37? HSlJ Ato. A9?o/67531/metadc331296/m2/1/high_re… · an appropriate Sobolev space H®(fl) or...
Transcript of 37? HSlJ Ato. A9?o/67531/metadc331296/m2/1/high_re… · an appropriate Sobolev space H®(fl) or...
3 7 ? HSlJ
Ato. A9?o
MINIMIZATION OF A NONLINEAR ELASTICITY
FUNCTIONAL USING STEEPEST DESCENT
DISSERTATION
Presented to the Graduate Council of the
University of North Texas in Partial
Fulfillment of the Requirements
For the Degree of
DOCTOR OF PHILOSOPHY
By
Terence W. McCabe, B.S., M.A.
Denton, Texas
August, 1988
•to/I
McCabe, Terence W., Minimization of a Nonlinear
Elasticity Functional Using Steepest Descent. Doctor of
Philosophy (Mathematics), August, 1988, 76 pp.,
bibliography, 21 titles.
The method of steepest descent is used to minimize
typical functionals from elasticity. These functionals are
coercive and bounded below but are not defined everywhere on
an appropriate Sobolev space H®(fl) or product of such
spaces. For the case ft = [a,b], suppose w € H2(ft),
u € H2(ft) * H2(ft)HHQ(0) w(a) < w(b), and u'+w' > 0 on [a,b].
The function u + w is called a deformation of the interval
[a,b]. The functional I is defined as follows:
I(u) - fb[[u'(x)+w'(x)]2 + * ldx J a*- [u1 (x)+w* (x) | J
rb
J W(x,u'(x)) dx a
where W is the stored-energy, I(u) is the total
stored-energy and the potential energy is assumed to be
zero. We are interested in minimizing I while preserving
the boundary conditions of w. In the case ft = [a,bj and
s = 2, I is convex; but for bounded regions ft in IR2 with
appropriate smoothness on #ft and s • 3, the corresponding I
is not convex. The choice of s in both cases is made to /
insure I is Frechet differentiable at many points in H2(ft)
and respectively.
The following results are proved. The gradient Vl and
a "modified gradient" Q are derived. It is shown that Q is
locally Lipschitz, Q(u) is a descent direction and Q(u) » 0
if and only if Vl(u) «= 0. The existence of solutions y and
z respectively to the equations y'(t) » -Vl(y(t)) with
y(0) = uQ and z'(t) = -Q(z(t)) with z(0) « uQ, is proved for
t € [0,#) where S > 0 and u£+ w' > 0 on 0. If z exists for
all t in [0,oo), then lim z(t) exists and is a local minimum t-ta>
of I. If ft C R2, if z exists for all t in [0,ao) and if the
range of z is bounded in H|xH®, then there is an increasing
sequence {tj} of numbers such that lim tj = oo , i-too
lim || Q(z(ti)) I] » 0 , and lim zft,) = <J) exists in i-too i-to
where <J> e Under certain conditions, Vl(<J>) = 0.
Finally, this steepest descent process is implemented
in a True Basic code for the ft2 case and graphical output is
displayed.
TABLE OF CONTENTS
Page
LIST OF ILLUSTRATIONS iv
CHAPTER
I. INTRODUCTION 5
II. THE ONE DIMENSIONAL CONVEX PROBLEM 19
III. THE TWO DIMENSIONAL NONCONVEX PROBLEM . . . . 32
IV. NUMERICAL SOLUTIONS OF THE TWO DIMENSIONAL PROBLEM FROM ELASTICITY 52
APPENDIX 68
BIBLIOGRAPHY 74
iii
LIST OF ILLUSTRATIONS
Figure Page
1. Graph of initial pair W and Y 62
2. Graph of approximate solution pair W+Ut and Y+Vt 62
3. Cross sections of initial W, W(I,5) and W(5,J) . 63
4. Cross sections of solutions (W+U|)(I,5) and
(H+Ut)(5rJ) 63
5. Cross sections of initial Y, Y(5,J) and Y(I,5) . 64
6. Cross sections of solutions (Y+Vi)(5,J) and
(Y+Vj)(I,5) 64
7. Initial pair W and Y 65
8. Deformations of W and Y at k = 15 65
9. Graphs of approximate solutions W+U2 and Y+V2 65
10. Cross section W(I,5) 66
11. Cross section of intermediate stage: (W+U)(I,5) 66
12. Cross section of solution: (W+U2)(I»5)
at k • 100 66
iv
CHAPTER I
INTRODUCTION
Solutions to many mathematical problems involve finding
a minimum or critical point of a functional. Such
functionals are often derived as variational principles for
differential equations. Proving the existence of critical
points can be difficult and has required the development of
many sophisticated mathematical methods, such as the
"mountain pass lemma" [14, pp. 7-20] and the theory
developed for convex functionals, earlier by Rosenbloom [16]
and later much more extensively by Brezis and others [5].
However, some problems involve functionals which are
nonlinear, not defined everywhere and not convex. Such
functionals arise in elasticity. In this paper, a typical
functional I from elasticity will be investigated, where
X €<(>) is the total stored-energy for the deformation <j) of the
set 0 C R B , The study will be conf ined to R1 and R2, where
I is convex in R1 and I is not convex in R 2 . A constructive
method using steepest descent will be used with an emphasis
on applicability to computing. The setting will be Sobolev
spaces (see Adams, [1]). Great use will be made of the
Sobolev gradients, an idea used by Neuberger [10, 11, 12,
13], Richardson [15] and Glowinski, Keller and Reinhart [8].
Because many of the obstacles to overcome are present in the
one dimensional case, the key ideas will be developed and
implemented for 0 « [0,1] in this introduction and Chapter
II. The more complicated, nonconvex case will be dealt with
in Chapter III.
Let H2(fi) denote the Sobolev space H2'2(Q) of square
integrable functions on the closed, bounded region Q with L2
second derivatives. When it is clear, H2(fl) will be written
as H2. Similarly, let H§ denote the subspace H2flH^ of H2
which contains all the elements of H2 which satisfy zero
boundary conditions. Denote by C(ft) the space of continuous
functions on 0 and by Cl(Q) the subspace of C(0) with
continuous first derivatives.
For the remainder of this chapter, 0 is assumed to be
[0,1]. Let w be a fixed function in H2(Q) such that
w(0) < w(l) and u be a function in H§(0) such that
w1(x) + u'(x) > 0 for each x € 0. The function u + w is
called a deformation of the interval [0,1]. The functional
1 is defined as follows:
(1.1) I(u) = f T(u'(x)+w> (x) )2 + J 1 dx J0 L Iu' (x) +w1 (x) |J I U1 ( X ) +W 1 ( X ) | •
f M(x,u'(xj)dx •'o
where W is the stored-energy density, I(u) is the total
stored-energy and the potential energy is assumed to be
zero. The goal is to minimize I while preserving the
boundary conditions of w. One of the characteristics of
elasticity problems is that minima are not unique. In the
case of 0 • [0,1] C R1, I is convex (see Ball, [2]).
Another way to show the convexity of I is to calculate the
first and second Frechet derivatives of I on a certain
natural open subset BZ of . A function v f H| is said to
belong to the set BZ if and only if v'(x)+w'(x) > 0 for all
x e 0.
LEMMA 1.1 BZ is an open set.
PROOF: Since HQ is compactly embedded in Cl, there is an
M > 0 such that if v € H$ then ||v|| < M*||vjJ , [1, pp.
C1 HQ
97-98]. Suppose v e and v has property BZ, that is
v'+w' > 0 on ft. Since v is in C1, there is a j > 0 such $
that S = min {v1 (x)+w' (x)}. Let r • W T m ' L e t ** e Ho suc*1
x € 0 that Ijv-qJJ < r. Me know that |!v• — • II < -
H Q Sup C1
M ||v-q|| S M*r • - |. Thus, | v' (x)-q' (x) | < | for
£ £ each x in 0 and so v' (x) ~ 2 < (2> a n d v' (x)+w' (x) - ^ <
s q'CxJ+w'tx). Since v'(x)+w'(x) > 6 then q'(x)+w*(x) > ^ .
Therefore q belongs to BZ and BZ is an open set.
The open set BZ is a natural set with which to begin a
study of I. The functional I is defined for each u in BZ
8
since the second term of the integral is bounded. Of course
there are functions for which I can be defined and which are
not in H§ and so not in BZ. For example, let w(x) « x for
each x € [0,1] and define u by the following: let u(x) » x
for x € [0,l/3)r let u(x) » -2x+l for x 6 {1/a,2/a] and let
u{x) * x-1 for x € (2/3#l]. This function u belongs to Hg
and u+w is said to have cracks or buckles. When following
the steepest descent process using an initial member of BZ,
it is possible that the trajectory leaves the set BZ and
descends to a region of HQ containing functions with such
cracks. From computations, this is not likely in R1 but has
been observed for many initial <|> in the R2 case. This seems
to indicate that some functions $ have boundary conditions
which force a buckling when the method of steepest descent
is used to minimize I.
Returning attention to I and the calculation of its
first and second Frechet derivatives, suppose u is in BZ.
Since BZ is an open subset of HQ and H2 is embedded in C1,
then for each g and h in
(1.2) I'(U)h =• f W + W l h ' - "'I JQ L IU'+W'|2 J
and
(1.3) I"(u)(h,g) - zf^h'g* + h' 'g' 1 Jq L |u1+w'|3J *
9
Jn lu'+W' 3J
Setting g = h,
I«'(u)(h,h) « 2
which is positive if h # 0. Hence, I is convex, although
not strongly convex as defined by Rosenbloom [16* p. 128].
Note also that I is twice continuously differentiable.
In order to calculate the gradient of I, first note
that for each h € H§, ||h||2 » j* |"ha + (h' )2 + (h")2l, the H2
norm for the subspace . Let P2 be the projection of
L2xL2xL2 onto the subspace * 2 consisting of triples of the
form (h,h',b") and h(0) « h(l) « 0. This subspace is
isomorphic to Hjjf. Since u € BZ, sign(u* (x)+w' (x)) • 1 for
each x in 0. Thus,
I'(u)h = C r2(u'+wl)hl + 1 1 J0 L |u»+w'|2J
- f l \ 0»h + r 2 ( u . + w . , + 1 lh' + 0»h" 1 J I L lu'+w'pJ
2 (u1 +w') - * I+w1 I 2 1»
2(U1+W1) - 1
Iu1+w'I2
0
rh '
h'
h" L 2xL 2xL 2
2
h '
h«
h" L2xL2xL2
10
0 h " 2(u1+w1) - 1
•*2 h' lu'+w'I2 •*2 h' o . h". L2xL2xL2
n p. 2(u'+w') - 1
k ' + w f 0 1 1
where H^(q,q*,q") «• q for each triple In the subspace H of
L2XL2XL2. Thus, II P (0, 2(u»+w«)-l/|u'+w«|2 ,0) is the Hg 1 2
gradient of I at u, denoted by (?HgI)u or (?I)u
Since I is twice continously differentiable at u € BZ
and Vl is locally Lipschitz, then the differential equation
( 1 . 4 ) y«(t) = - (?I)(y(t)) and y(0) - u0 € BZ
has a local solution using the usual Picard iteration
technique. Since Vl is a function from Hq to H$, an
examination of the iteration technique used to prove the
existence of y yields that trajectories y stay (locally) in
Hjj. This contrasts with the not uncommon difficulty in
nonlinear semi-group theory in which one starts with a
* "n
generator A and has to prove that the composite (I - A) p
stays in the domain of A. The process generates a nonlinear
semi-group which is not defined everywhere, i.e. defines a
local semi-group.
One may argue that insistence on the use of requires
too strong an hypothesis. Use of Hq would allow for the
buckling observed in elasticity problems in higher
11
dimensions [2, 3]. The functional I is only lower
semi-continuous on Hq, Gateau differentiable in some
directions and not Frechet differentiable anywhere.
However, the intent here is to establish some theorems, the
hypotheses for which are based largely upon observations of
numerical work. With this foundation, it is hoped to
eventually extend the numerical methods to boundary
conditions associated with buckling.
However, the space H$ will be used in a surprising way
to develop a modified steepest descent. Again starting with
an h € Hjf, u € BZ and using the H1 norm defined by
i
II h ||2 " f (h)2 + {h" )2 , Hq 0
it follows that
11 (u)h « fTafu'+w') - 1 lh' L L lu'+w'PJ
0 "h " 1
|u'+w'I2 # h'
L2xL2
2(u1+w1) - 1
u'+w'•2
where P is the projection of L2xL2 onto the subspace M
consisting of pairs (h,h') for which h(0) » h(l) •» 0, which
is isomorphic to H$. Define Q(u) to be
12
II P (O^fu'+w1 J-l/ju'+w' I2). II I f
Three facts will now be established about Q that will
be useful in the steepest descent scheme.
frgMMA 1.2 If u € BZ then Q(u) € Hg.
PROOF : Suppose u E BZ. Q(u) is the element h of for
which the distance between (0, 2(u'+w'J-l/Ju'+w'|2 ) and
(h,h') is minimum. Let g • 2(u1+w')-1/|u1+w'|2. Since u
and w are both in H2 and u'+w' is in C, g1 is seen to be
2 / til' -f-W1' ) 2fu,,+w") + — — and g* € L2 * Computing Q{u) is thus
J U1+w1 J 3
equivalent to minimizing the functional G(v) => f* (v -J o
0)2+(v*- g)2. Taking the Frechet derivative G'(v)h and
using a standard integration by parts argument, one gets
i
G'(v)h * 2*/Q(v - v" + g')h , where v 6 and h € C0. In
order for v to be the minimum of G, it must be the case that
G'{v)h « 0 for each h € C®. This is equivalent to v
satisfying the equation
(1.5) v - v" + g' = 0 and v(0) = v(1) = 0 .
This is an elliptic second order differential equation for
which it is well known that there is a unique solution.
Since g' € L2, it is known that the solution to this
equation belongs [8, pp. 169-175]. Hence Q(u) is the
solution to (1.5) and Q(u) is in H§.
13
The preservation of the smoothness by Sobolev gradients
is a major factor in the decision to use a steepest descent
method. The advantage of using a Sobolev gradient instead
of an L2 gradient is quite evident computationally and has
been confirmed theoretically by Neuberger [13]. The nature
and form of the projections used in these gradients have
also been investigated by Neuberger ([10, 11, 12]).
The second fact about Q is that for each u € BZ, - Q(u)
is a descent direction in Hjj. In order to show this one
assumes Q is locally Lipschitz (this will be proved in
Chapter II). With this assumption, the differential
equation
(1.6) z'(t) - - Q(z(t)) and z(0) * u0 6 BZ
has a local solution, that is, z is defined on [0,d) for
some d > 0. For t € [0,d), let f(t) » I(z(t)). Taking the
derivative of f,
f'(t) - I'(z(t))z'(t)
- <Q(z(t)),z'(t)>H£
- (Q(z(t))Q(z(t))>H£
- - II «<*<*» ll'j-
Thus, f is a non-increasing function. Unless there is a t
such that Q(z(t)) « 0, f is a decreasing function and as t
14
increases, f • I<z(•>) decreases. The function I provides a
Lyapunov function for solutions z to z1 = - Q(z) (see Brauer
and Nohel, [4 ] ) . Considerable use will be made of Lyapunov
functions throughout this work.
Finally, a close relationship will be established
between (Vl)u and Q(u) for u € BZ.
LEMMA 1.3 If u € BZ, then (Vl)u - 0 if and only if
Q(u) «= 0.
PROOF: Suppose u € BZ and (Vl)u • 0. Let g »
2(u'+w') - 1 . Then, u1+w112
0 - <(?I)u,Q{u))H| » I'(u)O(Q(u))
'o" Qu g # (Qu)' 0. JQu)".
flg'i Qu)« •'o
X'<ra•[,&•]> !.'<>, El-•(&•]> <Q(U),Q(U})HI
This implies that || Q(u) ||2 - 0 and Q(u) - 0 in H|. Thus,
Q(u) « 0 in H§.
15
Similarly, suppose Q(u) = 0. Then,
• • <{<!».<»»
(Vl)u " "o" t(Vl)u]' J 9 J(Vl)u]". 2 0_
<(VX)u,(Vl)u)H2 .
Hence, {Vl)u « 0.
As a result of this lemma, the search for critical
points for I becomes a search for zeroes of Q. Of course a
critical point of a nonconvex functional I is not
necessarily a minimum of I. But in using steepest descent
methods to minimize functionals, critical points that are
not local minimums are usually very unstable numerically.
Critical points computed in numerous examples have shown no
signs of such instability. This is a good indication that a
zero gradient obtained in a stable manner is a local
minimum. In computing for the R2 case in which I is not
convex, there is numerical evidence of local convexity near
stable functions where the gradient is close to zero.
In Chapter II, the existence of a local minimum of I,
namely lim z(t), is proved assuming only that the function z t-tao
exists for each t € [0,oo). In Chapter III, it is assumed
16
that 0 is a closed, bounded region of R2 with a smooth
boundary. A corresponding function Q is defined mid shown
to be locally Lipschitz. Again under the assumption that
the solution z to (1.6) exists for all t € [0,oo), it is
shown that there exists an infinite, unbounded and
increasing sequence of positive numbers {tj J, t such that
lim -f ||Q(as (11) )|l } » 0. Under the condition that the i-*oo 1 HqXHq j
range of z is bounded [10, 11] in it is shown that
there is a limit point 4> of {z(tj)> in Hj}*H$ so that if <J> is
in BZ then Q(4>) * 0.
Chapter IV describes how to implement the steepest
descent for these problems on a computer. A finite
difference scheme is used, employing the Sobolev norms. Key
parts of a True Basic code are discussed, including a
Gauss-Seidel type linear solver for the R2 case. The output
on several test problems is presented, including some
graphical output to illustrate the time-stepping process.
CHAPTER BIBLIOGRAPHY
1. Adams, R., Sobolev Spaces, Academic Press, New York, 1975.
2. Ball, J. M., Constitutive inequalities and existence theorems in nonlinear elastostat ics, Nonlinear Analysis and Mechanics: Heriot-Watt Symposium 1 (1977), 187-241.
3. , Singular Minimi zers and their Significance in Blasticity, Directions in Partial Differential Equations, Academic Press, 1987, 1-16.
4. Brauer, F. and Nohel, J., Ordinary Differential Equations, Preliminary Edition, W. A. Benjamin, Inc., New York, 1966.
5. Brezis, H., Operateurs maximaux monotones et semi-groupes nonlineares , Math. Studies 5, North Holland, 1973.
6. Curry, H. B., The method of steepest descent for nonl tnear minimi zation problems, Quart. Appl. Math., 2 (1944), 258-261.
7. Evans, L. C., Quasi convexity and partial regular* ty in the calculus of variations, (to appear).
8. Glowinski, R., Keller, H. and Reinhart, L., Continuatioi Conjugate Gradient Methods for the Least Square Solution of Nonlinear Boundary Value Problems, Rapports de Recherche, Institue National de Recherche en Informatigue et en Automatique, 1982.
9. Miranda, C., Partial Differential Equations of Elliptic Type, 2nd rev. ed., Springer Verlag, 1970.
10. Neuberger, J. W., Gradient Inequalities and Steepest Descent for Systems of Partial Differential Equations, Proceedings of the Conference on Numerical Analysis, Texas Tech University, 1983.
11. , Some Global Steepest Descent iesults
17
18
for Nonlinear Systems, Trends in Theory and Practice of Nonlinear Differential Equations, edited by V. Lakshmikantham, New York, Marcel Dekker, 1984, 413-418.
12. , Steepest Descent for General Systems of Linear Differential Equations in Silbert Space, Proceedings of a Symposium held at Dundee, Scotland March-July 1982, Lecture Notes in Mathematics, Vol. 1032, New York, Springer-Verlag, 1983, 390-406.
13. , Steepest Descent for Systems of Sonlinear Partial Differential Equations, Oak Ridge National Laboratory CSD/TM-161, (1981).
14. Rabinowitz, P. H., Kinimax Methods in Critical Point Theory with Applications to Differential Equations, Regional Conference Series in Mathematics, no. 65 (1984).
15. Richardson, W. B., Sonlinear Boundary Conditions in Sobolev Spaces, Dissertation, North Texas State University, 1984.
16. Rosenbloom, P. C., The method of steepest descent, Proc. Sympos. Appl. Math. VI, Amer. Math. Soc., 1956, 127-176.
CHAPTER II
THE ONE DIMENSIONAL CONVEX PROBLEM
The condition of convexity on functionals has been
widely utilized in minimization problems. Extensive work
has been done on convex elasticity problems by S. S. Antman
[2] and J. M. Ball [3, 4]. Using the strong assumption of
convexity, an extensive theory of semi-groups has been
developed by H. Brezis and others [6]. An early development
by P. C. Rosenbloom [10] proved asymptotic convergence of
trajectories of a semi-group satisfying equations of the
form y'(t) • - ( V j ) y ( t ) where y{0) • u and J satisfies the
strong convexity condition that JM(u)(h,h) > c*|| |{h ' o r eac*1
u and h in the Kilbert space H.
The steepest descent method has been adapted by
J. W. Neuberger for the following least squares problem [7,
8]. Suppose each of H and K is a Hilbert space, F is a
twice continuously differentiate function from H to K and
<j>:H -4 K such that $(u) » (1 /g)|}F(u)jj for each u € H. Under
the assumption that
( 2 . i ) ii<*4» uii„ 2 M - r < u > i i K
for some M > 0 and every u € B Q H, Neuberger proves there
exists a solution y:[0,oo) 4 I to the equation
19
20
y'(t) = -(V((>)y{t) and y(0) » u0, and if Range{y) C B then
lim y(t) exists and is a zero of <j> and of P. This gradient t-ta>
inequality is satisfied by many nonconvex functionals. The
set B is typically a bounded subset of a Sobolev space H and
the above result requires that the trajectory y remain
bounded. The results of this paper are an adaptation of
Neuberger's approach to the variational problem (1.1), but
without the advantage of a gradient inequality. In this
chapter, it is assumed that [0,1] » ft C R1. In this case
the functional I is convex. Although much is known for such
convex functionals, our constructive method may shed some
new light on non affine minimizers for 1 (see Ball [4,
pp. 4-5J). Most of the key difficulties in higher
dimensions are present and may be clearly seen in the one
dimensional case.
Attention will be focused on the "modified" gradient Q.
Recall from chapter I that - Q(u) is a descent direction,
Q(u) = 0 if and only if (Vl)u » 0 and if u € BZ then
Q(u) G H§. An important question arises as to the existence
of a solution z to the differential equation
(2.2) z'(t)"-Q(z(t)) and z(0) » u0 €
where t € [0,oo). In order to prove that there are local
solutions to such equations, it is sufficient to prove that
Q satisfies a Lipschitz condition.
21
THEOREM 2.1 Q is locally Lipschitz on BZ.
PROOF: Recall that I and I* are given by
I(u) • [(u'+w1)2 + 1 1 J 0 L ju'+w'|J
and
riuih. - si3n(u'+w,)-h']
J 0 L |u'+w'|2 J
2 c for each u and h in Ho and signfu'+w') • 1 since u BZ.
Let u € BZ. Since Hg is embedded in C1 and u'+w1 > 0 on
[0,1], there is a neighborhood B(u,rt) and positive numbers
and M2 such that for each v € B(u,rt)f JJl/tv'+w' )|j < Mt
and JJv'+w'Jj < M 2 . Let v € B(u,rt).
Let f » 2(u'+w1) - 1 and |u'+w'12
2(v'+wf) - * and note that f1 and g1 are in L2
|v'+w'|2
Let F - Q(u) ® II ? 0
x proof of Lemma 1.2, F and G satisfy the differential
and G » Q(v) = I! P [°j . From the 1 1 u
equations -FM+F+fl « 0 and -G"+G+g' • 0, respectively.
Using the equivalent norm of Hjj, ^ = /*(hw)2, it follows %
that
||Q(u)-Q(v)||^
22
n p "o" - n p 0 i i JP. .9.,
( 2 . 2 )
« f 1 (F,,-G" )2 = f 1 {F+f ' -G~g1 )2 •'O •'O
- f1<F-G)2 + 2(P-G)(f1-g') + (f'-g')2
•'o
< 2 /" * (F-G)2 + 2 /*1 (f'-g')2 »0 *0
The two terms of the right side of (2.2) will be
considered separately. Since is embedded in L2 and Hjj is
embedded in , there are positive constants Kt and K2 such
that for each h t€ and h a€ , jjh4 j|2 < Kt -||h| J|'
2 and L2
||h2|j2 < K2.||h2||
2 . For the first term from the right side "h§
of (2.2), it follows that
(2.3) 1 (F-G) 2 - HP-Gil^ < Kj •||F-G|(~
K, P "o" - f
f Sm
i
"o" "o'
.f ss.
HA
2
L2XL2
2
L2xL2
< K,-l
- K,/'(f-g)2
- K,J j2(u'-v'
< 2K, r ' m . - v i ' L r i»i-w'i-fiv'*w,irl
Jq \ [(u'+W1 ) a (V1 +w' ) 2J J
(u'-v1) (u'+v'+^w1 )"|2
(u'+w ')2(V
if1 +2w' )"|
•+W)'i J
23
l f u'-V •*0
< 2Kj f (u'-v')2C22+(2M1M24)2]
< 2Ki'Hu-vll2 [22 + (2M1M24)2]
H0
< 2Ki K2-||u-v|j2 [ 4+4M2M| ] < K-||u-vj|2
" "H§ " Hq
where K - 2KtK2[4+4M?M|].
Consider now the second term f (f'-g1)2 from the right Jo
side of (2.2). Differentiating f and g, gives that
f - 2u"+2w,,+2(u"+W") € L2 and a' - 2v"+2w,,+2*V ,+W") € L2. ]u'+w'|3 |v'+w'|»
Thus,
(2.4) rNf-g') 2
n2{u"-v'M+2r(u,,+w"Hv,+w,)8"(v,,+w,,>(u,+w,,3n2
L |u'+w'|3 Jv'+w113 J J < s f i i u•.-v>M2+r<
^,Hw,,>(v,•,^w,
JQ [ Iu'+w * J s1v1+w
) 3 _ { V . ( U . + W ' )8-|2
(2.5)
5 8'llu"vll^
+ 8-m|2 j*1 [(VL"+VI") (v'+w1 )-(v"+w")(u'+w1 )]2.
Let S » u'-v1, 9 • u'+w' and a • v'+w'. It is clear that
||ff|| < M and lUII < m . Since H(| is embedded in C1, Sup 2 Sup 2
there is a K3 > 0 such that ||tf|| < K3*||j|j * K3«J|u-vJ| Sup H|
24
< k 3 T j . Let L » K3*r1. Focusing on the integral of the
second term of the right side of (2.5), the function
pointwise satisfies the following:
t(u"+w"){v'+w1)-{v"+w")(u1+w') ]2
»|u" (v'+w1) 3+ w" (v1 +w') ® — v'^u'+w1 )3- w'^u'+w1)8!2
< j |-V"(T3+ u" ( f f - S ) 3 l + jw"!®3-^3] |J
< pu"-vH | |#|8 + |u" | I ( - 3 < r 2 6 + 3 f f $ 2 - $ 3 ) |
+ Jw" j | | (<r2+ff<r+*2 ) |J
< |ju"-v"|M§ + |u"||^|(3M^+3M2L+L2) + |wM|J #|3M|J 2
< 3J Ju"—v"|2M| + |u"|2|£|2(3Ml+3M2L+L2)2
+ |w" |a|tf|29M|J
Integrating, it is clear that
J[(u"+w")(v'+w1)-(v"+w")(u'+w1)]2
o
< (u"-V,,)2+ 3(3M1+3M2L+L2)2 (U")2j
<£ 3M|.[|u-v||2 + 3(3M|+3M2L+L2 )2||u||2 *K§-jju-vlj2
Hi UH§
FsM| + 3{3Ml4-3M2L+L2)2||u||2 K§
1
+ 9M24||»||^]||U-V||^
25
M-IML n -
2
L 0
where M is the constant from the previous step. Using this
inequality in (2.4), and then using (2.3) and (2.4), it
follows that
2 f 1 (F-G)2 + 2 f *(f•-g')2 •'O •'O
< 2-K.||u-v||^ + 2.[a-||u-v||^ + 8-Mj2-M-||u-v||^]
• (2»K + 16 + 16«M|2«M ) • j|u-v||2 . Ho
Thus,
IJQ{tl) —Q(V)|j2 < {2*K + 16 + 16'M|2«M )*||u-V||2 H0
for each v € B(u,rj). Therefore, Q is locally Lipschitz on
BZ and the theorem is proved.
Using this result and a standard existence argument
(see Brauer and Nohel, [5, pp. 364-417]), for each u0 € BZ
there is an e > 0 and a function z:[0,e) -> H2 such that
z'(t) « -Q(z (t)) and z(0) =» u0 . This is of course a local
result. But in numerical experiments for this steepest
descent problem, the global existence of solutions z to such
differential equations has always been observed. This
phenomenon can hold even for problems which have no
solution. For example, the least squares problem where
F:H^([0,1]) -+ L2([0,1]) such that F(u) « u'-l-u2 and
26
<|>(u) « { ^ / 2 ) | J ' F ( u ) ||2 produces a globally defined trajectory
z, but F has no zero for an Interval 0 of length more than
it . The global existence of trajectories shall be part of
the hypothesis of the following theorem.
THEOREM 2.2 Suppose z:[0,oo) -» H$ such that for each
t € [0,oo) z'(t) • -Q(z (t)) and z(0) = u0 . Then, 11m z(t) t-ta)
exists and is a member u of and lim ||Q(z(t))|| = 0. t-to Hq
PROOF: Let f: [0,co) -4 R such that for each t € £0,co),
f(t) • I(z(t)). Using the differentiability of I,
(2.6) f»(t) - l>(z(t))0z'(t) - <Q(z(t)),z'(t)>Hi
« <Q(z{ t)) ,-Q(z(t) )>H^ - - ||Q(z( t) )||2
and
(2.7) f"(t) = I"(z(t))[Q(z(t)),Q(z(t))]
- 2f 1[l+ - rl EQ(z{ t))1 J2
JQ L |Q(z(t))*+w'|8J
> 2^1[Q(z(t))']2 - 2 • Jj'Q (z {t}) ||
- 2•(- f'<t)).
Therefore,
(2.8) f"(t) > -2•f*(t)
for each t € [0,oo). Dividing both sides of (2.8) by the
27
negative quantity f1(t) and integrating, it is clear that f
satisfies the inequalities:
(2.9) f"(t) K -f'(t) < -21 7TJt) ~ 2 a n - f ' ( 0 ) - ®
and hence,
(2.10) 0 < -f'(t) < -f1(0)e"2t.
The following is an argument that lim z(t) exists in t-400
H|. Suppose s and t are in [0,co) and s < t. Let k and n be
the non-negative integers such k: 5 8 < an<* n~l < * 1 n.
Thus,
||z(t)-i(«)||Hj - || /a z' jfH 1
. r i»k K J
< if* f /1+1-f'(0)e_2i 1 / 2
i-k L 1 J
<ff l [f-(0)]i/2[-e-2i-2 ->• e-2i] i -k 2
l/2r - ~il/2 |,-1
i rnk < [ - n o , ] 1 ' ' ^ ] 1 ' 8 . ft- e"1
Since the series X e-i converges, then the trajectory z{t) i >o
28
is Cauchy in as t approaches oo. Therefore, the lim z(t) t-to
exists and is in . Call this limit u. Observe that
lim f'(t) and is zero because of (2.10). Since t-ta)
f ( t) a -||Q(z( t) )|j2 for each t € [0,oo}# then
lim ||Q(z(t} )||2 « 0. t-*oo
Even though the functional I is coercive in the Hg norm
but not in the H§ norm, it has been observed that for many
trajectories z the Range(z) is bounded. The following
theorem uses the hypothesis that z remains bounded in HQ.
THEOREM 2.3 Suppose z satisfies the hypothesis of Theorem
2.2 and u = lim z(t) € . Then, t-taj
(i) if Range(z) is bounded in H§, then u € , and
<ii) if u € B2I then Vl(u) = 0 and u is a local minimum of I.
PROOF; (i) Suppose the Range(z) is bounded in H§. Let
{tj> 00 be an increasing sequence in [0,oo) such that i»I
lim (t|> -» GO. There is a subsequence <z(Sj)} of (ztt;)} i-to
which converges weakly in H§ to a point v € HQ. Thus,
{z(S|)> converges weakly to v in . But, u is the strong
limit of z(t) in and so is the weak limit of {zfsj)}.
Since weak limits are unique, v = u and u G H§.
(ii) Suppose u € BZ. Since Hjf is compactly embedded
in C1 [1], there is a subsequence of {z(si)) which converges
29
strongly to ari element vi In C1. Strong convergence in C1
implies strong convergence in , so that vt = u. Suppose
{uk} is the convergent subsequence in C1. Since u € BZ and
<uk} converges strongly to u in C1, there is an M4 > 0 and
a positive integer n such that if k > n then
JujJ (x)+w* (x) | ~ an(* Ju1 (x)+w' (x) | - * o r ®acl1
x € [0,1]. Let H2 be the maximum of <||uk+wJjci}
a n d | j u + w | | c 1 '
Let h € such that ||h|jH = 1» ^or each k > n.
|<Q(uk)-Q(u),h>Hi|
< I f ^ - u ' J h ' + [("'-Vk'Xu'+Uk'+Zw'rtb, |J0 1 (UjJ+W1 )'2 (u« +w' )2 J
< 2* jj'Nu^-u'Jh' |
+ I f 1 j"[u'~uk ' 1 |u'+uk '+2w' |*[ jh,j Ko 1 (u£+W* ) * (u'+W« ) '* J
< 2-|K"UIIHJ IMIHJ + I-2Mj • |h* I -Ml
S 2-|K"UIIHJ IMIHJ + .WJ-IK IIH} IMIHJ =• (2+2H, ) - ||u-ut [|H, .
Since ||u-ut||H, < ||u'-ui||0, , then {| <Q(uk )-Q(u) ,h)H, | J
converges to 0. Hence, (Q(uk)} converges weakly to Q(u) in
H$. Since j||Q(uk -» 0 and ||Q(u)|| < lim inf {||Q(uk)||H J ,
then ||Q(u>)|H£ = 0 (see [9, p. 60]). Therefore, Q(u) • 0.
30
By lemma 1.3, this implies ( V l ) u = 0. Since 1 is convex,
then u is a local minimum for I on HQ.
CHAPTER BIBLIOGRAPHY
1. Adams, R., Sobolev Spaces, Academic Press, New York, 1975.
2. Antman, S. S., Monotonicity and invertibility conditions id one-dimensional nonlinear elasticity, Nonlinear Elasticity, ed. R. W. Dickey, Academic Press, New York,1973.
3. Ball, J. M., Constitutive inequalities and existence theorems in nonlinear elastostatics, Nonlinear Analysis and Mechanics: Heriot-Watt Symposium 1 (1977), 187-241.
4 . , Singular Minimizers and their Significance in Elasticity, Directions in Partial Differential Equations, Academic Press, 1987, 1-16.
5. Brauer, F. and Nohel, J., Ordinary differential Equat ions, Preliminary Edition, W. A. Benjamin, Inc., New York, 1966.
6. Brezis, H., Operateurs m x i m n u x mono tones et semi-QJdupes noftlin&iTt,S, Math. Studies 5, North Holland, 1973.
7. Neuberger, J. W., Gradient Inequalities and Steepest Descent for Systems of Partial Differential E q u a t i o n s , Proceedings of the Conference on Numerical Analysis, Texas Tech University, 1983.
8. , Some Global Steepest Descent iesults for Honlinear Systems, Trends in Theory and Practice of Nonlinear Differential Equations, edited by V. Lakshmikantham, New York, Marcel Dekker, 1984, 413-418.
9. Riesz, F. and Sz.-Nagy, B., Functional Analysis, Ungar, New York, 1955
10. Rosenbloom, P. C., The method of steepest descent, Proc. Sympos. Appl. Math. VI, Amer. Math. Sac., 1956, 127-176.
31
CHAPTER III
THE TWO DIMENSIONAL NONCONVEX PROBLEM
This elasticity problem in two dimensions has a
nonconvex functional I. However the main difficulties
encountered in two dimensions are of the same nature as
those in Chapter II for the one dimensional case. Let G be
a closed and bounded region in IR2 with a smooth boundary.
Let Hg(ft) » H3(fl)flH£(G). Suppose (w,y) € H3(fl)xH3(fl),
(u,v) € Hg(fl)xHjj(fl) and « u v(x) «= ff(x) • det D[j}ty]
- det["u*+w* u2 +w 21( x ) f o r ' Lvi +Yi v2+y2j
each x € (I. E'or each appropriate {u,v), define I(u,v) by
!u,v) - J (u I(U,V) - I (U|+w t)2+(U 2+W 2)
2+(V 1+Y 1)2+<v 2+y 2)
2+ 1
Q |ffU,vl
811 f W(Dd>)
where (j) = (u,v). Note that I is not defined for all
<u,v) € HjjxH|j.
An affine deformation <|> is one of the form
<|><x} • A0(x) + x0 where A0€ M+xn and x € Rn (see Ball, [4,
page 4]). W as defined above is H 1 , p quasiconvex for
1 < p < m. The condition of quasiconvexity, combined with
certain growth conditions on W, implies I is sequentially
32
33
weakly lower semicontinuous (swlsc). This Implies by
standard arguments that I has minimizers (Ball [3, p. 190]).
Ball has pointed out that, although non-affine minimizers
supposedly abound, little is known about such minimizers or
their smoothness [4, p. 5J. L. G. Evans [5, p. 3] has
proven partial regularity for Junctionals which satisfy a
uniformly strictly quasiconvexity condition, that is,
(3.1) jfw(A)+4 |D<J)|dY < jfw(A+D<j»dY
for some i >0, every smooth, bounded open O C R 2 , each 2X2
matrix A, and every <|> € Cl(Q#R2) with <|> » 0 on #Q. If
\ - 0, this is the definition of guasiconvexity. As Ball
points out, this does not allow for the singular behavior of
W(A) as det(A) -+ 0+ [4, p. 5]. In this paper the approach
is constructive with hypotheses suggested by observations of
computations. Conditions will be given that imply existence
of <|> in c^xc1 for which (Vl)«f» » 0. Many such functions
observed in computing are non-affine. Initial boundary
conditions which are planar do give rise, as expected, to
affine solutions.
Let BZ be the subset of H{xh{ such that (u,v) € BZ if
and only if #u v(x) > 0 for each x in ft.
LEMMA 3.1 BZ is an open subset of H(jxHj|.
PROOF: Using the fact that H§xe^ is compactly embedded in
34
the proof is similar to that of Lemma 1.1.
The choice of H|XH| is made so as to insure that X is
Frechet differentiate on BZ. The strategy for
investigating I will be the same as in Chapter II. For each
(u,v) and (g,h) in BZ, differentiating I yields
(3.2)
I'(u,v)(g,h)
' J -2 I (Uj +Wj Jgi + tUa+WaJga+tVt+yt Jh^va+ya )h2
0
J-sign(g) \« I2 I u,vl ft
t («i+wt )h2+(v2+y2)gt
(v1+y1 )g2-(u2+w2>hi 1
(Uj+Wt )' r - (v 2 +Y 2r ffl (u 2+w 2 ) + 1 + ( V i + y t ) 92 (v t+Yt ) +(u 2 +w 2 ) 9 hi
.(V2+Y2), CI - (Ui+W t )J ^2, (Uj+Wj )' (u2+w2) (V 1+y 1) ,(v2+y2)J
-(v2+y2)1
+{u2+w2) - (U i+Wj) )
where 9> is the projection from L2XL2XL2XL2XL2XL2 onto the
subspace consisting of elements of the form
(9rh,g< ,g2 ,h< ,h*} and n :R6-* IR such that i»l
II (xt ,x2 ,x3 ,x4 ,x5 ,x6 ) a (xj,x2>. Note that the first two l > i
components in the left side of the above inner product
expression have been deleted because they are (0,0) and the
two equivalent norms of H^xH^ have the same value in this
case. Let Q be the function from BZ to HqXH^ defined by
35
Q(U,V) = n v i * i
( ((Ui+W!)] 12 (U2+W2) 1 ItVi+Y!)
-(v2+y2
+ 1 +<v1+y1) T
&2 +(u2+w2) -(Ui+Wt )j J
(3.3) H 9> 1 # 1
o (JT 91 I l 34>J
''i^1 ;• yi (F,G) where is the projection of L2xL2xL2 onto the subspace
consisting of elements of the form (g>0i*g2)' ^ = ^iX9>i and
ir1 is the function on R3 such that 1t (Xi ,x2 ,x3) » x t. Since
Q is a projection, finding Q(u,v) is equivalent to
minimizing the functional y:BZ -f R defined by
?(P»q) = J*(P-0)2 + ( P i - a 4 > ) 2 + (p2-b<t>)2
a
+ {q-0)2 + {q1-c<j>)2 + {q2-d<j>)
2.
A variational argument gives that a minimum of f is a pair
(F,G) satisfying the pair of differential equations
(3.4) AF * 2A(u+w) + 2 det D
F P " [v+y] + F
with F « 0 on d(i, and
AG = 2A(v+w) + 2 ^ det + G (3.5)
with G = 0 on $0. By the theory of second order elliptic
partial differential equations (see Miranda [6]), for each
g 6 H1 there exists a unique solution y to the equation
36
Ay « y + g and y = 0 on $0
such that y € H$. Since (u,v) € BZ, then for each equation
(3.4) and (3.5) the corresponding g is in HQ. Thus, there
exists a unique pair of solutions (FrQ) to (3.4) and (3.5)
such that Q(u,v) • (F,G) € HgxH§. Hence, Q is a function
from BZ to HjjxHjf •
Following the same strategy as in Chapter II, the next
step is to show that Q is locally Lipschitz on BZ. Several
technical lemmas are required to begin this effort.
LEMMA 3.2 Suppose (u,v) = <j) € BZ. Suppose At , A2, A3 and
A4 are functions defined on HffxH|f by the following:
U> *!«)» - q = 2. 3 or 4
(ii) A2 (4>) - [det D(<(>) ]p = <tp , p - 1 or 2, ap -
(iii) A,(+) - det D[v*J
(iv) A4«j» » [A3«J»|p , p - 1 or 2 .
Then, each of At, A2, A3 and A4 is locally Lipschitz as a
function from H$XH§ to C(FL), H1(FT), H 2(0) and L2(FI)
respectively.
PROOF: Since H C R2, H1(0) and Hjj(Q) are algebras (see
Adams, [1]) and H$ is embedded in C1. By elementary
calculations, one has that each of Aj, A2, A3 and A4 is
37
/
Prechet differentiable. Thus, they are each locally
Lipschitz.
In showing the desired properties of Q, the norm of
HjjxHg must be considered carefully. As is well known, an
equivalent norm for H§(ft) is given by
where A is the Laplacian operator (see Agmon, Douglis and
Nirenberg, [2]}. A typical calculation shows that for Hjj,
the usual norm is equivalent to one that depends only on the
non-cross partial derivatives Ujjj and u222 * Thus, it is
easy to see that a third equivalent norm for H® is given by
INI2 38 / (Au)2 + ( [Au]j )2 + ([Au]2 )2 ,
with the first term being necessary. Use of this equivalent
norm will be made in showing that Q is locally Lipschitz.
Another technical but elementary lemma is needed to
show the consequences of products of such functions as At,
A2 » A3 and A4.
LEMMA 3.3 Suppose (U,v) - 4> € BZ, 0 > 0, ? :BZ -4 H2,
*8:BZ -4 H1, «:BZ -» H1 and each of r, * and at is Lipschitz on
the open ball B[<J>,y?] C BZ. Then, there is a N > 0 such that
for each ff € B[<J), ],
3 8
/ [ y { (O<§ U ) W ( < r ) - s r (<|>)<S < < j > ) * ( < j > ) ] 2 < N - | | < J > - f f | | 2
Q H g x H g
P R O O F : L e t M > 0 b e a L i p s c h i t z c o n s t a n t f o r e a c h o f ^
a n d w o n B [ ( | > , ^ ] . S i n c e t h e r a n g e s o f ? , "3 a n d # a r e
c o n t a i n e d i n H ^ f i ) a n d H ^ f Q ) i s e m b e d d e d i n L 4 ( f i ) , t h e n
t h e r e i s a M t > 0 s u c h t h a t | | < > H ' ) | | 2 < M t f o r e a c h 9 i n
B [ < f > , ) ? ] a n d A = 9 , <s o r » . S i n c e r a n g e o f y C H 2 ( f t ) a n d
H 2 ( f l ) i s e m b e d d e d i n C ( 0 ) , t h e n ? ( B [ < } ) , ; # ] ) i s u n i f o r m i l y
b o u n d e d o n U b y a c o n s t a n t M 2 • H e n c e , f o r e a c h 9 fc B [ < j > , ^ ]
f [* (*)* (*)»(*) -» ($)* ($)* ($) 1*
< 3 / [ y U ) - * « l » ] 2 [ < 9 U ) * U ) ] 2
• ' a
+ 3i\
0
+ | r < $ ) | 2 [ * ( 0 * ( $ > - * ( $ > * ( < • ) ] *
0
< 3 | | y ( f f ) - y ( < j > ) | | 2 E * U ) ] 4 ) *
+ 3 M 2 [ * f ( f f ) - « « l » ] 4 J T
+ 3 M 2 t < 9 ( i r ) - ^ ( < J » 3 4 ^
< 3 M | | 4 > - d | 2 ( M 1 ) 2 ( M 1 ) 2
H ^ x H g
+ 3 M 2 ( M t ) 2 | | » ( ) - W (4>) | | 2
H $ * H ©
39
+ 3M 2 (Mt > 2 ||«S {<T }-<« «|>) ||2
H^XH I
+ 3M 2(M t )2M||^-4>||2
H||XH§
» [3M(Mj ) 4 + 6M2 (Ma ) 2M3 H -4>1|2
Thus, N « 3M(M|)4 + 6M2(Mj[)2M and the lemma is proved.
A similar result can be proved in which rsBZ -* H2,
<«:BZ -» H2 and »:BZ -* L4 and each locally Lipschitz. These
results are essential for the following theorem.
THBQRBM 3t4 Q is locally Lipschitz on BZ.
PROOF: Let (u,v) « (J) € BZ. In analyzing Q» functions such
as kt , A2, A3 and A4 from lemma 3.2 and similar ones in
which u+w and v+y are interchanged, must be considered.
Choose 0 > 0 and N > 0 such that B[(4>),j?] C BZ and this ball
of radius $ and the constant N satisfies the hypotheses and
conclusions for each of the ten applications of lemma 3.3
that will used in the next paragraph. Let
(r,s) = ff € B[ (<J>),/? ]. Let Q(u,v) ® (F,G) and
Q(r,s) » (H,K) . Let ff„ „ • <r and ff_ * $. Thus, u>v r,s
2 i w . p n i
2
[aj LKJ|
40
• r-«n; 3 + II°-kii^
Considering the first term from above, it follows that
(3.6) I|*-hH2 " / C(AF)-(AH)]2
Hjj ''ft
+ / [ (AF) j - (AH) j ]2 + / [(AF)2-(AH)2]2,
Ju
Considering the first term on the right side of (3.6) and
using (3.4) for each of F and H,
/ [(AF)-(AH)]2
= jTl^Mu+wl+jJiyr d.t D[v'y]+F
- ^(r+wl-yfpr det D [s+y] ~H] 2
(3.7) < 3j£«<A»-Ar)» + 3^^-jip aet D ^ ]
" fijT let D[J y]]2 + 3jf(F-H)*.
If for each (p,q) € r(p,q) - ift 1 ia, *(p,q)(x) « 1 p-q|
and *(p,q) = det DJ p,q , then lemma 3.3 applies and the [q+y J
right side of (3.7) is less than or equal to
12 • ||u~r||2^ + 12jf «J»*(<j» - r (0*(/)j2+ 3 • ||F-H||
- 12-|^'"HSxhS + + 3-||P"H|1^
2
L2
41
(3.8) < 12(1+N) • IH*~'llggxjjg+ 3 - | M I ^
The second term of the right side of (3.8) will be dealt
with as it appears again in the second term of ||f-h||2 . " l!Hg
Next, considering the second term of |Jf-H||2 from
(3.6), it follows that
jfC(AF)j-(AH)|]2 ft
I
2 - 2A (r+w) i - [jj-pde tD [s+y] ] t ~
H l ]
(3.9) i 4jT[A(u+w)i-A(r+w)i ]2+ 4 f (Fl-Hi)2
II II
+ 4[jffr•" 1 • a®tD[vly] * | f p " V d e t D U y ] ] 2
+ " - l V [ d e t D U y ] ] J 2 '
Focusing on the third term of the right side of (3.9), if
At(r,s) = * 14 , A2 (r,s) » £det D[g+y]J • *i» a n d
A3 (r ,s) = det D|g+y] f o r each (r,s) € B[<[>, 3, then these
three functions satisfy the hypotheses of lemmas 3.2 and 3.3
with /? and N as the constants from lemma 3.3. Thus, the
third term of the right side of (3.9) is less than or equal
to 4*N'||<b- || " "HJxh$
42
Similarly, for the fourth term of the right side of
(3.9), if Aj(r,s) « i a 2 ia , A2(r,s)(x) - 1, and I r, s I
A4(r,s) * j det D[s+yJJ1 ^o r ®ach (r,s) € then these
three functions satisfy the hypotheses of lemmas 3.2 and 3.3
with and N as the constants from lemma 3.3. Thus, the
third term of the right side of (3.9) is less than or equal
to 4" N • ||<J) <r ||2 •
The first term of the right side of (3.9) satisfies
4/* [A(u+w) t -A(r+w) 1 ]2 - 4/*[A(u)1-A(r)1 ]
2
11 HH§xHg
This leaves only the second term of the right side of (3.9)
to be analyzed. Recall from (3.3) that Q(u,v) is defined as
Thus,
Q (u, v) » II 9" 111
I *0 »4>
ld<M
» fjr I 1 b* „b4>. "•"I])
4 / (Fj-H,)2 < 4[IF—H|j2
" V
- (F,G)
SB n '0 0 " a<l> - n a<r
K J 1 H'
SS 4 I V
L2XL2XL2
43
(3.10)
< 4 • 1 "o " "o
b4, L<i>J
a<r Lvl
L2XL2XL2
= 4j^2(u1+w1 )-Jr(v2+y2 )-2(r,+w1 )+$r(s2+y2 )J 2
+ j^2{u2+w2)+|T(v1+y1 )J
< sajf^-rj >2 + (u2-p2)2 + [^r(v2+y2) - *^2-(s2+y2)j
+ [jr(vi+Yi) " fr(si+Yi>]
< 3 2 . | | u - r | | » + 3Z.||vr-s||2 kl Q tlQ
+ 32jr[jT(V2+y2) - ^t(s2+Y2 )j 2
+ 32 [i?<vi+yi) - i?(si+Yi )]2•
Focusing on the third term of the right side of (3.10), if
A^r.s) * |fl 1 »y , A2(r,s)(x) = 1 and A8(r,s) » (s2+y2> for I r,sl
each (r,s) € B[<|>,4], then these three functions satisfy the
hypotheses of lemmas 3.2 and 3.3 with fi and N as the
constants from lemma 3.3. Thus, the third term of the right
side of (3.10) satisfies
3 2Jfl[i T { V 2 + Y 2 ) " i*(S2+Y2)]2
< 32J^[Ai <4>)A2 (4>)A3 (<J>) - At (<r)A2(<r)A3(<r)J2
< 32 'N* {kb-ffjl2
HjjxH I
44
Similarly, the fourth term of the right side of (3.10)
satisfies the following inequality
32jdD7<V,+Y'1 - ^ < s ' + Y i ' ] 2 S 8-N-||<t>-<r||^^.
Combining these inequalities involving the right side
of (3.10), it has been shown that
4f(Ft-Hj)2
* 32-|lu-rll',+ 32-IMI*,+
1*0 H0 W0xW§
- (32»N,+ 64 *N) •||<M||2
H§xH§
where N| > 0 is an embedding constant for H$XH|| in H^XH§.
These calculations and inequalities hold also for the second
term of the right side of (3.8).
The third term of the right side of (3.6) can be
bounded in the same way as the second term. Combining all
of these inequalities, it follows that there exists an
M_ > 0 such that for each # € B[<b,/] F
In an analogous way, there exists an MQ > 0 such that
for each r € B[<j), ]
45
IMI;I * VIMI H3
Combining these two inequalities, it follows that
CM; H ; x H ? <
Therefore, Q is locally Lipschitz on BZ and the theorem is
proved.
Using a standard existence argument, for each (u0,v0)
in BZ there exists an c > 0 and a function z:[0,e) BZ such
that z*(t) • -Q(z{t)} and z(0) » (u0,v0). In many numerical
experiments using steepest descent and an initial pair
(w,y), evidence of the global existence of z has been
observed. This has been the case even for (w,y) with
nonplanar boundary conditions. A more complete discussion
of these results will be given in Chapter IV. The global
existence of trajectories shall be part of the hypothesis of
the following theorem.
THEOREM 3.4 Suppose {u0#v0) = <j>0 € BZ and z:[0,oo)->BZ such
that for each t € [0,oo) z' (t) *» -Q(z(t)) and z(0) « <J)0.
Then,
(i) There exists an increasing, unbounded sequence
{11 >.00 C [0,co) such that lim ||Q(z(t{ )) || • 0,
1 - 1 i-to H Q XHQ
(ii) If Eange(z) is bounded in then there is a
46
sequence {t.-} 00 satisfying part (i) such that 1 i • 1
{z(tj )}.""^converges strongly in C^XCQ to a point <{> in
and
(iii) If (J> from part (ii) is in BZ, then Vl{<|>) « 0.
PROOF: Suppose the hypothesis of the theorem is true. If
there is a t0 € [0,oo) such that ]|Q(z( t0)) II • 0, then
for each t > t0, z'(t) = 0 and z(t) « z(t0). In this case
the theorem in clearly true. Therefore, suppose that
||Q{z(t0))|| > 0 for each t € [0,a>). H£XH£
Part (i): Let f:[0,oo)-»IR such that f{t) » I(z{t)).
Differentiating, it follows that
f'(t) « I1(z(t))z«(t) » <Q(z(t)),z»{t))H^XH^
« o X i i o
Thus, f is a decreasing function. Note also that f is
bounded below by 0 and — f • {t) > 0 for each t € [0,oo) .
Assume - f is bounded away from zero by some 8 > 0. Then
for each t € [0,oo)# f'{t) < -S and f(t) < -£t+f(0). But
then there must be an s € [ 0 , j such that f(s) = 0 . The
fact that f(t) > 0 for each t implies f(t) » 0 for each
t > s. Thus, f' (s+1) = 0. But f' (s+1) • -l|Q(z(s+l) ) II2
< 0. This is a contradiction and so - f' is not bounded
47
away from 0. Since - f'(t) > 0 for each t € [0,co), an
CO m
increasing unbounded sequence c a n be constructed
such that lim - f'ftj) » lim ||Q(z{ (ti )) II2 exists and is
i-to i-ta H^XHJ
0, which is the conclusion of part (i).
Part (ii): Suppose the Range(z) is bounded in
There exists a bounded sequence in H$xH$ satisfying part
(i). By a well known property of Hilbert Spaces [8, p. 89],
there exists a subsequence which converges weakly to an
element <|> in HjjxHjj. Since this sequence is also bounded in
and H$xHjj is compactly embedded in C^xc^, there exists
a subsequence {<j>j} of which converges strongly in CqXC$
to an element $ of C$xc^. Since «|>i) converges weakly in
H^XH| to <|>, it converges weakly to <f> in H^XH^. But, since
{<J>i > converges strongly to $ in C^XCq, it converges weakly
to $ in CQXCQ and hence it converges weakly to $ in H^xH^.
Thus, {(j } converges weakly to both q> and <p in HlxHl,
implying that $ = <J>. Therefore, {{j } converges strongly in
C^xc^ to an element (|> € which is the conclusion of
part (ii).
Part (iii): Suppose «j)i } • {z(tj)} = {(u1 ,v' )} is the
sequence generated in part (ii) and <|> » (u,v) € BZ. The
strategy is to show that Q{<|>) is the weak limit of the
sequence {Q(z(t-)> « {(Mu^v*)} in H^xHq. Let e > 0.
For each g € , ||g|| = Su^ j |gi (x) J , Jg2 (x) Jj. Define
48
F:IR5-*|? by F(Xj ,x2 ,x3 ,x4 ,x5 ) = (XJXJ-X2X4) 2x5. Since
{(u^v1)) converges to (u,v) in C Q * C Q , there is a compact
interval S containing the ranges of u2+w2, Vj+yj,
*f»
v2+Y2- u"+wt, u®+w2, v"+yt and V2+y2 for each m 6 Z . Thus,
F is uniformly continuous on S5. Since F is uniformly
continuous on S8 and € > 0, there exists a > 0 such that
if |jum-u|| < #1 and ||v"-vj| < S i t then for each x 6 0
(3.11) |F(ut+wt Ua+WafVt+Yi,v2+y2,a)(x)
-F(u?+w1,u^+w2,v?+yi,vf+y2,b)(x)| <
where (a,b) is one of the pairs (v2+y2,v2+y2),
(vf+yj jVi+yi), {u2+w2 ,u2+w2) or (uf+Wj ,VLt +wt) . Note that
(3.11) is the same as
(3.12) |U2a)<x)-(Uffl)2b)<x)| < l l H
for each appropriate pair (a,b). Let $ be the minimum of 6x
and e .
Since <{> € BZ, {z(ti)) converges to <J> strongly in CqXC^ C
and ^ > 0, there exist positive numbers fi and n such that if
jm *4" m € Z and m > n, then
I-<»I - H S : ; : 3 3 H >'•
K " " i • (—< [ 3 ~ , • > —
49
l | [ - ) I L , s • • I I S - I I L , HqXHQ
Let (g,h) € H^xHq such that
m > n,
C^XCq
= 1. For each
Hi*H &
/ (uf+Wj f
<2 (uf+w2)
\ (vf+Yt) \ l(v^+y2); (ffm) m » 2
'{Ui+wj r (U2+W2) (Vt+Yi ) {V2+Y2)J
H^xh^
-(v?+y2)' +(v?+Yi) + (uf+w2) -(uf+Wi )J
- ( v ^ y ^ +(V1+Y1) +{u2+w2) l-(u1+w1 )J
HoxHQ
H$xh£
|2Jq(U"~Ui)9i+(u2-u2)g2+(vi-vi)hi+(v§-v2)h2
+ ^ [ l ^ F ( V 2 " Y 2 ) + ^ ( V2 +Y2)]si
+ ^[uST2- ( V ? _ Y i ) " 7*<Vi+Yi>]ff2
~~ Y2 ) - J7(v2+Y2)]hl
+ * + F^ Vl + yl )]h2 I
« 2 | < P : v ] •[?]),
t jj-|i!i2!i^}jji2aa»i||Bt|
50
+ J R I , 2 < N ^ : > + W ' ' I I H
+ + L H I I + I H S I
< » - F 1 + I ( J [ J I A ' I 2 + R + ^ I H I R + fa\h*\*]
Thus {Q(z(tm))} converges weakly to Q(<j>) « Q(u,v). Using
the fact that ||Q(U,V)|| < 11m inf {||Q(z(tm) )|| }, [8, HoxH^ H^XH£
p. 60], and lim J|Q(z(tm))It • 0, It follows that nHoo HqXHQ
| |Q(U,V)|| » 0. Thus, Q(u,v) » 0. By a result analogous
H£xh£
to lemma 1.3, this implies (Vl)(u,v) = 0. Thus, (u,v) = (j>
is a critical point of I.
CHAPTER BIBLIOGRAPHY
1. Adams, R., Sobolev Spaces, Academic Press, New York, 1975.
2. Agmon, S., Douglis, A. and Nirenberg, L., Estimates near the boundary for solutions of elliptic partial differential equations satisfying general boundary conditions, I, Comm. Pure Appl. Math., 12 (1959), 623-727.
3. Ball, J. M., Constitutive inequalities and existence theorems in nonlinear elastostatics, Nonlinear Analysis and Mechanics: Heriot-Watt Symposium 1 (1977), 187-241.
Singular Unimizers and their Signif icance in Elasticity, Directions in Partial Differential Equations, Academic Press, 1987, 1-16.
5. Evans, L. C. , fjuas iconvexi ty and partial regularity in the calculus of variations, (to appear).
6. Miranda, C., Partial differential Equations of Elliptic type, 2nd rev. ed., Springer Verlag, 1970.
7. Rabinowitz, P. H., Mini max Methods in Critical Point Theory with Applications to Differential Equations, Regional Conference Series in Mathematics, no. 65 (1984).
8. Riesz, F. and Sz.-Nagy, B., Functional Analysis , Ungar, New York, 1955
51
CHAPTER IV
NUMERICAL SOLUTIONS TO THE TWO DIMENSIONAL
PROBLEM FROM ELASTICITY
Motivation for most of the results in the first three
chapters comes from numerical experiments using a finite
dimensional version of the problem. In some respects, the
theorems of chapters II and III are an indication of the
reliability of the numerical method used. In this chapter,
an explanation of this method will be given, as well as
results of computer runs using several initial pair (w,y).
A self contained True Basic program for the two dimensional
setting is given in appendix A.
Let n be a positive integer such that n > 4, 0 be
[0,ljx[0,l] C R2 and $ - £. Let G «• |(i,j>: i,;j€{0,1,••,n}j
and G' » i, j€{l,2, • • ,n-l}|. Let H^n+l) denote the
2
space of functions from G to R<n+1> and L2(n~l) denote the
space of functions from G' to R<n~l) . If u € H^n+l), u{i,j) = u. . will represent the value of the corresponding
1, j
function u at the point {g-, } € 0, where i, j€{0,l, • • ,n>.
In order to define a norm for K1(n+1), it is necessary
to define two difference operators. Let D1 :H* (n+l)-+L2 (n-1)
such that for each u € H1(n+1)
52
53
(4.1) Dlu(i,j) » n f > i + 1'^ 2] U(i-l>j)
Let DaiH1 (n+l)-»L2 (n~l > such that for each u € H1(n+1)
(4.2) Dlu(i, j) = U{1'^+1)2S
The norm for Hl(n+1) shall be the Sobolev norm defined In
the following way: if u € H1(n+1) a H* then
(4.3) jjujl2 » S 2 \ t) [u( i, j) ]2 + £ [Dlu(i,j)]2
H1 "-j »J -o i ' j • i
+ [D2u(i,j)]2J.
The norm for L2(n-1) is the Euclidean norm giveen by
||u|!2 - S2 nS1[u(i,j)32. L2 (n-1) i,j.i
It is easy to see that each of L2(n-1) and H1 Is a finite
dimensional Kilbert space with the appropriate inner
product. A special subspace of H1 is H^(n+1) = j u € H1:
u(i,j) = 0 if i or j belongs to (o,n}|. Denote H$(n+1) by
Ho •
The problem to be solved is to minimize a functional
I: HQXHQ -• E. Suppose (wfy) € HlxB*. For each
(u,v) € H^xh^, define I by
n - l
(4.4) I(u,v) - S 2 ^ [tDllu+w)(i,j)]2
i # j « 1
54
+ [D2(u+w)(i,j)J2 + [Dl<v+y)<i,j>]2
,2 . 1 + [D2(v+y)( i, j ) ] + • f • I J #j H
where a (u v) - d t[Dl(u+w)(i,j) D2(u+w)(i,j)l . . wnere ar1:Jtu,vj
aet|Dl(v+y)(i,j) D2(v+y) (i, J)J " T n e o n I y
problem with the differentiability of I at a point (u,v) in
this finite dimensional setting is if ar^(u,v) • 0 for any
(i,j) € G', which would also imply that I is not defined at
(u,v). Let BZ be the subset of HqXHq such that (u,v) 6 BZ
if and only if ®^(u,v) > 0 for each (i,j) € G* . Again, BZ
is an open subset of H^xh^. It is easy to see that I is
Frechet differentiable on BZ. Performing this calculation,
one gets
I'(u,v)(g,h)
(4.5)
'Dlg^ 'DlU+Wj • - X i j D2v+yjj D29ij D2U4-Wj j + X ij Bfvfyj j
Dlhj, 9 2 Dlv+y4 j + X ij D2u+Wjj
D2hj j D2v+yj j - X H Dlu+Wj j
(4.6) *2 X <[SaI*j]• [S*;]),ir3
+ <&£i j ] •&!]>, i»j =i
IR2
1
|R2
where for each (i,j) €6', = jj—j2 a n d Ai j ' Bi j » cij
and Pij are the obvious values from the right side of (4.5)
depending on (u,v). Let Auv - {A^}, Buv » {B^},
Cuv = {C£ j } and P u v = <E*£ j ), noting that each of these is a
55
member of L2(n-1).
Let B be the linear transformation from H1<n+1) to the
space K « L2(nfl)xL2(n-l)xL2(n-1) such that B { u ) is the
triple (u,Dlu,D2u). Let P be the orthogonal projection of K
onto its subspace K0 consisting of elements of the form
(g,Dlg,D2g). Note that K0 is isomorphic to HQ . Then (4.6)
becomes
(4.7)
I'(ufv)(g,h)
+ <
+ I Bh,P
0 " \ uv ) uv
U U
0 "
p Cuv i F
uv
Thus, the gradient
' " 0 " ' 0
t p Auv , * / Cuv
1 LV
l Fuv ^(H^XH^)1]!?]
Denote this gradient by (?HI)(u,v).
The strategy for the numerical scheme is to
investigate the differential equation
56
(4.8) z'(t) - -(VHI)(z(t)) and z(0) - (u0,v0)
for (u0,v0) € BZ and z: [0,ooHH^xH$ . Following the notation
used by Neuberger in [5], let E = 3r0fc/>i where sr0 is the
H
orthogonal projection of H1 onto . E is non-negative and
invertible. The following theorem gives an explicit
formula for (VHI)(u,v).
THEOREM 4.1 Suppose (u,v) € H^xH$. Then P « DE~li0Dt .
The proof involves showing that P satisfies the necessary
and sufficient conditions for an orthogonal projection,
that is, P 2 = P, /'(u,Dlu,D2u) = (u,Dlu,D2u) for
(u,Dlu,D2u) 6 K0, P * • P and the R(/) C K0. A careful
examination of P reveals that
(4.9)
and
(4.10)
* P i
uv B
JT P 1
UV
UV
UV B
P. uv
* E~XJ (yBt
uv
UV F_ uv
An equivalent norm for the subspace HqXH^ is given by
the following: for each (u,v) € HqXH* o~no
(4.11) n - 1
||(u,v)||2 - 62 E [Dlu( i, j)J 2 + [D2u( i, j) ]2
HqxH^ i t j •> i
57
+ 62 "S^Dlvd, j) ]2+ [D2v(i,j) ]2 . i ' ;j * 1
Consequently, I is coercive and as well as bounded below on
BZ. Although I is not defined on all of HqXH^, in
particular the boundary of BZ, steepest descent avoids
these places because I is lower semi-continuous [2] and
blows up near boundary points of BZ. Prom the theory of
ordinary differential equations arguments, there exists a
positive number S and a function z : [ 0 , £ s u c h that z
satisfies (4.8). Using these ideas and an argument similar
to that of lemma 3 from [4, p. 191], the following result
can be proved.
THEOREM 4.2 Suppose (u0,v0) G BZ. There exists a solution
z to (4.8) such that z: [O,oo)-*BZ.
A key to this argument is that I(z) is a decreasing
function for every such z. This means that I has some
properties of a Lyapunov function.
The problem of looking for the asymptotic limit of z
is reformulated numerically as constructing the sequence
(4.12) Zn + 1 - Zn - hn-(VHI)(Zn) n » 0,1,2,3,•••
or equivalently, using Zn • (UB«Vn) € H$xH$ and equalities
(4.9) and (4.10),
58
(4.13) Un + 1 » U„ - hn-[Fn]
and
(4.14) Vn+1 - Vn - hn.[Q„]
where for each n
F - E-'Ioi>«<0,A^ ,B ) n u n v n u n n
and
G,, - ,F ) U I I
v n n v n
and {hj} Is an sequence of positive time steps which must
carefully selected.
The program in Appendix A is divided into six
sections. Section 1 defines (W,Y), the initial (U0,V0) and
the desired accuracy for several processes in the code.
Section 2 contains the Call statement for the subrountines
which compute Fn = E" ifo9t (°»A
n v rBf| ) and Gn «• u n v n "n v n
E"l5r0fc {0,Cn v ,F„ v ). The description of Section 6 will
u n n n n
give more details on each subroutine. Section 3 contains
the step defining the new pair (Un+i»Vn+1) using (4.13) and
(4.14). It also provides an opportunity to monitor the
choice of the timestep for the first three steps in the
main process. This section calls a subroutine to check the
convexity of I at (Un,Vn), using the quantity
59
)(Zn)]. Section 4 implements the
graphic subroutines which help monitor the process.
Section 5 prints the approximate solution whenever the error
tolerances have been achieved. This section contains the
END statement.
Section 6 contains 4 subroutines, three of which are
used in computing Fn and G0. The subroutine "ROUGH_GRAD"
computes (0,A ,B ) and (0,C ,F_. .. ). The °n vn un va un vn ua vn
subroutine ,,TRANS_D" computes Bt(0,A.. v ,BtI „ ) and Uu vn un v«
St{0,Cn fFn ). The subroutine "INVJDTD" computes an un vn un vn
approximation to E" 1jr0i't (0,A. „ ,B. „ ) and
un vn un vn
E" (0,CTT u ,Fn v ). This is done using the Qauss-Siedel un "n °n vn
iteration technique (see Dahlquist C7, pp. 1016-1026]). In
this technique, E is decomposed into the sum of three
matrices, a strictly lower triangular matrix L, a diagonal
matrix M and the strictly upper triangular matrix Lfc.
Letting ? • ^ ( O ^ , „ ,Bri v ), the iterative scheme is un vn un vn
generally given by
(4.15) M(Xk+1) - M(Xk) - 7 - E(Xk)
and
(4.16) x k + 1 = Xk + M~1[^ - E(Xk)]
with an updating occuring after the calculation of each
component of Xk and (Xk) converges to x where E(^) =» 9. The
60
other component of (V^I)(Un+1,Vn+1) is calculated In the
same way.
Table I contains the graphic output of two runs of the
program in Appendix A. Each pair of pictures depicts z(tk)
for some k and the value of the norm of the gradient at that
point. The first pair of initial functions (W,Y) is a
discretization of a pair (w,y) that satisfies the hypotheses
in Chapter III and are given by w{S,T) » S2+S-T and
y(S,T) .» S+T+T2. Using the initial (U0,V0) " (0,0),
Figure 1 on page 62 is the initial pair (W,Y). Letting
lim z(t) « (Uj,Vj}, Figure 2 is the approximation of the
pair (W+U^Y+Vj). Figures 3 and 4 on page 63 are the middle
cross sections of functions W and W+Uj, respectively.
Figures 5 and 6 on page 64 are the middle cross sections of
the functions Y and Y+V$, respectively. Notice that the
graphs of W+Uj and Y+Vj indicate that the critical point is
not affine. The square of the norm of the gradient of this
approximate solution is 7.212 x 10"5.
The second pair of initial functions considered is a
discretization of a pair (w,y) that does not satisfy the
hypotheses of Chapter III. The trajectory z in this case
can be observed to crack and the lim z(t) = (U2,V2) belongs
to but not HfjxHjj. The formulas for this second pair
(w,y) are w(S,T) • S-S3 and y(S,T) • T. Again using the
initial pair (U0,V0) « (0,0), Figure 7 on page 65 is the
61
initial pair (W,Y). Figure 8 is the deformation of (W,Y)
after 15 time steps (k « 15). Figure 9 is the approximation
of the pair (W+U2fY+V2), for which k » 100. Figures 10, 11,
12 on page 66 are the corresponding middle cross sections of
the functions W, its deformation at step k = 15 and W+Ut,
respectively. The square of the norm of the gradient of
this approximate solution is 5.8:13 x 10~5.
62
Fig. 1—Graph of initial pair W and Y
Fig. 2—Graph of approximate solution pair W+Ut and Y+Vt
63
Pig. 3—Cross sections of initial W, W(I,5) and W(5,J)
Fig. 4—Cross sections of solutions (W+Uj)(I,5)
and (W+U,)(5,J)
64
Fig. 5—Cross sections of initial Y, Y(5,J) and Y(If5)
Fig. 6—Cross sections of solutions (Y+V!)(5,J)
and (Y+Vj)(I,5)
65
Fig. 7—Initial pair W and Y
Fig. 8—Deformations of W and Y at k = 15
Fig. 9—Graphs of approximate solutions W+U2 and Y+V2
66
Fig. 10—Cross section W(I,5)
Pig. 11—Cross section of intermediate stage: (W+U)(I,5)
Fig.12—Cross section of solution: (W+U2)(I,5) at k = 100
CHAPTER BIBLIOGRAPHY
1. Adams, R., Sobolev Spaces, Academic Press, New York, 1975.
2. Ball, J. M., Constitutive inequalities and existence theorems in nonlinear elastostaties, Nonlinear Analysis and Mechanics: Herlot-Watt Symposium 1 (1977), 187-241.
3. Glowinski, R., Keller, H. and Reinhart, L., Continuation Conjugate Gradient Methods for the least Square Solution of Nonlinear Boundary Value Problems, Rapports de Recherche, Institue National de Recherche en Informatique et en Automatique, 1982.
4. Neuberger, J. W., Steepest descent and differential equations, J. Math. Soc. Japan, Vol. 37, No. 2, 1985
5» » Steepest Descent for General Systems of Linear Differential Squat ions in iilbert Space, Proceedings of a Symposium held at Dundee, Scotland March-July 1982, Lecture Notes in Mathematics, Vol. 1032, New York, Springer-Verlag, 1983, 390-406.
6. , Steepest Descent for Systems of Honlinear Partial Differential Equations, Oak Ridge National Laboratory CSD/TM-161, (1981).
7. Young, D. M. and Gregory, R. T., A Survey of Numerical M a t h e m a t i c s , Yol. II, Addison-Wesley Publ. Co., 1973.
67
APPENDIX
APPENDIX
! PROGRAM NAME: 2D ELST ! ELASTICITY - PHANTOM GRID,ITERATIVE LINEAR SOLVER, GRAPHICS ! J(U,V)=* INTEGRAL [ ||D(U,V)||~2 - 1/(DET{U+W,V+Y)) J
SECTION 1
LIBRARY "3DLIB" PRINT "WHAT SIZE GRID? N=10" LET N=10 PRINT "WHAT ACCURACY FOR SOLUTION? El=.OOOOOOOl" LET El=.00000001 PRINT "WHAT ACCURACY FOR GRADIENT? E2=.00001" LET E2=.00001 PRINT "WHAT MESH IN T-DIRECTION"; INPUT H LET K=1 DIM U(0 TO 10,0 TO 10),V{0 TO 10,0 TO 10),W(0 TO 10,0 TO 10) DIM Y(0 TO 10,0 TO 10) DIM X1(0 TO 10,0 TO 10),DT_DU(9,9),RG1{10,10),RG2(10,10) DIM RG3(10,10),RG4(10,10),X2(0 TO 10,0 TO 10) LET S=N/2 FOR I«1 TO N-l
FOR J=1 TO N-l LET U(I,J)-0 LET V(I,J)=0
NEXT J NEXT I FOR I»0 TO N
FOR J=0 TO N LET W(I,J)«(I/N)"2+(I/N)-(J/N) LET Y(I,J)»(I/N)++(J/N)+(J/N)"2 ! LET W(I,J)«={I/N)-{I/N)~3 ! LET Y(I,J)°(J/N)
NEXT J NEXT I LET ERR1=-1 LET ERR2—1
SECTION 2
DO UNTIL MAX{ABS(ERR1),ABS{ERR2))<E1 CALL ROUGH_GRAD(N,S,U(,),V(,),W{,)rY(,),
RG1(,),RG2{,),RG3{,),RG4(,))
69
70
! COMPUTE FIRST COMPONENT OF ROUGH GRADIENT CALL TRANS_D(N,S,RG1(,),RG2(f),DT_DU{,)) ! COMPUTE FIRST COMPONENT OF SMOOTH GRADIENT LET ERROR®1 MAT X1»0 ! RESET X1=0 CALL INV_DTD(N,S,DT_DU(,),X1(,),E2) LET ERR1—1 LET GRAD1=0
! COMPUTE SECOND COMPONENT OF ROUGH GRADIENT CALL TRANS_D(N,S,RG3(,),RG4(,),DT_DU(,)) ! COMPUTE SECOND COMPONENT OF SMOOTH GRADIENT LET ERROR** 1 MAT X2=0 ! RESET X2=0 CALL INV_DTD(N,S,DT_DU(,),X2(,),E2) LET ERR2=-1 LET GRAD2-0
SECTION 3
! SELECT TIMESTEP IN RELATIONSHIP TO SIZE OF GRADIENT SELECT CASE K CASE 1,2 ! CHECK SIZE OF GRADIENT
PRINT "K-";K,"GRADIENT OF FIRST VECTOR - START" MAT PRINT XI PRINT "K«";K,"GRADIENT OF FIRST VECTOR - FINISH" PRINT "K»";K,"GRADIENT OF SECOND VECTOR - START" MAT PRINT X2 PRINT "K=";K,"GRADIENT OF SECOND VECTOR - FINISH" ! CHECK CONVEXITY OF I AT (U(K),V(K)) CALL CONVEX(XI{,),X2(,)fU(,),W(,),V(,),Y(,),N,S) PRINT "WHAT INITIAL TIME STEP"; INPUT H3 FOR 1=1 TO N-l
FOR J=1 TO N-l LET U(I,J)»U(I,J)-H3*X1(I,J) LET T»X1(I,J) LET ERR1=MAX(ERR1,ABS(T)) LET V(I,J)=V(I,J)-H3*X2(I,J) LET T=X2(I,J) LET ERR2-MAX{ERR2,ABS(T))
NEXT J NEXT I
CASE 3 PRINT "WHAT TIME STEP AT STEP 3"; INPUT H
CASE ELSE CALL CONVEX(XI{,),X2(,),U(,),W(,),V(,),¥(,) ,N,S) FOR 1=1 TO N-l
FOR J=1 TO N-l LET T=X1(I,J)
71
LET ERRl=MAX(ERR1,ABS(T)) LET T«X2(I, J) LET ERR2=MAX(ERR2,ABS(T))
NEXT J NEXT I LET TIMESTEP=MIN(1/ERR2,1/ERR1) LET H1-MIN(TIMESTEP».001,H) FOR 1=1 TO N-l
FOR J«1 TO N-l LET U (I, J)«U(I,J)-HI*X1(I,J) LET V(I,J)»V(I,J)-HI*X2(I, J)
NEXT J NEXT I
END SELECT
SECTION 4
! PLOT GRAPHS OF U{K)+W AND V(K)+Y SET MODE "HIRES" CALL PARAWINDOW (0, 1, 0, 1, 0, .5, WORK$) SETCAMERA (1, -2, 2) FOR 1-0 TO 10
FOR J=0 TO 10 CALL PL0T0N3 (1/10, J/10, W (I,J)+U(I,J), WORK$)
NEXT J PLOT
NEXT I
CALL PARAWINDOW (0, 1, 0, 1, -.5, 1, WORK$) FOR I»0 TO 10
FOR J=0 TO 10 CALL PL0T0N3 (1/10, J/10, Y(I,J)+V(I,J), WORK$)
NEXT J PLOT
NEXT I
! PRINT OUT U(K)+W AND V(K)+Y IN NUMERICAL TABLE IF INT<K/5)«K/5 THEN
PRINT "0(1,J) IS THE FOLLOWING AT STEP =";K FOR J»N TO 0 STEP -1
FOR 1=0 TO N STEP 1 PRINT USING "###.####":U(I,J)+W<I,J);
NEXT I PRINT
NEXT J PRINT " " PRINT PRINT "V(I,J) IS THE FOLLOWING AT STEP =";K FOR J=N TO O STEP -1
FOR 1=0 TO N STEP 1 PRINT USING "###.####":V(I,J)+Y(I,J);
72
#
NEXT I PRINT
NEXT J PRINT " "
END IF PRINT "STEP = ";K;"MAX GRAD1";ERR1;"MAX GRAD2";ERR2 LET K=K+1
LOOP
i SECTION 5
! PRINT THE SOLUTION PAIR (U+W,V+Y) PRINT PRINT "U(I,J) IS THE FOLLOWING" FOR J=»N TO 0 STEP -1
FOR I»0 TO N STEP 1 PRINT USING "###.####":U(I,J)+W(I,J);
NEXT I PRINT
NEXT J PRINT " " PRINT PRINT "V(I,J) IS THE FOLLOWING" FOR J-N TO 0 STEP -1
FOR 1=0 TO N STEP 1 PRINT USING "###.####":V(I,J)+Y(I,J);
NEXT I PRINT
NEXT J PRINT " PRINT END
I SECTION 6
SUB ROUGH_GRAD (N, S, U(,), V(,), W(,), Y{,), RG1(,), RG2<,),
RG3(,), RG4(,)) FOR 1-1 TO N
FOR J»1 TO N ! A=U1+W1 C=V1+Y1 B=U2+W2 D=V2+Y2 LET A»S*(U(IfJ-1)+W(I,J-l)-0(1-1,J-1)-W(I-1,J-l)) LET A-A+S*(U(I,J)+W(l,j>-U(l-I,jy-W(l-I,j)) LET B™S*(-U(I,J-l)-W(I,J-l)-U(1-1,J-l)-W{1-1,J-l)) LET B«B+S*(U(I,J)+W(I,J)+U(I-1,J)+W(I-1,J)) LET C"S*(V(I,J-l)+Y(I,J-l)-V(1-1,J-l)—Y(1-1,J—1)) LET C^C+S*(V(I,J)+Y(I,J)-V(1-1,J)-Y(1-1,J)) LET D"S*(—V(I,J-l)-Y(I,J-l)-V{1-1,J-l)-Y(1-1,J-l)) LET D»D+S*(V(IfJ)+Y(I,J)+V(I-1,J)+Y(I-1,J)) LET DET3«(A*D-B»C)"2
73
LET RG1(I,J)»2*(A-{D/DET3)) LET RG2(X,J)=2 *(B-(C/DET3)) LET RG3(I,J)*2 *(C-(B/DET3)) LET RG4(I,J)=2*(D-(A/DET3))
NEXT J NEXT I
END SUB
SUB CONVEX (XI( , ) , X2(,), U{,)f W(,), V( ,} , Y(,), N, S) LET DET2-1E+6 LET SUM1=0 LET SUM2=0 LET SUM3=0 LET SUM4-0 FOR 1=1 TO N
FOR J=1 TO N ! A-U1+W1 C=V1+Y1 B=U2+W2 D*=V2+Y2 I D1X1=FIRST PARTIAL OF GRAD OF XI ,ETC. LET A^S*(U(I,J-l)+W(I,J-l)-U{1-1»J-l)-W{1-1,J-l)) LET A=A+S*(U{I,J)+W{I,J)-U{I-1,J)-W{1-1,J)) LET B=S*(-U(I,J-l)-W(I,J-l)-U(1-1, J-l)-W(1-1, J-l)) LET B™B+S*(U(I,J)+W{I,J)+U{1-1,J)+W(1-1,J)) LET C=S*(V{I,J-l)+Y(I,J-l)-V(1-1,J-l)-Y(X-l, J-l)) LET C=C+S*(V(I,J)+Y(I,J)-V(1-1,J)-Y(X-l, J)) LET D*S*(-V(I,J-l)-Y(I,J-l)-V(1-1,J-l)—Y(1-1,J-l)) LET D»D+S*(V(I,J)+Y(I,J)+V(1-1,J)+Y(1-1,J)) LET DET1=(A*D-B*C) ! PRINT I ; J ; ; DET1, LET DIXIES*{XI(I,J-l)-XI(I—1,J-l)+X1(I,J)-XI{1-1,J)) LET D2X1»S*(X1(I,J)-X1(I,J-1)+X1(I-1,J)-X1(I-1,J-1>) LET D1X2-S*(X2(I,J-l)-X2(1-1,J-l)+X2(I,J)-X2(1-1,J)) LET D2X2a»S* (X2 (I, J) -X2{I, J-l)+X2(1-1, J) -X2 (1-1, J-l > ) LET DETGRAD«D1X1*D2X2-D2X1*D1X2 ! PRINT M DETGRAD( " ; I; J; " ) * " ;DETGRAD LET
SUM1=SUM1+(2/DET1"3)*(D1X1*D-D2X1*C+A*D2X2-B*D1X2)~ 2 LET SUM2»SUH2+{2/DET1"2)*DETGRAD LET SUM3=SUM3+(A"2+B"2+C *2+D * 2)+1/DET1~ 2 LET SUM4«SUM4+(D1X1)~2+(D2X1)~2+(D1X2)*2+(D2X2)*2 LET DET2-MIN(DET1,DET2)
NEXT J NEXT I PRINT PRINT "MIN DET =";DET2 PRINT "SUM1 • ";SUM1; PRINT "SUM2 =";SUM2 PRINT "NORM OF GRAD";SUM4 PRINT "I«'(U{K),V(K)>(GRAD,GRAD) - ";2*SUM4+SUM1-SUM2 PRINT "I(U(K), V ( K)) -";SUM3
END SUB
74
SUB TRANSJD (N, S, A(,), B(,), DT_DU(,)) ! D TRANSPOSE WITH ZERO BOUNDARY CONDITION ! THIS IS THE ROUGH GRADIENT FOR I«1 TO N-l FOR J=1 TO N-l LET T=S*(A(I,J)+A(I,J+1)-A(I+1,J)-A(I+1,J+l)) LET
DT_DU(I,j)=T+S*(B(I,J)-B(I,J+1)+B(I+1,J)-B(I+1,J+l)) NEXT J
NEXT I END SUB
SUB INV_DTD (N, S, DT_DU{,), X( , ) , E2)
! (DTD) • L + M + N ! M(X(K)-X(K-1)) « Y - (DTD)(X(K-l)) : ITERATION TO SOLVE
LINEAR SYSTEM
! A « X(K) - INV(M)[ X(K-l) + Y - DTD(X(K-1)] ! - INV(M)[ Y - (L+N)(X(K-1))] ! X(I,J) = THE GRAD OF U(I,J) OR V(I,J) LET L=1 LET E=10 DO UNTIL E<E2 LET ERROR-O IF INT(L/2)<L/2 THEN FOR 1=1 TO N-l FOR J=1 TO N-l LET A«DT_DU(I,J)+S*S*2*(X(I-1,J-1)+X(I-1,J+l)) LET A»(A+S*S*2*(X(I+1,J-l)+X(1+1,J+l)))/(8*S*S) LET B«A-X(I,J) LET X(I,J)-A LET ERROR^NAX(ERROR,ABS(B))
NEXT J NEXT I LET E=ERROR
ELSE FOR I=N-1 TO 1 STEP -1 FOR J-N-l TO 1 STEP -1 LET A=DT_DU(I,J)+S*S*2*(X(I-1,J-l)+X(1-1,J+l)) LET A»(A+S*S*2*(X(I+1,J-l)+X{1+1,J+l)))/(8*S*S) LET B=A-X(I,J) LET X(I,J)»A LET ERROR'MAX(ERROR,ABS(B))
NEXT J NEXT I LET E'ERROR
END IF LET L-L+l
LOOP END SUB
BIBLIOGRAPHY
Books
Adams, R., Sobolev Spaces, Academic Press, New York, 1975.
Brauer, P. and Nohel, J., Ordinary d i f f e r e n t i a l Equations, Preliminary Edition, W. A. Benjamin, Inc., New York, 1966.
Brezis, H. , (fperateurs maximaux monotones et seai-groupes aonl in tarts, Math. Studies 5, North Holland, 1973.
Miranda, C., Partial Differential Equations of Elliptic Type, 2nd rev. ed., Springer Verlag, 1970.
Rabinowitz, P. H., Minimax Methods in Critical Point Theory with Applications to Differential Equations, Regional Conference Series in Mathematics, no. 65 (1984).
Riesz, F. and Sz.-Nagy, B., Functional Analysis , Ungar, New York, 1955
Young, D. M. and Gregory, R. T., A Survey of Jfuaerical Mathematics, Vol. I I , Addison-Wesley Publ. Co., 1973.
Articles
Agmon, S., Douglis, A. and Nirenberg, L., Estimates near the boundary for solutions of e l l i p t i c partial d i f f e r e n t i a l equations satisfying general boundary conditions, I, Comm. Pure Appl. Math., 12 (1959), 623-727.
Antman, S . S . , Mono t on i c i t y and invertibility conditions in one-diaensional nonlinear elasticity, Nonlinear Elasticity, ed. R. W. Dickey, Academic Press, New York, 1973.
Ball, J. M., Constitutive inequalities and existence theorems in nonlinear elastostatics , Nonlinear Analysis and Mechanics: Heriot-Watt Symposium 1 (1977), 187-241.
Singular Minimizers and their Significance in El asticity, Directions in Partial Differential Equations, Academic Press, 1987, 1-16.
75
76
Curry, H. B., The method of steepest descent for nonlinear minimization p r o b l e m s . Quart. Appl. Math., 2 (1944), 258-261.
Evans, L. C. , fjuasiconvexity and partial regularity in the calculus of variations, (to appear).
Neuberger, J. W., Gradient Inequalities and Steepest Descent for Systems of Partial Differential Equations, Proceedings of the Conference on Numerical Analysis, Texas Tech University, 1983.
, Some Global Steepest Descent Mesults for Jfonl inear Systems, Trends in Theory and Practice of Nonlinear Differential Equations, edited by V. Lakshmikantham, New York, Marcel Dekker, 1984, 413-418.
Steepest Descent for General Systems of Linear Differential Squat ions in Milbert Space, Proceedings of a Symposium held at Dundee, Scotland March-July 1982, Lecture Notes in Mathematics, Vol. 1032, New York, Springer-Verlag, 1983, 390-406.
Steepest Descent for Systems of Jfonlinear Partial Differential Equations, Oak Ridge National Laboratory CSD/TM-161, (1981)
Steepest descent and differential equations, J. Math. Soc. Japan, Vol. 37, No. 2, 1985
Richardson, W. B., Jfonl inear Boundary Conditions in Sobole\ Spaces, Dissertation, North Texas State University, 1984.
Rosenbloom, P. C., Ike method of steepest descent, Proc. Sympos. Appl. Math. VI, Amer. Math. Soc., 1956, 127-176.
Reports
Glowinski, R., Keller, H. and Reinhart, L., Continuation Conjugate Gradient Methods for the Least Square Solution of Jfonl inear Boundary Value Problems, Rapports de Recherche, Institue National de Recherche en Informatique et en Automatique, 1982.