You are on page 1of 81

Chapter 10.

Sequences and Series of Functions


For the materials to be discussed below, we will need to introduce the extended real number system [
Definitions. (1) Let [ 1; C1] denote the set R [ f 1; C1g: We extend the ordering of R to [
defining 1 < x < C1 for every x 2 R: (Since 1; C1 62 R; they are not numbers! We treat
symbols.)

1; C1]:

1; C1] by
1; C1 as

Let S be a nonempty subset of [ 1; C1]: Then C1 is an upper bound of S in [ 1; C1] and 1 is a lower
bound of S in [ 1; C1]: (So every nonempty set S has at least one upper and one lower bound.) Define sup S to be
the least upper bound of S in [ 1; C1] and inf S to be the greatest lower bound of S in [ 1; C1]: (Note that for
nonempty S ; sup S and inf S always exist in [ 1; C1]:)
(2) For x 2 R [ fC1g; c > 0; let x C .C1/ D C1 D .C1/ C x ; c.C1/ D C1 D .C1/c; c. 1/ D 1
. 1/c: Similarly, for x 2 R[f 1g; c < 0; let x C. 1/ D 1 D . 1/C x ; c. 1/ D C1 D . 1/c; c.C1/
1 D .C1/c: Also, j C 1j D C1 D j 1j:

D
D

(3) For x 1 ; x 2 ; x 3 ; : : : 2 [ 1; C1]; we say lim x n D C1 iff for every real number r > 0; there exists a K 2 N such
n!1
that n  K implies x n > r: Similarly, we say lim x n D 1 iff for every real number r > 0; there exists a K 2 N
n!1
such that n  K implies x n < r:
Below x 1 ; x 2 ; x 3 ; : : : will denote a sequence of real numbers. If lim x n exists, then we can see where the sequence
n!1
is converging. If the limit does not exist, we still would like to get some information on the distribution of the sequence.
Recall a subsequence of x 1 ; x 2 ; x 3 ; x 4 ; : : : is a sequence x n1 ; x n2 ; x n3 ; x n4 ; : : : ; where the indices n 1 ; n 2; n 3 ; n 4 ; : : :
is a strictly increasing sequence of positive integers. For example, taking n i D 3i; we get the subsequence
x 3 ; x 6 ; x 9 ; x 12 ; : : : : Taking n i D i 2 ; we get the subsequence x 1 ; x 4 ; x 9 ; x 16; : : : : Taking n i to be the i-th prime number,
we get the subsequence x 2 ; x 3 ; x 5 ; x 7 ; x 11; : : : : For x 1 ; x 2 ; x 3 ; : : : ; consider the set of subsequential limits

LD z

2 [ 1; C1] : z D klim
x
!1 n

for some subsequence x n1 ; x n2 ; x n3 ; : : : of x 1 ; x 2 ; x 3 ; : : : :

If the sequence is bounded in R; then by the Bolzano-Weierstrass theorem, there is at least one subsequence that has
limit, i.e. L 6D ;: If the sequence is unbounded in R; then either 1 or C1 is in L; so again L 6D ;: Therefore sup L
and inf L always exist in [ 1; C1]:
Definitions. Define the limit superior (or upper limit) of x 1 ; x 2 ; x 3 ; : : : to be sup L and the limit inferior (or lower limit)
of x 1 ; x 2 ; x 3 ; : : : to be inf L: For limit superior, the notations are limsup x n or lim x n : For limit inferior, the notations
n!1

n!1

are liminf x n or lim x n : (Briefly, the limit superior is the largest subsequential limit and the limit inferior is the least
n!1

n!1

subsequential limit.)

Remarks. (1) The limit superior and limit inferior of x 1 ; x 2 ; x 3 ; x 4 ; : : : always exist in [
liminf x n D inf L  sup L D limsup x n :
n!1

1; C1]: We also have

n!1

(2) If fx 1 ; x 2 ; x 3 ; x 4 ; : : :g is not bounded above in R; then

C1 2 L; so limsup xn D sup L D C1: Similarly, if


n!1
fx1 ; x2; x3; x4; : : :g is not bounded below in R; then liminf
x n D 1:
n!1
(3) lim x n D z 2 [ 1; C1] () L D fz g () limsup x n D z D liminf x n :
n!1
n!1
(4) limsup. x n / D sup. L/
n!1

inf L D

and liminf.cx n / D c liminf x n : For every c


n!1

n!1

liminf.c C x n / D c C liminf x n :
n!1

n!1

D sup.cL/ D c sup L D c limsup xn


n!1
2 R; limsup.c C xn / D sup.c C L/ D c C sup L D c C limsup xn and

liminf x n : For c > 0; limsup.cx n /


n!1

n!1

n!1

n!1

78

n!1

Theorem. For x 1 ; x 2 ; x 3 ; x 4 ; : : : ; define


Mk

D supfxk ; xkC1 ; xkC2; xkC3; : : :g

and

mk

D inf fxk ; xkC1 ; xkC2; xkC3; : : :g:

 M2  M3  M4  : : : ; m 1  m 2  m 3  m 4  : : : ;
(2) limsup x n D lim Mk ; liminf x n D lim m k and both are limits of some subsequences of x 1 ; x 2 ; x 3 ; : : : :
n!1
k !1
k !1
n!1
Proof. (1) Since Mk  x kC1 ; x kC2 ; x kC3 ; : : : ; so Mk  supfx kC1 ; x kC2 ; x kC3 ; : : :g D MkC1 : The inequalities for m k s
(1) We have M1

are similar.

1; C1]: Let M D klim


M : We have
!1 k
2 L; z D jlim
x n  lim Mn D M : So sup L  M :
!1
j !1
Conversely, if M1 D C1; then x 1 ; x 2 ; x 3 ; : : : is not bounded above in R and so sup L D C1 D M : Otherwise, M1 2 R: By the supremum property, since M1 D supfx 1 ; x 2 ; x 3 ; : : :g; there exists some x n such that
M1 1 < x n  M1 : Next, consider M1Cn D supfx 1Cn ; x 2Cn ; x 3Cn ; : : :g D supfx k : k > n 1 g: By the supremum
1
property, there is n 2 > n 1 such that M1Cn
< x n  M1Cn : Keep on repeating this, we get strictly increasing
2
1
n j s such that M1Cn
< x n C  M1Cn : Then x n ; x n ; x n ; : : : is a subsequence. By the sandwich theorem,
jC1
lim x n D lim M1Cn D M 2 L: Then M  sup L: Therefore, sup L D M : The limit inferior case is similar.
j !1
j !1
(2) As both the Mk and the m k sequences are monotone, their limits exist in [

to show sup L D M : For every z

j 1

Corollary. If x n

 yn ; then limsup xn  limsup yn and liminf


x n  liminf yn :
n!1
n!1
n!1

n!1

Proof. This follows from observing that the Mk s and the m k s for x n s are less than or equal to those for yn s.
Examples. (1) Since lim

n!1

1Cn
D 14 ; so by remark (3), limsup 31CC4nn D 14 and liminf
D 14 :
n!1 3 C 4n
n!1

e n if n is odd
then 0 < x n < C1: So, the limits of the convergent subsequences are in [0; C1];
en
if n is even,
i.e. L  [0; C1]: Since lim x 2nC1 D 0 2 L and lim x 2n D C1 2 L; so by definition, liminf x n D inf L D 0 and
n!1
n!1
n!1
limsup x n D sup L D C1: (Alternatively, by part (2) of the theorem,
(2) Let x n

1Cn
3 C 4n

n!1

Mk

D
mk

supfe k ; ekC1 ; e .kC2/; ekC3 ; : : :g


supfek ; e .kC1/; ekC2 ; e .kC3/; : : :g

if k is odd
if k is even

inffe k ; ekC1 ; e .kC2/; ekC3 ; : : :g


inffek ; e .kC1/; ekC2 ; e .kC3/; : : :g

D C1 H)

if k is odd
if k is even

D 0 H)

limsup x n
n!1

liminf x n
n!1

D klim
M D C1;
!1 k

D klim
m D 0:/
!1 k

8 3C
<
3 14 ; : : : defined by x n D 0
: 1

if n D 3 j C 1
if n D 3 j C 2 :
if n D 3 j C 3
j C1
1
1
1
1
Observe x 1 D 4 < x 4 D 3 < x 7 D 3 <    < 0 D x 2 D x 5 D x 8 D    <    < x 9 D < x 6 D < x 3 D 1:
2
3
3
2
Then the Mk sequence is 1; 1; 1; 12 ; 12 ; 12 ; 13 ; 13 ; 13 ; : : : ; j C1 1 ; j C1 1 ; j C1 1 ; : : : : So by part (2) of the theorem, limsup x n
(3) Consider the sequence

4; 0; 1; 3 21 ; 0; 12 ; 3 13 ; 0; 13 ;

D klim
M D 0: Similarly, the m k sequence is 4; 3 12 ; 3 12 ; 3 21 ; 3 13 ;
!1 k
.3 C j C1 1 /; : : : : Again by part (2) of the theorem, liminf x n D lim m k D
n!1
k !1
79

1
j C1

n!1

3 13 ; 3 13 ; : : : ; .3 C j C1 1 /; .3 C j C1 1 /;
3:

The following are theorems involving limit superior and limit inferior.

6D 0 for all n ; we have


a
a
p
p
nC1
nC1 :

liminf
j
a
j

limsup
j
a
j

limsup
liminf
n
n
n!1
n!1
an
an
n!1
n!1
a
a
p
p
nC1
nC1 may not exist.)
(So, if lim
D
L
;
then
lim
j
a
j
D
L
:
However,
if
lim
j
a
j
exists,
then
lim
n
n
n!1
n!1
n!1
n!1 an
an
Proof. We will prove the rightmost inequality only. (The middle inequality follows from remark
(1) and the proof
anC1
: The case M D C1 is
of leftmost inequality is similar to that of the rightmost inequality.) Let M D limsup
an
n!1
trivial. So suppose M < C1: For every real number r > M ; since
 a a 
k C1 k C2
;
;::: ;
M D lim Mk ; where Mk D sup
k !1
ak akC1
a a
a a a a
k C1 k C2
n D kC1  kC2    n < r n k ;
there is Mk such that Mk < r: Then
;
;
:
:
:
<
r
:
So
for
n
>
k
;



a
a
ak
ak
akC1
an 1
k
k C1
p
p
p
k
which implies jan j < r jak jr : Since lim c D 1 for every positive constant c; so
n!1
p
p
p
limsup jan j  limsup r jak jr k D lim r jak jr k D r:
n!1
n!1
n!1
p
Taking limit as r ! M C; we get limsup jan j  M ; which is the rightmost inequality.
Theorem. For a sequence a1 ; a2 ; a3 ; a4 ; : : : of real numbers with an
n

n!1

Remarks. The theorem tells us that if the convergence of a series can be detected by the ratio test, then the convergence
of the series can also be detected by the root test, but not vice versa.
Theorem (Strong Form of Root Test). For a sequence a1 ; a2 ; a3 ; a4 ; : : : of real numbers,

jan j < 1; then n6D1 an converges absolutely.


n!1
p
1
(2) If limsup jan j > 1; then 6 an diverges.
n
D1
n!1
p
Proof. Let M D limsup jan j: If M < 1; then choose r so that M < r < 1: Since
n!1
p
p
M D lim Mk ; where Mk D supf jak j; C jakC1 j; : : :g;
k !1
p
p
there exists some M such that M < r: Then ja j; C ja j; : : : < r: So ja j < r n for n  k : By the comparison
(1) if limsup

test, the series will converge absolutely.

k 1

k 1

k C1

If M > 1; then since M is the limit of some subsequence of n jan j; there are infinitely many
there are infinitely many jan j > 1: So lim an 6D 0: By term test, the series diverges.

pja j > 1: Then


n

n!1

Theorem (Strong Form of Ratio Test). For a sequence a1 ; a2 ; a3 ; a4 ; : : : of real numbers with an
(1)
(2)

a
1
nC1
if limsup
< 1; then 6 an converges absolutely.

a
nD1
n!1
a n
1
nC1
If liminf
> 1; then 6 an diverges.

n!1
an
nD1

Proof. This follow from the theorem above and the strong form of the root test.
80

6D 0 for all n ;

Pointwise Convergence
Question. How do calculators compute sin x ; cos x ; e x ; x y ; ln x ; : : :?
For

1 < r < 1;

1
1

D 1 C r C r 2 C r 3 C    : Integration suggests

ln.1

r/

D r C r2 C r3 C r4 C    :

.x 1/2 .x 1/3
Setting x D 1 r suggests ln x D .x 1/ C
C    : As 1 < r < 1 is required, this formula
2
3
may only work for 0 < x D 1 r < 2: Note the formula is consisted of monomial terms, which may be computed by
hand in principle. Now the formula is a series of functions in x : What is the meaning of convergence if the terms are
functions?
Definitions. We say a sequence of functions Sn : E ! R converges pointwise on a set E to a function S: E ! R iff for
every x 2 E, lim Sn .x / D S .x / 2 R (i.e. for every x 2 E, for every " > 0, there exists K 2 N (K depends on x and
n!1
") such that n  K ) j Sn .x / S .x /j < ".) In that case, S .x / is said to be the pointwise limit of the sequence Sn .x /:

1
k D1

! R iff for every x 2 E,


lim . f1 .x / C f2 .x / C : : : C fn .x // D S .x /, i.e. the sequence of partial sums Sn .x / D f1 .x / C    C fn .x / converges
n!1
Similarly, a series of functions 6 fk .x / converges pointwise on E to a function S : E

pointwise on E to S .x /.

D .1 x 2 /n convergenpointwise on [0; 1] to some function S .x /:


0 if 0 < x  1
Solution. For every x 2 [0; 1], lim Sn .x / D lim .1 x 2 /n D
. So Sn .x / converges pointwise
n!1
n
!1
1 if x D 0
n 0 if 0 < x  1
on [0; 1] to the function S .x / D
.
1 if x D 0

Examples. (1) Show that the sequence Sn .x /

1
k D1

(2) Show that the series 6 ekx cos x converge pointwise on .

Solution. Recall the geometric series 6 r k converges (to


k D1

1; 0/; but not on .


r

) if and only if r

1
if and only if x < 0: So 6 ekx cos x converges pointwise on .
k D1
diverges when x D 0; it does not converge pointwise on . 1; 1]:
Definition. A power series is a function f .x / of the form a0 C a1 .x

1; 1]:

2.

1; 0/ to S .x / D
c / C a2 . x

1; 1/: Now r

D ex 2 .

1; 1/
1
e cos x
. Since 6 ekx cos x
1 ex
k D1
x

c/2 C : : :

D k6D0 ak .x

c/k ; where

c, a0 , a1 , a2 , : : : are numbers and c is called the center of the power series. Its domain (of convergence) is the set of
numbers, where the power series converges pointwise.
Power series are very important. For example, mathematical tables were built by using series such as
2

ex

D 1 C x C x2! C x3! C    D k6D0 xk!

cos x

D1

sin x

Dx

x2
2!
x3
3!

C x4!
5

C x5!

x6
6!
x7
7!

for

k 2k

k 2k C1

C    D k6D0 . .12k/ /x!

1 < x < C1
for

C    D k6D0 . .2k1/Cx 1/!

1 < x < C1
1 < x < C1:

for

Remarks. These series also converge for every complex numbers x by ratio test. If wet set x D i ; then we can check
that ei  D cos  C i sin : In particular, ei  D 1: Then ei  C 1 D 0; which is called Eulers formula. It is the most
beautiful formula in mathematics because it links up the five most important constants 1; 0; ; i; e in one equation.
Domain Theorem for Power Series. The domain of a power series f .x /
81

D k6D0 ak .x

c/k is a nonempty interval

with c as midpoint. (The half-length of the interval is called the radius of convergence of the power series, commonly
denoted by R.) We have
1
p :
RD
limsup k jak j
k !1

R ; c C R /; but it may or may not converge at

In fact, the power series converges absolutely on the open interval .c


either endpoint.

Proof. This follows by applying the strong form of the root test to 6 ak .x
k D0

c/k :

Remarks. Although we have a formula for the radius of convergence, it is often easier to compute the radius using the
ratio test (especially when the coefficients involve factorials).

D k6D0 xk! .
x kC1 k!
D lim jx j D 0 < 1; the series converges for every x : The
Solution. By the ratio test, since lim
k !1 .k C 1/! x k
k !1 k C 1
domain of f .x / is R D . 1; C1/: Here c D 0 and R D 1:

Examples. (1) Find the domain of f .x /

(2) Find the domain of f .x /

D k61D0 k!.x

/k :

.k C 1/!.x /kC1
0


.k C 1/jx  j D
Solution. By the ratio test, since lim
D klim
1
k !1
!1
k!.x /k
converges only for x D : The domain of f .x / is f g D [;  ]: Here c D  and R D 0:

(3) Find the domain of f .x /

D k6D1 .

1/k .x
k

25/k

if x
if x

D ; the series
6D ;

. 1/kC1 .x 25/kC1

k

D lim jx 25j k D jx 25j;
Solution. By the ratio test, since lim
k !1
kC1
. 1/k .x 25/k k!1
kC1
the series converges absolutely for jx 25j < 1 (i.e. 24 < x < 26) and diverges for jx 25j > 1 (i.e. x < 24
1 1
or x > 26/: When x D 24, the series becomes 6 which diverges by p-test. When x D 26, the series becomes
k D1 k

1 . 1 /k
which converges by the alternating series test. The domain of f .x / is the interval .24; 26]: Here c D 25
6
k
k D1
and R

D 1.

Question. How do we expand functions into series?


If f .x / D a0 C a1 x C a2 x 2 C a3 x 3 C    C an x n C    ; then formal computation yields a0 D f .0/; a1 D
0
f .0/; a2 D f 00.0/=2; a3 D f 000 .0/=6; : : : ; an D f .n/ .0/= n!; : : : : In case the function is not defined at 0 (for example,
log x), we can try to expand the function at any point c; where f is infinitely differentiable. Similarly, we can get the
coefficients an in terms of f .n/ .c/: So for an arbitrary (infinitely differentiable) function f .x / defined about c; we can
attempt to form the power series
f .c/ C f 0 .c/.x

c/ C

f 00 .c/
.x
2

c/2 C

f 000 .c/
.x
6

c/3 C    C

f .n/ .c/
.x
n!

c /n

which is called the Taylor series of f about c. (For the Taylor series of f about c
Maclaurin series of f :)

1 f .k/ .c/

C    D k6D0

k!

.x

c/k ;

D 0; it is traditionally called the

Remarks. Unfortunately, the Taylorseries of a function does not always equal to the function away from the center.
1= x 2
if x 6D 0 can be shown to be infinitely differentiable and f .n/ .0/ D 0
For example, the function f .x / D e
0
if x D 0
82

for every n

2 N: So the Taylor series of

D 0 is k6D0 0x k D 0; i.e.
f .x / only at the center c D 0:

f .x / about c

Therefore, the Taylor series of f .x / equals

the Taylor series is the zero function.

Nevertheless, the following theorems provide sufficient conditions for many common functions to equal their
Taylor series.

! R is infinitely differentiable , c 2.k/.a ; b/ and there are constants M ; > 0


1 f .c/
for every x 2 .a ; b/, n 2 N, then f .x / D 6
.x c/k for every x 2 .a ; b/:
k!
k D0

Taylor Series Theorem. If f : .a ; b/

such that j

j  M

f .n/ .x /

Proof. Apply Taylors theorem. Note j Rn .x /j 


yn
converges. By term test, lim
n!1 n!

Mn
n!

jx

1 yn
nD0 n!

cjn

y
D n!
; where y D M jx

cj: By example (1), 6

D 0: By sandwich theorem, nlim


R . x / D 0:
!1 n
 k
. 1/ cos x if n D 2k C 1 . So j f .n/ .x /j  1 D 1  1n for every x 2 R:
Examples.(1) For f .x / D sin x, f .n/ .x / D
. 1/k sin x if n D 2k
1 . 1/k x 2kC1
for all x 2 R.
Taking c D 0; by the Taylor series theorem, sin x D 6
k D0 .2k C 1/!


Remarks. For 0  x  ; j R18 .x /j


2
sin x to 10 decimal places.

18

18

< 6  10

13

: So x

x3
3!

17

x
C    C 17!
can be used to compute

D ex : So j f .n/ .x /j D ex  ew D ew  1n for every x 2 . w; w/: Taking


1 xk
1 xk
c D 0; by the Taylor series theorem, e x D 6
for all x 2 . w; w/. Since w is arbitrary, e x D 6
for all
k D0 k!
k D0 k!
x 2 R.

(2) For f .x /

D ex on .

xj
 j18!
 .=18!2/

w; w/, f .n/ .x /

Bernsteins Theorem. Suppose f and all its derivatives are nonnegative on a closed and bounded interval [c; c C r]:
1 f . k / .c /
For x 2 [c; c C r /; we have f .x / D 6
. x c /k :
k D0
k!
Proof. By substitution, we may assume c
formula with integral remainder Rn .x / D

0: The
Z case x
x

.x

t/

n 1

0 is clear. Suppose x
f .n / .t /

.0; r /: Apply Taylors

dt : The conclusion follows if we can show

.n 1/! 0
 x n
0  Rn .x / 
f .r /; since the expression on the right has limit 0 as n ! 1: Substituting t D x .1 u /; we get
r Z
1
n


x
Rn .x / D
u n 1 f .n/ x .1 u / du : Since f .nC1/  0 implies f .n/ increasing, we get f .n/ x .1 u / 
.n 1/! 0

Rn .x /
Rn .r /
f .r /
Rn .x /
.
n/
f
r .1 u / for all u 2 [0; 1]: Hence
is increasing. So 0 


; where Rn .r /  f .r /
n
n
n
x
x
r
rn
.
k
/
n 1 f
.0/ k
is because f .r / D 6
r C Rn .r /: Multiplying all terms of the inequalities by x n ; we get the result.
k!
k D0
Next we will extend the binomial theorem to cover all real exponents.
Binomial Theorem. For a

.1 C x /a

2 R; if

1 < x < 1; then

D 1 C ax C a .a2!

1/

x2 C

a .a

1/.a
3!

2/

1 a .a
k D1

x3 C    D 1 C 6

1/    .a
k!

k C 1/

xk:

D .1 C x /a ; thenZ xf .n/ .x / D a .a 1/    .a n C 1/.1 C x /a n : ByZ Taylors


formula with inte1
a .a 1/    .a n C 1/ x  x t n 1
n 1 .n/
gral remainder, Rn .x / D
.x t / f .t / dt D
.1 C t /a 1 dt :
.n 1/!
.n 1 /!
1Ct
Proof. Let f .x /

83

For x

2.

1; 1/; the function g .t / D

x t
1 x
has derivative g 0.t / D
< 0: On the closed interval with end1Ct
.1 C t /2
Z

x

 jg .0/j D jx j: Let k D .1 C t /a 1 dt ; then
0
1
j
a .a 1/    .a n C 1/kx n 1 j
bnC1
j
a n jjx j
j Rn .x /j 
: Since lim
D
lim
D
j
x j < 1; by ratio test, 6 bn conn!1 bn
n!1
.n 1/!
n
nD1
|
{z
}
call this b
verges. By term test, lim bn D 0: Then lim Rn .x / D 0 by the sandwich theorem.
n!1
n!1
points 0 and x ; g takes values between 0 and x : So there jg .t /j

Here are a few more common Taylor series. (Note the series only converge on a small interval to the functions.)
1
They are obtained by taking the cases a D 1; a D 1 (with x replaced by x 2 ) and a D
(with x replaced by x 2 )
2
of the binomial theorem and integrating term-by-term. These can be justified by applying Abels limit theorem and the
integration theorem for uniformly convergence to be covered soon. Alternatively, they can be justified by integrating
the Taylors formula with integral remainder for the corresponding binomial functions and using the bn estimates above
to squeeze the new remainder to 0.
x2
2

ln.1 C x / D x
tan
sin

x3
3

Dx

x4
4

C x3
5

C x5

x7
7

C    D k61D1 .

1 /k
k

1 k

1/k x 2kC1
2k C 1

C    D k6D0 .

for

1<x

1

for

1x

1

3
5
1
1/ x 2kC1
D x C 12 x3 C 12  34 x5 C    D x C k6D1 1 23 456  .2k
 .2k/ 2k C 1

1x

for

1

Power series are not only good for computing the values of functions, but they can be used to expand other
functions into series for limit, differentiation and integration purposes. Here we will practice expanding functions into
series of functions first.
Examples. (1) For x

(2) For x

2.

2 .0; 1/; x x D ex ln x D k6D0 .x lnk!x /

1; 1/; we have 1 <

(3) For x > 0; we have 0 < e


a

D x 3e

;r

De

; we have

x 2 < 1 and so

1
x2

D .1 C .

x 2 // 1=2

< 1: Using the geometric series formula


x3
ex

x
D
1
1

e x
e x

D x 3e x C x 3e

2x

D 1 C 12 x 2 C 38 x 4 C 165 x 6 C    :

a
1

C x 3e

3x

D a C ar C ar 2 C ar 3 C    with
1

C    D k6D1 x 3 e

kx

Uniform Convergence
Power series for sine and cosine allowed us to compute values of these functions. As another application of power
1 1 1
C
C    ; which
series, we will find the sum of certain convergent series of numbers. Consider the series 1
2 3 4
2
3
x
x
x4
converges by the alternating series test. Here is a clever way to find the sum. Let f .x / D x
C
C   :
2
3
4
(We can check that this power series converges pointwise on . 1; 1]:) The sum of the alternating series above is f .1/:
To find this value, we write f .1/
get f 0 .x /

D1

C x2

x3 C x4

f .1/

f .0 /

Z1

f 0 .x / d x : Now, one may try to differentiate term-by-term to

   D 1 C1 x : So f .1/ D
84

Z1
0

f 0 .x / d x

Z1
0

1
dx
1Cx

D ln 2:

Lets look at the argument more carefully. First, we see that every term of f .x / is differentiable, but does this
imply f .x / is differentiable? Next suppose f .x / is differentable. Can we simply differentiate term-by-term? Note
1 . 1/k 1 x k
d 1 . 1/k 1 x k
f .x / D 6
and f 0 .x / D
; but the assumed derivative is
6
k
d x kD1
k
k D1
1

Cx

1
   D k6D1. 1/k 1 x k

.

1
D k6D1 ddx

1/k 1 x k
:
k

If f .x / has finitely many terms, term-by-term differentiation is justified by mathematical induction. However, for
infinitely many terms, term-by-term differentiation needs to be justified.
Questions. (1) If Sn .x / converges pointwise on an interval to S .x / and each Sn .x / is continuous (or differentiable), is
it necessary that S .x / will also be continuous (or differentiable)?
(2) If gk .x / are continuous, differentiable, integrable on an interval [a ; b] for k
n

lim 6 gk .x /

x ! x0 k D1

g .x /;
D k6D1 xlim
!x k
n

d n
6 gk . x /
d x k D1

Zb

D k6D1 ddx gk .x /;
n

D 1; 2; 3; : : :, then
n

6 gk . x / d x

a k D1

D k6D1
n

Zb

gk . x / d x

for every positive integer n : If n is replaced by the symbol 1; will the equations still be true always?

0 if 0  x < 1
Examples.(1) Continuous functions Sn .x / D x n converge pointwise on [0; 1] to S .x / D
. However,
1 if x D 1
S .x / is not continuous on [0; 1]: This gives a negative answer to the first question above.
Let g1 .x /

D S1.x / and gk .x / D Sk .x /

1
x !1 k D1

lim 6 gk .x /

Sk 1 .x / for k > 1: Then

D xlim
lim S .x / D lim S .x / D 0; but 6 lim gk .x / D lim lim Sn .x / D lim 1 D 1:
n!1 x !1
n!1
!1 n!1 n
x !1
k D1 x !1

This gives a negative answer to the second question above.

1
x2
k D0 .1 C x 2 /k

(2) Consider the function S .x / D 6

D x 2 C 1 Cx x 2 C .1 Cx x 2 /2 C : : : : Now S .0/ D 0: For x 6D 0,




1
1
1
2
S .x / D x 1 C
C .1 C x 2 /2 C .1 C x 2 /3 C : : : D x 2 1 1 D 1 C x 2 :
1 C x2
1 1C x
0
1
x2
if x D 0
Thus S .x / D 6
converges pointwise on R to
. Although every term gk .x / D
2
2
k
1
C
x
if x 6D 0
.
1
C
x
/
k D0
2
3
n
x
2x 2.k 1/x
is differentiable on R with gk0 .x / D
so that the partial sum Sn .x / D 6 gk .x / is
2
k
2
k
1
.1 C x /
.1 C x /
k D0
differentiable, but S .x / D lim Sn .x / is not differentiable at x D 0: This gives a negative answer to the first
n!1
2

question above.

Note gk0 .0/

D 0 and Sn0 .0/ D 0: So at x D 0;

d 1
6 gk . x /
dx kD0

d
d
d
D ddx nlim
Sn . x / D
S .x / does not exist, but 6
gk .x / D lim
Sn . x / D 0:
!1
n!1 d x
dx
k D0 d x

This gives a negative answer to the second question above.


(3) Since Q \ [0; 1] is countable, its elements can ben arranged into a sequence r1 ; r2 ; r3 ; : : : ; where each element
1 if x D r1 ; r2 ; : : : ; rn
appears exactly one time. The functions Sn .x / D
are integrable on [0; 1] and converge
0 otherwise
85

1 if x 2 Q \ [0; 1]
pointwise on [0; 1] to S .x / D
; which is not integrable on [0; 1]: Again this gives a negative
0 otherwise
answer to the first question above.
Next, the continuous (hence integrable) functions

8 2
<n x
2
Sn . x / D
: 0 n x C 2n

n
y = Sn (x)

converge pointwise on [0; 1] to S .x /


1/ n

2/ n

2
for x 2 .0; 1]; Sn .x / D 0 when x >
n
Sk .x / Sk 1 .x / for k > 1; we have

1
6
k D1

Z1

gk . x / d x

D nlim
!1

Z1

, n > 2x

D nlim
1 D 1; but
!1

Sn . x / d x

D 0 because nlim
S .0/ D 0 and
!1 n

so that lim Sn .x /
n!1

Z1
0

if 0  x < 1= n
if 1= n  x < 2= n
if 2= n  x  1

D 0: Define g1.x / D S1.x / and gk .x / D

1
6 gk . x / d x
k D1

Z1

lim Sn .x / d x

0 n!1

Z1
0 dx
0

D 0:

Again this give a negative answer to the second question above.


Now recall that continuity, differentiability and integration are concepts involving limits:
g .x 0 C h /
h !0
h

g .x 0 / D lim g .x /;

g 0.x 0 / D lim

x ! x0

g .x 0 /

Zb

g .x / d x

D nlim
g
!1 k6
D1
n

aC

a
n

b

a
n

1
k D1

In general, for S .x / D lim Sn .x / or g .x / D 6 gk .x /, as the above examples showed, we have


n!1

lim lim Sn .x /

6D

d
lim Sn .x /
d x n!1

6D

lim Sn .x / d x

6D

x ! x0 n!1

Zb

a n!1

lim lim Sn .x /;

n!1 x ! x0

lim

n!1

Zb

lim

n!1 a

lim 6 gk .x /

6D

d 1
6 gk . x /
d x kD1

6D

x ! x0 k D1

d
Sn .x /;
dx

Zb

Sn . x / d x ;

1
6 gk . x / d x

a k D1

6D

1
gk . x / I
6 xlim
k D1 ! x 0
1 d
gk . x / I
6
k D1 d x
Z
1 b
gk . x / d x :
6
k D1 a

Question. Are there special conditions which ensure an interchange of limit operations is correct?
Notations. For a function f : E
called the sup-norm of f on E :

! R; the quantity supfj f .x /j : x 2 E g is often denoted by k f k E or k f k1 and is


Definitions. A sequence of functions Sn : E
on E to a function S: E ! R iff
lim k Sn

S(x) +

Sn (x)
E

D nlim
.supfj Sn .x / S .x /j : x 2 E g/ D 0
!1
(i.e. for every " > 0, there exists K 2 N (K depends only on ") such
that n  K ) j Sn .x / S .x /j < " for every x 2 E :)
1
n!1

S (x)
S(x)
)

! R converges uniformly

SkE

A series 6 gk .x / converges uniformly on E to S .x / iff its partial


k D1

sum sequence Sn .x / D 6 gk .x / converges uniformly on E to S .x /.


n

k D1

Theorem. If Sn .x / converges uniformly on E to S .x /, then Sn .x / converges pointwise on E to S .x /. Similarly, if

1
1
6 gk .x / converges uniformly on E ; then 6 gk .x / converges pointwise on E :
k D1

k D1

86

S k E : So lim k Sn

Proof. For all x 2 E ; j Sn .x / S .x /j  k Sn


x 2 E ; which implies lim Sn .x / D S .x /:

j S .x /
D 0 implies nlim
!1 n

SkE

n!1

S . x /j

D 0 for every

n!1

Uniform convergence is a sufficient condition that allows interchange of operations, as will be seen in the following
theorems. The proofs will be given in Appendix 4.
Continuity Theorem for Uniform Convergence. If every Sn .x / is continuous at c 2 E and Sn .x / converges uniformly
on E to S .x /; then S .x / is continuous at c; i.e.
lim lim Sn .x / D lim S .x / D S .c/

x !c n!1

x !c

D nlim
S .c/ D lim lim Sn .x /:
!1 n
n!1 x !c

Similarly, if every gk .x / is continuous at c 2 E and 6 gk .x / converges uniformly on E ; then 6 gk .x / is continuous


k D1
k D1
at c; i.e.
1
1
1
lim 6 gk .x / D 6 gk .c/ D 6 lim gk .x /:
x !c k D1

k D1 x !c

k D1

Integration Theorem for Uniform Convergence. If every Sn .x / is integrable on a bounded interval [a ; b] and Sn .x /
converges uniformly on [a ; b] to S .x /; then S .x / is also integrable on [a ; b] and

Z b

lim Sn .x / d x

n!1

Zb

S .x / d x

1
k D1

Zb

D nlim
!1

Sn . x / d x :

1
k D1

Similarly, if every gk .x / is integrable on [a ; b] and 6 gk .x / converges uniformly on [a ; b]; then 6 gk .x / is integrable

Z b1

on [a ; b] and

6 gk . x / d x
k D1

D k61D1

Zb

gk . x / d x :

At this point we expect a similar theorem may hold for differentiation. However, this is not the case as the
following example shows.
Example. Consider the function Sn .x / which equals jx j outside In

1 1i
; and on In ; we take the graph of Sn .x /
n n

2
2
to be the quarter circle with center at 0;
and radius
: Then the functions Sn .x / are differentiable on R and
np
n
2
2
; which has limit 0. So the functions Sn .x / converge uniformly on R to
supfj Sn .x / jx jj : x 2 Rg D Sn .0/ D
n
jx j; which is not differentiable.
Even if a sequence of differentiable functions converge uniformly on R to a differentiable function, limit and
x
differentiation may not be interchanged! For example, the functions Sn .x / D
converge uniformly on R to
1 C nx 2
2
1 nx
1
1
S .x / D 0 because Sn0 .x / S 0 .x / D
implies supfj Sn .x / S .x /j : x 2 Rg D Sn . p / D p ; which has
2
2
.1 C nx /
n
2 n
d
d
d
limit 0. However, at x D 0;
lim Sn .x / D
0 D 0; but at x D 0; lim
Sn .x / D lim 1 D 1:
n!1 d x
n!1
d x n!1
dx
Differentiation Theorem for Uniform Convergence. Let a ; b 2 R with a < b: If every Sn .x / is differentiable on
.a ; b/; Sn0 .x / converges uniformly on .a ; b/ to T .x / and Sn .x 0 / converges for at least one x 0 2 .a ; b/; then Sn .x /
converges uniformly on .a ; b/ to a differentiable function S .x / with
d
dx

lim Sn .x /

n!1

D ddx S .x / D T .x / D nlim
!1
87

d

dx

Sn . x / :

k D1

k D1

Similarly, if every gk .x / is differentiable on .a ; b/; 6 gk0 .x / converges uniformly on .a ; b/ and 6 gk .x 0 / converges

1
k D1

for at least one x 0 2 .a ; b/; then 6 gk .x / converges uniformly on .a ; b/ to a differentiable function with
d 1
6 gk . x /
d x kD1

D k6D1 ddx gk .x /:

Remarks. For unbounded intervals, in the differentiation theorem for uniform convergence, Sn .x / may not converge
uniformly on the interval to the pointwise limit function S .x /: To see this, consider Sn : R ! Rdefined by Sn .x / D x = n :
Then Sn0 .x / D 1= n converges uniformly on R D . 1; C1/ to T .x / D 0 and Sn .0/ D 0 converges to 0. However,
Sn .x / converges only pointwise on R to S .x / D 0; but not uniformly because k Sn S k. 1;C1/ D C1: Nevertheless
the other conclusions, namely the pointwise limit function S .x / exists, is differentiable and differentiation may be
done term-by-term at every point on the interval are true. This is because for every w on the unbounded interval we
can restrict the unbounded interval to an open bounded interval containing w and x 0 to get these conclusions.
Because of the above theorems, in order to interchange operations, we now need to know how to check uniform
convergence for sequence and series of functions. For sequences of functions, we will introduce the L-test below, and
for series, the M-test.
Theorem (L-test). If Sn .x / converges pointwise on E to S .x / and there are constants L n such that j Sn .x /
for every x 2 E and lim L n D 0; then Sn .x / converges uniformly on E to S .x /:
n!1

Proof. Since j Sn .x / S .x /j  L n for every x 2 E ; so k Sn


lim L n D 0; by sandwich theorem, lim k Sn S k E D 0:
n!1

SkE

D supfj Sn .x /

S . x /j : x

S .x /j

 Ln

2 E g  L n : Since

n!1

Examples. (1) Show Sn .x /

D sinpnx
converges uniformly on R.
n

Solution. Note that the pointwise limit function on R is S .x /

j Sn . x /

S .x /j D j Sn .x /j

(2) Show Sn .x /

De

lim

n!1

sin nx
pn

0. For every x

R;

 p1n and nlim


p1 D 0: By L-test, Sn .x / converges uniformly on R to S .x /:
!1 n

cos2 .1=x /= n converges uniformly on .0; 1/:

Solution. The pointwise limit function on .0; 1/ is S .x /

2
lim e cos .1=x /= n

n!1

D 1: For w 

0; by the

mean value theorem, je


1j D j. e /.w 0/j  jwj for some c between w and 0: For every x 2 .0; 1/;
2

2
L D 0; by the L-test, Sn .x / conj Sn .x / S .x /j D je cos .1=x /= n 1j  cos .n1=x /  L n D n1 : Since nlim
!1 n
verges uniformly on .0; 1/:
w

(3) Show Sn .x /

D x n converges uniformly on [0; t ] for t < 1, but Sn .x / does not converge uniformly on [0; 1].
Solution. The pointwise limit function on [0; t ] is S .x / D 0. For x 2 [0; t ];
j Sn .x / S .x /j  L n D t n : Now nlim
L D lim t n D 0 because t < 1: By the
!1 n n!1
(1, 1)

L-test, Sn .x / converges uniformly on [0; t ]:


Next suppose Sn .x / converges uniformly on [0; 1]: Since every Sn .x / is continuous on [0; 1]; by the continuity theorem for uniform convergence, the limit
function S .x / will
on [0; 1]: However the pointwise limit funcn 0 alsoif xbe6Dcontinuous
1
tion is S .x / D
; which is not continuous at x D 1; a contradiction.
1 if x D 1
So Sn .x / does not converge uniformly on [0; 1]:
88

D 1, 2, 3, : : :, there are constants Mk such that jgk .x /j  Mk for every x 2 E


1
1
and 6 Mk converges, then 6 gk .x / converges uniformly on E.
k D1
k D1
n
1
Proof. The n-th partial sum is Sn .x / D 6 gk .x / and the sum is S .x / D 6 gk .x /; which converges absolutely for
k D1
k D1
1
1
1
every x 2 E by comparing with 6 Mk : We have j Sn .x / S .x /j  6 jgk .x /j  6 Mk D L n for every x 2 E :
k D1
k DnC1
k DnC1

1
n
1
1
Since 6 Mk converges, so lim L n D lim 6 Mk
6 Mk D 0: By L-test, 6 gk .x / converges uniformly on E :
n!1
n!1 k D1
k D1
k D1
k D1
Theorem (Weierstrass M-test). If for k

1 sin kx
converges uniformly on R since for all x
k2
k D1

Examples. (1) 6

1 ln x
(2) Show 6
x
k D1

k

converges uniformly on [1; 1/.

ln x
D max ln x D 1 .
By calculus, max
x 2[1;1/
x x 2[1;1/ x
e

Solution.



1
2 R; sink2kx  k12 D Mk and k6D1 k12 converges.

 k  k
ln x  1 D Mk for every x 2 [1; 1/.
So
x e

Since 6 Mk converges by the geometric series test, Weierstrass M-test gives us the result.
k D1

Z1
(3) (Here is another application of power series!) Find

x x d x to five decimal places.

2 .0; 1]; x x D ex ln x : (As x ! 0C ; x ln x D .ln x /= 1x ! 0 by lHopitals rule. So at x D 0; we set


1 wk
for every w 2 R; we
x ln x D 0 and x x D 1 to be continuous. Then x x is integrable on [0; 1]:) Since ew D 6
k D0 k!
.x ln x /k
1 .x ln x /k
1
1
have x x D e x ln x D 6
: By calculus, max jx ln x j D max . x ln x / D : So
 Mk D k

Solution. For x

1
and 6 Mk
k D0

k D0
1 1
D k6D0 k!ek

x 2[0;1]

x 2[0;1]
e
k!
k!e
k
1
.
x
ln
x
/
D e1=e : By Weierstrass M-test, k6D0 k! converges uniformly on [0; 1] to x x : By the

k!

integration theorem for uniform convergence,

Z1

x dx
0

Z1

Z1
0

1 .x ln x /k
dx
6
k!
k D0

where we substituted u
find

x dx
0

 0:78343:

1
D k6D0

Z 1 .x ln x /k
0

k!

1
. 1/k
dx D 6
k D0 k!.k C 1/k C1

Z1

uk e

| 0 {z

Dk!

du

D k6D0 .k .C 11//kC1 ;
k

.k C 1/ ln x and integrated by parts k times in the computations. Taking 10 terms, we

(4) In this example, we will show the integration theorem


Riemann integrals. Let
(0
Sn .x / D .x n /= n 2
.x 3n /= n 2

for uniform convergence may not hold for improper


if x
if x
if x

2 [0; n / [ [3n ; 1/
2 [n ; 2n /
:
2 [2n ; 3n /

The graph of Sn .x / on [0; 1/ is consisted of the segments from .0; 0/ to .n ; 0/, from .n ; 0/ to .2n ; 1= n /; from
1
.2n ; 1= n / to .3n ; 0/ and the part of the positive x-axis starting at .3n ; 0/: Since 0  Sn .x /  for every x  0;
n
1
so the pointwise limit of Sn .x / is S .x / D lim Sn .x / D 0: Since lim k Sn S k[0;1/ D lim D 0; by definition
n!1
n!1
n!1 n
Sn .x / converges uniformly on [0; 1/ to S .x / D 0: Now

Z1
0

lim Sn .x / d x

n!1

Z1
0

S .x / d x

D 0;
89

Z1

but lim

n!1 0

Sn . x / d x

D nlim
1 D 1:
!1

Z1
(5) Find

x3

ex 1
Solution. (Note as x
0

2 .0; 1/; 0 < e x


x 2 .0; 1/. We have
Z1

For x

x3
ex

By calculus, for every x

dx

D wlim
!1

2 [0; w]; jx 3e kx j 
1
k D1

by p-test, Weierstrass M-test implies 6 x 2 e


for uniform convergence, we should get

Z1
0

Next substitute w
limC gk .t / D

Z1

4

:)
15
0C; the integrand tends to 0. So at 0, the integrand is set to 0 to be continuous.)
1
x3
x 3e x
< 1: So x
D
D
x 3 e x C x 3 e 2x C x 3 e 3x C    D 6 x 3 e kx for all
x
e
1
1 e
k D1

d x to five decimal places. (The exact answer is known to be

x3
ex

dx

D wlim
!1

Zw
0

x3
ex

 3 3
k

kx

3
e . k /k

dx

D wlim
!1

Zw

1
6 x 3e
k D1

kx

dx:

1 1
D e27
D Mk : Since k61D1 Mk D 27
6 3
3 k3
3
e kD1 k

converges

converges uniformly on [0; w]: Applying the integration theorem

Zw
0

1
6 x 3e
k D1

kx

1= t for t > 0; and define gk .t / D

dx

D wlim
!1 k6
D1

Z 1= t

x 3e

kx

Zw

x 3e

kx

dx:

d x : The value gk .0/ is defined to be

D k64 ; obtained by integrating by parts three times. Observe k6D1 gk .t / converges


t !0
0
uniformly on [0; 1/: (We check the M-test. Since x 3 e kx  0;
Z 1= t
Z1
6
jgk .t /j D
x 3 e kx dx 
x 3 e kx d x D 4
for every t 2 [0; 1/:
k
0
0
Z 1 x3
1 6
Finally 6

k D1

1
w!1 k D1
lim 6

k4

kx

dx

converges by p-test.) By the continuity theorem for uniform convergence,

Zw
0

x 3e

x 3e

kx

dx

D tlim
6 g .t / D 6 limC gk .t / D 6
!0C kD1 k
k D1 t !0
k D1

Z1

x 3e

kx

dx

ex

d x equals

D k6D1 k64 D 15  6:49394:


4

Other than the L-test and the M-test, there are a few useful facts for checking uniform convergence.
Facts. (1) (Subset Property) If Sn .x / converges uniformly on A to S .x / and E
on E to S .x / because 0  k Sn S k E  k Sn S k A ! 0:

 A; then Sn .x / converges uniformly

(2) (Union Property) If Sn .x / converges uniformly on A and B to S .x /; then Sn .x / converges uniformly on A [ B to


S .x / because 0  k Sn S k A[ B  k Sn S k A C k Sn S k B ! 0:
(3) (Bounded Multiplier Property) If Sn .x / converges uniformly on E to S .x / and g .x / is a bounded function on E ;
then g .x / Sn .x / converges uniformly on E to g .x / S .x / because 0  kg Sn g S k E  kg k E k Sn S k E ! 0:

(4) (Substitution Property) If Sn .x / converges uniformly on E to S .x / and h : E 0 ! E ; then Sn h .x / converges


uniformly on E 0 to S h .x / because E 0 D fh .x / : x 2 E 0 g  E implies 0  k Sn  h S  h k E0 D k Sn S k E0 
k Sn S k E ! 0: Briefly, this means we can do substitution for uniform convergent sequences and series and still
remain uniformly convergent.
Question. Are there special cases where pointwise convergence of nice functions implies uniform convergence?
For closed and bounded intervals, the following two theorems provide such special situations.
90

Dinis Theorem. If continuous functions Sn converge pointwise on a closed and bounded interval [a ; b] to a continuous
function S and for every x 2 [a ; b]; Sn .x / is an increasing sequence of real numbers, then the convergence is uniform
on [a ; b]:

1
k D1

For series, if each term gk .x / is nonnegative, continuous on [a ; b] and 6 gk .x / converges pointwise on [a ; b]


to a continuous function S .x /; then the convergence is uniform on [a ; b]: The statement is also true if increasing is
replaced by decreasing or nonnegative is replaced by nonpositive.
Proof. The continuous function fn D S Sn converges pointwise on [a ; b] to 0 and for every x 2 [a ; b]; fn .x /
decreases to 0. Since k Sn S k[a ;b] D k fn k[a ;b]; it suffices to show fn converges uniformly on [a ; b] to 0. Suppose
this is false, then there is " > 0 such that for every K ; there is n  K and x n 2 [a ; b] with j fn .x n /j  ": By
Bolzano-Weierstrass, x n has a subsequence x nk converging to some w 2 [a ; b]:
Now for every positive integer n ; consider those n k  n : Then fn .x nk /  fnk .x nk / D j fnk .x nk /j  ": Since fn is
continuous and x nk ! w; by the sequential continuity theorem, fn .w/ D lim fn .x nk /  ": Then 0 D lim fn .w/  ";
n!1

k !1

a contradiction to fn .w/ converging to 0 as n tends to infinity.

For the second statement, let Sn D g1 C    C gn and apply the first statement. For the third statement, multiply
every function by 1 and apply the first two statements.

Remarks. Dinis Theorem is false on intervals without an endpoint. For example, Sn .x / D x n converges pointwise
on [0; 1/ to S .x / D 0 and for every x 2 [0; 1/; Sn .x / is increasing, but lim k Sn S k[0;1/ D 1 so that the convergence
n!1

is not uniform on [0; 1/:

1
k D0

Abels Limit Theorem. If a power series 6 ak .x

1
k D0

to S .x /, then 6 ak .x

c/k converges pointwise on a closed and bounded interval [u ; v ]

c/k converges uniformly on [u ; v ] to S .x /. (By the continuity theorem for uniform convergence,

1
k D0

S .x / is continuous on [u ; v ]: Hence, 6 ak .u

D S .u / D xlim
S .x / and 6 ak .v
!uC
k D0

c/k

c/k

D S .v/ D xlim
!v

S .x /:/

1
k D0

Remarks. Abels Limit Theorem is false on intervals without an endpoint. For example, 6 x k converges pointwise
on [0; 1/ to S .x /

D 1 1 x ; but k Sn

x nC1
: x 2 [0; 1/g D 1 for every n so that the convergence is

S k[0;1/ D supf

1 x
not uniform on [0; 1/:
For a proof of Abels limit theorem, please see Appendix 4. This theorem has a useful consequence in differentiating power series.

1
k D0

Theorem (Differentiation of Power Series). If a power series 6 ak .x

c/k converges pointwise on .c

to S .x / where the radius of convergence R > 0; then S .x / is differentiable there and S 0.x /
x

2 .c

R ; c C R /.

Proof. The radius of convergence of S .x /

1
k D1

series. For the derived series 6 kak .x

1
limsup
k !1

jkak j

D k6D0 ak .x

c/k

c/k is R

1
limsup
k !1

p
k

jak j

D k6D1 kak .x

R; c C R/
c/k

for

by the domain theorem for power

, its radius of convergence is


1

limsup
k !1

p
k

ja k j

D R;

since

p
k

! 1 as k ! 1:

So both series converges pointwise on the interval .c R ; c C R /. For x 2 .c R ; c C R /; take " > 0 so that
[x "; x C "]  .c R ; c C R /: Abels limit theorem implies both series converge uniformly on [x "; x C "] (and
hence also on .x "; x C "/ by the subset property). Applying the differentiation theorem for uniform convergence
on .x

1
"; x C "/; we get S 0.x / D 6 kak .x
k D1

c /k 1 :

91

Examples. (1) (Here is yet another application of power series!) Find L


Solution. Recall ew

w3

x
D xlim
!0

L
1
Since
6

D 1 C w C w2! C    and sin w D w

C x C  D

3!

x 2 ex
sin.x 2 /
lim 6
:
x !0 x
ex 7 C 1
5

C    for 1 < w < C1: Hence,

.1 C x 5 C   / .x 2 x3! C   /
x 6 .1 C x 7 C   / C 1

1 6
1
x C x7 C   
C x C   :
6
6
D xlim
D
lim
6
7
!0 x
x !0 1
x C 
x C 

if x 6D 0 and 1 x C    D .x 6 e x C 1/=x 6
if x D 0
1
6

x 2 ex

sin.x 2 / =x 6
1=6

if x
if x

6D 0 ;
D0

both power series converge pointwise on [ 1; 1]; hence both are continuous at 0 by Abels limit theorem. So
1
1
1
lim . C x C   / D and lim .1 x C   / D 1: Therefore, L D :
x !0 6
x !0
6
6

(2) Show 1

1
2

C 13

1
4

C    D k6D1 .

1/k
k

D ln 2 and justify the steps clearly.

Solution. Step 1. The series converges by the alternating series test since 1 

1
2

1
 13     and klim
D 0:
!1 k

1 . 1/k 1 x k
: Then f .1/ converges by step 1. We have
k
k D1

Step 2. Define f .x / D 6

. 1/k x kC1
kC1 .
k !1
lim

k
1/k


k jx j
D jx j < 1
D klim
1xk
!1 k C 1

1 < x < 1:

if and only if

By the ratio test and step 1, we find that the domain of f is . 1; 1]: Applying differentiation of power series on
. 1; 1/;

1
f 0 .x / D 6 . 1/k 1 x k
k D1

x2
2

x4
4

D1

Cx

C  D 1 C x )
1

f .x /

f .0 / C

 ln.1 C x /
if 1 < x < 1
1 12 C 13 14 C : : : if x D 1.
Step 3. Using Abels limit theorem on [0; 1]; 1 12 C 13 14 C : : : D f .1/ D lim
x !1
So f .x / D x

C x3

x
0

f 0 .t / dt

D ln.1 C x /:

C D

D xlim
ln.1 C x / D ln 2:
!1
1
Remarks. (1) The series converges too slowly to ln 2. (For example, by the hundredth term, which is 100
D 0:01, it
still affects the second decimal place of ln 2.) To get ln 2, it is better that we consider for x 2 . 1; 1/;
1 C x 
 x2 x3 x4   x2 x3 x4   x3 x5 x7 
ln
D
ln.1C x / ln.1 x / D x
C
C  
x
   D 2 x C 3 C 5 C 7 C  
1 x
2 3 4
2 3 4
and set x

f .x /


 

3
5
7
D 13 to get ln 2 D ln 11 C ..11==33// D 2 13 C .1=33/ C .1=53/ C .1=73/ C    D 0:6931 : : : ; which is cor-

rect to four decimal places already by the fourth term.

(2) Differentiation of power series may be false if .c R ; c C R / is replaced by .c R ; c C R]; [c R ; c C R / or


1 . 1/k 1 x k
x2
x3
[c R ; c C R]: We saw 6
D
x
C
   converges pointwise on . 1; 1] to f .x / D ln.1 C x /;
k D1
k
2
3
1
but the derived series 6 . 1/k 1 x k 1 D 1 x C x 2    converges pointwise only on . 1; 1/ and diverges at x D 1

k D1
1
0
even though f .1/ D exist.

92

1 . 1/k 1 x k
converges pointwise on [0; 1] to ln.1 C x /: So by Abels limit
k
k D1
theorem, the convergence is uniform on [0; 1]: However, if you try to check this by Weierstrass M-test, then it will
1 1
1
1
lead to Mk D ; but 6 Mk D 6 diverges!
k
k D1
k D1 k
(3) Example (2) above showed that 6

1
2

(3) Find the value of 1

1
3

C 14 C 15

D k6D1.

1/kC1


1

1
2k

2k

1
7

C : : :.

 

1 1 1 1 1
1
C C
C : : : D 1 12
3 4 5 6 7
3
. 1/kC1
converges by the alternating series test.
.2k 1/.2k /
1
2

Solution. Step 1. The given series 1

1
6

D k6D1

1
4

 1
C 5

1
6

:::

1 . 1/kC1 x 2k
Step 2. Define f .x / D 6
: Then f .1/ D f . 1/ converges by step 1. We have
k D1 .2k
1/.2k /

. 1/kC2 x 2kC2 .2k 1/.2k /


.2k 1/.2k /x 2
D
lim

k
C
1
2k
k !1 .2k C 1/.2k C 2/ . 1/
k !1 .2k C 1/.2k C 2/
x
lim

D x2 < 1

1 < x < 1:

if and only if

By ratio test and step 1, it has domain [ 1; 1]: Applying differentiation of power series twice on . 1; 1/;
f 0 .x /

1/kC1 x 2k
2k 1

D k6D1 .

x3
3

Dx

f 0 .x /

f 0 .0/ C

D1

f 00 .x /

and

x2 C x4

Zx

D tan 1 x
Z
Zx
0
) f .x / D f .0/ C f .t / dt D tan 1 t dt D x tan

f 00.x /

:::

D 1 by the alternating series test.) So for

(Note the power series for f 0 .x / also converge at x

D 1 C1 x 2 )

C x5

::::

1 < x < 1, we have

f 00 .t / dt

0
x

1
ln.1 C x 2 /:
2

Step 3. Applying Abels limit theorem on [0; 1]; we get

1
2

1
3

C 14 C 15

1
3

C 15

1
7

C 19

1
6

1
7

C::: D

f .1/ D lim f .x /
x !1

D 4

: : : D f 0 .1/ D lim f 0 .x / D lim tan 1 x


x !1
x !1

1
ln 2
2

D tan

Remarks. This series converges too slowly. To compute  , one uses the relation tan. C /
x

D tan ; y D tan ; we get tan



4

D tan

So using tan

1 D tan

f 0 .x /

48 tan

48

3:1415926 : : : ;

1
18

1
18

1
2

C tan

1
3  183

x3
3

Dx

C 32 tan

C tan
1
3

D 2 tan
5

C x5

D C D tan
1

1
3

   for

C tan

1
7

 tan C tan 
1

D tan

tan tan

D : : : D 12 tan

1
18

1D


4


D 1tan tanC tan
: Taking
 x C y tan
1

C 8 tan

1
57

xy

: Then

5 tan

1
:
239

1 < x < 1 from above, we get

1
1
20 tan 1
57
239

1
1
:
:
:
C
32
5  185
57
1

1
3  573

93

C 5  1575

:::


20

1
239

1
3  2393

C:::

which is correct to seven decimal places with these eight terms.

Appendix 1: Space-Filling Curves


Do all curves have zero area? Our intuition will think so, but our intuition is wrong. In 1890, Peano showed
there is a continuous curve in R2 that passes through every point of the unit square [0; 1]  [0; 1]: We will present
I. J. Schoenbergs 1938 example of such a curve, which is easier to understand.

80
if t 2 [0; 1=3] [ [5=3; 2]
><
3t 1
if t 2 [1=3; 2=3]
On [0; 2]; define g .t / D
: Extend g to R by g .t C 2/ D g .t /: Next define
if t 2 [2=3; 4=3]
>: 1
3t C 5 if t 2 [4=3; 5=3]
g .3 j t / 1

1 g .32n 2 t /
1 g .32n 1 t /
;
y
.
t
/
D
:
Since
g
is
continuous
on
R
;
f .t / D x .t /; y .t / ; where x .t / D 6
2n  2n
6
2n
2n
nD1
n D1
1 1
on R and 6 n D 1; using the Weierstrass M-test and the continuity theorem, we see that x .t / and y .t / are continuous
nD1 2
on R: We will show S D f f .t / : t 2 [0; 1]g is the unit square [0; 1]  [0; 1]:
Since 0  x .t /; y .t /  1; S  [0; 1]  [0; 1]: Next, for .a ; b/ 2 [0; 1]  [0; 1]; write a and b in base 2 as
1 an
1 bn
a D .0:a1 a2 a3   /2 D 6 n and b D .0:b1 b2 b3   /3 D 6 n ; where an ; bn D 0 or 1. Let c D 2.0:a1 b1 a2 b2   /3
2
nD1
nD1 2
1 cn
1 1

D 2 n6D1 3n ; where c2m 1 D am and c2m D bm : Since 2 n6D1 3n D 1; so c 2 [0; 1]: We will show f .c/ D x .c/; y.c/ D
.a ; b/:

D cm for m D 1; 2; 3; : : : : (Then g .32n 2 c/ D c2n 1 D an ; g .32n 1c/ D c2n D bn and


m 1
1 ck
we will get x .c/ D a ; y .c/ D b:) Now 3m 1 c D 2.c1 c2    cm 1 :cm cm C1 cm C2   /3 D 2 6 ck 3m 1 k C 2 6 k m C1 :
k D1
k Dm 3
1
1
Note the first term is an even integer and the second term dm is in [0; 1]; since 2 6 k m C1 D 1: Further,
k Dm 3
We will show g .3m 1 c/

1 ck
dm D 2 6 k m C1
k Dm 3

 2 6k1Dm C1 1=3k
 2 =3

m C1

D 1=3

if cm
if cm

D0:
D1

D 0; then dm 2 [0; 1=3] and if cm D 1; then dm 2 [2=3; 1]: Since g .t C N / D g .t / for any even integer N ; so
0 if cm D 0
1
c/ D g .dm / D
D cm :
1 if cm D 1

So if cm
g .3m

Appendix 2: Everywhere Continuous, But Nowhere Differentiable Functions


Must every continuous function be differentiable at some place? Our intuition will think so, but our intuition is
wrong again. In 1872, Weierstrass proved the following result.
Theorem. There is a function continuous on R which is not differentiable at every x
Proof. On [0; 1]; define f0 .x /

only at the integers. Let fk .x /

x

2 R:

if 0  x < 1=2
: Extend f0 to R by f0 .x C 1/ D f0 .x /: Then f0 has roots
x if 1=2  x < 1
k
1
f0 .2 x /
1
and f .x / D 6 fk .x /: Since each fk is continuous on R; j fk .x /j  k on
k
2
2
k D0
1

1 1
converges, by Weierstrass M-test and the continuity theorem, f is a continuous function on R:
k D0 2k
Next we will show f is not differentiable on R: Since f ; like f0 ; is also periodic with period 1, it is enough to show
f is not differentiable at x 2 [0; 1/: Note that fk0 .x / D 1 if 2kC1 x is not an integer. Suppose f is differentiable at some

R and 6

94

x 2 [0; 1/: If x D .0:d1 d2 d3 : : :/2 ; then let an D .0:d1 d2 : : : dn /2 and bn


an increases to x while bn decreases to x : As n ! 1;
f .b n /
bn

f .an /
an

 f .bn /

f .x /
bn an
f .bn / f .x /
D b a
 f .bnn / n f .x /
D
bn x

f 0 .x / D

{z 0

! f 0 .x /

D .0:d1 d2 : : : dn 111 : : :/2 : So x 2 [an ; bn / and

 bn
f . x / f .a n / 
f 0 .x /
bn an
bn
b
x
f
.
x
/
f
.
an /
n
f 0 .x /
C
bn an
b
a
 bn x  n f .xn /
0
f .x /
C
bn an
x

f . x /D0

} | {z }
1

an 
an
n
x
an
f 0 .x /
bn an
 x
f .an /
f 0 .x /
an
bn
x
an

C bx

{z 0

! f 0 .x /

f . x /D0

an 
! 0:
an

} | {z }
1

f .bn / f .an /
: For k D 0; 1; : : : ; n 1; we have k C 1  n : So the interval .2kC1 an ; 2kC1bn / 
bn an

.d1 d2 : : : dkC1 /2 ; .d1d2 : : : dkC1 :111 : : :/2 ; which contains no integer. Hence, fk0.x / D 1 for x 2 .an ; bn /: By the
fk .bn / fk .an /
mean value theorem, there is cn 2 .an ; bn / such that
D fk0 .cn / D 1: For k D n ; n C 1; : : : ; 2k bn and
bn an
2k an are integers and so fk .bn / D 0 D fk .an /: Then
So f 0 .x /

D nlim
!1

f 0 .x /

D nlim
!1

f .bn /
bn

f .a n /
an

n 1
D nlim
!1 k6
D0

f k .b n /
bn

fk .an /
an

D k61D0 1;

which diverges by term test and gives a contradiction to f differentiable at x :


Appendix 3: Weierstrass Approximation Theorem
For an infinitely differentiable function, such as cos x ; we can compute the values of the function using its Taylor
series most of the time. However, if the function is only continuous, but not differentiable, then there is no Taylor
series. The following theorem asserts that for every continuous function f .x / on a closed and bounded interval, there
exists a sequence of polynomials Pn .x / converging uniformly on the interval to f .x /:
Weierstrass Approximation Theorem. Let f : [0; 1] ! R be continuous. For
every " > 0, there is a polynomial P .x / such that j f .x / P .x /j < " for all
x 2 [0; 1]. Taking " D 1= n ; we get a sequence of polynomials Pn .x / converging
uniformly on [0; 1] to f .x /:

f +

The theorem is also true for every closed and bounded interval [a ; b] by
substituting x D .t a /=.b a / and compact subsets of R (to be defined later).
(Note for " D 10 12, P .x / will agree with f .x / to at least 10 decimal places for
all x 2 [0; 1].)

f-
0

Remarks. Combining the continuous, nowhere differentiable function with the Weierstrass approximation theorem,
we can see that there is a sequence of polynomials converging uniformly on [0; 1] to a nowhere differentiable function.
So the uniform limit of continuous functions can be very bad in terms of differentiabilty!
To prove the Weierstrass approximation theorem, we will introduce the Bernstein polynomials of a continuous
function.
Definition. For each n

2 N, the n-th Bernstein polynomial of f .x / is



n
k n k
fn . x / D 6 f
x .1 x /n k ;
{z }
n k |
k D0
| {z } polynomial
number

95

deg fn

 n:

D 1; g .x / D x and h .x / D x 2 respectively.
n 
n
x
n
Solution. The binomial theorem asserts that .x C a / D 6
x k a n k : Differentiating x ; then multiplying by ;
n
k D0 k

n k n
n 1
k n k
and repeating these operations, we get .x C a / x D 6
x a
and
k D0 n k

n k2 n
n 2
n 1 x
[.n 1/.x C a / x C .x C a / ] D 6 2
x k an k :
n kD0 n k
n

n
n k n
Setting a D 1 x ; we get fn .x / D 6
x k .1 x /n k D 1; gn .x / D 6
x k .1 x /n k D x and
k D0 k
k D0 n k

n k2 n
.n 1 / x 2 C x
h n .x / D 6 2
x k .1 x /n k D
:
k
n
k D0 n
Examples. Find the n-th Bernstein polynomial of f .x /

As n

! 1; we get fn .x / ! f .x /;

n 
k 2 n k

Formula. 6 x
k D0

x .1

! g .x / and h n .x / ! h .x /:

   2 n  k
n
k
D k6D0 x 2 2x nk C kn
x .1
k
2
D x 2 2x 2 C .n 1/x C x

gn . x /
x /n

x /n

Proof of Weierstrass Approximation Theorem. For any " > 0, since f is uniformly continuous on [0; 1], there
"
is  > 0 such that for every x, w 2 [0; 1], jx wj <  ) j f .x / f .w/j < . By the extreme value theorem,
2
j f .x /j  M for some M. Choose integer n > M2 : Let P .x / be the n-th Bernstein polynomial of f .

"

./

2 [0; 1]; we have 0  .x 12 /2 D x 2 x C 14 ; which implies x x 2  14 : Next define S D


k o
k 2 ./
1
<  . Note k 62 S implies 2 .x
0; 1; : : : ; n and x
/  1: Then for every x 2 [0; 1];

n

n
n 



n
n
k n k
k
n k
x .1 x /n k
j f .x / P .x /j D f .x / k6D0 k x .1 x /
6 f. /
|
{z
} kD0 n k
n  D1 by examples
n
 

k  n k
k n k
n k


D k6D0 f .x / f . n / k x .1 x /  k6D0 f .x / f . n / k x .1 x /n k


 
 
k n k
k n k
n k


 k62S f .x / f . n / k x .1 x / C k662S f .x / f . n / k x .1 x /n k


 k62S "2 nk x k .1 x /n k C k662S 2M nk x k .1 x /n k


" n n k
1
k 2 n k
n k
 2 k6D1 k x .1 x / C 2M k662S  2 .x n / k x .1 x /n k
| {z }
For every x

by./

 "2 C 2M

x
n{z
2 }

by formula

 2" C 2nM 2 < "2 C 2" D ":


| {z }
by./

96

k: k

Remarks. In analysis, if you want to prove something is true for continuous functions, you should first see if it is true
for polynomials. Very often the truth of the polynomial case implies the truth for the continuous case by Weierstrass
approximation theorem.
Appendix 4 : Proofs of Theorems
Proof of the Continuity Theorem for Uniform Convergence. For every " > 0; let "0 D "=3: Since Sn .x / converges
uniformly on E to S .x /; there exists K 2 N such that n  K implies j Sn .x / S .x /j < "0 for every x 2 E : Since S K .x /
is continuous at c; so for this "0 ; there exists  > 0 such that for every x 2 E ; jx cj <  implies j S K .x / S K .c/j < "0 :
Then for every x 2 E ; jx cj <  implies

j S .x /

S .c/j  j S .x /

S K .x /j C j S K .x /

S K .c /j C j S K .c /

S .c/j < "0 C "0 C "0

D ":

The statement for series of functions follows by considering the partial sum sequence.
Proof of the Integration Theorem for Uniform Convergence. Let "n D k S
which is equivalent to Sn .x / "n  S .x /  Sn .x / C "n for all x 2 [a ; b]: So

Zb

Sn . x / d x

a/ D .L /

"n .b

 .U /
Hence, 0  .U /

Zb

Zb

. Sn . x /

Zb

"n / d x

S .x / d x

 .U /

S .x / d x

. Sn .x / C "n / d x

Zb

Sn . x / d x

S n . x /j

C "n .b

a /:

 "n ;

./

D 0; by sandwich theorem, we get




Z
Zb
b
b
b
.U /
S .x /d x D . L /
S .x /d x ; i.e. S .x / is integrable on [a ; b]: By (),
Sn . x / d x
S .x / d x  "n .b a /:
a
a
a
a
Zb
Zb
Zb
By sandwich theorem again, lim
Sn . x / d x D
S .x / d x D
lim Sn .x / d x :
n!1
n!1

S .x / d x

.L /

S .x / dx

 2"n .b

 .L /
Zb

Sn k[a ;b]: Then j S .x /

a /: Since lim "n


n!1

There is a Cauchy criterion for uniform convergence similar to the Cauchy criterion for a sequence of numbers.
Theorem (Uniform Cauchy Criterion). Sn .x / converges uniformly on E if and only if for every " > 0, there exists
K

./

2 N such that m, n  K ) j Sn .x /

Sm .x /j < " for every x

2 E:

Proof. For the if direction, the condition (*) implies Sn .x / is a Cauchy sequence for every x 2 E. So the sequence
Sn .x / converges to some number S .x / for every x 2 E. For every " > 0; by (*) again, there is K 2 N such that m ; n  K
implies j Sm .x / Sn .x /j < "=2 for every x 2 E ; which implies j Sn .x / S .x /j D lim j Sn .x / Sm .x /j  "=2 < "
m !1

for every x 2 E :
For the only if direction, given " > 0; there exists K such that n  K ) j Sn .x / S .x /j < "=2 for every x
Then m ; n  K ) j Sn .x / Sm .x /j  j Sn .x / S .x /j C j S .x / Sm .x /j < "=2 C "=2 D ":

2 E:

D 2.b "

: Since f Sn .x 0/g is a
a/
Cauchy sequence and Sn0 .x / satisfies the uniform Cauchy criterion, there is K 2 N such that m ; n  K implies
Proof of the Differentiation Theorem for Uniform Convergence. For " > 0; let "0

j Sn . x 0 /

Sm .x 0 /j <

Applying the mean value theorem to Sn

j Sn . x /

Sm . x /

"
2

and

j Sn0 .t /

Sm0 .t /j < "0

for every t

2 .a ; b/:

./

2 .a ; b/; there is t between x and y such that




"
S 0 .t / .x y /  jx y j"0 < .b a /"0 D :
./

Sm ; we get for every x ; y

Sn . y / C Sm . y /j D Sn0 .t /

97

Then m ; n  K implies j Sn .x / Sm .x /j  j Sn .x / Sm .x / Sn .x 0 / C Sm .x 0 /j C j Sn .x 0 / Sm .x 0 /j < " for every


x 2 .a ; b/: So Sn .x / satisfies the uniform Cauchy criterion and hence converges uniformly on .a ; b/ to some function
S .x /:

8 S .x / S . y /
< n
n
if x 2 .a ; b/ n f y g
is continuous at y because
Next, for every y 2 .a ; b/; the function gn .x / D
: Sn0 . yx/ y
if x D y
lim gn .x / D gn . y /: By (*) and (**), jgn .x / gm .x /j < "0 for every x 2 .a ; b/: So gn .x / satisfies the uniform Cauchy
x!y
( S .x / S . y /
if x 2 .a ; b/ n f y g
criterion and hence converges uniformly on .a ; b/: Let g .x / D lim gn .x / D
: By
x y
n!1
T . y/
if x D y
the continuity theorem for uniform convergence, g .x / is continuous at y : So
d
S . y/
dy

S .x /
D xlim
!y
x

S . y/
y

D xlim
g .x / D g . y / D T . y /:
!y

In computing an integral, the method of integration by parts allows us to switch a difficult integral to an easier
integral. Similarly, in summing a series, there is a method called summation by parts, which allows us to switch a
series to another more manageable series.

D k6Di ak D ai C aiC1 C : : : C a j and 1bk D .kbkCC11/ bkk D bkC1


j

Theorem (Summation by Parts). Let A j

6 ak bk
k Di
Proof. Note ai
m

6 ak bk
k Di

D Ai and ak D Ak

D Ai bi C . AiC1

Ak

D Am bm

m 1

6 Ak 1bk :

k Di

for k > i. So,

Ai /bi C1 C : : : C . Am

Am

/b m

D Am bm

Proof of Abels Limit Theorem. We first prove the special case c


such that j > n

 K ) jsj

Ai .bi C1

bi /

:::

Am

.b m

bm 1/:

D 0 and [u ; v ] D [0; 1]: Since k6D0 ak converges to

D a 1 C    C ak is a Cauchy sequence of numbers. So for every " > 0; there is K


j "
1
sn j D 6 ak < . We now show 6 ak x k converges uniformly on [0; 1] to S .x / by
2
k DnC1
k D0

S .1/, the partal sum sequence sk


./

bk ; then

D a0 C a1 x C : : : C an x n and Aj D kD6nC1 ak for j > n. For every


x 2 [0; 1], m > n > K , we apply summation by parts (where we take i D n C 1 in the formula) to get
m


m 1
m 1
k
m
k C1
k


j Sm .x / Sn .x /j D kD6nC1 ak x D Am x kD6nC1 Ak .x
x /  j Am jx m C 6 j Ak j.x k x kC1 /
k DnC1
by ./

 "2 x m C .x nC1 x nC2 / C : : : C .x m 1 x m / D 2" x nC1 < ":
j

verifying the uniform Cauchy criterion. Let Sn .x /

For the general case, depending on c < u or u  c  v or v < c; we have [u ; v ] is a subset of [c; v ] or [u ; c] [ [c; v ]
or [u ; c]; respectively. To show uniform convergence on [u ; v ]; it suffices to show uniform convergence on [c; v ] and
[u ; c] by the subset and union property. If c < v; then let r D v c and x 0 D h .x / D .x c/=r; ak0 D ak r k : We have
ak0 x 0k

D ak . x

1
k D0

c/k : So 6 ak .x

1
c/k converges pointwise on [c; v ] implies 6 ak0 x 0k converges pointwise on [0; 1]:

k D0
1 0 0k
1
Hence 6 ak x converges uniformly on [0; 1] by the argument above, which implies 6 ak .x c/k converges uniformly
k D0
k D0
on [c; v ] by substitution property. Similarly, if u < c; then letting r D c u and x 0 D .c x /=r; ak0 D ak . r /k ; we
will get uniform convergence on [u ; c]:

98

Chapter 11. Set Operations, Images and Inverse Images


Usually, a set contains elements that are numbers, for example ZD f: : : ; 1; 0; 1; : : :g: Later we will consider a
set whose elements are sets, for example # D fQ; R; C ; : : :g: We will say a collection (or class) to mean a set of sets.
Definitions. (a) The union of one or more sets is the set whose elements belong to at least one of these sets. The union
of sets S1 ; S2 ; : : : ; Sn may be denoted by S1 [ S2 [    [ Sn or
take the union of all sets in #; which will be denoted by

S 2#

n
[

Si : If we have a collection # of sets, then we may

i D1

S (or

S or

# ).

(b) The intersection of one or more sets is the set whose elements belong to every one of these sets. As usual, we will
use \ to denote intersection. There are similar notations for intersection of one or more sets as those for union.
(c) The complement of B in A (denoted by A n B) is the set whose elements belong to A; but not to B :
We have the following frequently used facts:
(1) If S

 T for every ; then

T and

T :

(3)

 R; R n .R n S / D S : If A  B; then for every set C; C n B  C n A:


[  [
\  \
(Distributive Law) A \
S D
S D
. A \ S / and A [
. A [ S /

(4)

[  \
\  [
(de Morgans Law) A n
S D
. A n S / and A n
S D
. A n S /

(2) For S

Definitions. For a function f : A ! B and S  A and C  B ; the set f . S / D f f .x / : x 2 S g is the image of S under
f and the set f 1 .C / D fx : f .x / 2 C g is the inverse image (or preimage) of C under f :
Remarks. In the definitions above, y

f .S/

() y D

f .x / for some x

2 S : Also, x 2

.C /

()

f .x /

2 C:

16
16
9
9

y = x2
4
1
-2 -1

-4 -3

Examples. Let f : R ! R be defined by f .x / D x 2 : Then f .[ 2; 1/ [ .3; 4]/


.1; 4] [ .9; 16] and f 1.[9; 16// D fx : x 2 2 [9; 16/g D . 4; 3] [ [3; 4/:
99

3 4

D fx 2 : x 2 [

2; 1/ [ .3; 4]g

The following are frequently used facts concerning images and inverse images of a function f : A
(5) If U
(6) U
(7) f
(8) f

 V  A; then f .U / 

 f 1. f .U // and f . f
[  [
S D
f . S /


\  \
S 
f . S /

(9) If X

 Y  B; then f

f .V /: If X

. X //  X :

[  [
S D
f


\
\
1
S / D
f

 A; then f . A n X / 

. X /  f 1.Y /:

f . A/ n f . X /:

If Y

 B; then f

. S /

. S /

! B:

. B n Y / D f 1. B / n f 1.Y /:

For the inclusion signs in facts (6), (8), (9), we now show there are examples where the two sides are not equal.
Let f : R ! R be defined by f .x / D x 2 :
(a) For fact (6), let U D .
f .R/ D [0; C1/ 6D X :

1; 0]; then

. f .U //

.[0; C1//

D R 6D U: Let X D R; then

(b) For fact (8), let S1 D . 1; 0] and S2 D [0; C1/; then f . S1 \ S2/ D f .f0g/
[0; C1/ \ [0; C1/ D [0; C1/: So, in this case, f . S1 \ S2/ 6D f . S1 / \ f . S2 /:

D f0g; but

. X //

f . S1 / \ f . S2 /

f.f

(c) For fact (9), let A D R and X D . 1; 0]; then A n X D .0; C1/; f . A/ n f . X / D [0; C1/ n [0; C1/ D ;; but
f . A n X / D f ..0; C1// D .0; C1/: So, in this case, f . A/ n f . X / 6D f . A n X /:
Next, we will demonstrate how to prove some of the facts and leave the rest to the readers.

Proof of A n

[  \
S D . A n S /:

Proof of f . f

2 An

[ 
S ()

()
()
()

[
2 A and x 62 S

x 2 A and x 62 S for every
x 2 A n S for every
\
x2
. A n S /
x

. X //  X :
y

f.f

. X //

() y D f .x /;
() y D f .x /;
H) y 2 X

where x 2 f 1 . X /
where f .x / 2 X

(Note in the last step, the (H direction is false if y cannot be written in the form f .x /:) In fact, f . f
and only if X  f . A/; where A is the domain of f :
Proof of f

[  [
S D
f

. S /:

[ 
[
S () f .x / 2
S


() f .x / 2 S for at least one
() x 2 [
f 1 . S / for at least one
() x 2 f 1. S /

100

. X // D X if

Proof of f

\  \
S 
f . S /:

y2 f

\ 

() y D f .x / for some x ;
() y D f .x / for some x ;
H) y 2 f\
. S / for every
() y 2 f . S /

\
2 S

x 2 S for every

where x
where

(Note in the third step, the (H direction is false in general because y


This x may not be the same for every :)

101

f . S / means y

f .x / for some x

2 S :

Chapter 12. Lebesgue Measure and Integral on R


There are two unsatisfactory features with Riemann integration:
(1) The integrable functions must be continuous
almost everywhere. In particular, functions that are discon
1 if x 2 Q
tinuous everywhere, such as f .x / D
; are not integrable. These include many important
0 if x 2 R n Q
functions in probability, physics and mathematics.
(2) For functions difficult to integrate, we use series expansion and properties of uniform convergence to compute
the integrals. Uniformly convergent sequences involving easily integrable functions are not always available.
These reasons are more than sufficient for the development of a more powerful integral with simpler conditions on
term-by-term integration. In 1902, such a theory was introduced by the French mathematician Henri Lebesgue. The
theory can be developed as follow:
Step 1. Study sets that are built from intervals which can be measured.
Step 2. Measure the length of as many sets as possible using the sets in Step 1.
Step 3. Identify functions having the property that the area under their graphs can be approximated by rectangles
whose bases are sets in Step 2.
Step 4. Assign an integral to each of the functions in Step 3 and study the basic properties of this integral.
Step 5. Obtain powerful theorems for term-by-term integration applicable to a wide collection of sequences and
series of functions.
Below we will develop the Lebesgue theory of integration on the real line following these steps.
Step 1: Open, Closed and Compact Sets
In R; a set is bounded iff it is bounded above and bounded below. A set is unbounded iff it is not bounded.
The open intervals in R can be divided into the following types:

.|a{z
; b/;}

the bounded type

|.a ; C1/; . 1{z; a /; . 1; C1/}; ;| D {z.a ; a}/


the unbounded types

the degenerate type

Similarly, the closed intervals can be divided into bounded and unbounded types. Now observe that R n [a ; b]
1; a / [ .b; C1/: This suggests that closed and open intervals are complement of each other.

Definitions. In R; (1) a set S is open iff it is the union of a collection of open intervals;
(2) a set W is closed iff R n W is an open set;

(3) a set K is compact iff K is a closed and bounded set.


Remark. Since R n .R n S / D S ; so a set S is open if and only if R n S is closed.
Examples. (1) Every open interval is an open set. In particular, R D .
102

1; C1/ and ; are open sets.

1; 1/ [ .2; 5/ [ .3; 7/ is an open set. Note the open intervals may overlap.
[

For f .x / D sin x ; f 1 ..0; C1// D   [ . 2; / [ .0; / [ .2; 3/ [    D
2n ; .2n C 1/ is open.

(2) The set .


(3)

n2Z

1; a]; [a ; C1/; R D . 1; C1/ and ; are closed because their complements


R n . 1; a] D .a ; C1/; R n [a ; C1/ D . 1; a /; R n . 1; C1/ D ;; R n ; D R are open.
(Exercise: Show that R and ; are the only open and closed sets in R:)
(Every finite set is compact.) If x 1 < x 2 <    < x n 1 < x n ; then the set fx 1 ; x 2 ; : : : ; x n 1 ; x n g is closed because
R n fx 1 ; x 2 ; : : : ; x n 1 ; x n g D . 1; x 1 / [ .x 1 ; x 2 / [    [ .x n 1 ; x n / [ .x n ; C1/ is open:

(4) The intervals .

(5)

The set is also compact because it is bounded above by x n and below by x 1 :


(6) The set W D f0g [ f 1n : n 2 Ng D f0; 1; 12 ; 13 ; 14 ; : : :g is closed because

Rn W

D . 1; 0/ [    [ . 14 ; 13 / [ . 13 ; 12 / [ . 12 ; 1/ [ .1; C1/

is open:

W is compact because in addition to being closed, it is bounded above by 1 and bounded below by 0.
(7) Any closed and bounded interval [a ; b] is compact. The Cantor set
K

D [0; 1] n

 1 2
; [
|3 3


1 2
7 8
1 2
7 8
19 20 
25 26 
; [ ; [
;
[
;
[
;
[
;
[



9 9
9 9
27 27 {z 27 27
27 27
27 27
}
call this set S

(always removing the open middle thirds of the remaining intervals) is a closed and bounded set because
R n K D . 1; 0/ [ S [ .1; C1/ is open and K is bounded above by 1 and bounded below by 0. Hence the
Cantor set is compact.
Note every element of an open interval is inside the interval and is not an end point.
Definitions. An element x of a set S is an interior point of S iff .x
(denoted by Int. S / or S  ) is fx 2 S : x is an interior point of S g:

r; x

C r /  S for some r > 0: The interior of S

Structure Theorem for Open Sets. For a nonempty set S in R; the following are equivalent.
(1) S is an open set.
(2) Every element of S is an interior point of S ; i.e. Int. S / D S :
(3) S is a countable union of pairwise disjoint open intervals. (Here pairwise disjoint means the intersection of
every pair is the empty set. The pairwise disjoint open intervals are called components of S.)
Proof. (1) ) (2) Suppose S is open, then S

2 S ; then x 2 .a ; b / for at least one : Let r be the


minimum of x a and b x ; then r > 0 and .x r; x C r /  .a ; b /  S : So every x 2 S is an interior point of S :
(2) ) (3) Suppose Int. S / D S : If x 2 S ; then .x r; x C r /  S for some r > 0: Now let Ix be the largest interval
containing x such that Ix  S : (Here Ix is an interval from m x D inffm : .m ; x]  S g to M x D supf M : [x ; M /  S g:)
If a D m x or M x is a real number, then a 62 S (otherwise .a r 0 ; a C r 0 /  S for some r 0 > 0; then Ix [ .a r 0 ; a C r 0 /

.a ; b /: If x

is a larger interval in S containing x :) So Ix is an open interval.


For x ; y 2 S ; if Ix \ I y 6D ;; then Ix D I y (otherwise Ix
the union of these pairwise disjoint open intervals.

[ I y is a larger interval in S containing x and y:) So S is

Finally there are only countably many of these intervals because in each of these intervals I ; we can pick a rational
number r I in the interval I : These rational numbers form a subset of Q: Since the assignment of r I to I is bijective, S
is the union of countably many pairwise disjoint open intervals.
103

(3) ) (1) This follows from the definition of open sets.


Definition. We define the length of a nonempty open set to be the sum of the lengths of the pairwise disjoint open
intervals in part (3) above. Also, we denote the length of an open set S by . S /: For the empty set, we define .;/ D 0:
Example. The set S D .0; 1/ [ .2; 4/ [ .3; 5/ is open. To express S as a countable union of pairwise disjoint open
intervals, we write S D .0; 1/ [ .2; 5/: So the length . S / D .1 0/ C .5 2/ D 4:
Concerning the lengths of open sets, we have the following useful facts.
Facts. (1) If In is a countable sequence of pairwise disjoint open intervals containing in .a ; b/; then 6 . In /  b
n

a:

(If there are finitely many intervals, then we can prove by induction. The case of one interval is easy. When there
is more than one intervals, let I j D .c; d / be the interval with the least left endpoint. Then 6 . In /  b d by
inductive step since the other intervals are in .d ; b/: Hence, 6 . In /  .b
n
infinite case, we take limit.)

d / C .d

c/ D b

n6D j

cb

a : For the

(2) If U ; V are open sets with U  V ; then .U /  .V /: Let the components of U and V be In and Jm respectively.
For each In ; take x 2 In : Then x 2 Jm for some m and Jm is the largest open interval in V containing x : Since
In is also an interval containing x in V ; we get In  Jm : For each m ; let Sm D fn : In  Jm g: By fact (1),
.V / D 6 . Jm /  6 6 . In / D .U /:
m

m n2 Sm

Theorem (Topological Properties of Open Sets).


(1)

; and R are open sets.

(2) The union of any collection of open sets is an open set.


(3) The intersection of finitely many open sets is an open set.
Proof. Part (1) is clear. For part (2), if S is open for a collection of s, then each S
intervals .a i ; b i / and so

[[

For part (3), if S1 ; : : : ; Sn are open and x

\n
k D1

\n
k D1

.a i ; b i / for some open

Sk ; then for each k ; there is a rk > 0 such that .x

Let r be the minimum of r1 ; : : : ; rn : Then r > 0 and .x


of

[
i

.a i ; b i / is open.

r; x

n
\
C r /  .x
k D1
n
\

Sk is an interior point. By the structure theorem for open sets,

k D1

rk ; x

C rk / 

n
\
k D1

rk ; x C rk /  Sk :

Sk : So every element

Sk is open.

Remarks. Part (3) above may be false if finitely is replaced by infinitely. For example, although each interval
1 1
\
1
1
1
; 1 C is open,
; 1 C D [0; 1] is not an open set (as 0 or 1 is not an interior point of [0; 1]).
k
k
k
k
k D1
Terminology. For any set X (eg. X
(1)

; and X are in C

D R), a collection C of subsets of X satisfying the properties:

(2) the union of any collection of elements of C is in C


(3) the intersection of finitely many elements of C is in C
is said to be a topology on X : By the topological properties of open sets, the class C
a topology on R: This terminology will come up in later courses.
104

D f S : S is an open set in Rg is

Theorem (Topological Properties of Closed Sets).


(1)

; and R are closed sets.

(2) The intersection of any collection of closed sets is a closed set.


(3) The union of finitely many closed sets is a closed set.
Proof. Part (1) is clear. For\
part (2),
[ W s are closed.\By de Morgans law and part (2) of the topological
 suppose
properties of open sets, R n
.R n W / is open, so W is closed. For part (3), suppose W1; : : : ; Wn are
W D

closed. Again by de Morgans law and part (3) of the topological properties of open sets, R n
is open, so

[n

k D1

n
n
[
 \
Wk D
.R n W k /
k D1

k D1

Wk is closed.

Remarks. Part (3) above may be false if finitely is replaced by infinitely. For example, although each interval
1 1
[
1
1
D .0; 1/ is not a closed set.
;1
is closed,
;1
k
k
k
k
k D3

1

Corollary. If S is open and W is closed, then S n W

D S \ .R n W / is open and W n S D W \ .R n S / is closed.

Question. Is every closed set a countable union of pairwise disjoint closed intervals? (If so, then we can define the
length of a closed set as the sum of the lengths of these intervals.)
Unfortunately, as we will see later, the answer is no. So it is more difficult to define the length of a closed set.
However, for compact sets, it is easier to define their lengths. For every compact set K ; since K is bounded, K  .a ; b/
for some open interval .a ; b/: So there are always bounded open set S containing K : By the corollary, S n K is open.
Definition. The length of a compact set K is

. K / D . S /


| {z }

 SnK ;
open

where S is any bounded open set containing K :


Remarks. In the definition, any bounded open set S containing K gives the same answer. To see that, write

D i[D1.ai ; bi / with pairwise disjoint .ai ; bi /: If .ai ; bi / \ K D ; for some i; then .ai ; bi / will be in S and S n K and
hence ..ai ; bi // will be cancelled in . S / . S n K /: So we may as well assume every .ai ; bi / \ K 6D ;:
S

Next, write .inf K ; sup K / n K

D j[D1.cj ; d j / with pairwise disjoint .cj ; d j /: If .cj ; d j / 6 S for some j ; then since
c j ; d j 2 K  S ; so c j 2 .am ; bm / and d j 2 .an ; bn / for some m ; n and c j < bm  an < d j : Then S0 D S [ .am ; bn /
is open, S0  S  K ; . S0 / D . S / C jbm an j and . S0 n K / D . S n K / C jbm an j: So . S0 / . S0 n K / D
. S / . S n K /: Thus, we may assume S  ..inf K ; sup K / n K / [ K D [inf K ; sup K ]:
Finally we can cancel . S \ . 1; inf K // and . S \ .sup K ; C1// from both . S / and . S n K / to arrive at
. K / D ..inf K ; sup K // ..inf K ; sup K / n K /:
Examples. (1) For every closed and bounded interval [x ; y]; we have [x ; y]  .x

.[x ; y]/ D . y C 1/

.x

1/

 .x

 D . y C 1/

1; x / [ . y ; y C 1/
105

1; y C 1/ and so

.x

1/

.1 C 1 / D y

x:

(2) Recall the Cantor set K is compact. Now K

. K / D 2

D3
D3

. 1/

.

1; 2/: So

 . 1; 2/ n K

1 2
1 2
7 8
; [ ; [ ;
3 3
9 9
9 9

1 2
7 8
19 20 
26 
[ 27 ; 27 [ 27 ; 27 [ 27 ; 27 [ 25
;
[



27 27

1 1 1
1
1
1
1
1C1C C C C
C C C C
3 |9 {z 9} |27 27 {z 27 27}

 . 1; 0/ [ .1; 2/ [

2 times

D3

4 times

1
1 
2 C 6 2n nC1 D 3
nD0
3

2C

1
3
2
3

D 0:

Answer to Question. No! The Cantor set K is closed. Since its length is 0, the only closed intervals contained in
K are single points fx g D [x ; x]: So every countable union of these (single point) closed intervals is a countable set,
which cannot be K as K is uncountable. To see K uncountable, consider the nonterminating base 3 representation of
x 2 [0; 1]: Let x D .0:a1 a2 a3 : : :/3 with all ai D 0; 1 or 2 and ai not eventually 0. Note
x

62 . 13 ; 23 / () a1 6D 1;

62 . 19 ; 29 / [ . 49 ; 59 / [ . 79 ; 89 / () a2 6D 1; : : : :

Hence x 2 K if and only if all ai D 0 or 2: Then f .x / D .a1 ; a2 ; a3 ; : : :/ is a bijection from K onto f0; 2g  f0; 2g 
f0; 2g     ; which is uncountable by example (10) of Chapter 3 and so is K :
To define the length of a closed set, we view the set as a limit of compact sets.
Definition. We define the length of a closed set W to be

.W / D lim  |W \ [{z x ; x]} :


x !C1
compact

Remark. As x increases, .W \ [ x ; x]/ increases. So the limit will always exist or it is C1: In particular, we
can compute it along any sequence fx n g tending to C1: If W is compact, then for x > j sup W j; j inf W j; we have
W \ [ x ; x] D W : So the length of W as a closed set is the same as the length of W as a compact set.
Examples. (1) For an unbounded closed interval [a ; C1/; by definition,

.[a ; C1// D lim .[a ; C1/ \ [ x ; x]/ D lim .[a ; x]/ D lim .x
x !C1
x !C1
x !C1
1
[

1
1
1
1
] D [1; 1 ] [ [2; 2 ] [ [3; 3 ] [   : Then W is closed because R n W
k
3
3
9
27
k D1
1
1
.1 3 ; 2/ [ .2 9 ; 3/ [    is open. Now the length
(2) Let W

a / D C1:

[k ; k C

.W / D lim .W \ [ n ; n]/ D lim 6 k


n!C1
n!C1 k D1 3
n 1

Step 2: Measurable Sets


Next we will try to measure arbitrary sets from inside and outside.
106

D k6D1 31k D

1
3

1
3

D 12 :

D . 1; 1/ [

Definitions. Let B be a subset of R:

(1) Define the Lebesgue outer measure of B to be m  . B / D inff. S / : S is open and B

 S g:

(2) Define the Lebesgue inner measure of B to be m  . B / D supf. K / : K is compact and K


(3) In case B is bounded, we say B is (Lebesgue) measurable iff m  . B /
the Lebesgue measure of B and denote it by m . B /:
(4) In case B is unbounded, we say B is (Lebesgue) measurable iff B
case, the Lebesgue measure of B is
m. B /

 B g:

D m . B /: In this case, we call this number

\ [a ; b] is measurable for every a  b: In this

D x !C1
lim m . B \ [ x ; x]/:
| {z }
bounded

(Note this equation also holds for bounded B since B \ [ x ; x] D B for x > j sup B j; j inf B j:)
Remarks. Most sets we can imagine are measurable. However, there still exist some extraordinary sets, which are
not measurable (see appendix 2 to step 2). In fact, there is a nonmeasurable set in every set with positive Lebesgue
measure! All known descriptions of these nonmeasurable sets are complicated. Outside set theory, they almost never
arise in practice.
Theorem. For any subset B of R; 0
m . B / D 0:

 m . B /  m . B /: In particular, if m . B / D 0; then B is measurable and

Proof. For every open set S and compact set K such that K  B  S ; T D S \ .inf K 1; sup K C 1/ is a bounded
open subset of S containing K : We have . K / D .T / .T n K /  .T /  . S /: This implies . S / is an upper bound
of f. K / : K is compact and K  B g: So m  . B / D supf. K / : K is compact and K  B g  . S /: This implies
m  . B / is a lower bound of f. S / : S is open and B  S g: So m  . B /  inff. S / : S is open and B  S g D m  . B /:
The second sentence of the theorem is clear.
Remarks. A set B such that m . B / D 0 is called a set of measure 0 or a null set. (For example, empty set, finite sets,
countable sets and the Cantor set K are sets of measure 0 as will be shown shortly.) These sets are regarded as small
sets in measure theoretical situations. If a property is true except on a set of measure 0, then we say the property is
true almost everywhere. (The abbreviation is a.e.) For example, if f .x / D g .x / except for x D 0; then f .x / D g .x /
a.e.. If h n .x / ! h .x / except for x in the Cantor set K ; then h n .x / ! h .x / a.e.. Also, Lebesgues theorem in Chapter
9 asserts that a bounded function on [a ; b] is Riemann integrable if and only if it is continuous almost everywhere.
Theorem. Every interval is measurable and the Lebesgue measure is the length of the interval. (Here an interval from
a to b is defined to have length b a :)
Proof. (Bounded Cases) First we will take care of some trivial cases. Since
m  .;/ D 0: By the theorem above, ; is measurable and m .;/ D 0:

; is open and .;/ D 0; it follows that

Next, if I D fa g, then I  .a "; a C "/ for every " > 0: It follows that m  . I /
By the theorem above, I is measurable and m . I / D 0:

 2" for all " > 0: So m . I / D 0:

For the nontrivial cases, let I be a bounded interval with endpoints a and b .a < b/: Since I  S" D .a "; b C "/
for all " > 0; it follows that m  . I /  b a C 2" for all " > 0: So m  . I /  b a : Similarly, using [a C "; b "]  I
for 0 < " < .b a /=2; we will get b a  m  . I /: By the theorem above, b a  m  . I /  m  . I /  b a : So
m  . I / D b a D m  . I /: Then I is measurable and m . I / D b a :
(Unbounded Cases) Clearly, the intersection of every unbounded interval with [a ; b] is a bounded interval, hence
the intersection is measurable. We will begin with R D . 1; C1/: We have m .R/ D lim m .R \ [ x ; x]/ D
x !C1

D x !C1
lim 2x D C1: Next, m .. 1; a // D lim m .. 1; a / \ [ x ; x]/ D lim m .[
x !C1
x !C1
x !C1
lim a C x D C1: Similarly, m .. 1; a]/ D C1; m ..a ; C1// D C1 and m .[a ; C1// D C1:
x !C1
lim m .[ x ; x]/

So in all cases, the Lebesgue measure of an interval is the length of the interval.
107

x ; a //

Remark. By the structure theorem for open sets, we also have m  . B /

D inf

n1

6 ja j
j D1

bj j : B

1
[
j D1

.a j ; b j / :

Theorem (Properties of Outer Measure).


(a) m  .;/ D 0:

 B; then m . A/  m . B /:
1
[
1
If A D
Ak ; then m  . A/  6 m  . Ak /:
k D1

(b) If A
(c)

k D1

Proof. Part (a) is proved in the last theorem. Part (b) follows because B

f. S / : S

is open and B

 S g  f. S / : S

 S ) A  S and so

is open and A

 S g:

 m . B /: For part (c), for every " > 0; by the remark above, there are open intervals
1
"
.ak j ; bk j / such that Ak  .ak j ; bk j / and 6 jak j bk j j < m  . Ak / C k : Then
2
j D1
Taking infimum, we get m  . A/

1
[

j D1

1[
1
[

1 1
.ak j ; bk j / and m  . A/  6 6 jak j
k D1 j D1
k D1 j D1

1
k D1

bk j j  6 m  . Ak / C ":

The result then follows by the infinitesimal principle.


Remarks. (1) There are similar properties for inner measure with similar proofs.
(2) In advanced courses, Lebesgue measure on subsets of R is generalized to subsets of an arbitrary set X : An outer
measure on X is defined to be a function on the subsets of X satisfying the three properties of outer measure above.
In showing a set is measurable, there is a theorem due to Caratheodory in 1914 which is often very useful.
Caratheodorys Theorem. A set S is measurable if and only if for every set X ; m  . X / D m  . X

\ S / C m . X n S /:

Remarks. (1) Since X D . X \ S / [. X n S /; so by part (c) of the properties of outer measure, we always have
m  . X /  m  . X \ S / C m  . X n S /: So to check measurability of a set, we have to prove only the reverse inequality.
(2) In advanced courses, this equivalent condition is used to define measurable sets. The advantages are that (i) we
do not need to consider bounded and unbounded sets separately, (ii) the condition does not involve inner measure and
(iii) when outer measures are defined by the properties above, there is no need to base the theory on open subsets
(i.e. topology).
The proofs of Caratheodorys theorem and the following useful properties of measurable sets will be presented in
appendix 1 to step 2.
Theorem (Properties of Measurable Sets).
(1) If S is open, then S is measurable and m . S / D . S /:
(2) If K is compact, then K is measurable and m . K /
.W /:

D . K /: If W is closed, then W is measurable and m.W / D

(3) If C is measurable, m .C / D 0 and C 0  C; then C 0 is measurable and m .C 0 / D 0: Also, if A; C are measurable


and m .C / D 0; then A [ C; A n C are measurable and m . A [ C / D m . A/ D m . A n C /:
(4) If Ak is measurable for every k
measurable) and

1
\
k D1

2 N; then

1
[

k D1

Ak is measurable ( i.e. countable union of measurable sets is

Ak is also measurable (i.e. countable intersection of measurable sets is measurable). If


108

A1 ; : : : ; An are measurable, then


above) and

\n
k D1

[n
k D1

Ak is measurable (by taking AnC1

Ak is measurable (by taking AnC1

AnC2

D AnC2 D    D R in the statement above).

(5) If A is measurable, then R n A is measurable. If A; B are measurable, then B n A

Terminologies. For any set X (eg. X


(1) X is in C
(2) if A is in C ; then X

D    D ; in the statement

D B \ .R n A/ is measurable.

D R), a collection C of subsets of X satisfying the properties:

n A is in C

(3) the union of countably many elements of C is in C


is said to be a  -algebra (or  -field) in X : By properties (1), (4), (5), the class C
a  -algebra in R: These terminologies will come up in later courses.

D f S : S is a measurable set in Rg is

When a set is known to be measurable, the natural problem then is to find its Lebesgue measure. The next few
facts will be helpful.
Theorem (Computation Formulas for Lebesgue Measure).
(1) (Translation Invariance) If A is measurable, then x C A D fx C a : a
measurable and m .x C A/ D m . A/ D m . A C x / for every x 2 R:
(2) (Countable Subadditivity) If A1 ; A2 ; A3 ; : : : are measurable sets, then m
AnC1

D AnC2 D    D ;; then we have m

n
[
 n
Ak  6 m . Ak /:
k D1

2 Ag; A C x D fa C x

:a

2 Ag are

1  1
[
Ak  6 m . Ak /: Also, if we take
k D1
k D1

k D1

Remark. This follows from part (c) of the properties of outer measure since m . Ak / D m  . Ak /:
(3) (Countable Additivity) If A1 ; A2 ; A3 ; : : : are pairwise disjoint measurable sets, then m
in countable subadditivity, 1 may be replaced by n :)

Remarks. As a consequence, if A  B and both are measurable, then m . A/


m . B n A/: Also, if A  B and m . A/ < 1; then m . B n A/ D m . B / m . A/:

1 
[

k D1

Ak

D k6D1 m. Ak /: (As

 m. B / since m. B / D m. A/ C

1 
[
 A2     ; then m Ak D klim
m . Ak /:
!1
k D1
1 
\
Ak D lim m . Ak /:
k !1

(4) (Monotone Set Theorem) Let A1 ; A2 ; A3 ; : : : be measurable. If A1


However, if A1

 A2     and m. Aj / < 1 for some j ; then m

1 is necessary.
1 
\
Ak D m .;/ D 0; but lim m . Ak / D 1:
k !1

Remarks. The condition m . A j / <


m

k D1

To see this, consider taking the sets Ak

k D1

Following these useful facts, we will present some examples.


Examples. (1) Since the Cantor set K is compact, by property (2), m . K / D . K / D 0:

109

D .k; C1/: Then

D fx1 ; x2; x3; : : :g1with no element repeated. Then


[  1
fxk g is measurable by property (4). By countable additivity, m. B / D m fxk g D k6D1 m.fxk g/ D 0: In fact,
BD
k D1
k D1
1  1
1 
[
[
if m . Ak / D 0 for every k 2 N; then by countable subadditivity, m
Ak  6 m . Ak / D 0; so m
Ak D 0:
k D1
(2) Let B be countable, eg. B

1
[

N; Z; Q: We can write B

k D1

Thus, m .Q/ D 0: By the remark to countable additivity, m .R n Q/ D C1 0


m .[0; 1] \ Q/ D 0 and so m .[0; 1] n Q/ D m .[0; 1]/ m .[0; 1] \ Q/ D 1 0 D 1:
(3) Let K be the Cantor set. Show that the set A
m . A/ D 0:

k D1

D C1: Also, by property (3),

D fx 2 R : x D y C z for some y 2 K ; z 2 Qg is measurable and

Solution. Since Q is countable, write Q as fr1 ; r2 ; r3 ; : : :g: Then A

1
[

C rn /: By property (2), K is measurable.


By translation invariance, K C rn is measurable and m . K C rn / D m . K / D 0 for every n 2 N: By property (4), A is
measurable. By example (2) above, m . A/ D 0:
n D1

.K

(4) Show that the set


A

D fx 2 [0; 1] : x has a decimal representation 0:a1 a2a3 : : :

such that a1 ; a2 ; a3 ; : : : are oddg

is measurable.

D fx 2 [0; 1] 1: x has a decimal representation 0:a1 a2a3 : : : such that a1 ; a2; : : : ; an are oddg: Then
\
A1  A2  A3     and A D
An : Let B D f.a1 ; a2 ; : : : ; an / : ai D 1; 3; 5; 7; 9 for i D 1; 2; : : : ; n g: Then B has
nD1
[
5n elements and An D
[0:a1 a2 : : : an 000 : : : ; 0:a1 a2 : : : an 999 : : :] is the union of 5n pairwise disjoint closed
Solution. Let An

.a1 ;:::;an /2 B

1
: By property (4), An and A are measurable. Since m . A1 /  m .[0; 1]/ D 1; by montone
10n
1 
set theorem, m . A/ D lim m . An / D lim 5n
D 0:
n!1
n!1
10n
intervals, each of length

Appendix 1 to Step 2 : Proofs of Properties and Computation Formulas


We will begin by proving a few lemmas for Caratheodorys theorem.
Lemma 1. If K is compact and O is open, then . K /
. K / C . O n K /:

. K

n O / C . O /: Also, if

Proof. Let P be a bounded open set containing K : By definition, . K / D . P /


P n .K

O ; then . O /

. P n K /: Note

n O / D P n K \ .R n O / D . P n K / [ P n .R n O / D . P n K / [ . P \ O /:


Now the sets P n K ; P \ O are open. So  P n . K n O /   . P n K / [ . P \ O /  . P n K / C . P \ O /: Hence

. K n O / C . O / D . P /  P n . K n O / C . O /
 . P / . P n K / . P \ O / C . O /
D . K / . P \ O / C . O /  . K /:
1
[
Let O D
.ak ; bk / with .ak ; bk / pairwise disjoint. If K  O ; then let inf K 2 .ai ; bi /; sup K 2 .a j ; b j / and
k D1
O 0 D .ai ; b j / \ O : Now O n [ai ; b j ] and O 0 n K are disjoint open sets and their union is O n K (as ai ; b j 62 O). So
. O / D . O 0 / C . O n [ai ; b j ]/ D . O 0 n K / C . K / C . O n [ai ; b j ]/ D . K / C . O n K /:
110

 R; we have m . A [ B /  m . A/ C m . B /  m . A [ B /:
Proof. Let O be open and B  O : Let K be compact and K  A [ B : By lemma 1, . K /  . K n O / C . O /: Since
K n O is compact in A; . K /  m  . A/ C . O /: Taking supremum over such K and infimum over such O ; we get
m  . A [ B /  m  . A/ C m  . B /:
For the inequality on the right, the case m  . A [ B / D 1 is clear. If m  . A [ B / < 1; then let O be open and
A [ B  O : For every compact K  A; by lemma 1, . O / D . K / C . O n K /: Since O n K is open and B  O n K
by disjointness, we have m  . B /  . O n K /: Then . K / C m  . B /  . O /: Taking supremum over such K and
infimum over such O ; we have m  . A/ C m  . B /  m  . A [ B /:
Lemma 2. For every pair of disjoint A; B

Lemma 3. If U is open with .U / < 1 and S is unbounded measurable, then m  .U

Proof. Let U
Let W

[n
i D1

1
[

\ S /  m .U \ S /:

1
.ai ; bi / with .ai ; bi / pairwise disjoint. Since .U / < 1; there is n 2 N such that 6 jai
i DnC1
i D1

bi j < ":

.ai ; bi /: Then
m  .U


\ S /  m .W \ S / C m .U n W / \ S  m .W \ S / C m .U n W /
1
n
 m .W \ S / C iD6nC1 jai bi j < i6D1 m .[ai ; bi ] \ S / C ":



D 1; 2; : : : ; n ; let "i D min jai 2 bi j ; 4n" : Since [ai C "i ; bi "i ] \ S is measurable, there is a compact
"
set K i  [ai C "i ; bi "i ] \ S such that m  .[ai C "i ; bi "i ] \ S /
< . K i /: Since K i  .ai ; bi /; the sets
2n
K 1 ; K 2 ; : : : ; K n are pairwise disjoint compact. So K D K 1 [ K 2 [    [ K n is a compact set in U \ S and
n
"
< . K /: Then
6 m .[ai C "i ; bi "i ] \ S /
2
i D1
For i

m  .U

\ S/

" < 6 m .[ai ; bi ] \ S /  6 m .[ai C "i ; bi "i ] \ S / C 2"i


i D1
i D1
n
 i6D1 m.[ai C "i ; bi "i ] \ S / C 2"  . K / C "  m .U \ S / C ":
Since " > 0 is arbitrary, we have m  .U \ S /  m  .U \ S /:
n

Lemma 4. For every open V ; we have m  .V / D .V / D m  .V /:

1
[
D .V / is clear. Let V D .ai ; bi / with .ai ; bi / pairwise disjoint.
i D1
[n
j
ai bi j
"i;n D
; then K n D [ai C "i;n ; bi "i;n ] is compact and K n  V : We have
2n
Proof. That m  .V /

For n

D 1; 2; 3; : : : ; let

i D1

lim . K n / D lim .1

n!1

n!1

1 n
/ 6 jai
n i D1

bi j D .V /:

So m  .V / D .V /:
Caratheodorys Theorem. A set S is measurable if and only if for every set X ; m  . X / D m  . X

\ S / C m . X n S /:
Proof. For the if direction, let X D [a ; b]; then ja bj D m  .[a ; b] \ S / C m  .[a ; b] n S /: By lemma 2, ja bj D
m  .[a ; b]/  m  .[a ; b] \ S / C m  .[a ; b] n S /: These imply m  .[a ; b] \ S /  m  .[a ; b] \ S /: As the reverse inequality
is clear, we get m  .[a ; b] \ S / D m  .[a ; b] \ S /: This yields S is measurable in the bounded and unbounded cases.
For the only-if direction, the case m  . X / D 1 is easy since property (c) of outer measure implies m  . X / 

m . X \ S / C m  . X n S /  1: So we may reduce to the case m  . X / < 1:
111

Assume S satisfies m  .U \ S /  m  .U \ S / for all open U with .U / <


an open U  X such that .U /  m  . X / C " < 1: Now

1: Then for every " > 0; there exists

m . X / C "

 .U / D m .U / .by definition of U and lemma 4/


 m .U \ S / C m .U n S /  m .U \ S / C m .U n S / .by lemma 2 and assumption/
 m . X \ S / C m . X n S /  m . X / .by U  X ; properties (b) and (c) of outer measure/:
Letting " ! 0; we get m  . X /  m  . X \ S / C m  . X n S /  m  . X /: Hence, m  . X / D m  . X \ S / C m  . X n S /:
Now every open set S satisfies the italicized assumption above by lemma 4 since U \ S is open. Also, by

lemma 3, every unbounded measurable set S satisfies the assumption. In the remaining case, S is not open and
is bounded measurable. Since U is open, by lemma 4, U satisfies the assumption as S : So for X D S ; we have
m  . S / D m  . S \ U / C m  . S n U / by the open case as U is open. Also, by lemma 2, m  . S /  m  . S \ U / C m  . S n U /:
Since S is bounded measurable, m  . S / D m  . S / < 1; so the last two sentences imply m  .U \ S /  m  .U \ S /:
Thus the remaing case also satisfies the assumption and we are done.
We will use Caratheodorys theorem to prove the following theorems, which yield properties (4), (5) of measurable
sets and countable additivity.
Theorem. If A and B are measurable, then R n A; A \ B ; A [ B are measurable.

Proof. Let A0 D Rn A: Since X n A D X \ .Rn A/ D X \ A0 ; so the Caratheodory condition is symmetric in A and A0 ; so


A measurable implies A0 D Rn A measurable. For A \ B ; since A is measurable, so m  . X / D m  . X \ A/ C m  . X n A/:
Since B is measurable,

\ A/ D m .. X \ A/ \ B / C m .. X \ A/ n B /;


Since X n . A \ B / D . X \ A/ n B / [ . X n A/ \ B [
m. X

measure that

n A/ D m .. X n A/ \ B / C m .. X n A/ n B /:

. X n A/ n B ; we get by part (c) of the properties of outer
m. X

n . A \ B //  m .. X \ A/ n B / C m .. X n A/ \ B / C m .. X n A/ n B /:
Combining with the equations above, we get m  . X /  m  . X \ A \ B / C m  . X n . A \ B //: Since X D . X \ A \
B / [ . X n . A \ B //; the reverse inequality is true by part (c) of the properties of outer measure. By Caratheodorys
theorem, A \ B is measurable.
Finally, R n . A [ B / D .R n A/ \ .R n B / is measurable and hence A [ B is measurable.
1
[
Ak ; then for every set S ; we have
Theorem. Let A1 ; A2 ; : : : be pairwise disjoint measurable sets and A D
m . X

k D1

1
k D1

m  . S / D 6 m  . S \ Ak / C m  . S n A/

D m . S \ A/ C m . S n A/:

D A1 [    [ An : Since An is measurable, by Caratheodorys theorem,


m  . S \ Bn / D m  .. S \ Bn / \ An / C m  .. S \ Bn / n An / D m  . S \ An / C m  . S \ Bn

Proof. Let Bn

/:

By induction, this gives m  . S \ Bn / D 6 m  . S \ Ak /: Since Bn is measurable by the last theorem and S

nA
! 1; we get m . S / 

k D1

n Bn ; so m . S / D m . S \ Bn / C m . S n Bn /  k6D1 m . S \ Ak / C m . S n A/: Letting n


1
6 m  . S \ Ak / C m  . S n A/  m  . S \ A/ C m  . S n A/  m  . S /; where the last two inequalities are by part (c) of
k D1
1
[
the properties of outer measure because S \ A D
. S \ Ak / and S D . S \ A/ [ . S n A/: Therefore, we have the
S

desired equation.

k D1

112

1
[
D Ak is measurable. (In case, the Ak s are not pairwise
k D1

disjoint, write A as Ai [ . A2 n A1 / [ A3 n . A1 [ A2 / [    ; then A will be measurable by the pairwise disjoint case.)
1
\
Remarks. (1) By Caratheodorys theorem, this implies A

Using de Morgans law, we see that


S

k D1

Ak is also measurable. Hence property (4) of measurable sets is true. Taking

D A; the equation yields countable additivity.

(2) If A1 ; A2 ; A3 ; : : : are measurable (not necessarily pairwise disjoint), then countable subadditivity follows by writing

1
[

k D1

m

Ak as A1 [ . A2

n A1 / [

A3


n . A1 [ A2 / [   ; which is the union of pairwise disjoint measurable sets so that

1 
[
Ak D m  . A1 / C m  . A2 n A1 / C m 

k D1

A3 n . A1 [ A2 /

 C    m . A / C m . A / C m . A / C   D
1

1
6 m  . Ak /:
k D1

(3) Monotone set theorem follows from countable additivity by taking B1 D A1 and Bn
first statement and B j D A j ; Bn D A j n An .n > j / in the second statement.

D An n An 1 .n > 1/ in the


(4) Translation invariance can be explained as follow. First note that  .x C ai ; x C bi / D jai bi j D  .ai ; bi / :
Hence, for S open, .x C S / D . S / and so m  .x C A/ D m  . A/ for every set A: Suppose A is measurable so that for
every set X ; m  . X / D m  . X \ A/ C m  . X n A/: If x 2 R; then for every set Y; let X D x C Y so that




m  Y \ .x C A/ C m  Y n .x C A/ D m  .x C X / \ .x C A/ C m  .x C X / n .x C A/


D m x C . X \ A/ C m .x C . X n A/
D m . X \ A/ C m . X n A/
D m . X / D m .Y /:
[


(5) Since every open interval .a ; b/ is measurable and m .a ; b/ D ja bj D  .a ; b/ ; so every open set S D
.ai ; bi /
i D1
1
 1
(with .ai ; bi / pairwise disjoint) is a measurable set and m . S / D 6 m .ai ; bi / D 6 jai bi j D . S / by countable
i D1
i D1
additivity. Hence property (1) is true. Since the complement of closed and compact sets are open, closed and
compact sets are measurable. For a compact set K ; if S is a bounded open set containing K ; then S n K is open and
. S / D m . S / D m . K / C m . S n K / D m . K / C . S n K / so that m . K / D . S / . S n K / D . K /: Hence for closed
set W ; by monotone set theorem, m .W / D lim m .W \ [ n ; n]/ D lim .W \ [ n ; n]/ D .W /:
n!1

n!1

(6) The first statement of property (3) of measurable sets follows from the first theorem of step 2, part (b) of the properties
of outer measure. For the second statement, if A; C are measurable and m .C / D 0; then m .C n A/ D 0 D m . A \ C /
and by countable additivity, m . A [ C / D m . A/ C m .C n A/ D m . A/ D m . A \ C / C m . A n C / D m . A n C /:
Appendix 2 to Step 2 : Proof of the Existence of Nonmeasurable Sets
Next we will show the existence of nonmeasurable sets. First we will recall the concept of an equivalent relation.
Definition. An equivalence relation R on a set E is a subset R of E

 E such that

2 E ; .x ; x / 2 R ;
(b) (symmetric property) if .x ; y / 2 R ; then . y ; x / 2 R ;
(c) (transitive property) if .x ; y /; . y ; z / 2 R ; then .x ; z / 2 R :
We write x  y if .x ; y / 2 R : For each x 2 E ; let
[ [x] D f y : x  yg: This is called the equivalence class
containing x : Note that every x 2 [x] by (a) so that
[x] D E : If x  y ; then [x] D [y] because by (b) and (c),
x2E
z 2 [x] () z  x () z  y () z 2 [y]: If x 6 y ; then [x] \ [y] D ; because assuming z 2 [x] \ [y] will
(a) (reflexive property) for every x

113

lead to x  z and z  y ; which imply x  y ; a contradiction. So every pair of equivalence classes are either the same
or disjoint. Therefore, R partitions the set E into mutually disjoint equivalence classes.
Below we will need the Axiom of Choice which states that if f S : 2 g is a nonempty collection of nonempty
disjoint subsets of a set X ; then there exists a set V  X containing exactly one element from each set S :
Theorem. Every measurable set A with m . A/ > 0 has a nonmeasurable subset.
Proof. If A is unbounded, then m . A/
enough to show this A
bounded.

\[

D x !C1
lim m . A \ [

\[
Replacing A by A \ [

x ; x]/ > 0 implies m . A

x ; x] has a nonmeasurable subset.

x ; x]/ > 0 for some x : It is


x ; x]; we may assume A is

For x ; y 2 R; define .x ; y / 2 R if y x 2 Q: We can easily check that  is an equivalence relation on R and


so we have a partition R D [ S ; where S s are disjoint sets such that x ; y are in the same S if and only if x  y :
2

Since Q is countable, each S is countable. By the axiom of choice, there is a set V containing exactly one element x
from each S :
For each r

2 Q; let V C r D fx C r : x 2 V g; then
R D fx

C r : x 2 V ; r 2 Qg D

r 2Q

C r:

If y 2 .V C i / \ .V C j / for some i; j 2 Q; then there exists x ; x 2 V such that y D x C i D x C j so that


x x D i j 2 Q: By the definition of V ; this implies x D x and so i D j : Hence .V C i / \ .V C j / D ; for
i 6D j :
Next we will show m  .V C i / D 0 for every i 2 Q: Let T D Q \[
. 1; 1/: For every compact K  V C i;
the sets K C r with r 2 T are disjoint as K C r  V C i C r: Since
. K C r /  .inf K 1; sup K C 1/; we
have m
K

r 2T

. K C r / < 1: Now m

r 2T

D

.K C r/

 V : Therefore, m .V C i / D 0: Note
A D R\ A D

[

6 m. K
r 2T

.V

r 2Q

r 2T

C r / D r62T m. K / implies m. K / D 0 for every compact


[
C r/ \ A D

r 2Q

.V


C r/ \ A :

Since m  . A/ D m . A/ > 0; by
 property (c) of outer measure,
 at least one of the set .V C r / \A has nonzero outer
measure, say m  .V C i / \ A > 0: Then m  .V C i / \ A  m  .V C i / D 0 < m  .V C i / \ A implies .V C i / \ A
is a nonmeasurable subset of A:
Remarks. In the case A D R; we see
[that one of the V C r is nonmeasurable and hence every V C r 0 D V C r C .r 0 r /
is also nonmeasurable. Then R D
V C r is measurable and the V C rs are pairwise disjoint, but none of them is
r 2Q
measurable.
Step 3: Measurable Functions

3

if x 2 Q
: On [0; 1]; every lower Riemann
2 if x 2 R n Q
sum equals 2 because every nonempty interval contains irrational numbers and every upper Riemann sum equals 3
because every nonempty interval contains rational numbers. So the function is not Riemann integrable.
Consider the function f : [0; 1]

! R defined by

f .x /

However, if instead of using the usual rectangles to approximate the area under the graph, we use rectangles whose
bases are measurable sets, then the area under the graph can be split into two such rectangles .[0; 1] \ Q/  [0; 3] and
.[0; 1] n Q/  [0; 2]: If the area of such a rectangle is defined by length of base times height, then the area under the
graph of f would be 0  3 C .1 0/  2 D 2: Lebesgue made use of this idea to extend integration to more functions.
114

More generally, let f : [a ; b] ! [0; k] be a function. If E 1 ; E 2 ; : : : ; E n are pairwise disjoint measurable sets such
that E 1 [ E 2 [   [ E n D [a ; b]; then the collection Q D f E 1 ; E 2 ; : : : ; E n g is called a measurable partition of [a ; b]:
Let m j D inf f f .x / : x 2 E j g and M j D supf f .x / : x 2 E j g for j D 1; 2; : : : ; n : The lower Lebesgue sum for f with

respect to Q is L. f ; Q /

D j6D1 m j m. E j /: The upper Lebesgue sum for f with respect to Q is U . f; Q/ D j6D1 M j m. E j /:


n

(The cases E j s are intervals lead to lower and upper Riemann sums.)

Question: What kind of functions will the supremum of their lower Lebesgue sums equal the infimum of their upper
Lebesgue sums?
Definition. Let A be a measurable set. We say f : A
measurable set for every a ; b 2 R:

! R is a (Lebesgue) measurable function iff

..a ; b// is a

Lebesgue Integral Theorem. Let f be a nonnegative valued bounded function on [x 0 ; x 1 ]: We have f : [x 0 ; x 1 ] ! R


is measurable if and only if
supfL. f ; Q / : Q is a measurable partition of [x 0 ; x 1 ]g D inffU . f ; Q / : Q is a measurable partition of [x 0 ; x 1 ]g:
(This common value is called the Lebesgue integral of f on [x 0 ; x 1 ] and is denoted by

[x0 ; x1 ]

f dm :)

Remarks. The proof will be presented in the appendix 1 to step 3. This theorem tells us that the bounded nonnegative
valued measurable functions on [x 0 ; x 1 ] have the property that the area under their graphs can be approximated by
rectangles whose bases are measurable sets. Our goal now is to know as many measurable functions as possible.
Theorem. The following are equivalent.
(1) f : A
(2) f

(3) f

(4) f

! R is measurable

.[a ; b// is measurable for every a ; b 2 R


.[a ; b]/ is measurable for every a ; b 2 R

..a ; b]/ is measurable for every a ; b 2 R

1; b]/ is measurable for every b 2 R


(6) f 1 .. 1; b// is measurable for every b 2 R
(7) f 1 .[a ; C1// is measurable for every a 2 R
(8) f 1 ..a ; C1// is measurable for every a 2 R:
1
\
1
Proof. (1) ) (2) We have [a ; b/ D
.a
; b/: So
n
(5) f

..

nD1

[a ; b/

1
\

which is measurable by (1).


(2) ) (3) We have f

(3) ) (4) We have f

(4) ) (5) We have f

.[a ; b]/ D
..a ; b]/ D
..

nD1

1
[

nD1

1; b]/ D

1
[

nD1

1
\
nD1

! \
1
D f

1 
;b
n

nD1

.a


1
; b/ ;
n

.[a ; b C //; which is measurable by (2).


1
n

.[a C ; b]/; which is measurable by (3).

1
n

.. n ; b]/; which is measurable by (4).


115

(5) ) (6) We have f

(6) ) (7) We have f

..

1; b// D

1
[

nD1

1; b

..

1
]/; which is measurable by (5).
n

.[a ; C1// D f 1.R n . 1; a // D A n f 1 .. 1; a //; which is measurable by (6).


1
[
1
f 1 .[a C ; C1//; which is measurable by (7).
(7) ) (8) We have f 1 ..a ; C1// D
n
nD1

(8) ) (1) We have


f

..a ; b// D f

.a ; C1/ n

1
\
nD1


1
; C1/ D f
n

.b

which is measurable by (8).

..a ; C1// n

1
\
nD1

1
; C1//;
n

..b

2 S is called the characteristic function of S : (In proba62 S


( R if a  0
bility, it is called the indicator function of S and is denoted by 1S .x /:) Note that XS 1 .[a ; C1// D S if 0 < a  1 :
; if 1 < a
Examples. (1) Let S be a set. The function XS .x / D

1 if x
0 if x

So if S is a measurable set, then XS is a measurable function.

(2) Every monotone function f on R is measurable. Just check all sets f


We leave this as an exercise.

..a ; C1// are intervals, hence measurable.

(3) Every continuous function on a measurable domain is measurable by the following theorem.
Definitions. Let U

 A  R: We say U is open in A if and only if U D V \ A for some open set V :


 R: The following are equivalent.

Topological Continuity Theorem. Let A; B


(1) f : A

! B is continuous.

(2) For every open set S ; f

. S / is open in A; i.e. f 1. S / D A \ V for some open set V :

(3) For every bounded open interval I , f

. I / is open in A; i.e. f 1. I / D A \ V for some open set V :

Proof. (1) ) (2) Suppose f is continuous. Let S be an open set and a 2 f 1 . S /: Then f .a / 2 S D Int. S /: So there
exists f .a / "; f .a / C "  S for some " > 0: Since f is continuous at a ; for this " > 0; there exists a > 0 such
that x 2 A and jx a j < a ) j f .x / f .a /j < ": Then
x

2 A \ .a

Hence, A\.a a ; a Ca /  f

 A\V D

a2 f

.S /

A \ .a

a ; a C a /
1

. S /: Letting V

a ; a C a /

2
[

f .x /

D
f

a2 f
1

"; f .a / C "

f .a /

.a

.S /

S )x2

. S /:

a ; a C a /; we have V is open and f 1. S / D

. S /: Therefore, f 1. S / D A \ V :

(2) ) (3) Since a bounded open interval I is an open set, so f

. I / is open in A by (2).

[
a2 f

fa g

.S /

(3) ) (1) Let a 2 A: For " > 0; consider the bounded open interval I D f .a / "; f .a / C " : By (3), f 1 . I / D A \ V
for some open V : Since f .a / 2 I ; so a 2 f 1 . I /  V : Then there exists  > 0 such that .a ; a C /  V : This
implies
x 2 A; jx a j <  H) x 2 A \ .a ; a C /  A \ V D f 1 . I /

H) f .x / 2 I D f .a / "; f .a / C "
H) j f .x / f .a /j < ":
Therefore, f is continuous at every a

2 A:

116

Remarks. (1) In the topological continuity theorem, if all the words open were replaced by closed, the statements
will remain true. This follows from W D R n S implies f 1 .W / D f 1 .R n S / D f 1 .R/ n f 1 . S / D A n f 1 . S / D
A \ .R n V /:
(2) Note the topological continuity theorem tells us that to define a continuous function, all we need to know is which
sets are open in the domain, i.e. a topology. In a course on topology where we deal with abstract spaces, there may not
be a distance between points. Continuous functions can still be defined using open sets on the spaces.
Next we will present basic properties of measurable functions. The purpose is to decide which functions are
measurable without having to use the definition. We remind the readers that for measurable set A and functions
f ; g : A ! R; we say f D g almost everywhere iff f .x / D g .x / except for x in a set of measure 0. In that case, we
denote it by f D g a.e.
Theorem (Properties of Measurable Functions). Let A be a measurable set.

! R is measurable, f . A/  B and g : B ! R is continuous, then g  f : A ! R is measurable.


1
(2) If f : A ! R is measurable, c 2 R and n 2 N, then c f ; j f j; f n ; e f ; cos f ; sin f ; (provided f .x / 6D 0 for all
f
p
x 2 A) and f (provided f .x /  0) are measurable functions.
f1
(3) If f1 ; f 2 : A ! R are measurable, then f1 C f2 ; f 1 f2 ; f1 f2 ;
(provided f2 .x / 6D 0 ), max. f1 ; f2 / and
f
(1) If f : A

min. f1 ; f2 / are measurable functions.


(4) Let f1 ; f2 ; f 3 ; : : : : A

! R be a sequence of measurable functions. For x 2 A; define the following functions

.sup fn /.x / D supf f1 .x /; f2 .x /; f3 .x /; : : : g;

.inf fn /.x / D inf f f1 .x /; f2 .x /; f3 .x /; : : : g;

.limsup fn /.x / D limsup fn .x /; .liminf fn /.x / D liminf fn .x /; . lim fn /.x / D lim fn .x /:


n!1
n!1
n!1
n!1
n!1
n!1
If each function .sup fn /.x /; .inf fn /.x /; .limsup fn /.x /; .liminf fn /.x /; . lim fn /.x / exists everywhere on A; then
n!1
n!1
n!1
each of the functions sup fn ; inf fn ; limsup fn ; liminf fn ; lim fn is measurable.
n!1
n!1
n1
(5) Let f ; g : A ! R be two functions such that f D g a.e.. If f is measurable, then g is measurable.
(6) If f : A

! R is a measurable function, then the function F : R ! R defined by


 f .x / if x 2 A
F .x / D
D . f X A /.x / is measurable.
0
if x 2 R n A

Remarks. (1) Property (4) remains true if exist everywhere on A is replaced by exist almost everywhere on A.
To see this, let E be the set where sup fn exists on A: Then m . A n E / D 0: So E D A n . A n E / is measurable.
Now sup fn X E exists everywhere on A and hence it is measurable. Since sup fn D sup fn X E a.e. on A; so sup fn is
measurable by property (5) if we consider it as sup fn X E : The other functions are similar.
(2) In place of measurable functions, if we consider Riemann integrable functions, then properties (1), (2) and (3) are
true and properties (4) and (5) are false.
For property (4), the limit of Riemann integrable functions may not be Riemann integrable when the limit exists.
For example, let Q \ [0; 1] D fr1 ; r2 ; r3 ; : : :g and fn D Xfr1 ;r2 ;:::;rn g : [0; 1] ! [0; C1/ is Riemann integrable because
it is discontinuous at only finitely many points. However, lim fn D XQ\[0;1] : [0; 1] ! [0; C1/ is not Riemann
n!1

integrable as in the beginning of step 3.

For Lebesgue integration, the limit function is measurable (hence can be integrated) whenever the limit exists by
the property (4) above. This is an advantage of Lebesgue integration over Riemann integration.
117

For property (5), the function f D 0 is Riemann integrable on every interval A


but XQ\[a ;b] is not Riemann integrable.

D [a ; b] and f D XQ\[a;b] a.e.,

Examples. (1) Show that f : R ! R defined by f .x /

D jex XQ.x / C max.cos x ; 12 /j is a measurable function.


Solution. Let f1 .x / D e x ; f2 .x / D XQ.x /; f3 .x / D cos x ; f4 .x / D 12 : Since Q is measurable, f2 is measurable. Since
f1 ; f3 ; f4 are continuous, they are measurable by the topological continuity theorem. Then f1 f2 and max. f3 ; f4 / are
measurable. Hence f1 f2 C max. f3 ; f4 / is measurable. Therefore, f D j f1 f2 C max. f3 ; f4 /j is measurable.
(2) Show that f : R ! R defined by f .x /

/.x /
D k61D1 X[k;C1
is measurable.
2k

1 1
: Now each term of the series is a constant
k D1 2k

Solution. The series converges pointwise on R by comparing with 6

times a characteristic function of an interval, hence it is measurable. The k-th partial sum function fk is measurable
because it is a sum of finitely many measurable functions. Since f D lim fk ; f is measurable.
k 21

( cos x

if x 2 [ 2; 0/
sin x if x 2 [0; 2 ] is measurable.
3
otherwise
Solution. We will express f using characteristic functions. Since [ 2; 3 ] D [ 2; 0/ [ [0; 2 ] [ .2; 3 ]; we
consider using X[ 2;0/; X[0;2 ]; X.2;3 ]: Now
(3) Show that f : [ 2; 3 ] ! R defined by f .x / D

.cos x /X[

2;0/ . x / C .sin x /X[0;2 ]. x /

( cos x C 0 C 0
3X.2;3 ] .x / D 0 C sin x C 0
0C0 3

if x
if x
if x

2 [ 2; 0/
2 [0; 2 ] D
2 .2; 3 ]

f .x /:

Since cos x ; sin x ; 3 are continuous, hence measurable, and X[ 2;0/; X[0;2 ]; X.2;3 ] are measurable, so sums and
products of these functions are measurable. Therefore, f is measurable.

D 0; then every function f : A ! R is measurable.


Since m . A/ D 0 and f 1 ..a ; b//  A; so f 1 ..a ; b// is measurable for every a ; b 2 R: Therefore,

(4) Show that if m . A/


Solution.
measurable.

(5) Show that if f : R ! R is continuous, then g : R ! R defined by g .x / D


Solution. By the fundamental theorem of calculus, g 0.x /
g is measurable by the topological continuity theorem.

f is

f .t / dt is a measurable function.

f .x /: Since g is differentiable, g is continuous. Therefore,

Remarks. This example shows the integral of a continuous function is measurable. If f is only Riemann integrable
(not necessarily continuous) on every interval [a ; b]; then the function g defined above will still be continuous, hence
measurable.
What about derivatives? As an exercise, show that if f : R ! R is differentiable, then f 0 will be measurable.
Note f 0 need not be continuous!
(6) Give an example of a nonmeasurable function.
Solution. To get such a function, take a nonmeasurable set S : Then the function XS is nonmeasurable because
XS 1 .[1; 1]/ D S is nonmeasurable.

118

Appendix 1 to Step 3 : Proofs of Properties and Theorems


Proof of Property (1). By the topological continuity theorem, for every a ; b 2 R; g
set S : By the structure theorem for open sets, S
x

2 .g  f /

1
[

nD1

1
 [
D f

.a ; b/

nD1

measurable sets.

.a ; C1/

D B \ S for some open

n D1

.an ; bn / ; which is measurable by the fact that f is measurable and property (4) of

1
x

and

Proof of Property (3). First we will show f1 C f2 is measurable. For every a

2 . f1 C f2 /

() .g  f /.x / D g . f .x // 2 .a ; b/

() f .x / 2 g 1 .a ; b/ D B \ S
1
[
() f .x / 2 S D .an ; bn /
nD1
1
1
[
 [

() x 2 f 1 .an ; bn / D f 1 .an ; bn / :

.a ; b/

Proof of Property (2). Just take g .x / to be cx ; jx j; x n ; e x ; cos x ; sin x ;


measurable functions.

.a ; b /

.an ; bn /: For every a ; b 2 R;

nD1

So .g  f /

px ; respectively in property (1) of


n

2 R;

 () f .x / C f .x / 2 .a ; C1/
1
2
() f1 .x / C f2 .x / > a
() f1 .x / > a f2 .x /
() there exists q 2 Q such that f1 .x / > q > a f2 .x /
() there exists q 2 Q such that f1 .x / 2 .q ; C1/ and f2 .x / 2 .a
1
1
() for some
[ q 21 Q; x 2 f1 .q ; 1C1/ \ f2 .a q ; C1/
f1 .q ; C1/ \ f2 .a q ; C1/ :
() x 2

q ; C1/

q 2Q

Therefore,

. f1 C f2 /

 D [  f 1 .q ; C1/ \ f 1 .a q ; C1/
| 1 {z } | 2
{z
}
q 2Q

.a ; C1/

measurable

measurable

is measurable by the fact that f1 ; f2 are measurable and property (4) of measurable sets.

Next f2 D . 1/ f2 is measurable by the property (2) of measurable


functions. So f1
f2 D f1 C . f2 / is

measurable by the first part. Next f1 f2 D 14 . f1 C f2 /2 . f1 f2 /2 is measurable by the property (2) of measurable
f1
1
D
f1
is measurable by the property (2) of measurable functions and
functions and the first two parts. Then
f2
f2
the third part.
Finally,

H)

max. f1 ; f2 / C min. f1 ; f2 / D f1 C f2
max. f1 ; f2 / min. f1 ; f2 / D j f1
f2 j
max. f1 ; f2 / D

1
. f1 C f2 C j f1
2

f2 j/ and min. f1 ; f2 / D

are measurable.
119

1
. f1 C f2
2

j f1

f2 j/

Proof of Property (4). First we will show sup fn is measurable. For every b 2 R;
x

2 .sup fn /

..

1; b]/ () .sup fn /.x /  b


() fn .x /  b for n D 1; 2; 3 : : :
() fn .x / 2 . 1; b] for n D 1; 2; 3; : : :
1
\
() x 2 fn 1.. 1; b]/:
nD1

1; b]/ D

1
\

1; b]/; which is measurable by the fact that fn s are measurable and property
(4) of measurable sets. So sup fn is measurable. Next inf fn D sup. fn / implies inf fn is measurable.
By the first part, Mk .x / D supf fk .x /; fkC1 .x /; fkC2 .x /; : : : g is measurable for every k : Since M1 .x /  M2 .x / 
M3 .x /  : : : ; limsup fn D lim Mk D inf Mk is measurable. Similarly, liminf fn D lim m k D sup m k is mean!1
k !1
k !1
n!1
surable, where m k .x / D inff fk .x /; fkC1 .x /; fkC2 .x /; : : : g: Finally, if lim fn .x / exists for every x 2 A; then
n!1
lim fn D limsup fn is measurable.
n!1
So .sup fn / 1 ..

fn 1 ..

nD1

n!1

Proof of Property (5). Let E D fx 2 A : f .x / 6D g .x /g; then m . E / D 0 and B D A n E is measurable. If x


then f .x / D g .x /: For every a 2 R; the set S D fx 2 E : g .x / > a g is a subset of E : So S is measurable. Now
g 1..a ; C1//

2 B;

D fx 2 A : g .x / > a g D fx 2 B : g .x / > a g [ fx 2 E : g .x / > a g


D fx 2 B : f .x / > a g [ S D . B \ f 1 ..a ; C1/// [ S

is measurable. Therefore, g is measurable.


Proof of Property (6). F is measurable because
F

..a ; b// D

f
f

1
1

..a ; b//
if 0 62 .a ; b/
..a ; b// [ .R n A/ if 0 2 .a ; b/

is measurable for every a ; b 2 R:


Next we will prove the Lebesgue integral theorem. We first make some remarks.
Remarks. (1) For a bounded function f , L. f ; Q /
mj

 U . f; Q/ because

D inff f .x / : x 2 E j g  supf f .x / : x 2 E j g D M j :

(2) If P D f E 1 ;    ; E m g and Q D f F1 ;    ; Fn g are measurable partitions of [a ; b]; then R D f E i \ Fj : i D


1; : : : ; m I j D 1; : : : ; n g is also a measurable partition of [a ; b] by the distributive law. This is a common
refinement of P and Q and has the property that L. f ; P /  L. f ; R /  U . f ; R /  U . f ; Q /: The proof is the
same as the proof of the refinement theorem for Riemann integral, just replace partitions by measurable partitions.
(3) (Integral Criterion) The Lebesgue integral of a bounded function f on [x 0 ; x 1 ] exists, that is
supfL. f ; Q / : Q is a measurable partition of [x 0 ; x 1 ]g D inffU . f ; Q / : Q is a measurable partition of [x 0 ; x 1 ]g
if and only if for every " > 0; there is a measurable partition Q of [x 0 ; x 1 ] such that U . f ; Q / L. f ; Q / < ": The
proof is the same as the proof of the integral criterion for Riemann integral, just replacing partitions by measurable
partitions.
120

! [0; M / is measurable.
 . j 1/ M j M 
1

Proof of the Lebesgue Integral Theorem. Suppose f : [x 0 ; x 1 ]


.x 1 x 0 / M
positive integer n >
: For j D 1; 2; : : : ; n ; let S j D f

"

a measurable partition of [x 0 ; x 1 ]: Since

.j

1/ M
n

f .x / <

jM
for all x
n

.x 1

m.Sj / D
L. f ; Q /  6
j D1 n
M

U . f; Q /

n
S j ; we get

For every " > 0; take

: Then Q

D f S1;    ; Sn g is

x0 / M
< ":
n

By the integral criterion, the Lebesgue integral of f on [x 0 ; x 1 ] exists.

2 N; by the integral criterion, there


1
is a measurable partition Q n D f E 1 ;    ; E k g of [x 0 ; x 1 ] such that U . f ; Q n / L. f ; Q n / < : Let m i D inf f f .x / :
n
k
k
x 2 E i g and Mi D supf f .x / : x 2 E i g; then gn .x / D 6 m i X E .x / and h n .x / D 6 Mi X E .x / are measurable and
i D1
i D1
gn .x /  f .x /  h n .x / for all x 2 [x 0 ; x 1 ]: Also g D sup gn and h D inf h n are measurable and g  f  h on [x 0 ; x 1 ]:
We will show g D h a.e., that is S D fx 2 [x 0 ; x 1 ] : h .x / g .x / > 0g is of measure 0. This will imply g D f a.e.,
Conversely, suppose the Lebesgue integral of f on [x 0 ; x 1 ] exists. For every n

which implies f is measurable by property (5).


Note S
Since gn .x /
Since

[k
i D1

1
[

j D1

S j ; where S j

D fx 2 [x0; x1] : h .x /

g .x / >

1
g: Let Tn; j
j

D fx 2 [x0; x1] : h n .x /

 g .x /  h .x /  h n .x /; so S j  Tn; j for all n ; j 2 N: For x 2 Ei \ Tn; j ; Mi

. E i \ Tn; j / D [x 0 ; x 1] \ Tn; j

m .Tn; j /
j
Since S j

mi

1
g:
j
1
gn . x / > :
j

gn . x / >

D h n .x /

D Tn; j and the Ei s are pairwise disjoint,

D i6Dk1 1j m. Ei \ Tn; j /  i6Dk1. Mi

m i /m . E i

\ Tn; j /  i6Dk1. Mi

m i /m . E i / D U . f ; Q n /

L. f ; Q n / <

 Tn; j for all n ; so m. S j /  m.Tn; j / < nj for all n : As n ! 1; we get m. S j / D 0: So m. S / D m

1 
[
j D1

Sj

1
:
n

D 0:

Appendix 2 to Step 3 : Cantors Function


Define a sequence of functions fn : [0; 1] ! R by
f1 . x / D x

and

fnC1 .x /

( fn .3x /=2

1=2
1=2 C fn .3x

if x
if x
2/=2 if x

By mathematical induction, we can easily checked that fn .0/


k fnC1 fn k[0;1] D 21n 3 for n D 1; 2; 3; : : : :

1
k D1

Consider the series f1 .x / C 6 . fkC1 .x /

D 0;

f n .1 /

for n

D 1;

D 1; 2; 3; : : : :

fn is continuous on [0; 1] and



f1 .x / C 6 fkC1 .x / fk .x / D lim fn .x /: Since


n!1
k D1
1
1
1
j fkC1 .x / fk .x /j  Mk D 2k 3 for every x 2 [0; 1] and k6D1 Mk D 3 by the geometric series test, so the series
converges uniformly on [0; 1] by the M-test. Therefore, f .x / D lim fn .x / is a continuous function on [0; 1] by
n!1
the continuity theorem of uniform convergence. By mathematical induction, we can also checked that every fn is
an increasing function, hence f is also an increasing function. For a graph of this function, please visit the website
http://facweb.stvincent.edu/academics/mathematics/Cantorm/cantorm.html .
fk .x //

D nlim
!1

2 [0; 1=3]
2 .1=3; 2=3/
2 [2=3; 1]

121

n 1

This function f is called Cantors function or the Cantor-Lebesgue function or Lebesgues singular function or
Devils Staircase. The last terminology is because the graph of the function is like a staircase with irregular steps. This
function exhibits many phenomena about Lebesgue measures and integrals.

Zb

f 0 .x / d x

Facts. (1) We will show the formula


a

f .b /

f .a / does not always hold. This does not contradict the

fundamental theorem of calculus because f is not differentiable everywhere on [a ; b]:


To see this, let K be the Cantor set. The function f is constant on each open interval in [0; 1] n K . The constant
1 2 1 2 7 8
1 2
7 8
19 20 25 26
values of f on the intervals . ; /; . ; /; . ; /; . ; /; . ; /; . ; /; . ; /; : : : are respectively
3 3 9 9 9 9 27 27 27 27 27 27 27 27
1 1 3 1 3 5 7
; ; ; ; ; ; ; : : :: So f 0 .x / D 0 for every x 2 [0; 1] n K : Since m . K / D 0; we may say f is differentiable
2 4 4 8 8 8 8
almost everywhere on [0; 1] and f 0 D 0 a.e. on [0; 1]: (Such a function f with f 0 D 0 a.e. is called a singular
function.) In particular, f 0 is measurable on [0; 1]: Now, the surprising fact is that
0D

Z1

f 0 .x / d x

6D

f .1/

f .0/ D 1

0 D 1:

This does not violate the fundamental theorem of calculus because f is not differentiable everywhere on [0; 1]:
(2) We will show continuous bijections do not necessarily take measurable sets to measurable sets and compositions
of two measurable functions may not be measurable.
We begin with the function g .x / D f .x / C x on [0; 1]: Then g is a continuous and strictly increasing bijection
from [0; 1] to [0; 2]: By the continuous inverse theorem, g has a continuous inverse function h : [0; 2] ! [0; 1]: Since
f is constant on every open interval .a ; b/  [0; 1] n K ; f .a / D f .b/ and so .g .a /; g .b// D . f .a / C a ; f .b/ C b/
has length ja bj: It follows that .[0; 1] n K / D 1 implies .g .[0; 1] n K // D 1: Then m .g . K // D 1 > 0: Therefore,
g . K / contains a nonmeasurable set V . Now W D h .V /  K and . K / D 0 implies W is measurable. So we see g
is a continuous bijection from a measurable set W onto a nonmeasurable set V : In particular, h D g 1 is continuous
(hence measurable) and W is a measurable set, but h 1 .W / D V is nonmeasurable!
Furthermore, consider the function XW  h : [0; 2] ! R If x 2 V ; then h .x / 2 W and so XW  h .x / D 1;
but if x 2 [0; 2] n V ; then h .x / 62 W and so XW  h .x / D 0: This implies XW  h D XV : Since XW is measurable,
h is continuous and XV is nonmeasurable, this provides an example of a composition of measurable function with a
continuous (hence measurable) function resulting in a nonmeasurable function!
(3) (American Mathematical Monthly, Problem 6378) We will show that the set of all compositions 

 : [0; 1] ! R and  : R ! R are measurable, is the set of all functions : [0; 1] ! R:

 ; where

To see this, we will start with the fact that the Cantor function f : [0; 1] ! [0; 1] is surjective as it is continuous
on [0; 1] with f .0/ D 0 and f .1/ D 1: So we can find a right inverse function e : [0; 1] ! [0; 1] of f by taking
e.x / D sup f 1 .fx g/ for example. So f  e is the identity function on [0; 1]: (Basically e.0/ D 0 and for every
x 2 .0; 1]; write the nonterminating base 2 representation 0:a1 a2 a3 : : : of x and we get e.x / equals to the base 3
number 2.0:a1 a2 a3 : : :/) This implies e.[0; 1]/  K . From f increasing, it follows that e is strictly increasing (for if
a < b; then assuming e.a /  e.b/; we will get a D f .e.a //  f .e.b// D b; a contradiction, so e.a / < e.b/). Hence,
e is measurable.
Now for every function : [0; 1] ! R; take 

D e : [0; 1] ! e.[0; 1]/ and


 . f .t // if t 2 e.[0; 1]/
 .t / D 0
:
if t 2 R n e.[0; 1]/
Then  : R ! R is equal to 0 almost everywhere on R because e.[0; 1]/  K . So  is measurable. Finally, for every
x 2 [0; 1]; we have .x / D e.x / 2 e.[0; 1]/ and so   .x / D . f .e.x /// D .x /:
122

Step 4: Lebesgue Integral and its Properties


Let f : [a ; b] ! [0; C1/ be a measurable function. The Lebesgue integral theorem is one approach to define the
Lebesgue integral for this type of function on [a ; b] in case it is bounded above. For unbounded nonnegative valued
function, we define the Lebesgue integral as follow.
Definition. Let f : [a ; b] ! [0; C1/ be measurable (and possibly not bounded above). We define

[a ;b]

f dm

D klim
!1

[a ;b]

fk dm

nZ

D sup

fk dm : k > 0 ; where fk .x / D

[a ;b]

 f .x /

if f .x /  k
if f .x / > k

D min. f; k/.x /:

(Note as k increases, the integral of fk will increase. So the limit is the same as the supremum, which is in [0; C1]:)
Once the Lebesgue integral of a nonnegative valued measurable function on a closed and bounded interval is
defined, the Lebesgue integral for a real valued measurable function on a measurable set can be defined as follow.
Definitions. (1) For f : R ! [0; C1/ measurable, define

f dm

D wlim
!1

D sup

f dm

[ w;w ]

nZ

[ w;w ]

f dm : w > 0 :

(As w increase, the integral will increase. So the limit is the same as the supremum, which is in [0; C1]:)
(2) For f : A

! [0; C1/ measurable, define

measurable functions.)

! R measurable, define
 f .x /
f C.x / D max. f .x /; 0/ D
0
Z
Z

f dm
A

f X A dm : (Note f X A is measurable by property (6) of

(3) For f : A

if f .x /  0
;
if f .x / < 0

f C dm ;

In case at least one of


A

dm is finite, define

f dm
A

f .x /

f C dm

min. f .x /; 0/

Z
f

f .x /
0

if f .x /  0
:
if f .x / > 0

dm :

Examples. (1) Let S be an measurable set. On an interval [a ; b]; considerZthe measurable partition Q
S ; [a ; b] n S g: Then L.XS ; Q / D m . S \ [a ; b]/ D U .XS ; Q /; which implies

[a ;b]

XS dm

D f[a ; b] \

D m. S \ [a ; b]/:

(2) By definition, example (1) above and the monotone set theorem, we have for c  0;

c dm
A

cX A dm

D nlim
!1

[ n;n]

cX A dm

D nlim
cm . A \ [
!1

n ; n]/ D cm . A/:

D 0 and m. A/ D C1; then cm. A/ should be interpreted as nlim


cm . A \ [ n ; n]/ D 0:
!1
Z
Z
Z
If c < 0; then cC D 0 and c D c > 0 and c dm D
cC dm
c dm D 0 . cm . A// D cm . A/:
CAUTION: If c

(3) Let m . A/ D 0 and f : A ! R be any function. Write f D f C f : On each [ w; w]; consider the measurable
partition Q A D f[ w; w] \ A; [ w; w] n Ag: SinceZm .[ w; w] \ A/ D 0 and X A D 0 on [ w; w] n A; we see that

L. fk X A ; Q A / D 0 D U . fk X A ; Q A /: These imply

f  dm
A

Z
R

f X A dm

D wlim
!1

Z
[ w;w ]

[ w;w ]

f X A dm

fk X A dm

D wlim
!1
123


lim

D 0: So
Z

k !1 [ w;w ]

fkX A dm

D 0 H)

Z
f dm
A

D 0:

! R be measurable and their integrals can be


Z
Z
Z
Z
Z
. f  g / dm D
f dm 
g dm (provided the right side is not of the form 1 1), c f dm D c
f dm
(1)
A
A
A
A
A
for every c 2 R (in case the right side is of the form 0  1; we will interpret 0  1 D 0).
Z
Z
Z
Z

(2) Let f  g a.e. on A: Then
f dm 
gdm ; with equality if and only if f D g a.e.. Also, f dm 
j f jdm:
A
A
A
A
Z
Z
Z
(3) If S ; T are disjoint measurable subsets of A; then
f dm D
f dm C
f dm : (In particular, if f is
S [ZT
ZS
ZT
Z
nonnegative valued and B is a measurable subset of A; then
f dm 
f dm C
f dm D
f dm :)
Theorem (Simple Properties of Lebesgue Integral). Let f ; g : A
defined.

Proof.
Using
Zf
Z

C
 D Zg

and g

An B

! [0; C1/: Note in this case



fk X A dm : Thus we may further reduce to the case A D [ w; w] and

g ; we can reduce to the case f ; g : A

f X A dm D lim lim
w!1 k !1 [ w;w]
R
the functions f ; g are nonnegative valued bounded functions. Then the proofs of these properties follow by replacing
partitions of [ w; w] by measurable partitions in the proofs of the simple properties of Riemann integrals. Taking
limits in k and w; we can get the equations and inequalities for the sets. The proof of the equality case of property (2)
will be left as an exercise.
f dm

fC

Question: How are Lebesgue integration and Riemann integration related?


If f is Riemann integrable on [a ; b]; then f is a measurable function and

ZTheorem (Equality
Z b of Integrals).
f dm D
f .x / d x :
[a ;b]

Proof. Let P be the partition fa D x 0 < x 1 <    < x n D bg of [a ; b]: Then the collection f E 1 ; E 2 ; : : : ; E n g; where
E 1 D [x 0 ; x 1 /; E 2 D [x 1 ; x 2 /; : : : ; E n D [x n 1 ; x n ]; is a measurable partition of [a ; b]: For the case f .x /  0 on [a ; b];

n
o
D sup L . f; P / : partition P D fa D x0 < x1 < : : : < xn D bg of [a ; b]
a
 supfL. f; Q/ : measurable partition Q D f E1 ; E2; : : : ; Ek g of [a ; b]g
 inffnU . f; Q/ : measurable partition Q D f E1 ; E2; : : : ; Ek g of [a ; b]g o
 inf U . f; P / : partition P D fa D x0 < x1 < : : : < xn D bg of [a ; b]
Zb
D .U / f .x / d x :
a
Zb
Zb
Zb
So if f is Riemann integrable on [a ; b]; then . L /
f .x / d x D
f .x / d x D .U /
f .x / d x and all inequalities
a
a
Za
Zb
above become equalities. By the Lebesgue integral theorem, f is measurable and
f dm D
f .x / d x : The
[a ;b]
a
case f is real valued follows by considering f C ; f  0 first. As above, they will be measurable and their Riemann
and Lebesgue integrals are the same. Then f D f C
f is measurable and the Riemann and Lebesgue integrals of
.L /

f .x / d x

f are the same.

Examples. (4) By property (3), example (3) above and equality of integrals,

[0;1]nQ

x 2 dm D

[0;1]

x 2 dm

{z

[0;1]\Q

x 2 dm D

by property .3/

124

}|

x 2 dm

{z

[0;1]

by example .3/

Z1
0

x2 dx

D 13 :

(5) Let f : [a ; C1/ ! [0; C1/ be locally Riemann integrable. Then

[a ;C1/

f dm

f X[a ;C1/ dm

D wlim
!1

[a ;w ]

f dm

D wlim
!1
D wlim
!1

Zw

[ w;w ]

f X[a ;C1/ dm

f .x / d x

Z1

f .x /:

Remarks. A similar result holds for any other unbounded interval in place of [a ; C1/: So for a nonnegative valued
function f ; the improper Riemann integral of f on an unbounded interval equals the Lebesgue integral on the same
interval.
(6) If f : R ! [0; C1/ is measurable,

Z p

f dm D

f dm C

{z

by property .3/

f dm 

Z p

}|

f dm C

{z

f 2 dm < 1; then show that

f dm < 1:
R
R
R
Solution. Let A D fx 2 R : 0  f .x /  1g D f 1 .[0; 1]/ andpB D fx 2 R : f .x / > 1g D f 1 ..1; C1//: Then A
and B are measurable, A [ B D R and A \ B D ;: Since f  f on A and f  f 2 on B ; so

f dm < 1 and

f 2 dm 

}|

by property .2/

Z p

f dm C

{z

by property .3/

f 2 dm < 1:

Remarks. If f is only Riemann integrable, then the sets A and B may not be unions of intervals. In that case, the
above integrals on A and B will not be defined for Riemann integrals!
There is a more elegant approach in defining the Lebesgue integral for nonnegative valued functions. It takes care
the bounded and unbounded cases at the same time.
The approach goes as follow. Any function  : A

! R of the form  D i6D1 ci XS ; where ci 2 R and Si is


n

a measurable subset of A for i D 1; 2; : : : ; n ; is called a simple function on A: (We remark that we may even
assume the sets Si are pairwise disjoint subsets of A because if the range of  consists of distinct values r1 ; r2 ; : : : ; r j ;

D i6D1 ri X .fr g/ and the sets  1.fri g/ are pairwise disjoint subsets of A.) For such a simple function ;
Z
n
 dm D 6 ci m . Si / by the remarks and properties above. The Lebesgue integral of a nonnegative measurable
i
D
1
A
Z

Z
function f : A ! [0; C1/ is taken to be
f dm D sup
 dm : 0    f;  simple (and for real valued f;
A
A
C
we consider f D f
f as before). The following theorem will show this approach agrees with our definition.
Z

Z
Theorem. For a measurable function f : A ! [0; C1/;
f dm D sup
 dm : 0    f;  simple :
A
A
Z

Z
Z
Proof. Let S D
 dm : 0    f;  simple : Since   f on A implies  dm 
f dm ; we have
A
A
A
Z
sup S 
f dm : For the reverse inequality, since
A
Z
Z
nZ
o
f dm D
f X A dm D sup
fk X A dm : w; k > 0 ;
A
R
[ w;w ]
Z
Z
Z
1
for every n 2 N; there are wn ; kn > 0 such that
f dm
<
fk X A dm 
f dm : Next, since an inten
then 

[ wn ;wn ]

gral is the supremum of lower sums, there exists a lower sum L. fkn X A ; Q n / D 6 m i m . E i /; where Q n
j

i D1

is a measurable partition of [ wn ; wn ]; such that

[ wn ;wn ]

fkn X A dm

1
< L. fkn X A ; Q n / 
n
125

Z
[ wn ;wn ]

fkn X A dm :

D f E1 ; : : : ; E j g

Let n

D i6D1 mi X E ; then 0  n 
j

f kn X A

f ; n is simple and

2
n

f dm
A

Z
By sandwich theorem,

f dm
A

D nlim
!1

Z
A

Z
A

n dm 

n dm D L. fkn X A; Q n /: Then

f dm :
A

n dm  sup S :

Step 5: Convergence Theorems.


When expanding a function into a series of functions, the series often does not converge uniformly on the domain.
So for functions difficult to integrate, the following theorems are more useful than the integration theorem for uniform
convergence. This again is another important advantage of Lebesgue integration over Riemann integration!

{z ! [0; C1}/ be measurable functions. If

Monotone Convergence Theorem(MCT). Let f1 ; f2 ; f3 ; : : : : A

fi . x /0

(1) lim fn .x / D f .x / a.e. on A and


n!1

(2) for almost every x

then

f dm
A

2 A;

f1 . x /

lim fn dm

A n!1

f2 . x /

D nlim
!1

f3 . x /

   ;

fn dm :
A

{z ! [0; C1}/ be measurable functions.

(Corollary to) Monotone Convergence Theorem. Let g1 ; g2 ; g3 ; : : : : A

1
6 gk .x / converges for almost every x
k D1

Z2
Examples. (1) Find lim

n!1 1

Solution. On [1; 2]; as n

x
1Ce

nx

1
A; then
6 gk dm
A k D1

gk . x /0
1Z
D k6D1 gk dm:
A

dx:

% 1; we have

nx

& 1; 1 C e

nx

& 1 and

fn . x /

continuous (hence Riemann integrable) on [1; 2]; by equality of integrals and MCT,

Z2
lim

n!1

If

x
1Ce

nx

dx

Z
x
D nlim
!1 [1;2] 1 C e
Z
x
D
lim
Z[1;2] n!1 1 C e
D
x dm
[1;2]
Z2
D x d x D 32 :
1

(2)
126

nx

dm

nx

dm

D 1 Cxe

nx

% x : Since

fn is

Z1

ee d x

Z
Z

[0;1]

ee dm

.ee continuous. Equality of Integrals/


x

1 .e x /k
1 wk
dm .Used ew D 6
for every w 2 R/
6
k! }
k D0 k!
[0;1] k D0 | {z
Z ekx0
1
D k6D0
dm .Terms continuous. Corollary to MCT/
[0;1] k!
Z
1 1 kx
D k6D0 ek! d x .Equality of Integrals/
0
1 k
D 1 C k6D1 e k!k 1 :

(3)Z

x 1 =3
1
ln dm
x x
.0 ;1 / 1

1
1
a
6 x .1=3/Ck ln dm .Used
1 r
.0;1/ k D0 |
{z x}

D k6D0 ar k

for r

2 .0; 1//

1Z
D k6D0
x .1=3/Ck ln x dm (Terms continuous. Corollary to MCT)
Z.0;1/
1 1 .1=3/Ck
D k6D0
x
ln x d x .8 > 0; lim x ln x D 0 ) Riemann integrable on [0; 1]/
x !0C
0


Z

1
.
4
=
3
/
C
k
.
1
=
3
1
1
x
x /Ck

(Integration by parts)
D k6D0 .4=3/ C k ln x 0C
dx
0 .4=3/ C k
1
D k6D0 .3k C9 4/2 :
>0 on .0;1/

(4) Let f : .0; 1] ! [0; C1/ be improper Riemann integrable. Observe that if A

X A .x / D
Since [ n1 ; 1]  [ nC1 1 ; 1]
of integrals,

H)

.0;1]

f dm

(1

if x
0 if x
0 if x

(1
2A
2 B n A  X B .x / D 1
62 B
0

if x
if x
if x

 B; then

2A
2BnA:
26 B

X[ 1n ;1]  X[ nC1 1 ;1]; so as n ! 1; f X[ 1n ;1] increases to f on .0; 1]: By MCT and equality

D nlim
!1

Z
.0;1]

f X[ n1 ;1] dm

D nlim
!1

Z
[ 1n ;1]

f dm

D nlim
!1

Z1
1= n

f dx

Z1

f dx:

Remarks. A similar result holds for any other bounded interval in place of .0; 1]: So for a nonnegative valued function
f ; the improper Riemann integral of f on a bounded interval equals the Lebesgue integral on the same interval.

Z
Definition. A measurable function f is Lebesgue integrable on A iff
A

j f j dm D

Lebesgue Dominated Convergence Theorem(LDCT). Let f1 ; f2 ; : : : : A


(1) lim fn .x / D f .x / a.e. on A and
n!1

(2) there exists a Lebesgue integrable function g : A

Z
A

f C dm C

Z
f
A

dm < 1:

! R be measurable functions. If

! R such that for all n ; j fn .x /j  g .x / a.e. on A;


127

Z
then

f dm
A

lim fn dm

A n!1

D nlim
!1

fn dm :
A

Remarks. The special case when m . A/ < 1 and g is a constant function is called the Lebesgue Bounded Convergence
Theorem (LBCT) in many books.
(Corollary
to) Lebesgue Dominated
Convergence
Z
Z
Z Theorem. Let g1 ; g2; : : : : A

1
6
k D1

jgk j dm < 1; then k61D1 gk dm


A
A

D k61D1

! R be measurable functions. If

gk dm :

The corollary was proved by Beppo Levi. So in many books, this corollary is also called the Beppo Levi theorem.

Z1

Examples. (1) Show that

sin x ln x d x
0

k C1

D k61D0 .2k C. 1/1!./2k C 2/2 :

sin x
(Note lim sin x ln x D lim
.x ln x / D 1  0 D 0: So sin x ln x can be extended to a continuous (hence Riemann
x !0C
x !0C x
integrable) function on [0; 1]:)
Solution. Recall sin x

k 2k C1

D k6D0 . .2k1/Cx 1/!

for all x

1/k x 2kC1 ln x
: Since lim x 2kC1 ln x
x !0 C
.2k C 1/!
function on [0; 1]; so by equality of integrals,

D.

gk . x /

Z1

x 2kC1 ln x
dx
.2k C 1/!

to LDCT,

 x 2kC2 1 Z 1 x 2kC1 
1
D .2k C 1/! 2k C 2 ln x 0C
dx D
.2k C 1/!.2k C 2/2
0 2k C 2
1

[0;1]

Z1

sin x ln x d x
0

Z
(2) Show that

[0;1/

1
D k6D0

1
6 gk dm
[0;1] k D0

cos

px d x D

gk . x /

D.

1/ x e
.2k /!

1
k D0

So 6

[0;1/

1
gk dm D 6
k D0
[0;1]

[0;1/

1
. 1/kC1
. 1/kC1 jgk j dm D 6
:
k D0 .2k C 1/!.2k C 2/2
[0;1]

1 . 1/k k!
:
6
k D0 .2k /!

1 . 1/k w2k
for every w
.2k /!
k D0

2 R: So on [0; 1/; e

cos

D k6D0 .

1/k x k e
.2k /!

: Next, let

: By example (5) of step 4,

jgk j dm D wlim
!1
1

[0;1/

1 1
< 6 2 < 1 (by the comparison and p-tests). By the corollary
k D1 k

Solution. Recall cos w D 6


k k

1/k x 2kC1 ln x
: Next, let
.2k C 1/!

D k6D0 .

R: So on .0; 1]; sin x ln x

D 0; gk can be extended to a continuous (hence Riemann integrable)

jgk j dm D
0
1Z
1
1
and so 6
j
gk j dm D 6
k D0
k D0 .2k C 1/!.2k C 2/2
[0;1]

Z
[0;w ]

jgk j dm D wlim
!1

Z w xke
0

.2k /!

dx

Z 1 xke
0

.2k /!

dx

D .2kk!/! :

jgk j dm D k6D0 .2kk!/! < k6D0 k!1 D e < 1 (by the comparison test). By the corollary to LDCT,
e

Z
p
cos x d x D

1
6 gk dm
k
[0;1/ D0

1Z
D k6D0
gk dm
[0;1/

128

1Z
D k6D0
. 1/k jgk j dm
[0;1/

D k6D0 . .2k1//!k! :
k

Z1
(3) Find lim

n!1

nx ln x
dx:
1 C n2x 2

 nx ln x

if 0 < x  1 ; then by lHopitals rule, lim f .x / D 0 on [0; 1]: Using the inequality
1Cn2 x 2
n
n!1
0
if x D 0

j ln x j if 0 < x  1
t
1
2
or by calculus,

; we have j fn .x /j  g .x / D
: Since g D jg j  0 on [0; 1] and the
2
1Ct
2
0
if x D
0
Z1
Z
1 1
1
improper Riemann integral
g .x / d x D
.x ln x x / D ; so
jg j dm D 12 and g is Lebesgue integrable
0
C
2
2
0
[0;1]
on [0; 1]: By LDCT,
Solution. Let fn .x / D

Z1

lim

n!1 0

nx ln x
dx
1 C n2x 2

D nlim
!1

[0;1]

fn dm

lim fn dm

[0;1] n!1

Remarks. The sequence fn does not converge uniformly on [0; 1] to 0 because k fn


So integration theorem for uniform convergence cannot be applied directly.

[0;1]

0 dm

D 0:

1
ln n
0k[0;1]  j fn . /j D
n
2

! 1:

The following remarks compare Riemann integrable functions with Lebesgue integrable functions.
Remarks. (1) For proper Riemann integrals, we have the following.
Theorem. If f : [a ; b] ! R is Riemann integrable, then (j f j is Riemann integrable on [a ; b]: By the equality of
integrals, the Lebesgue integral of j f j on [a ; b] is finite and so) f is Lebesgue integrable on [a ; b]:

However, the converse is false.


[a ;b]

f dm

D 0  1 C .b

The function f .x /

a /  0 D 0: Since j f j

1
0

if x
if x

2 Q \ [a ; b]
62 Q \ [a ; b]

is measurable on [a ; b] and

f ; f is Lebesgue integrable on [a ; b]: However, f is not

Riemann integrable on [a ; b]:


(2) For improper Riemann integrals, we have the following.
Theorem. If j f j is improper Riemann integrable on an interval I ; then f is (improper Riemann integrable by the
comparison test and) Lebesgue integrable on I and the two integrals give the same value. In particular, this holds true
for nonnegative valued functions.
The proof in the unbounded interval case follows from the ideas in example (5) of step 4 and in the bounded
interval case from example (4) of MCT. It is a consequence of equality of integrals, MCT and LDCT (using g D j f j).
If the function is real valued, then it is possible that the function be improper Riemann integrable, but not Lebesgue
integrable as the following examples show.
1 1 1 1
For the case of a bounded interval, say [0; 1]; we can partition [0; 1] with the points 0; : : : ; 16
; 8 ; 4 ; 2 ; 1:
k 1 k
. 1/ 2
1
1
Let g : [0; 1] ! R be defined by g .0/ D 0 and g .x / D
when k < x  k 1 for k 2 N; i.e.
k
2
2
k
1
k
1 . 1/ 2
g .x / D 6
X. 1k ; k1 1 ] .x /: Then g is improper Riemann integrable on [0; 1] because
2
2
k
k D1

Z1
0

g .x / d x

D wlim
!0C

Z1
w

g .x / d x

D k6D1 .

1 /k 1 2 k 1
k
2k 1

1 . 1 /k
1
D
6
2k
k
k D1

by the alternating series test. However, g is not Lebesgue integrable on [0; 1] because

[0;1]

< C1

jg j dm D k61D1 k1 D C1:

So some improper Riemann integrable functions are not Lebesgue integrable. Observe that g C and g has infinite
areas under the curves, so the Lebesgue integral for g cannot be defined.
129

1 . 1/k 1
. 1/[x]
X[k 1;k/ .x / D
; where [x]
k
[x]
C
1
k D1
Z1
Z
1 . 1 /k 1
jg j dm D
denotes the greatest integer less than or equal to x : Again
g .x / d x D 6
< C1 and
k
k D1
0
[0;C1/
1 1
sin x
on R:)
6 D C1: (Exercise: Another example is
x
k D1 k
For the case of an unbounded interval, say [0; C1/; take g .x / D 6

Appendix 1 to Step 5 : Proofs of Theorems and Corollaries


The two convergence theorems can be proved using a lemma due to Pierre Fatou in 1906.
Fatous Lemma. If f1 ; f 2 ; : : : : A

! [0; C1/ are measurable functions, then

liminf fn dm

A n!1

 liminf
n!1
Z

fn dm :
A

C1 for some x 2 A; we need to clarify the meaning of liminf


fn dm : Let
A n!1
S D fx 2 A : liminf fn .x / D C1g and T D fx 2 A : liminf fn .x / 2 Rg: Since fn .x /  0 on A; the function m k .x / D
n!1
n!1
inff fk .x /; fkC1 .x /; : : : g exists as real numbers everywhere on A and hence measurable. Then
1
1 \
1
 [
 [

.sup m k / 1 . 1; n] D
m k 1 . 1; n]
T D .liminf fn / 1 .R/ D .sup m k / 1 . 1; C1/ D
n!1
Since liminf fn .x / may be
n!1

nD1

nD1 k D1

D A n T : Now we can use the last theorem of step 4 to define


Z
nZ
o
liminf fn dm D sup
 dm : 0    liminf fn ;  simple :
n!1
n!1

is measurable and so is S

So in case m . S / > 0; the integral will equal C1:

D liminf
fn on A: Consider a nonnegative simple function  D 6 ci XS  f on A
n!1
i D1
with the Si pairwise disjoint. Then each ci  f D liminf fn on Si : For each t 2 .0; 1/; w 2 Si ; t ci < liminf fn .w/: This
n!1
n!1
1
\
implies t ci < inff fk .w/; fkC1 .w/; : : : g for some k : Then w 2 Bik D fx 2 Si : t ci < f j .x /g: Now Bi1  Bi2    
j Dk
1
[
and their union contains every w 2 Si : Hence their union is Si : By the monotone set theorem, lim m . Bin / D m .
Bin /
n!1
n D1


r
if w 2 Bin
D m. Si /: Define n D i6D1 t ci X B : Now for w 2 Si ; n .w/ D t0ci ifif ww 22 SBiinn Bin  ffnn .w/
.w/ if w 2 Si n Bin : So
n  fn on every Si (hence also on A). We have
Z
Z
r
fn dm 
n dm D 6 t ci m . Bin /:
i D1
r

Proof of Fatous Lemma. Let f

in

Taking the limit infima of both sides, we get

Z
r
r

liminf 6 t ci m . Bin / D t 6 ci m . Si / D t
 dm :
n!1
n!1 i D1
i D1
A
A
Z
Z
Now let t go to 1, then we get liminf
fn dm 
 dm : Taking supremum over all nonnegative simple  
n!1
A
A
get
Z
 Z
Z
Z
liminf
fn dm  sup
 dm : 0    f;  simple D
f dm D
liminf fn dm :
n!1
n!1
liminf

fn dm

130

f ; we

Z
Z
D 1; 2; 3; : : : ; so we have 0  fn dm  f dm: Hence
AZ
A
Z
Z
Z
Z
limsup
fn dm 
f dm : By Fatous lemma,
f dm D
liminf fn dm  liminf
fn dm : Since liminf 
n!1
n!1
A
A
A Z
A n!1
A
Z
limsup; the two inequalities above combine to give lim
fn dm D
f dm :
n!1
Proof of MCT. Since 0

fn

f a.e. on A for n

Proof of Corollary to MCT. The function f n D 6 gk is measurable. Since lim fn


n

k D1

fnC1 ; by MCT,

n!1

1
6 gk dm
k
A D1

Z
lim fn dm

A n!1

D nlim
!1

1
fn dm D lim 6
gk dm D 6
n!1 k D1 A
k D1
n

D k6D1 gk a.e. and f n  fn C gnC1 D

Proof of LDCT. Since j fn j  g a.e. on A implies g C fn ; g

gk dm :
A

 0 a.e. on A; by Fatous lemma,


Z
Z
Z
g dm C
f dm D .g C f / dm D
liminf.g C fn / dm  liminf .g C fn / dm D
g dm C liminf
fn dm :
n!1
n!1
A
A
A
A n!1
A
A
A
Z
Z
Z
Z
Z
Z
Z
g dm
f dm D .g f / dm D
liminf.g fn / dm  liminf .g fn / dm D
g dm limsup
fn dm :
n!1
n!1
A
A
A
A n!1
A
A
A
Z
Z
Z
Z
Z
Cancelling g dm < 1; we get
f dm  liminf
fn dm  limsup
fn dm 
f dm : The result follows.
n!1
n!1
A
A
A
A
A
Z
n
1
1
Proof of Corollary to LDCT. The function fn D 6 gk is measurable. Now 6
j
gk j dm < 1 implies 6 jgk .x /j
k D1
k D1 A
k D1
1
converges a.e. on A: (This is because if S D fx 2 A : 6 jgk .x /j D C1g has nonzero Lebesgue measure, then we get
k D1

the contradiction that

Z
1
6 jgk j dm D liminf
n!1
k D1 A

6 jgk j dm 
A k D1
n

fn

liminf 6 jgk j dm
n

A n!1 k D1

1
6 jgk j dm 
A k D1

1
1
So lim fn D 6 gk exists a.e. and j fn j  g D 6 jgk j a.e. on A: By MCT,
n!1
k D1
nD1
Lebesgue integrable on A: By LDCT,

1
6 gk dm
A k D1

lim fn dm

A n!1

D nlim
!1

1
6 jgk j dm D C1:/
S k D1

1
jg j dm D k6D1
A

1
fn dm D lim 6
gk dm D 6
n!1 k D1 A
k D1
n

jgk j dm < 1; so g is
gk dm :

Appendix 2 to Step 5 : Higher Dimensional Lebesgue Measure and Integral


For Rn ; there are also Lebesgue measure and integral. Define an n-dimensional open interval (or open rectangular
cell) to be any set of the form
I

D .a1 ; b1/  .a2 ; b2/      .an ; bn / D f.x1 ; x2; : : : ; xn / : ai < xi < bi

for i

D 1; 2; : : : ; n g:

Similarly, an n-dimensional semi-open interval is any set of the form

D [a1; b1/  [a2; b2/      [an ; bn / D f.x1 ; x2; : : : ; xn / : ai  xi < bi for i D 1; 2; : : : ; n g:


As usual, the n-dimensional volume of these intervals is . I / D . J / D .b1 a1 /.b2 a2 /    .bn an /: In Rn ;
J

open sets are sets that are unions of n-dimensional open intervals, closed sets are the complements of open sets and
compact sets are closed and bounded sets. (It is a good exercise for the reader to define the volume of an open set S
by first proving a structure theorem that S is a countable disjoint union of semi-open intervals, then taking the sum
131

of the volumes of the semi-open intervals to be the volume of S : The volume of compact and closed sets can be
defined as before.) However, to develop the theory so it can be used in more general settings, it is common to take the
Caratheodory approach. For a set B in Rn ; we define the outer measure of the set B by
m n . B /

D inf

n1

1
o
[
Ik ; Ik open interval :
6 . Ik / : B 
k D1
k D1

Caratheodorys theorem suggests that for a set S in Rn ; we can define S to be (Lebesgue) measurable if and only if for
every set X in Rn ; m n . X / D m n . X \ S / C m n . X n S /: For a measurable set S ; we define the Lebesgue measure m n of
S to be its outer measure, i.e. m n . S / D m n . S /:
Next we define a function f : A ! R to be (Lebesgue) measurable if and only if A is a measurable set in Rn
and f 1 ..a ; b// is a measurable set for every a ; b 2 R: For a nonnegative measurable functions f : A ! [0; C1/;
we define the Lebesgue integral on A as in the last theorem in step 4 via simple functions. All the properties of outer
measure, measurable sets, measurable functions, Lebesgue integral and convergence theorems are true by essentially
the same proofs.
There are additional facts about these Lebesgue measures and integrals.
Theorem. If X is a measurable set in Rn and Y is a measurable set in Rk ; then X  Y is a measurable set in RnCk :
If f : X  Y ! R is a measurable function, then for each x 2 X ; f x : Y ! R defined by f x . y / D f .x ; y / is a
measurable function and similarly, for each y 2 Y; f y : X ! R defined by f y .x / D f .x ; y / is also a measurable
function.
Furthermore, we have the following theorems for computing high dimensional integrals by multiple integrals
as in multivariable calculus. The first one is for nonnegative valued functions and the second one is for real valued
functions.

 Y ! [0; C1/ is measurable,


Z then the functions  : X ! [0; C1/ defined by
: Y ! [0; C1/ defined by . y / D
f y dm n are measurable and
X


Z
Z Z
Z Z
y
f x dm k dm n D
f dm nCk D
f dm n dm k :

Fubini-Tonellis
Theorem. If f : X
Z

.x / D

f x dm k and
Y

Fubinis Theorem. If f : X

X Y

 Y ! R is Lesbesgue integrable (i.e.

as in the Fubini-Tonellis theorem are true.

132

Z
X Y

j f j dmnCk < 1), then similar conclusions

Chapter 13. Vector-valued Functions


In this chapter we will be working with the n-dimensional vector space Rn and studying Rm -valued functions
F : Rn ! Rm :
n
Conventions. An element
0 x1 1of R may be viewed as a point .x1 ; : : : ; xn / or as a vector from the origin to the point or
:
as a column matrix @ :: A depending on the context. We will let O D .0; : : : ; 0/ denote the origin of Rn :
xn

Definitions. (1) Define the length (or norm) of v

D .x1 ; : : : ; xn / 2 Rn to be kv k D

x 12 C    C x n2 :

(2) For v

D .x1 ; : : : ; xn /; w D . y1 ; : : : ; yn / 2 Rn ; define the distance between v and w to be


p
d .v; w/ D kv wk D .x 1 y1 /2 C    C .x n yn /2 :

(3) For v

D .x1 ; : : : ; xn / 2 Rn and r > 0;


S .v; r / D fw 2 Rn : d .v; w/ D r g is the sphere centered at v of radius r,
B .v; r / D fw 2 Rn : d .v; w/ < r g is the open ball centered at v of radius r;
B .v; r / D fw 2 Rn : d .v; w/  r g is the closed ball centered at v of radius r:

(The open ball B .v; r / is also called the r-neighborhood of v and is also denoted by Nr .v/ in some books.)
Theorem (Triangle Inequality). For u ; v; w 2 Rn ; we have ku

wk  ku

v k C kv

wk: (Exercise.)

A vector-valued function is a function F : Rn ! Rm ; which assigns to every .x 1 ; : : : ; x n / 2 Rn a unique


value F .x 1 ; : : : ; x n / D . y1 ; : : : ; ym / 2 Rm : Each coordinate yi depends on x 1 ; : : : ; x n : Hence each yi is a function
fi : Rn ! R, i.e. yi D fi .x 1 ; : : : ; x n /: The function fi is called the i-th coordinate function of F for i D 1; : : : ; m :
Conventions. We will use capital letters for vector-valued functions and the corresponding small letters for their
coordinate functions.
Example. Let F : R2 ! R3 be given by F .x ; y / D .x cos y ; x sin y ; 0/: Then the coordinate functions of F are
f1 ; f2 ; f3 : R2 ! R given by f1 .x ; y / D x cos y ; f2 .x ; y / D x sin y ; f3 .x ; y / D 0: So

D . f1 .x ; y/; f2 .x ; y/; f3 .x ; y//:


For convenience, this is sometimes written as F D . f1 ; f 2 ; f3 /:
F .x ; y /

Below we will use the abbreviation iff for if and only if.

2 Rn and F : Rn ! Rm ; we define qlim


F .q / D L iff lim k F .q / L k D 0; i.e. for every
!p
kq p k!0
" > 0; there exists a  > 0 such that 0 < kq pk <  implies k F .q / L k < ": We say F is continuous at p iff
lim F .q / D F . p/: We say F is continuous on a set S iff F is continuous at every p 2 S :
q!p
q
Theorem (Useful Inequality). For i D 1; : : : ; m ; we have j yi j  y12 C    C ym2  j y1 j C    C j ym j: (Just square
Definitions. For p

all three expressions and compare.)


Limit Theorem. Let L

D .l1 ; : : : ; lm / 2 Rm and p 2 Rn : For a function F : Rn ! Rm ; we have


lim F .q / D L if and only if lim fi .q / D li for all i D 1; : : : ; m :
q!p
q!p
133

In the case L D F . p/; that means F is continuous at p if and only if all m coordinate functions fi : Rn
continuous at p:

! R are

Proof. Let p 2 Rn be fixed. For q 2 Rn ; let x D kq pk and F .q / L D . y1 ; : : : ; ym /: (Note both x and the yi s
are functions of q :) Taking the i-th coordinate of both sides, we get fi .q / li D yi for i D 1; : : : ; m :

q
k
F .q / L k D lim y12 C    C ym2 : By the useful inequality
kq p k!0
x !0
and sandwich theorem, we get 0 D lim j yi j D lim j fi .q / li j for i D 1; : : : ; m : So lim fi .q / D li : Therefore,
q!p
x !0
kq p k!0
the limit of fi is li for i D 1; : : : ; m :
Conversely, if the limit of fi is li for all i D 1; : : : ; m ; then 0 D lim j fi .q / li j D lim j yi j for i D 1; : : : ; m :
kq p k!0
x !0
Adding these m limits, we get 0 D lim .j y1 j C    C j ym j/: By the useful inequality and sandwich theorem, we get
x !0
q
2
2
0 D lim y1 C    C ym D lim k F .q / L k: Therefore, the limit of F is L :
x !0
kq p k!0
If the limit of F is L ; then by definition, 0 D

lim

Example. In the previous example, the coordinate functions f1 .x ; y / D x cos y ; f2 .x ; y / D x sin y ; f3 .x ; y /


continuous at every .x ; y / 2 R2 ; so F D . f1 ; f2 ; f3 / is continuous on R2 by the theorem above.
Question. What does it mean to say F : Rn
at a point p 2 Rn ?

D 0 are

! Rm is differentiable? If it is differentiable, what is the derivative of F

As in the theorem above, it will turn out that F : Rn ! Rm is differentiable at a point p 2 Rn if and only if every
coordinate function fi : Rn ! R is differentiable at p: For every fi ; we know the meaning of the partial derivatives
@ fi
@f
; : : : ; i : How are these partial derivatives related to the derivative of fi ? Geometrically, for n D 1; the
@ x1
@ xn
derivative is the slope of the tangent line. So for n D 2; we may think the derivative should measure the slopes of
the tangent plane. Since two lines are needed to determine a plane, so the derivative in the case n D 2 should involve
two numbers. In fact, it will turn out that the derivative of a vector-valued function is a linear transformation and the
partial derivatives are the matrix entries of the linear transformation. First we will do a review.
Review of Some Linear Algebra and Calculus Facts
(1) Recall a linear transformation T : Rn

! Rm is a function satisfying

T .v C w/ D T .v/ C T .w/

T .cv/

and

D cT .v/

for all c

2 R; v; w 2 Rn :

With respect to the usual basis


e1

D .1; 0; 0; 0; : : : ; 0/;

e2

D .0; 1; 0; 0; : : : ; 0/;

e3

D .0; 0; 1; 0; : : : ; 0/;

:::

of Rn and Rm ; the linear transformation T has a matrix representation

0 x1 1 0 c11   
:
: ::
T @ :: A D @ ::
:
xn
cm1   

c1n

::
:

1 0 x1 1
A @ ::: A

cmn

0 x1 1
:
for every @ :: A 2 Rn :

xn

xn

To get the j -th column of the matrix, we apply T to e j : We have

0 c1 j 1
T .e j / D @ ::: A :
cm j

Taking the i-th coordinate of both sides, we get ti .e j / D ci j for all i


134

D 1; : : : ; m and j D 1; : : : ; n :

(2) The equation of a nonvertical line passing through .x 0 ; y0 / 2 R2 is y D y0 C m .x x 0 /: Observe that the function
T : R ! R defined by T .v/ D m v is a linear transformation. In terms of T ; the equation of the line can be put
in the form y D y0 C T .x x 0 /:
Similarly, using a normal vector, we get the equation of a nonvertical plane passing through .x 0 ; y0 ; z 0 / 2 R3 is
z D z 0 C c1 .x x 0 / C c2 . y y0 /: In terms of the linear transformation T : R2 ! R defined by T .v1 ; v2 / D
c1 v1 C c2 v2 the equation of the plane can be put in the form z D z 0 C T .x x 0 ; y y0 /:

In Rn ; there are translates of .n 1/-dimensional linear subspaces passing through any point analogous to lines
in R2 ; planes in R3 ; etc. These are called hyperplanes. Using a normal vector, the equation of a nonvertical
hyperplane passing through .x 0 ; y0 ; : : : ; z 0; w0 / 2 Rn can be put in the form w D w0 C T .x x 0 ; y y0 ; : : : ; z
z 0 /; where T : Rn 1 ! R is a linear transformation of the form T .v1 ; : : : ; vn 1 / D c1 v1 C    C cn 1 vn 1 :
F . x / F .a /
D F 0 .a /: This
x a
is equivalent to the existence of a nonvertical tangent line to the graph of F at .a ; F .a //: The equation of this line
is y D F .a / C F 0 .a /.x a / (or y D F .a / C T .x a / if we let T .v/ D F 0 .a /v:) Now observe that

(3) For F : R ! R to be differentiable at a ; the definition requires the existence of lim

x !a

F .x /
x !a
x

F .a /
a

lim

D F 0 .a / ()
()
()

 F .x /

F .a /
F 0 .a / D 0
x !a
x a
F .x / F .a / F 0 .a /.x a /
D 0
lim
x !a
x a
j F .x / F .a / T .x a /j D 0:
lim
x !a
jx a j
lim

So F is differentiable at a if and only if there is a linear transformation T : R ! R such that


lim

x !a

j F .x /

a /j

F .a / T .x
jx a j

D 0:

In fact, the matrix of T is . F 0.a / / ; which contains the slope of the tangent line to the graph of F at .a ; F .a //:
Similarly, let F : R2 ! R be a function. We expect F differentiable at a point to be equivalent to the existence
of a tangent plane at the point. From multivariable calculus, we learn that if the graph of F has a nonvertical
tangent plane at .a1 ; a2 ; F .a1 ; a2 //; then the equation of this plane is z D F .a1 ; a2 / C c1 .x a1 / C c2 . y
@F
@F
a2 /; where c1 D
.a1 ; a2/; c2 D
.a ; a /: (Note c1 is the slope of the plane in the x-direction through
@x
@y 1 2
.a1 ; a2; F .a1; a2// and c2 is the slope of the plane in the y-direction through the point.) So we expect F to be
differentiable at .a1 ; a2 / if and only if there exists a linear transformation T : R2 ! R such that


F .x ; y/

lim

.x ; y /!.a1;a2/

F .a 1 ; a 2 /

k.x

a1 ; y

T
a2 /k

x
y

a1
a2

D0

and we also expect the matrix of T to be . @@Fx .a1 ; a2 / @@Fy .a1 ; a2 / / ; which contains all the slope informations of
the tangent plane. In general, this is how differentiability of vector-valued functions is defined.
Definitions. A vector-valued function F : Rn
transformation T : Rn ! Rm such that
lim

x !a

k F .x /

F .a / T .x
kx a k

a /k

D0

! Rm is differentiable at a D .a1; : : : ; an / 2 Rn iff there is a linear

or lim

t !O

k F .a C t /

F .a /
kt k

T .t /k

D0

by substituting t

Dx

a :

If such a T exists, we call it the derivative of F at a and denoted it by D F .a /:


Question. How do we come up with the linear transformation to check differentiability at a point? (This is answered
by the next theorem.)
135

! Rm is differentiable at a D .a1 ; : : : ; an / 2 Rn and f1 ; : : : ; f m : Rn ! R are


@ fi
.a / exist for i D 1; : : : ; m and j D 1; : : : ; n : In fact,
the coordinate functions of F ; then all the partial derivatives
@x
Differentiation Theorem. If F : Rn

the matrix of D F .a / with respect to the usual bases is

0 @ f .a /
@x
B
@ ::

:::
::
:
:
@ fm
@ x1 .a / : : :
1

@ f1
@ xn .a /

::
:

@ fm
@ xn .a /

1
C
A:

(This matrix is called the Jacobian matrix of F at a and is denoted by F 0 .a /:)


Proof. Since F is differentiable at a ; there exists a linear transformation T D D F .a / satisfying the condition in the
definition of differentiable function at a point. Let x D a C he j D .a1 ; : : : ; a j C h ; : : : ; an /; then

k F .x / F .a / T .x
0 D lim
x !a
F .a C hekx/ aFk.a /
j

D hlim
!0
h

F .a C he / F .a /
j

D hlim
!0
h

F
.
a
C hej /
T .e j /
) lim
a /k

h!0

hT .e j /

F .a /

D T .ej /:

Taking the i-th coordinate of both sides, we get

@ fi
f .a C he j /
.a / D lim i
h!0
@xj
h
Since ti .e j / is the .i; j /-entry of the matrix of T ;
i

D 1; : : : ; m and j D 1; : : : ; n :

fi .a /

D ti .ej /:

@ fi
.a / exists and its value is the .i; j /-entry of T
@xj

D F .a / for

Example. We will show that the function F : R2 ! R defined by F .x ; y / D x 2 C y 2 is differentiable at .0; 0/: We have
@f
@f
D
2x and
D 2y: By the differentiation theorem, if F is differentiable at .0; 0/; then T D D F .0; 0/ has matrix
@x
@y
k F .x ; y/ F .0; 0/ T .x 0; x 0/k D lim p x 2 C y2 D 0:
. @@Fx .0; 0/ @@Fy .0; 0/ / D . 0 0 / : Now lim
.x ; y /!.0;0/
.x ; y /!.0;0/
k.x ; y/ .0; 0/k
So, F is differentiable at .0; 0/:
Remarks. (1) The contrapositive of the differentiation theorem asserts that if some partial derivative does not exist at
a ; then the function is not differentiable at a :
(2) The converse of the differentiation theorem is false. It may happen that all partial derivatives exist at a point, but
the function is not differentiable there. For example, take
F .x ; y / D

n0

if x D 0 or y
1 otherwise.

D 0,



D 0 and F .0; y/ D 0; @@ Fx .0; 0/ D ddx F .x ; 0/ x D0 D 0 and @@Fy .0; 0/ D ddy F .0; y/ y D0 D 0: By the
differentiation theorem, if F is differentiable, then T D D F .0; 0/ has matrix . @@Fx .0; 0/ @@Fy .0; 0/ / D . 0 0 / :
k F .x ; x / F .0; 0/ T .x 0; x 0/k D p 1 does not have a limit. So, F is not
Since along .x ; x / ! .0; 0/;
2
k.x ; x / .0; 0/k
Since F .x ; 0/

2x

differentiable at .0; 0/:

Question. If F : Rn ! Rm is differentiable at a 2 Rn ; is it continuous at a? (Yes, we need a lemma before we can


prove it.) In remark (2), it is easier to see F is not differentiable at .0; 0/ by noticing that F is not continuous at .0; 0/
than checking the definition of differentiability.
136

Definitions. Let M be the m  n matrix of a linear transformation T : Rn ! Rm with entries ci j .i D 1; : : : ; m I j D


1; : : : ; n / under the usual bases of Rn and Rm : Writing all thesentries ci j in a row, we may consider M 2 Rmn and
define the norm of T and the norm of M to be kT k D k M k D

6 6 ci2j :
i D1 j D1

Continuity Lemma. For every linear transformation T : Rn ! Rm ; we have kT .v/k


v D .x 1 ; : : : ; x n / 2 Rn : Consequently, lim T .v/ D 0 by the sandwich theorem.
v!0

Proof. Let e1 ; : : : ; en be the usual basis of Rn and ci j


inequality and the Cauchy-Schwarz inequality,


kT .v/k D T

 kT kkv k for every

D ti .ej / be the entries of the matrix of T : By the triangle

s
rn rn
n
n

m n
2
6 x i ei D 6 x i T .ei /  6 jx i jkT .ei /k  6 x i 6 kT .ei /k2 D kv k 6 6 ci2j D kT kkv k:
i D1
i D1
i D1
i D1
i D1
i D1 j D1
n

! Rm is differentiable at a 2 Rn ; then F is continuous at a :


k F .x / F .a / T .x a /k D 0: So
Proof. Since F is differentiable at a, by definition, lim
x !a
kx a k
k F .x / F .a / T .x a /k kx a k D 0:
lim k F .x / F .a / T .x a /k D lim
x !a
x !a
kx a k
By the lemma, lim T .x a / D 0: By the triangle inequality, k F .x / F .a /k  k F .x / F .a / T .x a /kCkT .x
x !a
By the sandwich theorem, lim k F .x / F .a /k D 0: Therefore, lim F .x / D F .a /:
x !a
x !a
Theorem. If F : Rn

a /k:

Question. Are there easier methods to see if a function F : Rn ! Rm is differentiable at a point than to check the
definition? (The C 1 theorem below provides an easier way to check differentiabilty.)
Theorem. F : Rn ! Rm is differentiable at a 2 Rn if and only if all coordinate functions fi : Rn
differentiable at a : Also, D F .a / D . D f1 .a /; : : : ; D fm .a // in the case the functions are differentiable.
Proof. Since the i-th coordinate of
lim

x !a

F .x /

F .a / T . x
kx a k

a/

F .x /

DO

F .a / T .x
kx a k

if and only if

a/

lim

x !a

is

fi . x /

fi . x /

fi .a / ti .x
kx a k

fi .a /

kx

ti . x
ak

a/

a/

! R of F are

; by the limit theorem,

DO

for all i

D 1; : : : ; m :

So F is differentiable at a with D F .a / D T if and only if all fi s are differentiable at a with D fi .a / D ti ; i.e.


D F .a / D T
Definition. F : Rn

D .t1 ; : : : ; tm / D . D f1 .a /; : : : ; D fm .a //:

! Rm is continuously differentiable (or C 1 ) at a 2 Rn iff all partial derivatives @@xfi

exist in some

neighborhood of a (i.e. in some open ball B .a ; r /) and are continuous at a : Furthermore, we say F is C 1 near a iff F
is C 1 at every point in some open ball B .a ; r0 /: (In more advanced courses, we also say F is C n near a iff all partial
derivatives of order n or below are continuous in a neighborhood of a :)

! Rm is C 1 at a D .a1 ; : : : ; an /; then F is differentiable at a :


Proof. By the previous theorem, it suffices to show each coordinate function f : Rn ! R is differentiable at a : For
the derivative, the differentiation theorem suggests the linear transformation T : Rn ! R defined by
0 h1 1
@f
@f
:
@f
@f
T .h 1 ; : : : ; h n / D . @ x .a / : : : @ x .a / / @ :: A D
.a /h 1 C    C
.a /h n :
@x
@x
C 1 Theorem. If F : Rn

hn

137

Since f is C 1 at a ; all

@f
exist in some neighborhood B .a ; r /: For x
@xj
f .a / D f .x 1 ; : : : ; x n /

f .x /

f .x 1 ; x 2 ; : : : ; x n /

CC

f .a1 ; x 2 ; : : : ; x n /

f .a 1 ; a 2 ; : : : ; a n 1 ; x n /

D .x1 ; : : : ; xn / 2 B.a ; r /;

f .a1 ; : : : ; an /

f .a1 ; x 2 ; : : : ; x n /

f .a 1 ; a 2 ; : : : ; x n /

f .a1 ; a2 ; : : : ; an 1 ; an / :

Applying the mean value theorem, we have


f .x /

f .a / D

@f
.c1 /.x 1
@ x1

a1 / C    C

@f
.cn /.x n
@ xn

an /;

where c j is some point on the segment joining .a1 ; : : : ; x j ; : : : ; x n / and .a1 ; : : : ; a j ; : : : ; x n /: Then

j f .x /

a /j

f .a / T . x
kx a k

n
6
j D1

@f
.c j /
@xj

kx


@f
.a / .x j
@xj
ak


a j /


n
 j6D1 @@xf

.c j /

jx
@f
.a / j
@xj
kx

aj j
:
ak

| {z }
1

! a ; we have cj ! a : Since @@xf is continuous at a for j D 1; : : : ; n ; the right side of the inequality will go to
j
0 as x ! a : By the sandwich theorem, the left side will also go to 0 as x ! a : Therefore, f is differentiable at a :
As x

Examples. (1) Let F : R2 ! R2 be defined by F .x ; y / D .x 2 C y 2 ; sin x y /: Is F differentiable at .2; /? If so, what
are F 0 .2; / and D F .2; /?
Solution. The coordinate functions of F are f1 .x ; y / D x 2 C y 2 and f2 .x ; y /

@ f1
@x

D 2x ;

@ f1
@y

D 2y;

@ f2
@x

@ f2
@y

D y cos x y;

D sin x y: Now
D x cos x y

are continuous on R2 (in particular at .2; /). So F is C 1 at .2; /: By the C 1 theorem, F is differentiable at .2; /:
By the differentiation theorem,
F 0 .2; / D

 @ f .2; /
1

@x
@ f2
@ x .2; /

x   4
D F .2; /
D 
y

 

4 2
D  2 ;
  x   4x C 2 y 
2
D  x C 2y :
2
y
@ f1
@ y .2; /
@ f2
@ y .2; /

(2) Let T : Rn ! Rm be a linear transformation. Is T differentiable on Rn ? If so, what is the derivative of T at an


arbitrary a 2 Rn ?
Solution. (If T is differentiable, then the derivative of T is also a linear transformation. What would you guess the
derivative is?) We will show T is differentiable at every a 2 Rn and DT .a / D T : Since T .x a / D T .x / T .a /; so
lim

x !a

kT .x /

T .a / T .x
kx a k

a /k

D 0;

i.e.

DT .a / D T :

Remarks. The addition function P : R2 ! R defined by P .x ; y / D x C y is a linear transformation. The subtraction


function Q : R2 ! R defined by Q .x ; y / D x y is also a linear transformation. So both are differentiable and
D P .a / D P ; D Q .a / D Q for every a 2 Rn :
138

! R defined by R.x ; y/ D x y is differentiable and find D R.a /:


@R
(Note R is not a linear transformation!) We will check that R is C 1 at every a 2 R2 : Since
D y and
@x

(3) Show the multiplication function R : R2

Solution.
@R
D x are continuous on R2; by the C 1 theorem, R is differentiable everywhere. By the differentiation theorem,
@y

So D R .a /

x
y

R 0 .a1 ; a2 / D . @@Rx .a1 ; a2 /

D . a2

a1 /

x 
y

@ R .a ; a / /
@y 1 2

D . a2

a1 / :

D a2 x C a1 y :

(4) Show the division function S .x ; y / D x = y is differentiable for y 6D 0 and find DS .a / for a
a2 6D 0: (Although S : R  .R n f0g/ ! R; the definition of differentiation is the same!)

D .a1; a2 / 2 R2 with

Solution. For a D .a1 ; a2 / 2 R2 with a2 6D 0; we will check S is C 1 at a : This is true because the partial derivatives
@S 1
@S
x
and
D
D
exist in some neighborhood of a and are continuous at a : Now
@x
y
@y
y2

So DS .a1 ; a2 /

 
x
y

S 0 .a / D . @@ Sx .a /

a1
a22

1
a2

  x  D a2 x
y

a1 y
a22

@ S .a / /
@y

1
a2

D |DG

:

Theorem (Chain Rule). If F : Rn ! Rm is differentiable at a


b D F .a / 2 Rm ; then G  F : Rn ! Rk is differentiable at a and
D .G  F /.a /

a1
a22

  D F .a /
{z
}

F .a /

Rn and G : Rm

. @@Hs
So

@H
@s

@H
@t

x .s ; t /; y .s ; t / ; and G : R2

{z

matrix multiplication

! R are differentiable, then H D G  F :


/ D H 0.s ; t / D G 0 F .s ; t / F 0 .s ; t / D . @@Gx

D @@Gx @@ xs C @@Gy @@ sy and @@Ht D @@Gx @@xt C @@Gy @@ yt :

Rk is differentiable at


. G  F /0 .a / D G 0 F .a / F 0 .a / :

composition of linear transformations

Special Case: If F : R2 ! R2 ; F .s ; t /
R2 ! R is differentiable and

@G
@y

 @x

@s
@y
@s

@x
@t
@y
@t


D F .a /; h D F .x / F .a / D F .x / b; T D DG F .a / and S D D F .a /: Since F is
S .x a /
F .x / F .a / S .x a /
D
! O as x ! a : Also, since G is differentiable
kx a k
kx a k
G .b C h / G .b/ T .h /
at b; J D
! O as h ! O : For h 6D O ; let
kh k

K D G  F .x / G  F .a / DG F .a /  D F .a /.x a /

D G.b C h / G.b/ T S .x a /

D |G.b C h / {zG.b/ T .h}/ C |T h S{z.x a / }:
D J khk
Dkx a kT . I /
(
k K k  k J k kh k C kT . I /k if h 6D 0 Now
By the triangle inequality, we have
kx a k kT . Ik/xk a k
if h D 0.
kh k
k F .x / F .a /k k F .x / F .a / S .x a /k C k S .x a /k  k I k C k S k:
kx a k D kx a k 
kx a k
kx a k
Proof of Chain Rule. Let b
h
differentiable at a ; so I D

139

As x

! a ; we get h D F .x /

F .a / ! O ; J

! O ; I ! O ; T . I / ! O and kxkh ka k bounded, so kxk K ka k ! 0

by the sandwich theorem. Therefore, G  F is differentiable at a :


Theorem (Differentiation Formulas). If f ; g : Rn
(provided g .a / 6D 0) are differentiable at a and

! R are differentiable at a 2 Rn ; then f C g ; f

g ; f g and f =g

D . f  g /.a / D D f .a /  Dg .a /;
D . f g /.a / D g .a / D f .a / C f .a / Dg .a /;
f
g .a / D f .a / f .a / Dg .a /
D
:
.a / D
g
g .a /2

Equality of Second Derivatives


We learned in multivariable calculus that
Second Derivative Theorem. Let f : R2

@2 f
@x@ y

D @@y@fx
2

frequently. Here is a theorem about when this happens.

! R be such that @@ xf ; @@ fy ; @@y@fx


2

exist at every point of some ball B . p; r /

@2 f
@2 f
@2 f
@2 f
is continuous at p: Then
. p/ exists and
. p/ D
. p/:
@ y@ x
@x@ y
@x@ y
@ y@ x
@2 f
@2 f
@ @ f 
Proof. Let p D .a ; b/ and w D
. p/: By checking the definition of limit, we will prove
. p/ D
.a ; b/
@ y@ x
@x@ y
@x @y
.@ f =@ y /.a C h ; b/ .@ f =@ y /.a ; b/
D hlim
D w:
!0
h
@2 f 0
For every " > 0; since lim
. p / D w by continuity at p; there is 0 > 0 such that k p0 pk < 0 implies
p0 ! p @ y@ x

@ 2 f 0
@ y@ x . p / w < 2" : Let  D min.0 ; r /: Consider any point q D .a C h ; b C k/ in B. p; /: Applying the mean value
@f 0
theorem to u .x / D f .x ; b C k / f .x ; b/ and v. y / D
.a ; y / below, we have
@x
and

1 D f .a C h ; b C k /

D u .a C h /
D

v.b C k /

f .a C h ; b/

f .a ; b C k / C f .a ; b/

D @ x .a 0 ; b C k/ @@ xf .a 0 ; b/ h

@2 f 0 0
v.b/ h D v 0 .b0 /kh D
.a ; b /kh
@ y@ x

u .a / D u 0.a 0 /h

@ f

1
D .a 0 ; b0 / in B. p; / with a 0 betwee a and a C h and b0 between b and b C k: Then kh
@f
1 @f
lim
D
.a C h ; b/
.a ; b/: So 0 < jh j <  implies
k !0 k
@y
@y
.@ f =@ y/.a C h ; b/ .@ f =@ y/.a ; b/
1 "
w D lim
w  < ":

k !0 kh
h
2
for some p0

"
w < : Note
2

@2 f
is not continuous at p; then the conclusion may become false, i.e. the second derivatives above may
@ y@ x
x y .x 2 y 2 /
not be equal even if they exist. Here is an example. Let p D .0; 0/: Define f .0; 0/ D 0 and f .x ; y / D
x 2 C y2
for .x ; y / 6D .0; 0/:
Remarks. If

140

If .x ; y /

6D .0; 0/; then @@ xf .x ; y/ D

y .x 4 C 4x 2 y 2 y 4 /
@f
x .x 4 4x 2 y 2 y 4 /
and
.x ; y / D
: At .0; 0/;
2
2
2
.x C y /
@y
. x 2 C y 2 /2

@f
f .h ; 0/ f .0; 0/
.0; 0/ D lim
h !0
@x
h
So

D0

and

@f
f .0 ; k / f .0 ; 0 /
.0; 0/ D lim
k !0
@y
k

D 0:

@f @f
exist everywhere. Next, if .x ; y / 6D .0; 0/; then
;
@x @y
@2 f
x 6 C 9x 4 y 2 9x 2 y 4
.x ; y / D
@ y@ x
. x 2 C y 2 /3

y6

D @@x @fy .x ; y/;


2

.0; 0/ and has limit 1 as .0; y / ! .0; 0/: So these second derivatives cannot
@2 f
.@ f =@ y /.h ; 0/ .@ f =@ y /.0; 0/
h 0
be continuoust at .0; 0/: Finally, we have
.0; 0/ D lim
D hlim
D 1 and
h!0
!0 h
@x@ y
h
2
@ f
.@ f =@ x /.0; k / .@ f =@ x /.0; 0/
k 0
.0; 0/ D lim
D
lim
D 1:
k !0
k !0
@ y@ x
k
k
which has limit 1 as .x ; 0/

Inverse Functions
In one variable calculus, if y
then

1D

d
x
dx

For example, if we restrict y


we get

D ddx .g  f /.x / D g 0. f .x // f 0 .x / D dd xy dd yx
D

f .x /

D sin x to the interval .

d
arcsin y
dy
However, if we restrict y
since about x

D g . y/; which is also differentiable,


.
) dd xy D 1 dd xy :

f .x / is differentiable and has an inverse x

f .x /

D dd xy D

1
dy
dx

 
; /; then it has an inverse x
2 2

D cos1 x D  p

1
1

y2

D g . y/ D arcsin y: So

D sin x to an open interval containing 2 ; then there cannot be any inverse function,

D 2 ; the function is not injective.

In multivariable calculus, we are familiar with changing coordinates,


such as F .r; / D .x ; y /; where x D r cos 
p
and y D r sin : We can invert by G .x ; y / D .r; /; where r D x 2 C y 2 and  D arctan. y =x /; provided x 6D 0:

! Rn and q 2.Rn ; how can we know if the function has an inverse function near q?
dx
dy
If so, what is the analogous formula for
D
1
?
dy
dx
Question. For a function F : Rn

As the tangent hyperplane at q is close to the graph near q ; the injectivity of F nearq should be related to the
injectivity of D F .q /; which is equivalent to det F 0 .q / 6D 0: If F 1 exists, then F 1 F .x / D x : Applying the chain
0

0

1
rule, we can get the formula F 1 F .x / F 0.x / D I ; which implies F 1 F .x / D F 0 .x / :
Inverse Function Theorem. If F : Rn ! Rn is C 1 near q 2 Rn and det F 0 .q / 6D 0; then there exists an open ball
0

1
B .q ; r / such that F restricted to B .q ; r / has a C 1 inverse function F 1 and F 1 F .x / D F 0 .x /
for every
x 2 B .q ; r /:
We will give a sketch of the proof of the inverse function theorem in the appendix.
Example. In R2 ; define .u ; v/-coordinates by u D e y C x ; v D e x y ; where .x ; y / is the usual rectangular coordinates.
Is it possible to express .x ; y / as a differentiable function of .u ; v/ near the origin .x ; y / D .0; 0/? If so, what are
@x @x @y @y
; ; ; when .x ; y / D .0; 0/; i.e. at .u ; v/ D .1; 1/?
@ u @v @ u @v
141

Solution. Consider F : R2

! R2 given by F .x ; y/ D .u ; v/: Now @@ ux D 1;

@u
@y

D e y ; @@vx D ex ; @v
D 1 are
@y

e0
D 2 6D 0: By the inverse


D .0; 0/: So F is C 1 near q D .0; 0/: Also, det F 0 .0; 0/ D e10 1
function theorem, near q D .0; 0/; F has a C 1 inverse function F 1 .u ; v/ D .x ; y /: Now applying the formula in the
theorem at q D .0; 0/; we have
 @ x .1; 1/ @ x .1; 1/ 
 @u
1
@u
1 0
0 .q / 1 D @ x .0; 0/ @ y .0; 0/
@u
@v
D
.
F
/
.
F
.
q
/
/
D
F
:
@y
@y
@v
@v
|{z}
@ u .1; 1/ @v .1; 1/
@ x .0; 0/ @ y .0; 0/
continuous near q

Then

 @ x .1; 1/
@u
@y
@ u .1; 1/

@x
@v .1; 1/
@y
@v .1; 1/

 
D e10

@x
1
.1; 1/ D ;
@u
2
Remark. Note in the above example,

e0
1

D.1;1/

1
2

1
e0

@x
1
.1; 1/ D ;
@v
2
@x
@u

6D 1

. @u
@x

e0
: So
1

@y
1
@y
.1 ; 1 / D
and
.1 ; 1 / D
@u
2
@v

1
:
2

Implicit Differentiation
dy
by implicit differentiation, which assumes
In one variable calculus, given an equation F .x ; y / D 0: We can find
dx
the equation can be solved for y as a differentiable function of x : Below we will discuss implicit differentiation for
vector-valued functions. Recall O D .0; : : : ; 0/ is the origin of Rm : Consider a function F : Rn ! Rm with coordinate
functions f1 ; : : : ; f m : Rn ! R: The vector equation F .x / D O is equivalent to the following system of m equations
in n variables:
f 1 . x 1 ; : : : ; x n / D 0;

::
::
:
:
f m . x 1 ; : : : ; x n / D 0:

In the case when n > m ; i.e. more variables than equations, it is a common experience that we may be able to solve
for m of the variables in terms of the other n m variables. For example, in the linear case when fi .x 1 ; : : : ; x n / D
ai1 x 1 C    C ain x n (the ai j s are constants), the system becomes
a11 x 1 C    C a1n x n

D0

am1 x 1 C    C amn x n

D 0:

::
:

::
:

Moving the terms involving x m C1 ; : : : ; x n to the right sides, we have


a11 x 1 C    C a1m x m

am1 x 1 C    C amm x m

::
:

We can put this in matrix form as

0 a11   
@ ::: : : :
| am1 {z 

.a1m C1 x m C1 C    C a1n x n /;
::
:

.amm C1 x m C1 C    C amn x n /:

1 0 x1 1 0
:: A @ :: A D @
:
:
amm
xm
}
a1m

call this matrix M

142

.a1m C1 x m C1 C    C a1n x n / 1
::
A:
:
.amm C1 x m C1 C    C amn x n /

0 x1 1
:
We may solve for @ :: A by multiplying by M

if M is invertible (or equivalently if det M

xm
of the variables x 1 ; : : : ; x m will be a function of the remaining n

6D 0). In that case, each

m variables x m C1 ; : : : ; x n :

To deal with the case when the functions are not linear, we should first look at the condition det M
@ fi
of the fi s. Since ai j D
; so
@xj

0 @f   
@x
B
() det @ ::: : : :
@f

@x
1

det M

6D 0

m
1

@ f1
@ xm

::
:

@ fm
@ xm

6D 0 in terms

1
C
A 6D 0:

The matrix on the right above is important in the sequel. So we will introduce the following notation.

 m and p 2 Rn ; then we denote


   @@xf . p/ 1
:: C
::
:
: A
@f
   @ x . p/

Notation. If f1 ; : : : ; fm are functions of the variables x 1 ; : : : ; x n with n

0 @ f . p/
@x
@. f1 ; : : : ; f m /
. p/ D det B
@ ::
1

@.x 1 ; : : : ; x m /

@ fm
@ x1 . p /

and call it the Jacobian of f1 ; : : : ; fm with respect to x 1 ; : : : ; x m at p:


Question. In the general case, when a function F : Rn ! Rm (n > m) has a root p 2 Rn ; i.e. F . p/ D O ; and F is
C 1 near p (recall that means F is C 1 in some open ball B . p; r /), can we solve the vector equation F .x / D O
pfor some2
2
2
of the variables near p? For example, we can
solve
the
equation
x
C
y
25
D
0
near
.
3
;
4
/
to
get
y
D
25 x :
p
Near .3; 4/, we can solve it to get y D
25 x 2 : However, near .5; 0/; y cannot be expressed as a differentiable
function of x :
The following theorem tells us a sufficient condition when we can solve a (vector) equation for some of its
variables as a differentiable function of the other variables.
Implicit Function Theorem. Let

! Rm .n > m/ be C 1 near p D . p1; : : : ; pn / 2 Rn ;


F . p/ D O and
@. f1 ; : : : ; fm /
. p/ 6D 0:
@.x ; : : : ; x /

(1) F : Rn
(2)
(3)

Then there exist an open ball B . p; r / and a C 1 function G : B  ! Rm such that for every x
B . p; r /;
F .x 1 ; x 2 ; : : : ; x n / D O
if and only if
.x 1 ; : : : ; x m / D G .x m C1 ; : : : ; x n /:

(Here B  D f.x m C1 ; : : : ; x n /
coordinate space.)

2 Rn

: .x 1 ; : : : ; x n /

D .x 1 ; : : : ; x n / 2

B . p; r /g is the projection of B . p; r / onto the last n

In this case, we say x 1 ; : : : ; x m are implicit functions of x m C1 ; : : : ; x n near p:


Proof. Let k D n m : Define H : Rm Ck ! Rm Ck by H .x ; y / D . F .x ; y /; y / for x D .x 1 ; : : : ; x m /
.x m C1; : : : ; x n / 2 Rk : Since F is C 1 near p; H is also C 1 near p: From row operations, we can see that

F 0 . p/
det H 0. p/ D det
O

2 Rm ; y D

f1 ; : : : ; f m /
D @.
. p/ 6D 0:
@.x ; : : : ; x /
1

By the inverse function theorem, there exists an open ball B . p; r / such that H restricted to B . p; r / has a C 1 inverse
function H 1 : In order to be the inverse of H near p; we must have H 1 .x ; y / D . K .x ; y /; y / for some C 1 function
K defined near H . p/ D . O ; pm C1; : : : ; pn /:
143

Let  : Rm Ck ! Rm be the function .x ; y / D x for x 2 Rm ; y 2 Rk ; then   H D F : Define G . y / D K . O ; y /:


Since K is C 1 near H . p/ D . O ; pm C1; : : : ; pn /; it follows G will be C 1 near . pm C1 ; : : : ; pn /: Now if F .x ; y / D O
for .x ; y / 2 B . p; r /; then H .x ; y / D . O ; y /: So

.x ; y / D H 1. O ; y / D . K . O ; y /; y / D .G . y /; y /

D G. y/; then
F .x ; y / D F .G . y /; y / D F . K . O ; y /; y / D F . H

H)

D G. y/:

Conversely, if x

. O ; y // D .  H /. H 1 . O ; y // D . O ; y / D O :

Examples. (1) Given .1; 2/ is a solution of the equation x 6 C x 2 y 2 C y 6

D 69: Find dd yx at .1; 2/:

Solution. (Step 2) We perform implicit differentiation to get


6x 5 C 2x y 2 C 2x 2 y
Solving for

dy
dy
; we get
dx
dx

dy
dx

C 6y5 dd xy D 0:

dy
6x 5 C 2x y 2
: At .1; 2/; we get .1; 2/ D
2
5
2x y C 6y
dx

1
:
14

Comments: To do implicit differentiation means to apply chain rule to the terms x 2 y 2 and y 6 assuming y is a
differentiable function of x near .1; 2/: To see y is a function of x near .1; 2/, you may say we can sketch the graph
of the equation and see that a small piece of the graph near .1; 2/ is the graph of some function (because every vertical
line cuts the piece in exactly one point.) This is not too rigorous. Further, why is it differentiable? Also, it is hard to
draw graphs when there are more than three variables!
(Step 1) To justify y is a differentiable function of x near .1; 2/; we can use the implicit function theorem as follow.
Consider F : R2 ! R (here n D 2; m D 1) given by F .x ; y / D x 6 C x 2 y 2 C y 6 69 and p D .1; 2/: Then
F .1; 2/ D O : Near p;
@F
@F
D
6x 5 C 2x y 2
and
D 2x 2 y C 6y5
@x
@y
are continuous. So F is C 1 near p

D .1; 2/: Now @@Fy .1; 2/ D 196 6D 0: By the implicit function theorem, near .1; 2/;

there is a C 1 (in particular, differentiable) function G such that


x 6 C x 2 y2 C y6

69 D 0 if and only if

D G.x /:

D 0; w4 C x 4 C y4 C z4 D 18; is it possible to express .x ; y/ as a differentiable function of


@x
@y
.w; z / near the solution p D .w; x ; y ; z / D . 1; 0; 1; 2/? If so, what is
. p/ and
. p /?
@w
@w
Solution. (Step 1) Consider F : R4 ! R2 given by
F .w; x ; y ; z / D . f1 ; f2 /; where f1 .w; x ; y ; z / D w x yz and f2 .w; x ; y ; z / D w4 C x 4 C y 4 C z 4 18:
Since near p D . 1; 0; 1; 2/;
(2) For the system w x yz

@ f1
@w

D x yz; @@fx1 D w yz; @@fy1 D w x z; @@fz1 D w x y; @@wf 2 D 4w3; @@ fx2 D 4x 3 ; @@fy2 D 4y3 ; @@fz2 D 4z3
@ f . p/ @ f . p/

@. f1 ; f 2/
2 0
@
x
@
y
1



are continuous, so F is C near p: Now
. p/ D @ f
D
D 8 6D 0: By the implicit
@.x ; y /
. p/ @@fy . p/ 0 4
@
x
function theorem, near p; there is a C 1 (hence differentiable) function G such that F .w; x ; y ; z / D 0 iff .x ; y / D
G .w; z /:

144

(Step 2) Since near p; .x ; y / is a differentiable function of .w; z /; holding z constant in differentating the equations
with respect to w; we have
0D

@y
@
@
@x
.0/ D
.w x yz / D x yz C w yz C w x z ;
@w
@w
@w
@w

@
@ 4
@x
@y
.18/ D
.w C x 4 C y 4 C z 4 / D 4w3 C 4x 3
C
4y 3
:
@w
@w
@w
@w
@x
@y
At p D .w; x ; y ; z / D . 1; 0; 1; 2/; the equations simplifies to 0 D 2
. p/ and 0 D 4 C 4 . p/: Therefore,
@w
@w
@x
@y
. p/ D 0 and
. p / D 1:
@w
@w
0D

Appendix : Proof of the Inverse Function Theorem

! B.0; r / be such that k F .x / F . y/k  12 kx yk for every


1
x ; y 2 B .0; r /: Then F has a unique fixed point w 2 B .0; r /; i.e. F .w/ D w: (Here may be replaced by any
2

Contractive Mapping Theorem. Let F : B .0; r /

positive constant less than 1.)

Proof. Since F .x / D x and F . y / D y imply kx


i.e. x D y : So there is at most one fixed point.

y k D k F .x /

F . y /k 

1
kx
2

y k; which implies kx

y k D 0;

To get a fixed point, let x 1 D 0 and x nC1 D F .x n / 2 B .0; r / for n D 1; 2; 3; : : : : Now kx n


x nC1 k D
1
1
k F .xn 1 / F .xn /k  2 kxn 1 xn k: Repeating this n 1 times, we easily get kxn xnC1k  2n 1 kx1 x2 k: So
for n < m ;

kx n

x m k  kx n

x nC1 k C    C kx m

xm k 

1
2n 1


C    C 2m1 2 kx1

x2 k 

1
kx 1
2n 2

x 2 k;

which implies the sequence x 1 ; x 2 ; x 3 ; : : : is a Cauchy sequence. Hence the sequence has a limit w: Now kx n k
implies kwk  r; i.e. w 2 B .0; r /:
1
Next from 0  lim k F .x / F .a /k  lim kx a k D 0; we see that F is continuous. So x n
x !a
x !a 2
x nC1 D F .x n / ! F .w/; but x nC1 ! w: Therefore, F .w/ D w:

r

! w implies

Theorem (Mean Value Inequality). If F : Rn ! Rm is differentiable at every point of B . p; r / and k F 0 .a /k  k for


every a 2 B . p; r /; then k F .x / F . y /k  k kx y k for every x ; y 2 B . p; r /:
Proof. For x ; y

k F .x /

2 B. p; r / and t 2 [0; 1]; let .t / D .1

F . y /k D kg .1/

g .0/k  kg 0.c/k.1

t / y C t x and g .t / D F .t / : By exercise 13(b),

0/ D k F 0 .c/ .x

y /k  k F 0 .c/

Proof of Inverse Function Theorem. First we consider the special case q

kkx

y k  k kx

y k:

D O ; F . O / D O and F 0 . O / D I :

(Step 1) Since F is C 1 near O ; all the partial derivatives are continuous near O : Considering the n 2 partial derivatives
1
2
as coordinate functions, we see that the function F 0 : Rn ! Rn is continuous near x D O : So for " D > 0;
2
1
0
0
0
n
there is a  > 0 such that kx
O k <  implies k F .x / F . O /k D k F .x / I k < : Let G : R ! Rn be
2
145

D F .x / x : On B. O ; /; kG0 .x /k D k F 0.x / I k < 12 : By the mean value inequality, we get


1
G . y /k  kx y k for every x ; y 2 B . O ; /: By triangle inequality, for every x ; y 2 B . O ; /;
2

given by G .x /

kG .x /

k F .x /

F . y /k

D kx

G. y/

G .x /

k  kx

yk

kG . y /

G . x /k

 12 kx

y k:

./

(Step 2) Let 0 < r < =2: For each t 2 B . O ; r /; we will find a s 2 B . O ; / such that F .s / D t : This is the
same as finding a fixed point in B . O ; / for the function E : Rn ! Rn defined by E .s / D s C t F .s / : Now
k E 0 .s/k D k I F 0 .s/k  12 for every s 2 B. O ; /: By the mean value inequality, k E .x / E . y/k  12 kx yk
1
for every x ; y 2 B . O ; /: Since 2r < ; so for every s 2 B . O ; 2r /; we have k E .s / E .0/k  ks k: Then
2
1
k E .s/k  2 jsj C k E .0/k  r C jt j  2r: So E takes B. O ; 2r / to B. O ; 2r /: By the contractive mapping theorem,
there is a unique s 2 B . O ; 2r /  B . O ; / such that E .s / D s ; then F .s / D t : Note there is exactly one such
s 2 B . O ; / satisfying F .s / D t by (*).
(Step 3) Now for every t 2 B . O ; r /; define F 1 .t / D s ; where s 2 B . O ; / is as in step 2. In particular, F 1 . O / D O :
By (*), we have (**) k F 1 .t / F 1 .t0 /k  2kt t0 k; which implies F 1 is continuous. Since F is differentiable at
q D O ; F . O / D O and D F . O / D I ; so
r .x /
By (**), as x

D kF
1

F.O/

D F . O /.x

kx k

O/

F .x / x
kx k

! O;
kF

So F

F .x /

.x /

F .x /

.O/

kxk
r .x /kx k
kx k

I .x
F

O /k
F .x /

k

D kF

as x

! O:

r . x /k x k
O xk
kx k
r .x /kx k/ F .x /k
D 2kr .x /k
kx k

F .x /

 2k. F .x /

!O

is differentiable at O :

! 0:

(General Case) Define ; : Rn ! Rn by .x / D x q and . y / D D F .q / 1 y F .q / : Note that  and are


1
C 1 functions and have C 1 inverses given by  1 .x / D x C q and
. y / D D F .q /. y / C F .q /: Let H : Rn ! Rn be
1
1
defined by H .x / D  F   .x /: Then H . O / D . F . . O /// D . F .q // D O and
H 0. O / D

0 F .q / F 0 .q /. 1 /0 .0/ D . F 0 .q // 1 F 0 .q / I

D I:

By the special case above, H has an inverse near O ; which is differentiable at O : Then F has the inverse near q given
by F 1 D  1  H 1  and it is differentiable at F .q /: If in the above argument,we replace q by another point q 0
near q where F is C 1 ; then similarly we will get F 1 is also differentiable at F .q 0 /: Finally, F 1 is C 1 near F .q /

1
 1 if we set y D F .x /). So
because the chain rule gives . F 1 /0 F .x / D F 0 .x / ., . F 1 /0 . y / D F 0 . F 1 . y //
the partial derivatives of F 1 are polynomials of compositions of the partial derivatives of F with the function F 1 ;
which are continuous near q :
Remarks. If F is only C 1 at q 2 Rn ; then from the proof above, we will see that F has an inverse function near q ;
which is differentiable at F .q /: However, the inverse function may not be C 1 at F .q / as the partial derivatives of the
inverse function are not known to exist near F .q /:

146

Practice Exercises for Math 301


1. Prove the triangle inequality: for u ; v; w 2 Rn ; we have ku wk  ku
coordinates for u v and v w: Recall the Cauchy-Schwarz inequality.)

v k C kv

wk: (Hint: Write out

2. Let L D .l1 ; l2 ; : : : ; ln / 2 Rn : For i D 1; 2; 3; : : : ; let vi D .x i;1 ; x i;2 ; : : : ; x i;n / 2 Rn : For sequence v1 ; v2 ; v3 ; : : :


in Rn ; define lim vi D L if and only if lim jjvi L jj D 0: Use the useful inequality and the sandwich theorem
to prove that

i !1

i !1

lim vi

i !1

DL

D lj

if and only if lim x i; j


i !1

for all j

D 1; 2; : : : ; n :

3. Let F : Rn ! Rm be differentiable at p 2 Rn : Explain why there cannot be two different linear transformations
both satisfy the definition of derivative of F at p:
4. Show that F : R3 ! R2 given by F .x ; y ; z / D .x 2 C e y ; z cos x / is differentiable at every .a ; b; c/
F 0 .a ; b; c/ and D F .a ; b; c/ for .a ; b; c/ 2 R3 :

! R2 be defined by F .x ; y; z/ D .x y C cos. y C z/;


(i) Show that F is differentiable at every .a ; b; c/ 2 R3 :

5. (1997 Midterm) Let F : R3

(ii) Find F 0 .a ; b; c/ and D F .a ; b; c/:


6. Let F : R2

! R be defined by F .x ; y/ D

x 2 C y2
0

if x
if x

tiable? Explain your answers.

7. (2000 Midterm) Show that

8 2
< xy
f .x ; y / D x 2 C y 2
:0

the definition.
8. (2000 Midterm) Find all .x ; y /
sure to show work.
9. (2000 Final) Determine if
definition.

2R

ze x /:

6D 0 and y 6D 0 : At which .a ; b/ 2 R2 is F differenD 0 or y D 0

6D .0; 0/ is not differentiable at .0; 0/ by checking


if .x ; y / D .0; 0/

if .x ; y /

where the function

8 2 2
<x y
f .x ; y / D
: x x2 C 2xy

2 R3 : Find

if x

 1 x 2 C y2
f .x ; y / D 2 2 1 2
x C 2y

6D y

if x > 0
is differentiable. Be
if x  0

is differentiable at .0; 0/ or not by checking the

otherwise

10. (1999 Midterm) (a) Show that F .x ; y / D jx y j is not C 1 at .0; 0/:


(b) Show that F .x ; y /

D jx yj is differentiable at .0; 0/:

11. If f ; g : Rn ! R are differentiable at a 2 Rn ; then show that f C g ; f g ; f g and f =g (provided g .a / 6D 0) are


differentiable at a : Show that
D . f g /.a / D g .a / D f .a / C f .a / Dg .a /:
(Hint: Let F : Rn ! R2 be defined by F .x /
P .x ; y / D x C y ; then use the chain rule.)

D . f .x /; g .x //: For f C g ; consider P  F; where P is the function

12. For a ; b 2 Rm ; the line segment joining a to b consists of all points .t / D .1 t /a C t b 2 Rm for t 2 [0; 1]:
If F : Rm ! R is continuous on every point on the line segment joining a to b and F is also differentiable
on the line segment (except perhaps at the endpoints), show that there is a c on the line segment such that
F .b/ F .a / D F 0.c/.b a /:
147

13. (a) For F : R ! R2 defined by F .t / D .cos t ; sin t /; show that there is no c on [0; 2 ] such that F .2/
.2 0/ F 0 .c/:

F .0/

(b) Show that if F : R ! Rn is continuous on [a ; b] and differentiable on .a ; b/; then there is a c 2 .a ; b/ such
that jj F .b/ F .a /jj  jj F 0 .c/jj.b a /: (Hint: Consider .t / D . F .b/ F .a //  F .t / for t 2 [a ; b]: Here the
dot is the dot product in Rn :)
14. (2000 Midterm) Consider the system of equations
u

D x 2e y C cos x y C x

v D xe x y C y 2:

and

Show that near .x ; y ; u ; v/ D .0; 1; 1; 1/; .x ; y / can be expressed as a differentiable function of .u ; v/ and show
that
@u
1
:
.0; 1/ 6D @ y
@y
.1; 1/
@u
15. Consider the system of equations x

D u C v and y D u 2 C v: Using the inverse function theorem, show that near

.u ; v/ D .1; 1/; u and v can be expressed as differentiable functions of x and y : Find

@u @2u
;
in terms of u :
@x @x2

16. (1999 Midterm) Consider the system of equations


x

D v C 2w C evw

and

De

vw

C w C 2v:

Show that near .v; w; x ; y / D .0; 0; 1; 1/; .v; w/ can be expressed as a differentiable function of .x ; y / and

@v
@v
@w
@w
.1; 1/ C .1; 1/ D
.1; 1/ C
.1; 1/:
@x
@y
@x
@y
17. (1999 Midterm) Consider the system of equations
x

D w cos v C v cos w C ew

and

D w sin v C v sin w C ev :

Show that near .v; w; x ; y / D .0; 0; 1; 1/; .v; w/ can be expressed as a differentiable function of .x ; y / and find
the value of

@v
@v
@w
@w
.1; 1/ C .1; 1/ C
.1; 1/ C
.1; 1/:
@x
@y
@x
@y

18. (1999 Midterm) Consider the system of equations


x

D cos u C sin v

and

D sin u C cos v:

Show that near .u ; v; x ; y / D .0; 0; 1; 1/; .u ; v/ can be expressed as a differentiable function of .x ; y / and


2 C @v 2 :
@ u 2
C @@ uy 2 D @v
@x
@x
@y
19. (1997 Midterm) Given .1; 1/ is a solution of the equation x 3 C x y 5 C y 2
can be expressed as y

D 3: Show that near .1; 1/; the equation

f .x / for some differentiable function f : What is


148

dy
at .1; 1/?
dx

20. Given that p D .u ; v; w/ D .1; 0; 1/ is a solution of the system of equations euv C w D 0 and u sin vw D v:
Using the implicit function theorem, show that v and w can be expressed as differentiable functions of u near p:
dv
dw
Find
. p/ and
. p/:
du
du
21. Consider the system of equations e x

C ye

v D 1 and eu

ve

C y D 1:

(i) Show that near .x ; y ; u ; v/ D .0; 1; 0; 1/; .x ; y / can be expressed as a function of .u ; v/:
(ii) Find

@y
@x
.0; 1/; .0; 1/:
@u
@u

22. Suppose f1 ; f2 : R4

! R are differentiable functions. Consider the system of equations


f1 .w; x ; y ; z / D 0

f2 .w; x ; y ; z / D 0:

and

If near a solution .w0 ; x 0 ; y0 ; z 0 /; .x ; y / can be expressed as a differentiable function of .w; z /; then show that at
.w0 ; x 0; y0; z 0/;

@x
@w

@. f1 ; f2 / . @. f1 ; f2 /
@.w; y /
@.x ; y /

@y
@w

and

@. f1 ; f2 / . @. f1 ; f2 /
:
@.x ; w/
@.x ; y /

(Hint: Just use chain rule and Cramers rule.)

! R with variables x1 ; : : : ; xn ; w1; : : : ; wm ; the formula is


. @. f1 ; : : : ; fn /
@. f1 ; : : : ; fn /

Remarks. For n differentiable functions fi : RnCm

@ xi
@w j

@.x 1 ; : : : ; x i 1; w j ; x i C1 ; : : : ; x n /

@.x 1 : : : ; x n /

23. (2000 Midterm) Consider the system of equations


e x C y CzCw

D cos.x C y/ C sin.z C w/

Show that near .x ; y ; z ; w/

and

x y z w

D sin.x C y/ C cos.z C w/:

D .1;

1; 1; 1/; .x ; z / can be expressed as a differentiable function of . y ; w/: At


@x
@z
.x ; y ; z ; w/ D .1; 1; 1; 1/; find
and
:
@w
@w
24. (1999 Final) For the system x 3 C y 3 C z 3 C 3y

D x w3 and x 5 C y5 z3 D x yz C x 3 w5; show that .x ; y/ can be


@x
expressed as a differentiable function of .w; z / near the solution p D .w; x ; y ; z / D .1; 1; 0; 0/: Find
. p/ and
@w
@y
. p/:
@w

25. (1998 Midterm) (a) Given .w; x ; y ; z / D .0; 0; 0; 0/ is a solution of the system of equations
cos w C sin x

C tan y D z C 1

and

ew

Cx D yCe

Near .0; 0; 0; 0/; show that .x ; y / can be expressed as a differentiable function G of .w; z /:
(b) Find

@x
@y
.0; 0/ and
.0; 0/:
@w
@w

26. Given F : R3

! R is C 1 near .0; 0; 0/;


F .0; 0; 0/ D 0;

@F
@F
@F
.0; 0; 0/ 6D 0;
.0; 0; 0/ 6D 0;
.0; 0; 0/ 6D 0:
@x
@y
@z
149

Show that near .0; 0; 0/; the equation F .x ; y ; z /


y D g .z ; x /; z D h .x ; y /: Show that near .0; 0; 0/;

0 can be expressed in each of the form x

@x @y @z
@y @z @x
(Here

@x
@y

1:

D @@ fy ; @@ yz D @@gz ; @@ xz D @@ hx :)

27. Prove that if limsup x n


n!1

D x D liminf
x n (i.e. L D fx g), then lim x n D x :
n!1
n!1

Remarks. Think carefully. This was not explained in the notes!


28. Find the limit superior and the limit inferior of each of the following sequences.

D n 2.1 cos e n /I
(b) x n D f .en cos.n=2/ 1 /; where
(a) x n

f .x /

D x ln x :

29. (1997 Final) Find the limit superior and limit inferior of each of the following sequences:

D n .1 cos 1n / sin n for n D 1; 2; 3 : : : I


(ii) x n D .cos n / cos 2n5 for n D 1; 2; 3; : : : :
(i) x n

30. (1997 Final) Find the limit superior and limit inferior of each of the following sequences:

D [n =2] ncos n  for n D 1; 2; 3 : : : ; where [x] is the greatest integer less than or equal to x I
1 
(ii) x 1 D 1 and x nC1 D x n 1 C 2
for n D 1; 2; 3; : : : :

(i) x n

n 1

31. (1998 Final) Find the limit superior and limit inferior of each of the following sequences:

D n sin.e n / for n D 1; 2; 3 : : : I
(ii) x n D 2n cos.n=7/ for n D 1; 2; 3; : : : :
a
a
nC1
nC1 ; limsup pjan j and liminf pjan j; where
Find limsup
;
liminf

n!1
n!1
an
an
n!1
n!1
 k
k =2
if k is odd
ak D
:
k 2 =2k if k is even
(i) x n

32.

a
p
nC1
j
an j may exist, but lim
may not.)
n!1
n!1
a

(Remarks. This is an example to show lim

33. (1999 Midterm) (a) Let fan g and fbn g be sequences of real numbers. Show that
limsup.an
n!1

C bn /  limsup an C limsup bn ;
n!1

n!1

provided the left or right side is not of the form 1 1:


(b) Give an example of sequences fan g and fbn g of real numbers so that
limsup.an C bn / 6D limsup an C limsup bn
n!1

and the right side is not of the form 1

n!1

n!1

1: Give reasons to support your example.


150

f . y ; z /;

34. (a) Let fan g and fbn g be sequences of real numbers. Show that
liminf an C liminf bn
n!1

n!1

 liminf
.an C bn /  liminf an C limsup bn  limsup.an C bn /  limsup an C limsup bn ;
n!1
n!1
n!1

n!1

provided none of the expressions is of the form 1 1:


1
1
(b) If liminf an  ; liminf bn  and lim .an C bn / D 1; then prove that lim an
n!1
n
!1
n!1
n!1
2
2

n!1

n!1

D 12 D nlim
b :
!1 n

35. Let fan g and fbn g be sequences of nonnegative real numbers. Show that
limsup.an bn /  .limsup an /.limsup bn /;
n!1

n!1

n!1

provided the right side is not of the form 0  1:


36. Give an example of sequences fan g and fbn g of nonnegative real numbers so that
limsup.an bn / 6D .limsup an /.limsup bn /
n!1

n!1

n!1

and the right side is not of the form 0  1:


37. For a sequence a1 ; a2 ; a3 ; : : : of real numbers, let

n
Show that
liminf an
n!1

D a1 C a2 Cn    C an :

 liminf
n  limsup n  limsup an :
n!1
n!1

n!1

38. Find the domain of each of the following series of functions, i.e. the largest set where the series converges
pointwise. For those that are power series, state the radii of convergence.

1 .x 1/k
k D1 .2k /!
1 .2x 3/k
(c) 6
k D1
4k

1 xk
k D2 ln k
1 ekx
(d) 6
k D1 k C 1

(a) 6

(b) 6

39. Find the first three nonzero terms of the Taylor series of tan x with center at c D 0: Next, find the first three
nonzero terms of the Taylor series of sin x at c D 0 divided by the Taylor series of cos x at c D 0 by long division.
Remarks. Although the Taylor series of sin x and cos x at c D 0 have infinite radii of convergence, the Taylor
series of their quotient tan x at c D 0 does not have infinite radius of convergence since tan.=2/ D 1:

40. Show by an example that two power series with the same center may have radii of convergence equal 1, but their
sum may have infinite radius of convergence.
41. Find the Taylor series of cos2 x with center c
. 1; C1/ to cos2 x :

t !1

t2

0 and show that the Taylor series converges pointwise on

6D 0 : Show that f .n/ .0/ D 0 for n D 1; 2; 3 : : : : (Hint: First


D0

D 0 for n D 0; 1; 2; : : : by induction. Then, for x 6D 0; show f .n/ .x / D Pn 1x e 1=x for some

42. Let f : R ! R be defined by f .x / D


show lim t n e

2
e 1= x
0

if x
if x

polynomial Pn .x / by induction. Use the definition of derivative to compute f .n/ .0/:)


151

43. Show that each of the following sequences of functions converges uniformly on the intervals indicated.
xn
on [0; 0:99]
1 C xn
(c) h n .x / D cos x n on [0; r]; where 0 < r < 1

(a) fn .x / D

44. Does the sequence of functions Sn .x /


answer.

(b) gn .x /

D nxe

nx 2

on [0:1; 1/

D x n converge uniformly on [0; 1/?

Give explanations to support your

45. Show that each of the following series of functions converges uniformly on the intervals indicated.

1 sin kx
on . 1; C1/
ek
k D1
1
(c) 6 ke kx on [1; C1/
k D1
1
(e) 6 k 2 x k on [ r; r]; where 0 < r < 1
k D1
1 kx
(g) 6 kx on [r; C1/; where r > 0
k D1 e
(a) 6

1 xk
on [ 1; 1]
k D2 k .ln k /2
1
(d) 6 .x ln x /k on .0; 1]
k D1
1 xk
(f) 6
on [0; r /; where 0 < r < 1
k D1 k
(b) 6

46. Show that each of the following sequences or series of functions converges pointwise on the intervals indicated,
but fails to converges uniformly there.
(a) Sn .x / D sinn x on [0;  ]
(c) Sn .x / D nxe

nx 2

1
k D0

(b) 6 x k .1

x / on .0; 1]

on [0; 1]

47. Give an example of a sequence of functions Sn : [0; 1] ! R that does not converge uniformly on [0; 1]; but the
sequence of their squares Sn2 .x / converges uniformly on [0; 1]: Give reasons to support your example.
48. Construct sequences fn and gn which converges uniformly on some set E, but such that fn gn does not converge
uniformly on E : (Hint: Try fn .x / D x and gn .x / D n1 on R:)
49. If fn converges uniformly on E to f ; gn converges uniformly on E to g and both f and g are bounded on E ; then
show that fn gn converges uniformly on E to f g :
50. Prove that if fn converges uniformly on a set E to f and each fn is a bounded function on E (i.e. there is Mn 2 R
such that j fn .x /j  Mn for every x 2 E /; then the sequence fn is uniformly bounded on E in the sense that there
is M 2 R such that j fn .x /j  M for every n and every x 2 E :
51. If fn and gn converges uniformly on a set E to f and g ; respectively, then prove that fn C gn converges uniformly
on E to f C g : If, in addition, fn and gn are bounded functions on E ; then prove that fn gn converges uniformly
on E to f g : (Hint: For the second statement, use the previous exercise.)

1
1
k D1
k D1
1
uniformly on [0; 1]: Can 6 cos gk .x / converge pointwise on [0; 1]?
k D1
(b) If the sequence of functions Sn .x / converges uniformly on [0; 1] to S .x /; then show that cos Sn .x / must
converge uniformly on [0; 1] to cos S .x /:

52. (2000 Midterm) (a) If 6 gk .x / converges uniformly on [0; 1]; then show that 6 cos gk .x / cannot converge

53. Let F0 .x / beZ a bounded Riemann integrable function on [0; 1]: For n
by Fn .x / D

1; 2; 3; : : : ; define Fn .x / on [0; 1]

Fn 1 .t / dt : Prove that the sequence Fn .x / converges uniformly on [0; 1]: Prove that the series

1
6 Fk .x / also converges uniformly on [0; 1]:
0

k D0

152

n0

1
if x  0
; if x n is a sequence of points in interval .a ; b/ and if 6 cn converges absolutely, prove
1 if x > 0
nD1
1
that the series f .x / D 6 cn I .x x n / converges uniformly on .a ; b/ and f is continuous at every x 6D x n :

54. If I .x /

nD1

(Remarks: If .a ; b/ D . 1; C1/ and Q D fx 1 ; x 2 ; x 3 ; : : :g; then f is a function that is continuous at every


irrational number, but discontinuous at every rational number.)

11
i D1 j D1

55. Prove that 6 6 ai j

D j61D1

1 

6 ai j if ai j

i D1

 0 for all i and j (the case C1 D C1 may occur).


p

56. (1998 Midterm) Let [x] be the greatest integer less than or equal to x ; (for example, [ 2] D 1; [3 21 ] D 3 and
1 1
[5] D 5:) Define f .x / D x [x]: Consider the graph of f : Show that g .x / D 6 k f .kx / is (Riemann) integrable
k D1 2
on [0; 1]:

1
k D1

57. Prove that the series 6 . 1/k


absolutely for any value of x :

x2 C k
converges uniformly on every bounded interval, but does not converge
k2

D 1; 2; 3; : : : ; x real, put fn .x / D 1 Cxnx 2 : Show that fn converges uniformly on R to a function f


that the equation f 0 .x / D lim fn0 .x / is correct if x 6D 0; but false if x D 0:
n!1
Z 1 1 cos.x 2 /
1 . 1/kC1
1 cos.x 2 /
Find lim
and
show
that
d
x
D
:
6
x !0
x4
x4
3/
k D1 .2k /!.4k

58. For n

59.

and

ex

60. (1999 Midterm) Find lim

x !0

1
x

and show that

Z 1 ex

1
x

61. (2000 Midterm) (a) Find the maximum value of f .x /


to support your answer.
Z 12 x ln.2x /
1 . 1/kC1
(b) Show that
d x D 6 kC2
:
1Cx
.k C 2 /2
k D0 2
0
62. (a) Use

dx

D k6D1 k!.12k/ :

D jx ln.2x /j on the interval .0; 12 ]: Be sure to show work

d
1 
D .1 1 x /2 to find a power series for .1 1 x /2 on the open interval . 1; 1/:
dx 1 x

1
2

(b) Show that

.1

x/

dx

D k6D0 p

kC1

p
:
. 2 C k C 1/2 2CkC1

63. (a) Show that lim sin x ln x exists by considering sin x ln x


x !0C

D sinx x .x ln x /:

1 . 1/n x 2nC1 ln x
converges uniformly on [0; 1]: (Here at x D 0; we interpret x 2nC1 ln x as 0.)
.2n C 1/!
nD0
Z1
1
. 1/nC1
(c) Use the Taylor series of sin x to show that
sin x ln x d x D 6
:
nD0 .2n C 1/!.2n C 2/2
0

(b) Show that 6

64. (1998 Midterm) (a) Prove that the Taylor series of f .x /


every x 2 R:
(b) Prove that

Z 1 sin2 x
0

support the steps.

x2

dx

k C1 2k 1

D k6D1 ..2k1//!.2k2

1/

D sin2 x D 12 .1

cos 2x / about c

D 0 equals

f .x / for

: (At 0, the integrand is set to 1.) Be sure to give reasons to

153

2
3

(a) Show that

ln.1

x 2/
x2

1
3

dx

1 xk
for 1 < x < 1:
6
k D1 k

x/ D

65. (1999 Midterm) Given that ln.1

1 22k 1 1
:
6
1/32k 1
k D1 k .2k

1
1 x kC2
1
converges
and
find
its
sum
by
considering
f
.
x
/
D
:
6
k D1 k .k C 2/2k C2
k D1 k .k C 2/

(b) Show that 6

Z1

66. (2000 Midterm) (a) Find the value of

u 2 sin u du : Show work.

1
. 1/k
: Be sure to explain the steps. (Hint: Consider the integral in part (a).)
(b) Find 6
k D0 .2k C 1/!.k C 2/
tan x x
x sin x
and lim
(by power series method).
x !0 x .1
x !0 tan x
cos x /
x

67. Find lim

1
k D0

68. Suppose f .x / D 6 ak .x
k

D 1; 2; 3; : : : : (Hint:

c/k for every x

f .nC1/

D . f 0 /.n/ :)

2 .a ; b/; where c 2 .a ; b/: Show by induction that ak D

f .k/ .c/
for
k!

Remarks. This means if a function can be written as a power series on an open interval (with the center on the
interval), then it is the Taylor series of the function on the interval.

1 xk
1 1
converges
and
find
its
sum
by
considering
f
.
x
/
D
:
6
k D1 kek
k D1 k

69. Show that 6

70. (a) Find the domain of the power series f .x /


(b) Given arctan x

Dx

x3
3

x7
7

C x5

. 1; 1/ and justify.
1
. 1/nC1
(c) Show f .1/ D 6
nD1 .2n
1/.2n C 1/

D n6D1 .2n.
1

C    D n6D1 .

D 4

1/nC1 x 2nC1
:
1/.2n C 1/
1/nC1 x 2n
2n 1

for every x

2.

1; 1/: Find f 0 .x / for x

1
:
2

71. Given that all the solutions of the differential equation f .4/.x / f .x / D 0 are of the form f .x / D ae x C be x C
1 1
1 x 4k
c cos x C d sin x ; where a ; b; c; d 2 R: Show 6
converges and find its sum by considering 6
:
k D0 .4k /!
k D0 .4k /!
72. On E

D[

1; 1]; define polynomials P1 .x /

D 0 and for n > 1;

Pn .x /

D Pn

.x / C

x2

Pn2 1 .x /
:
2

Show that Pn .x / converges uniformly


 on E to jx j: (Hint: Show
 0  Pn .x /  PnC1 .x /  jx j on E by induction.
Use jx j Pn .x / D jx j Pn 1 .x / 1 12 .jx j C Pn 1 .x // to show jx j Pn .x /  jx j.1 12 jx j/n 1 < n2 on E :)
Remarks. This gives an example of a sequence of differentiable functions converging uniformly on an interval to
a nondifferentiable function. In this example, the limit function jx j is not differentiable at x D 0 only. Applying
the Weierstrass approximation theorem to a continuous, but nowhere differentiable function, it is possible that the
limit function can be nowhere differentiable!
73. If f is continuous on [0; 1] and if

Z1
0

f .x /x n d x

D 0 for n D 0; 1; 2; : : : ; prove that

f .x /

D 0 on [0; 1]: (Hint:

The integral of f with any polynomial is zero. Use the Weierstrass approximation theorem to show that the
integral of f 2 is 0.)
154

6D ; and S 6D R; then S cannot be both open and closed.

74. Show that in R; if S

75. For each part below, determine if the set is open, closed, compact or none of these. (You may use graphs.)

Df
(ii) B D g
(iii) C D h

(i) A

..2; 9]/; where f : R ! R; f .x / D x 2 :

.[ 1; 2/ [ .3; 4//; where g : .0; C1/ ! R; g .x / D e x :

. K /; where h : [0; 1] ! R; h .x / D x and K is the Cantor set.


1 1
[
1 
(iv) D D i
. ; 2 C 2 / ; where i : .0; 1/ ! R; i .x / D ln x :
n
n
nD1
(v) E

D j Rn

.2n ; .2n C 1// ; where j : R ! R; j .x / D sin x :

n2Z

 1 : n 2 Zand n 6D 0 :

(vi) F

(vii) G

D f0 g [


1
1
[ 2
; 2
] :
n C2 n C1
n2N

76. Find the length of each of the following sets.


(i) H

(ii) I

1
[

1
n
nD1

D [0; 1] n

1 1
;
20 n

C 201

:

(Be sure to explain where overlaps occur.)

1
[

1
1
; /:
2k
C
1
2k
k D1

77. Let A and B be sets in R:

(i) Show that Int. A \ B / D Int. A/ \ Int. B /:

(ii) Must Int. A [ B / D Int. A/ [ Int. B /?

1
78. In the construction of the Cantor set, at the n-th stage, let us remove an open interval of length n in the middle
3 2
of each remaining intervals instead, show that this new set is compact. Show that its length is positive, but it does
not contain any interval of positive length (i.e. it has no interior point).
Remarks. The existence of such a positive length Cantor-like set will be used in Math 501. From this, we can get
a sequence of continuous functions converging pointwise on [0; 1] to a function not Riemann integrable.

p
D fx 2 [ ;
p 2/ : x is irrationalg:
B D fx  C y 2 : x ; y 2 Qg:

79. For each part below, show that the set is measurable and find its Lebesgue measure.
(a) A
(b)

(c) C is the set of all numbers x such that x has a distance less than
m; n

2 N have highest common factor equal 1.

1
m
from some rational number ; where
5n
n

(d) D is the set of all numbers in the Cantor set that are irrational.
80. Show that the set
A

D fx 2 [0; 1] : x has a decimal representation not containing the digit 5g

is measurable. What is m . A/?


155

81. (1997 Final) (i) Show that if C is the Cantor set, then the set S
(ii) Show that m . S / D 0; where S is the set in part (ii).

D fx C n : x 2 C; n D 1; 2; 3; : : :g is measurable.

82. (1999 Final) Let G be a nonempty open set. Prove that not every number in G is equal to a number of the form
r C k ; where r is a rational number and k is a number in the Cantor set.
83. Use countable additivity to prove the first part of the monotone set theorem, namely for measurable sets A1 ; A2 ; : : : ;
if A1

 A2  : : : ; then m

1 
[
k D1

Ak

D klim
m . Ak /:
!1

84. Define the symmetric difference of two sets A and B to be A4 B


and m . A4 B / D 0; then B is measurable.

D . A n B / [ . B n A/: Show that if A is measurable

85. (1998 Final) Write the rational numbers Q D fq1 ; q2 ; q3 ; : : :g: Define G
(i) Show that G is measurable and m .G / < 1:

(ii) Show that if F is a closed set and Q  F ; then F


(iii) For every closed set F ; show that either m .G
86. (2000 Final) For X ; Y

1 
[
nD1

qn

1
1
; qn C 2
2
n
n

D R:

n F / > 0 or m. F n G/ > 0:

 R; define X C Y D fx C y

: x

2 X and y 2 Y g:

(a) Give the definition of an interior point of a set S in R:

(b) Let A be a nonmeasurable set and B be a nonempty open set. Show that A C B is a measurable set.

(c) Let A be a nonmeasurable set and C be a nonempty closed set. Give two examples, one showing A C C may
be a measurable set and one showing A C C may be a nonmeasurable set. Explain your examples.
(d) If X ; Y and X

C Y are nonempty measurable sets, must m. X C Y / D m. X / C m.Y /? Give explanations.

87. This is a structure theorem for measurable sets. Let A be a bounded measurable set, show that there is a sequence
of open set S1 ; S2 ; S3; : : : and a set C of measure 0 such that A D
A

 Sn and m. Sn /  m . A/ C n1 :)

1 
\
Sk n C: (Hint: Take open set Sn such that
k D1

Remarks. A set which is a countable intersection of open sets is called a G  set. Thus, the exercise says every
bounded measurable set is a G  set minus a set of measure 0. Similarly, there is a description of bounded measurable
sets using closed sets. It asserts that every bounded measurable set is a union of closed sets W1 ; W2 ; W3 ; : : :
and a set of measure 0. A set which is a countable union of closed set is called a F set. Thus, every bounded
measurable set is the union of a F set and a set of measure 0.
88. (1998 Final) Give an example of two functions f ; g : R ! R such that 0 < g .x /
is not measurable, but g is measurable.

f .x / for every x

2 R and

89. Let A be a measurable set. Show that the following are equivalent:
(a) f : A

! R is a measurable function.

(b) for every open set S ; f 1 . S / is measurable.


(c) for every closed set W ; f 1 .W / is measurable.
(d) for every compact set K ; f

. K / is measurable.

Remarks. There are measurable functions f ; g : A ! R such that for some measurable set B ; f
nonmeasurable and g . B / is nonmeasurable. See fact (2) of Appendix 2 to Step 3 of Chapter 12.
156

. B / is

90. (1997 Final) Let f : R ! R be a function and f 2 .x /

D . f .x //2 .

(i) Show that even if the function f 2 is measurable, f may still be not measurable.
(ii) Show that if the function f 2 is measurable and the set S
measurable. (Hint: XS is measurable.)

D fx 2 R :

f .x /

 0g is measurable, then

f is

91. For a ; b; c 2 R; define mid.a ; b; c/ to be the middle value when a ; b; c are arranged in increasing order. If
f1 ; f2 ; f3 : R ! R are measurable functions, show that g : R ! R defined by g .x / D mid. f1 .x /; f2 .x /; f3 .x //
is a measurable function.
92. (2000 Final) Let p; q ; r; s : R ! R be measurable. Show that the function t : R ! R defined by
t .x / D

r .x /
s .x /

if p.x / < q .x /
otherwise

is a measurable function.
93. Show that if f : R ! R is differentiable, then f 0 is a measurable function.
94. Show that if f : R ! R is a monotone function, then f is a measurable function.
95. (a) Show that if f ; g : R ! R are measurable functions, then S

D fx 2 R :

(b) Let f1 ; f 2 ; f3 ; : : : : R ! [0; 1] be a sequence of measurable


limn!1 fn .x / existsg is measurable.

96. (2000 Final) Let f ; g : R ! R be continuous. Show that A


example to show that the set A may also be a closed set.

D g .x /g is a measurable set.
functions. Show that S D fx 2 R :

D fx 2 R :

f .x /

f .x / < g .x /g is an open set. Give an

97. (1999 Final) Let f ; g : R ! R be measurable functions. Show that the set
S

D fx 2 R :

D n C g .x /

f .x /

for some n

2 Zg

is measurable.

98. (1998 Final) Let [x] be the greatest integer less than or equal to x : Let f : R ! R be a measurable function.
(i) Show that g .x / D [ f .x /] is a measurable function.
(ii) Show that h .x /

f .[x]/ is a measurable function.

99. (1999 Final) Let K be the Cantor set. Let f : R ! R be defined by f .x /


be continuous.

n1
0

if x
if x

62 K
2K

and let g : R ! R

(a) Prove that the composition function g  f is measurable.


(b) Prove that the composition function f

 g is measurable.

100. Let C be the Cantor set. Prove that C C C


Need to use base 3 representation.)

101. (1999 Final) Find lim

n!1

 3

[2;

D fx C y : x ; y 2 C g is a measurable set. (Hint: Show it is an interval.

sin x
dm : Be sure to give reasons, not just answer.
1 C e nx

102. (1997 Final) (i) For x > 0; show that lim 1 C

x n
n

D ex : (Hint: Let t D n1 and take log.)


 x n x
(ii) For x > 0 and positive integer n ; show that 1 C
e :
n
n!1

157

(iii) Find lim

Z 1

n!1 0

x n
e
n

1C

2x

dx:

103. (1998 Final) (i) Find the maximum of y .s /

Z1

(ii) Find lim

n!1 1

D s1=s for s 2 [1; C1/:

e x .nx / nx d x :
1

Z1

104. (1997 Final) Show that


0

Z
105. (1998 Final) Show that

.0;1/

Ze
106. (2000 Final) Show that

x
ex

dx

.ln x / ln.1

e x ln x d x

D n6D0 .n C1 1/2 : (Hint: Note e


x / dm

< 1 on .0; 1/:)

D n6D1 n .n C1 1/2 :

1 n C1
D n6D0 n!ne.n CC1/12 :

107. Use the Monotone Convergence Theorem to prove the fact that if A1 ; A2 ; A3 ; : : : are measurable sets, A1
A3

    ; then m

1 
[
An D lim m . An /:
n!1

 A2 

nD1

108. Use the Lebesgue Dominated Convergence Theorem to prove the fact that if A1 ; A2 ; A3 ; : : : are measurable sets,
A1

 A2  A3     and m. A1 / < 1; then m

109. Let g .x / D

0

if x
1 if x

2 [0; 1=2] : Let


2 .1=2; 1]Z

liminf fn .x / D 0 on [0; 1]; but


n!1

110. Show that f .x /

n sin x
x

if x
if x

f2k .x /

fn . x / d x

1 
\

nD1

An

D nlim
m . An /:
!1

g .x / and f2kC1 .x /

g .1

x / for x

[0; 1]: Show that

D 12 for every n : So the inequality in Fatous lemma can be strict.

6D 0 is a measurable, but not Lebesgue integrable function on R:


D0

111. Show by an example that Fatous lemma may become false if the functions are real-valued instead of nonnegative
valued.

158

You might also like