Professional Documents
Culture Documents
rel Rektorys
The Method of
Discretization in Time
and Partial Differential Equations
The Method of
Discretization in Time
and Partial DifFerential Equations
D. R E I D E L P U B L I S H I N G COMPANY
Dordrecht: Holland/Boston : U.S.A./London : England
Library of Congress Cataloging in Publication Data
Rektorys, Karel.
The method of discretization in time and partial differential
CE
equations.
(Mathematics and its applications. East European series; v. 4).
Translated from the Czech.
Bibliography: p.
Includes index.
1. Boundary value problems — Numerical solutions. 2. Differential equations,
Parabolic — Numerical solutions. 3. Evohition equations — Numerical solutions.
I. Title, n . Series: Mathematics and its applications (D. Reidel Publishing Company).
III. Series: Mathematics and its applications (D. Reidel Publishing Company). East
European Series; v. 4.
QA 379. R 4413 1982 515.3'5 82-7658
ISBN 90-277-1342-1 (Reidel) AACR 2
Editor's Preface IX
Preface XI
Notation Frequently Used XV
Chapter 1. Introduction 1
Chapter 2. A Brief Survey of Variational Methods 8
Part I. EXAMPLES 66
Chapter 3. Algorithm of the Method. Error Estimates 68
a) Algorithm of the method 68
b) Error estimates 81
Chapter 4. Linear Parabolic Problems 85
a) Illustrative examples 86
a) Homogeneous initial conditions 86
P) Nonhomogeneous initial conditions 94
b) A nontrivial example 107
c) An equation with time-dependent coefficients 118
Chapter 5. A Nonlinear Problem 127
Chapter 6. An IntegrodiflFerential Parabolic Problem 137
Chapter 7. A Problem with an Integral Condition 147
Chapter 8. Hyperbolic Problems 157
Chapter 9. An Application to Rheology 170
Part II. THEORETICAL ASPECTS OF THE METHOD OF DISCRE-
TIZATION IN TIME
Chapter 10. Introduction to Part II 185
Chapter 11. The Equation dujdt + Au = f. Existence and Conver-
gence Theorem 187
Chapter 12. Regularity Questions. Error Estimates 219
a) Smoother solutions 219
VII
VIII Contents
IX
X Editor's Preface
For details see [29] and Chapter 1 of the present book. His idea was
adopted in a large number of later works which treated, especially, more
general parabolic problems of the second order. See, in particular, the
comprehensive work [20] by O. A. Ladyzenskaja, the paper [21] by Iljin,
Kalasnikov and Oleinik, etc. In these treatises, classical methods or methods
based on fixed point theorems of the Schauder type have been used, as
usual. A basic work applying the Rothe method to equations of higher than
of the second order is the paper [22] by O. A. Ladyzenskaja (see also the
paper [23] by S. I. Ibragimov, etc.).
A rather new technique of treating the parabolic problems of higher orders
in this way was shown in my paper [1], pubhshed in 1971. It yields, in
a relatively very simple way, apriori estimates on the base of which existence
and convergence questions can be easily answered, including those which
arise when elliptic problems, generated by the Rothe method, are solved
approximately. This technique was followed in other texts — see e.g. J.
Necas [13], J. Kacur [14], [15], [17], [18], J. Kacur, A.Wawruch [16], etc. -
XI
XII Preface
and, in particular, served as the base for the intensive study of evolution
problems in my seminar held at the Technical University in Prague. The
developed method was shown to be applicable to the solution of a very
broad scale of problems (parabolic problems, including intergrodifferential
problems or problems with an integral condition describing complicated
processes in heat conduction, hyperboUc problems, problems in rheology,
etc.). Numerical aspects of this method have been examined, similarly as the
theoretical ones concerning existence and convergence, error estimates,
regularity properties of the solution, etc.
To point out that the technique of this method is in a way different
from the one used in the "classical" Rothe method (or method of lines,
as the Rothe method is frequently called) and that the range of the appli-
cability of this method is somewhat broader, we have given it a new nam.e:
The method of discretization in time.
The present book is devoted, in the first place, to results obtained by this
method in my seminar during the recent years. Most of these results are
mine. As far as some ideas due to my colleagues are utilized, the fact is
pointed out in the text (see, in particular. Chapters 15,17 and, to a certain
extent, 19, 20). Almost all of the presented results are published here for
the first time. In particular, the whole Part II of the book is practically new.
Our method is a numerical method. This fact makes it interesting for
mathematicians as well as for "consumers" of mathematics, especially
for engineers. The first part of the book is devoted to this second category
of readers. One finds here the routine of the method. Many numerical
examples are presented (with corresponding programs for convenience, if
useful). The classical Ritz method as well as the finite element method are
apphed to the approximate solution of elliptic problems obtained by way
of the application of the method of discretization in time. The efficiency
of error estimates, derived in Part II, is tested on practical examples. These
questions are of interest to a mathematician as well. On the other hand,
a reader interested in the practical aspects of the method only, can leave
the investigation concerning error estimates (which may seem rather lentghy
to him) aside; usually, approximations obtained_by the method of discretiza-
tion in time are quit£^ satisfactory even in cases^^^w^ derived error
estimates are riot.
For the convenience of the reader, basic concepts and results related
Preface XIII
XV
XVI Notation Frequently Used
Practically, these two assumptions are equivalent, because the way how
to verify them is the same.
The symbols Zj, or other symbols, as well as results, are often complet-
ed by "j = 1,..., p", etc. However, it is not the case everywhere, especially
where it is evident for which / s these symbols are considered or resuhs
valid.
Chapter 1
Introduction
z(x) - 2,-i(.x) „/ N
-''^^ J—L5^^ _ 2:"lx) = sin X (1.7)
be fulfilled f o r ; = I , . . . , 5.
Fig. 1.
with h = 0.2 stands here for the derivative dujdt in (1.1), while the boundary
conditions (1.3) are replaced by the conditions (1.8).
Replacing the derivative dujdt by (1.9), we establish a_^^discretization
of our problem in the direction of the time-axis". This gives the name
to our method.
For J = 1, (1.7) becomes
Zi(x) - Zo(x)
— z'i(x) = sin X ,
Introduction 3
we get
Ai{h + l)s'in X = h sin x ,
whence
Ai = = 1
I + h 1+ h
and finally
(The form
1- '
1 + b
has been chosen for Ai because of its usefulness to be seen later.)
Similarly, for 7 = 2, (1.7) yields
Z2(x) = A2 sin X
. . ( . + i) = * + ( i - ^ ) ,
h+I 1 , 1
A. = —• = 1
h+ I (1 + hf (1 + hf
and
Putting here h = 0.2 and rounding off to four decimals, we obtain finally
For comparison, the exact solution (1.4) gives for t = ti — 0.2, ...,t =
= t, = l:
u(x, /i) = 0.1813 sin x ,
u(x, ^2) = 0.3297 sin x ,
M(X, tj) = 0.4512 sin x ,
M(X, (4) = 0.5507 sin x ,
M(X, (5) = 0.6321 sin X . (1.19)
Introduction 5
The simple example, just presented, shows that the method of discretiza-
tion in time yields approximate solutions Zi(x), ..., z^(x)of the given prob-
lem at discrete values of the variable ( only. To get an approximation
in the whole domain Q, we may construct a function — let us denote it by
Mi(.x, t) — defined in the J-th subinterval
by
u,{x, t) = z,_,(x) + ^ ^ [z,(x) - zj..,{x)] . (1.20)
h
I For example, in the interval 12, the function «i(x, t) is of the form
estimate, are not all simple. For this reason, their investigation is postponed
to Part II of this book. For illustration, let us show here how do the appro-
ximations of the solution of the problem (l.l) —(1.3) at the points ti = 0.2,
^2 = 0.4 and fs = 1 become better if a finer division of the interval / into
100 subintervals of lengths h = 0.01 is chosen. We get (after rounding off
to four decimals):
at the point t = 0.2:
we write briefly
L M(X) dx .
the norm
\\ul = V(H,M)= /ru^(x)dx, (2.3)
the distance
e{u, v) = \\u -v\\= I [M(X) - v{x)Y dx . (2.4)
10 Chapter 2
R E M A R K 2.1. For two functions n(x), v{x) which need not be equal
at all points x e G, we may have
L
} G
[u{x) - v{x)Y dx = 0 .
According to (2.4), their distance in the space L2(G) is equal to zero. Such
functions are called equivalent. They diifer in G on a set of (Lebesgue) mea-
sure zero, or, as we say, they are equal almost everywhere in G. In parti-
cular, if two functions are different only at a finite number of points in G,
they are equivalent in LjfG).
All equivalent functions are considered as the same element of the space
For example, we consider as the zero function in L2(0, 1) not only the
function u(x) s 0 in [0, 1] but also every equivalent function, e.g. the
function v(x) defined in [0, 1] by v{x) = 0 in (O, 1] and v{x) = 1 for x = 0.
Having in mind the concept of equivalent functions, we can state the
following properties of the scalar product, norm and distance in the space
2,2(0) (u[x), v(x), r(x), M,(x), U2{x) are arbitrary functions from L2{G),
flj, 02 are arbitrary real numbers):
Scalar product:
{u, v) = (r, u) , (2.5)
(ajM, + fl2"2» v) = ai(ui, v) + 02(112, v), (2.6)
(u, M) ^ 0 , (2.7)
(t,, «) = 0 => M = 0 in L2{G)}) (2.8)
Norm:
|M| > 0 (2.9)
l"l = 0^ M == 0 in L2(G), (2.10)
I aw1 '= 1^1 I t^'li ' (2.11)
|(", 4 — u||'1 I t;'II (2.12)
jlw + v\ < II I<'II + (the triangle inequality), (2-13)
H - llrii < II" (2.14)
' ) By tt = 0 in L2(G) we denote the zero function in this space, thus (see Remark 2.1)
the function u(x) = 0 or an equivalent function. (Without the convention on equivalent
functions, (2.8) would not be true.)
Variational Metliods 11
Distance:
Q{U, C)^0, (2.15)
Q(II, U) = 0 <=> U = r in ^2(0), (2.16)
By (2.4), in the space Lj the so-called metric is defined, the space L2(G)
is an example of the so-called metric space ([3], Chap. 7). On this base,
we define convergence:
A sequence of functions «„ e L2(G) is said to converge in this space
(or to converge in the mean) to the function u e L2{G), and we write
M„ - u in L2(G), (2.21)
if
hm Q{U, U„) = 0 , (2.22)
or (see (2.4)) if
0 for
— < X ^ 7t
2n
(sketch the graphs of these functions) converges, in £2(0,7t), to the function
u(x) = 0, because
are called the Fourier coefficients, or the Fourier series of the function u(x)
with respect to the system (2.30), respectively.
The series (2.34) can be shown to be always convergent in the space
L 2 ( G ) . However, it need not converge to the function u{x). If the system
(2.30) is such that for every u e L2{G) the corresponding Fourier series
converges to M(X), this system is called complete.
An example of a complete orthonormal system in L2{0,1) is the system
of functions
^) A set M is called linear (in more detail: linear over the set of real numbers) if for
its elements the operations of the addition of two elements and of the multiplication
of an elemsnt by a real number are defined, with properties familiar from vector
algebra (of the form au + bu = (a + b) 11, etc., see [3], Chap. 6), while
u e M, V e M, a, b real numbers => au -{- bu e M.
Variational Methods 15
be assigned with properties (2.5) —(2.8).^) Let us define the norm and the
distance by
|Mi=V("'")' (2-43)
Q{U, V) - IIu - t;|| . (2-44)
•') For example, let M= C*'^ [0, 1]. This is a linear set provided the sum of two
functions and the multiplication of a function by a real number are defined in the usual
way. Let
.1 p
(«, v) = I iw dx -|- I u'v' dx . (2.41)
Jo Jo
Then all the properties (2.5) —(2.8) can be easily established, so that (2.41) is a scalar
product on C'^> [0, 1].
On the other hand, define
(u, v) = I u'v'
(u,v)= dx .
u'v'dx. (2.42)
Jo
(2.42) is not a scalar product on C*^' [0, 1]: The properties (2.5) —(2.7) are preserved,
however (2.8) is not valid, because
(«,«)= u'^
does not imply u(x) = 0, but only u'(x) s 0 and u{x) s const.
"*•) It need not be always the case. For example, it can be shown (see [3], Chap. 6)
that the space ^2 obtained when defining the scalar product (2.41) on the linear set
C<'* [0, 1] is not complete.
16 Chapter 2
' ) This assumption implies that the set D^ of all admissible functions (see (2.48)) is
a linear set, what makes all considerations much simpler. If the boundary conditions are
not homogeneous, we often succeed in constructing a sufficiently smooth function w(x)
satisfying these conditions. Then the solution of cur prcblem can be assumed in the
form
u = V + w, (2.46)
where the function v(x) has to satisfy the equation
Av = f- Aw (2.47)
with corresponding homogeneous boundary conditions.
On the construction of a function w^x) see [3], Chap. 11. See also [3], Chap. 46,
where the possibility of construction of such a function is discussed from the theo-
retical point of view.
On a more suitable formulation of the problem with nonhomcgeneous boundary
conditions see p. 52 ff.
Variational Methods 17
u^sin X
Fig. 2.
U'XfTC-x}
Fig. 3.
n X
Fig. 4.
of the conditions (2.50), however, it does not satisfy the first one. The
function sketched on Fig. 5 is not admissible, either: Indeed, it fulfils both
the boundary conditions (2.50), but it does not belong to C'^'[0, n] (not
even to C*^^[0, TT] because u'(x) is discontinuous at the point x = TI/2).
u=>i-x
Fig. 5.
i:
n
[u'v' + (1 + cos^ x) uv] dx . (2.61)
20 Chapter 2
P o s i t i v e n e s s : By (2.61), we have
I/O
.'2 dx = 0 => M'(X) = 0 => M(X) = const in [0,7t] .
u'^
we have
(Au, u) = 0 => «(x) = 0 ,
which proves (2.54). (2.52), (2.53) being fulfilled, the positiveness of the
operator A on the set D^ is established.
P o s i t i v e d e f i n i t e n e s s : We have obviously
REMARK 2.4. Naturally, from the just proved positive definiteness, posi-
tiveness follows immediately by Remark 2.3. Thus the proof of positiveness
was "useless". We have carried out this proof to show how to proceed if,
for example, the coefficient at u in the given equation is nonnegative (as
before) and at the same time, equal to zero at some points of the considered
interval, so that an inequality of the type (2.65) cannot be appHed. (This
occurs, for example, if the function 1 + cos^ x is replaced by the function
cos^ x). On how to prove positive definiteness in such cases, see [3], Exam-
ple 8.9.
Variational Methods 21
valid for functions g, h e C'-^\G) and for a bounded domain G with a Lip-
schitz boundary F. Here, v is the unit outward normal with components v,-.
(Thus V; are direction cosines of the outward normal.) Applying this formula,
we get (for details see [3], Chap. 8)
TABLE 2.1.
5- - L ^ r o - ^ + ^ =/' u = 0 on r
i , i = i OX; \ oXj/
N JV
REMARK 2.6. The problem (2.49), (2.50), p. 16, is a special case of Problem
1 from Table 1: We have a = 0, b = n, p{x) = 1, r(x) = 1 + cos^ x.
(Here, we can take po = !•)
The problem (2.66), (2.67) is a special case of Problem 5 for c(x) s 0,
a,j(x) = 1 for J = i, aij{x) = 0 for / 4= i. Here, we get
N N
I « o W a.-of; = E af >
This implies
H'(X) = 0 and M(X) = const. , (--^7)
as in Example 2.2 (see (2.64)), but nothing more, because boundary condi-
tions of the form (2.50) are not prescribed here. Thus [Au, u) = 0 implies
only w(x) s const., where this constant need not be equal to zero. Thus the
operator (2.74) is not on the set (2.75) positive.
The loss of positivity has curious consequences:
From (2.72), we get immediately
w'= - 2 x -I- C .
24 Chapter 2
3 2 6'
The following theorem is of fundamental importance:
) In this way, to every function u e D^ a real number (2.88) is assigned. (Note that
the scalar products of the functions/, « and Au, u are real numbers; the function/is
assumed to belong to £2(0}.) Such an operator is called a functional (defined on D^,
in our case).
26 Chapter 2
Proofs of both preceding theorems (for a more general case, when con-
sidering equation (2.87) in a Hilbert space) can be found in [3], Chap. 9.
In that chapter, a technical motivation of Theorem 2.2 with an example
is also presented.
The main sense of Theorem 2.2 consists in the fact that, on its base,
the problem of solving equation (2.87) on the set D^ of its admissible func-
tions is replaced by the problem of minimizing a functional on D^. For an
approximate solution of the latter problem effective methods are developed
(the Ritz method, etc.). However, note that Theorem 2.2 does not have the
character of an existence theorem. It asserts that if MQ ^ ^A is the solution
of equation (2.87), then MQ minimizes the functional Fu on D^, or if Fu
attains, on D^, its least value for an UQ, then this «<, is the solution of (2.87).
However nothing is said about the existence of a solution of the investigated
equantion or about the existence of a minimum of the considered functional,
on D^. In fact, neither the equation (2.87) need have a solution from the set
D^, nor the functional (2.88) need attain, on D^, its least value, even in very
simple cases: Consider the problem
0 for 0 g X ^ 0.5 ,
/W = ( (2-94)
1 for 0.5 < ;c < 1 .
Variational Methods 27
Obviously, fe Lj^, l). The problem (2.90), (2.91) cannot have a solution
UQ 6 D^: In fact,
so that, for MQ e ^y(, ^"o is a continuous function in the interval [0, 1],
while the function (2.94) does not possess this property (and, at the same
time, cannot be made continuous in that interval even when changing
its values on a set of measure zero). Thus the equation (2.90) cannot have
a solution UQ e D^. From the second assertion of Theorem 2.2 it then
follows that even the functional (2.88) cannot attain its minimum on D^
since the equation (2.87) should have a solution MQ e D^ in that case.
It follows from the preceding example that it is necessary to generalize
the concept of the solution in an appropriate way, because without doing
this we would not be able to include into our theory even very simple
problems from applications (deflection of a bar with a discontinuous
loading, etc.). The idea of how to perform this generalization is relatively
simple: To extend the domain D^ of definition of the considered operator A
by adding some suitable functions, and to extend the functional (2.88)
onto this extended domain. If we succeed in constructing this extension
in such a way that the so extended functional will actually attain its least
value, for a certain function UQ, we shall call this function a generalized
solution of our problem').
Such an extension can be constructed, in a relatively simple way, if the
operator A is positive definite on D^ (thus in the case most frequently
encountered in applications): Let us define, on D^, a new scalar product
(«, 1)4 by
) In a special case we may get «Q G D^. Then UQ is the classical solution of the problem,
according to Theorem 2.2. Such a case is encountered if the given data of the problem
are sufficiently smooth. Questions concerning smoothness of a generalized or of a weak
(see further) solution, especially in the closed domain G, are not simple. See [3], Chap.
46, and, in particular, [2], Chap. 4.
28 Chapter 2
see [3], Chap. 10.) On the base of this scalar product, the norm
are defined. In this way, we come to a metric space — let us denote it by S^.
This space can be shown to be not complete. In [3], Chap. 10, the manner
of achieving its completion is shown, adding the so-called ideal elements.
These ideal elements are functions of the space L 2 ( G ) ^ ° ) which cannot be
characterized very intuitively here. Let us note only that these functions
have, in G, the so-called generalized derivarives (see further, p. 38) up
to the /c-th order inclusive {2k being the order of the given equation) and
fulfil the prescribed boundary conditions in some generalized sense. From
the mathematical point of view they can be characterized as such functions
from the space £3(6) which are limit elements — in the metric of the space
£2(0) — of sequences which are Cauchy sequences in the space S^ but
which do not have a limit in that space.
On the thus extended set of functions, the scalar product (u, v)^, defined
originally by (2.95) for functions from D^ only, can be extended, as well
as the norm (2.96) and the distance (2.97). It can be shown ([3], Chap. 10)
that the so constructed space — let us denote it by H^ — is complete in this
metric, thus that it is a Hilbert space. The set of functions from D^ is dense
in this space.
The functional (2.88) is then extended to the whole space H^ by
") In the quoted Chap. 10 from [3] a more general case of a Hilbert space is considered.
Variational Methods 29
U = — , (2.99)
c
where C is the constant of positive definiteness from Def. 2.2 (see (2.55)).
Inequality (2.99) states the continuous dependence of the solution on the
right-hand side of the given equation: If f(x) changes "not too much"
in the norm of the space L2(G), then the solution Uo{x) changes "not too
much" in the energetic norm || ||^.
This fact is utilized if estimating the error of an approximate solution.
See further, p. 34.
The proof of the existence of the generalized solution UQ, i.e. of the fact
that the functional (2.98) attains its minimum on H^, is based on the so-
called Riesz theorem, which is an abstract one and gives no direct effective
way of getting this solution. For an a p p r o x i m a t e minimization of the
functional (2.98), many methods have been developed. Let us start with the
most typical of them, i.e. with the Ritz method.
Choose, in the space H^, a base
Vi,V2,...,v„,..}'). (2.100)
The requirement that (2.100) be a base, and thus a complete system, is very
important in direct variational methods (not only in the Ritz one) because
it means that an arbitrary function from the space H^ can be approximated
to arbitrary accuracy (in the metric of this space) by an appropriate
linear combination of the functions (2.100). Thus completeness of this
system plays an essential role in questions on convergence and represents
a basic requirement imposed on the sequence (2.100). Moreover, further
properties of the system (2.100) are required to ensure stability of the
numerical process. It follows that a suitable choice of a base is not at all
simple. Therefore, Chapters 20 and 26 of [3] are devoted to this question
and the most current types of bases are presented there, with practical
remarks on their properties. (See also Chap. 42 in [3] on the finite element
method.)
) Thus a linearly independent complete system in H^, see p. 14. For the base func-
tions very often admissible functions of the operator^ are chosen. Here, density of D^
in the space H^ is utilized.
30 Chapter 2
Thus let the base (2.100) be given. Take the first n terms of this base and
denote by S„ the n-dimensional subspace of the space tl^ formed by all
functions of the form
z „ = b^v, + ... + b„v„. (2.101)
The Ritz method consists in minimizing the functional (2.98) on S„. In more
detail: Look for an approximate solution M„ (called the n-th Ritz approxim-
ation) in the form
«„ = a^Vi + ... + a„v„, (2.102)
where the coefficients a j , . . . , a„ are to be determined from the condition
that the functional
f z„ = (z„, z„)^ - 2(/, z„) (2.103)
be minimal on the subspace S„ precisely for the function (2.102).
The functions Vi,...,v„ being fixed, the functional (2.103) becomes
a (quadratic) function of n variables b^,..., b„. The just mentioned condi-
tion on the minimum of this functional on S. leads to the conditions
dFz.
= 0,...,''^' = 0. (2.104)
dbi bi=ai,...,b„ = a„ ^b„
{vi, V„)A. Oi + {V2, v„)^ a2 + ... + {v„, v„)^ a„ = (/, v„). (2.105)
If, moreover, the base functions v„ are chosen from D^ (see Footnote
11, p. 29), the system (2.105) can be written in the form
(Avi, Vi) Oi + (Av2, Vi)a2 + ... + {Av„, v^) a„ = (/, v^),
{Avi, V2) ai + (Av2, V2) ^2 + • • • + (^!^n. V2) a„ = (/, V2),
[Av^, v„) fill + (Av2, v„) 02 + ... + {Av,„ v„) a„ = (/, i;„). (2.106)
Variational Methods 31
The matrix of the system (2.106) (and, of course, also of the system (2.105))
is symmetric because the operator A is positive definite — according to the
assumption — and, consequently, symmetric, so that we have {Av2, v^) =
== (Av,, V2), etc.
Since the system (2.100) is linearly independent, the determinant of the
system (2.105) (or (2.106)) — the so-called Gram determinant — is
different from zero. Thus there exists precisely one solution a , , . . . , a„
of the system (2.105) (or (2.106)) and, consequently, precisely one «-th Ritz
approximation u„.
For n -»• 00 we have
w„ -• I/Q in H^ as well as in ^-2(0). (2.107)
For details see [3], Chap. 13.
REMARK 2.7 (the Galerkin method). Note that since the operator ^4 is
linear, equations (2.106) may be obtained formally from the condition
that the function Au„ — / be orthogonal, in L2{G), to the first n functions
of the base (2.100), i.e. from the conditions
{Au„ - / , t'O = 0, ..., (.4«„ - / , ,;„) = 0 . (2.108)
By these conditions of orthogonality, the so-called Galerkin method is
characterized which thus leads, in the linear case (and under the condition
that the base functions belong to Z)^), to the same results as the Ritz method,
(See [3], Chap. 14.)
REMARK 2.8. (The finite element method.) This method is the Ritz
method, in essence, where the base is formed by functions of a very special
type. To be concrete, solve approximately the problem (2.49), (2.50), p. 16.
Divide the interval [0, TT] into ten subintervals of lengths h = njlO and
choose, for the base functions, the functions Ui(.x), V2{x), ..., i'9(x) sketched
A ,
i \
\ / \ .
32 Chapter 2
in Fig. 6. For example, the function f 2(.x) is defined in the interval [0, ic] by
for 0 ^ X ^ - ,
10
lOx 71 2n
1 for — s X ^ — ,
10 10
V2{x) = (2.109)
lO.Y 27t
3 - for < <3-
10
371
— 0 for < X < K.
10
The just presented functions are examples of the so-called spline functions.
In higher dimensions, these functions are defined similarly. For example,
if A' = 2, the "triangles" from Fig. 6 are replaced by "pyramids", see [3].
Chap. 42.
It is clear that if these functions are to be used for base functions, system
(2.105) (not (2.106)) should be applied: In the case of the above mentioned
problem (2.49), (2.50) we have
and
{Av„v,) = [ — v'l + (1 + cos^ x) I',] % dx (2.111)
For the considered functions f i , . . . , Vg, the integral (2.110) has sense because
these functions have piecewise continuous derivatives, while the integral
(2.111) loses its sense because second derivatives of these functions do not
exist.
For equations of higher than of the second order smoother functions
than piecewise linear spline functions are applied, namely piecewise poly-
nomial functions with polynomials of sufficiently high degrees.
Note that if functions of the type (2.109) are applied to the solution
of the problem (2.49), (2.50), the approximate solution is obtained in the
form of a piecewise linear (thus not of a smooth) function. (This circum-
stance does not appear only in the case of the problem (2.49), (2.50), of
Variational Methods 33
i"ol!. ^ ^ (2.112)
p. 29, can be utilized. To be concrete, let us consider the Ritz method with
base functions
",,...,f„ (2.113)
which belong to D^. Let
i/„ = ajt'i + ... + a„v„
bet the H-th Ritz approximation (see (2.102)). Because ujx) is an ap-
proximate solution of the equation
Ait=f (2.114)
34 Chapter 2
ii„ _ „ II < k ^
\r« "o|U = -
||u„-Uo||.^ll^^"^, (2.116)
where
A = ^ + i . (2.121)
Further,
(Au, u) = j [u'^ + x(l - x) u^] dx ^ j M'2 dx ^ 71^ j u^ dx
Jo Jo Jo
(2.127)
by (2.118), where a = 0, ft = 1. Thus the operator A is positive definite,
while for the constant C of positive definiteness the number n can be taken.
Analogously, in twodimensional problems, the inequality (2.120) can be
applied.
REMARK 2.10. Let us note that the coefficient (ft - afln^ in (2.118)
is the best possible in the sense that it yields the largest possible constant
of positive definiteness. In fact, for the function
. n(x — a)
u = sm —^
b - a
36 Chapter 2
The entire just presented theory, based on the theorem on the minimum
of the functional of energy, has the advantage that it is relatively simple
and rather intuitive from the technical point of view. However, it has some
drawbacks: Only symmetric problems are considered (moreover, some
smoothness of coefficients of the corresponding operators is needed);
the characterization of the space H^ (i.e. of its elements) is not simple;
nonhomogeneous boundary conditions make certain difficulties. Therefore,
another theory has been built up, based on the concept of the so-called
weak solution and on the Lax-Milgram theorem, which removes quite
a number of these drawbacks. We present a short survey of this theory
here. For details see e.g. [2], [3].
To begin with, let us define the so-called Sobolev space W^'XG): Denote
(see p. 9) by €^"'(0) the set of functions infinitely times continuously
differentiable in the closed region G (in the familiar sense of continuous
extension to the boundary). Let
i = (ii,...,i^) (2.129)
be a vector in £jy (the so-called multiindex), the components of which are
nonnegative integers. Denote, for u e C'-'^\G),
D'u=-^^ ^ , (2.130)
or, writing
|i| = ii + ... + /,., (2.131)
DM = —r ,
dxi dx2
Variational Methods 37
i f / = (0,2),
D'u =
dxl'
etc.
Let fc be a fixed positive integer. Define, on the set C^°°'(G), a scalar
product, denoted by (u, «)^2<''>(G)' ''>'
It can be easily shown that (2.133) has all the required properties of
(2.134)
a scalar product. If we define the norm and distance in the usual way by
' ^) The basic idea of this completion is similar as in the case of the construction of the
space H^. For details see [3], Chap. 29.
*^) Denoted also by i/*(C) in the literature.
38 Chapter 2
^'*) Thus for every function (p e C^'^H.G) with a compact support in G, see p. 9.
Variational Methods 39
and that, moreover, this limit is independent of the choice of the sequence
{u„{x)} satisfying (2.140). This limit, denoted by M(S), is called the trace
(on the boundary T) of the function u e W^^^Cy^)
It can be shown (see [3], Theorem 30.1) that a constant c exists, depend-
ing only on the considered domain G, such that for every function u e Wj' '(G)
we have
hW, ^ 4uw,<-^H0) • (2-I41)
(Roughly speaking: The trace u{S) does not change "very much" in the
metric of the space L2(r) if w(x) does not change "very much" in the metric
of the space Wi^\G))
If the function u{x) is sufficiently smooth in G, in particular, if u e C^'^''''{G),
the trace of this function coincides with the function u{S) in the ordinary
sense. In the general case, the trace of a function u e WJ'X^)'-^'^''^'^P'
proximated, to arbitrary accuracy (in the metric of the space L2(r)), by
the trace of a function from C^'^\G) which is "sufficiently close" to the
function u{x) (in the metric of the space W^^\G)y, see (2.141).
If II e WJ^'(G), it is possible to define, in a similar way, traces of the first
derivatives of this function. Moreover, the boundary F is Lipschitzian and —
consequently — has a unit ourward normal v almost everywhere on T, the
components v,(S) of which belong to L2(r). Thus it is possible to define,
'^) On the precise meaning of the symbol Jj./(5) dS and on the space Z.2(/") see [3],
Chap. 28.
'*) The assumption u e W'2^\G) is essential. In [3], Chap. 30, it is shown, on a counter-
example, that a concept similar to the just presented concept of a trace cannot be
introduced for functions which belong to the space ij^*^) only.
40 Chapter 2
in the sense of traces, the derivative dujdv of this function in the direction
of this normal,
dii ^ du , ^,,
- = I—^'o (2.142)
where dujdxlS) are traces of the functions dujdx^^x). Similarly, in the case
of functions from the space W2^\G) it is possible to speak abouttraces of their
second derivatives, and, in particular, about the derivative d'^ulSv^, etc.
For details see [3], Chap. 30.
The space W^'^^G) has been obtained by the completion — in the metric
(2.136) — of the space 5^''^(G), the elements of which were functions belong-
ing to €^"'(0). Let us denote by Df^G) a space with the same metric, the
elements of which are functions from C^Q'^^G) now^'). By its completion
we obtain another Flilbert space which is denoted by W2''\G). It can be shown
([3], Chap. 30) that for fucntions v e W^''''{G) we have
17
) Thus functions with compact support in G.
Variational Methods 41
ll"lk2<'>(G)
it follows that inequalities of the type (2.118), (2.120) are special cases of the
inequality (2.145) for w = 0 on T (and for N = I, ov N = 2). Of course,
the usefulness of the mentioned inequahties hes in the fact that the values
of the involved coefficients make them the best possible ones (cf. Remark
2.10).
The subspace of the space Wj^X^) of all functions which satisfy (in the
sense of traces) the corresponding h o m o g e n e o u s stable boundary condi-
tions is denoted by F.^')
Before discussing the given equation and the "admissible" form of the
boundary conditions as well as the concept of a weak solution in a sufficient
generality, let us present some rather simple examples which give a good
insight into the entire scope of problems and, at the same time, make the
idea of the concept of a weak solution more familiar.
^') For example, if considering the problem (2.148), (2.149), (2.150), we have
V= {v; V e W^'^\Gy, v = 0, dvjdv = 0 on /" in the sense of traces}.
(2.152)
•^°) Such a solution need not exist, even in very simple cases — cf. pp. 26, 27. Therefore
we are going to find a sufficiently general concept of a solution, to be able to ensure
existence.
Variational Methods 43
Now, the Green formula is valid for functions v^, fj e W^^\G) (see [3],
Chap. 31):
dx = i\V2\'i dS — • V2 dx , (2.158)
dx, G SXi
where V; is the I'-th component of the unit outward normal to F and v^iS),
1)2(5) are traces of the functions fi(x), vj^x) on F?^) Let i be fixed and put
du
Vi = V , V2 = .
dx,
Taking into account that t; e F, so that t' = 0 on T in the sense of traces
according to (2.155), we have
dv du
- v-—dx = dx.
}G SX^ G 3xi 8xi
(2.159)
i=ija8xidXi Ja
t \ T^ T^ dx = A{v, u) (2.160)
'•=1 JG ^Xi dXi
and write, as usual,
^') (2.158) represents a generalization of (2.69), p. 21, for functions from W^^^iG),
thus not requiring fj, Wj e C*^*(G).
44 Chapter 2
^^) By doing this, we eliminate the influence of boundary conditions from our consider-
ations and can investigate the differential equation itself.
^^) In LjiG); however, u e C ' ' ^ \ G ) = > AM is continuous in G, etc.
Variational Methods 45
REMARK 2.12. It follows that the concept of the weak solution of the
given problem is general enough: This solution is a function M(X) which
belongs to W2^\G) only and satisfies the given problem in the weak sense
(2.162), (2.168). At the same time, from the way of deriving the integral
identity (2.162) it is easily seen that if a solution u e C'-'^\G) of the problem
exists, then it is also its weak solution. Thus, in this sense, the weak solution
is a reasonable concept.
^*) Since K = lvi^\G) in our case, these identities are equivalent (because of the
density of C^Q-^HG) in wi^HG)). As seen from the following examples, it need not be
the case in general, and the functions v e Kare to be considered instead of (p e CQ^HG)
in order to take the given boundary conditions into account.
^^) On the construction of such a function (g{S) being given) see [3], Chap. 11. On
existence questions see [3], Chap. 46.
^®) It is shown ([3], Chap. 33) that the weak solution dees not depend on the choice
of the function w{x), the trace of which the function g{S) is.
46 Chapter 2
— = h on r (2.170)
dv
with/eL2(G), heL^Xr).
Since
du y Su ,
(2.171)
dv i^i dx,
(see (2.142)), derivatives of the first order are involved in the given boundary
condition so that the condition (of the so-called Neumann type) (2.169) is
unstable. Thus no stable boundary condition is present here. Consequently,
If we denote, as before,
i f ^^dx--Aiv,u) (2.174)
'•=1 J G oxi dXi
and put h{S) instead of du\dv{S) in the surface integral (2.173), according
to (2.170), we obtain
A condition similar to the condition (2.168) does not appear here. We define:
Variational Methods 47
REMARK 2.13. Let us note that here the boundary condition (2.170)
cannot be taken into account with the help of a function we W2^\G) as
in the foregoing example because traces of the functions dwjdxi, which are
needed in (2.171), do not exist, in general, for functions from W^^\G).
While in the weak formulation of the solution of the problem (2.153),
(2.154), the stable boundary condition (2.154) was expressed by (2.168)
(with the help of the function w), the function h[S) from the unstable
condition (2.170) appears on the right-hand side of the integral identity
(2.175) here.
EXAMPLE 2.7.
— Au = / in G, (2.176)
du
+ au = h on F (2.177)
dv
(problem with the so-called Newton boundary condition). It is assumed
that / 6 L2(G), h e Lj^f), c = const. > 0.
The condition (2.177) is unstable, thus
F= \\f\G). (2.178)
Let, first, u e C'-^\G) be a classical solution of the given problem. Multi-
plying (2.176) by an arbitrary function ve V, integrating over G and using
the Green formula, we get — at first for the left-hand side —
^ r di^du_
I: AM dx = — vh dS -f- avu dS +
i = i J e dxi dXi
Ax .
(2.179)
Denote, in this case,
'^ f du du
y dx = A(v, u), (2.180)
i=i JcdxiBxi
We define:
Let / e L 2 ( G ) , heL2{r),ff > 0. By the weak solution of the problem
(2.176), (2.177) we understand a function u e W2^\G) satisfying the integral
identity (2.183).
In the three foregoing simple examples, the basic idea and characteristic
features of the so-called weak solution have been shown. Let us turn to
a more general case.
respectively, where B,-, C, are some linear operators and gi{S), /i;(S) given
functions on / (to be specified later).^^)
The operator A is assumed to be of the form
A= I {-irD'ia,j{x)DJ). (2.187)
Here,
i = ((i, ..., (,v), J = (ii, • • •,;A) (2.188)
are multiindices (cf. (2.129)),
«,,(x) = 1 (2.193)
) The given boundary conditions cannot be "arbitrary". First of all, they are to be
linearly independent (none of them being a consequence of the others). Besides, they
are to have other properties, to be briefly clarified below.
If the differential equation (2.184) is an ordinary one, k boundary conditions are
prescribed at every boundary point of the interval considered.
28x
) In the classical form, equation (2.184) reads
8ii+ ... + ;»
..+ !>
z (-ly^ ,IN;JI ,jJ^) •
."-.JN
Assuming the equation in the form (2.191), or in the form (2.184) with the operator
(2.187), some smoothness of the coefficients afj(x) is needed. However, as seen below,
nothing similar is required in the weak formulation of the problem. It is sufficient if,
e.g., aij(x) are bounded measurable functions only. (Even boundedness can be general-
ized; we shall not go into details here.)
50 Chapter 2
for
/ = (1,0), i = (l,0) (2.194)
and for
i = (0, 1), i = (0, 1), (2.195)
and
a,j{x) ^ 0 (2.196)
in the remaining cases.•^^)
To the operator (2.187) the so-called bilinear^°) form
•^') Sincefc= 1, so that |i| g 1, |;| ^ 1, there are three possibilities for the vectors /', /:
(0,0), (1,0), (0,1).
If
''1,0,1,0, ^ 1 ' "0,1;0,1 ^ 1
while
"'i.iiijiji = 0 Otherwise,
we have, in fact,
I (-l)i^lD'(«,,(x)Z)0 =
\i\.\J\ = i
Let us assume, for a moment, that u(x) and aij{x) are sufficiently smooth
functions (this need not be made more precise here), choose an arbitrary
<p e CQ^\G) and integrate the product cpAu over G,
Lic
cpAii dx , (2.200)
using the Green formula. If we apply this formula |j|-times to each of the
integrals
since the integrals over F vanish in consequence of the fact that the traces
of the function co(x) as well as of its derivatives are zero functions on F (</>(x)
being a function with compact support in G). Thus if we multiply equation
(2.199) by an arbitrary function (p e C\^^{G), integrate over G and apply
the Green formula in the just described way (under the assumption of suf-
ficient smoothness of w(x) and fl,j(x)), we get
A{cp,u)=^{q>,f). (2.203)
Since this result holds for an arbitrary (p e C^°°'(G), (2.203) turns into
an integral identity
A{(p,u) = {(p,f) for every (peCi'"\G). (2.204)
Conversely, by the "inverse" process we come to the identity
which yields (2.199), in consequence of the density of the set C^^'^^G) in the-
space £.2(0).
Note that all the integrals (2.202) have sense if 11 e W^''\G) and if, e.g.,.
fl,y(x) are bounded measurable functions. Thus we define, without regard
to the smoothness of M(X) and a;j(x):
52 Chapter 2
For example, in the case of an equation of the eighth order (k = 4), the
following conditions on F may be given:
u = t/o , (2.210)
Variational Methods 53
A'u = hi , (2.212)
^ + M = /i2 . (2.213)
dv
The conditions (2.210), (2.211) are stable, the conditions (2.212), (2.213)
are unstable. Thus, n = 2, k - fi = 2 here. The conditions (2.210), (2.211)
are of the Dirichlet type (2.207) with s^ = 0, Sj = 2, the other two conditions
are conditions of the type (2.209).
In Example 2.5, the boundary condition (2.154) is stable (of the Dirichlet
type), thus we have /z = 1, A; — /( = 0. The boundary conditions (2.170),
(2.177) in Examples 2.6, 2.7 are unstable; /t = 0, fc — yti = 1 in both the cases.
B^u = V F^u = Qi ,
^') For simplicity, we write ^ j , ...,^^ here instead of ^ j , , ...,gs^ as should be written
in accordance with (2.207).
54 Chapter 2
F^u =
^v ds
is a "permitted" operator.
For a more detailed analysis and other examples see [2], Chap. 1, or [3],
Chap. 32.
^^) Nothingis said about the orders of derivatives in the "tangential direction", except
the requirement that the total order of the operators Fj, Fj be gj ^ — I.
Variational Methods 55
with
((z;, u)) = A{v, u) + a{v, ii) , (2.224)
where •4(v,u) was the bilinear form (2.180) corresponding to the given
56 Chapter 2
operator and
a(v, w) =• avu dS (2.225)
Note that a(v, u) (in contrast to A(v, u)) is a surface integral (i.e. an integral
over the boundary F) with an integrand bilinear in v and u (i.e. linear in v
as well as in w). Similarly, the term
in L^ir),
^ " + au = •h on F- (2.231)
dv
Variational Methods 57
Au =f in G, (2.232)
B-u = gi on F, i=l,...,n, (2.233)
C,u = hi on r , i = 1, ..., /c - /(, (2.234)
where A(v, ti) is the bilinear form (2.197), a{v, u) is a bilinear boundary
form (a sum of surface integrals with integrands bilinear in v and w) and
x{v, h) is a boundary form linear in v, assumed here in the form^^)
*"" r d''v
•4v,h)==Y. -T^'.-dS, (2.236)
i = i jr S^
where /i,(S), /' = I, ..., k — n, are functions from the boundary conditions
(2.234) and f,-, / = 1,...,/c —/(, are numbers from (2.208). However,
in this more complicated case, it is not possible, in general, to obtain the
integral identity (2.235), the (at least formal) consequences of which the
equation (2.232) and the conditions (2.234) should be, directly by multi-
plying the equation (2.232) by an arbitrary function ve V, integrating over G
and applying the Green formula in the famihar way (cf. (2.201), (2.202)).
Some "supplementary rearrangements" are needed, in the general case.
To make this problem clear, let us present the following examples (we have
chosen examples in ordinary differential equations to this purpose):
EXAMPLE 2.8.
u<*'=/, (2.237)
u(0) = 2 , u'(0) = 3 , (2.238)
u(l) - u'(l) = 5 , t("(l) - «'"(!) = 7 . (2.239)
) In order that all the boundary conditions (2.234) can be taken into account; see
below.
58 Chapter 2
EXAMPLE 2.9.
MW=/, (2.240)
M(0) = 2 , M'(0) = 3 , (2.241)
u{l) ~ u'(l) = 5 , u'{l) + i("(l) - !('"(!) = 7 . (2.242)
EXAMPLE 2.10.
ti^*'=f, (2.243)
M(0) = 2 , u'(0) = 3 , (2.244)
«(]) _ w'(l) = 5 , M"(1) = 7 , (2.245)
The given equation is of the fourth order, so k = 2. The stable boundary
conditions are the same in all the three examples and are of the "admissible"
type. At the point x — 0, the Dirichlet conditions are prescribed; at the
point X = 1, one stable and one unstable condition are given, thus /< = 1,
k ~ IX = 1; the stable condition is of the form (2.214) with Sj = 0 (corres-
ponding to v{\)) and t^ — \ (corresponding to v'{).)). Further,
V = {v; V e nj^\0, l), v{0) = 0, v'{0) = 0, r(l) - i''(l) = 0
in the sense of traces) . (2.246)
The Green formula, reduced to integration by parts here, gives for
VE V
This is the integral identity of the form (2.235) already, with a[v, M) = 0
and fi = 1 in (2.236) (note that in the case of ordinary differential equations
Variational Methods 59
In each of the last two cases, the "converse" integration by parts then
leads to the given equation and to the prescribed unstable boundary condi-
tions, as can be immediately verified similarly as in the case of Example
2.8.
REMARK 2.15. The presented examples show that the boundary form
a(v, u) plays an essential role in the formulation of unstable boundary
conditions. / / this form is found in such a way that, starting with the
integral identity (2.235) with A(v, u) given by (2.197) and y.(v, h) given by
(2.236), the unstable boundary conditions follow by formal application
of the Green theorem, or by further supplementary rearrangements, if
necessary^'*'), we say that a(v, u) is the bilinear boundary form correspon-
ding to the given problem (2.232) —(2.234). Or that
{{v, u)) = A{v, u) + a{v, u) (2.257)
is the bilinear form corresponding to this problem. We say also that the
form a{v, u) (or ({v, u))) corresponds to the operator A and to the boundary
operators Bj, C;.
The reader may object that this concept is rather unclear. In fact, questions
may arise on how to construct the form a{v, u) in more complicated cases
(the above mentioned rearrangements may be actually artificial enough),
even questions concerning its existence and uniqueness. Therefore, from the
mathematical point of view, it is more comfortable to take the form
a(v, u) as given and to derive, on its base and on the base of the form x(v, h),
the unstable conditions CjU = /i,-. Thus in this case, the forms a[v, u) and
x{v, h) are "primary concepts" from which the boundary conditions CiU = hi
are derived. This conception is followed, e.g., in the book [2] by Necas as
well as in the author's book [3]. However, in practice, the boundary con-
ditions are prescribed and the form a(v, u) is to be found. Therefore, in
spite of many preferences of the above mentioned conception, we give the
definition of a weak solution of the considered problem with the just in-
troduced concept of the form ((v, u)) corresponding to this problem, being
well aware of its disadvantages. The reader who is accustomed to the
above conception may assume the forms {(v, u)) and x(v, h) to be "primary"
throughout the book.
34
) The equation Au = /follows automatically.
Variational Methods 61
Here
k-n C at,,
<^'' h) = 1 T 7 ''•• dS (2.266)
DEFINITION 2.6. The form {{v, u)) is called bounded in the space
VFJ^X^)^*^) if a constant K > Q exists (independent of v and w) such that
we have
THEOREM 2.6. Let the form {(v, ii)) be bounded in H^i"(^)")""'' V-eUip-
tic. Then there exists precisely one weak solution of the problem (2.258) to
(2.260) in the sense of Definition 2.5.
For the proof see [3], Chap. 33.
I (Tvu dS (2.269)
i r
(TVU dS = a
r
vu dS ^'^•IIHlMO-lhllW)^
so that the surface integral on the right-hand side of (2.145) vanishes for
every v e V.'^^)
Many examples on how to establish boundedness and F-ellipticity can be
found in [ 3 ] , Chap. 33, simultaneously with an extensive table of (bounded)
l^-elliptic problems.'*^)
In particular, all problems of Table 2.1 (p. 22) are bounded in PVj*'(G)
and F-elliptic.
Note that the form (2.254), p. 59, is bounded in W^^\0, 1) (the proof is
similar to that in (2.27Q)), while the form (2.256),
^(».»)-i I if^Y<i«-o.
while
lrll»»'2<"(G) -* ^ •
On how to proceed in such cases see [3], Chap. 35.
Let the form ((r, u)) satisfy the conditions (2.267) and (2.268), ensuring
the (unique) weak solvabihty of the problem (2.258) —(2.260). Let it be,
moreover, F-symmetric, i.e. let
be vahd. Then ([3], Chap. 34) the weak solution of this problem mini-
mizes, in the space F, the functional
(where W(A) is the function from Definition 2.5). If the Ritz method (or the
finite element method as its special case) is applied, then for the coefficients
flj of the Ritz approximation
(2.276)
•*^) K-ellipticity implies positive definiteness. For details see [3], Chap. 36.
PART I. EXAMPLES
66
Examples 67
u(x) dx
G
instead of
etc.
Chapter 3
— + Au =f in g = G X (0, r ) , (3.1)
dt
uix, 0) = Uo(x) , (3.2)
BiU = 0 on r X (0, T) for / = 1, . . . , / t , (3.3)
Cju = 0 on r x ( 0 , T) for / = 1, ...,/c - / ( . (3.4)
Here, G is a domain mentioned on the foregoing page (thus bounded in Ef,,
with a Lipschitz boundary f), (O, T) is a bounded interval; A is a linear
operator of order 2k, of the form (2.187), p. 49, i.e.
A= I (-l)l'IZ)'(a,(x)DO, (3.5)
68
Algorithm of the Method 69
A = -^ =
dx\ dxl
d
(3.9)
dx \dxj dx2 \
EXAMPLE 3.2.
h = -, (3.14)
P
tj=jh, j = 0,l,...,p, (3.15)
and divide the interval
/ = [0, r ] (3.16)
into p subintervals
Jj-llj-uh]^ 7 = 1,...,P. (3.17)
Starting with the function
Zo(^) = "o(.^) (3-18)
from (3.2), we have to find, successively for7 = 1, ...,;;, the functions
z,(x) (3.19)
(the approximations of the solution u for r — tj) as solutions of the equations
In this way, the solution of the parabohc problem (3.1) —(3.4) is reduced
to the solution of (elliptic) problems (3.21) (or (3.20), which is the same),
(3.22), (3.23). In each of them, the difference quotient
Z; - ^J-l
(3.24)
h
stands for the derivative dulct in the original problem.
-[(l+sin^x)z;.]'+|(z, - z , _ 0 = l (3.30)
h
or
In very simple cases, the problems (3.21) —(3.23) can be solved "directly",
in a "closed" form. This occurs, for example, if A is an ordinary dif-
ferential operator with constant coefficients, the function / is of a special
form and the boundary conditions are simple. This was the case of the
problem 1.1 presented in Chap. 1. In more comphcated cases, the problems
(3.21)—(3.23) cannot be solved in such a simple way (they need not have
classical solutions, either), and approximate methods should be apphed,
most frequently the variational ones. Thus let us turn to the weak formula-
tion of the considered problems.
In Definition 2.5, p. 61, it is shown how to convert the elUptic problem
Au = / in G, (3.34)
B,M = 0 on r , ; = 1, . . . , / t , (3.35)
C,u = 0 on r , i = I,..., k — n , (3.36)
into a "weak" one. Note that in our case (homogeneous boundary condi-
tions), the term x{v, h) (see (2.266), p. 62) vanishes and that for the function
VV(A) from (2.264), p. 61, the zero function can be chosen. Thus, according
to the above mentioned definition, by the weak solution of the problem
(3.34) —(3.36) such a function
ueV (3.37)
is understood for which the integral identity
({v, M)) = \ vf dx
{v, «)) for every ve V (3.38)
integral
L vAu dx , ve V,
— Au = x\x2 in G, (-^-^l)
u = 0 on r. (3.42)
Here, /c = 1,
V = {v; WV\G), V =-0 on r in the sense of traces} . (3.43)
Multipying equation (3.41) by an arbitrary function v e V, integrating
over C and using formally the Green formula (2.158), p. 43, we obtain
(cf. (2.159)) the integral identity
dv dii dv 5u . ^ ,
1 -^— I dxi dx2 = x\x2V dxj dxj
Q \dXi dXi dXj 0X2) a
for every r e F . (3-44)
This is the integral identity (3.38) with
*) The second of the conditions (3.47) is unstable with respect to the given operator,
thus it does not appear in the definition of the space V.
74 Chapter 3
I 0
1 . u dx = t;[(l + sin^ x) u ' ] ' dx = -[_v{l + sin^ x) u'Jg +
f
Jo
1" dx = t'(7r) (1 + sin^ 7r) U'(K) -
D Jo0
v[{l + sin^ x) M']' dx .
(3.51)
Choosing for v(x) an arbitrary function (p e Co"\0, n) (thus with a compact
support in (O, n)), we get
Thus, in both the problems (3.41), (3.42) and (3.46), (3.47), the obtained
form (3.45) and (3.50) is the form corresponding to the given problem
in the sense of Remark 2.15, p. 60. Thus, by the weak solution of the prob-
lem (3.41), (3.42) such a function ue V is understood for which (3.44)
is satisfied, with K given by (3.43). Similarly, the weak solution u e F of the
problem (3.46), (3.47) should satisfy the integral identity (3.49), with V
given by (3.48).
A' = A + - , (3.54)
h
so that the form
f+tI^EL,{G). (3.56)
h
Then by a weak solution of the problem (3.21) —(3.23) we understand
such a function
ZjeV (3.57)
') vA'u Ax = r if
vAu dx -i— 1 vudx
•1G
*) The term
1 1f
-iv,u) = - 1 vu dx
n
''JG
does not change in the process of application of the Green formula. Thus if ((v, u))
corresponds to the problem (3.34) —(3.36) (in the sense of Remark 2.15, p. 60), then
(((f, u))) given by (3.55) corresponds to the problem (3.21) —(3.23).
76 Chapter 3
is satisfied.
Here, (((y, H))) is given by (3.55) and [{v, u)) is the form corresponding
to the problem (3.34)-(3.36).
f+^eL,{G). (3.61)
h ^ '
Then if the form {{(v, u))) is bounded in PFJ'''(G) and V-elliptic, there exists
precisely one weak solution of the problem (3.21) —(3.23), i.e. precisely
one function ZjS V satisfying (3.58).
How can these two properties be estabhshed in the case of the form
((r, u)) was shown in the concluding part of Chap. 2. In particular, bounded-
ness and F-ellipticity of the form (3,45) in the case of the space (3.43) has
been proved there (see the text related to (2.271), p. 63, and the correspond-
ing Footnote 40). Boundedness of the form (3.50) can be estabhshed by
the same argument as in the mentioned footnote (the function 1 + sin^ x
is bounded). This form is F-elliptic as well. However, to prove this, the
Friedrichs inequality (2.145) (for A' = l) cannot be applied in this case
since the boundary condition u'{n) = 0 gives no information about 1/(11).
But the first of the boundary conditions (3.47) yields, for v e V,
Thus,
1^ dr . I v {t)dt
0 Jo
by the Schwarz inequality, i.e.
Thus, by (3.50),
and finally
((t;,.)) = l((.,r)) + i((.,.))^
1 /I
> 1:^2(0 ,;•) + II^'IILJCO,:.)] ^ 2 ll ll'»'2('>(0,Il) » (3.62)
2V7t In
by which F-ellipticity of the form (3.50) with a = l/27r^ follows.
78 Chapter 3
LEMMA 3.1. Let the form {(v,u)) be bounded in [V^''\G) and V-elliptic.
Then the same holds for the form
^ ( 0 , 0 = 0, f^(l,0 = O (3.69)
ox ox
we come, applying the method of discretization in time, to the problems
Azj + ~ ^ = f + ^ , (3.70)
h h
Algorithm of the Method 79
Now, Theorem 3.1 and Lemma 3.1 yield a simple criterion on the solva-
bility of the problems (3.21)-(3.23):
Azj = (3.72)
and 7(,(x) = iio(x). Both the boundary conditions (3.71) are unstable with respect
to the operator A, so that
(3.73)
The corresponding form
is bounded in I f p ' ( 0 . 1)> but not K-elliptic. (For example, the function vix) = 1
belongs to H ' i ' ^ O , 1) and
with / I
c = niin(^l,-
Further, (((y, «))) being bounded, the problems to find the functions z^. e V,j = 1 ,...,p,
satisfying
are uniquely solvable if o n l y / e £2(0, 1), Ug e L^iO, 1). Thus the method of discretiza-
tion in time is applicable also in the case of the so-called problems of the Neumann
types.
80 Chapter 3
THEOREM 3.2. Let h> 0 be fixed, f e L^iG), UoeLjiG). Let the form
((f, M)) corresponding to the problem (3.34) —(3.36)^) be bounded in H^]*'(G)
and V-elliptic. Then each of the problems (3.21) —(3.23), solved successively
for j — \, ..., p, has precisely one weak solution, i.e. there exist precisely
p functions ZJE V, j = 1,.... p, satisfying, subsequently, (3.58).
In fact, ((v, uj) being bounded in W^''\G) and F-eHiptic, the same holds
for the form (((v, «))) according to Lemma 3.1. Further,/e 1-2(0), MQ e £-2(0)
imply
/+^eL,(G).
h
Thus, according to Theorem 3.1, there exists precisely one function Zj e F
satisfying (3.58) (for; = l). Now,
z, 6 W^'XG) => z, e L,{G) ,
so that
/ + ^ 6 L,{G).
h
Thus there exists precisely one function Zj e F satisfying (3.58) (for /" = 2),
etc.
In particular, the form (3.45) being bounded in W^''\G) and F-elliptic
with F given by (3.43), each of the problems (3.27), (3.28) with Zo(x) given
by (3.25) has precisely one weak solution. The same holds for the problems
(3.31)-(3.33) with Zo(x) given by (3.29).
REMARK 3.2. [Approximate solution.) If, moreover, the form (((f, u)))
is K-symmetric, i.e. if
(((f, u))) = (((«, t;))) for every v, u e V') (3.78)
is valid, current variational methods can be applied to approximate solution
of the problems
ZjeV, (3.79)
Let
Vi,...,v„ (3.81)
= ^ , „ / + ^ ) , (3.83)
b) ERROR ESTIMATES
feV, (3.87)
AfeL,{G), {{v,f)) = {v,Af) for all veV.'') (3.88)
Denote
\\Af\\ = M . (3.89)
|!.(x,/,)-z,(x)|| (3.90)
is valid:
!l«(x,0)-z,(x)i^^. (3.91)
||„(x,o)-zXx)|| = ^ ( l - e - ^ ^ - * ) . (3.93)
On how to estimate the error in the case that ||uo|| > 0(i.e. if (3.86)isnot
fulfilled) see in the next chapter (p. 99 IF).
REMARK 3.3. The error estimates (3.91) and (3.93) are derived under
the asumption that the problems (3.79), (3.80) have been solved exactly.
An approximate solution of these problems, e.g. in the sense of Remark 3.2,
is the source of further errors. Let us mention here the following fact:
If we solve the problem (3.79), (3.80) for 7 = 1 and use the Ritz method,
we get
z* = alvi + ... + alv„, (3.94)
84 Chapter 3
where a[, ...,al are determined by the Ritz system (3.83) with7 = 1. Here,
the difference between the exact solution Zi(x) and the function z*(x) is
caused only by the fact that the Ritz approximation is obtained instead of the
exact sohition of the considered problem since, if 7 = 1, z*_i(x) = z*(x) =
= ZQ[X). When computing z*{x) the situation is rather different because
here the difference Z2{x) — zl(x) is caused not only by the approximate
solution of the problem (3.79), (3.80) for j = 2, but also by the fact that
the just obtained approximation z*(x) stands on the right-hand sides of
(3.83) instead of the function ZI(A) which is not exactly known. Continuing
the numerical process in this way, the error "cumulates" in every step.
How to get a "total" estimate for \u{x, tj) — z*(x)\ in this case is well
seen from Example 4.6, p. 107. For the theoretical treatment of these questions
see Chap. 14.
Chapter 4
As pointed out on p. 66, the aim of Part I is to acquaint the reader with
the procedure of the method of discretization in time, including questions
concerning error estimates.
85
86 Chapter 4
I + A{t) u = f{t),
ct
thus with coefficients depending, in general, not only on the "space" va-
riables Xi,..., Xy but also on /, is considered.
a) ILLUSTRATIVE EXAMPLES
du d^u
= sin X in e=( ) X (0, 1 ) , (4.1)
dt dx^
t/(x, O) = Mo(x) = 0 , (4.2)
boundary conditions
z,(0) = 0 , 2,(71) ==0 (4.6)
(by (4.3)), with
Zo{x) = 0 (4.7)
(by (4.2)).
In Chap. 1, the explicit formula for Zj(x) was found:
Moo(^) = ( 1 - Y^°) ' ' " '"" ^ ^'^^^^ ' ' " ^ ^'^'^^^
(see p. 6). Because the exact solution is known in our case (see (1.4), p. 1),
H(X, f) = (1 - e - ' ) s i n x , (4.12)
it is possible to compare the functions (4.9) —(4.11) with the functions
u{x, 0.2) = (1 - e-°-') sin x = 0.1813 sin x , (4.13)
u(x, 0.4) = (1 - e""-'^) sin x = 0.3297 sin x , (4.14)
u(x, 1) = (1 - e-i) sin x = 0.6321 sin x . (4.15)
This comparison will appear to be very instructive for testing the efficiency
of the estimate (3.93), p. 83,
!]„(x,r,)-z,(x)i^^(l-e-^^-^). (4.16)
88 Chapter 4
Remind that this estimate was derived under the following assumptions:
The form ((v, u)), corresponding to the operator A and to the given boundary
conditions, is bounded in the space PFJ*'(G) and F-elliptic, i.e. (3.59), (3.60),
p. 76 hold,
\({v, u))\ ^ -Kj|£;||^^(t,(gj . |JM||vr2(W(G) for every two functions
v,ue \^f\G), (4.17)
and
((u, u)) ^ a|!i;!|^,(.,(C) for all veV; (4.18)
moreover, let the coefficient of positive definiteness C be found (as large as
possible) so that
((u, u)) ^ C'||y|P for all veV (4.19)
holds; further, let
feV (4.20)
and
AfeL^iG), {{v,f)) = {v,Af) for all VEV. (4.21)
In (4.16) we then put
M = \\Af\\ . (4.22)
In our case,
All = -u" , (4.23)
W^^\G)=Wi'\Q,n), (4.24)
V= {v;ve ^f\0, %), v{0) = 0, V{K) = 0
in the sense of traces}.^) (4-25)
Taking ve Varbitrary, we have, integrating by parts,
Now
((., r)) = iiiv, v)) + l{{v, v)) ^ i\\vV + m ' = mik^Ho.n, .
(4.31)
vf dx
v'u' dx = I vj for all ve V
Jo Jo
implies, formally, —«" = /(cf. the text preceding (3.52), p. 74), so that the form
(4.27) is actually the form corresponding to the problem
w(0) == 0 , u{n) = 0
in the sense of Remark 2.15, p. 60. In what follows, we are not going to point out
this fact in problems where it is obvious immediately.
•') This inequality has been derived for functions i; e C^^^ [a, b] satisfying v(a) = v{b) =
= 0. In consequence of the density of such functions in the space H^^')(a, b), it remains
valid for every function from this space.
90 Chapter 4
wherefrom the F-ellipticity of the form (4.27) follows. At the same time, we
can take, in (4.19),
C = 1 (4.32)
according to (4.29).
Further,
f{x) = sin X
in our case. Obviously,
sin X e C<^'[0, n] , sin 0 = 0 , sin n = 0 => sin x e F , (4.33)
and
Af = -(sin x)" = sin X e ^2(0, n) . (4.34)
so that
t
M^ = \\Af\\^ = I sin^ X dx = - , (4.35)
M = /^ (4.36)
|]u(x,0.2)-z,o(x)||^5^(l-e-''-^)^^ =
Kx,l)-z.ooWM^(l-e-)^^ =
At the same time, we get, by (4.9) —(4.11) and (4.13) —(4.15), the actual
errors:
Comparing the actual errors (4.41) —(4.43) with error estimates (4.37) to
(4.39). we can conclude that the efficiency of the error estimate (4.16) is
very high^) and that this estimate cannot be improved, practically. It does
not mean, of course, that in individual cases its efficiency could not be lower.
See Example 4.2 below.
REMARK 4.1. If the more rough estimate (3.91) is applied instead of the
estimate (4.16),
|l«(x,0)-zX.x)||g^, (4.44)
'^) We have
f n
Sin x^ = sin^ x d x = - . (4.40)
Jo 2
^) It becomes worse with increasing j ; this fact is well to be expected because, first,
the error becomes "cumulated" with increasing y, and, further, the norms of the solu-
tions themselves increase with j .
92 Chapter 4
1, . „ ,x I VI ^ 40 . O.OP
|]i((.x, 0.4) - Z4o(-'c)|| ^ / - = 0.002 51 , (4.46)
l, . .. , .. ^ 100.0.01^
/ - = 0.006 27 (4.47)
The following example shows that the estimate (4.16) loses its efficiency
in the case of "rapidly oscillating" functions/: ®)
'') Better said, the estimate (4.16) is derived for a relatively wide class of functions
requiring o n l y / e Kand Af& Z-2(G). It does not take into account their special proper-
ties to which also rapid oscillations belong. Such an estimate would be too cumber-
some and unefficient in "reasonable" cases.
Linear Parabolic Problems 93
with ^oi^) = 0 according to (4.49). Explicit solutions are (we shall not go
into details here; cf. Chap, l)
On the other hand, the corresponding bilinear form and its properties
remaining unchanged,
the estimiate (4.16) remains valid with the only difl'erence that instead of
2
by (4.36) we have here
M = 4/^, (4.60)
94 Chapter 4
because
M'2 =
_ li
AfAf\\2 16 sin' 2x dx = 8;r. (4.61)
0
The fact that in our case M is four times larger than in the preceding case,
and that, in consequence, the error estimate (4.16) is four times worse,
manifests that the efficiency of (4.16) will by far be not as high as before.
In fact, we get, by (4.16), for h = 0.01 a n d ; = 10,
|lKx.OJ)-=,4x)N?^^4^W30-e-«.) =
= 0.02(1 - 0.904 84) / - = 0.002 39 , (4.62)
which is by far not so close to the actual error as in the preceding case.
— + Au = 0 in O = G X Co, r ) , (4.63)
dt ^ . J
Also, for this case, the error estimate of the form (4.16) has been derived
(Theorem 13.6, p. 264) under the following assumptions: The form ((y, u))
corresponding to the operator A and to the given boundary conditions
' ) The problem (3.1) =(3.4), being linear, can be decomposed into a problem
with nonhomogeneous equation and homogeneous initial condition and a problem
with homogeneous equation and nonhomogeneous initial condition. Examples of the
first type have just been discussed, those of the second type are going to be discussed
in the text which follows.
Linear Parabolic Problems 95
In this case,
M = \A^UO\\ (4.68)
is to be taken in (4.16).
Of course, the assumptions (4.67) (in particular, the first and second one)
are rather restrictive. As concerns the efficiency of the estimate (4.16)
in this case, similar conclusions can be made as in the case / 4= 0, MQ = 0.
See the following examples.
This example is also an illustrative one because its solution is well known:
It was chosen in order that we had the possibility to compare the approxi-
mate solution, obtained by our method, with the exact one.
The method of discretization in time leads to the equations
- 4 + 7^ = ^ > (4.73)
h h
^2 == A2 sin X (4.80)
/, = 0.01 , 7 = 20 , (4.83)
Linear Parabolic Problems 97
we get
^ sin X = 0.8195 sin X, (4.84)
-'' " 1.01-
in very good agreement with
For the norm of the difference u{x, 0.2) — Z2o{x) in the space £2(0, n)
we thus get
|!u(x, 0.2) - Z2o(x)|| = 0.0008||sin x|| = 0.0008 / - = 0.001 01 .
(4.86)
Compare this actual error with the estimate (4.16). Here,
M = II^Vll - J l (4.91)
according to (4.68). Thus, we get the error estimate
REMARK 4.2. It is not surprising that the results (4.86) and (4.92) are
identical with the results (4.41) and (4.37), pp. 90, 91. In fact, by the substi-
tution
Zj = Sj + sin X
If the function MO(^) oscillates rapidly, the error estimate (4.16) loses its
efficiency:
EXAMPLE 4.4.
« > 1 an integer.
(1 + n hy
Similarly as in Example 4.2, a very good agreement between (4.96) and
(4.99) can be shown. However, the error-estimate (4.16) loses its efficiency.
This has happened, as we have seen, in Example 4.2 already, when sin nx
appeared on the right-hand side of the given equations. However, while
in that case M was n'^-times as larger than in the case of the right hand side
/(x) = sin X in consequence of
J - ^ = 0 in O = (0,7r)x(0,l), (4.102)
ct ex
ti{x,0) = Uo{x) = 1 , (4.103)
M(0, ?) = 0 , u{n,t) = 0. (4.104)
This example is also an illustrative one: The solution is
4 e~"''
u{x, t) = - Y. ^^^ ^^ • (4.105)
Tc n = i , 3 , 5 , . . . n
Zo(x) = 1 . (4.108)
h h
z, = 1 + C, sinh h C, cosh — .
cosh 1
C2= - ] , C i =
sinh —
and taking the first five, for example). However, having the results of the
examples solved above at our disposal, it is better to proceed on in the follow-
ing manner: As well known, we have for the function Ug[x) = 1
4 ^ sin(2n - l ) x . ^ .. -. /^ ,,r,\
Mo = - S —^^ '— in L2{0,n). (4.110)
itn=i 2n — 1
Linear Parabolic Problems 101
Take, for example, the first three terms of the series (4.110),
, s 4 /sin X sin 3x sin 5x\ , ,
The whole example being illustrative, the exact solution (4.105) of the
given problem (4.102)-(4.104) is known. In particular,
4 r _ e"'"''^ g-25.0.4
u(x, 0.4) = - e ° * sin X H sin 3x H sin 5x +
nl 3 5
' ) The reader may check, by direct computation, that the same result is obtained
(in a rather more labourious way) by the Ritz method, taking the first five terms of the
base (4.109).
102 Chapter 4
00 -(2n-l)2.0.4 -1
+ y^ sin (2n — l) x =
n'^4 2n — 1 J
00 -(2n-l)2.0.4 -I
where r(x, t) is the solution of the problem (4.102) —(4.104) with the initial
10 ) We use the orthogonality of the sine functions in i2(^' ^^ which yields
00 'oo TT °°
II y B„, sin m x P = V B^llsin^ wx|| = - Y Bf„ .
m=t m= 1 Z m=\
Linear Parabolic Problems 103
in our case. However, neither of these two estimates is of use: In fact, first,
the functions sin 3x and sin 5x in (4.111) oscillate rapidly. Thus, by (4.68),
we have
4
M = 11^^X3 II = - II sin X + 27 sin 3x -I- 125 sin 5x|| , (4.125)
n
and the obtained coefficients are too large in order that a reasonable estimate
can be expected. Similarly,
e-°-^ = 0.67 ,
and ||r3|i is not negligible at all as compared with unity. Thus, let us present
another idea which can be useful in many cases.
Let the values (4.122) be kept fixed and let only the first term in the series
(4.110) be taken into account (by which the rapidly oscillating terms are
eliminated). Thus
M(X, t) = u*{x, t) + r*(x, t) , (4.126)
104 Chapter 4
Let zto{^) be the function obtained when solving the problem (4.102) to
(4.104), with «o(^) replaced by Si(x), by the method of discretization in time
for the values (4.122). Since
where, by (4.128),
^o(^^^4sin(2n + l)x ") ^^^^^^
n 2/1 + 1
* *) Thus constructed by the method of discretization in time, at the same point t = 0.4.
*^) From the convergence Theorem 13.2, p. 255, it follows that the sequence of functions
(4.138) converges, in L2{G), for m-^co, to the function (i>„(x,OA). Consequently,
the sequence of corresponding norms converges to the norm of that function.
'^) In (4.128), the lower index of summation is equal to 2, not to 1.
106 Chapter 4
we obviously have (cf. (4.99) for the form of the functions (4.138))
1 1 ^'^)
': < ': '' (4 142)
[1 + 0.005(2n + Iff [1 + 0.01(2/1 + 1)^]*° ' ^ ' ^
n=i2n + 1
1
[1 + o.oi(2.n + xfY" rr^ ^ (2« + ifY^y-'
f-^'^TT
Linear Parabolic Problems 107
Taking (4.119) into account, this estimate is not very satisfactory (although
many times better than the one discussed in the text following (4.123)).
Nevertheless, it gives an idea what to do and what to expect in the case that
b) A NONTRIVIAL EXAMPLE
cu 4 +Z!\^W4 4 - ^ ) ^ 1 = 1 - ^ - - ^
c't ijdx^ V 25/^/J 16 4
in Q = G X (0, 1), (4.146)
u{x, y, 0) = 0 , (4.147)
M = 0 on r X (0, 1), (4.148)
where G is an ellipse with half-axes a = 4, ft = 2 in the coordinate axes
X, y; r is the boundary of G.
Equation (4.146) is an equation of the form
du
+ Au =f, (4.149)
dt
with
d_ y^\ du'
Au = 4+ (4.150)
dx S J 8x J dy LV 25; dyj
and
(4.151)
16 4
The method of discretization in time leads to the problems
^2\
with
zoix) = 0 (4.159)
where
y^\dvdu, /, x^\dvdu~\, ,
((f, «.))=£[(•4 + —8 /1 ex— dx—^ + I 4 -f 25j
—1^
dy dy_
i dx d>'
(4.160)
and
x^ v^
f{x,y)=l-^-y- (4.161)
16 4
It is not difficuh to prove that the form (4.160) is bounded in W^^\G)
and F-elhptic:
4 + —, 4 + —
8 25
in G. (See Chap. 2, p. 62.)
with
i[(IJ^(l^)] — I J t;^(x, >-) dx d j , (4.162)
« =4 + 4 (4.163)
'l '2
where l^, or / j . is the length of the projection of G into the x-, or j'-axis,
respectively. In our case
/i = 8 , /2 = 4 ,
so that
D _ J_ , J_ _ _L (4.164)
" 6 4 "'" 16 64
^*) Derived for v e C'''(G). However, the set of functions which belong to C'-^^(G)
and satisfy v = 0 on /" is dense in W'j''(G), so that this inequality remains valid for
ve fV^^\G).
110 Chapter 4
Thus
207t^
{{v, v)) ^ 4 dxdv > (4.165)
l\dxj \dy) 64
and
{{v, v)) = \{{v, v)) + \{{v, v)) ^
>
±whm-- ,2 /^.A2-l
+
10^2
64
\vV >
y
1071^ IOTT^
>
64 w+ 64
l»^2>"(C) • (4.166)
In this way, F-ellipticity of the form (4.160) is proved. At the same time,
we have found the constant of positive definiteness C:
{{{v,u)))^.{{v,u))+\{v,u) (4.168)
h
(Lemma 3.1, p. 78). Moreover,
so that (Theorem 3.2, p. 80) each of the problems (4.157), (4.158) (with
(4.159)), written in the form
ZjeV, (4.170)
j = l,...,p, has precisely one solution. Further, the form (4.160) and,
consequently, the form (4.168) being symmetric, the Ritz method can be
applied.
^^) This is why we have applied (4.162). Otherwise, we could have applied directly
(2.145), p. 41, with u = 0 on T.
Linear Parabolic Problems 111
"-i^-i-'i)- (4.172)
([3], p. 228). Take only the first of these functions, i.e. the function
Vy = \ (4.173)
16 4
(which obviously satisfies the condition v^ = 0 on t), and assume the appro-
ximate solution z*{x, y) of the first of the problems (4.170), (4.171) in the
form
x" y^
A{x, y) = aii^ - (4.174)
16 4
In this case, the Ritz system is reduced to one equation only,
lii-mH^wm dx dy +
+ - +
h
1 '•'"f/^ 4^2 sin^ (p
+ -
h L( 1 _ ^ _ Z_\ dx dy = 4 . 2
16 4
112 Chapter 4
+ - . 4 . 2 f f "(1 - r^)^rdrd(p =
^ JoJo
„ /TT 1 71 16 jr\
= 8 - + .-- + T: + .- +
V4 6.84 6.25 4/
+ ^/^i _ ? + l y 27: = 32.2171 + i . 8.3776 (4.177)
h\l 4 6j h
(rounded off to four decimals).
Further,
Put the just obtained function (4.181), instead of Zi(x, y), into (4.171)
written for j = 2. We obtain:
Z2 6 V, (4.182)
the whole procedure remains unchanged with the only difference that
instead of (4.178) we get (because z* = 0.009 63/according to (4.181))
(4.189)
E r r o r e s t i m a t e . As concerns the algorithm of the method of discreti-
zation in time, it remains, itself, relatively simple even in the just treated
more complicated case. As to the error estimate, the procedure is not so
simple any more, the errors becoming cumulated when using both the
approximate methods (of discretization and the Ritz method) simultane-
ously.
114 Chapter 4
Let us estimate, in this section, the error of the first and of the second
approximation, i.e. the numbers
\\u{x,y, 0.01) - z*(x,y) (4.190)
and
\\u{x, y, 0.02) - zj(x, >')|| . (4.191)
First of all, we have
\\u{x,y,Om)-zt{x,y)\\ ^
^ \\u{x, V, 0.01) - zi(x, >')|] + ||zi(x, y) - zf(x, y)\\ . (4.192)
Since, obviously, / e F ( / = 0 on T),
--[H)i-iyHH-':). e L,{G),
(4.193)
and (v, Af) = {(v,f)) for all v e V, the estimate (4.16) can be applied with
C^ = 3.084 by (4.167). Further,
+ — •- + 4 + dx d>' =
25/ 2
16
= 4.2 r^ sin^ (o + 2 -\ r^ cos^ (p r dr dcp =
Jo Jo [ 2 64 50
= 169.3637 ,
so that
M = 13.014.
For h = 0.01 and t = 0.01, (4.16) gives
||M(X, y, 0.01) - zi(x, >')|| g
where
A'z* = Az* +- z*
h
and
C +-
since
(((.,.))) = ((.,r)) + i(.,.)^(^(C^ +
Now,
dx dy =
aj . 169.3637 + ( ^ - i j . 8.3776 +
so that
\\A'z* - f\ = 0.075 . (4.197)
Thus, according to (4.195), we have
0.075
< 0.000 75 (4.198)
103.084
116 Chapter 4
= z Z2>-
F 2 - Z2 ^ F i - z (4.203)
Linear Parabolic Problems 117
(4.205)
Further,
||z2 - Z2II ^ |zi - z*\\ ^ 0.000 75 (4.206)
according to (4.203) and (4.198). To get an estimate for the last term in
(4.204), we use (4.195), where z*{x, y), or / ( x , y), is to be replaced by
ztix, y) = 1.963zf(x, y) (cf. (4.188)), or
| z t + y = 1.963/,
h
respectively. Thus we obtain, in this case, precisely 1.963-times the norm
ofzi(x, y) - z*{x,y),
\\z2 - z*|| ^ 1.963 . 0.000 75 < 0.0015 .
For the total error estimate we thus get
|lu(x, y, 0.02) - M*(x, y)\\ < 0.0014 + 00008 + 00015 =
= 0.0037 . (4.207)
Similar estimates can be obtained (in a relatively simple way, in our case),
for higher / s (j = 3,4,...). For example, for j = 3, we get, keeping the
same notation as before,
||u(x, y, 0.03) - z*(x, j;)|| ^ |«(x, y, 0.03) - Z3(x, y)\\ +
+ \\z,{x, y) - Z3(x, y)\\ + \\z,{x, y) - z*(x, >-)|! , (4.208)
where, in this case,
Iks - h\\ ^ \h - AW ^ Ik. - z,|| + i z , - z*!| ^
^ | z , - z\\\ + \\z2 - z*|i , (4.209)
etc.
118 Chapter 4
(„ = -£[(. + ,)£]
A(t) = («,3)
and
f(t) = {I + t) sin X . (4.214)
The solution of this problem is
u(x, 0 = (1 - e-*'' + ^'>'^) sin X . (4.215)
Thus the example is an illustrative one. It was chosen in order that its
approximate solution obtained by the method of discretization in time could
be compared with the exact solution and the efficiency of the error estimate
(15.147), p. 303, discussed.
Let us, first, verify the assumptions from Chapter 15. We have
V= {ve W^^\0, n), v{0) = 0, i<7i) = 0 '^)} = W^'\0, n), (4.216)
Thus
K = 2 (4.220)
can be taken in (15.12).
a= i (4.222)
in (15.13).
[{l + t + h) - (1 + 0] v'u' dx
etc., so that
K2 = 0, Ki = 0, K4 = 0 (4.225)
in (15.19), (15.134), (15.135).
(4.228)
Thus (see (15.22)),
||/|| = max ||/(f)|| = max (l + t) Isin x|| = 2.506 628 2 (4.229)
tsl tsj
and
||/(/)||^ = (1 + t) ||sin X[|K < 3.544 907 6 , (4.230)
Ml = 3.544 907 6 (4.231)
(see (15.139)).
Further,
\\f{t + h)-f{t)l ^ [(l + f + / , ) - ( l + 0]||sinx|| ^ ,, .^ ,, ^
Linear Parabolic Problems 121
I.e.
(4.232)
d = |]sinx| = 1.253 314 1
Thus, all the assumptions imposed on the form ((y, u)), and on the func-
tion /(r) in Chapter 15 are satisfied, so that there exists exactly one weak
solution of the given problem in the sense of Definition 15.1, p. 297. Con-
vergence of the method of discretization in time is ensured in the sense
of Theorem 15.1, p. 298. Moreover, the error estimate (15.147) is valid.
Zo(x) = 0 (4.238)
according to (4.211).
122 Chapter 4
(4.248)
whence
(l + h + 2h') A2 = i + h + 2h^ ,
I + h + h^
and
A, = 1
{i + h + h^){\ + h + 2/1^)
.,^1 sin X ,
' L (! + /! +h^) (1 + /i + 2/z^)... (1 + /i + 77i^)_
;• = 1 , • • • , ; ' (4.250)
In the following table, the functions Zj(x), j = 1,..., 10, are presented
for /) = 0.1 together with the functions u(x, t^, for comparison.
TABLE 4.1
As seen from Table 4.1, the results obtained by the method of discretiza-
tion in time are quite satisfactory. However, this cannot be said about the
error estimate (15.147), p. 303. In fact, according to (15.147), we have
\\u{x, tj) - zj{x)\\ ^ ^cjh' . (4.251)
Here, c^ is computed in the following way (see Chapter 15):
c. = Tll/ll .
C2= /r^(i/!l+ra)e^'^/"l
+ C3(2d + ra,)1e^*^'^4,
d = 1.253 314 1 ,
^2 = 0 ,
^3 = 0 ,
Ml = 3.544 907 6 ,
Mi = 5.013 256 4 .
According to the above formulae,
Ci = 2.506 628 2 ,
,^ = / p - 2 . 5 0 6 628_2 ^^^^^ ^^8 2 + 1.253 314 l) e^l =
= 16.690 142 ,
C3 = 7{2[2.506 628 2^ + 1.253 314 1^ 4- 2 . 16.690 142^] e^} =
= 91.374 358 ,
In the concluding part of Chapter 15, it is pointed out that the estimate
(4.251) is effective for "short" intervals only, and the reason of this fact is
explained there. Let us note here that (4.251), with c^ given by (4.252), is
valid for the whole interval [0, 1], since we have considered T = 1 v/hen
computing c^,..., c^. To obtain an error estimate valid at the point t = 0.1
itself, it is possible to put T— 0.1. When doing this, the efficiency of the
estimate in question will become substantially better.
First, some of the above constants will assume different values (because
1 + f ^ 1.1 then). In particular, we get
K = 1.1 ,
ll/ll = 1.378 645 5 ,
Ml = 1.949 699 1 ,
M2 = 1.516 510 0 .
The other constants remain unchanged. In this way, we get (not going into
more details)
Ci = T||/|| = 0.137 864 5 ,
C2 = 1.006 482 5 ,
C3 = 3.158 452 9 ,
C4 = 13.404 495 ,
c, = 19.589 76 .
According to (4.251) we then get the new error estimate
||M(.X, 0.1) - zi(x)|| g i . 0.1' . 19.589 76 = 0.097 948 8 . (4.255)
Comparing with (4.253), we see that this new error estimate is more than
1300 times better than the previous one.
Chapter 5
A Nonlinear Problem
(5.4)
1 X
Fig. 7.
') Notation used in the prceeding chapters is preserved. Thus the interval / = [0, T] =
= [0, 1] is divided, by the points of division tj, into p subintervals of lengths h = TJp.
ni
128 Chapter 5
of the problems
- z ; + 20zj + l ( z , - z ; _ , ) = 0 , (5.5)
h
z,(0) = 0 , z,(l) = 0 , (5.6)
with
Zo(x) = Mo(x) (5.7)
according to (5.2). In the weak formulation, the functions
Zje V (5.8)
are to be found, successively for j = 1, ..., p , satisfying the integral iden-
tities
{{v, zj)) + ~ {v, zj - zj_0 = 0 for all veV, (5.9)
h
with
2o(x) = Uo(x), (5.10)
V^{v-ve Wl'>(0, 1), z;(0) = 0, v{i) = O}^) = \\f\Q, l ) , (5.11)
Fu =\ u'^ dx 4- 5 M* dx (5.13)
0 Jo
'') The considered equations are of the second order, boundary conditions are Zj{0) = 0,
z/l)=0.
^) Multiplying, in (5.5), by an arbitrary function VE W'JHO, 1), integrating over the
interval (0,1) and applying integration by parts, in the usual way, one gets
which has, on F = rt'i'X^' -^)' ^'^ ^^^ needed properties and for its second
Gateaux differential the inequality
F'(r, t), i^) ^ cjjt'll^ (5.14)
is fulfilled. It then follows (cf pp. 316,317) that each of the problems (5.8), (5.9)
is uniquely solvable and that the function Zj minimizes, in V, the functional
1 C-
Hju = i u'^ dx + 5 u*dx +
-f"^
2h
'•h],
dx
''•'•' (5.15)
To the numerical solution, the finite element method will be applied,
in the modification of Remark 16.8, p. 329. For its convergence in the indi-
vidual problems (5.8) —(5.9) see Lemma 16.2, p. 328, and the text which
follows it, for convergence "in the whole" (the "Ritz-Rothe" method) see
Remark 16.9., p. 330.
V
i vJxl
v^fxt
' v,(x)
4 A • «
/ \ '1 /*
/ \ / \ / \
! / A A \
i // / \\ i'"^^ \I
/ # \ / ' •
0
^.- f y V—A X
Fig. 8.
z_,_i +f\udx.
In our case, we have f(x) ~ 0. On the other hand, UQ(X) ^ 0 is considered here (see
Remark 16.7, p. 327).
130 Chapter 5
^ = 0, ^ ^ - = 0, ^ = 0 (5.20)
da^i dai2 Sa^^
for the unknown coefficients a^, «i2> ''is ^^ (5.19). In consequence of
Lemma 16.2, p. 328, and of the linear independency of the functions (5.16),
these coefficients are uniquely determined.
Having obtained the function Zi(x), we put it in place of the function
Zi(x) into the functional Hj (thus into (5.15) written for j = 2). Let us denote
the obtained functional by H j . Instead of minimizing this functional on the
space V we shall do it on the space F3, thus assuming the approximation
z*{x) of the minimizing element in the form
Having obtained z*(x), we put it in place of Z2(x) into (5.15), written for
j = 3. The approximation Z3(x) of the minimizing element of the obtained
functional H^ on F3 will be assumed in the form
In fact, in (5.18) the coefficients at the functions Vi{x) and v^ix) are the
same. Replacing then u in (5.15), written for; = 1, by b^v^ + CjUj + d^v^,
the obtained function becomes symmetric in b^, d^, as seen from the form
of the functional (5.15) and from what has been said about Zo(^) = "o(-'f) •
Thus, minimizing that functional we get d^ = fci, etc.
1 f {bjVi
+— + CjV2 + bjv^f dx -
1 r^ dx.
- 7 (bjVi + CjV2 + bjV3){bj^iVi + Cj_iV2 + foy-ifs)
''Jo
(5.25)
etc. Further,
+
'1
+ ^[^(2b,^ + cj)j^?dx + 46,c, 0
1^11)2 dx
y* d x = ~ , vlvl d x = l i o , '^^'2 d x = ~
0 Jo Jo
yf dx = 8 , yjt)2 dx = — 4 . (5.29)
Choosing
h = 0.01 (5.30)
a n d inserting (5.29), (5.30) into (5.28), one obtains
Hjz* = b* + i 4 + ^bjcj + ib^bj + ibjc^ +
+ (8 + '-Dbj + (4 + f ) c j + (-8 + '^)bjcj -
- 100(ib;_l + ^ C ; _ 0 fc, - 100(^'lb;-l + i C , _ i ) C; . (5.3l)
A Nonlinear Problem 133
^ = 0, ^ = 0. (5.32)
dbj dcj
This system being nonlinear, its solution is rather labourious. For this
reason the method of minimization of Hj mentioned in Remark 16.8, p. 329,
will be applied (with one iterative step only): The values bj_i, Cj^i in (5.31)
being known, Hj is a function of two variables bj, Cj. The mentioned
iterative process consists in the following: Let us start with an arbitrary
vector
ib°, c°). (5.33)
Keep Cj fixed and minimize the function Hj{^i, c°) of one variable iJi only.
Obtaining b] in this way, minimize the function Hj{bj, ^2) of the variable
^2 to obtain c]. Then the function Hj{£,i, c]) is minimized, etc. Under the
assumptions of Lemma 16.1, p. 312 (which are fulfilled in our case), this
process is convergent.
bo^\, Co = i (5.38)
from (5.18). Substituting them for bj.y, Cj_i into (5.35), written for j = 1,
the equation for b^ is obtained:
It yields^)
bi = 0.245 38 . (5.40)
Then
bj-x = io = i , cj..^ - Co = i , bj = b,= 0.245 38 (5.41)
' ) The just described process gives approximate values of bj and Cj, thus approxim-
ations of the functions z* only. Nevertheless, we preserve the original notation bj, c,,
2*(x) for the sake of simplicity.
*) Uniquely, in accordance with the theory. (Note that the derivative of the left-hand
side of (5.39) is always positive, so that the considered equation has actually exactly
one real root.)
A Nonlinear Problem 135
giving
i)2 = 0.229 1 1 , C2 = 0.350 3 8 ,
so that
z* = 0.229 l l r i + 0.350 38i;2 + 0.229 Uv^
10 DIM B[30],C[30]
20 DEF FNF(X) = A0*Xt3 + A l * X t 2 + A 2 * X + A 3
20 DEF FNG(X) = A4*Xt3 + A5*Xt2 + A6*X+A7
40 WRITE (15,50)
50 FORMAT SX.'AO Al A2 A3 B A4 A5
A6 A7 C
60 B[l] = 0.25
70 C[l] = 0.5
80 FOR J = l TO 15
90 U1 = U4=0
100 U2=U5=1
110 A0=48
120 A1=18*C[J]
130 A2=12*C[J]t2+592
140 A3 = 6*C[J]f3 + 4*C[J]-100*(4*B[J]+C[J])
150 V1 = FNF(U1)
160 V2 = FNF(U2)
170 IF V1*V2<0 THEN 190
180 DISP "CHANGE INTERVAL U1,U2";
190 U3 = U1-(U2-U1)/(V2-V1)*V1
200 V3 = FNF(U3)
210 IF V1*V3<0 THEN 240
220 U1 = U3
230 GOTO 250
240 U2=U3
250 IF ABS(V3)<0.000001 THEN 270
260 GOTO 150
270 B[J+1]=U3
280 A4=24
290 A5=18*B[J+1]
300 A6=12*B[J+1112+296
310 A7=6*B[J+l]f3 + 4*B[J+l]-100*(B[J] + 2*C[J])
320 V4=FNG(U4)
330 V5 = FNG(U5)
340 IF V4*V5<0 THEN 360
136 Chapter 5
;• Ao Ai Aa A3 bj
u{x, 0) = 0 , (6.3)
«(0, f) = 0 , «(],?) = 0 . (6.4)
Equation (6.2) is really of the considered form (6.1), with
137
138 Chapter 6
(cf. e.g. Example 4.1, p. 88). The form (6.9) is bounded in W^^\0,1) and
F-elliptic (cf. the quoted example). It remains to verify assumptions (17.14)
and (17.15), p. 337, concerning the function K[x, t, u):
(i) A positive constant C exists such that
l\K{x, t2, lia) - K{x, t„ u,)\\ ^ C{\t^ - t,\ + ||M3 - «,||) (6.10)
holds, and
(ii) ueV^ K{x, t, u) e L2(0, l) . (6.11)
To verify condition (6.11) is very easy: We have
K{x, t,u)= I - u^{x). (6.12)
If u e V, then u e H'j''(0, 1), and u{x) is a continuous function (even abso-
lutely) in [0, 1]. Consequently, the function (6.12) is continuous in [0, 1]
and the more it belongs to ^2(0, 1).
Further,
\\K(x, t„ M2) - K{x, t„ u,)|| = i|l - u l - 1 + ulW =
= ||("i + «2)("i - "2)11 ^ j|"i + W2I! \\uz - w,|! . (6.13)
We show that we can consider
M ^ 1 (6-14)
elsewhere, so that in (6.10) it is possible to put C = 2. More precisely, we
prove that \\z"\\ ^ 1 for every division d„. Then, elsewhere in the proof
of convergence (Chapter 17), it is possible to limit oneself to such u's in
K{x,t,u) for which ||u|| ^ 1 .
) We write
I,
0
[1 - u^{x, T)] dT X h{l - zl)=^ h Ko{x),
2/1
[1 - u\x, T)] di « /i(l - zl) + /i(l - zl) =
= h[K^{x) + K,{x)] ,
etc.
3
) This fact follows from theorems on the regularity of a weak solutions. Moreover,
in such simple cases as are the condsidered ones, the existence of classical solutions
can be established directly using the Green function.
On this occasion, let us point out an additional useful property of the method
of discretization in time: It reduces the solution of parabolic problems to the solution
of elliptic problems which are often sufficiently simple to render an adequate notion
as concerns the behaviour of the original problem.
140 Chapter 6
solution at the same time). Then the same holds for the problem
- --;'(-V3) ^ 0 (6.25)
and by (6.18)
M = Zi(x3) g h^ . (6.26)
and, in general,
Thus, we have
lk"l ^ 1 (6-30)
"We shall actually apply the Ritz method here. In fact, although the
problem (6.18), (6.19) is very simple, the formula for its solution Zi(x) is
rather cumbersome (cf. Example 4.5, p. 100). Solving then the problem
(6.20), (6.21) for Z2{x) and successively the further problems, precise solu-
tions would be obtained in a very difficult manner.
*) Denoting
M = max Z2(x)
ie[0,l]
([3], Chapter 20), and take the first three terms, thus assuming approximate
solutions z*(x) of the problems (6.15), (6.16) in the form
z* = a*/' sin irx + 02" sin 2nx + a^^ sin 3rrx . (6.32)
Before passing to the numerical procedure, let us state that the problem
(6.18), (6.19) is "symmetric with respect to the point x = 1/2". In fact, the
left-hand side of (6.18) contains derivatives of even orders only, and its
right-hand side as well as the conditions (6.19) are symmetric with respect
to the point x = 1/2. Thus, the solution Zi(x) is a function symmetric with
respect to that point. It follows that the right-hand side of the equation
(6.20) will have the same property, and so will the function ^2(^)5 etc.
Thus, all the problems (6.15), (6.16) are symmetric with respect to that
point. Consequently, the middle terms in (6.32) can be dropped.*) In addition,
simplifying the notation, we shall consider the Ritz approximations z*{x)
in the form
zj = bj sin Kx + Cj sin 3KX . (6.33)
In the weak formulation, the problems (6.15), (6.16) read (for Fsee (6.8)):
To find, successively for 7 = 1,..., p, functions
2jeV,
((r,«))=jV«' dx
ric, and so is the foi
is symmetric, form
1 C^ If
{{{v, u))) = {(v, u)) + ~{v,u) = v'u' dx -I- - vu dx .
h Jo /ijo
Thus, the Ritz method can be applied.
*) Of course, it is possible, to investigate them, without any use, however, because all
the coefficients a^^ will be found to be equal to zero.
An Integrodifferential Parabolic Problem 143
The Ritz system for the unknowns bj, Cj is then of the form
= (smnx,]-z*_,+hl\l-zr)),
\ h i=o /
Now
4
sin^ 7tx dx = — , sin^ 3rtx dx =
Jo 37t Jo 9n'
36
j sin^ nx sin 37tx dx = , I sin nx sin^ 37tx dx =
Jo 153t Jo 3571
2v'^ n r - i . ^ - ' + T -
bi fo.C; + —,
n i = o \3 15 35
1 / . A 1 2jh
-l9n' + ~]cj = — C;_i + - ^ -
2V V 2/7 • 37C
Thus, starting with the values (6.40), bj, c,- are computed, successively,
by (6.39).
An Integrodifferential Parabolic Problem 145
50 + ^ V ^ = — '
2/ u
,„ 97i^\ 0.02
50 + — ci = ,
2 / 371
givmg
bi = 0.000 115 9 , Ci = 0.000 022 5 ,
so that
z\ = 0.000 115 9 sin 7CX + 0.000 022 5 sin 37tx .
( 50 + — j ^2 = 50 . 0.000 115 9 + — -
n
Below, numerical results for hi, Ci,..., b2o, C20 are presented, as obtained
on the Hewlett-Packard A 1030 computer, with a program in Basic Fortran.
146 Chapter 6
For example, at the point t = 0.1 (i.e. f o r ; = 10), the "Ritz-Rothe appro-
ximation" z*o(x) is
z*o = 0.004 929 1 sin nx + 0.000 424 1 sin 2KX .
10 DIM B[30],C[30]
20 WRITE (15,30)
30 FORMAT 7X."J",15X,"B",10X,"C"
40 B[1]=C[1]=0
50 FOR J = 1 TO 20
60 T=S=0
70 FOR 1=1 TO J
80 S = S + 4/3*B[I]t2-8/15*B[I]*C[I] + 36/35*C[I]t2
90 T=T-4/15+B[I]t2+72/35*B[I]*C[I] + 4/9*C[I]t2
100 NEXT I
110 B[J4-l] = (50*B[J]+0.02*J/PI-0.01*S/PI)*2/(100+PIt2)
120 C[J+l] = (50*C[J] + 0.02*J/3/PI-0.01*T/PI)*2/(100+9*PIt2)
130 WRITE (15,140)J,B[J+1],C[J+1]
140 FORMAT 5X,F3.0,9X,2F11.7
150 NEXT J
160 END
M(X, 0) = 0 , (7.3)
M(O,0 = O , ^ ( 7 1 , 0 = 0 . (7.4)
ox
From the theoretical point of view, problems of this kind are treated
in Chap. 18. From the practical point of view, the considered problem
describes a problem in heat conduction, where the intensity of inner sources
of heat depends also on the amount r of heat already developed. In parti-
cular, such a case is encountered in so-called problems of hydratational
heat. In these cases, the right-hand side of the considered differential equa-
tion is frequently of the form (7.1), thus decreasing exponentially with time t
and depending "slightly" on the temperature u and on r (the coefficients
at u and r are small in comparison with the coefficient at i).
A = -f,, (7.5)
dx^
/(x, f, «,r) = e-'-°-°>"-°-"% (7.6)
147
148 Chapter 7
' ) Even in the ordinary sense since a function from ff ;J''(0,7t) is absolutely continuous
in [0, jt].
The second of the conditions (7.4) does not appear in (7.8) because it is unstable
with respect to the operator A.
Thus all the requirements of Theorem 18.1, p. 376, are fulfilled. Conse-
quently, precisely one weak solution u{x, t), r{x, t) exists in the sense
of Definition 18.1, p. 374, and convergence of the method of discretization
in time is ensured. Moreover, the form ((i;, w)) being symmetric, the Ritz
method can be applied (see below).
- z ; ' + | z i = i z o + / o = l, (7.18)
h h
zi(0) = 0 , z;(jt) = 0 . (7.19)
In this way, we get the function Zi(x). At the same time,
r,{x) = hf,^h (7.20)
by (7.16). Thus, according to (7.15),
f,{x) = e-'--o-o>^'W-o-""<-). (7.21)
150 Chapter 7
h h ^ ^
2,(0) = 0 , z',{n) = 0. (7.23)
For the Ritz method, we choose the first two terms of the base
2 2 2
X 3x
fi = sin - , V2 = sin — . (7-31)
Thus, the Ritz approximations z*{x) of the functions Zj{x) will be of the form
z ; ( x ) = Z);sin^ + c , s i n ^ (7.32)
and the corresponding Ritz systems of the form
r*(x)=/,[/o*(x) + . . . + / ; L i ( x ) ] , (7.36)
(thus.
/*(x) = /o(x) = e° ^ 1) . (7.38)
NOW,
1 f" . , X , 71 TT
(7.40)
H — I sin - d x = — I ,
T 2 8 2/i
and similarly,
(7.41)
8 2«
{{{vu V2))) = 0 + 0 = 0 ; (7.42)
f X
(ui, 1) = sin - dx = 2 (7.43)
Jo 2
. 3x , 2
sin — dx = -
(-^a, 1) = 2 3 (7.44)
In what follows, let us consider
h = 0.01 .
(7.45)
Then (7.39)-(7.44) yield
40l7r
b,=2,
4097t (7.46)
giving
bi = 0.0127 , Ci = 0.0042 . (7.47)
Thus,
z*(x) = 0.0127 sin - + 0.0042 sin — (7.48)
^ ' 2 2
At the same time,
by (7.35).
When continuing in this line, the evaluation of the scalar products
in (7.33) would be rather lengthy. At the same time, the functions / / ( x ) are
"almost independent" of x because of the small coefficients 0.01 and 0.03
at the functions z*(x) and r*{x) in the exponent of (7.35). Thus, we introduce
the following simplification in what follows, without danger of committing
a considerable error: Having obtained a Ritz approximation we put in (7.35),
instead of this function, its mean value in the interval [0, n\. In this way,
we get the following process (we denote the obtained functions by Zj{x),
fj[x),fj(x) instead of the original functions denoted by a star):
where
{i{v„v,)))Bj + {iivuv,)))cj = ( ^ 1 , ^ ^ , - 1 +fj_,y (7.52)
1
zXx)dx, (7.54)
7 _ g-jfc-0.01rj-0.03fj (7.55)
f, = /,(/o + . . . + / , _ 0 , (7.56)
fo = 0 (7.57)
(thus,
/o = 1). (7.58)
154 Chapter 7
For the functions (7.31) and for h = 0.01 from (7.45), the system (7.52)
becomes
40l7r
Bj = {v„100zj^, +Jj^,)..
40971
cj = {v„mzj_, +fj^,). (7.59)
Now,
f . 3x ^ 2 ,
(.„/,_,) = / , - t sm — dx = - / , _ i . (7.63)
0 2 3'^ '
Putting these results into (7.59) and noting (7.55), (7.56), we obtain the
following recursion formulae:
(7.64)
40lTt
(507rc_,-_i + - | / j _ i ) , (7.65)
4097r
-J ' ) (7.66)
) Note that/y is a constant in the just considered process — this was the aim of the
chosen modification of the original process.
'*•) By (7.54) and (7.43), (7.44).
Problem with an Integral Condition 155
r, = O.Ol(/o + . . . + / , _ i ) , (7.67)
j = 1, 2,..., p with
5o = 0 , Co = 0 , /o = 1 , ro = 0 . (7.69)
Numerical results for hj, Cj, Zj, T-j and /,• are presented below for j = 1, ...
...,20, obtained on the Hewlett-Packard A 1030 computer with a program
in Basic Fortran.
For example, at the point f = 0.1 (i.e. for j = 10), we have
X 2x
z.Jx) = 0.118 820 sin - -f- 0.035 538 sin — ,
2 2
fio = 0.094 630 .
10 D I M B[30],C[3O],F[3O],R[30],Z[3O]
20 WRITE (15, 30)
30 FORMAT 5X,"J",UX,"B",11X,"C",11X,"Z",11X,"R",11X,"F"
40 B[1]=C[1] = R[1] = 0
50 F[l]=l
60 F O R J = l TO 20
70 S= 0
80 F O R 1 = ] TO J
90 S = S + F[I]
100 NEXT I
110 R [ J + 1 ] = 0.01*S
120 B [ J + 1 ] = (2*F[J] + 50*PI*B[J])*8/401/PI
130 C [ J + l ] = (2/3*F[J]+50*PI*C[J])*8/409/PI
140 Z[J+l] = (2*B[J+l]+2/3*C[J+l])/PI
150 F[J+1]=EXP(-0.01*(J+1)-0.01*Z[J+1]-0.03*R[J+1])
160 WRITE (15, 1 7 0 ) J , B [ J + 1 ] , C [ J + I ] , Z [ J + 1 ] , R [ J + 1 ] , F [ J + 1 ]
170 F O R M A T F7.0,7X,5F12.7
180 NEXT J
190 END
OVDOO-^a\t-^-P>'U
p p o p o o o o
'O'—'>—'i—^O'./i'—'^
• - - I — ' O O b J N J ^ N J O
O s C O O O t - f t ^ h J O O
p p o o o o o p
o o o b b b o_o
O O O C T N O J H - ^ O N -
o oooo ooo
l^ U U U
L/i ^ Ji.
4:^. 0 0 K-i
p p p p p p p p
^-^^UUU^U
'*0'—'OJ^yiC^OO'O'—
pppppppp
bo bo boboboooboo
O — K>O->'^4a.Crta
v O U J ^ ^ ^ J 0 0 ^ - / ^ O J ^
-J — L / i - ^ C 0 4 ; ^ ^ 0
Chapter 8
Hyperbolic Problems
u{x, 0) = 0 , (8.2)
- ( x , 0 ) = 0, (8.3)
dt
u(0, 0 = 0 , u{n, f) = 0 . (8.4)
The solution is
Au = - « " . (8.6)
157
158 Chapter 8
feV, (8.9)
and the error estimate
| | u ( x , / , ) - z , ( x ) | ^ V ( i ) ] | / | | w t ; + l)/'= (8.10)
(see (19.164)) can be appHed. Here
= sin X , (8.12)
*) The functions from W^^\0. n) being continuous (even absolutely) in [0, Jt], the
conditions v(0) = 0, 11(7:) = 0 are satisfied in the ordinary sense.
^) Multiplying Au = —u" by an arbitrary v eV, integrating between x = 0 and x = K
and applying integration by parts with the use of vifi = 0, v(ii) = 0. See Chapter 4.
Hyperbolic Problems 159
with
zo = 0 (8.14)
by (8.2) and
1\ 1
^1] 1 +
\ h'J 1
A,= , (8.19)
2(1 +h')' ^ ^
and
z, = sin X . (8.20)
2(1 + h') ^ ^
It is easily seen that the function Z2(x) can be assumed in the form A2 sin x
again, etc. In general, assuming Zj(x) in the form
Zj = Aj sin X , (8.21)
(8.12) yields
^; 1 +
or
1
Aj = ^-^^(lAj_,-Aj_2 + h'), (8.23)
160 Chapter 8
with
^0 = 0 , .4_i = ^ . (8.24)
TABLE 8.1
In Table 8.1, the functions 2i(x),..., Z^Q^X) are presented for h = 0.01
simultaneously with the functions u(x, O.Ol), ..., u(x, O.IO) computed by
(8.5).
Obviously, the results obtained by the method of discretization in time
are quite satisfactory. Note that the problem (8.1) —(8.4) is typical for
hyperbohc problems.
Let us turn to testing the efficiency of the error estimiate (8.10). If, for
example, J = 1, this estimate gives (see (8.11))
^(x,.v,0) = 0, (8.31)
ct
M = 0 on r X (0, 1), (8.32)
Here,
A = —J,— (4 -I- sm X sm y) — + dy (4 + sin X sin y) — i ,
ox
(8.33)
162 Chapter 8
In this way, F-ellipticity of the form ((v, w)) is proved. Its symmetry
is evident. Thus the method of discretization in time can be applied, the
existence of the solutions of individual eUiptic problems being ensured as
well as the convergence of the method. (See Chapter 19.) At the same time,
the coefficient of positive definiteness is found in this way, because by (8.36)
{iv,v))^C^lvr (8.38)
with
C^ = 8.
Applying the method of discretization in time, we have to solve, successi-
vely for J = 1, ..., p, the problems
or, denoting
2z,-_i - zj^^
^•-^*~. n- ;,. (8.40)
the problems
A'zj =-- / ; , (8.41)
with the boundary condition
Zj = 0 on r (8.42)
and with
/(x, y) = sin x sin >', (8.43)
of the sine base (25.26) from [3]. Thus, let us assume the Ritz approximation
z*{x, y) of the function Zi(x, y) in the form
Zj = Qi sm X sm y . (8.50)
1 1 (4 + s i n X sin ^ ) r ( ^ Y + ( | ^ Y 1 d x dy
l£|Wxsin^,dxd, = 4.(^ + ^) +
+ - . - + - . - + — - = — + 271^+—- (8.52)
,3 3 3 3/ Ah"- 9 4/i'
Further,
Thus
n'
8
fli — , • (8.54)
— + 27C' + -—
9 4h^
In particular, for
/) = O.OI , (8.55)
we get
n'
8
= 0.000 049 955 , (8.56)
— + 2n^ + 250071^
9
and
z* = 0.000 050 sin x sin y . (8.57)
Similarly, assuming z*{x, y) in the form
-2 = "2 sin X sin y , (8.58)
the Ritz method yields
{{{v„v,)))a,={v„n), (8.59)
where
= \ sin Xsin
si V. ( 1 H ^ I sin x sin v dx dy =
Jo Jo
/•it /•it
= (1 + 20 OOOfli) sin^ x sin^ y dx dy
Jo Jo
= (1 + 20 000ai). — . (8.60)
4
(8.59), (8.52) and (8.6O) with (8.56) then yield
02 = 0.000 199 730 . (8.61)
166 Chapter 8
Thus,
zt = 0.000 200 sin x sin j ^ . (8.62)
and both the terms on the right-hand side of (8.64) are to be estimated.
For the first one we have, using the analogue of (8.10),
A'zi ==fi
) We shall content ourselves with six decimals here — the reason will become clear
below.
Hyperbolic Problems 167
where
C'2 = C M - — ' = 10 008 (8.70)
+ sin X sin v ) ^ ^
dx
Now,
^ - 1 = -0.000 4 5 , (8.72)
h^ 2
(8.73)
0 < fli < 0.000 0 5 ,
thus, because of the values of the sine functions,
*n /*n
0 g \A'zX-f'f ^ (0.000 25 + 0.000 25 +
J 0 Jo
+ 0.000 45)^ dx dy = 0.000 95^ . n^ .
Thus,
\\A'z* - / ; ! ! ^ 0.000 95ii, (8.74)
((;;, zj)) + -*- (v, Z2 - 2zi -t- ZQ) = (vj) for all v e V, (8.77)
{{v, z,)) + ~ (v, Z2 - 2zt + Zo'))= (vJ) for all veV. (8.78)
Ir
In this way, a further error arises. We thus have
) ZQ = 0 in both cases.
Hyperbolic Problems 169
An Application to Rheology
170
Application to Rheology 171
with boundary conditions given by the bearing of the bar. Here, P is the
horizontal force, q{x) the vertical loading, £ the Young module of elasticity,
/ the moment of inertia of the cross-section of the bar with respect to the
bending axis.
In rheology, equation (9.6) becomes
with the operator F from (9.3). To the solution of this problem the method
of discretization in time will be apphed. Let us point out that our intention
in this chapter is not to build up a general theory of this method in rheology^).
Our aim is to obtain an idea about whether this method is suitable for
solving such problems or not. Therefore, a rather illustrative example
is chosen, with a simple law of creep given by
Thus, let us turn to the solution of our problem. Assume that the bar
is simply supported. Let us transform the length of the bar into the interval
[0, n] and let, first, the loading be given by a sine function. Thus, we have
to solve the problem
tj=jh, ;• = 0 , 1 , 2 , . . . . (9.12)
= (1 -e-'-)x'e^'-'^-^>^''cr,., (9.13)
1=0
whereCT,-stands for the approximation of u(t) at the point t = t^. If £j
stands for the approximation of £(/,) and if we denote, for brevity,
a = 1 - e"\ (9.14)
^) Note that
e^"'.
Application to Rheology 173
£,. = ^ + ^ ' X ' ' " • " " " ' " ' ^ i - (9-15)
E E i=o
Symbolically written,
E = F;'a, ;• = 0 , 1 , 2 , . . . . (9.16)
Having obtained, in this way, the "discrete approximation" of the opera-
tor f~^, we pass over to the discrete approximation of the operator F.
We show, first, the procedure for 7 = 0, 1, 2. If j = 0, then the sum
in (9.15) vanishes and we get
Eo = ^ . (9.17)
E
(At time t — 0, the law of creep becomes the classical Hooke law.) For
./ = 1, (9.15) yields
Ao = ~~^— (9.29)
IE - P ^ '
(we assume P < IE), and
C
Wo = sin X . (9.30)
IE - P
Thus, at time f = 0, we get the same result as in the classical theory
of elasticity. This is in accordance with (9.17), by which the classical Hooke
law is obtained for f = 0.
If 7 = 1, (9.25), or (9.26), turns into
IF^w^^^ + Pw'l = C sin X , (9.31)
or, by (9.21), into
/£(wi - awo)^'*' + Pw'i' = C sin X , (9.32)
respectively. Looking for the solution Wj(x) in the form
Wi = Ai sin X (9-33)
Application to Rheology 175
- ^ ' ^ = r (9.35)
IE - P
and taking (9.29) into account, we get, by (9.34)
W2 = A2 sin X (9.40)
^ = sj, ;• = 0 , 1 , 2 , . . . . (9.44)
E
176 Chapter 9
We assert that
J-i,
Sj=^Sj-aY,b'-'Ej_,, 7 = 0,1,2,.... (9.45)
we have (putting j = k + 1)
Sk + i - E.+ i - a{s, + e - \ _ i + ... + e^'-'^'s, + e ^ ^ ) . (9.48)
The assumption that (9.45) holds for all j ^ k then yields
s, + e~\_i + ... + e'l-'-'^Si + e - % =
= £4 - «(£*-! + bet-2 + fe\-3 + fe\-4 + ••• + b'^~*£o) +
+ e-''[c;,_i - a(e^_2 + be^.j + bh,_^ + ... fe'"'£o)] +
+ e"^''[ct-2 - a(£t-3 + be,_^ + ... + b ^ ' ^ ) ] +
+ e-3'-[£,_3 - o(e,_4 + ... + M - % ) ] +
+ +
+ e-'^'-eo = 8^ + £,-.,(e-'" - a) + e,_2[c--'- - a{t~'' + fe)] +
+ E.-sLe"" - ^ e - ^ " + e-'-Zj + fo^)] + ... +
+ eo[e-'''- - a(e-<''-i"" + e-'^-^"-^ + ... + fo"-!)] =
= £, + z,.,h') + 6,._2[e-^'- - (e-'- - b) (e-" + /,)] +
+ e.-sCe-'" - (e-'- - fo)(e-2'- + g-'-fo + b^)-\ + ... +
+ £o[e'"" - (e-" - b)(e-<^-i"' + Q'^'^-^^^'b + ... + b"-')] =
Then,
cj = E{ej - a t b'-hj_,) , j = 0, 1, 2 , . . . . (9.50)
wherefrom, by (9.53),
{IE - P) A J = C + IEa{Aj_, + bAj_. + ... + b^'^o) . (9.57)
Dividing by IE — P and using the notation of (9.35), we get (9.55).
178 Chapter 9
+ b''Ao . (9.60)
^) By (9.58).
Application to Rheology 17^
"•" f =h'-o
ft^o Ao ""^[2\vH
- T -2 ^
~ T ^''"^^ -T) + iT- J1)]4 =
Application to Rheology 181
T „<limln[e-''(2-r) + ( r - l ) ] / ) .
2 - T 2 - T
2 T
(9.73)
T 2 - T
by the I'Hospital rule.
Inserting the corresponding expressions for AQ and T, we obtain, finally,
^ _ ^^ l^ g-(J£-2P/J£-P)t /g J^\
IE - 2P {IE - P){IE - 2P)
and
w(x, f) = .4 sin X . (9.75)
It is not difficult to show that we have thus actually obtained the exact
solution of the problem (9.9) — (9.11). This solution can be received by
other methods as well (using the Laplace transform, etc.) in this simple
case.
, . 1.66C .
>Vi = /Ij sm X = sm x , (9.84)
IE
, . 1.95C .
vvj = Aj sm X = sm x (9.85)
and IE
4C .
iv^ = ^00 s'ri X = — sm X (9.86)
by (9.68).
Fig. 9.
The graphs of the functions (9.83) —(9.86) are sketched in Fig. 9. The
process of creeping is very well seen from this figure.
vvix, 01 =: sm X , (9.87)
IE
Application to Rheology 183
AC
w(x, oo) = — sin X . (9.90)
IE
For / = 0 and t = GO, the results obtained by the method of discretiza-
tion in time are identical with the exact solution. For t — 0.2 and t = 0.4
the difference is negligible.
Chapter 10
Introduction to Part II
185
186 Chapter 10
187
188 Chapter 11
-+ Z (-iy'+-+'--l ^.
8t i, + ... + iygk dxV...dx'^
The operators B,-, C,- are linear operators discussed in Chap. 2, the opera-
tors Bi containing derivatives of order gfc — 1 and thus stable with respect
to the operator A, the operators C; containing also derivatives of order
^ k (and ^2k — 1, of course) and thus unstable with respect to the operator
A. Moreover, the operators Bi are assumed to be of the form (2.214), p. 53,
and to satisfy the requirement (2.215). None of the operators B^, C; contains
derivatives with respect to t, their coefficients are independent of t. The
bihnear form
((t., u)) (11.8)
corresponding to the operator A and to the boundary operators B,-, Cj is
assumed to be bounded in the space ff j*'(G)^) and V-elliptic, with
1) Remind that wi''\G} is a Hilbert space the elements of which are functions having
generalized derivatives up to the order k including, with the scalar product
is the familiar scalar product in the space L2(G). For details see Chap. 2, p. 37.
The Equation dii/St + An = f 189
and
((r, t;)) ^ a|t;||^,,.,(C) for all veV.'f) (11.13)
As shown in Chaps. 1 and 3, the "classical" method of discretization
in time (the Rothe method, or the method of lines) consists in the following:
Divide the interval
/ = [0, T] (11.25)
by the points
tj^jh, j = l,...,p- 1, (11.26)
into p subintervals
/ , = [«;_!, f , ] , j = l,...,P, (11.27)
^) Let us recall, very briefly, some basic concepts discussed in detail in Chap. 2. Consider
the "corresponding" elliptic problem
Au = / , (11.14)
Bi«=0, 1=1,...,/;, (11.15)
Ci«= 0, 1 = l,...,k- n, (11.16)
where A is the investigated operator (11.5). In the weak formulation, the problem
(11.14)—(11.16) reads (cf. Definition 2.5, p. 61): To find such a function
ueV (11.17)
(because of the homogeneity of the conditions (11.15)) for which the integral identity
is satisfied. Here
is the so-called bilinear form corresponding to the operator A and to the operators B^, C;,
or, briefly, corresponding to the given problem. The term
is called the bilinear form corresponding to the operator A. It is given by the coefficients
of the operator A only, independently of the boundary conditions. (See however,
Remark 2.15, p. 60.) The term
a(v,u), (11.21)
190 Chapter 11
of lengths
h = - (11.28)
P
(denote this division of the interval / by d^) and find, successively for
j = 1,..., p, the functions Zj(x) as solutions of the problems
zo(^) = 0 (11.32)
according to (11.2).
dependent on the operator A and on the operators B^, C,-, is called the boundary form
corresponding to the given problem.
For details see Remark 2.15, p. 60. The term
yt{v,h) (11.22)
(see (2.236), p. 57) is not present here because of the homogeneity of unstable boundary
conditions (11.16).
In practice, the integral identity (11.18) is obtained if one multiplies the equation
(11.14) by an arbitrary function v e V, integrates over G,
(11.33)
h
(For details see Chaps. 1 and 3; see also Fig. 1, p. 2.)
In the weak formulation, the problems (11.29) —(11.31) read:
Starting with the function Zo(x) = 0, to find, successively for j = I,..., p,
the functions
ZjeV) (11.34)
are satisfied.
Here, ((., .)) stands for the above discussed bilinear form, correspond-
ing to the operator A and to the boundary operators B,-, C,- while ( . . . ) is
the scalar product in LiiG). (Instead of the integral identity (11.18), the
integral identities (11.35) are obtained here, of course, because of the terms
h h
in (11.29). Cf. Chap. 3.)
/+^6L2(G), (11.37)
h
one obtains immediately (cf. Theorem 3.2, p. 80) that the problem to find
ZieF (11.38)
such that
(((i;,z,))) = ( ' r , / + ^ ) V.EF (11.39)
Let us construct the function Ui{x, t), the so-called Rothe function,
defined for all ( e / by
/ and assumes the values z/x) at the points f = ?j. (Note that/^ — tj_i = h.)
See Fig. 10, where G = (0, n) and an jx e (O, n) is chosen to show the geo-
metrical meaning of w,(x. t) for this x.
ij = ['j-ut'n (11.44)
of lengths
K = (11.45)
Here
91.-1 (11.46)
t^^JK, ; = 0,1,
(where n stands for the superscript in /" and f", not an exponent, of course).
Similarly as above, we obtain, for every fixed n, the functions
911-1 (11.47)
';"eF, ; = 1, P,
satisfying the integral identities
(where
(((.,zj))) = ((.,z;.)) + l(.,z:.)
here), with Zo(x) = 0, and we can construct the Rothe function u„{x, t)
defined for every ( e 7 by
j= 1,...,2''-V
In this way, we come to a sequence
{u„{x,t)} (11.50)
of functions (11.49); let us call it the Rothe sequence corresponding to the
problem (11.l) —(11.4). It can be expected, intuitively, that this sequence
will converge (in a certain sense) to a function u(x, t) which will appear to be
a solution (also in a certain sense) of the problem (11.1) —(11.4). The follow-
ing investigations are devoted to the proof of the just expressed assertions.
The formulation of the thus obtained results is given in Theorem 11.1,
p. 217.
Let us investigate the first of the problems mentioned above i.e. the
question of convergence of the Rothe sequence (11.50). To this aim, let
us start by deriving some apriori estimates.
For this purpose, let us write the problems (11.34), (11.35) in the follow-
ing form, more convenient for our investigations: To find, successively
for j = 1, ..., p, such functions
ZjeV (11.51)
for which the integral identities
V = Zi . 11.54)
This is possible because Zj e V. We get
((z„z,)) + 7 ( z „ z O = ( z „ / ) . 11.55)
n
Now,
((zi,rO)^0 11.56)
by (11.13),
11.57)
and
(^i>/)glki||||/|l 11.58)
according to the Schwarz inequality (2.12), p. 10. So (11.55) yields
i|klNlklMI/IU 11.59)
wherefrom
11.60)
Subtracting from the identity (11.52) the same identity written for
we obtain (because of the hnearity of the form ((., .)), etc.)
the inequality
||z,-z,_,|l^^||z,-z,_,||||z,_, -z,_,l|, 11.64)
196 Chapter 11
yielding
\\z, - --_i|j ^ \\zj_, - z,_,!| , ;• = 2,..., p . (11.65)
Now, ro(x) = 0, so that, according to (11.60),
\\z, - z o l l = \z,\ ^h\\f\\. (11.66)
From (11.65) it then follows
||z,-z,_,llgMl/|
for all } = 2 , . . . , p, and, in consequence of (11.66), for all y = 1,..., p.
Thus, we have
| | . - - z , . _ i | | ^h\\f\, j = I,..., p. (11.67)
(11.67) then gives
||z,.|! ^ ||z,._i|| +h\\fl, j = l,...,p, (11.68)
wherefrom
|z,.|j ^ ; / j | j / | , ; = i,...,p, (11.69)
since ||zo|| = 0. Now,
P
implying finally
hj\\ ^ n n • (11-70)
Denote
Z.{^) = 'M^:ll^M, ; = i,...,p. (11.71)
h
(Thus, Zj{x) is the difl'erence quotient (11.33) "coresponding to dujdf'^)
The inequalities (11.67) yield, immediately,
What has been derived for the case of the division di can be easily derived,
in a quite similar way, for the case of an arbitrary division d„: From the
The Equation cujct + AU = f 197
integral identities
((., z;)) + 1 {v, 2," - z"j_ 0 = {v, f) V. e F (11.73)
K
which are analogues of the identities (11.52), we get for 7 = 1 and u = zj
(taking into account that zl{x) s 0 and ((z", z")) ^ 0)
f ik"r^ik-"iiii/ih (11-74)
K
giving
ll^'?ll ^ /'.||/i • (11-75)
Subtracting from (11.73) the same integral identity written for 7 — 1 and
putting V = z"j — z'j_y we get, similarly as in (11.65),
jl-j - =j-i|| ^ i|2j-i - ZJ-2II . 7 = 2, ..., « . (11.76)
This implies, simultaneously with (11.75),
V-A-A^K\I\, (11.77)
and finally
VA ^ Afl. (11-78)
li^;!! ^ i / i = (11-79)
7" = 1, ..., 2"~^p, where
The same conclusion follows easily even for the norms of the functions
z" in the space W^''\G), i.e. there exists a constant c- > 0 such that
fAz'-HG) ^C (11.81)
holds for all 7 and n. In fact, if we write the integral identity (11.73) in the
form
{{v,z'^)) + {v,Z"j) = {v,f) yveV (11.82)
198 Chapter 11
') In what follows, H= V OT H = £2(0) will be considered most often. Let us men-
tion that the definition of an abstract function introduced here may be generalized
substantially — see e.g. [6].
The Equation ciijct + Au = f 199
etc. Thus, an abstract function from / into V is given here, piecewise linear
obviously, as seen from (11.89). Its "symbolical graph" is sketched in Fig. 11.
r—iTti-^ Rg-ii-
The "symboHcal graph" of this function is sketched in Fig. 12. Even here
we have a mapping from / into V.
200 Chapter 11
Let us note that the functions :"ix), being from V, belong to L2{G) as
well. Thus, the presented functions are both also abstract functions from /
into 1.2(G).
An abstract function (11.88) is called continuous at an inner point
teJiito every s > 0 a ^ > 0 can be found such that
z;
Fig. 12.
o\ 1; t2 (; /
where is the norm in the space H. Similarly, the continuity from the
right, or from the left, is defined. The function (11.88) is called continuous
in the interval (0, T) if it is continuous at every point t e (0, T); it is called
continuous in the interval I = [0, T] if it is continuous in (O, T) and
continuous from the right at the point t = 0, from the left at t = T.*)
In a similar manner the limit (limit from the right or from the left)
is defined.
at an inner point t e I, if
(11.96) and (11.95) imply that the convergence in the space C(I,H) is
uniform convergence in H on the interval J, i.e. that to every s > 0 an
«o(£) (thus not depending on tel) can be found such that
" > "0 =* hit) - v„(i)|U < s Vr 6 / . (11.97)
For abstract functions the so-called Bochner integral can be defined,
which is an analogue of the Lebesgue integral of a real function and has
similar properties: First, a simple function is defined as such an abstract
function (11.88) which attains, on /, only a finite number of "values"
Qi,..., g„,e H, on (Lebesgue) measurable sets Ni,...,N„ with measures
Hi,...,fi„„ respectively. The Bochner integral of such a simple function
is defined by
[ y{t)dt = Y^ilh-
1= 1
I,,
u„{t)dt= X ''«2" = ''. I
i=l i=l
^".
Jf, moreover,
For the details and for the generalization of these concepts see [6].
An easy computation (which we are not going to carry out here) leads
to the conclusion that each of the Rothe functions u„(?) given by (11.89),
being piecewise hnear in t, is strongly measurable and Bochner integrable.
Moreover, we have
j;.('H'-'rjk- 1 +
^J-1
( ' - ' ; - i ) dt =
z",„-.,_0 + --- +
;=i 2
(11.105)
(since zS = 0). Thus (11.105) and (11.101) yield
The space £2(1, H) is the space of functions which are square integrable
in the Bochner sense, i.e. Bochner integrable and satisfying
1, ||y(0||^d(< + 0 0 ,
(3'i(0. 3^2(0)^
is the scalar product in H. The space H being a Hilbert space, it can be shown
that L2{l, H) is a Hilbert space as well.
A primitive function
Choose now, for H, the space Fdefined by (11.ll). In the space Ljf/, V),
the scalar product is given by (11.108), i.e. by
respectively.
In accordance with (11.115) convergence is defined:
;„->• in L2(/,F) (11.116)
means that
lim { \\y - y„l^dt = 0. (11.117)
w„(0 = z;-x + i ^ ^ ^ ^ ( r - ? ; - i ) in /; = K - i , r ; ] .
(11.119)
According to (11.87) we have
ilz}|y ^ c Vn = l , 2 , . . . and Vj = 0, 1, ..., 2"-^p . (11.120)
Now, „
0 ^ L~ ^J-^ ^ 1 in / ; , (11.121)
The Equation c'u;ct + An = f 205
t — t"
{^-•-^y--'- V
+ <
t - tj
<ii-^^^-^ C + - i
c = c . (11.122)
Thus, the Rothe sequence (11.118) is bounded in the space £2(1, V).
Since this space is a Hilbert space a subsequence
{«J0} (11-124)
can be found weakly convergent in this space to an abstract function
ueL2{l,V). (11.125)
We write
w„^-« in L,{I,V).'^) (11.126)
'^) We apply a well-known theorem here: If -Y is a Hilbert space and {x„] a bounded
sequence in this space (i.e.
UnWx^ C forall«),
then a subsequence {x„^ can be found which converges weakly in X to an element
xe X,
x„^-^x in X. (11.127)
(11.127) means that for every bounded linear functional g defined on X we have
g(x„) -> g(x).
According to the Riesz theorem, every bounded linear functional defined on the
Hilbert space X can be expressed as a scalar product with an appropriate (uniquely
determined) element y e X. Thus the weak convergence x„ -^ j ; in X can be charac-
206 Chapter 11
In what follows, we will show that precisely this function u{t) is the re-
quired solution (in the sense of Theorem 11.1, p. 217).
Using this notation, the functions (11.119) can be rewritten in the form
u„{t) = z%,+Z%t-fj_,) in / ; . (11.130)
Define the abstract functions
U„{t):I->L,{G), n = l,2,..., (11.131)
by
Z" for 1 = 0,
Un{t)( (11.132)
^ z ; for tei'; = {f;._„t'}], ; = i,...,2"-v.
From (11.79) it follows immediately (cf. (11.123)) that the sequence
{Unit)] (11.133)
is bounded in the space Lzih ^2(0)). This space being a Hilbert space,
a subsequence
{Udt)V') (11-134)
can be found converging weakly in this space to a function U e L2{l, L2.{G)),
U„^-U in L^ihL^iG)). (11.135)
terized by
(xn^,y)x^(^-yh Vj'G^. (11.128)
The weak limit is unique, i.e. a sequence cannot have two different weak limits.
Note that if a sequence has a limit in A" in the ordinary sense (called the strong limit,
if we wish to distinguish it from the weak one), this limit is also its weak limit. The
converse is not true.
^^) A subsequence with the same indices n^ as in (11.124) can be considered without
loss of generality.
The Equation cujct — An ^ f 207
exists (cf. (11.109), (11.110), where H = L^iG)). Now, from (11.132), ( U . l 30)
it follows immediately that in the space L2(/, ^2(G)) we have
J:
/o
t^J^)dT = t<JO, (11.137)
'*) In detail: The weak limit being unique it is sufficient to show that u„^ -^ w
in L2U, Z-^CG)), i.e. to show that
First, let, v(t) = ve LjiG) for all / e / . We have (see (11.110), p. 203)
Finally, from
we have
u(0) = 0 in C{I,L2{G)).'') (11.142)
Thus the function u(t) possesses several characteristic properties. First,
since u e L2{l, V) we have, for almost a\\ tel, ue V. In this sense the stable
boundary conditions of the given problem (11.1) —(11.4) are satisfied.
The initial condition (11.2) is fulfilled in the sense of (11.142). What is of
particular interest now is the question in which sense is the given differential
equation ( l l . l ) satisfied. To answer this question is the aim of the following
text.
'^) Note that the equality (11.138) was completed by "in L2{I, L2(G))". When doing
this we express that (similarly as in the Lebesgue space Lj of "ordinary" functions)
equality in the sense of "almost everywhere in the interval / " is understood. Thus, in
(11.139), we should have said that the function u{t) could be made absolutely continu-
ous in / when changed properly on a set of measure zero, if necessary. The "uiO) == 0
in C(/, Z,2(G))" means that the function u(t), as a contiriucus abstract function from /
into L2(G), assumes, for / = 0, the "value" zero in L2(G) (i.e. that «(0) is a function
equivalent to the function z s 0 in G).
The Equation Bu/dt + Au = f 209
However,
u - «„, = (u- uj + (M„, - aJ (11.148)
and w„^ -^ u in ^2(7, K), so that
Fig. 13.
of the Lebesgue integral, it can be shown that the set N is dense in L2([, V).
Since each of the functions from iV is a linear combination of functions
from M, it is sufficient to prove (11.150) only for the functions from M,
i.e. to prove that
« . - . - . - ) ( - ^ - ) =
f — t"''
= (z--z;i,)^-^ (11.154)
t;,(z^^-z-0-
J'7-1 \ "/Ik
~(f _ /"kV-I^nic
(V, Zj - Zj_i)y — - [V, ^j_y Zj )y
_ -"nt J'/-l
The Equation ciiict + Au = f 211
We thus have
Now,
/i„, -» 0 for n, -> CO ,
|(..z--z^^),|.^i|.||,!|z^--4-||,g2c||.||,
according to (11.87). Because i; is fixed, we get
However, the function v(t) from M has been chosen arbitrarily, where-
from (11.156) is proved for all ve M, and, in consequence of the density
of the functions from N (see p. 210), for all v e L2{l, V). Thus,
u„^--u in L,(/, F ) , (11.157)
which was to be proved.
Now we are prepared to show in which sense the function u[t) satisfies
the given equation (H.l):
Consider the integral identities (11.73) written for n^:
we have
{{v{t), u„M + « 0 . U„,it)) = Ht),m for almost all
telO,TY^). (11.161)
Integrate (11.161) between 0 and T. We obtain
L iHt),sit)))dt
i;
>r
I
Thus, (11.165) is actually a bounded linear functional in L2{l, V). Conse-
(11.166)
quently, from
u„^ — u in £2(7, V)
it follows (cf. Footnote 12, p. 205) that
1: {v{t),f)dt,
{v{t),u'(t)dt= f (KO>/)df
0 Jo
Vi'eL2(/, F) (11.170)
214 Chapter 11
In this sense, the function u{t) satisfies the differential equation (U-l).
u eAC{l,L^{G)), (U.172)
u'eL^iUL^iG)), (11.173)
u{t) for 0 g I^ a
v{t)( (11.178)
0 for a < t ^ T.
(«(/) belonging to L^Q, V), the same holds for v{t) obviously.) Now, u e
The Equation cujct 4- Aii = f 215
Before summarizing all the obtained results into a theorem, let us present
two remarks.
REMARK 11.1. Uniqueness of the weak solution u{t) implies that not only
the sequence {M„^(f)} converges weakly to u{t) in L2(/, F), but that the whole
sequence {«„(')} does. In fact, let {u„(f)} n o t converge weakly to u{t) in
h^, F ) . Then there exists at least one function v e L2(/, F) and an e > 0
such that
If
I Jo
{v{i), u{i) - M„(/))K d/ ^ £ (11.183)
(This result can be derived also in another way. See Remark 12.7, p. 243.)
Thus, we have proved the following theorem:
1^) We have
Thus, for every v e £2(1, L2(G)) fixed, (v, U)L2(I,L2(G)) is a bounded linear functional
on L2(I, V). Consequently, «„ ^ « in L2{I, V) implies
-^ + Au =f in 0 = G X (0, r ) , (11.187)
dt
u{x,0) = 0, (11.188)
BiU=0 on rx(0, r), i = ],...,/(, (11.189)
CiU = 0 on r x ( 0 , T), (• = 1, ...,/c - / i , (11.190)
iie given. Here,
feL^iG), (11.191)
^^= Z (-1)I'ID'K(A)D^), (11.192)
THEOREM 11.3. Let u(t), or u(t), be the weak solution of the problem
(11.l) —(11.4) with the right-hand side f, or f, respectively. Then
ll«(0-"(OI|c(/,L.(G))^r||/-/i|. (11.197)
Thus, if the difference of the right-hand sides of equation ( l l . l ) is small
in LJIG), the diiference of the corresponding weak solutions is small in
C{l, L2{G)), i.e. small in L2{C) uniformly for all teL
Chapter 12
Under the assumptions imposed on the form ((r, u))^) and on the function
/ in the foregoing chapter, not much more can be said about the weak
solution of the problem (11.1) —(11.4) than has been done in Theorem
11.1. In the present chapter, we show that under additional assumptions
stronger results can be obtained. Moreover, relatively simple error estimates
can be derived.
a) SMOOTHER SOLUTIONS
') Especially on the operator A (to have, for example, only bounded measurable
coefficients).
^) This assumption is satisfied if, for example, Afe L^iG) and ((f,/)) = (v, Af) Vf s V.
In this case M = \\Af\\. See Footnote 11, p. 82.
219
220 Chapter 12
^) When doing this, we extend the integral identities (12.3) —(12.7) by the "zero
identity"
(fr, 4 ) ) + (V, Zg) = (vj) (12.20)
(taking into account that Zg = 0), thus defining z'Lj = ~l'nf- ^r, which is the same,
(12.19) is the consequence of (12.20).
'^) Thus, s'j corresponds to d^ujdt'^ at the point t = t" = jh„.
222 Chapter 12
I|Z;||K^-. (12.24)
a
Thus, under the additional assumptions (12.1), (12.2), \\Z"\\y and \\sj\\ are
uniformly bounded with respect to both] and n.
This fact has interesting consequences: In the preceding chapter, we have
seen that sequence {U„{t)] of functions defined in the interval / by
Z" for f = 0
UXt) = ( (12.25)
^Z]{x) in / ; = (f;_,.f}]
(see (11.132), p. 206) converges weakly in L2(/, LaC^)) to a function U{t)
which is the derivative of the function u{t) in that space.*) Now, on the basis
^) More precisely, only a subsequsnce {Unjt)^ of the sequence {t/„(/)} was investigated
there. However, from the considerations of Remark 11.1, p. 215, it follows that this
assertion holds for the whole sequence {UJit)}.
Let us remind here that, speaking about the function u{t), we always mean the
absolutely continuous function, given by the integral (11.136), p. 207, thus the original
function (11.125) properly changed on a set of measure zero, if necerrary (cf. Footnote
15 on p. 208). Also in what follows, speaking about smoothness of the function «, we will
always mean smoothness of this function "properly changed on a set of measure zero",
without reminding this fact any more.
Regularity. Error Estimates 223
M e AC{I, V) . (12.27)
si for f= 0
s„{t) = ( (12.29)
^5; for tel)^{t''j.„f;]
converges weakly in £2(7,1'2(<5)) to the derivative of that function,
Consequently,
«' e AC{I, L^iG)) . (12.31)
II^^WIIK^- (12.32)
a
(see (12.24) and (11.122), p. 205). Thus, for every such t the sequence
{0„{t)] is compact in £2(0).') Further, the functions 0„{t) are uniformly
*) The sequences {0„(t)j and {U„(t)} are analogues of the sequences {«„(/) and {u„(,t)}.
^) Even in W^''~ ^ *(G), however, this fact is not of interest at this moment.
224 Chapter 12
g M (12.33)
and
V-elliptic . (12.36)
Moreover, let
feV (12.37)
and let a constant M (independent of v) exists such that
\{{v,f))\SM\\vl V.eF.^) (12.38)
Then
u e AC{1, V) , (12.39)
u' e L^{I, V) n AC{I, L^iG)), (12.40)
u" e L2{I, £2(0)) . (12.41)
Moreover, (12.34) holds.
(practically, A^fs 12(C) and {{v, Af)) = (v, A^f) holds for all veV) one
derives without difficulties that
I|s;ii = g M , (12.43)
K
yielding in the same way as above
Subtract the integral identity (12.5), p. 220, written for m and / instead
of n and j , from (12.6). One obtains
{{v, zT - 2r_0) + {^>, zr - zr_0 = 0 vr e K. (12.51)
Noting that
m _ m
•7m _ ±i ±i-l
K~
and taking into account (12.3) with z'^ — z" — z" because of z" = 0,
we get, successively,
/i„,((^,ZT)) + ( j ; , z r - / ) = 0 , (12.52)
h^{{v, Zfi) + {v, Z'"^-Z1) = Q, (12.53)
/.„i:((zr,zr))^iil/ir- (12.56)
i=l
«|Z?*.PS^|I/r. (12.64)
i = l,2, ....
according to (12.62).
At the same time
{^2j^ Si) ^ {slj-u sl_0 ^ • - ^ (^J+1. ^7+1) •'') (12-73)
Now, (12.73) applied in (12.72) gives
V-~Sj+2- .
"m
yields
(.7+2..7+2 - ^7+1) ^ o=> ||.7+2li s II.7+1II.
etc.
230 Chapter 12
Consequently for the functions s„(t) (see (12.29)) we have the estimate
lte.,ll-s;^ltel!s,^ll/!l. (.^.80)
I = 1, 2 , . . . , and
\K.4^~^JsU^~y\\fh (12-81)
j = 1, 2 , . . . (where, in accordance with our current notation,
1° <
2{q + 1)
Continuing the process leading to (12.79) and (12.82), after q steps we come
to the result that
i7'«> e AC{[ri, T], V). (12.83)
Hence we can present the following theorem:
Till now, regularity of the weak solution of the problem (11.1) —(11.4)
"with respect to the variable t" was in the center of our considerations. Let
us turn to questions on regularity "with respect to x".
Here, we shall follow, to a certain extent, the idea of the paper [16] by
J. Kacur and A. Wawruch, and make use of Theorem 4.1.3 of the mono-
graph by J. Necas [2], p. 200 (with x = 0 and I = k + I in the notation
of that theorem). We present this theorem in a special form here:^'*)
'*) Its general formulation would require the introduction of some new concepts.
Moreover, our aim is to present the following results in a form as simple as possible.
Note that different requirements may be imposed on the coefficients a;^(x) of the form
A{v, u) according to how many times should these coefficients be differentiated if we
write formally
^= X {-iy'^D'{a,j{x)DJ); (12.85)
\i\.\J\Sk
further, note that more general assumptions concerning g(x) may be considered, etc.
232 Chapter 12
converges weakly in L2{l, I'I'2^*^(G')) to the function u[t) which is the func-
tion u{t) restricted to x e G'.'^)
So we can present
^^) It actually converges to this function. In fact, the sequence {«„(?)} converges
weakly in L2(I, L2(G)) to u(t) (see Remark 11.2, p. 216). This insures the weak con-
vergence of {«„(')} precisely to u(t).
1^) Note that (12.106) insures A/e LiiG). If, moreover, ((u,/)) = (v, Af'^v E V (and
this case is very frequent), then (12.2) is fulfilled.
^') We denote the constant (depending on G') in (12.109) by the same symbol C as
in (12.90), although these two constants may be different.
Regularity. Error Estimates 235
REMARK 12.4. A similar results can be obtained when applying the above
procedure which leads to results on regularity "with respect to t" on an
interval [;?, T], thus without assumptions (12.1), (12.2), assuming F-sym-
metry of the form {(v, u)) only:
236 Chapter 12
Let ((v, w)) be F-symmetric. Then the inequahties (12.112) can be replaced
by the inequalities (12.65), (12.75) (p. 227 and 229) - let them be written
here in the form
vahd for t 2^ 2f°.'^°) Thus, if (12.106) holds, precisely the same procedure
as in Remark 12.3 can be applied, leading to the result
) Assuming that n is sufficiently large so that / " be a point of the division (/„ (and
2t° e{0, T)).
^') Hereby derivatives with respect to the variables Xj are understood.
Regularity. Error Estimates 237
At the same time, the norms of the functions z" are uniformly bounded
in this space. Thus the sequence {«„(?)} of the restrictions of the functions
u„(f) on G' (f is always kept fixed!) is bounded in W^^''^{G'). (This is easy
to calculate, see (11.122), p. 205.) Thus a subsequence {u„j{t)} can be found,
weakly convergent in this space to a certain function w(t) e W^^''\G'). Now,
M„ -> u in C(/, L 2 ( G ) ) , so that u„(t) -> M(() in L2(G). Consequently «„(r) -^
->• ti{t) in L2{G), and thus as well in L2(G'). So w(t) =^ u{t). Thus u(t) belongs
to W^^^^G').
If we remind that the number q has been chosen arbitrarily and apply
the well-known immersion theorem, we can complete Theorem 12.2 by the
following one:
b) ERROR ESTIMATES
Let us turn to the question of error estimates, i.e. to the problem of estimat-
ing the difference between the weak solution u(t) of problem (11.1)-(11.4),
p. 187, and its approximation obtained by the method of discretization
in time, taking the step h.
Under the assumptions of Theorem 11.1, p. 217, only, there is no possi-
bihty of obtaining an efficient error estimate. (Cf. also Remark 12.7, p. 243.)
In what follows, we will assume that besides these assumptions the assump-
tions (12.1), (12.2), p. 219, are fulfilled. We will show that in this case quite
satisfactory error estimates can be derived.
First, let us recall, that (i2.l), (12.2) imply
||s;|| g M (12.124)
^^) >i s (0, T) being arbitrary, this result holds for all / 6 (0, T].
Let us note here that there follows from Theorems 12.2, 12.4 and from Theorem
3.7.8 and Example 3.7.1 from [2], pp. 192 and 193, that the function u, as the function
of the variables x^, ..., Xj^, t, has in G' X [/;, T] continuous partial derivatives of all
orders (in the classical sense).
238 Chapter 12
(see (12.22). p. 221). Let us investigate the original division di of the interval
/ = [0, r ] into p subintervals of lengths h. At the point tj = jh we have
Denote
zlj-zj =qj, j = Q,l,...,p. (12.129)
(In particular,
g^ = z^ - z^ = 0 (12.130)
in consequence of the given zero initial condition.) The function ^X-"^)
satisfies the integral identity (see (12.127))
we get
I k l i ^ ^ - (12.132)
and, in general,
Analogously, denoting
z 2j 2j = rj, j = o,i,...,2p
and. in general,
Take a fixed k and consider the just discussed differences at the point
t^ - kh. Taking into account that for the division d„ we have to take
j = 2"~^k to get the point we obtain
^ ik + — + . . . +
4 V 2^ (2"~')'>
1 + - + ...+ ;]^ .
4 V 2 2"-7 2
(12.136)
240 Chapter 12
Now, the sequence {«„(<)} converges to u{t) in C{l, L2{G)), thus in L2(G),
at every point t^. Then (12.136) gives, for n -> oo ,
li«(.,)-".MI^^- (12.137)
This is the required estimate.
M = \\Af\l . (12.138)
(l+/iC^)li,J!|^l|,]_,||+^, (12.141)
j = 1,..., p, with
HI = 0 (12.142)
according to (12.130).
Let us majorize the ||gj|], j = 1, ..., p. First replace inequality in (12.141)
by equality,
dt 4
y{0) = 0 , (12.146)
nivi
y' = C'y,
4
Ik; I hM
- clkJ
h 4
Subtracting, we get
y _ M l ^ C \ \ q \ \ \ - y ) . (12.147)
h
Should ll^ill > 7j be true, then the right-hand side of (12.147) would be positive
on the interval [0, h]. Consequently, the left-hand side of (12.147) would be positive
on that interval as well, i.e. we would have
Thus
|U„^||g^^(l-e---). (12.149)
Similarly, we get
etc. Following the same idea as before, we come finally to the error estimate
ll"(0-"i(OII^^.(l-e-^^^*) (12.150)
which is still slightly better than the estimate (12.150). P'rom the practical
point of view, these two estimates are equivalent.
^'^) The estimate (12.137) is obtained from (12.150) by expanding the exponential
function into a series and taking its first two terms only.
Regularity. Error Estimates 243
//, moreover, the constant of positive definiteness C,for which we thus have
({v,v)) 2: C^||t;|f Vre F , (12.155)
is known, then (12.154) can be replaced by
|KO-«.(/,)||^^(l-e-^^*'), (12.156)
2C^
or by
< Ml
= 2C^
— ^1
(1 + hCjj
(12.157)
^^) As pointed out above (12.153) is fulfilled if ^ / e ijCG) and ((v,f))=(v, Af) for
all f 6 V. Then we take M = p / | | .
) I.e. the estimate of the difference between the solution u{t) and the approximation
Uj(/) obtained by the method of discretization in time at the points of division t^^ = kh,
where h is the step of the discretization.
244 Chapter 12
First, recall the definitions of the functions u„(t), u„[t) and U„(t):
«„(0 = z;_i +{t- t"j^,)z"j in / ; = K _ i , t-;], (12.158)
z1 for t = 0,
u„{t) = ( (12.159)
^z:; in /; = (r^ „/:;],
Z"i for f = 0,
C/„(0 = ( (12.160)
^ Z ; in I-}.
From j|z;|! ^ ||/|| (see (11.79)), it follows that
so that
il«„(0 - "«(OII ^ l|/i^« V/E/. (12.162)
Let u(/) be an arbitrary function from C(/, V). (We choose the space
C{l, V) here to avoid the lengthy writing "for almost all t e / " in what
follows.) Rewrite the integral identity (11.161), written for n instead of for
«t, in the form
{{v{t), i\{t))) =: {v{t),f{t)) - {v{t), UXt)) Vtel, (12.163)
where f[t) = / f o r all tel.Jn consequence of (12.161), we have
Now, M,„(/) -» u{t) in LjiG) for every t e I and h,„ -> 0 for m ^ oo, thus
\\u„{t) - u{t)\\ ^ 2||/|| v'(2T/7„) Vf e / . (12.172)
In particular, for « = 1, (12.172) yields the desired estimate
\\u{i) - u,{t)\\ ^ 2||/|| v'(2T/0 ytel. (12.173)
As error estimates (12.172) and (12.173) are quite unefficient from the
practical point of view, involving h to the power 1/2 only. Nevertheless,
the method presented is of particular theoretical interest. Among other
things, from (12.171) it follows directly (as proved in Chap. 11 in another,
undirected, way) that {i(„(/)} is a Cauchy sequence in C(l, L2{G)).
Chapter 13
~ + Au = f in 0 = G X (0, 7 ) , (13.1)
dt
M(X, 0) = »,(x), (13.2)
/?,.H = gi on r X (0, T ) , i = 1, . . . , / / , (13.3)
C,u = 0 on r x ( 0 , T ) , i = 1,...,/c - / < . ' ) (13.4)
As in Chap. 11, we assume that G is a bounded domain in £,v with a Lip-
schitz boundary F, the form [(v, u)) corresponding to the operator A and
to the given boundary conditions is bounded in W2''\G) and K-elliptic, i.e.
|((i;, u))\ ^ K|H|^,,.,(e) IHk.c-XG) Vr, u e nf\G) (13.5)
and
{{v, v))'^, (x\\v\\^ V u e F , (13.6)
where the notation
247
248 Chapter 13
3II
— + .4M = 0 in 0 = G X (0, T ) , (13.21)
u{x, 0) = 0 , (13.22)
B-u = Qi on r X (0, T), / = 1, . . . , / ( , (13.23)
C,M = 0 on r X (0, r ) , i = \,...,k- n. (13.24)
The problem (13.13) —(13.16) was treated in the two preceding chapters.
In this chapter, we shall be concerned with the problems (13.17) —(13.20)
and (13.21)-(13.24).
Nonhomogeneous Conditions 249
Thus, let us start with the first one, i.e. with the problem (13.17) —(13.20).
Applying the method of discretization in time, with step h — Tjp, one has
to find, successively for j = 1, ..., p, the functions
zj e V (13.25)
satisfying the integral identities
In the same way, we can construct, for the division d„, the Rothe function
u„(f) = M„(x, 0 defined in Fj = K - i , f,"], ; = 1..., 2"-'p, by
) Roughly speaking, the norm of the function Sujdt, obtained formally by (13.17)
for / = 0, may become arbitrarily large if the function UQ(X) is sufficiently "wiid".
Nonhomogeneous Conditions 251
paper [7] by the author and M. Ludvikova. In this chapter, we shall fol-
low, to a certain extent, the idea of that work while the results obtained
there will be strengthened substantially here.
First, let us return to the integral identities (13.26) written for the
division c/„:
-4 ^ "0 (13.41)
(uniformly for; and n). From (13.32) we then get by an easy computation
(cf. p. 205)
\\u„{t)\\ ^ Iwoli Vfe/. (13.42)
This inequality represents the continuous dependence, in C(/, L 2 ( G ) ) ,
of the norms of the Rothe functions M„(f) on ||u(,||. This fact will be utihzed
in what follows, simultaneously with an important assertion from [2]: Since
the form ((f, u)) is V-elliptic, a set M exists ([2], pp. 131, 132), dense
in V and consequently in L2(G), with the following property: If se M,
then precisely one g e L2(G) exists such that
is fulfilled?)
First, let
Mo = s e M . (13.44)
Similarly, putting
z; = z; + 5, ,/ = 2 , . . . , p , (13.48)
DEFINITION 13.1. The function u{t) is called the weak solution of the
problem (13.17)-(13.20) with
I/O = s e M . (13.64)
254 Chapter 13
DEFINITION 13.2. The function u{t) given by (13.71) is called the very
weak solution of the problem (13.17) —(13.20).
Nonhomogeneous Conditions 255
To establish this definition, we have to show that the function u{t) does
not depend on the choice of the sequence {s,} satisfying (13.66). However,
this is quite easy: Let S; e M, Sj e M, i = 1, 2 , . . . , be two such sequences,
thus satisfying
Hni Sj = MQ > ^™ ^i ~* "o in ^2(p) • (13.72)
•*) Recall that the functions z'jix) in (13.32) are the weak solutions of the problems
/? = - > 0 . (13.79)
3
*
To this r], a function s,- e M exists such that
hi -Uol<ri. (13.80)
Moreover, to the same tj, an HQ can be found, such that
n> no=> \\u'{t) - ul,{t)\\ <r] \/teI. (13.81)
(Having S; e M, {u^(i)} converges to u'{t) in C{l, ^2(0)).) Then
5
) The second of the conditions (13.93) is then fulfilled if e.g.
A^UoeL^iG) and ((v, Au^)) = (v, A^UQ) V y e F , (13.94)
vvithM= M^«oll-
Nonhomogeneous Conditions 259
In the same way, following (12.69)-(12.73), pp. 228, 229 (thus writing
(cf. (12.66), p. 228 and (12.76), p. 229)), etc. Thus, quite similar considerations
as above lead to the conclusion that
(i) for the functions u„(f) and U„{t) we have
"""^^
'"" - V(^"""''' il^"^'^l''-il'"°'l ^^^-"^^
for all t e [2/°, T], and that, consequently,
(ii) the restrictions of the functions u„{t) and [/„(/) on the interval [2r°, T]
converge weakly, in L2([2f°, T], V), and in L2([2r°, T], L2(G)), respectively,
to the restrictions u{t) and u'{t) of the functions u{t), or ii'{t) on that interval.
THEOREM 13.4 Let the form {{v, u)) be bounded in Wf\G), V-elliptic
and V-symmetric, MQ e -^2(0). Let rj e (0, T) be arbitrary. Then for every
te\r\,T'\ the very weak solution u{t) of the problem (13.17) —(13.20)
belongs to V. As an abstract function from \r], T] into V, u(t) has derivatives
«*'''(?) of all orders, and we have
Inserting v(t) = u^'^(f) and performing the integration in the second term,
we get
i: y>Udt=^
.,(0112 A, -. i"<"'"'
ll„(0„
h'nUv^^-^hV- (13-127)
2a
Then (13.66) implies the uniform boundedness of |M''^|j;,2m.)With respect
to ('. Thus, a subsequence {«'*(0} can be found, weakly convergent in
L2(l, V) precisely to the function u{t) (because (see above)
u'*->M in C{l,L2{Gy),
so that the more
«'*-« in L2{I,L2{G)),
and (11.185), p. 216, can be applied). Thus, the very weak solution satisfies
ueL2{l,V). (13.128)
Consequently, for every fixed v e L2{1, V), the limit process for ;; -* 0 in the
first term of (13.125) can be carried out. Since the equality (13.125) is valid
for every /; > 0, the same limit process in the second term can be performed
yielding finally (13.126).
REMARK 13.3. Regularity of the very weak solution "with respect to x "
can be treated similarly as in the preceding chapter:
For the coefficients of the operator A, let the condition
Moreover, let
Mo e V, {{v, uo)) = {v, Au^) Vi; e V. (13.132)
Nonhomogeneous Conditions 263
Thus, the conditions (13.86), p. 257, are satisfied and the integral identities
(13.87) can be converted, by the substitution (13.88), into the integral
identities (13.89) corresponding to the problem (13.13) —(13.16). In this
way, the uniform boundedness of ||Z"|| is ensured. Thus the procedure
similar to (12.96)-(12.102), p. 233, can be apphed leading to the assertion
of Theorem 12.3, p. 234:
ueL^iLW^^'XC')'). (13.133)
Here (see the quoted theorem) G' is a subdomain of G, G' cz G, and u(t)
is the function u(t) = u(x, t) restricted to x e G'.
What is of use here is to assume the F-symmetry of the form {(v, i/)),
UQ e L2(G) being sufficient, and to follow the procedure (13.101) —(13.110),
pp. 258-259, leading to the inequalities (13.114), (13.115), (13.120), (13.121),
etc., as well as the procedure of Remark 12.4, pp. 236 — 237, leading to
Theorem 12.4, p. 237. In this way we obtain:
THEOREM 13.6. Let the form ((«, «)) be bounded in ^Vf X<^) and V-elliptic,
with the constant of positive definiteness C (see (12.139), p. 240). Let
UoeV, AuoeV (13.134)
and
{{v, Mo)) = (v, Auo) , \{(v, Auo))\ S M\\v\\ ^v e V ^). (13.135)
Then
|KO-u,(0|lg^(l-e---). (13.137)
f - S = 0 in (0,.)x(0,l) (13.141)
ct ox
M(X, 0) = sin 30x , (13.142)
M(0) = 0 , u{n) = 0 . (13.143)
The function MO(X) = sin 30A- fulfils all the assumptions (13.134), (13.135),
with
A^Uo = 30*sin30x. (13.144)
8100 / ^
\\u{t,) - u,{t,)\\ ^ - J - ^ (1 - e - ° - ° " ) , (13.146)
for this case, would be valid no more. In fact, the explicit solution of the
problem (13.141) —(13.143) as well as its discrete approximations are known:
u = sin30xe"'°°', (13.148)
REMARK 13.6. [Rate of the decrease of the very weak solution of the
problem (13.17) —(13.20) with t.) Let the assumptions concerning the
boundedness and F-ellipticity of the form ({v, u)) be preserved and let the
constant of positive definiteness of this form be known, i.e. such a number
C > 0 for which
{{v, v)) ^ C^\\v\\^ WveV (13.152)
is vahd.^'). Then simple formulae can be derived showing the rate of decrease
of \\u{t)\\ with t:
^^) Here C need not be the largest possible one. For many examples see Chap. 4.
Nonhomogeneous Conditions 267
yielding
=1 ^
(13.155)
1 + C^h„
Similarly, z^ inserted for v into the second identity yields
and, in general,
Now, let
)t/i„ = f 6 [0, T] (13.158)
be fixed. We have
2j,\2"-».t/fc
(-S)
At the same point
t = 2fc/i„+i = 2 / c . ^ = kh„
2
268 Chapter 13
we get
<
'-2. n^z ::77rv7T7.-> (13-160)
C^hV"-""
(-?)
etc. Approaching the limit for n -• co and taking into account the uniform
convergence, in £2(6), of the sequence {u„(t)] to the very weak solution
u{t) of our problem in the interval / = [0, T], we obtain, finally,
^ - ^ = 0 in (0,K)x(0,l), (13.162)
dt dx
u(x, 0) = sin X , (13.163)
w(0) = 0 , U{K) = Q, (13.164)
~ + Au =0 in 0 = G X (0, T) (13.166)
ct
Nonhomogeneous Conditions 269
u(x,0) = 0 , (13.167)
B;u = y,- on f x (O, T ) , i = !,...,/(, (13.168)
CjW = 0 on r X (0, T ) . i = l,...,k ~ fi. (13.169)
Here, the existence of a function
weW^'XG) (13.170)
is assumed which satisfies
(13.173)
are fulfilled, with
zl = 0 (13.174)
(according to (13.167). The space Fis defined by
*^) The case we V can be excluded, of course, being of no interest. (The solution
of the considered problem would be a zero function.)
270 Chapter 13
the form {(v, u)) was assumed to be symmetric (in order that
u-weV, (13.177)
Since the form ((v, u)) is bounded in W^''\G) and F-elliptic, there exists
precisely one solution if this problem; let us denote it by
u*(x). (13.179)
Thus,
u* - we V, (13.180)
r ; = z"j + u* . (13.182)
We get
zJeF (13.183)
and
{{v, z"j + u*)) + ^ {v, z] - z;_i) = 0 Vi; 6 K. (13.184)
K
But
{{v, u*)) = 0 VD e F
with
z"o = - u * (13.187)
by (13.174) and (13.182).
DEFINITION 13.3. Let ii(t) be the very weak solution of the problem
(13.17)-(13.20) with
Mo(x) = -M*(x), (13.190)
where M*(X) is the weak solution of the problem (13.177), (13.178). The
function
u{t) = u{t) + u* (13.191)
is called the very weak solution of the problem (13.21) —(13.24).
^"- + Au = 0 (13.193)
8t
272 Chapter 13
Further, the form {{v, u)) is F-elliptic; thus, a constant c exists (indepen-
dent of vv) such that
^•') Thus, constructed from the functions z"(x) obtained by the solution of (13.172) to
(13.174).
Nonhomogeneous Conditions 273
"o = -u*eL2{G)
and u*{x) is the weak solution of the problem (13.177), (13.178). In particular
if the form ([v, u)) is symmetric. Theorems 13.4 and 13.5 are valid for the
function ii[t). Moreover,
(i) the function M*(X), being independent of /, has, taken as an abstract
function from / into W2'\G), derivatives of all orders with respect to t,
(ii) if the coefficients fli/x) have, in G, continuous derivatives of all
orders with respect to x (i.e. with respect to the variables Xj,j = I, ..., N),,
then the same holds for the function u*[x) in G.
Thus, we obtain
^*) This constant depends on the coefficient a of F-ellipticity of the form {(v, «)),
see [3], Chap. 33.
^^) Note that u e V, while u* e J^2***(G), however, u* ^ V.
274 Chapter 13
such that G' a G. Then, for every t e [;?, T], the function u(t) = u[x, t)
has, in G', continuous derivatives of all orders with respect to x.
See also Footnote 22, p. 237.
- H * e V).
See, however, Remark 13.10, p. 275.
DEFINITION 13.4. Let ^u(t), or ^u{t), or ^u{t) be the weak solution of the
problem (13.13) —(13.16), or the very weak solution of the problem (13.17)
to (13.20), or of (13.21)-(13.24), respectively. Then the function
u{t) = ^u{t) + ^u{t) + ^u{t) (13.198)
is called the very weak solution of the problem (13.1) —(13.4).
From the uniqueness of the (very) weak solutions of the corresponding
problems (Theorems 11.1, p. 217, 13.1, p. 255 and 13.7, p. 272), we have im-
mediately:
Note that when solving the problem (13.1) —(13.4) by the method of dis-
cretization in time with step h = Tjp we have to find, successively for
./• = 1,..., p, the functions
zj e Wf >(G) , Zj- weV, (13.199)
with
zoix) = uoix) , (13.200)
such that the integral identities
Let us note that it may occur for the problem (13.1) —(13.4) that regularity
results on the entire interval cannot be obtained when this problem is
"decomposed" into the problems (13.13)-(13.16), (13.17)-(13.20) and
(13.21)-(13.24) as accompHshed in this chapter. However, they may be
obtained when this problem is considered "on the whole". The situation
can be often simphfied if a suitable "particular solution" is found. Especially,
if
Ho ^ V, w iV, Uo - we V, (13.203)
then by the substitution
M = M* + z , (13.204)
where u* is the function (13.179) (cf. Remark 13.7. p. 271), the solution
of the given problem is reduced to the solution of a problem with homo-
geneous boundary conditions and with the initial function UQ — u* e V, etc.
Chapter 14
Let us return to the problem (13.1) —(13.4), i.e. to the parabolic problem
— + Au =f in 0 = G X (0, r ) , (14.1)
dt
u{x, 0) = Uo(x) , (14.2)
J5iM = ,9; on r X (0,T), (• = 1, . . . , / i , (14.3)
CiU = 0 on r X (0, T ) , J = 1,..., fc - /(.*) (14.4)
Let the assumptions of the preceding chapter be preserved; in particular,
let the corresponding bilinear form [{v, u)) be bounded in W^''\G) and
F-elliptic, with
V= {v;ve W^''\G), BiV = 0 on T, i = 1,..., / j , in the sense
of traces} , (14.5)
let a function w e IKj'X^) ^^^^^ ^"'''^ ^^^^
Bjw = Qi on r in the sense of traces , i = l , . . . , ,u, (14.6)
let
feL,{G), (14.7)
uo e L,{G). (14.8)
The solution by the method of discretization in time leads to the problems
(see (13.199)-(13.201), p. 275) of finding, successively for ; = 1, ..., .o,
such functions
ZjeW^'XG), Zj-weV, (14.9)
*) Recall that 2k is the order of the operator A.
277
278 Chapter 14
with
Zo(x) = «o(^), (14.10)
{{v,z,))+j{v,z,-z,)^{v,f) (14.13)
h
Thus, let
Vi{x),V2{x),...,v„{x),... (14.18)
be a base in V. Consider the first problem, i.e. the problem to find z^ e
6 VFf >(G) such that
Zi - weV, (14.19)
where the coefficients a^j are to be determined by the Ritz method. The
number Sj being given, such a function is unique.
Having obtained z*(x), insert it instead of Zi(x) into (14.12) and assume
the approximate solution z*{x) of the so obtained problem in the form
with coefficients 02; determined by the Ritz method again. Then insert
z*{x) instead of Z2(;i.) into (14.13) and find, by the Ritz method, the appro-
ximate solution
z*(x) = w(x) + X a^j v{x), (14.23)
j = i
etc.
Having received the approximations zj(x),j=l,...,p, construct
a function u*{x, t), the so-called Ritz-Rothe function, defined in Ij =
= [f;-i, tjl similarly as the function Ui{x, t):
Thus, let e > 0 be given. According to Theorem 13.12, p. 274 (cf. also
the text related to (14.17)), a sufficiently fine division of the interval /
exists — let the notation h be preserved for its step and, consequently,
the notation Ui(x, t) for the corresponding Rothe function — such that
can be attained for s^,..., Sp sufficiently large. Because of the form of the
functions Ui(x, t), u*{x, t) (being piecewise hnear in t), it is even sufficient
to show that
Denote
-^ = r] (14.29)
Putting z|(x) instead of 22(x) in (14.13) and denoting by Z3(A) the solu-
tion of the obtained problem, we get, quite similarly,
Iz, - z4 ^ \\z, - zt\\ < 2^ . (14.39)
Thus, choosing S3 in (14.23) sufficiently large, so that
|]Z3 - r*|| < r^ (14.40)
be fulfilled, we get
Thus (see the text preceding (14.28)), (14.27) is obtained, yielding together
with (14.26)
||I/(A-, t) - wf (x, Oil < E Vf 6 / . (14.44)
^) The case Bi{t) « = 0, C;(^) « = 0 can as well be considered if the space K(see (15.15))
remains independent of /.
283
284 Chapter 15
\v\ W2<'<)(C)
if u e V.
Moreover, the form ({v, w)) will be assumed F-symmetric for the sake
of simplicity, i.e.
((y, „)) = ((u, I,)) WV,UEV and Vt e I. (15.17)
As to the dependence of ((i;,«)) on t, we shall assume that positive
constants K^, K2 exist, independent of v, u, t, h such that
h
for all r, u e Fand f, ? + /?,< + 2h el,h > 0.^)
Further, let
f{t)eL2{G) \ftel, (15.20)
and let a constant d exist such that
fit + h) - f{t)
^d ^t,t + hel. (15.21)
h
Here, a remark similar to Footnote 2 can be added referring to the
function / ( x , t).
^) Practically, the assumptions (15.18), (15.19) mean that the functions aiAx,t)
and their first partial derivatives with respect to t satisfy in Q the Lipschitz condition
with respect to 1.
286 Chapter 15
M' (15.24)
in what follows, or by
ll^ll,^ (or briefly by ||!;||,) (15.25)
at the points of division tj. Thus, we have
IKil< = %%t;)),. (15.26)
Each of the problems (15.33), (15.34) has exactly one solution: First, let
j = 1. We have
Zo = 0 , / i e Li
(see (15.20)) so that
^+/ieL2(G).
/)
Since the form ((y, w))i is bounded and F-elliptic (see (15.12), (15.13)), so
is the form
{{v, M))I + -- (u, u)
h
(Lemma 3.1, p. 78). Thus, there exists precisely one Zi € V satisfying
(15.34) for j = 1. Now,
Zi 6 K=> Zi 6 L2{CJ) so that
— + /z e l^ii^^).
h
and there exists exactly one Zj e ^satisfying (15.34) (orj = 2, etc.
In this way, the functions Zj, ..., z„6 F are uniquely determined. It is
thus possible to construct the Rothe function Ui(x, t), or Ui(t) if it is con-
sidered as an abstract function from / into V, defined in / by
u,{t) = z,_i + Zj . {t - O-i) in / , = [ 0 - 1 , tj] , (15.36)
j = 1,..., p. Here
ZJ = ^J^LIJ^ . (15.37)
288 Chapter 15
and
j = 1,...,2"-V.
To be able to say something about the behaviour of the sequences
{u„(f)}, {u„{i)] and {Vn{t)} we have to derive some apriori estimates.
Consider the first of the integral identities (15.34) and insert z^ for v.
Inasmuch as ZQ ~ 0, we get
for
((-^i,z.))i^O
according to (15.13)
Similarly, putting r = c, in the second of the integral identities (15.34),
and taking into account that
{{Z„ Z2))2 ^ 0 ,
we obtain
-iz„z,-z,)^{z„f,) (15.47)
h
and
\m ^ 1^4 + Hf2\\ ^ M!l/ii + liAlD- (15.48)
In general, we get, in this way,
Thus, it is very easy to obtain an estimate for \\zj\\. However, the simple
way of estimating |]Zj|| shown in Chap. 11 cannot be apphed here, since
the difference of the consecutive integral identities gives
((z; - Zi-i.Zi))i =
by (15.13),
|(-../y)|^^i||/|!- (15.62)
Thus, (15.57) yields
a
+ 2ci||/|| + 2 c , 4 ; - l ) / 0 . (15.63)
Replacing [j — l)/i by T here, we find that (15.63) is an inequality of the
form
| ! z , | | ^ ^ ^ + B/,(||z,||^ + ... + ||z,.,||^) (15.64)
with
^ = ? ( | | / | + Td)c,, (15.65)
a
B =^ . (15.66)
a
According to relation (15.29) in Remark 15.3, we then have
|]z,||^^^e«<-''\ j=-l,...,p, (15.67)
and, finally, replacing (_/' — l)/i by Tonce more,
||z,||^^^e«^ (15.68)
with A, B given by (15.65), (15.66).
Or, brieHy written, we have
\2JIV^C2, j=h...,p, (15.69)
with
c2 / r ^ ( l | / | ! + W)e^-'^/^1. (15.70)
292 Chapter 15
((Z„z,)), = i - ( i z J j - | | z , _ , | j j + /,^!JZjj),
2h (15.75)
(15.76)
Thus,
{{Zj,Zj))j-{iZj,Zj_,))j_,=
h
^(IM + IM-O + ^dl-'Jj-
N!l-INIl-i> \N\l-Ml
and of terms of the form
-N|^,+2ilz,.||f-||z4f_,.
According to (15.18), (15.19), we have
||z..||f - ||z,l|f_,| ^ hK^z^Wr S hK,c\ , (15.81)
Mli - n4i + hiWl^ ^ h'K^z^y ^ h'KA . (15.82)
Thus we get, finally,
\Zjf ^ \\Z,f + {j - 1) hd' + 2K,cl + (j - 2) hKA +
+ /,(||Z,f + ... + | | Z , f ) . (15.83)
Now, putting V = Zi, (15.34) yields immediately
!iZii^i/ii^li/|l- (15.84)
Thus, if we replace, at the same time, (j — 1) h and (; — 2) h by T in
(15.83), we get
l l Z , P ^ ^ + B / , ( i Z , r + . . . + iZ,r) (15.85)
294 Chapter 15
with
A = | | / P + Td' +{2K^ + TK2)cl
and
B = 1. (15.86)
If convergence questions are discussed, h < ^ can be considered without
loss of generahty. (15.30) and (15.32) then yield
The constants c^, Cj, c^ in the estimates (15.51), (15.69), (15.88) do not
depend on h. Thus, these estimates remain valid for an arbitrary division
d„. Consequently, we may write
for every positive integer n and 1 ^ 7 ^ 2""^p, with the constants Ci, Cj, c^
given by (15.50), (15.70) and (15.89).
In virtue of (15.91), (15.92) the sequence of functions u„{i), or t7„(f),
given by (15.38), or (15.44), is bounded in the space £2(7, V), or L2(/, 1-2(0)),
respectively. Thus, a subsequence {u„J^i)], or {U„Jifj\, may be found,
converging weakly in L2{l, V), or in L2(/, ^2(0)), to a function u e L2{I, V),
OT U e L2{l, -^2(0)), respectively,
Thus,
u e AC{I, L2{G)) , (15.95)
U = u'e L^il, L2(G)) , (15.96)
u(0) = 0 in C(/, L2{G)). (15.97)
(cf. Chap. 11, pp. 208 — 211), where u„(t) are the functions (15.43). Further
(the same indices can be used without loss of generality),
Let us define the sequence of abstract function f„(t) from / into L^l^j) by
/," for r= 0
lit) = ( (15.100)
^/; for tei] = {fj_„f;\.
P'urther, let us denote by
\v,ul, (15.101)
the form {{v, u)), modified in the interval 7 in the following way: When
considering the division J,,, then in each of the intervals 7" the form (((,-, i;)),
is replaced by the form {{v, w)),-^)
Let v{t) be an arbitrary function from L2(7, V). In virtue of the integral
identities (15.40), of the definitions of the functions u„(t), lJ„(t),J„{t) and
of the form [v, u]„, we have
^) Practically, this means that the coefficients aij(x, t) of the operator A are replaced
by the function a;,-(;c, /") in these intervals.
296 Chapter 15
Now, Jo
r \y{t), u„Xt)\^ dt +
)
{v{t),u„j,t))dt:^
Jo
(KOJJO)d/.
(15.104)
i: {(v(t), u{t))) dt
*) QV(0.«(0))d^J^(£x||KO|K||"WiKd/J^
<K' \v{t)\\^dt. \u(t)\\'ydl.
Time-Variable Coefficients 297
We have
If
IJo
(WO. "JO].. - Mt), <7„i0))} dt <
/ Y r \[^^m..,.-{i^,^T))\<i'^l^n,TK,c4vl (15.111)
in virtue of (15.18). (For Cj see (15.70).) Thus, (15.109) is vahd for the case
v{t) = r. In a similar way, validity of (15.109) is proved for the case that
veN. (For N see Chap. 11, p. 209; thus, v{t) is a step function in /.) Since
the set A' is dense in £2(7, V), (15.109) is valid for every function v e £2(1, V).
u €L2{I,V), (15.113)
u'eL2{l,L,{G)), (15.115)
n
to be satisfied for all v e V. Here,
Zo = Mo e LiiG),
while
||uoi + 0 (15.120)
is assumed.
Insert v = Zj, i; = Zj, ..., tJ = Zj successively into (15.119) and perform
the summation. One gets
t{{^i^-.))i^^bor. (15.122)
1=1 2/7
i(||z2||^ + | | z 2 - Z , | | ^ - | | z , | | ^ ) + i | z 2 - Z , f = 0 ,
whence
N|^^iz,l|^ (15.123)
300 Chapter 15
Similarly, we get
ll-3i^^Ni^
lk;-2||^.^llz;-3i|j-.,
From assumption (15.120) it easily follows that ||z,|| 4= 0.') Thus, positive
numbers j?;, i — 1, ...,j — 1, exist such that
H\4\lx-\H"- (15.125)
Then we get
Pj-i
whence
||z,_,||j_,^^,_,||z,||j, (15.126)
and, similarly,
Nif.t" - ^
') The first of the integral identities (15.119) cannot be satisfied for Zj = 0, etc..
Time-Variable Coefficients 301
P , ^ l - ^ , / = l,...,j. (15.129)
a
Thus (we may assume that K^hla. < l),
ft-A.....fts(i-^)
\P
= fl-^Y"^fl-^Y. (15.130)
For all p sufficiently large we have
1 _:?^i.lY> 4e-*'''/".
ap
Assuming that /; possesses this property already,*) we get
((zi,zO)i + ( ( z „ z , ) ) , + ... + ((z,,z,)), ^
^ie-^'^/'X(z;.z.));- (15.131)
(15.122) then yields
^K4v\\ylu\\y, (15.134)
A g K4v\\y My (15.135)
h
where
A = \{{v, «)),+4A - 4((t', M)),+ 3ft -I- 6{{v, u)),^2H -
- 4((i;, u)),^„ + {{v, u)),\
for all v,ue V{h > 0). Further, positive constants dz, d^ exist such that the
inequaUties
1/(0 -f- 2h) - 2f{t + h) +/(0[ ^ ^^^ ^ (^3^3^^
1
h
|/(t + 3/2)-3/(f + 2/») + 3/(f + / 0 - / ( t ) | ^ ^^ ^^5_^3^-,
Time-Variable Coefficients 303
||Z;i|, g c ^ , (15.143)
where
C4 = / j - riCM?(l + TKi) + cl(^ + 2K2 + TK,\ +
\\4\ ^ cs , (15.145)
with
CK = \{M2 + df + ^ - K i A f J + 2cl{Ky + 2K2T+ KjT^) +
305
306 Chapter 16
u{x, 0) = 0 , (16.12)
u(0, 0 = 0 , M(1, f) = 0 . (16.13)
The "corresponding" elhptic problem is
— u" + 20u^ = sinKX , 0 < x < 1 , (16.14)
M(0) = 0 , u(l) = 0 . (16.15)
Choosing
F = {v,veW^'\0, 1), v{0) = 0, i;(l) = 0} = Wi'\0, l ) , (16.16)
multiplying (16.14) by an arbitrary function ve Fand integrating between
0 and 1, we get, in the usual way,
and
((., »)) v'u' dx + 20
L' vu^ dx (16.19)
REMARK 16.1. The theory of nonlinear elliptic problems is not simple. Al-
ready the formulation of such problems requires the introduction of a lot of
new concepts. The purpose of our book is not to acquaint the reader with
a general nonlinear theory. In this direction, the reader is referred to other
books, especially to [5]. Our aim here is to show the reader how can the
method of discretization in time be applied to the solution of nonlinear
parabohc problems. To keep the main ideas of the method as clear as
possible, a rather narrow class of such problems will be treated here, in order
that assumptions as simple as possible could be chosen which are sufficient
for deriving an existence theorem. (For more general results see the papers
referred to at the beginning of this chapter, in particular, the "classical"
paper [14] by J. Kacur.) In the first place, we shall assume that the space V
and the form {{v, u)) corresponding to the problem (16.5) —(16.7) are given,
((f, u)) being the Gateaux differential (see below) of a functional with some
special properties. This conception will make it possible, moreover, to give
some information concerning numerical aspects of the solution of elhptic
problems generated by the method of discretization in time.
') Thus a linear space with a norm i|.|| satisfying (2.9)—(2.13), p. 10. Examples of
normed spaces are Hilbert spaces, Banach spaces, etc. See, e.g., [3], Chapter 7.
308 Chapter 16
^) Note that, ug, v being fixed elements of V, F(UQ + tv) is a function of the real
variable /.
^) The common notation is F'(UQ, V). We use the notation F'(v, UQ) to be in conformity
with our standard notation ((v, «)) for integral forms. (As announced above, we shall
consider problems where ((y, «)) is the Gateaux differential of a functional.)
Nonlinear Parabolic Problems 309
REMARK 16.3. It can be shown that under these assumptions the relations
'^) We have
d ^,. X ,. F'(v, u + (t + At) z) - F'(v, u + tz)
— F'{v, u + tz) = hm —^ 5^ '-^ ^- =
df A(^o Af
_ F'{v, u + tz + z At) - F'{v, u + tz) _
Al-O At
(16.36)
because F"(v, z, u + tz) is a continuous function of the variable ; in consequence of the
assumption 3).
310 Chapter 16
^) Let us note that if 5) is omitted, 6) and 8) need not hold, in general. (It is sufficient
to consider the case of an unbounded linear functional on V; then F'{v, u) = Fv, while
F"(v, w, u) = 0, so that the assumptions 1)—4) are fulfilled.)
^) The xistence of the first integral follows from (16.39). Further, u e Kimplies ue AC.
• ([0, 1]) (what has been applied in the preceding text already), so that the second in-
tegral is well defined for every ue V.
Let us note further that there exists such a number y (see e.g. [2]) that the inequality
!l"l|c([o,i]) g 7|i«lk.u)(o.i) VME wf\Q, 1) (16.41)
holds, with
ll"i|c([o,)]) = "lax \u{x)\ . (16.42)
^cstO.l]
Nonlinear Parabolic Problems 311
Thus
F'{v,
^'{v, UQ)
Uo) == I V'U'Q
v'u'o dx + 20 rug dx (16.44)
Jo
The second d i f f e r e n t i a l :
2) The hnearity of the functional F"(v, w, u) in i' (u, w being fixed) as well
as in w (u, v being fixed) is obvious.
3) (16.41) yields
u„^u in V=> u„ -> u in C([0, l]) => u^-^ u^ in C([0,1]).
Denote
y - ul\\c = max \u'{x) - u'„{x)\ = k„. (16.47)
xe[0,l]
< eokJvL
because v and w are fixed elements from V. (16.48)
4) Further, let ue M, i.e. let lu\\y ^ k. Then by (16.41)
u e M => [|i/||c = max \u{x)\ ^ yk . (16.49)
xe[0,l]
Consequently
f1
tjwu^ d x <
r ^y| Iw] y^fc^ dx g y^/c^|]y||K |!vv|| (16.50)
«0 «0
In (16.32) it is thus sufficient to put
C = 1 + 60y^k^ . (16.51)
Fu = Fii - F(0) =
Then)
a) there exists exactly one point UQ e V, at which the functional G
assumes its minimum on V,
min Gu = GUQ ; (16.54)
Let us note that this lemma implies existence (and uniqueness) of a weak
solution of the problem (16.14), (16. 15). In fact, let
^ - ^ ^ • ^ ( O , 1). (16.56)
Gu = Fu — \ fu dx , u e V, (16.57)
where F is the functional (16.40) and / = sin 7tx. We have feL2{0, 1),
so that G is defined on the whole V. Now,
^) From (16.53) and from the assumptions of Remark 16.2 it follows that Fis a coer-
cive and strictly convex functional on V, and this ensures validity of Lemma 16.1.
314 Chapter 16
according to (16.46). Thus the functional G"{v, w, u) has all the properties
2) —4) from Remark 16.2. At the same time,
G'{v, 0) = — i; sin nx dx ,
Jo
and this is a bounded linear functional on V (because
— I V sin nx dx
0
|(i;, sin nx)\ ^ ||sin 7tx|| < sin nx\: H')-
Thus the functional G satisfies all the assumptions of Lemma 16.1.
Consequently, there exists exactly one Uo ^ I^for which its Gateaux difleren-
tial is equal to zero,
Let us turn to the solution of the problem (16. l) —(16.4) by the method
of discretization in time. As announced above, we assume everywhere that
the space Fas well as the form ({v, u)) are given.
Assumptions:
(i) Fis the space W2''\G) or some of its subspaces.^)
Fu = - f u'^ dx + 5 w* dx (16.70)
2J0 Jo
which satisfies, as shown above, all the assumptions l) —4) from Remark
16.2. Further, (16.66) follows from (16.60) and (16.61).
The assumptions (16.65) and (16.67) are obviously fulfilled.
' ) This not very essential assumption is introduced here in order to obtain the inequality
(16.72) and to simplify further considerations, in this way.
The reader may wonder why this assumption has not been introduced in Remark
16.2 already, because it implies the assumption 5) from the quoted remark. However,
it was necessary to consider a rather more general assumption 5) there, since the
functional (16.57), investigated in connection with J-emma 16.1, does not have the
property (16.65).
316 Chapter 16
WiM = Fu + — j u^ dx - j fu dx , (16.76)
Now, according to the assumption, the functional F has all the properties
2 ) - 4 ) from Remark 16.2. Since l//i > 0 is fixed, it follows from (16.79)
that the same holds for the functional Hi. Moreover, by (16.79) we have
ir, +/eL,(G)),
/?
to the conclusion that there exists precisely one Z2£V for which (16.74)
is satisfied if j = 2, etc.
with
zS = 0 . (16.89)
Denoting
Z;(x) = ^ f c l z i i l ^ l M , (16.90)
Z\ for ?= 0
(7„(0 =. ( (16.93)
^Z; for tel),
l|2iNi/||. (16.95)
(z,,z,-z,_o^o.
11) It is seen from the preceding text (as well as from the following one) that monotony
of the form {{v, u)) (implied by (16.66)) and the property 4) from Remark 16.2 play an
essential role in the whole theory. This is the case in more general problems (where,
for example, a Banach space is investigated instead of a Hilbert space considered here).
320 Chapter 16
from where
||Z;|| ^ ||Z,_i||, j = 2,...,p. (16.98)
Consequently,
||Z,|| ^ | | / | | , .7 = 1,..., p., (16.99)
by (16.95).
It follows immediately (we have ZQ = O) that
|jz,||^/<||Z,i+...+ |Z,ll)^;/,||/|i^T||/||. (16.100)
Finally, putting v = Zj into (16.74), we obtain
^•^) Note that the function (16.91) is piecewise linear in t and assumes the "values"
z" at the points t" of division.
Nonlinear Parabolic Problems 321
and
U = u' in L2(L L2(G)). (16.114)
Consequently,
u eAC{l,L2{G)), (16.115)
M'6 L2(/, L2(G)), (16.116)
t/(0) = 0 in C{I,L2{G)). (16.117)
Now, let i;(f) be an arbitrary function from £2(7, V). Let us define the
abstract function / from I into L2(G) by
f{t)==f Vie/. (16.118)
^^) See Remark 11.2, p. 216. Compactness of the identity mapping from Kinto L2(G}
is applied here as well as equicontinuity of the functions «„(r)in C(I, Z,2(<?))-
322 Chapter 16
Noting (16.74) and taking (16.92), (16.93) into account, we can write
^*) A question may arise concerning the existence of the first integral in (16.121).
However, this existence can be easily established: For n fixed, «„(/) is a step function
in L2U, V)- If v(t) is a step function too, then the considered integral is well defined.
Now, for every ue Vfixed, ((y, u)) is a linear (bounded) functional in V. Moreover,
the step functions form a set dense in L2U, V) (cf. the set A^ applied in Chap. 11, p. 209).
In this way, the existence of the first integral in (16.121) is ensured. For the two re-
maining integrals this fact is obvious.
* ^) Existence of the second of the considered integrals can be established without dif-
ficulties using step functions as in Footnote 14. (The so-called Minty trick can also be
applied to prove (16.123).)
Nonlinear Parabolic Problems 323
Now
\Uj gcj V«6 7. (16.125)
u,,^ -+ II in 1.2(0) uniformly in / (16.126)
Consequently
•r
I ( ( < ( 0 - u{t), u„^{t))) dt-^0 for n, CO . (16.128)
Further,
u„^ -- u in L2{I, V) (16.129)
by (16.108), so that
(16.133)
= "-irik "iiv'
324 Chapter 16
by (16.66). Consequently,
SC\\v{t)\\y\u„St)-u{t)\\y (16.136)
for almost all tel. Thus
\\{v{t),u„St)))At-[\{y{t),umdt <
Jo Jo
ueL2{l,V), (16.139)
ueAC{l,L2{G)), (16.140)
u'e L^il, L,{G)), (16.141)
is fulfilled.
Now,
u(a) =i 0 in L2{G).
Unicity imphes, in the usual way (see Remark 11.1, p. 215), that not
only the subsequence {M„,(/)} but the entire sequence {«„(?)} converges to
u{t) weakly in LzQ, V), and strongly in C(l, L2{G)).
REMARK 16.7. The existence of a weak solution for the case of nonhomo-
geneous initial conditions can be derived similarly, provided that the initial
function UQ{X) (and other data of the problem, if needed) are sufficiently
smooth. Thanks to the monotony of the form ((v, «)), continuous depen-
dence in L2(G) of the solution on the initial function UQ{X) can be easily
proved. By a limit process shown in Chap. 13 it is then possible to arrive
at the concept of a very weak solution for the case that MQ e ^2(0) only.
Such a problem (thus with an initial function UQ e L2{G)) is treated numeric-
ally in Chap. 5.
tb^v, (16.152)
i=l
holds for a c > 0 independent of v, u. Then there exists exactly one UQE V
for which G assumes its minimum on V,
min Gu = GUQ , (16.155)
ueV
LEMMA 16.2. Let the assumptions of Lemma 16.1 be fulfilled. Let {V„}
he a sequence of complete subspaces of V, satisfying (16.153). Then:
(i) there exists exactly one u„e V„ for which the functional G assumes
its minimum on V„,
min Gu = GM„ ; (16.157)
ueV„
^ ^) The simple applicability of Lemma 16.2 was one of the main reasons for the above
chosen conception of this chapter.
Nonlinear Parabolic Problems 329
was used (see pp. 316 — 317 where the functional H j , / / , were investigated
in detail) satisfying all the assumptions of that lemma provided the functional
F satisfied the assumptions (i) —(v), pp. 314, 315. In that way, the unique
solvability of the problems (16.73), (16.74) was proved by (16.156) because
REMARK 16.8. Thus, let us apply this method. Practically, this means
that we assume Zy ^ in the form (16.152), then put this sum for w into (16.160)
and look for the minimum of the so obtained function — let us denote it
briefly by
Hj{b„...,K)- (16.163)
on the set of vectors (foj,..., b^). This problem leads to the solution of the
330 Chapter 16
system
—^ = 0 , . . . , ^ = 0 . (16.164)
dbi dbs
Since this system is nonlinear its solution need not be simple. In this case,
the following procedure can be applied:
Choose
K...,b',) (16.165)
arbitrarily and minimize the function Hj(^i, b^, •••, b°) of one variable
Ci (keeping b^, •••, b° fixed). Obtaining <^° in this way, minimize the func-
tion Hj{tu £,2, bl,..., b°) (keeping ^?, fo°,..., fe° fixed), etc. This iterative
process can be shown to be convergent provided the functional //, satisfies
the assumptions of Lemma 16.1. (See [5].)
This approach was chosen in the numerical solution of the example
treated in Chap. 5.
16,,..,, (16.166)
i= 1
Vfe^^tv, (16.167)
•-1
Nonlinear Parabolic Problems 331
" i ( 0 = z;-i + - ' " / ' ' " ' {t - O-i) in ^- = [ 0 - 1 ' 0 ] .
/i
(16.168)
constructed on the basis of the functions Zj, let the function u*{t), defined
in I by
can be made arbitrarily small, for all tel, if h is sufficiently small and
Si,.... Sp are sufficiently large. In such a case, we shall say, briefly, that
the Riiz-Rothe method is convergent for our problem.
can be achieved.
In virtue of the monotony of the form {{v, «)) (see (16.71), p. 315), we
can proceed in almost hterally the same way as in Chap. 14.
Denote
n = ^ , (16.175)
2p
where, as usual, p = Tjh. Choose s^ in (16.166) so large that
llzi - ztll ^ /;. (16.176)
According to Lemma 16.2, such a choice is possible. ^^)
Insert z* into H2 instead of z^. According to the above text, the obtained
functional H2 attains its minimum, in V, at a point Z2. For the function
Z2, the integral identity
{{v,S2)) + -iv,Z2-z*) = {v,f) yveV (16.177)
h
holds. Subtract this identity from the identity for the function Z2,
(this even in the case that the iterative process mentioned in Remark 16.8 is applied).
Nonlinear Parabolic Problems 333
1; = Z2 - Z2 , u = Z2 ,
it follows that
Hence, by (16.179),
p2 - 4\\ ^ n • (16.181)
Then
IU2 - z*|| ^ In . (16.182)
\\h-z%\\^r,, (16.184)
we obtain
which implies (16.173) and with (16.172) the required result (16.171).
According to what has been said above we can summarize the obtained
results briefly into the following
The first steps dealing with the theoretical as well as numerical aspects of
these problems were undertaken in my seminar (A. Wohlmutova, F. Bubenik).
A more profound investigation of problems connected with equation (17.1)
and solved by the method of discretization in time can be found in [10].
This paper was taken as the basis for the conception of the present chapter.
335
336 Chapter 17
du
+ ^u = K{x, T, u{x, T)) dt + / ( x ) ' ) in Q = G x {0, T),
dt Jo
(17.4)
u(x, 0) = 0 , (17.5)
BiU = 0 on rx(0, r), i = l,...,n, (17.6)
CiU = 0 on r X (0, r ) , i = 1,..., fc - A(. (17.7)
A= 1 (-l)l'l7)'(fl,,(x)D^), (17.8)
(17.6) and (17.7) are, respectively, stable and unstable boundary conditions
with respect to this operator, the operators 6,-, C; satisfy conditions from
Chap. 2. (In particular, the operators JB,- are of the form (2.214), p. 53, and
satisfy (2.215).)
The bilinear form
{iv,u)) (17.9)
All = f in O,
B,M = 0 on r , ! = 1, ..., ju ,
CiU = 0 on r , i = 1, ..., k — n
*) The function/is assumed to depend on x only, in order that the idea of the method
be well visible on this more simple case. Extension of the obtained results to the case
f=f(x, t) makes no difficulties.
Integrodifferential Parabolic Problems 337
if V G K
Further,
feL^iG); (17.13)
the kernel K{x, t, u) is Lipschitz continuous in the variables t and u- in th&
following sense:
!|X(x, t2, U2) - K{x, tu u,)\\ S C{\t2 ~h\ + 1|«2 - Mill), (17.14)
and it satisfies
tel = [0,T], ueV=>KeL2{G). (17.15)
Under these assumptions, Theorem 17.1, p. 352, will be derived which is
an existence theorem for the given problem and a convergence theorem for
the corresponding method of discretization in time.
Kj{x) =K{x,jh,Zj{x)),')
(17.21)
^) Note that boundedness of ||«|| is not assumed so that some preliminary investigations
or estimates are to be carried out frequently. See the quoted chapter.
^) We take
{{''' ^j)) + 7 (^' ^j) = 7 (^' 2y-i) + K^' Ko + ...+ X,._i) + (vj)
h h
Vy e F ) . (17.25)
Note, first, that under the given assumptions each of the considered
problems has precisely one solution: Since the form ((f, w)) is bounded
and F-elliptic, so is the form
^Zo + hKo+feL^iG)
/•2h
K(X, T, ti(x, T)) di = K{X, T, U{X, T)) dt +
Jo
and there exists precisely one z^ e F satisfying (17.25) for 7" = 1. Similarly,
we have
zi e F=> zi 6 £2(0) and K^ e L^iG),
so that
|zi +b{Ko+K,)+feL,{G),
h
Zi6F,...,z,6F (17.27)
or
i/„(0 = z;.i + z ; . ( f - f ; _ o , (17.32)
Integrodiflferential Parabolic Problems 341
where
and z"(jc) are weak solutions of the analogues of the problems (17.22) to
(17.24)
Following the idea of Chap. 11, we show (for the precise formulation
of the results see Theorem 17.1, p. 352) that the sequence of Rothe functions
u„(i) is convergent in a certain sense and that its hmit u{t) is the solution
(in a weak sense) of the problem (17.4) —(17.7). Since many of the corres-
ponding investigations are similar to those carried out in detail in Chap. 11
we shall proceed slightly more rapidly, using some auxiliary results of the
quoted chapter.
To begin with, we derive an apriori estimate for the norms of the functions
Zj(x) in L2(G). This represents the fundamental step toward obtaining
apriori estimates needed for the proof of convergence mentioned above.
Consider the first of the integral identities (17.24). Put v = Z^. We get*)
Subtracting the integral identity (17.24) written for j = 1 from itself written
for i = 2, we obtain
((Zi,zi)) = -((zi,zi)):£ 0
h
by (17.12). Consequently,
( Z i , Z i ) ; g ( Z i , A ^ o + / ) = > | l Z i | | ^ \\hK^^-f\\
because
\{Z^,hK^+f)\-^ IIZJI llAiiTo +/II
by the Schwarz inequality.
342 Chapter 17
g \\K4 + ch (2 + I^^-::!^! + M] ^
\ h hJ
= |K!| + C/i(2+I|Z,|| + ||Z,||),
Thus, \\K""\ and \\z"j\\y are uniformly bounded with respect to n and j
as well.
/ = [0, r]
is used as above. Thus, the sequence of functions u„(t) is uniformly bounded
in Kfor all n and all t e I, and, the more, bounded in the space L2(/, V)-')
Consequently, a subsequence {w„^(f)} can be found converging weakly
in this space to a function u e 1.2(1, y),
Lei us define the "abstract step functions" Ujl) from / into ^2(0) by
Z"i for t = 0
Ht) = (
^ Z"j in the intervals ij = {fj.^, /;] , ./ = 1,... 2"- V ,
(17.52)
In consequence of (17.42), the norms |lL'„(t)|| are uniformly bounded with
respect to ;? and t, so that the sequence {U„(t)] is bounded in L2(/, L2(G)).
Thus, a subsequence {U„^(t)] (cf. Footnote 6) can be found converging
weakly, in this space, to a function U{t),
U„^^U in L,{1,L,{G)). (17.53)
Similarly as in Chap. 11, the function U(f) can be shown to satisfy
[JI
Jo
U{T) dr = u{t) (17.54)
To this aim, let us consider, first, the sequence {u„{t)} of abstract functions
from 1 into V defined by
z"i for / = 0
"„(0 = ( (17.58)
^z"j in 7; = ( f ; - i , t ; ] , 7 = i,...,2"-v.
where u(t) ~ u{x, t) is the function (17.50). Since ue F for almost all t e I
we have, by (17.15),
we L2(G) for almost all te 1. (17.62)
We show, first, that
||w(/) - w„J/)|| < 8 for almost all tel (17.63)
if n, > no{ey)
For a fixed division d„^ and for an arbitrary t£l"\i we have, by (17.14),
) I.e. for all sufficiently fine divisions of the interval /; here, w^ are indices from (17.50).
Integrodifferential Parabolic Problems 347
Moreover, since the functions w„{t) are simple functions (Chap. 11, p. 201)
from I into L2(G), (17.65), (17.63) imply that the function w[t) is Bochner
integrable as an abstract function from / into L2(C) (Chap. 11, p. 202).
Denote
H ; ( 0 = rvv„(T)dT, (17.67)
h„Kl for (= 0
\^„{t) = ( (17.68)
^hjKi + ... + K ; _ O in I-;.
(17.63) impHes, immediately, that
lim W„, = W in L^il, L,{G)) .') (17.69)
From (17.67), (17.68) we get, for (s /",
' ) Even in C(I, LjiG)). However, this fact will not be exploited in what follows.
*°) Here, we cannot write "in C{I, LiiG))", because the functions W„^<,t) are not
continuous in /. Nevertheless, the convergence in L2{G) is uniform in /.
348 Chapter 17
where
W{t) = f K{x, T, u{x, t)) d- (17.77)
by (17.66), (17.61).
Thus, the function u{t) satisfies (see (17.50), (17.56), (17.55), (17.57))
) The form ((v, u)) is bounded in V. For the details see Chap. 11, p. 213.
350 Chapter 17
" ^ (17.89)
According to (17.14), we have, first,
The function «(r) belongs to C(/, L2(G)) so that the (real) function ||u(f)||
is continuous in / and the more in the interval [0, / ] . Consequently, ||M(0!|
attains its maximum on this interval at a certain point 4 e [0, /J,'^)
max \\u{t)\\ = j|«(^t)l| . (17.92)
te[0,U
(17.90) yields
il"COil = 0 • (n.9E)
(17.92) then imphes
u{t) = 0 in L^iG) V? G [0, /] . (17.99)
Performing the same consideration in the interval [/, 2/] with the function
u{t) for te[0,^t]
v(t) = (^ (17.100)
^0 for tE{^t,T],
where ^Ms a point at which \\u{t)l attains its maximum in the interval
[/, 2/], and using the just obtained result (17.99), we get
u{t) = 0 in L2(G) \/t e [/, 2/] . (17.101)
After a finite number of steps we thus come to the conclusion that
w(r) == 0 in L2(G) "itel, (17.102)
which was to be proved.
REMARK 17.2. The uniqueness just proved imphes that not only the se-
quence {M,„XO but the whole sequence {u„{t)} converges to the weak solu-
tion u{t), weakly in L2{l, V) and strongly in C(/, ^2(0)). See Chap. 11, Re-
marks 11.1 and 11.2, pp. 215-216.
Let us summarize the obtained results into the following theorem:
2t=t"uVi, (17.108)
1=1
where the coefficients a^j are to be determined by the Ritz method. The
number .s, being given, the function z* is unique.
Consider the second of the integral identities (17.24) (thus for j = 2)
and replace Zj, and Ki, by z*, and K*, respectively. Here,
K*ix) = K{x, h, z*{x)). (17.109)
Denote by
Z2GF (17.110)
the (unique) function satisfying the thus obtained integral identity
2*=-|a,,r,, (17.112)
i=l
where (s2 being fixed) a2i are found by the Ritz method again. Then insert
z*, K*, KI, where
into (17.24), written for j = 3, instead of Zj, K^, K2, respectively, denote by
z^eV (17.114)
^-t^ziVi (17.116)
i=l
zl...,z*. (17.117)
We show that the weak solution u(t) = u(x, t) of the problem (17.4) to
(17.7) can be approximated, to arbitrary accuracy, by the function u*{t).
In detail, we show that e > 0 being given, the inequality
\\u{t) - u*{t)\\ < £ Vre7 (17.120)
can be satisfied if the numbers p, s^,..., Sp are sufficiently large.
In fact, from Theorem 17.1 it follows, first, that
Since the functions u^if), uX{t) are piecewise Hnear in t, it is sufficient to show
that we are able to achieve
Thus, let p be so large that (17.121) be fulfilled. Let ;; > 0 be given (to
be specified later). Choose Sj in (17.108) such that
Ijzi - z t l l ^r]. (17.124)
Such a choice is possible. Further, choose Sj so large that
+ /i^C(||zi-zt|| + ||z, - Z I I I + . . . +
+ I|z,_i-z*_J|) + ;,. (17.134)
Addition of the inequalities (17.124), (17.131), (17.133), ..., (17.134) yields
| z , - z ; | | g ^ + /i^C[(;-l)||z, - z t | l +
+(;-2)||z,-z*||+...+||z,_.-z^.||].
Now,
( ; - i ) ; i < T, ( ; - 2 ) ; j < T,...,h < T,
iQtegrodiflferential Parabolic Problems 357
so that, finally,
\\zj - z*\\ Sjl + hCT{\\z, - z*4 + \\Z2 - z1!| + ... +
+ ||z,_,-z*_,]|). (17.135)
Noting (17.124) and applying Remark 15.3, p. 286, we get
||z, - z*\\ S Pn e^^"-'" ^ Pil ^"', y/ = 1, ..., p . (17.136)
Thus, it is sufficient to choose
u{x, 0) = 0 , (18.3)
Biu = 0 on r X (0, 7"), i = l,...,fi, (18.4)
C,(M) = 0 on r X (0, T) , i - 1,..., fc - /(. (18.5)
As concerns the domain G, the operator A, the boundary operators B„ C,
and the bilinear form ((r, w)) corresponding to these operators (i.e. to the
elliptic problem
Au — f in G,
B^u = 0 on r . i = 1,...,/(,
C,« = 0 on r , i = 1, ...,/< — /(,
in the sense of Remark 2.15, p. 60) see the foregoing chapter.
Similarly as in that chapter, we shall assume that the form {{v, «)) is
bounded in W^^{G) and F-elliptic, i.e. that positive constants K and a
exist (independent of u, v) such that
358
Problem with an Integral Condition 359
and
{{v, v)) ^ o(\\v\\^ ^veV, (18.7)
where
V = {v; V 6 W^''\G), B,V = 0 on r in the sense of traces,
/=1,...,A']- (18-8)
(Here
stands for
if t; 6 V.)
As concerns the function / ( x , t, u, r), we shall assume that a constant
C > 0 exists such that
||/(.x, t2, U2, ra) - f{x, ti. Hi, ?-i)|| ^
^ C{\t2 - fi| + |i"2 - "ill + Ih - r.W) (18.9)
and that, moreover,
tel = lO,T], ue V, r e L2(G) =>fe L2(G). (18.10)
^z, + i z , = i z , _ i + / , _ ! , (18.11)
h h
BiZj = 0 on r, i=l,...,ti, (18.12)
CiZj = 0 on r , i = 1, ..., /c - ,« , (18.13)
to be solved successively for j — 1,..., p, with
Zo(x) = 0 (18.14)
360 Chapter 18
- Zo + /o 6 L2(G) .
h
The form ((r, M)) being bounded and F-elliptic, the same holds for the form
so that
y =1 + A 6 L^iG) ,
It
and there exists precisely one 22^ V satisfying (18.19). At the same time,
^^2 = / i ( / o + / i ) e L , ( G ) ,
etc.
Thus, it is possible to construct the (abstract) functions Ui(t), Ui{t) and
ri(r) defined in the interval
/ = [0, T] (18.23)
by
Z, =. i i ^ ' , (18,25)
h
by
u,{t) = Zj_,+{t~tj^,)Zj (18.26)
for 1 elj = [fj_i, fj], where tj — jb;
zi for t = 0
Slit) = ( (18.27)
^zj for telj=--{tj,utj],
Si for f = 0,
^'i(0 = { . (18-28)
Sj- for telj,
j = 1, • • . , p-
z"jeV (18.29)
Since these problems are uniquely solvable again, the abstract functions
u„{t), u„{t) and r„{t) can be defined in / similarly as before:
or, denoting
n n
7" — i J-1
K
uli) = 2)_, ^.{i-f._,)Z] (18.35)
Nortel) = \t]_„f^,;
z\ for r = 0
x\(t) = ( (18.36)
^z; for tel] = (t)_„f;\,
for t = 0
(18.37)
for
7 = 1, •.., 2"-V.
To be able to say something about the behaviour of the sequences {M„(t)}.
{t/„(?)} and {/'„(')}> we derive some apriori estimates.
Problem with an Integral Condition 363
Putting V = z" into (18.30), written for j = 1, and taking (18.31) into
account we get
Ikii ^ '^«l|/oll •
Similarly, we obtain, putting v = z^ into (18.30) forj = 2, etc.,
HI ^ \\z"\\ + h„\\f4,
\\zl\\ ^ \\zl\\ + K\\f4 .
we can write
]|z;|| ^ fl + c/,„x\lN"ll + ll^"il). (18-44)
- ^ - . = r i / o i + ^ . (18.46)
llz;i^||/o"|+iVr-/."-ill- (18-58)
By (l8.9), we have the more
7" I!
from where the uniform boundedness of |]Z"[j,
\\Z"A g c,, (18.61)
follows in a way similar to (18.49). From (18.30) and from the F-elhpticity
of the form ({v, u)) we then get, similarly as in the foregoing chapter, the
uniform boundedness of \\z",\\y ,
I-J!!''
(18.62)
It follows (cf. the preceding chapter) that the sequences {M„(0} ^^d
{f„{t)] are bounded in the spaces L2{l, V) and L2(l, L2{G)), respectively,
so that subsequences {u„^{t)] and {r„^{t)] can be chosen such that
u„, - u in L2{I, y) (18.63)
and
?„-/- in L2{I,L2iGy). (18.64)
Moreover, the sequence {u„^{t)] can be chosen in such a way that
u„^^u in C{I,L2{G)). (18.65)
Tn what follows, we show that the pair of functions
u e L2{I, V), re L2{l, LiiG)) (18.66)
is the required solution of our problem (in the sense given in Theorem 18.1).
366 Chapter 18
Because many steps of the proof are similar to those from the preceding
chapter, we shall proceed more rapidly here, without explaining the details
once more.
' ) Without loss of generality, the same indices tif^ can be chosen as in (18.63). The same
holds for selected sequences in the following text.
^) Or in AC{I, L^iG)), if required.
Problem with an Integral Condition 367
/;• for / = 0 ,
m =( (18.74)
^/(.x,/, zj, s^) for r e / ; ,
i = 1, ...,2""V> where
/;=/(x,A,4s;), (18.75)
s) = /i„(/o + . . . + / ; _ , ) , ^0 = 0 , (18.76)
see (18.32), (18.33). Thus, the difference between the functions (18.73),
(18.74) consists only in the second argument (in (18.73) ih„ stands for t
in (18.74)).
The functions (18.73) are simple functions') from / into £2(0). Thus,
they are Bochner integrable in /. An easy computation leads to the result
that the same is true for the functions (18.74)."*) Consequently, we can define
ali)=[)lx)dx'). (18.78)
m-f{x,iK,uXt),fn{t)), (18.79)
/„(/) = f{x, t, u„{t), Ut)) • (18-80)
For r = 0, we have
||a„(0) - r„(0)|| = /i„||/o"i ^ c,b„ . (18.81)
by (18.37), (18.78) and (18.53). Thus, let te{Q, T] be arbitrary. Let it lie
^) Seep. 201.
*) By (18.9), we have, for every t e I,
+ /(X, T, z"_i,s"_i)dT +
<"i-2
+ f{x,ii-l)h„,zUus1-t)] g
+ f' 'i|/(x,T,Z^i,S?_i)-
Jt"i-2
- / ( x , (i - 1) /7„, z ^ i , 5^l)i dr + /i„||/o"|! +
+ f ||/(x, T, z;;, si) - /(X, /;i„, Z?, Sl)\\ dT +
J<"i-l
and denoting
{CT+2c,)r
(18.83)
we have iinally
(18.84)
From the uniform convergence (in L2(G)) of {«„(')} in the interval / (see
(18.65)), the same property follows for the sequence {u„{t)] by a simple
computation.^) Thus,
u^„ ->• 0 for m, n -y CO . (18.86)
Divide the interval I into subintervals of the same length /, choosing /
so small that
C/ < 1 . (18.87)
Consider the first one i.e. the interval [0, / ] . Because the functions r„(f)
are piecewise constant in this interval, the norm \\r„{t) — r^{t)\\ attains
its maximum there. Let this occur at a point /j e [0, Z]. Denote
max \\f„{t) - r,„(0|| = \\f„{h) " Uh)\\ = dV • (18.88)
(e[0,/]
At the point f = /^ we have:
rl' = WUh) - KM ^ \Hh) - <T„(/:)|| +
+ \K{li) - 'Uh)\\ + IHh) - ^4'i)ll • (18.89)
The sum of the first two terms does not exceed fei/2"~^ + kijl'"'^ according
to (18.84). Further,
^) We have
'^i~l+~^^ + C'"^'
Ji) <
1 - CI
= ^ i ( ^ + ^ ) + B,«„ (18.92)
Similarly, let
A = /ci ( - ^ + ~ ) +
+ 2C/u„„ + CI Bi«.
L \2'""^ 2"~^
Problem with an Integral Condition 371
"^^^^^ A / I 1 \
•^
'^"^ - r r ^ - - ^ 4 ^ + _^) + ^^"- • ^''-''^
After a finite number of steps (depending on /, i.e. on C by (18.87)) we
come to the result that
r„„ = niax \\r„it) - r^{t)\\ S A ( ^ ^ + ^ ) + S"- • (18.97)
Now (see (18.86)), the right-hand side of (18.97) tends to zero for m, n -* co
independently of r e / and thus the same holds for r„„. This means that
{r„(f)} is a Cauchy sequence, and thus r„[t) -+ r{t) in L2(G) uniformly with
respect to f e /.
Because of our convention concerning the simplification of the notation,
we have finally the required result:
) We cannot speak about convergence in the space C(I, LjiG)) because the functions
?„ are not continuous in /.
^) As follows from (18.63), (18.64).
372 Chapter 18
where
g{t) = g(x, t) = f{x, t, u{x, t), r{x, t)) . (18.123)
DEFINITION 18.1. The pair of functions u{t) = u{x, t), r{t) = r(x, t)
satisfying (18.117) —(18.122) is called the weak solution of the problem
(18.1)-(18.5).
U n i q u e n e s s : Let
'u{t), V(0, (18.124)
^u{t), ^r{t) (18.125)
be two weak solutions of the given problem; then their difference
Divide the interval / = [O, T] into subintervals of the same lengths d, with
Since the functions u{t), r(t) are continuous (as functions from / into 1.2(G))
on /, the norms |«(0||5 ||K')II attain their maxima on the interval [0, d],
For |jw(')ii. let this maximum be attained at the point t = d^. For v in
Problem with an Integral Condition 375
H ^ Wo, (18.140)
376 Chapter 18
which yields
Mo=0, (18.141)
and
ro = 0 (18.142)
by (18.139). Thus,
u{t) = 0 , r(t) = 0 in [0, d] . (18.143)
Repeating this procedure for the interval [d, 2d] and using the just
obtained result (18.143), we get
u{t) = 0,' r{t) = 0 in [d, 2d] . (18.144)
After a finite number of steps we obtain
u(t)sO, r{t) = 0 in I. (18.145)
In this way, uniqueness of the weak solution is proved. Note that it implies
(cf. Remark 11.1, p. 215) the convergence of the entire sequences {«„(<)},
{r„{t)} in the corresponding spaces (thus, not only the convergence of the
subsequences {u„^{t)], {r„^{t)]).
Let us summarize the results obtained in this chapter into the following
theorem:
THEOREM 18.1. Let the problem (18.1)-(18.5) be given with the form
{{v, u)) satisfying the conditions (18.6), (18.7) of boundedness and V-ellip-
tictiy and with the function f(x, t, u, r) satisfying the conditions (18.9),
(18.10). Then there exists precisely one weak solution of this problem in
the sense of Definition 18.1. The function u(t) is the weak limit in Liif, V)
and the strong limit in C{l, L2(G)) of the Rothe sequence (18.34). The
function r{t) is the strong limit in ^2(0), uniformly with respect to t el, of
the sequence (18.37).
ftix)=f{x,h,z*ix),s*{x)), (18.148)
where
s*{x) = s,{x) = hfoix). (18.149)
Insert z*, f* instead of Zj, / j into the second of the integral identities
(18.19) (thus written for 7 = 2). Then there exists exactly one function
z^eV (18.150)
where (m2 being fixed) the coefficients aji are to be found by the Ritz
method again. Having found them, denote
put z*,/* instead of Zj, / j into the third of the integral identities (18.19),
thus satisfied by a unique function
heV,
378 Chapter IS
I.e.
{{v, Z3)) + i (v, 23 - 4 ) = {v,f*) V. e V. (18.155)
«t(0 = ^ ; - i + i ^ ^ {t - / ; - . ) , (18.157)
h
or
5* for f = 0,
f*=(^ (18.158)
^ s* for telj,
i — 1, ..., p. These functions are analogues of the functions (18.24), (18.28)
defined by
In what follows, we show that the functions u(i) and r(t) from Theorem
18.1 can be approximated to arbitrary accuracy by the function ii*(t) and
r*{t), respectively. More precisely: To every e > 0 is it possible to find
sufficiently large numbers p, m^,..., nip such that the inequalities
With regard to the form of the functions Ui{t), M*(t), ri(r), f\{t) it is even
sufficient to show that such numbers m^,..., nip can be found that the in-
equalities
be satisfied.
Let us choose ?/ > 0 (to be specified later). To this t] an Wi can be found
such that
||r, -z*\\^ri. (18.169)
Similarly, an nij can be found such that
1=2-41^^, (18.170)
where fj e I'is the function satisfying the integral identity (18.151), At the
same time, the function Zj e F satisfies
i z , - z * | ^ ; , + /Ki|A-/f|| +
.-||/2-/.*||+... + !|/,-X-/Ali). 18.178)
j; < 18.185)
2pe^
to ensure (18.167), (18.168), i.e. to ensure (18.165), (18.166), and, taking
account of (18.163), (18.164), also (18.161), (18.162).
u{x, 0) = 0 , (19.2)
— (x, 0) = 0 , (19.3)
382
Hyperbolic Problems. Homogeneous Initial Conditions 383
etc., basic results have been obtained in my seminar ( see the papers [9j
by J. Streiblova and [11] by F. Bubenik). In [9] an existence and convergen-
ce theorem is derived for the problem (19.1) —(19.5) (or, more precisely,
for a slightly more special problem) and an interesting error estimate
obtained. However, this estimate is unefficient for "small" t. The reason
of this inefficiency lies in an improper choice of the "initial function"
z_i (see Remark 19.2, p. 385), thus in an inappropriate difference formula-
tion of the homogeneous initial conditions. Therefore, in [11] a slightly
different approach is chosen. Moreover, unhomogeneous initial conditions
and convergence of the "Ritz-Rothe" method are investigated there.
In this chapter, as well as in the next one, many ideas of the just quoted
works are utilized although the approach chosen here is rather different
in some directions.
II^IIF (19.12)
equivalent with the norm (19.10). Let the new space, the elements of which
are the elements of the space Fand the scalar product on which is defined by
{v,u)y = {{v,u)), (19.13)
be denoted by V. In consequence of the equivalence of the norms (19.10),
(19.12), a sequence {v„] is convergent (or is a Cauchy sequence) in Fprecisely
if it is convergent (or is a Cauchy sequence) in V. This fact will often be used
in what follows without reminding it again.
In particular, V is complete, thus it is itself a Hilbert space.
Further,
feL,{G) (19.14)
is assumed.
REMARK 19.1. As concerns the case / = /(x, t), it does not represent
particular difficulties and is not considered here. Under some assumptions
concerning Lipschitz properties of / ( x , t) with respect to /, the whole
procedure described below can be applied. The corresponding calculations
are rather lengthy.
into p subintervals
^j = bj-utj'], ;• = l , . . . , p , (19.16)
of lengths h = Tjp (tj = jh) and replace the second derivative with respect
Hyperbolic Problems. Homogeneous Initial Conditions 385
REMARK 19.2. The choice of z_i is very important for obtaining the
"right" order of convergence. It would seem natural to choose
z_i(x) = 0 (19.25)
to express the condition (19.3) in an appropriate difference form. This is
reahzed in [9] and leads to the same concept of a weak solution of the given
problem (thus with the same properties) as when choosing z_i(x) by
^) By (19.4).
386 Chapter 19
(19.24)). However, the order of convergence is lower for small t (see the
quoted work [9]) than the order obtained in the second section of this
chapter (p. 403), where (19.24) is appHed.^)
Each of the problems (19.20), (19.22) (with (19.23), (19.24)) is uniquely
solvable: Since the form ((t>, w)) is bounded and F-elliptic so is the form
(((.,«))) = ( ( . , u ) ) + ! ( . , « ) (19.26)
h
for every fixed /i > 0 (cf. Lemma 3.1, p. 78). Take / = 1 in (19.22). We have
feL2{G), so that
h
and there exists exactly one z^eV satisfying (19.22), etc. Thus, we can
construct similar functions as in the preceding chapters (considered as
abstract functions from the interval / into For L2(G)):
Z. = ?i ^J^ , (19.28)
h
«i(') = 2,-1 + Zj. (t - tj_,) in IJ ; (19.29)
Zi for t = 0,
"i(0 = ( ^ (19.30)
or, denoting
(19.32)
h
17,(0 = Zj_, + sj. (t - lj_,) in /,. ; (19.33)
Zi for t = 0,
(7,(0 = ( (19.34)
^Z, in /,,
5, for t = 0,
Y,{t) = ( (19.35)
Sj m Ij,
j = l,...,p.
as above, we can define, in the interval / = [0, T], the abstract functions
;• = i,...,2"-V
To be able to describe the behaviour of the "Rothe sequence" («„(()} and
discuss the other sequences defined above we have to derive some apriori
estimates. These estimates will be derived for the division d^ with step h
and then easily generalized for the division i/„.
^) When doing this, we can assume, if we wish, that the entire process does not start
with the initial functions ZQ{X), Z_^{X), but that it begins with the functions z_^{x),
z_2(x), which uniquely determine the function ZQ(X) = 0 from (19.49). This will
appear convenient below when deriving the estimates for \\sj\\ y and |15^||.
Hyperbolic Problems. Homogeneous Initial Conditions 389
Z-2=2/i„7, (19.52)
or
S;llz,li!/|!- ("•««)
Taking into account that
h
and that (19.65) holds, we get, finally (multiplying (19.66) by two),
2\\z4^ + \\s,r^iifr + ifr. (19.6?)
Hyperbolic Problems. Homogeneous Initial Conditions 391
In the following text, apriori estimates for the norms of higher difference
quotients will be derived, although they will be needed only in Section b)
when deriving an error estimate. In contrast to the preceding text, feV
is assumed here.
' h
one obtains
{{sj - Sj_uSj)) + iSj,Sj -- s , - i ) = o (19.70)
In quite a similar manner as in which (19.59) has been obtained from
(19.56), the equahty (19.70) yields
and, finally,
||5,||^+|lS,r^i5,|!^+|S,f, j = 2,...,p'). (19.72)
Let us try to estimate the right-hand side of (19.72) with the aid of ||/||^
only. We have, by (19.23), (19.24), (19.48),
-0=0, z_i = Ih'f, z_2 = 2hV, (19.73)
^) Here we use the "additional" integral identity (19.49), see Footnote 3 on p. 388.
392 Chapter 19
so that
So = =^^^~^--^ = f, (19.74)
si - / = si - So = hS, (19.75)
by (19.51), and
/i
we get, by (19.81),
((^1 - if, s, - if)) ^ {{if, si - 1/)),
from where
Mv ^.V(f)l|/|lr, (19.89)
ll^'"i ^MWfWv (19.90)
From (19.87) and (19.23) we then get
M\v ^ K{\\Z"Ay + ... + \\Z1\\y) g ^ Tii/ll . (19.91)
Note that the inequalities (19.87), (19.88), (19.91) have been obtained
under the assumption/e L2(G) only.
In consequence of the K-eliipticity of the form {(v, u)) and of the inequali-
ties \\z"j\\ ^ Wz^Wy, ||Z;;|| ^ |]Zj|j^-, (19.91) and (19.87) imply uniform boun-
dedness-(with respect to both n and j) of |z"|| and ||Z"||.
Having obtained these results, all what follows is similar to the para-
bolic case (see the preceding chapters):
Consider the functions u„, U„ and Y„ given by (19.44), (19.46) and (19.47),
p. 388, the first two as abstract functions from / into V (or V), the third as
an abstract function from I into L2(G). The estimates (19.91), (19.87), (19.88)
394 Chapter 19
Thus,
M 6>1C(/, F ) , (19.102)
M' 6 L2(/, F) n ^C(7, L2(G)) , (19.103)
«" e L2(/, L2(G)) , (19.104)
u(0) = 0 in C{I,V), (19.105)
u'(0) = 0 in C(/, L2(G)). (19.106)
Hyperbolic Problems. Homogeneous Initial Conditions 395
In the preceding text, the existence of such a solution has been proved.
has the properties (19.102) —(19.106) and satisfies the integral identity
I 0
{{u{t), u'{t)y) dt + \u'{t), u"{t))dt = i\\u{a)\\^ - i-i|u(0)||f +
+ i||u'(a)l|-ii"'(0)i.
0
(19.114)
(19.113), (19.114) and (19.105), (19.106) yield
from where u{a) = 0 in K Since the function u{t) belongs to C(/, V) and
a e I is arbitrary, we have
w - 0 in C{I, V) (19.116)
which was to be proved.
REMARK 19.3. Uniqueness of the weak solution implies, in the usual way,
that Hj can be replaced by n in (19.95) —(19.97), i.e. that not only the sub-
sequences {jJ„J, {t7„J, {y„J but the entire sequences {«„}, {L/„}, {Y„}
converge weakly to the corresponding functions w, U, Y in L2{l, V) or
£2(7, L2{G)). The same holds for the subsequences (u„,} and {U„^}.
Now, from (19.37) and from the definition of the functions u„{t) and Y„(t),
it follows that
{{v{t), u„{t))) + m y„(0) = (K0> / ) (19.122)
holds for every v e L2{l, V) and for almost all t e I. Similarly, we have
{(v{t), uM) + {<t), Y„,{t) = {v{t),f). (19.123)
Subtracting and putting
v=U„-U„, (19.124)
we get
((«„ - M„„ V„ - U„)) + {U„ - [/„, 7„ - y„,) = 0 . (19.125)
Now,
tin - "m = "n- "m + ("« " "») + ("m " "m)
and ^
2 at
where at the points of division of the interval / one-sided derivatives are
to be understood, and similarly for the functions U„ — U^. Thus, we get,
by (19.125),
2 at 2 at
^ Ky{\\u„ - u„\\r + |u„ - ii^y) + K,{\\U„ - U„l +
+ ||L^„- U,„|)^). (19.126)
*) It follows immediately from the uniform boundedness of ||Z"|]j' and \\s"j\\.
' ) In consequence of the uniform boundedness of the norms
\\u„-ujL and ||y„-y„l|.
398 Chapter 19
It means that the sequence {u„(t)}, or {U„(t)} is a Cauchy sequence in C(/, V),
or C(i, 1.2(0)), respectively, from where (19.118) and (19.119) follow
immediately.
Moreover, when passing to the hmit for m -f 00 estimates of the form
can be obtained, which although inefficient from the practical point of view,
are of theoretical importance.
be satisfied. Let
f€L,(G). (19.132)
Then there exists exactly one weak solution of this problem in the sense
of Definition 19.1, i.e. exactly one abstract function u(t) with the properties
(19.102)-(19.106), (19.108). This function is a (strong) limit, in C{L V), of
the sequence of Rothe functions (19.43). Moreover, U„ -> u' in C(/, L2(C)),
where {U„[t)} is the sequence (19.45).
Hyperbolic Problems. Homogeneous Initial Conditions 399
b) ERROR ESTIMATE
(19.136)
Denote
p. = Pj~Pj-^ ^ (19.138)
h
) Thus, Pj means the difference between the first and second Rothe function at the
common point t = tj — t\j.
400 Chapter 19
In particular,
Po = 0 (19.140)
since zj = 0, ZQ = 0 and
n = 0 , (19.141)
because
Vt>6K. (19.142)
Let us note that
v2
Thus,
{{v, pj)) + (v, cjj) =-- -~ (v, 3Slj + Sl._,) Vr 6 F . (19.143)
o
^) The choice of the "initial function" z_i in [9] does not lead to (19.141). This is just
the reason why the error estimate in [9] is not efficient for small values of t.
Hyperbolic Problems. Homogeneous Initial Conditions 401
Insert
z
\\Pjr^h\'ii)\\f\\,Y.\\P,\\^.ch^Y^\\P,\\
1=1
(19.151)
with
c = ,/(f) ||/|!, . (19.152)
Denote
QL,-Q..,-''^^~^^0. (19.156)
has a positive discriminant, thus it has two distinct real roots (Qt+Ji <
< {Qi,+i)2- For the larger root we have
.,, ^ _ 1 + V [ l + 2fc(/c + 1)] , 1 + v'(4fc^ + 4k + l]
(Q/..i)2 - 2 - 2 -
= L ± ^ i = , + l. (19.157)
Mtl,)-u,{t))\\^ch^''^^- (19.160)
Hyperbolic Problems. Homogeneous Initial Conditions 403,
Similarly, we derive
etc., and that t^j = 'Ij = ijj = t] = tj according to the usual notation,
we can write
This "parabolic" type of approximation is very natural for the given hyper-
bolic equation with homogeneous initial conditions and makes it possible
to obtain a quite satisfactory error estimate (19.164) which is much better
than that derived in [9] when choosing
z-i = 0 . (19.169)
However, this approach is the source of the following difficulty: In the
quoted work by Bubenik, the hyperbolic problem with n o n h o m o g e n e o u s
initial conditions
Thus, let the problem (19.1)-(19.5) be given with / e L2(G) and with
the form ((r, u)) bounded in Wf\G), F-elliptic and F-symmetric. The method
of discretization in time leads to the integral identities (19.21),
Hyperbolic Problems. Homogeneous Initial Conditions 405
Z_i = Zi (19.175)
here. In particular, (19.174), (19.175) and (19.173) yield, for; = 1,
According to the given assumptions there exists exactly one z^e V satis-
fying (19.176). Then there exists exactly one Zj e F satisfying (19.173)
for; ~ 2, etc. Thus, it is possible to construct the functions (19.29), (19.30),
(19.31), (19.34), (19.35), pp. 386, 387, again, or, for the division d„, the
functions (19.43)-(19.47), p. 388, where z"e F satisfy the integral identities
Apriori estimates (19.68), p. 391, can be easily obtained for the choice
(19.175):
We have
Zi = ^ , Zo = - ^ , (19.180)
h h
^°) Note that the functions (19.45), defined in / by
L/„(/) = z ; _ i + 5 : ; ( t - t ; _ o in /; = K - i , ( ; ] ,
belong to L2(I, V) now, which was not the case when z_j was chosen by (19.167)
since then we had ZQ ^ Vii f$ V.
406 Chapter 19
from where
||Zo|| = ||Zi||. (19.181)
Inserting v =• z^ into (19.176), we get
2hM^h'\\f\\, (19.182)
thus
||ZoI| = | i Z i | | ^ i - / ' i / | l . (19.183)
Now, (19.60), p. 390, remains valid. The same holds on (19.66), p. 390, which
yields
As concerns the error estimate (19.164), p. 403, the same apriori estimates
as in (19.85), (19.86), p. 393, are obtained in the case of (19.175). In fact, let
feV. (19.187)
Consider the integral identities (19.50), p. 388. Let us write the first of these
in the form
from where
hi-^h'f\\rSih^f\\y. (19.190)
(19.189) then yields
l\\,^-lh'f\\^^lh^flyJ,h^f\\y,
or
Iz, - ^h'fl ^^ h^f\\y . (19.191)
2 V2
In order that (19.49), p. 388, be true (to ensure (19.72), p. 391), we have
to define
z_2=2zi+/iV- (19.192)
408 Chapter 19
Then
_ Zi - 3zo + 3z_i - z_2 _ 2zi - h^f
and by (19.191)
\\s4^-^A\f\\v- (19.193)
p ^ Po - P - i ^ _ P ^ ^ _ z i z - z - i ^
° /j h h
•1 ^^1 ^ /
4
l(2.;-„v)-l(2z;-^/)
/i 2/j
Hyperbolic Problems. Homogeneous Initial Conditions 409
(20.3)
) Thus with/== 0. If both the equation and the initial conditions are norhomogenecus,
then the problem (being linear) can be "decomposed" into a problem with homogeneous
initial conditions and with/=H 0 (considered in the preceding chapter) and a problem
with nonhomogeneous initial conditions and / = 0 (considered here). See also Remark
20.5, p. 424.
410
Hyperbolic Problems. Nonhomogeneous Initial Conditions 411
Under the conditions imposed on the form {{v, u)) and on the functions
UQ, t/o' the existence of exactly one function z\eV satisfying (20.12) is
ensured. Having found it, precisely one z^e V exists satisfying (20.9)
for J = 2, etc. Consequently, the famihar functions u„{t),..., Y„{t) (cf. p. 388)
can be defined in / = [0, T] by
^1 for t = 0,
ut)
nv }
/
\
(20.17)
^; in I"
412 Chapter 20
REMARK 20.1. It turns out that the assumptions (20.6), (20.7) are very
natural in the hyperbolic case. However, under these only assumptions,
certain difficulties arise when deriving apriori estimates — cf. similar
problems encountered in Chap. 13. Therefore, the case when the considered
functions are smoother is treated first — in more detail, if they belong to the
so-called set M (see p. 209). This assumption leads to the concept of a weak
solution of the problem (20.1)-(20.5). In Section b), (20.6), (20.7) is assumed
only, and the so-called very weak solution of this problem is obtained.
REMARK 20.2. If the data of the problem are sufficiently smooth, the
problem can be converted, by the substitution
u = r + Uo + tVo, (20.20)
into the problem with homogeneous initial conditions. This idea will be
followed, in a certain sense, in Section a), although not in such a "classical"
way.
In what follows, we use the result from the book [2] by J. Necas, pp. 131,
132 (or, rather, a special case of it), applied in Chap. 13, p. 251, already:
The form ((v, u)) being F-elliptic, a set M dense in Fexists such that for every
se M a unique h e L2{G) can be found that
(20.29)
;• = 1,..., 2 " " ^ . Moreover,
ro - 0 , (20.30)
r"-! = r\. (20.31)
-^ = r5 + Mo + /)„t/o ,
or
z"_i = r"_, + „ „ - / , „ L / o ,
t
414 Chapter 20
Thus, by (20.26), the problems in z" are converted into problems in r"
which correspond to the case of nonhomogeneous differential equation
with homogeneous initial conditions.
However, the right-hand side of this equation is not independent of t.
Thus, let us "decompose" the just obtained problem into two: Put
r) =-• f] + f] , (20.32)
where the functions
r"j e V (20.33)
satisfy the integral identities
(20.48)
J I. '
?" - ?"
(20.49)
K
we can rewrite, first, the integral identities (20.38) in the form
{{v, ;>;)) + {v, s]) = -}h„{v, G) yv e V. (20.50)
^) At the same time, the corresponding sequence (20.15) (with z" replaced by }")
converges, in C(I, L2(G)), to the function ?'(/), see Theorem 19.2, p. 406. We are not
going to recall this fact in what follows, formulating the final result in Theorem 20.1.
•*) These problems correspond to the problem (19.1)—(19.5) with the right-hand side
f(x) replaced by ~tG(x). This problem can be treated as well as a special case of the
problem (19.1)—(19.5) with/=/(A:, /), the solution of which was suggested in Remark
19.1. However, the direct way of its solution is much more simple.
416 Chapter 20
from where
PTv+lS-jr^Pj-dr+lSj-.r- (20.54)
If we define
r_2 = 2f"_i = 2r"i , (20.55)
so as to extend the integral identities (20.38) to the case; = 0, the inequali-
ties (20.54) can be written for all j = 2, ...,2"~^p, giving finally
l l ^ J l l F + i ^ J i ^ K l i . + ilS^i. (20.56)
However, by (20.39), (20.40), (20.55),
•fn
^1 —
' 1
' Z"o = n (20.57)
/'„ h,.
So = 0 . 1 (20.58)
2f\
(20.59)
h„ hi •
Hyperbolic Problems. Nonhomogeneous Initial Conditions 417
Choosing
^ = hi
-i
we then get
ll^;i^ ^ 2||G,f.
The same holds on uniform boundedness of s" in Lji^)- Writting
Having obtained these results, all the considerations carried out in the
preceding chapter can be reproduced. In particular, the sequence of Rothe
418 Chapter 20
coincides with the function WQ + tUo for all / e [0, T]. Thus, the sequence
of Rothe functions (20.13) converges in C(l, V) ' ) to the function
ueAC{l,V), (20.69)
u' e L^iL V) n AC{I, £2(6)) (20.70)
u" e £2(7, L2(G)), (20.71)
u(0) = Mo in C(/, F ) , (20.72)
M'(0) = b'o in C(/, L2(G)) (20.73)
In the above text, we have proved the existence of such a weak solution.
Uniqueness follows in quite a similar manner as in the preceding chapter.
We thus have:
THEOREM 20.1. Let the problem (20.1)-(20.5) be given. Let the bilinear
form ({v, u)) corresponding to the operator A and to the boundary operators
Bj, Cj (p. 383) be bounded in W2''(G), V-elliptic and V-symmetric (see
(19.7), (19.8), (19.11), pp. 383, 384), with Vgiven by (l9.9). Let MQ, L'O satisfy
(20.22). Then there exists exactly one weak solution of the given problem
in the sense of Definition 20.1, i.e. exactly one function u(f) satisfying
(20.69)-(20.74). This function is the (strong) limit, in C{l, V), of the se-
quence of Rothe functions (20.13), ii'(t) is the (strong) limit, in C(l, £2(0)),
of the sequence (20.15).
d^u
+ Au = f in e = G X (0, r ) , (20.75)
f
Jo
{{v{t), u{t))) dt + {v{t), u"{t)) dt = iv{t),f{t))dt
The problems (20.8) —(20.11) are replaced by the problems to find such
functions
z^eV, ; = 1,...,2n - l , (20.81)
which satisfy
2h„Uo . (20.84)
Hyperbolic Problems. Nonhomogeneous Initial Conditions 421
The weak solution u{t), or its derivative w'(t), is the (strong) hmit in C[l, V),
or in C(/, L2(G)), of the corresponding sequences (20.13), or (20.15), re-
spectively.
\\z"jUS:\\uorv + 2\\Uoj\
and the more
^r ^ \\uo\\y + ifUoW .
! • ' "
(20.91)
The Rothe function (20.13) being piecewise linear in t and assuming the
"values" z" at the points of division, we have
l!«„(Oik ^ i«o||7 + 2||(7o|| Wtel. (20.92)
p^ -* uo in V, P„ - Uo in L^iG). (20.95)
Hyperbolic Problems. Nonhomogeneous Initial Conditions 423
Denote by '"u(f) the weak solution of the problem (20.1)-(20.5) with the
initial conditions given by the functions p,„, P„„ In consequence of (20.95)
and of Theorem 20.2, {'"u[t)} is a Cauchy sequence in C(/, V). Thus, a func-
tion u e C(/, V) exists such that
lim "« = u in C{l, V) . (20.96)
From the just presented construction it follows that this function does not
depend on the choice of the functions p„, P„„ satisfying (20.95). Thus,
we are entitled to define:
We thus have:
THEOREM 20.3. Let the form [(v, «)) satisfy the assumptions of Theorem
20.1, let KQ e V, UQE L2{G). Then there exists exactly one very weak
solution of the problem (20.1) —(20.5).
In fact, let e > 0 be given. We have to show that WQ can be found such that
However,
||M(0 - uXt)\\v S. Ht) - -"uit^y + W"u{t) - '"«„(Oir +
+ f'ult) - u,it)\\y , (20.103)
Then
||u(f)-"•«(?)||p < - Vie/ (20.104)
REMARK 20.5. If both the initial conditions (20.2), (20.3) and the equa-
tion (20.1) are nonhomogeneous, Theorem 20.5 remains vahd, with the
only difference that integral identities (20.9) are replaced by the integral
Hyperbolic Problems. Nonhomogeneous Initial Conditions 425
identities
z*^tauv,, (20.111)
1=1
where 5j is a positive integer (see below). Using the Ritz method, the coeffi-
cients flj; in (20.111) are uniquely determined, and so is z*.
426 Chapter 20
Consider now the second of these problems, replacing Zj there by the just
obtained function z*. Assume the approximate solution in the form
4 = E«2it;,., (20.112)
1=1
find the coefficients AJ; by the Ritz method again and consider the third
of the problems (20.108), (20.109) with z* and z | substituted for r , and Zj,
etc. This process yields the functions
z*{x),...,z;(x). (20.113)
In such a case we say, briefly, that the Ritz-Rothe method applied to our
problem is convergent.
if Sj..... Sp are sufficiently large. Moreover, since the functions Mi(f) and
u*(() are of the form (20.114), (20.115), thus piecewise Hnear in t, it is
sufficient to prove that it is possible to satisfy
and replace z^ by the just obtained function z*. Denote by Z2 e Fthe func-
tion satisfying the thus obtained integral identity:
be valid for the Ritz approximation (20.112) of the function Zj. Then, by
(20.125), we have
|]z2 - z t W v ^ ^ l . (20.127)
Similarly, replace the integral identity
by the identity
Now, h being fixed (in order that (20.117) be fulfilled), p is fixed. After p
steps we come to the number c^ which is the largest of the numbers Cj.
Thus, it is sufficient to choose r] so small that
t] <— (20.136)
In accordance with what has been said above, we can express this result
briefly in the following form:
Concluding Remarks
|3) 1/ £ L^{I, V), u' e L^{I, L2{G)). First, let us recall that L^{I, H) is
the space of strongly intergrable abstract functions u from / into a normed
space H, with i|u(()||H bounded almost everywhere in / and with the norm
defined by
Y) I/„ -• u in C{l, V). Let us remind that, in the interval /, the function
u„[t) is defined by
.." _"
u„{t)^z';.,+'-^j^^it-fj^,) for tel-} = [f}.,,fpi.
K
Define, similarly,
respectively. Let us choose an arbitrary, but fixed t e I", multiply the first,
or second of these identities by the number (t — r"_ i)'/;,,, or 1 —(( — ?"_ i) :
h„, respectively, and make the sum. We obtain
Because the interval /" has been chosen arbitrarily, the obtained integral
identity is valid for all te I. Similarly, for the m-th division of the interval
/ the integral identity
valid for all tel, is obtained. Subtract this identity from that for the «-th
division, choose an arbitrary, but fixed t e / and put
V = u„{t) - Mm(0 •
We get
((u„(0 - uM <t) - "™(0)) + ("-.(0 - uM ^n{t) - v„it)) = 0.
Now
M\ ^ \\f\\
for all n a n d ; (see (11.79), p. 197), so that
— - c > 0, (21.5)
''0
and consider only such divisions of the interval / for which h ^ /IQ holds.
Then also
^ - c > 0 (21.6)
h
is valid.
Put u = Z, into (21.2) and divide by h. One gets
((Z„Z,)) + y ( Z „ Z , ) = 7 ( Z i , / ) ,
h h
) A typical case is that of the operator A — —S. with the Neumann boundary condi-
tion cuitv = 0 on r . The form (v, u)) itself is not K-elliptic in this case, while (21.1) is
fulfilled.
434 Chapter 2!
or
((Zi, Z,)) + c{Z,. Zi) + ([-') (2i, Z.) = ^ ( Z i , / ) , (21.7)
||Z,|| g J ^ (21.8)
1 — he
is obtained in the familiar way.
Subtracting (21.2) from (21.3), dividing by h and putting t- — Zj, we
obtain, similarly,
|jZ,||<J^<-^L, (21.9)
i — he (1 — he)
and, in general,
I N S ^ , . (2,.,0)
However,
1 <__1_ = 1 < (21.11)
(1 - he)' (1 - /ic)" / _ eTV
for all p sufficiently large (i.e. for all sufficiently fine divisions t/„ of the
interval /). In this way, uniform boundedness of ||Z"|j by the constant
2j|/j| e"^ follows for all such divisions.
This basic estimate being obtained, the whole further procedure (the
proof of uniform boundedness of ||2"|j as well as of |iz"!|F with the aid
of (21.1), etc.) remains unchanged.
l|z.NII/('Oi-
Subtracting (21.13) from (21.14), putting v = Zj and using (21.12), we obtain
| | Z , | | ^ | | Z , | | - t - | | / ( 2 / , ) - / ( / i ) | | ^\\fih)\\ +Ch.
In general, we have
||Z,|| g |I/(/,)|| + (j -l)Ch£M +TC, (21.16)
where
M - max ||/(0|! . (21.17)
tel
lkiNHI/WII>
|z2|i^||z,||+/,|]/(2/,)|g/,(||/(/0|| + 1/(2/011),
etc., and, in general,
= Irlkic/.K) i|/iiL2a,L2(G))
giving
li"!U.a.K,^-i/|k2(/,i.(G,,. (21.19)
a
This inequahty represents another form of the continuous dependence
of u on / , permitting to extend the concept of the solution to the case
o f / e L2(/, 1,(0)).
Such a function exists and is unique. Moreover, the form ((r, u)) being
F-elliptic, a constant c (independent of f e /) exists such that
Zj^Zj-TU*{jh), ./ = 1 , . . . , / ) . (21.36)
In this way, the problem (21.27) —(21.30) turns into that of finding such
functions
ZjeV, .7 = 1 , . . . , P , (21.37)
zo=-u*(0). (21.39)
VreF, (21.41)
with
fo = 0 . (21.42)
The first of these problems was treated in Chapter 13, Section a). Thus,
it is sufficient to investigate the problem (21.40)—(21.42). Denoting
2. = v ^ i i z i , j = i,...,p,
h
and inserting v = 2^ into the first of the integral identities (21.41), we get,
with regard to (21.34),
11:^1 I ^ c K , .
Subtracting then the first of these identities from the second one and putting
V = ^2, one obtains, using (21.35),
ll^jjl ^ {|Zi|| + cK^h ^ cXi + cKih .
In general, we get
\lj\\ ^ cK, + (j - 1) cK^h ^ c{Ki + K^T) . (21.43)
This inequality represents the fundamental estimate for the usual approach
to proving an existence and convergence theorem.
On the other hand, inserting, successively, v = Zj into the j-th. of the
integral identities (21.41), we get, using (21.34) only,
||zi|| ScK.h,
m S l|fx|| +cK,h^2cK,h,
(21.44)
II2,.II ^jcK,h^cK,T.
440 Chapter 21
y I jjM^S (21.46)
(see (2.266), p. 62; hieL2{r)) does not vanish. When considering the
problem
— + Au = 0 in e = G X (0, r ) , (21.47)
dt
u{x, 0) = 0 , (21.48)
Even in this very simple case, the usual approach to obtaining apriori
(21.59)
g cv (21.64)
dv'' i2(r)
holds for all ve V and for all i = I,..., k — n. (For the details see [3],
Chapter 33.) Finally, the well-known inequality
^44vl^'ih^in^'jHy^YJN\lin•
Consequently, such positive constants rji, r\2 can be found, with
holds.
Concluding Remarks 443
Let us assume, first, that the functions /?,• depend on t and satisfy the
following assumptions:
hi{0) - 0 , (21.67)
by (21.65),
i=X 1=1
i{lz4' + lZ,~Z,r~\\Z,r)^rj,C^,
from where
\\Z2f ^ | | Z i f +2;?2C'/J- (21.71)
Similarly, we obtain
IIZ3P g IJZall" + 2)72C2/T, (21.72)
which represents the basic estimate needed for the proof of existence and
convergence.
\\zjf ^|iz,_,|P+2^,H/,
which yield
\zjf S 2t]2HT. (21.75)
This inequality leads, in the usual way, to the conclusion on the continuous
dependence of the solution on H, i.e. on the given boundary functions
(in the L2(r) norm). This fact makes it possible to generalize assumptions
(21.67), (21.68), in particular to take into considerations such functions
which do not fulfil the condition (21.67).
There are too many works related to the problems of evolution equations,
more than can be presented here. Therefore, only such papers are contained
in this list which are quoted in the text, or are in close connection with
questions discussed in this book.
[1] Rektorys, K.: On AppHcation of Direct Variational Methods to the
Solution of Parabolic Boundary Value Problems of Arbitrary Order
in the Space Variables. Czech. Math. J. 21 (1971), pp. 318-339.
[2] Necas, J.: Les methodes directes en theorie des equations elliptiques.
Prague, Academia 1967.
[3] Rektorys, K.: Variational Methods in Mathematics, Science and
Engineering. 2-nd Ed. Dordrecht — Boston, Reidel 1979.
[4] Rektorys, K.: Survey of Applicable Mathematics. London, Iliffe 1969.
(2nd Ed.: Dordrecht-Boston-London, Reidel; in preparation)
[5] Fucik, S. - Kufner, A.: Nonlinear Differential Equations. Amsterdam
- O x f o r d - N e w York, Elsevier 1980.
[6] Kufner, A. - John, O. - Fucik, S.: Function Spaces. Prague, Academia
1977.
[7] Rektorys, K. - Ludvikova, M.: A Note on Nonhomogeneous Initial
and Boundary Conditions in Parabohc Problems Solved by the Rothe
Method. Aphkace matematiky 25 (1980), pp. 5 6 - 7 2 .
[8] Valesova, M.: Solution of the Equation dujdt + A[t) u = f{t) by the
Rothe Method. Habilitation. Bull, of the Faculty of Civil Engineering
in Prague. (To appear.)
[9] Streiblova, J.: Solution of Hyperbohc Problems by the Rothe Method.
Habilitation. Bull, of the Faculty of Civil Engineering in Prague
(To appear.)
446
References 447
[23] Ibragimov, S. I.: On the Analogy of the Method of Lines for Differen-
tial Equations in Abstract Spaces. Dokl. Azerb. Ak. Nauk XXI (1965).
(In Russian.)
[24] Dubinskij, J. A.: Quasilinear Elliptic and Parabolic Equations of
Arbitrary Order. Usp. Mat. Nauk 25 (1968), pp. 4 5 - 9 0 . (In Russian.)
[25] Mosolov, P. P.: Variational Methods in Nonstationary Problems.
Izv. Akad. Nauk 34 (1970), pp. 425-457. (In Russian.)
[26] Browder, F. E.: Existence Theorems for Nonlinear Partial Differential
Equations. Proceedings Symp. Pure Math. 16, pp. 1 — 60. Amer.
Math. Soc. 1970.
[27] Douglas, J. - Dupont, T.: Galerkin Methods for Parabohc Equations.
SIAM J. Num. Anal. 7 (1970).
[28] Raviart, P. A.: Sur I'approximation de certaines equations d'evolution
I., II. J. math, pure et appl. 46(1967), pp. 11-107, 108-183.
[29] Rothe, E.: Zweidimensionale parabolische Randwertaufgaben als
Grenzfall eindimensionaler Randwertaufgaben. Math. Ann. 102 (1930).
[30] Jerome, J.: The Method of Lines. J. Diff. Eqs. 30 (1978), No. 1.
[31] Lions, J. L. - Magenes, E.: Problemes aux limites non homogenes
et applications. Paris, Dunod 1968.
[32] Vejvoda, O. et al.: Partial Differential Equations. Alphen—Rockvill,
Noordhoff 1981.
Subject Index
449
450 Subject Index
K A R E L REKTORYS
Technical University, Prague
Audience
The book will be of interest to mathematicians (pure as well as applied),
scientists, and engineers.