FINITE ELEMENT DIFFERENTIAL FORMS ON CUBICAL MESHES

DOUGLAS N. ARNOLD AND GERARD AWANOU

Abstract. We develop a family of nite element spaces of dierential forms dened on
cubical meshes in any number of dimensions. The family contains elements of all polynomial
degrees and all form degrees. In two dimensions, these include the serendipity nite elements
and the rectangular BDM elements. In three dimensions they include a recent generalization
of the serendipity spaces, and new H(curl) and H(div) nite element spaces. Spaces in the
family can be combined to give nite element subcomplexes of the de Rham complex which
satisfy the basic hypotheses of the nite element exterior calculus, and hence can be used
for stable discretization of a variety of problems. The construction and properties of the
spaces are established in a uniform manner using nite element exterior calculus.
1. Introduction
In this paper we develop a family of nite element spaces o
r

k
(T
h
) of dierential forms,
where T
h
is a mesh of cubes in n 1 dimensions, r 1 is the polynomial degree, and
0 k n is the form degree. Thus, in 3 dimensions, the space o
r

k
(T
n
) is a nite
element subspace of the Hilbert space H
1
, H(curl), H(div), or L
2
, according to whether
k = 0, 1, 2, or 3. For n = 1 or 2, the spaces were previously known, while in three (or
more) dimensions, they are mostly new. Specically, our construction yields a new family of
H(curl) elements and a new family of H(div) elements on cubical meshes in three dimensions.
Our treatment in an exterior calculus framework allows all the spaces and their properties
to be developed together. The spaces combine together in complexes satisfying the basic
hypotheses of the nite element exterior calculus [5]. This means that, in addition to their
use individually, they can be used in pairs, o
r+1

k1
(T
h
) o
r

k
(T
h
) in a variety of mixed
nite element applications, with stability and convergence following from the abstract theory
of [5]. Element diagrams for some of the spaces are shown in Figure 1. The dimension of
the shape function spaces are given in Theorem 3.6 below and are tabulated for n 4 and
r 7 in Table 1.
Since the development of the rst stable mixed nite elements for the Poisson equation by
Raviart and Thomas in 1977, such elements have proven to be powerful tools for numerical
computation. Their paper [10] introduces a family of nite element discretizations of the
space H(div, ) for a two-dimensional domain , one for each polynomial degree. Together
with a corresponding discontinuous piecewise polynomial discretization of L
2
(), these RT
spaces stably discretize the mixed variational formulation of the Poisson equation on .
In [10], versions of the RT elements were given both for meshes of by triangles and
2000 Mathematics Subject Classication. Primary: 65N30.
Key words and phrases. mixed nite elements, nite element dierential forms, nite element exterior
calculus, cubical meshes, cubes.
The work of the rst author was supported in part by NSF grant DMS-1115291 and the work of the
second author was supported in part by NSF grant DMS-0811052 and the Sloan Foundation.
1
a
r
X
i
v
:
1
2
0
4
.
2
5
9
5
v
1


[
m
a
t
h
.
N
A
]


1
2

A
p
r

2
0
1
2
2 DOUGLAS N. ARNOLD AND GERARD AWANOU
r = 1 r = 2 r = 3
k = 0 o
1

0
o
2

0
o
3

0
k = 1 o
1

1
o
2

1
o
3

1
k = 2 o
1

2
o
2

2
o
3

2
k = 3 o
1

3
o
2

3
o
3

3
Legend: symbols represent the value or moment of the indicated quantities:
scalar (1 DOF); tangential vector eld (2 DOFs); vector eld (3 DOFs);
tangential component of vector eld on edge or normal component on face (1 DOF).
Figure 1. Element diagrams for o
r

k
in three dimensions for r 3.
by rectangles. Besides the original application of mixed nite elements for the Poisson
equation, the RT elements can be used together with the standard Lagrange nite element
discretization of H
1
() to give a stable mixed nite element discretization of the vector
Poisson equation curl curl u grad div u = f, in which the vector variable u is sought in
H(div) and approximated by RT elements, and the scalar variable = div u is sought
in H
1
and approximated by Lagrange elements. The RT elements were generalized to 3-
dimensions by Nedelec [8], with separate generalizations giving discretizations of H(div, )
and of H(curl, ), including both tetrahedral and cubic mesh generalizations of each.
In [7], Brezzi, Douglas, and Marini introduced a second family of nite element discretiza-
tions of H(div, ) in two dimensions, for both triangular and rectangular meshes. These
BDM element have also proven to be very useful. Nedelec [9] generalized the BDM family
FINITE ELEMENT DIFFERENTIAL FORMS ON CUBICAL MESHES 3
r
k 1 2 3 4 5 6 7
n = 1 0 2 3 4 5 6 7 8
1 2 3 4 5 6 7 8
n = 2 0 4 8 12 17 23 30 38
1 8 14 22 32 44 58 74
2 3 6 10 15 21 28 36
n = 3 0 8 20 32 50 74 105 144
1 24 48 84 135 204 294 408
2 18 39 72 120 186 273 384
3 4 10 20 35 56 84 120
n = 4 0 16 48 80 136 216 328 480
1 64 144 272 472 768 1188 1764
2 72 168 336 606 1014 1602 2418
3 32 84 180 340 588 952 1464
4 5 15 35 70 126 210 330
Table 1. Dimension of o
r

k
(I
n
).
to tetrahedral meshes in three dimensions, giving analogues both for H(div) and H(curl).
He also dened H(div) and H(curl) elements on cubic meshes in [9]. However these can-
not really be considered analogues of the BDM elements, as they do not seem to lead to
stable mixed nite element pairs. The generalization of the BDM elements to H(div) on
three-dimensional domains (but not H(curl)), was also made by Brezzi, Douglas, Duran, and
Fortin in [6].
1
The paper [6] also introduced an analogue of the rectangular BDM elements
to cubic meshes in three dimensions.
The nite element exterior calculus [4, 5] has greatly claried the relation of many of
these mixed nite element methods. In exterior calculus the space H(div) is viewed as a
space of dierential forms of degree n 1 in n dimensions, while H(curl) is a space of 1-
forms, H
1
a space of 0-forms, and L
2
a space of n-forms. These spaces are connected via the
de Rham complex, in which the fundamental dierential operators grad, curl, div (and others
in higher dimensions) are unied as the exterior derivative. We refer the reader to Table 2.2
in [4] for a summary of correspondences between dierential forms and vector elds. In
[4, 5] two fundamental families of nite element dierential forms are dened on simplicial
1
The nite element discretization of H(div) in [6] is identical with that of [9] in the case of tetrahedral
meshes. However the interior degrees of freedom dier. This does not aect the stability analysis for the
mixed Poisson equation, but only the ones given by [9] can be used to establish stability when discretizing
a mixed formulation of the vector Laplacian.
4 DOUGLAS N. ARNOLD AND GERARD AWANOU
meshes in n dimensions. Tables 5.1 and 5.2 in [4] summarize the correspondences between
these spaces of nite element dierential forms and classical nite element spaces in two and
three dimensions. The T

r

k
family specializes to the Lagrange elements, the RT elements,
and the fully discontinuous polynomial elements, for k = 0, 1, and 2, respectively, in two
dimensions, and to the Lagrange elements, Nedelecs generalizations of the RT elements to
H(curl) and to H(div), and the discontinuous elements for k = 0, . . . , 3 in three dimensions.
Taken together, these spaces form a complex which is a nite element subcomplex of the
de Rham complex:
0 T

r

0
d
T

r

1
d

d
T

r

n
0.
The second family of nite elements discussed in [4, 5] is the T
r

k
family. For 0-forms
and n-forms this brings nothing new, just recapturing the Lagrange and fully discontinuous
polynomial elements, but for 0 < k < n, this is a dierent family of nite element spaces.
For n = 2, k = 1 it gives the BDM triangular elements, and for n = 3, k = 1 and 2, it gives
Nedelecs generalizations of these to H(div) and H(curl). They combine into a second nite
element de Rham subcomplex
0 T
r

0
d
T
r1

1
d

d
T
rn

n
0.
Note that the degree decreases in this complex, in contrast to the preceeding one.
We now turn to the construction of analogous spaces and complexes for cubical meshes.
An analogue to the T

r

k
complex of elements for cubical meshes may be easily constructed
via a tensor product construction. For an explicit description in n dimensions and for all
form degrees k, see [3]. This includes the tensor product Lagrange, or Q
r
, elements for 0-
forms, the rectangular RT elements for 1-forms in 2-D, and the 3-D generalizations of them
given in [8]. The space for n-forms (L
2
) is the fully discontinuous space of tensor product
polynomials (shape functions in Q
r
).
This paper develops a second family of nite element spaces for cubical meshes. This
family may be viewed as an analogue of the T
r

k
family for cubical meshes. We denote
the new family of elements by o
r

k
and dene such a space for all dimensions n 1, all
polynomial degrees r 1, and all form degrees 0 k n. For 0-forms (discretization of
H
1
), S
r

0
is not equal to Q
r
, but rather to the serendipity elements in 2-D, and to a recent
extension of them for higher dimensions [2]. For n-forms, o
r

n
uses fully discontinuous
elements with shape functions in T
r
(not Q
r
).
For 1-forms in 2-D, the space o
r

1
coincides with the rectangular BDM elements of [7].
For 2-forms in 3-D, we believe that the space o
r

2
, which has not appeared before as far
as we know, is the correct analogue of the BDM elements on cubic meshes. It has the same
degrees of freedom as the space given in [6] but the shape functions have better symmetry
properties. For 1-forms in 3-D, o
r

1
is a nite element discretization of H(curl). To the
best of our knowledge, neither the degrees of freedom nor the shape functions for this space
have been proposed previously. Even for 0-forms, the spaces o
r

0
were only discovered
very recently in more than two dimensions. They are the generalization of the 2-dimension
serendipity spaces given by the present authors in [2]. That work was motivated by the
search for a nite element discretization of the de Rham complex on cubical meshes which
is completed in this paper.
In the remainder of the paper we will develop the o
r

k
spaces and their properties in the
setting of dierential forms on cubical meshes of arbitrary dimensions. Here we explicitly
FINITE ELEMENT DIFFERENTIAL FORMS ON CUBICAL MESHES 5
describe the spaces which arise in three dimensions in traditional nite element terminology
by giving their shape functions and degrees of freedom. See also Figure 1.
The space o
r

3
. This is simply the space of piecewise polynomials of degree at most r
with no continuity requirements. Obviously this gives a nite element subspace of L
2
(which,
unlike the remaining spaces, is dened also for r = 0). Its dimension is (r+1)(r+2)(r+3)/6.
The space o
r

2
. The shape functions for the space o
r

2
on the unit cube I
3
in three
dimensions are vector polynomials of the form
(1) u = (v
1
, v
2
, v
3
) + curl(x
2
x
3
(w
2
w
3
), x
3
x
1
(w
3
w
1
), x
1
x
2
(w
1
w
2
)),
with v
i
, w
i
T
r
(I
3
) and w
i
independent of x
i
. The dimension of this space is (r + 1)(r
2
+
5r + 12)/2. As degrees of freedom for u o
r

2
we take
u
_
f
u np, p T
r
(f), f a face of I
3
; u
_
I
3
u p, p [T
r2
(I
3
)]
3
.
These degrees of freedom
The space o
r

1
. The shape functions for the space o
r

1
in three dimensions are vector
polynomials of the form
(2) u = (v
1
, v
2
, v
3
) + (x
2
x
3
(w
2
w
3
), x
3
x
1
(w
3
w
1
), x
1
x
2
(w
1
w
2
)) + grad s,
with v
i
T
r
(I
3
), w
i
T
r1
(I
3
) independent of x
i
, and s is a polynomial of I
3
with super-
linear degree at most r + 1, where the superlinear degree of a polynomial is the ordinary
degree ignoring variables which enter linearly (e.g., the superlinear degree of x
2
1
x
2
x
3
3
is 5).
The dimension of this space is (r + 1)(r
2
+ 5r + 18)/2. As degrees of freedom for u o
r

1
we take
u
_
e
u t p, p T
r
(e), e an edge of I
3
;
u
_
f
u np, p [T
r2
(f)]
2
, f a face of I
3
; u
_
I
3
u p, p [T
r4
(I
3
)]
3
.
The space o
r

0
. Finally, the space S
r

0
is the generalized serendipity space of [2]. The
shape functions are all polynomials of superlinear degree at most r, and the degrees of
freedom are the values at the vertices and the moments of degree at most r 2, r 4, and
r 6 on the edges, faces, and interior, respectively. The dimension of this space is 8 for
r = 1, 20 for r = 2, and (r + 1)(r
2
+ 5r + 24)/6 for r 3.
As a consequence of the general theory below, the degrees of freedom given are unisol-
vent for all these spaces, and for any cubical decomposition T of , the assembled nite
element spaces o
r

k
(T ) have exactly the continuity required to belong to H
1
(), H(div, ),
H(curl, ), and L
2
(), for k = 0, . . . , 3, respectively. In other words, o
r

k
(T ) belongs to the
domain of the exterior derivative on k-forms. Moreover, exterior derivative maps o
r

k
(T )
into o
r1

k+1
(T ), so we obtain a nite element subcomplex
0 o
r

0
(T )
grad
o
r1

1
(T )
curl
o
r2

2
(T )
div
o
r3

3
(T ) 0
of the de Rham complex. Finally, if we dene projection operators
k
from smooth elds into
the nite element spaces S
r

k
using the degrees of freedom, these commute with exterior
6 DOUGLAS N. ARNOLD AND GERARD AWANOU
dierentiation. That is, the following diagram commutes:
0 C

()
grad
[C

()]
3
curl
[C

()]
3
div
C

() 0

_
1

_
2

_
3

_
0 o
r

0
(T )
grad
o
r1

1
(T )
curl
o
r2

2
(T )
div
o
r3

3
(T ) 0
where, for simplicity, we assume r 3 (otherwise some of the spaces are undened and some
parts of the diagram are not applicable).
The remainder of the paper consists of four sections. In Section 2, we recall some key
concepts from exterior calculus, particularly the Koszul dierential and Koszul complex of
polynomial dierential forms, which will be crucial to our construction. We introduce the
concept of linear degree, and show that the subcomplex obtained from the Koszul complex
by placing a lower bound on linear degree is exact. This is a key step in the unisolvence
proof in Section 3. In Section 3, we dene the spaces o
r

k
on an n-dimensional cube by
giving the shape functions and degrees of freedom. We derive a number of properties from
these denitions, leading to a formula for the dimension of the space of shape functions and
a proof of unisolvence of the degrees of freedom. We then dene projection operators
k
r
mapping smooth elds into the nite element spaces o
r

k
and show that the projections
commute with the exterior derivative. This accomplished, the new elements t squarely into
the framework of the nite element exterior calculus given in [5]. Therefore the application of
the elements to PDE problems and their numerical analysis does not require new ideas, and
so we do not discuss that here. The proof of unisolvence in Section 3 hinges on the special
case of functions with vanishing trace. The proof in that case is the topic of Section 4.
2. Notation and preliminaries
For n 1, the number of dimensions, let N
n
= 1, . . . , n, and let (k), 0 k n, denote
the set of subsets of N
n
consisting of k elements. For (k) we denote by

(n k)
its complement N
n
. For p we write p for p (k 1), and for q

we write + q for q (k + 1). If N

n
and q

we let (q, ) = (1)

l
where
l = #p [p < q, and set (q, p, ) := (q, )(p, + q p) for q

, p . For later
reference, we note that
(3) (q, p, ) = (p, p)(q, p),
which is easily veried by considering the cases q < p and q > p separately.
We now recall some basic tools and results of exterior algebra and exterior calculus. These
can be found, for example in [4, Section 2]. For each (k) we denote by
1
, . . . ,
k
its
elements in increasing order, and by
dx

= dx

1
dx

k
the corresponding basic alternator. A dierential k-form on a domain R
n
may be
written as
(4) =

(k)

dx

,
FINITE ELEMENT DIFFERENTIAL FORMS ON CUBICAL MESHES 7
with coecients

belonging to any desired space of functions on , e.g., L

2
(). A 0-form
is simply such a function. The exterior derivative of the dierential form (4) is
d =

(k)

qNn

dx
q
dx

,
where
q
= /x
q
.
A dierential k-form may be contracted with a vector eld on to give a dierential
(k 1)-form (or zero if k = 0). When the vector eld is simply the identity, the resulting
operator is the Koszul dierential. Equivalently, we may dene the Koszul dierential on
the basic alternators by
(dx

1
dx

k
) =
k

i=1
(1)
i+1
x

i
dx

1


dx

i
dx

k
,
and then extend it to a general dierential form by linearity:
(5) (

dx

) =

i=1
(1)
i+1
x

i
dx

1


dx

i
dx

k
.
The operator is a graded dierential, meaning that
() = 0, ( ) = + (1)
k
,
if is a k-form and an l-form.
The following lemma collects formulas for , d, and d.
Lemma 2.1. If is given by (4), then
=

(k1)

dx

, where

(q, )x
q

+q
, (6)
d =

(k+1)

dx

, where

q
(q, q)
q

q
, (7)
d =

(k)

dx

, where

_
x
q

p
(q, p, )x
q

+qp

. (8)
Proof. By denition
dx

q
(q, q)x
q
dx
q
,
so
=

(k)

q
(q, q)x
q

dx
q
.
Making the change of variable = q, so
(k), q (k 1), q

,
we obtain (6). The second result is proven similarly, and the third follows from the rst
two.
8 DOUGLAS N. ARNOLD AND GERARD AWANOU
We now turn to dierential forms with polynomial coecients. A monomial in n variables
is determined by a multi-index of n nonnegative integers: x

= x

1
1
x
n
n
. By a form
monomial in n variables, we mean the product of a monomial with a basic alternator: m =
x

dx

for some multi-index and N

n
. The polynomial degree and the linear degree of
m are dened
deg m = [[ :=

i
, ldeg m = # i

[
i
= 1.
Thus the linear degree of m is the degree of its polynomial coecient counting only those
variables which enter linearly, and excluding variables which enter the alternator. For or-
dinary monomials, i.e., 0-forms, the linear degree is equal to the dierence between the
polynomial degree and the superlinear degree which appeared in the introduction.
We dene 1
r

k
= 1
r

k
(R
n
) to be the span of the k-form monomials m with deg m = r,
and T
r

k
= T
r

k
(R
n
) =

r
s=0
1
s

k
to be the span of those with deg m r. If k = 0, we
may simply write 1
r
and T
r
. If is a subdomain of R
n
, we dene T
r

k
() to be the space
of restrictions to of the elements of T
r

k
(and similarly for other function spaces). Note
that d maps 1
r

k
into 1
r1

k+1
while maps 1
r

k
into 1
r+1

k1
. An extremely useful
identity is the homotopy formula ([4, Theorem 3.1]):
(9) (d + d) = (r + k), 1
r

k
(R
n
).
We extend the linear degree for form monomials to polynomial dierential forms by den-
ing ldeg for any 1
r

k
to be the minimum of the linear degree among all the monomials
m in . We say that is of homogeneous linear degree equal to l if ldeg m = l for every
monomial of . We denote by 1
r,l

k
the space of forms in 1
r

k
of linear degree at least l.
Obviously,
(10) 1
r,l

k
(R
n
) = 0, l > min(r, n k).
The exterior derivative d may decrease the linear degree of a polynomial dierential form,
but and d do not.
Lemma 2.2. For any T
r

k
, ldeg ldeg and ldeg d ldeg .
Proof. If m is monomial of and l = ldeg m, then it follows directly from the denition (5)
that the monomials of m are of linear degree l and/or l +1, so ldeg m l ldeg . Since
every monomial of is a monomial of m for some monomial m of , this implies the rst
inequality. For the second we use the dierential property of and the homotopy formula to
see that dm = (d+d)m is a multiple of m. Therefore ldeg dm = ldeg m ldeg ,
which gives the second inequality.
In view of Lemma 2.2, for each l 0, we obtain a complex
(11)

1
r1,l

k+1

1
r,l

k

.
When l = 0 this is the Koszul complex, and exactness follows from the homotopy formula.
In fact, the complex (11) is exact for all l 0.
Theorem 2.3. For r 1, 0 l < r, 0 k < n, the sequence
1
r1,l

k+1

1
r,l

k

1
r+1,l

k1
is exact. Equivalently, dim(1
r1,l

k+1
) + dim(1
r,l

k
) = dim1
r,l

k
.
FINITE ELEMENT DIFFERENTIAL FORMS ON CUBICAL MESHES 9
The proof is due to Scot Adams and Victor Reiner [1]. Its main ingredient is contained in
the following lemma.
Lemma 2.4. Let r 1, l 1, and 0 k < n. Suppose that 1
r

k
is of linear degree
at least l 1 and is of linear degree at least l. Then there exists 1
r1

k+1
such that
is of linear degree at least l. Further, if 1
r

n
is nonzero, then is of linear
degree 0.
Proof. For the nal statement, concerning n-forms, we write = p dx
1
dx
n
where
p 1
r
. Since r 1, p is not constant. But then =

i
(1)
i+1
px
i
dx
1

dx
i
dx
n
is easily seen to be of linear degree 0.
For 0 k < n, the proof hinges on a canonical form for an element of 1
r

k
which we
establish before proceeding. Let us say that a form monomial x

dx

is full if supp(),
the support of . To each of the monomials m = x

dx

of 1
r

k
, we associate the
increasing sequences and with = supp() and = supp(). Then m = dx

where = x

dx

is a full form monomial which is independent of the variables (that is,

supp() is disjoint from ). Note that ldeg = ldeg m. Finally, in the expansion of
as a linear combination of its monomials, we gather together the terms with the same
= supp(), and in this way write
(12) =

Nn

dx

,
where

1
r

k#
is independent of the variables, and has all of its monomials full. The
expression on the right-hand side of (12) is the desired canonical form of .
Now we proceed with the proof of the lemma. We consider rst the special case in which
is of homogeneous linear degree l 1, and, in this special case, we use induction on l, the
case l = 0 being known (exactness of the Koszul complex). Expressing in the canonical
form (12), we have ldeg

= l 1 for each (for which the coecient

does not vanish).

Now
(13) =

dx

dx

.
The rst sum is of homogeneous linear degree l 1 and the second of homogeneous linear
degree l. Since we assumed that is of linear degree at least l, the rst sum must vanish.
But this sum is in canonical form, so we conclude that

= 0 for each . Invoking the

inductive hypothesis, we can write

where

has linear degree at least l 1. Let

=

dx

1
r1

k+1
.
Then
=

dx

dx

dx

dx

.
Expanding the right-hand side into monomials, we see that each has ldeg at least l, so this
completes the proof under the assumption that is of homogeneous linear degree l 1.
Next we turn to the general case, in which is of linear degree at least l 1. We may
split as

with

of homogeneous linear degree l 1 and

of linear degree at least

l. Then

splits into a part of homogeneous linear degree l 1 and a part of homogeneous

linear degree l, while ldeg(

) l. Since, by assumption, ldeg() l, the part of

with
10 DOUGLAS N. ARNOLD AND GERARD AWANOU
linear degree equal to l 1 must vanish. That is, ldeg(

) l. Therefore, we may apply

the result of preceding special case to

to obtain 1
r1

k+1
such that ldeg(

) l.
Then = (

) +

is of linear degree at least l.

Finally, we give the proof of Theorem 2.3.
Proof. The result is certainly true for l = 0, so we may assume 1 l < r (and so r 2).
Suppose 1
r,l

k
with = 0. We must show that there exists 1
r1

k+1
with linear
degree at least l, such that = . Now := d/(r + k) 1
r1

k+1
is of linear degree at
least l 1 and satises = (d +d)/(r +k) = by (9). If k = n 1, the nal sentence
of the lemma insures that = 0. For 0 k < n 1, we apply the lemma with r and k
replaced by r 1 and k +1, respectively, and conclude that there exists 1
r2

k+2
such
that := is of linear degree at least l. Clearly = = .
3. The o
r

k
spaces
In this, the main section of the paper, we dene the polynomial spaces o
r

k
(R
n
) we shall
use as shape functions (see (17) below) and the degrees of freedom for these (see (21)). We
derive a number of properties of these polynomial spaces in Theorems 3.2 through 3.5 and
use them to verify unisolvence in Theorem 3.6.
The space of shape functions will consists of polynomials of a given degree plus certain
additional terms of higher degree which will be dened in terms of the following auxilliary
space:

k
(R
n
) =

l1
1
r+l1,l

k+1
(R
n
).
In view of (10), the sum is nite and
(14)
r

k
(R
n
) T
r+nk1

k
(R
n
).
Moreover, the sum is direct, since the polynomial degrees of the summands dier. The
following proposition, which follows directly from the denitions, helps to clarify the meaning
of this space.
Proposition 3.1. 1. The space

l1
1
r+l1,l

k
(R
n
) is the span of all k-form monomials
m with deg m r and deg mldeg m r 1.
2. The space
r

k
(R
n
) is the span of m for all (k+1)-form monomials m with deg m r
and deg mldeg m r 1.
For several values of k, this space can be described more explicitly. By (10),
(15)
r

k
(R
n
) = 0, for k = n or n 1,
while

n2
(R
n
) = 1
r,1

n1
(R
n
).
Now, 1
r,1

n1
(R
n
) is the span of the monomials x
i
w
i

i
, where w
i
1
r1
(R
n
) is independent
of x
i
and
i
:= (1)
i1
dx
1
. . .

dx
i
. . . dx
n
. We then have

i
=

j<i
x
j

j,i

j>i
x
j

i,j
,
FINITE ELEMENT DIFFERENTIAL FORMS ON CUBICAL MESHES 11
with

i,j
= (1)
i+j
dx
1
. . .

dx
i
. . .

dx
j
. . . dx
n
.
Therefore,
(16)
r

n2
(R
n
) =

i<j
x
i
x
j
(w
i
w
j
)
i,j
[ w
i
1
r1
(R
n
) independent of x
i
.
Finally, we identify
r

0
(R
n
). By Theorem 2.3 in the case k = 0, we see that
r

0
(R
n
) =

l1
1
r+l,l

0
(R
n
). By Proposition 3.1, this space is the span of monomials of degree > r
whose superlinear degree, that is, deg ldeg, is at most r.
We can now dene the space of polynomial k-forms which we use for shape functions:
(17) o
r

k
(R
n
) := T
r

k
(R
n
) +
r

k
(R
n
) + d
r+1

k1
(R
n
),
dened for all r 1 and all 0 k n. From (14),
(18) o
r

k
(R
n
) T
r+nk

k
(R
n
).
Note that, in case k = 0, the nal term in (17) vanishes, and, by the characterization of

0
(R
n
) just derived, o
r

0
(R
n
) consists precisely of the span of all monomials of superlinear
degree at most r. This is exactly the serendipity space as dened in [2]. In this case, (14)
gives the sharper degree bound
(19) o
r

0
(R
n
) T
r+n1

0
(R
n
).
Another case in which the expression for o
r

k
(R
n
) can be simplied is when k = n. By
(15),
o
r

n
(R
n
) = T
r

n
(R
n
).
When k = n 1, we have, by (15),
o
r

n1
(R
n
) := T
r

n1
(R
n
) + d
r+1

n2
(R
n
),
where the last space is characterized in (16). In the case of three dimensions, this is the
formula (1) given in the introduction, stated in the language of exterior calculus. In a similar
way, we recover the formula (2) for the 3-D H(curl) elements discussed in the introduction.
We now derive several properties of these polynomial spaces. The rst limits the mono-
mials that appear in the polynomials in o
r

k
(R
n
).
Theorem 3.2 (Degree property). For any n, r 1 and 0 k n, the space o
r

k
(R
n
) is
contained in the span of the k-form monomials m of degree at most r +nk
k0
for which
(20) deg mldeg m r + 1
k0
.
Proof. The bound r+nk
k0
on the degree is given in (18) for k > 0 and in (19) for k = 0,
so we need only show (20). If m is a monomial of an element of T
r

k
(R
n
), then deg m r
and ldeg m 0, so (20) holds with r on the right-hand side. If m is a monomial of an element
of
r

k
(R
n
), then m occurs in the expansion of p, where p is a (k +1)-form monomial with
deg p ldeg p r 1 (Proposition 3.1). Then deg m = deg p + 1 and, by Lemma 2.2,
ldeg m ldeg p, so again deg m ldeg m r. Finally, if k > 0 and m is a monomial of an
element of d
r+1

k1
(R
n
), then by the argument just given, m is a monomial of dq where
q is a (k 1)-form monomial with deg q ldeg q r + 1. Since deg m = deg q 1 and
ldeg m ldeg q 1, we get (20).
12 DOUGLAS N. ARNOLD AND GERARD AWANOU
A crucial property of these polynomial form spaces, is that they can be combined to form
a subcomplex of the de Rham complex.
Theorem 3.3 (Subcomplex property). Let n, r 1, and let 0 < k n. Then
do
r+1

k1
(R
n
) o
r

k
(R
n
).
Proof. With reference to (17), we note that dT
r+1

k1
(R
n
) T
r

k
(R
n
), and d(d
r+2

k
(R
n
))
vanishes, so it suces to prove that d
r+1

k1
(R
n
) o
r

k
(R
n
), which is immediate from
(17).
We next observe that the spaces increase with increasing polynomial degree.
Theorem 3.4 (Inclusion property). Let n, r 1, and let 0 k n. Then
o
r

k
(R
n
) o
r+1

k
(R
n
).
Proof. We must show each of the three summands on the right-hand side of (17) is included
in o
r+1

k
(R
n
). Clearly
T
r

k
(R
n
) T
r+1

k
(R
n
) o
r+1

k
(R
n
),
which establishes the rst inclusion. Next, we show that

k
(R
n
) T
r+1

k
(R
n
) +
r+1

k
(R
n
) o
r+1

k
(R
n
).
By Proposition 3.1, elements of
r

k
(R
n
) are of form m with m a (k + 1)-form monomial
with deg m r, deg m ldeg m r 1. By the homotopy formula (9), m is a constant
multiple of p with p = dm. We have deg p = deg m and ldeg p ldeg m 1. Then
deg p ldeg p deg mldeg m+1 r. If deg p = r, deg m = r +1 and m T
r+1

k
(R
n
).
On the other hand, if deg p r +1, by Proposition 3.1, m
r+1

k
(R
n
). This establishes
the second inclusion. To complete the proof, we show that d
r+1

k1
(R
n
) o
r+1

k
(R
n
).
Since
r+1

k1
(R
n
) o
r+2

k1
(R
n
) (by the inclusion just established), we infer from the
subcomplex property that d
r+1

k1
(R
n
) do
r+2

k1
(R
n
) o
r+1

k
(R
n
).
The third property of the o
r

k
spaces we establish concerns traces on hyperplanes. Con-
sider a hyperplane f of R
n
of the form x
i
= c for some 1 i n and some constant c.
The variables x
j
, j ,= i, form a coordinate system for f, so we may identify f with R
n1
and consider the space o
r

k
(f). It is a space of polynomial k-forms on f, and so vanishes
if k = n. Next, we consider the trace on f of a dierential form in n variables (dened as
the pullback of the form through the inclusion map f R
n
). Let (k), and let

be
a function of n variables. Then
tr
f
(

dx

) =
_
0, i ,
(tr
f

) dx

, i / .
In the last expression, tr
f

denotes the function of n 1 variables obtained by setting

x
i
= c and we view dx

as a basic alternator in the n 1 variables x

j
, j ,= i. The trace
property states that if u o
r

k
(R
n
), then tr
f
u, which is a polynomial k-form on f, belongs
to o
r

k
(f).
FINITE ELEMENT DIFFERENTIAL FORMS ON CUBICAL MESHES 13
Theorem 3.5 (Trace property). Let n, r 1, 0 k n, and let f be a hyperplane of R
n
obtained by xing one coordinate. Then
tr
f
o
r

k
(R
n
) o
r

k
(f).
Proof. Without loss of generality, we assume that f = x R
n
[ x
1
= c . First let us
comment on the Koszul operator applied to a polynomial dierential form on f. Such a
form may be written as a linear combination of monomials x

dx

where
1
= 0 and 1 / .
Referring to (5) we see that, if we view x

dx

as a form monomial in n variables and take

the Koszul dierential, the result is the same as if we view it as form monomial in n 1
variables and take the Koszul dierential. Thus we need not distinguish between the Koszul
dierential on R
n
and that on f.
We will prove the theorem by induction on k. For k = 0 we recall that o
r

0
(R
n
) is the
serendipity space spanned by the monomials of superlinear degree at most r, and, of course,
the superlinear degree does not increase when taking the trace. Hence, tr
f
o
r

0
(R
n
)
o
r

0
(f).
To prove the theorem for k > 0, assume that it holds with k replaced by k 1. In light of
(17) we need to show that the traces of each of the three spaces T
r

k
(R
n
),
r

k
(R
n
), and
d
r+1

k1
(R
n
) are contained in
o
r

k
(f) = T
r

k
(f) +
r

k
(f) + d
r+1

k1
(f).
For the T
r

k
(R
n
), this is evident, since tr
f
T
r

k
(R
n
) T
r

k
(f).
Next, we establish that tr
f
d
r+1

k1
(R
n
) o
r

k
(f). Indeed,
tr
f
d
r+1

k1
(R
n
) = d tr
f

r+1

k1
(R
n
) d tr
f
o
r+1

k1
(R
n
) do
r+1

k1
(f) o
r

k
(f),
where we have used, in turn, the commutativity of the trace with exterior dierentiation,
(17), the inductive hypothesis, and Theorem 3.3.
By Proposition 3.1, in order to show that tr
f

r

k
(R
n
) o
r

k
(f), and so complete the
proof, it suces to show that tr
f
m o
r

k
(f) whenever m is a (k+1)-form monomial with
deg mldeg m r 1. We write m as x

dx

, and consider separately the cases 1 / and

1 .
Assuming 1 / , let p be the (k + 1)-form monomial obtained restricting m to f, i.e.,
by setting x
1
= c in the coecient x

. Then tr
f
m = p. If m is linear in x
1
, then
deg p = deg m 1 and ldeg p = ldeg m 1. Otherwise deg p deg m and ldeg p = ldeg m.
In either event, deg p ldeg p deg m ldeg m r 1, so p o
r

k
(f) (again using
Proposition 3.1).
Assuming, instead, that 1 , we may write
m = x

dx

= x

1
1
x

dx
1
dx

,
where is a multi-index with
1
= 0 and (k) has
1
> 1. Then
p := tr
f
m = c

1
+1
x

dx

is a k-form monomial independent of x

1
with deg p deg m and ldeg p = ldeg m. We are
trying to show that p o
r

k
(f). This is obvious if deg p r, so we may assume that
deg p > r. We shall show that both dp and dp belong to o
r

k
(f), which suces by (9).
Now deg dp = deg p 1 deg m 1 and ldeg dp ldeg p 1 = ldeg m 1, so deg dp
ldeg dp deg mldeg m r 1. Therefore dp
r

k
(f) o
r

k
(f), as required.
14 DOUGLAS N. ARNOLD AND GERARD AWANOU
Finally, we show that p
r+1

k1
(f), whence dp o
r

k
(f) as well. By Proposi-
tion 3.1 this holds, since deg p r + 1 and deg p ldeg p deg m ldeg m r (even
r 1). This concludes the proof.
Having dened the space o
r

k
(R
n
) of polynomial dierential forms, we turn now to the
denition of the associated nite element space on a cubical mesh. As usual the nite element
space is dened element by element, by specifying a space of shape functions and a set of
degrees of freedom on each cube T in the mesh. (More generally the element T may be a
right rectangular prism, that is the Cartesian product of n closed intervals of positive nite
length.) As shape functions on T we use o
r

k
(T), the restriction of the above polynomial
space to the cube. The degrees of freedom have a very simple expression. Writing
d
(T) for
the set of d-dimensional faces of T, they are given by
(21)
_
f
tr
f
, T
r2(dk)

dk
(f), f
d
(T), k d min(n, r/2| + k).
Since
dimT
r2(dk)

dk
(f) = dimT
r2(dk)
(f)
_
d
d k
_
=
_
r d + 2k
d
__
d
d k
_
for f a face of dimension d, and since there are 2
nd
_
n
d
_
d-dimensional faces of an n-cube,
the number of degrees of freedom in (21) is given by
(22)
min(n,r/2+k)

d=k
2
nd
_
n
d
__
r d + 2k
d
__
d
k
_
.
We now turn to one of the main results of this paper, the proof that the degrees of freedom
(21) are unisolvent for o
r

k
(T).
Theorem 3.6 (Unisolvence). Let n, r 1 and 0 k n, and let T be a cube in R
n
. Then
(1) dimo
r

k
(T) is given by (22).
(2) If o
r

k
(T) and all the degrees of freedom in (21) vanish, then 0.
Using the trace property, we will reduce the proof of (2) to the case where belongs to
the space

o
r

k
(T) := o
r

k
(T) [ tr
f
= 0 on each face f
n1
(T) ,
the subspace with vanishing traces. This case is given in the following proposition.
Proposition 3.7. If

o
r

k
(T) and
(23)
_
T
= 0, T
r2(nk)

nk
(T),
then vanishes.
We defer the proof of this proposition to Section 4. Now, assuming this result, we prove
Theorem 3.6.
FINITE ELEMENT DIFFERENTIAL FORMS ON CUBICAL MESHES 15
Proof. We begin with the rst statement of the theorem. Since
r

k
(T) is in the range of
, and since the homotopy formula implies that no nonzero dierential form is in the range
of both and d, the sum on the left of (17) is direct. The homotopy formula implies as well
that d is injective on the range of . Therefore
dimo
r

k
(T) = dimT
r

k
(T) +

l1
dim1
r+l1,l

k+1
(T) +

l1
dim1
r+l,l

k
(T).
Applying Theorem 2.3, this becomes
dimo
r

k
(T) = dimT
r

k
(T) +

l1
dim1
r+l,l

k
(T) = dim[T
r

k
(T) +

l1
1
r+l,l

k
(T)].
The space in brackets is exactly the span of the k-form monomials m with deg mldeg m r,
and hence we need only count these monomials. This gives
(24) dimo
r

k
(T) =
_
n
k
_
#A(r, k, n),
where
_
n
k
_
is the number of basic alternators and A(r, n, k) is the set of monomials p in n
variables which are linear in some number l of the rst n k variables, x
1
, . . . , x
nk
, with
deg pl r. Now we count the elements of A(r, n, k). For any monomial p in n variables let
J N
nk
be the set of indices for which x
i
enters p superlinearly, let d 0 be the cardinality
of J, and for i N
nk
J, let a
i
= 0 or 1 according to whether p is of degree 0 or 1 in x
i
.
Then
p =
_

jJ
x
2
j
_
q
_

iN
nk
\J
x
a
i
i
_
,
where q is a monomial in the d variables indexed by J and the last k variables. With l
the number of a
i
equal to 1, we have deg p = 2d + deg q + l. Thus deg p l r if and
only if deg q r 2d. Thus we may uniquely specify an element of A(r, k, n) by choosing
d 0, choosing the set J consisting of d of the n k variables (for which there are
_
nk
d
_
possibilities), choosing the monomial q of degree at most r2d in the d+k variables (
_
rd+k
d+k
_
possibilities), and choosing the exponent a
i
to be either 0 or 1 for the n k d remaining
indices (2
nkd
possibilities). Thus
#A(r, k, n) =
min(nk,r/2)

d=0
2
nkd
_
n k
d
__
r d + k
d + k
_
=
min(n,r/2+k)

d=k
2
nd
_
n k
d k
__
r d + 2k
d
_
,
(25)
where the second sum comes from a change of the summation index (d dk). Substituting
(25) into (24) and using the binomial identity
_
n
k
__
n k
d k
_
=
_
d
k
__
n
d
_
,
16 DOUGLAS N. ARNOLD AND GERARD AWANOU
we conclude that
(26) dimo
r

k
(T) =
min(n,r/2+k)

d=k
2
nd
_
n
d
__
r d + 2k
d
__
d
k
_
.
This completes the proof of the dimension formula for o
r

k
(R
n
).
The proof of unisolvence is easily completed based on the trace property and Proposi-
tion 3.7. We use induction on the dimension n, the one-dimensional case being trivial.
Suppose o
r

k
(T) and all its degrees of freedom vanish. For any face f of dimension
n 1, tr
f
S
r

k
(f) and all the degrees of freedom for it vanish. By induction tr
f
0
on f. This implies that

o
r

k
(T), and we invoke Proposition 3.7 to conclude that
vanishes identically.
With the denition of the spaces complete, we use the subcomplex property, Theorem 3.3,
to dene a subcomplex of the de Rham complex on the cube:
R

o
r

0
(T)
d
o
r1

1
(T)
d

d
o
rn

n
(T) 0.
To show that this complex is exact, we dene the canonical projection

k
r
: C
k
(T) o
r

k
(T),
associated to the unisolvent degrees of freedom. That is,
k
r
o
r

k
(T) is determined by
the equations
_
f
tr
f
(
k
r
) =
_
f
tr
f
, T
r2(dk)

dk
(f), f
d
(T), k d min(n, r/2|+k).
Then, the following diagram commutes:
R

C

0
(T)
d
C

1
(T)
d

d
C

n
(T) 0

0
r

1
r1

n
rn
R

o
r

0
(T)
d
o
r1

1
(T)
d

d
o
rn

n
(T) 0
The proof of commutativity is based on two basic properties of dierential forms: (1) the
commutativity of trace and exterior dierentiation, tr
f
d = d
f
tr
f
, and (2) integration by
parts, which for dierential forms can be written
_

d = (1)
k1
_

d +
_

tr

tr

,
for a k-form and an (nk1)-form on an n-dimensional domain . See [4, Lemma 4.24]
for the same argument applied to simplicial elements. Since the top row of the diagram, the
de Rham complex on the cube, is exact, the commutativity of the diagram implies that the
bottom row is exact as well.
Having dened the nite element space o
r

k
(T) on a single cube T and established its
properties, the space o
r

k
(T
h
) associated to a cubical mesh T
h
is dened through the usual
nite element assembly. We have seen that the degrees of freedom associated to a face of
the cube and its subfaces determine the trace of the nite element dierential form on the
face. It follows that o
r

k
(T
h
) H
k
() (see [4, Section 5.1]).
FINITE ELEMENT DIFFERENTIAL FORMS ON CUBICAL MESHES 17
4. Unisolvence over the space with vanishing traces
We conclude the paper with the proof of Proposition 3.7, which is based on the following
lemma.
Lemma 4.1. Suppose that =

(k)

dx

where, for each (k),

is a ho-
mogeneous polynomial which is superlinear in all the

variables. Further suppose that

ldeg 1 and ldeg d 1. Then = 0.
Proof of Lemma 4.1. By (8) we have d =

(k)

dx

, where
(27)

_
x
q

p
(q, p, )x
q

+qp

.
Since ldeg d 1, each monomial of the polynomial

is linear in at least one

variable.
For any (k) and any subset of , let S

be the span of the (ordinary, 0-form)

monomials which are independent of the variables but depend on all the other variables,
and let S

denote the span of the monomials which are superlinear in all the

variables.
Denote by P

the projection onto S

. That is, if p =

p
m
m where the sum is over all
monomials m and the coecients p
m
are real numbers, all but nitely many zero, then
P

p :=

mS

p
m
m. Similarly denote by Q

the projection onto S

. We now calculate
the result of applying Q

to both sides of (27). First we note that

(28) Q

= 0,
since every monomial of

is linear in at least one

variable, and so none of them belong

to S

. Next, for each q

,
(29) Q

(x
q

) = P

(x
q

) = x
q

q
(P

),
with the rst inequality holding since each monomial of

, and so of x
q

is superlinear
in all the variables. Now we determine the action of Q

on the terms of the second sum

on the right-hand side of (27). For any q

and p , we claim that

(30) Q

(x
q

+qp
) =
_
0, p ,
x
q

p
(P

+qp

+qp
), p .
Indeed,
+qp
is superlinear in x
p
, so every monomial of x
q

+qp
depends on x
p
. This
implies that the projection is 0 in the case p . In case p , we write x
q

+qp
=

x
q

p
m, where the sum is over the monomials m of
+qp
. Since neither p or q belongs to
, the monomial m

:= x
q

p
m is independent of the variables if and only if same is true of
the monomial m, and, since m

always depends on x
p
, it depends on all the variables if
and only if m depends on all the ( p) variables. Further, m

is always superlinear in
all the

variables except possibly x

q
, and it is superlinear in x
q
if and only if m depends
on x
q
. In short, m

belongs to S

if and only if m belongs to S

+qp
. This completes
the verication of (30). Thus the application of Q

to (27) gives, in light of (28), (29), and

(30), that
(31)

_
x
q

q
(P

) +

p\
(q, p, )x
q

p
(P

+qp

+qp
)

= 0.
18 DOUGLAS N. ARNOLD AND GERARD AWANOU
We now claim that, for any (k) and , that
(32)

p\
(q, p, )x
q

p
(P

+qp

+qp
) =
_
c +

i\
x
i

i
_
(P

),
where c = #( ). Assuming this, we have from (31),
0 =
_
c +

i\
x
i

i
+

x
q

q
_
(P

) =
_
c +

x
i

i
_
(P

) = (c + r + 1)(P

),
where we have used Eulers formula for homogeneous polynomials (i.e., the homotopy formula
for 0-forms) in the last step. Thus P

vanishes, and so

also vanishes.
Since (k) was arbitrary, this implies that = 0.
Thus it remains only to prove (32). By the rst formula of Lemma 4, =

(k1)

dx

where

(q, )x
q

+q
.
Now let (k 1) and , and denote by R

the projection onto S

(the span
of monomials which depend on all the variables except the variables and which are
superlinear in the

variables). By the hypothesis that ldeg 1,

R

= 0.
Next we compute R

(x
q

+q
) for q

. If m is a monomial of
+q
, then m

= x
q

+q
depends on all the variables except for the variables if and only if the same is true of m.
Moreover, m

is superlinear in all the

variables if and only if m depends on x

q
(since m is
superlinear in all the

variables with the possible exception of x

q
). Thus
R

(x
q

+q
) = x
q
P

+q

+q
Combining the last three displayed equations, we obtain

(q, )x
q
P

+q

+q
= 0.
Now choose some p

and dierentiate this equation with respect to x

p
to get
(p, )
p
(x
p
P

+p

+p
) +

p
(q, )x
q

p
(P

+q

+q
) = 0,
or, after rearranging,
(33) P

+p

+p
+ x
p

p
(P

+p

+p
) =

p
(p, )(q, )x
q

p
(P

+q

+q
).
For any (k) and any p set = p (k 1). Then we can rewrite (33) in
terms of :
P

+ x
p

p
(P

) =

(p, p)(q, p)x

q

p
(P

+qp

+qp
).
Finally, taking any (k), we sum over p to obtain
cP

p\
x
p

p
(P

) =

p\

(p, p)(q, p)x

q

p
(P

+qp

+qp
).
FINITE ELEMENT DIFFERENTIAL FORMS ON CUBICAL MESHES 19
where c = #( ). In light of (3), this establishes (32), and so completes the proof of the
lemma.
Finally we give the proof of Proposition 3.7.
Proof. By dilating and translating, it suces to prove the result when T = I
n
with I =
[1, 1]. Let
=

(k)

dx

,
be a polynomial dierential form on the cube I
n
. Then tr
f
vanishes on the faces x
i
= 1
if and only if for each such that i / , 1 x
2
i
divides

. Now suppose that

o
r

k
(I
n
),
so that tr
f
vanishes on all the faces of the cube. It follows that

(1 x
2
i
)
for some polynomial

. The monomial expansion of then contains the form monomial

m

x
2
i
dx

, where m

is any monomial of highest degree of

. The linear degree of

this form monomial is 0, so, by the degree property (20), its degree is at most r + 1.
Having established that is of degree at most r + 1, let be its homogeneous part of
degree r+1. We have ldeg = 0. Now we may match terms in the denition (17) of o
r

k
()
to obtain that = + d, for some 1
r,1

k+1
(I
n
) and 1
r+1,1

k
(I
n
). Therefore,
ldeg = ldeg d ldeg 1 and ldeg d = ldeg d ldeg 1, where we have
used Lemma 2.2.
By Lemma 4.1, = 0, hence the monomial of highest order in the expansion of

is
of degree at most r. It follows that

is of degree r 2(n k) for each (k). We

can then choose the test function =

(k)
(1)
sgn(,

dx

in (23) to conclude that

vanishes.
References
1. Scot Adams and Victor Reiner, private communication.
2. Douglas N. Arnold and Gerard Awanou, The serendipity family of nite elements, Foundations of Com-
putational Mathematics 11 (2011), no. 3, 337344.
3. Douglas N. Arnold, Daniele Bo, and Francesca Bonizzoni, Approximation by tensor product dierential
forms, in preparation.
4. Douglas N. Arnold, Richard S. Falk, and Ragnar Winther, Finite element exterior calculus, homological
techniques, and applications, Acta Numerica 15 (2006), 1155.
5. , Finite element exterior calculus: from Hodge theory to numerical stability, Bulletin of the
American Mathematical Society 42 (2010), no. 2, 281354.
6. Franco Brezzi, Jim Douglas Jr., Ricardo G. Duran, and Michel Fortin, Mixed nite elements for second
order elliptic problems in three variables, Numerische Mathematik 51 (1987), no. 2, 237250.
7. Franco Brezzi, Jim Douglas Jr., and L. Donatella Marini, Two families of mixed nite elements for
second order elliptic problems, Numerische Mathematik 47 (1985), no. 2, 217235.
8. Jean-Claude Nedelec, Mixed nite elements in R
3
, Numerische Mathematik 35 (1980), 315241.
9. , A new family of mixed nite elements in R
3
, Numerische Mathematik 50 (1986), no. 1, 5781.
10. Pierre-Arnaud Raviart and Jean-Marie Thomas, A mixed nite element method for 2-nd order elliptic
problems, Mathematical aspects of nite element methods (Proc. Conf., Consiglio Naz. delle Ricerche
(C.N.R.), Rome, 1975), Vol. 606 of Lecture Notes in Mathematics (Berlin), Springer, 1977, pp. 292315.
20 DOUGLAS N. ARNOLD AND GERARD AWANOU
Department of Mathematics, University of Minnesota, Minneapolis, Minnesota 55455
E-mail address: arnold@umn.edu
URL: http://umn.edu/~arnold
Department of Mathematical Sciences, Northern Illinois University, Dekalb, IL, 60115
E-mail address: awanou@math.niu.edu
URL: http://www.math.niu.edu/~awanou