Adaptive Wavelet Collocation for Plane Elasticity

Problems
Silvia Bertoluzza

and Luis Castro

March 23, 2008

Abstract
After recalling the denition and main ideas of the adaptive collocation
scheme introduced in [28, 4], we apply it to some test problems arising in two
dimensional elasticity, and compare the results obtained with classical nite
element methods.
1 Introduction
The solution of many dierential problems is very smooth in a large part of the
domain, while being globally not as smooth. In order to approximate it in an optimal
way the need arises of using high order methods, capable of taking advantage of the
local smoothness of the solution by using a coarse grid in the greatest part of the
domain, but capable, on the other hand of coping with singularities.
Among the methods that display this potential one can count the methods based
on wavelet bases and multiscale analysis. Such bases allow on one hand to design
methods of arbitrarily high order, and on the other hand they display a local behav-
ior, allowing one to take into account the presence of singularities by locally rening
the approximation space.
Even more, such bases actually provide a singularity detection tool that has been
successfully employed in designing adaptive schemes for the numerical solution of
several classes of dierential equations [22, 4, 28, 27, 8, 25].

IMATI-CNR, v. Ferrata 1, 27100 Pavia (Italy)

ICIST-IST,DECivil, Av. Rovisco Pais, 1049-001 Lisbon (Portugal)

1
The fundamental idea underlying such bases is that, in a multiscale analysis, the
space V
j
corresponding to a uniform discretization with step-size h = 2
j
, can be
obtained as the span of two dierent bases. The nodal basis of scaling functions
{
jk
} and the hierarchical basis of wavelets {
mk
, m = j
0
, . . . , j 1}. Such a basis is
obtained by decomposing V
j
as V
j
= V
j1
W
j1
= V
j
0

j1
m=j
0
W
m
, where the detail or
complement spaces W
m
are dened as W
m
= (P
m+1
P
m
)V
m+1
, P
m
: L
2
(R) V
m
being the L
2
orthogonal projection, or in the bi-orthogonal case an L
2
bounded
projection operator. The change of basis between the two, can be performed by
applying the Fast Wavelet Transform, which costs O(2
j
) operations.
It is important to underline the known fact that when expressing a function in V
j
in terms of the nodal basis, generally, no matter how smooth it is, all the coe-
cients will be needed in order to get a good approximation, while, in expressing the
same function in the wavelet basis, in order to get an approximation of the same
order, one would usually need only a subset of the coecients, (essentially the ones
corresponding to those basis functions whose center is close to singularities).
In other words, the nodal basis naturally corresponds to taking a uniform discretiza-
tion and the wavelet basis to a non uniform one.
There is however the problem of selecting a grid which is well suited to the solution of
a given problem. This is usually done through the application of an adaptive scheme.
Adaptive renement based on dierent types of a posteriori error estimations has
been studied in several papers [16, 29, 24]. In particular, wavelet based adaptive
schemes have been studied in dierent papers and with dierent approaches [22, 8,
4, 28], which have in common the fact of taking advantage in one way or another
of norm equivalence properties of the form (2.17). In the framework of Galerkin
discretizations it is known that a hierarchical decomposition of the approximate
solution provides an a posteriori error estimator ([1]). Although such a theory does
not apply to collocation schemes, one can nevertheless try to apply the same principle
also to such a case. In particular, in the framework of collocation with Deslaurier-
Dubuc interpolating multiscale decomposition, the ideas of [22] have been tested on
Burgers equation in [2].
The application of such techniques in the framework of the solution of linear elliptic
PDEs, in dimension d, has been illustrated in [4]. This paper reports the application
of the adaptive collocation techniques to the solution of plane elasticity problems.
In section 2 the fundamental relations in elasticity are presented. In section 3, the
adaptive collocation technique is described in detail. The numerical applications
are presented and discussed in section 4. Three structures are analyzed to illustrate
and characterize the application of this adaptive technique. The rst numerical test
is used to assess the accuracy and eciency of the adaptive renement procedure
2
under analysis. The second example shows the capacity of such technique to detect
and treat singularities in the stress eld. Finally, the application of the algorithm
to the analysis if non-rectangular structures is illustrated.
2 Fundamental relations
This paper illustrates the application of the adaptive collocation techniques based
on the use of interpolating wavelets in the solution of plane elasticity problems. The
eld quantities involved in the description of the behavior of the structure depend on
two coordinates only, assumed here as (x, y). The boundary conditions are imposed
along a line (x, y) = 0 lying on the same xy plane.
Two types of problems may arise in such conditions: the plain stress and plain strain
problems [26, 3]. The same eld quantities are used to describe the behavior of both
problems. The equations relating those quantities are also the same, except for the
case of the elasticity conditions [26], as described below.
Figure 1: Plane elasticity problem
The domain of the structure is denoted by , while the boundary is represented
by . The boundary is subdivided into two parts. The rst is usually known as
kinematic boundary,
u
, and it is the portion of the enveloping surface where the
value of the displacements is prescribed (Dirichlet boundary conditions). The part
of the boundary where the loads are imposed (Neumann boundary conditions) is
called static boundary,

. It is important to stress that mixed boundary conditions

may occur, as illustrated in gure 1, where
u
represents the part of the boundary
3
where the displacements are imposed in one direction and the loads are prescribed
in the other. For the sake of simplicity we will however make throughout this paper
the assumption that
u
= , though the method that we propose easily generalize
to the case where
u
= .
Given the geometry of the structure, the elastic characteristics of the material,
and the loads applied both in the domain and on the boundary, the fundamental
problem of elasticity consists in the computation of all physical quantities necessary
for the characterization of the behavior of the structure. These quantities are the
displacements, the strains and the stresses.
To completely characterize the nal location of each point of the structure, it is
necessary to dene the two independent components of the displacement vector
eld, u
x
and u
y
. The deformation of the elastic body is completely dened by the
second order strain tensor,
ij
. The distribution of the stresses in the structure is
characterized by a second order tensor,
ij
.
The compatibility conditions in the domain relate the components of the strain
tensor and the displacement vector components as follows:

xx
=
u
x
x
,
yy
=
u
y
y
,
xy
= 2
xy
=
u
x
y
+
u
y
x
(1)
The kinematic (Dirichlet) boundary conditions may be written as
u
x
= u
x
on
u
, u
y
= u
y
on
u
(2)
where u
x
and u
y
denote the known values of the displacements on that portion of
the boundary.
The elasticity conditions relate the independent components of the stress and strain
tensors. For plane stress problems, it is possible to write:

xx
=
E
1
2
(
x
+
y
) ,
yy
=
E
1
2
(
x
+
y
) ,
xy
=
E
2 (1 + )

xy
(3)
In equations (3), parameters E and dene the elastic properties of an homo-
geneous, isotropic material. They correspond to the Young modulus and to the
Poisson ratio, respectively. These elasticity relations have to be modied to allow
for the consideration of plane strain problems [26]. However, the same equations
may be used if the real elastic parameters are redened according to the following
expressions [3]:
E

=
E
1
2
,

=

1
4
The static variables (stresses and applied loads) have to satisfy, both in the do-
main and on the static boundary, the so called equilibrium conditions. In , these
conditions are expressed as follows:

xx
x
+

xy
y
+ f
x
= 0 ,

xy
x
+

yy
y
+ f
y
= 0 (4)
The static (Neumann) boundary conditions correspond to

xx
n
x
+
xy
n
y
= t
x
,
xy
n
x
+
yy
n
y
= t
y
(5)
To express the stresses in terms of the components of the displacement eld, it is
necessary to replace the compatibility relations into the elasticity conditions. It is
then possible to write

xx
= e
1
u
x
x
+ e
2
u
y
y
,
yy
= e
2
u
x
x
+ e
1
u
y
y
,
xy
= e
3
_
u
x
y
+
u
y
x
_
(6)
where the parameters e
1
, e
2
and e
3
are computed using the values specied for the
Young modulus and Poisson ratio.
e
1
=
E
1
2
, e
2
=
E
1
2
, e
3
=
E
2 (1 + )
(7)
To obtain the system of dierential equations that rules the behavior of the structure,
it is necessary to replace equations (6) into the equilibrium conditions (4). For the
sake of simplicity we will introduce the notation
A =
_

_
e
1

2
x
2
+ e
3

2
y
2
(e
2
+ e
3
)

2
x y
(e
2
+ e
3
)

2
x y
e
2

2
x
2
+ e
1

2
y
2
_

_
, f =
_
f
x
f
y
_
B =
_

_
e
1
n
x

x
+ e
3
n
y

y
e
2
n
x

y
+ e
3
n
y

x
e
3
n
x

y
+ e
2
n
y

x
e
3
n
x

x
+ e
1
n
y

y
_

_
, t =
_
t
x
t
y
_
where n
x
and n
y
denote the components of the unit outward normal associated with
the considered boundary. The plane elasticity problem can be written as follows:
5
Find u = [u
x
, u
y
]
T
such that
_

_
Au = f in
u = g on
u
Bu = t on

.
(P)
3 The Adaptive Collocation Method
3.1 Interpolating MRA and Interpolating Wavelets
This section is devoted to recalling some properties of the the multiresolution frame-
work in which the method will be stated. In the following we will denote by
s,p,
the W
s,p
() norm, and by
s,
the H
s
() norm.
Let
L
be the Daubechies compactly supported scaling function of order L, [11].
The function
L
satises, among other properties,
(i) supp
L
= [0, 2L + 1].
(ii)
L
W
R/2,
for some R > 0 (R is proportional to L): |(d
s
/dx
s
)
L
| C, for
all integer s 0 s R/2;
(iii)
L
satises a two-scale relation (the renement equation) which in the
Fourier domain takes the form

L
() = m
0
(/2)
L
(/2),
(m
0
is a trigonometric polynomial of degree 2L + 1);
(iv)
L
is orthogonal to its integer translates:
_

L
(x)
L
(x k) dx =
0k
(v) Polynomials up to order L can be written as a linear combination of the integer
translates of
L
.
The Deslaurier-Dubuc fundamental function of order N = 2L + 1, is dened as
the autocorrelation of
L
:
(x) =
_
R

L
(y)
L
(y x) dy.
As a consequence of (i-v) the function satises the following properties:
6
1) supp = [N, N], and W
R,
;
2) the function clearly satises itself a two scale relation. In fact it satises

= |

L
|
2
, and hence (2.1) gives

() = |m
0
(/2)|
2

(/2).
3) As a direct consequence of the orthogonality of the translates of
L
the function
satises the following interpolation property:
(n) =
_
R

L
(y)
L
(y n) dy =
n0
.
4) As a consequence of (v), polynomials up to order N can be written as linear
combinations of the integer translates of .
Remark 3.1. The function was originally introduced by G. Deslaurier and S.
Dubuc [12] as the limit function of an interpolatory subdivision scheme. Its relation
with the minimal phase Daubechies orthonormal scaling functions was pointed out
by G. Beylkin and N. Saito ([7]).
Based on such a function we can easily build multiscale decompositions on R and,
by tensor product, on R
d
.
In order to construct multiscale decomposition on the interval ]0, 1[ for j j
0
=
[log
2
(N/2)] + 1 (where [] stands for the integer part) we dene the interpolating
basis functions on the interval by applying to the Deslaurier Dubuc functions the
ideas of [9, 13]. More precisely we let

jk
= (2
j
x k) +
1

n=N+1
a
nk
(2
j
x n), k = 0, .., L (8)

jk
= (2
j
x k), k = L + 1, ..., 2
j
L 1, (9)

jk
= (2
j
x k) +
2
j
+N1

n=2
j
+1
b
nk
(2
j
x n), k = 2
j
L, .., 2
j
, (10)
with extrapolation weights a
nk
and b
nk
dened by:
a
nk
= l
1
jk
(n2
j
), b
nk
= l
2
jk
(n2
j
),
7
where l
1
jk
and l
2
jk
denote the Lagrange interpolation polynomials of degree L dened
respectively as
l
1
jk
=
L

i=0
i=k
x i2
j
k2
j
i2
j
, l
2
jk
=
2
j

i=2
j
L
i=k
x i2
j
k2
j
i2
j
.
Remark that a
nk
and b
nk
are indeed independent of j.
A multiresolution {V
j
}
jj
0
of ]0, 1[ is then dened as
V
j
= span <
jk
, k = 0, .., 2
j
> L
2
(0, 1).
By tensor product we the easily dene a multiresolution on the square: introducing
the two dimensional scaling functions
j,k
, k = (k
1
, k
2
) G
j
= {0, .., 2
j
}
2
dened
as

j,k
=
jk
1

jk
2
we set
V
j
= span <
j,k
, k = (k
1
, k
2
) G
j
>
It is immediate to dene an interpolation operator L
j
: C
0
([0, 1]
2
) V
j
L
j
f =

kG
j
f(k/2
j
)
j,k
.
The wavelet basis for the complement space
W
j
= (L
j+1
L
j
)V
j+1
is composed by the functions

(1,0)
j,k
=
j+1,2k
1
1

j,2k
2
(11)

(0,1)
j,k
=
j,2k
1

j+1,2k
2
1
(12)

(1,1)
j,k
=
j+1,2k
1
1

j+1,2k
2
1
(13)
The grid-points corresponding to the scaling functions and the wavelets respectively
will be indicated as

j,k
= (k
1
2
j
, k
2
2
j
),
(1,0)
j,k
= ((2k
1
1)2
(j+1)
, k
2
2
j
),

(0,1)
j,k
= (k
1
2
j
, (2k
2
1)2
(j+1)
),
(1,1)
j,k
= ((2k
1
1)2
(j+1)
, (2k
2
1)2
(j+1)
).
8
In the following it will be practical to introduce the following compact notation on
the indexes: given = (, j, k) with = {0, 1}
2
\{0, 0}, j j
0
, and k such that

j,k
[0, 1]
2
, we set

j,k
,

j,k
.
Each continuous function f C
0
([0, 1]
2
) can be expanded as
f =

k{0,..,2
j
0}
2

j
0
k

j
0
k
+

,
where
= {(, j, k), , j j
0
, k such that

j,k
[0, 1]
2
}
denotes the set of compact indexes. For j j
0
xed, we set

j
= {(, j, k), , k such that

j,k
[0, 1]
2
}
so that =
jj
0

j
.
Remark that if a dyadic point p = (n)/2
m
[0, 1]
2
veries p {
j
0
,k
, k G
j
0
},
then there exists a unique (p) such that
p =
n
2
m
(p) : p =
(p)
. (14)
The spaces V
j
satisfy a Strang-Fix condition of order N, which implies that the
following approximation property holds: if f H
s
(]0, 1[
2
), 1 < s N + 1 then
r s, 0 r R we have
f L
j
f
r,]0,1[
2 C2
j(sr)
f
s,]0,1[
2.
Such a multiscale decomposition has been extensively investigated in [13], where
several properties which will play an important role in the following have been
stated. In particular we recall that it is possible to dene an interpolating wavelet
transform mapping any continuous function f into the sequence of coecients
w(f) = {{
j
0
,n
, n G
j
0
}, {

, }},
dened as

j
0
,n
= f(
j
0
,n
), and for
j
,

= f(

) L
j
f(

). (15)
9
Formula (15) clearly shows that the wavelet coecients
jk
measure the lack of
approximation of f by L
j
f at the point

. In fact, a stronger result holds. It

is possible to prove ([13, 14]) that the result of characterization of Besov spaces
that holds for orthonormal wavelets, carries on to interpolating wavelets on (0, 1),
provided that the Besov space considered is embedded in C
0
(0, 1). In particular in
dimension two this holds for the Sobolev spaces H
s
(0, 1) provided s > 1. In other
words, the size of the coecients

gives us information on eventual singularities

in all the derivatives of a function f up to a certain order. This is a key point in
the behavior of our method.
As in the usual framework of biorthogonal wavelets we have a fast wavelet transform
algorithm, acting on functions in V
j
. Given the coordinates of a function f V
j
with respect to the scaling functions basis {
j,n
, n G
j
} the FWT provides the
coordinates of f with respect to the wavelet basis {
j
0
,n
, n G
j
0
}
j1
m=j
0
{

j
}. However, unlike in the usual biorthogonal case, the coordinates with respect
to the scaling function basis are nothing but the values of f at the dyadic points.
Moreover, the matrix realizing such a change of basis turns out to be lower triangular.
This plays an important role in the ecient implementation of the method [4, 6]
3.2 Collocation on Non-Uniform Grids
Let us now briey recall how the above mentioned structure can be employed to
design a collocation method on non uniform dyadic grids. For any nite subset

h
of the index set , we dene a corresponding nite grid of dyadic points
G

h
as
G

h
:= {
j
0
,k
, k G
j
0
} {

,
h
}.
(in order to have a minimum degree of approximation the grid G
h
will be assumed
to contain the coarse grid of mesh size 2
j
0
), and we will denote by
V

h
= V
j
0
span <

,
h
>
the space spanned by the basis functions corresponding to the selected points.
It will also be convenient in the following to introduce the corresponding interpo-
lation operator L

h
: C
0
() V

h
. Given f C
0
(), L

h
f V

h
is the unique
function satisfying
L

h
f(p) = f(p) for all nodes p G
h
= G
j
0

h
.
10
By abuse of notation, for u = (u
x
, u
y
) C
0
()
2
we will denote by L

h
u the vector
(L

h
u
x
, L

h
u
y
) V
2

h
.
In order to take into account boundary conditions the grid G

h
is split as the union of
the set of interior nodes and the set of respectively Neumann and Dirichlet boundary
nodes:
G

h
= G
(i)

h
G
(N)

h
G
(D)

h
(4.5)
with
G
(i)

h
= G

h
]0, 1[
2
, G
(N)

h
= G

, G
(D)

h
= G

h

u
. (4.6)
Problem (P) can then be discretized as follows:
Given
h
, nd u V

h
such that
_

_
Au
h
(p) = f(p) for all nodes p G
(i)

h
u
h
(p) = g(p) for all nodes p G
(D)

h
Bu
h
(p) = t(p) for all nodes p G
(N)
h
(P
h
)
3.3 The adaptive collocation scheme
We can now briey describe the adaptive collocation scheme proposed in [4]. The
idea is to iteratively compute increasingly good approximations to the solution u of
Problem 1. and use the actual approximate solution at each iteration in order to
design a better grid to be used for computing the next approximate solution.
We select the rst grid G
0
h
= G

0
by requiring that the right hand side f is well
approximated on such grid, in the sense that
f L

0
f
t,
. (5.1)
Once the rst grid G
0
has been selected, we compute a rst approximate solution u
0
by applying a collocation method in the corresponding space U
0
= V

0
, according
to the formulation (P
h
).
Following the idea of [22, 28, 6, 4], we then analyze the computed solution in order
to design a new (better) grid. More precisely, at the nth step, given a grid G
n
,
11
and the relative approximate solution u
n
with coecients
w(u
n
x
) = {{c
x,n
j
0
,k
}, {e
x,n

,
n
}}, w(u
n
y
) = {{c
y,n
j
0
,k
}, {e
y,n

,
n
}}
we compute the next grid G
n+1
by removing the useless points and rening where
the approximation is bad. For each point

, = (, j, k), , j j
0
, k
{1, , 2
j
}
2
, dene a set U

of neighboring points
U

= [0, 1]
2
{
j+1
k
, k = (2k
1
+
1
, . . . , 2k
d
+
d
),
i
= 1, 0, 1},
and the corresponding index set U

as
U

= {(p), p U

}
(where, for a dyadic point p, (p) is dened by (14)). Dene then a set
sel
n
of
indexes selected for renement as

sel
n
= { : max{|e
x,n

|, |e
y,n

| >
a
/2
js
}.
We then introduce an updated index set

n+1
= { : max{|e
x,n

|, |e
y,n

|} >
r
/(2
js
j
2
)}

sel
n
U

,
and the corresponding rened grid
G
n+1
= G

n+1
.
The (n + 1) th approximation space is dened as
U
n+1
= V

n+1
.
The (n+1)th approximate solution is computed by solving the collocation Problem
B. on the space U
n+1
, with collocation grid G
n+1
.
We want to remark that one of the strength of the proposed rening procedure is its
extreme simplicity. In particular it does not need extra computation of any quan-
tities, the coecient of the approximate solution themselves being the parameters
driving the decision on whether to rene or not. Moreover, since the discretization
is not based on any triangulation, adding or removing each point has no inuence on
the other points of the grid, since there is no conformity requirement to fulll, which
causes some diculty when rening and derening a nite element triangulation.
12
4 Numerical applications
The analysis of three structures is used to illustrate the use of the interpolating
wavelets in the solution of plane elasticity problems. We always assume a physically
and geometrically linear behavior.
First, a square cantilever plate subjected to a uniform load acting on its upper edge
is used to illustrate the use of uniform grids. The solutions obtained are compared
with the ones resulting from the application of other numerical techniques. The
same structure is then used to show how ecient the use of the adaptive procedure
described in the previous sections can be.
The analysis of a rectangular plate with a central crack will show that the techniques
being tested are quite eective in what concerns the detection and treatment of
singularities in the stress eld. It shows also that the use of non-uniform grids
and the adaptive renement procedures lead to accurate numerical solutions with a
relatively small number of degrees of freedom.
The analysis of a simple gravity dam is used to illustrate the solution of problems
dened over non-rectangular domains.
4.1 Analysis of a cantilever plate with uniform grids
Three dierent discretizations are used in the analysis of the square cantilever plate
presented in gure 2. In each of them, a uniform grid is considered with N = 4
and j
0
= 3. The dierent discretizations take increasing values for parameter j
max
.
Table 1 lists the characteristics of all discretizations, including the number of grid
points in the mesh, n
grid
, the total number of degrees of freedom, n
dof
and the CPU
time (in seconds), T
CPU
, involved in the analysis. The examples presented in this
paper were run on machine with an AMD Atlhon processor running Linux.
Discretization N j
0
j
max
n
grid
n
dof
T
CPU
I 4 3 3 81 162 0.04
II 4 3 4 289 578 0.72
III 4 3 5 1089 2178 52.63
Table 1: Discretizations involved in the analysis of the square cantilever
Figure 3 shows the solution obtained with discretization I. The three stress and
13
Figure 2: Square cantilever plate
the displacement elds are plotted using the graphic routine package Janela [21].
For the representation of the stress elds, a ner uniform grid has been used. Each
independent component of the stress tensor is then plotted in a grid of 6565 dyadic
points in the domain. It is important to remind clearly that this ne grid is used
only to support the graphical representation of the solution; it is never used in the
computations.
The solution represented in gure 4 was obtained using a commercial nite element
code, COSMOS/M [10]. To obtain that solution, a mesh with 900 4-node bilinear
elements was used, yielding a total of 1860 degrees of freedom. The stress eld
is represented as it results directly from the computations without the use of any
smoothing techniques. Comparing gures 3 and 4 it is possible to observe qualita-
tively that the use of interpolating wavelets leads to accurate solutions with a small
number of degrees of freedom when compared to the classical displacement nite
element model.
The accuracy of the solutions obtained is illustrated in tables 2 and 3. Table 2 shows
the values computed for the components of the displacement eld at point D (see
gure 2).
I II III Mix F Mix L FE
d
x
-1.114 -1.057 -1.044 -1.036 -1.041 -1.039
d
y
-2.930 -2.941 -2.963 -2.981 -2.988 -2.979
Table 2: Components of displacement eld at point D.
14

xx

yy

xy
q
V
3.000E+00
-3.000E+00
0.000E+00
8.000E-01
-1.000E+00
0.000E+00
0.000E+00
-1.300E+00
= 2.50 p a / E
Figure 3: Square cantilever plate; solution with discretisation I

xx

yy

xy
Desl. Dominio
= 2.50 pa/E
3.000E+00
-3.000E+00
0.000E+00
8.000E-01
-1.000E+00
0.000E+00
0.000E+00
-1.300E+00
Figure 4: Square cantilever plate; solution with nite elements
15
The values presented in table 2 are normalized according to:
d
x
=
E
p a
d
x
; d
y
=
E
p a
d
y
The same table lists also the values obtained with two non-conventional hybrid-
mixed stress nite element formulations [19, 17, 18]. The Mix F and the Mix L
models use as approximation bases the trigonometric functions [20] and the or-
thonormal Legendre polynomials [23] and the solutions presented in table 2 are
obtained using 1544 and 1664 degrees of freedom, respectively.
Table 3 shows, for all test cases presented above, the values of the independent
components of the stress tensor on points A, B and C represented in gure 2. All
values are normalized according to the rule:

ij
=

ij
p
P I II III Mix L FE I
c
II
c
III
c

xx
2.402 3.463 4.726 9.559 7.623 8.387 10.588 13.180
A
yy
-1.000 -1.000 -1.000 -1.000 2.287 2.516 3.176 3.954

xy
0.000 0.000 0.000 0.000 -1.656 -2.527 -3.009 -3.571

xx
-2.708 -3.295 0.029 -6.237 -5.281 -5.934 -6.913 -8.088
B
yy
0.000 0.000 0.000 0.000 -1.584 -1.780 -2.074 -2.426

xy
0.000 0.000 0.000 0.000 -0.814 -1.357 -1.524 -1.753

xx
0.026 0.029 0.031 0.028 0.031 0.027 0.029 0.031
C
yy
-0.473 -0.472 -0.473 -0.489 -0.473 -0.473 -0.472 -0.473

xy
-0.670 -0.716 -0.736 -0.721 -0.747 -0.670 -0.716 -0.736
Table 3: Stresses on points A, B and C
It is important to discuss the values obtained for the stresses on points A and B. The
comparison of the results should not involve those points because it is known [3] that
singularities will occur precisely there. However, when comparing the values listed
in table 3, it is possible to verify that the solutions obtained with the interpolating
wavelets lead to smaller stresses in the corners. This is due to the fact that on points
A and B the Neumann boundary conditions are directly enforced in the collocation
procedure, in order to fully satisfy the equilibrium along the boundary. In structural
mechanics, it is always important to obtain equilibrated solutions. If the collocation
scheme is changed in order to impose the Dirichlet boundary conditions on points A
and B, the resulting stresses are closer to the nite element solutions, as illustrated
in three last columns of table 3 and the eect induced by the singularities is quite
clear.
16

xx

yy

xy
q
V
3.000E+00
-3.000E+00
0.000E+00
8.000E-01
-1.000E+00
0.000E+00
0.000E+00
-1.300E+00
= 2.50 p a / E
Figure 5: Square cantilever plate; compatible solution with discretization I
It is important to verify that the dierences between the compatible and equilibrated
solutions are concentrated around the corner points, as illustrated in gure 5. As
expected, the dierences tend to disappear if the degree of resolution is increased, as
illustrated in gure 6 where, for uniform meshes with j
max
= 5, the applied stresses
on the boundary are plotted for both compatible and equilibrated solutions.
4.2 Analysis of the cantilever plate with non-uniform grids
In the previous section only uniform grids were considered in the analysis. However,
this is not by far the most eective procedure to obtain an accurate solution for the
problems being solved. To fully exploit the advantages of the collocation technique
under analysis, a sequence of adaptive non-uniform grids must be used.
To illustrate the application of an adaptive renement procedure, lets take again
the square cantilever plate shown in gure 2. Lets consider an initial uniform grid
with j
0
= 3, j
max
= 4 and N = 4. This initial mesh is the one represented at the
top left corner of gure 7.
Then, the adaptive renement procedure is applied using the following values for the
parameters that control remeshing: t
a
= 1 10
5
and t
r
= 1 10
6
. The sequence
of non-uniform grids generated by the algorithm is presented also in gure 7. The
nal mesh is represented in its lower right corner. As expected, the grid points are
17
0 0.5 1 1.5 2 2.5 3 3.5 4
8
6
4
2
0
2
4
6
A
B C
D A
Diagram tx
"Equilibrated" solution
"Compatible" solution
0 0.5 1 1.5 2 2.5 3 3.5 4
1.5
1
0.5
0
0.5
1
1.5
2
2.5
3
A B C D
A
"Compatible" solution
"Equilibrated" solution
Diagram ty
Figure 6: Applied stress eld on the boundary
concentrated mostly near the corners of the clamped edge, where the singularities
of the stress eld occurs.
Mesh j
max
n
grid
n
dof
initial 4 289 578
2 5 597 1194
3 6 868 1736
nal 7 917 1834
Table 4: Adaptive non-uniform grids used in the solution of the square plate
Table 4 shows, for each grid, the number of points, n
grid
, the number of degrees
of freedom, n
dof
, and the highest level of resolution, j
max
. The nal solution is
computed in 12.50 seconds (except post-processing operations) and is represented
in gure 8. It must be stressed that the accuracy of this solution is better than the
accuracy of the uniform mesh with j
max
= 5 and it is much more economic in terms
of resources required for computation, both in terms of memory and CPU time.
It is interesting to analyze the evolution of the strain energy with the number of
degrees considered in the discretization. This information is presented in the graph
included in gure 9. It shows the evolution of the strain energy error for both uniform
and adaptive non-uniform meshes. For the computation of the strain energy error,
it was assumed that its exact value is given by U
exact
= 0.9865 [23].
18
0
0.2
0.4
0.6
0.8
1
0 0.2 0.4 0.6 0.8 1
"[Link]"
0
0.2
0.4
0.6
0.8
1
0 0.2 0.4 0.6 0.8 1
"[Link]"
0
0.2
0.4
0.6
0.8
1
0 0.2 0.4 0.6 0.8 1
"[Link]"
0
0.2
0.4
0.6
0.8
1
0 0.2 0.4 0.6 0.8 1
"[Link]"
Figure 7: Initial and nal grids

xx

yy

xy
q
V
3.000E+00
-3.000E+00
0.000E+00
8.000E-01
-1.000E+00
0.000E+00
0.000E+00
-1.300E+00
= 2.50 p a / E
Figure 8: Square cantilever plate; solution obtained with an adaptive renement
procedure
19
10
2
10
3
10
4
10
3
10
2
10
1
Adaptive refinement
Uniform grid refinement
Number of degrees of freedom
STRAIN ENERGY ERROR
error
Figure 9: Strain energy evolution for the uniform and non-uniform grids
4.3 Analysis of a rectangular plate with a central crack
The rectangular plate with a central crack shown in gure 10 is a typical example
where the use of an higher order scheme is highly recommended due to the existence
of a strong singularity at the tip of the crack. Due to the symmetry conditions, only
1/4 of the plate is analyzed.
The initial uniform grid is characterized by the use of j
0
= 3, j
max
= 4 and N = 4.
In the adaptive procedure, the adopted values for the remeshing parameters are the
following: t
a
= 1 10
5
and t
r
= 1 10
6
.
The nal solution is presented in gure 11. It is quite clear the accuracy of the
solution obtained and the capacity of the model to capture the singularity near the
crack tip. The nal grid is presented also in the same gure and involves a total of
1058 degrees of freedom.
Figure 12 shows the stresses applied on the boundary. It is possible to verify that
the equilibrium conditions are veried on the static boundary.
The sequence of non-uniform grids generated by the algorithm is presented in gure
13, where the
yy
stress eld distribution obtained for each case is also plotted. The
number of grid points, the total number of degrees of freedom and the maximum
level of resolution are given in table 5.
As explained in [4], the non-uniform renement of the mesh depends on the values
adopted for tolerances t
a
and t
r
. Dierent values for these parameters will lead to
dierent non-uniform grids. This aspect is illustrated in gure 14, where the nal
20
= 0.3
a/3 a/6
a
p
p
p
Figure 10: Rectangular plate with a central crack
0
0.5
1

xx

0
.
0
3
2
7
7
4

0
.
0
3
2
7
7
4
0.075574
0.18392
0
0.5
1
1.5

yy
0.3601
0
.
3
6
0
1
0
.
4
6
8
4
4
0
.
7
9
3
4
5
0.4
0.3
0.2
0.1
0

xy

0
.0
2
4
3
7
6

0
.0
8
8
9
1
5
0.12118

0
.1
5
3
4
5

0
.
2
5
0
2
6
Final grid
Figure 11: Square plate with central crack; nal solution
21
0 50 100 150 200 250 300 350
0.4
0.3
0.2
0.1
0
0 50 100 150 200 250 300 350
1.5
1
0.5
0
0.5
1
A
B C D
E
F
A C
D
F
A
A
tx diagram
ty diagram
Figure 12: Square plate with central crack; stresses on the boundary
Mesh j
max
n
grid
n
dof
initial 4 289 578
2 5 401 802
3 6 470 940
nal 7 529 1058
Table 5: Adaptive non-uniform grids used in the solution of the plate with a central
crack
22
0 20 40 60
0
20
40
60
80
100
120

yy
0
.
1
9
9
9
6
0
.
3
0
6
9
6
0.73496
0.19996
ndof=578
0 20 40 60
0
20
40
60
80
100
120

yy
0
.
3
3
6
7
5
0
.
4
7
0
4
2
0
.2
0
3
0
9
0
.0
6
9
4
1
9
0
.8
7
1
4
2
1
.
0
0
5
1
0
.3
3
6
7
5
0
.2
0
3
0
9
0
.4
7
0
4
2
ndof=802
0 20 40 60
0
20
40
60
80
100
120

yy
0
.
2
8
8
3
3
0
.1
6
8
1
8
0
.0
4
8
0
1
8
0
.8
8
9
1
2
0
.6
4
8
8
1
0
.4
0
8
4
9
0
.2
8
8
3
3
0.048018
0.40849
0
.5
2
8
6
5
ndof=940
0 20 40 60
0
20
40
60
80
100
120

yy
0
.9
0
7
8
2
0
.6
8
9
8
0
.0
3
5
7
6
4
0
.
2
5
3
7
8
0
.4
7
1
7
9
0
.
3
6
2
7
8
0
.2
5
3
7
8
0
.3
6
2
7
8
0
.4
7
1
7
9
ndof= 1058
Figure 13: Stress diagrams and collocation points localization
23
meshes obtained for dierent values of tolerances t
a
and t
r
are plotted.
Uniform grid Adaptive grid ta=10
5
tr=10
6
Adaptive grid ta=10
4
tr=10
6
Adaptive grid ta=10
4
tr=10
5
Adaptive grid ta=5x10
5
tr=10
5
Adaptive grid ta=5x10
4
tr=10
4
Figure 14: Square plate with central crack; inuence of parameters t
a
and t
r
It is interesting to investigate now what are the parameters that lead to the most
eective adaptive renement in the analysis of this structure. For this purpose,
gure 15 shows the evolution of the strain energy error for the sequence of meshes
generated during the adaptive procedure associated with each tested pair of toler-
ances t
a
and t
r
. From the analysis of this gure, it is possible to conclude that
the best choice corresponds to t
a
= 5 10
4
and t
a
= 1 10
4
. The same gure
shows also how ecient the adaptive procedures may be when compared to the use
of uniform mesh renements.
In the plots represented in gure 15, the exact value for the strain energy is consid-
ered to be equal to U = 0.0458402901388889 (p a)
2
/E. This value is given in [32]
and is obtained by applying a dual extrapolation technique to two series of statically
and kinematically admissible solutions, which provide upper and lower bounds for
the exact solution.
24
400 600 800 1000 1200 1400 1600 1800 2000 2200 2400
10
4
10
3
10
2
10
1
10
0
10
1
Uniform grid
ta=10
5
; tr=10
6
ta=10
4
; tr=10
6
ta=10
4
; tr=10
5
ta=5x10
5
; tr=10
5
ta=5x10
4
; tr=10
4
Number of degrees of freedom
Energy error
Figure 15: Square plate with central crack; energy error
4.4 Analysis of a gravity dam
The gravity dam presented in gure 16 is used to illustrate the application of the
adaptive collocation technique to the analysis of more general structures. As the
computational domain is not rectangular, it must be mapped onto a unit square by
means of a conformal mapping of class C
2
with a C
2
inverse, as explained in [4].
Figure 16: Gravity dam
Figure 17 shows the solution obtained with the nite element package ADINA [15].
In this computation a 100 eight-node rectangular mesh with a total of 640 degrees
of freedom was used.
25
TIME 1.000
X Y
Z
TIME 1.000
X Y
Z
STRESS-YY
RST CALC
TIME 1.000
1.33
0.00
-1.33
-2.67
-4.00
-5.33
-6.67
-8.00
-9.33
-10.67
-12.00
-13.33
-14.67
-16.00
-17.33
MAXIMUM
0.9847
EG 1, EL 95, IPT 31 (0.7250)
MINIMUM
-18.21
EG 1, EL 11, IPT 33 (-16.96)
TIME 1.000
X Y
Z
STRESS-ZZ
RST CALC
TIME 1.000
3.60
2.40
1.20
0.00
-1.20
-2.40
-3.60
-4.80
-6.00
-7.20
-8.40
-9.60
-10.80
-12.00
-13.20
MAXIMUM
3.091
EG 1, EL 1, IPT 33 (0.1324)
MINIMUM
-12.47
EG 1, EL 1, IPT 13 (-10.91)
TIME 1.000
X Y
Z
STRESS-YZ
RST CALC
TIME 1.000
28.00
26.00
24.00
22.00
20.00
18.00
16.00
14.00
12.00
10.00
8.00
6.00
4.00
2.00
0.00
MAXIMUM
27.88
EG 1, EL 1, IPT 33 (25.09)
MINIMUM
-0.3896
EG 1, EL 100, IPT 13 (-0.3283)
Figure 17: Solution with ADINA
26
Lets consider an initial uniform grid with j
0
= 3, j
max
= 4 and N = 4. This initial
mesh is the one represented at the top left corner of gure 18. The same gure shows
the sequence of non-uniform adaptive grids obtained. Table 6 shows the number of
grid points and the total number of degrees of freedom associated with each mesh.
To control the renement, the two tolerance parameters are given by t
a
= 5 10
4
and t
a
= 1 10
4
.
0
0.5
1
1.5
2
2.5
3
0 1 2 3 4 5
"Mesh 1 (289 nodes)"
0
0.5
1
1.5
2
2.5
3
0 1 2 3 4 5
"Mesh 2 (565 nodes)"
0
0.5
1
1.5
2
2.5
3
0 1 2 3 4 5
"Mesh 3 (714 nodes)"
0
0.5
1
1.5
2
2.5
3
0 1 2 3 4 5
"Mesh 4 (794 nodes)"
Figure 18: Non-uniform meshes
Mesh j
max
n
grid
n
dof
initial 4 289 578
2 5 565 1130
3 6 714 1428
nal 7 794 1588
Table 6: Adaptive non-uniform grids used in the solution of the gravity dam
The nal solution is presented in gure 19.
27

xx

yy

xy
q
V
0.000E+00
-1.800E+01
3.000E+00
-1.300E+01
0.000E+00
2.800E+01
0.000E+00
= 0.3333E+02 p a / E
Figure 19: Final non-uniform mesh in the analysis of gravity dam with a triangular
load
5 Conclusions and further developments
The examples presented in the previous section show clearly that the use of adaptive
collocation techniques based on the use of interpolating wavelets is an attractive tool
for the numerical analysis of plane elasticity problems.
In order to be able to deal with more realistic structural engineering problems,
several aspects of the implementation must however be improved. For this purpose
it is necessary:
To implement a scheme that allows for the sub-division of the structure in
sub-domains in order make possible the analysis of more complex geometries;
To develop a suitable preconditioner for the iterative solution of the governing
system that may be able to exploit the characteristics of the basis being used;
To implement ecient IWT algorithms for the post-processing operations.
To exploit the advantages of the numerical schemes being discussed, it will be con-
venient to generalize the application of these techniques to other types of structural
28
problems. The implementation of a similar adaptive collocation scheme for the so-
lution of Reissner-Mindlin plate bending problems, which is also a second order
problem, is currently under development. Also, the development of a collocation
technique for the solution of thin Kirchho plate bending problems (a fourth order
problem), is in program.
Finally, it will be necessary to include in the algorithm the possibility of taking into
account the physical non-linearities presented in the representation of the material
behavior. It is then necessary to use the Plasticity Theory for ductile materials
(such as metals) or the Continuum Damage Theory for quasi-brittle materials (such
as concrete).
Acknowledgements
This work has been supported by the EC through research programme TMR-
ERBFMRXCT-98-0184, Wavelets in Numerical Analysis and Simulation. It was
partially supported also by ICIST (of which Luis Castro is member) and by PRAXIS
XXI as part of research project 2/2.1/CEG/33/94.
References
[1] Bank, R.E., and Smith, R.K., Some a posteriori error estimates for elliptic
partial dierential equations. Comput. Meth. Appl. Mech. Engrg. 61:283301,
1987.
[2] Bertoluzza, S., Adaptive wavelet collocation method for the solution of Burgers
equation, Trans. Th. Stat. Phys. 25, 1996.
[3] J. R. Barber, Elasticity, Solid Mechanics and Its Applications, Kluwer Aca-
demic Publishers, Dordrecht, (1992);
[4] S. Bertoluzza, An Adaptive Collocation Method based on Interpolating Wavelets,
in Multiscale Wavelet Methods for Partial Dierential Equations, edited by
Dahmen, Kurdila and Oswald, Academic Press, 1997;
[5] S. Bertoluzza, J.C. Ravel, and Y. Maday. A dynamically adaptive wavelet
method for solving pertial dierential equations. Comput. Methods Appl. Mech.
Engrg., 1994.
29
[6] S. Bertoluzza and P. Pietra, Adaptive Wavelet Collocation for nonlinear BVPs,
Lecture Notes in Control and Information Sciences, Springer Verlag, London
(1996).
[7] G. Beylkin and N. Saito, Multiresolution representations using the auto-
correlation functions of compactly supported wavelets, IEEE Trans. Signal Pro-
cessing, 41:3584-3590, 1993.
[8] A. Cohen, W. Dahmen, and R. De Vore. Adaptive wavelet methods for elliptic
operator equations convergence rates. Technical report, IGPM - RWTH-
Aachen, 1998. to appear in Math. Comp.
[9] A. Cohen, I. Daubechies, and P. Vial. Wavelets on the interval and fast wavelet
transforms. ACHA, 1:5481, 1993.
[10] COSMOS/M, Finite Element System, Version 1.70, Structural Research and
Analysis Corporation, (1993);
[11] I. Daubechies, Orthonormal Bases of Compactly Supported Wavelets, Commu-
nications on Pure and Applied Mathematics, vol. 41, pp. 909-996, (1988);
[12] Interpolation Dyadique. In Fractals, dimensions non entieres et applications.
Masson, Paris, 1987, pp. 4455.
[13] Donoho, D., Interpolating Wavelet Transform, Department of Statistics, Stan-
ford University, 1992, preprint.
[14] Dahmen, W. Stability of Multiscale Transformation, Jour. of Fourier Anal. and
Appl.
[15] ADINA, Automatic Dynamic Incremental Nonlinear Analysis, Theory and
Modeling Guide, Volume 1: ADINA, Report ARD 01-7, Adina R & D, Inc,
Watertown, USA.
[16] Erikson, K., and Johnson C., Adaptive Streamline Diusion Finite Element
Method for stationary Convection-Diusion Problems, Math. of Comp. 60:167
188, 1991.
[17] J.A. Teixeira de Freitas, J.P. Moitinho de Almeida and E.M.B. Ribeiro Pereira,
Non-conventional formulations for the nite element method, Structural Engi-
neering and Mechanics, 4, pp. 655-678, 1996.
[18] J.A. Teixeira de Freitas, J.P. Moitinho de Almeida and E.M.B. Ribeiro Pereira,
Non-conventional formulations for the nite element method, in Modelling and
Simulation Based Engineering, pp. 254-259, edited by S.N. Atlury and P.E.
ODonoghue, Palmdale, 1998.
30
[19] J. A. T. Freitas, Duality and Symmetry in Mixed Integral Methods of Elasto-
statics, International Journal for Numerical Methods in Engineering, Vol. 28,
pp. 1161-1179, (1989);
[20] J. A. T. Freitas e E. M. B. R. Pereira, O Metodo dos Integrais de Fronteira Us-
ando Series de Fourier, MECOM89, X Congresso Ibero-Latino-Americano so-
bre Metodos Computacionais em Engenharia, Vol. 1, A221-A230, Porto, (1989);
[21] J. P. B. Moitinho de Almeida, Janela: Uma interface graca destinada `a
aplicacao em problemas de Mecanica Computacional, Internal Report, Insti-
tuto Superior Tecnico, Technical University of Lisbon, (1992);
[22] Y. Maday, V. Perrier, and J.C. Ravel. Adaptivite dynamique sur bases
dondelettes pour lapproximation dequations aux derivees partielles. C. R.
Acad. Sci Paris, 312, (Serie I):405410, 1991.
[23] E. M. B. R. Pereira and J. A. T. Freitas, A Mixed-Hybrid Finite Element Model
Based on Orthogonal Functions, International Journal for Numerical Methods
in Engineering, Vol. 39, pp. 1295-1312, (1996);
[24] Rannacher, R., The dual-weighted-residual method for error control and mesh
adaptation in nite element methods. The mathematics of nite elements and
applications, X, MAFELAP 1999 (Uxbridge), 97116, Elsevier, Oxford, 2000.
[25] Koster, F., Griebel, M., Kevlahan, N. K.-R., Farge, M., and Schneider, K. To-
wards an adaptive wavelet-based 3D Navier-Stokes solver. Numerical ow sim-
ulation, I (Marseille, 1997), 339364, Notes Numer. Fluid Mech., 66, Vieweg,
Braunschweig, 1998.
[26] S. P. Timoshenko and J. N. Goodier, Theory of Elasticity, 3
rd
edition, McGraw-
Hill International Book Company, Tokyo, (1982);
[27] Vasilyev, O.V. and Bowman, C., Second Generation Wavelet Collocation
Method for the Solution of Partial Dierential Equations. J. Comp. Phys.,
165:660-693, 2000.
[28] Vasilyev, O.V., and Paolucci, S., A Dynamically Adaptive Multilevel Wavelet
Collocation Method for Solving Partial Dierential Equations in a Finite Do-
main. J. Comp. Phys., 125:498-512, 1996.
[29] Verf urt, R., A Posteriori Error Estimation and adaptive Mesh Renement Tech-
niques, J. Comput. Appl. Math. 50:6783, 1994.
[30] J.L. Walsh, A closed set of orthogonal functions, Ann J Math, 55, pp. 5-24,
1923.
31
[31] O. C. Zienkiewicz e R. L. Taylor, The Finite Element Method, Volume 1 - Basic
Formulation and Linear Problems, 4
rd
edition, McGraw-Hill Book Company,
Berkshire, (1989);
[32] J. A. Teixeira de Freitas and J. P. Moitinho de Almeida and E. M. B. Ribeiro
Pereira, Non-conventional formulations for the nite element method, Compu-
tational Mechanics, 23:488-501, 1999.
32