Numerical Methods: Marisa Villano, Tom Fagan, Dave Fairburn, Chris Savino, David Goldberg, Daniel Rave

Numerical Methods
Marisa Villano, Tom Fagan,

Dave Fairburn, Chris Savino,
David Goldberg, Daniel Rave
An Overview
The Method of Finite Differences
Error Approximations and Dangers
Approxmations to Diffusions
Crank Nicholson Scheme
Stability Criterion
Finite Differences
Best known numerical method of
approximation
Marisa Villano
Finite Differences
Approximating the derivative with a

difference quotient from the Taylor series
Function of One Variable

Choose mesh size x
Then uj ~ u(jx)
First Derivative Approximations
Backward difference: (uj uj-1) / x
Forward difference: (uj+1 uj) / x
Centered difference: (uj+1 uj-1) / 2x
Taylor Expansion
u(x + x) = u(x) + u(x)x + 1/2 u(x)(x)

3
4
+ 1/6 u(x)(x) + O(x)
u(x x) = u(x) u(x)x + 1/2 u(x)(x)

3
4
- 1/6 u(x)(x) + O(x)
Taylor Expansion
u(x) = u(x) u(x x) + O(x)
x
u(x) = u(x + x) u(x) + O(x)
x
2
u(x) = u(x + x) u(x x) + O(x)
2x
Second Derivative Approximation
Centered difference: (uj+1 2uj + uj-1) / (x)
Taylor Expansion
2
u(x) = u(x + x) 2u(x) + u(x x) + O(x)

2
(x)
Function of Two Variables

n
u(jx, nt) ~ uj
Backward difference for t and x
u (jx, nt) ~ (ujn ujn-1) / t
t
u (jx, nt) ~ (ujn ujn-1) / x

x
Forward difference for t and x
u (jx, nt) ~ (ujn+1 ujn ) / t

t
u (jx, nt) ~ (ujn+1 ujn ) / x
x
Centered difference for t and x
u (jx, nt) ~ (ujn+1 ujn-1) / (2t)

t
u (jx, nt) ~ (ujn+1 ujn-1) / (2x)
x
Error
Truncation Error: introduced in the solution by the

approximation of the derivative
Local Error: from each term of the equation
Global
Error: from the accumulation of local
error
Roundoff Error: introduced in the computation by

the finite number of digits used by the computer
The Dangers of the Finite

Difference Method
Evidence from an example in 8.1
Dave Fairburn
Example from 8.1

Consider ut = uxx u(x,0) = h(x)
We will use the finite difference method to
approximate the solution
Forward difference for ut
Centered difference for uxx
Re-write equation in terms of the finite
difference approximations
Finite Difference Eqn.
ujn+1 - ujn = unj+1 - 2ujn + unj-1

t
( x) 2
Error: The local truncation error is O( t)

from the left hand side and is O( x)2 from
the right hand side.
Assumptions
Assume that we choose a small change in
x, and that the denominator on both sides
of the equation are equal.
We are now left with the scheme:
ujn+1 = unj+1 - unj + unj-1
Solving u with this scheme is now easy to
do once we have the initial data.
Initial Data
Let u(x,0) = h(x) = a step function with
the following properties:
h(x) = 0 for all j except for j = 5, so
hj = 0 0 0 0 1 0 0 0 0 0 0 .
Initially, only a certain section, which is
at j = 5 is equal to the value of 1.
j serves as the counter for the x
values.
How to solve?
We know u0j = 1 at j = 5 and 0 at all other j

initially (given by superscript 0).
We can plug into our scheme to solve for u1j at
all js.
u1j = u0j-1 - u0j + u0j+1
u15 = -1; u14 = 1; u16 = 1
Now we can continue to increase the # of
iterations, n, and create a table
Solution for 4 iterations

n
v
a
l
u
e
s
-4
10 -16 19 -16 10 -4
-3
-7
-3
-2
-2
-1
10
j values
Analysis of Solution
Is this solution viable?

Maximum principle states that the solution must
be between 0 and 1 given our initial data
At n = 4, our solution has already ballooned to u
= 19!
Clearly, there are cases when the finite
difference method can pose serious problems.
Charting the Error
Assume the solution is constant and equal to 0.5 (halfway between

the possible 0 and 1)
Lessons Learned
While the finite difference method is easy
and convenient to use in many cases,
there are some dangers associated with
the method.
We will investigate why the assumption
that allowed us to simplify the scheme
could have been a major contributor to
the large error.
Approximations of Diffusions
Neumann Boundary Conditions
and the Crank-Nicolson Scheme
Chris Savino
Errors have accumulated from the

approximations of the derivatives using the
previous scheme
The problem is the choice of the mesh t
to the mesh x
t
Let s= (x)2
can solve scheme
uj n 1 s(uj 1n uj 1n ) (1 2s)uj n
Neumann Boundary Conditions
ux(0, t ) g (t )
ux(l , t ) h(t )
Simplest Approximations are

u1n u 0 n
gn
x
uj n uj 1n
hn
x
To get smallest error, we use centered

differences for the derivatives on the boundary
Introduce ghost points
n
1
uj 1n
Boundary Conditions become
u u
2x
n
1
n
1
uj 1n uj 1n
hn
2x
Crank-Nicolson Scheme
Can avoid any restrictions on stability
conditions
Unconditionally stable no matter what the
value of s is.
Centered Second Difference:

uj 1n 2uj n uj 1n
2 n
u
)j
2
( x )
Pick a number theta between 0 and 1

Theta scheme:
uj n 1 uj n
(1 )( u 2 ) j n ( u 2 ) j n 1
t
We analyze the scheme by plugging in a

separated solution
uj n (eik x ) j ( (k )) n
Therefore
1 2(1 ) s (1 cos k x)
(k )
1 2 s (1 cos k x)
Must Check stability condition
If (k ) 1
Therefore
then
(k ) 1
s(1 2 )(1 cos k x) 1
1 2 0
is always true
1
1
2
If
then there is no restriction on the size
of s for stability to hold
The scheme is unconditionally stable
When theta = it is called the Crank-Nicolson
scheme
If theta < then the scheme is stable if
t
1
s
2
(x)
2 4
Stability Criterion
Approximations of the diffusion
equation, ut=uxx
David Goldberg
Stability Criterion
The method of finite differences gives an
answer, but it does not guarantee that this
answer is meaningful.
Values must be chosen appropriately, to
ensure that the results make sense and
are applicable to real world scenarios.
This condition, that values must satisfy in
order to be worthwhile, is called the
stability criterion.
Example
As per the book, take, for instance, the

diffusion problem:
=
for 0<x<, t>0
=0
at x=0, , that is
0, = , = 0
in
0,
2
, 0 = =
in ,
2
Example, continued
1.8
As can be easily shown, the graph of (x)

looks like this.
1.6
1.4
1.2
1
0.8
0.6
0.4
0.2
0
0
0.5
1.5
2.5
3.5
Example, continued
In attempting to use the method of finite
differences, we are using a forward
difference for ut and a centered difference
for uxx.
This means that +1 +1 2 + 1
It is important to note here that the

superscript n denotes a counter on the t
variable, and the subscript j denotes a
counter on the x variable.
Example, continued
In order to make the calculations a bit

cleaner, we are introducing a variable, s,
which is defined by s t
( x ) 2
Rearranging, we have
u nj 1 s u nj 1 2u nj u nj 1 u nj
u nj 1 s u nj 1 2s u nj u nj s u nj 1
u nj 1 s u nj 1 u nj 1 1 2s u nj
It would be nice if we could just plug in

values and get a valid result
Example, continued
However, putting in different values can lead to the

results being close to, or far from, that actual answer.
For instance, letting x=/20, and letting s=5/11, we get
a relatively nice result. Letting s=5/9 does not get such
a nice result.
2
1.5
1.5
0.5
0.5
0
0
So what, of significance, changes?
Example, Continued
As it turns out, changing the value of s can

significantly change the validity of the
solution. To see why, we return to our
...
equation.
u nj 1 s u nj 1 u nj 1 1 2s u nj
Seperate variables
u XT u nj X jTn
So, by combining like terms,
X j 1 X j 1
Tn 1
s
1 2s
Tn
Xj
Example, continued
Since the left hand side is a function of T

and the right hand side is a function of X,
they must be equal to a constant.
...
X j 1 X j 1
Tn 1
s
1 2s
Tn
Xj
Tn 1
Tn Tn 1 Tn nT0
Tn
and also
X j 1 X j 1
s
1 2 s
Xj
Example, continued
...
This is a discrete version of an ODE,

which when solved gives
1 2 s s(eik x e ik x )
1 2 s 2 s cos(k x)
Since, as discovered before, Tn nT0
if 1, T will grow without bound.
By above,
1 4s 1
1
So 1 4 s 1, s
2
Example, finished
t
Thus, to achieve stability, x

. This is
why setting s=5/9 didnt give a valid
result.
It is to be noted that usually the necessary
criterion is that 1 O(t ) instead of 1 , but
that in this case it was irrelevant.
So the stability criterion must be worked
out before one can effectively use the
method of finite differences.
Example from 8.2
Daniel Rave
Summary
Breif Review of Methods
Wide Applicability
Importance of Stability

Numerical Methods: Marisa Villano, Tom Fagan, Dave Fairburn, Chris Savino, David Goldberg, Daniel Rave

Uploaded by

Copyright:

Available Formats

Numerical Methods: Marisa Villano, Tom Fagan, Dave Fairburn, Chris Savino, David Goldberg, Daniel Rave

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

Numerical Methods: Marisa Villano, Tom Fagan, Dave Fairburn, Chris Savino, David Goldberg, Daniel Rave

Uploaded by

Copyright:

Available Formats

Numerical Methods

Marisa Villano, Tom Fagan,

Approximating the derivative with a

Function of One Variable

First Derivative Approximations

Backward difference: (uj uj-1) / x

Forward difference: (uj+1 uj) / x

Centered difference: (uj+1 uj-1) / 2x

u(x + x) = u(x) + u(x)x + 1/2 u(x)(x)

u(x x) = u(x) u(x)x + 1/2 u(x)(x)

Second Derivative Approximation

Centered difference: (uj+1 2uj + uj-1) / (x)

u(x) = u(x + x) 2u(x) + u(x x) + O(x)

Function of Two Variables

u (jx, nt) ~ (ujn ujn-1) / x

Function of Two Variables

Forward difference for t and x

u (jx, nt) ~ (ujn+1 ujn ) / t

Function of Two Variables

Centered difference for t and x

u (jx, nt) ~ (ujn+1 ujn-1) / (2t)

Truncation Error: introduced in the solution by the

Error: from the accumulation of local

Roundoff Error: introduced in the computation by

The Dangers of the Finite

Example from 8.1

Finite Difference Eqn.

ujn+1 - ujn = unj+1 - 2ujn + unj-1

Error: The local truncation error is O( t)

the right hand side.

We know u0j = 1 at j = 5 and 0 at all other j

Solution for 4 iterations

Is this solution viable?

Charting the Error

Assume the solution is constant and equal to 0.5 (halfway between

Errors have accumulated from the

can solve scheme

Neumann Boundary Conditions

Simplest Approximations are

To get smallest error, we use centered

Boundary Conditions become

Centered Second Difference:

Pick a number theta between 0 and 1

We analyze the scheme by plugging in a

Must Check stability condition

s(1 2 )(1 cos k x) 1

As per the book, take, for instance, the

for 0<x<, t>0

As can be easily shown, the graph of (x)

It is important to note here that the

In order to make the calculations a bit

It would be nice if we could just plug in

However, putting in different values can lead to the

So what, of significance, changes?

As it turns out, changing the value of s can

Since the left hand side is a function of T

This is a discrete version of an ODE,

Thus, to achieve stability, x