Вы находитесь на странице: 1из 25

MATH3068 LECTURE NOTES. HANDOUT 5.

FOURIER SERIES
DONALD I. CARTWRIGHT
1. Definitions
A trigonometric polynomial is a function s : R R (or R C) which can
be expressed by a formula
s(x) =
1
2
a
0
+
n

k=1
_
a
k
cos(kx) + b
k
sin(kx)
_
,
where the a
k
and b
k
are real or complex numbers. If a
n
and b
n
are not both zero,
then n is called the degree of s(x). The factor
1
2
in the constant term is there to
make some of our formulas below a little neater.
Examples 1.1. (a) Any constant function s(x) c is a trigonometric polyno-
mial of degree 0, with a
0
= 2c.
(b) e
ix
is a trigonometric polynomial of degree 1, because e
ix
= cos(x)+i sin(x).
(c) s(x) = sin(x)+
1
2
sin(2x)+
1
3
sin(3x)+
1
4
sin(4x) is a trigonometric polynomial
of degree 4.
A trigonometric series is a series
1
2
a
0
+

k=1
_
a
k
cos(kx) + b
k
sin(kx)
_
(1.1)
(which for a particular x R may or may not converge), where again the a
k
and
b
k
are real or complex numbers. The trigonometric polynomial
s
n
(x) =
1
2
a
0
+
n

k=1
_
a
k
cos(kx) + b
k
sin(kx)
_
(1.2)
is called the n-th partial sum of this trigonometric series.
Our goal in this chapter is to prove that under certain very general conditions, a
function f(x) can be represented by a trigonometric series in either the pointwise
or mean-square sense (see the denitions below). We shall apply this representa-
tion of f(x) to help us solve some partial dierential equations.
If L > 0, a function f : R R (or R C) is called L-periodic or of period L if
f(x+L) = f(x) for each x R. Since any trigonometric polynomial is 2-periodic,
it is natural to assume that the functions f(x) we are dealing with are also 2-
periodic. This is not essential, for if f(x) is L-periodic, then g(x) = f(Lx/2) is
2-periodic, and we can study g(x) instead of f(x).
Sometimes f(x) is merely dened on an interval of length 2, such as [, ].
But then (if f() = f()) we can dene f(x) in an obvious way for all x R in
such a way that f(x) becomes a 2-periodic function on R. This process is called
extending f(x) by periodicity.
1
2 D. CARTWRIGHT
Examples 1.2. (a) The function f(x) given on [, ] by f(x) = x
2
, then ex-
tended by periodicity, has graph:

2
Typical functions which arise in practice include
(b) the sawtooth function, with graph

3
and
(c) the square-wave function, whose graph is of the type:

3
To handle such possibilities, we shall allow our functions to be discontinuous
at some points, but not too many. For the sake of brevity, we shall call a real or
complex-valued function f(x) piecewise continuous if
f(x) is 2-periodic,
f(x) has at most nitely many discontinuities on any interval of length 2,
and
f(x) is bounded, i.e., there is a constant M such that |f(x)| M for all
x R.
Of course, to check that a 2-periodic function is bounded, one needs only check
that it is bounded on, say, [, ].
Given a (piecewise continuous) function f(x), how are we to nd numbers a
k
and b
k
so that the trigonometric series (1.1) sums to f(x)? It turns out that a good
choice to make for these numbers is the following:
a
k
=
1

f(x) cos(kx) dx k = 0, 1, 2, . . . (1.3)


MATH3068 LECTURE NOTES. HANDOUT 5. 3
and
b
k
=
1

f(x) sin(kx) dx k = 1, 2, 3, . . . (1.4)


These numbers are called the Fourier cosine and sine coecients of f(x). The
hypothesis of piecewise continuity ensures, amongst other things, that the integrals
in (1.3) and (1.4) are dened.
The series (1.1), with coecients given by formulas (1.3) and (1.4), is called the
Fourier series of f(x). Thus Fourier series are certain special trigonometric series:
those arising from piecewise continuous functions by specifying the coecients by
means of (1.3) and (1.4). Many trigonometric series are not Fourier series. For
example, it follows from Theorem 2.9 below that

k=1
1

k
sin(kx)
is not the Fourier series of any piecewise continuous function, even though it con-
verges for each x R.
Starting from a function f(x), why do we choose the a
k
and b
k
as above? The jus-
tication for making this choice depends on the orthogonality relations. These
say that if k, {0, 1, 2, . . .}, then
1

cos(kx) cos(x) dx =
_

_
0 if k = ;
1 if k = 1;
2 if k = = 0.
1

sin(kx) sin(x) dx =
_
0 if k = ;
1 if k = 1;
1

sin(kx) cos(x) dx = 0 for any k, .


These relations may be easily veried using such identities as
cos() cos() =
1
2
_
cos( + ) + cos( )
_
.
Suppose now that a function f(x) is given, and that there is a trigonometric
series (1.1) which converges to f(x) at each point x [, ] (equivalently, at
each point x R). Recall that this means that the n-th partial sum (1.2) of (1.1)
converges to f(x) for each x. If N, then
s
n
(x) cos(x) f(x) cos(x) as n
for all x [, ]. We might therefore expect that
1

s
n
(x) cos(x) dx
1

f(x) cos(x) dx as n . (1.5)


But the orthogonality relations imply that
1

s
n
(x) cos(x) dx =
_
0 if 0 n <
a

once n
So assuming that (1.5) holds,
1

f(x) cos(x) dx = a

must hold, i.e., the a


k
must be given by (1.3). Similarly, we might expect that
1

s
n
(x) sin(x) dx
1

f(x) sin(x) dx as n . (1.6)


4 D. CARTWRIGHT
The orthogonality relations imply that
1

s
n
(x) sin(x) dx =
_
0 if 0 n <
b

once n
So assuming that (1.6) holds,
1

f(x) sin(x) dx = b

.
So the b
k
must be given by (1.4). So the Fourier series of a function f(x) is the
only trigonometric series which could possibly converge to f(x) so nicely that (1.5)
and (1.6) are valid.
Remark 1. Let us emphasize here that we have not proved that the Fourier series
of f(x) converges to f(x). Indeed, there exist continuous 2-periodic functions
whose Fourier series do not converge at some points x. The above work is merely
a justication for making the choices (1.3) and (1.4) of coecients.
If we start from a trigonometric polynomial
s(x) =
1
2
a
0
+
n

k=1
_
a
k
cos(kx) + b
k
sin(kx)
_
and substitute in cos(kx) = (e
ikx
+ e
ikx
)/2 and sin(kx) = (e
ikx
e
ikx
)/2i, then
collecting terms, we nd that
s(x) =
1
2
a
0
+
n

k=1
_
a
k
_
e
ikx
+ e
ikx
2
_
+ b
k
_
e
ikx
e
ikx
2i
_
_
=
n

k=n
c
k
e
ikx
where
c
k
=
_

_
1
2
a
0
if k = 0,
1
2
(a
k
ib
k
) if k 1,
1
2
(a

+ ib

) if k = < 0.
(1.7)
The a
k
and b
k
can be expressed in terms of the c
k
:
a
k
= c
k
+ c
k
, if k 0, b
k
= i(c
k
c
k
) if k 1 (1.8)
Starting from a trigonometric series
1
2
a
0
+

k=1
_
a
k
cos(kx) + b
k
sin(kx)
_
(1.9)
we can form the series

k=
c
k
e
ikx
(1.10)
by dening the c
k
using the formulas (1.7). Conversely, starting from the se-
ries (1.10), we can form a trigonometric series (1.9), dening the a
k
and b
k
using
the formulas (1.8). A series of the form (1.10) is called a trigonometric series in
complex form, and the nite sum
n

k=n
c
k
e
ikx
MATH3068 LECTURE NOTES. HANDOUT 5. 5
is called the n-th partial sum of the series (1.10). No assertions are being made
here about convergence. The relationship between the series (1.9) and (1.10) is
purely formal. All that is true in general is that the series have equal partial sums:
1
2
a
0
+
n

k=1
_
a
k
cos(kx) + b
k
sin(kx)
_
=
n

k=n
c
k
e
ikx
for n = 0, 1, . . .
if the c
k
are related to the a
k
and b
k
via formulas (1.7) and (1.8).
When (1.9) is the Fourier series of a function f(x), the corresponding coecients
c
k
are called the complex Fourier coecients of f(x). There is a single formula
for these coecients:
c
k
=
1
2
_

f(x)e
ikx
dx k = 0, 1, 2, . . . (1.11)
To see this, suppose rst that k 1. Then
c
k
=
1
2
(a
k
ib
k
)
=
1
2
_
1

f(x) cos(kx) dx
1

f(x) sin(kx) dx
_
=
1
2
_

f(x)
_
cos(kx) i sin(kx)
_
dx
=
1
2
_

f(x)e
ikx
dx
The checking of the formulas when k 0 is left as an exercise.
The orthogonality relations can be neatly expressed by a single formula:
1
2
_

e
ikx
e
ix
dx =
_
1 if k = ,
0 if k = .
Of course, e
ix
here is another way of writing e
ix
, the bar over e
ikx
denoting
complex conjugation.
The Fourier cosine and sine coecients a
k
and b
k
of a function f(x) are related
to its complex Fourier coecients c
k
by (1.7) and (1.8). Notice that the a
k
and b
k
are real when f(x) is real-valued, in which case the a
k
, b
k
and c
k
are related by
a
k
= 2 Re c
k
b
k
= 2 Imc
k
when f is real valued. (1.12)
Here Re and Im mean real part of and imaginary part of, respectively.
Example 1.1. Suppose that R, and dene a function f(x) be setting
f(x) =
_
e
x
if 0 < x < 2
e
2
+1
2
if x = 0 or x = 2,
and then extending f(x) by periodicity. See the diagram below.
The easiest way to calculate the Fourier coecients of f(x) is to rst calculate
the complex Fourier coecients c
k
, and then to use (1.12). An easy calculation
shows that
c
k
=
1
2
_
2
0
f(x)e
ikx
dx =
1
2
_
2
0
e
(ik)x
dx =
1
2
e
(ik)x
ik

2
x=0
=
e
2
1
2( ik)
=
e
2
1
2
+ ik

2
+ k
2
6 D. CARTWRIGHT
Now using (1.12), we nd that
a
k
=
(e
2
1)
(
2
+ k
2
)
and b
k
=
k(e
2
1)
(
2
+ k
2
)
.

y = e
x
If instead of using (1.12) we use the usual formulas for a
k
and b
k
, we have to work
harder:
a
k
=
1

_
2
0
e
x
cos(kx) dx
=
1

_
2
0
e
x
d
dx
_
sin(kx)
k
_
dx if k = 0
=
e
x

_
sin(kx)
k
_

2
x=0

_
2
0
_
sin(kx)
k
_
(e
x
) dx (integrating by parts)
=

k
_
2
0
e
x
sin(kx) dx (which is

k
b
k
by the way)
=

k
_
2
0
e
x
d
dx
_
cos(kx)
k
_
dx
=
e
x
k
_
cos(kx)
k
_

2
x=0


k
_
2
0
_
cos(kx)
k
_
(e
x
) dx (int. by parts again)
=
(e
2
1)
k
2



2
k
2
1

_
2
0
e
x
cos(kx) dx
=
(e
2
1)
k
2



2
k
2
a
k
.
Adding

2
k
2
a
k
to both sides, we get
_
1 +

2
k
2
_
a
k
=
(e
2
1)
k
2

,
from which the formula a
k
=
(e
2
1)
(
2
+k
2
)
follows by a little algebra. We need to do
a separate calculation to nd a
0
:
a
0
=
1

_
2
0
f(x) dx =
1

_
2
0
e
x
dx =
e
2
1

.
To get b
k
, we would normally have to do a similar amount of work, but fortunately
we noticed that a
k
=

k
b
k
during the last calculation, and so the formula b
k
=

k(e
2
1)
(
2
+k
2
)
is obtained almost for free.
2. Convergence theorems for Fourier series
In this section we shall state the main theorems concerning convergence, and
give several examples. The proofs will be given later in the chapter.
MATH3068 LECTURE NOTES. HANDOUT 5. 7
Theorem 2.1. Let f(x) be a 2-periodic piecewise continuous function. Suppose
that f(x) is dierentiable at the point x
0
R. Then the Fourier series of f(x) at
x
0
converges to f(x
0
). That is,
1
2
a
0
+

k=1
_
a
k
cos(kx
0
) + b
k
sin(kx
0
)
_
= f(x
0
) (2.1)
More generally, the same is true if we merely assume that f(x) has left and right
derivatives at x
0
Example 2.1. Let f(x) be the function dened on [0, 2] by the following formula,
then extended by periodicity
f(x) =
_
x
2
if 0 < x < 2,
0 if x = 0 or x = 2.

2 3 4

2
3
4

An easy calculation shows that its Fourier cosine coecients a


k
are all 0 (because
f(x) is odd) and that its k-th Fourier sine coecient is 1/k. Thus its Fourier series
is

k=1
sin(kx)/k. Now f(x) is dierentiable at each point which is not a multiple
of 2. In particular, it is dierentiable at each point x of the interval (0, 2), and
so its Fourier series must converge to f(x) = ( x)/2 there:

k=1
sin(kx)
k
=
x
2
if 0 < x < 2
In this example, the theorem tells us nothing about what happens when x is a
multiple 2m of 2, because f(x) is not dierentiable at such x. But sin(kx) = 0
for all k N if x = 2m, so the Fourier series converges to 0 at such a point. But
f(x) is dened to be 0 if x is a multiple of 2. Thus, in this example, the Fourier
series of f(x) converges to f(x) at every point x R.
Example 2.2. Consider the function dened on [0, 2] by the following formula,
then extended by periodicity:
f(x) =
x
2
4

x
2
+

2
6
2 4 2 4
Being even, its Fourier sine coecients b
k
are all 0, while it is easy to calculate that
its k-th Fourier cosine ceocient a
k
is 1/k
2
if k 1, and that a
0
= 0. Thus its
Fourier series is

k=1
cos(kx)/k
2
. The function is continuous everywhere, and is
dierentiable at any point other than a multiple of 2. At any multiple x
0
of 2,
8 D. CARTWRIGHT
f(x) has left derivative /2 and right derivative /2. So Theorem 2.1 is applicable
at each point x
0
R. Thus

k=1
cos kx
k
2
=
x
2
4

x
2
+

2
6
for 0 x 2
In particular, taking x = 0, we obtain the famous formula

k=1
1
k
2
=

2
6
. (2.2)
When a piecewise continuous function f(x) has a discontinuity of a particularly
simple type at x
0
R, it is still possible to prove convergence of the Fourier series
at x
0
. We rst need to dene exactly what sort of discontinuities we can handle.
A function f : R C is said to have a simple jump discontinuity at the
point x
0
R if
(a) f(x) has left- and right-hand limits at x
0
. That is,
f(x
0
+) = lim
h0,h>0
f(x
0
+ h) and f(x
0
) = lim
h0,h>0
f(x
0
h) exist.
(b) f(x) has left- and right-hand slopes at x
0
. That is,
m
R
= lim
h0,h>0
f(x
0
+ h) f(x
0
+)
h
and m
L
= lim
h0,h>0
f(x
0
) f(x
0
h)
h
exist.

slope=mR
slope=mL
f(x
0
)
f(x
0
+)
f(x
0
)
x
0
In Example 2.1, if we take x
0
= 0, then f(x
0
+) = /2, f(x
0
) = /2 and
m
R
= m
L
= 1/2.
Theorem 2.2. Let f(x) be a 2-periodic piecewise continuous function. Suppose
that f(x) has a simple jump discontinuity at the point x
0
R. Then the Fourier
series of f(x) at x
0
converges to (f(x
0
) + f(x
0
))/2. That is,
1
2
a
0
+

k=1
_
a
k
cos(kx
0
) + b
k
sin(kx
0
)
_
=
f(x
0
+) + f(x
0
)
2
(2.3)
Example 2.3. In Example 1.1, we calculated the Fourier coecients of the function
dened by setting f(x) = e
x
for 0 < x < 2, and setting f(x) = (e
2
+ 1)/2 at
x = 0, then extending by periodicity:

y = e
x
MATH3068 LECTURE NOTES. HANDOUT 5. 9
Notice that f(x) has a simple jump discontinuity at x
0
if x
0
is any multiple of 2.
The value set for f(x) at x = 0 is chosen so that f(x
0
) = (f(x
0
+) + f(x
0
))/2
holds at such points x
0
. So Theorem 2.2 tells us that the Fourier series of f(x)
converges to f(x) at every point x. So
e
2
1
2
+

k=1
_
(e
2
1)
(
2
+ k
2
)
cos(kx)
k(e
2
1)
(
2
+ k
2
)
sin(kx)
_
converges to f(x) at each point x. So if 0 < x < 2,
e
2
1

_
1
2
+

k=1
cos(kx) k sin(kx)

2
+ k
2
_
= e
x
.
That is,
1
2
+

k=1
cos(kx) k sin(kx)

2
+ k
2
=
e
x
e
2
1
if 0 < x < 2.
On the other hand, the Fourier series, evaluating at x = 0 converges to (e
2
+1)/2.
That is,
e
2
1

_
1
2
+

k=1

2
+ k
2
_
=
e
2
+ 1
2
.
So
1
2
+

k=1

2
+ k
2
=

2
e
2
+ 1
e
2
1
.
Remark 2. Theorem 2.1 is a special case of Theorem 2.2. For if f(x) is dierentiable
at x
0
, then according to the above denition, it has a simple jump discontinuity
at x
0
, with f(x
0
+) = f(x
0
) = f(x
0
) and m
L
= m
R
= f

(x
0
). The jump is 0
here, of course.
Remark 3. It is easy to give examples of functions which fail to have left or right
derivatives at some points x
0
, or are discontinuous at some points, with disconti-
nuities which are more complicated than simple jump discontinuities. When this
happens, we cannot be sure that the Fourier series of the function converges to f(x
0
)
(or to anything). The theorem is not applicable.
Our next convergence theorem is of a completely dierent nature. It asserts that
the Fourier series of any piecewise continuous function always converges to f(x) in
the following mean square sense. Suppose that f(x) and f
0
(x), f
1
(x), f
2
(x), . . .
are piecewise continuous 2-periodic functions. We say that f
0
(x), f
1
(x), f
2
(x), . . .
converges to f(x) in the mean square sense if
_

|f(x) f
n
(x)|
2
dx 0 as n .
It is easy to give examples in which f
n
(x) f(x) as n for each xed x, but
f
n
(x) does not tend to f(x) in the mean square sense. In other words, pointwise
convergence does not imply mean square convergence. It is also easy to give
examples in which f
n
(x) f(x) as n in the mean square sense, but f
n
(x)
does not tend to f(x) at any particular x. In other words, mean square convergence
does not imply pointwise convergence.
Examples 2.1. (a) Let f
n
(x) be dened for x [0, 2] by
f
n
(x) =
_

n if
1
2n
x
1
n
,
0 for all other x [0, 2].
and then extended by periodicity.
10 D. CARTWRIGHT

n
1
2n
1
n
2
Then f
n
(x) 0 for each xed x [0, 2]. For if 0 < x 2, then f
n
(x) = 0 once
n > 1/x, since then 1/n < x 2. If x = 0, then f
n
(x) = 0 for all n, so that again,
f
n
(x) 0 as n . On the other hand,
_
2
0
|f
n
(x) 0|
2
dx =
_ 1
n
1
2n
n dx =
1
2
0.
(b) Let f
0
(x), f
1
(x), f
2
(x), . . . be the sequence of functions on [0, 2] whose rst few
terms have the following graphs:
1
2
y=f0(x) y=f1(x) y=f2(x) y=f3(x)
Then f
n
0 in the mean square sense, because, when we form
_
2
0
|f
n
(x) 0|
2
dx,
we see that we are integrating the constant 1 over an interval. For the rst function,
that interval is all of [0, 2], so that
_
2
0
|f
n
(x)|
2
dx = 2. For the next two
functions, the interval is half of [0, 2], and so
_
2
0
|f
n
(x)|
2
dx = . For the next four
functions, the interval is a quarter of [0, 2], and so
_
2
0
|f
n
(x)|
2
dx =

2
. For the
next eight functions, the interval is an eighth of [0, 2], and so
_
2
0
|f
n
(x)|
2
dx =

4
for these ns. Continuing like this, we see that
_
2
0
|f
n
(x)|
2
dx 0 as n , so
that f
n
(x) 0 in the mean square sense. On the other hand, f
n
(x
0
) 0 holds for
no x
0
[0, 1], because f
n
(x
0
) = 1 for innitely many ns, and also f
n
(x
0
) = 0 for
innitely many ns.
MATH3068 LECTURE NOTES. HANDOUT 5. 11
Theorem 2.3. Let f(x) be any 2-periodic piecewise continuous function. Let
s
n
(x) denote the n-th partial sum of the Fourier series of f(x). Then s
n
(x) f(x)
in the mean square sense:
_

|f(x) s
n
(x)|
2
dx 0 as n . (2.4)
Moreover, Parsevals Formula holds:
1
2
_

|f(x)|
2
dx =

k=
|c
k
|
2
=
1
4
|a
0
|
2
+
1
2

k=1
_
|a
k
|
2
+|b
k
|
2
_
.
(2.5)
We shall see below that Parsevals Formula is an immediate consequence of (2.4),
for it turns out that
_

|f(x) s
n
(x)|
2
dx =
_

|f(x)|
2
dx
_

|s
n
(x)|
2
dx
and it easy to show that
1
2
_

|s
n
(x)|
2
dx =
n

k=n
|c
k
|
2
=
1
4
|a
0
|
2
+
1
2
n

k=1
_
|a
k
|
2
+|b
k
|
2
_
Examples 2.2. (a) If we take f(x) to be the function of Example 2.1, and calculate
both sides of Formula (2.5), we nd that Parsevals Formula yields

k=1
1/k
2
=

2
/6, as we found before.
(b) If we take f(x) to be the function of Example 2.2, and calculate both sides
of Formula (2.5), we nd that Parsevals Formula yields

k=1
1/k
4
=
4
/90.
The proofs of the above theorems are quite short, and we shall learn a lot of
interesting extra facts about Fourier series along the way.
Aside: Some general facts about integration.
We often need to integrate functions. If f(x) is a bounded function dened on
a bounded interval [a, b], then f(x) is called Riemann integrable over [a, b]
if there is a number A with the following property: Given any > 0, there is a
partition a = a
0
< a
1
< < a
n
= b of [a, b], and there are numbers m
i
and
M
i
for i = 1, . . . , n so that m
i
f(x) M
i
for all x [a
i1
, a
i
], and so that
the lower Riemann sum

n
i=1
m
i
(a
i
a
i1
) and the upper Riemann sum

n
i=1
M
i
(a
i
a
i1
) are both within of A:
A <
n

i=1
m
i
(a
i
a
i1
) A and A
n

i=1
M
i
(a
i
a
i1
) < A + .
The number A is called the integral of f(x) over [a, b], and is denoted
_
b
a
f(x) dx.
This quantity has many interpretations, the main one being that it equals the area
under the curve y = f(x) (when f(x) 0 for all x [a, b]). The lower and upper
Riemann sums can be thought of as sums of areas of rectangles.
We illustrate this in the case n = 5. The sums of the rectangles in the two
diagrams are examples of lower and upper Riemann sums. In this example, the
partition of [a, b] is into 5 small intervals [a
i1
, a
i
] of equal length, so that a
i
a
i1
=
(b a)/5 for each i. In general, the partitions used need not be into subintervals
of equal length.
12 D. CARTWRIGHT
a0=a a1 a2 a3 a4
a5=b a0=a a1 a2 a3 a4
a5=b
The above is revision of material you learned about in rst year. In your rst
year course, the following basic fact was stated:
Theorem 2.4. Suppose that a function f(x) is dened and continuous at each
point of the bounded interval [a, b]. Then f(x) is Riemann integrable over [a, b].
The proof of this theorem is beyond the level of this course. In fact, we need a
slight generalization, whose proof we omit as well:
Theorem 2.5. Suppose that a function f(x) is dened and bounded on [a, b], and
has at most a nite number of discontinuities there. Then f(x) is Riemann inte-
grable over [a, b].
In particular, any piecewise continuous (2-periodic) function f(x) is Riemann
integrable.
We shall need to integrate complex-valued functions. If g(x) is a complex-valued
function on an interval [a, b], the integral
_
b
a
g(x) dx is dened in the obvious way:
one writes g(x) = u(x) + iv(x), where u(x) and v(x) have real values, and then
denes
_
b
a
g(x) dx =
_
b
a
u(x) dx + i
_
b
a
v(x) dx
provided that u(x) and v(x) are Riemann integrable. We call g(x) Riemann in-
tegrable if u(x) and v(x) are. The real and imaginary parts of
_
b
a
g(x) dx are
_
b
a
u(x) dx and
_
b
a
v(x) dx, i.e.,
Re
_
_
b
a
g(x) dx
_
=
_
b
a
Re
_
g(x)
_
dx Im
_
_
b
a
g(x) dx
_
=
_
b
a
Im
_
g(x)
_
dx
The usual rules for manipulating integrals remain valid for complex-valued func-
tions. For example, if g
1
and g
2
are Riemann integrable complex-valued functions
on [a, b], and if
1
,
2
C, then
1
g
1
+
2
g
2
is Riemann integrable on [a, b], and
_
b
a
_

1
g
1
(x) +
2
g
2
(x)
_
dx =
1
_
b
a
g
1
(x) dx +
2
_
b
a
g
2
(x) dx
One important fact has a slightly tricky proof.
Lemma 2.6. If g(x) is a Riemann integrable complex-valued function on an inter-
val [a, b], then

_
b
a
g(x) dx


_
b
a
|g(x)| dx
MATH3068 LECTURE NOTES. HANDOUT 5. 13
Proof. Write
_
b
a
g(x) dx = re
i
, where r 0 and R. Then

_
b
a
g(x) dx

= r = e
i
_
_
b
a
g(x) dx
_
=
_
b
a
e
i
g(x) dx
= Re
_
_
b
a
e
i
g(x) dx
_
because r R
=
_
b
a
Re
_
e
i
g(x)
_
dx

_
b
a
|g(x)| dx
because Re
_
e
i
g(x)
_
|g(x)| for all x [a, b], and because of the monotonicity
property of integrals of real-valued functions. (The function |g| is a real-valued
Riemann integrable function.)
Lemma 2.7. Let f(x) be a piecewise continuous 2-periodic function on R. Then
the integral of f(x) over an interval of length 2 does not depend on the particular
interval chosen. That is, for any a R we have
_
a+2
a
f(x) dx =
_

f(x) dx
Proof. The proof may be easily understood with the aid of the following diagram
2k 2(k+1)
a
a+2
Given a, pick k Z such that 2k a < 2(k + 1). Then
_
a+2
a
f(x) dx =
_
2(k+1)
a
f(x) dx +
_
a+2
2(k+1)
f(x) dx.
The second integral on the right is the right hand shaded area in the diagram. It
therefore equals the left hand shaded area in the diagram. Instead of appealing to
the picture, we can set x = t + 2 and use the periodicity of f(x). This means
_
a+2
2(k+1)
f(x) dx =
_
a
2k
f(t + 2) dt =
_
a
2k
f(t) dt.
Therefore
_
a+2
a
f(x) dx =
_
a
2k
f(x) dx +
_
2(k+1)
a
f(x) dx
=
_
2(k+1)
2k
f(x) dx
=
_
2
0
f(u) du setting x = u + 2k.
So I(a) =
_
a+2
a
f(x) dx does not depend on a, so that I(a) = I().
Now let f(x) be a piecewise continuous function, and let its Fourier cosine and
sine coecients be a
k
, k = 0, 1, . . ., and b
k
, k = 1, 2, . . ., respectively. Let s
n
(x)
14 D. CARTWRIGHT
be the n-th partial sum of the Fourier series of f(x). We substitute in the integral
formulas (1.3) and (1.4) for the a
k
and b
k
, and we nd that
s
n
(x) =
1
2
a
0
+
n

k=1
_
a
k
cos(kx) + b
k
sin(kx)
_
equals
1
2
_
1

f(t) dt
_
+
n

k=1
_
_
1

f(t) cos(kt) dt
_
cos(kx) +
_
1

f(t) sin(kt) dt
_
sin(kx)
_
which equals
1
2
_

f(t)
_
1 + 2
n

k=1
_
cos(kx) cos(kt) + sin(kx) sin(kt)
_
_
dt
This simplies to
1
2
_

f(t)
_
1 + 2
n

k=1
cos(k(x t))
_
dt =
1
2
_

f(t)D
n
(x t) dt
where
D
n
(x) = 1 + 2
n

k=1
cos(kx)
We have therefore found the following integral formula for s
n
(x):
s
n
(x) =
1
2
_

f(t)D
n
(x t) dt =
1
2
_

f(x t)D
n
(t) dt (2.6)
The second formula is obtained from the rst by substituting t = x u:
1
2
_

f(t)D
n
(x t) dt =
1
2
_
x
x+
f(x u)D
n
(u) (1)du.
=
1
2
_

f(x t)D
n
(t) dt, by Lemma 2
The function D
n
(x) is a trigonometric polynomial, called Dirichlets kernel. It
is easy to prove that
D
n
(x) =
sin
_
(n +
1
2
)x
_
sin
_
1
2
x
_ (2.7)
if x is not a multiple of 2 (while D
n
(x) = 2n + 1 if x = 2k, k Z).
3. Proof of the theorems on pointwise convergence
The proof of Theorem 2.2 follows fairly quickly after we have proved the following
key lemma.
Lemma 3.1. (The Riemann Lebesgue Lemma) Let f : [a, b] C be Riemann
integrable, where < a < b < . Then
_
b
a
f(x)e
ix
dx 0 as ( R) (3.1)
Thus
_
b
a
f(x) cos(x) dx 0 and
_
b
a
f(x) sin(x) dx 0 as (3.2)
MATH3068 LECTURE NOTES. HANDOUT 5. 15
Proof. Notice that (3.1) implies (3.2) and vice versa, because e
ix
= cos(x) +
i sin(x), while cos(x) = (e
ix
+ e
ix
)/2 and sin(x) = (e
ix
e
ix
)/2i. It is
easier to work with (3.1).
Let us rst see how easy this lemma is if we assume a bit more about f(x).
When f(x) has a continuous derivative, we can use integration by parts, and get
(if = 0)
_
b
a
f(x)e
ix
dx =
f(b)e
ib
f(a)e
ia
i

1
i
_
b
a
f

(x)e
ix
dx
Thus (because |e
ix
| = 1 for , x R)

_
b
a
f(x)e
ix
dx


1
||
_
|f(b)| +|f(a)| +
_
b
a
|f

(x)| dx
_
=
C
||
say
0 as ||
But we shall have to prove the lemma for very general functions f(x), and the
proof goes as follows:
(a). First assume that the interval is divided into r subintervals [a
j1
, a
j
], where
a = a
0
< a
1
< a
2
< < a
r
= b, and assume that f(x) takes the constant value c
j
throughout the interval (a
j1
, a
j
) (we dont care what values f(x) takes at the
points a
j
):
a0 a1 a2 a3 a4 a5 a6
Such a function is called a step function. Then
_
b
a
f(x)e
ix
dx =
r

j=1
_
aj
aj1
f(x)e
ix
dx
=
r

j=1
c
j
_
aj
aj1
e
ix
dx
=
r

j=1
c
j
e
iaj
e
iaj1
i
So using |e
it
| = 1 for t R, and the triangle inequality,

_
b
a
f(x)e
ix
dx


2
||
r

j=1
|c
j
| =
C
||
0
(b). Now let f(x) be any real-valued Riemann integrable function.
Lemma 3.2. If f(x) is any Riemann integrable function dened on [a, b], and if
> 0, then there is a step function g(x) such that
_
b
a
|f(x) g(x)| dx < .
16 D. CARTWRIGHT
If |f(x)| M for all x [a, b], then g(x) can be chosen so that |g(x)| M for all
x too.
Proof. According to the denition of Riemann integrability, if > 0, there is a
partition a = a
0
< a
1
< < a
r
= b of [a, b] such that
_
b
a
f(x) dx < L
_
b
a
f(x) dx, where L is a lower Riemann sum of f(x) for the partition. That is,
L =
r

j=1
m
j
(a
j
a
j1
)
where m
j
f(x) for all x [a
j1
, a
j
]. Notice that L equals
_
b
a
g(x) dx, where g(x)
is the step function taking the constant value m
j
on each interval (a
j1
, a
j
). Also,
g(x) f(x) holds for all x [a, b] (if g(x) is set equal to some large negative value
at each point a
j
). Thus
_
b
a
|f(x) g(x)| dx =
_
b
a
(f(x) g(x)) dx =
_
b
a
f(x) dx
_
b
a
g(x) dx
=
_
b
a
f(x) dx L < .
If |f(x)| M for all x [a, b], then the numbers m
j
can clearly be chosen to satisfy
|m
j
| M for all j too. Hence |g(x)| M for all x.
Now for R, part (a) of the proof, applied to g(x), shows that there is a

so large that

_
b
a
g(x)e
ix
dx

< once ||

. So

_
b
a
f(x)e
ix
dx

_
b
a
(f(x) g(x))e
ix
dx +
_
b
a
g(x)e
ix
dx

_
b
a
|(f(x) g(x))e
ix
| dx +

_
b
a
g(x)e
ix
dx

=
_
b
a
|f(x) g(x)| dx +

_
b
a
g(x)e
ix
dx

< +

_
b
a
g(x)e
ix
dx

< + = 2 for ||

The lemma is now proved.


Proof of Theorem 2.2.
Let f(x) be a piecewise continuous function having a simple jump discontinuity
at x
0
. We must show that s
n
(f(x
0
+) + f(x
0
))/2 as n , where
s
n
=
1
2
_

f(x
0
t)D
n
(t) dt
We start by turning this into an integral over [0, ]. Using the fact that D
n
(t) is
an even function of t,
_
0

f(x
0
t)D
n
(t) dt =
_

0
f(x
0
+ t)D
n
(t) dt
MATH3068 LECTURE NOTES. HANDOUT 5. 17
Also,
_

D
n
(t) dt = 2, and D
n
(t) is even, and so
_

0
D
n
(t) dt = . Thus
s
n

f(x
0
+) + f(x
0
)
2
=
1
2
_

0
_
f(x
0
+ t) + f(x
0
t)
_
D
n
(t) dt
f(x
0
+) + f(x
0
)
2
=
1
2
_

0
_
f(x
0
+ t) + f(x
0
t)
_
f(x
0
+) + f(x
0
)
__
D
n
(t) dt
=
1
2
_

0
g(t) sin
_
(n +
1
2
)t
_
dt
(3.3)
where
g(t) =
_
f(x
0
+ t) f(x
0
+)
t

f(x
0
) f(x
0
t)
t
_
t
sin
_
1
2
t
_ if 0 < t
Notice that t/ sin(
1
2
t) 2 as t 0. Thus g(t) 2(m
R
m
L
) as t 0. So
if we set g(0) = 2(m
R
m
L
), g(x) is continuous at 0, has only nitely many (if
any) discontinuities on [0, ], and is bounded on [0, ]. It is therefore Riemann
integrable, and so the Riemann Lebesgue Lemma shows that
_

0
g(t) sin((n +
1
2
)t) dt 0 as n .
In view of Formula (3.3), the theorem is proved.
4. Cesaro convergence of Fourier series
There are continuous 2-periodic functions f(x) for which s
n
(x) f(x) as
n at some points x. This is a complication, but it cannot be avoided unless
we place greater restrictions on our functions f(x). What is true, however, is a
somewhat weaker assertion: s
n
(x) f(x) in the Cesaro sense as n .
Denition 1. Let s
0
, s
1
, s
2
, . . . be a sequence of real or complex numbers. For
N = 1, 2, . . ., form the averages

N
=
1
N
N1

n=0
s
n
.
We say that (s
n
) converges to a limit in the Cesaro sense if
N
as N .
Example 4.1. Let s
n
= (1)
n
for each n 0. Certainly s
0
, s
1
, s
2
, . . . does not tend
to any limit. On the other hand,

N
=
_
0 if N is even;
1
N
if N is odd.
Thus (s
n
) converges to 0 in the Cesaro sense.
We next check that Cesaro convergence is really a weaker type of convergence.
Lemma 4.1. If a sequence (s
n
) converges to a limit in the usual sense, then it
converges to in the Cesaro sense.
18 D. CARTWRIGHT
Proof. Let > 0 be given. There is an n

such that |s
n
| < /2 once n n

.
Once N > n

, we have
|
N
| = |
1
N
N1

n=0
s
n
|
= |
1
N
N1

n=0
_
s
n

_
|

1
N
N1

n=0
|s
n
|
<
1
N
n1

n=0
|s
n
| +
1
N
N1

n=n

2
=
K
N
+
N n

2
, say,

K
N
+

2
< once N > 2K/
Thus
N
as N .
We next nd an integral formula for

N
(x) =
1
N
N1

n=0
s
n
(x)
This is easy; we just average the formulas
s
n
(x) =
1
2
_

f(x t)D
n
(t) dt
obtaining

N
(x) =
1
2
_

f(x t)F
N
(t) dt (4.1)
where
F
N
(t) =
1
N
N1

n=0
D
n
(t). (4.2)
The function F
N
(t) is called Fejers kernel. What makes Cesaro convergence easy
to work with is the remarkable fact that F
N
(t) 0 for all t R. In fact,
F
N
(t) =
sin
2
_
N
2
x
_
N sin
2
_
1
2
x
_ (4.3)
This may be seen by using the identity 2 sin(A) sin(B) = cos(AB) cos(A+B):
N1

n=0
sin
_
(n +
1
2
)t
_
sin
_
1
2
t
_
=
1
2
N1

n=0
_
cos
_
nt
_
cos
_
(n + 1)t
_
_
=
1
2
_
1 cos(Nt)
_
= sin
2
_
N
2
t
_
Identity (4.3) is now obtained by dividing by N sin
2
_
1
2
t
_
and using Formula (2.7).
MATH3068 LECTURE NOTES. HANDOUT 5. 19
Another basic fact about the Fejer kernel is that
1
2
_

F
N
(t) dt = 1. (4.4)
This holds because F
N
(t) is the average of the functions D
0
(t), . . . , D
N
(t), and
1
2
_

D
j
(t) dt = 1 for each j.
Theorem 4.2. Let f(x) be a piecewise continuous 2-periodic function, and as-
sume that f(x) is continuous at x
0
R. Then the Fourier series of f(x) at x
0
converges in the sense of Cesaro to f(x
0
).
Proof. As usual, let s
n
(x) denote the n-th partial sum of the Fourier series of f(x),
and let
N
(x) denote the average of s
0
(x), . . . , s
N1
(x). Using Formulas (4.1)
and (4.4), we have

N
(x
0
) f(x
0
) =
1
2
_

f(x
0
t)F
N
(t) dt f(x
0
)
=
1
2
_

_
f(x
0
t) f(x
0
)
_
F
N
(t) dt
Using the positivity of F
N
(t), we can write
|
N
(x
0
) f(x
0
)|
1
2
_

|f(x
0
t) f(x
0
)|F
N
(t) dt (4.5)
Now we use the continuity of f(x) at x
0
: given > 0, there is a > 0 such that
|f(x
0
t) f(x
0
)| < /2 if 0 |t| < . By reducing if necessary, we may assume
that < . We break the integral in (4.5) into two pieces: the integral over [, ],
and the integral over [, ] [, ]. Using (4.4) again, the rst integral is at most
1
2
_

2
F
N
(t) dt
1
2
_

2
F
N
(t) dt =

2
.
To deal with the second integral, remember that f(x), being piecewise continuous,
is assumed to be bounded. Thus there is a number M such that |f(x)| M for
all x R. Therefore, |f(x
0
t) f(x
0
)| 2M for all t R. Also, using (4.3) if
|t| we have
F
N
(t)
1
N sin
2
_
1
2
t
_
1
N sin
2
_
1
2

_.
Thus
1
2
_
|t|
|f(x
0
t) f(x
0
)|F
N
(t) dt
2M
N sin
2
_
1
2

_ =
K

N
, say.
Thus, once N > 2K

/
|
N
(x
0
) f(x
0
)|

2
+
K

N
< ,
so that
N
(x
0
) f(x
0
).
As we have mentioned, there are continuous 2-periodic functions whose Fourier
series fail to converge to f(x
0
) at some points x
0
. The next corollary shows that
the Fourier series cannot converge to any other limit:
Corollary 4.3. Suppose that the Fourier series of a piecewise continuous function
converges to some limit when x = x
0
. Suppose also that f(x) is continuous at x
0
.
Then must equal f(x
0
).
20 D. CARTWRIGHT
Proof. Let s
n
(x) be the n-th partial sum of the Fourier series of f(x), and let
N
(x)
be the average of s
0
(x), . . . , s
N1
(x). By hypothesis, s
n
(x
0
) . So Lemma 4.1
tells us that
N
(x
0
) as N . But Theorem 4.2 tells us that
N
(x
0
) f(x
0
).
Hence = f(x
0
).
Corollary 4.4. (The Uniqueness Theorem) Let f(x) and g(x) be two piecewise
continuous functions. Suppose that f(x) and g(x) have the same Fourier cosine
and sine coecients. Then on [, ], f(x) = g(x) for all but a nite number of x.
If f(x) and g(x) are actually continuous, then f(x) = g(x) for all x R.
Proof. Let h(x) = f(x) g(x). The hypotheses mean that h(x) is a piecewise
continuous function all of whose Fourier cosine coecients a
k
and sine coecients
b
k
are all zero. So the n-th partial sum of the Fourier series of h(x) is
s
n
(x) =
1
2
a
0
+
n

k=1
_
a
k
cos(kx) + b
k
sin(kx)
_
= 0
for all x R and all n 0. So the average
N
(x) of s
0
(x), . . . , s
N1
(x) is also 0:

N
(x) =
1
N
N

n=0
s
n
(x) = 0
for all x R and all N 1. But Theorem 4.2 tells us that
N
(x) h(x) as
N for each x R at which h is continuous. Thus h(x) = 0 for all x R at
which h(x) is continuous, so that f(x) = g(x) for any point x at which f(x) and
g(x) are both continuous.
5. Proof of Theorem 2.3
The goal of this section is to prove that if f(x) is any piecewise continuous
function, and s
n
is the n-th partial sum of the Fourier series of f(x), then
_

|f(x) s
n
(x)|
2
dx 0 as n
This theorem is of a quite dierent nature from that of our pointwise theorems
proved above. Instead of depending on the behaviour of f(x) near a particular
point, it is a global result.
One of the main steps in the proof is the following theorem (which gives more
justication for the denitions of the Fourier coecients a
k
and b
k
of f(x).
Theorem 5.1. Suppose that f(x) is a piecewise continuous function. Given any
integer n 1, and any trigonometric polynomial
t(x) =
c
0
2
+
n

k=1
_
c
k
cos(kx) + d
k
sin(kx)
_
(5.1)
of degree at most n, then the quantity
1
2
_

|f(x) t(x)|
2
dx
is minimized by taking c
k
= a
k
and d
k
= b
k
for all k (in which case t(x) is the n-th
partial sum s
n
(x) of the Fourier series of f(x)). In fact, for any t(x) as in (5.1),
1
2
_

|f(x) t(x)|
2
dx
=
1
2
_

|f(x) s
n
(x)|
2
dx +
|a
0
c
0
|
2
4
+
1
2
n

k=1
_
|a
k
c
k
|
2
+|b
k
d
k
|
2
),
(5.2)
MATH3068 LECTURE NOTES. HANDOUT 5. 21
so that
1
2
_

|f(x) t(x)|
2
dx >
1
2
_

|f(x) s
n
(x)|
2
dx
unless c
k
= a
k
and d
k
= b
k
for all k.
Proof. We write
f(x) t(x) = (f(x) s
n
(x)) + (s
n
(x) t(x))
= (f(x) s
n
(x)) +
a
0
c
0
2
+
n

k=1
_
(a
k
c
k
) cos(kx) + (b
k
d
k
) sin(kx)
_
.
We regard the sum on the right as a sum f
1
(x) + f
2
(x) + + f
2n+2
(x) of 2n + 2
functions, namely
f(x)s
n
(x),
a
0
c
0
2
, and the (a
k
c
k
) cos(kx) and (b
k
d
k
) sin(kx) for k = 1, . . . , n.
We are allowing the function f(x) and the coecients a
k
, . . . , d
k
to be complex. So
|f(x) t(x)|
2
= (f(x) t(x))(f(x) t(x))
equals
_
f
1
(x) + + f
2n+2
(x)
__
f
1
(x) + + f
2n+2
(x)
_
This equals
_
f
1
(x) + + f
2n+2
(x)
__
f
1
(x) + + f
2n+2
(x)
_
,
and when we multiply this out, it is the sum of (2n + 2)
2
terms
f
r
(x)f
s
(x), where r, s = 1, . . . , 2n + 2.
The key point is that
_

f
r
(x)f
s
(x) dx = 0 if r = s. (5.3)
Most of these equations are immediate from the orthogonality relations. For exam-
ple, if f
r
(x) = (a
k
c
k
) cos(kx) and f
s
(x) = (b

) sin(x), then
_

f
r
(x)f
s
(x) dx = (a
k
c
k
)(b

)
_

cos(kx) sin(x) dx = 0.
The trickiest of the equations (5.3) to check are those in which f
r
(x) = f(x)s
n
(x)
and f
s
(x) = (b

) sin(x), say. Then


_

f
r
(x)f
s
(x) dx = (b

)
_

(f(x) s
n
(x)) sin(x) dx.
So we have to show that
_

(f(x) s
n
(x)) sin(x) dx = 0.
This is because
_

f(x) sin(x) dx = b

, by denition of b

, and because the


integral
_

s
n
(x) sin(x) dx equals b

too (provided that n), again by the


orthogonality relations.
Because of (5.3), when we form the sum of all the integrals
_

f
r
(x)f
s
(x) dx,
we only need to consider the integrals with r = s. Hence
1
2
_

|f(x) t(x)|
2
dx =
2n+2

r=1
1
2
_

|f
r
(x)|
2
dx.
Since
1
2
_

|a
0
c
0
|
2
4
dx =
|a
0
c
0
|
2
4
22 D. CARTWRIGHT
and for k = 1, . . . , n,
1
2
_

|(a
k
c
k
) cos(kx)|
2
dx = |a
k
c
k
|
2
1
2
_

cos
2
(kx) dx
=
|a
k
c
k
|
2
2
and similarly
1
2
_

|(b
k
d
k
) cos(kx)|
2
dx = |b
k
d
k
|
2
1
2
_

cos
2
(kx) dx
=
|b
k
d
k
|
2
2
,
the result follows.
Corollary 5.2. If f(x) is a piecewise continuous function and n 1, then
1
2
_

|f(x)|
2
dx =
1
2
_

|f(x) s
n
(x)|
2
dx +
|a
0
|
2
4
+
1
2
n

k=1
_
|a
k
|
2
+|b
k
|
2
)
Proof. Simply take c
k
= 0 and d
k
= 0 for all k in (5.2).
Corollary 5.3. If f(x) is a piecewise continuous function, > 0, and n
0
1,
Suppose that t(x) is a trigonometric polynomial of degree n
0
, and that
1
2
_

|f(x) t(x)|
2
dx < .
Then, if s
n
(x) is as usual the n-th partial sum of the Fourier series of f(x), then
1
2
_

|f(x) s
n
(x)|
2
dx <
for all n n
0
.
Proof. If n n
0
, then t(x) is a trigonometric polynomial of degree at most n, and
so
1
2
_

|f(x) s
n
(x)|
2
dx
1
2
_

|f(x) t(x)|
2
dx
by Theorem 5.1.
So to prove Theorem 2.3, all we have to do is show that for each > 0, there is
a trigonometric polynomial t(x) such that
1
2
_

|f(x) t(x)|
2
dx < . (5.4)
We do this by rst nding a step function g(x) such that
1
2
_

|f(x)g(x)|
2
dx <

9
, then a 2-periodic function h(x) having a continuous second derivative such that
1
2
_

|g(x) h(x)|
2
dx <

9
, and nally a trigonometric polynomial t(x) such that
1
2
_

|h(x) t(x)|
2
dx <

9
. Then
|f(x) t(x)| = |(f(x) g(x)) + (g(x) h(x)) + (h(x) t(x))|
|f(x) g(x)| +|g(x) h(x)| +|h(x) t(x)|,
so that
|f(x) t(x)|
2
(|f(x) g(x)| +|g(x) h(x)| +|h(x) t(x)|)
2
3
_
|f(x) g(x)|
2
+|g(x) h(x)|
2
+|h(x) t(x)|
2
_
MATH3068 LECTURE NOTES. HANDOUT 5. 23
the last inequality holding because (a +b +c)
2
3(a
2
+b
2
+c
2
) for any a, b, c R.
Therefore
1
2
_

|f(x) t(x)|
2
dx
3
1
2
_

|f(x) g(x)|
2
dx + 3
1
2
_

|g(x) h(x)|
2
dx + 3
1
2
_

|h(x) t(x)|
2
dx
3
_

9
+

9
+

9
_
= .
Step (a). We nd the step function g(x) as follows. Recall that f(x), being
piecewise continuous, is bounded. That is, there is a number M > 0 so that
|f(x)| M for all x [, ]. By Lemma 3.2, there is a step function g(x) such
that |g(x)| M for all x [, ] too, and so that
1
2
_

|f(x) g(x)| dx <



18M
.
Since |f(x) g(x)| |f(x)| +|g(x)| 2M for all x, we have
1
2
_

|f(x) g(x)|
2
dx 2M
1
2
_

|f(x) g(x)| dx
and so
1
2
_

|f(x) g(x)|
2
dx <

9
.
Step (b). We nd the function h(x) having a continuous second derivative as
follows. Remember that the function g(x), being a step function, takes constant
values c
j
on intervals [a
j1
, a
j
], where = a
0
< a
1
< < a
n
= . All we have
to do is round out the function a little on each of these intervals. The graph
y = h(x), between a
j1
and a
j
is the curve indicated.
c
j
a
j1
a
j
Because h(x) = g(x) for most of the interval, the integral
1
2
_
aj
aj1
|g(x) h(x)|
2
dx
is small, and can be made as small as we please by making the rounded parts of h(x)
small enough. Hence
1
2
_

|g(x) h(x)|
2
dx,
which is the sum of the integrals over the intervals [a
j1
, a
j
], may be made as small
as we like, and so h(x) can be chosen so that this integral is less than /9.
Before we nd the the function t(x) of Step (c), we need a lemma which is useful
in other situations:
24 D. CARTWRIGHT
Lemma 5.4. Suppose that f(x) is a 2-periodic function which is dierentiable,
with f

(x) continuous. As usual, let a


k
and b
k
be the Fourier coecients of f(x),
and let a

k
and b

k
be the Fourier coecients of f

(x). Then a

0
= 0 and
a

k
= kb
k
and b

k
= ka
k
for k = 1, 2, . . . (5.5)
Proof. Firstly, by the Fundamental Theorem of Calculus, and since f(x) is periodic,
a

0
=
1

(x) dx =
1

_
f(2) f(0)
_
= 0.
Deriving the formulas (5.5) is just an exercise in integration by parts:
a

k
=
1

(x) cos(kx) dx =
1

_
f(x) cos(kx)
_

x=
x=

f(x)
_
k sin(kx)
_
dx
= 0 +
k

f(x) sin(kx) dx as f(x) is periodic


= kb
k
.
Similarly,
b

k
=
1

(x) sin(kx) dx =
1

_
f(x) sin(kx)
_

x=
x=

f(x)
_
k cos(kx)
_
dx
= 0
k

f(x) cos(kx) dx as f(x) is periodic


= ka
k
.

Corollary 5.5. Suppose that f(x) is a 2-periodic function which is twice dier-
entiable, with f

(x) continuous. As usual, let a


k
and b
k
be the Fourier coecients
of f(x). Then there is a number M, which does not depend on k, so that
|a
k
|
M
k
2
and |b
k
|
M
k
2
for all k 1. (5.6)
Proof. Let a

k
and b

k
be the Fourier coecients of f

(x) and let a

k
and b

k
be
the Fourier coecients of f

(x). Then using (5.5) twice, once on f(x) and once


on f

(x), we get
a

k
= kb

k
= k
2
a
k
and b

k
= ka

k
= k
2
a
k
.
Hence
a
k
=
1
k
2
a

k
=
1
k
2
2
_

(x) cos(kx) dx,


so that
|a
k
| =
1
k
2
|a

k
|
1
k
2
2
_

|f

(x) cos(kx)| dx
M
k
2
for M =
1
2
_

|f

(x)| dx. Similarly, |b


k
|
M
k
2
.
Corollary 5.6. Suppose that f(x) is a 2-periodic function which is twice dieren-
tiable, with f

(x) continuous. Let s


n
(x) denote the n-th partial sum of the Fourier
series of f(x). Then
|f(x) s
n
(x)|
2M
n
for all x R, (5.7)
where M is the number in the previous corollary.
MATH3068 LECTURE NOTES. HANDOUT 5. 25
Proof. Since f(x) is dierentiable everywhere, we know that its Fourier series con-
verges to f(x) at every point x. Hence
f(x) s
n
(x) =

k=n+1
_
a
k
cos(kx) + b
k
sin(kx)
_
.
Therefore
|f(x)s
n
(x)|

k=n+1

a
k
cos(kx)+b
k
sin(kx)

k=n+1
_
|a
k
|+|b
k
|
_
2M

k=n+1
1
k
2
,
where M is as in (5.6). Now we use the inequality

k=n+1
1
k
2

_

n
1
x
2
dx =
1
n
,
which is clear from the following diagram:
x n n+1 n+2
y
y =
1
x
2
This completes the proof.
Completion of Step (c). Suppose that h(x) is a 2-periodic function which
is twice dierentiable, with h

(x) continuous. Let t


n
(x) denote the n-th partial
sum of the Fourier series of h(x). Applying Corollary 5.6 to the function h(x),
we have |h(x) t
n
(x)| 2M/n for all x R, and for n = 1, 2, . . ., where M =
1
2
_

|h

(x)| dx, and so


1
2
_

|h(x) t
n
(x)|
2
dx
4M
n
2
.
So we can take t(x) = t
n
(x), where n is chosen large enough so that 4M/n
2
< /9.
This completes the last step in the proof that there is a trigonometric polyno-
mial t(x) so that (5.4) holds, and hence it completes the proof of Theorem 2.3.

Вам также может понравиться