Syllabus Fourier analysis

by T. H. Koornwinder, 1996
University of Amsterdam, Faculty of Science, Korteweg-de Vries Institute
Last modied: 7 December 2005
Note This syllabus is based on parts of the book Fouriertheorie by A. van Rooij (Epsilon
Uitgaven, 1988). Many of the exercises and some parts of the text are quite literally taken
from this book. Usage of this book in addition to the syllabus is recommended. The present
version of the syllabus is slightly modied. Modications were made by J. Wiegerinck and
by T. H. Koornwinder.
Contents
Part I. Fourier series
1. L
2
theory.
2. L
1
theory
3. The Dirichlet kernel
4. The Fejer kernel
5. Some applications of Fourier series
Part II. Fourier integrals
6. Generalities
7. Inversion formula
8. L
2
theory
9. Poisson summation formula
10. Some applications of Fourier integrals
References
[1] A. van Rooij, Fouriertheorie, Epsilon Uitgaven, 1988.
[2] A. A. Balkema, Syllabus Integratietheorie, UvA, KdVI, last modied 2003.
[3] J. Wiegerinck, Syllabus Functionaalanalyse, UvA, KdVI, last modied 2005.
[4] W. Rudin, Real and complex analysis, McGraw-Hill, Second edition, 1974.
[5] H. Dym & H. P. McKean, Fourier series and integrals, Academic Press, 1972.
[6] Y. Katznelson, An introduction to harmonic analysis, Dover, Second edition, 1976.
[7] T. W. Korner, Fourier analysis, Cambridge University Press, 1988.
[8] T. W. Korner, Exercises for Fourier analysis, Cambridge University Press, 1993.
[9] W. Schempp & B. Dreseler, Einf uhrung in die harmonische Analyse, Teubner, 1980.
[10] E. Stein & R. Shakarchi, Fourier analysis, an introduction, Princeton University Press,
2003.
[11] E. M. Stein & G. Weiss, Introduction to Fourier analysis on Euclidean spaces, Prince-
ton University Press, 1971.
[12] G. Weiss, Harmonic analysis, in Studies in real and complex analysis, I. I. Hirschman
(ed.), MAA Studies in Math. 3, Prentice-Hall, 1965, pp. 124178.
[13] A. Zygmund, Trigonometric series, Cambridge University Press, Second ed., 1959.
2
1 L
2
theory
1.1 The Hilbert space L
2
2
1.1 T-periodic functions. Let T > 0. A function f: R C is called periodic with period
T (or T-periodic) if f(t + T) = f(t) for all t in R. A T-periodic function is completely
determined by its restriction to some interval [a, a+T), and any function on [a, a+T) can
be uniquely extended to a T-periodic function on R. However, a function f on [a, a+T] is
the restriction of a T-periodic function i f(a+T) = f(a). When working with T-periodic
functions we will usually take T := 2.
1.2 The Banach spaces L
p
2
. Let 1 p < . We denote by L
p
2
the space of all 2-
periodic (Lebesgue) measurable functions f: R C such that
_

[f(t)[
p
dt < . This
is a complex linear space. In particular, we are interested in the cases p = 1 and p = 2.
The reader may continue reading with these two cases in mind. A seminorm | . |
p
can be
dened on L
p
2
by
|f|
p
:=
_
1
2
_

[f(t)[
p
dt
_
1/p
(f L
p
2
). (1.1)
By seminorm we mean that |f|
p
0, |f + g|
p
|f|
p
+ |g|
p
and |f|
p
= [[ |f|
p
for
f, g L
p
2
, C, but not necessarily |f|
p
> 0 if f ,= 0. The factor (2)
1
is included in
(1.1) just for cosmetic reasons: if f is identically 1 then |f|
p
= 1.
It is known from integration theory (see syll. Integratietheorie) that, for f L
p
2
, we
have that |f|
p
= 0 i f = 0 a.e. (almost everywhere), i.e., i f(t) = 0 for t outside some
subset of R of (Lebesgue) measure 0. It follows that, for f L
p
2
, the value of |f|
p
does
not change if we modify f on a set of measure 0, so | . |
p
is well-dened on equivalence
classes of functions, where equivalence of two functions means that they are equal almost
everywhere. Also, |f|
p
= 0 i f is equivalent to the function which is identically 0 on R.
Thus we dene the space L
p
2
as the set of equivalence classes of a.e. equal functions
in L
p
2
. It can also be viewed as the quotient space L
p
2
/f L
p
2
[ |f|
p
= 0. The space
L
p
2
becomes a normed vector space with norm | . |
p
. In fact this normed vector space is
complete (see syll. Integratietheorie, Chapter 6 for p = 1 or 2), so it is a Banach space.
If I is some interval then we can also consider the space L
p
(I) of measurable functions
on I for which
_
I
[f(t)[
p
dt < , and the space L
p
(I) of equivalence classes of a.e. equal
functions in L
p
(I). A seminorm or norm | . |
p
on L
p
(I) or L
p
(I), respectively, is dened
by
|f|
p
:=
__
I
[f(t)[
p
dt
_
1/p
. (1.2)
The linear map f f[
[,)
: L
p
2
L
p
([, )) is bijective and it preserves | . |
p
up to
a constant factor. It naturally yields an isomorphism of Banach spaces (up to a constant
factor) between L
p
2
and L
p
([, ]). Note that, when we work with L
p
2
rather than L
p
2
,
the function values on the endpoints of the interval [, ] do not matter, since the two
endpoints form a set of measure zero.
L
2
THEORY 3
Ex. 1.3 Show the following. If f L
1
2
then
_

f(t) dt =
_
+a
+a
f(t) dt (a R).
1.4 The Banach space (
2
. Dene the linear space (
2
as the space of all 2-periodic
continuous functions f: R C. The restriction map f f[
[,]
identies the linear space
(
2
with the linear space of all continuous functions f on [, ] for which f() = f().
The space (
2
becomes a Banach space with respect to the sup norm
|f|

:= sup
tR
[f(t)[ = sup
t[,]
[f(t)[ (f (
2
).
We can consider the space C([, ]) of continuous functions on [, ] as a linear
subspace of L
p
([, ]), but also as a linear subspace of L
p
([, ]).
Indeed, if two continuous functions on [, ] are equal a.e. then they are equal everywhere.
Hence each equivalence class in L
p
([, ]) contains at most one continuous function.
Hence, if f, g C([, ]) are equal as elements of L
p
([, ]) then they are equal as
elements of C([, ]).
It follows that we can consider the space (
2
as a linear subspace of the space L
p
2
, but
also as a linear subspace of L
p
2
.
The following proposition is well known (see Rudin, Theorem 3.14):
1.5 Proposition C([, ]) is a dense linear subspace of the Banach space L
p
([, ]).
Ex. 1.6 For p 1 prove the norm inequality
|f|
p
|f|

(f (
2
).
Ex. 1.7 Prove that the linear space (
2
is a dense linear subspace of the Banach space
L
p
2
(p 1).
Hint Show for f C([, ]) and > 0 that there exists g C([, ]) such that
g() = 0 = g() and |f g|
p
< .
Ex. 1.8 Prove the following:
Let V be a dense linear subspace of (
2
with respect to the norm | . |

. Then, for p 1,
V is a dense linear subspace of L
p
2
with respect to the norm | . |
p
.
1.9 For f a function on R and for a R dene the function T
a
f by
(T
a
f)(x) := f(x +a) (x R). (1.3)
If f is 2-periodic then so is T
a
f and, if V is any of the spaces L
p
2
or (
2
then T
a
: V V
is a linear bijection which preserves the appropriate norm.
4 CHAPTER 1
Proposition Let p 1, f L
p
2
. Then the map a T
a
f: R L
p
2
is uniformly
continuous.
Proof First take g (
2
. Then, by the compactness of [, ] and by periodicity,
g is uniformly continuous on R. Hence, for each > 0 there exists > 0 such that
|T
a
g T
b
g|

< if [a b[ < . Next take f L

p
2
and let > 0. Since (
2
is dense in
L
p
2
(see Exercise 1.7), we can nd g (
2
such that |f g|
p

1
3
. Hence
|T
a
f T
b
f|
p
|T
a
(f g)|
p
+|T
a
g T
b
g|
p
+|T
b
(f g)|
p

2
3
+|T
a
g T
b
g|

,
where we used Exercise 1.6. Now we can nd > 0 such that |T
a
g T
b
g|

<
1
3
if
[a b[ < .
1.10 The Hilbert space L
2
2
. We can say more about L
p
2
if p = 2. Then, for any two
functions f, g L
2
2
, we can dene f, g) C by
f, g) :=
1
2
_

f(t) g(t) dt. (1.4)

Note that f, f) = (|f|
2
)
2
. The form . , . ) has all the properties of a hermitian
inner product on the complex linear space L
2
2
, except that it is not positive denite, only
positive semidenite: we may have f, f) = 0 while f is not identically zero. However, the
right hand side of (1.4) does not change when we modify f and g on subsets of measure
zero. Therefore, f, g) is well-dened on L
2
2
and it is a hermitian inner product there. In
fact, the inner product space L
2
2
is complete: it is a Hilbert space.
For I an interval and f, g in L
2
(I) or L
2
(I) we can dene f, g) in a similar way:
f, g) :=
_
I
f(t) g(t) dt. (1.5)
The bijective linear map f f[
[,)
: L
2
2
L
2
([, )) preserves . , . ) up to a
constant factor. It naturally yields an isomorphism of Hilbert spaces (up to a constant
factor) between L
2
2
and L
2
([, ]).
The space L
2
2
is a linear subspace of L
1
2
. In fact, for f L
2
2
we have the norm
inequality
|f|
1
|f|
2
. (1.6)
(Prove this by use of the Cauchy-Schwarz inequality.) Also, L
2
2
is a dense linear subspace
of L
1
2
. Indeed, the subspace (
2
of L
2
2
is already dense in L
1
2
.
1.11 An orthonormal basis of L
2
2
. The functions t e
int
(n Z) belong to (
2
and
they form an orthonormal system in L
2
2
:
1
2
_

e
imt
e
int
dt =
m,n
(m, n Z). (1.7)
By a trigonometric polynomial on R (of period 2) we mean a nite linear combination
(with complex coecients) of functions t e
int
(n Z).
The following theorem has been mentioned without proof in syll. Functionaalanalyse.
L
2
THEORY 5
Theorem
(a) The space of trigonometric polynomials is dense in (
2
with respect to the norm | . |

.
(b) The space of trigonometric polynomials is dense in L
2
2
with respect to the norm | . |
2
.
(c) The functions t e
int
(n Z) form an orthonormal basis of L
2
2
.
Part (b) of the Theorem follows from part (a) (why?), and part (c) follows from part
(b). We will prove the theorem in a later chapter (rst part (b) and hence part (c), and
afterwards part (a)). However, in this Chapter we will already use the Theorem. So we
have to be careful later that circular arguments will be avoided.
1.2 Generalities about orthonormal bases
1.12 We now recapitulate some generalities concerning orthonormal bases of Hilbert
spaces, as given in syll. Functionaalanalyse. Let H be a Hilbert space. Denote the inner
product by . , . ) and the norm by | . |. Let / be an index set and consider an
orthonormal system E := e

A
in H, i.e., e

, e

) =
,
for , /. For convenience
we assume that the Hilbert space is separable. Thus the index set / of the orthonormal
system E will be countable.
Proposition (Bessel inequality)

A
[f, e

)[
2
|f|
2
(f H).
Corollary If f H then lim

f, e

) = 0, i.e., for all > 0 there is a nite subset

B / such that, if /B then [f, e

)[ < .
(The set /, equipped with the discrete topology, is locally compact. By adding the point
, we obtain the one-point compactication of /. This gives a further explanation of the
notion lim

.)
Denition-Theorem (orthonormal basis; Parsevals norm equality)
The orthonormal system E is called an orthonormal basis of H if the following equivalent
properties hold:
(a) Span(E) is dense in H.
(b)

A
[f, e

)[
2
= |f|
2
for all f H (Parseval equality).
(c) If f H and f is orthogonal to E then f = 0 (i.e., E is a maximal orthonormal
system).
1.13 Denition-Proposition (unconditional convergence)
Let Hbe a separable Hilbert space. Let / be a countably innite index set. Let v

A

H. We say that the sum

A
v

unconditionally converges to some v H if the following

two equivalent properties are valid:
(a) For each way of ordering /as a sequence
1
,
2
, . . . we have that v = lim
N

N
n=1
v

n
in the topology of H.
(b) For each > 0 there is a nite subset B / such that for each nite set ( satisfying
B ( / we have that |v

C
v

| < .
6 CHAPTER 1
Theorem Let H be a separable Hilbert space with orthonormal basis e

A
(/ a
countable index set). Let c

A

2
(/) (i.e.

A
[c

[
2
< ). Then there is a unique
f H such that f, e

) = c

( /). This element f can be written as f =

A
c

(with unconditional convergence).

Proposition (Parsevals inner product equality)
Let H be a separable Hilbert space with orthonormal basis e

A
(/ a countable index
set). Then
f, g) =

A
f, e

) g, e

) (f, g H)
with absolute convergence.
1.14 Theorem (summarizing this subchapter) Let H be a separable Hilbert space
with orthonormal basis e

A
(/ a countable index set). Then there is an isometry of
Hilbert spaces
T: f c

A
: H
2
(/)
dened by c

:= f, e

), with inverse isometry

T
1
: c

A
f:
2
(/) H
given by f :=

A
c

(unconditionally).
1.3 Application to an orthonormal basis of L
2
2
By Theorem 1.11 the functions t e
int
(n Z) form an orthonormal basis of L
2
2
. Let us
apply the general results of the previous subsection to this particular orthonormal basis of
the Hilbert space L
2
2
.
1.15 Denition Let f L
1
2
. The Fourier coecients of f are given by the numbers

f(n) :=
1
2
_

f(t) e
int
dt (n Z). (1.8)
Note that the integral is absolutely convergent since [f(t) e
int
[ [f(t)[.
We can consider

f as a function

f: Z C. The map T: f

f, sending 2-periodic
functions on R to functions on Z, is called the Fourier transform (for periodic functions).
Thus, by the Fourier transform of a function f L
1
2
we mean the function

f.
Remark In this subchapter we will consider the Fourier transform f

f restricted to
functions f L
2
2
. For such f we can consider the right hand side of (1.8) as an inner
product. Indeed, if we put e
n
(t) := e
int
(n Z) then

f(n) = f, e
n
) (f L
2
2
).
Proposition (Bessel inequality)

n=
[

f(n)[
2

1
2
_

[f(t)[
2
dt (f L
2
2
).
Corollary (Riemann-Lebesgue Lemma) If f L
2
2
then lim
|n|

f(n) = 0.
L
2
THEORY 7
1.16 The results of 1.15 did not depend on the completeness of the orthonormal system.
However, the proof of the next results uses this completeness, so we cannot be sure about
these results until we have proved Theorem 1.11(b).
Theorem (Parseval) Let f, g L
2
2
. Then
1
2
_

[f(t)[
2
dt =

n=
[

f(n)[
2
, (1.9)
1
2
_

f(t) g(t) dt =

n=

f(n) g(n). (1.10)

In (1.9) and (1.10), the integral on the left hand side and the sum on the right hand side
are absolutely convergent.
1.17 Theorem (Riesz-Fischer) Let
n

nZ

2
(Z), i.e.

n=
[
n
[
2
< . Then
there is a unique f L
2
2
such that

f(n) =
n
(n Z), and this f can be written as
f(x) =

n=

n
e
inx
(unconditional convergence in L
2
2
).
So, in particular,
f(x) = lim
N
N

n=N

n
e
inx
(convergence in L
2
2
),
i.e.,
lim
N
_

f(t)
N

n=N

n
e
int

2
dt = 0.
1.18 (summarizing this subchapter) The map T: f
n

nZ
: L
2
2

2
(Z) given by

n
:=
1
2
_

f(t) e
int
dt
denes an isometry of Hilbert spaces with inverse isometry T
1
:
n

nZ
f:
2
(Z) L
2
2
given by
f(x) :=

n=

n
e
inx
(unconditional convergence in L
2
2
).
1.19 Without proof we mention a more recent, very important and deep result: the
almost everywhere convergence of Fourier series of any function in L
2
2
.
Theorem (L. Carleson, Acta Mathematica 116 (1966), 135157) If f L
2
2
then
f(x) = lim
N
N

n=N

f(n) e
inx
with pointwise convergence for almost all x R.
8 CHAPTER 1
1.4 Orthonormal bases with cosines and sines
The orthonormal basis of functions t e
int
(n Z) for L
2
2
(see Theorem 1.11(c))
immediately gives rise to an orthonormal basis for L
2
2
in terms of cosines and sines. We
formulate it in two steps and leave the straightforward proofs to the reader.
1.20 Lemma For n = 1, 2, . . . the two-dimensional subspace of L
2
2
spanned by the
two orthonormal functions t e
int
also has an orthonormal basis given by the two
functions t

2 cos(nt) and t

2 sin(nt).
Proposition The functions t 1, t

2 cos(nt) (n = 1, 2, . . .), and t

2 sin(nt)
(n = 1, 2, . . .) form an orthonormal basis for L
2
2
.
Note that, in the orthonormal basis of the above Proposition, the cosine functions
(including 1) are even and the sine functions are odd. In fact, we can split the Hilbert
space L
2
2
as a direct sum of the subspace of even functions and the subspace of odd
functions, such that the cosines form an orthonormal basis for the rst subspace and the
sines an orthonormal subspace for the second subspace. Let us rst discuss the notion of
direct sum decomposition.
1.21 Denition Let H be a Hilbert space and let H
1
and H
2
be closed linear subspaces
of H (so H
1
and H
2
are Hilbert spaces themselves). We say that H is the (orthogonal)
direct sum of H
1
and H
2
(notation H = H
1
H
2
) if the two following conditions are
satised:
(i) The subspaces H
1
and H
2
are orthogonal to each other.
(ii) Each v H can be written as v = v
1
+v
2
with v
1
H
1
and v
2
H
2
.
Ex. 1.22 Let H
1
and H
2
be Hilbert spaces and make H := (v
1
, v
2
) [ v
1
H
1
, v
2
H
2

into an inner product space by the rules

(v
1
, v
2
) + (w
1
, w
2
) := (v
1
+w
1
, v
2
+w
2
), (v
1
, v
2
) := (v
1
, v
2
),
(v
1
, v
2
), (w
1
, w
2
)) := v
1
, w
1
) +v
2
, w
2
).
Show that H is a Hilbert space and that it is the direct sum of its two closed linear
subspaces (v, 0) [ v H
1
and (0, v) [ v H
2
.
An example of a direct space decomposition is given by the following Proposition.
1.23 Proposition The space L
2
2
is the orthogonal direct sum of the two closed linear
subspaces
L
2
2,even
:= f L
2
2
[ f(t) = f(t) a.e.,
L
2
2,odd
:= f L
2
2
[ f(t) = f(t) a.e..
The maps f f[
[0,]
: L
2
2,even
L
2
([0, ]) and f f[
[0,]
: L
2
2,odd
L
2
([0, ]) are
isomorphisms of Hilbert spaces, provided the inner product on L
2
([0, ]) is normalized as
f, g) :=
1

_

0
f(t) g(t) dt. (1.11)
Theorem The functions t 1 and t

2 cos(nt) (n = 1, 2, . . .) form an orthonormal

basis of L
2
2,even
, and hence also of L
2
([0, ]) (with inner product (1.11)). The functions
t

2 sin(nt) (n = 1, 2, . . .) form an orthonormal basis of L

2
2,odd
, and hence also of
L
2
([0, ]) (with inner product (1.11)).
L
2
THEORY 9
Ex. 1.24 Give the proofs of the above Proposition and Theorem.
1.5 Further exercises
Ex. 1.25 Let f be the sawtooth function, i.e. the 2-periodic function which is given on
(0, 2) by:
f(x) := x (0 < x < 2), (1.12)
and which is arbitarary for x = 0. Show the following:
(a) f L
2
2
and

f(n) = (in)
1
if n ,= 0 and

f(0) = 0.
(b)

n=1
n
2
=
2
/6.
Hint Apply formula (1.9).
Ex. 1.26 Show that

n=1
n
4
=
4
/90
by applying formula (1.9) to the function f L
2
2
such that f(t) := t
2
for t (, ).
Ex. 1.27 For 0 ,= a R let f
a
L
2
2
such that f
a
(t) := e
at
for t (, ).
(a) Compute the Fourier coecients of f
a
.
(b) Let a, b R such that a, b, a +b ,= 0. Derive the identity
coth(a) + coth(b) =
1

n=
_
1
a in
+
1
b +in
_
by applying (1.10) to the functions f := f
a
and g := f
b
.
(c) Show that the case a = b of this identity yields
coth(a) =
1
lim
N
N

n=N
1
a in
.
Does the right hand side converge? Is it allowed to replace lim
N

N
n=N
by

n=
? Conclude that

n=1
1
a
2
+n
2
=
coth(a)
4a

1
4a
2
(a ,= 0).
This becomes the identity in Exercise 1.25(b) as a 0.
Ex. 1.28 Put c
0
(t) := 1, c
n
(t) :=

2 cos(nt) (n = 1, 2, . . .), s
n
(t) :=

2 sin(nt) (n =
1, 2, . . .), and let
c
m
, s
n
) :=
1

_

0
c
m
(t) s
n
(t) dt (m = 0, 1, 2, . . . , n = 1, 2, . . .)
(inner products in L
2
([0, ])). Consider the matrix with (real) matrix entries
c
m
, s
n
), (row indices m and column indices n). Show that the columns of this matrix are
orthonormal, and that also the rows are orthonormal, i.e.,

m=0
c
m
, s
k
)c
m
, s
l
) =
k,l
,

n=1
c
k
, s
n
)c
l
, s
n
) =
k,l
.
Next compute the matrix entries explicitly.
10 CHAPTER 1
Ex. 1.29 Let f L
2
2
. For k = 1, 2, . . . put g
k
(t) := f(kt). Then g
k
L
2
2
. Express the
Fourier coecients of g
k
in terms of the Fourier coecients of f.
Ex. 1.30 Let , 1. Dene
L
2
2,,
:= f L
2
2
[ f(x) = f(x) a.e., f( x) = f(x) a.e..
a) Show that the Hilbert space L
2
2
is the direct sum of the four mutually orthogonal
closed linear subspaces L
2
2,,
(, 1).
b) For each choice of , nd an orthonormal basis of L
2
2,,
. (Start with the orthonor-
mal basis for L
2
2,even
or L
2
2,odd
given in 1.23.)
c) For each , give a Hilbert space isomorphism of L
2
2,,
with L
2
([0,
1
2
]).
Ex. 1.31 For F a function on R dene a function f on (0, 2) by f(x) := F(cot
1
2
x).
This establishes a one-to-one linear correspondence between functions f on (0, 2) and
functions F on R
(a) Show that the map f F is a Hilbert space isomorphism of L
2
((0, 2); (2)
1
dx)
onto L
2
(R;
1
(t
2
+ 1)
1
dt)., i.e.,
1
2
_

[f(x)[
2
dx =
1

[F(t)[
2
dt
t
2
+ 1
.
(b) Show that the orthonormal basis of L
2
((0, 2); (2)
1
dx) provided by the functions
x e
inx
(n Z), is sent by the map f F to the orthonormal basis of L
2
(R;
1
(t
2
+
1)
1
dt) given by the functions t (
t+i
ti
)
n
.
Ex. 1.32 Let V be the linear space of piecewise linear continuous functions on the closed
bounded interval [a, b]. So, f V i f C([a, b]) and there is a partition a = a
0
< a
1
<
a
2
. . . < a
n
= b such that the restriction of f to [a
i1
, a
i
] is linear for i = 1, . . . , n. Let
V
0
:= f V [ f(a) = f(b) = 0. Show that V is dense in C([a, b]) and that V
0
is dense
in L
p
([a, b]) (1 p < ).
11
2 L
1
theory
2.1 Growth rates of Fourier coecients
2.1 The Fourier coecients

f(n) (n Z) of a function f L
1
2
were dened in formula
(1.8). It follows from this formula that
[

f(n)[ |f|
1
(f L
1
2
, n Z). (2.1)
Hence
|

f|

|f|
1
where |

f|

:= sup
nZ
[

f(n)[.
So

f

(Z) if f L
1
2
. Here

(Z) := (
n
)
nZ
[ sup
nZ
[
n
[ < , a Banach space.
Ex. 2.2 Show that the map f

f: L
1
2

(Z) is a bounded linear operator. Also

determine its operator norm.
2.3 In 1.15 we gave the so-called Riemann-Lebesgue Lemma in the L
2
-case. It remains
valid for the L
1
-case:
Theorem (Riemann-Lebesgue Lemma) If f L
1
2
then lim
|n|

f(n) = 0.
Proof Let f L
1
2
. Let > 0. We have to look for a natural number N such that
[

f(n)[ < if [n[ N. Since L

2
2
is dense in L
1
2
(see 1.10), there exists a function g L
2
2
such that |f g|
1
<
1
2
. By the Riemann-Lebesgue Lemma for L
2
(see 1.15) there exists
N N such that: [n[ N [ g(n)[ <
1
2
. Hence, if [n[ N then
[

f(n)[ [

f(n) g(n)[ +[ g(n)[ |f g|

1
+[ g(n)[ <
1
2
+
1
2
= ,
where we used the inequality (2.1).
Ex. 2.4 Dene the linear space c
0
(Z) by
c
0
(Z) := c
n

nZ

(Z) [ lim
|n|
c
n
= 0.
Show that c
0
(Z) is a closed linear subspace of

(Z). So, in particular, c

0
(Z) becomes a
Banach space. Note also that f

f: L
1
2
c
0
(Z) is a bounded linear map.
2.5 In addition to the space (
2
we will deal, for k = 1, 2, . . . or , with the linear
space (
k
2
consisting of 2-periodic C
k
-functions on R. By (
0
2
we will just mean (
2
.
Elementary integration by parts yields:
(f

)(n) = in

f(n) (f (
1
2
, n Z). (2.2)
This is an important result. Corresponding to the dierentiation operator f f

acting
on 2-periodic functions we have on the Fourier transform side the operator multiplying
a function of n by in. Iteration of (2.2) yields:
(f
(k)
)(n) = (in)
k

f(n) (f (
k
2
, k = 0, 1, 2, . . . , n Z). (2.3)
For f (
2
the Riemann-Lebesgue Lemma gives that

f(n) = o(1) as [n[ , a
modest rate of decline for the Fourier coecients. It turns out that a higher order of
dierentiability of a 2-periodic function f implies a faster decline of

f(n) as [n[ :
12 CHAPTER 2
Theorem (a) Let k 0, 1, 2, . . . . If f (
k
2
then

f(n) = o([n[
k
) as [n[ .
(b) If f (

2
then

f(n) = O([n[
k
) as [n[ for all k 0, 1, 2, . . . .
So we see that for 2-periodic C

-functions the Fourier coecients decrease faster

in absolutele value to zero than any inverse power of [n[. We then say that the Fourier
coecients are rapidly decreasing.
Proof of Theorem Part (b) follows immediately from part (a). For the proof of part
(a) we use equations (2.3) and the Riemann-Lebesgue Lemma:
[

f(n)[ = [n[
k
[(f
(k)
)(n)[ = [n[
k
o(1) as n ,
since f
(k)
(
2
if f (
k
2
.
Ex. 2.6 Show that (2.2) still holds if f (
2
with piecewise continuous derivative.
Ex. 2.7 Let f (
2
. Suppose that f can be extended to a function analytic on the open
strip z C [ [Imz[ < K and continuous on the closed strip z C [ [Imz[ K. Show
that [

f(n)[ = O(e
K|n|
) as [n[ .
Hint If z C and [Imz[ K then f(z + 2) = f(z). Next show by contour integration
that

f(n) =
1
2
_
a+
a
f(z) e
inz
dz (n Z, [Ima[ K).
2.8 In the previous sections we derived the behaviour of the Fourier coecients

f(n) from
the behaviour of the function f. Now we consider the inverse problem: Let coecients
n
with certain behaviour be given. Find a 2-periodic function f such that

f(n) =
n
and
give the behaviour of f.
We say that the doubly innite sequence (
n
)
nZ
is in
1
(Z) if
|(
n
)|
1
:=

n=
[
n
[ < .
Theorem Let (
n
)
1
(Z). Put
f(x) :=

n=

n
e
inx
(x R), (2.4)
well dened because the series converges absolutely. Then f (
2
and

f(n) =
n
(n Z).
Also |f|

|(
n
)|
1
.
Proof The absolute convergence of the series on the right hand side of (2.4) is uniform
for x R because of the Weierstrass test, since [
n
e
inx
[ [
n
[ and

n=
[
n
[ < .
The sum of a uniform convergent series of continuous functions is continuous. Since the
terms of the series are 2-periodic in x, the same holds for the sum function. Thus f (
2
.
The inequality |f|

|(
n
)|
1
follows by taking absolute values on both sides of (2.4),
and by dominating the absolute value of the sum on the right hand side by the sum of the
absolute values of the terms. Finally, we derive that

f(n) =
1
2
_

m=

m
e
imx
e
inx
_
dx =

m=

m
_
1
2
_

e
i(mn)x
dx
_
=
n
,
where the second equality is permitted because the integral over a bounded interval of a
uniformly convergent sum of continuous functions equals the sum of the integrals of the
terms.
L
1
THEORY 13
Ex. 2.9 Let f(x) be given by (2.4). Prove the following statements. Use for (a) and (b)
the theorem about dierentiation of a series of functions on a bounded interval for which
the series of derivatives is uniformly convergent.
(a) If

n=
[n[ [
n
[ < then f (
1
2
.
(b) Let k 0, 1, 2, . . .. If

n=
[n[
k
[
n
[ < then f (
k
2
.
(c) Let > 1. If
n
= O([n[

) as [n[ then f (
k
2
for all integer k such that
0 k < 1.
(d) If
n
= O([n[
k
) as [n[ for all k 0, 1, 2, . . . then f (

2
.
(e) Let K > 0. If
n
= O(e
K|n|
) as [n[ then f can be extended to an analytic
function on the strip z C [ [Imz[ < K. Compare with Exercise 2.7
2.10 Remark Observe that the Fourier images of L
2
2
and of (

2
can be completely
characterized. Namely, for a given sequence (
n
)
nZ
we have:

n
=

f(n) (n Z) for some f L
2
2
i

n=
[
n
[
2
< ;

n
=

f(n) (n Z) for some f (

2
i
n
= O([n[
k
) as [n[ for all k 0, 1, 2, . . . .
However, such a characterization is not possible for the Fourier images of (
2
, (
1
2
, (
2
2
, . . .
and of L
1
2
.
Ex. 2.11 Let
0
:= 0 and
n
:= [n[

for n Z0. Then it is known in the literature

for which real values of there exists f L
1
2
such that

f(n) =
n
(n Z). For some
values of this follows from a non-trivial theorem (see sections 7.17 and 7.19 in van Rooij).
However, for some other values of the answer is immediate. Give the answer in these
cases.
2.2 Fubinis Theorem
2.12 For the treatment of convolution we will need Fubinis Theorem. The Proposition
below formulates Fubinis theorem for nonnegative measurable functions, see also syll.
Integratietheorie, Sections 5.65.8. Next the Theorem below will deal with complex-valued
measurable functions. Below we will work with -nite measure spaces (X, /, ) and
(Y, B, ) and their product space (X Y, /B, ), again a -nite measure space.
It is helpful to read rst the Theorem below for the case that X and Y are bounded
intervals, and are Lebesgue measure, and the function h: X Y C is continuous.
Then all integrals can also be considered as Riemann integrals and the assumption in the
Theorem concerning the absolute convergence of a repeated integral is automatic.
The general Proposition and Theorem below are much more delicate, because measure
spaces may be -nite rather than nite and because measurable rather than continuous
functions are considered.
2.13 Proposition Let the function h: X Y [0, ] be measurable with respect to
the -algebra /B. Then:
(a) The function y
_
X
h(x, y) d(x) is measurable on B;
(b) The function x
_
Y
h(x, y) d(y) is measurable on /;
14 CHAPTER 2
(c) We have
_
XY
h(x, y) d()(x, y) =
_
X
__
Y
h(x, y) d(y)
_
d(x) =
_
Y
__
X
h(x, y) d(x)
_
d(y).
(2.5)
Note that all integrals considered in (2.5) are integrals of measurable functions which
take values on [0, ]. The integral of such a function is well-dened and it will yield a
number in [0, ].
Consider next a function h: X Y C (so not necessarily non-negative) which is
measurable with respect to the -algebra / B. Then the nonnegative function [h[ is
measurable with respect to /B, so the above Proposition will hold with h(x, y) replaced
by [h(x, y)[. This we will need in the next theorem.
2.14 Theorem (see Rudin, Theorem 7.8) Let the function h: XY C be measur-
able with respect to the -algebra /B. Suppose that one of the two inequalities below
is valid.
_
Y
__
X
[h(x, y)[ d(x)
_
d(y) < . (2.6)
_
X
__
Y
[h(x, y)[ d(y)
_
d(x) < . (2.7)
(So the other inequality is also valid by the above Proposition.) Then:
(a)
_
X
[h(x, y)[ d(x) < for y outside some set Y
0
Y of -measure 0 and the function
y
_
X
h(x, y) d(x): Y Y
0
C, arbitrarily extended to a complex-valued function
on Y , is -integrable on Y .
(b)
_
Y
[h(x, y)[ d(y) < for x outside some set X
0
X of -measure 0 and the function
x
_
Y
h(x, y) d(y): XX
0
C, arbitrarily extended to a complex-valued function
on X, is -integrable on X.
(c) Formula (2.5) is valid.
The crucial condition to be checked in this Theorem is inequality (2.6) or (2.7). The
important conclusion in the Theorem is that (almost) all integrals in (2.5) are well-dened
and that the equalities in (2.5) (in particular the second one) hold.
2.3 Convolution
2.15 Denition The convolution product of two 2-periodic functions f, g is a function
f g on R given by
(f g)(x) :=
1
2
_

f(t) g(x t) dt (x R), (2.8)

provided the right hand side is well-dened. Then, clearly, f g is also 2-periodic and
f g = g f. Check these two properties by simple transformations of the integration
variable in (2.8)
For deriving further properties of the convolution product, the easiest case is when
f, g are continuous:
L
1
THEORY 15
2.16 Proposition If f, g (
2
then f g (
2
and
(f g)(n) =

f(n) g(n) (n Z). (2.9)
Proof For the proof of the continuity (in fact uniform continuity) of f g observe that

1
2
_

f(t) (g(x t) g(y t)) dt

|f|
1
sup
tR
[g(x t) g(y t)[.
Let > 0. Since any periodic continuous function is uniformly continuous on R (why?),
there exists > 0 such that [g(x t) g(y t)[ < for all t R if [x y[ < . Hence
[(f g)(x) (f g)(y)[ < |f|
1
if [x y[ < .
For the proof of (2.9) use Fubinis theorem in the case of continuous functions on a
bounded interval. Thus
(f g)(n) =
1
4
2
_

__

f(t) g(x t) dt
_
e
inx
dx
=
1
4
2
_

f(t) e
int
__

g(x t) e
in(xt)
dx
_
dt
=
1
4
2
_

f(t) e
int
g(n) dt =

f(n) g(n).
Ex. 2.17 Let f(x) := e
imx
, g(x) := e
inx
(m, n Z). Compute f g. Also check that the
result agrees with equality (2.9).
Ex. 2.18 Prove by Fubinis theorem in the case of continuous functions on a bounded
interval the following. If f, g, h (
2
then
(f g) h = f (g h) (associativity). (2.10)
2.19 Denition-Theorem Let f, g L
1
2
. Let (f g)(x) be dened by (2.8) for those
x R for which the integral on the right hand side of (2.8) converges absolutely.
(a) For almost all x R the integral on the right hand side of (2.8) converges absolutely.
Extend f g to a (2-periodic) function on R by choosing arbitrary values of (f g)(x)
on the set of measure zero where f(x) is not yet dened by (2.8). Then f g L
1
2
and
|f g|
1
|f|
1
|g|
1
. (2.11)
(b) The equivalence class of f g only depends on the equivalence classes of f and g. In
other words, for f, g L
1
2
the convolution product f g is well-dened as an element
of L
1
2
.
(c) For f, g L
1
2
, formula (2.9) is valid.
(d) For f, g, h L
1
2
, the associativity property (2.10) is valid.
Proof Let f, g L
1
2
. First observe that
1
4
2
_

__

[f(t) g(x t)[ dx

_
dt =
1
4
2
_

[f(t)[
__

[g(x t)[ dx
_
dt
=
1
2
_

[f(t)[ |g|
1
dt = |f|
1
|g|
1
< . (2.12)
16 CHAPTER 2
Hence, by Fubinis Theorem 2.14 we have that
_

[f(t) g(x t)[ dt <

for almost all x. Thus, (f g)(x) is well-dened by (2.8) for almost all x, and f g is
measurable on R. Again by Theorem 2.14, it follows that
1
4
2
_

f(t) g(x t) dt

dx < .
So f g L
1
2
. For the proof of (2.11) use that
|f g|
1

1
4
2
_

__

[f(t) g(x t)[ dt

_
dx =
1
4
2
_

__

[f(t) g(x t)[ dx

_
dt
by Proposition 2.13, and combine with (2.12).
For the proof of part (b) let f, g L
1
2
with |f|
1
= 0 or |g|
1
= 0. Then |f g|
1
= 0
because of (2.11).
For the proof of (c) repeat the proof of (2.9) for the continuous case. We can now
justify the interchange of integration order in the second equality of that proof by Fubinis
Theorem 2.14, because
1
4
2
_

__

[f(t) g(x t) e
inx
[ dx
_
dt < .
This last inequality is true because the left hand side equals | [f[ [g[ |
1
< .
Finally (d) can be proved similarly as in Exercise 2.18, with justication of the inter-
change of integration order by Theorem 2.14.
2.20 Proposition If f L
1
2
and g (
2
then f g (
2
and
|f g|

|f|
1
|g|

. (2.13)
Proof The proof given in Proposition 2.16 that f g (
2
if f, g (
2
, still works if the
assumption on f is relaxed to f L
1
2
. The proof of inequality (2.13) is straightforward.
Ex. 2.21 Show: If f, g L
2
2
then f g (
2
and
|f g|

|f|
2
|g|
2
. (2.14)
Hint First show that sup
xR
[(f g)(x)[ |f|
2
|g|
2
if f, g L
2
2
. Next use this inequality
and the fact that f g is continuous if f, g (
2
(see Proposition 2.16), together with the
density of (
2
in L
2
2
and the completeness of the Banach space (
2
with respect to the
sup norm.
L
1
THEORY 17
2.4 Further exercises
Ex. 2.22 Show that

n=0
2
n
cos(nx) =
4 2 cos x
5 4 cos x
(x R).
Compare with Exercise 2.9(e).
Ex. 2.23 Let A C such that [A[ , = 1. Let f(x) := (A+e
ix
)
1
(x R). Compute the
Fourier coecients

f(n).
Ex. 2.24 Let f L
1
2
. Express g(n) in terms of the

f(m) if:
(a) g(x) = f(x +a) (a R);
(b) g(x) = f(x);
(c) g(x) = f(x);
(d) g(x) = f(x).
Ex. 2.25 Let f L
1
2
, g (
1
2
. Show that f g (
1
2
and that (f g)

= f g

.
18
3 The Dirichlet kernel
3.1 Denition of the Dirichlet kernel
3.1 Let f L
1
2
. The formal doubly innite series

nZ

f(n)e
inx
(3.1)
is called the Fourier series of f. We will see that this series converges in a certain sense to
f(x), but the type of convergence will depend on the nature of f. Saying that the series
(3.1) converges in some sense to f(x) amounts to the same as saying that the sequence of
partial Fourier sums
S
N
(x) = (S
N
[f])(x) :=
N

n=N

f(n) e
inx
(N = 0, 1, 2, . . . , x R) (3.2)
converges in some sense to f(x) as N . Therefore it is important to examine the
functions S
N
[f] in more detail.
3.2 Denition-Proposition The partial Fourier sum (3.2) can be written as
(S
N
[f])(x) = (f D
N
)(x) =
1
2
_

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

f(x +t) D
N
(t) dt, (3.3)
where D
N
is the Dirichlet kernel:
D
N
(x) :=
N

n=N
e
inx
=
_

_
sin((N +
1
2
)x)
sin(
1
2
x)
(x / 2Z),
2N + 1 (x 2Z).
(3.4)
The Dirichlet Kernels D
2
, D
5
, D
10
, centered at 0, 2, 4 respectively.
(See also the graph of D
N
in van Rooij, p.13.)
DIRICHLET KERNEL 19
Proof Observe from (3.2) that
(S
N
[f])(x) =
1
2
_

f(t)
_
N

n=N
e
in(xt)
_
dt.
From (3.4) we see that D
N
(
2
and that it is an even function. The third equality in
(3.3) uses these last facts.
3.2 Criterium for convergence of Fourier series in one point
3.3 We will need the following three easy consequences of (3.4):
1
2
_

D
N
(t) dt = 1, (3.5)
[D
N
(x)[
1
[ sin(
1
2
x)[

[x[
(0 < [x[ ), (3.6)
and
D
N
(x) =
e
1
2
ix
2i sin(
1
2
x)
e
iNx

1
2
ix
2i sin(
1
2
x)
e
iNx
(x / 2Z). (3.7)
3.4 Theorem Let f L
1
2
, a R. Suppose that one of the two following conditions
holds:
(a) There are M, , > 0 such that
[f(a +t) f(a)[ M[t[

for < t < . (3.8)

(b) f is continuous in a and also right and left dierentiable in a.
Then we have:
lim
N
(S
N
[f])(a) = f(a).
Remark Condition (a) of Theorem 3.4 implies continuity of f in a. Functions f satisfying
condition (a) are said to be Holder continuous of order in a. (The terminology Lipschitz
continuous of order is equally common.) If (b) holds then
|f(a+t)f(a)|
|t|
is bounded for t
in some neighbourhood of 0 since it has a limit as t 0 and as t 0. Thus if (b) holds then
(a) is satised with = 1. So, we only need to prove the Theorem under condition (a).
Proof of Theorem 3.4 Assume condition (a). Let > 0. We want to show that
[S
N
(a) f(a)[ < for N suciently large. It follows from (3.3), (3.5) and (3.7) that
S
N
(a) f(a) =
1
2
_

(f(a +t) f(a)) D

N
(t) dt
=
_

g
a,+
(t)e
iNt
dt
_

g
a,
(t)e
iNt
dt, (3.9)
where
g
a,
(t) =
1
2
e

1
2
it
2i sin(
1
2
t)
(f(a +t) f(a)) (0 < [t[ < ). (3.10)
For 0 < [t[ < we can estimate [g
a,
(t)[ <
1
4
M[t[
1
(use (3.6) and (3.8)). For < [t[ <
we can estimate [g
a,
(t)[ < (4 sin(
1
2
))
1
[f(a+t) f(a)[. Hence the functions g
a,
are in
L
1
([, ]). By the Riemann-Lebesgue Lemma (Theorem 2.3) it follows that (3.9) tends
to 0 as N .
20 CHAPTER 3
3.5 Theorem Let f L
1
2
, let a R, and suppose that the limits
f(a
+
) := lim
xa
f(x), f(a

) := lim
xa
f(x)
exist. Suppose that one of the two following conditions holds:
(a) There are M, , > 0 such that
[f(a +t) f(a
+
)[ Mt

and [f(a t) f(a

)[ Mt

for 0 < t < .

(b) f is right and left dierentiable in a in the sense that the following two limits exist:
f

(a
+
) := lim
t0
f(a +t) f(a
+
)
t
, f

(a

) := lim
t0
f(a t) f(a

)
t
.
Then we have:
lim
N
(S
N
[f])(a) =
1
2
(f(a
+
) +f(a

)).
Proof Since D
N
is an even function, we conclude from (3.3) that
S
N
(a) =
1
2
_

1
2
(f(a +t) +f(a t)) D
N
(t) dt.
Hence
S
N
(a)
1
2
(f(a
+
) +f(a

)) =
1
2
_

_
f(a +t) +f(a t)
2

f(a
+
) +f(a

)
2
_
D
N
(t) dt.
Now put g(t) :=
1
2
(f(a + t) + f(a t)) for 0 < [t[ and g(0) := lim
t0
g(t) =
1
2
(f(a
+
)+f(a

)). Then condition (a) or (b) for f at the point a implies the corresponding
condition in Theorem 3.4 for g at the point 0. Now apply Theorem 3.4 to g at 0.
3.6 Corollary (localization principle) Let f, g L
1
2
, let a R, and suppose that
f(x) = g(x) for x in a certain neighbourhood of a. Then precisely one of the following two
alternatives holds for the two sequences
_
(S
N
[f])(a)
_

N=0
and
_
(S
N
[g])(a)
_

N=0
:
(a) The two sequences both converge and have the same limit.
(b) The two sequences both diverge.
Proof Apply Theorem 3.4 to the function f g at a. Since f g is identically zero in a
neighbourhood of a, we conclude that lim
N
(S
N
[f g])(a) = 0. Hence
lim
N
_
(S
N
[f])(a) (S
N
[g])(a)
_
= 0.
Thus the convergence and possible sum of the Fourier series of a function f L
1
2
at
a point a is completely determined by the restriction of f to an arbitrarily small neigh-
bourhood of f. If we leave a given function f unchanged on such a neighbourhood, but
change it elsewhere, then the Fourier coecients

f(n) may become completely dierent,
but convergence or divergence and the possible sum of the Fourier series at a will remain
the same. This explains why the above Corollary is called localization principle.
DIRICHLET KERNEL 21
Ex. 3.7 Let f be the sawtooth function dened in Exercise 1.25, formula (1.12). Put
f(x) := 0 for x 2Z.
(a) Prove that
f(x) = lim
N
(S
N
[f])(x) = 2

n=1
sin(nx)
n
(3.11)
with pointwise convergence for all x R.
(b) Show by substitution of x := /2 or x := /4 in (1.12), (3.11) that

4
= 1
1
3
+
1
5

1
7
+ , (3.12)

8
= 1 +
1
3

1
5

1
7
+
1
9
+
1
11

1
13
. . . . (3.13)
Ex. 3.8 Let 0 < < . Let f be a 2-periodic function which is given on [, ] by:
f(x) :=
_
1 if x [, ],
0 if < [x[ .
(a) Determine

f(n) (n Z).
(b) Determine

+ lim
N

0<|n|N
sin(n)
n
e
inx
provided the limit exists.
Ex. 3.9 Let f be a 2-periodic function which is C
1
outside 2Z and such that f(0
+
),
f(0

), f

(0
+
), f

(0

) exist in the sense of Theorem 3.5. Prove that

f(n) = O([n[
1
) as
[n[ , but not necessarily

f(n) = o([n[
1
) as [n[ .
Hint Use Theorem 2.5 and Exercise 1.25(a).
Ex. 3.10 Let f be a 2-periodic function which is C
1
outside a discrete set X having the
property that X [, ) is a nite set. Assume that f(x
+
), f(x

), f

(x
+
), f

(x

) exist
in the sense of Theorem 3.5 for each x X. Prove that

f(n) = O([n[
1
) as [n[ .
3.3 Continuous functions with non-convergent Fourier series
We will now show in a soft way (i.e., by using functional analytic methods) that there
exists a 2-periodic continuous function for which the Fourier series does not converge
everywhere. For this we need a hard estimate of the Dirichlet kernel in the two lemmas
below. This estimate will also be useful elsewhere. We refer to Stein & Shakarchi for
an explicit example of a 2-periodic continuous function with not everywhere converging
Fourier series.
3.11 Lemma Let
(x) := cot(
1
2
x) 2x
1
(0 < [x[ < 2). (3.14)
22 CHAPTER 3
Then
D
N
(x) =
2 sin(Nx)
x
+(x) sin(Nx) + cos(Nx) (0 < [x[ < 2). (3.15)
Proof It follows from (3.4) that, for 0 < [x[ < 2,
D
N
(x) =
sin((N +
1
2
)x)
sin(
1
2
x)
= sin(Nx) cot(
1
2
x) + cos(Nx).
Ex. 3.12 Show the following:
(a) lim
x0
(x) = 0;
(b) lim
x0

(x) =
1
6
;
(c) Put (x) := 0. Then is dierentiable and strictly decreasing on (2, 2).
(d) () = 2
1
, () = 2
1
, sup
x
[(x)[ = 2
1
.
3.13 Lemma We have
_

[D
N
(x)[ dx = 8
1
log N +O(1) as N .
In particular,
lim
N
_

[D
N
(x)[ dx = .
Proof It follows from (3.15) and Exercise 3.12(d) that

[D
N
(x)[ [2x
1
sin(Nx)[

[D
N
(x) 2x
1
sin(Nx)[ 2
1
+ 1 (0 < [x[ < ).
Hence
_

[D
N
(x)[ dx = 2
_

sin(Nx)
x

dx +O(1) as N .
Next we write
2
_

sin(Nx)
x

dx = 4
_
N
0
[ sin(x)[
x
dx
= 4
N1

n=0
_
1
0
sin(x)
x +n
dx
= 4
N1

n=1
_
1
0
sin(x)
x +n
dx +O(1) as N
(1)
= 4
1
N1

n=1
(n
1
+ (n 1)
1
) 4
1
N

n=1
_
1
0
cos(x)
(x +n)
2
dx +O(1) as N
(2)
= 8
1
N

n=1
n
1
+O(1) as N
(3)
= 8
1
log N +O(1) as N .
In equality (1) we used integration by parts. In equality (2) we majorized [
_
1
0
(x +
n)
2
cos(x) dx[ by n
2
and we used that

N
n=1
n
2
= O(1) as N . In equality
(3) we used that

N
n=1
n
1
= log N + O(1) as N (by comparing with the corre-
sponding Riemann integral, see the denition of Eulers constant ).
DIRICHLET KERNEL 23
Ex. 3.14 Let : [, ] R be continuous with only nitely many zeros. Dene the
linear functional L: (
2
C by
L(f) :=
1
2
_

f(x) (x) dx (f (
2
). (3.16)
Prove that L is a bounded linear functional with operator norm
|L| = ||
1
=
1
2
_

[(x)[ dx. (3.17)

3.15 Theorem For all x R there exists f (
2
such that the sequence
_
(S
N
[f])(x)
_

N=0
does not converge to a nite limit.
Proof Without loss of generality we may take x := 0. Dene the linear functional
L
N
: (
2
C by
L
N
(f) := (S
N
[f])(0) =
1
2
_

f(t) D
N
(t) dt.
It follows from Exercise 3.14 that |L
N
| = |D
N
|
1
, and it follows next from Lemma 3.13
that the sequence (|L
N
|)

N=0
is unbounded. Now suppose that lim
N
L
N
(f) exists for
all f (
2
. Then, for all f (
2
, the sequence (L
N
(f))

N=0
is bounded. Hence, by
the Banach-Steinhaus theorem (see syll. Functionaalanalyse), the sequence (|L
N
|)

N=0
is
bounded. This is a contradiction.
Ex. 3.16 Let f L
1
2
and let a R. Prove that for the two sequences
_
(S
N
[f])(a)
_

N=0
and
1

f(a +t)
sin(Nt)
t
dt, N = 0, 1, 2, . . .
precisely one of the following two alternatives holds:
(a) Both sequences converge with the same limit.
(b) Both sequences diverge.
Conclude that, for f satisfying one of the conditions of Theorem 3.5, we have
1
2
(f(a
+
) +f(a

)) =
1
lim
N
_

f(a +t)
sin(Nt)
t
dt. (3.18)
Ex. 3.17 Prove, by using (3.18), that
_

0
sin x
x
dx = lim
N
_

0
sin(Nt)
t
dt =
1
2
. (3.19)
24 CHAPTER 3
3.4 Injectivity of the Fourier transform
3.18 In this subsection we will prove that a function f L
1
2
which has all its Fourier
coecients

f(n) = 0, is almost everywhere equal to zero. For the case that f L
2
2
, this
was already implied by Parsevals identity (1.9). However, Parsevals identity depended
on Theorem 1.11, for which we postponed the proof. Let us rst consider the case that
f (
2
.
Proposition Let f (
2
such that

f(n) = 0 for all n Z. Then f = 0.
Proof Let F(x) :=
_
x
0
f(t) dt. Then F is a C
1
-function on R and, because

f(0) = 0,
the function F is 2-periodic. Thus F (
1
2
. Since F

= f, we have

f(n) = in

F(n) (see
(2.2)). Hence

F(n) = 0 if n ,= 0. It follows that (S
N
[F])(x) =

F(0). By Theorem 3.4 we
conclude that F(x) = lim
N
(S
N
[F])(x) =

F(0). Hence f(x) = F

(x) = 0.
For the similar result in the case that f L
1
2
we proceed as follows.
3.19 Lemma Let f L
1
2
such that

f(0) = 0. Put F(x) :=
_
x
0
f(t) dt. Then F (
2
and

f(n) = in

F(n).
Proof Continuity of F is a consequence of the dominated convergence theorem, applied
to the functions f
x,t
, where
x,t
is the characteristic function of the interval [x, x + t)
and letting t 0. Now we compute for n ,= 0:

F(n) =
_
2
0
_
x
0
f(t)e
inx
dt dx =
_
2
0
_
2
t
e
inx
dxf(t) dt
=
_
2
0
_
e
int
1
in
_
f(t) dt =
1
in
(

f(n)

f(0)) =
1
in

f(n).
We used Fubinis theorem for the second equality.
Remark It is a non trivial result of Lebesgue that for almost all x F

(x) exists and

equals f(x). (see Rudin, Theorem 8.17)
3.20 Theorem (a) Let f L
1
2
such that

f(n) = 0 for all n Z. Then f = 0 a.e. .
(b) Let f, g L
1
2
such that

f(n) = g(n) for all n Z. Then f = g a.e. .
Proof It is sucient to prove part (a). Let F be as in Lemma 3.19. Then it follows
that F (
2
and that

F(n) = 0 for n ,= 0. Thus F

F(0) = 0 by Proposition 3.18.
Since F(0) = 0, we conclude that F = 0. It follows that
_
b
a
f(t) dt = 0 for every choice of
a, b R. But then
_
E
f(t) dt = 0 for (bounded) open and also for (bounded) closed sets.
By regularity of Lebesgue measure, we conclude that
_
E
f dt = 0 for every Borel set. It
follows that f = 0 a.e. .
3.21 Corollary (a) The linear span of the functions t e
int
(n Z) is dense in L
2
2
with respect to the norm | . |
2
.
(b) The functions t e
int
(n Z) form an orthonormal basis of L
2
2
.
Proof Apply Denition-Theorem 1.12 to the case f L
2
2
of Theorem 3.20.
DIRICHLET KERNEL 25
So we have now proved parts (b) and (c) of Theorem 1.11, therefore we have also
denitely established all results in subchapter 1.2 which depended on Theorem 1.11, see
the summarizing 1.18. In the next chapter we will also prove part (a) of Theorem 1.11.
As a corollary of the injectivity result of Theorem 3.20 we can also give the following
addition to Theorem 2.8.
3.22 Corollary Let f L
1
2
. If

f
1
(Z) then f (
2
and, for almost all x R:
f(x) =

n=

f(n) e
inx
. (3.20)
The right hand side of (3.20) converges absolutely and uniformly for x R.
Proof Denote the right hand side of (3.20) by g(x). Then, by 2.8, g (
2
and g(n) =

f(n) for all n Z. Now apply Theorem 3.20(b).

3.5 Uniform convergence of Fourier series
We will now consider an analogue of Theorem 3.4 such that f satises not just a Holder
condition at one point, but a uniform Holder condition on some interval (a, b). Then it
will turn out that S
N
[f] converges uniformly to f on each compact subset of (a, b). For
f (
2
2
this result is almost immediate:
3.23 Proposition Let f (
2
2
. Then lim
N
S
N
[f] = f, uniformly on R.
Proof It follows from Theorem 2.5(a) that

f(n) = o([n[
2
) as [n[ . Now apply
Corollary 3.22.
3.24 The following lemma quickly follows from the well-known Arzela-Ascoli theorem
(see for instance A. Browder, Mathematical Analysis, Springer, 1996, Theorem 6.71 and
Corollary 6.73).
Lemma Let (X, d) be a compact metric set. Let (
n
)

n=1
be a sequence of complex-
valued functions on X which is equicontinuous on X, i.e., such that, for each > 0, there
is a > 0 with the property that [
n
(x)
n
(y)[ < for al n N if x, y X and
d(x, y) < . Suppose that lim
n

n
(x) = 0 for x X. Then lim
n

n
= 0 uniformly
on X.
Proof Suppose that the convergence
n
0 is not uniform on X. Then there exist > 0,
an increasing sequence of positive integers n
1
, n
2
, . . ., and a sequence x
1
, x
2
, . . . in X such
that [
n
k
(x
k
)[ . By compactness of X the sequence x
1
, x
2
, . . . has a subsequence
converging to some x
0
X. Without loss of generality we may assume that the sequence
x
1
, x
2
, . . . already converges to x
0
. Then
[
n
k
(x
k
)[ [
n
k
(x
k
)
n
k
(x
0
)[ +[
n
k
(x
0
)[.
By the assumptions, there exists K N such that [
n
k
(x
0
)[ <
1
2
if k K and [
m
(x
k
)

m
(x
0
)[ <
1
2
for all m N if k K. Hence
n
k
(x
k
) < if k K. This is a contradiction.
26 CHAPTER 3
3.25 Theorem Let f L
1
2
. Suppose that there exist an interval (a, b) and positive
real numbers M, such that
[f(x) f(y)[ M [x y[

for all x, y (a, b) (3.21)

(a uniform Holder condition on (a, b)). Then lim
N
S
N
[f] = f uniformly on each subin-
terval [c, d] with a < c < d < b.
Proof By the proof of Theorem 3.4 we can write
S
N
(x) f(x) =
_

g
x,+
(t)e
iNt
dt
_

g
x,
(t)e
iNt
dt (3.22)
with the functions g
x,
being given by (3.10). Put

N,
(x) :=
_

g
x,
(t) e
iNt
dt (x [c, d]).
By (3.21) and the proof of Theorem 3.4, the functions g
x,
belong to L
1
2
for x [c, d]. So
lim
N

N,
(x) = 0 for x [c, d] by the Riemann-Lebesgue Lemma (Theorem 2.3). We
will show that the functions
N,
are equicontinuous on [c, d]. Then we can apply Lemma
3.24 and thus prove the Theorem.
Let 0 < < such that < c a and < b d. Then we can estimate for x [c, d]
and 0 < [t[ < that [g
x,
(t)[ <
1
4
M[t[
1
(use (3.21) and (3.6)). We can estimate for
x, y [c, d] and < [t[ < that
[g
x,
(t) g
y,
(t)[ < (4 sin(
1
2
))
1
_
[f(x +t) f(y +t)[ +[f(x) f(y)[
_
.
Next, for x, y [c, d]:
[
N,
(x)
N,
(y)[
_
_
|t|<
+
_
<|t|<
_
[g
x,
(t) g
y,
(t)[ dt
M
_

0
[t[
1
dt +
1
4 sin(
1
2
)
_

[f(x +t) f(y +t)[ dt +

[f(x) f(y)[
2 sin(
1
2
)
.
Let > 0. We will nd > 0 such that the last expression is dominated by if x, y [c, d]
and [x y[ < . First take such that M
_

0
[t[
1
dt < /3. Then we can nd > 0 such
that the second and third term are dominated by /3 if [x y[ < . For the third term
this follows because f is uniformly continuous on [c, d]. For the second term this follows
from Proposition 1.9. This settles the equicontinuity of the functions
N,
on [c, d]. Thus
we can apply Lemma 3.24 and the Theorem will follow.
3.26 Corollary Let f (
1
2
, or let f (
2
with piecewise continuous derivative. Then
lim
N
S
N
[f] = f, uniformly on R.
Ex. 3.27 Let the 2-periodic function f be determined by f(x) := [x[ for x .
Determine the Fourier coecients

f(n). The uniform convergence of (S
N
[f])

N=0
on R is
implied by Corollary 3.26. Check this uniform convergence independently by using the
explicit values of

f(n).
DIRICHLET KERNEL 27
3.6 The Gibbs phenomenon
The approximation of a piecewise continuous 2-periodic function (for instance the saw-
tooth function) by its partial Fourier sum shows a remarkable overshooting behaviour
near the jumps of the function. This behaviour is called the Gibbs phenomenon.
3.28 Let the 2-periodic function f be given by f(x) := x for 0 < x < 2. This
function was studied in Exercises 1.25 and 3.7. We have
S
N
(x) = (S
N
[f])(x) = 2
N

n=1
sin(nx)
n
=
_
x
0
D
N
(t) dt x
= 2
_
x
0
sin(Nt)
t
dt +
_
x
0
(t) sin(Nt) dt +N
1
sin(Nx) x (0 < [x[ < 2), (3.23)
where we used (3.15) and where the function is given by (3.14).
Ex. 3.29 Let (a, b) be a bounded interval and let c (a, b). Let g C
1
((a, b)) and dene
G
n
(x) :=
_
x
c
g(t) e
int
dt (n Z, x (a, b)). Prove that G
n
(x) = O([n[
1
) as [n[ ,
uniformly for x in a compact subinterval of (a, b).
Hint Use integration by parts.
3.30 Let 0 < r < 2. We conclude from Exercises 3.29 and 3.12 that
_
x
0
(t) sin(Nt) dt = O(N
1
) as N , uniformly for [x[ r. (3.24)
The integral sine is the function dened by
Si(x) :=
_
x
0
sin y
y
dy (x R). (3.25)
This is an even C
1
-function (in fact it is an analytic function). It follows from (3.19) that
Si() := lim
x
Si(x) =
1
2
.
The absolute maximum of Si(x) for x [0, ) is reached for x = . A numerical compu-
tation shows that, approximately,
Si()
Si()
1.179 . . . .
Let 0 < r < 2. It follows from (3.23), (3.25) and (3.24) that
S
N
(x) f(x) = 2Si(Nx) +O(N
1
)
S
N
(x) f(x) = 2Si(Nx) + +O(N
1
)
_
as N , uniformly for 0 < x r.
This implies:
lim
N
_
S
N
(N
1
) f(N
1
)
_
= 2Si() 0.179 . . . ,
lim
N
_
S
N
(N
1
) f(N
1
)
_
= 2Si() + 0.179 . . . .
28 CHAPTER 3
Also, for 0 < < , we have
lim
N
_
sup
0<x
(S
N
(x) f(x))
_
= 2Si() , (3.26)
lim
N
_
inf
x<0
(S
N
(x) f(x))
_
= 2Si() +, (3.27)
but
lim
N
_
sup
x2
[S
N
(x) f(x)[
_
= 0
by Theorem 3.25.
Ex. 3.31 Let h C
1
([0, 2]) Let g be a 2-periodic function which coincides with h
on (0, 2). Write g(0
+
) := h(0) and g(0

) := h(2). Assume that g(0

+
) g(0

). Let
0 < < . Prove that
lim
N
_
sup
0<x
((S
N
[g])(x) g(x))
_
=
1
2
(g(0
+
) g(0

)) (2
1
Si() 1),
lim
N
_
inf
x<0
((S
N
[g])(x) g(x))
_
=
1
2
(g(0
+
) g(0

)) (2
1
Si() 1).
Hint Let the 2-periodic function f be given by f(x) := x for 0 < x < 2. Put
p(x) := g(x)
g(0
+
) g(0

)
2
f(x) (0 < x < 2).
Then p (
2
with piecewise continuous derivative. Now use (3.26), (3.27), and Corollary
3.26.
The Gibbs phenomenon.
Left: the sawtooth and S
64
, Right: (S
N
(N
1
x) f(N
1
x))/ for N = 1024
DIRICHLET KERNEL 29
3.7 Further exercises
Ex. 3.32 Prove that
sin x +
1
3
sin 3x +
1
5
sin 5x + = /4 if 0 < x < .
Hint Use Exercise 3.8 or use (1.12), (3.11).
Ex. 3.33 Let L
1
2
. Dene the linear functional L: (
2
C by (3.16). Prove that L
is a bounded linear functional with norm given by (3.17).
Hint Use Exercises 3.14 and 1.32.
Ex. 3.34 Prove that D
N
(x) can be written as a polynomial of degree 2N in cos(
1
2
x).
Ex. 3.35 Prove that between each two successive zeros of D
N
(x) there is precisely one
zero of D

N
(x).
Ex. 3.36 1) Let a
n

1
, a
n

1
be sequences of complex numbers. Prove (for N =
1, 2, . . .) the Summation by parts formula
N

n=1
a
n
(b
n+1
b
n
) = a
N+1
b
N+1
a
1
b
1

N

n=1
b
n+1
(a
n+1
a
n
).
2) Write down a version of this formula for the case where b
n
=

n
j=1
c
j
.
3) Give a direct proof of the convergence of the series in exercise 3.32 (without using
that this is a Fourier series of a smooth function)
4) Show that the series

n=2
sinnx
log n
converges pointwise for every x.
5) What can you say about uniform convergence?
30
4 The Fejer kernel
4.1 Ces`aro convergence
4.1 Denition Let (a
n
)

n=1
be a sequence of complex numbers. If the sequence does
not have a limit then we may try to nd a limit in a generalized sense by considering a
new sequence (b
n
)

n=1
with b
n
being the mean of the rst n elements of the sequence (a
k
),
i.e.,
b
n
:= n
1
(a
1
+a
2
+ +a
n
).
If a := lim
n
b
n
exists as a nite limit then we say that the sequence (a
n
) is Ces`aro
convergent with Ces`aro limit a. In that case we use the notation
(C) lim
n
a
n
= a. (4.1)
Similarly, we say that the series

n=1
a
n
is Ces`aro convergent with Ces`aro sum s if
the sequence (s
n
) of partial sums s
n
:=

n
k=1
a
k
is Ces`aro convergent with Ces`aro limit s.
In that case we use the notation
(C)

n=1
a
n
= s. (4.2)
More generally, Ces`aro convergence of the sum

n=n
0
a
n
with Ces`aro sum s means Ces`aro
convergence of the sequence (s
n
)

n=n
0
of partial sums s
n
:=

n
k=n
0
a
k
to Ces`aro limit s,
i.e., that lim
n
(n n
0
+ 1)
1
(s
n
0
+ +s
n
) = s.
4.2 Example
The sequence 1, 0, 1, 0, 1, 0, . . . does not converge but has Ces`aro limit
1
2
.
The sum 1 1 + 1 1 + 1 1 + does not converge but has Ces`aro sum
1
2
.
Ex. 4.3 Show the following. The sequence a
0
, a
1
, a
2
, . . . has Ces`aro limit a i the
sequence a
1
, a
2
, . . . has Ces`aro limit a.
Conclude that it is not necessary in the notation (4.1) to mention where the sequence (a
n
)
starts.
4.4 Proposition If the sequence (a
n
)

n=1
has limit a then it has also Ces`aro limit a.
Proof Let M, N N with M < N. For n N we have
[n
1
(a
1
+ +a
n
) a[ n
1
([a
1
a[ + +[a
M
a[) +n
1
([a
M+1
a[ + +[a
n
a[)
MN
1
sup
k=1,2,...
[a
k
a[ + sup
kM+1
[a
k
a[.
Let > 0. Because the sequence (a
n
) converges, we have A := sup
k=1,2,...
[a
k
a[ <
and we can choose M such that [a
k
a[ <
1
2
if k M. Hence
[n
1
(a
1
+ +a
n
) a[ N
1
MA+
1
2
if n N > M.
Choose N > M such that N
1
MA <
1
2
. We conclude that [n
1
(a
1
+ + a
n
) a[ if
n N.
FEJER KERNEL 31
Ex. 4.5 Give an example of an unbounded sequence which is Ces`aro convergent.
Ex. 4.6 Which of the following statements are true?
(i) If (C) lim
n
a
n
= a then (C) lim
n
(a
n
)
3
= a
3
.
(ii) If (C)

n=1
a
n
= s then (C) lim
n
a
n
= 0.
4.2 Denition of Fejer kernel
4.7 In (3.2) we introduced the partial Fourier sums S
n
[f] of a function f L
1
2
. In
Chapter 3 we examined convergence of the sequence
_
(S
N
[f])(x)
_

N=0
for xed x. More
generally we may examine Ces`aro convergence of this sequence, i.e., convergence of the
sequence
_
(
N
[f])(x)
_

N=0
, where

N
[f] :=
S
0
[f] +S
1
[f] + +S
N
[f]
N + 1
(f L
1
2
, N = 0, 1, 2, . . .). (4.3)
The function
N
[f] is called the N
th
Ces`aro mean of f.
4.8 Denition-Proposition The function in (4.3) can be written as
(
N
[f])(x) = (f K
N
)(x) =
1
2
_

f(t) K
N
(x t) dt =
1
2
_

f(x +t) K
N
(t) dt, (4.4)
where K
N
(x) is the Fejer kernel:
K
N
(x) :=
1
N + 1
N

n=0
D
n
(x) =
N

n=N
N + 1 [n[
N + 1
e
inx
=
_

_
1
N + 1
_
sin(
1
2
(N + 1)x)
sin(
1
2
x)
_2
(x / 2Z),
N + 1 (x 2Z).
(4.5)
Proof The second equality in (4.5) follows from the denition of D
n
(x) in (3.4). From
(3.4) we get also that
K
N
(x) =
sin(
1
2
x) + sin(
3
2
x) + sin((N +
1
2
)x)
(N + 1) sin(
1
2
x)
.
Now multiply numerator and denominator of this last expression by sin(
1
2
x) and use that
sin(
1
2
x) sin((k+
1
2
)x) =
1
2
(cos(kx)cos((k+1)x)) and 1cos((N+1)x) = 2 sin
2
(
1
2
(N+1)x).
This proves the last equality in (4.5). From (4.5) we see that K
N
(
2
and that it is an
even function. The third equality in (4.4) uses these last facts.
32 CHAPTER 4
The Fejer Kernels K
8
, K
16
, centered at 0, 4 respectively.
4.3 Ces`aro convergence of Fourier series
4.9 We need the following three easy consequences of (4.5):
K
N
(x) 0 (x R), (4.6)
1
2
_

K
N
(t) dt = 1, (4.7)
K
N
(x)
1
(N + 1) sin
2
(
1
2
x)

2
(N + 1) x
2
(0 < [x[ ). (4.8)
4.10 Theorem (Fejer) Let f L
1
2
, a R. Suppose that f is continuous in a. Then
lim
N
(
N
[f])(a) = f(a). (4.9)
Moreover, if f (
2
then the convergence in (4.9) is uniform for a R.
FEJER KERNEL 33
Proof Let > 0. We want to show that [
N
(a) f(a)[ < for N suciently large. It
follows from (4.4) and (4.7) that

N
(a) f(a) =
1
2
_

(f(a +t) f(a)) K

N
(t) dt.
For any [0, ) we split up the above integral as
_

=
_
|t|<
+
_
<|t|<
. Then, by use
of (4.6),
[
N
(a) f(a)[
1
2
_

[f(a+t) f(a)[ K
N
(t) dt +
1
2
_
<|t|<
[f(a+t) f(a)[ K
N
(t) dt.
Denote the two successive terms on the right hand side by I
1
and I
2
. First we estimate
I
1
. By the continuity of f in a we can choose such that [f(a +t) f(a)[
1
2
if [t[ .
Keep this value of . Then, by use of (4.7), I
1

1
2
. Next we estimate I
2
. It follows from
(4.8) that
I
2


2(N + 1)
_
<|t|<
[f(a +t) f(a)[ t
2
dt

2(N + 1)
2
_

([f(a +t)[ +[f(a)[) dt

2
(N + 1)
2
(|f|
1
+[f(a)[).
Hence I
2

1
2
for N suciently large. This proves the rst part of the theorem.
Next assume that f (
2
. Then f is uniformly continuous on R (by the compactness
of [, ] and by the 2-periodicity). Thus we can choose such that [f(a+t)f(a)[
1
2

for all a R if [t[ . With this choice of we have I

1

1
2
for all a. Then
I
2


2
(N + 1)
2
(|f|
1
+|f|

).
Hence I
2

1
2
for all a if N is suciently large. We conclude that [
N
(a) f(a)[ < for
all a if N is suciently large.
4.11 Corollary (Fejer) Let f L
1
2
, let a R, and suppose that the limits
f(a
+
) := lim
xa
f(x), f(a

) := lim
xa
f(x)
exist. Then we have:
lim
N
(
N
[f])(a) =
1
2
(f(a
+
) +f(a

)).
Ex. 4.12 Give the proof of Corollary 4.11 (compare with the proof of Theorem 3.5).
4.13 Theorem 1.11(a) is an immediate corollary of the second statement in Theorem
4.10:
34 CHAPTER 4
Corollary (Weierstrass) The space of trigonometric polynomials is dense in (
2
with
respect to the norm | . |

.
Proof Let f (
2
. Let > 0. Then, by the uniform convergence stated in Theorem
4.10, we can take N such that |
N
[f]f|

< . Now observe that

N
[f] is a trigonometric
polynomial.
4.14 Corollary Let f L
1
2
and a R. Suppose that f is continuous in a. If
lim
N
(S
N
[f])(a) exists then this limit is equal to f(a).
Proof By Proposition 4.4 the limit is equal to lim
N
(
N
[f])(a) and hence, by Theorem
4.10 to f(a).
4.15 Remark Observe that
N
[f] := f K
N
(f L
1
2
), where K
N
(N = 0, 1, 2, . . .)
is an even nonnegative function belonging to L
1
2
with properties (2)
1
_

K
N
(t) dt = 1
and:
For each (0, ): lim
N
K
N
(x) = 0 uniformly for [x[ .
Observe that the proof of Theorem 4.10 only uses these properties of K
N
and
N
[f].
Hence Theorem 4.10 remains valid for any choice of the functions K
N
satisfying the above
properties.
Ex. 4.16 Let f L
1
2
. Show that lim
N

N
[f] = f in L
1
2
.
Hint Show that lim
N
|
N
[f] f|
1
= 0 for f (
2
and use that (
2
is dense in L
1
2
.
4.4 Another approximation theorem of Weierstrass
Corollary 4.13 is an approximation theorem involving trigonometric polynomials. From
this we can derive an even more famous approximation theorem of Weierstrass involving
ordinary polynomials. (There exist many other proofs of this theorem.) As a tool for the
proof we introduce Chebyshev polynomials.
4.17 Denition-Proposition For each nonnegative integer n there is a unique poly-
nomial T
n
in one variable such that
T
n
(cos t) = cos(nt) (t R). (4.10)
The polynomial T
n
has degree n. It is called a Chebyshev polynomial (of the rst kind).
Proof For x [1, 1] T
n
(x) is uniquely determined by (4.10). Hence, if T
n
is a polyno-
mial then it is unique. Since
cos t cos nt =
1
2
cos(n + 1)t +
1
2
cos(n 1)t,
we have
T
n+1
(x) = 2x T
n
(x) T
n1
(x) (n = 1, 2, . . .),
while T
0
(x) = 1, T
1
(x) = x. Now it follows by induction with respect to n that T
n
(x) is a
polynomial of degree n in x.
FEJER KERNEL 35
Ex. 4.18 Show that
_
1
1
T
n
(x) T
m
(x)
dx

1 x
2
=
_
0, n ,= m,
/2, n = m ,= 0,
, n = m = 0.
Thus the T
n
form an orthogonal system in L
2
((1, 1); (1 x
2
)

1
2
dx). Show also that this
orthogonal system is complete.
Ex. 4.19 Show that for each nonnegative integer n there is a unique polynomial U
n
(x)
of degree n in x such that
U
n
(cos t) =
sin(n + 1)t
sin t
(0 ,= t R). (4.11)
Show that
U
n+1
(x) = 2x U
n
(x) U
n1
(x) (n = 1, 2, . . .),
while U
0
(x) = 1, U
1
(x) = 2x. The polynomial U
n
is called a Chebyshev polynomial (of
the second kind). Show that
_
1
1
U
n
(x) U
m
(x)
_
1 x
2
dx =
_
0, n ,= m,
/2, n = m.
Thus the U
n
form an orthogonal system in L
2
((1, 1); (1 x
2
)
1
2
dx). Show also that this
orthogonal system is complete.
4.20 Theorem (Weierstrass) Let f C([a, b]). Then there exists for each > 0 a
polynomial P in one variable such that [f(x) P(x)[ < for x [a, b].
Proof Without loss of generality we may assume that [a, b] = [1, 1] (why?). Put g(t) :=
f(cos t). Then g (
2
and g is even, so g(n) = g(n). Hence
(
N
[g])(t) = g(0) + 2
N

n=1
N + 1 n
N + 1
g(n) cos(nt)
and
N
[g] g for N , uniformly on R (see Theorem 4.9), so certainly uniformly on
[0, ]. Put
f
N
(x) := g(0) + 2
N

n=1
N + 1 n
N + 1
g(n) T
n
(x).
Then f
N
(cos t) = (
N
[g])(t) and f
N
(x) is a polynomial of degree N in x. Since
N
[g] g
for N , uniformly on [0, ], we have that f
N
f as N , uniformly on [1, 1].
Ex. 4.21 Let f(x) := [x[. Determine a sequence of polynomials which converges uni-
formly to f on [1, 1].
Ex. 4.22 The analogue of Theorem 4.20 does not hold on R. Give an example of a
continuous function on R which cannot be uniformly approximated by polynomials on R.
36 CHAPTER 4
4.5 Further exercises
Ex. 4.23 For 0 r < 1 let
P
r
(x) :=

n=
r
|n|
e
inx
(x R), (4.12)
the so-called Poisson kernel. For f L
1
2
put
A
r
[f] := f P
r
. (4.13)
The functions A
r
[f] are known as Abel means. Let a R and suppose that f is continuous
in a. Prove that
lim
r1
(A
r
[f])(a) = f(a). (4.14)
Moreover show that, if f (
2
, the convergence in (4.14) is uniform for a R.
Hint Use Remark 4.15.
Ex. 4.24 Prove that the Gibbs phenomenon cannot occur for the Ces`aro and Abel means.
That is, if m < f < M, then m < A
r
[f], C
N
[f] < M.
37
5 Some applications of Fourier series
5.1 The isoperimetric inequality
5.1 Theorem The area A of a region in the plane which is encircled by a closed non-
selntersecting C
1
-curve of length L satises A L
2
/(4). Equality holds i the curve is
a circle.
Proof Without loss of generality we may assume that L = 2, and that the curve is
positively oriented and parametrized by its arc length. We may also identify the plane
with C. Then the curve has the form t f(t) with f (
1
2
and with [f

(t)[ = 1 for all

t. Furthermore we may assume without loss of generality that

f(0) =
1
2
_

f(t) dt = 0.
Then we have to show that A with equality i f(t) = e
i(t+t
0
)
for some t
0
R. Now
we have
A
(1)
=
1
2
Im
_
2
0
f

(t) f(t) dt = Imf

, f) [f

, f)[
(2)
|f

|
2
|f|
2
(3)
= |f|
2
(4)
= |f

f(0)|
2
(5)
|f

|
2
(6)
= . (5.1)
Equality (1) follows from:
Ex. 5.2 Let t f(t) be a positively oriented closed non-selntersecting C
1
-curve in C.
Show that the area of the encircled region equals
1
2
Im
_
2
0
f

(t) f(t) dt.

Inequality (2) in (5.1) is the Cauchy-Schwarz inequality. Equalities (3) and (6) use that
|f

|
2
= 1 by the assumption [f

(t)[ = 1. Equality (4) uses the assumption

f(0) = 0.
Equality (5) follows from:
Ex. 5.3 Let f (
1
2
. Show that |f

f(0)|
2
|f

|
2
with equality i

f(n) = 0 for
n ,= 1, 0, 1.
The proof of the last part of Theorem 5.1 is also left as an exercise:
Ex. 5.4 Show that equality everywhere in (5.1) implies that f(t) = e
i(t+t
0
)
for some
t
0
R.
5.2 The heat equation
5.5 The partial dierential equation

t
F(t, x) =
1
2

2
x
2
F(t, x) ((t, x) (0, ) R) (5.2)
is called the heat equation. It is often considered with initial value
F(0, x) = f(x) (x R) (5.3)
38 CHAPTER 5
for a given function f on R. We will assume that f (

2
and we are interested in
a function F which is continuous on [0, ) R and C

on (0, ) R, which satises

F(t, x + 2) = F(t, x) for all t 0, and which is a solution of (5.2), (5.3). In a suitable
normalization of the variable t we can interprete this as the evolution in time of the
temperature on a ring which is idealized as the unit circle parametrized by its arc length.
Then F(t, x) is the temperature at position x and time t. The function f gives the initial
temperature distribution at time t = 0.
Put F
t
(x) := F(t, x) and

n
(t) :=

F
t
(n) =
1
2
_

F(t, x) e
inx
dx.
Ex. 5.6 Let f (

2
and let F be as above. Show the following.
(a)
n
is continuous on [0, ).
(b)
n
is C

on (0, ) with kth derivative given by

(k)
n
(t) =
1
2
_

k
F(t, x)
t
k
e
inx
dx.
(c) We have

n
(t) =
1
2
n
2

n
(t),
n
(0) =

f(n),
n
(t) =

f(n) e

1
2
n
2
t
.
Ex. 5.7 Let f (

2
. Show that
F(t, x) :=

n=

f(n) e
inx
1
2
n
2
t
(5.4)
is a well-dened C

function on (0, ) R, continuous on [0, ) R, 2-periodic in x

and satisfying (5.2), (5.3). Conclude that F(t, x) given by (5.4) is the unique solution of
(5.2), (5.3) of this type.
Hint For the rst part use Theorem 2.5(b).
Ex. 5.8 Let F be given by (5.4). Show that
_

F
t
(x) dx =
_

f(x) dx (t > 0)
(i.e., the total heat on the ring remains constant in time), and
lim
t
F(t, x) =

f(0)
(i.e., the temperature distribution tends to the constant distribution as t ).
Ex. 5.9 Dene the heat kernel by
H
t
(x) :=

nZ
e
inx
1
2
n
2
t
(t > 0).
Prove that H
t
(

2
. Show that, for F given by (5.4), we have F
t
= f H
t
(t > 0).
Ex. 5.10 Show that (f H
s
) H
t
= f H
s+t
(s, t > 0) and give a physical explanation
of this formula.
APPLICATIONS 39
5.3 Weyls equidistribution theorem
5.11 Theorem Let 0 < R Q and let 0 a < b 1. Then
lim
n
#r N [ 0 < r < n, < r > [a, b]
n
= b a.
Here # denotes the number of elements of a set and < x > denotes the fractional part
of a positive number x.
Ex. 5.12 Prove this theorem by completing the following outline:
1) Let f C
2
and R Q. Prove that
lim
n

n
j=1
f(2j)
n
=
1
2
_
2
0
f(t) dt.
a. In case f 1.
b. In case f(t) = e
ikt
, k Z.
c. In case f is a trigonometric polynomial.
d. For general f C
2
.
2) Prove that the formula is also valid if f is the periodic extension of the characteristic
function
[a,b]
of a closed interval [a, b] (0, 2). (Look for continuous functions f
1
, f
2
such that f
1
f f
2
and
_
2
0
(f
2
f
1
)(t) dt < ) and apply 1).
Derive the equidistribution theorem.
40
6 Generalities about Fourier integrals
This chapter is the beginning of Part II of the Syllabus. It deals with Fourier integrals.
6.1 In Part I we have seen that, for a 2-periodic function f which is suciently smooth,
for instance f (
2
2
, we can write
f(a) =

n=
_
1
2
_

f(t) e
int
dt
_
e
ina
(a R), (6.1)
where the doubly innite series converges absolutely. Now we will consider functions f on
R which are usually not periodic. The analogue of (6.1), valid for suciently nice functions
f, is
f(a) =
1
2
_

__

f(t) e
itx
dt
_
e
ixa
dx (a R). (6.2)
If f is a function on R which is suciently smooth (say C
2
) and which decreases suciently
fast in absolute value to 0 as t (say f(t) = O([t[
2
) as [t[ ), then (6.2) holds
with absolutely convergent inner and outer integrals. This we will prove later. For the
moment we will only prove (6.2) heuristically from (6.1).
6.2 Let f be a C
2
-function on R with compact support and let a R. We will give a
heuristic proof of (6.2). For > 0 such that supp(f) [, ] and [a[ dene a
function f

(
2
2
such that f

(t) := f(t) if t . It follows from (6.1) that

f(a) = f

(a/) =
1
2

n=
__

f(t) e
int
dt
_
e
ina/
=
1
2

n=
__

f(t) e
int/
dt
_
e
ina/

1
=
1
2

x
1
Z
__

f(t) e
ixt
dt
_
e
ixa

1
=
1
2
lim
N

x
1
Z[N,N]

f(x) e
ixa

1
, (6.3)
where we have put

f(x) :=
_

f(t) e
ixt
dt (x R). (6.4)
The sum in (6.3) can be considered as a Riemann sum for the partition of [N, N] into
equidistant points with distance
1
. Since the function x

f(x) e
ixa
is continuous, this
Riemann sum tends to the corresponding integral. So formally the limit of (6.3) as
equals
1
2
lim
N
_
N
N

f(x) e
ixa
dx,
which is equal to the right hand side of (6.2). It has to be justied that we may take the
limit for in (6.3) inside the limit for N . We will skip this here.
GENERALITIES 41
6.3 Denition Let f L
1
(R). Then the integral on the right hand side of (6.4) is
absolutely convergent since [f(t) e
ixt
[ [f(t)[. So (6.4) yields a well-dened function

f
on R. We call the transform f

f given by (6.4) the Fourier transform. The function

f
is called the Fourier transform of f. Note that

f only depends on the equivalence class of
f. So the function

f is well-dened for f L
1
(R).
Ex. 6.4 Let [a, b] be a closed bounded interval and let f :=
[a,b]
be its characteristic
function. Show the following: f L
1
(R) and

f(x) =
_
_
_
e
iax
e
ibx
ix
=
2 sin(
1
2
(b a)x) e

1
2
i(a+b)x
x
if x ,= 0,
b a if x = 0.
Observe from this result that

f is continuous and that lim
x

f(x) = 0.
Ex. 6.5 Let f(x) =
1
1+x
2
on R. Compute

f. (Use Contour integration.)
6.6 Let p [1, ), in particular we will consider p = 1 or 2. We write
|f|
p
:=
__

[f(t)[
p
dt
_
1/p
if f L
p
(R) or f L
p
(R),
and we put
|f|

:= sup
xR
[f(x)[ if f is a bounded continuous function on R.
Let C
c
(R) denote the space of complex-valued continuous functions on R with compact
support. The following Proposition is proved in Rudin, Theorem 3.14.
6.7 Proposition Let p [1, ). Then C
c
(R) is dense in L
p
(R).
Ex. 6.8 A function f: R C is called simple if it is a linear combination of characteristic
functions of bounded intervals. Equivalently, f is simple if there are real numbers a
1
<
a
2
< . . . < a
n
such that f vanishes outside [a
1
, a
n
] and f is constant on each interval
(a
i
, a
i+1
) (i = 1, . . . , n 1).
Let p [1, ). Prove, by using Proposition 6.7, that the space of simple functions is dense
in L
p
(R).
6.9 Theorem Let f L
1
(R). Then
(a)

f is a bounded continuous function on R.
(b) |

f|

|f|
1
.
(c) lim
x

f(x) = 0 (Riemann-Lebesgue Lemma).
42 CHAPTER 6
Proof For the proof of the continuity of

f note that, for x R,
lim
yx
_

f(t) e
iyt
dt =
_

_
lim
yx
f(t) e
iyt
_
dt =
_

f(t) e
ixt
dt
by Lebesgues dominated convergence theorem (syll. Integratietheorie) since [f(t) e
iyt
[
[f(t)[ and
_

[f(t)[ dt < . Clearly, it follows from (6.4) that

[

f(x)[
_

[f(t)[ dt = |f|
1
(x R).
This settles (a) and (b).
For the proof of (c) rst note that the result is valid if f is a simple function. This
follows from Exercise 6.4. Now let f L
1
(R). Let > 0. By Exercise 6.8 there is a simple
function g such that |f g|
1
< /2. By (b) we have that [

f(x) g(x)[ < /2. Hence

[

f(x)[ /2 +[ g(x)[ and there is M > 0 such that [ g(x)[ < /2 if [x[ > M.
Ex. 6.10 Let
C
0
(R) := f: R C [ f is continuous and lim
x
f(x) = 0.
Prove that C
0
(R) is a closed linear subspace of the Banach space of bounded continuous
functions on R with respect to the sup-norm, so C
0
(R) is a Banach space itself with respect
to the sup-norm. Also observe that the map f

f: L
1
(R) C
0
(R) is a bounded linear
operator.
Ex. 6.11 Let f L
1
(R), a R, c R0. Prove the following.
(a) If g(t) := f(t +a) then g(x) = e
iax

f(x).
(b) If g(t) := e
iat
f(t) then g(x) =

f(x a).
(c) If g(t) := f(ct) then g(x) = [c[
1

f(c
1
x).
(d) In particular, if g(t) := f(t) then g(t) =

f(t).
6.12 Proposition Let f, g L
1
(R). Then
_

f(x) g(x) dx =
_

f(x) g(x) dx.

(Both integrals converge absolutely because of Theorem 6.9(a).)
Proof We have to show that
_

x=
__

y=
f(x) e
ixy
g(y) dy
_
dx
is equal to
_

y=
__

x=
f(x) e
ixy
g(y) dx
_
dy.
GENERALITIES 43
This follows from the Fubini Theorem 2.14 since
_

x=
__

y=
[f(x) e
ixy
g(y)[ dy
_
dx < .
6.13 Proposition Let f C
1
(R) such that f, f

L
1
(R). Then
(f

)(x) = ix

f(x), (6.5)
and

f(x) = o([x[
1
) as x .
Proof Let M > 0. Integration by parts shows that
_
M
M
f

(t) e
ixt
dt = e
ixM
f(M) e
ixM
f(M) +ix
_
M
M
f(t) e
ixt
dt.
Then (6.5) will follow by letting M in the above formula, provided we can show
that lim
M
f(M) = 0. These last limits are indeed valid. For instance, f(M) =
f(0) +
_
M
0
f

(t) dt. Hence f() := lim

M
f(M) = f(0) +
_

0
f

(t) dt exists. Then

necessarily f() = 0, since otherwise
_

[f(t)[ dt = .
The second statement follows by combination with Theorem 6.9(c).
6.14 Proposition Let f L
1
(R) such that the function g: x xf(x) is in L
1
(R).
Then

f is dierentiable on R with (

f )

= i g.
Proof Let a, b R with a < b. Then, by Proposition 6.12 and Exercise 1.4,
_
b
a
g(x) dx =
_

g(x)
[a,b]
(x) dx
=
_

g(x) (
[a,b]
)(x) dx
=
_

xf(x)
e
iax
e
ibx
ix
dx
= i
_

f(x)(e
ibx
e
iax
) dx = i(

f(b)

f(a)).
Hence

f(a +h)

f(a)
h
= ih
1
_
a+h
a
g(x) dx.
Now let h 0 and use the fact that g is continuous.
6.15 Corollary
(a) Let f C
k
(R) such that f, f

, . . . , f
(k)
L
1
(R). Then (f
(k)
)(x) = (ix)
k

f(x) and

f(x) = o([x[
k
) as x .
(b) Let f be a Lebesgue measurable function on R such that
_

(1 +[t[)
k
[f(t)[ dt < .
Then

f C
k
(R) and (

f )
(k)
(x) =
_

(it)
k
f(t) e
ixt
dt.
(c) Let R with > 1. Let f be a Lebesgue measurable function on R such that
f(x) = O([x[

) for [x[ . Then

f C
k
(R) if k + 1 < .
44 CHAPTER 6
Observe the general rules:
Faster decrease of [f(t)[ to 0 as [t[ implies more smoothness of

f.
More smoothness of f implies faster decrease of [

f(x)[ to 0 as [x[ .
6.16 We introduce the following notation for some operators acting on functions on R:
(Df)(x) := f

(x) (dierentiation),
(Mf)(x) := x f(x) (multiplication),
(T
a
f)(x) := f(x +a) (translation),
(E
a
f)(x) := e
iax
f(x) (exponential multiplication),
(Pf)(x) := f(x) (parity operator).
Some of the previous results can now be summarized in the following table:
g g
Df iM

f
Mf iD

f
E
a
f T
a

f
T
a
f E
a

f
t f(ct) x [c[
1

f(c
1
x)
Pf P

f
6.17 Denition o is the set of all f C

(R) such that, for all nonnegative integers

n, m the function x x
n
f
(m)
(x) is bounded.
So o consists of the C

-functions f on R such that f and all its derivatives are

rapidly decreasing to 0 as the argument tends to . Briey we call such functions
rapidly decreasing C

-functions on R. This class of functions was introduced by Laurent

Schwartz. He used this as a class of test functions for his so-called tempered distributions.
The (immediate) proofs of the following statements are left to the reader.
6.18 Proposition o is a linear space. If f o then the functions Df, Mf, E
a
f, T
a
f,
x f(cx) and Pf are also in o. If f o then

f o.
Ex. 6.19 Let f(t) := e

1
2
t
2
(the Gaussian function). Show the following:
(a) f o.
(b) Df = Mf, D

f = M

f.
(c)

f(x) = C e

1
2
x
2
with C =
_

1
2
x
2
dx =

2.
Hint Solve the dierential equation for

f in (b).
Ex. 6.20 Let f(t) := e

1
2
t
2
. Then

f(x) =
_

1
2
t
2
e
ixt
dt = e

1
2
x
2
_

1
2
(t+ix)
2
dt (x R).
Derive from this formula the result of Exercise 6.19(c), by contour integration.
GENERALITIES 45
Ex. 6.21 Let f L
1
(R). Suppose that f has compact support, i.e., f vanishes outside
some bounded interval. Prove the following.
(a) If

f = 0 then f = 0 a.e..
Hint Reduce this to the case of periodic functions.
(b) The function

f is the restriction to R of an entire analytic function on C (the function
given by (6.4) for x C).
(c) If

f also vanishes on R outside some bounded interval then f = 0 a.e..
6.22 Denition-Theorem For f, g L
1
(R) the convolution product f g is dened
by
(f g)(x) :=
_

f(y) g(x y) dy (x R). (6.6)

Then the integral in (6.6) converges absolutely for almost all x R and f g L
1
(R).
Furthermore,
|f g|
1
|f|
1
|g|
1
(6.7)
and
(f g)(x) =

f(x) g(x). (6.8)
Proof Analogous to the proof of Denition-Theorem 2.19. Use the Fubini Theorem 2.14.
Ex. 6.23 Let p(x) be a polynomial with real coecients in the real variable x. Find
necessary and sucient conditions in order that the function x e
p(x)
belongs to o.
46
7 Inversion formula
7.1 (Dirichlet kernel) Let f L
1
(R), > 0. Put
(I

[f])(s) :=
1
2
_

f(x) e
ixs
dx (s R). (7.1)
In view of (6.2) we expect (but still have to prove) that I

[f] f as . The
advantage of (7.1) compared to the corresponding integral from to is that it can be
written as an integral transformation (of convolution type) acting on f with very simple
integral kernel: the Dirichlet kernel
sin(t)
t
=
1
2
_

e
ixt
dx. (7.2)
Proposition We have
(I

[f])(s) =
_

f(t +s)
sin(t)
t
dt, (7.3)

sin(t)
t

/ (t R), (7.4)
_

sin(t)
t
dt =
2

_

0
sin t
t
dt = 1 (7.5)
(the last two integrals converging but not absolutely converging).
The proof of these formulas is left to the reader. For the proof of (7.3) substitute (6.4)
into (7.1) and interchange the two integrals (allowed by Fubinis theorem; check this).
7.2 The following results are analogous to 3.2. We give a convergence criterium for
Fourier inversion in a certain point.
Theorem Let f L
1
(R), a R. Suppose that one of the following two conditions holds:
(a) There are M, , > 0 such that
[f(a +x) f(a)[ < M [x[

for [x[ < .

(b) f is continuous in a and also right and left dierentiable in a.
Then we have:
lim

(I

[f])(a) = f(a). (7.6)

Proof Condition (b) implies condition (a), so we may assume Condition (a). We have
(I

[f])(a) f(a)
_
1
1
sin(x)
x
dx
=
_
1
1
f(a +x) f(a)
x
sin(x) dx +
_
|x|>1
f(a +x)
x
sin(x) dx. (7.7)
INVERSION FORMULA 47
Since
_
1
1
sin(x)
x
dx =
_

siny
y
dy
2

_

0
sin y
y
dy = 1 as ,
it is sucient to prove that the right hand side tends to 0 as . However, we will
see that the right hand side is of the form
1
_

g(x) sin(x) dx with g L

1
(R), so it
tends to 0 as because of the Riemann-Lebesgue lemma (see Theorem 6.9(c)).
The claim about the right hand side is proved by splitting it up as a sum of three
terms (where we assume that < 1):
[x[ < : [g(x)[ =

f(a +x) f(a)

x

< M [x[
1
,
[x[ < 1: [g(x)[ =

f(a +x) f(a)

x

1
([f(a +x)[ +[f(a)[),
[x[ 1: [g(x)[ =

f(a +x)
x

[f(a +x)[.
7.3 Theorem Let f L
1
(R), let a R, and suppose that the limits
f(a
+
) := lim
xa
f(x), f(a

) := lim
xa
f(x)
exist. Suppose that one of the two following conditions holds:
(a) There are M, , > 0 such that
[f(a +x) f(a
+
)[ Mx

and [f(a x) f(a

)[ Mx

for 0 < x < .

(b) f is right and left dierentiable in a in the sense that the following two limits exist:
f

(a
+
) := lim
y0
f(a +y) f(a
+
)
y
, f

(a

) := lim
y0
f(a y) f(a

)
y
.
Then we have:
lim

(I

[f])(a) =
1
2
(f(a
+
) +f(a

)).
Ex. 7.4 Prove Theorem 7.3 as a corollary of Theorem 7.2, analogous to the proof of
Theorem 3.5.
7.5 Corollary (localization principle) Let f, g L
1
(R), let a R, and suppose
that f(x) = g(x) for x in a certain neighbourhood of a. Then precisely one of the following
two alternatives holds for the behaviour of (I

[f])(a) and (I

[g])(a) as :
(a) Both limits exist and are equal:
lim

(I

[f])(a) = lim

(I

[g])(a)
(b) Neither (I

[f])(a) nor (I

[g])(a) has a limit as .

48 CHAPTER 7
Thus, if for some f L
1
(R) and some a R the inversion formula (7.6) holds, then
this inversion formula remains valid if f is perturbed within L
1
(R) such that f remains
unchanged on some neighbourhood of a.
Ex. 7.6 Prove Corollary 7.5.
7.7 Corollary Let f L
1
(R) such that also

f L
1
(R) and f is dierentiable on R.
Then (

f )

is well-dened and
(

f )

(s) = 2 f(s) (s R).

Ex. 7.8 Prove Corollary 7.7. Can the dierentiability assumption be relaxed?
Ex. 7.9 (van Rooij, Voorbeeld 15.8) Dene f: R R by
f(x) :=
_
1 [x[ if [x[ 1,
0 if [x[ > 1.
(a) Show that f L
1
(R) and

f(x) = 2x
2
(1 cos x) (x ,= 0),

f(0) = 1.
(b) Show that

f L
1
(R) and that (

f )

(a) = 2 f(a) for all a R.

(c) Show that
_

1 cos x
x
2
dx = .
Ex. 7.10 Let
f(t) :=
_
_
_
1
2
t if 0 < t
1
2
,

1
2
t if
1
2
t < 0,
0, otherwise.
Compute

f(x). Examine the behaviour of

f(x) as [x[ .
Ex. 7.11 Let f be continuously dierentiable on R0. Suppose that f, f

L
1
(R) and
that the following four limits exist:
a := lim
t0
f(t), b := lim
t0
f(t), lim
t0
f(t) a
t
, lim
t0
f(t) b
t
.
(a) Show that

f(x) = O([x[
1
) as [x[ .
(b) Show that

f(x) = o([x[
1
) as [x[ i a = b.
Ex. 7.12 Let
f(t) :=
_
(1 t
2
)

1
2
if [t[ < 1,
0 if [t[ 1.
Assume that Re >
1
2
. Show that

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

k=0
(x
2
/4)
k
k! ( + 1)
k
,
where ( + 1)
k
:= ( + 1)( + 2) . . . ( +k).
What can be concluded about the speed with which [

f(x)[ tends to 0 as [x[ ?

49
8 L
2
theory
We will work now with the space L
2
(R) of square integrable functions on R and with the
Hilbert space L
2
(R) of equivalence classes of such functions. Inner product and norm are
given by
< f, g >:=
_

f(x) g(x) dx, |f|

2
:= (f, f))
1
2
=
__

[f(x)[
2
dx
_1
2
. (8.1)
The space o (see Denition 6.17) is a linear subspace of L
2
(R).
We will now denote by T a slightly renormalized version of the Fourier transform
f

f:
(Tf)(x) :=
1

f(x) =
1

2
_

f(t) e
ixt
dt. (8.2)
Here f L
1
(R), in particular we will take f o.
8.1 Proposition The map T: o o is a linear bijection, with inverse
(T
1
g)(s) = (Tg)(s) (g o). (8.3)
Furthermore T preserves the inner product (8.1) on o:
Tf, Tg) = f, g) (f, g o). (8.4)
For convenience we give also more extended versions of (8.3) and (8.4):
(T
1
g)(s) =
1

2
_

g(x) e
isx
dx, (8.5)
_

f(t) g(t) dt =
_

(Tf)(x) (Tg)(x)dx =
1
2
_

f(x) g(x) dx. (8.6)

The proof of this Proposition is immediate, in view of Corollary 7.7, Proposition 6.18
and Proposition 6.12.
8.2 We will now introduce some special C

functions, which we will need as tools (see

also van Rooij, 18.1).
Let a < b and put
f(x) :=
_
_
_
exp
_
1
x b

1
x a
_
if a < x < b,
0, otherwise.
Then f C

(R) and f(x) 0 for all x R.

Let
F(x) :=
_
_
x
a
f(t) dt
_
/
_
_
b
a
f(t) dt
_
.
Then F C

(R), F(x) = 0 if x a, F(x) = 1 if x b, and F is strictly increasing on

[a, b].
50 CHAPTER 8
Ex. 8.3 Let [a, b] be a closed bounded interval and
[a,b]
its characteristic function. Show
that
[a,b]
can be approximated arbitrarily closely in L
2
norm by functions in o.
8.4 Proposition The space o is dense in L
2
(R).
Proof By Exercise 6.8 the simple functions are dense in L
2
(R). Next use Exercise 8.3.
8.5 As a corollary of the previous propositions we obtain:
Theorem (Parseval-Plancherel) The map T: o o has a unique extension to a
continuous linear map T: L
2
(R) L
2
(R). This extension is an isometry of Hilbert spaces:
Tf, Tg) = f, g) (f, g L
2
(R)),
it is bijective, and it satises
(T
2
f)(s) = f(s) a.e.
We have obtained two denitions of the Fourier transform, which we will denote by
T
1
resp. T
2
for the moment:
(T
1
f)(x) :=
1

2
_

f(t) e
ixt
dt (f L
1
(R)),
(T
2
f)(x) :=(L
2
) lim
n
T
1
f
n
, where F
n
o and (L
2
) lim
n
f
n
= f.
For f L
1
(R) L
2
(R) it will turn out that T
1
f = T
2
f.
Lemma Let f L
1
(R) L
2
(R). Then there is a sequence (f
n
)

n=1
in o such that
|f f
n
|
1
0 and |f f
n
|
2
0 as n .
Proof Let > 0. There exists M > 1 such that |f
[M,M]
f|
p
< for p = 1, 2. Write
g := f
[M,M]
. Then there exists h o such that |g h|
2
< /(2

M) < . We can even

choose h in this way such that it has support within [2M, 2M] (why?). Thus
|g h|
1
=
_
2M
2M
[g(x) h(x)[ dx 2

M |g h|
2
< .
Hence |f h|
1
< 2 and |f h|
2
< 2.
8.6 Theorem If f L
1
(R) L
2
(R) then T
1
f = T
2
f.
Proof Take a sequence (f
n
)

n=1
in o such that |ff
n
|
1
0 and |ff
n
|
2
0 if n .
Then T
2
f = lim
n
T
1
f
n
in L
2
(R). Since

2 |T
1
f T
1
f
n
|

|f f
n
|
1
, it follows
that (T
1
f)(x) = lim
n
(T
1
f
n
)(x) pointwise. From integration theory we know that a
convergent sequence in L
2
(R) always has a subsequence which converges almost everywhere
to the limit function. Hence the sequence (T
1
f
n
)

n=1
has a subsequence (T
1
f
n
k
)

k=1
such
that
(T
2
f)(x) = lim
k
(T
1
f
n
k
)(x) = (T
1
f)(x) a.e.
L
2
THEORY 51
8.7 Theorem If f L
2
(R) then
(Tf)(x) = (L
2
) lim
M
1

2
_
M
M
f(t) e
ixt
dt,
which can be rewritten as
lim
M
_

(Tf)(x)
1

2
_
M
M
f(t) e
ixt
dt

2
dx = 0.
Proof |f f
[M,M]
|
2
0 if M . Hence |T
2
f T
2
(f
[M,M]
)|
2
0 as M .
Because f
[M,M]
L
1
(R) L
2
(R), it follows that T
2
(f
[M,M]
) = T
1
(f
[M,M]
).
Ex. 8.8 (van Rooij, Opgave 19.A) Show that not L
1
(R) L
2
(R). But show that all
bounded functions in L
1
(R) lie in L
2
(R).
Ex. 8.9 (van Rooij, Voorbeeld 19.9) Use the Fourier transform computed in Exercise
7.9 in order to show that
_

_
sin t
t
_
4
dt =
2
3
.
Ex. 8.10 (van Rooij, Opgave 19.B) Let f, g L
2
(R). Then fg L
1
(R), thus

fg
exists. Take representatives in L
2
(R) of

f, g L
2
(R) and denote these representatives also
by

f, g. Show that
2

fg(x) =
_

f(t) g(x t) dt.

Ex. 8.11 (van Rooij, Opgave 19.E) For n a nonnegative integer dene the function
h
n
o and the Hermite polynomial H
n
by
h
n
(x) := (x d/dx)
n
e
x
2
/2
, H
n
(x) := e
x
2
/2
h
n
(x).
Prove the following.
(a) We have
H
n+1
(x) = 2x H
n
(x) H

n
(x) and H
n
(x) = (1)
n
e
x
2
(d/dx)
n
e
x
2
.
H
n
is a polynomial of degree n with real coecients; the term of degree n in H
n
has
coecient 2
n
.
(b)

h
n
= (i)
n

2 h
n
. (Thus h
n
is an eigenfunction of the Fourier transform L
2
(R)
L
2
(R).)
(c) If n ,= m then h
n
, h
m
) = 0. (So the functions h
n
form an orthogonal system in
L
2
(R).)
Hint Let n > m. Substitute h
n
(x) = (1)
n
e
x
2
/2
(d/dx)
n
e
x
2
, h
m
(x) = e
x
2
/2
H
m
(x),
and perform integration by parts.
52 CHAPTER 8
(d) If f L
2
(R) and f, h
n
) = 0 for all n, then f(x) = 0 a.e.. (So the functions h
n
form
a complete orthogonal system in L
2
(R).)
Hint For each polynomial P we have
_

f(x) e
x
2
/2
P(x) dx = 0. By the domi-
nated convergence theorem this implies that
_

f(x) e
x
2
/2
e
ixy
dx = 0.
Ex. 8.12 (For this exercise use results from both Fourier series and Fourier integrals.)
Below dene x
1
sin x for x = 0 by continuity.
(a) Let t R. Show that for each x (, ) we have

n=
sin((t n))
(t n)
e
inx
= e
ixt
with pointwise convergence. What is the evaluation of the sum on the left hand side
for other real values of x?
(b) Show that, for all n, m Z, we have
_

sin((t n))
(t n)
sin((t m))
(t m)
dt =
n,m
,
where the integral converges absolutely.
(c) Does there exist f L
2
(R) with f ,= 0 such that
_

f(t)
sin((t n))
(t n)
dt = 0 for all n Z ?
(d) Let f L
2
([, ]). Extend

f, dened as a function on Z by (1.8), to a function on
R by

f(t) :=
1
2
_

f(x) e
ixt
dx (t R).
Show that

f(t) =

n=

f(n)
sin((t n))
(t n)
(t R)
with absolutely convergent sum.
53
9 Poisson summation formula
The Poisson summation formula is an equality of two doubly innite sums over sets of
equidistant points in R. The rst sum involves a function f on R, the second one its
Fourier transform

f.
9.1 Theorem Let f L
1
(R) C(R) such that f(t) = O([t[
2
) as [t[ and

f(x) =
O([x[
2
) as [x[ . (Take for instance f o.) Then

l=
f(x + 2l) =
1
2

k=

f(k) e
ikx
(x R), (9.1)
where the series on both sides pointwise and absolutely converge to a function in C
2
(i.e.,
to a 2-periodic continuous function).
Proof Put
g
N
(x) :=
N

l=N
f(x + 2l), g(x) := lim
N
g
N
(x) =

l=
f(x + 2l).
The last series converges pointwise and absolutely on R because f(x) = O([x[
2
) as [x[
. This series even converges uniformly for x in a compact subset of R. Hence g is
continuous on R. Clearly, g is also 2-periodic. Next observe that
[g
N
(x)[

l=
[f(x + 2l)[ and
_

l=
[f(x + 2l)[
_
dx =
_

[f(y)[ dy < .
This justies an application of the dominated convergence theorem as follows (k Z):
_

g(x) e
ikx
dx = lim
N
_

g
N
(x) e
ikx
dx
= lim
N
_
(2N+1)
(2N+1)
f(y) e
iky
dy =
_

f(y) e
iky
dy =

f(k).
Hence, because

f(k) = O([k[
2
) as [k[ , we conclude that g(x) equals the right hand
side of (9.1).
Specialization of (9.1) yields

l=
f(2l) =
1
2

k=

f(k). (9.2)
Here f is as in Theorem 9.1 and there is absolute convergence on both sides.
54 CHAPTER 9
Ex. 9.2 Let f be as in Theorem 9.1. Assume c > 0. Prove the following identities.
(a)

l=
f(x + 2c l) =
1
2c

k=

f(c
1
k) e
ikc
1
x
,
(b)

l=
f(2c l) =
1
2c

k=

f(c
1
k),
(c)

l=
e

1
2
(x+2c l)
2
=
1
c

k=
e

1
2
c
2
k
2
e
ikc
1
x
,
(d)

l=
e
(x+c l)
2
= c
1

k=
e
c
2
k
2
e
2ikc
1
x
,
(e)

l=
e
c
2
l
2
= c
1

k=
e
c
2
k
2
.
Ex. 9.3 Show that (d) and (e) of Exercise 9.2 remain valid if c, x C with [ arg c
2
[ < /2.
One of the theta functions is dened by

3
(z [ ) :=

n=
e
in
2
e
2inz
(Im > 0, z C).
Use (d) of Exercise 9.2 in order to show that

1
2
exp
_
z
2
/()
_

3
(
1
z [
1
i) =
3
(z [ i) (Re > 0, z C).
Ex. 9.4 (van Rooij, Opgave 14.A) Show that

nZ
1
n
2
+
2
=

cosh
sinh
( > 0).
Hint Apply (9.2) with f(x) := e
|x|
.
Ex. 9.5 Apply (9.2) with f as in Exercise 7.10 in order to show that

m=
1
(x +m)
2
=

2
(sinx)
2
(x RZ).
55
10 Some applications of Fourier integrals
10.1 Heisenbergs inequality
10.1 Lemma Let V be a complex vector space with hermitian inner product . , . ) and
with linear operators A, B: V V such that Af, g) = f, Ag) and Bf, g) = f, Bg)
for all f, g V . Then

1
2i
(AB BA)f, f)

|Af| |Bf| (f V ). (10.1)

For given f V equality holds in (10.1) i Af = iBf for some R or Bf = 0.
Proof Observe that

1
2i
(AB BA)f, f) =
1
2i
_
Bf, Af) Af, Bf)
_
= Im
_
Af, Bf)
_
and

Im
_
Af, Bf)
_

Af, Bf)

|Af| |Bf|.
10.2 We now discuss a special case of the above lemma. Take V := o, Af := f

,
(Bf)(x) := ixf(x). Then
1
2i
_
(AB BA)f
_
(x) =
1
2
f(x).
Hence
1
2

1
|f|
2
__

x
2
[f(x)[
2
dx
_1
2
1
|

f |
2
__

y
2
[

f(y)[
2
dy
_1
2
(f o). (10.2)
For given f o equality holds in (10.2) i f

(x) = x f(x) for some R, i.e., i

f(x) = const. e

1
2
x
2
for some > 0. (For 0 the function f is not in o.)
As a corollary of (10.2) we nd that for f o, x
0
, y
0
R the following inequality
holds:
1
2

1
|f|
2
__

(x x
0
)
2
[f(x)[
2
dx
_1
2
1
|

f |
2
__

(y y
0
)
2
[

f(y)[
2
dy
_1
2
. (10.3)
For xed f o the right hand side of (10.3) attains its absolute minimum for
x
0
= (f) :=
_

x [f(x)[
2
dx
_

[f(x)[
2
dx
, y
0
= (

f ) :=
_

y [

f(y)[
2
dy
_

f(y)[
2
dy
.
Here (f) is the mean value of the stochastic variable x with respect to the probability
distribution
__

[f(x)[
2
dx
_
1
[f(x)[
2
dx on R and (

f ) is the mean value of the stochastic

variable y with respect to the probability distribution
__

f(y)[
2
dy
_
1
[

f(y)[
2
dy on R.
Denote the corresponding variances by
(f) :=
1
|f|
2
__

_
x (f)
_
2
[f(x)[
2
dx
_1
2
,
(

f ) :=
1
|

f |
2
__

_
y (

f )
_
2
[

f(y)[
2
dy
_1
2
.
56 CHAPTER 10
Then we nally obtain Heisenbergs inequality
(f) (

f )
1
2
. (10.4)
Thus the two stochastic variables x and y cannot be concentrated both on arbitrarily small
intervals.
Heisenbergs inequality is better known as Heisenbergs uncertainty relation when in-
terpreted in quantum mechanics or in signal analysis. In quantum mechanics the variables
x and y are the position and impulse operator, respectively, for a one-dimensional particle.
Then position and impulse cannot be both measured with arbitrary accuracy. In signal
analysis x is the time and y is the frequency of a signal. Here time and frequency cannot
both be measured with arbitrary accuracy.
Ex. 10.3 Let f o, |f|
2
= 1, , , a, b > 0, ( + a
2
)( + b
2
) <
1
4
. Show that it is
impossible that the following two inequalities hold together:
_
|x|>a
x
2
[f(x)[
2
dx < and
_
|y|>b
y
2
[

f(y)[
2
< 2.
10.2 Fourier Transform on nite cyclic groups
The impact of the seemingly pure mathematical topic of this section on numerical and
applied mathematics cannot be overestimated. In the present section all relevant groups
will be nite. We can endow such group with the discrete topology (all subsets are open)
and obtain what is called a topological group.
For completeness, we mention here that a topological group is a group endowed with
a topology such that the group operations multiplication and inversion are continuous. In
particular the notion of a continuous function on a topological group is available. Important
examples of topological groups are R
n
and C
n
with the Euclidean topology. We will not
pursue the implications of this notion at all, but we will use the terminology in the following
sense. The denitions will always be for a topological group G, in the results the group
will always be nite. As a consequence, we will speak of continuous mappings of our nite
groups, although this is strictly speaking superuous: all mappings of a discrete topological
space are continuous. However, using the terminology as we do, it will t perfectly in the
general framework of topological and Lie groups.
10.4 Let N N, set
N
= e
2i/N
. We introduce
G
N
=
N
) =
k
N
, k = 1, . . . , N.
Thus G
N
is the multiplicative cyclic subgroup of order N of the circle group T = z C :
[z[ = 1 C. The relative topology on G
N
, inherated from the Euclidean topology on C
is discrete.
Next C(G) denotes the vector space of the continuous complex valued functions on a
topological group G. Of course C(G
N
) may be identied with C
N
.
A character on a group G is a continuous homomorphism from G to T. The set of
characters on G form a group, denoted by

G with multiplication
e
1
e
2
(g) = e
1
(g)e
2
(g), e
i


G, g G.
Ex. 10.5 Show that

G is indeed a group. What is its unit element? Can you describe
the inversion?
APPLICATIONS OF FOURIER INTEGRALS 57
10.6 Lemma

G
N

= G
N
.
Proof Dene : G
N


G
N
by
((
j
N
))(
k
N
) =
kj
N
.
One can check easily that is an injective homomorphism. It is also surjective: If e

G
N
,
then e(
k
N
) = (e(
N
))
k
T for all integers k. Since (e(
N
))
N
= e(id) = 1, we conclude
that e(
N
) =
j
N
for some j 1, .., N. It then follows that e = (
j
N
).
10.7 As C(G
N
)

= C
N
, it can be given the inner product of C
N
. We want the function
1 to have norm 1, therefore we normalize:
F, G) =
1
N

gG
N
F(g)G(g).
We set e
j
= (
j
N
)

G
N
and sometimes we will identify e
j
and
j
N
. We note that

N
=
1
N
.
10.8 Proposition

G
N
is an orthonormal basis of C(G
N
).
Proof Since

G
N
has N elements, which equals the dimension of C(G
N
), it suces to
show that

G
N
is an orthonormal system:
e
l
, e
j
) =
1
N
N

k=1

lk
N

jk
N
=
1
N
N

k=1

(lj)k
N
=
_
1 if l = j;

lj
N
1
(lj)N
N
1
lj
N
= 0 if l ,= j.
The proposition implies the following representation formula for f C(G
N
):
f(
l
N
) =

G
f, e)e(
l
N
) =

f(e)e(
l
N
) =
N

j=1

f(
j
N
)
lj
N
. (10.5)
Here

f is the nite Fourier transform of f. It is as expected, the function on

G
N
dened
by

f(e
j
) = (T
N
f)(e
j
) =
1
N
N

k=1
f(
k
N
)
jk
N
. (10.6)
10.9 For numerical purposes we are interested in the number of computations that are
necessary to compute

f(e
j
). Formula ((10.6)) gives 2N computations to compute

f(e
j
),
assuming knowledge of f and G
N
. Thus we need atmost cN
2
operations to compute T
N
f.
In fact we can do much better.
58 CHAPTER 10
10.3 Fast Fourier Transform
Theorem Suppose N = nm, n, m N and let f C(G
N
). Let F
j
C(G
m
) be
dened by
F
j
(
k
m
) = f(
nk+j
N
), j = 1, . . . , n, k = 1, . . . , m,
with
m
=
n
N
. Then we have

f(
l
N
) =
1
n
n

j=1
T
m
(
lj
N
F
j
)(
l mod m
m
).
Proof In the next computation we set p = nk +j, hence
p
N
=
k
m

j
N
. If p runs from 1
to N, then j runs from 1 to n and k from 1 to m.

f(
l
N
) =
1
N
N

p=1
f(
p
N
)
pl
N
=
1
n
n

j=1
1
m
m

k=1
F
j
(
k
m
)(
k
m

j
N
)
l
=
1
n
n

j=1
1
m
_
m

k=1

jl
N
F
j
(
k
m
)
kl
m
_
=
=
1
n
n

j=1
(T
m

jl
N
F
j
)(
l mod m
m
).
10.10 Thus, to compute

f(
l
N
) we need to compute T
m
F
j
. Suppose it takes A(m)
computations to compute the Fourier transform in G
m
(given F C(G
m
)). To compute

jl
N
F
j
takes m computations, to compute T
m
(
jl
N
F
j
) takes A(m) + m computations.
We need all n of these, yielding n(A(m) +m) computations, the result of which we store.
Finally, for each

f(
l
N
) we need n additions of our stored results. This gives the total of
computations
A(N) = Nn +n(A(m) +m) = N(n + 1) +A(m)n.
10.11 Theorem If N = n
1
n
p
, n
p
= 1, then A(N) N(1 +

i
(n
i
+ 1)).
Proof Using the previous calculation we compute
A(n
1
n
p
) N(n
1
+ 1) +A(n
2
n
p
)n
1
N(n
1
+ 1) +N(n
2
+ 1) +A(n
3
n
p
)n
2
n
1
. . . N

(n
i
+ 1) +N.
10.12 Corollary If N = q
p
we need at most O(N log N) computations to compute
Fourier transform in G
N
. (Here q is xed and p .)
Proof
A(n) N(2
p

j=1
(q + 1) +N) = N(2p(q + 1) + 1).
Note that p = log N/ log q, so that
A(N) 2N
log N
log q
(q + 1) +N = O(N log N).
APPLICATIONS OF FOURIER INTEGRALS 59
10.13 Next we indicate the relation of this nite Fourier transform with ordinary Fourier
transform on the line. Suppose f has compact support on [N, N] and the values of f
at the points j/K are known, j J = KN, KN + 1, . . . , KN 1. Then we can
approximate

f(x) as follows:

f(x) =
_
N
N
f(t)e
ixt
dt =
1
K
_
NK
NK
f(
s
K
)e
i
x
K
s
ds
1
K

jJ
f(
j
K
)e
ixj
K
.
We substitute x =
p
N
and nd

f(
p
N
)
1
K

jJ
f(
j
K
)e
i
2pj
2NK
=
2N
2KN

j
Jf(
j
K
)
pj
M
= 2N(T
M
g)(
p
M
).
Here we have written M = 2KN and dened g C(G
M
) by g(
j
M
) = f(j/K), (j J).
Thus we have found a way to compute the (approximate) Fourier transform of a
function of which the value is known only in nitely many points. Using the fast way
of computing it via Fourier transform on nite groups as we described here is called Fast
Fourier Transform. The paper of Cooley and Tuckey on this topic is in fact in the top ten
of most cited mathematical papers.
Ex. 10.14 Convolution on G
N
For f, g C(G
N
) the convolution f g is dened as
follows:
f g(
j
N
) =
1
N
N

k=1
f(
k
N
)g(
jk
N
).
Show that T
N
(f g)(
l
N
) = T
N
f(
l
N
)T
N
g(
l
N
).
Also show that f g = g f and f (g h) = (f g) h.
Ex. 10.15 Fast multiplication Let F, G be polynomials of degree, p, q respectively.
View F and G as elements of C(G
N
) with N = p +q +1, by declaring F(
k
N
) equal to the
coecient of z
k
of F; the same goes for G.
Give a rough estimate for the number of computations needed to compute F G
directly, and next, by observing that F G as element of C(G
N
) is the convolution of F
and G and using fast Fourier transform.
Apply this and describe a fast way for computing the product of two very large (10000-
digit) numbers.