Vous êtes sur la page 1sur 4

Numerical Partial Differential Equations Notes: 10/30/2014

Goal: Further discussion of the CFL condition.


We want to obtain better accuracy. To do this, we could try higher order
N
N
to create:
, UjN , Uj+1
interpolations. For example, we could use Uj1
N
p(x) = Uj1

(x xj1 )(x xj+1 )


(x xj )(x xj+1 )
(x xj )(x xj+1 )
N
+UjN
+Uj+1
(xj1 xj )(xj1 xj+1 )
(xj xj1 )(xj xj+1 )
(xj+11 xj1 )(xj+1 xj )

Then our new nuermcial method would be:


N (hak)(ak)
N (ak)(hak)
+ UjN (hak)(hak)
+ Uj+1
Ujn+1 = p(xj ak) = Uj1
(h)(2h)
h(h)
(2h)h
2 2

N ak
N a k
N
Ujn+1 = Uj1
2h + Uj1 2h2 + Uj

Ujn+1 = Uj

ak
N
2h (Uj+1

N
Uj1
)+

a2 k2 N
h2 Uj

a2 k2
N
2h2 (Uj1

ak N
2h Uj+1

a2 k2 N
2h2 Uj+1

N
2UjN + Uj+1
)

Note that this creates the Lax-Wendroff method!


We need that stability to be such that 1

ak
h

1 for xj1 xj ak xj+1

The point of this exervise is to show a general case of what is called the CFL
condition. Note in the above case, what we are essentially doing is creating a
three point interpolation of the step u(x ak, h) and then simplifying it in order
to get the Lax-Wendroff method.
What is the domain of dependence for the time-dependent PDE?
The mathematical domain of dependence of u(x, t) denot by D(x, t), is the
set of all points in space where the initial data at t 0 may have some effect
on the solution of u(x, t).
Say that you have a point in the future at a particular space. Displayed
below is a graphical representation of it:

Example:
For the linear advection equation
ut + aux = 0
We know hat the solution is:
u(x, t) = 0 (x a, t)
where 0 is the initial data.
Then D(x, t) = {X aT }
Example:
ut + Aux = 0
Suppose that A has a real and distinct set of eigenvalues: p , p = 1, 2, 3...s
Then we wold say that the domain of dependence is:
D(x, t) = {X p T, p = 1, 2, 3...s
Another example: The Heat Equation:
ut = uxx , u(x, 0) = u0 (x)
D(x, t) = (, )
The analytical solution for this is:
u(x, t) =

1
4t

e(x2)

/4t

u0 (x)dx

Infinite propogation speed. The data at any point affects the solution everywhere at any small time in the future.
We need to make sure that the numerical method captures this domain of dependence. If we are missing any of the data that influences the solution at a
later time then it wont get a correct answer because we will be eliminating
important information.
Note to self, be sure to study the characteristic method of lines that she
went over I am not entirely clear what is going on for it outside the use of the
Lagrange interpolant.
How do we do this?
Do we use the characterics for nonhyperbolic PDEs ?
We only use the characteristic if we only need to know a discrete number of
points behind you. We can assume that we would use characteristics for hyperbolic PDEs but we would not use them fr non-hyperbolic.

What is a numerical domain of dependence?


Each finite difference method D0 (x, t) is defined as:
D0 (x, t) = { all grid points x i at the initial point u0i has an affect on unj }
Lets calculate numerical domain of dependence for Lax-Wendroff method.
(1) Recall that the CFL condition is that D(x, t) D0 (x, t) for the limit as
k, h 0

laxwendroff.jpg

The stencil we will be using is:

0
0
Therefore, Ujn+1 depends on Uj(n+1)
, ..., Uj+(n+1))
Note: our result is the
same as the stability condition!

The CFL condition for the forward time centered space solution to the heat
equation behaves in a similar manner. We take three different point for N, N-1,
N+1 and propogate them base to the boundary. So this also has the same domain of dependence of the initial value.
Another note: for the above two examples we need r to go to zero so that
thedomain (our bracketed term ) will explode and take up the entire line.
In order for D0 (x, t) D(x, t) in the limit as k, h 0, we need the limit
as k and h approach zero of 1r =
Therefore, we also need for the limit of k,h as they approach zero, we need
the valuesin the relation hk = to be such that h is greater than k IE k decays to
zero faster than h does.
Passing the CFL condition does not guarantee that your method is stable. If satisfied, then the method might be convergent. We mainly use this to weed out
the methods that arent stable.

Vous aimerez peut-être aussi