Vous êtes sur la page 1sur 52

# Chapter

SERIES

OF FUNCTIONS

already run into series developments of functions several times : exponential, sine, cosine functions were expanded into power series; Taylor's theorem provides a way to develop series expansions for suitable functions; the exponential of a matrix gives us the only sure way to solve" We shall see in this chapter a system of constant coefficient linear equations. that a general technique for solving a differential equation involves approxi mation of the solution by series expansions. We shall begin by formulating the definition of convergence of a series of continuous functions and verifying the general criteria guaranteeing conver One of the most important of series expansions is that of power series. gence. We shall say that a function is analytic if it can be locally developed into a power series. We shall finally verify the fundamental theorem of algebra and complete the discussion of constant coefficient equations. We have delayed this until now because the kind of analytic techniques involved in the fundamental theorem are those which are most appropriately developed for the class of analytic functions. Further techniques for operating with power series will be explored, as well as the question of estimation of the error in replacing the power series by a partial sum.
We have

the

"

400

5.7

Convergence

401

5.1

Convergence
{/J be a sequence of continuous functions defined on a The series formed of the {fk} is the sequence of partial sums We say that the series converges if the sequence of these partial
(in the
sense

Definition 1.

Let

subset X of R".

(Zfc=i /*}
sum

converges

of Definition 19 of

Chapter 2),
to every

## and denote the

limitbyi?=iA.
Precisely then, /=
JV such that

Xfc=iA

if. corresponding

e >

0, there is

an

/(x)

k=l

ZA(x)
a

< e

for all

n >

N and

xe

## Since the limit of

uniformly convergent

(Theorem 2.14),
sequences,
we

we can

a

## Likewise, from the Cauchy criterion for

criterion for the convergence of series.

corresponding

Proposition 1. (Cauchy Criterion) Let {fk} be a sequence of continuous functions. The series /k converges if and only if, to each e > 0 there cor respond an N such that

X
Proof.
criterion.
m

<

for all

n,

m >

We must show that the sequence gn=^l=i fk satisfies the Cauchy For a given e > 0, let N be as in the proposition. Then, for any x,

>

:> N

\gm(x)

g(x)\
< e, so the

I /*(x)

<

2 /.
is proven.

<

Thus \\gm

g\\

proposition

guaranteed if the series of real numbers X*=i ||/J converges (for !!?=+ 1 /J !*=,,+ 1 ll/J). This gives us a powerful technique for verifying convergence of series.
Notice that the Cauchy criterion is Definition 2.
set X.

Let {fk} be a sequence of continuous functions defined The series is said to converge absolutely if ^f=1 ||/J < oo.

on a

402

Series

of Functions

Of course, as remarked above, an absolutely convergent series is conver gent. In the case of absolute convergence we can pose a comparison test,

just

as

on

Theorem 5.1.

sequence

sequence

numbers and

an

integer N

>

0 such that

(0 \\fk\\<Pn
00

fork>N

(ii)

Z Pk k
=

<

Then

/fc

converges

absolutely.
same as

Proof.

Examples
1.
r

zk
1.

VU

that domain, and

in

{|z|

<

r} for

any

<

For

\\zk\\ < rk in

Zr"
2.

Tzr

< oo

zk does
a

not

converge

uniformly

in

{\z\

<

1}.

In

fact, the

series is not

Cauchy

## sequence of functions, because for every n,

\fk

L,fk
k=l

m > n >

Thus, for \, say, there is no JV such that || "=,,/& II in n + 1 N, fact, not even for m
= =
.

<

i for all

3.

e7

X^=1 zk/k\
R"
and

converges

uniformly

in any disk

{|z|

<

R}

with

R finite.

Again, by comparison

Rk

*T\

zr!<0

5.1

Convergence

403

cos nx

4^ n
converges

uniformly

on

## the whole real line.

For any x,

cos nx

^n2
Since

1/rc2
{ak}

< oo

the

comparison

test

easily applies.

5. If

f(z)

=1

is any sequence of numbers such that \ak\ < oo, then on the closed unit disk. function is a continuous zk ak

The series converges uniformly since \\akzk\\ < \ak\. Finally, for the purpose of availability, we record the obvious extensions to series of the concerning integration and differentiation of

propositions

sequences of functions.

Proposition 2. (i) Let {/} be a sequence of continuous functions on the Let gn(x) \xafn- V tne series offunctions / converges, so
=

interval

[_a,b~].

## does the series

n=l

if/.=r(/.)
Ja
'a

<"> of continuously differ entiable functions on the If the series of functions # converges, and

\n=l

(ii)

Let

{/}
c,

be

for

some

/(c)

converges.

The limit is

## continuously differentiable and

(Z/,y =z/;
Examples
6.
CO

(5-2)

ln(l-x)=Z-r K
k
=

-1<*<1

404

Series of Functions

This follows
term
!*

by integrating

the

## geometric series (Example 1)

term

by

CO

.*

j0T^~t=S'oLt
co

in(i-x)= y

tk jfc+1
-

oo
=

k=ok + l

k^k

fk
-

"
,.

cos nx
-

fix)
is

k=i

n\

infinitely differentiable.
"

## For the differentiated series

sin

nx

is also convergent.

the series
S
"

(5.3)

can

be differentiated term

## By Proposition 2(ii) the sum is/'(x). Similarly, by term, and gives

n cos nx

(fi

1)!

which is
8. We

again convergent.
can

develop

following
1
i-

observations.

according

to the

r~x=^xk
x
k
=

",*
o

we

obtain

## by substituting -x2 for

1 + x

t=o

Integrate:
v2[+l
arctanx=

(-1)' n=o

zk + 1

5.1 EXERCISES
1. For what values of
x

Convergence

405

(a)

n=0

22V
cos nx

(d)

n=0

|
n

(jr+18)"

(b)

2
n=

(e>
1

2
=

""

(c)

n=0

(f )

| x<"2>
=

(a)

| hz=

(b)

|
=

-il
(Z)!
can

(c)

| -V
=

## 3. Which of these series

be differentiated

or

integrated

on

their

domain of convergence ?

(a) (b)

Exercise 1(a)

(c) (d)

Exercise

1(d)

cos nx

Exercise

1(b)

2
n
=

fl

1

expansion (c)

## for these functions

(a)

(l+x2)

Jo

e'1

dt

(b)

nm
Find
a

<

yx cos x.

PROBLEMS
1. (a) (Hint : (b)

expansion for sin 2x.) Find power series expansions for sin2 2. Prove Proposition 2.
power series
=

2 sin

x cos x

sin

and cos2

x.

3. Show that

lim
-.-l

n=l

log 2

L_ii=i0g2?

406

Series

of Functions

5.2

Algebra
we

## For the remainder of this

complex-valued
functions
are

functions of

chapter complex

restrict attention
The

variable.

the

## polynomial functions ; that is, functions of the form

+

P(z)

az"
are

a^iz"'1

+ axz + a0

complex numbers). We shall always assume a # 0; in degree of P. It is a basic fact of mathematics that every polynomial has a root; that is, there is a number c e C such that P(c) 0. The proof of this fact consists in a systematic investigation of the analytic properties of polynomials. First, we recall de Moivre's theorem.
this
case n

(where the at

is called the

Lemma.

Every
ce

nonzero

complex

number has

Proof.
em*
_

Let

C,

^ 0.

most convenient.

An nth root of

## For this purpose, the c is a number

polar representation
w
=

re'" is
r

pe'*
Let

such that

p"

and

n^

q js

an

2nk

277
i
=

47T
.

2tt(u
,
. .
.

1)
,

<*2

a*

a_!

2n

Then

a>i
=

expO'aj),
These
=

a>

exp(z'a)

are

all

distinct

and
p

have
=

the

property
=

(co)n
then

are
=

The numbers
c.

unity.
...,

Now, if

(pe'*)n

reie

c.

pe^ioi,

pe'^cu,,

...,

distinct, and

are

## all nth roots of

Now,
z-*co.

we

need two

polynomials.
place

The first is

deeper facts depending on the continuity properties intuitively clear: that \P(z)\ gets arbitrarily large

of
as

The second is the crucial fact for the fundamental theorem: the where a polynomial has a minimum modulus must be a root.

Lemma.

Let

(i)

lim
|z|->CO
>

P(z) \P(z)\
=

anz"
>

+ axz + a0 be a polynomial

of degree

n>0.

oo, that

>

0 there is

## K>0 such that

|P(z)| (ii)
are z

M whenever

\z\

K.

# 0, then z0 cannot be a minimum point for close to z0 such that \P(z)\ < |P(z0)|.

If P(z0)

5.2

## The Fundamental Theorem

of Algebra

407

Proof. (i) The point here is that the highest-degree term of P is regards the behavior of P as z^ oo. For z ^ 0,
the

dominating

term

as

1^)1

If |z|>ii:>l, then

k <n,

so

tf

>

!a|-i("j>l)
Mj\
)>-2M

Let M > 0 be

## given, and choose

max(l,2M|a|-1,2|fl|-1 (2

## Then, for \z\>K,

_l

"v

ak

>\a\[l

Thus

\P(z)\^\z\"-i\an\>.K-i\a\>M
(ii) Suppose
now

that

P(z0) ^0.

Let

## Q(z) =(P(z0)Y1P(z + z0)

Then G is also a polynomial, point for Q. Let

G(0)

1, and

we

## must show that 0 is not

minimum

Q(z)

1 +

2 * z*
as

! +

zm(am + z#(z^

where

# 0, and g(z) thus continuous (and that is all we is and a is polynomial 2s=m+i a*z"-(",+1). ^ to use the fact that for small z, zm need to know about g). Here again we want +* the to close polynomial 1 + zmam which has no minimum so Q is very dominates zm with r < 1). zm that z so -rjam 0 modulus at (choose choose an mth root of -a1 ; call it z, and consider the function
m

is chosen

the least

positive integer

k for which

ak

In

our

case,

we

408

Series

of Functions
r.

where
=

We have

l+rm(-l+rh(r))
is
a

h(r)

a^girzo)

## continuous complex-valued function.

Thus

\Q(rzo)\<\l-n+r+1\h(r)\
Now

limrrt(r) =0,
r-.0

so

we can

choose

r0 <

1 small enough

so

that

\r0h(r0)\ <i.

Then

I Q(rz0) I <, 1
which proves part

r0m +

i-o-G)

< 1

ir0m

< 1

(ii).
Let P be
=

Theorem 5.2.

## (Fundamental Theorem of Algebra)

There is
=

polynomial of

positive degree.

z0

C such that

P(z0)

0.

Proof. Let P(0) c0 By part (i) of the lemma, there is a K > 0 such that for \z\>K, \P(z)\>\c0\. Now A={zeC; |z| <K} is compact, so \P\ attains a minimum value on A, say at z0 But then z0 is a minimum point for all of C. For, since OeA, \P(z0)\ < |P(0)| and for z A, |P(z)| > c0 > |P(z0)|. Thus, even c0 for z A, we have |P(z0) | :< IP(z) | But then, by part (ii), there is no alternative : we must have P(z0) 0.
. .

Factorization

Theory
c

is

a zero

of the

(this is degree
zero

## proven below in Theorem 5.3). 1 less than that of P. If deg Q

Thus
>

polynomial P, then z c factors P P(z) (z c)Q(z) and Q has 0, Q has a zero c', which is also a
-

of P.

Further, Q(z)

(z

c')2'(z)
of P.

and

we can

order to find

exactly deg P

zeros

algebra.
Theorem 5.3.
n >

0.

There

are

## (Factorization Theorem) complex numbers a #0,

-

Let P be
z1;
.

polynomial of degree

z such that

P(z)
Proof.

a(z

zi)

(z

z)
on n.

The

proof is by induction

If

=1 the situation is

simple:

P(z)

OiZ

a0

ai

Iz

1)

5.2

## The Fundamental Theorem

of Algebra

409

(since ai ^ 0). Now we consider the case of general degree n, assuming the corollary for polynomials of degree n 1. By the theorem, there is a point c such that P(c) 0. Then
=

P(z)

P(z)

P(c)

2 *(z"

c*)
n

2 *(*

<o
\

( zW-1 )
"J

n-l/

on the right is a polynomial of degree n 1, so the induction assumption z-i. z_i) for suitable a^O, zu (z applies: it can be written as a(z Zi) we obtain z Thus, writing c

The factor

...,

P(z)=a(z-zi)---(z-z)
This factorization is

(5-4)

clearly unique, except for the order of the z;'s: a is the P and {zu of coefficient z} are the roots of P. Of course, leading If let not be need distinct; rs be the set of distinct roots. ru z Zjl, the list {zu zj, we let m,. be the number of occurrences of the root r{ in We can rewrite (5.4) as is called the multiplicity of the root r{
. . .

...,

mt

P(z)
and

a(z

rt)mi
=

(z

rs)m

(5.5)

on

real
=

polynomials.
a

Real

polynomials
come

Let

real

in

+ axz + a0 be

polynomial.
+

## conjugate pairs. If P(r) 0, then

=

P(z)

az"

-\

P(f)
so r

ani?)n +---+a1r
root of P.

a0=(anr"+---

+ a1z +

a0)~

PirY

is also

Since
+

(z
the

r)(z -r)

z2-(r

r)z

rr

z2-2 Re(r)z

\r\2

has real coefficients. Thus, if we rearrange the roots of P r, r, we into the real roots ru ...,rk and the conjugate pairs rk + l, rk+u and linear of a into quadratic product can rewrite the factorization (5.5)

polynomial

real

polynomials.
P(z)
=

a(z

rj"1

(z

rk)m\z2

Rfi(rt+1)z

k+il2) (z2-2Re(rt)z
+
' '

|rr|2)

410

Series

of Functions

PROBLEMS
4. Let w i, ...,co be the n nth roots of unity. Show that they are arranged at n equidistant points around the unit circle. Show that the sets {cui, oi}, {a>i, oil2, oj'c1} are the same, if u>, is the nearest such point to 1. 2 is 5. Let a>u Choose k so that kn w be the nth roots of unity. divisible by 4. Show that ikoju ikwn are the nth roots of 1. 6. Show that : (a) deg PQ deg P + deg Q. max(deg P, deg Q) if deg P # deg Q. (b) deg(P + Q) (c) When is the equation in (b) not true? 7. Given two polynomials P, Q show that there is a polynomial R which factors both P, Q and is factored by any polynomial which factors both P, Q. R is called the greatest common divisor of P and Q. 8. Show that a real polynomial of odd degree has a real root. 9. Prove that the polynomial 1 + zma (m > 0) has no minimum modulus
.

..,

.,

...,

at

0.

10. For

P(z)

2S'=o

az"

polynomial,

let

P'(z)=

Znanz"'1
n=

(a) Verify that the transformation P^P' is linear and satisfies (PQY=PQ' + P'Q (by induction
(P -*P' is
a
on

deg P).

r

## (b) Prove that

is

multiple

root of P if and

only if P(r)

0 and

P'(r)

0.

(c)

Define P"

=(P')',
m

P"

of at least multiplicity

if and

= =

r
=

is

root of P

Pfm"

*>(r)

0.

Now that
to
we

Equations

## know the factorization theorem for

complete
a

the

study
a

of constant coefficient

## polynomials we can return equations in one unknown func

operator of order k ; that is,

tion.
L is

Let L be

mapping L(f)
=

## from functions to functions defined

by

fm

'. f{i)
i
=

^C

(5-6)

5.3

Constant

Coefficient

Linear

Differential Equations

411

Corresponding

to L is the

polynomial

PLiX)

Xk

kZaiXi
>
=

## called the characteristic polynomial of L. know about such differential operators.

Theorem 5.4. Let L be
is
an

we

given by (5.6).
=

The collection

the equation

functions.

Lf=0 If r is a

n-dimensional

root

ofPL(X)

0,

vector

## Now if all the roots of the characteristic

solutions of Lf

polynomial are distinct, we have n 0, and it is easily verified (Problem 1 1) that they are inde
span

S(L). To examine the case of multiple roots, closely the relationship between the given differential operator and its characteristic polynomial. If P is a polynomial, we will let LP represent the corresponding operator; that is, for P(X) j=0 a,-*', FP is defined by

pendent.
we

Thus

they

## Now, from what

we

equations

we can
.

guess that the factorization of P will tell us all we want to know about LP In fact, we can factor the corresponding operator accordingly as the next

Lp+q

LP

Lq\ LPq

LpLq

Proof.

LP(LQ(f)).

=

composition

=

## of operators : (LPLQ)(f) We a little work.

= =

a0 then 0, that is, P(x) will prove it by induction on the degree of P. If deg P differentiable for sufficiently and any =LP(LQ(f)\ a0LQ(f) LPQ(f) PQ=a0Q function / Now suppose the lemma is true for all polynomials of degree n. Let P of degree n+ 1. If a is a root of P, we can write P(X) be a
,

Lsq

polynomial LsLq. We have left only to verify the lemma for polynomials of degree 1. That is, we must show that if R is a polynomial of degree 1 and T is any polynomial, X a, so that For once this is verified, we take R(X) then Lrt=LrLt.
S
a
n.
= =

## polynomial (X-a)S(X), where

is

of

degree

Thus, by hypothesis,

RS.

Then

412

Series
=

of Functions
=

## So, let R(X)

X-a, T(X)

2?U b, X'.
-

Then

RT(x)
Now

=f(b,- abl+1)X,+1
(
=

abo

we

compute LRLT :
\
in m

(m
(
=

IW" =2(w,0)'-^i,r
o

m
=

tibi-abi+l)f'^-ab0f
=

Q is
0.

as a

## of LPQ(f) We can, factors.

0 is

solution of LP(f)

Now let P be

by

the factorization

theorem,

write P

a

P(X)

with mt +

+ ms

deg P
(X

corresponding
-

to the factors

aj)1"1

are

in

Thus
=

we

## need to discover the solutions of the differential

=

0, where P(X) (X c)m. Consider, for example, the differential operator corresponding to (X c)2. We know one solution: ecx; we find another by the technique of variation of parameters. iX c)2 X2 2cX + c2. Test the operator on y zecx.
-

y'
Then

z'ecx + zcecx

y"

/'
Thus
z
=

2cy'

c2y

z"e"

or

z"

We

can

## guess then that the

general

situation is this.

Lemma 2.

The solutions

of LiX_c)m(f)

are

## spanned by e", xecx,

...,

xm~V*.
Proof. We by induction.
have to show that the named functions The
case m
=

are

solutions.

We do that Thus
we

1 is

## already known (by Lemma 1).

may

5.3

Constant

Coefficient

Linear

Differential Equations
m

413
we

assume

is
zero.

+ 1

By Lemma 2,

## need only verify that

Llx_c)m+i(xme")
=

But this is

L(x-c->"<Lix-c,(xmeC')

La.c)m(mxm-'e" +

cxme"

cxme")

La_c)m(mxm-1e")=0

by induction.
Theorem 5.5. Let

p(X)

X" +

j=o fl<-^f

oe a
=

coefficients. Let au ...,asbe the roots Then the space S(Lp) ms, respectively.
n-l

ofp(X) of solutions

0 with

## polynomial with complex multiplicities mt of the differential equation

lp(/)=/
is the linear span EXERCISES

+
i

Ei/(0 <
=

0 <j

<

m{

## (a) (b) (c) (d) (e) (f ) (g)

equations: 1 0, y"(0) 0, y'(0) 0, y(0) 4y 1 1 y'(0) 2, v"(0) 0, y(0) 5/ 3y y y" 1. 1, /(0) 0, y"(0) y- 6y" + 12/ ~Sy=0, y(0) 1 2, y"(0) 0, y(0) 3, /(0) y y~- 3y" + 3/ 2. 2, y '(0) 2, v"(0) 2, y'(0) yw + 2y" + ^ 0, v(0) "> 1, I2y' + 9y 0, y(0) /(0) y"(0) /4) + 4y('"> 2y<
y- 5y" + 8/
-

^"(0)=0. 3/ "' + 2y y^
-

0, y(0)

0, y'(0)

y"(0)

y "(0)

PROBLEMS
11.

(a)

Show that if ru 1
r

r are n

## distinct numbers, the matrix

r2 'It

is

nonsingular.

(Hint: If the

rows

are

dependent,

we

obtain

poly

1 with n distinct roots.) nomial of degree n functions The exp(/x) are independent. (Hint: exp(rix), (b) If these functions were dependent, we would be able to prove that the
..

.,

## columns of the above matrix

are

dependent.)

414

Series

of Functions

5.4

Solutions in Series
we are

If now
or

## given a linear differential equation which is not homogeneous,

has variable coefficients, we have a problem of a much different magnitude. In general, such problems cannot be solved explicitly. Thus, we must seek
ways to obtain

approximate solutions. This is one of the places where representations of functions are usable. The procedure of series approximation has two aspects. First, we must establish the theoretical validity of such a technique and, secondly (and this is essential from the practical point of view), we need a technique for effectively computing the In this section we shall describe this procedure, deferring these two error. essential points (which turn out to be the same!) until Section 5.7.
series

First,

an

example.

Suppose

we want

## the function / such that

=

f"(x)+gl(x)f'(x)+g0(x)f(x)
where g0 and gx

/(0)

a0

f'(0)

a,

that

they
us

are

give f(x)
Our

the

are defined in a neighborhood of 0. We shall assume sufficiently differentiable. Now our initial conditions first two terms of the Taylor expansion of/ at 0 :

a0 + atx +

higher-order
on

terms

(5.7)
"

assumption that the higherorder terms" are computable, and knowing enough of them will give a usable approximation to the solution. Now, evaluating the differ ential equation itself at 0 gives us the second-order term :

technique

will be based

the tacit

f"i0)+g1(0)f'(Q)+go(0)fi0)
or

f"(0)
so

-(g1(0)a1

go(0)ao)

f(x)

a0 + atx

i(aog0(0)

fli#i(0))x2
equation

## higher-order terms give

an

(5.8)

Differentiating the

differential

will

identity

express-

5.4

Solutions in Series

415

ing/'" Fix)
so

in terms of lower

derivatives,
+

so we

may
+

continue,

g[ix)f'(x)

<h(x)/"(x)

g'0ix)f(x)

g0ix)f'(x)

/'"(0)

-(^(0)

g0(0))ai
~

g'0i0)ao
+

^1(O)(01(O)a1
-

5o(0)ao)

(gi(0)2

9'M

9oiO))at

(0o(O)fif!(O)

g'oi0))a0

and

so we

## Taylor series of/:

/(x)

a0 + ajx
+ +

i(ao0o(O)
-

+
-

0i0i(O))x2
fif0(0))ai+ 0o(O)<h(O)
-

iCGhW2

3i(0)

3o(0))a]x3

higher-order

terms

9.

is in order.

We shall find

an

/'

(x2
=

1)/

+ xy
=

x2

(5.9)

X0)

y\0)

begins fix)

2x +

by substituting

## f (0) is easy Equation (5.9):

.

to calculate

f(x)

2x + x2 +
we

Differentiating (5.9),

obtain

y"

2x/

(x2

1)/

+ y +

xy'

2x

(5.10)
=

Evaluating at 0 we find /""(O) =/"(0) -/(0) 2. Differentiating (5.10) and evaluating at 0, we obtain /(4,(0) 2; once again gives 10. Thus, to five terms the Taylor expansion of the /(5>(0)
= =
-

desired solution is

fix)

2x + x2 +

\x3

h x4

A x5

higher-order
is

terms

## this is not very

glamorous, but it

computable!

The

phrase

416

Series

of Functions
represents the
error

"higher-order

terms"

## fifth-degree poly polynomial is com

pletely meaningless without some estimate on the error incurred. But our method gives no hint as how to estimate. So, in the hope of being able to give more form to the "higher-order terms," we will try a more brazen approach : we begin by assuming that the desired solution is the sum of a convergent power series (its full Taylor expansion") and we try to find the coefficients. If /(x) =0 ax", then differentiating term by term we
"
=

obtain

/'(x)=
/"(x)=

n=l

fn^x"-1
n(-lK'x""2
=

/w(x)

=
=

n(n
n

1)

(n

k +

\)anx"-k
given differential equations
and

{a} by equating
us

the coefficients

offm.
=

reconsider

(5.9).

Let

solution.

f(x) problem

"=0 ax"
becomes

be the desired

n(n
= n 1=2 2

lKx"-2
ay
=

(x2

1)
n=l

ax"-x

+
n=0

ax"

z2

(5.11)

a0

(5.11)

is

ik

2)(k
we

l)ak+2

ik- 1K_!

(k

l)ak+1

+ ak_x

Thus

a0 =0

equations

ay

2 0 0
=0

## 2a2~a1=2 3.2a3 + a0 2a2 4.3a4 + 2! 3a3

-

(initial conditions) ik 0) (k 1) (k 2)
=
= =

n(n

l)a

(n

(k

2)

5.4
We
can

417

can

-

n>2

(5.12)
we can we

However,
at
an

we

have

an

.

in that
In fact,

make

guess

## estimate for the general term a

2"

assert

k,

<

[n/3]!
largest integer less than or equal to x). This 0, 1, 2; we verify it in general by induction.
=

(5.13)
is in fact true for

([x]
n
=

Kl<

(fi-l)|aM-1|+(n-2)|fl.-3| nin 1)
-

<-(|an_!|
n

|a_3|)
+

W
<
n

2P_
-

\l(n

l)/3]!

[(n-3)/3]!

Now, since

in

3)

in

3)'

3
so

r?i ?
3

(?)

in

3)"

>
_

I [51 3

Similarly,

my
Thus,
1

2"

\a\<

3[n/3]!(

;_W3]!

418

Series

of Functions
tells
us a

This

now

lot.

to the

problem

in

(5.9)

differs from
+ + ^"X

ZX

T^X

by

at most

>6 ax",
by

satisfy (5.13).

Thus the

error

is

dominated

^3k\v\3k

## ^ Flrn W"= ^ a6[n/3]!

<

LnL+
fe!

<-)3k+l|v|3k+l
ka2

k!
-

[,

~)lk + 2\ v|3Ji+2
+

^
1
-

*!

(1

2|x|

4|x|2) (exp(2|x|)3

(2|x|)3)

## Hence in the interval 0

< x <

1 the solution is

polynomial except for our error of at most 7e2 The reader is forgiven if he is unimpressed with
should not go the paucity of of the
terms
so

our

estimate, but he
reason.

far

as

technique

for this

For

our

## results is due to laziness rather than the uselessness

le2
Let
us

If we pushed this procedure up to 1000 a task for (an easy computer), then the error would be at most which is less than (50)"90; a good estimate indeed. 2looo/l000!

Taylor development.

recapitulate
+
i

## the basic ideas.

We

are

given

an

initial value

problem :

ym
We

iW*)/0
=

m, yiO)

c0

y'iO)

Cl,

y*-x>(0)
"

ck_,

the #'s and h by power series expansions and test the solution" f(x) jj= 0 a x". The first k terms are found from the initial conditions, and the rest are found by equating the coefficient of x" on both sides of the

replace
=

equation.

This leaves

us

with these

problems

to

resolve:

(i) Can we represent the given #'s and /; by power series? (ii) Can we differentiate a power series term by term ? (iii) How do we multiply power series? (In the above illustration, the g's were polynomials, so there was not much difficulty.) (iv) Can the system of relations between the aks really be solved uniquely? (v) Can we effectively estimate the error between the solution and a finite part of its (supposed) Taylor expansion ?
Little

to

by little, we will resolve these problems. Suffice it to say that the (i) in general is No (see Section 5.8). However, in problems that

## 5.4 do arise series.

Solutions in Series

419

naturally, the given functions usually are sums of convergent power If this is the case, all other questions can be satisfactorily answered; that is, the solution also is the sum of a convergent power series whose co efficients can be determined by the above technique and the estimate on the
remainder
can

be

effectively computed.

Let

us

Examples
10.

x3y'"
y(0)

+
=

x2y"
1

xy'

+ y

ex

(5.14)
=

y'iO)

y"i0)

1/6
in

## Let the solution be

f(x)

^=0ax". Substituting
CO

(5.14),

we

obtain
QO OO

CO

n(n
n=0

l)(n

2)anxn

+
n
=

n(n
0

l)ax"

+
n=0

naxn

+
=0

anx"
CO
=

y"
-

t-0!
which

gives

these

equations
+

## for the coefficients :

ain(n
or

l)(n

2)

n(n

1)

1)

for all

a. "

(5.15)

n!(n3-2n2-n
we

l)

Notice, that

they

equation,
ditions.

have not used the initial conditions and fortunately conform to the requirements (5.15). That is, for this particular there is a unique solution independent of any initial con

This

does

not

any

apply (since

the

invertible).
11.

xy'
=

2y

0
=

(5-16)
0

yiO)

y'iO)

420

Series

of Functions
apply,
so we

should get
=

date.

## given (5.16) becomes

initial conditions.

Let/(x)

"=0

a

nin
n=0

l)flx"-2

nanxn
n=0

+
n=0

2ax"

or

a0

fli

0
+

(n
or

2)(n+l)an+2-nan

2an

n>0

(n
fl"+2

(n

2)a 2)(n+l)
-

=

0, aA

## 0 and thus all further coefficients

1

are

is/(x)

x2.

In the next

section,

we

shall
we

is most
domain.

have

fully develop the theory of power series. It already seen) to do so in the complex

EXERCISES 6. Find
an

=

y"

-xy

y'(0)

with

an error

same

[ 1, 1].

7. Do the

for

y-x2y
with

y(0)=0

/(0)=0

/'(0)=0

an error a

## of \0~* in [i, J].

a

recursive formula for the coefficients of the solution, and reasonable estimate:
8. Find

## (a) (b) (c) (d) (e)

2/ + v 1 /(0) 0. 0, v(0) /' 1 1 + /(0) y" 2/ xy e\ y(0) 1, arbitrary initial conditions. y(k) + v /'-/c2v=0. 0. / x2 + xy, y(0)
-

5.5

Power Series

421

PROBLEMS

Why doesn't Picard's theorem apply to Equation (5.14)? equation xy" + y'=0 seems to have only one solution by the series method, but two independent solutions by the method of separation of variables. Explain that.
13. The second-order

12.

5.5

Power Series

already discussed at length the power series expansion of the and trigonometric functions, the geometric series and some others. We have also seen that the Taylor formula produces a power series expansion exponential
for suitable functions.
We have observed that there is
a

We have

certain disk

corre

series, called the disk of convergence. The series We shall recollect all this converges inside that disk and diverges outside. information as the starting point of our discussion of complex power series.

sponding

to each power

Theorem 5.6.

Let cn be

sequence

of complex

numbers.

There is

a non-

z")

<

r)
with r<R.

and

uniformly

in any disk

{zeC:

## R has these two

descriptions:

(i) (ii)

R P

l.u.b.

{t: \c\t"

is

bounded}.

(limsup(|c|)1/")"1.
to

Proof. For at least part of the proof we could refer proposition we consider the set
[t
>0
:

Proposition

9.

As in that

there is

an

n)
upper

we can

take R

<x>,

(a)

Suppose \z\

>R.

Then there is

t,

have lim cz"

un

bounded.

=

n, we cannot

=0sojrz"
such that

diverges.

(b)

Let r<R, A

## (ze C:\z\ <r).

Then there is

t,r<t<R

422

Series

of Functions
say

{|c|/"} is bounded,

by

M.

If \z\ < r,
'

\cz"\

\c\t"

'

Thus, letting || || be the uniform norm for C(A), we have ||cz"|| < M(r/t)". Since rjt < 1, 2 (r/t)" < , so by comparison 2 c z" converges absolutely and uniformly
in

C(A).~

## Further, by definition, R is given by (i); the

shall leave
as

more

esoteric formulation

(ii)

we

Problem 14.

Examples
12.

If

cnz"

is

given

on

the circle

## of convergence P? The \z\

=

is that

practically anything

can

happen. \c\ < oo, then by 1}. Thus the series converges uniformly in

sequence

2 cz"

{c}

2 (z"/n2)

(c)
for any Since
we

## also has radius of convergence 1, but [( 1)"/"] does converge.

,

2 (l"/w)

does

z" has radius of convergence 1 but with \z\ 1 at all (lim z" # 0 if \z\
=

21 z"
=

## does not converge

!).

the circle of convergence is possible, needn't be concerned with the behavior of the series there (except
no on

general assertion

in

particular cases).
13.

The

geometric series

convergence 1.

absolutely

on

=0 z" is a power series with radius of This series converges to (1 z)_1 uniformly and Thus any disk {z e C: \z\ < r} with r < 1.
for

y~
1
-

z"

f0
a e

\z\

<

Now let 1
_

C,

^ 0.
1

Then
1
=

1
=
'

lz\"

z"
=

a~^~z

[1

(z/a)]

h \a~)

io ^
{z
6

14. The series

C:

|z|

<

|a|}.

=0 (z"/0

## has infinite radius of convergence.

5.5

Power Series

423

Thus the

continuous function on the whole plane, denoted does converge to the exponential function for real values of z. We have seen that, for real x, eix cos x + i sin x. We can use this (or to obtain series for the sine and Taylor's theorem) cosine:
sum a

is

ez since this

sum

cos x

i sin x

>
=

(ix)n
o

n!
x

^;

k%(4k)l
=
~

(4k
+

1)!

(4/c

2)!

(4k

3)!

(-l^x2"

(-l)fcx2t+1

4^o

(2k)!

lio

(2k

1)!
We
can use

These series also converge on the entire define the complex cosine and sine :
co

plane.

them to

-2k

oo

z2k+1

""-R-tm
We also have the
elz

to-(-*c2FM)i

(5-17)

equation
(5.18)
z

cosz

+ i sinz

for all
15.

complex

numbers

sum

## again that way).

obtain these other

## Replacing z by interesting equations :

ez
=

iz and

iz, alternately,

we

cos(

iz)

+ i
+ i

sin(

iz)

cos(i'z)

sin(z'z)

e~z Thus

cos(j'z)

sin(z'z)

ez + e"
=

2~
ez -e~
=

cos(iz) -isin(iz)

(5.19)

are

## For real values of z, the left-hand sides of

424

Series

of Functions
We
can use

hyperbolic cosine and hyperbolic sine, respectively. these expressions to define the complex cosh and sinh :
the
ez + e~z
.

cosh

sinh

ez-e~z
z
=

Because of
on

the
z

(5.19) and (5.20), the complex trigonometric imaginary axis, the hyperbolic functions :
cos(zz)
sinh
z
=

functions are,

cosh

-i

sin(z'z)
identities

## We should also note that the

trigonometric

bolic

ones.

Since
z

cos2(/z)
1

sin2(/z)

1, it follows from

cosh2

sinh2

## (see Exercise 10).

So does the series has radius of convergence 1. that the sums of see later k. We shall "=1nk (z"/n!), for any integer z" closed all these series can be given by expressions (such as

16.

"=1 (z"/!)

(1-z)-1).
fact, writing the polynomial p(z)

given by a power series. In "=0 anz" is the same as giving its power series expansion. What is more interesting is that any point in C can be chosen as the center of a power series expansion for
17.
A

polynomial

function in C is
=

p.

N

KZ)

=
=

aniZ
n

Z0 +

Z0)"

Using

P(z)=
All

(nW-z0)'zS-''
=

(5.23)
terms at will.

V /

sums

rewrite

Thus

we can

z0

Kz)=
n
=

(\ /kr"(z-z0)"
=

## which is the desired

expansion.

5.5

Power Series

425

generally, any series of the form =0 series Can we expansion centered at z0 power centered at a point other than the origin? The
More
.

-

15),

and the

proof

is like the

one

above for

## convergence intervenes after the

analog

of

Equation (5.23)

above.

It is

general fact that for any function given by a power series, we may move the center of the expansion to any other point in the disk of convergence. The
truly courageous student should try to prove this now; it can be done. We will give a proof later which is simple and avoids convergence problems but requires more sophisticated information about functions defined by power series expansions.

Multiplication of Power

Series
ex

Suppose /, g are complex-valued functions defined by power series Then we can find series expansions for pansions centered at a point z0 functions /+ g and/# also. Addition is easy: if, say
.

the

f(z)
then

an(z

z0Y

giz)

bn(z

z0)"

(f+g)iz)

yj(a

b)(z-z0y
for the

## But to find the series

=

expansion

product requires

little

more care.

0 (this involves no loss of generality). To say that Suppose that z0 that in a certain disk A, /is the limit of the polynomials is to z" say /(z) 2 a Thus, g is the limit of the polynomials ?=0 bnz". Similarly, az".
=

2^=o

fg is

we can

N

## ("=0 z")(^=0 bz").

Now,

/
\n

\
=

nz" )(lKz" / 0 / \n
=
=

afcmz"+'"
m
=

expression

## to form a series of powers of

z we

do not

the next few get a very aesthetic expression, but if we take some terms from we obtain a reasonable expression. the in sequence polynomials

k=0

g(

\n+m=k

anbm)zk
I

(5.24)

hope that fg is the limit of this sequence of polynomials. This is a reasonable hope; for even though we have modified the original sequence
We could

426
of

Series

of Functions
nor

polynomials

we

use

## by that sequence. In fact, by that fg is the limit of (5.21). Proposition

3.
=

making

careful

of this

fact,

we can

represented verify

## Let f(z) =0 az", g(z) of convergence of both series.

=0 bz"
Then

and suppose

is

(0 (/+ 9)iz) "=o ian + b)z* uniformly and absolutely in {zeC: \z\ <r), (ii) ifd)iz) f=o(.+m=jt anbm)zk uniformly and absolutely
= = =
=

in A.

Proof. Let pn(z) '2}=0akzk, q(z) =^l=0bkzk. By hypothesis p-+f, q -^g uniformly in A. Thus p + q ->/+ g, pq -+fg uniformly in A (Problem 2.55).
Since

Pn(z)

q(z)

k=l

2 ia- + b*)z"

rJLz)

0\t + J

we

-+fg uniformly

in A.
r.

## We know that pqn

But that is easy,

->fg

so

it would

seem

worth

our

while to compute pq

Pnq

rn
k

I
=

0\l + J

2
=

"'bX
k

t>n

Now, each

term on the
norms on

computing

## right is of the form aibjz'+J with i>n {z e C: \z\ <r},

or

j>n.

Thus,

\\pnq,~

r\\<y+\\aiz'\\\{ Ijl^^llj

+(,i0iia'z'ii)r_ifin^^ii)
<tl+Mr'\(f\bj\A (J>k')( _|+ \bj\A
+

5.5

Power Series

427

Now,

we

e

## 2?=o \a,\r', Jj=0

>

\bj\ r>

are

finite.

Let M be

number larger

than both.

0, there

are

(1) iVi>0 such that 2<=-.+ i \a,\r' < e if n > Nu (2) N2 > 0 such that 2j=+i \bj\ ri<eifn>N2, (3) N3 > 0 such that \\pq-fg\\ <e if n>N2.
These assertions follow from the known convergence of each n > max(7V~i, N2 Ns),
,

case.

Thus, if

## ll'-n-/^ll< llpn?n-/^ll+ II.P4,.-rJ|

<

|a,|r'

(21^1^
=

+21^1'-'

2 \bj

<

+ e-M+M-

(2M+ l)e

the

proposition

is concluded.

EXERCISES
9. Verify, in the way suggested in the text that cos2 z + sin2 z =1 is true 1 is always true. sinh2 z for all complex numbers z, and thus cosh2 z 10. Find a power series expansion for these functions:
=

(a)

exp(z2)
ez sin

(e)

e~'

J" ^Z
z

-dt
t

(b)

(f)

cosh

cos z

(c)

(g)
z

sinhz

(d)

Jo

exp(Z2)<fr
the power series that ez+w
=

11.

Verify by multiplying

ezew-

12. From the addition formula for the exponential the addition formulas for cos, sin, sinh, cosh.

PROBLEMS
14. If Show

{c} is
the

neighborhood
that

of which has

infinitely

many

## number, every members, is denoted lim sup c

.

of

convergence

of

the

series

2CZ"

is

(limsupflcl)1'")-1.

428

Series
15. 16.

of Functions
in
a

Expand ez Assuming

that

(1 +X2)"1'2

can

be

point z0 represented by
.

power series

centered at 0, find it. Find the power series for 17. Assuming that tan x can be represented by a power series centered sin x, find the power series ex at 0 and using the equation tan x cos x
arc cos x.
=

pansion of

tan

x.

5.6

Complex Differentiation
exponential function
is

lim
i-+0

=1
z

(5-25)

For

n=o

n\

so

e'-l

oo

_n-l

-V-1+'(^r)
power

oo

7n-2\

at 0.

ez
-

series,

so

is continuous

1
=

lim
2->0 z

1 + lim
z->0

zg(z)

From

(5.25)

and the

properties
z)
-

of the

exponential
.

it follows that

z-0

exp(z0) ^-^

exp(z0) lim
z->0

ez

I
=

exp(z0)

## The student of calculus will

recognize the limit on the left as a difference real ex quotient and the entire equation as a replica of the behavior of the more to consider idea a be It generally such good might ponential function. turns out to be a This the on a process of differentiation complex plane.

5.6

Complex Differentiation

429

very

significant idea,

because there
are so

are

Definition 3.

differentiable.

Let/be

## complex-valued function defined in a neighborhood

ofz0inC.
,. lim
z-zo

/is

differentiable at z0 if

/(z)-/(zo)
Z

Zq
case we

exists.
U.

In this

as

/'(z0).
we

If/ is

defined in

an

open

on

point

of U,

The usual

algebraic

facts

on

complex

domain.

are

differentiable

at z0

Then

so

are

f+g

and fg with

## the derivatives given

by

if+g)'izo)=f'izo)+9'izo)
ifg)'izo)
=

f'izoMzo) +fiz0)9'iz0)
is

entiable at z0 and
g

differentiable at z0 and /(z0) # 0. Then 1// is differ (l/f)'(z0) -f'(z0)/fiz0)2. (hi) Suppose f is differentiable at z0 and g is differentiable at f(z0). Then of is differentiable at z0 andig f)'(z0) g'(f(zQ))f'(z0).
(ii) Suppose f
= =

Proof.

These

## calculus that their

propositions are so much like the corresponding propositions proofs will be left to the reader.

in

Examples
18. The function
z0
.

is

z

1 for all

## A constant function is differentiable with derivative

zero.

Since
a suc

any

cession of
are

polynomial operations

is obtained from
as

## and constant functions by

described in

Proposition 4(i),

all

polynomials

differentiable.
z
-

## 19. The function

is nowhere differentiable.

## For the difference

quotient (z
and
as z
on

takes

z^)/(z z0), for zj= z0, is a point on the unit circle ranges through a neighborhood of z0 this difference quotient all values on the unit circle, so it could hardly converge.
-

430

Series

of Functions

Sum

of a

Power Series is

Infinitely Differentiable
was

## Our introduction to this section

essentially
We
now

proof
of
a

where differentiable.

sum

Theorem 5.7. Let

## verify this basic fact.

/(z)

j=0 anz"

of

convergence R.

Then

is

differentiable f'(z)=

at every

point

in the disk

n=l

Y,nanz"-1
of convergence
=

has the

same

as

"=0 az".
=

Proof,

lim

sup(|a|)1/n
same

## lim()1/n lim supGa,,!)1'"

,

lim

sup(|a|)1/n,

so

the series

radius of convergence as the given series. We must show 2 that it represents the derivative off. Fix a z0 \z0\<R and choose r > |z0|. The series 2"la"l'"""1 converges absolutely, so given e>0 there is an N such that

2">n "

lank""1

<e.

## quotient defining f'(z0) :

f(z)-f(z0)=
z

/z"-z"\
11
=

z0

z0

n=l

fai^z"-kzoA \k=l J
< r,

If \z\

<

r as

well

as

|z0|

then

n>N

2aitz"-kzk0A \fc= J
1
-

lt>N k
=

Similarly,

## |2> nanz0~ 1\<e.

Thus,
b-i

/(z)-/(z)
Zo

<

2e +

\Cn

z0

Now, by continuity, as z -> z0 the last term tends to zero. Thus, there is a 8 > 0 such that if | z z0 1 < 8, the last term is less than e. Thus, for \z\ <rand |z z0|
<

we

have

m-fco)
Z0

-2anzo"1

<3e

## which proves that the limit of the difference

quotient exists

and is

given by (5.26).

5.6

Complex Differentiation

431

In particular, since /' is given by a convergent power series, it also is differentiable, with derivative f"(z) ( \)az"~2, and so forth. We thus obtain these results, which form the complex version of Taylor's theorem
=
-

for

sums

## of convergent power series.

Let
=

f(z) Y^azn have radius of convergence R. infinitely differentiable in {z: \z\ <R). Furthermore, for every derivative, fm is given by a convergent power series,

Corollary 1.

Then

is

k the kth

/<(z)

n(n
n>k

1)
=

(n

k +

l)az"-"
in
a

Corollary 2.

Let
are

coefficients {an}
a
=

## f(z) j=0 az" be convergent uniquely determined by f:

The

/<>(0)

Notice that the definition of complex derivative is of the differentiation of functions of laries hold for functions of
a a

genuine generalization
Thus the
same

real variable.

corol

real variable

represented by
in
a

power series :

Corollary 3.
is

If f(x)

"=0

a(x

x0)"

neighborhood of

x0

then

infinitely differentiable

at x0

and

an

/(n)(*o)
:

/(x)

<n
n>k

1)

in

k)an(x

x0f-k

## left to the student.

proofs are implication of Corollary 2 is that the coefficients of a power series representation of a function are uniquely and directly determined by the function. In particular, a function cannot be
These corollaries
are

easily derived

written

as

the
us

sum

of

tion allows
cos
2

to

easily verify
z
=

## power series in more than the identity

one

way.

This observa

sin2

(5-27)

sums For the function cos2 z + sin2 z is a polynomial in functions which are can of power series and thus is the sum of a power series. Its coefficients

432
be

Series

of Functions
Corollary 3 just by letting z take on real values. (5.27) is the Taylor expansion of cos2 z + sin2 z, be the Taylor expansion for all z. Hence (5.27) is
to

But the

true.

of

## for real z, thus it must

always
The

Cauchy-Riemann Equation

complex differentiation with that R2, since R2 C. Suppose that is a function defined in a / complex-valued neighborhood of z0 x0 + iy0 is differentiable as a function of two real variables, then the differential If/ is defined and is a valued linear function on R2. Iff dfix0,y0) complexis also complex differentiable, then
of differentiation of functions defined
on
=
=
.

f'(z0)
exists.
Let

ru

r-

lim
z->zo

fiz)fizo)
Z

/e /,0-.

Zq

(5.28)
line. Then

->-z0

## along the horizontal

(5.28) specializes

to

ft \ f (z0)

x-*xo

i;fix>yo)-fixo>yo) hm X Xq
along
v

Sf
=

OX

(x0 v0)
,

(5.29)

If

we

let

-+z0

vertical,

we

also have

ft \ f (z0)

hm
y-j-o

fixo,y)-fixo,y0) Cv J'o)
~

ldf
=

Sy

(x0 y0)
,

(5.30)

Thus the
a

right-hand sides of (5.29) and (5.30) are the same. In conclusion, complex differentiable function must satisfy (when considered as a function of two real variables) this relation

(DM!)
This is called the Riemann

(5.31)

Cauchy-Riemann equation. More precisely, the Cauchyequations are found by writing /= u + iv and splitting into real and imaginary parts. Let us record this important fact.
Theorem 5.8. Let

f be

Split f

## complex differentiable function in a domain D. a function of two real

5.6

Complex Differentiation

433

variables

(z

iy).

Then these

hold in D:

dyji
dx
i

dy
du
=

(5.32)
dv dx
was

du dx

dv

dy

d~y

(5.33)
observed above.

du

Equation (5.33)

follows from

.dv

Bf

8u

,8v
'

d~x~~8~x

~dx

~d~y~~d~y

Hy
is

given by

dfizo ir
,

is))

^ (z0)r
f'(z0)r

-jt (z0>

+
+

if'iz0)s
is)
is
a

f'izo)ir

## complex differentiable function

complex

linear

complex-valued

function.

We shall show, via the techniques of the next few chapters, that a complex differentiable function can be written as the sum of a convergent power series. Thus, just by virtue of the differential being complex linear, the

PROBLEMS
18. Prove

of all

sum

of its power

series.

Proposition

4.
on an

## 19. Prove Corollaries 1 and 2 of Theorem 5.7.

20. Show that if / is an infinitely differentiable function ( , e), and there is an M > 0 such that

interval

|/w(x)|
then

<M

all of

allx

-s<x<e

/is the

sum

a a

## 21. Suppose /is plane. Show that:

the unit disk. power series which converges in complex differentiable function in a domain in the

(a) (b)

if

## if /is real-valued it is constant. |/| is constant, then /is constant.

434

Series

of Functions
Cauchy-Riemann equations in polar coordinates.
a

## 22. Write the

(Hint:
disk A.

Differentiate along the ray and circle through a point.) 23. Suppose / ,...,/, are given by convergent power series in If F is
a

polynomial

## in k variables such that

F(Mz),...,fk(z))=0
for real z, then (5.34) is true for all z in A. 24. Compute the limits of these quotients

(5.34)

as z

->

0:

arc

tan
z

cos z

(a)

(d)

sin hz

cos z

(b)

sin

(e)
z

TT72
sin
z

cos z

tanz

(c)

(f)

n=0, 1,2, 3, 4

complex- valued function of two real that/is a complex differentiable if and only if the differential df(z0) is complex linear for all z0 e D. 26. Suppose that / is twice differentiable in D, and is complex differ entiable. Show also that /' is complex differentiable. Supposing that that /is a quadratic polynomial in z. show 0, (/')' 27. If f=u+iv is complex differentiable in D and twice differentiable,
25.

Suppose / is
a

differentiable Show

variables in

domain D.

then

82u dx2

d2u

d2v dx2

82v

dy2

dy2

5.7

Differential

Equations
can

with

Analytic Coefficients
as

A function which

be

represented

the

sum a.

of

convergent power

series at

study posed

a

point

a e

We

the

some

of the

questions
verify
the

We

can use

In

and to

## provide following fact.

sought-for

estimates.

particular,

we

shall

5.7

Differential Equations
...,

with

Analytic Coefficients
are

435

Proposition
solution of the

5.

#fc_l5 gk

analytic

at 0.

Then the

ym
is also

+
i
=

g,yw
0

h(x)

y(0)

a0,...,

^""(O)

ak^

analytic

at

0; that is,

in

some

## vergent power series whose coefficients

disk centered at 0, it is the sum of a con can be recursively calculated from the

differential equation.
We

our

of the solution;

from Section 5.5, how to compute the Taylor coefficients business here is to show that the resulting series does in

producing

oo oo

are

## analytic in the interval |x|

< R.

Then

h(x)

n=0

" x"

g,(x)

n=0

a* x"

We some positive number M, \a\ <MR~", |a'| <MR'n for all ;' and n. shall obtain the desired estimate in terms of M, R and the initial conditions. For 0), leaving the general case simplicity, we shall do the homogeneous case only (h

and for

(Problem 27).
CO

If

fix)

2
n
=

^ X"

we

have

ak-i

Co

Oo

ck

=
. . . .

are

## found from these

equations :

2(n-l)---(n-k)cx"-"
=

+
(

## (fn(n-l)---(n- i)cx-) l(f "n'x") / / \n=l

=

(5.35)

0\n

now has a pain in his stomach similar to that of the author as this he wrote equation. Patience, dear readerthe fun has just begun ! Equating of equations for the coefficients of xm to zero, we obtain this recursive system

Surely,

436

Series

of Functions

coefficients :

m(m

\)

(m
+

k)cm
k-l m-k

2 2 aa'(m +
=

## (k+a)) (m (/c + a))cm +,_,*+) =0

or
1

k-l m-k

(m

k)

i=o a=o

2 2 ( + '-(* + ))
(m

(k

a))a'cm+i-(t+1,)
1 k
c0
,
.

(5.36)
1,
so

## By the restraints on i and the highest subscript of c

solve
purpose
assume

we

on

+ i the right is m
m

have

try to find
an

ck-i we can

## Equations (5.36) successively.

M>

We

now

estimate.

For this

1, and let

max

^,P(i?-l)-1(Pt-l),^,^y/J,0<y</c-l)
on

C is written

so as

to

assure

that

\cj\<iJ
We
now prove

(CM\J

(5.37)

this

## inequality for all

m,

by induction.

Thus

we use

(5.36), assuming

(5.37)

for all

n < m:

k-l m-k

I Cm I

<tn

i=o <z=o

2 2

\'\ km+l-(k + )l

-/Ml

0fb^

\~r)
(/n-k+l)

1 ATC"-1

"^

Now

J}=o

R-'

(R-"

1XP-1

I)"'

R(R 1)
1

l)-'(Rk

\)R-".

Thus

m-k+\ Mm
m

R(Rh
R

Rm

(~r
=

by definition of C. By this estimate we see that f(x) 2k*Lo c *" converges in the interval {x: |x| < R(CM)''}, and in that interval is the solution to our problem.

5.7
We
our we

Differential Equations

with

Analytic Coefficients

437

might perhaps
could

above estimates

usually

a better estimate by more clever substitutions ; but sufficient for the results desired. In any particular case be clever and obtain even better estimates.

were

Examples
20.

y'

exy'
a

+ xy

0, y(0)

0, /(0)

We will find

10~5
tion.

on

polynomial which approximates the solution to within [0.01, 0.01]. Let 2 cx" be the supposed solu By substituting the series we obtain
the interval

n(n
n
=

l)cx"-2

=

+
=

\=i

cx"+1
o

(5. 38)

-1
cm
=
-

give

c0

0,

1.

Equation (5.38)
\

becomes

lm-2 1

m(m
Thus

1)

-(m-l-i>m-i-i

+ cm_3

c2 c3
c4

-\cx

-\
+ ct +
= =

Cs

## c0) 0 \Qc2 -J1ji3c3 + 2c2 + Iq + cx) 5L -2Xo(4c4 + 3c3 + c2 + iq + C2)

TJo

so forth. The question is not really what the coefficients are (that is to be left to a machine) but how many coefficients need to be computed. The coefficients appear to be bounded (we could in fact show that they must be, cf. Problem 29). Let's try to prove that |cj < K for all n by induction. We have

and

m_2

1
-

|cJ

^73T) fS 7T('"
m

0|Cm-1-il

|Cm-31

my
as m >

so

long

4.

Thus
=

we

1/2.

## Then the difference between the solution

438

Series

of Functions partial
sum

of its

Taylor expansion

is dominated

by

n>k

|c||x|"<^|xr
Z >k

i|x|kl
L
1

|X|

for

|x|

<

1.

The interval
error

we are

<

10_1

so our

bound

on

the

is

i.i_.l
2

10"

iio
18
5

lv2 2X
-L

if k

"T

1 V4 34X

JT

JJ0A

i_ V5

1 y6 60A

by

at most

## 10-5 for all values of x in [-0.01, 0.01].

we

good an estimate in the interval just knowing that the coefficients are easy to see bounded is not good enough. We have to know that \c\ < Kr" for Let's try r some r < 1 and some K. \. That is, we attempt to < n. 2~" for all that Now, using the equation \c\ verify by induction defining {c},
Suppose [1, 1]. It is
21.

needed that
that

1
1

|Cml

\< S

m(m
K

1)
1

(my I ik

(m

1 i\
\

\ 2m-i-i^2n,-3J
1

/m-22;

e2
<

+ 4
m

K
7

## This is less than 2~mK

as soon asm>

2(e2
26

4),
we
=

orm>

26.

Thus choose
Thus

the induction step proceeds K so that the inequality holds for all

26 ;

need 2 will

(K

only do).

'

all

n.

## The desired solution differs in the interval

kth-order
1

Taylor polynomials by
1

c(D" n>k

lc*l

~fc-i

2"-1t'o2"^2''

on

ok-2

5.7

Differential Equations
5

with

Analytic Coefficients

439

if k is 19.

Thus

we

of the

Taylor expansion

approximation.

(l^y'
f(x)
=

exy

y(0)

y'(0)

to

an

accuracy of 10"

Let

=0

equations

0
n-2

c0=

1,

Ci

-1
c
=

n(n

1) \i=

(n-2

that the

-.cn-2-i) =oi! /
are

(5.39)
Suppose that

We will show

by inductions
n <m.

{c}

bounded.

Ic.l

<

K for all
1

Then
m-2 1
\

lm-2
.

|cm

<

i
m(m
K

1) \i=o

X(-i-0K+
.

>=o

t,k)
'!
/

/""- x
j
=

\
/

<

|"m(m
1)

1)

m(m
1 K 2

-1)1

m(m

<K

m(m
3.

1).
Thus
we can

as soon as w >

take A"

as a

so we

terms.

We have from

(5.39)

c2=

-\,c3= 0,

[-i, H

is

>k

klW<^ ^
=

n>k^

Z.

## This is less than 10"

when k

10,

so we

need 11 coefficients.

We

compute
c4
=

c\$ =20

c6

T20

etc.

440

Series

of Functions (giving
at most an error of

Up

to six terms

1/64),

our

solution is

~T

x^ + x^ + x^ 13x6 L2 20+720+'"

EXERCISES
13. Find
a

power series

## expansion equation, k(k

+

for the

general solution in

neighbor

(1

x2) /'

2x/
a

1) v

14. Find

power series 0

expansion

for the

general solution of

yh

Ixy' + 2k v

=/(x) such

that

## l, v(0) y + e-y x2, v, y(0)=0 (/)2 y(0)=0, (y')2=yy", / + 2x^2=0

=
=

v'(0)=0
y'(0)
=

16. How many terms of the power series for the solution y do we need: 3 (a) for an accuracy of 10" in the interval ( i, ) in Exercise 3(a)? 5 in the interval ( 10, 10) in Exercise 3(a) ? of 10" for an accuracy (b)

(c) 3(b)?

for

an

are

the solution of

## Equations (5.1), (5.2) polynomials?

PROBLEMS
in the text prove this theorem:

28.

...,gk-i

Theorem.

origin, then

any solution

## are analytic in an interval (R,R) about the of the differential equation

/k>

k2Zglyi"
r
=

can

be

expressed

as

the

sum

of a convergent power

series in

neighborhood
disk (|z| < R)

of the

g, h are

some

with R>\.

## Show that the solution of the linear differential

0

equation

y" + gy' + hy

(5.40)

5.8
is the

Infinitely

Flat Functions

441

sum of a convergent power series with bounded coefficients. 30. If the power series g(x) ~2,ax", h(x)=^bx" both have infinite radius of convergence, then so does the series expansion of the solution
=

of (5.40).

5.8

Infinitely

Flat Functions
are

which

we

have been

of

## is that the function have

derivatives of all order; even that however is insufficient. Another glance at Theorem 5.6 will remind the reader that there is a behavior requirement on these successive derivatives in order that the given function be the sum of its

Taylor expansion.

We shall show

by example

that there

are

infinitely
we

differentiable functions which are not sums of power series. make the notion of analyticity precise.
Definition 4. Let
a e

First,

shall

Let /be
sum

complex- valued
a

function defined in
ball

an

open set U.
a

U.

/is analytic
of

at a if there is
a

{z: \z

a\

<

r}

centered at

in U if /is

analytic

## convergent power series in this ball. at every point of U.

/ is analytic

We have deliberately stated this definition without reference to the domain of definition of the function; it applies equally well to functions of a real or are complex variable. The only functions which we know to be analytic For example, if/ is the sum of a convergent power the polynomials and ez. know that we can expand / in a series of we do not series at the

origin,
-

yet

a) with a any other point in the disk of convergence. We powers of (z shall see in the next chapter that this is the case. We have already seen that and now we will an analytic function has derivatives of all orders (is C) The clue to this function is a C function which is not analytic. produce given by
the

following fact,
6. lim
t.-0O

'
=

l'Hospital's

rule.

Proposition
Proof.

P(t)e~

## 0, for any polynomial P

Problem 31.
we

The function

have in mind

(Figure 5.1)
0

is defined

by

!exp(
0

*/
x^O

x >

(5.41)

442

Series

of Functions

Figure 5.1
a

is

certainly infinitely

differentiable at any

point

x0 #

0,

so we

need

only

Now,
0 all
n

(7(n)(x)

x <

exist and
are zero.

that for all n,

right

More

precisely

we

must prove

hm
x->0 x>0

<7<">(x)

ffM(Q)

(5.42)

We do this

by induction.

The

case n

0 is easy :

lim
X-+0 x>0

lim

x->0 X x>0

expl

X/

lim te~'
(->oo

To do the
x >

general

case we

<7(n)(x)

0.

Now

? (x)

exp( i)
-

'W
-

A pattern

seems

to be

developing.

5.8
For each
n

Infinitely Flat

Functions

443

there is

polynomial P

such that

ff(n)(x)
This
can

P (- exp (

^j

for

x >

(5.43)

be verified

by induction.

Assuming (5.43),

we

compute

'""<yyyy(yM-i)
'p-yi)
where

## Pn+1(X) -Z2(P'(Z) + Pn(X)). follows immediately from Proposition 6 :

= ,.

Now that

we

have this,

(5.42)

lim
x->0 x>0

<T(n)(x)

(T(n)(0)

,. =

1
-

lim

x-o X x>0

/1\
-

/
exp

1\
-

,. =

lim
f-oo

\X/

X]

rP(r)*T'

/N

a is also infinitely differentiable at 0. But it is certainly not analytic. Taylor expansion is "=0 0 x" which converges to o(x) only for x <, 0 and provides a poor means for approximating the value of <r(x) for x > 0. However, the fact that infinitely differentiable functions exist with this property has its bright side. The following construction will prove to be

Thus
Its

'

useful.

Lemma.

Given

<b, there
1

is

C00

function

Proof.

<

<

t^x)

ct(x(1

x < a.

## (See Figure 5.2.)

x))

has the

required properties

of

t0i.

We

then define

Toli(x)

T0

(H) "((K)('-r3)

444

Series

of Functions

Figure 5.2
Theorem 5.9.
tion

Let

## of open intervals covering

0

[a, b~] be a given interval and Uu Un a finite collec There exist C [a, b~\. functions pt such that
.

..,

<
= =

,

x e

Let Let

[/,

rt

Tib{

Then

t(x)

2 Tt(x) > 0

if

(ti(x)
Pix)
=

x 6

<*)
0
x

(a, bi)
,

^ (a, bi)
,

properties.

32. Let
a

## polynomial P, lim P(t)e~'

1-.QO

0.

be denned

by (5.31).

Define

\\t(\-t))dt
<d(x)
Jo

\\t(\-t))dt
= =

(a) w is C, (b) 0 ^ o(x) < 1, for all x, (c) to(x) 0, if x < 0, l,ifx>l. 33. Using Theorem 5.9 it can be shown that any continuous function is the limit of C" functions. Let/e C([0, 1]) and e > 0. Find a C function e. such that < # ||/ g\\ Here's how to do it. First pick an integer N>0 such that |/(x) f(y)\ < e/2 if \x y\<ljN. Now cover the interval [0,1] by the intervals

Show that

(d)a>(x)

5.9

Summary

445

U0,

...,

UN, where

u'-(n- n)
and let px,...,pN be the

/i-l /+1\

Let

^-l/^H
For any x,
x

is in

only

{U,},

say

UT, U,+1.
e
e

Then

l/(*)-*(*)l=M*)

**>-'()

+ Pr+1

/-,(^)

<2

5.9
Let

Summary
{/J
be
a

## The series formed from

the/k is the sequence of sums Q=1/J. If this sequence converges, we The series say that the series converges and denote the limit by = t fk
.

converges

absolutely

if

"=1||/t||

< oo.

arbitrarily

small

by choosing
If there is
a

m,

The

can

be

comparison test.

sequence

{pk}

of

positive

numbers and

an

integer
(0 ()
then

## N > 0 such that

IIAII<P*

fork>iV

YjPk

<

/k converges absolutely.
If

integration.

converges,

so

does

Jj/

and

j;(z/.)-i({
fundamental theorem of algebra.

If P is

polynomial :
(5.44)

P(z)

anz',

---

+ a1z + a0

446

Series

of Functions
.
.

with a # 0. Then P(z) has a complex root. If r1; rk are all the roots of P, then corresponding to each root there is a positive integer m( (called
.

the

multiplicity)
mt +

such that
+ mk

(i) (ii)

Piz)

(z

r^"'

iz

rk)m*
we

(5.45)
associate the constant

To the

coefficient

LP(f)

afM

---

a1f'

a0f
of LP
.

## P is called the characteristic

polynomial

These formulas

are

valid :

Lp+Q
If

LP

Lq

LPq

LpLq
of

solutions of LP(f)
erix

## (5.45) is the factorization of P, then the kernel 0, is spanned by the functions

=

LP the collection of
,

xmi~1erix

r2x
-ritx

xm2-ler2X mji-lgi-kjc
be
a

Let

{cn}

sequence of

number R

There is

nonnegative
c z") with

power series

## cz" diverges for \z\ cnz"

converges

>

R.

absolutely

in

{|z|

<

/}

for

r <

R.

(c)

P=[limsup(|C|)1/"]-1.
=

If/(z)

anz", g(z)
+

YJbnzn

in the disk

{\z\

<

r},

then

f(z)

g(z)=
n

t(an + bn)z"
=

fiz)giz)=
k=0

\n+m=k

abm)zk
1

in that

same

disk.

5.9

Summary
C,
we

447

,. hm

function defined
if

near

z0 in

say that

is

fiz)-fizo)
z
-

,,, =/(z0)

z0

exists.

The

sum

of

at

GO

sum

00

/(z)=az" n=0
Thus the
sum

/'(z)=

n=l

az"-1
A

of

## convergent power series is infinitely differentiable.

differentiable

complex-valued function of two real variables is complex differentiable if and only if it satisfies the Cauchy-Riemann equations :

f--4f) \dy!
dx in If h, g0 gk can be represented as sums of convergent power series a disk centered at zero, then the same is true for all solutions of the differ
,
. . .

ential

equation

ym

kZgiyw
i=0

+ h

Furthermore,

once

given

## the initial conditions

=

y(0)

ao,...,/k-i\0)

ak-1

the coefficients of the power series can be recursively calculated using the differential equation. Given any finite covering of the interval [a, b~] by open intervals Uu
...,

## U we can find C00 functions pt,

...,

p such that

(a)

(b) (c)

0<p,<l 0, outside Ut pt
=

Piix)

1 for all

x e

[a, U]
partition of unity
on

These functions
cover

are

called

[a, 6] subordinate

to

the

Uu

...,

U.

448

Series

of Functions

I. I.

## Hirshman, Infinite Series, Holt, Rinehart and Winston, New York,

Several
of view
on

1962.

Cartan, Elementary Theory of Analytic Functions of One or Complex Variables, Addison-Wesley, Reading, Mass., 1963. This text develops the subject of complex analysis from the point

H.

of power series. It also contains a complete discussion of the theorem existence of solutions of analytic differential equations. Further material
T. A. Bak and J.
can

be found in

Lichtenberg,

## Mathematics for Scientists, W. A.

Benjamin,

Inc.,

New

York, 1966.

Kreider, Kuller, Ostberg, Perkins, An Introduction to Linear Analysis, Addison-Wesley, Reading, Mass., 1966. W. Fulks, Advanced Calculus, John Wiley and Sons, New York, 1961. M. R. Spiegel, Applied Differential Equations, Prentice-Hall, Englewood Cliffs, N. J., 1958.
MISCELLANEOUS PROBLEMS 34. Find defined
x e on

sequence {/} of continuous nonnegative real-valued functions the interval (0,1) such that f(x) =2=i /(*) exists for all

35. For
oo

## [0, 1], but /is not continuous. |z| < 1, define

7k

lnz=2r k
k=l

Show that for all such z, z 36. Show that the series

exp(ln(l

z)).

i(z-n)2
converges to
a

C-{1,2,

...,,

## 37. Show that

exp(z)

lim

(1 + z/m)m.

(Hint:

Compute the

power

m-^oo

series expansion of (1 + zjm)m.) 38. Show that a real polynomial of odd degree always has 39. Let oj exp(277f'//j). Show that
=

real root.

z"

(1

ojz)(\

## a>2z) (!- wnz)

5.9
40. Let P be

Summary

449

a polynomial. Show that P is the square of another poly only if every root of P occurs with even multiplicity. 41 If P, Q are two polynomials, P divides Q if and only if every root of P is a root of Q with no larger multiplicity. 42. Suppose P is a polynomial of degree at least two. Show that there is a c such that P(z) c 0 has at least one multiple root. (Hint: Con sider P' as defined in Problem 10. If P'(d) 0, take c P(a).) 43. Let /i, / be functions in C(X). Show that these functions are independent if and only if there are points x,,...,x such that the matrix (f(xj)) is nonsingular. 44. Show that the functions e'x, xe", x"e'x are independent. 45. Show that if P, Q are polynomials, S(LP) c S(LQ) if and only if P divides Q. 46. If P is a polynomial of degree at least two, there is a c e C such that the equation LPf= c/has a solution of the form xe'x. 47. If a linear differential equation has polynomial coefficients, it has global solutions on all of R. 48. Let {c} be a sequence of complex numbers such that 2 knl < o. Let f(z) =2=o cz". Prove that |c0| is not a relative maximum of |/|,

nomial if and
.

..,

## unless all other coefficients vanish. 49. Suppose the function

f(x, y)

x2

y2 + iv(x, y)

is complex differentiable. Find v. 50. If / is a polynomial in x, y which is complex differentiable, then /(x, v) has the form Q(x + iy), where Q is a polynomial. (Hint: Substi
tute
x

51. in C.
52.

(z + z)/2, y (z z)/2, and use the Cauchy-Riemann equation.) Suppose /is a C2 complex-valued function defined on a domain D Show that if / and fz are both harmonic, then / is complex differ
= =

entiable.

If/

are

## complex differentiable and |/|2 + \g\2 is constant, then both

a

/,
A

are

constant. one-to-one

53.
<=

Suppose that / is

Let g : A -*- Z> be the inverse of differentiable, so is g. C. the 54. We may consider the function ez x + iy, u Re ez, v plane. Let z
=
=

as a

to

\mez; that is

e"

cos

e* sin y

that this

mapping

=

const,

on

the circles

centered at the

origin,

the lines y

<y < a

through the
takes

## Show that in any interval

{a

+ 2-n-} this

mapping

every value

precisely

once.

450

Series

of Functions
particular,
ez maps the horizontal

strip { tt <y < it) one-toplane negative real axis. Call this domain D. Define the complex logarithm log z: Z) - { tt <y <tt} as the inverse of this mapping. Show that log z is complex differentiable
one

(c)

In

and

(log)'

z can

be

represented by

## power series centered at 1

in the disk {|z 1| < 1}. (Recall Miscellaneous Problem 35.) that this provides a way for extending real functions to the

Notice

complex
.

expansion

domain besides that of power series. For example, the power series about 1 of log x extends it only to the unit disk centered at 1 The above extension of

log is defined

in the entire

## negative real axis.

55. Consider z2
maps the open

as a

This process is called analytic continuation. mapping of the plane into the plane. Show that it
one-to-one onto the domain D of Problem 54.

Let V
a

## be the inverse, and show that Vz is complex differentiable.

cos

Provide

similar discussion for the mapping z". 56. Discuss the mapping properties of
= =

z,

sin

z.

57. Show that the power series expansion of the solution of Exercise 8(d) with initial values y(0) 0 does not converge outside the unit disk. 1, y'(0) 58. I.

Suppose that/

are

## complex-valued functions defined

on

the interval

Show that

/(r)

j dt

hz-g(t)
is

complex differentiable and can be represented by a power series at any point of the image of g. 59. If /is C'inCx Xand for each fixed x,f(z, x) is differentiable in z,
then

F(t)=jf(z,x)dx
is also complex differentiable. 60. The equation of Exercise 6 is called Legendre's equation and the solutions {/*} for integral k are called the Legendre polynomials. They have this

interesting property:

fm(x)f(x)

dx

if

5.9
To prove this
we

Summary

451
as

must

=

((l-x2)/)'
Thus

k(k+l)v

\\

f-f-

^TT)

/.

[(1
i

*'>/<*>]'/<*>

dx

differentiable.
Show

same,

interchanging
=

and

n.

d.

## Show that /(z)

a

ep(2' is complex

that/(n)(z)e"',<z)
polynomial

is

polynomial

of

d"

exp(x2)

(exp(-x2))

equation y"

2/ + 2ny

0.

## (a) Find properties :

63.

C real-valued function

/ defined

on

## R" with these

(i) 0</(x)<l (ii) f(x)>0if \\x-x0\\<R (iii) /(x)=0if ||x-Xo!l>P (By C" we tinuous.) (b) Let
Bi,
. . .

mean

all

.

exist and

are con

A' be
<=

.

are

balls in
to

,

B is

## A partition of unity on X subordinate u B Bi u collection {/,...,/} of C" functions such that

(i)0</<l

(ii) /,(x)
(iii)

0 if

x
=

\$ Bi
lif*e*

2;=i/<
a

Find such

partition of unity.