The h-p Version of Finite Element Method in R3: Theory and Algorithm

Benqi Guo

Abstract
This paper gives a precise description of regularities of solutions and its derivatives of all order for elliptic problems on polyhedral domain in the frame of the countably normed spaces with weighted Ck -norms in neighborhoods of vertices, edges and vertex-edges. Under the guidance of the regularity theory the geometric meshes and P-Q distribution (bilinear-linear or linear-uniform) of element degrees are designed accordingly in each of singular neighborhoods. The algorithms combining the geometric mesh and corresponding P-Q distribution of element degree achieve the exponential convergence and e ciency of computation. The performance of the h, p and h-p versions for a benchmark elasticity problem on a polyhedral domain is given and analysed.

Key words: the h-p version, nite element method, ge-

ometric mesh, P-Q distribution, exponential convergence, vertex singularity, edge singularity, vertex-edge singularity.. 35J25.

AMS subject classi cations: 65N30, 65N15, 35B65,

1 Introduction
The h-p version, its theory and algorithm, is a new development of the nite element method (FEM) in 1980's and 1990's. It was originated in and oriented to the structural mechanical problems on nonsmooth domains in R2
Department of Applied Mathematics, University of Manitoba, Canada R3T 2N2 (bguo@newton.amath.umanitoba.ca).
This space left blank for copyright notice.

and R3, but now its use has been expanded to many other elds such as uid mechanics, thermal analysis, electronic engineering, etc.. The methodology developed in the past decades has signi cantly in uenced the theory and algorithm of FEM, the practices of engineering and scienti c computationm and the industry of commercial FEM codes, such as MSC/PROBE, PHLEX, MECHANICAL, STREE CHECK, and research code STRIPE (see 34,35,39]). The h-p version has been one of the most su cient achievements of FEM's history since 1970's. It is well known that the singularities of solutions for problems on nonsmooth domains may occur at the vertices and edges, which severely a ect the e ectiveness and e ciency of nite element solutions. The h-version, which reduces the element size h, and the p-version, which increase the element degree p, may not be able to achieve the desired accuracy in practical engineering range. Then the h-p version is the only reliable nite element approach which is able to provide e ective and e cient algorithms. It reduces the element size h and increase the element degree p simutaneously and selectively in order to achieve the optimal rate of convergence and the e ciency of computations. The h-p version of FEM in R2 was introduced in 1980's (see 2,25,26]) and has been well developed since then. It was originated in and oriented to the structure mechanical problems on nonsmooth domains. Under the guidance of the regularity theory in the frame of countably normed spaces for the problems on nonsmooth domains (see 3,4,27,28]), the geometric mesh and the P-distribution of element degrees are properly designed, which leads to the exponenetial rate of convergence with respect to cubic root of the number of degree of freedom. The exponenetial convergence has been seen in numerous computations by using commercial and research codes mentioned above. The theory and algorithms have been generalized from elliptic boundary value problems to interface problems, eigenvalue problems, highorder problems, parabolic and hyperbolic problems (see 5,7,13,16,19,29,31,38,40]). Thus the h-p version in R2 has already been well established in 1980's. For survey of the 1

2 h-p version in R2 we refer to 8,11]. Although the computation and implementation of the hp version in R3 was started in later 1980's (see 1,12,20]), there had been no progress on the approximation theory until the regularity theory in terms of countably normed spaces for elliptic problems on nonsmooth domains in R3 was established in early 1990's (see 20,22,23,24]). There are several quite di erent features of the h-p version in three dimensional setting from those in two dimensional setting, due to the complexity of singularity at vertices and edges. First of all, the geometric meshes are designed di erently in the neiborhoods of vertices, edges and vertex-edges. Secondly, in addition to the bilinear or linear P-distribution of element degrees, a linear or uniform Q-distribution of polynomials of one variable (in the direction of edges) have to be adopted in edge-neiborhoods and vertex-edge neighborhoods in orfer to obtain the e ciency of computations. As a major development of the nite element method, in theory and practices, the h-p version of FEM in R3 involves the regularity theory of PDE on nonsmooth domain, the approximation theory of the h-p version, the parallel and iterative solvers for large-scale systems resulted from the h-p nite element discretization, implementation, applications to structural mechanics and engineering computation, etc.. This paper will focus on the regularity, approximation and algorithm.

ICOSAHOM 95 subset of L such that Lm = fij 2 L j Am 2 ?i \ ?j = ij g. We denote by !ij the interior angle between ?i and ?j for i; j 2 L. By H 1 ( ) we denote the usual Sobolev space furnished with norm
8 < X kukH k ( ) = :
0

j j k

kD ukL2(

);

91=2 =

2 A Model Problem
ij ij j m i

where = ( 1 ; 2; 3) and D u = ux1 1 x2 2 x3 3 and 1 H0 ( ) = u 2 H k ( ) j uj?0 = 0 . We now consider a problem on the polyhedral domain 8 in ; < ? u=f uj?0 = 0; (2:1) : u j = g; . ?1 n . where ?0 = i2D ?i and ?1 = i2N ?i where D is a subset of I and N = InD. We assume that f and g are analytic on and ?1 respectively, then the singularity of the solution is caused solely by the unsmooth domain. To illustrate e ectively our main ideas, we consider here the Poisson equation with analytic data except the domain, but the regularity and approximation results and the algorithms presented in this paper are applicable to general elliptic problems with piecewise analytic data (see 9,20,21,22,23,24]). 1 There exists a unique (weak) solution u(x) 2 H0 ( ) satisfying the variational equation 1 B(u; v) = F(v); 8v 2 H0 ( ) (2:2)
1 1 where B(u; v) = ru rvdx on H0 ( ) H0 ( ) and R R 1 F(v) = fvdx + gvdS on H0 ( ). ?1 1 Let SN H0 ( ) be a properly selected piecewise polynomial subspace, and uN be the nite element solution in SN satisfying B(uN ; v) = F(v); 8v 2 SN : (2:3)

Fig. 2.1. Polyhedral Domain Let be a polyhedral domain in R3 shown in Fig 2.1, and let ?i ; i 2 I = f1; 2; 3; ; I g be the faces (open), Am ; m 2 M = f1; 2; 3; ; M g be the vertices. By ij we denote the edges which are the intersections of ?i and ?j . Let Im be a subset fj 2 I j Am 2 ?j g of I for m 2 M. Let L = fij ji; j 2 I ; ?i \ ?j = ij g, and let Lm denote a

There holds the error estimates ku ? uN kH 1 ( ) C vinf ku ? vkH 1 ( ): 2S

N

(2:4)

Therefore, the accuracy of the nite element solution depends solely on the preciseness of the description of the regularity of the solution, and the selection of the subspace SN which must re ect exactly the natures of the regularity of the solution. To this end we shall decompose the domain into various subregions, on which we introduce

The h-p Version of FEM in R3

the countable normed spaces to precisely describe the reg- C2 ij (U) with ij 2 (0; 1) satisfying ularity of the solution and its derivatives of all orders, and properly design the geometric mesh and P-Q distribution of the element degrees to achieve the exponential rate of ( convergence. if ?i ?0; ?j ?1 ij = 2!ij ij 1 ? ij ; otherwise: !ij (3:3) 2

3 Regularity and Approximation in Neighborhoods of Edges

We assume that the edge st = fx = (0; 0; x3) j a < x3 < bg lies on the x3-axis, and introduce a neighnorhood of the edge st shown in Fig. 3.1, U = U ; ( st) = fx 2 j 0 < r(x) = dist(x; st) < ; x3 2 I = (a+ =2; b ? =2)g where ; 2 (0; 1) are selected such that U ; ( st) \ ?` = for ` 2 I and ` 6= s; t. By C2 ij (U) we denote a countably normed space with weighted Ck -norm which is a set of continuous functions u(x) on U such that for a real number ij 2 (0; 1) and any ,

For the proof of the theorem we refer to 23]. Note that the space C2 ij (U) is an anisotropic space which the solution belong to. In the edge-neiborhood the solution behaves very di erently in the direction parallel to the edge and the directions perpendicular to the edge. To achieve the best approximation by a piecewise polynomial we have to de ne mesh and element degrees accordingly.

kr(x)( ij +
and

1+ 2

?1)D (u(x) ? u(0; 0; x3))k 0 C (U)

3

First we divide the neighborhood U = U ; ( st) into K levels along the edge with a uniform height H, which is not kuX3 3 (0; 0; x3)kC0 (I ) Cd3 3! (3:2) neccessary be small and will not be reduced when further re ning the mesh. Then, according to the distance to the Hereafter ! = 3=1 ` !; d = (d1; d2; d3) and d = edge we divide U into n geometric layers. By i;j;k we ` 3 d i ; C 1 and d` 1 are independent of . `=1 i denote an element in the i-th layer and the k-th level with 1 j J(i; k) J (uniformly bounded with respect to i X3 and k). The element i;j;k are hexahedral, or pentahedral, or tetrahedral, with hl denoting the dimensions in the xl A2 directions, l = 1; 2; 3. Select a mesh factor 2 (0; 1), the /2 the geometric mesh Un = f i;j;k; 1 i n; 1 k 12 2 K; 1 j J(i; k)g over the neighborhood U satisfying
12 1
A4 X2

Cd ! (3:1)

/2

h1 = h2 = c n?i ; h3 = H 1; dist( 1;j;k ; st) = 0; c2 n?i dist( i;j;k ; st) c3 n?i for 1 < i n; and all j; k:
X1

(3:4)

Fig. 3.1. Edge-Neighborhood U ; ( st )

Theorem 3.1 The (weak) solution u(x) of (2.2) 2

where c` ; ` = 1; 2; 3 are some constants indenpendent of i; j; k: A typical geometric mesh Un is shown in Fig 3.2. For a precise description of the geometric mesh Un we refer to 6,9,20].

4
X1 X2

ICOSAHOM 95
where N = O(n4) is the number of the degree of freedom of S P;Q;1 (Un ), bst depends on ; and st but not on N .

For the proof of the theorem we refer to 9].

1

12
0

X3

3 2

Fig. 3.2. Geometric Mesh Un We now de ne nite element space over the geometric mesh Un S P;Q (Un) = f (x) j (x)j i;j;k = 1(x1 ) 2(x3 ); 1(x) is a polynomial of degreepi;j;k ; and 2(x3 ) is a polynomial of degree qi;j;k in x3 g and S P;Q;1 (Un ) = S P;Q (Un) \ H 1(U). A linear P-distribution fpi;j;k; 1 i n; 1 k K; 1 j J(i; k)g and a uniform Q-distribution fqi;j;k; 1 i n; 1 k K; 1 j n J(i; k)g should be associated with the geometric mesh U , with pi;j;k = i] for all i; j; k (3:5) and qi;j;k = n] for all i; j; k (3:6) where > 0 is a degree factor. Hereafter a] denotes the smallest positive integer a. Then the combination of the geometric mesh and linear-uniform distribution of element degrees leads to the exponential convergence. Theorem 3.2 Let u 2 C2 st (U), and let Un be the geometric mesh de ned by (3.4), and P -Q distribution be a
linear-uniform distribution de ned by (3.5)-(3.6) with the degree factor satisfying

Remark 3.1 Algorithms combining the geometric mesh and linear-uniform distribution of element degree achieve the optimal convergence and the best e ciency of computations, because this combination re ect exactly the nature of singularity of the solution in the edge-neighborhood U. Algorithms, associated with meshes which is not re ned geometrically along the edge, is never able to reach the exponential convergence and computational e ciency. If a uniform P -distribution with pi;j;k = qi;j;k = n] is associated with the geometric meshes, the exponenetial convergence may hold but with much smaller bst in (3.8), which will severely a ect the e ciency of the computation of nite element solution. Remark 3.2 The design of long element
i;j;k near the edge re ect the fact that the solution is analytic along the edge. The polynomial 2 (x3) of high degree qi;j;k = n] is used to approximate u(0; 0; x3) e ectively. Although the degree of 2(x3) is higher than the degree of 1(x), the cost of computation for these polynomials 2(x) in x3 direction is very minor, and it can be ignored comparing the cost of computations for those polynomials 1(x). An alternative approach to use of high-degree 2(x3 ) is uniform re nement in the x3-direction when the meshes are geometrically re ned in the x1 -x2 plane, but it will a ect the computational e ciency and signi cantly increase the cost of computation.

st ) ln(1= > (1 ?ln(1=F ) ) (3:7) H (1 ? 1? where FH = 0 min1 (1 + )1+ ( dH) , d = max di . The Let Am be located in the origin. Then a neighborhood i ) O (Am ) of the vertex Am is de ned by there exists (x) 2 S P;Q;1 (Un) such that (3:8) ku ? (x)kH 1 (U) ce?bstN 1=4 O (Am ) = fx 2 j 0 < (x) = dist(x; Am) < g

4 Regularity and Approximation in Neighborhoods of VertexEdges

The h-p Version of FEM in R3

X3 13

5 kl. ( ; ; ) is the spherical coordinate with respect to Am and st. Similarly we introduce a countably normed space C2 m;st (V) with weighted Ck -norms, which is a set of continuous functions u(x) on V such that for a pair of real number m;st = ( m ; st); 0 < m < 1=2; 0 < st < 1 and for any ,

3 1 2 23

A1

;2 m;st (x)D

(u(x) ? u(0; 0; x3))kC 0 (V) cd !

ij + 1 +
2

Fig. 4.1. Vertex-Neighborhood O (Am ) k (x) m + 3 ?1=2(u(0; 0; x3)?u(0; 0; 0))x3 3 kC 0 (I ) cd3 3 3 !: (4:2) where 2 (0; 1) is selected such that O (Am ) \ ?l = for l 2 Im (see Fig 4.1). We further decompose O (Am ) = into an inner neighborhood and several neighborhoods of vertex-edge. To this end we assume the edge st; st 2 Lm lies in the positive x3-direction, and introduce a vertexTheorem 4.1 The (weak) solution u(x) of (2.2) belongs edge neighborhood shown in Fig 4.2 to C2 m;st (V) with st 2 (0; 1) satisfying (3.3) and m 2 V = V ; (Am ; ij ) = fx 2 O (Am ); 0 < < g (0; 1=2) satisfying
X3 12 12

12

;2 with m;st (x) = (x) m +j j?1=2 sin( (x)) for I = (0; ) and 3 0

?1 and

m > 1=2 ? m ;

1 q1 + 4 (m) ? 1 m=2 1

(4:3)

X2

( where 1m) is the smallest positive eigenvalue of the Laplace-beltramioperator on the polygon S , which is a portion of the unit sphere subtended by an in nite cone which coincides with in the neighborhood O (Am ). 2

1 A1

For the proof of the theorem we refer to 24]. The spaces C2 m;st (V), like the spaces C2 st (U), is an anisotropic space which the solution u(x) belongs to. The solution behaves in this neighborhood very di erently in X1 the direction of the edge and in the direction perpendicular to the edge. Furthermore, note that u(0; 0; x3) is no longer Fig. 4.2. Vertex-Edge Neighborhood V ; (Am ; st) analytic, instead, it belongs to a countably normed space where is the angle between st and radical from Am on an interval I = (0; ). This features of the regularity (the origin) to x, 0 < ; < 1 such that V ; (Am ; st) \ of solution must be fully considered when we design the V ; (Am; Lk;l) = Am for any st 2 Lm and kl 2 Lm ; st 6= mesh and distribution of element degrees.

V, (A1, 12 )

6
X3 1 12
3

ICOSAHOM 95

and
2
/6

bilinear-linear de ned by (4.5) and (4.6) with and satisfying and

qi;j;k = k] for all i; j; k (4:6) where p0 0 is a properly selected integer, ; > 0 are the degree factors. Theorem 4.2 Let u 2 C2 m;st (V), and let Vn be the geometric mesh de ned by (4.4) and P -Q distribution be > (1 ? st ) ln(1= ) ln(1=F1) (4:7)

> (1 ? m ) ln(1= ) (4:8) ln(1=F1) where F1 is the value of FH at H = 1 given in (3.7). Then, there exists a (x) 2 S P;Q;1 (Vn ) such that ku ? (x)kH 1 (V) ce?bm;st N 1=5 (4:9) Fig. 4.3. Geometric Mesh Vn The neighborhood V is divided into n geometric level where N = O(n5 ) is the number of degree of freedom, bm;st 2 according to the distance to the vertex Am and n geometric depends on ; m ; st; and , but not on N . layers according to angular distance to the edge st. The For the proof of the theorem, we refer to 9]. elements i;j;k located in i-th layer and k-th level with 1 j J(i; k) J (uniformly bounded with respect to Remark 4.1 The geometric meshes Vn are re ned both i; k) are hexahedral, or pentahedral or tetrahedral. Let 2 in and in , and P -distribution is bilinear with re(0; 1) be a mesh factor, and let hl denote the dimensions spect to the layer number i and level number k, because of element i;j;k in the xl direction, l = 1; 2; 3. Then the the solution possesses two di erent types of singularities: edge-singularity and vertex-singularity in the neighborhood geometric mesh Vn = f i;j;k; 1 i n; 1 k n; 1 V ; (Am ; st), which have been completely exposed in Thej J(i; k)g satis es orem 4.1. The exponential convergence and the e ciency h1 c1 2n?k?i; h2 c1 2n?k?i; h3 c2 n?k ; of computation are achieved only by those algorithms usdist( i;j;1; Am ) = 0; dist ( 1;j;k ; st) = 0; ing a proper combination of geometric meshes and bilineardist( i;j;k ; Am ) = c3 n?k ; for k > 1 and all i; j; linear distribution of element degrees. dist( i;j;k ; st) = c4 n?i; for i > 1 and all k; j; (4:4) Remark 4.2 Asymptotically the exponential rate with rewhere dist ( i;j;k ; st) = min (x) min sin (x) is spect to N 1=5 is the best approximation result we can prove, x2 i;j;k x2 i;j;k due to the re nement in two directions. But for practical an angular distance between the element i;j;k and st. A range of n e.g., n < 10, the exponential rate with respect geometric mesh Vn is shown in Fig. 4.3. For the precise to N 1=4 is possible by select suitable p0 , e.g., p0 = n. For description of geometric mesh on V we refer to 6,9,20]. details, see 9]. The corresponding nite element spaces over the geoRemark 4.3 The solution along the edge is not ananlytic, metric mesh Vn is de ned as instead, u(0; 0; x3) belongs to a countably normed space S P;Q (Vn ) = f (x) j (x)j i;j;k = 1(x) + 2(x3 ); over an interval. Hence, in order to achieve the exponen1 (x) is a polynomial of degree pi;j;k and tial convergence and e ciency of computation, the re ne2 (x3) is a polynomial of degree qi;j;k in x3g ment in x3-direction has to be carried out geometrically and a linear distribution of element degrees has to be adopted P;Q;1 (Vn ) = S P;Q (Vn) \ H 1 (V). and S for the polynomials 2 (x3) so that u(0; 0; x3) can be approxA bilinear P-distribution fpi;j;k 1 i n; 1 k imated e ectively. The increase of the number of degree of n; 1 j J(i; k)g and a linear Q-distribution fqi;j;k 1 i n; 1 k n; 1 j J(i; k)g should be associated freedom and of the computational cost by 2(x3 ) is very minor and ignorable comparing the total number of degree with the geometric mesh Vn with of freedom and the total cost of computation of the nite (4:5) element solution. pi;j;k = i + k ? p0] for all i; j; k
2
3 1

The h-p Version of FEM in R3

7 are hexahedral, or pentahedral or tetrahedral. A geomet~ ric mesh On = f j;k; 1 k n; 1 j J(k)g with a mesh factor 2 (0; 1) satisfying h1 h2 h3 c1 n?i ; dist( j;1 ; Am ) = 0; (5:2) dist( j;k ; Am ) = c2 n?i; for k > 1 and all j: ~ A geometric mesh On is showm in Fig 5.2. For the ~ precise description on the geometric mesh over O (Am ) we refer to 6,9,20].
13 23 1

5 Regularity and Approximation in Inner Neighborhoods of Vertices

We de ne an inner neighborhood of the vertex Am by excluding all vertex-edge neighborhoods V ; (Am ; st); st 2 ~ ~ Lm ; O = O (Am ) n st2Lm V ; (Am ; st), which is shown in Fig. 5.1. We assume that Am is located in the origin.
13
V, 1 , ) ( 23

X3
V (1 , ) 13
,

A1

23

V (1 , ) 12
,

~ Fig. 5.1. Inner Vertex-Neighborhood O (Am ) ~ The countably normed space C2 m (O) is de ned as a ~ set of continuous functions u(x) on O such that for a real number m 2 (0; 1=2) and any

Fig. 5.2. Geometric Mesh O ~ We de ne a nite element space over On by ~ S P (On) = f (x) j (x)j j;k is a polynomial of degree pj;k g ~ ~ ~ and S P;1 (On ) = S P (On) \ H 1(O). k (x) m +j j?1=2D (u(x) ? u(0; 0; 0))kC0 (O) cd !: ~ The character of singularity of the solution in the inner (5:1) vertex-neighborhood is also re ected in the designing of In this neighborhood the solution possesses only vertex- element degrees. Only a linear P-distribution fpj;k ; 1 singularity. k n; 1 j J(k)g is needed with pj;k = k]; for all j; k: (5:3) Theorem 5.1 The (weak) solution of (2.2) belongs to ~ C2 m (O) with m 2 (0; 1=2) satisfying (4.3). 2 where > 0 is the degree factor, because the solution behaves equally in all x`-direction and geometric mesh is For the proof of the theorem we refer to 24]. re ned in . 2 (U) and C2 Unlike the space C st m;st (V), the space The algorithms based on the geometric mesh and linear ~ C2 m (O) is an isotropic space, the solution is singular in P-distribution achieve the exponential convergence. , but behaves equally in the x` -direction, ` = 1; 2; 3. Due ~ ~n to this character of the singularity of solution, the mesh is Theorem 5.2 Let u 2 C2 m (O) and let O be the geometre ned in one direction, i.e., in . According to the dis- ric mesh de ned by (5.1) and the linear P -distribution be ~ with satisfying (4.8). Then there exists tance to the vertex Am we divide O (Am ) into n levels. de ned by (5.2) n P;1 (O ) such that ~ The elements in the k-th level with h` being the dimen- a (x) 2 S sions in the x` direction, ` = 1; 2; 3; 1 j J(k) J ku ? (x)kH 1 (O) ce?bm N 1=4 (5:4) ~ (uniformly bounded with respect to k), denoted by j;k

12

~n

12

8
where N = O(n4) is the number of degree of freedom, bm depends on m ; and , but not on N . 2

ICOSAHOM 95
domain is completely described by the local asymptotic expansion of singular functions, which contains the power of (x); r(x) and sin( (x)) (see 14,15,17,18,30,32,33,36, 37]). If the power and forms of leading singular functions are known, special element can be constructed. The algebraic rate of convergence of the conventional nite element solutions may be improved because of the special elements, but the singular basis functions will distroy the nice band structure of the sti ness matric and deteriorate the condition number. Moreover, in practice the strength of singularity and form of singular functions are unknown, additional e orts and computations for these information are required (see 12]). Hence practically and theoretically the h-p nite element solution are the only e ective and e cient numerical approach for the problems on nonsmooth domains, and the regularity theory in the frame of countably normed spaces is the only theory which is able to guide this computational practice and to lead to the exponential rate of the convergence.

Remark 6.2 The regularity of solution on the polyhedral

For the proof of the theorem we refer to 9].

6 Regularity and Approximation of the h-p Finite Element Solution on Polyhedral Domain
Theorem 6.1 The (weak) solution of (2.2) is ananlytic
on
0,

Let 0 = n m2M O (Am ) n st2L ; ( st ). Select and properly, will contain no edges and vertecies of the domain. 0 is called regular region in which the solution has no singularity and is analytic.
and for any

kD u(x)kC 0 (

0)

cd !

We neext design the h-p version nite element algoFor the proof we refer to 24]. rithms on whole polyhedral domain. By we denote a multi-index ( m ; st ; m 2 M; st 2 L), Since the solution u(x) is analytic on 0, we use a xed with m 2 (0; 1=2) and st 2 (0; 1), and by C2 ( ) we and coarse mesh 0 = f 0; 1 ` Lg: The elements H ` de ne a countably normed space with weighted Ck - norm, 0 's are hexahedral, pentahedral, or tetrahedral with h` ` namely, for u 2 C2 ( ), there holds H 1 being their dimensions in the x` -directions, 1 ` 3. A uniform P-distribution fp` = p; 1 ` Lg with i) u 2 C0( ); p coinciding with the highest degree used in elements in ii) u jU ; ( st) 2 C2 st (U ; ( st )); singular neighborhoods, is associated with the mesh 0 , H and the nite element space is de ned by ii) u jO (Am ) 2 C 2m (O (Am )); ~ ii) u jV ; (Am ; st ) 2 C 2m;st (V ; (Am ; st)); S P ( 0 ) = f (x) j (x)j 0 is a polynomial of degree p` = pg H ` ii) u j 0 2 C1 ( 0); and jD u(x)j cd !: and S P;1 ( 0 ) = S P ( 0 ) \ H 1( 0 ). H H Note that the accuracy of the nite element solution in Combining Theorem 3.1, 4.1, 5.1 and 6.1, we now have the this region is achieved by uniformly increasing the polynoregularity of the solution on whole polyhedral domain. mial degree, but not by reducing the element sizes. It is well-known that for the analytic function u(x) on 0 , there Theorem 6.2 The (weak) solution u(x) of (2.2) belongs to C2 ( ) with m 2 (0; 1=2); m 2 M and st 2 (0; 1); st 2 is a polynomial (x) 2 S P;1 ( 0 ) such that H L satisfying (3.3) and (4.3). 2 ku ? (x)kH 1 ( 0 ) ce?b0 N 1=3 (6:1)

Remark 6.1 The regularity results can be given in terms

of countably normed spaces with other types of weighted norm and in various coordinate systems, for instance, the spaces B2 ( ) in weighted Sobolev norms. For the purpose of numerical approximation we prefer to the description in terms of the countably normed spaces with weighted Ck norms. The regularity theorems in these countably normed spaces and the relations of these spaces have been given in 22].

where N is the number of degree of freedom, b0 is independent of N. ~ Let n be the union of all geometric meshes Un ; On ; Vn 0 , and let P-Q distribution of element and uniform mesh H degree on n be the union of the linear-uniform P-Q distributions on Un 's, the bilinear-linear P-Q distributions on Vn's, the linear P-distribution on On's and the uniform P-distribution on 0 . Further, by S P;Q ( n ) we denote 0 H

The h-p Version of FEM in R3 the nite element space over n of piecewise polynomi- i) On the faces ACNE, ABDC, AEFB, IJLM, ~ the traction T=(T1 , T2, T3)=0; als whose restrictions on Un ; On ; Vn and 0 belong to H P;Q (Un ), S P;Q (Vn ), S P (On ) and S P ( 0 ), ii) On the face DKLMNC, the displacement u1 = 0, ~ the spaces S H 1 the traction T2 = T3 = 0; respectively, and S P;Q;1 ( n ) = S P;Q ( n) \ H0 ( ). iii) On the face NMJHFE, the displacement u2 = 1, We now come to the conclusion of the approximation of the traction T1 = T3 = 0; the h-p nite element solution. iv) On the face KGIL, the displacement u2 = 0, the traction T1 = T3 = 0; Theorem 6.3 Let the geometric mesh n and the P -Q distribution associated with n de ned above. Then the h- v) On the face GHIJ, the traction T1 = 1; T2 = T3 = 0; p nite element solution uN 2 S P;Q;1 ( n) converges to the vi) On the face KDBFHG, the displacement u3 = 0, the traction T1 = T2 = 0. solution u(x) of (2.2) exponentially,

ku ? uN kH 1 (

ce?bN 1=5

(6:2)

where N is the number of degree of freedom of S P:Q;1 ( n ), b depends on ; ; ; ; but not on N .

Proof Due to the de nition of S P;Q;1 ( n ) and the combination of (3.8), (4.9), (5.4) and (6.1), there exists (x) 2
S P;Q;1 ( n ) such that

w=o K D

(11,11,11)

ku ? (x)kH 1 ( ) where b = minfbst; bm; bm;st; b0g depending on ; ; ; but not on N, which together with (2.4) leads to (6.2) 2
spect to N 1=5 is the best accuracy of nite element solution which we can prove. Although the rate is with respect to N 1=4 in neighborhoods of edges and inner neiborhoods of vertices, and N 1=3 in regular region 0 , the majority of the number of degree of freedom is concentrated in the neighborhoods of vertex-edges. Hence, as a total performance, the h-p nite element solution converges at the exponential rate with respect to N 1=5. In computational practices, if the integer p0 in the bilinear P -distribution is properly selected and the number of layers and levels are not large, then the exponential rate with respect to N 1=4 is possibly achieveable, as mentioned in Remark 4.3, and also as seen in practical computations.

ce?bN

1=5

1
T=0

1
H T=0 F

11 1 y
L C

A T=0

1
E J

Remark 6.3 Asymptotically the exponential rate with re-

z 1
N M
(0,0,0)

1 11
v=0

u=0

11

Fig. 7.1. Polyhedral Domain

The

h-version:

7 Numerical Example
A benchmark elasticity problem on a polyhedral domain with the modulus E = 100 and Poisson ratio = 0:3, shown is Fig. 7.1, is computed by the h, p and h-p version of nite element method. Let u = (u1; u2; u3) be the displacement and T = (T1 ; T2; T3) be the traction on the boundary. The follwing boundary conditions are imposed to this problem:

Uniform meshes of cubic elements with size h = 1; 1=2; 1=3; 1=4, shown in Fig. 7.2, are used with the uniform degree p = 2. Table 7.1 shows the relative error in energy norm 9% when h = 1=4 and N 300; 000 and CPU time t1(s) = 29; 800sec (single processor). Table 7.1 Performance of the h-version (p = 2) h No. of Elm N kekE;R % 1 331 6708 20.53 1/2 2648 42679 13.32 1/3 8937 131744 10.56 1/4 21184 297735 8.99 t1(s) 120 1890 10130 29800

10

ICOSAHOM 95 better than the h-version, but on quasi-uniform mesh, the p-version converges only as twice fast as the h-version. We will make more comments later. Table 7.2 Performance of the p-version (on Mesh 3, n1 = 3, n2 = 2) p N=DOF kekE;R % t1 (s) 2 2925 43.84 33 3 5135 19.28 56 4 9328 8.21 169 5 15549 5.04 481 6 24404 3.78 1308 7 36496 3.01 3305 8 52428 2.49 7408 9 72803 2.11 14860 10 98224 1.84 31842 Fig. 7.2. Uniform Mesh with h = 1
The

The

p-version:

The tensor product meshes with the following combination of n1 and n2 are used: Mesh 1: n1 = 1, n2 = 2; Mesh 3: n1 = 3, n2 = 2; Mesh 5: n1 = 5, n2 = 2; Mesh 7: n1 = 7, n2 = 2; Table 7.3 Performance of the h-p version (on Mesh 1,3,5, and uniform degree p = n1) Mesh(n1 ) p N=DOF kekE;R % t1 (s) 1 4 464 31.39 1 3 5 15549 5.04 481 3 6 24404 3.78 1308 5 7 67567 1.84 13337 5 8 97374 1.50 29392
100 x o o log(||e||)(%) x
DOF CPU

h-p

version:

Fig. 7.3. Geometric Mesh n1 ;n2 with n1 = 2; n2 = 1

The nite element solutions of the p-version are com10 puted on various geometric meshes. The geometric meshes o x o x of tensor product type are used, namely the meshes are heavily re ned near the vertex A with layer number n1 o x o x and slightly re ned at other vertices with layer number 1 1 2 3 4 5 6 7 8 9 10 n2 , because the singularity of solution near A is severe. N^(1/5) (resp. t1^(1/5) ) The mesh factor = 0:15. A geometric mesh n1;n2 with Fig. 7.4. Performance of the h-p Version n1 = 2; n2 = 1 is shown in Fig. 7.3. The degree p of element uniformly increases from 1 to 10. The performance Associated with a uniform P-distribution with p = of the p-version on Mesh 3 with n1 = 3 and n2 = 2 is given n1]; = 1:5 2:0. The Table 7.3 shows that the in Table 7.2. The relative error in energy norm reduces to 1:84% of relative error in energy norm is achieved at 3% when p = 7 and CPU time t1 = 3305sec (single proces- n1 = 5; n2 = 2 and p = 7; N = 67; 567, and the CPU sor). The p-version perform on the geometric mesh much time t1 = 13; 337sec: On the other hand, the p-version on

The h-p Version of FEM in R3

11

Mesh 3 needs p = 10; N = 98; 224 CPU time t1 = 31; 842. in R3. The practical strategy to achieve the optimal conThe relative error of the h-p version in energy norm vergence and e ciency for engineering applications in R3 kekE;R v.s. N 1=5 and t1 are ploted in Fig. 7.4. The com- needs to be developted further. parison between the h, p and h-p versions are shown in Fig. 7.5 and Fig. 7.6 where the relative error v.s. N and CPU time are ploted in log-log scales.

Acknowledgements

100 + * log(||e||)(%) +x x 10 + * + * + +

x + *

h-version p-version h-p version

This work was partially supported by National Science & Engineering Research Council of Canada under grant OGP0046726. The paper re ects the joint research with Dr.I. Babuska of TICAM, University of Texas and Dr. B. Andersson of Aeronautical Institute of Sweden. The computation was done by Dr. B. Andersson with code STRIPE.

1 100

+ * + * 1e+05 1e+06

1000

1e+04 Log(N)

References
1] B. Anderson, I. Babuska, v.T Petersdor and U. Falk, Reliable stress and fracture mechanics analysis of complex components using h ? p version of FEM, Int. J. Numer. Meth. Eng., 38(1995). 2] I. Babuska, and B.Q. Guo, The h-p version of the nite element method with curved boundary, SIAM J. Numer. Anal., 24(1988), pp.837-861. 3] I. Babuska, B.Q. Guo, Regularity of the solution of
elliptic problems with piecewise analytic data, Part 1: Boundary value problems for linear elliptic equations of second order, SIAM J. Math. Anal. 19(1988), pp.

Fig. 7.5. Comparison between h, p and h-p versions: Error v.s. DOF
100 + * log(||e||)(%) + 10 x x + * + * + + x x x + *
h-version p-version h-p version

+ *

1 1 10 100 1000 Log(CPU) 1e+04

+ * 1e+05

172-203.

Remark 7.1 The computations on this benchmark prob-

Fig. 7.6. Comparison between h, p and h-p versions: Error v.s. CPU

4] I. Babuska, B.Q. Guo, Regularity of the solution of

lem has shown that the h-p nite element solution can achieve the exponential rate in engineering practical range, e.g., 3%, as predicted by asymptotic analyses. Remark 7.2 The computation of the h-p version in R3 have not been exhaustive enough, more experience in computations and implementations are needed, e.g., how to generate a geometric mesh, how to implement the P -Q distribution of element degrees (no existing 3-dimensional code has this feature yet), and what are the optimal mesh factors and degree factors, etc.. Remark 7.3 The computation has shown that the pversion on an over-re ned geometric mesh converges to the desired accuracy expotentially before entering the asymptotic phase, it could be an alternative to the h-p version

elliptic problems with piecewise analytic data, Part 2: The trace spaces and applications to the boundary value problems with nonhomogeneous boundary conditions, SIAM J. Math. Anal. 20(1989), pp. 763-781. element method for problems with non-homogeneous essential boundary conditions, Comp. Meth. Appl.

5] I. Babuska, B.Q. Guo, The h-p version of the nite Mech. Engrg., 74 (1989), pp. 1-28.

6] I. Babuska I., and B.Q. Guo, Approximation properties of the h-p version of the nite element method, Technical Notes BN-1177, IPST, University of Maryland, College Park, 1995, to appear in Comp. Meth. Appl. Mech. Engrg. and The State of the Art of the h-p Versipon of the Finite Element Method edited by Babuska, I. in the series of Computational Method for Applied Mechanics and Engineering.

12 7] I. Babuska, B. Q. Guo and J. E. Osborn, Regular26(1989), pp. 1534{1560.

ity and Numerical Solution of Eigenvalue Problems with Piecewise Analytic Data, SIAM J. Numer Anal.,

ICOSAHOM 95 19] B.Q. Guo, The h ? p Version of the Finite Element Method for Elliptic Equation of Order 2m, Numer. Math., 53(1988), 199-224. 20] B.Q. Guo, The h-p Version of the Finite Element

8] I. Babuska and B.Q. Guo, The h, p and h-p Version of pp.159-174.

The Finite Element Method, Basic theory and Application, Advances in Engineering Software, 15(1992),

9] I. Babuska and B.Q. Guo, The h-p version of the nite element method in R3 , 1995. 10] I. Babuska and T. Janik, The p-version of the the 6(1990), pp. 343-369.
nite element method for parabolic equations. Part II : The h-p version in time, Numer. Meth. PDE.,

11] I. Babuska and M. Suri, The p and h-p versions of the nite element method, basic principles and properties, SIAM Review, 36(1994), pp. 578-632. 12] I. Babuska, v. T. Petersdor and B. Anderson, Nu31(1994), pp. 1265-1288
merical treatment of vertex singularities and intensity factors for mixed boundary value problems for the Laplace equation in R3, SIAM J. Numer. Anal.,

M. Costabel, M. Dauge and S. Nicaise, pp.101{120, Marcel Dekker Inc., (1994). 21] B.Q. Guo, Optimal nite element approach for the elasticity problems on nonsmooth domains in R3 , in Computational Mechanics'95, Eds. S.N. Atluri, G. Yagawa & T.A. Cruse, Vol. 1, pp.427-432, Springer. 1995. 22] B.Q. Guo, and I. Babuska, Regularity of the soluTechnical Notes BN-1181, IPST, University of Maryland, College Park, 1995. To appear in Proceeding of Royal Society of Edingburgh. 23] B.Q. Guo, and I. Babuska, Regularity of the soluNotes BN-1182, IPST, University of Maryland, College Park, 1995. To appear in Proceeding of Royal Society of Edingburgh. 24] B.Q. Guo, and I. Babuska, Regularity of the solutions pear. 25] B.Q. Guo, and I. Babuska, The h ? p version of the
for problem on nonsmooth domains in R3, Part 3: Regularity in neighborhoods of vertices, 1955. To aptions for problem on nonsmooth domains in R3 , Part 2: Regularity in neighborhoods of edges, Technical tions for problem on nonsmooth domains in R3 , Part 1: Countably normed spaces on polyhedral domains,

Method for Solving Boundary Value Problems in Polyhedral Domains, in Boundary Value Problems and Integral Equations in Nonsmooth Domains, eds:

13] K.S. Bey and J.T. Oden, h-p version discontinuous Galerkin methods for hyerbolic conservation laws, 1994, to appear in Comput. Meth. Appl. Mech. Engrg.. 14] M. Costabel, and M. Dauge, General edge asymptotics pp. 109-155, 157-184.
solutions of second order elliptic boundary value problems I & II , Proc. Roy. Soc. Edinburgh 123A(1993),

15] M. Dauge, Elliptic Boundary Value Problems on Corner Domains, Lecture Notes in Math. No.1341, Springer, New York (1988). 16] L. Domhowicz, J. T. Oden, K. Rachowicz and D. Hardy, An h-p Taylor-Galerkin nite element method for compressible Euler equations, Comput. Methods Appl. Mech. Engrg., 88(1991), pp. 363-396. 17] I. Grisvard, Singularities des problemes aux limites dans polyh^dres, Expose no VIII, Ecole Polytechnique e Centre de Mathematics, France ,1982. 18] I. Grisvard, Elliptic Problems in Nonsmooth Domains, Pitman, Boston, 1985.

nite element method, Part 1: The basic approximation results, Comput. Mech., 1(1986), pp. 21{41. 26] B.Q. Guo, and I. Babuska, The h ? p version of the nite element method, Part 2: General results and applications, Comput. Mech., 1(1986), pp. 203{220. 27] B.Q. Guo, and I. Babuska, On the regularity of interface problems in frame of countably normed spaces, to

appear. 28] B.Q. Guo, and I. Babuska, On the regularity of elasticity problems with piecewise analytic data, Adv. in Appl. Math., 14(1993), pp. 307-347. 29] B.Q. Guo, and H.S. Oh, The h-p version of the nite element method for problems with interfaces, Int. J. Numer. Meth. Engrg., 37(1994), pp.1741{1762.

The h-p Version of FEM in R3

behaviour of elastic elds and boundary integral Mellin techniques near conical points, preprint, 1994. 31] C.Y. Lee and J.T. Oden, A posteriori error estimation of h-p nite element approximations of frictional contact problem, Comput. Meth. Appl. Mech. Engrg.,

13

30] V. Kozlov, W.L. Wendland, and H. Goldberg, The

113(1994), pp.11-45. 32] J. M.-S. Lubuma, and S. Nicaise, Dirichlet problems in polyhedral domains, I: Regularity of the solutions, Math. Nachr., 168(1994). 33] V.G. Maz`ya, and B.A. Plamenevsii, On the Coef-

Co., Los Angeles, CA, (1990). 35] PHLEX User's Manual, Computational Mechanics Co., Austin, TX, (1990). 36] v. T. Petersdor , Boundary integral equations for
mixed Dirichlet, Newumann and transmission conditions, Math. Meth. Appl. Sci., 11(1989), pp.185{213. 37] v. T. Petersdor and E. P. Stephan, Regularity of mixed boundary value problems in R3 and and boundary element methods on graded meshes, Math. Meth.

cients in the Asymptotics of Solutions of Elliptic Boundary Value Problems in Domains with Conical points, Amer. Math. Soc., Transl., 2(1984), pp. 57-88. 34] MSC/PROBE, User's manual, MacNeal Schwendler

Appl., 12(1990), pp.229{249. 38] A. Safjan and J.T. Oden, High-order Taylor-Galerkin and adaptive h-p methods for second-order hyperbolic systems: Application to elastodynamics, Comput. Meth. Appl. Mech. Engrg., 103(1993), pp.187230. 39] STRIPE Users Manual, The Aeronautical Research Institute of Sweden (FFA), Bromma, Sweden, 1992. 40] B. A. Szabo, The p and h ? p versions of the nite element methods in solid mechanics, Comput. Methods Appl. Mech. Engrg. 80(1990), pp.185{195.