You are on page 1of 3

Lecture 1 1 Thu, 12/01/12

1 FEM Basics – Model problem • How to I know the corresponding numerical prob-
lem is well posed?
and variational formulation
• In what sense are the strong and weak forms equiv-
1.1 What is this? alent?
This is a graduate level course on the numerical solu- • What is a “weak” derivative, anyway?
tion of boundary value problems. We will be mainly
concerned with the solution by finite element methods, • How well can a function be approximated by piece-
although finite difference schemes will also be discussed. wise polynomials? Any how should this be mea-
For information of recommended text, resources, sched- sured?
ule, etc., please see the webpage:
• How can I turn this theory into a programmable
http://www.maths.nuigalway.ie/˜niall/MathsOfFEM numerical method?

These notes are intended just to provide a summary • And many more...
of what was covered in each class; they are incomplete.
Furthermore, the notes are not structured like a text- The rest of the notes in this section are based on ones
book. The development is not linear. Instead ideas are I’ve written for other courses over a number of years,
introduced, but only formally justified much later. This They owe more than a little to [?, Chap. 14]. Notes
is because, in my experience, having a strong motivation for other lectures in this course will not be nearly as
for a particular piece of theory, and a heuristic under- detailed.
standing of it, facilitates later deep learning. This is,
it is OK to defer technical details until a motivation is 1.4 Model boundary value problem
established, so long as this is done knowingly.
1.4.1 One dimension

1.2 Mathematical Preliminaries To formulate a model BVP we first define a differential


operator
The idea preparation for this course is to have taken up- L(u) := −u ′′ (x) + r(x)u(x). (1.1)
per undergraduate courses in linear algebra, functional
analysis, measure theory, numerical analysis, and dif- Then the general form of a BVP is: find a function u
ferential equations. However, since participants in this defined on the interval [a, b]
course have very different backgrounds, I’ll try to cover
L(u) = f(x) for 0 < x < 1,
all the fundamentals in class. (1.2)
u(0) = u(1) = 0.

1.3 Initial plan... For today we will make the assumptions that r and f are
continuous and have as many continuous derivatives as
Over the new two lectures, I’ll give a whirlwind tour of
we would like. Furthermore (today and every day) we
finite element methods. This will include
have that r(x) > 0 for all x ∈ (0, 1).
• Introducing boundary value problems in one, two
and n dimensions; 1.4.2 Two and more dimensions

• Variational (weak) formulations; Although we will concentrate on ODEs initially, our real
interest lies in PDEs. The simplest example is probably
• Best approximation;
Laplace’s equation:
• Piecewise polynomials;
¡ ∂2 ∂2 ¢
− + u=0 on Ω = (0, 1)2 .
• Implementation. ∂ x2 ∂ y2
The goal is not to present complete, correct theory, It can be written more succinctly as
but to motivate the deeper issues that we will study over
the next few weeks such as : −(uxx + uyy ) = 0 on Ω.

• Under what conditions is a boundary value prob- −∆u = 0 on Ω.


lem well posed?
Lecture 1 2 Thu, 12/01/12

Poisson’s equation is: (c) So we choose a much smaller subset — one with a
finite basis of size N — and look for the solution
−∆u = f(x, y) on Ω.
there.
Much of my own professional interest is in problems of
(d) Write down N versions of the integral equation that
the form:
must be satisfied.
−ε∆u + r(x, y)u = f(x, y) on Ω,
(e) We have N equations and N unknowns, so we can
(a “reaction-diffusion” problem), and solve to find the approximate solution.
−ε(uxx + uyy ) + a(a, y)ux = f(x, y) on Ω, (f) A simple but clever idea shows that the approxima-
(“convection-diffusion” or “advection-diffusion” problem). tion we find is the best possible one.

1.5 Solutions 1.7 Where the solution lives

We return our attention to ODEs. Some are reasonably The boundary value problem (1.2) above equates the
easy to solve by hand. That is, to write down a solution terms in u and u ′′ with a function f that is continuous
in terms of elementary functions. That is certainly the on (0, 1), so we must have that u, u ′ and u ′′ are all
case if the functions r and f are constant: continuous on the interval (0, 1). That is, u belongs to
the space of functions C2 (0, 1).
Example 1.1. Find the general solution to the DE:
So we can state the problem (1.2) more precisely:
−u ′′ (x) + 4u(x) = x for 0 6 x 6 3, (1.3) find u ∈ C2 (0, 1) such that

Unlike the example above, most boundary value prob- −u ′′ (x) + r(x)u(x) = f(x) for all x in (0, 1),
lems are difficult or impossible to solve analytically. (It’s
and u(0) = u(1) = 0.
not hard to check if a given u is a solution. We’ll come
back to this later.) So we need approximations. As mentioned previously, there are (uncountably) many
Two of the most popular numerical methods for con- functions in the space C2 (0, 1) (we say it is infinite di-
structing approximate solutions to BVPs are mensional ) – far too many to check one-by-one. So we
choose small subset of C2 (0, 1) that has only finitely
• Finite Difference Methods (FDMs) — where one
many members. The FEM will let us find the element
approximates the differential operator L, based on
of that subset that is “closest” to the true solution.
Taylor series expansions. We’ll discuss these for
one or two lectures.
1.8 The Weak Form
• Finite Element Methods (FEMs) – where to solu-
Before we give a numerical method for computing an
tion space is approximated, using ideas related to
approximate solution to a boundary value problem, we
interpolation. the main
will rewrite in as an integral equation.
Consider the Boundary Value Problem: find u ∈
1.6 Finite Element Methods 2
C (0, 1) such that
The key sequence of ideas for FEMs is
−u ′′ (x) + r(x)u(x) = f(x) on (0, 1),
(a) First replace the differential equation with an in- (1.4)
u(0) = u(1) = 0.
tegral equation, using integration by parts (in one
dimension) to reduce the order of the derivatives. If u solves this then,
This is called the “Variational ” or “weak ” form of
−u ′′ (x) + r(x)u(x) = f(x) on (0, 1).
the equation. We take care to note the space of
functions where the solution lives. So, for any (reasonable) function v,
(b) We could try to solve the problem by taking each −u ′′ (x)v(x) + r(x)u(x)v(x) = f(x)v(x), on (0, 1).
function from the space in turn, and substitute it
into the equation to see if it is a solution. But the Integrating both sides we get
set of possible solution is infinitely large, so we can’t Z1 Z1 Z1
check them all. − u ′′ (x)v(x)dx + r(x)u(x)v(x)dx = f(x)v(x)dx.
0 0 0
Lecture 1 3 Thu, 12/01/12

Using integration by parts:


Z1 1
Z1
− u ′′ (x)v(x)dx = −u ′ (x)v(x) + u ′ (x)v ′ (x)dx.
0 0 0

Now lets suppose that v(0) = v(1) = 0. So now the


problem is: Find u such that
Z1 Z1 Z1
′ ′
u (x)v (x)dx + r(x)u(x)v(x)dx = f(x)v(x)dx.
0 0 0

for all v.
Now lets introduce a little notation:
R1
• the inner product: (u, v) := 0 u(x)v(x)dx; and
the induced norm: kuk2 := (u, u)1/2 ,

• and H10 (0, 1) to be the space of (absolutely) contin-


uous functions on [0, 1] such that if w ∈ H10 (0, 1)
then w(0) = w(1) = 0 and kw ′ (x)k2 < ∞.

So now we have a more succinct statement: Find u ∈


H10 (0, 1) so that

(u ′ , v ′ ) + (ru, v) = (f, v) for all v ∈ H10 (0, 1).

However, in (1.4) we required that u be twice dif-


ferentiable; that seems unnecessary here where we have
the much weaker requirements that u ′ exist and be in-
tegrable.

Definition 1.1. The weak/variational formulation of


(1.4) is: Find u ∈ H10 (0, 1) such that

B(u, v) = l(v) for all v ∈ H10 (0, 1). (1.5)

where B(·, ·) is the (symmetric) bilinear functional

B(u, v) := (u ′ , v ′ ) + (ru, v).

and l(v) is the linear functional

l(v) = (f, v).

You might also like