You are on page 1of 20

# 4.

11

Lebesgue integrals

497

## b. Show that : Qa A is one to one.

>
c. What is A y |dx dy|?
4.10.18 What is the volume of the part of the ball of equation x2 + y 2 + z 2 4
where z 2 x2 + y 2 , z > 0?
4.10.19 Let Q = [0, 1] [0, 1] be the unit square in R2 , and let : R2 R2
be given by
  

u v2
u
=
and A = (Q).
2
v
u +v
a. Sketch A, by computing the image of each of the sides of Q (they are all
arcs of parabolas).
b. Show that : Q A is 11.
>
c. What is A x |dx dy| ?
Change of variables for exercise
4.10.20:
3

u
(w 1)u
v = (w3 + 1)v .
w
w

4.10.20

Solve exercise 4.5.19 again, using the change of variables in the margin.

## 4.10.21 The moment of inertia of a body X R3 around an axis is the integral

;
(r(x))2 |d3 x|, where r(x) is the distance from x to the axis.
X

## a. Let f be a nonnegative continuous function of x [a, b], and let B be the

body obtained by rotating the region 0 y f (x), a x b around the x-axis.
What is the moment of inertia of B around the x-axis?

2
2

## 4.11 Lebesgue integrals

This new integral of Lebesgue is proving itself a wonderful tool. I
might compare it with a modern Krupp gun, so easily does it penetrate
barriers which were impregnable.Edward Van Vleck, Bulletin of the
American Mathematical Society, vol. 23, 1916.
So far we have restricted ourselves to integrals of bounded functions with
bounded support, whose upper and lower sums are equal. In this section
we will study Lebesgue integration, which has three advantages:
1. Lebesgue integrals exist for functions that are unbounded and/or
have unbounded support (although for a function to be Lebesgue
integrable, the area under the graph must be nite).
2. Lebesgue integrals exist for functions plagued with the kind of local
nonsense that we saw in the function that is 1 at rational numbers
in [0, 1] and 0 elsewhere (example 4.3.3). The Lebesgue integral
ignores local nonsense on sets of measure 0.
3. Lebesgue integrals are better behaved with respect to limits.

498

## Our approach to Lebesgue integration is very dierent from the

standard one. The usual way of
dening the Lebesgue integral
;
f (x)|dn x|
Rn

## is to cut up the codomain R into

small intervals Ii = [xi , xi+1 ], and
to approximate the integral by

xi (f 1 (Ii )),
i

## where (A) is the measure of A,

then letting the decomposition of
the codomain become arbitrarily
ne. Of course, this requires saying what subsets are measurable,
and dening their measure. This
is the main task with the standard
approach, and for this reason the
theory of Lebesgue integration is
often called measure theory.
It is surprising how much more
powerful the theory is when one
decomposes the codomain rather
than the domain. But one pays a
price: it isnt at all clear how one
would approximate a Lebesgue integral: guring out what the sets
f 1 (Ii ) are, never mind nding
their measure, is dicult or impossible even for the simplest functions.
We take a dierent tack, building on the theory of Riemann integrals, and dening the integral
directly by taking limits of functions that are Riemann integrable.
We get measure theory at the end
as a byproduct.

Chapter 4.

Integration

## Remark. If Lebesgue integration is superior to Riemann integration, why

did we put so much emphasis on Riemann integration earlier in this chapter?
Riemann integrals have one great advantage over Lebesgue integrals: they
can be computed using Riemann sums. Lebesgue integrals can only be computed via Riemann integrals (or perhaps by using Monte Carlo methods).
Thus our approach is in keeping with our emphasis on computationally
eective algorithms. 
Before dening the Lebesgue integral, we will discuss the behavior of
Riemann integrals with respect to limits.

## Integrals and limits

The behavior of integrals under limits is often important. Here we give the
best general statements about Riemann integrals and limits.
We would like to be able to say that if fk is a convergent sequence of
functions, then, as k ,
;
;
lim fk = lim fk .
4.11.1
In one setting this is true and straightforward: when fk is a uniformly
convergent sequence of integrable functions, all with support in the same
bounded set. The key condition in denition 4.11.1 is that given
, the
same K works for all x.
Denition
in
4.11.1 (Uniform convergence). A sequence of functions
fk : Rn R converges uniformly to a function f if for every
> 0, there
exists K such that when k K, then, for all x Rn , |fk (x) f (x)| <
.
The three sequences of functions in example 4.11.3 do not converge uniformly, although they do converge. Uniform convergence on all of Rn isnt
a very common phenomenon, unless something is done to cut down the
domain. For instance, suppose that
pk (x) = a0,k + a1,k x + + am,k xm

4.11.2

## is a sequence of polynomials all of degree m, and that this sequence

converges in the obvious sense that for each degree i (i.e., each xi ),
the sequence of coecients ai,0 , ai,1 , ai,2 , . . . converges. Then pk does not
converge uniformly on R. But for any bounded set A, the sequence pk 1A
does converge uniformly.69
69

## Instead of writing the sequence pk 1A we could write pk restricted to A.

We use pk 1A because we will use such restrictions in integration, and we use the
indicator function 1 to dene the integral over a subset A:
;
;
p(x) =
p(x)1A (x).
A

Rn

4.11

## There are many reasons to

study Lebesgue integrals. An essential one is the Fourier transform, the fundamental tool of engineering and signal processing,
not to mention harmonic analysis. (The Fourier transform is discussed at the end of this section.)
Lebesgue integrals are also ubiquitous in probability theory.

Lebesgue integrals

499

Theorem
in
4.11.2 (Convergence for Riemann integrals). Let fk
be a sequence of bounded integrable functions, all with support in a
xed ball B Rn , and converging uniformly to a function f . Then f is
integrable, and
;
;
lim
fk (x) |dn x| =
f (x) |dn x|.
4.11.3
k

Rn

Rn

Proof. Choose
> 0 and K so large that supxRn |f (x) fk (x)| <
when
k > K. Then when k > K, we have, for any N ,

LN (f ) > LN (fk )
voln (B)

voln (B).

4.11.4

## Now choose N so large that UN (fk ) LN (fk ) <

; we get
Equation 4.11.4: If you picture
this as Riemann sums in one variable, is the dierence between
the height of the lower rectangles
for f , and the height of the lower
rectangles for fk , while the total width of all the rectangles is
vol1 (B), since B is the support for
fk .

## Exercise 4.11.13 asks you to

verify that mass is indeed lost in
the three sequences of functions in
example 4.11.3.

UN (f ) LN (f ) UN (fk ) LN (fk ) +2
voln (B),
\$
%&
'

4.11.5

<

## yielding U (f ) L(f ) <

(1 + 2 voln (B)). 
In many cases theorem 4.11.2 is good enough, but it cannot deal with
unbounded functions or functions with unbounded support. Example 4.11.3
shows some of the things that can go wrong.
Example 4.11.3 (Cases where the mass of an integral gets lost).
Here are three sequences of functions where the limit of the integral is not
the integral of the limit.
1. When fk is dened by

1
if k x k + 1
fk (x) =
4.11.6
0
otherwise,
the mass of the integral is contained in a square 1 high and 1 wide. As
k this mass drifts o to innity and gets lost:
;
;
;
lim
fk (x) dx = 1, but
lim fk (x) dx =
0 dx = 0. 4.11.7
k

## 2. For the function


fk (x) =

if 0 < x

otherwise,

1
k

4.11.8

the mass of the integral is contained in a rectangle k high and 1/k wide.
As k , the height of the rectangle tends to and its width to 0:
; 1
; 1
; 1
lim
fk (x) dx = 1, but
lim fk (x) dx =
0 dx = 0.
4.11.9
k

0 k

## 3. The third example is our standard example of a function that is not

integrable using the Riemann integral. Let us make a list a1 , a2 , . . . of the
rational numbers between 0 and 1. Now dene

1
if x {a1 , . . . , ak }
fk (x) =
4.11.10
0
otherwise.

500

Chapter 4.

Integration

Then we have

fk (x) dx = 0

for all k,

4.11.11

## but limk fk is the function that is 1 on the rationals and 0 on the

irrationals between 0 and 1, and hence not integrable. 

Figure 4.11.1.
Henri Lebesgue (18751941)
After Lebesgues father died of
tuberculosis, leaving three children, the oldest ve years old,
his mother cleaned houses to support them. Lebesgue later wrote,
My rst good fortune was to be
born to intelligent parents, then
to have been sickly and extremely
poor, which kept me from violent
games and distractions, . . . and
most of all to have an extraordinary mother even for France, this
country of good mothers.
When one of Lebesgues students apprehensively arrived for
her rst teaching job, in which
she was to replace a popular substitute teacher, she found a note
from him waiting for her. Faitesvous aimer l`
a-bas comme partout,
beloved there as you are everywhere). It was a ray of sunshine, she recalled in a note published in Message dun mathematicien: Henri Lebesgue, pour le
centenaire de sa naissance, Paris,
A. Blanchard, 1974.
Proposition 4.11.5 is proved in
appendix A.22.

The pitfalls of disappearing mass can be avoided by the dominated convergence theorem for Riemann integrals, theorem 4.11.4. In practice this
theorem is not as useful as one might hope, because the hypothesis that
the limit is Riemann integrable is rarely satised unless the convergence is
uniform, in which case the much easier theorem 4.11.2 applies. But theorem 4.11.4 is the key tool in our approach to Lebesgue integration. The
proof, in appendix A.22, is quite dicult and very tricky.
Theorem
in
4.11.4 (The dominated convergence theorem for
Riemann integrals). Let fk : Rn R be a sequence of integrable
functions. Suppose there exists R such that all fk have their support in
BR , and all satisfy |fk | R. Let f : Rn R be an integrable function
such that the set of x where limk fk (x)
= f (x) has measure 0. Then
;
;
lim
fk (x) |dn x| =
f (x) |dn x|.
4.11.12
k

Rn

Rn

## Dening the Lebesgue integral

The weakness of theorem 4.11.4 is that we have to know that the limit is
integrable. Usually we dont know this; most often, we need to deal with
the limit of a sequence of functions, and all we know is that it is a limit. But
we will now see that theorem 4.11.4 can be used to construct the Lebesgue
integral, which is much better behaved under limits.
We abbreviate Riemann integrable as R-integrable and Lebesgueintegrable as L-integrable.
Proposition
in
4.11.5 (Convergence except on a set of measure 0).
If fk for k = 1, 2, . . . are Riemann-integrable functions on Rn such that
;

|fk (x)||dn x| = A < ,
4.11.13
k=1

Rn

k=1

## (Recall that except on a set of measure 0 is also written almost

everywhere. So in proposition 4.11.5 we could simply say that the sum
converges almost everywhere (or a.e.).)
We can now dene equal in the sense of Lebesgue, denoted =.
L

4.11

This notion of Lebesgue equality is fairly subtle, as sets of measure 0 can be quite complicated.
For instance, if you only know a
function almost everywhere, then
you can never evaluate it at any
point: you never know whether
this is a point at which you know
the function. The moral: functions that you know except on a
set of measure 0 should only appear under integral signs.
Lebesgue integration is superior to Riemann integration. It
does not require functions to be
bounded with bounded support, it
ignores local nonsense on sets
of measure 0, and it is better behaved with respect to limits.
But if you want to compute
integrals, the Riemann integral is
still essential. Lebesgue integrals
are more or less uncomputable unless you know a function as a limit
of Riemann-integrable functions in
an appropriate sense in the sense
of proposition 4.11.5, for instance.

Lebesgue integrals

501

Denition
in
4.11.6 (Lebesgue equality). Let fk , gk be two sequences
of R-integrable functions such that
;
;


n
|fk (x)||d x| < and
|gk (x)||dn x| < . 4.11.14
k=1

Rn

Rn

k=1

fk =
L

k=1

gk

if

k=1

fk (x) =

k=1

gk (x)

a.e.

4.11.15

k=1

## Theorem 4.11.7 will make it possible to dene the Lebesgue integral. It

is proved after denition 4.11.8.
Theorem
in
4.11.7. Let fk , gk be two sequences of R-integrable functions
such that
;
;


n
|fk (x)||d x| < ,
|gk (x)||dn x| < ,
4.11.16
k=1

Rn

Rn

k=1

and

fk =
L

k=1

Then

;


Rn

k=1

fk (x)|d x| =

gk .

4.11.17

k=1
;

k=1

Rn

gk (x)|dn x|.

4.11.18

## Thus the integral of a function f that is the sum of a series of R-integrable

functions as in 4.11.13 depends only on f and not on the series. So we can
now dene the Lebesgue integral.
Note that the series on the right
of equation 4.11.20 is convergent,
since (by part 4 of proposition
4.1.14) it is absolutely convergent:
;



n 

f
(x)|d
x|
 n k

R
;

Rn

## now apply proposition 1.5.34.

Denition
in
4.11.8 (Lebesgue integral). Let fk be a sequence of Rintegrable functions such that
;

|fk (x)||dn x| < .
4.11.19
k=1

Rn

## Then the Lebesgue integral of f =

L

;
Rn

f (x)|dn x| =

fk is

k=1

;

k=1

Rn

fk (x)|dn x|.

4.11.20

502

Equation 4.11.18:
;

fk (x)|dn x|
n
k=1 R

;


n
k=1 R

gk (x)|dn x|.

Chapter 4.

Integration

l
Proof of theorem 4.11.7. Set hk = fk gk , and Hl = k=1 hk . To
prove equation 4.11.18 (repeated in the margin), we need to show that
;
lim
Hl (x) |dn x| = 0.
4.11.21
l

Rn

## The Hl form a sequence of Riemann-integrable functions converging to

0 almost everywhere (by equation 4.11.17); if in addition they all have
support in BR (0) and |Hl | R for all l, then Hl meets the conditions for
fk in theorem 4.11.4, so
;

## In equation 4.11.22, the function 0 plays the role of f in

theorem 4.11.4.
To go from equation 4.11.22 to
equation 4.11.23, we can switch
integral and sum because the sum
is nite. (Remember that Hl in
equation 4.11.22 is a nite sum.)
By the
denitions
of fk and gk ,
>
the series
|hk | is convergent,
k
so by choosing M large enough, we
can make the tail of the series as
small as we like. This is why we
choose M rst, then the R of the
truncation.
The denition of [Hl ]R may be
easier to understand if equation
4.11.26 is rewritten as follows:
[Hl ]R (x) =

0
if |x| > R

R
if |x| R, Hl (x) > R

R
if |x| R, Hl (x) < R

Hl (x) otherwise.
The virtue of dening [Hl ]R as in
equation 4.11.26 is that it shows
that [Hl ]R is Riemann-integrable
(see corollary 4.3.5).

lim

Rn

Hl (x) |dn x| = 0

4.11.22

which gives
lim

l ;

k=1

Rn

hk (x) |dn x| = 0,

4.11.23

## proving the result.

Our strategy will be to reduce the general case, where Hl is not bounded
with bounded support, to this one, by appropriately truncating the Hl :
we will dene the truncation [Hl ]R and we will consider Hl as the sum
[Hl ]R +Hl [Hl ]R and consider separately the integral of [Hl ]R (see equation
4.11.27) and the integral of Hl [Hl ]R (the remainder of the proof).
Choose
> 0 and choose M such that
;


|hk (x)||dn x| <
,
4.11.24
k=M +1

Rn

## so that for l > M we have

;
|Hl (x) HM (x)||dn x|
Rn

;
l

k=M +1

|hk (x)||dn x|

Rn

k=M +1

Rn

4.11.25
|hk (x)||d x| <
.
n

Next choose R such that sup |HM (x)| < R/2 and HM (x) = 0 when
|x| R. We dene the R-truncation of Hl , denoted by [Hl ]R , by the
formula



[f ]R = sup R1BR (0) , inf R1BR (0) , f ,
4.11.26
illustrated by gure 4.11.2, and shown in another form in the margin. These
[Hl ]R form a sequence of Riemann-integrable functions all with support in
BR (0), satisfying |[Hl ]R | R and (by equation 4.11.17) tending to 0 almost
everywhere, so, by theorem 4.11.4,
;
lim
[Hl ]R (x) |dn x| = 0.
4.11.27
l Rn
\$
%&
'
main motor of the proof

4.11

Lebesgue integrals

503

## The volume of the union A of

cubes is 2 /R:

We have now done most of the work (the hard part was proving theorem
4.11.4). But for l > M , we still need to deal with the dierence

> UN |Hl HM |

MC |Hl HM | voln C
=

## Hl [Hl ]R = (Hl HM ) ([Hl ]R HM )

CDN (Rn )

MC |Hl HM | voln C

CDN (Rn)
CA

R
R
voln C =
voln A.
2
2
(Rn)

CDN
CA

R
BR (0)

## By equation 4.11.25, the integral of |Hl HM | is less than

, so we only need
to consider the integral of |[Hl ]R HM |. Outside BR (0) we have HM = 0
and [Hl ]R = 0, so we need only consider
;


[Hl ]R HM (x)|dn x|.
4.11.29
BR (0)

To see that this integral is small, rst nd N such that UN (|Hl HM |) <
.
Then consider the union A of the cubes C DN (Rn ) that intersect BR (0)
and where MC (|Hl HM |) > R/2. Since the upper sum UN (|Hl HM |)
is small, A must have small volume; the computation in the margin shows
that it is at most 2
/R.
Let B be the union of the cubes C DN (Rn ) that intersect BR (0) and
such that MC (|Hl HM |) R/2; on these,
|Hl | |Hl HM | + |HM | R/2 + R/2 = R

Figure 4.11.2.

4.11.28

4.11.30

(remember that sup |HM (x)| < R/2), so on B, we have [Hl ]R = Hl . Thus70
;




[Hl ]R (x) HM (x)|dn x|
BR (0)

; 
; 




 n

]
(x)

H
(x)
x|
+
[Hl R
[Hl ]R (x) HM (x)|dn x|
|d
M
A
B
; 

3R

 n

4.11.31
voln (A) + Hl (x) HM (x)|d x| 3
+
= 4
.
2
B
\$
%&
'

## The R-truncation of a function;

see equation 4.11.26. The thin line
is the graph of f ; the dark line
inf(R1BR (0) , f ). We take the sup
of the dark line and R1BR(0) ,
to get the thick, light gray line
representing [f ]R .

## In summary: For any R,

;
;
n
n
lim
Hl (x) |d x| = lim
[Hl ]R (x) |d x| + lim
(Hl (x) [Hl ]R (x)) |dn x| .
l Rn
l Rn
l Rn
\$
%&
' \$
%&
'
;

0 by equation 4.11.27

4.11.32

0 as shown below

## For the second integral on the right, rst choose

> 0, then M satisfying equation 4.11.24, then R as above (text immediately before equation
4.11.26). For these choices we have shown that for all l > M , we have
;
;
;
n
n
Hl (x) [Hl ]R (x) |d x| =
Hl (x) HM (x) |d x| [Hl ]R (x) HM (x) |dn x| .
4.11.33
n
Rn
Rn
\$
%&
' \$R
%&
'
4 by eq. 4.11.31

So

;
lim

l
70

Rn

## (Hl (x) [Hl ]R (x)) |dn x| = 0.

Equation 4.11.31: To get the 3R/2 in the last line, remember that
sup |HM (x)| < R/2

and

|[Hl ]R (x)| R.

4.11.34

504

Chapter 4.

Integration

## Some examples of Lebesgue integrals

All the integrals we computed earlier in this book are examples of Lebesgue
integrals, because of the following result.
Proposition 4.11.9 justies using the same integration symbol
for Lebesgue integrals that we use
for Riemann integrals.

Proposition
in
4.11.9. If f is R-integrable, then it is L-integrable, and
its Lebesgue integral equals its Riemann integral.
Proof. Just take f1 = f , and set fk = 0 for k = 2, 3, . . . . Clearly
everywhere, and
;
;

n
|fk (x)||d x| = |f1 (x)||dn x| <
k=1
;

k=1

fk = f

4.11.35

;
fk (x)|dn x| =
%&

'

Lebesgue integral

f1 (x)|dn x| <
\$
%&
'

4.11.36

Riemann integral

We now have a denition of the integral for all sorts of functions that
we couldnt have previously thought of integrating.
Example 4.11.10: Actually, we
are being too fancy. We could just
take all fk = 0; they converge to
0 everywhere, hence they converge
to f except on the rationals, which
forms a set of measure 0. Thus the
sort of local nonsense involved
in this function does not aect the
Lebesgue integral.
The function in example 4.11.10
is what we consider a pathological function, not really a good
example of the power of Lebesgue
integration.
The functions given in examples 4.11.11 and 4.11.12 are closer
to our real motivation for introducing Lebesgue integrals. Such
integrals show up everywhere in
physics, for example.

## Example 4.11.10 (L-integrable function that is not R-integrable).

Let f be the indicator function of the rationals, i.e.,

1
if x Q
f (x) =
4.11.37
0
otherwise.

It can be written f =
fk , as follows. Let a1 , a2 , . . . be a list of the
rational numbers, and consider the functions

1
if x = ak
fk (x) =
4.11.38
0
otherwise.
All the fk have integral 0 (they are 0 except at one point), so f is Lintegrable and has integral 0. 
Example 4.11.11 (An integrable function with unbounded support). In one variable, you probably studied improper integrals of functions that are not bounded or do not have bounded support: integrals like
;
1

dx = [arctan x] =
4.11.39
2
1
+
x

xn ex dx = n!

4.11.40

1
dx = [2 x]10 = 2.
x

4.11.41

;
0

4.11

B2
B1

The function

B0

f (x) =

Figure 4.11.3.
Example 4.11.11: In R2 , the
boundaries of the Bi form a sequence of concentric circles: B0 is
the unit disc, B1 is the circular
strip between the circle of radius
1 and the circle of radius 2, and so
forth.

1
2

505

## These improper integrals have analogues in higher dimensions, like

;
1
|dn x|.
4.11.42
n+1
Rn 1 + |x|

B3

1
4

Lebesgue integrals

ln x

Figure 4.11.4.
Example 4.11.12: We consider
f (x) = 1[0,1] (x) ln x
L

## as a sum of functions: the function

that is ln x between 1/2 and 1 and
0 elsewhere, plus the function that
is ln x between 1/4 and 1/2 and 0
elsewhere, plus the function that
is ln x between 1/8 and 1/4 and 0
elsewhere, . . . .

1
1 + |x|m

4.11.43

## is L-integrable on Rn if and only if m > n. As shown in gure 4.11.3, dene


?
@
4.11.44
Bi = x Rn  2i1 < |x| 2i , for i = 1, 2, . . .
and set B0 to be the unit ball. (We could set 2i1 |x| 2i ; this would
not aect any integrals, since the overlap would be a set of measure 0.)
Then


1
fi , where fi = 1Bi
.
4.11.45
f=
1
+
|x|m
i=0
(The function f0 is identical to f on the unit ball, and 0 elsewhere; f1 is
identical to f on B1 , and 0 elsewhere, . . . .)
Clearly the fi are R-integrable: they are bounded with bounded support,
and continuous except on two spheres, which certainly have measure 0.
Thus the substance of this example is to show that
;

fi (x)|dn x| < .
4.11.46
i=1

Rn

## The map i : x  2i1 x, for i 1, is a change of variables that takes

B1 to Bi . The change of variables formula gives
;
;
;
1
|fi (x)||dn x| =
f (x)|dn x| =
|dn x|
m
Rn
Bi
Bi 1 + |x|

; 


1
det[Di (x)]|dn x|
=

(x)
i
1 + |x|m
B1
;
;
1
2(i1)n
(i1)n n
=
2
|d
x|

|dn x|
i1
m
(i1)m
m
1
+
|2
x|
2
|x|
B1
B1
;
1
(i1)(nm)
n
=2
|d x|.
4.11.47
m
B1 |x|

If m > n, this geometric series converges, since the sum i 2(i1)(nm)
converges when m > n. 
Example 4.11.12 (An unbounded, integrable function). The function f (x) = 1[0,1] (x) ln x is L-integrable, even though it isnt bounded. As
L
shown in gure 4.11.4, it can be written as a sum of bounded functions:
f (x) =
L


i=0

fi (x),



where fi = 1(2(i+1) ,2i ] (x) ln x.

4.11.48

506

Chapter 4.

Integration

; 1/2i

|ln x| dx
i=0

## Example 4.11.13 shows that Lebesgue integrals are not a strict

generalization of improper integrals: some integrals exist as improper integrals that do not exist
as Lebesgue integrals. But there
is no version of Fubinis theorem
or the change of variables formula
for improper integrals that depend
on cancellations, and Lebesgue integration wisely forbids them.

4.11.49

1/2i+1

## Example 4.11.13 (A function that is not Lebesgue integrable).

Some improper one-dimensional integrals do not correspond to Lebesgueintegrable functions: integrals whose existence depends on cancellations,
like
;
sin x
dx.
4.11.50
x
0
As you may recall from one-variable calculus, this improper integral is dened to be
; A
sin x
lim
dx,
4.11.51
A 0
x
and we can show that the limit exists, for instance, by saying that the series
; (k+1)

sin x
dx
4.11.52
x
k
k=0

## Exercise 4.11.16 asks you to

justify the argument in example
4.11.13.

## is a decreasing alternating series whose terms go to 0 as k . But this

works only because positive and negative terms cancel: the area between
the graph of sin x/x and the x axis is innite, and there is no limit

; A
 sin x 

 dx.
lim
4.11.53
A 0  x 
The integral in equation 4.11.50 does not exist as a Lebesgue integral. By
our denition, f is L-integrable implies that |f | is L-integrable, which is
not the case here. Note that we would not want this integral to exist, since
the change of variables formula fails for this sort of improper integral (see
exercise 4.11.17).

## Some elementary properties of the Lebesgue integral

Just as Riemann integrals are
used to dene volume, Lebesgue
integrals can be used to dene
volume in the sense of Lebesgue,
known as measure. This is discussed in appendix A.22.

Proposition
in
4.11.14 (The Lebesgue integral is linear). If f and g
are L-integrable and a, b are constants, then af + bg is L-integrable and
;
;
;
(af + bg)(x)|dn x| = a
f (x)|dn x| + b
g(x)|dn x|.
Rn

Rn

Rn




Proof. If f = fk and g = gk , then af + bg = (afk + bgk ). Indeed,
L
L
L

the series (afk + bgk ) will converge except on the union of the sets where


the series k |f (x)| and k |g(x)| diverge, and the union of two sets of

4.11

Lebesgue integrals

507

measure 0 is still of measure 0 (see theorem 4.4.4). For x not in this set, the
result follows from the corresponding statement for series of numbers. 
It is not true that the product of two L-integrable functions is necessarily
1[1,1] (x)

L-integrable; for instance, the function
is L-integrable, but its
|x|
square is not. However, we have the following result.
Proposition
in
4.11.15. If f is L-integrable on Rn , and g is R-integrable
n
on R , then f g is L-integrable.

Proof. Since f is L-integrable, we can set f = k fk , where the functions
L
fk are R-integrable and
;
|fk (x)| ||dn x| < .
4.11.54
k

Rn

## We have f g = k fk g, where fk g is R-integrable; since g is bounded, we

L
have
;

;
n
|fk g(x)| ||d x| sup(|g|)
|fk (x)| ||dn x| < .
4.11.55
k

Rn

Rn

Therefore f g is L-integrable. 
We use the symbol to mean less than or equal except for x on a set
L

of measure 0.
Proposition
in
4.11.16. If f and g are L-integrable and f g, then
;

Rn

f (x)|dn x|

Rn

g(x)|dn x|.

4.11.56

Proof. By proposition
> 4.11.14, the statement is equivalent to saying that
0 g f implies 0 Rn (g f )(x) |dn x|. Replace g f by f ; the statement
L
>
becomes 0 f = 0 Rn f (x)|dn x|. Assume 0 f , and write
L

f=
L

fk

k=1

;


k=1

Rn

4.11.57

Dene
Fm =

m


fk ,

H0 = 0,

k=1

hk = Hk Hk1 .

4.11.58

508

Chapter 4.

Integration

## We have four cases:

If Fk1 (x) 0, Fk (x) 0, then |hk (x)| = |fk (x)|

If Fk1 (x) 0, Fk (x) < 0, then |hk (x)| = |0 Fk1 (x)| < |Fk (x) Fk1 (x)| = |fk (x)|

If Fk1 (x) < 0, Fk (x) 0, then |hk (x)| = |Fk (x) 0| < |Fk (x) Fk1 (x)| = |fk (x)|
If Fk1 (x) < 0, Fk (x) < 0, then |hk (x)| = 0 |fk (x)|.

|fk | = |hk |

.
0

## In all four cases, |hk | |fk |, as shown in gure 4.11.5. Moreover, hk is

an R-integrable function for every k, so by theorem 4.11.4,

;


Fk1 = Hk1 Fk = Hk

case 1

k=1

|fk |

|hk |

Fk1 0 = Hk1

Fk = Hk

case 2

Fk

0 = Hk

Fk1 = Hk1

h2

hm1

4.11.59

h1

## hk (x). Since we have assumed 0 f we have

L

limm Hm (x) = f (x) a.e. Thus we have f (x) = k=1 hk (x), and
k=1

;
Rn

f (x)|dn x| =

;

k=1

= lim

Rn

and

## Hm = Hm Hm1 + Hm1 Hm2 + + H2 H1 + H1 H0 , 4.11.60

\$ %& ' \$ %& '
\$
%&
' \$
%&
'

Figure 4.11.5.
In all three cases above, since

Rn

## (The second inequality is from equation 4.11.57.) Since H0 = 0 we can

write Hm as

so limm Hm (x) =

|hk |

;

k=1

hm

|fk |

case 3

Rn

|hk (x)||dn x|

hk (x)|dn x|

m ;


k=1

Rn

4.11.61
;

hk (x)|dn x| = lim

Rn

Hm (x) 0. 

## |fk | = |Fk Fk1 |,

we have
|hk | |fk |.
In case 1, we could have
Fk < Fk1 ,
since the fk can be negative, but
this would not change the result.
Case 4 is obvious.

The important theorems about Lebesgue integrals are Fubinis theorem
(theorem 4.11.20), the change of variables theorem (theorem 4.11.21), the
monotone convergence theorem (proposition A22.1), and the dominated
convergence theorem (theorem 4.11.19). They are all easier to state for
Lebesgue integrals than for Riemann integrals.
We start with theorem 4.11.17 on series; this isnt the usual way to
approach the topic, but in our treatment of Lebesgue integrals, theorem
4.11.17 is the foundational result. It is proved in appendix A.22. Note
that we are not assuming (as we do in denition 4.11.8) that the fk are
Riemann-integrable. If the fk are Riemann-integrable, the statement is
simply the denition of the Lebesgue integral.

4.11

Lebesgue integrals

509

Theorem
in
4.11.17 (A rst limit theorem for Lebesgue integrals).
Let fk be a series of L-integrable functions such that
;

|fk (x)||dn x| < .
4.11.62
Rn

k=1

Then f (x) =

k=1

integrable, and
;
;
f (x)|dn x| =
Rn

Rn

fk (x)|dn x| =

k=1

;

k=1

Rn

fk (x)|dn x|.

4.11.63

## Our next theorem says that if fk is an increasing sequence of positive

L-integrable functions, and the increasing sequence of numbers
;
;
;
n
n
f1 (x)|d x|
f2 (x)|d x|
f3 (x)|dn x| . . .
4.11.64
Rn

## Of course, the monotone convergence theorem also applies to

monotone decreasing sequences
that are bounded below by an Lintegrable function.

Rn

Rn

is bounded, then f = supk fk is L-integrable. Such a function can have horrible local behavior; the theorem asserts that the horrible behavior occurs
only on a set of measure 0, and that the Lebesgue integral doesnt see it.
Theorem
in
4.11.18. (Monotone convergence theorem: Lebesgue
integrals). Let 0 f1 f2 . . . be a sequence of L-integrable nonL

negative functions. If

;
sup
k

Rn

fk (x)|dn x| < ,

4.11.65

then the limit f (x) = limk fk (x) exists for almost all x, the function
f is L-integrable, and
;
;
f (x)|dn x| = sup
fk (x)|dn x|.
4.11.66
Rn

Rn

## Proof. Apply theorem 4.11.17 to the sup, rewritten as the series

f = f1 + (f2 f1 ) + + (fk fk1 ) + = g1 + g2 + + gk + . 

## In theorem 4.11.19, the fk are

dominated by F . It and theorem 4.11.19 are proved in appendix A.22.

Theorem
in
4.11.19 (Dominated convergence theorem: Lebesgue
integrals). Let fk be a sequence of L-integrable functions that converges pointwise to some function f almost everywhere. Suppose there
is an L-integrable function F : Rn R such that |fk (x)| F (x) for
almost all x. Then f is integrable and its integral is
;
;
n
f (x) |d x| = lim
fk (x) |dn x|.
4.11.67
Rn

Rn

510

Chapter 4.

Integration

Theorem
in
4.11.20 (Fubinis theorem for the Lebesgue integral).
Let f : Rn Rm R be an L-integrable function. Then the function
;
y 
f (x, y)|dn x|
4.11.68
Rn

## is dened for almost all y and is L-integrable on Rm , and in that case


;
; ;
f (x, y)|dn x||dm y| =
f (x, y)|dn x| |dm y|. 4.11.69

y
1

+1

Rn Rm

0
1

+1

Figure 4.11.6.
Part 3 of Fubinis theorem:
Consider the function f : R2 R
that is +1 in the darkly shaded region, 1 in the lightly shaded region, and 0 elsewhere. For every
y, the function
x  f (x, y)
is integrable, and in fact
;
f (x, y) dx = 0.
R

## Further, the function

;
f (x, y)|dn x|
y 
Rn

is L-integrable on Rm , and

; ;
;
f (x, y)dx dy =
0 dy
R

= 0.
But f is not integrable on R2 , and
the iterated integral written the
other way:

; ;
f (x, y)dy dx
R

Rn

Rm

## 1. every point (x, y) Rn+m is the center of a ball on which f is

L-integrable
2. the function x  |f (x, y)| is L-integrable on Rn for almost all y
>
3. the function y  Rn |f (x, y)||dn x| is L-integrable on Rm
then f is L-integrable on Rn+m , and
;
;
n
m
f (x, y)|d x||d y| =
Rn Rm

Rm

;


f (x, y)|d x| |dm y|.
n

Rn

Note that the absolute values in parts 2 and 3 of the converse are necessary, as shown by the function f : R2 R represented in gure 4.11.6.
Theorem
in
4.11.21 (The change of variables formula: Lebesgue
integrals). Let U, V be open subsets of Rn , and let : U V be
bijective, of class C 1 , and with inverse of class C 1 , such that both
and 1 have Lipschitz derivatives. Let f : V R be dened except
perhaps on a set of measure 0. Then
f is L-integrable on V (f )|det[D] | is L-integrable on U .
In that case,
;
;


f (v)|dn v| =
(f )(u) det[D(u)] |dn u|.
V

4.11.70

## Theorem 4.11.21 is proved in appendix A.22. the theorem is true if is

of class C 1 but does not have Lipschitz derivative, but our version of the
implicit function theorem doesnt allow us to prove it.

## is nonsense; the inner integral is

identically + for x > 0 and identically for x < 0. The prob>
lem is that y  Rn |f (x, y)||dn x|
(note the absolute values) is not
integrable.

## The Gaussian integral

The integral of the Gaussian bell curve is one of the most important integrals in all of mathematics. The central limit theorem (see section 4.6)
asserts that if you repeat the same experiment over and over, independently
each time, and make some measurement each time, then the probability

4.11

Lebesgue integrals

511

## that the average of the measurements will lie in an interval [a, b] is

; b
(xx)2

1
2

e 2 dx,
4.11.71
2
a
where x is the expected value of x, and is the standard deviation. Since
most of probability is concerned with repeating experiments, the Gaussian
integral is of the greatest importance.
Example 4.11.22 (Gaussian integral). For the central limit theorem to
make sense, the integrand of equation 4.11.71 must be a probability density,
i.e., we must have
;
(xx)2

1
2

4.11.72
e 2 dx,
2

## Exercise 4.11.10 asks you to use a change of variables to derive equation

4.11.72 from the simpler equation
;

2
ex dx = .
4.11.73

We need to justify equation 4.11.73. The function ex doesnt have an antiderivative that can be computed in elementary terms.71 ;One way to com

By Fubini,
;
2
A =

x2

 ;
dx

y 2


dy

;
=

R2

e(x

+y 2 )

ex dx = A.

|d2 x|.

4.11.74

;
; 2 ;
2
(x2 +y 2 ) 2
e
|d x| =
er r dr d.
4.11.75
R2

3
4


; 2 ;
2
er
1
2
r 2
A =
re dr d = 2
= 2 0
= . 4.11.76
2
2
0
0
0

So A = . 

## When does the integral of a derivative equal the derivative of

an integral?
Very often we will need to dierentiate a function that is itself an integral.
This is particularly the case for Laplace transforms and Fourier transforms,
71

## This is a fairly dicult result; see Integration in Finite Terms by R. Ritt,

Columbia University Press, New York, 1948. Of course, it depends on your
denition of elementary;
an antiderivative can be computed from the error
>x
2
function erf(x) = 2 0 et dt, which is a tabulated function and can be looked
up in tables.

512

Chapter 4.

Integration

## as we will see. If we integrate a function with respect to one variable,

and dierentiate with respect to a dierent variable, under what circumstances does rst integrating and then dierentiating give the same result as
rst dierentiating, then integrating? The dominated convergence theorem
gives the following very general result.
Theorem 4.11.23 is a major
result and has far-reaching consequences.
We need the condition of equation 4.11.78 so that we can apply
the dominated convergence theorem in the proof.

Theorem
in
4.11.23 (Dierentiating under the integral sign). Let
f (t, x) : Rn+1 R be a function such that for each xed t, the integral
;
F (t) =
f (t, x) |dn x|
4.11.77
Rn

## exists. Suppose that Dt f exists for almost all x. If there exists

> 0
and an L-integrable function g such that for all s
= t,


 f (s, x) f (t, x) 

 g(x),
|s t| <
= 
4.11.78

st
then F is dierentiable, and its derivative is
;
DF (t) =
Dt f (t, x) |dn x|.
Rn

## It would be hard to exagger- Proof. Just compute

ate the importance of the Fourier
;
f (t + h, x) f (t, x) n
transform, in pure and applied DF (t) = lim F (t + h) F (t) = lim
|d x|
h0
h0 Rn
h
h
mathematics. We discuss it further in Volume 2. A very pleas;
;
f (t + h, x) f (t, x) n
ant treatment is given in Fourier
=
lim
Dt f (t, x) |dn x|.
|d x| =
h
Analysis by T. W. K
orner.
Rn h0
Rn
The Greek letter is xi, pronounced ksee or kseye.
Equation 4.11.82: Note that
this is an integral of a complexvalued function. So far, we have
real-valued functions, but complex-valued functions introduce no
new diculties. If f = f1 + if2 ,
with f1 and f2 real valued, so that
they are the real and imaginary
parts of f , then we dene
;
;
;
f (x) dx = f1 (x) dx+i f2 (x) dx.
Passing from f to f (f hat)
is one of the central constructions
of mathematical analysis.

4.11.79

4.11.80

Moving the limit inside the integral sign is justied by the dominated convergence theorem for Lebesgue integrals (theorem 4.11.19). 

## Applications to the Fourier and Laplace transforms

Fourier transforms and Laplace transforms are essential tools of analysis.
Here we will study them only as examples of dierentiation under the integral sign. We will see that these transforms turn dierentiation (an analytical construction) into multiplication (algebraic). This is a large part of
why these transforms are so important.
If f is an L-integrable function on R, then so is f (x)eix for each R,
since
|f (x)eix |=|f (x)|.

4.11.81

Denition
in
4.11.24 (Fourier transform). The function f dened by
;
f() =
f (x)eix dx.
4.11.82
R

4.11

## Equation 4.11.83: We could

also write Df() as (f)
.
Equation 4.11.84 is the condition in theorem 4.11.23 given by
equation 4.11.78. Here, the increment h (a variable tending to 0)
corresponds to s t in equation
4.11.78.

Lebesgue integrals

513

## According to theorem 4.11.23,

;
;
;


ix
ix

dx =
Df () =
D f (x)e
f (x)D (e ) dx =
ixf (x)eix dx
R

F (),
= ixf
4.11.83
provided that the dierence quotients
 i(+h)x
 ihx


e


eix
1 
ix  e


f (x) = |e | 

 |f (x)|
\$ %& '
h
h

4.11.84

## are all bounded by a single L-integrable function. Since

we see that

The more times f is dierentiable (i.e., the greater the number p in equation 4.11.87), the
smoother it is. But that requires
x  xp f (x) to be an L-integrable
function, which means f must decrease reasonably fast at innity,
to compensate for multiplication
by (ix)p .
You might think that if f is Lintegrable, then f must tend to
0 at innity. This isnt true; for
instance f could have spikes of
height 1 and width 1/n2 at all
the integers. But if f
is also Lintegrable, then f does have to
tend to 0 at innity.
Equations 4.11.88 and 4.11.89:
Exercise 4.11.8 justies this integration by parts for Lebesgue
integrals.

## |eia 1| = 2| sin(a/2)| |a| for any a R,

4.11.85

 ihx

e 1

 |f (x)| |hx||f (x)| = |xf (x)|,


h
|h|

4.11.86

## so this is satised if x  |xf (x)| is an L-integrable function. More generally,

if the function x  xp |f (x)| is L-integrable, then f is p times dierentiable,
and
;
;


pf (). 4.11.87

Dp f() =
Dp eix f (x) dx = (ix)p f (x)eix dx = (ix)
R

In other words, the faster f decreases at innity, the smoother the Fourier
transform f is. This brings out one feature of the Fourier transform: growth
conditions on the original function f get translated into smoothness conditions for the Fourier transform.
Rather than dierentiating the Fourier transform, we might want to
Fourier transform the derivative, which we can do if f is of class C 1 and
both f and f  are L-integrable. This is best done by integration by parts:
; A
;
F
 )() =
(f
f  (x)eix dx = lim
f  (x)eix dx
A
A
R
4.11.88
; A
7
8
A
ix
ix
= lim f (x)e
lim
if (x)e dx.
x=A
A

## Since f and f are continuous and L-integrable, limx f (x) = 0, so

7
8A
4.11.89
lim f (x)eix A = 0,
A

## This replacement of the analytic operation of dierentiation

by the algebraic operation of multiplication is one reason why the
Fourier transform is so important
in the theory of dierential equations, especially partial dierential
equations.

and
F
 )() = i
(f

;
R

f (x)eix dx = i f().

4.11.90

Thus the Fourier transform turns dierentiation into multiplication. Example 4.11.25 shows that it turns dierential operators into multiplication
by polynomials, and hence solving dierential equations into division by
polynomials.

514

## Example 4.11.25 is step zero

of an immense eld of mathematics. For instance, the zeros of
the polynomial in the denominator
in equation 4.11.91 pose serious
problems, and the analogues in
higher dimensions are much more
serious yet.

Chapter 4.

Integration

## Example 4.11.25 (Fourier transform of a dierential equation).

The Fourier transform of both sides of the dierential equation


ap Dp f + + a0 f = g is
ap (i)p + ap1 (i)p1 + + a0 f = g,
\$
%&
'
product of f and a polynomial

which gives
f =

(i)p ap

g
.
+ + a0

4.11.91

If you know how to undo the Fourier transform, you can compute f from
this formula for f. 

## The Laplace transform

Denition
in
4.11.26 (Laplace transform). The Laplace transform
L(f ) of f is dened by the formula
;
L(f )(s) =
f (t)est dt.
4.11.92
0

## Depending on the range of values of s you are interested in, the

Laplace transform L(f ) exists for
quite a broad range of functions f .
For instance, if f is L-integrable,
then L(f ) is a continuous function
of s [0, ), and if f grows more
slowly than some polynomial, then
L(f ) is dened and continuous on
(0, ).
Even if f grows as fast as eat ,
the Laplace transform Lf is dierentiable when s > a.

## Note that the integral is from 0 to , not to . Again, under

appropriate circumstances, we can dierentiate under the integral sign:
;
;




D(Lf )(s) =
Ds f (t)est dt =
tf (t)est dt = L(tf ) (s).
0

Let us spell out the dierentiability, under the hypothesis that for xed
s > 0, the function t  est/2 f (t) is L-integrable. We need to show that
the limit
; (s+h)t
Lf (s + h) Lf (s)
e
est
lim
= lim
f (t) dt
h0
h0 0
h
h
;
eht 1
= lim
f (t)est
dt
4.11.93
h0 0
h
exists. When 0 < |h| s/2, the
 ht

e
1 
|ht|

est/2 , so

e
h



ht
 st
1 
st/2
e f (t) e
f (t).

e
h

4.11.94

## Thus, by the dominated convergence theorem, we have



;
;
eht 1
eht 1

st
st
(Lf ) (s) = lim
e f (t)
lim e f (t)
dt =
dt
h0 0
h0
h
h
0
;
est tf (t) dt.
4.11.95
=
0

4.11

Lebesgue integrals

515

## Exercises for section 4.11

4.11.1 a. Let x R2 . For what values of p R is |x|p L-integrable over the
unit disc in R2 ? (Think of polar coordinates.)
b. For those values, compute the integral.
4.11.2
a. Let x R3 . For what values of p R is |x|p L-integrable over the
unit ball in R3 ? (Think of spherical coordinates.)
b. For those values, compute the integral.
;
4.11.3 a. For what values of p R does

R2 B1 (0)

## integral? (Think of polar coordinates.)

b. For those values, compute the integral.
;
4.11.4 a. For what values of p R does

R3 B1 (0)

## integral? (Think of spherical coordinates.)

b. For those values, compute the integral.
;
4.11.5 For what values of p R does

Rn B1 (0)

## integral? (The answer depends on n.)

Set Am = ;
{0 y xm , 0 x 1}, for m R, m 0. For what
1
values of p R does
|dx dy| exist as a Lebesgue integral?
2 + y 2 )p
(x
Am
4.11.6

4.11.7

F (x) =

f (t) dt,
;

0
x

g(t) dt

G(x) =

## Repeat exercise 4.11.6, setting Am = {0 y xm , 1 x }.

; x
a. Let f be L-integrable on R. Show that F (x) =
f (t) dt is continu-

4.11.8
0
ous.
b. Show that integration by parts holds for Lebesgue integrals: if f, g are
L-integrable, and F and G are the functions dened in the margin, then
; b
; b
f (t)G(t) dt = F (b)G(b) F (a)G(a)
F (t)g(t) dt.
a

4.11.8, part b.

lim F (x) = 0.

## d. Show that if f, F, g, G are as in part b, and in addition F is L-integrable

and G is bounded, then
;
;
f (t)G(t) dt =
F (t)g(t) dt.

4.11.9

## Compute the Fourier transform of the indicator function of [1, 1].

4.11.10
Use a change of variables and equation 4.11.73, to show that the
integrand of equation 4.11.71 is a probability density.
4.11.11

## Show that for all polynomials p(x), the Lebesgue integral

;
2
p(x)e|x| |dn x| exists.
Rn

516

Chapter 4.

Integration

 
x = xa y b integrable over the region
y
x 0, y 0, xy 1? (This isnt too hard to do directly, but using the change
of variables formula given in theorem 4.11.21 makes it easier.)
4.11.12

## For what values of a and b is f

4.11.13 a. Show that the three sequences of functions in example 4.11.3 do not
converge uniformly.
*b. Show that the sequence of polynomials of degree m
pk (x) = a0,k + a1,k x + + am,k xm
does not converge uniformly on R, unless the sequence ai,k is eventually constant
for all i > 0 and the sequence a0,k converges.
c. Show that if the sequences ai,k converge for each i m, and A is a bounded
set, then the sequence pk 1A converges uniformly.
4.11.14

For the rst two sequences of functions in example 4.11.3, show that
;
;
[fk ]R (x) dx = lim lim
[fk ]R (x) dx.
lim lim
k R

4.11.15

R k

;

i=0

1/2

1/2i+1

## |ln x| dx of example 4.11.12 converges.

>
4.11.16 Show that the integral 0 sinx x dx of equation 4.11.50 is equal to the
sum of the series
5;
6

(k+1)

sin x
dx ,
x
k
k=1
and that this series converges.
;
sin x
4.11.17 Make the change of variables u = 1/x in the integral
dx.
x
0
Does the resulting integral exist as an improper integral, as described in example
4.11.13?
Exercise 4.11.18: You will need
the dominated convergence theorem (theorem 4.11.4) to prove this.
Part b: Except at 0 means
except at 0 of Pk , i.e., the zero
polynomial.

## 4.11.18 Let Pk be the space of polynomials

of degree at most k. Consider the
>1
function F : Pk R given by p  0 |p(x)| dx.
>1
a.> Compute F when k = 1 and k = 2, i.e., evaluate the integrals 0 |a + bx| dx
1
2
and 0 |a + bx + cx | dx (the second one is hard).
b. Show that F is dierentiable except at 0. Compute the derivative.
*c. Show that if p has only simple roots between 0 and 1, then F is twice
dierentiable at p.

## 4.12 Review exercises for chapter 4

Exercise 4.1:
UN (1C ) = LN (1C ).

4.1

## Show that if C D(Rm ), then 1C is integrable.

4.2 An integrand should take a piece of the domain and return a number, in
such a way that if we decompose a domain into little pieces, evaluate the integrand on the pieces, and add, the sums should have a limit as the decomposition