FINITE ELEMENT METHODS:

IMPLEMENTATIONS

I
Vt DOLEJS

Charles University Prague

Faculty of Mathematics and Physics
Department of Numerical Mathematics
Sokolovsk
a 83, 186 75 Prague
Czech Republic

Praha

January 4, 2011

Contents
1 Implementation of FEM and DGM
1.1 P 1 solution of the model problem . . . . . . . . . . . . .
1.1.1 Triangulation . . . . . . . . . . . . . . . . . . . .
1.1.2 Finite element space . . . . . . . . . . . . . . . .
1.1.3 Discretization . . . . . . . . . . . . . . . . . . . .
1.1.4 Implementation . . . . . . . . . . . . . . . . . . .
1.1.5 Treatment of general boundary conditions . . . .
1.2 General FE solution of the model problem . . . . . . . .
1.2.1 Definition of reference elements . . . . . . . . . .
1.2.2 Evaluation of integrals on reference element . . .
1.2.3 Grids and functional spaces . . . . . . . . . . . .
1.2.4 Implementation . . . . . . . . . . . . . . . . . . .
1.3 Basis and shape functions . . . . . . . . . . . . . . . . .
1.3.1 The reference Lagrangian shape functions . . . .
1.3.2 The reference Lobatto shape functions . . . . . .
1.3.3 Global basis functions on Vh . . . . . . . . . . . .
1.4 Discontinuous finite elements . . . . . . . . . . . . . . . .
1.4.1 DG discretization of the model problem . . . . .
1.4.2 hp-discontinuous Galerkin method . . . . . . . . .
1.4.3 Definition of a DG basis . . . . . . . . . . . . . .
1.4.4 Construction of the DG basis functions . . . . . .
1.4.5 Evaluation of volume integrals . . . . . . . . . . .
1.4.6 Evaluation of face integrals . . . . . . . . . . . . .
1.4.7 Data structure . . . . . . . . . . . . . . . . . . .
1.5 Numerical quadratures . . . . . . . . . . . . . . . . . . .
1.5.1 Edge quadratures . . . . . . . . . . . . . . . . . .
1.5.2 Quadrilaterals quadratures . . . . . . . . . . . . .
1.5.3 Triangles quadratures . . . . . . . . . . . . . . . .
1.5.4 Data structure . . . . . . . . . . . . . . . . . . .
1.6 Basic visualization techniques . . . . . . . . . . . . . . .
1.6.1 Types of visualization . . . . . . . . . . . . . . .
1.6.2 Software for visualization . . . . . . . . . . . . . .
1.6.3 Visualization of higher order polynomial function

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

5
5
6
6
8
8
15
17
17
19
21
23
24
24
27
30
33
33
35
35
36
38
38
41
41
41
41
43
43
43
45
46
47

Chapter 1
Implementation of FEM and DGM
In this chapter we deal with the implementation of the finite element and the discontinuous Galerkin methods which were introduced in the previous chapters. However, the
implementation of these methods is rather complicated problem which can differ based
on the practical applications. Our aim is only to introduce some basic ideas. This
chapter is relatively autonomous, all necessary definitions and notations are recalled.
In order to state these notes more readable, we start in Section 1.1 with the implementation of a simple problem, namely P 1 continuous finite element solution of
the Poisson problem with homogeneous Dirichlet boundary condition. We describe
the implementation of this problem in details including the code subroutines which
are written in the Fortran 90 syntax. We suppose that the non-Fortran readers will
be able to understand this syntax too since the subroutines contain only standard if
then conditions and do cycles. The subroutines are commented in the standard way,
i.e., text behind the character ! is ignored.
In Section 1.2 we describe a general implementation of the conforming finite element
method. We employ the concept of the reference element which seems (in the authors
opinion) more suitable for the implementation. Moreover, we include a description of
the necessary data structures. In Section 1.3 we present a construction of the basis
functions based on two possible sets of shape functions, the Lagrangian and the Lobatto
shape functions. Section 1.4 contains an implementation of the discontinuous Galerkin
method with the emphasizes on the differences with the conforming methods. Since
DG methods allow a simple treatment of hp-methods, we consider an approximation
of different polynomial degrees on different elements. Finally, Sections 1.5 and 1.6
describes possible numerical quadratures and visualizations techniques, respectively.

1.1

P 1 solution of the model problem

In order to explain the basic aspects of the implementation of FEM and DGM, we
start with the model problem represented by the Poisson problem. We recall the weak
formulation and the finite element formulation of this elliptic problem.
Let be a bounded polygonal domain in Rd , d = 2, 3, with a boundary . We
5

seek a function u : R such that

u(x) = g(x), x ,
u(x) = 0, x ,

(1.1)
(1.2)

where g L2 ().
Let V := H01 () then the week formulation of (1.1) reads
find u V :

(u, v) = (g, v) v V.

(1.3)

Let Vh be a finite dimensional subspace of V than the finite element approximation

(1.3) can be written as
find uh Vh :

(uh , vh ) = (gh , vh ) vh V.

(1.4)

In the following, we describe in details the implementation of the method (1.4) with
the aid of a continuous piecewise linear approximation constructed over triangular grid.

1.1.1

Triangulation

Let Th (h > 0) be a partition of the closure of the domain into a finite number of
closed d-dimensional simplexes T with mutually disjoint interiors such that
[
=
T.
(1.5)
T Th

In two-dimensional problems (d = 2), the elements T Th are triangles and in threedimensional problems (d = 3) the elements T Th are tetrahedra. Figure 1.1 shows
two examples of triangular grid constructed over a square domain.
Moreover, we use the following notation for vertices and edges (faces for d = 3) of
Th . We denote by {vi }N
i=1 the set of all vertices of T Th lying inside of and by
N +Nb
{vi }i=N
the
set
of
all
vertices of T Th lying on . Finally, {ei }E
i=1 denotes the
+1
E+Eb
set of all edges (faces) of all T Th lying inside of and {ei }i=E+1 the set of all edges
(faces) of all T Th lying on .

1.1.2

Finite element space

Let
Vh := {vh , vh C() H01 (), vh |T P 1 (T ) T Th }

(1.6)

denote the space of continuous piecewise linear functions on Th vanishing on .

Obviously, each vh Vh is uniquely defined by its values in {vi }N
i=1 .
N
We denote by B := {i }i=1 the set of basis functions of Vh defined by
i Vh , i (vj ) = ij ,

i, j = 1, . . . , N,

(1.7)

where ij is the Kronecker symbol. Obviously, support of each i , i = 1, . . . , N is small

and consists of all triangles T Th having vi as vertex, see Figure 1.2.
6

0.5

0.5

-0.5

-0.5

-1

-1
-1

-0.5

0.5

-1

-0.5

0.5

Figure 1.1: Examples of triangular meshes constructed over = (1, 1) (1, 1): a
coarser grid (left) and a finer one (right)

suppi

vi

Figure 1.2: Support of the basis function i corresponding to the node vi

1.1.3

Discretization

The approximate solution uh Vh can be expressed by

uh (x) =

N
X

uj j (x),

uj R, j = 1, . . . , N.

(1.8)

j=1

Moreover, since i , i = 1, . . . , N form the basis of Vh , it is sufficient that the relation

(1.4) is valid for each i , i = 1, . . . , N . Then, using (1.8), the approximate problems
(1.8) is equivalent to
N
find {uj }N
j=1 R

N
X

uj (j , i ) = (g, i ) i = 1, . . . , N.

(1.9)

j=1

Let as define the stiffness matrix

S = {Sij }N
i,j=1 ,

Sij = (j , i ),

(1.10)

the solution vector u = {ui }N

i and the right-hand side
g = {gi }N
i=1 ,

gi = (g, i ), i = 1, . . . , N.

(1.11)

Then, the problem (1.9) is equivalent to the linear algebraic problem

Su = g.

(1.12)

Therefore, in order to implement the conforming piecewise linear FEM to (1.3), we

have to
evaluate S and g according to (1.10) (1.11),
solve the algebraic problem (1.12),
reconstruct the approximate solution uh Vh from u using (1.8).

1.1.4

Implementation

Data structures
In order to present algorithms, we introduce the global data structures defining the
triangulation Th . For simplicity, we restrict to d = 2.
Let N denote the number of inner vertices of Th and Nb denote the number of
vertices of Th lying on . Let the real arrays xp(1:N+Nb) and yp(1:N+Nb) be the x1 and x2 -coordinates is the nodes of mesh Th indexed by i = 1, . . . , N + N b, respectively.
We assume that first N components of xp(:) and yp(:) correspond to inner vertices
and the last Nb components of xp(:) and yp(:) to boundary vertices.
Furthermore, let the vertexes of triangle Ti Th , i = 1, . . . , M be indexed by the
integer (two-dimensional) array node(i,1:3). Obviously, if some vertex of Ti lies on
then we have for the corresponding value node(i,j)>N.
For simplicity, we assume that xp(:), yp(:), node(:,:) are global arrays so that
we have access to them from any subroutine.
8

suppi
vi

Dij

vj
suppi

Figure 1.3: Volume Dij = supp(i ) supp(j ) for nodes vi and vj connected by an
edge
Setting of the stiff matrix non-optimized algorithm
Here we present a non-optimized algorithm for the evaluation of the entries of the stiff
matrix S. According to (1.10), the entries of the stiff matrix satisfy
Z
Sij =
j i dx.
(1.13)

Since the support of each i , i =

vi , we can distinguish three cases:
R
|i |2 dx

supp(i )

R
j i dx
Sij =
Dij

1, . . . , N is located around the corresponding node

if i = j,
if vi and vj are connected by an edge,
otherwise.

Here, Dij = supp(i ) supp(j ), see Figure 1.3.

For the diagonal entries of the stiff matrix Sii , i = 1, . . . , N , we have
Z
X Z
X
2
Sii =
|i | dx =
|i |2 dx =
|T ||i |T |2 .
supp(i )

(1.14)

T supp(i )

(1.15)

T supp(i )

The last equality follows from the fact that i is linear on each T supp(i ) and thus
i is constant on T . Therefore, it is necessary to evaluate i on triangles having
vi as an vertex.
9

(x
1, y1) = 1

[x3, y3 ]

[x1, y1 ]

[x2, y2]

Figure 1.4: Triangle T and the function

Let us consider a triangle T with vertices [x1 , y1 ], [x2 , y2 ] and [x3 , y3 ]. Let (x,
y)
be the linear function on T such that
(x
1 , y1 ) = 1,
see Figure 1.4.
It is possible

x1
D = det x2
x3
Then

(x
2 , y2 ) = 0,

(x
3 , y3 ) = 0,

to derive the explicit relations for the gradient of .

x1
1 y1 1
y1 1
y2 1 , Dx = det 0 y2 1 , Dy = det x2
x3
0 y3 1
y3 1


(1.16)
Let us put

1 1
0 1 . (1.17)
0 1


Dx Dy
(y2 y3 , x3 x2 )
=
=
,
,
(1.18)
D D
2|T|
where |T| is the area of the triangle T.
Hence, based on (1.17) (1.18), we can define the following subroutine GRADn, which
evaluate the gradient of the linear function on the triangle with vertices [x1 , y1 ], [x2 , y2 ]
and [x3 , y3 ] such that this function takes value 1 in [x1 , y1 ] and vanish at [x2 , y2 ] and
[x3 , y3 ]. We use the convention that x1 = x1, y1 = y1 etc.
subroutine GRADn (x1,y1, x2, y2, x3, y3, Dx, Dy)
real, intent(in) :: x1,y1, x2, y2, x3, y3 ! input parameters
real, intent(out) :: Dx, Dy
! output parameters
real :: D
! local variable
D = x1*(y2-y3) + x2*(y3-y1) + x3*(y1-y2)
Dx = (y2 - y3) / D
Dy = (x3 - x2) / D
end subroutine GRADn
10

[x5 , y5 ]

[x4 , y4 ]
[xj , yj ]

[x6 , y6]

[x0 , y0 ]

[x3 , y3]

[xk , yk ]
[xl , yl ]

[x2 , y2 ]
[xi , yi ]
[x1 , y1]

Figure 1.5: Notation for the evaluation of the diagonal entries (left) and off-diagonal
entries (right) of the stiff matrix
Now, let us go back to the evaluation of Sii . Let Tij , j = 1, . . . , n be triangles
having the node vi as an vertex. Obviously, supp(i ) = nj=1 Tij . We assume that vi
has the coordinates [x0 , y0 ]. Moreover, the vertices of Tij , j = 1, . . . , n have coordinates
[xj , yj ], j = 1, . . . , n indexed counter-clock-wise, see Figure 1.5, left.
Finally, taking into account (1.15) we can evaluate Sii by the following subroutine:
subroutine STIFF_DIAGn(n, x(0:n), y(0:n), sii )
integer, intent(in) :: n
! input parameters
real, intent(in) :: x(0:n), y(0:n) ! input array
real, intent(out) :: sii
! output parameters
real :: D, Dx, Dy
! local variables
integer :: j, j1
! local variables
sii = 0.
do j=1,n
j1 = mod(j+1, n) + 1
call GRADn(x(0), y(0), x(j), y(j), x(j1), y(j1), Dx, Dy)
D = x(0)*(y(j)-y(j1) ) + x(j)*(y(j1)-y(0) ) + x(j1)*(y(0)-y(j) )
sii = sii + (Dx*Dx + Dy*Dy)*D/2
end do
end subroutine STIFF_DIAGn

! D/2 = area of triangle

Comparing the subroutines GRADn and STIFF_DIADn, we observe that the values
Dx and Dy are divided by D in GRADn, then we square this value and multiply by D in
STIFF_DIADn. It is more efficient to avoid the division and the following multiplication
by the same value. Hence, we can modify the subroutines GRADn and STIFF_DIADn by
their more efficient variants GRAD and STIFF_DIAD defined bellow.
subroutine GRAD (x1,y1, x2, y2, x3, y3, Dx, Dy)
real, intent(in) :: x1,y1, x2, y2, x3, y3
11

real, intent(out) :: Dx, Dy

Dx = (y2 - y3)
Dy = (x3 - x2)
end subroutine GRAD
and
subroutine STIFF_DIAG( n, x, y, sii )
integer, intent(in) :: n
real, intent(in) :: x(0:n), y(0:n)
real, intent(out) :: sii
real :: D, Dx, Dy
! local variables
integer :: j, j1
sii = 0.
do j=1,n
j1 = mod(j+1, n) + 1
call GRAD(x(0), y(0), x(j), y(j), x(j1), y(j1), Dx, Dy)
D = x(0)*(y(j)-y(j1) ) + x(j)*(y(j1)-y(0) ) + x(j1)*(y(0)-y(j) )
sii = sii + (Dx*Dx + Dy*Dy)/D/2 ! D/2 = area of triangle
end do
end subroutine STIFF_DIAG
Moreover, we have to evaluate the off diagonal terms of the stiff matrix, i.e., Sij
where vi and vj are connected by an edge of Th . We denote the coordinates of vertices
of triangles sharing the edge vi vj according to Figure 1.5, right. The we define the
subroutine fro the evaluation of the off-diagonal term of the stiff matrix:
subroutine STIFF_OFFDIAG(xi, yi, xj, yj, xk, yk, xl, yl, sij )
real, intent(in) :: xi, yi, xj, yj, xk, yk, xl, yl
real, intent(out) :: sij
real :: D, Dx, Dy
! local variables
sij = 0.
call GRAD(xi, yi, xk, yk, xj, yj, Dx, Dy)
D = xi*(yk-yj) + xk*(yj-yi) + xj*(yi-yk)
sij = sij + (Dx*Dx + Dy*Dy)/D/2 ! D/2 = area of triangle
call GRAD(xi, yi, xj, yj, xl, yl, Dx, Dy)
D = xi*(yj-yl) + xj*(yl-yi) + xl*(yi-yj)
sij = sij + (Dx*Dx + Dy*Dy)/D/2 ! D/2 = area of triangle
end subroutine STIFF_OFFDIAG
Finally, in order to evaluate all entries of the stiff matrix S we should call either
subroutine STIFF_DIAG or STIFF_OFFDIAG for all Sij 6= 0. However, it requires a
setting of the appropriate data structures, namely some link to the nodes connected
12

by an edge to a given node (see 1.5, left) and some link between elements sharing an
face (see 1.5, right.) Since we are going to introduce better algorithm in the following
section, we skip this one.
Setting of the stiff matrix optimized algorithm
We can notice that when we evaluate all non-vanishing terms of matrix S, we call
subroutine GRAD for each T Th several times. It is more efficient to go over all
T Th only once time. This means that we compute the gradients of all possible
linear functions on T vanishing in two vertices of T and equal to 1 in the rest node.
Then we evaluate the corresponding entries of the stiff matrix.
Here we introduce more efficient algorithm for the evaluation of S. It evaluates
all entries of S which correspond to inner vertices of Th given by (1.10). First, we
define a subroutine, which evaluate the gradient of three possible restrictions of the
test functions on the given triangle.
subroutine GRADIENTS( x, y, Dphi )
real, intent(in) :: x(1:3), y(1:3)
real, intent(out) :: Dphi(1:3, 1:2)
! Dphi(i,j) = derivative of i-th function with respect x_j
Dphi(1,1) = y(2) - y(3)
Dphi(1,2) = x(3) - x(2)
Dphi(2,1) = y(3) - y(1)
Dphi(2,2) = x(1) - x(3)
Dphi(3,1) = y(1) - y(2)
Dphi(3,2) = x(2) - x(1)
end subroutine GRADIENTS
Then we have the following subroutine for the evaluation of all entries of S. Let us
recall that t xp(:), yp(:), node(:,:) are global arrays storing the coordinates of
vertices and the indexes of vertices of triangles, see begin of Section 1.1.4. Hence, we
have access to them from the subroutines without their specification as input parameters.
subroutine STIFF(M, N, S )
integer, intent(in) :: M
! = number of triangles of triangulation
integer, intent(in) :: N
! = number of vertices of triangulation
real, dimension(1:N, 1:N), intent(out) :: S ! = stiff matrix
real :: Dphi(1:3, 1:2) , D ! local variables
integer :: i,j,ki, kj
S(1:N, 1:N) = 0

13

do k=1,M
x(1:3) = xp(node(k, 1:3) )
y(1:3) = yp(node(k, 1:3) )
D = x(1)*(y(2)-y(3)) + x(2)*(y(3)-y(1)) + x(3)*(y(1)-y(2))
call GRADIENT( x(1:3), y(1:3), Dphi(1:3, 1:2) )
do ki=1,3
i = node(k,ki)
if( i <= N) then
!! inner vertex?
do kj=1,3
j = node(k,kj)
if( i <= N) then
!! inner vertex?
S(i,j) = S(i,j) + (Dphi(ki,1)* Dphi(kj,1) &
+ Dphi(ki,2)* Dphi(kj,2) )/D/2
endif
enddo
endif
enddo
enddo
end subroutine STIFF
Remark 1 The subroutine STIFF evaluate only entries of the stiff matrix corresponding to the inner vertices of Th . It is assured by two if then conditions in subroutine
STIFF. If more general boundary condition than (1.2) is given also the entries corresponding to the boundary vertices of Th has to be evaluated. In this case, N denotes
the number of all vertices including the boundary ones, see Section 1.1.5. Moreover,
we can simply remove both if then conditions in subroutine STIFF.
Setting of the right-hand side
Here we present an algorithm which evaluate the entries of the vector g given by (1.11).
Let i = 1, . . . , N , we have
X Z
gi = (g, i ) =
g i dx,
(1.19)
T supp(i )

where supp(i ) is shown in Figure 1.2. Generally, the integrals in (1.19) are evaluated
with the aid of a numerical quadratures, which are discussed in Section 1.5. Since we
considered in this section the piecewise linear approximation, we will consider only the
following simple quadrature rule
Z

X
1
(xl )g(xl ),
g(x) (x) dx |T |
3
l=1

(1.20)

where xl , l = 1, 2, 3 are the vertices of an triangle T Th . This quadrature rule

integrate exactly the linear function, hence it has the first order of accuracy.
14

The application of (1.19) for the evaluation of (1.20) leads to a very simple relation
since i (x0 , y0 ) = 1 and i (xl , yl ) = 1, l = 1, . . . , n, see Figure 1.5, left. Then we have
1
gi meas(supp(i ))g(vi ),
3

i = 1, . . . , N

(1.21)

where vi , i = 1, . . . , N are inner vertices of Th .

Finally, let us note that the numerical quadrature (1.20) is not sufficiently accurate
generally.
Solution of the linear algebraic system (1.12)
Once we have evaluated the entries of matrix S and vector g, we have to solve the
linear algebraic system (1.12). In the case of the Laplace problem (1.1) the situation
is very simple. First we introduce the following term
Definition 1 Let Th be a mesh of . Let T and T be a pair of triangles from Th sharing an edge vi vj . Let vi , vj , vk and vj , vi , vl be the vertices of T and T , respectively.
We denote by SumAngles(T, T ) the sum of angles of T and T adjacent to vertices vk
and vl , respectively. We say that Th is of the Delaunay type if SumAngles(T, T )
for each pair of triangles T, T from Th sharing an edge.
It is possible to show that if Th is of Delaunay type then the corresponding stiff
matrix S is symmetric and diagonally dominant. In this case in fact arbitrary direct
or iterative solver can be used with success for the solution of (1.12).
Reconstruction of the approximate solution
Once we have found the vector u which is the solution of (1.12), we will reconstruct
the piecewise linear function uh Vh from u using (1.8). However, in practice we need
not know an explicit relation for uh . Usually, we require a visualization of the solution
and/or an evaluation of some characteristics. The basic visualization techniques are
mentioned in Section 1.6.

1.1.5

Treatment of general boundary conditions

In the previous sections, the homogeneous Dirichlet boundary condition was considered.
However, in practice, more general boundary conditions have to be taken into account.
Let be a bounded polygonal domain in Rd , d = 2, 3, with a boundary which
consists of two disjoint parts D and N . We seek find a function u : R such
that
u(x) = g(x), x ,
u(x) = uD (x), x D ,
u(x) n = gN (x), x D ,
where g L2 (), gN L2 () and uD is a trace of some u H 1 ().
15

(1.22)
(1.23)
(1.24)

Let
V := {v H 1 (), v|D = 0 in the sense of traces }

(1.25)

then the week formulation of (1.22) (1.24) reads

find u H 1 () :

u u V,
(u, v) = (g, v) + (gN , v)L2 (N )

(1.26)
v V.

Let Th be a mesh reflecting the consists of two disjoint parts D and N . Let
Xh be a finite dimensional subspace of H 1 () constructed over mesh Th and Vh Xh
its subspace such that vh |D = 0 vh Vh . The finite element approximation (1.26)
can be written as
find uh Xh :

(uh uD , vh )L2 (D ) = 0 vh Xh ,
(uh , vh ) = (g, vh ) + (gN , vh )L2 (N )

(1.27)
vh Vh .

The numerical implementation of (1.27) is more complicated than the implementation of (1.4). In this place we skip the evaluation of the boundary integral on the right
hand side of (1.27) and refer to Section 1.4.6. We focus on the assuring of the Dirichlet
boundary condition on D .
We use the following notation for vertices of Th which differs from the notation
introduced in the previous sections. We denote by {vi }N
i=1 a set of all vertices of
N +Nb
T Th lying inside of or on N and by {vi }i=N +1 a set of all vertices of T Th
lying on D .
Similarly as in the previous, we restrict to the piecewise linear approximations.
N +Nb
Therefore, each function from Xh is uniquely given by its values in {vi }i=1
and each
N
function from Vh is uniquely given by its values in {vi }i=1 . We denote by i Xh the
basis functions corresponding to vi , i = 1, . . . , N + Nb given by
i Xh ,

i (vj ) = ij ,

i, j = 1, . . . , N + Nb .

(1.28)

First, we introduce a piecewise linear approximation of uD from the boundary

condition. We define a function uD,h Xh \ Vh by
(uD,h uD , vh )L2 (D ) = 0 vh Xh \ Vh .

(1.29)

Such uD,h can be simply constructed since Xh is a finite dimensional spaces and then
(1.29) represents a system of linear algebraic equations. Namely, we have
uD,h (x) =

Nb
X

uD,j i (x),

(1.30)

j=N +1

therefore, we have from (1.29) (1.30) the following system of the linear algebraic
equations
Nb
X

uD,j (j , i )L2 (D ) = (uD , i )L2 (D ) ,

j=N +1

16

i = N + 1, . . . , Nb .

(1.31)

The evaluation of terms (j , i )L2 (D ) , i, j = N + 1, . . . , Nb is simple since these terms

are non-vanishing only if vi and vi are connected by an edge on D . Moreover, the
evaluation of (uD , i )L2 (D ) , i = N + 1, . . . , Nb reduces to an integration over two
edges on D having vi as an end-point.
The solution of the (1.27) can be written in the form
uh (x) =

N
X

uj i (x) +

j=1

Nb
X

uD,j i (x) =

N
X

uj i (x) + uD,h ,

(1.32)

j=1

j=N +1

where the coefficients uD,j , j = N + 1, . . . , Nb are known and the coefficients uj , j =

1, . . . , N are unknown. Inserting (1.32) into (1.27) and putting vh := i , we have
N
X

uj (j , i ) = (g, vh ) + (gN , vh )L2 (N ) (uD,h , i ),

i = 1, . . . , N,

j=1

which is very close to the problem (1.10), where the right-hand-side contains additionally terms arising from the Dirichlet and the Neumann boundary conditions. Moreover,
symbol N has little different meanings in (1.10) and in (1.33), in the former case denotes
the number of interior vertices, in the latter case the number of inner and Dirichlet
nodes. However, all subroutines presented in Section 1.1.4 can be employed with this
minor changes.

1.2

General FE solution of the model problem

In Section 1.1 we described the implementation of the conforming piecewise linear finite
element approximation of problem (1.1) (1.2). Here we consider a higher degree of
polynomial approximation. The implementation is more complicated since the gradient
of test function is not constant over T Th .
It is more comfortable employ the concept of the reference element. This approach
allows us to deal with more general elements than triangles, e.g., quadrilaterals, curvilinear elements, etc.

1.2.1

Definition of reference elements

In this lecture notes, we consider

the reference simplex (triangle for d = 2 and tetrahedron for d = 3)
(
)
d
X
= (x1 , . . . , xd )), xi 0, i = 1, . . . , d,
Tbt := x
xi 1

(2.33)

i=1

the reference hexagon (square for d = 2 and cube for d = 3)

Tbq := {
x = (x1 , . . . , xd )), 0 xi 1, i = 1, . . . , d, } .
17

(2.34)

 4 = [0; 1]
v

e3

3 = [0; 1]
v

3 = [1; 1]
v

Tq
e4

e3

e2

1 = [0; 0]
v

e1

1 = [0; 0]
v

2 = [1; 0]
v

e2
Tt

e1

2 = [1; 0]
v

Figure 1.6: Two dimensional reference elements: quadrilateral Tbq (left) and triangle Tbt
i the vertices and the edges ei
(right) with the vertices v
For simplicity, Tb denotes either the reference simplex or the reference hexagon based
on the context. We call Tb the reference element.
1, . . . , v
d+1 the vertices and by e1 , . . . , ed+1 the edges of
Moreover, we denote by v
b
1, . . . , v
sd the vertices and by e1 , . . . , e2d the edges of Tbq . For d = 2, we
Tt , similarly by v
assume that the numbering is oriented counterclockwise, see Figure 1.6, which shows
the two-dimensional reference triangle and the reference quadrilateral.
We assume that for each element T Th there exist a mapping
FT = FT (
x) = (FT,1 (
x), . . . , FT,d (
x)) : Tb Rd

such that FT (Tb) = T.

(2.35)

If FT is an affine (linear) mapping then T is either a triangle or a parallelogram based

on the type of Tb. Generally, the construction of the mappings FT is simple for several
types of elements:

Triangles: Let T be a triangle defined by its vertices x1 , x2 and x3 . Then the

linear mapping
FT (
x) = FT (
x1 , x2 ) = x1 + x1 (x2 x1 ) + x2 (x3 x1 )

(2.36)

maps Tbt on T and moreover, FT (0, 0) = x1 , FT (1, 0) = x2 and FT (0, 1) = x3 .

Tetrahedron: Let T be a tetrahedron defined by its vertices x1 , x2 , x3 and x4

Then the linear mapping
FT (
x) = FT (
x1 , x2 , x3 ) = x1 + x1 (x2 x1 ) + x2 (x3 x1 ) + x3 (x4 x1 ) (2.37)
maps Tbq on T and moreover, FT (0, 0, 0) = x1 , FT (1, 0, 0) = x2 , FT (0, 1, 0) = x3
and FT (0, 0, 1) = x4 .

Quadrilaterals: Let T be a quadrilateral defined by its vertices x1 , x2 , x3 and x4 .

Then the mapping
FT (
x) = FT (
x1 , x2 ) = x1 +
x1 (x2 x1 )+
x2 (x3 x1 )+
x1 x2 (x4 x2 x3 +x1 ) (2.38)
18

x1

x3
x2

x2,3

Figure 1.7: One-edge P 2 -curvilinear triangles

maps Tbq on T and moreover, FT (0, 0) = x1 , FT (1, 0) = x2 , FT (0, 1) = x3 and
FT (1, 1) = x4 . Let us note that if T is a parallelogram then x4 x2 x3 + x1 = 0
and (2.38) is identical with (2.36).
One-edge P 2 -curvilinear triangles: This types of elements are often used for an
approximation of non-polygonal boundaries. Let T be a one-edge P 2 -curvilinear
triangle, which is defined by its vertices x1 , x2 , x3 and the node x2,3 lying not
far from (x2 + x3 )/2. Element T is bounded by two straight lines x1 x2 and x1 x2
and one quadratic curve passing x2 , x2,3 , x3 , see Figure 1.7. Then the quadratic
mapping
FT (
x1 , x2 ) = x1 + x1 (x2 x1 ) + x2 (x3 x1 ) + 4
x1 x2 (x2,3 (x2 + x3 )/2) (2.39)
maps Tbq on T . Moreover, FT (0, 0) = x1 , FT (1, 0) = x2 , FT (0, 1) = x3 and
FT (1/2, 1/2) = x2,3 . Let us note that if x2,3 = (x2 + x3 )/2 then T is a triangle
and (2.39) is identical with (2.36).
It is possible to define mappings FK for more general elements. Let us note that in
the following sections we require that FK is invertible and continuously differentiable
on Tb. The existence of FK1 follows from the assumption that the Jacobi matrix
JFT (
x) :=

D
FT (
x)
D
x

(2.40)

does not change a sign on Tb. Very often, only polynomial mappings are considered
from practical reasons.

1.2.2

Evaluation of integrals on reference element

The concept of the reference elements is based in the Fubini theorem, which we present
in the following form.
19

Theorem 1 (Fubini theorem) Let T be the image of FT (Tb) where Tb is the reference
element and FT is the continuously differentiable mapping. Let f (x) : T R be an
integrable function defined on T . We define the function f : Tb R by f(
x) = f (x)
where x = FT (
x). Then
Z
Z
f (x) dx =
f(
x)| det JFT (
x)| d
x,
(2.41)
Tb

where JFT is the Jacobi matrix of FT .

Let us note that if FT is a linear mapping (i.e., T is an simplex or a parallelogram)
then JF is a constant matrix and | det JFT | = 2|T |, where |T | denotes the d-dimensional
Lebesgue measure of T .
Theorem 1 gives us a tool for the evaluation of the various type of integrals in the
implementation of finite element and discontinuous Galerkin methods. Instead of integration over T we integrate over Tb which is generally more easy. This is demonstrated
in the following paragraphs.
Let T Th and be a function defined on T . Our aim is to compute
Z
(x)
dx.
(2.42)
T xi
Let : Tb R be defined as usual by (
x) = (x) where x = FT (
x),
FT = (FT,1 , . . . , FT,d )

= FT1 (x) where

and T = FT (Tb). Moreover, we have x
1
1
)
, . . . , FT,d
FT1 = (FT,1

is the inverse mapping of FT . With the aid of Theorem 1 and the chain rule, we have
Z
Z

(x) dx =
(
x)| det JFT (
x)| d
x
(2.43)
T xi
T xi
Z X
Z X
d
d
1
(
x) xk
(
x) FT,k
=
| det JFT (
| det JFT (
x)| d
x=
x)| d
x.
k xi
k xi
T k=1 x
T k=1 x
Therefore, we have to evaluate the partial derivative of FT1 . It is possible to use the
following determination. Let i, k = 1, . . . , d, then
xk = FT,k (FT1 (x)),

xi

xk =
FT,k (FT1 (x)),
xi
xi
d
1
X
FT,k FT,l
,
ik =
xl xi
l=1
ik =

d
X

(JFT )kl (JF 1 )li ,

T

l=1

20

(2.44)

where (JFT )kl , k, l = 1, . . . , d and (JF 1 )li , l, i = 1, . . . , d denote entries of the Jacobi
T
matrices JFT and JF 1 , respectively. The relation (2.44) implies
T

I = JFT JF 1

JF 1 = (JFT )1 .
T

(2.45)

Therefore, instead of evaluation of the Jacobi matrix of FT1 it is sufficient to evaluate

(JFT )1 = the inversion of JFT . This is simple since JFT is (usually) 2 2 or 3 3
matrix.
Therefore, from (2.43) and (2.45) we have
Z

Z X
d
(
x)
(JFT (
x)1 )ki | det JFT (
x)| d
x
k
T k=1 x
Z X
d
(
x)
(JFT (
x)T )ik | det JFT (
=
x)| d
x

k
T k=1
Z 

b (
JFT (
=
x)T
x) | det JFT (
x)| d
x,

(x) dx =
xi

(2.46)

b )
where JFT
is the transposed matrix to JF1
, (JFT

i denotes the i-th component

T
T
T
T b
b and
b is the
of vector JFT arising from the matrix-vector product of JFT
by
T
reference gradient operator given by



b
.
(2.47)
,...,
:=
x1
xd
In the same manner it is possible to derive, e.g.,
Z 
Z
 

a (x) b (x)
b a (
b b (
dx =
JFT (
x)T
x)
JFT (
x)T
x) | det JFT (
x)| d
x
xi
xj
i
j
T
T
(2.48)
or more simply
Z
Z 
 

a b
T b
b a
dx =
JFT

| det JFT | d
x
(2.49)
b
FT
T
i
j
T xi xj
T
which is type of integral appearing in the evaluation of the stiff matrix.

1.2.3

Grids and functional spaces

We again consider the model Poisson problem (1.1) (1.2), the weak formulation (1.3)
and the finite element discretization (1.4). In contrast to Section 1.1, we consider
more general partitions of and finite element spaces. Let a grid Th be a set of
non-overlapping elements T Th such that
[
=
T.
(2.50)
T Th

We consider two following type of grids.

21

Triangular grid
Each element T Th is an images of the reference triangle (for d = 2) or the reference
tetrahedron (for d = 3) given by a polynomial mapping FT : Tbt R, T Th . If FT is
a linear mapping then T is a triangle (for d = 2) or a tetrahedron (for d = 3). If FT
has the form (2.39) then T is an one-edge P 2 -curvilinear triangle. The approximate
solution is sought in the functional space
Vht,p := {vh , vh C() H01 (), vh |T FT P t,p (Tbt ) T Th },
(2.51)
i1 +...+id p
X
t,p
ai1 ,...,id xi11 . . . xidd , ai1 ,...,id R},
P () := {
z : R, z(
x1 , . . . , xd ) =
i1 ,...,id =0

where Rd is an arbitrary domain Let us note that the notation vh |T FT P t,p (Tbt )
means that there exists a function vh P t,p (Tbt ) such that
vh (x) = vh (FT (
x)) = vh (
x)

x Tbt .

(2.52)

Obviously, if FT is a linear mapping then vh |T P t,p (T ). Otherwise, (e.g., T is a

curvilinear triangle), vh |T 6 P t,p (T ).
The spaces P t,p (Tbt ), p = 1, 2, 3 can be expressed (for d = 2) as
P t,1 (Tbt ) = span{1, x1 , x2 },
P t,2 (Tbt ) = span{1, x1 , x2 , x21 , x1 x2 , x22 },
P t,3 (Tbt ) = span{1, x1 , x2 , x21 , x1 x2 , x22 , x31 , x21 x2 , x1 x22 , x32 },

(2.53)

t,p := dim P t,p (Tbt ) =

where span{S} is the linear hull of the set {S}. Obviously, N
t,p := dim P t,p (Tbt ) = (p + 1)(p + 2)(p + 3)/6 for d = 3.
(p + 1)(p + 2)/2 for d = 2 and N
Quadrilateral grids

Each element T Th is an images of the reference square (for d = 2) or the reference

cube (for d = 3) given by a polynomial mapping FT : Tbt R, T Th . If FT is a
linear mapping then T is a parallelogram. The approximate solution is sought in the
functional space
(2.54)
Vhq,p := {vh , vh C() H01 (), vh |T FT P q,p (Tbq ) T Th },
p
X
ai1 ,...,id xi11 . . . xidd , ai1 ,...,id R},
P q,p () := { z : R, z(x1 , . . . , xd ) =
i1 ,...,id =0

where Rd is an arbitrary domain.

The spaces P q,p (Tbq ), p = 1, 2, 3 can be expressed (for d = 2) as

P q,1 (Tbq ) = span{1, x1 , x2 , x1 x2 },

P q,2 (Tbq ) = span{1, x1 , x2 , x1 x2 , x21 , x21 x2 , x21 x22 , x1 x22 , x22 },
P q,3 (Tbq ) = span{1, x1 , x2 , x1 x2 , x21 , x21 x2 , x21 x22 , x1 x22 , x22 ,
x31 , x31 x2 , x31 x22 , x31 x32 , x21 x32 , x1 x32 , x32 }.
22

(2.55)

 q,p := dim P q,p (Tbq ) = (p + 1)d for d = 2, 3.

Obviously, N
Since the most of the following considerations are valid for triangular as well as
quadrilateral grids we call Th a triangulation in both cases. Moreover, the corresponding reference element is denoted by Tb, P p (Tb) the corresponding space of polynomial
and the corresponding finite element space is
functions on Tb with the dimension N
denoted by Vh .

1.2.4

Implementation

Basis of Vh
Let N denote the dimension of the space Vh and Bh := {i (x)}N
i=1 be its basis. The
construction of Bh will be discussed in Section 1.3. Here we only note that we construct

:= {j , j : Tb R}N
the set reference shape functions B
j=1 such that
, N
is the dimension of P p (Tb).
j (
x) P p (Tb), j = 1, . . . , N
are linearly independent,
j (
x), j = 1, . . . , N
= P p (Tb).
span(B)

the reference basis.

We call B
Moreover, in virtue of (2.51) or (2.54), the basis functions i Bh satisfy: for each
such that
T Th , T supp(i ) there exists a function i,T B
i (x)|T = i (FT (
x))|T = i,T (
x)
More details are given in Section 1.3.

x Tb, i = 1, . . . , N.

(2.56)

A use of the concept of the reference elements

Our aim is to evaluate the integrals of type
Z
i j
dx,
i, j = 1, . . . , N, k, l = 1, . . . , d, T Th ,
T xk xl

(2.57)

which appears, e.g., in the definition of the stiff matrix.

A use of the approach from Section 1.1 requires an evaluation (and storing) of i |T
for all i = 1, . . . , N and all T Th . Let us note, that i |T = 0 for a lot of T Th
since supports of test functions contain only a few elements from Th .
It is more efficient to employ the concept of the reference elements, namely relation
(2.49), where the integration is carried out on the reference element Tb.
Therefore, using (2.49) and (2.57), we can evaluate (2.57) with the aid of
Z 
Z
 

i j
T b
T b
dx =
JFT i,T
JFT j,T | det JFT | d
x.
(2.58)
k
l
Tb
T xk xl
23

Data structures
In order to evaluate the right-hand side of (2.58) for all i, j = 1, . . . , N, k, l =
1, . . . , d, T Th it is enough to evaluate and store the following data.
(S1) for each T Th , the determinant of the Jacobi matrix det JFT and the transposed
matrix of the inversion of the Jacobi matrix JFT
,
T
, the gradient
b i on Tb,
(S2) for each i , i = 1, . . . N

(S3) for each pair (i , T ), T Th , T supp(i ), i = 1, . . . , N , an index j

} such that i (FT (
{1, . . . , N
x))|T = j (
x).
Let us note that (S1) is given by the geometry of the mesh, (S2) by the reference shape
functions and (S3) by a combination of both. However, in (S3) we have to store only
one integer index for any acceptable pair (i , T ).

1.3

Basis and shape functions

In this section we describe two possible way (among others) of the construction of basis
of spaces Vht,p and Vhq,p given by (2.54) and (2.52), respectively. In Section 1.3.1 we
describe the construction of Lagrangian basis, which is easy for the determination and
also for the implementation. However, in case when we use, e.g., different degrees of
polynomial approximations at different elements (hp-methods), the efficiency of this
approach is low. Therefore, in Section 1.3.2 we present the construction of the Lobatto
basis, which is more efficient generally.
As we have already mentioned in Section 1.2.4, first we define the reference shape

:= {j }N
b
functions B
j=1 defined on the reference element T . Then the basis of Vh is
We present the construction of the shape functions for d = 2. For
generated from B.
d = 3, it is possible use the similar (but technically more complicated) approach.

1.3.1

The reference Lagrangian shape functions

The reference Lagrangian shape functions are defined with the aid of a set of the
within the reference element.
reference Lagrangian nodes zi = (
zi,1 , zi,2 ), i = 1, . . . , N
. In the following we introduce
Then we simply put i (
zk ) = ij , i, j = 1, . . . , N
the reference Lagrangian nodes and the reference Lagrangian shape function for the
reference triangle and the reference square separately.
The reference Lagrangian shape functions on the square
Definition 2 (Lagrangian nodes) Let p 1, then the Lagrangian nodes on the reference square Tbq corresponding to the polynomial degree p is the set of nodes
q,p

q,p q,p
N
q,p
{zq,p
zk,1
, zk,2 )
k }k=1 , z
k = (

24

(3.59)

zq,1
4

zq,1
3

zq,2
7

zq,2
4

zq,1
2

zq,1
1

zq,2
1

zq,2
8

zq,2
9

zq,2
5

zq,2
6

zq,2
2

zq,2
3

zq,3
14

zq,3
13

zq,3
15

zq,3
16

zq,3
9

zq,3
10

zq,3
11

zq,3
12

zq,3
5

zq,3
6

zq,3
7

zq,3
8

zq,3
1

zq,3
2

zq,3
3

zq,3
4

Figure 1.8: The reference Lagrangian nodes on the reference square Tbq , p = 1 (left),
p = 2 (center) and p = 3 (right)
where

q,p
zj(p+1)+i+1,1
= i/p,

q,p
zj(p+1)+i+1,2
= j/p,

i, j = 0, . . . , p.

(3.60)

q,p = (p + 1)2 .
Obviously, N
Figure 1.8 shows the Lagrangian nodes for p = 1, 2, 3. We define the reference
q,p
N
b
Lagrangian shape functions {q,p
l }l=1 for p 1 on the reference square Tq by the
relation
q,p .
q,p
P p (Tbq ), q,p
zq,p
(3.61)
l
l (
k ) = lk , l, k = 1, . . . , N

It is possible to derive the following explicit relations for the Lagrangian shape functions
on the reference square Tbq

q,1
q,1
1 = (1 1 )(1 2 ),
2 = 1 (1 2 ),
q,1
q,1
3 = 2 (1 1 ), 4 = 1 2
1
1
p=2
q,2
1 = 4(1 1 )( 1 )(1 2 )( 2 ),
2
2
1
q,2
2 = 81 (1 1 )(1 2 )( 2 ),
2
1
1
q,2
3 = 41 ( 1 )(1 2 )( 2 ),
2
2
1
q,2
4 = 8(1 1 )( 1 )2 (1 2 ),
2
q,2
5 = 161 (1 1 )2 (1 2 ),
1
q,2
6 = 81 ( 1 )2 (1 2 ),
2
1
1
q,2
7 = 4(1 1 )( 1 )2 ( 2 ),
2
2
1
q,2
8 = 81 (1 1 )2 ( 2 ),
2
1
1
q,2
9 = 41 ( 1 )2 ( 2 ),
2
2
where 1 = x1 and 2 = x2 are the barycentric coordinates on Tbq .
p=1

25

(3.62)

zt,3
10

zt,2
6

zt,1
3

zt,3
8

zt,2
5

zt,2
4

zt,3
5

zt,1
2

zt,1
1

zt,2
2

zt,2
1

zt,2
3

zt,3
1

zt,3
9
zt,3
7

zt,3
6

zt,3
2

zt,3
3

zt,3
4

Figure 1.9: Lagrangian nodes on the reference triangle Tbt , p = 1 (left), p = 2 (center)
and p = 3 (right)
The reference Lagrangian shape functions on the triangle
Definition 3 (Lagrangian nodes) Let p 1, then the reference Lagrangian nodes on
the reference triangle Tbt corresponding to the polynomial degree p is the set of nodes
t,p

t,p
t,p
N
t,p
{zt,p
zk,1
, zk,2
)
k }k=1 , z
k = (

(3.63)

generated by the following algorithm

k := 0
for i = 0, . . . , p
for j = 0, . . . , p i
k := k + 1
t,p
zk,1
= j/p,
t,p
zk,2 = i/p.

(3.64)

t,p = (p + 1)(p + 2)/2.

Obviously, N
Figure 1.9 shows the Lagrangian nodes for p = 1, 2, 3. We define the reference
t,p
N
Lagrangian shape functions {t,p
l }l=1 by the relation
t,p
t,p
t,p
t,p
zt,p
l P (Kt ),
l (
k ) = lk , l, k = 1, . . . , N .

(3.65)

It is possible to derive the following explicit relation for the Lagrangian shape functions
on the reference triangle Tbt
p=1

t,1
t,1
t,1
1 = 1 ,
2 = 2 ,
3 = 3 ,

p=2

t,2
t,2
t,2
2 = 1 (21 1),
2 = 41 2 ,
3 = 2 (22 1),
t,2
t,2
t,2
4 = 42 3 , 5 = 3 (23 1), 6 = 43 1 ,

26

(3.66)

p=3

t,3
1
t,3
3
t,3
5
t,3
7
t,3
9

= 1 (31 1)(31 2)/2, t,3

2 = 91 2 (31 1)/2,
t,3
= 91 2 (32 1)/2, 4 = 2 (32 1)(32 2)/2,
= 93 1 (31 2)/2, t,3
6 = 271 2 3 ,
= 92 3 (32 1)/2, t,3
8 = 93 1 (32 1)/2,
= 92 3 (32 2)/2, t,3
10 = 3 (33 1)(33 2),

p=4

t,4
1
t,4
2
t,4
3
t,4
4
t,4
5
t,4
6
t,4
7
t,4
8
t,4
9
t,4
10
t,4
11
t,4
12
t,4
13
t,4
14
t,4
15

= 61 (41 1)(41 2)(41 3),

= 81 2 (41 1)(41 2)/3,
= 41 2 (41 1)(42 1),
= 81 2 (42 1)(42 2)/3,
= 62 (42 1)(42 2)(42 3),
= 83 1 (41 1)(43 2)/3,
= 321 2 3 (41 1),
= 321 2 3 (42 1),
= 82 3 (42 1)(42 2)/3,
= 43 1 (43 1)(41 1),
= 321 2 3 (43 1),
= 42 3 (42 1)(43 1),
= 83 1 (43 1)(43 2)/3,
= 82 3 (43 1)(43 2)/3,
= 63 (43 1)(43 2)(43 3).

Here 1 = x1 , 2 = x2 and 3 = 1 x1 x2 are the barycentric coordinates on Tbt .

1.3.2

The reference Lobatto shape functions

The disadvantage of the Lagrangian shape functions is that they are not hierarchical,
which means that
t,p

t,p+1

N
{q,p
q,p+1
}N
k=1
k }k=1 6 {
k

and

t,p

t,p+1

N
{t,p
t,p+1
}N
k=1 ,
k }k=1 6 {
k

p = 1, 2, . . . . (3.67)

This cause no troubles in case when our aim is to implement FEM for one given degree
of polynomial approximation only. In case, when we need to implement several degrees
of polynomial approximation (e.g., hp-FEM), then the use of the Lagrangian shape
functions requires an evaluation and a storing data structure (S1) for a lot of test
functions.
Therefore, it is more efficient to use Lobatto shape functions, which are hierarchical,
since they satisfy the conditions
t,p

t,p+1

N
{q,p
q,p+1
}N
k=1
k }k=1 {
k

and

t,p

t,p+1

N
{t,p
t,p+1
}N
k=1 ,
k }k=1 {
k

for triangles and quadrilaterals, respectively.

Now, we recall two well-known sets of functions.
27

p = 1, 2, . . . , (3.68)

Definition 4 (Legendre polynomials) Legendre polynomials Ln (x), n = 0, 1, . . . are

eigenvectors of the Legendre operator,


d
2 d
(3.69)
(1 x ) Ln (x) = n(n + 1)Ln , x (1, 1).

dx
dx
Legendre polynomials can be evaluated using the recursive formula

L0 (x) = 1,
L1 (x) = x,
k1
2k 1
Lk (x) =
xLk1 (x)
Lk2 (x), k = 2, 3, . . . .
k
k
Legendre polynomials form an orthogonal basis of the space L2 (1, 1), namely
 2
Z 1
for k = m,
2k+1
Lk (x)Lm (x) dx =
(3.70)
0
otherwise.
1

Definition 5 (Lobatto functions on (1, 1)) Let us define functions

1x
0 (x) =
,
(3.71)
2
x+1
,
1 (x) =
2
Z x
1

k (x) =
Lk1 ()d, k = 2, 3, . . . .
kLk1 k2 1
p
From (3.70) we have kLk1 k2 = 2/(2k 1). Obviously k (1) = 0 and k (1) = 0
since Lk1 , k 2 are orthogonal to L0 = 1, i.e.,
Z 1
Z 1
Lk (x)L0 (x) dx = 0, k = 2, 3, . . . .
(3.72)
Lk (x) dx =
1

Therefore, the Lobatto functions k , k = 0, . . . , p form a basis of the space Pp (1, 1).
For the purposes of our reference elements, we have to transform the Lobatto functions from interval (1, 1) onto (0, 1).
Definition 6 (Lobatto functions on (0, 1)) Let us define functions


x+1

k (x) = k
,
k = 0, 1, . . . ,
2

(3.73)

where k , k = 0, 1, . . . , are given by (5).

Obviously k (0) = 0, k (1) = 0 and the Lobatto functions k , k = 0, . . . , p form a basis

of the space Pp (0, 1).
For the definition of the shape function on triangles, it is convenient to decompose
the higher-order Lobatto functions k , k 2 from Definition 6 into products of the
form
k (x) = 0 (x)1 (x)k2 (x), k = 2, 3, . . . ,
(3.74)
where k2 P k2 ([0, 1]), k = 2, 3, . . . are the kernel functions.
Let us recall that we denote by {
vi } the vertices and by {
ei } the edges of the
reference element Tb.
28

Node type
Vertex
Edge
Bubble

Pol. degree
p1
p2
p3

# of shape functions # of objects

1
4
p1
4
(p 1)(p 2)
1

Table 1.1: Lobatto shape functions on Tbq

.
The reference Lobatto shape functions on the square
The Lobatto shape functions on the quadrilateral Tbq are constructed as tensor products
of the one dimensional Lobatto functions. They are grouped into 3 subsets according
to which node they belong.
l , l = 1, . . . , 4
The Lobatto vertex functions q,p
l , l = 1, . . . , 4 assigned to vertices v
v
l and vanishes at all remaining vertices. They are chosen bilinear,
are equal to one at v
defined via products of Lobatto functions as follows
x1 , x2 )
q,p
1 (
v
q,p
x1 , x2 )
v 2 (
q,p
v 3 (
x1 , x2 )
q,p
v 4 (
x1 , x2 )

=
=
=
=

0 (
x1 )0 (
x2 ),
1 (
x1 )0 (
x2 ),
1 (
x1 )1 (
x2 ),
0 (
x1 )1 (
x2 ),

(3.75)

for (
x1 , x2 ) Tbq .
For p 2, the Lobatto edge functions q,p
ej ,k , k = 2, . . . , p, j = 1, . . . , 4 are associated
with the corresponding edge ej . Their trace on ej coincides with the Lobatto functions
k and their trace vanishes on all remaining edges:
q,p
x1 , x2 )
e1 ,k (
q,p
e2 ,k (
x1 , x2 )
q,p
e3 ,k (
x1 , x2 )
q,p
e4 ,k (
x1 , x2 )

=
=
=
=

k (
x1 )0 (
x2 ),
1 (
x1 )k (
x2 ),
k (
x1 )1 (
x2 ),
0 (
x1 )k (
x2 ),

2k
2k
2k
2k

p,
p,
p,
p,

(3.76)

for (
x1 , x2 ) Tbq . Therefore, p 2 shape functions correspond to each edge of Tbq .
For p 3, the shape functions are completed by Lobatto bubble functions q,p
b,n1 ,n2 ,

b
which vanish everywhere on the boundary of Tq
q,p
x1 , x2 ) = n1 (
x1 )n2 (
x2 ),
b,n1 ,n2 , (

2 n1 p, 2 n2 p,

(3.77)

for (
x1 , x2 ) Tbq . The total number of the Lobatto shape functions on Tbq is summarized
in Table 1.1.
The reference Lobatto shape functions on the triangle
Lobatto basis of the space Vh consist again of vertex, edge and bubble shape functions
defined on Tbt .
29

Node type Pol. degree # of shape functions # of objects

Vertex
p1
1
3
Edge
p2
p1
3
Bubble
p3
(p 1)(p 2)/2
1
Table 1.2: Lobatto shape functions on Tbt
.
1, v
2, v
3 . Each function
t,p
t,p
Vertex functions t,p
3 , are assigned to vertices v
2 ,
1 ,
v
v
v
is equal to one at corresponding vertex and vanishes at remaining two vertices. Vertex
functions are linear, defined by following formulas
t,p
x1 , x2 ) = k ,
k (
v

k = 1, 2, 3,

(3.78)

where 1 = x1 , 2 = x2 and 3 = 1 x1 x2 are the barycentric coordinates on Tbt .

For p 2, the Edge functions t,p
ej ,k , k = 2, . . . , p, j = 1, 2, 3 coincide with onedimensional Lobatto functions on corresponding edge and vanish on all remaining
edges. They can be written using the kernel functions k2 , defined by (3.74), in
the form
t,p
x1 , x2 ) = 1 2 k2 (2 1 ),
e1 ,k (

2 k p,

t,p
x1 , x2 ) = 2 3 k2 (3 2 ),
e2 ,k (

2 k p,

t,p
x1 , x2 ) = 3 1 k2 (1 3 ),
e3 ,k (

2 k p.

(3.79)

b
For p 3, the bubble shape functions t,p
b,n1,n2 complete the basis on Tt . These
functions vanish on the whole element boundary. They will be defined using affine
coordinates and kernel functions as follows
t,p
b,n1,n2 = 1 2 3 n1 1 (3 2 )n2 1 (2 1 ),

1 n1 , 1 n2 , n1 +n2 p1, (3.80)

where n1 1 , n2 1 are the kernel functions defined by (3.74). Total number of the
Lobatto shape functions on triangular reference element is summarized in Table 1.2.
Remark 2 On both reference domains Tbq , Tbt shape functions coincide on edges with
Lobatto functions k , k = 0, 1, . . .. Moreover, the bubble shape functions does not affect
a possible compatibility between quadrilateral and triangular meshes This allows us to
combine triangular and quadrilateral elements in one hybrid mesh, see [1].

1.3.3

Global basis functions on Vh

In Sections 1.3.1 1.3.2, we introduced two possible sets of the shape functions defined
on the reference element Tb.
Here we describe a construction of a global basis of the finite-dimensional space
Vh , which is carried out by gluing together reference shape functions (transformed to
the physical element by mapping FT ). It is done in such a way, that resulting basis
functions satisfy conformity requirements of the space Vh H 1 ().
30

Remark 3 In order to fulfill the conformity requirement of global continuity of basis

functions, the orientation of physical mesh edges has to be taken into account. For
two-dimensional problems, it is possible to index edges of the reference element as well
as physical elements counterclockwise. Then the orientation of the reference element
iv
j Tb coincides with the orientation of the edge e = vi vj T such
edge e = v
that vi = FT (
vi ) and vj = FT (
vj ). In other case, when the orientations differ, all
reference edge functions have to be transformed by a suitable mapping, which inverts
the parameterization of the edge e.
In order to avoid a complication with a notation, we assume that Th is a triangular
grid. All statements are valid also for quadrilateral grid, only the superscript t has to
be replaced by the superscript q and the number of vertices and edges of T Th from
3 to 4.
Global Lagrangian shape functions

Let p 1 be a given degree of polynomial approximation and zt,p

i , i = 1, . . . , N be the
reference Lagrangian nodes defined by (3.64). For each T Th , we define the nodes

zt,p
zt,p
i ), i = 1, . . . , N ,
T,i = FT (

(3.81)

where FT maps Tbt onto T . Let us note that some of these nodes coincides, namely
t,p

zt,p
T,i = zT ,i for T 6= T sharing a vertex or an edge.
Moreover, we define the set of Lagrangian nodes
o
n
t,p
t,p

.
(3.82)
L := z such that (T, i) Th {1, . . . , N } satisfying z = z

It means that L t,p contains all zt,p

T,i , i = 1, . . . , N , T Th lying inside of (we exclude
nodes lying on ) and each node appears in L t,p only one times. We index the set
L t,p by L t,p = {zt,p
j , j = 1, . . . , NL }.
The global Lagrangian basis functions t,p
i , i = 1, . . . , NL is defined by
t,p t,p
t,p
i (x) Vh , i (zj ) = ij ,

i, j = 1, . . . , NL .

(3.83)

Let as note that each global Lagrangian basis function t,p

i , i = 1, . . . , NL can be,
due to (3.83), classified as
vertex basis function if the corresponding Lagrangian node is an vertex of Th , its
support is formed by a patch of elements T Th having this vertex common,
edge basis function if the corresponding Lagrangian node lies on an edge of Th ,
its support is formed by a patch of two elements T Th sharing this edge,
bubble basis function if the corresponding Lagrangian node lies in the interior of
a T Th , which is also its support.
The global Lagrangian basis function are defined in the following way.
31

Definition 7 (Vertex basis function) Let zt,p

i is a Lagrangian node which is identical
with the vertex vl of Th . Then the corresponding vertex Lagrangian basis function t,p
i
is a continuous function defined on which equals to one at vl and vanishes outside
of the patch Sl formed by all elements T Th sharing the vertex vl . For each element
T j Sl ,
t,p
t,p
x),
x = FT (
x),
(3.84)
i (x)|Tj =
k (
k = FT1 (zt,p
where t,p
i ).
k is the reference Lagrangian shape function such that z
Definition 8 (Edge basis function) Let zt,p
i is a Lagrangian node which lies on an edge
el of Th . Then the corresponding edge Lagrangian basis function t,p
i is a continuous
function defined on which vanishes outside of the patch Sl formed by all elements
sharing the edge el . For each element Tj Sl ,
t,p
t,p
x),
i (x)|Tj =
k (

x = FT (
x),

(3.85)

k = FT1 (zt,p
where t,p
i ).
k is the reference Lagrangian shape function such that z
Definition 9 (Bubble basis function) Let zt,p
i is a Lagrangian node which lies in the interior of an element T Th . Then the corresponding bubble Lagrangian basis function
t,p
i which vanishes outside of T ,
t,p
t,p
x),
i (x)|T =
k (

x = FT (
x),

(3.86)

k = FT1 (zt,p
where t,p
i ).
k is the reference Lagrangian shape function such that z
Global Lobatto shape functions
Let p 1 be a given degree of polynomial approximation and The global basis B of
Vh consists of three type of basis functions:
Vertex basis functions t,p
vl , l = 1, . . . , N , which are associated to each inner
vertex vl , l = 1, . . . , N of Th . The support of t,p
vl is formed by a patch of
elements T Th having vl as an vertex.
Edge basis functions t,p
el ,k , 2 k p, l = 1, . . . , E, which are associated to each
inner edge el , l = 1, . . . , E of Th . The support of t,p
el ,k , 2 k p is formed by
the patch of two elements from T Th sharing the face el .
Bubble basis functions t,p
T,n1,n2 , 1 n1 , 1 n2 , n1 + n2 p 1, T Th , which
are associated to each element T Th . Their support consists just from the one
element T .
The global Lobatto basis function are defined in the following way.
Definition 10 (Vertex basis function) The vertex basis function t,p
vl associated with
a mesh vertex vl , l = 1, . . . , N is a continuous function defined on which equals to
32

one at vl and vanishes outside of the patch Sl formed by all elements T Th sharing
the vertex vl . For each element Ti Sl ,
t,p
t,p
x),
vl (x)|Ti =
k (
v

x = FT (
x),

(3.87)

k of Tbt such
where t,p
k is the reference vertex shape function associated to the vertex v
v
k = FT1 (vl ).
that v

Definition 11 (Edge basis function) The edge function t,p

el ,k associated with a mesh
edge el and k = 2, . . . , p, is a continuous function defined on which vanishes outside
of the patch Sl formed by all elements sharing the edge el . For each element T Sl ,
t,p
t,p
x),
el ,k (x)|Ti =
ej ,k (

x = FT (
x),

(3.88)

where t,p
j of Tbt such that
ej ,k is the reference edge shape function associated to the edge e
1
ej = FT (el ).
Definition 12 (Bubble basis function) The bubble function t,p
T,n1 ,n2 associated with a
T Th and 1 n1 , 1 n2 , n1 + n2 p 1, k = 2, . . . , p, is a continuous function
defined on which vanishes outside of T ,
t,p
t,p
x),
T,n1 ,n2 (x)|T =
b,n1 ,n2 (

x = FT (
x),

(3.89)

b
where t,p
b,n1 ,n2 is the reference bubble shape function associated to Tt .

Remark 4 We already mentioned that a use of hierarchical finite elements is advantageous in case where different degree of polynomial approximation are used on different
elements. However, in case of conforming finite element methods, the basis functions
have to be continuous. Therefore, it is necessary to construct a transition from one
degree of polynomial approximation to the another one, see [1].

1.4

Discontinuous finite elements

In the previous sections of this chapter, we discussed implementations of conforming

finite elements methods, where the finite element spaces consists of continuous functions. In this section we focus on an implementation of methods based on discontinuous
approximation, namely the discontinuous Galerkin (DG) method.
Although the DG formulation is more complicated than for the conforming FEM
since additional terms appear there, its implementation is usually more simple since
the test functions are chosen separately for each T Th .

1.4.1

DG discretization of the model problem

In the following we recall the DG discretization of the model problem (1.1) (1.2). We
consider here the IIPG variant of the DG method since it has the simplest formulation.
33

Let Th (h > 0) be a partition of the closure of the domain into a finite number
of closed d-dimensional elements, where we admit a combination of triangular and
quadrilateral elements including hanging nodes. For simplicity, we write again,
[
=
T.
(4.90)
T Th

Moreover, Fh denotes the set of all edges (faces) of Th .

We assume that for each element T Th there exist a mapping
FT = FT (
x) = (FT,1 (
x), . . . , FT,d (
x)) : Tb Rd

such that FT (Tb) = T,

(4.91)

where Tb is either the reference triangle Tbt or the reference quadrilateral Tbq given by
(2.33) or (2.34), respectively.
We define the space of discontinuous piecewise polynomial functions Shp by
Shp = {vh , vh L2 (), vh |T FT P p (Tb) T Th },

(4.92)

where Tb denotes either Tbt or Tbq depending an the type of T . Furthermore, P p is the
space of polynomials of degree p on Tb defined by (2.51) and/or (2.54).
Since DG methods use a discontinuous approximation, it is possible to define the
space P p (Tb) in (4.92) by several manners, namely
P t,p (Tbt ) for triangles and P t,p (Tbq ) for quadrilaterals,
P t,q (Tbt ) for triangles and P t,q (Tbq ) for quadrilaterals,

P t,p (Tbt ) for triangles and P t,q (Tbq ) for quadrilaterals,

In this text we consider the first variants, the others can be obtain with a small modification.
Moreover, ne denotes a unit normal vector to e Fh , its orientation is arbitrary
but fixed. Furthermore, the symbols hvh ie and [[vh ]]e denote the mean value and the
jump in direction of ne of vh Shp on inner face e Fh (e ). For e ,
we
R put [[vh ]]e = hvh ie = vh |e . Finally, in case that ne or [[]]e or hie are arguments of
. . . d, e Fh , we omit very often the subscript e and write simply n or [[]] or hi,
e
respectively.
Now, we are ready to introduce the DG approximation of the model problem.
Definition 13 We say that uh Shp is the approximate DG solution of (1.1) (1.2)
if
XZ
XZ
uh vh dx
huh i n[[vh ]] d
(4.93)
T Th

eFh

eFh

XZ

[[uh ]][[vh ]] d =

where denotes the penalty parameter.

34

g vh dx

vh Shp ,

1.4.2

hp-discontinuous Galerkin method

Let us assume that the elements of Th are indexed by

Th = {T , = 1, . . . , M }.

(4.94)

DGM allows the use of different polynomial degrees over different elements without
a necessity to construct a transition from one degree of polynomial approximation to
the another one. In this section, we will consider this general case since its treatment
is relatively simple. We assign a positive integer p (local polynomial degree) to each
T Th . In case when the one uses the same degree of polynomial approximation in
the whole , we have p = p T Th . We define the vector
p = {p , T Th }.

(4.95)

Over the triangulation Th , we define the finite dimensional space of discontinuous

piecewise polynomial functions associated with the vector p by
Shp = {v; v L2 (), v| FK Pp (Tb) = 1, . . . , M },

(4.96)

where P p (Tb) denotes the space of all polynomials on Tb of degree p , K Th .

Now, we can simply formulate the hp-approximation of the model problem. We
replace
Definition 14 We say that uh Shp is the hp-approximate DG solution of (1.1)
(1.2) if
XZ
XZ
uh vh dx
huh i n[[vh ]] d
(4.97)
T Th

eFh

eFh

XZ

[[uh ]][[vh ]] d =

g vh dx

vh Shp ,

where denotes the penalty parameter.

In comparison with (1.4) the face integrals have to be implemented. On the other
hand, their implementation of DGM is easier since a discontinuous approximation is
used therefore it is possible constructed basis functions with a support consisting from
one T Th .
In the following we describe the construction of basis functions of Shp . First, we introduce a (global) basis of the space Shp as a composition of local shape constructed for
each T Th separately. Then, we introduce the construction of local shape functions.
Finally, we explain the evaluation of the face integrals.

1.4.3

Definition of a DG basis

Since Shp is a space of discontinuous piecewise polynomial functions, it is possible to

consider a set of linearly independent polynomial functions on T for each T Th
n
B =
,j ; ,j Shp , supp(,j ) T ,
(4.98)
o
,j are linearly independent for j = 1, . . . , dof ,
35

p
d=2
d=3

1
3
4

2
6
10

3
10
20

4
15
35

5
21
56

Table 1.3: Values of dof for p = 1, . . . , 5 and d = 2, 3

where
dof =

1
2
1
6

P2

j=1 (p

P3

+ j)

for d = 2
,

j=1 (p

+ j)

= 1, . . . , M,

(4.99)

for d = 3

denotes the number of local degrees of freedom for each element T Th . The values
of dof are shown in Table 1.3 for p = 1, . . . , 5 and d = 2, 3. We call B the local basis
on T . For the construction of the basis B , = 1, . . . , M , see Section 1.4.4.
A composition of the local bases B , = 1, . . . , M defines a basis of Shp , i.e.,
B = {j ; j Shp , j = 1, . . . , dof}.

(4.100)

By dof, we denote the dimension of Shp (=number of elements of the basis B) which
is equal to
M
X
dof =
dof .
(4.101)
=1

where dof is given by (4.99).

Therefore, a function uh Shp can be written in the form
uh (x) =

dof
M X
X

u,j ,j (x),

x , u,j R, j = 1, . . . , dof , = 1, . . . , M (4.102)

=1 j=1

where u,j R, j = 1, . . . , dof , = 1, . . . , M . Moreover, for uh Shp , we define the

vector of its basis coefficients by
dof
,
u = {u,j }=1,...,M
j=1,...,dof R

k = 0, 1, . . . , r.

(4.103)

Therefore, using (4.102) (4.103) we have an isomorphism

uh Shp

1.4.4

u Rdof .

(4.104)

Construction of the DG basis functions

In order to achieve a reasonable efficiency, hierarchical basis functions should be employed. Let
p = max p
=1,...,M

and let P p(Tb) denotes the space of all polynomials on Tb of degree pp. Let Sp be a
hierarchical basis of P p(Tb) which will be specified later. We call Sp the reference basis
and its elements the reference shape function.
36

Furthermore, let F := FT , = 1, . . . , M , be the mapping introduced by (4.91)

such that F (Tb) = T . We put
Tb, j = 1, . . . , dof },
B := {,j ; ,j (x) = ,j (F (
x)) = j (
x), x

(4.105)

which defines a basis B introduced in (4.98) for each element T Th separately.

Finally, (4.100) defines the global basis of Shp .
Reference shape functions
It is possible to use an arbitrary set of (hierarchical or non-hierarchical) shape functions
for the definition of Sp. An example are the Lagrangian and Lobatto shape functions
introduced in Sections 1.3.1 and 1.3.2, respectively. However, since Shp is the space
of discontinuous functions, it is not necessary to construct vertex, edge and bubble
functions in order to fulfil the continuity of the basis functions.
Therefore, it is possible to use the Taylor shape functions given by
(
)
d
X
i ,...,i (
x1 , . . . , xd ) = d (
xi xc )ii ; i1 , . . . , id 0,
ij p ,
(4.106)
1

i=1

j=1

where (
xc1 , . . . , xcd ) is the barycentre of Tb. (It is possible to use also xc1 = . . . = xcd = 0).
The advantage of the Taylor shape functions is a simpler implementation since we
need not to distinguish among the vertex, edge and bubble functions, one subroutine
is sufficient.
However, a direct use the Taylor shape functions is not too much practical since
these functions are not normalized which can cause problems in the solution of the
corresponding linear algebraic systems. Moreover, in many application, namely in
evolution problems, it is advantageous to employ an L2 -orthonormal basis because the
corresponding linear systems have favorable computational properties. Let us note
that a suitable preconditioning can reduce many drawback following from a possible
use of a non-normalized shape functions.
Therefore, we employ the local character of the shape functions and construct a
basis of Sp which is orthonormal with respect to the L2 -scalar product. The simplest
way, how to obtain the L2 -orthonormal basis is the performance of the Gram-Schmidt
orthogonalization process applied to the functions from (4.106). Then we obtain the
orthonormal set of the reference shape functions
Sp := {j , j = 1, . . . , dof },

(4.107)

where dof is given by (4.99) with p := p.

It is well known fact that the Gram-Schmidt orthogonalization is an unstable algorithm. However, in practice, this approach works since dof is usually not too large
number (between 10 and 200) and moreover, the instability of the Gram-Schmidt process cause a small violation of the L2 -orthogonality of the reference shape functions
which do not affect good computational properties of the corresponding linear algebraic
systems.
37

The Gram-Schmidt orthogonalization on the reference element can be carried out

symbolically by some suitable software (e.g., Maple) or numerically. Our numerical
experiments based on the numerical realization of the Gram-Schmidt orthogonalization
work with success for p = 10 dof = 45 (for d = 2).
Let us finally note, the global basis of Shp obtained from (4.100), (4.105) and (4.107)
is orthonormal if the mappings F := FT , = 1, . . . , M are linear. Otherwise, some
violations of the L2 -orthogonality are presented. However, the mappings F , =
1, . . . , M are close to linear mappings, hence these violations are small and (again) do
not affect good computational properties of the corresponding linear algebraic systems.

1.4.5

Evaluation of volume integrals

The evaluation of the volume integrals in (4.97) is very simple. Similarly as in the
conforming FEM, we have to compute the terms
XZ
,i ,j dx,
i = 1, . . . , dof , j = 1, . . . , dof , , = 1, . . . , M, (4.108)
T

T Th

where ,i , ,j B (= basis of Shp ). Since each ,i vanish outside of T , we have

Z
XZ
,i ,j dx, = ,
,i ,j dx,
(4.109)
T Th

for all i = 1, . . . , dof , j = 1, . . . , dof , , = 1, . . . , M . Therefore, the stiff matrix S

is block diagonal with blocks S , = 1, . . . , M given by
Z
dof
S = {S,i,j }i,j=1 , S,i,j =
,i ,j dx, i, j = 1, . . . , dof .
(4.110)
T

These integral can be simply evaluated wit the aid of (2.58), namely
Z

,i ,j dx =
T

Z X
d 
Tb l=1

b i

JFT
T

 

b j | det JF | d

JFT
x,
T
T
l

(4.111)

where i , j are the reference shape functions from Sp given by (4.107). In comparison
with the conforming FEM, situation is more simple, since we know that ,i corresponds
to i , i = 1, . . . , dof , compare with (2.56).

1.4.6

Evaluation of face integrals

In this section we deal with the evaluation of the face integrals in (4.97). Let us note
that the necessity to evaluate face integrals appears also in the conforming finite element
method, e.g., for the implementation of general boundary conditions, see Section 1.1.5.
Let e Fh be an edge of T Th . We call e the physical edge. Our aim is to
evaluate the integrals
Z
Z
f (x) d,
f (x) n d,
(4.112)
e

38

where n is the normal vector to e and f : e R, f : e R2 are given functions.

Let us recall the definition of the face integral. Let = (1 , 2 ) : [0, 1] e be a
parameterization of the edge e. Then
Z
Z 1
q

2

2
(4.113)
f (x) d =
f ((t)) 1 (t) + 2 (t) dt,
e

where i (t) , i = 1, 2 denotes the derivative of 1 (t) with respect to t.

In order to evaluate the integrals (4.112), we use the approach based on the reference
element. Let e be an edge of the reference element Tb such that T = FT (Tb) and
e = FT (
e). We call e the reference edge. Let
e(t) = (
x
xe,1 (t), xe,2 (t)) : [0, 1] e

(4.114)

be a parameterization of the reference edge e keeping the counterclockwise orientation

of the element. Namely, using the notation from Figure 1.6, we have
if Tb = Tbt then

e1 (t) := (t, 0), t (0, 1),

x
e2 (t) := (1 t, t), t (0, 1),
x
e3 (t) := (0, 1 t), t (0, 1),
x

if Tb = Tbq then

e1 (t) := (t, 0), t (0, 1),

x
e2 (t) := (1, t), t (0, 1),
x
e3 (t) := (1 t, 1), t (0, 1),
x
e4 (t) := (0, 1 t), t (0, 1).
x

(4.115)

by
Moreover, we define the infinitesimal increase of x
d
xe(t) :=

d
e(t) Rd ,
x
dt

(4.116)

namely
if Tb = Tbt then

d
xe1 := (1, 0), t (0, 1),
d
xe2 := (1, 1), t (0, 1),
d
xe3 := (0, 1), t (0, 1),

if Tb = Tbq then

d
xe1
d
xe2
d
xe3
d
xe4

(4.117)

:= (1, 0), t (0, 1),

:= (0, 1), t (0, 1),
:= (1, 0), t (0, 1),
:= (0, 1), t (0, 1).

Therefore, the physical edge e is parameterized by

e:

x = FT (
xe(t)) = (FT,1 (
xe(t)), FT,2 (
xe(t)))
= (FT,1 (
xe,1 (t), xe,2 (t)), FT,2 (
xe,1 (t), xe,2 (t))) ,
39

(4.118)
t [0, 1].

The first integral from (4.112) is given by

Z

f (x) d =
e

=
=

1
0
1
0
1

2
X
d
FT,i (
xe(t))
f (FT (
xe(t)))
dt
i=1

!1/2

dt

2
X
FT,i (
xe(t)) d
xe,j (t)
f (FT (
xe(t)))

dt
j
i,j=1

(4.119)
!1/2

dt

f (FT (
xe(t))) |JFT (
xe(t))d
xe| dt,
0

where JFT is the Jacobian matrix of the mapping FT multiplied by the vector d
xe (=
the infinitesimal increase) and | | is the Euclid norm of the vector. Let us note that if
FT is a linear mapping then e is a straight edge and |JFT (
xe(t))d
xe(t)| is equal to its
length.
Now, we evaluate the second integral from (4.112). For simplicity, we restrict only
to the case d = 2. Let r e be the tangential vector to e (if e is a straight line then
r e = e). Using (4.118) and (4.119), we evaluate r e at x(t) = FT (
xe(t), t [0, 1] by
r e (x(t)) = (r1 (x(t)), r2 (x(t)))


d
:=
FT (
xe(t)) = JFT,1 (
xe(t))d
xe, JFT,2 (
xe(t))d
xe .
dt

(4.120)

ne (x(t)) = (ne,1 (x(t)), ne,2 (x(t))) :

ne,1 (x(t)) = re,2 (x(t)), ne,2 (x(t)) = re,1 (x(t)),

(4.121)

Now, be a rotation we obtain the normal vector ne oriented outside of T , namely

where (te,1 , te,2 ) is the tangential vector to e given by (4.120). Here it is important that
the counterclockwise orientation of the elements is considered. Therefore, from (4.120)
and (4.121), we have


JFT,2 (
xe(t))d
xe, JFT,1 (
xe(t))d
xe .

ne (x(t)) =

(4.122)

Let us note that ne (x(t)) is not normalized, it is necessary to divide it by |ne (x(t))| =
|JFT (
xe(t))d
xe|. Finally, similarly as in (4.119), we obtain
Z

f (x) n d =
e

f (FT (
xe(t)))
0

ne (x(t))
|JFT (
xe(t))d
xe| dt,
|ne (x(t))|

(4.123)

f (FT (
xe(t))) ne (x(t)) dt,
0

where ne (x(t)) is given by (4.122). Let us note that if FT is a linear mapping then e
is a straight edge and |ne (x(t))| is equal to its length.
40

1.4.7

Data structure

In order to evaluate the integrals (4.111), (4.119) and (4.123), it is enough to evaluate
and store the following data.
(S1) for each T Th , the determinant of the Jacobi matrix det JFT and the transposed
matrix of the inversion of the Jacobi matrix JFTT ,
with the gradients
b i , i = 1, . . . N
on Tb.
(S2) the shape functions i , i = 1, . . . N

Comparison with the data structures of FEM implementation, we see that (S1) and
(S2) are in fact the same but (S3) is missing. This is caused by the fact that we have the
direct connection between the basis functions and the corresponding shape functions,
see relation (4.105).

1.5

Numerical quadratures

In the previous sections we assumed that we are able to evaluate all integrals (over
the reference element or the reference edge) exactly. It is true for the model problem
which is linear and therefore all integrands are polynomial functions. However, when a
problem is nonlinear, it is necessary to use an approximation, namely suitable numerical quadratures. Even in case, when the problem is linear, the implementation with
numerical quadratures is more simple.
There exists many numerical quadrature rules with many advantages and disadvantages. In the following we exhibits some basis quadrature rules for edge integrals and
volume integrals over the reference square and over the reference triangle, which exhibit
a reasonable compromise between accuracy and the implementation complexity.

1.5.1

Edge quadratures

We consider the Gauss quadrature rules given by

Z

f (x) dx Gk (f ) :=
0

k
X

wi f (xi ),

(5.124)

i=1

where wi , i = 1, . . . , k and xi , i = 1, . . . , k are the Gauss weights and the Gauss nodes,
respectively. Their values can be found in many textbooks, e.g., in [3]. Table 1.4 these
values for Gk , k = 1, . . . , 12.

1.5.2

Quadrilaterals quadratures

In order to integrate over the reference square Tbq , we employ the bi-Gauss quadrature
rules given by
Z
k X
k
X
bi
f (x) dx Gk (f ) :=
wi wj f (xij ),
(5.125)
Tbq

i=1 j=1

41

Gk
G1
G2
G3
G3
G4
G4
G5
G5
G5
G6
G6
G6
G7
G7
G7
G7
G8
G8
G8
G8
G9
G9
G9
G9
G9
G10
G10
G10
G10
G10
G11
G11
G11
G11
G11
G11
G12
G12
G12
G12
G12
G12

j
1
1
1
3
1
3
1
3
5
1
3
5
1
3
5
7
1
3
5
7
1
3
5
7
9
1
3
5
7
9
1
3
5
7
9
11
1
3
5
7
9
11

wj
1.00000000000000
0.50000000000000
0.27777777777778
0.27777777777778
0.17392742256873
0.32607257743127
0.11846344252809
0.28444444444444
0.11846344252809
0.08566224618959
0.23395696728635
0.18038078652407
0.06474248308443
0.19091502525256
0.19091502525256
0.06474248308443
0.05061426814519
0.15685332293894
0.18134189168918
0.11119051722669
0.04063719418079
0.13030534820147
0.16511967750063
0.13030534820147
0.04063719418079
0.03333567215434
0.10954318125799
0.14776211235738
0.13463335965500
0.07472567457529
0.02783428355809
0.09314510546387
0.13140227225512
0.13140227225512
0.09314510546387
0.02783428355809
0.02358766819326
0.08003916427167
0.11674626826918
0.12457352290670
0.10158371336153
0.05346966299766

xj
0.50000000000000
0.21132486540519
0.11270166537926
0.88729833462074
0.06943184420297
0.66999052179243
0.04691007703067
0.50000000000000
0.95308992296933
0.03376524289842
0.38069040695840
0.83060469323313
0.02544604382862
0.29707742431130
0.70292257568870
0.97455395617138
0.01985507175123
0.23723379504184
0.59171732124782
0.89833323870681
0.01591988024619
0.19331428364970
0.50000000000000
0.80668571635030
0.98408011975381
0.01304673574141
0.16029521585049
0.42556283050918
0.71669769706462
0.93253168334449
0.01088567092697
0.13492399721298
0.36522842202383
0.63477157797617
0.86507600278702
0.98911432907303
0.00921968287664
0.11504866290285
0.31608425050091
0.56261670425573
0.79365897714331
0.95205862818524

wj

xj

2
2

0.50000000000000
0.44444444444444

0.78867513459481
0.50000000000000

2
4
2
4

0.32607257743127
0.17392742256873
0.23931433524968
0.23931433524968

0.33000947820757
0.93056815579703
0.23076534494716
0.76923465505284

2
4
6
2
4
6

0.18038078652407
0.23395696728635
0.08566224618959
0.13985269574464
0.20897959183673
0.13985269574464

0.16939530676687
0.61930959304160
0.96623475710158
0.12923440720030
0.50000000000000
0.87076559279970

2
4
6
8
2
4
6
8

0.11119051722669
0.18134189168918
0.15685332293894
0.05061426814519
0.09032408034743
0.15617353852000
0.15617353852000
0.09032408034743

0.10166676129319
0.40828267875218
0.76276620495816
0.98014492824877
0.08198444633668
0.33787328829810
0.66212671170190
0.91801555366332

2
4
6
8
10
2
4
6
8
10

0.07472567457529
0.13463335965500
0.14776211235738
0.10954318125799
0.03333567215434
0.06279018473245
0.11659688229600
0.13646254338895
0.11659688229600
0.06279018473245

0.06746831665551
0.28330230293538
0.57443716949082
0.83970478414951
0.98695326425859
0.05646870011595
0.24045193539659
0.50000000000000
0.75954806460341
0.94353129988405

2
4
6
8
10
12

0.05346966299766
0.10158371336153
0.12457352290670
0.11674626826918
0.08003916427167
0.02358766819326

0.04794137181476
0.20634102285669
0.43738329574427
0.68391574949909
0.88495133709715
0.99078031712336

Table 1.4: Gauss quadratures Gk , k = 1, . . . , 12 with weights wi and nodes xi

42

where wi , i = 1, . . . , k are the Gauss weights introduced in (5.124) and

xij = (xi , xj ),

i, j = 1, . . . , k,

(5.126)

where xi , i = 1, . . . , k are the Gauss nodes introduced in (5.124). Obviously, the

bi-Gauss quadrature rules are derived by the integration with respect to the first variable and then with respect to the second one. Similarly, it is possible to derive the
quadrature rule for the reference cube for d = 3.

1.5.3

Triangles quadratures

In order to integrate over the reference triangle Tbt , we employ the Dunavant quadrature
rules given by
Z
mk
X
f (x) dx Dk (f ) :=
wi f (xi ),
(5.127)
Tbt

i=1

where mk is the number of integration nodes, wi , i = 1, . . . , k are the Dunavant weights

and xi = (i,1 , i,2 , i,3 ), i = 1, . . . , mk , are the Dunavant integration nodes. Their
values can be found, e.g., in [2], Table 1.5 shows these values for Dk , k = 1, . . . , 6.

1.5.4

Data structure

If the integrals appearing in FEM and/or DGM are evaluated with the aid of numerical
quadratures, we should revise the data structures (S1), (S2), (S3) introduced in Sections
1.2.4 and 1.4.7. Therefore, instead of (S1) and (S2), it is enough to evaluate and store
the following data.
(D1) for each T Th , the determinant of the Jacobi matrix det JFT and the transposed
matrix of the inversion of the Jacobi matrix JFTT in the used edge and volume
quadrature nodes,
with the gradients
b i , i = 1, . . . N
on Tb in
(D2) the shape functions i , i = 1, . . . N
the used edge and volume quadrature nodes.

However, there is a natural question which quadrature rules should be employed.

For linear or closely linear problem, it is sufficient to use the quadrature rules such
that integrate exactly polynomials of the degrees which arise in the volume and edge
integrals. For strongly nonlinear problems, more accurate quadrature rule should be
used, their degrees can be found empirically.

1.6

Basic visualization techniques

The result of a finite element computation is a (continuous or discontinuous) piecewise

polynomial function uh Vh . Usually, we are not interested of its analytical expression
but we need its visualization. In this Section we discuss these aspects.
43

Dk
D1
D2
D2
D2
D3
D3
D3
D3
D4
D4
D4
D4
D4
D4
D5
D5
D5
D5
D5
D5
D5
D6
D6
D6
D6
D6
D6
D6
D6
D6
D6
D6
D6
D7
D7
D7
D7
D7
D7
D7
D7
D7
D7
D7
D7
D7

j
1
1
2
3
1
2
3
4
1
2
3
4
5
6
1
2
3
4
5
6
7
1
2
3
4
5
6
7
8
9
10
11
12
1
2
3
4
5
6
7
8
9
10
11
12
13

wj
1.00000000000000
0.33333333333333
0.33333333333333
0.33333333333333
-.56250000000000
0.52083333333333
0.52083333333333
0.52083333333333
0.22338158967801
0.22338158967801
0.22338158967801
0.10995174365532
0.10995174365532
0.10995174365532
0.22500000000000
0.13239415278851
0.13239415278851
0.13239415278851
0.12593918054483
0.12593918054483
0.12593918054483
0.11678627572638
0.11678627572638
0.11678627572638
0.05084490637021
0.05084490637021
0.05084490637021
0.08285107561837
0.08285107561837
0.08285107561837
0.08285107561837
0.08285107561837
0.08285107561837
-.14957004446768
0.17561525743321
0.17561525743321
0.17561525743321
0.05334723560884
0.05334723560884
0.05334723560884
0.07711376089026
0.07711376089026
0.07711376089026
0.07711376089026
0.07711376089026
0.07711376089026

j,1
0.33333333333333
0.66666666666667
0.16666666666667
0.16666666666667
0.33333333333333
0.60000000000000
0.20000000000000
0.20000000000000
0.10810301816807
0.44594849091597
0.44594849091597
0.81684757298046
0.09157621350977
0.09157621350977
0.33333333333333
0.05971587178977
0.47014206410511
0.47014206410511
0.79742698535309
0.10128650732346
0.10128650732346
0.50142650965818
0.24928674517091
0.24928674517091
0.87382197101700
0.06308901449150
0.06308901449150
0.05314504984482
0.31035245103378
0.63650249912140
0.31035245103378
0.63650249912140
0.05314504984482
0.33333333333333
0.47930806784192
0.26034596607904
0.26034596607904
0.86973979419557
0.06513010290222
0.06513010290222
0.04869031542532
0.31286549600487
0.63844418856981
0.31286549600487
0.63844418856981
0.04869031542532

j,2
0.33333333333333
0.16666666666667
0.16666666666667
0.66666666666667
0.33333333333333
0.20000000000000
0.20000000000000
0.60000000000000
0.44594849091597
0.44594849091597
0.10810301816807
0.09157621350977
0.09157621350977
0.81684757298046
0.33333333333333
0.47014206410511
0.47014206410511
0.05971587178977
0.10128650732346
0.10128650732346
0.79742698535309
0.24928674517091
0.24928674517091
0.50142650965818
0.06308901449150
0.06308901449150
0.87382197101700
0.31035245103378
0.63650249912140
0.05314504984482
0.05314504984482
0.31035245103378
0.63650249912140
0.33333333333333
0.26034596607904
0.26034596607904
0.47930806784192
0.06513010290222
0.06513010290222
0.86973979419557
0.31286549600487
0.63844418856981
0.04869031542532
0.04869031542532
0.31286549600487
0.63844418856981

j,3
0.33333333333333
0.16666666666667
0.66666666666667
0.16666666666667
0.33333333333333
0.20000000000000
0.60000000000000
0.20000000000000
0.44594849091597
0.10810301816807
0.44594849091597
0.09157621350977
0.81684757298046
0.09157621350977
0.33333333333333
0.47014206410511
0.05971587178977
0.47014206410511
0.10128650732346
0.79742698535309
0.10128650732346
0.24928674517091
0.50142650965818
0.24928674517091
0.06308901449150
0.87382197101700
0.06308901449150
0.63650249912140
0.05314504984482
0.31035245103378
0.63650249912140
0.05314504984482
0.31035245103378
0.33333333333333
0.26034596607904
0.47930806784192
0.26034596607904
0.06513010290222
0.86973979419557
0.06513010290222
0.63844418856981
0.04869031542532
0.31286549600487
0.63844418856981
0.04869031542532
0.31286549600487

Table 1.5: Dunavant quadratures Dk , k = 1, . . . , 6 with weights wi and nodes xi

44

0.8

0.6

0.4

0.2

0
0

0.2

0.4

0.6

0.8

Figure 1.10: The basis graphical outputs: isolines (left) and a color map (right)

1.6.1

Types of visualization

There are several usual graphical outputs, which can be required based on the considered problem. Namely
isolines/isosurfaces of the solutions are the lines (for d = 2) or surfaces (for
d = 3) consisting of points of the computational domains where the approximate
solution equals to a given value, i.e., the set {x ; uh (x) = g} is the isoline
(isosurface) corresponding to the value g. Figure 1.10, left shows the isolines of
the piecewise linear approximation of the function
u(x1 , x2 ) = 2(x21 +x22 )1/4 x1 x2 (1x1 )(1x2 ),

(x1 , x2 ) (0, 1)(0, 1) (6.128)

for the values g = 0.01n, n = 0, 1, . . . , 18.

color maps of the solutions is the visualization technique where each element of
the mesh is drawn by a color which corresponds to a given value. Figure 1.10,
right shows the color map of the function (6.128). Let us note that color maps
very ofter smear details of approximate solution.
3D plot of the solution (only for d = 2) are the graph of the surface
{(x1 , x2 , x3 ); x3 = uh (x1 , x2 ), (x1 , x2 ) }.
Figure 1.11, left shows the 3D plot function (6.128). This graphical output contains the whole information of the approximate solution. On the other hand if
the approximate solution is very complex this type of output is hardly to see.
boundaries visualization are important, e.g., in aerodynamics, when we need to
know a pressure distribution on the airplane surface.
cuts of the solution help to see the solution inside of the computational domain,
Figure 1.11, right shows the diagonal cut from the left-below to right-up of the
function (6.128).
45

0.16
0.14

0.16
0.14
0.12
0.1
0.08
0.06
0.04
0.02
0
-0.02

0.12
0.1
0.08
0.06
0.04

10.8
0.60.4
0.2

0 0

0.2

0.4

0.6

0.8

0.02
0
-0.02
0

0.2

0.4

0.6

0.8

Figure 1.11: The basis graphical outputs: a 3D plot (left) and a diagonal cut (right)

1.6.2

Software for visualization

Within this section we briefly describe types of software which can be used for the
visualization of finite element solutions. Our aim is not to give a complete list of all
available software but give to the readers some overview about possibilities.
We suppose that the software for the visualization can be split into the following
groups.
Commercial graphical software usually represent a complete tool for visualization. Let us mention as an example the code Techplot which is able to draw a
continuous piecewise linear or piecewise constant approximate solution for d = 2
as well as d = 3 meshes. It allows a wide range of graphical tools, e.g., slices,
isosurfaces, cuts, animations etc. However, Techplot uses some postprocessing
when smooths the given approximate solution. This is advantageous for effect
of visualization but it brings some additional perturbation and/or errors. A
use of the commercial software does not require any pre-computing, in generally.
However, most of the commercial software do not allow a visualization of a higher
order approximation. In this case, some additional technique has to be used, see
Section 1.6.3.
Free graphical libraries represented, e.g., by PGPLOT, which is a set of Fortran
77 subroutines which are able to draw a given line, to fill a given polygon by a
give color, etc. Therefore, a use of this library requires a creation of an additional
code calling the PGPLOT subroutines. This represents an additional work but
you can draw exactly what you want. I.e., no postprocessing perturbs the results.
Figure 1.10, right was obtained by PGPLOT.
Basic graphical code as, e.g., gnuplot are able to carried out basis graphical task
as connecting a given set of nodes. This type of software can be used with
46

Figure 1.12: The sub-grids for the visualization on a triangle T for p = 2 (left), p = 3
(center) and p = 4 (right)
success for wall distribution, cuts and isolines. However, a visualization of isolines
on unstructured grids requires their pre-computation by some additional code.
Figures 1.10, left and 1.11 were created by gnuplot.

1.6.3

Visualization of higher order polynomial function

In this section we discuss some possibilities how to visualize a higher order piecewise
polynomial solution uh of an abstract problem. These visualization techniques are in
fact independent, if uh Vh is a continuous or uh Shp is discontinuous.
In order to visualize the any piecewise polynomial function, we need to evaluate its
value in given nodes from . If we use the data structure introduced in Section 1.5.4,
we can simply evaluate the solution in integrations nodes. However, it is not usually
enough since the Gauss and the Dunavant quadrature nodes do not lay at end points
of (0, 1) or at the boundary of the reference triangle, respectively. Therefore, for the
purposes of the visualization it is suitable to evaluate the solution at the Lagrangian
nodes introduced in Section 1.3.1. This means, the we have to evaluate and store also
on Tb in the nodes used for a visualization.
(D2a) the shape functions i , i = 1, . . . N

All graphical codes are able to draw a line between two nodes. Therefore, it is
possible to visualize without problems a piecewise linear approximate solution. A
visualization of higher order approximation is more complicated. It is possible to
calculate exactly the corresponding solution dot per dot, but it is usually very time
consuming.
More efficient and sufficiently accurate approach is the following. Let T Th be
an element and uh |T a polynomial function of degree p. Then we split T at several
p , see Figure
sub-elements by connecting the Lagrangian nodes FT (zpi ), i = 1, . . . , N
p
p and
1.12. We evaluate uh |T at each of the Lagrangian nodes FT (zi ), i = 1, . . . , N
carried out a piecewise linear visualization over this sub-grid.
Figure 1.13 shows the exact solution of a given problem which is visualized by the
isolines and the 3D plot. This problem was solved on a mesh consisting of four triangles
47

0.16
0.14
0.12
0.1
0.08
0.06
0.04
0.02
0
-0.02

0.8

0.6

0.4

10.8
0.60.4
0.2

0.2

0 0

0.2

0.4

0.6

0.8

0
0

0.2

0.4

0.6

0.8

Figure 1.13: The visualization of the given function, isolines (left) and 3D plot (right)
1

0.3
0.25
0.2
0.15
0.1
0.05
0
-0.05

0.8

0.6

0.4

10.8
0.60.4
0.2

0.2

0 0

0.2

0.4

0.6

0.8

0
0

0.2

0.4

0.6

0.8

Figure 1.14: Visualization of the P1 approximation, isolines (left) and 3D plot (right)
(arising by splitting the square by its diagonals) using Pk , k = 1, . . . , 7 approximations.
Figures 1.14 1.20 show these approximate solutions using the visualization on the
corresponding subgrids.

48

0.2
0.15
0.1
0.05
0
-0.05
-0.1

0.8

0.6

0.4

10.8
0.60.4
0.2

0.2

0 0

0.2

0.4

0.6

0.8

0
0

0.2

0.4

0.6

0.8

Figure 1.15: Visualization of the P2 approximation, isolines (left) and 3D plot (right)

0.16
0.14
0.12
0.1
0.08
0.06
0.04
0.02
0
-0.02
-0.04

0.8

0.6

0.4

10.8
0.60.4
0.2

0.2

0 0

0.2

0.4

0.6

0.8

0
0

0.2

0.4

0.6

0.8

Figure 1.16: Visualization of the P3 approximation, isolines (left) and 3D plot (right)

0.16
0.14
0.12
0.1
0.08
0.06
0.04
0.02
0
-0.02

0.8

0.6

0.4

10.8
0.60.4
0.2

0.2

0 0

0.2

0.4

0.6

0.8

0
0

0.2

0.4

0.6

0.8

Figure 1.17: Visualization of the P4 approximation, isolines (left) and 3D plot (right)

49

0.16
0.14
0.12
0.1
0.08
0.06
0.04
0.02
0
-0.02

0.8

0.6

0.4

10.8
0.60.4
0.2

0.2

0 0

0.2

0.4

0.6

0.8

0
0

0.2

0.4

0.6

0.8

Figure 1.18: Visualization of the P5 approximation, isolines (left) and 3D plot (right)

0.16
0.14
0.12
0.1
0.08
0.06
0.04
0.02
0
-0.02

0.8

0.6

0.4

10.8
0.60.4
0.2

0.2

0 0

0.2

0.4

0.6

0.8

0
0

0.2

0.4

0.6

0.8

Figure 1.19: Visualization of the P6 approximation, isolines (left) and 3D plot (right)

0.16
0.14
0.12
0.1
0.08
0.06
0.04
0.02
0
-0.02

0.8

0.6

0.4

10.8
0.60.4
0.2

0.2

0 0

0.2

0.4

0.6

0.8

0
0

0.2

0.4

0.6

0.8

Figure 1.20: Visualization of the P7 approximation, isolines (left) and 3D plot (right)

50

Bibliography
[1] L. Dubcova. hp-FEM for coupled problems in fluid mechanics. PhD thesis, Charles
University Prague, 2010.
[2] D. A. Dunavant. High degree efficient symmetrical gaussian quadrature rules for
the triangle. Int. J. Numer. Methods Eng., 21:11291148, 1985.
n, K. Segeth, and I. Dolezel. Higher-Order Finite Element Methods. Chapman
[3] P. Sol
& Hall/CRC Press, 2003.

51