MATH5004-LEC2: Outline

30.07.
2010
MATH5004-LEC2
FDM: Consistency, stability, converence, and error estimates
Curtin University is a trademark of Curtin University of Technology

CRICOS Provider Code 00301J
OUTLINE
 Global error
 Local truncation error
 Convergent method
 Grid refinement analysis
 Stencils for the 2D Laplacian
 Example
B Wiwatanapataphee 8/16/2020

CRICOS Provider Code 00301J 1
30.07.2010
Definition of the global error

Let 𝐔 = 𝑈 , … , 𝑈 be he solution of finite difference
scheme (assume no round-ff errors),
u = 𝑢(𝑥 ), … , 𝑢(𝑥 ) be the exact solution at grid points.
The global error vector is defined as 𝐄 = 𝐔 − 𝐮.
• The maximum norm (or the infinity norm) is
• The 1-norm is
• The 2-norm is
If 𝐄 ≤ 𝐶ℎ , 𝑝 > 0, C is a constant independent of h

the finite difference (FD) scheme is p-th order accurate
Definition A FD scheme is called convergent if lim 𝐄 = 0
→
Local truncation error
Consider the DE 𝑢 =𝑓 or 𝐿𝑢 = 𝑓 …..(1)
L is the differential operator 𝐿 =

By applying FDM
( )
DE(1) becomes =𝑓
or 𝐿 𝑢 = 𝑓
The local truncation error is 𝑇(𝑥) = 𝐿𝑢 − 𝐿 𝑢

30.07.2010
For the DE 𝑢 = 𝑓(𝑥)

and the three-point central difference scheme,
the local truncation error is
𝑇(𝑥) = 𝐿𝑢 − 𝐿 𝑢
𝑢 𝑥 + ℎ − 2𝑢 𝑥 + 𝑢(𝑥 − ℎ)
=𝑢 −
ℎ
𝑢 𝑥 + ℎ − 2𝑢 𝑥 + 𝑢(𝑥 − ℎ)
= 𝑓(𝑥) −
ℎ
which measures how well the FD discretisation
approximates the DE
Convergent Method
Definition A finite difference scheme is called consistent if
lim 𝑇 𝑥 = lim 𝐿𝑢 − 𝐿 𝑢 = 0
→ →
To check whether a FD scheme is consistent or not,

we use Taylor expansion.
For the DE 𝑢 = 𝑓 𝑥 , we have
𝑢 𝑥 + ℎ − 2𝑢 𝑥 + 𝑢(𝑥 − ℎ)
𝑇(x) = 𝑢 −
ℎ
ℎ
≈− 𝑢 𝑥 = Cℎ
12 So the FD scheme
lim 𝑇 𝑥 = lim Cℎ = 0 is consistent.
→ →

30.07.2010
As the consistence cannot guarantee a finite difference

work, the stability of a FD scheme is required.
For the model problem, we have
𝐴𝐮 = 𝐅 + 𝐓, 𝐴𝐔 = 𝐅
𝐴(𝐮 − 𝐔) = 𝐓
𝐴𝐄 = 𝐓
A is the coefficient matrix of the FD equations,

F is the modified source term by the BC,
T is the vector of the local truncation error at grid points
𝐴𝐄 = 𝐓
If A is non-singular, then
If A is singular, then
may be arbitrary large.
For the central FD scheme, we have
Thus, the global error depends on both the local

truncation error and

30.07.2010
Definition
A FD method for the elliptic DE is stable if A is
invertible and
where C and ℎ are constants
Theorem
A consistent and stable finite difference scheme is
convergent.
Grid refinement analysis

Assume that a method is p-th order accurate, i.e.,
We can divide h by half to get

We then have
10

30.07.2010
Improvement of Accuracy of Solution

Deferred Approach to the Limit
By reducing the value of k and h, i.e., through mesh refinement.
Denote u = true solution,
uh = finite difference solution,
then u = uh + error.
Let uh be the solution of the Laplace equation by using the 5
point scheme, then
u = uh +Ch2 .

11
 Based on the above formulae, to accelerate the convergence of

numerical solutions, we could use the so-called deferred
approach to the limit suggested by Richardson.
u = u1 +Chp . (1.1)
If mesh size h is reduced from h to h/2, then
u = u2 +C(h/2)p. (1.2)
where u2 is the approximation obtained with h/2.

12

30.07.2010
From (1.1) and (1.2), by eliminating ℎ , we obtain
2 p u2 - u1 (1.3)
u= p
.
2 - 1
For example, for p = 2,
4u2 - u1
u=
3

13
 If p is unknown, it can be estimated from three approximate

solutions at the same point obtained by using different mesh
sizes.
 Let u1 , u2 and u3 are approximations at the same point
obtained respectively using step size h1 , h2 and h3 where
1 1
h3 = h2 = h1
2 4
Then
𝑢 = 𝑢 + 𝐶ℎ ,
𝑢 =𝑢 +𝐶 ℎ ,
𝑢 =𝑢 +𝐶 ℎ ,

14

30.07.2010
from which, we have

15
Elliptic equations
In two space dimensions, a constant-coefficient elliptic
equation has the form
𝑎𝑢 + 𝑏𝑢 + 𝑐𝑢 + ℎ 𝑥, 𝑦, 𝑢, 𝑢 , 𝑢 = 𝑓(𝑥, 𝑦)
where the coefficients 𝑎, 𝑏, 𝑐 satisfy

𝑏 − 4𝑎𝑐 < 0.
Consider the problem

𝑢 +𝑢 = 𝑓(𝑥, 𝑦), 0 ≤ 𝑥, 𝑦 ≤ 1
16

30.07.2010
Stencils for the 2D Laplacian
The five-point stencil The nine-point stencil
17
Accuracy and Stability: the five-point stencil
Let ∆𝑥 = ∆𝑦 = ℎ
𝑇(𝑥) = 𝐿𝑢 − 𝐿 𝑢
𝑢 , +𝑢 , +𝑢, +𝑢, − 4𝑢 .
= (𝑢 +𝑢 ) −
ℎ
1
𝑇 =𝑓 − 𝑢 . +𝑢 . +𝑢. +𝑢 . − 4𝑢 .
ℎ
1
𝑇 = ℎ 𝑢 +𝑢 + 𝑂(ℎ )
12
The global error 𝐸 = 𝑢(𝑥 , 𝑢 ) − 𝑢

𝐴 E =T
18

30.07.2010
Accuracy and Stability: the nine-point stencil
∇ 𝑢(𝑥 , 𝑦 )
∇ 𝑢(𝑥 , 𝑦 )
It is noted that the error is still 𝑂 ℎ .
19
Since
From DE ∇ 𝑢 = 𝑓, we have
We can obtain a fourth-order accurate method of the

form
ℎ
∇ 𝑢 = 𝑓 𝑥 ,𝑦 + ∇ 𝑓 𝑥 ,𝑦 ,
12
20

30.07.2010
Example
Given
𝑢 +𝑢 =𝑥 𝑦 , 0 ≤ 𝑥, 𝑦 ≤ 𝜋
𝑢 0, 𝑦 = 0, 𝑢 𝜋, 𝑦 = 1
𝜕𝑢(𝑥, 0)
𝑢 𝑥, 𝜋 = 0, = 1.
𝜕𝑦
For 5 grid points in x-direction and 5-grid points in y-
direction,
• Derive the five-point FD equations of the PDE

• Derive the nine-point FD equations of the PDE
21


MATH5004-LEC2: Outline

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

MATH5004-LEC2: Outline

Uploaded by

Copyright:

Available Formats

30.07.

Curtin University is a trademark of Curtin University of Technology

Curtin University is a trademark of Curtin University of Technology

Definition of the global error

If 𝐄 ≤ 𝐶ℎ , 𝑝 > 0, C is a constant independent of h

L is the differential operator 𝐿 =

Curtin University is a trademark of Curtin University of Technology

For the DE 𝑢 = 𝑓(𝑥)

To check whether a FD scheme is consistent or not,

Curtin University is a trademark of Curtin University of Technology

As the consistence cannot guarantee a finite difference

A is the coefficient matrix of the FD equations,

For the central FD scheme, we have

Thus, the global error depends on both the local

Curtin University is a trademark of Curtin University of Technology

where C and ℎ are constants

Grid refinement analysis

We can divide h by half to get

Curtin University is a trademark of Curtin University of Technology

Improvement of Accuracy of Solution

Curtin University is a trademark of Curtin University of Technology

 Based on the above formulae, to accelerate the convergence of

Curtin University is a trademark of Curtin University of Technology

Curtin University is a trademark of Curtin University of Technology

From (1.1) and (1.2), by eliminating ℎ , we obtain

Curtin University is a trademark of Curtin University of Technology

 If p is unknown, it can be estimated from three approximate

Curtin University is a trademark of Curtin University of Technology

Curtin University is a trademark of Curtin University of Technology

from which, we have

Curtin University is a trademark of Curtin University of Technology

where the coefficients 𝑎, 𝑏, 𝑐 satisfy

Consider the problem

Curtin University is a trademark of Curtin University of Technology

Stencils for the 2D Laplacian

The five-point stencil The nine-point stencil

Accuracy and Stability: the five-point stencil

The global error 𝐸 = 𝑢(𝑥 , 𝑢 ) − 𝑢

Curtin University is a trademark of Curtin University of Technology

Accuracy and Stability: the nine-point stencil

It is noted that the error is still 𝑂 ℎ .

We can obtain a fourth-order accurate method of the

Curtin University is a trademark of Curtin University of Technology

• Derive the five-point FD equations of the PDE

Curtin University is a trademark of Curtin University of Technology

You might also like