Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
9JT
An Introduction to LINEAR ANALYSIS
DONALD L, KREIDER, Dartmouth College
ROBERT G. KULLER, Wayne State University
Introduction to
LINEAR
ANALYSIS
Copyright © 1966
Philippines Copyright 1966
For the student. Tradition dictates that textbooks should open with a few remarks
by the author in which he explains what his particular book is all about. This obli-
gation confronts the technical writer with something of a dilemma, since it is safe
to assume that the student is unfamiliar with the subject at hand; otherwise he
would hardly be reading it in the first place. Thus any serious attempt to describe
the content of a mathematics text is sure to be lost on the beginner until after he
has read the book, by which time, hopefully, he has discovered it for himself.
Still, there are a few remarks which can be addressed to the student before he
begins the task of learning a mathematical discipline. Above all, he should be told
what is expected of him in the way of prior knowledge, and just what special de-
mands, if any, are going to be made of him as he proceeds. In the present case the
first of these points is easily settled: We assume only a knowledge of elementary
calculus and analytic geometry such as is usually gained in a standard three-se-
mester course on these subjects. In particular, the reader should have encountered
the notion of an infinite series, and know how to take a partial derivative and eval-
uate a double integral. Actually almost two-thirds of this book can be read without
a knowledge of these last two items, while the first is covered quickly, but (for our
purposes) adequately, in Appendix I. we have
In short, kept formal prerequisites
to a minimum. At the same time, however, we demand
that the reader possess a
certain amount of that elusive quality called mathematical maturity, by which we
mean the patience to follow mathematical thought whither it may lead, and a will-
ingness to postpone concrete applications until enough mathematics has been done
to treat them properly.
This demand is a reflection of the fact that initially much of our work may
last
seem rather abstract, especially to the student coming directly from calculus. Thus,
even though we have made every effort to motivate our arguments by referring to
familiar situations and have illustrated them with numerous examples, it may not
be out of place to reassure those students not interested in mathematics per se that
every one of the topics we discuss is of fundamental importance in applied mathe-
matics, physics, and engineering. Indeed, without falsifying fact, this book could
have been entitled '*An introduction to Applied Mathematics" or "Advanced
Engineering Mathematics," and might have been save that the materia! covered is
of real value for the student of "pure" mathematics too. Nevertheless, most of
the ideas which we have treated grew out of problems encountered in classical
vi PREFACE
physics and ihe mathematical analysis of physical systems. As such these ideas lie
at the foundations of modern physics and, to a lesser extent, modern mathe-
matics as well.
But even more important, the subject which we have chosen to call "Linear
Analysis" is, when viewed as an entity, one of the most profound creations of the
human mind, and numbers among its contributors a clear majority of the great
mathematicians and physicists of the past three centuries. For this reason alone
it isworthy of study, and as our discussion proceeds we can only hope that the stu-
dent will come to appreciate the beauty and power of the mathematical ideas it
exploits and the remarkable creativity of those who invented them. If he does, his
efforts and ours will have been well rewarded.
For the instructor. As its title suggests, this book is an introduction to those
branches of mathematics based on the notion of linearity. The subject matter
in these fields is vast, ranging all the way from differential and integral equa-
tions and the theory of Hilbert spaces to the mathematics encountered in con-
structing Green's functions and solving boundary-value problems in physics. Need-
less to say, no single book can do justice to such a variety of topics, particularly
when, as in the present case, it attempts to start at the beginning of things. Never-
theless, it is the firm conviction of the authors that the notion of linearity which
underlies these topics and ultimately enables them to be classified as branches of a
single mathematical discipline can be developed in such a way that the student will
we are in a position to see this method as a technique for inverting linear differ-
ential operators. This leads naturally to the notion of Green's functions and their
associated integral operators, which we then treat in detail for initial-value prob-
lems. These ideas arise again in Chapter5 when we study the Laplace transform,
and our approach is such that we are able to give an integrated treatment of what
all too often strike the student as unrelated techniques for solving differential equa-
tions. The sixth, and last, in this sequence of chapters extends the survey of linear
PREFACE VH
differential equations beyond the customary beginning course by proving the Sturm
separation and comparison theorems and the Sonin-Polya theorem. Al this point
we anticipate our later needs by using these results to study the behavior of solu-
tions of Bessel's equation long before these solutions have been exhibited in series
form. Finally, continuing in the same spirit, we introduce the method of un-
determined coefficients for generating power series expansions of functions defined
by equations with analytic coefficients,
In Chapters 7 and 8 the setting changes to Euclidean spaces, and metric concepts
are introduced for the first time. We begin by proving the standard results for
finite dimensional spaces, and then proceed to discuss convergence in finite and
trate this theory by introducing several of the classical (Fourier) series of analysis,
first relative to the trigonometric functions, and then, in succession, relative to the
In Chapter 12, we pull the various threads of our story together by introducing
two- point boundary-value problems. We define eigenvalues and eigenvectors and
discuss the eigenvalue method for solving operator equations, As usual we begin
with the finite dimensional case, which is reduced to a problem in elementary alge-
bra via the characteristic equation, and then generalize to symmetric operators on
function spaces. The question of the existence of eigenfunction bases is treated in
a theorem unproved) of sufficient generality to cover the boundary-value prob-
(left
lems considered in the chapters which follow. We conclude this discussion by re-
turning to the subject of Green's functions to establish their existence and unique-
ness for problems with unmixed boundary conditions.
The last three chapters of the book use these results to solve boundary-value
problems involving the wave, heat, and Laplace equations. The physical signifi-
cance of these equations is discussed and the method of separation of variables
is applied to reduce the problems considered to Sturm-Liouville systems which
fall under our earlier analysis. The question of the validity of the solutions obtained
is settled for thewave equation by appeal to earlier results on the convergence of
Fourier series. Various forms of Laplace "s equation are then considered, and the
elegant theory of harmonic polynomials is introduced. Finally, cylindrical regions
make their appearance, and the Frobentus method is developed to the point where
Bessel's equation can be solved and orthogonal series involving Bessel functions
constructed.
The book ends with four appendices containing material which would have been
unduly disruptive in the body of the text. There we provide a discussion of point-
wise and uniform convergence which is sufficient for our needs, a brief treatment of
determinants, and a development of vector field theory to the point where unique-
ness theorems for boundary-value problems can be proved.
viii PREFACE
Having outlined what is in the book, a few words may be in order concerning
what is mt First, this is not a text in linear algebra. Thus, even though we do
present much of the material usually taught in a first course in linear algebra, a few
familiar topics have been omitted as unnecessary Tor the analysis we had in view.
We have also tried to keep the discussion sharply in focus at all times by giving
formal definitions of new terminology and precise statements of results being
proved or used. For the most part, theorems stated in the text are proved on the
spot. Those which are not comprise results whose proofs were felt to be either too
difficult for a book at this level or unenlightening in view of our objectives. Such
statements are usually accompanied by a reference to a proof in the literature.
In its present form this book is sufficiently flexible to be used in one of several
courses. For instance, Chapters through 6 plus parts of 7 and 15 provide mate-
1
rial for a combined course on (ordinary) differential equations and linear algebra
at the introductory level. On the other hand, Chapters 7 through 1 1 are logically
independent of everything which precedes them, save Chapter 1, and can be used
to give a course on series expansions and convergence in Euclidean spaces. By
omitting Chapter 10 and adding Chapter 12, the first few sections of Chapter 2,
and portions of Chapters 13 through 15, one obtains ample material for a one-se-
mester course in boundary-value problems suitable for students who are able to
solve elementary differential equations, Further there is more than enough mate-
rial (though not exactly of the traditional sort) for several of those courses which
go under the name of "engineering mathematics."' In fact, this book was written
primarily for such courses, and was motivated by the belief (or hope) that engineers
ultimately profit from mathematics courses only to the extent that these courses
present an honest treatment of the ideas involved.
For everyone. The internal reference system used in the text works as follows;
Items in a particular chapter are numbered consecutively as t for example, (3-1) to
(3-100). The first numeral refers to the chapter in question, the second to the num-
bered item within that chapter.
Throughout the book we have followed the popular device of indicating the end
of a formal proof by the mark | in the belief that students derive a certain comfort
from a clearly visible sign telling them how far they must go before they can relax.
PREFACE lx
As usual, sections marked with an asterisk may be omitted without courting dis-
aster. Everything so marked is cither a digression which the authors had not the
strength to resist, or material of greater difficulty than thai in the immediate vicinity.
As a gesture toward scholarly respectability, we have included a short bibliog-
raphy comprising those books which the authors personally found especially useful,
and for the convenience of those inveterate browsers of books we have prepared
an index of special symbols used in the text (see pp. xvi-xvii). Finally, a dia-
gram showing the logical interdependence of the various chapters appears after
the table of contents.
Debts and acknowledgements. Collectively and individually the authors are in-
debted to a large number of people who at long last can be publicly thanked:
First, the numerous students who have used portions of this material more or
less willingly at Dartmouth College and Indiana University over the past several
years, and whose comments have been far more valuable than they ever imagined.
Second, the surprisingly large number of professional colleagues whose advice
has been sought, sometimes unknowingly, and who have been more than gen-
erous in answering questions and furnishing criticism. In particular, special
thanks are due Professors H. Mirkil of Dartmouth College, G, Rota of Mas- C
sachusetts Institute of Technology, and M. Thompson of Indiana University, and
also Mr. L. Zalcman, presently at M. I. T.
And third, Mrs. Helen Hanchett of Hanover, New Hampshire, and Mrs.
Darlene Martin of Bloomington, Indiana, for their patience, good nature, and
unfailing accuracy in preparing typewritten versions of the manuscript too nu-
merous to count.
Thanks are also due. and hereby given, Dartmouth College for assistance ren-
dered in preparing a preliminary version of the manuscript and the Addison-
Wesley staff for seeing the book through press.
Lastly, thanks of a very special sort to our several wives for their constant sup-
port and encouragement as well as their equally constant insistence that we get on
with things and finish the job.
Conclusion. It seems to be one of the unfortunate facts of life that no mathe-
matics book can be published free of errors. Since the present book is undoubtedly
no exception, each of the authors would like to apologize in advance for any that
still remain and take this opportunity to state publicly that they are the fault of the
other three.
January, 1966 D. L. K.
R.G. K.
D. R. O.
F. W. P.
logical interdependence of chapters
Vector Spaces
2 7
Linear Euclidean
Tra nsfa rma t ions Spaces
3 S
Linear Differential Convergence in
Equations Euclidean Spaces
4 5 10 9
12
Boundary-Value
Problems
6 11
15 14
1-1 Introduction . 1
xl
xii CONTENTS
7 EUCLIDEAN SPACES
CONTENTS xiri
FOURIER SERIES
3 700
IV-l Surfaces in £R , surface area
IV-2 Surface integrals of vector fields 707
IV-3 The divergence theorem ............. 713
IV-4 Boundary-value problems revisited: uniqueness theorems .... 718
INDEX 767
.2 5 C i
JS
|l 6 e
a o
u t "3 '5
£ s.
CO •a *>
"
aw *
2
-
— a O
3J
&2 CJ
a 2 -£
1 E u
5P
o
I 1
5 I 2| Cl
E
o _ 1 ^3
e c
g G
«
B 1 1 s
OB c c
g
i o o l| >> 1 2
b* a T3 c
1 a, sr S
o vi b ™ 8 ° 8
o g
1 3 <u E
* £ S- 1 c a S J3
1 ! 3
5. J :-, Uj UJ U y: v <sj X w ?
JO
4 a B k
«
H Q
< 8
IS
i
<
2 §
MS i! .
M —
5 = =
V-j
™ <3 B M
H 2 8 2 fl
-2 O
3 i 1
•j Ia * 1 3 £ c u
£ .£
i
i a I 3 5 M I - I s
"5 § I o o ° =
a
£ <s
4) * 41 cj el
J £
a i 8 g i s*
t3 3i I J £ Q crt^.
- n id «n 00 o\ t»>
5 5!
el a
O fi
i
I
a 1 5,
s
§ _ |
1
II IS I1 M 'c
c '0
o =
e 53g 60
90S
1
m fit fi —
O
E
i_
s
c
a
Cl
^5
n a
2 I -
% 5
a
H * -3 A Q
tj 3
^= Hi
cj
u 1
U
."e
]u
1 ~r
1
rl s
SO CO
a £ I< E
T3 ^c 5 i it,
1 o ™ S
f** f»I** r-
a
f si S
J ti 2 5 X 13
c
-C 5
,2 c
% O
E "B
e 8
^
1 -0
!* a J2,
^ *s C fa
3 3 O
O fa
.5
3 ~-
3 * c B
y
3 1u l ^ a
M 5e a
| O u "8
x
1
f
,5 i
3 B c -
a
.9 M ^o
E £
c <•-
1c 1 -2
CJ
c Q
e O
ffl
s s
'J
6 a .2c B 1 I
2 'E. a
h. n
B
is
•5
_rt
'
<* a 3
D
IS
C
I 8
B c B
'j S O U
B a S3 I 3 H
s
c I 9
B B 'c a
E 5 a 3 S3 90 *-< P a £ z 5 s2
3 3 -h rv ^
3 ^
real vector spaces
1-1 INTRODUCTION
The Cartesian plane of analytic geometry, denoted by (R 2 is one of the most ,
x = (x u x2 )
Throughout this book we shall use boldface type (i.e., x, y, . . .) to denote vectors.
l
REAL VECTOR SPACES CHAP. 1
x + (y + z) = (x + y) + z, (1-2)
x + y = y + x. (1-3)
x + - x, (1-4)
and
x + (-x) = (1-5)
for every x. Taken together, Eqs. (1-2) through (1-5) imply that vector addition
behaves very much like the ordinary addition of arithmetic.
As well as being able to add vectors in (R 2 we can , also form the product of a
real number a and a vector x. The result, denoted ax, is the vector each ofwhose
components is a times the corresponding component of x. Thus, if x = (x 1} x 2 ),
then
ax = (ax i, ax 2 ). (1-6)
lx = x. (1-10)
lx=(-xi,-x2 )
The validity of each of these equations FIGURE 1-3
= a(p(x u x 2 j)
- a(/3x),
The reason for calling attention to Eqs. (1-7) through (1-10) is that they,
together with properties (1-2) through (1-5) for vector addition, are precisely
what make (R 2 a real vector space. Indeed, these equations are none other than the
axioms in the general definition of such a space, and once this definition has been
2
given, the above discussion constitutes a verification of the fact that (R is a real
vector space. But before giving this definition, we pause to look at another
example.
This time we consider the set Q[a, b] consisting of all real valued, continuous
functions defined on a closed interval [a, b] of the real line.* For reasons which
will shortly become clear we shall call any such function a vector, and, following
our general convention, write it in boldface type. Thus f is a vector in Q[a, b] if and
only if f is a real valued, continuous function on the interval [a, b].
At first sight it may seem that Q[a, b] and (R 2 have nothing in common but
the name "real vector space." However,
one of those instances in which
this is
first impressions are misleading, for as we
remarkably
shall see, these spaces are
similar. This similarity arises from the fact that an addition and multiplication
by real numbers can also be defined in e[a, b] and that these operations enjoy
2
the same properties as the corresponding operations in (R .
Turning first to addition, let f and g be any two vectors in Q[a, b]. Then their
sum, f +
g, is defined to be the function
(i.e., vector) whose value at any point x in
f n.^^/^.1
(f + g)(x) = f(x) + g(x). (1-11)
It is now easy to verify that apart from notation and interpretation Eqs. (1-2)
through (1-5) remain valid in Q[a, b]. In fact,
f + (g + h) = (f + g) + h (1-12)
and
f + g = g + f (1-13)
follow immediately from (1-11), while if denotes the function whose value is
* The
closed interval [a, b] is the set of all real numbers x such that a x < b; i.e., <
the interval from a to b, end points included. By contrast, if the end points are
[a, b] is
not included in the interval, we speak of the open interval from a to b, and write (a, b).
t For a proof of this fact see, for example, C. B. Morrey, Jr., University Calculus with
Analytic Geometry, Addison- Wesley, 1962, p. 89.
REAL VECTOR SPACES CHAP. 1
for every f in e[a, b] (Fig. 1-5). Finally, if — f is the function whose value at x is
f + (-f) = 0. (1-15)
2
We have seen that the sum of two vectors in (H is found by adding their cor-
possible in the present example and may be achieved as follows. If f is any vector
in Q[a, b], we agree to say that the "component" of f at the point x is its functional
value at x. Of course, every vector in Q[a, b] then has infinitely many components,
one for each x in the interval [a, b], but once this fact has been accepted it becomes
clear that Eq. (1-1 1) simply states that the sum of two vectors in e[a, b] is obtained
2
by adding corresponding "components," just as in (R .
2
The analogy with (R is now complete, for Eqs. (1-7) through (1-10) are also
valid in <3[a, b]. We restate them here as
If = f, (1-20)
that the space Q[a, b] is much more than an idle example. Indeed, a great deal of
our later work will be devoted to a study of this and similar vector spaces, and the
student will do well to understand it before reading further.
EXERCISES
In Exercises 1 through 5 compute the value of x y and a(x y) for the given vectors + +
x and y in (R 2 and real number a. Illustrate each of your computations with an appro-
priate diagram.
1. x = (0,2), y = (-1,1), o = 3
2. x = (i,l), y = (1,-2), a = -2
3. x = (-ii), y = (-2,-1), a = -1
4. x = (5,-2), y = (-3,2), a = \
5. x = (-5, -2), y = (-1, -1), a = -3
In Exercises 6 through 10 compute the value of f + g and a(f + g) for the given vectors
f and g in e[— 1, 1] and real number a. Illustrate each of your computations with an
appropriate diagram.
6. f(x) = 2x, g(x) = x 2 — x -\- \, a = 2
8. f(x) = e
x
g(x) = e~ x a = %
, , ,
x+3 x — 2 1
ft
9.
,,
«*) -
v
^33. g(*) = -^^'" =
5
11 •
11. sin
*-
+ 1
n
12. = I
x if x ^ °
x<0
i i
\x\
7 T [_ x if
13. tan x
14. tan (2* +D
15- In \x\ 16.
x2 3x + 2
REAL VECTOR SPACES | CHAP. 1
1 if X > I * - J if X >
,,
"• fW -
,, ,
i
1 if X <
18 - «*> - {-, - if » <
J
[x + 1 if x <
Solve each of the following vector equations for x.
21. 2x + 3(2, 1) = (0, 0) 22. 4x + 3x = 2(-l, 3)
25. 2f - 1 = sec 2
x 26. In f = 1
27. e
f
= x + 2 28. 2f + e~ x = e^
29. Prove Eqs. (1-2) through (1-5) and (1-7) through (1-10) of the text, and illustrate
With the examples of the preceding section in mind we now give the definition
of a real vector space.
Axioms for addition. Given any pair of vectors x and y in V there exists a
(unique) vector x + y in V called the sum of x and y. It is required that
x + (y + z) = (x + y) + z,
(iii) there exist a vector in 13 (called the zero vector) such that
x + = x
for all x in V, and
x + (-x) = 0.
1-2 | REAL VECTOR SPACES 7
Axioms for scalar multiplication. Given any vector x in V and any real num-
ber a there exists a (unique) vector ax in V called the product, or
scalar product, of a and x. It is required that
(viii) lx = x.
the case of the function space Q[a, b], and before embarking on the general study
of this subject we give several additional examples of such variety as to convince
the reader that real vector spaces are very common objects indeed in mathematics.
Still others will be found in the exercises at the end of this section. In each case
we leave the verification of Axioms (i) through (viii) as an exercise to aid the
beginner in assimilating the various requirements of the definition.
Example The real numbers, with the ordinary definitions of addition and
1.
multiplication, form a real vector space. In this case Axioms (i) through (viii)
are all familiar statements from arithmetic.
and
ax = (ax i, . . . , axn ). (1-22)
Example 3. Let (P denote the set of all polynomials in x with real coefficients,
and let polynomial addition and multiplication by real numbers be denned as in
high school algebra. Then (P is a real vector space.
This completes our list of basic examples. As we shall see, each of the real
n
vector spaces Q[a, has distinctive features not shared by either
b], (R , and (P
of the others, and for this reason these three spaces will occupy a special place in
our later work. And since we have already mentioned the future importance
n
of Q[a, b], it is only fair to add that (R and <P will in turn receive a great deal
of attention.
Before we continue, a few general remarks concerning the definition of a real
vector space are in order. First, since vector addition is associative, any finite
ambiguity. Moreover, the commutativity of vector addition implies that the value
of such a sum does not depend upon the order of the summands.
Secondly, in the interest of simplicity we shall write x — y from now on in
place of the unwieldy expression x + (— y). This, of course, involves the usual
change in terminology, for we then say that x — y is obtained by subtracting
y from x.
And finally, we call attention to the reasonably obvious fact that our insistence
upon using real numbers in the definition of a vector space is quite unnecessary.
Complex numbers, for instance, would do just as well, in which case we would
have what is known as a complex vector space. Even further
unsurprisingly
generalizations are possible.However, these more general vector spaces will not
appear in this book, and so we agree that the term "vector space" will always
mean real vector space, as defined above. Furthermore, the term scalar will be
used now and again as a synonym for the words "real number" in consonance
with the standard terminology of the subject.
EXERCISES
n
1. With addition and scalar multiplication as defined in the text, prove that (R is a real
vector space.
2. Prove that <P is a real vector space. (Recall that the vectors in (P may be written in
the form
flo + a\x + • • •
+ a nxn ,
«3 = l;
as = — 1.
5. Find the value of aipi + «2P2 + «3P3 in (P when
(a) pi(x) = x 2 — x + 1, cxi = 2,
P2(*) = 3X 3 + 2X — 1, OL2 = — 1,
P3(*) = — * 3 + 2x, a 3 = —2;
(b) P i(jc) = 2x* - Ax 2 + 1, 01 = |,
P2W = -x 4 + 2x 3 + x 2 - x + 2, a 2 = 2,
p 3 (x) = -x 4 + 4x 3 - 2x - f, a 3 = -1;
(c) pi(x) = ^x 3 — 2x 2 + ^, ai = 3,
p 2 (x) = 2x — 1, a 2 = -|,
p 3 (x) = x 2 a 3 = 2.
,
6. Does the set of all polynomials in x with integral coefficients form a real vector space
with the usual definitions of addition and multiplication by real numbers? Why?
7. Let C denote the set of all complex numbers, i.e., all numbers of the form a + bi,
x = (xi, x2 , . .
.)
x + y = (xi + yi, x2 + y2 , . .
.);
ax = (ax\, ax 2 , . . .).
(c) all functions which are different from zero at all but a finite number of points
of [a, b];
(f) all continuous functions which are zero on some closed subinterval of [a, b];
(g) all continuous functions which are zero on a fixed subinterval of [a, b].
10. Let (R+ be the set consisting of all positive real numbers, and define "addition" and
"scalar multiplication" in (R + as follows: If x and y belong to (R + let ,
x + y = xy,
where the product appearing on the right-hand side of this equality is the ordinary
product of the real numbers x and y; if a is an arbitrary real number, let
(w + x + y) + z = w + (x + y + z).
This gives
Ox + Ox - (0 + 0)x = Ox.
Now subtract Ox from both sides of this equation, and then use the fact that
1-3 I
ELEMENTARY OBSERVATIONS 11
Ox — Ox = to obtain
Ox + (Ox - Ox) = Ox - Ox
and
Ox + = 0.
Hence Ox = 0.
The proof of (1-24) is similar; this time set x = y = in a(x y) = ax ay. + +
Next, an equally elementary proof establishes the fact that the vector — x and
(— l)x are one and the same. Indeed, since lx = x and Ox = 0, we have
unique and —x is uniquely determined by x, in the sense that they are the only
vectors in V possessing their particular properties. This is the burden of
+ 0' = 0.
On the other hand, the zero vector has the property that + x = x for every x.
Hence
+ 0' = 0',
The second statement of the lemma follows from the sequence of equalities
EXERCISES*
*4. Prove that the commutativity of addition in a vector space is a consequence of the
remaining axioms, as follows
(a) Use Axioms (i), (iii), and (iv) to prove that x z = y z implies x+ + = y.
(This result is sometimes known as the Cancellation law of vector addition.)
(b) Justify each step in the following sequence of equalities:
(e) Use Axioms (v) and (vi) of Definition 1-1 to expand (1 + l)(x + y) in two
ways, and then apply (a) and (d) to deduce that x + y = y + x for all x and y.
1-4 SUBSPACES
Now that we are well armed with examples, we begin the systematic study of
real vector spaces. To do so, we introduce the important notion of a subspace of
a vector space.
to verify that the set of vectors in question satisfies all of the requirements of
Definition 1-1. However, the step by step verification of the axioms in this defi-
nition is both time consuming and tedious, and we now show how this procedure
can be substantially shortened. Specifically, we establish the following:
axX! + a 2 X2 (1-25)
t The term "subset" as used here and in similar contexts in the future means that every
vector in W also belongs to T). Moreover, we always assume that there is at least one
vector in W —a which fact is sometimes expressed by saying that is nonempty. W
1-4 I
SUBSPACES 13
(b) the subspace consisting of the zero vector by itself, called the trivial subspace.
A subspace of V which is distinct from 13 is called a proper subspace.
a continuously differentiable function and the sum of two such functions are
l
continuously differentiable. Hence Q [a, b] is closed under addition and scalar
multiplication and thus is a subspace of e[a, b]. More generally, if e n [a, b] denotes
the set of n times continuously differentiable functions on
all [a, b], then Q m [a, b]
n
is a subspace of Q [a, b] whenever m > n.
Example 4. Let <P W be the set consisting of all polynomials of the form
n—
aQ + a xx + • • '
+ a«-ix
m > n.
Now that we know what a subspace is, it is natural to ask how one might go
about finding all is a hard
subspaces of a given vector space. In general, this
problem, but for certain spaces the answer can readily be given. For instance,
3
it is not difficult to show that the only nontrivial, proper subspaces of (R are lines
and planes through the origin (see Exercise 4, below). And once this observation
* Mathematicians summarize these two facts by saying that is closed under vector W
addition and scalar multiplication. In this language the above criterion becomes the
statement that a nonempty subset of a vector space is a subspace if and only if it is closed
under vector addition and scalar multiplication.
14 REAL VECTOR SPACES | CHAP. 1
has been made one is struck by the fact that the intersection of any two subspaces
3 3
of (R is again a subspace of (R . Actually, this is true in general, as is shown in
Lemma 1-2. IfWi andV? 2 cire subspaces of V, then the set consisting
of all vectors belonging to both W i andV? 2 is a subspace ofV.
for any pair of real numbers a x and a 2 This implies that aix x + a 2 x 2 belongs .
criterion. |
of V. Then, as was noted above, there is at least one subspace of V containing 9C,
namely 13 itself. This being so, we attempt to find the "smallest" subspace of V
containing 9C, where by we mean that subspace of V which contains 9C, and
this
which in turn is contained in every subspace of V containing 9C. To show that
such a subspace actually exists, consider the totality of all subspaces of V which
contain 9C, and let S(9C) denote the set of vectors belonging to every one of these
subspaces ; i.e., S(9C) is the intersection of these subspaces. Reasoning as in the proof
of Lemma 1-2, we see that S(9C) is a subspace of V, and from its very definition it is
the problem of surveying the subspaces of V. Fortunately (and this is the reason
for introducing S(9C) in the first place) such a method is easy to derive. To do so,
we introduce the following definition.
«iXi + • • •
+ anxn ,
(1-26)
And now we can describe S(9C): it is the set of all linear combinations of the vectors
in 9C. Thus once 9C is known, so is S(9C), and (1-26) gives the form of each of its
vectors.
1-4 SUBSPACES 15
FIGURE 1-8
3
Before proving this assertion, let us look at some examples in (R . First let 9C
consist of a single nonzero vector x. Then S(9C) is the line through the origin
determined by x. But the points on this line are simply all scalar multiples ax
of x, and our assertion holds in this case. Next, let 9C consist of two nonzero,
noncollinear vectors, x x and x 2 In this case, S(9C)
is the plane through the origin
.
axXi + a2*2 certainly lies in this plane, and, conversely, we can "reach" any
vector y in this plane by a linear combination of x 1 and x 2 as indicated in Fig. 1-8. ,
Thus S(9C) is again the set of all linear combinations of the vectors in 9C.
We now prove the general result.
Theorem 1-1. Let X be a (nonempty) subset of a vector space 13. Then
the subspace of 13 spanned by 9C consists of all linear combinations of the
vectors in 9C.
Proof In the first place, the set of all linear combinations of vectors in 9C is closed
under addition and scalar multiplication, and hence is a subspace More- WofU
over, the equation x = lx shows that each x in 9C is a linear combination of
vectors in 9C, thus proving that 9C is contained in W. Finally, every subspace of 13
which contains 9C must contain all vectors of the form (1-26) by virtue of the fact
that a subspace is closed under addition and scalar multiplication. In other
words, "W is contained in every subspace of 13 containing 9C, and it follows that
W = S(9C). I
EXERCISES
1 . Determine which of the following subsets are subspaces of the indicated vector space.
Give reasons for your answers.
(a) The set of all vectors in (R 2 of the form x = (1, x 2 ).
(b) The zero vector together with all vectors x = (jci, x 2) in (R 2 for which x 2 /xi
has a constant value.
16 REAL VECTOR SPACES | CHAP. 1
(a) The subset of (P n consisting of the zero polynomial and all polynomials of
degree n — 1.
(b) The subset of (P consisting of the zero polynomial and all polynomials of
even degree.
(c) The subset of (P consisting of the zero polynomial and all polynomials of
degree 0.
(d) The subset of (P„, n > 1, consisting of all polynomials which have x as a
factor.
(a) The set of all functions in e[a, b] which vanish at the point xo in [a, b].
6. Consider the set of all infinite sequences {x u x„, . . . , . . .} of real numbers which
have only a finite number of nonzero entries. Prove that this set is a subspace of the
vector space defined in Exercise 8 of Section 1-2.
3
7. Determine which of the following vectors belong to the subspace of (R spanned
by (1, 2, 1) and (2, 3, 4).
(a) (4,7,6) <c) (4,1,1) (e) (2,9,4)
(b) (-± -i -f) (d) (2, 9, 5) (f) (0,
i,
-f
3
8. Determine which of the following vectors belong to the subspace of (R spanned
by (1, -3, 2) and (0,4,1).
(a) (3, -1, (c) (i 1, (e) (£, -1, f)
8) f)
(b) (2, -2, 1) (d) (2, -|, 0) (0 (|, 3, -|)
9. Determine which of the following polynomials belongs to the subspace of <P spanned
by x 3 2x 2 + 1, x 2 - 2, x 3
+ x. +
(a) x2 - x +3 (c) 4x 3 - 3x + 5 (e) ~4* 3 + fx 2 - x - 1
(b) x2 - 2x + 1 (d) x4 + 1 (0 x - 5
|0 if \ < x < 1.
4,
11. Let f be the function whose value is 1 at every point of the interval [a, b]. Find the
subspace of Q [a, b] spanned by f.
12. Find the subspace of (ft 3 spanned by each of the following sets of vectors.
(a) (2,1,3), (-1,2,1)
(b) x + 1, x2 - 1
(c) 1, x - 2, (x - 2)
2
14. Prove that the intersection of any collection of subspaces of a vector space V is a
subspace of 13. (By the intersection of a collection of subspaces of V we mean the
17. Let 9Ci and 9C2 be two sets of vectors in V, and let 9Ci n 9C2 be the set of vectors
belonging to 9C 1 and'X.2. Furthermore, let us agree that if 9Ci and 9C2 have no vectors
in common, S(9Ci D 9C2) is the trivial subspace of V. Show that S(9Ci f~l 9C2) is a
subspace of S(9Ci) n S(9C2). Give an example in (ft 3 where these two subspaces are
distinct, and one where they are identical.
18. Prove that the following two subsets span the same subspace of (ft 3 .
23. Let W
be the subset of (P„ consisting of all polynomials which have zero as a root,
i
and let W2
be the subset of (P„ consisting of the zero polynomial and all poly-
nomials of degree zero (i.e., constant polynomials). Prove that (P„ = Wi W2 +
(see Exercise 22), and show that each vector in (P n can be written in one and only
one way as the sum of a vector in W 1 and a vector in W2.
24. If 9Ci and 9C 2 are two sets of vectors in V, then 9Ci U 9C2 (read "9Ci union 9C2")
is the set of vectors belonging to 9Ci or to 9C2 (or both). Prove that
3 (see
and illustrate this result by examples chosen from (R Exercise 22).
25. Let a 1 and a 2 be real numbers and consider the linear equation
aixi + 0:2*2 =
in the unknowns xi and X2. A vector (ci, c£) of (R 2 is said to be a solution of this
equation if the substitution of c\ and a for x\ and X2 respectively reduces this
equation to an identity. Show that the set of solutions of the given equation is a
subspace of (R 2 . Describe this subspace graphically.
26. Show that the set of all simultaneous solutions of the pair of linear equations
27. (a) Suppose that the vector x = (ci, C2) is a solution of the pair of linear equations
01*1 + 02*2 = T2 .
Prove that every solution of this pair of equations is of the form x + y, where y is
a solution of
a 1*1 + a 2* 2 = °»
0i*i + 182*2 = 0.
(b) Conversely, with x and y as in (a), prove that every vector of the form x + y
is a solution of (I).
(c) Give a geometric description of the solutions of (II), and then use it and the
above results to obtain a description of the solutions of (I).
*28. Let 9C be an arbitrary subset of a vector space V, and let x and y be vectors in V.
Suppose that x belongs to the subspace S(9C, y) but not to S(9C). Prove that y then
belongs to S(9C, x). (This result is sometimes known as the exchange principle.)
3 perfectly clear
in (R , no two of which are collinear (Fig. 1-9). In this case it is
that the given set contains more vectors than are needed to span the plane
S(xi, x 2 x 3 ), since
, any two of them suffice in this respect. But at least two vectors
1-5 | LINEAR DEPENDENCE AND INDEPENDENCE; BASES 19
3
are always necessary to span a plane in (R and hence we obtain a "minimal" ,
vectors. This example suggests that it may be possible to reduce any finite set
of vectors Xi, x n to a minimal subset which continues to span S(xx,
. . . , xn ). . . . ,
This is in fact the case, as we show in Theorem 1-2, but before doing so we intro-
duce some useful terminology.
x = aiXi + • • •
+ a„x n ,
where the a; are scalars. If, on the other hand, no such relation exists,
x is said to be linearly independent of Xi, xn . . . , .
spanned by them (Theorem 1-1). In particular, each of the vectors X;, 1 < i < n,
is linearly dependent on x 1} x n since it belongs to the subspace spanned by
. . . , ,
these vectors. We also note that the equation = Ox implies that the zero vector
is linearly dependent on every vector.
FIGURE 1-9
<*iXi + a 2x 2 + 0:3X3 =
implies that
«i + 2a 2 — «3 = 0,
3«i + «3 = 0,
— «! + a2 = 0,
2«i + 3a 2 = 0,
We leave the proof of the above test as an exercise (see Exercise 9 below),
with the strong recommendation that it be done.
And now we are ready to show how one can weed the extraneous vectors from
any finite set x ls . . . , xn without disturbing S(x l5 x n ). The basic idea is . . . ,
x = jSiXi + • • •
+ /8 nxn
X = (jSl + «liSn)Xi + * * •
+ (|8 n _i + a _ii3n)Xn-l.
ra
This proves that x is already a linear combination of Xi, . . . , x n _i, and hence
that S(xi, . . . xn _i)
,
= S(x x , . . . , x„). In this case we drop the vector x n from
the set x l9 If, on the other hand, x n
...,x n . is not linearly dependent on
Xi, . .
.
, xn _i, we keep it.
If we repeat this procedure with each of the x* in turn, dropping x^ if it is linearly
dependent on the remaining vectors in the (possibly modified) set, keeping it
otherwise, it is clear that we obtain a linearly independent subset of x b xn . . . ,
which spans the subspace S(xi, x n ). This, of course, is what we started out . . . ,
(Note, however, that in general 9C contains many such subsets. This was the case
for instance in the example given at the beginning of this section. Other examples
will be found in the exercises below.)
Linearly inflependent sets enjoy a special status in the study of vector spaces,
and among such sets those which span the entire space are particularly important.
Such sets are named in
in one and only one way as a linear combination of these basis vectors, namely,
x = xj + x 2 j + * 3 k. This last property actually serves to characterize a basis
in a vector space, as we now show.
Proof First suppose that ei, e n is a basis for V. Then the . . . , e; span V, and
hence every vector in V can be written in at least one way as
x = «ie! + • • •
+ a n en . (1-27)
x = iSid + • • •
+ /3 nen (1-28)
But since ei, .e n is a basis for V, these vectors are linearly independent. Hence,
. . ,
by our test for linear independence, each of the coefficients in (1-29) is zero,
and it follows that an = jS 1} an = £„, as desired. . . . ,
prove their linear independence in order to show that they are a basis for V. To
accomplish this, we observe that = Oe x 0e w and that our assumption + • • •
+
concerning the uniqueness of such expressions implies that this is the only represen-
tation of as a linear combination of e ls . .
.
, en . Thus if o^ex + • * *
+ «wen = 0,
we must have ai = • • • = an = 0, and the test for linear independence now
applies. |
i
e2 = (0,1,..., 0),
en = (0, 0, . .
.
, 1)
n
are a basis for (R , since x = *iei + + xn en is the only way of expressing • • •
w
basis is called the standard basis for (R .
n
Example 2. Again in (R , let
ei = (1,0,... ,0),
e; = (1, 1, . .
.
, 1),
where, in general, e^ is the «-tuple having and 0's there- l's in the first i places
n
after.Then ej, . . . , e« is a basis forx = (x 1} xn ) (R . To prove this let . . . ,
be given, and let us attempt to find real numbers «i, ...,a n such that
x = aie[ + • • •
+ a n en. In order that such an equality hold we must have
= (a\ + a2 + • • •
+ oi n , «2 + ' * *
+ ot n , ... , a n ),
«1 + «2 + * * '
+ «n = *1,
Hence
«i = Xi — x2 ,
a2 = x2 — X3,
Qt w — = Xn — i Xn ,
&n = Xn ,
1-5 | LINEAR DEPENDENCE AND INDEPENDENCE; BASES 23
n
Thus the e t are a basis for (R
-
, as asserted.
~
Example 3. The polynomials 1, x, x2 ,
. .
.
, xn l
form a basis for the vector
space since each polynomial in this space can be written in
(Pn one and only one
~
way in the form a a^x an _\X n x
+ + • • •
+ .
Example 4. Let pi(x), p n (*) be any finite set of polynomials in (P, and let d
. .
. ,
be the maximum of the degrees of the Pi(x). Then no linear combination of these
polynomials is of degree greater than d, from which it follows that p! (x), pn (x) . .
. ,
is not a basis for (P, since (P contains polynomials of arbitrarily high degree. Thus
(P does not possess a basis in the sense of Definition 1-5.
n
In Examples 1 and 2 we exhibited two distinct bases for (R , and in each case
found that the number of vectors involved was the same. This was no
total
coincidence, for it can be shown that any two bases in a vector space 13 always
have the same number of elements. In other words, the number of vectors in a
basis for 13 (provided 13 has a basis) is an intrinsic property of 13 itself. We shall
prove this important result in Section 7, along with certain other facts about bases
in vector spaces, and mention it here only in order to justify the following
definition.
n
On the strength of the above examples we can assert that both (R and <?„ are n-
dimensional, and that (P is infinite dimensional.
EXERCISES
2. Find all linearly independent subsets of the following sets of vectors in (P4.
* By convention, the vector space consisting of just the zero vector is assigned di-
mension 0.
24 REAL VECTOR SPACES | CHAP. 1
3
3. Are the vectors (0, 2, -1), (0, \, -£), (0, §, -£) linearly independent in (R ? If
not, find a linearly independent subset which has a maximum number of elements.
2
4. Prove that each of the following sets of vectors is a basis for (Si .
i, J, i + J, «'i + /3'j.
7. Express (2, —2, 1, 3) as a linear combination of the vectors in the various bases of
Exercise 6.
2
8. (a) Show that the functions 1, sin 2 x, cos x are linearly dependent in C[— x, x].
(b) Show that the functions 1, cos x, cos 2x are linearly independent in C[— x, x].
9. Prove that the vectors xi, . . . , x„ are linearly independent if and only if the equation
aixi + • • •
+ a«x„ =
implies that ai = • • • = an = 0.
2
10. Prove that the vectors (a, b) and (c, d) are linearly independent in (R if and only if
ad - be 9^ 0.
*11. Prove that the vectors (jci, jc 2 , x 3 ), (yi, y2, V3), (zi, z 2; z 3 ) are linearly independent
3
in (Si if and only if
xi yi z\
X2 72 Z2 j£ 0.
*3 V3 Z3
*13. Show that the functions 1, sin x, cos x, sin 2x, cos 2x, sin nx, cos nx are linearly .
. ,
,
A, 2^ 2' 2 2
15. Express x 2 and x 3 as linear combinations of the basis vectors for (P4 given in
Exercise 14.
1-5 | LINEAR DEPENDENCE AND INDEPENDENCE; BASES 25
16. Assume that xi, X2, and X3 are linearly independent vectors in V. Prove that
xi + X2, xi + X3, and X2 + X3 are linearly independent.
17. Prove that the functions 1, ex , e 2x are linearly independent in C[0, 1]. [Hint: Differ-
entiate the expression a + fie
x
+ le 2x = 0.]
18. Prove that the functions 1, e x , xe x are linearly independent in 6[0, 1].
— — 2 — -1
1, x a, (x a) , . .
. , (x a)" ,
where a is an arbitrary real number, form a basis for (Pn [Hint: Consider the .
*22. Let 9C„ be any set of n polynomials in (Pn , one of each degree 0, 1, . .
.
, n — 1.
23. Let 9C4 be the basis for (P4 consisting of the cubic polynomial x 3 + 2x + 5 and
its first three derivatives (see Exercise 22). Write each of the following polynomials
9C is a as a linear combination of polynomials in 9C4.
(a) jc
3
+ 2x + 5
(b) x2 + 1
(c) 2x 3 - x2 + 10* + 2
24. Let 9C be a finite linearly independent subset of a vector space V, and suppose that
every finite subset of V which properly contains 9C is linearly dependent. Prove that
9C is a basis for 1).
25. Let 9C be a finite subset of a vector space T) which spans V, and suppose that no
proper subset of 9C spans V. Prove that 9C is a basis for V.
Use the fact that none of the et are zero to successively eliminate i, j, and k from
these equations, and conclude that n < 3. Now reverse the argument to show
that 3 < «.]
26 REAL VECTOR SPACES CHAP. 1
x = a^i + i
a n£n
be the unique expression for x in terms of this basis (Theorem 1-3). Then
the scalars <x\, . . . ,a n are called the coordinates or components of x with
respect to ei, . . . , e w and the basis vectors themselves are said to form
,
Thus a basis is a coordinate system, and the unique expression for a vector as a
linear combination of basis vectors is nothing other than the "decomposition"
of the vector into its components along the various coordinate axes. In these
terms the direct statement in Theorem 1-3 assumes the following eminently
reasonable form: The coordinates of a vector are uniquely determined by the coordi-
nate system.
At the same time, we caution the student not to expect too much from a co-
ordinate system, and especially not to fall into the error of expecting coordinate
axes to be mutually perpendicular. Strictly speaking, of course, the concept of
perpendicularity in a vector space has no meaning yet, but it will be defined in
Chapter 7. Nevertheless, it is common knowledge that certain coordinate axes,
n
such as the standard ones in (R , are mutually perpendicular. We merely wish
to emphasize the sometime nature of this phenomenon, and call attention to the
(4,2) = 4ei + 2e 2 .
However, if we use the coordinate system e( = (1,0), e'2 = (1,1), then the
«£ = (U)
e2 = (0,l)
ei = (l,0)
FIGURE 1-10
. :
2
Indeed, the vector in (R having coordi-
nates 4, 2 with respect to e[, e'2 is the or-
dered pair (6, 2). (See Figs. 1-10 and
1-11.)
Finally, we call attention to the fact FIGURE 1-11
that the operations of vector addition and
scalar multiplication are converted into ordinary addition and multiplication when
they are carried out with respect to a basis. For then these operations are per-
formed componentwise, irrespective of the nature of the vectors involved (n-tuples
of real numbers, polynomials, etc.). We prove this assertion as follows
Theorem 1-4. Let ei, e n be any basis for a vector space V. Then
. . . ,
x + y = («! + /3i)ei + • • •
+ (a„ + /3 n )e n ,
and that
ax = (aai)ei + • • •
+ (aa n )e n ,
as asserted.
Among its various implications, this theorem foreshadows the use of bases in
finite dimensional vector spaces whenever extensive numerical calculations are
in the offing. On the other hand, as long as we are concerned with the general
theory of vector spaces, bases are a distinct hindrance. This stems from the fact
that whenever a basis is used in the proof of a theorem which purports to be a
general statement about finite dimensional vector spaces, one must then prove
that the result in question is independent of the particular basis chosen to prove it.
And this is usually as difficult as it is to construct a coordinate free proof of the
original statement.
EXERCISES
Find the coordinates of each of the following vectors in 3 with respect to the basis
1 I
3. Prove that the vectors (2, 1, 0), (2, 1, 1), (2, 2, 1) form a basis for (R 3 Find the .
vectors in (R 3 which have the following coordinates with respect to this basis.
Find a basis in 4 with respect to which the vector (—3, 1,2, —1) has coordinates
5. (R
1,1,1,1.
6. Let ei, e2, e3 and ei, e2, e% be bases for a vector space V, and suppose that
d = + «2e2 + «3e3,
aie'i
e2 = + 2e'2 + foe^,
jSie'i /3
e3 = Tiei + Y 2 e 2 + 7 3 e 3 .
Find the coordinates of the vector x = x\ei + *2e2 + *3e3 with respect to the
basis ei, e2 , e3 .
7. Does there exist a basis for (R 2 with respect to which an arbitrary vector (xi, X2) has
coordinates 2*1 and 3^2?
8. Find a basis for (R 2 with respect to which an arbitrary vector (xi, X2) has coordinates
xi and x\ 2x2. +
9. Let ei, . . .
, e„ and ei, . .
. , e„ be bases for a finite dimensional vector space V,
and let
e2 = «i2ei + a22e2 + • • •
+ a n 2e n ,
e« = ai nei + a2 n e2 + "
"
•
+ «nrae ra .
1-7 DIMENSION
The technique used to prove this theorem has already been introduced in Exer-
cise 27 of Section 1-5 to treat a particular case. The reader may find it helpful to
keep that exercise in mind while reading the following proof.
* The statement of Theorem 1-5 also applies to the trivial space consisting of just the
zero vector, provided we agree that the empty set of vectors is a basis for this space. Such
an agreement is consistent with our definition of a basis for a vector space, and with the
convention that the dimension of the trivial space is zero.
1-7 I
DIMENSION 29
Proof. Let ei, . . . , e n be a basis for V, and suppose, contrary to the assertion of
the theorem, that V contains a linearly independent set e[, e'm in which . . . ,
m > n. Express each of the e^as a linear combination of the e z thereby obtaining ,
e'i = auei + a 2 ie 2 + • •
+ « nl e n ,
©2 = «12©1 + Oi22^2 + * • •
+ oi n2 en ,
:
(1-30)
Cm = «i wei + a 2m e 2 + • • •
+ « nm en ,
in which the an are Since none of the ey is the zero vector, at least one
scalars.
of the an from zero
is in each of these equations. (Recall that the zero
different
vector is linearly dependent on every vector in V.) Thus, by relabeling the e t if -
necessary, we may assume that an 5^ 0. This done, solve the first equation for
ei, and substitute the value obtained in the remaining m — 1 equations. This
eliminates ei from (1-30), and yields a system of equations of the form
e'3 = |823e 2 + i8 33 e3 + • • •
+ Pn&n + j8i 3 ei,
:
(1-31)
em = |82me 2 + |S 3 me 3 + • • •
+ Pnmtn + fi\m£\-
Focusing our attention on the first of these equations we note that the linear
independence of e[ and e 2 implies that at least one of the coefficients /3 22 ,
the system
e'3 = T 33 e 3 + • • •
+ Tn3 e n + 7i 3 e'i + T 23 e'2 ,
Let us now speculate on the effect of our assumption that m is greater than n.
A moment's thought will reveal that by continuing the above process of elimina-
tion we will eventually find ourselves confronted with a system of m — n equations
expressing each of the vectors en+i, . . . , e« as a linear combination of ej, , e'
n . . . .
Corollary 1. IfV has a basis containing n vectors, then every basis for V
contains n vectors.
Proof. If e x , . . . , en and e[, . . . , e m are bases for V, then the above theorem
implies that m < n, and n < m. Hence m = n. |
30 REAL VECTOR SPACES | CHAP. 1
This result furnishes the necessary justification for the definition of the dimension
of a vector space given in Section 1-5.
Theorem 1-5 also allows us to prove the reassuring fact that every subspace of a
finite dimensional vector space is finite dimensional, and that its dimension does
not exceed the dimension of the whole space. This is the content of
To prove the truth of this observation, we apply the test for linear independence
(p. 20) as follows. Suppose that
«iCi + • • •
+ amem + <x m+1 em+1 = 0. (1-32)
ai 0Cm
em+l — «
el ' '
' *
vmj
«m+l «m+l
and e m+ i is in W. Thus
aiei + • • •
+ amem = 0,
Hence all of the coefficients in (1-32) are zero, and ei, e m+ i are linearly . . . ,
independent.
We now repeat the above argument, this time starting with the subspace
S(ei, . . . , e w+ i). If S(ei, e OT+ i) is a proper subspace of
. . . , V we can enlarge
ei, . . . , e TO+ i to a linearly independent set in V containing m+ 2 vectors. But
1-7 | DIMENSION 31
Theorem 1-5 implies that this process must come to a halt after n — m steps,
atwhich point we will have a basis for V. With this we have proved the following
important and useful result.
V.
EXERCISES
4. Let*W be the set of all polynomials p(x) in (P n such that p(l) = p'(l) = 0.
(a) Prove that V? is a subspace of (P„, and find a basis for "W.
(b) Extend the basis for "W found in (a) to a basis for (Pn .
5. (a) Find the dimension of the subspace of Q[— w, x] spanned by the vectors 1,
sin x, sin 2 x, cos 2 x.
(b) Repeat part (a) for the vectors sin x cos x, sin 2x, cos 2x, sin 2 x, cos 2 x.
10. Let V be a vector space of dimension n. Prove that V contains a sequence of sub-
spaces
V ,V!,...,V n
having the following two properties
(a) dimt),- = /;
11. Let 1)i and 132 be finite dimensional subspaces of a vector space 13, and suppose
that 13 i and 132 have only the zero vector in common. Let ei, . .
. , e m be a basis
for 13 i, and ei, e» a basis for D2. Prove that ei,
. .
.
, em , ei, . . . , . .
. , e'n is a basis
for the subspace W = 13 1 + 13 2 of 13 (cf. Exercise 22, Section 1-4). Deduce that
dimW = dim 13 1 + dim13 2 .
*12. Let 13 1 and 13 2 be finite dimensional subspaces of a vector space 13. Prove that
13 1 + D2 is finite dimensional, and that
Informally, geometric vectors are arrows in the plane or 3 -space. As such, they
are familiar to anyone who has studied elementary physics, where they appear
as forces, velocities, accelerations, etc., i.e., quantities having a magnitude and
direction. In this section we propose to examine some of the vague ideas associated
with the use of such arrows, and make these ideas precise by constructing the
space of two-dimensional geometric vectors. Besides furnishing us with still an-
other example of a real vector space, this discussion will provide the link between
our definition of the term vector and the vectors introduced in elementary calculus
and physics.*
The geometric notion of an arrow in the plane finds its mathematical analogue
in the concept of a directed line segment. Specifically, the line segment between
2
two distinct points A and B in (R is said to be directed if the points are given a
definite order, say A, B. In this case we speak of the directed line segment from
A to B, which we denote by AB. A is then called the initial point of the segment,
and B the terminal point. We also agree to regard a single point as a directed line
segment, in which case the relevant symbol is AA.
Intuitively the directed line has a magnitude and direction. At
segment AB
first sight one might be tempted to define these concepts as length and angular
measure with respect to a coordinate system in the plane. However, any such
definition would have the grave defect of making the magnitude and direction of
^AB dependent upon the coordinate system used, in conflict with the intuitive de-
mand that they be intrinsically associated with AB itself. Unfortunately there is
* With obvious minor changes the following discussion can be adapted to 3-space or,
no way out of this dilemma so long as we continue to focus our attention upon a
2
single segment. But when we turn to the set of all directed line segments in (R ,
we observe that it is as easy to determine when two segments have the same mag-
nitude and direction as it is difficult to say what these terms mean. Indeed, using
the notion of parallel translation, we can say that AB and CD have the same mag-
nitude and direction if and only if they can be brought into coincidence by such a
translation. Furthermore, the entire theory of geometric vectors can be based
upon this simple observation.
We begin by giving the above discussion formal status in
For future use we note the following simple consequences of this definition:
Ai ~AB, (1-33)
Definition 1-9. The collection of all directed line segments in the plane
which have the same magnitude and direction as a given segment AB is,
At first sight this definition may seem somewhat bizarre, but it does in fact yield
precisely the sort of quantity we want in a geometric vector. For, whatever other
ideas one may have concerning geometric vectors, it is clear that every such vector
must be completely determined by its magnitude and direction. In other words,
it consists of nothing but a magnitude and direction. And when we consider equiva-
opically that segments having the same magnitude and direction are indis-
tinguishable.
But can we accept Definition 1-9 as it stands? Hardly; for we still lack the
necessary assurance that a geometric vector is unambiguously determined by any
one of its representatives. Phrased somewhat differently, the validity of Defi-
nition 1-9 depends upon the fact that no directed line segment can be a repre-
sentative of more than one geometric vector. Intuitively this is clear, but the student
should nevertheless appreciate the need for a proof based upon the definitions.
This is accomplished by the following theorem, which actually establishes a
somewhat stronger result.
2
Proof. If AB is a directed line segment in (R , then by Definition 1-8, \(AB) is
FIGURE 1-12
2
To define vector addition in we make use of the simple and geometrically
9
obvious fact that an arbitrary geometric vector, and P is any point in the
if v is
plane, then there exists precisely one representative of v with initial point P.
(To produce this representative, select any segment belonging to v and translate
2
its initial point to P.) This having been said, let v and w be vectors in Q and ,
the third side of the triangle formed from AB and liC, as shown in Fig. 1-12.
Here again we have a which requires justification, for it is predicated
definition
upon the fact that v + w remains the same regardless of which representatives of v
1-8 GEOMETRIC VECTORS 35
FIGURE 1-13
(v+w)+x = v+(w+x)
and ware used to compute it. But this, we assert, is obvious. For if A'B' ~ AB,
and B'C ~ BC, then triangles A'B'C and ABC will be congruent, and a parallel
translation will bring A'C into coincidence with AC (see Fig. 1-13).
With this difficulty out of the way, we invoke elementary geometry to prove
that v + (w + x) = (v + w) + and that v + w =
x, w + v (appropriate
diagrams appear above in Figs. 1-14 and 1-15). Moreover, if we set = y(AA),
then v + = v for every v in g
2
; while if v = v(AB), it is clear that the vector
—v = v(BA) has the property that v + (— v) = 0. Thus the additive axioms
2
for a vector space are satisfied in g .
2
Next let y(AB) be an arbitrary nonzero vector
and let a be a real number. in g ,
Choose a on the ray from A through B, and let \AB\ denote the
unit of distance
length of the segment AB. Then there exists a unique point C on this ray such
that |^C|/|y4fi| = \a\, where |a| denotes the absolute value of a. Ifa > 0, let
can again be used to prove that the required axioms are satisfied. Thus g 2 is a
real vector space.
36 REAL VECTOR SPACES | CHAP. 1
It is both interesting and instructive to compare this space with the vector space
2
51 defined in Section 1-1. To effect this comparison consider the totality of
directed line segments in the plane which have their initial point at the coordinate
2
origin in (31 . On the one hand, this collection of vectors is none other than the
2
set of vectors which comprise (ft However (and this is the crux of our argument),
.
2
(Recall that every vector in g has precisely one such representative, and con-
versely, every such directed line segment represents a unique vector in g 2 .) In
2
other words, the vectors in (ft are simply a particular set of representatives of the
2 2
vectors in g . From agree to replace each vector in g
this it follows that if we
2
by its unique representative emanating from the coordinate origin in (R , we find
that (R
2
and g 2 thenof precisely the same vectors. Moreover, addition and
consist
2
scalar multiplication in then become identical with the corresponding opera-
g
tions in (ft
2
Thus we conclude that (ft 2 and the space of two-dimensional geometric
.
2 2
vectors are essentially identical, g is simply (ft stripped of its coordinate system,
2 2
and conversely, (ft is g seen by means of a coordinate system.
EXERCISES
In Exercises 1 through 5 find two other directed line segments which have the same
magnitude and direction as AB for the given values of A and B.
In Exercises 16 through 20 find the value of a\(AB) for the given values of A, B, and a.
16. A = (-5, 7), B = (3, 10), a = 4 17. ^ = (9, 2), 5 = (5, -2), « = -J
18. A = (f, 6), B
= (4, |), a = 3 19. A = (0, 5), fi = (11,0), a = -1
20. ^ = (-3, -1), B = (-4, -5), a = £
1-9 I
EQUIVALENCE RELATIONS 37
25. Prove that the addition of geometric vectors is associative. [Hint: Use the result
of Exercise 23.]
26. Using the results of Exercises 22 and 23 prove that a(y w) = <xv aw and + +
(a + /3)v = a\ + j8v for any pair of geometric vectors v, w, and any pair of real
numbers a, /3.
27. Use the result of Exercise 22 to prove that (a/3)v = a(jSv) for any geometric vector v,
(iii) If (x, y) and (y, z) belong to (ft, then (x, z) belongs to (ft.
(ii) jc ~ y implies y ~ x,
One also says that an equivalence relation is reflexive, symmetric, and transitive;
these names being given respectively to properties (i), (ii), and (iii) above.
38 REAL VECTOR SPACES | CHAP. 1
Example 2. Let S be the set of all triangles in the plane, and letA x ~ A 2 mean
that A ! and A 2 are congruent. Then ~ is an equivalence relation on S.
Example 3. Let S be the set of directed line segments in the plane, and let
Example 4. Let S be the set of real valued functions which are continuous at
all but a number of points in an interval [a, b]. If /and g are two such func-
finite
tions, let / ~ g mean that f(x) = g(x) for all but a finite number of values of x
in [a, b]. Then ~ is an equivalence relation on S (see Exercise 1).
EXERCISES
2. Prove that the relation defined in Example 5 above is an equivalence relation. What
is the equivalence class determined by ^? By 5? Byf?
3. Let S be the of all integers, and let m be a fixed positive integer. If x and v belong
set
to S, set x ~y if and only if x — y is divisible by m. Prove that ~ is an equivalence
relation on S, and find all the equivalence classes for this relation when m = 2, 3,
and 4.
(a) If xi and X2 belong to 2D, set x\ ~ X2 if and only if f(x{) = f(x2). Prove
that ~ is an equivalence relation on 2D.
(b) Let (P be the partition of 2D defined by this equivalence relation (see Exercise 4).
Prove that there exists a one-to-one function g with domain <P and range (R such
that g([x]) = f(x) for each x in 2D. (Recall that a function / is one-to-one if
f(xi) = f(x 2 ) implies xi = x 2 .)
*6. The rational numbers. Let S be the set of all symbols a/b, b ^ 0, introduced in
Example 5 above, and let ~ be the equivalence relation that was defined on S; i.e.,
a/b ~ c/d if and only if ad — be. Let Q be the set of all equivalence classes of the
elements of S under this equivalence relation. Then the equivalence classes belonging
to Q are, by definition, rational numbers, and the set Q itself is called the set of all
rational numbers. Thus the rational number [a/b] is the equivalence class containing
the symbol a/b.
(a) To define addition of rational numbers set
ad + be
+ bd J
ac
bd
a'c' ac
Vd'~ ~b~d
and thus conclude that the above equation actually defines a multiplication of equiva-
lence classes.
2
linear transformations
and matrices
Up to this point our study of real vector spaces can best be described as a modest
generalization of some of the ideas implicit in analytic geometry. Although such
terms as linear dependence and independence, subspaces, bases, and the like, may
have been unfamiliar they actually add little to the knowledge of vector spaces
taught in elementary geometry. All this changes, however, as soon as these ideas
are,used to study functions defined on vector spaces. Here new and important
things do happen, and as the following discussion unfolds we shall find the con-
cepts introduced earlier taking on added meaning and significance.
The simplest yet most important functions which arise in the study of vector
spaces are known as linear transformations, and are defined as follows:
FIGURE 2-1
A(0) = 0. (2-3)
Another is that
A(a i x 1 + • • •
+ a n xn ) = a 1 y4(x 1 ) + • • •
+ «n^(x n ) (2-4)
The first of these assertions can be established by setting a = in (2-2), the second
by repeated use of (2-1) and (2-2) in the obvious fashion. In particular, when
n = 2, (2-4) becomes
are satisfied if and only if (2-5) is. (See Exercise 17.) From time to time we shall
use this fact when proving that a function is a linear transformation.
If A is a linear transformation from Vi to V 2 we write A:V\ — V2
> (read
"A maps Vi into V 2 "), and refer to Vi as
the domain of A. In this case the set of all
x=(x l
,x 2 )
= 0,1)
A(-e 2 ) =
(-1,-1)
A(x)=(x 1 ,-x 2 )
one and the same vector space. Indeed, this is one of the most fruitful settings in
which to pursue the study of linear transformations.
We conclude this section by giving a number of examples, several of which will
figure prominently in our later work. For the most part we simply state the
definition of the function in question, and omit the routine verification of linearity
in the expectation that the reader will supply the missing argument for himself.
2
Example 1. Let x = (xi, x 2 ) be an arbitrary vector in (R , and set
A(x) = (*i, -x 2 ).
2
Geometrically A can be described as the linear transformation mapping (ft onto
itself by reflection across the x r axis. (See Fig. 2-3 where, for generality, the
effect of A has been depicted relative to an oblique coordinate system.)
2
Example 2. Let A be the mapping of (R onto itself obtained by shearing the
plane horizontally so that the jc 2 -axis is shifted through a 45° angle as shown in
Fig. 2-4. Analytically, A is defined by the equation
A(x u X 2 ) = (*i + x 2 ,x 2 ),
and is clearly linear.
3
Example 3. Let <£ be any line through the origin in (R , and let A be a fixed
rotation about £. Then, arguing geometrically, it is easy to show that A is a linear
3
transformation mapping (R onto itself.
Example 4. The mapping which sends each vector in T>i onto the zero vector
in V2 is clearly a linear transformation from V 1 to V 2 for all Vi and V 2 It is .
called the zero transformation, and is denoted by the symbol O, irrespective of the
vector spaces involved.
/(x) = x
Example 6. Consider the space Q[a, b] of all real valued continuous functions
on the interval [a, b], and for each f in Q[a, b] set
Then since A(f) is continuous on [a, b], A can be viewed as a mapping of Q[a, b]
into itself. As such it is linear since
^(aifi + a 2 f2 ) = f [aifi(0 +
Ja
a 2 f2 (0] dt
X X
= f aifi(0 dt + f a 2 f2 (i) dt
Ja Ja
X X
= at [ fi(0 dt + a2
f
f2 (0 ^
Jo Ja
= ai A(ti) + a 2 ^(f2 ).
Example 7. For the same reasons as those just given, the mapping A : Q[a, b] — -> (R
x
defined by
6
^(f) = [ f(^) dx
Ja
is also linear.
imply that D is a linear transformation from 6 1 [a, 6] to C[a, 6]. More generally,
the operation of taking nth derivatives is a linear transformation mapping the
space of «-times continuously differentiable functions on an interval [a, b] into
the space e[a, b].
EXERCISES
2
Prove that each of the following equations defines a linear transformation from (R into
(or onto) itself, and describe the effect of the transformation in geometric terms.
1. A(xi, x 2 ) = — (*i, x 2 ) 2. A(xi,x 2 ) = (2*i, x 2)
3. A(xi,x 2 ) = 2(*i, x 2 /3) 4. A(xi, x 2 ) = 3(xi, x 2)
5. A(xi, x 2 ) = y/2(x\ — x 2 ,x\ + x 2) 6. A(x\, x 2 ) = (x 2 xi) ,
2-2 I
ADDITION AND SCALAR MULTIPLICATION 45
Determine which of the following equations defines a linear transformation on the space
of polynomials (P.
index /, 1 < / < n, let rn be an arbitrary real number. Prove that the function
A:V ^(R 1
defined by
A(aid + • • •
+ «„e„) = «i?7i + + an Vn • • •
forCR 1 .]
vector space V i into the same space D 2 The definition reads as follows: .
This, of course, is just the familiar addition of functions here applied to linear
transformations, and it is an easy matter to show that A + B is again a linear
transformation from V x to V 2 Indeed, if x x and x 2 belong to Vi, and a 1 and a 2
.
46 LINEAR TRANSFORMATIONS AND MATRICES | CHAP. 2
(D 2 + D)(y) = D 2y + Dy.
Example 2. Let K(7) be a fixed function in e[a, b], and let A be the linear trans-
formation mapping e[a, b] into itself given by
Then the sum A -\- I, I the identity transformation on Q[a, b] (see Example 5,
Section 2-1), is the linear transformation mapping e[a, b] into itself whose defining
equation is
X
(A + I)(f) =
Ja
[ K(t)f(t)dt + f.
A + (B + C) = (A + B) + C (2-7)
and
A +B = B + A (2-8)
A + O = O+ A = A (2-9)
We omit the proof that aA is linear, as well as the easy sequence of arguments
required to show that the remaining axioms in the definition of a real vector space
are now satisfied. Granting the truth of these facts, we have
EXERCISES
1. Cite the relevant axiom or definition needed to justify each step in the proof of the
linearity of A + B.
2. Prove that addition of linear transformations is associative and commutative.
3. Prove that the mapping —A defined in (2-10) is linear, and that
A + (-A) = -A + A = O.
5. (a) be a nonempty subset of a vector space V i, and let ($(9C) denote the set
Let 9C
of transformations A from Di to V 2 with the property that A(x) =
all linear for
all x in 9C. Prove that 0t(9C) is a subspace of the space of all linear transformations
*6. Let Vi and V 2 be finite dimensional vector spaces with bases ei, ,e„ and . . .
ei> • • • > ©to, respectively. pair of integers /, j, 1 < / < n, 1 < j < m,
For each
define Aij\ V i — V2 > by first defining A^ on the basis vectors of V i according to the
formula
if k = i,
Aij(e k ) =
\0 if k ^ i,
48 LINEAR TRANSFORMATIONS AND MATRICES CHAP. 2
and then use (2-4) to obtain the value of An for each x in 13 1. (See Exercise 20,
Section 2-1.)
(a) Prove that the Aij are linear transformations from "Ui to V2, and that they are
linearly independent in the vector space of all such transformations.
(b) Prove that the An span the space of linear transformations from V\ to V 2 and ,
hence deduce that this space is finite dimensional with dimension mn. [Hint: Two
linear transformations from Di tol>2 are identical if and only if they coincide on
a basis for Vi.]
FIGURE 2-5
A: v- V 2> B: V< V-
speak of applying B to A(x) to obtain the vector B(A(x)) in V 3 (see Fig. 2-5).
Thus A and B can be combined, or multiplied, to produce a function from V 1
to 1) 3 which will be denoted by BA, and called the product of A and B in that
order, \iz., first A, then B. This is the content of
Definition 2-4. If
C
A: U 1 — V2
> and B: V 2 —> V 3 are linear transforma-
tions, then their product, BA, is the mapping from Vi to T) 3 defined by
the equation
BA(x) = B(A(x)) (2-13)
for all x in V\.
2-3 I
PRODUCTS OF LINEAR TRANSFORMATIONS 49
The essential fact about such products is that they are always linear. Indeed,
ifx x and x 2 belong to "Ui, and a x and a 2 are arbitrary real numbers, then
BA(a 1 x 1 + a 2x 2 ) = 5[^(axXi + a 2 x 2 )]
= B[ ai A{TL{) + a 2 A(x 2 )]
= a 1 B(A(x 1 )) + a 2 B(A(x 2 ))
= axBAixJ + a 2 BA(x 2 ).
might seem more reasonable to denote the product of A and B by AB rather than
BA as above. The explanation for not adopting this notation is quite simple.
Were it used, (2-13) would have to be changed to read AB(x) - B(A(x)) and
the writing of equations would then be an open invitation to error.
Having established the convention that the symbol BA stands for the product
of and B, in that order, we observe that this product is defined only when the
A
image of A is contained in the domain of B. Thus one of the products AB or BA
may exist and the other not, a phenomenon which will reappear later when we
introduce the subject of matrices. But even when both A and B map a given vector
space into itself, in which case AB and BA are linear transformations on the same
space, it is by no means true that they must be equal. A simple example of this dis-
2
turbing fact can be given in by letting A be a counterclockwise rotation of 90°
(R
about the origin, and B a reflection across any line through the origin, say the
x-axis. Then, with e x and e 2 the standard basis vectors, AB(ei) = e 2 while
BA(e{) = — 2 and AB 9^ BA. (Picture?) In short, the multiplication of linear
,
transformations is noncommutative.
The foregoing exampleillustrates one of the ways in which this multiplication
differsfrom "ordinary" multiplication. Why then call it multiplication at all?
The answer is provided by the following identities which show that most of the
properties usually associated with the term multiplication are still valid when
phrased in terms of linear transformations. Specifically, assuming that all of the
indicated products are defined, we have
04 ! + A 2 )B = A B X + A 2 B, A{B 1 + B 2 = AB + AB 2
) X ,
(2-15)
Al = A '
/ the identity map. (2-17)
IA — A,
is associative, the next two that it is distributive over addition, and the fourth that
it commutes with the operation of scalar multiplication. Finally, (2-17) implies
A
that the identity transformation plays the same role in operator multiplication
that the number one plays in arithmetic. The reader should note, however, that
two different identity maps are usually involved here, and, strictly speaking, if
A: Vi — V 2 then (2-17) ought to be written
,
AI\) 1 = A,
/•o 2 v4 = A,
where lVl denotes the identity map on V x IV2 the identity map on V 2 But this
, .
notation is rarely used since the meaning of the unidentified symbol I is always
clear from the context.
The proof of each of the above identities is an easy exercise in the definitions
of the operations involved. Thus to establish (2-14) suppose that C: Vx —» V 2
,
B: V 2 -> V 3 and A: V 3 ,
— V4
> . Then each of the products A(BC) and (AB)C
isa linear transformation from V i to V 4 and to prove their , equality we simply
apply Definition 2-4, twice for each product. This gives
and (2-14) now follows from the equality of the right-hand sides of these expres-
sions.The remaining proofs have been left to the exercises.
on a fixed vector space V (i.e., A: V — V) we can form the product of A with >
itself any finite number of times, thereby obtaining a sequence of linear trans-
Thus
A 1
= A, A 2 = AA, A 3 = AA 2 .... ,
A mA n = A nA m = A m+n
( m \n jrnn
~1
n > 1, then D n and hence D itself, is certainly different from zero, and we
,
have therefore shown that a power of a nonzero linear transformation may be zero.
the fact that the property of being nilpotent actually depends upon the vector
space under consideration as well as the linear transformation involved. For
instance, D is nilpotent of degree n on (Pn , but is not nilpotent on (P. (Why?)
Example 3. Polynomials in A. If A is a linear transformation on a vector space
V, we can use the powers of A together with the operations of addition and scalar
multiplication to form polynomials in A. Thus if
n
p(x) = a + a xx + • • •
+ an x
p(A) = a I + a xA + • • •
+ an A n ,
or
n
p(A) = a + a xA + • • •
+ an A ,
the factor / being understood in the first term of this expression just as x° = 1 is
p(A)(x) = a x + ai^(x) + • • •
+ an A n (x).
the identity p(x)q(x) = q(x)p(x) for "ordinary" polynomials p and q implies that
p(A)q(A) = q(A)p(A). This, in turn, implies that such polynomials can be factored,
for, as the reader will remember, factorization of polynomials depends only on
the commutativity of multiplication and its distributivity over addition.
Example 4. Let e°°[a, b] denote the space of all infinitely differentiable functions
defined on the interval and again let D be differentiation. Then D maps
[a, b],
e°°[a, b] into itself, and we can therefore form polynomials in Z>, which in this
setting are expressions of the type
n- x
an D n + an _ x D + • • •
+ axD + a ,
linear differential operators, and it should be observed that they can also be
interpreted as linear transformations from Q n [a, b] to Q[a, b].) The polynomial
D2 + D — 2 is a typical example, and if y is any function in e x [a, b] (or Q 2 [a, b]),
then
= Tx(al- y) + 2 y
(fx- )
^ dx Zy
dx*
= (D + D -
2
2)y,
A = xD + 1, B = D - x;
that is,
= (** +d(!-*)
— A. (dy. — \ -l.
dy.
—
dx \dx / dx
= [xD 2 + (1 - x 2 )D - 2x]y,
and hence
(xD + 1)(D - x) = xD 2 + (1 - x 2 )D - 2x. (2-18)
2-3 PRODUCTS OF LINEAR TRANSFORMATIONS 53
BA(y) = (D - x)(xQ + y)
TX \ x % + n- x
\
x
tx + y l
d2 v , /-, 2^dy
= [xD 2 + (2 - 2
x )D - x\y,
and hence
(D - x)(xD + 1) = xD 2 + (2 - x 2 )D - x. (2-19)
In view of the examples just given it is clear that the time has come for us to
discuss the general problem of functional notation. All of the functions which
we shall encounter in this text will be elements in one of a number of vector
spaces, and hence should be denoted by such symbols as f or f(x) were we to be
inflexible in our use of bold face type. This, however, would ultimately involve
us in such unsightly (and confusing) expressions as x
11
, sin x, d 2 y/dx 2 fa ,
f(x)g(x) dx.
Such pedantry is pointless, and so we symbols / and f(x) when in
shall use the
our opinion the printed page or its reader would suffer from the use of bold face
type. As a general rule, when we wish to call attention to the fact that a function
is a vector, we shall emphasize it; otherwise, not.
Finally, to settle notational matters once and for all, we comment on our in-
tended use (and mild misuse) of the symbols / and f(x). Strictly speaking, /
should be used to denote a function, and f(x) its value at the point x. But here
again strict adherence to the letter of the law violates the spirit of clarity of expo-
n
sition, for then we would be forced, for example, to write sin and e where , ,
EXERCISES
1. Cite the relevant axiom or definition needed to justify each step in the proof of the
linearity of BA as given in the text.
2. Prove the distributivity formulas (2-15).
54 LINEAR TRANSFORMATIONS AND MATRICES | CHAP. 2
a n (x)D n + a n -i(x)D"-i + • • •
+ ai{x)D + a (x)
5. Write each of the following products in the form a2(x)D 2 + a\(x)D + ao(x).
(a) (xD + 1)2
(b) (2xD + 1)(Z> - 1)
(c) (D - l)(2xD + 1)
7. In each of the following find the result obtained by applying the given polynomial
in D to the indicated functions.
(a) D
2 -
2e x e~ x e x + e~ x
1 ; , ,
2-4 I
THE NULL SPACE AND IMAGE; INVERSES 55
all x in Di such that A(x) = 0. Then, as we have already observed, 31(A) always
contains the zero vector of D x . Actually we can say much more than this, for if
^(xO = A(x 2 ) = 0, then
for all scalars a 1} a 2 and , it follows that 31(A) is a subspace o/Vi. This subspace
is called the null space or kernel of A, and is of fundamental importance in studying
the behavior of A on 1) i.
Of equal importance with the null space of A is its image, 3(A), which, we recall,
is the set of all y in 1) 2 A(x) for some x in 1) x
such that y = . It too is a subspace
this time in V 2— since if yi and y 2 belong to d(A) with y x = A(x 1 ), y2 = A(x 2 ),
then
A(aix x + a 2x 2 ) = ai^(xi) + a 2 A(x 2 ) = atf! + a 2y 2 ,
with the zero polynomial, while its image consists of the zero polynomial and all
polynomials of degree < n — 1
Example 4. Let C 2 (— oo, oo) denote the space of all twice continuously differen-
tiable functions on (— oo, oo), and let A: e 2 (— oo, oo) —» e(— oo, oo) be the linear
transformation D —
2
I. Then
and the null space of A is the set of all functions y in e 2 (— oo, oo) for which
2? - y = a
Thus, 31(A) is the set of solutions of a certain differential equation, and the prob-
lem of finding all solutions of this equation is identical with that of finding the
null space of D2 — I.
Example 5.
00
Let (R be the space of all infinite sequences {x u x 2 x 3 .} of , ,
. .
real numbers, with addition and scalar multiplication defined termwise (see
56 LINEAR TRANSFORMATIONS AND MATRICES CHAP. 2
Exercise 8, Section 1-2), and let A and B be the linear transformations on (R°°
defined by
A{X U X 2 X 3 , , • • •} = {x 2 , X S , X 4 , . . .},
Then 31(A) is the subspace of (R°° consisting of all sequences of the forjn
{x lt 0,0, ...}, with xi arbitrary, while 31(B) = 0. On the other hand, 0(A) = (R°°,
while, by definition, 3(B) consists of all sequences whose first entry is zero.
Now that we have introduced the null space and image of a linear transforma-
tion we propose to take a closer look at those transformations A: °Q\ — V2
> for
which either
(i) 9100 = 0> or (») *(A) = v 2,
or both. The second of these equations asserts that A maps "0 1 onto V2 , and implies
that for each y in V2 there exists at least one xin1)i such that y = A(x). The
first, which says that the null space of A contains only the zero vector, turns out
to be equivalent to the assertion that A is one-to-one in the sense of the following
definition.
FIGURE 2-6
2-4 I THE NULL SPACE AND IMAGE; INVERSES 57
A~ l
(a x y l + a 2y 2) = a x x x + a 2x 2 = «i^ _1 (yi) + a 2 A~ 1 (y 2 ).
* These equations are the vector space analogs of such pairs of statements as
-1 -1 x)
sin (sin x) = x, sin (sin = x, — ir/2 < x < tt/2,
or
x, In (e*) = x,
A~\y) = x,
A~ A =l
IVl , AA- = 1
IV2 ,
(2-21)
where IVl and IV2 denote, respectively, the identity maps on 13 1 and 13 2 .
B(A(x)) = 5(0) = 0,
and on the other,
B(A(x)) = BA(x) = /(x) = x.
and it follows that y is the image under A of the vector B(y) in V i. Thus 3(A) = V2 ,
2
Example 6. If A is any rotation of (R about the origin through an angle 0,
Example 7. Let A : (R
3 — > (R
3
be defined by
A(x u x 2 x z ) ,
= (*i + x 2 ,x 2 ,x 3 ).
2-4 I
THE NULL SPACE AND IMAGE; INVERSES 59
The fact that such transformations do exist can be seen by looking at Example 5
above where
AB{xi, x 2 , x& . . .} = {xi, X2, X3, . .},
and
BA{x x x 2 x 3
, , , . ..} = {0, x2 x3
, , ...}.
Transformations of this sort are encountered fairly often in certain types of prob-
lems and are therefore distinguished by name according to the following definition.
Remark. If B is a right (left) inverse for A, then A is a left (right) inverse for B.
The example given a moment ago shows that a linear transformation may have
a right or left inverse without having an inverse. It is easy to show, however,
A has both a right inverse B
that if and a left inverse C, then A is invertible, and
B = C = A~K For then
AB = I and CA = I,
C(AB) = CI = C, (CA)B = IB = B,
dy
Then
LD(y) = FyXQdt = y(x) - y(a),
Ja
while
EXERCISES
1. Find the null space and, where applicable, the inverse of each of the following linear
transformations on (R 2 .
(a) >4(p) = ^ - 2-
(b) A(p) = xp(x)
(c) A(p) = />(*) - />(0)
where a 1, a 2 /3i, 182 are real numbers. Prove that A is linear, and find a necessary
,
7. Let .4 : 13 1 -> 13 2 be linear, and let W be a subspace of 13 2 . Prove that the set of all x
in 13 1 such that A(x) belongs to W a subspace of
is 13 1.
8. Let A: 13 1 —
be a one-to-one linear transformation, and let ei,
> 13 2 ,e„ be . . .
linearly independent vectors in 13 1. Prove that A(ei), A(e n) are linearly inde- . .
. ,
pendent in D2.
9. Let A : 13 1 — 13 2 be linear, and suppose that
Prove thatA is one-to-one if and only if it is onto. [Hint: See Exercise 8.]
10. Let A and B be invertible linear transformations mapping 13 onto itself. Prove that
AB and BA are also invertible, and that
(AB)- 1 = B-U-\ (BA)- 1 = A-iB-K
invertible.
2-5 |
LINEAR TRANSFORMATIONS AND BASES 61
*12. (a) Let V2 be finite dimensional, and let A: Vi —» 132 be one-to-one. Prove that
A has a left inverse but not a right inverse whenever $(A) 9^ X>2- [Hint: Choose
an appropriate basis in *U 2 .]
(b) Now let *U 1 be finite dimensional, and suppose that A is onto. Prove that A has
a right inverse but no left inverse whenever 31 {A) 9^ 0.
13. A linear transformation P: V — V * is said to be idempotent if and only if P2 = P.
(a) Prove that P is idempotent if and only if 7 — P is.
(b) Prove that 9l(P) = 4(1 - P), and that $(P) = 91(7 - P) whenever P is idem-
potent. [Hint: The image of P consists of all x inV such that P(x) = x.]
(c) Use the results of (b) to show that every x in V can be written uniquely in
the form
X = Xl + X2
x = xiei + • • •
+ xn e n
62 LINEAR TRANSFORMATIONS AND MATRICES |
CHAP. 2
is any vector in V i,
A(x) = XiAieJ + • • •
+ xn A(en ), (2-22)
and it follows that the value of A(x) is completely determined by the vectors
A(e x ), A(en ) in V 2 i.e., A is uniquely determined by its values on a basis
. . . , ;
A V —> V 2
:
x defined by setting
and then using (2-22) to compute the value of A(x) for every x in V i is clearly
linear.Thus Eq. (2-22) also tells us how to construct all linear transformations
from Vi to V 2 and we have proved ,
pair of real numbers ^(ex), A(e 2 ), and can therefore be represented by the ordered
pair (A(e{), A(e 2 j). Since distinct ordered pairs define distinct linear trans-
formations, it follows that there are exactly as many linear transformations from
2 1 2
(R to (R as there are vectors in (R .
Example 2. Let A: (R
2 — <R
2
be the linear transformation defined by
^(ei) = (oci,a 2 ),
A(e 2 ) = 03 1,182),
x = xiei + x 2e 2
2
be any vector in (R . Then
«i 0i 1
«2 fcj
and every such array can be viewed as the definition of a linear transformation
2-5 |
LINEAR TRANSFORMATIONS AND BASES 63
2
A of (R into itself where
that is,
Thus the ordered pair which describes A with respect to the basis e[, e 2 is (1, —2),
and we see that the description of a linear transformation by means of its values
on a basis changes with a change of basis.
EXERCISES
1. Let A:(H 2 — 1
be represented by the ordered pair of real numbers (ai,a 2 ) with
> (R
respect to the standard basis in (R 2 and let B: (R 2 —> (R 1 be represented by (/3i, /J 2 ).
,
there exist a basis ej, e2 in (R 2 with respect to which A is represented by (1, 1). As-
sume that this condition is satisfied, and find ei, e2 .
3. Let A: (R
2 —> (R
2
be represented by the array
T«i j8i'
and 2 -^ 2
let 5: (R (R be represented by
ax pi
> o'
JX2 P 2.
Find the representation of aA, A + B, and AB.
4. Let £((R 2 , (R )
1
denote the vector space of all linear transformations from (R 2 to (R 1 ,
and let
r^csi 2 ,^ 1 )-^^ 2
be defined by
T(A) = (ai,a 2 ),
where («i, a 2) is the ordered pair of real numbers which describes A with respect to
the standard basis in (R 2 .
64 LINEAR TRANSFORMATIONS AND MATRICES [
CHAP. 2
2 1 2
(a) Prove that T is a one-to-one linear transformation mapping £(<R (R ) onto (R , ,
*5. Generalize the technique used in the preceding exercise and show that there exists a
2 2 4 What is the dimension
one-to-one linear transformation mapping £((R (R ) onto (R , .
of£((R 2 ,(R 2 )?
2-6 MATRICES
We have seen that every linear transformation A:Vi-*V 2 can be obtained from
the formula
A(x) = ;M(ei) + • • •
+ xn A(e n ) (2-23)
by suitably choosing the v4(e y ) in V 2 and that (2-23) defines a linear transforma-
,
basis for V u and that x l5 xn are the coordinates of x with respect to this
. . . ,
basis.) We now use this observation to define the notion of a matrix for a linear
transformation, as follows.
Let f i, . . . , fm be a basis for V 2 and let
, A: V i
-+ V2 be given. Then, for each
integer j, 1 < j < n, there exist scalars an such that
^(ey) = 2 «<yf» (2
~24 )
A(e n ) = ai n fi + a 2n f2 + • • •
+ a mn fm .
whose columns are the coefficients of the various equations in (2-25). The reader
should note that the first subscript on an entry in (2-26) indicates the row in which
that entry appears, and the second indicates the column. With this convention in
force the entire array can be abbreviated (a^), it being understood that / and j
range independently over the integers 1, m and 1, n, respectively. When . . . , . . . ,
displayed as above, this set of scalars is called the matrix of A with respect to the
bases (S> 1 = {e 1? e n} and (S> 2 =
. {fi,
. .
f
,m} and is denoted by [A: (R lt (B 2 ], . . . ,
2-6 MATRICES 65
or simply by [A]. (In the special case where A maps V into itself and (Bi = = ®,
<B 2
the notation [A: (B] is also used.) When m = n, we say that (2-26) is a square
matrix; otherwise, rectangular. In general, a matrix consisting of m rows and n
columns will be referred to as an m X «-matrix (read "m by «").
The argument just given shows that every linear transformation from V x to V 2
determines a unique m X rc-matrix with respect to <Bi and (S> 2 But since the an .
in (2-25) uniquely determine the A(ej) and hence, by (2-23), A(x) for all x in V u
it follows that euery m X «-matrix determines a unique linear transformation from
It is important to realize that this theorem does not assert that every linear
transformation has a unique matrix. Indeed, any such assertion would be patently
false, for, as we have already seen, the matrix of a linear transformation can
change with a change of basis (Example 3, Section 2-5). Thus the several references
to bases which appear in Theorem 2-6 cannot under any circumstances be deleted.
2
Example 1. Let e x and e 2 be the standard basis vectors in (R , and let
A: (R
2
—> (R
2
denote the reflection across the ei-axis. Then
A(e{) = 1 •
ex + • e 2, A(e 2 ) = •
ei - 1 e2 ,
Lo -i
Example 2. Let A: (R
2 —> (R
2
be the (counterclockwise) rotation about the
origin through an angle 6, and again let
cos — sin
in0"|
sin OS0J
cos
66 LINEAR TRANSFORMATIONS AND MATRICES CHAP. 2
with ones along its principal diagonal and zeros elsewhere. For obvious reasons
this matrix is called the n X n-identity matrix.
Similarly the matrix of the zero transformation from 13 1 to V2 is always the
the m X n-zero matrix
0-- •
•• •
•• •
Example 4. Let
D:(Pn --><Pn
be differentiation, and let
(B = {1, x, . ..,X,n— It
~
D(x) = 1-1 + 0-X+ h • x nn—22 + i r\ .
• x„n—
n~J
D(x 2 ) = 0-l + 2-x+ h • jc + • x,71-1
~l = n~z
D(x n ) 0-l+0-x+ V in - \)x + • x,n— 1
and
1
•••
2 •••
[D: ffi] =
••• n - 1
•••
Example 5. Let
A: (P 3 ^ (?b
for a\lp(x) in (P 3 , and let (B x and (B 2 be the standard bases in (P 3 and (P 5 , respec-
tively. Then
,4(1) = 2x
2
- 3 = -3 • 1 + • x + 2 + •
*2 •
+
x3 • x
4
,
A(x) = 2x
3
- 3x = • 1 - 3 •
x + 2
x + 2• •
x +
3
• x
4
,
2 4
A(x ) = 2x - 3x
2
= • 1 + • x - 3 x2 + • •x3 + 2 •
x
4
,
and
3
-3
[^: (Bi, (B 2 ] = 2 -3
2
2_
EXERCISES
In Exercises 1 through 6 find the matrix of the given linear transformation with respect
to each of the given pairs of bases.
(0,1,0)
(b) ($>i the standard basis, (B 2 (1,0,1)
(0,0,1)
(0,1,0) 0,0,0)
f
(c) (Bi = {(1,0, 1), (B 2 (1,1,0)
(0,1,1) (1,1,1)
2. A: (R
3 -> (R
2
; 4(jci, x 2 x 3)
,
= (*i - jc 2 , 2x 2 - 3x 3 )
(a) (Bi and <B 2 the standard bases
(0,1,0)
(b) (Bi = j(l,0,l) (B 2 = the standard basis
1(0, 1, 1)
(0,1,1)
(3,1)
(c) (Bi ={(2,-1,-1), (B 2 =
d,f)
((3, 2, J)
Find the value of A(x) for any x in *Ui, given that A: 13 1 -» 13 2 is linear, and
[,4:(B *2] = (&/).
(a) Every linear transformation A 13 -> 13 has at least two invariant subspaces.
:
an "
• '
«lm ai,m+i • "
ai„
U: =
(B]
•
• a m +i, m +i ' a m-\-\,n
•
• •
a n TO +i
,
OCnn
an • •
ai m
am i • Oimm
[A:(S>] =
•
•• am + 1.TO+1 '
• a m +i, n
•
•• a„ ,m+l CX-nn
2-7 | ADDITION AND SCALAR MULTIPLICATION OF MATRICES 69
let3H wn denote the set of all m X ^-matrices. Then if dim 13 x = n and dim 13 2 = m,
Theorem 2-6 asserts that the function which associates each A in £(13 1, 13 2 )
with its matrix [A] with respect to a fixed pair of bases ($>i and (B 2 is a one-to-one
mapping of £(13 1} 13 2 ) onto 9TC mn . This simple fact allows us to translate algebraic
statements concerning linear transformations into statements concerning matrices,
and leads to the subject of matrix algebra. In particular, it allows us to convert
2fE wn into a real vector space by using the matrix analogs of the addition and
i=l i=\
and it follows that
(A + B)(ej) = Ate) + B(ej)
i=i t=i
m
and
((rA)(ej) = a
= ^ (<Tan)U-
Hence
M+ £] = (a,-y + /3,7 ),
[<rA] = (o-a.-y),
nition the m X n-matrix (a^ + ft,-); the product (T(a ) of a real number a i}
for all m X n-matrices (a,-,-), 0,,), and all real numbers (T.
and assert that matrix addition and scalar multiplication are performed entry by
entry, or termwise. Moreover, we now have
Theorem 2-7. The set 2fTl wn of all m X n-matrices is a real vector space
under the above definition of addition and scalar multiplication.
The student should appreciate that there is no need for a formal proof at this
point since the asserted result follows automatically from Theorem 2-6, the fact
that £(V U V 2 ) is a real vector space, and the way in which addition and scalar
multiplication were defined in 3Tl mn . Indeed, we can now assert that £(V U V 2)
and 3Tl wn are algebraically identical (or isomorphic), and that the function which
sends each linear transformation A: Vi — V2 > onto its matrix [A: <R lt (B 2 ] with
respect to a fixed pair of bases (Bi and (B 2 is an isomorphism of £(13 1, V 2) onto
Jl'-mn-
2-7 ADDITION AND SCALAR MULTIPLICATION OF MATRICES 71
As an illustration of the way in which this fact can be used to establish results
which are not otherwise obvious, we now propose to show that <£(T>i, V 2 ) is finite
dimensional and to compute its dimension. For this purpose we introduce the
special matrices (e^), 1 < i < m, 1 < j < n, each of which has the entry 1 at
the intersection of the rth row andyth column and zeros elsewhere:
= (en)- (2-29)
J
+ "21(^21) + + «2n(e2n)
+ «m i(ew i) + • • •
+ amn (emn ),
or
m,n
(2-30)
i,J=l
Thus the (e^) span M, mn , and since it is clear that (2-30) is the only possible way
of writing (o#) as a linear combination of the (e,-,-), it follows that these matrices
are a basis for 9!l wn . (This particular basis is called the standard basis for 9fTl mn .)
Hence £ (V 1 ,V 2 )
I is also finite dimensional with dimension mn, and we have
proved
Theorem 2-8. IfVi and V 2 are finite dimensional vector spaces, then so
is £(Vi, V 2 ), and
dim £(V 1, V 2) = (dim V i)(dim V 2 ). (2-3 1)
(e ln ) = (0,0,..., 1)
as a basis. In this case the (e -y) can be identified with the standard basis vectors
t
in (R
w
and when this identification is made 3Tli n becomes identical with (R n
, .
that there are exactly as many linear transformations from V to (ft as there are
1
w
vectors in <ft (cf. Example 1, Section 2-5).
71 LINEAR TRANSFORMATIONS AND MATRICES CHAP. 2
Example 2. Let
V* = Vo = (R'
2
Then dim £((R.
2
, (R ) = 4, and the (et-y) are four in number:
ri o] ro ii
(en) = («i2) =
0. LP oj
[0 o] ol
(«2l) = («22> = r°
_1 0. [p ij
/, A, A2 , A.
also map V into and thus belong to £(V, V). But by Theorem 2-8 this set
itself,
is linearly dependent in£(V, V). Hence there exists a smallest positive integer k
k~ x
such that A k is linearly dependent on I, A, ... A and it is now easy to show , ,
that these transformations are a basis for the subspace of £(V, V) spanned by
k
the powers of A (see Exercise 14). In particular, we can write A in the form
A = a k _iA + + axA + a I,
or
A — ak-iA axA — a = O, (2-32)
of lower degree having A as a root. For this reason m^(*) is called the minimum
polynomial of A. can be characterized as the polynomial of least degree with
It
2
leading coefficient 1 which has A as a root, and is clearly of degree <n when
dim V = n. Actually, it can be shown that the degree of m A (x) does not exceed
the dimension of V for any nonzero transformation A: V —» V. The proof,
however, is not easy.
2-7 ADDITION AND SCALAR MULTIPLICATION OF MATRICES 73
EXERCISES
In Exercises 1 through 5 compute the value of a[A] + 0[B] for the given scalars a, /3
1. a = 2, j8 = - 1,
1 -2 4 -1 2
2 -2 3 4 -2.
2.a = -*. = 4,
"2 -3~ -1 2"
[41
= 1 • [B] = 3 1
.4 -2. 4 -5_
3. a = §, = -2,
6 2 -3 -8 3
2
l"
[4] = ' [5] =
_1 4.
_A3 3J
4. a = hP = l
6 2
-3" -1
2 2
[A] = 4 1 » [5] = —2 1 —2 i
--1 1
3 3. -f 2
5. a = hP- ~h
5"
* -2 3 3 -4 1
[A] = § -1 -i <
3 . [B] = 6 — z2 A
2
.-3 1
L 4 i
2
2 3J L.3 3
2
6. Let /I denote the counterclockwise rotation of (R about the origin through the
angle w/4, and let B denote the reflection across the origin. Find the matrix of
A + B with respect to the standard basis ei = (1, 0), e2 = (0, 1), and with respect
to the basis ei = (1, 1), e'2 = (-1, 1).
Find the matrix of 2A — B with respect to the standard bases in (P3 and (P4, and
with respect to the bases
(Bi = {1, x - 1, (x - l) },
2
(B 2 = {l,x- 1, (x - l)
2
,
(x - 3
l) }.
9. Let ei, . .
.
, e„ and f i, . .
.
, fm be bases for 13 i and 13 2, respectively, and for each pair
of integers i,j with 1 < / < m, 1 < j < n, let E i} :Vi — > 13 2 be defined by
= if j ^ k,
Eij(e k )
[fi if j = k.
an «12 an ai2
Prove that A is linear, and find the matrix of A with respect to the standard basis
in 91X22.
an ai2 an ai2
12. What is the dimension of the vector space £(9TCm i,9Ei„)? Of £(9TCOT „,9re pg)?
2 2
13. (a) Prove that the functions sin x, cos x, sin x cos x, sin x, cos x are linearly
independent in C(— 00 , 00 ).
(b) Let V denote the subspace of C(— <» , 00 ) spanned by the functions in (a). Prove
that Dn , the «th power of the differentiation operator, maps V into itself for all n,
and find the matrix of D 2 — 2D + 1 with respect to the given basis for V.
14. Let A: V — V > be linear, with A
O, and let k be the smallest positive integer such
?*
15. Find the minimum polynomial of the linear transformation D: (P3 —> (P3.
be given, let
fc=i
t=l
2-8 MATRIX MULTIPLICATION 75
Then we have
ABiej) = A
(
\k=i
2 Mk) = 2 ^jA(h) / *;=i
fc=l v=l /
=
2 (2 a *'*^*y) 8»»
and the matrix of yi5 with respect to (Bi, (B 3 is the mX /"-matrix whose ijth entry
is£jk=i OtijSiky. But since
[A: « 2 «3 =
, ] (o«), 1 < i < m, 1 < k < n,
and
[5: (Bi, CB 2 ] = (fly), 1 < k < n, 1 < j < r,
It isimportant to notice that the product of two matrices is defined only when
the number of columns in the first matrix is equal to the number of rows in the second;
a restriction which is the matrix analog of the fact that the product of two linear
transformations is defined only when the image of the first is contained in the
domain of the second. When written in greater detail, Eq. (2-34) becomes
0=11011 + • • *
+ ttln0nl ail01r + ' • '
+ «ln0nr
OC21011 + * * *
+ «2n0nl «2101r + * * '
+ «2n0nr
Example 1. If
3"
1
'2-1
(<*ik)
= (&;) = -2 1
1 2 -3
4,
then (ancXPhj) is defined, and we have
=
"2-1 + (-l)(-2) + 0-0 2-3 + (-1)- 1 + 0-4
(aikXPkj)
.1-1 + 2 -(-2)+ (-3)-0 1 -3 + 2- 1 + (—3) •
4J
4 5
-3 -7
On the other hand, (0kj)(aik) is not defined since the number of columns in (j8 fc y)
Example 2. Let
3 1 ~-l 2
(«ii)
= (fty) =
-1 2 1 0_
Then
3 1 -1 2 -2 6
((*ij)(Pij)
=
-1 2. 1 0. 3 -2
-1 2"
3 1 -5 3
(Pij)(<Xij) =
1 -1 2 3 1
and are simply the powers of A with respect to (B. (Why?) This in turn implies
that the matrix of a polynomial
p{A) = ak A
k
+ ak-iA*-
1
+ h a xA + a I
in A is
Such expressions are called polynomials in [A], and are defined whenever [A]
is a square matrix. In particular, if
rriA(x) = xk — a k -ix
k 1
a\x a
and it follows that [A] is a root of m A (x). For obvious reasons this polynomial
is also called the minimum polynomial of the matrix [A].
whenever the sums and products appearing in these equations are defined.
Similarly, the identity matrices introduced in Example 3 of Section 2-6 play the
same role in matrix multiplication that the identity transformations play in operator
multiplication. And finally, we can use these identity matrices to define right,
left, and two-sided inverses for matrices by rewriting Eq. (2-21) and Definition 2-6
in matrix terms. The details have been left to the reader.
EXERCISES
(a) 1 -3 2 6 l"
4 1 -1 -2 3
2 -5 3. 3 -4.
4 5 3-2 1
.-10 2 4 L5 -1 3_
(c) a
3
i n
u £2 _1* _l"
2
6
J, 1 _J, -1 2 1
nil
6
L"
2
6
6
3 2 l 2-i
4 4
(d) [2 i -1 3] (e) [2 -1 3]
2 2
-1 -1
78 LINEAR TRANSFORMATIONS AND MATRICES CHAP. 2
an ai2 an a2i
and
«21 «22j a 12 a22.
to commute.
-1
4. An n X «-matrix [A] is said to be invertible with inverse [/4] if and only if
(a) Show that if [A] has a right inverse [B] and a left inverse [C] then
1 1
1 1.
has no left inverse in 9TC32, but has infinitely many right inverses in 3H32.
1 1 -2 1
1 -f
1
a
2 and when this condition is
is invertible if and only if jS ?* a , find its inverse satisfied.
7. Find all values of a and /3 for which each of the following matrices is invertible, and
compute their inverses. ,
"
2 2
fi 2_ a fl
p _
-1
(a) [A] - (M
_1 1.
2-8 I MATRIX MULTIPLICATION 79
1 -1 1 a 1
7 5
such that
2
a (3 1 0~
y *_ .0 1-
10. An n X n-matrix (an) is said to be nilpotent if and only if there exists an integer
k > such that (a,,)* is the n X w-zero matrix. The smallest positive integer for
which this is true is called the degree ofnilpotence of (a,,). Show that each of the
following matrices is nilpotent, and find their degree of nilpotence.
4 .-2 -1 -3. 1 -3 -4
11. Find all 2 X 2 nilpotent matrices with degree of nilpotence two. (See Exercise 10.)
12. Determine which of the following matrices are roots of the polynomial p(x) =
x3 — x2 + x — 1.
(a) TO 0] (b) 1
1 1
.1 0.
1 1
.1 -1 1 -1
13. Let A: 9E 2 2 -> 9H22 be defined by
"2
an a\2 l" an ai2
A .
Prove that A is linear, and find its matrix with respect to the standard basis in 9TC22.
an ai2 an ai2 1 -1
15. Solve each of the following matrix equations for [X], given that
2 1 -1 1 1
[A] = [B] = [C] = [O]
3 2 1 .-1 1.
16. Prove the associativity of matrix multiplication directly from Definition 2-8.
17. Find 2 X 2-matrices [A], [B] such that
inverse?
(c) Let T: (R 1 -» 3Tl nn be defined by
a •• •
n«) = a
for all a in (R 1 . Prove that T is one-to-one, linear, and that T(ci)T(0) = r(a/8) =
T(0)T(a) for all a, 0.
(d) Find the matrix of T with respect to the standard bases in (R 1 and 2fTC n „.
20. Prove that the only n X ^-matrices which commute with every matrix in 9TC nn are
the scalar matrices
" •••
a
Such equations are known under the generic name of operator equations, and will
appear throughout this book in a variety of forms. In general, of course, the
technique for solving a particular operator equation depends upon the operator
2-9 OPERATOR EQUATIONS 81
involved, and also upon the underlying vector spaces. Nevertheless there are a
number of such equations which can be proved without using
facts concerning
anything other than the linearity of A, and we propose to get them on record here
before going on to more specialized topics.
A vector x in Vi is said to be a solution of (2-37) if A(x ) = y. The totality
of such vectors is called the solution set of the equation. In the special case of a
homogeneous equation
Ax = (2-38)
whose right-hand side is zero, we know that this set is a subspace of 13 1. It is called
the solution space of the equation. One of the most important properties of
operator equations is that the problem of solving a nonhomogeneous equation
Ax = y,y 5* 0, can all but be reduced to that of solving its associated homogeneous
equation Ax = 0. In fact, if x p is a fixed solution of Ax = y, and if x^ is any
solution whatever of Ax = 0, then x p + Xh is also a solution of Ax = y, since
of Ax = y, and add. The reader will do well to keep both of these points of view
in mind as we continue.
Ay= 1,
a*y
dx*
-y=h (2-39)
and its solution set consists of all functions in C 2 (— oo, oo) which satisfy this
equation on the entire real line. In this case it is obvious that y = —1 is one such
function. Thus, to complete the solution of (2-39) it suffices to find all solutions
of the homogeneous equation
dx*
y
In Chapter 3 we will prove that the solution space of this equation has the functions
x
e and e~x as a basis, and hence as a corollary, that the solution set of (2-39) is
2
the totality of functions in C (— oo, oo) of the form
y = -1 + c xe
x
+ c 2 e~
x
,
&[x 1 +a2 X2 = P
<*\X\-\- 012X2 =Q
FIGURE 2-10
2 1
Example 2. Let A be a linear transformation from (R to (R , let e x and e 2 be
the standard basis vectors in (R
2
, and let A(ei) = ct\, A(e 2 ) = a2 , ai and a 2
real numbers. Then if x = x^! + x 2e 2 is any vector in (R
2
,
a i*i + a 2x 2 = 0. (2-40)
2
Since (2-40) is the equation of the line through the origin in (R with slope —ai/a 2 ,
2
the solution space of the equation Ax = is just the set of points in (R which
comprise that line. In this case the solution set of the nonhomogeneous equation
Ax = j8, j8 a real number, can be interpreted as a translation of the line described
by (2-40), as shown in Fig. 2-10.
tions. This is the so-called existence problem for operator equations, and theorems
which establish such conditions are called existence theorems.
Of equal, or even greater importance is the problem of ascertaining when Ax = y
admits at most one solution for any given y in V 2 This problem is known as the .
uniqueness problem for operator equations, and can always be answered by examin-
ing the homogeneous equation Ax = and using the following theorem.
The student should have no difficulty in convincing himself that this result is an
immediate consequence of Theorem 2-9 and the description of the solution set
of Ax = y given there.
In the case where A admits an inverse of one of the various types discussed in
Section 2-4, the equation Ax = y can be immediately solved. If, for instance,
A is invertible, then from Ax = y we deduce that
A~\Ax) = A~ l
y,
or
x = A~ x y,
and the solution (which in this case must be unique) has been described in terms
of A~ l
. Similarly, if B is either a right or left inverse for A we find that the solu-
tion set of Ax = y is the set of all x in Vi such that x = By. This technique for
solving an operator equation is known as inverting the operator, and is used
whenever an explicit formula for an inverse can be deduced from the definition
of A.
Example 3. Systems of Linear Equations. As our final example we apply the above
ideas to the study of systems of linear equations. Our motive for presenting this
somewhat extended example is twofold. First, the theory of linear equations is
;
important in its own right, and the results we are about to obtain will be needed
from time to time in our later work. Second, this material provides perhaps the
easiest application of linear transformations to the solution of a nontrivial prob-
lem, and should therefore help the student become familiar with such trans-
formations.
We begin by introducing some standard terminology. A system of m linear
equations in the n unknowns x x x 2 , , . . . , xn is a set of equations of the form
0=11*1 + «12*2 + * * *
+ OCi n Xn = 0i,
in which the an and & are real numbers. The an are called the coefficients of the
system, and have been so indexed that the first subscript on any coefficient indi-
cates the equation in which it appears, and the second the unknown with which
it is associated. Such a system is said to be homogeneous if all of the &• are zero
nonhomogeneous otherwise. Finally, a solution of (2-41) is an »-tuple of real
numbers (c x c 2 c„) with the property that when Ci is substituted for x lt
, ,
. .
.
,
c 2 for x 2 etc., each of the equations in the system becomes an identity. A system
,
M= «21 «22 *
* * 0C 2n
«mi ot m2 • • - amn _
n
be the m X H-matrix formed from the coefficients of (2-41), and let A: (R —* (R
m
be the linear transformation defined by this matrix relative to the standard bases.
n
Then if x denotes the vector (x lt . . . , xn ) in (R , and y the vector (0i, ... ,
|S W)
in (R
w (2-41) can be rewritten in operator form as
,
Ax = y. (2-42)
Furthermore, if
Ax = xtfi + x 2y 2 + •
*
' + xn yn
= n
for each x (x lt . . . , xn ) in (R . Thus Ax is a linear combination of y i, . . . , y„,
and it follows that these vectors span the image of A. This, combined with the
observation made a moment ago, yields our first existence theorem.
Theorem 2-11. The system of linear equations (2-41) has a solution if and
only if the vector y = (p u
m is linearly dependent on the
j3 OT ) in (R . . .
,
To answer the uniqueness problem for (2-41) we pass to the associated homoge-
neous system Ax = and apply Theorem 2-10.
0, We again use the fact that if
Ac = ctfi + • • •
+ cn yn .
ciyi + • • •
+ cn y„ = 0.
Thus Ax = has a nontrivial solution (i.e., a solution in which at least one of the
^ and only
m and we
C{ 0) if if the vectors y i, . . . , yn are linearly dependent in (R ,
have
Theorem 2-12. The system of linear equations (2-41) has a unique solution
{provided it has any solutions at all) if and only if the vectors formed from
m
the columns of the system are linearly independent in (R .
Corollary 2-1. A
homogeneous system of linear equations has nontrivial
solutions whenever the number of unknowns exceeds the number of equations.
]
We have already had occasion to remark that the operator D which maps a
differentiable functiononto its derivative is a linear transformation (see Sec-
tion 2-1). The same is true of polynomials in D, and of even more complicated
expressions such as xD
2
+
D + x. Linear transformations of this sort which
involve D and its powers are called linear differential operators. The study of such
operators leads naturally to the theory of linear differential equations, the subject
matter of this chapter and the chapters which follow.
To meaning to the term "linear differential operator," let I be
give a precise
an arbitrary and for each non-negative integer n, let e n (I)
interval of the real line,
denote the vector space of all real-valued functions which have a continuous «th
derivative everywhere in I. [Recall that the vectors in Q n (I) are real-valued func-
tions whose first n derivatives exist and are continuous throughout I, and that
vector addition and scalar multiplication in this space are defined by the equations
L = an (x)D n + an.xOc)/^-
1
+ • •
+ a x (x)D + a (x), (3-1)
Thus the image of a function / in e n (/) under the linear differential operator
Ly = an (x)y (n) + • •
+ mix)/ + a (x)y, (3-3)
where y', . .
. , yn derivatives of the function y = fix). Strictly
(n)
are the first
speaking, the left-hand side of (3-2) is the value of Lfat the point x, and a scrupu-
lous regard for accuracy would require that it be written (Lif))ix). For obvious
reasons the extra parentheses are almost always omitted. Moreover, we shall
fix)," thereby following the familiar custom of confusing a function with its
value at a point. This, of course, is just a linguistic convenience, and once under-
stood as such causes no difficulty.
just the identity transformation, and, in general, Dn can be viewed as the nth
power of the linear transformation D (see Section 2-3).
Example 2. Any polynomial in D of degree n, with real coefficients, is a linear
differential operator of order n on every interval of the real line.
L = a ix), (3-4)
where a ix) is continuous and not identically zero on /. Thus if/ is any function
in e(7),
Lfix) = a ix)fix),
xD 2 + 3Vx D - 1
*
Note that the expression aoix)fix) actually admits three different interpretations.
It can be viewed as the product of the functions aoix) and fix), or as the value of the
operator aoix) applied to the function fix), or as the product of the operators aoix) and
fix). The particular interpretation chosen, however, is usually a matter of indifference.
88 GENERAL THEORY OF LINEAR DIFFERENTIAL EQUATIONS J
CHAP. 3
(x + |jcj)Z>
2
- VxTT D + In (x + 1)
is of order 2 on (—1, 1), but of order 1 on the subinterval (—1,0] since x + |jc|
vanishes identically there. Thus the order of a linear differential operator may
depend upon the interval in which it is being considered, as well as on the algebraic
form of the operator itself.
= xD(2xy' + y) + 2(2*/ + y)
= x(2xy" + 3/) + 4xy' + 2y
= 2x 2y" + 7jc/ + 2j.
EXERCISES
The intervals [a, b) and (a, b] are defined, respectively, by the inequalities a < x < b,
*
a x < b. The first is said to be open on the right and closed on the left, the second
<
closed on the right and open on the left.
3-1 | LINEAR DIFFERENTIAL OPERATORS 89
2 2
(c) (4x D 2 + 4xD + 4x + ^^sinx
3. Find constants a, b, c such that a + b -f c = 1, and
4. Write each of the following linear differential operators in the standard form
(c) (xD + D) 22 2
(d) £> (xZ) - \)D
(e) D(Z)e* + 1) + e x
and thus deduce that the order of the product of two such operators need not be
the sum of the orders of the factors.
(b) Give an example to show that the product of two linear differential operators
on an interval I need not be defined on the same interval.
8. Prove that Dm (a(x)D n) is a linear differential operator of order m + n by expressing
thisproduct in standard form as a "polynomial" in D. [Assume the existence and
continuity of all the necessary derivatives of a(x).]
9. Find the sum Li + L2 of each of the following pairs of linear differential operators.
(a) Li = 2xD + 3, L 2 = xD - 1
(b) Lx = ex D 2 + A L 2 = e~ x D 2 - D
(c) L\ = xD + 1, L 2 = Dx
10. Prove that the sum of two linear differential operators defined on an interval / is
(b) Use (a) and the general distributivity formula for linear transformations that
was established in Section 2-3 to prove that the multiplication of constant coeffi-
cient linear differential operators is commutative. Deduce from this that the product
of two such operators can be obtained by treating them as ordinary polynomials
in D and using the usual rules of elementary algebra.
13. Factor each of the following linear differential operators into a product of irreducible
factors of lower order.
(c) AD 2 + AD + 1 (g) D* + 1
(d) Z) 3 - 3D 2 + 4 (h) D5 - 1
n n -k
Dn [f(x)g(x)] = £
fc=0
(k
)(D f(x))(D g(x)),
k
where
n(« - !)••• (n - k + 1)
o- k\(n - k)l k\
16. Use the result of the preceding exercise to express each of the following linear
differential operators in the form an (x)D n • • a\(x)D ao(x). + •
+ +
(a) D 3 (xD)
(b) Dm (xD)
(c) D 5 (xD 2 + ex)
17. Prove that for any pair of non-negative integers k and m,
k —m
! X m < k,
^**-r («)
lo, m > k.
"18. (a) Prove that
m m k
- — m+
(x D )x = k(k 1) •(k l)x
(a2X
2
D2 + a\xD + ao)x = [a2k(k — 1) + a\k + ad\x
(a) Compute the value of Lx k k an arbitrary real number, when Lis equidimensional.
,
(b) Prove that (xm Dm)(x n D n) = (xn D n)(xm Dm) for any pair of non-negative inte-
gers m, n, and hence deduce that the multiplication of equidimensional operators is
equation identically on /.
Thus an nth order linear differential equation is simply an equation of the
form
which is homogeneous, normal, and of order 2 on (— oo, oo) or any of its sub-
intervals, and
X ^ '
dX 3 dx
which is nonhomogeneous, normal, and of order 3 on (0, oo) and (— oo, 0), but
is non-normal on any interval containing the origin.
The primary objective in the study of linear differential equations is to find all
solutions of any given equation on an interval /. As might be expected, this is a
difficult problem, and a complete answer is known only for certain special types
independent solutions of (3-7), then every solution of that equation must be of the
form
y(x) = Cij>i(x) + • ' •
+ cnyn {x) (3-8)
for suitable real numbers a. Conversely, every function of this type is certainly
a solution of (3-7) whenever y x {x), yn (x) are, and for this reason (3-8),
. . . ,
with the Ci arbitrary, is called the general solution of (3-7). Finally, any function
obtained from the general solution by assigning definite values to the c» is called
a particular solution. We leave the reader to reflect upon the merits and short-
comings of this somewhat unfortunate choice of terminology.
By a familiar line of reasoning, these results are also pertinent to the study of
nonhomogeneous equations. Indeed, in Section 2-9 we saw that if yp (x) is any
solution of the nonhomogeneous equation
Ly = h(x) (3-9)
Example 1. The functions sin x and cos x are easily seen to be solutions of the
second-order equation
y" + y = (3-10)
on the interval (— oo, oo). Moreover, these functions are linearly independent
3-2 | LINEAR DIFFERENTIAL EQUATIONS 93
in 6 2 (— oo, oo ), since
Ci sin x + c 2 cos x =
where c x and c 2 are arbitrary constants. The reader should note that without a
theorem such as the one cited above there would be no guarantee whatever that
(3-1 1) includes every solution of the given equation.
y = x + Ci sin x + c 2 cos x.
Before leaving this section it may be instructive to compare the solution set
of a nonlinear differential equation with that of a linear equation. To this end we
consider
/ _ 3^2/3 = 0j (3
_ 13)
y = (x + cf (3-14)
as its "general" solution on the interval (— oo, oo). (See Fig. 3-1.) In particular,
the functions x
3
and (x + l)
3
But their sum is not, and
are solutions of (3-13).
hence the solution set of this equation is not a subspace of e x (— oo, oo), even
though the equation appears to be homogeneous. Moreover, all of the various
solutions obtained from (3-14) by assigning different values to c are linearly
independent in e 2 (— oo oo), and we conclude that a. first-order nonlinear differential
,
equation can actually have infinitely many linearly independent solutions. Finally,
(3-13) also admits an infinite number of from solutions which cannot be obtained
(x + c)
3
by specializing the constant somewhat peculiar solutions c. All of these
have the property that they are zero along an interval of the x-axis and are of the
following three forms
*
- fl)3 ' x < a, o, x < b,
yv = !
(
y ~ kx - bf, x >
0, x > a, b,
\x - af, x < a,
o, a < x < b,
(x - bf, x > b.
94 GENERAL THEORY OF LINEAR DIFFERENTIAL EQUATIONS |
CHAP. 3
y y
FIGURE 3-1
Thus, to use the term "general solution" in reference to (3-14) is in this case a
genuine misnomer. In short, every single one of the properties enjoyed by the
—
solution set of a linear differential equation fails to hold here a fact which, if it
does nothing should convince the student that linear differential equations
else,
EXERCISES
1. Determine the order of each of the following linear differential equations on the
indicated intervals,
(a) xy" - + l)y = 3, op (-oo, ») (b) (D + l) 3 = 0, on (0, 1)
(2x .y
(x + \x\)y'" + (sinjc)/ = 2e
x
(c) , on (-1, 1); on (0, °o)
(d) y/xy" — 2/ + (sinx)^ = lnx, on (1, »)
(e) (x + 1 + \x + 1|)/" + (x + \x\)y' + 2y = 0, on (-», oo); on (0, oo);
on (-1,0)
3-3 I
FIRST-ORDER EQUATIONS 95
2. In each of the following show that the given function is a solution of the associated
linear differential equation, and find the interval (or intervals) in which this is the
case.
2y" - xy'
(d) x + y = 1; 1 + 2*lnx
(1 - xV ~ 2xy' + 2y = 2; xtanh" x
1
(e)
3. (a) Show that e01 cos fo: and e? x sin Ax are linearly independent solutions of the
equation
(Z)
2 - 2aD + a2 + £ 2)y - 0, 6^0,
on (—oo, oo ).
(D - a) 2y = 0.
(b) Find the particular solution of this equation which satisfies the "initial" condi-
tions ^(0) = l,/(0) = 2.
5. (a) Verify that sin
3
x and sin — ^ sin 3x are solutions
x of
2 3 3
(b) x y" - 5xy' + 9y = 0; 2x In x, x
3
, x (2 In x - 1), on (0, °o
(c) y" + Ay = 0; sin 2x, —2 cos 2x, —cos (2x — 3), on (— oo, oo)
Let
fliW ^ + a (x)y = h(x) (3-15)
in the form
y = yP (x ) + yh(x), (3-16)
where yp (x) is any "particular" solution, and yh(x) is the general solution of the
homogeneous equation
1 dy a (x)
y * 0,
y dx a\(x)
or
11 _ —)\atf.x)\a x {x)\dx
Hence, by the theorem cited in the preceding section, the general solution of (3-17)
is
,, _ ro —I[aoW/ai(,x)]dx
dy a (x) h(x)
n 1R
.
(3_18)
Tx+ ^(x) y ^x)'
y g (3-19)
\dx fli(x) ) fli(x)
But
—
dx
°* ft, J"k»o(*)/«i(*>]*«^
J
_
~
(dy.
\<£t
+
_|_
a °(X ^
ai (x)
yi,l
,Jl<*o<*>/«i(*>]<**
rfx )
d_ / f[a (x)lai(x)]dx^ _ Kx )
e
Jla (x)/ ai (x)]dx
dx
Thus
_ —J[o (a;)/oi(a;)]da; h(-X) J\a {x)la l (x)\dx ,
/
3-3 | FIRST-ORDER EQUATIONS 97
y = J
h\X) f[a (x)/o 1 (a;)]da;
^ Q—
f[a (x)/ai(x)]dx
(3~20)
c an arbitrary constant.
Considering the simplicity of the technique underlying this result, it is not
recommended that the student memory. Instead he should re-
commit (3-20) to
member the general method, which can be described as follows: To find the general
solution of a normal first-order linear differential equation, rewrite the equation in
the form
dy ao(x) h(x)
,
= 3
dx a x (x) a\(x)
multiply by e
ha ° (x)la ^ x)]dx and integrate.
,
% + **-'
2xdx = x
In this case we multiply the equation by e$ e to obtain
X2
= xe .
Thus
X2 X2
ye = I
J
xe dx
I
+ c,
xfx + y = x. (3-21)
Since the leading coefficient of this equation vanishes when x = 0, the above
method applies only on the intervals (0, oo) and (— oo,0). There, however,
(3-21) may be rewritten
| + ^-l. 0-22)
e
fdxlx = e
\n\x\ = 1^
98 GENERAL THEORY OF LINEAR DIFFERENTIAL EQUATIONS | CHAP. 3
j>|jc| = / \x\dx + c
x2
y+ C, X > 0,
x2
- y+ C, X < 0.
Thus
xr + * >0 -
y = s
x <°>
i + r
and since c is arbitrary we have
y = x
ex
+ 2
We call the reader's attention to the fact that x/2 is the only solution of the given
equation defined on the entire real line. Nevertheless it is common practice to
call c/x +
x/2 the "general solution" of (3-21) without specifying the interval
in question—a practice which is admittedly convenient, but potentially misleading.
Example 3. We now use the above technique to solve the nonlinear equation
n
«iW| + a (x)y = h(x)y ,
(3-23)
equation,and here, as always, we assume that a (x), ai(x), and h(x) are con-
tinuous on an interval /, and that a\{x) 9^ on /.
Dismissing the cases n = and n = 1 which have been treated above, we
rewrite (3-23) as
a 1 {x)y~
n
& + a (x)y - 1 n
= h(x), (3-24)
which is a normal first-order equation on the interval /. We now solve this equa-
3-3 | FIRST-ORDER EQUATIONS 99
and then express the general solution of (3-23) as y = w 1/( 1_w) Finally,
tion for u, .
_2 dy .
_i 2
du 2
ue~
x
= - x 2 e~x dx + c.
J
Hence
u = 2 + 2x + x2 + ce
x
,
and y = 0.
EXERCISES
7. L—
at
+ Ri = E, L, R, E constants, L, R ^
2
8. (3x + 1)/ - 2*3> = 6x
2
9. (x + 1)/ - (1 - xfy = xe~x
2
10. (jc + 1)/ + xy = (1 - 2x)\/*2 + 1
11. x sin x —
ax
+ (sin x + x cos x)^ = xe
x
y
12. x ^
dx
+
\flx-\-
7
1
= 1 + V2x+ 1
100 GENERAL THEORY OF LINEAR DIFFERENTIAL EQUATIONS |
CHAP. 3
13. x (f
dx
+
y/i
,
y
_ X2
= (1 +VT^)^
2
14. sin x cos x -~
fi?X
+ y = tan x 15. (1 + sin x)-j-
ax
+ (2 cos x)y = tanx
,^ 0/1
16. 2(1
v
— 2
x N^
)/
//,2V
— — x (1 )y = xy 3 e -x ,-
17. / =
, / sin x :
- y cos x
2
sin x cos x
+ xy 2 - x =
18. yy'
2
19. (x + \Wy~y' = xe
2
3x/2
+ (1 - x)
20. (x + x + l)y/ + (2* + 1)/ = 2x -
2
^ 1
x(x + In x)
22. —— - y' + y = (1 + cos x).k
y __ sin 2x 2/3
21. x/ + ^— = -^-5-:
, .
/ , ,
2
lnx ^ lnx 6
+ l)lnx - x(3x + 4)/ (a:
23. (x - 1)/ - 2y = V(x2 - l)y 24. / =
(x3 + 2x2 - 1)>>2
2 3 2
25. (xy = (xy) (x + 1)
)'
26. Find the particular solution of the equation xy' — (sinx)^ = on the interval
(0, 00 ) which passes through the point (1, — 1). [Hint: Show that the general solution
dy -x 2 /2
x—
dx
+y = e
which passes through the point (2, -3). [Hint: Find the general solution and show
that it can be written in the form
C I -x*/2 .
y = - + -1 I e dx.\
-,
x x J2
(b) What is the ordinate of the point on the solution curvefound in (a) corresponding
to the point x = 1 ? (Consult a table of values for (l/V2ir)/f „ e~
t2/2
dt.) Find the
slope of the solution curve at this point.
^
dx
+ « (x)/ + ai{x)y + ao(x)
2 = 0, (3-26)
in which ao(x), ai(x), a2(x) are continuous on an interval / and a 2 (x) ^ on is called /,
a Riccati equation. A number of elementary facts concerning the solutions of such equa-
tions are given in the exercises which follow.
28. Let y\{x) be a particular solution of (3-26). Make the change of variable y =
yx +1/z to reduce (3-26) to a first-order linear equation in z, and hence deduce
that the general solution of a Riccati equation can be found as soon as a particular
solution is known.
Use the technique suggested in the preceding exercise to find the general solution of each
of the following Riccati equations.
29. y' — xy 2 + (2x — l)y = x — 1 ;
particular solution y = 1
3-3 I
FIRST-ORDER EQUATIONS 101
1
= cos X
33. y' — (sin
2
x)y 2 -\ y + cos 2 x = 0; particular solution .y
sin x cos x sin x
34. (a) Let yi(x) and >>2(*) be two particular solutions of Eq. (3-26). Show that the
general solution of the equation is
y - yi
(b) Let yi(x), y2(x), and j3(x) be distinct particular solutions of Eq. (3-26). Use the
result established in (a) to prove that the general solution of the equation is
( y - yiX-Vs - y2)
(y - y2)(y<i — yi)
c an arbitrary constant.
2
^
ax
+ ay + by + c =
(b) Use this result, together with Exercise 28 or Exercise 34(a), as appropriate, to
find the general solution of each of the following Riccati equations,
(i) y+ y
2
+ 3j + 2 = (ii) y + 4>> 2 - 9 =
(iii) /+ y
2 - 2y + 1 = (iv) 6/ + 6>> 2 + - >> 1 =0
36. (a) Prove that the change of variable v = y'/y reduces the second-order homoge-
neous linear differential equation
and hence deduce that the problem of solving (3-27) is equivalent to that of solving
the simultaneous pair of first-order equations
—- =
ax
vy, —
ax
= 2
-v - ai(x)v - ao(x). (3-29)
(b) What conditions ought one impose on (3-29) to correspond to the conditions
*(0) = >>o,/(0) = vi on (3-27)?
(c) Prove that every Riccati equation (3-26) in which a2(x) 9^ can be converted
to a second-order homogeneous linear differential equation by making the change of
variable y = v'/fazo).
37. Find the Riccati equation associated with y" — y = 0. Solve this equation, and
hence find the general solution of y" — y = 0.
38. Prove that whenever m\ and /«2 are distinct real roots of the quadratic equation
am 2 + bm + c = 0, a, b, c constants,
then emix and em2X are linearly independent solutions in ©(—«>, ») of the second-
order homogeneous linear differential equation
ay" + by' + c = 0.
/' - 2my' + m 2y = 0,
m a constant.
41 Use the result of the preceding exercise to find the general solution of each of the
following second-order linear differential equations.
(a) /' + 2/ + 1 = (b) Ay" - 12/ + 9y =
(c) (D - %) y
2 = (d) (36D 2 - 12D + l)y =
(e) (2D 2 - 2V2 D + l)y =
__ L, /
h{X) f[a X )/ ai (x)lrf3; —^[a Q (x)l ai (x)]dx
1
g
(
jx c (3-30)
in Fig. 3-2. Even more important, it is easy to see that there is a solution curve
passing through any preassigned point (x , y ) in this strip, since (3-30) can be
solved for c when x = x , y = y .
in question. In these terms our earlier results can be summarized by saying that
every initial-value problem involving a normal first-order linear differential
equation has at least one solution.
FIGURE 3-2
At this point it is only natural to ask whether or not such a problem can admit
more than one solution. This is the so-called uniqueness problem for first-order
linear differential equations, and is anything but an idle question. Indeed, in appli-
cations of differential equations to the natural sciences it is often essential to be
able to guarantee that the problem being investigated has a unique solution,
since any attempt to predict the future behavior of a physical system governed
by an initial value problem relies upon this knowledge. In the case at hand, it is
not difficult to show that the desired uniqueness obtains (see Exercise 14 below),
and hence the above assertion can be amended to read as follows:
The general theory of linear differential equations can properly be said to begin
with the theorem which generalizes this result to nth-order equations. In the
special case treated above, the theorem was proved by the simple expedient of
exhibitingall of the solutions at issue. Unfortunately, it is impossible to give an
argument of this type for equations of higher order, and though the asserted
theorem is true, its proof is not conspicuously easy. Thus, rather than become
104 GENERAL THEORY OF LINEAR DIFFERENTIAL EQUATIONS |
CHAP. 3
involved in a long and somewhat arid discussion at this time, we content ourselves
with a formal statement of the result.*
Theorem 3-2. (The existence and uniqueness theorem for linear differen-
tial equations.) Let
an (x)^n + •
+ a (x)y = h(x) (3-31)
there exists one and only one solution y(x) of (3-31) with the property that
~l
y(x ) = y ,
y'(x ) = yu . . .
,y
{n
\x ) = yn _ x .
It is also worth noting that Theorem 3-2 can be phrased in the language of linear
operators, in which case it assumes the following suggestive form:
w
IfL: e (7) — > G(I) is a normal nth-order linear differential operator, there exists
and
(ii) G(h)(x ) = yo, G(h)'(xo) = yu . .
. , ^^"-"(^o) = ^»-i-
When stated in these terms it is clear that the task of solving an initial-value
problem for a normal linear differential equation comes down to finding an explicit
form for the inverse operator G, since once G is known the problem
Ly = h;
n-1) = yn -u
y(x ) = y , . . • ,y (*o)
can be solved by computing the value of G(h). This point of view will be exploited
in later chapters where much of our work will be directed toward finding G for
specific classes of linear differential operators. As we shall see, G will turn out to
be an integral operator of the type considered in Example 2 of Section 2-2.
EXERCISES
Find the solution of each of the following initial-value problems and specify the domain
of the solution.
1. */ + 2y = 0, y(l) = -1
(sin*)/ + (cosx)y = y( =
2. 0,
~J 2
3. 2/ + 3y = e~x y(-3) = -3 ,
2 2
4. (x + 1)/ - (1 - x )y = e~\ y(-2) =
Use the given general solution to solve each of the following initial-value problems.
8. y" - k y = 0, = = a sinh kx +
2
y(0) y'(0) 1 ; y = c 2 cosh kx, k ^
* -
9. (1 — x 2 )y" - 2xy' = 0, y(-2) = 0, y'(-2) = 1; y = ci + c 2 In
1
x +
,
,
1
2
11. 4*y + 4xy' + {Ax - \)y = 0,
y(^j
- -l./(^) = 0;
—
irx
(ci sin x + c2 cos x)
y = c\ sin x + C2
14. (a) Let y i and y 2 be solutions of a normal first-order linear differential equation on
an interval /. Prove that y\ — y 2 is either identically zero or is different from zero
everywhere on /.
(b) Use the result in (a) to deduce that every initial- value problem involving a normal
first-order linear differential equation has at most one solution.
15. Give an example to show that the conclusion of Theorem 3-2 fails when the hypothe-
sis of normality is not satisfied.
16. Let y\ and y 2 be distinct solutions of a normal first-order linear differential equation
on an interval /. Prove that the general solution of the equation on / is
y - yi
= c,
yi - 72
where c is an arbitrary constant. [Hint: See Exercise 14(a).]
106 GENERAL THEORY OF LINEAR DIFFERENTIAL EQUATIONS |
CHAP. 3
show that c\ and c 2 can be chosen so that y(0) = w(0), /(0) = u'(Q), and then
apply Theorem 3-2.]
19. Show that two distinct solutions of a normal first-order linear differential equation
cannot have a point of intersection.
20. Prove that every nontrivial solution u(x) of a normal second-order linear differential
equation
a 2 (x)y" + a 1 (x)y' + a (x)y =
has only simple zeros. [A point xo is said to be a zero of a function u(x) if and only if
G(h)(x ) = y .
23. Given the linear differential operator D - k(k& constant), find the inverse operator
G which satisfies G(h)(xo) = yo.
"n(x)^ + -
+ a (x)y = (3-33)
Proof Let jc be a fixed point in I. Then by Theorem 3-2 we know that this equa-
conditions
yi(x ) = l,/i(*o) = 0, . .
.
.^"-"(xo) = °>
w_1)
y 2 (x ) = 0,y'2 (x ) = 1, . . . ,y2 (xo) = 0,
(3-34)
(n-l)/
,^"~ 1,
yn (.x ) = 0, j*(*o) = o, . .
. (*o) = 1
3-5 | DIMENSION OF THE SOLUTION SPACE 107
n .*
are the standard basis vectors in (R (See Fig. 3-3.) We assert that these solu-
tions are a basis for the solution space of (3-33).
FIGURE 3-3
ciyi(x) + • • •
+ cnyn (x) =
on I. Then this identity, together with its first n — 1 derivatives, yields the system
ciydx) + c 2y 2 (x) + • • •
+ cnyn (x) = 0,
ci/i(x) + c 2y 2 (x) + • • •
+ cny'n (x) = 0,
(3-35)
n~l
^-"W + c 2y 2 \x) + • • •
+ Cnjfr-VOc) = 0.
Setting x = x , we obtain
ciyi(x ) + c 2y 2 (x ) + • • •
+ cnyn (x ) = 0,
Ciy'i(xo) + c 2y 2 (x ) + • • •
+ cny^(x ) = 0,
(3-36)
Ciyf-'Kxo) + c 2 yf-
l
\x ) + • • •
+ cny^-
l
\x Q ) = 0,
y(x = a y'(x = ax /* u (x
(n-l)/ = an _i. (3-37)
) , ) , . . .
, )
* This choice of solutions has been illustrated in Fig. 3-3 for a second-order equation,
in which case
0>i(*o),/i(*o)) = (1,0),
(y2(xo),y 2 (x )) = (0,1).
108 GENERAL THEORY OF LINEAR DIFFERENTIAL EQUATIONS |
CHAP. 3
Then by the uniqueness statement in Theorem 3-2 we know that y(x) is the
solution of (3-33) which satisfies these particular initial conditions. But, using
(3-34) again, we see that the function
floJiW + aiJ>2(*) + * • •
+ an -iyn (x)
We call the reader's attention to the fact that the particular numerical values
used to fix the solutions yi(x), yn (x) did not really play an essential role in
. . .
,
the argument given above. Indeed, the success of the proof depended only on the
linear independence of the vectors
n
in (R and the choice made in (3-34) merely served to simplify our computations.
,
For as long as these vectors are linearly independent, the system of homogeneous
linear equations (3-36) will have a unique solution, and the a will be zero, as
required.
d 2y - =
y (3-38)
dx 2
is normal on the entire x-axis, and thus its solution space is a 2-dimensional
subspace of e(— go, oo). Moreover, it is easy to show that the functions
x x
yi(x) = \{e + e~ ) = cosh x,
= i( ex - ~x =
y 2 (x ) e ) sinnx
J>i(0) = 1, /i(0) = 0,
the argument given above implies that cosh x and sinh x are a basis for the solu-
tion space of this equation. Thus the general solution of (3-38) is
y = ci cosh x + c 2 sinh x,
x
yi(x) = e , y 2 (x) = e~
x
Ji(0) = 1, /i(0) = 1,
y 2 (0) = 1, / (0)
a = -1,
2 x
and since the vectors (1, 1) and (1, —1) are linearly independent in (R , e and
e~x also form a basis for the solution space of the equation. It follows that the
general solution of (3-38) may also be written
y = c xe
x
+ c 2 e~
x
,
g+ *y - o (3-39)
that sin 2x and cos 2x are linearly independent in e(— oo, oo). Hence they are a
basis for the solution space of (3-39), and the general solution of that equation is
n
are linearly independent in (R . Then yi(x), . . .
, yn (x) are linearly inde-
pendent in Q(I).
)
are linearly independent in e(-oo, oo) since the above test applied at x =
3
yields the linearly independent vectors (1, 1, 1), (0, 1, 2), (0, 0, 2) in (R .
EXERCISES
For each of the following homogeneous linear differential equations, (a) show that the
given functions span the solution space of the equation on an appropriate interval,
(b) choose a basis for the solution space from among these functions and use it to express
the general solution of the equation, (c) find a basis for the solution space of the
and
equation which satisfies initial conditions of the form (3-34) at the given point x .
2x
1. /" - /' - 2/ = 0; e~\ sinh x - \e\ 2e 1; x = ,
x + - x + x In x, - + x In - x(l - In x); xo = e
x
>
xx ^-^>
»
5. (x
2
D2 - 2)y = 0; 2x
2
-
xx
-» 3x
2
; x = 1
2* -1
6.
*v"
- ,
1 - x2
/ = 0; 1 - tanh x, 1 + In ^-^
1+*
'2; x =
7. Use the equation xy' + >> = to show that Theorem 3-3 fails if the hypothesis of
normality is not satisfied.
8. Show that the following functions are linearly independent in Q(J) for the given
interval /.
ax hx
(a)e , e {a*b)\ (-00,00)
2
(b) l,x,x ;
(-co,oo)
(d) x, x In x; (0, 00 )
ax ax
(e) e sin bx, e cos bx (b ^ 0); (- °o , 00
*10. Suppose that j>i(x), ,y n (x) are solutions of a normal homogeneous linear dif-
. . .
are linearly independent in 6(7), they are linearly independent in 6(7) for any sub-
interval 7 of 7. Give an example to show that this result fails if yi(x), . . . , y n (x)
do not satisfy such an equation.
3-6 THE WRONSKIAN 1 1
n
are linearly independent in (R . For our present purposes this result can be stated
more conveniently in terms of the determinant of a certain matrix, as follows:
Let yi, . .
, yn be arbitrary functions in C n_1 (7), and for each x in I consider
the matrix
y x (x) y 2 (x) • • •
yn (x)
y'i(x) y'2 (x) y'n {x) (3-42)
yr ^) ^- 1 i}
(jc) y'n~ \x\
l
,(w-l)/ v \ „(n-l)
y?- l
\x) yV-»{x) •••
y\r
l
\x)
For example,
x sin x
W[x, sin x] = = x cos x — sin x,
1 cos a:
and
x 2x
W[x, 2x] = = 0.
1 2
only if the columns of (3-42) are linearly independent when x = x But for -
each x in / the columns of (3-42) are none other than the vectors in (3-41), and
we therefore have the following theorem.
n_1
Theorem 3-4. The functions y i, yn are linearly independent in 6 (7),
. . .
,
and hence also in 6(7), whenever their Wronskian is not identically zero on I.
112 GENERAL THEORY OF LINEAR DIFFERENTIAL EQUATIONS CHAP. 3
Example 1. Since
x
e e
x x
W[e e~ , ] x
= -2,
e -e
x x
the functions e and e are linearly independent in e(7) for any interval /.
112 312
Example 2. The functions x, x , x are linearly independent in e(Y) for
any subinterval / of the positive x-axis since
.1/2 ^3/2
AI 1/2
L, ^3/2-1 -1/2 iv l/2
W[x, x \ x <] A 2
_l Y -3/2 Hy" 1/ 2
More generally, xa , x&, xy are linearly independent in ©(/), / as above, if and only
if a, |8, 7 are distinct real numbers (see Exercises 13 and 14 below).
3 3
Example 3. The functions jc and |x| are linearly independent in e(— oo, oo),
for if Cix
3
+ c 2 |*|
3
— 0, then
Cl (l)
3
+ c 2 |l|
3
= 0,
ci(-l)
3
+ c 2 |-l|
3
= 0,
„3
W[x 3
,
\x\
3
]
= = 0,
2 2
3x 3x
if x > 0, and
—x
W[x 3 ,
\x\
3
]
= = 0,
2 2
3x •3jc
This can in fact be done, simply by requiring that the functions be solutions of a
homogeneous linear differential equation. We prove this assertion as
an (x) g+ • • •
+ a (x)y = (3-44)
Ciy'iixo) + • • •
+ cny'n (x ) = 0,
(3-45)
n_1)
ctfi—'W + •
+ <W* (*o) = 0,
cally on / the determinant of (3-45) is zero, and the system has a nontrivial solution
(<?!, . .cn ). (See Theorem IH-9.) Thus the function
. ,
y(x) = J2 Ciyi{x)
n - 1}
X*o) = 0, y'(x ) = 0, . . . ,y (^ ) = 0.
But the zero function is also a solution of this problem, and hence Theorem 3-2
implies that
ciydx) + ••• + c nyn {x) =
Once again we have established a result which is stronger than the one adver-
tised. For the above proof only made use of the fact that the Wronskian of
y\, yn vanished at a single point in /, and hence the conclusion remains true
• • •
,
under this more restrictive hypothesis. Combined with Theorem 3-4 this obser-
vation yields
Theorem 3-6. A
of solutions of a normal nth-order homogeneous linear
set
differential equation is linearly independent in Q(I), and hence is a basis for
the solution space of the equation, if and only if its Wronskian never vanishes
on I.
114 GENERAL THEORY OF LINEAR DIFFERENTIAL EQUATIONS CHAP. 3
Example 4. By direct substitution the student can verify that sin 3 jc and 1/sin 2 x
are solutions of
^+
ax 2
tan x Q-
ax
- 6 (cot
2
x)y = (3-46)
on any interval I in which tan x and cot x are both defined. Moreover,
3 1
sin x
'
sin 2 x
W Sin
3 1
*' sin 2 x\
=
2 2 cosx
= 5 cos x.
[
3 sin x cos x
sin 3 jc
3 2
Since cos x is never zero on I, the above theorem implies that sin x and 1/sin x
are linearly independent in Q(I), and the general solution of (3-46) therefore is
3 c2
y — C\ sin x + sin 2 x
(Note that this result can also be obtained directly from Theorem 3-4.)
Example 5. The functions
3
ydx) sin x, y2(x) sin x j$ sin 3x
are solutions of
r2 j
^+(tan;c-2cotJc)^ (3-47)
sin
3
jc sin x — ^ sin 3a:
W\yi(x\ y 2 (x)] =
3 sin
2
jc cos jc cos jc — cos 3jc
= sin
3
jc(cos jc — cos 3jc) — 3 sin
2
jc cos jc(sin jc sin 3jc)
= sin
2
jc(sin 3jc cos jc — sinjccos3jc) — sin
2
jc(2 sin jccos jc)
= sin
2
jc sin 2jc — sin
2
jc sin 2jc = 0.
Hence y 1 and y 2 are dependent in 6(7), and do not form a basis for the
linearly
solution space of (3-47).* it is clear that any constant c
In this case, however,
3
is a solution of (3-47), and since c and sin jc are obviously linearly independent
y = C\ -f- c 2 sin
3
x.
EXERCISES
By computing Wronskians, show that each of the following sets of functions is linearly
independent in 6(7) for the indicated interval /.
1. 1, e~ x 2e 2x
, on any interval /
x
2. e , sin 2x on any interval /
3. 1, X, X , . . . , X on any interval I
4. In x, x In x on (0, oo )
5. X
\I2 ^1/3 on (0, oo
ax sin
6. e bx, e? x cos Z>x (6 * 0) on any interval /
8. e~ x xe~ x x 2 e~ x
, , on any interval /
x - 1
10. on (—oo, —1)
11. Vl — X2 , JC on (-1,1)
X
12. sin
•
- , cos z2 x on any interval /
13. Show that x", x&, x y are linearly independent in 6(0, oo) if and only if a, /3, 7 are
distinct real numbers. [Hint: If a, /3, 7 are distinct and cix" + C2X& + C3X y =
on (0, oo ), show that c\ = C2 = c% = by considering what happens as x tends
to oo.]
14. Show that x", x?, xy are
independent in 6(0, oo) if and only if they are
linearly
linearly independent in 6(/) for every subinterval / of (0, oo). [Hint: First establish
both of the following assertions, and then use Theorem 3-6:
(a) x", x&, xy satisfy the linear differential equation
1 1 1
a a+l}+y - 3
(b) W(x ,x ,x Y ) = x l3
a @ 7
and hence Wix", xP, xy ) either vanishes nowhere in (0, oo) or vanishes identically.]
15. Generalize the results of Exercises 13 and 14(b) to show that x" 1 , . . . ,xcln are linear-
ly independent in 6(7) for any subinterval I of (0, oo) if and only if a\, . .
.
, an are
distinct real numbers.
18. Suppose that /is an odd function in eH-a, a] (that is, f(—x) = —/(*)) and that
/(0) = /'(0) = 0. Show that
W[f(x), \f(x)\\ =
for allx in [—a, a], but that /and |/| are linearly independent in C 1 [— a, a] unless
/is identically zero. Compare this result with Example 3 in the text.
19. Let / g be any two functions in C 1 (7), and suppose that g never vanishes in /.
1 1
••• 1
a\ 02
W[e
aiX
,
a x
„ n l _ _(°H \- a n> x 2
02
2
n-l n—
«2
each index /, let u\ j} be the solution of the given equation which sal^sfies the initial
conditions u[%
xo
(xo) = 0, < j < n — 1, j /, and «{,*„(* o) = — 1. Then ^
Oi(Xo) = u i% (
(xo).]
23. Let u\ and U2 be linearly independent solutions of the normal second-order linear
differential equation
/' + ai(pc)y' + a (x)y = 0.
Express the coefficients ao(x), a\(x) in terms of u\ and U2. [Hint: Let y be an arbi-
trary solution of the equation, and consider the Wronskian of y, u\, U2-]
24. Generalize the result of the preceding exercise to an nth-order equation
>><"> + an-i(x)y< n -» + • • •
+ a (x)y = 0.
25. Use the results of Exercise 23 to find a homogeneous second-order linear differential
equation whose solution space has the following functions as a basis.
(a) x, xe x (b) x, x2
(c) sin x, cos x (d) x, sin x
(e) x, lnx
3-7 ABEL'S FORMULA 117
This fact can also be deduced from the following theorem, which gives an explicit
formula for the Wronskian in this case.
°2
^ dtf
+ fll
^ ~dx
+ a °(*)y = °
yi(x) J>2(*)
i<ny>oo,y>w = i y'i(x) y'2 (x)
ai(x)
myi(x),y 2 (x)].
02(*)
differential equation
4? ,
?iW y = *
dx a 2 (x)
y **" »
as asserted. |
If x is any fixed point in /, and W(x ) denotes the value of the Wronskian
if
of yi, • • •
, yn at x , then Abel's formula may be written in the form
This formula shows that the Wronskian of any basis for the solution space of a
normal homogeneous linear differential equation is determined up to a multi-
plicative constant by the equation itself, and does not depend upon the particular
basis used to compute it. This simple observation will be important later on.
Example 1. Since
2
dy ,
dy A
isnormal on (0, oo), the Wronskian of any two solutions y lt y 2 of this equation
must be of the form
C
W\y {x),y {x)\ = ce~hdxlx) = -- l 2
a b
c = x x (a bi — aib ),
and
x (a bi — aib Q )
W
* This result actually holds for all n > 2, provided that ai(x) and a 2 (x) are replaced
by a„_i(x) and a n (x), respectively.
3-7 | ABEL'S FORMULA 119
dy 2 —}l n
f[ai(x)lao(x)]dx
/
yi(x
\
~ /1WJ2
// \
= „
ce
2(
)~fa
y2 = cy^x)
J :r7
yi(x)
^ dx + kyi(x),
where k is an arbitrary constant, and since this formula is valid on any sub-
interval of / in which y x 9^ 0, it can be used to determine a second solution of
(3-51) on such a subinterval. In particular, the function
—Mx?
/-—f[a (x)la (x)]dx
1 2
dx
,—J[«l(*)/<*2(*)]<k ;
y = co>i + c 2y 2 ,
Hence a second linearly independent solution in (2(0, oo) can be found by solving
the first-order equation
jcV ~ 2xy = e~ x2 ' 2
.
4 x 2
-/ x- e~ <
dx,
on any interval in which tan jc and cot x are both defined. Applying the above
result we obtain a second solution
2
= J
sin x cos x dx
3
= % sin jc.
y = C\ + c 2 sin
3
x,
which agrees with the result obtained at the end of the last section.
EXERCISES
1 . For each of the following differential equations find the Wronskian of the solutions
y\, yiwhich satisfy the given initial conditions.
2 2
(a) x y" + xy' + (x + \)y = 0; j>i(l) = 0, ^(1) = 1, y 2 {\) = /2 (D = 1
2
(c) x y" - 3xy' + y = 0; yi (-l) = y\(-l) = 2, y 2 (-l) = 0,
/2 (_1) = -1
(d) /' + 2xy = 0; yi (0) = y\(0) = 1, y 2 (0) = 1, /2 (0) =
(e) y" - (sin x)y' + 3 (tan jc)y = 0;
>i(0) = 1, yi(0) = 0, > 2 (0) = 0, y2 (0) = 1
2
(f) Vl + Jtsy - x /+ y = 0; ji(l) = 1, /i(l) = 0, y 2 (l) = -1,
•ad) = 1
In Exercises 2 through 8, one solution of the differential equation is given. Find a second
linearly independent solution using the method of this section.
2. y" - 4/ + 4y = 0; e 2x
3. y" - lay' + a 2y = (a constant); e"*
4. 3x/' - / = 0; 1
5. y" + (tan*)/ — 6(cot 2 jc)^ = 0; sin 3 x
(1 - x )y" - 2x/ = 0; 1
2
6.
7. (1 - *V - 2xy' + 2y = 0; x
8. 2xy" - (e*)/ = 0; 1
9. (a) Prove that if fa is in e^/), 1 < h J < 2, then so is the function F defined by
fu(x) fi2(x)
F(x) =
f2\(,x) f22(x)
and that
(b) Generalize the result in (a) to the case of nth-order determinants; and show, in
particular, that
fu(x) fi 2 (x) • • •
j\n(,x)
can be expressed as the sum of n determinants, the fth of which is obtained from
|/i/(*)l by differentiating the functions in the /th column.
without using any information other than that provided by the equation itself
independent functions, C(x) and S(x), which are defined for all x and satisfy
the initial conditions
C(0) = 1, C'(0) = 0,
(3-57)
K J
S(0) = 0, S'(0) = 1.
valid for all x, we conclude that both C(x) and S(x) are infinitely differentiable
on (— oo , oo), and that all of their derivatives are solutions of (3-56). For example,
the identity C"(x) + C(x) = implies that C"(x) is diflferentiable, since C(x) is,
and that C""(x) + C'(x) = 0. But this is just (3-56) again with C'(x) in place
of y; S'(x) is treated similarly, and the argument can be repeated to establish
the assertion for still higher derivatives. In particular,
and the derivatives of C(x) and S(x) repeat in cycles of four, as expected. Finally,
(3-57) and (3-58) imply that
and it follows that C'(*) is the solution of (3-56) which satisfies the initial con-
ditions C"(0) = 0, C"(0) = — 1, while S'(x) is the solution which satisfies
S(x)
2
+ CO) 2 = 1. (3-59)
Indeed, since
^ '~ n2 A 21
/o/,
2C(x)C'(x) 0,
it follows that
S(x)
2
+ C(jc)
2
= k,
d2 2
—o
dx 2
C(a + x) = -C{a
—C(a + x) and —d 2 S(a +
-— x) = -S(a + x).
dx
Thus C(a + x) and .S(a + x) are solutions of (3-56), and as such must be linear
combinations of C(x) and S(x) that ; is,
and we conclude that the graph of C(jc) is symmetric about the y-axis, while that
of S(x) is symmetric about the origin (see Exercise 1).
At this point we could derive those long and all too familiar lists of trigonometric
identities involving C(jc) and S(x). However, it is much more instructive to prove
that these functions are periodic with period 2x. Here we begin by defining t/2
to be the smallest positive real number such that C(x) = 0. (The proof that such
a number exists has been left to the student in Exercise 2.) Then C(jc) is positive
on the interval (0, tt/2), and since S f
(x) = C(x), we conclude that S(x) is in-
creasing on that interval. But 5(0) = 0, and hence S(x) is also positive on (0, x/2).
,
Hence
C(x + 2w) = C(jc)C(2t) - <S(x)5(2tt) = C(x),
EXERCISES
1. Establish the identities (3-62) by showing that C(— x), S(—x) are solutions of (3-56)
and expressing them in terms of the basis C(x), S(x) for the solution space.
*2. Show that there is a least positive real number a such that C(a) = 0. [Hint: Assume
the contrary; then argue that
(a) C(x) > for < x < <»,
(b) S(x) > for < x < oo
(c) C(x) decreasing and concave downwards on (0, oo ),
is strictly
and derive a contradiction from (a) and (c).* Having established that C(x) has
positive zeros, consider the greatest lower bound a of the set of positive zeros of C(x).]
3. Let E(x) be the unique solution on (— oo , oo ) of the initial value problem y' — y = 0,
* To establish these facts rigorously, the intermediate value theorem and mean value
theorem from calculus must be applied. The student, however, may give an intuitive
argument based on a consideration of the graphs of C(x) and S(x).
3-8 | THE EQUATION y" +Y= 125
(e) E(— x) = -—r^ for every real number x; [Hint: Apply (d).]
E{x)
(0 < E(x) < 1 on (-oo,u);
4. (a) Prove that every solution of a homogeneous linear differential equation with
constant coefficients has derivatives of all orders at every point on the x-axis.
(b) Generalize this result to homogeneous equations with variable coefficients and to
nonhomogeneous equations.
4
equations with
constant coefficients
4-1 INTRODUCTION
Linear differential equations with constant coefficients, that is, equations of the
form
any
in)
+ fln-iy*-
1'
+ • • •
+ a y = h(x) (4-1)
in which a , . . . ,an 7^ are (real) constants, are in many respects the simplest
of all differential For one thing, they can be discussed entirely within
equations.
the context of linear algebra, and form the only substantial class of equations of
order greater than one which can be explicitly solved. This, plus the fact that
such equations arise in a surprisingly wide variety of physical problems, accounts
for the special place they occupy in the theory of linear differential equations.
We shall begin the discussion of this chapter by considering the homogeneous
version of Eq. (4-1), which can be written in normal form as
n~ x
(D n + an _ x D + • • •
+ a )y = 0, (4-2)
or as
Ly = 0, (4-3)
~1
where L is the constant coefficient linear differential operator Dn + an _iD n +
• • •
+ a . Algebraically such operators behave exactly as if they were ordinary
polynomials in D, and can therefore be factored according to the rules of ele-
EXERCISES
1. (a) Prove that the product of two complex numbers a + bi and c + di is real if
and only if either
(i) b = d = 0, or
(ii) a = c and b = —d.
[Hint: Recall that a complex number a + and only if b = 0, and that
bi is real if
a product of the form (a bi)(c di) + + is computed by using the distributive law
and the rule i
2 = — 1.]
(b) Let P(x) be a polynomial with real coefficients, and suppose that P(x) has
a + bi, b > 0, as a root; that is, P(a + bi) = 0. Prove that a — bi is also a root
ofP(x).
2. Let P(x) be a polynomial of degree n, n > 0, with real coefficients. Use the fact
complex number system to prove that P(x) can
that P(x) has exactly n roots in the
be factored into a product of linear and quadratic factors with real coefficients.
[Hint: See Exercise 1(b) above.]
3. Find the second-degree polynomial which has a + bi and a — bi, b > 0, as roots.
4. Prove that every polynomial of odd degree with real coefficients has at least one real
root. [Hint: See Exercises 1(b) and 2 above.]
5. Write each of the following linear differential operators as a product of operators of
degrees one and two.
(a) D 3 + AD 2 + 5D + 2 (b) D3 - D2 + D - 1
(D 2 - 4)y = (4-4)
may be rewritten
(D + 2)(Z> - 2)j> = 0.
2x
y = Cl e + c 2 e~
2x
,
m2 + aim + a = (4-6)
known as the auxiliary or characteristic equation of (4-5), and then rewrite (4-5) as
(D - ai)(Z) - a 2 )y = 0. (4-7)
This done, the argument falls into cases depending on the nature of a x and a 2 ,
as follows:
Case 1. «! awfif a 2 real and unequal. Here the reasoning used in the above ex-
ample carries over without change; the functions e a x and e a * x are linearly in- i
aix
y = Cie + c 2e
a 2x
(D - a)
2
y = 0, (4-8)
4-2 | HOMOGENEOUS EQUATIONS OF ORDER TWO 129
ax
and our earlier argument yields but one solution of the equation, namely e .
Using it, however, we can apply the method introduced in Section 3-7 to find a
second linearly independent solution by solving the first-order equation
2ax
W[e ax ,y(x)] = e .
ax
y = (ci + c 2 x)e .
b > and the above method apparently breaks down. Nevertheless, if we pre-
0,
a
tend that e" * and e * x continue to make sense when a.\ and a 2 are complex,*
1
the discussion under Case 1 would imply that the general solution of (4-7) is
aiX a *x
y = Cl e + c2e
{a+hi)x ~ hi)x
= d xe + c2e
(a
ax ibx ibx
= e Cl e( + c 2 e~ ).
e
lx
= cos x + / sin x
y = e
ax
[ci(cos bx + isinbx) + c 2 (cos bx — isinbx)]
= e
ax
[(ci + c 2) cosbx + i(ci — c 2 ) sin bx]
= ax ax
c 3e cos bx + c^e sin bx.
Thus, on purely formal grounds we are led to e ax cos bx and e ax sin bx as a basis
for the solution space of (4-7) when a x = a + bi and a 2 = a — bi. Of course,
we must now verify that these functions actually are solutions of the given equa-
tion,and that they are linearly independent in C(— oo, oo). But this is routine and
has been left as an exercise for the reader.
Since these three cases include all possible combinations of ai and a 2 we have ,
completed the task of solving the general second-order homogeneous linear dif-
ferential equation with constant coefficients. For convenience of reference we
conclude by summarizing our results.
To solve a second-order homogeneous linear differential equation of the form
(D 2 + axD + a )y = 0,
* Properly interpreted they do, as the reader may verify by consulting any text on the
theory of functions of a complex variable.
130 EQUATIONS WITH CONSTANT COEFFICIENTS | CHAP. 4
m2 + aim + a = 0.
Then the general solution of the given equation can be expressed in terms of a i and
a2 as follows:
Complex, ai = a + bi
x
e" (ci cos bx + C2 sin bx)
<X2 = a — bi
EXERCISES
1. /' + / - 2y = 2. 3/' - 5/ + 2y =
3. 8/' + 14/ - 15v = 4. /' - 2/ =
5. /' + Ay = 6. 3/' + 2j =
7. /' + 4/ + 87 = 8. Ay" - Ay' + 3y =
9. /' - 2/ + 2y = 10. 9/' - 12/ + Ay =
11. /' + 2/ + 4y = 12. 2/' - 2V2/ + 7 =
13. 2/' - 5V3/ + 67 = 14. 9/' + 6/ + y =
15. 64/' - 48/ + 17y =
In Exercises 16 through 25 find the solutions of the given initial- value problems.
x
28. Verify that e° cos bx and eax sin bx are linearly independent solutions of the
equation
(£> - ai)(D - a 2 )y =
29. Find a constant coefficient linear differential equation whose general solution is
30. For each of the following functions find a second-order linear differential equation
with constant coefficients which has the given function as a particular solution.
(a) x(l -\- e x) (b) 4 sin x cos x
(c) (1 + 2e )e 2x + 6x +
x
5 (d) cos x(l - 4 sin 2 x)
(e) e 3x
+ e 2x + xe 3 *
31. (a) Show that the general solution of the second-order equation
[D 2 - 2aD + (a
2
+ b 2 )]y =
y = cie^cos (bx + c 2 ),
If L = (D -
2
32. a) , a real, show that
Le = (k — a) e .
x x
Lxe = (k — a)[2e + k(k — a)e ],
(D 2 - 2D + 26)y =
(b) Solve the problem given in (a) again, this time writing the general solution in
the form
(a+bi)x (a — bi)x
y = cie +,
c2e
Evaluate c\ and c 2 formally, and then use Euler's formula to show that the resulting
solution can be transformed into the solution found in (a).
*34. The function ez z a
, complex number, is defined by the infinite series
*--l+z + £+-+£+-.
132 EQUATIONS WITH CONSTANT COEFFICIENTS | CHAP. 4
and it can be shown that this series converges absolutely for all values of z* Set
z = ix in this series, and use the fact that i
2 = —1 to prove Euler's formula.
[Hint: Since the series is absolutely convergent for all z, its terms may be rearranged
at will.]
(Z)
4
- 2D 3 + 2D 2 - 2D + \)y = (4-9)
we first decompose the operator into linear and quadratic factors, as suggested
in Section 4-1, to obtain the equivalent equation
(D - \)\D 2 + \)y = 0,
and then invoke Lemma 4-1 to assert that the solution space of each of the second-
order equations (D — \)
2
y = and (D 2 + l)y = is contained in the solu-
x x
tion space of (4-9). Thus e , xe , sin x, and cos x are solutions of (4-9), and since
these functions are linearly independent in <B(— oo, oo), the general solution of
the given equation is
y = (^1 + c 2 x)e
x
+ c3 sinx + c 4 cosa:.
This, in brief, is how all homogeneous constant coefficient equations are solved,
and save for the difficulty occasioned by equations such as
(D - \fy =
and
(D 2 + \)
2
y = 0,
where the above argument fails to yield the required number of linearly inde-
pendent solutions, we are done. But, recalling our experience with the equation
{D — a)
2
y = 0, it is not difficult to guess that the missing solutions for the
above equations are, respectively, x 2 e x x 3 e x and x
, , sin x, x cos x. Both of these
conjectures are correct, and we shall now prove the relevant generalization of
this fact for arbitrary equations of the form
(D n + an ^D n - + 1
• • •
+ a )y = 0, (4-10)
We begin by decomposing the operator into linear and quadratic factors, the
linear factors being determined by the real roots of the auxiliary or characteristic
equation
n~ x
mn + an _ x m + • • •
+ a = 0, (4-11)
2 m
{a
2
+
b )] corresponding to the pair of complex roots a bi, b > 0. This ±
we accomplish in the following theorem, which, it should be noted, also includes
the case (D —
2
a) discussed earlier.
Proof. To establish this result we must compute the value of the linear differential
operator Lm applied to the product x
m~ 1 and we therefore begin by giving a
y,
formula for evaluating all such expressions. Specifically, if L is any linear dif-
~
Lm (yxm 1
) = 0.
-1
M(yxm ) = (My)xm - 1
+ (M^Dx™- + ^ Wy^x™' + 1 1
• • •
.
r > m. Moreover, when r < m, the rth formal derivative of with respect to
m ~ r (see
M
D, Mir
\ consists of a sum of terms each of which contains the factor L
Exercise 20 below). Hence, since Ly = and since L is a constant coefficient
operator, Lemma 4-1 applies, and we find that
My = 0. Thus all of the terms M
in the above expression are zero, and the theorem is proved. |
As a consequence of this theorem we can now assert that the null space of the
operator {D — a)
m contains the functions
e
ax
sin bx, xe ax sin bx, . .
e
ax
cos bx, xeax cos bx, . .
And it is out of just such functions that the general solution of every homogeneous
constant coefficient linear differential equation is constructed. The construction
depends, of course, upon the fact that the various functions obtained in this way
are linearly independent in e(— oo, oo). They are, but unfortunately there is no
really brief proof of this assertion. One particularly elegant proof will be given
in Section 7-7 as an illustration of the ideas introduced there, and in the mean-
time we will content ourselves with indicating an alternate approach in Example
5 below.
(D 3 + \)y = 0. (4-13)
m3 + 1 =
are -1, \ + (V3/2)/, and \ - (\/3/2)/. Thus the general solution of (4-13) is
y = c x e~
x
+ e
xl2
lc 2 sin -^ x + c 3 cos -^- x) •
Example 2. Solve
y
(7) _ 2y5) +y = 3)
Q
(D 7 - 2D 5 + D 3 )y = 0,
and since
D 7 - 2D 5 + Z>
3
= D\D - \){D + l)]
2
,
the roots of the auxiliary equation are (a root of multiplicity 3), 1 (a root of
(D 2 + 1)(Z> - l)y = e
x
+ 2 - 7*sinx,
L = D(D - l)(D 2 + l)
2
.
Example 5. As our final example we shall prove that the various solutions of
the equation
(D - 2)(£> + 5) (Z>
3 2
- AD + 13)j; =
obtained using Theorem 4-1 are linearly independent in e(— oo, oo). In this case
the solutions are
e
2x
corresponding to the factor D — 2,
5x 5x 2 5x
e~ , xe~ , x e~ corresponding to the factor (D + 5) 3 ,
2x 2x
e cos 3x, e sin 3jv corresponding to the factor D 2 — AD + 13,
and obvious that they are somewhat too numerous to permit their Wronskian
it is
2x 2 5x 2x
c xe + (c 2 + c3 x + c 4 x )e~ + e (c 5 cos 3x + c 6 sin 3x) - (4-14)
Ci(D + 3
5) (Z>
2
- AD + 13)e
2x = 0.
c 4 (D + 5)\D 2 - AD + 13> 2 e- 5x = 0,
and since (D + 5)
2
(D 2 — AD + 13)x e
2 _5:i;
is not identically zero, we conclude
that c 4 = 0. Hence (4-15) becomes
2x
(c 2 + c 3 x)e~
5x
+ e (c 5 cos 3x + c 6 sin 3x) = 0.
In point of fact, this argument can be refined to give a general proof of the
linear independence of the functions appearing in the solutions of constant coef-
ficient homogeneous linear differential equations. We refrainfrom doing so,
however, since the problem will be considered in a later chapter where an entirely
different proof will be given.
EXERCISES
1. y">
_|_ 3/'
y _
3y = o ' - 2. y'" + 5/' - 8/ - My =
3. Ay'" + 12/' + 9/ = 4. /" + 6y" + 13/ =
5. 2/" + /' - 8/ - Ay = 6. /" + 3/' + / + 3y =
/iv) _ y = o
7. 8. / - 8/' + 16y =
iv)
9. /
iv)
+ 18/' + 81^ = 10. 4/ - 8/" - /' + 2/ =
iv >
H. yiv)
+ y »r _|_ y " = o 12. / = iv)
15. / +
v)
6/ iv)
+ + 36;;' + 24y
15/" + 26/' =
16. Find the general solution of (D 4
+ \)y = 0.
17. Find linear differential operators which annihilate each of the following functions,
2
(a) xV* +1) (b) 3e * cos 2x
(c) x(2x + 1) sin x (d) 3 + Ax - 2e~ 2x
2
(e) x 2 sin* cos* (f) xVsin x
4-3 HOMOGENEOUS EQUATIONS OF ARBITRARY ORDER 137
(b) Use the result in (a) to prove the assertion made during the proof of Theorem
4-1 to the effect that (r)
v = M
whenever r < m.
*21. (a) Show that the Wronskian of the functions ek i x e k ^ ,
e*n x is
k2
(*>! + +kn)x
••.
k
2
kl
, n— , n— • n—
K2
(d) What conclusion can be drawn about solutions of homogeneous constant co-
efficient linear differential equations from the results of (a) and (b)?
22. Prove that the solutions yi(x), y n (x) of a constant coefficient linear differential
. . . ,
equation are linearly independent in C(— w, oo) if and only if they are linearly
independent in 6(7) for every interval /. [Hint: Assume linear dependence on / and
use the uniqueness theorem.]
23. Prove that the functions e 2x xe 2x e 2x sin x, e 2x cos x are linearly independent in
, ,
6(7) for any /. [Hint: Show that this assertion is equivalent to asserting the linear
independence of 1, x, sin x, cos x in C(— oo, oo) (see Exercise 22), and then study
the behavior of c\ c 2x +
C3 sin x C4 cos x as x — oo .]
+ + »•
24. Establish the linear independence of the functions given in Exercise 23 by applying
the annihilator of sin x and cos x to the identity c\ + c 2x + C3 sin x + c± cos x = 0.
~
25. Prove that the functions e ax xe ax x m l e ax a real, are linearly independent in
, ,
. .
.
, ,
e(7) for any /. [Hint: A polynomial of positive degree has only a finite number of
zeros.]
138 EQUATIONS WITH CONSTANT COEFFICIENTS | CHAP. 4
26. (a) Let P(x) be a polynomial with real coefficients, and let L = P(D) be the asso-
ciated constant coefficient linear differential operator. Prove that
Le ax = p(a)e «*.
(b) Use the result in (a) to show that for any constant coefficient linear differential
operator L, Le ax = if and only if L has D — a as a factor.
27. Prove that
,n — .k m ax (0 if k > m,
(D a) x e =| m _k ax
\m(m — 1) • • •
(m — k + \)x e if k < m.
in the null space of L are linearly independent in 6(7) for any /. [Hint: Apply the
operators (£> n (D
a)" -
1-1
j3)
— ,
(D — j8)
n
(D — a) m_2 , . . . , in succession to the
identity
d ^x
~X
Cie ax _|_ . . .
+ CmXm-l e ax _|_ _|_ . . . _}_ dn X n ^x = 0,
where ai, . .
.
, a* are distinct real numbers.
Remark. The results of the last three exercises, when generalized to include complex
functions, furnish a proof of the linear independence of the solutions of the equation
Ly = obtained in this section.
differential equation
an (x) ^+ • • •
+ a (x)y = h(x) (4-16)
where yp isany particular solution of (4-16), and yn is the general solution of the
associated homogeneous equation
«»W %+ • • •
+ a (x)y = 0. (4-18)
4-4 | NONHOMOGENEOUS EQUATIONS: VARIATION OF PARAMETERS 139
Using the language of linear operators, the problem of finding a particular solution
of (4-16) —
which we assume defined and normal on an interval 7 consists of —
n
finding exactly one function in e (7) which satisfies the equation
Ly = h, (4-19)
And with this we have in fact defined a right inverse G for the operator L. Speci-
fically, be described as the (linear) mapping from 6(7) to 6 n (7) which sends
G can
each function h in 6(7) onto the solution of (4-19) which satisfies the initial con-
ditions given above. In this section we shall obtain an explicit formula for G in
terms of a basis for the solution space of the homogeneous equation Ly =
when L is an operator of order two. In the next section these results will be gen-
eralized to operators of arbitrary order, and once this has been done the study of
linear differential equations will have been reduced to the homogeneous case.
(The reader should note that this portion of our discussion is not restricted to
constant coefficient operators.)
Thus we begin by considering a normal second-order linear differential equation
= Kx)
d£ + + (4^2I)
° l(x) ao(x)y >
rfx
* This requirement can be viewed as restricting the domain of L in such a way that
L becomes one-to-one and has an inverse.
140 EQUATIONS WITH CONSTANT COEFFICIENTS | CHAP. 4
This identity, which must hold if (4-24) is to be a solution of (4-21), will obviously
be satisfied if Ci and c 2 can be chosen so that
for all x in /. Thus it remains to show that these equations serve to determine
c x (x) and c 2 (x), and that this can be done in such a way that the function
yi(x) J2OO
suitably choosing the limits of integration, the required initial conditions can also
be satisfied, and the argument is complete. For obvious reasons, this method of
constructing a particular solution for a nonhomogeneous linear differential equa-
tion out of the general solution of its associated homogeneous equation is known
as the method of variation of parameters.
Starting with (4-26), an easy calculation gives
Thus
and if these values are substituted in (4-24), and terms combined, we find that
One reason for calling attention to this expression is that it can be read as the
definition of a right inverse for the linear differential operator L = D2 +
a 1 (x)D + a (x), and, in fact, is the particular right inverse discussed earlier in
this section. For if h is any function in e(7), and if we set
* In the next section it will be shown that K(x, i) is also independent of the particular
basis chosen for the solution space of Ly = 0.
142 EQUATIONS WITH CONSTANT COEFFICIENTS | CHAP. 4
yh = Cx sin x + c 2 cos x
as its general solution. Thus we seek a particular solution of (4-31) in the form
where c^x) and c 2 (x) are determined from the pair of equations
defines a right inverse for L. (The reader should note that at this point we are in
a position to solve all linear differential equations involving D2 + 1, a fact which
* It is standard practice to omit the constants of integration at this point since without
them we still obtain a particular solution of the given equation.
4-4 | NONHOMOGENEOUS EQUATIONS: VARIATION OF PARAMETERS 143
vividly illustrates the economy of using Green's functions.) It now follows that a
particular solution of (4-31) can be obtained by applying G to the function tan x;
i.e., by computing
(*)=
yP (x) = f/ sin (x — t) tan * dt. (4-32)
Jo
This gives
sin t dt
,
cos x / dt
,
Jo Jo cost
= sin x — cos x In |sec x + tan x|.
xy" +y =
on (0, oo ) and (— oo, 0) is c x + c 2 In \x\. Hence the nonhomogeneous equation
xy + y = x + 1 (4-33)
has a particular solution of the form yp = d(x) + c 2 (x) In |x|, where c x (jc) and
c 2 (x) are determined from the equations
, .
c'2 (x)
.
• - =
X
1
—+
*
X
1
[Here h(x) = (x + 1)A since we must divide by x to put (4-33) in normal form.]
Thus
c\(x) = -(*+ 1) In 1*|,
c 2 (x) = x + 1,
and
This gives
=
^ = T+*~ (t + x
)
ln W + (t + x
)
ln W T+^
and the general solution of (4-33) is
x2
y = -r +x + C! + c 2 In Ixl
144 EQUATIONS WITH CONSTANT COEFFICIENTS | CHAP. 4
EXERCISES
1. (D
2
+ l)y = —!— 2. (Z)
2
- D - 2)y = e~ x sir
COS X
2 2x 2
3. (D + 4D + 4)y = xe 4. (Z) + 3D - A)y = x 2ex
2x
2 2 e
5. (AD + 4Z> + \)y = xe~ xl2 smx 6. (Z) + 4)y =
2
2x <N 2
7. (Z)
2
+ 10D - 12)3^ = (g + ,
l)
8. (Z> + 3)
2
>> = (jc + X)e
x
2 2x 2 x 2
9. (Z) - 2Z) + 2)>> = e sin x 10. (4Z> - SD + 5)y = <? tan |
For each of the following differential equations verify that the given expression is the
general solution of the associated homogeneous equation and then find a particular
solution of the equation.
16. D2 + 3 17. D2 - D - 2
18. Z)
2
+ 4Z> + 4 19. 4D - 8D + 5
2
25. Find a Green's function for the operator in Exercise 24, and use this function to
may be solved by setting (xD + 3)y = u, and then successively solving the first-order
equations
(Z) - \)u = ex and (xD + 3)y = u.
Use this technique to show that the general solution of this equation is
c\ ,
C2e .2 n . ~n , x
4-5 | VARIATION OF PARAMETERS; GREEN'S FUNCTIONS 145
27. Use the technique introduced in the preceding exercise to show once more that xe ax
is a solution of (D — a) 2y = 0.
28. Let K(x, t) denote the Green's function (4-30) for initial value problems involving the
operator L = D2 + a\(x)D + ao(x), and assume that L is defined on an interval /
of the x-axis.
(a) What is the domain of K(x, i) in the x/-plane?
(b) Prove that K(x, x) = and Kx (x, x) = 1 for all x in /. [Note: Kx denotes the
partial derivative of K(x, i) with respect to x.]
(c) Show that for each fixed t in I the function <p(x) = K(x, t) is a solution on / of
the initial value problem Ly = 0; <p(t) = 0, <p'(j) = 1.
(d) Use the results of (b) and (c) to deduce that K(x, i) is independent of the particular
basis yiix) and y2(x) chosen for the solution space of the homogeneous equation
Ly = 0.
29. With K(x, t) as in the preceding exercise, show that the function yp defined by
satisfiesthe initial conditions yp (xo) = yp (xo) = for all h in C(/). [Hint: Use
Leibnitz's formula for differentiating integrals,t namely,
r Hx) /• b(x)
*30. Find the Green's function for initial value problems on (0, <x>) for the operator
(n)
y + an _ !(*)/"- " + • • •
+ a (x)y = h(x) (4-34)
yn = c iyi (x) + • • •
+ cnyn (x) (4-35)
yP = ci(*)yi(*) + • • •
+ cn (x)yn (x), (4-36)
where, in addition to the requirement that yp satisfy (4-34), we impose the follow-
ing n — 1 conditions on the unknown functions Ci(x), , c n (x): . . .
c'iJ'i + • • •
+ nyn = c' 0,
^(n-2) + + =
c . . .
ctff-V
for all x in /. If the expression for yp is now substituted in (4-34) and the above
conditions are used, we obtain the additional equation
+ y^- =
l)
c'xyf-
1
'
+ • • • c'n
{
h(x), (4-38)
and, for each x in /, (4-37) and (4-38) may be viewed as a system of n linear
equations in the unknowns whose determinant is W[yi(x),
c[, . . . , yn (x)].
c'n . . .
,
Our earlier reasoning still applies, and we can obtain a particular solution for
(4-34) by solving this system for c[, c'n integrating, and then substituting . . . , ,
6
1
(See Exercise 17.) Hence, just as in the second-order case, the particular solution
may be written in integral form as
_ yi (x)VM + + y
;
MVM ,
( ^42)
4-5 | VARIATION OF PARAMETERS; GREEN'S FUNCTIONS 147
yi(t) •
yn (0
y'i(t) • y'nit)
yf~
2
\t) - yn'
( 2)
(t)
= yi(x) •
yn (x) (4-43)
K(x, t)
yi(t) • yn (t)
y'i(t) •• y'nit)
yf-
2)
(t) •
- y?- 2 \t)
y:~ \t)
1} { l
yr- (o •
••
The function K(x, t) defined here is called the Green's function for the operator
L = Dn + an ^i(x)Dn ~~ 1 + +a • • •
(x) (for initial value problems in the
interval I), and the expression
=
G(h)
r
/
Jx
K(x, t)h(t) dt (4-44)
n U
v(n-l) -
G(h){x ) = GW(*o) G(hy (x ) = (4-45)
yP = Ci(x) + c 2 (x)e~
2x
+ c 3 (x)e* 13 ,
2x
Ci(jc) + c'2 (x)e- + c 3 (x)ex 3 = '
0,
2x
-2c'2 {x)e- + ^c 3 (x)ex 3 = '
0,
Ac'2 {x)e-
2x
+ ic'3 (*K
/3
= ^ •
x
Thus
c'i(x) = - '
~Y
= i*e 2xr(x),
c'2 (x)
f -2x f *xl3 f
- x,3
yv », r(x)
J
dx
'v-.— +
.
^jf-
14 J
e
2x
r(x) dx + ^- J e r(x) dx
-
= / [~i + ^e~ 2{x + l)
^- x t)l3
]r(t)dt.
Alternately, we could have computed the Green's function K(x, i) for the (nor-
mal) operator D 3 + §D 2 — § D, and used (4-41) to express yv as an integral in-
volving K(x, i). Starting with the basis 1, e~ 2x , e
xl3
for the solution space of the
associated homogeneous equation, we then obtain
1 e~ 2t e
m
-2e~ 2t 3e
2x xl3
= 1 e~ e
K(x, t)
2t tl3
1 e~ e
-2e~ 2t 3e
2t
4e~ ±e t,a
T_
e
-5t/3 _ i
e
-2x
e
t/3 _ 2e
xl3
e~
2t
— 2-p-
9e
-5t/3 4 -5(/3
_ — a.
— i
-T 14^
_a_ -2(x-t)
T^
, 9
7e
(.x—t)/3
2
Thus
f
J
[-i ]r«)dt,
The remainder of this section will be devoted to taking a closer look at the Green's
functions for initial-value problems, and to establishing some of their more im-
portant properties. Throughout this discussion we shall assume that
L = Dn + a^xOc)!)"-
1
+ • • •
+ «o(*)
is a fixed linear differential operator on e n (I), and that K(x, i) is the function
obtained above from the general solution of the equation Ly = 0. It then follows
4-5 | VARIATION OF PARAMETERS; GREEN'S FUNCTIONS 149
from the very way in which K{x, i) was constructed (see Exercises 18-21 below)
that
(1) K(x, t) is defined throughout the region R of the xf-plane consisting of all
n n
d K/dx are continuous everywhere in R;
(3) For every x in I, and every h in 6(7),
the function
X
y(x) = ( K(x, i)h(i) dt
Jx
Ly = h;
y(x ) = /M = • • • = ^—"(jco) =
FIGURE 4-1
on I.
These properties are actually sufficient to characterize the Green's function for
initial- value problems involving L in the sense that K(x, i) is the unique function
defined in R which satisfies (1), (2), and (3). This assertion will be proved below
as Theorem 4-3, and is mentioned here only to motivate the following definition.
Definition 4-1. A function H(x, t) is said to be a Green 's function for initial
value problems involving the linear differential operator L if and only if
H{x, t) enjoys the three properties listed above for the function K(x, t).
This said, we proceed at once to give an alternate, and for our purposes much
more useful description of a Green's function for L. For convenience, we shall
denote the various derivatives dH/dx, d 2 H/dx 2 appearing in the following , . . .
argument by x H H
2 .... And with this notation in effect, we have
, ,
H(x, x) = 0,
H\(x, x) = 0,
(4-47)
Hn _ 2 (x, x) = 0,
Hn _ x (x, x) = 1,
and
Hn {x, + t) an _ !(*)#„_!(*, t) + • • •
+ a (x)H(x, t) = 0. (4-48)
150 EQUATIONS WITH CONSTANT COEFFICIENTS | CHAP. 4
Proof. First assume that H(x, i) is a Green's function for L. Then, by definition,
the function
y(x) = f"
H(x, t)h{i)dt (4-49)
for every x in I, and every h in 6(7); see (3) above. We now differentiate (4-49)
using Leibnitz's formula* to obtain
y'(x) =
Jx
(' Hiix, i)h(t) dt + H(x, x)h(x), (4-50)
which reduces to
H(x , x )h(x ) =
when x = x (recall that y'(xo) = 0). But, by assumption, this expression is
valid for all h in 6(7), and hence, in particular, for h = 1. Thus H(x , x ) = 0,
and since x can be chosen arbitrarily in I, we have
H(x, x) = 0,
and
y'(x) = f*
JXq
H (x, t)h(i)
x dt. (4-51)
x
y »(x) = f
JXn
IXq
H (x, f)h{i) dt +
2 Htix, x)h(x),
then
H\(x, x) = 0,
and finally
X
y "(x) =
Jx
f H (x, t)h(t)
2 dt.
Hn—zix, x) = 0,
and
- 1}
yn (x) = f*
Jx
Hn _ x {x,t)h{t)dt.
* Leibnitz's formula is
rHx)
D(X) r 0(,X)
r b(,x)
d_
dx J/a(x)
/(*, t)dt =
J a(x)
¥
OX
(x, dx + f(x, b(x))b'(x) - f(x, a(.x))a'(x).
4-5 | VARIATION OF PARAMETERS; GREEN'S FUNCTIONS 151
X
in
y \x) = [
Jx
Hn (x, t)h(t) dt + Hn-xix, x)h(x),
whence
/ n \x ) = H(x Q xoWxo). , (4-52)
(n
y \x) + tfn.xOc)/"-
1
^) + • • •
+ a (x)y(x) = h(x),
/ n \x ) = Kx ).
This, together with (4-52) and the fact that h and x are still arbitrary, implies
that
Hn —\{x, x) = 1,
and
X
y»>(x) = f Hn (x, t)h(t)dt + h(x).
Jxn
and the fact that this expression holds for all x in / and all h in 6(7) allows us
to conclude that the bracketed portion of the integrand is identically zero (see
Exercise 23). And first part of the proof is complete.
with this, the
As argument needed to show that (4-47) and (4-48)
for the remainder, the
imply that H(x, i) is a Green's function for L is an even more elementary com-
putation than the one just given, and has therefore been relegated to the exercises
(see Exercise 24). |
Among other things, this theorem asserts that for any fixed t in the interval I,
the function
k(x) = H(x, t )
But, as we know, the solution of this problem is unique. Thus the values of
H{x, t) are uniquely determined by the operator L on the line segment consisting
152 EQUATIONS WITH CONSTANT COEFFICIENTS | CHAP. 4
of those points (x, i) in R with t = t (see Fig. 4-2). However, / can be chosen
arbitrarily in /. and we therefore have our main result.
Theorem 4-3. The Green's function for initial value problems on I involving a
linear differential operator L is uniquely determined by L, and hence must
coincide with the function K{x, t) defined by (4-42) or (4-43). In particular,
K(x, i) is independent of the basis for the solution space of Ly = used in
computing it.
to
FIGURE 4-2
Everything that has been said up to this point in our discussion applies to
arbitrary linear differential operators. As might be expected, much more precise
information can be given in the case of operators with constant coefficients, and
we conclude this section with a theorem which describes the Green's functions
obtained in this special case.
Theorem 4-4. The Green's function for a constant coefficient linear differen-
tial operator L can be written in the form k(x — t), where k(x) is the solution
Ly = 0; XO) = /(0) = • = ^— al
,(n-2)/ryk
(0)
_
= 0,
,(n— 1)/A\ _
y»-"(0) = 1.
Proof The function H(x, t) = k(x - t) clearly satisfies the identities (4-47) and
(4-48) of Theorem 4-2, and hence, by Theorem 4-3, is the Green's function for L* |
EXERCISES
Use the method of variation of parameters (without employing Formulas (4-41) through
(4-43)) to find the general solution of each of the following differential equations.
* Note that the assumption of constant coefficients is needed to verify (4-48); see
Exercise 25.
4-5 | VARIATION OF PARAMETERS; GREEN'S FUNCTIONS 153
In Exercises 11 through 16 compute the Green's function K(x, i) for the given operator
(a) by using Formula (4-42) or (4-43) and (b) by applying Theorem 4-4.
19. K(x, i) is defined and has continuous derivatives through order n in the region R
described in the text.
20. The partial derivatives with respect to x of K(x, t) satisfy the identities
X
y(x) = [ K(x,t)h(t)dt
f(x)g(x)dx =
/
Ja
for every g in Q[a, b], then /= on [a, b]. [Hint: Assume /(*o) ^ and use the
continuity of /to obtain an interval (x — 8, x + 5) in which |/(x)| >|/(x )|/2.
Then find a function g in 6(7) for which the above integral is different from zero.]
23. Use the result of Exercise 22 to prove the assertion made in the text concerning the
bracketed term in (4-53).
24. Let H(x, t) satisfy the hypotheses of Theorem 4-2, and the identities given in (4-47)
and (4-48). Prove that for every xq in / and every h in C(/) the function
25. Show that the function H(x, t) = k(x — t) appearing in the proof of Theorem 4-4
satisfies (4-47) and (4-48).
The only values of K(x, which enter into the integration in (4-44) are those for which
t)
the point (x, t) lies R Xo of R shaded in Fig. 4-3. This suggests the possi-
in the subregion
bility of generalizing the notion of a Green's function as follows:
I I
XQ
Ly = h;
y(x )
=*•••= yi»-»(xo) = 0. FIGURE 4-3
In the exercises which follow we explore some of the properties of these functions, and,
in particular, show that under certain additional assumptions they coincide with K(x, t).
26. Let K(x, t) be as described above, and assume that L = D2 + a\(x)D + ao(x).
Suppose, in addition, that
(i) K(x, t), K\(x, t), K^ix, t) are continuous in R XQ ,
(ii) K(x, x) = on /,
(iii) K\(x, x) = 1 on /.
Prove that
K 2 (x, i) + a^xWiix, t) + a (*, t) = 0.
region R XQ .
2 O+
dx 2
X >%-2( + 1 X *)y (4-54)
4-6 I
REDUCTION OF ORDER 155
on the interval (0, oo). Here none of our earlier techniques is sufficient to obtain
y = x 2 c(x).
Then
/ = x 2 c'(x) + 2xc(x),
= x c"(x) + 4xc'(x) +
2
y» 2c(x),
and (4-54) yields
x*c" + (4*
3
+ x b )c' = x,
or
c" + XX
4 4- x
^
2
c'
But (4-56) may be viewed as a first-order equation in c', and as such can be solved
= 41
3
'
(4-56)
v '
by the technique introduced in Section 3-3. Indeed, using the integrating factor
e
f(4+x 2 )/xdx _ ^4^x2/2
we obtain
x2/2
e = [fc! + fxe dx]x- 4 e- x2 < 2
-
= [*i + ^ 2/2]^- 4e x2/2
c(x) = - ^+ k x j x-"e-x212 dx + *a ,
y = x 2 c(x ) = _L + k lX 2 x- 4 e- x2 ' 2
dx + fc 2 x2
J
* The phrase "discovered by inspection" is just a dodge to hide the fact that the process
was one of trial and error.
156 EQUATIONS WITH CONSTANT COEFFICIENTS | CHAP. 4
But since Lu = 0, the first term in this equation vanishes, and we may therefore
view (4-57) as a linear equation of order n — 1 in c'. This is the asserted reduc-
tion in order. In particular, this technique can always be used as it was above
to find the general solution of a second-order equation whenever one nontrivial
solution of its associated homogeneous equation is known.
Example 2. The second-order equation
(D - a)
2
y = (4-58)
has e ax as a solution. Using the above technique to find the general solution we
set
= ax
y c(x)e ,
c"e ax + 2ac'e
ax
+ a 2 ce ax - 2a(c'e
ax
+ ace ax ) + a 2 ce ax = 0.
c = k\X + k2 .
Thus
y= (k lX + k 2 )e ax ,
as expected.
EXERCISES
In Exercises 1 through 6, find the general solution of each differential equation using the
given solution of the associated homogeneous equation.
ai(x) —
dy
ax
+ floW)' = 0,
ai(x) —
dy
dx
+ ao(x)y = h(x).
(1 - x 2 )y" - 2xy' + 2y =
xy" + y' + xy =
has a solution, Jo(x), on the interval (0, °o ) which can be expanded in a power series as
2 4 6
Jokx) = 1 ~ + _
22 24 2 26 2 •
(2!) (3!)
Find a second solution, linearly independent of Jo(x) in 6(0, «> ), of the form
J x[Jo(x)] 2
and then prove that this solution can be written in the form
* This equation is known as Bessefs equation of order zero and will be studied in detail
in Chapter 15.
158 EQUATIONS WITH CONSTANT COEFFICIENTS | CHAP. 4
most efficient way of producing such a solution, and for certain equations the
work involved can be considerably lightened. For instance, it would be pointless
to use variation of parameters to find a particular solution of (D
2 — D + S)y = 3
since the solution yv = § is immediately evident. And even for an equation
such as
(D 2 + 3)y = e
x
yv = Ae x
L x Ly = 0. (4-60)
(D 2 + \)y = 3x 2 + 4, (4-61)
a particular solution of (4-61) can be found among the solutions of the homo-
geneous equation
D\D 2 + \)y = 0. (4-62)
4-7 |
THE METHOD OF UNDETERMINED COEFFICIENTS 159
yP = Ci + c 2x + c3x
2
+ c 4 sin x + c 5 cos x (4-63)
2
yP = Ci + c 2x + c 3x .
2c 3 + ci + c 2x + c 3x
2
= 3x 2 + 4,
ci + 2c 3 = 4, c2 = 0, c3 = 3.
yv = 3jc
2 - 2.
(D 2 - AD + 4)y = 2e
2x
+ cos x, (4-64)
(D - 2)
3
(D 2 + 1);; = 0.
2x 2x 2 2x
y = C\e + c 2 xe + c^x e + c 4 sin x + c 5 cos x,
2 2x
yp = c3x e + c 4 sin x + c 5 cos x.
In this case
y'
p = 2c s xe
2x
+ 2c 3 x 2e 2x + c 4 cos x — c 5 sin x,
j>p = 2c 3 e
2x
+ Sc s xe 2x + 4c 3 x 2e — c 4 sin x — 2a:
c 5 cos x,
2x 2x
2c 2 e + (3c 4 + 4c 5 ) sin x + (3c
5
— 4c 4 ) cos x = 2e + cos x.
160 EQUATIONS WITH CONSTANT COEFFICIENTS | CHAP. 4
Hence
2c 3 = 2, 3c 4 + 4c 5 = 0, 3c 5 — 4c 4 = 1,
and
c3 =1, c4 = -^, c5 = ^V-
2 2x
yp = x e — £$ sin x + ^ cos x.
EXERCISES
Use the method of undetermined coefficients to find a particular solution for each of the
following differential equations.
17. (Z) 3
+ Z) — \)y = sin x + cos x
18. Z> 2 (Z>
2
+ = 1 + 2xe*
1)>>
19. D(D 2 -
\){D - 2)y = x 2 + 2x + 3 - 2c*
(Z) + 5Z) + A)y = 2 cos
4 2
20. a:
Give the form of a particular solution for each of the following equations. The coeffi-
30. £>(Z)
2 - 4)
5>> = (x + 2
l) [(x + 1) + sinh 2x]
>
The reader should note that this equation on the entire x-axis, but is
is defined
normal only on intervals which do not contain the point x = 0. It is one of the
relativelyfew equations with variable coefficients that can be solved in closed
form in terms of elementary functions, and is important because its solutions are,
to some extent, typical of those of a large class of linear differential equations
whose leading coefficient vanishes at the origin.
As we can be converted into a (linear) equation with con-
shall see, Eq. (4-65)
stant coefficients by making the change of variable u = In x, and hence can be
solved by the methods introduced in this chapter.* Although this reduction can
be effected in a routine fashion, it is illuminating to consider it from the point of
view of linear operators by introducing the transformation T: e*(— oo, oo) —
e w (0, oo ) defined by
(Tg)(x) = g(lnx) (4-66)
n
for all g in e (— oo , oo). Thus Tmaps the function x onto In x, sin x onto sin (In x),
with (X~ f%x) =
l
etc., and is obviously linear. More important, it is invertible,
f(e
x
) for all fin e n (0, oo), and hence is a one-to-one linear transformation mapping
W
6 (— oo, oo ) onto e n (0, oo ). From this it follows that the problem of solving the
equation Ly = on the interval (0, oo), with L = xn D n + an _ 1 xn ~ 1 Z)n_1 +
• • •
+ a , is equivalent to the problem of finding all functions g in e M (— oo, oo)
such that LTg = 0. In other words, we must find the null space of the transforma-
tion LT.
To this end we begin by computing the various products DT, D 2 T, . .
.
, as
follows:
TD(D - l)g,
* In the following discussion we shall restrict our attention to the interval (0, oo ). On
(— oo , 0) the change of variable u = In (—x) must be used.
162 EQUATIONS WITH CONSTANT COEFFICIENTS | CHAP. 4
and, in general,
Thus
x
k
D kT = TD(D - 1) • • •
(D - k + 1),
L = xn Dn + a^!* 71
-1 ^- + 1
• • •
+ a x xD + a ,
(4-68)
we have
IT = TL, (4-69)
Z = Z)(D - 1) • • •
(D - n + 1) + an - X D(D - 1) • • •
(D - n + 2)
+ •
+ at/) + fl - (4-70)
Equation (4-69), together with the fact that T is one-to-one, implies that the
null space of the transformation LT coincides with the null space of L. This
establishes our contention that (4-65) can be reduced to an equation with constant
coefficients and also allows us to describe the solution of any (homogeneous)
Euler equation on the interval (0, oo) as follows:
y = <?i>>i(lnx) + • • •
+ cnyn (lnx), (4-71)
where y\(u), . . .
, y n (u) are a basis for the solution space of the constant
coefficient equation Ly = 0, L deduced from L by (4-70), and c x , . . . , cn
are arbitrary constants.
Finally, to remove the restriction on the interval, we note that for each non-
negative integer k,
x
k
D kg(-x) = A-DW*) = k
(-x) k D g(-x).
4-8 I
THE EULER EQUATION 163
the general solution y(x) of (4-72) is determined by the roots a\, a 2 of the quad-
ratic equation
m(m — 1) + ax m+ a = 0, (4-73)
a
y(x) = Cl |x|
ai
+ c 2 \x\ *;
(ii) if «x = a2 = a, then
y(x) = |*|
a
[ci sin (b In |*|) + c 2 cos (b In |*[)].
m(m — 1) + 2m — 6 = 0,
and has 2 and —3 as roots. Thus the general solution of (4-74) on (0, 00) is
-3
y = ci*
2
+ c 2* ,
and since the given initial conditions y(l) = 1, /(l) = imply that c x = f,
c2 = f, the required solution is
2 -3
y = f-* + I* -
Example 3. In Example 2 of Section 4-4 we said that the general solution of the
equation
xy" +/ = (4-75)
is
y = ci + c 2 ln |*|. (4-76)
To prove this assertion we multiply (4-75) by *, and obtain the second order
Euler equation x 2y" + xy' = 0. Here the indicial equation is m2 = 0, and
(4-74) now follows from the results of Example 1.
The general solution of a nonhomogeneous Euler equation Ly = h, defined on
an interval I not including the origin, can be obtained by using variation of param-
164 EQUATIONS WITH CONSTANT COEFFICIENTS | CHAP. 4
eters or the Green's function for L (see Example 2, Section 4-4). But here too the
problem can be reduced to one involving a constant coefficient equation, and is
usually easier to solve in this form. To do so we again use the linear transformation
T, this time to rewrite the equation
Ly = h (4-77)
as
Lr(r~V) = T(T~ l
h).
jC(r-V) = T~ l
h, (4-78)
of (4-77) on I.
u 2u
D 2y = e + e . (4-80)
2
x
y = x + -r
4
+ Ci + c 2 In x.
EXERCISES
7. x 3y'" - \lxy' - ly =
- 2x 2y" 8. x 3y"' - *x 2y" + 6xy' - 6y =
9. /" + 4x
jc
3 - 2y = 2/'
10. x /"
3
+ 4x 2/' - 8*/ + 8y =
11. *V + 6*V" + 7x 2/' + */ -
iy)
j;=
12. Let S be the linear transformation from © n (0, °o) to C n (— °o , 0) defined by
(Sf)(x) - /(-*).
(b) Use the results of (a) to prove that L(Sy) = if and only if Ly = 0, L as above,
and hence deduce that >>(— x) is a solution of Ly = on (— », 0) if and only if
14. Prove that (4-67) is valid for all positive integers k. [Hint: Use mathematical
induction.]
15. (a) Let T be the linear transformation defined in (4-66), and let L = D -j- ao.
Prove that
TLT' 1 = xD + ao.
L = xn D +
n
a,,-!*"- /)"- 1
1
+ • • •
+ a\xD + a
can be written as a product of operators of this form of orders one and two. [Hint:
Let L be the constant coefficient operator associated with L, and write
L = L1L2 '
' 'Lk,
where Li, i = 1, . .
.
, k, is a constant coefficient operator of order one or two.
Show that
L = (TLiT-^iTUT- 1 ) • •
(TUT- 1 ),
and use the results of Exercise 15.]
17. Use the results of Exercises 15 and 16 to factor each of the following operators,
(a) x2D2 + 2xD - 2 (b) x2D2 + D- 9
(c) x3 D3 - 2x 2 D 2 - llxD - 1 (d) x3 D3 - 3x 2 D 2 + 6xD - 6
(e) x4 D* + 6x 3 D 3 + lx 2 D 2 + xD - 1
future size of a population under the assumption that its rate of growth is known,
and arise in such diverse situations as the radioactive decay of matter and the
increase of bacteria in a culture. In such problems it is frequently assumed that
the rate of increase (or decrease) in population at time / is proportional to the
number of individuals present at that time.* Then if y(t) denotes the number of
individuals present at time t,
%-ky (4-80
for an appropriate constant k, and it follows that y obeys the well known law of
exponential growth
kt
y = ce ,
(4-82)
c a positive constant.
Hypotheses which result in something more realistic than prolonged exponential
growth are, of course, also used. One of the simpliest of these is the assumption
that the supply of necessities for life is constant and sufficient to support a total
population P. This implies that the factor of proportionality in (4-81) depends
upon y and P, and approaches zero as y approaches P. Thus y must now satisfy
a nonlinear equation of the form
= (4-83)
f t
f(y,P)y,
dy d 2
pt
^
at
+ Pu = 1, and u = ~ (1+
f
ce~ ).
food, living space, etc. For radioactive decay, short of an atomic explosion, the assump-
tion is in complete accord with experimental fact.
4-9 ELEMENTARY APPLICATIONS 167
Thus
y =
1 + ce-rt
= PPo
y P + (P - P )e~ Pt
In Fig. 4-4 we have sketched a number of these curves for various values of P .
FIGURE
F = fema. (4-84)
The reader should note that in 3-space this equation can be rewritten as a system
of three scalar equations
Fx = kmax ,
Fv = kmciy,
Fz = kmaz ,
For convenience the physical units are usually chosen so that k = 1 in these equa-
tions, and in the future we shall always assume this choice has been made.
As it stands, Eq. (4-84) assumes that m is constant. A slightly more sophisticated
formula which avoids this assumption is
F - (mT) -
(4-85)
St
168 EQUATIONS WITH CONSTANT COEFFICIENTS |
CHAP. 4
mL = ~ mgsi
sin < <
Jt( f) <p, <P
(4-86)
Ifwe now assume that <p is small so that sin <p « <p, the
above equation may be replaced by the linear equation
(4-87)
= FIGURE 4-5
<p(t) c x sin \/g/L t -+ c 2 cos
Finally, if we assume that the pendulum was initially released from rest at an
angle from the vertical, then <p(0) = <p <p'(0) = 0, and the corresponding
<p ,
particular solution of (4-87) is <p(f) = <p cos y/g/Lt. The student should recog-
nize this as the equation for simple harmonic motion whose period of oscillation
is 2-Ky/L/g.
EXERCISES
1. The half-life of a radioactive substance is defined as the length of time required for
half of the atoms in any sample of the substance to decay.
(a) The half-life of radioactive carbon 14 is 5600 years. Find the amount of carbon 14
remaining in a sample of amount xq at the end of t years. (Assume that the rate of
decay is proportional to the amount present.)
(b) If 90% of the carbon 14 in a given sample of carbon has decayed, find the age
of the sample.
4-9 ELEMENTARY APPLICATIONS 169
dy
= (*] k 2 )y.
dt
Determine k\ and & 2 if it is known that the colony doubles in size every 24 hours,
and would have its size halved in 8 hours were there no births.
5. If the population of a certain colony of bacteria doubles in 40 hours, how long will
it take for the population to increase 10-fold?
6. The population of a country increases 3% per year; its present population is 190
million.
(a) How many years will elapse before the population reaches 250 million?
(b) What will the total population be in 5 years? In 50 years?
7. Solve Exercise 6 under the additional assumption that the country admits 200,000
immigrants each year.
8. What is the annual rate of interest being paid on an account where interest is con-
tinuously credited at the rate of 5%?
9. Assume that evaporation causes a spherical raindrop to decrease in volume at a
rate proportional to its surface area, and find the length of time it will take a drop
of radius ro to evaporate entirely.
10 It can be shown that a body inside the earth is attracted
toward the center by a force which is directly proportional
to the distance from the center. Find the equation of mo-
tion of a ball dropped into a hole bored through the center 1
h
of the earth. When will the ball reach the opposite end of
the hole?
1
11 Liquid is U-tube as shown
oscillating without friction in a
in Figure 4-6. If the liquid was initially at rest with one
side ho inches higher than the other, determine the subse-
quent motion of the liquid by finding h as a function of
time. Show that the period of oscillation is rs/lL/g,
where L is the total length of liquid in the tube, and g the FIGURE 4-6
acceleration of gravity.
12. The angular momentum of a rotating body is given by the formula T = la, where T
is the torque applied, / the moment of intertia of the body about its axis of rota-
tion, and a = d2 <p/dt 2
is the angular acceleration. Suppose that a bob on a twisted
wire resists the twisting force with a torque k<p. Find / as a function of time if k = 1
and the period of rotation is half a second.
170 EQUATIONS WITH CONSTANT COEFFICIENTS |
CHAP. 4
Webegin by considering the R-L circuit shown in Fig. 4-7, where the symbol
— |
—
denotes a constant source of voltage E such as might be supplied by
|
a battery, and the arrows indicate the direction of the flow of current. By Kirchhoff's
second law we have
EL + ER - E = 0,
the negative sign being due to the fact that the voltage rises across the battery.
Hence, by (4-88),
and if we now assume that the circuit was energized at time t = 0, the flow of
current is obtained as the solution of the initial value problem
L^
at
+ Ri = E, i(0) = 0. (4-89)
E „ - _(JJ/L)t>
= Z
,•
(1 e-^ ^), 1
(4-90)
and we see that the current flow in this circuit is a sum of two terms, a time
* A closed curve is said to be oriented if a positive direction has been assigned for
traversing it.
4-10 SIMPLE ELECTRICAL CIRCUITS 171
whose effect diminishes with time. (See Fig. 4-8.) Since the inductance L appears
only in the latter term it follows that a simple R-L circuit operating under a con-
stant impressed voltage will eventually behave very much as if the circuit were
noninductive. The length of time required for the transient term to become
negligible sometimes called the delay time of the circuit, and furnishes a measure
is
Lj + t
Ri = E sin cot, i(0) = 0. (4-91)
= (aEL —fRiTAt
-(RIL)t . E —
i e ,
(R sin cat (aL cos (at) (4-92)
R 2 + w 2L 2 VR 2 + C0
2 L2
(see Exercise 2), which may be rewritten
= °> EL p -(RIL)t
i + = sin {(at — a) (4-93)
Z2
by setting R = Zcosa and (aL = Zsina, where Z= y/R 2 + « 2L 2 . Again
the flow is the sum of two terms, it and /„ the first of which is transient and dies
out as t increases. The second,
FIGURE 4-9
Finally, since
sin a = coL
and
,
cos a = R
.
_ dq
(4-94)
1
~ It
FIGURE 4-10
the voltage drop across a capacitor is (l/Qq, and Kirchhoff's second law leads to
the equation
d2q dq 1
L
^+R&+ c q E sin oit (4-95)
governing the accumulation of charge q(i) on the capacitor in the circuit. Thus
the behavior of this circuit can be determined as soon as the initial values of q
and i = dq/dt are known.
4-10 SIMPLE ELECTRICAL CIRCUITS 173
2L
± U
LC
\\2V
of the auxiliary equation
- .1
.2
Lm' + Rm
.
+ ^= 0.
\2L/ LC
L(
is positive, negative, or zero. However, it is easily seen that no matter what initial
ditions #(0) and i(0) imposed on the circuit, and depends only on R, L, C, and
E sin (at. [Why?] This observation
is in agreement with our intuition which
suggests that the long-range behavior of such a circuit ought to depend only upon
itscomponents, and not upon their state at time = 0. t
We now find the steady-state behavior of this system under the assumption
that R 5^ 0. In this case E sin cot cannot be a solution of (4-95) for any value of
E, and hence the equation has a particular solution of the form
When this expression is substituted in (4-95) and the resulting identity solved for
Ci and c 2, we obtain
Ci = E<*L- (1/coC)]
co{R 2 + [coL - (l/coC)]2}
c2 =
ER
co{R2 + [coL - (l/coC)]2}
7 = coL - l
z2 2
= * + (<^-sb)
CO C
the particular solution q s may be written
qs= -^
El .
Sinco
ER
'-^ COSC0 ''
174 EQUATIONS WITH CONSTANT COEFFICIENTS CHAP. 4
Z = VR 2 + [wL - (1/coC)] 2
in an R-L-C circuit depends upon o>, the maximum amplitude E/Z of the steady-
state current /, on co. It is clear that for fixed values of L and Cthe
also depends
quantity E/Z is a maximum when coL — 1/coC vanishes, i.e., when co = l/y/LC,
and for this value of «, E/Z = E/R. Thus, if we plot E/Z, the amplitude of is ,
as a function of co, the graph will attain its maximum value when co = l/y/LC
Furthermore, this maximum increases with decreasing R, as shown in Fig. 4-11.
Physically, these observations tell us that the circuit may react quite differently to
input voltages with different frequencies differs from
co. The more the frequency
l/y/LC, the smaller the amplitude of the steady-state current becomes, and hence
the voltage drop across the various components of the circuit is small. By minimiz-
ing its resistance, such a circuit can be made highly selective, in the sense that it
discriminates very sharply against inputs whose frequency differs from the circuit's
natural or resonating frequency l/\/LC.
Thus if a mixture of sinusoidal input
voltages of the same magnitude is applied
to an R-L-C circuit of high selectivity, so
that the impressed electromotive force is
y~l E sin co 4,
EXERCISES*
1. Find the current flow in a simple R-L circuit under a damped sinusoidal voltage
Eoe"01 sin bt, with Eo, a, b constants, a > 0. [Assume that at time t = 0, /(0) = z'o.]
2. Verify Eq. (4-92) of the text.
d*q 1
Solve this equation under the initial conditions i(0) = 0, #(0) = 0.03 coulombs.
Sketch the graph of the solution.
5. Add an inductance to the circuit in Exercise 3 and solve the resulting differential
equation, considering separately the three cases
Sketch the graphs of the three types of solutions obtained. (In this situation the
resistance R corresponds to a mechanical damping force, and these three cases yield,
respectively, overdamping, critical damping, and underdamping.)
6. Find the current flow in a simple R-L circuit under a constant voltage given that
R = 40 ohms, L = 8 henries, and that E = volts and i = 10 amperes when
t = 0. At what time will / = 5 amperes?
7. If the resistance is removed from the R-L-C circuit discussed in the text, q satisfies
the differential equation
v2
dq i
l
r _i_ n E sin cot.
Solve this equation in the two cases
(i) co * (ii) co =
* The units of measurement used in the following exercises are resistance in ohms,
inductance in henries, capacitance in farads, current in amperes, charge in coulombs.
176 EQUATIONS WITH CONSTANT COEFFICIENTS | CHAP. 4
and discuss the behavior of the solution in each case. Which of these cases exhibits
the phenomenon of resonance? [Note that in the resonant case the voltage drop
across the capacitor oscillates with the same frequency as the input voltage
(is sin t)/VLC. Its amplitude, however, is unbounded. This is the limiting case of
5-1 INTRODUCTION
In this chapter we shall for the first time make full use of the idea of a linear
operator and its inverse in solving initial-value problems involving linear differen-
tial equations. By contrast to the rather pedestrian methods developed in the last
chapter, our present investigations will yield an extremely efficient technique for
handling such problems. In addition, they will give a much deeper insight into the
role which operator theory plays in applied mathematics, and will serve as an
excellent introduction to a general method which will be used later to analyze
more difficult problems.
The particular linear transformation which we now intend to study is an integral
operator <£ known as the Laplace transform. Before giving the definition, however,
we introduce the notion of a piecewise continuous function, which will be needed
when we describe the domain of £.
Informally, a real valued function is
and the "square wave" function shown in Fig. 5-1 are piecewise continuous on
any finite interval of the f-axis. On the other hand, neither tan t nor sin \/t is
piecewise continuous on [0, x/2]; the first because of its behavior near w/2, the
second because it oscillates in such a manner that it does not approach a limit
as t — > (see Exercise 10).
* Note that only one of these limits is relevant when xo is an endpoint of the interval.
177
178 THE LAPLACE TRANSFORM CHAP. 5
For our purposes the essential fact concerning piecewise continuous functions
on a finite interval is that they are integrable. Indeed, if/ is piecewise continuous
on [a, b], with discontinuities at x X\, xn and possibly at a and b as well, , . . . , ,
[0, t ], t > 0, and for the sake of brevity shall say that such functions are piece-
wise continuous on [0, oo). Finally, the set of all such functions is obviously a
real vector spaceunder the definitions of vector addition and scalar multiplication
given in Chapter 1 (see Exercise 5).* We shall have more to say about this space
in the sections which follow.
EXERCISES
1. Which of the following functions are piecewise continuous on [0, oo)? Give reasons
for your answers.
t + 1 ... t - 2 i/t
(a) e (b) ln(r + 1) (c)
t - 1
(d)
fi - t - 2
(e) e
,
(a)
.
——
sin t
' na
...
positive integer
1
iff n = 1,2, an inte 8er
(d) fit) = n (e) f(i) = |
'
f
-J 1 otherwise
11 otherwise
3. Prove that for any real number a, e atf{t) is piecewise continuous on [0, oo) whenever
/is.
4. Prove that the product of two piecewise continuous functions on [0, oo ) is piecewise
continuous.
5. Prove that the set of all piecewise continuous functions on an interval / is a real
vector space under the "usual" definitions of addition and scalar multiplication.
\t - 1, 1 < t < 2
7. f(t) = cos \vt\
- i t, < t < %
8. /(,) =
(
Jl
lr
-
-
',
1, 1
<
<
/
/
<
<
1
2
9 m =
[
J _
(a_f
4 ,2
?
g/ + 3
i
<
f<,<-2
f <
o
10. Show that lim^ + (sin 1//) does not exist. [Hint: Sketch the graph of sin \/t in
the interval (0, oo ).]
11. Does lim _ o+ < > it sin 1//) exist? Is t sin 1/7 piecewise continuous on [0, oo)?
r e-
./o
et
fit)du (5-1)
8t
£[cos at\is) = / e~ cos at dt
Jo
rh
= lim / e~ cos
8t
at dt
< ->» J o
-st
= lim | 2
e
— g (a sin at — s cos af)
- st t0
re
Jo
fit)dt= lim f e-
•'°
st
fit)dt,
<o-»°°
and is said to converge for a particular value of s if and only if this limit exists.
180 THE LAPLACE TRANSFORM | CHAP. 5
.— sl
st
e- fii)dt (5-3)
Jo
exists for all t > 0. This can be accomplished by demanding that / be piecewise
continuous on every form [0, 1 ], t > 0, for then the integrand in
interval of the
(5-3) will be piecewise continuous, and the integral will exist. (See Exercise 4,
Section 5-1.) Piecewise continuity by itself, however, is not enough to guarantee
the existence of £[f] since (5-3) must also converge as t —> oo for at least one
value of s. One way of assuring this convergence is to require that f(t) be "domi-
st
nated" by some exponential function, thus in effect demanding that e~ f(t) ap-
proaches zero rapidly as t increases. To make this notion precise we lay down
the following definition.
at
|/(0| < Ce (5-4)
for all t > 0.*
In a moment we shall prove that the Laplace transform of any piecewise continuous
function of exponential order does in fact exist, but first some examples.
The constant function f(t) = 1 is of exponential order, as can be seen by
setting a = 0, C = 1 in (5-4). So too —and this is important —are the functions
n at n at n at
t , e , sin bt, cos bt, t e sin bt, t e cos bt,
familiar from the study of constant coefficient linear differential equations. For
n ai
instance, the proof that t e cos bt is of exponential order goes as follows: If
a > 0,
n at
t e cos bt
?2at g2at gat
* This inequality need only be satisfied, of course, at those points of the non-negative
/-axis where /is defined.
5-2 I
DEFINITION OF THE LAPLACE TRANSFORM 181
and L'Hopital's rule shows that this expression tends to zero as t — > oo. In
n at
particular, it is eventually less than 1, and hence \t e cos bt\ < e
2at
for
sufficiently large values of /. Thus there exists a constant C > such that
n at
\t e cos bt\ < Ce 2at for all t > (see Exercise 14 below). If a < the proof
is even easier, for then
n at n
\t e cosbt\ < t ,
n
and^he inequality t < l
e for large values of / implies the existence of a constant
n
C> such that \t e
at
cos bt\ < Ce* for all'/ > 0.
t
2
lim ^= lim e
<( '" a)
= <x>
for all a.
This having been said, we now prove the theorem which justifies introducing
functions of exponential order in the first place.
- 8t
r
Jo
e f{t)dt
—(8—a)t
e- s \Ce at )dt = C e dt
o Jo
-a)t
= lim ^—[1 - e-
{s
°]
= c .
f
if s > a,
st
and the comparison theorem implies that Jq e fif) dt exists for all s > a. |
On the strength of this result we can assert that the domain of definition of the
Laplace transform of a piecewise continuous function of exponential order always
The student who has not already met this theorem may be willing to accept it when
*
we point out that it is the analog for integrable functions of the comparison test for the
convergence of infinite series. For a proof see Appendix I.
182 THE LAPLACE TRANSFORM | CHAP. 5
with the possible exception of the point s itself, the domain of definition of <£[/]
is the open interval (s , oo), and for this reason s is known as the abscissa of
l
convergence of the function/. (Note that for certain functions such as e~ or 0,
s may be — oo .)
The Laplace transform of any function of exponential order exists this much ;
we know. But what about the converse? Is it true that a function whose Laplace
transform exists is necessarily of exponential order? The answer is no, and, as a
matter of fact, the function l/y/1 has a Laplace transform even though it is not
of exponential order. All of this is by way of saying that the set of functions
possessing Laplace transforms is larger than the set S of functions of exponential
order. How much larger we shall not say, for it is no easy matter to specify the
domain of £ precisely. Fortunately the set of functions of exponential order
contains all of the functions which arise in applications, and as such is large enough
for most purposes.
EXERCISES
Compute the Laplace transform and abscissa of convergence of each of the following
functions.
If, t > 1
14. Let /be piecewise continuous on [0, oo), and suppose that there exist constants C
and a such that |/(0 1 < Ce at whenever t > to > 0. Prove that /is of exponential
order.
15. Prove that the product of two functions of exponential order is of exponential order.
* Recall that b a lower bound of a nonempty set S of real numbers if and only if
is
.in, m = 0.
18. Is the function t* of exponential order on [0, <»)? [Hint: Use Exercise 16 and the
identity t
l
= e<
ln
'.]
19. Another version of the integral comparison theorem stated in the proof of Theorem
5-1 is the following. If /and g are integrable on [a, 1], < a < 1, and if \f(t)\ <
g(t) whenever < t < 1, then fl f(t) dt exists whenever fl g(t) dt exists. Use this
result to prove that 1/V7 has a Laplace transform.
— »t
[Hint: /
e
—-dt=\ ^—dt + ^—
e
dt.]
JO x/t J0 -x/t Ji
20. Let / be a function of exponential order, and let ao be the least real number such
that for some constant C,
1/(01 < Ce«<
for all a > ao-
(a) Show that ao > so, the abscissa of convergence of /.
(b) Show that there exist functions for which ao > .so. [Hint: Consider the function
= an mteger,i
f(t) r *f * *s
|o otherwise. J
Let S denote the set of all piecewise continuous functions of exponential order,
viewed as a real vector space under the usual definitions of addition and scalar
multiplication, and let J denote the set of all real valued functions defined on
intervals of theform (s oo) or oo), -so >: — <»• Then JF too can be made into
, ,
a real vector space provided we modify the addition used heretofore in function
spaces to accommodate the fact that the members of JF are not all defined on the
same interval. Specifically, if/ and g are any two functions in SF,/ + g is defined
to be the function whose domain is the intersection of the domains of /and g,
and whose value at any point s in that intersection is f(s) + g(s). Then, with
scalar multiplication as usual, 3" is a real vector space (see Exercise 1).
1 84 THE LAPLACE TRANSFORM | CHAP. 5
£[y] = <p(s)
can be solved uniquely for y when <p is given, and by now he should realize that
this is not an idle question. As in the discussion of the linearity of £, there is a
trivial difficulty which prevents us from giving an affirmative answer. For if/
and g are functions in 8 which differ only at their points of discontinuity, then
£[/] = £[#] even though/ ^ g. But two such functions are "very nearly" identi-
cal, and should this be the worst that can happen we would certainly be justified
in asserting that for all practical purposes £ is one-to-one. The following theorem
guarantees that it is, and for this reason is one of the most important results in the
theory of the Laplace transform.
* Recall that two functions are equal if and only if they have the same domain and
take the same value at each point in this domain.
t Strictly speaking, we should at this point replace $7 by the vector space 5* whose
dements are equivalence classes of functions in $ as determined by the above process of
identification (see Exercise 3). Then the mapping £* 8 —> JF* defined by £*[/] = £[/]*,
:
where £[/]* denotes the equivalence class in JF* containing the function £[/], is the
/linear transformation in question. It is common practice however to ignore this distinc-
tion and simply speak of £ itself as being linear.
5-3 I
THE LAPLACE TRANSFORM AS A LINEAR TRANSFORMATION 185
-1
<£ LV] = y if and only if £[y] = <p. (5-6)
Proof Indeed, in proving Theorem 5-1 we saw that there exist constants C and
a such that
i
£ wi < j&-a
for all s > a, and the desired result follows by taking the limit as s — > oo. |
On the strength of this theorem we can assert that such functions as 1, s, sin s,
and s/(s + 1) do not have inverse transforms in 8, since none of them approaches
zero as s — > oo
EXERCISES
1. (a) Prove that the set 8 of piecewise continuous functions of exponential order is a
real vector space under the usual definitions of addition and scalar multiplication,
(b) Using the definition of addition given in the text, prove that the set 5 is a real
vector space.
*2. Let /and g belong to $F, and define /~ g if and only if f(s) = g(s) on some inter-
val of the form so < s.
(b) Exhibitan equivalence class of functions in JF which does not contain the Laplace
transform of any function in 8.
(c) Can an equivalence class in If contain two different functions which are both
Laplace transforms of functions in 8? Why?
*3. Let SF* denote the set of all equivalence classes of functions in JF under the equivalence
relation defined in Exercise 2.
(a) Give an appropriate definition of addition and scalar multiplication for elements
of 5* so that JF* becomes a vector space.
(b) Define <£*: 8 —> JF* by <£*[/] = <£[/]*, where <£[/]* is the equivalence class in
JF* containing the function <£[/]. Prove that <£* is a linear transformation.
4. Prove that lims ^ w 5£[/] is bounded for any function /of exponential order, and
then use this result together with Theorem 5-3 to deduce that <£ -1 [$a ] does not exist
for any a > — 1. [Hint: See the proof of Theorem 5-3.]
£ tcos = ~7
"'3
52 + g2
' s > °- (5 )
In like fashion one can produce an almost endless list of elementary formulas,
among which are
£[1] = i . s > 0, (5-8)
n = ~^n
£[* ] ' s > 0, na non-negative integer. (5-11)
For example,
- 8t at
£ [eat ] = f" e e dt
Jo
to -a)t
= lim f e-
(s
dt
t —»oo JO
— s > a.
s a
* Recall that 0! = 1.
5-4 I
ELEMENTARY FORMULAS 187
This proves (5-9), and (5-8) can be obtained from it by setting a = 0. Formulas
(5-10) and (5-11) will be established presently, and a more comprehensive list is
Theorem 5-4. Let f be continuous on (0, oo), and suppose thatf is piecewise
continuous and of exponential order on [0, oo). Then
£[/"] = s
2
£[f] - sf(0+) - /'(0+), (5-13)
(n)
= n - n~ 1 +) - n~ 2 +) - ... - / (n - 1)
£[/ ] s £[f] 5 /(0 s f'(0 (0+).
(5-14)
= s£[f] + e- 8tf(t)\; ,
and the proof will be complete if we can show that e~ stf(t)\o = —/(0 +).
To this end we note that since/is of exponential order, e~stf(t) — as t —> oo
whenever s is sufficiently large. Thus e~ stf(t)\o vanishes at its upper limit, and,
taking account of the fact that / may have a jump discontinuity at the origin,
we have
e-8tf(t)\o = lim e~ 8tf(t) - lim e~ 8tf{t)
<->» t->o+
—St.
= -lim e- 8t
f(t)
<-*o+
= ~/(0 + ),
as required.
188 THE LAPLACE TRANSFORM | CHAP. 5
Formulas (5-13) and (5-14) can now be established by repeated use of (5-12),
and we are done (see Exercise 28).* |
Example 1. Let f{t) = — (l/a) cos at. Then f'(t) = sin at, and so, using
(5-12) and (5-7),
s 1
£[sin at] = £[cos at] -1
a a
£(*)
+
+
a \s 2 a 2/
!
a
s > 0.
s2 + a2
Example 2. Since Dn t
n = n\, (5-8) implies that
- -1 - -
£[D n t n ] = n n
s £[t ] s
71
• • • •
= n n
s £[t ].
n n =
Hence s £\t ] nl/s for every non-negative integer n, and
With the mechanics of using the differentiation formulas now out of the way,
we are in a position to illustrate the use of Laplace transforms in the solution of
an initial-value problem. Our example must perforce be a simple one since our
list of transforms is still rather limited. Nevertheless it will illustrate all of the
essential steps of the technique.
^- y = 1; y(0) = 0, /(0) = 1.
We begin by applying the operator £ to both sides of the given equation to obtain
(Note that this step depends upon the linearity of <£.) Using (5-13) and (5-8), this
equation may be rewritten
2
s*£[y]
r - 1 - £{y] = i
Thus
- 15)
^-icrh)' (5
and to complete the solution we must find a function whose Laplace transform is
given by this equation.To do so we use the method of partial fractions to rewrite
(5-15) as
m =
5—1
1
—1
S
j
y = e
l
- 1.
The above example illustrates the way Laplace transforms are used to solve
initial-value problems. In general, if we are given an wth-order linear differential
equation Ly = h(t) with constant coefficients and initial conditions y(Q) = y
= in ~ l) =
,
y'(0) yi, . .
. , y (0) yn _ 1} then (5-14) can be used to convert this initial-
value problem into an operator equation of the form
£[y] = <p(s),
y = £-%],
and the problem has been solved, provided £~ [<p] can be explicitly found. The 1
reader should appreciate that this argument depends upon two facts: first, that
the unique solution of every such problem is of exponential order, so that the
existence of £[y] is assured from the outset, and second, that the equation
-1
y = £ 0>] has a unique continuous solution. But both of these facts have already
been established; the first in the paragraph following Definition 5-1, the second
by Lerch's theorem.
We conclude this section by establishing the integral analogs of Formulas (5-12)
and (5-14) which are frequently useful in computing transforms and their inverses.
More generally
r rt
£ f(X)dX
fa"' ia
/<*><**••<**] = h £[f] ~hf
n times
ra rt ra rt rt
-M
Sn L
Jo J a
mdxdx s J
/
Ja
•..//(*)<&•«&.
Ja
(n — 1) times
(5-17)
Proof. The proof is based upon the observation that if /is of exponential order
then so is & f(x) dx (see Exercise 19). For using integration by parts with
rt
u (t) = /
f(x) dx and dv = e~ st dt,
we have
£ \f f(x)dx\ = I e-
8t
f f{x)dxdt
rt 00 /•»
= -\e- 8t
S Ja
f(x)dx +\S J o
e-
8t
f(t)dt.
But since J* f(x) dx is of exponential order, the first term in this expression tends
to zero as t —> oo provided s is sufficiently large, and hence
rt -, r0
r
f(x)dx\ = ijE|/l + f(x)dx.
£[ja I Ja
Except for obvious notational changes this is (5-16). Equation (5-17) is estab-
lished by iterating this result. |
f(x)dx^ = - (5-18)
£[f],
£[jo s
'" dX =
[f'"f ^^^
£ 7n
£W- (5-19)
\
Example 4. Since J*
'
cos ax dx = (I /a) sin at, (5-18) gives
-£[sina*l = -£[cosaf].
a s
5-4 I
ELEMENTARY FORMULAS 191
Thus
£[sin at] = - £[cos at]
s
= *( \
s \s 2 + a2 /
a ^
' s > n0.
s2 + a
1
Example 5. Use the integration formulas to compute £[te ].
Since
x
f xe dx = te
l
- e
l
+ \,
Jo
(5-18) gives
£[te
l
- e
l
+ 1] = -£[te'].
Site
1
] - £[e
l
] + £[1] = - £[te'].
Hence
**1 - 731 + i
= ^*1
and it follows that
£ lV] =
(5
J 1)2
' *> L
(The reader should note that this result can also be obtained by using the differen-
tiation formula.)
EXERCISES
Find the Laplace transform of each of the following functions. (In those problems
where it appears, a denotes a constant.)
1. sin {t + a) 2. (t + a) 2 3. (/ + a) n , n a positive integer 4. sinh at
5. cosh at 6. fV 7. sin
2
at 8. t
2
cost
9. t sin It 10. (2/ - 3)e« 2)/3
(5
2
+ 9 )(5 2 + 1)
12. Compute £[e' sin t]. [Hint: Use the differentiation formula.]
13. Compute the Laplace transform of the function u a (i) defined by the formula
1, / > a.
192 THE LAPLACE TRANSFORM I CHAP. 5
14. Compute the Laplace transform of the function whose graph is given in Fig. 5-3.
[Hint: Recall the formula for the sum of a geometric series.]
FIGURE 5-3
15. Suppose that the function /is continuous on (0, oo), and that its derivative /' is
of exponential order. Integrate the inequality
\~ stf\t)dt,
f
Jo
where /' was of exponential order. Prove that this was legitimate. (The problem
consists of investigating the behavior of this integral at zero and at the points of
discontinuity of /'.)
17. During the proof of Theorem 5-4 it was asserted that e~ stf(t)\o vanishes at its
19. Prove that J*j f(x) dx, a > 0, is of exponential order whenever /is of exponential
order.
4ff Ja Ja
f(x) dx dx £[f}-\
S* JO
f(x)dx--
S Jo
/
Ja
f(x)dxdx.
Solve each of the following initial- value problems using Laplace transforms.
21. y" - 3/ + 2y = = 4
0; >-(0) = 3, /(0)
22. ?' + y =y(0) =
t; -1, /(0) = 3
"29. If Ly = h(t) is a linear differential equation with constant coefficients and if hit)
is of exponential order, show that every solution of this equation is of exponential
order. [Hint: Each solution is of the form yh (t) + yv (f), where yh (t) satisfies the
homogeneous equation Ly = and yp (t) = /q G(t, l-)h(l-) d%. Recall how the
Green's function G(t, £) is constructed from solutions of Ly = (see Section 4-5).]
at
£[e f(t)] = <p(s - a). (5-20)
Proof.
at
£[e f(t)} = f e-
Jo
st at
e f(t)dt
{s -a)t
e- f(t)dt
-JJo
= £[/(')](* - a)
= <p(s - a). |
This result is sometimes known as the first shifting theorem (the second will be
given below), and may be written in terms of inverse transforms as
_1 - a -1
£ [^(5 a)] = c '£ [^(5)], (5-21)
or as
-1 _1
£ lVCs)] = e°'£ k(j + «)]• (5-22)
~2
£'e '
c ° s3 "= (,+y+9 -
Example 2. Compute
£~' 25+3
s2 - 4s + 20,
194 THE LAPLACE TRANSFORM | CHAP. 5
2s + 3
= 25 + 3
s2 - 4s + 20 - 2)2 + 16
(s
= 2(5
- 2) + 7
(s - 2)2 + 16
5-2
2
2)2 +
[(s- 2)2+ 16j
+ 4 - 2)2 +
16J '
4 l(s
L( 16J
But
£' i f 5-2 1
= e
2t
2t
cos 4t,
l(s - 2)2 + 16J
and
e
1
= e^sin4f.
[(s - 2)2 + 16j
Hence
>-i
(s -2)* + 16]
= 2e2t C ° S 4t
+ ^ Sin 4r -
In order to state our next result we must introduce the so called unit step function
ua (t), which is defined by the formula
KD •23)
(1, t > a.
FIGURE 5-4
/(0 = |°' ' ~ °>
(5-24)
(sin (t — a),
,
t > a,
v
shown in Fig. 5-5, in very simple form. Indeed, since ujj) is zero for t < a,
we have
fit) = ua (t) sin (t - a), (5-25)
fit) = ua (t)g(t - a)
0, t < a,
g(t -a), t > a,
* Some authors restrict the term "unit step function" to mean what we have called
«o(0- I n that case (5-25) becomes /(/) = uo(t — a) sin (t — a).
5-5 FURTHER PROPERTIES OF THE LAPLACE TRANSFORM 195
FIGURE 5-5
describes the function obtained by translating or "shifting" g(t) a units to the right
and then annihilating that portion to the left of a (Fig. 5-6). Such functions arise
in practice as time delayed inputs to physical systems (i.e., inputs occurring at
time t = a > 0) and are of considerable practical importance. The next theorem,
known as the second shifting theorem, gives a formula for the Laplace transform
of such a function.
FIGURE 5-6
(For physical reasons the factor e~as in this formula is known as a delaying
factor.)
8t
Proof. £[/]= f e~
Jo
f(t)dt
£[/] = C e-
Jo
8{x+a)
g(x)dx
e~ r e~
as 8X
= g(x) dx
Jo
= e~a8£{gl
and the theorem is proved.
1
or as
£ -1 [e- a V(*)] = u a (t)g(t - a), (5-27)
_as
= e £[sin (t + a)]
-as
= e £[sin cos a + cos sin a]
t t
as
= e~ {cos a£[sin + sin a£[cos t]} *]
as
_
~ e~ (cos a + sin a) -y
s + 2 1
Example 4. Let/ be the function whose graph is shown in Fig. 5-7. To compute
<£[/] it is convenient to think off as the sum of the functions
/_(/) =
/8W (0, /< 2,
l-(2 - 0, t>2.
FIGURE 5-7
Then
£[/] = £[/i] + £[/2 ] + £[/3 ]
2s
= - + e~ £[-t]
s
+ e~ £[/]
5 + e~
2s
- e~
s
Example 5. Find £T l
[e-*'/{s
2
+ 6s + 10)].
Using (5-27) we have
-3s
— = -
£ u 3 (t)g(t 3),
s2 + 6s + 10_
_1
whereof) - £ [l/(s 2 + 65 + 10)]. But, by (5-21),
£" 1
= £" 1
= e- 3 ^- 1
= e~ 3t sin *.
s2 + 6s + 10. L(s + 3)2 + 1 s2 + 1.
I
Hence
,—3*
-3(«— 3)
= u 3 (t)e~ sin (t — 3).
s2 + 65 + 10
Our next result is somewhat akin to Theorem 5-6 in that it permits us to compute
n
the Laplace transform of t f(J) when £[/] is known. Specifically, we have
n w
£[t f(t)l= (-l) J^(s). (5-28)
/»O0
at
<p(s) = / e- f(i)dt
Jo
A = st
e- f(t)dt
4-Ms)
ds iJ
ds o
^[e- stf(t)]dt
I.o ds
= - st
f e-
Jo
Jo
tf(t)dt
= -£[tf(t)],
and so forth. (For a justification of this differentiation under the integral sign see
Appendix I.) |
dn
<p(s) = (-lyrjB- 1
!^)]. (5-29)
dsn
Example 6.
A
ds \s 2 + l)
2s
(s 2 + l) 2
198 the laplace transform | chap. 5
Example 7.
(-D ds n \sj
n!
,s
n+1
I
= _ 1 d_ ( 1 \
(s 2 +f l) 2 2s ds\s 2 + 1/
£' f sin f.
[ds V + 2
1/J U + U
2
Hence
, dt
= \ I
t sin t dt
o
= — ht cos * + i sin t.
Our final formula is designed to reduce the computation of the Laplace trans-
form of a periodic function* to the evaluation of an integral over a finite interval,
and reads as follows:
Theorem 5-9. Iffis of exponential order and is periodic with period p, then
Proof. By definition,
st
£[/]= re~ f{t)dt
Jo
St 2P st
= f e~ f<J) dt+ f e- f(t) dt+ • •
•
Jo Jp
+ P
•(n+l)p „,
St
+ /' e- f(t)dt +
We now set x + np = t in the (n -f l)st integral of the above series to obtain
{n+1)P P
[
Jnp
e-*<f(t)dt =
Jo
[ e-«™f x + ( np)dx
P
= e-nps f e- sxf(x)dx,
Jo
Jo
nps P sx
+ e-
Jo
[ e- f(x)dx+ •••
ps 2ps P sx
= [1 + e~ + e~ + • • -]( e- f(x)dx.
Jo
1 + e~ ps + + e- +
is 1/(1 - e~ ps
), and it follows that
8tP
£[f] = S e- f(t)dt
_
1 — e—p*
as asserted.
Example 9. Find the Laplace transform of the function whose graph is shown
in Fig. 5-8. In this case /is periodic with period 2, whence
lie-'f{t) dt
£[/]
1
- Q—2 S
f 1 „-* l
Joe dt
1 - e -2s
1 - e~ s
S {\ - e -2 s )
EXERCISES
1. e '
sin 3t 2. 3e cos2f 3. t sin 3t
3t
4. t e cos t 5. <T cos(2f + 4) 6. te\d/dt)(sm2t)
2t
7. te f\t) 8. (Jf + l)/(0, where Z) = d/dt
f
t, f < 2
f
— 4, 3 < t < 4
.0, f > 4
l -3 2
19. 1 te sin t di 20. e / f cos 4f dt 21. f / f sin tdt
Jo Jo
2
d —
22. f6 'cos 3f dt 23.
«'J a
t
dt
(e sin j )rff
df* Jo
24. Find the Laplace transform of the stair-
case function
fit) = n + 1, 4 -
1
n < t < n + 1, h = 0, 1, 2, . . .
, 1
3
1
2
1
1 1
1
4a '2 4
a 2a 3a 1 3
1 1
25. 26. 27.
sis + 1) (s - 1)2 s(s + 2)2
5 1
28.
*2(s - 5)2
29.
(5 - a) n
, n>\
1 1
30. a, b constants 31.
(5 - a)(s - b) 52 + 45 + 29
2s
32. 33. [#m/: See Example 8.]
2s2 + 1 (*2 + 1)2
i 2
1 35
*34. 35. [Hint: First expand in partial fractions.]
0*2 + 4)3 (j2 + 1)2
3 <, —2a
2j 3e
"36. 37. 38.
(52 + 1)
3 352 + 1 5* + 1
1 4 4 2 2
39.
54
.
',
+ .
1
[#//tf: 5 + 1= (5 + 25 + 1) - 25 .]
35
40. [Hint: See Exercise. 35.]
(.S + I) 4
41. 42
s is - 1)(5 - 2)
2
1
43. 44.
53 + a3 53 + fl3
... 5+3
In —^r
rrr
[#m/:
.
— In 5+3
rf , 1
45. Apply Formula (5-29).]
s + 2 ^5 5 + 2 (5 + 2)(5 + 3)
2
46. In
5 + i 1
1
sis + 3)
*47. Describe a procedure for finding the inverse Laplace transform of any rational
function P(j)/Q(j), P, Q polynomials with degree P < degree Q. In particular,
show that it is sufficient to consider the special cases
1 1
(5
2 + a 2)n (5
2 _|_ a 2)n (s
_|_ fl
)n
s- 1
r
L(52 +
i
a2)»+lj
i -if
Jo
*-*
In
r
L(52 +
*
fl2)n
<//.
[* /' /"'
£—
1 1
_(52 + fl2)n+l
r:
l*an\
— Jo j / / /
Jo
t • • •
/
Jo
t
.
sm at dtdt- •
dt.
The use of the Laplace transform in solving initial -value problems was introduced
and justified in Section 5-4. In this section we shall illustrate how the formulas
just derived allow us to solve more elaborate problems.
Taking the Laplace transform of both sides of this equation, and applying the
given initial conditions, we obtain
.2„r„,
z n„r,, 2 3(5+2)
s £[y] + i
1 +, ,-r..-,
As£[y] +
,
\3£[y] = -=2
s
+ ,
'
(5 + 2) 2 + 9
Hence
m s2 + 4s + 13 ^ s 2 (s 2 + As + 13) ^ (s 2 + As + 13) 2
and we must now find the inverse transform of the various terms on the right-
hand side of this equation. The first can be disposed of without difficulty since
1 1
2 A
— —
B Cs + D
+
.
,
-r -52
s 2 (s 2 + 45+13) S '
5 '
52 + 45 + 13
whence
^5(5
2
+ 45 + 13) + B(s 2 + 45 + 13) + (Cs + Z>)5
2
= 2.
A + C = 0,
4^4 +5+ D= 0,
13,4 + AB = 0,
135 = 2,
5-6 LAPLACE TRANSFORM AND DIFFERENTIAL EQUATIONS 203
3
+ 9
\(s
3(169) V(5 + 2)2 + 9
9/
and
2<
= ~ T69 + 13 t + rfg e cos 3/ - ^V e 2<
sin 3*.
_s 2 (s 2 + 4s + 13)_
Finally, since
3(s + 2) 3 rf
_ I A/ 3 \
2 dy \(* + 2) 2 + 9,/
3(5 + 2) -2t
£" = Me sin 3/.
is 2 + % + 13) 2
Combining these results we see that the solution of (5-31) is
(D - a)
2
y = 0, (5-32)
y(0) = C\, /(0) = c2 Taking Laplace transforms of both sides of (5-32) and
.
s
2
£[y] - ds ~ c2 - 2a(s£[y] - c{) + oc
2
£[y] = 0.
204 THE LAPLACE TRANSFORM | CHAP. 5
Hence
Cl l
s — a
+ (c 2 + 3c ia )
(s - a) 2
But
— 1
= ai
£ — and <£" 1 te
s lis «) 2 J
Hence
a' a*
j = Cl e + (c 2 + 3cia)te
= c3 e
Of 2
+ 1
c<ite
j <xt
,
EXERCISES
3. ^+ 2
^+ 3y = 3t; y(0) = 0, /(0) = 1.
2 2(
4. (Z) - 4Z) + 4)y = 2e + cos t; y(0) = ft, /(0)
= <L.
25-
FIGURE 5-11
dy
5. ^
at
+ ky = h(j); y(0) = 0, with k a constant and the graph of h(i) given by
Fig. 5-11.
1
i3
'^~^ + %~ Ay=
2
A
^ + 4g2<;
m =
'
/(0) = 5
'
^'^ = 3'
8 '
^+ ^+S" 3 3
^"^ =
^ A0)
= y(0) = y (0) = ^'^^
'
= °-
*»
k_ _»
dt (l
r>2 .
10. -j-r + y — t +1; y(7r) = w , /(ir) = 2r. [Hint: First make the substitution
at 2,
X = t — IT.]
,2
11. -r^ - y = -10 sin 2t\ y(ir) = -1, /Or) = 0. [flwif: See Exercise 10.]
at z
12.
^+ y = I
'
U - 1,
'
f
"
>
1;
1;
y(l) = y'iX) = 1, /'(l) = /"(l) = 0.
-£ + 2y+ I
/ ><t)dt
y(t)dt = \2 - t, 1 < t < 2,
dt J
Jo°
0, / > 2,
14. (a) Suppose that y{t) is of exponential order and is a solution of the Euler equation
a, b constants. Show that £[y(t)] also satisfies an equation of the Euler type,
(b) Prove that the result in (a) is also valid for any solution (of exponential order)
of an Euler equation of order n.
15. Show that if f(t) and /'(/) are of exponential order, and if /(/) is continuous for all
/ > 0, then
Urn s£[f] = + ).
/(0
8 —>oo
*16. Bessel's function of the first kind of order zero, denoted Jo, is by definition that
solution of the differential equation
1
£[/o(0] = ,
•
V1+J2
206 THE LAPLACE TRANSFORM | CHAP. 5
(1 + 2
s )<p'(s) + s v (s) = 0,
£-V(*>Ks)] = f fit ~
./o
S)g(t)di, (5-34)
and in this form asserts that if we know the inverse transforms, / and g, of the
functions <p and \f/,
we can express the inverse transform of the product <p(s)^(,s) as
an integral involving /and g. The integral in question is called the convolution of
/and g and is denoted/ * g; that is,
f*g = JB-^WC')],
_1 _1
JB-^frM*)] = £ [^)] * JB [^)]. ( 5 -36)
5-7 | THE CONVOLUTION THEOREM 207
= I™ [' e- st
f(t- OgiOdtdt,
Jo Jo
where the integration is being performed over the region of the *£-plane described
by the inequalities
f
Jo
re-
Jk
8t
f(t- OgiOdtdt,
or
" 8t
-
/o &) '(/" e~ f(t £)di} rftf "G «* E 5-12
/""e-'/d
J%
- *)* =
JO
r e- s{u+i)
f(u)du.
Hence
"S(M+?)
£
t/o'
/( ' ~ 0g(0 *] = /o"
g(l)
(/o°°
e ^M >
jM
)
*
°° °°
= «""*«(«) «"~/(«) du) dZ
/o (/o
= re-8Uf(u)dur e - sig(0d^
Jo Jo
= £[f]£[gl
and the proof is complete. |
f fit - ZMOdZ =
Jo
['git
Jo
- £)/(£) </£.
Example 1. Find
£' 1
u* 2 + i)J
it can also be handled just as well by using (5-35) and (5-36) as follows:
£' = £' £—
[s(s 2 +
-
1)J
*
h]
= 1 * sin t
•t
= / sin £ d%
Jo
= 1 — cos t.
(D 2 + D- 6)y = h{t),
(5-37)
XO) = /(O) = 0,
y ' & 1
Kt).
U-2)(5+3)J
5-7 I THE CONVOLUTION THEOREM 209
But since
£"
[:(s
1
= £' if_i- L_l
- 2)(s + 3)J |_5(s 2) 5(5 + 3)J
= ie 2 < _ ie-3 '
we have
(5-38)
Jo
Jo
Jo Jo
The reader will recall from Section 4-5 .that the function K(t, £)
_
= 1*2 U-$)
\e _
±e~ 3(t ~® appearing in (5-38) is called the Green's function for the linear differ-
ential operator L = D2 + D— 6 (for initial-value problems at t = 0), and as
such completely determines an inverse for L. In the next section we shall have
more to say about the simple formula
= £—
1
G{tt
i
- 0*
s2 +s— 6
('
EXERCISES
Use the convolution formula to find the inverse Laplace transform of each of the following
functions.
3
£[f] e~ '£[f] 1
1. 3.
s2 + 1 S 2 (S + 1)
s 3/ 1
- - a ^ b
'
(s 2 + 1)2
*
(5 2 4- 1)2 (s a)(s b)
f At -
Jo
ZMOdi- = f*
Jo
git - *)/#)««.
./o Jo Jo
—
f
Jo
sin a(un —i un ) sin aun du n • • • du\
hf>f>f<
2 n n\
Jo Jo Jo
t sin at dt • • •
dt.
Use the result of Exercise 19 to compute the Laplace transform of each of the following
functions.
[
Jo
m *-[
t Jo
JO
£[f\ ds.
sin/ t
—-dt = ,
^
/. t 2.
The Gamma function T(jc) and Beta function B(x, y) are introduced in Exercises 25-33.
5-7 |
THE CONVOLUTION THEOREM 211
T(x) =
Jo
/ «~ V -1 dt. (5-39)
(It can be proved that this integral converges for all x > 0.) Use integration by
parts to show that
T(x + 1) = xT(x) for all x > 0.
26. Use the result of Exercise 25 together with the value of T(l) to prove that
T(n + 1) = n\, n = 0, 1, 2,
(Because of this property the gamma function is also called the generalized factorial
function.)
27. By differentiation of (5-39), show that T"(x) is non-negative on (0, »). Where,
approximately, does the minimum of T(x) lie? Draw an accurate graph of F(x).
28. Let n be an arbitrary real number greater than —1. Prove that
u
r(i) = 2 e~ du,
Jo
and hence that
r
00
-x r —Vxx
(a) /
—— : ax
dx mi
(b) i
/ e~
i dx
Jo y/x
a/y ^0
rcoroo x
'\\ - uf- 1 du
u
T(x + y)
I
Jo
whenever x, y > 0. This integral is known as the Beta function of x and y and is
[Hint: Use the result of Exercise 28 and the convolution formula to evaluate
£- 1 [l/s*+ 2
'] in two ways.]
212 THE LAPLACE TRANSFORM | CHAP. 5
dx Vx r(l//i)
/.o Vl - x* n r K» + 2)/(2/i)]
Then L can be viewed as a linear transformation from e n (7) to <B(I), and a solution
of (5-40) can be described as a linear transformation G from (3(7) to e n (7) which
acts as a right inverse for L, i.e., is such that L(G(h)) = /i for all h in e(7). In the
absence of any further conditions on the unknown y, G is not uniquely determined
by L since L is not one-to-one. Thus to construct G we must impose a sufficient
number of additional restrictions on y so as to make the solution of (5-40) unique.
For example, we might require, as in Section 4-4, that y be the solution of (5-40)
which satisfies the complete set of initial conditions
X'o) = /Co) = • • • = y {n - l
\t ) = (5-41)
at some point t in the interval I. Then the corresponding inverse G can be ex-
pressed as an integral operator
G(h)= ?K{t
Jt
t m&dl t
(5-42)
excellent example of the efficiency that accrues from using the "coordinate free"
ideas of linear algebra.
Thus, for the remainder of this discussion we shall consider the constant coeffi-
L = Dn + fln-i^""
1
+ • • •
+ a , (5-43)
Ly = h(t),
(5-44^1
l }
;K0) = /(0) = • • • = y»-»(0) = o,
p(s)£[y] = h(s),
where p(s) = s
n
+ an -i^ n_1 + * • •
+ a , and h(s) = £[h]. Thus
or n Ks)
p(s)
and hence if
and the function g(t — appearing in this integral is then, by definition, a Green's
function for L for problems at t =
initial-value (see Definition 4-2).
Fortunately we can say much more about g(t — £). For the function
g(t)
--a
=
Ly = 0,
tion of (5-47) on all of (— oo, oo). Thus by Theorem 4-3, K(t — £) is the Green's
function for L (for initial-value problems) on the entire interval (— oo, oo). When
Laplace transforms are used in a purely computational fashion it is common prac-
tice to ignore the distinction between K(t) and g(t), as we do, for example, in
stating the next theorem which summarizes the above results.
Theorem 5-11. Let L be the constant coefficient operator (5-43), and let
n -1
p(s) = s + fln-i^ + • ' •
+ flo
the function g(t — £) is the Green's function for L on the interval (— oo, oo).
L = D2 - 2aD + a
2
+ b
2
, b 9* 0,
we first set
8 ^ = £
[s 2 - las + a2 + b 2\
L(* - a) 2 + b 2}
at
= 1 e sin bt.
b
g(t- = e
ait ~^ sin b(t- a
I
(4£
4 - 4Z>
3
+ 5D 2 - AD + 1)^ = In t
L = D* - D 3 + fD 2 - D + a
Since
_s* - s* + fs 2 - s + J_
4
= £'
(2s - 1)
2 2
-f 1)
13. 22 16„A 12.
= £' 5 25 25 25
_(2.s - l) 2 2s-l T
+ s2 + 1
~ s2 + 1
= W ~W 12 12
+ if cos t - Hsin t,
~ = Ht - 0e (t -* )l2 - -° 12 -
g(t ¥te
{t
+ M cos (t - if sin (t - £).
The solution of the given initial-value problem can now be written in the form
m J>-ȴ dk,
since h(i) = (In t)/A is the right-hand side of the normalized differential equation
(Z>4 _ /)3 + ^2 _ ^ + i
)j;
= i
ln ,
As a final example let us use Green's functions and Laplace transform methods
to solve an initial-value problem with nonhomogeneous initial conditions, say,
Ly = h(t):
n - 1} (5-48)
y(to) = c , ... ,y (/ ) = cn _i.
Following the methods of Chapter 4 we can write the solution of (5-48) in the
form
y = yh + G(h),
(D 2 - 2aD + a2 + b )y
2
= h(t),
(5-49)
yW) = 2, /(,) = -3,
we have
G(h) = e
a «-* sin bit - mt)di.
lj^
216 THE LAPLACE TRANSFORM | CHAP. 5
(D 2 - 2aD + a
2
+ b )y
2
= 0,
This, of course, may be done with the methods of Chapter 4. But we can also
use Laplace transforms if we note that y h (j) = y(t - tt), where y(t) satisfies
(Z)
2 - 2aD + a
2
+ b )y
2
= 0,
(5-50)
X0) = 2, /(0) = -3.
(s
2 - las + a
2
+ 2
& )£[y] = 25 - 3 - 4a,
or
_ 2s - 3 - 4a
£ LJ>J - _ +
S2 2fl5 + a2 fe
2
2(5 - a) 3 + 2a
~ (s - a) 2 + b2 (s - a) 2 + b2
Thus
y(t) = 2e
at
cos bt - ^^e at
sin bt,
and
—
yhQ) = y(t - x) = e
a(t jt)
2 cos b(t — tt) r—- sin b(t — x)
—
a(t »•)
2 cos b{t - tt) r—- sin b(t - x)
EXERCISES
Determine Green's functions for each of the following linear differential operators in
three ways: (a) by applying Formula (4-42) or (4-43); (b) by applying Theorem 4-4;
and (c) by applying Theorem 5-11.
1. D 2 - 4D + 4 2. D2 + D 3. D 2 + 6D + 13
4. D 2 + i# - 4 5. Z>
2
- £/> + ^ 6 -
4£)3 - ^
4
7. D3 + 1 8. (Z>
2
+ l)
2
9. Z) + 1
10. (Z>
2 - 4Z> + 20)
2
5-8 |
GREEN'S FUNCTIONS FOR CONSTANT COEFFICIENT OPERATORS 217
(D
2
- l)y = f°>
° < ' < h
\t - 1, t > 1,
y(0) = /(u) = o,
in three different ways: (a) by using Laplace transforms directly; (b) by first deter-
mining a Green's function for D 2 — 1 and (c) by solving the initial-value problem
;
13. Show that the methods of Sections 5-4 and 5-5, when used to compute £- l [l/p(s)],
lead to the unique solution on the entire interval (— oo oo ) of the initial- value problem ,
of Exercise 12. (This justifies the statement that g(t — £) is the Green's function for
L on (-oo, oo ), where g(j) = £~ 1
[l/p(s)].)
14. Use Laplace transforms to derive (5-45) for any function hit) which is piecewise
continuous on [0, <»). [Hint: First consider the solution of (5^4) with hit) replaced
by
= h V>> ° ^ * ^ a
H{t) [
>
|u, t > a,
where a is a constant > 0.]
15. If L is a constant coefficient operator, show that yp (f) is a solution of the initial-
value problem
Ly = hit);
Ly = h{t + to);
Solve each of the following initial-value problems using the method given at the end of
this section.
21. (4Z)
2 - AD + 37)>> = e'
/2
cos3/; y(a) = /(a) = -2, a >
7,
Suppose that the spring-weight system is in equilibrium, and that the weight is
now subjected to an additional vertical force h(t) which may vary with time. Then
at time t, with the weight a distance X0 fr° m the equilibrium position, and the
positive ^-direction measured downward, the forces on the weight are mg due to
gravity, —k(y — ) due to the restoring force in the
y spring, and h(t). Hence,
m M+ ky== m (5-52)
Furthermore, since the system was initially at rest in equilibrium, y(t) must satisfy
and the motion of the weight is thus obtained as the solution of an initial-value
problem.
To solve this problem we use Laplace transforms and obtain
* The positive constant k is known as the spring constant. It depends upon the ma-
terial from which the spring is made, the dimensions of the spring, etc.
5-9 I
THE VIBRATING SPRING; IMPULSE FUNCTIONS 219
Hence
£L>]
= (5-54)
ms 2 + k
and
= £-'
1
y(t) * h(i)
ms 2 + k_
— v^7m h(t)
r 2
mk (s + (A:/m))_
—
"mk
sin J— t)
V™ )
* h(t)4
Thus the equation of motion for this system under the action of an arbitrary ex-
ternal force h can be expressed in integral form as
y(t) = 7=
1
f-sin [Vk/m (t -
/
S1 £)M£) d%. (5-55)
mk J o
In applications the impressed force h(t) is often of the form
where A and co are positive constants, in which case the equation of motion
becomes
A
= — == C sin [\/k/m (t — ,
yif) / £)] sin co£ rff
Vmk J o
Although this integral can be evaluated by elementary techniques it is instructive
to begin again with (5-54) and solve the problem directly. Thus
1
£\y] = £[A sin cot]
ms 2 + k
1 Act
ms 2 + k s2 + co 2
—
Aco
(_i s—\
k moo 2 \s 2 + co
2 ms 2 + k)
t Those of our readers who are familiar with the material of the preceding section will
recognize that we have now computed the Green's function g(t — £) for initial-value
problems involving the normalized operator L = 2
(k/m), and that D +
git — £) = y/m/k sin y/k/m (t — £).
220 THE LAPLACE TRANSFORM CHAP. 5
FIGURE 5-14
m = Ay/km
(jns 2 + k) 2
Ay/km s
s (ms 2 + k) 2
But since
£" = £
(ms 2 + k) 2 \ m
2m ds \ms 2 *-f k) \
1
2m ms 2 + k
sin y/k/m t,
2my/km
5-9 THE VIBRATING SPRING; IMPULSE FUNCTIONS 221
y = \ sin / — \ t cos t
FIGURE 5-15
_A_
-r sin y/k/m t — y/m/k t cos y/k/m t
2m
and we have
Therefore when the impressed frequency is equal to the natural frequency, the
amplitude of the oscillations increases with time and the spring is eventually
stretched beyond its elastic limit (see Fig. 5-15, sketched for A = k = m = 1).
This phenomenon is known as resonance, and is important in various physical
problems, f
A rather different situation arises if we attempt to find the response of this sys-
tem when the weight is struck a sharp blow in the vertical direction at time t = a,
0, < t < a,
Kt) -
T
' a < t < a + t, (5-59)
,0, a + r < t,
pulse of 1 to the system.* We now agree that the mathematical description of the
physical situation described by the words "sharp blow" is obtained by using a
force which acts throughout an arbitrarily short interval of time but imparts a
predetermined impulse, or change of momentum, to the system. Our problem
then becomes that of determining the behavior of the solution y(t) in (5-55)
when h is as above, and r — > 0.
0, < t < a,
=
1
sin \/k/m {t — £) d£, a < t < a + t, (5-60)
xo mk
a-\-r
1
in
sin y/k/m (t — dH t
a + t < t.
0, t < a,
yo(t) (5-61)
1
sin \/k/m (t — a), t > a,
(see Exercise 2). But y (t) is also the solution of the initial-value problem
2
dy
m-^ + ky 0;
y(a) = 0, y'(a) = - »
and as such can be interpreted as the response of a weighted spring which is given
unit momentum at t = a [i.e., my' {a) = 1], and left undisturbed thereafter. It
y it) may be interpreted
follows (see the preceding footnote) that in this situation
as arisingfrom an instantaneous unit impulse imparted to the system at t = a,
and the problem has been solved.
In certain circumstances it is convenient to think of such a unit impulse at f =
as arising from a fictitious function 8(t), called the Dirac Delta Function, whose
defining properties are
/""
5(0 = for all t 9* 0, S(t)dt = 1. (5-62)
J — 00
y(0) = /(0) = 0,
we have seen, has y (t) as its solution.
and, as
The foregoing discussion can easily be generalized to initial-value problems
of the form
Ly = 8(t - a);
( }
y(0) = /(0) = • • • = /"""(0) = 0,
~
where L = Dn + an _ 1 D n 1
+ • • •
+ a . To solve this problem we again use
Laplace transforms, first replacing 8(t — a) by the function h defined in (5-59),
and then passing to the limit as r — > 0. Thus
n ~1
with/?(s) = s + an _ 1 s n + • • •
+ a , and hence
= f'gO- Mi)#,
Jo
where g(t) = £ _1 [l/^(-s)]- Using the given value of hit) we thus obtain
0, t < a,
y (t) = <
- g(t - OdZ, a < t < a + t, (5-64)
J
+T
1
- f git- Od%, a + r < t,
\T J/ a
224 THE LAPLACE TRANSFORM | CHAP. 5
= !°'
'-"' (5-65)
*>(/)
W -a), t > a.
This result is valuable because it can be used to define the Laplace transform of
8{t — a). Indeed, by Formula (5-26) we have
£L>o] = e~ as £[g]
—as *
Pi*)'
Thus if £ [8(t — a)] exists at all it must have the value e~ as and hence we are ,
£[5(0] = 1. (5-67)
This apparent contradiction of Theorem 5-3, which asserts that £[/] ->0as
s — oo, is merely a reflection of the fact that 8(t) does not belong to the space of
piecewise continuous functions of exponential order. But this is hardly surprising
since, as was pointed out above, 8(t) is not even a function in the usual sense of
the term.
y(0) = /(0) = o.
Then
(ms
2
+ k)£\y] = ^^ - 2e-
5-9 THE VIBRATING SPRING; IMPULSE FUNCTIONS 225
and
m
JBM =
(ms 2 +
^
k)(s 2 + w2) ms 2
?
+ A:
e-
108
0, t < 10,
e(t) = 2 ,
{
sin Vk/m (t - 10), t > 10,
EXERCISES
where h(i) is defined by (5-59), leads to the same result as Formula (5-66).
2. Let y(t) be defined as in (5-64). Prove that
lim;y(/) = |
' '- fl '
4. Find the equation of motion of a weight having mass m which is initially in equilib-
rium at the end of a spring, and which, at time t = 0, is struck a sharp blow from
above which instantaneously imparts 1 unit of momentum to the system.
5. A weight of mass 1 is attached to a spring whose spring constant is 4, and at time
t = the weight is struck a blow from above which instantaneously imparts 1 unit
of momentum to the system. At time t = ir/2 a sinusoidal force of magnitude
—sin (t — (71-/2)) begins to act vertically on the system. Find the equation of motion
of the mass.
6. A mass m, hanging in equilibrium at the end of a spring, is struck from below and
instantaneously given two units of momentum. At time
a the mass t = is subjected
to the external force sin (t —
Assuming that the spring constant
a). is different
from m, find the equation of motion of the mass m.
226 THE LAPLACE TRANSFORM CHAP. 5
A unit mass is attached to a rigid spring whose spring constant is 3, and is then
mounted in an elevator as shown in Fig. 5-17. At time t = the elevator begins
to descend with a constant velocity of 2 ft/sec, and at that moment the mass is struck
a blow from above which instantaneously gives it one unit of momentum. Find the
equation of motion of the mass as a function of time.
v = 2ft/sec
FIGURE 5-18
FIGURE 5-17
8. A mass m is suspended on a spring beneath a car which moves with constant velocity v
along the track shown in Fig. 5-18. Suppose that at time t = the mass is struck
from below and instantaneously given one additional unit of momentum. Find the
equation for the motion of m in the vertical direction as a function of time.
9. Find the equation of motion of a mass m bobbing at the end of a spring with spring
constant k if the mass was originally displaced a units from equilibrium and given
10. It is observed that the spring of Exercise 9 oscillates with a period of 2 seconds when
m is 1 gram.
(b) Now suppose the mass on this spring is 9 grams and that the oscillations start
at a = 4 with initial velocity b = ir. Find the period and amplitude of the
oscillations.
11. Suppose that a spring is suspended in a resisting medium which opposes any motion
my" + X/ + ky = 0,
£ mas + mb + Xa
y =
1 ms 2 + Xs + k
12. Assume that the resisting force in Exercise 11 is sufficiently large to ensure that
X2 > 4km.
5-9 I
THE VIBRATING SPRING; IMPULSE FUNCTIONS 227
(a) Show that the quadratic equation ms 2 + As + k = then has distinct negative
roots, —a i, —a2 with <r\ < <T2, and that
y = e~" l \a + fie
~^~^\
where
a , 2bm + aX a 2bm + aX
2 - 4km 2
2>/X 2 2VX 2 — 4km
X - VX 2 - 4km x + Vx- 2 -- 4km
0"1 = ~ ' 0"2 =
2m 2m
(b) Show that except in the trivial case where y = 0, the mass in this problem is in
itsequilibrium position y = for at most one value of t. (The effect of the resistance
in this problem is often described by the word "damping," and the resulting motion
is known as damped harmonic motion.)
13. Suppose that the spring in Exercises 11 and 12 has spring constant 9, and that a unit
mass is attached to the spring, displaced one unit in the positive y-direction, and given
an initial velocity b = — 10. Suppose in addition that the medium offers a resistance
corresponding to X = 10.
(a) Find the equation of motion of the mass.
(b) Show that the mass is in its equilibrium position for exactly one instant. When
is this?
(c) What is the maximum height which the mass attains? (Recall that the positive
^-direction is downward.)
(d) Sketch the graph of the solution curve for this problem.
14. Show that a necessary and sufficient condition that the mass of Exercises 11 and 12
be in its equilibrium position at some one instant after the motion begins is that
< — a/j8 < 1, and hence deduce that if a j± and b = the mass approaches
its equilibrium position without ever passing through it.
15. Show that the quadratic equation ms 2 + Xs + k = for the vibrating spring prob-
lem described in Exercise 11 has a double root s = —a = — X/(2m) < when
X 2 = 4km. Find the equation of motion of the mass in this case. (This is an ex-
ample of what is known as critical damping.)
16. Let y = y(t) be the solution obtained in Exercise 15.
(a) Show thaty —» as t —* qo , and that the mass passes through its equilibrium
position at most once during the motion.
(b) Show that if b/a < — a, then y = when t = — a/(aa b) > 0, but that +
otherwise the mass approaches its equilibrium position without ever passing through
it.(Assume y ^ 0.)
17. Find the equation of motion for the system in Exercise 11 in the case where X 2 < 4km.
What is the behavior of y as t
—> oo ?
228 THE LAPLACE TRANSFORM | CHAP. 5
Function Transform
/•OO
8t
fit) £[/! = / e~ fit)dt
Jo
2
- + - +
fit) s £[f] sfi0 ) f'(0 )
n
- n + - j"- 2 +
f
W it) s £[f] 5 "V(0 ) /'(0 )
/-i>(0+)
*
/
fit)dt s
JO
* / •••/ fit)dt...dt n
Jo Jo s
n times
pQ,
/
Ja
•••/
n times
Ja
fit)dt...dt
-^LL md"" /*Qi pt pt
/ /
•••/ fii)dt...dt
SJo Ja Ja >
<
n — 1 times
* at
- =
e fit) f(s a), where f(s) £[/]
*
t
n
fit) ^ i«w n
0, t < a as
Uait)git) = e~ £[git + a)]
git), t > a
t The table consists of three parts. The first contains formulas of a general nature,
and the second the transforms of a small number of selected functions. When used in
conjunction, these two parts of the table will yield the Laplace transforms of most func-
tions which occur in practice. The third part of the table is primarily intended for com-
puting inverse transforms, and hence is designed to be read from right to left. The methods
of partial fractions, completing the square, etc., are, of course, indispensible in such
computations. Finally, those formulas in the first part of the table which are particularly
well suited to evaluating inverse transforms have been marked with an asterisk.
TABLE OF LAPLACE TRANSFORMS 229
Function Transform
* fa. t <
-
a as
, s ,
* u (t)g(t
a -«) = »
{' e- £[g]
[git -a), t > a
* £[/]£[«•]
fit- OgiOdZ
Jo
st
fit) periodic with period p (p > 0) Joe- fit)dt
1 - «-*•
M t
if ito
t^0+
^exists
f
/00
1
1
5
a< 1
e
j — a
n
t
a
sin at
52 + tf2
5
cos at
52 + fl2
a
sinh at
52 - a2
5
cosh a/
52 — a2
5(0 1
- —as
5(f a) e
n-l at
in
t
-
e
1)!
7
is
—-^r a)n
(« >
- 1)
1
^-r (sin a/ — at cos a?)
2a3 (j2 + a 2)2
—/
2a
.
sin at
(S
2 +
S
a 2)2
1
£ + a 2)»+l
Jo 2n [(S2 + fl2)nj^ (5
2
1 5
Tn^ (S
2 + a 2)» (5
2 + a 2)»+l
230 THE LAPLACE TRANSFORM | CHAP. 5
Function Transform
/» t pt pt
1
t t- - -
1 t sin atdt- • •
dt
2 n an\ Jo
J
Jo Jo 02 + fl
2)»+l
n times
S
1 1 t- • t sin atdt- • dt
2n
:
an\
/
Jo Jo
1
Jo (5
2 + fl2)»+l
n — 1 times
6
further topics in the theory of
Proof Let a and b (a < b) be two points of / such that y 2 (a) = y 2 (b) = 0, and
suppose that y 2 is nonzero everywhere between a and b. We must show that there
exists exactly one point c between a and b such that y\(c) = 0. (See Fig. 6-1.)
FIGURE fr-1
* None of the material in this chapter will be used in any essential way until Chapter 15
where equations with regular singular points are studied.
f A zero of a function y is a point xo at which y(xo) = 0. It may happen, of course,
that >>i and j 2 have no zeros on /, in which case the conclusion of the theorem is vacuously
satisfied.
231
232 THE THEORY OF LINEAR DIFFERENTIAL EQUATIONS |
CHAP. 6
nonzero, and
has the same algebraic sign everywhere in /. Moreover, the values of the
Wronskian at x = a and x = b are, respectively,
and hence y\(a), y\(b), y 2 (a), and y 2 (b) are all different from zero. But since a
and b are successive zeros of y 2 the derivative of y 2 must have opposite signs at a
,
and b (i.e., the graph of y 2 must be rising at a and falling at b, or vice versa).
Hence yi(a) and yi(b) have opposite signs, and it follows that yi(c) = for at
least one point c between a and b*
Finally, by reversing the roles of y x and y 2 in the above argument, we conclude
that y 2 has at least one zero between every pair of successive zeros of y x on /,
and we are done. |
y" + y = 0. (6-2)
A somewhat less obvious consequence is that any two functions of the form
have alternating zeros whenever a x b 2 ^ a 2 b u for all such pairs of functions are
also linearly independent solutions of (6-2). (See Exercise 1.)
Example The functions sinh x and cosh x are linearly independent solutions
2.
of y — y =
o, and hence their zeros alternate on (— oo, go). This, of course, is
obvious since sinh x has but a single zero, namely, x = 0, while cosh x has no
zeros at all, and is cited merely to emphasize that the separation theorem says
nothing at all about the number of zeros of a solution of (6-1).
Questions of the latter sort can sometimes be answered by using the following
theorem, also due to Sturm.
* The validity of this assertion depends upon the fact that a continuous function
assumes all values between its maximum and minimum on a closed interval [a, b].
6-1 I
THE SEPARATION AND COMPARISON THEOREMS 233
Proof. Let a and b be adjacent zeros of y 2 with a < b, and suppose that yi does
,
not vanish in the interval (a, b)* Since the zeros of a function y are the same
as those of — y, we may assume that yi and y 2 are both positive throughout
(a, b). Then, arguing as in the preceding proof, we have
where W(a) and W(b) denote the values of the Wronskian of yi and y 2 at a and b,
respectively. But
= - j 2 (xM(x)]
fx myi(x) yf 2 (x)] -^[y 1 (x)y 2 (x)
y2 (x)] = for all x in I. Thus yx must have a zero between a and b whenever
Pi(x) ^ p 2 (x). Moreover, the separation theorem implies that this result con-
tinues to hold when pi(x) = p 2 (x) unless y x and y 2 are linearly dependent in
e(7), and the proof is complete. |
has at most one zero on any interval in which p(x) < 0. For if we apply the
comparison theorem to this equation and
/' = 0, (6-5)
we conclude that on such an interval every solution of (6-5) must vanish at least
once between successive zeros of a solution of (6-4). The assertion now follows
from the fact that y" = has solutions (namely y = c) which do not vanish on
any interval.
* At this point we are making use of the fact that a nontrivial solution of a second-
order linear differential equation cannot have infinitely many zeros in any finite closed
interval of the *-axis. (See Exercise 6.)
234 THE THEORY OF LINEAR DIFFERENTIAL EQUATIONS |
CHAP. 6
EXERCISES
2. Prove that sin k\x has at least one zero between any two zeros of sin k^x whenever
k\ > k2 > 0.
3. (a) Show that every nontrivial solution of the equation y" + (sinh x)y = has at
most one zero in (— 00 , 0), and infinitely many zeros in (0, 00 ).
(b) Prove that the distance between successive zeros of any nontrivial solution of this
equation tends to zero asx-> 00
4. Every solution of the equation y" + xy = has infinitely many zeros in (0, 00 ).
equation can have infinitely many zeros in a closed interval / = [a, b]. [Hint: Assume
the contrary, and then show by partitioning / into arbitrarily small subintervals that
there would exist a point xo in / with the property that every open interval centered at
/contains infinitely many zeros of y(x). Use this fact to prove that
y(xo) = y'(x ) = 0,
x 2y" + x/ + (x
2
- p 2 )y = 0, (6-6)
Before the comparison theorem can be applied to (6-6) the term involving /
must be eliminated. This can be done by making the change of variable y = u/y/x
which transforms (6-6) into
without disturbing the zeros of its solutions. Thus it suffices to study the solutions
of (6-7). There are three cases to be considered.
theorem implies that every solution of Bessel's equation vanishes at least once
between any two zeros of a nontrivial solution of u" + u = 0. But, as we know,
the general solution of this equation is c x sin x c 2 cos x. Moreover, if a is an +
arbitrary real number, c x and c 2 can be chosen in such a way that the zeros of
ci sin x + c 2 cos x are a, a ± ir, a ± 2x, (See Exercise 2.) Thus we conclude
. . .
that every solution ofBesseVs equation of order p, p < \, has at least one zero in every
subinterval of the positive x-axis of length x; i.e., the distance between successive
zeros of these solutions does not exceed x. Finally, it is not difficult to show that
this distance is always less than x, and approaches x as x —» oo (Exercise 3).
Case 2. p = \. Here Eq. (6-7) reduces to u" + u = 0, and the general solution
of (6-6) can be written explicitly as
The remark made a moment ago concerning the choice of Ci and c 2 again applies,
and we can assert that the zeros of every nontrivial solution of BesseVs equation of
order % are equally spaced along the positive x-axis, successive zeros separated by
an interval of length ir.
apply the comparison theorem on the interval (x oo) to (6-7) and the equation ,
u" + \u = with general solution c x sin (y/2/2)x + c 2 cos (\/2/2)x. Since this
'
last function has infinitely many zeros in (x , oo) with successive zeros separated
by an interval of length a/2 x, it follows that every solution of BesseVs equation of
order p > % has infinitely many zeros, and that the distance between successive
zeros eventually becomes less than s/lir. By modifying the argument in the
obvious way, it can be shown that here too the distance between successive zeros
approaches x as jc —> oo
236 THE THEORY OF LINEAR DIFFERENTIAL EQUATIONS |
CHAP. 6
EXERCISES
" + (• + 4^)-*
2. Prove that for any real number a there exist constants c\ and C2 such that the zeros of
c\ sin x + C2 cos x are a, a ± ir, a ± 2ir, ....
3. (a) Prove that the distance between successive zeros of any nontrivial solution of
Bessel's equation of order p < % is always less than t. [Hint: For any fixed p < \,
and any jco > 0, 1 < 1 +«< 1 + (1 - 4p 2 )/(4x 2 ) on (0, x ).]
4. Prove that the distance between successive zeros of any nontrivial solution of Bessel's
equation of order p > § approaches ir as x —> <»
"5. (a) Let ao(x) be continuous on (0, a>), and suppose that there exist positive numbers
b, B such that b 2 < ao(x) < B 2 for all x > 0. Prove that every nontrivial solution of
/' + a (x)y =
has infinitely many zeros on (0, oo ), and that the distance d between successive zeros
can be estimated as
B ~ - b
[Hint:Use the comparison theorem on the given equation and each of y" + b 2y =
and y" + B 2y = 0.]
(b) Use the results in (a) to deduce the facts proved in this section concerning the
zeros of the solutions of Bessel's equation.
6. Let
be normal on an interval /. Show that there exists a function v(x) defined on / with
the property that the substitution y = uv reduces this equation to
u" + i(x)u = 0.
(c) (1
-
-
x
x
V
V
- 2xy' +p(j>+\)y =
- */ + p 2y = (d) y" - 2xy' + 2py =
8. Prove that every nontrivial solution of the Hermite equation
pa non-negative constant, has at most finitely many zeros on (— «> <*> ). [Hint: See
,
Example 3 and Exercise 6 of the preceding section, and Exercise 7 (d) above.]
*
Let
fl2(x)
S+ aiix)
tc
+ ao(x)y = ° (6_8)
p (x
\ _ gi [a 1 (x)/o 2 (x)] dx^ (6-9)
Then since
d ( , n dy\ , .d 2y . a\(x) , .dy
In addition to the self-adjoint form, there are a number of other special forms
for normal second-order equations which are sometimes useful. One of them
involving the invariant of the equation was given in the exercises at the end of
the last section. Another is a normalized version of the self-adjoint form in which
p(x) = 1. It can be deduced from (6-10) by making the change of variable
dx
= .
t
/ p(x)
For then
dy _ dy dt_ _ 1 dy
dx dt dx p(x) dt
A(P{X)\dy\
dx \
r
dx)
= A (dy\
dx \dt)
= ^L
dt
(ti\ dt_
\dt) dx
= J_ fy
p(x) dt 2
proves even more. For if the points x and t correspond under the change of
variable
dx
/
introduced in passing from (6-10) to (6-11), and if
(MO = pixtaiix),
then the fact that p(x) is positive throughout / implies that Qi(t ) > (M*o) if
- 12
s(**>&) + *<*»- and £(*'->$ + <*'»- °. (6 >
then Qi(t) > Q 2 (0 if and only if ^ x (a:) > q 2 (x). Hence the comparison theorem
stated in Section 6-1 is also valid for a pair of self-adjoint equations of the form
(6-12) whenever q\(x) >
q 2 (x) on /.*
As a final application of the ideas we have been exploring here we prove a rather
surprising result concerning the zeros of the derivative of any solution of a certain
class of self-adjoint equations.
* The reader should note that we have proved this assertion under the assumption
that the function p(x) is the same in both equations in (6-12). It can be shown, however,
that the conclusion of the comparison theorem continues to hold exactly as stated earlier
for the solutions of a pair of self-adjoint equations
d_
(p.M I) + ,.(»» = -
dx
0,
I (p 2 (x)
I) + q2 (X )y 0,
in which q\(x) > #2 00 and < pi(x) < P2(x). For a proof of this more general
theorem see G. Birkhoff and G. C. Rota, Ordinary Differential Equations, Ginn, Boston,
1962.
6-3 | SELF-ADJOINT FORM; THE SONIN-POLYA THEOREM 239
&{*<>£) + «*»- ^ 13 )
Then
FM = [yWf +
^ r)
[/'W/W]
!
f-^ + ^W-^to')*.
and since, by assumption, (py')' = —qy,
**<*> --©'!<«>•
Now suppose that pq is nonincreasing on /. Then d/dx(pq) < and F is non-
decreasing on / (i.e., F' > on I). Hence the same is true of any sequence of
values of F computed at points x1 < x2 < • • •
. If, in particular, xu x2 , . . are
the points at which y has a relative maximum or minimum, then /(*,-) = 0,
2
F(x{) = y(Xi) , and we have
2 2
X*i) < y(x 2 ) < • • •
Thus
b(^0l < b(* 2 )l < •
A
dx ('2) + (^'-* "
(6 14)
and since p(x)q(x) = x2 — p2 is increasing and positive on the interval (p, oo),
the Sonin-Polya theorem implies that the magnitude of the oscillations of such a
Note that the relative maxima and minima of a solution y of (6-13) are the zeros of /.
1 —
1
—
\
\ / —
' 1 1 1 1 1 1
1 1 1
xNv^
i 1 t^- 1
\ 5 / 10
_^/ 15
2
FIGURE 6-2
EXERCISES
d_
=
dx (**>!) + q(x)y 0,
dy
dP
+ Q(t)y =
d_
dx -T +
(*>£) 9(x)y =
5. Discuss the oscillatory behavior of the nontrivial solutions of each of the following
equations.
2*V + + (1 - =
(a) ' 6xy'
^j y 2 3
(b) y" -y + e\\ - x)y = (c) x y" + xy' + (x - 1)>> =
6. Prove that the sequence of absolute values obtained in the Sonin-Polya theorem is
7. (a) Discuss the oscillatory behavior of the nontrivial solutions of Airy's differential
equation
y" + xy =
on (—oo, oo ).
8. Let p(x) and q(x) > satisfy the hypotheses of the Sonin-Polya theorem, and let y(x)
be a nontrivial solution of the equation
=
l(' (x) l) +
q{x)y 0.
Prove that the values of \p{x)q(x)\ 1/2 \y(x)\ at the points where / = form an increas-
ing (decreasing) sequence if p(x)q(x) is an increasing (decreasing) function. [Hint:
One example of the way in which this point of view can be exploited was given
earlierwhen we derived all of the familiar properties of the functions sin x and
cos x from the fact that they satisfy the equation y" + y = (see Section 3-8).
Another is furnished by that treatment of the natural logarithm which sees this
= 1 subject to the initial
function as the solution of the first-order equation xy'
condition y{\) = 0. Together these examples pointedly illustrate the fact that
one of the most satisfactory ways of defining new functions in mathematics is as
solutions of differential equations.
In the remaining sections of this chapter we shall use the method of power
series expansions to study the solutions of a certain large class of linear differential
equations, or, rather, to study the functions defined by this class of equations.
Since this discussion depends upon certain results in the theory of power series
expansions of analytic functions we begin by reviewing some of the basic facts
concerning such series. assumed that the reader is already familiar with these
It is
results, and the real purpose of this resume is to fix terminology and provide
formulas for convenient reference.
In the first place, we recall that an expression of the form
00
22 a k (x - x )\ (6-15)
A;=0
with < r < oo . If, in the latter case, r is chosen as large as possible, the series
diverges when \x — x |
> r, and for this reason r is called the radius of con-
vergence of the series.*
Every power series with radius of convergence r defines a function / in the
interval \x — x \
< r, and we write
fix) = £
k=0
**(* ~ x o)
k
- (6-16)
a*) = 22 ka ^x - x of~\
k=\
* The behavior of a power series at the endpoints of its interval of convergence cannot
be predicted in advance.
6-5 | ANALYTIC SOLUTIONS 243
where the radius of convergence of each of these derived series is identical with
the radius of convergence of (6-16). In short, a power series may be differentiated
term-by-term without changing its radius of convergence.
Any function which can be represented by a convergent power series of the
form (6-16) in an open interval / about the point x is said to be analytic at x .
We have just seen that such a function must have derivatives of all orders every-
where in /, and, as might be expected, is actually analytic at each point of /. Thus
it is customary to speak of functions as being analytic on
an interval, and the
phrase "analytic at x " is used primarily to direct attention to the point about
which the series is expanded. Finally, the reader should be aware of the fact that
all of the elementary functions in mathematics—polynomials, exponentials,
trigonometric functions, etc. are analytic.— Indeed, this is one of the major
results of that chapter of calculus which deals with Taylor series expansions of
(analytic) functions, and will be used throughout the following discussion.
EXERCISES
1. Let a(l) denote the set of all functions analytic on an open interval / of the x-axis.
Prove that GL(I) is a real vector space under the "usual" definitions of addition and
scalar multiplication.
3. Prove that the power series expansion of an analytic function about the point xo is
unique, i.e., that such a function has precisely one such expansion.
Theorem 6-4. (The existence theorem for equations with analytic coef-
ficients.) Let
a* + *-iM sS + •
+ «<&» = *« (
6- 18 >
244 THE THEORY OF LINEAR DIFFERENTIAL EQUATIONS | CHAP. 6
an _ x (A:) and right-hand side h(x) are analytic in an open interval I. Let x
be an arbitrary point in I, and suppose that the power series expansions of
a (x), . h(x) all converge in the interval \x — x
. . , < r, r > 0. Then \
It is important to note that this theorem not only asserts the analyticity of
-2
£
fc=2
k(k - l)a k x
k
+ J] ka kx
fc=l
k
+ £
k=0
akx
k
= 0, (6-21)
and it follows that (6-20) will be a solution of the given equation if and only if
the a k are chosen so that the sum of the coefficients of like powers of x in this
expression are all zero (see Exercise 3 of the preceding section).
* For a proof of Theorem 6-4 the reader should consult E. A. Coddington, An Intro-
duction to Ordinary Differential Equations, Prentice-Hall, Englewood Cliffs, N. J., 1961.
f The student should appreciate that the choice xo = is one of convenience, not
necessity.
6-5 | ANALYTIC SOLUTIONS 245
-2 k
2 Kk -
fc=2
l)a kx
k
= X)
fc+2=2
(fe + 2)(k + l)a k+2 x
fc=0
whence
A + 2a 2 = 0,
The second of these equations is known as a recurrence relation (or /wi/te rfi/-
ference equation), and can be used to express the a k from A: = 3 onward in terms
of the preceding ones. Moreover, since a 2 = —(a Q /2), it follows that all of the
a k for k > 2 are uniquely determined by the values ofa and a x They fall into two .
k even k odd
ao
a2 = — a
2
fl i
as — — «i
3"
a* = 4^2 °5 " 5 3 •
°6 = ~ O
"7 _ _ q i
6-4-2 ~ 7-5-3
'
, ixfc #0 _
~ f_1\
k 9l
a 2k - (-1)
(2fcX2Jk - 2) • • 4 •
2
2fc+1 l
' (2* + 1X2* - 1) • • •
5 • 3
L*~T + 5^3~7.5-3"
(6-22)
+ fll h
To complete the problem it remains to show that (6-22) is the general solution
of the given equation on (—00, 00). To this end we introduce the series
^)=l + g(-D fc
(2 , )(2 ,
_*
2) ... 4-2
(6-23)
2k+1
x
k=1
(2k + \)(2k- l)-.-5-3
An easy computation using the ratio test now shows that y and yx both converge
for all Hence so does (6-24), and we conclude that this expression is a
values of x.
solution of (6-19) on the entire *-axis, just as predicted by Theorem 6-4. Finally,
we note that y and y 1 are themselves solutions of (6-19), and that
J>o(0) = 1, / (0) = 0,
J>i(0) = 0, y\(0) = 1.
Hence y and y x are linearly independent in e(— 00, 00) and therefore span the
solution space of (6-19) (cf. Theorem 3-3). This implies that (6-24) is the general
solution of the given equation, and we are done.
Example 2. The differential equation
y(x) = £
fc=0
akx
k
(6-26)
of its general solution about the point x = 0. Before we begin, however, we note
that since the leading coefficient of (6-25) vanishes when x = ±1, Theorem 6-4
only guarantees that the series in question will converge in the interval (—1, 1).
This said, we substitute (6-26) and its first two derivatives in (6-25) to obtain 1
-2 ~l
(1 - x 2 ) J^ k(k
k=2
- \)a k x
k
- 2x jj ka k x
k=l
k
+ X(X + 1) ^
k=0
"*** = °>
6-5 ANALYTIC SOLUTIONS 247
or
~2
2 Kk -
k=2
l)a kx
k
_]£*(*_
k=2
\)a k x
k
- £ 2ka x
fc=l
k
k
00
k
+ 2] MX + l)a kx = 0.
k =o
By shifting the index of summation on the first series and consolidating the last
three, this expression may be rewritten
To complete the solution we set the various coefficients in this last series equal
to zero and solve the resulting equations. This gives
and it is now an easy matter to express all of the ak in terms of a and a x Indeed,.
a2 = - —+j-^-
-(X 1)X
a ,
etc., while
a _ (X + 2)(X- 1)
«s - _ (X + 4)(X- 3)
03
_ (X + 4)(X + 2)(X - 1)(X - 3)
« l5
5 .
4 5|
248 THE THEORY OF LINEAR DIFFERENTIAL EQUATIONS |
CHAP. 6
etc. In general,
a2k = (-I)*
where
.. t
(X+1)X 2 ,
(X + 3)(X + 1)X(X - 2) 4
(6-29)
yo(0) = 1, /o(0) = 0,
yi(0) = 0, /i(0) = 1,
given above has only a finite number of nonzero terms, and hence is a poly-
y
nomial. In fact,a polynomial of degree In involving only even powers of x.
it is
In +
1 involving only odd powers of x. But since j> and y x are themselves solu-
tions of (6-25) we conclude that Legendre's equation has polynomial solutions for
each non-negative integral value of the parameter X. These polynomials will be
studied in considerable detail in Chapter 11, and we therefore say no more about
them here other than to remark that they provide examples of power series solutions
by Theorem 6-4.*
EXERCISES
Express the general solution of each of the following equations as a power series about the
point x = 0.
1. y" + y = 2. y" - y =
3. y" - 3xy = 4. y" --y = 3xy'
5. (x + 1)/' - 6y =
2
6. (x + 1)/' - 8xy' +
2
15y =
7. (2x 2+ 1)/' + 2xy' - lSy = 8. 2y" + 2x 2/ - ay =
9. y" + *V + 2*y = 10. /' - x 2y = 6x
11. 3y" - xy' + y = x 2 + 2x + 1 12. /" + 3a 2/' - 2y =
13. /" - 3xy' - y = 14. /" + x 2y' - xy =
Use the method of undetermined coefficients to express the general solution of each of the
following equations as a power series about the point x = 0, and specify an interval in
which the solution is valid.
2
15. (x - \)y" + xy' - 4y =
2 2
16. (x - 2)y" + xy' - y = x
v
17. y
// ,
x
y
25y 1 + 2x
x2 - 4 X2 - 4 x^ - 4
3 2
18. (x - S)y" + jc /+ xy = 16
3 2
19. (x + 2)^" + 6x y' - 6xy =
4 2
20. (a - 4)/" - 36a- / - 48xy =
21. Use the method of undetermined coefficients to solve the initial- value problem
y" + xy' - 2y = e x
;
y(0) = /(0) = 0.
[Hint: Expand e x
as a power series about x = 0.]
22. Find the general solution of Airy's equation/' + xy = 0.
23. Find a necessary and sufficient condition that a differential equation of the form
(x 2 + a)y" + fixy' + ly =
has a polynomial solution of degree n.
24. Let yo and yi be the series solutions of Legendre's equation given in (6-29).
(a) Find the radius of convergence of these series. [Hint: Use the ratio test.]
(b) Prove that up to constant multiples Legendre's equation of order n (n a non-
negative integer) has only one polynomial solution.
(c) Prove that the function
HNn)-
is a solution of Legendre's equation of order one on the interval (—1, 1), and use
this fact to write the general solution of the equation in closed form.
250 THE THEORY OF LINEAR DIFFERENTIAL EQUATIONS |
CHAP. 6
(b) Use the method of undetermined coefficients to find a basis for the solution space
of Hermite's equation.
(c) Show that Hermite's equation of order n has polynomial solutions of degree n
for each integer n > 0, and that up to constant multiples there is precisely one such
solution for each n.
*26. Use the method of undetermined coefficients to show that Bessel's equation of order
zero has a solution Jo which is analytic on the entire x-axis and satisfies the
condition /o(0) = 1.
immediately to a general formula for the coefficients of the series being sought.
In many cases neither of these simplifications occur, and in order to put this
discussion in a more reasonable perspective we now present a number of less
elementary examples.
By Theorem 6-4 we know that the desired solution can be expressed in the form
k
y(x) = J2 akx (6-31)
fc=0
for suitable constants a k , and that the resulting series will converge on (- oo, co).
Thus we substitute (6-31) and its first two derivatives in the given equation to
obtain
~1
k=2
£ 3*(* - I****"
2
~ E
fc=l
kakxk + £=°
fc
akxk+1 + E
k=0
«*** = l >
,
6-6 |
FURTHER EXAMPLES 251
fc=l fc=0
Collecting terms we have
6a 2 — «i + a
00
+ E
fc=l
t3 ^+ 2 X* .+ l)**+2 ~ (* + 1H+1 + «t + fl*_!]X* = 1,
6a 2 — a1 + a = 1,
In addition, by setting a: = in (6-31) and its first derivative, and using the
given initial conditions, we find that a = ax = 0. Hence
a = 0, ax = 0, a2 = £,
and, in general,
k+2
a fc+i afc + Qfc-i ,
K
. .
Here, for the first time, we are confronted with a recurrence relation which
cannot be solved for a k as a function of k alone. As suggested above, this is not
at all uncommon, and when it occurs we have no choice but to compute a few
terms of the series involved and then use Theorem 6-4 to determine a (minimal)
interval of convergence. In point of fact, this is usually sufficient for most pur-
poses, since it is always possible to develop the series to the point where it can be
used in numerical work.
In the present case we have
fl
3 — 54> #4 = — 324» a5 = ~ i&S >
and
y(x) = ±x 2 + &X - 3
^x 4
- T2 zx \ 5
+ • •
•
XI) =
xy» +?+
0, /(l)
xy =
=
0;
-1. ^ Z)
, , .
y(u ) = 2 akU
"2 '1
(« + i) £
fc=2
*(* - iK« fc
+ S
fc=l
kakuh + (M + !)
fc=0
S flfc "
fc
= °'
«o + fli + 2a 2
fc
it=i
Thus
flO+a + = l 2fl 2 0,
2
(k + 2)(A: + l)fl*+2 + (* + l) «fc+i + «fc + «*-i = °> ^ > 1.
02 = 2> fl 3 = — £> fl 4 = h
Hence
Xw) = -u + it/
2
- i«
3
+ i«
4
+ •
'
•
and, setting u = x — 1,
X*) = i - x + K* - i)
2
- K* - i)
3
+ K* - D 4 + • •
•
From Theorem 6-4 we conclude that this series converges at least in the interval
< x < 2, since this is the largest interval centered at x = 1 in which the
equation is normal.
Example 3. For our final example we solve the initial value problem
y(*) = 2
k=0
a **
k
-2
£
fc=2
k(k - l)a k x
k
- e
x
J2
&=0
ak x
k
= 0,
which we rewrite as
£
k=0
(k + 2)(/c + l)a k+2 x
k
- f E IrVZ
\fc=0 **/V=o
«**^ = 0.* (6-35)
/
In order to put this expression in the form required by the method of undetermined
coefficients (i.e., a power series in x), we now use the theorem which asserts
that power series may be multiplied according to the usual rules of algebra within
their common intervals of convergence. Thus
(flo + «i* + a 2x
2
+ • •
•) (l + x + ~+ • •
•)
= a + (a + a x )x + (f[
+ ai + a z) x2
1 / fl o , . \ a + «i
x
Recall that e = 1 + jc + x 2 /2! + *3/3! +
254 THE THEORY OF LINEAR DIFFERENTIAL EQUATIONS |
CHAP. 6
etc., and in principle all of the a k can now be computed in terms of a and a x .
#2 = 2">
n — i
«4 = 6"'
and
4
x2 x2 x
The validity of the above computations obviously depends upon the fact that
convergent power series can be multiplied as though they were polynomials. Since
we shall have occasion to refer to this result again we now state it precisely and
formally as
J2a k x
fc=0
k
and £
fc=o
b kx
k
(6-36)
£
fc=0
c k x\ (6-37)
with
k
ck = 2
3=0
ai b *-i> (6-38)
known as the Cauchy product of the series in (6-36), also converges for
\x\ < r, and
(£
\fc=0
**A ( S b ^)
\fc=o
/ /
= E
fc=o
c **
fc
<6
-39
>
When phrased in terms of analytic functions this theorem asserts that the
product of two functions/and g which are analytic on an interval lis itself analytic
on /, and that its power series expansion about any point x in / is the Cauchy
*
product of the power series expansions of/ and g about x
* For a proof the reader should consult Buck, Advanced Calculus, 2nd Ed., McGraw-
Hill, New York, 1965.
6-6 | FURTHER EXAMPLES 255
EXERCISES
Find the first four nonzero terms in the series expansion of the solution of each of the
following initial value problems, and determine a (minimal) interval of convergence for
the series.
1. /' + (sinx)y = 0; 2. 2y" -/+(* + l)y = 0;
y(0) = l,/(0) = 0. v(0) = 0, /(0) = 1.
7. (* - 3)y" + jc /+
2
y = 0; 8. [1 + ln(l + x)]y" - xy' +y = sin x;
j/(0) = 0, /(0) = 6. y(0) = = 1.
0, v'(0)
What is the interval of convergence of the power series expansion of this integral?
14. Let
00
k=0
be a solution of the equation
/>(*) = 2
fc=0
****** ^*) = S
k=0
0***
fork = 0, 1, 2,
.
euclidean spaces
Much of the content of elementary geometry depends upon the ability to measure
distance between points. In this chapter we shall show how distance, together
with such related concepts as length and angular measure, can be generalized to
arbitrary real vector spaces. These so-called "metric" concepts are the foundation
of Euclidean geometry, and from them flow a wealth of results in both geometry
and analysis.
In order to introduce a metric into a real vector space we must first choose a
unit of distance for our measurements. This can be done most easily by defining
what is known as an inner product on a vector space. The logic behind taking the
notion of inner product as primitive will become compelling once we have used it
to define length, angular measure, and distance. For each of these concepts then
appears as a natural consequence of the notion of inner product, and the student
is led to appreciate them as an elaboration of a single idea.
x y • = y • x, (7-1)
(ax) •
y = a(x •
y) for every real number a, (7-2)
(xi + x 2) y• = X! •
y + x2 •
y, (7-3)
x •
x > 0, and x • x = if and only if x = 0. (7-4)
* Some authors call x • y the "scalar product" of x and y. We shall avoid this termi-
nology, however, because of the possibility of confusing it with scalar multiplication as
introduced in Chapter 1
256
7-1 | INNER PRODUCTS 257
If we apply Eq. (7-1) to (7-2) and (7-3), we see that an inner product also
satisfies
x •
(yi + y 2) = x •
y2 + x y2
• and x (ay)
• = a(x •
y). (7-5)
(aiXi + a 2 x 2) •
CSiyi + 2 y 2) = «i/Si(xi •
yi) + aij8 2 (xx •
y 2)
(/—o)
+ a 2jSi(x 2 •
y x) + a 2 p 2 (x2 '
y 2 ),
where a\, a 2 , j8i, /8 2 are arbitrary scalar s. For future reference we also note that
x y • = whenever x = or y = 0. (7-7)
[The proofs of (7-6) and (7-7) have been left to the reader in Exercises 6 and 7
below.]
Equation (7-1) in the above definition asserts that an inner product is a com-
mutative or symmetric operation on pairs of vectors. Equation (7-2) may be
interpreted as an associativity requirement, this time with respect to scalars, while
(7-3) requires that the operation be distributive. These two conditions, together
with their analogs given in (7-5), are said to make the inner product bilinear.
Finally, (7-4) is referred to by saying that an inner product is positive definite,
the allusion here being to the fact that the inner product of a vector with itself is
always greater than zero unless the vector involved is the zero vector. Thus one
frequently hears an inner product called a real-valued, symmetric, bilinear, positive
definite operation on pairs of vectors.
Example 1. Let x = = n
(jc 1s . .
.
, xn ) and y (y 1} . .
. , yn ) be vectors in <5t , and
define x y by •
x y • = *iJ>i + • • •
+ xnyn . (7-8)
n
Then becomes a Euclidean space, and as such is called Euclidean n-space.
(R
2
In and (R 3 this inner product is none other than the familiar "dot product"
(R
of physics, where the definition is usually phrased geometrically as the product
of the length of x, the length of y, and the cosine of the angle between them. The
equivalence of these definitions will become apparent in the next section (see
Eq. 7-21).
b
fg = [ f(x)g(x)dx. (7-9)
Ja
It is not difficult to show that (7-9) satisfies all of the requirements for an inner
product. Indeed, it is perfectly obvious that f •
g is both real-valued and sym-
258 EUCLIDEAN SPACES |
CHAP. 7
and
b b
b
2
f.f = f f(x) dx > 0,
Ja
and
f • f = if and only if f = 0.
n we
Hereafter whenever we refer to (R or e[a, b] as Euclidean spaces, shall
assume that we are using the inner products defined in the above examples unless
express mention is made to the contrary.
b
f-g= [
f(x)g(x)r(x)dx (7-10)
Ja
product defined on W coincides with the inner product defined onl). clear It is
in this sense provided we use as the inner product on W the one that defined is
EXERCISES
y = (i,i,3)
2. Find f • g for each of the following pairs of vectors in C[0, 1]. (Recall that the inner
product is defined by (7-9).)
(a) f(x) = x (b) f(x) = x
gix) = 1 - x2 g(x) = 1 - x
(c) f(x) = sinTTAr/2 (d) f(x) = ex
gix) = costtx/2 g(x) = sin*
(e) f(x) = |jc- i|
six) = 4 - I* - 4|
3. Find f • g for each of the following pairs of vectors in 6[0, 1] when the inner product
is defined with respect to the weight function r(x) = e x (see (7-10)).
(a) fix) = 1 - 2x (b) fix) = x 2
gix) = e~ x gix) = e
x
(a) x y = xiyi
•
(b)x y = 2ix x yi + x 2 yz)
•
l
f-g = C
Ja
Rx)g{x)dx.
i
p •
q = aobo + aibi + • • •
+ a n bn ,
(Note that by adding terms with zero coefficients we can make any two polynomials
in (P have the same apparent degree, as above.)
(b) Is (P with this inner product a subspace of the Euclidean space Q[a, b]l Why?
11. Let 13 be a Euclidean space with inner product x •
y.
xoy = 2(x-y).
xoy = a(x •
y).
Determine those values of a for which this definition yields an inner product on 13.
*12. (a) Let /be a continuous function on the interval [a, b], and suppose f(x ) > for
some xo in this interval. Use the definition of continuity to prove that fix) >
for all values of x in some subinterval of [a, b] containing the point *o-
(b) Use the result in (a) to prove that in C[a, b] f f > 0, and that f f = if and • •
only if f = 0.
14. Will (7-10) define an inner product on Q[a, b] if we merely require r to be non-
negative on [a, b\l Why?
15. Let r be any function in Q[a, b] which vanishes for at most finitely many values in the
interval [a, b]. Prove that
/an «i2\
=
(fl ^ L
\a21 a )
#22/
7-2 I LENGTH, ANGULAR MEASURE, DISTANCE 261
(a) Show that this definition satisfies Eqs. (7-2) and (7-3) of Definition 7-1 for
every 2x2 matrix (a„).
(b) Show that Eq. (7-1) is satisfied if and only if a 12 = 021 [i.e., if and only if (a,-,-)
2
is a symmetric matrix], and hence deduce that (7-11) defines an inner product on (R
if and only if (an) is a 2 X 2 symmetric matrix such that
2 2
aii*i + (ai2 + a2i)xiX2 + a22*2
is non-negative for every choice of x 1 and X2, and is zero if and only if x 1 = X2 = 0.
2
(c) Find a matrix (an) which reduces (7-11) to the ordinary inner product on (R .
(d) Determine which of the following matrices can be used to define an inner product
on(R 2 :
(::)•(-::>(::)
*17. Generalize the preceding exercise to (R n .
X=(*!,X2 )
(see Fig. 7-1). But this expression may be
rewritten in terms of the inner product
2
on (R (Formula 7-8) as
l|x|| = Vx^x,
Ixll = Vx • x. (7-12)
262 EUCLIDEAN SPACES |
CHAP. 7
= w
Thus, in particular, the length of a vector x (xi, . . . , xn ) in (R is
||x|| = Vxi + • • •
+ £, (7-13)
llfll
= (jjixfdx) 112 . (7-14)
2
Next, we observe that if x and y are any two nonzero vectors in (R , the formula
cos v = *,7 I,
' < e < x, (7-15)
is an immediate consequence of the law of cosines (see Exercise 6). But the ex-
pression
xy
|x|| ||y||
IMI llyll
for every pair of nonzero vectors in a Euclidean space, since, of course, any
definition of cos must satisfy the inequality -1 cos 1. This fact will < <
emerge as a consequence of the following important result, known as the Schwarz
or Cauchy-Schwarz inequality.
Theorem 7-1. (Schwarz inequality.) Ifx and y are any two vectors in a
Euclidean space, then
(x •
y)
2
< (x • x)(y •
y). (7-17)
= 2
a (x • x) - 2a/3(x •
y) + /3
2
(y •
y),
whence
2a/3(x •
y) < a 2 (x • x) + /r(y •
y)
We now set
= Vy •
y an d = Vx • x.
7-2 I LENGTH, ANGULAR MEASURE, DISTANCE 263
This gives
2Vx-x Vyy (x •
y) < 2(x • x)(y •
y),
or
x y
• < Vx» x Vy -y.
Squaring, we obtain (7-17). |
n
In (R the Schwarz inequality assumes the form
,2 / n \ / n
-
* E*? E-rf < 7 18 >
(IH' L i=i / \t=i
2 2
< (£ /W 2 dx)
(/aV(*M*) ^) (£ g(x) </*)
•
(7-19)
and asserts that the absolute value of the inner product of two vectors does not
exceed the product of the lengths of the vectors. Thus
whenever x and y are nonzero. But this is just another way of writing (7-16),
the inequality needed to justify using (7-15) as a definition of cos 6, and we can
now state
cose = *'* •
(7-21)
I,
Ml llyll
If, on the other hand, one of the vectors is zero, we set cos 6 = 0.
It goes without saying that in defining the cosine of the angle between x and y
we have, by implication, also defined the angle in question; just take the principal
value of the inverse cosine.
264 EUCLIDEAN SPACES [ CHAP. 7
At this point all that remains of our original program is to define the distance
between any two points (i.e., vectors) in a Euclidean space. Again this is done
2
simply by copying the definition from (R where the distance between x and y
is the length of the vector x — y (Fig. 7-2). Thus
But is this a reasonable definition of the term "distance"? In order to answer this
question we must first decide what properties we require of distance in general.
The first three of these properties follow immediately from the definition of
length and the axioms governing an inner product. The last, however, is not
quite so obvious. To prove it we first establish an inequality which is of some
Lemma 7-1. Ifx and y are arbitrary vectors in a Euclidean space, then
||x + y|l
< [|x|| + ||y||. (7-27)
7-2 | LENGTH, ANGULAR MEASURE, DISTANCE 265
Proof.
1/2
||x + y||
= [(x + y).(x + y)]
= Vx-x + Vy «y
= NI + ||y||.l
The triangle inequality follows at once from this result. Indeed,
and
d(x + z, y + z) = d(x, y). (7-30)
EXERCISES
4
1. Find the length of each of the following vectors in (ft .
x y •
cos 6 =
Ml ||y||
2
for any pair of nonzero vectors x, y in (R (see Fig. 7-3).
7. (a) Compute f •
g for each of the following pairs of vectors in C[— x, w]:
(i) f(x) = sinmx, g(x) = sinnx,
(ii) f(x) = sin/nx, g(x) = cosnx,
(iii) f(x) — cos mx, g(x) — cos nx,
where m and n are arbitrary non-negative integers,
(b) What can you say about the functions
9. Let a i, .
.
, a n be positive real numbers. Prove that
(ax + • • •
+ an ) (!+... + 1) > n
2
*10. Let a,b,c be positive real numbers such that a + b + c = 1. Use Cauchy's
inequality to prove that
e-)<s-)c-)»
11. Prove that the following inequality holds for any collection of real numbers
a\, .an . , \
(a+^y < 2
fll +,
valid forany pair of vectors x and y in a Euclidean space. Expand this inner product,
and derive the Schwarz inequality by examining the discriminant of the resulting
inequality in t (This is another very popular way of deriving the Schwarz inequality.)
.
*14. Set x =
z in (7-26) and use (7-24) and (7-25) to deduce (7-23), thus showing that
this relation is actually implied by the other three.
15. Show that ||ax|| = \a\ ||x|| for all real numbers a.
*19. Let
(an ai2\
£21 A22/)
be a 2 X 2 matrix whose entries are real numbers, and suppose that (a,,) is so chosen
that (7-11) an inner product on (R 2 (see Exercise 16 in the preceding
is section).
Use the Schwarz inequality to deduce that
«12
2 ^
< #11^22.
Conversely, show that if (a,,) satisfies this inequality, and ai2 = «2i, then (7-11)
defines an inner product on (R 2 .
268 EUCLIDEAN SPACES CHAP. 7
7-3 ORTHOGONALITY
Two vectors in a Euclidean space are said to be orthogonal or perpendicular if the
cosine of the angle between them is zero. Referring to Definition 7-3 we see that
the zero vector othogonal to everything, and, in general, that x and y are orthog-
is
l|x + y||
2 = ||x|j
2
+||y||
2
.
Proof.
I|x + y||
2
= (x + y)-(x + y)
= x x + 2(x y) +
• •
y •
y
= 2
+ 2(x.y)+
||x|| ||y||
2
.
Xi-xj = (7-31)
xt - •
Xi = 1 (7-32)
= (0
= L
if /'
^ j, (7-33)
Xi-xj 8 i3; where 8 {j .
= .
if
for introducing orthonormal sets at all is for the convenience which sometimes
results from working with unit vectors.
Before giving examples, two points in the above definition.
we call attention to
(see Example 3 below). Such sets will occur repeatedly in Chapters 9 and 11.
f+g=l+ sinx/^ n,
\. / /% = sin x \ ,
— 7T 7P
FIGURE 7-5
Example 1. In (R
3
the vectors (1, 0, 0), (0, 2, 0), (0, 0, — $) form an orthogonal
set, while the standard basis vectors (1,0, 0), (0, 1, 0), (0, 0, 1) form an ortho-
normal set. More generally, the set consisting of the standard basis vectors in
n
(R is orthonormal.
+ bi sin x + b 2 sin 2x + • • •
+ bn sin nx, (7-34)
denote the set of all trigonometric polynomials of degree < 2/i 1, together +
with the zero polynomial. We make 3 n a Euclidean space by defining addition
and scalar multiplication of trigonometric polynomials termwise, as with ordinary
.
T
fg =
f_J(x)g{x)dx.
(7-35)
J
sin mx sin nx dx = 0, if m j± n,
J
cos mx cos nx dx = 0, if m 5^ n.
J
1 dx = 2ir,
*
and (7-38)
2
f
sin
2
mx dx = f cos mx dx — v, if m > 0.
Example 3. It follows from the preceding example that the (infinite) set
every one of its finite subsets is linearly independent in the earlier sense of the
term. Thus, for example, the vectors 1, x, x2 , ... are linearly independent in the
space of polynomials (P. (Proof?)
linearly independent.
.
a lXl + • •
+ a nx n = 0, (7-40)
(aiXi + • • •
+ a„X n ) • x, = • Xt - = 0.
a t (x t x t )
• = 0,
(7-40) is zero, and the assertion follows from the test for linear independence. |
P •
q = j P(x)q(x) dx,
the polynomials 1, x, x2 — % are mutually orthogonal, and hence are a basis for
this space.
Example 5. We saw above that the functions 1, cos x, sin x, ... , cos nx, sin nx
are mutually orthogonal in 3 n , the space of all trigonometric polynomials of
degree < 2n 1. +
Hence these functions are linearly independent in 3 n More- .
over since every vector in this space is a linear combination of these functions it
follows that they form a basis for 3 n , and hence that dim 3n = 2n + 1.
in e[— 7r, ir] implies that this set is also linearly independent. Combined with the
fact that any n + 1 vectors in an n-dimensional space are linearly dependent, we
conclude that Q[—ir, ir] is infinite dimensional.
EXERCISES
1. Verify that the Pythagorean theorem holds for the orthogonal functions sin x, cos x
in C[— 7r, 7t].
*2. Let xi, . .
.
, x„ be mutually perpendicular vectors in a Euclidean space. Prove that
||Xl+---+X„|| 2 = ||
Xl ||2 + ...
+ || Xn p.
3. Prove that the polynomials 1, x, x 2 — ^ are mutually orthogonal in (P3 with the
inner product defined as in Example 4 of the text.
4. Let x be a nonzero vector in a Euclidean space. Show that the vector x/||x|| is of
length 1.
6. Find a polynomial of unit length in (P3 which is orthogonal to 1 and x 2 . (Use the
inner product of Example 4 of the text.)
7. Find a polynomial of degree 3 which is orthogonal to l,x, x 2 in (P4. (Use the
inner product of Example 4 of the text.)
8. Find a vector of unit length in (R 3 which is orthogonal to the vectors x = (1,-1,0),
y = (2,1,-1).
9. Find a vector of length 2 in (R 4 which is orthogonal to the vectors x = (1, 0, 3, 1),
y = (—1, 2, 1, 1), z = (2, —3, 0, —1), and has as its second component.
10. Let x and y be linearly independent vectors in (R 3 Prove that there exist precisely .
12. Let x and y be arbitrary vectors in a Euclidean space, and suppose that ||x|| = ||y||.
Prove that x y and x — y are orthogonal. Interpret this result geometrically.
+
*13. Suppose xi, . .
.
, xn is a finite orthonormal set in a Euclidean space V. Prove that
for any vector x in V
n
X~^ t \2 ^ II II
2
1=1
This inequality is a special case of BesseVs inequality which will be proved in general
in Section 8-4.
*14. Let xi, . .
.
, x„ be an orthonormal basis for a Euclidean space V.
(a) If x is any vector in 13, prove that
n
2 2
||x|| - X>'X<) .
This result is known as ParsevaVs equality, and will be proved in more general terms
in Section 8-4.
(b) If x and y are any two vectors in V, prove that
x y =•
2 (x • x»)(y • x t ).
Euclidean spaced, and if the equality in (b) is valid for every pair of vectors x, y inT),
then x 1, . .
.
, x„ is a basis for 1).
7-4 |
ORTHOGONAUZATION 273
7-4 ORTHOGONAUZATION
We now know that every orthogonal set of vectors in a Euclidean space is linearly
independent (Theorem 7-3). However, would be of only passing
this result
interest were it not for the fact that a Euclidean space also contains "enough"
orthogonal vectors to enable us to replace a given linearly independent set by an
equivalent orthogonal one. More precisely, in this section we shall prove that any
(finite or infinite) linearly independent set X in a Euclidean space can be con-
verted into an orthogonal set which spans the subspace S(9C). This process of
orthogonalizing a linearly independent set, as it is called, has a number of im-
portant and useful consequences, not the least of which are the computational
simplifications which result from working with orthogonal vectors.
e2=x2 — 0*1
FIGURE 7-6
Rather than begin with the most general situation, we shall introduce the
orthogonalization process by two examples. The drawn from (R 2 where
first is
2
a basis for (R ,and in this case our problem becomes that of replacing x x and x 2
with an orthogonal basis e
ei, 2 constructed out of x x and x 2 in some reasonable
way. Figure 7-6 suggests the most natural solution of our problem; simply take
ei = xi, and then let e 2 be the "component" of x 2 perpendicular to Xi. Thus
we write e 2 in the form
e 2 = x 2 — 0*1,
With this, e 2 has been determined in terms of Xi (= ei) and x 2 and the basis
,
(0,1) .(1,1) 1
(1,1) -0,1) 2
and so
ei = (1,1), e2 = (-4,4).
Figure 7-7 shows that this is exactly the result one would expect on the basis of
our earlier remarks.
I /
i /
Xi = e! = (U),
x2 = (0,l)
e
2=(-!J)^
I
FIGURE 7-7
The procedure is essentially the same as that used above in (R 2 and is started by ,
e 2 'ei = 0, e2 = x2 — ae u
which gives again
a = x 2 -ei •
ei-d
It is clear that e 2 is not the zero vector (why?), and also that ei and e 2 both
3
belong to the subspace of (R spanned by x x and x 2 Hence S(ei, e 2 ) is a subspace .
of S(x l5 x 2 ). Moreover, since the orthogonal vectors ei, e 2 are linearly inde-
pendent, S(e i, e 2 ) has the same dimension as S(x i, x 2 ). Thus
S(ei, e 2 ) = S(xi,x 2 ).
3
Combined with the x 2 x 3 form a basis for (R this equality implies
fact that xi, , ,
3
that x 3 does not belong to the subspace of (R spanned by ei and e 2 Referring to .
e3 = x3 — «ie! — a 2e 2 ,
e
e3 = x 3 — («l e l + «2 e 2>
FIGURE 7-8
and
= x3 •
2 — a 2 (e 2 • e 2 ),
whence
x 3 -e 2
<*i
= x3 -ei
«2"
ei •
ei e2 • e2
e3 = (0,0,1) /
/
/x3 = (i,i,D
i
/ J^^-H* >
y\ = (0,1,0)
Xl = ei = (1,1,0)
FIGURE 7-9
276 EUCLIDEAN SPACES | CHAP. 7
We can now quickly dispose of the general situation in which we are required
to orthogonalize an arbitrary set of linearly independent vectors x l5 x 2 , . . .
eM+ i = x n+ i — a^ei — a 2e 2 — • • • — an e n ,
x n +i • e2 — a 2 (e 2 • e 2) = 0,
and so to
— Xn+i
X •
ei
= x„+i
t:
• e2
= x„+i
— • en
ai ,
a2 , . . . ,
an '
ei •
ei e2 • e2 en • en
which determines en+ i. Again the linear independence of xi, xn+ i implies . .
.
,
(See Exercise 9 below.) Thus the orthogonalization process has been continued,
as required, and we can now state the following important result.
ei = xx, (7-41)
and,
en+ i = x w+ — i ouei — • • • — anen ,
(7-42)
* The knowledgeable reader will recognize that we are giving a proof by mathematical
induction at this point.
7-4 I ORTHOGONALIZATION 277
where
x fi x n+ i e2
= —n: +i — • •
a\ ;
i
' a2 = (7-43)
d-ei e2 • e2
X j Jv 9 A 5 • • •
P •
q = P(x)q(x) dx. (7-44)
f
The orthogonalization goes as follows:
(0 ei = 1
ei • ei = /
dx = 2.
(ii) e2 = x — a, where a = ^ }_ 1 x dx — 0.
Thus
e<> = x
=
e2 • e2 = £/ dx §.
(iii) e3 = x2 — ai — a 2 x, where
ai = i /*
jc
2
dx = £, a2 = f| x 3 dx = 0.
Thus
e3 — * 3
f\x -&
2 2
e 3 -e 3 = dx = &.
1
3, Av 3 Hy v-*
5A, A 3.Y
7 A 1
~T 35
3
(7-45)
in (P.
278 EUCLIDEAN SPACES I CHAP. 7
X = aiei + • • •
+ a n en I
•
is any vector in V, then since e* • e,- = 5 t y, /
s *3 s \
/ X 1
s •
/
x • e,- = on
form 1 •
s
1
s
x = (x •
eOex + • • •
+ (x • e„)e* /*\ 1 x
xy = aijSi + • • •
+ a„/3n . (7-47)
Verbally, this result may be expressed by saying that the inner product of two
vectors in a finite dimensional Euclidean space is the sum of the products of their
EXERCISES
p •
q = aob + a\b\ -\ + a„6„,
where
p(x) = a + a\x -\ + anx n ,
q(*) = b + bix H \- bnxn .
(See Exercise 10, Section 7-1.) If a is an arbitrary constant, show that the polyno-
mials
1, x - a, (x - a) 2 , . .
. , (x - a) n , . .
e£ = (1,1,..., 0),
e'n = (1, 1, . .
.
, 1),
in(R\
9. In passing from step n to n + 1 in the orthogonalization process (Theorem 7-4) we
had the following situation:
(i) S(ei, . . e„) = S(xi,
. , . . . ,x„), with ei, . . . , e„ mutually orthogonal, and
(ii) e„ + i = xn+ i — aid — • • — a»e„, where ai,...,a„ are computed
according to (7-43).
.
(a) e n+ i ?* 0, and
(b) S(ei, . .
.
, e n +i) = S(xi, . . . ,x n+ i). Prove these statements.
10. Write out a proof of Eq. (7-47) of the text.
Find a vector of unit length in (ft which is orthogonal to the subspace spanned by
3
13.
the vectors (1, 2, -1) and (-1,0, 2). (See Exercise 12.)
2
14. (a) Orthogonalize the set of vectors 1, sin x, sin x in C[-t, tt].
(b) Use the result in (a) to find a unit vector which is orthogonal to the subspace of
e[—n-, it] spanned by 1, sin x. (See Exercise 12.)
15. (a) Let ei, e2 be an orthonormal set in a Euclidean space V, and let be the sub- W
space spanned by ei and e2. If x is an arbitrary vector in V, prove that there exists
precisely one vector y in such that x - y is orthogonal to (see Exercise 12). W W
(b) Find y if V = (ft
3
, ei = (1, 0, 0), e 2 = (0, V2/2, V2/2), and x = (1, 1, 1).
16. Let po(x) pi(x) ..., Pn(x), ... be the sequence of polynomials obtained by ortho-
t t
2 in C[— 1, 1].
gonalizing the sequence 1, x, x , . . .
17. Legendre polynomials. The Legendre polynomial Pn (x) may be computed by the
following general formula where n successively assumes the values 0, 1, 2, .
.
:
n(n — 1) n-2
^nW 2(2« - 1) X
2«(/z!)2
polynomials.
(i) 3x 2 - 2x + 1
Proof. Since V? is finite dimensional, we can apply Corollary 7-2 and find an
orthonormal basis e ls , en for W. Then if it exists at all, the vector y of
. .
.
and it remains to show that the a* can be so determined that the vector d = x — y
is orthogonal to each of the basis vectors e z-.t But if we substitute (7-50) into
of that step.
282 EUCLIDEAN SPACES CHAP. 7
(7-49) to obtain
x = axd + • • •
+ a„e n + d,
d • et = 0, ez - • ey = 5,-y,
we find that
This set of equations determines y uniquely in terms of x and the basis vectors
ei, . .
.
, en as
y = (x •
ei)ei + • • •
+ (x • e n)e n . (7-52)
We now show that the vector d, which is called the component of x perpendicular
to W, canbe used to measure the distance from x to W. To do so, let z 9^ d be
any vector in V such that x — z belongs
to W. (One can view z as a vector from
W to the point x; see Fig. 7-12.) Then
since z — d can be written as the differ-
ence of two vectors in *W, it too belongs
to "W, and so is orthogonal to d. It now
follows from the Pythagorean theorem
that
|2 _
= ||d||
2
+ Z-d| FIGURE 7-12
and hence, since ||z - d|| > 0, that ||z|| > ||d||. In other words, of all vectors z
any other vector in «W, then ||x - w|| is greater than ||d||.
x — d = a^i -f + a«e n .
This time the requirement that d be orthogonal to each of the e t leads to the -
(en •
ei)«i + (e„ • e 2 )a 2 + • • •
+ (e n • e„)an = en • x,
in the unknowns a ls . .
. , a n which must be solved
, in order to find d. Our earlier
results guarantee that this system has a unique solution since we know that d is
uniquely determined by x and W.
Digressing for a moment, we recall that a system of n linear equations in n
unknowns has a unique solution // and only if the determinant of its coefficients
is different from zero.* Hence
ei-ei ei • e2 ei • e„
e 2 «ei e2 • e2 e2 • e„
^ 0. (7-54)
en -ei e2
construction of such a basis a tedious job at best. In essence, the issue here is
is
that one ought to attack a simple problem directly, rather than try to adapt it to
fit some general formula. Each of the following examples illustrates this point.
* The student who is unfamiliar with determinants should omit this paragraph, and
resume reading after the statement of Theorem 7-6.
284 EUCLIDEAN SPACES CHAP. 7
= d •
y = x •
y - a(y •
y),
and
a = xy •
yy
It follows that
1/2
|d|| = (d-d)
= [(x - ay) • (x - ay)]
1/2
= [x • x - 2a(x y) • + a
2
(y •
y)]
1/2
2 2 1/2
(x-y) (x-y)
x x • 2 j
yy yy J
(x • x)(y -
y) - (x •
y)
1/2
= (7-55)
lldll
yy
for the distance from x to the line determined by the vector y ^ in any Euclidean
space.
2
This formula assumes a particularly simple form in (R . For if x = (x lt x 2 )
and y = (yi,y 2 ), then
1/2
\x\ + xl)(yl + yl) - (xij>i + x 2y 2 f
Idll =
+ y%z yi
2 i
1/2
\x\y\ - 2x y 2 x 2yx + x%y\
x
2, .2
2 2
yi + y2
i
1/2
(x x y 2 — x 2 yxf
2 2
yi + i
y2
j>2 + 1
j>3 "J
A similar formula can be established for (ft", but it is obviously easier to use
(7-55) directly.
x yi • - «(yi •
y0 - 0(y 2 •
yi) = 0,
x y2 • - «(yi •
y2) - /3(y 2 • y2) = 0.
Computing the values of the various inner products involved, we find that these
equations become
1 - a - /3 =
6 - a - 3/3 = 0,
d " V4 + "
4 2
Example 3. The Fourier coefficients offix). Let e[— x, t] be the space of con-
tinuous functions on the interval [— t, tt] with the usual inner product
T(x) = -y + ai cos x + • • •
+ an cos nx
+ bi sin x + " • •
+ bn sin nx,
is orthogonal to each of the functions 1, cos x, sin x, ... , cos nx, sin nx. But, by
(7-37), these functions are mutually orthogonal in e[— w, t], and so, by taking
inner products, we have
=
J_ T f(x)dx-^J_Jx
0,
J
fix) sin xdx — bn I sin
2
x dx = 0,
/ f(x) cos nx dx — an I
2
cos nx dx = 0,
Since
f
/ dx = 2ir,
and
2
/ sin
2
mxdx = J
cos mxdx = x
a = - f(x)dx,
7T J _,
an = - I
f(x ) cos nxdx, bn = - /
/(x)sinwxdx.
7T 7 -t * J —r
7-5 I
PERPENDICULAR PROJECTIONS; DISTANCE TO A SUBSPACE 287
These coefficients are known as the Fourier coefficients of the function / (The
student should now appreciate the reason for associating the factor \ with the
constant term of the trigonometric polynomials in 3 n It was done simply to .
assure that each of the Fourier coefficients has the same constant before the
integral, for without the \ the formula for a would have been (l/2ir)f_ r f(x) dx.)
We have now shown that the trigonometric polynomial T whose coefficients are
the Fourier coefficients of/ is the best approximation in 3 n to the function/, in
the sense that of all the functions P belonging to 3„, T is the one which minimizes
the integral
T
lV(*) - P(x)] dx.
2
(7-59)
/_ t
The value of this integral is often called the mean deviation (or mean square de-
viation) of P from / In these terms, T is that trigonometric polynomial in 3 n
with minimum mean deviation from /
These considerations lead one naturally to the problem of determining whether
the mean deviation of T from / tends to zero as n —> oo In other words, can f .
EXERCISES
1. Find the distance from each of the following points x in (R 2 to the line (through the
origin) determined by the point y.
(a) x = (1, 1, 1); y = (1, 1,0) (b) x = (0, 1, 1); y = (1, 1,0)
(c) x = (2,1,-3); y = (-1,2,2)
(d) x = (-£, 1, 3/V3); y = (4, 3, 5V3)
3. Find the distance from each of the following points x in 4
(R to the line (through the
origin) determined by the point y.
7. Find the distance from each of the following points x in (R 3 to the plane (through the
origin) determined by the given points y and z.
9. Find the coordinates of the point in (R 3 which lies on the plane determined by the
origin, (1, 2, 2), and (2, —2, 1) and is at a minimum distance from the point (3, 1, 1).
2 x 2 ), and
10. (a) Show that the area of the triangle in (R with vertices at (0, 0), (*i,
(yi,y2)is %\xiy2 x 2 yi\. -
(b)More generally, show that the area of the triangle in (R n with vertices at 0, x, and
yisH(x-x)(y.y) - (x-y) 2 ] 1 ' 2 .
11. Find the perpendicular projection of each of the following vectors in &[—ir, ir] onto
the indicated subspace W, and compute the distance from the vector to the subspace.
(a) f(x) = x, W = S(l,cosx,sinx)
(b) f(x) = cos
2
*, W = S(l,cos2x)
(c ) f( x)
= x2 , w = S(l,cosx,cos2x)
12. Repeat Exercise 11 for the following vectors and subspaces of C[— ir, ir].
(a) f(x) = x 2
, W = S(sin x, sin 2x)
(b) f(x) = jc
3
, W = S(l, cos x, cos 2x)
(c) f(x) = x
3
, W = S(sin x, sin 2x)
In Exercises 13-19 find the Fourier coefficients (for all values of n) for the given function
in C[— 7r, ir].
13. fix) = x 14. f{x) = x
2 15. f(x) = sin 2 x
m ma, positive integer =
16. fix) = x ,
17. fix) \x\
(a) f(x) = x
n
, n an integer (b) fix) = sin x
(c) fix) = |xl (d) fix) = 3x 2
22. Let ei, . .
.
, e„ be an orthonormal basis in a Euclidean space V, and let W be the
m-dimensional subspace of V spanned by ei, em m < n. . .
. , ,
23. Let 9C be an arbitrary nonempty subset of a Euclidean space V, and let 9C- - (read 1
"9C perp") denote the set of all vectors in 13 which are orthogonal to every vector in 9C.
Show that 9C-1 is a subspace of U What is 9C
X if 9C = V? If 9C contains only the
zero vector?
24. With 9C and 9C X as in Exercise 23, let (9C- -)- - denote the set of all vectors in
1 1
V which
are orthogonal to every vector in 9C
X .
s = {ao,ai, . . . ,a„, . .
.}
of real numbers having only a finite number of nonzero entries, with addition and
scalar multiplication defined termwise. If
and
t = {bo, b\, ...,£»,...},
define
00
S • t = 2^i a nbn-
n=0
Show that with this definition V becomes an infinite dimensional inner product space,
(b) Let W be the subspace of V consisting of all sequences of the form
26. Let V be a finite dimensional Euclidean space, and let W be a subspace of V. Show
that
v = *w + w- 1
,
and that WnV 1 contains only the zero vector (see Exercise 22, Section 1-4).
27. Find W- where 1
, W
is the subspace of
(a) (1,1,0)
would equal c, but in practice, of course, none of them will, and we are thus faced
with the problem of finding the "best approximation" to c available from our
experimental data.
To do so, we view the n experimental measurements as a vector x = (xi, . . . ,xn )
w
in (R , and let y be the n-tuple (1, . .
.
, 1), so that
cy = (c, . .
. , c).
xry
yy y,
so that
Xi + + Xn
yy FIGURE 7-16
This, of course, is none other than the arithmetic average of the xt and we now ,
X2 ~ (-X21j *22)j
Xn — (-Knlj Xn 2)
7-6 I
THE METHOD OF LEAST SQUARES 291
of c. Again we wish to use the x t to obtain the best possible approximation c' =
(ci, c'2 ) to c.
This time we view the experimental results as a vector
2n
in (R , and let yi and y 2 be, respectively, the orthogonal vectors
2n
in (R . Proceeding as before, we take as our approximation the scalars c\, c 2
This gives
C\
•
= -^11 + *21 + * ' '
+ Xn \
' C2
,
=
X\2 + ^22 + * ' '
+ *n2 '
and
c' = i Xl + • • •
+ x n ). (7-60)
(
This vector may be familiar to some of our readers as the centroid of the vectors
Xi, . .
.
, xn . We note that it may be characterized as the vector in (R
2
which
minimizes the quantity
E llx,-
- c'||
2
(7-61)
be handled in exactly the same way, and it is not difficult to show that Formulas
(7-60) and (7-61) continue to describe the best approximation.
A related problem of this type occurs when one is given a scalar y which is
y = ex,
yi = cxi,
y2 = cx 2 ,
(7-62)
yn = C-*n>
tions will not be compatible (i.e., will not admit a unique solution), and our prob-
x = (*i, . .
.
, xn ) and y = (y u . . . , yn).
In this context our problem can be rephrased as follows: Find a scalar d such that
the vector c'x in the subspace S(x) is as close as possible to y. This, of course,
requires that d be so chosen that the vector c'x — y has the smallest possible
must minimize the quantity \\dx — y||
2
length. Thus d But .
= J2 (dxi - yi)
2
,
and we see that the best approximation to c is the scalar which minimizes the sum
For rather obvious reasons this method of approximation (and its generalization
below) is called the method of least squares.
In the present example it is easy to compute the value of d explicitly. Indeed,
since c'x — y must be perpendicular to x we have (c'x — y) x = 0, and hence •
X
d
X X
or
c =
"-"-
xiyi ^
+ •
• •
•
• ^ Xny-n
+ 7' . (7-64)
v 2
+ •
•
+ x2
Example 1. Use the method of least squares to find the best approximation to c
2 = 3c,
5 = Ac,
6 = 6c.
7-6 I THE METHOD OF LEAST SQUARES 293
C =
2 + 6 + 20 + 36
= 32
1 + 9 + 16 + 36 31
The type of approximation which has just been considered often arises as a
problem in curve fitting. In this context one is asked to find a straight line y = c'x
2
through the origin (in (R ) which best fits the points (x lf y{), (xn yn ). The . . . , ,
accepted method of solution is to let A,- denote the difference c'Xi — yit as shown
in Fig. 7-17, and then choose c' so that the sum of the squares of the A,- is a mini-
mum.* But this sum is just the quantity given in (7-63), and the present problem
is identical with the one solved above.
Finally, we consider the general situation in which a scalar y is an unknown
linear combination of the scalars x x ,
. . . , xm ; i.e.,
y = ci*i + i Cm.Xn
Again we imagine that n experiments have been performed (where n > m), and
that the rth experiment has yielded the values
xn, Xi m and yi, respectively, . .
.
, ,
yi = ci*n + c 2 x 12 + •
~r" Cm,X Inn
yn = C\Xn i + C 2 Xn 2 + '
"T" Vnmi
(7-65)
* It makes excellent sense to use the quantities Af for this purpose rather than the A»
themselves, since in the latter case a large positive difference would obliterate the effect
of several small negative differences, contrary to the requirement that each point be
given equal weight in the fitting process.
294 EUCLIDEAN SPACES | CHAP. 7
formed from the columns of (7-65), and let be the subspace of (R w spanned by W
Xi, x.
m (recall
.
.
that
, m < n). We now make the assumption that our ex-
periments were so designed that Xi, x m are linearly independent, so that they . . . ,
form a basis for V?. The c[, c'm are then chosen so that the vector . . . ,
C\X\ + • • •
+ cm x m
is the perpendicular projection of y onto W. Thus the c t' must minimize the
length of the vector
(cix x + • • •
+ m xm )
c' - y,
2K
1=1
c i**i + • • •
+ c'mXim) - ytf, (7-66)
[(c'iXi + • • •
+ c'mXm) ~ y] • X< = 0,
(x 2 •
xi)ci + (x 2 • x 2 )c 2 + • • •
+ (x 2 • xm )c'm = x 2 •
y,
(X m •
Xi)ci + (X m • X 2 )c 2
/
+ * • •
+ (X» • Xm m y = Xm • V,
been discussed in the preceding section [see (7-53)]. These equations are called
the normal equations for the approximation in question.
15 = ci + 2c 2,
12 = 2c x + c2 ,
10 = ci + c2 ,
= Cy — C2 .
Use the method of least squares to find the best approximation to c x and c2 .
4
(Note that x x and x 2 are linearly independent in (R , as required.) Thus the normal
equations for this approximation are
Example 3. Find the equation of the parabola in the ;cy-plane which passes
through the origin of coordinates and has vertical axis, and which best fits the
points (—1, 3), (1, 1), (2, 5) in the sense of the method of least squares.
The general equation of such a parabola is
y = cix
2
+ c 2 x,
3 = ex — c2 ,
1 = Ci + c2 ,
5 = 4c i + 2c 2 .
8c'i + 6c 2 = 8,
and their solution is c[ = fr» c 2 = — if- Hence the desired parabola is lly =
20a:
2
— 12x, and as can be seen in Fig. 7-18, this curve really does fit the given
points extremely well.
FIGURE 7-18
: : :
EXERCISES
(a) Find the best approximation to the constant which can be made by each student.
(b) Which set of values is the more consistent?
2. Let (xi, .x m) be the vector whose components represent the results of a series of
. .
,
(b) Consider the sequence obtained by combining the two sets of data in (a). Find
the best approximation for this sequence, and compare its variance with the vari-
ances for the two approximations in (a).
4. Prove that the assertion made in the text concerning expression (7-61) is true.
vectors m x,. Find the center of mass of each of the following systems in (R 3 .
t
7-6 | THE METHOD OF LEAST SQUARES 297
9. (a) Show that the center of mass x (Exercise 7) of k particles, each of mass m; located
at the point xi} = 1, 2, ik, minimizes the quantity \\Mx - ]Ci=i m Xi|| 2
. . . , t ,
where M= 2Zi=i m i-
(b) Letxi = (0,0,0),x 2 = (1, -1, 3), x 3 = (1, l,5),x 4 = (2, -8, -10). Find
a set of masses mi, m 2 m 3 m^ such that ||10x , ,
— 52f=i miXi\\
2
is minimized by the
vector x = (1,-1,2).
10. Let mi = 4, xi = (0,0,0); m2 = 3, x2 = (1,1,1); m3 = 2, x3 = (1,1,0);
«i4 = 1. Determine X4 so that ||10x — £* = i miX,||
2
is minimized by x = (1,2,1).
(See Exercise 9(a).)
11 Let T be a linear transformation mapping CR n into itself, and let x be the centroid of k
vectors xi, . . . ,x k in (R n . Show that T(x) is the centroid of the vectors T(xi),
• .
. , T(x k ).
12. Let T be the transformation which maps each vector in a Euclidean space V onto its
best approximation (in the sense of least squares) in a given finite dimensional sub-
space of V. Prove that T is a linear transformation. What are the null space and
image of T?
13. Find the y = ex in (R 2 which best fits each of the following
line sets of points.
Sketch the graph of each of these lines, and plot the given points relative to the
standard coordinate system in (R 2 .
(a) (5, 9), (10, 21), (15, 19) (b) (-4, -11), (1, 3), (5, 16), (10, 29)
(c) (-6, 10), (2, -2), (8, -11), (10, -16)
14. Repeat Exercise 13 using the following data.
(a) (-4, 9), (4, -10), (12, -25) (b) (-5, -16), (10, 31), (15, 31), (20, 40)
(c) (-4,-5), (8,9), (12,16), (20,24)
298 EUCLIDEAN SPACES | CHAP. 7
In Exercises 15-18 suppose that y = cxxx + C2X2, and use the method of least squares to
obtain an approximate formula for y from the given data.
15. xii = 1, X12 = 0, y\ = 2 16. *11 = 1, X12 = 0, yx = l
(a) (1, 4), (2, 5), (-1, -3) (b) (1, 3), (2, 7), (-1, 2), (-2, 8)
(a) (1, -2), (-1, 1), (2, -5), (3, -20) (b) (1, 2), (-1, 1), (2, 7), (-2, -4)
y = ax 4 + bx 2 + ex
2
which, in the sense of least squares, best fits the following points in (R :
real vector space 6(7), where 7 is an arbitrary finite or infinite interval. We re-
frained from doing so, however, because such a proof is both tedious and unin-
teresting. But now that we have the notion of an inner product available, we
shall give a particularly elegant proof of this result which has the added virtue of
serving as an excellent example of the interplay between analysis and the theory
of Euclidean spaces.
Specifically, we must prove that every set of functions of the form
where a and b are real numbers (b > 0) and m is a non-negative integer, is linearly
independent in the real vector space To this end suppose that F
C(— oo, oo).*
where the a k 0k are real numbers and the a k b k are non-negative real numbers.
, ,
with a k b k positive.
,
The essential step in proving that all of the coefficients of F are zero is fur-
nished by the following lemma.
Lemma 7-2. IfTis a trigonometric sum with the property that lim^*, T(x) =
0, then all of the coefficients ofT are zero.
* Note that it is sufficient to consider linear independence in 6(— qo, oo) since by
Theorem 3-2 (the uniqueness theorem) such a set of solutions is linearly independent
in C(— oo oo ) if and only if it is linearly independent in 6(7) for every subinterval 7 of
,
Proof. Let 3 be the real vector space consisting of all trigonometric sums, with
the usual definitions of addition and scalar multiplication, and define an inner
product on 3 by
for any pair of trigonometric sums /and g. It is easy to see that (7-71) does in
fact definean inner product on 3 provided that the limit in question always exists
(see Exercise 1). To prove the existence of this limit we first observe that since
integration is a linear operation (i.e., is performed term by term) it suffices to
consider the case in which / and g are "monomials." In other words, we need
only establish the existence of (7-71) when
Case 1. f(x) = cosaxx, g(x) = cos a 2 x, a x ,a 2 > 0;
Case 1.
sin 2a\t 1
f'g= {
lim t + ^'
.
if
f
ai = a. i* o,
2t 2«i
lim - dt = 1, if ax = a2 = 0.
<->00 *
Case 2.
bx — ai
t
bi + fli
+ &i - «i
f'g= { if fli 9± bu
= 0, if 0i = b\
*, ,_>«, 2t\_ ax _
Case 3.
^^2/L
1 r sin(^
&1
- *,)* _
- 62
»h(»i
*>1
+
+ &2
M J
- o, if ^^
f'g =
lim^f I * - xf-sin26i* if b\ = fc 2 -
Hence the necessary limits exist, and (7-71) defines an inner product on 3. More-
over, from the above computations we see that whenever a and b are positive real
numbers the functions
1, cos ax, sin bx
7-7 |
AN APPLICATION TO LINEAR DIFFERENTIAL EQUATIONS 301
|1||
= 1, ||cosajc|| = —— » ||sinfejc|| = — (7-72)
y2 v2
Now suppose that
lim T(x) = 0.
Making use of the orthogonality just established, together with the relations listed
we have
in (7-72),
rt
T(x) -cosakx . .. 1
«fc
= — n-^
rnr~
2
= 2 lim -
/
/
, .
T(x) cos a k x ax,
, .
k = 0,
.
1, .... n,
\\cosa k x t ^M t Jo
8* = •
z.
—\^~ = 2
2 lim - / T(x) sin b kx dx, k = 1, 2, . . . , n,
"—.Ml
i ii
sin I-*- * ./o
and to prove the lemma we must show that all of these coefficients are zero.
Let us examine the a k By assumption, T(x) —» as jc — oo. Hence, given any
.
real number e > 0, there exists a real number r > such that
Then
lim -
t
/
J
T(x) cos a kx dx + -
t J
/ T(x) cos a fc .x fifo
r
1
< lim-
Jo
) cos a k x dx + lim,
/ T(x) cos afcA ah;
But
while
lim -
t—»QO ^ J
/
r
r(x) cos a k x dx < lim -
t—*ao t
/
Jr
\T{x) cos a^| ^
< lim - / e dx
t->x 7* Jr
_ im l
*- r)
_
302 EUCLIDEAN SPACES | CHAP. 7
Hence |a&| < e for any positive number e, and it follows that ak = 0. A similar
argument shows that 0* = for all k, and the proposition is proved. |
It is now relatively easy to establish our main result, which we state formally as
Theorem 7-7. The set of all functions of the form (7-67) is linearly inde-
pendent in C(— oo, oo ).
Proof Let F be a finite linear combination of these functions such that F(x) = 0.
We shall assume that F has at least one nonzero coefficient, and deduce a
contradiction.
Indeed, if such is the case there exists at least one integer / such that Pi has
F(x) = e
aiX
Pi(x) + e
a ^ P i+1 (x) +
x
• • •
+ ea^P (x), r
and hence
- ai)x
p.(x) = e- aiX F(x) - e
{ai + l
P i+1 (x) - e
{a '- ai)x
Pr (x).
Since F(x) = 0,
aiX
lim e' F{x) = 0,
and since
(o '- Oi)
lim e */>y(j0 =
lim Pi(x) = 0.
X—»00
But Pi has at least one nonzero coefficient. Hence there exists at least one
index j in (7-74) such that the trigonometric sum Ttj has a nonzero coefficient. If
we choose j as large as possible, then
Since it is clear that each of the terms on the right-hand side of this equation tends
to zero as x — oo, we conclude that
lim T{j(x) = 0.
£—00
follows that all of the coefficients of F must be zero, and the theorem is proved. |
EXERCISES
1. Verify that (7-71) defines an inner product on the vector space 3 of all trigonometric
sums. In particular, if
n
show that
2
a
f-f-f
4
+ 5i;<«*
2
+ /ft
*=i
2. Suppose that the roots of the auxiliary polynomial for a constant coefficient linear
differential equation are 2, 2, 2 db 2/, 2 ± 2i, 2 ± 3i, 2 ± 3/, 2 ± 3/, 1 ± /, 1 ± /.
Write out F as in the text, and group the terms in the way described.
Urn e^-^Pjix) =
x—»oo
hm —
,•
X—»x
T(x)
^=
•*
n
In this case x is said to be the limit of {x k } , and we denote the fact that the
lim {Xfc} = x.
k— ><x>
* Although this chapter is logically self-contained, we assume that the reader is familiar
with the notions of sequential convergence and infinite series as studied in elementary
calculus. A review of this material can be found in Appendix I.
304
8-1 I
SEQUENTIAL CONVERGENCE 305
Lemma 8-1. If
lim {x } fc
= x and lim {x }
fc
= y,
fc-»oo k—»
then x = y.
Proof. Let e > be given. Then, by definition, there exists an integer K such
that
= e.
Since e > was arbitrary, this inequality implies that ||x — y|| = 0. Hence
x = y, as asserted. |
1
Example 1. Let {ak} be a sequence of real numbers, viewed as vectors in (R .
Then {ak } converges to the real number (i.e., vector) a if and only if
1
Recalling that the inner product in (R is just ordinary multiplication of real
numbers, we find that
[fob - «)
2 1/2
= ]
= \a k — a\.
lim \a k — a\ = 0.
k—><x
But this equation is none other than the definition of sequential convergence given
1
in elementary calculus, and it therefore follows that sequential convergence in (R
is identical with the usual convergence of sequences of real numbers.
306 CONVERGENCE IN EUCLIDEAN SPACES I CHAP. 8
an arbitrary vector in n
is (R , and if, for each
integer k,
we have
ix* — x||
2
= (a i k
- ai y
+ • ' •
+ (<Xnk ~ OL n y
Hence
lim ||x fc — x|| =
k—>00
FIGURE 8-1
{<Xnk} = {«nlj «n2, • • •}»
= +i
Jfc+1,
{x*}
y=i* + c-ir e2
= {ei + e 2 |ei
,
- e2, &i + e2, . .
.}
does not converge because the sequence {1, — 1, 1, .} formed from the com- . .
{(1/2"- 1 *! + e 2}
The preceding examples show that the study of convergence infinite dimensional
Euclidean spaces is essentially the same as the study of convergence of sequences
of real numbers. However, in infinite dimensional spaces such as Q[a, b] the situa-
tion becomes much more complex, and correspondingly more interesting. For
then the type of convergence defined above is radically different from that studied
8-1 SEQUENTIAL CONVERGENCE 307
momentarily that in a function space with an integral inner product the assertion
that
b
lim |f* - f|| = lim ( f [fk (x) - AxtfdxY 12 =
k—>oo k—>oo \Ja /
is not at all the same as saying that the sequence {ffc} converges to the function f
at every point of [a, b]. In analysis such convergence is known as mean convergence
to emphasize that it is computed by integration, which, in a sense, is a generalized
averaging process.
x4 x3 x2
FIGURE 8-2
lim = lim
k—>oo
||** 0||
k—>Q0 (/><*r
= • ( 2 V' 2
lim
k
= 0.
This not withstanding, {*, * 2 , * 3 , . . .} does not converge to zero at each point in
the interval [—1,1]. In fact, at * = 1 the sequence converges to 1, while at
The example just given shows that mean convergence is different from pointwise
convergence. The one which follows shows how different it really is.
Example 4. Let (?G[a, b] denote the set of all piecewise continuous functions on
[a, b] ; that is, the set of all functions which are continuous everywhere on [a, b]
except (possibly) at a finite number of points where they have jump discontinuities
(see Section 5-1). In Section 9-2 we will prove that (PQ[a, b] can be regarded as
a Euclidean space under the usual definitions of addition, scalar multiplication,
* Recall that a sequence of functions {/*} defined on [a, b] is said to converge pointwise
to the function/ if lim*-** fk (x) = f(x) for each point x in [a, b]. (See Appendix I.)
308 CONVERGENCE IN EUCLIDEAN SPACES CHAP. 8
FIGURE 8-3
b
f-g [ Kx)gOO dx.
Ja
Accepting the truth of this assertion, let [a, b] be the unit interval [0, 1], and let
h = [0, 41 h = ih 11
h = [0, il
-*6 = L2j ?1 h = [|, 1],
FIGURE 8-4
(See Fig. 8-4.) For each integer k > 0, let /*(*) be the function in (PC[0, 1]
which has the value 1 when x is in Ik and elsewhere on [0, 1] (Fig. 8-5), and
consider the sequence {fk } (The function/* is known as the characteristic function
.
of the interval h.) We assert that {fk} converges in the mean in (Pe[0, 1] to the zero
lim ||/fc
- 0|| = 0.
k—»oo
But
1-
FIGURE 8-5
and since fc is identically 1 on Ik and is zero elsewhere, the value of this integral is
just the square root of the length of Ik . However, the length of Ik tends to zero
with increasing k, and it follows that
2
lim (f*[fk(x)] dxy* = o,
k—»oo
as asserted.
Now let x be a fixed point in [0, 1], and consider the sequence of real numbers
ifk(xo)}. We contend that this sequence does not converge. For, by definition,
fk(x ) is either 1 or depending on whether x is or is not in Ik But the Ik were .
constructed in such a way that there exist arbitrarily large values of k for which
fk(x ) = and arbitrarily large values of k for which fk (x ) = 1. Thus {fk (x )}
contains both zeros and ones no matter how far out in the sequence we go, and
hence does not converge.
EXERCISES
1
Determine which of the following sequences {a k} converge in (R , and find the limit of
each convergent sequence.
(-1)"
1. ak = 2. ak =
2k + 1
3. ak = 1 + (-1)* 4. ak =
k + 1
A: + 1
5. ak = 1 + e 6. ak
(k + 1)(* - 1)
2
fc
+1 + (_i)
fe
In A;
7. ak = i. ak =
2 fc
310 CONVERGENCE IN EUCLIDEAN SPACES | CHAP. 8
Determine which of the following sequences {x*} converge in (R 3 , and find the limit of
each convergent sequence. (In each case the vectors ei, e2, e3 are an orthonormal basis
for (R 3 .)
9. x fc =
(*>
+
[ 1 t_ J
ei + e2 + , ,
2
e3 10. x* = — ei + k
2
e2 + (-l)*e 3
11. x fc
= (1 - 2~ fc
k In A: , 1
j^TT ei + T- +
,
12 - x* = e2 e3
*TnT
13. Let {xfc} be a convergent sequence of vectors in a Euclidean space. Prove that for
every real number e > there exists an integer such that K
||x m - x„|| < e
forany function / in 6[— x, x]. Use this fact to prove that the sequence {sin kx},
k = 1,2,..., does not converge in the mean in C[— x, x].
15. Prove that the sequence {(sin kx)/k} converges in the mean in 6[— x, x] to the zero
function. Prove that this sequence also converges to zero for each x in the interval
[-x,x].
*16. (For students familiar with uniform convergence.) Let {fk } be a sequence of func-
tions in C[a, b], and suppose that {fk} converges uniformly on [a, b] to a function fin
Q[a, b]; i.e., given any e > there exists an integer k such that \fk(x) — f(x)\ < e
for all k > K and all x in [a, b]. Prove that {fk} also converges in the mean to /.
given by
lim {x fc
•
yjfc}
= x • y.
k—yoo
lim |(x fc
•
y fc ) - (x •
y)| = 0.
k—>oo
= |(u* •
Vfc) + (x • vjt) + (y • ujb)|
< \n k - Vfc| -F |x • v fc | + |y • u h \,
lim|(x .y fc fc
)- (x.y)| = 0.1
A;—+oo
{Xi, Xi + X2 , Xi + x2 + x3 , ...},
and then take the convergence of this sequence as the criterion for the convergence
of the series in question. This is the content of
x = 2x
k=l
*> (8
~3
)
and say that x has been expanded as an infinite series. In greater detail,
XX=i Xfc converges to x if and only iffor each real number e > 0, there exists
an integer K such that
£x fc < e (8-4)
&=i
whenever n > K.
n
Example. Let e l5 e n be an orthonormal basis for
. .
.
, (R , and suppose Y,t=i *fc
71
is an infinite series of vectors in (R Then .
00 00 00
^2 x k = X/
/b=l fc=l
aifcei + ' ' *
+ Sa
A;=l
nk^k,
where the o^ are the coordinates of X& with respect to the given basis, and it
follows from Example 2 of the preceding section that £fc=i x*; converges to the
vector x = <*iei + + anen if and only if
• • •
00 CO
fc=l fc=l
In particular, this result implies that the usual tests for the convergence of a series
of real numbers such as the comparison test, the ratio test, etc., can be applied
n
to determine whether or not an infinite series in (R converges. The only difference
is that the tests must be applied n times, component by component.
EXERCISES
(b) Let {a k } and {/3 fc } be convergent sequences of real numbers with limits a and /?,
respectively, and let {x } and {y k } be as in Lemma 8-2. Prove that {a kx k
fc fikYk} +
converges to ax + /3y.
lim [(x k - x) •
y] =
k—>oo
(b) Let V
be a finite dimensional Euclidean space, and let {x k} be a sequence of
vectors in V. Suppose that there exists a vector x in V such that
lim [(x* - x) •
y] =
k—->oo
and that every vector in such a space can be written uniquely in the form
x = (x •
eOei + • • •
+ (x • e n)e n . (8-5)
But the Gram-Schmidt process can be applied equally well in an infinite dimen-
sional Euclidean space where it can be used to produce an infinite orthonormal set
ei, e 2 , .This fact suggests that an extended version of our earlier results may
. . •
be within our grasp, and the remaining sections of the present chapter are devoted
to realizing this suggestion.
Before we begin, we emphasize that the ideas we are about to develop are
natural and inevitable consequences of the study of finite dimensional Euclidean
spaces, and should be understood as such. This, however, is not to deny their
importance or to suggest that they are being presented as an academic exercise
in generalization. Quite the contrary; for, as we shall see, these ideas lie at the
heart of much of and its applications to mathematical physics.
analysis
The most obvious way in which to attempt to generalize (8-5) in the presence
of an orthonormal set ei, e 2, ... in an infinite dimensional space V is to replace
its right-hand side by the infinite series
^2 (x • e fc >fc. (8-6)
fc=i
* Such a series is sometimes called a formal series, to emphasize the fact that on the
face of things it is nothing more than an expression which actually may be devoid of
meaning.
314 CONVERGENCE IN EUCLIDEAN SPACES I CHAP. 8
(8-6) is deduced from x, albeit in a purely formal way. The one most; commonly
used is
x ~ ]T (x • e k )e k , (8-7)
fc=i
the symbol ~ (which, by the way, is no relative of the one used earlier for equiva-
lence relations) being used to emphasize that the series in question may not con-
verge to x. Of course, if it does, we write
x = ^ (x • e A )e fc , (8-8)
and say that the series converges in the mean to x. In either case, the inner products
x e*; are called the coordinates or (generalized) Fourier coefficients of x with
•
It is clear that the Fourier coefficients of x depend upon the orthonormal set
with respect to which they are computed. Not quite so clear, but equally true, is
that (8-6) may converge in the mean to x for one orthonormal set but not for
another. The following examples illustrate both of these points.
f(x) = x, —it<x<t,
in C[— t, 7r] with respect to the orthonormal set
x-e =fc
sin kx dx, k = 1,2,
Vf J —*
and, using integration by parts, we obtain
—= / x sin kx dx = — z
x cos kx
cos kx dx
sin
I sii
( x ^
,
1 = .
. . l .
In the next chapter we shall see that this series actually converges in the mean to
x, so that we are justified in writing
* = 2f;(-i)*-'^
fc=l
1 cos jc cos 2x
> 5 J • • • 5
(8-10)
V2x Vx s/r
But all of these integrals are zero (see Exercise 1), and hence (8-7) becomes
x ~ 0.
(x-ei)e! + •• •
+ (x-e n )e n ,
if w is any vector in V n ,
This simple observation furnishes the key to the proof of the following theorem.
316 CONVERGENCE IN EUCLIDEAN SPACES CHAP. 8
Theorem 8-2. Let J2k=i QkPk be any infinite series which converges in the
mean to x; i.e.,
y^ y
ak e k . (8-12)
fc=i
This theorem asserts that whenever x can be expressed as an infinite series the
coefficients of the series are uniquely determined and must be the Fourier coeffi-
cients of x with respect to the orthonormal set e l5 e 2 , . . . . This, of course, is
the infinite dimensional analog of the uniqueness of (8-5) in the finite dimensional
case.
Since
Proof.
x = ^
k=l
ak e k ,
we have
lim
n—>oo
x — ^
fc=i
ake k = 0.
x — ^
fc=i
(x • e k )e k < X — ^T, Qk * k
fc=l
lim
n—»oo
x - 2
fc=l
( x * efc ) e *
x = 2 ( x * e *)efc -
00
and we must now show that this equation implies ak — (x • ek ) = for all k.
^ =
lim
n—>oo
2
fc=i
fa* - ( x * e k)]e k - = lim X)
fe=i
[
fl
* - <x *
e
2
°-
The desired conclusion now follows from the fact that each of the terms
[a k — (x • e k )]
2
in this sum is non-negative. |
8-3 | BASES IN INFINITE DIMENSIONAL EUCLIDEAN SPACES 317
In view of this result seems reasonable to extend the meaning of the term
it
x = 2 «**• (8
~ 14
)
The reader should note that the vectors in a basis for 13 are only required to be
orthogonal, not orthonormal. Of course, any basis for 13 can be normalized in
the standard fashion, but would be a hinderance to restrict the definition in this
it
This would be somewhat closer to the spirit of the definition in Chapter 1, but
since any such set can be orthogonalized by the Gram-Schmidt process, it would
not result in any significant gain in generality.*
For all we know at the moment, we may never encounter an infinite dimensional
Euclidean space with a basis. This notwithstanding, we shall assemble a number
of results about such spaces pending the time when we finally meet one. However,
in the absence of an immediate example it is only fair to state that these results
are more than speculations in a vacuum since all of the Euclidean spaces which
appear in analysis do, in fact, have bases.
We begin by making a number of simple observations, the first of which we
state formally as follows:
Lemma 8-3. The zero vector is the only vector which is orthogonal to
every vector in a basis for a Euclidean space 13.
sions relative to a basis,and the fact that 2jk = i ai&k converges in the mean to
zero whenever ak = for all k. Despite its simplicity, Lemma 8-3 is often useful,
particularly in showing that an orthogonal set is not a basis for a given Euclidean
*A word of warning. Definition 8-3 is not the only meaning assigned to the term
"basis" for infinite dimensional spaces. There is another in common use which is not
restricted to Euclidean spaces, but applies to arbitrary vector spaces as well. We mention
this point only to alert the student to the fact that when he encounters this term he must
check the definition to avoid serious errors.
318 CONVERGENCE IN EUCLIDEAN SPACES | CHAP. 8
space. For instance, when this result is combined with Example 2 above, it allows
us to assert that the set of functions in (8-10) is not a basis for C[— t, t].
oo
x = ^
fc=i
(x • e k )e k .
For then, by Theorem 8-2, this expression must be unique, and Definition 8-3 is
therefore satisfied.
And while on the subject of orthonormal bases, there is another point it would
do well to settle. Recall that if ei, . .
.
, en is such a basis in a finite dimensional
space, and if
n n
x = J]
fc=l
(x • e k )e k and y = ^
k=l
(y • e k )e k
That is, the inner product of x and y is the sum of the products of their cor-
oe oo
x = J] (x • e k )e k and y = ]T (y • e k )e k
k=l k=l
n n
xn = S
fc=i
(x • e fc )e fc and yn = ^
fc=i
(y •
ejfc> fc .
{x n .y n }^x.y. (8-15)
n
x n yn • = 2Z ( x * e *Xy '
e *)'
x y • = X) (x • e*)(y • e fc ). (8-16)
Jk=l
— »
EXERCISES
1. Verify that
—— / x dx = and —— x cos kx dx = / 0, k = 1, 2,
%/2ir J
-x Vx J -*
2. Compute the Fourier coefficients of the function
— 1 cos x cos 2x
• • •
> > j
sin x sin 2x
^
fc=i
aifih = 0.
furnish the justification for most of our work with orthogonal sequences in func-
tion spaces.
oo
2
A;=l
(x-e*)
2
< ||x||
2
(BesseTs inequality). (8-17)
00
2 2
^2 (x-e*) = ||x|| (ParsevaVs equality). (8-18)
k=l
320 CONVERGENCE IN EUCLIDEAN SPACES | CHAP. 8
x - J] (x • e k )e k
k=i
x - ^2
k=l
(x • e k )e h = (
x - 2
k=l
( x '
e *)e * I' (
x — 2
fc=l
( x ' ek ^k
= x x • - 2 ^j (x • e A )(x • e fc ) + ( ^ (x • e y )ey ) •
( X) ( x ' e *) e * )
(We have changed the index of summation on the first factor of the last term for
convenience in computation.) But since the e^ are orthonormal, ey • e& = 8jk,
n \ / n \ n n
(
J] (x • e^e, )
•
( X) (x * ek ^k ) = 22 ( x ' e >')( x * e *Xe ' * efc )
= Z(x
k=l
-
e *)
2
Thus
x - ^2 (x • e^ = iixf-x;(x -^ (8-19)
fc=i fc=i
< x - J]
fc=i
(x • e fc )e fc = N 2
- Z>- e ^
fc=i
2
X)(x.e fc ) < ||xf
2
for all h, and the partial sums of the series £*=i (x • e fc ) form a bounded non-
decreasing sequence of non-negative real numbers. By a well-known theorem
from calculus (see Appendix I) we conclude that this series converges and that
To finish the proof we now suppose that ei, e 2 , ... is an orthonormal basis
for V. Then by Theorem 8-2 we know that
x = ^ (x • e >, fc
Hence
n
lim
n—»oo
x — ^
A;=l
(x • e fc )efc = o,
-X;(*-<rf =
2
lim |x|| 0.
n—yao
fc=l
Thus
k=i
which is Parseval's equality. Finally, since the steps in this chain of reasoning are
reversible, we conclude that e ls e 2 , . . . is a basis whenever Parseval's equality is
is equal to the sum of the squares of the lengths of its components relative to an
for every xinV. That is, the Fourier coefficients ofx tend to zero as k—» oo
for any orthonormal set in V.
2
Indeed, Bessel's inequality implies that £t-i (x • e*) is a convergent series of
real numbers. Hence the individual terms in this series approach zero with in-
~h**'*> ~h w 2
(8_21)
where the
x-e fc
2
IMI
EXERCISES
1. Let ei, e 2 , . . .be an orthogonal set in a Euclidean space V. Prove that this set is a
basis for V if and only if its associated orthonormal set is' a basis.
2. Prove Formula (8-21).
3. Assume that the functions 1, sin x, cos x, . . , sin kx, cos kx, . . . are a basis for the
Euclidean space Q[— x, x]. (They are.) Find the Fourier coefficients of fin Q[— x, x]
in terms of this basis. What is the series expansion of/?
(In those cases where the context is clear we shall omit the reference to V, and
simply say that "W is a closed subspace.)
As with every new definition, one's first thought is to produce examples. In
this case we can furnish a plentiful supply by examining finite dimensional spaces,
for there we can prove
d, . .
.
, e m are a basis for W.* It follows that an arbitrary vector x in V belongs
to *W if and only if its last n — m components with respect to this basis are zero.
Now suppose that {x k } isa sequence of vectors in V?, and that {xk } — x. »
Then, by Example 2 of Section 8-1, we know that the sequences of real numbers
formed from the components of the x k converge to the components of x. It
follows that the last n — m components of x are zero, and hence x belongs to W,
as asserted. |
arbitrary Euclidean space is closed. The proof is almost identical with the one
just given, and has been left as an exercise.
This lemma tells us that the only interesting illustrations of Definition 8-4 are
to be found in infinite dimensional spaces, if any exist at all. They do, as the
following example shows.
Example. Let l 2 (read "little I two") denote the set of all infinite sequences
of real numbers
x = C*l> *2j *3> • • •)
^2 Xk < 00.
Then t2 becomes a Euclidean space if, with x as above, a a real number, and
y = Oi, J>2, ys, • • •), we define
ax = (ax i, aX 2 ax 3 , , . . .)
x + y = (*i + yu x 2 + y2 x 3 + y3
, , • •)
x y • = xiyi + x 2y 2 + x 3j> 3 + • •
•
* This assertion can be proved at once by applying the Gram-Schmidt process to the
basis described in Theorem 1-7.
324 CONVERGENCE IN EUCLIDEAN SPACES | CHAP. 8
(see Exercise 2). Moreover, since l 2 contains the infinite orthornormal sequence
et = (1,0,0,...),
e2 = (0,1,0,...),
e3 = (0,0,1,...),
it is infinite dimensional.
Now W
be the set of all vectors in l 2 which have only a finite number of
let
xi= (£.0,0,...),
x2 = \2i ?>">•• •)>
x3 = (25 4) 8> • -h
fc
each of whose vectors belongs to VP, converges to the vector (J, 5, i, . . . , l/2 , . .
.)
which is not in W.
(The Euclidean space l 2 is an example of what mathematicians call a Hilbert
space.)
Having shown that there exist subspaces which are not closed, and hence that
there is good reason for studying closed subspaces as entities in themselves, we
begin with the usual elementary observations. First, every Euclidean space V,
viewed as a subspace of itself, is closed, and hence every subspace of V (closed
or not) is contained in at least one closed subspace of V. Next, the intersection
of any collection of closed subspaces of V is itself a closed subspace. (Recall that
this intersection consists of those vectors which belong to every one of the sub-
spaces in question. Again the proof is deferred to the exercises.) Conclusion: If
W is an arbitrary subspace of a Euclidean space V, there exists a smallest closed
subspace of V containing V? namely the intersection of all the closed subspaces
;
of V which contain VP. This subspace is denoted VP, and, for obvious reasons,
is called the closure ofVP in V. Finally, VP can be described as follows.
The sequence of observations leading to Lemma 8-5 admits a useful and natural
generalization. Suppose that instead of starting with a subspace of V we begin
with an arbitrary nonempty subset 9C in V. The n we can form the subspace S(9C)
spanned by 9C, and follow this with the closure S(9C) of S(9C) in V, in which case
we say that I(9C) is the closed subspace of V generated by 9C. Note that the sub-
shows that when 9C is an infinite subset of an infinite dimensional space, S(9C) will
in general be different from S(9C). In Chapter 1 we gave a description of the vectors
belonging to S(9C) in terms of the vectors belonging to 9C (see Theorem 1-1) When .
combined with Lemma 8-5, that description allows us to assert that S(9C) consists
of all vectors in V which are limits of linear combinations of vectors in 9C. Stated
more formally, we have
V?ofa Euclidean space V. Then, ife is any positive real number, and x any
vector in W, there exists a finite linear combination Ya=\ «fc x of vectors x
fc fc
Finally, suppose that 9C is a set of generators for V itself, and that the vectors in 9C
are also linearly independent.* On the strength of our experience with finite dimen-
sional spaces the reader could hardly be blamed for assuming that every vector
x in V could then be written in the form
00
X = ^ a kX k>
where the a k are scalars, and the x k belong to 9C. Surprisingly, this is not the case,
and an example of this apparently paradoxical situation can be found in Exercise
13 below. However, things improve remarkably when 9C is also an orthonormal (or
orthogonal) set, for then we can prove
Proof. We must show that every vector x in V can be written in the form
00
x = J] ( x • e *)e fc-
* It can be shown that such a set always exists in any Euclidean space V. The proof,
however, is not easy.
326 CONVERGENCE IN EUCLIDEAN SPACES | CHAP. 8
To do so, let € > be given. Then, since the e^ generate V we can find a linear
combination X)*Li ak ek such that
x — ^2 ak e k < €.
N
x - J]
k=i
(x * ^k < x - ^ (x • e k )ek
N
< < e.
*=1
x = ^
fc=i
(x • e k )e k ,
This result explains why orthonormal sets enjoy a special status in the theory of
infinite dimensional Euclidean spaces beyond what they have in finite dimensional
spaces. In the latter their use is simply a matter of convenience, in the former it
EXERCISES
1. Prove that every finite dimensional subspace "Wofa Euclidean space V is closed.
[Hint: If x is the limit of a sequence of vectors in W, apply Lemma 8-4 to the sub-
space of V spanned by W and x.]
8-5 | CLOSED SUBSPACES 327
v / I
and then use the fact that x and y belong to 1 2 to obtain an upper bound, independent
of n, for the right-hand side of this inequality.]
(b) Prove that x •
y is defined. [Hint: For any two real numbers x, y, we have
xy = %[(x + y)
2 — x2 — y
2
]. Hence
n / n n n \
^2 wi*
fc=i
= i ( 2
\a;=i
( Xfc + y^ ~
2
2
«:=i
x* ~ ]C ^*
fc=i
/
4. In the example above it was asserted that a certain sequence in 1 2 converged to the
vector (^, 5, i, . . .). Prove this assertion.
in the definition of a closed subspace reference must be made to the space in which
the closure is taking place.)
7. Prove that = Wfor any subspace W Euclidean space 13. (This result Wofa is
sometimes expressed by saying that "the closure of the closure is the closure.")
8. Prove Lemma 8-5.
9. Prove thatW is a closed subspace of 13 if and only if W = W. [Hint: See Exercise 7.]
verges to x, and let y be any vector in W. To prove that x • y = 0, use the identity
x y = y • • (x — x k)
13. Let 1 2 be the Euclidean space defined above, and for each integer k > 2 let x*,
where the ones occur in the first and A:th positions. Let 9C denote the set consisting
of the vectors X2, X3, . .
(c) Show that the vector ei = (1, 0, 0, . . .) belongs to S(9C). [Hint: Compute the
value of
ei J^ x,
*=2
and show that
°°
1
lim ei Y) xjt = 0.
n—»oo Jfc=2
(d) Prove that S(9C) = I2. [Hint: Begin by showing that each of the vectors ei,
e2, . . • as defined in the text belong to S(9C), and then show that these vectors are
a basis for ^2-]
(e) Show that ei cannot be written in the form £ST=2 a kx k thereby furnishing the ,
converges. Prove that X)"=i a *>/k also converges. Give an example to show that
the converse of this statement is false.
9
fourier series
9-1 INTRODUCTION
Although we now know a great deal about the general theory of orthogonal
series in infinite dimensional Euclidean spaces we have yet to produce a single
concrete example of such a series. To fill this gap in our knowledge we propose
to devote the next three chapters to a detailed study of several types of series
expansions in infinite dimensional function spaces. In each case we shall begin,
as of course we must, by exhibiting a basis for the particular space in which the
series is tobe constructed. Once this has been done the series in question will
appear as a special version of our earlier results, and for this reason the following
discussion can be viewed as a collection of elaborate examples in illustration of
material that is already known.
At the same time, however, it is only fair to warn the reader that each of the
series which we shall construct is of significant importance in physics and applied
mathematics. Indeed, since 1822 when Jean Baptiste Fourier first solved the
problem of heat flow in solid bodies by means of those series which now bear his
name, this subject has grown until, at present, it is an entire branch of mathe-
matics and mathematical physics. Later in this book we shall examine some of
the applications of this theory, but first we discuss the series themselves.
Given the importance of the various series which we are about to consider, it is
and
(ii) the limits
fix$) = lim f(x + h),
exist at each point x in [a, b]. (Note that only one of these limits is relevant
More generally, the requirement that both of these limits be finite everywhere in
[a, b] implies that the only discontinuities of/ are "jump discontinuities" of the
f(xt) ~ flxo)
x, < x < 1,
/(*) =
1 - x, 1 < x < 2,
(see Fig. 9-2) is piecewise continuous on the interval [0, 2] and has a jump dis-
FIGURE 9-4
rb
/ f{x)dx
Ja
exists, and is independent of whatever values (if any) / assumes at its points of
discontinuity. In particular, if /and g are identical everywhere in [a, b] save at
their points of discontinuity, then
rb rb
/ f(x)dx = / g(x)dx.
Ja Ja
2. If / and g are piecewise continuous on [a, b], then so is their product fg.
(See Exercise 11.) This, together with (1) implies that the integral of the product
of two piecewise continuous functions always exists.
This said, we now turn our attention to the problem of converting the set of
piecewise continuous functions on [a, b] into a Euclidean space. In view of the
fact that this set includes the continuous functions on [a, b] it is only reasonable
332 FOURIER SERIES | CHAP. 9
f'g= [ f(xMx)dx.
Ja
(9-2)
But does (9-2) actually yield an inner product on the set of piecewise continuous
functions on [a, b]l The unpleasant answer is, No. To see what goes wrong, let
n(x) be a function which is zero everywhere in [a, b] except at a finite number of
points (see Fig. 9-5). Such a function
is said to be a null function, and has the .
" *
annoying property that
look the fact that a null function is not identically zero, and treat it as if it were.
But then, to be consistent, we must also regard any two piecewise continuous func-
tions as identical whenever they a finite number ofpoints. And this is
differ at only
just what needs to be done to make (9-2) yield an inner product.
As usual, there are a number of facts which have to be verified before such an
assertion can be accepted. Among them we call attention to the need to ascertain
that (9-2) respects our identification of functions. In other words, we must show
that whenever /i is identified with/, and g x with g, then
[ f\(x)gi(x)dx
= f
f(x)g(x)dx.
Ja Ja
For only then will (9-2) unambiguously define an inner product on the set of
piecewise continuous functions identified as above. This, however, is easy to
prove, and has been left to the reader as an exercise (see Exercise 5).
all real numbers a. Thus functional addition and multiplication by real numbers
also respect our identification of functions, and it follows that the set of piecewise
continuous functions on a fixed interval [a, b], identified as above, is a Euclidean
space. We shall denote this space by (PQ[a, b], and assert without proof that it
Now
that we have rigorously constructed the space (?e[a, b] and insisted that
itsvectors are collections of functions identified with one another, it is perfectly
clear that we can ignore this fact and treat these vectors as though they were
9-2 I
THE SPACE OF PIECEWISE CONTINUOUS FUNCTIONS 333
ordinary functions. This is precisely what we shall do henceforth, but always with
the tacit understanding that the facts of the matter are as outlined above, and that
all of our arguments can be rigorously restated, if necessary. As with all such
abuses of terminology, this involves no real danger of error, and has the positive
result of simplifying language and notation.
EXERCISES
1. For each of the following functions evaluate the right- and left-hand limits at all
points of discontinuity, and state whether or not the function is piecewise continuous
on [0, 2].
^ (o, x =
(a) /(*)=*>
i
0<x<l
n *
ib) fix) ={ l
\x - 2, l < x < 2 I- > < x < 2
_ 2 i i
{2 < < ^ 1
= 1 " *' ^ x < J
x -l '
(d) fix) f
°
1, 1 < x < 2
U - 1, 1 < x < 2
< x < i i x
(1, < x < 1
0, } < x < f (f) fix) = x - 1
2. (a) Suppose a piecewise continuous function /is identified with a continuous func-
tion g in (PQ[a, b]. What can be said about the magnitudes of the jump discon-
tinuities of/?
(b) Use the result in (a) to prove that Q[a, b] is a subspace of (PQ[a, b].
3. Determine whether or not the following functions are piecewise continuous on [0,1].
0, if x is rational
4. Prove that/i and/2 are identified in (?Q[a, b] if and only if/i — /2 is a null function.
5. Let /and g be piecewise continuous on [a, b], and suppose that/i is identified with
/,and g\ with g. Prove that
,b ,b
/ fi(x)gi(x) dx = / fix)gix)dx.
Ja Ja
1, cos —a
TTX
'
.
sin —a
TX
' cos
nirx
a
»
.
sin
tvkx
a
j
11. Let /and g be piecewise continuous on the interval [a, b], and let fg denote their
product; that is, fg is the function defined by
fg(.x) = f(x)g(x)
for all x in [a, b\. Prove that fg is piecewise continuous on {a, b], and thus deduce
that the integral in (9-2) exists. [Hint: Use the fact that the product of continuous
functions is continuous.]
about the origin (Fig. 9-7). Thus, for integral values of is even if n is even, n, xn
odd if n is odd, a fact which helps explain the particular terminology used in
this context.
The importance of even and odd functions for our work stems from the equalities
whenever /is odd and integrable. Both of these assertions are easy consequences
of the above definition (see Exercise 2), and are also evident from the geometric
interpretation of the definite integral as area.
An equally elementary observation is that the product of two functions is even
whenever both of the functions are even or both are odd, and is odd whenever
one of the functions is even and one odd. In short, the multiplication of even
and odd functions obeys the rules
f—a f(x)g(x)dx =
J
whenever / and g have the opposite parity. Hence even and odd functions in
if/ is even. The value of these results for computing Fourier coefficients is too
obvious to need comment.
336 FOURIER SERIES | CHAP. 9
Lemma 9-1 .
Every function on the interval [—a, a] can be written in exactly
one way as the sum of an even function and an odd function.
f / . _ f(x) + f(-x) _
— f(x) - /(-*)
jekx) 2 jo{x) 2 yy~i)
fE — gE = go — fo-
But the difference of two even functions is even, and the difference of two odd
functions is odd. Thus the function defined by the above identity is simultaneously
even and odd, and so must be the zero function. In other words, fs — gE =
go — fo = 0, and it follows that gE = fE, go = fo, as desired. |
The functions fs and fo defined in (9-7) are known respectively as the even
and odd parts off
Example 2. If f(x) = e
x
, then
x ,
-x x _ -x
e e e e
fs(x) = = and f (x) = =
Thus the even and odd parts of the exponential function are the hyperbolic cosine
and hyperbolic sine, respectively.
EXERCISES
1. Classify each of the following functions as even, odd, or neither even nor odd.
.2
-1 -1
(e) 7 r^r? (0 sin x ve/ cos
(g) x (h)
(x + l)(;c - 7^
1) (x -f- l)(x - 1)
(i)
X(
^^ (J) /(W). /defined in [0, a]
3. Establish the multiplicative rules for even and odd functions given in the text.
9-4 | FOURIER SERIES 337
4. Show that
for any piecewise continuous function /on the interval [— it, t].
5. Let /be differentiable in the interval [-a, a]. Prove that/' is odd whenever /is even,
and even whenever /is odd.
6. Let /be an integrable function defined in the interval [—a, a], and let
7. Decompose each of the following functions into its even and odd parts.
2
(a)
X
x — ,
1
(b) ——
x + 1
(c) x cos x - cos 2x (d)
x
_i_
— r
i
1
f(x ) = ^2 a * x *V a * constants.
10. Let 8[— a, a] denote the set of all even piecewise continuous functions on the interval
[—a, a], and let ©[— a, a] denote the set of all odd piecewise continuous functions on
[—a, a]. Prove that 8[— a, a] and G[— a, a] are subspaces of (?Q[— a, a].
11. Prove that the zero function is the only function which is simultaneously even and
odd on [—a, a].
12. Let poix), Piix), ... be the sequence of polynomials obtained by applying the Gram-
Schmidt orthogonalization process to l,x, x 2 ...in [—1,1] (see Example 3, ,
Section 7-4). Show that/?2fc(*) is an even function, and thatp2k+iix) is an odd func-
tion for all values of k. [Hint: Use mathematical induction.]
13. Let /be piecewise continuous on [—a, a], and suppose that /is orthogonal to every
even function in <P6[— a, a]. Prove that /is odd. [Hint: Use Lemma 9-1.]
We have already seen that these functions are mutually orthogonal in (P(B[— r, t],
and we shall prove shortly that they are a basis as well. Granting the truth of this
fact, we can then use Formula (8-22) to write any piecewise continuous function
m = 4il2 +
Hi
±
»_,
/•cos kx
||cosA:a:|| 2
cos kx
,
+
'
f'sinkx
-Tp-: ——=
||sin/:jc||
n-^
:
sin
. ,
kx (mean) (9-9)
where the notation "(mean)" indicates that the series in question converges in
the mean to /. But since
ii if = r dx = iw,
J TT
2
||cosfcx||
2
= I cos kx dx = %,
J X
2 2
||sin &a:|| = f sin kx dx = tt,
J — jr
where
for all k. This particular representation of/is known as its Fourier series expansion
on the interval {
— *, tt], and the a k and b k are called the Fourier or Euler-Fourier
coefficients off.
Once again we emphasize that (9-10) must be read as asserting that the series
in question converges in the mean to/ not that it converges pointwise in the sense
that
a
f(x ) = ?? +
T J] (flfc cos fcxo + 6fc sin kx )
fc=l
for all jc in [— ir, tt]. Indeed, since the value of/ at x can be changed arbitrarily
without changing the value of Fourier coefficients, this would be entirely too
its
bk = - / sin kx dx
x Jo
2
= 7— (1 — COS kir)
kx
\±> k = 1,3,5,...,
sin 3x sin 5x
/(*) = sin a;
,
^ ^
,
1 ?
,
fc=i
of its Fourier series, which, as can be seen, already furnishes a fairly good ap-
proximation to /throughout the interval [— r, *]. This approximation improves
considerably if we use the first four terms in the series (Fig. 9-10), and it is not
so doing one quickly becomes convinced that the series actually converges point-
wise to / every where in [— *-, t] where /is defined. Moreover, when x = and
4 r . , sin 3x ,
sin 5x . sin 7x1
±ir, the series obviously converges to zero even though /is not defined at those
points. Here then, in our very first example, we have come upon a Fourier series
which converges pointwise in the interval [— *-, ir], and which represents the func-
tion from which it was derived at each point in the domain of that function.
It would be easy to multiply the number of such examples indefinitely until the
reader became convinced (erroneously, as it turns out) that all Fourier series con-
[— ir, *]. Instead, however, we prefer to cite a
verge at each point in the interval
theorem which will account for this phenomenon within the framework of the
general theory of Fourier series. For the present we content ourselves with a
statement of the result in question, leaving any attempts at a proof to the next
chapter.
which we mean thatf has a piecewise continuous first derivative on [—ir, ir].
f(xt) + f(xp)
/(-*+) + /(O _
(9 14)
at ±ir.
This theorem is one of the most important in the entire theory of Fourier series,
and the student should make certain that he understands it thoroughly before
going on. In particular, he should note that the expression
Axf) + f(xo)
is none other than the average of the right- and left-hand limits of /at x , and is
equal to f(x ) whenever x is a point of continuity of/ Hence the Fourier series
expansion of a piecewise smooth function f in (PQ[— ir, ir] converges to f(x ) when-
ever x is a point of continuity off On the other hand, if/has a jump discontinuity
at x , (9-13) implies that the Fourier series for / converges at x to the value
located at the midpoint of the jump, as shown in Fig. 9-11.
FIGURE 9-11
9-4 FOURIER SERIES 341
When these results are applied to Example 1 above they allow us to assert that
the series
4T . . sin 3x sin 5x ] / Q i c\
sin x H ^ 1
^ h • •
•
(9-15)
if X = — 7T, 0, 7T,
1 if < x < x.
Thus, for example, when x = ir/2, the value of this series is 1, and we conclude
that
= ^
1
-a 3^5 7
or
= ^
1
3^5 7
V2 + V2 V2 V2
x 2 2-3 2-5 2-7 +
and we have a second representation of t/4 as
V? 1+1-1-1+
2 ( ^3 5 7 ^ )
Before continuing, it may be appropriate to remark that there exist continuous
functions whose Fourier series diverge at finitely many points in [— t, it]. Thus
the requirement that / have a piecewise continuous first derivative is imposed in
order to guarantee the pointwise convergence of its Fourier series. Incidentally,
the problem of determining whether there exists a continuous function whose
Fourier series diverges everywhere in [— ir, x] is still unsolved. Needless to say,
such a function, if it exists, will be rather bizarre, and is not likely to arise in
applications.
By now the reader has undoubtedly observed that our description of the point-
wise behavior of a Fourier series is incomplete. For if a trigonometric series
00
FIGURE 9-12
consequence of the fact that the functions sin kx and cos kx are periodic, with 2w
as a period, and allows us to make the following important observation: If (9-16)
converges pointwise to a function f everywhere in [— ?r, x), then it actually converges
on the entire x-axis to the function F obtained by repeating f successively along the
x-axis in intervals of length 2x (see Fig. 9-12). It is obvious that the function F
obtained in this way is periodic with 2r as a period, in the sense that
F(xt) + F(xo)
In particular, we note that the Fourier series for / will converge pointwise to a
continuous function on the entire x-axis if and only if /is continuous on [— tt, it]
and /(— tt) = /(tt). For only then will F be free of jump discontinuities. We
shall have occasion to use this fact later in the chapter.
At the risk of belaboring the obvious, we point out that Theorem 9-2 can be
used to sketch the graph of the Fourier series of any piecewise smooth function/
in (?e[— t, x]. The procedure is as follows: First sketch the graph of the periodic
extension F off; then plot the midpoint of each jump discontinuity of F. The
resulting picture, with these isolated points included, will be the graph of the series
9-4 FOURIER SERIES 343
FIGURE 9-13
4T sin 3x
+
•
+
.
sin x
f(x) = I*!,
— < 7T x < IT,
sin kx dx
a = - / x dx = 7r,
* Jo
344 FOURIER SERIES CHAP. 9
FIGURE 9-14
cos 3x cos 5x
x = (cos a; H To I « T (mean)
32 52
in (?e[—T, *]. The graph of this series is shown in Fig. 9-14, and, for comparison,
we have sketched the graph of the sum of the first three terms of the series in
Fig. 9-15.
FIGURE 9-15
g = Mi + /] = \ + */,
sin 3x , sin 5x ,
)
,
* + §(
9-4 FOURIER SERIES 345
• • • • • •
h 1
|
1
FIGURE 9
The moral of this example is that a Fourier series can sometimes be found without
recourse to integration. We refer the reader to Exercise 21 for a discussion of
this technique, and to Fig. 9-16 for the graph of the series.
EXERCISES
In Exercises 1-10 find the Fourier series expansion of the given function. Sketch the
graph of the series obtained, paying particular attention to its values at any points of
discontinuity.
(b) What is the value of the Fourier series for /when x = kir, k an integer? When
x = ilk + 1)
^,
k an integer ?
• :
12. (a) Sketch the graph of the Fourier series expansion of the function
fix) = <
|> -1 < X < 1,
1, 1 < X < T,
x = 1, x = x, x = —6, x = 3, x = 7, x = — 57r.
13. (a) Sketch the graph of the Fourier series expansion of the function
2t
X + 7T, — T<X< — »
2T
- -j < x < 0,
fix)
0<x<3>
IT
'•
f < * <
(b) What is the value of the Fourier series for this function when x = kw, k an
integer? When x = (4k +
1)(tt/3), k an integer?
£!_ 1+
x
^
± +±+±+
22 ^ 32 ^ 42
^ ....
6
!L _
^,±4.1 + 1 +r " ..
52 ^ 72
1
"*
8 32
fix)
0, —W < X < 0,
1 1
_ y^ 2a
cot air = -
k *=.
t2
for all x in the domain of/. The smallest positive real number p with this property
(if such exists) is called the fundamental period of/; otherwise p is simply called a
period.
(b) Give an example of a periodic function which does not have a fundamental
period.
(c) Let/i and/2 be periodic functions with fundamental periods p\ and p%, respec-
tively. Prove that ai/i +
0:2/2 is periodic for every pair of real numbers a 1 and a 2
if and only if pi/p2 is a rational number.
(d) Generalize the result in (c) to linear combinations of n periodic functions with
fundamental periods p\, . .
, pn .
19. Determine which of the following functions are periodic, and find the fundamental
period of each if it has one. (See Exercise 18.)
(a) sin —
1TX
(b) 2 sin 3x — cos 2x (c) tan x
2
, s r, x 10 if x is rational
(g) fix) = \
. .
[I if x is irrational
*20. Let
/(x) =
ew'
where Pix) and Qix) are polynomials and (?(0) 5^ 0. Prove that if /is periodic it is
constant. [Hint: Let /(0) = a, and consider the equation Pix) — aQix) = 0.]
21. Let
00 . 00
be, respectively, the Fourier series expansions of the functions /and g in (PC[— w, ir].
+
aao f3Ao
+ S^ aak + ^^ cos kx "*" ^°** "*" P Bk ) sin ^
is the Fourier series expansion of the function af + fig for any pair of real numbers
22. Use Exercise 21 and the series found in Examples 1 and 2 and Exercise 1 above to
obtain the Fourier series expansions of the functions shown in Fig. 9-17.
FIGURE 9-17
E (-ir sin kx
t ( v\ = 0, — 7T < X < 0, - ,
=
X
—t<x<tt.
liW {/ n .
,
and h(x)
N -•>
1, < X < T, X
Use these series, and Exercise 21, to find the Fourier series expansion of each of the
following functions.
(a) f(x) =
—w < x <
\ f
< x < ir
24. What is the Fourier series expansion of 2 + 7 cos 3x — 4 sin 2x, considered as a
function in (PC[-7r, x] ?
—
9-5 I
SINE AND COSINE SERIES 349
25. Without resorting to integration, find the Fourier series expansion of each of the
following functions in (PC[— ir, ir].
2
(a) sin * (b) sin x cos x
3 2
(c) sin x (d) sin x ( cos -
J
2
I/' f(x) dx
2
= ^+J^(al + bt),
T J-
with respect to the basis 1, cos x, sin x, ... in (PQ[— ir, ir], where the ak and bk are
the Fourier coefficients of/.
27. (a) Find the Fourier series expansion of the function /( x) in terms of the expansion
of/(;c).
(b) Use the series found above, together with the appropriate result from Section 9-3
to find the Fourier series expansions of fs and/o, the even and odd parts of/.
(c)Under the assumption that the functions 1 cos x, sin x, form a basis for
, . . .
(PC[— ir, ir], use the results in (a) and (b) to find bases for the subspaces £[— ir, ir]
and 0[-ir, ir] of (PC[—w, ir].
28. Use the results in Exercise 27, together with Exercise 21 and the Fourier series
expansion of e x ,
—t < x < ir, to find the Fourier series expansions of sinh x and
cosh x.
*29. (a) Prove that the functions 1, cos x, sin x, ... , cos kx, sin kx, . . . are orthogonal
in (PQ[a, a + 2ir] for any real number a.
(b) Under the assumption that the functions in (a) are a basis for (PC[— ir, ir],
prove that they also form a basis for (PQ[a, a +
2ir] for any real number a. [Hint:
In the examples of the last section we took advantage of the fact that the functions
considered were even or odd to simplify the task of finding their Fourier series ex-
pansions. This technique can be exploited more often than one might expect,
and is of sufficient importance to be brought out into the open.
Specifically, if /is an even function in (?Q[— ir, ir], then, for all values of k,
f{x) cos kx is even, and f(x) sinkx is odd. Thus, by (9-5) and (9-6),
/ f(x) cos kx dx = 2
J
f(x) cos kx dx,
l f{x) sin kx dx = 0,
and it follows that the Fourier series expansion of an even function in (PC[— ir, ir]
.
involves only cosine terms and may be computed according to the formula
00
where
ak = -2 /
f /(*) cos foe dx. (9-18)
x Jo
A similar argument shows that the Fourier series expansion of an odd function in
(PC[— x, x] involves only sine terms, and may be computed according to the formula
00
where
Actually these results are more than mere formulas. For if we combine them
with the fact that the functions
l,cos x, coslx, . .
is a basis for the space of piecewise continuous even functions on [— t, t], and that
isa basis for the space of piecewise continuous odd functions on [— t, it]. Indeed,
this is the import of the assertion that the series in (9-17) and (9-19) converge
in the mean.
In applications of the theory of Fourier series one frequently needs to obtain
a series expansion for a piecewise continuous function / which is defined only on
the interval [0, *]. One way of doing this is to extend f to the entire interval
[-7T, tt] (where by this we mean that a function F is defined on [— *-, tt] in such
a way that F coincides with /on [0, *-]), and then expand F as a Fourier series. In
those cases where /is reasonably well-behaved, Theorem 9-1 guarantees that the
Fourier series expansion of F will be a good approximation to /on [0, x].
The crux of this method concerns the manner in which /is extended to [— x, *].
This, of course, can be done in any way whatsoever (so long as the resulting func-
tion belongs to <P6[— x, x]), but the following two extensions are the most con-
venient and important. The first is the so-called even extension of/, denoted f E ,
and defined by
yv
\f(-x), -ir < X < 0,
9-5 I
SINE AND COSINE SERIES 351
while the second is the odd extension off, denoted Of, and defined by
(9-22)
Figure 9-18 illustrates the even and odd extensions of a particular/, and furnishes
visual evidence of the easily proved assertion that Ef is even and Of is odd for
any/. This being the case, we can use Formulas (9-17) through (9-20) to obtain
the Fourier series expansions of Ef and Of. They are
OD
Ef {x) = y + ^
fc=i
a k cos kx (mean),
(9-23)
Ofc
2
= - f f(x) cos kxdx,
/
IT Jo
and
00
(9-24)
bk = -2 f f(x)
I sin /ex dx.
IT Jo
These series are called, respectively, the Fourier cosine and Fourier sine series
expansions off, a mild misuse of terminology which rarely causes any misunder-
standing. Unfortunately, the somewhat benighted term "half-range expansion"
is also used in this context.
As an example, we compute the Fourier cosine and sine series for the function
Ok = - I
2
x cos kx dx
IT JO
= r / x sin kx dx
irk Jo
T nit
x cos kx
irk o
+ kJo
cos kx dx
Finally,
2
^T
a = 2- /
/
j
x 2 dx = —^
Li
tt Jo j
and we have
To compute the series expansion of Of we use (9-24) with f(x) = x2 . This gives
h = -2 rx
/
2 .
sin
si foe afrc
IT Jo
2 r x 2 cos
co: kx
o
+ v
k Jo
/ x cos kx dx
2t cos for . 4
- cos kx ,
k
-,
h tt
k3 T
2ir COS kir , 4 , , , N
2tt 8_
k k*ir
= <
2x
fc = 2, 4, 6,
—
FIGURE 9-20
Thus
^ , x
= ~ / .
x
sin 2x , sin 3*
....)
f (x) 2tt I sin 2 1
3
sin
.
x H
sin 3a:
r^— ^
, sin
—5x ,
(mean).
-K- 33 53 f-
)
The graphs of these two series are shown in Figs. 9-19 and 9-20, respectively.
EXERCISES
1. (a) Find the even and odd extensions of each of the following functions in (PC[0, ir],
(b) Sketch the graphs of the Fourier sine and cosine series for each of the functions
in (a).
354 FOURIER SERIES | CHAP. 9
2. Let /be an arbitrary function in (PC[0, 7f]. Prove that Ef is an even function in
(?e[-7r, x], and that Of is odd.
3. Find the Fourier sine series expansion of the function
\Z2tt 7__
16
=
~ 22
1
- 1 62
3__
- 1
+ __5
102 - 1 142 _ i + " '
•
8. Let /be a function in (PC[0, x] which is symmetric about the line x = ir/2. Show
that the only nonzero terms in the Fourier sine series expansion of /are of the form
(^ )'
Show that the only nonzero terms in the Fourier cosine series expansion of /are of
the form
Ak cos kx, k odd,
and find a formula for Ak.
12. Use the results of Exercise 11 to find the Fourier cosine series expansion of the
function
1, < x < ir/2,
/W (-1, ir/2 < x < x.
13. Use the results of Exercise 11 to find the Fourier cosine series expansion of the
function
(b) Use the fact that the functions 1, cos x, sin x, . . . form a basis for <PC[— ?r, it]
to prove that the set of functions in (a) is a basis for <PQ[0, ir].
,
1, cos —
ttx
' sin
.
—
tx
' cos
2irx
'
.
sin
Ittx
» • • •
, n -_ N
(9-25)
P P P P
are mutually orthogonal in (P6[ —p,p] (Exercise 1 below).* Moreover, just as in
the case where p — can be shown that these functions are a basis for this space,
-k, it
and hence that their associated orthogonal series (which, by the way, are still
called Fourier series) converge in the mean. And finally, with due allowance being
made for the length of the interval, all of our earlier remarks concerning pointwise
convergence are valid in this setting.
To obtain formulas for the Fourier coefficients of a function in (pe[ ~p,p]
we note that
/p ax = 2p,
-v
and
cos
>2 kirx
P
dx
,
=
J-p
/
f sin
• 2 —P
kirx
- dx =
fix) = ?r
2
+ S
feV(
Uk cos ~7T
P
+ b k sin ~r
P
)
(mean), (9-26)
* Actually, the entire issue comes down to making a change in scale on the x-axis by
substituting irx/p for x in the functions used earlier.
356 FOURIER SERIES | CHAP. 9
where
ak = — f f(x)cos^—dx,
(9-27)
/p f(x) sin -^— dx
-p P
for all k. And with this we are done.
The above discussion can easily be adapted to handle the Euclidean space
(?Q[a, b]. Indeed, if we set 2p = b — a, so that [a, b] = [a, a + 2p], the func-
tions in (9-25) are also a basis for (?e[a, a + 2p\ This leads at once to the fol-
lowing formulas for computing the Fourier series expansion of a function / in
(?G[a, b\.
x
fix) =
f+ ak cos u
b — ^
a
+ b k sin ,
^_ J
(mean), (9-28)
&=i N
where b
v 2knx ,
ak ) cos t dx,
b — a
(9-29)
rb
Ikirx
b, = x
sm
.
r
,
dx,
)
b — a
for all k.
/•i
/•i
0o = 1, ak = 0, k ^ 0, bk = —
kr
Hence
_ sin 4xjc sin 6irx
= I-i(si
. , . .
FIGURE 9-21
9-6 CHANGE OF INTERVAL 357
/4
flfc = / /(x) cos kirx dx
r3 z-4
= / (x — 2) cos /:7rx dx + I (4 — x) cos &7rx dx,
for any real number a. [At this point we are using the obvious fact that if g is
'b+2p
f g{x)dx = f g(x)dx
Ja Jb
for any pair of real numbers a, b.] We now set a = — 1 in (9-30) to obtain
But on the interval [—1, 1], F coincides with the even function |jc|. Hence bt =
for all k,and a^ = 2 f x cos kirx dx. Thus
k odd,
\ 0, A: even, k ^ 0,
fix)
x
= ^ - ^2 I
cos x*
,
"•
32
,
'
52
,
r- • • • J (mean).
The student will find this technique useful in some of the following exercises.
EXERCISES
1 , cos —
kx
>
.
sin —
tcx
» • •
' cos
kirx
P
'
.
sin
kirx
P
> • •
•
P P
are mutually orthogonal in (PC[— p, p\.
2. Let /be piecewise continuous and periodic on the entire x-axis, and suppose that/
has period 2p. Prove that
ra+2p rb+2p
f(x)dx = f(x) dx
J jb
0, 1 < x < 2.
\x - 2, 2 < x < 3.
5. Find the Fourier series expansion of sin x as a function in (PC[0, t/2], and sketch
the graph of the series.
6. Find the Fourier series expansion of cos x as a function in (PQ[tt/2, 2>tt/2], and sketch
the graph of the series.
7. Find the Fourier series expansion of the function
1, 2 < x < 3,
- x 4 3 < x < 4>
fM = l
nx) >
X 5 < x < 6.
9. Find a Fourier series which contains only sine terms and which converges pointwise
to the function x — 1 for 1 < x < 2.
10. Find a Fourier series which contains only cosine terms and which converges point-
wise to the function x — lforl<Jc<2.
11. Find a Fourier series which contains only sine terms of odd "degree" and which
converges pointwise to the function x 2 — 4 when 2 < x < 3. [Hint: See Exercise
8, Section 9-5.]
12. Let /be a piecewise continuous function of period 2x defined on the entire jc-axis,
and suppose that the Fourierseries expansion of /is
-y + 22
-r- (a fc cos k x + ^* sm kx).
Show that
A k = {-\fa k
and
B k = (-l) fc
6 fc .
13. Given that the Fourier series expansion of x, —t < x < ir, is
sin 2x sin 3x
sin x •
2 3
14. Let /be piecewise continuous on (— °o, <*>), and assume that /(x + tt) = —fix)
for all x where /is defined.
(b) Show that the Fourier series for /has only odd terms.
(c) Prove, conversely, that fix) = —fix + ir) whenever the Fourier series for /
has only terms of odd degree.
(d) What can be said about a function whose Fourier series contains only even
terms ? Why is this situation not particularly interesting ?
Remark: Functions satisfying fix) — —fix + w) on (— oo, oo) are said to possess
half-wave symmetry and are of interest in electrical engineering.
360 FOURIER SERIES CHAP. 9
We have already had occasion to remark that the entire theory of Fourier series
such that
|r(x) - f(x)\ < e
for every x in [
— x, x]
Descriptively, this theorem asserts that the graph of T(x) lies between the graphs
of/+ e and/ — e throughout the entire interval [— x, x], as illustrated in Fig.
9-24. The notation 7V(e) in the formula for T(x) is a reflection of the fact that in
general the number of terms in this trigo-
nometric polynomial will depend upon e,
increasing as e becomes small.
Theorem 9-3 is often stated in the fol-
lowing terms: "Any continuous periodic
function can be uniformly approximated by
trigonometric polynomials," phraseology
which requires some explanation. In the
first place, we have already observed that
t The discussion in this section uses the notion of a closed subspace defined in Sec-
tion 8-5.
9-7 | THE BASIS THEOREM 361
periodic extension F of / will be continuous on the whole real line (Fig. 9-25). It
FIGURE 9-25
As far as the word "uniform" is concerned, it will suffice for now to remark
that it is used to call attention to the fact that the approximation of / by T holds
for all x in the interval [— x, t].
We are now ready to prove that the set 1, cos x, sin x, ... , which we shall
denote by <B,a basis for (PC[— x, 71-]. To do so we introduce the subspace
is
(P[— 7r, x] consisting of all continuous functions in (PC[— x, x] which satisfy the
equation /(— x) = /(x) 5
and establish two preliminary results concerning this
subspace. The first is
Theorem 9-4. Every function in (P[ — x, it] belongs to I((B), the closure
of($> in (Pe[-x, x].
Proof. Recall that S((B) is, by definition, the subspace of (P6[— x, x] consisting
of those functions which belong to S((B) or are limits in the mean of sequences of
functions in S((B), or both.* Thus to prove the theorem we must show that every
fin <P[— x, x] can be approximated arbitrarily closely in the mean by a trigono-
metric polynomial.
To accomplish this we apply the approximation theorem and, for each integer
k > 0, find a trigonometric polynomial T& such that
1
\Tk(x) - f(x)\ <
kV2i
for all x in [— t, x]. Then
h
m-n= [Tk (x) - f(x))
2
dxj
<
\J_r2Trky
Theorem 9-5. Given any real number e > 0, and any f in (P6[ — ir, tt],
\f~ < €.
tion obtained by redefining / near — tt, x and t as shown by the broken lines ,
in Fig. 9-26. (Note that this has been done in such a way that g(— tt) = g(ir), so
that g does indeed belong to <?[— it, x].)
Calculating \\f
- g\\
2
we find that
2
11/ g\\
= f [/(*)- gix)fdx
J T
T+d 2
x +S 2
= f~ if(x)- g(x)] dx+ r [f{x)- g(x)] dx
J —TT Jx
X —5
+ f
Jt-8
[fix)- g{x)fdx,
To make this quantity less than any preassigned e > 0, it obviously suffices to
2
choose 8 < e/l6M Thus the theorem holds for functions with a
.
single dis-
continuity. To give the proof for any fin (P6[— t] just repeat this jt, process at
each point of discontinuity. |
9-7 I
THE BASIS THEOREM 363
Before stating the basis theorem, let us pause a moment to take stock of the
situation. The preceding any function in (Pe[— t, t] can be
result tells us that
approximated arbitrarily closely in the mean by a function in (P[— ir, it]. But
Theorem 9-4 asserts that every function in (?[— t, t] can in turn be approximated
arbitrarily closely in the mean by a trigonometric polynomial. Taken together
these two statements ought to yield a proof of the fact that (B is a basis for
(PQ[ — t, it]. They do, in the following manner.
Proof. We must show that every /in <PC[— t, it] is the limit of a sequence of
trigonometric polynomials. But, by Theorem 9-5 we can find functions fk in
(P[ — 7r, 7r] such that
1
— Tk <
\fk \\
^
Hence, by the triangle inequality,
^ 2k ^ 2k k
For later purposes we also record the following easy consequence of the basis
theorem:
then
1 C 2
2 00
Proof Since the functions 1, cos x, sin x, . . . are a basis for (PC[ — ir, tt], Parseval's
equality is satisfied, and we have
2 2
ii/ii
2
= H# 1 +
fc==l
(/• cos kx)
|cosA:^|| 2
(/• sin kx)
||sinA;x|| :
(9-32)
364 FOURIER SERIES |
CHAP. 9
2 2
(/• cos kx) (J-r f(x) cos foe Jx) 2
||COS/cx|| 2 7T
2 2
(/•sin /ex) _
= (f _ r f(x) sin kx dx) _
" , 2
TiinMI 2 *
With an eye to later applications we now sketch the theory of series expansions
for functions of two variables. A corresponding theory also exists for functions
of any finite number of variables, but since we shall have no occasion to use these
results their formulation is left to the reader.
Our first task, of course, is to define a function space in which to conduct the
discussion. This accomplished by generalizing the notion of piecewise continuity
is
of/ and (x, y) approaches (x , yo) from the interior of any one of the regions into
which R is separated by the arcs of discontinuity.
Any function which is continuous in and on the boundary of R is piecewise con-
tinuous, so that, in particular, such functions as sin {mx) sin (ny), sin (mx) cos (ny),
etc., m
and n integers, are piecewise continuous on any rectangle in the plane.
However, the set of piecewise continuous functions is clearly larger than the set
of continuous functions for any rectangle R. In Fig. 9-27 we have illustrated a
rather general piecewise continuous function, in order to dispel any doubts about
the nature of such functions. We also remark that it is quite legitimate to consider
piecewise continuous functions in regions other than rectangles. The basic defini-
tion remains unchanged, except that we replace R by a region whose boundary
A plane curve
x = x(t)
y = y(0\
is said to be a differentiable arc if the functions x(j), y(f) have continuous derivatives with
respect to t. A differentiable arc which does not intersect itself is said to be simple.
9-8 ORTHOGONAL SERIES IN TWO VARIABLES 365
FIGURE 9-27
(The student will recall that when R is the rectangle a < x < b, c < y < d,
this double integral may be evaluated by computing either one of the iterated
integrals
[ fix, y)g(x, y) dy dx
[ Jc
Ja
or f [ fix, y)g(x, y) dx dy)
Jc Ja
Theorem 9-7. Let {f(x)} and {gj(y)} be orthogonal bases for the Euclid-
ean spaces (PQ[a, b] and (PQ[c, d], respectively. Then the set of all products
The proof of this theorem is given below, following the examples. First, however,
we observe that the generalized Fourier coefficients of any function F in (9Q(R),
computed with respect to the functions in (9-34), are
ai3
(figj) •
(figj)
or, in greater detail,
00
£ ^fMsjiy), (9-36)
and Theorem 9-7 allows us to assert that this series converges in the mean to F.
(The order of summation in such a series is a matter of indifference since the
assertion that the functions figj are a basis for (PC(/?) is not affected by the order
in which they are displayed.)
Example 1. Let R be the rectangle — ir < x < ir, — ir<y<ir. Then the
set of functions
sin mx sin ny, sin mx cos qy,
cos px sin ny, cos px cos qy,
where m and n range independently over the integers 1,2,..., and p and q over
the integers 0, 1, 2, . . . , is a basis for (PQ(R). More generally, the set of functions
sin
.
a
cos -^
b
>
a b
(9-38)
p-irx
cos -
a
sin —r~
b
—
niry .
>
pwx
cos -
a
—
qiry
cos J-rL-
b
F(x,y) = xy
in the rectangle — ir < x < ir, — ir < y < ir relative to the basis given in the
preceding example.
Here we must evaluate the coefficients
of Formula (9-35) for the various functions in (9-37) and the given function F.
But since x cos px and y cos qy are odd functions of x and y,
/T rir
Thus all of the Fourier coefficients of F except the a mn are zero. To evaluate them
we note that
2
I sin tnx dx = j
2
cos ny dy = w
J X J X
J_ x J_ x sin
2
mx cos 2 ny dx dy
x sin mx dx y sin ny dy
TV*
=
~
A_
2
fx
j
•
sin mx dx f
I
J y sin ny dy.
IT Jo Jo
But
t sin kt dt = (— 1)
and it follows that
to+1 JT
(-D m (-1)
n
m-\-n
= (-D
mn
Thus
sin x sin 2y sin 2x sin 3;
xy = 4 sin x sin jy
1 -2 2-1
sin x sin 3y sin 2x sin 2^ sin 3x sin ^
H
To ' 2^2 ' yvi
= y* ^ OT+W sin mx sin «j
4
Series of this sort are called double Fourier series, and, as we shall see, arise in the
study of boundary value problems involving partial differential equations.
We now return to the proof of Theorem 9-7. For the sake of simplicity we shall
assume that {fi(x)} and {gj(y)} are orthonormal bases for (S>Q[a, b] and (?Q[c, d\
As we know, this involves no loss of generality since an orthogonal set can always
368 FOURIER SERIES | CHAP. 9
be normalized in the usual fashion. In this case the set of products {fi(x)gj(y)}
isalso orthonormal, and hence, of course, orthogonal, as asserted in the theorem.
We leave the task of establishing this fact as an easy exercise, and turn to the
problem of proving that these functions are a basis for (PQ(R).
By Theorem 8-3 we know that the set {fi(x)gj(y)} will be a basis if and only
if it satisfies Parseval's equality. Thus it suffices to prove that
r
F{x,yfdR = £ «Jy (9-39)
b
2
F(x,y) dx = JliF.fi?
/. i=l
r- rb
-E F(x, y)f(x) dx
hi(y) = / F(x,y)Mx)dx,
Ja
rb *
/ F{x,yfdx= f^hiiyf. (9-40)
Ja i=i
We now call result from the theory of infinite series which says that a
upon the
series of positivecontinuous functions which converges pointwise on a closed
interval to a continuous function may be integrated term-by-term.* Thus, inte-
grating (9-40), we obtain
* This result is a consequence of Dini's Theorem (see R. Courant, Calculus II, Inter-
science, New
York) which guarantees that such a series is uniformly convergent and
can be integrated term-by-term. (See Appendix I.)
9-8 |
ORTHOGONAL SERIES IN TWO VARIABLES 369
But since {gj(y)} is a basis for (PC[c, d], Parseval's equality implies that
2
hi(y) dy= YjQii.gif
3=1
h %{y)g,{y)dj[
3 =1 Je
for each integer /. If we substitute these values in (9-41), and recall the definition
of hiiy), we find that
*-
j- pa
rd pb
=
.
2
i,j=l, L*' c
/
J/a
F x 3y)fi(x)gj(y)dxdy
(.
00
i,j=l
Thus (9-39) holds, and the proof is complete in the case where F is continuous.
The proof when F is piecewise continuous, but not continuous, requires a more
sophisticated version of the "mending of discontinuities" theorem of Section 9-7
to show that Fcan be approximated arbitrarily closely in the mean by a continuous
function. Although the procedure necessary to prove this result is conceptually
clear, its details are both complicated and unenlightening, and we therefore omit
the argument. But once this fact has been accepted, the general conclusion follows
from the continuous case proved above. |
2
dF dF d F
and
dx dy dx dy
exist and are bounded everywhere in R. Then the double Fourier series for F
converges pointwise to F everywhere in R*
With obvious modifications this theorem remains true for any rectangle.
* See E. W. Hobson, The Theory of Functions of a Real Variable, Second Ed., Cam-
bridge University Press, 1921, 1926.
370 FOURIER SERIES |
CHAP. 9
EXERCISES
1. (a) Verify that the set of functions {fi(x)gj(y)} of Theorem 9-7 is orthogonal in
(PC(/?), and that this set is orthonormal whenever {fi(x)} and {gj(y)} are ortho-
normal.
(b) Find the norms of these functions in terms of the norms of the functions ft(x) and
gj(y); i.e., find a formula for \\fi(x)gj(y)\\ in terms of ||/»(*)|| and \\gj(y)\\.
In each of the following exercises find the double Fourier series expansion of the given
function in (?Q(R); R the rectangle —k < x < tt, —t < y < t.
5. F(x, y) = x
6. F(x, y) the function which is 1 when x and y are both positive or both negative, and
— 1 otherwise.
9. F(x, y) = xy 2
10. F(x, v) = \xy\
10
10-1 INTRODUCTION
In this chapter we shall investigate some of the convergence problems which
arise in the study of Fourier series. Our first efforts will be devoted to proving
the theorem cited in Section 9-4 describing the pointwise behavior of the Fourier
series for a piecewise smooth periodic this has been done, we
function. Once
shall consider the more delicate (and problem of uniform conver-
interesting)
gence, and the related questions of term-by -term differentiability and integra-
bility of Fourier series. Finally, we shall introduce the important notion of "sum-
mability" for infinite series, and use it to extend these results to arbitrary functions
in (PC[— 7r, t]. Throughout this discussion we shall assume that the reader is
familiar with the notion of uniform convergence, and the results contained in
Sections 1-3 and 1-4 of Appendix I. They will be essential in all that follows.
lim j
g(x) sin (Kx) dx = lim f g(x) cos (\x) dx = 0. (10-1)
X—>oo Ja X—»oo Ja
The reader should note that (10-1) has already been established when X —» oo
through the values 2wk/(b — a), k = 1,2,... (see Corollary 8-1). The burden
of the present assertion is that this result is still valid as X tends continuously to
infinity. Intuitively, of course, this is reasonable, since the positive and negative
portions of the area under each of the curves g(x) sin \x and g(x) cos Xjc tend to
cancel one another as X —> oo.
Proof. Since the argument is similar for both functions we shall give the proof
only for g(x) sin Ax. Here if
rb
IQ$ = /
8(x) sm 0^x ) dx, (10-2)
Ja
and if e > is given, we must show that there exists a constant X such that
|/(X)| < e for all X > X To this end we assume for the moment that g is con-
.
rb— t/X i v
/(X) = -/
Ja—Kl\
g[t + \)
A/
sin (\t)dt,
\
•6— jr/A
rO— ir/X r v
2/(X) = - i
g ( jc + y)
sin (\x) dx + / g(x) sin (Xx) dx
./ a tt/X 6-t/X
•b-WX
sin (Xx) dx.
Hence if M denotes the maximum value of the function \g\ on [a, b], and if t/\ <
b — a (which, of course, we may assume), then
• b
2|/(X)| < Ml
'a— *7X
|sin Xx| dx + M I
J b—ir/X
[sin Xjc| dx
6-x/X
b-x/X
2A^r
< g(*) - £ dx
X
To complete the proof in the case under consideration we now use the fact that
g is uniformly continuous on [a, b] to find a constant X such that
!««-*(* + SI <E-r7i
10-3 I
POINTWISE CONVERGENCE OF FOURIER SERIES 373
for all X > X ,and all x in [a, b].* In addition, we suppose that X is chosen
so that, at the same time, Mir/X < e/2 whenever X > X . Then
i/(x)i <! + f
= €
for all X > X , as required. Finally, to establish the assertion for an arbitrary
function in (PQ[a, b], we merely apply the above argument to each of its con-
tinuous pieces. |
In this section we use the Riemann-Lebesgue lemma to establish our first con-
vergence theorem. The main step in the proof consists of deriving a formula for
the partial sums of an arbitrary Fourier series
00
where
and /is any function in (PC[— w, *]. The derivation goes as follows.
Let
Then
= ~
J
f(f) \j + ^ (cos kx cos kt + sin kx sin kt) dt
=
\ j_J(t) [j + £ cosA: (' - *) *•
sin (« + I) 5 — sm 9 =
2
2 sin ^
2
^
k=\
cos ks,
or
/c=l
Hence
Finally, if we now assume that /is periodic on (—00, 00) with period 2t (i.e., if
we replace / by its periodic extension to the entire real line), then Sn (x) is also
v
ds. (10-5)
' tt J- T 2 sin (5/2)
This is the desired result, which is known as Dirichlef s formula for Sn Moreover, .
for future reference, we also note that when (10-4) is integrated from — ?r to t,
we have
''rin£+^ &=ir . (10-6)
f
_ T 2sin(V2)
Now that these facts have been established we can easily prove
Theorem 1 0-1 Let f be piecewise continuous on {—go, go), with period 2tt,
.
and suppose that f(x) = M/(x + ) + /CO] for al1 x Then the Fourier -
provided these limits exist. The reader should note that these definitions reduce to the
usual ones at those points where /is continuous. Moreover, if both of these limits exist
and are equal and if xo is a point of continuity of/, then /is differentiable at xq, and the
value of its derivative is the common value of the above limits. Finally, we recall that a
function is said to be piecewise smooth if it has a piecewise continuous first derivative.
10-3 | POINTWISE CONVERGENCE OF FOURIER SERIES 375
lim
fl*> + *) - JM and lim
JXx^-J,)-/^)
jy y
k J- T 2 sin (s/2)
1 f(x + s) - f(x ) is
7T 7— s sin ^5
sin (« + ^)s g?5.
Moreover, since /has both a right- and left-hand derivative at x , the function
00
where au and && are the Fourier coefficients of/ converges to/(jc ) = M/C*^) +
/(XcT)]- But this is equivalent to showing that the Fourier series expansion of
the function
G(x) = f{x + *„) - /(*o)
converges to zero at x = (see Exercise 4). To do so we decompose G into its
and observe that the Fourier series for Go converges to zero at x = 0. Hence it
is prove that the same is true of the Fourier series for Ge- However,
sufficient to
it follows at once from the definition of G that Ge is continuous at x = 0, and
376 CONVERGENCE OF FOURIER SERIES CHAP. 10
has a right- and left-hand derivative there (see Exercise 4). Thus, by the continuous
case treated above, the Fourier series for Ge converges to %[Ge(0 +) Ge(0~)] = +
G^(0) = 0. From this we conclude that the Fourier series for G, which is just
the sum of the corresponding series for Ge and Go, converges to zero at x = 0,
and the proof is complete. |
In Section 10-7 we
prove a far reaching generalization of this result. In
will
the meantime, however, we conclude our
discussion of pointwise convergence by
evaluating a certain improper integral which will be encountered shortly.
Proof.
Lemma 10-2
f Jo
sin
x
jc
= {x Sm 2
—T < X < X, X J* 0,
f(x)
i
2> x = 0,
and let Sn denote the nth partial sum of the Fourier series for /. Then since / is
piecewise smooth on [— w, t], Theorem 10-1 implies that Sn (x ) —>/(xo) as
n — > go for all jc in [— t, t]. Setting x = 0, and using Formula (10-5), we
therefore have
as n — oo. We now use the fact that the integrand appearing in this expression
is even to deduce that
sin (n + %)s
ds —* ^
as n — > oo . Hence
(n+l/2)x
sin x
dx
/.
sin x ,
= wir
f
Jo x
dx
2
I sin*
Ak = dx
J (k— l)x X
.
FIGURE 10-1
k = 1,2,.... Then, referring to Fig. 10-1, we see that the A k measure the area
between the x-axis and successive oscillations of the curve sin x/x. Thus A i >
A2 > and Ak —> as k — oo (see Exercise 2). From this it follows that the
' • , *•
alternating series
A\ — A2 + A3 — - •
< \A N+ i\.
k=N+l
/•KT
sin* / sin x
dx < dx +
T X JT X j=k+l
< \A k \ + \A k+1 \,
EXERCISES
1. Prove that the function g(s) defined by Formula (10-7) is piecewise continuous on
[-7T, 71"].
2. Let
•Jfcir
sinx
Ak = dx k = 1,2,...
I.(k—l)r X
3. Show that
sint sin
<^
I , I t ,
dt
Jo
/
t -Jot / <fr
for all x.
378 CONVERGENCE OF FOURIER SERIES CHAP. 10
4. (a) Let au, bk denote the Fourier coefficients of/. Prove that the series
ao
-r-
2
+ ^2 iflk cos kxo + b k sin kxo)
converges to zero at x = 0.
(b) Show that the second of the above series is the Fourier series expansion of the
function G(x) = fix + xo) — f(xo). [Hint: Make the change of variable t =
x + xq in the formulas for the Fourier coefficients of G on [— x, ir] and use the
periodicity of/.]
(c) Show that
(d) Show that the right- and left-hand derivatives of GE exist at x = 0. [Hint: Note
that
5. A continuous function /is said to satisfy a Lipschitz condition of order a if there exist
positive constants M and a such that
|/(*i) - f(x 2 )\ < M|xi - x 2 |«
Mir" Mxa
la* < \bk\ <
—
for all k > 0. [Hint: Make the change of variable x = t + (r/k) in the formulas for
a and
fc bk, and deduce that
6. (a) Prove that every continuously differentiable function /on a closed interval [a, b]
a Lipschitz condition of order one on that interval. (See Exercise 5 above.)
satisfies
[Hint: Use the mean value theorem and set = max\f'(x)\ on [a, b].] M
(b) Prove the following generalization of Theorem 10-1. Let f be piecewise continuous
on (— co , oo ) with period 2-k, and suppose that
/(*) = -lim
7T \__xx,
/
J -„
/(x + s) —S
rfs.
sin Xs
f(x + s) ds
OS f(x
0
+ s) — f(x
+
) .....
sin (Xs) ds x+
+.
+ /OO
,
)f
-,
Jo
/
/
sin Xs
s
</s
,
—_).,..,
/.0
m
/(* + s) — f(x
jk -. sin Xs
+, i
.
J —
/
S
sm ^+
/
(X5)
, .,
/(X )
/
/
S
,
ds,
IT J T
8. Prove that
1
-
v^ COS KX =
+ fs
. , sin (n
-—— + i)x
/ ,
:
, ...
2 2sin(x/2)
as follows.
Use Euler's formula
cos x + / sin x
to deduce that
n
Us
1
k=x
cos kx
380 CONVERGENCE OF FOURIER SERIES | CHAP. 10
fc=l
Evaluate this expression by using the formula for the sum of the first n terms of a
geometric series, and then find its real part.
9. Let /be piecewise continuous and periodic on (— °°, °°), with period 2x. Prove
that the same is true of the function
J (x + S)
SH
L (AL +^)£
2sin(*/2)
Now that we have settled the question of pointwise convergence for the Fourier
series expansions of piecewise smooth functions, we propose to determine con-
ditions under which this convergence is uniform on a closed interval [a, b]. Here,
of course, we must impose additional hypotheses on the functions considered,
and at first sight one might expect them to be rather stringent. Surprisingly, how-
ever, we need only demand that the functions be continuous in order to guarantee
both uniform and absolute convergence; an assertion which we prove as
Theorem 10-2. Let f be a continuous function on (— oo, oo) with period 2x,
and suppose that f has a piecewise continuous first derivative. Then the
Fourier series for f converges uniformly and absolutely to f on every closed
interval of the x-axis.
Proof Let
00 / <x>
be the Fourier series for/and/', respectively. Then, since/is periodic with period
27T,
«'o
= - / f\x)dx = -[/(tt) - /(-*)] = 0,
TT J -re TT
= - f(x) cos kx + k I
J —V
f(x) sin kx dx
IT |_ —TT
= -[ T fix) sin kx dx c)
= kb k ,
10-4 UNIFORM CONVERGENCE OF FOURIER SERIES 381
and
bi -if
=
K
fix)
J —,
sin kx dx
Em 2 2
+^)^ - / a*) <&<«>
2
Thus
E t^«f + «]
Jfc=l
2
< «>
and we conclude that the sequence \k\^al + bis, k = 1,2,..., belongs to the
Euclidean space / 2 of all "square summable" sequences of real numbers intro-
duced in Section 8-5. But the sequence {l/k}, k = 1, 2, ... also belongs to t 2 , .
Hence the inner product of these two sequences exists, and it follows that the
series
E U-kVJ+~bl) = ± Val +
k=l \* * fc=l
bl
must converge.
Now, given any pair of real numbers a and b, the Cauchy-Schwarz inequality
2
in Gt applied to the vectors ae x + be 2 and (cos kx)ei + (sin kx)e 2 e 1} e 2 the
, ,
= Va? + b2
-+-
E
fc=i
\
a k cos kx + bk sin kx\
|¥|
1
z '
+ E ^i +
fc=i
*.
-y
2
+ E
fc=i
( flfc cos kx + 6*; sin fac)
382 CONVERGENCE OF FOURIER SERIES CHAP. 10
The reader should note that the hypotheses of Theorem 10-2 merely require/'
to be piecewise continuous, and hence allow/' to be undefined at isolated points
of the x-axis. Thus the Fourier series expansion of a function such as the one
shown in Fig. 10-2 converges uniformly and absolutely in every closed interval
of the x-axis in spite of the fact that /' does not exist at the points lirn, n an in-
teger. On the other hand, the theorem is certain to fail in any interval where /
itself is discontinuous, since it is well known that the limit of a uniformly con-
vergent sequence of continuous functions (in this case the Sk(xj) is continuous.
Thus the above result is we demand uniform con-
the best that can be expected if
Lemma 10-3. Let <p be the piecewise smooth, periodic function on(— oo, oo)
whose definition in the interval [
— t,t] is
0,
<p(x) = 0, x = 0, (10-8)
< <
K-0- x IT,
(see Fig. 10-3). Then the Fourier series for <p converges pointwise to <p for
all x, and the convergence is uniform on any closed interval which does not
contain a point of the form 2irn, n an integer.
10-4 |
UNIFORM CONVERGENCE OF FOURIER SERIES 383
Granting the truth of this assertion, Theorem 10-3 is proved in the following
manner.
Let xi, x 2 , . . . , x m be the points in (— t, t) where /is discontinuous, and for
each i, i = 1, . . . , m, let co t-
denote the magnitude of the jump discontinuity at
Xi [that fix*) — f(x~)]. Then the function u>np(x
is, coi = — X;) also has jump
discontinuities of magnitude co; at the points Xi + 2-wn, but is continuous for all
is continuous both at the points where/is continuous, and at the points Xi + 2im.
In short, by subtracting coi<p(x —
X{) from / we have removed the discontinuities
at the points x{ + 2irn without introducing any new points of discontinuity. We
now repeat this process for each index i, to obtain the function
which is piecewise smooth on (— oo, oo), periodic with period 2t, and continuous
everywhere except possibly at the points ±7r, ±3t, .... [Such discontinuities
will occur whenever /(—7r) ^ /(x).] To remove these last discontinuities we
set co x = /(t + ) — f(r~), and construct the function
m
F(x) = /(*) - 2 w ^(x — **0 ~ u v <p{x - tt).
Then F satisfies the hypotheses of Theorem 10-2, and its Fourier series therefore
converges uniformly to F in every closed interval of the x-axis. Moreover, Lemma
10-3 allows us to assert that the Fourier series for the function
m
$(*) = 2 awC* - **•) + "> T ip{x — tt)
proved.
it^' ("H»
and hence to prove the lemma it suffices to show that this series converges uniformly
on every closed subinterval of [ — tt, x] not containing the origin.
384 CONVERGENCE OF FOURIER SERIES | CHAP. 10
sin 2* sinnx
Son (x)
/ \
= +
• . .
sin x H = h
n
and let
(Note that Sn is the nth partial sum of the Fourier series for the function ir<p.)
Then, since sin kx = Tjdx) — T]c _ 1 {x) for all k > 1, we have
Sn (x) = Tl (x) +
T* (X) ~ ™ + +
Tn (x) — Tn-xix)
or
sb(j0 = 1-2 TM + ^)
2-3
+ '
'
... + r-iW
(n - l)n
'
+ W
Moreover, by summing the trigonometric identity
1 v y|
I
sin ^jc]
(ax — 1)« m
+!
1 1 1
<
|sin(x/2)| m(m + 1)
+ •••
+ (n — \)n
+ *-
n m
Now let x be restricted by the inequalities < 8 < \x\ < ir. Then
. x
sin > sm .
^ 5
and we have
1
\Sn (x) - SUx)\ <
sin (g/2) {[—(m + 1}
+ •••
+ (n — l)nj
+ U-J-
« m
But since the series
^fc(fc+ 1)
10-4 UNIFORM CONVERGENCE OF FOURIER SERIES 385
1 1
m(m + 1)
+ •••
+ (n — l)n_ n m
can be made arbitrarily small by taking m and n sufficiently large. Hence, given
any e > 0, there exists an integer such that N
\Sn {x) - Sm {x)\ < €
for all m, n > N, and all x with 8 < \x\ < t. This, of course, implies that
Y>v sin kx
fc=i
K
is uniformly convergent whenever < 5 < \x\ < ir. The same is therefore true
of the series for <p, and since the choice of 5 in (0, ir) was arbitrary, we are done. |
We
have already observed that the Fourier series for a piecewise smooth func-
tion cannot converge uniformly on any closed interval [a, b] containing a point
of discontinuity of the function in its interior. To complete our discussion
of uniform convergence, it remains to consider the behavior of Fourier series
(a) (b)
in the vicinity of such discontinuities. The diagrams in Fig. 10-4, depicting the
function
-tt < x < 0,
{-1,
0, x = TT, 0, TT,
1, < x < T,
and several of the partial sums of its Fourier series, are typical of the situation
which obtains. From these diagrams it is apparent that the oscillations in the
partial sums, Sn of the Fourier series for / do not decrease at a uniform rate on
,
the interval (0, -n) as n —» oo On the contrary, the oscillations toward either end
.
of this interval remain rather large for all values of n, and though these exceptional
oscillations move toward the ends of the interval as n increases, they do not die
out in the process. This peculiar behavior of the partial sums of a Fourier series,
wherein these sums seem to gather momentum before plunging across a jump
discontinuity, is known as the Gibbs phenomenon, after the American mathema-
tician and physicist, J. W. Gibbs, who first discovered it. In the next section we
shall analyze this phenomenon in some detail, and obtain a limiting value for the
amplitude of the oscillations involved.
EXERCISES
Each of the following exercises refers to an arbitrary trigonometric series of the form
00
Y + k=l
Yj ( ak cos kx + bk sin kx) (10-11)
3. Suppose that
E M + N>
^=1
converges. Prove that (10-11) converges uniformly and absolutely on every closed
interval of the x-axis, and is the Fourier series expansion of its limit.
4. Prove that the conclusion of Exercise 3 holds whenever
N<p- \h\<-&
for all k > 0.
10-5 THE GIBBS PHENOMENON 387
<p(x) = 0, x = 0, (10-12)
7T
< X < 7T.
Esin kx (10-13)
FIGURE 10-5
and by our earlier results we know that this series converges to <p(x) for all x,
the convergence being uniform on any closed interval not containing a point of
the form 27ra, n an integer. Figure 10-6 shows the graph of <p on the interval
[— 7r, ir], together with the graph of the sum of the first six terms of its Fourier
series, and again furnishes visual evidence of the fact that the partial sums of the
series tend to "overshoot" the values of the function near a point of discontinuity.
FIGURE 10-6
Thus, the values of «p between any two jump discontinuities range through the
interval (— ir/2, Sn the nth partial sum of (10-13), range
w/2), while those of ,
determines what is known as the Gibbs interval for <p, and our first objective is to
obtain a precise description of this interval.
To this end we observe that a real number y will belong to the Gibbs interval
for <p if and only if there exists an increasing sequence of positive integers {rik}
388 CONVERGENCE OF FOURIER SERIES | CHAP. 10
k = 1, 2, ... , and a sequence of real numbers {xk} converging to zero, such that
n
1
^ ^ = sin (« + h)t
~ + ^2 cos
o 2 sin (t/2)
dt
1
/**
sin (« + j)t
dt
t/2
Thus
«*>--£+u*^*
+ ^/ lsirrk)-^] sin(rt + ^^ (1(M5)
.(n+l/2)x
_1 / sin (n + 2)* j
"~_ / sinM
fifo.
2J0 V2 «
1 1
lim = 0,
sin (t/2) t/2_
1 1
* The reader who is so inclined may take this statement as a formal definition of the
Gibbs interval for <p.
,
10-5 I
THE GIBBS PHENOMENON 389
is continuous on the closed interval [0, x], provided it is assigned the value
when t = 0. Moreover, if
x
C
1 1
in;x)=X
2Jo sin (//2) t/2_
sin (n + %)t dt,
lim F(n; x) =
x->0
uniformly in n; i.e., given any e > 0, a 5 > can be found such that for all n,
\F(n; x)\ < e whenever |x| < 5. (See Exercise 1 below.) Thus (10-15) may be
rewritten
Mn + l/2)x
Sn (x) = - \ + Jo
^dt + F(n;x),
and follows that if {jt&} is any sequence of real numbers which converges to
it
•h
/•h
lim
k—yao
Snk (x k ) =
Jo
J
—^
~T
dt,
(Note that every real number h can be expressed as a limit of this form, and that in
certain cases hmay assume the values oo .) Hence every point of the form ±
h
sin t ,
at, — oo < n < oo
/.o t
Conversely, if j; belongs to the Gibbs interval for <p, there exist sequences
{xk} and {rik} with the property that Sn (xu) —* yo as k —> oo. But then there
exists a subsequence {«&'*&'} of {rikX k } which either converges or else diverges
to ±oo.* If h denotes the limit of this subsequence, then (10-15) (in its amended
form) implies that
/•h
•h
sin t
——dt ,
Jo
Theorem 10-4. The Gibbs interval for the function <p defined above is the
set of all real numbers y of the form
I sin
~7~J dt (10-16)
Jo
with < h < oo.
Actually, we can say even more than this. For, referring to Fig. 10--8, and using
the interpretation of the integral as area under the curve, we see that (10-16)
assumes its maximum value when h = t. Thus the Gibbs interval for <p is
/•7T rtc
/ sin t , / sin t ,
sin t
sin t
dt
~~t~
x = 0. Once again this implies that the Fourier series for <p is not uniformly
convergent in any open interval having a point of discontinuity of <p as one of
its end points.
It is now a simple matter to pass to the case of an arbitrary piecewise smooth
function. is piecewise smooth on [— ir, x] with a jump discontinuity
Indeed, if/
at x , then, arguing as in the proof of Theorem 10-3, the function
and has a uniformly convergent Fourier series on that interval. But since
,+ ~
fix) = g(x) + /W) f(xo)
<p(x — x ),
we conclude that the Gibbs interval for /at x must be the same as the Gibbs
interval for the function
Thus, on the vertical line x = x the Gibbs interval for /consists of all points y
such that
= \f(xt) ~ f(xo)\
j 08 949
In short, the length of the Gibbs interval for /at jc exceeds the magnitude of the
jump discontinuity in /at that point by the factor 1.089490 ....
EXERCISES
1. Let
i i
foi x)
-'
Proof. Let
y+ X)
k=i
( ak cos kx + bk sin kx) (10-17)
and
~2 + 2
k=i
(°* cos kx + b sin
'k kx ) (10-18)
be, respectively, the Fourier series for /and/'. Then, as was shown in the proof
of Theorem 10-2, a = 0, and, for k > 0,
Thus, (10-18) is
00
Corollary 10-1. Let f be continuous and periodic on (— oo, go), with period
2ir, and suppose that f has and piecewise con-
a continuous first derivative
tinuous second derivative. Then the Fourier series for is uniformly and f
absolutely convergent on every closed interval of the x-axis, and can be
obtained by differentiating the Fourier series for f term-by-term. More
(n_1) {n)
generally, iff, /', . . .
,/ are continuous, while f is piecewise con-
tinuous, then the Fourier series for f \j =
(i
1, . . . ,n — 1 converges uni-
formly and absolutely to f on every closed (j)
interval of the x-axis, and can
be obtained by differentiating the Fourier series for f term-by-term j times.
-~ + ( Qk cos kx + bk sin kx ^
2 X)
*=i
fix) dx = -£(b- a)
^ a (sin kb
fc
— sin ka) — b k (cos kb — cos ka)
k ~1
(10-19)
Proof Set
Fourier series as
°°
A
F(x) = -y + ^ (A
fc=i
k cos kx + Bk sin kx),
Ak = -
r (10-21)
Bk = %> k>\.
Hence
**-** cos **
f'mdt = ^x + ^+E fl * sin
.
(10-22)
•6
/•o pb ra
r
/(*) </* = -± ib - a) + 2_j £
Ja
00
sin ka — b k cos
-E flfc fca
Finally, since both of these series are absolutely convergent, the necessary re-
arrangement of terms leading to (10-19) can be effected, and we are done. |
In much the same way we can also prove the following theorem.
00
has a Fourier series which converges pointwise for all x in the interval (— t, t),
and
x
[ at) dt = y^+ y~ 1
bk cos kx + [ak + ( ~ 1)fc+lflo] sin kx
(10-24)
Remark. The reader should note that since /is arbitrary in (PC[— t, t], the series
in (10-23) need not converge tof(x) everywhere in [— x, t]. This not withstanding,
the series for \lf{i) dt does converge pointwise as asserted in the theorem.
I „ a A x^ a k sin kx — b k cos kx
=
T+g
N , ,
.
f(t)dt
T* +
'
J *
—t < x < it. (See Formula 10-22.) We now set x = and find that
4?=t,T' d°- 25 )
*-!*
and hence that
/"
a ^ ^ bk a k sinkx — b k cos fcx
A mdl 2r+£
.
h k
EXERCISES
1. Verify that the function defined by Eq. (10-20) is periodic with period 2x, and com-
pute the coefficients Ak and Bk (k > 0) of its Fourier series.
Prove that
fit) dt
,
= ,
/
J — T
2_j J.
4. Let /be a piecewise continuous odd function on (— oo , oo ) with period lie. Prove that
for all x.
~-
*+i sin kx
= 2-f( -1 > '
_7r<x<7r '
2 jfe
k=i
*3
for —t < x < 7r. See Exercises 14(b) and 15(b) of Section 9-4.
6. Show that the trigonometric series 2~lk=2 sm kx/\n k is «o/ the Fourier series of any
function (PC[-x, it]. [Hint: See (10-25).]
lim /
fk (x)g(x)dx = / f(x)g(x)dx
k—»O0 J X J —
for all g in (PC[— x, 7r]. [Hint: Apply the Cauchy-Schwarz inequality to the functions
f-fk and g in (PC[-t, x].]
(b) Use the result in (a) to deduce that if £*=i /t converges in the mean to /in
(PC[-7r, *], then
/
f(x)g(x)dx = J2 / /*(*)*(*)<**
(c) Let /and g be piecewise continuous on (— °°, °°) with period 27r, and let au, bk
and a*. /8a= be, respectively, the Fourier coefficients for /and g. Prove that
/x
f(x)g(x) dx = ——^ + y~] (a k a k + 6*/3/fc).
-r 2 f—
o
though the Fourier series for / is divergent in the usual sense. Needless to say,
the method of "summation" which is used to accomplish this feat must be quite
different from the standard one involving the partial sums of the series, and yet,
at the same time, must yield the value which would be obtained by the method
of partial sums whenever the latter do converge. In this section we shall introduce
this new technique of summation, and then use it to prove Fejer's theorem —one
of the most important results in the entire theory of Fourier series.
Let
a + fli + a2 + • •
•
(10-26)
Sk = «0 + «1 + -
' '
+ ak,
ak =
s
° + Sl + '
' + Sk -'
(10-27)
k
for all k > 0. If this sequence converges to a limit cr as k —» oo, that is, if
we then say that (10-26) is summable by the method of arithmetic means, or Cesar
summable, and <r is called the Cesaro sum of the series.
1 - 1 + 1 - 1 + • •
•
(10-28)
with partial sums
S = 1, Si = 0, S2 = 1, ^3 = 0,
fl 1 2 2. 3 3. \
10-7 |
SUMMABILITY OF FOURIER SERIES; FEJER'S THEOREM 397
and has | as a limit. Thus (10-28) is Cesaro summable, with sum \, even though
the series diverges in the usual sense of the term.
00 1
l +4 + i + ••
k=0
with partial sums
S = 1, Sl = 1 + 2' ' Sk = 2 - 1
2*
since, by (10-27),
<Tk = £ C?0 + Sl + • • •
+ Sfc-l)
*-!
2 - -V-
1 1
= 2
k \ 2*~V
and
limcr fc
= 2.
fc—>oo
a + ai + a2 + • '
'
(10-29)
converges to the value <r, then the series is Cesaro summable, and its Cesaro
sum is a.
Proof. Let e > be given. Then since (10-29) converges to <r, there exists an
integer N such that
\sn ~ <r\ < 2
398 CONVERGENCE OF FOURIER SERIES | CHAP. 10
i n—l
= n E ('* - *)•
Wn — <T = E
k=0
(** - o-) + Y^
k=N+l
(s k - <r)
1 .
N . 1 .
n .
^ ^ E
&=o
l
j* - °1 + n
" fc=iV+l
E
k* — <^l
lj; h -«^©
-'-? 5
<
Moreover, since JV is fixed, the quantity
^
1
-
"fc=0
E I
5* - *i
can also be made less than e/2 by choosing n sufficiently large, say n > N', and
when this is done we have
Wn ~ <r\ < + = e
2 2
E
k=0
fl fcW (10-30)
In this case we say that the series is Cesaro summable at the point x if the nu-
merical series
00
E
k=0
°*(*o)
10-7 |
SUMMABILITY OF FOURIER SERIES; FEJER'S THEOREM 399
is Cesaro summable in the sense of the above definition, and that (10-30) is uni-
k=0
converges uniformly to <r(x) on an interval [a, b], where the ak(x) belong to
(PQ[a, b], then the series is uniformly summable on [a, b] to the same func-
tion a(x).
With these preliminaries out of the way, we now state the theorem which
justifies introducing the notion of Cesaro summability.
We have already stated that this theorem is one of the most important in the
theory of Fourier series, and before giving a proof we point out some of its far-
reaching consequences. In the first place, the conditions imposed here are not
sufficient even to guarantee that the Fourier series for /converges pointwise, much
less that the convergence be uniform. In spite of this, Fejer's theorem asserts
that we can "sum" the series in question, and thereby recover the function /.
Moreover this summability is sufficiently well behaved so as to proceed uniformly
—
on closed intervals a truly remarkable fact. Hence one can legitimately say that
the Fourier series for a continuous periodic function in (PC[— t, w] does serve to
determine the function from which it was derived. (At the end of this section we
shall generalize these results in the usual way to piecewise continuous functions as
well.)
Finally, we note that Fejer's theorem also implies the following fact.
2
°
+ ^2 (Okcoskx + bksinkx) (10-31)
k=i
Proof. By Fejer's theorem, (10-31) is Cesaro summable to f(x) for all x. If, in
addition (10-31) converges pointwise
5
when x = x then, by Theorem , 10-8, the
value of the series must be the same as its Cesaro sum, namely f(x ). |
At first sight Theorem 10-11 may seem to be stating the obvious, since the
reader has probably long since assumed that a Fourier series must converge to
the function from which it was obtained if it converges at all. Until now, however,
we have proved this fact only for piecewise smooth functions, and there is nothing
in our earlier results to prevent the Fourier series for a continuous function from
converging pointwise to an entirely different function. This, as we now see, is
impossible.
Turning to the proof of Fejer's theorem
we begin by establishing an elementary lem-
ma on approximating continuous functions,
which, though obvious, does stand in need
of proof. The result in question asserts that
every continuous function on a closed inter-
val can be uniformly approximated by a
"broken line function," meaning a continuous figure io-9
function whose graph is made up of a finite
number of line segments as shown in Fig. 10-9. Since we shall have occasion to
refer to this result in our later work, we state it formally as follows.
Proof The proof follows the construction illustrated in Fig. 10-10, and goes as
follows.
Since /is uniformly continuous on [a, b] there exists a 8 > such that
whenever \x x
- x 2 < 5. (Theorem
\
1-13, Appendix I.) Moreover, since [a, b]
a = x < xi < • • •
< xn = b
in [a, b] such that each of the intervals Ik = [x k _ u x k ] has length less than 5.
The function B is now constructed by successively joining the points (x f(x )), ,
(jci, /(*i)), . . .
,
(xn /(*„))
,
on the graph of/ by straight line segments. Thus
if x belongs to the interval Ik ,
B(x) = /(x _0 + _ (x - x fc _0
Xk _ i
fc
Xk
1
(xk-i,A.Xk--i))
a — XQ xi x2 x n-2 x n-\ X n = b
FIGURE 10-10
\f(x) - B(x)\ =
*fc •*&—
But by (10-32),
— < 1,
\Xk Xk—i\
we conclude that
Our next result is the key to the proof of Fejer's theorem, and is of sufficient
independent interest to be stated separately.
Lemma 10-5. Let f be piecewise continuous on (—00, 00) with period 2t,
and let M denote the maximum value of \f(pc)\ for all x. Then if<r n denotes
the nth arithmetic mean of the Fourier series for f
K(x)\ < M
for all n, and all x.
Proof If s n (x) denotes the wth partial sum of the Fourier series for/, then by (10-5),
/
sn (x)
\
= -1 / r,
Ax + ,
s)
n sin (n
~
i)^ ,
, ,Z.
+
ds.
/ jy .
irj_ r
' 2 sin (s/2)
Hence
2 sin
Z" 1
k=0
= 2 sin -= s,
Now when/(x) = 1, it is clear that s k (x) = 1 for all k (consider the Fourier
series for/). Thus, in this case, a k (x) = 1 for all k, and (10-33) yields
/ m ^ M / sin («/2> ,
v
" ~ mr J- T 2 sin 2 (j/2)
as asserted.
With these facts in hand, the proof of Fejer's theorem is all but obvious. Indeed,
let /be continuous and periodic on (— oo, oo) with period 2%, and let e > be
given. We must show that if n is sufficiently large,
for all x, where an is the «th arithmetic mean of the Fourier series for/
But by Lemma 10-4, we can find a continuous broken line function B, with
period 2t, such that
so that \g(x)\ < e/3 for all x, we let <r'n and a'n denote, respectively, the nth
'
arith-
we have
<rn (x) = <r'n(x) + <(x)
10-7 |
SUMMABILITY OF FOURIER SERIES; FEJER'S THEOREM 403
But since \g(x)\ < e/3 for all x, Lemma 10-5 implies that the same is true of
n (x)\.
\<T' Finally, since the Fourier series for B converges uniformly to B on every
closed interval of the x-axis, Theorem 10-9 allows us to conclude that its as-
- =€
\f(x) <Tn (x)\
<f + f
+ f
>
Functions with jump discontinuities can be handled in much the same way, in
which case we have the following generalized version of Fejer's theorem.
Theorem 10-12. Iff is piecewise continuous on (— oo, co), with period 2ir,
then the Fourier series for f is Cesar o summable for all x, with sum
+
f(x ) + f(x~)
2
The proof has been left to the reader as an exercise (Exercise 8).
At this point we can truthfully say that the theory of Fourier series for piece-
wise continuous functions is complete. Indeed, we now know that every such
series is "summable," either in the standard fashion when the function involved
is piecewise smooth, or by the method of arithmetic means. Moreover, the series
will always converge pointwise in either the standard or Cesaro sense to the func-
tion from which it was derived (so long as the obvious proviso is made for points
of discontinuity), and this convergence will be uniform on any closed interval
inwhich the function is continuous. Truly, then, there is little more to be said.
At the same time, however, the reader should not be misled into thinking that
we have uttered the last word on the subject of Fourier series. This is far from
being the case. Nevertheless, at this point the direction of inquiry changes abruptly,
and addresses itself to the task of generalizing the above results to wider classes
of functions. Such generalizations do exist, but are based upon an entirely dif-
ferent type of integral than the one we have been using, and are beyond the reach
404 CONVERGENCE OF FOURIER SERIES | CHAP. 10
of an introductory text. Thus, as soon as we have settled the one point which still
EXERCISES
1. Let £fc=o a k and XX=o b k be Cesaro summable, with sums o- and t, respectively.
Prove that
J] (aa k + &b k )
(a)l+0-l + 0+l+0-l+---
(b)l+0 + 0-l+0 + 0+l+0 + 0-l+---
4. (a) Show that the series
is Cesaro summable to zero in the interval (0, lid. [Hint: Use Formula (10-10).]
is Cesaro summable to zero in the interval (0, 2ir). [Hint: Use Formula (10-4).]
Tk = <y\ + o"2 +- • • •
+ q~fc
The sequence {r k } obtained in this way is known as the sequence of second arithmetic
means associated with the given series, and the series is said to be summable to the
value r by the method of second arithmetic means if lim^c, r k = r. (Higher orders
of summability by arithmetic means can also be defined.)
(a) Prove that the series
1 - 2 + 3 - 4 + ••
is not Cesaro summable, but is summable by the method of second arithmetic means.
10-8 |
THE WEIERSTRASS APPROXIMATION THEOREM 405
Cesaro summable.
(b) Use the result in (a) to deduce that the series
l
2 - 2 2
+ 3
2 - 42 +
is not Cesaro summable.
8. Prove Theorem 10-12. [Hint: First show that it is sufficient to prove
and
£ = /> + «»->^£^*-
2
/0 xn -m sin (n/ 2)s
[fix +
. >^
s) - jy
f{x )] ds =
-, 2sin^ (s/2)
for all x. Let e > be given, and divide the integral appearing in this expression
2
— 1
mc J
f
/
r
[fix
rt
+. ^
s) - ft
fix^)]
+m
2
sin
sm J
Q*/2)s
(s/2)
ds
o
sin
+ —
mr Js / [/(* + *) - /(^
+
)i
in/2)s
ri-nr/em
2 2
sin' (5/2)
ds
where 5 is chosen so that \f(x + s) — /(*)[ < e for all s in the interval (0, 8].
Theorem 10-13. Let f be continuous and periodic on (— oo, oo) with period
2ir, and let e > be given. Then there exists a function g which is continu-
ously differentiable and periodic on (— oo, oo) with period 2ir, such that
\fix) - gix)\ < e for all x.
F ix) =
yJ_ Ax+
d i)dt.
406 CONVERGENCE OF FOURIER SERIES | CHAP. 10
(Note that F§(x) is the average value of/(x) in the interval x — 5 < t < x + 5.)
Then Fs is continuous and periodic on (—00, 00) with period 2w, and if we set
u = x -f- /, so that
+8
^O) = ±f T5 /
25 Jx—5
/(") du t
we find that
(see Appendix I). Thus F'6 is also continuous and periodic on (— 00, 00).
Next, we observe that
-x-\-8
/•x-f-o x+S
r-X-^-0
r
[/(")- Ax)]du
25 x —
x+b
< 1/(M)_ KX)ldU '
fsJx-5
But since /is continuous and periodic on (—00, 00), it is uniformly continuous,
and hence there exists a number 5 = 5(e) such that \f(u) — f(x)\ < e whenever
\x — u\ < 5(e). We now set g = Fs (e) and use the above inequality to deduce
,
that
rx+d(e)
N
Tn(x) = A + ^
k=i
(A k cos kx + Bk sin kx)
such that
\f(x) - TN (x)\ < €
for all x.
Proof According to the preceding theorem we can find a function g such that g
and g' are continuous and periodic on (—00, 00) with period 2ir, and \f(x) —
g(x)\ < e/2 for all x. Thus if Tn denotes the trigonometric polynomial consist-
ing of the nth partial sum of the Fourier series for g, Theorem 10-2 implies that
there exists an integer N such that \g(x) — TN (x)\ < e/2 for all x. Hence
In the next chapter we shall have occasion to refer to the version of the Weier-
strassapproximation theorem which uses ordinary polynomials, and hence we
now prove this result as well. Our argument is based on Theorem 10-14, and
begins with the following lemma.
Lemma 10r-6. If
n
T(x) = a + ^ an cos kx (10-36)
T(x) = Ao + A i cos x + • • •
+ A n cos n x.
T(x) = a + ( ^2 dk cos kx
J
+ an cos nx,
+ n_1 x
T(x) = Aq Ax cos x + • • •
+ An _ 1 cos + an cos nx.
But
cos nx = 2 cos [(« — l)x] cos x — cos (n — 2)x,
and, applying the induction assumption once more, we can write cos (n — \)x
and cos — 2)x as polynomials involving powers
(« of cos jc. This, of course,
implies that T can be written in the desired form. |
Theorem 10-15. Let f be continuous on the interval [—1, 1], and let e >
be given. Then there exists a polynomial P such that
for —t < t < t. Then F is continuous, periodic, and even. Hence the Fj
approximations of Theorem 10-13 are also even, and it follows that their Fourier
series contain only cosine terms. Thus, by the preceding lemma a partial sum
T(t) = Aq +
Ax cos t +
A n cos n / of the Fourier series for some F$
• • •
+
will satisfy
for all t in [— t, t]. Setting x = cos t yields \f(x) — P(x)\ < e for all x in
[— 1, 1], with
P(x) = A + A x x + •••
+ A n x n .\
for all x in [a, b]. In other words, a continuous function on a closed interval
can be uniformly approximated by polynomials.
The asserted result follows immediately from Theorem 10-15 and the fact
1 1-1 INTRODUCTION
In this chapter we continue the study of series expansions in Euclidean spaces
by introducing three classic orthogonal series expressed in terms of polynomials.
At the moment, of course, it is not at all clear that there is anything to be gained
by using polynomials in place of trigonometric functions, and we shall have to
ask the reader to reserve judgment on this point until we come to the study of
boundary value problems. In fact, this entire chapter can be omitted without
serious prejudice until Chapter 13 has been read, and even then it will be possible
to continue with no more than a knowledge of Legendre polynomials and series
(Sections 11-2 through 11-4). Nevertheless, the reader who pursues this dis-
cussion to its conclusion will enhance his appreciation for the scope and subtility
of the theory of orthogonal series, and will be that much better prepared for the
material which follows.
Before we begin, it may be appropriate to remark, once more, that our present
investigations are a natural outgrowth of the ideas developed in the study of
Euclidean spaces. Thus, although certain portions of the following discussion are
technically involved, the issues at stake are simple and familiar. The student who
keeps this point firmly in mind as he reads on should then be able to see the forest
while among the trees.
sequence of polynomials
As we shall see, these polynomials actually form a basis for (PC[— 1,1], and were
itnot for the complications involved in applying the orthogonalization process
we could proceed directly to the study of series expansions relative to (11-1). But
these complications do exist, and are serious enough to make it easier to start
409
410 ORTHOGONAL SERIES OF POLYNOMIALS CHAP. 11
\ 1 Port y
p ^x
y I
\
\ w
p*w / / //
\
-l
1 \
Y
/
\
/
/ // \ \ / II 1
_1 ~
A
FIGURE 11-1
afresh with a slightly different polynomial basis. This, then, is the reason for
introducing the so-called Legendre polynomials —they are a basis for (PC[— 1, 1],
PoW = 1. (11-2)
and
^*> = 24^ -^ 2
(11-3)
for n > 0.Then Pn (x) is called the Legendre polynomial of degree n, and
(11-3) is known as Rodrigues' formula for these polynomials.*
Po(x) = 1, Pi(x) = x,
P 2 (x) = fx 2 - h P 3 (x) = fx 3 - fx,
P 4 (X) = 3AX 4 - ¥X 2 + f, P 5 (X) = ^X 5 - ^X S + tf*.
Ri, . . . , R m -i
an orthogonal is basis for (P m and every polynomial
, Q of degree
<m can be written in the form
Q = <xqRo + • • •
+ am -iR m —i,
where
«* = k = 0,...,m- L
I^IT'
Hence
Q Rn — '
«o(^0 Rn) "
+ ' ' "
+ «m— lC^m-l * Rn),
and since Rk Rn =
• if k ?* n, it follows that Q Rn =
•
0, as asserted. |
Theorem 1 1-1. Let {Q n (x)} and {R n (x)}, n = 0,1,2, ... ,be orthogonal
sequences of polynomials in (?Q[a, b] indexed by degree. Then, for each n,
Q n and R n are scalar multiples of one another.
Proof. Since Q , . . . , Qn is an orthogonal basis for <P n+ i,
RnW =
|^gj Qn ( X ),
Finally, ifwe accept the assertion that the Legendre polynomials are mutually
orthogonal in <PC[— 1,1], Theorem 1 1-1 implies that each Pn is a constant multiple
of the corresponding polynomial in (11-1). This is, in fact, the case, and we
shall see later that
= (11-4)
Pn(x)
£0y p 2 n (x),
EXERCISES
n 9* 1.
4. Prove that if (PC[— 1 1] has a basis (B consisting of polynomials, then (B must contain
,
exactlyone polynomial of each degree. Hence deduce that up to scalar multiples the
sequence of polynomials given in (11-1) is the only polynomial basis for (PC[— 1, 1].
Our first task is to prove that the Legendre polynomials are mutually orthogonal
in (P6[- 1,1]. To this end we begin by showing that Pn is a solution of a certain
linear differential equation, a fact which in itself is of considerable importance.*
Thus let (x 2 l) —
n =
w, and let w
{k)
denote the /cth derivative of w. Then
* A direct proof of orthogonality can also be given along the lines suggested in Exer-
cise 5 below.
—
11-3 |
ORTHOGONALITY: THE RECURRENCE RELATION 413
and, multiplying by x
2
— 1, we have
O - 2
l)w
(1)
- 2nxw = 0. (11-5)
(x
2 - l)w
(2)
- 2x(n - l)w (1) - 2nw = 0,
(x
2 - 1)m/
3)
- 2x(n - 2)h>
(2) - 2[n + (n - l)]w
(1)
= 0,
{x
2
- l)w
(k+2) - 2x[n - (k + l)]w
(k+1)
-2[n + (n - 1) + h (« - &)]h>
(A:)
= 0.
But since
(2» - fc)(fc + 1)
+
n _l (n
/ - 1
1)
^_+L • • •
+r /
(n - ia
k) = ^
{x
2
- \)w
(k+2) - 2x[n - (k + l)]w
a+1) - (2n - k){k + l)w
(k)
= 0.
Pn = ^.
2"n\
(1 - x
2
)P'r[ - 2xP'n + n(n + \)Pn = 0,
(1 - x
2
)P'rl - 2xP'n + n(n + \)Pn= 0,
(1 - 2
x )P% - 2xP'm + m(m + \)Pm = 0,
1
(1 - 2
X JPm PX - P'r^Pn] ~ 2x[Pm Pn ~ P^n] = PmPn[m(m + 1) - «(« + 1)].
(11-7)
But the left-hand side of this equation is just the derivative of
1 1
(1 - x 2 )[Pm P'n - P'mPni = [m(m + 1) - n(ii + l)]/" Pm {x)Pn {x)dx.
l-i J-i
Finally, since 1 — x 2 vanishes at the upper and lower limits of integration, we have
and thus, if m 5* n,
1
Pm (x)Pn (x) dx = 0.
J
This completes the proof of
(PC[-1, 1].
Here we begin by considering the function xPn (x), which is obviously a poly-
nomial of degree n 1. +
Thus, since P P u Pn +i, is an orthogonal basis , . . . ,
*f,M = "fP^W-
p p
*=„
"- "
p
r n + \ -*n+l
f v\
-
This equation can be simplified still further if we note that xPn {x) 2 is an odd
function (it contains only odd powers of jc), for it then follows that
the coefficient of xk in Pk is
1
2k(2k - 1) • • •
[2k - (k - 1)] = ^^ •
(11-9)
Secondly, since
(x
2
- if = x
2k
- kx
2k ~ 2
+ Kk ~ l)
x
2k ~* ,
~2
the coefficient of xk in Pk is
( fc
-
_j>_ (-,0(2* - 2) ••• [2k - (k + 1)] = - ! (U-IO)
2Hfc (f_ 2) ,
•
We now use (11-9) to compute the coefficients of x n+1 on both sides of (11-8)
and equate the results, obtaining
CL = n + 1
»
2/i + 1
~x
To find |8 we
use (11-9) and (11-10) to compute the coefficients of xn on both
sides of (11-1 1), again equating results. This gives
_ ( 2yi ~ 2> !
= _ "+ 1 [2(« + 1) - 2]! [2(/t - 1)]!
2»(/i - !)!(« - 2)! 2n + 1 2*+i«!(n - 1)! + ^2»-i[(/i - l)!] 2 '
1
P =
2/1+1*
We now substitute this value in (11-11) and solve the resulting expression for
Pn +\ to obtain
Equation (11-12) is known as the recurrence relation for the Legendre poly-
nomials and can be used to deduce properties of Pn +i from those of the two
immediately preceding polynomials. For example, given that P (x) = 1 and that
Pi(jc) = x, we can compute P 2 from (11-12) by setting n = 1. This gives
i^C*) = 2X ~ 2-
Ps(x) ~ 3 X \2 X ~~ 27
— 3*
— 2A — JS-y-O 3
2 Ay
Proof The assertion obviously holds for P and P^ Moreover, if we assume that
P m _ LnIn ++ n
t-
1
1
i
_
n
"
+ 1
= i
n
Inmuch the same way it can be shown that Pn (— 1) = (— \) .
We now use the recurrence relation to compute the value of ||P n ||. Again a
trick is needed, and this time it is furnished by the polynomial
A simple calculation using (11-9) reveals that this polynomial is of degree less
than n (see Exercise 8), and hence, by Lemma 11-1, is orthogonal to Pn in
(P6[— 1, 1]. Thus, if we multiply (1 1-13) by Pn and integrate from — 1 to 1 we get
i /.i
Pnipcfdx = 2n ~ l
/ xPn _i(x)Pn (jc) dx. (11-14)
—l n y_i
This is half of what we need. The other half is found by multiplying the recurrence
relation by Pn -i and then integrating. This gives
n +
, ,
1 J _i
,
Pn-l{xfdx = ^^
n -\- l
/
J_i
xPn _ l (x)Pn (x)dx.
Thus, by (11-14),
In -
/ Pn {xf dx = s!
Pn -i(*) Jx,
In + 1 J
(
or
2 In - 1
l|2 =
\\Pn\\ - \\Pn- ill n 1,2,
In + 1
\\P,f — 1.0
3 z
ll^ll
2 — 5 3 z
In - 1 In - 3 2n — 5
ll^.f In + 1 2n — 1 In - 3
1-^
— —
them interesting as that may be we let matters rest where they are, and bring
this discussion to a close with a theorem which will be of some importance in the
following chapters.
1 1
Pn (x)dx = Pn (x)P (x)dx = 0,
f J
1 1
by orthogonality, and it follows that Pn must change sign at least once in the interval
— 1 < x < 1, and hence has a root in this interval. Now let x lf x 2 , . . . , xm be
the roots of Pn in (—1, 1), and consider the polynomial
Q(x) = (x - xO(x - x 2) • • •
(x — xm ).
Pn (x)Q(x) dx 9± 0.
/—
However, by Lemma 11-1, Pn is orthogonal to every polynomial of degree less
than n, and hence would be orthogonal to Q were m < n. The preceding in-
equality excludes that possibility, and forces us to conclude that m = n, as
asserted. |
EXERCISES
for all n, and use this result together with Theorem 11-5 to deduce that
n
Pn(-l) = (-D .
3. Compute /V
4. Prove that
2
Pn (x) = [(2n)\/2 n (n\) ]p n (x),
f^Pn^x" dx =
for every non-negative integer m < n. [Hint: Use integration by parts, and the result
in (a).]
(c) Use the result in (b) to deduce that the Legendre polynomials are mutually
orthogonal in (PC[— 1, 1].
J
11-3 |
ORTHOGONALITY: THE RECURRENCE RELATION 419
6. (a) With w and w (n) as in Exercise 5, use integration by parts to prove that
l 1
n
w
{n)
w
{n)
dx = (1 - x) (l + xfdx.
f (2/i)!
P„(x) dx =
-i 2n + 1
- n « =
- k) (2/i - *)(* + 1)
+
« _u in
/
+
1) j_ • • •
+
_l (n
r
8. Prove that P„(jc) — [(2n — l)/n]xPn -i(x) is a polynomial of degree < /*.
9. Use mathematical induction to prove that the nth derivative of the product of two
functions is given by
=
a»)
g (t
) . . ,
where
^n+1
2(n + 1)
U 1)Fn + /z + 1
Xtn + 2
Fn -
(b) Use the result just obtained and the differential equation for Pn to prove that
P'n + 1 = XK + (« + DPn.
11. Differentiate the recurrence relation for Pn +i, and use the result obtained together
with that of Exercise 10(b) to prove that
f In
(a) xPn (x)Pn -i(x) dx = _ > n > 1
J 4n2 1
14. (a) Use Exercise 12(b) and some other suitable identity to prove that
2
— 1
Pn = xPn ^ + P'n-\ '
(b) Replace n by n — 1 in the equation of Exercise 10(b), and square the result.
Square the identity in 14(a), and use these two relations to establish
-~^
l
[P'nf + [Pn]
2
= ±=/- [P'n-lf + [Pn-^f
for n > 1.
2
1 —
=—^- [PLf + [Pnf < 1
n2
\Pn(x)\ < 1
Show that Q(xi) = a» for < i < m, and then prove that no other polynomial of
degree less than or equal to m has this property.
17. Let F(x) be any polynomial of degree < 2m + 1, and let xo,...,xm be the roots
of Pm+i. Divide F(x) by Pm+ i(x) to obtain
where P(x) is of degree < m, and i?(x) is either zero or of degree <m.
(a) Prove that
f F(x)dx = f R(x)dx.
(b) Prove that R(x) is the polynomial Q(x) constructed in Exercise 16(b) if a{ is
taken as F(x,-), / = 0, 1, . . . , m.
11-4 | LEGENDRE SERIES 421
-i
i
\P n (x)\ dx <
\/n
—
for all n > 1. [Hint: Use the Cauchy-Schwarz inequality.]
19. Prove that none of the Legendre polynomials has a repeated root. [Hint: Observe
that any repeated root of P n
is also a root of P'n and then use the uniqueness theorem,
is only natural to ask if they form a basis for (Pe[— 1, 1]. The answer is that
they do, and a proof of this fact will be given toward the end of the present section.
For the moment, however, let us accept the truth of this assertion, and consider
the series expansion of an arbitrary function / in (P6[— 1, 1] computed with
respect to the Legendre polynomials. By Formula (8-22), this series assumes
the form
n=0 H^ 71 ' 1
Since /is an odd function on [—1, 1], so is f(x)P 2n (x) for all n, and we have
a 2n = 0, n = 0,1,2,
Moreover,
«2n+l = (4« + 3)
Jo
/ ^n+lW dx.
To evaluate this integral we write the differential equation for Pn in the form
1 *
Hence
a 2n +\ =
4« + 3 p " = 0, 1,2,
20s + n 2n(0),
Finally, since
^
«1 —
— 3
2>
and we have
(2« - 1)
2 ,
V, n» 4n+3 3 5 1 • • • • -
(
,
f(x
+ + f{x-)
)
2
This said, we now turn to the one major item of unfinished business —a proof of
the fact that the Legendre polynomials form a basis for (PC[— 1, l].f Again the
most important step in the proof is furnished by the Weierstrass approximation
theorem, phrased this time for ordinary polynomials.
Theorem 1 1-9. Let f be a continuous function on a closed interval [a, b], and
let e be a positive real number. Then there exists a polynomial p such that
* In fact, can be shown that the Legendre series for a function converges pointwise
it
(P6[— 1, 1], must also contain all polynomials. This fact, combined with the
Weierstrass approximation theorem and Theorem 9-5 will allow us to show
that every function in (PC[— 1, 1] is the limit in the mean of a sequence of functions
in S(cE), which, of course, will prove the basis theorem.
Indeed, if /is a continuous function in (PC[— 1, 1], Theorem 11-9 provides a
sequence of polynomials {Qk}, k = 1, 2, ... (the subscript does not indicate ,
l
l/(*) ~ Qk(x)\ <
VSk
for all x in [— 1, 1]. But then
and we conclude that {Qk} converges in the mean to /. Hence S(£) contains all
II*
~ gk\\ < 2k"
Since each gk is continuous we can apply the preceding argument to find a poly-
nomial Qk such that
< 2k
+ 2k
~ k*
and the sequence of polynomials {Qk} converges in the mean to g. Thus g also
belongs to§>(£), and we have proved
Theorem 1 1-10. The Legendre polynomials form a basis for the Euclidean
space (Pe[— 1, 1].
11-5 I CONVERGENCE OF LEGENDRE SERIES 425
EXERCISES
1. What is the formula for the Legendre series expansion of an even function in
(PC[- 1,1]? Of an odd function ?
2. Prove that P 2 «+i(0) = 0, n = 0, 1, 2, . .
. , and that
JV.OT-C-D'
1 '.3 '.5 "'^" ' «-l.2
2-4-6 (2«)
3. Find the Legendre series expansion of the function \x\ in (?G[— 1, 1]. [Hint: Use
integration by parts and Exercise 12(a) of the preceding section to evaluate the
coefficients.]
4. Prove that P'n = (In - l)Pn -i (2« - 5)Pn _ 3 + + (2/i - 9)P„_ 5 H , for all
n > 1, by expanding P'
n in a Legendre series.
5. Find the Legendre series expansion of each of the following functions in (?Q[— 1, 1].
(a) jc
3 (b) x b - x 3
+ 2 (c) 4x 4
+ 2* 2 - x.
1, -1 < x < 0,
/(*)
0, < x < 1.
Sn (x) = J] a k Pk(x),
k=0
where
• 1
2k + 1
f(t)P k (t)dt.
«fc
— 1
Then
2k +
Sn(.x) = ^
fc=0
1
f(t)Pk (t) dt Pk(x)
2k + 1
P k (t)P k (x) f(t)dt
fc=0
where
Kn {t,x) = ^ ^~^P
k=0
k (t)P k (x).
We now interchange the roles of x and ? in this expression and subtract (11-18)
from the result to obtain
or
Kn (t,x) =
rs , n n +
— =
1 >w+1 (QPn (x) - Pn (t)Pn+1 (x)
(11-19)
t — x
The two formulas just derived allow us to express the difference between Sn (x)
and f(x) in integral form. The following lemmas will enable us to show that this
difference approaches zero with increasing n whenever /is reasonably well-behaved,
and therefore play the same role in the present argument that the Riemann-
Lebesgue lemma played in the study of Fourier series.
11-5 |
CONVERGENCE OF LEGENDRE SERIES 427
lim »
1/2
|
T g(x)Pn (x)dx = 0. (11-22)
0-ig(x)pn (x)dX y
,
g(x)
,2
dx
.
= v^
2^ -,
£, n=o J_j Pn (xy dx
= Z [/_
n=0
P^ x f dx
Hence
2
= ?!L li)"' £(x)Pn (.x) dx =
Si (sttt)" kl SL ( i
0,
—i
which, in turn, implies the desired result. |
Id ( \\
\Pn(x)\ < — ^ (11-23)
/2(1 _ x ^ l/2
for all n > 0, and all x in the interval (—1, 1).
The proof of this inequality is long and somewhat involved. Hence, rather than
run the risk of obscuring our present argument, we have deferred the proof to
the exercises where sufficient directions have been given to enable the student to
work the details through for himself.
/
Now let 'be an arbitrary function in (PQ[— 1, 1], and let x be a point in the open
interval (—1, 1) at which /is continuous and has a right- and left-hand derivative,
i.e., a point where
-
lim
f0)
t
~
—
/(X
Xq
^ and lim
At)
t —
/fro)
Xq
<-»x+
1 ,
- _ n + 1 f{t) - f(x
Sn(xi) f(x )
—
)
[Pn+l (t)Pn (x ) - Pn (t)Pn+l (x )) dt
—1 t Xq
1
—±~ P
n
2
l
n (x ) I
^ ^^r-^- pn+i(» dt
- n
~±^P
L
n+l ix«)f
J_ 1
^-^P
— I Xq
n (t)dt.
But, by Lemma 11-3, there exists a constant M such that, for all n > 1,
0±A pnK x ( \
0) <r
M IL±1
2 ^ 2(1 - x|)i/2 „l/2
<
M 2n
2(1 - *2)l/2 «l/2
= Kn 112 ,
where
K= M/{\ - xl)
112
.
Moreover (and this is the critical point in the argument), the assumption that /
has a. right- and left-hand derivative at x implies that the function
f(0 - f(x ) . , . , .-
i .^ ^: i,
/ — x
'
lim
"+ l
Pn (x f^-^P^Od,
n—>oo )f
r1
< Klim n
112
f
At)-,fix ) P (t)dt
- n+1
n—+oo J -i t .
= 0.
Similarly,
lim
n—>x
—+~- Pn +
n
2
1
l(Xo) I
— 1
^—^ t Xo
Pn(t) dt = 0,
liml5n (x ) - /(x )| = 0,
11-5 I CONVERGENCE OF LEGENDRE SERIES 429
Theorem 11-12. The Legendre series for a function fin (PC[ —1,1] converges
Lemma 11-4. Let x be an arbitrary point in the open interval (—1, 1).
Then
«x
Pn+l(t)Pn (x ) ~ Pn (t)Pn + l (x ) ^
.. ... -1 t — Xo
and
~ Pn (t)Pn +
m !±A ^
(x
li
n—>oo -^ J
f Pn + l(t)Pn(x
Xn
*
)
-*()
1 )
f(x )= \U -1 <*<*o,
(0, xo < x < 1.
Since
pn(x ) =
2n~\^[
[K+l(x) ~ K ~^ x)]
for all n > 1 (Exercise 12(a), Section 11-3), the general coefficient in this series is
/x
Pn (x) dx
[Pk+i(x) - Pk-i(x)] dx
i
Moreover,
*x
1 /
a =
and it follows that the value of the nth partial sum of this series at the point x is
= 2 + 2^n+l{Xo)Pn (.Xo).
But by (11-20),
/xo
Pn+\it)Pn {Xo) — Pn (t)Pn+ i(Xo) ,
-1 t — Xq
Hence, by (11-23),
x
Pn+1 (t)Pn (x ) - Pn (t)Pn+1 (x ) ^
n— >oo / -1 t - Xq
= \ + \ lim P„+i(^ )^n(^o)
This proves the first statement in the lemma, and the second follows by sub-
tracting this result from (11-21). |
Now let /be any function in (P6[— 1,1] with the property that
m . /(*
+
> + /<*->
for all a: in (— 1, 1), and let x be any point at which/has a right- and left-hand
derivative. Then, if Sn (x ) denotes the value of the nth partial sum of the Legendre
series for/ (11-20) implies that
Sn (x ) = h+h
where
— XQ
1
^
J Xn t
11-5 I CONVERGENCE OF LEGENDRE SERIES 431
Thus
+>
f(xo) Kxt)
SnOo) - /(*o) = h ~ + h -
and to prove that
lim h ~
Axo) = lim h - /W) = o.
n—>oo
f(xo)
lim
n—»oo
h ~
Now let
, x
/(0 - /W) , \ <t<x
lo, x < / < 1.
and we have
lim
n—>qo
h ~
f(Xoj = v
lim
n =
—+ 1
/:
g(t)[Pn+1 (t)Pn (Xo ) - Pn (t)Pn+1 (x )] dt.
f(xo) =
lim 7i
n—>oo
- 2
0.
+>
lim
/«) = o,
n—>oo
thereby proving
f(xt) + fixo)
at each point x in the open interval (—1, 1) where f has a right- and left-
hand derivative.
EXERCISES
lim (
—— - ) / g(x)Pn (x)dx = o,
then
/2
lim n / g(x)Pn (x) dx = 0.
n —>oo I J —
n
Mx) = -
7T
/
Jq
[x + O2 - 1)
1/2
cos <p] d<p.
2 2
= - / {— n(x 2 — l)sin <p + (x — 1) '
cos<p[x + (x — 1) '
cos <p]f
TT Jo
X [* + (* — 1 ) COS if] dip.
u = [x + (x 2 — 1)
1/2
cos <£>]", dv = cos <pd<p
11-5 I CONVERGENCE OF LEGENDRE SERIES 433
to prove that
/ (x — 1)
r
cos<p[x +,
(x — 1) cos<p] dip
,
Jo
x 2 2 ,n— 1 ,
f 2 1^1/2
= / w(x
,
— 1N
l)sin
.
r
<p[x +, /
(x — 1) cos<p] a<p.
Jo
(d) Use the results of (a), (b), and (c) to deduce that
1' n
Pn (x) = - [x + (* - I) ' cos <p] d<p.
x Jo
4. (a) Rewrite the formula in Exercise 3(d) as
\Pn(x)\ < 1
/•ir/2 r<f
^«(x) < -
7r
/
Jo
c dip < —
irn l/z u
r^--
Jo
/ e <fr.
,-' 2/2
/
^ < 00
Jo
We now know a great deal about the behavior of polynomials in the Euclidean
space (PC[— the way from some very special properties of the Legendre
1, 1], all
polynomials to the central result which states that the orthogonalized sequence
obtained from 1, x, x 2 ... is a basis for the space. In point of fact, however, the
,
essential portions of earlier discussion remain valid over any finite interval
our
[a, b], and thus, for theoretical purposes at least, we need not consider the ap-
parently more general space <?Q[a, b]. But when we come to study piecewise
continuous functions on the entire real line, (— oo, oo), the situation becomes much
different, and requires special consideration.*
In the first place, if / and g are arbitrary piecewise continuous functions on
(—oo, oo ), there is no guarantee that the improper integral
f—
J oo
f(x)g(x)dx = lim f
a —>» J —a
f(x)g(x)dx (11-24)
b —>»
converges. Indeed, (1 1-24) is undefined even when /and g are polynomials. Thus
our usual definition of an inner product is not valid, and this, in turn, implies that
if we wish to use the general theory of Euclidean spaces in this context we must
either consider another set of functions, or another inner product, or both.
As our point of departure, let us insist that any function space which we consider
contain all polynomials, and that its inner product be defined by means of an
improper integral over the entire real line. This being the case, it is clear that
we must introduce a weight function w into our integral in order to guarantee that
/oo
w(x)p(x)q(x) dx
— oo
exists for every pair of polynomials p, q. (See Example 3, Section 7-1). This
requirement suggests that w be piecewise continuous on the entire real line, and
that it —> oo. Moreover, it is convenient to require that
tend to zero as \x\
n = (Why?) Our experience from
limizi^oo w(x)x any positive integer n.
for
calculus suggests that we try an expo-
nential function (with negative exponent),
and since we allow x to assume negative
as well as positive values, the exponent
should involve only even powers of x.
Thus we are led to try the weight function
w{x) = e~ x /2
. (11-25) FIGURE 11-2
(See Fig. 11-2.)* The following lemma guarantees that this choice will be
successful.
^2n+l = Oj
(2«)!
2n
Io.n. — ./~z
V2x
2"n!
for all n.
2n+1
/2 »+i= lim I* e-* 2l2 x dx.
a—»oo J —
b—>a>
b e~x2
f e-
x2,2
x 2n+1 dx = f— e-
x2 ' 2
x
2n+1
dx+ [* ' 2
x 2n+1 dx
J — J a J a,
b
= x2 ' 2 2n+1
f e- x dx.
Ja
But
b b
2n+1 x2 ' 2 x2 ' 2
f x e~ dx= f x 2 \xe- ) dx
Ja Ja
" rb
= - x2n e - x -l
b „ , 2,
l2
+ 2nf x 2n e~
x l2
dx.
a Ja
Thus
2,„ b ~x
I2n +i = lim \-x
2n
e~
x l2
+ 2n[ x 2n e~ x l2
dx]
a—>x> L a Ja J
b—>oo
2n - x
b
= 2«lim x e~ x2l2 dx
a—>oo J\a
6—>oo
= 2«limrr x 2n - 2n - l
b
1
e-* 2l2 dx+ \ x e-x2 ' 2
dx\
a—>oo L./ — Ja J
2n - x
b
= 2rtlim x e- x2l2 dx,
a—KX Jf—a
b—>oo
* Someauthors set w(x) — e~ x Except for the obvious modifications this necessi-
.
of the results in this section still remain valid. However, the Hermite poly-
tates, all
nomials defined below then do not have 1 as their leading coefficient.
a 1
and it follows that I2n +i = 2nl 2n -i for all n > 0. Moreover, by direct compu-
tation, we find that Ii = 0. Hence I 2 n+\ = for all n > 0, as asserted.
To evaluate I 2n we must use the fact (proved in Exercise 3 below) that
7 = r—
J 00
e- x2/2 dx = -v/2^.
Then if « > 0,
J 00
= limT- x 2n -
a —>oo L
l
e-
x2 ' 2
\
I
h
—
+ (2n - 1)
J
/"*
—a
x 2n
-2
^ 2/2
del
J
6—»oo
= (2rt - 1) /2 „_ 2 .
Henc;
/2 = (2 - 1)V2tt,
(2» - l)(2n - 3) • • •
(2 - 1) = ^ if ii > 0,
From the way things are developing, it is obvious that we stand in imminent
prospect of having to consider various integrals of the form
x2 ' 2
f
J 00
e- f{x)g{x)dx.
This being so, it will be well to dispose of convergence questions once and for all,
which we do by proving
e-
x2/2
f(xfdx < and [" e~ x2l2 g{x) 2 dx < oo.
f—
J 00
oo
J — oo
Then
oo 2
x l2 2
(0 [ e' [af(x)] dx < oo for all real a;
J — 00
(ii)
J
r 00
e'
x2/2
[f(x) + g(x)]
2
dx < *o;
2
let
/°° x ,2
e' f(x)g(x)dx < oo.
— 00
11-6 HERMITE POLYNOMIALS 437
f
J 00
e-
x2,2
[af(x)fdx = a2 T
J 00
e~
x l2
f{x) dx.
2
To prove (ii), we introduce the function max (/, g) which, for each x, is de-
two numbers f(x),
fined to be the larger of the g(x). It is clear that max (f g) is
2 2
[/+s] <[|/l + \g\]
= 4max(/ 2 ,g 2 )
< 4[/
2
+ g
2
].
FIGURE 11-3
Thus
x2 ' 2 2 x2/2 2
f— e- [f{x) + g(x)] dx < 4f e- f(x) dx
j 00 — J 00
x2 ' 2
+ 4f
J— 00
e- g(x?dx.
J
r— 00
e-
x2 ' 2
[f{x)+ g(x)]
2
dx < oo.
Finally, since
2
f(x)g(x)
[f(x) + g(x)] fixY + g(xY
2 2
2
e- x2l2 f(x)g(x)dx = e~ x2l2 [f(x) +
f if g(x)] dx
- if e-
x2/2
f(x) dx
2
- if* e-^/
2
^) 2
^.
7— 00 y — 00
But we now know that all of these integrals are finite, and we are done. |
These technical details out of the way, we consider the set of all piecewise con-
tinuous functions on (—00, 00), identified in accordance with the convention in
Section 9-2, which have the property that
J
r— 00
e-
x2 ' 2 2
f(x) dx < 00. (11-26)
rather than/, appears in (11-26), the superscript recalling that a weight function
is involved, and d standing for "integrable." The functions in d%_ are sometimes
referred to as being "square integrable with respect to the weight function e _x2/2 ."
1
f'g= J
r— oo
e-
x2 ' 2
Kx)g(x)dx. (11-27)
Proof. From Lemma 1 1-6 we know that f -\- g and of belong to #2 whenever
/ and g do, and also that the integral defining / •
g is finite. This disposes of all
but the straightforward details in the proof, and these we leave to the reader. |
P—
J 00
e-
x2/2
f(x) dx
2
< M T— 2
J 00
e-
x2 ' 2
dx = M V^r2
< oo.]
lim 11/- pn \\
= lim ([* e~ x2 ' 2
[f(x) - pn {x)f dx)
112
= 0.
culties, and since Theorem 1 1-1 holds, mutatis mutandis, in this situation, we are
free to seek another basis composed of polynomials, one of each degree. This
time the most convenient basis consists of the so-called Hermite polynomials,
which are defined by the formula
H (x) = 1, H^x) = x,
H 2 (x) = x2 - 1, H 3 (x) = x 3 - 3x,
H 4 (x) = x4 - 6.x
2
+ 3, H 5 (x) = x 5
- lOx
3
+ 15*.
11-6 I HERMITE POLYNOMIALS 439
#2 with
=
~
\\Hn \\* nWlir. (11-29)
Proof. Let
/oo
2
—00
e-' ' 2
Hm (x)Hn (x)dx
/OO
x
Hm (x)^Je- x2 ' 2
)dx,
/oo tn—m 2
J = (-l) 2 V.r e-
x2/2
dx
J — 00
= nXs/li,
One of the principal reasons for the importance of Hermite polynomials is that
they appear as solutions of a certain linear differential equation. To derive this
equation we first establish the following recurrence relation.
Proof. The proof is similar to the one given in Section 1 1-3 for the Legendre
polynomials. Since xHn (x) is a polynomial of degree n + 1, its Hermite series is
of the form
Hn xHk = if k < n — 1
• . Moreover,
xHn • Hn =
J
r— xe-
x2 ' 2
Hn {xf dx
=
since the integrand is an odd function (see Lemma 11-5). Thus (11-31) is of the
form
xHn (x) = H
a n n+1 (x) + /3 n ^n _i(x).
If we equate the coefficients of x n+l in this identity, and recall that the leading
_ xHn • Hn +\ _ j
-"n + 1 *
-"n + 1
But then
a — xHn
_ Hn _\ '
Pn
Hn _ Hn _\
i
•
s/I-k n\
V2tt (n - 1)!
Thus
xHn — Hn+ + i
nHn _i,
as asserted. |
11-6 J
HERMITE POLYNOMIALS 441
To derive the differential equation we start with the defining formula for Hn :
2
£-e-* ' 2
= (-\Te- x2l2Hn (x\
^e-*
dx +
n 1
!* = (-lf[- xHn (x) + m(x)]e-
x2 > 2
.
n+l
d
n +*
-e-**
12
= (-l) n+1 e- x2l2Hn+1 (x).
dx
Thus
Hn+1 = xHn — Hn , (11-32)
Hn — nHn —\ = 0,
or
Hn+1 - {n + \)Hn = 0.
Hn +\ = xHn + Hn — H^,
Hn - xHn +
'
nHn = 0.
y" - x? + ny = 0. (11-33)
As with the Legendre polynomials and their differential equation, it can be shown
that Hn is the only polynomial solution of (11-33) with leading coefficient 1. We
encounter this equation again in Chapter 13 when we discuss the Schrodinger
shall
wave equation from quantum mechanics.
1
EXERCISES
(a) e*
2/4
(b) x sin x (c) |jc| (d) f(x) = n,n<x<n+\
(e) xe*'>
2/8
8 ,rx
(f)
^/.a ln W»
/(*)«'" I 1*1 ^
- 1
. (g) *
[0, |x| < 1
-x ./ — 00
Change to polar coordinates, evaluate this integral, and thus show that / = y/lir.
4. Prove that
iln - 1) (2n - 3) • • •
(2 - 1) = ^. R > 0.
5. Given that Hq(x) = 1, use (11-32) and mathematical induction to prove that
Hn {x) is a polynomial of degree n with leading coefficient 1.
2
6. Prove that the functions 1, x, x , . . . are linearly independent in 3™-
7. (a) Use the recurrence relation to verify that the polynomials listed in the text as
H2, . .
.
, H5 are correct.
for all n.
9. Let #2 denote the set of all piecewise continuous functions /on (— 00 , 00 ), identified
Prove that this defines an inner product on #2, and that with the usual definitions of
(b) Prove that #2 contains all functions of the form e~ x2/4p(x), where p is a
polynomial.
d -x 2 /2 1/4
hn(x) = (-!)* nB e
dx
11. Show that the Hermite functions satisfy the same recurrence relation that the Her-
mite polynomials do.
(see Exercise 9), and suppose that pn and qn are polynomials of degree n. Prove that
for each n, p n is a scalar multiple of q n .
converges. It is then easy to show that ^ is a Euclidean space under the usual
addition, scalar multiplication, and inner product, the last being defined by the
formula
f-g = re-xf(x)g(x)dx.
Jo
Moreover, since
r e~ x dx
Jo
x n
= «!, 7i = 0, 1, 2, ... (11-34)
1
(Exercise 1), $%, contains all polynomials, and hence, as before, we have the fol-
lowing theorem.
From this it follows that any mutually orthogonal set of polynomials in #2>
one of each degree, will be a basis for this space. One such set is the sequence
{L n (x)}, n = 0, 1, 2, ... of Laguerre polynomials, where
,
n
L n {x) = (- 1)V £- (x e- x ). (11-35)
Indeed, when the argument used to establish Theorem 11-16 is adapted to this
situation, we find that
L m L n = 0, m 7* n, •
n~x
= (-l) n [xn -
^ n
{x
n
e-
x
) n x
2
+ • • -]e~
x
(11-37)
(see Exercise 9, Section 1 1-3) where the terms omitted are of degree < n — 1 in x.
The lemma now follows by substituting this expression in (11-35). |
where
_ (xL n ) • Lk
ak " INI
2
But, by Lemma 11-1, (xL n ) - Lk = Ofor all k < n - 1. Hence (11-38) reduces to
(xL n ) • L n +\
1,
and we have
2
(xL n )-L n+1 = [(«+ l)!]
fln-l
[("
- I)!] 2 [(»- I)!] 2
~
2
n\
— -n 2 ,
L(« 1)'J
Starting with the value L (x) = 1 derived from (11-35), the above formula
yields
L (x) = 1,
Li(x) = x — 1,
L 2 (x) = x — 4x + 2,
2
L 3 (x) = x 3 - 9x 2 + 18* - 6,
L 4 (;t) = x 4 - 16x 3 + 72x 2 - 96* + 24,
L 5 (x) = x 5 - 25x 4 200jc
3 -
600jc + 600x -
-(-
2
120.
n+l
d d
dx
Ln (x) = (-l)V^(xV-*) + L n (x)
= (-l)V^K- ^ - 1 n
x e~
x
) + L n {x)
= (-l)W (x*
dx n
1
Similarly,
£ £»-i(*) = (-lr-v^c*-- 1
*-*) + l^x),
and we therefore have
But, by (11-40),
we find that
2
xL'n = nL n + n Ln _ x . (11-42)
2
We now use (11-41) again to deduce that n L'n _ x
2
+ nL'n = n Ln _ and
1 hence,
by (11-42), that
2
n L'n _ x + nL'n = xL'n - nL n . (11-43)
or, by (11-43),
Xl^n i
**, n = -X-^n ^^-"nt
thereby proving
(Exercise 11, Section 15-3), L n is the only polynomial solution of this equation with
leading coefficient 1.
EXERCISES
1. Prove that
x n
/ e~ x dx = n\, n = 0,1,2,
Jo
4. Prove that
= |0
ifm^n,
|(n!)
2
if m = n.
5. Use the differential equation for the Ln to prove the orthogonality of these polyno-
mials in #2 [0> °°). [Hint: Mimic the argument used in the case of the Legendre
polynomials.]
In the following exercises we give the student the opportunity to explore for himself an
interesting variation of the material in this section.
n+a x
L :\x) = (-lfx-V
(
(x e-
J^ ),
8. Prove that
||L^||
2
= n\T(n + a+ 1),
where T(a) denotes the gamma function. (See Exercises 25 and 26, Section 5-7.)
9. (a) Prove that the Laguerre polynomials {L^} satisfy the recurrence relation
which we expand as a power series in r under the assumption that x and r are chosen
so that the resulting series converges. This allows us to write
= jjPn(*> B ,
(H-46)
n=0
where P Pu
, . . . are functions of x alone. These coefficients can be computed
directly from (1 1-45) and (1 1-46) in the usual fashion, and it is not difficult to show
thai: Pn must be a polynomial in x of degree n. For instance, setting r = in
P (x) = 1.
For our present purposes, however, it is sufficient to observe that this series is
uniformly and absolutely convergent whenever \2xr — r
2
< 1, and hence, under \
dG _ n -\
(11-47)
dr f^ nPn {x)r
n=\
But, by (11-45),
dG _ r — x
dr (1 - 2xr + r 2 ) 312
or
(1 - 2xr + /-
2
)^+ (r - x)G = 0.
(1 - 2xr + 2
r ) £ nPnixy- 1
+ (r - x) £P n (x)r
n
= 0,
11-8 GENERATING FUNCTIONS 449
00 00
-1
X
n=l
nPn (x)rn - xP (x) _ £
n=l
(2n + \)xPn {x)r
n
n+1
+P (x)r + X (n + l)Pn (x> = 0,
or
n=0 n=l
+ X ^n_l(x>
w=2
n
= 0.
Thus
+ X K" + O^n+lC*) -
n=2
(2/1 + 1 )*/>„(*) + /lP»_i(*)>
n
= 0,
and we have
Pi - */>o = 0,
But the second of these expressions is none other than the recurrence relation for
the Legendre polynomials, and since P (x) = 1 and Pi(x) = x, we conclude
that the coefficients in (11-46) are, in fact, the Legendre polynomials. Thus the
function
1
G(x, r) =
(1 - 2xr + r2 Y 12
is a generating function for the Legendre polynomials.
In physics, the function G(x, r) is encountered in the study of planetary motion
and electrostatic potential, among other places. In fact, Legendre's original
memoir on this subject, published in 1785,
was devoted to the study of the gravitational
attraction of "spheroids," and it is only
fitting that we conclude our discussion of the
Q of the plane arising from a pair of point charges of magnitudes +<r and — a
located as shown in Fig. 11-4. Since the potential at Q due to a single point
cha rge a located R units away is <j/R, the potential at Q in this case is
r2 = r
2
+ x2 - 2rjccos0,
R'
2
r
2
+ x 2
— 2rx cos (x — 0),
and
-1/2
-(2^cos(tt-0)-^)
R'
Save for notation, each of these expressions is the generating function for the
Legendre polynomials, and we therefore have
R =l±Pn(cOSe)(£f
n=0
and
^7 = \ £
n=0
P»(C0S (r - »))
(f)"
X *
- \ £ P„(-COS
n =0
0)
(ff.
|n=0
Finally, recalling that Pn is even if n is even, and odd if n is odd, the above ex-
pression reduces to
go
/v\ 2n + 1
V = vS^n
n=0
+ l(cOS0)^j
•
xr - r2 2
G(x,r) = e
'
. (11-48)
Indeed, since the series
e
z
= 1 + z + |j- + • • •
2
G(x,,)=l + (xr-0 + l(,r-0 + ...
= 1 + xr + ^(;c
2
- 1>
2
+ •••
for all x and all r. Differentiating this series term-by-term we find that
dG _
= y\
00
Hn+1 (x) ^n
dr ^
n=0 n\
dr
- W r)
2^ n y
r
n=0
= xH xH"(x) Hn -^x>>
(x) + X) -f r".
w=l
Thus
71 =1 '
n= l
and it follows that
#! - *7/ = 0,
as required.
n=l
Moreover, by (11-49),
V-r?~={l-x-r)G,
dr
- rf £ (-l)^n(x) -^y + 7
(x - 1 + r) £ (-!)"«*)£ = 0.
'(11-51)
EXERCISES
1. Let
G(x,r) =
(1 - 2xr + r2)i/2
Prove that
(a) (1 - 2xr + r
2
)—
ox
- rG^O;
... 3G
, dG_
(b)(x-r)---r--=0;
<c)r£(^-(l-r*)f -0.
2 (a) Substitute the series
00
into the identities in 1(b) and 1(c), and deduce in turn that
and
(« + l)Pn (x) - P'n+1 (x) + xP'n (x) = 0.
3. Use the identities in the preceding exercise to derive the differential equation for Pn .
[Hint: Replace n by n — 1 in the second identity in 2(a) and then substitute the value
of P'n -\ found in 2(b). Differentiate, and again eliminate P'n -\.\
G (x n = /,
>
n=0
— r" r 3
prove that
Hn(x) = (-1) e
2
— Jn
e
2
~ r2/2 = 2/2
[Hint: e xr e x2/2 e-^-^ .]
5. (a) With G as in Exercise 4, compute dG/dr, and show that Hi(x) = nHn -\(x).
(b) Use the result in (a) and the recurrence formula for the Hn to obtain the dif-
ferential equation for Hn .
6. Show that the recurrence relation for the Laguerre polynomials follows from (1 1-51).
r-( \ = * — rz/(l— r)
G(x, r) e
1 — r
(1 - r)--+ rG=0,
ox
and
Kl-r)^-(*-l+r)^0.
or dx
oo n
xy" + (1 - x)y' + ny = 0.
1
Ln (x) = (-1) e — (x e ),
show that
n 2 n—k
X
Ln(x) = X)^ 1 ^
fc=0
|_(* - k)\\ k\
S(-i)U)^
and show that
n=0
a)
isa generating function for the Laguerre polynomials I4 . (See Exercises 6 through
10 of the preceding section.)
The following table summarizes the results obtained in this chapter. For con-
venience of reference we have also included the corresponding information for
Bessel functions (see Chapter 15).
LEGENDRE POLYNOMIALS
Definition
* Recurrence P_l = 0, P = 1
Differential
(1 _ x 2)y » - 2xy' + n(n + l)y =
equation
Generating 1
Orthogonal in (PC[-1, 1]
o, m 9^ n
Orthogonality
"m ' "n
» m — n
12/1 + 1
* The index —1 is used as a subscript only when working with recurrence relations.
11-8 |
GENERATING FUNCTIONS 455
HERMITE POLYNOMIALS
tt / \ / t\ n x /2
2 A""
a —X 2 /2
Definition
* Recurrence /f_! = 0, Ho = l
Generating pxr—r 12
function
-* /2
Orthogonal in $%(— °° , °°); weight function e
Orthogonality
«m "n — * <
l\/27r w!, m = n
LAGUERRE POLYNOMIALS L n
Definition f -t\ n x
Lt n /(x)\ = (-1) e —
a f
(x e
n ~x \
)
* Recurrence L_i = 0, Lo = l
Generating l — n/(l— r)
function l -r C
Orthogonal in g%[0, °°); weight function e~*
Orthogonality r . r _ J°>
m 9± n
l(n!) , m — n
* The index —l is used as a subscript only when working with recurrence relations.
1
LAGUERRE POLYNOMIALS L^
* Recurrence L^\= 0, L ( a)
= l
relation
4ti + (2« + a + l - *)4a) + n(n + a)L^! s 0, n >
Differential
equation
*/' + (a + l - *)/ + ny =
Generating l — rz/U— r)
c
function a - r)«+i
\n\T(n + a + l), m = n
BESSEL FUNCTIONS
Recurrence
relation
xJp+ i — 2pJp -+- xJp -i = 0, p an arbitrary real number
Eifferential
jcV + jc/ + (x 2 - p 2 )y =
equation
Crenerating
e (x/2)«-l/0
function
* The index — l is used as a subscript only when working with recurrence relations.
12
where the a;, &, and 7 t are constants. The problem, of course, is to find all func-
2
tions y in Q [a, b] which simultaneously satisfy (12-1) and (12-2).*
For instance, the equation
y" + y = (12-3)
is a problem of this type on the interval [0, ir]. To solve it we simply apply the
boundary conditions to the general solution Cisinx + c 2 cos x of (12-3) to
y = c sin x,
ferential equations.
In order to exclude certain trivial cases from the following discussion, we
demand that at least one of the a; and one of the fa appearing in (12-2) be dif-
ferentfrom zero, and that the left-hand sides of these equations be linearly inde-
pendent in the sense that they are not constant multiples of one another. Further-
more, to ensure that we are actually looking at a boundary-value problem, and not
an initial-value problem, we also require that (12-2) contain nonzero terms in-
volving each of the endpoints of the interval. Finally, we shall say that the given
boundary conditions are homogeneous whenever y 1 = 7 2 = 0. In this case the
set of twice continuously differentiable functions on [a, b] which satisfy (12-2)
2
is a subspace S of Q [a, b], and by viewing L as a linear transformation from S to
e[a, b] the problem in question becomes one of solving a certain operator equation
to wit, the equation Ly = h, where h is a known function in e[a, b], and
L:S — e[a, b]. Despite its simplicity this observation is important because it shows
thai: the boundary conditions influence the problem only to the extent of deter-
mining the domain space for L. Strictly speaking, some symbol other than "L"
ought to be used to represent the operator from S to Q[a, b], since heretofore we
2
have considered L as acting on all of e [a, b]. (Recall that operators with dif-
ferent domains are different, even though the operators themselves are defined
by the same formula.) However, such changes in notation would be as confusing
as they are correct, and will therefore be avoided. But by the same token, it then
becomes mandatory to specify the domain of the operator being considered when-
ever there is any possibility of confusion.
At this point we invoke a familiar argument to reduce the study of boundary-
value problems with nonhomogeneous boundary conditions to the homogeneous
case (see Section 2-9 and Exercise 8 below). Hence, unless otherwise stated, we
shall assume from now on that all boundary conditions imposed are homogeneous.
The solutions of a boundary-value problem involving a linear differential
operator L: S — e[a, b] are intimately related to the solutions of the equation
Ly = \y, (12-5)
where X isan unknown parameter. In this setting we are required to find all values
of X for which (12-5) admits nontrivial solutions in S, and then find the solutions
12-1 | DEFINITIONS AND EXAMPLES 459
corresponding to these X. The reader should note that (12-5) may be rewritten
(L - XI)y = 0, (12-6)
where / denotes the identity transformation which sends each function in e[a, b]
onto itself, or as
a 2 (x)/' + a^x)y' + [a (*) - X]y =
if L =
a 2 (x)D
2
ai(x)D + +
a (x). Thus, for each value of X, (12-5) is a homo-
geneous second-order linear differential equation, and the solution set of any
(homogeneous) boundary-value problem involving this equation is the null space,
in S, of the operator L — XI.
Lx = Xx
has nontrivial solutions; then find all solutions corresponding to these values
of\.
And this is the problem with which we shall begin our investigations. But first,
an example.
y y '
(12-7)
X0) = 0, jM = 0.
Here S is the space of all twice continuously differentiable functions on [0, ir]
which vanish at the endpoints of the interval, and L is the second-order linear
differential operator — D
2
(The minus sign has been introduced merely to sim-
.
plify the final results. Without it the relevant values of X would be negative.)
We distinguish three cases, according as X = 0, X < 0, X > 0.*
_ ^ _Xx
Case 2: X < 0. In this case y = cie^~ Xx + c 2e , and the boundary con-
ditions again yield y = 0.
* At this point we are tacitly assuming that X must be real. This assumption will be
justified later.
. ,
c2 = 0, C\ sin \/\ ir = 0.
sinVXir = 0;
more, for each of these values of X the constant c x in (12-8) remains arbitrary,
and it follows that the solution space corresponding to X n is the one-dimensional
subspace of C[0, x] spanned by the function sin nx.
The numbers \ n n 2 which determine the cases in which (12-7) has non-
=
trivial solutions are called the eigenvalues for this problem, and each nontrivial
solution corresponding to the eigenvalue \ n is called an eigenvector or eigen-
function belonging to X n . In the next section this terminology will be generalized
to i nclude a much wider class of problems, and for the moment we merely ask
the reader to note that any set of eigenfunctions, such as sin x, sin 2x, sin 3x, . . .
one for each eigenvalue, is orthogonal in e[0, ir]. This, as we shall see, is no accident,
and when properly generalized will be of fundamental importance in the study of
boundary-value problems.
EXIERCISES
5)
y" + 9y = 0; /(0) = 0, /(*) =
Bi = 7i, B2 = T 2 ,
be a boundary-value problem of the type described by Eqs. (12-1) and (12-2) above.
Prove that the solution set of this problem consists of all functions in Q 2 [a, b] of the
form y p +
va, where y p is a fixed solution of the problem, and y h a solution of the
12-2 I
EIGENVALUES AND EIGENVECTORS 461
Ly = 0,
Bi = B2 = 0.
9. Find all eigenvalues and eigenfunctions for the boundary- value problem
y" + Xy = 0,
y" + 4/ + (4 + 9\)y = 0,
HO) = 0, y(a) = 0,
with X real.
(a) Show that this problem has no nontrivial solutions for A < 0. [Hint: Consider
the cases X < and X = separately.]
(b) Show that the only positive values of X for which this problem has nontrivial
solutions are
2 2
9a 2
Definition 12-1. The values of X for which Eq. (12-9) has nonzero solu-
tions are called the eigenvalues (or characteristic values) of L, and for each
eigenvalue X the nonzero vectors in S which satisfy the equation Lx = X x
are called the eigenvectors (or characteristic vectors) of L belonging to X .
462 BOUNDARY- VALUE PROBLEMS CHAP. 12
The reader should observe by definition, the zero vector is never an eigen-
that,
vector for L. Furthermore, zero an eigenvalue for L if and only if the equation
is
In other words, the zero vector together with the eigenvectors for L which belong
to > constitute a subspace of S (and hence, by implication, of V as well). We
shall denote this subspace by S Xo and observe in passing that dim S Xo > 1 for
,
FIGURE 12-1
We have just seen that Lx belongs to S Xo for all x in S Xo . This fact is some-
times expressed by saying that S Xo is "invariant" under L in accordance with the
following definition.
We hasten to point out that there is nothing in this definition to imply that the
nonzero vectors in an invariant subspace for L need be eigenvectors for L. Indeed,
as we shall see momentarily, such a conclusion is false. Rather, the implication
goes the other way: the S Xo are invariant subspaces for L consisting of vectors
with the special property that Lx = X x.
Having introduced the notion of invariant subspace, we can now rephrase
the definition of an eigenvector for a linear transformation L: $ — > V to read as
follows: A nonzero vector x in S is an eigenvector for L if and only if the one-di-
mensional subspace of S spanned by x is invariant under L. This observation is
frequently useful in the search for eigenvectors.
12-2 |
EIGENVALUES AND EIGENVECTORS 463
Lei = ei, Le 2 = — 2.
Then, from geometric considerations alone, it is clear that the only subspaces of
2
(ft
2
which are invariant under L are (i) the trivial subspace, (ii) (ft itself, and (iii)
the two one-dimensional subspaces spanned by e x and e 2 By the remark made a .
moment ago, the last two must be eigenspaces for L, and are the only such. Further,
it is obvious from the definition of L that these subspaces are associated with the
eigenvalues 1 and —1, respectively.
Example 2. If L: (ft
2 — » (ft
2
is reflection across the origin, then every subspace
of (ft
2
is invariant under L. In this case Lx = — x for all x, and it follows that — 1
2
is the only eigenvalue for L, and that S_i = (ft . The reader should note that
here the invariant subspace associated with the eigenvalue is fwo-dimensional.
2
Example 3. Let L be a rotation of (ft about the origin through an angle 0.
Then, if is not an integral multiple of w, there are no one-dimensional invariant
subspaces, and L has no eigenvectors at all.
differentiable functions y such that y(0) = y(r) = 0, and let L: S —» e[0, ir] be
the operator — D2 . Then, by the example in the previous section, L has an in-
finite sequence of eigenvalues
2
1,4, 9,..., « ,...
All this is simple enough, but hardly explains why these notions were introduced
in the first place. The following theorem furnishes a partial answer to this question,
and gives an indication of the importance of eigenvectors in the study of linear
transformations.
(Note that this result would fail if were allowed to be an eigenvector. Thus
the prejudice against the zero vector found in Definition 12-1.)
Proof. The theorem is obviously true when applied to a single eigenvector. Beyond
this we reason by induction, as follows.
Assume that the theorem has been proved for every set of n — 1 eigenvectors
for L, n > 1, let Xi, . . . , x„ be « eigenvectors belonging, respectively, to distinct
eigenvalues X l5 . . . , X n and
, let
aiXi + • • •
+ a n _iX n _i + (XnXn = 0. (12-10)
464 BOUNDARY- VALUE PROBLEMS | CHAP. 12
a i Lx 1 + *
-
'
+ OLn—lLXn-X + ct n Lx n = 0,
or
«iOiXi) + • • •
+ a n -iOn-iXn-i) + ot n (\ n Xn) = 0. (12-11)
We now multiply (12-10) by X n and subtract the resulting equation from (12-11)
to obtain
ai(\i — X n )xi + • • •
+ a w _i(X n _i — X n )xn _i = 0.
EXERCISES
1. Every vector space V has at least two subspaces which are invariant under a linear
transformation L: V —> 1). What are they?
X = Xi + X2
4. (a) Show that the null space of a linear transformation L : V — V is invariant under L.
>•
Piy(b) + 02/0) = 0,
Ly = (py')' + gy,
12-3 |
EIGENVECTORS IN FINITE DIMENSIONAL SPACES 465
vanishes.
Xi
An -
Of course, such bases need not exist for a given L: V —> V (see Examples 2 and
3 above). When they do,
however, a number of pleasant things happen. For one,
we can then solve operator equations involving L; and rather efficiently too. The
following example will illustrate the technique.
3
Example 1. Let L be a linear transformation mapping 01 into itself, and suppose
that L has distinct eigenvalues Let ei, e 2 e 3 be eigenvectors belonging
Xi, X 2 , A3. ,
Lx = y, (12-12)
3
y known, x unknown. Then, since the vectors ei, e 2 , e 3 are a basis for (R , we have
x = *xei + x 2e 2 + *3e3,
LOid + x 2e 2 + x 3 e 3) = y^ + y 2e 2 + y&z-
Hefrce
Ai A2 A3
whenever the \ are different from zero. If, on the other hand, one of the X»-, say
Xi, is zero, (12-12) has no solutions at y x = 0. In the latter case the
all unless
equation x{K x = y 1 is satisfied for all values of x u and the solution set of (12-12)
then consists of all vectors of the form
x = xiei + ^ e + ^ e 2 3,
A 2 A3
with xi arbitrary.
The generalization of these results to w-dimensional spaces is obvious, and has
been left to the reader.
«ni*i "+" oi n2 x 2 + • • •
+ oc nn xn — yn ,
where
«11 «12 "In"
«21 «22 «2n
«wl a n2
12-3 I
EIGENVECTORS IN FINITE DIMENSIONAL SPACES 467
all with respect to the chosen basis. In particular, this is true of the equation
Lx = Xx,
(L - X/)x = 0,
«21 a 22 — X «2n
«nl «n2 - X
we therefore conclude that the eigenvalues for L are simply the values of X for
which the system of homogeneous equations
has nontrivial solutions. But this will occur if and only if the determinant of the
coefficients of (12-13) vanishes (see Appendix III), i.e., if and only if
an — X a 12
Thus the eigenvalues for L can be computed by solving (12-14) for X, and since
the left-hand side of this equation is an nth degree polynomial in X, this can even
be done by the methods of elementary algebra (at least for small values of n).
The polynomial appearing in (12-14) is known as the characteristic polynomial of
the linear transformation L, and the equation itself is called the characteristic
equation of L. As its name suggests, the characteristic polynomial is independent
of the particular basis used to compute it a fact which is also proved in Ap- ;
pendix III.
46H BOUNDARY- VALUE PROBLEMS I CHAP. 12
Example 2. Find the eigenvalues and eigenvectors for the linear transformation
2 2
L: '
31
v
(R ,
given that the matrix of L with respect to the standard basis e it e 2 is
2 1
1 - X
= 0,
1 - X
or
(1 - X)
2 _ o,
and it follows that X = 1 is the only eigenvalue for L. Hence a nonzero vector
x .Xxei + x 2e 2 will be an eigenvector for L if and only if Lx = x. Rewriting
this equation in matrix form as
2 1
.
x 2.
Xi = Xi,
2xi -\- x2 = x2 .
Thus xi must be zero, while x 2 is arbitrary, and the eigenvectors for L are of the
form x 2 e 2 x 2 t^ 0. Finally, the eigenspace is the one-dimensional subspace of
,
spanned by e 2 .
-1
2 2 1
-X 1
Lx = Xx (12-15)
when X = 2 and X = — 1.
In the first case, (12-15) becomes
l" Xl 2*1
-1 x2 = 2x 2
2 2 1 -*3_ 2x 3
and we have
x3 = 2a: i,
—x2 = 2x 2 ,
cos 6 — sin 6
sin 6 cos 6
cos 6 — X —sin 6
= 0,
sin e cos 6 — X
and we have
X
2
- 2(cos 0)X + 1 = 0.
Thus
X = cos 6 ± / sin 6,
and it follows that Xand only if 6 = nir. Moreover, when this is the case,
is real if
X assumes one of the values ± 1, and has all of (R 2 as its invariant subspace.
Otherwise, X is complex, and L has no eigenvectors.
470 BOUNDARY- VALUE PROBLEMS |
CHAP. 12
EXERCISES
Find 2
1. all eigenvalues and eigenvectors for the linear transformations on (R defined
by the following matrices.
2 1 1 1 1
3
2. Find all eigenvalues for the linear transformations on (R defined by the following
matrices, and in each case find the eigenvectors belonging to the real eigenvalues.
(a) 2 1 (b) 1 2
-1 2 3 1 1
1 2 1 1
(c) -1 1 (d) 2
1 2 2 -1
-1 1 1
3. Exercise 2 for
(a) 1 1 (b) 1 1
1 2
1 1 1 -1
(c) 5 -6 -6 (d) 1 2 1
-1 4 2 1 2 1
3 -6 -4 1 2
4 defined
Find the eigenvalues and eigenvectors for the linear transformations on (R
by the matrices
(a) 1 -1 (b) -1 2 1 3
1 1 -2 1
-1 1 2 1 2 -3
1 -1 4
Let L: (ft
3 —
> (ft
3
be the linear transformation whose matrix with respect to the
standard basis ei, e2, e3 is
1 2
2 1_
12-4 | SYMMETRIC LINEAR TRANSFORMATIONS 471
Use the eigenvalue method to solve the equation Lx = y for x, given that
2 r
1 3
.0 2
and
(a) y = -2ei + e 2 ;
(b) y = 4ei + 4e 2 + 2e 3 ;
(c) y = ei + 9e 2 + 2e 3 .
"9. (a) Let L be a linear transformation on a finite dimensional vector space V, and
let Xo be a real eigenvalue for L of multiplicity m, by which we mean that (X — Xo) TO
is a factor of the characteristic polynomial for L. Prove that the dimension of the
invariant subspace associated with Xo is a? most m. [Hint: Consider the character-
istic polynomial of the linear transformation obtained by restricting L to the sub-
space S Xo .]
(b) Give an example to show that this dimension can, in certain cases, be less than
m. [Hint: Consider the operator — D on the space of polynomials (P„.]
Before giving any examples, we prove that this definition accomplishes our ob-
jective by establishing
Proof Let Xi and x 2 be nonzero vectors in S, and suppose that Lxi = XiXi,
Lx 2 = X 2 x 2 with Xi
, X 2 Then ^ .
(Z,X!)-x 2 = Xi(xi -x 2 ),
Xi • (Lx 2 ) = X 2 (xi • x 2 ),
(Xi - X 2 )(xi -x 2 ) = 0.
basis in V. Then if
Ley = aiyei + • * *
+ oc nj e n , j = 1, . . . , n,
we have
e* • (Ley) = e{ •
(aiyei + • • •
+ anjen )
= aij{ei • e0 + • • •
+ ciij(ei • e t-) + • • •
+ a n ;(ej • en)
JO, i*j,
}, i = J-
(Le {) • ey = (ai,ei + • •
+ ot ni e n ) • ey
= aii(ei • ey) + • • •
+ aji(ej • ey) + • • •
+ a ni (e n • ey)
= OLji,
and the equality (Le t ) ey = e» (Ley) implies that a»y = ay; for all i and all j.
• •
L*j = aiyei + • • •
+ a nj e n , j = 1, ,n.
Then
(Le^) • ey = aji = ctij = e; • (Ley)
for all / and j, and hence if x and y are arbitrary vectors in V with
x = Xiei +
y = J>iei + ~~T~ yn$ni
we have
(Lx) •
y = ( J2 XiLti \-\Yj yjtj
J=l
n n
i=i y=i
= 22 *^ e* • ( Le ^
= [Y^XieA-lj^yjLej
\i=i j \j=\
= x • (Ly).
3 3
Example 2. Let L: (R -> (R be defined by the matrix
1 f
1
1 1
with respect to the standard basis ei, e 2 e 3 Then, by the above theorem, L is a , .
symmetric linear transformation, and an easy computation reveals that its charac-
teristic equation is
X(X - 1)(X - 2) = 0.
474 BOUNDARY-VALUE PROBLEMS CHAP. 12
Hence the eigenvalues for L are 0, 1, 2, and Theorems 12-2 and 12-3 imply that
3
any complete set of eigenvectors for L will be an orthogonal basis for (R a fact ;
which can be readily verified by direct computation. Finally, we note that the
matrix of L with respect to such a basis assumes the diagonal form
o"
2_
and
and L is symmetric onHere again Theorem 12-3 applies, and we can assert
S.
that any complete set of eigenvectors for L is an orthogonal set in e[0, ir]. This
agrees with the results obtained in Section 12-1 where we found that the eigen-
values for L are the integers X n = n
2
n = 1, 2, and that the corresponding
, . . . ,
eigenvectors (or eigenfunctions as they are called in this case) are cn sin nx, c n an
arbitrary nonzero constant.
The reader may have noticed that all of the eigenvalues for the linear trans-
formations considered in the last two examples were real. This, as it turns out, is
always true of symmetric linear transformations; a fact which we state formally as
EXERCISES
2
1. Let L be the symmetric linear transformation on (R defined by the matrix
a b
b c
2. Let Li and Li be symmetric linear transformations mapping 13 into itself. Prove that
the transformationL1L2 is symmetric if and only if L1L2 = L2L1.
3. Let S denote the subspace of 6[0, ir] consisting of all twice continuously differentiable
functions v such that y'(Q) = y'(ir) = 0, and let L:$ —> G[0, 7r] be the operator 2
—D .
4. Determine whether or not the following linear transformations are symmetric on the
Euclidean space of polynomials with inner product p- q = f_ 1 p(x)q(x) dx.
(a) Lp(x) = xp(x); (b) Lp(x) = p'(x);
p(x) ~ x)
(c) Lp(x) = p{x + 1) - p(x); (d) Lp(x) - ~^ {
.
5. Let £F denote the set of all real- valued functions of an integer variable k,
— 00 < k < 00 , and let A 2 J —> : ff be defined by
'6. Let JFo denote the space of all real- valued functions defined on the finite set of integers
0,1, ... ,N,N + 1, and for each pair of functions F, G in Jo, set
iV+l
F>G = j^F^Gik).
k=0
where p any function in Q x [a, b] such that p(x) > 0, or p(x) < 0, for
is
all x in the open interval (a, b), and q is an arbitrary function in e[a, b].
and
= a °( X ) hai(x)la 2 (x)]dx
n(x) c
a 2 (x)
(see Section 6-3 and Exercise 1 below). Thus, without any real loss of generality,
we can (and shall) restrict ourselves to the study of self-adjoint operators, and to
differential equations of the form
Finally, we note that the function p appearing as the leading coefficient in this
equation is allowed to vanish at the endpoints of [a, b~\. This fact will be of some
importance later.
Our immediate objective is to determine conditions under which a self-adjoint
operator will be symmetric when viewed as a linear transformation from S to
2
Q[a, b], S a subspace of Q [a, b] determined by a pair of homogeneous boundary
conditions
ai y(a) + a 2 y(b) + a 3y'(a) + a 4/(b) = 0,
(12-19)
Piy(a) + P 2 y(b) + 0a/(a) + /3 4/(fc) = 0.
L = D(pix)D) + <?(*)
is any self-adjoint linear differential operator on [a, b], and if yi and y 2 are
twice differentiable on [a, b], then
= yiipyW - y2(pyi)'
= yiipyV + p'yfi - y&py" + p'yi]
= p'\yiy 2 - j>2j>i] + Piyiy'2 - wi']
r
= [piyiy'2 - y2y'i)Y- 1
Formula (12-20) can be written in much more suggestive form by integrating
from a to b. For then its left-hand side becomes >>i (Ly 2 ) — (Ly{) y 2 and we • •
,
therefore have
Before giving any examples, we use Theorem 12-6 to determine several of the
more obvious and important boundary conditions which lead to symmetric
operators.
+
«iX«) « 2 /(a) = 0,
(12-23^
Mb) + /3 2 /(Z>) = 0,
K >
478 BOUNDARY- VALUE PROBLEMS | CHAP. 12
to force at least one of the as and one of the |8's to be different from zero.) Then,
if yi and y 2 are any two functions in S,
which involves the values of y and / at only a single point is said to be unmixed.
In these terms the above argument asserts that a self-adjoint linear differential
2
operator is symmetric on every subspace ofe [a, b] described by a pair of unmixed
boundary conditions. (Note that only one such condition need be given if p vanishes
at a or at b.)
Case 3. Assume that p(a) = p(b), and let S be the subspace of e 2 [a, b] consist-
Then (12-22) is obviously satisfied for all y x and y 2 in S, and L is again symmetric.
This is known as the case of periodic boundary conditions.
2
Example Let S be the subspace of e [0, *] consisting of
1. all functions y satisfy-
X0) = yi/) = 0,
2
e [0, 2x] described by the periodic boundary conditions
To find its eigenvalues and eigenfunctions we again apply the given boundary
conditions to the general solution of
y" + \y = 0. (12-26)
When X < 0,
y = c x e^-
Xx
+ c2 e~^ x
,
which can be satisfied with y ^ by setting y/\ = 1, 2, 3, ... . Thus the integers
i2 n2 t2
1 5 *- 5 J j • • •
2
are eigenvalues, and the invariant subspace associated with n is the two-dimen-
sional subspace of e[0, 2ir] spanned by the functions sin nx and cos nx.
d
dx (/>(*);|)
+ fo(*)- *J> = 0,
are orthogonal.* In the next chapter we shall see that such systems arise naturally
in the study of boundary-value problems involving partial differential equations,
and for this reason are extremely important in physics and applied mathematics.
EXERCISES
= / [o i (l)/a 2 (l)1<il
p(x) e
Q\X)
_ ao(x)
J[ ai (x)/a 2 (x)]dx
t \
az(x)
(a) D2 + - D+ >
1, x
x
(b) (cosx)D 2 + (sinx)D - 1, -tt/2 < x < ir/2
(c) x 2 D 2 + xD + (x 2 — p 2 ), x > 0, a real number /?
(d) (1 — x 2
)D 2 —
2xD + n(n + 1), — 1 < * < 1, n a non-negative integer
y" + Xv = 0,
/(-tt) = 0, /Or) = 0.
/' + \y = 0,
y(0) = 0, y'(T) = 0.
/' + Xy = 0,
j(0) = 0, y(L) = 0.
*,2 2
"n = r 2 » ^ = *J ^J • • • J
12-6 |
FURTHER EXAMPLES 481
and functions
yn (x) = sin — > n = 1,2, . . .,
/' + \y = 0,
/(0) = 0, /(L) = 0.
A computation similar in all respects to the one used to solve Example 1 reveals
that the eigenvalues for this problem are the non-negative constants
M2 2
X„ = 2 ' n = 0,1,2, ... ,
and that
/' + = \y 0,
y(0) = 0,
hy(L) + y\L) = 0,
given that h andL are positive constants. (Note that the boundary conditions are
unmixed, and that the problem falls under Case 2 above.)
As usual, we argue by cases, depending upon the algebraic sign of X, and again
find that there are no eigenvalues < 0. On the other hand, when X > 0,
y = Ci sin VX x + c 2 cos Vx x,
and the first boundary condition implies that c2 = 0. Thus it remains (if possible)
to choose X so that the function
y = Ci sin \/x x,
tan fx = - -^ /x (12-28)
482 BOUNDARY-VALUE PROBLEMS CHAP. 12
FIGURE 12-2
its solutions can be visualized as arising, via (12-28), from the points of inter-
section of the graphs of the functions tan /x and —/x/hL. As indicated in Fig.
12-2, there are infinitely many such points
located symmetrically across the origin. Thus the given problem has an infinite
2
^n ==
Y2 ' n = 1,2, . .
.
with X n < X n+ i for all n, and lim^* X n = 00. [Also note that from the geom-
etry of the situation we have lim n _+oo (X n+X - X n) = x.] Finally, the functions
HnX
y n (x) = sin
= sin a/\^ x, n = 1, 2, . .
.
A o-* >i 2
+ x^ = (12-29)
dx
on the interval [—1, 1].
The leading coefficient of this equation vanishes at the endpoints of the interval,
and hence, by Case 1 above, the operator D[{\ — x )D] is symmetric on all of
2
2
e [— 1, 1]. Moreover, since (12-29) is the self-adjoint form of Legendre's equation
(of order X), our earlier results imply that the integers n{n + \),n = 0, 1, 2, ... ,
* This example should be omitted by anyone who is not familiar with the material in
Sections 11-2 through 11-4.
.
12-6 I
FURTHER EXAMPLES 483
we know, the Legendre polynomials are a basis for e[— 1, 1], and hence, by
Lemma 8-3, y = 0. Since this cannot be, no such eigenvalue exists.
EXERCISES
1. Verify that the eigenvalues and eigenvectors listed in Example 2 above are correct.
2. (a) Show that the boundary-value problem consisting of the fourth-order differential
equation
,4
dy 2 n
cos vw = 7=
*
cosh Vco
(b) Use the technique introduced in Example 3 above to prove that the boundary-
value problem in (a) has infinitely many non-negative eigenvalues w„, n = 0, 1,
2, . . . . How do these eigenvalues behave as n — » °o ?
(c) What is the general solution of the boundary-value problem in (a) correspond-
ing to the eigenvalue co n ?
3. Let L denote the fourth-order linear differential operator D4 , and let S denote the
subspace of Q 4 [a, b] consisting of all functions y such that
(b) Use the result in (a) to prove that eigenfunctions belonging to distinct eigen-
values for the boundary-value problem L: S — 6 [a, b]
> are orthogonal.
* Note that at this point Theorem 12-6 allows us to assert, without further proof, that
the Legendre polynomials are mutually orthogonal in (PC [—1, 1].
484 BOUNDARY- VALUE PROBLEMS | CHAP. 12
Let
^0j ^1> ^2> • • •
be a complete set of eigenfunctions belonging to the X n . Then, since the & are a
basis for <3[a, b], we have
00
Kx) = 2
n=0
C«Vn(*),
where
h '
<p n _ }a h(x)<pn (x) d.X
'
2
IWI ja MX)] 2 dx
and the series converges in the mean to h. We now set
00
y(X ) = 2 "nVnOO,
n=0
(12-31)
^
n—0
oc n <p n (x) = ^2
n=0
cn <pn (x).
(recall that L<pn = X n <p n), and it follows that (12-31) will be a solution of the
given equation whenever
(i) the a n can be chosen so that
00
^2 (*n<Pn(x)
n=0
2
defines a function in <5 [a, b] whose first two derivatives can be computed by
termwise differentiation.
It is clear that the first of these requirements can be met by setting a n = cn / \ n
so long as X n ^ for all n (i.e., so long as L
one-to-one). Furthermore, the
is
resulting solution is then unique. If, on the other hand, one of the X n say X is , ,
zero, the problem has no solution at all when c ^ 0, and an infinite number of
solutions when c = 0.
Unfortunately, no such simple analysis can be used to dispose of (ii), since here
we must investigate the convergence of the series
n=0 Aw
that the resulting series has the required properties. This method will be illustrated
in some detail in the next chapter.
2
Xn = n , (Pn(x) = sin nx,
n = 1,2,..., and the <pn are a basis for e[0, *]. (See Section 9-5.) Hence the
boundary-value problem
y
' (12-32)
X0 = >>« = o
486 BOUNDARY- VALUE PROBLEMS |
CHAP. 12
n=\
with
cn = ~
TT
I
Jo
Kx ) si n nx dx.
-/' = x,
(12-34)
X0) = y(T) = 0,
and we have
00
where
r = - / x sin nx dx.
ir J o
whence
Of course, (12-34) can also be solved in closed form by applying the given
boundary conditions to the general solution of /' = x. Lest the reader feel —
that we have been somewhat dishonest in using Fourier series when this easier
method was at hand we point out that frequently no such option exists, and the
only available solutions are those expressed as series in terms of eigenfunctions
for the problem.
* The reader should note that (12-33) will satisfy the given boundary conditions and
reduce to the value of h at and ir only if /r(0) = h(ir) = 0. Thus, in general, we can
neither demand nor expect the solution of (12-32) to satisfy the differential equation on
the closed interval [0, x].
12-7 I
BOUNDARY- VALUE PROBLEMS AND SERIES EXPANSIONS 487
The technique used in the above example was successful precisely because the
operator — D had a sufficient number of mutually orthogonal eigenfunctions
2
and
M< |Xi| < M< ,
lim |X n l = oc.
Moreover, the invariant subspaces of Q[a, b] associated with the \ n are all
one-dimensional; any complete set of eigenfunctions for L, one for each
eigenvalue, is a basis for Q[a, b]; and the series expansion of any piecewise
smooth function y on [a, b] relative to such a basis converges uniformly and
absolutely to y on any closed subinterval in which y is continuous*
EXERCISES
Find the formal series expansion of the solution of the boundary-value problems in
Exercises 1-6 in terms of the eigenfunctions for the associated Sturm-Liouville system.
1. y" = x(x - 2x), y(0) = 0, /(*) =
2. /' = X 2 - 7T 2 /(0) = 0, jKx) =
,
3. /' = sin
Try
—
/(0) = 0, /(L) =
,
For a proof see E. L. Ince, Ordinary Differential Equations, Dover, New York, 1956.
488 BOUNDARY- VALUE PROBLEMS |
CHAP. 12
T
f h(x) dx = and f h(x) dx ^
Jo Jo
separately.]
*8. By citing appropriate theorems in the text, verify the assertion made above concern-
ing the convergence of the series given in (12-33).
^(/>Mg + ?M-M*)]y [
= (.2-35)
defined on an and
interval [a, b],
(ii) a pair of homogeneous boundary conditions which serve to determine the
domain space for the operator
L = D(p(x)D) + q(x).
Lyi = Air(*)yi(*),
Ly 2 = \2r(x)y 2 (x) t
whenever the boundary conditions are such that the expression on the right-hand
side of (12-36) vanishes. Assuming this to be the case, (12-37) allows us to assert
that the functions \fry\ and \Zry 2 are orthogonal in e[a, b], or, equivalently,
that y x and y 2 are orthogonal in e[a, b] with respect to the weight function r. (See
Example 3, Section 7-1.) This latter terminology has the advantage of banishing
the cumbersome factor \Jr from the discussion of orthogonality, and amounts to
redefining the inner product on e[a, b] to be
P(x){yi(x)y2(x) - y 2 (x)y[(x)] =
Ly = \ry
We call the reader's attention to the fact that in general the operator L will not
be symmetric on S with respect to the weighted inner product defined by (12-38).
Nevertheless, Theorem 12-8 asserts that eigenfunctions belonging to distinct
eigenvalues are still orthogonal, and this
is really what is needed to construct
eigenfunction bases. Moreover, since the conditions required to ensure orthog-
onality here are the same as those imposed in Section 12-5, we see that the
conclusion of Theorem 12-8 is assured whenever S is described by boundary
conditions of Type 1, 2, or 3 of that section. And finally if L: S -> e[a, b] is
both one-to-one and normal, and if r(x) > for all x in [a, b], it can be shown
that Q[a, b] has a basis composed of eigenfunctions for L. Again we omit the
proof.
4-U
dx f) +
x
Ae* y=
x
A(9e
4a
>,
XO) = 0, y(d) = 0,
x in [0, a], the result cited a moment ago guarantees the existence of an eigen-
function basis for e[0, a]. To compute such a basis we argue as follows.
y = c xe + c2e
Ci + c2 = 0,
(-2 + 3VX)a C2e (-2-3VX)« =
Cie + Q
Thus c x
= c2 = 0, and (12-39) has no positive eigenvalues.
2x
Case 2. X = 0. This time the general solution of the equation is (c x + c 2 x)e~ ,
c2 = 0,
Cx sin 3V— X a = 0.
Hence (12-39) has nontrivial solutions if and only if X satisfies the equation
sin 3V— X a = 0, and it follows that the eigenvalues for this problem are
2„2
n
xn = ~
it
-gj2 '
"
= ,
l
>
2'
EXERCISES
Compute the eigenvalues and eigenfunctions for the boundary-value problems in Exercises
1-8, and in each case determine a Euclidean space in which a complete set of eigenfunctions
for the given problem is an orthogonal set.
2
x y" - xy' + (1 + \)y = 0,
given that y(l) and lim _>o+ \y(x)\ < °° How does the set of eigenvalues of
= 0, a; .
this problem from those encountered earlier in this chapter? [Hint: Note that
differ
the given equation is an Euler equation, and recall that when its indicial polynomial
has complex roots a ± j3i, the solution space on (0, oo ) is spanned by the functions
x sin 08 In x), x" cos (8 In x).]
01
Lx = y, L: S -> V, (12-41)
The best place to begin, perhaps, is at the point where we left the discussion of
initial-value problems in Chapter 4. As the reader will recall, we saw there that if
y(x ) = y ,
y'(x ) = y lt
Moreover, the function K(x, t), known as the Green's function for L for initial-
way from the coefficients
value problems, can be constructed in a perfectly definite
2
of L and a basis for the null space of L in Q [a, b]. Our present objective is to
2
prove that a similar construction is possible whenever S is a subspace of Q [a, b]
determined by a pair of unmixed boundary conditions
+
cny(d) oi 2 y'(a) = 0,
(12-42)
Mb) + 02/0) = 0,
Obviously then, our first task is to devise a criterion which will guarantee the
one-to-oneness of L. In other words, we must impose restrictions on the boundary
conditions appearing in (12-42) which will ensure that the only solution of the
equation
Ly = (12-43)
be the general solution of (12-43). Then y(x) will be identically zero if and only
ifci = c 2 = 0, andy(x) will belong to S if and only if
Then using the functions 1 , x as a basis for the null space of L, the above deter-
minant becomes
1 a
= b — a,
1 b
In this case we can also prove that L maps S onto Q[a, b] by the simple expedient
of constructing L~ l
directly from the formula for L. Indeed, since the equation
Ly = h is simply y" (x) = h(x), two integrations yield
y'(s) =
/.
h(t)dt + c,
and
y(x) = + — +
f[f hit) dt ds c(x a) d, (12-47)
where c and d are arbitrary constants. Applying the given boundary conditions
we find that d = and that
/ / h(t)dt ds + c(b - a) = 0.
Ja \-J a
Thus
rb i- rs
ds,
494 BOUNDARY- VALUE PROBLEMS |
CHAP. 12
= 0, s < 0,
"oO)
1, s > '0,
y(x) = u Q {x - s) h{t) dt ds
X -aC h(i) dt ds
-J af b — a Ja
b , rs
- X
— a
Kt) u (x — s)
— dt\ ds.
b a_
'b rb r
rir woO , -|
X*) = a \J a
- "oO - s) - lZa\ h ^ dt )
ds
•6 / rb
x — a\
u Q (s - U (X — S) — ds\ h(t) dt,
~b^~a\
where
x —
K(x, t) u (s — i) u (x — s) — - i ds
b
(x — a)(t - b)
— x < t,
b a
= {
(12-49)
(x-bXt-a), x>L
b — a
(See Exercise 3.) The function K(x, i) defined by this formula is known as the
operator L = D
2
Green's function for the for the given boundary-value problem,
and Eq. (12-48) can be read as the definition of L~ e[a, b] -> S with
l
:
b
L- l
h{x) = [ K(x, t)h(t) dt. (12-50)
Ja
12-9 |
GREEN'S FUNCTIONS: AN EXAMPLE 495
In the next section we shall prove that an analogous result is valid for any normal
second-order linear differential operator L acting on S so long as S is determined
by a pair of unmixed boundary conditions, and L is one-to-one when restricted
toS.
EXERCISES
Determine which of the following operators are one-to-one when restricted to the
given subspace S of Q 2 [a, b].
(a) L = D 2 + 4D + 4 (b) L = D2 + 1
S:y(l) = 0, y(e') =
Show that if the condition in Lemma 12-3 is satisfied for a pair of functions yi and
j 2, it will also be satisfied for every other pair of linearly independent functions of the
form
Yi(x) = Jiyi(x) + y 2 y2(x),
Y2 (x) = 7ayi(jt) + T 4 y 2 W.
3. Prove that
(x - a)(t - b)
— X < t,
b a
uo(s — u (x — s) — ds =
t)
b — a
\
(x - b)(t - a)
— X > t,
b a
where «o is the unit step function, and a < x < b,a < t < b.
4. Prove that the first derivative with respect to x of the Green's function K(x, i) con-
structed above has a jump discontinuity of magnitude 1 along the line x = t, but is
continuous at all other points in the region a<x<b,a<t<b.
5. Let be a fixed point in the interval [a, b], and let K(x, i) be the Green's function
to
constructed above. Prove that the function K(x, to) is a solution of the boundary-
value problem
D 2y = 0,
y(a) = y(b) = 0,
6. Use the technique introduced above to find the Green's function for the operator D 2
on the subspace of Q 2 [0, 1] defined by the boundary conditions >>(0) = /(l) = 0.
7. Repeat Exercise 6 for the boundary conditions
y(0) - y'(0) = 0,
y(l) + /(l) = 0.
496 BOUNDARY- VALUE PROBLEMS |
CHAP. 12
8. Let S denote the subspace of Q 2 [a, b] determined by the periodic boundary conditions
y(fl) = y(b),
y'(a) = y'(b),
and let L = D(p(x)D) + q(x) be normal on [a, b]. Prove that L is one-to-one when
restricted to S if and only if
h
L- X
h(x) = f K(x, t)h(t) dt
Ja
for all h in <B[a, b]. In order to motivate the axiomatic definition of the function
K(x, t) given below, begin by presenting a heuristic argument which will
we
_1
simultaneously suggest the existence of L the validity of the above formula, ,
Ly = h (12-51)
describes the behavior of a physical system under the influence of a given input
function h, and assume, in addition, that the response y(x) of the system at the
point x due to the unit input
. fl x =
if t,
(pt{x) ~ (0 ifx^ t,
is K(x, 0-t Then, in view of the linearity of L, it is reasonable to expect that the
* The following argument is essentially the one given by R. Courant and D. Hilbert in
Methods of Mathematical Physics, Vol. I, Interscience Publishers, Inc., New York, 1953.
t For instance, (12-51) might be the equation of equilibrium of an elastic string
stretched along the interval [a, b] and subjected to a continuously distributed force
h = h(x). In that case, <p (x) represents a unit force applied at the point t, and K(x, i)
t
b
y(x) = f K(x,t) dt.
Ja
In the case of a general input function h, this reasoning leads to the formula
b
y(x) = f K(x, t)h(t) dt, (12-52)
Ja
where the integrand is now viewed as the response at x to that portion of the
input applied at the point t.
Ly = <p t (x),
and hence, for each fixed t in [a, b], the function K(x, t ) satisfies the homo-
geneous equation
Ly =
/ ft Q (x)dx = 1.
Jt -(
(We assume that e has been chosen sufficiently small so that the interval [t — e,
to + e] is contained in [a, b].) Then, if K(x, t ) denotes the response of the sys-
tem at x to f we have
to ,
L[R(x, t )] =f t0 (x),
4: [p{x)R'{x, t )] dx + / q(x)K(X, t ) dx = 1,
J/ t —e «-* J t —e
or
We now make the not unreasonable assumption that as e — 0, K(x, > t ) —> K(x, t )
d
P(to) K(x, t )
dx
Thus
t+
°
d
K{x,t Q ) (12-55)
Tx pOo)
an equation which asserts that at the point t the derivative of K{x, t ) has a jump
discontinuity of magnitude l/p(t ).
(1) K(s, t) is defined and continuous for a < x < b, a < t < b, and,
as a function of x, is twice continuously differentiable except when x = t;
(2) For each fixed t in [a, b], K(x, t ) belongs to the subspace S (i.e.,
satisfies the boundary conditions imposed on the problem), and, in addi-
tion, is a solution of the equation Ly = 0, except at the point x = t \
t+
1
(3)-^K(x,t )
P(to)
With this as our definition, we now state the following basic theorem.
Ly = h
is given by the formula
rb
y(x) = / K(x, t)h(t) dt,
Ja
mined by the operator L and the boundary conditions which define the sub-
space §>.
12-10 GREEN'S FUNCTIONS: UNMIXED BOUNDARY CONDITIONS 499
We defer the proof of this result to the next section in favor of showing how the
Green's function for L can be explicitly computed once L and S are known. Here
we argue as follows.
Let ji and y 2 be solutions of the homogeneous equation Ly = chosen so
that y\ satisfies the boundary condition imposed at x = a, and y 2 the boundary
condition imposed at x = b; that is,
«iJi(«) + =
"2/1(0) 0,
0U>20) + foyM) = 0.
c ?± such that y 2 {x) = cyi(x), and the function cyi(x) would be a nontrivial
solution of Ly = satisfying both of the boundary conditions imposed on S.
This, however, contradicts the assumption that L is one-to-one when restricted
to S, and is therefore impossible.
We now use the fact that the Wron- i^^iW
skian of y x and y 2 never vanishes on the 2 '*^
interval [a, b] to find functions A x (t) and 1
A 2 (0 such that
^2(0^(0 - ><i(0/i(0 = 1
Pit)
for all t in [a, b]. The first of these equations guarantees that for each t in the
interval (a, b) the curves A i(t )yi(x) and A 2 (t )y 2 (x) intersect at the point x = 1
(see Fig. 12-3), while the second guarantees that the slope of A 2 (t )y 2 (x) differs
from the slope of Ai(t )y 1 (x) by l/p(t ) at x = t . Thus the function
^i(0j>i(*)> x < t,
K(x, t) = (12-57)
A 2 {t)y 2 {x), x > t,
satisfies the various conditions imposed in Definition 12-5, and is the Green's
function for L. Finally, by solving (12-56) for A and A 2 we
x obtain the formula
yi(x)y 2 (t)
X < t,
Remark: The argument just given provides a rigorous proof of the existence
of a Green's function for the problem under consideration. Uniqueness will be
established in the next section.
500 BOUNDARY- VALUE PROBLEMS | CHAP. 12
y(a) = y(b) = 0.
y x (x) = x — a, y 2 (x) = x — b
in (12-58), to obtain
'(*-«X/-»),
b — a
K(x, t) = {
(t-
b — a
EXERCISES
1 Prove that the equation Ly = 0, L as above, has a pair of solutions y i and y% such that
a\y\(,a) + a2y'i{a) = 0,
j8 iy 2 (6) + fa-tito = 0.
2. Use the method of this section to obtain the Green's functions for the boundary- value
problems of Exercises 6 and 7 of the preceding section.
3. (a) Find the Green's function for the boundary- value problem
// + k 2y = 0, y(0) = >>(!) = 0.
Continuing with the notation introduced above, we now show that Definition
12-5 does, in fact, uniquely characterize the Green's function for L. This is the
content of
b
y(x) = I K(x, i)h(t) dt
Ja
I
(See Appendix I.) Thus
ociy(a) + a 2y'(a) =
Ja
ai K(a, t) + cc 2 —d K(a, t) hit) dt
= 0,
* Strictly speaking, F' has a removable discontinuity at x = to, which we can ignore
because the discontinuities in and H
cancel by subtraction. K
*
the last step following from the fact that, as a function of x, K(x, t) belongs to S
for each t in [a, b]. Similarly
Mb) + W &) = 0,
o
y K(x, t)h(t) dt + J^ (*> x ~) h (x)
+ /' ^
Jx
K{x, t)h{t) dt-j± (x, x + )h(x)
dK . _. dK . _i_.
K(x, i)h(i) dt + h(x)
La dX 2
dK . _N dK , 4. N
FIGURE 12-4
and it follows that
dK a*
;>(*)
= A(*) + /»(x) 4k
d* 2
K(*> + />'(*) £
dx
*(*» + «(*)*(*» < )j
w
= h(x) + t)]h(t) dt,
Ja
12-11 | GREEN'S FUNCTIONS: PROOF OF THE MAIN THEOREM 503
L[y(x)] = h(x),
Proof. Let s and t be fixed points in [a, b] (with 1 < s ), and set u = K(x, s ),
v = K(x, t ). Then Lu = Lv = for all x in [a, b] different from s and 1 and ,
-jr[/>(wi/ - u'v)] = 0.
6
rn r\V hi
+ p(x)
I
Jf
^— (x, to) • K(x, so) - — (x, s ) • K(x, *
II
)j| +
= 0.
Thus
\ A IT rkTC I
+ p(so)
[^ (s$, So) • K(s , to) - "^ (sh~, *o) • K(s ,
*o)J
-
+ P(b)
[||
(b, to) • K{b, so)
H (b, s ) • Kib, /
)]
But, using the known jumps in dK/dx, this expression can be rewritten
dK
aiK(a, t ) = -a 2 -^ (a, t ),
axK(a, s ) = — a2 — (a, s ),
and
P Mb, t ) = -0 a |£(Mo),
^K{b,So)= -fo~(b,to),
we see that the bracketed terms vanish. Thus K(s , t ) = K(t , s ), as asserted.
13
13-1 INTRODUCTION
Historically the theory of boundary-value problems grew out of the study of certain
partial differential equations encountered in classical physics, and many of the
ideas treated in this book originated in attempts to solve these problems. For this
reason, if for no other, any introduction to the subject of boundary-value problems
would be incomplete without a discussion of partial differential equations. But
there are, in fact, compelling reasons for pursuing this subject which are quite
unrelated to any feelings of historical nicety. And though these reasons are
bound to become obvious as the chapter unfolds, it may not be out of place to
mention some of them before we begin.
For one thing, this discussion will serve to unite the various results on eigen-
functions and orthogonal series expansions that have been obtained in the pre-
ceding chapters, bringing them into sharper focus, and reenforcing the point of
view which sees them as a unified body of mathematical thought. For another,
we will at last be in a position to consider nontrivial physical problems, and the
fact that they can be solved with comparative ease should increase the student's
appreciation for the power of the techniques we now have at hand. And finally,
this material will in its turn suggest further problems leading to new results in the
subject.
defined in a region R of the xy-plane, the most general first-order linear differential
operator defined on e^/?) has the form
where a(x, y), b(x, y), c(x, y) are continuous everywhere in R. Here L may be
viewed as a linear transformation from e 1 (R) to <5(R), the space of all functions
continuous in R, and if h is any preassigned function in Q(R), the equation Lu = h,
u unknown, is a first-order (linear) partial differential equation.
It is clear that analogous, but more cumbersome formulas can be given for linear
differential operators (and equations) of higher order involving any number of
variables. Moreover, it is equally clear that all of the standard facts pertaining
to linearity continue to hold in this more general setting. This not withstanding,
the general theory of linear partial differential equations has very little in common
with that of ordinary equations for the simple reason that the solution space of
every (homogeneous) linear partial differential equation is infinite dimensional.
For instance, it is not difficult to show that the general solution of the first-order
equation
|+|= 03-,)
is a solution of (13-1), and it is clear that these functions are linearly independent
in Q(R) for any R. The fact that even so simple an equation as this has such a
wealth of linearly independent solutions gives some indication of the difficulties
13-2 I
PARTIAL DIFFERENTIAL EQUATIONS 507
As used here the word "region" is a technical term reserved to describe a connected
subset of the plane each point of which can be surrounded by a circle lying entirely
within the set in question.* Thus the upper half plane, an infinite or semi-infinite
vertical strip of the plane, the interior of a rectangle, or the annulus between two
concentric circles are all regions in this sense, and as such are typical of the two-
dimensional regions in which boundary-value problems involving partial differential
equations are defined. In addition, we shall assume henceforth that the boundary
of each of the regions we consider is made up of a finite number of simple differ-
a. For each point b on B and each p in R, the limiting value of u(p) as p ap-
proaches b along any smooth curve in R is f(b )- (See Fig. 13-1.)
b. For each point b on B and each p in R the limiting value of u(p) as p ap-
proaches b along any smooth curve in R which is normal to B at b is f(b )-
* A
subset R of the plane is said to be connected or pathwise connected if every pair of
points in R can be joined by a smooth curve lying entirely in R.
f The boundary of a region R is, by definition, the set B of all points in the plane with
the property that every circle centered at a point of B contains points in R and points
not in R.
508 THE WAVE AND HEAT EQUATIONS | CHAP. 13
EXERCISES
1. (a) Give the formula for the most general linear differential operator L: Q 1 (R) —
Q(R) when R is a region of xyz-space.
(b) Give the formula for the most general linear differential operator L: Q 2 (R) —
Q(R) when R is a region of the .xy-plane.
2. Determine which of the following partial differential equations are linear.
.
N d u ,
d u , d u „ ,, . d u ,
du du , d u . , ^
du du
(c)
fa* +
2
*ai + ^" " (d)
\a* +
W>)
u -
dx dy
a
d W d W
=0_
(C)
^+^ +(X>,M)
2
2
•*\ d u d u iou
(du ou\
du\ o
x 2x>'l-+ -) = (^>
a^ + + M M
(f)
>'^2 +
(a) Show that u(x — y) is a solution of the partial differential equation u x + uy =
whenever m is a differentiable function of a single variable.
(b) Let F(x, y) be a solution of u x + «y = 0. Set /? = x + y, # = x — y, and
write F(x, y) as
Show that G is actually a function of q alone by computing dG/dp, and then deduce
that every solution of u x + m„ = can be written in the form w(x — y), where u
is a differentiable function of a single variable.
Throughout the next three chapters, we shall, with but one exception, be ex-
problems involving various forms of the
clusively concerned with boundary-value
following second-order linear partial differential equations
2 2 2 2
d u d u d u d u
^+^ ^ ^a^' a>0
1 , .
(13 "2)
.
+
.
,
= >
(i3 - 4)
§+&"+f?-°-
Each of these equations first arose in classical (i.e., Newtonian) physics as the
mathematical description of a particular type of physical system, and the first two
are still known by the names of the simplest systems they describe. Thus (13-2)
13-3 | THE CLASSICAL PARTIAL DIFFERENTIAL EQUATIONS 509
is called the wave equation, and (13-3) the heat equation. Equation (13-4), on the
other hand, is known as Laplace's equation in honor of one of the mathematicians
who first studied it. For the present we shall content ourselves with the observa-
tion that Laplace's equation can be viewed as the time independent version of the
heat equation, and shall turn our attention to (13-2) and (13-3).
FIGURE 13-3
elastic string of arbitrary length which is vibrating vertically in the xw-plane and
whose position of rest lies along the x-axis (Fig. 13-3). Throughout this discussion
we shall make the following simplifying assumptions
(a) The amplitude of vibration of the string is small, and each point on the
string moves only in a vertical direction
(b) All frictional forces (both internal and external) may be neglected
(c) The mass of the string per unit length is sufficiently small in comparison
with the tension in the string that gravitational forces may be neglected.*
In Fig. 13-4 we have isolated a small segment of the string and have indicated
by T and T' the forces of tension acting on its endpoints. Since the string moves
only in the vertical direction, the horizontal components of T and T' must cancel,
and we have
2 X + AX
d u
T' sin a' — T sin a = p Ax (13-7)
dt 2 FIGURE 13-4
* The question as to whether these assumptions are permissible from a physical point
of view and do not prejudice the solution obtained is one which must be settled by the
physicist in his laboratory. In most situations they are, in fact, physically acceptable.
510 THE WAVE AND HEAT EQUATIONS CHAP. 13
X + AX
FIGURE 13-5
where p denotes the mass per unit length of the string, Ax the length of the seg-
ment in question, and d 2 u/dt 2 the acceleration of the segment at an appropriate
point between x and x + Ax. Using (13-6), this equation may be rewritten
~k~ 'dt 2
Hence
J_ du du p d u
Ax dx x+Ax dX k a72
2
<Tw 1 d u
dx 2 a 2 dt 2
where a = y/k/p.
The two-dimensional wave equation
2 2 2
d u d u _ 1 d u
(13-8)
dx 2 dy 2 ~~ a 2 dt 2
arises in physics as the differential equation governing the motion of a thin flexible
membrane of constant density which is tightly stretched and then fixed along its
boundary, and which vibrates in the w-direction from its position of rest in the
xy-plane. In this case the simplifying physical assumptions under which the equa-
tion is derived are as follows:
(a) the amplitude of vibration is small, and every point of the membrane moves
only in the it-direction;
13-3 | THE CLASSICAL PARTIAL DIFFERENTIAL EQUATIONS 511
is small, we can replace sin a x and sin a 2 by tan ai and tan a 2 respectively.* ,
TAx(tanai — tan« 2 )-
r Ay (tan a 3 — tan a 4 )
2
d u
TAx(tancn — tana 2 ) + TAy(tan<x 3 — tan« 4 ) = pAxAy-^* (13-9)
where p is the mass of the membrane per unit area, and d 2 u/dt 2 is computed at
some point in the region under consideration. But
tan «x = du
tan a 2 = —
du
dy
Ty (x^y+Ay) (x 2 ,V)
tana 3 = -
du
tana 4 = -
du
(3J,J/i) (x+Ax.j/2)
where x\ and x 2 lie between x and x + Ax, y± and y 2 between y and y + Ay.
Thus (13-9) may be rewritten
1 du du du du p d u
Ay\_dy (x t ,y+Ay) fry (x 2 ,i/)-
+ Ax dx dx (.x+Ax,v 2 )l T dt 2
(z,2/i)
Compare
3 5
X X
sinx = x- — +—- ,
with
tan x = x +
, x
—+
— + 2x
Their difference is of the order of magnitude of x 3 /2, which is small if x is near zero.
512 THE WAVE AND HEAT EQUATIONS | CHAP. 13
2 2 2
d u d u _ 1 d u
where a = y/T/p.
Finally, the three-dimensional wave equation arises, among other places, in that
branch of physics which deals with electric and magnetic fields in space. In fact,
by using Maxwell's equations from electromagnetic field theory it can be shown
that each of the components of both the electric and magnetic field strengths in a
region of space are governed by Eq. (13-2).
Next we consider the heat equation, and show that under appropriate assump-
tions it serves to describe the temperature distribution in material bodies as a
function of position and time. To obtain the one-dimensional version of this
equation we consider a slender homogeneous rod, lying along the x-axis, and
insulated so that no heat can escape across its longitudinal surface. In addition,
we make the simplifying assumption that the temperature in the rod is constant
on each cross section perpendicular to the x-axis, and thus that the flow of heat
in the rod takes place only in the ^-direction.
Now, for empirical reasons it is assumed that the quantity of heat, AH, which
flows across any cross section of the rod is proportional to the rate of change of
temperature u on that cross section. In other words,
where the minus sign is introduced because heat flows in the direction opposite
to the positive direction of du/dx.But it is also known that the amount of heat
which accumulates in any portion of the rod is proportional to the product of its
mass and the (average) time rate of change of temperature in that mass. Hence
AH = cmy t
(13-11)
for an appropriate positive constant c, known as the specific heat of the material
in question.
To obtain the heat equation we now focus our attention on the portion of the
rod between the points x and x +If p denotes the mass of the rod per unit
Ax.
length, then by (13-11) the amount of heat accumulating in this portion of the
rod per unit time is
.„ =
AH cp
.
Ax —
du
'
where du/dt is computed at some point between x and x Ax. But by (13-10) +
the amount of heat flowing across the two faces of this portion of the rod is
du du
AH = -k dx x+Ax
+ dx xj
13-3 I THE CLASSICAL PARTIAL DIFFERENTIAL EQUATIONS 513
du cp du
J_ du
Ax dx x+Ax dx ~k ~dl
d u du
= a
dx 2 Tt
2
where the constant cp/k has been replaced by a to emphasize that it is positive.*
An argument similar in almost all respects to the one just given can be used to
show that the temperature distribution in a thin rectangular plate, insulated so that
no heat flows across its faces, is governed by the two-dimensional heat equation
2 2
d u d u du
= '
dx 2 + w a
Tt
1. u(A, t) = 0;
2. ux (A, = 0;
3. u{A, t) = (l/h)u x (A, t), h a constant.
The first of these conditions simply means that the string is held fixed at the end-
point A, while the second is known as a. free-end condition. Here the string can
move in a vertical direction at A, but is constrained to do so in such a way that it
always remains horizontal there (see Fig. 13-6). The third condition says that
at x = A the displacement of the string is proportional to its slope. But for small
vibrations we know that ux (A, i) = tan a « sin a, where a is the angle which the
string makes with the horizontal at A. Thus u(A, t) ~ (l/h) sin a, and the string
behaves as though it were attached to the end of a spring as shown in Fig. 13-7.
Indeed, the requirement that such a system be in equilibrium at a given displace-
ment u{A, t) is T sin a =
ku{A, t), where k denotes the spring constant, and
(3) now follows by setting h= k/T. Finally, if the coupling constant h is very
small we approach the free-end condition given in (2).
When solving the one-dimensional heat equation it is customary to start with
a given temperature distribution u(x, 0)
initial = f(x) along the conducting rod,
and then choose the boundary conditions from among the following
1. u(A, t) = k, k a constant;
2. u x (A, = 0;
3. u x {A, i) = hu(A, t), h a constant.
The first of these conditions means that the end of the rod at A is maintained at
the constant temperature k, the second that the rod is insulated at A and neither
gains nor loses heat at that end. This time the third condition may be read as
asserting that the rate at which heat passes through the end at A is proportional
to the temperature at A.
EXERCISES
d u d u 2 du
dx 2 dy 2 dt
2. Show that the equation governing the temperature distribution in a thin homo-
geneous rod is
2
2 du d u , .
N
+ , , ,
a ^7 = ^~^2
,
when heat is being generated in the rod (say by an electric current) at the rate g(x, i)
3. Suppose that the assumptions under which the equation of motion of the vibrating
string was derived are modified to include a retarding force due to air resistance
which is proportional to the velocity of the string. Show that the equation governing
the motion then becomes
2 d u du d u
U = + '
dx^ dt d~fi
4. Suppose that an external force of magnitude G(x, i) per unit length acts on a vibrating
string. (This is the case of "forced vibrations.") Show that the equation governing
13-3 | THE CLASSICAL PARTIAL DIFFERENTIAL EQUATIONS 515
the motion is
d u 2 d u 1 . .
is a solution of the one-dimensional wave equation whenever /and g are twice dif-
ferentiable functions of a single variable.
6. (a) Let G(p, q) and its derivatives G p G q G pp Gpq and G qq be continuous through-
, , , ,
out the pq-p\ane. Prove that there exist twice continuously differentiable functions
d and (7 2 of a single variable such that
(b) Let F(x, be a solution of the one-dimensional wave equation, and suppose
t)
7. Let u(x, t) = f{x + at) + g(x — at) be any twice continuously differentiable
solution of the one-dimensional wave equation, and note that
Set
r(x) = f{x) + gix),
8. (a) Show that the physical units of the constant appearing in the equation
d u _
= 1 d u
dx2 a2~dt2
516 THE WAVE AND HEAT EQUATIONS | CHAP. 13
are those of velocity; i.e., length/time. [Hint: Recall that a = VWp> where the
units of T and p are, respectively, force/length and mass/length, and then use
Newton's second law.]
(b) Use the above result together with Exercise 6(b) to show that every twice con-
tinuously differentiable solution of the one-dimensional wave equation can be
interpreted as the superposition of two standing waves, one of which moves to the
right, and one to the left, both with velocity a.
FIGURE 13-8
2 2
d u d u
=
^ 1
i/(0, t) - u(ir, i) = 0,
u t (x, 0) = g(x),
where / and g are assumed known. Physically this problem consists of finding
the equation of motion of an elastic string which is stretched along the x-axis from
to X, clamped at its endpoints, given initial position fix), initial velocity g(x),
and then allowed to vibrate freely.
We begin by seeking solutions of (13-12) of the form
u(x, i) = X(x)T(t), (13-14)
,
o u _ Y"T '
= XT"
dx 2 dt 2
X"T = \ XT",
whence
Y" \ T"
whenever XT ^ 0. At this point we make the crucial observation that the left-
hand side of this expression is a function of x alone, while the right-hand side
involves only t. Thus each is a constant, A, and (13-15) is equivalent to the pair of
ordinary linear differential equations
X" ~ XX = °*
(13-16)
T" - \a
2
T = 0,
the first of which must satisfy the endpoint conditions X(0) = X(ir) = 0.
where A n and Bn are arbitrary constants. Hence, forming the product of the
functions in (13-17) and (13-18), we see that each of the functions
This done, we now attempt to use these functions to construct a solution u(x, t)
In general, of course, no one of the un (x, t) by itself will satisfy these conditions.
Neither, for that matter, will any finite sum of them, and it therefore appears that
the only possible choice is an infinite series of the form
n=l
00
= 2= sm nx
n l
iAn sin not + Bn cos nat) (13-21)
u{x, 0) = 2s
71=1
n sin nx, (13-22)
and we see that the first condition in (13-20) will be satisfied if the Bn are chosen
in such a way that (13-22) converges (pointwise) to the function /in the interval
[0, w]. But this is a familiar problem which, as we know, can be solved in either
of the following equivalent ways.
I. Let 0/ denote the odd extension of/ to the interval [
— tt, tt] (see Section 9-5),
and let Bn be the «th Fourier coefficient of Of. Then
2
Bn = -
r f(x)
/ sin nx dx, (13-23)
TT JO
and whenever / is sufficiently well behaved, the series obtained from (13-22)
using these values for the coefficients will converge to /everywhere in [0, t].
II. The functions sin nx, n = 1,2,..., form a complete set of eigenfunctions
for the Sturm-Liouville problem
and are an orthogonal basis for the space (PC[0, x] (Theorem 12-7). Thus the
Bn in (13-22) may also be computed as the coefficients of the expansion of/ in
terms of the eigenfunctions sin nx. The student should realize that the same series
is obtained in either case.
for determining the A n is much the same. We differentiate (13-21)
The technique
term-by-term with respect to t (under the assumption, of course, that this can be
done), and then set t = 0. Using the fact that u (x, 0) = g(x), we obtain t
00
and it follows that naA n must be the nth coefficient of the eigenfunction expansion
of g. Thus
An = /
/
sii nx dx,
g(x) sin (13-25)
Jo
u{x, i) = ^B
n=l
n sin nx cos nat,
three harmonics. The reader should note that the different overtones are char-
acterized by the appearance of nodes or stationary points which may be regarded
as visual manifestations of the eigenvalues for this problem. In the general case
where g^O the situation is much the same, but does not lend itself to such an
easy graphical realization.
t =
"2 Cm)
* The frequency v of the wave described by the function cos at is, by definition, a/2ir.
Physically, v may be interpreted as the number of waves which pass a given point per
unit time.
520 THE WAVE AND HEAT EQUATIONS | CHAP. 13
EXERCISES
In Exercises 1 through 5 find the solution u(x, i) of the one-dimensional wave equation
on the interval [0, L] subject to the endpoint conditions w(0, t) = u(L,i) = and initial
conditions as given.
4. uix, 0) = u (x, 0)
for < x < L/3 and 2L/3 < x < L
0, t
n u 2
o _ 2 o u
~ a
dfi dx~2
(See Exercise 4 of the preceding section.) Find the time independent solution of
this equation, and show how to use it to treat the general case.
7. Find the equation of motion of a vibrating string with fixed ends if it is released
from rest in the plucked position shown in Fig. 13-10. Sketch the position of the
string at times L/2a, L/a, 3L/2a, 2L/a.
8. Generalize the results of the preceding exercise to the case where the string is plucked
as shown in Fig. 13-11.
m(0, t) = u(L, t) = 0,
u(x, 0) = 0,
a2 a2
a* - a* -
a **• °'
w(0, = u(L, t) = 0,
w(x,0) = f{x),
u (x,0)
t
= g(x),
2 a u 2
2 di1 u a
d i . .
a ^r^~ ^r =
^c 2 ~ ~dt 2
F(x, t)
2 a u 2
2 da* U
u O
d I
a T-^r —
~ -7~z =
dx* Bt 2
such that w(0, t) = w(L, i) = 0, w(x, 0) = f(x) - v(x, 0), w t (x, 0) = g(x) -
v t (x, 0).
(b) Use the technique suggested in (a) to solve the boundary-value problem
2 a u 2
da u a
r — -t—r = Sin X COS t,
dx 2 dt 2
u(0, i) = u(JL, i) = 0,
u(x,0) = f(x),
u (x,0) t
= g(x).
11. Under suitable assumptions it can be shown that the torsional vibrations of a
homogeneous metallic shaft of uniform circular cross section are governed by the
partial differential equation
x
o
2 z 2<p
o
<p _ a
2
~dt 2 ~ drf'
where a a positive constant, and <p(x, t) is the angular displacement from equi-
is
librium at time t of the cross section of the shaft at x. (See Fig. 13-12.) Assume a
shaft of length L with <p(x, 0) = f(x), <p t (x, 0) = g(x).
FIGURE 13-12
(a) Find <p(x, t) if the ends of the shaft are clamped, i.e., if <p(0, t) = <p(L, t) = 0.
(b) Find <p(x, i) if <p x (0, t) = <p x (L, t) = 0. (These conditions obtain when the
ends of the shaft are free to twist and no torque is transmitted across them.)
00
dx 2 ~ a 2 dt 2 '
w(0, = "O, = 0,
and that the initial conditions u(x, 0) = f(x) and u t (x, 0) = g(x) can be formally
satisfied by choosing A n and Bn as the coefficients in the orthogonal series ex-
pansions
CO CO
2 Bn
n=l
sin nx = f(x) and 2
n=l
na^n sin nx = g(x) (13-27)
on [0, 7t], To complete the discussion of this problem we now impose conditions
on / and g which are sufficient to guarantee the validity of these results. Our
choice in this respect will be guided by the various convergence theorems for
Fourier series proved in Chapter 10, and we shall assume that the reader is familiar
with these results.
The first step in our argument consists of the simple observation that with An
and Bn as above
00
U\(x, = 2 ^n
n=\
sin nx cos nat (13-28)
2 2
_ 1 d u
d u
'
^
dx ~ a 2 dt 2
2
and
00
u 2 (x, = 2^
n=l
n s* n nx s* n nat (13-30)
dx 2 ~ a 2 dt 2 '
Thus (13-26) can be viewed as the sum of the solutions of two simpler problems,
and it suffices to direct our attention to them.
Beginning with the first, that is with (13-28) and (13-29), we recall that the
series for /given in (13-27) will converge (pointwise) to f(x) for each x in [0, tt]
whenever
(ii)/(0) = /W = 0.
to rewrite (13-28) as
- ~ 33
Ml (x , t) = ^^2 Bn sin n(x at) + £^ Bn sin "(* + at ^ ( 13 )
n=l n=l
(The rearrangement of terms here is justified by the fact that under the hypotheses
in effect this series is absolutely convergent.) Thus
= 0,
and
m(*,0) = Unx) + F(x)]
= F(x)
= f(x)
Under this assumption it is easy to show that F' exists and is continuous for all x
(see Exercise 3), whence
and
An-
dUi
= h[-aF'(x) + aF'(x)] = 0,
~dt t=0
as required.
Finally, to complete the argument and prove that U\ is also a solution of the one-
dimensional wave equation we impose the following additional restrictions on/:
(iv) f" is continuous on [0, x], and
(V) /"(0) = /"Or) = 0.
For then, arguing as above (Exercise 3 again) we find that F" is everywhere con-
tinuous,and that
a m = a
[F „ (x _ at) + F(x + at)l
dt 2 2
iifl - %F"(x
ir/r/: - at) + F"(x + at)].
dx 2
Thus
2 2
d Uj _ J_ d Uj ^
dx 2 ~ a 2 dt 2
g(x) = X) Cn sin nx
* In particular, this means that /has a right-hand derivative/^ at zero, a left-hand de-
rivative fi at x, and that
converges uniformly and absolutely to g(x) for all x in [0, *•]. Since A n = CJna,
we have
C
M „(v f) = -
1
Y—
°°
1
—dt
- = YC
Z-j n sin nx cos nat
n=l
^ = *£ C
d'
n=l
n sin n(x - at) + *
n=l
£C n sin n(x + at). (13-35)
But the assumptions imposed on g imply that these two series are uniformly and
absolutely convergent on [0, «•]. Hence so too are
and
l
dt/2
t=0
on [0, x]. Thus the function defined by (13-36) satisfies the initial conditions
prescribed in (13-31). Furthermore, since G is odd and periodic with period 2t,
it satisfies the end conditions w 2 (0, /) = u 2 (t, t) = as well. Hence, to com-
plete the argument we need only show that this function is also a solution of the
A X^
a ^-J n
— sin nx sin nat,
n= l
as required by (13-34).
526 THE WAVE AND HEAT EQUATIONS | CHAP. 13
rx-\-at
u 2 (x, t) = x- / G(s) ds
la J x—at
rt
— ^2 Cn sin nx J
cos nas ds
J°
n=l
= 1
- >
a *-*{ n
n\-^
,
—
C
sin
.
nx sin «af,
n=l
d
-^ = ±[G(x + at) - G(x - at)],
Finally, to assure the existence of the second partials of w 2 we now assume that
This implies that G' exists and is continuous on (— oo, oo), and that
Thus
2 2
d u2 _ 1^ d u2
3
~~
dx 2 a2 dt 2
00
An = —2 - r g(x)
nica J o
/
sin nx dx,
2
Bn = - f f(x) sin hx dx,
I
T Jo
w(0, = m(tt, = 0,
whenever
and
g(0) = g(ir) = 0.
Although Theorem 13-1 is sufficient for most purposes, it is clearly not strong
enough to handle every boundary-value problem involving the one-dimensional
wave equation that arises in practice. Perhaps the best known example which
falls outside the scope of this theorem is that of the "plucked string," and involves
finding the equation of motion of a string which is released from rest in the position
shown in Fig. 13-13. Here too a formal solution can be obtained by the method
of separation of variables (see Exercise 8, Section 13-4), but a much more careful
analysis than the one given above is required to treat those points where and f
f" have discontinuities. We shall omit this
discussion since it would carry us further
afield than we care to go, and content our-
selves with the remark that in most cases the
formal solution can in fact be proved valid.*
FIGURE 13-13
EXERCISES
1. Give a rigorous discussion modeled on the one appearing above to establish the
validity of the formal solution of the boundary-value problem
a u 2 a 2
2 o a u
a = '
8x2 d/2
w(0, t) = 0, ux (t, =
when
(Physically this is the problem of a torsionally vibrating bar with one fixed end and
one free end. See Exercise 11, Section 13-4.)
w*(0, /) = 0, u x (ir,t) =
when
(Physically this is the problem of a torsionally vibrating bar with both ends free.)
3. Let /be continuously differentiable on the interval [0, a), and suppose that
<^ = a
2 ~' (13-37)
K
2
dx dt
13-6 |
THE ONE-DIMENSIONAL HEAT EQUATION 529
u(x,0) = f(x).
Physically these equations describe a slender insulated rod of length L, with initial
temperature distribution f(x), whose left-hand end is kept at 0°, and which loses
(or gains) heat through its right end at a rate proportional to the temperature at
that end. The problem is to find the temperature u(x, t) in the rod as a function
of position and time.
We begin by seeking solutions of (13-37) of the form
EL = 2 T_
~ a
x T
whenever XT ^ 0, and it follows that
X" - \X = 0,
X (13-40)
r - 4t = o,
X a constant. Moreover, it is easy to see that the only way in which XT can be
nontrivial and yet satisfy the required boundary conditions is for X itself to satisfy
those conditions. Hence we must begin by solving the Sturm-Liouville system
X" - XX = 0;
X(0) = 0,
hX(L) + X'(L) = 0.
But this problem has already been discussed in Section 12-6, and we know that
its eigenfunctions are
Tn (i) = A n e -(KJa)t
530 THE WAVE AND HEAT EQUATIONS | CHAP. 13
Hence the only nontrivial solutions of (13-37) which are of the form XT and
which satisfy the first two boundary conditions in (13-38) are
(x2ja2)t
un (x, t) = An sin (X n x)e- , n = 1,2,..., (13-41)
An an arbitrary constant.
now remains to use these functions to construct a solution u(x, t) of (13-37)
It
which also satisfies the boundary condition u(x, 0) = f(x). To this end we form
the series
u(x,t)= ^A n e
-^ la2)t
sin (X„x)
n=l
and set t = to obtain
00
u(x, 0) = ^P A n sin \ n x.
n=l
From this it follows that the A n must be chosen as the coefficients of the series
expansion of/ in terms of the eigenfunctions sin \ n x.* Thus
Jo /(*) sin (\ n x) dx
A = 2
^
Jo sin (X n x) dx
and the series
„ (x> . v g/M^MA
x)dx
,-,.:,.',,
sin (XnX)> (13 _ 42)
n=1 Jo sin (\ n
which converges in the mean in (P6[0, L], is a formal solution of the given problem.
As oar second example we solve the one-dimensional heat equation given that
t/(0, = 100,
u(x, 0) = f(x).
X" - \X = 0,
r-\
a
t= z
o,
* The reader should note that there is no possibility of using a Fourier series here.
The expansion must be given in terms of the sin X n x.
13-6 | THE ONE-DIMENSIONAL HEAT EQUATION 531
« (x f = Kx + 100, K = - j^Thl
is a solution with the desired properties.
This done,we now observe that the sum of w and anY solution of (13-37) which
satisfies the homogeneous boundary conditions u(0, t) = 0, u x (L, t) =
— hu(L, i)
will in turn satisfy the first two boundary conditions in (13-43). In particular, our
earlier results imply that
= (Kx + + -^ la2)t
u(x, t) 100) J£ A n e sin (\ n x) (13-44)
n=l
is such a solution, and the problem will be solved as soon as the An are chosen so
that w(x, 0) = /(*). But when / = 0, (13-44) becomes
00
= (Kx + + X* x
u(x, 0) 100) X) An sin
n=l
Thus
Sj [f(x) - Kx - 100] sin (\ n x) dx
A = 2
Jo sin (\ n x) dx
EXERCISES
In each of the following exercises find the solution u(x, i) of the one-dimensional heat
equation which satisfies the given boundary conditions.
(a^ Find the temperature in the rod as a function of position and time.
(b; What is the steady-state temperature in the rod (i.e., the temperature as t —* oo ) ?
12. Find the steady-state temperature in a slender insulated rod of length L given that
13. Suppose that the temperature in the rod described in Exercise 12 is allowed to
reach equilibrium and that the temperature at the endpoints and L is then suddenly
changed to and 100, respectively. Find the new temperature distribution in the
rod as a function of x and t.
14. The equation governing the temperature distribution in a slender insulated rod in
which heat is being generated at a constant rate per unit length is
2
a u
du a
dt dx*
+ k,
k a constant. (See Exercise 2, Section 13-3.) Solve this equation given that
di2 u
.,2
o u 2 du
~~v
ox :
+ ay 2
a
Yt
(13-45)
FIGURE 13-14
13-7 |
TWO-DIMENSIONAL HEAT EQUATION; BIORTHOGONAL SERIES 533
We now propose to solve this equation in the presence of the following boundary
conditions:
u(x,y,t) = X(x)Y(y)T(t)
which are nonzero in the region under consideration. Substituting this expression
Y"
>Ljr
±Y" = a
„ T7
2l -
(13-49)
Y f
Thus
Y" T' X"
-yr = 2
a j; ^ =X '
Xa constant >
and we have
Y" - \Y = 0.
The boundary conditions (13-46) imply that y(0) = Y(M) = 0, and lead to the
familiar set of eigenvalues —n\ 2 /M 2 , and eigenfunctions
from which it follows that ^T'V^is also a constant, /*• Taking (13-47) into con-
sideration we find that we must now solve the Sturm-Liouville system
Leaving the details to the reader, we assert that in this case the eigenvalues are
„
2
m2 2
"
Mm =
-k -k
Mo = 0, Ml = — £2 ' ' • • '
jy~ ' ' •
534 THE WAVE AND HEAT EQUATIONS | CHAP. 13
Xm (x) = Bm cos ^~ r
» m = 0,1,2, ... (13-52)
a
T~ ~* \U + W)
and
2 2
_ -(x/a) [(m/L) +(n/M) 2 ]<
J- e (13-53^
is a soli tion of the two-dimensional heat equation which also satisfies the boundary
conditions given in (13-46) and (13-47).
To c ^mplete the solution of the given problem we must now choose the A mn
so that the (double) series
00
mirx niry
f(x, y) = 2 Amn cos
"X
.
sin
M
w=l
and hence must be the coefficients of the double Fourier series expansion of f
in the rectangular region R under consideration (see Section 9-8). Thus
• Af r L
AA n
_ _?_
~ LM '
A*.J0rin(5g)<fc*,
Jo L
and, when m ^ 0,
4
LM Jo Jo
1 ^^™^^™^)**^
and has been completely determined.
u(x, y, i)
on the vertical edges of the plate.) Such sets of eigenfunctions lead to the notions
of biorthogonal sets of functions and biorthogonal series expansions. The basic
theorem concerning biorthogonal sets in rectangular regions was given in Section
9-8, where it was shown that whenever {fm } and {gn } are orthogonal bases in
(9Q[a, b] and (PG[c, d], respectively, {fmgn } is a (bi)orthogonal basis in 6>e(R), R
the rectangle a < x < b, c < y < d. The import of this theorem is now obvious:
it allows us to apply the theory of eigenfunction expansions in rectangular regions,
and use them to solve three-dimensional boundary-value problems such as the
one discussed above.
EXERCISES
1 Verify that the eigenvalues and eigenfunctions for the Sturm-Liouville problem
4. Solve the two-dimensional heat equation in the rectangular region < x < t,
< y < t given that
«(0, y, t) = u(t, y, i) = u(x, 0, i) = u(x, t, i) = 0,
u(x,y,0) = f(x,y).
dx 2
M
_i__u—
dy dz
n
2 2
(b) Generalize the result in (a) to the case where u assumes arbitrary boundary
values on all six sides of the region.
«(0, y, t) = u(t, y, t) - 0,
7. Solve the boundary-value problem in Exercise 6 when the last boundary condition
is replaced by u (x, y, 0)
t
= g(x, y).
research in this field. Unfortunately most of these results are beyond the scope of
this book, and any attempt to solve (13-55) as it stands would be quite
out
|!i_
a* 2
c V* = ^> dt
(13-56)
which we propose to solve in the region — oo <*< oo, * > 0, subject to the
restriction that the solutions tend to zero as |x| -> oo. Applying the method
of
separation of variables with $(x, i) = iK*M0, we find that <p and ^ satisfy the
equations ,
and
g+ (X - cV)* = 0, (13-58)
we attempt to solve this problem by using the Hermite polynomials Hn (x) defined
in Section 1 1-6. With this in mind we each non-negative integer n
recall that for
the function e~ x2,4 Hn (x) is a solution of the differential equation
y" + n + l-^)y = ° 3 - 59 )
(
(see Exercise 12, Section 11-6). The strong similarity between this equation and
(13-58) leads us to seek a solution of the latter in the form
- {ax)2l4
S(x ) = e Hn (ax) (13-60)
S'( x ) = e- (a2x2)l4[aH^(ax) - ±a
2
xHn (ax)],
S"(x) = a e-
2 laV)l4
\HU(qx) - (ax)H'n (ax) - \
(l - ^p)Hn (ax)
But since Hn is a solution of
y" — xy' + ny = 0,
we have
H'rliax) — (ax)Hn(ax) + nHn (ax) = 0,
and it follows that
S"(x) = aV
2 a
= a [-n-l+ *f\s(x).
Thus
S"(x) + 2
a (n + \
- ^p) S(x) = 0,
Xn = (2/i + \)c, n = 0, 1, 2, . . .
, (13-61)
and
*„(*) = e-
{cx2)l2
Hn (^/Tcx). (13-62)
538 THE WAVE AND HEAT EQUATIONS | CHAP. 13
(It can be shown that up to multiplicative constants these are the only eigen-
functions for this problem.) Moreover, with X n as above,
<PnV) — e ,
(i) monotonic nonincreasing if Gn (t) > Gn+X {t) for all t in / and all n;
(ii) uniformly bounded if there exists a real number M such that \G < M n (t)\
for all t in / and all n. (The word "uniform" is used here because M simultaneously
bounds all of the Gn .)
Perhaps the simplest nontrivial example of a sequence with these properties is
{t }, n = 1, 2,
n
on the unit interval [0, 1]. In this case monotonicity is
. .
.
,
n
obvious, and uniform boundedness follows from the inequality \t < 1 whenever \
n
|/| <
1. This last inequality actually shows that {t } is uniformly bounded on
n
the larger interval [—1, 1], though it is no longer monotonic there since the t
alternate in sign to the left of the origin. For our purposes a more pertinent
example is furnished by the sequence with
\ n as in Section 13-6. Indeed, since X x < \ 2 < Gn (t) > Gn+ i(t) for all • • •
,
t > 0, and {Gn (i)} is monotonic nonincreasing. In addition, |Gn (/)| < e° = 1
for all / > 0, and hence the sequence is uniformly bounded on the semi-infinite
(13-63)
J2 Fn(x)Gn (t)
n=l
Proof. The proof is based upon the fact that an infinite series is uniformly con-
vergent if and only if its associated sequence of partial sums is uniformly a Cauchy
sequence (see Appendix I). Thus if
denotes the kth partial sum of (13-63), we will be done if we can show that for each
€ > there exists an integer K such that whenever k > K, p > 0,
\s k+p (x, i) - s k (x, i)\ < e (13-64)
<Tk(x) =
n=l
^F n (x)
be the kth partial sum of the uniformly convergent series £ n =i Fn (x). Then
Fk(x) = (Tk(x) — <Tk-i(x) for all k > 1, and we have
Sk+p — Sk = Fk+pGk+p + • • •
+ Ffc +1 Gfc +i
= (<Tk+p — <Tk+p—l)Gk+p + * •
*
p p—\
= 2^1 ( a k+n — <7k)Gk+n ~ ^J i&k+n ~ <Tk)Gk+ n +l
ns=l = n l
P-l
= 2_j (?k+n
— <Tk)(Gk+ n ~ Gfc+n+l) + fak+p ~ <Tk)Gk+p-
n=l
Thus
\Sk+p — Sk\ < 22 \<Tk+n — &k\ \Gk+n ~ ^fc+n+ll + Wk+p ~ <*k\ \G k +p\-
n=1
(13-65)
Now let e > be given, and let M be such that |<j n (Ol < M. for all n and all /
in the interval [c, d]. Then since £"=i Fn (x) is uniformly convergent on [a, b]
for all k > K, n > 0, and all x in [a, b]. Substituting this inequality in (13-65)
we obtain
"p-i
\Sk+p — Sk\ < 3M /] \Gk+n — <Jjfc+n+l| + |Gfc+p| (13-66)
n=l
But since {Gn (t)} is monotonic nonincreasing, G k+n (t) > Gk+n+1 (t) for all t, and
p— i p—
/) \Gk+n — Gk+n+ i\ = 2J (Gk+n — Gk+n+1 )
n=l n=l
= (Gjfc+i — <J7fc +2 ) + (G k+ 2 ~ Gk+ %) + • •
•
+ ((jfc+p_I — (jfc+p)
= Gfc+l — Gk+p .
Hence
With this result in hand we are ready to investigate the formal solution
oo
2 2
x>r (X " /a n
<x, t)=Y J
An sin (X n (13-67)
n=l
00
In this case the restrictions needed to guarantee the validity of (13-67) are
and we now assume that they are in force. The first step in the proof consists of
an argument to show that (13-67) is uniformly and absolutely convergent in the
region < x < L, t > 0. We use the Weierstrass Af-test (Appendix I), as
follows.
From the general theory of Sturm-Li ouville series we know that (13-69) con-
verges (pointwise) to fix) for all x in (0, L). Moreover, by Theorem 8-3 the series
2
x: ki
n=l
is convergent. Hence the sequence {\A n \} is bounded, and there exists a positive
constant K such that
\A n \
< K
for all n. From this it follows that for any t >
a
— (X n+1 /a )t
2
2
-alla )t
But from the way in which the X n were determined in the preceding chapter it is
— <*n+l/«
lim
n—><x>
p
— (X„/a
—— = )t Q
)'o
0,
- \A n \n sm(\ nX )e-^
2 la2)t
°\ < 4x^- (X " /aVo
2
d u
= - ]T^x2sin(X n x)e- (X » /a2) <
(13-72)
dx 2 n=l
and we have therefore proved that under the hypotheses imposed above the
function
2 2
= ^2 A n sin (\ n x)e -(\ Ja )t
u(x, t)
n=l
«(0, =
for all t > 0. Moreover, since
= ^2 A n \ n cos (\ n x)e~
dx
n=l
* At this point the reader would do well to recall the remarks made in Section 13-2
concerning solutions of boundary-value problems, and, in particular, that they need not
satisfy the differential equation on the boundary of the region in question.
1 3-9 |
THE HEAT EQUATION; VALIDITY OF THE SOLUTION 543
/aV
ux (L, = E A ^n
n=l
cos (X n L)]e-
(X »
and
But
h sin (X n L) + X n cos (X n L) =
for all n (see Eq. 12-27), whence
t/ x (L, r) = —hu(L, i)
u(x, 0) = f(x)
and the only troublesome step in the proof. At first sight it might seem that we
need only set t = in (13-67) and watch the series reduce to (13-69) to complete
the argument. Unfortunately this will not do. The error in such a naive argu-
ment arises from the fact that we are not in a position to assert the continuity of
u(x, t) for t > 0, and thus cannot guarantee that u(x t) approaches f(x ) as ,
(x , approaches (x
t) , 0). To establish this fact we use Theorem 13-2, as follows.
The sequence
-cx*/« 2 )*}
{e
2A
n=l
n sin (X n x )
n=l
{x2ja2)t
u(x, t) = ^2 A n sin (X n x)e-
n=l
with
So f(x) sin (X w x) dx
=
2
So sin (\ n x) dx
converges uniformly and absolutely in every region of the form < x < L,
Remark. The reader should appreciate that under the hypotheses imposed here
this theorem is valid only if the last boundary condition is interpreted
to mean
for all x in (0, L). However, when /is continuous on [0, L] and/(0) = f(L),
the series
fix) = 2 An
n=l
sin XnX
= W)t
u(x , t) Y,A n sin (\ n x)e-*
n=l
converges uniformly and absolutely for < x < L, t > 0. In this case u{x, t)
is continuous for < x < L, t > 0, and the last boundary condition is satisfied
EXERCISES
1. Give a rigorous discussion to establish the validity of the formal solution of the
boundary-value problem
d u _ 2 du
= a
dtf fr'
3. Prove that Theorem 13-2 remains valid if the hypothesis "nonincreasing" is replaced
by "nondecreasing."
14
boundary-value problems
for laplace's equation
14-1 INTRODUCTION
In this chapter we restrict our attention to boundary-value problems involving
Laplace's equation
S+g+S- <"-»
in rectangular, circular, and spherical regions. The student will observe that (14-1)
can be viewed as the special case of the heat equation in which du/dt = 0, and
with this interpretation in effect its solutions are called steady-state solutions since
they describe temperature distributions which are independent of time. Other
interpretations of (14-1) and its solutions are given in various exercises throughout
this chapter.
Before Eq. (14-1) can be solved in a circular or spherical region it must be
transformed to polar or spherical coordinates. The computations involved in
making these changes are somewhat lengthy, and so to avoid interrupting our
later work we dispose of them here and now.
The polar coordinate form of Laplace's equation is derived by introducing the
change of variables
x = r cos 0, y = r sin 0, (14-2)
in
as follows. From (14-2) and the chain rule for differentiation we obtain
—
du
dr
= cos 6
n
—
du
dx
+ ,
—
du
dy
.
sin 6
n '
(14_3)
—
du
dd
= —r sin 6n .
—
du
dx
+ cos —
,
du
dy
r
Q
6 >
546
14-1 INTRODUCTION 547
and since the determinant of this pair of equations is nonzero everywhere in the
punctured plane r ^ 0, (14-3) can be solved for du/dx and du/dy whenever r ^ 0.
This gives
—
du
= cos 6
„du 1 .
sin
„
—
du
dx dr r dd
(14-4)
—
du
dy
= .
sin 6
„
—
du
dr
+ - cos —
du
,
dd
1
r
.
But these formulas are valid for any differentiable function u = u(x, y), and
hence can be applied to du/dx and du/dy themselves to obtain
„ d (du\ 1 „ d (du\
= cos e Smd
.
dx 2 Fr\te)--r d-8\d-x)
(14-5)
2
d u .
n /du\
d (du ,1 n d (du\
If the derivatives
are now computed from (14-4), substituted in (14-5), and the resulting equations
added, we find that
2 2 2 2
d u d u d u 1 du 1 d u
+ =
~ + +
dx 2 dy 2 dr 2 r dr r2 dd 2
2
d u 1 du d u
Jr
1
= °- (14^6)
2
dr- +-rTr 7 dT
2 2
y = rsin<p sin 0,
z = r cos <p.
(See Fig. 14-1.) A computation entirely similar to the one given above now yields
the equality
^^d^w =
-T J[j)/ 2 chA
+
1 d_ ( Sm . d«\
+
1 c^m
d*2 d>,2
-I"
az 2 r2 5r ^ dr// r 2 sin ^ d(p \
<P
d(p ) r 2 sin 2 ^ 302
'
548 BOUNDARY- VALUE PROBLEMS FOR LAPLACE'S EQUATION |
CHAP. 14
dr V
2
Tr)
.
+
1
sin^ * V ^j + sln^ W= 2
°" (14~ 7)
.
study of gravitational attraction in three-space, and it was only later that he found
the version
2 2 2
d u d u d u
+ + ^
dx 2 dy 2 dz 2
introduced earlier.
EXERCISES
2
d u d\ _d\ lfla
+,}_dji.
dx 2
+ dy 2
~~
dr2
+ r dr r 2 dd 2
'
2. Use the transformation formulas given in the text to derive Eq. (14-7) from (14-1).
g + g-0 04-8)
in a rectangular region of the form < x < L, < y < M, and begin by
imposing the boundary conditions
u(0,y) = 0, u(L,y) = 0,
^
u(x, M) = 0, u{x, 0) = f{x).
In physical terms this problem requires that we find the steady-state temperature
in a thin rectangular plate three sides of which are held at 0°, and the fourth
at
Weagain find the solution by separating variables, and thus set u(x, y) = XY,
where X
and Y are, respectively, functions of x and y alone. This implies that
u xx = X"Y, u yy = XY", and hence (14-8) becomes
XI - _ EL
Y ~ X
14-2 | LAPLACE'S EQUATION IN RECTANGULAR REGIONS 549
X" + XX = °>
(14-10)
Y" - \Y = 0,
X a constant, and the given boundary conditions plus the requirement that
XY 9* imply
X(0) = X(L) = 0,
Y(M) = 0.
Hence we can again begin with the solutions of the Sturm-Liouville system
X" + \X = 0,
*(0) = X(L) = 0,
2 2
n T
Y" — Y=
and has the general solution
Bn = — cosh-y-M, Cn = sinh-y-Af.
Yn (y) = sinh
^ (M - y). (14-13)
* In this problem it is more convenient to work with hyperbolic functions than with
exponentials.
550 BOUNDARY-VALUE PROBLEMS FOR LAPLACE'S EQUATION |
CHAP. 14
00
u(x, y) = 2
n=\
un(x y)>
= ]T A n
n=l
sin
(j^j sinh
^ (Af - y).
nirx
u(x, o) = 2^ ^ smh I "~ ZT /
sin
L
ra=l
and will represent the function / on the interval (0, L) if the A n sinh (rnrM/L) are
the coefficients of the Fourier series expansion of the odd extension of /to [— L, L\.
Thus L
. , rncM ( 2 r, \ • (nirx\ ,
t
^ n sinh -j- = j /(*) sm \^-^y <**
Jo /
and
u(x y,
ki*,.)';
£
y
- 1 Zj /°/W^(""/f)^ sin (V) sinh ^
sinh(nirM/£) \t/ £
(M - ,)•
w=l
(14-15)
Once again it can be shown that this solution is valid whenever / is sufficiently
smooth (see Exercise 10), and that it is the only possible solution of the problem
in question.
that the above argument can be applied to solve any boundary-value
It is clear
problem for Laplace's equation in a rectangular region given that the solution is
to vanish on three sides of the rectangle. But if u u u 2 u 3 w 4 denote the solutions , ,
corresponding to the four possible cases with one nonzero boundary condition,
the function
4
u(x, y) = ^ u k (x, y)
14-2 |
LAPLACE'S EQUATION IN RECTANGULAR REGIONS 551
will be a solution assuming nonzero values on all four sides of the rectangle, and
it follows that we can actually solve all problems involving Laplace's equation
and boundary conditions of the form
EXERCISES
In each of the following exercises find the solution of Laplace's equation in the rectangle
< x < L, < >> < M which satisfies the given boundary conditions.
2. u(L, y) = u(x, 0) == 0,
m
"(0, y)
v
= U < y < M/2
[M - y, M/2 < y < M
3. u(0,y) = u(L,y) = 0,
(
L
0, < x <
2
u(x, M) = u(x, 0) = <
2 3L L2 L
6. Find the series solution of the equation u xx + u yy = in < x < L, < y < M
given that
u(0,y) =fi(y), u(L,y) = f2 (y),
u(x, 0) = h{x), u(x, M) = Mx).
7. Derive the formula for the steady-state temperature in a ir X ir square plate whose
faces are insulated so that no heat flows across them, under the assumption that
8. Solve u xx + Uyy = in the region < x < ir, < y < t, subject to the bound-
ary conditions
9. Find the steady-state temperature in the infinite plate shown in Fig. 14-2 if the
vertical edges are kept at 0°, the edge on the x-axis at 100°, and if the temperature
throughout the plate is bounded.
100°
FIGURE 14-2
*10. (This Exercise assumes a knowledge of the material in Section 13-9.) When M=
L = tt the boundary-value problem discussed in this section is
d u d u _
dx 2 dy 2
u(x,0) = /(*),
\7* x
sinhn(7r-->0
/
«(*, y)
x
= 22 Aa " sin t(«*>
sinh^x '
where Y,n=\ A n sin nx is the Fourier series expansion of the odd extension of / to
the interval [— t, it].
sinh «(x — y)
n = 1,2,
sinh mr
sinh niit — y)
<
— < 1
sinh nir
for all n.)
(b) Assuming that / and /' are piecewise continuous on < x < tt, show that
the formal solution given above is uniformly and absolutely convergent in the
rectangle < x < w, y < y < t for any y > 0. [Hint: Use the methods of
Section 13-9, including that of Exercise 3.]
(c) Show that under the assumptions given in (b) the series obtained by twice
differentiating the formal solution term-by-term with respect to x and with respect
to y are also uniformly and absolutely convergent on < x < ir, yo < y < if.
14-3 |
LAPLACE'S EQUATION IN A CIRCULAR REGION 553
(d) Using the results of (b) and (c), prove that the series solution u(x, t) satisfies
(e) With /and/' as above, prove that lim y ^o+ u(x y) = f(x ) for each x in ,
(0, ir), and thus show that u(x, i) is a solution of the given boundary-value problem.
(0 Prove that the solution u(x, y) satisfies the boundary condition u(x, 0) = f(x) in
the stronger sense
lim u(x, y) = f(x )
V->0+
for all xq in (0, it) whenever / is continuous and /' piecewise continuous on [0, ir].
2 2
d u 1 du 1 d u n nA 1A
.
dr 2 r dr r2 dd 2
Here, in addition to the usual series solutions obtained by the method of separa-
tion of variables, there are certain exceptional solutions whichdepend upon only
one of the variables, and we begin by examining them.
First, if u is a function of r alone, (14-16) becomes
~ + ~f =
dr 2 r dr
(14-17)
or
(du\
ll'Sl-o.
and two integrations yield
u = c In (kr), (14-18)
where c and k > are arbitrary constants. In this case the solutions are undefined
when r = 0, and hence are valid only in regions of the form r > where they
provide solutions of (14-16) which are constant on circles centered at the origin.
On the other hand, if u is a function of 6 alone, we have
^=
dd 2
significance), then u(6 + 2ir) = u(6), A = 0, and it follows that (14-16) has no
nonconstant solutions which depend only upon the polar angle 0.
where / is a continuous function such that f(6 + 2x) = /(0) for all 0.
and we have
r
2
R" + rR' 0" >
-^ = -pr- = X,
>
X a constant.
R ©
This yields the pair of ordinary differential equations
0" + x@ = 0,
(14-20)
r
2
R" + rR' - \R = 0,
Save for notation, this problem was solved as Example 2 in Section 12-5, where
we saw that its eigenvalues and eigenfunctions are
Xw = «
2
, n = 0, 1,2, . . .
and
0(0) = A n cosnd + Bn sinnd. (14-21)
Moreover, when \ = \n = n
2
, the second equation in (14-20) becomes
r
2
R" + rR' - n
2
R = 0, (14-22)
u n (r, 6) = r
n
(A n cos nd + Bn sin n&),
each of which is continuous, satisfies Laplace's equation, and is periodic in 6 with
14-3 | LAPLACE'S EQUATION IN A CIRCULAR REGION 555
period 2ir. To satisfy the boundary condition w(l,0) = f(6), we now form the
series
u{r, d) = 4 +
2
2 '"V* cos nd + Bn sin nd>>> (H-24)
i/(l, 0) = 4s + Z)
Z
(^» cos w * + B" sin "*)'
n=l
and it follows that the constants A n and Bn must be the Fourier coefficients of/; i.e.,
Bn = - I
f(0) sin nd dd.
T J —n
Finally, we leave the reader the task of verifying that whenever / is sufficiently
Interchanging the order of integration and summation (an operation which can
be shown to be valid here), this expression may be rewritten
T
°°
f
u(r, 6) = —1 I f(s)
l~l
-= + V^ n
r (cos ns cos nd + sin ns sin nd) ds
=
\\_ Ks)\~ + 2 r
n
cosn{s - d)\ds.
556 BOUNDARY-VALUE PROBLEMS FOR LAPLACE'S EQUATION |
CHAP. 14
\ + 2 '
,n
cosn(5 - 0) (14-26)
appearing in this integral. The easiest way to go about this is to introduce complex
numbers, and set
Then
z
n = r
n |-
CQS n (s _ fl) _|_ i sm n (s _ q^
i
"
Z
n=l
\
2
+ 2
.-i
*
n
= - \ + + z + z' + * * •>
--U 2 '
1
l
- z
= _ I +
^
1
_ . 1 — r cos (s — 0)
Re
2
+ 1 - zj 2 1 - 2r cos (5 - 0) + r2
2(1 - 2r cos (s - 0) + r2 )
This expression is known as the Poisson Integral form of the solution of Laplace's
equation in the unit circle with boundary condition u{\, 0) = /(0), and, as men-
tioned above, appears in advanced work in this subject.
14-3 I LAPLACE'S EQUATION IN A CIRCULAR REGION 557
EXERCISES
1. Use the technique suggested in Example 3 of Section 4-8 to show that u = c In (kr)
d u 1 du _
dr*^~ ~r dr
~ '
n
2. Verify that r n and r~ are solutions of
r 2 R" + rR' - n 2R = 0.
Q\
P'< Mi
V.
R2 V
<\*V<
Wire
dx
4. If a doubly infinite straight wire carries a uniform static electric charge a per unit
length (o- > 0), then the potential at a point P due to the charge on an element of
the wire of length dx is defined to be
where i?i and R2 are as shown in Fig. 14-4, and the distance OQ = 1. By experi-
ment it can be shown that the potential V(P) at P is the "sum" of the potentials due
to the various elements of length, i.e., that
V(P) = 2a dx.
Jo \Ri R2 )
Set OP — r, and evaluate this integral. Compare the result obtained with the
^-independent solution of the two-dimensional Laplace equation.
5. Refer to Appendix I to obtain sufficient conditions on f(6) to permit the necessary
interchange of the order of integration and summation in the derivation of the
Poisson integral.
558 BOUNDARY- VALUE PROBLEMS FOR LAPLACE'S EQUATION | CHAP. 14
6. Verify directly that the Poisson integral satisfies Laplace's equation when r < 1.
(See Appendix I for the necessary rule for differentiating the integral.)
*7. (This exercise assumes a knowledge of the material in Section 13-9.) Assume that
fid) and fid) are piecewise continuous and periodic on [0, 2w] with period 2ir, and
that
fie) = h[fid+) + fie-)]
°°
a
uir, 6) = -j + ^
n=l
r
n
(A n cos nd + Bn sin nd),
where
A
fie) = -j + X) ( An cos nd + Bn sin n6 ^
(c) Show that lim r ^i- uir, do) = Wo) for all O in [0, 2*-].
(d) Prove that if fid) is also continuous then the boundary condition is satisfied in
a (I4" 28)
('S)+ib£(""S) + iik8f-
As usual we our attention to solutions which are single-valued,
shall restrict
continuous, and bounded under consideration, and which as a conse-
in the region
quence are periodic in d with 2tt as a period. With these assumptions in force it is
easy to show that (14-28) has no nonconstant solution involving only one of the
variables. Under less however, solutions of the latter
restrictive hypotheses,
type do exist, but since they can be found without difficulty they have been left
to the exercises.
14-4 | LAPLACE'S EQUATION IN A SPHERE 559
In this section we
solve (14-28) under the assumptions that the solution is
independent of 0,and that u is known when r = 1; i.e., «(1, <p) = f(<p). Then
2
d u/dd
2
= 0, and (14-28) assumes the somewhat simpler form
,2. ^.. *.. *2.
2# u
+ cot^ + ^i = 0. (14-29)
We now apply the method of separation of variables, this time with u(r, <p)
=
R(r)$(<p), to obtain the pair of equations
r
2
R" + 2rR! - \R = 0, < r < 1, (14-30)
and
<*>" + cot <p <*>'
+ X<l> = 0, < <p < ir, (14-31)
cfo
d<p
_ _ . d&
ds
d% =
d<p*
Sm 2
.
^- d'Q
C0S<f>
d$
ds
and we have
or
(1
-^ 2>
) ^-
d$ z
„
2
^+ g&
x$ = '
-^^ L (14-32)
But this is none other than Legendre's equation, and hence the eigenvalues for the
problem under consideration are the integers X n = n(n + 1), n = 0, 1, 2, . . .
(see Example 4, Section 12-6). Since the corresponding eigenfunctions for (14-32)
are the Legendre polynomials Pn (s), it follows that the functions
are the eigenfunctions for (14-31) under the given boundary conditions.
Next we observe that (14-30) is an Euler equation. Moreover, when X = \ n =
n(n + 1), an easy computation shows that its solution space is spanned by the
pair of functions
n
R n (r) = r and R n (r) = r~ (n+1) ,
and since the second of these solutions must be rejected because of its discontinuity
at the origin, we conclude that the relevant solutions of (14-29) are
Finally, to satisfy the boundary condition w(l, <p) = f(<p) we form the series
00
f(<p)
= Y,A n Pn (cos<p) (14-36)
n=0
for all <p in the interval [0, tt]. To this end we again set s = cos <p and rewrite
(14-36) as
00
/(cos
-1
s) = 2dP
n=0
n n (s),
inwhich form it is obvious that the A n must be chosen as the coefficients of the
-1
Legendre series expansion of /(cos s); i.e.,
In + 1
/(cos
i
s)Pn (s) ds
2 1
(14-37)
2n + 1
f(<p)Pn (cos <p) sin <p d<p.
Using these coefficients, (14-35) will converge in the mean to f{<p) when = 1. /•
Moreover, as in all of the preceding examples, it can also be shown that this series
respect to each variable whenever /is sufficiently smooth (see Exercise 13). This,
of course, implies that (14-35) is a solution of the given boundary-value problem,
and problem has a unique solution (see Appendix IV), we are done.
since this
Before going on, there is one rather interesting aspect of these results which
deserves closer attention. It concerns our apparent success in obtaining an
orthogonal series expansion for the function f(<p) defined on the surface of the
unit sphere, which, after all, is the import of the assertion that
00
71
2= ^n(cOS <p)
addition and scalar multiplication, and also becomes a Euclidean space if we set
/•*= fff(<p)g(<p)dS,
04-38)
s
where the integral in question is taken over all of S. Integrals of this sort are known
as surface integralsand are discussed in Appendix IV. For now we need know
only that they enjoy all of the standard properties of linearity (a fact which is
implicit in the statement that / g is an inner product), and that they can be
•
evaluated as ordinary iterated integrals. Indeed, when 5 is the surface of the unit
sphere, it can be shown that
1-2 1 r*
f'g=fJo [ K<p)g(<p) sin
Jo
<P d(P dd
.>Jde .
i
if m t* n,
!f m = n.
'fcrr) FIGURE 14-5
* This equality can be motivated by observing that the integral on the left is taken
over the surface of the unit sphere, described by letting <p vary from to x, 6 from to 2ir.
Since an element of surface dS on a sphere of radius r is given by the expression
n= «=1
n=2 «=3
FIGURE 14-6
(See Theorem 11-6.) Thus the functions Pn (cos <p), n = 0, 1,2, ..., are mutually
orthogonal in Q V (S), and the series given in (14-36), with coefficients as in (14-37),
is the orthogonal series expansion off in terms of the Pn (cos <p). From this point
of view the statement that this series converges in the mean to / is equivalent to
the assertion that the Pn (cos <p) form a basis for Q V (S), a fact which is proved in
exactly the same way as the corresponding statement for the Legendre polynomials
in the space (PQ[— 1, 1].
The functions Pn (cos <p) appearing in this discussion are known as surface zonal
harmonics, or simply zonal harmonics* The use of the term "harmonic" is ex-
plained by the fact that any solution of Laplace's equation, and thus, in particular,
Pn (cos <p), is said to be a harmonic function. The term "zonal" is applied because
the curves on the surface S of the unit sphere along which Pn (cos <p) vanishes are
parallel to the equator of S and thus separate S into "zones." Indeed, using the
fact that the Legendre polynomial of degree n has roots x u . . . , xn in the open
interval (—1, 1) (Theorem 11-7), it follows that the zeros of Pn (cos <p) consist
-1
of the n parallels of latitude on the unit sphere given by v?i = cos xu ,<p n = . . .
-1
cos xn , < ^i < 7r. These curves determine the zones alluded to in the name
zonal harmonic (see Fig. 14-6).
n
* Some authors reserve the latter term for the functions r Pn (cos <p).
14-4 | LAPLACE'S EQUATION IN A SPHERE 563
EXERCISES
r
2
R" + 2rR' - \R =
is spanned by the functions /"and r
_(n+1) when X = n(n + !)•
2. Find the general solution of Eq. (14-28) which depends only on r, and which
vanishes at infinity.
3. Show that (14-28) has no nonconstant single-valued solution which depends only
on0.
4. Find the general solution of (14-28) which depends only on <p.
5. Use the results of the preceding exercises to show that under the conditions assumed
in the text (14-28) has no nonconstant solutions which involve only one of the
variables r, <p, 6.
z
6. A semi-infinite wire coincides with the posi-
tive z-axis, and carries a static electric charge
a per unit length (a > 0), while a similar
wire, oppositely charged, coincides with the
negative z-axis. Given that the potential at
a point P in space due to the charge element
a dz is a dz/R (see Fig. 14-7), perform the
appropriate integration, and find the poten-
tial at P due to the entire (doubly infinite)
wire. Compare your answer with the result
obtained in Exercise 4. x figure 14-7
7. Find a solution u(r, <p) of Laplace's equation in the region \r\ > 1 given that
"(!,*>) = f(fP).
8. Let u(r, <p) be the steady-state temperature in a spherical shell of inner radius a,
outer radius b. Find u(r, <p) if u(a, <p) = f(<p), u(b, <p) = 0.
*=i dxk
where u = «i(xi) • • •
w„(xn ). Prove that
—
u"
u k
= A*, X* a constant, k = 1, . . . , n,
and that Xi + • • •
+ X„ = 0.
12. Solve the problem in Exercise 11 when f(<p) = cos 3 <p — cos <p.
"(r, <p) = ^A
n=0
nr
n
Pn (cos <p)
constructed in the text is uniformly and absolutely convergent for r < tq < 1,
< (p < ir. [Hint: Recall that under these assumptions the Legendre series for /
converges in the mean to /, and hence that there exists a real number M such that
\A n \
< M for all n. Then use the fact that |PTC (*)| < 1 for -1 < x < 1. (See
Exercise 15, Section 11-3.)]
(b) Given that < n 2k for -1 < x < 1, where P™ denotes the £th
\P%*(x)\
derivative of prove that the series in (a) may be twice differentiated term-
Pn ,
by-term with respect to either variable, and that the resulting series still converge
uniformly and absolutely for r < ro < 1, < <p < ir.
(c) Use the result in (b) to show that this series is a solution of Laplace's equation
in the region r < 1
whenever YJZ=o A nPn {cos <p) converges uniformly to/(<p) on the interval [0, *].
only to the requirement that the solution be continuous, single-valued, and take
on preassigned values when r = 1. Thus we consider the equation
f^^ + JLafrin*^
\
+ 4-^-0
Bd sin 2 2
(14-40)
\ Br) sin ip d<p B<p) <p
r
2
R" + 2rR' $" + cot ip <£' 1 ©'
=
R + $ sin 2 <p
0. (14-41)
14-5 I LAPLACE'S EQUATION; SPHERICAL HARMONICS 565
Since the first term in this equation depends only on r, while the second is inde-
pendent of r, we have
r
2
R" + 2rR!
R
= X
X a constant, or
r
2
R" + 2rR! - \R = 0. (14-42)
for the solution space of this equation (see Exercise 1). Thus if we demand that
a be an integer (which we do for the sake of simplicity) it follows that
X = ra(ra + 1), ra = 0, 1, 2, ... , and that a = ra, or a = — (ra + 1).* Hence,
in this instance, the relevant solutions of (14-42) are
— (m + 1)
the solutions r being rejected for reasons of continuity.
Next, we use the chosen values of X to rewrite (14-41) as
+ ^COt 0"
+ n
<S>" $>' 1
if
m{m
, ,
1) H
,
r
$ sin 2 <p &
or as
0^
+ /
ra(ra + n
,
1) sin
2 •
<p -\
,
*" + —COt r <p
—$'
sin
. 2
<p 0. (14-44)
Thus 0"/0 is also constant, and the requirement that be periodic with period
2t implies that
—= -n , n = 0, 1, 2,
— n
$" + cot <p <*>' + ra(ra + 1)
sin J ip_
$ = 0, (14-46)
* The reader who feels that we have been somewhat ruthless in discarding potential
values of X should realize that we are only trying to find a single solution of (14-40) which
satisfies the given boundary conditions, and the fact that the above choices lead to such
a solution will furnish a postiori justification for making them. Of course, the heart of
the matter lies in a uniqueness theorem which can be cited to prove that nothing is lost
by this line of reasoning.
566 BOUNDARY- VALUE PROBLEMS FOR LAPLACE'S EQUATION | CHAP. 14
$ = (1 _ sT'V
2\n/2,
Then
*. (1 _, V /.[:dw ns
- w
ds 1 s2
2
d *
ds 2
= (1 - 5*>
2yi/2
<ft
2 \
2M5
—
1
s2
dw
ds
— n —— —-
I
-,
(1
s (n
„,„
— —I)
s2 )2
- w
dw
(1 - s
2
)^
ds
- 2
2(n + 1>^
ds
+ [m(m + 1) - n(n + l)]w = 0. (14-48)
(1 - 2
s )y" - 2j/ + m(m -f \)y = (14-49)
on [— 1, 1]. Hence P%\ the nth derivative of Pm , satisfies the equation obtained
by differentiating (14-49) n-times with respect to s. Performing these differentia-
tions we obtain, successively,
(1 - *V 3)
- 2(2s)y
{2)
+ [m(m + 1) - 2]y
{1)
= 0,
(1 _ 2
5 )/4) _ 3(2s)/
3)
+ [m(m + 1) - 2(3)]/
2)
= 0,
(1 _ 5 2)/»+2) - („ + l)(2s)y
(n+1)
+ [m(m + 1) - «(« + l)]/
n)
= 0,
and since the last of these equations is just another form of (14-48), it follows that
P£° is a solution of that equation on [— 1, 1]. When rewritten in terms of <p we
find that
n
*mn (*) = sin <p P%\cos v ) (!4-50)
is a nontrivial solution of (14-46) on [0, ir] for each pair of non-negative integers
m and n, n < m*
Combining the solutions found in (14-43), (14-45), and (14-50) we obtain the
functions
m n n)
r (A mn cos nd + Bmn sin nd) sin <? /4 (cos <p),
n = 0, 1, . . . , m, m = 0, 1, 2, . . .
,
f-g = fff(<p,e) g
(<p,e)ds
s
= /"
Jo
V Jo
f(#> e ^(<P, 0) sin ip dip dd,
and let
where m and n are as above. These particular functions are known as spherical
harmonics, and for each fixed value of m the 2m + 1 functions
2 2
||WmO|| UmO ' UmO
* n+1 n)
A — /• Wmn Jo Jo' /O, 0) cos (nd) sin ? P^ (cos *>) dip dd
Mmn ' Mmn
* w+1
-0*W.TJ.
f • vm n Jo Jo' f(<P, 0) sin (n0) sin <p /*f (cos <p) d<p dd
n 9± 0.
Vmn Vmn '
.
As usual, when /is sufficiently well behaved this series is twice differentiable term-
by-term with respect to each variable, and the series
00 r a m
= m
u(r, (p, 6) 2_j r —2~ um0 + 2_^ (A mn umn + Bmn vmn )
m=0 L n=l
EXERCISES
r 2 R" + 2rR! - \R =
on (0, <x>).
(b) Prove that this equation has solutions of the form rk k a non-negative integer, ,
2
d F - 2 Jd n 2 dF
+ + =
W2 + k
(l s) ^ n(n \)F)
Solve this equation by the method of separation of variables, and compare the
results with those in the text.
To complete the discussion of the last section we must show that the spherical
harmonics
n
umn (<p, 0) = cos nd sin <p P%\cos <p),
(14-53)
n
vmn (<p, 6) = sin nd sin <p P£°(cos <p\
Q(S). We begin with the proof of orthogonality, which in this case requires
14-6 | ORTHOGONALITY OF THE SPHERICAL HARMONICS 569
that we show
umn • urs = whenever m 9^ r or n 9^ s (or both),
Recalling that inner products in Q(S) are computed as twofold iterated integrals
according to the formula
f'g = Jo
[^ f /(^ 0)S^> 0) sin * ^ ^>
Jo
( 14_55 )
we have
«mn '
Urs = [^ COS «0 COS 50 </0 /" sinn+S+1 <p P%\cOS <p)Pr (S
\cOS if) dip.
Jo Jo
1
2 n
(1 - x ) P^\x)P rn \x) dx =
{
0. (14-56)
J
To establish this equality we recall that P™ and PJ.
n)
satisfy the identities
(1 - x )P% +2)
2
- 2(» + l)xP% +1) + [m{m + 1) - n(n + 1)JP£° = 0,
- n+2) n+1) -
(1
2
x )Pr (
- 2(« + l)xPr {
+ [r(r + 1) n(n + l)]P<
n)
= 0,
(14-57)
throughout the interval [—1, 1]. We now multiply the first of these expressions
(1 - Jc ) Pr -
2 M (n)
by the second by , (1 x 2 ) nP%\ and subtract to obtain
- - - w+1)
+ x f[P£ +1)Pi
2 n) )
2(« 1)jc(1 P< />£ ]
= W+ l) - m{m + l)](l - x 2
)
n
P™P™
But by setting P^ +1) Pr ( ( n)
- P (n+1) i>£° =
r y, the left-hand side of this identity
may be rewritten
d^
[(1 - x 2 ) n+1 y],
dx
and we have
and since m 9* r,
1
- 2 n n)
(1 x ) P^Pl. dx = 0,
f
as desired.
To complete the proof we observe that the above argument applies equally well
to the v mnand shows that v mn -v rs =
, whenever m 9^ r or n ^ s. Finally, the
vanishing of u mn v ra for all m, n, r, s is an immediate consequence of the or-
•
For future use we take formal note of the following obvious consequence of this
result.
r m.. r m ii r Tn it i,rn
U m i, r^il
r u mm
' "mOi ' "ml » • • • j ' "mm, r i'i
. . . ,
are linearly dependent in 6(5). But this contradicts the known orthogonality of
these functions, and the corollary follows.
On the strength of Theorem 14-1 we are justified in introducing the formal series
expansion of a continuous function /on the unit sphere in terms of the umn v mn , ,
even though we cannot as yet assert that this series converges in the mean to /.
Using the notation of Section 8-3 we therefore have
- j m
(14-58)
2
m=0 L n=l
where A m o/2, A mn Bmn are the generalized Fourier coefficients of/ and are com-
,
Am o /• Ww o
2 Um Q ww0
'
(14-59)
J ' Umn = J ' Vmn
mn ~~ "mn
i* mn 'U
U Uffin
'
Vmn ' Vm n
14-6 I
ORTHOGONALITY OF THE SPHERICAL HARMONICS 571
This series is known as the Laplace series expansion of/, and in the next section
we shall prove that it does in fact converge in the mean to/. But first we evaluate
the various inner products u mn • u mn and v mn • vmn appearing in (14-59). The
argument goes as follows.
From (14-53) and (14-55) we have
2 2n+1 2
umn • umn = / cos nB dd / sin tp [/>£°(cos ip)] d<p,
Jo Jo
2 2n+1 2
Vmn ' Vj = / sin nddd sin <p [P^\cos <p)] d<p.
Jo Jo
Thus if
n
(1 - x 2 ) [P%Xx)] 2 dx = f sin^+VtP^cos^)] 2 ^,
Jf—i
-•ran
Jo
we see that
and it remains to find the value of Imn for given integers m and n (n < m). To
this end we observe that
_ r ,2
= -~
2
ImO =
— [Pm(x] dx
[Pm(x)Y
.
(14-61)
i0 j
/
Jm r X
(see Section 11-3), and then use integration by parts to express Im ,n+i in terms
of Imn by setting
- n+1 n+1)
u = (1 x2) /4 , dv = P% +1) dx
in
T +1 [P% +1 \X)]
2 2
/m..+l = J^O ~ X dx.
This gives
1Jm,n+l
—
— (\
Kl ~ xY 2\n+ln(n+l)n(n)|l
) rm *m |—
-\_\<\ - * )P%
2 +2)
~ 2(n + l)^n+1)](l - x
2 n
) P^dx
= -/.jK - * 2 1 A n+2)
- 2(» + l)^n+1)](l - x 2 ) nP™dx.
But since
(1 - x )/C +2)
2
- 2(« + l)xP% +1) + [m(m + 1) - fi( R + 1)]P<T> =
572 BOUNDARY- VALUE PROBLEMS FOR LAPLACE'S EQUATION | CHAP. 14
/m n+ i
,
= [m(m + 1) - n(n + l)]j_ (1 - 2
x f[P^\x)fdx
Hence
Im ,n+i = (m - n){m + n + l)4m, (14-62)
and an easy computation starting with the known value of Im0 now yields
T
/», =
(m
t
(m
+
—
— «)!
£r ~
n)\
x~t
2m +1
2
'
_ .
< n < m
.
/t .
(14-63)
,. N
Finally, if this result is substituted in (14-60) and (14-59) we obtain the formulas
r2ir rr
Am n = ^m +
l-K\m
l
^-
+
lH
n)\
/
Jo
/
Jo
f((f , d) COS (Jiff) Sin
w+1
<p /£>(C OS <p) d<p dd,
^+ /"2x /•«
EXERCISES
(b) by writing
f(<p, 8)
= \ cos
2
cp sin 2 <p sin 28,
2
Pk (cOS if).
d(cos <p)
4. Prove that
1 3
Pf (COS
•
6930
P^CCOS tp) + t^:
1540
*>) sin 30 sin (p
(b) Find the solution of the boundary- value problem of Section 14-5 when f(<p, 0)
6. Solve Laplace's equation in the spherical shell with inner radius a and outer radius
b (b > a), given that u =
for r = b, and u = f(<p, 0) for r = a. [Warning: Since
- (m+1)
the point r = is not in the shell, both of the functions rm and r must be used.]
7. (a) Use the results of the preceding exercise to find a solution for the shell problem
when u = for r = a, and u = g(<p, 0) for r = b.
(b) Solve the shell problem when u = f(<p, 0) for r = a, and u = g(<p, 0) for r = b.
given any function/in e(5), and any real number e > 0, there exists a finite linear
combination L(<p, 0) of the u mn v mn such that ,
for all < <p < ir, < < 2x. Indeed, the passage from such an approximation
to the assertion that the u mn , v mn generate e(5) is all but identical with the proof
of Theorem 9-6 and will be left to the reader. With this understanding, we turn
our attention to the following theorem.
Theorem 14-2. Any continuous function on the surface of the unit sphere
can be uniformly approximated by a finite linear combination of spherical
harmonics.
In broad outline, the proof consists of an argument to show that the spherical
harmonics can be replaced by certain polynomials in x, y, z, which are then used
574 BOUNDARY- VALUE PROBLEMS FOR LAPLACE'S EQUATION | CHAP. 14
Proof. We begin by rewriting umn and v mn as polynomials in cos 0, sin 0, cos <p,
we see that cos nd is the real part, and sin nd the imaginary part of the expansion
of (cos + / sin 0)
n
. But, by the binomial theorem,
+ n = n
+ n_1
(cos /'
sin 0) cos «i cos 0(/ sin 0)
n~ 2
+ a 2 cos 0(/ sin 0)
2
+ • • •
+ (/ sin 0)
n
,
Combined with (14-64), this implies that each of the spherical harmonics u mn ,
n- n m - n - 2i fc fc
a jk cos sin sin <p cos <p
where the c ijk are real numbers. Thus x 2 — 2y + xz and 2xy — 3yz z 2 are homo- +
2
for suitable constants a jfc and integers j and k. Thus the functions r m u mn and
m term appearing as
r v mn have a similar decomposition, with a typical
n -k n - fc k m - n - 2j cos m - n - 2y
= a jk r 2j[r cos"-* sin <p\r sin* 5 sin
fe
<p\r *>].
(14^66)
To complete the proof we recall that
oijkK.x -r y + 2 ) x y z ,
For our purposes this lemma is important because it allows us to assert that the
equation
2 2 2
d u d u d u n ,_
(14- 67)
,
S3 + a^ + aP = °
has polynomial solutions which are homogeneous of degree m for each integer
m > 0. Any such solution of Laplace's equation is called a homogeneous har-
monic polynomial of degree m, and the linearity of (14-67) implies that the set 3CW
of all homogeneous harmonic polynomials of degree mis a real vector space.
Moreover, it follows from Corollary 14-1 that the 2m + 1 functions
mu m mm m
r mQ) r u ml , ..., r Umn, r vml , ..., r v mn (14-68)
u(x,y,z)=
i+j-\-k=m
^ c ijkx
i j h
yz (14-69)
2 2 2
d u d u d u
dx*
+ dy*
~*~
dz*
2 2 2
d u d u d u
dx~*
+ dy*
+ dz*
If this array is examined closely, it soon becomes apparent that the entires along
the two enclosed diagonals on the upper right determine, via (14-71), all of the
other coefficients in the triangle. Indeed, the coefficients along the uppermost
diagonal determine those along the third, which in turn determine those along
the fifth, etc., while the coefficients along the second diagonal determine, succes-
sively, those along the fourth, sixth, and so forth. Thus two harmonic polynomials
in 5C m are identical if and only if the 2m +
1 coefficients appearing along the upper
coefficients c m - k ,p,y, c m - k -i,p,y zero, and let Rk , 1 < k < m, denote the poly-
14-7 | HARMONIC POLYNOMIALS AND THE BASIS THEOREM 577
In this case an easy computation using (14-71) shows that the particular homoge-
neous harmonic polynomials constructed above are
Thus
0o = *
3
- 3xz 2 , J\\ — X Z — jj z ,
Qi = x 2y — yz 2 , R2 = xyz,
Q 2 = -xz + xy
2 2
, Rs = ~iz 3 + y
2
z,
Q 3 = -3yz 2 + y 3 ,
578 BOUNDARY- VALUE PROBLEMS FOR LAPLACE'S EQUATION | CHAP. 14
and the preceding lemma implies that these polynomials are a basis for the space
3C 3 Using them we can easily construct all homogeneous harmonic polynomials
.
of degree 3 in x, y, z.
With these hand we are now in a position to prove that every con-
results in
tinuous function defined on the surface of the unit sphere can be uniformly
f(cp, 6)
approximated by a linear combination of spherical harmonics (Theorem 14-2).
Once again we use the Weierstrass approximation theorem, which in this context
asserts that every continuous function on the closed unit sphere r < 1 can be
uniformly approximated by polynomials* Thus, if e > is given there exists a
polynomial p(x, y, z) such that
also valid
1
a=0
we have dim V = (m + l)
2
.
EXERCISES
2. (a) Use the technique of Example 1 to find a basis for the space 3C 2 of homogeneous
harmonic polynomials of degree 2.
(b) Determine which of the following polynomials belong to the space 3C2:
x{y - 2z) + y(x - 2z) + <x - y), y(y - 2x) - x(x + z).
x3 — xz 2 — 2xy 2
is harmonic, and express it as a linear combination of the basis for 3C3 found in
Example 1.
^2 = 3V 3r 3,2.
5. Express each of the functions
in terms of x, y, z.
6. Let Vn denote the vector space of all harmonic polynomials of degree n (n > 2) in
x and y, and let each of the polynomials in V„ be written in standard form as
n
n ~k
P(,x, y) = 22 c n-k,kX y .
Jfc=0
(a)Prove that dim V n = 2, and that the real and imaginary parts of (x + iy)
n
7. Use the result of Exercise 6(b) to find the value of each coefficient in P(x, y) in
terms of the coefficient cn ,oorcB _i,i.
9. (a)
contains the term x 6 but no term in x 5y.
(b) Repeat (a) for the polynomial which contains the term x 5y but no term in x 6 .
10. Prove that the number of terms with nonzero coefficients in a homogeneous har-
monic polynomial in x and y of degree n > 2 is n/2, (n/2) 1, or n 1 if n + +
is even, or (n l)/2 or n + +
1 if n is odd. (See Exercises 6 and 7.)
11. Use the technique introduced in this section to find a basis for the space 3C4-
12. (a) Prove that if f(x, y, z) is a harmonic function in a sphere (or in all of three-
space), then so is the function g(x, y, z) = /(— x, y, Use this result to show that
z).
for fixed values of any two variables, the even and odd parts of f(x, y, z) with
respect to the third variable are harmonic whenever /is harmonic.
(b) Use the above result to prove that whenever a polynomial in 3C2 is even (or odd)
with respect to a particular variable, it is a linear combination of those members
of the standard basis for 3C2 which are even (or odd) with respect to that variable.
(c) Show that a polynomial in 3C2 cannot be odd with respect to each of the three
variables.
2
r u 2 ,n (« = 0,1,2),
2
r v2 n,
(" = 1,2)
in terms of the basis for 3C2 found in Exercise 2(a). [Hint: Use the properties noted
in Exercise 12.]
(b) Use the result in (a) to express each of the basis polynomials found in Exercise
2(a) in terms of the basis consisting of r U2, n (n = 0, 1, 2) and r V2, n (n = 1, 2).
2 2
14. Let P(x, y, z) be an arbitrary polynomial in 3C3. Prove that dP/dx belongs to 5C2,
and express this polynomial as a linear combination of the basis for 3C 2 found in
Exercise 2(a).
*15. Prove that the number of nonzero terms in Pm (xi, X2, • • • , xk ), the general homo-
geneous polynomial of degree m > 1 in k > 1 variables, is
(m + k - l\
\ k-\ )
[Hint: For k > 1 write
Recall that
-(,:,) Mi - y)'.
.
c(mi, . . . , m k _u m k + 2)
k ~X
E {rrii
7
— +—-1)
+—2)(m; c(mi, . . . , m,_i, im + 2, m t+ i, . . . , mk ).
i=1
(m fc + + 1)
2)(/llA;
the polynomial in which xk appears to the first degree, and suppose that Pm is har-
monic. Prove that the coefficients of Pm are uniquely determined by the coefficients
of Q m and Q m -\.
*17. Let X
denote the set of all homogeneous harmonic polynomials of degree m > 1
in k >
2 variables xi, xk with the property that each polynomial in 9C con-
. . ,
tains exactly one term (and that with coefficient 1) for which the exponent of xk is
less than 2. Prove that 9C is a basis for the space 3C m k of all homogeneous harmonic ,
*18. Can the following assertion be deduced from the results of this section?
Any harmonic function in the closed unit sphere r < 1 can be uniformly approxi-
mated by harmonic polynomials.
Why?
19. Use Theorem 14-2 to prove that the spherical harmonics are a basis for the Euclidean
spaceC(5).
15
boundary-value problems
involving bessel functions*
15-1 INTRODUCTION
After the succession of examples in the preceding chapters the student has probably
come to the conclusion that the techniques we have developed are adequate to
solve any boundary-value problem involving the wave equation, heat equation, or
Laplace's equation, at least when the underlying region and boundary conditions
are reasonably simple. This, however, is false, as can be seen by considering
Laplace's equation in cylindrical coordinates (r, 6, z), where x = r cos 6, y =
r sin d, z = z. For then
2 2
d u d u d^u _ -
dx*
+ dy 2
+ dz 2
- u
becomes
^ + 1^+1^
dr 2 ^ ^^=
+ r dr r2 dd 2 dz 2
0-
' U
(15-1)}
and if we attempt to find solutions of (15-1) which are independent of 6 and have
the form u(r, z) = R(r)Z(z), we see that R and Z must satisfy the equations
Z" - X
2
Z= (15-2)
and
r
2
R" + rR' + 2
\ r
2
R = 0, r > 0. (15-3)
2
(The motive behind writing the constant as —X will become clear later.)
Although the first of these equations can be handled with ease, the second
causes trouble because its leading coefficient vanishes at r = 0. Here the imagina-
tive student might suggest using the power series method to solve (15-3) about
r = a, a > 0, since its coefficients are analytic whenever r ^ 0. Unfortunately,
this willnot do, for the simple reason that the resulting series need not converge
outside the interval (0, 2a), whereas we seek solutions valid for all r > 0. (See
Theorem 6-4.) Clearly, then, if we are to make any progress in solving this prob-
lem, or others of the same ilk, we must devise a method for studying the solutions
of a linear differential equation near points where its leading coefficient vanishes.
This is precisely what will be done in the sections which follow, where we introduce
the celebrated method of Frobenius generalizing the power series technique to a
large number of non-normal has been done,
linear differential equations. Once this
we shall use the technique in question to study the solutions of Bessel's equation
and the last important class of elementary boundary-value problems.
EXERCISES
1. Use the coordinate transformations given above to show that Laplace's equation
becomes
d u 1 du 1 d u d_u _
in cylindrical coordinates.
2. Assume that Laplace's equation in cylindrical coordinates has a solution of the form
u(r, z) = R(r)Z(z), and show that R and Z must satisfy Eqs. (15-2) and (15-3).
r 2 R" + rR' + \ 2r 2R =
p( x ) = 2 a^x ~ x °)
k
>
valid in some interval about the point x Moreover, since p(x ) = 0, the leading
.
coefficient in this series must vanish, and there exists a positive integer m such that
m —m
p(*) = 2
A;=m
ak ( x ~ *°)
fc
= (•* ~ *°) 2
k=m
° k (x ~ *°)
k
- (x - XoTpiix),
.
where analytic at x
/>i is and/?i(x ) 5^ 0. We now divide the
, coefficients of
(15-4) by pi to obtain an equation of the form
pendent upon the exponent m and the value of q x at x and leads to a classification ,
where q and r are analytic at x . All other singular points are said to be
irregular.
x\x + l)
3
(x - 1)/' + xy' - 2y = 0, (15-7)
and
are undefined at x = —1
As the terminology suggests, regular singular points are relatively easy to handle.
In fact, with appropriate modification, the method of undetermined coefficients
introduced in Chapter 6 can be used to obtain a basis for the solution space of
any second-order linear differential equation about a regular singular point. Not
. :
so when x is an irregular singular point. Here the situation is far more com-
plicated, and it is not difficult to exhibit equations which fail to have series solu-
tions of any form about such a point (see Exercise 11 below). Fortunately, all of
the boundary-value problems we shall encounter involve equations whose only
singularities are regular, and hence we shall have no more to say about irregular
singular points.
Finally, for simplicity, we shall limit the following discussion to singularities at
The reader should appreciate that this involves no loss of generality since the
EXERCISES
Find and classify all of the singular points for the equations in Exercises 1 through 10.
1. x 3
(x 2 - 1)/' - x(x + 1)/ - (x - \)y =
2. (3x - 2)
2
xy" + xy' - y =
3. (x 4 - \)y" + xy' =
4. (x + 4
l) (x - l)
2/' - (x + l)
3
(x - 1)/ + y =
5. x 3
(x ix - 1)/
- l)y" + + 2xy =
6. x (x + 1)V
3 - y =
7. x(l - x)y" + (1 - 5x)y' - Ay =
8. Legendre's equation:
x 2y" + xy' + (x 2 - \ 2 )y =
10. Laguerre's equation
xy" + (1 - x)y' + \y =
1 1 Prove that the equation
x *y" + y =
has no nontrivial series solution of the form
ak x
T
x 2y" + axy' + by = 0,
a and b (real) constants, studied in Chapter 4. At that time we proved that the
solution space of this equation is spanned by a pair of functions ji and y 2 con-
structed from the roots v 1} v 2 of the equation
v(y - 1) + av + b = 0,
as follows. If Vi 9^ v 2 ,
if ^i = v2 = v,
yiix) = \x\
v
, y 2 (x) = \x\'ln\x\*
Save for the fact that yi and y 2 here appear in closed form, these expressions are
typical of the solutions of second-order homogeneous linear differential equations
with a regular singular point at the origin. Indeed, we shall find that about x =
the solution space of such an equation is always spanned by a pair of functions
which depend upon the roots of a polynomial equation of degree two, that these
functions involve powers of \x\, and under certain circumstances a logarithmic
term as well. The following example will illustrate these remarks, and introduce
the technique which is used to handle the general case.
y(x) = x Y, v
fc=0
«*** = E
k=0
akxk+ "> (15_9)
then
a = a
y x {x) = \x\ sin (j8 In |*|), y 2 (x) \x\ cos (j3 ln \x\).
15-3 | EXAMPLES OF SOLUTIONS ABOUT A REGULAR SINGULAR POINT 587
k+v - 2
/'(*) = 2
k=0
(k + v\k + v - l)a k x ,
E
fc=0
(* + *)(* + v - l)a k x
k+v
+ 2
fc=0
(* + ^)^fc+ +1 "
- 1 E
fc=0
( fe + ") a **
fc+
"
+ \ E^
fc=0
fc+
" = °-
But since
E
fc=0
<k + v)« fc
/+" +1 =
E
fc=l
(* + " - iK-i^n
(15-11)
(Recall that we assumed a ^ 0.)
Equation (15-10) determines the admissible values of v for this problem as \
and 1, and is known as the indicial equation associated with (15-8). We now set
I{v) = v{v - 1) - \v +h
and rewrite (15-10) and (15-11) as
I(y) = 0, (15-12)
and
I(k + v)a k + (k + v - l)« fc _i = 0, k>\. (15-13)
588 BOUNDARY- VALUE PROBLEMS INVOLVING BESSEL FUNCTIONS | CHAP. 15
1
Hk = — T Qk—1-
•
Thus
01 = — «0,
«0
00
" 3= - 3!'
fc = (-1) fc0O
k\
Similarly, when v = 1,
0fc = — 0fc-l3
2A: + 1
and
fli = -§0o,
02 = 57300,
03 = - 7-5-3 0o,
ak = (-1)*
- 0o-
(2A: + l)(2fc l)---5-3
We now set a = 1 and substitute the above values in (15-9) to obtain the two
series
k
kX_
yi(x) = * 1/2 X;(-i> (15-14)
k=0
and
]
{2x)
y 2 (x) = xj^i-lf - (15-15)
k=0
(2k + 1)(2A; 1) • - •
5 • 3
Leaving the latter point as an exercise, we establish convergence by using the ratio
test with
,ft+3/2 -fc+1/2
Pi
(*+ 1)! k\ k+ 1
15-3 | EXAMPLES OF SOLUTIONS ABOUT A REGULAR SINGULAR POINT 589
for all *, and the series do converge as desired. Thus the general solution of
(15-8) on (0, oo ) is
X*) - C ,M'"
£ (~i?E + ^ £ (-'f g* +\f-
fc=0
c
fc=0 v ' 5 3
'
EXERCISES
1. Prove that the functions y\ and y 2 defined by (15-14) and (15-15) are linearly inde-
pendent in 6(0, oo ). [Hint: Consider the behavior of ciy[(x) c2 y2 (x) as x —> 0.] +
2. The equation
x 2y" + x(x - 4)/ + y =
has solutions which are defined for all x. What are they ?
3. Find the indicial equation associated with the regular singular point at x = for
each of the following equations.
(a) x 2y" + xy' - y =
(b) x 2y" - + 1)/ + (* -
2x(x l)y =
(c) x 2 /' -
2xy' + y =
(d) x 2y" —
jcv' + (x
2 —
X 2 )y = 0, X a constant
(e) xy" + (1 — x)y' + Xy = 0, X a constant
5. Find the indicial equation associated with each of the regular singular points x = 1
Use the method introduced in this section to find two linearly independent solutions
of the equations in Exercises 6 through 10. In each case verify that the solutions obtained
converge whenever \x\ > 0, and that they are linearly independent in 6(0, °o) and
C(-oo,0).
xy" + (1 - x)y' + \y = 0,
has a solution which is analytic for all x, and which reduces to a polynomial when
X is a non-negative integer.
In this section we shall indicate how the technique introduced above can always
be used to find at least one solution of
00
y(x) = x v
2
fc=0
akx (15-17)
with a 9^ 0. Then
~\
/(*) = *"
E
k=0
(* + v)a k x
k
00
00 00 00
2
fc=0
(k + */)(& + v - \)a k x
k
+ <?(*) J]
k=0
(k + *>Hx
fc
+ K*) £
k=0
a*** = 0.
(15-18)
)
q(x) = E
fc=0
qk x\ r(x) = E
k=0
r kx\ (15-19)
where both series converge in an interval \x\ < Ro, Ro > 0, centered at x = 0.
+ ^
00 / 00 v / 00 \
E
fc=0
(* + "X* + * - i**** + (
\/e=0
E ( fc x
)
/ \fc=0
( E ****
/
+ (£>***)(£ '***) = o,
\fc=0 / \fc=o /
and if we now carry out the indicated multiplications according to the formula
given in Section 6-6 we obtain
E
00
fc=C L
fc=0
r
( fc + ^+ v -!>»*+ E
k
fc
i=0
0" + v)ajqk _j +
J=0
k
fc
E W-i
Hence (15-17) will formally satisfy the given differential equation in the interval
< x < R if and only if
fc
v{v - 1) + ^+ r = (15-21)
(15-22)
The first of these relations is known as the indicial equation associated with (15-16),
and its roots, which determine the admissible values of v in (15-17), are called
the characteristic exponents of that equation. We direct the reader's attention to
the fact that since q and r are, respectively, the constant terms in the series
expansions of q and r, (15-21) may be rewritten
(cf. Eq. (15-19)). Thus when q and r have been explicitly given, the indicial equa-
tion can be obtained directly from (15-16) without undertaking the above
computations.
_
I(v) = v{y - 1) + q Qv + r ,
and let v x and v2 denote the roots of the equation I(v) = 0. Moreover, for con-
venience we suppose that v x and v2 have been labeled in such a way that
Re (>i) > Re (v 2 )* Then, when v = v u (15-22) becomes
k-i
I(k + v x )a k + 2
y=o
tO' + "i)?*-i + /*— y]flt y = 0, (15-24)
and, by the way in which v\ was chosen, we know that I(k + v{) 9^ for k = 1,
1
_
—
ak = ~ jfk , „ x J2 [C/ + ^Oflfc-i + '
,
*-y]fly» ^ > 1, (15-25)
with this we have succeeded in producing a. formal solution of Eq. (15-16), valid
in the interval < x < R Moreover, if xv is replaced by \x\ v throughout these
-
computations, this result obviously holds in the interval — R < x < Oas well.
Finally, it can be shown that the series obtained in this way always converges if
< |jc| < -R -t Hence the function
y(x) = |xp 2
fc=0
a ^ k
(15-26)
fc-i
I{k + v 2 )a k + 2
i=o
tO" + v 2 )qk-j + r k -j}aj = 0, (15-27)
which can again be solved for a k provided I(k + v 2 ) 9^ for all k > 1. But
when k > 0,
I(k + v 2) =
y(x) = WZ 2
fc=0
a ^ fl o
* °» ( 15
" 28
>
valid for < |jc| < R whenever the roots of the indicial equation I{v) = do not
differ by an integer. In this case it is easy to show that the (particular) solutions
where c x and c 2 are arbitrary constants. This, for instance, is precisely what
happened in the example given in the preceding section.
EXERCISES
Find two linearly independent solutions about x = for each of the following equations,
2* 2
5. 3x 2/' - ~y' + 7>> =
x — 1 x — 1
+—
x(7x + 1) 1
10. 2x 2y" + T-~y' 7T,y =
+ X +
AC 1 1
Compute the values of the coefficients a\, a 2 , as in the series solutions of each of the
following equations. (Assume a q =1.)
11. x 2y" + x(x + 1)/ + y =
12. 16a:V - 4x(x - 4)/ - = 2
>>
13.
2
x (x 2 - \)y" - xy' - 2y =
594 BOUNDARY-VALUE PROBLEMS INVOLVING BESSEL FUNCTIONS | CHAP. 15
*17. (a) Show that x = 1 is a regular singular point for Legendre's equation
fc=0
teger. Our experience with the Euler equation suggests that a solution involving
a logarithmic term should arise when Vi = v 2 and, as we shall see, this can also
,
happen when v x
9^ v 2 . The following theorem gives a complete description of the
situation, both in the general case treated above, and in each of the exceptional
cases.
and suppose that Vi and v 2 have been labeled so that Re (vi) > Re (y 2 ).
Then (15-29) has two linearly independent solutions yi and y 2 valid for ,
y x (x) = \x\
Vl
2
fc=0
ak x
k
, a = 1,
y 2 (x) = \x\
V2
J2 b k x\ b = 1.
fc=0
1
Case 2. V\ = v2 = v. Then
00
k
y^x) = \x\" ^2 akx , a = 1,
fc=0
hW = \x\
v
2
fc=i
**** + JiWln |x|.
yi(x) = \x\
n ]T a k xk > oo = l,
A;=0
k
y 2 (x) = \x\
V2
J] hx + cji(x) In |x|, bQ = 1, c a (/ix«f) constant.*
k=0
In the remainder of this section we sketch the argument leading to the solution
involving a logarithmic term when v x = v 2 The reasoning in the case where .
x
v
k=0
^ <>kx
k
(15-30)
so that the resulting expression satisfies (15-29) for < x < R . This time,
however, we also regard v as a variable, and write
Moreover, we assume from the outset that a = 1. Then if L denotes the linear
differential operator x D
2 2
xq{x)D r(x), we have + +
—
Ly(x, v) = I{v)x
v
+ x Yj
v
oo
k=l
W
/
'
+ v)a k +
k
]T
j=0
[(J + v)q k _ j + n^aA x k
\
l
,
(15-32)
where the q k _j and rk _j are the coefficients in the power series expansions of q and
r about the origin. [See (15-21) and (15-22).] We now use the recurrence relation
to determine a l5 a2 ,
. . . in terms of v in such a way that every term but the first
y x {x, v) = x" i + E
k=l
a ^ xk (15-33)
Lyi(x, v) = I{v)x
v
. (15-34)
(Recall that, by assumption, a = 1.) But since v x is a double root of the indicial
equation I(y) = 0, I(v) = (y — v x)
2
, and (15-34) may be written
Lyi(x, v) = (y — vl fx
v
(15-35)
But since
d
yi(x, v) =
dv
k
y2(x,v 1 ) = ^-yi(x,v) = xVl J2 a 'k(yi)x + yi(x, vO In x,
fc=l
which is precisely the form of the second solution given in the statement of
Theorem 15-1 under Case 2.
15-6 | BESSEL'S EQUATION 597
EXERCISES
Find two linearly independent solutions on the positive x-axis for each of the equations
in Exercises1 through 10.
1. x 2y" + x(x -
+ (1 -
1)/ x)y = 2. xy" + (1 - x)/ - y =
3. x 2y" 3xy' + (x + l)y =
+ 4. x 2y" + 2x 2y' - 2y =
11. Use Theorem 15-1 to determine the form of two linearly independent solutions
12. Prove that the solutions y\ and j>2 given in Cases 2 and 3 of Theorem 15-1 are
independent in 6(0, Rq) and C(— Rq, 0).
linearly
d
[Ly(x,v)] = L
av ov
x 2y" + xy + (x
2
- p
2
)y = (15-36)
about the point x = under the assumption that p is real. We recall (see Sec-
tion 6-2) that (15-36) is known as BesseVs equation of order p, and, as we shall
see, it arises in the study of boundary-value problems involving Laplace's equation
and the wave equation.
Since the indicial equation associated with (15-36) is
„2 - p2 == 0, (15-37)
and has roots ±p, Theorem 15-1 guarantees that Bessel's equation of order p
possesses a solution of the form
00
k
yi(x) = xp ]T akx , p > 0,
k=0
598 BOUNDARY- VALUE PROBLEMS INVOLVING BESSEL FUNCTIONS | CHAP. 15
(*
2
- p
2
)yi (x)
= xp £
k=2
ak _ 2 x
k
- xp JP
fc=0
2
a k x\
00
xyi(x) = x p J] (k + />>****,
fc=0
00
*Vi'(*) = *P 2
fc=0
(fe + />)(* + /» - l)fl*x*.
A:=0 fc=2
or
00
k
(2/? + l) fll x + J] L*(2/> + *>** + a k _ 2 ]x = 0,
fc=2
fli = o3 = fl 5 = • • • = 0,
a
a2 = -
2(2p + 2)
a* ~ 2 •
4(2/> + 2)(2/> + 4)
and, in general,
= (—1)
«2fc
2 • 4 • 6 • • •
(2k)(2p + 2)(2p + 4) • • • (Ip + 2fc)
= 1c-n} & ^
2**k\(p+ 1)Q> + 2)---Q> + *)
^ ^
Hence
J". W- g ^ «•
2 !(f, + 1)( r+2)...( P + fe )
1
a =
V>T{p + 1)
15-6 BESSEL'S EQUATION 599
T(p) = f\- f- l x
dU p > 0. (15-39)
Jo
It can be shown that this integral converges for all p > 0, diverges to + oo when
p = 0, and has the values
r(D = l,
T(p) = T(p + 1)
(15-40)
read as the definition of T(p) for — 1 < p < 0, since T(p + 1) is already defined
in that interval. This done, we use (15-40) and the values just obtained to extend
the definition of T(p) to the interval -2 < p < - 1. Continuing in this fashion
we obtain a real-valued function defined for all values of the independent variable
p, save p = 0, — 1, —2, The graph of the resulting function is shown in
Fig. 15-1.
I
FIGURE 15-1
2k+p
MX) "
S ( 1)
*2"+'A:!(p+ !)(/> + 2) (p + k)T(p + 1)
'
600 BOUNDARY- VALUE PROBLEMS INVOLVING BESSEL FUNCTIONS CHAP. 15
1.0
0.9
0.8
\Jo
0.7
0.6
k<>i
0.5
sJ2
0.4
0.3
0.2
0.1
A Y .8 y5.1 A.i 7.C / 8.4v G.7
] I \ 3 ( 7
V \
10
0.1
0.2
-0.3
-0.4
FIGURE 15-2
oo / i\fc /J\2k+p
(-1V -,©' (15-41)
fc=0
r(fc + \)r(p + k + l
(15-43)
* Note that
Case 1. p >2p not an integer.* Here the roots of the indicial equation asso-
0,
ciated with (15-36) do not differ by an integer, and a second solution can be
obtained by repeating the above argument with —p in place of p. Obviously this
will lead to a series whose coefficients have the same form as before, and since the
gamma function is defined for nonintegral negative values of its argument, the
solution in question can be written
for x > 0. (For negative values of x we must replace x~ p by \x\~ p . From now
on, however, we shall restrict our attention to the positive x-axis.) we
Finally,
observe that (15-44) is defined even when p is of the form n +
J, n an integer,
and again yields a solution which is linearly independent of Jp (See Exercises .
6 and 1 1 below.) Hence we conclude that the general solution of BesseVs equation
of order p is
and its indicial equation has zero as a repeated root. Hence, by Theorem 15-1,
we can find a second solution of the form
00
K {x) = Yj b kX
k
+ / (*)ln*, (15-46)
k=i
00
~1
xK (x) = ^
fc=3
b k _ 2x
k
+ xJ (x) In x,
k ~l J (x)
Ko(x) = J2 kb k x + 7o(*)ln* +
k=l
~X - J (x)
xK'qXx) = Y, k ( k ~ x )°kX
k
+ xJK(x)\nx + 2J' (x)
k=l X
* Note that the difference of the roots of (15-37) be an integer and only
will if if p is
an integer or half an integer.
602 BOUNDARY- VALUE PROBLEMS INVOLVING BESSEL FUNCTIONS |
CHAP. 15
2
bi + Ab 2 x + ]T [k b k + bk^x*' 1
fc=3
2 _1
bx + Ab 2 x + Yj t* ** + ^-2l^ = -2^'oW-
fc=3
Finally, by (15-42),
( \ — V* ( n fc
^ v 2/c-1 '
'oW
/
- Z-< v ^ 2 2k (W) 2
whence
fc=3 fc=l
2k+1
bix + E [(2* + lVW+i + b 2k ^]x + 4b 2 x
2
00 °°
4^
+ £ [( 2*yw + 6 2fc _ 2 ]x
2fc
- x
2
+ E (- 1 >
fc+1
2^^
Thus Z>i = 63 = b5 = • • • = 0, while
46 2 = and (2k) b 2k
2
+ ^-2 = (-1)* +1 ' k > L
1,
22^)2
Hence
. 1
b* = 22'
^4 = — 42 C1 + 2) = — 24(2! )2
^ ~^" 2)'
22 .
*oM =
&=1
Z^(l + \ + • + j[)(f)" + •«*) in*
15-6 BESSEL'S EQUATION 603
0.6
*o
0.5
Yt
0.4
0.3
0.2
0.1
/ .
if J 4 i
V ( 7, Ii J 9 10
0.1 /
0.2 /
3
/
0-1
0.5 /
06 1 /
FIGURE 15-3
lnj + 7
(15-48)
**** + "
*»(*) = E
k=0
+ cJn (x) In x,
where c is a constant. Here too the b k and c can be evaluated by the method of
undetermined coefficients, but the argument is now exceptionally long and involved.
Fortunately, we shall not have to use these functions in any of our later work,
i + EO + m^)
604 BOUNDARY- VALUE PROBLEMS INVOLVING BESSEL FUNCTIONS |
CHAP. 15
a -n
*»(*) = - £
n-l
E
k=0
(n - k - 1)!
(i)
2k+n
-m
+ .40) In x,
where Hn = 1 + \ + + l/«.
In* + 7
x fc=i
fc!(/i + A:)!
EXERCISES
r(p + 1) = pT(p).
2 - (x+ »
TO)] = 4/ Jo
)
c rfx^,
Jo f
5. Prove that
Jo(x) = -Ji(x).
Jx
6. (a) Prove that the functions Jp and J_ p are linearly independent in 6(0, oo) for all
nonintegral values of p.
(b) Show that the function Y defined in the text is a linear combination of J
and Kq.
7. Prove that
/_„(*) = (-l) n/„(x)
for all integers n.
15-6 |
BESSEL'S EQUATION 605
is
y = xaZa (j3x),
Ji/2(x) = x
\TX
—
sinx.
VxZi /4 (y
j
10. (a) Prove that the function y = eax Zp (fix) is the general solution of the differential
equation
/- + (i_ 2B y+ a. + ^_|_^),.o
) (
whenever Zp is the general solution of Bessel's equation of order p, and a and /3
*11. (a) Prove that the Wronskian of Jp and 7_ p satisfies the differential equation
-^[xW(Jp ,J_ p )] = 0.
T(l - p)T(p)
whenever p is not an integer. (Note that this argument also provides a proof of the
fact that Jp and /_ p are linearly independent when p is not an integer.)
606 BOUNDARY- VALUE PROBLEMS INVOLVING BESSEL FUNCTIONS |
CHAP. 15
Now that we have the series expansions for Jp and Yp we are in a position to ,
derive a number of important formulas involving Bessel functions and their deriv-
atives. The first two are immediate consequences of (15-41), and read as follows:
J£
[*%(*)] = A-i(4 05-49)
+ p + 1)
X
2k+2 P
_ ^ (-lf2(p + k) 2k+2p -i
^2 2 k+Pk\T(k
+p+ 1)
_ (~0 2k+(p-l)
- x p Y^
Lj 22k+ P -ik\T(k + p)
= x /p_i(x).
This proves (15-49) and, with obvious modifications, (15-50) as well. A similar
pair of formulas holds for Yp but we omit the proof.
,
When the derivatives appearing in (15-49) and (15-50) are expanded, these
formulas become
xJ'p + pJp = xJp -i, (15-51)
and
xJ'p — pJp = — x/p+i, (15-52)
Theorem 15-2. The Bessel functions of the first kind satisfy the recurrence
relations
xJp+1 - IpJp + x/p_i = 0, (15-53)
and
Jp+i + 2J'p - Jp-i = 0. (15-54)
(15-41) becomes
But since
*_ X) (-1? „2fc
A/2W = ,-
---( 2/c
V2r(f) t'o
fc
2fc/c!3 -
5 + *>
2-
+ 2-4-3-5 2-4-6-3-5-7 +
V2r(|) L
3
1 r _^ jc
5
x7
^ ^
rm L
V2xT($) 3! 5! 7!
1
sin x.
'2*r(f)
r(f) = Vt/2,
whence
Ji/2(x) = a
\ TX
/ — sin x. (15-55)
and it now follows from (15-53) that every Bessel function of the first kind of
half-integral order (i.e., of order n + \, n an integer) can be expressed in finite
form in terms of elementary functions. For instance,
2_ smx
cos a:
TX
3 sin x 3 cos x
— sin a;
theorem we can even assert that the magnitude of their oscillations decreases with
increasing jc. In short, the graphs of Jp and J_ p (or Yp ) have a damped oscillatory
character; a fact which was borne out in the case of Jn +i/2 b Y the results of the
preceding example. Actually, this rough description of the behavior of Bessel
functions can be made much more precise. For instance, it is not too difficult
to show that every solution of Bessel's equation of order p can be written in the
form
y(x) = ^sin(x
x
+ <) + ^
r
y/x
) -
where A p and wp are constants whose values depend upon p, and rp is a function,
again dependent on p, which is bounded as x — > oo Thus, for large values of x,
.
Lemma 15-1. The zeros of Jp and Jp+ \ are distinct, and alternate on the
positive x-axis.
Proof. By (15-52),
Thus if Jp (x ) and Jp+1 (x ) were to vanish for some x > 0, Jp (x ) would also
15-7 I
PROPERTIES OF BESSEL FUNCTIONS 609
vanish, and the uniqueness theorem for initial-value problems involving second-
order linear differential equations would then imply that Jp = 0. This is nonsense,
and it therefore follows that the zeros of Jp and Jp+ 1 are distinct.
To complete the proof, let Xi < X 2 be consecutive positive zeros of Jp . Then
by (15-52),
7p+ i(Xi) = -7p(X0 and >/p+i(X 2 ) = -JP (X 2 )-
But, by assumption, Jp(\{) and «/p(X 2 ) have opposite signs, and the above equalities
then imply thatJp+ \ must vanish at least once between Xi and X 2 A similar .
argument using (15-51) shows that Jp must vanish between consecutive zeros of
Jp+u and we are done. |
has infinitely many zeros on the positive x-axis for all values of p and all
constants a and /?.
we have
^(Xi) < 0, y^(X 2 ) > 0, 7^(X 3 ) < 0,
EXERCISES
4. Prove that
1/2W =
J-1/2M */
\\-
\ TX
.'
— cosx
71
5. Show that
and that
c an arbitrary constant.
610 BOUNDARY- VALUE PROBLEMS INVOLVING BESSEL FUNCTIONS | CHAP. 15
6. Prove that
Mx)dx = Ji(x) + ^ X
+ ^/
X
W+ 3 • •
•
7. Prove that
00
/ Jn (x)dx = 1
Jo
for all integers n > 1. (Remark. This result is also valid when n = 0.)
- =
[JP -i(x)f [JP+ i(x)f
^ £.\J p (x)f.
r it yi
2 V^ 2(/> + 2fc) rf 2
k—0
11. Show that
2
k
Jn\x) = Jn-k(x) - kJn -k+2<,x) -\
&(& —— 1)
Jn-k+^x) + • •
•
k
+ (-l) Jn +k(x),
where
,m-
JnXx)
-
= ^ d
/»(*).
15-7 |
PROPERTIES OF BESSEL FUNCTIONS 611
— [x%(x)J q (x)]
= (t + p + q)x
l
~ J (x)J (x)
p q
- x
l
[Jp (x)Jq+ i(x) + Jq (x)Jp+ i(x)],
(b) Add the two identities in (a) and then integrate to deduce that
/p+l(x) g+l(x)
+ {[(/> + 1) + (<7 + D]/ / ^ - /p+1 (x)7g+1 W}.
(b) Iterate the result in (a) to obtain the formula
{p + q)\
J
-WW x
dx = 2^Jp+k (x)Jq+k (x) - Jp (x)Jq (x)
fc=0
Jp+n(x) +n(x)
- Jp+n (x)Jq+n (x) + + + 2n)j
(p q
^ dx.
V^£ dx=-± 2
- [Ux)f +
j {[/o(*)] 2
J^ [Jk (x)]j
16. Discuss the behavior of the positive zeros of the function J".
'
1 8. (a) Prove that the Laplace transform of the function x pJp (\x) is given by the formula
p
£[xpJp (\x)] = \ r(2 P + i)
2PT(p + 1)(*2 + X2)(2 P +l)/2
612 BOUNDARY- VALUE PROBLEMS INVOLVING BESSEL FUNCTIONS |
CHAP. 15
for all non-negative values of x and X and all non-negative integral p. [Hint: Show
that y = x pJp (\x) is a solution of the equation
xy" + (1 - 2/>)/ + XV = 0,
— — +
d£[y]
- (1 + 2p)
s ds
= 0.
£[y] s2 + X2
Solve this equation for £[y], and complete the proof by evaluating the constant of
integration.]
X"
£[/„(X*)] = + \2)V2]n
'
(52 + X 2)l/2[j + ( j2
19. (a) Use the convolution theorem and Exercise 18(a) to show that
by Laplace transform methods, and then use the result to deduce that
x
f
/ sin (t — X)7o(X) d\ = xJi(x).
Jo
certainly the most natural way of approaching the study of Bessel functions of
integral order, it is also possible to define these functions by means of a generating
function G(x, t), in which case Jn (x) appears as the nth coefficient in the series
and use the well-known power series expansion for the exponential function to
write
X _X_ 2 _*!_
xt/2
{
,
'2
, .
. . ,
f ,
2 2 2! 2 n n\
e
—x\2t _
- ,
1 —
X
^
i ,
+222!'
X 2 _ . . .
+ i /_
C
1\ n
1)
— t~
^
n 4-
+ • • • •
2 2B|l!
15-8 |
THE GENERATING FUNCTION 613
Hence
n=0 J ln = J
and since each of these series is absolutely convergent for all x and all t 9^ we
can perform the indicated multiplication and rearrange terms to obtain
G(x, t) = £ J *W>
where the Jn (x) are functions of x alone. Moreover, an easy computation reveals
that
Jn (x)
2
x4
2fc
=©• "l
_n\ 22(«+l)
x
'
242!(n + 2)! ' ^ ^
fc
22^!(A:
x
+ «)!
+
{x,2)[t {llt)]
Theorem 15-3. The function e generates the Bessel functions of
integral order of the first kind in the sense that
c
(«/2)[l-(l/l>]
= £
n= — 00
jn(x)t
n
(15 _ 58)
(Incidentally, this result motivates the choice of the coefficient a in the series
expansion for Jn that was made in Section 15-6.)
Having proved this theorem it is impossible to resist the temptation to do
something with it, since we are now but a step away from a number of important
results. To obtain them we make the change of variable t = e
ld
in (15-57), and
use the identity e ie
= cos 6 + i sin to write
e
(*/2)[«-(i/m = e
ix S ine
= cos(xsin0) + / sin (x sine).
and by equating real and imaginary parts and using the identity
we find that
00
(15-59)
sin (x sin fl) = 2 J /2 n+iW sin (In + 1)0.*
Continuing, we now multiply the first of these formulas by cos 2kd, the second
by sin 2kd, and integrate the resulting expressions term-by-term over the interval
< 6 < it (an operation which is legitimate here) to obtain
J2 k+i(x) = -
7T
/
Jo
sin (jc sin 0) sin (2k + 1)6 dd.
cos (nd — x sin 0) = cos nd cos (jc sin 0) + sin nd sin (x sin 0),
we deduce
Actually, the argument we have just given proves more than this. For if x is
the functions cos (x cos 0) and sin (x sin 0) on the interval < < tt, then
(15-60) implies that
J2k(x) = ~y
and (15-61) that
(y-\ - b 2h±l
I
* The validity of these computations depends upon the fact that (15-58) is absolutely
convergent for all t ^ or complex.
0, real
.
15-8 I
THE GENERATING FUNCTION 615
Theorem 1 5-5. For each fixed x > 0, Jn (x) approaches zero as n — > oo
EXERCISES
1. Deduce (15-60) from (15-59) under the assumption that termwise integration is
legitimate.
2. Prove that \Jn (x)\ < 1 for all integers n, and all x.
e
( * /2)[ '- (1/ ' )]
= 2 '»<*>'*
g
(x/2)[i-(l/0] (v/2)[«-(l/0] _ l(x+y)/2][t-a/t)]
to prove that
00
MX + y) = ^—
k= oo
Jk(x)Jn -k{y).
&=i
(d) Prove that
00 00
6 cos <tf0
Tip + q) Jo
for all p > 0, q > 0.
616 BOUNDARY- VALUE PROBLEMS INVOLVING BESSEL FUNCTIONS CHAP. 15
(b) Use the result established in (a) to rewrite the series expansion for Jp as
Jp (x) = y -^rrrr y- x J
sin cos 6 dd,
ni)T(p+h)^ (2k)\
j
Jo
i
2k \
£ (2k)\
p
(x/2)
(*<***?' (-D
Jp(x)
r(*)r(p + i)./o
^'e^itxr
{2k)
fc=0
<*>•
(x/2Y 2p
Jp (x) = sin 8 cos (x cos 0) dd.
r(*)ro> + %)Jo
(This result is known as the Poisson Integral Form for Jp .)
The following exercises introduce the so-called modified Bessel functions of integral
order, /„(*).
{x/2)U+{1/t)] in powers of
6. (a) Expand the function e t as
e
(./2)[l+(l/l)]
= £ /„(*)/»,
In(x) =
2 n n\
1 + 2(/i + 2)
+ 2 •
4(2* + 2)(2/i + 4)
+
n+2k
fc=0
I- n (x) = /„(*).
(b) Prove that the functions /„ satisfy the following recurrence relations
/ —
In - 1
*n4-l
-—I*-nt
—
X
r —
in - i'
In .l^t*«)
\
In+1 = In In-
X
8. Show that the modified Bessel function /„(*) is a particular solution of the equation
x 2y" + xy' + (X
2
jc
2 - p 2 )y = 0, (15-63)
or
* The reader can verify that (15-63) is a variant of Bessel's equation by making the
2 -
change of variable s = \x in s 2y" sy' (s
2
+
p )y = 0. Equation (15-64) is, of
+
course, the self-adjoint form of (15-63).
.
since the leading coefficient of (15-64) vanishes when x = 0, we need only impose
a boundary condition at the point x = 1
d_
(i5 - 66)
dx (*s)+(^-£)"- -
|0i| + Ifol 5* 0. By the results in Section 12-8 we can assert in advance that
eigenfunctions belonging to distinct eigenvalues for this problem are mutually
orthogonal in (PC[0, 1]. Furthermore, the eigenfunctions belonging to a positive
eigenvalue X will be the nonzero multiples of Jp (\ x), while the eigenfunctions
belonging to the eigenvalue (if is be the nonzero multiples of
an eigenvalue) will
jcV + x/ - p 2y = 0.
Thus, under the hypotheses imposed upon p and X we need only examine the
functions x p and J (Kx), X > 0, as potential eigenfunctions. And here we argue
p
as follows.
XI) = 0, (15-68)
and it follows that Jp (\x) will be an eigenfunction if and only if Jp (\) = 0. Hence
the positive zeros X x < X2 < • • •
of Jp are eigenvalues and
Jp(Ux), k — 1,2,...,
Case 2. /3i
= 0. Here (15-67) becomes
/(l) = 0, (15-69)
and
1, JoQi k x), k = 1,2,..., when/? = 0.
. ,
15-9 |
STURM-LIOUV1LLE PROBLEMS FOR BESSEL'S EQUATION 619
Case 3. For various reasons the case fix ^ 0, j8 2 ^ is not particularly interest-
ing, and is usually replaced by the more general requirement that when jc = 1,
Jp(\x) satisfy the equation
(As we shall see, this type of boundary condition arises in the study of heat flow
in cylindrical regions, and is not as one might think.) Since (15-70)
artificial as
has a variable coefficient X, it does not under any of the several types of bound-
fall
satisfy (15-66) and (15-70), and as a consequence are "eigenfunctions" for this
problem. Moreover, reasoning as in Section 12-8, we find that these functions
are mutually orthogonal in <PC[0, 1] with respect to the weight function x. (See
Exercise 2 below.) Thus, on the face of things, the situation here would appear
to be identical with that discussed in each of the preceding cases. This, however,
is not quite true, for it turns out that whenever p < h, Eq. (15-70) has other roots
in addition to the v k . Indeed, when/7 = h, Eq. (15-70) becomes
X/p+ i(X) = 0,
and has X = as a root. When p < h the situation is much more complicated,
and cannot possibly be treated with the tools we have available. Suffice it to say
that the equation then admits a pair of imaginary roots ±«/ a phenomenon —
which does not occur with any of the boundary conditions we have considered
heretofore. In the next section we will find that the existence of these additional
roots introduces difficulties in the study of series expansions relative to the orthog-
onal set {Jp(vkX)}
In view of the now obvious fact that soon be computing series expan- we will
sions relative to the eigenfunctions found above, we conclude this section by
evaluating their norms in (PC[0, 1]. The basic formula for all of these computa-
tions is given in the following lemma.
More generally*
C 2 2/2 2
[Jp (\x)]
2
xdx = y[^(X6)] 2 +
X ~ P [Jp QJ>)f. (15-72)
J 2
[x
2
(/)
2
]' + [(jc
2 - p
2
)y
2
]' - 2xy 2 = 0,
we conclude that
2x[Ux)]
2
= ± {*V P (*)]
2
+ (x
2
- 2
p XJP {x)f}.
Hence
x iX
,X
f 2
2/ [^(x)] **/*
2
= xV^)] 2 + (/ - /FpW]
Jo lo lo
2 2 2 2 2
= x W(x)] + (x - p XU*)] ;
the last step following from the fact that pJp (0) = for all p > 0. The desired
result now follows by setting jc = A/in the above integral. |
When rewritten in terms of the inner product on <?e[0, 1], Eq. (15-71) becomes
2
2 2 X2
\\JP M\\ = M4(x)] + 2 x^ ^ (x)]2> [
(15_73)
= «^(X*)]
a ~74
l|/p(Xfcx)|| . ( 15 )
The reader should note that by using Formula (15-52) this result can also be
written
a 2
l|/p(X*x)|| = M/p+i(X*)] . (15-75)
Vk_— F
p
]Jp<Pk)V \
.2
VpMY. (15-79)
EXERCISES
1. Verify that Eq. (15-63) is Bessel's equation of order p in the variable \x.
2. Apply the argument given in Section 12-8 to prove that the functions Jp (y k x),
k = 1,2,..., discussed in Case 3 above are mutually orthogonal in (PC[0, 1] with
respect to the weight function x.
Now that we have established the existence of an infinite set of mutually orthog-
onal eigenfunctions for each of the boundary-value problems introduced above
we are faced with the task of determining whether these sets of functions are bases
for (PC[0, 1]. With but one exception the answer is yes, and in general the results
in this connection are analogous to those we have encountered in similar situations
in the past. This time, however, we can do no more than state the relevant the-
orems, since their proofs are far too difficult to be given here. The first and most
important of them reads as follows.
Wxt) + /W)]
for each x in the interval (0, 1), and the convergence is uniform on every
closed subinterval of(0, 1) which does not contain a point of discontinuity off
of p.
The series described in this theorem is known as the Bessel or Fourier-Bessel
series for f of the first kind with respect to the functions Jp (Kkx). The reader should
note that this result actually provides us with infinitely many different series of
this kind, one for each admissible value of p.
Example 1. Find the Bessel series expansion with respect to Jp (\kx) for the
function x p
, p > 0.
In this case the coefficients in the series are given by the formula
x p+1Jp (\ k x)dx = -^ 2 J/
? + %{t)dt
Jfo \l+ o
d r.p+l
P+1
[t Jp + 1 (t)]dt
P+2
A£ dt
+1 J
[f P + l{t)]
P+2
X*
= y- Jp+i&k)'
Thus
Ck k= 1,2,...,
Wp+l(A*;)
and
where the series converges in the mean in (P6[0, 1], pointwise in (0, 1), and uni-
formly on any closed subinterval of (0, 1).
15-10 I
BESSEL SERIES OF THE FIRST AND SECOND KINDS 623
V- /o(X**) 1
< x < 1. (15-83)
Xfc/iCX/;) 2
j^J
Ck = x,3JoiXkX) dx,
[MW] 2 Jo
and, reasoning as in the preceding example, we find that
x Jo(\]cX) dx \ / t%(t)dt
= ~ t
2
4:[tMt)]dt
dt
>l J °
x! lo Jo
2
t J x {t) dt.
Xfc x|^o
But
/V!(0* = - r-fJoMdt
o dt'
-2f o ^[^W]*
= 2\ kJi(\ k ),
and it follows that
x J (\icX) dx = MW 4y 2 (x*)
Xfc
x2
Thus
Ck J_ _ 4
^i(Xft) x*
XlJ
624 BOUNDARY- VALUE PROBLEMS INVOLVING BESSEL FUNCTIONS | CHAP. 15
and
x2 = 2
Z T7^(v ~ 4W XfcX >» ° < x < L
Then
JpfokX), k=l,2,..., p > 0,
and
1, J (ix k x), k = 1,2,...,
ore bases for (PC[0, and series expansions relative to these bases converge
1],
in exactly the same way as the series described in the preceding theorem.
In particular, it now follows from Formulas (15-76) and (15-77) that a function
Ax) = £
k=i
c kJp (tx k x), p>0, (15-84)
with
with
Co = I
Jo
f(x)xdx, ck = rr
yo\^k)r
^ Jo
f(x)Jo(n k x)xdx, k>0. (15-87)
A series of this type is usually said to be a Bessel series expansion- off of the second
kind, or a Dini series expansion off
Finally, a similar result holds for the functions Jp (y k x) discussed in Case 3
of the preceding section provided we insist that p be greater than h. On the other
hand, when p < h, an additional function must be added to the Jp iy k x) to obtain
p
a basis for (PC[0, 1]. When/? = h, the function in question is x or any nonzero ,
EXERCISES
^ r1
/•i
2
n -2
Jo
n
x J (\x)dx = + ^'o(A) - ijL
^T-) x Jo(\x)dx.
°° r i
|
f
x = 2 X) "T~TFr\ i i / ^o(XfcO dt Jo(X x), fc
r xJo(kx)Jo
I
Jo
(\ k x)dx = \ kJo(\k)Mk)
k
22
- Xfc
Mkx) = 2Joik)YJ —— 2
^| Jo(\ k x),
fc=i (X fc k )/i(Xjb)
for < x < 1.
4. Prove that
l-* 2
= 8£4^-
k=\ X*;/i(Xfc)
5. Expand the function xp ,p > 0, in (PC[0, 1] as a series involving the functions Jp(ji k x).
6. Prove that
Jo ^>'
lnx= -2^ 2 2
k =\ Xfc[.7l(Xfc)]
/ x In x7o(X&x) dlx.]
./o
fc=i Xi/p+i(X/t)
for all /? > 0, and all x in (0, 1).
for < x < 1. [Hint: Multiply the formula in (a) by x~ (p+1) and differentiate.]
626 BOUNDARY-VALUE PROBLEMS INVOLVING BESSEL FUNCTIONS CHAP. 15
At this point it should be abundantly clear that we have assembled more than
enough information to solve boundary-value problems involving Bessel's equation.
Indeed, given what we now know, this is largely a matter of routine computation,
and involves little that could not be left to the reader's imagination and the —
exercises. But, in the interest of completeness, we shall devote the following pages
to a brief discussion of several problems of this type, producing a formal series
solution for each of them. Of course, once this has been done we are still faced
with the task of determining conditions under which these series actually satisfy
the problems they purport to solve. Here, however, the technical difficulties are
formidable, and force us to be content with the vague statement that all of our
results are valid whenever the functions involved are sufficiently smooth.
This said, we turn our attention to Laplace's equation in cylindrical regions,
which we first propose to solve under the assumption that the solutions are in-
dependent of the polar angle 0. In this case the relevant version of Laplace's
equation is
2 2
d u 1 du d u
+
.
+
.
= (15-88)
dr 2 dr dz 2
(see Section 15-1), and we remind the reader that its solutions can be interpreted
as steady-state temperature distributions in the region in question.
Example 1. Solve Eq. (15-88) in the cylindrical region r < 1, < z < a
(Fig. 15-4) under the assumption that
u(l,z) = 0,
u(r, a) = 0, (15-89)
u(r,0) =f(r).
R" + -r R +
f
\
2
R = 0,
(15-90)
Z" - X Z =
2
0,
*(1) = o,
(15-91)
Z(a) = 0.
the admissible values of X are the positive zeros of J , and we have shown that up
to constant multiples R must be one of the functions
are solutions of (15-88) and satisfy the boundary conditions u(l,z) = u(r, a) = 0.
To accommodate the remaining boundary condition we now form the series
By our earlier results we know that this equation can be satisfied for any / in
(Pe[0, 1] by letting A k sinh (K k a) be the Ath coefficient in the series expansion of/
relative to the functions JoO^kr). Hence
«(',0) =/(r).
Physically these conditions assert that the base of the cylinder is maintained at a
known temperature /(/•), the top at zero, and that the lateral surface exchanges
heat freely with the surrounding medium at a rate proportional to the temperature
of the surface.
628 BOUNDARY- VALUE PROBLEMS INVOLVING BESSEL FUNCTIONS | CHAP. 15
This time we must solve the pair of equations in (15-90) in the presence of the
endpoint conditions
*'(1) - hR(\) = 0,
Z(a) = 0.
R(r) = J (\r),
Again our earlier results allow us to assert the existence of infinitely many such
roots v\ < v2 < • • • , all of which are positive, and with them, solutions
we find that
00
oo
(Note that the existence of this series for any fin (PC[0, 1] is guaranteed only if
h < 0. Otherwise a more complicated series must be used.) Thus, by Formula
(15-79), we have
lvl
Ak = .
2 a
2 \ Ary (y k r)rdr, (15-96)
[h
2
+ v k][J (y k )] smhO> k a)Jo
Example 3. As our final example of this type we again solve Eq. (15-88) in the
region r < 1, < z < a, but now impose the boundary conditions
u(r, 0) = 0,
u(r,a) = 0, (15-97)
ii(1, z) =f(z).
,
This time the method of separation of variables leads to the pair of equations
2
Z" + X Z= 0,
(15-98)
R" + - R' - \
2
R = 0,
R" + - R' - \
2
R = (15-99)
r
for each of these values of X. To this end we make the change of variable t = i\r
(i = y/—l), and rewrite (15-99) as
or
d2R dR
¥ + 7I + * 1
- °'
v2fc
= E o*r
*=o ^ 22 !
)
2
and itfollows that J (i\r) is a real-valued function after all. In view of this fact
it is reasonable to adopt a notation which does not (misleadingly) involve /, and
so we set
oo 2fc
o(*)=E^w*
7 (15 ~ 100)
630 BOUNDARY- VALUE RPOBLEMS INVOLVING BESSEL FUNCTIONS |
CHAP. 15
The function I (which the student should regard as being defined by this series)
is called the modified Bessel function of order zero of the first kind (see Exercise 6,
Section 15-8), and in the present case allows us to write
R{r) = 7 (Xr).
*«-/„(**).
and we conclude that the solution of the boundary-value problem under discussion
isof the form
Thus the A k are determined by the requirement that A k I (kT/a) be the kth coeffi-
off on the interval [0, a], and we have
cient in the Fourier sine series expansion
A * = -^-nrja\\
aIo(Kir/a) Jo
m^^dz.
a
05-102)
EXERCISES
1. Find the steady-state temperature in the cylindrical region r < 1, < z < a, given
that the solution u is independent of 8, and
(The last boundary condition asserts that the lateral surface of the cylinder is in-
-rlfr
+ T-Wz =
. 1 1
«>-efi'
a>0n -
(l5
c >k
w(l, 0, t) = 0,
T^
dr 2
+ -^
dr
r
= -
a z at 1
2^' a > °> (15-105)
and
u{\, t) = 0,
u t
{r, 0) = g{r).
Arguing in the usual fashion we set u(r, t) = R(r)T(t), and find that (15-105)
becomes
R" + (\/r)R' j_ r; = _ 2
R a2 T
where X is a positive constant. (The choice of sign here is forced upon us by the
requirement that T be periodic.) Thus R and T must be solutions of the equations
R" + - R' + \
2
R = 0, (15-107)
and, in addition, R must satisfy the boundary condition R(l) = 0. Starting with
the general solution
R(r) = AJ (\r) + BY (Xr)
where \ k is the A;th positive zero of J . Moreover, with these as the values of X
the solutions of (15-108) are
A k Bk
, constants, and it follows that the functions
00
u(r,0) =f(r),
u t (r, 0) = g(r).
fir) = X) AkU^ur),
We now turn to the general case with solutions dependent upon 0. Here we
start by 0, t) = R(r)®(d)T(t) in (15-103) to obtain
setting u(r,
R + r R + r2 © a2 T
where, as before, X is a positive constant. From this we obtain the pair of equations
T" + X Vr = 0,
i{
+ a il
+ r2
2
_ R" + (\/r)R' + \ R = ©^ = _ 2
M
(1/r 2 )/? © '
where n > is again a constant. (The choice of sign is dictated by the fact that ©
must be periodic with period 2x.) Thus we must solve the equations
+ XVr = 0,
T" (15-112)
r
2
R" + rR' + (A 2 2 - 2 )/? = /- /i 0, (15-113)
0" + m 2 = 0, (15-114)
R(l) = 0. (15-115)
Starting with (15-114), we use the periodicity of © to deduce that the only
admissible values of /x are
Hn = n, n = 0, 1, 2, . .
.
where A nk and Bnk are arbitrary constants. Putting these results together, we
conclude that the functions
and
v n k(r, 6, i) = [A nk cos (a\ nk t) + Bnk sin (a\ nk t)] sin (nd)Jn (\ nk r),
n = 0,1,2,..., k = 1,2,...,
00 00
u(r, M) = 2= £
71 fc=l
["»*(r » ' ') + Vnk (r > e 0],>
(15-116)
OO 00
fir, J)=2E
=
&
n fc=l
nk (r > e > °) + °»*fr '» °M
or
oo / oo oo \
/fo *) = 2 12
n=0 Ia;=1
[^n^n(Xn^)] cos «0 + J] P„*/„(\ n *r)]
&=1
sin «0
'
• (15-1 17)
with
floW = Kr,e)de,
^/_
an(r) = 2A
fc=l
nkJn0^nkr), and bn (r) = ^A
/c=l
nkJn (\ nk r h
.
15-12 |
THE VIBRATING CIRCULAR MEMBRANE 635
for all n. Hence the A nk and A nk must be the coefficients in the Bessel series expan-
sions of a n (r) and b n (r) with respect to the functions Jn (Kkr), and we have
1
2 f
A nk = 7i ,x v, 9 /
an (r)Jn (\ nk r)r dr,
Un+liKk)] 2
A nk = 71 r\ M2 Jo
/
bn (ryn (\ nk r)r dr.
Vn+i{Kk)r
Thus
1 rr
B °k = J_j(r>Wo^okr)rd8dr,
«Xod/i(Xa*)P i
Bnk = .
r/ ,. ,
12 / /
g(r, 6) cos (nd)Jn (\ nkr)r dd dr,
Bnk = >
ry —n~ via / / #(r > ) sin ("0)4( W> ^ <fr,
EXERCISES
1 An object located at the point x = xo starts from rest and moves along the x-axis
under the action of a force directed toward the origin whose magnitude is propor-
tional to the distance of the object from the origin and to its mass, and initially is
moxo. Determine the motion of the object if its mass m varies with time according
to the formula m = mo(l 0- +
2. A uniform, flexible cable of length L is suspended vertically as shown in Fig. 15-5.
At timet = that portion of the cable between a: = Oandx = aL is given a uniform
horizontal velocity v = f(x). Describe the subsequent behavior of the cable, given
636 BOUNDARY-VALUE PROBLEMS INVOLVING BESSEL FUNCTIONS CHAP. 15
d y d
g
dl* dx (-D
where g is the acceleration due to gravity. [Note. The
x-axis is directed upward with the cable suspended from
-
*
the point (L, 0).] v= Ax).
FIGURE 15-5 X&
3. Find the steady-state temperature distribution u(r, 6, z) in the cylindrical region
r < 1,0 < z < a, given that
4. Find the temperature u(r, 6, t) in the two-dimensional region shown in Fig. 15-6,
given that the boundary of the region is maintained at a temperature of 0°, and that
at time t = the interior is at a uniform temperature of 100°.
5. Find the equation of motion of a vibrating membrane of the shape shown in Fig.
15-7, under the assumption that the membrane is held fixed along its boundary and
is released from rest at time / = from a known position.
6. Solve Exercise 5 when the membrane is also given a known initial velocity.
infinite series
1-1 INTRODUCTION
The objective of this appendix is to provide a reasonably complete account of
the material relating to the convergence of sequences and series that was used
in the body of the text. From the standpoint of logical completeness this discus-
sion ought to begin with a detailed study of the real number system, including its
In these terms the least upper bound principle —which, by the way, is actually a
theorem concerning the real numbers —reads as follows:
Least upper bound principle. Every (nonempty) set of real numbers which is
bounded from above has a least upper bound. (Again there is a companion state-
ment concerning lower bounds which we omit.)
And once this statement has been accepted we are back on solid ground where
theorems can be proved and definitions given without further gaps in the reasoning.
637
638 INFINITE SERIES | APPENDIX I
Needless to say, we shall not attempt to give a complete treatment of the several
topics mentioned above, since this would entail writing an entire text, or more on
advanced calculus. Neither shall we prove every assertion that is made as we
progress, since this too would result in a labored discussion. We do, however,
insist upon the fact that these proofs are now within our reach, and want only
time and patience to present.
We assume that the reader is already familiar with the notion of a sequence {ak }
of real numbers, which, we recall, is simply an ordered list
{a i, a 2 , . . . , ak , . . .}
of real numbers indexed by the positive integers, or, more formally, a real valued
function F whose domain is the positive integers, and whose value F(k) at k is ak .
Actually, there is no reason to insist that the indexing always begin with the sub-
script one, and when convenient we shall change it without comment.
This said, we now introduce the concept of sequential convergence, as follows.
for all k > K. When this happens we say that a is the limit of {ak } , and
write
lim a k = a, or {a k } —» a.
If, on the other hand, no such number exists, {ak } is said to diverge.
Implicit in the statement of this definition is the assertion that the limit of a con-
vergent sequence is unique. To see this, suppose that {a k } converges to a, and let
a' 9± a. Then if e = \a — a'\/3 and if K is chosen so that la* — a\ < e for all
k > K, the only entries in {ak } which do not lie in the interval (a — e, a + e)
are a u a 2 ax, and it follows that {ak } does not converge to a'.
, . . . ,
Conversely, if {ak} is not bounded from above, then for each real number a
we can find an integer K such that a < ax- Setting e = \a — ok\, we have
\a — a k > e for all k > K. Thus {a k } does not converge to a, and since a was
\
Using this result, it is now relatively easy to establish a criterion which will
\a m — an \
< e (1-2)
for all m,n>K.
The convergence criterion we now propose to establish asserts that the class of
Cauchy sequences is identical with the class of convergent sequences. This is easily
the most important single result on sequential convergence.
Proof Suppose that {a k } is convergent, with a as its limit. Then, given any e > 0,
K
we can find an integer such that \a — ak < e/2 for all k > K. Thus, if m and \
\a m — an \
= \(a m — a) + (a — an )\
< \a m — a\ + \a - an \
<! +! = €,
Then b\ > b 2 > b 3 > and each bk is at least as large as the greatest lower
• • • ,
Ki be chosen so that \a m — a n < e/3 for all m, n > K\. (The existence of such
\
an integer follows from the assumption that {ak} is a Cauchy sequence.) Let K2
be chosen so that \a — b k < e/3 for all k > K2 and let K be the larger of AT X
\
,
\a - ak \
<\a — bk \ + \b k - ak \
< 3 + \bk - a k \.
But since b k is the least upper bound of {ak , ak+1 , . . .}, there exists an index
p > k such that \b k — ap \
< e/3. Hence
\b k — ak \
< \b k — ap \ + \a p — ak \
<- +-
^3^3 = -•
3
\a — ak \ < ^ +y= e,
1 1 1
2 3 k
Example 2. The sequence {1, - 1, 1, - 1, ...} is not a Cauchy sequence, and hence
does not converge.
The argument given in this section can be summarized by saying that the least
upper bound principle implies that every Cauchy sequence of real numbers is
convergent. It is also possible to turn things around, and deduce the least upper
bound principle from Theorem 1-2. In short, these two facts concerning the real
number system are equivalent, and either can be taken as the starting point for
the study of infinite series, and, in fact, the entire theory of real valued functions
of a real variable. We omit the proof.
We conclude this section by proving an elementary computational theorem
which will be needed below.
lim a k = a, lim bk = b.
k—>oo k—>oo
Then
(i) {aak + pbk} is convergent for all real numbers a and @, and
Proof. We leave to the reader the easy task of verifying that {aak } — > aa when-
ever {ak} — a. This proved, (i) will follow as soon as we show that {ak} — > a
and {b k} — » b imply {a k + bk} — > a + b. Let € > be given. Then
and, by assumption, we can find integers Ku K2 such that \a k — a\ < e/2 for all
k > K x, \b k - b\ < e/2 for all k > K2 . Thus if K is the larger of #i and K2 ,
we have
Now let e > be given, and choose integers K and K 2 such that x
€
\a k — a| <
2|6|
for all A: > K x , and
l*» " *l < 2S
for all k > K2 . (If b = 0, the second term in (1-3) vanishes, and we need only
choose K x .) Then with K the larger of K x and K2 , and k > K,
In this section we review the elementary facts concerning infinite series of con-
stants, including several well-known tests for the convergence of such series.
00
^
fc=i
ak = ax + a2 + • • •
+ ak + • •
•
(1-4)
be an infinite series of real numbers, and let {sk} be the associated sequence
of partial sums
Sx = ax ,
s2 = «i + a2 ,
sk = ax -\- a2 + — + • ak .
a = ^2 ak ,
k=l
and say that a is the sum of the series. Otherwise, (1-4) is said to diverge.
Perhaps the most familiar example of a convergent infinite series is the geometric
series
a + a r + a r
2
+ • • •
,
(1-5)
1-3 |
INFINITE SERIES 643
whose ratio r satisfies the inequality - 1 < r < 1. Indeed, in this case
so = a ,
Si = a (l + r),
k
sk = fl0 (l + r + r
2
+ • • •
+ r ).
(1 + r + • • •
+ r^Xl - r) = 1 - r\
we have
- k+1
1 r
Sk = ^o _ ' r 9± 1,
1 r
— r
fc~»oo l
provided |r| < 1. On the other hand, if \r\ > 1, the sequence {sk } is divergent,
and hence so is (1-5).
Example 1. The real number 0.33333 ... is the sum of the geometric series
_^
3
_j
3
^
I
3
^
i_ . .
10 10 2 103
whose ratio is y^. In this case the formula given above for lim^*, sk yields
a _ 3 1 _ 1
1 - r
~ 10*1 _ jl_
~ 3'
1 10
as expected.
Theorem 1-4. The series £)j°=i ak converges if and only if for each e >
there exists an integer K, depending on e, such that
2
k=m
ak < e (1-6)
In fact we note that the expression £*=m <**> is just the difference sn — sm _ x of
the partial sums sn and jto _i. Hence the theorem states that ]C*=i a k converges
if and only if its sequence of partial sums is a Cauchy sequence, and this is the
Theorem 1-4 provides a general convergence criterion for series, but unfor-
tunately it is difficult to apply in practice. We devote the remainder of this section,
therefore, to the derivation of several consequences of the theorem which, although
lacking the generality of Theorem 1-4, do provide convenient tests for convergence
in a large number of cases.
Proof. Since
M= \Sn — S„_i|,
It is useful to restate the last result in the following form: If ak does not tend to
zero ask — > oo , then X)£=i a k diverges. Thus each of the series
oo oo oo / i \n
Ettv
n=l
2>
=
n
n
l
"' 2(« +
n=l N ;)
'
diverges, the first because lim„_., n/(l + n) = 1, the second because lim n _^„ (sin n)
does not exist, and the last because
lim ( 1 + -
n))
= e.
n^oo \
It must be emphasized, however, that the converse of Theorem 1-5 does not hold,
for, as we shall see presently, the so-called harmonic series
Zr
fc=i
i + i + i + ---
Theorem 1-6. (Comparison test.) Let YI=\ ak and £,k=i bk be series with
positive terms.
(X) If Zfc=i <*k converges and b k < ak for every k, then Y,k=\ b k also
converges.
(2) IfHk=\ ak diverges andb k > ak for every k, then J2k=i b k also diverges.
1-3 I
INFINITE SERIES 645
Proof. (1) Since bk > for all k, the partial sums of b k form a monotoni-
2~Lk=i
cally nondecreasing sequence. But this sequence is also bounded from above, since
n n oo
^2
k=i
bk - 2
k=i
ak - S
fc=i
ak = s >
where S denotes the sum of the convergent series £*=i a^ Hence by Theorem 1-1
the sequence of partial sums of ££=i bk converges.
(2) In this case 2~Ifc=i bk > H£=i ak and it follows that these partial sums are
unbounded since otherwise the series JLk=i a* would converge. |
Theorem 1-7. (Ratio test.) Let £]a=i ak be a series of positive terms and
suppose that
L = lim
k—>Q0
^ @k
exists. Then
Proo/. (1) Suppose that Z, < 1 and that r is chosen to be a fixed real number
with L < r < 1. Then for sufficiently large values of n, say n > N, we have
an+ i/an < r. Thus
2
tf.v+2 < '"o^+i < r aN ,
k
ciN+k < r au.
2J aN+k = 22 a k>
fc=0 k=N
(2) The proof of this case is similar: this time Jlk=i ak is compared with a
divergent geometric series (ratio r > 1). We omit the details. |
Theorem 1-8. (Integral test.) Let 2Za=i a k be a series of positive terms and
assume that there exists a function f, continuous and monotonically nonin-
creasing on 1 < t < cc, such that f(k) = ak for k = 1,2, Then
the series 2~lk=i a k and the improper integral
J" f(f) dt converge or diverge
together.
646 INFINITE SERIES |
APPENDIX I
Proof. Assume first that J\ f(t)dt converges. Then a reference to Fig. I- 1(a)
makes it clear that
rn + 1 f°o
Thus the sums YLk=i a k of the given series form a monotonically nonde-
partial
creasing sequence whichis bounded above by the real number a x + J\ f(t) dt.
FIGURE 1-1
If, on the other hand, the integral j^ fit) dt diverges, then the partial sums
5Z3b=i «fc are unbounded, for in this case (see Fig. I-lb)
r n+1
2>*> /
Jl
fd)dt,
JfcTi
£;-•+*+*+
diverges, for the integral test may be applied in this case with /(f) = \/t to obtain
lim /
f t
_p
dt = lim
t- p+1 — T /> 7* 1,
» *
and since this limit exists if and only if p > 1, it follows that a p-series
££=1 \/k p converges ifp > 1 and diverges ifp< 1. In particular the series
00 -t 00 1
E^
fc=l
and
EfcToi
fc=l
E-7^ and
Efciu9
Vk
fc=1 fc =i
diverge. The /(-series and the geometric series constitute useful classes of series
for which the question of convergence is completely settled. By using these series
together with the comparison test, a large number of additional examples may be
treated.
^Kk + \) 1-2
1
+A + A +
' 2-3 ' 3-4
00 1
y—
*-!*'
t
fc=0
= l
+ h + h + h+-
Since
1 1 1
<
~
(k + 1)! 1 •
2 •
3 • • •
(fc + 1) 2*
k
the given series converges by comparison with the geometric series Xlfc=o (¥) -
Note that the ratio test may also be applied in this case, since
oo i, & ^2 t3
^ 2! ^ 3!
^
£i fc!
i
"» (
* + + 1)! ,. (At + 1)*+' A;!
'C'/^ (k + 1)!
£l( +
= £)'" , = e > 1.
Since such series converge if p > 1 and diverge if /? < 1, the ratio test cannot
possibly give any information in the case L = 1 (see Theorem 1-7).
Theorem 1-9. 7/ELi M converges, then so does Zfc=i «fc- Briefly, ab-
Proof. Under the assumption that £fc=i \a k converges, we shall show that the \
partial sums of £fc=i a k form a Cauchy sequence. But this follows immediately
from the relation
n n
X) °k ^ X) l°*l' m ~ "»
k=m k=m
because the right member can be made arbitrarily small by choosing m sufficiently
large. |
Since many important tests for convergence of series apply directly only to
series whose terms are positive, Theorem 1-9 provides a way, sufficient for many
1-4 I ABSOLUTE CONVERGENCE 649
applications, of applying these tests to arbitrary series. The ratio test for instance,
Cln+l
L = lim
n—><»
exists. Then
then \a n+ x > \a n for sufficiently large n. Thus a n does not tend to zero as n
| \
00.
converges.
Since ak = (— l) kk(x/2) 3k we have ,
3n+3
an+l (-l) n+1 (« + l)(x/2) n + 1 /sV « + 1
(-l)»«0/2)3*
Hence
lim
^n+1
= lim
n + 1
n—»oo n—»oo
z
and the given series converges if |jc/2|
3
< 1 and diverges if \x/2\ > 1. Thus
the series converges for values of x lying in the interval
-2 < x < 2
and diverges if \x\ > 2. Finally, in order to determine the behavior of the series
at the points x = ±2, we note that the general term becomes ±(— l)*/c in this
case, and since this quantity does not tend to zero as k —» go , the series diverges
at both points.
650 INFINITE SERIES | APPENDIX I
+1
(-o* + i_
,_ i2^3 i +
f-
/ j h.
K
—.
l
4 T ...
J
fc=l
whose partial sums do approach a limit (see Theorem I— 11), despite the fact that
the corresponding series of absolute values is divergent. This example is typical
of alternating series where we have the following useful criterion for convergence.
Theorem 1-11. If the terms of the series £*=i ak alternate in sign and satisfy
(ii) lim \a k \
= 0,
k—>oo
then the series converges. Moreover, if S is the sum of the series and Sn is
\S - Sn < \
\a n \.
= S2n -1 + ( a 2n + a 2n + l)
< $2*, — 1j
2n 2n— 2
S 2n = 2
fc=l
ak = S
fc=l
° fc + ( a 2n-l + fl 2n)
> ^to— 2j
since a 2n -i +
a 2 n > 0. It follows that the odd-numbered partial sums form a
nonincreasing sequence bounded below by S 2 and that the even-numbered par- ,
tial sums form a nondecreasing sequence bounded above by Si (see Fig. 1-2).
we conclude that SE = So, and the common limit 5 is the desired sum of the series.
Moreover from the inequalities
we find that
\S — S2 k\ < 152*— i — S2 k\ = \a2k\,
and
\S — -S 2 fc+l| < |'S'2fc+l
— $2k\ — |«2fc+l|j
l«6l
— • m • —• •—
S2 54 56 • • • S5 S3 S {
FIGURE 1-2
QD
(x + If
E
fc=0
2k + 1
d-7)
to show that it converges absolutely if Jjc -f- 1 1 < 1 and diverges if -{-
jjc > 1 1 1
that is, the series converges absolutely for values of x in the interval —2 < x < 0.
At the endpoints of this interval the ratio test gives no information. However, if
(-0*
A;=0
2k + 1
and E 2k +
k=0
1
This is the familiar notion of continuity basic to any elementary calculus course.
Not usually introduced at that level, however, is the following notion of uniform
continuity.
but not on x, such that |/(*i) - f(x 2 < )\ e whenever x ls x 2 are in / and
continuous on that interval. The following example shows, however, that the
converse is false.
Example 1. Let
f(x) = l/x, < x < 1;
let x be any point in this interval, and let e > be given. If x > x /2, then
Xq
< — \x - x \.
Xq XXq Xq
If, further, \x - x \
< (x 2 /2) e, then
< €.
Xq
Thus with
Xq Xq
8 = min
It follows that /is continuous in the interval < x < 1. However, the given
1-5 BASIC NOTIONS FROM ELEMENTARY CALCULUS 653
Ax) =
i
2e<f(x )
FIGURE 1-3
the fact that the interval < x < 1 under consideration is not closed. This is
Six
- y ||
= V(x! - yi y + ... + (*„- yn y < 8.
for every e > there exists a 8 > depending in general on e but not
0,
on x, such that |/(x) — /(y)| < e whenever x, y are in 3D and ||x — y|| < 8.
654 INFINITE SERIES | APPENDIX I
If we consider functions defined on closed and bounded regions 3D, we have the
following theorem.*
Two properties of continuous functions which are of the greatest interest from the
viewpoint of elementary calculus are given by the next two theorems.
bounded region 3D of (R n
then there exist real numbers m, and points
,
M
y = 0>i, • • • , yn ), z = Oi, . . . , zn ) in 3D such that
lim
^X + ® ° ~ f^
fc->o h
= Vxf + • • •
+ x'i < M.
1-5 | BASIC NOTIONS FROM ELEMENTARY CALCULUS 655
fib) - /(a)
f'(x
b — a
).
The geometric content of this result is illustrated in Fig. 1-4. It states that there
is at least one point in the open interval a < x < b where the tangent to the
curve y = f(x) is parallel to the secant line connecting the points (a, /(a)), (b,f(b)).
FIGURE 1-4
Turning now to integration, we assume that the student already has some
6
intuitive feeling for the definite integral J f(x) dx of a continuous function /.
A detailed definition would be too lengthy to present here. Nevertheless, given a
continuous function/, it is useful to recall the following terminology:
(1) Jaf(x)dx is called the definite integral of/ on the interval a < x < b,
whereas
Later we shall need to extend this notion to include the endpoints of an interval as
well.
656 INFINITE SERIES |
APPENDIX I
It follows almost immediately from the mean value theorem (Theorem 1-17) that
two indefinite integrals off in a < x < b differ by at most an additive constant.
Thus we are led to the formula
b
f f(x)dx = F(b)- F{a),
Ja
b b
Theorem 1-20. (Mean value theorem for integrals.) Iff is continuous for
a < x < b, then there is an x in the open interval a < x < b such that
rb
/
f{x)dx = (b - a)f(x ).
Ja
lim fk (x Q )
exists for each x in I.
The following examples illustrate this definition and point out some of the reasons
for introducing a stronger type of convergence below.
1-6 I SEQUENCES AND SERIES OF FUNCTIONS 657
we have
lim fk (x ) = lim Xq = 0,
k—>oo fc*-»00
if < x < 1,
Ax) 1 if x = 1.
(See Fig. 1-5.) Note that whereas each of the/ fc is continuous on the entire in-
terval < x < 1, the limit function /is not continuous in this interval (namely,
it is discontinuous at x = 1).
0,1)
Example 2. Let fk (x) be the function defined on < x < 2 whose graph is
indicated in Fig. 1-6. Clearly ^(0) = timk^MO) = 0. More-
for every k, so
over if < x < 2, then there is some value of k, say K, such that 2/K < x ,
Proof We must show that for any x Q a < x Q < b, and any e
, > 0, there is a
5> such that \f(x) — f(x )\ < e whenever a < x < b and \x — x \
< 8.
Now
|/(x) - f(x )\
= \f(x) - fk (x) + fk (x) - fk (x ) + fk(x ) - f(x )\
But since the convergence is uniform, we can choose an integer k so that the first
and third terms on the right-hand side of (1-8) are each less than e/3 for every x
in the interval a < x < b. Moreover, since the function fk thus chosen is
continuous, we can also choose 8 > such that \fk(x) — fk(xo)\ < «/3 whenever
a < x < b and \x — x < 8. The desired conclusion is now immediate. |
\
Proof. We must show that for any e > there exists a K such that
x
\[ Mt)dt- r mdt < e
I Ja Ja
I
C h(t)dt -
Ja
f f(t)dt
Ja
= I
I
f [/*(/) -
Ja
f(t)]dt\
I
<
Ja f |/*(0 - /(Ol dt
b
< [
Ja
\M0 - f(t)\ dt
< (» - «)jr-5 - e
It would be useful to have a result similar to Theorem 1-22 which would apply
to differentiation instead of integration. Unfortunately this is impossible since
tions which converges pointwise to a limit f for a < x < b, and if the se-
quence {/ '0c)} converges uniformly on a < x < b, then
fc
exists for all f
x in the interval and
Proof Let g be the limit function to which {f'k (x)} converges uniformly. Then,
applying Theorem 1-22, we have
X
f g(t)dt = lim (* f&i)dt
Ja k—>oo Ja
= lim [fk (x) - fk {a)]
k—>30
= Ax) - f{a),
and it follows from the fundamental theorem of calculus that f'(x) = g(x) =
lirrifc^oo/fcCx), as desired. |
All of the above results may be recast in the context of series of functions rather
than sequences. As usual we will write
Ax) = E
fc=i
/*(*)' a < x <b, (1-9)
and say that the series converges (pointwise) to /, if its associated sequence of
partial sums converges pointwise to/. The series is said to converge uniformly if
the same is true of its sequence of partial sums. In this case it follows immediately
from Theorems 1-21 and 1-22 that if each term of the series is continuous, then
their sum /is also continuous, and
r At)dt
Ja
= e
k=1 Ja
r MOdt
Finally, if each term of the series (1-9) is continuously differentiate, and if the
series 2w*=i ./*(*) obtained by differentiating each term is uniformly convergent,
then
fix) = E/*W-
/t=i
Proof. It follows from the comparison test that for any Xo in the given interval
the series Sa=i/a;(^o) converges absolutely. Thus the series converges pointwise
to a limit function /on a < x b. < Now
/(*) - S
k=l
/*(*)
fc=n+l
k=n+l
< J) M k
k=n+l
= EM * - E M*
and this latter expression tends to zero as n — > oo . Since it is independent of jc,
=
X,
fc=i
~W~
k*
sinx • • • (1-10)
Since
2
sin k x
k*
2
^
/or a// x, and since £*=i 1/& converges, it follows from the Weierstrass Af-test
that the given series converges uniformly on — oc < x < oo . Let / be the limit
function; i.e.,
2
a \ V^ sin/c A
k2
^/ x j ^ / sin /c
2
*
_ ^ cos /c
2
a
If, on the other hand, we differentiate the terms of (I— 10), we obtain the series
00
y^ cos k
2
x = cos a + cos 4a + • • • (I-H)
Of particular importance among series of functions are the so-called power series
00
2
^2 a kx
k
= a + aix + a 2x + • • •
,
fc=0
k
Theorem 1-25. If the power series IX=o GkX converges for some value of x,
say x = a then it converges absolutely for every x satisfying \x\ < \x
, \,
and it converges uniformly on every interval defined by \x\ < \xi\ < |a |.
Proof.
k
Since Jlk=o a k x converges, we know that ak xl >
— as k — oo and hence
that there exists a number such that |fl*Xo| < M, k M = 0, 1, 2, . . . . Now
662 INFINITE SERIES APPENDIX I
In particular, for any fixed x lf \xi\ < the series £fc=o \akx\\ converges. Thus
\x \,
k
Now for any power series Yik=o QkX , one of the following is certainly true:
k
(0 Sr=o cikX converges for every value of*.
(2)
k
Hfc=o OkX converges only for x = 0.
k
(3) Sr=o QkX converges for some nonzero value of x but not for all values.
k
In the third case the set of positive numbers x for which Y2=o akX converges
is bounded above, for otherwise, by the theorem, case (1) would apply. Letting
R be the least upper bound of this set, we conclude that J2k=o <*kX
k
converges if
\x\ < R and diverges if \x\ > R.
Combining the above cases, we have
Theorem 1-26. For any power series 52k=o QkX k there is a nonnegative
number R(R = and R = ao are included), called the radius of convergence
of the series, such that the series converges (absolutely) if \x\ < R and
diverges if \x\ > R. Moreover, if Ri is any number such that < Ri < R,
k
then J2k=o QkX converges uniformly on the interval —R\ < x < Ri.
k
Example 1. Consider XX=o (1 /k\ )x . Applying the ratio test we have
+1
[l/(k + 1)!]^ \x\
(!/*!)** k + 1
and for any x, this ratio tends to zero as k —> oo Thus the given series converges .
absolutely in — oo < x < oo and uniformly on every finite interval — JRi < x < Ri.
k
Example 2. If for Ea=o ak x the limit
L = lim
k—»oo dk
exists, then the radius of convergence of the series is R = \/L. For in this case
we obtain the ratio
k+\
Qk+iXr Qk+1
OkX* dk
and this tends to L\x\ as k — > oo. Thus the series converges if L|jc| < 1 and
diverges if \Lx\ > 1, i.e., converges if |x| < R and diverges if |*| > R.
1-7 | POWER SERIES 663
00
F(x) = 2
k=0
°***
has radius of convergence R, then fa F(x)dx exists for ~R < a < b < R
and
"«
/
F(x)dx = £ /
******* = £ ak
b
• (1-12)
/Voqf. According to Theorem 1-26, the series converges uniformly on a < x < b.
Hence (1-12) follows from the general results on integration of uniformly con-
vergent series. |
F'(x) = ]T ka k x
k -\ (1-13)
fc=0
k
< k\a k \R r l
M „k _ x
k
\Xi\
-k M
fc-1
*1
Xi
However, the series
±k M
fc-1
fc=0 l*il
converges by the ratio test, and the uniform convergence of (1-13) now follows
from the Weierstrass M-test. |
Stated informally, the two preceding theorems assert that a power series may be
integrated or differentiated term by term without affecting the radius of conver-
664 INFINITE SERIES |
APPENDIX I
(a) £
A;=0
xk = l
+ x + x2 + *
3
+ ' '
'
00
v fc
+X X2 Y3 X*
o» Er+r-'
, + + T + T + T- --'
A: 1
1
4
x
fc+2
x
2
x3 x
<c > £ ft + IV* + 2) 2
+
,
2 •
3
+
,
3 •
4
+
fc=0
Hx) = S akx
k
.
F'(x) = £ ka x -\
fc=0
k
k
-R < x < R.
k ~n
F^Xx) = Y, Kk ~ l)(fc - 2) • • • (fc - /i + \)a k x ,
fc=0
for n = 1, 2, 3, . . . , and hence F (n) (0) = n\a n . With this we have proved
The theorem asserts the uniqueness of the power series expansion of a function
F on a given interval —R < x < R. The existence of such a series is a more
difficult problem, and we investigate it in the next section. We note, of course,
1-7 |
POWER SERIES 665
Theorem 1-30. If
00 00
k k
/(*) = 2] a kX and g(x) = ]T b kx
fc=0 fc=0
on —R < x < R, then for any constants a, /3 the series Yik=o («#* + Pbk)xk
has radius of convergence at least R and represents the function af(x) + @g(x)
on —R < x < R.
Theorem 1-31. If
f(x ) = 2
k=0
a kx
k
and g(x) = ]T
fc=0
b kx
k
ck = J] dibk-i = a bk + a^k-i + • • •
+ a 6 fc , (1-14)
The reader will note that the coefficients (1-14) are exactly the ones obtained by
"formal multiplication" of the given series, treating them as polynomials.
As a final comment, we note that so far we have been discussing power series
k
Z^=o ak x whose interval of convergence was centered at the point x = 0. The
entire discussion may be carried through equally well, however, for series of the
form X!fc= i a k (x — a)
k
. The interval of convergence for such series is of the form
a — R < x < a +R (with or without the endpoints) and the radius of con-
ax) = £
A;=0
a ^x - °)
fc
Ax) = j£ a k {x - of
fc=0
ak = ~f m (a), k = 0,1,2,....
Theorem 1-32. (Taylor's formula with remainder.) Suppose that f and its
first n + 1 derivatives are defined and continuous on the interval I defined by
\x — a\ < R. Then for all x in I, we have
Ax) = S^rr1
(*
k\
" ")" + *»(*>* ( M5 >
k=0
where
f(x)-f(a)= f f'(t)dt.
Ja
* Recall that the formula for integration by parts may be written in the form
fu dv = u(v +c) f(p — +
c) du, where c is an arbitrary constant.
1-8 | TAYLOR SERIES 667
Example 1. The function /(*) = x 8/3 has continuous derivatives through order
two on the interval — oo < x < oo Thus, when a = . 8, Taylor's formula yields
Iff has derivatives of all orders at the point a, it is only natural to consider the
infinite Taylor series
- f \a) k
Z^ix-af,
fc=0
(1-17)
and ask whether this series converges to fin an interval \x — a\ < R. Applying
Theorem 1-32 we can assert that (1-17) converges to /whenever \x — a\ < R if
and only if
lim R k (x) =
k—>oo
for every x in the interval. Thus to settle this question it suffices to determine the
behavior of Rk(x) as k — > oo . To this end the following result is particularly
useful.
(n +
a\
1)!
(1-18)
\Rn(x)\ = ± J
(x- t)
n
f
{n+1
\t)dt
{x - t)
n
dt
n\
n+1
= M (xin-+a) 1)!
668 INFINITE SERIES | APPENDIX I
Example 2. If f(x) = e
x
, Taylor's formula yields
x
e = i + ^ + ^r+--- + ^+ *»(*)»
with
Rn(x) =
hi {X ~ °V *'
w+1 (jc)| = x
< e
R on the interval < R, Theorem 1-33 yields
Since |/ \e \
|jc|
lim R n (x) = 0*
n—>a>
i.e. that e is represented by its Taylor series on the entire real line.
e = e
1
= ^ h + *»0)
fc=0
and
I^ (1) l^ e <
(^FT)! (^FT)T
-7 =
This latter expression may be made smaller than 10 by choosing n 10.
*-ov.+ D!
converges by the ratio test. Hence its general term tends to zero.
1-9 FUNCTIONS DEFINED BY INTEGRALS 669
Example 3. Since the derivatives of sin x and cos x are each bounded by = 1 M
on — oo < x < x the remainder terms of their Taylor series are bounded by
,
n+1 /(n
\x\ + 1)!. Thus in each case the Taylor series converges to the respective
function on the interval — oo < x < oo The reader can easily show that the .
x3 ,
x5
cos* = 1 - ~+^ •
x x
sinhx = \{e - e~ ) = x + ^+^+ • • •
,
^+~+
x x
cosh x = \{e + e~ ) = 1 + • • •
.
d
F(s) = [ f(s,t)dt
Jc
Proof. We have
d
\F(s + h) - F(s)\ = I f [f(s + h, t) - f(s, /)] dt I
\
Jc I
d
<
J
f \f(s + h,t)- f(s, 01 dt.
c
Proof. We calculate
d
f
+ - = + h,t)-
\
[F(s h) F(s)]
Jl \f(s f(s, t)] dt.
1
+ - = V.(s +
h
\f(s h, t) f(s, 0]
fs
eh, t),
FXs) _
h-+o
lim
F(s + *)
h
- m _ lim [
h->oJc
d
3j (s
ds
+ „,,()*,
and since df/ds(s, t) was assumed to be continuous, Theorem 1-34 yields the de-
sired result. |
The formula of Theorem 1-35 can be extended to allow variable limits on the
defining integral, as follows.
d
Ns
V^^! )
tf(.) / \c(s)
t(s,t)dt
c
G(s,u,v) = f f(s,t)dt.
Ju
Then F(s) = G(s, c(s), d(s)), and the chain rule for functions of several variables
yields
= +
F'(s)
|f
(s, c(s), d(s))
H (s, c(s\ d(s))c'(s)
+ ^(s,c(s),d(s))d'(s). (1-21)
^(s,c(s),d(s)) = f(s,d(s)),
These equations together with (1-21) give the desired formula known in the liter-
ature as Leibnitz's formula. |
We now turn our attention to improper integrals and let / be piecewise con-
tinuous on < x < oo (see Section 9-2).
r B
- / /(*) dx < €
Jo
whenever B > M.
In a similar fashion, j_ aB f(x) dx is defined for a piecewise continuous function
on — oo < x < oo , by the double limit
lim
a-*-* [ f(x)dx.
B-><*> JA
672 INFINITE SERIES APPENDIX I
on this interval. The situation here is analogous to that which arose in defining a
function as the pointwise limit (sum) of an infinite series; for example the con-
tinuity of/(s, on the region < / < oo, s in I, does not imply the continuity
of F(s) on I. For this reason we extend the notion of uniform convergence, as
follows.
B
m- f
Jo
/(s,t) dt <j
for every s in /. Then
+ f(s,t)dt - F(s)
^+ f(s + h, t) dt f(s,t)dt
Now choose 8 > so that the latter term is less than e/3 whenever \h\ < 8.
f
f Jc /0, t)dtds = f
f f(s, t) ds dt
Ja Jc Ja
(provided, of course, that /is continuous). However, this result is in general false
if the integrations are carried out over unbounded intervals. To examine this
situation more closely, we must define what is meant by improper double integrals
and explore their relation to the improper iterated integrals.
Definition 1-15. Let R be the first quadrant of the s, t plane. We say that
the improper double integral
/]>'> ds dt
converges to L if for every e > there is a positive number M such that
rB r A
L - [ f f(s, t)dsdt\ < e
Jo Jo I
Thus if we let
*B rA
F(A,B)= f [ f(s,t)dsdt,
Jo Jo
then
r f'/(*.
Jo Jo
ds dt = lim
B—»oo A—>oo
lim F(A B)> >
r r f(s
JO JO
> o dt ds = iim
A-*ao 5-»oo
iim p( A B^ >
The question of equality of the three integrals is then just a special case of the
corresponding problem for limits. Thus the following result is of interest in this
context.
.
Theorem 1-38. Let the double limit L = \\mA-+n,B-+*> F(A, E) exist; i.e.,
assume that for every e > there is an M> such that \L — F(A, B)\ < e
whenever A > M
and B > M.
lim L(B) = L.
lim L(A) = L.
Thus, given > 0, choose M so that \L — F(A, B)\ < e/2 whenever A > M
e x x
and B > Mi. Now if B > M is held fixed we can find an M > Mi such that
x
\F(A, B) - L(B)\ < e/2 whenever A > M. Hence \L — L{B)\ < e, and since
this can be done for any B > M x, the proof is finished. |
Restated in terms of integrals, Theorem 1-38 asserts that if the double integral
L = Xfi?/(5, ds dt exists, then the existence of j f(s, t) dt for every s implies
that the iterated integral jo fo f(s, t) dt ds exists and equals L (and similarly with
the opposite order of integration). More interesting, however, is the converse
problem of inferring the existence of the double integral from the existence of one
of the iterated integrals. For this, the notion of uniform convergence enters once
again, and we accordingly restate Definition 1-14 in a form appropriate to limits.
whenever A > M
We can now state the main result.
(2) if lim^^oo F(A, B) = L(A) exists for every A, then linu^ L(A) = L.
Given e > 0, choose M x so that the first term on the right-hand side is less than
e/2 whenever B > M± and choose M 2 (using uniform convergence) so that the
second term is less than e/2 whenever A > 2 With = max {M 1} M . M M 2} we
have \L — F(A, B)\ < e whenever A, B are both greater than M. (2) The
second result now follows from Theorem 1-38. |
For the special case of improper iterated integrals, the above theorem takes the
following form:
ff f{s,t)dsdt = f f f(s,t)dtds.
JJr Jo Jo
X
f F(s)ds
= f f f(s, t) dt ds = f* f f(s,t)dsdt.
Ja Ja Jo Jo Ja
Proof We need only apply the theorem to the function f(s, t) defined on
0<s<ao,0<f<ooby
/y„ a = f/(
5 > 0, if a < s < b,
JK ' ' otherwise.
(0,
The above theorems treat the problem of integrating functions defined by im-
proper integrals. A useful result concerning the differentiation of such functions
is the following:
rtx>
F'(s) = / ?f(s,t)dt.
Jo ds
U(u,t)- f(a,t)]dt
= F(u) - F(a\
One of the main examples in the text of a function defined by an integral is the
Laplace transform
st
£L/TO = re- f(t)dt = F(s) (1-23)
Jo
ht
= I
C(e~ - X)e- f{t)dt\
St
I Jo l
< f(l - ht
e- )e-
st
\f(t)\ dt.
Jo
Thus if a, M are any constants chosen so that \f(t)\ < Me ai , it follows for s > a
that
ht ~ a)t
\F(s + h) - F(s)\ < MP(1 - e- )e-
is
dt
Jo
Ls — ot h + S — a
r
1-9 I
FUNCTIONS DEFINED BY INTEGRALS 677
and hence that F(s -\- h) — F(s) tends to zero as h —> through positive values.
A slight modification of the argument yields the same result if h — » through
negative values, and therefore the continuity of F(s) is established for s > a. We
have thus proved the following theorem.
/oo /•oo
cT st n st
e- f(t)dt= (-D"J o t e- f(Odt
ds r j
given in Section 5-5 of the text by showing that each of the integrals
r t
n
e-
st
f(t)dt, n = 0,1,2,..., (1-24)
Jo
converges uniformly on a < s < oo where a > a (see Theorem 1-42 for the
definition of a ). In fact, choosing Si (a < Si < a) and M so that
1/(01 < Me s i\
we have, for s > a,
- Sl)t
I
I
f
JA
t
n
e- stf(t)dt\
I
< MA J
t
n
e~
{s
dt
- Sl)t
< M TA J
t
n
e-
{a
dt.
* The number ao, sometimes called the order of f, is greater than or equal to the
abscissa of convergence sq of /. As was shown in the text proper, however, we may
have 5q < ao-
APPENDIX II
lerch's theorem
Let/(r) be defined on < t < oo, and be piecewise continuous on every finite
interval t< <
Assume,
A. moreover, that f(t) is of exponential order, i.e.,
that there exist constants a and such that \f{i)\ < Me
at
M
< t < oo. It is ,
for all s > So, s some constant, then fif) is identically zero {except possibly
at its points of discontinuity).
P(x) = £
fc=0
«***
/
J°
st
e- P(e- )f(t)dt= /
l
Jq
e-
st
£a e- k
kt
f(t)dt
k=0
/OO
kt
= !>*/ e-%- f(t)]dt
n
= ^2 a k <t>(s + k) = 0, s > s .
k=0
1
_1
f jt' P(jc)/(-ln x)dx=0, s > s .
Jo
^i-i|/(_l nx) |
< Mx s *- l
e
a( - lnx)
= MxSl - {a+1 \
and it follows that the function
_1
G(x) = Jc
si
/(-lnjc), < x < 1,
678
APPENDIX II 679
G(x)P(x)dx =
/.o
for every polynomial P. We shall deduce from these conditions that G(x) =
for < x < 1 (except possibly at its points of discontinuity). In fact, let us
choose a complete orthogonal basis for the vector space (PC[0, 1] with inner
product f g = J f(x)g(x) dx*
• Then any piecewise continuous function g
which satisfies
1
f g(x)P(x) =
Jo
Except for the fact that G may have infinitely many discontinuities, the proof
would be complete. Fortunately, it is not difficult to show that a complete ortho-
normal basis for (P(B[0, 1] is also a complete basis for the slightly larger class of
functions which, like G, may have infinitely many jump discontinuities but are
bounded, f We thus conclude that G is identically zero wherever it is continuous,
= *s -1
and hence since G(x) i
/(— In x), the same must be true of/. |
g(x)Pn (x) dx = 0, n = 0, 1, 2, . . .
J
Thus the piecewise continuous function g must be zero (except at its points of discon-
tinuity) and hence the same is true of g.
f The function G(x) is, in fact, piecewise continuous in every interval A < x < 1
(for < A < 1) and is continuous at x = 0.
APPENDIX III
determinants
lll-l INTRODUCTION
We shall present here a brief introduction to determinants, with sufficient attention
if and only if a 1? . .
.
a n are linearly dependent. Such a function, when defined,
,
n
for any real numbers a, and any vectors a(, a^ . .
.
, a re in (R .
680
IH-1 INTRODUCTION 681
Similarly if n
, = 2 and the vectors a x = 011^1 + 02^2 and a 2 = 012^1 + 022^2
are given, then I and II yield
JD(a 1 ,a 2 ) = D(a xl ei + 2X e 2 a l2 e + a 22 e 2 ) , x
= fluZ)(ei, ai 2 e + a 22 e 2 ) + a 21 D(e 2
x , a 12 e x + 22 e 2 )
= aii[ai 2 D(e ,e ) + 22 £>(ei, e 2 )]
l 1
+ a 21 [a 12 D(e 2 e x ) , + a 22 D(e 2 e 2 )] ,
= a 11 a 22 D(e 1 e 2 ) , + a 21 a i2 D(e 2 , e x ).
Z>(a l5 a 2 ) = a xl a 22 — a 21 a 12 .
ai = (
: ' • • • > an = ( :
0nl "
' '
®nn/
In this case we follow the usual custom of denoting its value by det M or by
\an\. Thus
«11 012 01n
021 022 02n = D
0nl 0»i2 " " "
Q-nn ^0nl
The results of the example above then take the more familiar forms
|011 012|
|0l| = 01 and = 011022 021012-
|021 022l
D(sl 1 , . .
.
, a,-, . . . , a„) = D(a u . .
.
, • a,-, . . . , an )
= 0- D(a lf . . . ,a„ . . . , an ) = 0. |
= D(a x ,
. . . , a,- + a y ... , , a{ + ay, . .
.
, an )
= D(a.i, . .
.
, a{, . .
.
, ai, . . . , an )
~r "(ai, . .
.
, 3j, . .
.
, ay, . . . , an)
+ D{a x ,
. . . , ay, . . . , ay, . . . ,a n ).
But the first and fourth terms of this sum vanish (by Condition I) and the sum of
the second and third terms is therefore zero as desired. |
Now suppose that we take some permutation e p(1) ep(2 ), , e p(n) of the , . • .
Ill— 2 |
BASIC PROPERTIES OF DETERMINANTS 683
the plus or minus sign being chosen according as the number of interchanges
required for this rearrangement is even or odd. It is of course an essential fact
(which we shall not prove here) that the number of interchanges is either always
even or always odd for all possible ways of carrying out the above rearrangement.
The permutation p itself is accordingly said to be an even permutation or an
odd permutation depending on which of these two possibilities holds.* Let us put
.
__ (+ 1 if p is an even permutation,
{
— 1 if p is an odd permutation.
n
an = a ln e l + a 2n ^% + • • •
+ a nn e n =
3=1
^ a jn tj,
n n n \
Z>(a 1? . . . , an ) =
Z
( y=i
a J'
ie J» S
y=i
a & e J> • • • ' £
j=i
a i**i
)
'
n / n n \
= 2
j i=i
a hi D \*h>
\
Yj
y=i
a i&i> • •
•
' 2
y=i
a '» e
>
/
n n j n \
= 2
y,=i
fl
J'ii 2
y 2 =i
a i22^(ei
\
1 ,
ey 2 , • • • , 2
j=i
fl y« e i)
'
= Z) a iii
Z) ^ 22
• • •
J2 a Jnn D{e h e J2 , , . . . , e jn )
3i=i h=i y n =i
n n
=
3l
S= S l 32=1
" "
3n=l
S a hi a h2 • ' ' a Jnn D{e h e J2 , , . . . , ey n ).
In this last sum, however, the only terms which are different from zero are those
for which the sequence j u ,jn is a permutation of the numbers 1, ...,«; for
. . .
integers for which / < j and p(i) > p(j). The permutation p is even or odd according
as I is even or odd. For example the list 3, 1, 2, 5, 4 contains three inversions, hence an
odd number of interchanges is required to rearrange this list into the natural order.
1 1
684 DETERMINANTS |
APPENDIX III
where the notation indicates that the sum is extended over the «! permutations of
The derivation of Eq. (Ill— 3) from Properties I through III of Definition III—
demonstrates that // there is a function D which satisfies these properties, then its
values must be given by Eq. (IH-3). Thus there most one function (for each ri) is at
satisfying the conditions of Definition III— 1. Moreover, it is not difficult to show
that the function defined by Eq. (Ill— 3) does in fact satisfy I through III. We shall
omit these details and merely summarize the results in the following theorem.
Theorem III — 3. For each n, there is one and only one function D satisfying
Properties I through III of Definition III— 1. Its values are given by the formula
an 012 * *
#ln
we note that the right-hand side of (III—4) consists of the sum of n\ terms, each
one a product of n factors chosen from among the entries of in such a way M
that no two of these factors occur in the same row or the same column of M.
Example. We can also use (III—4) to obtain the result of the example in the
preceding section. For the value of the 2 X 2 determinant is given by (III— 4) as
D(a u a 2 ) = «U 012|
021 022
= a(pi)ana 2 2 + <^{p 2)^21^12 = 011022 — 021012,
the signs of <r(pi) and (r(p 2 ) being determined by counting in their respective terms
the number of inversions in the arrangement of (first) subscripts.
For practical purposes, Eq. (Ill—4) is of little use for determinants of order
greater than three. Indeed the expansion of a 4 X 4 determinant would have
24 terms, that of a 10 X 10 determinant would have 10! = 3,628,800 terms,
and so forth. The remainder of this section and the next are devoted, therefore,
to properties of determinants which lead to simpler procedures for computing
their values.
From this point on we shall use in most cases the matrix notation
a nl an2 '
' ' ann
as its rows. Multiplying a column (row) by a real number a or adding two columns
(rows) is to be interpreted, therefore, as performing these operations on the cor-
responding vectors. We shall allow ourselves the usual misuse of language which
confuses a function with its values, and in this case we shall often speak of the
determinant |a^| when we are really speaking about the matrix (an) or about
the value of the function D on this matrix. Context will always provide the exact
meaning.
— 2~t
<T \P) a ip(.l) a 2p(2) ' ' '
Clnp(n)-
Thus the n\ products which enter into the expansion of det are precisely those M
which occur in the expansion of det M' and we need only show that they occur ,
686 DETERMINANTS | APPENDIX III
#p( 1)1^2(2)2 ' ' ' ap{n)n and a \q(l) a 2q{2) ' ' '
^nq{n)
be two corresponding products, i.e. products which differ only in the order of their
factors. If we apply the permutation p to the second product (or more exactly to
the subscripts of its factors) we obtain
But by definition of the transpose this must agree with the first product, and hence
p(q(i)) =
i for / 1, 2, =
n. This implies that either p and q are both even
. . . ,
permutations or else both are odd. Thus a(p) = cr(q). and we are done. |
The theorem may be stated informally by saying that the value of a deter-
last
minant is unchanged if its rows and columns are interchanged. This result permits
us to concentrate on just the columns of a determinant. Each theorem that we
prove will remain true if the word "column" is replaced by "row" and vice versa.
Certain useful properties of determinants are immediate consequences of Defi-
nition III— 1.For example, it follows from Property I that the value of a determin-
ant is multiplied by the real number k if each of the entries in a single column or
row are multiplied by k. We also easily obtain the following useful result.
D(a 1? . .
. , ^+ knj, . .
.
, ay, . .
.
, a„)
= D(a lf . .
.
, a,-, . .
.
, a,-, . . . , a„) + kD(a 1} . .
.
, ay, . .
.
, ay, . .
.
, a n ),
an * " " a \n
Z>(a 1? . .
.
, an ) = :
... J =0.
Un i Linn
Z>(ai, . . . , a l5 . . . , an ) = D(a u . . .
, ^
j=i
c i a J> • • • > a «)
— / j
CjD(&i, . . . , ay, . . . , a n ).
3
But in this sum, the /th term vanishes because c* = 0, and each of the other terms
also vanishes because in each case two of the vectors to which D is applied are
equal. |
any n — \ vectors of <R n Applying Theorem III—3, we write the value of this
.
1 a 12 din
a 22 = z2 G(p)aP(in aP(2)2 ' • • a p{n)n . (Ill— 5)
an2
But the only terms in this sum which are different from zero are those for which
the permutation p satisfies p(l) = 1. Moreover, for every such permutation, the
number of inversions in the sequence p(l),p(2), . . . ,
p(ri) is the same as the number
of inversions in the shorter sequence p(2), pi}), . . . ,
p(n). The value of (III— 5),
therefore, reduces to
2 <1
ff(?K(2)2««(3)3 Qq{n)n>
&nl '
' '
Onii—l Qn ,j + i
(III-7)
By applying the corollary to Theorem III—2, we may move the y'th column j — 1
places to the left thus multiplying the value of the determinant by (— l) y_1 Then .
applying this corollary once again (in conjunction with Theorem III—4), we move
the z'th row up i — 1 places, this time multiplying the value of the determinant by
688 DETERMINANTS |
APPENDIX III
1-1
(— l) . The resulting determinant now has the form (III— 5) and hence can be
reduced to an (« — 1) X (n — 1) determinant by applying Eq. (Ill— 6). The
value of D = Z)(ai, . .
.
, ay_ l5 e*, ay +1 , . .
.
, an) is thus given by
D = (-ly+JMv, (III-8)
where Aft-y is the determinant obtained from (III—7) by deleting the rth row and
they'th column. We give formal status to this new (« — 1) X in — 1) determinant
in the following definition.
and
D = £ aijAij, i = 1, 2, . .
.
, n. (111-10)
D = D(a t ,
. .
.
, ay, . .
.
, an )
n
= ^ aijD(*u
i=l
. • • , e i5 . . . , an )
n
= / J
aijAij.
i=l
This proves (HI-9), and (111-10) now follows from Theorem HI-4, for if Bi} is
important corollary of the last theorem, and apply these results to obtain the
converse of Theorem III— 6.
and
2=
3 1
aijAki = 0, i, k = 1, • • , n, i*k. (111-12)
But the latter sum vanishes since it is the expansion by cofactors of the kth column
of a determinant whose y"th and kth columns are equal. Similar reasoning applies
Proof. The necessity of the condition has already been established in Theorem
III-6. To prove the sufficiency we must show, conversely, that if det (A) = 0,
Consider first the case in which the cofactor A pq of some entry a pq of A is differ-
/ j
Apj&j = 0.
3 =1
Since at least one of the coefficients (viz. A pq ) in this linear combination is not
zero, the vectors a l5 . . . , a n are linearly dependent, and we are done.
If the cofactor of every element of A is zero, then (except in the case in which
a_, = for all j, where the theorem is obviously true) there exists an integer r,
but for every larger submatrix M' we have det (M') = 0* Since the order of
rows and columns immaterial, we may assume that M
is in the rows lies first r
Now choose k so that r -\- I < k < n and let M ' be the submatrix
(an ai r tfi, r +l
'
M' Q rr ar , r ^-i
r+l
y^ ciijAj = 0, / = 1, . . . , r, or / = k. (111-14)
Since k was chosen arbitrarily from {r + 1, ...,«}, the relation (III- 14) holds
for / = 1,2, ... ,n, and hence is equivalent to
r+l
22 A J a J = °-
y=i
Since
A r+1 = det(Af) 5* 0,
we thus conclude, that ai, . . . , a r+i are linearly dependent and hence that the
same is true of a l5 . . . , an .
|
\aa\ = (111-15)
#nl an2
\
a a\ = 2 <r(p) a P an a P(2)2 ' '
'
a vWn, (111-16)
(1) If a single column (row) of |a z-y| is multiplied by the constant k, the value
of the determinant is multiplied by k.
(2) If two columns (rows) of \aa\ are interchanged, the value of the determinant
is multiplied by — 1
(3) The value of the determinant |a»y| is unchanged if a multiple of one column
(row) is added to another column (row).
(4) The determinant |a f vanishes if and only if its columns (rows) are linearly
,-|
multiple of another.
(5) The value of the determinant |a t y| is unchanged if its rows and columns are
interchanged.
(6) If each entry in a fixed column of |a t-,-| is multiplied by its cofactorf and the
results are added, the sum is the value of the determinant; i.e., for any fixed j,
l<j<n,
\a {j \
= £ auAiy (IH-17)
* It has been estimated that the fastest modern electronic computer would require sev-
eral centuries to compute and add the 100! terms of (111-16).
| See Definition III—2 for the meaning of the term cofactor.
692 DETERMINANTS APPENDIX
The corresponding statement for rows is also true. Namely for any fixed /',
1 < i < n,
Equation (III- 17) is called the expansion of |a*j| by cofactors of the jth column
and (III— 18) the expansion of |a^| by cofactors of the rth row. One consequence of
(4) and (6) is the following property:
each entry of a fixed column (row) of |a^| is multiplied by the cofactor
(7) If
of the corresponding entry of a second column (row), and the results are added,
the sum is zero. More precisely,
2 a a A ik 0, j, k = 1, . . . , n, yV k,
(111-19)
0, i,k=\,...,n, i t± k.
3=1
Example 1. For determinants of order <3, it is not difficult to apply (HI- 16)
directly. For the first-order determinant \a\ we obtain, of course, \a\ = a, and
for determinants of orders 2 and 3 we obtain, respectively,
The last of these formulas is frequently remembered by forming the six products
indicated by the arrows in Fig. Ill— 1 and assigning a plus or minus sign according
FIGURE III
111-4 I SUMMARY AND EXAMPLES 693
-1 3
1 -6 = 2-1-7 + 4-(-l)-(-6) + 3-0-2
7
- 4-1-3 - (-6)-2-2 - (-l)-0-(7)
= 14 + 24 + - 12 + 24 - = 50,
while
-7
= 3-1 - (-7) -4 =3 + 28 31.
1
Example 2. The rule for expansion by cofactors may be used to evaluate the
third-order determinant of Example 1. Expanding by cofactors of the 1st row,
for example, we have
-1 3
1 -6
1
2
-6 =
7
2
2 7
+ (-!)'
(-t 1) + 3
= 2 •
19 + 24 - 3 •
4 = 50.
2 -1 3
-1
-6 3
1 -6 = 2
7
+ 4
1 -6
4 2 7
= 2(7 + 12) + 4(6 -
3) = 50.
3 7 9 11
1 5 -7 -2
D = 6 2 4 9
8 2 -5
Multiplying the third column by —4 and adding it to the first column introduces
a second zero into the fourth row. We then add twice the third column to the
fourth column, introducing a — 1 into the lower right corner. Thus
33 7 9 29 -33 7 67 29
39 5 -7 -16 39 5 -39 -16
D 10 2 4 17 -10 2 38 17
2 -1 -1
694 DETERMINANTS APPENDIX III
the second determinant being obtained from the first by adding twice the fourth
column to the third column. If we now expand by cofactors the fourth row, we
obtain
-33 7 67 -33 7 67
D = 4+4,
(_!)*-*-*(_ i) 39 5 -39 = — 39 5 -39
-10 2 38 -10 2 38
-33 7 67
D - 2 39 5 -39
-5 1 19
2 -66
- 2 39 5 -39
-5 1 19
2 -66
= - 2 64 -134
-5 1 19
2 -66
D= (-2)(-l)
64 -134
= 2(-268 + 4224) = 2 • 3956 = 7912.
anXi + a 12 x 2 + • ' •
+ a ln xn = b\,
OnlXl + fln2*2 +
and let
D (111-21)
#rol #n2
(^ «*l^ilVl + (£ <*i2AiijX + 2 • •
+ f S a inAiijXn = ^ b i A H-
.
But the coefficient of xi is then just the value D of \aij\ itself, and the coefficients
of x 2 , .
.
, xn are (by Eq. Ill— 19) all zero. Thus we have
n
Dxi = J] M^
and if D ^ we have determined x 1 . Of course, instead of An we could have
used the cofactors of they'th column of (III—21). We would then have obtained
the formulas
Dxi = 2
i=l
biAi i> J = l >
2 '
,n
independent (see Theorem 2-12 in the text). But, by Property (4), this is equiv-
alent to the nonvanishing of D. Thus we have the following theorem.
2 Mi
j = 1,2, (111-22)
D
It is worth noting that the expression in the numerator of (HI-22) is just the
expansion (by cofactors of the y'th column) of the determinant obtained from the
coefficient determinant |a t-y| by replacing they'th column by the vector (bu . .
.
, b n )',
that is,
an ••• bi -•• a^
an \ bn
an a\i a lr j= 1,2, ,n. (111-23)
an \
When written in this form, (III—23) is called Cramer's rule for solving the given
system of equations.
x + 2y = 3
3 -5
D = 1 2
6 + 5 = 11,
696 DETERMINANTS APPENDIX III
14 -5
= 3 2 28 + 15
= 43
*i ~
3 -5 11 11
1 2
3 14
1 3 9-14 = -5
x2
3 -5 11 11
1 2
It is clear, of course, that the method of Examples 4 and 5 can be applied only
to systems of equations for which the coefficient matrix is a square matrix, and
for this reason solutions given by (III—23) are largely of theoretical interest.
anl '
' ' ann • • •
kd (111-24)
c ll ' ' ' c ln bu •
'
bi n
cn\ bnl
\
e ij\ = Z~i
<T (P) e Pd)i eP(2)2 ' " '
^p(2n)2n> (111-25)
where p ranges over the (2«)! permutations of {1, 2, ... 2n}. Since e^ = for ,
1 < i < n, n + 1 < j < 2«, the only terms of (III—25) which are different from
p(l),p(2),...,p(n)
and
pin + l),p(n + 2), . . . ,p(2n).
Ill— 5 I
MULTIPLICATION OF DETERMINANTS 697
<r(r)<T(s)a r (i)lClr(2)2 ' ' ' a r (n)nb s (l)lb s (2)2 ' ' ' b s (n)n,
= e P^n)2n
\
e ii\
Zl viP^PWl ' ' '
= ^ ^^ r s
cr (. r )°'( S ) a ran ' ' ' a r(n)nbs{l)l ' ' ' bs{n)n
= Wij\ '
\bij\,
flnl '
' ' a nn "
\<*n\
' \bij\
= (111-27)
2>n bin
-1
— 1 Ki
Denoting the first n columns of (111-27) by C l5 . . . , Cn , and adding the sum
2 bud =
i=l
611C1 + b 2 xC 2 + • • •
+ bnl C n
«11 * • •
din YL a \ibi\
an \ • • •
ann YL amb%\ • • •
\a%j\ ' h a\ =
\
-1 ••• b 12 ---b ln
-1 :
• • • — 1 bn2 - •
-bnn
flu *
•
' «in Hauba • • •
YLaubi
Y,an ib n
\bij\ = (111-28)
-1
... -i o •••
-1 ••• •••
•-1
\aa\-\bij\ = (-Dn 011 • ' a ln Ha ubn • • •
J2 a iib
-1 Haiibn Ha ubi
w
Kl-M = (-l)
-1 L,Qnibn H^nibi
Finally noting that
_1 ...
(-D
n = (-i) n (-ir = (-i^+d = i,
-1
we obtain
Haiibn Ha ubi
Wa\ •
\bn\ = (111-29)
zL^nibil
'
' ' a ln '4,1 •• bin
(^11
B
an\ ' * ' ann> \b n i •
•
"nn/
the determinant on the right of (III—29) is seen to be the determinant of the matrix
product of A and B, and we have
det(^) = ^ (A)
-1 -1 = =
For det 04 )- det 04) = det(^ ^) det (/) 1.
We
conclude our discussion of determinants with one further application of
Theorem III- 10. Namely, let A = (a,-y) be any n X n matrix, and consider its
characteristic polynomial*
det (PAP' 1
- XI) = det (PAP' 1 - XPIP~ l )
- _1
= det (P(A X/)/> )
_1
= det (P) det (A - XI) •
det (P )
= det (A - XI).
Theorem 111-11. If A and B are similar matrices, i.e., if there exists a non-
-1
singular matrix P such that B = PAP then A and B have the same char-
,
acteristic polynomial.
It is not difficult to show that two n X n matrices are similar if and only if they
represent the same linear transformation L: (R
n — > (R
n
(relative to possibly dif-
ferent bases). Since Theorem III— 1 1 asserts that the characteristic polynomial of
two such matrices is the same, it follows that the characteristic polynomial which
is associated with choosing a matrix representation for L is independent of
L by
the particular matrix used. This allows us to conclude that the eigenvalues for L
can be found by representing L in matrix form and solving the resulting charac-
teristic equation for X.*
APPENDIX IV
uniqueness theorems
IV-l SURFACES IN (R
3
, SURFACE AREA
Our objective in this appendix is to prove the uniqueness of the solutions obtained
in Chapters 13 through 15 for various boundary-value problems involving partial
differential equations from mathematical physics. For this purpose it will be
necessary to introduce several concepts from vector field theory, among them the
notion of surface integral. However these ideas will be developed only to the
extent required for the stated goal.
A surface S in (R 3 is said to be represented parametrically if it is the range of a
function r : (R
2 — > (R
3
. In this case S is the set of points (x, y, z) in (R
3
satisfying
x = x(u, v),
y = y(u,v), (IV-l)
2 = z(ll, V),
where x(u, v), y(u, are the coordinate functions of r and (u, v) lies in a
v), z(u, v)
2
region R uv of (R . Equations (IV-l) are called parametric equations for S. For
3
example the sphere of radius R in (R centered at the origin is represented para-
metrically by
x = R cos u sin v
y = i? sin w sin v,
z = R cos v,
where {u, v) lies in R uv : < u < 2w, < v < t (see Fig. IV-l).
A surface S may also be represented explicitly in the form
F(x,y,z) = 0. (IV-3)
We are familiar with each of these forms in the case of the sphere described above.
For example, the equation
x2 + y
2
+ z
2
= R2
700
3 SURFACE AREA 701
IV-1 I
SURFACES IN (R ,
(0,0,1)
(1,0,0)
defines this sphere implicitly, while the upper and lower hemispheres are explicitly
z = VR -2 x2 - y2 , x2 + y2 < R2 ,
and
= -VW - x2 - y
2
, x2 + y
2
< R2 .
It is usually the particular application at hand that determines which of the above
representations is used.
We shall say that a given surface S has a unique tangent plane at
is smooth if it
each of its points P and if this tangent plane varies continuously as P ranges over
the surface. In this case we can erect at each point P of S a vector N which is normal
to S at P (i.e., perpendicular to the tangent plane at P). For the explicit represen-
tation (IV-2) of S, the property of smoothness amounts simply to the continuity
of the partial derivatives df/dx and df/dy, and a normal vector N is then given by
(IV-4)
dx dy
3
where i, j, k are the standard basis vectors of (R . For implicit and explicit repre-
^
NT =
dF.
—I
dx
+
.
—
dF.
dy
dF,
+ -r-k
dz J
.
(IV-5)
provided that this vector does not vanish, while in the case of the parametric repre-
sentation (IV-1), the vectors
du ~ du dw
J
du '
(IV-6)
dr _ dx . . dy . . dz
dv
~ dv
l
au
J
au
702 UNIQUENESS THEOREMS APPENDIX IV
are tangent to the surface S and hence determine the tangent plane only if they are
linearly independent. In the latter case a normal N to S at the point P = r(u, v)
is orthogonal to each of the vectors (IV-6) and hence may be expressed in the
form
i J k
dx dy dz
N = — X-
du dv du du du
dx dy dz
dv dv dv
dy dz dz dx dx dy
du du du du du du
dy dz
i + dz dx
j + dx dy
(IV-7)
dv dv dv dv dv dv
i J k
2
Here N is easily seen to be the inward pointing normal at P and ||N|| = R sin v.
Employing, instead, the explicit or implicit representations of this sphere, Eqs.
i J k
Xi yi Zl
X2 yi z2
istwice the area of the triangle determined by ai and a2- These properties, together
with the fact that the vectors a 1, a 2, a 1 X
a2 (in that order) forma right-handed triple,
serve to determine ai X
a 2 uniquely.
3 703
IV-1 I
SURFACES IN (R , SURFACE AREA
FIGURE IV-2
i +,
J + k
^#2 _ X 2 _ y 2
"
v^ - 2 x2 - y
2
or
N = 2xi + 2y\ + 2zk.
We now consider the notion of surface area. Here we begin with a smooth
surface S and a parametric representation r = x(u, v)i + y(u, v)j + z(u, v)k as
described above. To obtain an approximation to the area of S we subdivide the
surface into a finite number of pieces AS; by means of a network of intersecting
curves corresponding to a grid of horizontal and vertical lines in the domain R uv
of r (see Fig. IV-2). The approximation is now found by replacing each of the
ASi by an appropriate portion of the tangent plane to S at a point in AS t and taking
area on the tangent plane as an approximation to area on the surface.
Specifically, let w and r(w) be as shown in Fig. IV-2 and let
dr
and U (w)
dv
be the tangent vectors to S (Eq. IV-6) at the point r(w). Then if AS; is the image
under r of the indicated rectangle with sides Aw,, Ai?», the area of the parallelogram
formed by the vectors
dr
ti Aui = AW,; t2 AVi
du dv
whenever Aw* and Av { are small. But, as was pointed out in the footnote on
page 702, the area of this parallelogram is just the magnitude of the vector
and hence, using (IV-7), we have the required approximation in the form
2 2 1/2
dy dz dz dx dx dy
du du du du du du
|N|| AuiAVi = < + + \ Au { AV{.
dy dz dz dx dx dy
, dv dv dv dv dv dv
(IV-8)
We now repeat the above computation for each of the AS;, add the results, and
take the limit as the number of pieces AS; is allowed to increase in the usual fashion.
The value of this limit(which can be shown to exist under the hypotheses in force)
is, by definition, the area of S, and we therefore have
2 2 2} 1/2
dy dz dz dx dx dy
du du du du du du
a(s) + + dudv.
dy dz dz dx dx dy
dv dv dv dv dv dv
(IV-9)
1/2
= d(y, zf d(z, xf d(x, yf du dv, (IV-10)
a(s) ~*~ "*"
J L d{u, v) d(u, v) d(u, v)
where
dX dy
d(x, y) du du
d(u, v)
dX dy
dv dv
d(xi, . . . , xn )
d(ui, . . . , un )
3.Xi dx 2 dxn
dwi dui dUi
dX\ dx 2 dxn
du 2 du 2 du 2
d*i dx 2 dxn
dun du n dun
3 SURFACE AREA 705
IV-1 I
SURFACES IN (R ,
(R
3
and a smooth two-dimensional
,
surface S in that region. Let S be subdivided
Fds -
If'
jj
s
is defined to be
lim 2 F(P?)a(ASO,
t=i
where the limit is taken in such a way that the diameter of each of the AS,- tends
to zero.* Here again it can be shown that this limit exists, and that its value is
independent of the manner in which S is subdivided and the P* chosen in AS,-.
Finally, to obtain a formula for evaluating this integral, let
x = x(u, v),
y = y(u,v),
z = z(u, v)
Fd$
2 2 1/2
d(y, if d(z,x) d(x, y)
F(x(u, v), y(u, v), z{u, vj)
d(u,v)
+ ,
^7^
d(u,v)
+
,
d(u,v)
dudv,
R,
(IV-11)
where the integral on the right is an ordinary double integral over R uv Integrals .
of this type are encountered in physical problems dealing with surface distributions
of matter where they admit interpretations as mass, moments, etc.
* By definition, the diameter of ASi is the least upper bound of the set of real numbers
|Pi - P2II with Pi and P 2 in ASi.
706 UNIQUENESS THEOREMS APPENDIX IV
R,
Example 3. Find the total mass M of a hemispherical shell S of unit radius, given
that the density of S is proportional to the dist ance from the b ase of the hemisphere.
Let S be represented (explicitly) by z = Vl — x 2 — y
2
, x2 + y
2
< 1. Then
parametric equations for S may be given in the form
x = u,
y = v,
z = a/1 — w2 — v2,
with u
2
+ v
2
< l. Since the density of S is given by the scalar function p = kz,
k a (positive) constant, Formula (IV-l 1) yields
M pd§>
1/2
d(y,zf d{z,xf d(x,y)
k\/\ — u2 — v2 +,
+ ,
dudv.
d(u, v) d(u, v) d{u, v)
But
whence
M= k J J
dudv = kir.
u
2 ,
-\-v
2 ^
S 1
1
The example suggests that we simplify Formula (IV-11) (and IV-10) for
last
surfaces S which are represented explicitly as follows. If
z = /(•*> y)> (X y) in R *
X = u,
y = v,
we have
d(y, z)
= _ df s
'
d(z, x)
= _df, d(x, y)
= {
d(u, v) du d{u, v) dv a(w, v)
IV-2 |
SURFACE INTEGRALS OF VECTOR FIELDS 707
3
over a smooth surface S in To do so we use the integral of scalar functions
(R .
introduced in the preceding section, and tentatively define the required integral to be
ff(F-n)d&, (IV-12)
s
where n = n(P) is a unit normal vector at the point P of S chosen in such a way
that F n
•
is a continuous function on S (Fig. IV-3). Informally this requirement
means that we must designate one side of the surface as positive, and let n(P) be
the unit normal at P which points in that direction. For many surfaces such a
choice of positive direction is easily made. On the sphere (or any closed surface),
for example, we may choose n(P) to be the outward pointing normal. And on a
surface defined explicitly by
=
- (df/dy)j + k
-(df/dx)i
n
VTTXdf/dxy + (df/dyy
3
Unfortunately, there are smooth surfaces in (R , so-called one-sided surfaces, for
which not possible to assign a positive direction. The most notorious example
it is
of such a surface (and aside from trivial modifications the only one) is the Mbbius
strip m, shown in Fig. IV-4.* It is clear that 9TI is genuinely one-sided in the
F(P)
sense that anyone who agreed to paint a single side of the strip would find that he
had actually contracted to paint the entire surface.More to the point is the ob-
servation that we can pass from one of the two unit normal vectors at P to the
other by moving the normal continuously along a smooth curve on 9TI. Thus
there is no natural meaning that can be attached to the integration of F n over am, •
and we must accordingly exclude the Mobius strip and its relatives from further
consideration. To this end we introduce the following definition.
3
Let us consider, now, an orientable smooth surface S in (R with positive unit
normal n, and let
If (F • n) d§.
from S), then the surface integral of F on S can be evaluated as an ordinary double
integral over the domain D uv of r by the formula
[**'>>&3+<*<*'>>&3
r
du dv. (IV-13)
IV-2 |
SURFACE INTEGRALS OF VECTOR FIELDS 709
para-
Before considering examples, several remarks are in order. First, if the
metric representation of S is such that the vector = t x X t 2 does not point N
from the above formula must be modified to read
in the positive direction S,
ff (F • n) dS = (F • N) du dv. (IV-14)
-ff
d
(F • n) c& = -P(x, y, f{x, y)) £- Q(x, y, f{x, y))
fy
+ R(x,y,f(x,y)) dudv. (IV-15)
(This formula will prove quite useful later.) an alternative and Finally, there is
very convenient notation for the surface integral of F on S which can be obtained
by writing
ff (F
• n) d$ =
ff
(Pi + &+ Rk)-n d§,
s s
For then
i-n = cos7 1} j-n = cos7 2 , k-n = cos7 3 ,
(i • n) d§>, (j • n) d§>, (k • n) dS
(i • n) d§> = dy dz,
(j • n) d§> = dz dx,
(k • n) dS = dx dy, FIGURE IV-5
710 UNIQUENESS THEOREMS | APPENDIX IV
The value of this expression is, of course, still given by (IV-13) or (IV-14).*
ff (F • n) dS
s
with < u < 2t, < v < ir, we obtain as in Example 1 of the last section
i J k
dx dy dz
N = du du du
dx dy dz
dv dv dv
But this vector is clearly an inward normal to S, and hence we must use (IV-14)
to evaluate £fs (F • n) dS. This gives
2
ff (F • n) dS = ff (sin
2
u cos u sin
3
vi + k) • (cos u sin v\
+ sin u sin
2
vj + sin v cos uk) c?m c/u
= r f
Jo Jo
(sin
2 2
wcos wsin
5
y + smvcosv)dudv
W
[ sin v dv [* sin +
2 2
= 5
u cos udu f sinucosurfu/ du
Jo Jo Jo Jo
15*
* The reader should not take these remarks too seriously; they have been given only
1 1 xz dx dy
xz dx dy jc(1 — x — y) dy dx
0^0
1 -il— x
2
xy x y dx
L
Example 3. Let F be the vector field associated with a time independent flow of
fluid in a region of 3-space [i.e., F(x, y, z) is the velocity vector of the fluid at the
point (x y, z)], and let S be an orientable smooth surface lying in this region.
Then // F is continuous and n
is the positive unit normal to S at a point P, the
unit time in the positive direction across a surface element containing P, i.e. to
the, flux across AS in the positive direction. The usual limiting argument now
applies and allows us to assert that
//(F-n)</S (IV-17)
z
'
(0, 0, 1)
Ay
ixQ ,yQ ,ZQ)
and suppose the box B to have edges of length Ax, Ay, Az situated as shown in
Fig. IV-7. Then the net amount of fluid leaving B through the shaded sides is
(approximately)
where < < 1 . Similar results express the rate of flow through the remaining
two pairs of parallel sides of B using dQ/dy, dR/dz instead of dP/dx. Hence if
+—
dQ
(dP
(
\dx
t
,
h ^
dy
dR\
dzj
Ax Ay ,
)
A A A
y Az
divF = ^
dx
+^
dy
+ ^.
dz
ov-18)
If the fluid is incompressible, then div F must measure the rate (per unit volume)
at which fluid is being introduced at the point x. Therefore, if S is a closed surface,
enclosing a region R, the integral
fff (div F) dx dy dz
R
represents the total amount of fluid introduced into the region R, and hence must
also represent the total flux across the boundary surface §. Comparison of this
result with (IV- 17) leads to the equation
This which would almost be self-evident if one were willing to accept the
result,
physical argument just given, is one of the basic relations in vector field theory.
One finalremark before leaving this section. Thus far we have limited our
is no need to be quite this
treatment of integration to smooth surfaces. There
we want to integrate over surfaces such as the surface
restrictive, and in fact shall
surface of a closed cylinder (Fig. IV-8). Although not smooth,
of a cube, or the
surfaces are constructed from a finite number of nonoverlapping smooth
these
pieces and are accordingly called piecewise smooth surfaces. The various integrals
surface S simply by
of these two sections are extended to a piecewise smooth
integrals over the smooth pieces of S. The only question which presents
adding the
real difficulty in this connection is that of orientation. However, for a piecewise
3
smooth surface S which bounds a finite region of (R i.e. which has an inside and,
In this section we shall establish the equality given in Eq. (IV- 19) of the preceding
section, known in the literature as the divergence theorem, or as Gauss' theorem*
ff (F
• n) dS = (div F) dV. (IV-20)
fff
dV
dP dQ dR
P dydz + Qdzdx + Rdxdy = dx
~*~
dy
~*~
dz.
dV,
dV 'v
dP
P dy dz = dV,
dx
dV
dQ
Q dz dx dy
dV,
dV
dR
Rdxdy = fl J
j^dV. FIGURE IV-9
dV
And finally, since these equations are all of the same type, it obviously suffices
to establish just one of them.
This said, let V be an "elementary" region bounded above and below by smooth
surfaces Si and S 2 described, respectively, by the functions z =/2(x,y) and
z = fi(x, y) where (jc, y) ranges over a region D in the xj-plane. Let S 3 denote
the lateral surface (if any) of V. Then, by Formula (IV- 15),
Rdxdy = / /
(Rk) • n d$
§2
//(*">•(- i'-f^")^
R(x > y> f2(x, y)) dA.
Similarly
(where the minus sign occurs because the positive normal to S i points toward the
interior of V), and
Rdxdy = / /
(Rk) • n ds = 0,
JJ
Rdxdy =
JJ
[R(x, y, f2 (x, y)) - R(x, y, Mx, y))] dA
f2( x <y'>
d
R(x, y, z) dz dA = -*dV '
fi(x,y) dz dz
the theorem to each of them in turn. Then, since the integrals over those portions
II (F-n)d$ = ^2
JJ
(F-n)^ ^^ ~^%k
n fff / Mj
= E
1=1
///
(divF)JF / ^ y
(div F) dV, X^^— «A
Example 1. Use the divergence theorem to evaluate the surface integral con-
ff xy dy dz
2
+ dx dy,
2
x2
2
where S is the surface of the unit sphere + y + z = 1-
By (IV-20) we have
2
2
xy dydz + dxdy =
fff
y dV,
ff
s v
2 2
2
xy dy dz + dx dy = / / (r sin <p sin d) r sin <p dr d<p dO
/2t pT /*1
2 3 4
/ sin Odd- I sin cp d<p •
I r dr
Jo Jo Jo
= TrdXi)
_ 4ir
~ 15
'
2
// xy dy dz + x dz dx + (x + z) dx dy
To evaluate this integral let V denote the region bounded by S and the plane
z = 0, and let §>i denote the closed disc x 2 + y
2
< 1 in the xy-plane. Then we
write
2
If
xy dy dz + xdzdx + (x + z) dx dy
s
2
= [J
xy dy dz + x dz dx + (x + z) dx dy
dV
+ ff
2
xy dy dz + x dz dx + (x + z) dx dy.
§i
(The reader should note that the plus sign appearing here is not a misprint. Why?)
But by the divergence theorem, we have
2
ft xy dydz
2
+ x dz dx + (x + z) dx dy =
fff
(y + 1) dV
dV v
2
= I f f t( r s * nv s * n 0) + W r2 s n ^ ^r d*P d®
*
Jo Jo Jo
2* 12 l
r2 * r*l 2 l
o ^ r a r r* f 2
= / sin^ Odd- I sin
d
<pdp- I r*dr + I
Jo
dd
Jo
sm <p d<p
Jo
r dr
Jo Jo Jo
_ 2tt 2t _ Air
~ TI + T= T'
Finally,
//
xy dy dz + x dz dx + (x + z) cfx d[y
[xy
2
i + jcj + (x + z)k] • k </S
- //< + z)dA
x +j/ < 1
2=0
Vl-X 2
jc dy dx = 2 / x\/l — x 2 dx =
-1 ^-Vl-x 2 -/-l
IV-3 1 THE DIVERGENCE THEOREM 717
and we have
2
xy dy dz + xdzdx + (x + z)dxdy = -y
divergence of a con-
In Eq. (IV- 18) of the preceding section we defined the
tinuously differentiable vector field by the equation
div F = —+
3a;
-^7
dj>
+ -T-
dz
'
where P, Q, and are the coordinate functions of F with respect to the standard
i?
coordinate system, its physical interpretation suggests that its values remain un-
changed under a change of basis. We are now in a position to prove this fact by
3
giving a description of div F which does not involve a coordinate system in (R .
To this end we recall that if/is real valued and continuous in a region R of 3-space,
the mean-value theorem for integrals asserts the existence of a point (x*, y*, z*)
in R such that
ffff(x,y,z)dV=f(x*,y*,z*)V,
R
where V denotes the volume of R. Thus if x is any point in the interior of the
domain of F, and (B e is the sphere of radius e about x, then
of div F(x) given in the preceding section. Indeed, if F is the velocity field of a
time independent flow of fluid,
ff (F-n)rfS
718 UNIQUENESS THEOREMS APPENDIX IV
is, as we have seen, the flux crossing the surface of (B e in the positive direction.
Hence, in the limit, (IV-21) measures the amount of fluid diverging from x per
unit volume.
+
. .
The rules for manipulating this symbol are perfectly simple and go as follows.
If i/' is a differentiable scalar function, then
dP dQ dR
VF dx By Bz
divF,
and finally
B_ B_ B_
V X F
Bx By Bz
P Q R
(BR _ BQ\ 1+,(BP_ BR\ } ,(dQ_ BP\
\By Bz~/ \Bz BxJ ^\Bx By/*'
a quantity called the curl of F. Finally, carrying this notation to its logical con-
clusion, we set
2 2 2
B B B
T
,
,
Bx 2 ^ By 2 Bz 2
2
and write V V = V2
• . (The operator V is sometimes called the Laplacian.)
IV-4 |
BOUNDARY-VALUE PROBLEMS REVISITED: UNIQUENESS THEOREMS 719
Since V F is just
• another way of writing div F, the divergence theorem may be
rewritten in terms of V as
we find that
V F = . u(V 2 v) + (V«) • (Vv) (IV-24)
This is Green's first identity. To derive the second we simply interchange u and v
in (IV-25) and subtract to obtain
V2 u = in V,
u = <p on S.
Proof. Let u x and u 2 be two solutions of the given problem and set U= u\ — u2 .
V2U =0 in V,
U= on S,
and we will be done if we can show that U= in V. To this end we note that when
u = v Green's first identity becomes
[u(Vu) • n] dS = [u(V u)
2
+ (Vh) • (Vw)] dV. (IV-27)
ff fff
av v
720 UNIQUENESS THEOREMS | APPENDIX IV
jlj{(yu).(yu)]dv = 0,
v
or
///[©' +©*-© dV = 0;
and since the integrand appearing here is continuous and nonnegative everywhere
in V, we have
©' +($+©-«
in V. This, in turn, implies that
<W _ dU = dU =
~~
dx dy dz
Theorem IV-3. Jf$ and V are as above, there exists at most one solution
of the three-dimensional heat equation
Proof As in the preceding proof let U denote the difference between any two
solutions of this problem. Then U is a solution of the problem
9 dU __9
a
a,
~ v "•
u(x, y, z, t) = on S, u(x, y, z, 0) = 0,
2
But V U = 2
a (dU/dt). Thus
d 2
a
2
U -£+
at
vU \\ \\
dV=0,
'
v"
or
///|(i^VK=-///l|V^K.
v v
i\\\
dt
i^U dV= - 2
) \\VU\\
2
dV. (IV-28)
Now set
2 2
/(0= fff(ia U )dV.
Then 7(0) = 0, while 7(0 > and 7'(0 < for all / > 0. A straightforward
application of the mean value theorem then implies that /(/) = 0. Indeed, from
the mean value theorem we deduce that
and since 7(0 > and I'(t) < 0, we must have 7(0 = 0. Thus the left-hand
side of (IV-28) vanishes, and it follows that
2
\vU\\ dV = 0.
///
(Note that this argument can be used equally well for any of the other boundary
conditions considered earlier.)
Finally we consider the uniqueness problem for the wave equation.
Theorem IV-4. With S and V as above, there exists at most one solution of
the three-dimensional wave equation
initial conditions
du
u(x, y, z, 0) = G(x, y, z), (x, y, z, 0) = H(x, y, z)
Tt
2d U _2
u(x, y, z, i) = on S,
u{x, y, z, 0) = —
du
(x, y, z, 0) = in V,
dU
dt
V2 U + V {—) •
Vt/J dV =
JJ
~ VU- n d% = 0. (IV-29)
But
dU
dt
and
(VC/) V^1|||V^;
so that (IV-29) becomes
Dmes
+ vi/ JK = 0. (IV-30)
Thus the integral in (IV-30) does not depend on t, and since the initial conditions
require it to be zero when t = 0, we conclude that it is identically zero. This
implies that
2(dU\
and hence U must be a constant. Again using the initial conditions, this constant
must be zero, and we are done. |
further reading
LINEAR ALGEBRA
Halmos, P., Finite Dimensional Vector Spaces, 2nd Ed.; Van Nostrand, Princeton, 1958.
N.J., 1961.
DIFFERENTIAL EQUATIONS
Birkhoff, G., and G. C. Rota, Ordinary Differential Equations ; Ginn, Boston, 1962.
Coddington, E., Introduction to Ordinary Differential Equations; Prentice-Hall, Engle-
wood Cliffs, N.J. , 1961.
Pontryagin, L., Ordinary Differential Equations; Addison- Wesley, Reading, Mass., 1962.
Weinberger, H., A First Course in Partial Differential Equations; Blaisdell, New York,
1965.
Yoshida, K., Lectures on Differential and Integral Equations; Interscience, New York,
1960.
Churchill, R., Fourier Series and Boundary Value Problems, 2nd Ed.; McGraw-Hill,
New York, 1963.
Davis, H., Fourier Series and Orthogonal Polynomials; Allyn and Bacon, Boston, 1963.
Jackson, D., Fourier Series and Orthogonal Polynomials; Mathematical Association of
America (Carus Monograph), 1941.
Lebedev, N., Special Functions and Their Applications; Prentice-Hall, Englewood Cliffs,
N.J., 1965.
Sansone, G., Orthogonal Functions, Revised English Ed.; Interscience, New York, 1959.
Tolstov, G., Fourier Series, Prentice-Hall, Englewood Cliffs, N.J., 1962.
APPLIED MATHEMATICS
Friedman, B., Principles and Techniques of Applied Mathematics; Wiley, New York, 1956.
Fleming, W., Functions of Several Variables; Addison- Wesley, Reading, Mass., 1965.
answers to odd-numbered exercises
Chapter 1
Section 1-1
10x + 5 2x + 1
11. No; undefined at x = 1
+ X + x - 6
Section 1-2
Section 1-4
1 (a) No; not closed under addition (b) Yes; satisfies subspace criterion
(c) Yes; satisfies subspace criterion (d) Yes; satisfies subspace criterion
(e) No ; not closed under addition
3. (a) Yes; satisfies subspace criterion
(b) No ; not closed under scalar multiplication
(c) Yes; satisfies subspace criterion (d) No; not closed under addition
(e) Yes; satisfies subspace criterion
7. (a), (b), (0 9. (a), (0 11. All constant functions
13. (a) x(ax + /?) (b) (x + l)(ax + (3) (c) ax 2
+ (3x + y
17. If 9C consists of the two vectors and<y the two vectors (1, 0, 1),
(0, 1, 1), (0, -1, 1)
(—1, 0, 1), then S(9C n <y) is the trivial subspace of V, but S9C fl S^ comprises all
vectors of the form (0, 0, X3). If 9C consists of the two vectors (0, 0, 1), (0, 1, 1)
and'y the two vectors (0, 0, 1), (1, 0, 1), then S(9C D^) and S9C n S'y each comprise
all vectors of the form (0, 0, X3).
725
726 ANSWERS TO ODD-NUMBERED EXERCISES | CHAP. I
Section 1-5
3. No; (0, 2, -1) and (0, *, -£), or (0, \, -£) and (0, f, -±)
5. i = 1(1, 0) + 0(0, -1) = cos 0(cos 0, sin 0) - sin 0(-sin 0, cos 0)
= -
a
(a, 0) + \ (0, 0) = 1(1, 1) + 0(0, 1),
p
a'i + 0'j = a'(l, 0) - 0'(O, -1) = (a' cos + 0' sin 0)(cos 0, sin 0)
23. (a) x3 + 2x + 5 = + 2x + 5)
l(x 3
(b) x2 + 1 = K3* 2 + 2) + ^(6)
(c) 2* 3 - x2 + lOx + 2 = 2(* 3 + 2x + 5) - J(3x
2
+ 2) + 1(6*) - ^"(6)
Section 1-6
(0 -2, 1, 2 (0 (10, 6, 3)
5. One such basis is (-3, 0, 0, 0), (0, 1, 0, 0), (0, 0, 2, 0), (0, 0, 0, -1).
Section 1-7
1. (a) 2 (b) 3
5. (a) 3 (b) 2 7. 2
Section 1-8
7.
Section 1-9
Chapter 2
Section 2-1
Section 2-2
5. (b) If 9C = {0}, then Ct(9C) is the space of all linear transformations from V i to V2
if SC = V 1, then G(9C) contains only the zero transformation.
Section 2-3
2
n n -
(b) L D y(x) = y(x) y(a) + /(«)(* - a) + ^r <* ~ a> +
(n— 1), N
^ (n - 1)P '
(c) 2xD 2
+ 2(1 - x)Z) - 1 (d) x 2 Z) 2 + 2x(l - x 2 )Z> - 6x 2
(e) x 2 Z) 2 + 2x(2 - x 2 )D + 2 - 4x 2
7. (a) 0, 0,
(b) 0,-6 sin 2x, 2e x
(c) -3 sin x - cos x, -2e x , 2 - 2x - 2x 2
2x
(d) 0, 0, 2e~
(e) Ax 2 Ax 2
,
- 21 x, e*(x
2
+ 2x - 2)
Section 2-4
1. (a) 9104) = 0; A~ x
(yi,y2) = (ivi, -&V2)
(b) 9104) is the line x2 = 0; there is no inverse.
(c) 9104) is the line xi + x2 = 0; there is no inverse.
y
(d) 9104) = 0; A-\yuy 2 ) = (^y^> -^~)
3. (a) 9104) comprises all constant polynomials; there is no inverse.
(A is not onto.)
5. ai/32 — «2|Si 9*
Section 2-5
3. Representation of aA is I
a^j
•
Representation of A + B is
^+^ ft + ^
5.
Representation of
dim £((R 2
,
2
(R ) =
AB is
4
^ + ^ ^+^
2-7 ANSWERS TO ODD-NUMBERED EXERCISES 729
tic m 2-6
1 1 o" 1 1 1 -1 o"
1. (a) 1 1 (b) 1 -1 (c) 1 2 1
-1 1
'tioi2 2-7
-2 1 -6
-11
20 -4
1. 3 7 -
3. 4 16
4 -8 8 L 3 3 .
2 1 1 8~
9 4 3 3
1 5 1 1
5.
3 2 2 6
35 7
. 18 18 i o
-i -2 -2
l 1
7. For standard bases,
1 1
0"
1 2
for bases »1, t»2,
1 2
1
11. If (Bx = {(e n ), (<? 12 ), (e 21 ), (e22 )}, CB 2 = {(e'u), (e[ 2 ), (e{ 3 ), (e'21 ), (e'22 ), (e23 )},
then [A:(S,u (B 2 ] = 1
2
-2
13. (b) -3 -4 15. mx)(x) = xv-3
2 -1
-2 2
730 ANSWERS TO ODD-NUMBERED EXERCISES | CHAP. 2
Section 2-8
18 -16 1
31 -25 1
8 2-4 12
_2_ 1
a/3 /3
inverse
1
fl 1
a(/3 — a) a(/3 — a)
inverse
a _l
j8(j8 - a) 0O - a)
(c) Invertible if and only if a 9^ 0;
£ I
a a
inverse -1
1 _^ J_
.a a' a2
a "1 0~ "-1 0~
1 - a2 a, arbitrary except that /3 ^ 0; also and
a , fi
7 -ij |_
7 1
1 -1
7 arbitrary; and
1 -1
a
™2 also
11. a, arbitrary except that j8 5^ 0;
, j8
7
0'
2
2 1
13.
-1
-1 -6 1
15. (a)
£ -1* (b)
3
(c) ± 9
.2
2
(d) The matrix is an n X 1 matrix with entries corresponding to the n 2 matrices
(eij) of the image basis. Those entries which correspond to (en), O22), (£33),
• • • ,
(e nn ) are each 1 and all other entries are 0.
Chapter 3
Section 3-1
Section 3-2
3. (b) y = cie ax
cosbx + C2e ax
sinbx
y = 6e aa: cos bx
<*(b + 1)
(c) e ax sin bx
b
) ) . —
7. (a) y = c\ sinh x +
C2 cosh x
(b) j = cix 3 + C2X 3 ]nx
(c) y = ci sin 2x +
C2 cos 2x
+
1 + x
(d) y = c\x C2X In
1 - x
Section 3-3
1. y = —Q ' on (— oo , 0) or (0, oo
3. y =
sin x
on far < x < (k + 1)tt, k = 0, ±1, ±2, . . .
~x
5 . y = e + ce-^ l2) \ on (-00,00)
-(fi/£)
7. /
R
+ ce , on (-00, go)
x
= ce — (2x + l)e / n
9 _y ^\ <x
> on (— °o , 00
4(*2 + 1)
t
e{x - 1) + c
11. y = far < x < (k + 1>, A: = 0, ±1, ±2, . . .
x sin x
= + *2 *
.13. y
C ^ )
(<? + c)
'
° n (_i o) ° r ' (0, i}
sin x + In 1 — sin x\ + c
= -
1
15. y
(1 + sin x)2
on —- 2k 1
7r
^ x ^
< <
2k —+ -
1
ir,
,
k = n
U, ±1,
t ,
.,
±A
_1
17, ^ = (sin x In |csc 2x + cot 2x| c sin x) +
For a given c the solution is valid on any interval which does not include any point
far/2, k = 0, ±1, ±2, or at which the quantity in the parentheses vanishes;
. .
.
,
2 3'2 2'3
19. y = e\\ + c(x + i)- ] , on (-00, 00)
1/3
21. y = 3
* +
_u 1 A 2
~ 6x\ _ 3 /x
2
- 12x\
+,1/3^
72x
(In x) 3
+ c
2 V In x / 2 V On x)2 / 4 V
on (0, l)and (1, 00)
2
23. y = {(1 - x)[c + £ln(-x + V* 2 - 1)] - Vx 2 - l} on (-00, -1); ,
_1 2
y = -{(1 - x)[c + £cos x] + Vl - * }
2 on (-1, 1); ,
2
y = {(x - l)[c + £ln (x + V* 2 - 1)1 - Vx 2 - l} on (1, 00); ,
5
25. y = - 7=
x3 + 5x + cV|jc|
For a given c the solution is valid on any interval on which the denominator
nowhere vanishes; y = is also a solution, on (— <x> 00). ,
3-4 I ANSWERS TO ODD-NUMBERED EXERCISES 733
x2/2
27. (a) v = - - + - / e~ dx; on (-oo, 0) and (0, »)
X X J2
(b) y(l) = -6.3407, y'(l) = 6.9472
1 2x
29. y = 1 + —
31. y = x + —
1 x + ce J c In |x|
( ce ~ 2) 1
sin x
ce — 2
35. (b) (i) y also y = —1
1 — cex
;
3
my 2 \ce 12 * - 1/
'
also y
1
(iii) y = 1 + ; also y = 1
x + c
1 (2ce^ x + 3\ .
37.,' + ,
2
_ 1 =0,0 = 1 + ce2x _ x
> also v = 1 ; y = cie* + c 2e
Section 3-4
1
1. y = -' on x >
x2
3. y = e~
x
- (,- 3/2 + 3,- 9/2 )e- (3/2)x on (-co, co) ,
1
5. y = tfx, on (— t, 0)
x sin x J _i/2 x
= 1 1 + sin x
7. y
sin x
1 + (1 + sin x) In , on (0, 7r)
x — 1
3(x + 1)
1L ' =
9 + \/3ir
sin
.
x H
, v3 — :
7r
cos x > on (0, 00 )
V^L 12 - 4
13. y = (1 + 21n2) - 21n|x|,on(-oo,0)
15. The equation + y = is of the first order on (—00, 00) but is not normal
xy'
on any which contains x = 0. If y(x) is any solution then the equation
interval
implies that y(0) = 0, and so if yo ^ there is no solution which satisfies the initial
condition y(0) = yo.
~ x o\
23. G is the linear operator which transforms h(x) into e Hx) J*^ h(x)e~ kx dx +y e Hx
734 ANSWERS TO ODD-NUMBERED EXERCISES |
CHAP. 3
Section 3-5
y = ci 1 + c 2 e~ x + c^e 2x
• .
yx = 1, y 2 = \ - %e~* + \e
2x
(c) ,
^3 = -\ + &-' + fe 2x
(c) >>i = 2e
x —
2xe x — e^cosx + e*sinx,
y 2 = —2e + 4xe + 2e cosx — S^sinx,
x x a:
x — x —
y3 = e 3xe e x cos x + 3e* sin x,
= x — x
y± xe e sin x
5. (a) Interval is (0, <»).
(b) One representation of y is
ci
x
I
2
'a
+
*' +
1
2
- ) c 2 (3x
x
2
2
).
1
1
(c)yi =-3 + 3x'
y2==
T-3x
Section 3-6
1. = -\2e x
W[l, e~ x , 2e 2x ]
3. ^[l,x, x 2 ...,x n ] = 1!
,
•
2! • 3! •
•;
Section 3-7
M -1 +|ln
1 + x
3. Jce 5. -J--
2
7.
* - x
sin x 1
4-3 I ANSWERS TO ODD-NUMBERED EXERCISES 735
Chapter 4
Section 4-1
3. x2 - lax + (a
2
+ b 2) =
5. (a) (D + l)
2 (D
+ 2)
(b) (Z) - 1)(Z>
2
+ 1)
(c) (D + V5)(D - V5)(Z) 2 + 2D + 5)
(d) (D + 1)(Z) - 1)(Z> + 2)(Z> -2)
(e) (Z>
2
+ V2VT0 - 2 Z> + vlO )(Z> 2 - V2\/I0 - 2 Z) + %/TO )
Section 4-2
\. y — c\e~ 2x + C2e x
3. y = ae
3x/4
+ c 2 e' 5x/2
5. y = ci cos 2x + C2 sin 2a:
7. j> = cie"~
2x cos2x
+ C2e~ 2x sin 2x
9. >> = cie
x
cosx + C2£ x sinx
y = c ie~ cos Vl x + cie~ sin a/3 x
x x
11.
"
y = — §e- sin3jt
2x
21.
y = 3xe^
x
23.
25. y = \/2e
(V2/2)x (cos — x + sin — x)
29. (a) /' + 6/ + 9y = (b) /' - 2? + 5y =
(c) /' + 4/ + Ay = 4 (d) /' + Ay' + 3>> = 3x + 16
(e) /' + 9y = 3x
31. (b) y = ci cos (2x + C2)
33. (a) y = e cos 5a: + ^e x sin 5a:
x
" 50 *
(b) y = (4 - ;rV)e<
1+5i
+ (i + ^iV >* 1
Section 4-3
11. y = ci + c 2 A- + c 3 e-
(1/2)x
cos —
V3
x + c 4 e-
(1/2)x sin —
V^
x
17. (a) (D - l)
3 (b) D2 - 4Z> + 8 (c) (D 2 + l)
3
(d) Z)
2
(D + 2) (e) (D 2 + 4)
3
(f) (D 2 - 2D + 5)
3
(g) (Z)
2
+ l)
2
(h) (D - - 6D + 10) 3
3)
3 (Z> 2
21. (d) The results of (a) and (b) show that the solutions are linearly independent in
C(— oo, oo ).
Section 4-4
{Note. Sometimes a simplified particular solution may be obtained by deleting from the
solution given by variation of parameters, or the use of Green's functions, any terms
which satisfy the homogeneous equation.)
1. y = (ci + In |cos x\) cos x + (c2 + x) sin x
2* — 1 o , O
3. y e 2x + (ci + C2X)e~ Zx
32
v p = x (ilnx - |)
3
11.
cos 2x
13- yp = T? hi|tanx| + (cos2x)ln
16 1_
(1 + cos 2x)|sin 2x\
2{x ' (x - l)
15. >>
p = —x 17. K(x, t) = U.e
l)
~ e- ]
x_< x(x - t)
19. #(*,/) = 2£> sin K* - 21. K(x,t)
(1
+
-
x)(l
x)(l
-
+ /)
2 2 2
25. If a 5^ or if co 9^ b —
- a , then
+r
2c £1 Z>2
where
c = (b
2 - a 2) 1 ' 2 , Di = [a
2
+ (w + c)
2
]
1/2
, £> 2 = [a
2
+ (« - c)
2
]
1 '2
,
If a = and co
2 = b 2
, then
y = —
2 co
o (sin w/ — (at cos wf).
4-8 I ANSWERS TO ODD-NUMBERED EXERCISES 737
Section 4-5
= + + C2 + C3X + c±e~
(t - C1
y e*
9. **
)
~ - - -
11. #(;t, = ex l
(x t) 1
*- J)
13. #(x, /) = - 2e
\\e*- 1 2 (*-»
+ e 3( ]
Section 4-6
l m
y = 2x + /afe-(
x2/2)
dx + k2
3. y = kix + k 2 2
= e — /[ao^/ajCz)]^
^JX) /[« (x)/a jCi)]^
7. )>
I c + k
7 tfi(x)
^olnx
—= 1
Section 4'-7
1. ^p - 23 e x _j_ x 2 _ 2x
_ 3- y P = £ cos x — \ sin x
-2
yP = (x - 3x + ^&)e*
2
5. JP = -|*e >
7.
2 3
9. y P = eh ( 899 + I 305 * + 675 * + 125x )
11. yP = ^_
(2374 - 2940* + 3200x - 1500a: + 625x )
2 3 4
23. yp = en?-* + x 3
[(c 2 + c 3 x + c 4 x 2
)cosx + (c 5 + c 6 x + c 7 x 2
)sinx]
1. y = ci\x\ + c 2 |x|- 2
a 2 = i[l
2U — ai
«1 T
+ V « J» then
V</]; «-"«-•" #(*,
-»vv-v, f)
»y — /- a —
'? 2
y/df
a
x [ln x — In /]
(ii) If
1
a = 0, set a = (1 - a)/2; then #(x, ?)
/a-l
(e) (x
2
D 2 + xD + l)(xZ) + 1)(*Z> - 1)
Section 4-9
11. h = h cosV2g/Lt
Section 4-10
3. (a) q = 0.03e
(c) 3.5676 volts
2 1/2
5. If R > 2(L/C)
1/2
, set A = [R - (4L/C)] ; then
[R/(2L)]t
q = 0.03e- cosh-, + -sinh-f
1/2 [RI{2L)]t
If* = 2(L/C) then q = , 0.03[1 + (R/2L)t]e- .
5-4 I ANSWERS TO ODD-NUMBERED EXERCISES 739
Vlc 1 'LC
Chapter 5
Section 5-1
(c) Not piecewise continuous on [0, °o), since function becomes infinite at t = 1
- 2 = t 1
Note that the function, as given, is not continuous on [0, 00 ), since it is unde-
fined at t = 2.
(e) Not piecewise continuous on [0, 00), since it becomes infinite as t —>
7.1 9. -;
Section 5-2
3. £[/](*) = - (1 - e~
s
), if s * 0; £[/](0) = 1; s = -«
5
5. £[sin atf] = ——
s2
-
+
— - ;s
a2
=
Section 5-4
1
cos a + 5 sin a
52 + 1
r n n— 1 n-2 „ 1
° U
* -•
3. n\
_n!s
I
(n - l)\s 2
I
(n - 2)!j3
"*" I
^I
l!^ n ^
I
j»+ij
2
s 2a
5
-s* - a2 ' s(s 2 + 4a2)
740 ANSWERS TO ODD-NUMBERED EXERCISES CHAP. 5
4s
13. -e
'
(s 2 + 4)2 a
a
17. If |/(0| < Ce \ then lim e~"f(t) = if s > a.
t—>00
21. y = e + 2e* 2<
y = -| + \e- + t qCOs2? +
3
23. %
§sin2f
25. y = -it + fe
2 3 - fe~ 2< '
Section 5-5
72j(j - 9)
1. 3.
(5 -2)2+9 (j2 + 9)4
5.
(5 + 3) cos 4 - 2 sin 4
(^ + 3)2 + 4
2,y2 s2 + 1
2
—n — —2s —3s
— e—
I 4s,
1
13.
-»
s2
[1
yi
Ae + i /i
4e ]
15.
(s - 2)2 + 4
1 l + e~"\ 2(s-- 2)
17. ( 19.
-
j2 + 1 \1 - C-'V S[(S 2)2 + 1]2
(s2 + 1)4 s3
„„ 3s
4
- 16s 3 + 96s - 108 1 e
2a
(2 sin a — 5a sin a — cos a)
23.
- ttt^ »x„ «,„ +
,
- + -
l)2[(s - + l) 2 l) 2
,
25. 1 - e 27. ±
4 -
— ie
4e
-2t
"* —
- ite~
2'
1 n —1 at
31. i-e~
zt
-2t
sin5/
29. t e
(1 - D!
33. / sin t 35. § sin ? + fr cos t
(0 if r ^
37. V3 W2 (r)sin —l
(/ - 2) = i/^ J_
2,
(
, _ 2) if , > 2
39.
1
cosh
,
—-sin ——
t . t
— sinh
. ,
——cos—-
t t
V2 V2 V2 V2 a/2
= 1 if t ^ 1,
41. 1 + «i(0
2 if t > 1
5-7 |
ANSWERS TO ODD-NUMBERED EXERCISES 741
„,
43. —
3a 2
1
e
-at
- — 1
3a 2
<?
at/2
cos
V3
—- at
2
+
4\/3
-^pr- *
9a 2
at/2
sin
.
—
V3 at
2
Ae 1 —2t —3u
45. - ,
(e — e )
\
n —1 times
Section 5-6
1. y = \e - \e~
l
- \te~
l
3. y = %e- cosV2t l
+ ^e^siny/lt + t - f
~ j)
+ I E ("D^Wtl - e-
k(t
]
'- x) - -»
11. y = 2sin2r + fe- (
fe
(
(-D
!7 -'°« = Z4^<
fc=0
2 2 *(A:!) 2
,2*
Section 5-7
-<
1. f/C - £)sinU£ 3. e + / - 1
Jo
6t at
e e
5. f[rcos/ + sin/] 7. ~ \{ a ^ £ ,e
af
if a = £
o — a
;
n O 2 2
_ (cos bt — cos at) ifa ?* b ; ^/sinar ifZ> = ±a
a2 £2
n_1
,, -« _ _
•'+ 1
,„ - Oe-tgWdt ,. /
11. ej f(t 17.
(« - D!
«, t ii 3tt 1 /, 9 \ „ _i s
-tan
T -j,ln^+ ?j-3tan
, ,
21. j 23. j
j
Section 5-8
7. K(t, e-
(< -e -^- e/2 V~3
cos^(r-{) + >/3e<-*'* sin ^(,-0
9.^(^) = ^j,-
V_2( ^ /2
sin^(f- f) + cosYC- *)
+ e
V2(«-f)/2
sin^(f - f)-cos^(/- f)
~ _1 -1)
11. y = hi(0O ' + i[c' - c- ('
]}
_ (0 if< f < 1,
|l _ t+ i[ t-i
e _ e
-(«-i)] if/ > 1
19. y = f^- + 2)
^e-(«-2) + ^-t + ^e4 '- 10 - Ate~ {
21. y = e'
/2
[^(/ - a) sin 3/ - -^ sin 3a sin 3(/ - a)]
Section 5-9
. t , i -t ,
2\/3 t/2 . V3,
3 2
+ ^e
'- 1)/2
i„-«-» _ i^-^cos^C/- l)
(
sin^O - 1)
V + y-* + ^e^sin^^ if ^ / ^ 1,
= <
fc' + fcr
l
+
2\/3 «/2
e' sm
•
—\/3
t + i
$e
-0-D -\e
tjt-D/2
cos — - 1)
f^sin2/ if ^ t ^ tt/2,
lm
— a/3 r- fk
+ fk
.
L
7. y = 2t -^ sin V 3
.
t 9. y = a cos a /— /
,
b A /-r- sin A /
(b) r = i In 3 = 0.27465
(c) \/3/27 = 0.064150 above position of equilibrium
6-3 ANSWERS TO ODD-NUMBERED EXERCISES 743
(d) Mass starts at time t — at the point 1 unit below the position of equilibrium
with velocity 10, and moves upward. It passes through the position of equi-
librium att = 5 In 3, and at / = ^ In 3 it reaches its maximum height, x/3/27,
above the position of equilibrium, after which it descends asymptotically
toward the position of equilibrium.
15. y = e-<"[a + {b + <ra)t]
Chapter 6
Section 6-1
5. (b) No; if /(x) = on (0, oo) except possibly on a finite closed interval, then the
equation cannot have infinitely many solutions on (0, «> ).
Section 6-2
2
/ 1
1 — An
P \
7. (a) u" + (l + 4 J ) " = 0, [— °o 0) and
on (- , (0, <x>
(b) «" +
!+/?(/> + Dd ~ * )
« = 0, on (-oo,-l), (-1,1) or (1,00)
(1 - x2 ) 2
2 2
x + 2 + 4/(1 - x )
=
(c) u" + 4(1 - X2 )2 u 0, on (-oo,-l), (-1, l)or (l,oo)
Section 6-3
= on (— oo, oo)
^i'- 0,
0, on(-oo,-2 1/3 ),
(-2,1/3 ,2-1/3 )or(2
1/O 1/O 1/O
1/3
,oo)
1/2
3/2 (-In x) sin x
(d) (-lnx) ^ + ,
^ y, on (0, 1),
dx dx 2x
d_
n \l/2
(In x) sin
•
x
3/2 ffr
(lnx) >>, on (l,oo)
dx dx 2x
d_ 1 dy 2x
(e) + y = 0, on(— oo, — l)or (— 1, oo
dx x -+- 1 dx (x + D2
,2
2t 2
3. ?£+ (e - p )y = 0, on (-co,0) or (0,oo)
at*
5. (a) Every solution has infinitely many zeros on every interval of the form (— oo — a) ,
or (a, oo ), where a > 1, and does not oscillate on (—1, 0) or (0, 1).
(b) Every solution has infinitely many zeros on every interval of the form (— oo a) ,
(c) Every solution has infinitely many zeros on every interval of the form (a, <x>
),
7. (a) Every solution has infinitely many zeros on every interval of the form (a, oo ),
Section 6-5
= f-(-D*
2k
1-3' = aoyo + divr, vo Is oft)! x '
fc=0
00
(— 1) 2fc+l x
x , on ,(—oo, oo
k=0
(2k + 1)!
3*
= owo + = +
El
3. y ao'UJ'o 1
J^tti
^ ft![2 •
5 •
8 • • •
(3k - 1)]
00 1
3&+1 , >.
(-D
5. >> = tfi(x + x ) + 3a ^
A;=0
(2Jfc - 3)(2k - 1)
a:
2/fc
, on (—1, 1)
7. y = ai(x + §x'
- 5)!(2fc+ 1)
+ ao
L
1 + 9x
2
+ ^x4 + 3 ^
A=3
(-1)
fc (2fc
2*-3[A:!(ft - 3)!]
*
2fc
on
(-^
9. y = a yo + ai^i, where y = ^ (-D *
fc=0
"TuT"
£!3*
3k
3i
'
(-I)'
3
1
!
-St ft=0
(1 + 3 •
1)(1 + 3 •
2) • • •
(1 + 3*)
x , on (— oo , oo
1 6-1 2*
+ flo 1 — - Y^ riV
Kg}
'
; + a\x, on (—oo, oo
6^[k\(2k - \)
10 19 - -
£ 1 (9A: 8)
• • •
3*
>>o = 1 + (3A:)!
* '
yi -x + (3* + 1)!
y2 ^ 2
+ 2E- *=i
7 •
16 •
25
(3*
•
+
•
(9*
2)!
- 2)
— x 3k+2 , on (—00,
.
00),
)
6-5 I
ANSWERS TO ODD-NUMBERED EXERCISES 745
y\ = x +
2
*=i
[(!)(-!) - 4][(3)(1) - 4][(5)(3)
(2k +
-
1)!
4] • • •
[(2k - \)(2k - 3) - 4] 2 *+i
on (-1,1)
24(*-i)A:!(ik - 4)!
= 1 _^8* 2 4_1Z^
+ 128*
4 _ J^L * 6 _ cf ^- 7)!(2fc + D(2k-
yo
= x - x 3 5
1024
^ 24<*-»fc!(ife - 4)!
yi + £x , on (-2, 2)
m
19.
V^ ( i^+l
= ao2^(-\)
A: + 1
x
3fc
+.
„ «„
aix, on
/ -,1/3
(-2 ,2
l/3,
>;
k _ )
fc=0
+-^ X"^
+ V^ +
4 2fr+l
= -x 2
1 1 3 1
21. y
1
+-x
"
,
/4
+2^ . .
fl 2fcX
2fc ,
2^ »2fc+ix ,
,
where
fc=3 fc=2
2 3
a2 * = t^T [1 - 2(* - 2) + 2 (A: - 2)(* - 3) - 2 (fc - 2)(k - 3)(k - 4)
(2A:)!
2
+ • • •
+ (-2f- (k - 2)!],
on (—oo, oo
23. There exists a solution which is a polynomial of degree zero (a nonzero constant)
if and only if 7 = 0; there exists a polynomial of degree 1 which is a solution if
and only if /3 +
7 = 0; there exists a solution which is a polynomial of degree
n > 2 if and only if n(n - 1) fin 7 = and either a = or k(k - 1) + + +
file + 7 5^ for all nonnegative integers £ which are less than n and differ from
n by an even integer.
f. . *
fc
(2kTT)l
X '
k =l
on (—oo 5
oo )
746 ANSWERS TO ODD-NUMBERED EXERCISES |
CHAP. 7
Section 6-6
1. y = - i* 3 + j±oX 5 + j^* 6 +
1 • • • , on (-oo, oo)
3. y = — 1 — x + \x 2 — \x z + • • • , on (— 1, 1)
5. y = 1 + x + x 2 + jix 3 + on (-*, J) • • • ,
7. j> = 6x + i*
3
+ f* + 27)* +
4 5
on (-3, 3) • • • ,
13. x - \x z + i^x
5 - ^x 7 + • • • , on (-00, 00)
Chapter 7
Section 7-1
5. (b)
Section 7-2
3. (a)
—
V3
(b) VT^l (c)
2V30
-^- (d) —
V2
(e) V2(l - In 2)
(b) Orthogonal, that is, the cosine of the angle between any two distinct vectors
(functions) is zero.
Section 7-3
7-5 |
ANSWERS TO ODD-NUMBERED EXERCISES 747
Section 7-4
5 '
w
(a) e*
'
e~* - -=
e2
2
- 1
ex (b) e
1
, e 2x
2(e 2
—++ +—
3(e
e
1)
1)
- ex
, ,
11. (a) at
/ x
=
x—
— • e,
-
>
.
i = .,
„ „
1, 2, 3, . .
. , n
,,.
(b) x y = v^
}^
•
— (x • e*)(y
(tk
_,
• e*)
13. ±-^(4,-1,2)
Section 7-5
l. (a)
—
V2
(b) 2 (c)
C1
f± (d)
30V29
-^~
3. (a) l (b)
—
\/43
(c) &Vm5 ,
(d) 3
Q
y (11 4 19.\
- v 9 > 9> 9 /
io
ly -
fl
» -
<-nv
(n
2
+
-
1),
o ' bn ~ (-lr+w
(„2 +
-
l )7r
o
*
+
•
1 +
2n
^ ^"
+ 3)
2(/i + 2)
]
x + 4(/i + ^
^|+ +t}*
3
1)(« 3)
- i);
1/2
lldll
= [1 + (-1)"]
v
- [1 - (-1)"]
12/i + 1 (n + 1)2(« + 3)2 L ~ '
J
"' (« + 2)2
i 2
- = V6
(c) Projection, •
1 + if(x £); ||d||
24
(d) Projection, 1 + 3(x - i); ||d|| =
23. If 9C = V, then SC^ contains only 0; if 9C contains only 0, then XL = V.
27. (a) ay + /3z, where y and z are linearly independent and orthogonal to (1,1,0)
and a, j3 are arbitrary reals; for example:
a(l, -1,0) + 0(0,0,1)
(b) a(l, 1, 0), a an arbitrary real
(c) ay + jSz, where y and z are linearly independent and orthogonal to (1, 2, —1)
and a, are arbitrary reals; for example:
a(l, 0, 1) + 0(-2, 1, 0)
Section 7-6
3. (a) A: 16.06, variance 0.00236; B: 16.12, variance 0.00092; fi's results are the more
consistent,
(b) 16.09, variance 0.00254, greater than for either A's or Z?'s data alone
5. (2.25, 2)
7. (a) (-0.8, 0.4, 1.7) (b) (3.4, -1.3, 3.6) (c) (1.9, 0.4, 0.8)
13. (a) 54* - 35v = (b) 417* - 142y = (c) 26* + lly =
15. y = fxi + %X2 17. y = x\ — x2
19. y = %xi — \x2 — |*3 - (a) y = -£x 2 + ffx
(b) y =
= 9
2
-
2 -
M A
(c) ^ -WfX
37' ff;
"• v = _3y4
23 y 8*
i 15„2
* 1_
' 8 10
Chapter 8
Section 8-1
Section 8-3
ck = ;
2
—— ie - e~
T
), k = 1, 2, 3,
(i + k W^r
With respect to the second set
k
ck = (-l)
fc+1
-(/ - O, k = 1,2,3,
(1 + k^W^
Section 8-4
bk i/'
= - J fix) sin A:x t/xr, A: = 1, 2, 3,
Section 8-5
3. (b) / 2
13. (b) The vector y = iyi,y2,)>3, . . .) is in S(9C) if and only if for some integer
k > 2, yk+ i = yk+2 = yk +3 = • • = and yi = y 2 + ^3 + + yk • • • -
Chapter 9
Section 9-2
7. Yes; a linear combination would vanish identically only if all the coefficients were
zero.
Section 9-3
1. (a) Odd (b) Even (c) Neither (d) Even (e) Odd
(f) Odd (g) Neither (h) Even (i) Neither (j) Even
750 ANSWERS TO ODD-NUMBERED EXERCISES | CHAP. 9
2 j
,
2x
7. (a) fE (x) =^^4>
x - 2 l
f (x) =
X^ - 1
1
(b) fE (x) = 1 _ x2
'
•—
Mx) =
1 _ X2
(c) fE (x) = — cos 2x, fo(x) = X COS X
Section 9-4
fc+i
(-D kx
1-22] sin
k=X
3.
(-D* (cos Arx — k sin £x)
+ £2
1 1 fl « „+i2 1 n 1 n
f. . .
ifc +1
7.§.
2
+ 42 (-1)£2 cos Ax
00 j
9.-2 jP -sin&x
5t\ \
— and — ,3t the
/5tt
11. (a) On 2x, 1 I ) series converges to —1 and 1, respectively;
atx = 5x/2 and x = 3x the series converges to zero. On [— 2x, — 3x/2) the
converges to 1, on (— 3x/2, — x/2) to —1, and on (— 7r/2, 0] to 1. At
series
x = — 3x/2 and x = — x/2 the series converges to zero,
(b) When x = kir, the series converges to ±1, according as k is even or odd;
when x = (2k + l)x/2, the series converges to zero.
13. (b) When x = k-w, the series converges to x/3; when x = (4k l)x/3 the series +
converges to x/6 or x/3, according as k is or is not an integral multiple of 3.
k+l
-1) " 1 , , (-D—
15. (a) cos kx H sin kx
k 2 ir )
k=\ x
"1 +1
(-i)*
17. (a) cos ax = -
a
+ ,
, v-
2a >
*—( k 2
— — a 2
cos kx
. 1
J
9-5 | ANSWERS TO ODD-NUMBERED EXERCISES 751
sin fcx
(c) T + 2 JP
25. (a) i — ^ cos 2* (b)isin2x
(c) f sin jc — i sin 3* (d) sin x + 5 sin 2x
\
(e) \ cos x + \ cos 5.x (0 f + i cos 2x + •§ cos 4x
00
00
00 00
(b) fE (x) ~ y+ J]
*=i
a* cos £x, /o(x) ~ ^
fc=i
Z> fc sin kx.
(c) 1, cos x, cos 2x, . . . and sin x, sin 2x, . . . are bases for S[— x, it] and 0[— x, x],
respectively.
Section 9-5
L-JC,0<JC<T x
e < x < ir
,
7. 2 ^- sin Arx
9.4^;
i
i~\f
+l ~ sin (2k — 1 )x
^ (2* - l) 2 ir(2k - 1).
/2
.0 if A: = 0,2,4,6,...
752 ANSWERS TO ODD-NUMBERED EXERCISES | CHAP. 9
Section 9-6
3. fix) = \ + 2 Ak cos
kirx
—r- » where
ifk = 1,5,9,
kir k 2 ir 2
ifk = 2,6,10,...,
Ak = < k 2 it 2
ifk = 3,7,11,.
kir k 2 ir 2
ifk = 4,8,12,16,...
7-i + s — -
(-D* 1
cos kirx
.
+ ,
— sm
1 - .
kirx , 8 < x < 10
k'lr* kir
A=l L
2 i 1
9. /^ -
*— sin At7ta:, 1 < x < 2
*£l*
ir A:'
k+1
-l) 3ir . (2k - l)irx „
11.
- sm 2 < x < 3
(2k 1)2 (2A: - Dv
,
Section 9-8
3. If F(— x, >>) = F(x, >-) and F(x, —y) = F(x, >0, then
00
5. 2 /]
*— sin mx
ro>=l
' m
^•2 + 2 cos 2* cos 2> — ^ sin 2x sin 2y
2 oo
°o /
/- <
i \»™+l
\»»+l oo /
f 1 \»H
A -jm-\-q-\-l
2x
9. sin mx cos #y
wi,g=l
11-4 | ANSWERS TO ODD-NUMBERED EXERCISES 753
Chapter 10
Section 10-3
00
1. A = 2
J] bk /k\ A k = -b k /k, k * 0; Bk = a k /k
k=\
Section 10-6
3. No; the formula x/2 = £"=i (-1)* +1 (sin fc*A) is invalid outside (--tt, tt).
Section 10-7
3. (a) (b)
Chapter 11
Section 11-2
Section 11-4
/(*) = S
k=0
a 2kP2k(x) (mean),
where a 2 * = (4k +
l)f£ f(x)P 2k (x) dx.
If/(x) is an odd function in (PC[-1, 1], then
00
Section 11-5
1. n
1/2
g(x)Pn (x) dx (2 \
1/2
/2 -I- A 1/2
f
1
Section 11-6
Section 11-7
( 2) 2
9. (b) Lo = 1, L? = x - 8x + 12,
4 2)
= x - 3, L3
2)
= x 3
- 15x
2
+ 60* - 60
Chapter 12
Section 12-1
9. X n = 2 « = 0, 1, 2,
/7 , ; Jn ci cos nx + C2 sin nx
Section 12-2
Section 12-3
X = 5 + V5
, "°
with eigenvectors
(
of the form X2 ei + e2 +
- (
—
y/5
-e 3 j ;
2 '
"V* " '
2
5 - V5 / 1 +~ V5
X = , with eigenvectors of the form X2 I ei + e2 —
)
5. (a) x = -|ei + + |e3 e2 (b) x = ^ei + |e 3
(c) x = -fei - 2e 2 + fe 3 -x
Section 12-4
[Afote. If 6 = 0, one of these vectors is the zero vector; in this case the required
basis is {ei, e2}.]
3. X = n2 n ,
= 0, 1, 2, . .
5. (b) The null space is the set of all functions F{k) in $7 of the form
F{k) = cik + c2 , k = 0, ±1, ±2,
where c\ and C2 are arbitrary real constants.
Section 12-5
3. X = n2 n ,
= 0, 1, 2, 3, . .
.
, with eigenfunctions of the form
y = c cos nx (c t* 0),
=
(2/i + l)
2
=
X , « 0, 1,2,3,... with eigenfunctions of the form
{In + \)x
y = c sin (c ^ 0)
Section 12-7
1. * = 128
~ Z2*— (2«—+TT^
7T
n=U
'
7^
l) 5
sin o
2
— x 3 - No solution; Xo = 0, c = — -
7T
^
fc+1 fc+i
(-l) 4 1
— 1 + (-D
sin {2k )x + sin 2&x
5. y
_ _(2* - 1)% (2/t - 1)3
1
8A:3
2*
7. If Jo' «(*) <** ^ 0, there is no solution; if f h{x) dx - 0, then
00
where
2* 2a-
1 1
an = — I h{x) cos nx f/x, bn = — I
h{x) sin nx i/x,
7T Jo 7T Jo
and c is arbitrary.
756 ANSWERS TO ODD-NUMBERED EXERCISES | CHAP. 1
Section 12-8
5. X„ = „2^2
« 2 2 '-a »
x' 2
sin —
rnrx
^
, « = 2, 4, 6,
7r , <?„(*)
/0
nwx
?
x/z
cos -r- n = , 1, 3, 5,
2
orthogonal in Q[— 1, 1] with respect to the weight function e~ x
7. Xn = —n 2
,(p n (x) = e~ (ncosnx x
+ sinnx), n = 1,2, 3,. . . ,andX = l,f(x) = 1;
orthogonal in 6[0, 7r] with respect to the weight function e 2x
9. Here the set S of eigenvalues is not a sequence but a region of the complex plane,
comprising those points X = £ + /17 such that — 1 < £ and |i;| < 2(£ -f- 1) 1/2 ,
with the addition of one boundary point, X = —1. To obtain the eigenfunction
<p(h; ^corresponding to a given X ^ of S, set X = pe ie (where p = [£ 2 + n 2 ] 1/2 ,
Section 12-10
1
- (cot A: sin fo — cos kt) sin Arx, x < t,
[Note. The constant k may be chosen as any real number except that k must
not be an integral multiple of t, since then L = D 2 -\- k 2 is not one-to-one
when restricted to S.]
x
1 r
(b) y(x) = - {(cot k sin kx — cos kx) / hO) sin kt dt
k Jo
+ (sin kx) J
A(0(cot k sin kt — cos £r) dt}
JX
Chapter 13
Section 13-2
(b) a (*, y)Dxx + 6(x, j;)/)^ + c(x, y)Dyy + </(*, y)£>x + e(x, y)Dy + /(*, y)
Section 13-4
.
1.
.
u(x, t)
.
= cos—
«7raf
— sin—=- . n7rx
— »
^-Z> + (-D^'lcos — sm —
at 2 J^ i _ , i mrat . rmx
3. u(x,,) -
n=l
^
5. «(*,/) = 2 Ucos— + —
n=l N
&/v^
L
/ J
1
—
.
mrat
nir
,
wraf
ft. sin
.
.
nirx
^
/ittaA
/JTTflM
sin
. AI7TX
7. "(*,0 = ^i;-^n
v 2Z-j n T2 cos--sin
L2 L
n=l
. rmx
sin-yr-
n=l
When / = L/(2a) or t = 3L/(2a), the string is in the position of equilibrium
the reflection in the x-axis of
(y = 0); when f = L/a, it is in a position
which is
its initial position; when t = 2L/a, its position is identical with its initial position.
9. (a) u(x,y) =
2L v^ ~
-^ £ rm-~ n=l
1
2
sin
.
sin
.
—
rmb .
sin —
mrat
sin
.
—
rmi
= 2^
^v /
V
Cn C0S
mrat
~L~
+
_ .
Sm
rmat\
mrat\
~W S "
sin
.
.
—
rmx
n=l ^
where £ .z,
L Jo
m
Cn = - I f(x)sm ^dx, D n = rma
L
— Jo
g(x)sin
m^dx
L
n sin
mrat\
—J cos —
mrx
where
A> = 7 ( /Wcos^Jx,
L
» = 0,1,2,...
L Jo
r
= 7— sm ^x
*2»-i 7;
— / *(*) 2Z^
LL
(2n l)ira Jo
Section 13-6
, x
4L^-% 1
*
-u2m+i)2**tiL*a*] .
(2m + 1)ttx
1. m(x, = -jL
^2^
-i=U
(2m + 1)2
' J —
3. «Cx.O = 4 t TT^TIk
8
+
ir 3
e«»*W>
*" (2m'
l)3
sin
(2m
+
L
1)rx
5"(*.<> = -|l-2X;z^rT
=
e cos—
1
T *\
/-
i. u(x, t) = —
Aq
+ X^
2^ A
I A
ne
-(n 2 r 2 tJL 2 a 2 )
cos_
rVKX
zr
ra=l
where
sin -—
«7T*
(b) w(x, t) = -
L, IT •*~^
n—L
n L L
13. w(x,
lOOx
= — \-
2
- > ^ & + (-l)"
+1
(&i - 100)
e
_ (n 2,2 (/L 2 a2)
sin
. mrx
—
L 7r
-'— j
' n L
Section 13-7
2 \ oo ^-[(2fe + l) 2 T 2 t)/(a 2 Af 2
3. „(*, ,, ,) - ^-j- e
f o „ o
cos
T) g ^ + )y
)
sm ^—
00
ui = ^
7»=1
F„,„ sinh \/m 2 + « 2 z sin mx sin «y,
71 =1
where
f ffl » =
7r 2 sinhxv/m 2
-==
+ «2
J
-^o 7o
I gi(x, y) sin mx sin nydxdy.
The formulas for « 2 , «3 are obtained from that for u\ by permuting cyclically
the subscripts 1, 2, 3 and the variables x, y, z. Similarly the formulas for u 2 , v-s
are obtained from that for v\. The solution of the given boundary-value prob-
lem is U = «l + #2 + «3 + l?l + V2 + U3.
—
7. u(x, y,t) = 22 (Emn cos a\J m 2 + n2 t + Fmn sin a\/m 2 + n2 t) sin mx sin ny,
m=l
n=l
where
F = —
** J
/(x, >>) sin mx sin ny dx dy,
./
F — /
g(x, y) sin mx sin ny rfx </y
/
ir 2 a\/m 2 + n 2 Jo Jo
Section 13-9
"7T* 2 WTX
/
m(x,
,\
f) = V* D -(n 2 T 2 *)/(L 2 a 2 sin——»
2_j Bn e
) •
where 2?„ = - /
/ /(x) sin
.
-— efr.
Chapter 14
Section 14-2
where
161/
t-m —
(2m + 1)3^3 si n h [(2m + 1)ttM/L]
nir(M — v) rnry
= Zc„ ,
. , . .
sinh
, .
h sinh
.
X sin
(2m + 1)ttx
where
sr
(2m + 1)3^3 sinh [(2m + l}irM/L]
7. w(x, y) an
=l sinhnw
n
where
) cos nx dx
760 ANSWERS TO ODD-NUMBERED EXERCISES | CHAP. 14
9. u(x, = >
1
cosh
,
(2m += l)iry
smh
.
,
(2m +
=
1)^
y)
X sin
(2m + 1)tx
Section 14-3
n
^2 r f(d) cos n(s - 6),
Theorem 1-22 to the sequence of partial sums, and combine the results for the
subintervals.
Section 14-4
00
A n = ~^~
2
— /
Jo
f(<p)Pn (cos <p) sin <p d<p
9. (a)u(r,<p) = J + r 2 (cos 2 ^ - J)
(b) u(r,<p) = -i - fr 2 + 2r 2 cos 2 ^
(c) u(r, <p) = i + f r 2 - 2r 2 cos 2 p
(d) u(r,<p) = i%- + ^-r 2 (l - 3 cos 2 <p) + ^(3 - 30 cos 2 <p + 35 cos 4 <p)
2fc+1
11. w(r, 9?) = 2^ fi fc r P2 fc+i(cos^),
fc=0
where
/•t/2
Bk = (4k + 3)/ f((p)P 2 k+i(cos(p)sm<pd<p
Jo
Section 14-5
_i + vr+s>/2 (-i-vrRx)/2
cir( + C2r if x > _ i;
_1/2
1. (a) R = < /•
(ci + C2lnr) ifX = —5;
j~ 1/2
(a cos [-h/-(l + 4X)] In r + c 2 sin [£\/-(l + 4X)] In r)
ifX < -i
3. «oo = 1 «io = cos <p «ii = cos 6 sin <p
un = sin sin <p U20 = f cos 2 <p \ — W21 = 3 cos 6 sin <p cos <p
"22 = 3 cos 26 sin 2 45 U21 = 3 sin sin <p cos <p U22 = 3 sin 2d sin 2 <p
Section 14-6
2 „ .1 .1 «22 1 °°
4A: + 1
1
3. COS
2
(p COS ~ 1
- «00 + -Z "20 +^ - «2fc,2
O 3 24
4f^2 k(k+ 1)(4*2 1)
14-7 I ANSWERS TO ODD-NUMBERED EXERCISES 761
5. (a) fbp, 0) ~
1
- woo ^
+ 2^ 22*+i(4A:2
72k+iuk2 _
(-D*(4A: +
ix(fc +
nrrt-u
1)
1)!]2
fc=l
{lk 2)! 2
X t M 2fc ,o - (2k - 2)\(2k + k + 2)u 2k 2,
+
2k
l)i]2
(2k 2)! 2
X t u 2k ,
- (2k - 2)l(2k + k + 2) M2fc ,2
»+ 1 r - (w+1)
7. (a) ii(r, <p, 0) = ]T
i=0
[r
ro
- a
2
] ^ m0"mO + £ v4 mn « mn + Bmnv n
where
,-TO +l
Slmn — Simnx!*, u s/ £,2m+l _ ^mn(g),
a 2"»+l
i
,m+l
"mn = Bmn (a, b; 5 '
_ ^mn(g),
£2»»+l a 2m+l
with
2m + 1 (m — n)l
A mn (g)
2tt (m + n)!
/•2t
rZ-K /i
rT
n+
X / / *fo 0) cos /i0 sin <p/C (cos (p) dip dd,
./o ^0
2m + 1 (m - n)!
Bm n(g) =
27r (m -f- «)•
*2t /•*
n+
X g((f, 6) sin nd sin <pP„T (cos <p) d<p dd
o Jo
(b) u(r,<p,6) = a
2_j \2 rno(r)UmO + 22 [«mn(r)««» + jS mn (r)i?mn ]}
'
m=0 * n=l
where
2m
a^nO) = 4n(i>, # J
/)[r" — /, +V-(m+1) + ^ m ] „(fl, b\ g)
m - 2m+1 {m+1)
X [r a r- ],
Section 14-7
3. Qo - 202
5. r 3 U3o = — §x 2 z — \y 2 z + z 3 3
r V3\ = —%x 2y — fy 3 + 6yz^
r «3i = — f* 3 — fx>> + 6xz = 30x>>z
3 2 2 3
r vs2
r 3 «32 = 15(x 2 z — y 2 z) r 3 £?33 = 15(3* 2.y — >> 3 )
r 3 K33 = 15(* 3 — 3xy )
2
= — 1) (« - D!
C„_2i-i,2y+i
- - Cn — i,i
(
(2y + 1)!(« 2y 1)!
762 ANSWERS TO ODD-NUMBERED EXERCISES I CHAP. 15
11. x - 6x z + z
4 2 2 4
x 3 y - 3xyz 2
x 2y 2 — x 2 z 2 — y,2,2
2 2
z 1,4
j_ ^z-
+ ^3 _ 3xyz 2
^
4 _ 6>,2 Z
2
+ 24 X 3 Z7 _
2 VZ — ivr 3 xy 2z zr — j^xz'
3
y z — yz°
13. (a) r
2
u 20 = -K* 2 - z 2) ~ Uy 2 ~ z 2) (b) x2 - z2 = -r 2 « 20 + ir 2 « 22
r 2 M9i
U2i = 3fxz)
3(xz) ^
xv = 2
Z
^r V22
\r o<>9
r 2 u 22 = 3(x 2 - z 2) - 3(y
2 - z 2) y
2 — z2 = —r 2U20 — \r 2 U22
2
r U2i = 3(yz) xz = §r 2 U2i
2
r t;22 = 6(x>>) yz = §r 2 v 2 i
m + * ~ 2 + A: - 3
17. dim 3C m , fc
= +
f
\ k - 2 k -1 /
Chapter T5
Section 15-1
3. Set x = Xr
Section 15-2
Section 15-3
3. (a) v
2 - 1 = (b) ?
2 - 3? - 1 =
(c) v
2 - 3v + 1 = (d) v
2 - 2v - X2 =
(e) v
2 =
2 —
5. v 0, in each case.
(-0*
7. yx = |x|
1/3
(l + £*), J2
1-1/3
E
fc=0
A:!3*(3A: - 5)(3/t - 2)
9.^ = ixi
i/2
^ 2 = ixi
i/4
2:
fc=0
fc!2
2fc
(4A: - 1)
*X(X - 1)---(X A: 4- 1) *
11. y = ao i + E(-d
fc=i
(fe!)2
x
_
-
*
X .
(X ife)!(fc!)2
fc=0
15-4 I ANSWERS TO ODD-NUMBERED EXERCISES 763
Section 15-4
fc-i
-
E
00
2y 15
= \x\
1/2
^2a kx k
where ao = 1, a* = 4*-'- 1 «i>
1. yi
fc=0
,
/c(2A: + 1)
j^Q "
V2
00
^2a k xk , where ao = 1, ak =
J
_
^
z2~^T
/ — 8
a i> k ^ 1; \x\ < 4
k(2k l^
fc=0
00
k— i
(_1) "
3. yi = |x|
5/4
^
*=o
a kx
k
, where ao = 1, ak
4*+*A:(4A:
2(-l) 4 (4y +
+ 5) ^=0
y y
7) fl y,
\x\
,
< 4
fc— i
2
5 ^l = * y^ akx
k
where ao = 1> ak = — X>.*s ;w< i 1 ;
-
k=0
J ,
3t(t + 1)
i-o
)t-i
V2 = \x\
2/3
J]
fc=0
flfcX*, where a = 1, a* = -
3^(3^ - 5) ^^
i=0
7 " 5 ** y '
7. 71 = x ^2 a ^ fc
» where ao = 1, ak = .... ,
3)
/] (— l)'(4y + 3)ay,
fc=0
*^1;M<1;
V2 = Ixl
11 *
f^a k x\ where ao = 1, ak = Jtlil^
- (-!)>* A: ^ 1;
A:(4A: 3)
fc=0 ;=o
< |jc| < 1
9. yi = x + 2i ' \X < 00
+
\
• 10 13
^Jfc![7- • - •
(3A: 4) J
00
1—1/3
V2 = \X » < (jc| < °o
^A:![(-l)-2-5---(3A: - 4)]
fc
vi = (1 - fx + ^x - tVo* +
2 3 •) cos (In [*|) • •
3
+ (£* - ^ro*
2
+ nhi* + • • •) sin ( ln W)>
v2 = (-£* + ^x - 2
y^o*
3
+ cos (ln \x\)
• • •)
3
+ d-f* + TO"* 2 " Tiro* + * * •> sin < ln l*D-
r ~ -4 + V2i
13. vi = V2/, ao = 1 giveai = 0, a2 = ^ ' «3 = 0;
4 -
= - r^V2
/- ~ /
= —V 2 = = =
•
a 1 give a\ 0, a2 ' «3 U.
j/
2 /,
.
y/2 2
y\ = (1 - ix
2
+ • • •) cos (\/21n|x|) + (- ^x*
12
+ •J
sin(\/21n |*|),
2 2
- +
(^y^ +
/
y2 = •••)cos(v 21n|x|)+ (1 ±x • • •) sin (a/2 In |x|).
yi = (1 - fx + ^x + sffo* +
2 3
-)cos (In \x\) • •
2
+ (£* + fo* ~ imo^ 3 + •
'
sin 11
W).
v2 = (-^x - ^x + t^oX +
2 3 • • •) cos (In |x|)
+ - (i
fx + im*
2
+ sfk* 3 + • • sin d11 W)-
This solution is valid for < |jc| < 00
where
a* =
^^E y=o
(-l)'" 2 ^+ *> - #** k = h
(JVote. It can be shown that, with a = 1, the formulas above imply that
X (x - 1)"
Section 15-5
(— 1) fc+i
1. v.i = x, y2 = 2^-jTTZ x + x In x ,
^-^[i + Ec-d*^*
fc+i Ly-iVy^* -i
= 2*- X)(-l) lnx
1
y2
_|_ x
(A!) 2
yi = 4!
A + 1 jfc+4
"
y2 = 1 + §x + 1
|*
2
5.
J] (A: + 4)! ,
3k+ 2
7. yi = x2 + £
fc=i
(-1)*
5 •
8 • • •
1
(3A + 2)
x , y2 -ti-^m 3/fc
2 S
9. yi = 1 -3x + ix - ix ,
4)! (A- - * 2 3
+ Hx - fiE^OkoT ^ +
3
y2 = Ix - \*x 2 1
(1 - 3x + fx - |x )ln*
15-11 I
ANSWERS TO ODD-NUMBERED EXERCISES 765
11. If 2/7 is not an integer, Case I; if p = 0, Case II; if 2p is an integer other than zero,
Case III.
Section 15-6
Section 15-7
Section 15-9
Section 15-10
^+i(M*)
5. If /> > 0, x* = 2
£
*=l [Mfc
2
— /> ][JP (j*k)]
2
/p(MfcX)> o < x < 1,
where /x*is the fcth positive zero of /£(*). If p = then x p = 1 and the Dini series
contains a single nonvanishing term, 2co = 1.
Section 15-11
00
f(r)M* k r)rdr,
^"[/ift^unhX*,'
2
g(r)J (\ k r)r dr
[/i(X*)P sinh X*.
ti* Jo
00
Section 15-12
1. x = XQ
n(l)/o(l + - J'o(X)Yo(\ + tj
/o(i)na) -Mi)Yo(i)
00 00
3. u(r, d, z) = ]jP
n=0
2
fc=l
l Ank cos nd ~^~ Ank sin /I ^»( A ~* r) sinh \ nk (a — z),
2
•^rafc — f(r, d)J„(X nk r)r cos nd dr dd, n > 0,
7r[Jre+ i(X n fc)] 2 sinh \ nk a J-* Jo
2
^nfc — / f(r,d)Jn (X nhk r)r sin nd dr dd
7r[/„ + i(X nfc )] 2 sinh \ nka Jo
00 00
5. «(r, M) = JJA
i=l fc=l
2ntkJ2n(^2n,kr) cos (X2n,fc a/) sin 2n0,
s*2n,k = r , 7T
—— / / f(r, d)J2 n(\2n, k r)r sin 2nd drdd,
7nV2n + HA2 n,k)\ z J ./O
where \ nr i a k is
,
the &th positive zero of /n w«(X) and
^4 n fc — 2
f(r, d)Jnr/a (\nir/ a ,kr)r sin (nird/a) dr dd,
tt[/(a+«T)/«(Xnir/a,fc)3 Jo Jo
4
#nfc = g(.r, d)JnT/a (\n*j a ,kr)r sin (nw6/a) dr dd,
aoi\ nr / atk [J( a + nr ) a (\nT/a,k)] 2 JO Jo
Convergence E
absolute, 648
conditional, 650 Eigenfunction, 460, 474
in a Euclidean space, 304 Eigenvalue, 460, 461
of integrals, 179, 671, 672, 673, 674 method, 466, 484
mean, 307, 314 Eigenvector, 460, 461
in (R", 306 Endpoint conditions, 457
pointwise, 306, 338, 656 Equivalence
of a sequence of real numbers, 305, 638 class, 38
of a series, 242, 311, 314, 642, 661 relation, 37
uniform, 658, 661, 672, 674 Euclidean space, 256
Continuity, 652, 653 Euler
uniform, 652, 653 equation, 161, 586
Convolution formula, 206 formula, 129
Coordinate axes, 26 Euler-Fourier coefficients, 338
Coordinates of a vector, 26 Even extension, 350
Coordinate system, 26 Exchange principle, 18
Cramer's rule, 695 Existence
Critical damping, 227 problem, 83
Curl of a vector field, 718 theorems, 104, 243
idempotent, 61 N
identity, 43
Nilpotent, linear transformation, 51
image of, 42
matrix, 79
invertible, 57
Node, 519
matrix of, 64 equations, 294
Normal
minimum polynomial of, 72
Normal vector to a surface, 701, 708
nilpotent, 51
null space of, 55
polynomials in, 51 O
powers of, 50
products of, 48
Odd extension, 351
Operator equation, 81
one-to-one, 56
Orthogonal, 268
onto, 42
Orthogonalization, 277
range of, 42
Orthonormal, 268
scalar multiples of, 47
symmetric, 471
zero, 43
Lipschitz condition, 378
Lower bound, 182, 637 Parallelogram law, 1
greatest, 182, 637 Parseval's equality, 272, 319, 322, 363
Particular solution, 81, 92
Partition of a set, 39
Pendulum, simple, 167
M Permutation, 682
Matrix even, 683
addition, 70 odd, 683
characteristic polynomial of, 699 Periodic
definition of, 64 extension, 342
diagonal, 80, 465 function, 198, 347
identity, 66 Perpendicular
of a linear transformation, 64 projection, 282
minimum polynomial of, 77 vectors, 268
multiplication, 75 Phase angle, 171, 174
nilpotent, 79 Piecewise continuous function, 177, 330,
nonsingular, 689, 699 364
principal diagonal of, 77 Pointwise convergence, 306, 338, 656
similarity,699 Poisson integral form, 556, 616
symmetric, 472 Polynomial
transpose, 685 characteristic, 467,699
triangular, 471 Hermite, 438, 537
zero, 66 Laguerre, 444, 447
Maximum-minimum property, 654 Legendre, 278, 280, 410, 483, 559
Mean convergence, 307, 314 minimum, 72, 77
Mean square deviation, 287 space of, 8, 13
Mean value theorem Population growth, 166
for derivatives, 655 Positive definite, 257
for integrals, 656 Power series, 242, 661
Minor, 688 differentiation of, 242, 662
Mobius 707
strip, integration of, 663
Momentum, 168 Product
Multiplication of linear transformations, 48
of linear transformations, 48 of matrices, 75
of matrices, 75 of power series, 254
772 INDEX
field, 707
product, 704
Taylor series, 667
Vector space
Taylor's formula, 666
of functions, 3, 13, 183, 332
Test for linear independence, 20
of linear transformations, 47
Transient
of matrices, 70
charge, 173
of polynomials, 8, 13, 575
current, 171
Vibrating spring, 218
Transitive, 37
Transpose of a matrix, 685
Triangle inequality, 264 W
Trigonometric polynomial, 269
Wave equation, 509
d'Alembert's solution of, 515
U one-dimensional, 509, 516
Undetermined coefficients, method of,
Schrodinger, 510
158, 244
two-dimensional, 510
Uniform continuity, 652, 653 Weierstrass
Uniqueness problem, 83 approximation theorem, 360, 406,
Uniqueness theorem, 104, 718 408, 578
Upper bound, least, 637 M-test, 660
Weight function, 258
Wronskian, 111
Variance, 296
Variation of parameters, 141
Vector Zero vector, 6
addition, 1, 3, 6 Zonal harmonics, 562
CDE69
i
K.