Optional additional material for G12MAN

Professor J.K. Langley

November 10, 2014

Introduction

The extra material included here is entirely OPTIONAL, and some of it is quite difficult. Material
included here but not in the lectures slides or problem sheets will definitely NOT be required for
the G12MAN examination.

1.1

A Fourier series formula

We take < x < and look at

S(x) =

X
(1)n+1 sin nx
n=1

= sin x

sin 2x sin 3x
+
....
2
3

The aim is to show that S(x) = x/2 for < x < .

To do this we take N N and set
YN = YN (x) = sin x

sin 2x
(1)N +2 sin(N + 1)x
+ ... +
.
2
N +1

By Eulers formula (G11LMA)

eit = cos t + i sin t,
we have YN = Im ZN , where
ei2x
(1)N +2 ei(N +1)x
ZN = eix
+ ... +
N +1 
 2 ix
e
(1)N eiN x
= eix 1
+ ... +
2
N +1


i(x+)
e
eiN (x+)
ix
= e
1+
+ ... +
2
N +1
Z 1
= eix
1 + sei(x+) + . . . + sN eiN (x+) ds.
0

Here we have used de Moivres formula (eit )n = eint for n Z and the fact that ei = 1. Now
0 < x + < 2, and so for 0 s 1 we have u = sei(x+) 6= 1 and hence
u(1 + u + . . . + uN ) = u + . . . + uN +1 = (1 + u + . . . + uN ) + uN +1 1,
1

which gives
1 + u + . . . + uN =

1 uN +1
.
1u

This means that

1

1 sN +1 ei(N +1)(x+)
ZN = e
ds =
1 sei(x+)
0
and YN is the imaginary part of this integral.
Now the complex number 1 + seix has modulus
ix

Z
0

eix + sN +1 ei(N +2)(x+)

ds,
1 + seix

|1 + seix | Re (1 + seix ) = 1 + s cos x.

Recall again that 0 s 1 and < x < . So if cos x 0 then 1+ s cos x c(x) = 1,
while if cos x < 0 then 1 + s cos x 1 + cos x = c(x) > 0. Since |eit | = cos2 t + sin2 t = 1 for
t R this means that the complex number
ei(N +2)(x+)
1 + seix
has modulus at most 1/c(x), and so has its imaginary part. Therefore we get

Z 1
Z 1 N +1 i(N +2)(x+) 
 N +1 i(N +2)(x+) 




s
e
s
e
Im



=
Im
ds
ds




ix
ix
1
+
se
1
+
se
0
0
Z 1 N +1
s
1

ds =
0
c(x)
(N + 2)c(x)
0
as N . This means that as N we have

 Z 1

Z 1


Z 1
eix
eix
1
Im
ds =
Im
YN Im
ds =
ds
ix
1 + seix
s + eix
0
0 1 + se
0
and so
Z 1
sin x
ds.
YN I(x) =
2
0 1 + s + 2s cos x
So we just need to show that I(x) = x/2.
To compute I(x) we can note that the substitution s = 1/t gives
Z
Z
sin x
dt
sin x
I(x) =
=
dt
2
1
2
2
1 + t + 2t cos x t
1 + t + 2t cos x
1
1
and so
Z
Z
sin x
sin x
ds =
2I(x) =
ds.
2
1 + s + 2s cos x
(s + cos x)2 + sin2 x
0
0
For 0 < x < we now substitute s + cos x = u sin x, ds = du sin x to get
Z
du
2I(x) =
= tan1 () tan1 (cot x)
2
cot x u + 1


 

= tan1 () tan1 tan
x =
x = x.
2
2
2
Clearly I(0) = 0, while if < x < 0 we can write I(x) = I(x) = (x/2) = x/2.
We conclude this section by remarking that, although this proof is evidently hard, Fourier
series are extremely powerful and versatile objects and the module G12DEF will focus not on
proof but rather on how to compute and use them.
2

1.2

Proof of the monotone sequence theorem (from G11ACF)

If the real sequence (xn ) is non-decreasing (i.e. xn xn+1 ) for n N then xn tends to a limit
(finite or +).
(You need to know the result, but will not be asked to reproduce this proof.)
Proof. We look at the set {xn : n N } = A. Suppose first that A is not bounded above.
This means that if we are given some positive number M , then no matter how large M might
be, we can find some member of the set A, say xm , such that xm > M . But then, because the
sequence is non-decreasing, we have xn > M for all n m. So we obtain:
some n1 such that xn > 10 for all n n1 ;
some n2 such that xn > 100 for all n n2 ;
some n3 such that xn > 1000 for all n n3 ;
some nq such that xn > 10q for all n nq .
This tells us that limn xn = +.
Now suppose that A is bounded above, and let s be the supremum (least upper bound) of
the set A. Suppose we take a real number t < s. Then t is not an upper bound for A, and so
some xm has xm > t, which implies that t < xn s for all n m. So this time we obtain:
some n1 such that s 1/10 < xn s for all n n1 ;
some n2 such that s 1/102 < xn s for all n n2 ;
some nq such that s 1/10q < xn s for all n nq .
This tells us that limn xn = s.
The corresponding statement for non-increasing sequences (yn ) (namely, that (yn ) either converges or tends to ) is deduced by looking at xn = yn .

1.3

Limits of subsequences

The sequence (sin n) (n = 1, 2, . . .) is clearly bounded, and so has a convergent subsequence

by the Bolzano-Weierstrass theorem, but it is not obvious what the limit of a convergent subsequence would be. The answer is given by the following:
Fact A: Let [1, 1]. Then there exists a subsequence of (sin n) with limit .
To prove this it is enough to establish:
Fact B: Let [1, 1], let > 0, and let N N. Then there exists n N with n N and
| sin n | < .
To get Fact A from Fact B, take n1 1 with | sin n1 | < 1, then n2 1 + n1 with
| sin n1 | < 1/2, and so on.
How to prove Fact B? First of all, we take a large integer M and divide the interval [2, 2]
into 2M subintervals I1 , . . . , I2M each of length 2/M . If we choose M large enough then for
3

any x and y in the same Ij we will have | sin x sin y| < . This can be seen from the graph
and follows from the fact that
Z x




| sin x sin y| = 
cos t dt |x y|.
y

Now choose Ip [2, 0] and Iq [0, 2] such that each of Ip and Iq contains a point where
the sine function takes the value . So now to prove Fact B it is enough to show that we can
find an integer n N and an integer m(n ) such that n + m(n )2 belongs to either Ip or
Iq ; this is because sin(n + m(n )2) = sin n .
To do this we look at all real numbers of form n + m2 with m and n integers. If m, n, p, q
are all integers and n + m2 = q + p2 then we must have n = q and m = p: this is because is
irrational. So for each positive integer n we choose an integer q(n) such that n+q(n)2 [0, 2).
There are infinitely many of these numbers and for different n their values are different. So we
can find two of these numbers which are as close together as we like, and by subtracting these
two we can find a positive integer r and an integer s(r) such that |r + s(r)2| is as small as we
like.
So we can find a sequence of positive integers rk and integers s(rk ) such that 0 6= xk =
rk + s(rk )2 0 as k . We can assume all these xk are different (since none of them
are 0 but they tend to 0) and that the integers rk are different (because for a given rk N
there is only one s(rk ) Z such that rk + s(rk )2 [1/2, 1/2]). So this gives us an integer
r N and an integer s(r) such that r + s(r)2 [2/M, 0) (0, 2/M ]. Now by taking
r + s(r)2, 2r + 2s(r)2, 3r + 3s(r)2, . . ., we can find a positive integer t such that tr + ts(r)2
belongs to Ip or Iq , and our required n is n = tr.
The final comment is that not much in this argument depends on the function in question
being sine; something similar can be done for cosine, and indeed the argument can be adapted
for any continuous function h : R R which has an irrational period, in which case we need to
take h(R).

1.4

Distance in Rd , the Cauchy-Schwarz inequality, and a proof

of the triangle inequality

Take x and y in Rd . If we choose R then x + y Rd and

0 kx + yk2 =

d
X

(xj + yj )2 .

j=1

This gives, for every R,

0

d
X
j=1

x2j

+ 2

d
X
j=1

xj y j +

d
X

yj2 = C + B + A2 = Q().

j=1

Since this quadratic Q() is at least 0 for all real we get

!2
! d
!
d
d
X
X
X
B2 = 4
xj y j
4AC = 4
x2j
yj2 .
j=1

j=1

j=1

This yields the famous Cauchy-Schwarz inequality (G11ACF): for real xj , yj ,

!2
! d
!
d
d
X
X
X
xj y j

x2j
yj2 .
j=1

j=1

j=1

Taking square roots gives

 v

! d
!
d
d
 u
X
X
X
u


xj y j  t
x2j
yj2 .



j=1

j=1

j=1

Now we take = 1 to get

kx + yk2 =

d
d
d
d
X
X
X
X
(xj + yj )2 =
x2j + 2
xj y j +
yj2 ,
j=1

j=1

j=1

j=1

and so
kx + yk2



d
d
X
 X


2

xj + 2 
xj y j  +
yj2


j=1
j=1
j=1
v
!
!
u
d
d
d
d
X
X
X
X
u
2
2
2
t

xj + 2
xj
yj +
yj2
d
X

j=1

j=1

j=1

j=1

v
2
v
u d
u d
uX
uX
x2j + t
yj2 = (kxk + kyk)2 .
= t
j=1

j=1

This proves the triangle inequality.

1.5

Expressing real numbers in binary

Let 0 x 1. We can write x in binary (base 2) form

x = 0 a1 a2 a3 . . . =

a1 a2 a3
+ 2 + 3 + ...,
2
2
2

in which each aj is either 0 or 1. Clearly we can write 0 = 0 0. We also have

1=

1 1 1

+ + + . . . = 0 1.
2 4 8

Now take any x with 0 < x < 1. To show that x has a base 2 expansion we first let a1 be the
largest member of {0, 1} such that a1 /2 x. This of course implies that (a1 + 1)/2 > x. The
expansion is then generated inductively using the following.
Claim. For n N there are integers a1 , . . . , an , each of which is 0 or 1, such that
a1
an
a1
an
1
+ ... + n x <
+ ... + n + n.
2
2
2
2
2
5

(1)

This is proved by induction on n, and we have already seen this for n = 1. Assuming that (1) is
given we reach the next stage by letting an+1 be the largest integer such that
a1
an an+1
+ . . . + n + n+1 x.
2
2
2
Then (1) shows that an+1 is either 0 or 1. The fact that an+1 is chosen as large as possible
ensures that
an an+1
an an+1
a1
a1
1
+ . . . + n + n+1 x <
+ . . . + n + n+1 + n+1 ,
2
2
2
2
2
2
2
and this completes the induction.

1.6

Monotone surjective functions and continuity

Theorem 12.4: Let I and J be intervals in R (not necessarily bounded) and let the function
f : I J be non-decreasing and onto (i.e. surjective i.e. f (I) = J). Then f is continuous on I.
To prove Theorem 12.4 take any in I, and any sequence (xn ) in I with limit . We have
to show that limn f (xn ) = f (). So take > 0.
Suppose first that f () is not an endpoint of J. Then we can find A and B in J such that
f () < A < f () < B < f () + .
Because J = f (I), we can find s and t in I such that
f () < f (s) = A < f () < B = f (t) < f () + .
Because f is non-decreasing on I we must have s < < t. But xn , and so there exists
some integer N such that s < xn < t for all n N . This gives, for all n N , since f is
non-decreasing,
f () < f (s) f (xn ) f (t) < f () + ,
and hence |f (xn ) f ()| < . Since can be chosen arbitrarily small, we get limn f (xn ) =
f ().
Now we take the case where f () is the right-hand endpoint of J. If J consists of one point
only then obviously f is constant on I and so continuous. Assuming that J does not consist of
one point only, we can find A in J such that
f () < A < f ().
Because J = f (I), we can find s in I such that
f () < f (s) = A < f (),
and we must have s < . But xn , and so there exists some integer N such that s < xn
for all n N . This gives, for all n N , since f is non-decreasing and f () is the right-hand
endpoint of J,
f () < f (s) f (xn ) f (),
and hence |f (xn ) f ()| < . Since can be chosen arbitrarily small, we get limn f (xn ) =
f ().
Similar modifications apply when f () is the left-hand endpoint of J.
6

1.7

A non-decreasing function on R which is discontinuous at every rational x

The set Q is countable, and so it is possible to form a sequence (pn ) (n = 1, 2, . . .) in which

each pn is a rational number, and every rational number appears at least once.
Let fn (x) = 0 if x < pn , with fn (x) = 1/3n if x pn . Now let
f (x) =

fk (x).

k=1

This converges by the comparison test from G11ACF, since 0 fk (x) 3k for each k. Also f
is non-decreasing, because all of the fk are.
Now take any n N, and put
X
gn (x) =
fk (x).
kN,k6=n

Again, gn is non-decreasing, and so limxpn gn (x) exists and is at most gn (pn ) (see Ch. 12).
Since limxpn fn (x) = 0, we have
lim f (x) = lim gn (x) gn (pn ) < gn (pn ) + fn (pn ) = f (pn ),

xpn

xpn

and this proves that f is discontinuous at pn .

1.8

A continuous nowhere differentiable function

This example is due to van der Waerden (1930).

Let n be a non-negative integer. For any real number x, define fn (x) to be the distance from
x to the nearest rational number of the form m/10n , with m an integer. Now we define
f (x) =

fn (x).

n=0

Note that |fn (x)| < 10n for all x and for all n N. So the sum converges, and gives a
non-negative function f which is positive at a lot of points. Also f is continuous by the M -test.
Take any real number x. We will show that f is not differentiable at x.
For each positive integer q, we choose a new number yq as follows. The real number x belongs
to some interval of the form [m/10q , (m + 1)/10q ), with m Z, and this interval can be divided
into halves, namely [m/10q , (m + 1/2)/10q ) and [(m + 1/2)/10q , (m + 1)/10q ). Now we put
yq = x 1/10q+1 , with the chosen so that yq lies in the same half as x.
The idea is to compare fn (x) and fn (yq ). If n > q then since yq is x shifted by 10q1 ,
which is an integer multiple of 10n , you do not change the distance to the nearest point of form
( integer )/10n , so fn (x) = fn (yq ) for n > q.
But x and yq belong to the same half of the interval [m/10q , (m + 1)/10q ). Hence for n q
these points x and yq both belong to the same interval [r/10n , (r + 1)/10n ) (where r Z) and
7

indeed to the same half of this interval. So for n q the nearest point of form ( integer )/10n
is the SAME for x and yq , and hence fn (yq ) fn (x) = (yq x) for n q. This means that
q

f (yq ) f (x) X
1 = m(q).
=
yq x
n=0
We do not know the value of m(q), but it is an integer, and m(1) is 0, 2 or 2. Moreover, each
time we increase q by 1, we add 1 to m(q) and so the parity of m(q) changes. Indeed, m(q)
must be even if q is odd and odd if q is even.
Now let q . We find that yq x, but
f (yq ) f (x)
yq x
cannot tend to a finite limit, as it is alternately even and odd!

1.9

Theorem 13.3 (the product rule etc.)

Suppose that f and g are differentiable at a, and that R. Then:

(i) (f + g)0 (a) = f 0 (a) + g 0 (a) ;
(ii) (f )0 (a) = f 0 (a) ;
(iii) (f g)0 (a) = f 0 (a)g(a) + f (a)g 0 (a) ;
(iv) if g(a) 6= 0, then (1/g)0 (a) = g 0 (a)/g(a)2 .
Proof. Parts (i) and (ii) are very easy.
(iii) As x a, we have
f (x)g(x) f (a)g(x) f (a)g(x) f (a)g(a)
f (x)g(x) f (a)g(a)
=
+
=
xa
xa
xa




f (x) f (a)
g(x) g(a)
= g(x)
+ f (a)
g(a)f 0 (a) + f (a)g 0 (a).
xa
xa
(iv) Again, as x a, we have
(1/g(x)) (1/g(a))
g(a) g(x)
g 0 (a)
=

.
xa
(x a)g(x)g(a)
g(a)2

1.10

An explicit calculation of a Riemann integral without using

antiderivatives

Consider the curve y = x2 , 0 x 1, and the area A bounded by this curve, the positive x-axis,
and the line from (1, 0) to (1, 1) (the area under the curve). We determine this area without
using calculus. Of course, calculus tells us to expect the answer
x3
A=
3


1
0

1
= .
3

We proceed by dividing [0, 1] into n equal sub-intervals, each of length 1/n, this for n N.
Then the part of the required area between x = (k 1)/n and x = k/n may be enclosed in a
rectangle of base 1/n and height (k/n)2 . Thus
!
n
n
X
X
(k/n)2
n(n + 1)(2n + 1)
A
=
k 2 n3 =
.
3
n
6n
k=1
k=1
(The last fact used here is easily proved by induction on n.) Now n(n + 1)(2n + 1)/6n3 is always
greater than 1/3 and tends to 1/3 as n .
We can also see that the part of the required area between x = (k 1)/n and x = k/n
encloses a rectangle of base 1/n and height ((k 1)/n)2 . Thus
!
n
X
(n 1)n(2n 1)
2
.
A
(k 1) n3 =
6n3
k=1
The last quantity is always less than 1/3 and tends to 1/3 as n . Thus A must equal 1/3.

1.11

The effect of refinement on Riemann sums

Part of Lemma 14.1: Let f be a bounded real-valued function on I = [a, b].

If P, Q are partitions of I and Q is a refinement of P , then
L(P, f ) L(Q, f ),

U (P, f ) U (Q, f ).

Proof. We first prove this for the case where Q is P plus one extra point. The general case then
follows by adding points one at a time. So suppose that Q is the same as P , except that it has
one extra vertex c, where xk1 < c < xk . Then
U (Q, f ) U (P, f ) = (c xk1 ) sup{f (x) : xk1 x c} +
+(xk c) sup{f (x) : c x xk }
(xk xk1 ) sup{f (x) : xk1 x xk }.
This is using the fact that all other terms cancel. Since the first two suprema are not greater
than the third we see that U (Q, f ) U (P, f ) 0. The proof for the lower sums uses the same
idea.

1.12

Proof of Theorem 14.3

Theorem 14.3: Let a < c < b and let f : [a, b] R be a bounded function which is Riemann
integrable on [a, c] and [c, b]. Then f is Riemann integrable on [a, b] and
Z b
Z c
Z b
f (x) dx =
f (x) dx +
f (x) dx.
a

Proof. Take > 0. Since the Riemann integral is the infimum of the upper sums we can find a
partition P of [a, c] and a partition Q of [c, b] such that
Z c
Z b
U (P, f ) <
f (x) dx + , U (Q, f ) <
f (x) dx + .
a

But if we combine the vertices of P and Q we can make a partition S of [a, b] which satisfies
U (S, f ) = U (P, f ) + U (Q, f ).
So we have
b

Z
f (x) dx U (S, f )

Z
f (x) dx +

f (x) dx + 2.
c

In the same way, since the Riemann integral is the supremum of the lower sums we can find a
partition P 0 of [a, c] and a partition Q0 of [c, b] such that
Z c
Z b
0
0
L(P , f ) >
f (x) dx , L(Q , f ) >
f (x) dx .
a

Combining these gives a partition S 0 of [a, b] with

L(S 0 , f ) = L(P 0 , f ) + L(Q0 , f )
and so
Z

f (x) dx L(S , f )
a

f (x) dx 2.

f (x) dx +
a

Hence
Z

Z
f (x) dx

Z
f (x) dx

f (x) dx + 4,
a

and since we can choose arbitrarily small we must have

Z

1.13

Z
f (x) dx =

f (x) dx =
a

Z
f (x) dx +

f (x) dx.
c

Riemann integration on a subinterval

Let a < c < b and let f be a bounded real-valued function on [a, b] which is Riemann integrable
on [a, b]. We prove that f is Riemann integrable on [a, c] and on [c, b].
To do this, take > 0 and partitions R and S of [a, b] such that
b

Z
U (R, f ) <

f (x) dx +
a

and
Z

f (x) dx .

L(S, f ) >
a

We can do this because the upper integral is the infimum of the upper sums and the lower integral
is the supremum of the lower sums. We can assume that c is a vertex of R and S, and indeed that
R and S have the same vertices, because adding more vertices cannot increase U nor decrease
L. Since the upper and lower integrals are the same this gives us
U (R, f ) L(R, f ) < 2.
10

But we can now split R into a partition P of [a, c] and a partition Q of [c, b], and
U (R, f ) = U (P, f ) + U (Q, f ),

L(R, f ) = L(P, f ) + L(Q, f ).

Hence
U (P, f ) L(P, f ) + U (Q, f ) L(Q, f ) = U (R, f ) L(R, f ) < 2.
In particular, we get
0 U (P, f ) L(P, f ) < 2.
But U (P, f ) is at least the upper integral from a to c of f , and L(P, f ) is at most the lower
integral. Hence we get
Z c
Z c
f (x) dx
f (x) dx < 2.
0
a

Since can be as small as we like we must have

Z c
Z c
f (x) dx =
f (x) dx,
a

so f is Riemann integrable on [a, c], and the same argument works for [c, b].

1.14

The Riemann integral of a sum

Let f and g be bounded real-valued functions on [a, b] R which are both Riemann integrable
on [a, b]. We prove that so is f + g and that
Z b
Z b
Z b
f (x) + g(x) dx =
f (x) dx +
g(x) dx.
a

Take a partition P = {x0 , x1 , . . . , xn } of [a, b]. Now

sup{f (x)+g(x) : xk1 x xk } sup{f (x) : xk1 x xk }+sup{g(x) : xk1 x xk }.
Through this argument, and a similar one using the infimum, we obtain
Mk (f + g) Mk (f ) + Mk (g),

mk (f + g) mk (f ) + mk (g).

So we have
U (P, f + g) U (P, f ) + U (P, g),

L(P, f + g) L(P, f ) + L(P, g).

Take > 0, and choose partitions R and S of [a, b] such that

Z
U (R, f ) <

Z
f (x) dx + ,

U (S, g) <

g(x) dx + .
a

Let P be the partition formed by combining the vertices of R and S. Since refinement cannot
increase upper sums, this yields
Z b
Z b
Z b
f (x) + g(x) dx U (P, f + g) U (P, f ) + U (P, g) <
f (x) dx +
g(x) dx + 2.
a

11

In the same way we find that

Z b
Z b
Z b
f (x) + g(x) dx L(P, f + g) L(P, f ) + L(P, g) >
f (x) dx +
g(x) dx 2.
a

But
b

f (x) dx,

f (x) dx =
a

and the same is true of g. Hence we have

b

Z
f (x) + g(x) dx

Z
f (x) + g(x) dx <

f (x) + g(x) dx + 4.

Because can be chosen arbitrarily small, we must have

Z

1.15

g(x) dx.

f (x) dx +

f (x) + g(x) dx =

f (x) + g(x) dx =
a

Uniform convergence and the Riemann integral

The pointwise limit of Riemann integrable functions does not have to be Riemann integrable (see
the problem sheets). However, the following is true.
Theorem: Let [a, b] be a closed interval in R and let (fn ) be a sequence of (bounded) real-valued
functions on [a, b], each of which is Riemann integrable there. Suppose that as n the fn
converge uniformly on [a, b] to a function f : [a, b] R. Then f is Riemann integrable on [a, b]
and
Z b
Z b
lim
fn (x) dx =
f (x) dx.
n

Proof. We note first that each fn has to be bounded (say |fn (x)| Mn < for all x [a, b])
in order to make all the suprema Mk (fn ) and infima mk (fn ) finite, so that the upper and lower
sums exist. Here we are not assuming that all the Mn are the same.
On the other hand, by the uniform convergence, we can take N N such that
sup{|fn (x) f (x)| : x [a, b]} 1
for all n N . So for all x [a, b] we have
|f (x)| |fN (x)| + |f (x) fN (x)| MN + 1
and so the limit function f is bounded on [a, b]. This implies that the upper and lower integrals
Z

Z
f (x) dx,

f (x) dx
a

definitely exist, and the upper integral is at least the lower integral.
12

Now take any real > 0. By the uniform convergence, we can take S N such that
sup{|fn (x) f (x)| : x [a, b]}
for all n S. If P = {x0 , . . . , xm } is a partition of [a, b] and n S then we have
Mk (f ) Mk (fn ) + ,

U (P, f ) U (P, fn ) +

m
X

(xk xk1 ) = U (P, fn ) + (b a).

k=1

But fn is Riemann integrable, so we can choose P with U (P, fn ) close to

gives
Z b
Z b
fn (x) dx + (1 + b a).
f (x) dx U (P, f ) <

Rb
a

fn (x) dx, and this

Similarly, we have, for n S,

mk (f ) mk (fn ) ,

L(P, f ) L(P, fn ) (b a),

and we can choose P with L(P, fn ) close to

Z

Z
f (x) dx L(P, f ) >

Rb
a

fn (x) dx, which gives

f (x) dx 2(1 + b a).

fn (x) dx (1 + b a) >
a

Since we can choose as small as we like, we deduce that the upper and lower integrals of f are
the same, so that f is Riemann integrable, and that
Z b

Z b



 (1 + b a)
f
(x)
dx

f
(x)
dx
n


a

for all n S, which finishes the proof.

13