FEniCS Course

Lecture 0: Introduction to FEM

Contributors
Anders Logg

1 / 33
What is FEM?

The finite element method is a framework and a recipe for

discretization of differential equations

• Ordinary differential equations

• Partial differential equations

• Integral equations

• A recipe for discretization of PDE

• PDE → Ax = b

• Different bases, stabilization, error control, adaptivity

2 / 33
The FEM cookbook

(i)

Au = f Partial differential equation

(ii)
v
by a(u, v) = L(v) Continuous variational problem
iply
Mult
(iii)
V a(uh , v) = L(v) Discrete variational problem
⊂
Vh
ake (iv)
T
φ j
Uj AU = b System of discrete equations
j
P
=
tu
h
Le

3 / 33
The PDE (i)
Consider Poisson’s equation, the Hello World of partial
differential equations:

−∆u = f in Ω
u = u0 on ∂Ω

Poisson’s equation arises in numerous applications:

• heat conduction, electrostatics, diffusion of substances,

twisting of elastic rods, inviscid fluid flow, water waves,

magnetostatics, . . .

• as part of numerical splitting strategies for more

complicated systems of PDEs, in particular the

Navier–Stokes equations
4 / 33
From PDE (i) to variational problem (ii)
The simple recipe is: multiply the PDE by a test function v and
integrate over Ω:
Z Z
− (∆u)v dx = f v dx
Ω Ω

Then integrate by parts and set v = 0 on the Dirichlet

boundary:
Z Z Z
∂u
− (∆u)v dx = ∇u · ∇v dx − v ds
Ω Ω ∂n
| ∂Ω {z }
=0

We find that: Z Z
∇u · ∇v dx = f v dx
Ω Ω

5 / 33
The variational problem (ii)

Find u ∈ V such that

Z Z
∇u · ∇v dx = f v dx
Ω Ω

for all v ∈ V̂

The trial space V and the test space V̂ are (here) given by

V = {v ∈ H 1 (Ω) : v = u0 on ∂Ω}
V̂ = {v ∈ H 1 (Ω) : v = 0 on ∂Ω}

6 / 33
From continuous (ii) to discrete (iii) problem

We approximate the continuous variational problem with a

discrete variational problem posed on finite dimensional
subspaces of V and V̂ :

Vh ⊂ V
V̂h ⊂ V̂

Find uh ∈ Vh ⊂ V such that

Z Z
∇uh · ∇v dx = f v dx
Ω Ω

for all v ∈ V̂h ⊂ V̂

7 / 33
From discrete variational problem (iii) to
discrete system of equations (iv)
Choose a basis for the discrete function space:

Vh = span {φj }N
j=1

Make an ansatz for the discrete solution:

N
X
uh = Uj φj
j=1

Test against the basis functions:

Z N
X Z
∇( Uj φj ) · ∇φi dx = f φi dx
Ω j=1 Ω
| {z }
uh

8 / 33
From discrete variational problem (iii) to
discrete system of equations (iv), contd.
Rearrange to get:
N
X Z Z
Uj ∇φj · ∇φi dx = f φi dx
j=1 |Ω {z } | Ω {z }
Aij bi

A linear system of equations:

AU = b
where
Z
Aij = ∇φj · ∇φi dx (1)
ZΩ
bi = f φi dx (2)
Ω
9 / 33
The canonical abstract problem
(i) Partial differential equation:

Au = f in Ω

(ii) Continuous variational problem: find u ∈ V such that

a(u, v) = L(v) for all v ∈ V̂

(iii) Discrete variational problem: find uh ∈ Vh ⊂ V such that

a(uh , v) = L(v) for all v ∈ V̂h

(iv) Discrete system of equations for uh = N

P
j=1 Uj φj :

AU = b
Aij = a(φj , φi )
bi = L(φi )

10 / 33
Important topics

• How to choose Vh ?

• How to compute A and b

• How to solve AU = b?

• How large is the error e = u − uh ?

• Extensions to nonlinear problems

11 / 33
How to choose Vh

12 / 33
Finite element function spaces

uh

13 / 33
The finite element definition (Ciarlet 1975)

A finite element is a triple (T, V, L), where

• the domain T is a bounded, closed subset of Rd (for

d = 1, 2, 3, . . . ) with nonempty interior and piecewise

smooth boundary

• the space V = V(T ) is a finite dimensional function space

on T of dimension n

• the set of degrees of freedom (nodes) L = {`1 , `2 , . . . , `n } is

a basis for the dual space V 0 ; that is, the space of bounded
linear functionals on V

14 / 33
The finite element definition (Ciarlet 1975)

T V L

v(x̄)

v(x̄) · n
R
T v(x)w(x) dx

15 / 33
The linear Lagrange element: (T, V, L)

• T is a line, triangle or tetrahedron

• V is the first-degree polynomials on T

• L is point evaluation at the vertices

16 / 33
The linear Lagrange element: L

17 / 33
The linear Lagrange element: Vh

uh (x, y)

T = {T }

18 / 33
The quadratic Lagrange element: (T, V, L)

• T is a line, triangle or tetrahedron

• V is the second-degree polynomials on T

• L is point evaluation at the vertices and edge midpoints

19 / 33
The quadratic Lagrange element: L

20 / 33
The quadratic Lagrange element: Vh

21 / 33
Families of elements

22 / 33
Families of elements

23 / 33
Computing the sparse matrix A

24 / 33
Naive assembly algorithm

A=0

for i = 1, . . . , N

for j = 1, . . . , N

Aij = a(φj , φi )

end for

end for

25 / 33
The element matrix

The global matrix A is defined by

Aij = a(φj , φi )

The element matrix AT is defined by

AT,ij = aT (φTj , φTi )

26 / 33
The local-to-global mapping
The global matrix ιT is defined by

I = ιT (i)

where I is the global index corresponding to the local index i

4 4

3 3
8
7
5 9 4
4
9 5
1
1
6 2
6
5 2 1 3
6
2 2

27 / 33
The assembly algorithm

A=0

for T ∈ T

Compute the element matrix AT

Compute the local-to-global mapping ιT

Add AT to A according to ιT

end for

28 / 33
Adding the element matrix AT

ι2T (1) ι2T (2) ι2T (3) 1 2 3

ι1T (1) 2

3
ι1T (2)

AT,32
ι1T (3)

29 / 33
Solving AU = b

30 / 33
Direct methods

• Gaussian elimination
• Requires ∼ 2 3
3N operations

• LU factorization: A = LU
• Solve requires ∼ 2 3
3N operations
• Reuse L and U for repeated solves

• Cholesky factorization: A = LL>

• Works if A is symmetric and positive definite

• Solve requires ∼ 1 3
3N operations
• Reuse L for repeated solves

31 / 33
Iterative methods
Krylov subspace methods
• GMRES (Generalized Minimal RESidual method)
• CG (Conjugate Gradient method)
• Works if A is symmetric and positive definite

• BiCGSTAB, MINRES, TFQMR, . . .

Multigrid methods
• GMG (Geometric MultiGrid)

• AMG (Algebraic MultiGrid)

Preconditioners
• ILU, ICC, SOR, AMG, Jacobi, block-Jacobi, additive

Schwarz, . . .
32 / 33
Which method should I use?

Rules of thumb

• Direct methods for small systems

• Iterative methods for large systems

• Break-even at ca 100–1000 degrees of freedom

• Use a symmetric method for a symmetric system

• Cholesky factorization (direct)

• CG (iterative)

• Use a multigrid preconditioner for Poisson-like systems

• GMRES with ILU preconditioning is a good default choice

33 / 33