Presentation Draft FMEFM

A Survey on a Mixed Finite Element Method
David Sacco
December 8, 2016
David Sacco Mixed Finite Element Method December 8, 2016 1 / 26

What is FEM?
Start with a PDE you wish to solve (in our case, Laplace’s
equation)
Use a specific (or multiple) Hilbert Space to approximate solutions
to PDE’s
Find the variational problem to the model problem.
Assuming the variational problem has an existing unique solution,
we use the basis of a subspace of our Hilbert Space(s) to rewrite
the weak form into a discrete form.
Write down the discrete form into a matrix equation, then solve
numerically with a triangulation of the domain.

What is FEM?
For this project, we aim to solve the following problem numerically via
a mixed method.
Model Problem
Let Ω ⊂ R2 be a bounded polygonal domain. We consider the following
problem.
−∆u = f in Ω, u = 0 on ∂Ω
where f ∈ L2 (Ω).

What makes it a Mixed Problem is by introducing a new variable
σ = −∇u to rewrite the Model Problem.
“Mixed” Model Problem
Let Ω ⊂ R2 be a bounded polygonal domain, and σ = ∇u. Then the
model problem becomes
∇ · σ = f and σ = −∇u in Ω, u = 0 on ∂Ω
where f ∈ L2 (Ω).

The Approximation Spaces
We need to have two Hilbert Spaces, since we wish to approximate u
and σ. For σ, we use
H(div; Ω) = p ∈ L2 (Ω) : ∇ · p ∈ L2 (Ω)

In order to approximate u, we will be using a subspace of L2 (Ω). The

approximation spaces are then the so called Raviart-Thomas Subspace
RT0 (Ω) ⊂ H(div; Ω), and space of piecewise constant polynomials on
T , Qh ⊂ L2 (Ω).
The Approximation Space for σ
Let a ∈ R2and b ∈ R, then
RT0 (T ) = p ∈ L2 (T ) : p(x) = a + bx, and [p]E · νE = 0

for all T , x ∈ T , and all E ∈ EΩ . Here, EΩ is the set of all interior edges
of the triangulation, [p]E = p|T+ − p|T− is the jump of p across the
edge shared by the two neighboring triangular elements, and νE is the
outward pointing normal at the midpoint of the edge E.
The Raviart-Thomas Element
Basis for RT0 (Ω)
The basis functions for the Raviart-Thomas Space, RT0 (Ω), are given
by (
|E|
± 2|T ±|
(x − P± ) for x ∈ T±
ψE (x) =
0 elsewhere
The main property of these functions is that ψE · νE = 1 along the

midpoint of an edge of a given triangular element, T , and ∇ · ψ is
constant in T± . Here, we choose the orientation of νE to always point
from T+ to T− . This determines the sign of ψ, negative only when
pointing inwards.
David Sacco Figure 1

Mixed Finite Element Method December 8, 2016 6 / 26
Properties of the Triangulation
Let Ω be a polygonal domain, we require that the triangulation Th be a
quasi-uniform triangulation. Define h = max{diam(T ) : T ∈ Th }.
Quasi-uniformity of T
Let Ω be a polgyonal domain. Th is said to be quasi uniform if there
exists ρ > 0 such that
min{diam(BT ) : T ∈ Th } ≥ ρhdiam(Ω)
where BT is the largest ball contained in T .

Error Estimates
We regard um m
h and σh to be the finite element method solutions, and
k · ks the L2 norm when s = 0, and the H s norm for s > 0.
Theorem 1
Let Ω be a polygonal domain. For the axillary function σ corresponding
to mixed model problem, we have that
||σ − σhm ||0 ≤ Ch||σ||1 ≤ Ch||u||2
Where u is the solution to the model problem, and C > 0 is

independent of h.

Error Estimates
Theorem 2
Let Ω be a polygonal domain. For the solution u to the model problem,
with f the forcing term, and um
h the mixed finite element solution, then
ku − um 2
h k0 ≤ C(hkuk1 + h (kuk2 + kf k1 ))
For some C > 0 independent of h.
The proofs of these theorems require properties of the subspaces RT0

and Qh , and properties involving the Galerkin Method solution, uG
h.

Variational Formulation
Manipulating the mixed model problem, we may arrive to the following.

The Variational Problem
For given f ∈ L2 (Ω), find σ ∈ H(div; Ω) and u ∈ L2 (Ω) such that
Z Z
σ · q dx + u∇ · q dx = 0 ∀q ∈ H(div; Ω)
Ω Z Ω Z
v∇ · σ dx = vf dx ∀v ∈ L2 (Ω)
Ω Ω
Here, q and v may be regarded as test functions.

Discrete Formulation
Now, we write the problem framed in the approximation spaces, RT0

and Qh .
The Discrete Problem
For given f ∈ L2 (Ω), find σh ∈ RT0 (Th ) and uh ∈ Qh such that
Z Z
σh · qh dx + uh ∇ · qh dx = 0 ∀qh ∈ RT0
Ω ZΩ Z
vh ∇ · σh dx = vh f dx ∀vh ∈ Qh
Ω Ω
Here, qh and vh may be regarded as test functions in their respective

spaces.

The Linear System
We must use the basis of RT0 to now rewrite the discrete problem into
a system of linear equations. Suppose in a triangulation of Ω, that
there are M interior edges, and L triangular elements,
Th = (T
P1 , T2 , ..., TL ). By properties of the subspace’s basis, we have
σh = M k=1 xk ψk , so we may rewrite the discrete problem.
Linear System
M
X Z L
X Z
xk ψj · ψk dx + xM +l ∇ · ψj dx =0
k=1 Ω l=1 Tl
M
X Z Z
xk ∇ · ψk dx = f dx
k=1 Tl Tl
with j = 1, 2, ..., M and l = 1, 2, ..., L.

The Linear System
From the linear system, we have M + L equations with M + L

unknowns. Let xψ = (x1 , x2 , ..., xM ), and xu = (xM , xM +1 , ..., xM +L ),
then we may cast the linear system into the matrix equation

B C xψ 0
=
CT 0 xu bf
R
Here, bf comes from the approximation of the integral T f dx on each
triangle. B and C are blocks, representing each of the sums from the
Linear System. The key to their computation is in the evaluation of
the integrals, which may be computed exactly.

In order to implement the mixed finite element method, we need the
following:
A quasi-uniform triangulation
keep track of the various data – edges, nodes, elements, boundary
conditions
Form a relationship (function) between the data
Build a Local Stiffness Matrix to construct the Global Stiffness
Matrix
Approximate the right hand side
Solve the matrix equation

For the triangulation, we use an external package, iF EM , to generate
our mesh. The package’s squaremesh function yields the following
triangulation that we will use.

The triangulation function squaremesh yields files coordinate.dat
and element.dat, which consists of the coordinates of nodes, as
well as numbering of the elements.
In our figure the node 1 has the coordinates (0, 0), while element 1
is composed of nodes {(1, 4, 5)}.
dirichlet.dat is also generated from iF EM , and enumerates the
edges that are on the boundary. Edges are just denoted by pairs of
nodes. For example, { (1, 4) } would be a row in the file.

Nodes to Elements Matrix
Let N be the number of nodes, then nodes2element is a N × N sparse

matrix assigning pairs of nodes to a triangular element. We define it by
Matrix Definition
For a pair of nodes (k, l), and element j,

j if (k, l) are nodes of the element number j
nodes2element(k, l) =
0 otherwise
For us, N 2E(1, 5) = 1, but note N 2E(5, 1) = 5, by definition N 2E(k, l)

is assigned to T+ , while N 2E(l, k) is assigned to T− .

Nodes 2 Edges Matrix
nodes2edge is a N × N sparse symmetric matrix, assigning nodes to

edges.
Matrix Definition
For a pair of nodes (k, l),

j if edge Ej is connected by nodes (k, l)
nodes2edge(k, l) =
0 otherwise
To create this matrix, let B = N 2E + N 2E T , so that every pair of

nodes is assigned a number in B. Since B is symmetric, find all of the
indices in the upper triangular portion where B holds a nonzero value.
These nodes that hold a nonzero value in B are edges, are then
assigned number 1 through M in the sparse matrix at those indices.

Edge to Element Matrix
edge2element is a M × 4 sparse matrix, assigning edges (pairs of

nodes) to elements. Each row of this matrix consists of a pair of nodes
(k, l), with a pair of elements (m, n). Recall in matlab notation that
arrays may be input to a matrix to focus on a submatrix.
Matrix Definition
Let (k,l) be labels of nodes, (m, n) be labels of elements. Then

[m, n] if edge j belongs to elements [m, n]
edge2element(j, [3, 4]) =
[m, 0] if edge j belongs to only m
This definition focuses on the 3rd and 4th columns of Ed2El. It also is
worth noting that this defines elements in the 3rd column as T+ , while
the 4th column describes T− . This also tells how to choose the ± in the
basis functions.

Local Stiffness Matrices
It is customary to write a smaller matrix to ‘generate’ the bigger

matrix. Denote BL and CL the local 3 × 3 matrices for B and C, and
|T | the area of T , |E| the length of E. For a given triangular element
T , with nodes {P1 , P2 , P3 }. Let
 
2 0 1 0 1 0
0 2 0 1 0 1  

1
 0 P1 − P2 P1 − P3
0 2 0 1 0
M := 
0 and N := P2 − P1 0 P2 − P3 
1 0 2 0 1

1
 P3 − P1 P3 − P2 0
0 1 0 2 0
0 1 0 1 0 2
1 T T
Then we may write BL = 48|T | CL N M N CL .

Local Stiffness Matrices
R
CL requires us to compute ∇ · ψ dx. Fortunately ∇ · ψ is constant,
T
since ( |E|
± |T± | on T±
∇ · ψE =
0 elsewhere

Global Stiffness Matrix
To form the larger matrix, we note

X X
B= B(T ), C = C(T )
T ∈T T ∈T
If (m1 , m2 , m3 ) are edge numbers for T , and σEmk = ±1, then we

compute
B(T ) = diag(σEm1 , σEm2 , σEm3 )BL diag(σEm1 , σEm2 , σEm3 )
and
C(T ) = diag(σEm1 , σEm2 , σEm3 )(|P3 − P2 |, |P1 − P3 |, |P2 − P1 |)T
Hence the global matrix is formed.

Forcing Term
In order to assemble the right hand side, we approximate the integral

by a one point quadrature rule,
Z
f dx = |T |f (z)
T
where z is located in the centroid of each triangle.

Numerical Experiment
Let Ω = [0, 1] × [0, 1]. For this numerical example, we consider the
problem (
−∆u = (2(x2 − x) + 2(y 2 − y)) in Ω
u = 0 on ∂Ω
The exact solution is
u(x, y) = (x2 − x)(y 2 − y)
with
σ(x, y) = ((2x − 1)(y 2 − y), (x2 − x)(2y − 1))

Figure 2: The plot of the approximate displacement, u.
Figure 3: The plot of the x and y direction of σ, denoted px and py .

The error for the experiment coincides with the theoretical estimates as
well.
mk
kσ−σh ku−um
h k0
h ku − um m
h k0 kσ − σh k0 m k
kσ−σh/2
0
0 ku−um k
h/2 0
.5 0.015539 0.061658 1.737737 1.81479
.25 0.008562 0.035482 1.928159 1.96097
125 0.004366 0.018402 1.981351 1.99115
.0625 0.002192 0.009287 1.995273 1.99795
.03125 0.001097 0.004654 1.998797 1.99942
.015625 0.000548 0.002328 1.999656 1.99985
0.0078125 0.000274 0.001164 − −


Presentation Draft FMEFM

Uploaded by

Copyright:

Available Formats

Presentation Draft FMEFM

Uploaded by

Document Information

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

Presentation Draft FMEFM

Uploaded by

Copyright:

Available Formats

A Survey on a Mixed Finite Element Method

David Sacco Mixed Finite Element Method December 8, 2016 1 / 26

David Sacco Mixed Finite Element Method December 8, 2016 2 / 26

David Sacco Mixed Finite Element Method December 8, 2016 3 / 26

David Sacco Mixed Finite Element Method December 8, 2016 4 / 26

In order to approximate u, we will be using a subspace of L2 (Ω). The

The main property of these functions is that ψE · νE = 1 along the

David Sacco Figure 1

where BT is the largest ball contained in T .

David Sacco Mixed Finite Element Method December 8, 2016 7 / 26

||σ − σhm ||0 ≤ Ch||σ||1 ≤ Ch||u||2

Where u is the solution to the model problem, and C > 0 is

David Sacco Mixed Finite Element Method December 8, 2016 8 / 26

For some C > 0 independent of h.

The proofs of these theorems require properties of the subspaces RT0

David Sacco Mixed Finite Element Method December 8, 2016 9 / 26

Manipulating the mixed model problem, we may arrive to the following.

Here, q and v may be regarded as test functions.

David Sacco Mixed Finite Element Method December 8, 2016 10 / 26

Now, we write the problem framed in the approximation spaces, RT0

Here, qh and vh may be regarded as test functions in their respective

David Sacco Mixed Finite Element Method December 8, 2016 11 / 26

with j = 1, 2, ..., M and l = 1, 2, ..., L.

From the linear system, we have M + L equations with M + L

David Sacco Mixed Finite Element Method December 8, 2016 13 / 26

David Sacco Mixed Finite Element Method December 8, 2016 14 / 26

David Sacco Mixed Finite Element Method December 8, 2016 15 / 26

David Sacco Mixed Finite Element Method December 8, 2016 16 / 26

Let N be the number of nodes, then nodes2element is a N × N sparse

For us, N 2E(1, 5) = 1, but note N 2E(5, 1) = 5, by definition N 2E(k, l)

David Sacco Mixed Finite Element Method December 8, 2016 17 / 26

nodes2edge is a N × N sparse symmetric matrix, assigning nodes to

To create this matrix, let B = N 2E + N 2E T , so that every pair of

David Sacco Mixed Finite Element Method December 8, 2016 18 / 26

edge2element is a M × 4 sparse matrix, assigning edges (pairs of

David Sacco Mixed Finite Element Method December 8, 2016 19 / 26

It is customary to write a smaller matrix to ‘generate’ the bigger

David Sacco Mixed Finite Element Method December 8, 2016 20 / 26

David Sacco Mixed Finite Element Method December 8, 2016 21 / 26

To form the larger matrix, we note

If (m1 , m2 , m3 ) are edge numbers for T , and σEmk = ±1, then we

B(T ) = diag(σEm1 , σEm2 , σEm3 )BL diag(σEm1 , σEm2 , σEm3 )

C(T ) = diag(σEm1 , σEm2 , σEm3 )(|P3 − P2 |, |P1 − P3 |, |P2 − P1 |)T

Hence the global matrix is formed.

David Sacco Mixed Finite Element Method December 8, 2016 22 / 26

In order to assemble the right hand side, we approximate the integral

where z is located in the centroid of each triangle.

David Sacco Mixed Finite Element Method December 8, 2016 23 / 26

u(x, y) = (x2 − x)(y 2 − y)

David Sacco Mixed Finite Element Method December 8, 2016 24 / 26

Figure 2: The plot of the approximate displacement, u.

Figure 3: The plot of the x and y direction of σ, denoted px and py .

David Sacco Mixed Finite Element Method December 8, 2016 25 / 26

David Sacco Mixed Finite Element Method December 8, 2016 26 / 26

You might also like