Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
July 15, 2016 13:43 WSPC/103-M3AS 1650040
Mathematical Models and Methods in Applied SciencesVol. 26, No. 9 (2016) 1649–1669c© World Scientific Publishing CompanyDOI: 10.1142/S0218202516500408
Finite element approximations of symmetric tensorson simplicial grids in Rn: The lower order case
Jun Hu
LMAM, School of Mathematical Sciences, andBeijing International Center for Mathematical Research,
Peking University, Beijing 100871, P. R. [email protected]
Shangyou Zhang
Department of Mathematical Sciences, University of Delaware,Newark, DE 19716, USA
Received 6 May 2015Revised 11 March 2016Accepted 27 April 2016Published 18 July 2016
Communicated by D. Arnold
In this paper, we construct, in a unified fashion, lower order finite element subspacesof spaces of symmetric tensors with square-integrable divergence on a domain in anydimension. These subspaces are essentially the symmetric tensor finite element spacesof order k from [Finite element approximations of symmetric tensors on simplicial gridsin Rn: The higher order case, J. Comput. Math. 33 (2015) 283–296], enriched, for each
(n − 1)-dimensional simplex, by(n+1)n
2face bubble functions in the symmetric tensor
finite element space of order n + 1 from [Finite element approximations of symmetrictensors on simplicial grids in Rn: The higher order case, J. Comput. Math. 33 (2015) 283–
296] when 1 ≤ k ≤ n − 1, and by(n−1)n
2face bubble functions in the symmetric tensor
finite element space of order n + 1 from [Finite element approximations of symmetrictensors on simplicial grids in R
n: The higher order case, J. Comput. Math. 33 (2015) 283–296] when k = n. These spaces can be used to approximate the symmetric matrix field ina mixed formulation problem where the other variable is approximated by discontinuouspiecewise Pk−1 polynomials. This in particular leads to first-order mixed elements onsimplicial grids with total degrees of freedom per element 18 plus 3 in 2D, 48 plus 6 in3D. The previous record of the degrees of freedom of first-order mixed elements is, 21plus 3 in 2D, and 156 plus 6 in 3D, on simplicial grids. We also derive, in a unified waywhich is completely different from those used in [D. Arnold, G. Awanou and R. Winther,Finite elements for symmetric tensors in three dimensions, Math. Comput. 77 (2008)1229–1251; D. N. Arnold and R. Winther, Mixed finite element for elasticity, NumberMath. 92 (2002) 401–419], a family of Arnold–Winther mixed finite elements in anyspace dimension. One example in this family is the Raviart–Thomas elements in onedimension, the second example is the mixed finite elements for linear elasticity in two
1649
Mat
h. M
odel
s M
etho
ds A
ppl.
Sci.
2016
.26:
1649
-166
9. D
ownl
oade
d fr
om w
ww
.wor
ldsc
ient
ific
.com
by U
NIV
ER
SIT
Y O
F D
EL
AW
AR
E o
n 10
/17/
17. F
or p
erso
nal u
se o
nly.
July 15, 2016 13:43 WSPC/103-M3AS 1650040
1650 J. Hu & S. Zhang
dimensions due to Arnold and Winther, the third example is the mixed finite elementsfor linear elasticity in three dimensions due to Arnold, Awanou and Winther.
Keywords: Mixed finite element; symmetric finite element; first-order system; simplicialgrid; inf–sup condition.
AMS Subject Classification: 65N30, 73C02
1. Introduction
The constructions, using polynomial-shape functions, of stable pairs of finite ele-ment spaces for approximating the pair of spaces H(div, Ω; S)×L2(Ω; Rn) in first-order systems are a long-standing, challenging and open problem, see for instance,Refs. 4 and 6. For mixed finite elements of linear elasticity, many mathematicianshave been working on this problem and compromised to weakly symmetric or com-posite elements, cf. Refs. 3, 7, 8, 34, 36, 37, 39–42 and 45. It is not until 2002that Arnold and Winther were able to propose the first family of mixed finiteelement spaces with polynomial-shape functions in two dimensions.10 Such a two-dimensional family was extended to a three-dimensional family of mixed elements,6
while the lowest order element with k = 2 was first proposed in Ref. 2. We referinterested readers to Refs. 2, 5, 6, 10, 12, 17–19, 11, 24, 31, 35, 43, 44, 9, 13, 20, 25,26, 27, 30 and 29, for recent progress on mixed finite elements for linear elasticity.See Ref. 14 for the DPG method for the linear elasticity problem.
In Refs. 32 and 33, Hu and Zhang proposed new ideas to design discrete stressspaces and analyze the discrete inf–sup condition. In particular, they were able toconstruct suitable H(div, Ω; S)-Pk space, namely, Σk,h defined in (4.1) below, withk ≥ 3 for 2D, and k ≥ 4 for 3D, finite element spaces for the stress discretizationin both two and three dimensions. In Ref 28, Hu constructed, in a unified fashion,suitable H(div, Ω; S)-Pk space with k ≥ n + 1, and proposed a set of degrees offreedom for the shape function space, in any dimension.
The purpose of this paper is to extend those elements in Ref. 28 to lower ordercases where 1 ≤ k ≤ n. Since it is, at moment, very difficult to prove that the pairof the space Σk,h and the L2-Pk−1 space, namely, Vk,h defined in (4.2), is stable, theΣk,h space has to be enriched by some higher order polynomials whose divergenceare in Vk,h. Thanks to Ref. 28, it suffices to control the piecewise rigid motion space.Hence, we only need to add, for each (n − 1)-dimensional simplex, (n+1)n
2 simplexbubble functions in Σn+1,h when 2 ≤ k ≤ n−1, and (n−1)n
2 simplex bubble functionsin Σn+1,h when k = n. This in particular leads to first-order mixed elements onsimplicial grids with total degrees of freedom per element 18 plus 3 in 2D, 48 plus 6in 3D. The previous record of the degrees of freedom of first-order mixed elementsis 21 plus 3 in 2D, and 156 plus 6 in 3D, on simplicial grids. These enriched bubblefunctions belong to the lowest order space from a family of Arnold–Winther mixedfinite elements in any space dimension which, together with the Vk,h space, form astable pair of spaces for first-order systems. Note that these spaces in this family ofArnold–Winther mixed finite elements are constructed in a unified and direct way
Mat
h. M
odel
s M
etho
ds A
ppl.
Sci.
2016
.26:
1649
-166
9. D
ownl
oade
d fr
om w
ww
.wor
ldsc
ient
ific
.com
by U
NIV
ER
SIT
Y O
F D
EL
AW
AR
E o
n 10
/17/
17. F
or p
erso
nal u
se o
nly.
July 15, 2016 13:43 WSPC/103-M3AS 1650040
Finite element approximations of symmetric tensors 1651
which is completely different from those used in Refs. 6 and 10. One example inthis family is the Raviart–Thomas elements in one dimension, the second exampleis the mixed finite elements for linear elasticity in two dimensions due to Arnoldand Winther,10 the third example is the mixed finite elements for linear elasticityin three dimensions due to Arnold, Awanou and Winther.6
We end this section by introducing first-order systems and related notations.We consider mixed finite element methods22 of first-order systems with symmetrictensors: find (σ, u) ∈ Σ × V := H(div, Ω; S) × L2(Ω; Rn), such that
(Aσ, τ) + (div τ, u) = 0 for all τ ∈ Σ,
(div σ, v) = (f, v) for all v ∈ V.(1.1)
Here the symmetric tensor space for the stress Σ is defined by
H(div, Ω; S) :=
τ =
τ11 · · · τ1n
......
...
τn1 · · · τnn
∈ H(div, Ω; Rn×n)
∣∣∣∣∣∣∣∣∣τT = τ
, (1.2)
and the space for the vector displacement V is
L2(Ω; Rn) :=(u1, . . . , un)T
∣∣ ui ∈ L2(Ω), i = 1, . . . , n. (1.3)
This paper denotes by Hk(T ; X) the Sobolev space1 consisting of functions withdomain T ⊂ R
n, taking values in the finite-dimensional vector space X , and with allderivatives of order at most k square-integrable. For our purposes, the range spaceX will be either S, Rn, or R. Let ‖ · ‖k,T be the norm of Hk(T ), and S denote thespace of symmetric tensors, and H(div, T ; S) consist of square-integrable symmetricmatrix fields with square-integrable divergence. The H(div)-norm is defined by
‖τ‖2H(div,T ) := ‖τ‖2
0,T + ‖div τ‖20,T .
Let L2(T ; Rn) be the space of vector-valued functions which are square-integrable.Here, the compliance tensor A = A(x) : S → S, characterizing the properties of thematerial, is bounded and symmetric positive definite uniformly for x ∈ Ω.
The rest of the paper is organized as follows. In the next section, we presentsome preliminary results from Ref. 28; see also Refs. 32 and 33, for the cases n = 2and n = 3, respectively. In Sec. 3, based on these preliminary results, we propose afamily of Arnold–Winther mixed finite elements in any space dimension. In Sec. 4,we present lower order mixed finite elements and analyze the well-posedness of thediscrete problem and error estimates of the approximation solution. In Sec. 5, wepresent the first-order mixed elements. The paper ends with Sec. 6 which lists somenumerics.
2. Preliminary Results
Suppose that the domain Ω is subdivided by a family of shape regular simplicialgrids Th (with the grid size h). For any edge xixj of element K, i = j, let ti,j denote
Mat
h. M
odel
s M
etho
ds A
ppl.
Sci.
2016
.26:
1649
-166
9. D
ownl
oade
d fr
om w
ww
.wor
ldsc
ient
ific
.com
by U
NIV
ER
SIT
Y O
F D
EL
AW
AR
E o
n 10
/17/
17. F
or p
erso
nal u
se o
nly.
July 15, 2016 13:43 WSPC/103-M3AS 1650040
1652 J. Hu & S. Zhang
associated tangent vectors, which allow for us to introduce the following symmetricmatrices of rank one
Ti,j := ti,jtTi,j , 0 ≤ i < j ≤ n. (2.1)
For these matrices of rank one, we have the following result from Ref. 28; seealso Refs. 32 and 33, for the cases n = 2 and n = 3, respectively.
Lemma 2.1. The (n+1)n2 symmetric tensors Ti,j in (2.1) are linearly independent,
and form a basis of S.
With these symmetric matrices Ti,j of rank one, we define an H(div, K; S) bub-ble function space
ΣK,k,b :=∑
0≤i<j≤n
λiλjPk−2(K; R)Ti,j, (2.2)
where λi, i = 0, . . . , n, are the barycenter coordinates of element K. Define the fullH(div, K; S) bubble function space consisting of polynomials of degree ≤k:
Σ∂K,k,0 := τ ∈ Pk(K; S), τν|∂K = 0. (2.3)
Here ν is the normal vector of ∂K. We have the following result due to Ref. 28.
Lemma 2.2. It holds that
ΣK,k,b = Σ∂K,k,0. (2.4)
Let ΣK,b,h denote the sum of these H(div, K; S) bubble function spaces, namely,
Σk,b,h :=∑
K∈Th
ΣK,k,b. (2.5)
We need an important result concerning the divergence space of the bubblefunction space. To this end, we introduce the following rigid motion space on eachelement K:
R(K) := v ∈ H1(K; Rn), ε(v) := (∇v + ∇vT )/2 = 0. (2.6)
It follows from the definition that R(K) is a subspace of P1(K; Rn). For n = 1,R(K) is the constant function space over K. The dimension of R(K) is n(n+1)
2 . Fortwo dimensions, the rigid motion space R(K) is
R(K) :=(
a1
a2
)+ b
(−x2
x1
), a1, a2, b ∈ R
; (2.7)
for three dimensions, the rigid motion space R(K) reads
R(K) :=
a1
a2
a3
+ b1
−x2
x1
0
+ b2
−x3
0x1
+ b3
0−x3
x2
, ai, bi ∈ R, i = 1, 2, 3
.
(2.8)
Mat
h. M
odel
s M
etho
ds A
ppl.
Sci.
2016
.26:
1649
-166
9. D
ownl
oade
d fr
om w
ww
.wor
ldsc
ient
ific
.com
by U
NIV
ER
SIT
Y O
F D
EL
AW
AR
E o
n 10
/17/
17. F
or p
erso
nal u
se o
nly.
July 15, 2016 13:43 WSPC/103-M3AS 1650040
Finite element approximations of symmetric tensors 1653
This allows for defining the orthogonal complement space of R(K) with respectto Pk−1(K; Rn) by
R⊥(K) := v ∈ Pk−1(K; Rn), (v, w)K = 0 for any w ∈ R(K), (2.9)
where the inner product (v, w)K over K reads (v, w)K =∫
K v · wdx. When k = 1we have R⊥(K) = 0.
Lemma 2.3. For any K ∈ Th, it holds that
div ΣK,k,b = R⊥(K). (2.10)
Proof. The proof can be found in Ref. 28; see also Refs. 32 and 33, for the casesn = 2 and n = 3, respectively.
We need a classical result and its variant.
Lemma 2.4. It holds the following Chu–Vandermonde combinatorial identity andits variant
n∑=0
C+1n+1C
k−1 =
n∑=0
Cn−n+1C
k−1 = Cn
n+k, (2.11)
andn∑
=0
C+1n+1C
k−1C
2+1 =
(n + 1)n2
Cnn+k−2, (2.12)
where the combinatorial number Cmn = n···(n−m+1)
m···1 for n ≥ m and Cmn = 0 for
n < m.
3. A Family of Arnold–Winther Mixed Elementsin Any Space Dimension
3.1. The lowest order Arnold–Winther mixed
elements in any space dimension
To define lower order mixed finite elements with k ≤ n, we extend the lowest orderArnold–Winther mixed elements in 2D10 and 3D2,6 to any space dimension. To thisend, we introduce the following divergence-free space for element K ∈ Th,
Σ3→n+1,DF (K; S) := τ ∈ Pn+1(K; S)\P2(K; S), div τ = 0, (3.1)
where
Pn+1(K; S)\P2(K; S) := τ ∈ Pn+1(K; S) and τ ∈ P2(K; S). (3.2)
For any τ ∈Pn+1(K; S)\P2(K; S), its divergence div τ is an n-dimensional vector-valued polynomial of degree ≤n. This implies that the dimension of div Pn+1(K; S)\P2(K; S) is
n
((2n)!n!n!
− (n + 1))
Mat
h. M
odel
s M
etho
ds A
ppl.
Sci.
2016
.26:
1649
-166
9. D
ownl
oade
d fr
om w
ww
.wor
ldsc
ient
ific
.com
by U
NIV
ER
SIT
Y O
F D
EL
AW
AR
E o
n 10
/17/
17. F
or p
erso
nal u
se o
nly.
July 15, 2016 13:43 WSPC/103-M3AS 1650040
1654 J. Hu & S. Zhang
which is equal to the number of the divergence-free constraints in (3.1). Since thedimension of the space Pn+1(K, S)\P2(K, S) is(
(2n + 1)!n!(n + 1)!
− (n + 2)!2!n!
)n(n + 1)
2,
it follows that the dimension of the space Σ3→n+1,DF (K; S) is((2n + 1)!n!(n + 1)!
− (n + 2)!2!n!
)n(n + 1)
2− n
(2n)!n!n!
+ n(n + 1). (3.3)
Then we can define the following enriched P2(K; S) space
P ∗2 (K; S) := P2(K; S) + Σ3→n+1,DF (K; S). (3.4)
The dimension of P ∗2 (K; S) is equal to the dimension of P2(K; S) plus the dimension
of Σ3→n+1,DF (K; S). Thanks to (3.3),
the dimension of P ∗2 (K; S) =
(2n + 1)!n!(n + 1)!
n(n + 1)2
− n(2n)!n!n!
+ n(n + 1). (3.5)
To present the degrees of freedom of P ∗2 (K; S), we define
M2(K) := τ ∈ P ∗2 (K; S), div τ = 0 and τν|∂K = 0, (3.6)
where ν is the normal vector of ∂K. For the space M2(K), we have the followingimportant result.
Lemma 3.1. The dimension of M2(K) is
(2n − 1)!n!(n − 1)!
n(n + 1)2
+n(n + 1)
2− n
(2n)!n!n!
. (3.7)
Proof. By Lemma 2.1, Ti,j, 0 ≤ i < j ≤ n, are linearly independent, which impliesthat the dimension of the bubble function space ΣK,n+1,b defined in (2.2) is
(2n − 1)!n!(n − 1)!
n(n + 1)2
. (3.8)
Since the dimension of R(K) is n(n+1)2 , the dimension of R⊥(K) (with respect to
Pn(K; Rn)) is
n(2n)!n!n!
− n(n + 1)2
. (3.9)
Thanks to Lemma 2.2, the H(div) bubble function space ΣK,n+1,b is the full H(div)bubble function space of symmetric tensor-valued polynomials of degree ≤n+1 overK. Then, the definitions of P ∗
2 (K; S) and M2(K) imply that M2(K) is identical tothe space of all the divergence-free functions in ΣK,n+1,b. Therefore, its dimensionof M2(K) is equal to
the dimension of ΣK,n+1,b − the dimension of R⊥(K)
which completes the proof.
Mat
h. M
odel
s M
etho
ds A
ppl.
Sci.
2016
.26:
1649
-166
9. D
ownl
oade
d fr
om w
ww
.wor
ldsc
ient
ific
.com
by U
NIV
ER
SIT
Y O
F D
EL
AW
AR
E o
n 10
/17/
17. F
or p
erso
nal u
se o
nly.
July 15, 2016 13:43 WSPC/103-M3AS 1650040
Finite element approximations of symmetric tensors 1655
Before presenting the degrees of freedom of P ∗2 (K; S), we introduce a more
notation with 0 ≤ ≤ n which denotes an -dimensional simplex of K. For = 0, 0 is a vertex of K; for = 1, 1 is an edge of K.
Theorem 3.1. A matrix field τ ∈ P ∗2 (K; S) can be uniquely determined by the
following degrees of freedom:
(1) the mean moments of degree at most n − over , of tTl τνi, νT
i τνj, l =1, . . . , , i, j = 1, . . . , n− , (C2
n+1− + (n− ))Cn = (n−)(n++1)
2 Cn degrees of
freedom, for each -dimensional simplex of K, 0 ≤ ≤ n−1, with linearlyindependent tangential vectors t1, . . . , t, and n− linearly independent normalvectors ν1, . . . , νn−;
(2) the average of τ over K, n(n+1)2 degrees of freedom;
(3) the values of moments∫
K τ : θdx, θ ∈ M2(K), (2n−1)!n!(n−1)!
n(n+1)2 + n(n+1)
2 −n (2n)!n!n!
degrees of freedom.
Proof. We assume that all degrees of freedom vanish and show that τ = 0. Notethat the mean moment becomes the value of τ for a zero-dimensional simplex 0,namely, a vertex, of K. The first set of degrees of freedom implies that τν = 0 on∂K while the second set of degrees of freedom shows div τ = 0. Then the third setof degrees of freedom proves that τ = 0. Next we shall prove that the sum of thesedegrees of freedom is equal to the dimension of the space P ∗
2 (K, S). In fact the sumof the first set of degrees of freedom is
n−1∑=0
C+1n+1
(n − )(n + + 1)2
Cn,
we refer interested readers to Ref. 28 for a detailed proof of the numbers of degreesof freedom in the first set. By the Chu–Vandermonde combinatorial identity (2.11)and its variant (2.12), see more details from Ref. 28,
n−1∑=0
C+1n+1
(n − )(n + + 1)2
Cn =
(2n + 1)!n!(n + 1)!
n(n + 1)2
− (2n − 1)!n!(n − 1)!
n(n + 1)2
.
Hence the desired result follows from (3.5), and the sum of the second and thirdsets of degrees of freedom.
We denote by Σ∗2,h the space of symmetric tensor fields that belong piecewise
to P ∗2 (K; S), and with the continuity conditions induced by the degrees of freedom
defined in Theorem 3.1. In particular, for τ ∈ Σ∗2,h, the normal components τν are
continuous across all the internal (n − 1)-dimensional simplices n−1 and, henceΣ∗
2,h ⊂ H(div, Ω; S).
Remark 3.1. For n = 2, Σ∗2,h is the lowest order Arnold–Winther element stress
space proposed in Ref. 10; for n = 3, Σ∗2,h is the discrete stress space defined in
Ref. 2, see also Ref. 6.
Mat
h. M
odel
s M
etho
ds A
ppl.
Sci.
2016
.26:
1649
-166
9. D
ownl
oade
d fr
om w
ww
.wor
ldsc
ient
ific
.com
by U
NIV
ER
SIT
Y O
F D
EL
AW
AR
E o
n 10
/17/
17. F
or p
erso
nal u
se o
nly.
July 15, 2016 13:43 WSPC/103-M3AS 1650040
1656 J. Hu & S. Zhang
To define a family of first-order mixed elements, we need a family of simplifiedlowest order mixed elements, which is defined by
P ∗2 (K; S) := τ ∈ P ∗
2 (K; S), div τ ∈ R(K). (3.10)
Since the dimension of R(K) is n(n+1)2 , the equation gives n(n+1)
2 constraints on
P ∗2 (K; S). Hence the dimension of P ∗
2 (K; S) is
(2n + 1)!n!(n + 1)!
n(n + 1)2
− n(2n)!n!n!
+n(n + 1)
2.
A complete set of degrees of freedom for P ∗2 (K; S) is obtained by removing the
n(n+1)2 average values over K for P ∗
2 (K; S). The global space Σ∗2,h is defined in a
similar way as Σ∗2,h.
Remark 3.2. For n = 2, 3, Σ∗2,h are the simplified lowest order element stress
spaces in Refs. 10 and 6, respectively.
3.2. Higher order Arnold–Winther mixed elements
To define Arnold–Winther mixed elements of order k > 2, we introduce the followingdivergence-free space for element K ∈ Th,
Σk+1→k+n−1,DF (K; S) := τ ∈ Pk+n−1(K; S)\Pk(K; S), div τ = 0. (3.11)
Here Pk+n−1(K; S)\Pk(K; S) and Pk+n−2(K; R)\Pk−1(K; R) are defined ina similar way as Pn+1(K; S)\P2(K; S) defined in (3.2). The dimension ofΣk+1→k+n−1,DF (K; S) can be analyzed by a similar argument as that forΣ3→n+1,DF (K; S) in the previous subsection. In fact, since the dimension of thespace Pk+n−2(K; R)\Pk−1(K; R) is
((k + 2n − 2))!n!(k + n − 2)!
− (n + k − 1)!n!(k − 1)!
,
the number of the divergence-free constraints imposed in (3.11) is
n
(((k + 2n − 2))!n!(k + n − 2)!
− (n + k − 1)!n!(k − 1)!
).
In addition, the dimension of the space Pk+n−1(K; S)\Pk(K; S) is((k + 2n − 1)!n!(k + n − 1)!
− (n + k)!k!n!
)n(n + 1)
2.
It follows that the dimension of the space Σk+1→k+n−1,DF (K; S) is((k + 2n − 1)!n!(k + n − 1)!
− (n + k)!k!n!
)n(n + 1)
2− n
(((k + 2n − 2))!n!(k + n − 2)!
− (n + k − 1)!n!(k − 1)!
).
(3.12)
Then, we define the following enriched Pk(K; S) space
P ∗k (K; S) := Pk(K; S) + Σk+1→k+n−1,DF (K; S). (3.13)
Mat
h. M
odel
s M
etho
ds A
ppl.
Sci.
2016
.26:
1649
-166
9. D
ownl
oade
d fr
om w
ww
.wor
ldsc
ient
ific
.com
by U
NIV
ER
SIT
Y O
F D
EL
AW
AR
E o
n 10
/17/
17. F
or p
erso
nal u
se o
nly.
July 15, 2016 13:43 WSPC/103-M3AS 1650040
Finite element approximations of symmetric tensors 1657
It follows that the dimension of P ∗k (K; S) is equal to the sum of the dimension of
Pk(K; S) and the dimension of Σk+1→k+n−1,DF (K; S), namely,
(k + 2n − 1)!n!(k + n − 1)!
n(n + 1)2
− n
((k + 2n − 2)!n!(k + n − 2)!
− (n + k − 1)!n!(k − 1)!
). (3.14)
To present the degrees of freedom of P ∗k (K; S), we define
Mk(K) := τ ∈ P ∗k (K; S), div τ = 0 and τν|∂K = 0, (3.15)
where ν is the normal vector of ∂K. For the space Mk(K), we have the followingimportant result.
Lemma 3.2. The dimension of Mk(K) is
(k + 2n − 3)!n!(k + n − 3)!
n(n + 1)2
+n(n + 1)
2− n
(k + 2n − 2)!n!(k + n − 2)!
. (3.16)
Proof. The dimension of the space ΣK,k+n−1,b reads
(k + 2n − 3)!n!(k + n − 3)!
n(n + 1)2
. (3.17)
Since the dimension of R(K) is n(n+1)2 , the dimension of R⊥(K) (with respect to
Pk+n−2(K; Rn)) is
n(k + 2n− 2)!n!(k + n − 2)!
− n(n + 1)2
. (3.18)
It follows from the definition of P ∗k (K; S) and Lemma 2.2 that Mk(K) contains
all divergence-free tensor-value functions of ΣK,k+n−1,b. Then the desired resultfollows from Lemma 2.3.
Theorem 3.2. A matrix field τ ∈ P ∗k (K; S) can be uniquely determined by the
following degrees of freedom:
(1) the mean moments of degree at most k + n − − 2 over , of tTl τνi, νT
i τνj ,
l = 1, . . . , , i, j = 1, . . . , n−, (C2n+1−+(n−))C
k+n−2 = (n−)(n++1)2 C
k+n−2
degrees of freedom, for each -dimensional simplex of K, 0 ≤ ≤ n−1, with
linearly independent tangential vectors t1, . . . , t, and n− linearly independentnormal vectors ν1, . . . , νn−;
(2) the values∫
K τ : θdx for any θ ∈ ε(Pk−1(K; Rn)), nCnn+k−1 −
n(n+1)2 degrees of
freedom;(3) the values
∫K τ : θdx for any θ∈Mk(K), (k+2n−3)!
n!(k+n−3)!n(n+1)
2 + n(n+1)2 −n (k+2n−2)!
n!(k+n−2)!
degrees of freedom.
Proof. We assume that all degrees of freedom vanish and show that τ = 0. Notethat the mean moment become the value of τ for a zero-dimensional simplex 0,namely, a vertex, of K. The first set of degrees of freedom implies that τν = 0 on
Mat
h. M
odel
s M
etho
ds A
ppl.
Sci.
2016
.26:
1649
-166
9. D
ownl
oade
d fr
om w
ww
.wor
ldsc
ient
ific
.com
by U
NIV
ER
SIT
Y O
F D
EL
AW
AR
E o
n 10
/17/
17. F
or p
erso
nal u
se o
nly.
July 15, 2016 13:43 WSPC/103-M3AS 1650040
1658 J. Hu & S. Zhang
∂K while the second set of degrees of freedom shows div τ = 0. Then the third setof degrees of freedom proves that τ = 0.
Next we shall prove that the sum of these degrees of freedom is equal to thedimension of the space P ∗
k (K; S). In fact, it follows from the Chu–Vandermondecombinatorial identity (2.11) and its variant (2.12) that the number of degrees inthe first set is
n−1∑=0
C+1n+1
(n − )(n + + 1)2
Cn+k−2 =
n(n + 1)2
(Cnk+2n−1 − Cn
k+2n−3); (3.19)
we refer interested readers to Ref. 28 for a detailed proof of the numbers of degreesof freedom in the first set. The desired result follows from (3.14) and (3.16).
We denote by Σ∗k,h the space of symmetric tensor fields that belong piecewise
to P ∗k (K; S), and with the continuity conditions induced by the degrees of freedom
defined in Theorem 3.2. In particular, for τ ∈ Σ∗k,h, the normal components τν are
continuous across all the internal (n − 1)-dimensional simplices n−1 and, henceΣ∗
k,h ⊂ H(div, Ω; S).
Remark 3.3. For n = 2, Σ∗k,h is the higher order mixed element stress spaces in
Ref. 10; for n = 3, Σ∗k,h is the higher order mixed element stress spaces in Ref. 6.
4. A Family of Lower Order Mixed Elements
4.1. Mixed methods
For 2 ≤ k ≤ n, we follow the idea of Refs. 28, 32 and 33 to define the followingdiscrete stress space:
Σk,h := σ ∈ H(div, Ω; S), σ = σc + σb, σc ∈ H1(Ω; S),
σc|K ∈ Pk(K; S), σb|K ∈ ΣK,k,b, ∀K ∈ Th, (4.1)
which is an H(div) bubble enrichment of the H1 space
Σk,h := τ ∈ H1(Ω; S), τ |K ∈ Pk(K; S), ∀K ∈ Th.
For τ ∈ Σk,h, the degrees of freedom on any element K are: the mean moments ofdegree at most k − − 1 over each -dimensional simplex of K, 0 ≤ ≤ n of τ .A standard argument is able to prove that these degrees of freedom are unisolvent.
In order to get a stable pair of spaces, we take the discrete displacement spaceas the space of discontinuous piecewise vector-valued polynomials of degree ≤k−1,namely,
Vk,h := v ∈ L2(Ω; Rn), v|K ∈ Pk−1(K; Rn) for all K ∈ Th. (4.2)
Unfortunately, we cannot establish the stability of the pair of spaces Σk,h and Vk,h.We have to enrich the discrete stress space by some bubble functions. More precisely,
Mat
h. M
odel
s M
etho
ds A
ppl.
Sci.
2016
.26:
1649
-166
9. D
ownl
oade
d fr
om w
ww
.wor
ldsc
ient
ific
.com
by U
NIV
ER
SIT
Y O
F D
EL
AW
AR
E o
n 10
/17/
17. F
or p
erso
nal u
se o
nly.
July 15, 2016 13:43 WSPC/103-M3AS 1650040
Finite element approximations of symmetric tensors 1659
the displacement space Vk,h can be decomposed as
Vk,h = V Rk,h ⊕ V R⊥
k,h ,
where
V Rk,h := v ∈ L2(Ω; Rn), v|K ∈ R(K) for all K ∈ Th
and
V R⊥k,h := v ∈ L2(Ω; Rn), v|K ∈ R⊥(K) for all K ∈ Th.
By Lemma 2.3,
div Σk,b,h = V R⊥k,h ,
where Σk,b,h is defined in (2.5). That is to say, to get a stable pair of spaces, we onlyneed to enrich the discrete stress space Σk,h by some higher order bubble functionswhich are able to control the piecewise rigid motions in V R
k,h. We shall select thesebubble functions from the lowest order Arnold–Winther element stress space Σ∗
2,h
defined in the previous section. To this end, given an (n−1)-dimensional simplex F
of Th, let ωF := K−∪K+ denote the union of two elements that share F . We recallthat R(ωF ) is the rigid motion space over ωF while R(ωF )|F is the restriction onF . Further we let (R(ωF )|F )⊥ denote the orthogonal complement space of R(ωF )|Fwith respect to P1(F ; Rn) which allows to define the following (n− 1)-dimensionalsimplex H(div) bubble function space:
B1F :=
τ ∈ Σ∗
2,h, τ = 0 on Ω\ωF ,
∫F
τν · pds = 0 for any p ∈ (R(ωF )|F )⊥,
the averages of τ over both K− and K+ vanish,
the values of∫
K
τ : θdx vanish for any θ ∈ M2(K), K = K− and K+
.
(4.3)
Here ν is the normal vector of F . We also need a subspace of B1F defined by
B2F :=
τ ∈ B
1F ,
∫F
τν · pds = 0 for any p ∈ P0(F ; Rn)
. (4.4)
Hence we define the following enriched stress space
Σ+k,h = Σk,h +
∑F
B1F for 2 ≤ k ≤ n − 1; (4.5)
and
Σ+k,h = Σk,h +
∑F
B2F for k = n. (4.6)
Lemma 4.1. The space Σ+k,h is a direct sum of the spaces Σk,h and ΣF B1
F for2 ≤ k ≤ n − 1, and it is a direct sum of the spaces Σk,h and ΣF B2
F for k = n.
Proof. We only prove the first part of the theorem since the proof of the secondpart is similar. In fact, on the one hand, given F , it follows from the definition of
Mat
h. M
odel
s M
etho
ds A
ppl.
Sci.
2016
.26:
1649
-166
9. D
ownl
oade
d fr
om w
ww
.wor
ldsc
ient
ific
.com
by U
NIV
ER
SIT
Y O
F D
EL
AW
AR
E o
n 10
/17/
17. F
or p
erso
nal u
se o
nly.
July 15, 2016 13:43 WSPC/103-M3AS 1650040
1660 J. Hu & S. Zhang
B1F that for any τ ∈ B1
F it vanishes on the following degrees of freedom:
• the mean moments of degree at most n− over , of tTl τνi, νT
i τνj , l = 1, . . . , ,
i, j = 1, . . . , n − , (C2n+1− + (n − ))C
n = (n−)(n++1)2 C
n degrees of freedom,for each -dimensional simplex of K, 0 ≤ ≤ n − 2, with linearly indepen-dent tangential vectors t1, . . . , t, and n− linearly independent normal vectorsν1, . . . , νn−,
for any K ∈ Th. On the other hand, for any τ ∈ Pk(K; S) with 2 ≤ k ≤ n − 1, ifit vanishes on the above degrees of freedom, τν = 0 on ∂K where ν is the normalvector of ∂K; see Ref. 28 for more details. This indicates that τν = 0 on F whichimplies the first part of the theorem.
It follows from the definitions of Vk,h (Pk−1 polynomials) and Σ+k,h (enriched Pk
polynomials) that
div Σ+k,h ⊂ Vk,h.
This, in turn, leads to a strong divergence-free space:
Zh := τh ∈ Σ+k,h | (div τh, v) = 0 for all v ∈ Vk,h
= τh ∈ Σ+k,h | div τh = 0 pointwise. (4.7)
The mixed finite element approximation of problem (1.1) reads: find (σh, uh) ∈Σ+
k,h × Vk,h such that(Aσh, τ) + (div τ, uh) = 0 for all τ ∈ Σ+
k,h,
(div σh, v) = (f, v) for all v ∈ Vk,h.(4.8)
4.2. Stability analysis and error estimates
The convergence of the finite element solution follows the stability and the standardapproximation property. So we consider first the well-posedness of the discreteproblem (4.8). By the standard theory, we only need to prove the following twoconditions, based on their counterpart at the continuous level.
(1) K-ellipticity. There exists a constant C > 0, independent of the meshsize h
such that
(Aτ, τ) ≥ C‖τ‖2H(div) for all τ ∈ Zh, (4.9)
where Zh is the divergence-free space defined in (4.7).(2) Discrete B-B condition. There exists a positive constant C > 0 independent of
the meshsize h, such that
inf0=v∈Vk,h
sup0=τ∈Σ+
k,h
(div τ, v)‖τ‖H(div)‖v‖0
≥ C. (4.10)
Mat
h. M
odel
s M
etho
ds A
ppl.
Sci.
2016
.26:
1649
-166
9. D
ownl
oade
d fr
om w
ww
.wor
ldsc
ient
ific
.com
by U
NIV
ER
SIT
Y O
F D
EL
AW
AR
E o
n 10
/17/
17. F
or p
erso
nal u
se o
nly.
July 15, 2016 13:43 WSPC/103-M3AS 1650040
Finite element approximations of symmetric tensors 1661
Theorem 4.1. For the discrete problem (4.8), the K-ellipticity (4.9) and the dis-crete B-B condition (4.10) hold uniformly. Consequently, the discrete mixed problem(4.8) has a unique solution (σh, uh) ∈ Σ+
k,h × Vk,h.
Proof. The K-ellipticity immediately follows from the fact that div Σ+k,h ⊂ Vk,h.
To prove the discrete B-B condition (4.10), for any vh ∈ Vk,h, we will construct aninterpolation operator Πh : H1(Ω; S) → Σ+
k,h such that, for any τ ∈ H1(Ω; S),∫K
div(τ − Πhτ) · vdx = 0 for any v ∈ Vk,h, (4.11)
for any K ∈ Th. Further, if τ ∈ Hk+1(Ω; S), it holds
‖τ − Πhτ‖0 + h‖div(τ − Πhτ)‖0 ≤ Chk+1|τ |k+1. (4.12)
We only show the above result for the cases 2 ≤ k ≤ n−1 since the proof for the casek = n is similar. First let Ih : H1(Ω; S) → Σk,h be a Scott–Zhang38 interpolationoperator such that
‖τ − Ihτ‖0 + h‖∇Ihτ‖0 ≤ Ch‖∇τ‖0. (4.13)
Since Ih preserves symmetric Pk functions locally,
‖τ − Ihτ‖0 + h‖∇(τ − Ihτ)‖0 ≤ Chk+1|τ |k+1, (4.14)
provided that τ ∈ Hk+1(Ω; S). See Ref. 28 for more details.Second, these enriched bubble functions in
∑F B1
F on the (n − 1)-dimensionalsimplices F allow for defining a correction δF
h ∈ B1F such that∫
F
δFh ν · pds =
∫F
(τ − Ihτ)ν · pds for any p ∈ R(K)|F . (4.15)
For these corrections δFh , we have
‖δFh ‖0,ωF + h‖div δF
h ‖0,ωF ≤ C(‖τ − Ihτ‖0,ωF + h‖∇(τ − Ihτ)‖0,ωF ). (4.16)
Finally, we take
Π1hτ = Ihτ +
∑F
δFh . (4.17)
We get a partial-divergence matching property of Π1hτ : for any p ∈ R(K), as the
symmetric gradient ε(p) = 0,∫K
(div Π1hτ − div τ) · pdx =
∫∂K
(Π1hτ − τ)ν · pds = 0. (4.18)
Next we make a correction δKh ∈ ΣK,k,b on each element K such that∫
K
div δKh · pdx =
∫K
div(τ − Π1hτ) · pdx for any p ∈ R⊥(K). (4.19)
The existence of δKh follows from Lemma 2.3, which also implies that
‖div δKh ‖0,K ≤ ‖div(τ − Π1
hτ)‖0,K . (4.20)
Mat
h. M
odel
s M
etho
ds A
ppl.
Sci.
2016
.26:
1649
-166
9. D
ownl
oade
d fr
om w
ww
.wor
ldsc
ient
ific
.com
by U
NIV
ER
SIT
Y O
F D
EL
AW
AR
E o
n 10
/17/
17. F
or p
erso
nal u
se o
nly.
July 15, 2016 13:43 WSPC/103-M3AS 1650040
1662 J. Hu & S. Zhang
In addition, the δKh can be selected such that
‖δKh ‖0,K = min‖δK‖0,K , div δK = Π⊥
K div(τ − Π1hτ), δK ∈ ΣK,k,b, (4.21)
where Π⊥K : L2(K, Rn) → R⊥(K) denotes the L2 projection operator. It follows
that ‖div δKh ‖0,K defines a norm for it. Then, a scaling argument proves
‖δKh ‖0,K ≤ Ch‖div δK
h ‖0,K ≤ Ch‖div(τ − Π1hτ)‖0,K . (4.22)
Finally, we define
Πhτ = Π1hτ +
∑K
δKh . (4.23)
Hence Eq. (4.11) follows from (4.18) and (4.19). The estimate (4.12) follows from(4.14), (4.16), (4.20) and (4.22). In addition, by (4.13), (4.20) and (4.22),
‖τ − Πhτ‖0 + h‖∇Πhτ‖0 ≤ Ch‖∇τ‖0. (4.24)
In the sequel, we use the interpolation operator Πh to prove the discrete inf–supcondition (4.10). Indeed, by the stability of the continuous formulation, there is aτ ∈ H1(Ω; S) such that,
div τ = vh and ‖τ‖1 ≤ C‖vh‖0.
In this paper, we only consider the domain such that the above stability holds. Werefer interested readers to Ref. 23 for the classical result which states it is true forLipschitz domains in Rn; see Ref. 21 for more refined results.
It follows from (4.11) and (4.24) that
div Πhτ = vh and ‖τ − Πhτ‖0 + h‖∇Πhτ‖0 ≤ Ch‖vh‖0, (4.25)
which shows (4.10) and completes the proof.
Theorem 4.2. Let (σ, u) ∈ Σ × V be the exact solution of problem (1.1) and(τh, uh) ∈ Σ+
k,h × Vk,h the finite element solution of (4.8). Then, for 2 ≤ k ≤ n,
‖σ − σh‖H(div) + ‖u − uh‖0 ≤ Chk(‖σ‖k+1 + ‖u‖k), (4.26)
and
‖σ − σh‖0 ≤ Chk+1|σ|k+1. (4.27)
Proof. The estimate (4.27) follows from the stability of the elements in Theorem4.1 and the standard theory of mixed finite element methods.15,16 Let the interpo-lation operator Πh be defined in (4.11). It follows
div Πhσ = div σh. (4.28)
This leads to
(A(σh − Πhσ), σh − Πhσ) = (A(σh − σ), σh − Πhσ)
+ (A(σ − Πhσ), σh − Πhσ)
= (A(σ − Πhσ), σh − Πhσ).
Mat
h. M
odel
s M
etho
ds A
ppl.
Sci.
2016
.26:
1649
-166
9. D
ownl
oade
d fr
om w
ww
.wor
ldsc
ient
ific
.com
by U
NIV
ER
SIT
Y O
F D
EL
AW
AR
E o
n 10
/17/
17. F
or p
erso
nal u
se o
nly.
July 15, 2016 13:43 WSPC/103-M3AS 1650040
Finite element approximations of symmetric tensors 1663
Hence
‖σh − Πhσ‖0 ≤ C‖σ − Πhσ‖0.
Then the estimate (5.8) follows from the triangle inequality and (4.12).
5. First-Order Mixed Elements
In order to get first-order mixed elements, we propose to take the following discretedisplacement space
V1,h := v ∈ L2(Ω; Rn), v|K ∈ R(K) for any K ∈ Th. (5.1)
To design the space for the stress, we define
Σ1,h := τ ∈ H1(Ω; S), τ |K ∈ P1(K, S) for any K ∈ Th. (5.2)
Since the pair (Σ1,h, V1,h) is unstable, we propose to enrich Σ1,h by some (n − 1)-dimensional simplex bubble function spaces. Given an (n− 1)-dimensional simplexF of Th, we define
BF :=
τ ∈ Σ∗2,h, τ = 0 on Ω\ωF ,
∫F
τν · pds = 0 for any p ∈ (R(ωF )|F )⊥,
the values of∫
K
τ : θdx vanish for any θ ∈ M2(K), K = K− and K+
.
(5.3)
This allows for defining the following enriched stress space
Σ+1,h = Σ1,h +
∑F
BF . (5.4)
For this enriched space Σ+1,h, the number of degrees of freedom on each simplex
is 18 and 48 for n = 2, 3, respectively, which are the simplest conforming mixedelements so far. A similar argument of Lemma 4.1 shows that Σ+
1,h is a direct sumof Σ1,h and ΣF BF .
The mixed finite element approximation of problem (1.1) reads: find (σh, uh) ∈Σ+
1,h × V1,h such that(Aσh, τ) + (div τ, uh) = 0 for all τ ∈ Σ+
1,h,
(div σh, v) = (f, v) for all v ∈ V1,h.(5.5)
It follows from div Σ+1,h ⊂ V1,h that div τ = 0 for any τ ∈ Zh, which implies the
above K-ellipticity condition (4.9). A similar proof of Theorem 4.1 shows the dis-crete inf–sup condition (4.10). In particular, there exists an interpolation operatorΠh : H1(Ω, S) → Σ+
1,h such that
‖τ − Πhτ‖0 + h‖div(τ − Πhτ)‖ ≤ hk‖τ‖k, k = 1, 2, (5.6)
and∫K
div(τ − Πhτ) : pdx =∫
∂K
(τ − Πhτ)ν · pds = 0 for any p ∈ R(K), (5.7)
Mat
h. M
odel
s M
etho
ds A
ppl.
Sci.
2016
.26:
1649
-166
9. D
ownl
oade
d fr
om w
ww
.wor
ldsc
ient
ific
.com
by U
NIV
ER
SIT
Y O
F D
EL
AW
AR
E o
n 10
/17/
17. F
or p
erso
nal u
se o
nly.
July 15, 2016 13:43 WSPC/103-M3AS 1650040
1664 J. Hu & S. Zhang
for any K ∈ Th. A summary of these results leads to the error estimates in thefollowing theorem.
Theorem 5.1. Let (σ, u) ∈ Σ × V be the exact solution of problem (1.1) and(τh, uh) ∈ Σ+
1,h × V1,h the finite element solution of (5.5). Then,
‖σ − σh‖H(div) + ‖u − uh‖0 ≤ Ch(‖div σ‖1 + ‖u‖1) (5.8)
and
‖σ − σh‖0 ≤ Ch2‖σ‖2. (5.9)
6. Numerical Test
We compute a 2D pure displacement problem on the unit square Ω = [0, 1]2 witha homogeneous boundary condition that u ≡ 0 on ∂Ω. In the computation, we letthe compliance tensor in (1.1):
Aσ =12µ
(σ − λ
2µ + nλtr(σ)δ
), n = 2,
where δ =(
1 0
0 1
), and µ = 1/2 and λ = 1 are the Lame constants. Let the exact
solution be
u =
(ex−yx(1 − x)y(1 − y)
sin(πx) sin(πy)
). (6.1)
The true stress function σ and the load function f are defined by the equationsin (1.1), for the given solution u.
In the computation, the level one grid consists of two right triangles, obtainedby cutting the unit square with a north-east line. Each grid is refined into a half-sized grid uniformly, to get a higher level grid. In all the computation, the discretesystems of equations are solved by Matlab backslash solver.
We use the bubble enriched P2 symmetric stress finite element with P1 discon-tinuous displacement finite element, k = 2 in (4.2) and in (4.6), and k = 2 in (4.1).That is, three P3 bubbles are enriched each edge. In Table 1, the errors and theconvergence order in various norms are listed for the true solution (6.1). The opti-mal order of convergence is observed for both displacement and stress, see Table 1,as shown in the theorem.
Table 1. The errors, eh = σ − σh, and the order of conver-gence, by the 2D k = 2 element in (4.6) and (4.2), for (6.1).
‖u − uh‖0 rate ‖eh‖0 rate ‖div eh‖0 hn
1 0.27452 0.0 1.24637 0.0 6.97007772 0.02 0.07432 1.9 0.18054 2.8 2.13781130 1.73 0.01959 1.9 0.02429 2.9 0.57734125 1.94 0.00497 2.0 0.00314 2.9 0.14709450 2.05 0.00125 2.0 0.00040 3.0 0.03694721 2.0
Mat
h. M
odel
s M
etho
ds A
ppl.
Sci.
2016
.26:
1649
-166
9. D
ownl
oade
d fr
om w
ww
.wor
ldsc
ient
ific
.com
by U
NIV
ER
SIT
Y O
F D
EL
AW
AR
E o
n 10
/17/
17. F
or p
erso
nal u
se o
nly.
July 15, 2016 13:43 WSPC/103-M3AS 1650040
Finite element approximations of symmetric tensors 1665
Table 2. The errors, eh = σ − σh, and the order of conver-
gence, by the Arnold–Winther 21/3 element, for (6.1).
‖u − uh‖0 rate ‖eh‖0 rate ‖div eh‖0 hn
1 0.30554 0.0 1.58058 0.0 10.31991249 0.02 0.22589 0.4 0.89927 0.8 6.81340378 0.63 0.10922 1.0 0.25584 1.8 3.61633797 0.94 0.05354 1.0 0.06633 1.9 1.83690959 1.05 0.02661 1.0 0.01674 2.0 0.92212628 1.0
As a comparison, we also test the Arnold–Winther element from Ref. 10, whichhas a same degree of freedom as ours, 21, on each element. But the displacementin that element is approximated by the rigid-motion space only, instead of the fullP1 space, i.e. 3 dof versus 6 dof on each triangle. The total degrees of freedom forthe stress for the new element are 3|V| + 3|E| + 3|K|, where |V|, |E| and |K| arethe numbers of vertices, edges and elements of Th, respectively, while those for theArnold–Winther element are 3|V| + 4|E|. Since the three bubble functions on eachelement can be easily condensed, these two elements almost have the same com-plexity for solving. The errors and the orders of convergence are listed in Table 2.Because the new element uses the full P1 displacement space, the order of con-vergence is one higher than that of the Arnold–Winther element. Also as the newelement includes the full P2 stress space, the order of convergence of stress is oneorder higher, see the data in Tables 1 and 2.
Appendix A. The Basis Functions of Σk,h in Two Dimensions
Let x0x1x2 =: K ∈ Th with three edges Ei and corresponding three barycentricvariables λi. Here λi is a linear function which vanishes on edge Ei and assumesa nodal value 1 at the opposite vertex xi. Given Ei = −−−−−−→xi−1xi+1, its two endpointsare xi−1 and xi+1, which allows for defining its k − 1 interior nodal points by
xEi,j =j
kxi−1 +
k − j
kxi+1, j = 1, . . . , k − 1. (A.1)
We also define (k−1)(k−2)2 nodal points inside K by
xK,l,m =l
kx0 +
m
kx1 +
k − l − m
kx2, 1 ≤ l, m and l + m ≤ k − 1. (A.2)
Then the nodes for the Lagrange element of order k is
XK = xi, i = 0, 1, 2 ∪ xEi,j , i = 0, 1, 2, j = 1, . . . , k − 1∪ xK,l,m, 1 ≤ l, m and l + m ≤ k − 1.
Given node xEi,j on edge Ei, j = 1, . . . , k−1, let φEi,j ∈ Pk(K; R) be its associatednodal basis function of the Lagrange element of order k such that
φEi,j(xEi,j) = 1 and φEi,j(x′) = 0 for any x′ ∈ XK other than xEi,j. (A.3)
Mat
h. M
odel
s M
etho
ds A
ppl.
Sci.
2016
.26:
1649
-166
9. D
ownl
oade
d fr
om w
ww
.wor
ldsc
ient
ific
.com
by U
NIV
ER
SIT
Y O
F D
EL
AW
AR
E o
n 10
/17/
17. F
or p
erso
nal u
se o
nly.
July 15, 2016 13:43 WSPC/103-M3AS 1650040
1666 J. Hu & S. Zhang
Let ni = 〈ni,1, ni,2〉T and n⊥i = 〈−ni,2, ni,1〉T be normal and tangent vectors on
edge Ei, respectively. We define a matrix of rank one by
TEi = n⊥i n⊥
i
T. (A.4)
With these (k− 1)-edge bubble functions φEi,j on each edge and the matrix TEi ofrank one, we can define exactly (k − 1) stress functions τEi,j by
τEi,j = φEi,jTEi , j = 1, 2, . . . , k − 1, i = 0, 1, 2. (A.5)
By the definition, we have
τEi,j · nl|El= 0, i, l = 0, 1, 2, j = 1, . . . , k − 1, (A.6)
which implies that they are H(div) bubble functions on element K.Given E we need a basis which takes TE for S. To this end, we let
TE,1 = nEnTE and TE,2 =
12(n⊥
EnTE + nE(n⊥
E)T ).
It is straightforward to see that TE , TE,1 and TE,2 are linearly independent andtherefore form a basis of S. The canonical basis of S reads
T1 =
(1 0
0 0
), T2 =
(0 11 0
), and T3 =
(0 00 1
). (A.7)
Let XE denote all interior nodes, defined in (A.1), of all the edges, XK denote allinterior nodes, defined in (A.2), of all the elements, and XV denote all the verticesof Th. Define the Lagrange element space of order k by
Ph := H1(Ω; R) ∩ v ∈ L2(Ω; R), v|K ∈ Pk(K; R), ∀K ∈ Th.
Given node x ∈ XV ∪ XE ∪ XK, let φx ∈ Ph be its associated nodal basis function,which is similarly defined as φEi,j in (A.3).
The basis functions of Σk,h can be classified into four classes:
(1) Vertex-based basis functions: Given vertex x ∈ XV, its three associated basisfunctions of Σk,h read
τV,x,i = φxTi, i = 1, 2, 3.
(2) Volume-based basis functions: Given node x ∈ XK inside K, its three associatedbasis functions of Σk,h read
τK,x,i = φxTi, i = 1, 2, 3.
(3) Edge-based basis functions with nonzero fluxes: Given node x ∈ XE on edge E,its two associated basis functions with nonzero fluxes of Σk,h read
τ(nb)E,x,i = φxTE,i, i = 1, 2.
(4) Edge-based bubble functions: Given node x ∈ XE on edge E which is shared byelements K1 and K2, its bubble functions in Σk,h read
τ(b)E,x,i = φx|KiTE , i = 1, 2.
Mat
h. M
odel
s M
etho
ds A
ppl.
Sci.
2016
.26:
1649
-166
9. D
ownl
oade
d fr
om w
ww
.wor
ldsc
ient
ific
.com
by U
NIV
ER
SIT
Y O
F D
EL
AW
AR
E o
n 10
/17/
17. F
or p
erso
nal u
se o
nly.
July 15, 2016 13:43 WSPC/103-M3AS 1650040
Finite element approximations of symmetric tensors 1667
It is straightforward to see that these functions defined in the above four termsform a basis of Σk,h, which are very easy to construct.
Acknowledgment
The first author was supported by the NSFC Projects 11271035, 91430213 and11421101.
References
1. R. A. Adams, Sobolev Spaces (Academic Press, 1975).2. S. Adams and B. Cockburn, A mixed finite element method for elasticity in three
dimensions, J. Sci. Comput. 25 (2005) 515–521.3. M. Amara and J. M. Thomas, Equilibrium finite elements for the linear elastic prob-
lem, Numer. Math. 33 (1979) 367–383.4. D. N. Arnold, Differential complexes and numerical stability, in Proc. of the Interna-
tional Congress of Mathematicians, Vol. I: Plenary Lectures and Ceremonies (HigherEd. Press, 2002), pp. 137–157.
5. D. N. Arnold and G. Awanou, Rectangular mixed finite elements for elasticity, Math.Models Methods Appl. Sci. 15 (2005) 1417–1429.
6. D. Arnold, G. Awanou and R. Winther, Finite elements for symmetric tensors in threedimensions, Math. Comput. 77 (2008) 1229–1251.
7. D. N. Arnold, F. Brezzi and J. Douglas Jr., PEERS: A new mixed finite element forplane elasticity, Jpn. J. Appl. Math. 1 (1984) 347–367.
8. D. N. Arnold, J. Douglas Jr. and C. P. Gupta, A family of higher order mixed finiteelement methods for plane elasticity, Numer. Math. 45 (1984), 1–22.
9. D. N. Arnold, R. Falk and R. Winther, Mixed finite element methods for linear elas-ticity with weakly imposed symmetry, Math. Comput. 76 (2007) 1699–1723.
10. D. N. Arnold and R. Winther, Mixed finite element for elasticity, Numer. Math. 92(2002), 401–419.
11. D. N. Arnold and R. Winther, Nonconforming mixed elements for elasticity, Math.Models Methods Appl. Sci. 13 (2003) 295–307.
12. G. Awanou, Two remarks on rectangular mixed finite elements for elasticity, J. Sci.Comput. 50 (2012) 91–102.
13. D. Boffi, F. Brezzi and M. Fortin, Reduced symmetry elements in linear elasticity,Commun. Pure Appl. Anal. 8 (2009) 95–121.
14. J. Bramwell, L. Demkowicz, J. Gopalakrishnan and W. F. Qiu, A locking free hpDPG method for linear elasticity with symmetric stresses, Numer. Math. 122 (2012)671–707.
15. F. Brezzi, On the existence, uniqueness and approximation of saddle-point problemsarising from Lagrangian multipliers, Rev. Francaise Automat. Informat. RechercheOperationnelle Ser. Rouge 8 (1974) 129–151.
16. F. Brezzi and M. Fortin, Mixed and Hybrid Finite Element Methods (Springer, 1991).17. C. Carstensen, M. Eigel and J. Gedicke, Computational competition of symmetric
mixed FEM in linear elasticity, Comput. Methods Appl. Mech. Engrg. 200 (2011)2903–2915.
18. C. Carstensen, D. Gunther, J. Reininghaus and J. Thiele, The Arnold–Winther mixedFEM in linear elasticity. Part I: Implementation and numerical verification, Comput.Methods Appl. Mech. Engrg. 197 (2008) 3014–3023.
Mat
h. M
odel
s M
etho
ds A
ppl.
Sci.
2016
.26:
1649
-166
9. D
ownl
oade
d fr
om w
ww
.wor
ldsc
ient
ific
.com
by U
NIV
ER
SIT
Y O
F D
EL
AW
AR
E o
n 10
/17/
17. F
or p
erso
nal u
se o
nly.
July 15, 2016 13:43 WSPC/103-M3AS 1650040
1668 J. Hu & S. Zhang
19. S. C. Chen and Y. N. Wang, Conforming rectangular mixed finite elements for elas-ticity, J. Sci. Comput. 47 (2011) 93–108.
20. B. Cockburn, J. Gopalakrishnan and J. Guzman, A new elasticity element made forenforcing weak stress symmetry, Math. Comput. 79 (2010) 1331–1349.
21. R. G. Duran and M. A. Muschietti, An explicit right inverse of the divergence operatorwhich is continuous in weighted norms, Studia Math. 148 (2001) 207–219.
22. B. M. Fraejis de Veubeke, Displacement and equilibrium models in the finite elementmethod, in Stress Analysis, eds. O. C. Zienkiewics and G. S. Holister (Wiley, 1965),pp. 145–197.
23. V. Girault and P. A. Raviart, Finite Element Methods for Navier–Stokes Equations(Springer, 1986).
24. J. Gopalakrishnan and J. Guzman, Symmetric nonconforming mixed finite elementsfor linear elasticity, SIAM J. Numer. Anal. 49 (2011) 1504–1520.
25. J. Gopalakrishnan and J. Guzman, A second elasticity element using the matrixbubble, IMA J. Numer. Anal. 32 (2012) 352–372.
26. J. Guzman, A unified analysis of several mixed methods for elasticity with weak stresssymmetry, J. Sci. Comput. 44 (2010) 156–169.
27. J. Hu, A new family of efficient conforming mixed finite elements on both rectangularand cuboid meshes for linear elasticity in the symmetric formulation, SIAM J. Numer.Anal. 53 (2015) 1438–1463.
28. J. Hu, Finite element approximations of symmetric tensors on simplicial grids in Rn:
The higher order case, J. Comput. Math. 33 (2015) 283–296.29. J. Hu, H. Man, J. Wang and S. Zhang, The simplest nonconforming mixed finite
element method for linear elasticity in the symmetric formulation on n-rectangulargrids, Comput. Math. Appl. 71 (2016) 1317–1336.
30. J. Hu, H. Y. Man and S. Zhang, A simple conforming mixed finite element for linearelasticity on rectangular grids in any space dimension, J. Sci. Comput. 58 (2014)367–379.
31. J. Hu and Z. C. Shi, Lower order rectangular nonconforming mixed elements for planeelasticity, SIAM J. Numer. Anal. 46 (2007) 88–102.
32. J. Hu and S. Zhang, A family of conforming mixed finite elements for linear elasticityon triangle grids, preprint (2014), arXiv: 1406.7457v2.
33. J. Hu and S. Zhang, A family of conforming mixed finite elements for linear elasticityon tetrahedral grids, Sci. China Math. 58 (2015) 297–307.
34. C. Johnson and B. Mercier, Some equilibrium finite element methods for two-dimensional elasticity problems, Numer. Math. 30 (1978) 103–116.
35. H.-Y. Man, J. Hu and Z.-C. Shi, Lower order rectangular nonconforming mixed finiteelement for the three-dimensional elasticity problem, Math. Models Methods Appl.Sci. 19 (2009) 51–65.
36. M. Morley, A family of mixed finite elements for linear elasticity, Numer. Math. 55(1989) 633–666.
37. W. F. Qiu and L. Demkowicz, Mixed hp-finite element method for linear elasticitywith weakly imposed symmetry, Comput. Methods Appl. Mech. Engrg. 198 (2009)3682–3701.
38. L. R. Scott and S. Zhang, Finite-element interpolation of non-smooth functions sat-isfying boundary conditions, Math. Comput. 54 (1990) 483–493.
39. R. Stenberg, On the construction of optimal mixed finite element methods for thelinear elasticity problem, Numer. Math. 48 (1986) 447–462.
Mat
h. M
odel
s M
etho
ds A
ppl.
Sci.
2016
.26:
1649
-166
9. D
ownl
oade
d fr
om w
ww
.wor
ldsc
ient
ific
.com
by U
NIV
ER
SIT
Y O
F D
EL
AW
AR
E o
n 10
/17/
17. F
or p
erso
nal u
se o
nly.
July 15, 2016 13:43 WSPC/103-M3AS 1650040
Finite element approximations of symmetric tensors 1669
40. R. Stenberg, Two low-order mixed methods for the elasticity problem, in The Math-ematics of Finite Elements and Applications, Vol. 6, ed. J. R. Whiteman (AcademicPress, 1988), pp. 271–280.
41. R. Stenberg, A family of mixed finite elements for the elasticity problem, Numer.Math. 53 (1988) 513–538.
42. V. B. Watwood Jr. and B. J. Hartz, An equilibrium stress field model for finite elementsolution of two-dimensional elastostatic problems, Int. J. Solids Struct. 4 (1968) 857–873.
43. S. Y. Yi, Nonconforming mixed finite element methods for linear elasticity usingrectangular elements in two and three dimensions, Calcolo 42 (2005) 115–133.
44. S. Y. Yi, A new nonconforming mixed finite element method for linear elasticity,Math. Models Methods Appl. Sci. 16 (2006) 979–999.
45. O. C. Zienkiewicz, R. L. Taylor and J. Z. Zhu, The Finite Element Method: Its Basisand Fundamentals, 6th edn., Vol. 1 (Elsevier, 2005).
Mat
h. M
odel
s M
etho
ds A
ppl.
Sci.
2016
.26:
1649
-166
9. D
ownl
oade
d fr
om w
ww
.wor
ldsc
ient
ific
.com
by U
NIV
ER
SIT
Y O
F D
EL
AW
AR
E o
n 10
/17/
17. F
or p
erso
nal u
se o
nly.