6G6Z3004 Numerical Methods for Partial Differential

Equations II

Finite-Difference Methods
Dr Jon Shiach

2019 – 2020

© Department of Computing and Mathematics

Contents

1 Partial Differential Equations 1

1.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Classifying PDEs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Domain of dependence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4 Using PDEs to model real world phenomena . . . . . . . . . . . . . . . . . . . . . . . . 7

2 Finite-Difference Approximations 11
2.1 The Taylor series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2 “Big-oh” notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.3 Deriving finite-difference approximations . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.4 The finite-difference toolkit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.5 Example of a finite-difference scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

3 Elliptic Partial Differential Equations 31

3.1 Discretising a two-dimensional domain . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.2 Boundary conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.3 Deriving a FDS to solve Laplace’s equation . . . . . . . . . . . . . . . . . . . . . . . . 34
3.4 Solving Laplace’s equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.5 The Jacobi method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.6 Convergence criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.7 Solving a linear system using an indirect method . . . . . . . . . . . . . . . . . . . . . 39
3.8 The Gauss-Seidel method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.9 The Successive Over Relaxation (SOR) method . . . . . . . . . . . . . . . . . . . . . . 45
3.10 Line SOR method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

4 Hyperbolic Partial Differential Equations 53

4.1 The advection equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.2 The First-Order Upwind (FOU) scheme . . . . . . . . . . . . . . . . . . . . . . . . . . 54
4.3 Validating the time marching schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.4 The Forward-Time Central-Space (FTCS) scheme . . . . . . . . . . . . . . . . . . . . 60
4.5 Convergence, consistency and stability . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.6 The Lax-Friedrichs scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.7 The Leapfrog scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.8 The Lax-Wendroff scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.9 The Crank-Nicolson scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.10 Analysis of finite-difference schemes used to solve the advection equation . . . . . . . . 72

5 Parabolic Partial Differential Equations 75

5.1 Heat diffusion equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
5.2 Two dimensional heat diffusion equation . . . . . . . . . . . . . . . . . . . . . . . . . . 77
5.3 Grid refinement study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80

6 Solving Multidimensional Partial Differential Equations 85

i
Contents

6.1 The two-dimensional advection equation . . . . . . . . . . . . . . . . . . . . . . . . . . 85

6.2 Operator splitting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
6.3 Dimensional splitting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
6.4 Term splitting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
6.5 Operator splitting for stiff problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95

7 Solving Systems of Partial Differential Equations 97

7.1 The Shallow Water Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
7.2 Estimating the wave speeds of the SWE . . . . . . . . . . . . . . . . . . . . . . . . . . 99
7.3 Solving the shallow water equations using finite-difference schemes . . . . . . . . . . . 101
7.4 Two-dimensional shallow water equations . . . . . . . . . . . . . . . . . . . . . . . . . 109
7.5 Validating the two-dimensional SWE solver . . . . . . . . . . . . . . . . . . . . . . . . 111

References 113

A Derivation of the Thomas algorithm 117

Finite-Difference Methods ii Dr Jon Shiach

Part II - Finite Difference Methods

In part I of this unit you were introduced to methods used find the analytical solution to a PDE.
Unfortunately, for most practical applications attempting to find the analytical solution is either very
difficult or not possible so we need to use a numerical approach instead. There are several different
numerical methods that can be used to solve a PDE and in this unit we will be focusing on two of
the most common of these: finite-difference methods (taught in term 1) and finite-volume methods
(taught in term 2).

Term 1 Teaching Schedule – Dr Jon Shiach

Week Date (w/c) Material
5 21/10/2019 Chapter 1 Partial Differential Equations:
Definitions and terminology, classifying PDEs, use of PDEs in modelling,
Eulerian and Lagrangian methods.

Chapter 2 Finite-Difference Approximations

The Taylor series; the truncation error; first and second-order finite-difference
approximations; method of undetermined coefficients; the finite-difference
toolkit.
6 28/10/2019 Chapter 3 Elliptic PDEs:
Laplace’s and Poisson’s equations; discretising a 2D domain; Jacobi, Gauss-
Seidel and SOR methods; the L2 error.
7 04/11/2019 Chapter 4 Hyperbolic PDEs:
The advection equation; the first-order upwind and forward-time central-
space schemes
8 11/11/2019 Chapter 4 Hyperbolic PDEs continued:
Convergence, consistency and stability; von Neumann stability analysis; Lax-
Friedrichs, Leapfrog, Lax-Wendroff and Crank-Nicolson schemes
9 18/11/2019 Chapter 5 Parabolic PDEs:
The heat diffusion equation; grid refinement study
10 25/11/2019 Chapter 6 Solving Multidimensional PDEs:
Differential marching operator; dimensional splitting; term splitting

Coursework assignment handed out

11 02/12/2019 Chapter 7 Systems of PDEs:
The Shallow Water Equations; linearising a system of PDEs; the Jacobian
matrix
12 09/12/2019 Consolidation of Finite-Difference Methods

iii
Chapter 1

Partial Differential Equations

This chapter will introduce the concept of partial differential equations, the different types and the
general methodology used to solve partial differential equations using numerical methods.

1.1 Definitions
1.1.1 Partial derivatives
A partial derivative of a function of several variables is the derivative of that function with respect to
one of those variables, the other variables being treated as constants. The partial derivative is denoted
by the symbol ∂ instead of the ‘d’ that is used for the derivative of a single variable function. The
same rules for derivatives of single variable functions apply to partial derivatives.
For example, let U (x, y) = x2 + xy be a two-variable differential function then U can be differentiated
with respect to x and y, e.g.,
∂U
= 2x + y,
∂x
∂U
= x.
∂y
The partial derivatives are commonly denoted using subscripts, i.e.,
∂U
= Ux = ∂x U.
∂x
Higher order partial derivatives can be represented by repeating the variable in the subscript, i.e.,

∂2U
= Uxx = ∂xx U.
∂x2
Mixed partial derivatives can be formed by differentiating a derivative with respect to a different
variable, i.e.,    
∂ ∂U ∂ ∂U ∂2U
= = = Uxy = Uyx = ∂xy U = ∂yx U.
∂x ∂y ∂y ∂x ∂x∂y

1.1.2 Partial differential equations

A Partial Differential Equation (PDE) is a differential equation that is expressed in terms partial
derivatives of a function with two or more variables. For example,

aUx + bUy + cUxx + dUyy + eUxy = f,

here U is an unknown function of the independent variables x and y and a, b, c, d, e and f are either
known functions of x and y or known constants.

1
Chapter 1. Partial Differential Equations

1.1.3 Order of a PDE

The order of a PDE is the order of its highest derivative. For example,

aUx + bUy = 0,

is a first-order PDE whereas

aUxx + bUy = 0,

is a second-order PDE.

1.1.4 Linear PDEs

A PDE that can be expressed as a linear combination of first-order partial derivatives is said to be a
linear PDE. For example,
aUx + bUy + cUz = f,

is a linear PDE.

1.1.5 Domain
The domain is defined as the set of all values of the independent variables for which there exists a
solution to a PDE. In most cases the domain can be split into the spatial domain which is the subset
of the domain for the spatial variables, x, y, z etc., and the time domain which is the subset of the
domain for the difference values of the time variable t. Often the spatial domain is simply referred to
as the domain even if time is also an independent variable of the PDE.

1.1.6 Initial conditions

The initial conditions is a function U of the known solution at the initial point in time. For PDEs
that do not have time as an independent variable the initial conditions be any function. If U (t, x) is a
function of two variables where x is spatial position and t is time then the initial conditions at t = 0
is the function U (0, x).

1.1.7 Boundary conditions

The boundary conditions determine the value of U at the boundary of the spatial domain. If U (t, x)
is a function of two variables where a ≤ x ≤ b is the spatial position and t is time then the boundary
conditions are described by the functions U (t, a) and U (t, b).

1.1.8 Solution of a PDE

The solution of a PDE is the function U that satisfies:
• the PDE itself;
• any boundary conditions;
• any initial conditions.
Most PDEs do not have exact analytical solutions so a numerical approach is required. Numeri-
cal methods for solving PDEs calculate an approximate solution to the PDE at discrete values of
the independent variables using a numerical method that is usually implemented using a computer
program.
Since a numerical solution is only an approximation of the exact solution U , it is common to denote
the numerical solution using the lowercase character u.

Finite-Difference Methods 2 Dr Jon Shiach

 Chapter 1. Partial Differential Equations

Example 1

Show that the following PDE

Ut = kUxx , (1.1)

with the following initial and boundary conditions

 πx 
U (0, x) = 2 sin ,
L
U (t, 0) = 0,
U (t, L) = 0,

has the solution

 πx   
kπ 2 t
U (t, x) = 2 sin exp − 2 . (1.2)
L L

Solution: We need to show that the function in equation (1.2) satisfies the PDE in equa-
tion (1.1), the initial conditions and the boundary conditions. The PDE is expressed in terms
of the derivatives Ut and Uxx , evaluating these gives
 πx   
2kπ 2 kπ 2 t
Ut = − 2 sin exp − 2 ,
L L L
2    
2π πx kπ 2 t
Uxx = − 2 sin exp − 2 .
L L L

Substituting into equation (1.1) gives

 πx     πx   
2kπ 2 kπ 2 t 2kπ 2 kπ 2 t
− 2 sin exp − 2 = − 2 sin exp − 2 ,
L L L L L L

since LHS = RHS then the function in equation (1.2) satisfies the PDE in equation (1.2). Now
consider the initial conditions
 πx   
kπ 2 (0)
U (t = 0, x) = 2 sin exp −
L L2
 πx 
= 2 sin exp(0)
L
 πx
= 2 sin ,
L
therefore equation (1.2) satisfies the initial conditions. Finally consider the boundary conditions
     
π.0 kπ 2 t kπ 2 t
U (t, x = 0) = 2 sin exp − 2 = 2 sin(0) exp − 2 = 0,
L L L
     
π.L kπ 2 t kπ 2 t
U (t, x = L) = 2 sin exp − 2 = 2 sin(π) exp − 2 = 0,
L L L

which are both satisfied. Since equation (1.2) satisfies the PDE, the initial conditions and the
boundary conditions then we can say that this is a solution to the PDE in equation (1.1).

Dr Jon Shiach 3 Finite-Difference Methods

Chapter 1. Partial Differential Equations

1.2 Classifying PDEs

Linear second-order PDEs can be classified as one of three types: parabolic, hyperbolic and elliptic.
Consider the second-order linear PDE

aUxx + bUxy + cUyy + dUx + eUy + f = 0. (1.3)

A PDE of this form can be classified by the following



< 0, elliptic,
b − 4ac = 0, parabolic,
2


> 0, hyperbolic.

The terminology used to classify PDEs comes from that used to classify conic sections (Hoffman, 2001)
(figure 1.1). Conic sections are described by the equation

ax2 + bxy + cy 2 + dx + ey + f = 0, (1.4)

and the type of curve represented by equation (1.4) is classified as follows



< 0, ellipse,
b − 4ac = 0,
2
parabola,


> 0, hyperbola.

It should be obvious that there are similarities between the types of PDEs and the different conic
sections. However, there is no significance between the names of the types of PDEs and the behaviour
of their solutions, they are simply names given to the different types and do not share any properties
with the different conic sections.

(a) Ellipse (b) Parabola (c) Hyperbola

Figure 1.1: Conic sections

Example 2

For each of the PDEs given below, classify them as either an elliptic, parabolic or hyperbolic
PDE.
(i) Uxx + Uyy = 0;
(ii) Ut = kUxx ;
(iii) Utt = c2 Uxx .

Solution:
(i) Comparing the PDE to equation (1.3) we have a = 1, b = 0 and c = 1 therefore

b2 − 4ac = 0 − 4 = −4 < 0,

so this is an elliptic PDE.

Finite-Difference Methods 4 Dr Jon Shiach

 Chapter 1. Partial Differential Equations

(ii) Here we have a = −k, b = 0 and c = 0 therefore

b2 − 4ac = 0 + 0 = 0,

so this is a parabolic PDE.

(iii) Letting t = y we have a = −c2 , b = 0 and c = 1 therefore

b2 − 4ac = 0 + c2 > 0

for all c ∈ R so this is a hyperbolic PDE.

1.2.1 Elliptic PDEs

Elliptic PDEs are associated with steady state behaviour where the behaviour of the solution does not
depend upon a time variable. A steady state problem is defined by the elliptic PDE and the boundary
conditions that are applied to every point on the boundary. Any initial conditions defined in the
domain will not affect the solution. Solutions to elliptic PDEs are smooth with no discontinuities.
Examples of elliptic PDEs are Laplace’s equation

Uxx + Uyy = 0,

and Poisson’s equation

Uxx + Uyy = f (x, y),

which can be used to model magnetic, electrical and gravitational fields.

1.2.2 Parabolic PDEs

Parabolic PDEs are associated with problems that include the diffusion of some quantity across the
domain, i.e., a propagation problem. Like elliptic PDEs, the solutions are smooth with no discontinu-
ities but unlike elliptic PDEs, parabolic PDEs allow for transport to be modelled with the inclusion
of the time variable. Therefore the solutions are not steady state problems and will evolve over time
and require initial conditions that satisfy the PDE at the lower bound of the time domain.
An example of a parabolic PDE is the heat diffusion equation

Ut = αUxx ,

where α is some heat diffusion coefficient that determines the conductivity of a material. Since t is an
independent variable the solution to this equation will evolve over time.

1.2.3 Hyperbolic PDEs

Hyperbolic PDEs are associated with propagation problems where some quantity or change in quantity
travels across the domain. However, unlike parabolic PDEs where the solution at a point in the domain
depends upon the solution at all other points in the domain, in the hyperbolic case the solution only
depends upon points in the neighbourhood region. Examples of hyperbolic PDEs include the wave
equation

Utt = c2 Uxx ,

where c is the speed of sound and is used in the modelling of acoustic waves and the advection equation

Ut + vUx = 0,

where v is the advection velocity which is used to transport of a pollutant along a channel.

Dr Jon Shiach 5 Finite-Difference Methods

Chapter 1. Partial Differential Equations

1.3 Domain of dependence

The differences between the three types of PDEs defined in section 1.2 can be illustrated by considering
what is known as the domain of dependence. Let U (t, x) be the solution of a PDE at a given point in
space x and timet.

Definition 1: Domain of dependence

The domain of dependence is defined by the set of all points in the domain that influences the
value of U (t, x).

Definition 2: Range of influence

The range of influence is defined as the set of all points in the domain for which U (t, x) is
a member of the domain of dependence, i.e., the set of all points that the value of U (t, x)
influences.

Consider an elliptic PDE where the spatial domain is defined in the xy-plane with a ≤ x ≤ b and
c ≤ y ≤ d (figure 1.2). As mentioned in section 1.2 elliptic PDEs have a steady state solution and do
not change over time. As such, for elliptic PDEs the domain of dependence is the same as the range
of influence and includes all points in the domain.

y boundary

range of influence

boundary U (x, y) boundary

domain of dependence

c x
a b
boundary

Figure 1.2: Domain of dependence and range of influence for an elliptic PDE.

The solution to a parabolic PDE must satisfy the PDE, the boundary conditions and, because parabolic
PDEs include a time variable, the initial conditions for the PDE. Since parabolic PDEs model propa-
gation, the solution U (t, x) is dependent upon the solution at previous time values. Also, the solution
across the spatial domain depends upon all other points at the same time level. Therefore the domain
of dependence for a parabolic PDEs is bounded by the spatial boundaries x = a and x = b and the
initial conditions at t = tmin (figure 1.3). The range of influence is the set of all points at later time
levels up until t = tmax .

Finite-Difference Methods 6 Dr Jon Shiach

 Chapter 1. Partial Differential Equations

t
tmax

range of influence boundary

U (t, x)
tn

boundary domain of dependence

tmin x
a b
initial conditions

Figure 1.3: Domain of dependence and range of influence for a parabolic PDE.

Hyperbolic PDEs model propagation of a quantity across the spatial domain but unlike elliptic and
parabolic PDEs, the solution of U (t, x) is not influenced by all points in the spatial domain at once. For
example, if a disturbance occurs at a particular point, the effect of the disturbance will propagate away
from this point causing wave like structures. Only those points within the region of the propagating
waves after a certain amount of time will be affected by the disturbance. Therefore the domain of
dependence of a hyperbolic PDE is dependent upon the speed of propagation. Let c be the fastest
propagation speed permitted by the hyperbolic PDE, then the domain of dependence is bounded by
two lines called characteristics that extend back through time from U (t, x) to the initial conditions at
t = tmin . The range of influence is also bounded by the characteristics and extends forward in time
until t = tmax (figure 1.4).

t
tmax
range of
influence boundary

tn U (t, x)
x0 + ct x1 − ct
boundary domain of
dependence
tmin x
a x0 x1 b
initial conditions

Figure 1.4: Domain of dependence and range of influence for a hyperbolic PDE.

1.4 Using PDEs to model real world phenomena

The methodology for using PDEs to model real world phenomena will depend on the problem or
scenario that is being investigated, but the for most cases the following steps are used:
1. Build the model – an investigation will usually be prompted by a brief that outlines the
purpose of the study and what questions need answering. In almost all cases, assumptions that
simplify the development of the model are applied, for example, if modelling heat transfer of an
object we could assume that the material has uniform heat conduction properties throughout.
These assumptions should simplify the model but still retain the behaviour of the phenomena
that is being modelled. The model can then be created using PDEs, or systems of PDEs where

Dr Jon Shiach 7 Finite-Difference Methods

Chapter 1. Partial Differential Equations

the dependent variables provide information of the state of the system. In almost all instances,
the PDEs that are used in modelling have already been derived previously and are then applied
to the specific case.

Mathematical modelling usually begins with a very simplified approach that provides a basic
model. This is then improved by including additional terms or equations that enable the model
to include other properties of the system. For example, the modelling of water waves may begin
with a model that can approximate the depth and velocity of the water over a flat bed surface,
this can then be improved by including terms that model bed topography, bed friction, wave
breaking etc.

2. Derive the numerical scheme – the numerical method chosen is based upon the PDE being
solved, the geometry of the physical domain, the accuracy required, the computational resources
available and the behaviour of the phenomena being modelled. There are several different
methods currently available to do this, examples include finite-difference methods, finite-volume
methods, finite-element methods, spectral methods and particle based methods. The complete
numerical model may use a single method or combination of two or more methods.

3. Write the program – a computer program is written to perform the calculations of the numer-
ical scheme. This can be time consuming and frustrating step. Do not expect to write a working
program straight away, even experienced computer programmers rarely write an error-free pro-
grams on the first attempt. The process of writing a program should be done in steps, once a
new part of the program is written it should be tested to see whether it performs as required.
Only when you are satisfied that the new part performs as required you should proceed to write
the next step in the program.

4. Verify the program – just because a program is working and produces an output does not
mean that the output is what the numerical scheme should produce. Minor errors called ‘bugs’
in the program will affect the numerical solution and render the model useless. To verify that
the program is performing as expected, the numerical solution is obtained using pen and paper
(know as performing a ‘hand calculation’) and compared to the results from the program. Since
performing the calculations by hand can be a time consuming process, the program is tested
using a small number of solution points. Care should be taken to ensure that every variable
and array calculated in the program matches that of the hand calculation. It is important that
you know what every line in the program does and understand the structure of the program.
It can be difficult to work with a program written by someone else but by stepping through
the program line-by-line will give you an understanding of what it does (this is when a well
commented program is useful).

5. Validate the model – once the program is performing as expected the next step is to validate
the model by comparing the output against real world data, either from controlled experiments
or real life measurements. Remember that a mathematical model is only an approximation of a
real world scenario and the assumptions and compromises made in developing the model may
mean that the model solutions may not exactly replicate the real world data. What is important
is that the behaviour of what is being modelled is replicated.

6. Present the model – the final step in the modelling process is to present the model, the results
and any conclusions that have been arrived at during the investigation. This usually takes the
form of a report or a paper published in a journal or conference proceedings. It is important
that the model is explained in enough detail so that the reader can reproduce the results for
themselves and apply the model to other cases. This step can be daunting when first encountered
and takes practice to get good at it. In this unit you will be asked to produce reports on work
in the computer labs as part of formative assessment in addition to your summative coursework
assessment. Guidance material on how to write mathematical reports is provided on Moodle.

Finite-Difference Methods 8 Dr Jon Shiach

 Chapter 1. Partial Differential Equations

1.4.1 Eulerian and Lagrangian methods

Consider the solution U to a particular PDE with the spatial domain Ω. If we were to use a numerical
method to solve U then it stands to reason that we will need to calculate our solution at a finite
number of discrete points in Ω else it would take forever to calculate the solution. Numerical solutions
methods are classified into two types based on the method used to discretise the domain: Eulerian
methods and Lagrangian methods.
Eulerian methods calculate the solution at points in the domain that have a fixed position (figure 1.5).
The values of the variables in the PDE such as pressure, density, temperature and velocity etc. are
calculated based on these fixed points. Examples of Eulerian methods include finite-difference, finite-
volume and finite-element methods.
y Ω

Figure 1.5: An Eulerian method calculates the solution at points in the domain whose
positions are fixed.

Lagrangian methods calculate the solution at points that move around the domain based on the
flow conditions of the solution (figure 1.6). The values of the variables such as pressure, density,
temperature etc. are calculated at these points and the velocities are used to determine the position.
Examples of Lagrangian methods, which are also known as mesh free methods include Smoothed
Partial Hydrodynamics (SPH) and Moving Particle Semi-implicit (MPS) methods.

y Ω

Figure 1.6: An Lagrangian method calculates the solution at points in the domain whose
positions change depending on the flow conditions.

1.4.2 Exact solution versus numerical approximation

It should always be noted that when using a numerical method to calculate the solution of a PDE,
we are not actually solving the PDE, rather we are calculating an approximation of the solution. Of
course we aim to make our numerical approximation as close a possible to the exact solution. There
are three factors that determine the accuracy of a numerical approximations
• The accuracy of the numerical method. There are many different methods that we can employ
to solve a PDE with varying degrees of accuracy. In general, more accurate methods require

Dr Jon Shiach 9 Finite-Difference Methods

Chapter 1. Partial Differential Equations

more complicated derivations and more computational resources. In practice we often reach a
compromise between the accuracy of our chosen method and the effort required to derive an
implement a more accurate method.
• The number of discrete points where we are calculating the solution. The more points we use
to calculate a solution the more accurate our solution will be. We are limited to the number of
points by the computational resources available.
• Computational rounding errors. There is a limit to the accuracy of a floating point operation
(i.e., operations involving decimals) performed by a computer. Computational rounding errors
can compound to affect the accuracy of a numerical solution.

Finite-Difference Methods 10 Dr Jon Shiach

Chapter 2

Finite-Difference Approximations

Finite-difference methods are one of the simplest and most common methods used to solve PDEs.
Approximations of partial derivatives that use the difference of the function value at two or more
points are derived from the well known Taylor series. The derivatives in the PDE are replaced by these
difference approximations to form an algebraic system, the solution of which gives an approximation
to the solution of the PDE. This chapter will introduce finite-differences, show how to derive finite-
difference approximations and how we can use them to solve PDEs.

2.1 The Taylor series

The Taylor series, named after English mathematician Brook Taylor (1685–1731), is an infinite series
expansion of a function written in terms of the derivatives of the function about a single point. Let
f (x) be a differentiable function and h be some small positive quantity then the Taylor series is written
as
∞
X hn
f (x + h) = f (x),
n! (n)
n=0

where f(n) (x) denotes the nth derivative of f (x) with respect to x. Writing out the first few terms
gives
Forward Taylor Series

h2 h3 h4
f (x + h) = f (x) + hfx (x) + fxx (x) + fxxx (x) + f(4) (x) + · · · (2.1)
2! 3! 4!

The quantity h is known as the step length since we are approximating the function f a small step
along the number line from x. As h is a positive quantity then x + h is said to be forward of x so
equation (2.1) is colloquially known as the forward Taylor series. We can also approximate f at x − h
simply by replacing h by −h in equation (2.1) to give

Backwards Taylor Series

h2 h3 h4
f (x − h) = f (x) − hfx (x) + fxx (x) − fxxx (x) + f(4) (x) + · · · (2.2)
2! 3! 4!

Since x − h is a backwards step along the number line from x equation (2.2) is known as the backwards
Taylor series. Note how in the backwards Taylor series the odd order terms have negative coefficients
whilst the even order terms have positive coefficients, i.e.,
∞
X hn
f (x − h) = (−1)n f (x).
n! (n)
n=0

11
Chapter 2. Finite-Difference Approximations

2.1.1 Truncating the Taylor series

The Taylor series will give an exact values of f (x + h) if we sum over an infinite number of terms. This
is not possible in practice so we use the Taylor series to give an approximation of f (x + h) by only
including the first few terms in the series summation. In doing so we are truncating the Taylor series
so that we omit the higher order terms. For example, truncating equation (2.1) after the first-order
term gives
f (x + h) ≈ f (x) + hfx (x). (2.3)
This is known as the first-order forward Taylor series since the highest order of the derivative in the
series is a first-order derivative.

2.1.2 Truncation error

Truncating the Taylor series is useful as it greatly reduces the amount of computations we have to
do, however we pay the price of a reduced accuracy in our approximation. We need to be able to
gauge how inaccurate our series approximations are as a result of truncation and to do so we use the
truncation error which is the error that is attributed to the omission of the higher order terms when
truncating the Taylor series, i.e.,

f (x + h) = series approximation + truncation error.

In most cases we cannot calculate an exact value of the truncation error (if we could do that there
would be no need to truncate the Taylor series in the first place) but we can examine the behaviour
of the truncation error. Consider the truncation error for the first-order forward Taylor series

h2 h3
f (x + h) = f (x) + hfx (x) + fxx (x) + fxxx (x) + · · · .
| {z } |2! 3!
{z }
1st order approximation truncation error

Notice that the values of the denominators are factorials so the first term omitted will have a much
larger value than any of the other terms and is therefore the dominant component of the truncation
error. For example, consider the first-order Taylor series approximation of cos(0.1) using equation (2.3)
with x = 0 and h = 0.1
cos(0 + 0.1) ≈ cos(0) − 0.1 sin(0) = 1.
Since cos(0.1) = 0.995004 (correct to 6 decimal places) then our truncation error is |0.995004 − 1| =
0.004996. The absolute value of the lowest order omitted term is
 2 2   
h ∂   0.12 
   
 2! ∂x2 cos(x) = − 2! cos(0) = 0.005,

so the error attributed to the other omitted terms is just |0.004996 − 0.005| = 4 × 10−6 .
The truncation error is dominated by lowest order term omitted we can use this to express the
truncation error as a polynomial function of h such that

f (x + h) = series expansion + E(h) + higher order terms.

For an nth order Taylor series expansion

 n+1 
 h 

E(h) =  f(n+1) (x) .
(n + 1)!

E(h) only includes the largest term from those omitted in the truncation of the Taylor series the actual
truncation error will always be less than E(h). Therefore we can say that

“As h → 0 the truncation error will tend to zero at least as fast as E(h).”

Finite-Difference Methods 12 Dr Jon Shiach

 Chapter 2. Finite-Difference Approximations

The truncation errors for the first-order Taylor series approximation of cos(x + h) using values of
h = 0.5, 0.25, 0.125, 0.0625 and the function E(h) for a first-order approximation have been plotted in
figure 2.1. Note is that as the value of the step length h decreases, so does the truncation error. This
is an important result that applies to all numerical methods, i.e.,

“The smaller the step length, the more accurate the approximation.”

0.12
truncation error
E(h) = |− 12 h2 cos(1)|
0.1
truncation error

8 · 10−2

6 · 10−2

4 · 10−2

2 · 10−2

5 · 10−20.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

h

Figure 2.1: Plot of truncation errors for the first-order Taylor series expansion of cos(x + h)
for h = 0.5, 0.25, 0.125 and 0.0625 compared with E(h).

The truncation errors demonstrate quadratic behaviour as they tend to zero as h decreases. This is
because the dominant term in the truncation error is a quadratic polynomial in terms of h. Finally
the truncation errors are always less than the values of E(h). In most practical cases it will not be
possible to calculate the truncation error so we use E(h) as a worst possible scenario to examine the
behaviour of the truncation errors in numerical schemes.

2.2 “Big-oh” notation

We have seen in the previous section that instead of calculating a value for the truncation error, we
examine the behaviour of the dominant term in the truncation error for different values of h. For
example when comparing numerical methods, the method with a truncation error that tends to zero
fastest as we decrease the step length will provide the more accurate approximation. Comparison of
this convergence behaviour is done using “big-oh” notation.

Definition 3: O(hn )

Let f (h) be a function of h and define f (h) = O(hn ) such that the following is satisfied

f (h)
lim = C, (2.4)
h→0 hn

where C is a non-zero positive constant.

Dr Jon Shiach 13 Finite-Difference Methods

Chapter 2. Finite-Difference Approximations

Consider the case when h is small

f (h)
≈C
hn
∴ f (h) ≈ Chn .

What this tells us is that since C > 0, the values of f (h) will always be less than that of Chn and “as
h → 0, f (h) → 0 at least as fast as hn → 0”. If this is true then we can write f (h) = O(hn ) and we say
that “f (h) is order h to the n”. For example, consider the function f (h) = 2h3 , using equation (2.4)

2h3
lim = C,
h→0 hn

and since C is a scalar quantity n must be 3 in order to cancel out the h in the numerator so we can
say that 2h3 = O(h3 ). Also consider g(h) = sin(h), using the series expansion of sin(h) we have

h3 h5 h7
sin(h) h− + + + ···
lim = lim 3! 5! 7! =C
h→0 hn h→0 hn
h3 h5 h7
∴h− + + + · · · ≈ Chn .
3! 5! 7!
h3 h5 h7
Since h − + + + · · · < h then n = 1 and we can say that sin(h) = O(h).
3! 5! 7!

2.2.1 Expressing the truncation error as O(hn+1 )

We use big-oh notation to denote the errors attributed to the omission of the higher order terms in
the truncation of of the Taylor series. Let E(h) = O(hn ) then the value of n will be one more than
the order of the Taylor series expansion, e.g.,

h2 hn
f (x + h) = f (x) + hfx (x) + fxx (x) + · · · + f (x) + O(hn+1 ).
2! n! (n)

For example, a first-order Taylor series expansion, n = 1 and the truncation error is represented by
O(h2 ), i.e.,
f (x + h) = f (x) + hfx (x) + O(h2 ),
and similarly for a second-order Taylor series

h2
f (x + h) = f (x) + hfx (x) + fxx (x) + O(h3 ).
2
This means that for a first-order Taylor expansion, reducing the step length h by a factor of 21 will
result in the error being reduced by a factor of at least h2 = ( 21 )2 = 14 . For a second-order Taylor
series expansion, reducing h by the same factor will result in the the error being reduced by a factor
of at least h3 = ( 12 )3 = 81 .

2.2.2 Properties of O(hn )

The following lists some useful properties of O(hn ):
• Multiplying the truncation error by a scalar does not change the convergence behaviour of the
truncation error, i.e.,

kO(hn ) = O(hn ).

Finite-Difference Methods 14 Dr Jon Shiach

 Chapter 2. Finite-Difference Approximations

• When summing series expansions of different orders, the truncation error of the lower order
expansion dominates, i.e., if m > n

O(hm ) + O(hn ) = O(hn ).

• Dividing O(hn ) by hm where m < n reduces the value of n by m, i.e.,

O(hn )
= O(hn−m ). (2.5)
hm

f (h) m
Proof. Assuming equation (2.5) is true, let = O(hn ) and using equation (2.4)
h
f (h)
m f (h) f (h)
lim hn−m
= lim n−m+m = lim n = C.
h→0 h h→0 h h→0 h

2.3 Deriving finite-difference approximations

Finite-differences approximate the derivatives of a function using the values of that function at equally
spaced discrete points in the domain. For example consider the derivative of a function f (x), we can
approximate the value of the derivative at the point x by using the values of f (x − h), f (x) and
f (x + h).

h
x
f (x − 2h) f (x − h) f (x) f (x + h) f (x + 2h)

Figure 2.2: Neighbouring values are used to approximate the derivative of f (x).

2.3.1 First-order accurate finite difference approximation of fx (x)

The derivation of a finite-difference approximation is achieved by truncating the Taylor series and
rearranging to make the desired derivative the subject of the equation. Consider the first-order forward
Taylor series
f (x + h) = f (x) + hfx (x) + O(h2 ).
Rearranging to make fx (x) the subject gives
f (x + h) − f (x) O(h2 )
fx (x) = + .
h h
O(h2 )
Since = O(h) by equation (2.5) then
h
First-order forward difference

f (x + h) − f (x)
fx (x) = + O(h). (2.6)
h

Equation (2.6) is known as the first-order forward difference approximation of fx (x). Alternatively
we can also use the first-order backwards Taylor series
f (x − h) = f (x) − hfx (x) + O(h2 ),

Dr Jon Shiach 15 Finite-Difference Methods

Chapter 2. Finite-Difference Approximations

and doing similar gives the first-order backward difference approximation of fx (x)

First-order backward difference

f (x) − f (x − h)
fx (x) = + O(h). (2.7)
h

2.3.2 Second-order accurate finite-difference approximation of fx (x)

Equations (2.6) and (2.7) are both first-order accurate approximations as denoted by O(h). To con-
struct a second-order accurate approximation we use the second-order forwards and backwards Taylor
series
h2
f (x + h) = f (x) + hfx (x) + fxx (x) + O(h3 ), (2.8a)
2
h2
f (x − h) = f (x) − hfx (x) + fxx (x) + O(h3 ). (2.8b)
2
Since we wish to approximate fx (x) we need to eliminate the fxx (x) terms. This can be done by
subtracting equation (2.8b) from (2.8a) to give

f (x + h) − f (x − h) = 2hfx (x) + O(h3 ),

which can be rearranged to give the second-order central difference approximation of fx (x).

Second-order central difference

f (x + h) − f (x − h)
fx (x) = + O(h2 ). (2.9)
2h

2.3.3 Second-order accurate finite-difference approximation of fxx (x)

To derive a finite-difference approximation of a higher order derivative we need to truncate the Taylor
series to an order greater than the order of the derivative. For example, to derive a finite-difference
approximation of fxx (x) we require the third-order forward and backward Taylor series

h2 h3
f (x + h) = f (x) + hfx (x) + fxx (x) + fxxx (x) + O(h4 ),
2 6
h2 h3
f (x − h) = f (x) − hfx (x) + fxx (x) − fxxx (x) + O(h4 ).
2 6
Here we want to approximation fxx (x) so we need to eliminate the fx (x) and fxxx (x) terms. Since
these have different signs in the forward and backward Taylor series expansions we can do this by
summing the two, i.e.,

f (x − h) + f (x + h) = 2f (x) + h2 fxx (x) + O(h4 )

Rearranging to make fxx (x) the subject we have

Second-order symmetric difference

f (x − h) − 2f (x) + f (x + h)
fxx (x) = + O(h2 ). (2.10)
h2

This is the second-order symmetric difference approximation of fxx (x). Note that when summing the
forward and backwards Taylor series expansion the odd order terms cancel out. This means that using
an odd order Taylor series expansion will result in an approximation that has an order of accuracy

Finite-Difference Methods 16 Dr Jon Shiach

 Chapter 2. Finite-Difference Approximations

one less than the order of expansion. For this reason, apart from the first-order approximations, the
order of finite-difference approximations increase as even numbers.
The forward, backward and central difference formulae have been used to approximate the slope of the
tangent to a function f (x) in figure 2.3. Note that the forward and backward differences give a poor
approximation of the tangent whereas the central difference is much more accurate. This is because
the central difference is second-order accurate whereas the forward and backward differences are only
first-order accurate.
f (x) f (x) f (x)

xi xi + h x xi − h xi x xi − h xi xi + h x

(a) forward difference (b) backward difference (c) central difference

Figure 2.3: Difference approximations of the tangent slope of f (x) at x = xi (dashed line).

Example 3

Use first and second-order accurate finite-differences to approximate the first and second deriva-
tive of f (x) = cos(x) at x = π4 using a step length of h = 0.1. Calculate the absolute error
between your approximations and the exact value.

Solution: Using the first-order forward difference from equation (2.6) we have
π  π 
 π  cos + 0.1 − cos
fx ≈ 4 4 = −0.741255.
4 0.1
Using the first-order backward difference from equation (2.7) we have
π  π 
 π  cos( − cos − 0.1
fx ≈ 4 4 = −0.670603.
4 0.1
Using the second-order central difference formula from equation (2.9) we have
π  π 
 π  cos + 0.1 − cos − 0.1
fx ≈ 4 4 = −0.705929.
4 2(0.1)
The actual value is fx ( π4 ) = − sin( π4 ) = −0.707107 (correct to 6 decimal places) so the error
for the first-order forward approximation is 0.034148, the error for the first-order backward
approximation is 0.036504 and the error for the second-order approximation is 0.001178.

The second-order approximation of fxx (x) is (a first-order approximation does not exist)
f (x − h) − 2f (x) + f (x + h)
fxx (x) ≈ ,
h2
so we have
π  π  π 
π  cos − 0.1 − 2 cos + cos + 0.1
fxx ≈ 4 4 4 = −0.706518,
4 0.12
which when compared to the actual value of fxx ( π4 ) = − cos( π4 ) = −0.707107 (correct to 6
decimal places) gives an error of 0.000589.

Dr Jon Shiach 17 Finite-Difference Methods

Chapter 2. Finite-Difference Approximations

2.3.4 Estimating the order of an approximation

Given a known truncation error it is possible to estimate the order of an approximation. We have
seen earlier that for E(h) = O(hn )
E(h) ≈ Chn .
Taking logarithms of both sides gives

log |E(h)| ≈ n log(h) + log(C),

which is a linear function. Therefore we can approximate n in O(hn ) by calculating the gradient of
log |E(h)| for a series approximation for the different values of h, i.e., if hmax and hmin are the largest
and smallest values of h for which the series approximation has been calculated then

log |E(hmax )| − log |E(hmin )|

n≈ .
log(hmax ) − log(hmin )

For example consider the forward, backward and central difference approximations of the first deriva-
tive of f (x) = cos(x) at x = π4 using step lengths h = 0.1, 0.05, 0.25 and 0.0125 shown in table 2.1.

π
Table 2.1: Finite-difference approximations of the first derivative of f (x) = cos(x) at x = 4
using step lengths h = 0.1, 0.05, 0.25 and 0.0125.

h forward backward central

0.1000 −0.741255 −0.670603 −0.705929
0.0500 −0.724486 −0.689138 −0.706812
0.0250 −0.715872 −0.698195 −0.707033
0.0125 −0.711508 −0.702669 −0.707088

The absolute errors between the finite-difference approximations and the exact value of − sin( π4 ) =
−0.707107 are given in table 2.2.

Table 2.2: Truncation errors for the Taylor series approximations of cos(1 + h) for h = 0.5,
0.25, 0.125 and 0.0625.

h forward backward central

0.1000 0.034148 0.036504 0.001178
0.0500 0.017379 0.017969 0.000295
0.0250 0.008765 0.008912 0.000074
0.0125 0.004401 0.004438 0.000018

To estimate the order of the forward difference approximation we have

log(0.034148) − log(0.004401)
n≈ = 0.985305 ≈ 1,
log(0.1) − log(0.0125)

so we can say that the truncation error for the forward difference finite-difference approximation is
O(h). Doing similar for the backward and central difference approximations we get n ≈ 1.0134 and
n ≈ 1.9998 which suggests truncation errors of O(h) and O(h2 ) respectively.
A plot of the errors for the forward, backward and central difference approximations using a loglog
scale is shown in figure 2.4. By using a loglog scale we can easily compare the behaviour of the
errors as h decreases. Note that the forward and backward approximations have very similar errors

Finite-Difference Methods 18 Dr Jon Shiach

 Chapter 2. Finite-Difference Approximations

where as the central difference approximation is significantly more accurate. Furthermore the line
representing errors in the central difference approximation has a gradient of 2 and is steeper than
those of the forward and backward differences which have gradients of 1, this shows that as the step
length decreases, the central difference approximation will converge to the exact value at a faster rate
than the forward and backward difference approximations.

Figure 2.4: Plot of the error for the forward, backward and central difference approximations.

2.3.5 Finite-difference approximations of mixed derivatives

Finite-difference approximations of mixed derivatives where partial derivatives a function with respect
to two or more independent variables are combined can be derived by applying finite-difference ap-
proximations of the function with respect to a single independent variable. For example consider fxy
where x and y are independent variables of f , we can write this as
   
∂ 2f ∂ ∂f ∂ ∂f
fxy = = = .
∂x∂y ∂x ∂y ∂y ∂x

Let ∆x and ∆y be the step lengths for the x and y variables respectively and using a first-order
forwards approximation for fx (x, y) and fy (x, y)

f (x + ∆x, y) − f (x, y)
fx (x, y) = + O(∆x),
∆x
f (x, y + ∆y) − f (x, y)
fy (x, y) = + O(∆y),
2∆y

then

∂
fxy (x, y) = (fy (x, y))
∂x
fy (x + ∆x, y) − fy (x, y)
= + O(∆x)
∆x
f (x + ∆x, y + ∆y) − f (x + ∆x, y) f (x, y + ∆y) − f (x, y)
−
∆y ∆y
= + O(∆x, ∆y)
∆x
f (x + ∆x, y + ∆y) − f (x + ∆x, y) − f (x, y + ∆y) + f (x, y)
= + O(∆x, ∆y)
∆x∆y

Dr Jon Shiach 19 Finite-Difference Methods

Chapter 2. Finite-Difference Approximations

2.3.6 Deriving finite-difference formulae using the method of undetermined coef-

ficients
Perhaps the best method for deriving finite-difference approximations of higher order derivatives is
the method of undetermined coefficients. The approximation of a derivative is written as a sum of
f (x), f (x − h), f (x + h) etc. each multiplied by some coefficient. The value of these coefficients are
determined by replacing f (x + h), f (x − h) etc. with the corresponding Taylor series expansion and
then ensuring that the values coefficients match with the derivative that is being approximated.
For example, let us assume that we want a finite-difference approximation that uses the nodes f (x),
f (x + h) and f (x − h), i.e.,

f(n) (x) = c1 f (x − h) + c2 f (x) + c3 f (x + h), (2.11)

where ci are the undetermined coefficients. Substituting in the second-order forwards and backwards
Taylor series in place of f (x + h) and f (x − h) and factorising gives
   
h2 h2
f(n) (x) = c1 f (x) − hfx (x) + fxx (x) + c2 f (x) + c3 f (x) + hfx (x) + fxx (x) + O(h3 )
2 2
1
= (c1 + c2 + c3 )f (x) + (−c1 + c3 )hfx (x) + (c1 + c3 )h2 fxx (x) + O(h3 ).
2
If we require a finite-difference approximation of fx (x) then we need the right-hand side to equal fx (x)
so the coefficients of f (x) and fxx (x) should be zero and the coefficient of fx (x) should be 1. This
leads to the following linear system of equations

c1 + c2 + c3 = 0,
1
−c1 + c3 = ,
h
c1 + c3 = 0.

The solution to this linear system is c1 = − 2h 1

, c2 = 0 and c3 = 1
2h . Substituting back into equa-
tion (2.11) gives the following finite-difference approximation

1 1 O(h3 ) f (x + h) − f (x − h)
fx (x) = − f (x − h) + f (x + h) + = + O(h2 ),
2h 2h 2h 2h
which is the same as equation (2.9). Alternatively if we required a finite-difference approximation of
fxx (x) then the coefficients of f (x) and fx (x) need to be zero and the coefficient of fxx (x) should be
1, i.e.,

c1 + c2 + c3 = 0,
−c1 + c3 = 0,
2
c1 + c3 = 2 .
h

Solving gives c1 = 1
h2
, c2 = − h22 and c3 = 1
h2
so the second-order finite-difference approximation of
fxx (x) is

f (x − h) − 2f (x) + f (x + h)
fx (x) = + O(h2 ),
h2
which is the same as equation (2.10).
The method of undetermined coefficients also allows us to derive skewed or one-sided finite-difference
approximations which use more of the values of the neighbouring nodes to one side of f (x) than the
other. These can be useful for approximating derivatives close to the boundaries of the domain.

Finite-Difference Methods 20 Dr Jon Shiach

 Chapter 2. Finite-Difference Approximations

Example 4

Use the method of undetermined coefficients to derive a one-sided second-order finite-difference

approximation of fx (x).

Solution: A one-sided finite-difference approximation means that the other nodes used in the
approximation are only on one side of the f (x) node, e.g.,

f(n) (x) = c1 f (x) + c2 f (x + h) + c3 f (x + 2h).

Here we need the Taylor expansion of f (x + 2h) which is easily obtained by substituting 2h in
place of h in the second-order forward Taylor series

(2h)2
f (x + 2h) = f (x) + 2hfx (x) + fxx (x) + O(h3 )
2
= f (x) + 2hfx (x) + 2h2 fxx (x) + O(h3 ).

Substituting in the second-order Taylor series for f (x + h) and f (x + 2h) and factorising gives
 
h2  
f(n) (x) = c1 f (x) + c2 f (x) + hfx (x) + fxx (x) + c3 f (x) + 2hfx (x) + 2h2 fxx (x)
2
+ O(h3 )
 
1
= (c1 + c2 + c3 )f (x) + (c2 + 2c3 )hfx (x) + c2 + 2c3 h2 fxx (x) + O(h3 )
2

For an approximation of fx (x) we need

c1 + c2 + c3 = 0,
1
c2 + 2c3 = ,
h
1
c2 + 2c3 = 0,
2
which can be written in matrix form as
    
1 1 1 c1 0
0 1 2 c2  =  1  .
h
0 21 2 c3 0

The inverse of the coefficient matrix is

 
1 − 32 1
0 2 −2
0 − 12 1

so the solution to this linear system is

      3 
c1 1 − 32 1 0 − 2h
   1  2 
c2  = 0 2 −2  h  =  h  ,
c3 0 − 12 1 0 − 21 h

and the one-sided finite-difference approximation of fx (x) is

3 2 1 O(h3 )
fx (x) = − f (x) + f (x + h) − f (x + 2h) +
2h h 2h h
−3f (x) + 4f (x + h) − f (x + 2h)
= + O(h2 ).
2h

Dr Jon Shiach 21 Finite-Difference Methods

Chapter 2. Finite-Difference Approximations

2.4 The finite-difference toolkit

Finite-difference methods solve PDEs by replacing the partial derivatives with finite-difference ap-
proximations. The type and order of the finite-difference approximations used will depend upon the
accuracy required and the computational resources available. We can build what is known as the
finite-difference toolkit of common derivatives for convenience (see table 2.4) and we can pick and
choose the approximations in order to suit our needs. Sometimes a non-symmetric finite-difference
approximation is required, this usually happens near to the boundaries where the number of nodes each
side of f (xi ) may not be the same. In these cases we need to derive a finite-difference approximation
using the method of undetermined coefficients (section 2.3.6).

We will consider PDEs written in terms of the function f (t, x) where t usually denotes time and x
denotes spatial positioning. If one of the independent variables is fixed, the partial derivative of f (t, x)
can be approximated using finite-difference approximations. For example, consider the approximation
of fx (t, x) using the second-order central difference approximation in x with a step length of h = ∆x
results in

∂ f (t, x + ∆x) − f (t, x − ∆x)

f (t, x) = ,
∂x 2∆x

or alternatively, ft (t, x) with step length h = ∆t is

∂ f (t + ∆t, x) − f (t − ∆t, x)
f (t, x) = .
∂t 2∆t

For convenience, a subscript/superscript notation is employed. Let the subscript i denote the spatial
positioning and the superscript n denote the time level, i.e.,

fin = f (tn , xi )

then we can write

∂ f n − fi−1
n
f (tn , xi ) = i+1 ,
∂x 2∆x
∂ f n+1 − fin−1
f (tn , xi ) = i .
∂t 2∆t

Finite-Difference Methods 22 Dr Jon Shiach

 Chapter 2. Finite-Difference Approximations

Finite-Difference Toolkit

Order
Derivative Finite-difference approximation
of accuracy

fi+1 − fi fi − fi−1
1st (forward) or (backward)
h h
fx fi+1 − fi−1
2nd (central)
2h
fi−2 − 8fi−1 + 8fi+1 − fi+2
4th
12h
fi−1 − 2fi + fi+1
2nd (symmetric)
fxx h2
−fi−2 + 16fi−1 − 30fi + 16fi+1 − fi+2
4th
12h2
−fi−2 + 2fi−1 − 2fi+1 + fi+2
2nd
fxxx 2h3
fi−3 − 8fi−2 + 13fi−1 − 13fi+1 + 8fi+2 − fi+3
4th
8h3
fi−2 − 4fi−1 + 6fi − 4fi+1 + fi+2
2nd
fxxxx h4
−fi−3 + 12fi−2 − 39fi−1 + 56fi − 39fi+1 + 12fi+2 − fi+3
4th
6h4

2.5 Example of a finite-difference scheme

Now that we can approximate partial derivatives using our finite-difference toolkit the solution of a
simple PDE is going to be used as an example. The PDE in question is known as the advection
equation that models pure advection (the transport of a substance in space) is defined by

Ut + vUx = 0. (2.12)

The independent variables are t that denotes time and x that denotes space. The function U = U (t, x)
is the concentration of some substance and v is the velocity that the substance travels along the domain
which is defined by a ≤ x ≤ b. In order to solve equation (2.12), we need the initial conditions that
define the values of U at the start of the solution (t = 0), i.e.,

U (0, x) = f (x). (2.13)

The solution of the advection equation is the function U (t, x) which satisfies:

• equation (2.12) for all values of a ≤ x ≤ b at time t;

• the initial conditions in equation (2.13)

• any conditions that are applied at the boundaries x = a and x = b.

An exact solution is defined at an infinite number of values for the independent variables t and x. We
will construct a Finite-Difference Scheme (FDS) to approximate U at a finite set of discrete points in
t and x.

Dr Jon Shiach 23 Finite-Difference Methods

Chapter 2. Finite-Difference Approximations

2.5.1 Spatial discretisation

The FDS solves a PDE at discrete points along the domain called finite-difference nodes. For simplicity
it is common to locate the finite-difference nodes at equally spaced intervals across the domain creating
what is known as a uniform grid. This process is called discretising the domain.
Consider the diagram in figure 2.5. Here, the x domain is the range a ≤ x ≤ b and is discretised using
a uniform grid of N nodes. The node at the left-hand boundary is given the label x0 and we increment
the subscript for each successive node until we get to node at the right-hand boundary that is given
the label xN −1 . The first node in the finite-difference grid is located at x0 = a, the second node is
located at x1 = a + ∆x where ∆x is the spatial step length between nodes, therefore the ith node in
the grid is located at xi = a + i∆x.

a a + ∆x a + 2∆x a + (N − 2)∆x b
x
x0 x1 x2 ··· xN xN −1
∆x

Figure 2.5: The x domain discretised using a uniform grid of finite-difference nodes

Using a uniform finite-difference grid means that the step length ∆x and the number of nodes N are
related. Since we require xN +1 = b then

a + (N − 1)∆x = b

which gives

b−a
∆x = . (2.14)
N −1

If however the spatial step length ∆x is defined then it is easy to see that
b−a
N =1+ .
∆x

2.5.2 Deriving a finite-difference scheme

The next step is to derive a finite-difference scheme for the PDE. We do this by replacing the partial
derivatives in the PDE with appropriate finite-difference approximations from our toolkit. The choice
of a different finite-difference approximations will result in a different FDS and these depend on the
order accuracy required. For example, if we want a first-order FDS then we can replace Ut and Ux in
equation (2.12) with a forward and backward differences respectively, i.e.,

Ut + vUx = 0
Uin+1 − Uin Uin − Ui−1
n
+v = 0.
∆t ∆x
We now rearrange this equation so that Uin+1 is the subject

v∆t n
Uin+1 = Uin − (U − Ui−1
n
). (2.15)
∆x i
This is a FDS to solve the advection equation equation (2.12). Since the finite-difference approxima-
tions for derivatives in space and time are O(∆x) and O(∆t) respectively we can say that this FDS
has formal accuracy of first-order in space and time. A FDS of the form of equation (2.15) is called a
time marching scheme since the solution at the next time level, n + 1, is calculated using values from

Finite-Difference Methods 24 Dr Jon Shiach

 Chapter 2. Finite-Difference Approximations

the current time level, n. Since U (t, x) is only known exactly at the initial time level, equation (2.15)
is rewritten replacing Uin with uni to reflect that it is an approximation of the exact solution

v∆t n
un+1 = uni − (u − uni−1 ) + O(∆t, ∆x). (2.16)
i
∆x i

It is of interest to know which nodes are required to calculate the value of un+1
i for a given FDS.
In equation (2.16) the calculation of the value of the node un+1
i is dependent upon two nodes at the
current time level uni and the neighbouring node to the left uni−1 . This information can be represented
in a diagram known as a finite-difference stencil shown in figure 2.6.

un+1
i
n+1

∆t
uni−1 uni
n

∆x
x
i−1 i i+1

Figure 2.6: The stencil the first-order finite-difference scheme, equation (2.16), used to solve
the advection equation.

2.5.3 Computational and boundary nodes

Looking at the finite-difference stencil in figure 2.6 for the FDS we will encounter a problem when
trying to evaluate un+1
0 since it requires the value of the node un−1 which doesn’t exist. To counter this
problem we only apply the FDS to the computational nodes that can be evaluated, i.e., un1 , · · · , unN −2 .
The boundary node, un0 , is removed from the calculation of the FDS and its value is determined using
some boundary condition that is implemented at the boundary (boundary conditions are discussed in
section 3.2 on page 32). The FDS used here only encounters a problem at the left-hand boundary,
however if we were to use a forward or central discretisation Ux then we would also encounter this
problem at the right-hand boundary. Although it may not always be necessary, it is common practice
to use boundary nodes at either end of the spatial domain (figure 2.7). For this example we will
assume the boundary conditions always have a value of zero, i.e.,

U (t, a) = U (t, b) = 0.

boundary boundary

computational nodes

u0 u1 u2 u3 ··· uN −2 uN −1 x
boundary boundary
node node

Figure 2.7: The boundary nodes are removed from the calculation of the FDS.

Dr Jon Shiach 25 Finite-Difference Methods

Chapter 2. Finite-Difference Approximations

2.5.4 Hand calculation

To demonstrate how the FDS in equation (2.16) is used to calculate an approximation of the solution
to the PDE in equation (2.12) we are going to calculate the values at the finite-difference nodes by
hand∗ . This is a time consuming and labourious exercise and is not recommended practice for solving
PDEs numerically. It is done here to help understand the process and can be useful for verifying and
debugging programs.
For this example we are going to use a spatial domain of 0 ≤ x ≤ 1 which will be discretised using 11
nodes. The velocity is set at v = 1 and the time step is ∆t = 0.05. First we calculate the spatial step
using equation (2.14)
1−0
∆x = = 0.1.
11 − 1
Using xi = a + i∆x then the co-ordinates of the finite-difference nodes are
x = (0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0),
i.e.,
u0 u1 u2 u3 u4 u5 u6 u7 u8 u9 u10

0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 x

The initial conditions are given by the following exponential function

U (0, x) = exp[−100(x − 0.4)2 ], (2.17)
and the boundary conditions are
U (t, 0) = 0,
U (t, 1) = 0.
Calculating the initial values of the first few nodes gives
u00 = 0, (boundary node),
u01 = exp(−100(0.1 − 0.4)2 ) = 0.0001,
u02 = exp(−100(0.2 − 0.4)2 ) = 0.0183,
u03 = exp(−100(0.3 − 0.4)2 ) = 0.3679,
..
.

For this FDS, ∆t, ∆x and v are constant so we can pre-calculate v∆t ∆x in order to save time later on. It
is good practice to pre-calculate any values that you know will not change over the course of applying
a numerical scheme
v∆t 1 × 0.05
= = 0.5,
∆x 0.1
so the FDS is
un+1
i = uni − 0.5(uni − uni−1 ).
Calculating the solution at the first time step for the first few nodes
u11 = u01 − 0.5(u01 − u00 ) = 0.0001 − 0.5(0.0001 − 0) = 0.0001,
u12 = u02 − 0.5(u02 − u01 ) = 0.0183 − 0.5(0.0183 − 0.0001) = 0.0092,
u13 = u03 − 0.5(u03 − u02 ) = 0.3679 − 0.5(0.3679 − 0.0183) = 0.1931,

The values of the first three iterations of the FDS are given in table 2.3.
∗
When we say we are going to perform a hand calculation it means that we are going to perform each calculation
using pen and paper and the help of a calculator.

Finite-Difference Methods 26 Dr Jon Shiach

Dr Jon Shiach
Table 2.3: Values of the first three iterations of the FDS used to solve the advection equation

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

xi 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0

u0i 0.0000 0.0001 0.0183 0.3679 1.0000 0.3679 0.0183 0.0001 0.0000 0.0000 0.0000

27
u1i 0.0000 0.0001 0.0092 0.1931 0.6839 0.6839 0.1931 0.0092 0.0001 0.0000 0.0000

u2i 0.0000 0.0000 0.0046 0.1012 0.4385 0.6839 0.4385 0.1012 0.0046 0.0000 0.0000

u3i 0.0000 0.0000 0.0023 0.0529 0.2698 0.5612 0.5612 0.2698 0.0529 0.0023 0.0000

Finite-Difference Methods
Chapter 2. Finite-Difference Approximations
Chapter 2. Finite-Difference Approximations

2.5.5 Solution procedure

The solution procedure for solving the advection equation FDS given in equation (2.15) is summarised
by the flow chart shown in figure 2.8. Flow charts are useful for understanding the procedures and
processes involved when using numerical methods. Here we can see that following the initialisation
of the variables and arrays, a loop is required to iterate the FDS through the number of time steps.
Within this loop the FDS is used to calculate un+1 (this requires another loop to loop through all of
the computational nodes to calculated the updated solution), un and t are updated ready for the next
time step and the current solution is outputted.

start

initialise variables

calculate node co-ordinates

x and initial conditions u0

set step counter n to 0

yes
n > nsteps end

no
calculate one step of the FDS
for each computational node
and store the values in un+1
i

update solution array, time

variable and step counter
un ← un+1
t ← t + ∆t
n←n+1

output un

Figure 2.8: Flow chart for the FDS used to solve the advection equation.

2.5.6 MATLAB code

It should be obvious to the reader that calculating the solution to a PDE using a FDS is not usually
done by hand. There are far too many calculations required which means there are more opportunities
to make mistakes and it soon becomes quite tedious. Performing the calculations using a computer
program saves time and effort and changes made to the problem can be implemented quickly and with
the minimum of fuss. We can use the flow chart in 2.8 to help in writing a MATLAB program to
perform the calculations. The MATLAB code shown in listing 2.1 was used to calculate the values in
table 2.3.
Listing 2.1: MATLAB program to calculate the solution of the advection equation using the
first-order FDS in equation (2.16).

Finite-Difference Methods 28 Dr Jon Shiach

 Chapter 2. Finite-Difference Approximations

1 % fds_example . m by Jon Shiach

2 %
3 % This program solves the linear advection equation using a first - order FDS
4
5 % Clear workspaces ( should always be done at the start of an m - file )
6 clear
7 clc
8
9 % Define variables
10 N = 11; % number of nodes
11 xmin = 0; % min value of x
12 xmax = 1; % max value of x
13 dt = 0.05; % time step
14 v = 1; % velocity
15 t = 0; % initial value of t
16
17 % Calculate node positions
18 dx = ( xmax - xmin ) / ( N - 1) ;
19 x = xmin : dx : xmax ;
20
21 % Calculate initial conditions
22 u = exp ( -100 * ( x - 0.4) .^2) ;
23
24 % Output column headings and initial conditions to the command window
25 hline = repmat ( ' - ' , 1 , 9 * ( N + 1) ) ;
26 fprintf ( '\ n % s ' , hline )
27 fprintf ( '\ n t ')
28 fprintf ( '| u_ %1 i ' , [0: N -1])
29 fprintf ( '\ n % s ' , hline )
30 fprintf ( '\ n %6.2 f ' , t )
31 fprintf ( ' |%7.4 f ' , u )
32
33 % pre - calculate C = v * dt / dx
34 C = v * dt / dx ;
35
36 % Perform 3 iterations of the FDS
37 for n = 1 : 3
38
39 % Calculate boundary conditions
40 u (1) = 0;
41 u ( N ) = 0;
42
43 % Calculate new values of u
44 unew = u ;
45 for i = 2 : N - 1
46 unew ( i ) = u ( i ) - C * ( u ( i ) - u (i -1) ) ;
47 end
48
49 % Update t and u
50 t = t + dt ;
51 u = unew ;
52
53 % Output current solution
54 fprintf ( '\ n %6.2 f ' ,t )
55 fprintf ( ' |%7.4 f ' ,u )
56 end
57 fprintf ( '\ n % s \ n \ n ' , hline )

The MATLAB program shown in listing 2.1 is explained below:

Dr Jon Shiach 29 Finite-Difference Methods

Chapter 2. Finite-Difference Approximations

• Lines 6 and 7 – since this is an m-file it has access to any previously defined variables or arrays
which might cause errors in the program, therefore it is always a good idea to clear the memory
using clear and the command window using clc. Note that this step isn’t necessary if you have
used a function instead of an m-file.
• Lines 10 to 15 – defines all of the variables used in the program at once in the beginning. It is
good practice to use comments to explain what each variable represents.
• Lines 18 and 19 – calculates the spatial step and uses an implicit for loop to generate an array
of x values going from xmin to xmax in steps of dx
• Line 22 – calculates the initial values of u0i using equation (2.17) and stores them in the array u.
• Lines 25 to 31 – outputs the column headings using formatted output. This is purely cosmetic
and not really necessary of the successful calculation of the solution, however sometimes it is
useful to be able to output the solution in a readable format.
• Line 34 – The value of v∆t
∆x is pre-calculated since it is constant throughout. This should be done
for all values that do not change over the iterations since computational effort is not wasted
calculating the values when they are already known.
• Line 37 – a for loop is used to calculate 3 iterations of the FDS. The variable n is the iteration
counter.
• Lines 40 and 41 – the values of the boundary nodes are calculated using the boundary conditions
for the current problem.
• Lines 44 to 47 – a for loop is used to loop through the nodes and calculate the value of un+1
i
using the FDS in equation (2.16). Since array indices in MATLAB begin at 1 rather that 0,
the first computational node u1 is the second node in u therefore we add 1 to the array indices,
i.e., i = 2 : N - 1 in the code represents the computational nodes i = 1, · · · , N − 2 in the
finite-difference discretisation.
• Lines 50 and 51 – u is updated by setting it equal to unew ready for the next iteration. t is also
updated by incrementing by dt, although not used in this program it is useful to keep track of
the t value for output purposes.

Finite-Difference Methods 30 Dr Jon Shiach

Chapter 3

Elliptic Partial Differential Equations

Elliptic PDEs are used to model problems where the independent variables are restricted to spatial co-
ordinates only. Since time is not a variable the solution of an elliptic PDE is constant and referred to
as steady state solutions. Examples of elliptic equations are Laplace’s equation and Poisson’s equation
that are used to model electrical and gravitational fields. Laplace’s equation is the simplest elliptic
PDE and was derived by French mathematician Pierre-Simon Laplace (1749–1827). The class of
analytical solutions to Laplace’s equation are called harmonic functions which have applications in
many fields of science such as magnetism, astronomy and fluid dynamics. Laplace’s equation is
Laplace’s equation

Uxx + Uyy = 0, (3.1)

where x and y are the spatial co-ordinates. Poisson’s equation named after French mathematician
Siméon Denis Poisson (1781–1840) is an extension of Laplace’s equation

Poisson’s equation

Uxx + Uyy = f (x, y), (3.2)

where f (x, y) is some function of x and y. This chapter will focus on the solution of Laplace’s equation
using finite-difference methods and introduces the reader to the discretisation of a two-dimensional
spatial domain, calculation of boundary conditions and solving linear systems.

3.1 Discretising a two-dimensional domain

Since Laplace’s equation is written in terms of x and y we need to discretise both spatial dimensions
to produce a two-dimensional grid. If the spatial domain is defined by xmin ≤ x ≤ xmax and ymin ≤
y ≤ ymax . NX and NY nodes are used to discretise the domain in the x and y directions respectively
(the subscripts X and Y have been written in uppercase to avoid confusing it with first-order partial
derivatives), then the spatial steps are
xmax − xmin
∆x = ,
NX − 1
ymax − ymin
∆y = .
NY − 1
The position of the nodes in the finite-difference grid are the co-ordinates (xj , yi ) where
xi = xmin + i∆x,
yj = ymin + j∆y.

31
Chapter 3. Elliptic Partial Differential Equations

To identify individual nodes in a two-dimensional finite-difference grid a subscript notation is employed

where the index i denotes the position of the node in the x direction and j denotes the position of the
node in the y direction, i.e., ui,j (figure 3.1)∗ .

y
ui−1,j+1 ui,j+1 ui+1,j+1
j+1

ui−1,j ui,j ui+1,j

j

ui−1,j−1 ui,j−1 ui+1,j−1

j−1
x
i−1 i i+1

Figure 3.1: Node numbering convention for a two-dimensional finite-difference grid.

3.2 Boundary conditions

We saw in section 2.5.3 that it may not be possible to calculate the values of the nodes on the
boundaries using a particular FDS. When this is the case, these boundary nodes are omitted from
the calculation of the FDS and assume some value that depends on the behaviour of the system at
the boundary. Determining the value of the boundary nodes is done by invoking boundary condi-
tions. These notes will discuss two types of boundary conditions: Dirichlet and Neumann boundary
conditions.

3.2.1 Dirichlet boundary conditions

Dirichlet boundary conditions are named after German mathematician Peter Gustav Lejeune Dirichlet
(1805–1859) and specify the value of U at the boundary of the solution domain. This is useful in cases
where the points on the boundary are not influenced by the behaviour of the points within the domain.
Dirichlet boundary conditions are applied to the boundary nodes that are positioned on the boundaries,
i.e., u0 and uN −1 , and the FDS are applied to the computational nodes u1 to uN −2 (figure 3.2).

boundary boundary

computational nodes

u0 u1 u2 u3 ... uN −2 uN −1 x
boundary boundary
node node

Figure 3.2: When using Dirichlet boundary conditions the nodes on the boundaries are
excluded from the computation and determined by some known scalar or function.

The are three main uses for Dirichlet boundary conditions:

∗
Be careful to avoid confusing the node indexing convention used for finite-differences grids with the indexing used
for matrices.

Finite-Difference Methods 32 Dr Jon Shiach

 Chapter 3. Elliptic Partial Differential Equations

• Values on the boundaries are constant, e.g.,

u0 = α,
uN −1 = β,

where u0 and uN denotes the boundary nodes and α and β are constants. This is used in
cases where the solution on the boundary is known and constant. For example, modelling heat
diffusion across a domain where the boundary temperature is always zero.
• Values on the boundaries adhere to some function, e.g.,

u0 = f (t, x, y),
uN −1 = g(t, x, y),

where f (t, x, y) and g(t, x, y) are some functions of the independent variables. These are used
in cases where the solution at the boundary is not constant but known to behave in a certain
way. For example, modelling wave propagation where the boundary condition is used to generate
waves passing into the domain.
• Boundaries are used to model periodic behaviour, e.g.„

u0 = uN −2 ,
uN −1 = u1 ,

where u1 and uN −2 are the values of the first and last computational nodes (figure 3.3). This
is used where we want to model an infinite domain. For example, modelling wave propagation
where waves passing out through the right-hand boundary will pass in from the left boundary.

x
u0 u1 u2 u3 ... uN −2 uN −1

Figure 3.3: Periodic boundary conditions

3.2.2 Neumann boundary conditions

Neumann boundary conditions are named after German mathematician Karl Gottfried Neumann
(1832–1925) and specifies the derivative of the solution across the boundary. Boundary conditions
of this type are used where the points on the boundaries are influenced by the behaviour in the do-
main. When using Neumann boundary conditions, additional nodes called ghost nodes are added to
each end of the domain so that we can calculated finite-difference approximations for the nodes on the
boundary (u−1 and uN +1 in figure 3.4) and the FDS is applied to the computational nodes u0 to uN .

boundary boundary
ghost node ghost node

x
u−1 u0 u1 u2 ... uN −2 uN −1 uN

computational nodes

Figure 3.4: Neumann boundary conditions use the derivative across the boundary to deter-
mine the values of the ghost nodes.

Dr Jon Shiach 33 Finite-Difference Methods

Chapter 3. Elliptic Partial Differential Equations

To specify a Neumann boundary condition we need to approximate the derivative across the boundary
using a finite-difference. Consider the Neumann boundary condition

Ux = f (U ),

where f (U ) is a function of U . Approximating Ux using a central difference we have

u1 − u−1
= f (U )
2∆x
u−1 = u1 − 2∆xf (U ),

so the value of the ghost node u−1 is an expression involving the second computational node u1 and
f (U ). There are two main uses for Neumann boundary conditions:
• Transmissive boundary conditions: the derivative across the boundary is zero, i.e., f (U ) = 0 so

u−1 = u1 ,
uN = uN −2 .

This has the effect that any flow will simply pass out of the boundary as if it wasn’t there. For
example, consider the modelling of water waves down a channel, by specifying a transmissive
boundary at one end the waves will simply pass through as if the channel had an infinite length.
• Reflective (solid) boundary conditions: no flow through a reflective boundary. This used where
one of the dependent variables of the PDE is velocity and the velocity of the ghost node is set
to equal that of the velocity of a computational node with the sign reversed, e.g.,

v−1 = −v1 ,
vN = −vN −2 ,

where vi is the velocity at node i.

3.3 Deriving a FDS to solve Laplace’s equation

We will now proceed to derive a FDS to solve Laplace’s equation. In equation (3.1) there are two
second-order partial derivatives that need approximating. Looking at the finite-difference toolkit
(page 23) we can use the second-order symmetric finite-difference approximations
ui−1,j − 2ui,j + ui+1,j
Uxx ≈ ,
∆x2
ui,j−1 − 2ui,j + ui,j+1
Uyy ≈ .
∆y 2
The FDS is derived by replacing the derivatives in equation (3.1) by the finite-difference approxima-
tions

Uxx + Uyy = 0
ui−1,j − 2ui,j + ui+1,j ui,j−1 − 2ui,j + ui,j+1
+ = 0.
∆x2 ∆y 2
We want an expression that gives the solution for the node ui,j , therefore rearranging to make ui,j the
subject results in
Second-order finite-difference scheme to solve Laplace’s equation

∆y 2 (ui−1,j + ui+1,j ) + ∆x2 (ui,j−1 + ui,j+1 )

ui,j = . (3.3)
2(∆x2 + ∆y 2 )

Finite-Difference Methods 34 Dr Jon Shiach

 Chapter 3. Elliptic Partial Differential Equations

This is a second-order accurate FDS that approximates the solution to Laplace’s equation. The stencil
for this FDS is shown in figure 3.5.

ui,j+1

ui,j
ui−1,j ui+1,j

ui,j−1

Figure 3.5: The finite-difference stencil for the second-order FDS used to solve Laplace’s
equation

3.4 Solving Laplace’s equation

In order to demonstrate the solution of Laplace’s equation using finite-differences we are going to use
a simple example with a coarse grid shown in figure 3.6. The domain is 0 ≤ x ≤ 4 and 0 ≤ y ≤ 3
has been discretised using 5 nodes in the x direction and 4 nodes in the y direction so that ∆x =
∆y = 1. Dirichlet boundary conditions are implemented at all four boundaries so that the values of
the boundary nodes are constant. We need to calculate the solution to Laplace’s equation using the
FDS in equation (3.3) for the six computational nodes u1,1 , u2,1 , u3,1 , u1,2 , u2,2 and u3,2 .

8.9 8.9 8.9 8.9 8.9

8.4 u1,2 u2,2 u3,2 9.2

7.2 u1,1 u2,1 u3,1 9.4

6.1 6.8 7.7 8.7 6.1 x

Figure 3.6: Finite-difference grid used to solve Laplace’s equation.

Using equation (3.3) and considering the computational nodes going from left-to-right along each row

Dr Jon Shiach 35 Finite-Difference Methods

Chapter 3. Elliptic Partial Differential Equations

and up the rows from top-to-bottom we have

7.2 + u3,1 + 6.8 + u2,1

u1,1 = ,
4
u1,1 + u3,1 + 7.7 + u2,2
u2,1 = ,
4
u2,1 + 9.4 + 8.7 + u3,2
u3,1 = ,
4
8.4 + u2,2 + u1,1 + 8.9
u1,2 = ,
4
u1,2 + u3,2 + u2,1 + 8.9
u2,2 = ,
4
u2,2 + 9.2 + u3,1 + 8.9
u3,2 = .
4
These equations can be written as the linear system by moving all of the unknown values to the
left-hand side and the known values to the right-hand side

4u1,1 − u2,1 − u1,2 = 14,

−u1,1 + 4u2,1 − u3,1 − u2,2 = 7.7,
−u2,1 + 4u3,1 − u3,2 = 18.1,
−u1,1 + 4u1,2 − u2,2 = 17.3,
−u2,1 − u1,2 + 4u2,2 − u3,2 = 8.9,
−u3,1 − u2,2 + 4u3,2 = 18.1,

or alternatively using matrix notation of the form Au = d such that

    
4 −1 0 −1 0 0 u1,1 14
−1 4 −1 0 −1 0  u2,1   7.7 
    
 0 −1 4 0 0 −1    
  u3,1  = 18.1 .
−1 0 0   
4 −1 0  u1,2  17.3

 
 0 −1 0 −1 4 −1 u2,2   8.9 
0 0 −1 0 −1 4 u3,2 18.1

This linear system can be solved using direct methods such matrix inversion or Gaussian elimination
but these are inefficient for large systems. Instead we can use indirect methods which use an iterative
process where an initial estimate of the solution at the nodes u is improved to form a better estimate.
This estimate in turn is improved and the process continues until two successive estimates agree to a
required accuracy.

3.5 The Jacobi method

The simplest indirect method is called the Jacobi method after German mathematician Carl Gustav
Jacob Jacobi (1804 – 1851). If uki,j denotes the current estimate of Ui,j and uk+1
i,j denotes the improved
estimate then using equation (3.3) we have

Jacobi method for computing the solution to Laplace’s equation

∆y 2 (uki−1,j + uki+1,j ) + ∆x2 (uki,j−1 + uki,j+1 )

uk+1
i,j = (3.4)
2(∆x2 + ∆y 2 )

Note that all of the values on the right-hand side of equation (3.4) have superscript k meaning that
they are all known values from the current iteration.

Finite-Difference Methods 36 Dr Jon Shiach

 Chapter 3. Elliptic Partial Differential Equations

Example 5

Perform two iterations of the Jacobi method to solve Laplace’s equation for the finite-difference
grid shown in figure 3.7 with zero starting values for the computational nodes.

8.9 8.9 8.9 8.9 8.9

8.4 0.0 0.0 0.0 9.2

7.2 0.0 0.0 0.0 9.4

6.1 6.8 7.7 8.7 6.1 x

Figure 3.7: The values of the initial conditions and boundary nodes for solving Laplace’s
equation.

Solution: The values of the computational nodes after one iteration of the Jacobi method are

7.2 + u02,1 + 6.8 + u01,2 7.2 + 0 + 6.8 + 0

u11,1 = = = 3.5,
4 4
u01,1 + u03,1 + 7.7 + u02,2 0 + 0 + 7.7 + 0
u12,1 = = = 1.925,
4 4
u02,1 + 9.4 + 8.7 + u02,3 0 + 9.4 + 8.7 + 0
u13,1 = = = 4.525,
4 4
8.4 + u02,2 + u01,1 + 8.9 8.4 + 0 + 0 + 8.9
u11,2 = = = 4.325,
4 4
u01,2 + u03,2 + u02,1 + 8.9 0 + 0 + 0 + 8.9
u12,2 = = = 2.225,
4 4
u02,2 + 9.2 + u03,1 + 8.9 0 + 9.2 + 0 + 8.9
u13,2 = = = 4.525.
4 4
The second iteration of the Jacobi method uses the values from the first iteration
7.2 + u12,1 + 6.8 + u11,2 7.2 + 1.925 + 6.8 + 4.325
u21,1 = = = 5.0625,
4 4
u11,1 + u13,1 + 7.7 + u12,2 3.5 + 4.525 + 7.7 + 2.225
u22,1 = = = 4.4875,
4 4
u12,1 + 9.4 + 8.7 + u12,3 1.925 + 9.4 + 8.7 + 4.525
u23,1 = = = 6.1375,
4 4
8.4 + u12,2 + u11,1 + 8.9 8.4 + 2.225 + 3.5 + 8.9
u21,2 = = = 5.7562,
4 4
u11,2 + u13,2 + u12,1 + 8.9 4.325 + 4.525 + 1.925 + 8.9
u22,2 = = = 4.9188,
4 4
u12,2 + 9.2 + u13,1 + 8.9 2.225 + 9.2 + 4.525 + 8.9
u23,2 = = = 6.2125.
4 4

Figures 3.8 and 3.9 contain the values of the computational nodes after the first and second iteration

Dr Jon Shiach 37 Finite-Difference Methods

Chapter 3. Elliptic Partial Differential Equations

of the Jacobi method respectively.

8.9 8.9 8.9 8.9 8.9

8.4 4.325 2.225 4.525 9.2

7.2 3.500 1.925 4.525 9.4

6.1 6.8 7.7 8.7 6.1 x

Figure 3.8: The values of the nodes for the first iteration of the Jacobi method.

8.9 8.9 8.9 8.9 8.9

8.4 5.7562 4.9188 6.2125 9.2

7.2 5.0625 4.4875 6.1375 9.4

6.1 6.8 7.7 8.7 6.1 x

Figure 3.9: The values of the nodes for the second iteration of the Jacobi method.

3.6 Convergence criteria

Iterations for an indirect method continue until we are satisfied that the current estimate is accurate
enough for our needs. This is done by calculating an error for the current estimate, err, and comparing
it to a predetermined tolerance denoted by tol. The tolerance is some small number that we choose
depending on the accuracy required. The smaller the value of tol, the more accurate the solution will
be but will be will require more iterations† . If err < tol then it is assumed that the last estimate has
converged to the exact solution. The problem is since we do not know the exact solution we cannot
calculate an exact value of err. However, assuming the sequence u0 , u1 , u2 , . . . converges to the exact
solution, then |uk+1
i − uki | → 0 as k → ∞ so |uk+1i − uki | can be used as a measure of the accuracy of
uk+1
i .
Another issue is that the solution to a PDE is approximated at multiple points in the domain so the
calculation of the error must take into account the solution at all of these points. To do this we can
use the L2 norm of the solution vector u which is defined as
†
Note that this refers to the accuracy of the solution of the linear system resulting from the FDS. The accuracy of
the FDS used to solve the PDE is dependent upon the order of the finite-difference approximations used to derive it.
Just because the error between two estimates is small does not mean that the solution obtained is accurate for the PDE

Finite-Difference Methods 38 Dr Jon Shiach

 Chapter 3. Elliptic Partial Differential Equations

Definition 4: The L2 norm

sX
∥u∥ = u2i . (3.5)
i

This provides a single value that can be used for the error at each iteration. There are two main ways
of calculating an error using the L2 norm: the absolute error and the relative error. The absolute error
is a measure of distance between two estimates of the solution

Definition 5: Absolute error

errabs = ∥uk+1 − uk ∥. (3.6)

The relative error is the measure of the distance between two sets of solutions relative to the most
recent estimation

Definition 6: Relative error

∥uk+1 − uk ∥
errrel = . (3.7)
∥uk+1 ∥

The choice of whether to use the absolute or relative error will depend on the size of the numbers you
are dealing with. For example, consider err calculated using the values u0 = 0.01 and u1 = 0.001,

errabs = |0.001 − 0.01| = 0.009,

|0.001 − 0.01|
errrel = = 9.
|0.001|

So the absolute error between these values is small suggesting convergence but the relative error is
large. Doing the same of the values u0 = 100000 and u1 = 100100

errabs = |100100 − 100000| = 100,

|100100 − 100000|
errrel = = 9.99 × 10−4 .
|100100|

In this case the absolute error is large whereas the relative error is small.

3.7 Solving a linear system using an indirect method

The general process of calculating the solution to a linear system using an indirect method is sum-
marised in the flow chart shown in figure 3.10. Once the variables and arrays have been initialised a
while loop is required to test for convergence. If the solution hasn’t converged yet, the new solution
uk+1 is calculated, then the error between the current solution and new solution and then the current
solution is updated by setting it equal to the new solution.

Dr Jon Shiach 39 Finite-Difference Methods

 Chapter 3. Elliptic Partial Differential Equations

start

initialise variables

calculate ∆x,
∆y, xi , yj and u0

set err ← 1 and k ← 0

no
err > tol end

yes

calculate uk+1
i,j for each
computational node

calculate err us-

ing uk+1 and uk

update uk ← uk+1
and k ← k + 1

Figure 3.10: Flow chart for solving an elliptic PDE using an indirect method.

The MATLAB code for calculating the Jacobi iterations for the Laplace equation example is given in
Listing 3.1.

Listing 3.1: MATLAB program to calculate the Jacobi iterations for the Laplace equation
example.
1 % laplace . m by Jon Shiach
2 %
3 % This program calculates the solution to the Laplace equation using the
4 % Jacobi method
5
6 % define variables
7 Nx = 5; % no . of nodes in the x direction ( inc ghost nodes )
8 Ny = 4; % no . of nodes in the y direction ( inc ghost nodes )
9 lenx = 4; % length of the x domain
10 leny = 3; % length of the y domain
11 err = 1; % error between successive iterations
12 tol = 1d -6; % convergence tolerance
13 k = 0; % iteration counter
14
15 % Calculate x and y node co - ordinates
16 dx = lenx / ( Nx - 1) ;
17 dy = leny / ( Ny - 1) ;
18 x = 0 : dx : lenx ;
19 y = 0 : dy : leny ;
20 [x , y ] = meshgrid (x , y ) ;
21

Finite-Difference Methods 40 Dr Jon Shiach

 Chapter 3. Elliptic Partial Differential Equations

22 % Initialise solution array

23 u = [ 8.9 , 8.9 , 8.9 , 8.9 , 8.9 ;
24 8.4 , 0.0 , 0.0 , 0.0 , 9.2 ;
25 7.2 , 0.0 , 0.0 , 0.0 , 9.4 ;
26 6.1 , 6.8 , 7.7 , 8.7 , 9.8 ];
27 u = flipud ( u ) ; % invert u to be consistent with matrix indexing
28

29 % output column headings and initial values

30 hline = repmat ( ' - ' , 1 , 12 * (( Nx - 2) * ( Ny - 2) + 2) ) ;
31 fprintf ( '% s \ n k | ' , hline )
32 for j = 1 : Ny - 2
33 for i = 1 : Nx - 2
34 fprintf ( ' u (%1 i ,%1 i ) | ' , i , j )
35 end
36 end
37 fprintf ( ' L2 error \ n % s \ n %4 i ' , hline , k ) ;
38 fprintf ( ' |%10.6 f ' , u (2: Ny -1 , 2: Nx -1) ')
39 fprintf ( ' |%10.6 f \ n ' , err )
40

41 % perform iterations until convergence

42 while err > tol
43
44 % calculate improved estimate using the Jacobi method
45 unew = u ;
46 for j = 2 : Ny - 1
47 for i = 2 : Nx - 1
48 unew (j , i ) = ( u ( j +1 , i ) + u (j , i +1) + u (j -1 , i ) + u (j , i -1) ) /
4;
49 end
50 end
51

52 % calculate absolute error

53 err = sqrt ( sum ( sum (( unew - u ) .^2) ) ) ;
54
55 % update u
56 u = unew ;
57 k = k + 1;
58
59 % output current solution
60 fprintf ( ' %4 i ', k);
61 fprintf ( ' |%10.6 f ' , u (2: Ny -1 , 2: Nx -1) ') ;
62 fprintf ( ' |%10.6 f \ n ' , err ) ;
63 end
64
65 fprintf ( '% s \ n ' , hline )
66
67 % Plot solution
68 surf (x , y , u ) ;
69 xlabel ( '$ x $ ' , ' fontsize ' , 16 , ' interpreter ' , ' latex ')
70 ylabel ( '$ y $ ' , ' fontsize ' , 16 , ' interpreter ' , ' latex ')
71 zlabel ( '$ U (x , y ) $ ' , ' fontsize ' , 16 , ' interpreter ' , ' latex ')

The MATLAB code shown in listing 3.1 is explained below:

• Lines 7 to 13 – initialises the variables that define the problem. It is good practice to list all
of the variables at the top of the program and use comments to explain what each variable
represents.

• Lines 16 to 20 – calculates the spatial steps and the node co-ordinates. Note the use of the
meshgrid command to produce two-dimensional arrays for plotting purposes.

Dr Jon Shiach 41 Finite-Difference Methods

Chapter 3. Elliptic Partial Differential Equations

• Lines 23 to 27 – defines the initial values for the finite-difference nodes. Note that this array is
inverted using the flipud command so that it is consistent with matrix indexing (i.e., element
u(1,1) corresponds to node u1,1 ). This isn’t strictly necessary and only done here so that we
can directly compare the results to a hand calculation.

• Lines 30 to 39 – prints the column headings and initial values to the command window. This
code includes fprintf commands that outputs the current iteration values. It isn’t strictly nec-
essary to do this but it can help in verifying the code, debugging and analysing the convergence
behaviour of the method.

• Line 42 – a while loop is used to iterate the Jacobi method until err > tol.

• Lines 45 to 50 – the values of uk+1 are calculated using row-by-column order. Note that the
subscript notation used in the FDS is not the same as standard matrix subscript notation, i.e.,
u(j, i) corresponds to the element ui,j .

• Line 53 – calculates the absolute L2 error.

• Lines 56 and 57 – the values of u and k are updated ready for the next iteration.

• Lines 60 to 62 – the iteration counter, current estimate and absolute L2 error are outputted to
the command window.

The Jacobi method took 33 iterations to converge to the solution shown in figure 3.11 using tol = 10−6 .
The iteration values for the Jacobi method are shown in table 3.1.

8.9 8.9 8.9 8.9 8.9

8.4 8.3800 8.5807 8.8666 9.2

7.2 7.6391 8.1764 8.7858 9.4

6.1 6.8 7.7 8.7 6.1 x

Figure 3.11: The final values of the nodes calculated using Jacobi method with tol = 10−6 .

Finite-Difference Methods 42 Dr Jon Shiach

 Chapter 3. Elliptic Partial Differential Equations

Table 3.1: Iteration values for the Jacobi method applied to solve the Laplace equation
example.

k uk1,1 uk2,1 uk3,1 uk1,2 uk2,2 uk3,2 L2 error

0 0.000000 0.000000 0.000000 0.000000 0.000000 0.000000 1.000000
1 3.500000 1.925000 4.525000 4.325000 2.225000 4.525000 8.975696
2 5.062500 4.487500 6.137500 5.756250 4.918750 6.212500 4.874463
3 6.060937 5.954688 7.200000 6.820312 6.339063 7.289062 2.930346
4 6.693750 6.825000 7.835938 7.425000 7.241016 7.909766 1.768262
5 7.062500 7.367676 8.208691 7.808691 7.764941 8.294238 1.067219
6 7.294092 7.684033 8.440479 8.031860 8.092651 8.518408 0.644121
7 7.428973 7.881805 8.575610 8.171686 8.283575 8.658282 0.388761
8 7.513373 7.997040 8.660022 8.253137 8.402943 8.739796 0.234638
9 7.562544 8.069085 8.709209 8.304079 8.472493 8.790741 0.141617
10 7.593291 8.111062 8.739956 8.333759 8.515976 8.820426 0.085473
.. .. .. .. .. .. .. ..
. . . . . . . .
32 7.639088 8.176397 8.785755 8.379958 8.580744 8.866625 0.000001
33 7.639089 8.176397 8.785755 8.379958 8.580745 8.866625 0.000001

3.8 The Gauss-Seidel method

The Jacobi method calculates the values of the next estimate using the previous estimate for all
computational nodes. This is inefficient since the values of nodes that have already been updated in
the current iteration are not used. The Gauss-Seidel method named after German mathematicians
Carl Friedrich Gauss (1777–1855) and Philipp Ludwig von Seidel (1821–1896) improves the rate of
convergence of the Jacobi method by using values updated during the current iteration. Recall that
the Jacobi method for solving Laplace’s equation, equation (3.4), is

∆y 2 (uki−1,j + uki+1,j ) + ∆x2 (uki,j−1 + uki,j+1 )

uk+1
i,j = .
2(∆x2 + ∆y 2 )

If the values of ui,j are calculated along each row from left to right before moving up to the next row,
the nodes below and to the left of ui,j have already been updated in the current iteration. Using these
updated values in equation (3.4) gives

Gauss-Seidel method for computing the solution to Laplace’s equation

∆y 2 (uk+1 k 2 k+1 k
i−1,j + ui+1,j ) + ∆x (ui,j−1 + ui,j+1 )
uk+1
i,j = , (3.8)
2(∆x2 + ∆y 2 )

which is the Gauss-Seidel method used to solve Laplace’s equation (3.3).

Example 6: Gauss-Seidel method

Perform two iterations of the Gauss-Seidel method to solve Laplace’s equation for the finite-
difference grid shown in figure 3.7 with zero starting values for the computational nodes. Cal-
culate the absolute L2 error after each iteration.

Dr Jon Shiach 43 Finite-Difference Methods

Chapter 3. Elliptic Partial Differential Equations

Solution: The first iteration of the Gauss-Seidel method for this problem is

7.2 + u02,1 + 6.8 + u01,2 7.2 + 0 + 6.8 + 0

u11,1 = = = 3.5,
4 4
u11,1 + u03,1 + 7.7 + u01,2 3.5 + 0 + 7.7 + 0
u12,1 = = = 2.8,
4 4
u12,1 + 9.4 + 8.7 + u03,2 2.8 + 9.4 + 8.7 + 0
u13,1 = = = 5.225,
4 4
8.4 + u02,2 + u11,1 + 8.9 8.4 + 0 + 3.5 + 8.9
u11,2 = = = 5.2,
4 4
u11,2 + u03,2 + u12,1 + 8.9 5.2 + 0 + 2.8 + 8.9
u12,2 = = = 4.225,
4 4
u12,2 + 9.2 + u13,1 + 8.9 4.225 + 9.2 + 5.225 + 8.9
u13,2 = = = 6.8875.
4 4
Calculating the absolute L2 error using iterations 0 and 1 gives
p
err = (3.5 − 02 ) + (2.8 − 0)2 + (5.225 − 0)2 + (5.2 − 0)2 + (4.225 − 0)2 + (6.8875 − 0)
= 11.820275.

The second iteration of the Gauss-Siedel method for this problem is

7.2 + u12,1 + 6.8 + u11,2 7.2 + 2.8 + 6.8 + 5.2

u21,1 = = = 5.5,
4 4
u21,1 + u13,1 + 7.7 + u11,2 5.5 + 5.225 + 7.7 + 5.2
u22,1 = = = 5.6625,
4 4
u22,1 + 9.4 + 8.7 + u13,2 5.6625 + 9.4 + 8.7 + 6.8875
u23,1 = = = 7.6625,
4 4
8.4 + u12,2 + u21,1 + 8.9 8.4 + 4.225 + 5.5 + 8.9
u21,2 = = = 6.75625,
4 4
u21,2 + u13,2 + u22,1 + 8.9 6.75625 + 6.8875 + 5.6625 + 8.9
u22,2 = = = 7.051563,
4 4
u22,2 + 9.2 + u23,1 + 8.9 7.051563 + 9.2 + 7.6625 + 8.9
u23,2 = = = 8.203516.
4 4
Calculating the absolute L2 error using iterations 1 and 2 gives
s
(5.5 − 3.5)2 + (5.6625 − 2.8)2 + (7.6625 − 5.225)2
err =
+ (6.75625 − 5.2)2 + (7.051563 − 4.225)2 + (8.203516 − 6.8875)2
= 5.502598.

The iteration values for the Gauss-Seidel method for this problem are shown in table 3.2. The Gauss-
Seidel method converges to the solution significantly faster than the Jacobi method. For this example,
18 iterations were required for the Gauss-Seidel method to converge using tol = 10−6 compared to 33
iterations for the Jacobi method.

Finite-Difference Methods 44 Dr Jon Shiach

 Chapter 3. Elliptic Partial Differential Equations

Table 3.2: Iteration values for the Gauss-Seidel method applied to solve the Laplace equation
example.

k uk1,1 uk2,1 uk3,1 uk1,2 uk2,2 uk3,2 L2 error

0 0.000000 0.000000 0.000000 0.000000 0.000000 0.000000 1.000000
1 3.500000 2.800000 5.225000 5.200000 4.225000 6.887500 11.820275
2 5.500000 5.662500 7.662500 6.756250 7.051562 8.203516 5.502598
3 6.604687 7.254687 8.389551 7.739062 8.024316 8.628467 2.525223
4 7.248437 7.840576 8.642261 8.143188 8.378058 8.780080 1.064415
5 7.495941 8.054065 8.733536 8.293500 8.506911 8.835112 0.396726
6 7.586891 8.131835 8.766737 8.348451 8.553849 8.855146 0.145074
7 7.620071 8.160164 8.778828 8.368480 8.570948 8.862444 0.052882
8 7.632161 8.170484 8.783232 8.375777 8.577176 8.865102 0.019266
9 7.636565 8.174243 8.784836 8.378435 8.579445 8.866070 0.007018
10 7.638170 8.175613 8.785421 8.379404 8.580272 8.866423 0.002557
.. .. .. .. .. .. .. ..
. . . . . . . .
17 7.639088 8.176397 8.785755 8.379958 8.580745 8.866625 0.000002
18 7.639089 8.176397 8.785756 8.379958 8.580745 8.866625 0.000001

3.9 The Successive Over Relaxation (SOR) method

The convergence of the Gauss-Seidel method can be increased by extrapolating the solution using a
weighted average between the current solution uki,j and the Gauss-Seidel solution for the next iteration,
ūk+1
i,j ,

i,j = (1 − ω)ui,j + ωūi,j ,

uk+1 k k
(3.9)

where ω is called the relaxation parameter. The value of the relaxation parameter is restricted to the
range 0 < ω < 2. When ω = 1 it is easy to see that equation (3.9) reduces to the Gauss-Seidel method.
When 0 ≤ ω < 1 the convergence of the Gauss-Seidel method is slowed down, this is called under
relaxation and is used when the convergence behaviour of the indirect method oscillates about the
solution. When 1 < ω ≤ 2 the convergence of the Gauss-Seidel method increased, this is called over
relaxation and is used to accelerate the convergence when the convergence behaviour is monotonic.
The optimal value of ω will depend upon the problem being solved.

Writing out equation (3.9) in full gives

Point SOR method for computing the solution to Laplace’s equation

!
∆y 2 (uk+1 k 2 k+1 k
i−1,j + ui+1,j ) + ∆x (ui,j−1 + ui,j+1 )
uki,j = (1 − ω)uki,j +ω . (3.10)
2(∆x2 + ∆y 2 )

Equation (3.10) is known as the point SOR method since it updates each point in the finite-difference
grid at a time. The choice of ω will determine the convergence rate of the point SOR method. There
are methods that can be used to find the optimal value of ω for Laplace’s and Poisson’s equations (see
Yang and Gobbert (2009) as an example) but these are outside the scope of this unit. A numerical
investigation where solutions are calculated using different values of ω and choosing the one that uses
the fewest iterations can be used to approximate an optimal ω value (but this may be different when
using a different number of nodes).

For this example the optimal value was found to be ω = 1.12 and the point SOR method required 10
iterations to converge to tol = 10−6 (table 3.3).

Dr Jon Shiach 45 Finite-Difference Methods

Chapter 3. Elliptic Partial Differential Equations

Table 3.3: Iteration values for the point SOR method applied to solve the Laplace equation
example.

k uk1,1 uk2,1 uk3,1 uk1,2 uk2,2 uk3,2 L2 error

0 0.000000 0.000000 0.000000 0.000000 0.000000 0.000000 1.000000
1 3.920000 3.253600 5.979008 5.941600 5.066656 8.160786 13.757648
2 6.024256 6.545146 8.468180 7.236463 8.027872 8.707600 5.674735
3 7.055940 7.965140 8.720186 8.199092 8.492768 8.842715 2.074922
4 7.599272 8.147607 8.778868 8.365880 8.572604 8.865286 0.605516
5 7.631864 8.173621 8.785430 8.377345 8.579838 8.866441 0.044353
6 7.638447 8.176206 8.785689 8.379838 8.580715 8.866620 0.007557
7 7.639079 8.176391 8.785760 8.379962 8.580747 8.866627 0.000674
8 7.639089 8.176400 8.785756 8.379959 8.580747 8.866626 0.000015
9 7.639090 8.176398 8.785756 8.379959 8.580746 8.866625 0.000003
10 7.639089 8.176398 8.785756 8.379959 8.580745 8.866625 0.000001

3.10 Line SOR method

The point SOR method given in equation (3.10) can improved upon by taken advantage of the efficiency
of solving tridiagonal systems. The idea is that each row of computational nodes is considered one
at a time using the point SOR method which can be written as a tridiagonal linear system for the
current row. The values of the points along the row are updated by solving the tridiagonal system
using the very efficient Thomas algorithm.

uki,j+1
j+1
sweep by rows
uk+1
i−1,j uk+1
i,j uk+1
i+1,j
bottom to top
j solving for the
nodes on each
uk+1 row
i,j−1
j−1

x
i−1 i i+1

Figure 3.12: The line SOR method solves for each row of the finite-difference at a time.

Consider the finite-difference grid in figure 3.12. If we want to update each row of nodes going from
bottom to top then the nodes ui,j , ui−1,j and ui+1,j will be updated for the (k+1)th iteration. The node
ui,j−1 has already been updated in the previous row and node ui,j+1 is yet to be updated. Rewriting
equation (3.10) for ∆x = ∆y in this way gives
!
uk+1 k+1 k+1 k
i−1,j + ui+1,j + ui,j−1 + ui,j+1
i,j = (1 − ω)ui,j + ω
uk+1 k
.
4

Finite-Difference Methods 46 Dr Jon Shiach

 Chapter 3. Elliptic Partial Differential Equations

Rearranging so that the nodes in row j with the superscript k + 1 are on the left-hand side gives

i−1,j + 4ui,j − ωui+1,j = 4(1 − ω)ui,j + ωui,j−1 + ωui,j+1 .

−ωuk+1 k+1 k+1 k k+1 k

Considering each computational node i = 1, 2, . . . , N − 2 in row j we have

−ωuk+1
0,j + 4u1,j − ωu2,j = 4(1 − ω)u1,j + ω(u1,j−1 + u1,j+1 ),
k+1 k+1 k k+1 k

−ωuk+1
1,j + 4u2,j − ωu3,j = 4(1 − ω)u2,j + ω(u2,j−1 + u2,j+1 ),
k+1 k+1 k k+1 k

..
.
−ωuk+1
N −4,j + 4uN −3,j − ωuN −2,j = 4(1 − ω)uN −3,j + ω(uN −3,j−1 + uN −3,j+1 ),
k+1 k+1 k k+1 k

−ωuk+1
N −3,j + 4uN −2,j − ωuN −1,j = 4(1 − ω)uN −2,j + ω(uN −2,j−1 + uN −2,j+1 ).
k+1 k+1 k k+1 k

Since uk+1 k+1

0,j and uN −1,j are the boundary nodes their values will be known so moving these over to the
right-hand side gives

1,j − ωu2,j = 4(1 − ω)u1,j + ω(u1,j−1 + u1,j+1 + u0,j ),

4uk+1 k+1 k k+1 k k+1

−ωuk+1
1,j + 4u2,j − ωu3,j = 4(1 − ω)u2,j + ω(u2,j−1 + u2,j+1 ),
k+1 k+1 k k+1 k

..
.
N −4,j + 4uN −3,j − ωuN −2,j = 4(1 − ω)uN −3,j + ω(uN −3,j−1 + uN −3,j+1 ),
−ωuk+1 k+1 k+1 k k+1 k

N −3,j + 4uN −2,j = 4(1 − ω)uN −2,j + ω(uN −2,j−1 + uN −2,j+1 + uN −1,j ).
−ωuk+1 k+1 k k+1 k k+1

which can be written as the matrix equation

   k+1   
4 −ω u1,j 4(1 − ω)uk1,j + ω(uk+1 k k+1
1,j−1 + u1,j+1 + u0,j )
−ω 4 −ω  uk+1   4(1 − ω)uk2,j + ω(uk+1 k 
  2,j 


 2,j−1 + u2,j+1 ) 

 .. .. ..   ..   . .
 . . .  .  =  .. 
   k+1   
 −ω 4 −ω  uN −3,j   4(1 − ω)uN −3,j + ω(uN −3,j−1 + uN −3,j+1 )
k k+1 k

−ω 4 uk+1
N −2,j 4(1 − ω)u k
N −2,j + ω(u k+1
N −2,j−1 + u k
N −2,j+1 + u k+1
N −1,j )

This matrix equation can be solved using any linear solver but since it is a tridiagonal system, the
Thomas algorithm is by far the most efficient for doing this.

3.10.1 Thomas algorithm

The Thomas algorithm is a well known highly efficient solution algorithm for solving tridiagonal
systems. It works by using LU decomposition where the system is manipulated so that it is an upper
triangular form which can then be solved using back substitution. Consider a linear system of N
equations of the form
    
b1 c1 x1 d1
a2 b2 c2  x2   
d2
    
 .. .. ..  ..   ..
 . . .  . = ,.
    
 aN −1 bN −1 cN −1  xN −1  dN −1 
aN bN xN dN

where ai , bi and ci are elements of the lower, main and upper diagonals respectively. The Thomas
algorithm proceeds by performing a forward sweep modifying the upper diagonal coefficients ci and

Dr Jon Shiach 47 Finite-Difference Methods

 Chapter 3. Elliptic Partial Differential Equations

the right-hand side values di

c

 ,
i
i = 1,
bi
c′i = ci

 b − a c′ , i = 2, . . . , N,
i i i−1


 di
 , i = 1,
b
d′i = dii − ai d′

 i−1
 , i = 2, . . . , N.
bi − ai c′i−1
The solution is obtained using back substitution
xN = d′N ,
xi = d′i − c′i xi+1 , i = N − 1, . . . , 1.
The derivation of the Thomas algorithm can be found in appendix A on page 117 and the MATLAB
code is given in listing 3.2.
Listing 3.2: MATLAB function for the Thomas algorithm
1 function x = Tridiag (A , d )
2
3 % This function solves a tridiagonal linear system of the form Ax = d using
4 % the Thomas algorithm
5
6 % Calculate size of the system
7 N = length ( d ) ;
8
9 % Determine lower , main and upper diagonal elements
10 a = [0 ; diag (A , -1) ];
11 b = diag ( A ) ;
12 c = [ diag (A , 1) ; 0];
13
14 % Forward sweep
15 c (1) = c (1) / b (1) ;
16 d (1) = d (1) / b (1) ;
17 for i = 2 : N
18 c ( i ) = c ( i ) / ( b ( i ) - a ( i ) * c (i -1) ) ;
19 d ( i ) = ( d ( i ) - a ( i ) * d (i -1) ) / ( b ( i ) - a ( i ) * c (i -1) ) ;
20 end
21
22 % Backward sweep
23 x(N) = d(N);
24 for i = N - 1 : -1 : 1
25 x ( i ) = d ( i ) - c ( i ) * x ( i +1) ;
26 end
27
28 end

3.10.2 Optimal value of the relaxation parameter

The optimal value of the relaxation parameter is the value of ω for which the SOR method takes the
fewest number of iterations to converge to a solution. For tridiagonal systems of the form Au = d,
such as the one seen in the line SOR method, it can be shown that the optimal value of ω is
2
ωopt = p . (3.11)
1 + 1 − ρ(TJ )2
TJ is the iteration matrix for the Jacobi method defined by
TJ = D−1 (L + U ),

Finite-Difference Methods 48 Dr Jon Shiach

 Chapter 3. Elliptic Partial Differential Equations

where A = D − L − U , i.e., D are the diagonal elements of A and L and U are the a lower and upper
triangular elements of A respectively with their signed reversed, ρ(TJ ) the spectral radius of A which
is defined as the largest absolute eigenvalue of TJ .
Since A that is used to calculate ωopt contains ω we have an implicit relationship. In order to determine
ωopt , we first calculate equation (3.11) using some starting value of ω and then use this to calculate
A and the next estimate of ωopt . The iterations continue until two successive estimates agree to some
tolerance. For this Laplace equation example, ωopt = 1.036 and the line SOR method took just 8
iterations to converge to the solution (table 3.4).

Table 3.4: Iteration values for the line SOR method applied to solve the Laplace equation
example.

k uk1,1 uk2,1 uk3,1 uk1,2 uk2,2 uk3,2 L2 error

0 0.000000 0.000000 0.000000 0.000000 0.000000 0.000000 1.000000
1 4.825480 4.739728 5.881230 7.647557 7.573769 8.125412 16.192856
2 7.472641 7.934598 8.619766 8.346390 8.531502 8.833439 5.159570
3 7.633276 8.167999 8.780027 8.379191 8.579668 8.865868 0.332282
4 7.639055 8.176354 8.785722 8.379981 8.580778 8.866648 0.011751
5 7.639100 8.176413 8.785766 8.379962 8.580750 8.866629 0.000095
6 7.639090 8.176399 8.785757 8.379959 8.580746 8.866625 0.000021
7 7.639089 8.176398 8.785756 8.379959 8.580745 8.866625 0.000002
8 7.639089 8.176398 8.785756 8.379959 8.580745 8.866625 0.000000

Table 3.5 contains an analysis of the rate of convergence for each indirect method discussed in this
chapter. The CPU time was recorded by placing the tic and toc MATLAB commands at each end
of the program and taking the average time over several runs. The normalised column was calculated
by dividing the CPU time taken for each method by the time taken for the Jacobi method. The
Jacobi method is the most computationally expensive with the Gauss-Seidel, point SOR and line SOR
methods taking 1.7%, 1.3% and 1.3% of the time taken by the Jacobi method respectively. Usually it
is expected that the line SOR method would offer more of a saving than the point SOR method but
in this example there aren’t enough nodes in each row of the finite-difference grid for the increased
efficiency to take effect.

Table 3.5: Analysis of the rate of convergence of different indirect methods applied to the
Laplace equation example.

Method Iterations CPU time (s) Normalised CPU time

Jacobi 33 3.16 × 10−5 1.000
Gauss-Seidel 18 1.82 × 10−7 0.017
SOR 10 1.37 × 10−7 0.013
Line SOR 8 1.35 × 10−7 0.013

Example 7

An electrical field over a two-dimensional domain is to be modelled using Poisson’s equation

Uxx + Uyy = f (x, y),

where f (x, y) = x + y. The domain of 0 ≤ x, y ≤ 1 is discretised using 4 nodes in the x and y

directions and Dirichlet boundary conditions provide the values of the nodes on the boundary
as shown in the diagram below.

Dr Jon Shiach 49 Finite-Difference Methods

Chapter 3. Elliptic Partial Differential Equations

4 3 2 1

3 u1,2 u2,2 1

2 u1,1 u2,1 1

1 1 1 1 x

Compute the first two iterations of the Jacobi, Gauss-Seidel and point SOR methods (with
ω = 1.1) for this problem.

Solution: Using the second-order symmetric differences from the finite-difference toolkit to
approximate Uxx and Uyy ,
ui−1,j − 2ui,j + ui+1,j
Uxx ≈ ,
∆x2
ui,j−1 − 2ui,j + ui,j+1
Uyy ≈ ,
∆y 2
then Poisson’s equation becomes

Uxx + Uyy = x + y
ui−1,j − 2ui,j + ui+1,j ui,j−1 − 2ui,j + ui,j+1
+ = x i + yj .
∆x2 ∆y 2
1
Since ∆x = ∆y = 3 then multiplying both sides by ∆x2 = ( 13 )2 = 1
9 gives
x i + yj
−4ui,j + ui−1,j + ui+1,j + ui,j−1 + ui,j+1 = ,
9
and rearranging to make ui,j the subject gives
ui−1,j + ui+1,j + ui,j−1 + ui,j+1 xi + yj
ui,j = − .
4 36
The Jacobi method for solving this finite-difference scheme is

uki−1,j + uki+1,j + uki,j−1 + uki,j+1 xi + yj

uk+1
i,j = − .
4 36
1
The co-ordinates of the computational nodes are x1 = y1 = 3 and x2 = y2 = 32 . Calculating
the first iteration:
2 + u02,1 + 1 + u01,2 1
+ 13 2+0+1+0 2
u11,1 = − 3
= − = 0.731481,
4 36 4 108
u01,1 + 1 + 1 + u02,2 2
+1 0+1+1+0 1
u12,1 = − 3 3 = − = 0.472222,
4 36 4 36
3 + u02,2 + u01,1 + 3 1
+2 3+0+0+3 1
u11,2 = − 3 3 = − = 1.472222,
4 36 4 36
u1,2 + 1 + u02,1 + 2
0 2
+2 0+1+0+2 4
u12,2 = − 3 3 = − = 0.712963
4 36 4 108

Finite-Difference Methods 50 Dr Jon Shiach

 Chapter 3. Elliptic Partial Differential Equations

The values of uki,j for the first two iterations are shown in the table below.

k u1,1 u2,1 u1,2 u2,2 L2 error

0 0.000000 0.000000 0.000000 0.000000 -
1 0.731481 0.472222 1.472222 0.712963 1.853055
2 1.217593 0.833333 1.833333 1.199074 0.856394

The Gauss-Seidel method for solving this finite-difference scheme is

uk+1 k k+1 k
i−1,j + ui+1,j + ui,j−1 + ui,j+1 xi + yj
uk+1
i,j = − .
4 36
Calculating the first iteration:

2 + u02,1 + 1 + u01,2 1
+ 13 2+0+1+0 2
u11,1 = − 3
= − = 0.731481,
4 36 4 108
u1,1 + 1 + 1 + u02,2
1 2
+1 0.731481 + 1 + 1 + 0 1
u12,1 = − 3 3 = − = 0.655093,
4 36 4 36
3 + u02,2 + u11,1 + 3 1
+2 3 + 0 + 0.731481 + 3 1
u11,2 = − 3 3 = − = 1.655093,
4 36 4 36
u11,2 + 1 + u12,1 + 2 2
+2 1.655093 + 1 + 0.655093 + 2 4
u12,2 = − 3 3 = − = 1.290509.
4 36 4 108
The values of uki,j for the first two iterations are shown in the table below.

k u1,1 u2,1 u1,2 u2,2 L2 error

0 0.000000 0.000000 0.000000 0.000000 -
1 0.731481 0.655093 1.655093 1.290509 2.317101
2 1.309028 1.122106 2.122106 1.524016 0.907904

The SOR method for solving this finite-difference scheme is

!
uki,j+1 + uki+1,j + uk+1 k+1
i,j−1 + ui−1,j xi + yj
uk+1
i,j = (1 − ω)uki,j +ω − .
4 36

Dr Jon Shiach 51 Finite-Difference Methods

Chapter 3. Elliptic Partial Differential Equations

Calculating the first iteration (using ω = 1.1):

!
2 + u02,1 + 1 + u01,2 1
+1
u11,1 = −0.1u01,1 + 1.1 − 3 3
4 36
 
2+0+1+0 1
= 0 + 1.1 − = 0.804630,
4 54
!
u 1 + 1 + 1 + u0 2 1
1,1 2,2 +
u12,1 = −0.1u02,1 + 1.1 − 3 3
4 36
 
0.804630 + 1 + 1 + 0 1
= 0 + 1.1 − = 0.740718,
4 36
!
3 + u02,2 + u11,1 + 3 1
+ 2
u11,2 = −0.1u1,2 + 1.1
0
− 3 3
4 36
 
3 + 0 + 0.804630 + 3 1
= 0 + 1.1 − = 1.840718,
4 36
!
u11,2 + 1 + u12,1 + 2 2
3 + 3
2
u12,2 = −0.1u2,2 + 1.1
0
−
4 36
 
1.840718 + 1 + 0.740718 + 2 2
= 0 + 1.1 − = 1.494154.
4 54

The values of uki,j for the first two iterations are shown in the table below.
k u1,1 u2,1 u1,2 u2,2 L2 error
0 0.000000 0.000000 0.000000 0.000000 -
1 0.804630 0.740718 1.840718 1.494154 2.610906
2 1.434061 1.250632 2.240632 1.594941 0.909003

Finite-Difference Methods 52 Dr Jon Shiach

Chapter 4

Hyperbolic Partial Differential

Equations

Hyperbolic PDEs are used to model propagation problems where a quantity travels across a domain.
Common applications are the modelling of water flow, weather forecasting, blast wave analysis and
aerodynamics. Unlike elliptic equations where the solutions are steady state, hyperbolic PDEs contain
a time derivative term so the solution will evolve over time from a set of initial conditions. Any
perturbation in the dependent variables will propagate outward over time, however, unlike elliptic
PDEs, not every point in the domain will feel this disturbance at once. For example, consider dropping
a stone in a pool of water initially at rest. Waves caused by the disturbance will propagate out from
where the stone was dropped and not all parts of the pool will be affected by the disturbance at once.
As the waves propagate across the surface of the pool, the region that is affected by the disturbance
expands until all of the pool has been affected. This chapter will discuss the use of FDS to solve the
advection equation and introduces the concept of the stability of a numerical scheme.

4.1 The advection equation

The simplest example of a hyperbolic PDE is the one-dimensional advection equation that models
pure advection along a one-dimensional channel. The one-dimensional advection equation is written
as

The advection equation

Ut + vUx = 0 (4.1)

where t and x are independent variables denoting time and space respectively and v is the velocity.
If the initial conditions are described by the function, U (0, x), then the exact solution at time t is
U (t, x − vt).

Consider figure 4.1 where an initial profile is defined that consists of a triangular perturbation (dashed
red line). If v is the advection velocity, then at time t the triangular profile has travelled distance vt
in the x direction. Note that the triangular perturbation has not deformed through any other process.

53
Chapter 4. Hyperbolic Partial Differential Equations

U (t, x) vt

x0 x1 x2 x0 + vt x1 + vt x2 + vt x

Figure 4.1: Pure advection of an initial profile. Note that the shape is unchanged is it
propagates along the x domain.

The advection equation is useful for developing FDSs to solve hyperbolic PDEs since it is simple to
calculate an exact solution for a problem and discontinuities which are challenging for a numerical
scheme to resolve can be easily specified. Any numerical scheme that performs poorly when solving
an advection problem will be unsuitable for other applications involving more complicated hyperbolic
PDEs.

4.2 The First-Order Upwind (FOU) scheme

The derivation of a finite-difference scheme (FDS) for solving the advection equation involves approx-
imating the partial derivatives using the finite-differences from the toolkit (table 2.4 on page 23). For
example, using a first-order forward difference to approximate Ut and a first-order backward difference
for Ux , i.e.,
un+1 − uni
Ut ≈ i
,
∆t
un − uni−1
Ux ≈ i ,
∆x
results in
un+1 − uni un − uni−1
i
+v i = 0.
∆t ∆x
Rearranging to make un+1
i the subject gives
v∆t n
un+1 = uni − (u − uni−1 ).
i
∆x i
Since both the time and space derivatives are approximated using first-order differences then we say
that this FDS is first-order accurate in time and space and we write this using O(∆t, ∆x)

The FOU scheme for the advection equation when v > 0

v∆t n
un+1 = uni − (u − uni−1 ) + O(∆t, ∆x). (4.2)
i
∆x i

Assuming that v in equation (4.1) is positive then the direction of flow is from left-to-right along the
x domain. To capture this flow, the discretisation of Ux needs to use the neighbouring node in the
upwind direction (the node to the left). Therefore equation (4.2) is known as the first-order upwind
scheme. If v is negative, then flow travels from right-to-left, therefore the spatial discretisation uses
the node to the right, and the FOU is
The FOU scheme for the advection equation when v < 0

v∆t n
un+1 = uni − (u − uin ) + O(∆t, ∆x). (4.3)
i
∆x i+1

Finite-Difference Methods 54 Dr Jon Shiach

 Chapter 4. Hyperbolic Partial Differential Equations

Equations (4.2) and (4.3) are examples of time marching schemes where the solution of u(t, x) is
marched forward in time by one time step of ∆t through a single application of the scheme. The
stencil for the FOU scheme when used to solve the advection equation can be seen in figure 4.2.

un+1
i un+1
i

uni−1 uni uni uni+1

(a) v > 0 (b) v < 0

Figure 4.2: Finite-difference stencils for the FOU scheme.

4.2.1 Solution procedure

The solution procedure for solving the advection equation is summarised in the flow chart in figure 4.3.
Once the variables and arrays have been initialised, a loop is required to march the solution forward
through time until t = tmax . Within the time step loop, the boundary conditions are calculated prior
to calculating un+1 using the chosen FDS. The current solution un is updated by setting it equal to
the updated solution and the time step is incremented by ∆t. Finally the current solution is plotted
to produce an animation.

start

calculate x and u0

no
t < tmax end

yes
calculate boundary conditions

calculate un+1 using FDS

Update un ← un+1 and t ← t + ∆t

output un

Figure 4.3: Flow chart for applying a time marching scheme to solve the advection equation.

Example 8

The concentration of a pollutant in a river is to be modelled using the advection equation. The
stretch of river under study is of length 10m with an average flow velocity of v = 5ms−1 . At
time t = 0s a pollutant is released into the river such that the concentration of pollutant in the

Dr Jon Shiach 55 Finite-Difference Methods

Chapter 4. Hyperbolic Partial Differential Equations

river, U (t, x), can be described by

(
1, 2 ≤ x ≤ 4,
U (0, x) =
0, elsewhere.

The domain is discretised using a finite-difference grid of six nodes and first-order zero gradient
boundary conditions are used to calculate the ghost nodes such that

u−1 = u0 , uN = uN −1 .

Use the FOU scheme with ∆t = 0.25 to solve the advection equation to model the concentration
of the pollutant after t = 1s.

Solution: The first thing we need to do is discretise the spatial domain. Since 0m ≤ x ≤ 10m
and we are using 6 computational nodes the spatial step is
10
∆x = = 2m,
6−1
so the co-ordinates of the nodes are

x0 = 0, x1 = 2, x2 = 4, x3 = 6 x4 = 8, x5 = 10.

The finite-difference grid for this problem can be represented using the diagram below. The
computational nodes are represented by filled circles and the ghost nodes u−1 and u6 are repre-
sented by unfilled circles. The finite-difference schemes will be applied to calculate the solutions
for the computational nodes whereas the ghost nodes will be calculated using the zero gradient
boundary condition given above.
u−1 u0 u1 u2 u3 u4 u5 u6
x
0 2 4 6 8 10
The concentration of the pollutant at t = 0s is given by a case statement where the initial values
of the nodes is 1 where the node co-ordinates are between 2 and 4 inclusive and 0 elsewhere,
therefore

u00 = 0, u01 = 1, u02 = 1, u03 = 0, u04 = 0, u05 = 0,

and the ghost nodes are u−1 = u0 = 0 and u6 = u5 = 0. We now need to compute the
solution using the FOU scheme for each computational node. Since v, ∆t and ∆x are constant
throughout we can precalculate v∆t
∆x

v∆t 5(0.25)
= = 0.625.
∆x 2
The first step of the FOU is

u10 = u00 − 0.625(u00 − u0−1 ) = 0 − 0.625(0 − 0) = 0,

u11 = u01 − 0.625(u01 − u00 ) = 1 − 0.625(1 − 0) = 0.375,
u12 = u02 − 0.625(u02 − u01 ) = 1 − 0.625(1 − 1) = 1,
u31 = u03 − 0.625(u03 − u02 ) = 0 − 0.625(0 − 1) = 0.625,
u14 = u04 − 0.625(u04 − u03 ) = 0 − 0.625(0 − 0) = 0,
u15 = u05 − 0.625(u05 − u04 ) = 0 − 0.625(0 − 0) = 0.

The values for the computational nodes for the first four time steps are shown in the table
below.

Finite-Difference Methods 56 Dr Jon Shiach

 Chapter 4. Hyperbolic Partial Differential Equations

n tn un0 un1 un2 un3 un4 un5

0 0.00 0.000000 1.000000 1.000000 0.000000 0.000000 0.000000
1 0.25 0.000000 0.375000 1.000000 0.625000 0.000000 0.000000
2 0.50 0.000000 0.140625 0.609375 0.859375 0.390625 0.000000
3 0.75 0.000000 0.052734 0.316406 0.703125 0.683594 0.244141
4 1.00 0.000000 0.019775 0.151611 0.461426 0.695801 0.518799

4.3 Validating the time marching schemes

To validate a numerical scheme (test whether it provides suitable approximations) we can make use
of test cases. These are often simplified scenarios where problems are defined using idealised values
that can often mean we have a known exact solution. To validate the FDS used to solve the advection
equation, a test case that involves the advection of a Gaussian curve profile along a channel is used.
The spatial domain is defined as 0 ≤ x ≤ 1 and the initial conditions are described by

U (0, x) = f (x) = exp(−200(x − 0.25)2 ),

which produces a Gaussian curve centred at x = 0.25. The exact solution for the advection equation
is U (t, x − vt) so for this test case

U (t, x) = f (x − vt) = exp(−200(x − vt − 0.25)2 ).

The FDS is calculated using 101 nodes so that ∆x = 0.01 and a time step of ∆t = 0.01 is used.
Transmissive boundary conditions are employed at either end of the domain so that a quasi-infinite
domain is modelled. This is achieved by using the following values for the ghost nodes

u−1 = u1 , uN = uN −2 ,

The velocity is set at v = 0.5 and the FDS is marched forward in time until t = 1 when the centre of
the Gaussian curve is at x = 0.75.

The computed solutions for the FOU scheme are plotted against the exact solution in figure 4.4 for
increasing values of t. The numerical solutions show a marked reduction in the height and width of
the Gaussian curve which is due to the FOU scheme only being first-order accurate in space. This
phenomena is called numerical dissipation and is a common feature, although undesirable, of first-order
methods.

Dr Jon Shiach 57 Finite-Difference Methods

 Chapter 4. Hyperbolic Partial Differential Equations

Figure 4.4: The solution of the advection equation using the FOU scheme.

The MATLAB code used to perform the calculations is given in listing 4.1.

Listing 4.1: MATLAB code for the FOU scheme used to solve the advection equation.
1 % advection . m by Jon Shiach
2 %
3 % This program solves the advection equation using the FOU scheme
4
5 % Clear workspaces
6 clear
7 clc
8
9 % Define variables

Finite-Difference Methods 58 Dr Jon Shiach

 Chapter 4. Hyperbolic Partial Differential Equations

10 N = 101; % number of nodes

11 xmin = 0; % lower bound of x
12 xmax = 1; % upper bound of x
13 v = 0.5; % velocity
14 t = 0; % time variable
15 tmax = 1; % max value of t
16 dt = 0.01; % time step
17
18 % Discretise the domain ( remember to include ghost nodes )
19 dx = ( xmax - xmin ) / ( N - 1) ;
20 x = xmin - dx : dx : xmax + dx ;
21
22 % Define initial conditions
23 u = f(x); % solution array u ^ n
24
25 % Time marching loop
26 while t < tmax
27
28 % Ensure that t does not exceed tmax
29 dt = min ( dt , tmax - t ) ;
30
31 % Calculate boundary conditions
32 u (1) = u (3) ;
33 u ( N +2) = u ( N ) ;
34

35 % Calculate values of u ^( n +1) using the FOU scheme

36 unew = u ;
37 C = v * dt / dx ;
38 for i = 2 : N + 1
39 unew ( i ) = u ( i ) - C * ( u ( i ) - u (i -1) ) ;
40 end
41
42 % Update u and t
43 u = unew ;
44 t = t + dt ;
45
46 % Calculate exact solution
47 uexact = f ( x - v * t ) ;
48
49 % Plot the numerical against the exact solution
50 plot (x , uexact , 'r - ' , ' linewidth ' , 2)
51 hold on
52 plot (x , u , 'bo - ' , ' markerfacecolor ' , 'b ')
53 hold off
54 axis ([ xmin , xmax , -0.5 , 1.5])
55 xlabel ( '$ x $ ' , ' fontsize ' , 16 , ' interpreter ' , ' latex ')
56 ylabel ( '$ u (t , x ) $ ' , ' fontsize ' , 16 , ' interpreter ' , ' latex ')
57 title ( sprintf ( '$ t = %1.3 f $ s ' , t ) , ' fontsize ' , 16 , ' interpreter ' , ' latex
')
58 shg
59 pause (0.001)
60 end
61
62 % Add legend to plot
63 leg = legend ( ' exact solution ' , ' numerical solution ') ;
64 set ( leg , ' fontsize ' , 12 , ' interpreter ' , ' latex ')
65
66 % -------------------------------------------------------------------------
67 function u = f ( x )
68
69 % This function defines the initial profile .

Dr Jon Shiach 59 Finite-Difference Methods

 Chapter 4. Hyperbolic Partial Differential Equations

70 u = exp ( -200 * ( x - 0.25) .^2) ;

71
72 end

The MATLAB code shown in listing 4.1 is explained below:

• Lines 10 to 16 – initialises the variables used to define the problem.

• Lines 19 and 20 – calculates the spatial step ∆x and the x co-ordinates of the finite-difference
nodes. Note that the co-ordinates of the ghost nodes (x−1 = xmin − ∆x and xN +1 = xmax + ∆x)
have also been included in the x() array

• Line 23 – initialises the solution array u() using a Gaussian curve function f(x) that is defined
at the bottom of the m-file.

• Line 26 – a while loop is used to loop through time until t = tmax.

• Line 29 – the value of dt is checked to ensure that t does not exceed tmax.

• Lines 32 and 33 – the values of the ghost nodes are calculated so that transmissive boundary
conditions are employed.

• Lines 36 to 40 – a for loop is used to loop through the values of u() and calculate unew().
Since we have added two ghost nodes to the finite-difference grid and arrays in MATLAB start
at 1, the index 2 : N + 1 correspond to nodes 0, . . . , N − 1.

• Lines 43 and 44 – u are t are updated ready for the next iteration.

• Line 47 – the exact solution to the advection equation is calculated using the same function that
was used to define the initial conditions.

• Lines 50 to 59 – the numerical and exact solutions are plotted for the current time step. This
produces an animation which shows the evolution of the wave which will loop back around and
finish where it started due to the use of the periodic boundary conditions. Plotting the solution
at every time step is useful for seeing the evolution of the solution over time but can cause the
program to slow down considerably.

• Lines 67 to 72 – a function f(x) is defined that calculates the initial profile given the array
containing the node co-ordinates x(). This function is used twice in the main program, first
it is used to define the initial values of u() and then again to calculate the exact solution
U (t, x) = f (x − vt) at each time step.

4.4 The Forward-Time Central-Space (FTCS) scheme

We have seen in figure 4.4 that the FOU schemes gives a relatively poor approximation of the exact
solution due to a first-order finite-difference approximation being used to approximation Ux in equa-
tion (4.1). Using a second-order central difference approximation should improve the accuracy of our
numerical solutions. Recall that that central difference approximation is

ui−1 + ui+1
Ux = + O(∆x2 ),
2∆x

and substituting this along with a forward difference approximation for Ut into equation (4.1) gives

un+1 − uni un − uni−1

i
+ v i+1 = 0,
∆t 2∆x

which is rearranged to give

Finite-Difference Methods 60 Dr Jon Shiach

 Chapter 4. Hyperbolic Partial Differential Equations

The FTCS scheme for solving the advection equation

v∆t n
un+1 = uni − (u − uni−1 ) + O(∆t, ∆x2 ), (4.4)
i
2∆x i+1

This FDS is known as the Forward-Time Central Space (FTCS) scheme. The FTCS scheme is first-
order accurate in time and second-order in space. The stencil for the FTCS scheme is given in
figure 4.5.

un+1
i

uni−1 uni uni+1

Figure 4.5: Finite-difference stencil for the FTCS scheme.

A plot of solution for the FTCS scheme when applied to solve the test problem given in section 4.3
at t = 1 is shown in figure 4.6. Here it is clear that something has gone wrong and we are seeing
the solution oscilate behind the Gaussian curve. If the calculations were allowed to continue these
oscilations would increase and the program would eventually crash. This is an example of an unstable
scheme and it is well known that the FTCS scheme is unstable when applied to the advection equation.
It was included here to demonstrate that not all FDS are applicable for certain PDEs. It is important
that we can determine whether a FDS is stable for a particular PDE.

Figure 4.6: The numerical and exact solutions of the advection equation at t = 0.1 using
the FTCS scheme.

4.5 Convergence, consistency and stability

We have seen in figure 4.6 that not all FDS are suitable for solving the advection equation. Once we
have derived a FDS we need to determine whether it can be applied to solve the PDE. One way to do
this is to apply the scheme to a test problem to see if it approximates the known solution. However,
this is time consuming and may not give you the correct result, for example a suitable scheme may
be deemed unsuitable due to errors in the code.

Dr Jon Shiach 61 Finite-Difference Methods

Chapter 4. Hyperbolic Partial Differential Equations

There are three important characteristics that a FDS must have in order for it to be suitable for
solving a PDE. These are consistency, stability and convergence.
Definition 7: Consistency

A FDS is said to be consistent if as ∆t, ∆x → 0 then the numerical scheme is equivalent to the
PDE.

Definition 8: Stability

A FDS is said to be stable if the errors at each step of the time marching scheme caused by
the finite-difference approximations and computational rounding remain bounded and do not
increase.

Definition 9: Convergence

A FDS is said to be convergent if as ∆t, ∆x → 0, u(tn , xi ) → U (t, x) (the numerical solution

approaches the exact solution of the PDE).

These three characteristics are related by the Lax equivalence theorem which is stated below
Theorem 1: Lax equivalence theorem

consistency + stabilily ⇔ convergence

i.e., a scheme is convergent if it is consistent and stable (Lax and Richtmyer, 1956).

Convergence is essential if we want to apply our FDS to solve a PDE, therefore we must check to see
if a FDS is both consistent and stable.

4.5.1 Checking for consistency

To check that a scheme is consistent we replace the terms involving the individual nodes, uni , with the
corresponding Taylor series expansion for that node. Then we can write the resulting expression in
the form of the PDE that is being solved plus the truncation error. If the truncation error tends to
zero as ∆t, ∆x → 0 then the scheme is consistent. Furthermore, the formal accuracy of the scheme
can be determined by the truncation error.
Example 9

Show that the FOU scheme is consistent for the advection equation and has formal accuracy
O(∆t, ∆x).

Solution: The FOU scheme is

v∆t
un+1
i = uni − (ui − ui−1 ),
∆x
so we need to replace ui with U and un+1
i and uni−1 with Taylor series expansions for U (t+∆t, x)
and U (t, x − ∆x) respectively

uni = U,
∆x2
uni−1 = U − ∆xUx + Uxx + O(∆x3 ),
2
∆t2
un+1
i = U + ∆tUt + Utt + O(∆t3 ).
2

Finite-Difference Methods 62 Dr Jon Shiach

 Chapter 4. Hyperbolic Partial Differential Equations

Therefore the FOU scheme becomes

 
∆t2 v∆t ∆x2
U + ∆tUt + Utt + O(∆t ) = U −
3
U − U + ∆xUx − 3
Uxx + O(∆t ) .
2 ∆x 2

Simplifying
∆t v∆x
Utt + O(∆t2 ) = −vUx +
Ut + Uxx + O(∆x2 )
2 2
∆t v∆x
Ut + vUx + Utt + O(∆t2 ) − Uxx + O(∆x2 ) = 0.
2 2
For the FOU scheme to be consistent we need this equation to be equivalent to the advection
equation plus the additional truncation errors. As it stands we have two terms involving second-
order derivatives Utt and Uxx , however, these can be omitted by increasing the truncation errors
from O(∆t2 ) and O(∆x2 ) to O(∆t) and O(∆x) so we now have

Ut + vUx + O(∆t) + O(∆x) = 0.

We have shown that the FOU scheme is consistent for the advection equation and is first-order
accurate in time and space.

4.5.2 The CFL condition

The program used to solve the advection equation with the FOU scheme in listing 4.1 uses a small
time step of ∆t = 0.01. Since each step of the time marching scheme calculates the solution over
a time bound of ∆t, the larger the value of ∆t, fewer time steps are required and therefore fewer
computations are needed to calculate u(t, x). Attempting to run the program with a larger value of
the time step, ∆t = 0.1 say, results in the solution becoming unstable. There exists a condition on
the value of ∆t which is known as the CFL condition ∗ (Courant et al., 1967) which must be satisfied
in order for a numerical scheme to be stable.
Consider figure 4.7(a) that shows the advection of a Gaussian curve in space and time. We can define
a curve in the t–x space by tracking the propagation of a single point on the curve over time (e.g., the
peak of the curve). This curve is known as a characteristic curve which can the described in this case
by x0 + vt (figure 4.7(b)).

t x0 + vt

x0 x
(a) Gaussian curve. (b) Characteristic curve.

Figure 4.7: The advection of a Gaussian curve in the xt domain and its corresponding
characteristic curve.
∗
Named after German mathematicians Richard Courant (1888–1972), Kurt Otto Friedrichs (1901–1982) and Hans
Lewy (1904–1988).

Dr Jon Shiach 63 Finite-Difference Methods

Chapter 4. Hyperbolic Partial Differential Equations

Definition 10: The CFL condition

When using a fixed grid to calculate the solution to a PDE, the characteristic of the PDE cannot
travel more than one spatial step in a single time step. This can be written as

|v|∆t ≤ ∆x,

This can be rearranged to give an expression for the Courant number

|v|∆t
C= ≤ 1, (4.5)
∆x
which according to the CFL condition must satisfy C ≤ 1.

Definition 11: Numerical domain of dependence

The numerical domain of dependence is a set of all nodes that are required to compute the
value of the node uni for a given finite-difference scheme.

uni

un−1
i−1 un−1
i

Figure 4.8: Numerical domain of dependence for the FOU scheme.

Consider figure 4.8 that shows the numerical domain of dependence for the FOU scheme used to solve
the advection equation where node uni has been calculated using the two nodes un−1
i−1 and ui
n−1
which
are each calculated using two nodes at the previous time step and so on. For the CFL condition to
be satisfied, the characteristic represented by the dashed line must be contained within the numerical
domain of dependence.

The CFL condition is useful for calculating a value of the time step. Rearranging equation (4.5) gives

∆x
∆t ≤ . (4.6)
|v|

4.5.3 von Neumann stability analysis

The CFL condition is a necessary condition for stability but not a sufficient one. This means that the
CFL condition must be satisfied for a numerical scheme to be stable but just because it is satisfied
does not mean that the numerical scheme is stable. The purpose of stability analysis is to determine
whether a numerical scheme is stable.

The stability of a numerical scheme can be analysed using von Neumann stability analysis named after
Hungarian mathematician John von Neumann (1903–1957). This assumes that the PDE is linear and
periodic boundary conditions are employed but its implementation is often applied to other non-linear
PDEs with non-periodic boundary conditions.

Finite-Difference Methods 64 Dr Jon Shiach

 Chapter 4. Hyperbolic Partial Differential Equations

Consider the FTCS scheme

C n
un+1 = unj − (u − unj−1 ),
j
2 j+1

where C = v∆t∆x is the Courant number. Note that the subscripts are written in terms of j and not
i as used elsewhere so to avoid confusion when using complex numbers. The error in the numerical
approximation, εnj = Ujn − unj , of the approximations also satisfy the FTCS scheme, i.e.,

C n
εn+1 = εnj − (ε − εnj−1 ). (4.7)
j
2 j+1

For linear PDEs with periodic boundary conditions, assuming that the errors grow or shrink expo-
nentially over time then we can write the error at each node as a Fourier mode†

εni = eat eikm x , (4.8)

√
where a is some constant, km is a wave number and i = −1 is the imaginary number. In this
equation the eat term describes the growth of the error at a single node as a function of time. The
growth of the error in a numerical scheme can be determined by substituting equation (4.8) into (4.7)

C at ikm (x+∆x)
ea(t+∆t) eikm x = eat eikm x − (e e − eat eikm (x−∆x) )
2

Dividing throughout by eat eikm x gives

C ikm ∆x
ea∆t = 1 − (e − e−ikm ∆x ) (4.9)
2

The term on the left-hand side is known as the growth factor which gives the growth in the error over
one time step of ∆t so we use G = ea∆t for convenience. For stability we require

|G| ≤ 1, (4.10)

i.e., the errors do not increase over a single step. Using the exponential trigonometric identity

eix − e−ix
sin(x) = ,
2i

we can combine equations (4.9) and (4.10) to give the following condition for stability

1 ≥ |1 − iC sin(km ∆x)|.
√
The modulus of a complex number is defined by |a + bi| = a2 + b2 so squaring both sides gives

1 ≥ 1 + C 2 sin2 (km ∆x).

Since sin2 (x) is bounded by the range [0, 1] and C 2 is always positive then this inequality only holds
when C = 0 = v∆t ∆x . This would imply that ∆t = 0 which means there is no feasible value of the
time step for which the FTCS scheme remains stable. We have shown that the FTCS scheme is
unconditionally unstable when applied to the advection equation (note that this is not a general result
and the FTCS scheme may be stable for other PDEs).

†
I have skipped over some explanations here for brevity. Readers seeking further explanation are advised to read
Hirsch (1994).

Dr Jon Shiach 65 Finite-Difference Methods

Chapter 4. Hyperbolic Partial Differential Equations

von Neumann stability analysis

In summary, the von Neumann stability analysis is carried out using the following steps:

1. Replace each instance of unj in the FDS with its corresponding Fourier mode, e.g.,
unj = eat eikm x .

2. Rearrange to obtain an expression for the growth factor G = ea∆t .

3. The condition G ≤ 1 to determine for what values of C is the scheme stable. This can
be a tricky stage requiring use of trigonometric identities and/or properties of moduli. If
equation (4.10) cannot be satisfied for C > 0 then the FDS is unconditionally unstable;
else if it is satisfied for certain values of C then the FDS is conditionally stable; and if it
is satisfied for all values of C then the FDS is unconditionally stable.

Note that it is not recommended in practice to use the largest value of C for which a scheme remains
stable in a computer program. This is because small errors due to the limit of machine precision for
floating point operations can cause the calculation of C to exceed the limit for stability. For example,
if a particular scheme is stable for C ≤ 1 then we could multiply the maximum value of ∆t for which
a FDS is stable by a safety factor of 0.9 to ensure that we always have a stable scheme.
Example 10

∆x
Use von-Neumann stability analysis to show that the FOU scheme is stable for ∆t ≤ .
|v|

Solution: The FOU scheme is

v∆t n
un+1 = unj − (u − unj−1 ),
j
∆x j
v∆t
substituting the Courant number C = ∆x
un+1
j = (1 − C)unj + Cunj−1 .
Substituting in the following Fourier modes
unj = eat eikm ∆x ,
un+1
j = ea(t+∆t) eikm x ,
unj−1 = eat eikm (x−∆x) ,
then we have
ea(t+∆t) eikm x = (1 − C)eat eikm ∆x − Ceat eikm (x−∆x) .
Dividing throughout by eat eikm x gives
G = (1 − C) + Ce−ikm ∆x .
For stability we require |G| ≤ 1 so
1 ≥ |(1 − C) + Ce−ikm ∆x |
≥ |1 − C| + C|e−ikm ∆x |.
Since |eiθ | = 1 by Euler’s identity then
1 ≥ |1 − C| + C.
This is satisfied for C ≤ 1 so the maximum allowable time step for the FOU scheme is
∆x
∆t ≤ .
|v|

Finite-Difference Methods 66 Dr Jon Shiach

 Chapter 4. Hyperbolic Partial Differential Equations

4.6 The Lax-Friedrichs scheme

The Lax-Friedrichs scheme named after American mathematician Peter Lax (1926–1969) and Kurt
Otto Friedrichs is a FDS that attempts to overcome the instabilities in the FTCS scheme by artificially
dampening the oscillations that cause the scheme to become unstable. The uni term in the forward
difference for the time derivative in the FTCS scheme is replaced by an average of the two adjacent
nodes, i.e.
uni−1 + uni+1
un+1
i − un − uni−1
2 + v i+1 =0
∆t 2∆x
which results in the following FDS
The Lax-Friedrichs scheme
1 v∆t n
un+1 = (uni−1 + uni+1 ) − (u − uni−1 ) + O(∆t, ∆x). (4.11)
i
2 2∆x i+1

The Lax-Friedrichs scheme is a first-order in space and time with the stencil given in figure 4.9 (note
that even though a central difference was used to approximate Ux , the Lax-Friedrichs scheme is still
only first-order accurate in space). Use of von Neumann stability analysis shows that the Lax-Friedrichs
scheme is stable for C ≤ 1.

uni+1

uni−1 uni uni+1

Figure 4.9: Finite-difference stencil for the Lax-Friedrichs scheme.

The Lax-Friedrichs solution to the test problem is shown in figure 4.10. Here the solution has dissipated
even more than the FOU scheme due the averaging of the uni term used to remove the oscillations in
the FTCS scheme.

Figure 4.10: The numerical solution of the advection equation using the Lax-Friedrichs
scheme compared to the exact solution.

Dr Jon Shiach 67 Finite-Difference Methods

Chapter 4. Hyperbolic Partial Differential Equations

4.7 The Leapfrog scheme

The Leapfrog scheme is a second-order scheme in time and space derived by approximating both the
time and space derivatives using second-order accurate central differences i.e.,

un+1 − un−1 un − uni−1

i i
+ v i+1 = 0,
2∆t 2∆x

which is rearranged to give

The Leapfrog scheme

v∆t n
un+1 = un−1 − (u − uni−1 ) + O(∆t2 , ∆x2 ). (4.12)
i i
∆x i+1

The stencil for the Leapfrog scheme is shown in figure 4.11. Note that it is the node at the previous
time level un−1
i that is updated to calculate un+1
i and not uni so we say that uni has been ‘leapfrogged’
hence the name of the scheme. Using von Neumann stability analysis it can be shown that the Leapfrog
scheme is stable when C ≤ 1.

un+1
i

uni−1 uni uni+1

un−1
i

Figure 4.11: Finite-difference stencil for the Leapfrog scheme.

We have a problem with the Leapfrog scheme in that attempting to calculate the solution at the first
time level, u1i , we need values from a time level prior to the initial conditions, u−1
i , which do not exist.
A solution to this problem is to solve the first time step using a first-order in time scheme which only
requires the values at one time level u0i , and then continue with the Leapfrog scheme for all other time
levels. The FTCS scheme can be used to calculate u1i , even though it is unconditionally unstable since
we are only using it for one step.

The solution to the test problem using the Leapfrog scheme is shown in figure 4.12. Here the numerical
solution quite closely replicates the exact solution and the dissipation that was evident for the first-
order in time schemes is not present. However, the Leapfrog scheme is notoriously prone to oscillations
in the solution for problems with rapidly varying profiles (the Gaussian curve used here has a relatively
gentle change in the gradient). Also, the calculation of the Leapfrog scheme requires the values of
nodes at two different time steps which requires a computer program to keep track of two solution
arrays, un−1 and un , as opposed to just one.

Finite-Difference Methods 68 Dr Jon Shiach

 Chapter 4. Hyperbolic Partial Differential Equations

Figure 4.12: The numerical solution of the advection equation using the Leapfrog scheme
compared to the exact solution.

4.8 The Lax-Wendroff scheme

The Lax-Wendroff scheme named after Peter Lax and American mathematician Burton Wendroff
(1930–present) is an extension of the Lax-Friedrichs scheme but uses a second-order discretisation of
the time derivative (Lax and Wendroff, 1960). This is done by calculating the Lax-Friedrichs scheme
over a half time step for “dummy” nodes that are assumed to be located halfway between two finite-
difference nodes. The solution over a full time step is then calculated using the Leapfrog scheme of a
full time step using the half time step solution, therefore the full solution is second-order in time and
space.
The Lax-Wendroff scheme can be written as a two-step process, the first step (called a predictor step)
uses a Lax-Friedrichs scheme to solve over a half time step
n+ 1 1 n
v∆t n
ui+ 12 = ui + uni+1 − (u − uni ), (4.13)
2 2 2∆x i+1
n+ 1 ∆x
where the ui+ 12 refers to the dummy node placed a half spatial step from node uni , i.e., x + 2 . The
2
second step (often called a corrector step) uses the Leapfrog scheme to update the solution over a full
time step (figure 4.13)  
v∆t n+ 12 n+ 12
n+1
ui = ui −
n
ui+ 1 − ui− 1 . (4.14)
∆x 2 2

un+1
i

n+ 1 n+ 1
ui− 12 ui+ 12
2 2

uni−1 uni uni+1

Figure 4.13: Finite-difference stencil for the two-step Lax-Wendroff scheme.

Equations (4.13) and (4.14) can be combined to give a single expression for the Lax-Wendroff scheme

Dr Jon Shiach 69 Finite-Difference Methods

Chapter 4. Hyperbolic Partial Differential Equations

 
v∆t n+ 21 n+ 21
un+1
i = uni − ui+ 1 − ui− 1
∆x 2 2
 
v∆t 1 n v∆t n 1 n v∆t n
= uni − (u + ui+1 ) −
n
(u − ui ) − (ui−1 + ui ) +
n n
(u − ui−1 )
n
∆x 2 i 2∆x i+1 2 2∆x i
v∆t n v 2 ∆t2 n
= uni − (ui + uni+1 − uni−1 − uni ) + (u − uni − uni + uni−1 )
2∆x 2∆x2 i+1
v∆t n
v 2 ∆t2 n

= uni − ui+1 − uni−1 + 2
ui−1 − 2uni + uni+1 .
2∆x 2∆x

Analysis of the consistency of the Lax-Wendroff scheme shows that it is second-order in time and
space

The Lax-Wendroff scheme

v∆t n
v 2 ∆t2 n

un+1
i = uni − ui+1 − uni−1 + 2
ui−1 − 2uni + uni+1 + O(∆t2 , ∆x2 ). (4.15)
2∆x 2∆x

The stencil for the Lax-Wendroff scheme is shown in figure 4.14. The solution to the test problem
using the Lax-Wendroff scheme is shown in figure 4.15. The numerical solution closely approximates
the exact solution and there is no numerical dissipation of the curve.

un+1
i

uni−1 uni uni+1

Figure 4.14: Finite-difference stencil for the single step Lax-Wendroff scheme.

Figure 4.15: The numerical solution of the advection equation using the Lax-Wendroff
scheme compared to the exact solution.

Finite-Difference Methods 70 Dr Jon Shiach

 Chapter 4. Hyperbolic Partial Differential Equations

4.9 The Crank-Nicolson scheme

The Crank-Nicolson scheme named after English mathematicians John Crank (1916–2006) and Phyllis
Nicolson (1917–1968) uses a combination of two first-order in time methods to produce a method that
is second-order in time (Crank and Nicolson, 1947). The first-order methods used are the FTCS
scheme and the Backwards-Time, Central-Space (BTCS) scheme. The BTCS scheme is the same as
the FTCS scheme except the spatial differencing uses values from the (n + 1) time level

un+1 − ui un − uni−1
i
+ v i+1 = 0, (FTCS), (4.16)
∆t 2∆x
un+1 − ui un+1 − un+1
i
+ v i+1 i−1
= 0, (BTCS). (4.17)
∆t 2∆x
The Crank-Nicolson scheme uses the average of equations (4.16) and (4.17)
!
v uni+1 − uni−1 ui+1 − ui−1
n+1 n+1
un+1 − uni
i
=− + ,
∆t 2 2∆x 2∆x

which is rearranged to give

The Crank-Nicolson scheme
v∆t n

un+1
i = uni − ui+1 − ui−1
n
i+1 − ui−1 + O(∆t , ∆x ).
+ un+1 n+1 2 2
(4.18)
4∆x

Equation (4.18) is an implicit equation where the solution at the next time step un+1
i is dependent on
values of nodes from the current time level, ui±1 and from those at the next time level, un+1
n
i±1 which
v∆t
are unknown (see the stencil in figure 4.16). Let r = 4∆x and moving all un+1 terms to the left-hand
side of equation (4.18) gives

− run+1 n+1
i−1 + ui i+1 = rui−1 + ui − rui+1 .
+ run+1 n n n
(4.19)

un+1
i−1 un+1
i
un+1
i+1

uni−1 uni uni+1

Figure 4.16: Finite-difference stencil for the Crank-Nicolson scheme.

The linear system in equation (4.19) can be solved to determine the values of the nodes within the
domain, un+1
i where i = 1 . . . N − 1. The nodes on the boundaries, un+1
0 and un+1
N , need to take
into account the boundary condition that is applied at that boundary. For example, consider the
calculation of equation (4.19) for the node on the left-hand boundary

− run+1 n+1
−1 + u0 + run+1
1 = run−1 + un0 − run1 . (4.20)

If zero gradient boundary conditions are used then un+1 n n

−1 = u−1 = u1 so

−run1 + un+1
0 + run+1
1 = run1 + un0 − run1
∴ un+1
0 + run+1
1 = un0 + run1 .

Doing similar for the right-hand boundary gives

−run+1
N −2 + ruN −1 = −ruN −2 + uN −1 .
n+1 n n

Dr Jon Shiach 71 Finite-Difference Methods

Chapter 4. Hyperbolic Partial Differential Equations

The solution of the advection equation using Crank-Nicolson scheme given in equation (4.19) with
transmissive boundary conditions at each end can be written as the following tridiagonal matrix
equation
     
1 r un+1
0 1 r un0
   n+1    n 
−r 1 r   u1  r 1 −r   u1 
     
 .. .. ..   .     .. 
 . . .   ..  =  . . . . . . . . .  . .
     
 −r 1 r   n+1   1 −r   n 
  uN −2   r  uN −2 
−r 1 un+1
N −1 −r 1 unN −1

This linear system can be solved using the Thomas algorithm discussed in section 3.10.1 on page 47
to calculate the values of un+1 . Performing a von Neumann stability analysis shows that the Crank-
Nicolson scheme is unconditionally stable for all values of ∆t. This means if we wanted to we could
use a single step to update the solution to any value of t. However, since the approximation of the
time derivative is still second-order, choosing a high value for ∆t will increase errors even though the
scheme remains stable.
The Crank-Nicolson scheme was used to model the propagation of the Gaussian curve is shown in
figure 4.17.

Figure 4.17: The numerical solution of the advection equation using the Crank-Nicolson
scheme compared to the exact solution.

4.10 Analysis of finite-difference schemes used to solve the advection

equation
A number of finite-difference schemes have been presented in this chapter and applied to solve the
advection equation for a Gaussian curve initial conditions. The advection equation is a very simple
example of a hyperbolic PDE but has little use when modelling physical phenomena. However, it
is useful to know how FDS perform when applied to the advection equation since if a FDS gives a
poor approximation of the exact solution, then it is unlikely to perform well for more complicated
hyperbolic PDEs.
One of the features of hyperbolic PDEs is that the solutions can admit discontinuities, therefore in
order for a numerical scheme to be applicable it needs to be able to handle these discontinuities in
addition to very step gradients. The Gaussian curve test used to examine the FDS is an easy test for

Finite-Difference Methods 72 Dr Jon Shiach

 Chapter 4. Hyperbolic Partial Differential Equations

a numerical scheme since there is no abrupt changes in the gradient. A tougher test of a numerical
scheme is the propagation of a ‘top hat’ function defined using the following case statement
(
1, |x − vt − 0.25| ≤ 0.1,
U (t, x) = (4.21)
0, otherwise.

The FDS presented in this chapter with that exception of the FTCS scheme have all been applied to
solve this problem and the solutions are shown in figure 4.18.
The first-order in time schemes both exhibit numerical dissipation but it is more pronounced in the
Lax-Friedrichs scheme (figures 4.18(a) and 4.18(b)). This is a common feature of first-order methods.
The second-order in time schemes all exhibit over/undershoots where the profile makes an abrupt
change in the gradient. This is known as Gibbs phenomenon and is characteristic of second-order
methods. The Leapfrog and Crank-Nicolson schemes exhibit severe oscillations in the upwind region
behind the tophat (figures 4.18(c) and 4.18(e)) whereas the Lax-Wendroff scheme (figure 4.18(d))
does not have these oscillations and is clearly the best performing of all of the methods seen here. A
summary of the FDS described in this chapters is given in table 4.1.

Table 4.1: Summary of the finite-difference schemes presented in this chapter.

Scheme O(∆tn ) O(∆xn ) Stability Comments

FOU 1st 1st C≤1 Dissipative
FTCS 1st 2nd unconditionally Cannot be used to solve advection equation
unstable
Lax-Friedrichs 1st 1st C≤1 Dissipative
Leapfrog 2nd 2nd C≤1 Oscillations in upwind region when applied
to discontinuities
Lax-Wendroff 2nd 2nd C≤1 Over/undershoots at discontinuities
Crank-Nicolson 2nd 2nd unconditionally Implicit method. Oscillations in upwind re-
stable gion when applied to discontinuities

Dr Jon Shiach 73 Finite-Difference Methods

Chapter 4. Hyperbolic Partial Differential Equations

(a) FOU (b) Lax-Friedrichs

(c) Leapfrog (d) Lax-Wendroff

(e) Crank-Nicolson

Figure 4.18: The solutions of the various FDS used to model the propagation of a ‘top hat’
profile

Finite-Difference Methods 74 Dr Jon Shiach

Chapter 5

Parabolic Partial Differential Equations

Parabolic PDEs combine second-order spatial derivatives seen in elliptic PDEs with first-order deriva-
tives in time and space that model propagation in hyperbolic PDEs. Parabolic PDEs are most com-
monly used to model heat diffusion and acoustic problems. This chapter uses the heat diffusion
equation to demonstrate the solution of one and two-dimensional parabolic PDEs in addition to per-
forming a grid refinement study.

5.1 Heat diffusion equation

The heat diffusion equation is the simplest parabolic PDE and models the diffusion of heat across the
domain. The one-dimensional form of the heat diffusion equation is

The heat diffusion equation

Ut − αUxx = 0, (5.1)

where α is the thermal diffusivity of the material defined by

k
α= , (5.2)
ρcp

and k is the thermal conductivity of the material, ρ is the density and cp is the specific heat capacity.

5.1.1 One-dimension test problem

The use of equation (5.1) can be demonstrated by the modelling of the heat diffusion along a metal
bar. Consider figure 5.1 that shows a bar of length L that is perfectly insulated along its length so
that no heat is transferred in the radial direction. The temperature of at each end of the bar set at a
constant 0◦ C.

insulation
U (t, 0) = 0 U (t, L) = 0

x=0 x=L
insulation

Figure 5.1: Heat diffusion along a metal bar.

75
Chapter 5. Parabolic Partial Differential Equations

The initial temperature profile across the length of the bar is given by the function
 πx 
U (0, x) = sin . (5.3)
L

It can be shown that the exact solution to this problem with initial conditions given by equation (5.3)
is
 πx   
απ 2 t
U (x, t) = sin exp − 2 . (5.4)
L L

5.1.2 Solving the one-dimensional heat diffusion equation

Equation (5.1) can be solved by applying finite-difference approximations from the finite-difference
toolkit (table 2.4 on page 23) and rearranged to give expressions for the updated approximation uni
similar to what was done for the advection equation in chapter 4. For example, using a forward
difference in time and a symmetric difference in space gives

un+1 − uni un − 2uni + uni+1

i
− α i−1 = 0,
∆t ∆x2

therefore

The FTCS for the heat diffusion equation

α∆t n
un+1 = uni + (u − 2uni + uni+1 ) + O(∆t, ∆x2 ). (5.5)
i
∆x2 i−1

This is the FTCS scheme for the heat diffusion equation. In order to determine the value for ∆t for
α∆t
which equation (5.5) is stable we need to conduct a von Neumann stability analysis. Let r = ∆x2 then
equation (5.5) can be written in compact form as

un+1
i = (1 − 2r)uni + r(uni−1 + uni+1 ).

Substituting in the Fourier modes and rearranging gives (see section 4.5.3)

G = (1 − 2r) + 2r cos(km ∆x).

For stability we require |G| ≤ 1

1 ≥ |1 + 2r(cos(km ∆x) − 1)|,

which is satisfied when r ≤ 1

2 so the FTCS scheme is stable when

∆x2
∆t ≤ . (5.6)
2α

Note that unlike the advection equation where the FTCS scheme was unconditionally unstable, the
FTCS scheme is conditionally stable for the heat diffusion equation.

The test problem has been solved using the FTCS scheme. The length of the bar was set at L = 1
and thermal diffusivity α = 1 was used. The domain was discretised using 11 nodes so that N = 10
and ∆x = 0.1. Dirichlet boundary conditions U (t, 0) = U (t, 1) = 0 were employed at both ends of
the domain. The FTCS scheme was iterated until t = 0.05 using equation (5.6) to calculate ∆t. The
solution is plotted against the initial profile and exact solution in figure 5.2.

Finite-Difference Methods 76 Dr Jon Shiach

 Chapter 5. Parabolic Partial Differential Equations

Figure 5.2: The solutions to the one-dimensional heat diffusion equation using the FTCS
scheme with N = 10 at t = 0.05.

The temperature profile along the bar remains centred as the scheme is marched forward in time.
This is because the initial profile was symmetrical and the boundary conditions are the same at either
end of the domain. Dirichlet boundary conditions ensure the temperature at the boundary nodes
remains zero and this has a cooling effect on the nodes in the domain. The FTCS scheme gives very
good agreement with the exact solution where the absolute L2 error between the exact and computed
solutions is 0.0095.

5.2 Two dimensional heat diffusion equation

The one-dimensional heat diffusion equation equation (5.1) can be extended into two spatial dimensions
by the inclusions of double derivative terms for the y variable, i.e.,

Two-dimensional heat diffusion equation

Ut − α(Uxx + Uyy ) = 0. (5.7)

The process of solving equation (5.7) is similar to the one-dimensional case. If a FTCS scheme is
employed then using a forward difference in time and symmetric differences in both spatial directions
then
 n 
un+1 − uni ui−1,j − 2uni,j + uni+1,j uni,j−1 − 2uni,j + uni,j+1
i
−α + =0
∆t ∆x2 ∆y 2

which gives

FTCS scheme for the two-dimensional heat diffusion equation

α∆t n α∆t n
un+1 n
i,j = ui,j + (u − 2uni,j + uni+1,j ) + (u − 2uni,j + uni,j+1 ). (5.8)
∆x2 i−1,j ∆y 2 i,j−1

Performing a von Neumann stability analysis on equation (5.8) gives

2α∆t 2α∆t
G=1− 2
[1 − cos(km ∆x)] − [1 − cos(ℓm ∆y)]
∆x ∆y 2

Dr Jon Shiach 77 Finite-Difference Methods

Chapter 5. Parabolic Partial Differential Equations

where ℓm is the wave number in the y direction. For stability we require |G| ≤ 1 and G is at a
maximum when km ∆x = ℓm ∆y = π which gives 1 − cos(π) = 2 so
  
 
1 − 4 α∆t + α∆t  ≤ 1.
 ∆x2 ∆y 2 
This is satisfied when the term in brackets is less than or equal to 12 , therefore
α∆t α∆t 1
2
+ 2
≤
∆x ∆y 2
∆x2 ∆y 2
∆t ≤ .
2α(∆x2 + ∆y 2 )
Note that if ∆x = ∆y then
∆x2
∆t ≤
.
4α
Comparing this expression to equation (5.6) we see that the two-dimensional scheme requires a time
step half of that of the one-dimensional scheme.

5.2.1 Modelling heat diffusion across a metal sheet

The FTCS scheme for the two-dimensional heat diffusion equation is applied to model the temperature
across a metal sheet. The sheet is a unit square that is discretised using 51 nodes for both the x and y
domains giving ∆x = ∆y = 0.02. The initial temperature across the domain is zero with the exception
of a C-shaped region which has a temperature of 1 (figure 5.3).

y
Uy (t, x, 1) = 0

1
U (0, x, y) = 1
0.8

0.6
Ux (t, 0, y) = 0 Ux (t, 1, y) = 0
0.4

0.2
U (0, x, y) = 0

0 x
0 0.2 0.4 0.8 1
Uy (t, x, 0) = 0

Figure 5.3: Modelling heat diffusion across a metal sheet: initial conditions.

Zero gradient Neumann boundary conditions using ghost nodes are implemented at all four boundaries,
i.e.,
un−1,j = un1,j , unNX ,j = unNX −2,j , uni,−1 = ui,1 , uni,NY = uni,NY −2 .
This means that the temperature gradient across the boundaries is zero so that the metal sheet is
perfectly insulated (i.e., the temperature outside of the domain has no effect on the temperature within
the domain).
The solutions for this problem at times t = 0.00, 0.25, 0.75, 1.00 and 1.50 have been plotted and
shown in figure 5.4. As the solution is marched forward in time the area surrounding the C-shape
begins to heat up whilst C-shape itself begins to cool. This diffusion continues until the initial shape
is no longer recognisable.

Finite-Difference Methods 78 Dr Jon Shiach

 Chapter 5. Parabolic Partial Differential Equations

(a) t = 0.1 (b) t = 0.25

(c) t = 0.50 (d) t = 0.75

(e) t = 1.00 (f) t = 1.50

Figure 5.4: Solutions to the two-dimension heat diffusion problem.

Dr Jon Shiach 79 Finite-Difference Methods

Chapter 5. Parabolic Partial Differential Equations

5.3 Grid refinement study

It was discussed previously in section 4.5 that a convergent scheme is one where the error between the
numerical approximations to a PDE and the exact solution of a PDE converge to zero as ∆t, ∆x → 0.
So choosing a smaller value of ∆x will give a better approximation to the exact solution, but this
improvement in the accuracy comes at an additional computational cost since we are calculating the
solution using more grid points and we will need to use a smaller value of ∆t for stability. In practice
there is a compromise between accuracy and computational speed and we need to ask at what point
is it no longer worthwhile reducing ∆x to gain a more accurate result? This is the purpose behind a
grid refinement study.
A grid refinement study is performed using the following steps:
1. The PDE is solved using a finite-difference grid with the fewest number of nodes that can be
reasonably expected to capture the solution.
2. The solution is recalculated using a finer finite-difference grid with more nodes than before.
3. Steps 1 and 2 are repeated increasing the number of nodes used in the finite-difference grid.
Metrics used to analyse the performance of the model are compared and a decision is made to
the optimum number of nodes.
The metrics used to analyse the performance of the numerical model will depend upon the phenomenon
begin modelled, but in most cases will involve the accuracy of the numerical solution and the time
taken to compute this solution.

5.3.1 Grid convergence index

In the majority of cases there are no known exact solutions for a PDE so we cannot compute the error
between the numerical solution and the exact solution. What we can do is calculate the difference
between the solutions using different sized grids and see whether the solution is converging at a rate
expected for the order of the method. To do this we calculate the Grid Convergence Index (GCI)
which is a measure of the convergence to an estimate of the exact solution between two different sized
grids.
Let E be the error between the exact solution fexact and the numerical solution using grid spacing h,
f (h) such that
E = f (h) − fexact = Chp , (5.9)
where p is the order of convergence of the numerical method and C is some constant. In practice
factors such as boundary conditions, grid spacings and sink and source terms mean that the actual
order of convergence is less than that of the theoretical order of convergence. So we need to calculate
the observed order of convergence for the numerical method.
The observed order of convergence can be calculated using equation (5.9) for three different grid sizes.
Lets say we have a fine, medium and coarse grid with grid spacing h1 , h2 and h3 respectively where r
is the grid refinement ratio such that
h2 h3
r= = ,
h1 h2
writing out equation (5.9) for each of these grids and solving for p gives
 
f3 − f2
ln
f2 − f1
p= ,
ln(r)
where fi is the numerical solution using grid spacing hi . Applying Richardson extrapolation using f1
and f2 gives
f1 − f2
fexact ≈ f1 + p . (5.10)
r −1

Finite-Difference Methods 80 Dr Jon Shiach

 Chapter 5. Parabolic Partial Differential Equations

The second term on the right-hand side of equation (5.10) is the fractional error for f1 . Calculating
the percentage error between f1 and fexact gives
   
fexact − f1 f1 − f2 1 ε21
100 ≈ 100 = p ,
f1 r − 1 f1
p r −1

where ε21 is the percentage relative error between the grids using h1 and h2 , i.e.,
 
f2 − f1
ε21 = 100 .
f1

The GCI gives an estimate of the discretisation error in the fine grid solution (f1 ) relative to the
extrapolated solution fexact . Roache (1994) defines the GCI index calculated using the fine and
medium grids as
1.25|ε21 |
GCI21 = p . (5.11)
r −1
and similar for the medium and coarse grids GCI32 . A solution is said to be in the asymptotic range
of convergence if
rp GCI21 ≈ GCI32 . (5.12)
If equation (5.12) is satisfied then the solution using the h2 grid has achieved grid convergence and no
further refinement of the grid will provide more accurate results.
To perform a grid refinement study using the GCI we do the following (Roache, 1994):

Grid refinement study

1. Calculate the solution of the PDE using three different finite-difference grids with grid
spacings h1 (fine grid), h2 (medium grid) and h3 (coarse grid) where h1 < h2 < h3 and
h2 h3
r= = .
h1 h2
2. For a sample of points in the grids calculate the observed order of convergence using
  
  f3 − f2 
ln  
  f2 − f1 
p= , (5.13)
ln(r)

where f1 , f2 and f3 are the solutions at the points for grid spacings h1 , h2 and h3
respectively. The use of the moduli in equation (5.13) is so we can deal with oscillatory
convergence.

3. Calculate the grid convergence index for the fine and medium grid solutions and the
medium and coarse grid solutions.
     
 f − f   f − f 
1.25 100  1.25 100 
2 1 3 2
f1  f2 
GCI21 = , GCI 32 = .
rp − 1 rp − 1
4. If rp GCI21 ≈ GCI32 then the medium grid solution has reached asymptotic convergence.
If not, refine the grid further and repeat steps 1–3.

The FDS described previously to solve the heat diffusion equation has been applied using square
finite-difference grids with N3 = 10, N2 = 20 and N1 = 40 nodes such that r = 2 and h3 = 0.1,
h2 = 0.05 and h1 = 0.025. The GCI was calculated for a sample of four points from the domain
located at (0.3, 0.3), (0.5, 0.3), (0.7, 0.3) and (0.3, 0.5) and these are shown in the table below. Note
that for the point at (0.7,0.3), GCI32 ≪ rp GCI21 which indicates that the h2 grid has not reached
asymptotic convergence.

Dr Jon Shiach 81 Finite-Difference Methods

Chapter 5. Parabolic Partial Differential Equations

(x, y) fexact f1 f2 f3 p GCI21 GCI32 rp GCI21

(0.3,0.3) 0.6210 0.7221 0.7564 0.8022 0.42 17.50 22.37 23.43
(0.5,0.3) 0.7073 0.7045 0.6966 0.6658 1.95 0.49 1.92 1.90
(0.7,0.3) 0.5893 0.5932 0.6050 0.6079 2.00 0.83 0.20 3.30
(0.3,0.5) 0.7315 0.7279 0.7178 0.6797 1.91 0.63 2.40 2.36

The finite-difference grid was refined further so that the fine, medium and coarse grids use N1 = 80,
N2 = 40 and N1 = 20. The GCI for the same sample of points are shown in the table below. Here, all
of the points sampled satisfy equation (5.12) (or close enough) which suggests that the discretisation
using N = 40 nodes is sufficient in this case.

(x, y) fexact f1 f2 f3 p GCI21 GCI32 rp GCI21

(0.3,0.3) 0.6792 0.7030 0.7221 0.7564 0.85 4.24 7.43 7.63
(0.5,0.3) 0.7070 0.7064 0.7045 0.6966 2.05 0.11 0.45 0.45
(0.7,0.3) 0.5675 0.5852 0.5932 0.6050 0.54 3.76 5.41 5.48
(0.3,0.5) 0.7310 0.7302 0.7279 0.7178 2.09 0.12 0.53 0.53

(a) N = 10 (b) N = 20

(c) N = 40 (d) N = 80

Figure 5.5: Solutions to the heat diffusional example at t = 0.5 using four different grid
sizes.

Finite-Difference Methods 82 Dr Jon Shiach

 Chapter 5. Parabolic Partial Differential Equations

Example 11

A numerical solver was used to calculate the solution to a particular PDE using five different
grid sizes. The values of the spatial step used and solutions at two points in the finite-difference
grid are shown in the table below.
h point 1 point 2
0.1520 0.5549 0.5432
0.0760 0.5533 0.5577
0.0380 0.5462 0.5668
0.0190 0.5513 0.5719
0.0095 0.5490 0.5693

Use the grid convergence index to determine the optimum grid spacing that provides the best
compromise between accuracy and computational time.

Solution: Considering the first three grids we have h1 = 0.038, h2 = 0.076, h3 = 0.152 and

point 1: f3 = 0.5549, f2 = 0.5533, f1 = 0.5462,

point 2: f3 = 0.5432, f2 = 0.5577, f1 = 0.5668.

Calculate the ratio of successive step lengths:

h2 0.038
r= = = 2.
h1 0.076
Calculate the observed order of convergence:
  
  0.5549 − 0.5533 
ln  
  0.5533 − 0.05462  | ln(0.2254)|
point 1: p= = = 2.1497,
ln(2) ln(2)
  
  0.5432 − 0.5577 
ln  
  0.5577 − 0.5668  | ln(1.5934)|
point 2: p= = = 0.6721.
ln(2) ln(2)

Calculate GCIs:
  
 0.5533 − 0.5462 
1.25 100 
0.5462
point 1 : GCI21 = = 0.4727,
 22.1497 − 1
 
 0.5549 − 0.5533 
1.25 100 
0.5533
GCI32 = = 0.1052,
 22.1497 − 1
 
 0.5577 − 0.5668 
1.25 100 
0.5668
point 2 : GCI21 = 0.6721 − 1
= 3.3820,
 2
 
 0.5432 − 0.5577 

1.25 100
0.5577 
GCI32 = = 5.4768.
20.6721 − 1
Checking for asymptotic convergence:

point 1: rp GCI21 = 22.1497 (0.4727) = 2.0976 ̸= 0.1052 = GCI32 ,

point 2: rp GCI21 = 20.6721 (3.3820) = 5.3889 ≈ 5.4768 = GCI32 ,

Dr Jon Shiach 83 Finite-Difference Methods

Chapter 5. Parabolic Partial Differential Equations

so for the solution using h = h2 = 0.076 point 2 has reached asymptotic convergence point 1 has
not. Therefore we need to use a finer grid, considering the next three grids where h1 = 0.019,
h2 = 0.038, h3 = 0.076 and

point 1 : p = 0.4473, GCI21 = 2.9487, GCI32 = 4.1434,

point 2 : p = 0.8354, GCI21 = 1.4212, GCI32 = 2.5588.

Repeating the calculations for these values gives

point 1 : p = 0.4473, GC121 = 2.9487, GCI32 = 4.1434,

point 2 : p = 0.8354, GCI21 = 1.4212, GCI32 = 2.5588.

Checking for asymptotic convergence:

point 1: rp GCI21 = 20.4473 (2.9487) = 4.1051 ≈ 4.1434 = GCI32 ,

point 2: rp GCI21 = 20.8354 (1.4212) = 2.5359 ≈ 2.5588 = GCI32 .

So for the solution using h = h2 = 0.038 both sample points have reached asymptotic conver-
gence so this is the optimum grid spacing.

Finite-Difference Methods 84 Dr Jon Shiach

Chapter 6

Solving Multidimensional Partial

Differential Equations

For most real world practical applications we require methods that can model systems over two or
three-dimensional spatial domains. To simplify our numerical models we can split the problem up
so that we solve for multiple models each with single spatial dimension. This method is known as
dimensional splitting and can also be used for PDEs including multiple terms.

6.1 The two-dimensional advection equation

Consider the two-dimensional form of the advection equation seen in equation (4.1) on page 53

Two-dimensional advection equation

Ut + vUx + wUy = 0, (6.1)

where v and w are the velocities in the x and y directions respectively. Equation (6.1) models pure
advection so the shape of the initial profile is unchanged as it propagates though the spatial domain.
If the initial conditions are defined by

U (0, x, y) = f (x, y),

then the exact solution is

U (t, x, y) = f (x − vt, y − wt).

6.1.1 Deriving an unsplit FDS for the two-dimensional advection equation

The derivation of an unsplit FDS to solve the two-dimensional advection equation proceeds in the
same way as seen previously, i.e., we approximate the partial derivatives using finite-differences from
the finite-difference toolkit (table 2.4 on page 23). Recall that when v > 0 the FOU scheme uses a
backwards difference to discretise the spatial derivative, assuming that v, w > 0 then

i,j − ui,j
un+1 uni,j − uni−1,j uni,j − uni,j−1
n
+v +w = 0,
∆t ∆x ∆y
which is rearranged to give
Unsplit FOU scheme for the two-dimensional advection equation

v∆t n w∆t n
i,j = ui,j −
un+1 (u − uni−1,j ) − (u − uni,j−1 ) + O(∆t, ∆x, ∆y).
n
(6.2)
∆x i,j ∆y i,j

85
Chapter 6. Solving Multidimensional Partial Differential Equations

Equation (6.2) is the FOU scheme for the two-dimensional advection equation which is first-order in
time and space. To test this scheme the advection of a two-dimensional Gaussian curve is used. The
domain is defined by the range 0 ≤ x, y ≤ 1 which is discretised using a finite-difference grid with
spatial steps ∆x = ∆y = 0.02. Dirichlet boundary conditions are used at all four boundaries where
the boundary nodes remain a constant zero. The initial profile is defined by

U (0, x, y) = exp[−200((x − 0.25)2 + (y − 0.25)2 )]

and the velocities are v = w = 1. The initial profile the numerical solution at t = 0.5 is shown in
figure 6.1. The numerical scheme exhibits dissipation perpendicular to the main direction of flow given
by the vector v = vi + wj.

(a) t = 0.0 (b) t = 0.5

Figure 6.1: Solutions of the two-dimensional advection equation using the FOU scheme.

Performing a von Neumann stability analysis on the FDS in equation (6.2) shows that it is stable
when

v∆t w∆t
+ ≤ 1,
∆x ∆y

so the maximum allowable time step is

∆x∆y
∆t = . (6.3)
v∆y + w∆x

This is a very restrictive constraint on the value of ∆t and the analysis requires difficult algebra.
In general, unsplit schemes are unwieldy, have restrictive stability constraints, are computationally
expensive and are rarely used for parabolic and hyperbolic PDEs.

6.2 Operator splitting

Instead of deriving a single scheme to solve a multi-dimensional PDE which can be complicated,
we can derive simpler schemes to solve PDEs that are related to the original PDE that we want to
solve. This method is called operator splitting since we are splitting a complicated problem up into
several simpler problems. The schemes used for the related PDEs are applied in sequence so that
the completed sequence is equivalent to the unsplit scheme. The manner of splitting can be done by
spatial dimensions, components of the PDE or a combination of both.

Finite-Difference Methods 86 Dr Jon Shiach

 Chapter 6. Solving Multidimensional Partial Differential Equations

6.2.1 Differential marching operator

In order to ensure that the sequence of split schemes is equivalent to the unsplit scheme, we need a
way of comparing the two types. This can be done by considering the differential marching operator
for the schemes. The idea is to write the solution of the PDE at the next time step in the form of
some operator that is applied to the current solution. This is done by expressing the time derivatives
in the Taylor series using spatial derivatives. For example, consider the second-order Taylor series
expansion for one step forward in time

∆t2
U (t + ∆t) = U + ∆tUt + Utt + O(∆t3 ). (6.4)
2
We want to replace the Ut (t, x) and Utt (t, x) terms with spatial derivatives which will depend on the
PDE begin solved. Let ∂t and ∂tt denote be two differential operators for partial derivatives for the t
variable such that ∂t U = Ut and ∂tt U = Utt , equation (6.4) can then be written as

∆t2
U (t + ∆t) = U + ∆t∂t U + ∂tt U + O(∆t3 ) (6.5)
2
We require expressions for ∂t U and ∂tt U which will depend upon the PDE being solved. For example,
consider the one-dimensional advection equation equation (4.1) written using differential operator
notation

∂t U + v∂x U = 0.

Therefore

∂t U = −v∂x U,

and

∂tt U = ∂t (∂t U ) = ∂t (−v∂x U ) = −v∂x (∂t U ) = −v∂x (−v∂x U ) = v 2 ∂xx U.

Substituting ∂t U and ∂tt U into equation (6.5) gives

v 2 ∆t2
U (t + ∆t) = U − v∆t∂x U + ∂xx U + O(∆t3 ),
2
factorising out the U terms results in
 
v 2 ∆t2
U (t + ∆t) = 1 − v∆t∂x + ∂xx U + O(∆t3 ).
2

Let LX (∆t) be the term in the square brackets then

U (t + ∆t) = LX (∆t)U + O(∆t3 ), (6.6)

where
Differential marching operator for the one-dimensional advection equation

v 2 ∆t2
LX (∆t) = 1 − v∆t∂x + ∂xx + O(∆t3 ). (6.7)
2

LX (∆t) is the differential marching operator for the advection equation (an uppercase X has been used
in the subscript so as to avoid confusing it with a partial derivative). The different finite-difference
schemes presented in chapter 4 can be derived by replacing ∂x and ∂xx by different finite-difference
approximations from the toolkit, i.e.,

un+1
i = LX (∆x)uni .

Dr Jon Shiach 87 Finite-Difference Methods

Chapter 6. Solving Multidimensional Partial Differential Equations

6.2.2 Differential marching operator for the two-dimension advection equation

Following a similar process seen previously we can derivate the differential marching operator for
the two-dimensional advection equation. Using differential operator notation, equation (6.1) can be
written as
∂t U + v∂x U + w∂y U = 0,
therefore
∂t U = −v∂x U − w∆y U,
∂tt U = ∂t (∂t U ) = ∂t (−v∂x U − w∂y U ) = −v∂x (∂t U ) − w∂y (∂t U )
= −v∂x (−v∂x U − w∂y U ) − w∂y (−v∂x U − w∂y U )
= v 2 ∂xx U + 2vw∂xy U + w2 ∂yy U.
Substituting ∂t U and ∂tt U into equation (6.5) gives
U (t + ∆t) = U − ∆t[v∂x U + w∂y U ]
∆t2 2
+ [v ∂xx U + 2vw∂xy U + w2 ∂yy U ] + O(∆t3 )
 2 
∆t2 2
= 1 − ∆t(v∂x + w∂y ) + (v ∂xx + 2vw∂xy + w ∂yy ) U + O(∆t3 ).
2
2
So the second-order differential marching operator for the two-dimensional advection equation is
Second-order differential marching operator for the two-dimensional advection
equation

∆t2 2
LXY (∆t) = 1 − ∆t(v∂x + w∂y ) + (v ∂xx + 2vw∂xy + w2 ∂yy ), (6.8)
2

and the FDS can be written as

i,j = LXY (∆t)ui,j .
un+1 n
(6.9)
Note that equation (6.8) assumes that the spatial steps ∆x and ∆y are constant throughout the
domain.

6.3 Dimensional splitting

Dimensional splitting splits a PDE with an n-dimensional spatial domain into a system of n PDEs
each having a single spatial domain. Each of these one-dimensional PDEs can be solved using our
one-dimensional schemes and then combined in sequence to provide the solution to the n-dimensional
PDE. For example, we can split the two-dimensional advection equation equation (6.1) into two one-
dimensional PDEs
Ut + vUx = 0, (6.10a)
Ut + wUy = 0. (6.10b)
The differential marching operator for equation (6.10a) was seen in equation (6.7). Doing similar for
equation (6.10b) we have
w2 ∆t2
LY (∆t) = 1 − w∆t∂y + ∂yy + O(∆t3 ). (6.11)
2

Using dimensional splitting we solve a series of one-dimensional problems in one direction before
solving another series of one-dimensional problems in the other direction(s). We can write this as

i,j = LY (∆t)LX (∆t)ui,j .

un+1 n

Finite-Difference Methods 88 Dr Jon Shiach

 Chapter 6. Solving Multidimensional Partial Differential Equations

In this case we have chosen to solve in the x direction first (since LX (∆t) is next to uni,j ) and then we
use those solutions to solve in the y direction. This is shown pictorally in figure 6.2.

y y

x x
(a) x sweep (b) y sweep

Figure 6.2: Dimensional splitting using the sequence LY (∆t)LX (∆t).

In order for the split scheme to be equivalent to an unsplit scheme we need to show that LY (∆t)LX (∆t) =
LXY (∆t)
  
w2 ∆t v 2 ∆t
LY (∆t)LX (∆t) = 1 − w∆t∂y + ∂yy 1 − v∆t∂x + ∂xx + O(∆t3 )
2 2
∆t2 2
= 1 − ∆t(v∂x + w∂y ) + (v ∂xx + 2vw∂xy + w2 ∂yy ) + O(∆t3 )
2
= LXY (∆t).

So the split scheme is equivalent to the unsplit scheme. Note that if we applied the one-dimensional
scheme in the y direction first we would still get the same result so the order in which the individual
dimensional operators are applied doesn’t matter. The choice of which one-dimensional scheme is used
in the x and y directions will depend on the PDE being solved and the problem that it is applied to.
For example, using the one-dimension FOU for both dimensions gives
v∆t n
u∗i,j = uni,j − (u − uni−1,j ), (6.12a)
∆x i,j
∗ w∆t ∗
un+1
i,j = ui,j − (u − u∗i,j−1 ). (6.12b)
∆y i,j

The application of equation (6.12a) creates new values of the dependent variable at each node denoted
by u∗i,j . equation (6.12b) then uses these new values to update the values of the dependent variable
for one full time step. The choice of the one-dimensional schemes do not have to be the same for each
dimension although accuracy considerations may require that they be of the same order.
For stability, the maximum allowable time step for the split scheme is the smallest of the time steps
for the individual one-dimensional schemes, e.g.,
 
∆x ∆y
∆t ≤ min CX , CY .
|v| |w|

where CX and CY are the Courant numbers that ensure the stability of the schemes used in the x
and y directions respectively. This stability criterion is much less restrictive than that of the unsplit
scheme equation (6.3).
The split FOU scheme has been applied to the same test problem described in section 6.1.1 and the
solution is shown in figure 6.3. Here the numerical scheme exhibits the dissipation that is characteristic
of first-order schemes, however, the anti-dissipation seen in the direction of flow using the unsplit
scheme is no longer present.

Dr Jon Shiach 89 Finite-Difference Methods

Chapter 6. Solving Multidimensional Partial Differential Equations

Figure 6.3: Solution of the two-dimensional advection equation using the split FOU scheme.

6.4 Term splitting

PDEs may contain several terms that model different physical processes. Term splitting can be used
to split up the PDE and treats each term separately. For example, consider the one-dimensional
advection-diffusion equation
The advection-diffusion equation

Ut + vUx − αUxx = 0. (6.13)

This PDE combines the advection equation equation (4.1) with the diffusion equation equation (5.1) to
model transport and diffusion at the same time. Equation (6.13) is a parabolic PDE since it contains a
Uxx term. Assuming v > 0 and using first-order differences for Ut and Ux and second-order difference
for Uxx results in
un+1 − uni un − ui−1 un − 2uni + uni+1
i
+v i − α i−1 = 0,
∆t ∆x ∆x2
which gives the following unsplit scheme
v∆t n α∆t n
un+1 = uni − (ui − uni−1 ) + (u − 2uni + uni+1 ) + O(∆t, ∆x). (6.14)
i
∆x ∆x2 i−1
To perform a von Neumann stability analysis on this scheme we can write it in the form
un+1
i = (C + D)uni−1 + (1 − C − 2D)uni + Duni+1 ,

where C = v∆t ∆x and D = ∆x2 . Substituting the Fourier nodes and applying the condition |G| ≤ 1
α∆t

results in (this bit is difficult)

1 − C − 2D ≥ 0,
so the maximum allowable time step is
∆x2
∆t ≤ . (6.15)
v∆x + 2α
Note that when α = 0 we have the stability constraint for the heat diffusion equation equation (5.6)
and when v = 0 we have the stability constraint for the FOU scheme when applied to the advection
equation equation (4.6).

Finite-Difference Methods 90 Dr Jon Shiach

 Chapter 6. Solving Multidimensional Partial Differential Equations

6.4.1 Differential marching operator for the advection-diffusion equation

To establish whether a term splitting is valid for the advection-diffusion equation we need to compare
the differential marching operators for the unsplit equation against the combined differential marching
operators for the advection and diffusion equations. Following the steps described in section 6.2.1 for
equation (6.13)
∂t U = α∂xx U − v∂x U,
∂tt U = v 2 ∂xx U − 2vα∂xxx U + α2 ∂xxxx U.
Substituting these into the second-order Taylor series gives
∆t2
U (t + ∆t) = U + ∆t∂t U + ∂tt U + O(∆t3 )
2
∆t2 2
= U + ∆t(α∂xx − v∂x )U + (v ∂xx − 2vα∂xxx + α2 ∂xxxx )U + O(∆t3 )
 2 
∆t2 2
= 1 + ∆t(α∂xx − v∂x ) + (v ∂xx − 2vα∂xxx + α ∂xxxx ) U + O(∆t3 ),
2
2
therefore the differential marching operator for the advection-diffusion equation is
∆t2 2
LAD (∆t) = 1 + ∆t(α∂xx − v∂x ) + (v ∂xx − 2vα∂xxx + α2 ∂xxxx ) + O(∆t3 ). (6.16)
2
The differential marching operators for the advection equation and the diffusion equation are
v 2 ∆t2
LA (∆t) = 1 − v∆t∂x + ∂xx + O(∆t3 ),
2
α2 ∆t2
LD (∆t) = 1 + α∆t∂xx + ∂xxxx + O(∆t3 ).
2
Combining these two and truncating terms higher than second-order gives
  
v 2 ∆t2 α2 ∆t2
LA (∆t)LD (∆t) = 1 − v∆t∂x + ∂xx 1 + α∆t∂xx + ∂xxxx + O(∆t3 )
2 2
v 2 ∆t2 α2 ∆t2
= 1 + α∆t∂xx − v∆t∂x − vα∆t2 ∂xxx + ∂xx + ∂xxxx + O(∆t3 )
2 2
∆t2 2
= 1 + ∆t(α∂xx − v∂x ) + (v ∂xx − vα∂xxx + α2 ∂xxxx ) + O(∆t3 )
2
= LAD (∆t).
Therefore the term split scheme is equivalent to the unsplit scheme. For stability we choose a value
of the time step that satisfies both the advection scheme and the diffusion scheme. If we choose the
FOU scheme to solve the advection term and the FTCS scheme to solve the diffusion term so to be
consistent with equation (6.14), then we have seen previously in equations (4.6) and (5.6) that the
time steps for the separate schemes are
∆x
∆tA = ,
v
∆x2
∆tD = .
2α
Comparing these time steps with equation (6.15) we can see that
∆tA ≥ ∆tAD ,
∆tD ≥ ∆tAD ,
so the split scheme will have a higher maximum allowable time step than the unsplit scheme. When
using the split scheme we choose
∆t = min(∆tA , ∆tD ).

Dr Jon Shiach 91 Finite-Difference Methods

Chapter 6. Solving Multidimensional Partial Differential Equations

6.4.2 Solving the advection-diffusion equation using a split scheme

Splitting equation (6.13) by the advection and diffusion terms gives

Ut + vUx = 0, (6.17a)
Ut − αUxx = 0. (6.17b)

Applying the FOU scheme for equation (6.17a) and the FTCS scheme for equation (6.17b) results in
the following split scheme

v∆t n
u∗i = uni − (u − uni−1 ), (6.18a)
∆x i
α∆t ∗
un+1 = u∗i + (u − 2u∗i + u∗i+1 ). (6.18b)
i
∆x2 i−1

where u∗i denotes the values of the dependent variable updated using the first scheme as before.
Equations (6.18a) and (6.18b) have been applied model advection-diffusion of the same Gaussian
curve used in section 4.3. All of variables remain the same and the diffusion coefficient of α = 0.01
was used. The solution for times t = 0.0, 0.1, 0.2, 0.3, 0.4 and 0.5 is shown in figure 6.4.

Figure 6.4: Solutions to the one-dimensional advection diffusion equation at times t = 0,

0.1, 0.2, 0.3, 0.4 and 0.5.

As the Gaussian curve travels down the channel the diffusion term causes it to dissipate. However, it
was seen in figure 4.4 that the FOU scheme exhibits numerical dissipation when solving the advection
equation so how much of this dissipation is due to the diffusion term and how much is because of the
first-order scheme? The second-order Lax-Wendroff scheme has been used to solve the advection term
in the split scheme and the numerical solutions are compared to the FOU scheme at time t = 0.5 in
figure 6.5. This shows that at the peak of the curve, the first-order dissipation causes at 0.05 reduction
in height.

Finite-Difference Methods 92 Dr Jon Shiach

 Chapter 6. Solving Multidimensional Partial Differential Equations

Figure 6.5: Comparison between the FOU and Lax-Wendroff schemes for solving the advec-
tion part of the split advection-diffusion equation.

The maximum time step for the unsplit scheme is ∆t = 0.0033 and the split scheme is ∆t = 0.005
which means that the unsplit scheme requires 50% more time steps than the split scheme.

Example 12

Consider the following two-dimensional PDE

Ut + αUx + βUyy = 0.

(i) Derive the unsplit second-order differential marching LXY (∆t) for this PDE.
(ii) Use dimensional splitting to split this 2D PDE into two 1D PDEs.
(iii) Derive the two second-order differential marching operators LX (∆t) and LY (∆t) for each
of the split PDEs.
(iv) Show that the split operator sequence LX (∆t)LY (∆t) is equivalent to the unsplit operator
LXY (∆t).

Solution:
(i) Using differential operator notation

∂t U + α∂x U + β∂yy U = 0,
∴ ∂t U = −α∂x U − β∂yy U,
∂tt U = −∂t (α∂x U ) − ∂t (β∂yy U )
= −α∂x (∂t U ) − β∂yy (∂t U )
= −α∂x (−α∂x U − β∂yy U ) − β∂yy (−α∂x U − β∂yy U )
= α2 ∂xx U + 2αβ∂xyy U + β 2 ∂yyyy U.

Dr Jon Shiach 93 Finite-Difference Methods

Chapter 6. Solving Multidimensional Partial Differential Equations

Substituting ∂t U and ∂tt U into the second-order Taylor series expansion

∆t2
U (t + ∆t) = U + ∆t∂t U + ∂tt U + O(∆t3 )
2
∆t2 2
= U + ∆t(−α∂x U − β∂yy U ) + (α ∂xx U + 2αβ∂xyy U + β 2 ∂yyyy U )
2
+ O(∆t3 )
 
∆t2 2
= 1 + ∆t(−α∂x − β∂yy ) + (α ∂xx + 2αβ∂xyy + β ∂yyyy ) U + O(∆t3 )
2
2
therefore
∆t2 2
LXY (∆t) = 1 + ∆t(−α∂x − β∂yy ) + (α ∂xx + 2αβ∂xyy + β 2 ∂yyyy ) + O(∆t3 ).
2
(ii) Using dimensional splitting the PDE can be written as the following system of two PDEs
Ut + αUx = 0, (6.19)
Ut + βUyy = 0. (6.20)
(iii) Marching operator for equation (6.19):
∂t U = −α∂x U,
∂tt U = −α∂t (∂x U ) = −α∂x (∂t U ) = −α∂x (−α∂x U ) = α2 ∂xx U,
∆t2 2
∴ U (t + ∆t) = U + ∆t(−α∂x U ) + (α ∂xx U ) + O(∆t3 )
 2 
α2 ∆t2
= 1 − α∆t∂x + ∂xx U + O(∆t3 ),
2
so
α2 ∆t2
LX (∆t) = 1 − α∆t∂x + ∂xx + O(∆t3 ).
2
Marching operator for equation (6.20):
∂t U = −β∂yy U,
∂tt U = −β∂t (∂yy U ) = −β∂yy (∂t U ) = −β∂yy (−β∂yy U ) = β 2 ∂yyyy U,
∆t2 2
∴ U (t + ∆t) = U + ∆t(−β∂yy U ) + (β ∂yyyy U ) + O(∆t3 )
 2 
β 2 ∆t2
= 1 − β∆t∂yy + ∂yyyy U + O(∆t3 ),
2
so
β 2 ∆t2
LY (∆t) = 1 − β∆t∂yy + ∂yyyy .
2
(iv) Combining the split marching operators LX (∆t) and LY (∆t)
  
α2 ∆t2 β 2 ∆t2
LX (∆t)LY (∆t) = 1 − α∆t∂x + ∂xx 1 − β∆t∂yy + ∂yyyy + O(∆t3 )
2 2
β 2 ∆t2 αβ 2 ∆t3
= 1 − β∆t∂yy + ∂yyyy − α∆t∂x + αβ∆t2 ∂xyy − ∂xyyyy
2 2
α2 ∆t2 α2 β∆t3 α2 β 2 ∆t4
+ ∂xx − ∂xxyy + ∂xxyyyy + O(∆t3 )
2 2 4
∆t2 2
= 1 + ∆t(−α∂x − β∂y ) + (α ∂xx + 2αβ∂xyy + β 2 ∂yyyy )
2
∆t3 ∆t4 2 2
+ (−αβ 2 ∂xyyyy − α2 β∂xxyy ) + α β ∂xxyyyy + O(∆t3 ).
2 4

Finite-Difference Methods 94 Dr Jon Shiach

 Chapter 6. Solving Multidimensional Partial Differential Equations

Since the terms involving ∆t3 and ∆t4 are not O(∆t3 ) they can be omitted and we have

∆t2 2
LX (∆t)LY (∆t) = 1 + ∆t(−α∂x − β∂y ) + (α ∂xx + 2αβ∂xyy + β 2 ∂yyyy ) + O(∆t3 )
2
which is the same as LXY (∆t) from part (i).

6.5 Operator splitting for stiff problems

Cases may arise in split systems where the maximum allowable time step for one scheme may be much
less than that of the other schemes. For example, if dimensional splitting was used to split up a PDE
that models some kind of flow in a two-dimensional domain then the stability criteria are of the form

∆x
∆tX = ,
|v|
∆y
∆tY = .
|w|

If v ≫ w (≫ means ‘much greater than’) and if the spatial steps are roughly equal then it follows that
∆tX ≪ ∆tY . If the sequence LY (∆t)LX (∆t) is used then for stability both schemes will need to use
the same value of ∆t which in this cases will be the much more restrictive ∆tX . So even though the
scheme for solving in the y direction is stable for a much larger value of ∆t, it is constrained by the
stability of the scheme in the x direction. This is an example of a stiff problem.
Fortunately, we can form a sequence where each scheme uses a different value of ∆t whilst the whole
sequence remains stable. Suppose that ∆tY is some multiple of ∆tX , i.e., ∆tY = n∆tX then the
following sequence can be used,
      n
∆tY ∆tY ∆tY
LY (∆tY ) LX . . . LX = LY (∆tY ) LX , (6.21)
n n n
| {z }
n times

i.e., a single application of LY (∆tY ) followed by n applications of LX ( ∆tnY ). To show that this is a
valid sequence consider the first-order differential marching operator for the one-dimensional advection
equation

L(∆t) = 1 − v∆t∂x .

Applying this to equation (6.21) and multiplying out using the binomial theorem
  n    n
∆tY ∆tY
LY (∆tY ) LX = (1 − w∆tY ∂y ) 1 − v ∂x
n n
   
∆tY
= (1 − w∆tY ∂y ) 1 − nv 2
∂x + O(∆t )
n
= (1 − w∆tY ∂y )[1 − v∆tY ∂x + O(∆t2 )]
= 1 − ∆tY (v∂x + w∂y ) + O(∆t2 )
= LXY (∆tY ),

which is the same as the first-order form of equation (6.8(). Therefore the sequence in equation (6.21)
is valid. If LX (∆tX ) and LY (∆tY ) are second-order and n is even then it can be shown that
   n    n
∆tY 2 ∆tY 2
LXY (∆tY ) = LX LY (∆tY ) LX . (6.22)
n n

Dr Jon Shiach 95 Finite-Difference Methods

Chapter 6. Solving Multidimensional Partial Differential Equations

An operator sequence of the form given in equation (6.22) is called a symmetric operator sequence.
One particular example of a symmetric operator uses n = 2, i.e.,
   
∆t ∆t
LXY (∆t) = LX LY (∆t)LX , (6.23)
2 2

which is known as Strang splitting after American mathematician Gilbert Strang (1934–present) (fig-
ure 6.6).

y y y

x x x
∆t ∆t
(a) x sweep using 2
(b) y sweep using ∆t (c) x sweep using 2

Figure 6.6: Strang splitting using the sequence LX ( ∆t ∆t

2 )LY (∆t)LX ( 2 ).

Finite-Difference Methods 96 Dr Jon Shiach

Chapter 7

Solving Systems of Partial Differential

Equations

The universe is governed by physical laws where variables are intrinsically linked. For example, the
behaviour of a fluid depends upon the velocity, density, pressure, temperature, viscosity etc. Any
change in one of these variables will have an effect on all of the others. Also in order to model
physical phenomena we need to make sure that the relationships between the variables satisfy certain
requirements. For example, at its most basic level, the modelling of fluid flow needs to ensure the
conservation of mass and momentum in the system, i.e., mass and energy can’t be added or taken
away∗ . This means that instead of using a single PDE to model physical behaviour, a system of PDEs
is required. We can use the numerical schemes and splitting methods developed for single PDEs on
a system of PDEs with some minor changes. This chapter will use the shallow water equations to
demonstrate the solution of a system of PDEs but the concepts are similar for other systems.

7.1 The Shallow Water Equations

The Shallow Water Equations (SWE) are a hyperbolic system of PDEs that model water flow in
shallow regions. The derivation of the SWE is achieved by taking the Navier-Stokes system of PDEs
and depth-averaging by assuming the flow in the vertical direction is negligible. This may seem like an
odd assumption but it is valid due to the physics of water flow in shallow water. Consider figure 7.1
that shows the motion orbits of a neutrally buoyant particle in different water depths. The water
depth d is defined as the distance between the bed surface and the mean water level. The wavelength
λ for a regular wave is defined as the distance between the corresponding points on two successive
waves (i.e., the troughs). The depth-to-wavelength ratio is used to classify water in three regions.
When λd > 12 the water is considered deep and the motion orbits are roughly circular with the radius
decreasing quadratically as the distance under the surface increases. When 20 1
≤ λd < 21 the water is
considered to be in intermediate depth water and the motion orbits become elliptic with the minor
axis decreasing as the depth under the water surface increases. When λd < 20 1
the water is considered
shallow and the motion of the particle follows a horizontal orbit with very little motion in the vertical
direction. Therefore the assumption that vertical velocity is negligible in shallow water is valid.

∗
This only applies to the modelling of the behaviour of the fluid and not necessary the complete system being modelled.
For example, when modelling water waves we may need to add mass or momentum into the system (using sink or source
terms) to generate the waves. The conservation of mass and momentum in the governing equations ensures that the
fluid behaves as required.

97
Chapter 7. Solving Systems of Partial Differential Equations

λ λ λ

d d d

d 1 1 d 1 d 1
(a) Deep: ≥ (b) Intermediate: ≤ < (c) Shallow: <
λ 2 20 λ 2 λ 20
Figure 7.1: The motion of a particle depends on the depth to wavelength ratio.

In one-dimension, the SWE are a system of two PDEs which are written as

The one-dimensional Shallow Water Equations (SWE)

ht + (hu)x = 0, (7.1a)
 
1 2
2
(hu)t + hu + gh = −ghzx . (7.1b)
2 x

where h is the water depth, u is the horizontal velocity, g = 9.81ms−2 is the acceleration due to gravity
and z is the bed surface elevation (figure 7.2). Equation (7.1a) ensures that we have conservation of
mass, h, and that no mass is added or removed from the system. Equation (7.1b) ensures that we
have conservation of momentum, hu, which means that no energy is added or taken away. The term
on the right-hand side of equation (7.1b) is the bed source term and ensures that change in velocity
caused by changes in the bed topography is balanced by the change in momentum.

g = 9.81ms−2
z

water surface

u
h
bed surface

z
x

Figure 7.2: Definition of variables used in the shallow water equations.

It is more convenient to write the SWE in vector form as

Ut + Fx = S,

where U is the vector of conserved variables, F is the flux vector that models the changes in the
conserved variables in the x direction and S is the vector of source terms, i.e.,
     
h hu 0
U= , F = , S= .
hu hu2 + 21 gh2 −ghzx

Finite-Difference Methods 98 Dr Jon Shiach

 Chapter 7. Solving Systems of Partial Differential Equations

In addition to modelling shallow water flows, the SWE can be used to model tsunami propagation.
A tsunami is generated by seismic activity on the ocean floor that causes the displacement of a large
volume of water. This results in a solitary wave with a wavelength up to λ = 200km and a height of
just 2m travelling at approximately 800kms−1† . Since the average depths of the oceans are typically in
region of 4km, the depth-to-wavelength ratio for a tsunami is 0.02, well within the range of applicability
of the SWE. As the tsunami propagates into shallower water it begins to steepen and a bore wave
forms which can be up to 5m in height. The damage that a tsunami can cause to coastal regions is
catastrophic, one of the worst examples in recent years was the Indian ocean tsunami on Boxing day
2004 that claimed the lives of more than 280,000 people.

7.2 Estimating the wave speeds of the SWE

We saw in section 4.5.2 that in order for a scheme to remain stable, the CFL condition states that the
time step cannot exceed the spatial step divided by the velocity. For linear PDEs such as the advection
equation the velocity is constant, the SWE is a non-linear system meaning that the velocity is variable
across the domain. We need to determine the fastest and slowest possible velocities permitted by the
SWE to ensure that the calculation of the time step never violates the CFL condition. To do this we
linearise the homogeneous form of the SWE (i.e., ignoring the source terms)

Ut + Fx = 0, (7.2)

where
       
U1 h F1 hu
U= = , F = = .
U2 hu F2 hu2 + 12 gh2
by writing it in the form
Ut + JUx = 0, (7.3)
where J is the Jacobian matrix defined by
Definition 12: The Jacobian matrix
 
∂F1 ∂F1
∂F  ∂U ∂U2 
J= =
 ∂F
1 . (7.4)
∂U 2 ∂F2 
∂U1 ∂U2

To evaluate J we can write F using U1 and U2 such that

   
F1 U2
F = = ,
F2 U22 U1−1 + 12 gU12
Evaluating the partial derivatives in the Jacobian gives
∂F1 ∂U2
= = 0,
∂U1 ∂U1
∂F1 ∂U2
= = 1,
∂U2 ∂U2
 
∂F2 ∂ 1 2
= U22 U1−1 + gU = −U22 U1−2 + gU1 = −u2 + gh,
∂U1 ∂U1 2 1
 
∂F2 ∂ 1 2
= U22 U1−1 + gU = 2U2 U1−1 = 2u.
∂U2 ∂U2 2 1
†
A feature of tsunamis is that if you were on a small boat out at sea a tsunami wave would be barely noticeable. It
is only when the tsunami reaches shallower regions that it steepens to form a bore wave. Fisherman would sometimes
return from being out at sea to find that their homes and communities had been destroyed by a tsunami when they
didn’t notice anything unusual when at sea.

Dr Jon Shiach 99 Finite-Difference Methods

Chapter 7. Solving Systems of Partial Differential Equations

so the Jacobian is
 
0 1
J= ,
−u2 + gh 2u

and the SWE can be written as

    
h 0 1 h
+ = 0. (7.5)
hu t −u2 + gh 2u hu x
√ √
J has two eigenvalues λ1 = u + gh and λ2 = u − gh. These are real and distinct as long as h ≥ 0
which classifies equation (7.2) as a system of hyperbolic PDEs. We want to split equation (7.5) into
two separate PDEs that are analogous to the advection equation. To do this we diagonalise J to
produce a matrix D where the eigenvalues of J form the main diagonal, i.e.,
 
λ1 0
D= .
0 λ2


Let P = p1 p2 where p1 and p2 are the column eigenvectors corresponding to λ1 and λ2 then


JP = J p1 p2


= Jp1 Jp2


= λ1 p1 λ2 p2 (since Jp = λp)
 

λ1 0
= p1 p2
0 λ2
= PD
−1
P JP = D.

Therefore P −1 JP is the transformation that diagonalises J. Introducing a change in variable where

W = (W1 , W2 )T such that U = P W then equation (7.3) becomes

P Wt + JP Wx = 0,

premultiplying by P −1 and simplifying gives

P −1 P Wt + P −1 JP Wx = 0
Wt + DWx = 0.

D is a diagonal matrix so we can write this as a system of two linear equations

∂  p  ∂
(W1 ) + u + gh (W1 ) = 0, (7.6a)
∂t ∂x
∂  p  ∂
(W2 ) + u − gh (W2 ) = 0. (7.6b)
∂t ∂x
√ √
Equations (7.6a) and (7.6b) are two linear advection equations with wave speeds u + gh and u − gh
(the eigenvalues of J). These wave speeds are the fastest and slowest speeds permissible by the SWE.
Since we are interested in the maximum absolute wave speed for stability purposes we can see that
 p p  p
max |u + gh|, |u − gh| = |u| + gh.

7.2.1 Calculating the time step for the SWE

Any explicit scheme to solve the SWE must take into account these wave speeds when determining
the maximum allowable time step for stability. Recall that the maximum allowable time step for the
advection equation is
∆x
∆t ≤ . (7.7)
|v|

Finite-Difference Methods 100 Dr Jon Shiach

 Chapter 7. Solving Systems of Partial Differential Equations

√
Since the wave speeds for the SWE are |u| + gh then the maximum allowable time step for the SWE
is

∆x
∆t ≤ √ . (7.8)
|u| + gh

Equation (7.8) calculates a separate value of ∆t for each point in the domain. We require a single
value of ∆t so that a numerical scheme is stable for all points in the domain, therefore we choose the
smallest value of ∆t, i.e.,

Maximum allowable time step for the SWE

∆x
∆t ≤ p . (7.9)
max(|ui | + ghi )
i

Since the values of h and u will change over time, equation (7.9) is calculated at the beginning of each
time step (note that in practice the maximum allowable time step is multiplied by some safety, e.g.,
0.9, factor to ensure stability).

7.3 Solving the shallow water equations using finite-difference schemes

The FDS developed for the advection equation in chapter 4 can be modified to solve the SWE. Recall
that the Lax-Friedrichs scheme for solving the advection equation is

1 ∆t n
un+1 = (uni+1 + uni−1 ) − v (u − uni−1 ) + O(∆t, ∆x).
i
2 2∆x i+1

The velocity in the advection equation, v, is constant in space and time. The velocity in the SWE is
variable in space and time and is governed by the flux vector F that is differentiated with respect to
x. Therefore we can replace the uni±1 terms in the spatial derivative with the corresponding flux terms
and remove v giving the Lax-Friedrichs scheme for solving the SWE

The Lax-Friedrichs scheme for solving the one-dimensional SWE

1 n ∆t
Uin+1 = (Ui+1 n
+ Ui−1 )− (F n − Fi−1
n
) + O(∆t, ∆x). (7.10)
2 2∆x i+1

7.3.1 Solution procedure

The solution procedure for solving the SWE is similar to that used for the advection equation with a
couple of exceptions. Unlike the advection equation where the velocity, and therefore the maximum
allowable time step ∆t, is constant, the velocities in the SWE are variable so ∆t is calculated at the
beginning of each time step. Also the flux vector is calculated before the finite-difference scheme is
applied to march the solution forward in time.

Dr Jon Shiach 101 Finite-Difference Methods

Chapter 7. Solving Systems of Partial Differential Equations

start

initialise variables

calculate U 0

no
t < tmax end

yes

calculate ∆t

calculate boundary conditions

calculate F n

apply FDS to calculate U n+1

update U n ← U n+1 and t ← t + ∆t

plot current solution

Figure 7.3: Flow chart for solving the SWE.

Finite-Difference Methods 102 Dr Jon Shiach

 Chapter 7. Solving Systems of Partial Differential Equations

Example 13

The SWE are being used to model a one-dimensional dam break. The domain is 2m in length
with a partition located at x = 1m separating two bodies of water with heights 5m and 2m
for the upstream and downstream bodies respectively. First-order transient flow boundary
conditions are employed at both ends of the domain that is discretised into a reduced grid with
∆x = 0.5m for verification purposes. Acceleration due to gravity is assumed to be g = 9.81ms-2 .

Calculate two steps of the Lax-Friedrichs scheme for this problem.

Solution: Since ∆x = 0.5m the co-ordinates of the finite-differences nodes are

x0 = 0.0, x1 = 0.5, x2 = 1.0, x3 = 1.5, x4 = 2.0,
so the initial conditions are
h0 = 5, h1 = 5, h2 = 5, h3 = 2, h4 = 2,
u0 = 0, u1 = 0, u2 = 0, u3 = 0, u4 = 0.
Since zero gradient boundary conditions are used at both boundaries, we have
h−1 = h1 = 5, h5 = h3 = 2,
u−1 = u1 = 0, u5 = u3 = 0.
We form an array where each column contains the vector U = (h, hu)T for each node (including
the ghost nodes)
 
0 5 5 5 5 2 2 2
U =
0 0 0 0 0 0 0

and do similar for the flux vector F = (hu, hu2 + 12 gh2 )T

 
0 0 0 0 0 0 0 0
F =
122.625 122.625 122.625 122.625 19.62 19.62 19.62
Calculate the maximum allowable time step (using a safety factor of 0.9)
   
∆x 0.5
∆t = 0.9 √ = 0.9 √ = 0.0643.
maxi (|ui | + ghi ) 5g
Now perform one step of the Lax-Friedrichs scheme equation (7.10)
         
1 5 5 0 0 5
U01 = + − 0.0643 − = ,
2 0 0 122.625 122.625 0
         
1 5 5 0 0 5
1
U1 = + − 0.0643 − = ,
2 0 0 122.625 122.625 0
         
1 5 2 0 0 3.5
1
U2 = + − 0.0643 − = ,
2 0 0 19.62 122.625 6.6184
         
1 5 2 0 0 3.5
1
U3 = + − 0.0643 − = ,
2 0 0 19.62 122.625 6.6184
         
1 2 2 0 0 2
U41 = + − 0.0643 − = .
2 0 0 19.62 19.62 0
U2
The conserved values can be extracted from U = (U1 , U2 )T using h = U1 and u = therefore
U1
h0 = 5, h1 = 5, h2 = 3.5, h3 = 3.5, h4 = 2,
u0 = 0, u1 = 0, u2 = 1.8910, u3 = 1.8910, u4 = 0.

Dr Jon Shiach 103 Finite-Difference Methods

Chapter 7. Solving Systems of Partial Differential Equations

Moving onto the second step we first calculate the maximum allowable time step
 
0.5
∆t = 0.9 √ = 0.0581,
1.8923 + 3.5g

and using the boundary conditions to calculate the values of the ghost nodes

h−1 = h1 = 5, h5 = h3 = 3.5,
u−1 = u1 = 0, u5 = u3 = 1.8910.

The U 1 and F 1 arrays are

 
5 5 5 3.5 3.5 2 3.5
U1 = ,
0 0 0 6.6184 6.6184 0 6.6184
 
1 0 0 0 6.6184 6.6184 0 6.6184
F = .
122.625 122.625 122.625 72.6014 72.6014 19.62 72.6014

Performing another step of the Lax-Friedrichs scheme

         
1 5 5 0 0 5
U02 = + − 0.0581 − = ,
2 0 0 122.625 122.625 0
         
1 3.5 5 0 0 3.8657
2
U1 = + − 0.0581 − = ,
2 6.6184 0 72.6014 122.625 6.2136
         
1 3.5 5 0 0 3.8657
2
U2 = + − 0.0581 − = ,
2 6.6184 0 72.6014 122.625 6.2136
         
1 2 3.5 0 6 3.1343
2
U3 = + − 0.0581 − = ,
2 0 6.6184 19.62 72.6014 6.3853
         
1 3.5 3.5 6.6184 6.6184 3.5
U42 = + − 0.0581 − = .
2 6.6184 6.6184 72.6014 72.6014 6.6184

The values of the conserved variables after two steps of the Lax-Friedrichs scheme are

h0 = 5, h1 = 3.8657, h2 = 3.8657, h3 = 3.1343, h4 = 3.5,

u0 = 0, u1 = 1.6073, u2 = 1.6073, h3 = 2.0373, h4 = 1.8910.

Finite-Difference Methods 104 Dr Jon Shiach

 Chapter 7. Solving Systems of Partial Differential Equations

7.3.2 The one-dimensional dam break test

To test a numerical scheme for solving the SWE we can use the one-dimensional dam break test. This
consists of a one-dimensional channel with two bodies or reservoirs of water separated by a partition
or dam (figure 7.4(a)). At time t = 0 the partition is instantaneously removed and the two bodies
of water are allowed to interact. The collapse of the upstream reservoir causes a bore wave to form
the travels downstream (figure 7.4(b)). The loss of mass due to this bore wave causes an upstream
travelling rarefaction wave, or depression wave, to form.

partition

hup

hdown
x
(a) initial conditions
h
rarefaction wave

bore
wave

x
(b) solution at time t

Figure 7.4: The one-dimensional dam break problem.

The dam break problem is a tough test for FDS since the initial conditions contain a discontinuity
which is also present in the solutions in the form of a bore wave. The one-dimensional dam break
problem is one of the few problems for the SWE where it is possible to calculate an analytical solution.

The Lax-Friedrichs scheme shown in equation (7.10) has been applied to the one-dimensional dam
break problem. The domain 0 ≤ x ≤ 1 is discretised using N = 101 nodes giving ∆x = 0.01 with a
partition located at x = 0.5 separating the reservoirs to the upstream (left) and downstream (right)
sides of the partition with water depths hup = 1 and hdown = 0.5 respectively. Transmissive boundary
conditions using ghost nodes were employed at both ends of the domain so that waves will pass through
the boundary without reflection, i.e.,

h−1 = h1 , hN = hN −2 ,
u−1 = u1 , uN = uN −2 .

The solutions for the water depth h and velocity u are shown in figure 7.5. Similar to the advection
equation, the Lax-Friedrichs scheme exhibits numerical dissipation which means the numerical solution
does not capture the discontinuity at the bore wave and smooths out the rarefaction wave. In order
to capture these phenomena we require a higher order scheme.

Dr Jon Shiach 105 Finite-Difference Methods

 Chapter 7. Solving Systems of Partial Differential Equations

(a) water height (b) velocity

Figure 7.5: Solution to the one-dimensional dam break at time t = 0.1s using the Lax-
Friedrichs scheme

7.3.3 MATLAB code

The MATLAB code for solving the one-dimensional SWE using the Lax-Friedrichs scheme is given
in listing 7.1. Note that this program makes use of functions to calculate the time step, boundary
conditions, flux vector and the Lax-Friedrichs scheme. It is good practice to structure a program in
this way as it breaks down the program into manageable chunks and avoids repetition of code when
performing the same computation multiple times.

Listing 7.1: MATLAB code for solving the SWE using the Lax-Friedrichs scheme
1 % oned_swe . m by Jon Shiach
2 %
3 % This program solves the SWE using the Lax - Friedrichs scheme
4
5 % Clear workspaces
6 clear
7 clc
8
9 % Define global parameters
10 global gravity
11 gravity = 9.81; % acceleration due to gravity
12

13 % Define variables
14 N = 101; % number of nodes
15 xmin = 0; % lower bound of x
16 xmax = 1; % upper bound of x
17 hL = 1; % water height to the left of the dam
18 hR = 0.5; % water height to the right of the dam
19 t = 0; % time variable
20 tmax = 0.1; % max value of t
21
22 % Define spatial array ( including ghost nodes )
23 dx = ( xmax - xmin ) / ( N - 1) ;
24 x = xmin - dx : dx : xmax + dx ;
25
26 % Define initial conditions
27 U = zeros (2 , length ( x ) ) ;
28 U (1 , x <= 0.5) = hL ;
29 U (1 , x > 0.5) = hR ;
30

Finite-Difference Methods 106 Dr Jon Shiach

 Chapter 7. Solving Systems of Partial Differential Equations

31 % Time marching loop

32 while t < tmax
33
34 % Calculate maximum allowable time step
35 dt = calculate_dt (U , dx , t , tmax ) ;
36
37 % Calculate boundary conditions
38 U = calculate_BC ( U ) ;
39
40 % Calculate F vector
41 F = calculate_F ( U ) ;
42
43 % Calculate one step of the Lax - Friedrichs scheme
44 U = Lax_Friedrichs (U , F , dt , dx ) ;
45
46 % Update t
47 t = t + dt ;
48
49 % Plot current solution
50 num_plot = plot (x , U (1 , :) , 'o - ' , ' markerfacecolor ' , ' blue ') ;
51 axis ([ xmin , xmax , 0 , 1.2])
52 xlabel ( '$ x $ ' , ' fontsize ' , 16 , ' interpreter ' , ' latex ')
53 ylabel ( '$ h $ ' , ' fontsize ' , 16 , ' interpreter ' , ' latex ')
54 title ( sprintf ( '$ t = %1.3 f $ ' , t ) , ' fontsize ' , 16 , ' interpreter ' , ' latex ')
55 shg
56 pause (0.001)
57 end
58
59 % Plot numerical solution against exact solution
60 load dambreak_exact
61 hold on
62 exact_plot = plot ( xexact , hexact , 'r - ' , ' linewidth ' , 2) ;
63 hold off
64 leg = legend ([ exact_plot , num_plot ] , ' exact solution ' , ' numerical solution ')
;
65 set ( leg , ' fontsize ' , 14 , ' interpreter ' , ' latex ')
66

67 % -------------------------------------------------------------------------
68 function dt = calculate_dt (U , dx , t , tmax )
69
70 % This function calculates the maximum allowable time step
71 global gravity
72 dt = 0.9 * dx / max ( abs ( U (2 , :) ./ U (1 , :) ) + sqrt ( gravity * U (1 , :) ) ) ;
73 dt = min ( dt , tmax - t ) ;
74
75 end
76
77 % -------------------------------------------------------------------------
78 function U = calculate_BC ( U )
79
80 % This function calculates the values of the ghost nodes using transmissive
81 % boundary conditions
82 U (: , 1) = U (: , 3) ;
83 U (: , end ) = U (: , end -2) ;
84

85 end
86
87 % -------------------------------------------------------------------------
88 function F = calculate_F ( U )
89
90 % This function calculates the flux vector F from U .

Dr Jon Shiach 107 Finite-Difference Methods

 Chapter 7. Solving Systems of Partial Differential Equations

91 global gravity
92 F (1 , :) = U (2 , :) ;
93 F (2 , :) = U (2 , :) .^2 ./ U (1 , :) + 0.5 * gravity * U (1 , :) .^2;
94
95 end
96
97 % -------------------------------------------------------------------------
98 function Unp1 = Lax_Friedrichs (U , F , dt , dx )
99
100 % This function calculates a single step of the Lax - Friedrichs scheme
101
102 N = size (U , 2) - 2;
103 Unp1 = U ;
104 C = 0.5 * dt / dx ;
105 for i = 2 : N + 1
106 Unp1 (: , i ) = 0.5 * ( U (: , i +1) + U (: , i -1) ) - C * ( F (: , i +1) - F (: , i -1) )
;
107 end
108

109 end

7.3.4 The MacCormack scheme

The Lax-Wendroff scheme given in equation (4.15) can not be applied to solve the SWE in its current
form since it can only handle flow in one direction (i.e., linear PDEs). Instead, we can write it as an
equivalent two-step method which is known as the MacCormack scheme (MacCormack, 1969).

The MacCormack scheme for solving the one-dimensional SWE

n+ 12 ∆t

Ui = Uin − n
Fi+1 − Fin , (7.11a)
∆x
1 n n+ 1 ∆t n+ 1 n+ 1
Uin+1 = (Ui + Ui 2 ) − (Fi 2 − Fi−1 2 ) + O(∆t2 , ∆x2 ). (7.11b)
2 2∆x

The predictor step equation (7.11a) updates the solution over a partial time step and the corrector
step equation (7.11b) uses the predictor values to update the solution over a full time step. The
two-step MacCormack scheme is equivalent to the Lax-Wendroff scheme and has the same order and
stability properties. Since the predictor step changes the values of the variables used in the corrector
step, the boundary conditions and flux vector need to be calculated using the predictor values prior
to the corrector step.

The MacCormack scheme was applied to the one-dimensional dam break problem and the solutions
can be seen in figure 7.6. Unlike the Lax-Friedrichs scheme, the MacCormack scheme is able to capture
the discontinuity at the bore wave and does not suffer from numerical dissipation. However, there
are noticeable oscillations in the solution caused by the abrupt changes in the gradient of the surface.
This is a common feature of second-order methods. There are methods that can be applied to the
scheme in order to minimise these oscillations but this are outside the scope of this unit (interested
readers are advised to search for ‘artificial dissipation’, ‘TVD schemes’ and ‘MUSCL schemes’).

Finite-Difference Methods 108 Dr Jon Shiach

 Chapter 7. Solving Systems of Partial Differential Equations

(a) water height (b) velocity

Figure 7.6: Solution to the one-dimensional dam break at time t = 0.1s using the MacCor-
mack scheme

7.4 Two-dimensional shallow water equations

The two-dimensional SWE include a third equation that models flow in the y direction.
The two-dimensional Shallow Water Equations

ht + (hu)x + (hv)y = 0,
 
1 2
2
(hu)t + hu + gh + (huv)y = −ghzx ,
2
 x

1 2
2
(hv)t + (huv)x + hv + gh = −ghzy .
2 y

The variables h, u, g and z have been defined previously with the exception of v which is the horizontal
flow velocity in the y direction. These can be written more succinctly in vector form as

Ut + Fx + Gy = S,

where U is the vector of conserved variables, F and G are flux vectors that model change in the
conserved variables in the x and y directions respectively and S are the source terms that model bed
topography
       
h hu hv 0
U = hu , F = hu2 + 21 gh2  , G= huv , S = −ghzx  .
hv huv 2 1
hv + 2 gh 2 −ghzy

7.4.1 Solving the two-dimensional shallow water equations

Applying the second-order MacCormack scheme with dimensional splitting means we can use the
following symmetric operator sequence
       
∆t ∆t ∆t ∆t
Uin+1
= LX LY LY LX Uin .
2 2 2 2

The solution procedure for solving the two-dimensional SWE is summarised in the flow chart in
figure 7.7.

Dr Jon Shiach 109 Finite-Difference Methods

Chapter 7. Solving Systems of Partial Differential Equations

start

initialise variables

calculate U 0

yes
t ≥ tmax end

no
calculate ∆t

calculate bound-
ary conditions

calculate F n or Gn
use 1D Lax-Wendroff in
the x direction with ∆t
2 1
calculate U n+ 2

use 1D Lax-Wendroff in the calculate bound-

in the y direction with ∆t
2 ary conditions

calculate
1 1
use 1D Lax-Wendroff in the F n+ 2 or Gn+ 2
in the y direction with ∆t
2

calculate U n+1
use 1D Lax-Wendroff in
the x direction with ∆t
2

update U n ← U n+1
and t ← t + ∆t

plot current solution

Figure 7.7: Flow chart for solving the two-dimensional SWE using the MacCormack scheme.

Finite-Difference Methods 110 Dr Jon Shiach

 Chapter 7. Solving Systems of Partial Differential Equations

7.5 Validating the two-dimensional SWE solver

Adding a another spatial dimension to the SWE provides further complications when writing the
computer program to perform the calculations. As always, it is a good idea to validate the program
to test whether it produces expected results.

7.5.1 Pseudo one-dimensional dam break

The first test that is being used is a pseudo one-dimensional dam break problem where the one-
dimensional dam break test described in section 7.3.2 is applied to a two-dimensional grid (figure 7.8).
The domain is discretised using 50 nodes in both the x and y directions. The solution to this problem
should be the same as that described for the one-dimensional case along the direction of flow and
horizontal in the other direction. The numerical solutions for the pseudo one-dimensional dam break
for both the x and y directions can be seen in figure 7.9. Taking a cross-section of the solutions shows
that they are comparable to the one-dimensional solutions seen previously (figure 7.10).

partition
y y
1 1

hup = 1

hup = 1 hdown = 0.5

hdown = 0.5

0 0
0 1 x 0 1 x
(a) x direction (b) y direction

Figure 7.8: Plan view of the pseudo one-dimensional dam break problem.

(a) x direction (b) y direction

Figure 7.9: Pseudo one-dimensional dam break test.

Dr Jon Shiach 111 Finite-Difference Methods

Chapter 7. Solving Systems of Partial Differential Equations

(a) water height (b) velocity

Figure 7.10: Comparisons of the solutions for the water height and velocity for the pseudo
one-dimensional dam break problem in the x and y directions.

7.5.2 Collapse of a cylindrical column

The pseudo one-dimensional dam break test does not test the numerical solutions for flow in both
spatial dimensions at the same time. One test that will show up any flaws in the solver is the modelling
of the collapse of a two-dimensional cylindrical column. The problem consists of a two-dimensional
domain 0 ≤ x, y ≤ 1 with a water depth of hdown = 0.5. The initial conditions consist of a perturbation
in water depth in the form of a circular region centred at (0.5, 0.5) with radius 0.2 in which the water
height is hup = 1.0 (figure 7.11). The velocities in the x and y directions are initialised to zero.

1
hdown = 0.5

hup = 1.5

0 x
0 1

Figure 7.11: Initial conditions for the collapse of a cylindrical column problem.

The values of the initial conditions can be calculated using the following
( p
1, (x − 0.5)2 + (y − 0.5)2 < 0.1,
h(x, y) = ,
0.5, elsewhere,
u(x, y) = 0,
v(x, y) = 0.

Solid wall boundary conditions are employed at all four boundaries and the values of the ghost nodes

Finite-Difference Methods 112 Dr Jon Shiach

 Chapter 7. Solving Systems of Partial Differential Equations

are calculated using the following

h−1,j = h1,j , hNX ,j = hNX −2,j , hi,−1 = hi,1 , hi,NY = hi,NY −2 ,

u−1,j = −u1,j , uNX ,j = −uNX −2,j , ui,−1 = ui,1 , ui,NY = ui,NY −2 ,
v−1,j = v1,j , vNX ,j = vNX −2,j , vi,−1 = −vi,1 , vi,NY = −vi,NY −2 .

The collapse of the cylindrical column causes a circular bore wave to form propagating in the radial
direction away from the centre. If there are any problems in resolving the spatial derivatives caused
by the dimensional splitting it will be noticeable in the solution. A contour plot of the values of h at
t = 0.075 is shown in figure 7.12. Here we can see the a circular bore wave is propagating at equal
velocities away from the centre of the cylinder thus indicating that the two-dimensional solver can
resolve the spatial dimensions equally.

Figure 7.12: Contour plot of the collapse of a cylindrical column showing the bore waves
travelling with equal velocities away from the centre of the cylinder.

The solutions at various times in the simulation are shown in figure 7.13. At time t = 0.05 the
cylindrical column has collapsed in on itself and the outward travelling bore wave has started to form.
At time t = 0.10 the bore wave has almost reached the boundaries and a depression has formed where
the centre of the cylinder was. At time t = 0.15 the bore waves have reached the boundaries and are
reflected back towards the interior of the domain. At time t = 0.25 and t = 0.3 the reflected waves
have started to interact with each other causing the water surface to break up into lots of individual
waves.

Dr Jon Shiach 113 Finite-Difference Methods

Chapter 7. Solving Systems of Partial Differential Equations

Figure 7.13: Surface plots of the solution to the collapse of a cylindrical column problem at
t = 0.00, 0.05, 0.10, 0.15, 0.25 and 0.30.

Finite-Difference Methods 114 Dr Jon Shiach

References

Courant, R., Friedrichs, K., and Lewy, H. (1967). “On the partial difference equations of mathematical
physics”. In: IBM Journal of Research and Development 11.2. (English translation from German
of the original paper published in 1927), pp. 215 –234. url: http://web.stanford.edu/class/
cme324/classics/courant-friedrichs-lewy.pdf.

Crank, J and Nicolson, P. (1947). “A practical method for numerical evaluation of solutions of partial
differential equations of the heat conduction type.” In: Proc. Camb. Phil. Soc. 1, pp. 50–67.

Hirsch, C. (1994). Numerical Computation of Internal and External Flows. Vol. 1: Fundamentals of
Numerical Discretization. John Wiley & Sons. Chap. 8, pp. 283–288.

Hoffman, J.D. (2001). Numerical Methods for Engineers and Scientists. 2nd ed. CRC Press.

Lax, P.D. and Richtmyer, R.D. (1956). “Survey of the stability of linear finite-difference equations”.
In: Comm. Pure Appl. Math. 9, pp. 267–293.

Lax, P.D. and Wendroff, B. (1960). “Systems of conservation laws”. In: Comm. Pure Appl. Math. 13.2,
pp. 217–237.

MacCormack, R.W. (1969). “The effect of viscosity in hypervelocity impact cratering”. In: pp. 69–354.

Roache, P.J. (1994). “Perspective: a method for uniform reporting of grid refinement studies”. In: J.
Fluids Eng. 116, pp. 405–412.

Yang, Shiming and Gobbert, Matthias K (2009). “The optimal relaxation parameter for the SOR
method applied to the Poisson equation in any space dimensions”. In: Applied Mathematics Letters
22.3, pp. 325–331.

115
Appendix A

Derivation of the Thomas algorithm

Consider a tridiagonal linear system expressed in the form Ax = d

    
b1 c1 x1 d1
a2 b2 c2   x2   d2 
    
 a3 b3 c3   x3   d3 
    
 .. .. ..   ..  =  ..  (A.1)
 . . .  .   . 
    
 an−1 bn−1 cn−1  xn−1  dn−1 
an bn xn dn

This linear system is converted into upper triangular form U x = d′ where the main diagonal only
contains 1’s. Consider row 1 of equation (A.1)

b1 x1 + c1 x2 = d1

Divide by b1 so that the main diagonal is equal to 1

c1 d1
x1 + x2 = .
b1 b1
c1 d1
Let c′1 = and d′1 = then
b1 b1

x1 + c′1 x2 = d′1 . (A.2)

Consider row 2 of equation (A.1)

a2 x1 + b2 x2 + c2 x3 = d2 (A.3)

Eliminate x1 using a2 multiplied by equation (A.2)

a2 x1 + b2 x2 + c2 x3 − a2 (x1 + c′1 x2 ) = d2 − a2 d′1

(b2 − a2 c′1 )x2 + c2 x3 = d2 − a2 d′1

Divide by b2 − a2 c′1 so that the main diagonal element is equal to 1

c2 d2 − a2 d′1
x2 + x 3 = .
b2 − a2 c′1 b2 − a2 c′1

c2 d2 − a2 d′1
Let c′2 = ′ and d′2 = then
b2 − a2 c1 b2 − a2 c′1

x2 + c′2 x3 = d′2 . (A.4)

117
Appendix A. Derivation of the Thomas algorithm

Consider row 3 of equation (A.1)

a3 x2 + b3 x3 + c3 x4 = d3
Eliminate x1 using a3 multiplied by equation (A.4)

a3 x2 + b3 x3 + c3 x4 − a3 (x2 + c′2 x3 ) = d3 − a3 d′2

(b3 − a3 c′2 )x3 + c3 x4 = d3 − a3 d′2

Divide by b3 − a3 c′2 so that the main diagonal element is equal to 1

c3 d3 − a3 d′2
x3 + x 4 = .
b3 − a3 c′2 b3 − a3 c′2
′
Let c′3 =
c3
and d′ = d3 − a3 d2 then
b3 − a3 c′2 3
b3 − a3 c′2

x3 + c′3 x4 = d′3 . (A.5)

Consider an ith row of equation (A.1). Using a similar method shown for rows 2 and 3 above the ith
row can be written as

xi + c′i xi+1 = d′i (A.6)

where
ci
c′i = (A.7)
bi − ai c′i−1
di − ai d′i−1
d′i = . (A.8)
bi − ai c′i−1

Finally, consider the nth row of equation (A.1)

an xn−1 + bn xn = dn .

Eliminate xn−1 using an multiplied by equation (A.6

an xn−1 + bn xn − an (xn−1 + c′n−1 xn ) = dn − an d′n−1

(bn − an c′n−1 )xn = dn − an d′n−1

Divide by bn − an c′n−1 so that the main diagonal element is equal to 1

dn − an d′n−1
xn = .
bn − an c′n−1

The matrix equation can now be expressed in the form U x = d′

    ′ 
1 c′1 0 0 0 0 x1 d1
0 1 c′2 0 0 0   x2   d′2 
    
0 0 1 c′3 0 0     ′ 
   x3   d3 
 .. ..   ..  =  ..  . (A.9)
 . .  .   . 
    
0 0 0 0 1 cn−1  xn−1  d′n−1 
′

0 0 0 0 0 1 xn d′n

Once the system is written in upper triangular form the solution of x can be found by back-substitution.
Starting with row n of equation (A.9) the solution for xn is simply d′n

xn = d′n . (A.10)

Finite-Difference Methods 118 Dr Jon Shiach

 Appendix A. Derivation of the Thomas algorithm

Consider a row i of equation (A.9) when i = n − 1, n − 2, . . . , 1

xi + c′i xi+1 = d′i ,

therefore

xi = d′i − c′i xi+1 . (A.11)

The Thomas algorithm is written as follows

• Calculate modified coefficients c′ and d′ using
c


i
i = 1,
′ bi
ci = ci

 b − a c′ i = 2, 3, . . . , n
i i i−1


 d i
 i = 1,
b
d′i = dii − ai d′

 i−1
 ′ i = 2, 3, . . . , n
bi − ai ci−1

• Calculate solutions of xi using

(
d′i − c′i xi+1 i = n − 1, n − 2, . . . , 1,
xi =
d′i i = n.

Dr Jon Shiach 119 Finite-Difference Methods