Finite element characteristic methods

requiring no quadrature
Olivier Pironneau
LJLL, Universite Pierre et Marie Curie, Place Jussieu, Paris F-75005
E-mail: pironneau@ann.jussieu.fr
in honor of M. Tabata for his sixtieth birthday
Abstract
The characteristic methods are known to be very ecient for
convection-diusion problems including the Navier-Stokes equations.
Convergence is established when the integrals are evaluated exactly,
otherwise there are even cases where divergence has been shown to
happen. The family of methods studied here applies Lagrangian con-
vection to the gradients and the function as in Yabe[?]; the method
does not require an explicit knowledge of the equation of the gra-
dients and can be applied whenever the gradients of the convection
velocity are known numerically. We show that converge can be sec-
ond order in space or more. Applications are given for the rotating
bell problem.
Keywords Finite Element Method, Characteristic-Galerkin, convection-
diusion equation.
1 Introduction
Let u be solution of the convection-diusion-dissipation equation in a bounded
open set over a time interval (0, T)

t
u +a u (u) +bu = f, u(0) = u
0
in , t < T (1)
with a smooth positive diusion , smooth convection velocity a, dissipa-
tion coecient b and source f. For simplicity we assume a, b independ of t,
constant, and with homogenous Neumann boundary conditions,
n
u = 0
on (0, T).
1
To obtain a second order method in time when f = 0 one may use Strangs
operator splitting as in Beale-Majda[1] and compute
u
m+1
= D
t/2
oC
t
oD
t/2
u
m
where D

u
0
is the solutions at time of

t
u (u) +bu = 0, u(0) = u
0
, (2)
and C

u
0
is the solution of

t
u +a u = 0, u(0) = u
0
(3)
As this last equation says also that the total derivative of u((t), t) along
the characteristic curve (t) is zero, a discretization is u
m+1
(x) = u
m
(X(x))
where X(x) is (t
m
) when (t
m+1
) = x.
Alternatively one may use
1
t
(u
m+1
u
m+1
) (u
m+1
) +bu
m+1
= 0, u(0) = u
0
,
with u
m+1
(x) = u
m
(X(x)) where X(x) = x ta(x
t
2
a(x)) (4)
The scheme is O(t
2
) + O(Ct) but C tends to zero with + b. In a
large class of applications and b are small and so in practice it might be
enough. Still, the diculty then is to nd a second order nite element
approximation of this scheme, stable even if = 0.
Many research papers deal with this problem (see [5] and the references
therein and more recently [3] [2] [6]) . The usual approach is to write the
scheme in weak form and discretise by the nite element method. However
all proofs require exact quadrature for the term containing u
m+1
which
in turn implies that the nite element mesh and the convected mesh are
intersected. Since this is too costly, we propose here an alternative based
on an idea due to Yabe[?] which has been shown very accurate numerically
but which has not yet been proven to work theoretically.
Consider (3) in (0, T) with a, u
0
smooth a = 0 and a n = 0 on the
boundary. The semi-analytical solution of this problem can be written in
terms of the Lagrangian coordinate (t), solution of
(t) = a((t)), (t
0
) = x
0
(5)
where t
0
and x
0
are arbitrary time and point. The solution is
u((t), t) = u(x
0
, t
0
),
a formula which is based on the observation that d
t
u((t), t) = 0. Thus
with a nite dierence of the time derivatives and by choosing t
0
= t
m+1
and x
0
= x we build the following second order scheme:
u
m+1
(x) = u
m
(X(x)) with X(x) = x a(x a(x)
t
2
)t (6)
We recall the denition of an m
th
order scheme:
2
Denition If u(, ) denotes the exact solution of the partial dierential
equation and if u
m
() denotes the numerical solution at time t
m
by some
numerical scheme, if there exists C such that
|u
m
u(, t
m
)|

Ct
p
then the scheme is said to O(t
p
), or p-order accurate.
Lemma 1.1 Scheme (6) is second order accurate
Proof
To check the consistency one replaces u
m
in (6) by u(x, t
m
), the exact
solution at time t
m
and performs a Taylor expansion of the right hand
side. This shows that the scheme is second order consistent; since stability
is obvious from (6), by the von Neumann lemma the scheme is second order
accurate.
1.1 Standard nite element discretization
The simplest discretization in space is to appy the scheme at the center of
gravity q
k
c
of each triangle T
k
:
u
m+1
h
(x) = u
m
h
(X(q
k
c
)) x T
k
k (7)
Proposition 1.2 Scheme (7) is O(
h
t
+ t
2
) and best at t = h
1/3
in the
sense that for some constant C independent of h and t
|u
m
h
u(, t
m
)|

|u
m
h
u
m
|

+ |u
m
u(, t
m
)| C(
h
t
+t
2
) (8)
Proof
Applying a Taylor expansion to u
m+1
gives
[u
m+1
h
(x) u
m+1
(x)[ = [u
m
h
(X(q
k
c
)) u
m
(X(q
k
c
)) + u
m+1
()(x q
k
c
)[
|u
m
h
u
m
|

+ |u|

h |u
0
h
u
0
|

+ |u|

hT
t
(9)
Hence (8) holds
The same argument works for a P
1
approximation and convection at
the vertices
Proposition 1.3 Let
i

3
1
be the barycentric coordinates of x T
k
(see
(22)). The scheme
u
m+1
h
(x) =

i=1,2,3

i
u
m
h
(X(q
i
)) x T
k
k (10)
3
is O(
h
2
t
+t
2
) and best at t = h
2/3
The proof is also given in Lemma 4.1. The fundamental remark for this
paper is that the argument does not work for P
2
element, however by
applying it to the gradients we gain on order of convergence in space.
2 Convection of the gradients
To obtain a second order scheme in space for (3) we intend to convect the
function u and its gradients v(x, t) =
x
u(x, t). For clarity we consider
only the case b = f = 0. Notice that v satises a vector equation like (30)
with = 0 and b = a:

t
v +av + (a)v = 0, v(0) = u
0
(11)
So in principle a characteristic nite element scheme based on an analytic
formula for (11) could be used; however our numerical tests indicate that
it is better to dierentiate (6) rather than to use (11). This leads to:
v
m+1
(x) = X(x)v
m
(X(x)) (12)
Lemma 2.1 The solution of (11) satises
v(x, t
m+1
) := u(x, t
m+1
) = X(x)v(X(x), t
m
) +O(t
3
) (13)
so Scheme (12) is second order accurate because
|v
m+1
v(, t
m+1
)|

(1 +ct)|v
m
v(, t
m
)|

+Ct
3
(14)
Proof
v(X(x), t
m
) = [v ta(x a
t
2
)v +
t
2
2
v

aa][
x,t
m +O(t
3
)
and
X(x) = I a(x)t +
t
2
2
((a

)a + (a)
2
)(x) +O(t
3
) (15)
Therefore placing the exact solution into the scheme gives
v(x, t
m+1
) = (I at +
t
2
2
((a

)a + (a)
2
))

v tav +
t
2
2
aav +
t
2
2
v

aa

[
x,t
m +O(t
3
)
= v(x, t
m
) t(av +av)[
x,t
m
4
+
t
2
2
(a

av + (a)
2
v + 2aav
+aav +v

aa)[
x,t
m +O(t
3
)
= [v +t
t
v +
t
2
2

tt
v][
x,t
m +O(t
3
)
which is correct because by (3) for v:

tt
v = a(av) +v

aa +aa

v + a(av) + (a)
2
v
Finally (14) is found by subtracting (13) from (12).
3 Discretisation with P1 discontinuous ele-
ments
To solve numerically (3) in (0, T) we use u
m
, X dened in (6). As
mentionned earlier this time scheme is O(t
2
) .
Scheme (6) is discretized in space by a discontinuous piecewise linear
approximation on a triangulation T :
u
m+1
h
(x) = X(q
c
)u
m
h
(X(q
c
))(x q
c
) +u
m
h
(X(q
c
)) (16)
for all x T the triangles of the triangulation; q
c
denotes the center of
gravity of T. Notice that this is equivalent to saying that u
m+1
is recon-
structed from its gradient and its value at q
c
and that these are obtained
by the method of characteristics from the values at the previous time step.
Proposition 3.1 When h Ct, the method is L

-stable in the sense

that |u
m
h
|

C
h
t
+ |u
0
h
|

.
Proof
By construction u
m+1
h
(x) = X(q
c
)u
m
h
(X(q
c
)) for all x T with center
q
c
; hence by (15)
|u
m+1
h
|

(1 +ct)|u
m
h
|

(1 +ct)
T
t
|u
0
h
|

.
Similarly [u
m+1
h
(x)[ |u
m
h
|

[x q
c
[ + |u
m
h
|

so
[u
m+1
h
[

h(1 +ct)
T
t
|u
0
h
|

+ |u
m
h
|

which proves the proposition.

5
Theorem 3.2 The method is O(h) in the sense that
|u
m
h
u(, t
m
)|

C(
h
2
t
2
+t
2
)
hence optimal at t = O(h
1
2
).
Proof
Let v
m
= u
m
and v
m
h
= u
m
h
; then for some between x and q
c
[v
m+1
h
v
m+1
[(x) = [v
m+1
h
(q
c
) v
m+1
(q
c
) u

m+1
()(x q
c
)[
= [X(q
c
)(v
m
h
v
m
)(X(q
c
)) +O(|u

h)
(1 +ct)|v
m
h
v
m
|

+C|u

h (17)
Hence
|u
m
h
u
m
|

< C(h +
h
t
) (18)
Similarly, using the following Taylor expansion (which relies on the fact
that x and q
c
are in the same triangle)
u
m+1
(x) = u
m+1
(q
c
) + u
m+1
(q
c
)(x q
c
)
+
1
2
u
m+1

)(x q
c
)(x q
c
) (19)
leads to
[ u
m+1
h
u
m+1
[(x) = [u
m+1
h
(q
c
) u
m+1
(q
c
)
+(u
m+1
h
(q
c
) u
m+1
(q
c
))(x q
c
) +
1
2
[u

m+1
(

)(x q
c
)(x q
c
)[
|u
m
h
u
m
|

+ |u
m+1
h
u
m+1
|

h + |u

h
2
2
C
h
2
t
2
(20)
4 Discretisation with P2 discontinuous ele-
ments
Convection of the gradients The scheme (12) for v is discretized in
space by continuous nite elements of degree one on triangles:
v
m+1
h
(x) =

i=1,2,3

i
X(q
i
)v
m
h
(X(q
i
)) (21)
where q
i
are the vertices of the triangulation and
i

i=1,2,3
the barycen-
tric coordinates of x, i.e.
x =

i=1,2,3

i
q
i

i=1,2,3

i
= 1 (22)
6
Lemma 4.1 Let v
m
h

m
be given by scheme (21) and let u
m

m
be dened
by the time scheme (6), for (11). Then
|v
m+1
h
u
m+1
|

(1 +ct)(|v
m
h
u
m
|

+Ch
2
Proof
Dene the space error = v
m
h
v
m
, then

m+1
(x) = v
m+1
h
(x) v
m+1
(x) =

i=1,2,3

i
w
m
h
(q
i
) w
m
(x)
where w
m
(x) = X(x)v
m
(X(x)), w
m
h
(q
i
) = X(q
i
)v
m
h
(X(q
i
)).
Again using the fact that x, q
i
are in the same triangle, a Taylor expan-
sion gives
w
m
(q
i
) = w
m
(x) + (q
i
x) w
m
(x) +
1
2
(q
i
x)
T
w

m
(
i
)(q
i
x)
for some
i
. Hence

i=1,2,3

i
w
m
(q
i
) = w
m
(x) +

i=1,2,3

i
2
(q
i
x)
T
w

m
(
i
)(q
i
x)
Therefore, we have
|
m+1
|

max
k
[

i=1,2,3

i
(w
m
h
(q
i
) w
m
(q
i
))[
T
k
+sup
x,k
[

i=1,2,3

i
2
(q
i
x)
T
w

m
(
i
(x))(q
i
x)[
T
k
max
k
max
i,q
i
T
k
[X(q
i
)(v
m
(X(q
i
)) v
m
h
(X(q
i
)))[
T
k
+sup
x,k
[

i=1,2,3

i
2
(q
i
x)
T
w

m
(
i
(x))(q
i
x)[
T
k
(1 +ct)|
m
|

+C|v

h
2
2
(23)
where c and C are functions of |a|

and of |a

.
Remark 1 Note that we use explicitly the fact that the hat function
i
is
positive, and this is why this proof does not work for nite elements with
non-positive hat functions such as the P
2
element.
7
Convection of the function The approximate gradients v
m
h
, can be
computed on (0, T) independently of u but we need to be able to construct
at any time t
m
= mt an approximate value, u
m
h
of u(, t
m
), with the best
accuracy.
On each triangle the function u
h
which approximates the solution of (3)
is reconstructed from its approximate gradient v
m
h
and a value at one point,
say
k
, for instance the center of gravity, of each triangle T
k
. Hence at each
time step u
m
h
is a second order polynomial on each triangle, discontinuous
at the edges.
More precisely, v
m
h
is computed by (21) and being linear in x there
exists A R
22
such that, for all x T
k
v
m
h
(x) = A(x
k
) +v
m
h
(
k
) (24)
Then we compute u
m
h
by
u
m
h
=
1
4
(x
k
)
T
(A
m
k
+(A
m
k
)
T
)(x
k
) +v
m
h
(
k
)(x
k
)
+u
m1
h
(X(
k
)) (25)
Note that u
m
h
is never used and that v
m
h
,= u
m
h
.
Lemma 4.2 Equation (26) written at the 3 vertices has one and only one
solution A.
Proof
Let us choose a coordinate system where q
1
= (h
1
, h
2
)
T
, q
2
= (0, h
3
)
T
,
(q
3
= (0, 0)
T
; let b
i
j
= v
m
h
(q
i
)
j
v
m
h
(q
3
)
j
. Then the linear system for A is
a
11
h
1
+a
12
h
2
= b
1
1
a
21
h
1
+a
22
h
2
= b
1
2
a
12
h
3
= b
2
1
a
22
h
3
= b
2
2
(26)
The solution exists and is unique for all non degenerate triangles.
Lemma 4.3
|u
m
h
u
m
|

C
h
2
t
Proof
|u
m
h
u
m
| |u
m
h
v
m
h
| + |v
m
h
u
m
| (27)
Lemma 4.1 gives a bound for the last term on the right, so let us focus on
the rst term on the right. By construction
v
m
h
(x) = A
m
k
(x
k
) +v
m
h
(
k
)
8
=
1
2
(A
m
k
+ (A
m
k
)
T
)(x
k
) +v
m
h
(
k
) +
1
2
(A
m
k
(A
m
k
)
T
)(x
k
)
= u
m
h
(x) +
1
2
(A
m
k
(A
m
k
)
T
)(x
k
) (28)
Let u
m
h
be the P
2
interpolate of u
m
. Then [u
m
u
m
h
[ ch
2
and
u
m
h
(x) = B(x
k
) + u
m
h
(
k
)
for some 2 by 2 symmetric matrix B. Therefore
1
2
(A
m
k
(A
m
k
)
T
)(x
k
) =
1
2
((A
m
k
B) (A
m
k
B)
T
)(x
k
)
This is an expression which involves only the values of w(x) := v
m
h
(x)
u
m
h
(x) at the vertices.
By inspection of (26), there exists a constant C such that [A
m
k
B[
C|w|

/h, which, by (28) and with C denoting

C
2
max
|x
k
|
h
implies that
|v
m
h
u
m
h
|

C|v
m
h
u
m
h
|

Finally, replace u
m
h
by u
m
at the cost of a ch
2
to obtain in (27)
|u
m
h
u
m
|

(1 +C)|v
m
h
u
m
|

+Cch
2
which, by Lemma 4.1, is less than C
h
2
t
.
Theorem 4.4 The absolute error between the exact solution of (3) and
the one computed by convecting the gradients at the vertices and function
at the center of each triangle, i.e. (21,25), is of order 3/2 in the sense that
whenever u and a are twice continuously dierentiable in space and time.
|u
m
h
u(, t
m
)|

C(
h
3
t
2
+t
2
)
which is optimal at t = h
3/4
.
Proof
Let u
m
be the solution of the scheme in time (6). A Taylor expansion shows
that, for some T
k
,
(u
m+1
h
u
m+1
)(x) = (u
m+1
h
u
m+1
)(
k
)
+ (u
m+1
h
u
m+1
)() (x
k
)
= (u
m
h
u
m
)(X(
k
)) + (u
m+1
h
u
m+1
)() (x
k
)(29)
9
The rst term on the right is bounded by |u
m
h
u
m
|

; the last term

is bounded by h|u
m+1
h
u
m+1
|

, which, according to Lemma 4.3 is

O(h
2
/t), therefore
|u
m+1
h
u
m+1
|

(1 +ct)|u
m
h
u
m
|

+C
h
3
t
and the nal result stems from the bound |u
m
u(, t
m
)|

Ct
2
.
Remark 2 When f ,= 0 it suces to add on the right hand side (f(q
i
) +
f(X(q
i
))/2. When b ,= 0 v and u are coupled so it needs some eort to
adapt the proof but it can be done. Notice also that there are elds for im-
provement, for instance by choosing
k
to be the points which convects into
the local max or min of u on each T
k
the method becomes non-oscillatory.
Notice also that it would have made more sense to replace v
m
h
by u
m
h
once
u
m
h
is computed, but the proof of convergence with such change is open.
By convecting u and u and each component of u

at the centers of gravity

and then reconstructing a P
2
polynomial in each triangle, we may be able
to achieve a better error.
5 Discretisation with Q2 elements
When the domain is quadrangulated, the gradients are approximated by
a standard continuous Q
1
vector valued function on the quadrilateral ele-
ment, the function on each quadrilateral is
u
h
[
R
k
= ax
2
y +bxy
2
+cx
2
+dy
2
+fx +gy +e
where x, y is in a reference frame whose origin is at the center
k
of the
rectangle R
k
so that e = u
h
(
k
) and all other parameters are determined by
the 2 gradients at the 4 vertices (there is a compatibility condition so one
data is redundant, namely a=b). Thus, with q
i
= (x
i
, y
i
)
T
, any solution of
(30) will do:

2x
1
y
1
y
2
1
2x
1
0 1 0
x
2
1
2x
1
y
1
0 2y
1
0 1
2x
2
y
2
y
2
2
2x
2
0 1 0
x
2
2
2x
2
y
2
0 2y
2
0 1
2x
3
y
3
y
2
3
2x
3
0 1 0
x
2
3
2x
3
y
3
0 2y
3
0 1
2x
4
y
4
y
2
4
2x
4
0 1 0
x
2
4
2x
4
y
4
0 2y
4
0 1

a
b
c
d
f
g

u
h
(q
1
)
u
h
(q
2
)
u
h
(q
3
)
u
h
(q
4
)

(30)
We conjecture that a similar proof as for the triangular elements may be
contructed to show that the method is O(t
2
+
h
3
t
2
). If such is the case
then the method is optimal at t = h
3/4
.
10
6 The convection-diusion equation
When f = b = 0 the fully discretized scheme is
1
t
(u
m+1
h
u
m+1
h
, w
h
) +(u
m+1
, w
h
) = 0 (31)
for all w
h
V
2
h
, the space of P
2
piecewise quadratic functions. Here to
avoid quadrature on the term (u
m
h
oX, w
h
) (the usual characteristic nite
element method) we replaced u
m
h
oX by u
m+1
h
computed by one time step
of the algorithm of the previous section with u
m
h
as initial condition.
Proposition 6.1 Let t

> h and assume that we could prove that:

| u
m+1
h
u
m
oX
m
| |(u
m
h
u
m
)| +C(h|u
m
h
u
m
| +t
3
+h
3
) (32)
where the norms are L
2
and u
m

M
1
is the solution of
u
m+1
u
m
oX tu
m+1
= 0,
u
m+1
n
[

= 0 (33)
(this corresponds to (23) with L
2
norms instead of L

). Then the scheme

(31) would be O(
h
3
t
+t
2
).
Proof
The scheme (33) is O(t
2
+

t). Let
h
be the projection operator on V
2
h
with the scalar product (, )+t(, ). Then with w
h
= u
m+1
h

h
u
m+1
in (31):
|u
m+1
h

h
u
m+1
|
2

= ( u
m+1
h
u
m
oX, u
m+1
h

h
u
m+1
)
| u
m+1
h
u
m
oX||u
m+1
h

h
u
m+1
|

where | |

| |
2
+t| |
2
. Furthermore
|u
m
h
u
m
| +Ch|u
m
h
u
m
| (1 +
C
2
h
2
t
)
1
2
|u
m
h
u
m
|

Now when all is smooth |u

m+1

h
u
m+1
|

C(h
3
+

th
2
), so the
result follows.
7 Numerical Tests
In all cases here f = 0. The rotating bell test case [5] has a = (x
2
, x
1
)
T
and u
0
= e
r|xx0|
2
. Here the domain is truncated to the unit disk
(the unit square when quadrilateral elements are used) and r = 20, x
0
=
11
(0.35, 0.35)
T
. At t = 2, u = u
0
so the analytical solution is known when
= 0.
To test the case ,= 0 we chose a problem with an analytical solution
u
a
= exp(
r
t+1
[x

[
2
) with x

= x x
0
(t), where x
0
(t) rotates as in the no
diusion case; it corresponds to = 0.08 and a time dependent dissipation
term b = 1/(t + 1); the wrong sign is not a problem because the bilinear
form is still coercive; boundary conditions are needed in this case and we
chose u = u
a
on .
Numerical tests with discontinuous P1 elements Resuts for both
problems, = 0 and ,= 0, are on gure 1. The triangulation has 5300
vertices and 50 time steps are used to cover the full turn. To study the
error decay as a function of h we use the automatic mesh generator of
freefem++ where the number of vertices is controlled by the number M of
boundary points. By taking t = c/

M, for some constant c we achieve

t = O(

h).
Scheme (16) ( = 0) is somewhat disappointing: to observe an error
decay in O(h) we need to take c very small, leading to 300 time steps for
the triangulation with 5300 vertices ; for larger time steps the results are
more like O(

h). In the diusion case with scheme (31) and P

1
elements,
the results are excellent and the decay of the error is O(h
2
).
Numerical tests with discontinuous Q2 elements Figure 2 displays
a perspective view of the exact and computed solution for a mesh of 6320
elements and 80 time steps. The error decay, as a function of h, while
keeping the ratio h/t constant, is also shown on gure 2
Results with discontinuous P2 elements We tried two implementa-
tions to compute v
m+1
h
A/ by convecting v
m
h
as in the theory or B/ by
convecting u
m
h
; the numerical results are almost identical. The errors are
shown in table 1. The optimal time step was search and seems to be O(h)
rather than O(h
3/4
) and 50 is best when there are 150 vertices on the circle,
giving 5300 vertices. For the problem with diusion the method (31) works
very well and seems to be O(h
2
) even when = 0. A zoom of the solution
is shown on the left side of gure 3 and compared with the exact solution.
On the right side the L
2
-error is shown as a function of h on a log-log scale
and compared with the solution of (31).
12
IsoValue
-0.0789317
-0.014729
0.0494737
0.113676
0.177879
0.242082
0.306284
0.370487
0.43469
0.498892
0.563095
0.627298
0.6915
0.755703
0.819906
0.884108
0.948311
1.01251
1.07672
1.14092
IsoValue
0.0270912
0.0769788
0.126866
0.176754
0.226642
0.276529
0.326417
0.376305
0.426192
0.47608
0.525967
0.575855
0.625743
0.67563
0.725518
0.775405
0.825293
0.875181
0.925068
0.974956
Figure 1: Left pure convection: zoom of the solution u
h
computed with P1
discontinuous elements with a 5300 vertices and 50 time steps; there is no
phase error and the exact solution is at the same position as the computed
one; the peak of the discrete solution is at 1.14. Right convection-diusion:
computed solution and exact (the perfect circles) solution; here the L
2
error
is 0.009 after one turn and the peak of the discrete solution is at 0.94.
13
"rese.txt"matrix
"res.txt"matrix
0
10
20
30
40
50
60
70
80
0
10
20
30
40
50
60
70
80
-0.2
0
0.2
0.4
0.6
0.8
1
1.2
0.0001
0.001
0.01
0.1
1
10 100 1000
"errorq2.txt"using 1:2
"errorq2.txt"using 1:3
"errorq2.txt"using 1:4
"errorq2.txt"using 1:2
Figure 2: Top: the solution u
h
computed with Q
2
elements with a uniform
mesh 8080 and 80 time steps . Bottom: The error as a function of h
1
on a log-log scale. The O(h
2
) and O(h) lines are drawn for comparison.
14
0.001
0.01
0.1
10 100
"errorp1.txt"
50/x/x
15/x/sqrt(x)
"errorp1.txt" using 1:3
Figure 3: Top: the solution u
h
of (31) after one turn with a mesh with 3003
vertices and 90 time steps; the analytical solution u
a
is also plotted (the
perfect circles). Notice how each level line of u
h
is close to the corresponding
level line of u
a
The L2 error is 0.006. Bottom: The error as a function of h
1
on a log-log scale. The O(h
2
) and O(h
3
2
) lines are drawn for comparison.
The last curve is the error from the standard P2 method with 6 inner
quadrature points; though the errors are smaller, the error decay is not as
good.
15
Table 1: Errors with P2dc elements by method A and method B.
C/h 1 2 3 4
c/dt 25 50 66 120
A 0.078 0.017 0.0088 0.0055
B 0.079 0.015 0.0088 0.0053
8 Conclusion
By convecting the gradients at each time step we have shown that an
explicit convection strategy converges to order h
n
where n = 1 for P
1
approximation and n =
3
2
for P
2
; this is an improvement over the classic
convection of vertices whereby n =
4
3
at best (Proposition 1.3).
Convergence of the implicit method for the convection diusion case is
still an open problem. The numerical results are excellent though not
outstandingly better than the classic characteristic nite element method.
But in the small velocity cases, where the classic method could diverge here
it seems to converge with this new quadrature idea.
Acknowledgements We wish to thank Frederic Hecht for his implemen-
tation method with freefem++ (www.freefem.org).
References
[1] J. T. Beale, A. Majda, Rates of convergence for viscous splitting of
the Navier-Stokes equations, Math. Comp. 37 (1981), 243-259.
[2] A. Bermudez, M. R. Nogueiras, and C. Vazquez: Numerical solution
of (degenerated) convection-diusion-reaction problems with higher
order characteristics/nite elements. Part II:fully discretized scheme
and quadrature formulas. SIAM J. Numer. Anal. Vol 44, no 5, pp
1854-1876 (2006)
[3] K. Boukir, Y. Maday, B. Metivet, A high order characteristics method
for the incompressible Navier Stokes equations Comp. Methods in
Applied Mathematics and Engineering 116 (1994), 211 218.
[4] F. Hecht, O. Pironneau and A. Leyaric and K. Ohtsuka: freefem++:
the documentation, www.freefem.org (2006).
[5] O. Pironneau. Finite Element Methods for Fluids. Wiley (1989).
16
[6] H. Rui and M. Tabata: A second order characteristic nite element
method. Numer Math, 92, pp 161-177 (2002).
[7] T. Yabe, T. Ishikawa, P.Y. Wang, T. Aoki, Y. Kadota and F. Ikeda: A
universal solver for hyperbolic equations by cubic-polynomial interpo-
lation. II. Two- and three-dimensional solvers. Comput. Phys. Comm.
66 , no. 2-3, pp 233242 (1991).
17