, we get
= 1 sin(kh). (15.19)
The stability criteria 1 require
(
)2 . 2 kh 0
h Sill .
(15.20)
However, this inequality is impossible to satisfy and this scheme
is thus unconditionally unstable.
To avoid the difficulty of instability, we can use other schemes
such as the upwind scheme and Lax scheme. For the upwind
scheme, the equation becomes
 u"! u"!  u"!
1
1 1
+ c[
1
h
1
 ] = 0,
(15.21)
214
Finite Difference Method 15.2 Hyperbolic Equations
whose stability condition is
=II [1 cos(kh) + isin(kh)]l 1,
which is equivalent to
(15.22)
(15.23)
This the wellknown CourantFriedrichsLewy stability condi
tion, often referred to as the Courant stability condition. Thus,
the upwind scheme is conditionally stable.
15.2.2 SecondOrder Wave Equation
Higher order equations such as the secondorder wave equation
can be written as a system of hyperbolic equations and then be
solved using numerical integration. They can also be solved by
direct discretization using a finite difference scheme. The wave
equation
(15.24)
consists of second derivatives. If we approximate the first
derivatives at each time step n using
u' uft1 uf
i
, uf uf1
ui1 =
(15.25)
then we can use the following approximation for the second
derivative
"
U· =......;........;........;.
t
uf+1 2uf + uf1
(15.26)
This is in fact a central difference scheme of second order accu
racy. If we use the similar scheme for timestepping, then we
get a central difference scheme in both time and space.
215
15.3 Parabolic Equation Finite Difference Method
Thus, the numerical scheme for this equation becomes
ui+1 2uf + ui1 2 uf+1 2uf + uf1
= c
(15.27)
This is a twolevel scheme with a second order accuracy. The
idea of solving this difference equation is to express (or to solve)
ui+
1
at time step t = n+ 1 in terms of the known values or data
uf and ui
1
at two previous time steps t = n and t = n  1.
15.3 Parabolic Equation
For the parabolic equation such as the diffusion or heat con
duction equation
8u =
8t 8x 8x '
(15.28)
a simple Euler method for the time derivative and centered
secondorder approximations for space derivatives lead to
(15.29)
The stability requirement ::; 1 leads to the constraint on the
time step (see the example),
(15.30)
This scheme is shown in Figure 15.1 and it is conditionally sta
ble.
0 Example 15.2: From equation (15.29), we can apply the von
Neumann stability analysis by assuming uj = f.neikhj, we have
Dividing both sides by f.neikhj, we have
Dt::..t ·kh ·kh
f. = 1 + """h2"[e
1
+ ez  2].
216
Finite Difference Method 15.3 Parabolic Equation
t
Q
t=n+1
I
I
I
0 6
Ot=n
j1 j j+1
x·
J
Figure 15.1: Central difference in space and explicit Euler time
stepping.
Using cosx = (eix + eix)/2 and sin
2
(x/2) = (1 cosx)/2, we obtain
4DD.t .
2
kh
Since sin( X) 1' thus e 1 requires
4DD.t
1 < 1 < 1.
 h2  0
or
0
A typical feature of a solution to the diffusive system is
that the profile is gradually smoothed out as time increases.
The timestepping scheme we used limits the step size of time
as larger time steps will make the scheme unstable. There
are 1nany ways to improve this, and one of most widely used
schemes is the implicit scheme.
To avoid the difficulty caused by very small timesteps, we
now use an implicit scheme for time derivative differencing, and
thus we have
DL).t
 uf! = + +
J J h2 J+l J J1 .
Applying the stability analysis, we have
1
= 1 + 4D.6.t · 2 kh'
fiT""sm
2
217
(15.31)
(15.32)
15.4 Elliptical Equation Finite Difference !v!etbod
whose norm is always less than unity ( 1 ~ 1 ~ 1). This means
the implicit scheme is unconditionally stable for any size of
time steps. That is why in1plicit methods are more desired in
simulations. However. there is one disadvantage of this method,
which requires more programming skills because the inverse of
a large matrix is usually needed in implicit schen1es.
15.4 Elliptical Equation
In the parabolic equation, if the time derivative is zero or u
does not change with time Ut = 0, then it becomes a steady
state problem that is governed by the elliptic equation. For
the steady state heat conduction problem, we generally have
the Poisson problem,
V · [K(u, x, y, t)Vu] = f, (15.33)
If K is a constant, this becomes
f
q=.
K
(15.34)
There are many methods available to solve this problems such
as the boundary integral method, the relaxation method, and
the multigrid method. Two major ones are the longtime ap
proximation of the transient parabolic diffusion equations, the
other includes the iteration method.
The long time approximation method is essentially based
on fact that the parabolic equation
(15.35)
evolves with a typical scale of ../'Kl. If VKJ, >> 1, the system is
approaching its steady state. Assuming t + oo and K >> 1, we
then have
2 f 1
V U =   Ut + 0.
K K
(15.36)
218
Finite Difference Method 15.4 Elliptical Equation
The longtime approximation is based on the fact that the
parabolic equation in the case of K = const degenerates into the
above steadystate equation (15.33) because Ut + 0 as t + oo.
This approximation becomes better if K >> 1. Thus, the usual
numerical methods for solving parabolic equations are valid.
However, other methods may obtain the results more quickly.
The iteration method uses the secondorder scheme for space
derivatives, and equation (15.34) in the 2D case becomes
Ui+l,j  2Ui,j + Uil,j Ui,j+l  2ui,j + Ui,j1
(Ax)2 + (Ay)2 = q. (15.37)
If we use Ax = Ay = h, then the above equation simply be
comes
(15.38)
which can be written as
Au=b. (15.39)
In principle, one can solve this equation using Gauss elimina
tion; however, this becomes impractical as the matrix becomes
large such as 1000 x 1000. The GaussSeidel iteration provides
a more efficient way to solve this equation by splitting A as
A=L+D+U, (15.40)
where L, D ~ U are the lower triangle, diagonal and upper tri
angle matrices of A ~ respectively. The iteration is updated in
the following way:
(15.41)
This procedure stops until a prescribed error or precision is
reached.
219
Chapter 16
Finite Volume Method
16.1 Introduction
The finite difference method discussed in the previous chapter
approximates the ordinary differential equations and partial
differential equations using Taylor series, resulting in a system
of algebraic equations. The finite volume method resembles
the finite difference method in certain ways but the starting
point is the integral formulation of the problem. It uses the
integral form of the partial differential equations in terms of
conservation laws, then approximates the surface and boundary
integrals in the control volumes. This becomes convenient for
problems involving flow or flux boundaries.
For a hyperbolic equation that is valid in the domain n with
boundary an'
(16.1)
or written in terms of flux function F = F(u) = K.'\Ju, we
have
8u
at + v · F = q.
(16.2)
The integral forn1 of this equation becomes
In ';;: dQ + In V' . F = In qdf!.
(16.3)
221
16.2 Elliptic Equations Finite Volume !v!etbod
If the integral form is decomposed into many small control vol
umes, or finite volumes, n = U ~
1
ni and ni n ni = 0. By
defining the control volume cell average or mean value
(16.4)
where \ti = IOil is the volume of the small control volume ni,
the above equation can be written as
(16.5)
By using the divergence theorem
fv \7 · F = £ F · ndA,
(16.6)
we have
(16. 7)
where dS = ndA is the surface element and n is the outward
pointing unit vector on the surface ri enclosing the finite vol
ume ni. The integration can be approximated using various
nun1erical integration schemes. In the sin1plest 1D case with
h = ~ x , the integration
11(i+l/2)h
ui = udx,
h (i1/2)h
(16.8)
is a vertexcentred finite volume scheme. In the following sec
tions, we will discuss the three major types of partial differ
ential equations (elliptic, parabolic and hyperbolic) and their
finite volume discretizations.
16.2 Elliptic Equations
Laplace's equation is one of the most studied elliptic equations
\7
2
u(x, y) = 0, (x, y) E 0, (16.9)
222
Finite Volume l'vlethod 16.3 Parabolic Equations
its integral form becomes
'\7 udO. =  · dS = 0.
1
2 1 8u
n r 8n
(16.10)
For the simple regular grid points the control vol
ume in this case is a cell centred at ( j with a size of
(along xaxis) and (along yaxis), and the boundary
integral on any cell consists of four parts integrated on each
of the four sides. By using the simple approximation with
= (ui+l,j and t = (ui,j+l we have
1
8u

8
dO.= A (ui+l.i + Uit,j 2ui,j)
f2·. n uX
t,]
+ (ui,j+l + ui,j1  2ui,j) = 0.
(16.11)
Dividing both sides with and letting = = h, we
obtain
(ui+l,j + Ui.j+l + Uil,j + Ui,jt) 4Ui,j = 0,
(16.12)
which resembles finite difference methods in many ways. In
fact, this is exactly the Laplace operator for a 5point differ
encing scheme.
16.3 Parabolic Equations
For the case of heat conduction
(16.13)
we have its integral form
[ [ {)u 8
2
u
lt ln ( 8t  k {)x2  q)dxdt = 0.
(16.14)
223
16.4 Hyperbolic Equations Finite Volume !v!etbod
If we use the control volume from ( i 1 /2)h to ( i + 1 /2)h where
h = D.x, and with time from step n to n + 1, we have
{(n+1)At 1(i+1/2)h {Ju fJ2u
lnAt (i1/2)h ( {)t  k 8x2  q)dxdt = O.
(16.15)
By using the midpoint approximation
l '1/J(x)dx = '1/J[(a; b)](b a),
(16.16)
and the DuFortFranke! scheme where we first approximate the
gradient
[)2u uf+1  2uf + uf1
{)x2 = h2
(16.17)
then replace 2uf with (uj+
1
+ uj
1
), we have
u ~ + 1  u ~  1
t t
=
2D.t
(16.18)
where we have used the central scheme for time as well. The
finite volume scheme is more versatile in dealing with irregular
geometry and more natural in applying boundary conditions.
Following the stability analysis, we get 1 ~ 1 < 1 is always true
and thus the DufortFranke! scheme is unconditionally stable
for all .Dot and D.x.
16.4 Hyperbolic Equations
For the hyperbolic equation of the conservation law in the one
dimensional case
{)u 8w(u) _
0
{Jt + ax  '
(16.19)
we have its integral form in the fixed domain
dx= udx
1
Xb {)u {) 1Xb
Xa {Jt {Jt Xa
224
Finite Volume l'vlethod 16.4 Hyperbolic Equations
= {w[u(xb)] w[u(xa)]} = 0. (16.20)
If we use the midpoint rule u* to approximate the integral, we
have
(16.21)
If we choose the control volume [(i (i + cen
tred at the mesh point Xi = = ih with the approximation
ui ui in each interval, and using the forward differencing
scheme for the time derivative, we have
(16.22)
By further approximation of the flux \ll(xi+l/2) \ll(xi), we
have the upward scheme
(16.23)
which is conditionally stable as we know this in the finite dif
ference method. For the simplest flux w(u) = cu, we have
n+l n ( n n )
u = ui  h ui  ui1 '
and its stability requires that
0 < 1
< h  .
225
(16.24)
(16.25)
Chapter 17
Finite Element Method
In the finite difference method, we approximate the equations
at a finite number of discrete points, and there are many limi
tations in finite difference methods. One of such disadvantages
is that it is not straightforward to deal with irregular geometry.
lviore versatile and efficient n1ethods are highly needed. In fact,
the finite element method is one class of the most successful
methods in engineering and have a wide range of applications.
The basic aim of the finite element method is to formulate
the numerical method in such a way that the partial differential
equation will be transformed into algebraic equations in terms
of matrices. For timedependent problems involving partial dif
ferential equations, the equations can be converted into an ordi
nary differential equation, which will in turn be discretized and
converted into algebraic equations by timestepping or some it
eration techniques. For example, a linear elastic problem can
be formulated in such a way that it is equivalent to the equation
of the following type
Ku=f, (17.1)
where K is the stiffness matrix, and f is a vector corresponding
to nodal forces and some contribution from boundary condi
tions. u is the unknown vector to be solved and it corresponds
to the nodal degree of freedon1 such as the displacement.
227
17.1 Concept of Elements Finite Element Metbod
17.1 Concept of Elements
17.1.1 Simple Spring Systems
The basic idea of the finite element analysis is to divide a model
(such as a bridge and an airplane) into many pieces or ele
ments with discrete nodes. These elements form an approxi
mate system to the whole structures in the domain of interest,
so that the physical quantities such as displacements can be
evaluated at these discrete nodes. Other quantities such as
stresses, strains can then be evaluated at certain points (usu
ally Gaussian integration points) inside elements. The simplest
elements are the element with two nodes in 1D, the triangular
element with three nodes in 2D, and tetrahedral elements with
four nodes in 3D.
In order to show the basic concept, we now focus on the
simplest 1D spring element with two nodes (see Figure 17.1).
The spring has a stiffness constant k (Njm) with two nodes i
and j. At nodes i and j, the displacements (in metres) are Ui
and Uj, respectively. /i and /j are nodal forces.
k
Figure 17.1: Finite element concept.
From Hooke's law, we know the displacement ~ u = Uj  Ui
is related to f, or
f = k ( ~ u ) . (17.2)
At node i, we have
(17.3)
and at node j, we get
(17.4)
228
Finite Element 1\lethod 17.1 Concept of Elements
These two equations can be combined into a matrix equation
or Ku = f. (17.5)
Here K is the stiffness matrix, u and f are the displacement
vector and force vector, respectively. This is the basic spring
element, and let us see how it works in a spring system such
as shown in Figure 17.2 where three different springs are con
nected in series.
Figure 17.2: A simple spring system.
For a simple spring system shown in Figure 17.2, we now
try to determine the displacements of Ui ( i = 1, 2, 3, 4). In order
to do so, we have to assemble the whole system into a single
equation in terms of global stiffness matrix K and forcing f.
As these three elements are connected in series, the assembly
of the system can be done element by element. For element E1,
its contribution to the overall global matrix is
(
k1 k1 ) ( Ul ) = ( fi ) '
k1 k1 U2 /2
(17.6)
which is equivalent to
(17.7)
where
(17.8)
229
17.1 Concept of Elements Finite Element Metbod
and !'f
1
= (ft, /2, 0, 0). Similarly, for element E2, we have
( k2 k2 ) ( U2 ) = ( /2 ) '
k2 k2 U3 /3
(17.9)
or
KFU
0 0
!}
k2 k2
k2 k2
0 0
(17.10)
where we have used the balance at node 2. For element E3, we
have
(17.11)
or
(17.12)
where /4 = /. has been used. We can now add the three sets
of equations together to obtain a single equation
( kl
k2
0 0 ) ( UJ ) ( !J )
t
kt + k2 k2 0 U2 /2+/2
k2 k2 + k3 k3 U3 =  j3 + /3 '
0
k3 k3 U4 f*
or
Ku=f, (17.13)
where
K = Kt +K2 +K3
( kl
kl 0
_t}
= T
kt + k2 k2
(17.14)
k2 k2 +k3
0 k3 k3
230
Finite Element 1\lethod 17.1 Concept of Elements
and
(17.15)
In general, the matrix K is singular or its rank is less than the
total number of degrees of freedom, which is four in this case.
This means that the equation has no unique solution. Thus, we
need the boundary conditions to ensure a unique solution. In
this spring system, if no boundary condition is applied at any
nodes, then the applied force at the node 4 will make the spring
system fly to the right. If we add a constraint by fixing the left
node 1, then the system can stretch, and a unique configuration
is formed.
In our case where there are no applied forces at nodes 2 and
3, we have
(17.16)
0 Example 11.1: For k
1
= 100 Njm, k
2
= 200 N/m, and k
3
=
50N/m, and J. = 20 N, the boundary at node 1 is fixed (u
1
= 0).
Then, the stiffness matrix is
and the force column vector
100
300
200
0
0
200
250
50
fT = (0, 0, 0, 20).
_t )'
50
The rank of K is 3, therefore, we need at least one boundary condi
tion. By applying u
1
= 0, we now have only three unknown displace
ments u
2
, u
3
, u
4
• Since u
1
= 0 is already known, the first equation for
Ut becomes redundant and we can now delete it so that the reduced
stiffness matrix A is a 3 x 3 matrix. Therefore, we have
(
300
A=  ~ 0 0
and the reduced forcing vector is
200
250
50
gT = (0, 0, 20).
231
~ ),
50
17.1 Concept of Elements Finite Element Method
The solution is
(
0.2)
u =A lg = 0.3 .
0.7
Therefore, the displacements are u
2
= 0.2 m, u
3
= 0.3 m, and u
4
=
0.7 m.
Theoretically speaking, the force should be 20N eve1ywhere in
the spring systems since the mass of the springs is negligible. Let us
calculate the force at nodes 2 and 3 to see if this is the case. At the
node 2, the extension in element E
1
is ~ u = u
2
 u
1
= 0.2 m, thus
the force at node 2 is
f2 = k t ~ U = 100 X 0.2 = 20N.
Similarly, at node 3 of element E2, we have
which is the same at node 3 of element E
3
So the force is 20 N everywhere. 0
17.1.2 Bar and Beam Elements
The spring system we discussed earlier is limited in many ways
as a spring does not have any mass and its cross section is not
explicitly included. A more complicated but realistic elen1ent
is the bar element as shown in Figure 17 .3, which is a uniform
rod with a cross section area A, Young's elastic modulus E,
and a length L. A bar element can only support tension and
compression, it cannot support bending. For this reason, it is
also called a truss element.
The displacements at nodes i and j are Ui and Uj, respec
tively. The forces at the corresponding nodes are fi and fi.
Now we have to derive its stiffness matrix. Assuming the bar
is linearly elastic, the stress a is thus related to strain f via
232
Finite Element .Atfethod 17.1 Concept of Elements
a= EE. Since f = (uj ui)/L and a= f /A where F is the
force in the bar element, we have
EA
f = y ( ~ u ) = k ( ~ u ) , (17.17)
where ~ u = Uj  Ui is the extension or elongation of the bar
element. Now the equivalent spring stiffness constant is
k
= EA
L"
Therefore, the stiffness matrix K for this bar becomes
K = ( k k ) = EA ( 1 1 ) .
k k L 1 1
u·
t
A,E,L
Figure 17.3: Bar element.
j
u·
J
(17.18)
(17.19)
We have up to now only discussed 1D systems where all
displacements Ui or Uj are along the bar direction, and each
node has only one displacement (one degree of freedom). We
now extend to study 2D syste1ns. In 2D, each node i has two
displacements ui (along the bar direction) and vi (perpendicu
lar to the bar direction). Thus, each node has two degrees of
freedom.
If we rotate the bar elen1ent by an angle (} as shown in
Figure 17.4, we cannot use the standard addition to assemble
the system. A transformation is needed between the global
coordinates (x, y) to the local coordinates (x', y'). From the
geometrical consideration, the global displacements ui and vi
233
17.1 Concept of Elements Finite Element Metbod
at node i are related to the local displacement and (usually)
= 0.
( ) = ( :: ) ( ).
(17.20)
y
Figure 17.4: 2D transformation of coordinates.
Using the similar transformation for Uj and Vj, we get the
transformation for the twonode bar ele1nent
(
sinO 0
(:}
I
ro
cosO 0
U=
0 cosO
0 sinO COS 0 Vj
which can be written as
u' = Ru, (17.21)
where
R== (
cosO sinO 0
sinO cosO 0
0 0 cosO
0 0 sinO cosO
(17.22)
The same applies to transform the force,
f' = Rf, (17.23)
234
Finite Element 1\lethod 17.2 Finite Element Formulation
and the stiffness matrix in local coordinates is
K'u' = f'. (17.24)
As the calculation is n1ainly based on the global coordinates,
and the assembly should be done by transforming the local
systems to the global coordinates, by combining the above two
equations, we have
K'Ru=Rf,
or
R
1
K'Ru = Ku = f,
which is equivalent to a global stiffness matrix
The stiffness matrix K is a 4 x 4 matrix in 2D.
(17.25)
(17.26)
(17.27)
Bar elements can only elongate or shrink, they do not sup
port bending or deflection. For bending, we need the beam ele
ments which include a rotation around the end nodes ()i and ()i.
In this case, each node has three degrees of freedom ( Ui, Vi, Oi),
and the stiffness matrix is therefore a 6 x 6 matrix in 2D.
For more complicated elements, it is necessary to use a formal
approach in terms of shape functions and weak formulations.
17.2 Finite Element Formulation
17.2.1 Weak Formulation
l\llany problems are modelled in terms of partial differential
equations, which can generally be written as
.C( u) = 0, X E !l, (17.28)
where .C is a differential operator, often linear. This prob
lem is usually completed with the essential boundary condition
£(u) = (u u) = 0 for X E {)!lE and natural boundary condi
tions B(u) = 0 for X E anN. Assuming that the solution can
235
17.2 Finite Element Formulation Finite Element !v!etbod
be approximated by uh over a finite element mesh with an av
eraged element size or mean distance h between two adjacent
nodes, the above equation can be approximated as
(17.29)
Multiplying both sides of the equation by a test function or
a proper weighting function, integrating over the domain and
using associated boundary conditions, we can write the general
weak formulation of Zienkiewicztype as
fn C( uh)widn
+ f B(uh)widr + f &(uh)widrE ~ o, (17.30)
lanN lanE
where (i = 1, 2, ... , M). If we can approximate the solution uh
by the expansion
M M
uh(u, t) = L ui(t)Ni(x) = L UjlVj, (17.31)
i=l j=l
it requires that Ni = 0 on 8f2E so that Wi = 0 on {}O.E. Thus,
only the natural boundary conditions are included since the
essential boundary conditions are automatically satisfied. In
addition, there is no much limitation on the choice of Wi and
wi. If we choose wi = wi so as to simplify the formulation,
we have
(17.32)
17 .2.2 Galer kin Method
There are many different ways to choose the test functions Wi
and shape functions Ni. One of the most popular methods
is the Galerkin method where the test functions are the same
as the shape functions, or Wi = Ni. In this special case, the
formulation simply becomes
(17.33)
236
Finite Element 1\lethod 17.2 Finite Element Formulation
The discretization of this equation will usually lead to an alge
braic matrix equation.
On the other hand, if we use the Dirac delta function as
the test functions Wi = 8 ( x  ~ ) , the method is called the
collocation method which uses the interesting properties of the
Dirac function
fo f(x)8(x Xi)dn = f(xi)· (17.34)
together with 8(x Xi) = 1 at x = Xi and 8(x Xi) = 0 at
xfXi·
17.2.3 Shape Functions
The n1ain aim of the finite elen1ent method is to find an ap
proxinlate solution uh(x, t) for the exact solution u on some
nodal points,
M
uh(x, t) = L Ui(t)Ni(x)
i=l
(17.35)
where ui are unknown coefficients or the value of u at the dis
crete nodal point i. Functions Ni (i = 1, 2, ... , M) are linearly
independent functions that vanish on the part of the essential
boundary. At any node i, we have Ni = 1, and Ni = 0 at any
other nodes, or
M
LNi = 1, Ni(Xj) = 8ij·
(17.36)
i=l
The functions Ni ( x) are referred to as basis functions, trial
functions or more often shape functions in the literature of
finite element methods. For the simplest 1D element with two
nodes i and j, the linear shape functions can be written as
(17.37)
where L = lxi xil, which is shown in Figure 17.5.
237
17.2 Finite Element Formulation Finite Element !v!etbod
j
~ = xflxi xil
Figure 17.5: The 1D linear shape functions.
X
Using the assumptions that ui(t) does not depend on space
and lVi(x) does not depend on time, the derivatives of u can
be approximated as
M
8u 8uh """' 1
a ~ fJx = ~ Ui(t)N (x),
X i=l
M
. OUh """' . N( )
u ~ ~ = ~ u i x,
vt i=l
(17.38)
where we have used the notations:
1
= dj dx and· = 3t. Higher
order derivatives are then calculated in a similar way. The
ultimate goal is to construct a method of computing ui such
that the error uh  u is minimized. Generally speaking, the
residual R varies with space and time, so we have
(17.39)
There are several methods to minimize R. Depending on the
scheme of minimization and the choice of shape functions, var
ious methods can be formulated. These include the weighted
residual method, the method of least squares, the Galerkin
method and others.
238
Finite Element 1\lethod 17.3 Elasticity
17.3 Elasticity
17.3.1 Plane Stress and Plane Strain
The stress tensor u and strain tensor e are not written as tensor
forms but vector forms u = (axx, ayy, a:;::, axy, ay::, a::x)T, and
€ = (f.xx' f.yy, f.:;:;, "fxy, 'Yyz, 'Yzx)T. The strain tensor is usually
defined as
1 oui 8u·
cij = ( + _J ),
2 8xj 8xi
(17.40)
where one applies the engineering shear strain f.xy = 2cxy·
Hooke's elasticity can be expressed as
(17.41)
where Dis a 6 x 6 symmetrical matrix as functions of Young's
modulus E and Poisson's ratio v.
Two special cases that are commonly found in many appli
cations are the plane stress (a zz = 0, but f.z:; =/= 0) and plane
strain (f.:::: = 0, but azz =/= 0). The commonly used formulation
is the displacementbased formulation or ubased formulation.
In the 2D case, the displacen1ent u = (u, v)T and the strain f.
and stress a are defined as
(17.42)
Now the stressstrain relationship becomes
a= D(f. f.o), (17.43)
where D is a 3 x 3 matrix. The strains are given by
au 8v au 8v
f.x=, f.y= l'xy=+,
ox 8y ' {)y OX
(17.44)
where f.o is the initial strain due to temperature change or ther
mal loading. If there is no such change, then the initial strain
can be taken to be zero in most applications.
239
17.3 Elasticity Finite Element Metbod
The equilibrium of force in elasticity leads to
V' · u+ b = 0. (I7.45)
where b = [fx /y]T is the body force. In the case of plane
stress, we have
E ( I v 0 )
D=
2
v I 0 .
I  v 0 0 (I  v)/2
(I7.46)
In the case of plane strain, we have
E
(
~ ~ ~ i ~ v 0 )
v 1v 0
D = (I  2v) l ~ v l ~ v (12v) .
2(1+v)
(I7.47)
Clearly, for the ID case plane stress when v = 0, /y = 0 and
uy = Txy = 0, the equation of force balance simply becomes
E 8
2
u
I2 8 2 + fx = 0,
v x
(I7.48)
where we have used the stressstrain relationship. This ID
equation is essentially the same as the ID heat transfer equa
tion u" + Q = 0 that will be discussed in detail later, so the
solution technique for the ID heat transfer shall equally ap
ply. Therefore, we shall focus on the 2D case in the rest of
this section.
Displacements ( u, v) in a plane eletnent can be interpolated
from nodal displacements. For example, using a triangular ele
ment (i,j, m) with three nodal points as shown in Figure I7.6
(xi, Yi), (xj, Yi), and (xm, Ym), we have
u = (:) = (N;I,N;I,Nml]
=Nd, (I7.49)
240
Finite Element 1\lethod 17.3 Elasticity
y
X
Figure 17.6: Schematic triangular mesh and the layout of a
triangular element.
where I is a 2 x 2 unitary matrix, i.e.,
and
N = ( N
0
i 0
Ni
N·
J
0
By defining a differential operator
(
fx 0 )
Ld= 0 iu,
{) w
i)y 'ljX
we can rewrite the above forn1ulation as
where
B =LdN.
Now the equation (17.45) becomes
where
Ku=f,
K = In BTDBdV,
241
(17.50)
(17.51)
(17.52)
(17.53)
(17.54)
(17.55)
(17.56)
17.3 Elasticity Finite Element Metbod
and
(17.57)
where r is the surface traction (force per unit area).
17 .3.2 Implementation
In the case of 2D elastic problems, the simplest elen1ents are
linear triangular elements, thus we have Bi = LdNi, or
1 c;
0 b· 0
bm
J
B = LdN =
0 Ci
0 c· 0
J
c· b· c· b·
Cm
bm
l l
J J
(17.58)
where
1
1
Xi
Yi
1 x·
Yi
2
J
1
Xm
Ym
(17.59)
is the area of the triangular element. The linear elen1ent im
plies that the strain is constant in the element. The stiffness
matrix (i,j = 1, 2, ... , 6) of each triangular element can be
expressed as
(17.60)
For a triangular element with three nodes (i,j, m), each node
has two degrees of freedom (ui, Vi), then the stiffness matrix
K(e) for each element is a 6 x 6 matrix. In general, if each
element has r nodes, and each node has n degrees of freedom,
then the local stiffness matrix is a rn x rn matrix. If the region
has M nodes in total, then M n equations are needed for this
problem. For the present case (n = 2), we need 2/v/ equations
for plane stress and plane strain.
In order to calculate the contribution of each element to
the overall (global) equation, each node should be identified in
some way, and most often in terms of the index matrix. The
242
Finite Element 1\lethod 17.3 Elasticity
nodal index matrix for three nodes ( i, j, m) can be written as
(
(i, i) (i,j) (i, m) )
IDnode = (j, i) (j,j) (j, m) .
(m, i) (m,j) (m, m)
(17.61)
The nodal index matrix identifies the related nodes in the global
nodenumbering system. However, the main assembly of the
stiffness matrix is about the corresponding equations and the
application of the boundary conditions, thus we need to transfer
the nodal index matrix to the equation index matrix in terms of
the global numbering of equations with two degrees of freedom
for each node (e.g., equations 2i 1 and 2i for nodal i, equation
2j  1 and 2j for node j, etc). Thus, for each entry in the
stiffness matrix, say (i,j), in the nodal index matrix, we now
have four entries, i.e.,
(
.. ) ~ ( (2i 1,2j 1) (2i 1,2j))
Z,J (2i, 2j  1) (2i, 2j) '
etc (17.62)
The equation index matrix now has 6 x 6 entries, and each
entry is a pair such as (2i 1, 2j 1), ... , (2m, 2m), etc. By
writing it as two index matrices (ID = JDT), we now have
IDequ = J D ~ u ,
2i 1 2i 2j 1 2j 2m1 2m
2i 1 2i 2j 1 2j 2m1 2m
JDequ =
2i 1 2i 2j 1 2j 2m1 2m
(17.63)
2i 1 2i 2j 1 2j 2m1 2m
2i 1 2i 2j 1 2j 2m1 2m
2i 1 2i 2j 1 2j 2m1 2m
So that the contribution of K ~ ) to the global matrix Kij is
sin1ply
K  K + K(e)
(J Dequ(/.J),J Dequ(/,J)]  (J Dequ(I,J),JDequ(/.J)] (/.J)'
I,J = 1,2, ... ,6. (17.64)
243
17.4 Heat Conduction Finite Element Metbod
Similarly, the contribution of the body force and external force
can be computed
J(l) = f(l) + j(e)(l)
where l = 2i 1, 2i, 2j 1, 2j, 2m 1, 2m etc.
17.4 Heat Conduction
Heat transfer problems are very common in engineering and the
geometry in most applications is irregular. Thus, finite elen1ent
methods are especially useful in this case.
17 .4.1 Basic Formulation
The steadystate heat transfer is governed by the heat conduc
tion equation or Poisson's equation
V · (kVu) + Q = 0, (17.65)
with the essential boundary condition
u =u, x E anE, (17.66)
and the natural boundary condition
(17.67)
Using the formulation similar to the formulation (17.33) in
terms of u ~ uh, we have
f [V · (kVu) + Q]1VidO. f [k
8
8
u q]lVidr = 0. (17.68)
ln lanN n
Integrating by parts and using Green's theorem, we have
244
Finite Element !vletbod 17.4 Heat Conduction
+ f Q1Vidn f [k
8
8
uh  q]JVidr = o.
ln lanN n
(17.69)
Since lVi = 0 on {)O.E, thus we have
f [ ]Nidr = f [ ]Nidr.
lan lanN
(17.70)
Therefore, the above weak formulation becomes
Substituting uh = E j \ 1 ~
1
UjlVj(x) into the equation, we have
 [ qNidr = 0.
lanN
This can be written in the compact matrix form
M
L KijUj = fi, KU = f,
j=l
(17.72)
(17.73)
where K = [Kij], (i,j = 1, 2, ... , M), uT = (ut, U2, ... , UM ), and
cT = (ft, !2, ... , !M ). That is,
Kii = In k"V Ni "V lVjdO.,
(17.74)
fi = f QNidn + f qNidr.
ln lanN
(17.75)
0 Example 17.2(a): As a simple example, we consider the 1D
steadystate heat conduction problem,
u"(x) + Q(x) = 0,
witb boundazy conditions
u(O) = /3, u'(l) = q.
245
17.4 Heat Conduction Finite Element Method
For a special case Q(x) = rexp(x), we have the analytical solution
u(x) = ({3 r) + (re
1
+ q)x +rex. (17.76)
Then equation (17. 75) becomes
For the purpose of demonstrating the implementation procedure, let
us solve this problem by dividing the interval into 4 elements and 5
nodes. This will be discussed later in more details. D
17 .4.2 ElementbyElement Assembly
The assembly of the linear matrix system is the popular element
byelement method. The stiffness matrix Kin equations (17.73)
and ( 17.75) is the summation of the integral over the whole
solution don1ain, and the don1ain is now divided into m ele
ments with each element on a subdomain ne (e = 1, ... , m).
Each element contributes to the whole stiffness matrix, and in
its contribution is a pure number. Thus, assembly of the
stiffness matrix can be done in an elen1entbyelement manner.
Furthermore, Ki,j =I= 0 if and only if (or iff) nodes i and j be
long to the same elements. In the 1D Ki,j =I= 0 only for
j = i  1, i, i + 1. In finite element analysis, the shape func
tions Nj are typically localized thus the 1natrix K is
usually sparse in most cases.
The elementbyelement formulation can be written as
(17. 77)
and
m
fi = L fi(e),
e=l
246
Finite Element !vletbod 17.4 Heat Conduction
In addition, since the contribution of each element is a sinl
ple number, the integration of each elen1ent can be done using
the local coordinates and local node numbers or any coordi
nate system for the convenience of integration over an element.
Then, the nonzero contribution of each element to the global
system matrix K is simply assembled by direct addition to the
corresponding global entry (of the stiffness n1atrix) of the cor
responding nodes or related equations. In reality, this can be
easily done using an index matrix to trace the element contri
bution to the global system matrix.
0 Example 17.2(b): The assembly of the global system matrix
for the example with 4 elements and five nodes is shown below. For
each element with i and j nodes, we have
L= he,
Kljl = [f kN;Njdx] = :. ( !
1
1)
1 '
f
~ e) = Q he ( 1 )
z 2 1 '
so that, for example in elements 1 and 2, these can extend to all
nodes (with hi = xi+l xi, i = 1, 2, 3, 4),
(
k/ht
k/ht
K<'l = ~
k/ht 0 0
k/ht 0 0
0 0 0
0 0 0
0 0 0
and so on. Now the global system matrix becomes
K=
247
17.4 Heat Conduction Finite Element Method
(
k/ht k/ht 0 0 0
}
k/hl k/hl + k/h2 k/h2 0 0
0 k/h2 k/h2 + k/h3 k/h3 0
0 0 k/h3 k/h3 + k/h4 k/h4
0 0 0 k/h4 k/h4
U = U ~ }
( Qh./2 ) Q(hl + h2)/2
I= Q(h2 + h3)/2 ,
Q(h3 + h4)/2
Qh4/2 + q
where the last row of I has already included the natural bounda1y
condition at u' ( 1) = q. D
17.4.3 Application of Boundary Conditions
Boundary conditions can be essential, natural or mixed. The
essential boundary conditions are automatically satisfied in the
finite elen1ent formulation by the approxin1ate solution. These
include the displacement, rotation, and known value of the so
lution. Sometimes, they are also called the geometric boundary
conditions. In our example, it is u(O) = /3. Natural boundary
conditions often involve the first derivatives such as strains,
heat flux, force, and moment. Thus, they are also referred to
as force boundary conditions. In our example, it is u'(1) = q.
The natural boundary conditions are included in the inte
gration in the finite elen1ent equations such as (17.75). Thus no
further in1position is necessary. On the other hand, although
the essential boundary conditions are automatically satisfied
in the finite element formulations, they still need to be im
plemented in the assembled finite element equations to ensure
unique solutions. The imposition of the essential boundary con
ditions can be done in main several ways: a) direct application;
b) Lagrangian n1ultiplier and c) penalty method. To show how
these methods work, we use the 1D poisson equation on the
distinct points xi(i = 1, 2, ... , 1\11) E (0, 1] to aid our discussion.
248
Finite Element 1\lethod 17.4 Heat Conduction
Direct Application
In this method, we simply use the expansion uh = Et!
1
uilVi,
and apply directly the essential boundary conditions at point i
to replace the corresponding ith equation with ui = ui so that
ith row of the stiffness matrix K in equation (17. 73) becon1es
(0, 0, ... , 1, ... , 0) and the corresponding /i = f(i) = ui. All
other points will be done in the similar manner. For example,
the boundary conditions u(O) = a and u(M) = (3 in the 1
D case mean that the first and last equations are replaced by
u1 =a and UM = {3, respectively. Thus, Ku = 1, It =a (all
other coefficients are set to be zeros: K12 = ... = K1M = 0,
and KMM = 1, !M = /3 with Ki\111 = ... = KM,M1 = 0).
Then, the equations can be solved for (ub u
2
, .•• , uM )T. This
method is widely used due to its simplicity and the advantage
of timestepping because it allows bigger time steps.
Lagrangian Multiplier
This method is often used in the structure and solid mechanics
to enforce the constraints ( Ui = Ui). The variation is added by
the extra term .X(ui ui) where A is the Lagrange multiplier.
Now we have
(17.79)
whose variation 8II = 0 leads to
(17.80)
Because 8u and 8X are arbitrary, we have
where ei = (0, 0, ... , 1, 0, ... , O)T (its ith entry is equal to one).
This method can be extended tom Lagrangian multipliers.
249
17.4 Heat Conduction Finite Element Method
Penalty Method
One of the most widely used methods of enforcing the essential
boundary conditions is the socalled penalty method in terms
of a very large coefficient I' Ui = ft at Xi E an E' so that
/Ui = '")'U can be directly added onto Ku =f. In the 1D ex
ample, it simply leads to Ku = Ku +!', KMM = KMM+/, and
It =It+ /O::, fA,/ = fM + '"'(/3. The common rule for choosing
I' is that '"'f >> max IKiil· Usually, '"Y ~ lOOOmax IKiil should
be adequate. The penalty method is widely used in steady
state problems. However, it may affect the efficiency of time
stepping since it increases the maximum eigenvalue of the stiff
ness matrix, and thus very small time steps are required for
convergence. The advantage of the penalty method is that the
handling of the essential boundary conditions becomes simpler
from the implementation point of view. The disadvantage is
that the conditions are only satisfied approximately.
0 Example 17.2(c): Following the same example of the 1D
steady state beat conduction discussed earlier, we now use the direct
application method for the essential boundary conditions. lVe can
replace the first equation E;=l K
1
iui = f
1
with u
1
= (3, so that the
first row becomes K
1
j = (1 0 0 0 0) and f
1
= (3. Thus, we have
K=
(
1 0 0 0 0
}
k/ht k/ht + k/h2 k/h2 0 0
0
k/h2 k/h2 + k/h3 k/h3 0
0 0
k/h3 k/h3 + k/h4 k/h4
0 0 0
k/h4 k/h4
U= ( ~ ~ }
( ~ )
Q(ht + h2)/2
I= Q(h2 + hJ)/2 ,
Q(h3 + h4)/2
Qh4/2 + q
For the case of k = 1,Q = 1,h
1
= ... = h
4
= 0.25, f3 = 1 and
250
Finite Element !vletbod 17.5 TimeDependent Problems
q = 0.25, we bave
( 1
0 0 0
~ }
( 1 )
4 8 4 0 0.25
K= ~
4 8 4 f = 0.25
0 4 8 4 0.25
0 0 4 4 0.375
Hence, tbe solution is
0.72
(
1.00 )
U = K
1
J = 0.50 .
0.34
0.25
17.5 TimeDependent Problems
D
The problems we have discussed so far in our finite element
analysis are not timedependent, and the solutions obtained are
the steadystate solutions. However, most realistic problems
involve time, and thus we will now discuss the tin1edependent
problems.
17.5.1 The Time Dimension
As the weak formulation uses the Green theorem that involves
the spatial derivatives, the time derivatives can be considered
as the source term. Thus, one simple and yet instructive way
to extend the finite element forn1ulation to include the time
dimension is to replace Q in equation (17.65) with Q GUt
/3Utt = Q  GU  /3il so that we have
V' · (kV'u) + (Q GUt f3utt) = 0. (17.81)
The boundary conditions and initial conditions are u(x, 0) =
4J(x), u = u,x E {)O.E, and k g ~  q = O,x E {)O.N. Using
251
17.5 TimeDependent Problems Finite Element Metbod
integration by parts and the expansion uh = E J ~
1
uiNi, we
have
M M
'Lrj (k'VN/vlVj)dn] + L J [(NiaNi)ui + (Ni/3Ni)uj]dn
i=l n i=l n
 f NiQdn  f Niqdr = o,
ln lanN
which can be written in a compact form as
where
and
Mii+ Cu+Ku = f,
Kii =In [(k'V Ni 'V Ni )]dO.,
fi = [ NiQdn + [ Niqdr,
ln lanN
(17.82)
(17.83)
(17.84)
(17.85)
(17.86)
The matrices K, M, C are symmetric, that is to say, Kij =
Kji, Mij = Mji, Cij = Cji due to the interchangeability of the
orders in the product of the integrand k, Ni and Nj (i.e., 'V Ni ·
k · 'VNi = k'VNi'VNi, Nia1Vi = 1ViaNi =aNiNi etc). The
matrix C = [Cij) is the damping matrix similar to the damping
coefficient of damped oscillations. M = [Mii] is the general
mass matrix due to a similar role acting as an equivalent mass
in dynamics. In addition, before the boundary conditions are
imposed, the matrix is usually singular, which may imply many
solutions. Only after the proper boundary conditions have been
enforced, the stiffness matrix will be nonsingular, thus unique
solutions may be obtained. On the other hand, M and C will
be always nonsingular if they are not zero. For example, for
the 1D elen1ents (with nodes i and j),
1)
1 '
(17.87)
252
Finite Element .Atfethod 17.5 TimeDependent Problems
but
= ( ) , det[M(e)] # 0,
= o.he ( 2 1 ) det[C(e)] # 0,
t} 6 1 2 '
(17.88)
Clearly, if M = 0, it reduces to the linear heat conduction. If
C = 0, it becomes the wave equation with the source term.
17 .5.2 TimeStepping
Fron1 the general governing equation
Mii+ Cu+Ku = f, (17.89)
we see that it is an ordinary differential equation in terms of
time and matrices. Thus, in principle, all the timestepping
methods developed in the standard finite difference method
can be used for this purpose. For a simple center difference
scheme, we have
.. (un+1 _ 2un + un1)
u =
(17.90)
so that equation (17.89) becomes
(un+1 _ 2un + un1)
M
(17.91)
Now the aim is to express un+
1
in tenns of un and un
1
.
17.5.3 1D Transient Heat Transfer
In the case of heat conduction (M = 0), we have
Cu+Ku=f, (17.92)
253
17.5 TimeDependent Problems Finite Element Method
or
. c
1
(f K )
U=  U. (17.93)
Using the explicit timestepping method, we can write it as
n+l n
u  u = ct (f K n)
tl.t u '
(17.94)
so that we have
(17.95)
0 Example 17.9: For a transient heat conduction problem, we
have
OUt = kuxx + Q,
and
u(x, 0) = 0, u(O, t) = 1, u'(l) = q.
The formulation with 5 nodes and 4 elements leads to
C=
( 2h,
ht 0 0
0 )
ht 2(ht + h2) h2 0 0
~ 0
h2 2(h2 + ha) ha 0
6 0
0
ha 2(ha + h4) h4
0 0 0
h4 2h4
For the case of a= 6,k = 1,Q = 1,h
1
= ... = h
4
= 0.25, we
have
( 0.5
0.25 0 0
0 )
0.25 1 0.25 0 0
C= 0 0.25 1 0.25 0 .
0 0 0.25 1 0.25
0 0 0 0.25 0.5
0
17.5.4 Wave Equation
For the wave equation (C = 0), we have
Mii+Ku=f. (17.96)
254
Finite Element 1\lethod 17.5 TimeDependent Problems
Using
.. un+1  2un + un1
u = ( ~ t ) 2
(17.97)
we have
where I is an identity or unit matrix. For example, the 1D
wave equation
8
2
u 8
2
u
8t2 = c {)x2'
with the boundary conditions
u(O) = u(1) = 0, u(x, 0) = e(x
1
/
2
)
2
,
can be written as
(17.99)
(17.100)
Mii = fo
1
NiNjdx, Kii = l cNfNjdx, f = 0, (17.101)
and u
0
is derived from the u(x, 0) = exp[(x 1/2)
2
].
The finite element methods in this book are mainly for lin
ear partial differential equations. Although these methods can
in principle be extended to nonlinear problems, however, some
degrees of approximations and linearization are needed. In ad
dition, an iterative procedure is required to solve the resultant
nonlinear n1atrix equations. The interested readers can refer to
many excellent books on these topics.
255
Chapter 18
Reaction Diffusion
System
The partial differential equations we solved in the previous
chapters using the three major numerical methods are linear
equations. We know that the generalized forn1s of parabolic
equations are nonlinear reactiondiffusion equations. l\1athe
matically speaking, nonlinear equations are far more difficult
to analyse if it is not impossible. From the numerical point
of view, some extra linearization and approximations should
be used for the nonlinear terms. However, the finite differ
ence scheme should still be useful for most nonlinear equations
though they should be implemented more carefully. Before we
proceed to study the nonlinear system, let us review the funda
mental characteristics of linear parabolic equations by solving
the linear heat conduction equations.
18.1 Heat Conduction Equation
18.1.1 Fundamental Solutions
From the similarity solution in section 10.3, we know that both
diffusion equation and heat conduction equation may mathe
257
18.1 Heat Conduction Equation Reaction Diffusion System
matically have a similarity variable defined by
X
(= ym'
(18.1)
where k is either the thermal diffusivity or diffusion coefficient.
In engineering, the coefficient k has the unit [n1]
2
/[s], thus the
unit of kt is [1n]
2
, which means the variable ( is dimensionless.
Any two combinations of x and kt giving the same ( will have
similar solutions.
Using the similarity variable (, the heat conduction equa
tion can be transformed into an ordinary differential equation
!"(() = 2(/', or (ln !')' = 2(. (18.2)
Integrating it once, we have
(18.3)
where C is an integration constant. Integrating it again, we
have
1
xo
2
u = C x e< d( + D, (18.4)
which is the general solution of the heat conduction equation.
If the domain is semiinfinite or infinite so that x
0
~ 0, then
we get
X
=Aerf( ~ ) + B .
v4kt
(18.5)
0 Example 18.1:
domain, we have
For the heat conduction in a semiinfinite
au= k8
2
u
8t 8x
2
'
with the boundazy condition
U = UQ, (X ~ 0),
and an initial condition
u = 0, (x > 0) at t = 0.
258
Reaction Diffusion System 18.2 Nonlinear Equations
The general solution is
X
u = Aerf( /"A"U) + B.
v4kt
For x ~ oo, erf(x/ ../4kl) ~ 1, we have
A+B = 0.
At x = 0, erf(O) = 0, we get B = uo. The solution is
X X
u = uo[l  erf(. 17'i:!')] = uoerfc r:rr:r·
v4kt v4kt
However, if u is constant (u = uo) in the initial region x E [h, h),
then we have
uo hx h+x
u = [erf + erf].
2 J4kt J4kt
0
The solutions of heat conduction do not always involve the
error function because error functions only occur when the inte
gration involves semiinfinite or infinite domains. If the domain
has a finite length, then the solutions often consist of power se
ries or even special functions. For example in heat conduction
through a plane sheet with zero initial temperature, its two
surfaces are held at constant temperatures with the boundary
conditions u = u
0
at x = 0 for (t ~ 0), and u = 0 at x = L for
(t ~ 0). The general solution can be written as
(1
x) 2 ~ uo . n1rx kn21r2t;L2
u = uo   + L....,  sm e ,
L 1T' n=l n L
(18.6)
which is a slowly convergent series.
18.2 Nonlinear Equations
18.2.1 Travelling Wave
The nonlinear reactiondiffusion equation
8u 8
2
u
8t = D {)x
2
+ f(u),
(18. 7)
259
18.2 Nonlinear Equations Reaction Diffusion System
can have the travelling wave solution under appropriate condi
tions of f(O) = f(1) = 0, f(u) > 0 for u E [0, 1], and f'(O) > 0.
For example, f(u) = ')'u(1  u) satisfies these conditions, and
the equation in this special case is called the Kolmogorov
PetrovskiiPiskunov (KPP) equation. By assuming that the
travelling wave solution has the form u( () and ( = x  vt, and
substituting into the above equation, we have
Du"(() + vu'(() + f(u(()) = 0. (18.8)
This is a secondorder ordinary differential equation that can
be solved with the appropriate boundary conditions
u( oo) + 1, u( oo) + 0. (18.9)
The KPP theory suggests that the limit of the speed of the
travelling wave satisfies
v 2?: 2J Df'(O). (18.10)
18.2.2 Pattern Formation
One of the most studied nonlinear reactiondiffusion equations
in the 2D case is the KolmogorovPetrovskiiPiskunov (KPP)
equation
{)u fPu 8
2
u
{)t = D( ax2 + {)y2) + '"YQ(u),
(18.11)
and
q(u) = u(1  u). (18.12)
The KPP equation can describe a huge number of physical,
chemical and biological phenomena. The most interesting fea
ture of this nonlinear equation is its ability of generating beau
tiful patterns. We can solve it using the finite difference scheme
by applying the periodic boundary conditions and using a ran
dom initial condition u = random( n, n) where n is the size of
the grid.
260
Reaction Diffusion System 18.2 Nonlinear Equations
1
0.5
0
200
0 200
0
Figure 18.1: 2D pattern formation forD= 0.2 and 'Y = 0.5.
Figure 18.1 shows the pattern formation of the above equa
tion on a 100 x 100 grid for D = 0.2 and 'Y = 0.5. We can see
that rings and thin curves are formed, arising from the random
initial condition. The landscape surface shows the variations in
the location and values of the field u(x, y) can be easily demon
strated.
The following simple 15line Matlab program can be used
to solve this nonlinear system.
% Pattern formation: a 15 line matlab program
% POE form: u_t=O*(u_{xx}+u_{yy})+gamma*q(u)
%where q(u)='u.*(1u)';
% The solution of this POE is obtained by the
% finite difference method, assuming dx=dy=dt=1
% Written by X S Yang (Cambridge University)
% Usage: pattern(100) or simply >pattern
function pattern(time)
%
line 1
% Input number of time steps
if nargin<1, time=100; end
%
line 2
261
18.2 Nonlinear Equations Reaction Diffusion System
% Initialize parameters n=100;
% 0=0.2; gamma=0.5;
n=200; 0=0.2; gamma=0.5; % line 3
%Set initial values of u randomly
u=rand(n,n); grad=u*O;
% Index for u(i,j) and the loop
I = 2:n1; J = 2:n1;
% Time stepping
% line 4
% line 5
for step=1:time, % line 6
% Laplace gradient of the equation % line 7
grad(I,J)= u(I,J1)+u(I,J+1)+u(I1,J)+u(I+1,J);
u ={14*0)*u+O*grad+gamma*u.*(1u); %line 8
% Show results
pcolor(u); shading interp;
%
line 9
% Coloring and colorbar
colorbar; colormap jet;
%
line 10
drawnow;
%
line 11
end
%
line 12
% plot as a surface
surf(u);
%
line 13
shading interp;
%
line 14
view( [ 25 70]);
%
line 15
%  End
of Pragram

If you use this program to do the simulations, you will see that
the pattern emerges naturally from the initially random back
ground. Once the pattern is formed, it evolves gradually with
time, but the characteristics such as the shape and structure
of the patterns do not change much with time. In this sense,
one can see beautiful and stable patterns.
262
Reaction Diffusion System 18.3 ReactionDiffusion System
18.3 ReactionDiffusion System
The pattern formation in the previous section arises naturally
from a single equation of nonlinear reactiondiffusion type. In
many applications, we often have to simulate a system of non
linear reactiondiffusion equations, and the variables are cou
pled in a complicated manner.
The pattern formation in the previous section comes from
the instability of the nonlinear reaction diffusion system. In
order to show this, let us use the following mathematical model
for enzyme inhibition and cooperativity.
For example, the following system consists of two nonlinear
equations
(18.13)
(18.14)
and
bu cv
f(u, v) = au(1  u) 
1
,
+u+v
(18.15)
g(u,v) =
vd
(18.16)
1+u+v'
where Du and Dv are diffusion coefficients, while a, b, c, d are all
constants. This reaction diffusion system may have instability
if certain conditions are met.
The steady state solutions are obtained from f(uo, vo) = 0
and g(uo, vo) = 0. They are
b ~
uo =
2
a [ V 1 + 4/;2  1],
vo = 0. (18.17)
Let 1/J = (u uo, v  vo) be the small perturbation, then 1/J
satisfies
81/; = D\12,.,, + M"''
at ~ ~ ,
(18.18)
263
18.3 ReactionDiffusion System Reaction Diffusion System
where
(18.19)
and
M = 1 ( (2au
0
+b) a(1 duo)+ c ) .
(1 + uo) 0
(18.20)
Writing 'l/J in the form of
(18.21)
where the sumn1ation is over all the wavenumbers k, we have
(18.22)
where I is a 2 x 2 unity matrix. This eigenvalue equation has
two roots. Since ~ ( A ) > 0 implies that instability, this requires
that
Dv d
 < ~      : ~
Du (2auo +b)
(18.23)
The range of unstable wavenumbers between the two roots of
k
2
at the bifurcation point is given by
k2 = dDu  Dv(2auo +b) [
1
± J
1
D D
8
]
± 2DuDv ( 1 + Uo) +
4
u v '
(18.24)
with
8
_ (2auo +b)
 [dDu Dv(2auo + b)]2 ·
(18.25)
If the unstable criteria are satisfied, any small random pertur
bation can generate complex patterns.
Similar to the nonlinear KPP equation (18.12), beautiful
patterns also arise naturally in the following nonlinear system
8u 8
2
u 8
2
u 
8t = Da ( 8x2 + 8y2 ) + 'Y f ( u, v)'
(18.26)
8v 8
2
v 8
2
v _
8t = Db( 8x2 + 8y2) + f3g(u, v),
(18.27)
264
Reaction Diffusion System 18.3 ReactionDiffusion System
and
](u, v) = u(1 u), g(u, v) = u u
2
v, (18.28)
for the values Da = 0.2, Db = 0.1, "( = 0.5 and {3 = 0.2.
With different functions j ( u, v) and g( u, v), these equations
can be used to sin1ulate the pattern formation in a wide range
of applications where nonlinear reactiondiffusion equations are
concerned.
265
Chapter 19
Probability and
Statistics
All the mathematical models and differential equations we have
discussed so far are deterministic systems in the sense that
given accurate initial and boundary conditions, the solutions
of the system can be determined (the only exception is the
chaotic system to a certain degree). There is no intrinsic ran
domness in the differential equations. In reality, randomness
occurs everywhere, and not all models are deterministic. In
fact, it is necessary to use stochastic models and sometimes
the only sensible models are stochastic descriptions. In these
cases, we have to deal with probability and statistics.
19.1 Probability
19.1.1 Randomness and Probability
Randomness such as rouletterolling and noise arises from the
lack of information, or incomplete knowledge of reality. It can
also come from the intrinsic complexity, diversity and pertur
bations of the system. The theory of probability is mainly the
studies of random phenomena so as to find nonrandom regu
larity.
267
19.1 Probability Probability and Statistics
For an experiment or trial such as rolling dices whose out
come depends on chance, the sample space n of the experi
ment is the set of all possible outcomes. The sample space
can be either finite or infinite. For exan1ple, rolling a sixsided
die will have six different outcomes, thus the san1ple space is
n = {1, 2, 3, 4, 5, 6}. The elements of a sample space are the
outcomes, and each subset of a sample space is called an event.
For example, the event S = {2, 4, 6} is a subset of n. In a
sample space n, the outcomes of an experiment is represented
as numbers ( 1 for heads and 0 for tails for tossing coins). A
realvalued variable that is defined for all the possible outcomes
is referred to as a random variable, which is a function that as
sociates a unique numerical value with every outcome of an
experiment, and its actual value varies from trial to trial as the
experiment is repeated. The values of a random variable can
be discrete (such as 1 to 6 in rolling a single die) or continuous
(such as the level of noise). If a random variable only takes
discrete values, it is called a discrete random variable. If its
values are continuous, then it is called a continuous random
variable.
Figure 19.1: Venn Diagrams: An Band AU B.
Two events A and B can have various relationships and
these can be represented by Venn diagrams as shown in Figure
19.1. The intersection An B of two events means the outcome
of the random experiments belongs to both A and B, and it
is the case of 'A AND B'. If no event or outcome belongs to
the intersection, that is An B = 0, we say these two events are
268
Probability and Statistics 19.1 Probability
mutually exclusive or disjoint.
The union A U B denotes the outcome belongs to either
A or B or both, and this means the case of 'A ORB'. The
complement A = n  A (or not A) of the event A is the set of
outcomes that do not belong to A but in the sample space n
(see Figure 19.2). The A B means the outcomes in A only.
Figure 19.2: Venn Diagrams: A = n  A and A  B.
Probability P is a number or an expected frequency as
signed to an event A that indicates how likely the event will
occur when a random experiment is performed. This probabil
ity is often written as P( A) to show that the probability P is
associated with event A. For a large number of fair trials, the
probability can be calculated by
P(A) = NA(number of outcomes in the event A).
N n (total number of outcomes)
(19.1)
0 Example 19.1: If you tossed a coin 1000 times, the head (H)
occurs 511 times and the tail (T) occurs 489 times. The probability
P(H) and P(T) are
511
P(H) =
100
= 0.511,
and
489
P(T) =
1000
= 0.489.
There are three axioms of probability, and they are:
Axiom I : 0 ~ P(A) ~ 1.
269
19.1 Probability Probability and Statistics
Axiom II: P(f!) = 1.
Axiom III: P(A U B)= P(A) + P(B), if An B = 0.
The first axiom says that the probability is a number be
tween 0 and 1 inclusive. P(A) = 0 corresponds to impossibility
while P(A) = 1 corresponds to absolute certainty. The sec
ond axiom simply means that an event must occur somewhere
inside the sample space. The third axiom is often called the ad
dition rule. Since A and A are mutually exclusive (An A= 0),
we have
P(A) + P(A) = P(A u A) = P(fl) = 1, (19.2)
or
P(A) = 1  P(A), (19.3)
which is usually called the NOT rule. The third axiom can be
further generalized to any two events A and B
P(A U B) = P(A) + P(B)  P(A n B). (19.4)
In a special case when events Ai ( i = 1, 2, ... , n) exhaust the
wholesamplespacesuch that A= Uf=tAi = AtUAiU ... UAn =
nand Ai n Ai = 0(i =I= j),
n
P(A n B)= L P(Ai n B). (19.5)
i=l
Since n n B = B, we also get
n
P(n n B) = P(B) = L P(Ai n B), (19.6)
i=l
which are the useful properties of the total probability.
For example, if you randomly draw a card from a standard
pack of 52 cards, what is the probability of it being a red king
or a diamond with a face value being a prime number (if its
270
Probability and Statistics 19.1 Probability
face value is counted from 1 to 13). The prime numbers are 2,
3, 5, 7, 11, 13, therefore they are 6 cards that forms the primes.
The possibility of event (A) of drawing a red king is P( A) =
~ = ,Js. The probability of event (B) of drawing a prime
number is P(B) =
5
6
2
= :f
6
• As a diamond king (13) is also a
prime, this means P( A n B) = ~ . Therefore, the probability
P(A U B) = P(A) + P(B)  P(A n B)
1 3 1 7
= 26 + 26  52 = 52.
Two events A and B are independent if the events has no
influence on each other. That is to say, the occurrence of one
of the events does not provide any information about whether
or the other event will occur. In this case, the probability of
both occuring is equal to the product of the probabilities of the
two individual events P(A) and P(B)
P(A n B)= P(A) * P(B). (19.7)
This can be easily extended to n mutually independent events
Ai(i = 1, 2, ... , n). The probability of all these events happening
is
n
P(L nAi) = Ilf::
1
P(Ai) = P(Al)P(A2) · · · P(An)· (19.8)
i=l
0 Example 19.2: The probability of drawing a king from a pack
of cards (Event A), and showing an even number of rolling a six
sided die (event B) is P(A n B). We know P(A) = 4/52 = 1/13,
and P(B) = 3/6 = 1/2. Since these two events are independent, the
probability that both events occur is
1 1 1
P(A n B) = P(A)P(B) =  ·  = 
13 2 26.
0
If the two events are not independent, then one may affect
the other event, in this case, we are dealing with the conditional
probability which will be discussed later in the next section.
271
19.1 Probability Probability and Statistics
In calculating the probabilities, it is useful to know the pos
sible combinations and permutations of certain objects. Sup
pose you have 5 pairs of shoes, 4 pairs of trousers, 7 shirts and
2 hats. This is equivalent to the lineup problem from your feet
to your head. In this case, as the event of selecting each thing
to wear is in a similar manner of putting it into slots in succes
sive stages, the total number of all possible ways is simply the
multiplication of all the possible choices for each stages. All
possible outfits you can wear form a permutation problem, and
the total number is 5 x 4 x 7 x 5 = 700.
In order to line 5 objects marked A, B, C, D, E, in the first
place, there are 5 possible choices, the second place has only 4
options, the third place 3 choices, the fourth place has 2 choices,
and there is only one left for the last place. Thus the number
of all possible permutations is 5 x 4 x 3 x 2 x 1 =51. Following
this line of reasoning, n objects can in general be pern1utated
. I
Inn. ways.
Suppose there are n = 20 students in a class (named 81,
8
2
, .... , 82
0
), we want to select 5 students at random to fonn
a 5student team to work on a project. This is different from
the lineup problem because once you have selected any five
students (say) 8 ~ , 81, 8w, 81s, 819, it does not n1atter what
order you selected them, the final formed team is the same.
There are 51 permutations within the same tea1n. Order does
not count in this case. This is a combination problem (also
called a comn1ittee problem). As before, there are 5 places to
line up the students, and the total number of all permutations
for selecting 5 students is 20 * 19 * 18 * 17 * 16. Therefore, the
total number of combinations (of selecting 5 students) is
20c = 20 * 19 * 18 * 11 * 15 = 201 =
15504
5
51 5!15! .
(19.9)
In general, the total number of all possible combinations of
selecting k objects from n is
nck = ( ~ ) = k l ( n n ~ k)!" (19.10)
272
Probability and Statistics 19.1 Probability
The consistency requires OJ= 1.
0 Example 19.9: A club of 5 members is chosen at random
from 8 female students, 10 male students, and 7 teachers. What is
the probability of the club consisting of 2 female students, 2 male
students, a.nd 1 teacher? The total number of clubs is
25
Cs. If two
female students are selected, we have
8
C2. Similarly,
10
C2 for select
ing 2 male students, and
7
C
1
for selecting one teacher. Therefore,
the total number N of forming the 5member club is
N = sc2IOC27Cl = ~ ~ 0166
2SC
5
253 ·
0
There is an interesting 'birthday paradox' which is related
to this context. The birthday paradox was first proposed in
1939 by Richard von l\1ises, which states that what is the prob
ability of two people having the same birthday in a group of n
people. For a group of 367 people, it is certain that there must
be two people having the san1e birthday as there are only 365
(or 366 if someone was born in a leap year) possible birthdays.
Ignoring 29 February and the year of birth and assuming that
the birthdays are evenly distributed throughout the year, we
only have 365 different birthdays (days and months only). If
the event A denotes that all the n people will have different
birthdays (no birthday matching), the first person can have
any date as his or her birthday, 365/365. The second person
must be in other 364 dates, which is 364/365, and the kth per
son has (365 k + 1)/365. Therefore, the probability of no two
people having the same birthday is
365 364 (365 n + 1)
P( A, n) = 365 x 365 x .. · x 365
365 * (364) * ... * (365 n + 1) _ 365!
365n  (365 n)f365n ·
(19.11)
Now the probability of two people with the same birthday is
 365!
P(A, n) = 1 P(A, n) = 1 (365 n)!365n.
(19.12)
273
19.1 Probability Probability and Statistics
The factorial 3651 is a large number, but you do not have to
deal with such large numbers. You can use a simple calculator
to estimate it. For five people, the probability of two people
with the same birthday is
P(.A
5
) =
1
_ 365 * 364 * 363 * 362 * 361 ~
0 027
(
19
.
13
)
' 365
5
. '
which seems insignificant. However, the interesting thing is
that for n = 23, the probability becomes
 365!
P(A, 23) = 1  (365  23)136523 ~ 0.507.
(19.14)
This means that you have slightly more than a 5050 chance of
finding two people sharing the same birthday. If you increase
n, you get P(A, 30) ~ 0. 706 for n = 30, P(.A, 40) ~ 0.891 for
n = 40, and P(A, 50) ~ 0.970 and P(A, 70) ~ 0.9992 (aln1ost
certainty) for n = 70.
It is worth noting that there is some difference in combi
nations when the member drawn is placed back or not. Sup
pose there are 10 red balls and 10 white balls in bag. If we
draw a ball (say a red, event A) from the bag and then put it
back (with replacement), then we draw another ball (event B).
P(A) = 1/20 and P(B) = 1/20. The probability of getting two
red balls are P(A n B) = P(A) * P(B) = 1/400. We call this
easel.
For a second case (Case II), if we do not put it back after
we have drawn the first ball (without replacmnent), then the
probability of event B is now different P(B) = 1/19 as there
is now only 19 balls in the bag. The probability of getting two
red balls now becomes P(A n B) = .Jrr x .fg = ~ ~ which is
different from 1/400.
The reason here is that the two events are not independent
in the case of noreplacement. If we use notation 'BIA' which
means that event B occurs given that event A has occurred,
then we can use P(BIA) to denote the probability of event
B when there is no replacement in event A in the scenario
274
Probability and Statistics 19.1 Probability
described in Case II. Now P(B) becomes P(BIA). Hence, we
have
P(A n B) = P(A)P(BIA), (19.15)
which is often called the multiplication rule in probability the
ory. Similarly, we can get
P(A n B) = P(B)P(AIB). (19.16)
This is essentially a conditional probability problem which forms
the main topic of the next section.
19.1.2 Conditional Probability
In calculating the probabilities, we often assume that all pos
sible outcomes of an experiment such as drawing a card are
equally likely. Probabilities can change if additional informa
tion is known or some other event has already occurred and
thus P(BIA) denotes the probability that event B will occur
given that event A has already occurred. The conditional prob
ability can be calculated by
Conversely, we have
P(BIA) = P(B n A)
P(A) .
P(AIB) P(A n B)
P(B) .
(19.17)
(19.18)
Using equation (19.15), we can write the above formulation as
P(AIB) = P(A)P(BIA)
P(B)
P(A)P(BIA)
P(A)P(BIA) + P(A)P(BIA)'
(19.19)
which is the Bayes' theorem. Here have used A u A = 0 and
P(A) = 1  P(A).
275
19.1 Probability Probability and Statistics
As an example, we consider the drug test in sports. It is
believed that the test is 99% accurate if athletes are taking
drugs. For athletes not taking drugs, the positive test is only
0.5%. It is assumed that only one in 1000 athletes takes this
kind of drug. Suppose an athlete is selected at random and
the test shows positive for the drug. What is the probability
that the athlete is really taking the drug? If event A denotes
an athlete is taking the drug, and B denotes the event that
the individual tests positive. Thus, P(A) = 1/1000, P(BIA) =
0.99 and P(BIA) = 0.005. The probability that the athlete is
actually taking the drug is
P(AIB) _ P(A)P(BIA)
 P(A)P(BIA)+P(A)P(BIA)
0.001 * 0.99 ~ 0.165.
0.001 * 0.99 + 0.999 * 0.005
(19.20)
This is surprisingly low in probability.
0 Example 19.4: The classical problem of three cards consists of
three cards: one blue card (B) is blue on both sides, one white card
(W) is wbite on both sides, and one mixed card (A1) is wbite on one
side and blue on the other. If you draw one card at random from a
bag and place it on a table, suppose that the side you can see is blue,
what is the probability of other side is also blue? This a conditional
probability problem. There are 3 blue faces and 3 white faces, thus
tbe total probability of showing a blue face (F) is P(BF) = 1/2, and
probability of pull the blueblue card out is P(BB) = 1/3, while the
probability of showing a blue face is P(BFIBB) = 1 if the pulled
card is blueblue one. Then, the probability of other side being also
blue is
P(BBIBF) = P(BFIBB)P(BB) = 1 X t = ~
P(BF) ! 3·
(19.21)
!\lost people will intuitively guess that the probability is t, which is
not correct. 0
Another related problem is the socalled l\1onty Hall prob
lem (or three door problem) in a classical game show. Suppose
276
Probability and Statistics 19.1 Probability
that you are given the choice of three doors. There is an ex
pensive car behind one door, behind other two doors are goats.
If you choose one door (say, A) at random, then the host opens
one of the other door (say, B), which he knows there is a goat
behind it, to reveal a goat. Now you have a choice either to
stick with your original choice or swap with the other door
(say, C). What is your best strategy based on probability? Ini
tially, the prize car behind any door (Y) has a priori probability
P(any) = 1/3, so your initial choice P(A) = 1/3. As the host
knows where the prize is, if the car is behind A, the host will
open B or C so 1/2 each. If the car is behind B, the host
will never open B, and if the car is behind C, the host will
surely open B. Nlathematically, this gives P(OpenBIA) = 1/2,
P(OpenBIB) = 0, and P(OpenBIC) =I.
So the total probability of opening door B is
P(OpenB) = P(A)P(OpenBIA) + P(B)P(OpenBIB)
1 1 1 1 1
+P(C)P(OpenBIC) = '3 x
2
+
3
x 0 +
3
x 1 =
2
. (19.22)
Now the probability of the car behind door Cis
P(CIOpenB) = P(OpenBIC)P(C) = 1 x ~ = ~ ' (19.23)
P(OpenB) ~ 3
which is greater than your initial choice 1/3. Therefore, the
best strategy is to switch your choice. This game has other
variations such as the three envelope problem and others, but
the analysis and strategy are the same.
19.1.3 Random Variables and Moments
Random Variables
For a discrete random variable X with distinct values such as
the number of cars passing through a junction, each value Xi
may occur with a certain probability p(xi)· In other words,
277
19.1 Probability Probability and Statistics
the probability varies with the random variable. A probabil
ity function p(xi) is a function that defines probabilities to all
the discrete values xi of the random variable X. As an event
must occur inside a sample space, the requirement that all the
probabilities must be summed to one leads to
n
LP(Xi) = 1. (19.24)
i=l
The cumulative probability function of X is defined by
P(X ~ x) = L p(xi)· (19.25)
Xi<X
For a continuous random variable X that takes a continuous
range of values (such as the level of noise), its distribution is
continuous and the probability density function p( x) is defined
for a range of values x E [a, b] for given limits a and b [or even
over the whole real axis x E (oo,oo)]. In this case, we always
use the interval (x, x + dx] so that p(x) is the probability that
the random variable X takes the value x < X ~ x + dx is
~ ( x ) = P(x < X ~ x + dx) = p(x)dx. (19.26)
As all the probabilities of the distribution shall be added to
unity, we have
l p(x)dx = 1.
(19.27)
The cumulative probability function becomes
~ ( x ) = P(X ~ x) = J.x p(x)dx, (19.28)
which is the definite integral of the probability density function
between the lower limit a up to the present value X= x.
Mean and Variance
Two main measures for a random variable X with a given prob
ability distribution p(x) are its mean and variance. The mean
278
Probability and Statistics 19.1 Probability
J1. or the expectation value of E[X] is defined by
J1. = E[X] =<X>= j xp(x)dx, (19.29)
for a continuous distribution and the integration is within the
integration limits. If the random variable is discrete, then the
integration becomes the sumn1ation
E[X] = L XiP(Xi)· (19.30)
The variance var[X] = a
2
is the expectation value of the
deviation squared (X J..L)
2
. That is
(19.31)
The square root of the variance a = Jvar[X] is called the
standard deviation, which is simply a.
This simply becomes a sum
(19.32)
for a discrete distribution. In addition, any other formulas for
a continuous distribution can be converted to their counter
part for a discrete distribution if the integration is replaced by
the sum. Therefore, we will mainly focus on the continuous
distribution in the rest of the section.
Other frequently used measures are the mode and median.
The mode of a distribution is defined by the value at which
the probability density function p( x) is maxin1um. For an even
number of data sets, the mode may have two values. The me
dian m of a distribution corresponds to the value at which the
cumulative probability function <l>(m) = 1/2. The upper and
lower quartiles Qu and QL are defined by <I>(Qu) = 3/4 and
<I>(QL) = 1/4.
279
19.1 Probability Probability and Statistics
Moments and Moment Generating Functions
In fact, the mean is essentially the first moment if we define
the kth moment of a random variable X by
E[Xk] = Jlk =I xkp(x)dx, k = 1, 2, ... , N. (19.33)
Similarly, the kth central moment is defined by
E[(X Ill] = vk =I (x p,)kp(x)dx, k = 1, 2, ... , N.
(19.34)
Obviously, the variance is the second central moment. From
these definitions, it is straightforward to prove
(19.35)
and
var[ax + .6] = a
2
var[X]. (19.36)
where a and .6 are constants.
Ivlost probability functions can be expressed in terms of
moments and moment generating functions. The moment gen
erating function is defined by
(19.37)
where v E 'R is a real parameter. By expanding exp[vx] into
power series and using the definition of various moments, it is
straightforward to verify that
(19.38)
and
2
_ ~ G x ( O ) _ [dGx(0)]
2
a  dv2
dv ·
(19.39)
280
Probability and Statistics 19.1 Probability
19.1.4 Binomial and Poisson Distributions
Binomial Distribution
A discrete random variable is said to follow the binomial dis
tribution B(n,p) if its probability distribution is given by
B(n,p) = nCxpx(l Ptx, nc  n! (19.40)
x xl(nx)l'
where x = 0, 1, 2, ... , n are the values that the random variable
X may take, n is the number of trials. There are only two
possible outcomes: success or failure. p is the probability of
a socalled 'success' of the outcome. Subsequently, the proba
bility of the failure of a trial is q = 1  p. Therefore, B(n,p)
represents the probability of x successes and n x failures in
n trials. The coefficients come from the coefficients of the bi
nomial expansions
n
(p+qt = LnCxpxqnx = 1,
(19.41)
x=O
which is exactly the requirement that all the probabilities should
be summed to unity.
0 Example 19.5: Tossing a coin 10 times, the probability of
getting 7 heads is B(n, 1/2). Since p = 1/2 and x = 7, then we have
10c (.!.)1(.!.)
3
= ~ ~ o 111
7
2 2 128 . .
0
It is straightforward to prove that J1. = E[X] = np and
a
2
= npq = np( 1  p) for a binomial distribution.
Another related distribution is the geometric distribution
whose probability function is defined by
P(X = n) = pqn1 = p(1 p)n1'
(19.42)
where n ~ 1. This distribution is used to calculate the first
success, thus the first n 1 trials must be in failure if n trials
are needed to observe the first success. The mean and variance
of this distribution are J1. = 1/p and a2 = (1  p)jp2.
281
19.1 Probability Probability and Statistics
Poisson Distribution
The Poisson distribution can be thought as the limit of the
binomial distribution when the number of trial is very large
n + oo and the probability p + 0 (small probability) with the
constraint that A = np is finite. For this reason, it is often
called the distribution for smallprobability events. Typically,
it is concerned with the number of events that occur in a certain
time interval (e.g., number of telephone calls in an hour) or
spatial area. The Poisson distribution is
A> 0, (19.43)
where x = 0, 1, 2, ... , n and A is the mean of the distribution.
Using the definition of mean and variance, it is straightforward
to prove that J.L = A and a
2
= A for the Poisson distribution.
The parameter A is the location of the peak as shown in Figure
19.3.
0 Example 19.6: If you receive 3 calls per hour on your mobile
phone on the average. If you do not switch your phone off, what is
the probability that it begins to sound (one call is enough) during
any onehour class? \'Ve know that .,\ = 3. The probability of no
phone call at all is
Thus, the probability of sounding is P(X > 0) ~ 1  0.0498 ~ 0.95.
In fact, the probability of receiving one call is
31e3
P(X = 1) = 
1
,::::::: 0.149,
and the probability of receiving two calls is
0
282
Probability and Statistics 19.1 Probability
0.05r.....;===::::::;,
0.04
,,
I \
I \
0.03 II \
'iC' I \
rt :' \
0.02 : \
I \
I \
I \
0.01 :'
I
,
 A=S
 · A=10
 A=15
25
X
Figure 19.3: Poisson distributions for different values of .X =
5, 10, 15.
The moment generating function for the Poisson distribu
tion is given by
(19.44)
19.1.5 Gaussian Distribution
The Gaussian distribution or normal distribution is the most
important continuous distribution in probability and it has a
wide range of applications. For a continuous random variable
X, the probability density function (PDF) of a Gaussian dis
tribution is given by
1 (x,_.)2
p(x) = . m = e  ~ ,
ay27T
(19.45)
where a
2
= var[X] is the variance and Jl = E[X] is the mean
of the Gaussian distribution. From the Gaussian integral, it is
easy to verify that
/
00
00
p(x)dx = 1,
283
(19.46)
19.1 Probability Probability and Statistics
0.1
0.08
0.02
/.'
,':'
I I
I I
, ,' ,·'
____ ,. ...... _ ..... ,
20 10 0
X
 <J=5
 · <J=7
 <J=10
10 20 30
Figure 19.4: Gaussian distributions for a = 7, 10.
and this is exactly the reason that the factor 1/ .;2i comes
from the normalization of the all probabilities. The probability
function reaches a peak at x = J..l and the variance a
2
controls
the width of the peak (see Figure 19.4).
The cumulative probability function ( CPF) for a normal
distribution is the integral of p( x), which is defined by
1 !X (<w2
4>(x) = P(X < x) = e 2a d(.
21ra
2
oo
(19.47)
Using the error function defined by Chapter 1, we can write it
as
1 XJ..t
4>(x) = ro[1 + erf( ro )].
v2 v2a
(19.48)
The moment generating function for the Gaussian distribution
is given by
(19.49)
The Gaussian distribution can be considered as the limit
of the Poisson distribution when A >> 1. Using the Sterling's
approximation xl rv ..f2;iX(xje)x for x >> 1, and setting J..l =A
and a
2
=A, it can be verified that the Poisson distribution can
284
Probability and Statistics 19.1 Probability
be written as a Gaussian distribution
(19.50)
where J..t = A. In statistical applications, the normal distribu
tion is often written as N (J..t, a) to emphasize that the proba
bility density function depends on two parameters J.L and a.
The standard normal distribution is a normal distribution
N(J..t, a) with a mean of J.L = 0 and standard deviation a = 1,
that is JV(O, 1) This is useful to normalize or standardize data
for statistical analysis. If we define a normalized variable
XJ.L
a
(19.51)
it is equivalent to give a score so as to place the data above or
below the mean in the unit of standard deviation. In terms of
the area under the probability density function, sorts where
the data falls. It is worth pointing out that some books define
z = = ( x  J.L) /a in this case, and call the standard normal
distribution as the Z distribution.
Table 19.1: Function¢> defined by equation (19.53).
¢>
0.0 0.500 1.0 0.841
0.1 0.540 1.1 0.864
0.2 0.579 1.2 0.885
0.3 0.618 1.3 0.903
0.4 0.655 1.4 0.919
0.5 0.692 1.5 0.933
0.6 0.726 1.6 0.945
0.7 0.758 1.7 0.955
0.8 0.788 1.8 0.964
0.9 0.816 1.9 0.971
285
19.1 Probability Probability and Statistics
Now the probability density function of standard normal
distribution becomes
(19.52)
Its cumulative probability function is
= = 
2
1
[1 + erf(.
v27r oo v2
(19.53)
As the calculations of ¢ and the error function involve the nu
merical integrations, it is usual practice to tabulate ¢ in a table
(see Table 19.1) so that you do not have to calculate their values
each time you use it.
19.1.6 Other Distributions
There are a dozen of other important distributions such as the
exponential distribution, lognormal distribution, uniform dis
tribution and the x
2
distribution. The uniform distribution has
a probability density function
1
p= , X= (a:,b],
{3a:
(19.54)
whose mean is E[X] = (a: + {3) /2 and variance is a
2
= (/3 
a:)
2
/12.
The exponential distribution has the following probability
density function
f(x) = ..\e.Xx (x > 0),
(19.55)
and f(x) = 0 for x:::; 0. Its mean and variance are
(19.56)
The lognormal distribution has a probability density func
tion
1 (lnx JL)
2
f(x) =
2
2
],
x 21ra
2
a
(19.57)
286
Probability and Statistics 19.1 Probability
whose mean and variance are
(19.58)
The x
2
distribution, called chisquare or chisquared dis
tribution, is very useful in statistical inference and method of
least squares. This distribution is for the quantity
2 _ ~ ( X i  /li)2
XnL ,
i=l ai
(19.59)
where thenindependent variables Xi are normally distributed
with means Jli and variances a'f. The probability density func
tion for x
2
distribution is given by
p(x) = 2n/2:(n/2) x1'lexf2 (19.60)
where x 2:: 0, and n is called the degree of freedom. Its cumu
lative distribution function is
<I>( ) = "'f(n/2, x/2)
X f(n/2) '
(19.61)
where "Y(n/2, x/2 is the incomplete gamma function. It can be
verified that the mean of the distribution is n and its variance
is 2n.
For other distributions, readers can refer to any books that
are devoted to probability theory and statistical analysis.
19.1.7 The Central Limit Theorem
The most important theorem in probability is the central limit
theorem which concerns the large number of trials and explains
why the normal distribution occurs so widely. This theorem is
as follows: Let Xi(i = 1, 2, ... , n) ben independent random vari
ables, each of which is defined by a probability density function
Pi ( x) with a corresponding mean J..Li and a variance a'f. The sum
of all these random variables
n
e = Lxi = Xt +X2 + ... +Xn,
i=l
287
(19.62)
19.1 Probability Probability and Statistics
0.2
>0.15
C)
i
:::s
tr
! 0.1
LL
0.05
1 2 3 4 5 6
Outcome
Figure 19.5: A uniform distribution.
7
is also a random variable whose distribution approaches the
Gaussian distribution as n ~ oo. Its mean E[8] and variance
var[8] are given by
n n
E[8] = L E[Xi] = L Jli,
(19.63)
i=l i=l
and
n n
var[8] = L var[8] = La?.
(19.64)
i=l i=l
The proof of this theorem is out of the scope of this book
as it involves the n1oment generating functions, characteristics
functions and other techniques. In engineering mathematics,
we simply use these important results for statistical analysis.
In the special case when all the variables Xi are described
by the same probability density function with the same mean
Jl and variance a
2
, these results become
E[8] = np,, var[8] = na
2
.
By defining a new variable
8 np,
{n = ..jn '
a n
288
(19.65)
(19.66)
Probability and Statistics 19.2 Statistics
0.2
0.15
0.05
I I
2 4 6 8 10 12 14
Outcome
Figure 19.6: A bilinear distribution.
then the distribution of ~ n converges towards the standard nor
mal distribution N(O, 1) as n ~ oo.
Let us see what the theorem means for a simple experiment
of rolling a few dice. For a fair sixsided die, each side will
appear equally likely with a probability of 1/6 ~ 0.1667, thus
the probability function after rolling it 15000 times approaches
a uniform distribution as shown in Figure 19.5.
If we now roll two independent dice 15000 tin1es and count
the sum (112) of the face values of both dice, then the sum
obeys a bilinear distribution as shown in Figure 19.6. If we
roll n = 15 independent dice, the sums of the face values vary
from 1 to 90. After rolling the 15 dice 10,000 times, the distri
bution is shown in Figure 19.7 and it approaches to a normal
distribution as n ~ oo.
19.2 Statistics
Statistics is the mathematics of data collection and interpreta
tion, and the analysis and characterisation of numerical data
by inference from sampling. Statistical methods involve reduc
289
19.2 Statistics Probability and Statistics
0.1 r...rr.,
0.08
>
~ 0 . 0 6
CD
::I
117'
~ 0 . 0 4
0.02
~ 0 30 50 60 70 80
Outcome
Figure 19.7: An approximate Gaussian distribution (the out
comes of the sum of face values in rolling 15 dice).
tion of data, estimates and significance tests, and relationship
between two or more variables by analysis of variance, and the
test of hypotheses.
19.2.1 Sample Mean and Variance
If a sample consists of n independent observations Xt, x2, ... , Xn
on a random variable x such as the price of a cup of coffee, two
important and commonly used parameters are sample mean
and sample variance, which can easily be estimated from the
sample. The san1ple mean is calculated by
1 1 n
X =<x>= (Xt + X2 + ... + Xn) = LXi,
n n i=l
(19.67)
which is essentially the arithmetic average of the values Xi.
Generally speaking, if u is a linear combination of n inde
pendent random variables Yt, Y2, ... , Yn and each random vari
able Yi has an individual mean J.li and a corresponding variance
290
Probability and Statistics
a[, we have the linear combination
n
U = L O:iYi = O:tYl + G2Y2 + ··· + GnYn,
i=l
19.2 Statistics
(19.68)
where the parameters ai(i = 1, 2, ... , n) are the weighting coef
ficients. From the central limit theorem, we have the mean J.lu
of the linear combination
n n
J.lu = E(u) = E(L GiYi) = L aE(yi) = L GiJ.li· (19.69)
i=l i=l
Then, the variance u ~ of the combination is
n
u ~ = E[(u J.lu)
2
] = E[L o:i(Yi J.li)
2
],
i=l
which can be expanded as
n
a ~ = L a ~ E[(Yi  J.li)2]
i=l
n
+ L O:iGjE[(Yi J.li)(Yj J.lj)],
i,j=l;i#j
(19.70)
(19.71)
where E[(Yi  J.Li)
2
] = ar Since Yi and Yi are independent,
we have E[(Yi  J.li)(Yi  J.li )] = E[(Yi  J.li)]E[(Yi  J.li )] = 0.
Therefore, we get
n
u ~ = L:a'faJ. (19.72)
i=l
The sample mean defined in equation (19.67) can also be
viewed as a linear combination of all the Xi assuming each of
which has the same mean J.li = J.l and variance u'f = u
2
, and the
same weighting coefficient Gi = 1/n. Hence, the sample mean
is an unbiased estimate of the sample due to the fact J.lx =
Ef=t J.L/n = J.l. In this case, however, we have the variance
n 1 0"2
0 " ~ = ~ 0"2 = 
x ~ n2 n'
t=l
291
(19.73)
19.2 Statistics Probability and Statistics
which means the variance becomes smaller as the size n of the
sample increases by a factor of 1 ln.
The sample variance 8
2
is defined by
2 1 ~ 2
S = 
1
L...,(xi x) .
n i=l
(19.74)
It is worth pointing out that the factor is 1l(n 1) not 1ln be
cause only 1 I ( n  1) will give the correct and unbiased estimate
of the variance. From the probability theory in the earlier sec
tions, we know that E[x
2
] = J..L
2
+ a
2
. The mean of the sample
variance is
1 ~ 2 1 ~ 2 2
J..ls2 = E[
1
L...,(xix) ] = 
1
L E[(xi nx )]. (19.75)
n i=l n i=l
Using E[x
2
] = J..L
2
+ a
2
In, we get
Jls2 = n ~
1
~ { E [ x ~ ]  nE[x
2
]}
1 a
2
= { n(J..l2 + a2)  n(J..l2 + )} = a2.
n1 n
(19.76)
0 bviously, if we use the factor 1 In instead of 1 I ( n  1), we
would get J..l
8
2 = n ~
1
a
2
< a
2
, which would underestimate the
sample variance. The other way to think the factor 1 I ( n  1)
is that we need at least one value to estimate the mean, we
need at least 2 values to estimate the variance. Thus, for n
observations, only n  1 different values of variance can be
obtained to estimate the total sample variance.
19.2.2 Method of Least Squares
Maximum Likelihood
For a sample of n values Xt,X2, .•• ,Xn of a random variable X
whose probability density function p( x) depends on a set of k
parameters f3t, ... , f3k, the joint probability is then
292
Probability and Statistics 19.2 Statistics
= p(xt, f3t, ... , !3k)p(x2, f3t, ... , !3k) · · · p(xm f3t, ... , f3k)· (19.77)
The essence of the maximum likelihood is to maximize 4> by
choosing the parameters /3i. As the sample can be considered
as given values, the n1aximum likelihood requires that
84>
0/3i = 0, (i = 1, 2, ... , k), (19.78)
whose solutions for /3i are the maximum likelihood estimates.
Linear Regression
For experiments and observations, we usually plot one variable
such as pressure or price y against another variable x such as
time or spatial coordinates. We try to present the data in a
way so that we can see some trend in the data. For n sets
of data points (Xi, Yi), the usual practice is to try to draw a
straight line y =a+ bx so that it represents the major trend.
Such a line is often called the regression line or the best fit line
as shown in Figure 19.8.
y
Figure 19.8: Least square and the best fit line.
The method of least squares is to try to determine the two
parameters a (intercept) and b (slope) for the regression line
from n data points. Assuming that Xi are known more precisely
and Yi values obey a normal distribution around the potentially
best fit line with a variance a
2
. Hence, we have the probability
(19.79)
293
19.2 Statistics Probability and Statistics
where A is a constant, and f(x) is the function for the re
gression [f ( x) = a + bx for the linear regression]. It is worth
pointing out that the exponent L:i=dYi f(xi)]
2
ja2 is similar
to the quantity X ~ defined in the x
2
distribution.
The essence of the method of least squares is to maximize
the probability P by choosing the appropriate a and b. The
maximization of P is equivalent to the minimization of the
exponent 'l/J
n
'l/J = L)Yi  f(xi)f ·
(19.80)
i=l
We see that 'l/J is the sum of the squares of the deviations f.r =
(Yi f(xi))
2
where f(xi) =a+ bxi. The minimization means
the least sum of the squares, thus the name of the method of
least squares.
In order to minimize 'l/J as a function of a and b, its deriva
tives should be zero. That is
8'1/J n
7) = 2L[Y (a+ bxi)] = 0,
a i=l
(19.81)
and
(19.82)
By expanding these equations, we have
n n
na+bLxi = LYi,
(19.83)
i=l i=l
and
n n n
a LXi + b L : x ~ = LXiYi,
(19.84)
i=l i=l i=l
which is a system of linear equations for a and b, and it is
straightforward to obtain the solutions as
(19.85)
294
Probability and Statistics 19.2 Statistics
where
1 n
y = LYi·
n i=l
If we use the following notations
and
n
Kxx = L X ~ ,
i=l
n
Kxy = LXiYi,
i=l
then the above equation for a and b becomes
KxxKy KxKy
a= K (K )2 '
n XX X
The residual error is defined by
fi = Yi  (a + bxi),
whose sample mean is given by
1 n 1 1 n
Jte =  L fi = Yi  a  b L Xi
n i=l n n i=l
= y  a  bx = [y  bx]  a = 0.
The sample variance 8
2
is
(19.87)
(19.88)
(19.89)
(19.90)
(19.91)
(19.92)
(19.93)
where the factor 1/ ( n  2) comes from the fact that two con
straints are need for the best fit, and the residuals therefore
have a n  2 degrees of freedom.
295
19.2 Statistics Probability and Statistics
Correlation Coefficient
The correlation coefficient rx,y is a very useful parameter to
find any potential relationship between two sets of data Xi and
Yi for two random variables x and y, respectively. If x has a
mean J..lx and a sample variances;, andy has a mean J..ly and
a sample variance s ~ ' the correlation coefficient is defined by
(19.94)
where cov(x, y) = E[(x  J..lx](Y  J..ly) is the covariance. If
the two variables are independent cov(x, y) = 0, there is no
correlation between them (rx,y = 0). If ri,y = 1, then there is a
linear relationship between these two variables. r x,y = 1 is an
increasing linear relationship where the increase of one variable
will lead to increase of another. r x,y = 1 is a decreasing
relationship when one increases while the other decreases.
For n sets of data points (Xi, Yi), the correlation coefficient
can be calculated by
or
nKxy KxKy
rx,y = ;.===::::::::::=======,
J(nKxx K';)(nKyy K ~ )
(19.96)
where Kyy = Ei=l yr
0 Example 19.7: Is there any relationship between shoe size
and height among general population? By collecting data randomly
among our friends, we have the following data:
Height (h): 162, 167, 168, 171, 174, 176, 183, 179 (em);
Shoe size (s): 5.5, 6, 7.5, 7.5, 8.5, 10, 11, 12.
From these data, we know the sample mean J.lh = 172.5, J.Ls = 8.5.
296
Probability and Statistics 19.2 Statistics
0.15.....rr.
~ 0.1
Q.
c::
.!!
~
c!o.os
Figure 19.9: Confidence interval 'Y = 1 a.
The covariance cov(h, s) = E[(h J..lh)(s J..ls)] = 13.2. \Ve also have
the standard deviation of height sh = 6.422 and the standard devia
tion of shoe size Ss = 2.179. Therefore, the correlation coefficient r
is given by
r = cov(h, s) ::::::: 13.2 :::::::
0
_
94
_
ShSs 6.422*2.179
This is a relatively strong correlation indeed. D
19.2.3 Hypothesis Testing
Confidence Interval
The confidence interval is defined as the interval fh ~ X ~ 0
2
so that the probabilities at these two limits fh and fh are equal
to a given probability 'Y = 1 a (say, 95% or 99%). That is
(19.97)
The predetermined parameter 1' is always near 1 so that it can
be expressed as a small deviation a << 1 from 1 (see Figure
19.9). If we choose 1' = 95%, it means that we can expect that
about 95% of the sample will fall in the confidence interval
while 5% of the data will not.
297
19.2 Statistics Probability and Statistics
For the standard normal distribution, this means P( 8
8) = 1 a, so that
a
8) = 1  2" (19.98)
If a = 0.05, we have </>( 8) = 0.975 or 8 = 1.960. That is to
say, 8 8 or J..L  8a x J..L + 8a. We also know that
if you repeat an experiment n times, the variance will decrease
from a
2
to a
2
jn, which is equivalent to say that the standard
deviation becomes a/ .Jii for a sample size n. If a = 0.01, then
8 = 2.579, we have
a a
J..L 2.579 Vn X J..L + 2.579 .;n· (19.99)
On the other hand, for 8 = 1, we get J..L a x J..L +a and
1' = 0.682. In other words, only 68.2% of the sample data will
fall in the interval [J..L a, J..L +a] or
x = J..L ±a,
with a 68.2% confidence level.
(19.100)
It is conventional to use 1 = 0.95 for probably significant,
0.99 for significant, and 0.999 for highly significant.
0 Example 19. 8: The sample data of the time taken for a quick
lunch at a restaurant are as follows: 19, 15, 30, 20, 15, 23, 28, 22,
23 minutes. Suppose you want to attend a lecture at 12:30, at what
time you should start your order if you want to take 5% chance of
being late? The sample mean is
1
J.L = x = g(19 + 15 + 30 + 20 + 15 + 23 + 28 + 22 + 23) = 21.67.
The sample variance is
which gives a standard deviation of a = 5.15 minutes. If you are
willing to take 5% chance, then = 0.95, it = 1.645. So
you shall start
x = J.L = 30.15,
298
Probability and Statistics 19.2 Statistics
which is about 30 minutes earlier or at about 12:00. 0
Student's tdistribution
The Student's ttest is a very powerful method for testing the
null hypothesis to see if the means of two normally distributed
samples are equal. This method was designed by W. S. Gosset
in 1908 and he had to use a pen name 'Student' because of his
e1nployer's policy in publishing research results at that time.
This is a powerful method for hypothesis testing using small
size samples. This test can also be used to test if the slope
of the regression line is significantly different from 0. It has
become one of the most popular methods for hypothesis testing.
The theoretical basis of the ttest is the Student's tdistribution
for a sample population with the unknown standard deviation
a, which of course can be estimated in terms of the sample
variance 8
2
from the sample data.
For n independent measurements/data Xt, x2, ... , Xn with an
estimated sample mean x and a sample variance 8
2
as defined
by equation (19.74), the tvariable is defined by
XJ.L
t = (Sf Vn).
(19.101)
The Student's tdistribution with k = n  1 degrees of freedom
is the distribution for the randon1 variable t, and the probability
density function is
(19.102)
It can be verified that the mean is E[t] = 0. The variance is
a
2
= kf(k 2) fork> 2 and infinite for 0 < k ~ 2.
The corresponding cumulative probability function is
r( k+t) 1t (2 k 1
F(t) = r [1 + ]+d(.
v'f:rrr(k/2) oo k
(19.103)
299
19.2 Statistics Probability and Statistics
This integral leads to a hypergeometric function, and it is not
straightforward to calculate, that is why they are tabulated in
many statistical tables. For a confidence level of y = 1  a, the
confidence interval is given by
Q
F(O) = 1 2"'
(19.104)
which is usually tabulated. For a = 0.05 and 0.01 (or 1a/2 =
0.975 and 0.995), the values are tabulated in Table 19.2.
Table 19.2: Limits defined by F(O) = 1  a/2 in equation
(19.104).
k
F(O)o.975 F(O)o.995
1 12.7 63.7
2 4.30 9.93
3 3.18 5.84
4 2.78 4.60
5 2.57 4.03
6 2.45 3.71
7 2.37 3.50
8 2.31 3.36
9 2.26 3.25
10 2.23 3.17
20 2.09 2.85
50 2.01 2.68
100 1.98 2.63
00 1.96 2.58
Suppose we are dealing with the 95% confidence interval,
we have p( 0 ~ t ~ 0) = 1  a = 0.95 or p(t ~ 0) = 1 
a/2 = 0.975, we have 0 = ta,k = 12.70(k = 1), 4.30(k = 2),
3.18(k = 3), ... , 2.228(k = 10), ... , 1.959 for k+ oo. Hence,
s s
II  0 < t < II+ 0.
,..., rn _,..., vn
300
(19.105)
Probability and Statistics 19.2 Statistics
This is much more complicated than its counterpart of the
standard normal distribution.
Student's ttest
There are quite a few variations of the Student's ttest, and
most common are the one sample ttest and the two sample
ttest. The one sample ttest is used for measurements that
are randomly drawn from a population to compare the sample
mean with a known number.
In order to do statistical testing, we first have to pose pre
cise questions or form a hypothesis, and such hypothesis is
conventionally called the null hypothesis. The basic steps of a
ttest are as follows:
1. The null hypothesis: Ho: J.L = J.Lo (often known value) for
one sample, or H
0
: f..Lt = f..L2 for two samples;
2. Calculate the ttest statistic t and find the critical value
() for a given confidence level 1 = 1  o by using F( t ::s;
0) = 1 of2;
3. If ltl > 0, reject the hypothesis. Otherwise, accept the
hypothesis.
0 Example 19.9: A group of candidates (say, more than 100
students) have claimed to have an averaged IQ of 110 (or J..Lo = 110).
Then, you randomly sample 11 students to do the IQ test and results
are: x = IQ = 106,112, 103,108,108, 109,100, 106,106,99,101. Test
the hypothesis:
Ho: J..L= J..Lo,
at a confidence level of95%.
From the data, we know that n = 11, x = 105.273, S = 4.077.
Then, we have
t = (x J..L) = (105.273 110) ~ _
3
_
846
_
(S/.fii) 4.077/JIT
301
19.2 Statistics Probability and Statistics
We only use the positive value if we look at the statistical tables.
lVe also know for k = n 1 = 10 degrees of freedom at a 95%
confidence level, (} = 2.228. At a 95% confidence level, the probability
oft>(} is 0.025 (or 2.5%) and the probability t < 8 is also 0.025.
Thus, the hypothesis is not valid at a 95% confidence level. At the
same level of confidence, the true mean J.Lo lies in in the range of
x 2.228 * Sj.J[f ~ J.Lo ~ x + 2.2288/.Jff. or
102.53 ~ J.Lo ~ 108.00.
0
Another important ttest is the twosample paired test. As
suming that two pairs of n sample data sets Ui and Vi are in
dependent and drawn from the same normal distribution, the
paired ttest is used to determine whether they are significantly
different from each other. The tvariable is defined by
n(n 1)
(19.106)
" " " ~ (U. _ iT.)2'
6t=l t. Vi
where {fi = Ui  [J and 1% = Vi  V. In addition,
2 1 ~   2
sd = 
1
L.)Ui Vi) .
n i=l
(19.107)
This is equivalent to apply the onesample test to the difference
ui  Vi data sequence.
0 Example 19.10: A novel teaching method of teaching children
science was tried in a class (say class B), while a standard method was
used in another class (say class A). At the end of the assessment,
8 students are randomly drawn from each class, and their science
scores are as follows:
Class A: Ui = 76, 77, 76, 81, 77, 76, 75, 82;
Class B: Vi= 79, 81, 77, 86, 82, 81, 82, 80.
At a 95% confidence level, can you say the new method is really
better than the standard method?
If we suppose that the two methods do not produce any differ
ence in results, that is to say, their mean are the same. Tl1us the null
302
Probability and Statistics 19.2 Statistics
hypothesis is:
Ho: J..L.\ = JlB·
lVe know that V = 77.5, V = 81. The combined sample VBriance
Sd = 2.828. We now have
t = [J v = 77.581 = 3 5
Sd/Vn 2.828/v'8 ..
We know from the statistical table that the critical value 8 = 2.37
for F(8) = 1  o:/2 and k = n  1 = 7. As t < 8 or t > 8, we
can reject the null hypothesis. That is to say, the new method does
produce better results in teaching science. 0
The variance analysis and hypothesis testing are important
topics in applied statistics, and there are many excellent books
on these topics. Readers can refer to the relative books listed
at the end of this book. It is worth pointing out that other
important methods for hypothesis testing are Fisher's Ftest,
x
2
test, and nonparametric tests. What we have discussed in
this chapter is just a tip of an iceberg, however, it forms a solid
basis for further studies.
303
References
Abramowitz M. and Stegun I. A., Handbook of Mathematical
Functions, Dover Publication, (1965).
Arfken G., Mathematical Methods for Physicists, Academic
Press, ( 1985).
Ashby lvi. F. and Jones D. R., Engineering Materials, Perga
mon Press, (1980).
Atluri S. N., Methods of Computer Modeling in Engineering
and the Sciences, Vol. I, Tech Science Press, 2005.
Audoly B. and Neukirch S., Fragmentation of rods by cas
cading cracks: why spaghetti do not break in half, Phys.
Rev. Lett., 95, 095505 (2005).
Bathe K. J., Finite Element Procedures in Engineering Anal
ysis, PrenticeHall, (1982).
Carrrier G. F. and Pearson C. E., Partial Differential Equa
tions: Theory and Technique, 2nd Edition, Academic
Press, 1988.
Carslaw, H. S., Jaeger, Conduction of Heat in Solids, 2nd
Edition, Oxford University Press, (1986).
Chriss N., BlackScholes and Beyond: Option Pricing Models,
Irwin Professional Publishing, (1997).
305
19.2 Statistics Probability and Statistics
Courant R. and Hilbert, D., Methods of Mathematical Physics,
2 volumes, Wileylnterscience, New York, (1962).
Crank J., Mathematics of Diffusion, Clarendon Press, Oxford,
(1970).
Devaney R. L., An Introduction to Chaotic Dynamical Sys
tems, Redwood, (1989).
Drew, D. A., Mathematical modelling of twophase flow, A.
Rev. Fluid Mech., 15, 261291 (1983).
Fenner R. T., Engineering Elasticity, Ellis Horwood Ltd, (1986).
Fowler A. C., Mathematical Models in the Applied Sciences,
Cambridge University Press, (1997).
Farlow S. J., Partial Differential Equations for Scientists and
Engineers, Dover Publications, (1993).
Fletcher, C. A. J., Fletcher C. A., Computational Techniques
for Fluid Dynamics, Vol. I, SpringerVerlag, GmbH, (1997).
Gardiner C. W., Handbook of Stochastic Methods, Springer,
(2004).
Gershenfeld N., The Nature of Mathematical Modeling, Cam
bridge University Press, (1998).
Goldreich P. and S. Tremaine, The dynamics of planetary
rings, Ann. Rev. Astron. Astrophys., 20, 24983 (1982).
Goodman R., Teach Yourself Statistics, London, (1957).
Gleick J., Chaos: Making a New Science, Penguin, (1988).
Hinch E.J., Perturbation Methods, Cambridge Univ. Press,
(1991).
Hull J. C., Options, Futures and Other Derivatives, Prentice
Hall, 3rd Edition, (1997).
306
Probability and Statistics 19.2 Statistics
Jeffrey A., Advanced Engineering Mathematics, Academic Press,
(2002).
John F., Partial Differential Equations, Springer, New York,
(1971 ).
Jouini E., Cvitanic J. and l\.fusiela NL, Handbook in Mathe
matical Finance, Cambridge Univ Press, (2001).
Kardestruncer H. and Norrie D. H., Finite Element Handbook,
l\.fcGrawHill, (1987).
Keener J., Sneyd J., A Mathematical Physiology, Springer
Verlag, New York, (2001).
Korn G. A. and Korn T. M., Mathematical Handbook for Sci
entists and Engineers, Dover Publication, (1961).
Kreyszig E., Advanced Engineering Mathematics, 6th Edition,
Wiley & Sons, New York, (1988).
Kant T., Finite Elements in Computational Mechanics. Vols.
1/11, Pergamon Press, Oxford, (1985).
Langtangen, H P, Computational Partial Differential Equa
tions: Numerical Methods and Diffpack Programming,
Springer, (1999).
LeVeque R. J., Finite Volume Methods for Hyperbolic Prob
lems, Cambridge University Press, (2002).
Lewis R. W., Nlorgan K., Thomas H., Seetharamu S. K., The
Finite Element Method in Heat Transfer Analysis, Wiley
& Sons, (1996).
Mandelbrot B. B., The Fractal Geometry of Nature, W.H.
Freeman ( 1982).
Mitchell A. R. and Griffiths D. F., Finite Difference Method in
Partial Differential Equations, Wiley & Sons, New York,
(1980).
307
19.2 Statistics Probability and Statistics
Ivloler C. B., Numerical Computing with MATLAB, SIAM,
(2004).
I\1urray J. D., Mathematical Biology, SpringerVerlag, New
York, (1998).
Ockendon J., Howison S., Lacey A., and lVIovchan A., Applied
Partial Differential Equations, Oxford University Press,
(2003).
Pallour J.D. and lVIeadows D. S., Complex Variables for Sci
entists and Engineers, lVIacmillan Publishing Co., (1990).
Papoulis A., Probability and statistics, Englewood Cliffs, (1990).
Pearson C. E., Handbook of Applied Mathematics, 2nd Ed,
Van Nostrand Reinhold, New York, (1983).
Press W. H., Teukolsky S. A., Vetterling W. T., Flannery B.
P., Numerical Recipe in C++: The Art of Scientific Com
puting, 2nd Edition, Cambridge University Press, (2002).
Puckett E. G., Colella, P., Finite Difference Methods for Com
putational Fluid Dynamics, Cambridge University Press,
(2005).
Riley K. F., Hobson M. P., and Bence S. J., Mathematical
Methods for Physics and Engineering, 3rd Edition, Cam
bridge University Press (2006).
Ross S., A first Course in Probability, 5th Edition, Prentice
Hall, (1998).
Selby S. I\1., Standard Mathematical Tables, CRC Press, (1974).
Sicardy B., Dynamics of Planetary Rings, Lecture Notes in
Physics, 682, 183200 (2006).
Strang G. and Fix G. J., An Analysis of the Finite Element
Method, PrenticeHall, Englewood Cliffs, NJ, (1973).
308
Probability and Statistics 19.2 Statistics
Smith, G. D., Numerical Solutions of Partial Differential Equa
tions: Finite Differerence Methods, 3rd ed., Clarendon
Press, Oxford, (1985).
Thomee V., Galerkin Finite Element Methods for Parabolic
Problems, SpringerVerlag, Berlin, (1997).
Terzaghi, K., Theorectical Soil Mechanics, New York, Wiley,
(1943).
Turcotte. D. L. & Schubert, G., Geodynamics: Application
of Continuum Physics to Geological Problems, 1st. ed.,
John Wiley, New York, (1982).
Weisstein E. W., http:/ jmathworld.wolfram.com
Wikipedia, http:/ jen.wikipedia.com
Wylie C. R., Advanced Engineering Mathematics, Tokyo, (1972).
Versteeg H. K, Malalasekra W., 1vialalasekra W., An Intro
duction to Computational Fluid Dynamics: The Finite
Volume Method, Prentice Hall, (1995).
Yang X. S., Young Y., Cellular automata, PDEs and pattern
formation (Chapter 18), in Handbook of Bioinspired Al
gorithms, edited by Olarius S. and Zomaya A., Chapman
& Hall I CRC, (2005).
Yang X. S., An Introduction to Computational Engineering
with Matlab, Cambridge Int. Science Publishing, (2006).
Zienkiewicz 0 C and Taylor R L, The Finite Element Method,
vol. 1/11, IvicGrawHill, 4th Edition, (1991).
309
Appendix A
Mathematical Formulas
This is a summary for the mathematical formulas that have ap
peared in various sections in this book. We list these formulas
for your easy reference.
A.l Differentiations and Integrations
Differentiation Rules:
(uv)
1
= u
1
v + uv
1
I I
(
~ ) I = u v uv
v v2
{f[g(x)]}
1
= !
1
[g(x)] · g
1
(x)
Leibnitz's Theorem:
d ~ n (uv) = u(n)v + nu<n!)v' + ... + ( ; )u(nr)v(r)
+ ... + uv<n),
(
n) n!
r = r!(n r)!
Integration by parts
b dv b du
l
b
1 u dx dx = [uv] a + 1 v dx dx
311
A .2 Vectors and !via trices !v.Iathematical Formulas
Differentiation of an integral
d 1b(x) db da 1b(x) {)u(x, y)
d u(x,y)dy = [u(x,b)d u(x,a)d ]+
0
dy
X a(x) X X a(x) X
Power Series
~ z2 zn
e = 1 + z + 
21
+ ... +
1
... (z E C)
. n.
z3 zs z2 z4
sin z = z  3J + 5f  ... , cos z = 1  2f + 41  ...
z3 zs
sinh z = z + 3J + 5f + ... ,
Complex Numbers
ei
0
=cosO+ isinO, [ei1r + 1 = 0].
z = x + iy = rei
0
= r(cosO + isinO)
De Jvloivre's formula:
[r( cos 0 + i sin O)t = rn( cos nO+ i sin nO)
A.2 Vectors and Matrices
Dot Product
Cross Product
i j k
a x b = nlallbl sin 0 = Eijkajbk = a1 a2 a3
bt b2 b3
312
1\t/athematical Formulas A .2 Vectors and lvlatrices
Vector Triple
a1 a2 a3
[a, b, c] = a · (b x c) = bt b2 b3
CJ C2 C3
a· (b x c)= b · (c x a)= c ·(ax b)= a· (c x b)
ax (b x c)= (a· c)b (a· b)c
Divergence Theorem of Gauss
fffv V·udV= jfsu·dS
Stokes's Theorem
Green's Theorems
Identities
f
If
8v au
( udx + vdy) = (ax  ay )dxdy
v. v Xu= 0, v XV¢>= 0
V X (if>u) = (j>V Xu+ (V¢) xu
V · (¢u) = 4>V · u + (V4>) · u
V x (V x u) = V(V · u)  V
2
u
Inverse, Trace and Determinants
IAB .... ZI = IAIIBI ... IZI, IAI = detA
A vi= Aivi, eig(AB) = eig(BA)
313
A.3 Asymptotics !v.Iathematical Formulas
tr(AB) = tr(BA), tr(A +B) = tr(A) + tr(B)
1
det(A 
1
) = det(AB) = det(A)det(B)
det(A)'
Exponential Matrices
00 1 ~
etA= L
1
(tAt = l+tA+ 
2
A
2
+ ...
n=O n.
oo (1)n1 1 1
ln(IA) = L nl An= A 2A2 + 3A3 + ...
n=1
eAeB = eA+B (if AB = BA)
~ e t A = AetA = etA A
dt
(eA)1 = eA, det(eA) = etrA
A.3 Asymptotics
Gaussian Distribution
1 (x Jt)
2
p(x; Jt, a)= ~ exp[
2
2
] + <5(x)
27ra2 a
1
00 QX2d If
e x= 
oo 0:
l1+ 00.
1
00
2n ax2 (1)n ·1· 3· · · (2n 1) r:;r
oo x e dx =
2
n V 7n+T' (n > 0).
Binomial Distribution
B(k )
n! k(1 )nk (k 0 1 2 )
; n,p = (n k)tkt p  p ' = ' ' ' ... , n
314
1\t/athematical Formulas A .4 Special Integrals
B(x; n oo, p)l "'p(x; J.L, a), p, = np, a
2
= np(1  p).
np»l
Poisson Distribution
e.XAx
f(x;A) = 
1
, x EN
X.
f(x; A>> 1) ::::::: p(x; p,, a), p, = A, a
2
=A
B(k x; n oo,p) "'f(x; A= np), (lim np =A)
noo
A.4 Special Integrals
Gamma Function
r(x) = fooo txletdt = fooo et+(x1) Intdt
1 1
r(n + 1) = nl, r(
2
) = 2y'7r, r(
2
) = v'1r
r(x + 1)::::::: (=)xJ27rX, (x oo).
e
Stirling's Formula
nl::::::: n >> 1
e
Error Functions
315
Index
1D, 209, 254, 255
2D, 261
Airy stress function, 192
algorithms, 210
analytic function, 67
analytical function, 67
assembly by element, 246
asymptotic, 151
error function, 314
Gamma function, 314
Stirling's formula, 315
Bessel equation, 88
Bessel function, 24, 151
bifurcation, 99
binomial distribution, 281
birthday paradox, 273
BlackScholes equation, 205
boundary condition, 252
essential, 248
natural, 248
multiple variables, 165
pendulum, 159
shortest path, 156
central difference, 212
central limit theorem, 287
chaos, 99
complex integral, 70
residue, 70
complex variables, 62
coordinates
cylindrical, 15
polar, 15
spherical, 15
correlation coefficient, 295
cross product, 30
cumulative probability function,
284
curl, 38
curvature, 153
determinant, 49
difference equation, 95
calculus of variations, 153 differential operator, 84
brachistochrone problem, 164lifferentiation, 1
constraint, 160
curvature, 153
Dido's problem, 163
hanging rope problem, 163
316
implicit, 4
partial, 9
rule, 2
vector, 32
INDEX
diffusion equation, 146
displacement, 239
divergence, 38
divergence theorem, 313
dot product, 28, 30
DuFortFrankel scheme, 224
dynamic reconstruction, 102
dynamical system, 102
gradient, 38
Green's function, 148
Green's identity, 41
Green's theorem, 313
INDEX
harmonic motion, 109. 159
heat conduction, 139, 223, 253
hybrid method, 149
hyperbolic equation, 224
elastic wave, 203 firstorder, 214
elasticity, 181, 240 secondorder, 215
beam bending, 197 hyperbolic function, 65
CauchyNavier equation, 19J.hypothesis testing, 297
elastostatic, 185, 198
EulerBernoulli theory, 196 index matrix, 243
Hooke's law 181 inner product, 28
'
MaxwellBetti theorem, 185 integral
multiple, 12
strain tensor, 182
stress tensor, 182
stressstrain relationship, 184
elliptic equation, 218
Bessel function, 88
Cauchy's theorem, 71
differentiation, 12
Gaussian. 18
error function, 20
Euler scheme, 210
EulerLagrange equation, 154
exponential distribution, 286
finite difference method, 209
finite element method, 227, 248
finite volume method, 221
FokkerPlank equation, 205
Galerkin method, 238
Gamma function, 22
Gauss's theorem, 41
GaussSeidel iteration, 219
Gaussian distribution, 283
line, 38
residue theorem, 73
special, 17
integral equation, 153, 167
displacement kernel, 169
Fredholm equation, 167
sparable kernel, 169
Volterra equation, 168, 170
integral transform
Fourier, 125
Fourier transform, 144
Laplace, 131
Laplace transform, 143
wavelet, 134
317
INDEX
integration, 5
by parts, 6
iteration method, 219
Jacobian, 13
kinematics, 33
Lagrangian, 158
Lame constant, 184
Laplace's equation, 139, 201
Laurent series, 69
leapfrog scheme, 212
leastsquare, 238
Leibnitz theorem, 3
linear difference equation, 95
linear system, 56
lognormal distribution, 286
Lorenz attractor, 103
mathematical model, 201
matrix, 47
exponential, 52
Hermitian, 53
inverse, 50
Maxwell's equations, 204
mean, 278
INDEX
ordinary differential equation
complementary function, 81
general solution, 81
homogenous equation, 81
linear system, 85
particular integral, 81
oscillation
damped, 112
forced, 109
natural frequency, 112
small amplitude, 119
undamped, 109
outer product, 30
parabolic equation, 202, 216
pattern formation, 260
bifurcation, 264
instability, 263
pattern formation , 261
PDE, 138, 141, 203, 213
Poisson distribution, 281
Poisson's equation, 201, 244
probability, 267
axiom, 270
conditional, 271, 275
distribution, 279
method of least square, 292
moment generating function, 280
independent events, 271
moment, 280
NavierStokes equation, 206
NavierStokes equations, 206
node, 242
normal distribution, 283
normal modes, 116
null hypothesis, 301
ODE, 77, 80, 81
l\1onty hall problem, 276
permutation, 272
random variable, 268, 277
randomness, 267
probability density function, 283
quadratic form, 55
random variables, 277
318
INDEX
reactiondiffusion, 257, 262
reactiondiffusion equation, 204
recurrence equation, 95
Residue theorem, 73
Riccati equation, 77
Riemann (function, 68
Rien1ann hypothesis, 69
RungeKutta method, 210, 213
Saturn's rings, 43
Schrodinger equation, 206
selfsimilarity, 105
separation of variables, 141
series
asympottic, 17
power, 8
Talor, 8
shape functions, 238
similarity solution, 145
SineGordon equation, 207
soliton, 147
INDEX
SturmLiouville equation, 86
Taylor series, 69
tensor, 173
analysis, 175
Cartesian, 175
notations, 173
rank, 175
vector, 176
three card problem, 276
timedependent, 251
timestepping, 217, 253
explicit, 254
implicit, 211
transient, 254
Travelling wave, 147
travelling wave, 259
triangular element, 240
uniforn1 distribution, 286
upwind scheme, 214
stability condition, 211, 215
variance, 278
standard normal distribution,
vector, 27, 29
285
statistics, 289
confidence interval, 297
linear regression, 293
maximum likelihood, 292
sample mean, 289
sample variance, 289
steady state, 245
stiffness matrix, 242
Stokes's theorem, 41
stress intensity factor, 195
Student's tdistribution, 299
Student's ttest, 301
triple, 31
vector calculus, 32
Venn diagram, 268
vibration, 109
wave equation, 139, 202, 203,
215, 255
weak forn1ulation, 236
Young's modulus, 181
319
Applied Engineering Mathematics
Applied Engineering Mathematics
XinShe Yang
University of Cambridge, Cambridge, United Kingdom
CAMBRIDGE INTERNATIONAL SCIENCE PUBLISHING
Published by Cambridge International Science Publishing 7 l\!Ieadow Walk, Great Abington, Cambridge CBl 6AZ, UK http:/ fwww .cisppublishing.com
First Published 2007
@Cambridge International Science Publishing @XinShe Yang
Conditions of Sale
All rights reserved. No part of this publication may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopy, recording, or any infornlation storage and retrieval system, without permission in writing from the copyright holder. British Library Cataloguing in Publication Data
A catalogue record for this book is available from the British Library
ISBN 9781904602569 Cover design by Terry Callanan
Printed and bound in the UK by Lightning Source UK Ltd
Preface
Engineering mathematics including numerical methods and application is the essential part of key problemsolving skills for engineers and scientists. Modern engineering design and process modelling require both mathematical analysis and computer simulations. Vast literature exists on engineering mathematics, mathen1atical modelling and numerical methods. The topics in engineering mathematics are very diverse and the syllabus of mathematics itself is evolving. Therefore, there is a decision to select the topics and limit the number of chapters so that the book remains concise and yet comprehensive enough to include all the important mathematical methods and popular numerical methods. This book endeavors to strike a balance between mathematical and numerical coverage of a wide range of mathematical methods and numerical techniques. It strives to provide an introduction, especially for undergraduates and graduates, to engineering mathematics and its applications. Topics include advanced calculus, ordinary differential equations, partial differential equations, vector and tensor analysis, calculus of variations, integral equations, the finite difference method, the finite volume method, the finite element method, reactiondiffusion system, and probability and statistics. The book also emph~ sizes the application of important mathematical methods with dozens of worked examples. The applied topics include elasticity, harmonic motion, chaos, kinematics, pattern formation and hypothesis testing. The book can serve as a textbook in engineering mathematics, mathematical modelling, and scientific computing. XinShe Yang Cambridge, 2007
A C :rvlclntosh. Seffen. Dr J l\II Lees and Dr C T Morley at Cambridge University. I also thank Dr G. I thank many of my students. XinShe Yang ii . Helen. I also would like to thank my publisher. Love. R W Lewis. I thank my wife. C J :Mcdiarmid at Oxford University. Prof. J Brindley. Parks. I would like to thank my mentors. and Prof. Dr Victor Riecansky. Dr S. and son. A C Fowler and Prof. Dr K.Acknowledgements First and foremost. Prof. for his kind help and professional editing. and many colleagues for their inspiration. at Cambridge University who have indirectly tried some parts of this book and gave their valuable suggestions. tutors and colleagues: Prof. Dr T. D T Gethin. especially Hugo Whittle and Charles Pearson. Prof. for their help and support. Last but not least. Guest. Andre Revil for their help and encouragement. Young. Prof.
and conference proceedings. solar eclipses. published in 2006 by Cambridge International Science Publishing Ltd.Phil in applied n1athematics from the University of Oxford. His research interests include asymptotic analysis. book chapters. iii . bioinspired algorithms. He is also the author of a book entitled: "An Introduction to Computational Engineering with Jvlatlab".About the Author XinShe Yang received his D. Dr Yang has published extensively in international journals. computational engineering. He is currently a research fellow at the University of Cambridge. engineering optimization. combustion. scientific programming and pattern formation.
.
.4.1.1 Differentiations .1 Asymptotic Series 1.2 Integrations .... . . . 1.1 Vectors . 1. .2 Jacobian ... . 1.3. . .. ..5..4..3 Taylor Series and Power Series 1. .. . .2 Gaussian Integrals 1.5. . 1.. . .4 Multiple Integrals .. 1. .3 In1plicit Differentiation .2. 1. ..Contents 1 Calculus 1.2 Differentiation Rules 1.5.3 Error Functions ..2. 1..1... 1. 2. .1 Partial Differentiation .2 Integration by Parts . 1.5.1 Definition . .3 Partial Differentiation . .. .1. .1 l\1ultiple Integrals 1.. .2. ... . .. .2 Differentiation of an Integral 1. 1. .1 Definition .1...1 Dot Product and Norm v 27 27 28 . .3.5. . . .5 Bessel Functions 1 1 1 2 4 5 5 6 8 9 9 12 12 12 13 16 17 18 20 22 24 2 Vector Analysis 2..4 Gamma Functions 1. 1.5 Some Special Integrals . .. 1. .
. . . 5. 2. 4. . 2. . .3 Analytic Functions 4. .2 37 38 40 41 41 42 2.2.1 :rviatrix .4 Three Basic Operators . . .3 3 Matrix Algebra 3. .1. .3. 3. . . .1. .2. 5. .2. 2. .3 Inverse . 4..4 Complex Integrals . . .2.1 Conservation of 1vlass 2.2 Kinematics . . 3. 2. .2 Hyperbolic Functions .1 Complex Numbers and Functions .1 Linear ODEs . . .1 Introduction .3 Vector Triple .5 Hermitian and Quadratic Forms 3. 5. . .2 Determinant. .2 First Order ODEs . .2 Saturn's Rings CONTENTS 30 31 32 32 33 2.2.2 Nonlinear ODEs 5. . . . .3 Higher Order ODEs ..1 General Solution 5. . . 2. .4 :rviatrix Exponential.4 Linear System . . 5. 3.3 Line Integral .6 Solution of linear systems 4 Complex Variables 4.5 Son1e Important Theorems Applications . . .CONTENTS 2.3. .2 Cross Product 2.5 SturmLiouville Equation vi 77 77 78 78 80 81 81 84 85 86 . Vector Algebra .2. . 47 47 49 50 52 53 56 61 61 65 67 70 5 Ordinary Differential Equations 5. .2.1 Differentiation of Vectors 2. . . 5.3.. 5.2 Differential Operator . 3.3. .
112 . .1. .3 Normal Ivlodes .5. 116 . .1 Separation of Variables . . 102 .1 Undamped Forced Oscillations 7.3 Selfsimilarity and Fractals . 6. .5. 10 Techniques for Solving PDEs 10.2 5. 7. .2 Classification 9. 6.1 Fourier Transform 8. 145 . 139 . Euler Buckling . . 7 Vibration and Harmonic Motion 7.3 Similarity Solution . . . .2 Chaos and Dynamical Systems 6. 143 . 126 126 128 129 131 134 137 . . . . . 9 Partial Differential Equations 9.3 ~avelet . 147 .3 Lorenz Attractor . . 141 .1.1 Linear Difference Equations 6. 7. 10.3 Fourier Transform 8. .1 First Order PDE 9. vii 95 95 98 99 . 119 125 . .1. . . 139 141 .2 Damped Forced Oscillations .2.CONTENTS 5. 103 . .4 Small Amplitude Oscillations 8 Integral Transforms 8. . . .3 Bessel Equation . 8. .5. 8.2. . . .2. . . 10. . . . .1 Fourier Series . .1 5. . .1 Bifurcations and Chaos 6. . . 8.2 Transform l\1ethods . . 105 109 . 138 . . . .2 Laplace Transforms. 10. . .2 Fourier Integral . Nonlinear SecondOrder ODEs CONTENTS 88 90 91 6 Recurrence Equations 6.4 Travelling ~ave Solution .. .3 Classic PDEs . 109 .. . .2 Dynamic Reconstruction.
14. .3.2 Tensors . 189 .2.5 EulerBernoulli Beam Theory 14 Mathematical Models 14.3 Tensor Analysis . . . .2 l\tlaxwell's Equations viii 153 . .2 Displacement Kernels 11.1.2 Other PDEs . . .1. 167 . 203 . . . . .1 Separable Kernels . 11. 174 . 11.2.3. .2 EulerLagrange Equation 11.2 l\1axwell's Reciprocal Theorem 13. 12. 12. 196 201 . .1 Classic l\1odels . .3 Equations of l\1otion .2 Integral Equations . .. 201 . 149 11 Integral Equations 11.3. 148 . . . 13 Elasticity 13.3 Solution of Integral Equations . . 169 . .1 Laplace's and Poisson's Equation .4 Airy Stress Functions .1 Linear Integral Equations 11.CONTENTS 10. . 153 . . 13. . 11.1 Hooke's Law and Elasticity 13.1. 170 .1. . 204 .1 Calculus of Variations . . 14. . . . . 170 173 . 14.5 Green's Function 10. . 153 . . .1 Curvature .3 Volterra Equation 12 Tensor Analysis 12. . . 14. . . 13. 202 . . . 167 . 203 . . . . .3 Wave Equation . 181 . .3 Variations with Constraints 11. 14. .1. 160 .1 Elastic Wave Equation . 11. . 14. 169 .6 Hybrid Method . . . 11. . CONTENTS . . . 202 .2. . . . 173 .2 Parabolic Equation . 165 .1. 203 .4 Variations for l\1ultiple Variables 11.1. . 175 181 . . 185 .1 Notations . . . . 192 . 154 . .
1 Introduction . ix .6 14. 222 . 17.2. 221 .4 Hyperbolic Equations 17 Finite Element Method 17. . 236 .1. . .3 RungeKutta Jviethod 15. . 17. 239 . 17. NavierStokes Equations . . .7 14. 235 . . 244 15 Finite Difference Method 15. 210 .2 SecondOrder Wave Equation 15.2 Implementation .2. . 207 209 . . 235 . . 213 . 17. .4 14.3 Parabolic Equations 16.1 Integration of ODEs . 209 . 17.2. . 239 .2 LeapFrog Jviethod 15. .4 Heat Conduction .2.2 Hyperbolic Equations . 15.1.2.2.2. 218 221 . 205 .3 14. 242 . 17. 228 .1 Euler Scheme . . . 215 . 16.8 ReactionDiffusion Equation. 228 . 212 . ..3. SineGordon Equation CONTENTS .2 Bar and Beam Elements . . 223 . 16. 232 . .3 Parabolic Equation . 16 Finite Volume Method 16. . . 206 .4 Elliptical Equation .2. .CONTENTS 14. 214 .2.1 Concept of Elements .1 Simple Spring Systems . 204 . 15. .1. .1 Weak Formulation 17. 237 .2 Elliptic Equations . 213 .2.3 Elasticity .2 Finite Element Formulation 17. 205 . FokkerPlank Equation BlackScholes Equation .1.2 Galerkin Jviethod 17. 224 227 . 216 .5 14.1 Plane Stress and Plane Strain . 15.2.1 FirstOrder Hyperbolic Equation 15. Schrodinger Equation . . 17.3. .3 Shape Functions . 206 . 15.1.
.6 Other Distributions . 257 .1. 19. 283 . . 287 .2 Iviethod of Least Squares 19. . 260 . . .2. . . 267 . .1 Randomness and Probability 19. . 19. . 17.4.2 Pattern Formation .5 TimeDependent Problems . 253 . . . 281 . . 18. .1. .2 ElementbyElement Assembly .7 The Central Limit Theorem . . . . .1. . 315 17. 267 . 254 257 .5. . 17. A Mathematical Formulas A. 257 . 312 . . 297 311 .1. . . 275 . 18 Reaction Diffusion System 18.1 Probability .3 Random Variables and Ivloments 19.2 Nonlinear Equations .1 Basic Formulation . .3 Hypothesis Testing . .2.2 Conditional Probability . 244 .2. 290 . 263 267 .5. . . .1 Travelling Wave . .1.4 Special Integrals . . .2 Statistics .2 TimeStepping . . 19. 17. 17. .5. .3 Asymptotics A. .2. . . . .2. 18. . X .1 Fundamental Solutions .5 Gaussian Distribution . . 259 .CONTENTS CONTENTS . . . 251 . 19. 286 . . 17. . A. . 18. . 246 . . 292 . . 19. 251 . . 259 .5. 277 . . . 289 . .4. 248 . . . 314 . 311 .. 19.3 Application of Boundary Conditions 17.3 1D Transient Heat Transfer . 18.1 Heat Conduction Equation 18.4 Wave Equation . . .3 ReactionDiffusion System .1. 17.2 Vectors and Matrices . . . .1.1. 19. 19 Probability and Statistics 19.1 The Time Dimension. . 253 .4. .1 Sample Ivlean and Variance 19.1 Differentiations and Integrations A.4 Binomial and Poisson Distributions.
Therefore. This gradient or limit is the first derivative of the function f (x) at P. y) is defined as dy dx = df(x) = !'(x) = dx lim f(x .dx+0 + ~x). ~x (1.1. Thus. the slope or the gradient of the curve at the point P( x.f(x).Chapter 1 Calculus The prelin1inary requirements for this book are the precalculus foundation mathematics. then we say that the function is nondifferentiable at P. 1. and readers can refer to any book that is dedicated to these topics. the above definition can also be written 1 . the limit of the infinitesimal change ~x is denoted as the differential dx. By convention. 1) on the condition that there exists such a limit at P. we will only review some of the basic concepts of differentiation and integration.1.1 1. If the limit does not exist at a point P. We assume that the readers are familiar with these preliminaries.1 Differentiations Definition For a known function or a curve y = f (x) as shown in Figure 1.
and the choice is purely out of convenience..=l(n) dxn lf'ly = dl(nl) . (1.1: Gradient of a curve The second derivative of l(x) is defined as the gradient of l 1(x).2) which can be used to calculate the change indy caused by the small change of dx.4) 1.1.3) The higher derivatives can be defined in a similar manner. we can derive a differentiation rule using the definition from the first princi 2 .. d y dx 3 3 =I"'( ) =  x dl"(x) dx ' . 2 Differentiation Rules If a more complicated function I (x) can be written as a product of two simpler functions u(x) and v(x). . . or 2 d y =I"( ) = dl(x) dx2x dx · (1. The primed notation 1 and standard notation d~ can be used interchangeably. dx (1. Figure 1.1 Differentiations as Calculus dy = dl = ~dx dl(x) =I1(x)dx. Thus.1.
u(x)v(x) and subtracting and adding u(x + ~x)v(x) [= 0] terms. (1... Using f(x + ~x)  f(x) = u(x + ~x)v(x + ~x) . we have + u(x + ~x)v(x) df d[u(x)v(x)] dx = dx . we can gen eralize this rule. f(x) = eg(x) and g(x) = xn].7) where the coefficients are the same as the binomial coefficients nc = ( r n) = rl(nr)!' nl r (1.5) which can be written in a contract form using primed notations J'(x) = (uv)' = u'v + uv'.1 Differentiations pies. we finally get the Leibnitz's Theorem for differentiations + .Calculus 1. + uv(n). [ ( A )v(x + ~x).6x+O ~g ~X' ~~ ~g (1.9) 3 . or f(x) = f[g(x)] [for example. f(x) = exn] can be written as a function of another function g(x).u(x)] I1m ux+~x +vx ~ ~ ~~ X = u(x) dx dv + dx v(x).6) If we differentiate this equation again and again.v(x) ( )u(x + ~x). du (1.l .8) If a function f(x) [for example. then we have )! '( X I' i D . (1.
we have 2y ~~ + cos(x)eY + sin(x)eY: = . Differentiating term by term with respect to x. 2: Find the derivative ::. 1. dgdx (1.1.12) 0 Example 1.1.cos(x) = + sin(x)[ecos(x))' + sin(x)e.1: The deriva.1.6) and (1/v)' = v1 fv 2 .10) (1. If one use 1/v instead of v in the equation (1. In this case.3 Implicit Differentiation The above differentiation rules still apply in the case when there is no simple explicit function form y = f (x) as a function of x only.tive of f(x) = sin(x)ecos(x) can be obtained using the combination of the above differentiation rules. we have the following differentiation rule for quotients: (!)I v = u v . 0 Example 1. y + sin(x) exp(y) = 0.cos(x))' cos(x )e.cos(x). we can differentiate the equation tern1 by tern1 with respect to x so that we can obtain the derivative dyfdx which is in general a function of both x and y. For example.11) {f[g(x)]} 1 = J'[g(x)] · g1(x). we have f' (x) = (ex")' = ex" nxn l.cos(x)[. In our example.sin(x). f'(x) = [sin(x))'ecos(x) = cos(x)e. v2 I I (1. if y 2 + sin( x )eY = cos( x). The derivatives of various functions are listed in Table 1.uv .1 Differentiations Calculus which leads to the following chain rule f 1(x) or = df dg. dy dx cos(x)eY + sin(x) 2y + sin(x)eY · 0 4 .cos(x) + sin2(x )e.
1 Integrations Definition Integration can be viewed as the inverse of differentiation.13) F(x) = 1x !(~)~. and the integration starts from xo (arbitrary) to x.2 Integrations Table 1. ~) in the integrand.1 x tan. As we know.14) where f (x) is called the integrand. the 5 .Calculus 1. xo (1.2. The integration F( x) of a function f (x) satisfies d~~x) or = f(x).1: First Derivatives f(x) xn ex ax(a > 0) lnx logax sinx cosx tanx sin. In order to avoid any potential confusion. it is conventional to use a dummy variable (say.v'1x 2 1 1+x2 coshx sinhx 1. (1.1 x cos.1 x sinhx coshx f'(x) nxn1 ex axlna ! X 1 xlna cosx sinx sec 2 x 1 ~ 1 .2 1. the geometrical meaning of the first derivative is the gradient of the function f (x) at a point P.
Integration by Parts From the differentiation rule (uv )' = uv' + u' v.F(a). lb 1b a a X X 6 .15) The difference F (b) . (1.2. As F'(x) = f(x).1.19) in the indefinite form. Integrating both sides. we can also write the above equation as = l f(x)dx = l F'(x)dx = F(b) .2 Integrations Calculus geometrical representation of an integral f(~)d~ (with lower integration limit a and upper integration limit b) is the area under the curve f(x) enclosed by xa. It can also be written in the definite form as dv du (1.18) I b 1 a dv udx = uv .20) ud dx = [uv] + vd dx.17) where cis the constant of integration.xis in the region x E [a.u'v. In this case.ldu vdx dx dx ' (1. we have la J: fb f(x)dx = 1b f(x)dx 1a f(x)dx = F(b).F(a). (1.F (a) is often written in a compact form Fl~ F(b). we have (1. When the lower limit is not explicitly given. the integral is called a definite integral as the limits are given.2 f(x)dx = F(x) + c. The integrals of some of the con1mon functions are listed in Table 1. we have uv' = (uv)'.F(a). the integral is called an indefinite integral I 1. the change or shift of the lower limit will lead to an arbitrary constant c.2. b]. For the definite integral. xo xo (1.16) Since the lower limit x 0 is arbitrary.
1 ~] coshx sinhx lncoshx The integration by parts is a very powerful method for evaluating integrals.a2) sinhx coshx tanhx (or cosh.fany complicated integrands can be rewritten as a product of two simpler functions so that their integrals can easily obtained using integration by parts. a a 1 In a+x 2a 2a 1 .Calculus 1. 0 Example 1.1 i!. Hence. 1\.dx 2 2 X J J J 0 7 .1 ~ a ax 1 v'a1x1 1 ~ 1 Jx2a2 ln(x + v'x2 + a2) [or sinh.1 ~] ln(x + v'x2 .2: Integrals f(x) xn(n tf 1) 1 X f f(x) xn n+1 lnlxl ex ex sinx cosx 1 a2tx2 a2x2 x2a2 cosx sinx 1 tan..lnx.!.In~ x+a sin.9: The integral of I= x lnx dx can be obtained 2 by setting v' = x and u = lnx.. v = x2 and u' = ~· \Ve now have 2 2 x x 1 I= xlnxdx = .2 Integrations Table 1.
we have the firstorder approximation to f (xo + h) f(xo +h)~ f(xo) + hf'(xo). If we repeat the procedure for f'(x). any change from xo to xo+h is approximated by a linear term hf'(xo).21) and 1:£ = F' = f(x). (1.F(a). /'(x)dx = f(xo +h). (1. (1. we can approximate the integral as 1 xo xo f'(x)dx ~ !'(xo)h. we have J'(xo +h)~ !'(xo) + hf"(xo). Following the same procedure for higher order derivatives.24) Thus.2. we have (1. (1. 1.23) If h is not too large or f' (x) does not vary dramatically.26) which is a better approximation than f'(xo +h) ~ f'(x 0 ).2 Integrations Calculus Other important methods of integration include the substitution and reduction methods. we can reach the nth order approximation f(xo +h)= f(xo) + hf'(xo) + 8 ~~ !"(xo) + .3 From Taylor Series and Power Series l 1 xo xo+h f(x)dx = F(b) .22) which means that f(xo +h) = f(xo) + 1 xo xo+h f'(x)dx. Readers can refer any book that is dedicated to advanced calculus.f(xo). (1.25) This is equivalent to say.1.~ !"'(xo) .
the function l(x. we have implicitly assumed that all the derivatives l'(x). In deriving this formula. 5! x4 xs (x E 'R). this is indeed the case..1 Partial Differentiation Partial Differentiation The differentiation defined above is for function I (x) which has only one independent variable x.3. and the gradient will generally depend on the location x.. (x E 'R). The 9 .27) where Rn+t (h) is the error of this approximation and the notation means that the error is about the same order as n + 1th term in the series.. . the gradient or rate of change will also depend on the direction (along xaxis or yaxis or any other directions). The end results are the Maclaurin series.2 has different gradients at (0. 0) along xaxis and yaxis.Calculus 1. l(n)(x) exist. (1. y) = xy shown in Figure 1. This is the wellknown Taylor theorem and it has many applications. In almost all the applications we meet. (1. . we then reach the infinite power series and the error limnoo Rn+ 1 + 0 if the series converges.30) cosx = 1 .. .28) sin x = x . sin( x) and ex. . l"(x).. For example. For example. all the orders of the derivatives exist. If we continue the process to infinity.29) (1..3! x3 x2 + . their gradient will depend on both x andy in general. For functions I (x. 1.. In addition..3 Partial Differentiation (1.. For example.3 1. y) of two variables x andy.+2! 4! and .
6. y + ~y) .yayx lim f(x. af(x. :Mathematicians like to 10 .yO .y) = f = 8!1 ay . we have af(x.f(x. y). we have U %. gradients along the positive x.xO. y) ~X (1. Thus.32) Similarly. y) ~y .33) The notation fxly emphasizes the fact that y is held constant. The partial derivative of f (x.directions are called the partial derivatives respect to x and y.y) = f = 8!1 8x  x 8x Y = . They are denoted as and respectively.1. y) with respect to x can be calculated assuming that y =constant. y) = xy.2: Variation of f(x.6.f(x.3 Partial Differentiation Calculus 1 1 Figure 1.x=const (1. respectively.and y. The subscript notation fx (or /y) emphasizes the derivative is carried out with respect to x (or y).y=const lim f(x + ~x.
then we have the generalized chain rule . the total infinitesimal change df can be written as (1.av dx 11 dg ag ag du ag dv (1.. and fxy = fyx = 1. u. (1. The second partial derivative of f(x. For example. we have fxy = !yx· 0 Example 1. then the above equation leads to the following chain rule ~ df af dx = 8xd~ + 8y~' af dy (1.36) which is very useful in calculating the derivatives in parametric form or for change of variables.sin(x)eY.cos(x)eY.f(x. 0 For any small change Llf = f(x+flx. .. y) due to flx and fly.3 Partial Differentiation use the subscript forms as they are simpler notations and can be easily generalized. y) is fxx =.34) Since flxfly = flyflx.35) If x and y are functions of another independent variable ~. If a complicated function f (x) can be written in terms of simpler functions u and v so that f(x) = g(x.Calculus 1. y) = xy + 8! + Cos(x)eY.4: sin(x)eY are fx = ax = y The first partial derivatives of f(x. v) where u(x) and v(x) are known functions of x.37) .=ax +au dx + dx. y+fly).
T)dT = w[x. v(x)]. (1.ou . y) dy. 1. we have the total differential as df = op dt 8t + op dx + op dy + 8Pdz. this 12 .39) with fixed integration limits a and b. u(x)] . I(x) = ru(x) lv(x) 1/J(x.=8x + dx + dx' . t) such as the pressure in a fluid.4.38) 1. For a function p(x. X X du dv (1. (1. y)dy.4 !v!ultiple Integrals Calculus The extension to functions of more than two variables is straightforward. v(x))d ].w[x.2 Differentiation of an Integral When differentiating an integral <11(x) = l Ja <P(x.ov dx where ~ dl ol ol du 8 I dv (1. 8x oy 8z (1. we have 84>(x) = [b 8</J(x.41) the following formula is useful: dl d = X 1u(x) v(x)  o'l/J dT 8X + [1/J(x.1. z.3.42) This formula can be derived using the chain rule .4 1.1 Multiple Integrals Multiple Integrals As the integration of a function f (x) corresponds to the area enclosed under the function between integration limits. OX OX (1.40) When differentiating the integrals with the limits being functions of x. u(x))d 1/J(x. v(x)). y.43) = 1/J(x. u(x)) and §f = 1/J(x.
45) 0 Example 1.4.2 Jacobian Sometimes it is necessary to change variables when evaluating an integral. The simplest form of dA is dA = dxdy in Cartesian coordinates. (1. y)dxdy. Here the extra factor dxjdv in the integrand is referred to as the Jacobian. dv a (1.4 A1ultiple Integrals can extend to the double integral and multiple integrals. For a simple onedimensional integral. the double integral is defined as F =In f(x. In order to emphasize the double integral in this case. with the length 2a and the width 2b.44) where dA is the infinitesin1al element of the area.Calculus 1. and n is the region for integration. y)dA. and we have 1 Xa xb f(x)dx = 1b f(x(v))d dv. This is relatively sin1ple as dv = ~~dv. is defined by I= J In y dS = J In y dxdy. (1. and we have 0 1. the change of variables from x to a new variable v (say) leads to x = x(v). the integral is often written as I = J In 2 f(x. For a function f(x.46) X where the integration limits change so that x(a) Xa and x(b) = Xb. 2 The plate can be divided into four equal parts. 13 . y). 5: The area moment of inertia of a thin rectangular plate.
it is more complicated. 17).47) where J is the Jacobian. y) to polar coordinates (r. an integral in (x. 14 (1. 8)./Cd = I ax ay . Thus. 1J) is just a useful shorthand.4 !v!ultiple Integrals Calculus x= For a double integral. J 0 In a similar fashion. y) will be transformed into /! V (j>(x.50) . This equivalent to say that the change of the infinitesimal area dA = dxdy becomes   I 0~ §JL {)x ~ {)x "lJ7i §JL I. y = rsin8. y = y(~. 17).I 0~ {)x {)x §JL 0~ {JiJ I• 0~ ~ d d = I a(x. y) a(~. (1.y. the Jacobian is J _ a( X.48) "lJ7i The notation a( x. y)dxdy jj !(~1J)IJid~d1J.ar a8.1.z)dxdydz = JL 1P(~.y)dxdy = J 4>(rcos8.ax ay 1.49) 0 Example 1.1J) (1. Assuming x(~. y) 1. That is J = a(x.8) . y) _ ax ay ax ay . the change of variables in triple integrals gives = jjfn c/J(x.rsin8)rdrd(). = (1.1J. 1J) ~ 1J a~ a1] a1] a~ ~ 1]. 6: \Vhen transforming from (x.a(r. we have the following relationships x = rcos8. we have jj f(x./Cd X y a(~.()IJI~d1]d(. y) /a(~.ae ar = cos 8 x r cos() . Thus.( r sin 8) x sin 8 = r[cos 2 8 + sin 2 8) = r.
51) F F 0~ 0~ 0< 0< m: For cylindrical polar coordinates (r. 8(r.53) z Figure 1. (1. (1. 4>) as shown in Figure 1. (1. The Jacobian is therefore J y = rsin¢.3: Cylindrical polar coordinates. and 4> is the azimuthal angle. the Jacobian is J = sin 8 cos 4> sin 8 sin 4J cos 8 r cos 8 cos 4J r cos 8 sin 4> r sin 8 r sin 8 sin 4J r sin 8 cos 4J 0 15 = r 2 sin 8.54) Therefore. Z) = rsin¢ rcos¢ 0 =r. where 8 is the zenithal angle between the zaxis and the position vector r.4.3. Z =Z.52) = 0( X. 1J. y.z) 0 0 1 cos 4> sin¢ 0 (1. z = r cos 8. () J = {)(X. y. (jJ. y = r sin 8 sin¢. For spherical polar coordinates (r.(jJ. we have x = rsin8cos¢. we have {)(~. 8.Calculus 1. z) as shown in Figure 1.4 !v!ultiple Integrals and (1. Z) = Ox ~ Ox iJ1] Ox ~ Oy 0::: ~ 0::: x = rcos(jJ.55) .
4: Spherical polar coordinates.5 Some Special Integrals Calculus z Figure 1. Since the infinitesimal volume element dV = r 2 sin( O)drdOd¢ in spherical coordinates r ~ 0.1. 7: defined as = r 2 sin OdrdOdc/>.56) The volume of a solid ball with a radius R is V= !!In dV. the ball can be divided into two equal parts so that V = 2 Jo = 2 {R {rr/2 {}0 {21r sin0[} 0 d¢]dO}dr Jo {R {27r Jo {rr/2 sin(O)dO}dr 0 1.5 Some Special Integrals Some integrals appear so frequently in engineering mathematics that they deserve special attention. the volu1ne element change in the spherical systen1 is dxdydz 0 Example 1. (1. Thus. 0 ~ () ~ 1r and 0 ~ ¢ ~ 21r. l\1ost of these special 16 .
though they might give the same results. 17 (1. That is + f g 0. The notation ~ does not necessarily mean ~ in general. sin x ~ x and sin x ~ x if x + 0].58) The big 0 notation means that f is asymptotically equivalent to the order of g(x).5 Some Special Integrals integrals are also called special functions as they have certain varying parameters or integral limits. Obviously. In this special case. X+ XQ. x 0 can be any value. The small o notation is used if the limit tends to 0. including 0 and oo. we write f = O(g).57) where K is a finite.59) which is equivalent to f fg + 1 and gf f+ 1 as x + xo. this does not mean f ~ 100.60) or f = o(g). If the limit is unity or K = 1.5. We only discuss four of the n1ost common integrals here. (1. nonzero limit. for two functions f (x) and g( x). When we say f is order of 100 (or f ~ 100).Calculus 1. especially in the case when x + 0 [for example. we say f(x) is order of g( x).61) .1 Asymptotic Series Before we discuss any special functions. (1. 1. let us digress first to introduce the asymptotic series and order notations because they will be used to study the behaviours of special functions. we write J~g. Loosely speaking. (1. (1. if f(x) g(x) + K ' X+ XQ. but it can mean that f is between about 50 to 150.
. which is still very large.1. However. we have n! = 9. Another classical example is the Stirling's asymptotic series for factorials n >> 1. Since the Gaussian integral is a definite integral and must give a constant value.63) This is a good example of asymptotic series.. while the leading approximation is V21m(nfe)n = 9.1rn(.5.5 Some Special Integrals Calculus If g > 0. It can be defined by /(a) = ! oo ea:x 00 2 dx. we can change the dummy variable as we wish.3326 x 10 157 . We have / 2 =[ 00 100 100 1oo ea:x dxf = oo ea:x dx oo ea:y dy 2 2 2 18 .. (1.3248 x 10157 .2 Gaussian Integrals The Gaussian integral appears in n1any situations in engineering mathematics and statistics. though three orders smaller than the leading approximation.64) In order to evaluate the integral. for n = 100. (1. the error of the truncated series Rk decreases and gets smaller con1pared with the leading term [here V21m(nfe)n].::. The difference between these two values is 7.~ n n 1 v .62) In fact.51480n3 . let us first evaluate / 2 . In fact. it can be expanded into more terms nl 'V .. 7740 x 10154 . R2 = V21m(nje)n is still very large as R2 + oo if n >> 1. for Vx E 'R. Rn does not necessarily tend to zero. f = o(g) is equivalent to f << g. For example. For standard power expansions. but for an asymptotic series.) ( 1 + 12n + 288n2 1 139 . ). (1. the error Rk(hk) + 0. For example. tin· 1. we have ex~ 1 + x + O(x 2 ) ~ 1 + x + ~ + o(x).
a a 00 (1. 19 (1.67) n times with respect to a.65) Changing into the polar coordinates (r.69) where n > 0 is an integer. a Since a is a parameter. a) is a zeromean Gaussian probability function. 0) and noticing r 2 = x 2 + y 2 and dxdy = rdrdO. we can differentiate both sides of this equation with respect to a.71) . and we have oo ()= 1 00 ! X 2e oo ox2dX =1 ~ .67) can be rearranged as 00 / 00 f(x. For a special case when a = ~ and n = 0. for xi= 0. we have I2 = roo dr Jo rear dO [21r 2 = 21r Therefore.Calculus 1. As a~ 0. a)dx = 1.68) By differentiating both sides of the Gaussian integral (equation 1.70) The function f(x. and we get the generalized Gaussian integral In In = 1 00 X 2n e ox 2 oo (1. 2a a (1. f(x) ~ 8(x) where 8(x) is the Dirac 8function which is defined by 8(x) i= 0 (at x = 0). the equation v2u"" (1.67) eax2dx = ~ . I a loo oo 1 7r e ar2d( ar 2) = . but 8(x) = 0. (1.66) (1. (1. 5 Some Special Integrals = i:i: Jo eo(x2+y2)dxdy.
we have erf(O) 0.78) . which appears frequently in heat conduction and diffusion problems. and erf( oo) = 1.. Error Functions The error function. .1. Using Taylor series for the integrand e _112  .73) where f(x) is a function.. (1.:i.erf(x) = y'i 1 x 00 et2 dt.5.1 .3 f(x)o(x. is defined by erf(x) = v'i Jo 2 rx e 11 2 d'f/.5 Some Special Integrals Calculus and 00 /00 o(x)dx = 1. Both the error function and its complementary function are shown in Figure 1.75) The error function is an odd function: erf( x) = erf(x).]. (1.(3)dx = !((3)...6"' + . (1.74) Its complementary error function is defined by 2 erfc(x) = 1.72) It has an interesting property that J 1. 1 4 1 6 (1. Using the results from the Gaussian integral 1 00 e 11 2 d'TJ = ..77) and integrating term by term.'TJ 2 + 2"' .42 + . (1. The error function cannot be easily evaluated in closed form.3 20 + x5 x 7 10.5. (1. we have 2 x3 erf(x) = yli"[x.76) oo together with the basic definition. (1.
79) The integrals of the complementary function are defined by ierfc(x) = and (1. or erf(x) =  n x2n+ L 1.fi n=O 2n + 1 n. oo 1 2 ( 1) (1.Calculus 1..5 0 0. if we replace x in the error function by lim [3oo 2 [1 + erf(/3x)] 1 ? H(x). (1.. .84) .~o 5 0 X 5 10 Figure 1.5 Some Special Integrals 2. we can prove the following asymptotic series ex2 erf(x) {3x.82) On the other hand.81) Loo erfc(TJ)dTJ..!So. (1. H(x) 21 = 0 (for x < 0). (1. "t: 0. (1.' Xy7r (x? oo).5  erf(x) .80) Using integration by parts.erfc(x) CD 1 .5: Error functions.83) where H (x) is a Heaviside function or a unit step function which is defined by H(x) = 1 (for x > 0).5 1~ 1. we have "J 1 r..5r~~r=========~ 2 1.:.
For n >> 1. If X= n is an integer (n E JV). (1.87) When x = 1. (1. we have r(x + 1) = fooo txetdt = txetl: + fooo xtxletdt = xr(x). then r(n + 1) = n!. Thus.91) 22 .1.90) 8t ' The maximum is !max = exp[n ln n.86) Using integral by parts. (1. we have n! = /_: e{n lnln(l+()Jn(l+<)} dr = /_: e{(n In nn)+nlln(l+()<J} dr. That is to say. (1.n). we now can set t = n + r = n(1 + () so that r = n( varies around n and ( around 0.5 Some Special Integrals Calculus At x = 0. (1.6. n! = r(n + 1) = enlnttdt.89) fooo The integrand f(n. we have The variation of r(x) is shown in Figure 1. t) = exp[n ln tt] reaches a tnaximum value at af =O or t =n.5. 1.4 Gamma Functions (1. it is discontinuous and it is convention to set H(O) = 1/2.85) The special function is the gamma function which is defined by r(x) = fooo txletdt = fooo et+(xl)lntdt. Its relationship with the Dirac 8function is that d~~x) = 8(x).
r( 2) = y7r. The integration limits for r = n( (not () are from oo to oo.93) From the Gaussian integral with a= 1/(2n) + () = (. a (1.96) The standard gan1ma function can be decomposed into two incomplete functions: the lower incomplete gamma function 23 .. r(~2) = ..2" + "3 .. it can be shown that 1 r.6: Variation of f(x). Using ln(l we have (1..fi27r. where we have used ln[n(l + ()] = Inn+ ln(l + ().5 Some Special Integrals 20 10 10 20 5 0 X 5 Figure 1. (2 (3 (1.94) we now obtain the Stirling's asymptotic formula From the basic definition.. .Calculus 1.92) 1 00 ear 2 dr oo = ~ = /2rn.. (1.
101) r(x +Y).x) ~ r(a) as X~ have 1 As r(!) = . l(o:. The beta function is linked to r function by = r(x)r(y) B( x.100) From the definition. The solution (see later chapters in this book) can be expressed as Taylor's series.5 Bessel Functions Bessel functions come from the solution of the Bessel's equation (1. x) so that r(x) = 'Y(O:.y) (1. y) = B(y.5 Some Special Integrals Calculus /'(0:. we know that the beta function is symmetric.97) while the upper incomplete gamma function is defined by r(a. x) + r(a.x ). B(x. x) and the upper incomplete gamnla function r(a. and the Bessel function associated with this equation can be defined by (1.5.102) which arises from heat conduction and diffusion problems as well as wave propagation problems. 00.103) 24 . x).Jif. 2 2 Another related function is a beta function (1.x) = 1 1 00 to:letdt. x). (1.99) erf(x) = fi'Y( .98) Obviously. The lower incomplete gamma function is defined by (1. we (1.1. 1.
J_>... It can also be defined by the Bessel integral J>. With these fundamentals of preliminary mathematics..2 . 2 (1.=O . they have the following properites The Bessel functions of the first kind are plotted in Figure 1. These are the Bessel functions of the first kind. 7..Calculus 1.105) When A = k is an integer.. 1 0. cos(A7r).104) The Bessel functions of the second kind are related to J>.. sin(A7r) · (1.1 . where A is a real paran1eter. we are now ready to study a wide range of mathematical methods in engineering.(x) = 1 [27r 7r Jo cos[ At..1o!::=51~0=:'15' X Figure 1.x sin t]dt..5 ..7: Bessel functions.5 Some Special Integrals  A. >. 25 . and can be defined by y _ J>..
.
2. . velocity. Xn) andy = (Yt. . x 2.Chapter 2 Vector Analysis 1\tiany quantities such as force.1) and the sum is also a vector. Zero vector 0 is a special vector that all its components are zeros. Y2. The manipulation of vectors is often associated with matrices. All these vectors form a ndimensional vector space vn. we simply add their corresponding components.. (2. . . Z =X+ Y = (xt + Yt. To add two vectors x = (xb x2.X2 + Y2. and deformation in engineering and sciences are vectors which have both a n1agnitude and a direction. Yn). we will introduce the basics of vectors and vector analysis. .. .1 The addition of vectors has comn1utability (u + v = v + u) and associativity [(a+ b) + c = a+ (b +c)].. Xn)... This follows the vector addition parallelogram as shown in Fig 2. . where its components x 1 . .1 Vectors A vector x is a set of ordered numbers x = (xt. .. x2. In this chapter. ... Xn are real numbers.Xn + Yn)... The multiplication of a vector x with a scalar or constant a: is carried 27 .
b. y2. {3 are not all zeros..1.· : Figure 2. oy2.. The length or norm of a vector x is the root of the dot product of the vector itself.4) .•..1 Vectors Vector Analysis out by the multiplication of each component. ..: : : .•. Yn)· In addition. If a. 2..·.. . c are in the same plane..· :" :· // .. Three linearly dependent vectors a. lxl = llxll = VX:X= 28 ~~xr (2. then these two vectors are linearly dependent.. ay = (ay1. ayn)· (2.2. (2....~·········· ')/ .3) which is a real number. y = (y~. (a{3)y = a(f3y) and (a+ f3)y = ay + {3y...· .1 Dot Product and Norm The dot product or inner product of two vectors x and y is defined as X· Y= XtYl + X2Y2 + ··· + XnYn = L i=l n XiYi.. . Two nonzero vectors a and b are said to be linearly independent if aa + {3b = 0 implies that a= {3 = 0.2) Thus. 4 . Two linearly dependent vectors are parallel (a//b) to each other..1: Vector addition. .
. (2. Rearranging equation (2.1 Vectors When llxll = 1. . y. then it is a unit vector. 7r. (2. 1. j = (0. then we get the useful CauchySchwartz inequality: (2. e2 . 29 . O:n are the components of a relative to the basis e1. 0. {3 are constants. 7). en.. 1) for three x. 0. X· (y+z) =x·y+x·z.7) If the dot product of these two vectors is zero or cos(8) = 0 (i. .. 0.. i=l n (2. and thus x = x1i + x2.9) Any vector a in a ndimensional vector space vn can be written as a combination of a set of n independent basis vectors or orthogonal spanning vectors e1. () ::. The three unit vectors satisfy i · j = j · k = k · i = 0. then we say that these two vectors are orthogonal. If() is the angle between two vectors a and b. 0). + O:nen = L O:iei. () = 1r /2). The most common basis vectors are the orthogonal unit vectors.Vector Analysis 2. they are i = (1. In a threedimensional case.10) where the coefficients/scalars 0:1.e. . .i + x3k.0:2.. zaxis. we obtain a formula to calculate the angle () between two vectors cos(()) = a·b II all lib II (2. 1. It is straightforward to check that the dot product has the following properties: X·Y = y ·X.6) where a:.. e2. en.8) Since cos(8) ::. so that a= a:1e1 + a:2e2 + . (2.5) and (a:x) · (f3y) = (a:f3)x · y.. then the dot product can also be written a· b = llall llbll cos(8). k = (0.. . 0 ::..
14) and (ax)x (. (2.B)xxy.1 Vectors Vector Analysis 2. 30 . 0) and b = (2.2 Cross Product The dot product of two vectors is a scalar or a number. 1.1: For two 3D vectors a = (1. their dot product is a· b = 1 X 2+1 X ( 1) +0 = 1. .13) In fact. following a righthand rule. (2.15) A very special case is axa = 0. 1. On the other hand. It is straightforward to check that the cross product has the following properties: xxy = yxx.By) = (a.16) 0 Example 2. (2. ) The angle between a and b can also be expressed as (2. (x + y)xz = xxz + yxz. ( 2 12 . The vector c =ax b is perpendicular to both a and b. lla x bll sin f) = llallllbll. the norm lla x bll is the area of the parallelogram formed by a and b. For unit vectors.1. we have ixj = k. the cross product or outer product of two vectors is a new vector c=axb which is usually written as i a X j X2 k X3 b = Xt Yt Y2 Y3 = I:: :: li +I :: :: lj +I :: :: lk. 0). jxk = i. kxi = j.2.
c can be either a scalar a· (b x c)= or a vector ax ay a:: bx by b:: Cy Cz ' (2.17) ax (b x c)= (a· c)b. 3).56 .1.0 X 11 X 0.1 Vectors we can calculate the angle(} between the two vectors.18) . Similarly. the product of triple vectors a.2 X 1) = (0.1 X (1). 0 Their cross product is V =a X b = (1 X 00 X (1).3 Vector Triple For two vectors.(a· b)c. which is a vector pointing in the negative zaxis direction. b. The vector v is perpendicular to both a and b because a ·V = 1 and b ·V = 2 X X 0+1 X 0+0 X ( 3) = 0. 31 (2. 0 2. We have cosO= or (} a·b 1 llallllbll = J2J5' l = cos vlO 1'17\ ~ 1 71.0. 0 + (1) X 0+0 X ( 3) = 0.Vector Analysis As their moduli are 2. their product can be either a scalar (dot product) or a vector (cross product).
2 Vector Algebra Vector Analysis As the dot product of two vectors is the area of a parallelogram.1 Vector Algebra Differentiation of Vectors The differentiation of a vector is carried out over each component and treating each con1ponent as the usual differentiation of a scalar. Conversely. for a position vector P(t) = x(t)i + y(t)j + z(t)k.2.22) we can write its velocity as v = ~ = ±(t)i + y(t)j + z(t)k. (2.25) where cis a vector constant.21) 2.24) where () = d()jdt. (2. the scalar triple product is the volume of the parallelepiped formed by the three vectors.2 2.20) a and X (b X c) #(a X b) XC. 32 . (ax b)· (c x d) =(a· c)(b ·d). (2.23) and acceleration as a= ~t~ = X(t)l + ij(t)j + Z(t)k.19) (2. (2.2. Thus. From the definitions. the integral of vis P = J vdt+c.(a· d)(b ·c). (2. (2. it is straightforward to prove that a· (b x c)= b · (c x a)= c ·(ax b)= a· (c x b).
The first one is the Cartesian coordinates (x.26) d(axb) _ da b db dt .27) 2. y) with two unit vectors i (along positive xaxis) and j (along positive yaxis).2. and the second one is the polar coordinates (r. y Figure 2. it is easy to check that the differentiation of vectors has the following properties: d(a:a) ~= and 0 da dt~ d(a ·b) da b db dt = dt · +a·dt' (2. 0). The position vector r = x(t)i + y(t)j at point P at any instance t in the Cartesian coordinates can be expressed as (r. their unit vectors and their relationship with Cartesian coordinates.2 Vector Algebra From the basic definition of differentiation.2: Polar coordinates.28) 33 .Vector Analysis 2. The velocity vector is (2. 0) with two unit vectors er and ee as shown in Figure 2. let us study the motion along a curved path. In mechanics~ there are three coordinate systems which can be used to describe the motion uniquely.2.dt x + ax dt · (2.2 Kinematics As an application of vector algebra.
2. Figure 2. point 0) and 'l/J is the angle of the tangent at the point P (see Figure 2.3).31) where pis the radius of the curvature at point P. .2 (frO )er + (rO + 2rO)eo.3: Intrinsic coordinates and their unit vectors. . For the circular n1otion such as a moving bicycle wheel as shown in Figure 2. In a rotating reference frame with an angular velocity 34 . p .. 82 (2. the velocity is simply (2. 1/J). (2.29) The third coordinate system is the intrinsic coordinate system (s.30) The acceleration becomes a= set+ en. In the intrinsic coordinates.. the three coordinate systems are interconnected. and the velocity is always along the tangent. the position is uniquely determined by (s. 'l/J) where sis the arc length from a reference point (say. Naturally. The two unit vectors for this systems are et along the tangent direction and en which is the unit normal of the curve.4.2 Vector Algebra Vector Analysis and the acceleration is a= v= .
In addition. and acent = W X ( W X r).35) is the centripetal acceleration.32) and the acceleration is ap = aA +d dt2 2 rl A+ . The velocity is Vp = VA + dt drl X A +W X r.Jf dw acor r + acor + acent. (2. the differentiation of the unit vectors are connected by (2. the velocity and accelera tion at any point (say) P can be calculated using another fixed point A on the rotating body (or wheel). (2. (2. w = Ok where k point to the zaxis.4: Three coordinate systems for a wheel in circular motion.36) 35 . It is worth noting that the velocity v A and acceleration aA is the velocity and acceleration in a nonrotating frame or an inertia frame.2 Vector Algebra Figure 2.Vector Analysis y 2.33) where = 2w X drl dt A' (2.34) is the Coriolis acceleration.
(2. '2 (2.38) Differentiating it with respect to time and using the relationships of unit vectors.p = a sin() = 5.2: A car is travelling rapidly along a curved path with a speed of 30 mjs at a given instance. Thus. and let a be the magnitude of the total acceleration..2. The accelera. + R</> en. the velocity is v =set= 30et. The accelerometer also indicates that the component of the acceleration perpendicular to the travelling direction is 5 mjs2. 36 s2 302 . en = sin </>i +cos </>j. The velocity for this circular motion is simply (2. we have a = v = R</>et • .tion is given by a= set+ ~en= a(cos8et p ·2 + sin8en)· Therefore. we have p = .40) 0 Example 2. The car is fitted with an accelerometer and it shows that the car is accelerating along the curved path at 2 m/~. is there any danger of toppling over? Let () be the angle between the acceleration vector and the velocity vector.37) In the intrinsic coordinates.6 meters. 8 = R~. In the intrinsic coordinates. lVhat is the direction of the total acceleration at this instance? lVhat is the radius of the curvature? Suppose the car has a height of 2 meters and a width of 1.39) where the unit vectors are et =cos </>i +sin </>j.2 Vector Algebra Vector Analysis and (2. we have s = R</> where R =constant is the radius of the wheel in circular motion.
0. \Ve 2 + 52 = J29. If we take the moment about the axis through the two contact points of the outer wheels. Thus.8A1 < 0. However. where J\tl is the mass of the car. know that the magnitude of the acceleration is a = The angle is v2 (} = tan.1 '5 ~ 68. z) = xi+ yj + zk.Ildy dx ds = Jdx'l + dy'l Figure 2. There is no danger of toppling over.5: Arc length along a curve.8m from the edges of the outer wheels. if the car speeds up to v = 42 m/s (about 95 miles per hour). there is a danger of toppling over when the moment of the weight is just balanced by the moment of the centripetal force.3 Line Integral I ________ _. y. we can assume that the centre of gravity is approximately at its geometrical centre. For example. 2 In addition.8Mg p v2 ~ 2. we have the total moment 1 X 1\t[.20°.2. the centre is 1m above the road surface and 0. in order to calculate the arc length L of curve r 37 .2 Vector Algebra or the instantaneous radius of curvature is p = 30 2 /5 = 180m. 0 2.Vector Analysis 2. An important class of integrals in this context is the line integral which integrates along a curve r(x.
Sometimes. L = [s ds = [s dx2 + dy2 = 1x }~ ~0 xo J (2. ax ay az 38 (2. 2. 8k 1+J+.2 Vector Algebra Vector Analysis as shown in Figure 2.cos1) is a unit vector and o. (2.cos{3. are the gradient operator (grad or V). z) in the direction of n an = n · V¢ = a¢ a¢ a¢ 8¢ ax cos( a)+ ay cos(f3) + 8z cos( I). (2. z can be written in a sin1ilar way.1 are the directional angles.f3.4 Three Basic Operators Three important operators commonly used in vector analysis. y.2. the gradient of any scalar function¢ of x.2. Generally speaking. we have to use the line integral.43) This is equivalent to applying the del operator V to the scalar function¢ r7 v = 8.5. 8. y.42) where n = (cosa.9: The arc length of the parabola y(x) = tx 2 from x = 0 to x = 1 is given by L= 1' v't + 1 y' 2dx = 1' v't + x2dx = ~[x~ + ln(x + ~>{ = 2 [V2 ln(V2 1)] ~ 1. the divergence operator (div or V ·) and the curl operator (curl or Vx). especially in fluid dynamics. it is useful to calculate the directional derivative of a function ¢ at the point (x.41) 0 Example 2.44) .14779..
z)k. For a vector field u(x. d lVU= t"'7 v 8u3 ·U=++ 8ut 8x 8u2 oy oz' (2. (2.47) and the curl of u is the cross product of the del operator and the vector field u i j 1:) k {) curl u = \1 x u = 1:) (JX Ut (JY IE (2. 4> are scalar functions. \1 x (\1 x u) = \1(\1 · u) .51) (2.Vector Analysis 2.46) the application of the operator \1 can lead to either a scalar field or a vector field. z) = ut(x.54) .2 Vector Algebra The direction of the gradient operator on a scalar field gives a vector field. 39 82 w 82 w 82 w (2.49) \11/J = 0. (2. y.\7 2u. z)j + u3(x. (2. y.48) u2 u3 It is straightforward to verify the following useful identities associated with the \1 operator: \1 · \1 \1 X X U = 0.45) where a. \7(1/J¢) = 1/J\14>+¢\11/J. y. ('Vw) 2 = 8x2 + 8y2 + 8z2 . The gradient operator has the following properties: \l(o:'l/J+/3¢) = o:\11/J+/3\1¢. z)i + u2(x. The divergence of a vector field is the dot product of the del operator \1 and u . 'V· (1/Ju) = 1/J\1· u + (\11/J) · u.53) One of the most common operators in engineering and science is the Laplacian operator is \1 w = \1.52) \1 X (1/Ju) = 1/J\1 Xu+ (\11/J) XU. y.50) (2. (3 are constants and 1/J. depending on how the del operator applies to the vector field. (2. (2.
</>. In cylindrical polar coordinates (r.2.57) In spherical polar coordinates (r. Gauss's theorem states: Jjfn(v. u = ~ 8(rur) + ~ 8ucp + 8u:: r 8r r 8</> 8z · (2.5 Some Important Theorems The Green theorem is an important theorem. especially in fluid dynamics and the finite element analysis.56) The Laplace equation becomes (2. u = 1 8 (r ur) 2 2 + 1 r2 8r2 r sin 8 8(sin8ue) 88 + _1_8uct> (2. Q)dn = Jfs Q. ¢).60) which connects the volume integral to the surface integral. (2.59) 2. The theorems connect the integrals of divergence and curl with other integrals. ndS.55) In engineering mathematics. we have \7.2 Vector Algebra Vector Analysis for Laplace's equation (2. 8. it is sometimes necessary to express the Laplace equation in other coordinates. we have \7.2. z).58 ) r sin 8 8<1> · The Laplace equation can be written as (2. 40 . For a vector field Q = ui + vj in a 2D region n with the boundary r and the unit outer normal nand unit tangent t.
kdS = £Q. the useful Green's first identity can be written as J 'l/JV</>dr = ln { (.61) which connects the surface integral to the corresponding line integral. 2. In our simple 2D case.3.63) where dO. y.3 Applications Another important theorem is Stokes's theorem: Jfs(v x Q). For any enclosed volume n.62) For any scalar functions 'l/J and ¢.PV 2 ¢ + V'l/1 · V¢)d0..Vector Analysis 2.3 Applications In order to show the wide applications of vector analysis. z. the total mass which leaves or enters the surface S is fspu · dA. we get Green's second identity 2. tdr = £Q. t) are the density and the velocity of the fluid. y. By using this identity twice. dr. (2. (2. (2. respectively. let us apply them to study the mechanical and flow problems. this becomes f !on (udx + vdy) = Jln (0 x [[ 8v 8u {)y)dxdy. The rate of change of mass in n is .1 Conservation of Mass The mass conservation in flow mechanics can be expressed in either integral form (weak form) or differential form (strong form). = dxdydz. t) and u(x. where p(x. z.
which is the differential form or strong form of mass conservation. yet its thickness does not exceed 100 meters. The ring systen1 has a diameter of 270. One natural question is why the formed structure is a ring system.000 km. The sizes of particles in the rings vary from centimeters to several meters. Interchange of the integration and differentiation in the second term. This is the integral form or weak form of the conservation of mass. while the differential form is more natural for mathmnatical analysis.3. The integral form is more useful in numerical methods such as finite volume methods and finite element methods. we have In V · (pu)dV + ! In pdV = 0. The ring system has very complicated structures. why not a 42 . we have 1 n 8t [. we have Using Gauss's theorem for the surface integral.2. Therefore.2 Saturn's Rings We all know that Saturn's ring system ranks among the most spectacular phenomena in the solar system. and this size distribution is consistent with the distribution caused by repeated collision. This is true for any volume at any instance.+ V · (pu)]dV 8p = 0. and subsequently the only way that this is true for all possible choice of n is ap 8t+V·(pu)=0. 2.3 Applications Vector Analysis The n1ass conservation requires that the rate of loss of mass through the surface S is balanced by the rate of change in n.
67) 43 . the total angular momentum is then h = k · [L miri i=l N N x vi] N = L mi(ri i=l N X Viz)· k + L mi(ri x Vir)· k + L mi(ri x Vie)· k. If we now decompose the velocity of each particle into Vi = (vir. Vie. under some reasonable assumptions.. The interparticle collisions dissipate energy while conserving the total angular momentum. Suppose the whole particulate system consists of N particles (i = 1. Viz). and we have h = L mirivie· i=l N (2. N). i=l i=l (2. therefore. 2. So only the tangential terms are nonzero. the rings must be particulate. . When the debris particles surrounding a planet will ultimately settle into flat circular rings or disks. However. . we can understand why the ring system is so.3 Applications spherical shell system? This is a challenging topic which has not yet fully understood. and the zaxis (along k direction) is normal to this plane. Niaxwell in 1890 showed that a fluid or gaseous ring was unstable.. Later.66) The total mechanical energy is (2.65) The first two terms disappear because Viz is parallel to k and axial velocity does not contribute to the angular momentum. y) plane coincides with the plane of the rings. Laplace in 1802 showed that such rings could not be solid because the tensile strength of the known materials was too small to resist tidal forces from Saturn. which are the natural consequence of energy dissipation in rotating systems. Its total angular momentum is h. By choosing a coordinate system so that (x.Vector Analysis 2.
Now let us see why the minimization of the rotational energy will also lead to the same conclusion of ring formation. the angular momentum becomes N h = :Lmirlw = Iw. This means that the collisions dissipate energy while flattening the system into a disk or rings.3 Applications Vector Analysis where 1/J(ri) is the potential per unit mass due to Saturn's gravity. we have Vie = TiW. Loosely speaking. subsequently redistributing or changing its angular velocity. This will essentially lead to a quasisteady state. With these assumptions. one particle may move faster and ultimately collides with other particles. we have to maximize I because h is constant. The total rotational energy is T= 2 ~miriw = 2Iw = 2T ~ Tmin· l=l 1 N 2 1 2 1 h2 In order to minimize T.2. therefore. Therefore. prdr dr 2 4 44 . we can see that the minimization of E while h is held constant requires that Viz ~ 0 and vie ~ 0. the system will evolve towards an energy minimum. we can assume that the angular velocity w = iJ is almost the same for all particles as t ~ oo (or any reasonable long time) so that collisions are no longer significant or the rate of energy dissipation is small. i=l I= :Lmirl. From both expressions for h and E. If we further assume that the potential energy does not change significantly (this is true if the particles do not n1ove significantly along the radial direction). thus the minimization of total energy leads to the minimization of the total rotational energy. For a disk with a radius a. i=l where I the moment of inertia of the particulate system. we have I= !v 2 r dm = t foR r 2 fo " dO= 21ftp foR r 3 = 1ftp ~ . The interparticle collisions will dissipate the energy. a thickness t <<a and the total mass m. If there are different angular velocities.
we have 2 2 !ball= 5mR. we have Iring > /disk > !sphere > !ball· This means that the total rotational energy is minimized if the particle system evolves into a ring or at least a disk. For a spherical shell. we have Similarly. 45 . 1 2 If all the mass is concentrated at a ring. we have /disk= 2mR.3 Applications Using the density p = mj(t1r R 2 ).Vector Analysis 2. we have !sphere = 2 2 3mR · Therefore. for a solid ball with the same mass and same radius R. This is probably the main reason why the planetary system and rings are formed.
.
1 2 3) A= ( 4 5 6 ' and B= ex ( sinx) eiO ' icosx (3. 47 . and u is a 3 x 1 column matrix or column vector. A matrix A has m rows and n columns. A matrix is a table or array of numbers or functions arranged in rows and columns. The elements or entries of a n1atrix A are often denoted as aii. (3.3) where A is a 2 x 3 matrix.2) (3. A is square if m = n. B is a 2 x 2 square matrix.1 Matrix 1\tiatrices are widely used in almost all engineering subjects. or m x n.Chapter 3 Matrix Algebra 3.1) we say the size of A is m by n. For example.
We can multiply a matrix A by a scalar a: by multiplying each entry by a:.7) and (3. . That is to say.4) where (i = 1. However. 2.j = 1.. The multiplications of matrices are generally not commutive. then the product Cis mxr.n). for a 2 x 2 matrix Adt  dA . n). the multiplication has associativity A(uv) = (Au)v and A(u + v) =Au+ Av. and thus AT will be nxm if A is mxn. if A is mxn and B is nxr. Its diagonal 48 .2.j = 1. (aT)ij = aii. The product of two matrices is only possible if the number of columns of A is the same as the number of rows of B. . For example.. In general. and their sum. m... . i. ..3. is obtained by adding corresponding entries C=A+B. which is also m x n. . .... The transpose AT of A is obtained by switching the position of rows and columns.2. Ctj = (AB)ii = L aikbki· k=l n ~ n (3.e. A. (3. . .(i = 1.. 2.1 !v!atrix 1\llatrix Algebra The sum of two matrices A and B is only possible if they have the same size m x n. we have (3. AB "# BA.6) The differentiation and integral of a matrix are done on each member element..m..5) If A is a square matrix. then we have An = AA .8) A diagonal matrix A is a square matrix whose every entry off the main diagonal is zero (aii = 0 if i "# j). ( da21 (If'"" dau dt ~ da12 ) dt dt ' (3.
2 Determinant elements or entries may or n1ay not have zeros.10) A zero or null matrix 0 is a matrix with all of its elements being zero. we have AI=IA=A. and is nonsingular if and only if IAI =/= 0. In general. 3. the matrix I= (0 1 0 0) 0 1 0 0 1 (3. its determinant simply becomes (3.11) where Mij is the determinant of a minor matrix of A obtained by deleting row i and column j. A nxn square matrix is singular if IAI = 0. (3. The trace of a square matrix tr(A) 49 . For a simple 2 x 2 matrix. expanding by row k. For example. For example.2 Determinant The determinant of a square matrix A is a number or scalar obtained by the following recursive formula or the cofactor or Laplace expansion by column or row.1\t/atrix Algebra 3.13) where A and Bare the same size (nxn). (3.9) is a 3 x 3 identity or unitary matrix. IABI = IAIIBI.12) It is easy to verify that the determinant has the following properties: laAI =alAI. we have n det(A) = IAI = L(1)k+iakjMkj' j=l (3.
3 Inverse The inverse matrix At of a square matrix A is defined as A .19) (3. + G. it is straightforward to prove that the inverse of a matrix has the following properties (A 1)1 =A.3. IZI.16) (3.3 Inverse 1\llatrix Algebra is defined as the sum of the diagonal elements. n tr(A) = L~i = i=l a11 + a22 + . tr(A) = tr(AT)... From the basic definitions..22) It is worth noting that the unit matrix I has the same size as A. From the basic definitions.. tr(A +B) = tr(A) + tr(B). n). The inverse of a square matrix exists if and only if A is nonsingular or det(A) =I= 0. tr(AB) = tr(BA). det(A 1 (3.. z)T = zT . (3. the rank of A is rank( A) ~ min( m. it is nonsingular if rank( A)= n.nn· (3.... Generally.21) 1 ) = det(Ar det(AB) = det(A)det(B).20) (3.17) (3.18) (3. (3.23) .1 A = AA . BT AT.1 = I. 50 (3. 3.14) The rank of a matrix A is the number of linearly independent vectors forming the tnatrix. ZI = IAIIBI .. For a n x n square matrix A.15) IAB . it is straightforward to prove the following (AB ...
2 2 B = ( ~1 ~ 7 1 ) .1: l 1 ( d =(adbe) c (3. The transpose matrices of A and B are AT= 0 ~1 ~). V21 = 1 X 1 + 1 X 2 + 0 X 1 = 1. BT = ( 1 1 2 1 ) 3 7 . we have 51 .25) For two matrices A= ( we have ~1 3 i ~). Thus. 9 17 However. V32 = 3 x (1) + 2 x 3 + 2 x 7 = 17.24) A simple useful formula for obtaining the inverse of a 2 x 2 matrix is a b) ( c d 0 Example 9. AB=V= ( 8 26) 1 4 . V31 = 3 x 1 + 2 x 2 + 2 x 1 = 9. Similarly.1\t/atrix Algebra 3. V22 = 1 X ( 1) + 1 X 3+0 X 7 = 4. BA does not exist. V12 = 1 x ( 1) + 2 x 3 + 3 x 7 = 26. where V11 = 1 x 1 + 2 x 2 + 3 x 1 = 8.3 Inverse and (3.
~).... and A 2 = AA and so on. In this case.4 !\latrix Exponential The inverse of A is Matrix Algebra and the determinant of A is detiAI = 9. 0 Example 9. The exponential of a square matrix A is defined as e A ~ 1 1 =L. 0 3.2: For a simple matrix A=U we have eA n.4 Matrix Exponential Sometimes.  ( e' 0 For A=(: :). 52 .An =I+A+. we need to calculate exp[A]. 2 n=O nJ (3. The trace of A is tr(A) = A11 + A22 + Aaa = 1 + 1 + 2 = 4. we have to deal with matrix exponentials.A 2 + .3..26) where I is a unity matrix with the same size as A. The matrix exponentials are very useful in solving systems of differential equations.. where A is a square matrix. This (rather odd) definition in fact provides a n1ethod to calculate the matrix exponential.
29) (3.5 Hermitian and Quadratic Forms we have eA . In general.1) !(1 + e2t) ~(e2t. Fortunately.1\t/atrix Algebra 3.5 Hermitian and Quadratic Forms The matrices we have discussed so far are real matrices because all their elements are real. (3.31) (3. . and the matrix becomes a complex matrix. For a matrix A. the entries or elements of a matrix can be complex numbers. it is quite complicated but still straightforward to calculate the matrix exponentials.( ~(1 + e2t) !(e2t.28) (if AB !!:_etA = BA). 3. it can be easily done using a computer. (3.11 (tA) n =I+ tA + 2A2 + ..1) ) .32) dt = AetA = etA A ' (eA)1 = eA. det( eA) = etrA. we can prove the following useful identities e tA ~ i = L. By using the power expansions and the basic definition.30) (3. 2 n=O n.27) (3.. t e et cosw et sinw et sinw ) et cosw · 0 As you see. its complex conjugate A* 53 . For a slightly complicated matrix A= ( t w we have A= ( w )·.
it is called a unitary matrix. It is an antiHermitian matrix if A* =A.3. (3. If the Hermitian matrix of a square matrix A is equal to the in verse of the matrix (or At 0 Example 9.33) =~~ :::). :~~· 1 ::: ) . For a matrix ( 2~i et11' a= 3 . it is said to be an Hermitian matrix. cos 9 · 54 .1 = AT.i1r ' its complex conjugate a• and Hermitian conjugate at are a· = ( 2 :.34) and (3. (3. The Hermitian conjugate At is obtained by taking the transpose of its complex conjugate matrix.i e 111' 3 + 2i 0 1 ) 1 + i1r ' For the rotation matrix A =( cos 9 .35) A square matrix A is called orthogonal if and only if A . That is to say. 9: = A .sin 9 sin 9 ) . If a square matrix A satisfies A* = A.1). for A= ( we have :~ 1.5 Hermitian and Quadratic Forms Matrix Algebra is obtained by taking the complex conjugate of each of its elements.2i 0 1 ) 1 .
Q2. 55 . i=l n (3.. a quadratic form 'l. Qn. . An important result concerning the quadratic form is that it can always be written through appropriate transformations as 1/J(q) =I: AiQr =AtQI + A2Q~ + . ..b might be more easily evaluated in certain intrinsic coordinates Qb Q2..37) Since 'l/J is a scalar. For a real vector qT = ( Q1. it should be independent of the coordinates.. In the case of a square matrix A.. 'l. 0 A very useful concept in engineering mathematics and computing is quadratic forms. we have AT = A l.38) The natural extension of quadratic forms is the Hermitian form that is the quadratic form for complex Hermitian matrix A.b(q) is a scalar function defined by 'f.5 Hermitian and Quadratic Forms sin() ) cos() ' and Since cos 2 () + sin 2 () = 1.b(q) = qT Aq An = ( Ql Q2 ••• Qn ) ( ~~~ Ant )(l:} n (3. qn) and a real square matrix A. q3.36) which can be written as 1/J(q) = LL QiAijQj· i=l j=l n (3.1\t/atrix Algebra its inverse and transpose are 3.AnQ~. .. Therefore. the original matrix A is orthogonal.
56 (3. A.6 Solution of linear systems !via trix Algebra Furthermore.1 b.1 Au=A.39) can be written in the compact form as (3. 0 Example 9.4: For a vector q A= 2 = (q1.40) or simply Au=b. q2) 1 ' and the square matrix ( 1 2) we have a quadratic form 0 3.41) In the case of m = n. the matrix A can be linear operators and functionals in addition to numbers.6 Solution of linear systems A linear system of m equations for n unknowns (3. (3. we multiply both sides by A .42) .3.l (this is only possible when m = n).
1 . z)..45) A at2 a22 A .. Ai 57 = (3. Eigenvalues have the interesting connections with the matrix. a. (3.46) A which is equivalent to (3.Al)v = 0.. (3.1\t/atrix Algebra 3.1b.50) eig(A)..43) A special case of the above equation is when b =Au. the characteristic equation has n solutions.44) (A. tr( A) =L i=l n B.. and this becomes an eigenvalue problem.1 = z.47) In general. An eigenvalue A and corresponding eigenvector v of a square matrix A satisfy Av or = Av. (3. Any nontrivial solution requires aua21 (3.1 A ... Some useful identities involving eigenvalues and inverse of matrices are as follows: (AB .48) For a symmetric square matrix. the two eigenvectors for two distinct eigenvalues Ai and Aj are orthogonal vTv = 0.6 Solution of linear systems we obtain the solution u=A. .1 .ii = A1 + A2 + .. ann =0. + A 0 • (3.49) (3.
5: For a simple 2 x 2 matrix its eigenvalues can be determined by 1.3. In order to obtain the eigenvector for each eigenvalue.A)(4. (3. 1 or {1 .A) . 5 = 0.52) (3.A = O. The trace of A is tr(A) =Au + A22 = 1 + 4 = 5 =At + A2.2 X which is equivalent to (A+ 1){A.53) L Aii = L Ai. or which is equivalent to 58 .51) (3. we plug this into IAAllv = 0. Thus. 0 Example 9.6) = 0. ~ I( ~: ) = o. and we have 1{1) 2 5 4{1) I( VV2t ) = 0.A 5 I 2 4.6 Solution of linear systems Matrix Algebra = eig(AB) tr(A) = eig(BA). we assume For the eigenvalue At = 1. or I . the eigenvalues are At = 1 and A2 = 6.
2We have VI = 5/v'29~ v2 = eigenvalue and eigenvector At= 1. The direct inversion becomes impractical if the n1atrix is very large (say. l\1any efficient algorithms have been developed for solving such systems.)2 + v2. if n > 1000). the solution u =Alb generally involves the inversion of a large matrix. Gauss elimination and L U decomposition are just two examples.6 5 2 46 I ( Vt ) 1 v2 Using the normalization condition v'f has the following solution + v~ = 1. 59 . 5v2 ( 2.54) Similarly. the above equation For a linear system Au= b.X2 = 6 gives 1.6 Solution of linear systems This equation has infinite solutions. V29 = 0. each corresponds to the vector parallel to the unit eigenvector. As the eigenvector should be normalized so that its modulus is unity. this additional condition requires v~ which means + v~ = 1.1. we have the first set of Vt to ). = ( Thus.1\t/atrix Algebra 3. (3. 2/v'29. the second eigenvalue .
.
3) . 4.1 Complex Numbers and Functions lviathematically speaking. it is son1etimes easy or even necessary to use complex variables in mathematics and engineering. b). z =' which consists of the real part ~(z) =a and the imagery part SJ<( z) = b. b).2) The multiplication and division of two complex numbers are in the similar way as expanding polynon1ials (a+ ib) · (c + id) = (ac.bd) 61 + i(ad +be). z =a+ ib. It can also be written as the order pair of real nunlbers using the notation (a.1) a. the techniques based on complex variables are among the most powerful methods for mathematical analysis and solutions of mathematical models. however.Chapter 4 Complex Variables Although all the quantities are real variables in the physical world. a complex number z is a generalized set or the order pair of two real numbers (a. (4. In fact. (4. written in the form of ·2 1 (4. The addition and substraction of two complex numbers are defined as (a+ ib) ± (c + id) =(a± c)+ i(b ±d). bE 'R.
() = arg z. 0) to the point (a.1). This representation is often called the Argand diagram (see Figure 4. (4. The angle () that the vector makes with the positive real axis is called the argument (see Fig 4. b). and it is 62 .5) The order pair (a. b). This plane is called the complex plane. 1r is called the principal argument of z.ad c_+_i_d = c2 + d2 + z:c2~+d::::2 · (4. Figure 4. The value range 1r < () :::. we may replace () by () + 2n7r (n E N).1: Polar representation of a complex number. similar to a vector. The length of the vector is called the magnitude or modulus or the absolute value of the complex number 2 (4. (4. be . implies that a geometrical representation of a complex number a + ib by the point in an ordinary Euclidean plane with xaxis being the real axis and yaxis being the imaginary axis (iy). The complex conjugate or simply conjugate z (also z*) of z =a+ ibis defined as z =a. The vector representation starts from (0.4.1).ib.1 Complex Numbers and Functions Complex Variables and a + ib ac + bd .4) Two complex numbers are equal a + ib = c + id if and only if a= c and b =d.7) In fact.6) r = lzl = v' a + b2 .
...14) For 0 = 1r...0 02 i03 = (1 2f + . . + ~ + .i sin 0.10) and for a special case z = iO. (4.... 51 os (4. this leads to a very interesting formula (4.12) 02 04 cos 0 = 1 ..8) This polar form of z and its geometrical representation can result in the Euler's formula which is very useful in the complex analysis ei6 = cos(O) + isin(O).3f + .2! + 31 .Complex Variables 4. the Euler's formula becomes eilJ = cos( 0) + i sin( 0) 63 = cos 0 . In the complex plane. (4. ..13) 2! 4! we get the wellknow Euler's formula or Euler's equation eilJ = cos 0 + i sin 0.11) Using the power series sin 0 and =0 + . (4... ) + i(O..+ .. 03 3! 02 03 (4. (4. 1 .1 Complex Numbers and Functions usually denoted as Argz.. z2 zn (4..15) If we replace 0 by 0.16) . we have e ilJ = + z . ).9) The Euler formula can be proved using the power series. ... For any z E C... the complex number can be written as z = rei() = r cos( 0) + ir sin( 0). we have the power series ez = 1 + z + 2f + . . (4.
4.1 Complex Numbers and Functions
Complex Variables
Adding this equation to (4.14), we have
ei0
+ eiO = 2 cos 8,
eiO
(4.17)
or
cos8 =
+ eiO
2 Similarly, by deducting (4.16) from (4.14), we get
eiO eiO
(4.18)
sin8 =
i 2
(4.19)
For two complex numbers z1 = r 1ei 01 and z2 = r2ei 02 , it is straightforward to show that ZJZ2 = r1r2ei(o 1 +02 ) = r1r2[cos(a1
+ a2) + isin(at + a2)],
(4.20) which can easily be extended to get the wellknown de l\1oivre's formula
[cos( B)+ isin(8)r = cos(n8) + isin(n8).
0 Example 4.1: direct calculation
(4.21)
Find z 4 if z
= 1 + J3i.
lVe can evaluate it by
z4
= (1 + J3i) 4 = [(1 + J3i) 2 ) 2 = [1  3 + 2J3i) 2
3 2v'3)
= 22 ( 1 + v'3i) 2 = 4(1 
= 8 8v'3i.
lVe can also use Moivre's formula. The modulus of z is r =
lzl
=
+ J3 = 2. The argument()= tanz = 2e7rf 3 . We now have
2
V1
2
1
f
= 7r/3 = 60°.
3
Thus,
z 4 = 24 e 4 7r/J = 16(cos 4 11"
3
+ isin 411")
= 16( ~
v;
i)
= 8 8v'3i,
which is exactly the same result as we obtained earlier. The second method becomes much quicker if you want to evaluate (say) z 100 .
0
64
Complex Variables
4.2 Hyperbolic Functions
5~,~~~~~~
\
\
~
\ \
''
''
u;
'''
...
>C ~ 0 c
··~···

5 3
sinh cosh ·· tanh

2
1
0
X
1
2
3
Figure 4.2: Hyperbolic functions.
4.2
Hyperbolic Functions
Hyperbolic functions occur in many applications and they can be thought as the complex analogues of trigonometric functions. The fundamental definitions are sinhx and
=
ex ex
2
h~
coshx
=
ex+ ex , 2
(4.22)
1 (4.23) COS X tan x Figure 4.2 shows the variation of sinh, cosh, and tanh. If we replace x by ix and use Euler's formula, then we have tanhx = coth x
sinhx
= h.
eix _ eix
sinh ix = 2
. . ZSlnX )] . . = ZSlllX.
=
2
1 [( COSX
. . + ZSlnX ) 
(COSX
(4.24)
65
4.2 Hyperbolic Functions
Complex Variables
Similarly, we have 1 cosh ix = (ezx 2
°
+ ezx)
0
=
~[(cosx + isinx) + (cosx isinx)] =
cosix = coshx,
cosx.
(4.25)
In a similar fashion, we can also prove that sin ix = i sinh x.
(4.26)
Some identities are as follows: cosh2 x sinh2 x = 1, sinh 2x = 2 sinh x cosh x, and cosh 2x = sinh2 x
0 Example 4.2: definitions, we have
(4.27) (4.28) (4.29)
+ cosh 2 x.
Prove that cosh 2 x sinh 2 x = 1. From the
and Thus, we have
= 4[2 (2)) = 1.
0
1
The inverses of hyperbolic functions are defined in a similar way as trigonometric functions. For example, y = cosh x, its inverse is defined as x = cosh t y. From the basic definitions, we have sinhx + coshx =ex. (4.30)
66
Complex Variables
4.3 Analytic Functions
Using sinhx
= Vcosh2 x
1, we have (4.31)
Vcosh 2 x 1 + coshx =ex, or
(4.32)
which gives
4.3
Analytic Functions
Analytic Functions
Any function of real variables can be extended to the function of complex variables in the same fonn while treating the real numbers x as x + iO. For example, f (x) = x 2 , x E 'R becomes f (z) = z 2 , z E C. Any complex function f (z) can be written as
f(z)
=
f(x + iy)
=
= ~(f(z))
+ ir;s(J(z))
(4.34)
u(x, y) + iv(x, t),
where u(x, y) and v(x, y) are realvalued functions of two real variables. A function f (z) is called analytic at z0 if f' (z) exists for all z in smne €neighborhood of zo, that is to say, it is differentiable in some open disk lz z 1< f. If f(z) = u + iv is analytic at 0 every point in a domain n, then u(x, y) and v(x, y) satisfying the CauchyRiemann equations
8u 8x
8v
8y'
(4.35)
Conversely, if u and v of f (z) = u + iv satisfy the CauchyRiemann equation at all points in a domain, then the complex function f(z) is analytic in the same domain. For example, the 67
4.3 Analytic Functions
Complex Variables
elementary power function w = zn, (n > 1) is analytic on the whole plane, w = pei(j>, z = rei 9 , then
p = rn, 4> = n8.
(4.36)
The logarithm is also an elementary function w = In z In z =In lzl
+ i arg(z) =In r + i(8 + w1rk),
(4.37)
which has infinitely many values, due to the multiple values of 8, with the difference of 21rik(k = 0, ±1, ±2, ... ). If we use the principal argument Argz, then we have the principal logarithm function Ln(z) =In lzl + Argz. (4.38)
If we differentiate the CauchyRiemann equations, we have fJ2uj8x8y = 8 2 uj8y8x. After some calculations, we can reach the following theoren1. For given analytic function f(z) u + iv, then both u and v satisfy the Laplace equations
(4.39)
This is to say, both real and imaginary parts of an analytic function are harmonic. A very interesting analytical function is the Riemann zetafunction ((s), which is defined by
((s)
=
n=l
~ , L ns
00
1
(4.40)
where s is a complex number with its real part more than unity. That is s E C and ~(s) > 1. This function (infinite series) is analytic, and it can be extended for all complex numbers s =I= 1. For example,
((2) =
but
1 1 + 22
1 + 32
+ ... = 6'
00.
7r
2
(4.41)
1 1 1 ((1) = 1 + 2 + 3 + 4 + ... =
68
(4.42)
Complex Variables
4.3 Analytic Functions
This ((s) function has trivial zeros s = 2, 4, 6, ... and it also has nontrivial zeros. There is a famous unsolved problem, called the Riemann hypothesis, related to this function. The Riemann hypothesis conjectured by Bernhard Reimann in 1859 states that all real parts of any nontrial zero of the Riemann zetafunction ((s) are That is to say, all the nontrivial zeros should lie on a straight lines= + iy. This is amilliondollar open problem as the Clay Mathematical Institute in 2000 offered a million dollars to search for a proof, and yet it still remains unsolved.
!·
!
Laurent Series
For an analytic function p(z), one of important properties is the singularity such as the pole. If p( z) can be written as
q(z) p(z) = ( z zo )n'
(4.43)
where n > 0 is a positive integer while q(z) =/= 0 is analytic everywhere in the neighbourhood containing z = zo, we say that p( z) has a pole of order n at z = z0 . The above definition is equivalent to say that the following limit is finite
zzo
lim [p(z)(z zot] = (,
11(11 < oo,
( E C.
(4.44)
Any analytic function f (z) can be expanded in terms of the Taylor series
(4.45)
This expansion is valid inside the analytic region. However, if the function f (z) has a pole of order n at z = zo and it is analytic everywhere except at the pole, we can then expand the function p(z) = (zz 0 )n f(z) in the standard Taylor expansion.
69
are called the analytic part. fr f(z)dz.zo) + . is called a contour or line integral of f (z).49) .4 Complex Integrals Given a function f (z) that is continuous on a piecewise smooth curve r. The tern1s with the inverse powers an/(zzo)n+ ... while the usual terms ao(z. In this series... (4.zo)k. + ok(z. + ok(z.zo)k + . the Laurent series can be written as f(z) = L 00 ok(z. This integral has similar properties as the real integral frlof(z) + fig(z)]dz = o £ f(z)dz + ...8£g(z)dz.4 Complex Integrals Complex Variables This means that original function f (z) can be written as a power series ) !( z=( O_n O_t ) zzon + . ..46) which is called a Laurent series.48) If F(z) is analytic and F'(z) = f(z) is continuous along a curve r..F[z(a)]. and it is an extension of the Taylor series. Furthermore. 70 (4.zo) + .4.. +at/(zzo) are called the principal part of the series... In general. 4.zo)k + . (4.. (4. the most in1portant coefficient is probably Ot which is called the residue of f(z) at the pole z = zo.47) k=n where n may be extended to include an infinite number of terms n+ oo. it is often assumed that On =F 0. +( zzo ) oo(z. then the integral over r. then l f(z)dz = F[z(b)].
vdy) +i fr (vdx + udy). we have Cauchy's Integral Theorem. y). fr SS(z) Figure 4. this becomes I = Jn (. [ 8u (4. which states that the integral of any analytic function f(z) on a simply closed path r in a simply connected domain n is zero. Thus.50) By using the Green theorem.Complex Variables 4.3: Contours for Cauchy integrals. 8v (4.ay . If the close path is decomposed into two paths with 71 .4 Complex Integrals Cauchy's Integral Theorem We say a path is simply closed if its end points and initial points coincide and the curve does not cross itself. That is f(z)dz = 0.51) Fron1 the CauchyRiemann equations. the integral on a sin1ply closed path I= = lr fr f(z)dz = fr (u + iv)(dx + idy)] (udx.ay )dxdy. For an analytic function f(z) = u(x. y) + iv(x. we know that both integrals are zero. [ 8u av .ax )dxdy + zJn (ax . This theorem is very important as it has interesting consequences.
(4.53) since the contribution from the cut is zero. The only nonzero contribution over a small circular contour is the residue a1· We have I= f k p(z)dz = 21ria1 = 21ri Res[p(z)]j . ~ 72 (4..3). we can use the Laurent series for p( z) (4. this can be extended to the integrals over two closed paths r and y such that y is a very small circular path inside r. then form a close path. +ak ( zzo) k + . lr2 That is to say that the integrals over any curve between two points are independent of path. .zo +ao(zzo ) + . I= frp(z)dz. In fact.56) .52) lr1 { f(z)dz = f f(z)dz. the Cauchy integral theorem implies that fr f(z)dz = L f(z)dz..4 Complex Integrals Complex Variables reverse directions r 1and r 2(see Figure 4. Using a small cut with two curves C1 and C2 so that these two curves combine with r and y form a closed contour (see Figure 4.. we can make the contour y sufficiently small to enclose just around the pole.55) so that the expansion can be integrated term by term around a path. and this makes the calculation of the integral much easier in some cases. (4.54) p(z) = ( a1 ) z. For an analytic function with a pole.zo) over any simply closed path r enclosing a point zo in the domain n.4. For the integral of p( z) = f (z) / (z ..3). which leads to r1 and r2 (4. This property becomes very useful for evaluation of integrals.
1)1 }. .58) The residue theoren1 serves a powerful tool for calculating some real integrals and summation of series..4: Contour for the integral J(o. the residue theorem states i f(z)dz = 21fi r L k=t N Resf(z)lzk' where r is a simple closed path enclosing all these isolated points. nl 00 00 Figure 4..d z = f(zo).4 Complex Integrals which can be written in terms of f(z) as f(z) .(N.1 . the following forn1ula gives a quick way to calculate the residue Resf(z)lzo . (4. (3).. this can be extended for higher derivatives.~o _ 1 . ZN.!. especially when 73 . Residue Theorem For any analytic f (z) function in a domain n except isolated singularities at finite points Z1. . If f(z) has a pole of order Nat zo. z2. 21rz r z. dNt [(z. .57) Similarly. and we have Jr (z..zo i (4.zo)n+l j f(z) dz = 21rij(n)(zo)..zo)N f(z)] dzN1 .Complex Variables 4.
Using the residue theorem.59) is that this mapping preserves the angles between curves and their orientations. we have eio2c.w EC. 0 Example 4. we have f oe 1 2 00 ( + j34 d( = fJ /l2. One of the widely used mappings is 1\tlobius linear fractional transformation W= az+f3 ' "fZ + 8 (4.. a contour is closed. /(( + i/32 ). Combining the arc with the horizontal line from the integral limits from oo to oo along the (axis.60) 74 . Hence. The essence of a conformal mapping w = f(z). we have In a special case when o = 0. (4.i(32 eio c. As the function 02 </> = ei < (/34 + ( 2 ) has two poles ( = +i/32 and i/3 2 and only one / pole ( = +i/32 is in the upper half plane.9: In order to evaluate the integral it is necessa1y to construct a contour (see Figure 4.4). we can construct a contour to encircle the pole at ( = i{J2 by adding an additional arc at the infinity (( + oo) on the upper half plane. z. 1r Another important topic in complex variables is the conformal mapping. /(( + ij32) f(() = ( .4.4 Complex Integrals Complex Variables the integrand is a function of sin and cos that can be changed into a complex integral.ij32 ' 4> = where f(() = 2 ( . The real integral f~oo 1/J(x)dx becomes 27ri multiplying the sun1 of the residues of 1/J(x) at the poles in the upper halfspace.
"/. Conformal mappings are useful in solving steadystate problems involving harmonic functions by transforming the problem from a complicated geon1etrical domain to a regular domain such as circles and rectangles. rotations.4 Complex Integrals By choosing the appropriate coefficients o. inversion.Complex Variables 4. and expansions and contractions. 75 .. this mapping can include all major geometrical transformations such as translations. {J E C. (3. and subsequently the techniques based on conformal mapping are widely used in solving Laplace's equation in engineering.
.
the Riccati equation: y' + a(x)y 2 + b(x)y = c(x).1 Introduction An ordinary differential equation (ODE) is a relationship between a function y( x) of an independent variable x and its derivatives y'. their relationship can in general be written as a differential equation or even a set of differential equations.y(n)) I _ '..y.2) 77 .y. .1) The solution of the equation is a function y = f (x). temperature.. If the variables or quantities (such as velocity.y. . satisfying the equation for all X in a given domain n. (5. 5. . y".l' Q.. The order of the differential equation is equal to the order n of the highest derivative in the equation.. It can be written in a generic form . y(n).Chapter 5 Ordinary Differential Equations 1viost n1athematical models in engineering are formulated in terms of differential equations. . pressure) change with other independent variables such as spatial coordinates and tin1e. (5.y II . Thus.( x.
(5. If all these forms are not possible. An equation is called linear if it can be arranged into the form an(x)y(n) + .. the equation is homogeneous. = 5.5) 78 .. If the right hand side is zero or ¢(x) 0. and the following equation of Eulertype (5. the alternatives are to use approximate and numerical solutions. not on y or any derivatives. the solutions can only be obtained for a few simple types.1 First Order ODEs Linear ODEs The general form of a first order linear differential equation can be written as y' + a(x)y = b(x). series form and integral form. then the equation is nonlinear. The solution of an ordinary differential equation is not always straightforward. Even for linear equations. and it is usually very complicated for nonlinear equations. The solution of a differential equation generally falls into three types: closed form. (5. Therefore. The degree of the equation is defined as the power to which the highest derivative occurs.3) is a second order. It is called nonhomogeneous if ¢( x) =I= 0. A closed form solution is the type of solution that can be expressed in terms of elementary functions and some arbitrary constants.2.5. both Riccati equation and Euler equation are of the first degree. If any of the coefficients is a function of y or any of its derivatives. + a1 (x)y' + ao(x)y = ¢(x).2 First Order ODEs Ordinary Differential Equations is a first order ODE.2 5.4) where all the coefficients depend on x only. Series solutions are the ones that can be expressed in terms of a series when a closedform is not possible for certain type of equations. The integral form of solutions or quadrature is sometimes the only form of solutions that are possible.
7) yei a(x)dx = So its solution becomes J b(x)ei a(x)dxdx +C.I a(x)dx j b(x)ei a(x)dxdx + ce. (5. 0 Example 5. However. one integration constant usually requires one condition to determine it.1: lVe now try to solve the ordina1y differential equation ~~ + ty(t) = t with an initial condition y(O) = 0.I a(x)dx.I tdt 79 . then y(x = 0) =Yo is boundary condition at x = 0. As a(t) = t. then y(t = 0) =Yo can be thought of as the initial condition at t = 0. we have y'ei a(x)dx + a(x)yei a(x)dx = b(x)ei a(x)dx.I tdt j (t)ei tdtdt + ce. Multiplying both sides of the equation by explf a(x)dx]. if x = t means tilne. By simple integration.2 First Order ODEs This equation is always solvable using the integrating factor and it has a closed form solution. we have (5. and these extra requirements are usually the initial condition when time is zero or boundary conditions at some given points which are at the domain boundary.6) which can be written as [yei a(x)dx]' = b(x)ei a(x)dx.8) y(x) = e. which is often called the integrating factor. its general solution is y(t) = e.Ordinary Differential Equations 5.9) where C is an integration constant. However. (5. If xis spatial coordinate. Nevertheless. b(t) = t. The integration constant can be determined if extra requirements are given. (5. the classification of conditions may also depend on the meaning of the independent x.
. 0 Example 5.2: To solve y'(x) + xy = y 20 . the solution becomes 0 5.12) which is a standard first order linear differential equation whose general solution is given earlier in this section. (5. it reduces to a standard first order linear ordinary differential equation.n)q(x). the Bernoulli's equation can be written in the generic form y' + p(x)y = q(x)yn. For example. By dividing both sides by yn and using the change of variables 1 u(x) = .11) u' + (1. or C=l.10) In the case of n = 1.n)p(x)u = (1.5). sometimes a transform or change of variables can convert it into the standard first order linear equation (5.2 First Order ODEs t2 t2 t2 Ordinary Differential Equations t2 = e"TeT +GeT= 1 +GeT. Thus. 80 . (5.n)y' ' yn (5. whose general solution is u(x) = Ae 19x + 1. and we u' = 19y' jy 20 .19xu = 19. From the initial condition y(O) = 0 at t = 0.2 Nonlinear ODEs For some nonlinear first order ordinary differential equations.2. yn1 we have U= 1 (1. we first use u(x) = 1/y 19 . we have 0 = 1 +C. The original equation becomes u'.5.
Ordinary Differential Equations 5..(x). We have y(x) = Yc(x) + y. . the corresponding linearly independent terms can then be replaced by erx[Acos(wx) + Bsin(wx)].15) Assuming y = AeAx.(x). .. at.16) which has n roots in general.13) its general solution y( x) consists of two parts: the complementary function Yc( x) and the particular integral or particular solution y. the solution to the original equation becomes 19 1 = Ael9x y + 1' or 0 5. . For the special case of higherorder ODEs where all the coefficients an. Then.3 Higher Order ODEs Therefore. and complex roots always occur in pairs A= r ± iw. we get the polynomial equation of characteristics (5. (5... An. ao are constants.14) 5.3.1 General Solution The complementary function is the solution of the linear homogeneous equation with constant coefficients and can be written in a generic form (n) anYc + anlYe(n1) + ··· + alYc + ao I 0· (5. the solution can be expressed as the summation of various terms Yc(x) = Ek=l CkeAkx if the polynomial has n distinct zeros At... 81 . + aty' + aoy = f(x). For complex roots. (5.3 Higher Order ODEs Higher order ODEs are more complicated to solve even for the linear equations. any(n) + .
Assuming that y = AeAx.2y"(x). ex and e 2 x.\ + 1)(.2(asinx. ekx. + Qxn (polynomial). N).1)(.. we have (acosx + bsinx)...bcosx) + 2(asinx + bcosx) 82 = sinx. 2. For f(x) = sin(ax) or cos(ax).2) = 0.13) so that the coefficients A and B can be determined. . Similarly.2y"(x)y' (x) + 2y( x) = sin x.3 Higher Order ODEs Ordinary Differential Equations The particular solution y.13).bsinx) . 1. we first tzy to solve its complementary equation or homogeneous equation y"'(x). Thus..(x) is any y(x) that satisfies the original inhomogeneous equation (5. For f(x) = ekxxn. (acosx.9: In order to solve the equation y"'(x). For most of the combinations of basic functions such as sin x. Qxn)ekx. f (x) = ekx sin ax or f (x) = ekx cos ax.. = (A+ Bx + Cx 2 + . the particular solutions can take various forms.X. y.. The general complementazy function becomes As the function f (x) = sin x. we have the characteristic equation or (.y'(x) + 2y(x) = 0. then we can try = A sin ax+ B sin ax. we can use = ekx (A sin ax + B cos ax).\. 0 Example 5. y.. Substituting this into the original equation. y. we have to find its complementary function Yc(x) and its particular integral y*(x). the method of the undetermined c~ efficients is widely used. three basic solutions are ex. we then try = A+ Bx + Cx 2 + . and xn. thus we can assume that the particular integral takes the form y*(x) = asinx + bcosx. y. cos x. Depending on the form of the function f (x). We then substitute it into the original equation (5.5. For a polynomial f(x) = xn(n = 0. . l\!Iore general cases and their particular solutions can be found in various textbooks.
we have the characteristic equation Its complementary function is As the right hand side f (x) = ex is of the same form as the first term of Yc.2a + 4b = 0.a+ 2b) cosx = 0.x. whose solution becomes a= 5' 1 1 b = 10' Now the particular integral becomes y*(x) = Finally. Let us see an example..4: Consider the equation y"(x). .cosx + Aex + Bex + Ce 2 x. there are some problems in the case when the righthand side of the differential equation has the same form as part of the complementary function. Thus.. 0 Example 5. However.1) sinx + (a+ 2b.. 0 ~sinx + .Ordinary Differential Equations 5. Usingy(x) = Ke>. 5 10 0 The methods for finding particular integrals work for most cases. In this case.!:. the general solution y= ~ sinx + 11 cosx. we have 4a + 2b = 1. then standard trial function aex cannot be a particular integral 83 . the trial function should include one higher order term obtained by multiplying the standard trial function by the lowest integer power of x so that the product does not appear in any term of the complementary function.3 Higher Order ODEs or (b + 2a + b + 2a.3y'(x) + 2y(x) =ex.
3(a + b + bx) + 2(a + bx) = 1. As there is no constraint for a. then solve each factor separately. On the other hand.x = Ae>.x = Ane>.x.. any nonzero aex can be absorbed into Aex. An. The beauty of using the differential operator form is that one can factorize it in the same way as for factorizing polynomials. the general solution becomes 0 5. A.5.5: To find the particular integral for the equation y""' + 2y = 84 17e2 x. Thus. = (a+ bx)ex first..17) Since we know that De>. 0 Example 5. so they are equivalent to D ..dx· d (5.. the integral operator nt = J dx is just the inverse of the differentiation.3. In fact. or b = 1.3 Higher Order ODEs Ordinary Differential Equations as it automatically satisfies the homogenous equation y"(x)3y'(x)+ 2y(x) = 0. and we have Dividing both sides by ex.x and Dne>. differential operators are very useful in finding out both the complementary functions and particular integral. A differential operator D is defined as D=. we have (a+ 2b + bx). and nn .. thus we take it to be zero (a= 0 )..2 Differential Operator A very useful technique is to use the method of differential operator D. Thus. We have to tzy y. Thus..... any polynomial P(D) will map to P(A). .
For a secondorder differential equation.x via sin() = ~ ( ei 9 eie) and coshx =(ex+ ex)/2. dv (5.. x). v. + a1(x)y1 + ao(x)y + ¢(x). an nthorder linear equation can always be written as a linear system of n firstorder differential equations. Higher order differential equations can conveniently be written as a system of differential equations. sinh x and others.15).. Ynt]T. cos x.18) which is a system for u = [y Y1 Y2 . (5.. v. x).. 5. we have 17e2x y* e2x = 25 +2 = 2" 0 This method also works for sin x.Ordinary Differential Equations we get or y = D5 +2e * 5. then the system is said to be autonomous.. and this is because they are related to e>.4 Linear System 17 2x Since D 5 ~+ . In fact. Such a system 85 .4 Linear System For a linear n order equation (5. A linear system of ODEs is more suitable for mathematical analysis and numerical integration.19) If the independent variable x does not appear explicitly in f and g. we can always write it in the following form du dx = f(u.\ 5 = 25 . dx = g(u. it can be always written as a linear system an(X)Y~1 = an1(X)Yn1 + .
.25) where the known function p( x) is differentiable.5. . if the problen1 has solutions. (5. then this becomes an eigenvalue problem for matrix A.. ann Un or ii =Au. y(b) = 0. For the obvious reason. (5. (5. The para1neter A to be determined can only take certain values An.23) 5. . A general linear system of nth order can be written as l( ( u2 ~~ . r(x) are continuous. 86 .5 SturmLiouville Equation One of the commonly used secondorder ordinary differential equation is the SturmLiouville equation in the interval x E [a. ' (5.5 SturmLiouville Equation Ordinary Differential Equations has important properties. . (5. (5. and the known functions q(x). . this problem is called the SturmLiouville eigenvalue problem.24) with the boundary conditions y(a) + ay'(a) = 0.. called the eigenvalues. For simplicity and in keeping with the convention.. . b] d dy dx [p(x) dx] + q(x)y + Ar(x)y = + (3y'(b) 0.Al)v = 0.20) Un ant an2 . atn a2n  .21) If u = vexp(At). we use t = x and u = dufdt in our following discussion. l( l Ut u2 . . . au a21 a12 a22 .. (A.22) which will have a nontrivial solution only if det(A AI) = 0. .
33) .Ordinary Differential Equations 5. (x)r(x)dx = Omn· At < A2 < .n.27) It is possible to arrange the eigenvalues in an increasing order + oo. (5... r(x) =/= 0.~ (x)l/1>. (5. called eigenfunctions. there is a corresponding solution 1/J>.~(x)•h. That is l.P>.28) Sometimes.(x)r(x)dx = 0. the Riccati equation can be written in the generic form y' = p(x) + q(x)y + r(x)y 2 . For example. it is possible to transform a nonlinear equation into a standard linear equation. By using the transform y(x) = or u'(x) r(x)u(x)' e.26) ll/1>. where (5.31) u".5 SturmLiouville Equation For each eigenvalue An. (5. < An < ...J r(x)y(x)dx. The SturmLiouville theory states that for two different eigenvalues Am =/= Am their eigenfunctions are orthogonal. then it reduces to a first order linear ODE. or more generally (5.. (5.29) If r(x) = 0.P(x)u' + Q(x)u = 0.30) u(x) we have = (5.32) P(x) = r'(x) + r(x)q(x) r(x) 87 Q(x) = r(x)p(x)... (5.
Thus.34) is in fact an eigenvalue problem as it can be written as (xy')' + (x.5.)y = 0. but we only focus on the case when the values of v are integers.39) . so that the first term of an is not zero.1 Bessel Equation The wellknown Bessel equation (5.36) where s is a parameter to be determined. (5.37) (5.l)anxn+s + L(n + s)anxn+s n=O n=O +L 00 00 anxn+s+2 l/2 n=O 88 L anxn+s = 0. we can always change the value of s. If a0 = 0. Thus. X l/2 (5.35) Although v can be any real values. This method is often called the Frobenius method which is essentially an expansion in terms of a Laurant series. we assume that the solution can be written as a series expansion in the form y(x) =xs I: anxn =I: anxn+s.5 SturmLiouville Equation Ordinary Differential Equations 5.38) Substituting these expression into the Bessel equation. we assume in general ao =/= 0.5. In order to solve this equation. we have 00 00 L(k + s)(k + s. we have dy = dx L an(n + s)xn+s1' n=O 00 (5. n=O (5. n=O n=O 00 00 ao =/= 0.
40) Since a 0 =F 0.Ordinary Differential Equations 5. It has many interesting properties: (5.. we finally have = (5.5 SturmLiouville Equation Equating the coefficients of the same power xn(n = 0... 3. This is the Bessel function of the first kind. we can get some recurrence relationships. we thus have s = ±v.1) + an(n + s) or + an2.. . .. or a1 (5. 1 2 (5.45) Since we now know that a1 = 0... Jv = ~ ~22 nnl(~n+. 2.. . 4. oo (l)n so that k = 0. (s =F . (5.. we have an(n + s)(n + s.41) (5.48) 89 . 1. we have (5.. From n = 1 terms.. 2.44) an2 (n + s) 2 . 1.v_)!x2~n+v ' ~ (5. = a1 0.v 2 = (5.. For n = 0..46) where we have used n The function J v = 2k = 0. we have at(2s + 1) = 0. 4. .43) For the rest of terms n = 2...v 2an = an2 n(n + 2v)' 0.). All the even terms contain the factor a0 .. .. 2. thus aa =as= a7 = . ). 3.42) = 0...47) is called the Bessel function of the order v..
51) Euler Buckling As an example. let us look at the buckling of an Euler column which is essentially an elastic rod with one pinjointed end and the applied axial load P at the other end. = 0). Its general solution is u = Asinax + Bcosax. (5.54) Applying the boundary conditions.55) (at x = L).52) or 2 p = EI' a (5. (5.5. the Euler bean1 theory gives the following governing equation (5.49) and fox uJo(u)du = xJ1 (x). Its Young's modulus is E and its second moment of area is I= J y 2 dA =canst (for a given geon1etry).5 SturmLiouville Equation Ordinary Differential Equations I: v=oo 00 J.56) . Let u(x) be the transverse displacement.5. = 1. (5. (5. 90 Asin(aL) = 0. Jv(x) = ( 1)"' J.53) which is an eigenvalue problem.(x). we have at the fixed end (at x and at the free end u=O. (a :/= /3). The column has a length of L. (5. (5.2 xJv(ox)Jv(f3x)dx = 0. B=O.50) There are other properties as well such as the orthogonality l 5.
(5.61) . . different equations will usually require different n1ethods. Even for relatively simple secondorder ODEs. its total angular momentum L is conserved.riJ.L = 0 (trivial). the force is purely radial in the polar coordinates.3 Nonlinear SecondOrder ODEs For higherorder nonlinear ordinary differential equations. ..n27r2EJ ' £2 (n=1.5.a2E/. therefore. This can be beautifully demonstrated by the solution process of finding the orbit of a satellite.. As the satellite orbits the Earth.' 7r2EJ (5. 27r. and there is no guarantee that you can find the solution. (5. n1r. 1r.. ). or r 2 iJ = Lfm = h = const.5 SturmLiouville Equation Thus we have two kinds of solutions either A = 0 or sin( o. L = mr 2 !f/f = const.3. we have m[d r _ r(d8) 2 ] = dt 2 dt 91 2 _ GMm r2 ' (5.57) Therefore. (5.L) = 0.Ordinary Differential Equations 5.. we have u(x) = 0 which is a trivial solution. . we have P .59) The solutions have fixed mode shapes (sine functions) at some critical values (eigenvalues Pn)· The lowest eigenvalue is p* = y. there is no general solution technique.2.L) or o. Using Newton's second law of motion and Newton's law of gravity. 5. The radial acceleration is ar = f ..60) which is the Euler buckling load for an elastic rod. One of the best methods is the change of variables so that the nonlinear equation can be transformed into a linear ordinary differential equation or one that can be solved by other methods. So the nontrivial solution requires that sin( o. ..58) = 0. For A = 0.
(5.70) .5. we have dr _ _ _ 2 du _ h du dt u dt d(J ' d u = dr = h ~u d(J = h2u2 ~u dt 2 dt d8 2 dt d8 2 . m we then have L (5. Now the governing equation becomes 2 (5.62) d2 r dt2 . it is straightforward to write down its solution u = S + Acos8 + Bcos8 = S[l + ecos(8 + 'l/7)].r3 h2 + .66) (5.68) Since this is a secondorder linear ordinary differential equation.S[l+ecos(B+'l/1)]' 92 . Then.69) where A and B are integration constants. which can be converted into the eccentricity e and the initial phase 'l/J. G is the universal constant of gravitation.63) which is a nonlinear equation.64) which is equivalent to dt = d(J / ( hu2 ) and this can be used to eliminate t.65) and (5. the conservation of angular momentum becomes (5.r = GM 0.67) or (5.5 SturmLiouville Equation Ordinary Differential Equations where M and m are the masses of the Earth and the satellite. respectively. Using the conservation of angular momentum so that h=. By using the change of variables u = lfr. The final solution is 1 r~~~ (5. (5.
along the n1ajor axis) and one focus at the origin.71) which is the orbit for satellites and planets.5 SturmLiouville Equation which corresponds to an ellipse where e is the eccentricity of the orbit. If we set the polar coordinates in such a way that 'l/J = 0 (say. then the equation simply becomes h2 r= GM[l + ecos8] ' (5.Ordinary Differential Equations 5. 93 .
.
.. we say that the difference equation is homogeneous. 1. we need the difference equation or the recurrence equation as the counterpart in differential equations. and even the independent variables such as time are not continuously counted or measured (in seconds or years). In this book. the equation is called the linear equation with constant coefficients (as the counterpart in the differential equations). .Chapter 6 Recurrence Equations 6. 1V) are coefficients which are not functions of y. A linear difference equation of Norder can generally be written as aoYn + atYn1 + a2Yn2 + . In fact... If all the coefficients ai are constant. especially the linear ones. . If f(n) = 0. some problems such as finance are concerned with quantities (say.1) where ai(i = 0. .. (6. we only focus on 95 .. + aNYnN = f(n). For this type of problem..1 Linear Difference Equations Differential equations always concern the quantities that vary continuously. interest rate) that are discrete and do not vary continuously. Yn [= y(n)] is the value of the variable y at n = 0. there many similarity between difference equations and differential equations. 2. However.
. Thus. 2. (6.6) It has two solutions A1 and A2 in general.5). we have ayn + bYn1 + CYn2 = f(n). we reach a characteristic equation aA 2 + bA + c = 0. ). we assume that Un = o:An (similar to Yc = Ae>. 3.2) which can also be written as (6. 1. The most fan1ous difference equation is probably the recurrence relation Yn = Yn1 + Yn2. If f(n) = 0. The recurrence equation is valid for n = 2.6.5) while Pn is any particular solution of (6. 1.4) for the Fibonacci sequence (0. In order to obtain un.7) Find the solution of Yn + 3Ynl + 2Yn2 = 0.1 Linear Difference Equa. (6. 3. Substituting into (6. (6.x for differential equations). .. 8. Therefore. y(1) = 1. the general solution of a difference equation Yn = Un + Pn where Un is the complementary solution to the homogeneous equation (6..1). we have un = AA} 0 Example 6. 5. ..tions Recurrence Equations the secondorder linear difference equation. we say the equation is homogeneous. and the initial conditions are y(O) = 0. Similar to the solution procedure of linear ordinary differential equations..2). 96 . (6.3) where g(n) = f(n.6: + BA2.
Therefore. Now we have A = Yn 1/ v'5. we try Pn =a. we have (A+ 1){A + 2) = 0. we first use Yo = 1. tl1e general solution can be written as Yn = A(1t + B(2t. we try 97 . 0 Example 6.1 Linear Difference Equations This is a homogeneous equation. For n = 1. The general solution is therefore In order to determine A and B. we assume Un = or O'An. For f(n) = kn. = 0. homogeneous equa whose solution is A= 1 ±2v's. we can use the similar technique used in ordinary differential equations.Recurrence Equations 6. 0 For given initial values. The general solution becomes y'5 = _1 ( 1 + v'5)n __ ( 1 . we try Pn =a+ {3n where a and {3 will be determined. B y'5 = 1/ ..5. This is a A 1 = 0. Thus. we can determine the constant in the general solution so that an exact expression can be obtained. Y1 = 1 gives 1 = A( 1 +2 y5) + B( 1 2 y5 ). For f (n) = k = const.y5)n 1 2 2 . 7: difference equation The Fibonacci sequence is governed by the Yn with initial conditions Yo tion. 0 For finding a particular solution Pn. The characteristic equation is A2  = Yn1 + Yn2. For f(n) = k1n. y 1 = 1. we have A1 = 1 and A2 = 2. we get O=A+B.
6 6.8: For the equation Yn+t. 6 .6Ynt = n. the general solution is k· y(n) = A(2)n + B3n ~ . . or (. and the characteristic equation is .6[a + b(n. the analysis is not easy. 98 .(a+ bn). its complementary equation is Yn+l . Other forms such as polynomials can be done in a similar fashion.6 = 0. what happens to the nonlinear equations? The main problem with nonlinear equations is that there is no general solution techniques available for most cases. The general complementary solution can now be written as Finding any pa.\ 2  . this equation is valid.X. \Ve now have a+ b(n + 1).Yn .3) = 0.Yn. nonlinear equations may have chaotic behaviour under appropriate conditions..2 Chaos and Dynamical Systems As you may have noticed that the above analysis is mainly about the linear equations. As for any n. Even for the simplest case.X+ 2)(..6bn = n. and b= Finally. therefore we have a = .rticular solution requires that we t1y Pn = a+ bn.6Yn1 = 0.2 Chaos and Dynamical Systems Pn = Recurrence Equations O:')'n. In addition. = n. even for the simplest equations. 0 Example 6. the behaviour of nonlinear equations is very complex. Often.\.6.1)] or 6a + 7b.
0198.6 0. .01. y9 = 0.. YB = 0..2381. . (6.4972. If the initial value Yt = 0 or Yt = 1.5000. 6.Recurrence Equations 6. . This is the wellstudied logistic map. y 10 = 0.1: Variations of Yn for v = 2 and v = 3. then the system seems to be trapped in the state Yn = 0 (n=2.0388. However. Y4 = 0. For a given value v = 2. it does not work.01. Y3 = 0. Y7 = 0. if we use a slight difference value (say) Yt = 0.3628.2 Chaos and Dynamical Systems 0. we now briefly introduce the concept of chaos by studying the following nonlinear difference equation Yn+l = VYn(1. which is essentially an iteration process because all the values Y2.1 Bifurcations and Chaos In order to show this.5000.8) where v is a constant parameter. can be determined for a given parameter v and an initial condition Yt· This is one of the simplest dynanlical systems.1381.3.8 0. Y6 = 0. Y2 = 0... then we have Yt = 0. Ys = 0...2 00 20 40 >o c n 60 80 100 20 40 n 60 80 100 Figure 6. we can use a computer or a pocket calculator to do these calculations. Y3.Yn). This seemingly simple equation is in fact very complicated.0746.9) 99 . ). (6.8 ~u1 0.. y 11 = 0.4623. If you try to use the method to solve the linear difference equations discussed earlier.2.
If we do the same simulations again using a different value v = 3. Now we have a difference phenomenon. If we use a different parameter v = 3 and run the simulations again from the same initial value YI = 0.2). the system bifurcates into two states.2 n n Figure 6.1. the results are also plotted on the right in Figure 6. The final values do not settle to a single value. they oscillate between two states or two final values Yoo = 0.01.3.001 to v = 4. From the chaotic map. It seems that the system is attracted 100 . It gives a detailed map about how the system behaves.5000. the values are trapped or attracted to a single value or state Yoo = 0.6.2: Variations of Yn for v = 3. The iteration system bifurcates into two states as the parameter v increases. every values seems difference. there are four final states.6560. Then. the system settles (or attracts) to a single state.45.6770 and Yoo* = 0.000.5 (shown in Figure 6. Instead. the final state is zero (the system is attracted to a stable state Yoo = 0). we see that for v < 1. the system is chaotic and the values looks like a randmn noise. For v = 4.1. we can plot out the number of states (after 1V = 500 iterations) and then we get a bifurcation map shown in Figure 6. For 3 < v < 3.2 Chaos and Dynamical Systems Recurrence Equations 0. Following exactly the same process but using different values of v ranging from v = 0. All the values are then plotted in a graph as shown in Figure 6.5 and v = 4. For 1 < v < 3.
. •. 0.. by these discrete states.···1 3 3.Recurrence Equations 6... say.9 3.6 .2 0 .J..4 3.4 3.Yn) + Yn+ 1 because you may try many times to simu late the same system at the same initial value Yo = 0.::: v* where v* ~ 3..8 Figure 6... we should get the same set of values Yt. i 3. the map is also called the attractors of the dynamical system. The system becomes chaotic for v . '• 0... So where is the chaos anyway? The problem is that this system is very sensitive to the small variation in the initial value y 0 ..··..57 is the critical value.5 3. For this reason.7 v lift: 3. Then.3: Bifurcation diagram and chaos .. You are right. 101 .2 3.. It is a bit surprising for a seemingly determined system VYn ( 1 .01 (say) and parameter v.2 Chaos and Dynamical Systems Figure 6..·= 4 .8 3. ···. .6 ··. If there is a tiny different...75 O. Y2.4: Bifurcation and similarity.:~.:·.
2 Dynamic Reconstruction When a nonlinear system becomes chaotic. 2 Yn. . 6. This selfsimilarity is one of the typical behaviours of chaotic systems and it also occurs in other nonlinear systems such as (6. it seems that it is very difficult to understand the behaviour.ooo). If we study Figure 6.. (6. there may be some regularity such as attractors and selfsimilarity as we have seen earlier. we can see there is a similarity between the whole map and its certain part (enclosed by a box) which is enlarged and plotted on the right in the same figure. then the set of values you get will be completely different. In addition. even the computer simulations can only use a finite number of digits.2. the ratio between the lengths of the parameter intervals for two successive bifurcation approaches the Feigenbaum constant 8F = 4.5. However.6. but it does provide some vivid picture for the nature of chaos and sensitivity to the initial uncertainty in chaotic systems.01 Recurrence Equations ± 0. Ysoo or Yt. Obviously.000000001 or even 10.1000 difference..ooo. In some case. Since there is always uncertainty in the real world. In fact. but we do know it only depends on one parameter 102 . there is a famous 'butterfly effect'. and there is no way of predicting the final values (say. it is even possible to reconstruct the system itself.2 Chaos and Dynamical Systems y 0 = 0.10) and Yn+l = ASlll \ . this is exaggerated too much. It says that the wing flip of a butterfly in Africa can cause a tornado in America or anywhere.4 closely. so the chaos is intrinsic for nonlinear dynamical systems. which is universal for the chaotic systen1s.669 . Suppose we do not know the exact form of the dynamical system.11) which are plotted in Figure 6.
we can actually reconstruct the system using the sorted data and plotting Yn versus Yn.. For example.6 2.Xsin2 Yn· Bifurcation for Yn+1 = Asin Yn and Yn+1 = and we can observe the states Y1 . We plot U versus V. If there are N observations.Yn1). we obtain the dynamical system Yn = VYn1(1. from the 100 data for the nonlinear system discussed in the previous section. YN as one set and V = Yb Y2. From our observations. . Yn (to a certain degree).4 A.. . Yn1 as another set.8 3 Figure 6. Y2.. .Recurrence Equations 6..y). . The mapping is then y ~+ y(1.y). With a free parameter v and discrete time n.12) We see that even a simple nonlinear equation in 1D can show the rich complexity of dynan1ical behaviour. 2. 6.Yn)· (6. or Yn+1 = VYn(l . The parabola is essentially the original function y(1.1.2 Chaos and Dynamical Systems 2. Now we briefly look at a Lorenz system as a classical example.5: . The Lorenz equa103 .2 2.5 0.5 Yn+1=A sin2 (yn) 0 2 2. the constructed system is a parabola as plotted in Figure 6. ..5 2 1. .6.3 Lorenz Attractor The Lorenz attractor was first discovered by Edward Lorenz when he studied the weather model in 1963.2.. then the system can be dynamically reconstructed. we have U = Y2.
6 0.7: Lorenz strange attractor. tions are x= il = f3x + yz. the system becomes chaotic.8 Figure 6. ~ is the Prandtl number.7. For certain ranges of parameters.6. The system moves around in a curious orbit in 3D as shown in Figure 6.13) z= ("Y.2 Chaos and Dynamical Systems Recurrence Equations \ 0. Specifically. and the 104 .2 '\ \ 0.z. (6. "Y = 28. and "Y is the Rayleigh number. for example f3 = 8/3.6: Dynamic reconstruction of the nonlinear function. and /3. ~' "Y are parameters. where·= ft. ~(z y). Figure 6. ~ = 10.x)y.4 0.
3 Selfsimilarity and Fractals orbit is bounded. we know that a point is zero dimension and a line is one dimension.8. Two classical exa1nples are fern leaves and the pattern of lightening. By starting with this generator (left) denoted as 8 1 . The basic idea of fractals can be demonstrated using the generation of the Koch curve (also called the Koch snowflake curve). This is because there are 4 more segments added. This is where the fractals arise. For the nth generation. but each segment is only 1/3 of the previous one in length. and thus chaos is very different from the random noise. we get the Koch curve. and thus the Lorenz strange attractor. and it is also linked with the concept of fractals. But there are other possibilities in nature that the dimension can be a fraction or a real number. If we continue this procedure many times. 6.8. but not periodic or convergent.Recurrence Equations 6. Shnilarly. we have the total length 105 . The generator or seed of this curve is the four straight line segments as shown in Figure 6. This strange characteristic gives the name 'strange' to the attractor. we get the second generation curve (middle. 3). All these dimensions are integers (0.3 Selfsimilarity and Fractals In the bifurcation map of chaotic systems. The chaotic map looks like a butterfly. even though they sometimes may look the same. selfsimilarity and fractals occur frequently in nature. This selfsimilarity is quite universal in many phenomena. This observation suggests that even if the system is chaotic. 2. In fact. 1. we have seen the selfsimilarity. 8 2 ). there is still some pattern or certain regularity in its behaviour. In geometry. A snapshot at generation 5 is shown on the right in Figure 6. for the 82 curve. the total length is 42 /3 2 . the dimension for a plane is two and the dimension for a volume is three. and replacing each straight line segment by the generator itself. The total length of the seed 81 is Lt = 4/3 since it has four segments and each segment has a length of 1/3.
Zn + C. and the total number is N(f) = 4n. the iterations continue until certain criteria are met.14) For the Koch curve. we say that c belongs to the Mandelbrot set. which is related to the iterative dynamical systems. (6. In this set. the smallest number N (f) of balls to cover the curve r will varies with f. where z E C is in the complex plane. For a curve r in a compact set of a metric space and for each f as the radius of the small balls. and c is a complex number to be tested. the magnitude of z during the iterations changes. it either stays small ( < 2) or it will eventually surpass two (> 2). The fractal dimension can be defined by covering the curve using small balls (circles in a plane) and counting the number of balls with radius r = f > 0. for each given c.h m . Hence. otherwise.= E+0 1 3n n n 1 3 ~ 1.16) Zn+l. ln4n ln4 d= . That is to say. it is not 106 . e+0 In f (6. or 2 (6. the minimum radius of the balls is f = 1/3n.2619. we have the fractal dimension of the Koch curve .15) Another famous example of fractals. The fractal dimension is defined by the limit d =lim lnN(f). If you start with zo = 0 (say).8: Generation of the Koch curve. If lzl < 2. ( ~ )n with 4n segments in total. is the Mandelbrot set z = z 2 + c.3 Selfsimilarity and Fractals Recurrence Equations Figure 6.6. the stopping criterion is simply lzl 2:: 2.
Vast literature exists on this subject. and follow the same iterative process again.9: Fractal nature of the l\1andelbrot set: z = z 2 +c.9). 107 . After many iterations across a region of the complex plane. iy) if it is not part of the Mandelbrot set. and it is still an active research area. and the point cis marked on the complex plane (x. The iterations will go on until the modulus of z reach 2.3 Selfsimilarity and Fractals Figure 6. the results become the wellknown picture of the 1\!Iandelbrot set (shown in Figure 6. part of the l\1andelbrot set. we change a different value of c. Then. You can view this system as a dynamical system. This simple iterative system can produce beautiful patterns. It is an usual practice to mark the points with colours depending on the number of iterations for each point to reach the modulus lzl = 2. but it is a very complicated system.Recurrence Equations 6.
.
(7.1) where wfi = kjm.2) where wo is the natural frequency of the system. which can be written as y" + wfiy = f(t). it is time to see how they are applied to engineering problems.Chapter 7 Vibration and Harmonic Motion After the studies of the complex numbers and ordinary differential equations. and f(t) is the a known function oft. 7. This chapter concerns the vibration and harn1onic motion of mechanical systems. 109 .1 Undamped Forced Oscillations The sin1ple system with a spring attached with a mass m is a good example of harmonic motion (see Figure 7. we have y" + w~y = acoswt. In the case of f(t) = acoswt. If the spring stiffness constant is k. (7.1). and a is the amplitude of external forcing. then the governing equation of the oscillations is a secondorder ordinary differential equation for undamped forced harn1onic motion.
With these conditions. (7 .5) w0 w 2 Therefore. Its general solution is (7. we assume that Yp = asinwt + bcoswt.7) 110 . (7. we have A = 0 and B = aj(w~ .cosw0 t). (7.1 Undamped Forced Oscillations Vibration and Harmonic A1otion Figure 7. the general solution y( t) = A sin w0 t + B cos w0 t + • Q' 2 w0 w 2 cos wt. It needs some modifications. we have to consider two different cases w =I= wo and w = wo because for w = wo the standard particular a sin wt + b cos wt does not work. we have the initial conditions y(O) = 0 and y' (0) = 0. We now have y(t) = Q' w2 0 w2 (coswt.7. The general solution y( t) = Yc + Yp consists of a complementary function Yc and a particular integral Yp· The complementary function Yc satisfies the hon1ogeneous equation y" +w~y = 0. For w =I= wo.6) If we further assume that the system is initially at rest when the force starts to act. (7.w 2 ) in the general solution. We thus obtain Q' Yp = 2 coswt.4) For the particular integral Yp.3) Yc(t) = A sin wot + B cos wot.1: Undamped forced oscillations.
8) where 2a . (w. = .1 Undamped Forced Oscillations 5 (l)=:(l) 0 =2 5 10 time (t) 15 20 Figure 7. For the special case of w = wo..10) .wo) /2.wo )t ._~. while its amplitude or envelope slow oscillates with a frequent 8w = (w . (w + wo) t s1n 2 sm 2 2 0 w = = A(t) sin (w + wo)t 2 _ A( t ) 2 A(t) sinwt. the complementary function is the same as before. we can see that the oscillator oscillates with a major and fast frequency w = (w + wo)/2. 111 (7. but the particular solution should take the following form Yp = t(asinwt + bcoswt)..cos D we have y (t ) = . (7.2 sm (C+D) s1n (CD) 2 2 .. Using cos C . This phenomenon is called beats.. . w0 w (7..9) As lw...wo)t 2 sm 2 .. w2 2a . (w .wol < lw + wol.2: Variation of an1plitude at w = wo.Vibration and Harmonic !vlotion 7.
the equation of forced oscillations becomes y"(t) + 2Xy'(t) + w5y(t) = 112 acoswt. As the amplitude A(t) increases with time as shown in Figure 7. As all the real systems have a certain degree of friction.14) . this phenomenon is called resonance. (7. 7.13) where A(t) = atj(2wo).12) Similarly.7.5.3. the initial solution y(O) = y' (0) = 0 in1plies that A = B = 0. thus damping should be included. (7. . (7. With damping.2 Damped Forced Oscillations Vibration and Harmonic A1otion which gives 2wo The general solution is therefore Yp(t) =  Q t sm wot. An example of damping is shown in Figure 7. and the external forcing cause the oscillations grow out of control when the forcing is acted at the natural frequency wo of the system. We now have y(t) = 2wo ° tsinwot = A(t)sinwot.11) y(t) = Asinw0 t + Bcosw0 t + ~ tsinw0 t.2 Damped Forced Oscillations k y~ m I Ty(t) Figure 7.3: Damped harmonic motion. 0 (7.
we can multiply (7. In fact.16) where z(t) = y(t) + i17(t). and we have z"(t) + 2Az + w5z = aezwt. (7. For the transient part. In both cases ~(fl. The quality factor Q = 2~ is also commonly used. but it may become a little awkward as it involves complex numbers.= wo({ ± FI). then Jl < 0.19) For ~ = 0.15) Since eiwt = coswt + isinwt.) < 0. The complementary function corresponds to the transient part while the particular function corresponds to the steady state.Vibration and Harmonic 1\lfotion 7. For~< 1. For { = 1. In principle. which corresponds to the harmonic oscillations without damping.= A+ iJwfi.A2 and ~(1l) < 0.wfi. then fl. The amplitude decreases exponentially at just the slowest possible manner without any oscillations.14).2 Damped Forced Oscillations where A is the damping coefficient. we write the companion equation for equation (7. We now have fl. If A2 < w5. we have fl. (7.17) or Jl = A ± J A2 . we get Jl = w0 { +iwo~113 . there is quite an elegant method using the complex variables.'(t) + w51J(t) = asinwt.t)t + 0.15) by i. and add it to (7. the characteristic equation gives (7. one can try to solve this equation using the standard method. it is conventional to define a case of critical damping when { = Ajwo = 1.14) and (7. (7. it is critical damping as the imaginary term is zero. In order to do this. In engineering. we essentially solve both equations (7. Jl = wo. = iwo.14) with a different forcing term rl'(t) + 2Ar. (7.15) at the same time if we can separate the real and imaginary parts. By solving this equation.18) If A2 ~ w5. thus the solution Zc oc e~(J.
Figure 7. the transient part Yc will becomes negligible as t increases. As the an1plitude decreases much faster than that at the critical damping. . (7.20) and P(iw) = (iw) 2 + 2X(iw) + w5 = (w5. The general solution becomes Q Q z(t) = P(iw) ezwt = [(wfi. and overdamping (~ = 1. we only need to find the particular solution Zp· If we try the particular solution in the form z = zoeiwt. t (7. (7.21) which is essentially the characteristic polynomial.5). Therefore. The imaginary part is zero (no oscillation). If time is long enough (t >> 1). ~ > 1 leads to 1L = wo( ~ ± ~) < 0.1).2 Damped Forced Oscillations Vibration and Harmonic A1otion 1r~~~~~. 10 15 time (t) 20 25 Figure 7. For this reason.22) .4: Critical damping(~= 1). this case is thus called overdamped. underdamping(~= 0. it is called underdamped. The real part corresponds to the exponential decrease of the amplitude and the imaginary part corresponds to oscillations.Xw] ezw . we have z" + 2Xz' + wfi = P( iw )z.w2) + 2Xwi.4 shows the characteristics of these three cases.w2) 114 + 2i. Finally.7.
2 Damped Forced Oscillations  2.X.5.5 1 4 5 Figure 7.= a A J(w5 . where A= lzl is the modulus and¢ is the phase shift.w2)2 + 4.14) is the real part. the solution of the original equation (7. 1 .5 ro0:2 0. We can always write the general solution z = Aei(wt+<l>).O 0. It is conventional to call H(iw) = 1/ P(iw) the transfer function.25) 2 w0 w 2 As the amplitude of the forcing is a. Therefore.27) 115 . (7. the gain G(w) of the oscillation is _1 G(w) = . we have z = Aei(wt+<l>)' (7.1.24) and ¢=tan 2Xw .26) which is shown in Figure 7. (7.:O.5 0.75 ·. That is y(t) = (7.5: Variations of amplitude with frequency w and damping coefficient . Finally.Vibration and Harmonic !vlotion 3 7.1.23) where (7.0  s c.5 2  A.25 0.X2w2 Acos(wt + ¢).
6. we know wo is the natural frequency of the concerned system.7. 7. In general.2A 2 . Let u 1. wo coswt.3 Normal Modes In the above harmonic oscillations. we have a A+ 2Awo' and (7.29) The system is in the same phase with the forcing. the driving force is at very low frequency. For w << wo. u2. (7.31) y(t) = : 2 sinwot. m 3 . w. 116 . m 2 . 0 (7. Now let us study a more complicated system with three mass blocks attached in by two springs as shown in Figure 7. If w ~ wo. which is relatively simple.28) That is a y(t) ~. the forcing is at very high frequency. We have </>+ 7r. Then.30) The oscillator is completely out of phase with the forcing. the amplitude of the oscillations increases dramatically.1 . (7. If w >> wo. there may be many natural frequencies or modes in a system. (7.32) At resonance frequency = w~. their accelerations will be ii. we have </>+ 0. and the natural frequencies are in fact determined from the eigenvalue problem resulting from the system. respectively.3 Normal Modes Vibration and Harmonic A1otion Some special cases where w+ 0 and w+ oo are very interesting. u3 be the displacement of the three mass blocks m 1 . This system can be thought of as a car attached to two caravans on a flat road.
2 m2 0 ) ( Ut ) 0 m3 u3 ( kt . = d 2 uj dt 2 • From the balance of forces and Newton's law.kt(U2.Vibration and Harmonic 1\lfotion 7. m2u2 = k2(u3. 2.1 kt kt + k2 k2 ~:2)(::)=(~} Mii+Ku = 0.36) (7.Ut).3 Normal !v!odes u2.34) (7.33) (7.37) is a secondorder ordinary differential equation in terms of matrices.38) and the stiffness matrix is Equation (7. These equations can be written in a matrix form as (I + or 0 0 ii. Now we have w[MUi cos(wt) + KUi cos(wt) 117 = 0. The mass matrix is (7.u2). we have mtii.40) . (7. (7. u3).u2).35) ffi3U3 = k2(u3. 3) are constants and w2 can have several values which correspond to the natural frequencies.37) where uT = (ut. This homogeneous equation can be solved by substituting Ui = Ui cos(wt) where Ui(i = 1. ih where ii. 0 (7.t =kt(u2ul). u2.
7. 1). the three natural frequencies are w~ = 0.41) This is essentially an eigenvalue problem because any nontrivial solutions for ui require (7. Therefore. For w~ = 0. Dividing both sides by cos(wt). 2.1: For the simplest case when m 1 = m 2 = m 3 = m and k 1 = k 2 = k. or w 2 (k. where i = 1.42) Therefore. This is a.6: Harmonic vibrations. cubic equation in terms of w2 . we have (U1 .· 2 3k k k k 118 ..w 2 m 2 ) = 0. which is the rigid body motion.3 Normal Modes Vibration and Harmonic A1otion ~1 ~3 Figure 7. and it has three solutions. 3. U2 : U3 ) = 1. w3 = :. we have =0. the eigenvector is determined by *(1. For w 2 = k/m.w2 m)(3km. the eigenvalues of this equation give the natural frequencies. we have (7. 0 Example 7.
The general coordinates can be distance and angles. the 119 .4 Small Amplitude Oscillations For a mechanically conservative system.0. U3 ) = ~(1. That is to say.Vibration and Harmonic 1\lfotion 7.43) J where the stiffness matrix is K l. the velocities of the system will be q = 41... U3 ) = "*(1. qn).. (7. we can always take q = 0 at the equilibrium so that we can expand V in terms of q as a Taylor series V(q) = Vmin +I: ~~ q.. we can thus take the potential energy at equilibrium V min to be zero..qi +. tinIf we consider the system consists of many small particles or even imaginary parts.. Thus.qj=O 2 8qi8qj .. 1). we have (U1 . otherwise. it becomes (U 1 . and the force Fi = g~ shall be zero. 2. 0 7. This is correct because the resultant force must be zero at equilibrium. +I: I. where Tis its total kinetic energy and V is its total potential energy. This means that block 1 moves in the opposite direction away from block 3. ti2.iq. U2. K. the equilibrium or the minimum value of V requires ~ = 0 at the equilibrium point qi = 0. As we are only concerned with small amplitude oscillations near equilibrium Vmin = V(O) = Vo. block 2 moves in the different direction from block 3 which is at the same pace with block 1. and block 2 remains stationary. . its total energy E = T + V is conserved.. The configuration of the mechanical system can be described by its general coordinates q = (q1. \Vritten in normalized form. then the total kinetic energy T is a function of q and sometimes q. but the potential energy V is mainly a function of q only. . In addition.4 Small Amplitude Oscillations which leads to U2 = 0. and U1 = U3 . . U2 . 1).44) Since potential energy is always relative to an arbitrary reference point. l l (7. For wi = 3k/m. q2. . _ ~ 8 Vo J2 I qi=O.
and thus the corresponding kinetic energy is ~mql. the total potential energy is now simplified as v = L L qiKijqj = qTKq. dt 0 .. .45) which is a quadratic form. and thus timeindependent.. . Un coswt) and substituting it into the above equation. let us solve the same system of three mass blocks discussed earlier as shown in Figure 7.. So we have d(T dt + V) = !!:_ [q T M q + q TK q ] = .6. . .7. (7. L. the total mechanical energy E = T +Vis conserved. The total kinetic energy is the sum of all the components over all particles or parts.47) Since M and K are symmetric matrices. (7. . T = """' """' ffiijqiqj = qTM q. For a conservative system. . Assuming the solution in the form qT = (q1 . we have (7.46) j where M = [mij] is the mass matrix. forming a quadratic form. j (7. That is to say. . this above equation becomes Mq+Kq=O. the velocity is linear in terms of Qi.. For any small oscillation. As an application. The total potential energy T is the sum of each mass block 120 . .48) This is a second order ordinary differential equation for matrices.49) which is an eigenvalue problem. The component of the resultant force along the general coordinate qi should also be zero. . Therefore.Li (7. q2 . qn) = (Ut coswt.4 Small Amplitude Oscillations Vibration and Harmonic A1otion systen1 will be driven away by the resultant force. U2 coswt.
the force is f = kx.55) For example.52) We see that M is a symmetric matrix.56) and 121 .53) Therefore. which implies that Kij should be symmetric as well.4 Small Amplitude Oscillations 0 (7. thus the potential energy stored in a spring is (7.Vibration and Harmonic 1\lfotion 7. (7. For a spring. u2 etc. The stiffness matrix K = [Kij] can be calculated using (7.50) which can be written as a quadratic form T ·TM· =U U. (7. the total potential energy of the twospring system is (7.51) 0 (7.54) Since interchange of u 1 and u 2 does not change V. it is thus symmetric in terms of u1.
7.4 Small Amplitude Oscillations Vibration and Harmonic A1otion
= 2{)ul [kt (u2 ut) + k2(u3 u2)] =  2 .
Following the similar calculations, we have
1 {)
kt
(7.57)
(7.58)
which is exactly 1/2 multiplying the stiffness matrix we obtained earlier in equation (7.39). Thus, the equation for small amplitude oscillation is
Mii+Ku= 0.
For the special case of mt = m2 = m3 its eigenvalues are determined by
(7.59)
= m and
kt
= k2 =
k,
=0,
(7.60)
which is exactly the problem we solved in the previous section (see example 7.1). For a simple system such as a pendulum, equation (7.39) is equivalent to the following simple formula for calculating the natural frequency V"(q) (7.61) w=
M(q)'
where V" > 0 because the potential energy at equilibrium is minimum.
A simple pendulum with a mass m is hanged vertically from a ceiling with a distance L from the fixed point. Let 8 be the small angle from its equilibrium, then the kinetic energy is 2 2 T= = (0) 2 • The potential energy is
0 Example 7.2:
!mv
!m£
V
= mgL(1 cosO).
122
Vibration and Harmonic 1\lfotion 7.4 Small Amplitude Oscillations Therefore, the stiffness is K = !V"(8) = !mgLcos8le=O = mgL/2. The equivalent mass is 1\tf ( 8) = !mL 2 • The governing equation becomes 1 2 .. 1 mL 8 + mgL8,
2
2
or
8+ 8 = 0.
g
..
L
The natural frequency for small oscillations is
The period of this pendulum is
T
=
2 1r = 21r w
!I.. y;
0
123
Chapter 8
Integral Transforms
The mathematical transform is a method of changing one kind of functions and equations into another kind, often simpler or solvable one. In general, the transform is essentially a mathematical operator that produces a new function F( s) by integrating the product of an existing function f(t) and a kernel function K(t, s) between suitable limits
F(s) =
J
K(t, s)f(t)dt.
(8.1)
In the Laplace transform, the kernel is simply exp( st) and integration limits are from 0 to oo. In the Fourier transform, the kernel is exp(±ist) with a normalized factor 1/../2ff. The Fourier transform maps the time domain of a timedependent series such as a signal into a frequency domain, which is common practice in signal processing. The Laplace transform is a very powerful tool in solving differential equations. In this chapter, we will focus on the three major transforms: Fourier, Laplace and Wavelet transforms. They are commonly encountered in engineering and computational sciences.
125
8.1 Fourier Transform
Integral Transforms
8.1
8.1.1
Fourier Transform
Fourier Series
For a function f (t) on an interval t E [T, T], the Fourier series is defined as
where
and
bn = 
11T 11T = 11T f(T) . T T
ao = T an
T
T T
T
f(t)dt,
n1rt
(8.3)
f(t) cos( T
)dt,
(8.4) (8.5)
n1rt s1n()dt, (n = 1, 2, ... ).
Here an and bn are the Fourier coefficients of f (t) on [ T, T]. The function f (t) can be continuous or piecewise continuous with a jump discontinuity. For a jump discontinuity at t =to, if !'(to) and !'(to+) both exist, but !(to) f f(to+ ), then the Fourier series converges to [!(to)+ !(to+ )]/2. The Fourier series in general tends to converge slowly. From the coefficients an and bn, one can easily see that bn = 0 for an even function f( t) = f(t). Similarly, ao = an = 0 for an odd function f ( t) =  f (t). In both cases, only oneside [0, T] of the integration can be used due to the symmetry. Thus, for an even function f (t), we have the Fourier cosine series on [0, T]
(8.6)
For an odd function f (t), we have the Fourier sine series
f(t) =
L
00
sin( T
n1rt
).
(8.7)
n=l
126
Integral Transforms
8.1 Fourier Transform
f
t
4
2
0
2
4
Figure 8.1: Triangular wave with a period of 2.
0 Example 8.1: The triangular wave is defined by f(t) = ltl for t E [1~ 1] with a period of2 or f(t+2) = f(t). Using the coefficients of the Fourier series, we have
ao =
Since both
11 2
1
ltldt =
1
2[
110 (
1
1
t)dt +
11
0
tdtJ =
2·
1
~
It I and cos(n1rt) are even functions,
1 1
we have for any n
1,
an = {
ltl cos(n7rt)dt = 21 t cos(n7rt)dt }_1 0
1 
= 2tsin(n1rt)l
n1r
0
~} n1r
{
sin(n1rt)dt =
2
n 2 7r 2
[cos(n1r) 1].
0
Because ltl sin(n1rt) is an odd function, we have
1
bn = /
1
ltl sin(n1rt)dt = 0.
Hence, the Fourier series for the triangular wave can be written as f(t) =
2 + 2 L
1
~ cos(n1r) 1
n 2 1r2
n=O
cos(n1rt).
0
The nterm of the Fourier series, that is an cos(n1rtjT) + bn sin(n1rtjT), is called the nth harmonic. The energy of the n harmonic is defined by A~ = a~ + b~. The sequence of An forms the energy or power spectrum of the Fourier series. The energy spectrum of the triangular wave is shown in Figure 8.2.
127
2 Fourier Integral For the Fourier coefficients of a function defined on the whole real axis [oo.8.. 0. We have ao + 0 if (8.1.9) In this case.05 Figure 8.15 ! i. and n1f' Wn=T' (8.1 (8.2 Integral Transforms 0. 8.11) (8.1 0.1 Fourier Transform 0. oo].2: Energy spectrum of the triangular wave.12) . oo 128 (8.8) as the limits T + oo and Wn + 0. oo !00 f(t) sin(wt)dt. the Fourier series becomes the Fourier integral f(t) = where fooo [a(w) cos(wt) + b(w) sin(wt)dw. a(w) = .10) 1 b(w) = 1f' !00 f(t) cos(wt)dt. we can take the limits a(wn) = /_: f(t) cos(wnt)dt.
1"[/ (t)] of f (t) is defined as F(w) = .1 Fourier Transform Following the similar discussion for even and odd functions. we know that even functions lead to Fourier cosine integrals and odd functions lead to Fourier sine integrals.14) where exp[iwt] = cos(wt) + i sin(wt). V !:: Joe F(w)eiwtliv. = dwn ' dnF(w) . The transform can have different variations such as the Fourier sine transforn1 and the Fourier cosine transform. Hence. 8..16) (8.15) (8.rl [F(w)] = . and it has most of the properties of the differential operator and the integral operator.Integral Transforms 8.1"[/(t)] + .1. its Fourier transform is = (iwt F(w). it is straightforward to prove that it has the following properties: F[f(t) + g(t)] = .18) = o. 27r oo (8. V !:: Joo f(t)eiwtdt.. 27r co (8.1"[/(n)(t)] if f(t + ±oo) = f'(t + + ±oo) + 0. ±oo) = .1"[/(t)] = .2: For the triangle function f(t) and f (t) = 0 for It I > 1.17) (8. = f(nl)(t = 1ltl for (It I ~ 1) \1'21rF(w) = /_: (1.ltl)eiwtdt 129 .1. F[of(t)] (8.1"[( it)n f(t)] and . D Example 8.3 Fourier Transform The Fourier transform .F(g(t)].1"[f(t)]. The Fourier transform is a linear operator.13) and the inverse Fourier transform can be written as f(t) = . The Fourier transforms of some common functions are listed in the following table 8.
0 Table 8. The discrete Fourier transform (DFT).[8(w + wo). 0.1f w w2 Hence.::o w2 (a > 0) 1 f(at) a2+t2 cos(wot) sin(wo) sinxax J1[8(w.8(w.1: Fourier Transforms f(t) f(t. we have etW + iw 1 = 2{1.wo) 1/v"Fi v"Fi8(w) . 130 . (lwl > a) The most useful Fourier transform for engineering and computational science is probably the discrete form. especially in digital signal processing.wo) + 8(w + wo)] i ~' vt· @ealwl a ! F(~) .8. (w/ 2)2 = . tw 2o a2+w2 1 ""'7'f!r"e 7:'2" 4a v:. 2(W /2) = 1 . 1 sin 2 (w/2) 1 . 2 F(w) = .smc (w/2).eiw _ v £.to) j(t)eiw0 t 8(t) 1 sign(t) eoltl e(at)2 F(w) = F[f(t)] F(w)etwto F(w.1 Fourier Ti·ansform Integral Transforms = f 0 {1 + t)eiwtdt + t r (1.1.t)eiwtdt.cosw) w 2 w2 .j2. 2 we have rn=F( ) = iw+ 1.cosw .j2. lo 1 Sill Integrating by parts and using .wo)] (a > 0) (lwl < a).
thus for N values of k.1 (F(8)] is f(t) or f(t) = .2 Laplace Transforms The Laplace transform .1 additions for each X[k]. The inverse Laplace transform . 2 (8.20) N k=O L A periodic signal x( n + N) = x( n) has a periodic spectrum X[k + N] = X[k]. In fact. 2 8 . we have . many of the DFT calculations are not necessary. (8. by rearranging the formula. also called the signal reconstruction. However.Integral Transforms 8.19) n=O and the inverse transform.C(t] = 1 t lo co testdt = loco estdt + [estl 00 o o 8 8 o 1 = . This type of algorithms is called Fast Fourier Transform (FFT).2 Laplace Transforms for a periodic discrete function or signal x( n) with a period N. is defined by 1 N1 x[n] = X[k]ei~. The discrete Fourier transform consists of N multiplications and N. if N =2m (mE JV). filter design and signal reconstruction. The Laplace transforms of most simple functions can be obtained by direct integration.C(eo:t] = {oo eo:testdt = [1e<so:)tl oo = _I__ lo 8 a o 8. For simple functions t and eo:t. 8. is defined by N1 X[k] = L x[n]ei ~.c. the computational complexity is of O(N 2 ).21) where 8 > 0.c. one can get the complexity of O(N log2 N).a 131 .1 [F(8)].".C [f (t)] of a function f (t) is defined as F(8) =. (8.C[f(t)] =loco f(t)estdt. Vast literature exists on the signal processing such as FFT.
1 [so] = eat. s The Laplace transform pairs of common functions are listed below in table 8.C[f(t)] + /3. It often involves the partial fractions of polynomials and usage of different rules of Laplace transforn1s.22) However.9: In order to obtain the Laplace transform of f(t) = coswt.a)] .26) (8. we have 1 .8.C[d'(t)] = lo rX) d'(t)estdt = estl t=O = 1.iw + s + iw] s s2 +w2 0 132 . 0 Example 8. 1 = ~[ [ e<s+iw)tdt + foe esiw)tdt) 2 Jo lo 00 1 = 1 1 = 2[s. (8.25) (8.a).C[e0 t f(t)] = = a. the inverse of a Laplace transfonn is usually more complicated.C[g(t)]. it is straightforward to prove that the Laplace transform has the following properties: . . = eos ..23) (8.C[f' (t)] = s.C[af(t) + fjg(t)] . we have its Laplace transform . From the basic definition.24) (8. we shall first write f(t) Then.f(O). For the Dirac ST ' <5'function.c.c[j].2 Laplace Transforms Integral Transforms 1 Conversely' .2.27) F(s.C[ ft f( r)dr] lo = .1 [ 1 ] = t .C[f(t)] .!.C[f(t)]. = coswt = '2(e .wt + ezw t ).C[f(t. s >a.c. . (8.
For two functions f and g..29) (8. their convolution f * g obeys (8.2 Laplace Transforms Table 8.Integral Transforms 8.28) and their Laplace transforms follow .a)g(a)do:.C[f(t) * g(t)] C 1 [F(s)G(s)] = = F(s)G(s)..2: Laplace Transform Function f(t) 1 8(t) tn.30) fo' f(t. (8..:.n > 0 cos( at) sin( at) eat t1/2 t1/2 Laplace Transform F( s) s 1 s s:l+a:l a s:l+a:l 1 sn+T n! sa '2'S3' ( s 1( 11" ) 1/2 tn f(t) cos( at+ !3) sinh(ot) cosh( at) erfc( 2 ~) _1_e:rr a2 1)ndnF(s) dsn s2+a2 ~ s cos(~)a sin(~) a s1a:l s2a2 leavs s . 133 . sinoVt 1~at (a> 0) a~ J3 (eat _ ePt) (0 =f= /3) Both Fourier and Laplace transforms follow the convolution theorem.
v~ (n = 1.3). As a result. a transient signal is decomposed into elementary components of wavelets 134 .velet Integral Transforms The Fourier transform has the similar properties f(t) * g(t) = /_: = f(t)g(t.33) The Fourier transform is very useful to analyse stationary signals where a stationary signal means that the frequencies of the signal do not change with time. ~w~t ~ 2. ~' v~ ~ cos(nt). 1 (8..3 Wavelet The Fourier transform is an ideal tool for studying the stationary time signal whose properties are statistically invariant over time. F(w)G(w). the wavelet transform is very powerful in dealing with nonstationary signals because the Fourier transforn1 is not suitable for such signals. in the Fourier transform there is a tradeoff between frequency resolution and time resolution.34) which is similar to the Heisenberg uncertainty principle for spatial and velocity intervals.3 lVa. The wavelet transform is an alternative approach to the Fourier transform to overcome the resolution problen1 using the Ivlother wavelet 1/J or prototype for generating the other windows functions. In the Fourier transform.u)du. and all the used windows are in the form of either dilated/compressed or shifted. (8. (8.8. .. 2. For nonstationary signals whose frequencies f = w/2~ vary with time (see Figure 8. In addition. the Fourier transform does not work well.32) F[j(t) * g(t)] 8.31) (8. the stationary signal is decomposed into linear combinations of sine and cosine waves v2~ . ~sin(nt). In the wavelet transform. ).
Wavelets are defined as a realvalued function 'ljJ(t) (t E 'R) in terms of the generator wavelet or n1other wavelet.36) 135 . There are three major type of wavelets: Grossmannl\1orlet wavelets.Integral Transforms 8.5 Hz and /2 =2Hz) and a nonstationary signal whose frequency varies with time. or wavelet packets. a. Q Q 1 t(3 Q > 0. To generate other wavelets. (8. (3. GrossmannIvlorlet wavelets are of the form 'l/1(). t) is used by translating in time and change of scales.3 Wavelet s =0 u.3: A stationary signal with two fixed frequencies (ft = 0. as 1 1 0 10 2 f:ro/21t varies with time (t) 4 6 8 10 Figure 8.b E 'R. oo and oscillating in a wavery manner. The Daubechies wavelets have the form m. Daubechies wavelets and GaborIvlalvar wavelets. The function 'l/J is both well localized. decreasing rapidly as t .35) where 'l/J a generator wavelet.nEZ.. 'l/J( a.. (8. as >..
Wavelet analysis has vast literature and it is still an active research area in signal processing.3 lVa. s) = 1 JJST J tT f(t) · 1/J()dt.8.37) The continuous wavelet transform can be defined by \II J(T.m)).velet Integral Transforms The GaborlVIalvar wavelets are in the form w(t. 136 .n EN. (8.38) where T is the translation of the location of the window and s is the scale where s = 1 is for the most compressed wavelet. 1 mE Z. Readers can search the latest research journals to follow the latest developments. 8 (8.m) cos[1r(n + '2 )(t.
y). A partial differential equation (PDE) is a relationship containing one or more partial derivatives. ) = O. There is no universal solution technique for nonlinear equations.. Similar to the ordinary differential equation.... 8x' ay' 8x 2 ' ay2' 8xay' . even the numerical simulations are usually not straightforward.Chapter 9 Partial Differential Equations Partial differential equations are much more complicated compared with the ordinary differential equations. the highest nth partial derivative is referred to as the order n of the partial differential equation. where u is the dependent variable and x. . y. y) ay = f(x. .2) a(x. . y) 8x + b(x. The general form of a partial differential equation can be written as 8u 8u 8 2u 8 2u 8 2u (9 · 1) 1/J(x. 137 . are the independent variables. Thus.. which can be written as 8u 8u (9. y. we will mainly focus on the linear partial differential equations and the equations of special interests in engineering and computational sciences. A simple example of partial differential equations is the linear first order partial differential equation..
6) This is equivalent to the following equation in terms of parameter s (9. y)uy = f(x. compact subscript forms are often used in the literature.9. If the right hand side is zero or simply f(x.1 First Order PDE The first order partial differential equation of linear type can be written as a(x.4) In the rest of the chapters in this book. y)ux + b(x. y) = 0. 9. For simplicity in notations in the studies of partial differential equations. not u itself. then the equation is said to be homogeneous. They are Uyy ={)y2 ' {) 2 u Uxy =8x8y' {]2u (9.2) as aux +buy= f. y).1 First Order PDE Partial Differential Equations for two independent variables and one dependent variable u. y only.5) which can be solved using the method of characteristics (9. 138 . (9. b and f are functions of x.7) which essentially forms a system of firstorder ordinary differential equations. The equation is said to be linear if a. we will use these notations whenever no confusion occurs.3) and thus we can write (9. (9.
(9.2 Classification The simplest example of a first order linear partial differential equation is the first order hyperbolic equation Ut +cux = 0. t) = 1/J(xct) at timet.Partial Differential Equations 9. g. h.2 Classification A linear secondorder partial differential equation can be written in the generic form in terms of two independent variables x andy. it is parabolic. 9. One example is the wave equation Utt = c2Uxx· If ~ = 0. y.3 Classic PDEs Iviany physical processes in engineering are governed by three classic partial differential equations so they are widely used in a vast range of applications.O) = 1/J(x).ct).8) where cis a constant. the equation is elliptic.9) which is a travelling wave along xaxis with a constant speed c. k and f are functions of x and y only. it is hyperbolic. then u(x.4ac < 0. then we say that this equation is quasilinear.10) where a. Diffusion and heat conduction equations are of the parabolic type Ut = KUxx· 9. c. If ~ = b2 . It has a general solution of u = . b. auxx + buxy + CUyy + gux + huy + ku = /. u) is also a function of u. If the initial shape is u(x. (9. If~ > 0. One famous example is the Laplace equation Uxx + Uyy = 0. therefore the shape of the wave does not change though its position is constantly changing.P(x. 139 . If f (x. (9.
this becomes the Laplace equation.y)En. The heat conduction equation (9. Here k is the thermal diffusivity and f(x.13) Ut = kuxx. (9. is a famous example. are governed by parabolic equations. the steady state of heat conduction with a source is governed by the Poison equation (x. The solution or a function is said to be harmonic if it satisfies the Laplace equation. k is replaced by the diffusion coefficient D. For diffusion problem.11) + Uyy = q(x. y. We will give a brief description of some of these equations in later chapters. The 1D wave equation in its simplest form is Utt = C Uxx. t) is the heat source. Heat Conduction Equation Timedependent problems.12) for two independent variables x and y. If there is no heat source (q = 0). 140 .3 Classic PDEs Partial Differential Equations Laplace's and Poisson's Equation In heat transfer problems. y. 2 (9.14) where cis the speed of the wave. There are other equations that occur frequently in mathematical physics. engineering and computational sciences.9. or Uxx (9. such as diffusion and transient heat conduction. t). Wave Equation The vibrations of strings and travelling sound waves are governed by the hyperbolic wave equation.
However. hybrid methods. As a classic example. and these ordinary differential equations often appear as eigenvalue problems. and integral transform methods. respectively. the partial differential equation is usually transformed into two or three ordinary differential equations (ODEs). T(t) are functions of x. The final solution is then obtained by solving these ODEs.Chapter 10 Techniques for Solving PDEs Different types of equations usually require different solution techniques. As a result. similarity solutions. we now try to solve 141 . 10. y. series expansions. These methods include separation of variables. Y(y). In order to determine these functions.1) where X(x). there are some methods that work for most of the linearly partial differential equations with appropriate boundary conditions on a regular domain. (10. they have to satisfy the partial differential equation and the required boundary conditions.1 Separation of Variables The separation of variables attempts a solution of the forn1 u = X(x)Y(y)T(t). t.
we have sinAL= 0.2) with the initial value and boundary conditions u(O. (10. 7) From the boundary condition u(O. t) = 0. (10. called eigenvalues.3) y. and it only takes an infinite number of discrete values.1 Separation of Variables Techniques for Solving PDEs the 1D heat conduction equation in the domain x E [0. t) = 0. Therefore. we now get two ordinary differential equations X"(x) + A2 X(x) = 0.= X"(x) T'(t) kT. we have u(x. Letting u(x. 0) = 1/J. (10. t) = 0. both sides must be equal to the same constant. (10. Hence.8) which requires that AL = n1r.6) The solution for X(x) is in a generic form X(t) =a cos Ax+ . we have a = 0. Each eigenvalue A = An = n1r / L. The negative sign is just for convenience because we will see below that the finiteness of the solution T(t) requires that eigenvalues A2 > 0 or A are real. therefore. (n = 142 . t) = X(x)T(t). t) = u(L. and the constant can taken to be as A 2 . A cannot be continuous.10. The solution for T(t) is (10. From u(L.5) where A is the eigenvalue.BsinAx. L] and t>O (10. Please note that n f 0 because the solution is trivial if n = 0. (10.4) As the left hand side depends only on x and the right hand side only depends on t. T'(t) + kA 2T(t) = 0.
(10. (10.9) By superimposing Un = XnTn and expanding the initial condition into a Fourier series so as to determine the coefficients.13) 143 . Substituting into the solution for T(t). we now solve the 1D heat conduction in semiinfinite interval [0. the equation then becon1es (10. n1rd~ an = L Jo 1/. 2 00 n=l 2 [L . it transforms the partial differential equation into an ordinary differential equation. )has a corresponding eigenfunction Xn = sin(.)~.Techniques for Solving PDEs 10. we have u(x.!fethods 1.2 Transform Methods Laplace Transform The basic idea of the integral transform method is to reduce the number of the independent variables. (10. By solving the ordinary differential equation. we have (10.(x) s1n( y. t). As an example. t) = T 0 . t) = L:: an sin( n~x )e< nz > kt. 2. T(O. oo ). s) = f 000 T(x.11) with the boundary conditions T(x.10) 10.2 Transform }.Xnx). .. t)estdt be the Laplace transform of T(x. 0) = 0. For the 1D timedependent heat conduction. the solution to the original problem is obtained by inverting back from the Laplace transform..12) Let T(x.
Let u(w. __ . t)eiwxdx be the Fourier transform of u(x. 0) = 0. 0) = 1/. /U). The finiteness of the solution as x + oo requires that B = 0. lEx By the inversion of the Laplace transform. 2vkt where erfc( x) is the complementary error function. By using the inverse Fourier transform. we have T = Toerfc( .( ) du(w.a) 2 ]. !7C V 27r !00 1/. (10.(x) = exp[(x . The general solution to the ordinary differential equation can be written as IBx · 1'Fx T=Aev"f· +BeV"f·. t). The famous example is the classical wave equation Utt = V Uxx' 2 (10.(w) cos(vwt)ezwxdw . oo 144 .16) The general solution in tenns of the parameter w is u(w. 0) _ dt 0 .2 TI·ansform 1\tlethods Techniques for Solving PDEs and the boundary condition at x = 0 becomes f'x=O = To/ s.14) with the initial conditions u(x.15) dt2 = v w u. t) = f~oo u(x. t) = . X Fourier Transform Fourier transform works in the similar manner as the Laplace transform. Jk: with UtpW..T. t) = ijj(w) cos(vwt). and Ut(x.10. This transforms the PDE problem into an ODE d2u 2 2(10. we finally have 1 u(x. and the boundary conditions lead to To T = e s V"k'.
In this case.22) or f" = (f'.18) can be solved by using the similarity method by defining a similar variable {10.23) . one can assume a= 0 so that u = f(() (see the following example for details). the diffusion equation {10. we have {10. For example. the coefficients a and f3 can be detern1ined.19) or (=X #" (10. (10. 10.3 Similarity Solution = [1P(x + vt) + 'l/1(x. we can assume that the solution to the equation has the form 0 (10.Techniques for Solving PDEs 10.vt)] 1 2 _  2 1 [e (xa+vt)2 + e (xavt)2] .20) In general.17) which implies two travelling waves: one travels along the xaxis and the other along the negative xaxis direction.3 Similarity Solution The essence of similarity solution is to use the socalled sinlilarity variable ~ = xjtf3 so as to reduce the partial differential equation to an ordinary differential equation. For most applications.B(Kt) 2/3l. 145 (10. By substituting it into the diffusion equation.21) U = {Kt) j[ {K~)f3].
we obtain u =A 1 ~ et.12)..24) and (10. au = .w)] = T0 erfc( .erf( . we have used the chain rules of differentiations fx = ~ and = ~ so that *: £ & . ~ = Cerf( .27) Integrating it again and using the ( 2 = 4{2 . this means that all the exponents for any tterms must be zero. the boundary condition as x+ oo implies that C+D = 0. Therefore.{3kf'(()x = 8t (~t)f3t {3(J'(()(~t)1 (10. In deriving this equation. (10. Since 146 .26) !"If' and integrating the above equation once.10.. For the same problem as (10..3 Similarity Solution Techniques for Solving PDEs where !' = df I d(.. !"(() Using (ln !')' = we get = ~f'. while u(O.1: For the similarity solution of the diffusion equation Ut = KUxx' we assume u = (Kt)o: !(. v4~t ·2 X (10. we finally have u = T0 [1. .w)· X X (10.o where C and D are constants that can be determined from appropriate boundary conditions. Therefore.}) 13 • Now we want to know why a = 0.1'7'::7) + D. the diffusion equation becomes !. we have 2{3 = 1.28) t. or {3 = Now.25) Since the original equation does not have timedependent terms explicitly. t) =To means that D = C =To.29) 0 Example 10.. or (10.
The interesting feature of this equation is that it has a solitary wave solution or soliton. thus f3 = ~. Q' = 0.Techniques for Solving PDEs the diffusion equation becomes 10. 10.8 + 1 = 0. Now we seek the travelling wave solution in the form u = <f>(x. Q' .32) . 147 (10.vt). Nowadays.31) which is a thirdorder nonlinear partial differential equation.4 Travelling lVave Solution The requirement that no explicit (time t) appears in the equation leads to Q' = 0.30) or (10. then we do not impose any requirement on a. telecommunications use solitons to carry signals in optical fibres.This nonzero a indeed appears in some nonlinear diffusions equations where the diffusion coefficient is not a constant and K{ u) may depend on u itself.2.4 Travelling Wave Solution The travelling wave technique can be demonstrated using the famous Kortewegde Vries (KdV) equation (10. The soliton phenomenon was first observed by John Russell in 1834 when he travelled along the Union canal in Scotland. You may think that why not divide both sides of the above equation by (Kt )o:. Thus.
10. waves can travel in both directions.vt .38) 148 . That is to say.5 Green's Function Techniques for Solving PDEs where v is the speed of the travelling wave.10. We can see that the speed of the wave depends on the amplitude or height of the wave. 2 (10. we get '1/J d¢ d'lj. By substituting into the KdV equation.v¢ = 0.34) where A is an integration constant.<5) ]. ¢" + 0 at far field x + ±oo leads to A = 0. For linear wave equations. The requirement of c/>. but here it is only possible for the soliton to travel in one direction. big waves travel faster than smaller waves.37) where <5 is a constant and v /2 is the amplitude of the wave. T 2 (10.5 Green's Function The method of Green's function is very powerful in solving elliptic equations. (10.36) Integrating it again and substituting back to u.35) Integrating with respect to¢.33) = c/>c/>1 and integrating the above equation once.v¢ = 0. we have the travelling wave solution u = v sech 2[~ (x . cl>'. Let '1/J = cl>'. we get ~'l/J2 = ¢3 + ~v¢2. A Green's function inside the domain n is defined as (10. Using (¢ 2 /2)' we have (10. ¢" + 3¢2  vel> = A. + 3¢ . that is along xaxis direction in this scenario. (10. we have ¢"' + 6c/>c/>1 .
For the hyperbolic equations such as the wave equation. ~.~)8(y. t) can be concentration or any other quantity. 18 8u (10.41) where ll: is the diffusion coefficient.()8(t.43) First. the Crank's diffusion problem in an infinite cylinder is governed by the following equation: 8u {)t =. The fundamental Green's function for this case is G(x.1etbod where 8(x) is the Dirac delta function. we have (10. c r (10.c2 ~G = 8(x. In this case.17)8(z. but often a single method simply does not work.]. Readers can refer to the literature listed at the end of the book.(. r E (O. z. For example. G = G(x.z. y. (10.y. but it can be very neat in obtaining solutions. u(r. we can define the Green's function as Gtt .6 Hybrid Method Some differential equations can be solved using one of the methods described above.Techniques for Solving PDEs 10. a hybrid method that combines several methods is needed. (). 17.a).6 Hybrid J. (10.40) The Green's function method is very complicated. 10.~.r).T) =  (10. 8r [rll: 8r].1J. 0) = 1/J(r). we use the separation of variables.44) 149 . Generally speaking. t) = 0.42) u(r.39) 1 41rc2 r 8[(t. Now we want to solve this equation with the following boundary conditions: u(r. and r =a.t.r). and r is the distance in the polar coordinates. It usually requires that G = 0 on the boundary surface r.
t) = L DiJo(Air)e\rKt.46) where Di are undetermined coefficients.10.6 Hybrid Method Techniques for Solving PDEs then we have the equation for v( r) v" + v' + A2 v = 0.47) which means that paratneter Ai are the roots of the Bessel function J0 ( Aia) = 0.49) and (10.48) Using the basic properties of the Bessel functions (10. r 1 (10. The initial condition gives 1/J(r) = L DiJo(Air). t) has the following form u(r.45) where v' = dv / dr and A is a parameter to be determined. i=l 00 (10. This equation is essentially the Bessel equation of the zeroorder v = 0.50) loa rJo(Air)dr = . The boundary condition [u(r =a) = 0] requires that (10.i Jt(aAi). o (10. Thus. i=l 00 (10. the general solution can be written as (10.51) u(r.52) 150 . we can now assume that the general solution for u(r. t) = 2 L 2 ~ Jo(rAi) \~Kt i=l a J 2 ( A·) e • 1 a z Ina r'l/J(r)Jo(rAi)dr.
0) = ¢(r) .uo = 1/J(r). and power series. By substituting u = w + uo and 1/J(r) = ¢(r) . 0 Example 10. u(r.i)dr.uo and using the properties of Bessel functions. we have to make a transformation u = w + uo. we finally obtain where hi = foa r¢(r)Jo(r>. which is the problem we have just solved. but now the boundary conditions for w become w(r =a. asymptotic methods. These include series methods. approximate methods.1ethod The solution procedure shows that it requires a combination of separation of variables. For a very special case when </>(r) = 0.2: If we now want to solve the same diffusion equation in the cylindrical coordinates with slight different boundary conditions: u(r =a. t) = 0. perturbation n1ethods and naturally the numerical methods. t) = uo = const. Both u and w satisfy the same diffusion equation. we have There are other important methods for solving partial differential equations. 151 .6 Hybrid P.Techniques for Solving PDEs 10. Bessel functions.. w(r. 0) = ¢(r).
.
In general. 11. integral equations are a different type of equation and they frequently occur in applied mathematics and natural sciences. In this chapter.1. making the value of the integral a local maximum or minin1um. For example. On the other hand. we will briefly touch these topics. 11.Chapter 11 Integral Equations The calculus of variations is important in many optin1ization problems and computational sciences. especially the formulation of the finite element methods. in mechanics we may want to find the shape y( x) of a rope or chain when suspended under its own weight from two fixed points. namely the curvature of a curve. In this case. the calculus of variations provides a method for finding the function y(x) so that the curve y(x) minimizes the gravitational potential energy of the hanging rope system. let us first discuss an important concept.1 Curvature Before we proceed to the calculus of variations. a curve y(x) can be described in a parametric form 153 .1 Calculus of Variations The main aim of the calculus of variations is to find a function that makes the integral stationary.
11.1 Calculus of Variations
Integral Equations
in terms of a vector r( s) with a parameter s which is the arc length along the curve measured from a fixed point. The curvature K of a curve is defined as the rate at which the unit tangent t changes with respect to s. The change of arc length is
(11.1)
We have the curvature

dt
ds
=Kn= n.
1 p ·
(11.2)
where p is the radius of the curvature, and n is the principal normal. As the direction of the tangent is defined by the angle 8 made with the xaxis by t, we have tan() = y'. Hence, the curvature becomes
K==
d() ds
dO dx dx ds ·
(11.3)
Fron1 () = tan 1 y'(x), we have
dO [ _ 1 ( ')]' y" dx = tan y = (1 + yl2
r
(11.4)
Using the expression for dsjdx, the curvature can be written in terms of y(x), and we get
(11.5)
11.1.2
EulerLagrange Equation
Since the calculus of variations is always related to some minimization or maximization, we can in general assume that the integrand 1/J of the integral is a function of the shape or curve y( x) (shown in Figure 11.1), its derivative y' (x) and the spatial
154
Integral Equations
11.1 Calculus of Variations
coordinate x (or time t, depending on the context). For the integral
I
=
l
'1/;(x, y, y')dx,
(11.6)
where a and bare fixed, the aim is to find the solution of the curve y( x) such that it makes the value of I stationary. In this sense, I[y(x)] is a function of the function y(x), and thus it is referred to as the functional.
y
a
'__,. X
Figure 11.1: Variations in the path y(x). Here, stationary n1eans that the small change of the first order in y( x) will only lead to the secondorder changes in values of I[y(x)], and subsequently, the change 8I of I should be virtually zero due to the small variation in the function y(x). 'Iranslating this into the 1nathematical language, we suppose that y(x) has a small change of magnitude off so that
y(x)
+
y(x) + f((x),
(11.7)
where ( (x) is an arbitrary function. The requirement of I to be stationary means that
8I
or more accurately,
= 0,
(11.8)
di
df 'E=0 = 0,
for all ((x).
(11.9)
155
11.1 Calculus of Variations
Integral Equations
Thus I becomes
I(y, <) =
t
,P(x, y + <(, y1 + <(')dx
= [ 1/J(x,y,y')dx+ [[<((: +(':Jdx+0(<2 ). (11.10)
The first derivative of I should be zero, and we have
(11.11)
which is exactly what we mean that the change ~I (or the first order variation) in the value of I should be zero. Integrating this equation by parts, we have
(11.12)
If we require that y(a) and y(b) are known at the fixed points x = a and x = b, then these requirements naturally lead to ((a) =((b)= 0. This means that the above right hand side of the equation is zero. That is,
81/J] b [( 8y' a= O,
which gives
(11.13)
dx 1 [81/J ~ 81/J](dx
b
a
8y
8y'
=
0.
(11.14)
As this equation holds for all ( (x), the integrand must be zero. Therefore, we have the wellknown EulerLagrange equation
(11.15)
It is worth pointing out that this equation is very special in the sense that 1/J is known and the unknown is y(x). It has many applications in mathematics, natural sciences and engineering.
156
Integral Equations
11.1 Calculus of Variations
The simplest and classical example is to find the shortest path on a plane joining two points, say, (0,0) and (1, 1). We know that the total length along a curve y( x) is
L=
l
J!+y'2dx.
(11.16)
Since '1/J = J1 + y'2 does not contain y, thus~= 0. Fron1 the EulerLagrange equation, we have
(11.17)
its integration is
8'1/J
ay'
Rearranging it as
'2
=
v1 +
or
y'
yn = A.
(11.18)
y
= 1
A2 A2 '
I
A
y = v'1 A 2 '
(11.19)
and integrating again, we have
y
= kx +c,
k
 v'1 A 2 •
A
(11.20)
This is a straight line. That is exactly what we expect from the plane geometry.
D Example 11.1: The Euler Lagrange equation is very general and includes many physical laws if the appropriate form of 1/J is used. For a point mass m following under the Earth's gravity g, the action (see below) is defined as
1/J =
~mv 2  mgy = ~m(y) 2 
mgy,
where y(t) is the path, and now x is replaced by t. v = velocity. The EulerLagrange equation becomes
y is
the
81/J d 81/J 8y = dt ( 8v ),
157
11.1 Calculus of Variations or
Integral Equations
d mg = dt (mv),
which is essentially the Newton's second law F = ma because the right hand side is the rate of change of the momentum mv, and the left hand side is the force.
Well, you may say, this is trivial and there is nothing new about it. This example is indeed too simple. Let us now study a more complicated case so as to demonstrate the wide applications of the EulerLagrange equation. In mechanics, there is a Hamilton's principle which states that the configuration of a mechanical system is such that the action integral I of the Lagrangian .C = T  V is stationary with respect to the variations in the path. That is to say that the configuration can be uniquely defined by its coordinates qi and timet~ when moving from one configuration at time t 0 to another time t = t*
I
=
fo'• .C(t, q;, tj;)dt,
i =
1,2~
... ,N,
(11.21)
where T is the total kinetic energy (usually, a function of Qi), and V is the potential energy (usually, a function of q). Here Qi means . 8qi (11.22) qi =
m·
In analytical mechanics and engineering, the Lagrangian .C (=Kinetic energy  Potential energy) is often called the action, thus this principle is also called the principle of least action. The physical configuration or the path of movement follows such a path that makes the action integral stationary. In the special case, x + t, the EulerLagrange equation becomes (11.23) which is the wellknown Lagrange's equation. This seems too abstract. Now let us look at a classic example.
158
Integral Equations
11.1 Calculus of Variations
Figure 11.2: A simple pendulum.
0 Example 11.2: For a simple pendulum shown in Figure 11.2, we now try to derive its equation of oscillations. We know the kinetic energy T and the potential energy V are
V = mgh = mgl(l cos 8).
Using C, = T  V, q = 8 and q =
iJ, we have
which becomes
mgl sin 0  dt (ml 2 0) = 0.
d
.
Therefore, we have the pendulum equation
This is a nonlinear equation. If the angle is very small (8 << 1), sin 8 ~ 8, we then have the standard equation for the linear harmonic motion
0
159
0).. We have I= subjected to the constraint Q= t t 1/J(x. 160 (11.24) cf>(x. there is no additional constraint. y')dx.Q or J = t[.¢]dx. sometimes these variations have certain additional constraints. 0) and (d.. In this sense.<P] in the EulerLagrange equation or following the same procedure of the derivations. (11.y. we now define its gravitational potential energy EP as Ep = j_:[pgy(x)ds] = pg j_:yJl + y12dx.27) Now we can come back to our example of the hanging rope problem with two fixed points. (11.28) . y. The total length of the rope is L. the sliding of a bead on a hanging string. we have (11. it requires that 2d < L. the calculus of variation discussed up to now is unconstrained.P + >. the method of Lagrange multipliers can transform the constrained problem into an unconstrained one by using a combined functional J = I + >. In order to find the shape of the hanging rope under gravity. Now we want to make the integral I stationary under another constraint integral Q that is constant.1 Calculus of Variations Integral Equations 11. However. (11.1. and it hangs frmn two fixed points ( d.11. is the undetermined Lagrange multiplier.3 Variations with Constraints Although the stationary requirement in the calculus of variations leads to the minimization of the integral itself.y')dx.26) where ). Replacing 1/J by [1/J + >..25) As for most optimization problems under additional constraints. From the geometric consideration. for example.
then the EulerLagrange equation can be reduced into a simpler form in this special case. we have (11.oy' dx = 0 + y oy + y oy'' .30) Jd [pgy +X] VI+ y'2 dx. Q= Jd d J1 +y12 dx = L.Integral Equations 11.ow (11. Using vi . (11.29) By using the Lagrange multiplier A.35) 161 .32) which can again be written as d ..34) Substituting the expression of \II into the above equation.8\11 8 y' = A = const.8\11 dx[w. we have J = Ep J = + .\Q.y o. (11. or ~! = 0.33) . d Since \II = [pgy + X] + y'2 does not contain X explicitly.Y 8y'] = The integration of this equation gives \II .31) d\11 8\11 ow dy ow dy' and the EulerLagrange equation~= fx(~ ). the stationary values of J requires (11.1 Calculus of Variations The additional constraint is that the total length of the rope is a constant (L). or (11. Thus. (11.=8x +8y dx + dx .ow .
We assume that free end is at (0. we then have 8\J! oy' From J = Ep = o.37) Using the boundary conditions at x = ±d and the constraint Q = L. 0). y) where y acts like a free parameter to be determined. what happens if we only fix one end at (a. (11.39) 0 Example 11.11.1 Calculus of Variations Integral Equations Multiplying both sides by vf1 A cosh(= pgy +A.) J 1 + y'2 • Thus. Since the variation of 81 = 0.()](dx 8\J! 8y d 8\J! dx 8y' + [(]b 8\J! oy' a = 0. As the variation ( is now nonzero at the free end point. 162 . we have \J! = (pgy + >.Q. while allowing the free end of the hanging rope to slide on a vertical pole? lVell. (11.cosh(y)]. we have 8J = l b a [ .9: For the hanging rope problem..36) whose solution is cosh t[PYY A +A]= xpg + K A . we get 8y' [(pgy + >...38) Finally. + >.)vfl a + y'2 ] = 0. (11. we have + y''l and using the substitution (11. this forms a variation problem with variable endpoint(s).. Now the boundazy condition at the free end is different. h(pgd) _ pgL SlllA2A. we have K = 0 and in1plicit equation for A . the curve for the hanging rope becomes the following catenary A pgx pgd y(x) = pg[cosh(A).
0 In other words. Legend says that Dido was promised a piece of land on the condition that it was enclosed by an oxl1ide.L stationary. She had to cover as much as land as possible using the given oxhide. we have or 163 . where X a and xb are two end points (of course they can be the same points). Thus. \rVe also ha. Suppose the total length of the oxhide strip is L.. the whole curve of the hanging rope with one free end forms half the catenary.)/ 11. Using the EulerLagrange equation.ve the additional constraint 1 Jl + Xa Xb y 12dx = L = const. Such a boundary condition of y' = 0 has the real physical meaning because any nonzero gradient at the free end would have a nonzero vertical component. She cut the oxhide into narrow strips with ends joined. the slope is zero at the free end.1 Calculus of Variations Jt + y' 2 = 0. or y' = 0. The enclosed area A to be maximized is A= 1 Xa xb y(x)dx. 0 Example 11. This forms an isoperimetric variation problem. and a whole region of a hill was enclosed. thus causing the vertical slip along the rope due to the tension in the rope.Integral Equations or y'(pgy + >..jl +ifl)dx.4: Dido's problem concerns the strategy to enclose a maximum area with a fixed length circumference. thus the maximization of A is equivalent to make I = A + >. As L is fixed. The zerogradient leads to the static equilibrium. That is I= A+ AL = { ' [y + A.
the most area that can be enclosed by a fixed length is a circle. In 1696. we can determine the speed of the bead from the equation ~mv 2 + mgy = mgh. Suppose we want to design a sn1ooth (frictionless) slide. From the conservation of energy. where B is another integration constant. also called the shortest time problem or steepest descent problem.K. which initiated the development of the calculus of variations. B) and a radius >.40) . 164 (11. By rearranging..>.11. we have y' = J>. >. ... Therefore..y' Jt + y'2 ± where K is the integration constant. h) to a lower point (a.1 Calculus of Variations Integral Equations which becomes 1 . and we have v = J2g(h. 0).y)._ ( dx Integrating it once. It was believed that Newton solved it within a few hours after receiving it.(x + K)2 X+ K Integrating this equation again. what is the best curve/shape the slide should take so that a child can slide down in a quickest way? This problem is related to the brachistochrone problem.. Johann Bernoulli posed a problem to find the curve that n1inimizes the time for a bead attached to a wire to slide from a point (0.2. we get Jt + y'2 y' ) = 0 · ====x+K.. 0 An interesting application is the design of the slides in playgrounds. 2  (x + K)2 + B. we get y(x) = =F)>. This is equivalent to which is essentially the standard equation for a circle with the centre at (.!!.
As the curve must pass the point (0. we have y '2 = Bh+y hy ' 1 B = 2gA2.1. we get k = 0. 0) is t = lo 0 ads = loa 1 v 0 )1 + y'2 dx. we have x= ~ [0  sin 0] + k.41) Using the simplified EulerLagrange equation (11.44) where 0 < 1r and k is an integration constant. The bead travels a longer distance. or at least it is rather counterintuitive. thus has a higher average velocity and subsequently falls quicker than traveling in a straight line. (11. 11. (11.42) By differentiation and some rearrangements.y)  y I 8 8y' [ (1 + y12) 2g(h.1 Calculus of Variations So the total time taken to travel from (0.sin 0). h). J2g(h ."2(1.Integral Equations 11.y = ~(1 cos 0) and integrating.34) because the integrand \ll = )1 + yrl / J2g(h.y) (11. (11. which seems a bit surprising.4 Variations for Multiple Variables What we have discussed so far mainly concerns the variations in 2D. and subsequently the variations are in terms y(x) or curves only.45) This is a cycloid. (11.cosO). What happens if we want to study a surface in the full 3D configuration? The principle in the previous sections can be 165 . h) to (a. we have (1 + y 12 ) 2g(h.y)  ] A . not a straight line. B y = h.43) By changing of variables TJ = h.y) does not contain x explicitly. So the parametric equations for the curve become x = ~ (0 .
we can approximately assume that Ux and uy are small.47) In this case. and in fact we know that all bubbles are spherical.49) or (1 + u~)Uxx. we get = wu = 0 (11.1 Calculus of Variations Integral Equations extended to any dimensions with multiple variables. It has been proved that the fundamental solution to this equation is a sphere.11.2uxUy + (1 + u. the principle of least action leads to the minimal surface problem. y (11. (11. thus the above equation becomes Laplace's equation Uxx + Uyy 166 = 0. then the total surface area is where W= {)u 1 + ()2 ax au + ()2 = {)y v+ 1 u2 X + u2. y). This nonlinear equation has been one of the active research topics for more than a century.)uyy = 0. we will focus on the minimization of a surface here. however. the extended EulerLagrangian equation for two variables x and y becomes (11. Now we assume that the shape of the bubble is u( x. The surface integral of a soap bubble should be stationary. For some problems.51) . (11.48) Substituting winto the above equation and using ~! since w does not contain u explicitly.50) This is a nonlinear equation and its solution is out of the scope of this book. Suppose we want to study the shape of a soap bubble.
we know that the unknown y( x) to be optimized is inside the integrand of I. (11.1 Linear Integral Equations Fredholm Integral Equations A linear integral equation for y( x) can be written in the following generic fonn u(x) +A l K(x. TJ)Y(TJ)dTJ = v(x)y(x). Thus. the action principle. we will mainly focus on the simplest types of integral equations. SturmLiouville problem.52) where K(x. this is an integral equation.2. many physical processes and laws of conservation are expressed in terms of integral forms rather than their differentiation counterparts. then the unknown y( x) appears only once in the integral equation. In fact. surface shape minimization. even the numerical simulations are usually not straightforward.2 Integral Equations The calculus of variations has many applications. and it is under the integral 167 . If the function v( x) = 0. In certain sense. Integral equations are much more complicated compared with the differential equations.2 Integral Equations From the calculus of variations. 11. rJ) is referred to as the kernel of the integral equation. There is no universal solution technique for nonlinear equations. The parameter A is a known constant. The other classical examples include Fermat's principle in optics. the equation is inhomogeneous. Naturally. 11. and of course the finite element analysis. one of the ways of constructing an integral equation is to integrate from a differential equation. If the function u( x) = 0. If u( x) =/= 0.Integral Equations 11. the equation is then called homogeneous.
A kernel is called a displacen1ent kernel if it can be written as a function of the difference (x .52) becomes the integral equation of the second kind u(x) +A l K(x. TJ) = L.2 Integral Equations Integral Equations sign only. The integral equation becomes singular and at least one of its integration limits approaches infinite.11. The kernel is said to be separable or degenerate if it can be written in the finite sum form K(x. is called a Fredholm equation. TJ)Y(TJ)dTJ = v(x)y(x). Volterra Integral Equation In general. TJ) = K(x.TJ) of its two arguments K(x. fi(x)gi(TJ). then the equation becomes a Volterra equation. This is called the linear integral equation of the first kind u(x) +At K(x. 168 (11. (11. respectively.54) An integral equation with the fixed integration limits a and b.53) On the other hand.56) where fi(x) and gi(TJ) are functions of x and TJ.TJ). 1/)Y('I)d'l = 0. (11. 1/)Y('I)d'l = y(x). (11. i=l N (11. equation (11. If the upper integration limit b is not fixed. the Volterra integral equation can be written as u(x) +A lx K(x.57) .55) The first kind [or v(x) = 0] and second kind [or v(x) = 1] are defined in the similar manner. if v(x) = 1.
these coefficients can be written as simple explicit expressions.60) we now have the solution in the form y(x) = u(x) +A Oi L aifi(x). we can substitute the kernel (11.59) Because the integration limits are fixed. i=l N (11.3 Solution of Integral Equations Iviost integral equations do not have closedform solutions.(x)g. tr b N N J.62) fori= 1. Only for a few special cases.(17)d17 = y(x).(1J)Y(1J)d1J. 11. By defining = t g. (11.3. For linear integral equations. /. 2.(17)d1J = y(x).56) into the equation and we have u(x) +A which becomes J.(x) <>.58) u(x) +A tr J.. (11. b g.Integral Equations 11. 169 . N.. the closedform solutions are only possible for the special cases of separable and displacement kernels. .3 Solution of Integral Equations 11.1 Separable Kernels For a Fredholm integral equation of the second kind with separable kernels.61) where the N coefficients are determined by (11. the integrals over 1J should be constants that are to be determined. (11. .
3 Solution of Integral Equations Integral Equations 11. we have (11.11.65) The solution y( x) can be obtained using the inverse Fourier transform y(x) = _1_ V(w) eiwxdw.3.3.69) .Av'2iK(w)] roo (11. we have U(w) + Av'2?fK(w)Y(w) = Y(w). we have u(x) +A/_: K(x 1J)Y(1J)d1J = Y(1J).66) 11.2 Displacement Kernels For a singular integral equation with a displacement kernel. (11.Av'2iK(w) . the equation can be solved by Fourier transforms if both integration limits of the integral are infinite. In the case of a simple degenerate kernel K(x. 170 (11.64) which is an algebraic equation for Y(w). (11.63) Using the Fourier transforms and the convolution theorem. (11.67) y(x) = u(x) +A which becomes fox f(x)g(1J)Y(1J)d1J.68) y(x) = u(x) + Aj(x) fox g(1J)Y(1J)d1J. In this case. 17) = f(x)g(1J).3 Volterra Equation A Volterra equation with separable kernels may be solved by transforming into a differential equation via direct differentiation. Its solution is simply Y(w) = U(w) 1 . (11. y'2i 1oo [1.
We get y(x) = f(x)[e.73) P(x) = .tion of Volterra type y(x) =ex+ lax ex sin(()y(()d(.Xf(x)g(x)\ll(x) = ¢'(x). we have (11. (11.J P(x)dx]{j [Q(x)ef P(x)dx]dx }.3 Solution of Integral Equations If f (x) =/= 0.Xf(x)g(x).71) w'(x).70) Putting ¢(x) = u(x)/ f(x) and differentiating it. This is equivalent to the standard form w' + P(x)\ll = and Q(x). (11.5: Let us tzy to solve the integral equa. the exact form of the solution can be obtained.72) which is a firstorder ordinary differential equation for W(x).74) We can use the standard technique by n1ultiplying the integrating factor exp[J P(x)dx] to obtain the solution..Integral Equations 11. (11. we get y(x) ex = 1 + rx sin(()y(()d(. 0 Example 11. By letting W = y(x)j f(x). (11. Jo 171 .75) With appropriate boundary conditions. Q(x) = ~~~:~]'. we divide both sides by ex. it can be written as y(x) f(x) = u(x) rx f(x) +A lo g(TJ)Y(TJ)dTJ. we have ~~~:~]' = 4>'(x) + Ag(x)y(x). First. (11.
we finally obtain y(x) + 1. However.y(x)ex = y(x) sin(x). ex = exp[x. 0 There are other n1ethods and techniques of solving integral equations such as the operator method.11. = y(x)sm(xL . In this case.sinx)]. 172 ..!.2 (cosx..y'(x). series method and the Fredholm theory. most integral equations do not have closedform solutions. we have [In y(x)]' =ex sin x By direct integration. ex Divide both sides by y(x) and using [Iny(x)]' = y'(x)fy(x). we have lny(x) = xThus. ~ex cosx +~ex sinx.3 Solution of Integral Equations whose differentiation with respect to x leads to Integral Equations [&] or y(x) ' . numerical methods are the best alternative.
in the threedimensional case. Any lowercase subscript that appears exactly twice in any term of an expression means that sum is over its all possible values of the subscript. Tensors are the extension of vectors.1 Notations In tensor analysis. the Einstein sun1mation convention or Einstein notations 1 and notations for subscripts are widely used. For example. we have GiXi =L i=l 3 aiXi = GtXt + a2x2 + G3X3. (12. though most commonly used tensors are second order tensors. 12.1) ~jBjk 1 =L j=l 3 AijBjk = AitBlk + Ai2B2k + Ai3B3k· (12.Chapter 12 Tensor Analysis Ivlany physical quantities such as stresses and strains are tensors. and they can have any number of dimensions and any orders. Vectors are essentially firstorder tensors.2) This notation convention was introduced by Albert Einstein in 1916 when formulating the theory of General Relativity. 173 .
2. (12. x2. For example.2 Tensors When changing the bases from the standard Cartesian e 1 = i. k) is an even permutation of (1. can alternatively be written as (i=1. 2.4) For a tensor with three subscripts sin1ilar to 8ij. with an angle of () around a fixed axis. the matrix for a simple rotation. k) is an odd permutation of (1.2 Tensors Tensor Analysis OUi = \7 . x~. 3). { 1 if (i. the LeviCivita symbol or tensor is defined as Eijk = +1 if (i. i= j).j. x 3 ) in the old bases is related to the new vector x' = ( x!. (12.n).2.. translation. e~. e~. 0 (otherwise).12. e2 = j. x~) in the new bases by a coefficient matrix Sij. is defined as 8 ij = 1 (if i { 0 (if i = j).6) Using the summation conventions. .8) (12. 3). the matrix equation Ax=b.5) 8ij The tensors and Eijk are related by 8iq8jp· fijkfkpq = 8ip8jq  (12. e3 = k to a new set of bases ei.9) . u = OUt 8xi OXt + OU2 + OU3 8x2 8x3 · (12.7) 12. Sij can be the rotation. Sij becomes Sij= ( ~:o ~: 174 D· (12.3) The Kronecker delta 8ij which is a unity tensor (like the unity matrix I in matrix analysis). (12. a position vector x = (x 1. . enlargement or any of their combinations.j..
3) are said to form a firstorder Cartesian tensor (or vector in this case). vectors have rank 1. the rank is associated with the number of indices to describe a tensor in terms of a multidimensional array. If components of a variable such as strains aij are transformed as (12. then there is one Sij extra in the product for transforming.Tensor Analysis 12. Therefore.11) then ui(i = 1. In this sense. any 175 . In engineering and computing. In a similar fashion.14) 12. Scalars have rank zero. and for each order increase.3 Tensor Analysis The orthogonality of Sij requires that ssT= sTs =I or (12.13) and (12. and secondorder tensors have rank 2.12) we say these components form a secondorder tensor. higherorder tensors can be defined. 2. a secondorder tensor is equivalent to a twodimensional array or a matrix.10) If the components ui of any variable u are transformed to the components u~ in the new bases in the same manner as (12.3 Tensor Analysis Tensors in Cartesian Coordinates One of the main advantages of tensors is that a tensor is independent of any chosen frame of reference. The order of a tensor is also called its rank. but no subscripts are allowed to appear more than twice (12.
In the similar way as n1ultidimensional arrays or matrices.]. r& spectively. the divergence theorem can be rewritten as the following form f 8 ui dV = J uinidS. the stress strain relation in linear elasticity is independent of frame of reference (for details see next chapter) (12. 8uq 8XjXp Similarly.Tji)[antisym.20) lv 8 xi !s The tensor forn1s are sometimes useful to the proof of the complex relationship among vectors and tensors.12.l and A are Lame constants.. For secondorder tensors. a tensor Tij is said to be symmetric if Tij = Tji.] + ~(Tij.3 Tensor Analysis Tensor Analysis physical laws or equations that are formulated in terms of tensors should be independent of frame of reference.16) All the formulas in vector analysis can be rewritten in the tensor forms using the summation convention and notations (12. 176 .19) \7 X (\7 X u)i = fijkfkpq.17) \1 2 1/J = {)21/J 8xi8Xi = Oij {)1/J 8xi8Xj' (12. For example.15) where aij and Cij are the stress tensor and strain tensor.18) (12. two tensors can be added or subtracted componentbycomponent if and only if they are the tensors of the same order. and antisymmetric if Tij = Tji· An interesting property of a tensor Tij is that it can always be written as a sum of a symmetric tensor and an antisymmetric tensor Tij = ~(Tij + Tji)[sym. J. They also become handy for the implementation of numerical algorithms.. (12. (12.
we know that a X (b X c) = Eijkaj(Ekpqbpcq) = EijkEkpqajbpCq = (8ip8jq.c(tSiPaibp)· By renaming the indices (j . Y' 2 ¢ = ¢.c(a ·b).23) (12.22) ' UXi where the index of the spatial component Xi is denoted by a comma to avoid any potential confusion with other indices. the important operators involving the V'operator can be written as grad¢ = ¢. i = V' ¢.1: Let us now use the tensor notations to prove ax (b x c) = b(a ·c) .v (12.8iqcq(8iPaibp)· Using 8ipbp = b and 8iqcq = c.X3 2 · UXl UX2 8 (12.21) Another way of denoting the derivatives of tensors a and v is to use following notations ai=~.24) 82 u1 82 u2 82 u3 = Vi. aa (12. we have a(b x c) = b(tSjqajcq) .ii = ~¢.26) 0 Example 12.c(a ·b).8iq8jp)ajbpcq = 8ipbp(8iqajcq). With these notations.j) so that 8iqajcq 8ijaici = a • c and tSiPaibP = a • b. we finally obtain ax (b x c)= b(a ·c). Fl·om above expressions for cross products. and V'.25) A very special case is that (12.Tensor Analysis 12. we have the identity oiiAi = Ai· The cross product Ax B can be expressed as (12.i = !i""""2 + '£l"2 + .i and q . 0 177 .3 Tensor Analysis Using the tensor notations.
where 8j acts in the similar way as 8ii· In many books on tensor analysis. 2.27) where vi(i = 1. (12.3 Tensor Analysis Tensor Analysis Tensors in NonCartesian Coordinates The tensors we have discussed so far are expressed in Cartesian coordinates. Tensor analysis is very important in theoretical physics and differential geometry where formal mathematical theory is required. (12. a tensor such as a vector v can be expressed as the sum of its components multiplying by the basis vectors (12. we can write a secondorder tensor in tern1s of covariant components aii and contravariant components aii as (12. q3 ).30) 178 . the tensor product (also called outer product) of two tensors is rather complicated. In fact. 3) are called the covariant components of v in the contravariant basis vectors ei. 2. the basic vectors are given by (i=1. 3) are called the contravariant components of v in the covariant basis vectors ei. many books on tensor analysis use a modern approach in terms of tensor duality.2. For example.28) where ei and ei are reciprocal systems of vectors. and ei · ei = 8j. covariance and contravariance concepts.3). In nonCartesian coordinates.12.29) U3V1 U3V2 U3V3 Using these basis vectors. they are more complicated. ei is also written as €i = ei. In the simplest term. the tensor product u Q9 v = uvT of two vectors u and vis given by U@V = ( ~~ )@(Vt U3 V2 V3) = ( ~~~~ ~~~~ ~~~: ) . q3) at any point P on a position vector r(q1 . Furthermore. Q2. while vi(i = 1. q2 . For the curvilinear coordinates (Ql .
the formal approach is preferred. Mathematically speaking. however. especially from the computational point of view. In the next chapter.ui = ui.Tensor Analysis 12. its determinant g = l9ij I = det(gij) is related to the Jacobian J = Vff· The derivatives in nonCartesian coordinates are far more complicated. For example.ui.31) This tensor is always symmetric 9ii = 9ji.] •1 become identical and r~j = 0 in Cartesian coordinates. the divergence of a vector is defined by the covariant differentiation u1i or V · u = u~i = ~u~ +r~.+r1 . That is why we have used an oversimplified approach here. the simple formulation in terms of multidimensional arrays in Cartesian coordinates is a more convenient approach. the fundamental metric tensor is defined by (12. we will study the theory of linear elasticity as an application of tensor analysis. and they usually involve the Christoffel coefficients rt = ek · ~.3 Tensor Analysis In a given frame of reference. 179 . and it is therefore not necessary to distinguish the contravariant and covariant vectors and components. It is worth pointing out that ei and ei vq1 l] . In engineering mathematics.
.
(13. The unit of stress is N jm 2 .1) where k the spring constant. this equation only works for 1D deformations. (13. that is F=kx.Chapter 13 Elasticity 13.1 Hooke's Law and Elasticity The basic Hooke's law of elasticity concerns an elastic body such as a spring.2) where ~Lis the extension. L' 181 . For a bar of uniform crosssection with a length L and a cross section area A. and it states that the extension x is proportional to the load F. the stressstrain relationship is u=Ec. However. The stress and strain are defined by u = A. it is n1ore convenient to use strain e and stress u. we have EA k= EA (13. though it is conventionally expressed in mjm or% (percentage) in engineering. For the elastic bar.4) F= L~L=kx.3) where E is the Young's modulus of elasticity. while the strain is dimensionless. F e = L' ~L (13. Written in terms F and x = ~L.
= c::.).z)].x. it is useful to (13. we need to generalize Hooke's law. This equation is still only valid for any unidirectional compression or extension. and x3 = z.x c::.9) where write Xt = x.8) Therefore.5) and strain tensor e = ( ::: ::: ::: ) = ( c::. The strain tensor is defined by the displacement u T = Cij =  1 8ui (2 8Xj 8uj + . x2 = y.10) 1 e = 2'(V'u + V'uT). 8 Xi (13.6) it can be proved later that these tensors are symmetric. (13. The generalized Hooke's law can be written as Cxx = E[axx. we only have 6 independent components or unknowns for stresses and 6 unknown strain cmnponents. O"xz = O":. For the 2D and 3D deformation.7) and cxy cxz cy::: (13.11) (13. For the general stress tensor (also called Cauchy stress tensor) (13. + O"zz)].12) cyy = E[ayy.y c::. which leads to O"xy = O"yx' = cyx. that is u = uT and e = eT.v(ayy 1 + O"::.13.1 Hooke's Law and Elasticity Elasticity where k is the equivalent spring constant for the bar. = O"::. Sometimes. :~~ :~~ :~: ) ' c31 c32 c33 O"yz (13.x.::.y' = c:::y· (13.v(axx 1 182 .
It can be defined as the ratio of the transverse contract strain (normal to the applied load) to the axial strain in a stretched cylindrical bar in the direction of the applied force.13) (13. and negative one in compression. For example.:: = gO"x.3. and v = 0 rv 0. v = 0.:. Sometimes.25 rv 0.. 1+v where v is the Poisson's ratio.17) where we have used the Einstein's summation convention O"kk = O"xx+O"yy+O"::::· Another related quantity is the pressure.1 Hooke's Law and Elasticity 1 c::::: = E [u:::: v(uxx + O"yy)]. and it measures the tendency of extension in transverse directions (say.Elasticity 13. (13.16) cxy = gO"xy. This generalized Hooke's law can concisely be written as (13. it is more convenient to express the stress tensor in terms of pressure and devitoric stress tensor Sij (13.5 for most common materials. x and y) when the elastic body is stretched in one direction (say. steels have v = 0.19) 183 . which is defined by p = O"kk 1 3 = (13. For a perfectly incompressible material. 1+v ex. Some auxetic material such as polymer foams or antirubbers have a negative Poisson's ratio v < 0. while the positive pressure acts in compression. z).14) (13.15) (13.5.18) The negative sign comes from the conventions that a positive normal stress results in tension.
L is called the shear modulus. we have (13. known as the dilation.2v) vE .1 Hooke's Law and Elasticity Elasticity akk If we want to invert equation (13.22) CJii = CJkk = _ v 1 2 Ckk· Substituting it into equation (13.26) This stressstrain relationship can also be written as u = 2Ge + A(V · u)l5. while K = 3 ( 1 ~2 . (13. (13.21) means the fractional change in volume. (13.27) In engineering.25) where J. the quantity Ckk = Cxx + Cyy + Cz::: = Bx2 + By2 + Bz 2 fPut 8 2 u2 82 u3 = V · u.20) aii = akk· where we have used 8ii = 811 + 822 + 833 = 1 + 1 + 1 = 3 and In engineering. we have first express in terms of Ckk so that the right hand side of the new expression does not contain the stress akk· By contraction using j + i. G = J. we have Cij = gaii .23) or after some rearrangement 1+v gaii = Cij + 1 _ 2v Ckk8ij. This gives that E (13. v (13. 184 .l = 2(1 + v)' A= (1 + v)(1 .17).24) which can be written as (13. They are E G = J.13.) is called the bulk modulus which is the ratio of pressure p to the volume change ~ V. (13.17).E ( 1 +v v 1 E _ 2 v Ckk)8ij.L and A are La1ne constants.
31) which means the stress tensor is symmetric.1). the balance of force leads to V ·U+ b = 0.32) 185 . thus the result moment must be zero. respectively. we only have to consider the two faces that are far from this point because the two faces through this point do not contribute to the moment. (13.2 P. and the dimensions of the elements are 8x.Elasticity 13. The forces on the two faces are O'xydAtdY and ayxdA2dx where the surface areas are dAt = 8x8z and dA2 = 8y8z. 8y and 8z. Alternatively. (13. a corner of the cube).30) Since there is no relative rotation of the cube element because the cube element must be at rotational equilibrium. This leads to (13. If we take the moment about a line which is parallel to zaxis and goes through one corner point.2 Maxwell's Reciprocal Theorem For an elastostatic problem. For a small cube volun1e element. We consider the forces along the four faces that are parallel to zaxis (shown in Figure 13. OXi (13. the total force along the jaxis is d/j = oaiidV. This is in fact the compatibility condition for stresses.28) where b is the body force or force per unit volume. The total moment about point D is (13.29) 0x· J here we have used the index summation conventions.tfaxwell's Reciprocal Theorem 13. Taking the moment of the two force components about any point (say. Similarly. the total body force dfi along the xiaxis is 8aii dfi = dbi = dV. consider a cube element at rotational equilibrium with a volume dV = 8x8y8z.
Using the stressstrain relationship.ail)ft = 0.2 !viaxwell's Reciprocal Theorem Elasticity Thus.34) provides the the principal stresses ai(i = 1. we know that a square matrix can always be expressed in terms of eigenvalues and eigenvectors. it is straightforward to prove that the strain tensor is also symmetric (cij = cji)· In engineering. (13. and there are no shear stress components. In the coordinate system formed by the three eigenvectors. From the matrix algebra. 3) (eigenvalues) and their principal directions ftT = n1. The nontrivial solutions require the determinant of the co186 . this tensor is often written as (13. n3 (eigenvectors). For the stress tensor.13. axy = ayx· Similar arguments for other faces along xand ydirections about any other points and we have aii = Uji· z y 8y Figure 13. the eigenvalue problen1 (u.1: A cubic element in an elastostatic body.33) so as to emphasize that the nondiagonal elements are for shear components Txy etc. 2. the stress is expressed by the three principal stresses along three principal directions. n2.
(r.)a2 (13.41) where a1.ax. + r.:.)]a [axaya::.::.:. Tx::.).ax.Tx::.36) + a::.35) a3.::.(axr. That is (Jx.::. + r. It can be expanded into a cubic equation = 0. + 2TxyTy:.(r. + a::. .Elasticity 13.(J Txy Tx:. (J:.38) (13. a2. this tensor can be transformed into a diagonal form (13. ay.y + r. and these invariants satisfy the characteristic equation (13. the three invariants become (13.a Txy Ty:..37) where I1 = tr( a ii) = ax + a y + a::.43) 187 .:.tfaxwell's Reciprocal Theorem efficient matrix must be zero.(J Ty:. (13. + ayr.y)] = 0.y + r. a3 are principal stresses..(ax+ ay +[axay + aya:.2 P. Written in terms of principal stresses. + a::. (13. Under appropriate transformations.T.39) h and = axay + aya:: + a:.. There are three invariants (It. h and I3) for the secondorder symmetric stress tensor aii.42) (13.
2 !viaxwell's Reciprocal Theorem Elasticity and (13.47) where Cij are the influence coefficients or flexibility matrix. we first apply /i slowly (so as to reduce the dynan1ical effect) with fi = 0. f 2 .45) (13.. In the case of only two forces. It is called Maxwell's reciprocal theorem or MaxwellBetti theorem. the principle of superposition applies and we have (13. . The work 188 .. Q2. In other words. Qn. then fi. the displacement at point i due to a unit load at another point j is equal to the displacement at j due to a unit load at point i. the final displacements are the same where fi is applied first. fn acted upon the body at n different points.. the system has no memory of the load history. or even both are applied at the same time.13. This theorem is essentially equivalent to say the displacements are pathindependent and independent of the order of the loads applied upon the elastostatic body.46) (13. The displacements at each point in the direction of the corresponding force are Ql. or fi is applied first then f i. . The total work done due to this set of loads is (13. which states that the influence coefficients (or flexibility) matrix is symmetric Cii = Cji· That is to say. For a linear elastic body. .44) Now consider an elastic body when n concentrated load f 1 . the displacement of point i is Cii/ i and the displacement of point j is Ciif i..48) There is a very useful theorem concerning these coefficients. . For two forces f i and f i.
51) Since fi · fj is a dot product and thus a scalar. 13.3 Equations of !vlotion done is !Ciifl where Jl = fi · fi.50) As the total work done should be independent of the order in which the loads are applied.52) This completes the proof of the :rviaxwell's reciprocal theorem. This theorem is the important basis for boundary element analysis and virtual work method in computational engineering. we have (13. Together with the generalized Hooke's law and relationship with displacement u. then f i. Now with /i kept fixed. which leads to (13.Elasticity 13.49) If follow the same procedure but slowly apply I i first. we apply I i slowly. the total work done is now (13.53) where p is the density of the elastic body. 189 . the force f means the force per unit mass.3 Equations of Motion For a general solid where the inertia is not negligible. in this case. 1 2 1 (13. the additional displacement at point i is Cij f i and the additional displacement at point j is Cjj f j. In some books. The total work done at the final state is W = 2ciifi + 2ciifi2 + Cii(fi · fj). this requires W = W. the extra work done is Cijfi · fj + !Cii!J. the following form of body force b = pf is used. we now get (13. In this case.
58) (13. [G(Vu + VuT) + . If we substitute the generalized Hooke's law and displacement into the first equation (13.56) There are 15 equations (6 for stresses. 6 strain components and 3 displacements).X(V · u)d. we have V · [2Ge + .60) V.55) (13.57) (13.13.3 Equations of Motion Elasticity we have the following set of equations of motion for an elastic body.57).54) (13. 6 for strains. the elastic field should be uniquely determined if appropriate boundary conditions are given. we have V·u+b=p8t 2 ..X(V · u )«5] + b = P 8t2 or {)2u . (13. There are other compatibility equations as well. If we write the equations of motion using the bold font notations. {)2u (13.62) .X(V · u)«5] + b = P {)t 2 which leads to (G + .61) [)2u • (13.59) u = 2Ge + . 3 for displacement) and we have 15 unknowns (6 stress con1ponents. (13. Therefore. and we will briefly discuss them later. 1 e = 2'(Vu + VuT). (13.X)V(V · u) + GV u + b = P {)t 2 190 2 {)2u .
Furthermore..66) Then.67) Since A+ 2G > G. i.3 Equations of !vlotion Using G +A= G/(1.. the displacement field is nonrotational.2v) and after some rearrangements. we differentiate Cxx with respect 191 .>.e. we have Cxx =ax' au Cyy = 8v oy' By assuming the displacements are continuous and differentiable functions of positions. u2. from the definitions of the strain components in terms of displacements uT = (ut. V x (V xu)= 0.65) For the Pwave in 1D.+p2G). v. w). (13.Elasticity 13. the speed of P. Pwaves always travel faster than Swaves. u3) = (u.wave is Vp = J(. we can take V · u = 0 for Swave. (13.64) thus its wave speed is vs=~· (13.63) which is the wellknown CauchyNavier equation. From the identity V(V · u) = V x (V x u) + V 2 u. the 1D CauchyNavier equation becomes P 82u1 fJt2 = (A + 2G) a2ut ax2 . we have (13. therefore. In the simplest 1D case without any (external) body force b. the equation is simplified as (13. This equation supports both longitudinal wave (P wave) and transverse wave (S wave).
we are dealing with plane strain and plane stress problems.4 Airy Stress Functions Elasticity to y twice.70) Now differentiate cxy with respect toy once. In the same fashion.13.72) which is the compatibility equation. We have only three independent stress 192 . In this case. =I= 0). we have (13. and with respect to x once. then the plane stress problem involves no stress components depending on z.74) 13.4 Airy Stress Functions For certain engineering problems. the solutions in a plane is of concern. we can derive other compatibility equations (13.:. For a plane stress problem. we have (13. This can be rearranged as 2 8 cxx 8y2 + 82cyy 8x 2 = 28 2 cxy 8x8y ' (13. 73) (13.69) Similarly.71) where we have used the interchangeability of partial derivatives fPvj8x8y = 8 2vj8y8x. we assume that a:::: = 0 (but c:. That is to say Uxz = Uyz = Uzz = 0. differentiate cyy with respect to x twice. we have (13.
called Airy's stress function.Elasticity 13. however. = v(axx + ayy) is not zero. Thus.vayy). it becomes 82 [8r2 18 + .80) In this case.81) which is a biharmonic equation and can be written as (13. (13.79) For plane strain problems with no body forces. = . The Airy's stress function is defined by (13.77) 1+v However.83) 193 .75) Cxx = E(axx.. the compatibility equation becomes (13.4 Airy Stress Functions components axx' aYY' and axy· The generalized Hooke's law reduces to 1 (13.::. the equilibrium equations are automatically satisfied if the stress components are related to a scalar function q. z). the stress a::. = 0. Cyy = E(ayy. The compatibility equation becomes 82Cxx + fPcyy 8y 2 8x 2 = 2 fPcxy 8x8y · (13. Cxy = gaxy· 1 (13. 8r + 182 2 r2 882] q. (13.82) In cylindrical polar coordinates (r.(Cxx v 1 v + Cyy).76) (13. 8.::.78) which is not zero in general. there are only three independent strain components Cxx' Cyy' and Cxy.::. For plane strain problems. . it is assumed that C::. C::.O.vaxx).
4 Airy Stress Functions Elasticity Figure 13. {)O) = 0. (13. The surfaces of the crack shall be stress free.2.1n(n + 1)/(0).88) .2: A crack in an elastic plate.86) As the secondorder equation y" + >. which leads to the following boundary conditions U()() = [)2 cp {)r2 = 0. Now consider a semiinfinite crack in an infinite solid as shown in Figure 13. we get [d02 d? + (n + 1) 2 ][dd?. we can here use this method twice. Ur() = {) 1 {)cp {)r (. 2 (13. 2 y = 0 has a general solution y = A sin >.1)0. at 0 = ±7r. (13. 194 (13..87) +C cos(n.O + B cos >.84) Let us try a solution of the form cp = rn+l f(O). The boundary conditions become aeo = rn.. the general solution takes the following form f(O) = Acos(n + 1)0 + Bsin(n + 1)0 (13. The far field is subjected to bounded stress at infinity.. assuming the elastic body deforms in plane strain.13.O.85) and substitute it into the governing biharmonic equation.+ (n02 1) ]f(O) = 0.1)0 + D sin(n.
but the solutions has singularity as r+ 0.95) (13. we have sin(2n7r) = 0. ro 6=0 Finally..1)8.92) which is for the opening (model I) of the crack..93) K1 =lim uee(r. We know n = 0 is trivial.Elasticity 13.1)[C sin(n.D = 0. From the first equation. 0.1)8]}. therefore.4 Airy Stress Functions and Ure = rnln{(n + 1)[Asin(n + 1)8. ±2. (13. . . The constraint now becomes n ::. ±1. This is however acceptable in the crack propagation as the stress concentrations do physically exist. (13. 8 8 2 2 (13. 2 )COS.91) By defining the stress intensity factor K1 for the crack. we get 3A+C= 0.D cos(n.90) and rn(n ~ 1) does not converge. Substituting the general solution into the boundary conditions with n = 1/2 and 8 = ±1r. B.Bcos(n + 1)8] (13. 8)1 . M=:(1 }(/ v 21rr + Slll . 8 . at 8 = ±1r. the solution of stresses can be written as Urr = .94) K1 U()() = v'27fTCOS }(/ 2 3 2' 8 (13.89) +(n. they are not suitable solutions.96) Ure = v'21fT cos 2 s1n 2· 195 8 . K _ 3A/2i I 4 ' (13. It is a limit of stress at 8 = 0 (13. 1 3 n = ±2.
5 EulerBernoulli Beam Theory Elasticity Once we have the stress distribution. 2) the beam deformation is dominated by bending. Under these assumptions. and f(x. The first assumption implies that the bending moment M is proportional to the curvature K of the bending.97) and ue = K1(1 .98) 13.5 EulerBernoulli Beam Theory The EulerBernoulli beam theory is a simplified theory for calculating the deflection of beams under a distribution of load force using the linear isotropic theory of elasticity. 30 E V.100) 196 . and we have (13. 3) the beam is long and slender with a constant cross section along the axis. (13. That is M=EIK. 0 (13. Then the displacements are the integration of the strains.13.99) where E is the Young's modulus and I is the area moment of the beam's cross section.3). we can get the strains. In mechanics. It is worth pointing out that the area moment about a horizontal axis through the centroid is defined by (13.v) fiT . Let u(x. I is also called the second moment of area or the area moment of inertia. t) be the force per unit length. we can now derive the governing equations.r[sin 2 .4v) sm 2]. . t) be the deflection of the beam (shown in Figure 13. The basic assumptions for the beam theory are: 1) the beam is isotropic and elastic. and distortion and rotation are negligible.(5. A be the area of the cross section.
we have I = bh3 /12. but we use J here) about an axis.102) and the third assumption means tx << 1. V(x)/ X . Therefore. and it should not be confused with the mass moment of inertia J (also often denoted as I. and the V::::::: ~(EI82u)..3: Beam bending. t)dx .t) Figure 13.:~ rI ~ /v(x+dx) f(x.. 197 . _____ _J Tu(x..r===:.::.104) 8x 8x 2 For a small volume element (also shown in Figure 13.. we have (13.101) with a unit of [Kg] [m] 2 . For a cylindrical rod with a radius of R...Elasticity 13. we have I= 7rR4 j4.3). the mass of the element is pAdx where p is the density.. For a rectangular beam with a base width of band a depth of h. which is defined by (13..103) 82u M::::::: EI 8x2' or (13..5 EulerBernoulli Beam Theory which has a unit of [m] 4 . Both E and I do not change along the xaxis. The second assumption means that the shear V (x) is related to the bending moment 8M 8x = V(x). (13...
resulting in more fragmented pieces. t). This equation can essentially explain why spaghetti and dry pasta almost always break into more than two fragments. we have f(x. When a spaghetti rod snaps. the equation becomes a homogeneous form {)2u pA fJt2 + {)x2 [EI {)x2] = {)2 a2u 0. t) = 0. it will generally break three to pieces.107) which is the EulerBernoulli equation. a rod or any slender column.105) V(x). Substituting the above expression for V.ax ]dx. t)dx = 8V [f(x. we have 82 u (13. ()2 For the elastostatic problem.ax = pA at2 . 8V Using the Newton's second law of motion.13. 82 82 u (13. It governs the waves that travel along a beam. This phenomenon is very interesting and once puzzled the famous physicist Richard Feynman for quite a while.106) 82 u pA at2 + ax2 [EI ax2] = f(x). and it was recently studied by two scientists that the brittle fragmentation process is virtually governed by this CauchyNavier equation. You can try in your kitchen to break a slender spaghetti by holding its two ends and gradually form an arc and bend beyond its curvature limit.5 EulerBernoulli Beam Theory Elasticity acceleration is The shear force variation V(x V (x) + ~dx. If there is no force f(x. we have 82 {)2u ax2 [EI {)x2] = q(x). + dx) (13. t) .109) . and these waves locally increase the curvature in the rod. 7Jfi : : : : 0. They found that the sudden relaxation of the curvature by first breaking will lead to a burst of flexural waves along the spaghetti rod. 198 (13. (13. and the total force is fff.108) which is a fourthorder wave equation.V(x + dx) + f(x.
5 EulerBernoulli Beam Theory where q(x) = f(x) is the applied force per unit length. This equation will be used to determine the deflection of a bean1. we have du q EI. If the length of the cantilever is L. Integrating once again. we have D = 0. thus q = ~. At the free end. 0 Example 19. Therefore.Elasticity 13. These conditions lead to A = qL and B = qL 2 /2. Therefore.x2 2 + x+C· 2 qL 2 As the beam is fixed at x = 0.1: Let us now use the EulerBernoulli theory to calculate the shape of a heavy cantilever with a uniform cross section under its own gravity.= x3 dx 6  qL . which implies that M(L) = 0 and V(L) = q*L+A = 0. we have EI dx 2 = cPu 2x q 2 +Ax+B. Thus we have C = 0 from u = 0. the beam cannot support bending moment and/or shear. we have d4 u W EI dx4 = q = T where we have use EI =canst. the force is constant q(x) = pgA per unit length where g is the acceleration due to the Earth's gravity. we have u = 0 and :: = 0 at x = 0. the final deflection The end deflection 8 = u(x = L) is 0 199 . then the total weight is W = pgAL. For a beam under its own gravity. Integrating the above equation again. we have As Ux = 0 at x curve becomes = 0. Integrating it twice.
.
there are so many different kinds of phenomena that require both mathematical modelling and computer simulations as well as experimental studies. 201 . and subsequently we have to focus on the relevant equations and to introduce them very briefly. almost all books or studies on the partial differential equations will have to deal with these three types of basic partial differential equations. In most cases. we have to broaden our view to study other kinds of partial differential equations.Chapter 14 Mathematic al Models What we have discussed so far in terms of differential equations is very limited. In engineering and natural sciences. 14. mathematical modelling per se is a subject with vast literature. the classical models (using the heat conduction equation and the wave equation and others) are simply not adequate to described these phenomena. Therefore. let us first remind us the three types of classic partial differential equations because they are widely used and occur in a vast range of applications.1 Classic Models Before we introduce more complicated partial differential equations. To a certain extent. In fact.
. if the temperature is not known. the appropriate boundary conditions are needed. this becomes the Laplace equation. the steady state of heat conduction with a source is governed by the Poison equation (x. t) is the heat source. A simple boundary condition is to specify the temperature u = uo on the boundary an. This type of problem is the Dirichlet problem.2) for two independent variables x and y. y. such as diffusion and transient heat conduction.. some problems may have a mixed type of boundary conditions in the combination of au + (3 g~ = 1'. Here k is thermal diffusivity and I (x.1. . which naturally occur as a radiation or cooling boundary condition.!fathematical l.!fodels 14. In order to determine the temperature u completely. x2 = y. X3 = z.y) En.4) where I is the reaction rate. but the gradient 8uf 8n is known on the boundary where n is the outwardpointing unit normal. FUrthermore. t). (14. 202 . The solution or a function is said to be harmonic if it satisfies Laplace's equation. and this forms the Neumann problem. or Uxx (14. If there is no heat source (q = 0).1. 14.2 Parabolic Equation Timedependent problems.3) Written in thendimensional case Xt = x.1) + Uyy = q(x. it can be extended to the reactiondiffusion equation (14.14. On the other hand. are governed by the parabolic equation (14.1 Laplace's and Poisson's Equation In heat transfer problems. .1 Classic Models l. y.
t) where f and g are and x . the wave equation can be written u~ 11 = 0.1. arbitrary functions of x + ct = f(x (14.6) fort> 0 and oo as < x < oo. + ct) + g(x.1\t/athematical 1\t/odels 14. respectively. we have u(x. (14. 14.2. 'f/ = x.7) Integrating twice and substituting back in terms of x and t.2 Other PDEs 14. engineering and computational sciences.ct). Using a transformation of the pair of independent variables ~ = x+ct.ct. One wave moves to the right and one travels to the left at a constant speed c. The 1D wave equation in its simplest form is Utt.1 Elastic Wave Equation As we have seen in the linear elasticity. the wave in an elastic isotropic homogeneous solid is governed by the following 203 . We can see that there are two directions that the wave can travel.2 Other PDEs We have shown examples of the three major equations of secondorder linear partial differential equations. There are other equations that occur frequently in mathematical physics.8) 14.ct. (14. We will give a brief description of some of these equations.C Uxx'  2 (14.3 Wave Equation The vibration of strings and travelling sound waves are governed by the hyperbolic wave equation.5) where cis the velocity of the wave.
(14.11) (14. One example is the combustion equation Ut = D Uxx + Que >.2 Other PDEs !Ylathematical A1odels equation in terms of displacement u. B is the magnetic field and J is the current density. 14. The first equation is the electrostatic equation. z. A and J.14) where D is the diffusion coefficient and f is the reaction rate.3 ReactionDiffusion Equation The reactiondiffusion equation is an extension of heat conduction with a source Ut = Dtl.L)\7(\7 · u) + b. E is the electric field.12) \7 · B = 0.L\7 2 u +(A+ J.2. 14..at' 8B (14.9) where p is density.15) 204 . Finally. c is the speed of light.10) VxE=.13) where Pe is the charge density.lo are the permittivity and permeability of the free space. respectively. Such an equation can have two types of wave: transverse wave (Swave) and longitudinal or dilatational wave (Pwave). (14. u).u + f(x.2 Maxwell's Equations The scientific essence of modern wireless communications is governed by the :Maxwell's equations for electromagnetic waves (14. the second one is the Faraday's law and the last equation is the AmpereIviaxwell's law. b is body force. eo and J.14.ju . (14.LoJ + c2 7K' 1 8E (14. \7 x B = J.l are Lame constants. y.2. p 82 u 8t2 = J.
•••• 1kp ' ..2. } (14.2.L. x 2 . Xn) is the drift vector. (14..19) where D[k] (x 1 .4 FokkerPlank Equation The time evolution of the probability density function of position and velocity of a particle system is described by the FokkerPlank equation =[L:1 +L:L: 8t i=l 8p n aD!ll axi n n 82 n!~l lJ ]p.. ..lk .2 Other PDEs where D. t) at time t is governed by the wellknown Nobelwinning BlackScholes equation au {)t = ru . 8Xii 8Xik k=lt}.5 BlackScholes Equation In the option pricing model.1)k_!__ ./V). ... the value of an option u( 8. nl 11(xt' . a is the stock volatility or the standard deviation of 205 . Q and A are constants.20) where 8 is the current stock price of the underlying stock and r is the riskfree interest rate.. 14.r 8 a8 au  2a 1 2 28 2 8 882 ' u (14. where /' > 0. . 14. N ZI .1\t/athematical 1\t/odels 14. In the special case of N = 2. These equations are sometimes also called the equations in excitable media.16) (14. t is the time until option expiration. 0 < a < 1 and E << 1. The other example is the FitzHughN agumo equations for transport of a nerve signal Ut = Uxx + u( 1  U) ( U  Q)  V. xn) are tensors.17) Vt = f(U. _!__{D(kJ L.. "" (.. while Dl21 is the diffusion tensor. .18) i=l i=l 8xi8xi which can be generalized as 8p at ="" . (14.
2m ':l'+ u·T· ':l' ' (14.14. (14. 14.2. (14.25) Ut + (u · V)u = 1 2 ReV u .6 Schrodinger Equation The famous Schrodinger equation is the revolutionary equation in quantum mechanics and molecular dynamics ZHv ·i:_ aw 8t tt2 t"'72 .E is the energy potential.2. V = uer(Tt) (14.21) and x = ln(S/E) + (2 2r (J 1)r. (14. This equation can be obtained from the energy formE=::+ U (where pis the momentum) using differential operator mapping E + itt£ and p + itt V. the change of variables T = (T.t)a 2 /2. we can use the standard methods such as integral transforms to solve this equation.26) 206 .23) Then. The interesting feature for a call option with an exercise price E and expiry time T. W is the probability wave function and U = V (r) .24) where tt is a Planck constant. 14.y. This is a complex partial differential equation. The BlackScholes equation is very similar to the extended version of the diffusion equation.22) can transform it into a standard diffusion equation (14.Vp.2 Other PDEs !Ylathematical A1odels stock returns. 7 NavierStokes Equations The NavierStokes equations for an incompressible flow can be written as V·u=O.
2 Other PDEs where Re = pUL/J. respectively. >> 1.28) We can see that the equations are still nonlinear even in this simplified case. (14. U is the typical velocity and L is the length scale. In computational fluid dynamics. Therefore.29) which can generally be written as Utt = Uxx + asin(wu). we have the inviscous flow Ut + (u · V')u = V'p.27) In another limit Re V'. p and J. we have the Stokes flow (slow flow) governed by (14. In the limit of Re << 1. (14. Almost all these equations are very difficult for mathematical analysis.1\t/athematical 1\t/odels 14. u = 0.8 SineGordon Equation Another important equation that appears in a wide range of applications in physics and many other fields is the SineGordon equation Utt. we will introduce various numerical methods in details. the numerical methods are a good alternative in this case. (14. 14.30) This is a nonlinear hyperbolic equation. and most of them do not have closedform solutions under most common boundary conditions.Uxx + sin(u) = 0.2.l are the density of the fluid and its viscosity.l is the Reynolds number. In the next few chapters. 207 . most simulations are mainly related to these equations.
.
The finite difference method essentially transforms an ordinary differential equation into a set of algebraic equations by replacing the continuous derivatives with finite difference approximations on a grid of mesh or node points that spans the domain of interest based on the Taylor expansions. 209 (15.1 Integration of ODEs The secondorder or higher order ordinary differential equations can be written as a firstorder systen1 of ODEs. It has the advantage of simplicity and clarity. 15.1) . especially in 1D configuration and other cases with regular geon1etry.y). In general. Since the technique for solving a system is essentially the same as that for solving a single equation dy dx = f(x.Chapter 15 Finite Difference Method The finite difference method is one of the n1ost popular methods that are used commonly in computer sin1ulations. the boundary conditions and boundary nodes need special treatment.
RungeKutta method. y)dx ~ Yn + hf(xn. Thus.2) but in practice it is usually impossible to do the integration analytically as it requires the solution of y( x) to evaluate the righthand side. y(x))dx. and many others.Yn ~X (15. stability is an important issue in numerical analysis. y(x) =Yo+ {x f(x. 15. These schemes include the simple Euler scheme. the Euler method is first order accurate. . some approximations shall be utilized.3) This is a forward difference method as it is equivalent to the approximation of the first derivative 1 Yn = Yn+l. the solution can be obtained by direct integration. 2. (15.1 Integration of ODEs Finite Difference Metbod then we shall focus on the firstorder equation in the rest of this section.1 Euler Scheme Using the notations h = ~x = Xn+l . Yn) + O(h2 )..15. and ' = djdx for convenience. lxo (15. then the explicit Euler scheme can simply be written as Yn+l = Yn + 1 Xn Xn+l f(x. Xn = x 0 + n~x (n = 0. 1. the algorithm must be stable in order to reach convergent solutions. Relaxation method.5) Thus.4) The order of accuracy can be estimated using the Taylor expansion h2 Yn+l = Yn + hY'I n + TY "I n + ··· ~ Yn + hf(xn. Numerical integration is the most common technique to obtain approximate solutions.Xn.1. In principle. Yn = y(xn).. Thus. N). . For any numerical algorithms. Defining 8y as the 210 . Yn)· (15. There are various integration schemes with different orders of accuracy and convergent rates.
the step size cannot be too large as the accuracy reduces as the step size increases. One alternative that can use larger h is the implicit Euler scheme. This means that any step size is acceptable.1: U ::. Thus. However.1 Integration of ODEs discrepancy between the actual numerical solution and the true solution of the Euler finite difference equation. The implicit scheme generally gives better stability.Finite Difference Method 15.7) The stability condition becomes (15. The stability restricts the size of interval h. To solve the equation dx = f(y) = ey. 0 Example 15. it requires the following stability condition 1~1 ::. and the iteration starts with a guess value which is usually taken to be zero for most cases.y. an iteration procedure is needed to march to a new value Yn+l.6) In order to avoid the discrepancy to grow. dy we use the explicit Euler scheme. we have (15.Ynd/h and the righthand side of equation (15. 1. for most problems such as nonlinear ODEs. which is usually small. Another practical issue is that. and we have 211 . and this scheme approximates the derivative by a backward difference y~ = (Yn. the evaluation of y' and f'(y) requires the value of Yn+l which is unknown.8) which is always stable if f'(y) = 0. Now the scheme can be written as (15.2) is evaluated at the new Yn+l location.
The stability condition becomes 2 O<h<~ . Since f(y~) = f(Yn) Yn+l + :~8yn. Therefore. Yn)· (15.0). =1h(eYn In order for the scheme to be stable (or en.10) The central difference method is second order accurate.Yn1 The Leapfrog scheme is the central difference 1 Yn = 2Ax ' (15.15.12) or 8yn+1 = ~ 8Yn1' 2 where ~ = 1 + 2hf'(y)~. it requires that or .11) (15. and a special case is 1~1 = 1 when f'(y) is purely in1aginary.ey. then the real solution satisfies y~+l = y~ + hf(y~). 212 .1 Integration of ODEs Finite Difference Method Suppose the discrepancy between real solution y~ and the numerical Yn is 8yn so that y~ = Yn + 8yn.2 LeapFrog Method Yn+1 . the above equation becomes + 8yn+l = Yn + 8yn + h[f(Yn) + J'(Yn)8yn]· 8yn+l Together with the Euler scheme.1 ~ 1 + hf' + 1) ~ 1. + 1 0 15. This scheme is stable only if 1~1 ::.6). the central scheme is not necessarily a better scheme than the forward scheme. then we have en+l =en+ h/'en' e= 1 + hf'.1.9) which leads to Yn+1 = Yn1 + 2hf(xn. the leap frog method becomes 8yn+1 = 8Yn1 + 2hf'(y)8yn. 2 (15. 1. we have = 8yn + J'8Yn· or Suppose that 8yn ex en. In a similar way as equation (15.
Yn). (15. the higherorder scheme is better than the lower scheme. One can view this scheme as a predictorcorrector method. The popular classical RungeKutta method can be written as a= hf(xn. Usually. Yn+l = Yn + 15. Yn). a+ 2(b +c)+ d d = hf(xn + h.13) (15.yn +c).. In addition. Now we start with the simplest first order equations and then move onto more complicated cases. b = hf(xn + h/2. but not always. 213 . (15.1. f(xn. Yn C = + a/2). Generally speaking. In fact.Finite Difference Method 15.14) This scheme is second order accurate with higher stability compared with previous shnple schemes. The RungeKutta method uses a trial step to the midpoint of the interval by central difference and combines with the forward difference at two steps Yn+l/2 = Yn + . we can use multisteps to devise higher order methods if the right combinations are used to eliminate the error terms order by order.15) ' 6 which is fourth order accurate..3 RungeKutta Method We have so far seen that stability of the Euler method and the central difference method is limited. nonlinear problems are very common in engineering applications. hf(xn + h. boundary conditions are more complex.2 Hyperbolic Equations 15.2 Hyperbolic Equations Nun1erical solutions of partial differential equations are more complicated than that of ODEs because it involves time and space variables and the geometry of the domain of interest. Yn + b/2).
u"!1 h 1  1 ] = 0. Assun1ing the independent solutions or eigenmodes (also called Fourier modes) in spatial coordinate x in the form of uj = ~neikhi. (15.. we get exp( ikhj) and ~= 1. j1] = 0.21) 214 .u"!J ~t J + c[ un un i+1. x = x 0 + jh.19) The stability criteria 1~1 ~ 1 require . h (15.18) Dividing both sides of the above equation by using sinx = (eix. . (15.~n ikhj e ~e = ~n c~t eikh(j+1) ~ h 2 e ihk(j1) .15. (15. n = 0.17) where t = n~t. ( c~t )2 Sill 2 kh ~ 0 .eix)/2i.. . we have ~ ~n+ 1 ikhj .. 2.16) where c is a constant or the velocity of advection. the equation becomes u~+ 1 . we have u~+ 1 . ~n (15. 1. this inequality is impossible to satisfy and this scheme is thus unconditionally unstable. 2.17). (15. we start with the onedimensional scalar equation of hyperbolic type.2.ic~t sin(kh). .2 Hyperbolic Equations Finite Difference Metbod 15. we use the von Neumann stability analysis. By using the forward Euler scheme for time and centeredspaced scheme. 1.1 FirstOrder Hyperbolic Equation For simplicity.. . we can use other schemes such as the upwind scheme and Lax scheme.. To avoid the difficulty of instability. and h = ~x.20) However.u"!1 1 ~t + c[ u"!. In order to see how this method behaves numerically. and substituting into equation (15.j = 0. For the upwind scheme.
(15.22) (15..uf1 ~x (15. The wave equation (15..cos(kh) + isin(kh)]l ~ which is equivalent to 0<..2uf + uf1 (~x)2 (15..u~1 U· =. the upwind scheme is conditionally stable... 15.Finite Difference Method 15.... 215 . Thus..... t ~X uf+1..uf i~X .23) This the wellknown CourantFriedrichsLewy stability condition...:~l. c~t 1.24) consists of second derivatives.2 SecondOrder Wave Equation Higher order equations such as the secondorder wave equation can be written as a system of hyperbolic equations and then be solved using numerical integration.2 Hyperbolic Equations whose stability condition is 1~1 =II..... If we approximate the first derivatives at each time step n using u'.26) This is in fact a central difference scheme of second order accuracy. often referred to as the Courant stability condition....uft1.c~t [1.. ui1 = uf.. They can also be solved by direct discretization using a finite difference scheme.. then we get a central difference scheme in both time and space.2. If we use the similar scheme for timestepping.25) then we can use the following approximation for the second derivative " u~.
we can apply the von Neumann stability analysis by assuming uj = f. = 1 + """h2"[e Dt::.29). we have ·kh 1 f.2uf + ui1 2 uf+1. 15.t + ez·kh  2].neikhj.1. (15.neikhj.15.27) This is a twolevel scheme with a second order accuracy.2: From equation (15. 216 . 1 leads to the constraint on the time step (see the example).1 and it is conditionally stable.3 Parabolic Equation For the parabolic equation such as the diffusion or heat conduction equation 8u = ~(D8u) 8t 8x 8x ' (15.1 at two previous time steps t = n and t = n .30) This scheme is shown in Figure 15.29) The stability requirement ~ ::. 0 Example 15. we have Dividing both sides by f.2uf + uf1 (~t)2 =c (~x)2 (15.28) a simple Euler method for the time derivative and centered secondorder approximations for space derivatives lead to (15.3 Parabolic Equation Finite Difference Method Thus.. The idea of solving this difference equation is to express (or to solve) ui+ 1 at time step t = n+ 1 in terms of the known values or data uf and ui. the numerical scheme for this equation becomes ui+1.
< h2 < 0 or 0 A typical feature of a solution to the diffusive system is that the profile is gradually smoothed out as time increases. 4DD.6.Finite Difference Method t 15.3 Parabolic Equation Q I I I t=n+1 Ot=n x· J 0 j1 6 j j+1 Figure 15. Using cosx = (eix + eix)/2 and sin2 (x/2) = (1.1 . 2 kh Since sin( X) ~ 1' thus e~ 1 requires 4DD..t h2 J+l + 2u~+l + u~+ 1 ) . To avoid the difficulty caused by very small timesteps. we have ~ = 1 + fiT""sm 2 kh' 4D.= (u~+l J J DL).t 1 .1. The timestepping scheme we used limits the step size of time as larger time steps will make the scheme unstable. and one of most widely used schemes is the implicit scheme.31) Applying the stability analysis. we obtain e=1~sm(2). There are 1nany ways to improve this.t . we now use an implicit scheme for time derivative differencing.t · 2 217 (15. and thus we have u~+l .1: Central difference in space and explicit Euler timestepping..uf!..cosx)/2. J J1 1 (15.32) .
which requires more programming skills because the inverse of a large matrix is usually needed in implicit schen1es. x. then it becomes a steadystate problem that is governed by the elliptic equation. Assuming t + oo and K >> 1. V · [K(u. t)Vu] If K is a constant. and the multigrid method. K f (15.35) evolves with a typical scale of ..36) K K 218 . (15. the system is approaching its steady state..4 Elliptical Equation Finite Difference !v!etbod whose norm is always less than unity (1~1 ~ 1)./'Kl. For the steady state heat conduction problem. if the time derivative is zero or u does not change with time Ut = 0. Two major ones are the longtime approximation of the transient parabolic diffusion equations. y. the other includes the iteration method.33) q=. this becomes = f. (15. we then have 2 f 1 V U = .15. 15. we generally have the Poisson problem.34) There are many methods available to solve this problems such as the boundary integral method. However. >> 1. This means the implicit scheme is unconditionally stable for any size of time steps. The long time approximation method is essentially based on fact that the parabolic equation (15. there is one disadvantage of this method.4 Elliptical Equation In the parabolic equation. If VKJ. That is why in1plicit methods are more desired in simulations. the relaxation method.Ut + 0.
The GaussSeidel iteration provides a more efficient way to solve this equation by splitting A as A=L+D+U.33) because Ut + 0 as t + oo. Thus.37) If we use Ax comes = Ay = h. (15. the usual numerical methods for solving parabolic equations are valid.j (Ay) 2 + Ui. other methods may obtain the results more quickly. this becomes impractical as the matrix becomes large such as 1000 x 1000.39) In principle. 219 .41) This procedure stops until a prescribed error or precision is reached.j + Ui. however. The iteration is updated in the following way: (15. diagonal and upper triangle matrices of A~ respectively. one can solve this equation using Gauss elimination.j .2Ui. then the above equation simply be(15.38) which can be written as Au=b. This approximation becomes better if K >> 1. However.34) in the 2D case becomes Ui+l. D~ U are the lower triangle. and equation (15. (15.4 Elliptical Equation The longtime approximation is based on the fact that the parabolic equation in the case of K = const degenerates into the above steadystate equation (15.Finite Difference Method 15.j1 = q.j+l . The iteration method uses the secondorder scheme for space derivatives.40) where L.j (Ax)2 + Uil. (15.2ui.
.
'\Ju. resulting in a system of algebraic equations. then approximates the surface and boundary integrals in the control volumes.Chapter 16 Finite Volume Method 16. It uses the integral form of the partial differential equations in terms of conservation laws.1) or written in terms of flux function F have = F(u) = K. This becomes convenient for problems involving flow or flux boundaries.3) 221 .. F q. The finite volume method resembles the finite difference method in certain ways but the starting point is the integral formulation of the problem.2) The integral forn1 of this equation becomes In '. For a hyperbolic equation that is valid in the domain n with boundary an' (16. (16. we 8u at + v · F = V' .1 Introduction The finite difference method discussed in the previous chapter approximates the ordinary differential equations and partial differential equations using Taylor series. (16.: + In dQ = In qdf!.
8) is a vertexcentred finite volume scheme. The integration can be approximated using various nun1erical integration schemes. 222 Laplace's equation is one of the most studied elliptic equations (x. we have (16. y) E 0. n = U~ 1 ni and ni ni = 0. y) = 0. (16.9) .2 Elliptic Equations Finite Volume !v!etbod If the integral form is decomposed into many small control volumes. the integration ui = h 11(i+l/2)h (i1/2)h udx. or finite volumes. parabolic and hyperbolic) and their finite volume discretizations.2 Elliptic Equations \7 2 u(x. In the following sections. we will discuss the three major types of partial differential equations (elliptic. 7) where dS = ndA is the surface element and n is the outward pointing unit vector on the surface ri enclosing the finite volume ni. In the sin1plest 1D case with h = ~x.6) (16.16. 16. (16. By defining the control volume cell average or mean value n (16.4) where \ti = IOil is the volume of the small control volume the above equation can be written as ni. (16.5) By using the divergence theorem fv \7 · F = £F · ndA.
In fact.q)dxdt = 0.ui.] 8u . the control volume in this case is a cell centred at (i~x.14) .= 8n ~X ~y A uX (ui+l. 223 82 u (16.3 Parabolic Equations For the case of heat conduction (16.13) we have its integral form lt ln ( 8t  [ [ {)u k {)x2 .j+l + Uil.ui.j ~x~y.j + Ui.j)/~x and = (ui. (16. j ~y) with a size of ~x (along xaxis) and ~y (along yaxis). which resembles finite difference methods in many ways. (16.j1 Dividing both sides with obtain (ui+l.12) + Ui.3 Parabolic Equations its integral form becomes 1 n '\7 2 udO. = 18n · 8u r dS = 0.11) and letting ~x = ~y = h. 2ui.j.j) + ~y (ui.10) For the simple regular grid points (i~x.i + Uit. 16.j+l. this is exactly the Laplace operator for a 5point differencing scheme. and the boundary integral on any cell consists of four parts integrated on each of the four sides.j = 0. we have t 1 f2·.j 2ui.jt). By using the simple approximation ~ with g~ = (ui+l. t.j+l + ui.j~y).j) = 0.j)j~y. we (16.4Ui.dO.Finite Volume l'vlethod 16.
we have u~+1. 8w(u) _ 0 ' (16.1/2)h to (i + 1/2)h where h = D. (16.18) where we have used the central scheme for time as well.4 Hyperbolic Equations Finite Volume !v!etbod If we use the control volume from (i .u~1 t t 2D. and with time from step n to n + 1.x.a).17) {)x2 = h2 then replace 2uf with (uj+ 1 + uj. Following the stability analysis. we get 1~1 < 1 is always true and thus the DufortFranke! scheme is unconditionally stable for all .t = (16. we have {(n+1)At 1(i+1/2)h {Ju fJ2u lnAt (i1/2)h ( {)t . b)](b.15) By using the midpoint approximation l '1/J(x)dx = '1/J[(a.16) and the DuFortFranke! scheme where we first approximate the gradient [)2u uf+1 .k 8x 2  q)dxdt = O.2uf + uf1 (16.4 Hyperbolic Equations For the hyperbolic equation of the conservation law in the onedimensional case {)u {Jt + ax.19) we have its integral form in the fixed domain 1 Xb . (16. The finite volume scheme is more versatile in dealing with irregular geometry and more natural in applying boundary conditions. 16.d x ={) 1Xb udx {)u {Jt Xa {Jt Xa 224 .1).Dot and D.x.16.
Finite Volume l'vlethod 16.w[u(xa)]} = 0. h (16.20) If we use the midpoint rule u* to approximate the integral. (16.23) which is conditionally stable as we know this in the finite difference method.25) 225 .21) If we choose the control volume [(i 1/2)~x.22) By further approximation of the flux \ll(xi+l/2) have the upward scheme ~ \ll(xi). (i + 1/2)~x] centred at the mesh point Xi = i~x = ih with the approximation ui ~ ui in each interval. we have (16. we have u n+l = ui  n c~t h (uin  n ) ui1 ' (16.1. we (16. and using the forward differencing scheme for the time derivative.4 Hyperbolic Equations = {w[u(xb)]. For the simplest flux w(u) = cu.24) and its stability requires that 0 < < c~t . we have (16.
.
For example. and there are many limitations in finite difference methods. For timedependent problems involving partial differential equations. One of such disadvantages is that it is not straightforward to deal with irregular geometry. the equations can be converted into an ordinary differential equation. The basic aim of the finite element method is to formulate the numerical method in such a way that the partial differential equation will be transformed into algebraic equations in terms of matrices. lviore versatile and efficient n1ethods are highly needed. 227 .1) where K is the stiffness matrix. which will in turn be discretized and converted into algebraic equations by timestepping or some iteration techniques. u is the unknown vector to be solved and it corresponds to the nodal degree of freedon1 such as the displacement.Chapter 17 Finite Element Method In the finite difference method. a linear elastic problem can be formulated in such a way that it is equivalent to the equation of the following type Ku=f. and f is a vector corresponding to nodal forces and some contribution from boundary conditions. (17. we approximate the equations at a finite number of discrete points. the finite element method is one class of the most successful methods in engineering and have a wide range of applications. In fact.
we have = k(~u).4) 228 . the triangular element with three nodes in 2D.2) (17.Ui f At node i.1 Concept of Elements Simple Spring Systems The basic idea of the finite element analysis is to divide a model (such as a bridge and an airplane) into many pieces or elements with discrete nodes.1: Finite element concept. or ~u = Uj .1 17. we know the displacement is related to f.1. From Hooke's law. /i and /j are nodal forces. k Figure 17. strains can then be evaluated at certain points (usually Gaussian integration points) inside elements. we get (17. The simplest elements are the element with two nodes in 1D. These elements form an approximate system to the whole structures in the domain of interest. respectively.1 Concept of Elements Finite Element Metbod 17. In order to show the basic concept. The spring has a stiffness constant k (Njm) with two nodes i and j.17. At nodes i and j. we now focus on the simplest 1D spring element with two nodes (see Figure 17.1). so that the physical quantities such as displacements can be evaluated at these discrete nodes. and tetrahedral elements with four nodes in 3D. (17. Other quantities such as stresses. the displacements (in metres) are Ui and Uj.3) and at node j.
4). we have to assemble the whole system into a single equation in terms of global stiffness matrix K and forcing f. (17.5) Here K is the stiffness matrix. 2. respectively. For element E1.7) where (17.6) (17. u and f are the displacement vector and force vector. For a simple spring system shown in Figure 17. the assembly of the system can be done element by element. its contribution to the overall global matrix is k1 ( k1 which is equivalent to k1 ) ( k1 Ul ) = ( U2 fi ) ' /2 (17.2.2 where three different springs are connected in series.2: A simple spring system. we now try to determine the displacements of Ui ( i = 1.8) 229 .1 Concept of Elements These two equations can be combined into a matrix equation or Ku = f. and let us see how it works in a spring system such as shown in Figure 17.Finite Element 1\lethod 17. In order to do so. As these three elements are connected in series. 3. Figure 17. This is the basic spring element.
For element E3. has been used.j3 + /3 ' f* (17. 0.13) or where K = Kt +K2 +K3 0 k2 k2 +k3 k3 230 ( = T kl kl kt + k2 k2 0 _t} k3 (17.12) where /4 = /.11) (17. we have ( k2 k2 ) ( U2 ) k2 k2 U3 0 k2 k2 0 = ( /2 ) ' /3 (17.1 Concept of Elements Finite Element Metbod and !'f 1 = (ft. ) ( = /2+/2 ) !J . we have (17. for element E2.17.9) or KFU or 0 k2 k2 0 !} (17.14) . /2. Similarly. 0).10) where we have used the balance at node 2. We can now add the three sets of equations together to obtain a single equation ( t kl k2 kt + k2 k2 0 k2 0 k2 + k3 k3 0 ) ( U2 UJ k3 U3 k3 U4 Ku=f.
the stiffness matrix is 100 300 200 0 0 200 250 50 _t )' 50 and the force column vector fT = (0. By applying u 1 = 0. if no boundary condition is applied at any nodes. u 4 • Since u 1 = 0 is already known. then the system can stretch.1: For k 1 = 100 Njm. Then. the boundary at node 1 is fixed (u 1 = 0). we have (17. therefore. 20). then the applied force at the node 4 will make the spring system fly to the right. and the reduced forcing vector is gT = (0. we need at least one boundary condition.15) In general. This means that the equation has no unique solution. The rank of K is 3. u 3 . and J. If we add a constraint by fixing the left node 1. we need the boundary conditions to ensure a unique solution. = 20 N. In this spring system. and a unique configuration is formed. 0. we have 300 A= ( ~00 200 250 50 50 ~ ). Therefore.Finite Element 1\lethod 17. 0. 231 . which is four in this case. we now have only three unknown displacements u 2 . 0. Thus. and k3 = 50N/m. k2 = 200 N/m.1 Concept of Elements and (17. the first equation for Ut becomes redundant and we can now delete it so that the reduced stiffness matrix A is a 3 x 3 matrix. In our case where there are no applied forces at nodes 2 and 3. 20).16) 0 Example 11. the matrix K is singular or its rank is less than the total number of degrees of freedom.
1 Concept of Elements Finite Element Method The solution is u =A lg = ( 0. Let us calculate the force at nodes 2 and 3 to see if this is the case. The forces at the corresponding nodes are fi and fi. it is also called a truss element. the force should be 20N eve1ywhere in the spring systems since the mass of the springs is negligible.17. it cannot support bending.3 .3 m. Now we have to derive its stiffness matrix. the stress a is thus related to strain f via 232 . A bar element can only support tension and compression. we have which is the same at node 3 of element E 3 So the force is 20 N everywhere.2) 0. which is a uniform rod with a cross section area A. Young's elastic modulus E. Theoretically speaking. A more complicated but realistic elen1ent is the bar element as shown in Figure 17.2 m. For this reason. At the node 2. the extension in element E 1 is ~u = u 2 .7 Therefore.3.2 Bar and Beam Elements The spring system we discussed earlier is limited in many ways as a spring does not have any mass and its cross section is not explicitly included. The displacements at nodes i and j are Ui and Uj. 0 17.2 m.7 m. respectively. Assuming the bar is linearly elastic. thus the force at node 2 is f2 = kt~U = 100 X 0. Similarly. the displacements are u 2 = 0.1. u 3 = 0. and a length L.2 = 20N. 0. and u 4 = 0.u 1 = 0. at node 3 of element E2.
each node i has two displacements ui (along the bar direction) and vi (perpendicular to the bar direction).Atfethod 17. In 2D. (17. 1 (17. If we rotate the bar elen1ent by an angle (} as shown in Figure 17.Finite Element .Ui is the extension or elongation of the bar element. Thus.L j u· t u· J Figure 17. y) to the local coordinates (x'.19) A. and each node has only one displacement (one degree of freedom). A transformation is needed between the global coordinates (x.ui)/L and a= f /A where F is the force in the bar element.17) where ~u = Uj . each node has two degrees of freedom.3: Bar element. Since f = (uj. we have f = y(~u) EA = k(~u).4. From the geometrical consideration.E. We have up to now only discussed 1D systems where all displacements Ui or Uj are along the bar direction. y'). We now extend to study 2D syste1ns.18) Therefore.1 Concept of Elements a= EE. the stiffness matrix K for this bar becomes K = ( k k ) = EA ( 1 k k L 1 1 ) . the global displacements ui and vi 233 . we cannot use the standard addition to assemble the system. Now the equivalent spring stiffness constant is k = EA L" (17.
20) Figure 17.21) cosO sinO 0 sinO cosO 0 R== ( 0 cosO 0 0 sinO cosO 0 The same applies to transform the force. f' = Rf.4: 2D transformation of coordinates. (17. we get the U= I (~) ro ( ~o sinO 0 cosO 0 0 cosO sinO 0 COS si~O) (:} 0 Vj which can be written as u' where = Ru. Using the similar transformation for Uj and transformation for the twonode bar ele1nent Vj. ). and (usually) ( ~~ ) = ( y ~::o :: )(~.17. (17. si~O} (17.22) (17.1 Concept of Elements Finite Element Metbod u~ at node i are related to the local displacement v~ = 0.23) 234 .
(17. each node has three degrees of freedom (Ui. often linear.1 Finite Element Formulation Weak Formulation l\llany problems are modelled in terms of partial differential equations. 17.26) (17. by combining the above two equations. and the stiffness matrix is therefore a 6 x 6 matrix in 2D. In this case.24) As the calculation is n1ainly based on the global coordinates. (17.u) = 0 for X E {)!lE and natural boundary conditions B(u) = 0 for X E anN.Finite Element 1\lethod 17. For more complicated elements. and the assembly should be done by transforming the local systems to the global coordinates. X E !l.C is a differential operator.27) The stiffness matrix K is a 4 x 4 matrix in 2D.2 Finite Element Formulation and the stiffness matrix in local coordinates is K'u' = f'.28) where .2. For bending. Vi.C(u) = 0. (17. we have K'Ru=Rf. Bar elements can only elongate or shrink. they do not support bending or deflection. which can generally be written as . it is necessary to use a formal approach in terms of shape functions and weak formulations. Oi). Assuming that the solution can 235 . we need the beam elements which include a rotation around the end nodes ()i and ()i.25) or R.1 K'Ru = Ku = f. which is equivalent to a global stiffness matrix (17.2 17. This problem is usually completed with the essential boundary condition £(u) = (u.
2 Finite Element Formulation Finite Element !v!etbod be approximated by uh over a finite element mesh with an averaged element size or mean distance h between two adjacent nodes.2.. Thus.31) i=l j=l it requires that Ni = 0 on 8f2E so that Wi = 0 on {}O. we have (17.2 Galerkin Method There are many different ways to choose the test functions Wi and shape functions Ni.. . M). integrating over the domain and using associated boundary conditions. One of the most popular methods is the Galerkin method where the test functions are the same as the shape functions. In this special case. the formulation simply becomes (17.29) Multiplying both sides of the equation by a test function or a proper weighting function. If we choose wi = wi so as to simplify the formulation. In addition. the above equation can be approximated as (17. If we can approximate the solution uh by the expansion uh(u.32) 17 . there is no much limitation on the choice of Wi and wi. or Wi = Ni. we can write the general weak formulation of Zienkiewicztype as fn C( uh)widn (17. (17.17. 2. . lanN lanE M M where (i = 1. t) =L ui(t)Ni(x) =L UjlVj.33) 236 .E.30) + f B(uh)widr + f &(uh)widrE ~ o. only the natural boundary conditions are included since the essential boundary conditions are automatically satisfied.
For the simplest 1D element with two nodes i and j.Xi)dn = f(xi)· (17.Xi) = 0 at 17.Xi) = 1 at x = Xi and 8(x.. if we use the Dirac delta function as the test functions Wi = 8(x . we have Ni = 1. t) = L Ui(t)Ni(x) i=l (17.34) together with 8(x. . 237 . . trial functions or more often shape functions in the literature of finite element methods.. which is shown in Figure 17. the method is called the collocation method which uses the interesting properties of the Dirac function fo xfXi· f(x)8(x. M uh(x.Finite Element 1\lethod 17.~). At any node i. 2.5. and Ni = 0 at any other nodes.35) where ui are unknown coefficients or the value of u at the discrete nodal point i.xil.37) where L = lxi.2. i=l Ni(Xj) = 8ij· (17. On the other hand.2 Finite Element Formulation The discretization of this equation will usually lead to an algebraic matrix equation. Functions Ni (i = 1. M) are linearly independent functions that vanish on the part of the essential boundary.3 Shape Functions The n1ain aim of the finite elen1ent method is to find an approxinlate solution uh(x. t) for the exact solution u on some nodal points. or M LNi = 1.36) The functions Ni (x) are referred to as basis functions. the linear shape functions can be written as (17.
various methods can be formulated. Depending on the scheme of minimization and the choice of shape functions. u~~=~ui .2 Finite Element Formulation Finite Element !v!etbod X j ~ = xflxi xil Figure 17.5: The 1D linear shape functions. X i=l 8u M M . . so we have (17. These include the weighted residual method.38) i=l where we have used the notations: 1 = dj dx and· = Higher order derivatives are then calculated in a similar way. The ultimate goal is to construct a method of computing ui such that the error uh . the derivatives of u can be approximated as 8uh """' 1 a~ fJx = ~ Ui(t)N (x). the method of least squares. OUh """' vt N( x. 238 3t.39) There are several methods to minimize R. ) (17. the Galerkin method and others. the residual R varies with space and time.u is minimized. Generally speaking. Using the assumptions that ui(t) does not depend on space and lVi(x) does not depend on time.17.
ay::.3.( .1 Elasticity Plane Stress and Plane Strain The stress tensor u and strain tensor e are not written as tensor forms but vector forms u = (axx. but azz =/= 0). In the 2D case. The strain tensor is usually defined as 1 oui 8u· cij = .41) where Dis a 6 x 6 symmetrical matrix as functions of Young's modulus E and Poisson's ratio v. The strains are given by f. Two special cases that are commonly found in many applications are the plane stress (a zz = 0.x=.o). "fxy.yy. The commonly used formulation is the displacementbased formulation or ubased formulation. au {)y 8v OX (17.. If there is no such change. a:.xx' f. a::x)T. (17.:. ayy.42) Now the stressstrain relationship becomes a= D(f. =/= 0) and plane strain (f.3 17. axy. (17. f.44) where f. then the initial strain can be taken to be zero in most applications. and € = (f.40) 2 8xj 8xi where one applies the engineering shear strain f.z:.43) where D is a 3 x 3 matrix. v)T and the strain f.+ _ J ). 239 .xy = 2cxy· Hooke's elasticity can be expressed as (17.::. and stress a are defined as (17.Finite Element 1\lethod 17.y= au 8v ox 8y ' l'xy=+.:. f. 'Yyz.:::: = 0.o is the initial strain due to temperature change or thermal loading.3 Elasticity 17.f. the displacen1ent u = (u. but f.. 'Yzx)T.
Ym). /y = 0 and = Txy = 0. we have (I7.6 (xi. Yi).49) 240 . so the solution technique for the ID heat transfer shall equally apply. v) in a plane eletnent can be interpolated from nodal displacements.I. using a triangular element (i. and (xm.3 Elasticity Finite Element Metbod The equilibrium of force in elasticity leads to V' · u+ b = 0.N. (I7. Therefore. For example. v x (I7. we shall focus on the 2D case in the rest of this section. Displacements (u.v)/2 (I7.Nml] =Nd. for the ID case plane stress when v = 0.45) /y]T is the body force.17. Yi).48) where we have used the stressstrain relationship.j.v 2 I. (xj. (I7. (I .2v) ( l~v l~v ~~~ 1v i~v v 0 0 (12v) 2(1+v) ) . where b = [fx stress.I. In the case of plane E ( I D=. we have E D = (I . This ID equation is essentially the same as the ID heat transfer equation u" + Q = 0 that will be discussed in detail later.46) In the case of plane strain.v 0 v I 0 0 ) 0 . we have u = (:) = (N. m) with three nodal points as shown in Figure I7. the equation of force balance simply becomes E 82u I2 8 2 + fx = 0.47) uy Clearly.
241 (17. (17..53) where B =LdN.50) and N = ( Ni 0 0 Ni N· J 0 (17.54) Now the equation (17.55) where K = In BTDBdV.52) 'ljX we can rewrite the above forn1ulation as (17. where I is a 2 x 2 unitary matrix. i.6: Schematic triangular mesh and the layout of a triangular element.3 Elasticity X Figure 17.Finite Element 1\lethod y 17.51) By defining a differential operator Ld= fx iu.45) becomes Ku=f. (17.56) .e. (17. 0 ) 0 ( {) w i)y (17.
each node should be identified in some way.j.57) where r is the surface traction (force per unit area).3.. bm (17.60) For a triangular element with three nodes (i.1 x· J 2 1 Xm Yi Yi Ym (17.2 Implementation In the case of 2D elastic problems. if each element has r nodes.17. and each node has n degrees of freedom. each node has two degrees of freedom (ui.3 Elasticity Finite Element Metbod and (17. 17. m).58) where 1 Xi 1 ~=. 6) of each triangular element can be expressed as (17. If the region has M nodes in total.j = 1. and most often in terms of the index matrix. then the stiffness matrix K(e) for each element is a 6 x 6 matrix. . thus we have Bi = LdNi. then M n equations are needed for this problem. Vi). the simplest elen1ents are linear triangular elements. In general. or B = LdN = 21 ~ c. The 242 . The linear elen1ent implies that the strain is constant in the element. The stiffness matrix K~j) (i. In order to calculate the contribution of each element to the overall (global) equation. . 0 0 Ci b· J 0 c· J 0 c· J bm 0 Cm c· l b· l b· J ~). 2. For the present case (n = 2).59) is the area of the triangular element. then the local stiffness matrix is a rn x rn matrix.. we need 2/v/ equations for plane stress and plane strain.
etc).J)' (17. m) can be written as (i.. By writing it as two index matrices (ID = JDT). in the nodal index matrix. . the main assembly of the stiffness matrix is about the corresponding equations and the application of the boundary conditions. 2i 1 2i 1 2i 1 2i 1 2i 1 2i 1 2i 2i 2i 2i 2i 2i 2j 2j 2j 2j 2j 2j 1 1 1 1 1 1 2j 2j 2j 2j 2j 2j 2m1 2m 2m1 2m 2m1 2m 2m1 2m 2m1 2m 2m1 2m JDequ = (17. ( .. i) (j.e. m) (m.Finite Element 1\lethod 17.2j)) (2i. for each entry in the stiffness matrix.J)] K (J Dequ(I. 2j) ' etc (17.JDequ(/. i.2.j) (i. we now have IDequ = JD~u. (17. 2m). equation 2j .3 Elasticity nodal index matrix for three nodes (i.J). 2j .J).J Dequ(/.. equations 2i 1 and 2i for nodal i.61) The nodal index matrix identifies the related nodes in the global nodenumbering system. .2j 1) (2i 1. etc. thus we need to transfer the nodal index matrix to the equation index matrix in terms of the global numbering of equations with two degrees of freedom for each node (e..1) (2i.J = 1.. and each entry is a pair such as (2i. 243 .62) The equation index matrix now has 6 x 6 entries. Thus.j) (m.6. i) (i... j. 2j.J)] Kij is + K(e) (/. m) ) IDnode = ( (m.1). .j) (j. i) (j.1 and 2j for node j. (2m.j).64) I. . ) Z. we now have four entries. m) . However.1.g.J ~ ( (2i 1.63) So that the contribution of K~) to the global matrix sin1ply K (J Dequ(/. say (i.
(17. (17.17. we have f [V · (kVu) + Q]1VidO.66) and the natural boundary condition (17.f [k 8u.4 Heat Conduction Finite Element Metbod Similarly.65) with the essential boundary condition u =u. x E anE.q]lVidr = ln lanN 8n 0. 2m.1. (17.1.67) Using the formulation similar to the formulation (17. 17 . Thus. 2j. we have 244 .1. 2m etc.4.68) Integrating by parts and using Green's theorem. 2j. finite elen1ent methods are especially useful in this case. 17.33) in terms of u ~ uh. 2i..1 Basic Formulation The steadystate heat transfer is governed by the heat conduction equation or Poisson's equation V · (kVu) + Q = 0. the contribution of the body force and external force can be computed J(l) = f(l) + j(e)(l) where l = 2i.4 Heat Conduction Heat transfer problems are very common in engineering and the geometry in most applications is irregular.
Kii = In k"V Ni "V lVjdO.. we have  lanN [ qNidr = 0. (i.70) Substituting uh = Ej\1~ 1 UjlVj(x) into the equation. !M ). That is. u"(x) + Q(x) = 0.4 Heat Conduction + f Since lVi = ln Q1Vidn f [k 8uh lanN 8n  q]JVidr = o. and (ft. KU = f. we consider the 1D steadystate heat conduction problem. . (17.E.75) fi = f QNidn + f qNidr. .Finite Element !vletbod 17. (17. ln lanN 0 Example 17.73) where K cT = = [Kij]. . . u'(l) = q. 2. uT = (ut. .74) (17.. UM ).j = 1.72) This can be written in the compact matrix form L KijUj = j=l M fi. !2.2(a): As a simple example. lan lanN Therefore..69) 0 on {)O. 245 witb boundazy conditions u(O) = /3. (17. thus we have f [ ]Nidr = f [ ]Nidr. .... .. U2. the above weak formulation becomes (17. M). (17.
1. . 75) becomes For the purpose of demonstrating the implementation procedure.1 + q)x +rex. m).r) + (re. i + 1. This will be discussed later in more details.75) is the summation of the integral over the whole solution don1ain. Each element contributes to the whole stiffness matrix. In finite element analysis.4. 2~ .17.j =I= 0 only for j = i .j =I= 0 if and only if (or iff) nodes i and j belong to the same elements. 77) and fi = L fi(e). In the 1D case~ Ki. assembly of the stiffness matrix can be done in an elen1entbyelement manner. The stiffness matrix Kin equations (17.2 ElementbyElement Assembly The assembly of the linear matrix system is the popular elementbyelement method. and in fact~ its contribution is a pure number.76) Then equation (17. Furthermore..4 Heat Conduction Finite Element Method For a special case Q(x) = rexp(x). we have the analytical solution u(x) = ({3. and the don1ain is now divided into m elements with each element on a subdomain ne (e = 1. The elementbyelement formulation can be written as (17.. D 17. Thus. e=l m 246 .73) and ( 17. Ki. let us solve this problem by dividing the interval into 4 elements and 5 nodes. the shape functions Nj are typically localized functions~ thus the 1natrix K is usually sparse in most cases. i. (17.
2(b): Kljl = [f kN. 2. For each element with i and j nodes. The assembly of the global system matrix for the example with 4 elements and five nodes is shown below.4 Heat Conduction In addition. these can extend to all nodes (with hi = xi+l xi. the nonzero contribution of each element to the global system matrix K is simply assembled by direct addition to the corresponding global entry (of the stiffness n1atrix) of the corresponding nodes or related equations.Finite Element !vletbod 17.Njdx] = :. In reality. for example in elements 1 and 2. 3. this can be easily done using an index matrix to trace the element contribution to the global system matrix. i = 1. k/ht k/ht K<'l = ( ~ k/ht k/ht 0 0 0 0 0 0 0 0 0 0 0 0 0 and so on. Now the global system matrix becomes K= 247 . ( ! 1) ' 1 1 fz~ e) = Q he ( 1 ) 2 1 ' so that. we have L= he. 4). the integration of each elen1ent can be done using the local coordinates and local node numbers or any coordinate system for the convenience of integration over an element. 0 Example 17. since the contribution of each element is a sinlple number. Then.
The natural boundary conditions are included in the integration in the finite elen1ent equations such as (17. On the other hand. To show how these methods work. Thus no further in1position is necessary. These include the displacement.75). rotation. 1] to aid our discussion. b) Lagrangian n1ultiplier and c) penalty method./2 Q(h2 + h3)/2 .. . 248 . it is u(O) = /3. Q(h3 + h4)/2 Qh4/2 + q where the last row of I has already included the natural bounda1y D condition at u' ( 1) = q.4.17. we use the 1D poisson equation on the distinct points xi(i = 1. and moment. although the essential boundary conditions are automatically satisfied in the finite element formulations. they are also called the geometric boundary conditions. and known value of the solution. it is u'(1) = q.. they still need to be implemented in the assembled finite element equations to ensure unique solutions. heat flux. Application of Boundary Conditions Boundary conditions can be essential. Sometimes. In our example. natural or mixed. force. 2.4 Heat Conduction Finite Element Method ( k/ht k/hl 0 0 0 k/ht k/hl + k/h2 k/h2 0 0 0 k/h2 k/h2 + k/h3 k/h3 0 0 0 k/h3 k/h3 + k/h4 k/h4 0 0 0 k/h4 k/h4 } U=U~} 17.3 I= ( Q(hl + h2)/2 ) Qh. The essential boundary conditions are automatically satisfied in the finite elen1ent formulation by the approxin1ate solution. The imposition of the essential boundary conditions can be done in main several ways: a) direct application. Natural boundary conditions often involve the first derivatives such as strains. In our example. 1\11) E (0. . Thus. they are also referred to as force boundary conditions.
For example. 73) becon1es (0.. . Now we have (17. the boundary conditions u(O) = a and u(M) = (3 in the 1D case mean that the first and last equations are replaced by u1 =a and UM = {3. !M = /3 with Ki\111 = . respectively. and apply directly the essential boundary conditions at point i to replace the corresponding ith equation with ui = ui so that ith row of the stiffness matrix K in equation (17. . 0.. .. .4 Heat Conduction Direct Application In this method. we simply use the expansion uh = Et! 1 uilVi. This method is widely used due to its simplicity and the advantage of timestepping because it allows bigger time steps. and KMM = 1. .•• . we have where ei = (0. 249 . Thus. . the equations can be solved for (ub u 2 . Lagrangian Multiplier This method is often used in the structure and solid mechanics to enforce the constraints (Ui = Ui).80) Because 8u and 8X are arbitrary. . The variation is added by the extra term . ... Ku = 1. 1.. .. uM )T.. 0) and the corresponding /i = f(i) = ui.Finite Element 1\lethod 17. Then.ui) where A is the Lagrange multiplier.M1 = 0)... O)T (its ith entry is equal to one). All other points will be done in the similar manner. 0. 1.. It =a (all other coefficients are set to be zeros: K12 = . = K1M = 0. = KM. This method can be extended tom Lagrangian multipliers.79) whose variation 8II = 0 leads to (17. 0.X(ui..
we now use the direct application method for the essential boundary conditions. KMM = KMM+/./ = fM + '"'(/3. so that the first row becomes K 1j = (1 0 0 0 0) and f 1 = (3. lVe can replace the first equation E. and It =It+ /O::. and thus very small time steps are required for convergence. For the case of k = 1. = h 4 = 0. Thus.25..4 Heat Conduction Finite Element Method Penalty Method One of the most widely used methods of enforcing the essential boundary conditions is the socalled penalty method in terms of a very large coefficient I' Ui = ft at Xi E an E' so that /Ui = '")'U can be directly added onto Ku =f.Q = 1.h 1 = . it may affect the efficiency of timestepping since it increases the maximum eigenvalue of the stiffness matrix. we have K= 1 k/ht 0 0 0 0 0 k/h2 k/h2 + k/h3 k/h3 0 0 0 k/h3 k/h3 + k/h4 k/h4 0 0 0 k/h4 k/h4 ( k/ht + k/h2 k/h2 0 0 } U= ( ~~} ( I= 250 Q(ht + h2)/2 Q(h2 + hJ)/2 Q(h3 + h4)/2 Qh4/2 + q ~ ) .. f3 = 1 and . However. The penalty method is widely used in steadystate problems.17. In the 1D example. The disadvantage is that the conditions are only satisfied approximately.2(c): Following the same example of the 1D steady state beat conduction discussed earlier.=l K 1iui = f 1 with u 1 = (3. it simply leads to Ku = Ku +!'. fA. The advantage of the penalty method is that the handling of the essential boundary conditions becomes simpler from the implementation point of view. 0 Example 17. '"Y ~ lOOOmax IKiil should be adequate. The common rule for choosing I' is that '"'f >> max IKiil· Usually.
81) The boundary conditions and initial conditions are u(x. one simple and yet instructive way to extend the finite element forn1ulation to include the time dimension is to replace Q in equation (17.375 Hence. 17.00 0. and kg~.1 J = 1.Finite Element !vletbod 17. However.GUt.q = O.50 . Thus. (17.x E {)O. tbe solution is U = K./3il so that we have V' · (kV'u) + (Q.25 D 17. ( 0.25.5 TimeDependent Problems q = 0. most realistic problems involve time. and thus we will now discuss the tin1edependent problems.65) with Q.f3utt) = 0.N.GU . and the solutions obtained are the steadystate solutions.25 0.72 ) 0. 0) = 4J(x). u = u.x E {)O.25 0.25 ) 1 0.5 TimeDependent Problems The problems we have discussed so far in our finite element analysis are not timedependent. Using 251 .5.1 The Time Dimension As the weak formulation uses the Green theorem that involves the spatial derivatives.34 0. we bave 0 ( K= 0 4 ~ 1 8 4 8 0 0 4 0 0 4 8 4 0 4 4 4 ~} f = ( 0. the time derivatives can be considered as the source term.GUt/3Utt = Q .E.
On the other hand. the matrix is usually singular. which may imply many solutions. that is to say. M = [Mii] is the general mass matrix due to a similar role acting as an equivalent mass in dynamics. for the 1D elen1ents (with nodes i and j). C are symmetric. For example. 'V Ni · k · 'VNi = k'VNi'VNi..83) Kii =In [(k'V Ni 'V Ni )]dO.85) fi and = ln [ NiQdn + [ Niqdr. we have 'Lrj (k'VN/vlVj)dn] + i=l J[(NiaNi)ui + (Ni/3Ni)uj]dn L i=l n n M M . Ni and Nj (i. the stiffness matrix will be nonsingular. (17. Kij = Kji.17. Mij = Mji.86) The matrices K. Cij = Cji due to the interchangeability of the orders in the product of the integrand k.e. The matrix C = [Cij) is the damping matrix similar to the damping coefficient of damped oscillations. thus unique solutions may be obtained. In addition. where = o.84) (17..f Niqdr ln lanN Mii+ Cu+Ku = f.82) which can be written in a compact form as (17. M.f NiQdn . M and C will be always nonsingular if they are not zero. Only after the proper boundary conditions have been enforced.5 TimeDependent Problems Finite Element Metbod integration by parts and the expansion uh = EJ~ 1 uiNi.87) 252 . Nia1Vi = 1ViaNi =aNiNi etc). (17. 1) 1 ' (17. lanN (17. before the boundary conditions are imposed.
5.92) 253 .3 1D Transient Heat Transfer In the case of heat conduction (M = 0). we have Cu+Ku=f. it reduces to the linear heat conduction. Thus.89) we see that it is an ordinary differential equation in terms of time and matrices. (17. all the timestepping methods developed in the standard finite difference method can be used for this purpose. 17.he ( 2 1 ) 6 1 2 ' det[C(e)] # 0. 17 .1.5 TimeDependent Problems but M.. C~~) = t} det[M(e)] # 0.90) M (~t)2 (17.89) becomes (un+1 _ 2un + un1) (17. (un+1 _ 2un + un1) u= (~t)2 so that equation (17.88) Clearly.91) Now the aim is to express un+ 1 in tenns of un and un. it becomes the wave equation with the source term. If C = 0.2 TimeStepping Fron1 the general governing equation Mii+ Cu+Ku = f. in principle. (17.~) = /3~e ( ~ ~ ) .Finite Element .5.Atfethod 17. if M = 0. (17. we have . o. For a simple center difference scheme.
25 0 0 0. we OUt = kuxx + Q. 0 0 0. we have 0.17.25 0 1 0 ) 0. The formulation with 5 nodes and 4 elements leads to C= ( 2h. we have Mii+Ku=f.9: have For a transient heat conduction problem. t) = 1.u n K U. and u(x.25 0.t = ct (f.25. 254 (17. ) (17. = h 4 = 0.Q = 1.25 0 0 ( 0.94) so that we have (17.25 0. we can write it as u n+l tl.25 0 1 0.5 0 0 17.4 Wave Equation For the wave equation (C = 0). u'(l) = q.25 C= 1 0 .93) Using the explicit timestepping method.5 0 0.96) .. c.25 0. ht ~ 0 6 0 0 ht 2(ht + h2) 0 h2 2(h2 + ha) 0 0 h2 0 0 ha 2(ha + h4) 0 0 ) 0 ha 0 h4 2h4 h4 For the case of a= 6.95) 0 Example 17.k = 1.1 (f ..5. 0) = 0. u(O.h 1 = .K n) u ' (17.5 TimeDependent Problems Finite Element Method or U= .
(17. (17. 255 . the 1D wave equation 82u 82u (17. Although these methods can in principle be extended to nonlinear problems. For example.5 TimeDependent Problems Using un+1 .99) 8t2 = c {)x2' with the boundary conditions u(O) = u(1) = 0. however.2un + un1 u = (~t)2 . The finite element methods in this book are mainly for linear partial differential equations.97) we have where I is an identity or unit matrix. (17. f = 0. The interested readers can refer to many excellent books on these topics. 1 Kii = l cNfNjdx. 0) = e(x.101) and u 0 is derived from the u(x. some degrees of approximations and linearization are needed. 0) = exp[(x.1/ 2 ) 2 .Finite Element 1\lethod 17.100) Mii = fo NiNjdx. an iterative procedure is required to solve the resultant nonlinear n1atrix equations..1/2) 2 ]. can be written as u(x. In addition.
.
1 18. Before we proceed to study the nonlinear system.1. However. From the numerical point of view. l\1athematically speaking. 18. We know that the generalized forn1s of parabolic equations are nonlinear reactiondiffusion equations. the finite difference scheme should still be useful for most nonlinear equations though they should be implemented more carefully.1 Heat Conduction Equation Fundamental Solutions From the similarity solution in section 10.Chapter 18 Reaction Diffusion System The partial differential equations we solved in the previous chapters using the three major numerical methods are linear equations. we know that both diffusion equation and heat conduction equation may mathe257 . nonlinear equations are far more difficult to analyse if it is not impossible. some extra linearization and approximations should be used for the nonlinear terms. let us review the fundamental characteristics of linear parabolic equations by solving the linear heat conduction equations.3.
(18. In engineering. the coefficient k has the unit [n1] 2 /[s]. Any two combinations of x and kt giving the same ( will have similar solutions. If the domain is semiinfinite or infinite so that x 0 ~ 0. which means the variable ( is dimensionless. then we get X (18. .3) where C is an integration constant. we have For the heat conduction in a semiinfinite au= k8 2 u 8t with the boundazy condition U= and an initial condition UQ. v4kt 0 Example 18.2) Integrating it once. Using the similarity variable (. Integrating it again. or (ln !')' = 2(.1: domain. 8x 2 ' ~ (X 0). the heat conduction equation can be transformed into an ordinary differential equation !"(() = 2(/'. thus the unit of kt is [1n] 2.5) =Aerf( ~)+B. we have xo 2 u = C x e< d( + D.1 Heat Conduction Equation Reaction Diffusion System matically have a similarity variable defined by (= ym' X (18. (x > 0) 258 at t = 0.18.4) 1 which is the general solution of the heat conduction equation.1) where k is either the thermal diffusivity or diffusion coefficient. u = 0. (18. we have (18.
+ erf]. (18. For x ~ oo. its two surfaces are held at constant temperatures with the boundary conditions u = u 0 at x = 0 for (t ~ 0). For example in heat conduction through a plane sheet with zero initial temperature.2. The solution is u = uo[l . then we have uo hx h+x u = [erf. 7) 259 . 18..2 18. X At x = 0. erf(O) = 0..6) which is a slowly convergent series.Reaction Diffusion System The general solution is u = Aerf( /"A"U) v4kt X 18. n1rx = uo (1 .. h). and u = 0 at x = L for (t ~ 0). If the domain has a finite length. we get B = uo.. then the solutions often consist of power series or even special functions. L 1T' n=l n L (18. 17'i:!')] = uoerfc r:rr:r· v4kt v4kt However.1 Nonlinear Equations Travelling Wave The nonlinear reactiondiffusion equation 8u 8t = D {)x 2 82 u + f(u).L.L2 . 2 J4kt J4kt 0 X The solutions of heat conduction do not always involve the error function because error functions only occur when the integration involves semiinfinite or infinite domains.2 Nonlinear Equations + B. uo sm e kn21r2t. erf(x/ . we have A+B = 0.erf(. if u is constant (u = uo) in the initial region x E [h.. The general solution can be written as u 2~ ./4kl) ~ 1.x) +.
chemical and biological phenomena. f(u) = ')'u(1 .11) {)t = D( ax2 + {)y2) + '"YQ(u).2 Pattern Formation One of the most studied nonlinear reactiondiffusion equations in the 2D case is the KolmogorovPetrovskiiPiskunov (KPP) equation {)u fPu 8 2 u (18.u). we have Du"(() + vu'(() + f(u(()) = 0. and q(u) = u(1 . (18.vt. 1].10) 18. (18. n) where n is the size of the grid.12) The KPP equation can describe a huge number of physical. We can solve it using the finite difference scheme by applying the periodic boundary conditions and using a random initial condition u = random( n. By assuming that the travelling wave solution has the form u( () and ( = x . The most interesting feature of this nonlinear equation is its ability of generating beautiful patterns. 260 . (18. For example.9) The KPP theory suggests that the limit of the speed of the travelling wave satisfies v 2?: 2JDf'(O). u( oo) + 0.u) satisfies these conditions. and the equation in this special case is called the KolmogorovPetrovskiiPiskunov (KPP) equation. f(u) > 0 for u E [0.8) This is a secondorder ordinary differential equation that can be solved with the appropriate boundary conditions u( oo) + 1.2. (18.18. and substituting into the above equation.2 Nonlinear Equations Reaction Diffusion System can have the travelling wave solution under appropriate conditions of f(O) = f(1) = 0. and f'(O) > 0.
end 261 % line 1 % line 2 .5.2 and 'Y = 0. y) can be easily demonstrated.Reaction Diffusion System 18.2 Nonlinear Equations 1 0. The landscape surface shows the variations in the location and values of the field u(x. time=100. The following simple 15line Matlab program can be used to solve this nonlinear system.1 shows the pattern formation of the above equation on a 100 x 100 grid for D = 0. % Pattern formation: a 15 line matlab program % POE form: u_t=O*(u_{xx}+u_{yy})+gamma*q(u) %where q(u)='u.5 0 200 0 0 200 Figure 18.5. % The solution of this POE is obtained by the % finite difference method. We can see that rings and thin curves are formed. arising from the random initial condition.2 and 'Y = 0. Figure 18.1: 2D pattern formation forD= 0. assuming dx=dy=dt=1 % Written by X S Yang (Cambridge University) % Usage: pattern(100) or simply >pattern function pattern(time) % Input number of time steps if nargin<1.*(1u)'.
2 Nonlinear Equations Reaction Diffusion System % Initialize parameters n=100.*(1u).End of Pragram  % line 13 % line 14 % line 15 If you use this program to do the simulations. % line 3 % line 4 % line 5 % Time stepping for step=1:time. % 0=0. gamma=0.J1)+u(I.J). grad=u*O. % line 11 end % line 12 % plot as a surface surf(u). %Set initial values of u randomly u=rand(n. J)+u(I+1. J = 2:n1. % line 9 % Coloring and colorbar colorbar. one can see beautiful and stable patterns.j) and the loop I = 2:n1. n=200. In this sense.J+1)+u(I1.5. u ={14*0)*u+O*grad+gamma*u. 0=0. shading interp.n). shading interp.2.18. % . but the characteristics such as the shape and structure of the patterns do not change much with time. gamma=0.J)= u(I. view( [ 25 70]). % line 10 drawnow. colormap jet. it evolves gradually with time. %line 8 % Show results pcolor(u).2.5. Once the pattern is formed. you will see that the pattern emerges naturally from the initially random background. 262 . % line 6 % Laplace gradient of the equation % line 7 grad(I. % Index for u(i.
then 1/J satisfies 81/.. vo = 0. and the variables are coupled in a complicated manner. The steady state solutions are obtained from f(uo.15) (18. b. They are uo b~ = 2a [V1 + 4/. c.uo..1]. we often have to simulate a system of nonlinear reactiondiffusion equations. at 263 .vo) be the small perturbation. vo) = 0.cv 1 +u+v .. the following system consists of two nonlinear equations (18.17) Let 1/J = (u. This reaction diffusion system may have instability if certain conditions are met. d are all constants. For example.3 ReactionDiffusion System 18.14) and f(u.18) ~ ~.u) g(u. while a.3 ReactionDiffusion System The pattern formation in the previous section arises naturally from a single equation of nonlinear reactiondiffusion type. = D\12. In order to show this.v) = bu.Reaction Diffusion System 18.13) (18. + M"'' (18.16) vd 1+u+v' where Du and Dv are diffusion coefficients. let us use the following mathematical model for enzyme inhibition and cooperativity. (18. vo) = 0 and g(uo. v . v) = au(1 . (18. The pattern formation in the previous section comes from the instability of the nonlinear reaction diffusion system.2 . In many applications.
24) with 8 _ (2auo +b) . beautiful patterns also arise naturally in the following nonlinear system 8u 8t 8v 8t = Da ( 8x2 82 u + 8y2 ) + 'Yf (u.26) = 82v 82v _ Db( 8x2 + 8y2) + f3g(u.18.20) Writing 'l/J in the form of (18.22) where I is a 2 x 2 unity matrix.3 ReactionDiffusion System Reaction Diffusion System where (18.21) where the sumn1ation is over all the wavenumbers k.< ~:~ (18. Similar to the nonlinear KPP equation (18. v)' 82 u  (18.19) and M = 1 (1 + uo) ( (2au0 +b) a(1 duo)+ c ) . 0 (18.[dDu. v). any small random perturbation can generate complex patterns. This eigenvalue equation has two roots. 264 (18. this requires that Dv d . Since ~(A) > 0 implies that instability.25) If the unstable criteria are satisfied.Dv(2auo + b)]2 · (18.27) .23) Du (2auo +b) The range of unstable wavenumbers between the two roots of k2 at the bifurcation point is given by k 2 = dDu .Dv(2auo +b) [1 ± ± 2DuDv ( 1 + Uo) J1 +4 D D 8] u v ' (18. we have (18.12).
these equations can be used to sin1ulate the pattern formation in a wide range of applications where nonlinear reactiondiffusion equations are concerned.2.u 2 v.2.3 ReactionDiffusion System and ](u. g(u. Db = 0. 265 . v) = u.28) for the values Da = 0.u). v). (18. With different functions j (u. v) and g( u.5 and {3 = 0. v) = u(1.1.Reaction Diffusion System 18. "( = 0.
Chapter 19
Probability and Statistics
All the mathematical models and differential equations we have discussed so far are deterministic systems in the sense that given accurate initial and boundary conditions, the solutions of the system can be determined (the only exception is the chaotic system to a certain degree). There is no intrinsic randomness in the differential equations. In reality, randomness occurs everywhere, and not all models are deterministic. In fact, it is necessary to use stochastic models and sometimes the only sensible models are stochastic descriptions. In these cases, we have to deal with probability and statistics.
19.1
19.1.1
Probability
Randomness and Probability
Randomness such as rouletterolling and noise arises from the lack of information, or incomplete knowledge of reality. It can also come from the intrinsic complexity, diversity and perturbations of the system. The theory of probability is mainly the studies of random phenomena so as to find nonrandom regularity.
267
19.1 Probability
Probability and Statistics
For an experiment or trial such as rolling dices whose outcome depends on chance, the sample space n of the experiment is the set of all possible outcomes. The sample space can be either finite or infinite. For exan1ple, rolling a sixsided die will have six different outcomes, thus the san1ple space is n = {1, 2, 3, 4, 5, 6}. The elements of a sample space are the outcomes, and each subset of a sample space is called an event. For example, the event S = {2, 4, 6} is a subset of n. In a sample space n, the outcomes of an experiment is represented as numbers ( 1 for heads and 0 for tails for tossing coins). A realvalued variable that is defined for all the possible outcomes is referred to as a random variable, which is a function that associates a unique numerical value with every outcome of an experiment, and its actual value varies from trial to trial as the experiment is repeated. The values of a random variable can be discrete (such as 1 to 6 in rolling a single die) or continuous (such as the level of noise). If a random variable only takes discrete values, it is called a discrete random variable. If its values are continuous, then it is called a continuous random variable.
Figure 19.1: Venn Diagrams: An Band AU B. Two events A and B can have various relationships and these can be represented by Venn diagrams as shown in Figure 19.1. The intersection An B of two events means the outcome of the random experiments belongs to both A and B, and it is the case of 'A AND B'. If no event or outcome belongs to the intersection, that is An B = 0, we say these two events are
268
Probability and Statistics
19.1 Probability
mutually exclusive or disjoint. The union A U B denotes the outcome belongs to either A or B or both, and this means the case of 'A ORB'. The complement A = n  A (or not A) of the event A is the set of outcomes that do not belong to A but in the sample space n (see Figure 19.2). The A B means the outcomes in A only.
Figure 19.2: Venn Diagrams:
A = n  A and A  B.
Probability P is a number or an expected frequency assigned to an event A that indicates how likely the event will occur when a random experiment is performed. This probability is often written as P( A) to show that the probability P is associated with event A. For a large number of fair trials, the probability can be calculated by
P(A) = NA(number of outcomes in the event A). N n (total number of outcomes)
(19.1)
0 Example 19.1: If you tossed a coin 1000 times, the head (H) occurs 511 times and the tail (T) occurs 489 times. The probability P(H) and P(T) are
P(H) =
and
511 = 0.511, 100 489 = 0.489. 1000
P(T) =
There are three axioms of probability, and they are: Axiom I : 0
~
P(A)
~
1. 269
19.1 Probability
Probability and Statistics
Axiom II: P(f!) = 1. Axiom III: P(A U B)= P(A)
+ P(B), if
An B = 0.
The first axiom says that the probability is a number between 0 and 1 inclusive. P(A) = 0 corresponds to impossibility while P(A) = 1 corresponds to absolute certainty. The second axiom simply means that an event must occur somewhere inside the sample space. The third axiom is often called the addition rule. Since A and A are mutually exclusive (An A= 0), we have
P(A)
or
+ P(A) = P(A u A) = P(fl) = 1,
P(A) = 1  P(A),
(19.2)
(19.3)
which is usually called the NOT rule. The third axiom can be further generalized to any two events A and B
P(A U B) = P(A)
+ P(B) 
P(A n B).
(19.4)
In a special case when events Ai (i = 1, 2, ... , n) exhaust the wholesamplespacesuch that A= Uf=tAi = AtUAiU ... UAn = nand Ai n Ai = 0(i =I= j),
P(A n B)=
Since n n B = B, we also get
L P(Ai n B).
i=l
n
(19.5)
P(n n B) = P(B) =
L
i=l
n
P(Ai
n B),
(19.6)
which are the useful properties of the total probability. For example, if you randomly draw a card from a standard pack of 52 cards, what is the probability of it being a red king or a diamond with a face value being a prime number (if its
270
Probability and Statistics
19.1 Probability
face value is counted from 1 to 13). The prime numbers are 2, 3, 5, 7, 11, 13, therefore they are 6 cards that forms the primes. The possibility of event (A) of drawing a red king is P( A) = ~ = ,Js. The probability of event (B) of drawing a prime number is P(B) = 562 = 6 • As a diamond king (13) is also a prime, this means P( A n B) = ~. Therefore, the probability
:f
P(A U B) = P(A) 1
= 26
+ P(B) 
P(A n B)
+ 26
3
1 7  52 = 52.
Two events A and B are independent if the events has no influence on each other. That is to say, the occurrence of one of the events does not provide any information about whether or the other event will occur. In this case, the probability of both occuring is equal to the product of the probabilities of the two individual events P(A) and P(B) P(A n B)= P(A)
* P(B).
(19.7)
This can be easily extended to n mutually independent events Ai(i = 1, 2, ... , n). The probability of all these events happening is
n
P(L nAi) = Ilf:: 1 P(Ai) = P(Al)P(A2) · · · P(An)·
i=l
(19.8)
0 Example 19.2: The probability of drawing a king from a pack of cards (Event A), and showing an even number of rolling a sixsided die (event B) is P(A n B). We know P(A) = 4/52 = 1/13, and P(B) = 3/6 = 1/2. Since these two events are independent, the probability that both events occur is
1 1 1 P(A n B) = P(A)P(B) =  ·  = 13 2 26.
0
If the two events are not independent, then one may affect
the other event, in this case, we are dealing with the conditional probability which will be discussed later in the next section.
271
as the event of selecting each thing to wear is in a similar manner of putting it into slots in successive stages. the total number of all possible ways is simply the multiplication of all the possible choices for each stages. n objects can in general be pern1utated . . the total number of combinations (of selecting 5 students) is 20 * 19 * 18 * 11 * 15 = 201 = 15504 (19. the final formed team is the same. 820 ). Inn. there are 5 places to line up the students. C. Suppose you have 5 pairs of shoes. there are 5 possible choices. 4 pairs of trousers.10) . As before.19. Thus the number of all possible permutations is 5 x 4 x 3 x 2 x 1 =51. There are 51 permutations within the same tea1n. Suppose there are n = 20 students in a class (named 81. we want to select 5 students at random to fonn a 5student team to work on a project. In this case. Following this line of reasoning. 819. and the total number of all permutations for selecting 5 students is 20 * 19 * 18 * 17 * 16. In order to line 5 objects marked A. Therefore. the second place has only 4 options. D.I ways. 8 2. 81.. the fourth place has 2 choices. This is equivalent to the lineup problem from your feet to your head. This is different from the lineup problem because once you have selected any five students (say) 8~.9) 51 5!15! . the total number of all possible combinations of selecting k objects from n is 20 nck c5 = = ( ~ ) = kl(nn~ k)!" 272 (19. This is a combination problem (also called a comn1ittee problem). it does not n1atter what order you selected them. it is useful to know the possible combinations and permutations of certain objects. . 7 shirts and 2 hats. 8w. and the total number is 5 x 4 x 7 x 5 = 700. the third place 3 choices. B. 81s. Order does not count in this case. and there is only one left for the last place.. All possible outfits you can wear form a permutation problem. E.. In general.1 Probability Probability and Statistics In calculating the probabilities. in the first place.
a. and the kth person has (365. What is the probability of the club consisting of 2 female students. If the event A denotes that all the n people will have different birthdays (no birthday matching). Therefore. the first person can have any date as his or her birthday.Probability and Statistics 19. Therefore.11) .1 Probability The consistency requires OJ= 1. * (365. For a group of 367 people.P(A. If two female students are selected. 10 male students. we have 8 C2. 10 C2 for selecting 2 male students. The birthday paradox was first proposed in 1939 by Richard von l\1ises. which is 364/365. Ignoring 29 February and the year of birth and assuming that the birthdays are evenly distributed throughout the year. Similarly. n) = 1. · x 365 364 (365. 2 male students.n)f365n · 365! Now the probability of two people with the same birthday is P(A.n + 1) 365n _ 365! (19. n) = 365 x 365 x . n) = 1.9: A club of 5 members is chosen at random from 8 female students. The second person must be in other 364 dates. the total number N of forming the 5member club is N = sc2IOC27Cl = ~ ~ 0166 2SC5 253 · 0 There is an interesting 'birthday paradox' which is related to this context. which states that what is the probability of two people having the same birthday in a group of n people. and 7 teachers.n 365 + 1) 365 * (364) * . and 7 C 1 for selecting one teacher.(365. 365/365. 273 (19..(365. 0 Example 19... we only have 365 different birthdays (days and months only). the probability of no two people having the same birthday is P( A.k + 1)/365.n)!365n. it is certain that there must be two people having the san1e birthday as there are only 365 (or 366 if someone was born in a leap year) possible birthdays.nd 1 teacher? The total number of clubs is 25 Cs.12) .
If we use notation 'BIA' which means that event B occurs given that event A has occurred. 30) ~ 0. P(A) = 1/20 and P(B) = 1/20. For a second case (Case II). We call this easel. (19. the probability becomes 365! P(A.23)136523 ~ 0. but you do not have to deal with such large numbers. You can use a simple calculator to estimate it. 23) = 1 .(365 . you get P(A. 40) ~ 0. The probability of getting two red balls are P(A n B) = P(A) * P(B) = 1/400.A.891 for n = 40. the interesting thing is that for n = 23. However. then we can use P(BIA) to denote the probability of event B when there is no replacement in event A in the scenario 274 .fg = ~~ which is different from 1/400. Suppose there are 10 red balls and 10 white balls in bag. then the probability of event B is now different P(B) = 1/19 as there is now only 19 balls in the bag. It is worth noting that there is some difference in combinations when the member drawn is placed back or not. ' ( 19 .14) This means that you have slightly more than a 5050 chance of finding two people sharing the same birthday. P(.Jrr x .19. and P(A. The reason here is that the two events are not independent in the case of noreplacement. For five people. if we do not put it back after we have drawn the first ball (without replacmnent). then we draw another ball (event B).507. The probability of getting two red balls now becomes P(A n B) = .A 5 ) ' = 1 _ 365 * 364 * 363 * 362 * 361 365 5 ~ 0 027 .970 and P(A. 706 for n = 30. event A) from the bag and then put it back (with replacement). the probability of two people with the same birthday is P(.1 Probability Probability and Statistics The factorial 3651 is a large number. 70) ~ 0.9992 (aln1ost certainty) for n = 70. If we draw a ball (say a red. 50) ~ 0. 13) which seems insignificant. If you increase n.
Now P(B) becomes P(BIA).1 Probability described in Case II.2 Conditional Probability In calculating the probabilities. (19.Probability and Statistics 19.15) which is often called the multiplication rule in probability theory. we often assume that all possible outcomes of an experiment such as drawing a card are equally likely.19) = which is the Bayes' theorem. (19. (19. The conditional probability can be calculated by P(BIA) = P(B n A) P(A) . Similarly. we have P(AIB) P(A n B) P(B) . 19.P(A). Hence. we can get P(A n B) = P(B)P(AIB). (19.1.15).16) This is essentially a conditional probability problem which forms the main topic of the next section. we can write the above formulation as P(AIB) = P(A)P(BIA) P(B) P(A)P(BIA) P(A)P(BIA) + P(A)P(BIA)' (19. 275 Au A 0 and . we have P(A n B) = P(A)P(BIA).18) Using equation (19. Probabilities can change if additional information is known or some other event has already occurred and thus P(BIA) denotes the probability that event B will occur given that event A has already occurred.17) Conversely. Here have used P(A) = 1 .
and probability of pull the blueblue card out is P(BB) = 1/3. For athletes not taking drugs. (19. 0. suppose that the side you can see is blue. Suppose 276 . Thus.005.4: The classical problem of three cards consists of three cards: one blue card (B) is blue on both sides. the probability of other side being also blue is P(BBIBF) = P(BFIBB)P(BB) = 1 X P(BF) ! t =~ 3· (19. P(BIA) = 0.1 Probability Probability and Statistics As an example.20) 0 Example 19. and one mixed card (A1) is wbite on one side and blue on the other. Then.5%.001 * 0.99 ~ 0. What is the probability that the athlete is really taking the drug? If event A denotes an athlete is taking the drug.165. one white card (W) is wbite on both sides.P(A)P(BIA)+P(A)P(BIA) 0. what is the probability of other side is also blue? This a conditional probability problem. The probability that the athlete is actually taking the drug is P(AIB) _ P(A)P(BIA) . the positive test is only 0.99 + 0. P(A) = 1/1000.999 * 0. thus tbe total probability of showing a blue face (F) is P(BF) = 1/2. Suppose an athlete is selected at random and the test shows positive for the drug. while the probability of showing a blue face is P(BFIBB) = 1 if the pulled card is blueblue one.001 * 0. t. It is believed that the test is 99% accurate if athletes are taking drugs. It is assumed that only one in 1000 athletes takes this kind of drug. we consider the drug test in sports. If you draw one card at random from a bag and place it on a table. There are 3 blue faces and 3 white faces. and B denotes the event that the individual tests positive.99 and P(BIA) = 0.19.21) !\lost people will intuitively guess that the probability is not correct. which is 0 Another related problem is the socalled l\1onty Hall problem (or three door problem) in a classical game show.005 This is surprisingly low in probability.
to reveal a goat. 277 . the best strategy is to switch your choice. A) at random. this gives P(OpenBIA) = 1/2. This game has other variations such as the three envelope problem and others. so your initial choice P(A) = 1/3. B). each value Xi may occur with a certain probability p(xi)· In other words.1.Probability and Statistics 19. but the analysis and strategy are the same. P(OpenBIB) = 0. the prize car behind any door (Y) has a priori probability P(any) = 1/3. So the total probability of opening door B is P(OpenB) = P(A)P(OpenBIA) +P(C)P(OpenBIC) + P(B)P(OpenBIB) 1 1 = '3 x 1 2 + 3 x 0 + 3 x 1 = 2. if the car is behind A.3 Random Variables and Moments Random Variables For a discrete random variable X with distinct values such as the number of cars passing through a junction. = 1 x ~ = ~' ~ 1 1 (19. C). the host will surely open B. and P(OpenBIC) =I. then the host opens one of the other door (say. Nlathematically. behind other two doors are goats. As the host knows where the prize is. What is your best strategy based on probability? Initially. If the car is behind B.23) 3 which is greater than your initial choice 1/3.22) Now the probability of the car behind door Cis P(CIOpenB) = P(OpenBIC)P(C) P(OpenB) (19. If you choose one door (say.1 Probability that you are given the choice of three doors. which he knows there is a goat behind it. the host will open B or C so 1/2 each. There is an expensive car behind one door. Now you have a choice either to stick with your original choice or swap with the other door (say. Therefore. 19. and if the car is behind C. the host will never open B.
As an event must occur inside a sample space. (19. we always use the interval (x. A probability function p(xi) is a function that defines probabilities to all the discrete values xi of the random variable X. i=l (19.oo)].27) The cumulative probability function becomes ~(x) = P(X ~ x) = J. (19. The mean 278 .25) For a continuous random variable X that takes a continuous range of values (such as the level of noise).x p(x)dx.28) which is the definite integral of the probability density function between the lower limit a up to the present value X= x. the requirement that all the probabilities must be summed to one leads to n LP(Xi) = 1.1 Probability Probability and Statistics the probability varies with the random variable. In this case. b] for given limits a and b [or even over the whole real axis x E (oo.26) As all the probabilities of the distribution shall be added to unity. (19. Mean and Variance Two main measures for a random variable X with a given probability distribution p(x) are its mean and variance.24) The cumulative probability function of X is defined by P(X ~ x) = L Xi<X p(xi)· (19. we have l p(x)dx = 1.19. x + dx] so that p(x) is the probability that the random variable X takes the value x < X ~ x + dx is ~(x) = P(x <X~ x + dx) = p(x)dx. its distribution is continuous and the probability density function p( x) is defined for a range of values x E [a.
31) The square root of the variance a = Jvar[X] is called the standard deviation.1 Probability J1. any other formulas for a continuous distribution can be converted to their counterpart for a discrete distribution if the integration is replaced by the sum. In addition. (19. This simply becomes a sum (19. Other frequently used measures are the mode and median. For an even number of data sets. The median m of a distribution corresponds to the value at which the cumulative probability function <l>(m) = 1/2.L) 2 .Probability and Statistics 19. The upper and lower quartiles Qu and QL are defined by <I>(Qu) = 3/4 and <I>(QL) = 1/4. 279 . which is simply a.32) for a discrete distribution. the mode may have two values. =E[X] =<X>= j xp(x)dx. If the random variable is discrete.29) for a continuous distribution and the integration is within the integration limits. or the expectation value of E[X] is defined by J1. we will mainly focus on the continuous distribution in the rest of the section. Therefore. The mode of a distribution is defined by the value at which the probability density function p( x) is maxin1um. That is (19. then the integration becomes the sumn1ation E[X] = L XiP(Xi)· (19.J.30) The variance var[X] = a 2 is the expectation value of the deviation squared (X..
36) where a and . k = 1. . (19.34) Obviously. = vk k = 1..p.6] = a 2 var[X]. From these definitions. N. the kth central moment is defined by E[(X.33) Similarly. N.38) and a  2 _ ~Gx(O) _ [dGx(0)] 2 dv 2 dv · (19. . By expanding exp[vx] into power series and using the definition of various moments. it is straightforward to verify that (19. it is straightforward to prove (19.)kp(x)dx. the mean is essentially the first moment if we define the kth moment of a random variable X by E[Xk] =Jlk =I xkp(x)dx.6 are constants.19. 2. The moment generating function is defined by (19.. the variance is the second central moment. .35) and var[ax + . ..39) 280 . (19. (19.Ill] =I (x. Ivlost probability functions can be expressed in terms of moments and moment generating functions. 2..37) where v E 'R is a real parameter.1 Probability Probability and Statistics Moments and Moment Generating Functions In fact.
p) if its probability distribution is given by B(n.) 3 = ~ ~ o 111 2 2 128 .41) which is exactly the requirement that all the probabilities should be summed to unity.x failures in n trials.p) = nCxpx(l. 2. then we have 10c7 (. B(n. p is the probability of a socalled 'success' of the outcome. = 1/p and a2 = (1 . There are only two possible outcomes: success or failure. This distribution is used to calculate the first success.!.42) where n ~ 1.Ptx.!. The coefficients come from the coefficients of the binomial expansions n (p+qt = LnCxpxqnx = 1. .1. thus the first n. Therefore.5: Tossing a coin 10 times. .p. the probability of getting 7 heads is B(n. x=O (19.40) where x = 0.p)n1' (19. nc  x n! xl(nx)l' (19. n is the number of trials.p)jp2. Subsequently.p) for a binomial distribution. 0 Example 19.)1(.1 trials must be in failure if n trials are needed to observe the first success. The mean and variance of this distribution are J1. 281 .. .1 Probability 19.p) represents the probability of x successes and n. Since p = 1/2 and x = 7. 1. 1/2).Probability and Statistics 19.4 Binomial and Poisson Distributions Binomial Distribution A discrete random variable is said to follow the binomial distribution B(n. = E[X] = np and a 2 = npq = np( 1 .. the probability of the failure of a trial is q = 1 . 0 It is straightforward to prove that J1. n are the values that the random variable X may take. Another related distribution is the geometric distribution whose probability function is defined by P(X = n) = pqn1 = p(1.
3. The Poisson distribution is A> 0.149.\ = 3. it is often called the distribution for smallprobability events. 1. 31e3 P(X = 1) = . Typically. it is concerned with the number of events that occur in a certain time interval (e..0498 ~ 0.g. For this reason.6: If you receive 3 calls per hour on your mobile phone on the average. The probability of no phone call at all is Thus. If you do not switch your phone off. The parameter A is the location of the peak as shown in Figure 19. number of telephone calls in an hour) or spatial area. 1 and the probability of receiving two calls is 0 282 ..1 Probability Probability and Statistics Poisson Distribution The Poisson distribution can be thought as the limit of the binomial distribution when the number of trial is very large n + oo and the probability p + 0 (small probability) with the constraint that A = np is finite. Using the definition of mean and variance. n and A is the mean of the distribution.43) where x = 0..95. it is straightforward to prove that J. 2. . what is the probability that it begins to sound (one call is enough) during any onehour class? \'Ve know that . the probability of receiving one call is ~ 1 . 0 Example 19. the probability of sounding is P(X > 0) In fact.L = A and a 2 = A for the Poisson distribution.0. (19.19.. .::::::: 0..
For a continuous random variable X.X = 5.. the probability density function (PDF) of a Gaussian distribution is given by p(x) = ay27T ..Probability and Statistics 19.3: Poisson distributions for different values of . m=e~..03 'iC' II I rt 0.· A=10 0._..A=S .1.1 Probability 0.. 10.45) where a 2 = var[X] is the variance and Jl = E[X] is the mean of the Gaussian distribution.44) 19.04 A=15 .01 :' . I 25 X Figure 19. .)2 (19.05r. 1 (x. 15. The moment generating function for the Poisson distribution is given by (19.46) . 283 (19. I \ I \ \ \ 0. it is easy to verify that 00 / 00 p(x)dx = 1. From the Gaussian integral.5 Gaussian Distribution The Gaussian distribution or normal distribution is the most important continuous distribution in probability and it has a wide range of applications.===::::::...02 I I :' : I \ \ \ \ \ 0.
and this is exactly the reason that the factor 1/ . Using the Sterling's approximation xl rv .iX(xje)x for x >> 1.l =A and a 2 =A.' 0 X ____ .f2.. ...t (19. 10.49) The Gaussian distribution can be considered as the limit of the Poisson distribution when A >> 1. The cumulative probability function (CPF) for a normal distribution is the integral of p( x). v2 v2a The moment generating function for the Gaussian distribution is given by (19. which is defined by 4>(x) = P(X < x) = ~ 1 21ra 2 !X oo e (<w2 2a d(.02 I I .08 <J=5 ..1 r. .48) 4>(x) = ro[1 + erf( ro )]. .1 Probability Probability and Statistics 0.· <J=7 ....<J=10 0.. it can be verified that the Poisson distribution can 284 ..':' I I /. we can write it as 1 XJ._ . (19. and setting J.4: Gaussian distributions for a = 5~ 7.' .... .·' 20 10 10 20 30 Figure 19...::::======il 0.l and the variance a 2 controls the width of the peak (see Figure 19.47) Using the error function defined by Chapter 1.19. The probability function reaches a peak at x = J.~...2i comes from the normalization of the all probabilities.4)..
655 0.540 0.758 0.1 0.618 0.2 0. The standard normal distribution is a normal distribution N(J.50) where J.841 0.53). XJ.788 0.971 285 .0 0.8 1.t = A.919 0. a) to emphasize that the probability density function depends on two parameters J.5 0.t.9 0.964 0. the normal distribution is often written as N (J.816 1. that is JV(O.0 1.4 0.1: Function¢> defined by equation (19.692 0.L and a.L a (19.J..579 0.945 0.3 1.7 0.7 1.885 0.6 1. It is worth pointing out that some books define z = ~ = (x .L) /a in this case. ~ ¢>(~) ~ ¢> 0. a) with a mean of J. 1) This is useful to normalize or standardize data for statistical analysis.4 1. Table 19. If we define a normalized variable ~=. ~ sorts where the data falls.9 0.903 0.1 1.933 0.1 Probability be written as a Gaussian distribution (19. and call the standard normal distribution as the Z distribution.864 0...500 0.2 1.5 1. In terms of the area under the probability density function.51) it is equivalent to give a score so as to place the data above or below the mean in the unit of standard deviation. In statistical applications.6 0.t.726 0.8 0.L = 0 and standard deviation a = 1.955 0.Probability and Statistics 19.3 0.
.JL) 2 f(x) = ~exp[ (19. ~)].\e. (19.b]. Its mean and variance are (19.52) Its cumulative probability function is 1 ¢(~) = ~ ~~ e~: d~ = .56) The lognormal distribution has a probability density function 1 (lnx. (19. {3a: 1 X= (a:.53) As the calculations of ¢ and the error function involve the numerical integrations. The exponential distribution has the following probability density function f(x) = .1 Probability Probability and Statistics Now the probability density function of standard normal distribution becomes (19. 2a 2 x 21ra2 286 .57) ]. uniform distribution and the x2distribution.1. 2 v27r oo v2 (19..54) whose mean is E[X] = (a: + {3) /2 and variance is a 2 = (/3 a:) 2 /12.[1 + erf(. The uniform distribution has a probability density function p= . it is usual practice to tabulate ¢ in a table (see Table 19. lognormal distribution.55) and f(x) = 0 for x:::.Xx (x > 0).6 Other Distributions There are a dozen of other important distributions such as the exponential distribution.19. 0.1) so that you do not have to calculate their values each time you use it.. 19.
This theorem is as follows: Let Xi(i = 1.62) .Li and a variance a'f.58) The x2distribution.. x/2 is the incomplete gamma function. readers can refer to any books that are devoted to probability theory and statistical analysis. . . 19.1.Probability and Statistics 19. It can be verified that the mean of the distribution is n and its variance is 2n.61) where "Y(n/2. x/2) f(n/2) ' (19. each of which is defined by a probability density function Pi (x) with a corresponding mean J./li)2 i=l ai . 2. and n is called the degree of freedom. 287 (19. is very useful in statistical inference and method of least squares.60) where x 2:: 0.. The sum of all these random variables e= Lxi = i=l n Xt +X2 + .7 The Central Limit Theorem The most important theorem in probability is the central limit theorem which concerns the large number of trials and explains why the normal distribution occurs so widely.. The probability density function for x2distribution is given by p(x) = 2n/2:(n/2) x1'lexf2 (19. (19...1 Probability whose mean and variance are (19.59) where thenindependent variables Xi are normally distributed with means Jli and variances a'f. +Xn. n) ben independent random variables. This distribution is for the quantity XnL 2 _~(Xi. called chisquare or chisquared distribution. For other distributions. Its cumulative distribution function is <I>( ) X = "'f(n/2.
1 0.64) The proof of this theorem is out of the scope of this book as it involves the n1oment generating functions.jn n ' (19. Its mean E[8] and variance var[8] are given by E[8] = and var[8] L E[Xi] = L Jli.1 Probability Probability and Statistics 0..5: A uniform distribution. i=l n (19.65) = 8. we simply use these important results for statistical analysis.. In engineering mathematics.np. (19.66) 288 . a .15 C) i :::s LL ! 0.19. is also a random variable whose distribution approaches the Gaussian distribution as n ~ oo.2 >0. In the special case when all the variables Xi are described by the same probability density function with the same mean Jl and variance a 2 .63) = L var[8] = i=l La?. characteristics functions and other techniques. i=l i=l n n n (19.05 tr 1 2 3 4 Outcome 5 6 7 Figure 19. these results become E[8] = np. By defining a new variable {n var[8] = na 2 .
1667.6: A bilinear distribution.000 times.15 0. then the distribution of ~n converges towards the standard normal distribution N(O. the sums of the face values vary from 1 to 90.2 Statistics Statistics is the mathematics of data collection and interpretation. 1) as n ~ oo. the distribution is shown in Figure 19. each side will appear equally likely with a probability of 1/6 ~ 0. After rolling the 15 dice 10. If we now roll two independent dice 15000 tin1es and count the sum (112) of the face values of both dice. 19. and the analysis and characterisation of numerical data by inference from sampling. Let us see what the theorem means for a simple experiment of rolling a few dice.2 Statistics 0. then the sum obeys a bilinear distribution as shown in Figure 19. For a fair sixsided die.5. thus the probability function after rolling it 15000 times approaches a uniform distribution as shown in Figure 19.7 and it approaches to a normal distribution as n ~ oo.2 19. Statistical methods involve reduc 289 . If we roll n = 15 independent dice.6.05 2 I 4 Outcome 6 8 10 12 I 14 Figure 19.Probability and Statistics 0.
. 0. tion of data.67) n i=l which is essentially the arithmetic average of the values Xi. Xn on a random variable x such as the price of a cup of coffee. Y2.LXi.04 0. + Xn) n 1 1 n = . two important and commonly used parameters are sample mean and sample variance.7: An approximate Gaussian distribution (the outcomes of the sum of face values in rolling 15 dice)..1 Sample Mean and Variance If a sample consists of n independent observations Xt. (19..06 ~0. and the test of hypotheses.08 >CD ::I ~0. which can easily be estimated from the sample.1 r. and relationship between two or more variables by analysis of variance.. x2.. .. Generally speaking.02 117' ~0 30 Outcome 50 60 70 80 Figure 19.rr.2..li and a corresponding variance 290 . if u is a linear combination of n independent random variables Yt.19.. estimates and significance tests. . 19.. The san1ple mean is calculated by X =<x>= (Xt + X2 + . ..2 Statistics Probability and Statistics 0. Yn and each random variable Yi has an individual mean J.
li = J.J.67) can also be viewed as a linear combination of all the Xi assuming each of which has the same mean J.lu of the linear combination n n J.2 Statistics a[. and the same weighting coefficient Gi = 1/n. (19.j=l. From the central limit theorem. we have the linear combination n U = L O:iYi i=l = O:tYl + G2Y2 + ··· + GnYn.68) where the parameters ai(i = 1.l and variance u'f = u 2 .lx = Ef=t J.li)(Yi .J. 2. i=l (19.Li) 2 ] = Since Yi and Yi are independent.l. however.J. we have the variance 0"~ = ~ 0"2 = x n 1 0"2 ~ n2 t=l n' (19.73) 291 . Hence. we get n ar u~ = L:a'faJ.L/n = J.J.70) which can be expanded as n a~ =La~ E[(Yi . In this case. .Probability and Statistics 19.72) The sample mean defined in equation (19.lu) 2] = E[L o:i(Yi.lj)]. we have E[(Yi . .69) Then.J.lu = E(u) = E(L GiYi) = L aE(yi) = L GiJ. (19.li) 2].J. we have the mean J. n) are the weighting coefficients..li)2] i=l n + L i. Therefore. the variance u~ of the combination is u~ = E[(u..J.i#j O:iGjE[(Yi.li )] = 0.li· i=l i=l (19. the sample mean is an unbiased estimate of the sample due to the fact J.71) where E[(Yi .li )] = E[(Yi .li)]E[(Yi .J.li)(Yj.J. i=l n (19.J.
(19. we know that E[x 2] = J.•• .x) . ...ls2 = 1~ E[. ... The sample variance 8 2 is defined by S2 = .X2... The mean of the sample variance is J.. we need at least 2 values to estimate the variance..2 Method of Least Squares Maximum Likelihood For a sample of n values Xt. the joint probability is then 292 . f3k.2 Statistics Probability and Statistics which means the variance becomes smaller as the size n of the sample increases by a factor of 1ln.1) is that we need at least one value to estimate the mean.. Thus.(xix) 2] n.1).L = L 1~ 2 2 E[(xi nx )]... .1) will give the correct and unbiased estimate of the variance.74) It is worth pointing out that the factor is 1l(n 1) not 1ln because only 1I (n ..19.1 different values of variance can be obtained to estimate the total sample variance. we get Jls2 = n ~ 1 ~ { E[x~] . From the probability theory in the earlier sections. which would underestimate the sample variance. (19.2.L.1 i=l J.L. The other way to think the factor 1I (n . we would get J. for n observations.75) 1 i=l n Using E[x 2] = 2 + a 2In. if we use the factor 1In instead of 1I(n .76) n1 n 0 bviously.l2 + )} = a2..l 82 = n~ 1 a 2 < a 2.nE[x2 ]} 1 a2 n(J. = { n(J. 1 ~ n.1 i=l 2 (19.(xi..L 2 + a 2.l2 + a2)  19.Xn of a random variable X whose probability density function p( x) depends on a set of k parameters f3t.. only n .
!3k)p(x2. !3k) · · · p(xm f3t.78) whose solutions for /3i are the maximum likelihood estimates. For n sets of data points (Xi..79) 293 . . f3t. 84> (i = 1.. . . 2.8: Least square and the best fit line. k).2 Statistics = p(xt. Such a line is often called the regression line or the best fit line as shown in Figure 19. f3k)· (19.Probability and Statistics 19.. Yi). y Figure 19... . (19. The method of least squares is to try to determine the two parameters a (intercept) and b (slope) for the regression line from n data points. As the sample can be considered as given values. we usually plot one variable such as pressure or price y against another variable x such as time or spatial coordinates. . Linear Regression For experiments and observations.. . We try to present the data in a way so that we can see some trend in the data. Assuming that Xi are known more precisely and Yi values obey a normal distribution around the potentially best fit line with a variance a 2 ... the usual practice is to try to draw a straight line y =a+ bx so that it represents the major trend. the n1aximum likelihood requires that 0/3i = 0.77) The essence of the maximum likelihood is to maximize 4> by choosing the parameters /3i. we have the probability (19. . . f3t. Hence.8.
It is worth pointing out that the exponent L:i=dYi.19. The maximization of P is equivalent to the minimization of the exponent 'l/J n 'l/J = L)Yi . In order to minimize 'l/J as a function of a and b.r = (Yi. thus the name of the method of least squares. The essence of the method of least squares is to maximize the probability P by choosing the appropriate a and b.f(xi)) 2 where f(xi) =a+ bxi.83) and n n a LXi i=l + b L:x~ = LXiYi.82) By expanding these equations. and f(x) is the function for the regression [f (x) = a + bx for the linear regression].81) and (19. a i=l (19.84) which is a system of linear equations for a and b. That is 8'1/J n 7) = 2L[Y. i=l i=l (19.(a+ bxi)] = 0. we have n n na+bLxi = LYi. and it is straightforward to obtain the solutions as (19.85) 294 . i=l i=l n (19. The minimization means the least sum of the squares.2 Statistics Probability and Statistics where A is a constant.f(xi)f · i=l (19.80) We see that 'l/J is the sum of the squares of the deviations f.f(xi)] 2 ja2 is similar to the quantity X~ defined in the x2distribution. its derivatives should be zero.
295 .b.92) (19.(KX )2 ' The residual error is defined by fi (19.a .87) If we use the following notations (19.89) then the above equation for a and b becomes KxxKy KxKy a= n K X X .L Xi n n i=l 1 1 n = y .(a + bxi).2 Statistics where 1 n y = .bx] .2 degrees of freedom.LYi· n i=l (19.88) and n n Kxx =LX~.a . The sample variance 8 2 is (19. i=l Kxy = LXiYi.93) where the factor 1/ (n .91) whose sample mean is given by Jte = L n i=l 1 n fi = Yi . i=l (19.90) = Yi .a = 0.Probability and Statistics 19.bx = [y .2) comes from the fact that two constraints are need for the best fit. (19. and the residuals therefore have a n .
J.K~) (19. From these data. 296 .lx](Y . we have the following data: Height (h): 162.94) where cov(x. If x has a mean J. 12. y) = E[(x .5. 7.y where Kyy = . the correlation coefficient can be calculated by or rx. 176..2 Statistics Probability and Statistics Correlation Coefficient The correlation coefficient rx. If ri. Shoe size (s): 5.Ls = 8..===::::::::::=======..19. y) = 0.J. r x. then there is a linear relationship between these two variables.5.7: Is there any relationship between shoe size and height among general population? By collecting data randomly among our friends.lh = 172.. 11. 183.)(nKyy. nKxy KxKy J(nKxx. Yi).y = 1 is an increasing linear relationship where the increase of one variable will lead to increase of another. 8. andy has a mean J. 168.96) = Ei=l yr 0 Example 19. 7. 171.y is a very useful parameter to find any potential relationship between two sets of data Xi and Yi for two random variables x and y. respectively. 10.5.y = 1.. we know the sample mean J. 6.K'.y = 0). For n sets of data points (Xi.ly) is the covariance.lx and a sample variances. 167. r x. If the two variables are independent cov(x. 179 (em).5..5.ly and a sample variance s~' the correlation coefficient is defined by (19. J.y = 1 is a decreasing relationship when one increases while the other decreases. 174. there is no correlation between them (rx.5.
\Ve also have the standard deviation of height sh = 6.179.2.ls)] = 13. 297 . it means that we can expect that about 95% of the sample will fall in the confidence interval while 5% of the data will not.a (say.1 Q.9: Confidence interval 'Y = 1 a.Probability and Statistics 19. s) = E[(h.97) The predetermined parameter 1' is always near 1 so that it can be expressed as a small deviation a << 1 from 1 (see Figure 19.179 This is a relatively strong correlation indeed. 95% or 99%)...422*2. The covariance cov(h.422 and the standard deviation of shoe size Ss = 2. If we choose 1' = 95%. D 19. . ~ 0..2 :::: _ _ r = cov(h.3 Hypothesis Testing Confidence Interval The confidence interval is defined as the interval fh ~ X ~ 02 so that the probabilities at these two limits fh and fh are equal to a given probability 'Y = 1.lh)(s.2.. Therefore.9). s) : : : : 0 94 ShSs 6.rr.os Figure 19. the correlation coefficient r is given by 13.15.J.J. That is (19...!! c:: ~ c!o.2 Statistics 0.
98) If a = 0.999 for highly significant.n· (19.645.L = x= 1 g(19 + 15 + 30 + 20 + 15 + 23 + 28 + 22 + 23) = 21.05.a. 20.960. It is conventional to use 1 = 0.. 28.Jii for a sample size n. In other words. which is equivalent to say that the standard deviation becomes a/ .a. 298 . for 8 = 1. this means P( 8 ~ 8) = 1. So you shall start x = J.L +a] or = 0. 15. then 8 = 2. we have </>( ~ ~ 8) say.19.95. 23 minutes. we have a a J. it gives~ = 1.99 for significant.579.2 Statistics Probability and Statistics ~ ~ For the standard normal distribution. If a = 0.99) On the other hand..01.67.a ~ x ~ J. the variance will decrease from a 2 to a 2 jn.975 or 8 = 1.682. 8: The sample data of the time taken for a quick lunch at a restaurant are as follows: 19. and 0.15. If you are willing to take 5% chance..2" (19.. we get J. (19.579 Vn ~ X ~ J. only 68.L +~a = 30.8a ~ if you repeat an experiment n times.15 minutes.L + 2...L.2% confidence level.L ±a.579 .. 8 ~ ~ ~ 8 or J. so that a 4>(~ ~ 8) = 1 . at what time you should start your order if you want to take 5% chance of being late? The sample mean is J. 22. The sample variance is which gives a standard deviation of a = 5.2% of the sample data will fall in the interval [J.L +a and 1' = 0.2. then <!>(~) = 0..100) with a 68. 0 Example 19. That is to x ~ J. 0.L .L.L. Suppose you want to attend a lecture at 12:30. 30...L + 8a. 15.95 for probably significant. J. We also know that x = J. 23.
101) The Student's tdistribution with k = n . the tvariable is defined by XJ. It has become one of the most popular methods for hypothesis testing.1 degrees of freedom is the distribution for the randon1 variable t. k (2 k 1 (19.103) oo 299 . This test can also be used to test if the slope of the regression line is significantly different from 0. S. 19.Probability and Statistics which is about 30 minutes earlier or at about 12:00.L t = (Sf Vn). (19. .2) fork> 2 and infinite for 0 < k ~ 2. This is a powerful method for hypothesis testing using smallsize samples.74). and the probability density function is (19.102) It can be verified that the mean is E[t] = 0. x2. This method was designed by W. The variance is a 2 = kf(k.. . For n independent measurements/data Xt. Gosset in 1908 and he had to use a pen name 'Student' because of his e1nployer's policy in publishing research results at that time. which of course can be estimated in terms of the sample variance 8 2 from the sample data.. The corresponding cumulative probability function is F(t) = r( k+t) v'f:rrr(k/2) r 1t [1 + ]+d(.2 Statistics 0 Student's tdistribution The Student's ttest is a very powerful method for testing the null hypothesis to see if the means of two normally distributed samples are equal. The theoretical basis of the ttest is the Student's tdistribution for a sample population with the unknown standard deviation a. Xn with an estimated sample mean x and a sample variance 8 2 as defined by equation (19.
975 12. Hence. we have p( 0 ~ t ~ 0) = 1 . ..30(k = 2).995 63. .93 5.98 1.. and it is not straightforward to calculate. ..60 4. .0 .05 and 0.25 3. we have 0 = ta.959 for k+ oo.975. 3.68 2.7 9.19.78 2. rn..30 3.a = 0.70(k = 1)..105) .2"' Q (19.II.995).a/2 0.50 3. k F(O)o.37 2.71 3.104) which is usually tabulated.45 2.26 2. 2..18 2.03 3. 300 s s (19.2: Limits defined by F(O) = 1 . .96 F(O)o.2.a. For a = 0.a/2 in equation (19. the confidence interval is given by F(O) = 1.k = 12.84 4.09 2. that is why they are tabulated in many statistical tables.2 Statistics Probability and Statistics This integral leads to a hypergeometric function.95 or p(t ~ 0) = 1 a/2 = 0. the values are tabulated in Table 19.. For a confidence level of y = 1 . .31 2. 1.104).975 and 0.23 2.63 2. vn .85 2.7 4._..36 3. = Table 19.58 1 2 3 4 5 6 7 8 9 10 20 50 100 00 Suppose we are dealing with the 95% confidence interval.17 2.< t < II+ 0. 4.18(k = 3)..01 1.228(k = 10).57 2.01 (or 1.
J.Lo = 110).273. 0 Example 19.fii) 4. The null hypothesis: Ho: J.L= J. and most common are the one sample ttest and the two sample ttest.99. The one sample ttest is used for measurements that are randomly drawn from a population to compare the sample mean with a known number. 3. 103. you randomly sample 11 students to do the IQ test and results are: x = IQ = 106. we have = 11.106. more than 100 students) have claimed to have an averaged IQ of 110 (or J.2 Statistics This is much more complicated than its counterpart of the standard normal distribution. The basic steps of a ttest are as follows: 1.101. S = 4.100. x = 105..of2.108. Test the hypothesis: Ho: J.L = J.Lo. we first have to pose precise questions or form a hypothesis. or H0 : f. 106.o by using F( t ::s. reject the hypothesis.273.077/JIT 301 . Otherwise.Probability and Statistics 19..077.108.. Then.. 0) = 1. and such hypothesis is conventionally called the null hypothesis.112. In order to do statistical testing. From the data. we know that n Then.L) = (105. at a confidence level of95%..9: A group of candidates (say. t = (x. 109.110) ~ _ _ _ 3 846 (S/.L2 for two samples. accept the hypothesis.Lo (often known value) for one sample. 2.Lt = f.. Student's ttest There are quite a few variations of the Student's ttest. Calculate the ttest statistic t and find the critical value () for a given confidence level 1 = 1 . If ltl > 0.
1) """~ (U.Vi data sequence. the probability oft>(} is 0. _ iT.5%) and the probability t < 8 is also 0. 75.025 (or 2. 77. Thus.228 * Sj. 76. can you say the new method is really better than the standard method? If we suppose that the two methods do not produce any difference in results.L.1 = 10 degrees of freedom at a 95% confidence level. the hypothesis is not valid at a 95% confidence level.2.variable is defined by n(n. The t.. At the same level of confidence. (} = 2.Lo lies in in the range of x. their mean are the same. Assuming that two pairs of n sample data sets Ui and Vi are independent and drawn from the same normal distribution. 8 students are randomly drawn from each class.Lo ~ 108.10: A novel teaching method of teaching children science was tried in a class (say class B). 77.Jff. At a 95% confidence level. 81.1 i=l This is equivalent to apply the onesample test to the difference ui .)Ui 1 ~ n.J[f ~ J.00. 82. 81.228. 82. 81. that is to say.2 Statistics Probability and Statistics We only use the positive value if we look at the statistical tables. lVe also know for k = n. 76. the paired ttest is used to determine whether they are significantly different from each other.19. At a 95% confidence level. (19. 82. At the end of the assessment. Vi where {fi = Ui . 80. 0 Another important ttest is the twosample paired test.53 ~ J. 0 Example 19. Tl1us the null 302 . and their science scores are as follows: Class A: Ui = 76.025. 77. In addition.)2' 6t=l t.2288/.[J and 1% = Vi 2 sd = (19.106) V. or 102. while a standard method was used in another class (say class A). 86.Lo ~ x + 2. Vi) 2 .107) . Class B: Vi= 79. the true mean J.
5. It is worth pointing out that other important methods for hypothesis testing are Fisher's Ftest.581 = 3 5 2. We now have V = 81. however. What we have discussed in this chapter is just a tip of an iceberg.. and there are many excellent books on these topics. Readers can refer to the relative books listed at the end of this book. That is to say.\ = JlB· lVe know that V = 77.828. The combined sample VBriance t = Sd/Vn [J v = 77. the new method does produce better results in teaching science.37 for F(8) = 1 . and nonparametric tests. As t < 8 or t > 8.2 Statistics Ho: J. we can reject the null hypothesis.o:/2 and k = n . x2test.. Sd = 2.Probability and Statistics hypothesis is: 19. 0 The variance analysis and hypothesis testing are important topics in applied statistics.828/v'8 .1 = 7. it forms a solid basis for further studies.L. We know from the statistical table that the critical value 8 = 2. 303 .
.
Vol.. BlackScholes and Beyond: Option Pricing Models. Academic Press. (1982). F. Finite Element Procedures in Engineering Analysis. Oxford University Press. and Neukirch S. Conduction of Heat in Solids. ( 1985).. Engineering Materials. J.. PrenticeHall. S. Pergamon Press. 2nd Edition. A. F. Irwin Professional Publishing. 2nd Edition. Arfken G. Dover Publication. (1980). Partial Differential Equations: Theory and Technique. Bathe K. Rev. 2005. (1965)... (1997).. E.. and Stegun I. Carslaw. Jaeger. Phys. 95. Tech Science Press. Lett. Academic Press. Handbook of Mathematical Functions. Atluri S. 305 .References Abramowitz M. Carrrier G. (1986). Chriss N. 1988. Methods of Computer Modeling in Engineering and the Sciences. H. and Pearson C. Audoly B. Mathematical Methods for Physicists. R. Fragmentation of rods by cascading cracks: why spaghetti do not break in half. Ashby lvi.. N. and Jones D... I. 095505 (2005).
Methods of Mathematical Physics.. 306 . Mathematics of Diffusion. Wileylnterscience. A. C.. 20. PrenticeHall. Astron. Perturbation Methods.. A. and S. The dynamics of planetary rings. Devaney R. Mathematical Models in the Applied Sciences. Options. Penguin. Drew. Gleick J. J. Cambridge Univ. Dover Publications.. Cambridge University Press. and Hilbert.. D. 3rd Edition. (1962). Ellis Horwood Ltd. Chaos: Making a New Science. J. (1988). (1993). Fowler A. Astrophys. I.19... C. Fletcher C. L. Fluid Mech. (1957). Hull J.. Fenner R. Press. Fletcher. Redwood. SpringerVerlag. Farlow S. Springer. Computational Techniques for Fluid Dynamics. Futures and Other Derivatives. Rev. London. An Introduction to Chaotic Dynamical Systems. Crank J. Mathematical modelling of twophase flow. (1970). Cambridge University Press. 24983 (1982). (1997).2 Statistics Probability and Statistics Courant R. Ann. D.. A. T. GmbH. Vol. 15. The Nature of Mathematical Modeling.. Goldreich P. 2 volumes. Handbook of Stochastic Methods.. Gershenfeld N. (1989). Rev. (1998). Teach Yourself Statistics. (2004).J. 261291 (1983). W. Tremaine. (1997). Clarendon Press.. Goodman R.. A. Hinch E. (1997)... Gardiner C. Engineering Elasticity.. (1986). Oxford. New York. (1991). C.. Partial Differential Equations for Scientists and Engineers.
. The Finite Element Method in Heat Transfer Analysis. Advanced Engineering Mathematics.. Korn G. A. Springer.. 6th Edition. B. 307 .. M. Handbook in Mathematical Finance. R. Springer. W.fcGrawHill. Thomas H. New York.. Wiley & Sons. (1961).. J..fusiela NL. (1971 ). Kreyszig E. Nlorgan K. (1985). Mathematical Handbook for Scientists and Engineers. Seetharamu S. Finite Volume Methods for Hyperbolic Problems. Cambridge Univ Press. Jouini E. New York. Freeman ( 1982)... Finite Element Handbook. Cambridge University Press. Kardestruncer H. Sneyd J. Cvitanic J. Keener J. (1988).. (1996). (2002). Pergamon Press. Advanced Engineering Mathematics. Partial Differential Equations.. and l\.H. John F. Finite Difference Method in Partial Differential Equations. and Norrie D..Probability and Statistics 19. W. A Mathematical Physiology. New York. Oxford. K.. H. New York. The Fractal Geometry of Nature. (2002).2 Statistics Jeffrey A. H P. LeVeque R. (2001). Academic Press. (2001). Mandelbrot B.. SpringerVerlag. (1987). Computational Partial Differential Equations: Numerical Methods and Diffpack Programming. and Korn T. Langtangen. Wiley & Sons. (1999). F.. Vols. 1/11. Kant T. Lewis R.. (1980). Wiley & Sons. and Griffiths D. l\. Dover Publication. Mitchell A. Finite Elements in Computational Mechanics.
. Applied Partial Differential Equations. Englewood Cliffs. P. 183200 (2006)... F.. Numerical Recipe in C++: The Art of Scientific Computing.. Pallour J.. NJ. (1983). (1998). A first Course in Probability.. CRC Press. Selby S. D. New York.. 308 . Cambridge University Press. Pearson C. Ross S. 682. Cambridge University Press (2006). (2004).. Standard Mathematical Tables.. Mathematical Methods for Physics and Engineering. Strang G. PrenticeHall.19.. S.. Press W. Englewood Cliffs.. Teukolsky S. J. Complex Variables for Scientists and Engineers. New York. Van Nostrand Reinhold. (1974).. (1990). B. 3rd Edition. (1973).D. (1998). Lacey A. Ockendon J. lVIacmillan Publishing Co. E... and Fix G. T. Puckett E. Vetterling W.. SIAM. H. 2nd Ed. Oxford University Press. Sicardy B. Lecture Notes in Physics... SpringerVerlag. (2002).. An Analysis of the Finite Element Method.. Numerical Computing with MATLAB. Howison S. Mathematical Biology. (2005).2 Statistics Probability and Statistics Ivloler C. 5th Edition. Probability and statistics. and lVIeadows D.. P. I\1. Papoulis A. (2003). Riley K. A. (1990). J. G. and Bence S.. Finite Difference Methods for Computational Fluid Dynamics. 2nd Edition. P. Flannery B. PrenticeHall. Hobson M. Handbook of Applied Mathematics. Dynamics of Planetary Rings. and lVIovchan A. Cambridge University Press. Colella. I\1urray J.
(1943). (1972). Cellular automata. http:/ jen. (1982). Weisstein E. John Wiley. Oxford.2 Statistics Smith. New York.wolfram. Zienkiewicz 0 C and Taylor R L.. W. Yang X. The Finite Element Method. D. G. (2005). (1997). K. & Schubert. New York.. D. 1st. An Introduction to Computational Fluid Dynamics: The Finite Volume Method.. (2006). Malalasekra W. (1991).com Wikipedia.wikipedia. 1vialalasekra W... Versteeg H. Tokyo. IvicGrawHill. G.. Terzaghi. 309 . edited by Olarius S. S. S. Theorectical Soil Mechanics... Numerical Solutions of Partial Differential Equations: Finite Differerence Methods. An Introduction to Computational Engineering with Matlab. Berlin. Cambridge Int.Probability and Statistics 19. PDEs and pattern formation (Chapter 18). Chapman & Hall I CRC.. Clarendon Press. Geodynamics: Application of Continuum Physics to Geological Problems. in Handbook of Bioinspired Algorithms. ed. and Zomaya A. Turcotte. Thomee V. SpringerVerlag. R. Prentice Hall. Galerkin Finite Element Methods for Parabolic Problems. Science Publishing. vol. L. 3rd ed. http:/ jmathworld.. Advanced Engineering Mathematics. 4th Edition. Wiley. (1995). Yang X..com Wylie C. 1/11. Young Y... K. (1985)..
.
We list these formulas for your easy reference... + uv<n). +. A..Appendix A Mathematical Formulas This is a summary for the mathematical formulas that have appeared in various sections in this book..l Differentiations and Integrations 1 1 Differentiation Rules: (uv) = u v + uv v2 1 ( ~)I= v {f[g(x)]} = Leibnitz's Theorem: 1 u v.r)! 1 b dv u dx dx = [uv] a + l 1 b du v dx dx 311 . Integration by parts )u(nr)v(r) ( n) r b = n! r!(n.uv 1 I I ! [g(x)] · g1(x) d~n (uv) = u(n)v + nu<n!)v' + . + ( .
A .3J + 5f . + .y)dy = [u(x.. z [ei1r +1= 0].... z3 zs z3 = 2f + 41 .a)d ]+ dy X a(x) X X a(x) 0X Power Series esin z sinh z ~ = 1 + z + .Iathematical Formulas Differentiation of an integral d 1b(x) db da 1b(x) {)u(x.. . y) d u(x.. ..2 Vectors and !via trices !v. 21 n. = x + iy = rei 0 = r(cosO + isinO) De Jvloivre's formula: [r( cos 0 + i sin O)t = rn( cos nO+ i sin nO) A. z2 z4 Complex Numbers ei 0 =cosO+ isinO. zs cos z z2 zn (z E C) 1 = z .b)d u(x.2 Vectors and Matrices Dot Product Cross Product i j a2 a x b = nlallbl sin 0 = Eijkajbk = a1 bt b2 k a3 b3 312 ... + 1 ... = z + 3J + 5f + ...
eig(AB) = eig(BA) 313 .(a· b)c Divergence Theorem of Gauss fffv V·udV= jfsu·dS Stokes's Theorem Green's Theorems f (udx + vdy) Identities = If 8v au (ax .V 2 u Inverse. c] = a · (b x c) = bt CJ b2 C2 b3 C3 a· (b x c)= b · (c x a)= c ·(ax b)= a· (c x b) ax (b x c)= (a· c)b...1\t/athematical Formulas A . v XV¢>= 0 V X (if>u) = (j>V Xu+ (V¢) xu V · (¢u) = 4>V · u + (V4>) · u V x (V x u) = V(V · u) .. IZI. v Xu= 0..2 Vectors and lvlatrices Vector Triple a1 a2 a3 [a.ay )dxdy v.. Trace and Determinants IAB . IAI = detA A vi= Aivi. b. ZI = IAIIBI .
r V7n+T' (n > 0). (if AB = BA) n=1 eAeB = eA+B ~etA dt = AetA = etA A (eA)1 = eA... .Jt) 2 2a 2 Gaussian Distribution p(x.. 1 1 00 00 oo e QX2dx= If 0: oo 2n ax2 x e dx = (1)n ·1· 3· · · (2n 1) 2n r:.p ) = (nn! k)tkt p k(1 . n. det(A . n=O n. n ) 314 ..p )nk ' (k = 0' 1' 2' . oo L 00 1 ~ ln(IA) =L (1)n1 nl 1 1 An= A. a)= ~ exp[27ra2 1 ] + <5(x) l1+ 00.Iathematical Formulas tr(AB) = tr(BA).3 Asymptotics !v. det(eA) = etrA A. Binomial Distribution B(k .1) = 1 tr(A +B) = tr(A) + tr(B) det(AB) = det(A)det(B) det(A)' Exponential Matrices etA= 2 1 (tAt = l+tA+ 2 A + ..A..2A2 + 3A3 + . Jt.3 Asymptotics (x.
A>> 1) : : : : p(x. 1 r( 2) = v'1r 1 r(x + 1)::::::: (=)xJ27rX. a). B(k ~ p. p)l np»l "'p(x.p). A. J. a 2 = np(1 .L. = A. p. Poisson Distribution e. Stirling's Formula nl::::::: (x ~ oo). r(e Gamma Function 2) = 2y'7r. p. e n >> 1 Error Functions 315 . = np.XAx f(x. x EN a 2 =A (lim np =A) noo f(x...1\t/athematical Formulas A .. 1 X. (~tV27ffi.p) "'f(x. a). x. n ~ oo.A) = .4 Special Integrals B(x. A= np).4 Special Integrals r(x) = fooo txletdt = fooo et+(x1) Intdt r(n + 1) = nl. n ~ oo.
151 bifurcation. 281 birthday paradox. 153 brachistochrone problem. 160 partial. 67 analytical function. 38 curvature. 156 central difference. 15 correlation coefficient. 284 curl. 192 algorithms. 15 spherical. 248 natural. 70 complex variables. 62 coordinates cylindrical. 159 shortest path. 248 multiple variables. 30 cumulative probability function. 210 analytic function. 2 problem. 67 assembly by element. 246 asymptotic. 165 pendulum. 261 Airy stress function. 88 Bessel function. 9 curvature. 254. 252 essential. 314 Gamma function. 84 calculus of variations. 205 boundary condition. 4 constraint. 49 difference equation. 15 polar. 287 chaos. 212 central limit theorem. 95 differential operator. 163 Dido's hanging rope problem.Index 1D. 295 cross product. 315 Bessel equation. 24. 209. 99 complex integral. 273 BlackScholes equation. 1 implicit. 163 vector. 153 rule. 314 Stirling's formula. 164lifferentiation. 151 error function. 255 2D. 153 determinant. 70 residue. 32 316 . 99 binomial distribution.
248 sparable kernel. 286 integral equation. 71 differentiation. 297 elastostatic. 223. 224 elastic wave. 19J. 198 EulerBernoulli theory. 168. 38 Green's function. 12 elliptic equation. 146 displacement. 144 Gauss's theorem. 73 EulerLagrange equation. 196 index matrix. 17 exponential distribution. 181. 185. 219 Laplace transform. 41 Green's theorem. 134 317 diffusion equation. 154 special. 210 residue theorem. 131 GaussSeidel iteration. 149 hyperbolic equation. 184 Cauchy's theorem. 243 Hooke's law 181 inner product. 238 Fourier. 215 beam bending. 30 DuFortFrankel scheme. 12 strain tensor. 102 gradient. 167 finite element method. 148 Green's identity. 240 secondorder. 214 elasticity. 28 ' theorem. 41 Laplace. 28. 125 Gamma function. 182 stressstrain relationship. 102 dynamical system. 313 . 221 Volterra equation. 143 Gaussian distribution. 205 integral transform Galerkin method. 167 displacement kernel. 65 CauchyNavier equation. 253 hybrid method. 182 Bessel function. 313 dot product.INDEX INDEX harmonic motion. 224 dynamic reconstruction. 38 Euler scheme. 20 line. 153. 197 hyperbolic function. 239 divergence. 209 Fredholm equation. 109. 88 stress tensor. 227. 18 error function. 283 wavelet. 203 firstorder. 169 finite difference method. 159 heat conduction. 139. 185 integral MaxwellBetti multiple. 38 divergence theorem.hypothesis testing. 218 Gaussian. 22 Fourier transform. 169 finite volume method. 170 FokkerPlank equation.
213 matrix. 277 318 . 203. 6 iteration method. 270 Maxwell's equations. 276 NavierStokes equation.INDEX INDEX integration. 267 inverse. 95 linear system. 103 ordinary differential equation complementary function. 283 normal modes. 206 random variable. 56 lognormal distribution. 81 oscillation damped. 264 instability. 244 Hermitian. 280 moment. 85 particular integral. 52 Poisson's equation. 263 pattern formation . 206 permutation. 277 node. 204 conditional. 158 Lame constant. 260 bifurcation. 271. 201. 278 distribution. 238 Leibnitz theorem. 81 homogenous equation. 272 NavierStokes equations. 283 probability density function. 139. 267 normal distribution. 116 null hypothesis. 119 undamped. 5 by parts. 81 general solution. 109 outer product. 112 forced. 81 linear system. 201 PDE. 275 mean. 77. 286 Lorenz attractor. 271 moment generating function. 81 random variables. 201 Laurent series. 301 quadratic form. 69 leapfrog scheme. 281 exponential. 80. 112 small amplitude. 268. 3 linear difference equation. 138. 280 l\1onty hall problem. 53 probability. 141. 219 Jacobian. 292 independent events. 55 ODE. 109 natural frequency. 33 Lagrangian. 279 method of least square. 212 leastsquare. 50 axiom. 30 parabolic equation. 184 Laplace's equation. 13 kinematics. 202. 216 pattern formation. 47 Poisson distribution. 261 mathematical model. 242 randomness.
297 linear regression. 173 Residue theorem. 242 Stokes's theorem. 204 Taylor series. 289 sample variance. 29 triple. 262 SturmLiouville equation. 27. 173 Rien1ann hypothesis. 245 stiffness matrix. 289 steady state. 86 reactiondiffusion equation. 254 power. 43 three card problem. 41 stress intensity factor. 259 shape functions. 69 recurrence equation. 175 RungeKutta method. 175 Riemann (function. 77 Cartesian. 211 asympottic.INDEX INDEX reactiondiffusion. 17 transient. 214 soliton. 141 explicit. 147 Talor. 109 wave equation. 292 sample mean. 105 timestepping. 147 stability condition. 251 selfsimilarity. 203. 31 vector calculus. 8 travelling wave. 139. 8 Travelling wave. 238 triangular element. 211. 206 timedependent. 285 statistics. 176 Saturn's rings. 240 similarity solution. 207 upwind scheme. 257. 195 Student's tdistribution. 278 vector. 69 rank. 286 SineGordon equation. 95 tensor. 276 Schrodinger equation. 215. 175 Riccati equation. 254 series implicit. 210. 217. 289 confidence interval. 181 319 . 299 Student's ttest. 236 Young's modulus. 213 vector. 255 weak forn1ulation. 202. 268 vibration. 73 analysis. 301 variance. 293 maximum likelihood. 145 uniforn1 distribution. 215 standard normal distribution. 32 Venn diagram. 68 notations. 253 separation of variables.
This action might not be possible to undo. Are you sure you want to continue?
We've moved you to where you read on your other device.
Get the full title to continue reading from where you left off, or restart the preview.