univ new haven ct dept of computer science local … · 2014. 9. 27. · ad-aiis 432 yale univ new...
TRANSCRIPT
AD-AIIS 432 YALE UNIV NEW HAVEN CT DEPT OF COMPUTER SCIENCE F/S 12/1LOCAL-MESH, LOCAL-ORDER. ADAPTIVE FINITE ELEMENT METHODS WITH A--ETC(U)DEC 81 A WEISER N00014-76-C 0277
UNLSSIFIED TR-213 N
2N
11112.
111111.25 11111J.4 lI 8
MICROCOPY RESOLUTION TEST CHARTNAT IONAL BURiEAU Of STANOARDS 1963-A
LCID ~
or~ Pulilc asLU D ti~uon Ufmja
YALE UNIVERSITYSDEPARTMENT OF COMPUTER SCIENCE
82 04 'RO 127
Local-Mesh, Local-Order, Adaptive Finite Element
Methods with A Posteriori Error Estimators
for Elliptic Partial Differential Equations
Alan Weiser"
Technical Report #213
December, 1981
IDepartment of Computer Science, Yale University, New Haven, CT 06520
This research was supported in part by ONR Grant N00014-76-C-0277, !SU-ONRGrant FI.N00014-80-C-0076, the Department of Energy (DOE Grant DE-AC02-77ET53053),The National Science Foundation (Graduate Fellowship), and Yale University.
" i o V.IM.
A3STACT
Local-esho Local-Order, Adaptive Finite ElementMethods with A Posteriori Error Estimatorsfor Elliptic Partial Differential Equations
Alan Veiser
Yale University. 1981
The traditional error estimates for the finite element solution of
elliptic partial differential equations are a priori, and little
information is available from them about the actual error in a specific
approximation to the solution. us, in the traditional finite element
method, the choice of discretization i eft to the user, who must base
his decision on experience with similar equations.
- In recent years, locally-computable a posteriori error estimators
have been developed, which apply to the actual errors committed by the
finite element method for a given discretization. These estimators lead
to algorithms in which the computer itself adaptively decides how and
when to generate discretizations. So far, for two-dimensional problems.
the computer-generated discretizations have tended to use either local
mesh refinement, or local order refinement, but not both.
\n this thesis, we present a new class of local-mesh, local-order,
square finite elements which can easily accomodate computer-chosen - ,-_
I, I .io, IIl~
I4 .4I
[~~0 i ... .. ... .. .__ . .. 1- I
- discretizations. We present several now locally-computable a posteriori
error estimators which, under reasonable assumptions, asymptotically
yield upper bounds on the actual errors omlitted, and algorithms in
which the computer uses the error estimators to adaptively produce
sequences of local-mesh, local-order discretizations. We present
theoretical and numerical results indicating the ex c d and actual
behavior of our methods.
I
Table of -Cote.ts .
CAPR1: Introduotion ....... . .. . . 1
CHAPTR2: Local mosh refinement .. . .. . . . 6
2.1 Introduction . . . . . . . . . ..... 62.2 Terminology . . ........... . 72.3 The 1-irregular rule and some of its consequences . . 112.4 Alternative mesh refinement rules . . . . . . . 172.5 A mesh data structure . . . . . . . . . . . 21
CHAPR 3: Local order refinement ... ....... 25
3.1 Introduction ...... .............. 253.2 The Lagrange case ..... ............. 263.3 The Hermite case . . . ............. 29
3.4 A mesh data structure for local order refinement . 29
CHAPTER 4: A posteriori error estimators , . , ..... 33
4.1 Introduction . . . ......... . 334.2 Terminology ...... .............. . 354.3 A Dirichlet a posteriori error estimator ..... 404.4 A Neumann a posteriori error estimator ...... 424.5 Properties of the Neumann error estimator ..... 444.6 Alternative error estimators . . . .... . S1
4.6.1 Ignoring irregular corners . .. .. 514.6.2 Using the matrix for a local Poisson problem . 524.6.3 Using a smaller matrix . . . . . 54.6.4 Using the residuals and Jumps in normal derivatives
574.7 Extensions . . . . . . . . . . . . . . . 604.8 Summary . . . . . . . . . . . . . . . . 64
CHAPTER 5: Heuristic refinement criteria . . . . . . . . 67
5.1 Introduotion . . . . . . . . . . . . . . 675.2 A model of work . ........... 685.3 A model of predicted error behavior ... ... 705.4 A symetized model of predicted error behavior . . . 765.5 Choosing the refinement cutoff . . . . . . . . 785.6 Drawbacks ... . . . . . . . . . . .. 79
-iv-
5.7 Expected error behavior . . . . . . . . . . . 815.7.1 A smooth solution . . . . . . . . . . 825.7.2 A point singularity . . . . . . . . . . 835.7.3 A line singularity . . . . . . . . . . 84
CfhAFM 6: Computational aspects . . • . . . . . 85
6.1 Introduction . . . . . . . . . . . ... 856.2 Assembling the linear system . . . . . . . .. 876.3 Solving the linear system . . . . . . . . . . 90
6.3.1 A smooth solution . . . . . . . . . . 906.3.2 A point singularity . . . . . . . . . . 926.3.3 A line singularity . . . 9 . . . . . . 946.3.4 Disoussion . . . . . . . . . . . . 95
6.4 Computing the error indicators 96
C1APTM 7: Numerical results . . . . ..... . . . 101
7.1 Introduotion .. ............ 1017.2 The problem set ............. 1017.3 Error behavior . . . . . . . . . . . . . 1047.4 Error estimator behavior . . . . . . . . . . 1077.5 Error estimator convergence . . . . . . . . . 108
List of Figures
2-1: A bisection-type loeally-refined noes . . . . . . 6
2-2: Lagrange versus tonsor-product basis functions .. . . 11
2-3: Configurations of eleensts in su3p(B(V)) .. . .. . 13
2-4: f(E) .. . . . . . ..... 14
2-5: Minimum and alnum N(Y) for interior vertices .... 15
2-6: 1 in 1(L+1) . . ... . . 16
2-7: SinSularity at P - (2/3.2/3). IlX - 4 . . . . . . . 19
2-8: A modified basis function ... . . ..... . 19
2-9: LUX - 4: alternative nosh refinmeont rules . ., . . . 22
2-10: Nodes in ICT and IVURT .. ...... ... 23
3-1: A side basis function ............ 28
3-2: Losal-order basis funetions .......... 30
3-3: Basis functions mapped to vortiees ........ 31
4-1: Ignoring the irregular crners ......... 52
5-1: Early order refinement penalty . . . ...... 81
6-1: An element with one irregular corner ....... 89
6-2: Nested dissection ordering with a smootk solution . . . 91
6-3: Reosrsive ordering vith a point singularity . . . . . 93
6-4: Nested-disseotion-type ordering with a line singularity . 95
7-1: Problem - tru erro . . ..... . ., . 111
7-2: Sample grid for problem 1 (method +, residuals and Jumps) 112
-vi-
7-3: Problem1 ,_.approxhiito error , . * . . . . . , 113
7-4: Problem 1 - residuals ad Jumps .... . . . . . 114
7-5: Problem 2 - tru error ,. . . * . * 115
7-6: Sample $rid for problem 2 (method . residuals sad Juva) 116
7-7: Problem 2 - approximate error ..... . . . o . 117
7-8: Problem 2 - residualssadJws * * * . * *. 118
7-9: Problem 3 - true error . . . * . , ..... 119
7-10: Sample grid for problem S (method 2. residuals asd Jumps) 120
7-11: Sample grid for problem 3 (method . residuals and Jumps) 121
7-12: Problem 3 - approximate error . . . . . . . . 122
7-13: Problem 3 - residuals sad Jumps .. .. .. . . . 123
7-141 Problem 1 - approzimate error . . . . . . . . . 124
7-15: Problem 1 - residuals sad Jumps . * . , . . . . 125
7-16: Problem 2 - approximate error . .. . .. . 126
7-17: Problem 2 - residuals andJups . ...... * a 127
7-18: Problem 3 - approximate error . . . . . . . . 128
7-19: Problem 3 - residualsaadJumps . .... . . . . 129
List of Tables
7-1: Test problems S . . . 102
7-2: Refinement metkeds 0 . 0 0 . 0 . 0 0 * a 0 0 103
7-3: Iffiaiezy sunary .. a 0 o 0 00 0 0 * 107
7-4: Error estimator behavioz 0 110
List of "yols
0, so 7
Il, z E Iy, 1 8
M 8
admissible mesh 8
(iW)regular vertex 9
k-irregular mesh 9
S, BI (piecovise bilinear) 10
1-irregular rule 11
f( ) 12
k26
Pi(x), pi(z) pa(y) 26
S, BI (local-order) 27
k 28
% (.,.k), 1111 is36
a* 36
-ix-
.,'
7 W!dl
Symbol -_.lu
SE 37
SE ' SE-S97 SI (.), I(.) 3E
N(E), ya) 39
up U, 0. U 39
, e 39
y 40
h 40
FE(.), [aaU/n], FE(-) 42
E 43
*I 44
koax 47
5o
~so
S"E(., I11 , (E) 52
! ._ 53
~ss-E -
55
so AE 60
-P., P.5 65
-: .... ________ --. . '- '- .-. ..... . . ... .. . . ..
z 69
MO.) p(z) (p.). v(p(z)) (WE) 69
m(z) , p(z) 70
C(;), G(z.P) 71
4B , 471
(h) ()72
4E 73
PCUT 7
~CUT 78BE,(B' 86
IE, IE (B,) 86
G(E) . ws is t 8
vL 89a ,
w 99p.
CN.FlR 1
Introduotion
The traditional error estimates for the finite element solution of
elliptic partial differential equations are a priori, and little
information is available from them about the actual error in a specific
approximation to the solution. Thus, in the traditional finite element
method, the choice of discretization is left to the user, who must base
his decision on experience with similar equations.
In recent years, locally-computable a posteriori error estimators
(e.g., Babutka and Rheinboldt [8]) have been developed, which apply to
the actual errors committed by the finite element method for a given
discretization. These estimators lead to algorithms in which the
computer itself adaptively decides how and when to generate discreti-
zations. So far, for two-dimensional problems, the computer-generated
discretizations have tended to use either local mesh refinement (e.g.,
Babulka and Rheinboldt [10], Bank and Sherman [13]), or local order
refinement (Babulka, Szabo, and Katz [111), but not both.
In this thesis, we present a new class of local-meeh, local-order,
1
-.. -4, vv ~mU~ d bud e vie
2
square finite elements which can easily accomodate computer-chosen
discretizations. We present several new locally-computable a posteriori
error estimators which, under reasonable assumptions, asymptotically
yield upper bounds on the actual errors comitted, and algorithms in
which the computer uses the error estimators to adaptively produce
sequences of local-mesah, local-order discretizations. We present
theoretical and numerical results indicating the expected and actual
behavior of our methods.
We begin in Chapter 2 by considering bisection-type local mesh
refinement with square elements. We present the class of f1-irregular"
meshes, a variant of a class of meshes suggested by Bank and
Sherman [14], such that any bisection-type locally-refined mesh can be
converted into a 1-irregular mesh with no more than a constant times as
many elements, and the resulting finite element matrix (with bilinear
basis functions) has no more than a constant number of nonzeroes in any
row. Conversely, we show that there are simple bisection-type
locally-refined meshes with essentially dense finite element matrices.
We discuss simple data structures for handling 1-irregular meshes.
In Chapter 3, we extend the finite element spaces considered in
Chapter 2 to include basis functions with locally-varying polynomial
orders.
In Chapter 4, we develop new a posteriori error estimators. We
present a new error estimator which has two main advantages over the
3
estimator presented in Babulka and Rheinboldt [8]:
1. Under reasonable assumptions, the new error estimator is
asymptotically an upper bound on the norm of the true error.
2. The new error estimator can be computed locally in each
element. (The error estimator in [8] must be computed over
more complicated local regions.)
Under suitable assumptions, we show that, like the estimator in [8], the
new error estimator is no larger than a constant times the norm of the
actual error. We also present several cheaper error estimators, and
discuss their properties. Under suitable assumptions, Babulka and
Miller [4] have recently shown that the error estimator used in their
code (with piecewise bilinear basis functions) converses to the norm of
the true error. We have not yet proven convergence for our error
estimators.
In Chapter 5, we develop several heuristic models of error behavior
which lead to adaptive refinement procedures (of. Brandt [16], Babulka
and Rheinboldt [71). We present the asymptotic expected error and work
behavior for three problem classes consisting of problems with:
1. smooth solutions.
2. solutions with point singularities.
3. solutions with line singularities.
Except in the presence of line singularities, the expected error
converges to zero faster than inverse polynomially with respect to work.
4
In Chapter 6, we discuss some of the computational aspects of the
methods presented in Chapters 2-5. We consider efficient assembly of
the finite element system (of. Bisenstat and Schultz [181, Weiser,
Eisenstat, and Schultz [31]), and efficient construction of the error
estimators. We present the asymptotic complexity of nested-dissection-
type Gaussian elimination for the three classes of problems considered
in Chapter 5. The operation counts and storage for the problems with
sinsularities are linear (optimal-order) in the number of elements N, in
contrast to the operation counts and storage for uniform meshes, which
are proportional to N3 1 2 and N los(N) respectively (e.g., George [231).
In Chapter 7, we present numerical results obtained using prototype
codes. We present the resulting error and error estimator behavior for
problems from the three classes considered in Chapter 5. The error
estimators are usually accurate to within a factor of two. In some
cases, the error estimators appear to converge to the norm of the true
error as the mesh size decreases.
I an deeply grateful to my advisor, Professor Randolph E. Bank, for
suggesting the topic of this research and working closely with me on its
realization. I am honored to know him and to have worked with him. I
would like to thank my other advisors, Professors Martin N. Schultz and
Stanley C. Eisenstat, for helping to guide my research, and for their
unfailing intelligence, integrity, and support.
Many other past and present members of the numerical analysis group
at Yale have contributed much to my enjoyment of graduate school and my
appreciation and understanding of numerical analysis. They are (in
order of appearance at Yale) Andy Sherman, John W. Lewis, inti Chandra,
Rob Schreiber, lack Perry, Dave Fyfe, Trond Steihaug. Craig Douglas,
Howie Elman, Zen Jackson, Tony Chan, John Kindle, and Denren Zhu.
I an grateful to many other members-at-large in the numerical
analysis comunity. I = especially grateful to Professor Mary
F. Wheeler, for her encouragement and guidance when I was an
undergraduate, and Professor Ivo Babulka, for his excellent and
influential work.
I a grateful to the Office of Naval Research CONK Grant
N00014-76-C-0277, FSU-ONR Grant Fl .N00014-8O-C-0076). the Department of
Energy (DOE Grant DE-ACO2-77ET5353), the Notional Science Foundation
(Graduate Fellowship), and Yale University for their financial support.
Most of all, I -grateful to my family, and my loving wife Mary
Ann.
CRLAiPTm 2
Looal mesh refinement
2.1 Introduction
In this chapter, we consider bisection-type local mesh refinement
with square elements. A saiple mesh of this type is shown in Figure
2-1.
I I I I lI I I l II I I I I-- 0----- -.. 0--0
SI I I III II I -- o---.I I I I I I
0- .-- 0--- •-V--OI I II I Il I II I II I II I I0-- 0- 0
Figure 2-1: A bisection-type lovally-refined mesh
This type of mesh refinement has been intensively investigated in
recent years, mainly by Babulka and Rheinboldt and their students
t6
7
(e.g. [4]). The data structures and computational algorithms required
to implement the finite element method with general noshes of this type
are fairly complicated (see Rheinboldt and Mesztenyi [251 and
Gannon [21]). Thus, Bank and Sherman [14] and Van Rosendale [30]
suggest using restricted classes of meshes which can be implemented with
much simpler data structures.
In Section 2.2, we present some standard terminology for dealing
with bisection-type local mesh refinement, and define the class of
"l-irregular" meshes, a variant of a class of meshes suggested by Bank
and Sherman [14]. In Section 2.3, we consider a variant of a mesh
refinement rule suggested by Bank and Sherman which can convert any
bisection-type locally-refined mesh into a 1-irregslar mesh with no more
than a constant times as many elements, such that the resulting finite
element matrix (with bilinear basis functions) has no more than a
constant number of nonzeroes in any row. Conversely, in Section 2.4. we
show that there are simple bisection-type locally-refined noshes such
that the resulting finite element matrices are essentially dense. We
also consider alternative mesh refinement rules. In Section 2.5, we
present data structure details for handling 1-irregular meshes.
2.2 Temiznology
We now oonsider bisection-type local mesh refinement with square
elements. For simplicity, we take the domain to be the unit square,
B := C0,1]X[01.] = ((Z.y): 01z.l, 0lyll),
!1
with boundary 8Q. The square region B - [xE, xe+h]X[yzyi+hh] is called
an element. The elements
[xE xE+hE/2 lX [y+/ 2 ,y+yh], [£+h/2oE+h£]X[y+hI/2,y£+N],
axExe+hE/2]X[yE.yE+hE/2]. [zx+hE/2,z£+hEX[yEye+h/21
are called the sons of E, and E is called the father of its sons. A
o M is defined to be an unordered set of elements. An element in K
is called refined if its four sons are in N. and unrelied otherwise.
The class of bisection-type locally-refined noshes with square
elements, or admissible noshes (Babulka and Iheinboldt [8. 101), is
recursively defined by the following two rules:
1. M) is an admissible mesh.
2. If a mesh N is an admissible mesh, and E is an unrefined
element in N, then the mesh (N, the sons of El is an
admissible mesh.
The level L of R is defined to be the number of generations between
Q and B. Thus, in an admissible mesh on B, hE - (1/2)L . A noizhbor of
£ is defined to be another element at the same level as E that shares a
side with B.
A corner of an unrefined element is called a vertex. A vertex
which it a corner of each unrefined element it touches is called a
9
xauar vertex. All other vertices are called igula . The
.irULILity iadex (Babulka and Rheinboldt [101) of a mesh is defined to
be the maximum number of irregular vertices on a side of an unrefined
element. We ocall a mesh with irregularity index I k a k-ieulaz nesh.
A (regular) vertex on 80 is called a boundary vertex. All other
vertices are called interior.
In figures depicting meshes,
0 denotes a regular verteg.
o denotes a vertex (r,'gplsr t irregular),
* denotes an irregular vertex,
(v) at a vertex denotes function value.
For example, Figure 2-1 depicts an admissible 2-irregular mesh with 4
refined elements, 13 unrefined elements (one labelled K), 12 regular
boundary vertices, 5 regular interior vertices, and 6 irregular vertices
(one labelled V).
This is the name used by Babuaka and Rheinboldt [10). Bank and
Sherman [141 use the name "green vortex". Gannon [211 uses the name
"inactive vertex".
10
Let M be an admissible mesh. A function S is called viecewise
bilinear on M if, on any unrefined element E in M,
S(x.y) s(zE.yE) (xB+hZ-x)(yE+hE-y) /E
+ g(zhEyE) (xE ) (yE+hey) S h
2+ S(xEoyE+hE) (xE+h -x)(y-YE) /I
+ S(zE+hEyE+h) (xxE)(yyE) / 2
Let the finite element space S (= S(M)) be the space of continuous
piecewise bilinear functions on M. Let the regular vertices for I be
Vl.....VN . Let B (x,y) (= B(VI)(xy)) be the function in S which
satisfies the Lagrange conditions
BI(VJ) - ij f1 if I j
0 if I #,
for J'1,...,N. BI is well-defined using induction on elemnet level.
Moreover, (BI) is a basis for S, since each BI is in S. the BIIs are
linearly independent, and each function in S is a linear combination of
the Bi'a.
The suipozt of B., supp(B1 ), is defined to be the union of all
unrefined elements E such that BI is not identically zero in R. In
general, the support of BI may be non-convex or non-simply-connected.
Alternately, the standard tensor-product basis functions {1I may be
11
defined for S, so that each supp(p_(V)) is a rectangle. Most of the
considerations applying to (BI ] also apply to (1). However, supp(B(V))
is always contained in supp(p(V)) (see Figure 2-2), and we know of no
computational advantage for (1i). Thus, we restrict our attention to
(B).
(o -0 . .(0) O(0) -O -()IIJ I III I
(0)-(0)(1/2) I (0)(1/4)(1/2)I I I I III I(0)(1/2)(1)----() (0)(1/2)(1)- -(0)I I I I III I I I II
(0)---(o) ---- (0) ()--()-(o)
B(1/2,1/2) D(1/2.1/2)
Figure 2-2: Lagrange versus tensor-product basis functions
2.3 The 1-irregular rule and some of its consequences
Given an admissible mesh, we force possible additional refinement
by applying, as many times as possible, the following variant of a rule
of Bank and Sherman [14):
refine any unrefined element with more than one (2.1)
irregular vertex on one of its sides.
We call rule (2.1) the 1-irruuu Lg". The 1-irregular rule must
terminate, since it never increases the maximum element level in the
mesh. After the 1-irregular rule terminates, the resulting mesh is
1-irregular. Conversely, if any refinement forced by the 1-irregular
12
rule is omitted, the resulting mesh is not 1-irregular. Thus, the
resulting mesh has the fewest elements of any 1-irreSular mesh
containing the original mesh.
The resulting configurations (up to symetry) of the support of
B(V) for interior V are depicted in Figure 2-3. Because of the
1-irregular rule, further refinement of an element in one of cases I-VI
produces another of cases I-VI, possibly at the next finer level.
Each unrefined element E in the support of B(V) satisfies either
A : V is a corner vertex of B,
or
B : V is not a corner vertex of K, but
V is a corner vertex of the father of E, and
V is next to an irregular corner vertex of E.
Thus, it is simple to determine which basis functions are nonzero
in B. Figure 2-4 depicts a sample B in its father, f(B), and numbers
the 6 relevant vertices. B3 and B4 are nonzero in B (V4 is regular
since f(E) is refined). B2 (B1 ) is nonzero in B if V2 is regular
(irregular). B$ (B6) is nonzero in B if V5 is regular (irregular).
There are exactly four basis functions nonzero in K. Since the
element stiffness matrices are small (44), and the mappings from local
to global basis functions are simple, the resulting finite element
system is straightforward to assemble.
13
.. .. 0 .. .. 00--0----- 0
I IBI II A I A I o-o---o A II I I IBlAI I
IAIAI I A IAII I I I I I0-- 0--0------ 0
I l BI I A Io----o A I o--o A I
IBIA I I IAI IIII :o--o-.--V--o-.--o IV: o--V- -o
I A o-o---o o-o A II IB I IBI I----- 0 0- 0- 0
0- 0-- 0
IBI I iBI Io-o A I o-o A IIAI I IAI I
V o-V--o--o VI ?ooV-o---oIA I A I B o- AI BO----O-- O O'--
Figure 2-3: Configurations of elements in supp(B(V))
14
2 3
I ElI I I
Figure 2-4: f(E)
Furthermore, the basis functions are locally independent:
in each unrefined element E, (2.2)
(BIlE : E in supp(BI)) are linearly independent,
where "1E denotes restriction to element E. We can show that (2.2)
holds if and only if
no irregular vertex touches unrefined elements (2.3)
at three different levels.
For example, in the mesh shown in Figure 2-1, irregular vertex V touches
unrefined elements at levels 1, 2, and 3, and the restrictions to
element E of the five basis functions nonzero in E are linearly
dependent.
Let N(V) be the number of basis functions with support intersecting
supp(B(V)), where V is a regular vertex in a 1-irregular mesh. N(V) is
the number of nonzeroes in the row of the finite element matrix
corresponding to B(V). The minimum possible N(V) for an interior vertex
is 5, and the maximum possible N(V) is 13 (Figure 2-5). N(V) is 9 for
interior vertices in a locally uniform mesh. N(V) for boundary vertices
can be as small as 4. N(V) is 6 for boundary side vertices in a locally
uniform mesh.
O--O- -0O--o-- OI I I I I I
0_-_.. &--o---. I 0- II I I I I I I I I I-0-- -.-- V-- .-- 0 0- -V-- -- 0
I I I I I I I 0-0-. I Io--o--o I .--- o---o o.-o--o.--oI I I0__OO- ... _O O 0-0 0__
N(V) = 5 N(V) = 13 N(V) = 13
Figure 2-5: Minimum and maximum N(V) for interior vertices
Many refinement sequences arising in practice naturally yield
1-irregular meshes, so the 1-irregular rule introduces no additional
refinement. We now show that, given an arbitrary admissible mesh, and
applying the 1-irregular rule as many times as possible,
the number of elements forced to refine by the (2.4)
1-irregular rule is no more than eight times the number
of refined elements in the original mesh.
Hence the number of basis functions introduced by the 1-irregular rule
is no more than a constant times the number of original basis functions.
Suppose there are LMAX levels in the mesh. Let
16
O(L) : (original refined elements at level L),
R(L) := (elements at level L refined by the 1-irregular rule).
It suffices to show that each element E in R(L) shares a corner vertex
with at least one element F in O(L). We do this using induction from
Lx LMAX back down to 0. There are clearly no elements in R(LMAX).
Suppose the statement holds for all levels with index greater than L.
Since E is in R(L), the 1-irregular rule forces E to refine because G, a
neighbor of E, has a son, H, which is refined at level L+l, and shares a
side with E. If H is in O(L+l), then G-F is in O(L), and (2.4) holds.
Otherwise, H is in R(L+1). By induction, there is an element I in
O(L+1) which shares a corner vertex with H. Then the father F of I is
in O(L) and shares a corner vertex with E (Figure 2-6). So. (2.4) holds
in all cases.*
I I I I Io-F--o -- oIl lI I I0--0------0I Im I Io--G--o ] II I I I0-0 00- -- o
Figure 2-6: H in R(L+1)
We do not know if (2.4) is sharp. The largest ratio of enforced
refinement to original refinement we have encountered is 5.75.
. -I - - - ' " "- . . . . . . .. ...
17
In summary, given any admissible mesh, the 1-irregular rule forces
additional refinement such that
the resulting 1-irregular mesh has no more than a (2.5)
constant times the original number of basis functions,
and
the resulting element stiffness matrices are of (2.6)
bounded size (4X4), and simple to assemble,
and
the number of nonzeroes in any row of the resulting (2.7)
finite element system is bounded independent of the mesh.
2.4 Alternative mesh refinement rules
In this section, we show that a refinement rule like the
1-irregular rule is needed for properties (2.6) and (2.7). We also
consider alternative mesh refinement rules.
Suppose, because of a sharp point singularity, that a mesh results
from refining only elements which contain the point P - (2/3,2/3).
Since
2/3 = I - 1/2 + 1/4 - 1/8 + 1/16-...,
the mesh is formed by successively refining the alternately upper right
or lower left son of the smallest refined element. If there are
LMAX >) I levels in the mesh, the active vertices are the ten boundary
vertices and the LMAX interior vertices
VL PL PL PL L (-1/212 L-lI...., XA"L L . . . ..L .L II -M II "O
ris
Let element E be the smallest element containing P. For any VL , it cas
be shown by induction on I that B(VL ) is nonzero at the points (P1 _I,P 1 )
and (Pl,PI_1 ) if and only if I > L. Setting I - LKAX, since
(PI.X_,PIAX) and (PIf,Pj/X_l) are corners of E, E must be in the
support of each B(VL). Thus, the element stiffness matrix for E, and
the resulting finite element system, are essentially dense. Since the
mesh is 2-irregular, this example disproves the conjecture (Babulka and
Miller [4], page 17) that (2.7) holds for k-irresular meshes with k
bounded.
Assembly of the finite element sytem is fairly complicated (e.g.,
Gannon [21], Rheinboldt and Mesztenyi [251). If Gaussian elimination is
used to solve the system, the solution time is asymptotically
proportional to LEA.. By contrast, if the i-irregular rule is applied,
assembly is simple. If the vertices are ordered according to the
maximum levels of the elements in their supports, solution time is
asymptotically proportional to LMAX (see Section 6.3).
In the previous example, (2.6) and (2.7) can be insured without
enforcing additional refinement, by using modified basis functions which
are identically zero in the element containing the singularity. Figure
2-8 depicts the modified basis function 1(V1 ). When there are Jam sharp
point sinsularities at
P1 = (9/16 + (2/3)116, 10116 + (2/3)/16)
and
P 2 (9116 + (2/3)116. 13/16 + (2/3)/16),
19
I I I I V : regularS I I I interior vertex
-- 0
(0) -V0- ..... (0)
I iI I II IiI I iII I
II I
-VoFigure 2-7: Singularity at P = (2/3,2/3), IJ(AX = 4
(0) ..... (-1)- (0)- (0)II I iII I 01I :(
I (o)-()-()-(o)I 1 lo1 I
(.S)-(O)-(o) III I I I(0)- (1)(.5) -(0)--(-1)
II II.I III III III III III I
(0) (0) (0)
Figure 2-8: A nodified basis function
however, the support of the basis function for V1 must asymptotically
intersect the supports of roughly half the basis functions in S, since
20
it can only be zero at one point along the line segment from (1/2.1/2)
to (1/2,1). Thus, with aU rule for forming basis functions, it may be
advantageous (purely from matrix sparsity considerations) to force
additional refinement.
We briefly mention alternatives to the 1-irresular rule. Another
rule which insures properties (2.5). (2.6), and (2.7) (when applied as
many times as possible) is:
refine all unrefined side- and corner- neighbors of any (2.8)
element which has grandsons (sons of a son).
Rule (2.8) is equivalent to a condition in Van Rosendale [301, page 59.
With this rule, only cases I-V in Figure 2-3 are possible, and (2.5),
(2.6), and (2.7) hold for both (Ut) and %fi). However, rule (2.8) may
force more additional refinement, and hoe require more work, than the
1-irregular rule, and is no easier to implement.
Conversely, any rule which may force less additional refinement
than the 1-irreSular rule does so only when neighboring elements with
size ratios at least 4:1 occur. For example, if we apply the rule
if an irregular vertex touches unrefined elements at three (2.9)
different levels, refine the element at the middle level,
as many times as possible (refining elements with lowest level first),
the resulting mesh will satisfy (2.6). and nay have fewer elements then
21
the resulting 1-irregular mesh. Unfortunately, (2.5) and (2.7) may not
hold: for a problem with a sharp point singularity at (1/2+s,1/2+s), as
c approaches zero, the number of nonzeroes in the row of the finite
element system for B(1/2,1/2) may be arbitrarily large.
Figure 2-9 depicts the results of applying the 1-irregular rule,
rule (2.8), and rule (2.9) to the mesh in Figure 2-7.
2.5 A mesh data structure
The data structure we use to represent 1-irregular meshes is a
variant of a data structure presented in Bank and Sherman [141. The
data structure has two main components: a refined-element tree IRCT,
stored in a 4XNIRCT array, and a vertex list IVET, stored in a 2XNIVUT
array. For each refined element 9, there is a 4X4 node in IRCT
containing pointers to the father of E, to any of the four sons of E
which are refined, and to the nine vertices which are corners of the
sons of E. For each interior vertex V which bisects an element side,
there is a 2XI node in IVUT containing pointers to the elements with
sides bisected by V. Other types of vertices have other information
stored in IVIUT. Figure 2-10 depicts a sample refined element E, its
node in IRCT, and a sample node in IVURT.
The neighbors of any element can be easily found. For instance, in
Figure 2-10, the pointer location for the bottom neighbor of S1 is in
the node for E. while the pointer location for the right neighbor of S1
is in the node for the right neighbor E' of E. The node for K' (if it
- .----- !~~.
22
0-----------------o 0-------o----o--o
0- I.a-0 ----.--. 0III I 0 0 -.- 0II0
.-r 0-o- 0 0 ---0
0- -0------O0 0-0-------0-the 1-irregular rule rule (2.8)
0-
0I -- 0- 0rul i (29
Fiur 2-9 l-; 4: atr tiems rinm true
23
I f(E) I - - I C
Is IS I I V V I v I-0 1--- w Ef I_ 1 2 1 WIEs .I I w2
1s1 2I vi W 2 1 1V3 W2 V2 - 0 - -
I so I s1 S2 I S3
element E, with sons Sop ... S3 node for E in IRCT
I E lI I ElII---I III E I I null I
node for W. in IVERT node for W1 in IV MTE' refined E' not refined
Figure 2-10: Nodes in IRCT and IVEUT
exists) is pointed to in the node for W1 in IVUT.
Moreover, suppose the entire mesh is rotated 90, 180, or 270
degrees. Then IV T and the top row of IRCT remain the same, and the
last three rows of each node in IRCT are rotated. For example, if the
mesh is rotated 90 degrees counterclockwise, the node for E becomes
If(E)l - - I C
I V1 I v2 v3 vo
ueW1 W2 W3 dWe n eI S 2 S 3 1S O
Thus, the neighbors of any son of E can be found with the same procedure
used to find the neighbors of $T aae long as indices into the last
three rows of IRCT are incremented by the appropriate value sood 4.
24
This mesh data structure makes it relatively easy to construct the
finite element system for 9, to evaluate the app:xoximate solution. and
to refine the nesh. More data structure considerations are in Dank and
Sherman [141.
CNAPMIR3
Local order refinement
3.1 Introduction
Traditional finite element methods for elliptic problems employ
finite element spaces with low-order (e.g.. bilinear) basis functions.
These spaces are appropriate when the solutions to the elliptic problems
are not smooth. When the solutions are smooth, standard approximation
theory indicates that high-order piecewise polynomial basis functions
can yield much faster convergence. Thus, it is advantageous to be able
to use both low-order and high-order basis functions.
Local-order finite element spaces have received recent attention by
Babu'lka, Katz, and Szabo [3, 11, 12] and Babulka and Dorr [2). Full
local-order, local-mesh methods in more than one dimension have not yet
been considered.
In this chapter, we extend the finite element spaces with
1-irregular meshes considered in Chapter 2 to include basis functions
with locally-varying polynomial orders. In Section 3.2, we consider the
C case, and in Section 3.3, we briefly consider spaces of smoother
25
26
functions. In Section 3.4, we present simple data structures for the
spaces considered.
3.2 The Lagrange ease
Let M be a given 1-irregular mesh. In this section, we construct a
space S (= S(M)) of C0 piecewise polynomials such that, in each
unrefined element E in X,
smooth functions are locally approximated to order k 2 2. (3.1)
Our local order refinement follows the basic approach of
Taylor [29] and Babulka, Katz, and Szabo [3]. On [0,1], let the
polynomials (pi(x): i=l...,kmax) satisfy
Pl(X) = -x,
p2(x) =x,
and for i > 2,
Pi(x) is a polynomial of order exactly i,
Pi(c ) e P(1) - 0.
Since (Pl ..... ap )is a basis for the polynomials of order k > 1. the p
are called (C0) hierahc v9. lm23Jla.
On element E, Pi(z) and p (y) are defined by mapping [0,1J onto
(e,,Eehul and t[yy+h,]:
27
D .p.(x) DJpi ((-xE)/hE) hE-j,
and
DJpE(y) := DJp M((y-yE)/hE) hE-,
j=O,...,kmax-1.
Let pi(t) be the Legendre polynomial on [0,1] of order i. Since
Pi(t) = 1, and
Opi p 0O iJ,
a convenient choice for [pi is
= f; Pig1 (t dt i 2 2. (3.2)
We define S to be the span of three types of basis functions
(B (x,y)): vertex, side, and element.
The vertex basis functions are the piecewise bilinear basis
functions from Chapter 2, with B I(V) M 61i for all regular vertices V1.
The side basis functions are products of piecewise linear functions
in one direction with quadratic or higher-order hierarchic polynomials
in the other direction. Each side a of an unrefined element, not
strictly contained in a side of another unrefined element, has side
basis functions with hierarchic polynomials of order up to
28
(3.3)k : max(k E s contains a side of unrefined element E).
A basis function for side s in Figure 3-1 is h()pE(y).
II all
II B s-....II all
hlx)y.
h(x)
Figure 3-1: A side basis fuhction
The element basis functions for an unrefined element B are the
functions
Element basis functions for E are nonzero only in E, and only occur when
k 2 5.
Since the mesh is 1-irregular, IL2c basis function B. for S is of
the form p (x) Ey(y) in each unrefined element E, where Ex is either E
or f(E), and Ey is either E or f(E). Moreover, (3.3) and (3.4) imply
that in each element E, any monomial z y with i+u-I I is a linear
combination of restrictions to E of basis functions, so (3.1) holds.
-! - .... T
29
Figure 3-2 depicts the values of the basis functions in an element
in a sample local-order finite element. Note that many of the basis
functions are of order greater than kE = 3.
3.3 The Hermits case
The construction in the last section extends immediately to spaces
of globally Ca- 1 functions for any integer a 1 1. The hierarchic
polynomials satisfy Hermite interpolation conditions on their first a
derivatives at 0 and 1, so kE 2 2a. Vertex, side, and element basis
functions are defined as in the C case. There are
2a vertex basis functions for each regular vertex,
(kE-4a)(kE-4a+l)/2 element basis functions for element E
(none if kE I 4a),
and
a (ks-2a+1-i) side basis functions for side s.
and condition (3.1) holds as before.
3.4 k soah data struoture for local order refinement
In this section, we extend the mesh data structure from Section 2.5
to handle local-order finite element spaces.
For each unrefined element E, we store -kE in IRCT, in the location
in the node for f(E) which will contain a pointer to the node for E if E
is refined.
30
mesh:
O---...0-----0- ,0
I I I I
o-------- k-4 II El
I k-3 I k=3 I0----0--0 0
I Ik- I I0----II I I II I I I0--0-0- 0
basis functions in element E :
E E *1P3 (2)p 2 (y)E E f( ] )Ipl~x)P(y~l ' . , f(E(y)p2xp
0
0-E1 E
B 1P4(x)P1 (y P
p E BX~ E ( ()
Figure 3-2: Lool-order basis functions
Let the function NDV map the basis functions for S to the regular
31
vertices in the mesh, such that
1. MBV maps each vertex basis function to its vertex.
2. MRV maps each side basis function for side a to the
last-created regular vertex at an end of a.
3. RE3V maps each element basis function for element E to the
center vertex of f(E).
A sample mesh indicating MBV is shown in Figure 3-3.
~Iv-S-v--- V---S- V V :vertex basis functions
S E S E S : side basis functions
V--S---V--S--S E S E : element basis functionsI I I IS E S E II
I4 I : basis functions sappedV -S--V---S-- V- ----- S_ba V to vertex VI IjSE sE II I IV -S--V--S- - S E S
S E S E I
[V -s---v--s-= v-..
Figure 3-3: Basis functions mapped to vertices
We order the vertices and basis functions for S so that the basis
functions mapped by NOV to vertex V areIV
BMVB(IV) ' 8 MVB(IV)+1 . , BMVD(IV+l)-1
32
and we associate a nuamber IBFT(IBF) with each basis function BIBF
encoding the spatial relation to the vertex NDV(BB F) (o.S., left, lower
right, etc.).
With the data structure from Section 2.5, and two additional arrays
containing MVB and IBFT, we can easily determine the basis functions
which are nonzero in a given unrefined element.
As in Chapter 2, this mesh data structure makes it relatively easy
to construct the finite element system for S. to evaluate the
approximate solution, and to refine the mesh.
C(APERU 4
A posteriori error estimators
4.1 Introduotion
The traditional error estimates for the finite element method for
elliptic problems are a priori (e.g., Ciarlet (17]), predicting the
expected rate of convergence of the error to zero, but not saying much
about the actual error for a given finite element space.
In recent years, locally-computable a posteriori error estimators
* have been developed, mainly by Babulka and Rheinboldt, which estimate
the actual error for a given space. The landmark paper in this area is
by Babu'ka and Rheinboldt [81, giving a method of constructing a
locally-computable error estimator which is within a multiplicative
constant of the norm of the actual error. Further results for
one-dimensional problems are given by Babulka and Rheinboldt [6, 7. 51
and Reinhardt [24]. Further results for two-dimensional problems are
given by Babulka and Miller [4]: under suitable assumptions, they prove
that the error estimator used in their code (with piecewise bilinear
basis functions) converges to the norm of the actual error as the mesh
size goes to zero.
33
34
In Section 4.2, we present some terminology for discussing
a posteriori error estimators for two-dimensional Neumann problems. In
Section 4.3, we briefly describe the error estimator presented in [8].
In Section 4.4, we present a new error estimator which has two main
advantages over the estimator in [8]:
1. Under reasonable assumptions, the new error estimator is
asymptotically an upper bound on the norm of the true error.
2. The new error estimator can be computed locally in each
element. (The one in [81 must be computed over more
complicated local regions.)
In Section 4.5, we prove the upper-bound property of the new error
estimator, and, under suitable assumptions, we show that the now error
estimator is no larger than a multiplicative constant times the norm of
the actual error. In Section 4.6, we present several cheaper
alternative error estimators. In Section 4.7, we extend the estimators
to handle more general boundary conditions and differential operators.
In section 4.8, we summarize the results in this chapter.
In Chapter 7, we present numerical evidence that, in some cases,
our error estimators converge to the norm of the actual error as the
nosh size goes to zero. We have not been able to prove this
convergence. Currently, we can only prove convergence for our error
estimators for one-dimensional problems [15].
35
4.2 Terminology
Consider the model Neumann problem
Lu := - D (a(x,y)Dxu) - D (a(xy)D yu) + b(x,y)u f(x,y) (4.1)
in the interior of f0,
au/8n = 0 on 80,
where D , D , and a/8n denote derivatives in x, y, and the outwardx y
normal to 0 respectively, a(xy) and b(xy) are in L(0), f(xy) is in
2L (0), 0 < a . a(xzy) i a, and 0 < b S b(x,y) S b. Variants of (4.1)
are discussed in Section 4.7.
Let w be an open subset of 11 with piecewise smooth boundary aw.
For a non-negative integer k, and a function v in CO(w), we define the
norm
11V112 != ko J-O (D• 'Di-Jv(x'Y))2 dx dy.
Let BkW) denote the completion of Z(w) with respect to the norm
11.11 iksw) is equivalent tok,w
(v in L2(w) : IVikw <.1,
if derivatives are defined in the distributional sense (e.g..
Ciarlet [171, page 114). Let k(s) denote the completion of the
functions in C(w) with compact support in a with respect to the norm
11'11 When k is omitted, k = 0 is assumed. For v,w in 0(a)k,.1
36
(= L2(w)), let
(vw) : f1 v(x,y) w(x,y) dx dy.
For v,w in H 1w), let
a (vw) - (aD xv,D w) + (aDy vD w) + (bv,w)
If w is the interior of Q, a (u,v) is obtained by multiplying (4.1) by a
function v, and integrating over w using Green's theorem. Let the
n.eraz lllvill := a (vv). By the assumptions on a(x,y) and0
b(x,y), there exists C > 0 independent of v, such that
C 1 Ilvll S. IIIvll 5.C 11vll', (4.2)
for all v in Hl(w).
On the boundary 8w, for v and w in L2 (w), let
v'w> ae: w Lv(o) w(a) do,
where do is arc length. We define the norm
SC denotes a generic constant, not necessarily the same in each
instance.
Si6
37
Iv12 : <v,va V)
Let a/an denote differentiation in the direction of the outward normal
to o, with the limit in the differentiation taken over points inside w.
For brevity, we sometimes replace v by its closure in names for
spaces, norms, and inner products: for example, I1VIlk,E stands for
"v''k,i nterior(E) . When the subscript w is omitted in norms and inner
products, w = interior(Q) is assumed.
Given a 1-irreSular mesh N with unrefined elements (E), we choose a
finite element subspace S (= S(M)) of ff1(Q) as in Chapters 2 and 3. For
simplicity, we assume S is based on C hierarchic polynomials, as in
Section 3.2. We also choose a larger finite element space S ( S(M))
containing S, constructed in the same manner as S, such that in each
unrefined element E,
smooth functions are locally approximated to order kF ) kE
Since S contains S, and the basis functions (B1) are hierarchic, the
basis
(B1 : BI in SI
for S contains the basis
(B1 B1 in S]
38
for S. For each unrefined element E, we define SE to be the space with
basis functions
(B 11 E in supp(B1) B I in )
where (B I E in supp(B1) are linearly independent because of (2.2).1 EI
(3.3), and (3.4), and we define S E to be the space with basis functions
(B I 1 E in supp(B 1), B I in S).
We define S CSEto be the span of the functions
(B IE: E in supp(B1) B1 in 8,. B1 not in S).
For any function
V B 'S I B IB
in SE , we define
Yv Ia vIBI IE(4.3)
in SE Then v-1E (v) is in 'E7 ad for any function
V BI i vi B I
in S, the function
I(v) B I as v B I
is in S, and is equal to I (v) in E.
39
Let N(E) denote the set of unrefined elements which intersect with
E at more than one point. For example, in Figure 3-2. N(E) is the set of
labelled elements.
For convenience in discussing functions in the product space
M (a) (v: v in H (E) for all E),
let
a(vw) E E aE(vw) for v and w in HM
and
IIvI12 V E l E for v in HM(a).
The weak form of (4.1),
a(uv) = (fv) for all v in ]Jm)
has a unique solution u in H1 () (e.g., Ciarlet [173, page 19). We
impose the weak form of (4.1) in S to get a finite elental arnroximation
U in S. where
a(Uv) = (fv) for all v in S.
The ±uu for S is e :- u-U. Similarly, U in S is required to satisfy
A(lv) = (f'v) for all v in S.
1i
40
and the corresponding error for S is e :- u-U.
In this chapter, we consider ways of estimating the norm of the
error 1llelll. Let
a := U-U.
We assume the saturatio condition that there exists 0 y< 1 such that
lll~l l lll lll.(4.4)
If u is sufficiently smooth, then y goes to zero as h goes to zero.
Thus, since
we concentrate our attention on estimating IIIlt.
4.3 A Diriehlet a posteriori error estimator
Traditionally, the error estimates for the finite element method
have been a priori, predicting the expected rate of convergence of
IIleill to zero (usually as h :- max(h.) goes to zero), but not saying
much about the actual error for a given S.
In [8], Babulka and Rheinboldt present a posteriori error
estimators constructed basically as follows. Lot (B) be a subset of
the basis functions for S which form a partition of unity on 9. For
each 3, let :supp(B and let e be the solution in fO(Q) of the
local Dirichlet problem
41
a0 (e,v) = (f'v)O a. (U.,v) for all v in R10(0")
Under suitable assumptions, Babulka and lheinboldt show that there exist
C1 > 0 and C > 0 independent of S such that2
cI l 2 Illel111.1112.
In practice, since H0(Q ) is infinite-dimensional, e is not
computable. The most straightforward computable approximation of a3 is
@J in
S := (p(xy) : p in S, supp(p) contained in 03)
defined by the equations
a (siov) (f'v) - a (U,v) for all v in S (4.5)
Then
c m i2 C2 llll (4.6)
The quantity (TIlls1112 )112 is called an e*zor eiMas .
Each a in (4.5) is a projection of a in the energy norm onto a set
of functions, so the right hand inequality in (4.6) arises naturally,
while the left hand inequality is harder to prove.
42
4.4 A Neumann a posteriori otter estimater
The motivation for a Neumann a posteriori error estimator is that
it is more Important to have an upper bound on ilsli than a lover
bound. We construct an error sinulator a, such that a is the projection
of s onto S with respect to the energy norm. Then 1iJJ S1 111;111
arises naturally. Following Babulka and Rheinboldt [10], we call
I I an error estimator, and for each element E, we call 1116-111 E an
error indicator.
To motivate our construction of a, suppose u is in 12(E). By
Green's theorem on element E,
ao(e,v) - FLIV) :- (fv)E + <aau/n,v> - a(U,v) (4.7)
for all v in H1(E). For a given v, we can compute every quantity in
(4.7) except au/On on the boundary of E. Suppose E shares side a with
element E'. On a, a possible approximation to aau/SnIE is the average
value of aU/OfnE on E and E', where n. is the outward normal from E.
Since
LSj~1j3 - 'Is '
we define the approximation
[aOU/On] IE : 1/2 ( aU/Dnl - aOU/nl, , a not in 8a. (4.8)
10, B *in Ol
43
Note
[aOU/n]I E - [aBU/8n] EP
The resulting approximation to FE(v) is
FE(v) :- (f'v)E + <[a8U/8n],v)SE - aE(U,v). (4.9)
Since aE(.,.) is elliptic, we can define an error simulator for e in E
by replacing FE with FE on the right hand side of (4.7). However, as
approaches zero, the finite element matrices for the error simulator
approach an indefinite matrix corresponding to the case when b(x,y) = 0.
Thus, we formulate error simulator equations which are
automatically consistent when b(x,y) = 0 and aE(',') is indefinite. We
approximate s by the function aE in SE defined by the equations
aE(E*V) = FE(v-IE(v)) for all v in SE (4.10)
where IE(v) is defined by (4.3). Since a E(.,) is elliptic, aE exists
uniquely. Since v need not vanish on DE, (4.10) is a Neumann problem.
Moreover, since IIE is in SE
IE(1) 1,
so the consistency condition aE( EO) - 0 is automatically satisfied.
Equations (4.10) are equivalent to choosing aE in SE such that
44
aE E tE(v) for v in
for v in S3
The corresponding matrix equation is A , f for the coefficient vector
of ;E where
A:= A I 2 r:- I . f:I 2
the coefficients of the basis functions for SE are in ;1 and the
coefficients of the basis functions for SE-SE are in 2
We define the value of the error simlator ; at (xy) in Q to be
8(zy) :- a (Xy),
where E is an unrefined element containing (xy). When there is no
danger of confusion, we refer to a3 as a. Note that a is in %(1), but
not necessarily in Hl(Q).
4.5 Properties of the Neumama error estimator
The most important property of ; is that
Thmorn. A: IIslll 11 ll;Ill.
Proof: Suppose v is in S. Since (aOU/ni - - (aOU/nJi, and
[aU/nJ - 0 on 80,
45
E <[aaU/anv> DE = 0.
Thus,
B F(v') - E V - aB(U'V)) (f,v) - (U'v).
Since I(v) (page 38) is in S.
E FE(IE(v)) - E FE(I(v)) " (f,I(v)) - a(Ui(v)) 0. (4.11)
Thus,
a(sv) - a(ev) = (fv) - a(Uv) = E E(v-IE(v)) = a(;,v), (4.12)
and, in particular,
ll s l 2 = a(ss) = a e : l l l l l l l l
By (4.4). we have
Corollary 4.2: (1-T)Illl I1 1.
This is a jIokl inequality. We cannot generally expect
IllolllE S C I1l;lllE : suppose u is smooth, S contains bilinears,
hb nh in region R in O. and hE h >) h away from R. In region R, U is
a finite element approximation to the solution to
L(u.) = f(xy) in the interior of R, (4.13)
ul U on aR.
46
Let E be an element in the interior of R. Since the error simulator
equations in E are local, they are the same for (4.13) as for (4.1).
Since (Theorem 4.6) 1I1;11E is bounded by localE de by ocalquantities behaving
like IIIU-uRIIIE , and the boundary error in (4.13) pollutes the
solution throughout R, we expect
E >> III 11 E
(In practice, this may be acceptable, since the error is bigger away
from R due to the coarse mesh, and the small size of the error simulator
is in accord with the fact that no more mesh refinement should occur.)
By the triangle inequality and (4.12),
Ililill - Ilillll S Il-sill - inf(IIl-flll: f in 5),
so we can get an a posteriori upper bound for IIlll-IIlllI by
measuring the difference in the energy norm between ; and any function
in S. For instance, by averaging ;, we can construct an approximation
ave to 6 in S: s-s approximates the size of the Jumps in a across
element boundaries. If
AJ veJ JlJ
with y ( 1. then
Of course, the Jumps in a can be made as small as desired by adding
47
penalty terms to equations (4.10). but the resulting equations are
non-local.
We now prove a local a priori upper bound for IIIs Ill' in terms of
quantities depending on e. We assume there exists a constant kmax such
that
kE S. kaiax independent of S. (4.14)
We start with a few lemmas.
Leoma 4.3: 1111E(S)IIIESC IllEIIIE.
Proof: Suppose v is in S. Since the dimension of S E is finite andI
is linear,
WbIE(v).IE (0))E I C1 (bv,v) E
where C 1is independent of v and hE . (Otherwise, there would exist w
in S Esuch that (bww)E - 0 and WbIE w)IE (w))E 0 0. which leads to a
contradiction.) Similarly, since the dimension of BE is finite, X is
linear, and IEM =1,
(aD I (v),Dx1E(0) + (aDyIE(v),Dy1 (0)
C 2 ( (aD xvD v) E + (a y vD yv)1 E
where C 2 is independent of v and h3E Thus,
48
1102
2 2IIz~lE _B mICl2) IIvl E .
Let v=
Lemma 4.4: l-1R;)l ,idini~ j~: 1 E 7E(E) 1 aE C hi' 'E"'E
Proof: The trace inequality
- E)2' <- 1 E1,)(.5IE- E a - hIlE_ chEE-E EIE + E I EEA h 14
follows directly from Theorem 3.10 in Agnon [1].
Supposev is in E For 0 S xny < 1, let
v,(x,y) := v(x, + hE z, YE + hE Y) "
a (x,y) =E(v)(xE + hE x, YE + hE y).
Since the dimension of SE is finite, I E is linear. and IE( - 1,
11 12 SC(Iv112 + I 12
and thus
1IT-1(V)112l2 vII2tiV*-*I(v)t ES h 1.3t,,l
where C is independent of v and hE . The result follows by letting
v = 'B in (4.15) and using (4.2) and Loima 4.3.
vjft J.5: If u is in S2(3),
.. .. .... . .. --- . . ... .
rI
49
laae/nlaB SchE112 1 1Il1l. + 2 11.112, ./
Proof: Since u is in 92 (E), e is in 92 (B). The result follows directly
from Theorem 3.10 in Agmon [1].
The following theorem shows that the error is large in local
regions where the error indicators are large (and u is sufficiently
smooth).
Theorem 4.6: If u is in H2 (E' ) for all E' in N(W), then
l 1,2 1,12 I12.
IIElllB Ci E'sN(E) Ill, eE + ' l ll2,E'
Proof: For any v in 'E
a E(EV) -(fv-E(v))E + <([aU/n]'V-IE(V)>,E - aE(Uv-IE(v))
- FE(v-IE(v)) - (au/an'v-I,(v)>,E + ( [aaU/Dn ] .v-iE(v)>)B
SaE(e,v-IE(V)) - ([abo/8n],V-IE(v))BE
Let v = " By the Cauchy-Sohwarz inequality and the definition of
[-],
i t ~ ~~~labe/DnIDB s ;-+e > o
EsIN(E) SEP /8n1.E, )
By Lma 4.3 and Lema 4.4,
- w
50
-1/2 (4.16)I'I EIE l l + C hE 1 E E'sN(E) OaBe/OnI " 4
By Lema 4.5, noting that (since the mesh is 1-irregular)
hE1 /2 hE - 1 /2 C for all E' in N(E),
Il;ElllE C IllelllE+
11. 2 1l 2 )1/2.
E'sN(E) Iell + hi 2,E'
The result follows by noting that E is in N(E) and using (4.2).
If u is in H2 (E), we expect the right hand side of Theorem 4.6 to
converge to zero at the same rate as 11le11 If the coefficients
a(xy) and b(x,y) are sufficiently smooth, u is in B2 (C) (e.g.,
Ciarlet (17], page 138).
We now prove a slobal a priori upper bound for 1l1J11. We assume
that u is in 92 (E) for all E, and we assume the saturation condition
that there exists 0 j 7 ( - such that
B"B la8 a/nla ) S y Jlell. (4.17)
If u is sufficiently smooth, then 7 goes to zero if h goes to zero.
Theorn4.,1: There exists C independent of S and S such that
JJJahJiJ C 111o1JJ.
51
Proof: It follows from Theorem 3.10 in Agnon [1 and (4.14) that
* -1 85/OIOE .C 1 1.1d2.. + 2 11.112,
-1 2
C -1 IIsII81.9
Then, using ahe/On - ae/Sn + abc/On in Equation (4.16),
III;EIIIE 1 C 1110lllE + C hEl12 L E'sN(E) la./OnlB' +
C I sN() IIEll,
Squaring, summing over E, using the triangle inequality, and noting that
the maximum number of N(.)'s in whioh any element appears is bounded,
JJJ;iJJ SC 111e1i1 + C 118111 + C h lsaho/Sna12. 11/2.
The Theorem follows using (4.2), noting that 111*111 11 0lel, and
using (4.17).
4.6 Alternative *mz estimaters
In practice, there are several alternativ, error estimators that
are loss costly to compute than IlleIhl.
4.6.1 Ipnring irregular •0aeon
To can compute an error simulator in eloent 3 by replacing ZE(v)
with an analogous function I (v) which iaterpolates v locally in R. The
resulting error simulator, in effet, ignores the irregularity of the
corners of , and its system is slightly cheaper to assemble than the
52
system A v. f. In practice, the resulting error estimator appears to
be at least as accurate as 11lla111. However, Equation (4.11) in the
proof of Theorem 4.1 break do"e when v Is a side basis function for a
side s containing an irregular vertex (Figure 4-1): 1 (v), the analogue
to I(v), is not continuous across s. Thus, we do not consider this
alternative error estimator further.
(0) (0)- (0) (0)
)-( *() I(0)-(O) -- I0
I 11 ()
Figure 4-1: Ignoring the irregular corners
4.6.2 Using the matrix for a local Poisson problem
We can form and solve error simulator equations using a matrix
corresponding to an associated local Poisson problem. For v and w in
jj( ):- jL(E) ( (DxvDxw) E + (D yvDw) E)
where jj(E) Uinf(a(z,y) in E).
thtInstead of choosing ;E in 8Eby (4.10), we can choose aRin SEsuch
53
5I ( E v) - FE(v) for all v in S SE (4.18)
Lo for all v in SE #
and
(0E,1)K 0. (4.19)
(We impose (4.19) because &,(1.1) 0.) Ve construct the error
simulator a and the error indicators (111111 from se in the same way
we constructed a and (111;111E ) from ;E "
Equations (4.18) correspond to the matrix equation A f for the
coefficient vector v of a (xy), where
A, A EvA I I T v -II
I A I I _v I I f2 I
the coefficients of the basis functions for SE are in , and the
coefficients of the basis functions for S-7S are in _2 . Since
11Is1J111 III6+c1 R for any constant C. we can set
reorderinS the unknowns if necessary, so that A is positive definite.
For fixed kE , A depends only on the reSularity of the ceorners of R, and
not on h.
54
Ij. .u. 4.8: If a(x,y) is Lipschitz-continuous, then
II 1 E X ! ! E (1+o(h)) I111
as h oes to zero.
Proof: By (4.10) and (4.18), since QE(vJv) 5 aE(VV) for all v in SE
II IE -&.E(6,8) .. ..18ll 1llE 1_I1; I1E lllsll E III;III E .
Similarly,
2
-E--E8)1116111 18.*)II81E A E(s, =E( I E ll~l E •
and if a(x,y) is Lipschitz-continuous in E, so that
a(x,y) I #(E) + O(hE) in E, then by (4.19),
11111 E I C hE 1 I E11, ,
and
I11 1 1E (1+o(hE)) IIIllE
as hE goes to zero.
Since mesh refinement occurs in a very regular fashion, we can
precompute the factorization of A. and only incur the cost of a
backsolve plus computation of the right hand side for each element. If
the hierarchic polynomials (pi) are chosen by (3.2), then since
0 pip; 0 , i j J , max(ij) > 2,
'S
A is sparse. In practice, we replace the quantity A(E) with the more
convenient quantity a(xE+/2,yE-2). Then
Il;IlS IES (1+O(hF)) 1115111E (1(h ) 111;II .
4.6.3 Using a smaller matrix
We can also compute aE in SE-SE by the equations
AE(aE'v) - gE(v) for all v in SE-SE . (4.20)
constructing the error simulator a and the error indicators (Il1111 )
from *E in the same way we constructed a and (III;I) from a-,
Equations (4.20) correspond to the matrix equation A2 v 2 f2 * or
v = f, for the coefficient vector v of 's. where
We now show
Thoren 4,9: There exists a constant C independent of h such that
Proof: Th. left hand inequality follows from
l- Tj = ;Tf - T - £(s,5) j ilSlIlI l!ll.
11,11E vvAYi~~)jIIal 1811
56
The following approach to the right hand inequality is due to
Eisenstat (19]. Since A is positive definite, using block Gaussian
elimination, A4 Y2 - f2 , where
T -1A4:= (A2-A 3 A1 A3 )
is symmetric and positive definite. Then
2 T -TIlIsillI f 2 A4' f2E 24 2
Similarly, A2 is positive definite, and
2 T -TIIIEIII = f2 2 f2
Since A4 and A2 are finite-dimensional,
2 vT AT vT A2TC :=maxv v v A v<
v0
In particular, let v = f
For example, if E has no irregular corners, a(E) 1, SE contains
bilinears, and S-SE contains the biquadratic basis functions
(XE+hEx) (YE+hEY)(YYE) / h 3
(x-zE) ( h (Y-Y) / 3
Eh ., (IRE) (y +h y, / h
(z +h
E....) (..zE) (Y.YE) ..
57
then
6 0 0 0 o 0 0 0I 0 4 -1 -2 I I 1 1 -1 -1
A =10-1 4-1 I/6, A3 = -1 1 1-1 1112,1 0 -2 -1 4 I-1 -1 1 1
113 0 2 0 I1 0 13 0 2 1
A=1 2 0 13 0 1/90,I 0 2 0 13 I
and C (11/6)1/2 ; 1.354.
4.6.4 UsinS the residuals and Jumps in normal derivatives
We may choose not to solve any linear systems.
By Green's theorem, for v in SE7SE
a(Uv) = (LUv)E + (aaU/8n,v>8E
so
'E(sE.v) = (f-LU.v)B + ([a8U/an]-aaU/8n,v>OE.
In particular,
II; 1112 - (f-LU, )E + ([aU/On]-S8U/8n,sE >E S
and so
I-
58
111-8E IIE C ~1 (h E'-&(E)) hif-Lull E + (4.21)
C 2 (hEA(E)) l[a8U/8n]-aaU/anl.aE
where
C 1 hEA E)) sup Ilvhl /h1lvhllveS+ E-
C 2 (hEA(E)) :sup IVI aE -IIII
8+
and S+ :=(v :v v v0(x,y)+C, v0 in SE7SE ,v # 0, and (v1l)E 01.
For example, if E has no irregular corners, S E contains bilinears.
and SE-SE contains the biquadratic basis functions on page 56, then S+
is the span of
hE 3 (xehE~x) (yE+hEy) (yyE - 1/12,
hE (xKE) (yhgy (YY - 1/12,
*~~~~ h 3 (zE.Ex)( ye) Ehy) - 1/12,
h E3 (xEhE-x) (x-zE) (yyE - 1/12,
* I and
1/2C I(hE4j(E)) - hE / (22 ACE))
C2(hE..A(E)) - C 3 hE /(11 &CE)) )1/
Babulka and Rheinboldt [9) suggest error indicators when S contains
piecewise bilinear basis functions which amount to approximating
11,1,12 h 2/(24jL(E)) 11f-Lull 2 + (4.22)
8E
for interior elements. Since 1/22 : 1.1 X 1/24 and 3/11 :1.6 X 1/6,
the two schemes are in rough agreement.
If the mesh is uniform, S contains piecewise bilinear basis
functions, S contains piceewise biquadratic basis functions, u is
sufficiently smooth, and h goes to zero, then combining Corollary 4.2,
Theorem 4.8, Theorem 4.9, and (4.21),
... 112 -2 j (12A(E)) 1 1/2 hllf-LU1 +
(2jk(E))- -/2 h1/2, I[aSUf n3-aOU/OnI ) 2,
where y goes to zero. Similar results hold for higher-order elements
and nonuniform grids.
In numerical tests, we have observed that quantities of the form
C (h'j,()) lfLUI1 + C2 (h 1 9 aj(E)) I[aOU/On]-a8U/On 12
or
E (C I(h,,a(E)) hlf-Lull5 + C2(hR,gj(E)) l[aOU/Dn1-a&U/DnI, )2
only seem to approach 11l@1112 to within a multiplicative constant
60
depending on u. In Chapter 7, we present numerical results using the
error estimator
, :: E 2 1/2
where
2 K(k )(h2/(E) hf-Lull 2 + 4hE//j(E) i[aU/8n]-aU/8n12 ) ,(4.23)
=E E E E O
and
K(kE ) := (kE(kE+l)(kE+2 ))-l. (4.24)
E is chosen to extend formula (4.22). and to perform well on the
problem
D2u - D2 u = f(x,y) in 0 , u = 0 on 80,
x y
when u = sin(nx)sin(ny), k E and hE are uniform, and h goes to zero.
4.7 Extensions
If b(x,y) = 0 in 0, in order to insure nonsingular linear systems,
we impose the Neumann consistency conditions (u,1) = 0, (U,1) = 0, and
(,1)- E = 0. We still have
II1vlllE .i c Ilvlll,E for all v in H (E).
and globally,
1 0 1 C 1 11
61
Since (;E'1 )K 0 implies 11;E'E I C h "'l
LE I13 C 111;111E
for h sufficiently small. (Similarly, 1 ,9I~ C '1 ";1 1E . Thus'
Theorem 4.6 and Theorem 4.7 hold for h sufficiently small.
If a E (-,) is of the form
a(v~w) - (az,y) D Zv,D 2) E + (a~zry) D yvD yw) E +
(bx,y) vDV) E + (o(x~y) Dxv~w) E + (dxy) D yv~ w)E
then, under suitable assumptions, a112 .. ca(, and
liii j C a E(.es) as h goes to zero (see Schatz [271). Thus, Theorem
4.6 and Theorem 4.7 hold for h sufficiently small.
For problems with non-homogeneous Neumann boundary conditions
a~uISn - g(z,y) on 80 . £ in H (80),
if we set
a(u,v) - (f,,) + (s,v> for all v in Elm0,
a(U,v) C f,) + <gSv0 for all v in S.
and
[a#U/8a1 :-U~,y) if a is on 0.
62
then we obtain the same results as before.
For problems with homogeneous Dirichlet boundary conditions
u = 0 on fl,
1the solution u in H (0) satisfies
a(uv) = (fv) for all v in Hl0(Q).0
Similarly, the equations for U, ;, a, and a are now posed on subspaces
of KI(N). Since all the functions of interest now vanish on a0, the0
other results follow as before. Problems with non-homogeneous Dirichlet
boundary conditions which are exactl satisfied are handled similarly.
When Dirichlet boundary conditions are not exactly satisfied by U,in element E touching 80, we define a E in SE by the equations
aE(;EV) = FE(-IE(v)) for all v in SE which vanish on W0,
E on 80.
We define a from aE as before. Let ab be any function in S such that
Equation (4.19) is no longer needed in elements touching 80.
63
;0 "" -b vanishes on 80, a let 10 -;b B y the ame proof asfor Theorem 4.1,
Iol0111 . 111;0111,
so
111-11 1le l 11 II;0111 + 111;bill
Similarly, under the assumptions of Theorem 4.6,111 21612 )112
lllSlll~ S C ' |t'sN(E) 1 ]lll t, + h1 , l 2,. , +/
lll;blll n
and under the assumptions of Theorem 4.7.
; " 11 B, x
where (B } is the basis for S, we choose
;b B I b I I
where
Sb: span(B B I in S. BI(x'y) 0 0, (zy) in 80).
Since the area of supp(b) is O(h), we expect IlbiII - o(1ll'IIl) as h
64
goes to zero. Since a is nonzero on 80, s should be in S rather thr-
s -S when E touches 80.SESE
Analogous results hold for a and s.
As in Babulka and fleinboldt [10], on an element side s of E which
touches 801, we replace the term
(kE) 4 hE/A(E) I[aaU/an]-&aU/anl25
with the term
3a(E)/h 1.12E a
in formula (4.23).
4.8 Sugary
The error indcators discussed in this chapter are:
Ill;s III * where E is in SE and
a(e(;e'v) f s (v) for all v in SE-SE
for allyv in S E
IllEslIIE , where a is in S8 ( Ie'l)E - 0. and
S(B$E'v) P FEWv for all v in %SE
0 for all v in SE
65
111R IIIE .vhor o E is in SE-SE and
-4E( E V) F FE(v) for all v in k-S,
and s where
2 2221F SM= ) h-E/jL(E Ijf I2 + 4 hB/A(E) I[aOU/On]-a8U/nI
The corresponding error eiatou are
t - E s111;1,12 )1/2.
1118111 - B I'3"'1 )1/2,
I~~l~~lLI 2 ( ~ ~ l l 1/2
and
- ' 2 )1/2.
We define the corresponding relative erors in the estimators
eBabulka and Rheinboldt call (the relative error + 1) the "efficiency
index" [101 or the "effectivity index* [9).
66
and
We have shown, for sufficiently smooth u, that as h goes to zero,
(14y) F1 1111S lili~C1111
and
where the positive constants C, C, and C are independent of h, and If u
is sufficiently smooth, y goes to zero if h goes to zero.
Heuristic refinement criteria
5.1 Introduetioa
The guiding principle in choosing a finite element space is to
minimize the norm of the error as a function of work. However, there is
not usually enough information available to choose an appropriate space
a priori. Instead, it is necessary to choose a sequence of spaces
adativel . Each succeeding space in the sequence is chosen on the
basis of criteria computable from the preceding space, with the criteria
tending to satisfy the guiding principle when a heuristic model of error
behavior holds.
When the polynomial orders of the basis functions are uniform, the
resulting refinement algorithms tend to approximately equalize the sizes
of the errors in the elements in the next space. In finite element
methods, for example, this heuristic refinement criterion is used in
Fried and Yang [20], Babulka and Rheinboldt [8], and Bank and
Sherman [13]. When the polynomial orders of the basis functions vary
locally, the heuristic refinement criteria must be somewhat more
complicated (e.g., Brandt [16]).
67
68
In Sections 5.2 - 5.5, we develop two heuristic models of error
behavior for local-mesh, local-order finite element spaces, and an
adaptive refinement algorithm which uses the models. In Section 5.6, we
discuss drawbacks inherent in the models. In Section 5.7, we present
the asymptotic expected error and work behavior for three problem
classes consisting of problems with:
1. smooth solutions.
2. solutions with point singularities.
3. solutions with line singularities.
Except in the presence of line singularities, with appropriate local
mesh and local order refinement, the expected error converges to zero
faster than inverse polynomially with respect to work.
5.2 A model of work
The guiding principle in choosing a "good" finite element space S
can be stated as
minimize 111oll as a function of W, (5.1)
where Ille1l is the error in the energy norm, and W is the work.
However, there is not usually enough information available to choose an
appropriate space a priori. Thus, we must choose a sequence of nested
spaces
S contained in S contained in ... Sn
with finite element solutions U1 ... ,Un , taking S - Sn (for instance,
because the work limit is reached). We choose each Si+l adaptivelz , on
the basis of information obtained from Si Since the spaces are
0-
69
nested, the errors (ei) and work counts (WI satisfy
Ills0 lI I Ill21 ... 2 IlianlIl
and
W1 < W2 < ... ( Wn
The work counts can represent either storage or computer time. If
the work counts represents storage, V V . If the work countsn
represents time, V - n and if
wi S T wi+1 , 0 J y < 1 , I 5 i 5 n, (5.2)
then W I (l-y)- i Wn . Thus, V is usually proportional to Wn . Since it
is possible that n = i+l, in the remainder of this chapter, we take as
the "local guiding principle" a local version of (5.1):
given Si , choose Si+ 1 to minimize tilei+lIIl (5.3)
as a function of Wi+I
(subject to the possible enforcement of (5.2)).
Let S be the current finite element space S in the sequence, with
mesh M, unrefined elements (R), and error e. Let z :- (,y) denote a
point in 0. Let h(), p(z), and w(p(z)) be piecewise constant functions
defined by
70
'x(z) hE the local mesh size,
p(z) PE kE-l = the local polynomial degree,
w(p(z)) WE = the local work count,
if z is in element E. We make the heuristic assumption that the work W
for S is
W = EwE (5.4)
This is reasonable if V is the dimension of S, or (since the mesh is
1-irregular) the storage for the stiffness matrix for S, or the number
of arithmetic operations needed to solve the finite element system for S
when the work is proportional to the number of elements in the mesh (see
Section 6.3). It is not reasonable if the work grows faster than the
number of elements: then, W depends on S globally as well as locally.
5.3 A model of predicted error behavior
In this section, we construct a model of predicted error behavior
based on standard a priori error estimates, and derive heuristic
refinement criteria which tend to satisfy (5.3) when the model holds.
Let m(z) and ;(z) be piecewise constant functions defined by
m(z) : max( k : Hull kE -
p(z) := min(p(z),m(z)-l)
'- I i °- . . o .. ..
71
if z is in element E. m indicates the maximum local smoothness of u in
E, and p indicates the expected convergence rate in the energy norm
using polynomials of degree p (order kE). Standard a priori error
estimates (e.g., Ciarlet [17], Theorem 3.1.5 and proofs of Theorems
3.2.1 and 3.2.2) yield
2 C(p(z)) a(z) G(zp(z)) h(z) dz, (5.5)
where a(z) is the coefficient in (4.1), and
G~z~) :'S ( DJ Dp+'11 u(z) ) 2G(Z'1=J0 x Y2)
is in L2 (0) by the definition of p.
Following Brandt [16], Chapter 8, we make the crucial assumption
that, for heuristic purposes, (5.5) holds in each element, with
eauality:
* IeIll : E :4E f C(j(z)) a(s) G(zp(z)) h(z)ZP(z) dz, (5.6)
so that
111e1112 C B
We also assume, for heuristic purposes, that
a and G are approximately constant in each E (5.7)
and
72
C, G, and V are (formally) differentiable in h and p. (5.8)
Assumptions (5.4). (5.6), (5.7), and (5.8) comprise our first heuristic
model of predicted error behavior. The form of (5.6) is important, but
not the specific values of C(p) and G(z,p). We make assumption (5.8)
only for convenience: all derivatives with respect to h and p can be
replaced by divided differences.
In the following paragraphs, using (5.4), (5.6), (5.7), and (5.8),
we estimate the marginal efficiency of refinement in each element E,
obtaining
Xh) =- (/O/hE) / (OW/ahE) (5.9)
for mesh refinement (h-refinement), and
I - (E() at/p) / (OW/apE) (5.10)
for order refinement (p-refinement). Let
X E:mar( X(h) X (p )
E: E E
p - - - -
73
We use these quantities in the following adaptive refinement algorithm:
1. Find the unrefined element 3 with the largest 1K "
2. Determine the refinement cutoff X CUT < XK (e.g., as in
Section 5.5).
3. For each unrefined element E with 2E I XCUT
a. If Xi X ) refine the mesh in E.
b. If )< X(p), refine the order in E.
4. Take Si+ to be the resulting finite element space.
Mb (p)We now determine 'i and kP). First, we estimate p in each
unrefined element E. Suppose the order in E has never been refined, and
(from a previous finite element space in the sequence) we have an
approximation 4f(E) to the square of the error in f(E) before f(E) was
refined. By (5.6) and (5.7),
4f(E) tE 2
so we estimate
p = ( ln(Cf(1 )) - ln(QB) ) / ln(4) - 1.
On the other hand, suppose the order in B has been refined, and we have
an approximation CE to the square of the error in element 9 with kS one
less than its current value. If
74
4E' tE iPC=T(.1
for some constant pCUT < 1 (e.g., as in Section 5.4), there may still be
higher derivatives of u in E which can be exploited, so we estimate
- PE " Otherwise, we retain our previous estimate of p. In order for
our estimate of p to be logically consistent, we restrict it to the
interval (0,PE
Using our estimate of p, by (5.7) and (5.9), we estimate
[h) : 2 2 a C G h2p)/8h / 2 w h-2)lDh} (5.12)
=- P2 (a C Gh 2p) h- / ((-2) wh - 3)
-(P / w ) C
in element E.
There are two cases to consider in estimating )p). If p 2 m-l,
refining the order will not change the error at all (in (5.6)). Thus,
i) = 0. On the other hand, if p ( m-l, by (5.7) and (5.10),
(P) = ((E) / w,
where w' :- aw/Op. If u(z) C sin(Ox) sin(Oy) in E, where 0 is the
approximate "frequency" of u in E, then
1luf 12 C 0 2(p+l) ,
p+l, Ie
and by (5.6), since C(p) (l - for Taylor's theorem,
75
-2 2 2(p~)24E C2 (p0 h * 0 l h2 ~(.3
=C 2 (0 h) 2 (p~l (p 1) 2
Since e is independent of h and p. by (5.13),
e p12h . (5.14)
Using the crude bound a(pf)/dp I p p 1/2,
O(C /ap: 4F (2 WnO h) -p)
so we estimate
() Z 4 t: (-2nWeh)+ p) /1w' if p m-1 (.5
0 ~if p2.-i.
In practice, we estimate 4E by the squared error indicator ttIIE
(or ai!i~ I~II or 62) in formulae (5.12), (5.14). (5.15). and
the formulae f or ;
Expressions (5.12) and (5.15) indicate that X.tends to decrease as
E Is refined. Thus, the adaptive refinement algorithm tends to
approximately equalize U.). When the order is uniform, this reduce* to
"asymptotic equidistribution" of the errors amo&$ the elements (eog.,
Fried and Yang [201).
76
5.4 A symetrized model of predicted error behavior
In this section, we construct a second model of predicted error
behavior based on syumetrized a priori error estimates.
In considering combined h and p refinement, Babulka and Dorr [21
present an a priori error estimate of the form
I1e11 2 Cknax) G(z,m(z)) h(z) 2P(z)P(z)-2(m(z)-l)dz.(5.16)
Since C and G are independent of p, this estimate is more symmetric in h
and p than (5.5). Moreover, it indicates the manner in which
convergence occurs when p > w-1. Theorems and numerical evidence in
[11] show the exponent in p is optimal in the presence of singularities
of u not located at mesh nodes, and that if all the singularities of u
are located at mesh nodes, the exponent of p can double. The effect of
this doubling will be considered at the end of this section.
We make the heuristic assumption that (5.16) holds in eanh element,
with eauallit:
eillE (5.17)
:= C(knax) fialz) G(z,m(z)) h(z) 2(z) plz)-2(m(z)-l) dz.
Assumptions (5.4), (5.7), (5.8), and (5.17) comprise our second
heuristic model of predicted error behavior. The resulting estimates of
) , ( p) , and (X can be used in the refinement algorithm in
Section 5.3.
4eL
77
In our second model, if the order in unrefined element 9 has never
been refined, then by (5.17) and (5.7), we estimate
In (l(f(E)) - n(Q) )/ln(4) - 1.
On the other hand, if the order in B has been refined, then if < in-I,
E- h (P/(P-1))
else if p -,
4E/ 4 : (P/(P-1))-2(n-)
Since the function (p/(p-1))-2(p-l) equals 1/4 at 2, and is monotone
decreasing with limit e- 2 as k goes to * we decide whether to increase
our estimate of p on the basis of (5.11), with pCUT ° 2 In order for
our estimate of p to be logically consistent, we restrict it to the
interval (DIP1 ].
As in Section 5.3, we estimate
(h) w ) .18)
If p < in-, we estimate
P) (-2 In(h) + 2 (m-l)lp) I w'. (S.1)
Otherwise, we estimate
(P) (" 2 (mrl)/p) /'.(5.20)
IE
78
In practice, we approximate CE by the squared error indicator
I11;1 E (or 1111 1 _ or 82) in formulae (5.18), (5.19),
(5.20), and the formulae for p.
Suppose the exponent of p in (5.17) is doubled. Then our estimate
(h)(p)of X is exactly as before, and our estimate of XB changes by a
factor of 2 when p r a-i, and by a small factor when p < m-1 and hE is
small. The main effect is to encourage order refinement when a is small
and p > m-1. But then (p)E is usually much smaller than X , so the
practical effect on refinement behavior is slight. In general, the form
of (5.17) is more important than the values of the exponents.
5.5 Choosing the refinement cutoff
A simple choice for XCUT is
X U C X
for some constant 0 < C < 1, where
u= ax( ).
If X is too large, (5.2) may not hold. If X is too small, weCUTCUmay refine an element with small XE when it may be more efficient to
refine an element with large XE twice. We can avoid this situation by
letting XCUT approximate the largest XE for a once-refined element.
This quantity can be estimated by the marginal efficiency of refinement
for element R after it has been refined once:
79
If (h), and the order in B has never been refined, we can estimate
~ 2
C U T = ( p K / vX ) C 2 / f B
If )X = and the order in E has been refined, then by (5.13) and
(5.12) (or (5.17) in our second model), we can estimate
4 -(p +I) h)CUT
If -x - 4 P), then in our first model, by (5.13) and (5.15), we can
estimate
x UT X(P 0 33)2 / (p3+1) ( w,( 3j+1)Iw'(p+1)
and in our second model, by (5.17)-(5.20), we can estimate
XEST- (p) 2 -2 1)( )G w'(3 +1)
5.6 Drawbacks
The heuristic assumptions leading to our estimates of X) may be
too crude to yield useful refinement information. We base an alternate
refinement algorithm solely on M h) and p:
8o
1. Find the unrefined element 2 with the largest (h)
2. Determine X < Xh)CUT 9
3. For each unrefined element E with XE 2 CUT
a. If p > p, refine the mesh in E.
b. If p 5 p, refine the order in E.
4. Take S to be the resulting finite element space.i+l
If the order in element E is refined, p is estimated by (5.11).y o ll(P1)12 11e12
Then if 4E and 4E behave similarly to IIle(p-l)lII and IIletII, p
tends to be one more than its (usually) optimal value in-i. Conversely,
E e l E , then p tends to be less than m-1. The latter
situation arises most frequently when ks > kE on the sides of E. Thus,
in our current code, if k 2 kE+l on any three sides of E, or k 2 kE+2s E
on any side of E, then the next time E is refined, the order is refined,
regardless of the size of p.
Lack of de-refinement penalizes early order refinement. For
example, suppose h j 1/4 and k 2 3 are desired near the origin. If the
order is refined first, then k 2 3 everywhere (Figure 5-1).
The resulting finite element space may not satisfy (5.3). even if
the heuristic models of error behavior hold. For instance, if
Ai) XiP), it may still be more efficient to refine K in h if the
resulting decrease in is greater with mesh refinement than with order
81
0. 0 0II II I I II I I I I k-3 I k-3 II I I I I I I
k-2 I I k-3 0 oI I I I I I II I I I I k-3 I k-3 II I I I I I I
0 0 0 O0
order refinement first
o 0 O 0-0 0 - CI I I I I I I II I I k-2 I k-2 I I k-2 I k-2 II I I I I I I II k-2 I o -I I I I TIIIII I I k-2 I k-2 1 I k-3 I k-2I I I I I I I I00 O -0 O 0 0 0
mesh refinement first
Figure 5-1: Early order refinement penalty
refinement, and if the other 's are very small.
Rigorous models are preferable to the heuristic models used in this
chapter. A rigorous approach for o&-dimensional problems is developed
in Babulka and Rheinboldt [7).
5.7 Expected error behavior
In this section, we model the asymptotic behavior of error with
respect to work for three problem classes consisting of problems with:
1. smooth solutions.
2. solutions with point singularities.
3. solutions with line singularities.
82
We use the heuristic model from Section 5.3. We expect similar behavior
using the heuristic model from Section 5.4.
5.7.1 A smooth solution
Suppose u is smooth everywhere in L, and h(z) and p(z) are uniform.
By assumption (5.6),
~C(p) h2
and by assumption (5.4),
p A p h72 , ft > 2.
Minimizing 4 while holding W fixed, 2 p/h - 2/h T = 0 and
((C'/C) + 2 ln(h))t + X P/p W = 0 imply
h Z-/2
and
C e- ¥ p ,
where
:= p + C'/C.
Thus, it is best to fix the mesh and only refine the order. Suppose
C(p) = A? , A > 1.
Since C'/C - ln(A), then A e = A e- P-ln(A) < 1, so
= (A e-Y) p
is exponentially decreasing in p. But V = pp e is polynomial in p. so
if there exists A < * such that C(p) S Ap for sufficiently large p. then
is exponentially decreasing with respect to W.
*~*~ -
83
5.7.2 A point siagularity
Suppose u has an isolated singularity at a point a in 0, such that
near a, u has l+a derivatives in L2 , and u is smooth away from a.
Suppose there are unrefined elements at levels 1 to L in the mesh for
SL, and at each level only the elements which touch the point a are
originally refined. Suppose p(z) uniformly equals L. Then
C~)2+2aCE C ha) + if E touches a
and
C(p) hi 2 if E does not touch a.
By the proof of (2.4), there are no more than a constant number of
elements at each level, so
C (p) (( 2-L) 2+2a + L-l ( 2 -I)2+P)
C1 (L) 2- L
and
V 2 = 1
- C2 LP +l.
If CIL) AL for some A 2 as L goes to, then q is exponentially
decreasing with respect to W. Babulka and Dorr [2] make a similar
84
conjecture, and under suitable assumptions, prove that convergence is
faster than polynomial.
5.7.3 A line singularity
Suppose u has a line singularity at a line a in 0, such that near
2a, u has 1+a derivatives in L , and u is smooth away from a. To fix
ideas, suppose p(z) is uniform, a traverses 0 from top to bottom, and
there are elements at levels 1 to L in the mesh for SL # where at each
level the elements which touch the line a are present in SL Then
2+2atE C(a) hi if E touches a
and
2+2p
CE Clp) hE if E does not touch a.
There are at least 2L elements with level L touching a, so
SCl (a) 2L (2-L 2+2a
and
W C2(a) 2
Then
C3 (a) W-1-2a
regzdlest of how p is chosen. Convergence cannot be faster than
polynomial, since the complexity of geometrically approximating a
asymptotically dominates the cost.
AD-AI15 432 'ALE UNIV NEW HAVEN CT DEPT OF COMPUTER SCIENCE F/A 12/1LOCALMESH. LOCALORDER, ADAPTIVE FINITE ELEMENT METHODS WITH A--ETC(U)DEC 81 A WEISER N0001 -76-C-0277
UNCLASSIFIED TR-213 NL2fl/lllllllll
EI//I/EE///EEIEEE/I//I///EEIE//EEE//IL
lii. 128 11.liiU; mm*3
1111.2 '*.40 Jl -6IIIIIj8
MICROCOPY RESOL011ON TEST CHARTNATIONAL 1ALPAlT At11 0 AN[PARPI11-
RAFIPU 6
Computational aspoets
6.1 Introduotiom
In this chapter, we discuss some of the computational aspects of
the methods presented in Chapters 2-5. Since we have not yet
implemented all the procedures discussed, detailed operation counts are
omitted.
In Section 6.2, we consider the efficient assembly of the finite
element system (of. Eisenstat and Schultz [18], Weiser, Kisenstat, and
Schultz [31]). In Section 6.3, we investigate the complexity of sparse
direct elimination with nested-dissection-type ordering of the unknowns,
for the three classes of problems discussed in Section 5.7. The
operation counts and storage for the problems with singularities are
linear (optimal-order) in the number of elements. In Section 6.4, we
consider the efficient computation of the a posteriori error estimators.
We now outline the computational tasks to be performed.
There are four computational steps to perform for each finite
element space S in the sequence (Si ).
i$
nl
86
1. Form the linear system A U = f, where
fB1) is the basis for S,
U(xy) = .u3 r(x,y),
A.A E E (I
11 E -BE
I = E f : (f'BI)"- L E II
After initializing A and f to zero, for each E,
a. Assemble A and fE. the element stiffness matrix and
right hand side for the basis (IE) , where (I is the
basis which would result if E had no irregular corners.
b. Obtain AE and fE from A and ?E, by changing variables
from (B to (B
c. Add A and f to A and f.
2. Solve A U - f.
3. For each B, compute Il'IllE , IIsI6_lE, Il!is ,or &E*
4. Depending on the error indicators computed in Step 3, either
stop, or adaptively choose Si+1 , and go to Step 1.
Each integral fj v(z,y) dx dy is approximated by
G(E) G)(E) t=E ws wt v(x sy), (6.1)8- I
87
whete (w,,,z) and (w tlyd are the weights and points for the G(E-point
Gauss-Legendre integration rules on Cxj+h1 and [yE9y1*h3 ]
respectively. Choices for G(E) are specified in Chapter 7.
6.2 Assembling the linear system
Consider Step l.a (Section 6). Fot simplicity. suppose
E (v~w) - (a(14") VMw)
the bilinear form for a weighted least squares problem, since
considerations for general a~% are almost identical.
Since (i1)is a subset of
sparse versions of the tensor-product assembly procedures in Eisenstat
and Schultz [18] and Weiser, Bisenstat, and Schultz [31] are
appropriate. If
B3- i(KP~)
B - p (z)p (y),I J n
then, depending on whether a(z,y) and f(z,y) are non-separable,
separable, or constant, Aand f3 can be computed using the formulae
88
(a~zy B1,J0
(w p (z )p (x )> a (x ) ~ #w~
3 2 1 W5 0I~P x 3 t(t,)ny)) aytut
(f W().5
The quantities in angle brackets 0> can be precomputed independent of E.
The quantities in squatse brackets 11 should be computed as intermediate
sims in each B. Since
NB ' - ~- -EAlla (i~s)(j'n) Ain(Jm (jA,m)(i.n) MA(Jon)(im)
only entries in A with i I .1 and a I a should be computed explicitly.
Consider Step 1.b (Section 6). For simplicity, we discuss Step 1.b
only for element E in Flours 6-1, when kE is uniform. Other cases are
handled analogously.
Figure 6-1t An element with one irregular corner
The [B I) nonzero in B have the values
E a
E W f(E) Y
Suppose
f(E), ns.8(2.3)U (y) - (v). (6.2)
Since refinement is regular, the values (vL can be precomputed
EEindependent of E, and Aand f can be obtained from the formulae
90
E -EA(i,m)(J,n) A(i,m)(J,n) i02. J02,
V,L A ) i-2, J02,
EA(jn)(im) i2, J=2
and
fE ?Em)i(i'm) in
The resulting computation is relatively expensive when kE is small.
When a(x,y) is separable, fever multiplies are required if Step l.b is
omitted, and AE and fE are formed directly.
6.3 Solving the linear system
Consider Step 2 (Section 6), solving the linear system A U - f. In
this section, we investigate the complexity of sparse direct elimination
with nested-dissection-type ordering of the unknowns for the three
classes of problems discussed in discussed in Section 5.7.
6,3.1 A smooth solutioa
As in Section 5.7, suppose U is smooth everywhere in Q. Suppose k
and h are uniform, and there are N :_ h72 elements. There are
(k-4)(k-3)/2 element basis functions nonzero in each E. In sparse
A
91
Gaussian elimination, if the element basis functions are ordered first,
their elimination induces no fill-in (i.e., they are zstnaticall
condensed 1281). This initial elimiation requires N((k-4)(k-3)/2)3/16
multiplies and :NC(k-4)(k-3)/2) 2/2 storage. Although this cost
dominates when k )) N. in solving problems to usual accuracy.
elimination of the element basis functions Is relatively inexpensive.
There are : (2k-3)N remaining basis functions Mk-2 side basis functions
associated with each distinct element side, and I vertex basis function
associated with each distinct vertex).
Consider a nested dissection ordering of the side and vertex basis
functions, In conjunction with sparse elimination of A (e.g., George and
Lin [22]). A sample nested dissection ordering is shown in Figure 6-2.
0- c--a 0 00
I _ I I I I I I I "I0-0 0 -- '- -0----O 0I 1 1
O-0 0 c- o 0 0--c--0
c-o-0 0
0 0 I r I a I I 'I4 3
I----- 0-0'--- -- __7_
92
1-4 correspond to basis functions on uniform submeshes with N14
elements: in each group the unknown& are recursively ordered using
nested dissection on the submesh. Since group 5 separates the unknowns
in the first four groups, the number of multiplies to eliminate the
unknowns in groups 1-5 is. to dominant-order terms, bounded by U(N),
where
1(N) - 4 M(N/4) + C (kNI/2 )3 .
The rightmost term is a bound on the number of multiplies needed to
densely eliminate the unknowns in group 5. The solution to this
recurrence relation is O(k3N/ 2) (e.g., Rose and Whitten [26). Theorem
1), as is the operation count to eliminate the remaining unknowns in
group 6. Thus, the expected operation count, neglecting static
condensation cost, is O(k3N3 / 2). Similarly, the expected storage is
O(k 2Nlog(k 2N)).
6.3.2 A point singularity
As in Section 5.7. suppose u has an isolated singularity at the
point ar(0.0). Suppose there are unrefined elements at levels 1 to L in
the mash, and at each level only the elements which touch a are present.
Suppose k is uniform.
There are N : 3L unrefined elements in the mesh. After static
condensation, : (2k-3)N side and vortex basis functions remain.
Consider a recursive ordering of the side and vertex basis
93
functions, in conjunction with sparse elimination of A. A sample
ordering is shown in Figure 6-3.
3 0
I I II I II I I
I 2 II I I
0------0C-
I II I I II Ii ' I I
0 I
I II I I0-o-0---o I I00-0 I I o o-o--o---- 0
Figure 6-3: Recursive ordering with a point singularity
The unknowns are ordered according to the groups numbered 1-3. Group 1
corresponds to basis funotions on a submesh with N-3 elements: the
unknowns in group 1 are recursively ordered on the submesh. The number
of multiplies to eliminate the unknowns in groups 1 and 2 is, to
dominant-order terms, bounded by 1(N), where
N(N) - I(N-3) + C k3 .
The rightmost term is a bound on the number of multiplies needed to
densely eliminate the unknowns in group 2. The solution to this
recurrence relation is 0(WN), and the operation count to eliminate the
remaining unknowns in group $ Is O(W3 ). Thus, the expeoted operation
94
count, neglecting static condensation cost, is O(k3N). Similarly, the
expected storage is O(k 2N).
6.3.3 A line singularity
As in Section 5.7, suppose u has a line singularity along the line
a - ((O,y)). Suppose there are unrefined elements at levels 1 to L in
the mesh, where at each level the elements which touch the line a are
present. Suppose k is uniform.
- LThere are N - 3(2 ) unrefined elements in the mesh. After static
condensation, ; (2k-3)N side and vertex basis functions remain.
Consider a nested-dissection-type ordering of the side and vertex
basis functions, in conjunction with sparse elimination of A. A sample
ordering is shown in Figure 6-4. The unknowns are ordered according to
the groups numbered 1-4. Groups 1 and 2 correspond to basis functions
on submeshes with (N-2)/2 elements: in each group the unknowns are
recursively ordered on the submesh. Since group 3 separates the
unknowns in the first two groups, the number of multiplies to eliminate
the unknowns in groups 1-3 is, to dominant-order terms, bounded by U(N),
where
31(N) = 2 1(N/2) + C (klogN)
The rightmost term is a bound on the number of multiplies needed to
densely eliminate the unknowns in group 3. The solution to this
recurrence relation is O(k 3N) (e.g., Rose and Whitten [261). and the
iT ~ a 2
95
---- 0o-4---o0-0-0 1o-o-ol I I 1
0-0-0---- 0o--0 I Ioo-o--o I I
0-0-0 I Io-o-o---o------ o-3 --
-0-0 1,-0-0 i : r -.0
0-0-0-0 "o-o-o 2 I0-0-0 - - -
0-0 --00-0-0--'0
0-0 1 Alio -oo--o---- 0-- .. 0
Figure 6-4: Nested-dissection-type ordering with a line singularity
operation count to eliminate the remaining unknowns in group 4 is
O((klogN) 3). Thus, the expected operation count, neglecting static
condensation cost, is O(k3N). Similarly, the expected storage is
2O(k N).
6.3.4 Discussion
The operation counts and storage for the problems with
singularities are optimal-order in the number of elements N. We expect
similar operation counts for more general problems with point or line
singularities. The optimality for the line singularity problem depends
on the fact that we can systematically find sets of O(N1 13-) elements
(a ) 0) which separate the remaining elements into two or more disjoint
sets of equal size.
96
We do not know an efficient (i.e., O(N)), automatic, way of finding
orderings such as those presented. See George and Liu [22], Sections
7.2 and 8.2, for work in this direction.
In the example problems, the orderings of the interior unknowns
correspond to the reverse of the order in which the edges associated
with the unknowns are created. This correspondence does not always
hold. (For example, suppose u has a point singularity at (1/2,0).)
6.4 Computing the error indicators
Consider Step 3 (Section 6), the computation of the error
indicators iia-III E IIiI E , l lil1l , or sE * Recall
FE(v) :- (f'v)E + <[aO nl,v> - aElUv).
First, suppose E has no irregular corners. Then the possible quantities
to be computed in Step 3 are
1. (a (B,B)), where (B) is the basis for SE which would
result if E had no irregular corners. These quantities can
be computed as in Step l.a.
2. [(f,BI)E : in Si-SE). These quantities can be computed
as in Step l.a.
3. [E(UaBI) BI in 1j-SE First, convert the representation
U(XY)l - I Us BI(X'Y)
* t
97
into the form
For example, for element E in Figure 6-1, when k is uniform,
if
U(x'y)I E 2 #2 m U(i,m) PEtx) pE(y) +
p2U E pm(E) (y),m U(2,m) P2E Pm
then
Ui)-{ U (j,) 1 2,
ilL=L(m) vLm (Ni,L) i=2
where
Other cases are handled analogously.
Then, for 111;111IE 0 compute IaE (U. Il with a matrix-vector
multiply. For Isj Eor 11811 , compute (U(x,,Yt) by
the formula
U(xs#yt) m P3(yt) I i Pi(x) 8 (im) (6.3)
98
and then use the formula
I ~ ~(a(x,y) U,Bi) <w (WtnYt
[ p ( px)> [ U(xsyt)a(xs.yt)
4. {(<[aU/8u IuiI>E• In a preprocessing step, compute and
store the coefficients of the one-dimensional polynomials
U/n on each element side. Then, for example, on the left
side of element E, compute (a8U/an(x EYt )IE and
(a8U/an(xE.yt)IE,), where element E', if it exists, shares
part of the left side of E. Form
((sOg/OnJ IBI)aE(left) -
t <Wtpm (yt)/2>[aaU/an(xEYt)HE -aU/Dn(x Eyt) E#]
for each B I pi(x)pm(y). Other cases are handleda
analogously.
5. E ,1 , or 111-1112 . Suppose the system for
The storage required can be reduced by intermixing the preprocessing
step with error indicator computation.
99
the coefficients of aBis A v f. Compute a lower
triangular matrix L such that L L T- A, and solve L y - f.
(For and 'KL is independent of hE. Then compute
8' E
6. 11fLUII2 . for E First compute (Ux# sin :6.3),
lefthsieo compute
~~~~' (a1>(aU8(xsyt)~sp 8K+aOJO(Etyl)
No. s~aupposn as n ifrreuar corerors simplc ity heol
co deometin fgr 6-1.8' Ihn > SE are unin, for m exml, other
SNce refineme is airegular cvaler~.p can bemprcomputwedol
independent of E, and the right hand side components have values
100
whenB is in SE-SE and i 2, and
f(i~m) L P-k+1 wmP LP £Li vp,L PE'(iL) )
when k < a n. k and 1 2. The rest of the computations are unchanged.
CRAP=l 7
Numerisal results
7.1 Introduction
In this chapter, we present numerical results obtained using
prototype codes written in FORTRAN. which implement some of the methods
discussed in Chapters 2 - 5. The problem set, containing one problem
from each of the three classes considered in Section 5.7, is specified
in Section 7.2. In Section 7.3, we present the error behavior, which is
in fair accord with Section 5.7. In Section 7.4, we present the error
estimator behavior: the error estimators are usually accurate to within
a factor of two. In some oases, some of the estimators appear to
converge to the norm of the true error. In Section 7.5, we present more
numerical evidence of error estimator convergence.
7.2 The problem set
Prototype codes were written in FORTRAN, implementing some of the
methods discussed in Chapters 2 - 5. Numerical tests were conducted to
investigate the behavior of the methods.
The test problem set consisted of three Poisson problems with
101
, oW 1 , -
102
Dirichlet boundary conditions,
D2u - D2u - f(zy) in the interior of A,x y
u(xy) = g(xy) on 80,
with solutions specified in Table 7-1.
II Problem Solution class II Solution IIII II II III
II 1 smooth II u(.y) - cos(e +Xy) IIII II II IIII I I - - I I----**--IIII II II 12IIII 2 point II u(xy) r sin(012) , IIII II I IIII Isingularity II - r cos(e) . y - r sin(C) IIII II II IIII II II
II 3 line II u(x.y) - 0 t -1/2 II
II IIsingularity II sin(wt) -1/2 < t ( 1/2 IIII II II It1I 1/2 S t * IIII II II
S II II t = Sx-y-1 IIII II II II
Table 7-1: Test problems
The solution to Problem 1 is in fk(f) for all k. The solution to
Problem 2 is the fundamental esenfunction for a crack problem with
interior angle 2w; this function is in B312-*(G) for any a ) 0. but not
in 3312(a). The solution to Problem 3 is a smeared shock wave with
shock-width 1/5 in z; this function is in N2(0). but not in 22 +(Q) for
103
any s ) 0.
We tested the seven refinement methods specified in Table 7-2. All
adaptive sequences of finite element spaces started out with 2X2 meshes.
The adaptive sequences for methods + and X started out with ks- 2.
Method Ih IIk_II I
U uniform (1/2,...,1/32) II - 2II II II IK - 1/2 II adaptive (Section 5.4)2 adapv ( o 52 adaptive (Section 5.3) II - 2
3 adaptive (Section 5.3) II - 3
II 4 IIadaptive (Section 5.3) II - 4 I+ adapte a v (+ adaptive II adaptive (Section 5.6)
II I adaptive II adaptive (Section 5.4) I
Table 7-2: Refinement methods
We used three heuristic refinement procedures, based on the
formulae in Chapter 5. In the "true error" procedure, we set
= 11181o112 In the "approximate error" procedure, we set
M= I!,I!I , with S composed of piecewise polynomials one order
higher than S. In the "residuals and Jumps" procedure, we set 4B 2
For smplioity, we got xiS - X1/2. To avoid ill-conditioned linear
systems, we restricted k1 . 7. Our tests were run in single precision
on a DEC-System 2060 computer.
104
G(E)-point Gauss-Lesendre quadrature rules (6.1) were used to
approximate integrals in each element E. In the original assembly
phase, G(E) = kmaxE , where
kmaxE : max(kE, : El is in N(E)).
In the error indicator phase, G(E) - kmazE+l. In the evaluation of
111111E , G(E) - kmaxE+3 . With these values, for with smooth
solutions, the errors due to numerical quadrature are asymptotically
negligible (e.g., Ciarlet [17], Section 4.1).
7.3 Error behavior
In this section, the resulting errors for the adaptive sequences of
finite element spaces are depicted on 1og10-1og10 scales. The
x-coordinate of each data point is the number of nonzeroes in the
stiffness matrix , and the y-coordinate is 11le101. Since the true work
generally grows faster in k than the number of nonzeroes, the measure of
work is biased in favor of high-order methods.
Figure 7-1 presents the errors due to applying the "true error"
Our prototype codes should be made more efficient before computer
time is used as the measure of work.
105
procedure to Problem 1. Sines u(xvy) is smooth, nonuniform meeh
refinement does not improve the asymptotic oonversence rate. The slopes
hk-i)srran ~7of the lines correspond to Oh - ) error and O(h 2k4 ) work. Since k
increases for methods K, +, and I, these methods achieve faster than
polynomial convergence. Methods + and X perform some initial mesh
rfinesent (see Figure 7-2, with k. labelled in each element), so they
are somewhat less efficient than method K.
Figures 7-3 and 7-4 present the errors due to ippiyJ.u the
"approximate error" and "residuals and Jumps" jroedu-1s to Problem 1.
The errors are similar to the ones shown in 1-1,
Figure 7-5 presents the errors due to appl-ing the "true error"
procedure to Problem 2. The slope of the line for method U corresponds
to an O(h 1 / 2 c) onvergence rate and an O(h -2 ) work count. Method K has
almost the same line, with slope corresponding to an 0(k -1 ) convergence
4rate and an 0(k ) work count. The slopes of the lines for methods 2. 3,
With the work count proportional to the number of unknowns, method K
would appear twice as efficient as method U (see [111). With a more
realistic work count, method K would asymptotically be the least
efficient method considered.
II
106
and 4 are the same as the slopes for the same methods in Figure 7-1.
Local mesh refinement, in effect, transforms variables so the problem
looks like it has a smooth solution on a uniform mesh. Methods + and X
are fairly efficient in all cases, refining the mesh near the
singularity, and refining the order where the solution is smooth (see
Figure 7-6).
Figures 7-7 and 7-8 present the errors due to applying the
"approximate error" and "residuals and Jumps" procedures to Prot-lem 2.
The errors are similar to the ones shown in Figure 7-5.
Figure 7-9 presents the errors due to applying the "true error"
procedure to Problem 3. The slope of the line for method U corresponds
to an O(h) convergence rate and an O(h -2 ) work count. The line for
method I appears to level out, reflecting an O(k 4 ) work count and slow
convergence in k. The slope of the line for method 2 reflects an
improving (( N-112 ) convergence rate and an O(N) work count, where N is
the number of elements. At the modest accuracies required, the mesh is
refined not only along the line singularities at the edges of the
boundary layer, but in the boundary layer as well (see Figure 7-10). We
expect the slope of the line for method 2 to approach -1 asymptotically,
reflecting an O(NO) convergence rate. The slopes of the lines for
methods 3, 4, +, and I reflect convergence rates tending toward O(N-).
Figure 7-11 depicts a sample grid for method +. Note that the shape of
the boundary layer cannot be discerned from the grid.
. . .. ... . . u m m lm m m wm m m omm •
r1
107
Figures 7-12 and 7-13 present the errors due to applying the
"approximate error" and "residuals and jumps" procedures to Problem 3.
The errors are similar to the ones shown in Figure 7-9.
In this test, methods + and X appear to be the most robust, since
they are fairly efficient for all three problems, regardless of whether
the errors are large or small. Each of the other methods does poorly
for some problems in some cases. Table 7-3 s=-arizes the observed
efficiency a, where
Illeli (the number of nonzeroe8)-a .
Method J Problem 1 II Problem 2 JJ Problem3 JIII II-J It
U J 1/2 JJ 1/4 JJ < 1/2 JJ
,I K (k-1)/2 II 1/4 II : 1/2 IIII I II II2 II 1/2 II 1/2 II > 1/2 IIIIII-- - II II II
II 3 II 1 II 1 II 1 IIIIII II II -,II
4 II 3/2 II 3/2 It : 1 II
+ II (k-1)/2 II (k-1)/2 II - 1 IIIIII - II II IIX II (k-X)/2 II (k-1)/2 I 1 IIIIII II II It
Table 7-3: Efficiency susmary
7.4 Error estimator behavior
Figures 7-14 - 7-19 depict the relative errors p and & in the error
estimators for the tests run using the "approximate error" and
* -
108
"residuals and jumps" procedures respectively.
In the "approximate error" procedure, the error is usually
estimated within a factor of two (-.5 j p j 1), and always
-.7 1 - 1.6
p appears to converge to zero in methods U and 2 for Problem 1, and in
methods 2 and 3 for Problem 2 (see Section 7.5). For Problem 1. p
varies somewhat erratically in methods + and X, especially when k, is
large. This may be partly due to ill-conditioning.
In the "residuals and jumps" procedure, the error is always
estimated within a factor of two. In particular, in method 3,
-.23 Sa .20,
and g appears to converge to zero. In methods + and X, 2 varies more
smoothly than p. Thus, in this test, the "residuals and Jumps"
procedure appears to be slightly more robust than the "approximate
error" procedure.
7.5 Error estimator convergence
As further evidence of apparent error estimator convergence, we now
present the behavior of the four error estimators 111,111. 111111,
1lll1l, and a. with relative errors p. p, p. and & respectively, for a
sample variable-coefficient problem with a smooth solution. k and
h - 1/n are uniform. S contains piecewise polynomials one order higher
-- i it " II -I. ... .. . " -1- '-
109
than S. We have observed similar error estimator behavior for other
problems with smooth solutions.
Table 7-4 depicts the relative errors in the error estimators for
the Dirichlet problem
- D (exyD u) - D (exyD u) + u/(l+x+y) - f(xy)x x y y
in the interior of 0,
u(x,y) = 0 on 80,
where f(x,y) is chosen so that u(x,y) = exy sin(nx) sin(xy).
In Table 7-4, the error estimators are always within a factor of
two of Illelll. None of the error estimators converge to Illelil with k
increasing and h fixed, but some appear to converge with h decreasing
and k fixed. 111;111 and 111.111 appear to converge when k j 4.
Iligill appears to converge when k ( 4. but does not converge when
k - 4. The convergence rates are not clear. When k 2 or 3. i behaves
like a constant times Illelil.
110
if n 1f1116.111 if-- l-I-l III I
k=2 2 .181(+1) .053 .026 .026 II .264 .877 .028 .024 .022 II .378 .436 .012 .011 .010 II .40
16 .218 .0039 .0037 .00331I .4032 .109 .0011 .0010 .0009 II .40
1- - I 1llell pI I-In I11.1 II II
k-3 2 I .515 -.46 -.45 -.45 II -.154 I .857(-1) II -. 12 -.11 -. 12 II -.00888 i .194(-1) II -.022 -.022 -.022 II .018
16 I .477(-2) II -. 0041 -.0038 -.0043 II .013
In pI1111 I
k=4 2 I .443 .33 II .29 .29 -.44 II4 I .426(-1) .12 II .10 .039 -.39 II8 I .446(-2) .043 II .038 -.11 -.35 II
16 I .508(-3) .019 II .018 -.25 -.32 II
II n II lllelll II P II £ II P II 2 II
k-5 II 2 II .671(-l) II .35 II .34 II .34 II -.36 IIII 4 II .426(-2) II .23 II .22 II .18 I-.42 IIII 8 II .256(-3) II .20 II .19 II -.026 II -.48 II
II n II lllelll II II II P II 2II----II I-- I I--**~-I I II I I---I I
k-6 II 2 II .118(-1) II .31 II .29 II .29 II -.25 IIII 4 II .508(-3) II .23 II .22 II .099 II -.29 IIII 8 II .156(-4) II .32 II .32 II .077 II -.26 II
II n II1lllell II p II 2 II P II 2 III II--I I - I ----- II II II
k-7 II 2 II .213(-2) II .36 II .35 .35 3 II -.14 IIII 4 II .380(-4) II .36 II .35 II .31 II-.19 II
Table 7-4t Error estimator behavior
1* -- *.iA0l2
111
Figure 7-1: Problem 1 - true error
SMOOTH - TRUE ERROR
1. DE-O2 x-+
I.OE-03
I.-01. OE-0
1.ED . .... ' .... ..... : I ....
1.OE01 !.OE02 1. OE*O3 1. OE04 .OE+05NONZEROES
U H UNIFBORM K=2K H = .5 . K RDPTIVE2 HRRDPTIVE , K = 23 H A PTIVE .K : 34 HRDPTIYE K = 44 H RDRPTIVE , K RDAPTIVE. DISCRETEX H ADAPTIVE , K FDRPTIVE. SYMMETRIC
112
Figure 7-2: Sample grid for problem 1 (method +,residuals and Jumps)
VT 34 -313
4
4 4
113
Figure 7-3: Problem 1 - approximate error
SMOOTH - RPPROXIMRTE ERROR1. OE+OO
1. OE-O
1.OE-02
S1. OE-03LI
1. OE-04
,x xi 1. OE-05 ~X+)4+
1.OE+O1 1.DE+02 1.OE+03 1.OE+04 1.OE+0S
NONZEROES
U H UNIFORM K = 2K H = .5 K ADAPTIVE2 H ADAPTIVE . K = 23 H ADAPTIVE . K = 34 H ADAPTIVE . K = 4+ H ADAPTIVE . K ADAPTIVE. DISCRETEX H ADAPTIVE , K ADAPTIVE. SYMMETRIC
I
114
Figure 7-4: Problem 1 - residuals and jups
SMOOTH - RESIDURLS RNO JUMPS1. OE-O00'
1. OE-02
1.OE-03La-
1. OE-04 +X+X
1. OE-05 X
xx
1.OE-06 . . ...... ' . . . . . . .. -I , . . .....1.OE+O 1. OE+02 1. OE+03 1.OE+04 1. OE+05
NONZERBES
U H UNIFORM . K = 2K H = .5 . K ADAPTIVE2 H ADAPTIVE . K = 2S H ADAPTIVE . K = 34 H RDAPTIVE .K = 4+ H ADAPTIVE . K ADAPTIVE. DISCRETEX H ADAPTIVE . K ADAPTIVE. SYMMETRIC
115
Filgre 7-5: Problem 2 - true error
POINT SINGULRRITY - TRUE ERROR1. DE+DO
1. CE-02
I.OE-02 -
. 4 x
I.DE-D2
1.OEO1 1. OE.02 1. OE 03 1.OE+04 1. OE 05
NONZEROES
U H UNIFORM K = 2K H = .5 , K ADAPTIVE2 H ADAPTIVE, K = 23 H ADAPTIVE K = 3I H ADAPTIVE K= 4+ H ADAPTIVE , K ADAPTIVE. DISCRETEX H ADAPTIVE , K ADAPTIVE. SYMMETRIC
116
Figtr 7-6: Sample grid for problem 2 (method +, residuals and Jumps)
3 2
22
2 2
222 2
117
Figure 7-7: Problem 2 - approximate error
POINT SINGULARITY - RPPROXIMATE ERROR1.OE+O0
1. oE-oI
1. OE-01LUJ
1.OE-O3
1.OE-04 I .
1.OE+O1 1. OE+02 1. OE+03 1. OE+04 1.OE05
NONZERSES
U H UNIFORM . K = 2K H = .5 . K RRPTIVE2 H ADAPTIVE . K = 23 H RDPTIVE , K = 34 H RDRPTIVE . K = 4+ H ADAPTIVE . K ADAPTIVE. DISCRETEX H ADAPTIVE . K RDAPTIVE. SYMMETRIC
118
Figure 7-3; Problem 2 - residuals and jumps
POINT SINGULARITY - RESIDUALS AND JUMPS1. OE O0 I. . .pl . .. .. .. . . .. .
I. OE-01
1. OE-O2
1.OE-03
1.OE-041.OE+O1 1.OE+02 1.OE+03 1.0E,04 1.OE+OS
NONZEROES
U H UNIFORM . K = 2K H = .5 . K ADAPTIVE2 H ADAPTIVE . K = 23 H ADAPTIVE , K = 3I H ADAPTIVE . K = 4+ H ADAPTIVE . K ADAPTIVE. DISCRETEX H ADAPTIVE . K ADAPTIVE. SYMMETRIC
119
Figure 7-0: Problem 3 -true error
LINE SINGULARITY -TRUE ERROR1. OE+01
1. OE.OO
1.OE-02
1.OE.O1 1.0E+02 1.OE+03 I.0E+04 1. OE+05NONZEROES
U H UNIFORM K =2K H = .5 ,K ADAPT IVE2 H ADAPTIVE . K =23 H ADAPTIVE . K =34 H ADAPTIVE . K =4+ H ADAPTIVE . K ADAPTIVE. DISCRETEX H ADAPTIVE . K ADAPTIVE. SYMMETRIC
120
Figusre 7-10: Sample grid for problem 3 (method 2, residuals ant jumps)
OWN"
121
Figure 7-11: Sample grid for problem 3 (method +o residuals and jumps)
3 3 44
3 3
3 3 4
3 3 3 3 3 33 3
3 3 3 3 3 3
33 3 3 3 3 3
3 3 33 3 3 3 3 3
2 3 3 3 32 2 2
_ _ _ _3 3 3 32~2
3 2 3 32 22 2 2 2
S 22
2 2
2 2 2 2 3 322 22 2 2 3 3 3
2 2 2 2
3 3 3 3 3 3
3 3 3 3 3 3
1
122
Figure 7-12: Problem 3 -approximate error
1.O+O1 LINE SINGULARITY A PPROXIMATE ERROR
1.0E+00
1. OE-01
i.OE.1 1.~eO2 NONZEROES
KH = .5 .K ADAPTIVE2H ADAPTIVE. K =2
3H ADAPTIVE . K =
+H ADAPTIVE . K ADAPTIVE. DISCRETEX H ADAPTIVE . K ADAPTIVE. SYMMETRIC
123
Figure 7-13: Problem 3 - residuals and Jumps
LINE SINGULARITY - RESIDUALS AND JUMPS1.OE+O1. ......... ,
1. OE.OO
oU I.OE-O01 +
I.OE-O02
1.OE+O1 I.OE+02 1.OE+03 1.OE+04 1.OE+05
NONZEROES
U H UNIFORM K : 2K H = .5 .K RDAPTIVE2 H ADAPTIVE K = 23 H RDRPTIVE K = 34 H ADAPTIVE, K = 4+ H ADAPTIVE , K ADAPTIVE, DISCRETEX H ADAPTIVE . K ADAPTIVE. SYMMETRIC
I.
124
FiSure 7-14: Problem 1 - approximate error
SMOOTH - RPPROXIMRTE ERROR2.000. ......................
X
0.000
+x
1.000
0.000 ' .- J . +
-1.000. . .. . . .. . . . . ., , , . . .1.OE+O1 1.OE+02 1. OE+03 1.0E+04 1. OE05
NONZEROES
U H UNIFOR , K = 2K H = .5 . K ADAPTIVE2 H ADAPTIVE . K = 23 H ADAPTIVE , K = 34H ADAPTIVE K =4
+ H ADAPTIVE . K ADAPTIVE. DISCRETEX H ADAPTIVE . K ADAPTIVE. SYMMETRIC
125
Fiure 7-15: Problem 1 - residuals and Jumps
SMOOTH - RESIDUALS RND JUMPS2.000
1.000
0.000
X + X x *x
-1.000l.OE+01 1. E+02 1.OE+03 1. OE+04 1. OE+05
NONZEROES
U H UNIFOR ,K : 2K H = .5 K RDRPTIVE2 H RPTIVE, K = 23 H RPTIVE K = 34 H ADAPTIVE K = 4+ H ADAPTIVE , K ADAPTIVE. DISCRETEX H ADAPTIVE . K ADAPTIVE. SYMIETRIC
126
Figure 7-16: Problem 2 -approximate error
POINT SINGULRRITY -RPPROXIMRTE ERROR
2.00
1.000
0.000
-1.00......I1.OE+01 1. OE+a2 1. 0E403 1. OE404 1.ODE+05
NONZEROES
U H UNIFORM . K =2K H =.5S SKADAPTIVE
2 HRADAPTIVE.K =23 H ADPTIVE K =34 M ADAPTIVE. K =4+ H ADAPTIVE . K ADAPTIVE. DISCRETEX H ADAPTIVE , K ADAPTIVE. SYMMETRIC
127
Figure 7-17: Problem 2 - residuals and Jumps
POINT SINGULARITY - RESIDUALS RND JUMPS2.000. ........ ,. ....
1.000
-0.000
1.OE+01 1.OE+02 1.OE+03 1. OE+04 1.OE.05NONZEROES
U H UNIFORM . K : 2K H : .5 . K OAPTIVE2 H RDTIVE .K = 23 H DAPTIVE , K a 34I HADRPTIVE . K : 4+ H ADRPTIVE . K ADRPTIVE. DISCRETEX H ADAPTIVE . K ADAPTIVE. SYMMETRIC
128
Figure 7-18: Problem 3 - approximate error
LINE SINGULARITY - RPPROXIMRTE ERROR2.000
1.000
-I.000
-1.000 ... . . .......... . ..... .I . . . . .I . . . .
1. OE+01 1. OE+02 1. OE+O3 1. OE+04 1. OE+05
NONZEROES
U H UNIFIRM o K : 2K H : .5 , K ADAPTIVE2 H ADAPTIVE .K = 23 H ADAPTIVE . K = 34 H AOPTIVE , K = 4+ H AOPTIVE . K ADAPTIVE, DISCRETEX H ADAPTIVE . K ADAPTIVE. SYMMETRIC
.. .. , .I;,
129
Figuze 7-19: Problem 3 -residuals and jumpa
LINE SINGULARITY -RESIDUALS AND JUMPS2. ID[
~w -1.D14 .OE.01 I.0E402 1.OE.03 1.OE+04 1.OE+0S
NONZEROES
U H UN!FU -K .(2K H =.5 K ADAPTIVE2 HRADAPTIVE.K =23 H ADAPTVE ,K z34 M ADAPTIVE ,K z
XH ADAPTIVE .~~ K DPIE YMTI
Bib 1 lography
[1] S. Agmon.Lectures on Ellivtic Doundary Value Problems.Van Nostrand, New York, 1965.
[2] I. Babulka and M. R. Dorr.Error estimates for the combined an p versions of the finite
element method.Technical Report BN-951, Institute for Physical Science and
Technology, University of Maryland, 1980.
[31 I. Babulka, I. N. Katz, and B. A. Szabo.
Hierarchic families for the p-version of the finite elementmethod.
In Proceedings of the Third IMACS International Jyos- 2nComnuter Methods for Partial Differential Eguation., pages278-286. Lehigh University, Bethlehem, Pennsylvania, 1979.
[4] I. Babulka and A. Miller.A Posteriori error estimates and adaptive techniaues for thie
finite element method.Technical Report BN-968, Institute for Physical Science and
Technology, University of Maryland, 1981.
[51 I. Babulka and W. C. Rheinboldt.A posteriori error analysis of finite element solutions for
one-dimensional problems.SIAM Yournal on Numerical Analysis 18:565-589, 1981.
[6] I. Babulka and W. C. Rheinboldt.A posteriori error estimates for the finite element method.International Zournal for Numerical Methods " Engineering
12:1597-1615, 1978.
[7] I. Babulka and W. C. Rheinboldt.Analysis of optimal finite-element meshes in R1.Maknatio-s gf omouttii a 33:435-463, 1979.
[8] 1. Babulka and W. C. Rheinboldt.Error estimates for adaptive finite element computations.JIA MJournal 9n Numicls AauliuLS 15:736-754, 1978.
130
131
[9] 1. Babulka and W. C. heinboldt.On the reliability and optimality of the finite element method.foraulx ad Structures 10:87-94, 1979.
(101 I. Babulka and W. C. Rheinboldt.Reliable error estimation and mesh adaptation for the finite
element method.In .onstatiLa3lMthds nlinar Mechanics, pales 67-108.
North-Rolland. New York, 1980.
[11] I. Babulka, B. A. Szabo, and I. N. Katz.The p-version of the finite element method.x JoLunal gn Nuuziaul Analyis 18,515-S45, 1981.
[123 I. Babulka and B. A. Szabo.oh the rates gf sacyn9erUn if fhaite elemnt method.Technical Report 80/2, Center for Computational Mechanics, 1980.
[13] R. E. Bank and A. Sherman.A multi-level iterative method for solvin8 finite element
equations.In r,esAinau of .th Fft1 fmlaSLm= IS Rezsvoir IJSaLsLoAn.
pages 117-126. Society of Petroleum Engineers of AINE, Dallas,1979.
[14] R. E. Bank and A. Sherman.A refnem alzrithm and AINa dAt s.trus.tuar I"finite
element ms.Technical Report 166. University of Texas Center for Numerical
Analysis, 1980.
[151 R. E. Bank and A. Veiser.In preparation.
[16] A. Brandt.Multi-level adaptive solutions to boundary-valuo problems.Mateats. Ll Cast tatiJLJsl 31:338-390, 1977.
[171 P. 0. Ciarlet.
North-Iolland, Amsterdam, 1978.
[18 6. C. Eiseastat and M. 1. Behalts.Computatiomal aspects of the finite element method.In lh Na kmaial Enudutlm ok 1J I nit Bl al I&S, zLtAA1 AU u .12 ZALUAL &uLU &L, pages 05-54.Academie Press, Collee Park, Narylaxd, 1972.
132
[19] S. C. Eisenstat.Private comunication.
[20] I. Fried and S. K. Yang.Best finite elements distribution around a singularity.AIM JournAl 10:1244-1246, 1972.
[21] D. B. Gannon.Self Adaptive .Methods fo Parabolic Partial Differetial
Eauations.
PhD thesis, University of Illinois, 1980.
[22] A. George and 3. W. 1. Liu.Cmuter Solution of _Iau A spA fPosiltvc Definite Systems.Prentice-Hall, New Jersey, 1981.
[23] A. George.Nested dissection of a regular finite element mesh.SIAMournal _n Numerical Analysis 10:345-363, 1973.
[24] H.-I. Reinhardt.A posteriori error estimates for the finite element solution of a
singularly perturbed linear ordinary differential equation.SIAM Journal .o Numerical Analysis 18:406-430, 1981.
[25] W. C. Rheinboldt and C. K. Mesztenyi.On a data structure for adaptive finite element mesh refinements.AM Inact~iA l2 =n Mathosatcau Software 6:166-187, 1980.
[26] D. 3. Rose and 6. F. Whitten.A recursive analysis of dissection strategies.In Sipr0e Matrix Cosnutaons, pages 59-83. Academic Press, New
York, 1976.
[27] A. Schatz.An observation concerning Ritz-Galerkin methods with indefinite
bilinear forms.MlJku1 o.1ti fj Comutation 28:959-962, 1974.
[28] G. Strang and G. 3. Fix.A Aol s IL Ih FLte Elemfnt Method.Prentice-Hall, New Jersey, 1973.
[291 R. L. Taylor.On completeness of shape functions for finite element analysis.
417-atio2,l ;1 u a r u Methods t4:17-22, 1972.
133
E301 3. R. Van Rossudals.Iai Solution ol Fiait Jam1t Zonati sa Legally bLht,GridLs kX M2ti-iot.l tkod,.PhD thesis, University of Illinois. 1980.
[31] A. Weiser. S. C. Riseastst, and N. I. Sohultz.On solving elliptic equations to moderate acouracy.I" Jounal an Noazn AnalyAi 17:9)8-929. 1980.
PATE
IILMEI