Finite Element Method I Finite Element Method II

Example Definition (Finite Element Method)

Lets consider the O.D.E. on the interval t ∈ ∆ = [0; 1] Consider an equation (O.D.E.) f ( t, y, y′ ) = 0 of a function y( t) defined on the
interval t ∈ ∆:
f ( t, y, y′ ) = y′ ( t) + y( t) = 0 1 Define a trial solution of the form:
n
with the initial condition y(0) = 1. ỹ( t) = φ0 +
X
a i φ i ( t)
i =1
We can solve this O.D.E.:
where the functions {φ i ( t)} are basis functions, and φ0 refers to a constant
analytically giving the exact solution:
that usually can be inferred from initial conditions.
′
y( t) = exp(− t) 2 Replace y( t) by ỹ( t) in the O.D.E. f ( t, ỹ( t), ỹ ( t)) = R ( t, a) (R is called the
residual) and solve the coefficients {a i } i=1···n .
numerically giving an approximate solution using:
The estimated ỹ( t) gives a numerical solution to the O.D.E. defined on the
Ï Finite Difference Method,
interval ∆.
Ï Finite Element Method.

Solving y′ + y = 0 using collocation method I Solving y′ + y = 0 using collocation method II

Definition (Collocation method) Example

The collocation method imposes that the residual vanishes at n points i.e. Lets take t1 = 31 and t2 = 32 such that we impose R ( t1 , a) = R ( t2 , a) = 0. This gives
R ( t i ; a) = 0 for t i ∈ ∆ and i = 1, · · · , n. us 2 equations: ³ ¡ ¢2 ´
1 + 1 + 31 a 1 + 23 + 13
 ¡ ¢

 a2 = 0

Example  ³ ¡ ¢2 ´
1 + 1 + 32 a 1 + 43 + 23
 ¡ ¢
a2 = 0

i
We choose the basis φ i ( t) = t and n = 2 so:
Solving in a 1 and a 2 gives:
ỹ( t) = 1 + a 1 φ1 ( t) + a 2 φ2 ( t) = 1 + a 1 t + a 2 t2
27 9 2
ỹ( t) = 1 − t+ t
29 29
and
R ( t, a) = 1 + (1 + t) a 1 + (2 t + t2 ) a 2 Compare that with the true function y( t) = exp(− t).

Sub-domain method I Sub-domain method II

Example
We choose ∆1 = 0; 21 and ∆2 =
£ ¤ £1 ¤
2;1 and :
· ¸
Definition (subdomain method) 1 1 5 7
Z
R ( t; a) dt = 2 + a1 + a2 = 0
∆1 ∆1 2 8 24
Divide the interval ∆ of definition of the O.D.E. into subdomains {∆ i } i=1,··· ,n , and
and
solve a such that : 1
Z ·
1 7 25
¸
1
Z
R ( t; a) dt = 2 + a1 + a2 = 0
R ( t; a) dt = 0 ∀ i = 0, · · · , n ∆2 ∆2 2 8 24
∆i ∆i
Solving this system gives the solution
18 6 2
ỹ( t) = 1 − t+ t
19 19

Compare ỹ( t) estimated using subdomain method, with the exact solution y( t).
Least-Squares Method I Least-Squares Method II

Definition (Least-Squares Method)

We can use least square estimation to find the optimum values of the parameters
a: ½Z ¾
â = arg min (R ( t; a))2 dt Example
∆

or solve Exercise: Try to find ỹ with the Least squares method.

∂
Z
(R ( t; a))2 dt = 0 ∀i
∂a i ∆
or
∂
Z
(R ( t; a))2 dt = 0 ∀i
∆ ∂a i
or
∂R ( t; a)
Z
2 R ( t; a) dt = 0 ∀i
∆ ∂a i

Galerkin Method I Exercises

Definition (Galerkin Method) Solve on the interval [0; 1] the O.D.E.:

The Galerkin method imposes the following criterium to estimate a:
y′′ ( t) + y( t) = 1
Z
R ( t; a) φ i ( t) dt = 0 ∀ i = 1, · · · , n
∆ with the boundary conditions y(0) = 1 and y(1) = 0:
1 analytically.
Example 2 numerically using
1 the collocation method
Find that the estimate of ỹ( t) using Galerkin method is:
2 the subdomain method
32 2
ỹ( t) = 1 − t + t2 3 the Galerkin method
35 7
3 Draw all the solutions you found on a graph (use matlab).
Compare that solution with the exact one y( t).

Method of Weighted residuals I Method of Weighted residuals II

Weighting function Wi for:
1 the Collocation method:
Definition (Weighted residuals)
Wi ( t) = δ( t − t i )
In fact all the previous method (collocation, galerkin, subdomain, Least squares) 2 the Sub-domain method:
can be understood as special cases of the Method of Weighted residuals (MWR). (
The notion of MWR is to force the residual R ( t; a) to zero in some average sense 1 if t ∈ ∆ i
Wi ( t) =
over the domain. That is: 0 otherwise
Z
3 the Least Squares method:
R ( t, a) Wi ( t) dt = 0 i = 1, · · · , n
∆
∂R ( t; a)
Wi ( t) = 2
where the number of weight functions Wi ( t) is exactly equal to the number of ∂a i
unknown constant a i in ỹ. 4 the Galerkin Method:
∂ ỹ
Wi ( t) = = φ i ( t)
∂a i
Remarks on FDM and FEM Which basis functions? I

Finite Difference Method: FDM
approximates an operator (e.g., the
derivative) and solves a problem on a
set of points (the grid).
Finite Element Method: FEM uses
exact operators but approximates
the solution basis functions. Also,
FE solves a problem on the interiors
of grid cells (and optionally on the
gridpoints as well).
Amongst the methods of weighted residuals, the Galerkin method is the most
commonly used.
MWR apply equally well to any family of basis functions {φ i }. However to be
sucessful, the basis functions need to be chosen with care
Ï Global basis functions perform badly in fitting local behaviour, e.g.:

φ i (t) = t i

Which basis functions? II Which basis functions? III

Ï More local basis functions can be defined so that there are non-zero only on
small regions, e.g. piecewise linear (or hat) functions:

t− t i−1



 t i − t i−1 t i−1 ≤ t < t i






t i+1 − t
φ i (t) = t i+1 − t i t i ≤ t < t i+1









0 otherwise