Você está na página 1de 8

Comput. Methods Appl. Mech. Engrg.

253 (2013) 455462

Contents lists available at SciVerse ScienceDirect

Comput. Methods Appl. Mech. Engrg.


journal homepage: www.elsevier.com/locate/cma

Virtual Element Methods for plate bending problems


Franco Brezzi a,b,c, L. Donatella Marini d,
a

Istituto Universitario di Studi Superiori, Pavia, Italy


King Abdulaziz University, Dept. of Math., POBox 80203, Jeddah 21589, Saudi Arabia
c
IMATI-CNR, Via Ferrata 1, 27100 Pavia, Italy
d
Dipartimento di Matematica, Universit di Pavia, Via Ferrata 1, 27100 Pavia, Italy
b

a r t i c l e

i n f o

Article history:
Received 19 June 2012
Accepted 18 September 2012
Available online 27 September 2012
Keywords:
High-order MFD
Plate bending problems
Virtual Elements

a b s t r a c t
We discuss the application of Virtual Elements to linear plate bending problems, in the KirchhoffLove formulation. As we shall see, in the Virtual Element environment the treatment of the C 1 -continuity condition is much easier than for traditional Finite Elements. The main difference consists in the fact that
traditional Finite Elements, for every element K and for every given set of degrees of freedom, require
the use of a space of polynomials (or piecewise polynomials for composite elements) for which the given
set of degrees of freedom is unisolvent. For Virtual Elements instead we only need unisolvence for a space
of smooth functions that contains a subset made of polynomials (whose degree determines the accuracy).
As we shall see the non-polynomial part of our local spaces does not need to be known in detail, and
therefore the construction of the local stiffness matrix is simple, and can be done for much more general
geometries.
2012 Elsevier B.V. All rights reserved.

1. Introduction
The Finite Element literature of the last 50 years contains a big
variety of H1 -conforming (in practice, C 0 ) elements of various
degrees, with different features. On the other hand, the list of available H2 -conforming (in practice, C 1 ) elements is much more slim.
Among the most commonly used are the composite Hsieh
CloughTocher element (with cubic accuracy) and its reduced version (with quadratic accuracy), the Argyris element (with quintic
accuracy) and its reduced Bell version (with quartic accuracy). But
a much bigger effort has been devoted to alternative formulations
that could avoid the use of C 1 nite elements, including mixed formulations (essentially, based on the HellingerReissner principle)
or the use of the ReissnerMindlin model for thin plates and shells
(instead of moderately thick). Dual hybrid elements of Pian and coauthors (see e.g. [32,33]) could be seen as intermediate C 1 formulations, as the displacements are indeed C 1 but they are dened only
at the interelement boundaries (see also [11] and in particular [17]).
And several among the most common and useful applications of
nonconforming elements (as for instance the Morley element) can
indeed be seen as an effort to bypass C 1 continuity.

Corresponding author at: Dipartimento di Matematica, Universit di Pavia, Via


Ferrata 1, 27100 Pavia, Italy. Tel.: +39 0382 985638; fax: +39 0382 985602.
E-mail addresses: brezzi@imati.cnr.it (F. Brezzi), marini@imati.cnr.it
(L.D. Marini).
0045-7825/$ - see front matter 2012 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.cma.2012.09.012

We refer for instance to the books [1,24,6,39] (or, for more


mathematically oriented ones, to [18,10,9]) and to the references
therein.
In this paper we want to show that the newly born Virtual Element
method [4] is able to tackle the construction of C 1 -approximations in
a much swifter way, and on very general polygonal elements, and we
propose a few examples of elements that could surely be interesting
in the approximation of the solution of plate bending problems in the
KirchhoffLove formulation.
Virtual Element methods could be seen as an evolution of Mimetic Finite Differences and related methods (see e.g.
[23,25,28,29], or the more recent [8,19,20,37]). In particular they
are strictly related to the more theoretical versions of MFD
[26,14,16,15,12], and specially to their last versions concerning
higher order methods [31,22,3,2].
As we shall see, the basic idea consists in choosing rst some degrees of freedom at the interelement boundaries that could identify in a unique way the traces of globally C 1 functions that are
polynomials of degree 6 r on each edge, with normal derivative
of degree 6 s on each edge. Then we add a suitable amount of
internal degrees of freedom, and we dene the discrete subspace
inside the elements by means of a local plate bending problem.
Needless to say, the actual solution of these local problems will
not be required, not even in an approximate way.
Actually, we will show that, having constructed the discrete
subspace V h as above, if the boundary and internal degrees of
freedom are chosen properly, then

456

F. Brezzi, L.D. Marini / Comput. Methods Appl. Mech. Engrg. 253 (2013) 455462

(a) in each element K, the restriction of the discrete space to K


contains the space Pk of polynomials of degree 6 k, where k
depends on r and s,
(b) for every polynomial pk 2 Pk K, and for every non-polynomial element v h of V h , the contribution of K to the energy
bilinear form apk ; v h can be computed exactly, using only
the degrees of freedom of v h (and not the fact that v h solves
a local plate problem),
(c) we can then easily complete the computation of the energy
bilinear form av h ; wh (when neither v h nor wh is a polynomial in Pk ) in an almost arbitrary way, in order to ensure
stability.
Note that, if we have (a)(c) we can ensure a sort of kth order
patch test, meaning that, on any patch of elements: if the true
solution is a global polynomial of degree 6 k, then our discrete
solution will coincide exactly with the true solution.
For other attempts to deal with elements of a general shape we
refer to [5,7,21,30,3436,38] and the references therein. We point
out, however, that here we do not use numerical integration.
An outline of the paper is as follows. In Section 2 we recall the
continuous problem and x some notation. In Section 3 we present
in an abstract framework the Virtual Element approach, state the
basic assumptions and prove a convergence theorem. In Section 4
we explicitly show how to construct a C 1 VEM-approximation of
the plate bending problem with the optimal convergence rate,
namely, order k, with k P 2 whenever the discrete space V h contains locally the polynomials of degree k.
Throughout the paper we shall use the common notation for the
Sobolev spaces Hm D for m a nonnegative integer and D an open
bounded domain. In particular (see e.g. [27,18]) the L2 D scalar
product and norm will be indicated by ; 0;D or ; D and k  k0;D
or k  kD , respectively. Moreover, for m a nonnegative integer, the
mth seminorm of the function u will be dened by

j2m;D

ju

2


jaj
X
@ u
:
 a1 a2 
@ x1 @ x2 
jajm

1:1

0;D



Z
Z
aw; v D 1  m w=ij v =ij dx m DwDv dx :
X

In (2.4) v =i @ v =@xi ; i 1; 2, and we used the summation convention of repeated indices. Thus, for instance:

w=ij v =ij w=11 v =11 2w=12 v =12 w=22 v =22 :

9 M > 0 such that au; v 6 MkukV kv kV


9 a > 0 such that av ; v P akv k2V

DD2 w f

in X;

2:1

where D Et3 =121  m2 is the bending rigidity, t the thickness, E


the Young modulus, and m the Poissons ratio.
Assuming for instance the plate to be clamped all over the
boundary, Eq. (2.1) is supplemented with the boundary conditions

@w
w
0 on C:
@n

2:2

The variational formulation of 2.1,2.2 is:

Find w 2 V : H20 X solution of


aw; v f ; v 8v 2

H20

X;

2:3

where, as we said, ;  is the usual scalar product in L2 X, and the


energy bilinear form a;  is given by

u; v 2 V;

v 2 V:

2:6
2:7

Hence, (2.3) has a unique solution, and (see, e.g. [27])

kwkV 6 Ckf kL2 X :

2:8

Before going to the discretization of (2.3) let us x some notation


that will be used later on. On a domain D  R2 , with boundary @D,
we denote by n n1 ; n2 the outward unit normal vector to @D,
and by t t 1 ; t2 the unit tangent vector in the counterclockwise
ordering of the boundary. For v 2 H2 X, let M M ij v i; j 1; 2 be
the moment tensor given by the stressstrain relation

3
2
32
3
v =11
1 m 0
M 11
6
7
6
76
7
4 M 22 5 D4 m 1 0
54 v =22 5;
v =12
0 0 1m
M 12

2:9

and let Mn Mij nj be its normal vector. We then denote by


Mnn v : Mij ni nj the normal bending moment, by Mnt v : M ij ni t j
the twisting moment, and by Q n v : Mij=i nj the normal shear force.
Always for a generic domain D we also set, as in (2.4)



Z
Z
aD w; v D 1  m w=ij v =ij dx m DwDv dx ;
D

2:10

keeping the simpler notation aw; v only when D X.


After integrating by parts twice we have

where for the nonnegative multi-index a a1 ; a2 we denoted as


usual jaj a1 a2 .
When D  X the subscript D will often be omitted.

Let X  R2 be a convex polygonal domain occupied by the plate,


let C be its boundary, and let f 2 L2 X be a transversal load acting
on the plate. The KirchhoffLove model for thin plates (see e.g.
[18]) corresponds to look for the transversal displacement w solution of

2:5

Setting kv kV : jv j2;X , it is easy to see that, thanks to the boundary


conditions in V and to the Poincar inequality, this is indeed a norm
on V. Moreover

aD w; v

2. The continuous problem

2:4

Z
@v
DD2 w v dx
M nn w
dt
@n
D
@D


Z
@Mnt w
Q n w
v dt:

@t
@D

2:11

When D X and v 2 V the last two terms in (2.11) disappear, due


to the boundary conditions (2.2), but (2.11) will be useful later on
when dealing with the discrete problem (and, in any case, whenever
boundary conditions of a different type are considered on some part
of the boundary). In the sequel we shall often use v =n and v =t for
@ v =@n and @ v =@t, respectively.
3. The discrete problem. Abstract framework
Let fT h gh be a sequence of decompositions of X into elements K,
and let E h be the set of edges e of T h . We want to construct a nite
dimensional space V h  H20 X, a bilinear form ah ;  : V h  V h
! R, and an element fh 2 V 0h such that the discrete problem:

Find wh 2 V h such that


ah wh ; v h hfh ; v h i 8 v h 2 V h ;

3:1

has a unique solution wh , and good approximation properties hold.


Namely, if k P 2 is the target degree of accuracy, and the solution w
of (2.3) is smooth enough, we want to have

kw  wh kV 6 C h

k1

jwjk1;X ;

3:2

where C, here and in the following formulae, is, as usual, a positive


constant independent of h. Note that (3.2), as written, implies our
k  th order Patch-Test, since when w 2 Pk then jwjk1 0.

457

F. Brezzi, L.D. Marini / Comput. Methods Appl. Mech. Engrg. 253 (2013) 455462

Let us rst recall the basic assumptions that we need (see e.g.
[12]).
H0 - There exists an integer N and a positive real number c such
that for every h and for every K 2 T h :
 the number of edges of K is 6 N,
 the ratio between the shortest edge and the diameter hK of
K is bigger than c, and
 K is star-shaped with respect to every point of a ball of
radius chK . h

ah u; v 6 a M kukV kv kV

for all u and

in V h :

3:11

We have the following convergence theorem.


Theorem 3.1. Under the Assumptions H1H2 the discrete problem:
Find wh 2 V h such that

ah wh ; v h hfh ; v h i 8 v h 2 V h ;

3:12

has a unique solution wh . Moreover, for every approximation wI of w in


V h and for every approximation wp of w that is piecewise in Pk , we
have


kw  wh kV 6 C kw  wI kV kw  wp kh;V kf  fh kV 0
h

H1 We assume that we are given, for each h:

where C is a constant depending only on a; a ; a ; M and, with the usual


notation, the norm in V 0h is dened as


 a space V h  V; for each element K we will denote

V Kh restriction of V h to K

3:3

 a symmetric bilinear form ah from V h  V h to R that can be


split as

ah uh ; v h

X
aKh uh ; v h 8 uh ; v h 2 V h ;

3:4

hf  fh ; v h i
:
v h 2V h kv h kV

kf  fh kV 0 : sup
h

3:13

Proof. Existence and uniqueness of the solution of (3.12) is a consequence of (3.8) and (2.7). Next, setting

dh : wh  wI

where
is a bilinear symmetric form on V Kh  V Kh ;
 an element fh 2 V 0h . h
aKh

we have

Similarly, we split the bilinear form a;  and the norm k  kV as

au; v

X
aK u; v 8 u; v 2 V;
K

kv kV

X 2
jv jV;K

a akdh k2V 6 use 2:7 and 3:10


6 a adh ; dh 6 ah dh ; dh use 3:14
ah wh ; dh  ah wI ; dh use 3:12 and 3:4
X
hfh ; dh i 
aKh wI ; dh use wp

!1=2

8v 2 V:

3:5

hfh ; dh i 

Since in what follows we shall also deal with functions in


Q
H2 T h : K H2 K, we also need to dene a broken H2 semi-norm:

jv jh;V :

X 2
jv jV;K

3:14

!1=2

8v 2

hfh ; dh i 
hfh ; dh i 

X
K
X

K
X



aKh wI  wp ; dh aKh wp ; dh use 3:7

aKh wI  wp ; dh aK wp ; dh use w and 3:5

aKh wI  wp ; dh aK wp  w; dh  aw; dh

H2 K:

3:15

3:6

Now we can use (2.3) in (3.15) to get

a akdh k2V

3.1. An abstract convergence theorem

6 hfh ; dh i 
Together with H0 and H1 we further assume the following
properties.
H2 There exists an integer k P 2 (that will determine our order of accuracy) such that for all h, and for all K in T h ,

X


aKh wI  wp ; dh aK wp  w; dh  f ; dh

hfh ; dh i  f ; dh 

X

aKh wI  wp ; dh aK wp  w; dh :
K

3:16
K

Finally we use (3.13), (3.9), and the continuity of each a in (3.16) to


obtain

 k-Consistency: 8p 2 Pk ; 8v h 2 V h

aKh p; v h aK p; v h :

3:7

 Stability: 9 two positive constants a and a , independent of h


and of K, such that

8v h 2 V h a aK v h ; v h 6 aKh v h ; v h 6 a aK v h ; v h : 


kdh k2V 6 C kf  fh kV 0 kwI  wp kh;V kw  wp kh;V kdh kV
h

3:17

for some constant C depending only on a,a ; a , and M. Then the result follows easily by the triangle inequality. h

3:8
K

We note that the symmetry of ah , (3.8) and the continuity of a easily imply the continuity of ah with

aKh u; v 6 aKh u; u

1=2 

1=2

1=2  K
1=2
aKh v ; v
6 a aK u; u
a v ; v

6 a M kukV;K kv kV;K

for all u and

in V h :
3:9

4. Construction of V h ; ah , and fh
We shall construct now a space V h , a bilinear form ah , and a
right-hand side fh satisfying assumptions H1H2. Our family of elements will depend on three integer indices r; s; m, related to the
degree of accuracy k P 2 by:

r maxf3; kg;

s k  1;

m k  4:

4:1

In turn, (3.8) and (3.9) easily imply

8v 2 V h a av ; v 6 ah v ; v 6 a av ; v :
and

3:10

Remark 4.1. As we shall see in a while, the indices r and s will be


related, respectively, to the polynomial degree of the functions in
V h , and to the polynomial degree of their normal derivative, on

458

F. Brezzi, L.D. Marini / Comput. Methods Appl. Mech. Engrg. 253 (2013) 455462

each edge of the decomposition. On the other hand, the index m


will be related (in a way to be made precise) to degrees of
freedominternal to each element. h
Remark 4.2. As it will be clear when constructing the spaces, condition (4.1) will not allow, for instance, to take the simple choice

r 5; s 3; m 0 and k 4

4:2

(that might produce a Bell-like type of element). What we really


need is actually

r P maxf3; kg;

s P k  1;

m P k  4:

4:3

On the other hand, the present choice allows to consider a family of


discretizations that depends only on one parameter (here, k). Choosing (4.3), instead, we should actually deal with a family of spaces V h
depending on four parameters r; s; m, and k, with a considerable
complication in the notation and little conceptual gain. However,
in developing our theory we should keep an eye on more general
cases, like (4.3) or other possible approaches for which our theory
would apply almost unchanged (but whose comprehensive treatment would imply a much more complicated notation). h

Fig. 1 shows the rst two elements of the family determined by


the d.o.f. (4.7)(4.11): for both elements we have that v 2 P3 e on
v 2 P e for k 2, and @ v 2 P e for k 3. These
each edge, but @@nje
1
2
@nje
elements are the extension to polygonal domains of the HsiehClough-Tocher triangle (k 3) and of its reduced version (k 2).
We note that the degrees of freedom (4.7) and (4.8) are always
needed to ensure C 1 -continuity at the vertices. Moreover, on each
edge, they provide v and v =t at the endpoints of the edge: enough
information to identify uniquely a polynomial of degree 6 3. This,
together with (4.9), explains the requirement r P 3. Indeed, to
identify a polynomial of degree r, together with (4.7) and (4.8)
we need r  3 additional information (as for instance (4.9)). At
the same time, conditions (4.8) provide, for every edge, v =n at the
endpoints of the edge, thus allowing to identify uniquely a polynomial of degree 6 1. For s > 1, to identify a polynomial of degree 6 s,
we would need the additional s  1 conditions (4.10). On the other
hand, the d.o.f. (4.11) are equivalent to prescribe PKm v in K, where,
for m a nonnegative integer,

PKm v is the L2 K  projector operator onto the space Pm K: 4:12


We summarize the above discussion in the following Proposition.

4.1. Local construction of V h

Proposition 4.1. In each element K the d.o.f. (4.7)(4.9) uniquely


determine a polynomial of degree 6 r on each edge of K, the degrees of
freedom (4.8) and (4.10) uniquely determine a polynomial of degree
6 s on each edge of K, and the d.o.f. (4.11) are equivalent to prescribe
PKm v in K. h
Let be the number of edges (and hence of vertices) of K, and let
GK be the set of degrees of freedom corresponding to K, whose total
number amounts to

Let K be a generic polygon in T h , let k P 2, and let r; s; m be given by (4.1). We dene V Kh (that will be the local version of our discretized space V h ) as

T : T @ T 0 ; with T @ f3 r  3 s  1g  r s  1;


m 1m 2
m2 3m 2

and T 0
:
4:13
2
2

Remark 4.3. We can easily attach to each vertex n a characteristic


length hn , taken, for instance, as the average of the diameters of the
elements having n as a vertex. To an edge e we can attach, as characteristic length he the length jej of the edge itself, while to an element K we can attach as characteristic length hK its diameter. h

V Kh : fv 2 H2 K s:t: D2 v 2 Pm K; v je 2 Pr e; v =n je
2 P s e; 8e 2 @Kg:

4:4

We shall make use of the following notation: for t a nonnegative


integer, and e an edge with midpoint xe , we denote by Met the set
of t 1 normalized monomials

Met :


2

t )
x  xe x  xe
x  xe
:
1;
;
;...;
he
he
he

4:5

Similarly, for a two-dimensional domain K with diameter hK and


barycenter xK we denote by MKt the set of t 1t 2=2 normalized monomials

(
b
x  xK
K
Mt :
;
hK

jbj 6 t ;

4:6

where, as usual, for a nonnegative multiindex b b1 ; b2 we set


jbj b1 b2 and xb xb11 xb22 . In K we dene the following degrees
of freedom:

 The value of v n 8 vertex n:

Now we have to show how to associate to the above degrees of freedom a function in K. It is important to point out from the very
beginning that we do not need to be able to compute such a function, but just to show that it exists and it has certain properties.
Proposition 4.2. The degrees of freedom (4.7)(4.11) are unisolvent
in V Kh (as dened in (4.4)).
Proof. According to Proposition 4.1, to prove the proposition it is
enough to show that, for each K 2 T h , a function u in V Kh such that

u 0;

@u
0 on @K
@n

4:14

and

PKm u 0 in K

4:15

4:7

4:8
 The values of hn v =1 n and hn v =2 n 8 vertex n:
Z
1
e
 For r > 3; the moments
qnv ndn 8q 2 Mr4 ; 8 edge e: 4:9
h
Ze e
 For s > 1; the moments qnv =n ndn 8q 2 Mes2 8 edge e: 4:10
e
Z
1
4:11
 For m P 0; the moments 2 qxv xdx 8q 2 MKm ::
hK K
Note that all the quantities ((4.7)(4.11)) have the same dimension,
and hence in a homothetic blow up of K would scale in the same
manner.

Fig. 1. Local d.o.f. for the lowest-order element: k 2 (left), and next to the lowest
k 3 (right).

F. Brezzi, L.D. Marini / Comput. Methods Appl. Mech. Engrg. 253 (2013) 455462

is actually identically zero in K. To this end, we show that D2 u 0 in


K, which joined with (4.14) gives u  0. We rst solve, for every
q 2 Pm K, the following auxiliary problem: Find w wq 2 H20 K
such that:

aK w; v q; v 0;K

8v 2 H20 K;

4:16

that could also be written as

DD2 w q in K;

wj@K

@w
0;
@n j@K

or else w DD20;K 1 q:

4:17

Next, we consider the map R, from Pm K into itself, dened by

Rq : PKm DD20;K 1 q  PKm wq 8q 2 Pm K:

4:18

We claim that the operator R dened by (4.18) is an isomorphism


from Pm K into itself. Indeed, from (4.18) and the denition of P Km
we have, for every q 2 Pm K:

Rq; q0;K PKm DD20;K 1 q; q0;K PKm wq; q0;K wq; q0;K

459

The dimension of V h , that is, the total number of degrees of freedom


in V h is then given by

G 3NV NE r  3 s  1 NK T 0 ;

4:28

where N V is the number of internal vertices of T h ; N E the number of


internal edges, N K the number of elements, and T 0 is still given by
(4.13). Needless to say, in the construction of v h 2 V h one should
take into account the values of v h and v h=n at the boundary as well,
setting them equal to zero.
Remark 4.5. At this point we can order (in a rather arbitrary way)
the degrees of freedom of V h

g1 ; g2 ; . . . ; gG :

4:29

Accordingly we will have a characteristic length hi attached to each


degree of freedom g i , depending on its nature, according to Remark
4.3. It is obvious that, in the computer code, we will have chosen,
once and for all, an orientation for the normal unit vector, say ne ,
to each internal edge e. h

aK wq; wq:
Since w 2 H20 K we have then that

fRq 0g ) faK wq;wq 0g ) fwq 0g ) fq 0g:

4:19

We note that, if u 2 V Kh is such that u @u=@n 0 on @K, then

PKm u PKm DD20;K 1 DD2 u RDD2 u:

4:20

Hence, if P Km u 0 as in (4.15), we can say that

PKm u 0 ) RDD2 v 0 ) DD2 v 0;


and the proof is concluded.

x21

1
3

x31 n1

4:21

and so on. h
4.2. Global construction of V h
For every decomposition T h , and for every k P 2 we are now
ready to dene the space V h as

D2 v jK 2 Pm K8K 2 T h g;

4:30

Moreover, by the usual BrambleHilbert/DenyLions technique (see


e.g. [10]) and using a scaling argument to get around the variability
in the geometry (see e.g. [13]) it is not difcult to see that, for a and
b nonegative indices, we can prove that

kw  wI ka;X 6 C h

@K

V h fv 2 V : v je 2 Pr e;

g i w  wI 0 8i 1; 2; . . . ; G:

ba

jwjb;X

a 0; 1; 2; a 6 b 6 k 1

4:31

Remark 4.4. We point out once more that we will not try to compute the solutions of the local bi-harmonic problems involved in
the denition of (4.4). On the other hand we will always assume
that we can compute polynomials on @K, and hence also integral
of polynomials over K. Indeed, for instance,

Remark 4.6. It follows easily from the above construction that for
every smooth enough w there exists a unique element wI 2 V h such
that

@v
2 Ps e8e 2 E h ;
@n je
4:22

where r; s, and m are always given by (4.1) (although, actually, they


could be any triple that satises (4.3)). It follows from the above
discussion that the global degrees of freedom in V h could then be
taken as
The value of v n 8 internal vertex n
4:23
4:24
The values of hn v =1 n and hn v =2 n 8 internal vertex n
Z
1
e
qnv ndn 8q 2 Mr4 ; 8 internal edge e4:25
For r > 3; the moments
h
Ze e
qnv =n ndn 8q 2 Mes2 8 internal edge e 4:26
For s > 1; the moments
e
Z
1
qxv xdx 8q 2 MKm 8 element K:
4:27
For m P 0; the moments 2
hK K

(with a constant C independent of h) as in the usual Finite Element


framework. h
We point out that (as it happens also for the classical nite elements), different choices of degrees of freedom might be used for
the same space. However a different choice of the local degrees of
freedom can induce a different choice of the global degrees of freedom, that in turn could alter the continuity property of the resulting
functions, giving rise to a different global space. Let us see a couple of
examples that, in our opinion, are particularly meaningful. Let us
consider the value k 5 (and, according to (4.1), r; s; m
5; 4; 1), and the corresponding degrees of freedom in our construction (4.7)(4.11). They are: the values of w; wx ; wy at the vertices, plus, on every edge, the moments up to the order two of w
and of wn , plus the moments up to the order one of w inside the
element. Taking instead k 4 with r 5; s 3, and m 0 (using
this time (4.3)) we would have the same degrees of freedom minus
the moments of order one of wn on each edge. The two cases are
presented in Fig. 2, with obvious meaning of the symbols.
Note that, however, we could as well have used as local degrees
of freedom the ones in Fig. 3, where (with classical notation) the
circles at the vertices represent the six degrees of freedom values
of w; wx ; wy ; wxx ; wxy ; wyy at the vertices.
The theoretical treatment of the case of Fig. 3 could be done in a
way practically identical to the one followed here. As already
pointed out in Remark 4.2 we chose to skip the comprehensive
treatment of all possible cases allowed by our theory, in order to
avoid a too cumbersome notation.
4.3. Construction of ah
We are left to show how to construct a (computable!) symmetric bilinear form ah satisfying (3.7) and (3.8). We shall work on a

460

F. Brezzi, L.D. Marini / Comput. Methods Appl. Mech. Engrg. 253 (2013) 455462

generic element K 2 T h , and we recall that we denoted by V Kh the


restriction of V h to K. First of all, we observe that the local degrees
of freedom allow us to compute exactly aK p; v for any p 2 Pk K
and for any v 2 V Kh . Indeed, recalling (2.11), we have

D2 p v dx

 
Z 
@v
@M nt p
M nn p
v dt:
Q n p

@t
@n
@K

aK p; v D

4:32

We note then that D2 p belongs to Pk4 K; for m P k  4, the integral over K that appears in (4.32) is then computable using only the
values of the internal degrees of freedom of v, without actually
knowing v. On the other hand, on each edge of K both M nn p (which
belongs to Pk2 e) and Q n p @Mnt p=@t (which belongs to
Pk3 e) are polynomials, as well as v and v =n . From the degrees of
freedom on the boundary we can easily compute both v and v =n ,
and in conclusion all the terms in the right-hand side of (4.32)
can be easily computed without knowing v inside.
b as the
Now for every u 2 C 0 K we dene its quasi-average u
constant function (over K) whose value

1X
b :
u
uxi
i1

4:33

is given by the average of the values that u assumes at the vertices


ni , (i 1; 2; . . . ; ) of K. Next, we introduce the operator
PKk : V Kh ! Pk K  V Kh dened as the solution of

aK PKk w; q aK w; q 8w 2 V Kh ; 8q 2 Pk K
d
K
d
b
P
rd
PKk w r
w:
k w w;

4:34

We note that for v 2 Pk K the rst equation in (4.34) implies


PKk v =ij v =ij for i; j 1; 2, that joined with the second equation
gives easily

PKk v v 8v 2 Pk K:

4:35

Choosing aKh u; v aK PKk u; PKk v would now ensure property (3.7),


but not, in general, property (3.8). Hence we need to add a suitable
term capable to ensure (3.8). Let then SK u; v be a symmetric positive denite bilinear form, to be chosen to verify

c0 aK v ; v 6 SK v ; v 6 c1 aK v ; v ; 8v 2 V Kh with PKk v 0;

4:36

for some positive constants c0 ; c1 independent of K and hK . Then set

aKh u; v : aK PKk u; PKk v SK u  PKk u; v  PKk v :

4:37

Proposition 4.3. The bilinear form (4.37) constructed with the above
procedure satises both the consistency property (3.7) and the
stability property (3.8).
Proof. Property (3.7) follows immediately from (4.35) and (4.34):
for p 2 Pk K, (4.35) implies SK p  PKk p; v  PKk v 0 8v , and
hence

Fig. 3. Alternative d.o.f. for the elements of Fig. 2. We have an Argyris-like


element on the left and a Bell-like element on the right.

aKh p; v aK PKk p; PKk v aK PKk p; v aK p; v :


K

4:38

K
k w; q

Moreover, from (4.34) we have a w  P


0 for all q 2 Pk K
and for all w 2 V Kh , so that by Pythagoras theorem we have

aK v  PKk v ; v  PKk v aK PKk v ; PKk v aK v ; v 8v 2 V Kh :

4:39

Property (3.8) follows then easily from (4.36) and (4.39). Indeed
setting a : maxf1; c1 g we have

aKh v ; v 6 aK PKk v ; PKk v c1 aK v  PKk v ; v  PKk v


6 maxf1; c1 g aK PKk v ; PKk v aK v  PKk v ; v  PKk v
a aK v ; v ;
4:40
and similarly, setting a : minf1; c0 g:


aKh v ; v P minf1; c0 g aK PKk v ; PKk v aK v  PKk v ; v  PKk v
a aK v ; v : 
4:41
4.4. Choice of SK
In general, the choice of the bilinear form SK might depend on the
problem and on the degrees of freedom. From (4.36) it is clear that
SK must scale like aK ;  on the kernel of PKk . Since in (4.23)(4.27)
we took care to have degrees of freedom of the same dimension,
then all the elements of the canonical basis u1 ; . . . ; uT dened by

g i uj dij ;

i; j 1; T;

4:42

will scale in the same way (and in particular as the ones associated
to point value degrees of freedom). Hence the choice

SK v ; w D

T
X
g i v g i whi 2 ;
i1

(with hi as in Remark 4.5) would clearly do the job.


Remark 4.7. Clearly, is we decide to use degrees of freedom
involving second derivatives, as the ones in Fig. 3, we should scale
2
them by hn in order to apply the above argument h

4.5. Construction of the right-hand side


In order to build the loading term hfh ; v h i for v h 2 V h in a simple
and easy way it is convenient to have internal degrees of freedom
in V h , and this means, according to (4.1) and (4.11), that we need
k P 4. In this case we dene fh on each element K as the L2 K
projection of the load f onto the space of piecewise polynomials
of degree m k  4, that is,
Fig. 2. Local d.o.f. for the (5,4,1) element with k 5 (left), and for the (5,3,0)
element with k 4 (right).

fh PKk4 f

on each K 2 T h :

Then, always for k P 4, the associated loading term

F. Brezzi, L.D. Marini / Comput. Methods Appl. Mech. Engrg. 253 (2013) 455462

hfh ; v h i

XZ

fh v h dx 

K2T h

XZ
K

K2T h

XZ
K2T h

Acknowledgements

PKk4 f v h dx

f PKk4 v h dx

can be exactly computed using the degrees of freedom for V h that


represent the internal moments, see (4.11).
For k > 4, that is, m k  4 P 1, standard L2 orthogonality and
approximation estimates on star-shaped domains yield

hfh ; v h i  f ; v h

XZ

6C

P Kk4 f  f v h  PK1 v h dx

k3
hK jf jk3;K

2
hK k h k2;K

K2T h

6 Ch

k1

X

jf j2k3;K

1=2

kv h kV ;

4:43

K2T h

and thus, recalling (3.13),

kf  fh kV 0 6 Ch

k1

!1=2
jf j2k3;K

4:44

K2T h
k1

Hence, for k > 4 Theorem 3.1 ensures the optimal Oh error bound.
For k 6 4 we need some additional tricks. In view of Remark 4.4 we
assume that we are able to compute exactly the integral

Z
K

p1 q1 dx

4:45

for every pair p1 and q1 of polynomials of degree 6 1 on K. In each


case k 2; 3; 4 the term hfh ; v h i will then be dened by

hfh ; v h iK

Z
K

bk v h
ek f P
P
1
1

4:46

e k and P
b k are projectors on the space of polynomials of dewhere P
1
1
gree 6 1 that will change from one value of k to another.
For k 2 and for k 3 we take

e k f : PK f ;
fhjK  P
k2
1

bk v h  v
d
e h : v
ch k  2x  xK  r
P
vh
1

on each K:

4:47

e h 2 Pk2 K.
Note that for both values k 2 and k 3 we have that v
Hence on each K we have

fh K  f ; v h K
hfh ; v h iK  f ; v h K PKk2 f ; v
fh K f ; v
fh  v h K
PKk2 f  f ; v
k1

fh  v h K 6 ChK kf k0;K kv h kk1;K :


f ; v

4:48

For k 4 we take instead

e k f : PK f ;
fhjK  P
1
1

bk v h  v
e h : PK0 v h x  xK
P
1

d
r
vh

This paper was funded by King Abdulaziz University, under


grant No. (38-3-1432/HiCi). The authors, therefore, acknowledge
technical and nancial support of KAU. The authors acknowledge
as well support of MIUR (Italian Ministery of University and
Research), under project PRIN2008. The rst author also thanks
Ahmed Alsaedi and Bashir Ahmad for the fruitful conversations.
References

K2T h

461

on each K:

4:49

We have on each K

e h K  f ; v h K
hfh ; v h iK  f ; v h K PK1 f ; v
e h K f ; v
e h  v h K
PK1 f  f ; v
e h  v h K 6 Ch3K jf j1;K kv h k2;K
f  PK0 f ; v

4:50

where, going from second to third line we took advantage of the fact
e h  v h has zero mean value over K. Note that, for k 4, we
that v
have m 0 in (4.1), and from (4.11) we have the right to use P K0 v h
instead of v^ h , since P K0 v h is now part of the degrees of freedom.

[1] K.J. Bathe, Finite Element procedures, Prentice Hall, 1996.


[2] L. Beiro da Veiga, G. Manzini, A higher-order formulation of the mimetic nite
difference method, SIAM J. Sci. Comput. 31 (1) (2008) 732760.
[3] L. Beiro da Veiga, K. Lipnikov, G. Manzini, Convergence analysis of the highorder mimetic nite difference method, Numer. Math. 113 (3) (2009) 325356.
[4] L. Beiro da Veiga, F. Brezzi, A. Cangiani, G. Manzini, L.D Marini, A. Russo, Basic
Principles of Virtual Element Methods to appear in Math, Models Methods
Appl. Sci. 2013. http://dx.doi.org/10.1142/S0218202512500492.
[5] T. Belytschko, Y.Y. Lu, L. Gu, Element-free Galerkin methods, Int. J. Numer.
Meth. Engng. 37 (1994) 229256.
[6] T. Belytschko, W.K. Liu, B. Moran, Nonlinear Finite Elements for Continua and
Structures, J. Wiley, 2000.
[7] D.J. Benson, Y. Bazilevs, E. De Luycker, M.-C. Hsu, M. Scott, T.J.R. Hughes, T.
Belytschko, A generalized nite element formulation for arbitrary basis
functions: From isogeometric analysis to XFEM, Int. J. Numer. Meth. Engng
83 (6) (2010) 765785.
[8] P. Bochev, J.M. Hyman, Principle of mimetic discretizations of differential
operators, in: D.N. Arnold, P. Bochev, R. Lehoucq, R. Nicolaides, M. Shashkov
(Eds.), Compatible Discretizations. Proceedings of IMA hot topics workshop on
compatible discretizations, IMA Vol. 142. Springer-Verlag, 2006.
[9] D. Braess, Finite Elements: Theory, Fast Solvers and Applications in Solid
Mechanics, Cambridge University Press, 1997.
[10] S.C. Brenner, L.R. Scott, The Mathematical Theory of Finite Element Methods,
Texts in Applied Mathematics, vol. 15. Springer-Verlag, New York, 2008.
[11] F. Brezzi, Sur la mthode des lments nis hybrides pour le problme
biharmonique, Numer. Math. 24 (1975) 103131.
[12] F. Brezzi, A. Buffa, K. Lipnikov, Mimetic nite differences for elliptic problems,
Math. Model. Numer. Anal. 43 (2009) 277295.
[13] F. Brezzi, M. Fortin, Mixed and Hybrid Finite Element Methods, SpringerVerlag, New York, 1991.
[14] F. Brezzi, K. Lipnikov, M. Shashkov, Convergence of mimetic nite difference
method for diffusion problems on polyhedral meshes, SIAM J. Num. Anal. 43
(2005) 18721896.
[15] F. Brezzi, K. Lipnikov, M. Shashkov, V. Simoncini, A new discretization
methodology for diffusion problems on generalized polyhedral meshes,
Comp. Meth. Appl. Mech. Engrg. 196 (2007) 36823692.
[16] F. Brezzi, K. Lipnikov, V. Simoncini, A family of mimetic nite difference
methods on polygonal and polyhedral meshes, Math. Models Methods Appl.
Sci. 15 (2005) 15331553.
[17] F. Brezzi, L.D. Marini, On the numerical solution of plate bending problems by
hybrid methods, R.A.I.R.O. R-3 9 (1975) 550.
[18] P.G. Ciarlet, The Finite Element Method for Elliptic Problems, North-Holland,
1978.
[19] J. Droniou, R. Eymard, T. Gallout, R. Herbin, A unied approach to mimetic
nite difference, hybrid nite volume and mixed nite volume methods, Math.
Models Methods Appl. Sci.(M3AS) 20 (2) (2010) 265295.
[20] R. Eymard, T. Gallout, R. Herbin. Discretization of heterogeneous and
anisotropic diffusion problems on general non-conforming meshes, SUSHI: a
scheme using stabilisation and hybrid interfaces, 2009. Doi:10.1093/imanum/
drn084. Published online June 2009.
[21] T.-P. Fries, T. Belytschko, The extended/generalized nite element method: An
overview of the method and its applications, Int. J. Numer. Meth. Engng 84
(2010) 253304.
[22] V. Gyrya, K. Lipnikov, High-order mimetic nite difference method for
diffusion problems on polygonal meshes, J. Comput. Phys. 227 (2008) 8841
8854.
[23] R. Hiptmair, Discrete Hodge operators, Numer. Math. 90 (2) (2001) 265289.
[24] T.J.R. Hughes, The Finite Element Method: Linear Static and Dynamic Finite
Element Analysis, Dover Publications, 2000.
[25] J.M. Hyman, M. Shashkov, Adjoint operators for the natural discretizations of
the divergence gradient and curl on logically rectangular grids, Appl. Numer.
Math. 25 (4) (1997) 413442.
[26] Yu. Kuznetsov, S. Repin, New mixed nite element method on polygonal and
polyhedral meshes, Russ. J. Numer. Anal. Math. Modelling 18 (3) (2003) 261
278.
[27] J.-L. Lions, E. Magenes, Problmes aux limites non homognes et applications,
Vol. I, Dunod, Paris, 1968.
[28] L. Margolin, M. Shashkov, P. Smolarkiewicz, A discrete operator calculus for
nite difference approximations, Comput. Meth. Appl. Mech. Engrg. 187
(2000) 365383.

462

F. Brezzi, L.D. Marini / Comput. Methods Appl. Mech. Engrg. 253 (2013) 455462

[29] C. Matiussi, An analysis of nite volume, nite element, and nite difference
methods using some concepts of algebraic topology, J. Comput. Phys. 133
(1997) 289309.
[30] S.E. Mousavi, N. Sukumar, Numerical integration of polynomials and
discontinuous functions on irregular convex polygons and polyhedrons,
Comput. Mech. 47 (2011) 535554.
[31] J.B. Perot, V. Subramanian, Higher-order mimetic methods for unstructured
meshes, J. Comput. Phys. 219 (1) (2006) 6885.
[32] T.H.H. Pian, P. Tong, Basis of nite element methods for solid continua, Int. J.
Numer. Meth. Engng. 1 (1969) 328.
[33] T.H.H. Pian, K. Sumihara, Rational approach for assumed stress nite elements,
Int. J. Numer. Meth. Engng. 20 (1984) 16851695.
[34] I.E. Sutherland, G.W. Hodgman, Reentrant polygon clipping, Commun. ACM 17
(1) (1974) 3242.

[35] N. Sukumar, E.A. Malsch, Recent advances in the construction of polygonal


nite element interpolants, Arch. Comput. Meth. Engrg. 13 (1) (2006) 129
163.
[36] A. Tabarraei, N. Sukumar, Extended nite element method on polygonal and
quadtree meshes, Comput. Methods Appl. Mech. Engrg. 197 (5) (2008) 425
438.
[37] K.A. Trapp, Inner products in covolume and mimetic methods, Math. Model.
Numer. Anal. 42 (2008) 941959.
[38] E. Wachspress, A rational Finite Element Basis, Academic Press, New York,
1975.
[39] O.C. Zienkiewicz, R.L. Taylor, The Finite Element methods, Vol. 13,
Butterworth, Heinemann, Oxford, 2000.