Numerical solution of partial differential

equations

Dr. Louise Olsen-Kettle

The University of Queensland
School of Earth Sciences
Centre for Geoscience Computing

Email: [email protected]
Web: http://researchers.uq.edu.au/researcher/768
@DrOlsenKettle

ISBN: 978-1-74272-149-1
 Acknowledgements

Special thanks to Cinnamon Eliot who helped typeset these lecture notes in
LATEX.

Note to reader
This document and code for the examples can be downloaded from

http://espace.library.uq.edu.au/view/UQ:239427.

Please note if any of the links to code are not working please contact my
email address and I can send you the code.
Contents

1 Overview of PDEs 9
1.1 Classification of PDEs . . . . . . . . . . . . . . . . . . . . . . 9
1.1.1 Elliptic . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.1.2 Hyperbolic . . . . . . . . . . . . . . . . . . . . . . . . 9
1.1.3 Parabolic . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.2 Implicit Vs Explicit Methods to Solve PDEs . . . . . . . . . . 10
1.3 Well-posed and ill-posed PDEs . . . . . . . . . . . . . . . . . 10

I Numerical solution of parabolic equations 12

2 Explicit methods for 1-D heat or diffusion equation 13

2.1 Analytic solution: Separation of variables . . . . . . . . . . . 13
2.2 Numerical solution of 1-D heat equation . . . . . . . . . . . . 15
2.2.1 Difference Approximations for Derivative Terms in
PDEs . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2.2 Numerical solution of 1-D heat equation using the
finite difference method . . . . . . . . . . . . . . . . . 16
2.2.3 Explicit Forward Euler method . . . . . . . . . . . . . 17
2.2.4 Stability criteria for forward Euler method . . . . . . . 20
2.3 Method of lines . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.3.1 Example . . . . . . . . . . . . . . . . . . . . . . . . . . 21

3 Implicit methods for 1-D heat equation 23

3.1 Implicit Backward Euler Method for 1-D heat equation . . . . 23
3.1.1 Numerical implementation of the Implicit Backward
Euler Method . . . . . . . . . . . . . . . . . . . . . . . 24
3.1.2 Dirichlet boundary conditions . . . . . . . . . . . . . . 24
3.1.3 Mixed boundary conditions . . . . . . . . . . . . . . . 24
3.2 Crank-Nicolson Scheme . . . . . . . . . . . . . . . . . . . . . 26

2
4 Iterative methods 28
4.1 Jacobi method . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.1.1 Example . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.1.2 Applying the Jacobi method . . . . . . . . . . . . . . . 30
4.2 Gauss-Seidel Method . . . . . . . . . . . . . . . . . . . . . . . 31
4.2.1 Example: using Gauss-Seidel method to solve a
matrix equation . . . . . . . . . . . . . . . . . . . . . . 31
4.3 Relaxation Methods . . . . . . . . . . . . . . . . . . . . . . . 32

5 2-D Finite Difference 34

5.1 2-D Poissons equation . . . . . . . . . . . . . . . . . . . . . . 34
5.2 2-D Heat (or Diffusion) Equation . . . . . . . . . . . . . . . . 38
5.2.1 Alternating Direct/Implicit method for the 2-D heat
equation . . . . . . . . . . . . . . . . . . . . . . . . . 39
5.3 Cylindrical and spherical polar co-ordinates . . . . . . . . . . 40
5.3.1 Example: Temperature around a nuclear waste rod . . 42

II Numerical solution of hyperbolic equations 47

6 Analytical solutions to the 1-D Wave equation 48

6.1 1-D Wave equation . . . . . . . . . . . . . . . . . . . . . . . . 48
6.2 dAlemberts solution . . . . . . . . . . . . . . . . . . . . . . . 48
6.3 Separation of variables . . . . . . . . . . . . . . . . . . . . . . 49

7 Flux conservative problems 51

7.1 Flux Conservative Equation . . . . . . . . . . . . . . . . . . . 51
7.2 Stability analysis of numerical solutions of the first order
flux conservative or 1-D advection equation . . . . . . . . . . 51
7.3 Forward Time Centred Space (FTCS) . . . . . . . . . . . . . 52
7.3.1 von Neumann stability analysis of FTCS method . . . 52
7.4 Lax Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
7.4.1 von Neumann Stability Analysis of Lax Method . . . 53
7.5 Courant Condition . . . . . . . . . . . . . . . . . . . . . . . . 54
7.6 von Neumann Stability Analysis For Wave Equation . . . . . 55
7.6.1 Lax method . . . . . . . . . . . . . . . . . . . . . . . . 55
7.7 Other sources of error . . . . . . . . . . . . . . . . . . . . . . 56
7.7.1 Phase Errors (through dispersion) . . . . . . . . . . . . 56
7.7.2 Dispersion in the numerical solution of the 1-D
advection equation using the Lax method . . . . . . . 57
7.7.3 Error due to nonlinear terms . . . . . . . . . . . . . . . 59

3
 7.7.4 Aliasing error . . . . . . . . . . . . . . . . . . . . . . . 59

8 Numerical Solution of 1-D and 2-D Wave Equation 61

8.1 Explicit Central Difference for 1-D Wave Equation . . . . . . 61
8.1.1 Example: plucking a string . . . . . . . . . . . . . . . 61
8.1.2 1-D Wave Equation with Friction . . . . . . . . . . . . 65
8.2 2-D Wave Equation . . . . . . . . . . . . . . . . . . . . . . . . 68
8.2.1 Example: vibrations of a thin elastic membrane fixed
at its walls . . . . . . . . . . . . . . . . . . . . . . . . . 68
8.2.2 Examples of wave equation . . . . . . . . . . . . . . . 71

9 Finite element method 73

9.1 An introduction to the Finite Element Method . . . . . . . . 73
9.2 Comparing FEM solution to FD solution for our example . . 77
9.2.1 FD solution . . . . . . . . . . . . . . . . . . . . . . . . 77
9.3 2-D Finite Element Method . . . . . . . . . . . . . . . . . . . 78
9.3.1 2-D hat functions . . . . . . . . . . . . . . . . . . . . 79
9.3.2 Example: 2-D Finite Element Method using eScript
for elastic wave propagation from a point source. . . . 80

10 Spectral methods 83
10.1 An introduction to spectral methods . . . . . . . . . . . . . . 83
10.1.1 Example 1: Comparing the accuracy of solutions of a
variable speed wave equation with either the spectral
or finite difference method . . . . . . . . . . . . . . . . 84
10.1.2 Example 2 Comparing spectral and finite difference
methods with constant wave speed conditions and
initial conditions of a non-smooth pulse . . . . . . . . . 87

III Nonlinear partial differential equations 89

11 Shock wave 90
11.1 Analytical solution: Method of characteristics . . . . . . . . . 90
11.1.1 Example 1: Using method of characteristics to solve
the linear 1-D advection equation . . . . . . . . . . . . 91
11.1.2 Example 2: Using method of characteristics to solve
the nonlinear inviscid Burgers equation . . . . . . . . 91
11.2 Numerical Solution for nonlinear Burgers Equation . . . . . . 94
11.2.1 Example I: Finite difference solution with Lax Method 95
11.2.2 Example II: Solution using Method of Lines . . . . . . 97

4
 11.2.3 Example III: Solution using Spectral Method . . . . . . 97

12 Korteweg-de Vries Equation 100

12.1 Solitons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
12.2 Analytical solution . . . . . . . . . . . . . . . . . . . . . . . . 100
12.3 Numerical solution of KdV Equation . . . . . . . . . . . . . . 103
12.3.1 Solving directly with Spectral Method . . . . . . . . . 104
12.3.2 Modifying Uxxx term causing instabilities in direct
spectral method . . . . . . . . . . . . . . . . . . . . . . 105
12.3.3 Interacting Solitons . . . . . . . . . . . . . . . . . . . . 106

5
List of Figures

2.1 Initial conditions in (a) and matlab solution using Forward

Euler method for temperature distribution along rod with
time in (b) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

3.1 Initial conditions in (a) and matlab solution using Backward

Euler method for temperature distribution along rod with
time in (b) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

4.1 Plot of residual using the Gauss-Seidel method after each

iteration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

5.1 3-D Cylindrical Co-ordinates . . . . . . . . . . . . . . . . . . . 41

5.2 3-D Spherical Polar Co-ordinates . . . . . . . . . . . . . . . . 41
5.3 Initial conditions in (a) and matlab solution using Backward
Euler method for temperature distribution near nuclear rod
at different time intervals in (b) . . . . . . . . . . . . . . . . . 46

7.1 Solution at t = 1 using the Lax method with different time

steps, (a) t = x/2c where dispersion is present but the
pulse matches the analytical solution for the speed of the
wave, (b) t = x/c where no dispersion is present and
numerical solution matches analytical solution exactly, and
(c) t = 1.001x/c where the Courant condition is not met
and solution is becoming unstable. . . . . . . . . . . . . . . . 59
7.2 Aliasing error occurs when the mesh spacing x is too large
to represent the smallest wavelength 1 and misinterprets it
as a longer wavelength oscillation 2 . . . . . . . . . . . . . . 60

8.1 Initial conditions in (a) and matlab solution using explicit

central difference method for 1D wave equation in (b) . . . . 64
8.2 DAlemberts solution in (a) and error using numerical
matlab solution using explicit central difference method for
1D wave equation in (b) . . . . . . . . . . . . . . . . . . . . . 65

6
8.3 Initial conditions in (a) and matlab solution using explicit
central difference method for 1D wave equation with friction
in (b) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

9.1 FEM mesh with triangles . . . . . . . . . . . . . . . . . . . . 74

9.2 FEM mesh with triangles . . . . . . . . . . . . . . . . . . . . 78
9.3 2D hat function
(j (xj , yj ) = 1, j (xi , yl ) = 0 if i 6= j and j 6= l) . . . . . . . 80
9.4 Plot of Euclidean normal of the displacement at t > 0 for a
point source using eScript. . . . . . . . . . . . . . . . . . . . 81

10.1 Numerical solution for 1D advection equation with initial

conditions of a smooth Gaussian pulse with variable wave
speed using the spectral method in (a) and finite difference
method in (b) . . . . . . . . . . . . . . . . . . . . . . . . . . 87
10.2 Numerical solution for 1D advection equation with initial
conditions of a box pulse with a constant wave speed using
the spectral method in (a) and finite difference method in (b) 88

11.1 The analytical solution U (x, t) = f (x U t) is plotted to

show how shock and rarefaction develop for this example . . . 95
11.2 Initial conditions in (a) and solution for nonlinear Bugers
equation using the Lax method in (b) . . . . . . . . . . . . . . 96
11.3 Initial conditions in (a) and solution for nonlinear Bugers
equation using the method of lines in (b) . . . . . . . . . . . . 98
11.4 Initial conditions in (a) and solution for nonlinear Bugers
equation using the spectral method in (b) . . . . . . . . . . . 99
q
12.1 The initial conditions U (x, 0) = f (x) = Asech2 ( 12 A
x). . . . . . 102
12.2 Initial conditions in (a) and solution for nonlinear KdV
equation using the direct spectral method in (b) . . . . . . . . 105
12.3 Initial conditions and final solution after one period in (a)
and solution for nonlinear KdV equation using a modified
spectral method in (b) . . . . . . . . . . . . . . . . . . . . . . 106
12.4 The initial conditions U (x, 0) = f (x) is plotted to show the 2
solitions and their speeds . . . . . . . . . . . . . . . . . . . . . 107
12.5 Initial conditions and final solution after one period in (a)
and solution for nonlinear KdV equation for two interacting
solitons in (b) . . . . . . . . . . . . . . . . . . . . . . . . . . . 107

7
Numerical solution of parabolic
and hyperbolic PDEs

This document and code for the examples can be downloaded from

http://espace.library.uq.edu.au/view/UQ:239427.
Please contact me ([email protected]) if you are unable to download
the code and I can send it to you.
References:
Applied Numerical Methods for Engineers using Matlab and C, R. J. Schilling
and S. L. Harris.

Computational Physics Problem solving with computers, R.H. Landau

and M. L. Paez.

An Introduction to Computational Physics, T. Pang.

Numerical Recipes in Fortran (2nd Ed.), W. H. Press et al.

Introduction to Partial Differential Equations with Matlab, J. M. Cooper.

Numerical solution of partial differential equations, K. W. Morton and

D. F. Mayers.

Spectral methods in Matlab, L. N. Trefethen

8
Chapter 1

Overview of PDEs

1.1 Classification of PDEs

The classification of PDEs is important for the numerical solution you choose.

A(x, y)Uxx + 2B(x, y)Uxy + C(x, y)Uyy = F (x, y, Ux , Uy , U )

1.1.1 Elliptic

AC > B 2

For example, Laplaces equation:

Uxx + Uyy = 0

A = C = 1, B = 0

1.1.2 Hyperbolic

AC < B 2

For example the 1-D wave equation:

1
Uxx = Utt
c2

A = 1, C = 1/c2 , B = 0

9
1.1.3 Parabolic

AC = B 2

For example, the heat or diffusion Equation

Ut = Uxx

A = 1, B = C = 0

1.2 Implicit Vs Explicit Methods to Solve PDEs

Explicit Methods:
possible to solve (at a point) directly for all unknown values in the
finite difference scheme.

stable only for certain time step sizes (or possibly never stable!). Sta-
bility can be checked using Fourier or von Neumann analysis. Time
step size governed by Courant condition for wave equation.

Implicit Methods:
there is no explicit formula at each point, only a set of simultaneous
equations which must be solved over the whole grid.

Implicit methods are stable for all step sizes.

1.3 Well-posed and ill-posed PDEs

The heat equation is well-posed Ut = Uxx . However the backwards heat
equation is ill-posed : Ut = Uxx at high frequencies this blows up!
In order to demonstrate this we let U (x, t) = an (t) sin(nx)
then:

Uxx = an (t)n2 sin(nx), and Ut = an (t) sin(nx)

U =U an (t) sin(nx) = an (t)n2 sin(nx)

| t {z xx}
Heat Equation

10
 2t
an = an n2 an (t) = an (0)en

For the heat equation the transient part of the solution decays and this has
stable numerical solutions.

U = Uxx an (t) sin(nx) = an (t)n2 sin(nx)

|t {z }
Backwards Heat Equation

2t
an = an n2 an (t) = an (0)en

For the backwards heat equation the transient part of the solution blows up
and the numerical solution would fail! In general it is difficult or impossible
to obtain numerical solutions for ill-posed PDEs.

11
 Part I

Numerical solution of parabolic

equations

12
Chapter 2

Explicit methods for 1-D heat

or diffusion equation

We will focus on the heat or diffusion equation for the next few chapters.
This is an example of a parabolic equation.

2.1 Analytic solution: Separation of variables

First we will derive an analtical solution to the 1-D heat equation. Consider
the temperature U (x, t) in a bar where the temperature is governed by the
heat equation, Ut = Uxx . The ends of the bar are cooled to 0 C and the
initial temperature of the bar is 100 C.

U (0, t) = 0 C -  U (L, t) = 0 C
6
U (x, 0) = 100 C
We want to solve Ut = Uxx using separation of variables. We assume
that the solution can be written as the product of a function of x and a
function of t, ie. U (x, t) = X(x)T (t) then:
T 2X
Ut = X = T = Uxx divide by XT
t x2
1 T 0 (t) X 00 (x) 2
= = (2.1)
T (t) X(x) | {z }
| {z } | {z } constant
function of t only function of x only

13
The only way the LHS and RHS of equation 2.1 can be a function of t and
x respectively is if they are both equal to a constant which we define to be
2 for convenience.

2
T 0 + 2 T = 0 T = e t
X 00 + 2 X = 0 X = A sin x + B cos x

Use boundary conditions U (t, 0) = 0 X(0) = 0 = B

U (t, L) = 0 X(L) = 0 = A sin

n
t = n = , n = 1, 2, . . .
L

U (x, t) = X(x)T (t)

X nx 2 t
= An sin( )e
n=1 L

2
e t is a transient solution and decays in time to boundary conditions.
Use initial conditions U (0, x) = 100 C to find An :

X
U (0, x) = T0 = An sin (nx/L)
n=1

Use orthogonality: 0L sin( nx ) sin( mx

R
L L
)dx = nm
and cos(m) 1 = 0, for m = 0, 2, 4, . . .
and cos(m) 1 = 2, for m = 1, 3, 5, . . .

Am = T0 [L/m(cos(m) 1)]
2L
= T0
m

for m=1,3,5,. . .

14
2.2 Numerical solution of 1-D heat equation
2.2.1 Difference Approximations for Derivative Terms
in PDEs
We consider U (x, t) for 0 x a, 0tT
Discretise time and spatial variable x:

T a
t = , x = ,
m n+1

tk = kt, 0 k m xj = jx, 0j n+1

Let Ujk = U (xj , tk )

Consider Taylor series expansion for Ujk+1 :

Ujk t2 2 Ujk
Ujk+1 = Ujk + t + + O(t3 ) (2.2)
t 2 t2
If we only consider O(t) terms in equation 2.2 then we arrive at the forward
difference in time approximation for Ut :

Ujk Ujk+1 Ujk

= + O(t)
t t
We can also derive a higher order approximation for Ut if we consider the
Taylor series expansion for Ujk1 as well:

Ujk t2 2 Ujk
Ujk1 = Ujk t + + O(t3 ) (2.3)
t 2 t2

Ujk Ujk+1 Ujk1

2.2 2.3 = + O(t2 ) leap-frog (or centred difference) in time.
t 2t

This gives higher order accuracy than forward difference.

15
 We can also perform similar manipulations to arrive at approximations
for the second derivative Utt :

2 Ujk Ujk+1 + Ujk1 2Ujk

2.2 + 2.3 2Ujk = + O(t2 ) central difference
t2 t2

The finite difference method makes use of the above approximations to

solve PDEs numerically.

2.2.2 Numerical solution of 1-D heat equation using

the finite difference method

Ut = Uxx

Initial conditions

U (0, x) = f (x)

Types of boundary conditions

Neumann boundary conditions

Ux (t, 0) = g1 (t)
Ux (t, a) = g2 (t)

Dirichlet boundary conditions

U (t, 0) = g1 (t)
U (t, a) = g2 (t)

Mixed boundary conditions

U (t, 0) = g1 (t)
Ux (t, a) = g2 (t)

16
2.2.3 Explicit Forward Euler method
or FTCS (Foward Time Centred Space)
We want to solve the 1-D heat equation:
Ut = Uxx . (2.4)
We solve this PDE for points on a grid using the finite difference method
where we discretise in x and t for 0 x a and 0 t T :
xn+1 = a
xn

..
.

x3
x2
x1
x0
t0 t1 t2 t3 ... tm1 tm = T

We discretise in time with time step: t = T /m and in space with grid

spacing: x = a/(n + 1), and let tk = kt where 0 k m and xj = jx
where 0 j n + 1.
Let Ujk = U (xj , tk ) then the finite difference approximations for equation
(2.4) are given by:
U (xj , tk ) Ujk+1 Ujk
= + O(t), Forward Euler method for time derivative,
t t
k
2 U (xj , tk ) Uj+1 2Ujk + Uj1
k
= + O(x2 ),
x2 x2
Central difference method for spatial derivative.
Our discretised PDE (equation (2.4)) becomes:
Ujk+1 Ujk  k k k

= 2
Uj+1 2U j + Uj1 ,
t x
 
or Ujk+1 = s Uj+1
k k
+ Uj1 + (1 2s)Ujk ,
t
where s = .
x2
17
Ujk+1 is the solution for the temperature at the next time step.
Suppose we have initial conditions U (x, 0) = Uj0 = f (xj ), and mixed
boundary conditions:

Dirichlet boundary conditions at x = 0: U (0, t) = U0k = g1 (tk ),

k
Un+1
Neumann boundary conditions at x = a: Ux (a, t) = = g2 (tk ),
x

Numerical implementation of Explicit Forward Euler method

Solving equation (2.4): Ut = Uxx with:
initial conditions: Uj0 = f (xj ) = U (x, t = 0),
Dirichlet boundary conditions at x = 0: U (x = 0, t) = U0k = g1 (tk ) and
k
Neumann boundary conditions at x = a: U (a, t)/x = Un+1 /x = g2 (tk ).
To solve using the Neumann boundary condition we need an extra step:
k k
Un+1 Un+1 Unk
= g2 (tk ),
x x
k
or Un+1 = xg2 (tk ) + Unk . (2.5)

We can write out the matrix system of equations we will solve numerically
for the temperature U . Suppose we use 5 grid points x0 , x1 , x2 , x3 , x4 = xn+1 ,
ie. n = 3 in this example:

x0 = 0 x1 x2 x3 x4 = xn+1 = a

We let:
U1k

~k
U
k ~ k at time tk .
= U2 , solution for temperature vector U
k
U3

The boundary conditions give U0k = U (x = 0, tk ) and Un+1

k
= U4k = U (x =
a, tk ).  
We can rewrite Ujk+1 = s Uj+1
k k
+ Uj1 + (1 2s)Ujk in matrix form:

U1k+1 U1k sU0k

1 2s s 0
~ k+1
U
k+1
= U2 =

s 1 2s s k
U2 + 0 (2.6)

U3k+1 0 s 1 2s U3k sU4k

18
 Using boundary conditions: U0k = g1 (tk ) and U4k = xg2 (tk )+U3k equation
(2.6) becomes:

1 2s s 0

U1k+1 U1k

~ k+1 =
s 1 2s s
U U2k+1 = U2k

0 s s}
1| {z

U3k+1 U3k

| {z
Neumann bc }
A

Dirichlet bc
z }| {

sg1 (tk )

+ 0




sxg2 (tk )

| {z }
|
Neumann
{z
bc }
~b
~ k+1
or U = AU + ~b
~k (2.7)

The term (1 s) in the matrix A above and the term (sxg2 (tk )) in
vector ~b above are from the Neumann boundary condition given using the
approximation in equation (2.5).

Matlab code for Explicit Forward Euler method

The matlab code for this example is ForwardEuler.m.
Solving equation (2.4): Ut = Uxx with 0 x 1, = 105 , and
0 t 12, 000, using m = 600 time steps, and n = 39 for 41 grid points in
x.
Initial conditions: U (x, t = 0) = 2x + sin(2x) + 1,
and Dirichlet boundary conditions at x = 0: U (x = 0, t) = 1 and Neumann
boundary conditions at x = 1: U (x = 1, t)/x = 2.
With these boundary conditions we can check that the numerical solution
approximates the steady state solution U (x, t) = 2x + 1 as t . Your
solution using the finite difference code can also be checked using Matlabs
PDE solver (using pdex1).

Figure 2.1 shows the initial conditions in (a) and matlab solution for tem-
perature distribution along rod with time in (b). The numerical solution
matches the analytical solution reasonably well: U (x, t) = 2x + 1 at the final
time. However the Neumann boundary condition at x = 1 introduces error

19
 Initial condition for Temperature distribution Variation of Temperature distribution with time
2.8

2.6
3
2.4
2.5
2.2

U
2
U

1.5
1.8

1.6 1
1
0.8 15000
1.4 0.6
10000
0.4
0.2 5000
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 x 0 0
x t

Figure 2.1: Initial conditions in (a) and matlab solution using Forward Euler
method for temperature distribution along rod with time in (b)

through the approximation to the spatial derivative in this boundary condi-

tion so the numerical solution at x = 1 is not exactly U = 3 but close to
it.

2.2.4 Stability criteria for forward Euler method

Suppose U (x, 0) = f (x) = cos (x/x) where these initial conditions data
oscillate with the same frequency as the grid, x = a/n + 1, As before we
let xj = jx, 0 j n + 1

f (xj ) = cos(j) = (1)j

Using finite difference discretisation of (2.4) Ut = Uxx :

Ujk+1 = s(Uj+1
k k
+ Uj1 ) + (1 2s)Ujk

At first time step (k = 1):

Uj1 = (1 2s)(1)j + s [(1)j+1 + (1)j1 ]

| {z }
2(1)j

= (1 4s)(1)j
At k = 2:

Uj2 = (1 2s)Uj1 + s(Uj+1

1 1
+ Uj1 )

20
 Uj2 = (1 2s)(1 4s)(1)j + s [(1 4s)(1)j+1 + (1 4s)(1)j1 ]
| {z }
2(14s)(1)j

= (1 4s)2 (1)j

Therefore at k = n

Ujn = (1 4s)n (1)j

NB: the term (1 4s) determines stability.

This solution for Ujn will become unbounded as n if |1 4s| 1 or
s > 1/2.
We know the exact solution |U (x, t)| |U (x, t0 )| = x, t.
The Forward Euler method is only stable if s (known as the gain parameter)
satisfies 0 s 1/2 or equivalently the time step satisfies: t x2 /2.
You can check that using the matlab code ForwardEuler.m that when the
time step exceeds this value the numerical solution becomes unstable.

2.3 Method of lines

There are other explicit numerical methods that can be applied to the 1-
D heat or diffusion equation such as the Method of Lines which is used
by Matlab and Mathematica. The trick with the Method of Lines is that
it replaces all spatial derivatives with finite differences but leaves the time
derivatives. It is then possible to use a stiff ordinary differential equation
solver on the time derivatives in the resulting system.

2.3.1 Example
The matlab code for this example is MethodOfLines.m and Uprime.m
We are solving the same system again with the method of lines: Ut = Uxx
where the initial conditions are U (x, 0) = sin(2x) + 2x + 1
0 x 1, = 105 , 0 t 12, 000.
boundary conditions are U (0, t) = 1 and Ux (1, t) = 2
Again we get:
~
U ~ + ~b
= AU
t
How? Replace
Uj+1 2Uj + Uj1
Uxx = ,
x2
21
where U (xj , t) = Uj (t), xj = jx, 0 j n + 1
x = a/(n + 1) = 1/(n + 1) (a = 1)
with boundary conditions: U (0, t) = U0 (t) = 1

U Un+1 Un+1 Un
(1, t) = (t) ' = 2 Un+1 = Un + 2x
x x x

In matrix form for n = 3 elements:

x0 = 0 x1 x2 x3 x4 = 1

~ U1 2 1 0 U1 1
U
= U2 =

1 2 1
U2

+ 0
t x2 x2

U3 0 1 1 U3 2x

~ = AU
U ~ + ~b

~ using ode45 and matlab code MethodOfLines.m and

We solve for U
Uprime.m.

22
Chapter 3

Implicit methods for 1-D heat

equation

3.1 Implicit Backward Euler Method for 1-D

heat equation
Unconditionally stable (but usually slower than explicit methods).
implicit because it evaluates difference approximations to derivatives
at next time step tk+1 and not current time step we are solving for tk .

k+1
Uj+1 2Ujk+1 + Uj1
k+1
Uxx (tk+1 , xj ) =
x2
k+1
Uj Ujk
Ut (tk+1 , xj ) =
t

Ut = Uxx becomes:

t
Ujk = Ujk+1 k+1
[Uj+1 2Ujk+1 + Uj1
k+1
]
x2
= (1 + 2s)Ujk+1 s(Uj+1
k+1 k+1
+ Uj1 ) (3.1)

t
where s = x 2 as before.

We still need to solve for Ujk+1 given Ujk is known This requires solving a
tridiagonal linear system of n equations.
a
Again we let Ujk = U (xj , tk ); xj = jx, j = 0, ..., n + 1, x = n+1 ; tk =
T
kt, k = 0, ..., m, and t = m .

23
3.1.1 Numerical implementation of the Implicit Back-
ward Euler Method
Again we are solving the same problem: Ut = Uxx , U (x, 0) = Uj0 = f (xj )

3.1.2 Dirichlet boundary conditions

U (0, t) = 1 = U0k , k
U (a, t) = U (1, t) = 3 = Un+1 = U4k

For simplicity we consider only 4 elements in x in this example to find

the matrix system we need to solve for:

x0 = 0 x1 = 0.25 x2 = 0.5 x3 = 0.75 x4 = 1

k+1 k+1
Rewriting s[Uj+1 + Uj1 ] + (1 + 2s)Ujk+1 = Ujk as a matrix equation:

U1k+1 U1k U0k+1

1 + 2s s 0
k+1 k

s 1 + 2s s
U2 = U2 + s 0

0 s 1 + 2s U3k+1 U3k U4k+1

| {z } | {z } | {z }
Tridiagonal matrix Solution U at next time step given from b.c.

~ k + ~b
~ k+1 = U
AU
~ k + ~b]
~ k+1 = A1 [U
U

3.1.3 Mixed boundary conditions

The matlab code for this example is BackwardEuler.m.
k
Un+1
U (0, t) = 1 = U0k , Ux (1, t) = 2 = x

Using a leap-frog approximation for the spatial derivative:

Ujk k
Uj+1 k
Uj1
= + O(x2 )
x 2x
24
This is more accurate than the forward approximation we used previously
(see section 2.2.1):

Ujk k
Uj+1 Ujk
= + O(x)
x x

So for the Neumann boundary condition we have:

k k
Un+1 Un+2 Unk
Ux (1, t) = =2 (3.2)
x 2x
With n = 4, U5k is called a ghost point because it lies outside the bar. So
using equation 3.2 we define U5k :

U5k = 4x + U3k
k
Because we have Neumann boundary conditions at x = a(= 1), Un+1 = U4k
is unknown, and given by equation 3.1:

U4k = s[U5k+1 + U3k+1 ] + (1 + 2s)U4k+1

use U5k+1 = 4x + U3k+1
U4k = s[4x + 2U3k+1 ] + (1 + 2s)U4k+1

Our system of equations becomes:

1 + 2s s 0 0
U1k+1 U1k

s 1 + 2s s 0

U2k+1 U2k


0 s 1 + 2s s =

U3k+1 U3k

0 0 2s 1 + 2s

U4k+1 U4k
| {z }
Neumann b.c. | {z } | {z }
| {z } ~ k+1
U ~k
U
A

Dirichlet b.c.
z }| {

sU0k+1

0

+



0

4sx

| {z }

|
Neumann
{z
b.c. }
~b

~ k + ~b = ~c
~ k+1 = U
AU
~ k+1 = A1~c
U

25
Using an implicit solver means we have to invert the matrix A to solve for
~ k+1 which is a lot more computationally expensive than the matrix multi-
U
~ k+1 for explicit solvers in section 2.2.3.
ply operation in equation 2.7 to find U
This method is stable for s 0 so larger time steps can be used for implicit
methods than explicit methods.

The matlab code for this example is BackwardEuler.m.

Initial condition for Temperature distribution Variation of Temperature distribution with time using Backward Euler method
3

2.8
3
2.6

2.4 2.5

2.2 U 2
U

2
1.5
1.8
1
1.6 1
0.8 15000
1.4 0.6
10000
0.4
0.2 5000
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 x 0 0
x t

Figure 3.1: Initial conditions in (a) and matlab solution using Backward
Euler method for temperature distribution along rod with time in (b)

Figure 3.1 shows the initial conditions in (a) and matlab solution for tem-
perature distribution along rod with time in (b). The solution using the
Backward Euler method in figure 3.1 is stable even for large time steps and
this matlab code uses a time step 6 greater than the solution using the For-
ward Euler method in figure 2.1. So even though there is more work in each
time step (inverting a matrix) using the implicit Backward Euler method it
allows larger time steps than the explicit Forward Euler method.

3.2 Crank-Nicolson Scheme

Average of the explicit (forward Euler) and implicit (backward Euler)
schemes.

Uses:
Ujk+1 Ujk
Ut =
t
26

k+1
1 Uj+1 2Ujk+1 + Ujk+1 Uj+1
k k
2Ujk + Uj1


Uxx = 2
+ 2

2 |

x
{z } | x
{z }

implicit explicit

Often used for simple diffusion problems.

27
Chapter 4

Iterative methods

Implicit methods are stable - however they can take much longer to compute
than explicit methods. We saw that the Backward Euler method requires a
system of linear equations to be solved at each time step:

~ k + ~b)
~ k+1 = ~c (where: ~c = U
AU

where A is a tridiagonal matrix. How can we speed up this calculation? By

using iterative methods.
Iterative methods:

improve the solution of A~x = ~b.

use a direct method or a guess for an initial estimate of the solution.

useful for solving large, sparse systems (eg. tridiagonal matrix A in

Backward Euler scheme).

many different methods such as Jacobi, Gauss-Seidel, relaxation meth-

ods.

iterative methods are not always applicable and convergence criteria

need to be met before they can be applied. However they are ideal for
finite difference methods (involving solution of large sparse matrices).

Iterative methods begin with an initial guess for the solution ~x0 to the
matrix equation we are trying to solve: A~x = ~b. Each iteration updates
the new k th estimate (~xk ) which converge on the exact solution ~x. Different
methods have different convergence times and for big inverse matrix problems
are much faster than direct matrix inverse methods.

28
4.1 Jacobi method
A is decomposed into a sum of lower-triangular (L), diagonal (D) and upper-
triangular terms (U ):
A = L+D+U
 
@
@ @ U
@ @
@ @
@ @
A= @D @
@ @
@ @
@
@
L @

 

4.1.1 Example
A for Backward Euler Method with Dirichlet boundary conditions:

1 + 2s s 0
s 1 + 2s s


.. .. ..

A= . . .




s 1 + 2s s

0 s 1 + 2s

1 + 2s 0

0 0
1 + 2s
...
s


D = ,L = ,


..
.. ..

.

. .

0 1 + 2s 0 s 0

0 s 0

... ...

U =

...


s


0 0
We want to solve A~x = b

(D + L + U )~x = ~b

29
 or D~x = b (L + U )~x

x1
x2

~x = ..

.

xn

If ~xk is k th estimate of solution A~x = ~b then the (k + 1)th estimate is:

D~xk+1 = b (L + U )~xk (Jacobi Method)

Since D is diagonal (Dij = ij Aij ), we can write the vector equation above
for ~xk+1 for each component (xk+1 k+1
1 , ...xn ).


1

xk+1 k
X
i = bi Aij xj , 1in (4.1)
Aii
j6=i
|{z} | {z }
D1
L+U part

4.1.2 Applying the Jacobi method

To start the scheme use an initial guess ~x0 , (eg. ~x0 = ~0). The iterations are
repeated until A~xk ~b or the residual:

|~b A~xk | < error tolerance (eg.105 )

Jacobi method converges to correct solution xk x as k if :

kD1 (L + U )k < 1
X
|Aii | > |Aij | , 1in
j6=i
| {z }
A is strictly diagonally dominant
where kBk is the row-sum norm defined below:
n n
X X |aij |
kBk = max |Bij | = max < 1.
j=1,j6=i |aii |
1in 1in
j=1

The degree to which the convergence criteria:

X
|Aii | > |Aij |, 1 i n
j6=i

30
 holds is a measure of how fast the estimate ~xk converges to actual solution
~x.

Look for matlab code on this free source website: http://www.netlib.org/

which implements the Jacobi method and try for yourself.

4.2 Gauss-Seidel Method

improves convergence of Jacobi method by simple modification

in Jacobi method the new estimate, xk+1

i is computed using only the
current estimate, xkj

Gauss-Seidel method uses all the possible new estimates (j i 1)

xk+1 k+1 k+1 k+1
i1 , xi2 , ..., x1 , x0 when updating the new estimate xk+1
i :

i1 n
1
xk+1 Aij xk+1 Aij xkj ,
X X
i = bi j 1 i n.
Aii j=1 j=i+1

This is an improvement over the Jacobi method because it uses the new
estimate xk+1
j when it can. In vector form: ~xk+1 = D1 (bL~xk+1 U~xk )
or (D + L)~xk+1 = b U xk .

solution converges xk x as k if: k(D + L)1 U k 1.

4.2.1 Example: using Gauss-Seidel method to solve a

matrix equation
The matlab code for this example is GaussSeidel.m.

Solve A~x = ~b for Res = |b Axk+1 | < 1e3 with:

5 0 2 7 0
0
A= 3 5

1
, b= 2

, x = 0


0 3 4 4 0

0
0
takes 9 iterations. The Jacobi method needs
with initial guess ~x = 0

0
17 iterations to converge so takes nearly twice as long as the Gauss-Seidel
method.

31
 Residual after each iteration using GaussSeidel method
9

6
Residual

0
0 1 2 3 4 5 6 7 8 9
no of iterations

Figure 4.1: Plot of residual using the Gauss-Seidel method after each itera-
tion

Figure 4.1 shows the residual using the Gauss-Seidel method after each iter-
ation, it takes 9 iterations for the residual error to be less than 0.001.

4.3 Relaxation Methods

Relaxation methods generalise Gauss-Seidel method by introducing a relax-
ation factor, > 0. If is optimised for the system this can increase the
rate of convergence of the solution xk by modifying the size of the correction:

i1 n

xk+1 = xki + Aij xk+1 Aij xkj ,
X X
i bi j 1 i n. (4.2)
Aii j=1 j=i

This is called the successive relaxation (SR) method and for:

0 < < 1 under-relaxation

= 1 Gauss-Seidel method

> 1 over-relaxation

32
We can re-write equation 4.2:

i1 n

xk+1 = (1 )xki + Aij xk+1 Aij xkj ,
X X
i bi j 1in
Aii j=1 j=i+1

Solution converges, ie xk x as k if: k(D+L)1 [(1)DU ]k < 1.

33
Chapter 5

2-D Finite Difference

5.1 2-D Poissons equation

Solving Laplaces (f = 0) or Poissons equation in 2-D:

Uxx + Uyy = f (5.1)

We discretise in x and y-directions:

yn+1 = b
yn

..
.

y3
y2
y1
y0
x 0 x1 x2 x3 ... xm xm+1 = a

We discretise in x-direction with grid spacing: x = a/(m + 1) and in

y-direction with grid spacing: y = b/(n + 1), and let xk = kx where
0 k m + 1 and yj = jy where 0 j n + 1. We let Ukj = U (xk , yj )
and fkj = f (xk , yj ) We are solving equation (5.1) using Dirichlet boundary

34
conditions:

U (0, y) = U0,j = g0,j (yj )

U (a, y) = Um+1,j = gm+1,j (yj )
U (x, 0) = Uk,0 = gk,0 (xk )
U (x, b) = Uk,n+1 = gk,n+1 (xk )

Using central difference approximations for Uxx and Uyy then the finite dif-
ference approximations for equation (5.1) are given by:

2U Uk+1,j 2Uk,j + Uk1,j

2
= Uxx (xk , yj ) = ,
x x2
2U Uk,j+1 2Uk,j + Uk,j1
2
= Uyy (xk , yj ) = .
y y 2

Our discretised PDE (equation 5.1) becomes (if x = y = h):

Uk+1,j 2Uk,j + Uk1,j Uk,j+1 2Uk,j + Uk,j1

+ = fk,j ,
h2 h2
or Uk+1,j + Uk1,j 4Uk,j + Uk,j+1 + Uk,j1 = h2 fk,j . (5.2)

Since we have Dirichlet boundary conditions: the outer boundaries of the

region we are solving for are known: U0,j , Um+1,j , Uk,0 , Uk,n+1 , and we need to
find the interior values: Uk,j for 1 k m and 1 j n.

For example: m = 3 and n = 3:

35
y4 = yn+1 = b

y3 h h h
U13 U23 U33

y2 h h h
U12 U22 U32

y1 h h h
U11 U21 U31

y0
x0 x1 x2 x3 x4 = xm+1 = a

Thus we need to solve for the interior values marked with a circle above
as the boundary values are already given. We let the vector of interior values
we are solving for be defined as:

U11


U12

U13




U21

~ =
U

U22 ,

vector of interior values we are solving for.


U23


U31

U32


U33

~ using equation (5.2):

Thus the matrix system for U

36
Uk+1,j + Uk1,j 4Uk,j + Uk,j+1 + Uk,j1 = h2 fk,j becomes:

4

k=1 1 0 1 0 0 0 0 0
U11
j=1
Uk,j Uk,j+1 Uk+1,j

k=1 1 4 1 0 1 0 0 0 0


U12
j=2
Uk,j1 Uk,j Uk,j+1 Uk+1,j



k=1 0 1 4 0 0 1 0 0 0

U13


j=3
Uk,j1 Uk,j Uk+1,j



k=1
1 0 0 4 1 0 1 0 0
U21


j=1
Uk1,j Uk,j Uk,j+1 Uk+1,j



k=1 0 1 0 1 4 1 0 1 0

U22
j=2 Uk1,j Uk,j1 Uk,j Uk,j+1 Uk+1,j



k=2 0 0 1 0 1 4 0 0 1
U23
j=3 Uk1,j Uk,j1 Uk,j Uk+1,j



k=2 0 0 0 1 0 0 4 1 0
U31
j=1
Uk1,j Uk,j Uk,j+1



k=3 0 0 0 0 1 0 1 4 1


U32
j=2
Uk1,j Uk,j1 Uk,j Uk,j+1



k=3 0 0 0 0 0 1 0 1 4

U33

j=3 Uk1,j Uk,j1 Uk,j

U10 + U01
f11

Uk,j1 + Uk1,j

U02


f12

Uk1,j





U14 + U03

f13



Uk,j+1 + Uk1,j






U20
f21



Uk,j1





0
+ = h2 f22




U24
f23
Uk,j+1



U30 + U41
f31

Uk,j1 + Uk+1,j





U42


f32

Uk+1,j





U34 + U43

f33

Uk,j+1 + Uk+1,j

37
ie:
4 1 0 1 0 0 0 0 0 U11


1 4 1 0 1 0 0 0 0
U12

0 1 4 0 0 1 0 0 0 U13




1 0 0 4 1 0 1 0 0
U21


0 1 0 1 4 1 0 1 0
U22

1 4 0


0 0 1 0 0 1
U23


0 0 0 1 0 0 4 1 0
U31

0 0 0 0 1 0 1 4 1 U32


0 0 0 0 0 1 0 1 4 U33

h2 f11 g10 g01

h2 f12 g02

h2 f13 g14 g03


h2 f21 g20



=

h2 f22 ,

2
h f23 g24




h2 f31 g30 g41

h2 f32 g42


2
h f33 g34 g43
or ~ = f~.
AU (5.3)

If m n 100 then direct matrix elimination methods can be used. Oth-

erwise iterative methods such as Jacobi, Gauss-Seidel, relaxation methods
~ , as discussed in previous chapter 4. Because A
should be used to solve for U
is sparse and diagonally dominant, iterative solutions are ideal here.
However we see in the next section 5.2.1 that when we solve 2-D parabolic
equations (2-D heat or diffusion equations) that the numerical solution re-
quires tweaking since the matrix A is no longer tridiagonal for 2-D as it was
for 1-D.

5.2 2-D Heat (or Diffusion) Equation

We consider the 2-D heat equation:

Ut = (Uxx + Uyy ) for 0 x a, 0 y b and 0 t T.

with initial conditions: U (0, x, y) = f (x, y) and boundary conditions: U (t, x, y) =

g(t, x, y) for (x, y) on boundary.

38
 If we use explicit forward Euler scheme as we did for the 1-D heat
equation the stability criteria is even stricter than for 1-D:

x2 + y 2
t
8
this is less attractive because the time step is much smaller.

if we use backward Euler or the Crank-Nicolson methods they are no

longer as attractive because the matrix systems to be solved are much
larger and no longer tridiagonal.

We will look at an alternative finite difference method specifically tai-

lored to the 2-D heat equation: the alternating direction implicit (ADI)
method.

5.2.1 Alternating Direct/Implicit method for the 2-D

heat equation

Ut = (Uxx + Uyy )

We let t = T /m, x = a/(n + 1), y = b/(p + 1) tk = kt, 0 k

m, xi = ix, 0 i n + 1, yj = jy, 0 j p + 1, and:

U (tk , xi , yj ) = Uijk

The time derivative, Ut , is approximated using a leap-frog step in time about

the mid-point tk+1/2 , where tk+1/2 = (tk + tk+1 )/2 using a time step of t/2:
k+1/2
Uij Uijk+1 Uijk1
= at time tk+1/2 .
t t
The spatial derivatives, Uxx and Uyy , are approximated using central differ-
ences:

2 Uijk k
Ui+1,j 2Uijk + Ui1,j
k
= at time tk ,
x2 | x
{z
2
}
EARLY STEP
2 Uijk+1 k+1
Ui,j+1 2Uijk+1 + Ui,j1
k+1
= at time tk+1 .
y 2 y 2

39
This introduces an early bias which evaluates Uxx (tk ) at an earlier time
than Uyy (tk+1 ). This bias is compensated by also evaluating Uxx (tk+2 ) at
tk+2 ; Uyy (tk+1 ) again at tk+1 and Ut (tk+3/2 ) at mid-point between tk+1 and
tk+2 :
k+3/2
Uij Uijk+2 Uijk+1
= ,
t t
2 Uijk+2 k+2
Ui+1,j 2Uijk+2 + Ui1,j
k+2
=
x2 | x{z
2
}
LATE STEP
2 Uijk+1
and as before.
y 2
k k
The boundary conditions specify Uoj , Um+1,j , Uiok , Ui,p+1
k
, and initial condi-
0
tions specify Uij . So we are solving for 1 k m, 1 i m, 1 j p.
ie. for each interior point at time tk .
We solve the problem for both time steps tk+1 and tk+2 at the same time
using early and late definitions. If we let sx = t/x2 , sy = t/y 2
then Ut = (Uxx + Uyy ) becomes:
Early step:

Uijk+1 (1 + 2sy ) sy [Ui,j+1

k+1 k+1
+ Ui,j1 ] = Uijk (1 2sx ) + sx [Ui+1,j
k k
+ Ui1,j ]

This is solved first for ith row of U k+1 matrix, for 1 i n.

Late step:

Uijk+2 (1 + 2sx ) sx [Ui+1,j

k+2 k+2
+ Ui1,j ] = Uijk+1 (1 2sy ) + sy [Ui,j+1
k+1 k+1
+ Ui,j1 ]

This is solved for j th column of U k+2 matrix, for 1 j p.

These are solved using LU decomposition. (For more details see Schilling
and Harris, p. 445).

5.3 Cylindrical and spherical polar co-ordinates

Spherical and cylindrical symmetry in problems are often exploited to
reduce 2-D 1-D or 3-D 1-D.

3-D Cylindrical Co-ordinates

x = r cos
y = r sin

40
z=z
r = x2 + y 2
= arctan( xy )

Figure 5.1: 3-D Cylindrical Co-ordinates

3-D Spherical Polar Co-ordinates

x = r sin cos
y = r sin sin
z = r cos

Figure 5.2: 3-D Spherical Polar Co-ordinates

2-D Polar Co-ordinates

x = r cos
y = r sin

41

r = x2 + y 2
= arctan( xy ).

r sin
= + = cos
x x r x r r
r cos
= + = sin +
y y r y r r

5.3.1 Example: Temperature around a nuclear waste

rod
The matlab code for this example is NuclearWaste.m.
y
r
6

'$
- x
&%
r=a
Nuclear rod buried in ground
We consider the temperature increase due to storage of nuclear rods which
release heat due to radioactive decay:

1 T
(r, t) 52 T (r, t) = S(r, t)
t | {z }
source term
| {z }
2 -D heat equation

where the source term due to the radioactive decay of rod is given by:
(
Trod et/0 /a2 for r a
S(r, t) =
0 elsewhere.

where a = 25cm, = 2 107 cm2 /year, Trod = 1K, 0 = 100years, rc =

100cm, TE = 300K, 0 < r < 100cm and 0 < t < 100years. Initially
T (r, t = 0) = 300K.
Because the problem has circular symmetry (ie. no dependence)
2-D problem in (x, y) reduced to 1-D problem in r. 52 T = Txx + Tyy is 2-D
in Cartesian co-ordinates. However if we choose to use polar co-ordinates
then the temperature, T (r, t) is a function of r and t only because the rod
circularly symmetric and there is no dependence. This reduces the original

42
2-D problem to 1-D!
How do we evaluate 52 T in polar co-ordinates?
2T sin T sin T
Txx = 2
= (cos )(cos )
x r r r r
2 sin cos sin2 2 cos sin sin2
= cos2 Trr Tr + Tr + T + T
r r r r

cos T cos T
Tyy = (sin + )(sin + )
r r r r
2 cos sin 2 cos sin cos2 cos2
= sin2 Trr + Tr T + Tr + T
r r2 r r2

1 12
and Txx + Tyy = Trr + Tr + T
r r

using cos2 + sin2 = 1.

Since the temperature T (r, t) has no dependence then T = 0 and we
are solving the 1-D heat equation in polar co-ordinates:

1 T 2T 1 T
2 = S(r, t)
K t r r r
We know that in the steady state solution eventually the nuclear rod is no
longer radioactive and stops releasing heat: S(r, t) 0 as t , and
further enough away from the rod the temperature equals the environment
temperature, T (r = rc , t) = 300K. So the solution should approach the
environmental temperature T (r, t) = 300K once rod has finished radioactive
decaying.
We use finite differences to solve:
1 T 2T 1 T
2 = S(r, t) (5.4)
K t r r r
We observe that there is a singularity at r = 0 in the above equation where
special care needs to be taken so that the numerical solution is stable.
Initial conditions T (r, 0) = 300K.
Neumann boundary conditions at r = 0 (temperature cannot flow into r = 0
region)
T
(r = 0, t) = 0
r
43
Dirichlet boundary conditions at r = rc

T (r = rc , t) = 300K

Again we discretise space and time: r = rc /(n + 1), t = Tf /m, rj =

jr, 0 j n + 1, tk = kt, 0 k m, T (rj , tk ) = Tjk , and S(rj , tk ) =
Sjk .
Discrete Neumann boundary conditions at r = 0 become:

Tjk T0k T k T0k

(r = 0, t) = =0 1 T0k T1k
t t t
Discrete Dirichlet boundary conditions at r = rc become:

Tjk (r = rc , t) = Tn+1
k
= 300

We will use the backward Euler method (implicit) to solve the PDE. This
means evaluating the spatial derivatives in r at the future time step tk+1 :

Tjk+1 Tjk
Tt (tk+1 , rj ) =
t
Tj+1 2Tjk+1 + Tj1
k+1 k+1
Trr (tk+1 , rj ) = (centred difference at tk+1 )
r2
k+1 k+1
Tj+1 Tj1
Tr (tk+1 , rj ) = (leap-frog in space)
2r

Using rj = jr our discretised PDE 5.4 becomes:

1 1
Tt Trr Tr = S(r, t)
r
k+1 k+1 k+1 k+1 k+1
1 T 2Tj + Tj1 1 Tj+1 Tj1
[Tjk+1 Tjk ] [ j+1 2
] [ ] = Sjk
t
| {z } | r
{z } jr 2r
| {z }
Tt / Trr Tr /r

We let s = t/r2 and we arrive at:

k+1 s k+1 s
Tj+1 [s ] + Tj1 [s + ] + Tjk+1 [1 + 2s] = Tjk + Sjk t (5.5)
2j 2j

This is a tridiagonal matrix for 1 j n.

44
Numerical solution of the 1-D heat equation in polar co-ordinates
using the Backward Euler method
For n = 3:

 -
r0 = 0 r1 r r2 r3 rn+1 = rc = r4

The boundary conditions give T0k T1k using T r

(r = 0, t) = 0) and T4k =
300K using (T (r = rc , t) = 300) and the initial conditions are Tj0 = 300K.
We solve equation 5.5 for T1k , T2k , T3k at each time step (tk ):

1 + 2s (s 2js ) 0 T1k+1 (s + 2js )T0k+1

(s + s ) (s 2js ) k+1

2j
1 + 2s T2 + 0

s k+1 s k+1
0 (s + 2j ) 1 + 2s T3 (s 2j )T4
T1k S1k

k k
= T2 + t S2
T3k S3k

Using the boundary conditions: T0k+1 T1k+1 , T4k+1 = 300K

(1 + s + 2js ) (s 2js ) 0 T k+1

s s 1k+1
(1 + 2s) (s 2j )

(s + )
2j T2
0 (s + 2js ) (1 + 2s) T3k+1
T1k S1k

0
k k
= T2 + t S2 0

k k s k+1
T3 S3 (s 2j )T4
(1 + s + 2s ) (s 2s ) T1k+1

0
(s + 4s ) (1 + 2s) (s 4s ) T2k+1

0 (s + 6s ) (1 + 2s) T3k+1
T1k S1k

0
k k
= T2 + t S2 0

k k s
T3 S3 (s 6 )300

Or to simplify we are solving the following matrix equation for the vector of
unknown temperatures T~ k+1 :

~ k + ~b
AT~ k+1 = T~ k + tS

45
 The matlab code NuclearWaste.m can be used to check that solution
for T 300K as t (steady state approaches environment temperature,
300K).

Initial condition for Temperature distribution

301 301
Temp after 1 year
300.8 300.9 Temp after 10 years
Temp after 50 years
Temp after 100 years
300.6 300.8

300.4 300.7

300.2 300.6
Temperature

Temperature
300 300.5

299.8 300.4

299.6 300.3

299.4 300.2

299.2 300.1

299 300
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
r r

Figure 5.3: Initial conditions in (a) and matlab solution using Backward
Euler method for temperature distribution near nuclear rod at different time
intervals in (b)

Figure 5.3 shows the initial conditions and temperature distribution near the
nuclear rod at different time intervals.

46
 Part II

Numerical solution of
hyperbolic equations

47
Chapter 6

Analytical solutions to the 1-D

Wave equation

6.1 1-D Wave equation

Utt c2 Uxx = 0

or ( c )( + c )U = 0 (6.1)
t x t x
This is a hyperbolic equation since A = 1, C = c2 , B = 0 so that AC < B 2

6.2 dAlemberts solution

We introduce a change of variables:
= x + ct
= x ct
Then:
x t 1
= + = +
x t x c t
x t 1
= + =
x t x c t
So equation 6.1 becomes:

c (c )U = c2 U = 0

U (, ) = g() + f ()
= g(x + ct) + f (x ct)

48
 g(x + ct) defines waves that travel in left direction with speed c

f (x ct) defines waves that travel in right direction with speed c.

The pulses move without dispersion and the initial pulse breaks into a
left and right pulse.

6.3 Separation of variables

This time we will derive the analytical solution using the separation of vari-
ables technique as we did for the 1-D heat equation in section 2.1. We want
to solve the 1-D heat equation:

Utt = c2 Uxx (6.2)

with periodic boundary conditions U (0, t) = 0 = U (L, t)

Again we assume U (x, t) = X(x)T (t) then substitute into equation 6.2:

X(x)T 00 (t) = c2 X 00 (x)T (t)

then divide by XT
00
T X 00
= c2 = 2 (constant)
T
|{z} X
| {z }
function of t only function of x only

Solving X 00 = k 2 X, where k = /c for X(x) gives:

X = A sin(kx) + B cos(kx)

The boundary conditions give X(0) = B = 0 and X(L) = A sin kL = 0

k = kn = n/L, n = 0, 1, . . .. Thus the general solution for X(x) is:
X nx
X(x) = an sin( )
n L

Similarly if we solve T 00 = n2 T (where n = ckn ) we find the general

solution:

T (t) = C sin(n t) + D cos(n t)

. So the solution for U (x, t) is:

U (x, t) = X(x)T (t)

X
= [an sin(n t) + bn cos(n t)] sin(kn x)
n

49
where an , bn are given by initial conditions:

U
U (x, 0) = U0 (x), (x, 0) = V0 (x)
X
t X
U0 = bn sin(kn x) V0 = an n sin(kn x)
n n

RL
using orthogonality of sine functions: 0 sin(km x) sin(kn x)dx = nm

2ZL
bm = U0 (x) sin(km x)dx
L 0
2 ZL
am = V0 (x) sin(km x)dx
wm L 0
where kn = n/L and km = m/L.

50
Chapter 7

Flux conservative problems

7.1 Flux Conservative Equation

A large class of PDEs can be cast into the form of a flux conservative equation:

~
U f ~ ~ ~
= (U , Ux , Uxx , ...)
t x

Example: flux conservative form for the wave equation

We consider the 1-D wave equation Utt = c2 Uxx . If we let:
!
r U U
w
~= , where r = c , and s = .
s x t

This means that:

! !
w
~ r s
t
c x
= s = r
t t
c x
!
w
~ 0 c
or = ~ =
w f (w)
~
t x c 0 x

7.2 Stability analysis of numerical solutions

of the first order flux conservative or 1-D
advection equation
U U
= c (7.1)
t x
51
We introduce a change of variable = x ct and:

= = c , = =
t t x x

We see that equation 7.1 holds: c U

= c U
. So, U (x, t) = U () =
f (x ct) is the analytic general solution of equation 7.1, which is a wave
propagating in the right (positive x) direction.
We study the stability of different finite difference schemes in solving the
flux conservative or 1-D advection equation:

Ut = cUx , x0 x x1 , t0 t T
1 x0
Again we discretise problem x = xn+1 , t = T t m
0
and let xj = x0 +
jx, j = 0, . . . , n + 1, tk = t0 + kt, k = 0, . . . , m, and Ujk = U (xj , tk ).

7.3 Forward Time Centred Space (FTCS)

Forward Euler method in time:
Ujk Ujk+1 Ujk
= + O(t)
t t
Leap-frog or centred difference in space:

Ujk k
Uj+1 k
Uj1
= + O(x2 )
x 2x
Using FTCS method: Ut = cUx gives:
ct k
Ujk+1 = Ujk k
[U Uj1 ] (7.2)
2x j+1

7.3.1 von Neumann stability analysis of FTCS method

FTCS is unstable! Why?
We assume that independent solutions (eigenmodes) of equation 7.2 (or any
difference equation) are of the form:

Ujk = k eipjx (7.3)

where p is a real spatial wavenumber and = (p) is a complex number that
depends on p.
Equation 7.3 shows that the time dependence of a single eigenmode Ujk is

52
only through successive powers of ( k ). Difference equations are unstable
if |(p)| > 1 for some p. is called the amplification factor.
To find (p) for FTCS method substitute Ujk = k eipjx into equation 7.2:

ct ipx
k+1 eipjx = k eipjx eipx )

1 (e
2x | {z }
2i sin(px)
ict
(p) = 1 sin(px)
x
and |(p)| 1 p FTCS scheme is unconditionally unstable for solving
Ut = cUx .

7.4 Lax Method

Again we are solving the flux conservative equation: Ut = cUx . The in-
stability in the FTCS method is removed in the Lax method by using the
U k +U k
average for Ujk = j+1 2 j1 instead of Ujk in approximating Ut :

Ujk Ujk+1 12 [Uj+1

k k
+ Uj1 ]
=
t t
and centred difference again for Ux . Then Ut = cUx becomes:
1 k ct k
Ujk+1 = [Uj+1 k
+ Uj1 ] k
[Uj+1 Uj1 ] (7.4)
2 2x

7.4.1 von Neumann Stability Analysis of Lax Method

The Lax method is conditionally stable. To see substitute Ujk = k eipjx into
equation 7.4:

1 ct ipx
[eipx + eipx ]
k+1 eipjx = k eipjx (e eipx )

2
} 2x
| {z }
| {z
2i sin(px)
cos(px)
ct
= cos(px) i sin(px)
x
Lax method stable when ||2 1 :

2 c2 t2
| cos (px) + 2
sin2 (px)| 1
x
53
 c2 t2
or |1 (1 ) sin2 (px)| 1
x2
c2 t2
1 0
x2
c2 t2
or 1
x2
x
or t (c > 0)
| {z c }
COURANT CONDITION

7.5 Courant Condition

The Courant condition means Lax method is stable when t x/c

The physical meaning is that value Ujk+1 is computed from information

at points j 1 and j + 1 at time k in a stable scheme, where the
wave speed is less that the mesh spacing divided by time. ie. in a
continuum wave equation information propagates at maximum speed
c, so Lax method is stable when x
t
c. This is shown in the plot below:

t 6
s s s s s
6
Ujk+1 x t
s s s s
 -s?

J

J
s s s s s

J

J
k k k
Uj1
Uj J U

J j+1

J -

J

wave speed c
 J
- x

54
 Unstable schemes arise when x t
c ie. when the time step t be-
k+1
comes too large because Uj requires information from points outside
k k
[Uj1 , Uj+1 ] as shown in the plot below. (see Press et al, Numerical
Recipes, p. 825-830)
t 6
Ujk+1
s s s s
 -s

J

J
x 6

J

J

J

J t

J

J

J
s
s s s J s?

J

Uk Ujk k
Uj+1 J

j1 J

J -

J
x
wave speed c

J-

J

7.6 von Neumann Stability Analysis For Wave

Equation

Utt = c2 Uxx
! !
r cUx
let w
~ = =
s Ut

We saw earlier that:

! ! !
w
~ r s
t
c x s
= s = r =c
t t
c x x r

7.6.1 Lax method

We will solve the wave equation using the Lax method:
rt = csx becomes:

1 k ct k
rjk+1 = [rj1 k
+ rj+1 ]+ (sj+1 skj1 ) (7.5)
2 2x
55
st = crx becomes:

1 ct k
sk+1
j = [skj1 + skj+1 ] + k
(r rj1 ) (7.6)
2 2x j+1

where rjk = r(xj , tk ) and skj = s(xj , tk )

For von Neumann stability analysis assume eigen-modes for rjk and skj are
of the form:
! !
rjk k ipjx rj0
= e solutions stable if || 1
skj s0j

Equation 7.5 and 7.6 give:

! ! !
cos(px) ict
x
sin(px) r0 0
=
ict
x
sin(px) cos(px) s0 0

This has a solution only if determinant = 0.

This gives = cos(px) i ct

x
sin(px).

x
This is stable if ||2 1 which gives same Courant condition t c
.

7.7 Other sources of error

7.7.1 Phase Errors (through dispersion)
Fourier analysis of the Lax method shows how phase errors arise.

The Fourier mode U (x, t) = ei(px+t) is an exact solution of Ut = cUx

if and p satisfy the dispersion relation = cp, then U (x, t) =
eip(xct) = f (x ct) gives the exact solution of Ut = cUx .

ie. this mode is completely undamped and the amplitude is constant (no
dispersion) for the numerical solution using a time step which satisifes
this dispersion relation.

We will show the effects of phase errors by studying the numerical

solution of the 1-D advection equation using different time steps which
lead to dispersion being absent or present in section 7.7.2.

56
Dispersion relation for the Lax Method
The dispersion relation is only satisfied if: t = x c
.
k k ipjx
Why? Consider Uj = e
In section 7.4.1 we found:
ct
= cos(px) i sin(px)
x
ct
= eipx + i(1 ) sin(px)
x
If we let t = xc
= eipx and Ujk = k eipjx = eip(kx+jx)
When we substitute xj = jx, tk = kt and the dipsersion relation x =
ct then: Ujk = eip(ckt+jx) = eip(xj ctk ) = f (xj ctk ) .
| {z }
exact solution
Thus the Lax method has no dispersion present when the time step sat-
isfies the dispersion relation exactly: t = x
c
. We will show this in the next
section. x

7.7.2 Dispersion in the numerical solution of the 1-D

advection equation using the Lax method
The matlab code for this section is Lax Flux.m.
Use Lax Method to solve:
Ut + Ux = 0, 0 x 2 , 0t1
initial conditions:
(
1, 0.2 x 0.4
U (x, 0) = = U 0 (x)
0, otherwise
boundary conditions:
U (0, t) = U (2, t) = 0

Exact solution
initial pulse final pulse
t=0 t=1

-
0.2 0.4 1.2 1.4
U (x, t) = U 0 (x ct)

57
 = U 0 (x t) (c = 1)

In the next section we compare the above exact solution with the numerical
solution using the Lax method with different time steps:
x
t = c
no dispersion matches analytic solution.
x
t = 2c
dispersion present but pulse matches speed of wave.
1.001x
t = c
courant condition not met unstable!

Lax Method for Ut + Ux = 0

Equation 7.4 gives: Ujk+1 = 12 [Uj+1
k k
+ Uj1 ct
] 2x k
[Uj+1 k
Uj1 ]
ct
let s = x
Ujk+1 = 21 (1 s)Uj+1
k k
+ Uj1 + 21 (1 + s)Uj1
k

Again for simplicity we only consider 4 elements in x:

x0 x1 x2 x3 x4

Solve for Ujk+1 for 0 k m, 1 j 3 with boundary conditions:

U0k = 0, U4k = 0. We have:

U1k+1 1
U1k 1
(1 + s)U0k

0 2
(1 s) 0 2
k+1 1 1
U2 = 2 (1 + s) 0 U2k +
(1 s) 0

2
k+1 1 k 1 k
U3 0 2
(1 + s) 0 U3 2
(1 s)U4
or U ~ k + ~b
~ k+1 = AU

Dispersion means the initial pulse changes shape (unlike analytical solution)
because wave components with different frequencies travel at different speeds.
The matlab code is Lax Flux.m.
The numerical solution changes for different time steps depending on
whether or not the scheme is stable or dispersion is present. Figure 2.1 shows
how the solution changes for different time steps depending on whether or
not the scheme is stable or dispersion is present.

58
 Final solution at t=1 Final solution at t=1 Final solution at t=1

1 1 1

0.8 0.8 0.8

0.6 0.6 0.6

U

U
0.4 0.4 0.4

0.2 0.2 0.2

0 0 0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
x x x

Figure 7.1: Solution at t = 1 using the Lax method with different time
steps, (a) t = x/2c where dispersion is present but the pulse matches
the analytical solution for the speed of the wave, (b) t = x/c where
no dispersion is present and numerical solution matches analytical solution
exactly, and (c) t = 1.001x/c where the Courant condition is not met
and solution is becoming unstable.

7.7.3 Error due to nonlinear terms

Example
Shock wave equation:

Ut + U U
| {z x}
=0
nonlinear term
nonlinear term causes wave profile to steepen resulting in a shock.

schemes stable for linear problems can become unstable.

this will be discussed later in chapter 11

7.7.4 Aliasing error

Example
Alising error occurs when a short wavelength (1 ) is not represented well by
the mesh-spacing (x), and may be misinterpreted as a longer wavelength
oscillation (2 ).

59
 1

0.5

-0.5

-1 x
1
2
-1.5
0 1 2 3 4 5 6
x

Figure 7.2: Aliasing error occurs when the mesh spacing x is too large to
represent the smallest wavelength 1 and misinterprets it as a longer wave-
length oscillation 2

60
Chapter 8

Numerical Solution of 1-D and

2-D Wave Equation

8.1 Explicit Central Difference for 1-D Wave

Equation

Utt = c2 Uxx , 0 t T, 0 x a
T a
Discretise: t = m , x = n+1 ,
tk = kt, 0 k m, xj = jx and 0 j n + 1.

8.1.1 Example: plucking a string

The matlab code for this example is Wave1D.m.

Q
 Q
Q
 Q
 Q
  Q
 Q
 Q
0 0.8a a

A string is initially plucked or lifted from rest:

boundary conditions: U (0, t) = 0, U (a, t) = 0 or U0k = 0, Un+1
k
=0
initial conditions: string is plucked or lifted 1mm at x = 0.8a:
(
1.25x
a
, for x 0.8a
U (x, t = 0) = f (x) =
5(1 xa ), for x > 0.8a

61
Plucked string is released from rest:
U
(x, 0) = g(x) = 0
t

U (x, t = 0) = f (x) Uj0 = fj = f (xj )

U Uj0 U 1 Uj1
(x, t = 0) = g(x) j = gj = g(xj )
t | t {z 2t }
leap-frog in time
We can solve for ghost point Uj1 :
Uj 1 = Uj1 2tg(xj )
We approximate Utt and Uxx using central differences:
Ujk+1 2Ujk + Ujk1
Utt =
t2
Uj+1 2Ujk + Uj1
k k
Uxx =
x2
c2 t2
Using Utt = c2 Uxx and s = x2
, we solve for Ujk+1 at time step k + 1:

Ujk+1 = Ujk1 + 2Ujk (1 s) + s(Uj+1

k k
+ Uj1 )
| {z } | {z }
solution at tk1 solution at tk

In order to find Uj2 we need to know Uj0 and Uj1 .

We consider n = 3:

x0 x1 x2 x3 x4

boundary conditions: U0k = 0, U4k = 0

initial conditions: Uj0 = fj , Uj1 = Uj1 2tg(xj ) = Uj1 , since g(xj ) = 0.
1
U1
~ 1 1
First find U = U2
U31

U11 U10

2(1 s) s 0
~1
U = U21 =

s 2(1 s) s 0
U2
1
U3 0 s 2(1 s) U30
| {z }
A

62
 U00 U11

1
+ s 0 U2

U40 U31
| {z }
b

Use Uj0 = fj and Uj1 = Uj1 2tgj

U11

2(1 s) s 0 f1
~1 1 1
U = U2 = s 2(1 s) s f2

2
U31 0 s 2(1 s) f3
U0

g1
s 0
+ 0 + t g2

2
U40 g3

~ 1 = 1 AU
U ~ 0 + 1~b + d~
2 2
0 0
For this example, U0 = 0, U4 = 0 and:
Uj0
(x, t = 0) = g(xj ) = 0 d~ = ~0
t
~ 2, . . . , U
for U ~ m we have:

Ujk+1 = 2Ujk (1 s) + s(Uj+1

k k
+ Uj1 ) Ujk1
for 1 k m:
U1k+1 U1k

2(1 s) s 0
~ k+1
U k+1
= U2 =

s 2(1 s) s k
U2
U3k+1 0 s 2(1 s) U3k
| {z }
A
U0k U1k1

k1
+ s 0 U2


U4k U3k1
| {z }
b

U ~ k + ~b U
~ k+1 = AU ~ k1

The matlab code is Wave1D.m.

In our example U0k = 0, U4k = 0 and ~b = ~0, since U0k = 0 = U4k
At fixed boundaries U (0, t) = 0 = U (a, t) wave is reflected. We plot the
numerical solution in figure 8.1.

63
 Initial condition for plucked string
1

0.9

0.8

0.7

0.6

0.5
U

0.4

0.3

0.2

0.1

0
0 5 10 15 20 25 30 35 40 45 50
x

Figure 8.1: Initial conditions in (a) and matlab solution using explicit central
difference method for 1D wave equation in (b)

We can compare with DAlemberts solution which gives:

1
U (x, t) = [f (x ct) + f (x + ct)] since Ut (x, 0) = 0
2
where U (x, 0) = f (x) (initial conditions) for < x <
What if we want to solve the wave equation for 0 x a, with fixed bound-
ary condition U (t, 0) = 0 = U (t, a)? We can extend DAlemberts general so-
lution for Utt = c2 Uxx with initial conditions:U (x, 0) = f (x) Ut (x, 0) = g(x):
f (x + ct) + f (x ct) 1 Z x+ct
U (x, t) = + g(z)dz
2 2c xct
for x
In our example we have initial conditions:
(
1.25x
a
, 0 x 0.8a
Ut (x, 0) = 0, U (x, 0) = f (x) = x
5(1 a ), for x 0.8a

0xa
with fixed boundary conditions:
The boundary condition U (0, t) = 0 is equivalent to f and g being odd
functions:
U (0, t) = 0 f (x) = f (x)
g(x) = g(x)
(f and g are odd functions)

64
 The boundary condition U (a, t) = 0 is equivalent to f and g being peri-
odic with period 2a:

U (a, t) = 0 f (x + 2a) = f (x)

g(x + 2a) = g(x)
(f and g are periodic with period 2a)

Since Ut (x, 0) = g(x) = 0 the analytical solution for our example:

f (x + ct) + f (x ct)
U (t, x) =
2
and we can compare the analytical solution with the numerical solution in
figure 8.2.

Figure 8.2: DAlemberts solution in (a) and error using numerical matlab
solution using explicit central difference method for 1D wave equation in (b)

8.1.2 1-D Wave Equation with Friction

The matlab code for this example is Wave1DFriction.m.
We consider friction due to viscosity of medium and density of string.
Suppose we are solving:

U + 2 U = c2 Uxx , 0 x a = 50, 0 t T = 20

The friction term opposes motion of string and means that eventually
vibrations decay with time.

65
Suppose string is initially plucked in 2 places:

@ @
@ @
@ @
0 0.1a 0.3a 0.7a 0.9a a

We have initial conditions:

0, 0 x 0.1a
5(10x a), 0.1a x 0.2a





5(10x + 3a), 0.2a x 0.3a




U (x, 0) = 0, 0.3a x 0.7a
5(10x 7a), 0.7a x 0.8a





5(10x + 9a), 0.8a x 0.9a





0, x 0.9a

Ut (x, 0) = 0
and boundary conditions: U (x, 0) = 0, U (x, a) = 0.
Again we use central difference for Uxx and Utt as in section 8.1.1.
We use a leap-frog step for Ut

Ujk Ujk+1 Ujk1

=
t 2t
Now we substitute difference approximations into Utt + 2 Ut = c2 Uxx

Ujk+1 2Ujk + Ujk1 Ujk+1 Ujk1 c2 (Uj+1

k
2Ujk + Uj1
k
)
+ =
t2 t x2
2 2
let s = cx
t
2

Rearranging for Ujk+1 gives:

1 n o
Ujk+1 = 2(1 s)Ujk (1 t)Ujk1 + s(Uj+1
k k
+ Uj1 )
1 + t
Special care is again needed to solve for Uj1 which needs Uj0 and the ghost
point, Uj1 . To find Uj1 we use initial condition:
U Uj0 Uj1 Uj1
(x, t = 0) = =0=
t t 2t
or Uj1 = Uj1 (since Ut (x, 0) = 0)

66
We evaluate Uj1 :




1



Uj1 = 2(1 s)Uj0 (1 t) Uj1 +s(Uj+1
0 0
Uj1 )
1 + t

| {z }


=Uj1

2 1 n o
Uj1 = 2(1 s)Uj0 + s(Uj+1
0 0
Uj1 )
1 + t 1 + t
1 n o
Uj1 = 2(1 s)Uj0 + s(Uj+1
0 0
Uj1 )
2

Example n = 3

x0 = 0 x1 x2 x3 x4 = a

U0k = 0 = U (0, t), k

Un+1 = U4k = U (a, t)
~ 1 first:
Again we solve for time step k = 1, U
U11 U10 U00

2(1 s) s 0
~1 = 1 s
U21 =
0
U

s 2(1 s) s U2 + 0

2 2
U31 0 s 2(1 s) U30 U40
~ k+1 are given by:
and the solution for time steps, k 1, U

U1k+1 U1k

2(1 s) s 0
~ k+1 1
= U2k+1 =
k
U

s 2(1 s) s U2
1 + t

k+1
U3 0 s 2(1 s) U3k
| {z }
A
k1
U0k

U1
s 1 t
+ 0 U2k1

1 + t 1 + t

U4k | {z } U3k1
| {z } e
b
~ k + ~b eU
= AU ~ k1
The numerical solution is plotted in figure 8.3 below.

The matlab code is Wave1DFriction.m

67
 Initial condition for plucked string
250

200

150
U

100

50

0
0 5 10 15 20 25 30 35 40 45 50
x

Figure 8.3: Initial conditions in (a) and matlab solution using explicit central
difference method for 1D wave equation with friction in (b)

8.2 2-D Wave Equation

Utt = (Uxx + Uyy ), 0 x a, 0 y b, 0 t T

8.2.1 Example: vibrations of a thin elastic membrane

fixed at its walls
We discretise in x and y-directions:
yp+1 = b
yp

..
.

y3
y2
y1
y0
x 0 x1 x2 x3 ... xn xn+1 = a

T a b
We discretise: t = m , x = n+1 , y = p+1 , tk = kt, xi = ix, yj =
jy0 k m, 0 i n + 1, 0 j p + 1, and let Uijk = U (tk , xi , yj )

68
 Suppose we solve for n = 3 and p = 3 and have Dirichlet boundary
conditions:
k k k
U (0, y, t) = 0 = Uoj , U (a, y, t) = 0 = Un+1,j = U4j , U (x, 0, t) = 0 =
k k k
Ui0 , U (x, b, t) = 0 = Ui,p+1 = Ui4
and initial conditions:
U (x, y, 0) = f (x, y) = fij Ut (x, y, 0) = g(x, y) = gij .
Since we have Dirichlet boundary conditions: the outer boundaries of the
k k k k
region we are solving for are known: U0,j , Un+1,j , Ui,0 , Ui,p+1 , and we need to
k
find the interior values: Ui,j for 1 i n and 1 j p.
y4 = yp+1 = b

y3 h h h
U13 U23 U33

y2 h h h
U12 U22 U32

y1 h h h
U11 U21 U31

y0
x0 x1 x2 x3 x4 = xn+1 = a

We will use the 2-D Central Difference Method

Uijk+1 2Uijk + Uijk1

Utt = ,
t2
k
Ui+1,j 2Uijk + Ui1,j
k
Uxx = ,
x2
k
Ui,j+1 2Uijk + Ui,j+1
k
Uyy =
y 2

t2 t2
We let sx = x2
, sy = y 2
and substitute the central difference approx-

69
imations into our PDE, Utt = (Uxx + Uyy ) we solve for Uijk+1 :

Uijk+1 = 2Uijk (1 sx sy ) Uijk1 + sx (Ui+1,j

k k
+ Ui1,j k
) + sy (Ui,j+1 k
+ Ui,j1 )

computing U ~ k+1 uses the solution at U

~ k and U
~ k1 .
1 0 1
For first time step Uij needs Uij and Uij . Again we need to use the initial
conditions to find the ghost point, Uij1 :

Uij0 Uij1 Uij1

= Ut (x, y, 0) = = g(xi , yj ) = gij Uij1 = Uij1 2tgij
t 2t
Solution at first time step k = 1:
sx 0 sy 0
Uij1 = Uij0 (1 sx sy ) + tgij + 0
(Ui+1,j + Ui1,j ) + (Ui,j+1 0
+ Ui,j1 )
2 2
k
U11

Uk
12
k
U13
Uk

21
~k =
If we let U U22k

k
U
23
Uk
31
k
U32
k
U33
then for time steps, k > 1, the solution is:

Uijk+1 = 2Uijk (1 sx sy ) Uijk1 + sx (Ui+1,j

k k
+ Ui1,j k
) + sy (Ui,j+1 k
+ Ui,j1 )

and we can write this in vector form:

U ~ k + ~b U
~ k+1 = AU ~ k1

where:
sy 0 sx 0 0 0 0 0

s
y sy 0 sx 0 0 0 0
0 sy 0 0 sx 0 0 0


s
x 0 0 sy 0 sx 0 0
A =

0 sx 0 sy sy 0 sx 0

0 0 sx 0 sy 0 0 sx

0 0 0 sx 0 0 sy 0

0 0 0 0 sx 0 sy sy

0 0 0 0 0 sx 0 sy
and = 2(1 sx sy )

70
 k k
sx U01 + sy U10

k

sx U02

k k
sx U03 + sy U14


k


sy U20

b =
0

k


sy U24

k k

sx U41 + sy U30

k
sx U42


k
sx U43 + sy sk34

8.2.2 Examples of wave equation

1. Elastic wave propagation through rocks in 1-D

xx,x = Utt (8.1)

where

xx = Exx , xx = stress, xx = strain

U
= E
x

E
8.1 EUxx = Utt or Utt = Uxx

q
E
elastic waves propagate with speed

2. Electromagnetic Wave Equation

c2 2 E = E and c2 2 B = B (8.2)

From Maxwells equations where E is electric field, B is magnetic field.

Derived using:
B
E = , E = (8.3)
0 t

E
B = 0, B = 0 0 (8.4)
t
~ ~ 2~
 8.4 and using ( V ) = ( V ) V
taking curl of 8.3 and
and ( E) = 0 = 0, ( B) = 0 gives Equation 8.2 where
q
1
c= 0 0
= 3 108 m/s.

71
3. Schrodingers Wave Equation

ih = H
t
for a wavefunction of a quantum system defined by Hamiltonian,
H.
2 2

eg. H = KE + P E = h2m + V (r)
R
numerical solutions also need to satisfy |(x)|2dx = 1

72
Chapter 9

Finite element method

9.1 An introduction to the Finite Element

Method
Finite difference (FD) method is an approximation to the differential
equation.

Finite element method (FEM) is an approximation to its solution.

FD methods are usually based on the assumption of regular domains

eg line in 1-D, rectangle in 2-D with regular elements

FEM is better for irregular regions as the domain can be partitioned

into any simple subregion such as triangles or rectangles in 2-D or
bricks and tetrahedra in 3-D. Figure 9.1 shows a finite element mesh
with triangles for an irregular domain.

Example: Solving Poissons equation in 1-D using FEM

Uxx = q, 0xL (9.1)

We consider Dirichlet boundary conditions: U (0) = U (L) = 0. A weak

solution of (9.1) considers the variational form of (9.1):
Z L Z L
Uxx (x)(x)dx + q(x)(x) = 0, (9.2)
0 0

where (x) satisfy the boundary conditions: (0) = (L) = 0.

73
 Figure 9.1: FEM mesh with triangles

We can integrate the first term by parts:

Z L Z L
Uxx (x)(x)dx = Ux (x)(x)]x=L
x=0 Ux (x)x (x)dx
0 0
Z L
= Ux (x)x (x)dx
0

using (0) = (L) = 0.

Then (9.2) becomes:
Z L Z L
Ux (x)x (x)dx = q(x)(x)dx (9.3)
0 0

Equation (9.3) holds for all functions (x) which are piece-wise continous
and satisfy the bc: (0) = (L) = 0.
To solve equation (9.3) using the FEM we again introduce a mesh (as in
FD) on the interval [0, L] with mesh points xj = jx, j = 0, . . . , n + 1 where
L
x = n+1 . To complete the discretisation we must choose a basis for (x).
The most common basis chosen for (x) are the hat functions, j (x).
We solve (9.3) using these:
n
X
(x) = aj j (x)
j=1

where

74
 0 x xj1


0, for
1
(x xj1 ), for xj1 x xj


j (x) = x
1


1 x (x xj ), for xj x xj+1
0, for x xj+1

6
1

A



A
A
j (x)

A

A

A

A

A

A -
xj1 xj xj+1

 -

with this construction: j (xi ) = ij and:

0, for 0 < x < xj1
j 1
, for xj1 < x < xj


0
j (x) = = x
1
x

x , for xj < x < xj+1
0, for x > xj+1

We let (x) = nj=1 aj j (x) and (xi ) = ai for i = 1, . . . , n, and (0) =

P

1 (0) = 0 and (L) = n (L) = 0 so that (x) satisfies boundary conditions.

The hat functions
RL
are advantageous as a basis as they are nearly or-
thonormal, ie. 0 j (x)k (x)dx = 0 when |j k| > 1.
Using FEM we seek an approximate solution to (9.3) which is satisfied
for all basis functions, i (x), for i = 1, . . . , n:
Z L Z L
Ux (x)x (x)dx = q(x)(x)dx
0 0

and require that 9.3 be satisfied for = i , i = 1, . . . , n. We also expand

the solution U (x) using the hat functions i as a basis:
n
X
U (x) Uh (x) = bj j (x)
j=1

75
 This simplifies equation 9.3 and we solve for = i , i = 1, . . . , n:
ie.
Z L Z L
0 0
Uh (x)i (x)dx = q(x)i (x)dx, for i = 1, . . . , n
0 0
0 f
where f (x) = x
.
Z L
0 0
LHS = Uh (x)i (x)dx
0
n
Z LX
0 0
= bj j (x)i (x)dx
0 j=1
n
X
= Ci,j bj
j=1
0 0
where Ci,j = 0L j (x)i (x)dx. Ci,j is known as the stiffness matrix in me-
R

chanics.
To find the coefficients bj which define our solution U (x) we must solve
n linear equations:
n
X Z L
LHS = Ci,j bj = RHS = q(x)i (x)dx = qi (9.4)
j=1 0

for i = 1, . . . , n with qi = 0L q(x)i (x)dx).

R

We approximate the solution by expanding in the basis of hat functions:

U (x) nj=1 bj j (x). Thus we only need to know the coefficients bj to
P

define our solution U (x) and FEM solves the following equation for vector
~b = (b1 , . . . , bn ):
n
X Z L Z L
bj j,x (x)i,x (x)dx = q(x)i (x)dx,
j=1 0 0
n
X
or bj Ci,j = qi
j=1

for i = 1, . . . , n.
We will show that the stiffness matrix C is tridiagonal for this example.
We are solving the above system for coefficients bj , thus we are solving C~b = ~q
and can use iterative methods in FEM solutions too.
We can show that the stiffness matrix is tridiagonal:
1

x
, i=j1
Z L 2 ,

i=j
Cij = j,x (x)i,x (x)dx = x
1
0 , i=j+1
x


0, elsewhere

76
We approximate qi using:
Z L Z L
qi = q(x)i (x)dx q(xi ) i (x)dx
0 0
Z xj Z xj+1 !
1 1
= q(xi ) (x xj1 )dx + 1 (x xj )dx
xj1 x xj x
= xq(xi )

We can substitute the above simplifications into equation 9.4 and arrive
at C~b = x~q or x
1
C~b = ~q:
2 1

x2 x2
0 ...
b1

q1


1 2 1 ...


x2 x2 x2
b
2

q2


. = .

.. .. .. . .


0 . . . . .



.. .. 1 2

bn qn
. . x2 x2

Thus the matrix system to solve is the same as FD solution in this example
and the solution involves inverting the stiffness matrix C:
~b = C 1 x~q

Iterative methods are useful in FEM too as it involves inverting large, sparse
matrices.
Once ~b is known, the solution U to the PDE is given by:
n
X
U (x) bj j (x)
j=1

This is a weak solution of the PDE Uxx = q.

9.2 Comparing FEM solution to FD solution

for our example

Uxx = q, 0 x L, U (0) = U (L) = 0

9.2.1 FD solution
L
Discretise using xj = jx, j = 0, 1, . . . , n + 1 where x = n+1
, U0 = 0 =
Un+1 (using boundary conditions).

77
 We let U (xj ) = Uj , q(xj ) = qj . The central difference approximation to
the PDE is:
Uj+1 2Uj + Uj1
Uxx =
x2
and Uxx = q becomes
Uj+1 + 2Uj Uj1
= qj
x2
We solve for U1 , . . . , n since U0 and Un+1 given from boundary conditions
and we can rewrite in matrix form:
2 1

0 ...

Ux
0 (=0)

x2 x2 U1 2 q1

1 2 1 ..
.

x2 x2 x2



U2 0
q2

.. + ..

= ..

... ... ...

0 . . .






.. ... 1 2

Un Un+1 =0
qn
. x2 x2 x2
| {z }
same coefficient matrix for FD as FEM
In this example FEM and FD methods solve the same matrix system.

9.3 2-D Finite Element Method

Figure 9.2: FEM mesh with triangles

78
 We consider a triangular mesh (could also be rectangular) shown in figure
9.2 where G is the domain inside the circle and G is the domains boundary.
We are solving:
2 U = q in G (9.5)
with boundary conditions, U = 0 on G.
The weak solution for U satisfies the variational form for Equation 9.5:
Z Z Z Z
2 U (x, y)(x, y)dxdy = q(x, y)(x, y)dxdy (9.6)
G G

where (x, y) = 0 on G (satisfies boundary conditions).

Using Greens first identity:
Z Z I Z Z
(2 U )dxdy = (U n)dS ( U )dxdy
G G G
| {z }
=0 since =0 on G
Z Z
= ( U )dxdy
G

Thus equation 9.6 becomes:

Z Z Z Z
( U )dxdy = qdxdy (9.7)
G G

which holds G where = 0 on G.

Similarly to the 1-D case we seek an approximate solution to equation 9.7
by expanding U (x, y) in a basis of 2-D hat functions:
n
X
U (x, y) Uh (x, y) = bj j (x, y)
j=1

where Uh (xi , yi ) = bi and Uh = 0 on G.

9.3.1 2-D hat functions

The 2-D hat functions satisfy j (xj , yj ) = 1, j (xi , yl ) = 0 if i 6= j and j 6=

l at all other vertices. The 2D hat function is plotted in figure 9.3.
We require that equation 9.7 holds for all (x, y) and solve for =
1 , 2 , . . . , n :
Z Z Z Z
Uh i dxdy = qi dxdy, for i = 1, . . . , n (9.8)
G G

79
Figure 9.3: 2D hat function (j (xj , yj ) = 1, j (xi , yl ) = 0 if i 6= j and j 6=
l)

Z Z n
X
LHS = Uh i dxdy = Ci,j bj
G j=1

where Ci,j = G j i dxdy is called the stiffness matrix.

RR

Equation 9.8 becomes:

n
X Z Z
Ci,j bj = qi , for i = 1, . . . , n where qi = qi dxdy.
j=1 G

If C is symmetric and positive definite then the system has a unique solution.

9.3.2 Example: 2-D Finite Element Method using eS-

cript for elastic wave propagation from a point
source.
eScript is a general PDE solver which implements the finite element
method written in python
(see https://launchpad.net/escript-finley)

eScript can be applied to any problem of the form:

(Ajl a,l + Bj a),j + Cl a,l + Da = Xj,j + Y

where a is the scalar we are solving for in this example.

(eScript can also solve for a vector ~a)

80
We are using Einstein notation and according to this convention if an index
appears twice in a single term it implies we are summing over all possible
values:
2 fi 2 f1 2 f2
ai f,ii = ai = a 1 + a 2
x2i x21 x2
We will see that the FEM takes care of spatial derivative in the problem
below. However we still need to approximate time derivatives.
We want to solve the 2-D wave equation for a point source:

tt = Vp2 (xx + yy ) + FPS

where p is the wave speed, is the wave-field and FPS is the force due to
the point source.
In eScript this becomes:

Da = Xj,j + Y
where a = tt
D = 1
Xj = Vp2 ,j
Y = FPS

Figure 9.4: Plot of Euclidean normal of the displacement at t > 0 for a point
source using eScript.

We solve for ak at each time step tk . Once ak is known we use it to

calculate the solution at the next time step, k+1 using the central difference

81
formula:
2 k k+1 2k + k1
ak =
t2 t2
or

k+1 = 2k k1 t2 ak

The eScript python code is 2Dpointsource.py and the output from this
code is shown in figure 9.4.

82
Chapter 10

Spectral methods

10.1 An introduction to spectral methods

remove spurious dispersion and are highly accurate

exponential convergence for smooth functions (smooth functions have

rapidly decaying Fourier transforms)

usually involve calling a Fast Fourier Transform (fft) subroutine.

good for smooth solutions.

Like the FEM, the spectral method also approximates the solution
U (x).

FEM approximates the solution as a linear combination of piece-wise

functions that are non-zero only on small subdomains (hat functions)-
local approach.

spectral methods approximate the solution as a linear combination of

continuous functions that are generally nonzero throughout the domain
(usually sinusoids or Chebychev polynomials)- global approach.

We will show an example using the spectral method where U (x) is ex-
panded as a Fourier series and this series and its spatial derivatives are then
substituted into the PDE resulting in a system of ODEs in time.

83
10.1.1 Example 1: Comparing the accuracy of solu-
tions of a variable speed wave equation with
either the spectral or finite difference method
Spectral method for variable speed wave equation
In this example we will compare the accuracy of either the spectral or finite
difference method when solving the 1D advection equation with a variable
wave speed, c(x) = 51 + sin2 (x 1). First we will derive the solution using
the spectral method:

Ut + c(x)U (x) = 0, 0 x 2 and 0 t 9

U (x, 0) = exp(100(x 1)2 )
1
c(x) = + sin2 (x 1)
5
U (0, t) = U (2, t) periodic boundary condition

The matlab code is spectral variable wave speed.m

Again we discretise in space and time: x = 2
2n
= n , t = m
T

xj = jx, j = 0, 1, 2, . . . , 2n 1
tk = kt, k = 0, 1, 2, . . . , m
U (xj , tk ) = Ujk
The spectral method uses the discrete Fourier transform of U (xj , t):

U = F (U ),
2n1
X
= U (xj , t) exp(ixj )
j=0
2n1
X
= U (xj , t) exp(i2j/(2n)), using xj = jx = j2/2n
j=0
2n1
X
= U (xj , t) exp(ij/n)
j=0

for = n + 1, . . . , n.
U (xj , t) is then defined as the inverse discrete Fourier transform of U :

U (xj , t) = Uj = F 1 (U ),
n
1 X
= U exp(ixj )
2n =n+1
n
1 X
= U exp(i2j/(2n)
2n =n+1

84
where j = 0, . . . , 2n 1.
With this definition the spatial derivatives are:
n
U (xj , t) 1 X
= i U exp(ixj )
x 2n =n+1
= F 1 (i U )
= F 1 (iF (U ))

We solve the advective equation with variable wave speeds and compare
the solution with either FD or spectral method:

Ut + c(x)Ux = 0
1
where c(x) = + sin2 (x 1)
5
ic: U (x, 0) = exp(100(x 1)2 )
periodic bc: U (0, t) = U (, t)

The central difference approximation is used for Ut and spectral method

for Ux :
Ujk+1 Ujk1
+ c(xj ) F 1 (iF (Ujk )) = 0
| 2t
{z } | {z }
leap-frog for Ut spectral method for Ux
or Ujk+1 = Ujk1 + 2tc(xj )F 1 (iF (Ujk ))

For Uj1 need Uj0 and Uj1

U (x, 0) = Uj0 = exp(100(x 1)2 )

since c(x) 51 we can assume a constant wave speed of 1/5 to calculate

Uj1 at t = t.

1 t
Uj1 = U (x, t) = U (x + c(x)t) U 0 (x + t) = exp(100(x + 1)2 )
5 5
The matlab code is spectral variable wave speed.m.

Comparing accuracy of solution with spectral method vs. finite

difference method
Solve again using finite difference.

Ut + c(x)Ux = 0

85
This matlab code is fd variable wave speed.m.
The Lax method is used for Ut and central difference method for Ux :

U Ujk+1 21 (Uj1
k k
+ Uj+1 )
=
t t
k k
U Uj+1 Uj1
=
x 2x
c(xj ) = cj .

Plug the formulas into the PDE:

Ujk+1 21 (Uj1
k k
+ Uj+1 ) k
U k Uj1
+c(xj ) j+1 = 0
| t
{z } | 2x
{z }
Ut Ux
1 1
or Ujk+1 k
= (1 + scj )Uj1 k
+ (1 scj )Uj+1
2 2

where s = t/x. Using 4 elements:

x0 x1 x2 x3 x4

U0k = 0 = U4k given by boundary conditions

Uj0 = U (x, 0) given by initial conditions

U1k+1 U1k (1 + sc1 )U0k

0 1 sc1 0
~ k+1 k+1 1 k 1
U = U2 = 1 + sc2 0 1 sc2 U2 + 0
2 2

U3k+1 0 1 + sc3 0 U3k (1 sc3 )U4k

~ k + ~b
~ k+1 = AU
or U

Figure 10.1(b) shows that the solution using finite differences is much worse
than the spectral method because dispersion is introduced in FD method
when a variable wave speed is applied. However figure 10.1(a) shows the
spectral method performs well when smooth initial conditions of a Gaussian
pulse are used and there is very little dispersion present.

86
 Variation of U with time Variation of U with time
1.5
1
1
0.8
0.5
U

0.6
0

U
0.5 0.4

10
0.2
8
0
6 10

4
5
2
t
0 t 5 6 7
6 7 0 2 3 4
0 1 2 3 4 5 0 1
x x

Figure 10.1: Numerical solution for 1D advection equation with initial condi-
tions of a smooth Gaussian pulse with variable wave speed using the spectral
method in (a) and finite difference method in (b)

10.1.2 Example 2 Comparing spectral and finite dif-

ference methods with constant wave speed con-
ditions and initial conditions of a non-smooth
pulse
We solve the advective equation with constant wave speed (c = 1) with initial
conditions of a box pulse and compare the solution with either FD or spectral
method:

Ut + Ux = 0
(
1, 0.5 x 1.
ic: U (x, 0) =
0, otherwise
periodic bc: U (0, t) = U (, t)

The matlab code for the spectral solution is spectral c 1 box.m.

The matlab code for the FD solution is fd c 1 box.m. In this code we

have chosen the time step carefully so that no dispersion is present for
a constant wave speed of c = 1. Please see section 7.7.2 for a discussion
on dispersion in finite difference methods.

Figure 10.2(a) shows that the solution with an initial condition which is not
smooth like the box pulse we used here using the spectral method is much
worse than the finite difference method. This is because the spectral method
uses Fourier series to approximate the initial conditions and is unable to

87
 Variation of U with time Variation of U with time

1.5 1

0.8
1
0.6
0.5
U

U
0.4
0
0.2
0.5
0
5
5

4
4

3 3

2 2

1 1
t 6 7 t 7
4 5 5 6
0 2 3 0 3 4
0 1 1 2
0
x x

Figure 10.2: Numerical solution for 1D advection equation with initial con-
ditions of a box pulse with a constant wave speed using the spectral method
in (a) and finite difference method in (b)

approximate non-smooth initial conditions accurately. It is important to

note that for the numerical solution using the FD method in figure 10.2(b)
that we were able to remove dispersion in this example by carefully choosing
t = x/c (see section 7.7.2).

88
 Part III

Nonlinear partial differential

equations

89
Chapter 11

Shock wave

11.1 Analytical solution: Method of charac-

teristics
Analytical solution to the shock wave equation is given by the method of
characteristics. We will illustrate this method first for a linear first-order
PDE:
can be applied to first order PDEs:
a(x, t)Ux + b(x, t)Ut + c(x, t)U = 0 (11.1)
with initial conditions U (x, 0) = f (x)
We change co-ordinates from (x, t) to (x0 , s) so that our PDE 11.1 becomes
an ODE for certain characteristic curves in the x-t plane. The new variable,
s, will vary along the characteristic curves, whereas x0 will remain constant.
How does it work?
dx dt
We let: = a(x, t), = b(x, t)
ds ds
dU dx U dt U
Then = + = aUx + bUt (11.2)
ds ds x ds t
We substitute 11.2 into 11.1:
dU
+ c(x, t)U = 0
ds
This is an ODE along the characteristic curves satisfying the characteristic
equations:
dx dt
= a(x, t) and = b(x, t)
dt ds

90
11.1.1 Example 1: Using method of characteristics to
solve the linear 1-D advection equation

Ut + cUx = 0
initial conditions: x(s = 0) = x0 , t(s = 0) = 0
U (x, t = 0) = f (x) or U (s = 0) = f (x0 )

dx
= c x = cs + k1 , use x(0) = x0 = k1
ds
x = cs + x0

dt
= 1 t = s + k2 , use t(0) = 0 = k2
ds
t=s
and since t = s x = ct + x0
Solve for U :
dU dt dx
= Ut + Ux = Ut + cUx = 0
ds ds ds
ie dU
ds
= 0 U = k3 (U is constant along characteristic curves).
Use initial conditions U (s = 0, x0 ) = f (x0 ) = k3
ie U = f (x0 ) = f (x ct) since x0 = x ct.
This is the same as DAlemberts solution for a wave moving to the right at
speed c. The characteristic curves are given by: x = x0 + ct or t = 1c (x x0 ).

t 6

-
1
slope (= c
) and U are constant along curves x

11.1.2 Example 2: Using method of characteristics to

solve the nonlinear inviscid Burgers equation
Shock waves result when solving the nonlinear inviscid Burgers equation:
Ut + U Ux = 0
initial conditions: x(s = 0) = x0 , t(s = 0) = 0
U (x, t = 0) = f (x) or U (s = 0) = f (x0 )

91
Now the wave speed is not constant but depends on the amplitude U (x, t).
The characteristic equations are:

dt
= 1 t = s (using t(0) = 0)
ds
dx
= U x = U t + x0 (using x(0) = x0 and t = s)
ds
Again:

dU dt U dx U
= +
ds ds t ds x
= Ut + U Ux = 0
U = k3 = f (x0 ) = f (x U t)

So U = f (x U t) is given implicitly since U is a function of itself. The

characteristic curves given by

1 1
t= (x x0 ) = (x x0 )
U f (x0 )

The characteristic curves no longer have constant slope - they may cross
(meaning U is multiply defined shock waves) or be discontinuous (regions
with no solution for U expansion waves) as we will see in the next example.

Example
Solving Ut + U Ux = 0 with the following initial conditions:
(
U1 , x > 0
U (x, t = 0) = f (x) =
U2 , x < 0

( 1
U1
(x x0 ), x > 0 or x = U1 t + x0
t= 1
U2
(x x0 ), x < 0 or x = U2 t + x0

2 cases

U1 < U2 - compression wave shock

U1 > U2 - expansion wave rarefaction

92
Case 1: Shock wave U1 < U2
6
t
x = U1 t x = U2 t

   
  
   
   
   
   
   
   
    -

x<0 x>0 x
1 1
t= U2
(x x0 ) t= U1
(x x0 )

In the fan bounded by x = U1 t and x = U2 t the characteristic curves are

multi-valued leading to shocks (breaking waves). We illustrate this below:
U t=0 U '$ t>0
U2 U2
6 6 shock
U1 U1
&%
- -
x centred compression wave x
with overlap

Case 2: Rarefaction or expansion wave U1 > U2

6
t
x = U2 t x = U1 t
no solution

   
   
  
   
  
   
 
    -

x<0 x>0 x
1 1
t= U2
(x x0 ) t= U1
(x x0 )

The solution is single-valued for t > 0 unlike the shock wave case. How-
ever in wedge between x = U2 t and x = U1 t there is no information. We
assume x = Ut in wedge since U2 t x U1 t and speeds vary U2 U U1

93
and add solution to the wedge.

Adding solution to wedge:

6
t
x = U2 t x = Ut x = U1 t

 

    
   
    

   

  

   

  


    -

U2 , xt < U2



x
Thus U = t
, U2 < xt < U1
U1 , xt > U1

U t=0 U t>0
6 U1 6'$U1

U2 U2 rarefaction
&%
- -
x centred expansion wave x

11.2 Numerical Solution for nonlinear Burgers

Equation

Ut + U Ux = 0, 0 x 1, 0t1

U (x, 0) = f (x) = exp(10(4x 1)2 )

Solution given implicitly by U (x, t) = f (x U t) so speed depends on ampli-

tude, U .
We study the numerical solution using 3 methods but we will see in each
case that the numerical solution fails to produce a shock wave because we
are unable to produce multi-valued solutions.

94
 1
t=0
0.9 t>0
0.8
0.7
0.6
U

0.5
0.4 rarefaction shock
0.3
0.2
0.1
0
0 0.1 0.2 0.3 0.4 0.5
x

Figure 11.1: The analytical solution U (x, t) = f (x U t) is plotted to show

how shock and rarefaction develop for this example

11.2.1 Example I: Finite difference solution with Lax

Method
The matlab code is Shock Lax.m. We are solving:

Ut + U U x = 0

Ujk Ujk+1 21 (Uj1

k k
+ Uj+1 )
= (Lax method for Ut )
t t
Ujk k
Uj+1 Uj1k
= (leap-frog for Ux )
x 2x
The Courant condition only holds for linear wave equation. A good guess is
x
t << max(U )
.(Waves travel at a maximum wave speed U = 1.)
Put difference equations into PDE:

Ut + U Ux = 0 becomes:
1n k o
Ujk+1 = Uj+1 (1 sUjk ) + Uj1
k
(1 + sUjk ) (11.3)
2
t
Where s = x
Use boundary conditions U (0, t) = U (1, t) = 0 and for 4 elements:

x0 x1 x2 x3 x4

95
 So U0k = 0 = U4k given by boundary conditions and we can rewrite Equa-
tion 11.3 as a matrix system of equations:

U1k+1 1 sU1k U1k

0 0
~ k+1 k+1 1
U = U2 = 1 + sU2k 0 1 sU2k U2k

2

k+1 k k
U3 0 1 + sU3 0 U3
(1 + sU1k )U0k

1
+ 0
2

k k
(1 sU3 )U4
1 ~ k 1~
= AU + b
2 2

A varies with time because of Ujk term in matrix!

We can compare the difference between the matlab code for the linear
1D advective equation (Ut + Ux = 0), Lax Flux.m in section 7.7.2 and the
shock wave equation (Ut + U Ux = 0) above, Shock Lax.m.

Initial condition for U

0.8

0.6
U

0.4

0.2

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x

Figure 11.2: Initial conditions in (a) and solution for nonlinear Bugers equa-
tion using the Lax method in (b)

When we compare the analytical solution given by the method of character-

istics to the numerical solution given by the Lax method we can see that the
numerical solution is accurate for the linear 1D advection equation (see nu-
merical solution in figure 7.1) but fails to give a shock wave for the nonlinear
Burgers equation in figure 11.2. The Lax method introduces dispersion into
the numerical solution and in the nonlinear case this removes the shock
wave instability and flattens the wave front.

96
11.2.2 Example II: Solution using Method of Lines
The matlab code is Shock mol.m and Uprime2.m.

U
= U Ux
t
Using the method of lines solution (as demonstrated in section 2.3) we only
replace spatial derivative Ux with FD approximation.

Uj Uj+1 Uj1 Uj Uj+1 Uj1

 
= , = Uj
x 2x t 2x
and again we show the case with 4 elements:

x0 x1 x2 x3 x4

U0 = 0 = U4 from boundary conditions.

~ U1 0 U1 0 U1 UU
U 1 1 1 0
= U2

= U2 0 U2
U2 +

0
t 2x 2x

U3 0 U3 0 U3 U3 U4

or

~ = 1 ~ + ~b
U A(U )U
2x

When the shock develops in figure 11.3(b) the numerical solution becomes
unstable using the method of lines.

11.2.3 Example III: Solution using Spectral Method

The matlab code is Shock spectral.m.

Ut = U Ux , 0 x 2
x
(we can change variable: = 2
later so that the range for is 0 1)

97
 Initial condition for U
1

0.9

0.8

0.7

0.6

0.5
U

0.4

0.3

0.2

0.1

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x

Figure 11.3: Initial conditions in (a) and solution for nonlinear Bugers equa-
tion using the method of lines in (b)

Spectral method
We let U (xj , tk ) = Ujk , xj = jx, j = 0, 1, . . . , 2n 1, tk = kt, k =
T
0, 1, . . . , m, and t = m .
Take the discrete Fourier transform of U :
2n1
X
U = F (U ) = U (xj , t) exp(ixj ) for = n + 1, . . . , n
j=0
j
where xj = jx =
n
then
n
1 X
Uj = F 1 (U ) = U exp(ixj )
2n =n+1
for j = 0, 1, . . . , 2n 1

and

Ujk 1 X n
= i U exp(ixj )
x 2n =n+1
= F 1 (i U )
= F 1 (i F (U ))

Use leap-frog for Ut :

Ujk+1 Ujk1
Ut =
2t
98
then Ut = U Ux becomes:

Ujk+1 = Ujk1 2tUjk F 1 (i F (Ujk ))

To find Uj1 for spectral method we assume wave speed 1 and:

Uj1 = U (x, t) = U 0 (x U t) = f (x U t)
f (x + t) = exp(10(4(x + t) 1)2 )

Initial condition for U Variation of U with time

0.8

1
0.6
U

0.8

0.6
U

0.4 0.1
0.4
0.08
0.2
0.06
0.2
0
0.04
0
0.2
0.4 0.02
0 0.6
0.8 x
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0
1
x
t

Figure 11.4: Initial conditions in (a) and solution for nonlinear Bugers equa-
tion using the spectral method in (b)

Again as in the method of lines the numerical solution becomes unstable as

shock develops in figure 11.4 using the spectral method.

99
Chapter 12

Korteweg-de Vries Equation

12.1 Solitons
solitons or solitary waves result from solution of the KdV equation

KdV equation is a model for shallow water waves:

Ut + U Ux + Uxxx = 0 nonlinear PDE

analytical solutions exist

solitons move in isolation and propagate without changing form. Ve-

locity is amplitude dependent (linearly proportional to maximum am-
plitude).

the nonlinear term causes waves to steepen (U Ux

the dispersive term causes waves to disperse (Uxxx )

these effects are in exact balance for solitons waveform maintains

its size, shape and speed as it travels.

solitons pass through each other without change of form except shifted.

12.2 Analytical solution

Ut + U Ux + Uxxx = 0

100
Let U = f () = f (x V t) where = x V t then:

= = V , = =
t t x x
df
Let f 0 () =
then Ut + U Ux + Uxxx = 0 becomes: (using U = f () =
f (x V t))

V f 0 + f f 0 + f 000 = 0
d f2
We integrate once: (use f f 0 = ( ))
d 2

Z
0
Z
d f2 Z
V f d + ( )d + f 000 d = 0
d 2
f2
V f + + f 00 = C (12.1)
2
Multiply by f 0 and integrate again:

2
0f
Z Z Z Z
0 0 00
V f f d + f d + f f d = Cf 0 d + c0
2
Z Z Z
V 2
Term 1 = V f f 0 d = V f 2 V f f 0 d V f f 0 d = f
2

2
Z
0f f3 Z 2 0 Z
f2 f3
Term 2 = f d = f f d f 0 d =
2 2 2 6

Z
0 00 0 2
Z
00 0
Z
0 00 f 02
Term 3 = f f d = (f ) f f d f f d =
2
So multiplying 12.1 by f 0 and integrating again gives:
V 2 f 3 f 02
f + + = Cf + C0 (12.2)
2 6 2

We assume boundary conditions f () 0, f 0 () 0 as .

So Equation 12.2 C0 = 0 and Equation 12.1 C = 0.
We assume initial conditions for soliton:
s
A
U (x, 0) = f (x) = Asech2 ( x)
12

101
 100
90
80
70
60
f(x)

50
40
30
20
10
0
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2
x

q
Figure 12.1: The initial conditions U (x, 0) = f (x) = Asech2 ( A
12
x).

So f (0) = A and f 0 (0) = 0 and f 0 () 0 for 0.

Rearranging Equation 12.2 gives

f3
f 02 = V f 2 (12.3)
3
Use
A
f (0) = A, f 0 (0) = 0 0 = A2 (V )
3
A
V =
3
Since f 0 () 0 for 0 we take the negative square root of 12.3:

f q
0
f = Af
3
This can be integrated analytically by making a change of variable if we let
f = Asech2 then df = 2Asech2 tanh d and integrating:
f q
0
f = Af
3
Since we assumed 0 we integrate from = 0 in integration limits:
df
Z
d
d 1 Z
= d
0 f Af 3 0
Z f () Z f
df df
= =
3 f (0) f Af A f Af

102
 now substitute f = Asech2
Z
2Asech2 tanh d
= q
0
Asech2 A Asech2
Z
2 tanh d
= q
0
A 1 sech2
now use 1 sech2 = tanh2
Z
2 2
d = =
0 A A 3
s
A
=
12
U (x, t) = f () = Asech2
s
A
= Asech2
12
s
A A A
U (x, t) = Asech2 (x t) (using V = )
12 3 3
| {z }
soliton travelling to right
1
where sechx = coshx
and coshx = 12 (ex + ex ).

12.3 Numerical solution of KdV Equation

Ut + U Ux + Uxxx = 0, 0 x 2
with periodic boundary conditions U (0) = U (2).
and initial conditions:
s
A
U (x, 0) = Asech2 ( (x )), A = 100
12
The analytical solution is:
s
A A
U (x, t) = Asech2 (x t)
12 3

If we apply the spectral method directly we find that the linear term, Uxxx ,
involves high frequencies making the numerical solution unstable as we will
see in section 12.3.1. Section 12.3.2 shows how to modify this term to gain
stability using a modified spectral method.

103
12.3.1 Solving directly with Spectral Method

Ut + U Ux + Uxxx = 0

The matlab code is SpectralDirectSoliton.m.

Take discrete Fourier transform of U :
2n1
X
U = F (U ) = U (xj , t) exp(ixj ), for = n + 1, . . . , n
j=0

and the inverse discrete Fourier transform of U :

n
1 1 X
Uj = F (U ) = U exp(ixj ), for j = 0, . . . , 2n 1
2n =n+1

We calculate spatial derviatives using spectral method:

Ujk 1 X n
Ux = = U (i) exp(ixj )
x 2n =n+1
= F 1 (i U ) = F 1 (iF (U ))

and:
3 Ujk 1 X n
Uxxx = = U (i 3 ) exp(ixj )
x3 2n =n+1
= F 1 (i 3 U ) = F 1 (i 3 F (U ))

At high wavenumbers, , this term causes instabilities in solution.
We use a leap-frog approximation for Ut :

Ujk U k+1 Ujk1

= j
t 2t
Plug approximations into PDE: Ut + U Ux + Uxxx = 0,
 
Ujk+1 = Ujk1 2t Ujk F 1 (iF (U )) + F 1 (i 3 F (U )) solution blows up!

Figure 12.2 shows that the numerical solution for the KdV equation blows
up using a direct spectral method. In the next section we modify the Uxxx
term causing instabilities.

104
 Initial condition for U Variation of U with time
100

90

80

70 300

60 200

50 100
U

U
40

30 100
3

20 200
2
5
10 300 x 10
1
0
2
0 4
0 1 2 3 4 5 6 7 6 0 t
8
x
x

Figure 12.2: Initial conditions in (a) and solution for nonlinear KdV equation
using the direct spectral method in (b)

12.3.2 Modifying Uxxx term causing instabilities in di-

rect spectral method
The matlab code is SpectralModifiedSoliton.m.
The direct method solves:

Ujk+1 = Ujk1 + 2t[Ujk F 1 (ivF (U )) + F 1 (iv 3 F (U ))]

The last term approximating Uxxx makes PDE very stiff at high wavenum-
bers.
To remove this instability for high wavenumbers we replace the last term
with:

sin(v 3 t) v 3 t + 0(t3 ) as t 0 this is satisfied

x3 x5
Using sin x = x 3!
+ 5!
. . . and re-solve with this approximation:

Ujk+1 = Ujk1 + 2tUjk F 1 (ivF (U )) + 2F 1 (i sin(v 3 t)F (U ))

This numerical solution is stable!

(See Fornberg and Whitham, Philos. Trans. Roy. Soc. London (1974))
Again use the same initial conditions:
A
 
Uj1 = U (xj , t) = U 0
x + t
3
A
 
= f x+ t
3
s
A A
 
2
= Asech ( x + t )
12 3

105
 Variation of U with time
120
Final solution at t=one period
Initial Condition at t=0
100
100
80

60
80

U
40

20

60 0
U

0.2

40 0.15

0.1
20

0.05 8
6 7
4 5
0 t 3
0 1 2 3 4 5 6 7 0 1 2
0
x x

Figure 12.3: Initial conditions and final solution after one period in (a) and
solution for nonlinear KdV equation using a modified spectral method in (b)

Figure 12.3 shows that the numerical solution for the KdV equation is sta-
ble using a spectral method where we have modified the Uxxx term causing
instabilities.
The method of integrating factors can also be used to remove the insta-
bility due to Uxxx term (see Trefethen).

12.3.3 Interacting Solitons

The matlab code is InteractingSoliton.m.
When 2 solitons travelling at different speeds collide their waveform main-
tains same size, shape and speed but the smaller (and slower) soliton is back-
ward shifted and the taller (and faster) soliton is forward shifted.
To show this feature of solitons we begin with intial conditions of two
solitons with speeds of V = 2A/3 and V = A/3:
s s
A 3 2A
U (x, 0) = f (x) = Asech2 (x ) + 2Asech2 (x )
12 2 12 2

where A = 100.
Figure 12.5 shows the numerical solution for the KdV equation for 2
interacting solitons using the modified spectral method. In figure 12.5(a) we
plot the initial conditions and final solution after one period. We see that
after the interaction the smaller soliton is backward shifted and taller soliton
forward shifted in time.

106
 200 V=2A/3

150
f(x)

100 V=A/3

50

0
0 1 2 3 4 5 6
x

Figure 12.4: The initial conditions U (x, 0) = f (x) is plotted to show the 2
solitions and their speeds

Variation of U with time

250
Final solution at t=one period
Initial Condition at t=0

200
250

200
150
150

100
U

100
U

50

0
50
50
0.2
0.15 8
0
6
0.1
4
0.05
2
50
0 1 2 3 4 5 6 7 t 0 0 x
x

Figure 12.5: Initial conditions and final solution after one period in (a) and
solution for nonlinear KdV equation for two interacting solitons in (b)

107