A New Family of Mixed Finite Elements In) R 3: Numerische Mathematik
A New Family of Mixed Finite Elements In) R 3: Numerische Mathematik
A New Family of Mixed Finite Elements In) R 3: Numerische Mathematik
J.C. N6d~lec
Math~matiqucs Applique6cs, Ecole Polytechniquc, F-91120 Plaiseau, France
1. Introduction
The mixed finite element was introduced first in several papers of Fraeijs De
Veubeke. In 1977, Raviart and Thomas [8] used these elements for solving
second order equations in two dimension. In 1980 [6], we introduced two
families of mixed finite elements in three dimension. The first family generalizes
that of P.A. Raviart-J.M. T h o m a s and is conforming in the space H(div). The
second one appears to be completely new and is conforming in the space
H(curl).
In 1982, in the reference [7], we used the new finite elements to introduce
an approximation of the Stokes equations that generalizes to the 3D case, the
(~, co) approximation.
in 1984, Brezzi et al. [2] introduced a new mixed finite element conforming
in H(div) in two dimensions. This last paper is the starting point of on search
for new families of mixed finite elements in three dimension. We introduce here
two families of such finite elements. The first one is conforming in the space
H(div) and in fact this family is split in three corresponding to the case of
tetrahedrons, cubes and prisms. The second family is conforming in the space
H(curl) and is also split into three parts. We describe these elements, prove the
unisolvence and estimate the interpolation error. When writing this paper, we
learned that Brezzi, Douglas, Duran and Fortin obtained similar results for
some finite elements in H(div). They obtained results for the case of tetrahedra
and cubes. Apparently their degrees of freedom are different from ours.
In the last chapter, we describe the way our elements can be used to
approximate the Stokes' system.
58 J.C. Nhd61ec
Notations
We say that a finite element is conforming in some functional space V if, when
f2 is the union of two elements K 1 and K 2 with a c o m m o n face f, the function
defined by nzP on K1 and nzP on K 2, belongs to the functional space V(f~).
It follows from L e m m a 1 that a finite element is conforming in H(div) if
and only if, the nullity of the degrees of freedom depending only on the face f
implies the nullity of the normal component of p on this face.
A New Family of Mixed Finite Elements in 3 59
In the previous work [6], we have introduced some spaces of polynomials and
some associated finite elements on tetrahedra.
We introduce the spaces ~k and ~k
~ k = ( P k - 1 ) 3 1 ~ 4 - 1 "r (2.1)
Proof We first have to check that the total number of degrees of freedom is
the dimension of the space P. The dimension of P is
(k + 1)(k + 2)(k + 3)
dim(P) = 2 (2.10)
60 J.C. N6delec
The number of degrees of freedom of type (2.7) is 4 dim P~ that is 2(k + 1)(k + 2).
The number of degrees of freedom of type (2.8) is the dimension of ~ k - a that
(k - 1)(k + 1)(k + 2)
is . To prove the unisolvence now, it suffices to show that p
2
is zero when all the degrees of freedom are zero.
The degrees of type (2.7) imply
we conclude that
~divpqgdx=O; Vq)EPk_ 1 ;
K
which implies
divp---0. (2.13)
S(p'q)dx=O; Vqe(Pk_1)3;
r
The property of unisolvence of the finite element of Definition 5 in reference
[6] and (2.11) associated to the last equality implies p = 0.
It remains to prove (2.9). Using the Green formula (2.12) for p and rcp
yields:
S div(p-~zp)q~dx= - ~ (p-zrp.grad~o)dx + ~ (p-ztp).nq~dy. (2.15)
K K OK
This relation is true for every r in the space Pk-1 and this property is exactly
the identity (2.9). []
Proposition 1. Let rr be the interpolation operator associated to the finite element
of Definition 1. Let h be the diameter of the tetrahedron K and p the diameter of
the inscribed sphere, then we have:
IIP - l t p l{(L~(K))~< Ch k + 1]pl(H,,+'(K))~; (2.16)
hk+ 1
IlD(p-ltp)ll(L2(r))~<c--[pl(Hu+l(r))3 [] (2.17)
P
A New Family of Mixed Finite Elements in 1~ 3 61
1
~(p. v)qd7 (2.18)
mes(f) I
1
V= ~ a l a 2 A a l a 3.
Ue, L D
h
-<c. (2.21)
P
where c is a constant independant of h.
We introduce the following space, which "approximates" the space H(div):
( k + 1)(k+2) (k-1)(k+l)(k+2)
dim(Vh) = 2 Ny -~ 2 NT (2.23)
1 1
(DF(~)~. (DF~ ^ DF~2) ) det(DF) (c~o F - 1)d 7. (2.24)
mes(f) /
Theorem 2. 7he finite element given by the Definition 2 is unisolvent and confor-
ming in the space H(div).
Proof We have to check first that the dimension of P is the total n u m b e r of
degrees of freedom.
The n u m b e r of degrees of freedom of type (2.26) is six times the dimension
of Qk(f) that is: 6 ( k + 1) 2.
The n u m b e r of degrees of freedom of type (2.27) is: 3 ( k - 1)(k+ l) 2.
The dimension of P is:
dim(Qk) 3 = 3(k + 1) 3
( l + 1)(l+2)
dim(P~,m) = (m + 1)
2
64 J.C. N6d61ec
~t,,, (resp. ~/,m) will denote the spaces of pairs of polynomials which are in ~t
(resp. ~l) for x 3 fixed, and are of degree m in x 3.
We can defined the new finite element:
Definition 3. 1~ K is a prism whose base is a triangle in the (x~, x2) plane, with
three vertical edges parallel to the x 3 axis.
2~ The space P of polynomials is
P =(Pk,k)3
3~) The degrees of freedom are of four types:
The number of degrees of freedom of type (2.30) is three time the dimension of
Q,: 3 ( k + l ) 2. The number of degrees of freedom of type (2.31) is:
(k +2)(k + 1)(k - 1)
9 The dimension of ~k-l,k is the dimension of ~ k - 1 in 2D
2
multiply by the dimension of Pk in 1D that is: (k+ 1)2(k-I).
The result on equality is just an easy verification.
To prove anisolvence, it is sufficient to prove that, when all the degrees of
freedom are zero, the vector p is also zero.
We use first the degrees of type (2.29) which implies:
P3=0, (2.34)
which is zero for qo in the space Pk-1,k because then the first two components
of gradqo are in ~k-l,k.
Thus, as divp is also in this space, we have proved
divp=0. (2.37)
We know that in 2D, (Pk_ ~)2 is the direct sum of ~k ~ and the grad (Pk)- As a
consequence, we obtain that (Pk- ~,k)2 is the direct sum of ~ k - a,k and the space
grad(Pk,k), for the first two components.
Using this property, (2.38) and the degrees of type (2.32) we obtain
Now (2.35) and (2.39) constitute the degrees of freedom of the H(div) finite
element in 2D from reference [6] for the degree k + l so that they imply
p-0. []
We can prove, for this finite element, some error estimates similar to those
contained in Proposition 1. We can also associate a curved finite element. This
is particularly usefull in the case k = 1 using distorded prisms.
Remark 3. Using the Green' formula (2.36), we obtain:
Proof. We first check that the number of degrees of freedom is exactly the
dimension of the space P.
p. n = 0. (2.45)
P3 - 0. (2.46)
We then use the formula (2.36) and the degrees of type (2.42) to find:
Pl = ~ x 2 .
c3q~ ' qo~Pk.k- 1 (2.48)
P2 = ~ X 1
and from (2.45) on the vertical faces, we have ~0 = 0 ; on the three vertical faces;
so that
q0 = 2 1 2 2 2 3 ~ , ~ P k - 3,k- 1- (2.49)
The degrees of freedom are such that the right handside is zero when ~p is in
the space Pk- 1,k- 1. This p r o p e r t y is exactly the expression of (2.44).
where a is any of the six edges of the tetrahedron and z is the unit tangent
vector along the edge a;
~(p.q)d?; Vqe~a_l(f) (3.2)
f
and tangent to the face f for each of the four faces of the tetrahedron; ~k- 1 is
defined by (2.1);
~(p.q)dx; Vqe~k_2. (3.3)
K
The number of degrees of type (3.1) is six times the dimension of Pk that is:
6(k + 1).
The number of degrees of type (3.2) is four times the dimension of @k-l(f)
that is:
4(k + 1)(k - 1).
The number of degrees of freedom of type (3.3) is the dimension of ~a-2 that
is:
(k + 1)(k - 1)(k - 2 )
2
We can now just check that the total number of degrees is dimP.
To prove the conformity in the space H(curl), we use the degrees of type
(3.1) and (3.2) relative to one face.
The degrees of type (3.1) imply
We have the Stokes' formula on this face (n is the normal to the face):
The relation (3.5) show that ~ is constant along Of and so can be chosen to be
zero of 0 f We can then write ~ as a product of the barycentric coordinates on
Of by a polynomial of degree k - 2 :
~/---=).12223 (p ; cpePk_ 2 ( f ) . (3.9)
pAn=0; on f. (3.11)
Using next the property that div(N k_ 2) is Pk-3, we obtain that q~ is zero an the
proof is complete. []
From Theorem 4, we know that there is an interpolation operator denoted
by re, associated to the finite element of Definition 4.
We have some links between this finite element and the finite element of
Definition 1 for the degree k - 1 . Let re* be the interpolation operator associate
to the latter element. We have
70 J.C. N6d61ec
~(curl(p-np)'q)dT=O; Vqe(Pk_2) 3.
K
They can be replaced by two values of (p. ~) taken at two points of the edge a.
A New Family of Mixed Finite Elements in ]R3 71
p. z =0, (3.23)
Pl =X2(1 --X2)q~l ;
(3.24)
P2 =Xl(1 - x 0~o2.
72 J.C. N6d61ec
Then the degrees of type (3.21) implies now that q)l and ~o2 are zero. This
proves the conformity of the element. We now have
Proof Each face of this finite element is either a triangle or a rectangle. The
degrees on the triangular faces are exactly those of the corresponding tetrahe-
dron element and the degrees on the rectangular faces are exactly those of the
corresponding cubic element. It result from Theorem 4 and 5 that the nullity of
these degrees implies
p ^ n =0, on these faces. (3.31)
Pl =X3(1 --x3)q)l ;
(3.32)
P2 =X3(1 --Xl)q?2 ;
Using the relation (3.33) and the degrees of type (3.29), we obtain:
P3 =0. (3.34)
Pl -c?xl.
(3.38)
g~o '
P2 -63x2
Using (3.31), we can choose q~ so that
Using the 2D result that div(~ k_ ~) is Pk-2, we can conclude that O = 0 and that
completes the proof. []
One might think that a property similar to that of Proposition 2 is true for
this finite element and that of Definition 3. This is not the case and similarly it
is also not true for the finite elements of Definition 2 and Definition 6.
The finite element of Definition 7 is in the spirit of those of Definition 5
and 6. We introduce here a different one that belongs to the family of the
reference [6].
Definition 8
1~ K is the prism of Definition 3
2~ The space P of polynomials is
! )
~3Pl qadx=O;
~3xz Vq3~Pk_ 1,k_2- (3.49)
\~X 1
Pl-dxl.
d, Ca~,k- (3.51)
Pz-dx 2
76 J,C. N6d61ec
Proof. For every face of the prism K, we use the identity (3.14). This identity
yields for the two horizontal faces
to obtain
(curl(p-top). q)dx =0; (3.56)
K
for every q =(ql, q2, q3), (q l, q2)e(Pkk- 2,k- 1)2, q3 ePk- l,k-- 2"
This last identity finishes the proof of the proposition.
-vAu+gradp=f, in f2,
divu =0, in f2, (4.1)
ULr=O; on 0~.
We introduce
w =curlu (4.1)
Then the system (4.2) has a anique solution when the domain f2 is connected
and simply connected, and we have
Then a variational formulation of the equation linking 05 and ~o, given by (4.1)
and (4.3), is:
v~(curloo.curl~b)dx=~(f.curlt~)dx; VqJEJg;
o (4.4)
~(co.zOdx-~(cur105.curln)dx=O; VTz~H(curl).
f2 ~2
~(o).~)dx-~(u.curlTt)dx=O; VneH(curl).
12 12
In order to approximate the solution of the Equation (4.4) using the finite
element introduced in this work, we suppose that we have a mesh ~, covering
the domain f2. We will suppose here that we are using tetrahedrons.
We introduce some finite element spaces:
Oe={Oe~Hl(f2);Oelr~Pk+,;VKeC~e}; O~
h
-< C for every K in (~h'
P
and if the diameters h are uniformly equivalent.
We have the
Theorem 7. When the triangulation is regular, for every element vh in the space
U* (respectively Uh), there exist a unique element wh in the space Wh (respectively
Who) such that:
curl wh =Vh; (4.7)
Moreover, we have
IlWhHmcum< C I[Vhll(L2(a))3. (4.9)
Proof We first prove the uniqueness. We just have to show that the system of
equations
curlWh=0; WhEWh(res p. 1410);
(4.10)
~ (wh.gradOh)dx=O; VOhE6)h(resp" 6)O);
fk
ahlr~Pk+ 1. (4.12)
This property implies that eh is in the space Oh(res p. 6)~ So that we can
choose Oh as % in (4.10) which yields w h =0.
We have proven that the operator curl is injective from the subspace of
vectors in Wh(resp. Wh~ satisfying (4.8) into the space U*(resp. Un).
In order to prove the surjectivity, it is sufficient to prove that these two
spaces are of same dimension. The dimension of Wh is
(k + 1)(k - 1)(k - 2 )
dimWh=(k+l)N"+(k+l)(k-1)Nr 2 Nr. (4.13)
A New Family of Mixed Finite Elements in ~3 79
k(k-1)
dimO.=N~ + kNo+ - - 2 ~ U~~ k(k-1)(k-2)N r. (4.14)
6
The dimension of the subspace of Wh is
M k = d i m Wh - d i m Oh + 1
(k-1)(k+2) (k-1)(k-2)(2k+2)
= -Ns+N"-t- 2 Nf~- 6 NT+I
k(k+ 1) k(k+ 1 ) ( 2 k - 5 )
dim Uh* - 2 N, -~ 6 Nr" (4.15)
In the case of the spaces W~ and Uh, we must modify the above numbers, n s is
the n u m b e r of vertex of g, on 0Q; n a is the n u m b e r of edges of cgh on &2; n s is
the n u m b e r of faces of Zh on 0(J.
We have
dim Wh~ = d i m Wn - ( k + 1)n, - ( k + 1)(k - 1)hi; (4,17)
k(k-1)
dimOh~ = dim Oh --ns -kna 2 nf. (4.18)
(k - t)(k + 2)
M~176176 2 nf-1 (4.19)
So that we obtain
M ~ - d i m Uh = M k - d i m U~ + n ~ - n . + n I - 2 . (4.22)
The proof of inequality (4.9) is very similar to the same proof in reference [7]
and will not be given here. []
F r o m the Theorem 7, we can derive some error estimates for the approxi-
mation of the Stokes' equation (4.5).
Proposition 3. L e t u, w be the s o l u t i o n o f (4.5) a n d (uh, wh) t h e s o l u t i o n o f (4.6),
we have
11u - u h II~L2~))~ + IIW -- W h II~L2~))3 < C h k - 1 (4.24)
curl w = v h;
div w = 0 ; (4.25)
w /x n l r = O .
The domain t2 being connex, connected and simply connected, there exist a
unique solution of this equation in the space (HZ()) 3. Now, the interpolate
nhW is such that (by the Proposition 2)
W =TZh w + g r a d ~ h ; OChE O h
References
1. Brezzi, F.: On the existence, uniqueness and approximation of saddle point problems arising
from Lagrangian multipliers. RAIRO 8, 129-151 (1974)
2. Brezzi, F., Douglas, J., Marini, LD.: Two families of mixed finite elements for second order
elliptic problems. (To appear in Numer. Math.)
A New Family of Mixed Finite Elements in 1t 3 81
3. Ciarlet, P.G.: The finite element method for elliptic problems. Amsterdam: North Holland 1978
4. Ciarlet, P.G., Raviart, P.A.: A mixed finite element method for the biharmonic equation.
Math6matical aspects in finite element method (C. de Boor ed.), pp. 125-145. New York:
Academic Press 1974
5. Fortin, M.: An analysis of the convergence of mixed finite element method. RAIRO 11,341-354
(1977)
6. N6d61ec, J.C.: Mixed finite element in 11.3. Numer. Math. 35, 315-341 (1980)
7. N6d61ec, J.C.: Elements finis mixtes incompressibles pour l'equation de Stokes darts 113 Numer.
Math. 39, 97-~ 12 (1982}
8. Raviart, P.A., Thomas, J.M.: A mixed finite element method for 2nd order elliptic problems. In:
Mathematical aspects of finite element methods (A. Dold and B. Eckmann, eds.) Lect. Notes
606. Berlin, Heidelberg, New York: Springer 1977
9. Thomas, J.M.: Thesis Paris (1977)