Fourier Analysis

Tom Leinster

2013–14
Contents

A Preparing the ground 2

A1 The algebraist’s dream . . . . . . . . . . . . . . . . . . . . . . . . 2
A2 Pseudo-historical overview . . . . . . . . . . . . . . . . . . . . . . 5
A3 Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
A4 What does it mean for a sequence of functions to converge? . . . 12
A5 Periodic functions . . . . . . . . . . . . . . . . . . . . . . . . . . 16
A6 The inner product . . . . . . . . . . . . . . . . . . . . . . . . . . 18
A7 Characters and Fourier series . . . . . . . . . . . . . . . . . . . . 20
A8 Trigonometric polynomials . . . . . . . . . . . . . . . . . . . . . . 22
A9 Integrable functions are sort of continuous . . . . . . . . . . . . . 24

B Convergence of Fourier series in the 2- and 1-norms 28

B1 Nothing is better than a Fourier partial sum . . . . . . . . . . . . 28
B2 Convolution: definition and examples . . . . . . . . . . . . . . . . 31
B3 Convolution: properties . . . . . . . . . . . . . . . . . . . . . . . 33
B4 The mythical delta function . . . . . . . . . . . . . . . . . . . . . 36
B5 Positive approximations to delta . . . . . . . . . . . . . . . . . . 38
B6 Summing the unsummable . . . . . . . . . . . . . . . . . . . . . . 41
B7 The Fejér kernel . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
B8 The main theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 47

C Uniform and pointwise convergence of Fourier series 49

C1 Warm-up . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
C2 What if the Fourier coefficients are absolutely summable? . . . . 52
C3 Continuously differentiable functions . . . . . . . . . . . . . . . . 53
C4 Fejér’s theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
C5 Differentiable functions and the Riemann localization principle . 58
C6 Weyl’s equidistribution theorem . . . . . . . . . . . . . . . . . . . 61

D Duality 66
D1 An abstract view of Fourier analysis . . . . . . . . . . . . . . . . 66
D2 The dual of a finite abelian group . . . . . . . . . . . . . . . . . . 70
D3 Fourier transforms on a finite abelian group . . . . . . . . . . . . 73
D4 Fourier inversion on a finite abelian group . . . . . . . . . . . . . 77

1
Chapter A

Preparing the ground

A1 The algebraist’s dream

For the lecture of 13 January 2014

The algebraist thinks: ‘Analysis is hard. Can we reduce it to algebra?’

Idea Use power series.

1 Many functions f : R → C can be written as a power series,

∞
X
f (x) = cn xn . (A:1)
n=0

If we can deal with the sequence (cn ) rather than the function f , everything
will be much easier (and more algebraic).
2 If we can express f in this way, then the coefficients cn must be given by

f (n) (0)
cn = .
n!
(Proof: differentiate each side of (A:1) n times, then evaluate at 0.)

cn xn might not converge. Or, it might converge, but not to f (x).

P
3 Problem:
E.g. put ( 2
e−1/x if x 6= 0,
f (x) =
0 if x = 0

(Figure A.1). It can be shown that f (n) (0)/n! = 0 for all n (which you
might
P∞ guess from the flatness of the function near 0). But of course f (x) 6=
n
n=0 0 · x except at x = 0. This brings to light:

4 Another problem: different infinitely differentiable functions can have the

same power series. (E.g. this is the case for the f just defined and the
constant function 0.)

2
 1
exp(-1/(x**2))

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
-4 -2 0 2 4

2
Figure A.1: The function x 7→ e−1/x

5 Yet another problem: the power series method only captures infinitely differ-
entiable functions. (For recall that every power series is infinitely differen-
tiable inside its disk of convergence.) However, many interesting functions
are not infinitely differentiable.
6 In this power series approach to functions, the number 0 is given a special
role. This isn’t necessarily good or bad, but P
might seem a little suspicious.
We could equally well consider power series cn (x − a)n centred at a, for
any other value of a.
P f (n) (0) n
7 Also suspicious: the power series n! x of f depends only on the values
of f near 0. In the jargon, it’s locally determined. You can’t reasonably
expect to predict the value of f for large |x| given only the value of f for
small |x|.
In other words, if you had two functions f and g that were the same
throughout the interval (−δ, δ), then they’d have the same power series.
However, they might have very different values outside that interval.

So for various reasons, the algebraist’s dream can’t be realized using power
series. However. . .

Alternative idea Use Fourier series.

1 Many 1-periodic functions f : R → C can be written as a Fourier series:

∞
X ∞
X
f (x) = an cos 2πnx + bn sin 2πnx
n=0 n=0

(an , bn ∈ C). 1-periodic means that f (x + 1) = f (x) for all x ∈ R.

(Clearly any function expressible in this way is 1-periodic. You might be
more used to dealing with 2π-periodic functions and Fourier series involv-
ing the terms cos nx and sin nx, but the difference is purely cosmetic.)
Since
cos θ = (eiθ + e−iθ )/2, sin θ = (eiθ − e−iθ )/2i,

3
 we can rewrite this more efficiently as
∞
X
f (x) = ck e2πikx
k=−∞

for certain ck ∈ C. (Note that the sum starts at −∞.)

2 If we can express f in this way, then the coefficients ck must be given by
Z 1
ck = f (x)e−2πikx dx.
0

(We’ll see why later, or you can try proving it now.)

ck e2πikx might not converge. Or, it might converge, but not to
P
3 Problem:
f (x). (There are examples of both these phenomena.)
4 But unlike for power series, different continuous functions always have differ-
ent Fourier series.
5 And unlike for power series, functions of many kinds can be captured using
Fourier series. Far from having to be infinitely differentiable, even some
discontinuous functions can be captured.

6 And again unlike for power series, in the definition of ck , no point in R is

given a special role.
ck e2πikx of f
P
7 And once more unlike for power series, the Fourier series
R1 −2πikx
depends on all of f (since ck = 0 f (x)e dx, and the periodic func-
tion f is determined by its restriction to [0, 1]). In the jargon, it’s globally
determined. So it’s not unreasonable to expect to be able to predict the
value of f for all x given only the Fourier coefficients ck . We’ll see that
we often can.

The algebraist’s dream can’t be fully realized. But Fourier series come closer
to realizing it than power series do, at least for periodic functions on R. They’re
not as obvious an idea as power series, but in many ways they work better.
We’ll spend a lot of the course exploring the extent to which functions can be
understood in terms of their Fourier series. There are many analytic subtleties,
which we’ll have to think hard about.
The development of Fourier theory has been very important historically. It
has been the spur for a lot of important ideas in mathematics, not all obviously
connected to Fourier analysis. We’ll meet some along the way.

4
 Figure A.2: Excerpt from the index of Tom Körner’s book Fourier Analysis

A2 Pseudo-historical overview
For the lecture of 16 January 2014

Most mathematicians are terrible historians. They can’t resist recounting what
should have happened, not what did happen. I can’t claim to be any better—
hence the ‘pseudo’ of the title.
This lecture is all about the ‘excessive optimism’ and ‘excessive pessimism’
mentioned in the index of Körner’s book (Figure A.2).
A 1-periodic function on R is determined by its values on [0, 1) (or any
other interval of length 1). We say that a 1-periodic function is integrable if
its restriction to [0, 1) is integrable in the usual sense. Let f : R → C be an
integrable 1-periodic function.
For k ∈ Z, the kth Fourier coefficient of f is
Z 1
fˆ(k) = f (x)e−2πikx dx.
0

The Fourier series of f is the doubly infinite series Sf given by

∞
X
(Sf )(x) = fˆ(k)e2πikx .
k=−∞

It’s not entirely clear what kind of thing Sf is. For instance, it might not
always converge, and it’s not even obvious what convergence should mean. But
the following definition is perfectly safe: for n ≥ 0, the nth Fourier partial
sum is the function Sn f given by
n
X
(Sn f )(x) = fˆ(k)e2πikx .
k=−n

Classical question: for ‘nice’ functions f , is it true that

(Sn f )(x) → f (x) as n → ∞

for all x ∈ [0, 1)? For ‘most’ x ∈ [0, 1)? For some x ∈ [0, 1)?

5
 This is the question of pointwise convergence. There are other kinds of
convergence, as we’ll see, some of which are important in ways that pointwise
convergence is not.
Fourier himself (1768–1830) thought that yes, it’s always true. But he was
imprecise about this (and most other things). It’s not even clear what he would
have taken the word ‘function’ to mean.
To start with a point that was certainly clear to Fourier himself, it can’t
be true for all x and all f . For instance, suppose it’s true for all x for some
particular f . Define g to be the same as f except at one point of [0, 1), where
it takes some different value. Then ĝ(k) = fˆ(k) for all k, so Sn g = Sn f for all
n, so it can’t also be true that (Sn g)(x) → g(x) as n → ∞ for all x.
(For instance, take g to be the 1-periodic function given by g(x) = 1 for
x ∈ Z and g(x) = 0 for x 6∈ Z. Then for all n, the function Sn g has constant
value 0, so (Sn g)(x) 6→ g(x) whenever x ∈ Z.)
Backing up Fourier’s intuition, Dirichlet proved:
Theorem A2.1 (Dirichlet, 1829) Let f : R → C be a 1-periodic, continu-
ously differentiable function. Then (Sn f )(x) → f (x) as n → ∞ for all x ∈ [0, 1)
(or equivalently, for all x ∈ R).

Once Dirichlet had proved that, it was generally believed that ‘infinitely
differentiable’ could be relaxed to ‘continuous’—that it was just a matter of time
before someone figured out a proof that would work in this wider generality.
Most of the most prominent mathematicians of the day believed this. Dirich-
let believed it, Cauchy, believed it, and Riemann, Weierstrass, Dedekind and
Poisson all believed it. Cauchy even claimed he’d proved it. (Standards of
rigour were lower in those days.) Dirichlet promised he’d prove it, but never
did.
But then came a bombshell.
Theorem A2.2 (du Bois-Reymond, 1876) There is a 1-periodic, contin-

∞ f : R → C such that for some x ∈ [0, 1), the sequence
uous function
(Sn f )(x) n=0 fails to converge.

Using this example, it takes comparatively little effort to construct a 1-

periodic continuous function such that ((Sn f )(x))∞
n=0 fails to converge at 10
points, or 100, or 1000:
Theorem A2.3 Let E ⊆ [0, 1) be a finite set. Then there is a 1-periodic,
con-
∞
tinuous function f : R → C such that for all x ∈ E, the sequence (Sn f )(x) n=0
fails to converge.

Widespread pessimism ensued. The pendulum swung from the general con-
sensus that Fourier series behave perfectly for all continuous functions to the
opposite extreme. After du Bois-Reymond’s theorem appeared, it came to be

∞ f : R → C with the
believed that there was some 1-periodic continuous function
property that for all x ∈ [0, 1), the sequence (Sn f )(x) n=0 fails to converge.
This pessimism was reinforced by a further example:
Theorem A2.4 (Kolmogorov, 1926) There exists a 1-periodic, Lebesgue in-
tegrable function

∞ f : R → C such that for all x ∈ [0, 1), the sequence
(Sn f )(x) n=0 fails to converge.

6
 We won’t be doing Lebesgue integrability in this course, and I won’t assume
you know anything about it. But as general cultural background:

Remark A2.5 Every Riemann integrable function is Lebesgue integrable. (In

particular, a Lebesgue integrable function need not be continuous, and Kol-
mogorov’s example certainly wasn’t.) Moreover, the Lebesgue integral of a
Riemann integrable function is equal to its Riemann integral. However, there
are functions that are Lebesgue integrable but not Riemann integrable. So, the
theory of Lebesgue integration extends, and is more powerful than, the Riemann
theory.

This situation persisted until relatively recently. For instance, one of the
books on my own undergraduate reading list was Apostol’s 1957 text Mathe-
matical Analysis. In most respects, it is like today’s textbooks, but he states in
it that it’s still unknown whether for a continuous function, the Fourier series
has to converge at even one point.
The turning point came in the 1960s.

Theorem A2.6 (Carleson, 1964) Let f : R → C be a 1-periodic function.

• If f is continuous then (Sn f )(x) → f (x) as n → ∞ for at least one
x ∈ [0, 1).
• Better still, if f is Riemann integrable then (Sn f )(x) → f (x) as n → ∞
for at least one x ∈ [0, 1).

• Better still, if f is Riemann integrable then (Sn f )(x) → f (x) as n → ∞

for almost all x ∈ [0, 1). (‘Almost all’ will be defined later.)
• Even better still, if |f |2 is Lebesgue integrable (which is true if f is Rie-
mann integrable) then (Sn f )(x) → f (x) as n → ∞ for almost all x ∈ [0, 1).

All known proofs of Carleson’s theorem are hard, still far too hard for a
course such as this.
(I say ‘still’ because most hard proofs of important facts are simplified over
time. This was a big theorem that attracted a lot of attention, so lots of people
have put in lots of work to simplify the proof. They have simplified it, but
it’s still well beyond our reach. This is true even for the very watered-down
version of Carleson’s theorem in the first bullet point: that for merely continuous
functions, there is merely one point at which the Fourier series behaves well.)
Carleson’s theorem is best possible, in a sense that can be made precise
(Theorem A3.18). Roughly, this means that you can’t do any better than ‘al-
most all’. It almost completely answers the question of pointwise convergence of
Fourier series. But there are other types of convergence too, and the behaviour
of Fourier series from those points of view is interesting too, and has provoked
a lot of mathematical developments—as we shall see.

7
A3 Integration
For the lecture of 20 January 2014

We won’t be doing any Fourier analysis for the first couple of weeks. The course
pulls together several strands of analysis, and we’re going to look at them one at
a time before attempting to bring them together. This has the disadvantage of
postponing the moment when we first see any Fourier theory, but the advantage
that we can concentrate on one thing at a time.
For the rest of this section, let I ⊆ R be a bounded interval.
Definition A3.1 i. A function f : I → R is integrable if it is bounded and
Riemann integrable.
ii. Let f : I → C be a function. Writing f1 , f2 : I → R for its real and
imaginary parts (so that

f (x) = f1 (x) + if2 (x)

for all x ∈ R), we say that f is integrable if f1 and f2 are both integrable.
In that case, we define the integral of f by
Z Z Z
f (x) dx = f1 (x) dx + i f2 (x) dx.
I I I

Part (i) is a declaration about how the word ‘integrable’ will be used in this
course. Other people use it differently. For example, a more permissive meaning
(which you don’t have to know anything about) would be ‘Lebesgue integrable’.
Examples A3.2 i. Bounded continuous functions are integrable. (If I is
closed then ‘bounded’ follows automatically from ‘continuous’.)
ii. Let J ⊆ I be an interval. The characteristic function (or indicator
function) of J is the function

χJ : I → C

defined by (
1 if x ∈ J,
χJ (x) =
0 if x ∈
6 J.
R
Then χJ is integrable, and I χJ (x) dx = |J|. Here |J| is the length of
the interval J, defined by |J| = sup J −inf J (or as 0 if J = ∅). Concretely,
if J is [a, b] or [a, b) or (a, b] or (a, b) then |J| = b − a.

Notation A3.3 Let f, g : I → C. We write f + g : I → C for the function

defined by
(f + g)(x) = f (x) + g(x)
(x ∈ I). We define f · g similarly, f¯ by f¯(x) = f (x) (where the bar means
complex conjugate), and |f | by |f |(x) = |f (x)|. Given c ∈ C, we write c : I → C
for the function with constant value c. So, for instance, c · f (or cf ) is the
function I → C given by (cf )(x) = c · f (x).

8
Lemma A3.4 i. If f, g : I → C are integrable then so is f + g, with
Z Z Z
(f + g)(x) dx = f (x) dx + g(x) dx.
I I I

ii. If f : I → C is integrable and c ∈ C then cf is integrable, with

Z Z
(cf )(x) dx = c f (x) dx.
I I

iii. If f : I → C is integrable then so is f¯, with

Z Z
¯
f (x) dx = f (x) dx.
I I

Proof Parts (i) and (ii) follow from the corresponding properties of real-valued
integration, and part (iii) follows directly from the definitions. 

I’ll occasionally state ‘Facts’, without proof. Some of these facts were proved
in previous courses (such as PAA). Others weren’t, and I’ll be asking you to take
those on trust.
Fact A3.5 Let f : I → C be an integrable function and φ : C → C a continuous
function. Then the composite φ ◦ f : I → C is integrable.

Example A3.6 If f is integrable then so is |f |p , for any real p > 0. (Take

φ(z) = |z|p .)

Lemma A3.7 If f, g : I → C are integrable then so is f · g : I → C.

Proof See Sheet 1, q.3. 

(If you’ve come across Lebesgue integration, you may be aware that both
this lemma and Example A3.6 fail for Lebesgue integration. This is one respect
in which the Riemann theory makes life simpler.)
Fact
R A3.8 IfR f, g : I → R are integrable and f (x) ≤ g(x) for all x ∈ I then
I
f (x) dx ≤ I g(x) dx.

Note that this is for functions into R, not C; inequalities makes no sense in
C.

Lemma A3.9 (Triangle inequality for integration) If f : I → C is inte-

grable then |f | is integrable, with
Z  Z
 
 f (x) dx ≤ |f |(x) dx.
 
I I
R
Proof That |f | is integrable is a special case of Example A3.6. Since I
f (x) dx
is a complex number, we may write
Z Z 
 
f (x) dx = u  f (x) dx

I I

9
for some u ∈ C with |u| = 1. Now
Z  Z   Z Z
 f (x) dx = Re  f (x) dx = Re 1 1
    
f (x) dx = Re f (x) dx

I
 
I
 u I I u
Z 1 
= Re f (x) dx (A:2)
u
ZI 
1

≤  f (x) dx (A:3)

u
ZI Z
= |f (x)| dx = |f |(x) dx.
I I

Here Re(z) denotes the real part of a complex number z, equation (A:2) follows
from the definition of complex-valued integration (Definition A3.1(ii)), the in-
equality (A:3) follows from Fact A3.8 and the fact that Re(z) ≤ |z| for all z ∈ C,
and the rest follow either from earlier lemmas or directly from the definitions.

We now come to a very important concept. It was introduced by Lebesgue,

and forms part of his theory of integration; and although we are not studying
that theory, we will need this one particular concept.
Definition A3.10 A subset E of R has measure zero (or is null) if for all
ε > 0, there exists a sequence of intervals (Jn )∞
n=1 in R such that
∞
[ ∞
X
E⊆ Jn and |Jn | ≤ ε.
n=1 n=1

Examples A3.11 i. Any countable set E ⊆ R has measure zero. This is

clear for E = ∅, so assume E 6= ∅. We can choose a sequence (xn )∞ n=1 such
−(n+1) −(n+1)


that E = {x S1 , x2 , . . .}.
P Let ε > 0. Put
P∞ −n Jn = x n −2 ε, xn +2 ε .
Then E ⊆ En and |Jn | = n=1 2 ε = ε.
ii. There also exist uncountable sets of measure zero (such as the Cantor set).
iii. If E has measure zero and F ⊆ E then F has measure zero.
iv. Let J be an interval with |J| > 0. (This condition just means that J is
not empty or of the form {a}.) Then J does not have measure zero. This
is not as obvious as it sounds. If you do think it’s obvious, try proving it!

Definition A3.12 Let S ⊆ R. A property P of points x ∈ S holds almost

everywhere (a.e.), or for almost all x ∈ S, if there is some E ⊆ S of measure
zero such that P holds for all x ∈ S \ E.

For example, if f, g : I → C then

‘f = g a.e.’
and
‘f (x) = g(x) for almost all x’
both mean: there exists a subset E ⊆ I of measure zero such that f (x) = g(x)
for all x ∈ I \ E.
What’s this got to do with integration?

10
 R
Fact
R A3.13 If f, g : I → C are integrable and f = g a.e. then I
f (x) dx =
I
g(x) dx.
R
Fact A3.14 If h : I → R is integrable, h(x) ≥ 0 for all x ∈ I, and I
h(x) dx =
0, then h = 0 a.e..

Proposition A3.15 Let f, g : I → C be integrable functions with f = g a.e..

Assume that |I| > 0.
i. Let x ∈ I. If f and g are both continuous at x, then f (x) = g(x).

ii. If f and g are continuous then f = g.

Proof For (i), suppose for a contradiction that f (x) 6= g(x). By continuity, we
can find some δ > 0 such that f (t) 6= g(t) for all t ∈ (x − δ, x + δ) ∩ I. Also,
since f = g a.e., we can choose E ⊆ I of measure zero such that f (t) = g(t) for
all t ∈ I \ E. Then (x − δ, x + δ) ∩ I ⊆ E, so (x − δ, x + δ) ∩ I has measure zero
by Example A3.11(iii). But |I| > 0, so (x − δ, x + δ) ∩ I is an interval of length
> 0, contradicting Example A3.11(iv).
Part (ii) follows immediately. 

The next result makes no mention of the concept of measure zero.

PropositionRA3.16 Let h : I → R be an integrable function with h(t) ≥ 0 for

all t ∈ I and I h(t) dt = 0. Assume that |I| > 0.
i. Let x ∈ I. If h is continuous at x, then h(x) = 0.
ii. If h is continuous then h = 0.

Proof We could deduce this from Fact A3.14 and Proposition A3.15. This is
unnecessarily complicated, though. Sheet 1, q.4 asks you to find a direct proof.

∗ ∗ ∗

The last two results of this lecture are non-examinable, and are included just
for background.
Fact A3.17 A function f : I → C is integrable if and only if it is bounded and
continuous a.e. (that is, the set {x ∈ I : f is not continuous at x} has measure
zero).

For example, χQ∩[0,1] : [0, 1] → C is not integrable, since the set of disconti-
nuities is [0, 1], which does not have measure zero.
At the end of the last lecture, I said that in a certain sense, Carleson’s
theorem (Theorem A2.6) cannot be improved. Here is the precise statement.
Theorem A3.18 (Kahane and Katznelson, 1960s) Let E ⊆ [0, 1] be a set
of measure zero. Then there is a 1-periodic,
continuous function f : R → C such
∞
that for all x ∈ E, the sequence (Sn f )(x) n=0 fails to converge.

11
Figure A.3: The area between f and g is small, but the largest difference between
them is large.

A4 What does it mean for a sequence of func-

tions to converge?
For the lecture of 23 January 2014

There is no single right answer to this question. Compare the question ‘how big
is a person?’ I might interpret that as a question about height, but you might
interpret it as a question about weight. Neither of us would be right or wrong.
Height and weight are correlated, but not logically related in an absolute sense.
Related to the question of the title is another question: what does it mean
for two functions to be ‘close’ ?
We might say that twoR functions f and g are close if the area between their
graphs is small (that is, |f (x) − g(x)| dx is small). Or, we might say that
they are close if their values are never too different (that is, supx |f (x) − g(x)|
is small). Both ideas are sensible, but they are not the same, as Figure A.3
demonstrates.
For the rest of this section, let I ⊆ R be a bounded interval of length > 0.
Definition A4.1 Let f : I → C be an integrable function. Define:
R
• kf k1 = I |f (x)| dx, the 1-norm of f ;
qR
• kf k2 = I
|f (x)|2 dx, the 2-norm of f ;

• kf k∞ = sup |f (x)|, the ∞-norm or sup-norm of f .

x∈I

Remarks A4.2 i. In the definition of kf k2 , it really makes a difference that

we write |f (x)|2 , not f (x)2 , since f is a complex -valued function.
ii. The definition of kf k∞ is only really appropriate for continuous functions
f . We will mostly stick to continuous functions when we speak of kf k∞ .
iii. The word ‘norm’ is actually not quite right, for reasons that will soon be
explained.
R 1/p
iv. For 1 ≤ p < ∞, put kf kp = I
|f (x)|p dx . It can be shown that
lim kf kp = kf k∞ for continuous f , which explains why it’s called k · k∞ .
p→∞

12
 If we want to refer to the 1-norm in the abstract, we sometimes write it as
k · k1 . The dot is a blank or placeholder, into which arguments can be inserted.
The same goes for k · k2 and k · k∞ .
Lemma A4.3 Let k · k stand for any of k · k1 , k · k2 or k · k∞ . Then for all
integrable functions f, g : I → C and all c ∈ C:
i. kf k ≥ 0;
ii. kcf k = |c| · kf k;
iii. kf + gk ≤ kf k + kgk.
Proof All easy except (iii) for k · k2 , which is Sheet 1, q.6(iv). 
A norm on the set {integrable functions I → C} is an operation satisfying
conditions (i)–(iii) and one further condition: that if kf k = 0 then f = 0. This
further condition fails for k · k1 and k · k2 (as the next example shows), so strictly
speaking we should call them ‘seminorms’. But I will abuse terminology and go
on calling them the 1-norm and 2-norm. (The ∞-norm really is a norm.)
Example A4.4 Define f : [0, 1) → C by
(
1 if x = 0,
f (x) =
0 otherwise.

Then kf k1 = kf k2 = 0 but f 6= 0.
Lemma A4.5 i. For an integrable function f : I → C,
kf k1 = 0 ⇐⇒ f = 0 a.e. ⇐⇒ kf k2 = 0.

ii. For an integrable function f : I → C, if kf k1 = 0 or kf k2 = 0 then

f (x) = 0 for all x ∈ I such that f is continuous at x.
iii. For a continuous function f : I → C,
kf k1 = 0 ⇐⇒ f = 0 ⇐⇒ kf k2 = 0.

iv. For an integrable (or just bounded) function f : I → C,

kf k∞ = 0 ⇐⇒ f = 0.
Proof i. Follows from Facts A3.13 and A3.14.
ii. Follows from Proposition A3.16(i) (taking h = |f | or h = |f |2 ).
iii. Follows from (ii).
iv. Follows from the definition. 
We’ve seen that kf k1 , kf k2 and kf k∞ are three different measures of the
size of a function f .
So, kf − gk1 , kf − gk2 and kf − gk∞ are three different measures of the
distance between functions f and g.
They genuinely measure different things! Look back at the opening para-
graphs of this lecture and Figure A.3.
We now consider the three resulting notions of convergence, plus two more.

13
Definition A4.6 Let (fn ) be a sequence of integrable functions from I to C,
and let f be an integrable function from I to C. We say:

• fn → f in k · k1 if kfn − f k1 → 0 as n → ∞;
• fn → f in k · k2 (or in mean square) if kfn − f k2 → 0 as n → ∞;
• fn → f in k · k∞ (or uniformly) if kfn − f k∞ → 0 as n → ∞;
• fn → f pointwise if for all x ∈ I, fn (x) → f (x) as n → ∞;

• fn → f a.e. if for almost all x ∈ I, fn (x) → f (x) as n → ∞.

All five types of convergence are useful and meaningful.

We now work out the logical relationships between them. First, here’s some-
thing in common between the first three.
Lemma A4.7 Let k · k stand for k · k1 , k · k2 or k · k∞ . If fn → f in k · k then
kfn k → kf k as n → ∞.

Proof For all n, we have

kfn k − kf k ≤ kfn − f k

by Lemma A4.3(iii), and similarly

kf k − kfn k ≤ kf − fn k = kfn − f k,

so  
 kfn k − kf k  ≤ kfn − f k.

But kfn − f k → 0 as n → ∞, so kfn k − kf k → 0 as n → ∞, as required. 

Here’s a classic fact from PAA:

Fact A4.8 Let f, f1 , f2 , . . . : I → C be functions such that fn →R f uniformly.
Suppose
R that each fn is continuous. Then f is continuous, and I fn (x) dx →
I
f (x) dx as n → ∞.

In words: a uniform limit of continuous functions is continuous, and the integral

of a uniform limit is the limit of the integrals.
Now let’s begin to record the implications between the different types of
convergence.
Lemma A4.9 Let f, f1 , f2 , . . . : I → C be integrable functions. Then

fn → f uniformly
⇒ fn → f pointwise
⇒ fn → f a.e..

Proof Immediate from the definitions. 

We now restrict ourselves to the case I = [0, 1), since that’s what we’ll need
once we start to consider 1-periodic functions.

14
Lemma A4.10 i. Let f : [0, 1) → C be an integrable function. Then

kf k1 ≤ kf k2 ≤ kf k∞ .

ii. Let f, f1 , f2 , . . . : [0, 1) → C be integrable functions. Then

fn → f in k · k∞
⇒ fn → f in k · k2
⇒ fn → f in k · k1 .

Proof i. We prove kf k1 ≤ kf k2 later. (Reader: fill in the reference once

we’ve got to it!) For the other inequality:
s s
Z 1 Z 1
kf k2 = |f (x)|2 dx ≤ kf k2∞ dx = kf k∞ .
0 0

ii. Follows from (i) (with fn − f in place of f ). 

Remark A4.11 Here’s a summary of the implications:

k · k∞
tt NNNN
u} tttt #+
k · k2 pointwise
uu LLLLLLL
v~ uuuu LLL
"*
k · k1 a.e.

That is, k · k∞ -convergence implies k · k2 -convergence, and so on. Uniform

convergence is the strongest type there is.
No further implications hold: there are examples of sequences of functions
that converge in k · k1 but not in k · k2 , etc.

15
A5 Periodic functions
For the lecture of 27 January 2014; part one of two

You know what it means for a function f on R to be 1-periodic: f (x + 1) =

f (x) for all x. But in fact, there are at least three ways to think about a periodic
function: as a. . .

i. 1-periodic function R → C: R
−1 0 1

ii. function [a, a + 1) → C a

[a, a + 1) a+1
(for any a ∈ R):
a a+1

iii. function T → C,
T
where T is the circle:
0

In (i), it’s best to think of the real line R as coiled up into a spiral of period
1, so that x, x ± 1, x ± 2, . . . all lie on the same vertical line.
In (ii), it’s best to think of the interval [a, a + 1) as bent round into a circle,
as shown. (The values of a we most commonly use are 0 and −1/2.)
The viewpoint in (iii) is ultimately the most satisfactory. To get from (i)
to (iii), think of pushing down on the coil to squash it into a circle. To get
from (ii) to (iii), join the two ends of the arc.
Formally, T is the quotient group R/Z. (In case you’ve forgotten what this
means, there is an equivalence relation ∼ on R given by x ∼ y ⇐⇒ x − y ∈ Z;
then R/Z is the set of equivalence classes.) The elements of T are the elements
of R but with x regarded as the same as x + n in T whenever x ∈ R and n ∈ Z.
So,
. . . , −1.9, −0.9, 0.1, 1.1, 2.1, . . .
are all names for the same element of T. (Compare the fact that

. . . , −19, −9, 1, 11, 21, . . .

are all names for the same element of Z/10Z, i.e. the same integer mod 10.)
There are one-to-one correspondences between functions as in (i), functions
as in (ii), and functions as in (iii). We will switch freely between the three
ways of thinking about periodic functions, but most often, we will adopt the
viewpoint of (iii).
We will also take for granted some easy lemmas about periodic functions,
e.g. that any linear combination or product of 1-periodic functions is again 1-
periodic.
Definition A5.1 A function T → C is continuous if the corresponding 1-
periodic function R → C is continuous.

16
 Note that a 1-periodic function f : R → C is continuous if and only if its
restriction fe to [0, 1) is continuous and lim fe(x) = fe(0). (Here lim means
x→1− x→1−
the limit as x tends to 1 from below.) To see why the second condition is needed,
consider the function f : R → C defined by f (x) = x − bxc, where bxc is the
integer part of x:

−1 0 1 2

The restriction of f to [0, 1) is continuous, but f itself is not continuous, because

lim f (x) 6= f (0).
x→1−

Definition A5.2 A function f : T → C is integrable if the corresponding

function fe: [0, 1) → C is integrable. We then define
Z Z 1
f (x) dx = fe(x) dx, kf k1 = kfek1 , etc.
T 0

Remarks A5.3 i. Definition A5.2 is unchanged if we replace 0 by a and 1

by a + 1, for any a ∈ R.
ii. A little common sense is called for. When we say that a 1-periodic function
f : R → C is integrable, it does not mean that f itself is integrable! For
example, if f : R → C has constant value 6, then f is not integrable as
an ordinary function, but it is integrable as a 1-periodic function (and its
integral is 6). In practice, confusion shouldn’t arise.

Functions on a half-open interval can be continuous but not integrable, since

they might fail to be bounded by shooting off to ±∞ or oscillating wildly near
the open end. But things are easier in the world of periodic functions:
Lemma A5.4 Every continuous, 1-periodic function is integrable.

Proof Let f : R → C be a continuous 1-periodic function. We have to prove

that its restriction fe: [0, 1) → C is integrable. We know that fe is continuous,
so it is enough to prove that fe is bounded. Indeed, f |[0,1] is a continuous
function on a closed bounded interval, and therefore bounded; hence f |[0,1) = fe
is certainly bounded. 

17
A6 The inner product
For the lecture of 27 January 2014; part two of two

You’rePfamiliar with the scalar product v.w of two vectors v, w ∈ Rn , defined

by v.w = i vi wi . Sometimes this is written as hv, wi. PerhapsP you’re familiar
with the complex version: given v, w ∈ Cn , we put hv, wi = i vi wi . There is
also a version for complex-valued functions:
Definition A6.1 Let f, g : T → C be integrable functions. We define
Z
hf, gi = f (x)g(x) dx ∈ C.
T

In order for this definition to make sense, we need to know that the function
f · g is integrable. Lemmas A3.4(iii) and A3.7 guarantee this.
Lemma A6.2 Let f, g, h : T → C be integrable functions, and let a, b ∈ C.
Then:
i. hg, f i = hf, gi;
ii. haf + bg, hi = ahf, hi + bhg, hi and hf, ag + bhi = ahf, gi + bhf, hi;
iii. hf, f i = kf k22 ≥ 0.

Proof Sheet 1, q.6. 

The properties of h·, ·i stated in this lemma nearly say that it is an inner
product. All that prevents it from being one is that hf, f i = 0 does not quite
imply f = 0; it only implies that f = 0 almost everywhere (by Lemma A4.5(i)).
However, I will abuse terminology slightly by referring to hf, gi as the inner
product of f and g anyway.
Here are some further properties of h·, ·i.
Lemma A6.3 Let f, g : T → C be integrable functions. Then:
i. kf + gk22 = kf k22 + kgk22 + 2Rehf, gi (‘cosine rule’);
ii. if hf, gi = 0 then kf + gk22 = kf k22 + kgk22 (‘Pythagoras’).

Proof For (i), use khk22 = hh, hi. Part (ii) follows immediately. 

Here is the most fundamental result about the inner product.

Theorem A6.4 (Cauchy–Schwarz inequality) Let f, g : T → C be inte-
grable functions. Then
|hf, gi| ≤ kf k2 kgk2 .

Proof Sheet 1, q.6. 

Note the modulus sign on the left-hand side. Without it, the inequality
would not even make sense, since hf, gi is not usually a real number.
From this, we deduce that kf + gk2 ≤ kf k2 + kgk2 (Sheet 1, q.6 again; see
also Lemma A4.3). We also deduce the following, which completes the proof of
Lemma A4.10:

18
Lemma A6.5 Let f : [0, 1) → C be an integrable function. Then kf k1 ≤ kf k2 .

Proof Let us apply the Cauchy–Schwarz inequality to |f | and the constant

function 1. We have

h|f |, 1i = kf k1 , k |f | k2 = kf k2 , k1k2 = 1,

giving kf k1 ≤ kf k2 · 1 = kf k2 , as required. 

There is a cousin of the Cauchy–Schwarz inequality that is also useful:

Lemma A6.6 Let f, g : T → C be integrable functions. Then

|hf, gi| ≤ kf k1 kgk∞ .

Proof We have
Z  Z
   
|hf, gi| =  f (x)g(x) dx ≤ f (x)g(x) dx

ZT T
Z
= |f (x)| · |g(x)| dx ≤ |f (x)| · kgk∞ dx = kf k1 kgk∞ ,
T T

using Lemma A3.9 in the first inequality. 

Remark A6.7 (Non-examinable.) Both the Cauchy–Schwarz inequality and

Lemma A6.6 are special cases of a result known as Hölder’s inequality, which
states that |hf, gi| ≤ kf kp kgkq whenever 1/p + 1/q = 1 (with 1 ≤ p ≤ ∞,
1 ≤ q ≤ ∞). Here kf kp and kgkq are defined as in Remark A4.2(iv).

Lemma A6.6 easily implies two more small results, both of which will be
useful later:
Lemma
p A6.8 Let f : T → C be an integrable function. Then kf k2 ≤
kf k1 kf k∞ .

Proof Put g = f in Lemma A6.6. 

Lemma A6.9 Let f1 , f2 , . . . , f, g : T → C be integrable functions. Suppose that

fn → f as n → ∞ in k · k1 . Then hfn , gi → hf, gi as n → ∞.

Proof For each n, we have

|hfn , gi − hf, gi| = |hfn − f, gi| ≤ kfn − f k1 kgk∞ ,

and kfn − f k1 → 0 as n → ∞, so the result follows. 

19
A7 Characters and Fourier series
For the lecture of 30 January 2014; part one of two

Among all periodic functions, certain ones are special. These are the so-
called ‘characters’. Fourier theory can be seen as an attempt to build all periodic
functions out of characters.
Let k ∈ Z. We define ek : R → C by ek (x) = e2πikx . This function is
1-periodic, so can be seen as a function ek : T → C, the kth character of T.
Remarks A7.1 i. The notation ek is not standard. No one outside this
class will know what you mean by ‘ek ’ unless you define it.
ii. The apparently strange terminology ‘character of T’ will be put into con-
text in the very last part of this course.
Here are some elementary properties of the characters.
Lemma A7.2 Let k ∈ Z and x, y ∈ T. Then:
i. ek is continuous;
ii. |ek (x)| = 1;
iii. ek (x + y) = ek (x)ek (y), ek (−x) = 1/ek (x), and ek (0) = 1.
Note that in (iii), it does make sense to add and subtract elements of T,
because T is by definition a group (the quotient R/Z).
Proof Straightforward. 
Some further elementary properties:
Lemma A7.3 Let k, ` ∈ Z. Then:
i. ek+` = ek · e` , e−k = 1/ek , and e0 = 1;
ii. e−k = ek ;
iii. ek = ek1 .
Proof Straightforward. 
We now come to a crucial property of the characters.
Lemma A7.4 The characters (ek )k∈Z are orthonormal. That is, for k, ` ∈ Z,
(
1 if k = `,
hek , e` i =
0 if k 6= `.
Proof We have
Z Z 1 Z 1
hek , e` i = ek (x)e` (x) dx = ek (x)e−` (x) dx = ek−` (x) dx,
T 0 0

using Lemma A7.3. If k = ` then the integrand is e0 = 1, so hek , e` i = 1. If

k 6= ` then
 1
1 2πi(k−`)x
hek , e` i = e = 0,
2πi(k − `) 0
as required. 

20
 You can think of the characters as analogous to the standard basis vectors
in Rn (which are also orthonormal). When we express a point of Rn in terms of
its coordinates, we are viewing it as a linear combination of the standard basis
vectors. Similarly, in Fourier theory, we seek to view any periodic function as a
linear combination of the characters. The analogy is not exact, because there
are infinitely many characters, so we have to take infinite linear combinations
of characters. This is what gives the subject its subtlety.
Here are the central definitions of this course.
Definition A7.5 Let f : T → C be an integrable function.
i. For k ∈ Z, the kth Fourier coefficient of f is

fˆ(k) = hf, ek i.

ii. For n ≥ 0, the nth Fourier partial sum of f is the function

n
X
Sn f = fˆ(k)ek : T → C.
k=−n

iii. The Fourier series of f is the expression

∞
X
Sf = fˆ(k)ek .
k=−∞

Remarks A7.6 i. Explicitly,

Z 1
ˆ
f (k) = f (x)e−2πikx dx (k ∈ Z),
0
Xn
(Sn f )(x) = fˆ(k)e2πikx (n ≥ 0, x ∈ T),
k=−n
X∞
(Sf )(x) = fˆ(k)e2πikx .
k=−∞

P∞
ii. The Fourier series of f is k=−∞ hf, ek iek . Compare:
Pn if u1 , . . . , un denote
the standard basis vectors of Rn , then v = k=1 (v.uk )uk for all v ∈ Rn .
So, we might guess that f is ‘equal’ to its Fourier series (whatever that
means). The central question of this subject is whether, and in what sense,
this is actually true.

We finish by recording two basic properties of Fourier coefficients.

Lemma A7.7 Let k ∈ Z. Then:
\
i. af + bg(k) = afˆ(k) + bĝ(k) for all a, b ∈ C and integrable f, g : T → C.
ii. Let f1 , f2 , . . . , f : T → C be integrable functions such that fn → f as
n → ∞ in k · k1 . Then fc ˆ
n (k) → f (k) as n → ∞.

Proof Part (i) follows from Lemma A6.2, and part (ii) from Lemma A6.9. 

21
A8 Trigonometric polynomials
For the lecture of 30 January 2014; part two of two

As we saw in the last section, Fourier theory asksPwhether a periodic function

∞
f can be expressed as a linear combination Sf = k=−∞ fˆ(k)ek of characters
ek . This is, in general, an infinite linear combination (whatever that means).
But to get started, it is useful to consider the finite linear combinations of
characters. These are called trigonometric polynomials.
Definition A8.1 A function g : T → C is a trigonometric Pnpolynomial if
there exist n ≥ 0 and c−n , . . . , c0 , . . . , cn ∈ C such that g = k=−n ck ek . The
degree of g is the least n for which this is possible.

Example A8.2 For Pnany integrable f : T → C and n ≥ 0, the nth Fourier

partial sum Sn f = k=−n fˆ(k)ek is a trigonometric polynomial of degree ≤ n.

It is perhaps not obvious that the coefficients of a trigonometric polynomial

are unique. The next two results show that, in fact, they are.
Pn
Lemma A8.3 Let g = k=−n ck ek be a trigonometric polynomial. Then
(
ck if |k| ≤ n,
ĝ(k) =
0 if |k| > n

(k ∈ Z).

Proof We have
* n
+ n
X X
ĝ(k) = hg, ek i = c` e` , ek = c` he` , ek i
`=−n `=−n
(
ck if − n ≤ k ≤ n
=
0 otherwise,

where in the last step we used the fact that the characters are orthonormal. 
Pn Pn
Corollary A8.4 If k=−n ck ek = k=−n dk ek then ck = dk for all k ∈
{−n, . . . , 0, . . . , n}. 

In other words, the characters are linearly independent.

Example A8.5 Let f : T → C be an integrable function. Then for all n ≥ 0
and k ∈ Z, (
[ fˆ(k) if |k| ≤ n,
Sn f (k) =
0 otherwise.
In other words, Sn f and f have the same kth Fourier coefficients for |k| ≤ n.

22
 Here is a baby version of the whole of Fourier theory.
Proposition A8.6 Let n ≥ 0. Then the functions

{trigonometric polynomials of degree ≤ n}  C2n+1

given by

g 7→ ĝ(−n), . . . , ĝ(0), . . . , ĝ(n)
Pn
k=−n ck ek ←[ (c−n , . . . , c0 , . . . , cn )
are mutually inverse.
Pn
Proof • Let g = k=−n cP k ek be a trigonometric polynomial of degree ≤ n.
n
We must show that g = k=−n ĝ(k)ek . This follows from Lemma A8.3.
Pn
• Let (c−n , . . . , cn ) ∈ C2n+1 and put g = k=−n ck ek . We must show that
ĝ(k) = ck for all k ∈ {−n, . . . , n}. This also follows from Lemma A8.3. 

Fantasy A8.7 We can fantasize about extending Proposition A8.6 to a pair of

mutually inverse functions

{nice functions T → C}  {nice double sequences in C}

given by
∞
f 7→ fˆ(k) k=−∞
P∞
k=−∞ ck ek ←[ (ck )∞
k=−∞ .

The rest of the course explores this fantasy.

23
 f (· + t) f

−t 0

Figure A.4: Translating a function.

A9 Integrable functions are sort of continuous

For the lecture of 3 February 2014

Integrable functions are not necessarily continuous. However, integrable

functions do satisfy a continuity-like condition of a not-so-obvious kind, as fol-
lows.
Let f : T → C be a function. Given t ∈ T, we obtain a new function

f (· + t) : T → C

defined by
x 7→ f (x + t).
Geometrically, this means shifting the graph of the function by t units to the
left (Fig. A.4). (Or really, since T is a circle, it means rotating the graph by a
fraction t of a revolution.) It’s not so hard to see (try it!) that

f is continuous ⇐⇒ f (· + t) → f pointwise as t → 0

—in other words, for each x ∈ T, f (x + t) → f (x) as t → 0. Similarly, it’s not

hard to see that

f is uniformly continuous ⇐⇒ f (· + t) → f in k · k∞ as t → 0

—in other words, kf (· + t) − f k∞ → 0 as t → 0. Obviously, neither condition

is satisfied by an arbitrary integrable function. However, it is true that for any
integrable function,
f (· + t) → f in k · k1 as t → 0. (A:4)
Proving this will take the rest of this lecture, and will require us to go right
back to the definition of integrability.
Here’s the plan. First we’ll introduce a class of particularly simple functions,
the so-called ‘step functions’. We’ll show that every integrable function can
be closely approximated by a step function. The proof of (A:4) for arbitrary
integrable functions f proceeds in two steps: (i) prove it for step functions
(which is relatively easy, as step functions are so simple); (ii) extend the result
to an arbitrary integrable f by approximating it with step functions.
Definition A9.1 Let I ⊆ R be aPbounded interval. A step function is a
n
function f : I → C such that f = k=1 ck χJk for some n ≥ 0, c1 , . . . , cn ∈ C,
and bounded intervals J1 , . . . , Jn ⊆ I.

24
 0 1 2 3

Figure A.5: A step function.

In other words, a step function is a finite linear combination of characteristic

functions of intervals.
Example A9.2 2χ[0,3] − χ[1,3) is a step function (Figure A.5).

Step functions are integrable, by Example A3.2(ii) and Lemma A3.4. Indeed,
Z X X
ck χJk (x) dx = ck |Jk | .
I k k

Of course, not every integrable function is a step function. However, there is

a sense in which every integrable function can be well approximated by a step
function. This is somewhat similar to the fact that arbitrary images can be
displayed on a computer screen, which is a grid of discrete pixels; the finer the
grid is, the better the quality of the display.
To make ‘well approximated’ precise, we introduce some more terminology.
Definition A9.3 Let I be a bounded interval, and let k · k stand for k · k1 , k · k2
or k · k∞ . Let F be a set of functions I → C, and let G ⊆ F. Then G is dense
in F with respect to k · k if:
for all f ∈ F, for all ε > 0, there exists g ∈ G such that kf − gk < ε.

(You may be familiar with something like this definition from the theory of
metric spaces; for example, Q is dense in R.)
Proposition A9.4 Let I be a bounded interval. Then {step functions I → C}
is dense in {integrable functions I → C} with respect to k · k1 .

Proof We prove it just for I = [0, 1), since this is the case that will matter
most to us and the proof for other bounded intervals is very similar.
First take a real -valued integrable function f : [0, 1) → R, and let ε > 0. By
definition of integration, we can choose a partition P of [0, 1) such that
Z 1
L(f, P ) > f (x) dx − ε.
0

Here L(f, P ) is the lower Darboux sum: if P is the partition

0 = x0 < x1 < · · · < xn = 1

and we put mk = inf{f (x) : xk−1 ≤ x ≤ xk } (for 1 ≤ k ≤ n), then

n
X
L(f, P ) = mk (xk − xk−1 ).
k=1

25
Put Jk = [xk−1 , xk ) and put
n
X
g= mk χJk .
k=1
R1
(Draw a picture!) Then g is a step function, with L(f, P ) = 0
g(x) dx. We
have g(x) ≤ f (x) for all x ∈ [0, 1), so
Z 1 Z 1
kf − gk1 = (f (x) − g(x)) dx = f (x) dx − L(f, P ) < ε,
0 0

as required.
Now take an arbitrary integrable function f : [0, 1) → C, say f = f1 +if2 with
f1 , f2 : [0, 1) → R. Let ε > 0. By definition of integrability of complex-valued
functions (Definition A3.1(i)), f1 and f2 are integrable. So by the previous
paragraph, we can choose step functions g1 , g2 : [0, 1) → R such that

kf1 − g1 k1 < ε/2, kf2 − g2 k1 < ε/2.

Put g = g1 + ig2 : [0, 1) → C, which is also a step function. Then

kf − gk1 = k(f1 − g1 ) + i(f2 − g2 )k1 ≤ kf1 − g1 k1 + kf2 − g2 k1 < ε/2 + ε/2 = ε

(using Lemma A4.3), as required. 

We are now ready to prove that integrable functions are ‘sort of continuous’,
in the sense of (A:4), following the plan described above.
Theorem A9.5 Let f : T → C be an integrable function. Then f (· + t) → f in
k · k1 as t → 0 (that is, kf (· + t) − f k1 → 0 as t → 0).

Proof For the duration of this proof only, let us say that an integrable function
g : T → C is ‘good’ if g(· + t) → g in k · k1 as t → 0. We will prove that every
integrable function is good.
First, for any interval J ⊆ [0, 1), the characteristic function χJ is good.
Indeed, write a = inf J and b = sup J. If a = b, or if a = 0 and b = 1, then
kχJ (· + t) − χJ k1 = 0 for all t. Otherwise, we can show that

kχJ (· + t) − χJ k1 = 2|t|

whenever |t| is sufficiently small (exercise; see Figure A.6). Hence kχJ (· + t) −
χJ k1 → 0 as t → 0. Pn
Second, every step function g : [0, 1) → C is good. Indeed, if g = k=1 ck χJk
(in the usual notation) then

n
X

kg(· + t) − gk1 =
c k χ Jk (· + t) − χ Jk

k=1 1
n
X
≤ |ck | χJk (· + t) − χJk 1
k=1
→0

26
 a−t a b−t b

Figure A.6: The area between a characteristic function and a translation of it

by a small distance t is 2|t|.

as t → 0, by the first part.

Finally, every integrable function f : [0, 1) → C is good. Indeed, let ε > 0.
By Proposition A9.4, we can choose a step function g : [0, 1) → C such that
kf − gk1 < ε/3. This implies that for all t ∈ R,
Z
kf (· + t) − g(· + t)k1 = |f (x + t) − g(x + t)| dx
ZT
= |f (y) − g(y)| dy = kf − gk1 < ε/3,
T

where the second equality is by substitution. (Intuitively, the area between the
graphs f and g is unchanged if we shift everything horizontally by t.) Also,
by the second part, we can choose δ > 0 such that for all t ∈ (−δ, δ), we have
kg(· + t) − gk1 < ε/3. Now for all t ∈ (−δ, δ), we have

kf (· + t) − f k1 ≤ kf (· + t) − g(· + t)k1 + kg(· + t) − gk1 + kg − f k1

< ε/3 + ε/3 + ε/3 = ε,

as required. 

27
Chapter B

Convergence of Fourier
series in the 2- and 1-norms

In this chapter, we’ll prove:

for any integrable function f : T → C, the Fourier series of f con-
verges to f in both k · k2 and k · k1 .

It’s as simple as that. Compare the long and complicated saga of pointwise
convergence recounted in Section A2. In contrast, this result is clean, easily
stated, and not too difficult to prove.
The 2-norm will play a much greater role than the 1-norm in this chapter. Be-
cause k·k2 goes hand in hand with inner products (recalling that kf k22 = hf, f i),
working in the 2-norm has a great deal in common with ordinary Euclidean ge-
ometry. It’s the most easily visualized context to work in.

B1 Nothing is better than a Fourier partial sum

For the lecture of 6 February 2014; part one of two

The title of this section is made precise by part (i) of the following lemma,
illustrated in Figure B.1. Recall that given an integrable function f , its nth
Fourier partial sum Sn f is a trigonometric polynomial of degree ≤ n.

Lemma B1.1 Let f : T → C be an integrable function, and let n ≥ 0. Then:

i. kf − Sn f k2 ≤ kf − gk2 whenever g is a trigonometric polynomial of degree
≤ n.
ii. kf k22 = kSn f k22 + kf − Sn f k22 .
Pn
iii. kSn f k22 = k=−n |fˆ(k)|2 .

Proof Before we prove any of the three parts, we show that hf − Sn f, hi = 0

for all trigonometric polynomials h of degree ≤ n. (Figure B.1 makes this look

28
 f

Sn f

g
0
trig poly
s of deg
≤n
Figure B.1: Approximating a function f by a trigonometric polynomial of degree
≤ n.

plausible.) Indeed, by linearity, it is enough to prove this when h = ek for some

k ∈ {−n, . . . , 0, . . . , n}, and

hf − Sn f, ek i = hf, ek i − hSn f, ek i = fˆ(k) − Sd

n f (k) = 0

by Example A8.5, as required.

Now let g be a trigonometric polynomial of degree ≤ n. By what we have
just shown, hf − Sn f, Sn f − gi = 0, so

kf − gk22 = k(f − Sn f ) + (Sn f − g)k22

= kf − Sn f k22 + kSn f − gk22 (B:1)

by Pythagoras (Lemma A6.3(ii)). Part (i) follows, then part (ii) by putting
g = 0 in (B:1).
For part (iii), we have
* n n
+
X X
2
kSn f k2 = ˆ
f (k)ek , ˆ
f (`)el
k=−n `=−n
n
X
= fˆ(k)fˆ(`)hek , e` i
k,`=−n
n 
 ˆ 2
X 
= f (k) ,
k=−n

where the first two equalities use Lemma A6.2 and the third uses the orthonor-
mality of the characters (Lemma A7.4). 

To elaborate a little on the title of this section: among all trigonometric

polynomials of degree ≤ n, nothing approximates f better than the Fourier
partial sum Sn f .
Proposition B1.2 Let f : T → C be an integrable function. Then:
 ˆ 2 converges; indeed, ∞  ˆ 2 ≤ kf k22 .
P∞   P  
i. k=−∞ f (k) k=−∞ f (k)

ii. (Riemann–Lebesgue Lemma) fˆ(k) → 0 as k → ±∞.

29
 Pn
Proof By parts (ii) and (iii) of Lemma B1.1, we have k=−n |fˆ(k)|2 ≤ kf k22
for all n ≥ 0. The results follow. 

The Riemann–Lebesgue lemma tells us that not every double sequence

(ck )∞
k=−∞ of complex numbers arises as the sequence of Fourier coefficients of
some function. This is already a substantial result. Looking back again at
Fantasy A8.7, we begin to get a sense of what a ‘nice’ double sequence might
be.
The next lemma will help us to prove that (Sn f ) always converges to f in
the 2-norm.
Lemma B1.3 Let f : T → C be an integrable function. The following are equiv-
alent:
i. Sn f → f in k · k2 ;
ii. gn → f in k · k2 for some sequence (gn ) of trigonometric polynomials.

Proof (i) ⇒ (ii) is trivial, since each Sn f is a trigonometric polynomial.

For (ii) ⇒ (i), let ε > 0. Choose m such that kf − gm k2 < ε. Put N =
deg(gm ). Then
ε > kf − gm k2 ≥ kf − SN f k2 ,
using Lemma B1.1(i) in the second inequality. But then

kf − SN f k2 ≥ kf − SN +1 f k2

by Lemma B1.1(i) again, noting that SN f has degree ≤ N + 1. Continuing like

this, we find that

ε > kf − SN f k2 ≥ kf − SN +1 f k2 ≥ kf − SN +2 f k2 ≥ · · · ,

and in particular, kf − Sn f k2 < ε for all n ≥ N . 

Our strategy for proving that Sn f → f in k · k2 will be to prove that gn → f

in k · k2 for some other sequence (gn ) of trigonometric polynomials—a sequence
that is easier to work with than (Sn f ) itself.
In order to carry out this strategy, we need to come up with some convenient
sequence (gn ) of trigonometric polynomials. How can we cook up new trigono-
metric polynomials? By convolution, as we’ll see in the next two sections.

30
B2 Convolution: definition and examples
For the lecture of 6 February 2014; part two of two

You’ve already met convolution of functions on R. Convolution of functions

on T, which is what we’ll mostly be concerned with, is very similar.
Definition B2.1 Let f, g : T → C be integrable functions. The convolution
f ∗ g : T → C is the function defined by
Z
(f ∗ g)(x) = f (t)g(x − t) dt
T

(x ∈ T).

Remark B2.2 Recall from Section A5 that the circle T is a group (the quotient
R/Z). The group operation is addition mod 1. So, it does make sense to add
and subtract elements of T, as we did in the definition of convolution.

Our first example of convolution is important enough to be stated as a

lemma.
Lemma B2.3 For any k ∈ Z and integrable function f : T → C,

f ∗ ek = hf, ek iek = fˆ(k)ek .

Proof For all x ∈ T,

Z
(f ∗ ek )(x) = f (t)ek (x − t) dt
T
Z
= f (t)ek (x)ek (−t) dt
T
Z
= ek (x) f (t)ek (t) dt
T
= hf, ek iek (x)
= fˆ(k)ek (x)

by Lemmas A7.2 and A7.3. 

This is remarkable. It tells us that when we convolve a character ek with

anything (anything! ), the result is a scalar multiple of ek . This is a very special
property of the characters.
Convolution has a smoothing effect. For the purposes of the following ex-
ample, let us consider functions on R rather than T.
Example B2.4 Given an integrable function f : R → C, we have
Z x+1/2


f ∗ χ[−1/2,1/2] (x) = f (t) dt.
x−1/2

This is a ‘moving average’ of f : the value of f ∗ χ[−1/2,1/2] at x is the mean

value of f over the interval [x − 1/2, x + 1/2] (Figure B.2).

31
 f

f ∗ χ[−1/2,1/2]

1 1 1

Figure B.2: Convolution has a smoothing effect.

32
B3 Convolution: properties
For the lecture of 10 February 2014; part one of two

Example B2.4 showed the smoothing effect of convolution. In that particular

example, the convolution of two discontinuous functions was continuous. This
surprising behaviour is, in fact, a completely general phenomenon:
Lemma B3.1 Let f, g : T → C be integrable functions. Then their convolution
f ∗ g : T → C is continuous.
This is far stronger than we might have guessed in advance. It’s not merely
true that the convolution of two integrable functions is integrable, or that the
convolution of two continuous functions is continuous. In fact, for any two
integrable functions f and g, no matter how discontinuous they may be, f ∗ g
is continuous.
Proof Let x, h ∈ T. Then
Z 
 
|(f ∗ g)(x + h) − (f ∗ g)(x)| =  f (t)[g(x + h − t) − g(x − t)] dt

ZT
≤ |f (t)| |g(x + h − t) − g(x − t)| dt (B:2)
T
Z
≤ kf k∞ |g(x + h − t) − g(x − t)| dt
ZT
= kf k∞ |g(u + h) − g(u)| du (B:3)
T
= kf k∞ kg(· + h) − gk1
→ 0 as h → 0. (B:4)
Here (B:2) is by the triangle inequality for integration (Lemma A3.9), equa-
tion (B:3) comes from substituting u = x − t, and (B:4) is by Theorem A9.5
(‘integrable functions are sort of continuous’). 
Here are some basic properties of convolution. In future, we will use them
without explicitly referring back to this lemma.
Lemma B3.2 Let f, g, h : T → C be integrable functions, and let c ∈ C. Then:
i. f ∗ g = g ∗ f ;
ii. (f ∗ g) ∗ h = f ∗ (g ∗ h);
iii. f ∗ (g + h) = (f ∗ g) + (f ∗ h);
iv. f ∗ (cg) = c(f ∗ g).
Proof For (i), let x ∈ T. Then
Z
(f ∗ g)(x) = f (t)g(x − t) dt
ZT
= f (x − u)g(u) du
T
= (g ∗ f )(x),
by substituting u = x − t. The other parts are similarly straightforward. 

33
Remark B3.3 This lemma tells us that + and ∗ give the set

{integrable functions T → C}

the structure of a ‘commutative algebra’ over C, that is, both a vector space
over C and a ring.
Or nearly. The only missing part is that it has no multiplicative identity.
Nowadays, the definition of ‘ring’ is usually taken to include the existence of a
multiplicative identity; but in analysis especially, there are important rings that
do not have one.
We’ll see very soon that if the identity did exist, it would be the mythical
‘delta function’.

Lemma B3.4 Let f : T → C be an integrable function, and let g : T → C be a

trigonometric polynomial. Then f ∗ g is a trigonometric polynomial.

Again, this is stronger than might be expected. It’s not merely true that
the convolution of two trigonometric polynomials is a trigonometric polyno-
mial. In fact, the convolution of anything with a trigonometric polynomial is a
trigonometric polynomial.
Pn
Proof We may write g = k=−n ck ek for some n ≥ 0 and ck ∈ C. Then
n
X n
X
ck fˆ(k) ek ,


f ∗g = ck (f ∗ ek ) =
k=−n k=−n

which is a trigonometric polynomial. Here the second equality uses

Lemma B2.3. 

Remark B3.5 This lemma says that {trigonometric polynomials} is an ideal

in the ring {integrable functions T → C}.

Example B3.6 Every Fourier partial sum of a function f is a convolution of

f with a trigonometric polynomial. More exactly, using Lemma B2.3 again,
n
X n
X n
X
Sn f = fˆ(k)ek = f ∗ ek = f ∗ ek .
k=−n k=−n k=−n
Pn
Since k=−n ek is a trigonometric polynomial, Lemma B3.4 tells us in this case
that Sn f is also a trigonometric polynomial—which of course we already knew.

Since Fourier
Pn partial sums are important in Fourier theory, this example
suggests that k=−n ek is also important. It is, and it has its own name:
Definition
Pn B3.7 Let n ≥ 0. The Dirichlet kernel of order n is Dn =
k=−n e k .

So, Example B3.6 can be restated as:

Lemma B3.8 Sn f = f ∗ Dn , for all integrable f : T → C and n ≥ 0. 

Remark B3.9 We can dream P∞of taking n → ∞ in LemmaPB3.8, so that

∞
Sf = f ∗ D∞ where D∞ = k=−∞ ek . However, the sum k=−∞ ek does
not converge anywhere, so D∞ does not really exist.

34
 Here are two further properties of convolution.
Lemma B3.10 For integrable functions f, g : T → C,

kf ∗ gk∞ ≤ kf k∞ kgk1 .

Proof For all x ∈ T,

Z 
 
|(f ∗ g)(x)| =  f (x − t)g(t) dt
ZT
≤ |f (x − t)| |g(t)| dt
T
Z
≤ kf k∞ |g(t)| dt = kf k∞ kgk1 . 
T

The final property should remind you of an important fact about Fourier
transforms.

Lemma B3.11 Let f, g : T → C be integrable functions, and let k ∈ Z. Then

∗ g(k) = fˆ(k)ĝ(k).
f[

Proof Sheet 2, q.5. 

35
 1/ε

−ε/2 ε/2

Figure B.3: Approximation to the mythical delta function.

B4 The mythical delta function

For the lecture of 10 February 2014; part two of two

This short section is largely intended as motivation for the rest of Part B. It
contains no definitions or theorems, but the ideas are important for what follows.
Fact: there is no integrable function δ : T → C such that
Z
for all continuous f : T → C, f (x)δ(x) dx = f (0). (B:5)
T

(Compare Sheet 1, q.5.) But we can get close. Non-rigorously, for a ‘small’
ε > 0, put
1
∆ε = χ[−ε/2,ε/2] : [−1/2, 1/2) → C
ε
(Figure B.3). Then for any continuous function f : T → C,
Z Z ε/2 Z ε/2
1 1
f (x)∆ε (x) dx = f (x) dx ≈ f (0) dx = f (0),
T ε −ε/2 ε −ε/2

using continuity in the approximate equality.

Imagine now that there is a function δ satisfying (B:5).
Let f : T → C be a continuous function. Then

f ∗ δ = f,

since Z
(f ∗ δ)(x) = f (x − t)δ(t) dt = f (x − 0) = f (x)
T

for all x ∈ T. (So δ is an identity for convolution with continuous functions;

compare Remark B3.3.)
If δ was actually a trigonometric polynomial, then the equation f ∗ δ = f
together with Lemma B3.4 would imply that every continuous function f was
a trigonometric polynomial. This is obviously false.

36
 If δ is merely a limit of trigonometric polynomials Kn , say Kn → δ in k · k2 ,
then perhaps it follows that f ∗ Kn → f ∗ δ in k · k2 . In that case, we would have
f ∗ Kn → f in k · k2 . Lemma B1.3 would then imply that Sn f → f in k · k2 .
This is the result we’re aiming for.

We know that no δ satisfying (B:5) exists. However, the previous paragraph

suggests a strategy:
Look for a sequence (Kn ) of trigonometric polynomials such that
for all continuous (or even integrable) functions f : T → C,
f ∗ Kn → f in k · k2 .
Our fantasies about the delta function suggest that the sequence (Kn ) should
somehow ‘converge to δ’. So, the plan now is to look for such a sequence (Kn ),
and show that it does what this informal chain of reasoning leads us to hope it
will.

37
 Kn

−1/2 −δ δ 1/2
Figure B.4: Axiom PAD3 states that for large n, the shaded area is small.

B5 Positive approximations to delta

For the lecture of 13 February 2014

The last section culminated in a plan: to show that Sn f → f in k · k2 for all f ,

• look for a sequence (Kn ) of trigonometric polynomials that in some sense
‘converges to the (non-existent) delta function’; then
• show that f ∗ Kn → f in k · k2 for all f .
We’ll carry out this plan. In this section, we give a precise meaning to the
phrase ‘converges to the delta function’. We also show that if (Kn ) has this
property, then f ∗ Kn → f in k · k2 for all f . Actually finding such a sequence
(Kn ) is left until later.
Definition B5.1 A positive approximation to delta (PAD) is a sequence
(Kn )∞
n=0 of integrable functions T → R such that:

PAD1 for all n ≥ 0 and t ∈ T, we have Kn (t) ≥ 0;

R
PAD2 for all n ≥ 0, we have T Kn (t) dt = 1;
R
PAD3 for all δ ∈ (0, 1/2), we have lim δ<|t|≤1/2 Kn (t) dt = 0.
n→∞

R
Remarks B5.2 i. PAD2 is inspired by the thought that if T δ(t)f (t) dt =
f (0) for all continuous f (as in the previous section),R then in particular
this is true when f is the constant function 1, giving T δ(t) dt = 1.
ii. In PAD3, the ‘δ’ mentioned is a real number, not the delta function!
iii. PAD2 tells us that the area under the graph of Kn is always 1. PAD3 says
that as n gets larger, that area gets concentrated into an ever-narrower
strip around the y-axis (Figure B.4).
iv. For this part of the theory, it’s convenient to view functions on T as
functions on [−1/2, 1/2).

38
 v. The name ‘approximation to delta’ is non-standard. The usual name is
‘approximation to the identity’, since delta is the (non-existent) identity
for convolution.
vi. Almost every definition in this course is conceptually well-motivated. This
one, however, is not. The conditions PAD1–3 are just what is needed in
order to make the arguments work. (Other variants are possible.) How-
ever, it’s only a stepping stone, which we’ll use to reach theorems that are
clean and free of arbitrary conditions.

∞
Examples B5.3 i. nχ[−1/2n,1/2n) n=1 is a PAD. (Check!) A typical ele-
ment of this sequence is shown in Figure B.3, taking ε = 1/n.
ii. (Dn )∞
n=0 is not a PAD. First, it fails PAD1, since (for instance) D1 (1/2) =
−1 < 0. More seriously, it can be shown that kDn k1 → ∞ as n → ∞,
whereas if (Kn ) is a PAD then
Z
kKn k1 = Kn (t) dt = 1 (B:6)
T

for all n (by PAD1 and PAD2).

The rest of this section is devoted to showing that when (Kn ) is a PAD,
f ∗ Kn → f in k · k2 for any integrable f . The proof is quite delicate.
The first step is to prove the weaker result that f ∗ Kn → f in k · k1 . (To
see why it’s ‘weaker’, recall Lemma A4.10.)
Proposition B5.4 Let (Kn )∞n=0 be a PAD and let f : T → C be an integrable
function. Then f ∗ Kn → f in k · k1 .
Proof We begin by finding an upper bound for kf ∗ Kn − f k1 .
For all x ∈ T and n ≥ 0,
Z 1/2 
 
|(f ∗ Kn )(x) − f (x)| = 
 (f (x − t) − f (x))Kn (t) dt
−1/2
Z 1/2
≤ |f (x − t) − f (x)|Kn (t) dt,
−1/2

using PAD2 in the first line and PAD1 in the second.

(The idea now is roughly as follows. Let δ > 0 be small. Then |f (x−t)−f (x)|
is small for |t| < δ, at least if f is sort of continuous. Also, Kn (t) is small for
|t| > δ, since Kn is something like the delta function.)
It follows that for all x ∈ T, n ≥ 0 and δ ∈ (0, 1/2),
|(f ∗ Kn )(x) − f (x)|
Z Z
≤ |f (x − t) − f (x)|Kn (t) dt + |f (x − t) − f (x)|Kn (t) dt,
|t|<δ δ<|t|≤1/2

so by the triangle inequality,

|(f ∗ Kn )(x) − f (x)|
Z Z
≤ |f (x − t) − f (x)|Kn (t) dt + 2kf k∞ Kn (t) dt. (B:7)
|t|<δ δ<|t|≤1/2

39
Integrating (B:7) with respect to x shows that for all n ≥ 0 and δ ∈ (0, 1/2),
kf ∗ Kn − f k1
Z 1/2 Z Z
≤ |f (x − t) − f (x)|Kn (t) dt dx + 2kf k∞ Kn (t) dt (B:8)
−1/2 |t|<δ δ<|t|≤1/2
Z Z
= kf (· − t) − f k1 Kn (t) dt + 2kf k∞ Kn (t) dt. (B:9)
|t|<δ δ<|t|≤1/2
R
(To get (B:8), we used the fact that 2kf k∞ δ<|t|≤1/2 Kn (t) dt is independent of
x, so that integrating it with respect to x over an interval of length 1 leaves it
unchanged.)
We now prove convergence. Let ε > 0. By Theorem A9.5 (‘integrable
functions are sort of continuous’), we can choose δ ∈ (0, 1/2) such that for all
t ∈ (−δ, δ), kf (· − t) − f k1 < ε/2. By PAD3, we can then choose N ≥ 0 such
that for all n ≥ N , Z
ε
Kn (t) dt < .
δ<|t|≤1/2 4kf k∞
So by (B:9), for all n ≥ N ,
Z
ε ε
kf ∗ Kn − f k1 ≤ Kn (t) dt + 2kf k∞
2 |t|<δ 4kf k∞
Z 1/2
ε ε
≤ Kn (t) dt + (B:10)
2 −1/2 2
= ε, (B:11)
using PAD1 in (B:10) and PAD2 in (B:11). 
We now know that f ∗ Kn → f in k · k1 for any integrable f . It is relatively
easy to deduce the stronger result that f ∗ Kn → f in k · k2 . The key is
Lemma A6.8, which gives an upper bound on the 2-norm in terms of the 1- and
∞-norms.
Proposition B5.5 Let (Kn )∞n=0 be a PAD and let f : T → C be an integrable
function. Then f ∗ Kn → f in k · k2 .
Proof For all n ≥ 0, we have
kf ∗ Kn k∞ ≤ kf k∞ kKn k1 = kf k∞
by Lemma B3.10 and (B:6). So for all n ≥ 0,
kf ∗ Kn − f k∞ ≤ kf ∗ Kn k∞ + kf k∞ ≤ 2kf k∞ .
(Idea: we now have control over the ∞-norm of (f ∗ Kn − f ). The previous
proposition gives us control over its 1-norm. Putting them together will give us
control over its 2-norm.)
Hence for all n ≥ 0, using Lemma A6.8 and Proposition B5.4,
p
kf ∗ Kn − f k2 ≤ kf ∗ Kn − f k1 kf ∗ Kn − f k∞
p p
≤ kf ∗ Kn − f k1 2kf k∞
→ 0 as n → ∞. 

40
B6 Summing the unsummable
For the lecture of Monday 24 February 2014

Before I explain the title, recall: we’re looking for a sequence K0 , K1 , . . . of

trigonometric polynomials that approximate the mythical δ function increas-
ingly well.
How can we find one?

Idea If δ existed, its Fourier coefficients would be given by

Z
δ̂(k) = δ(t)e−2πikt dt = e−2πik0 = 1
T
Pn
for all k ∈ Z, and so Sn δ = k=−n ek = Dn . If also Sn δ → δ as n → ∞, then
Dn → δ. Since Dn is a trigonometric polynomial, we might try Kn = Dn .

Problem (Dn ) is not a PAD, as noted in Example B5.3(ii). Also, the sequence
∞
(D
P∞ n )n=0 does not converge in any of our usual five senses. In other words,
k=−∞ ek is thoroughly unsummable.

So, we want to be able to sum unsummable series. This sounds impossible,

but it’s not: you just need to be more generous about what ‘summable’ means.
To explain this, it’s useful to step back from our particular problem—indeed,
step back from Fourier analysis entirely—and think about the general question:
how can we sum a divergent series?
Example B6.1 The series
∞
X
(−1)n = 1 − 1 + 1 − 1 + · · ·
n=0

PN n
does not converge, since the partial sums SN = n=0 (−1) are

S0 = 1, S1 = 0, S2 = 1, S3 = 0, ...,

and the sequence (SN ) does not converge. P∞ n

Nevertheless, there are non-rigorous ways of evaluating S = n=0 (−1) .
For instance:
i. Alice thinks that

S = (1 + −1) + (1 + −1) + (1 + −1) + · · · = 0 + 0 + 0 + · · · = 0.

Bob thinks that

S = 1 + (−1 + 1) + (−1 + 1) + · · · = 1 + 0 + 0 + · · · = 1.

Alice and Bob agree to split the difference, and so conclude that S = 1/2.

41
 ii. Consider two copies of S lined up in columns:

2S = (1 − 1 + 1 − 1 + · · · )
+ (1 − 1 + 1 − · · · )
= 1 + 0 + 0 + 0 + ···
= 1.

So S = 1/2, the same answer as obtained by Alice and Bob.

iii. Apply the formula for the sum of a geometric series (even though it’s
invalid, as that formula requires the ratio to have modulus strictly less
than 1):
1
S= = 1/2.
1 − (−1)
Again, this is the same answer.
Those three methods were just for fun. Each of them can actually be made
respectable, but I won’t show you how. Instead, here’s a fourth way, which is
the one that will matter to us the most.
iv. The sequence of partial sums (1, 0, 1, 0, . . .) doesn’t converge, but it oscil-
lates evenly around 1/2. The centre of mass is 1/2, if you like. Since a
centre of mass (or centroid) is a mean, this suggests considering the mean
of all the partial sums so far.
Put
1
AN = (S0 + · · · + SN ).
N +1
A short calculation shows that
(
1 1
+ if N is even,
AN = 12 2(N +1)
2 if N is odd.

So limN →∞ AN = 1/2, as expected.

There is some general terminology for this fourth method.

Definition B6.2 i. Let (sn )∞
n=0 be a sequence in C. Its N th Cesàro mean
(N ≥ 0) is
1
aN = (s0 + · · · + sN ) ∈ C.
N +1
If aN → s as N → ∞, we say that s is the Cesàro limit of (sn ).
P∞
ii. The Cesàro sum of a series n=0 xn is the Cesàro limit of the partial
PN
sums SN = n=0 xn , if it exists.

We have just seen an example of a series that is Cesàro-summable but not

summable in the usual sense. Now we show that the method of Cesàro sum-
mation extends the method of ordinary summation, in the sense that if the
ordinary sum exists then so does the Cesàro sum, and the two sums agree.
Proposition B6.3 i. Let (sn )∞
n=0 be a sequence in C, and let s ∈ C. If s is
the limit of (sn ) then s is also the Cesàro limit of (sn ).

42
 P∞
ii. Let n=0 xn be a series
P in C. If the sum exists, then the Cesàro sum
exists and is equal to xn .

Proof We prove (i); then (ii) follows immediately.

Let aN be the N th Cesàro mean of (sn ). For all N ≥ 0,
 
 (s0 − s) + · · · + (sN − s)  |s0 − s| + · · · + |sN − s|
|aN − s| = 
 ≤ .
N +1  N +1

(The idea now: we want |aN − s| to be small. Let L be a large integer. Then
|sn − s| is small when n ≥ L, and |s0 −s|+···+|s
N +1
L−1 −s|
is small when N is much
greater than L, since the numerator does not depend on N .)
Let ε > 0. Since sn → s, we can choose L such that |sn − s| < ε/2 for all
n ≥ L. We can then choose an integer
n 2
o
M ≥ max L, |s0 − s| + · · · + |sL−1 − s| − 1 .
ε
For all N ≥ M ,

|s0 − s| + · · · + |sL−1 − s| |sL − s| + · · · + |sN − s|

|aN − s| ≤ +
N +1 N +1
(ε/2)(M + 1) (N − L + 1)(ε/2)
< + (B:12)
N +1 N +1
≤ (ε/2) + (ε/2) = ε,

where in (B:12), we used the definition of M in the first summand and the
definition of L in the second. 
Pn
Remark B6.4 The partial sums Dn = k=−n ek are too wild to form a PAD.
The proposition we have just proved suggests that the Cesàro means N1+1 (D0 +
· · · + DN ) might be tamer. It turns out that they are; indeed, they form a PAD.
This will allow us to complete the plan described at the beginning of Section B5,
thus proving that in the 2-norm, Fourier series always converge.

43
B7 The Fejér kernel
For the lecture of Thursday 27 February 2014; part one of two

At the end of the last section, we expressed the hope that although the Dirichlet
kernels Dn do not form a PAD, perhaps their Cesàro means
1
(D0 + · · · + Dn )
n+1
do. Here we show that this is indeed the case.
Definition B7.1 Let n ≥ 0. The Fejér kernel of order n is
1
Fn = (D0 + · · · + Dn ) : T → C.
n+1
Note that the Fejér kernel, like the Dirichlet kernel, is a trigonometric poly-
nomial.
Way back in Section A1, we abandoned the sine-and-cosine formulation of
Fourier series, choosing to work with the more elegant exponential formulation.
But it will be useful to have expressions for Dn and Fn in traditional trigono-
metric form.
Lemma B7.2 Let n ≥ 0 and 0 6= t ∈ T. Then

sin((2n + 1)πt) 1 sin2 ((n + 1)πt)

Dn (t) = , Fn (t) = .
sin πt n+1 sin2 πt
Proof We have
n
X
Dn (t) = e1 (t)k .
k=−n

Since t 6= 0, we have e1 (t) 6= 1, and we may therefore apply the formula for
summing a geometric series. After some routine algebra, we get

e1 (t)n+1/2 − e1 (t)−(n+1/2)
Dn (t) = . (B:13)
e1 (t)1/2 − e1 (t)−1/2

Noting that e1 (t)α = e2πiαt and applying the formula eiθ = cos θ + i sin θ, we
obtain the result on Dn (t).
We now use a trick: multiply the top and bottom of (B:13) by the bottom.
This gives

e1 (t)n+1/2 − e1 (t)−(n+1/2) e1 (t)1/2 − e1 (t)−1/2

  
Dn (t) =  2
e1 (t)1/2 − e1 (t)−1/2
e1 (t)n+1 + e1 (t)−(n+1) − e1 (t)n + e1 (t)−n
   
=  2 .
e1 (t)1/2 − e1 (t)−1/2

44
 PN
So for any N ≥ 0, the sum n=0 Dn (t) telescopes, giving
N
e1 (t)N +1 + e1 (t)−(N +1) − e1 (t)0 + e1 (t)−0
   
X
Dn (t) =  2
n=0 e1 (t)1/2 − e1 (t)−1/2
2
e1 (t)(N +1)/2 − e1 (t)−(N +1)/2

=  2
e1 (t)1/2 − e1 (t)−1/2
[2i sin((N + 1)πt)]2
= ,
[2i sin(πt)]2

hence the result on FN (t). 

This explicit formula helps us to prove our main result on the Fejér kernel.
Proposition B7.3 (Fn )∞ n=0 is a PAD. In particular, there is a PAD consisting
of trigonometric polynomials.

Proof PAD1: ByR Lemma B7.2, FN (t) ≥ 0 for all N and t.

PAD2: First, T Dn (t) dt = 1 for all n, since
Z n
X
Dn (t) dt = hDn , e0 i = hek , e0 i = 1.
T k=−n

Hence
Z Z Z 
1
FN (t) dt = D0 (t) dt + · · · + DN (t) dt
T N +1 T T
1
= (1 + · · · + 1) = 1.
N +1
PAD3: Let δ ∈ (0, 1/2). We must prove that
Z
lim FN (t) dt = 0.
n→∞ δ<|t|<1/2

We use the fact that sin θ ≥ π2 θ for all θ ∈ [0, π/2]. (This can be proved using
convexity; see Figure B.5.) Now
1/2
sin2 ((N + 1)πt)
Z Z
2
FN (t) dt = dt (B:14)
δ<|t|≤1/2 N +1 δ sin2 πt
Z 1/2
2 1
≤ dt (B:15)
N +1 δ (2t)2
1  1 
= −2 (B:16)
2(N + 1) δ
→ 0 as N → ∞. (B:17)

Here (B:14) follows from Lemma B7.2 and FN being even, (B:15) is because
sin2 ((n + 1)πt) ≤ 1 and | sin πt| ≥ π2 · πt, and (B:16) is a routine calculation. 

45
 1

sin θ

2
πθ

0
0 π/2
2
Figure B.5: sin θ ≥ πθ for all θ ∈ [0, π/2].

46
B8 The main theorem
For the lecture of Thursday 27 February 2014; part two of two

We can now prove the main theorem of Part B. It states that Fourier’s idea
works perfectly when we use the 2- or 1-norm.
Theorem B8.1 Let f : T → C be an integrable function. Then Sn f → f in
both k · k2 and k · k1 .

Proof For all n, the Fejér kernel Fn is a trigonometric polynomial, so f ∗ Fn

is a trigonometric polynomial (Lemma B3.4). Also, (Fn ) is a PAD (Proposi-
tion B7.3), so f ∗ Fn → f in k · k2 (Proposition B5.5). Hence f is the limit in
k · k2 of a sequence of trigonometric polynomials. So by Lemma B1.3, Sn f → f
in k · k2 . Finally, by Lemma A4.10, Sn f → f in k · k1 . 

Corollary B8.2 The set {trigonometric polynomials} is dense in {integrable

functions T → C} with respect to both k · k2 and k · k1 .

Proof This follows from Theorem B8.1, noting that the Fourier partial sums
Sn f are trigonometric polynomials. 

In other words, if f is a point in the space of all integrable functions on

T, then there are trigonometric polynomials arbitrarily close to f . (Compare
Proposition A9.4 on step functions.)
Theorem B8.3 (Parseval) Let f : T → C be an integrable function. Then
v
u ∞ 
u X
fˆ(k)2 .

kf k2 = t
k=−∞

Proof Sn f → f in k·k2 , so kSn f k2 → kf k2 by Lemma A4.7, so kSn f k22 → kf k22 .

But
n
fˆ(k)2
X
kSn f k2 =
 
2
k=−n

by Lemma B1.1, so the result follows. 

In a suitable sense, Parseval’s theorem says that the map f 7→ fˆ is an

isometry (distance-preserving). It is also angle-preserving:
Corollary B8.4 Let f, g : T → C be integrable functions. Then
Z ∞
X
f (x)g(x) dx = fˆ(k)ĝ(k).
T k=−∞

Proof Sheet 3, q.5. 

Finally, the map f 7→ fˆ is, essentially, injective:

47
Corollary B8.5 Let f, g : T → C be integrable functions such that fˆ(k) = ĝ(k)
for all k ∈ Z. Then:

i. f = g a.e.;
ii. f (x) = g(x) for all x ∈ T such that f and g are both continuous at x;
iii. f = g if f and g are both continuous.

Proof We have f[ − g(k) = 0 for all k ∈ Z (using Lemma A7.7(i)), so kf −gk2 =

0 by Parseval’s theorem. All three parts now follow from Lemma A4.5. 

This result does not mention the 2-norm (or indeed any norm) in its state-
ment, although it does use the 2-norm in its proof. It answers a very fundamental
question about Fourier series, telling us:
Different functions have different Fourier series.

Here ‘different functions’ must be understood as ‘functions that are not almost
everywhere equal’, since it’s a basic fact that if f = g a.e. then f and g have
the same Fourier coefficients; see A3.13.
This encourages us to believe in Fantasy A8.7.

48
Chapter C

Uniform and pointwise

convergence of Fourier
series

In Remark A4.11, we met this diagram:

k · k∞ N
tt
ttt NNN
u} t #+
k · k2 pointwise
uu LLLLLLL
v~ uuuu LLL
"*
k · k1 a.e.

Part B covered the lower-left region of the diagram: convergence of Fourier series
in the 2- and 1-norms. In Part C, we look at the upper-right region: uniform
and pointwise convergence. (We will never look seriously at almost everywhere
convergence.)

C1 Warm-up
For the lecture of Monday 3 March 2014; part one of two

In Part B, we showed:
for all integrable f : T → C, Sn f → f in k · k2 and k · k1 .
Is it also true in k · k∞ ?
No, since if Sn f → f in k · k∞ then f is continuous (by Fact A4.8), whereas
not every integrable function is continuous.
But this leaves open the possibility that:
for all continuous f : T → C, Sn f → f in k · k∞ .
However, this is not true either. By du Bois–Reymond’s example (Theo-
rem A2.2), there is some continuous f : T → C such that Sn f does not even
converge pointwise to f , let alone uniformly.

49
 So if we’re hoping for all Fourier series to converge in either the uniform or
pointwise sense, we’ll be disappointed. But perhaps it’s true if we dilute our
hopes, asking only for convergence almost everywhere. In other words, we might
hope that:
for all continuous f : T → C, Sn f → f almost everywhere.
This is true, by Carleson’s theorem (Theorem A2.6). But as previously men-
tioned, all known proofs are much too hard for this course.
However, there are some easier results available. For example, we will be
able to prove the theorem of Dirichlet that if f is continuously differentiable,
then Sn f → f uniformly (hence in all five senses). We will also look at an
application that seems to have nothing to do with Fourier theory: the so-called
equidistribution theorem of Weyl, concerning the statistical behaviour of mul-
tiples of irrational numbers.

We begin by collecting a few basic facts.

As described in Fantasy A8.7, the theory of Fourier series is about the back-
and-forth between functions on T and doubly infinite sequences. So far, we have
concentrated on the passage in one direction, starting with a function f on T
and deriving the double sequence (fˆ(k))k∈Z . Let us now consider the opposite
direction. Write

CZ = double sequences c = (ck )∞


k=−∞ in C .

Lemma C1.1 Let c ∈ CZ . Then

∞
X ∞
X
|ck | < ∞ ⇒ |ck |2 < ∞ ⇒ sup |ck | < ∞.
k∈Z
k=−∞ k=−∞

(This should remind you of the situation with convergence in the 1-, 2- and
∞-norms—but it’s the other way round. Notice that squaring a small positive
number decreases it.)
P∞
Proof For the first implication, suppose that k=−∞ |ck | < ∞. Then certainly
the set {|ck | : k ∈ Z} is bounded; put kck∞ = supk∈Z |ck |. For all n ≥ 0, we
have
Xn n
X ∞
X
|ck |2 ≤ kck∞ |ck | ≤ kck∞ |ck |,
k=−n k=−n k=−∞
P∞
so k=−∞|ck |2 < ∞. P∞
For the second implication, if k=−∞ |ck |2 < ∞ then {|ck |2 : k ∈ Z} is
bounded, so {|ck | : k ∈ Z} is bounded too. 

Remark C1.2 No further implications hold. To see that the converse of the
first implication fails, consider
(
1/k if k ≥ 1
ck =
0 if k ≤ 0.

To see that the converse of the second implication fails, take ck = 1 for all k.

50
 Here is another basic fact. It is about convergence in the 1-norm, but will
be useful for arguments about uniform and pointwise convergence.

Pn C1.3 Let c ∈ C and let f : T → C be an integrable function. Suppose

Z
Lemma
that k=−n ck ek → f in k · k1 as n → ∞. Then ck = fˆ(k) for all k ∈ Z.
Pn
Proof For each n ≥ 0, write gn = k=−n ck ek . Let ` ∈ Z. By Lemma A7.7(ii),
gc ˆ
n (`) → f (`) as n → ∞. But by Lemma A8.3, gcn (`) = c` whenever n ≥ |`|, so
ˆ
c` = f (`), as required. 

We can now answer a subtle question that has been present from the begin-
ning. Can it happen that the Fourier series of f converges, but not to f ? For
convergence in the uniform sense, no.
Lemma C1.4 Let f : T → C be an integrable function. Suppose that the se-
quence (Sn f )∞
n=0 converges uniformly (not necessarily to f ). Then:

i. (Sn f )(x) → f (x) for almost all x ∈ T;

ii. (Sn f )(x) → f (x) for all x ∈ T such that f is continuous at x;
iii. Sn f → f uniformly if f is continuous.

Part (iii) says that for a continuous function f , if Sn f converges uniformly

to something, that something must be f .

Proof Let g be the uniform limit of the sequence (Sn f ). By Fact A4.8, g is
continuous. Certainly Sn f → g in k · k1 , so fˆ(k) = ĝ(k) for all k ∈ Z by
Lemma C1.3 (taking ‘ck ’ to be fˆ(k) and ‘f ’ to be g). The result now follows
from Corollary B8.5. 

51
C2 What if the Fourier coefficients are abso-
lutely summable?
For the lecture of Monday 3 March 2014; part two of two

P∞
We saw in Proposition B1.2 that k=−∞ |fˆ(k)|2 < ∞ for any integrable func-
P∞ f : T ˆ→ C. However, f may or may not have the stronger property that
tion
k=−∞ |f (k)| < ∞. (To see why this is stronger, recall Lemma C1.1.)
P∞ ˆ
In this section, we will see that if k=−∞ |f (k)| < ∞ then the Fourier series
of f behaves well, in a sense that will be made precise. In the next section, we
will see that many common functions do have this property.
We begin with a result on double sequences.
P∞
Proposition C2.1 Let c ∈ CZ . Suppose that k=−∞ |ck | < ∞. Then the
P ∞
n
sequence k=−n ck ek converges uniformly to a continuous function g.
n=0
Moreover, ĝ(k) = ck for all k ∈ Z.
Pn
Proof For n ≥ 0, write sn = k=−n ck ek .
First we show that the sequence (sn ) converges
P pointwise. Indeed, let x ∈ T.
Given ε > 0, we can choose N such that k : |k|≥N |ck | < ε; then for all m ≥
n ≥ N,
 
 X  X
|sm (x) − sn (x)| =  ck ek (x) ≤ |ck | < ε.
 
 
k : n<|k|≤m k : n<|k|≤m

So the sequence (sn (x))∞

n=0 is Cauchy, and therefore convergent, as claimed.
Next, for each x ∈ T, put
∞
X
g(x) = lim sn (x) = ck ek (x).
n→∞
k=−∞

This defines a function g : T → C. Now (sn ) converges uniformly to g, since

 
 X  X
ksn − gk∞ = sup ck ek (x) ≤ |ck | → 0
 
x∈T  
k : |k|>n k : |k|>n

as n → ∞. By Fact A4.8, g is therefore continuous, and by Lemma C1.3,

ĝ(k) = ck for all k ∈ Z. 
We now come to the most important result of Part C so far.
Theorem C2.2 Let f : T → C be an integrable function. Suppose that
P∞ ˆ
k=−∞ |f (k)| < ∞. Then:

i. (Sn f )(x) → f (x) as n → ∞ for all x ∈ T such that f is continuous at x;

ii. Sn f → f uniformly if f is continuous.
Proof By Proposition C2.1 (taking ck to be fˆ(k)), the Fourier partial sums Sn f
converge uniformly to some continuous function g : T → C, and ĝ(k) = fˆ(k) for
all k ∈ Z. The result follows from Corollary B8.5. 

52
C3 Continuously differentiable functions
For the lecture of Thursday 6 March 2014; part one of two

P∞
We have just shown that if k=−∞ |fˆ(k)| < ∞ then the Fourier series of
f behaves well (Theorem C2.2). If you had a particular function f in front of
you, you might be able to calculate its Fourier coefficients, you might be able
to calculate the sum of their absolute values, and that sum might be finite.
In that case, you could conclude that the Fourier series of that particular f
was well-behaved. But are there general conditions on f guaranteeing that
P ∞ ˆ
k=−∞ |f (k)| < ∞?
The answer turns
P out to be yes. But beware that not all integrable functions
f : T → C satisfy |fˆ(k)| < ∞. In fact, there exist quite ‘nice’ functions f
P ˆ
with |f (k)| = ∞:
Example C3.1 Define f : T → C by
X sin(2πnx)
f (x) = .
n log n
n≥2

P ˆthat Sn f → f uniformly, and so f is continuous. One can also

One can show
show that |f (k)| = ∞. (The proof of both these statements is omitted and
P ˆ
non-examinable.) So a continuous function f need not satisfy |f (k)| < ∞.

However, we will seePthat if f is smooth enough, the Fourier coefficients

decay fast enough that |fˆ(k)| converges.
Definition C3.2 Let n ≥ 0. We write C n (T) for the set of functions T → C
whose corresponding 1-periodic function g : R → C is n times continuously
differentiable (that is, g (n) exists and is continuous). We also write

C(T) = C 0 (T) = {continuous functions T → C}.

For example, a function T → C belongs to the set C 1 (T) if and only if the
corresponding 1-periodic function R → C is continuously differentiable.

Lemma C3.3 Let n ≥ 0, f ∈ C n (T), and k ∈ Z. Then

(n) (k) = (2πik)n fˆ(k).

fd

Proof Sheet 3, q.6(ii). 

In other words, differentiating f amounts to multiplying its kth Fourier

coefficient by 2πik.
We can now prove the 1829 theorem of Dirichlet, one of the early landmarks
of Fourier analysis. In fact, what we prove is stronger than the version stated
earlier (Theorem A2.1): there, we only asserted pointwise convergence, but here
we prove uniform convergence.
P∞
Theorem C3.4 (Dirichlet) Let f ∈ C 1 (T). Then k=−∞ |fˆ(k)| < ∞ and
Sn f → f uniformly.

53
 Since uniform convergence is the strongest kind of convergence, this tells us
that the Fourier series of a continuously differentiable function converges in all
five senses. Thus, it behaves as well as any function could.

Proof For k 6= 0, we have

1
|fˆ(k)| = |fc0 (k)| · (C:1)
2π|k|
1  c0 1 
≤ |f (k)|2 + (C:2)
2 (2π|k|)2
2
|fc0 (k)| 1
= + 2 2, (C:3)
2 8π k
where in (C:1) we used Lemma C3.3 and in (C:2) we used the fact that uv ≤
1 2 2
P∞ ˆ
2 (u + v ) for all u, v ∈ R (proof: exercise). So to show that k=−∞ |f (k)| is
finite, it suffices to show that the sum of (C:3) over all k ∈ Z is finite. First, f 0 is
P∞
continuous and so integrable; hence k=−∞ |fc0 (k)|2 < ∞ by Proposition B1.2.
Second,
X 1 X 1
2
=2 < ∞.
k k2
06=k∈Z k≥1
P∞
Hence k=−∞ |fˆ(k)| < ∞, giving Sn f → f uniformly by Theorem C2.2. 

Remark C3.5 We have just shown that

X
f ∈ C 1 (T) ⇒ |fˆ(k)| < ∞.

But the converse fails, even for continuous functions. Here is an example due
to Weierstrass:
∞
X
f (x) = 2−n cos(2n πx).
n=1
P ˆ
It can be shown (non-examinably) that |f (k)| < ∞ and that f is continu-
ous. But it can also be shown that f is not differentiable anywhere, let alone
continuously differentiable everywhere.

54
C4 Fejér’s theorem
For the lecture of Thursday 6 March 2014; part two of two

Du Bois–Reymond’s example (Theorem A2.2) tells us that for
a continuous

∞
function f : T → C and a point x ∈ T, the sequence (Sn f )(x) n=0 need not
have a limit. However, Fejér showed that it always has a Cesàro limit, namely,
f (x). His result is the centrepiece of this section.
In Part B, we proved that Sn f → f in k · k2 by proving that f ∗ Fn → f in
k · k2 . We did that by proving that (i) (Fn ) is a PAD, and (ii) f ∗ Kn → f in
k · k2 for any PAD (Kn ). We adopt a similar strategy here, but with pointwise
and uniform convergence in place of mean square convergence.
For integrable f : T → C, we have
1
f ∗ Fn = f ∗ (D0 + · · · + Dn )
n+1
1
= (f ∗ D0 + · · · + f ∗ Dn )
n+1
1
= (S0 f + · · · + Sn f ).
n+1
This is nothing but the nth Cesàro mean of the sequence S0 f, S1 f, . . .. We will
write
1
An f = (S0 f + · · · + Sn f ) = f ∗ Fn .
n+1
To say that f (x) is the Cesàro limit of the sequence ((Sn f )(x)) is to say that
(An f )(x) → f (x) as n → ∞, or equivalently that (f ∗ Fn )(x) → f (x) as n → ∞.
Here is a summary of the notation and terminology:
Ordinary convergence Cesàro convergence
nth term in sequence Sn f An f
nth kernel Dn (Dirichlet) Fn (Fejér)
To prove Fejér’s theorem, we will need to know a little more about PADs.
We showed in Section B5 that for any PAD (Kn ) and integrable function f , we
have f ∗ Kn → f in both k · k2 and k · k1 . That result cannot be strengthened
to convergence in k · k∞ for arbitrary integrable functions f , since each f ∗ Kn
is continuous and a uniform limit of continuous functions is continuous. So the
best we could hope for is that if f is continuous then f ∗ Kn → f in k · k∞ .
We might also hope to be able to prove a result about convergence at particular
points. Both these hopes are fulfilled:
Proposition C4.1 Let (Kn ) be a PAD and f : T → C an integrable function.
Then:
i. (f ∗ Kn )(x) → f (x) as n → ∞ for all x ∈ T such that f is continuous at
x;
ii. f ∗ Kn → f uniformly if f is continuous.

The proof is similar to the proof of Proposition B5.4, the corresponding

result for the 1-norm.

55
Proof For (i), let x ∈ T. In the proof of Proposition B5.4, equation (B:7)
stated that for all n ≥ 0 and δ ∈ (0, 1/2),
Z Z
|(f ∗Kn )(x)−f (x)| ≤ |f (x−t)−f (x)|Kn (t) dt+2kf k∞ Kn (t) dt.
|t|<δ δ<|t|≤1/2

Suppose that f is continuous at x. Let ε > 0. Choose δ ∈ (0, 1/2) such that
|f (x − t) − f (x)| < ε/2 for all t ∈ (−δ, δ). By PAD3, we can also choose N such
that for all n ≥ N , Z
ε
Kn (t) dt < .
δ<|t|≤1/2 4kf k∞
Then for all n ≥ N ,
Z
ε ε
|(f ∗ Kn )(x) − f (x)| ≤ Kn (t) dt + 2kf k∞ ·
2 |t|<δ 4kf k∞
Z 1/2
ε ε
≤ Kn (t) dt + = ε,
2 −1/2 2

as required.
For (ii), suppose that f is continuous. Let ε > 0. Every continuous map
from a compact metric space to a metric space is uniformly continuous, and T
is compact, so f is uniformly continuous. Hence we can choose δ ∈ (0, 1/2) such
that for all x ∈ T and t ∈ (−δ, δ), we have |f (x − t) − f (x)| < ε/2.
(In case you have not previously encountered this theorem about metric
spaces, an alternative argument is available. Here we use instead the theorem
that any continuous function on a closed bounded interval is uniformly contin-
uous. Since f is continuous on [−1, 1], it is uniformly continuous there. So by
interpreting T as [−1/2, 1/2), we can choose δ as in the last paragraph.)
By PAD3, we can choose N as in the proof of (i). For all x ∈ T, for all
n ≥ N , we calculate that |(f ∗ Kn )(x) − f (x)| < ε, again as in the proof of (i).
So kf ∗ Kn − f k∞ ≤ ε for all n ≥ N , as required. 

Theorem C4.2 (Fejér) Let f : T → C be an integrable function. Then:

i. (An f )(x) → f (x) as n → ∞ for all x ∈ T such that f is continuous at x;
ii. An f → f uniformly if f is continuous.

Proof (Fn ) is a PAD and An f = f ∗ Fn , so this follows from the last proposi-
tion. 

Looking back at du Bois–Reymond’s example (Theorem A2.2), we see that

the sequence (An f ) is more likely to converge than the sequence (Sn f ).

Remark C4.3 If f, g : T → C with fˆ(k) = ĝ(k) for all k, then An f = An g for

all n, so by Fejér’s theorem, f (x) = g(x) for all x ∈ T such that f and g are both
continuous at x. This gives an alternative proof of Corollary B8.5(ii), (iii). But
unlike that corollary, Fejér’s theorem gives an explicit way of reconstructing
f from its Fourier coefficients (at least for continuous f ): from knowing the
Fourier coefficients, we can compute An f for each n, and we then obtain f (x)
as limn→∞ (An f )(x).

56
Corollary C4.4 The set {trigonometric polynomials} is dense in C(T) with
respect to k · k∞ .

Proof Let f ∈ C(T). For each n ≥ 0, the function An f is a trigonometric

polynomial, since S0 f, . . . , Sn f are. The result follows from part (ii) of Fejér’s
theorem. 

By Lemma A4.10(i), density with respect to k · k∞ is the strongest type of

density (that is, stronger than density with respect to k · k2 or k · k1 ). Corol-
lary C4.4 can be compared to Corollary B8.2, which stated that the set of
trigonometric polynomials is dense in the set of integrable functions with re-
spect to k · k2 (hence k · k1 ).
(Faced with this, it is natural to ask whether Corollary B8.2 can be improved
to density with respect to k · k∞ . The answer is no, again because a uniform
limit of continuous functions is continuous.)

Corollary C4.5 (Weierstrass approximation theorem) Let I ⊆ R be a

closed bounded interval and f : I → C a continuous function. Then for all
ε > 0, there exists a polynomial p over C such that supx∈I |f (x) − p(x)| ≤ ε.

Proof Sheet 4. 

This result is important in many ways. For example, it is a basic fact in

numerical analysis, where we might wish to approximate an arbitrary continuous
function by a polynomial. However, we will not need it for anything in the rest
of this course.

57
C5 Differentiable functions and the Riemann lo-
calization principle
For the lecture of Monday 10 March

This section contains two major results. First, there is a convergence theorem
for differentiable functions, to add to our existing convergence theorem for con-
tinuously differentiable functions (Dirichlet’s Theorem C3.4). Then, we meet
the deep and shocking ‘localization principle’ of Riemann.
Dirichlet’s theorem states that if f is continuously differentiable then Sn f →
f uniformly. We will weaken both the hypothesis and the conclusion, proving
that if f is differentiable (but not necessarily continuously so) then Sn f → f
pointwise (but not necessarily uniformly).
To prove this, we need another fact about integration.
Fact C5.1 Let I ⊆ R be a bounded interval and x ∈ I. Let g : I → C be a
bounded function such that for all δ > 0, the restriction g|I\(x−δ,x+δ) is inte-
grable. Then g is integrable.

The proof is omitted and non-examinable, but can be found in the appendix
of the book Fourier Analysis by Stein and Shakarchi.
Theorem C5.2 Let f : T → C be an integrable function. Then:
i. (Sn f )(x) → f (x) for all x ∈ T such that f is differentiable at x;
ii. Sn f → f pointwise if f is differentiable.

Proof We just have to prove (i), since this immediately implies (ii). Take x ∈ T
such that f is differentiable at x.
For all n ≥ 0,
Z Z
f (x) − (Sn f )(x) = Dn (t)f (x) dt − Dn (t)f (x − t) dt (C:4)
T T
en+1 (t) − e−n (t)
Z
= (f (x) − f (x − t)) dt (C:5)
e1 (t) − 1
ZT
= (en+1 (t) − e−n (t))g(t) dt, (C:6)
T
R
where in (C:4) we used the fact that T Dn (t) dt = 1, in (C:5) we used equa-
tion (B:13) (from the proof of Lemma B7.2), and in (C:6) we put

f (x) − f (x − t)
g(t) =
e1 (t) − 1

for 0 < |t| ≤ 1/2. If g is integrable then

f (x) − (Sn f )(x) = ĝ(−(n + 1)) − ĝ(n) → 0 − 0 = 0

as n → ∞, by the Riemann–Lebesgue lemma (Proposition B1.2), completing

the proof. So it is enough to prove that g is integrable. To do this, we will need
the assumption that f is differentiable at x, which we have not used yet.

58
 We prove that g is integrable using Fact C5.1.
First we show that for each δ > 0, the function g|[−1/2,1/2)\(−δ,δ) is integrable.
Let δ > 0. By Lemma A3.7, it is enough to show that the restrictions of the
functions
t 7→ 1/(e1 (t) − 1), t 7→ f (x) − f (x − t)
to [−1/2, 1/2) \ (−δ, δ) are both integrable. And indeed, the first restricted
function is integrable because it is continuous and bounded, and the second is
integrable because f is.
Next we show that g is bounded. For 0 < |t| ≤ 1/2, we have

f (x) − f (x − t) t 1
g(t) = · → f 0 (x) · as t → 0,
t e1 (t) − 1 2πi

using l’Hôpital’s rule in the second factor. Define g(0) = f 0 (x)/2πi; then g is
continuous at 0, so there exists η > 0 such that g|(−η,η) is bounded. But also,
g|[−1/2,1/2)\(−η,η) is bounded (since it is integrable). Hence g itself is bounded.
It follows from Fact C5.1 that g is integrable, as required. 

It once seems to have been believed that although a continuous function

need not be differentiable everywhere, it must be differentiable somewhere. If
this were the case, Theorem C5.2 would tell us that the Fourier series of a
continuous function f must converge to f (x) for at least one value of x. However,
Weierstrass’s example of a continuous function that is nowhere differentiable
(Remark C3.5) shows that this strategy will not work.
We now use Theorem C5.2 to deduce the second major result of this sec-
tion, the Riemann localization principle. A good way to prepare for this is to
look back at Section A1 (the algebraist’s dream), and especially the two points
labelled ‘7’ there.
Remark C5.3 The Taylor series
∞
X f (n) (0) n
(T f )(x) = x
n=0
n!

of an infinitely differentiable function f is ‘determined locally’ at 0. That is, if

f, g : R → C with f |(−δ,δ) = g|(−δ,δ) for some δ > 0, then T f = T g. In contrast,
the Fourier series of f : C → C is not determined locally at 0 (or indeed at any
other point), since each coefficient
Z
ˆ
f (k) = f (t)e−2πikt dt
T

involves all of f . Thus, given f, g : T → C with f |(−δ,δ) = g|(−δ,δ) for some

δ > 0, we can not deduce that Sf = Sg.
The next result therefore comes as a surprise.

Corollary C5.4 (Riemann localization principle) Let f, g : T → C be in-

tegrable functions and x ∈ T. Suppose there exists δ > 0 such that f |(x−δ,x+δ) =
g|(x−δ,x+δ) . Then

(Sn f )(x) → f (x) as n → ∞ ⇐⇒ (Sn g)(x) → g(x) as n → ∞.

59
 In other words, whether or not the Fourier series of f at x converges to f (x)
depends only on the behaviour of f near x.

Proof The integrable function f − g : T → C has constant value 0 on some

neighbourhood of x, so is differentiable at x. Hence by Theorem C5.2,

(Sn (f − g))(x) → (f − g)(x) as n → ∞.

Equivalently,
(Sn f )(x) − (Sn g)(x) → 0 as n → ∞.
The result follows. 

The theme of ‘local versus global’ becomes increasingly visible at this level
of mathematics. You may have met it in differential geometry; for example,
a surface is locally like R2 , but globally may be quite unlike it. In number
theory, there is the local-to-global principle of Hasse, which has to do with p-adic
numbers and is helpful in determining the solvability of polynomial equations
over the integers. The interplay between local and global is captured formally
by the notion of sheaf, which is especially prominent in algebraic geometry.

60
 x1 x3

x2

Figure C.1: The beginning of the sequence (xn )∞

3 
n=1 , where xn = 8n .

C6 Weyl’s equidistribution theorem

For the lecture of Thursday 13 March

This section is an application of our theory. We will use what we have learned
to solve a problem that appears to have nothing to do with Fourier analysis.
Question:

Which sequences x1 , x2 , . . . in [0, 1) are ‘evenly spread’ ?

More fundamentally:
What does it mean for a sequence x1 , x2 , . . . in [0, 1) to be ‘evenly
spread’ ?

We answer the second question immediately. Write #S for the cardinality of a

finite set S. (You might be more used to |S|, but that could be confused with
the notation |I| for the length of an interval I.)
Definition C6.1 A sequence (xn )∞
n=1 in [0, 1) is equidistributed if for all
intervals I ⊆ [0, 1),
1 
· # j ∈ {1, . . . , n} : xj ∈ I → |I| as n → ∞.
n
So, roughly speaking, for (xn ) to be equidistributed means that for large n,
about 1/2 of x1 , . . . , xn are in (0, 1/2], about 1/5 are in [3/5, 4/5], and so on.
Given x ∈ R, write hxi = x − bxc. Thus, hxi is the unique element of [0, 1)
such that x−hxi ∈ Z (or equivalently, such that hxi represents the same element
of T as x).
i. For n ≥ 1, put xn = 38 n (Figure C.1). Then xn ∈


Examples C6.2
{0, 1/8, . . . , 7/8} for all n, so
1 
· # j ∈ {1, . . . , n} : xj ∈ (0, 1/8) = 0
n
for all n, whereas |(0, 1/8)| = 1/8. So (xn ) is not equidistributed.

∞
ii. Similarly, for any α ∈ Q, the sequence hnαi n=1 is not equidistributed.

61
 So far, we haven’t seen any examples of equidistributed sequences, only non-
examples. It’s not obvious that there are any equidistributed sequences at all!
But using Fourier analysis, we’ll construct lots.
When we defined equidistributed sequence, we were choosing a precise mean-
ing for the phrase ‘evenly spread’. But here’s a different, less obvious interpre-
tation. We could say that a sequence (xn ) in [0, 1) is ‘evenly spread’ if for all
integrable functions f : T → C,
n Z 1
1X
f (xj ) → f (t) dt as n → ∞. (C:7)
n j=1 0

Think for a while about why this is a sensible idea—understanding the rest of
this section depends on it.
There are variations on this idea in which we use different classes of func-
tions. For instance, we could regard (xn ) as ‘evenly spread’ if (C:7) holds for
all continuous (rather than integrable) functions f , or all step functions, or all
characters, etc. Actually, one of these variations produces the notion of equidis-
tributed sequence itself:
Lemma C6.3 Let (xn )∞ n=1 be a sequence in [0, 1). Then (xn ) is equidistributed
if and only if (C:7) holds whenever f is the characteristic function of an interval
in [0, 1).

Proof For an interval I ⊆ [0, 1), and for n ≥ 1, we have

n
1X 1 
χI (xj ) = # j ∈ {1, . . . , n} : xj ∈ I
n j=1 n

and Z 1
χI (t) dt = |I| .
0
The result follows. 

We now show that all these notions of ‘evenly spread’ sequence produced by
choosing different classes of function are, in fact, the same.
Theorem C6.4 Let (xn )∞
n=1 be a sequence in [0, 1). The following are equiva-
lent:
i. (C:7) holds for all integrable f : T → C;

ii. (C:7) holds for all characters f = ek : T → C, where k ∈ Z;

iii. (C:7) holds for all trigonometric polynomials f : T → C;
iv. (C:7) holds for all continuous f : T → C;
v. (C:7) holds for all characteristic functions f = χI : T → C, where I ⊆
[0, 1) is an interval;
vi. (C:7) holds for all step functions f : T → C.

62
Proof We show that (i) ⇒ (ii) ⇒ (iii) ⇒ (iv) ⇒ (v) ⇒ (vi) ⇒ (i).
(i) ⇒ (ii): trivial. Pm
(ii) ⇒ (iii): assume (ii). Let f = k=−m ck ek be a trigonometric polynomial.
Then
m m n m Z 1 Z 1
1X X 1X X
f (xj ) = ck · ek (xj ) → ck ek (t) dt = f (t) dt
n j=1 n j=1 0 0
k=−m k=−m

as n → ∞.
(iii) ⇒ (iv): assume (iii). Let f : T → C be a continuous function. Let
ε > 0. By Corollary C4.4, we can find a trigonometric polynomial g such that
kf − gk∞ < ε/3. By assumption, we can choose N such that for all n ≥ N ,
 n Z 1 
1 X 

n g(xj ) − g(t) dt < ε/3.
j=1 0

Then for all n ≥ N , using the triangle inequality,

 n Z 1   n n   n Z 1 
1 X  1 X 1X  1 X 

n f (xj ) − f (t) dt ≤ 
  f (xj ) − g(xj ) + 
  g(xj ) − g(t) dt
j=1 0 n j=1 n j=1 n j=1 0
Z 1 Z 1 
 
+  g(t) dt − f (t) dt
0 0
n
1X
< kf − gk∞ + ε/3 + kf − gk1
n j=1
≤ kf − gk∞ + ε/3 + kf − gk∞ < ε.

(iv) ⇒ (v): assume (iv). Let I ⊆ [0, 1) be an interval, and let ε > 0. We
can find a continuous function gU : [0, 1) → R such that gU (t) ≥ χI (t) for all
t ∈ [0, 1) and
Z 1 Z 1
gU (t) dt − χI (t) dt < ε/2
0 0

(Figure C.2). By assumption, we can choose N ≥ 1 such that for all n ≥ N ,

 n Z 1 
1 X 

n gU (xj ) − gU (t) dt < ε/2.
j=1 0

So for all n ≥ N ,
n Z 1
1X
χI (xj ) − χI (t) dt (C:8)
n j=1 0
n Z 1
1X
≤ gU (xj ) − χI (t) dt
n j=1 0
 X n Z 1  Z 1 Z 1 
1
= gU (xj ) − gU (t) dt + gU (t) dt − χI (t) dt
n j=1 0 0 0

< ε/2 + ε/2 = ε.

63
 χI

gU

Figure C.2: The characteristic function χI of an interval, and a continuous

function gU approximating it. (It is intended here that gU (x) = χI (x) for most
values of x; the graphs are drawn a small distance apart for visual clarity only.)

Similarly, by approximating χI from below instead of above, (C:8) is greater

than −ε.
(v) ⇒ (vi): this is the same argument as in (ii) ⇒ (iii). (Every step function
is a finite linear combination of characteristic functions of intervals, just as every
trigonometric polynomial is a finite linear combination of characters.)
(vi) ⇒ (i): assume (vi). First take a real -valued integrable function f : T →
R. Let ε > 0. In the proof of Proposition A9.4, we showed that there is a
step function gL (called g there) such that gL (t) ≤ f (t) for all t ∈ [0, 1) and
R1 R1
0
f (t) dt − 0 gL (t) dt < ε/2. Similarly, there is a step function gU such that
R1 R1
gU (t) ≥ f (t) for all t ∈ [0, 1) and 0 gU (t) dt − 0 f (t) dt < ε/2. Arguing as in
Pn R1
(iv) ⇒ (v), we get n1 j=1 f (xj ) → 0 f (t) dt as n → ∞.
Now take any integrable function f : T → C. Then f = f1 + if2 where
f1 , f2 : T → R are integrable functions. The result follows by linearity, using the
argument of (ii) ⇒ (iii) again. 
Corollary C6.5 Let (xn )∞ n=1 be a sequence in [0, 1). Then (xn ) is equidis-
tributed if and only if it satisfies Weyl’s criterion:
n
1 X 2πikxj
for each integer k 6= 0, e → 0 as n → ∞.
n j=1

Proof This is essentially (v) ⇐⇒ (ii) of Theorem C6.4. Indeed, condition (v)
is equivalent to (xn ) being equidistributed, by Lemma C6.3. Condition (ii) holds
if and only if for each k ∈ Z,
n Z 1
1 X 2πikxj
e → ek (t) dt as n → ∞. (C:9)
n j=1 0

When k = 0, (C:9) holds for any sequence (xn ) (check!). On the other hand,
R1
e (t) dt = hek , e0 i = 0 whenever k 6= 0. The result follows.
0 k

This enables us to find, at last, some examples of equidistributed sequences.
Corollary C6.6 (Weyl’s

∞ equidistribution theorem) Let α ∈ R \ Q. Then
the sequence hnαi n=1 in [0, 1) is equidistributed.
Proof We verify Weyl’s criterion. Let 0 6= k ∈ Z. For all n ≥ 1,
n n
1 X 2πikhjαi 1 X 2πikjα
e = e
n j=1 n j=1

64
since hjαi − jα ∈ Z. We sum this geometric series using the standard formula,
which is valid as long as the ratio is not equal to 1. Here, this means e2πikα 6= 1,
or equivalently kα 6∈ Z, which is true as k 6= 0 and α 6∈ Q. Thus, the formula is
valid and the sum of the geometric series is

1 2πikα e2πiknα − 1
e .
n e2πikα − 1
So for all n ≥ 1,

 1 n 2πikhjαi 
 X   2πiknα 
 + |1|
 ≤ · e
1 1 2

e  n e2πikα − 1 = n · e2πikα − 1 → 0

n
j=1

as n → ∞. 

Example C6.7 Consider the sequence

e = 2.718 . . . , 2e = 5.436 . . . , 3e = 8.154 . . . , . . . .

∞
The sequence hnei n=1 of fractional parts is

0.718 . . . , 0.436 . . . , 0.154 . . . , . . . .

Weyl’s equidistribution theorem tells us that ‘in the long run’, about 1/100 of
these fractional parts lie between 0.12 and 0.13 (for instance). In particular,
there are infinitely many natural numbers n such that

0.12 ≤ hnei < 0.13.

Could you prove this without Weyl’s theorem? Could you prove that there is
even one number n with this property?

65
Chapter D

Duality

This course is called Fourier Analysis, but we have focused throughout on

Fourier series. You’re aware that Fourier transforms are also important. What
is the relationship between the two?
Section D1 gives the answer. We’ll see there that by looking at Fourier
analysis from sufficiently high altitude—by taking a bird’s eye view—it becomes
apparent that Fourier series and Fourier transforms are two special cases of the
same general construction.
But it also becomes apparent that there are other interesting special cases.
One case that is both interesting and easy is Fourier analysis on finite abelian
groups. That’s what we’ll study for the rest of Part D.

D1 An abstract view of Fourier analysis

For the lecture of Monday 17 March

This section (D1) is not directly examinable, but reading it is recommended: it

should help you to understand the ideas behind Part D and how it relates to the
rest of the course.

The right general context for Fourier analysis is that of topological groups. To
understand what a topological group is, you first need to know roughly what
a topological space is. So here is a short, informal introduction to topological
spaces.
Roughly speaking, a set X is called a topological space if we know what
it means to ‘move gradually’ within X. For instance, Rn is a topological space
because we know what it means for one point to be close to another, and we
know what a ‘continuous path’ in Rn is (namely, a continuous map [0, 1] → Rn ).
Similarly, the sphere is a topological space: you know what it means to move
gradually on the surface of the earth. The circle T is a topological space for the
same reason.
Some non-mathematical examples to help your intuition: the set of all
colours is a topological space, because you know what it means for a colour
to change gradually. The set of all possible human faces is a topological space,
because you know what it means for a face to change gradually (e.g. as you

66
age) or for one face to be similar to another. Your own changing face defines a
continuous map f from [0, 1] to the space of all faces, with f (0) being your face
at birth and f (1) your face at death.
Formally, a topological space is a set X equipped with some extra data
(specifying what it means to ‘move gradually’). That extra data is called a
‘topology’ on X. We will not need the formal definition.
Some topological spaces are rather trivial. For instance, you can’t move
gradually from one integer to another without passing through non-integers, so
the set Z is a topological space in a trivial way: the only way of moving gradually
is to stay still. Formally, this is called the discrete topology on Z. (If you know
the definition of topological space, the discrete topology is the topology in which
all subsets are open.) Any set can be given the discrete topology.
Topological spaces are the right context for the definition of continuous map.
You know what continuity means for a function R → R, and more generally for
a function Rn → Rm , and also for a function T → C. Given topological spaces
X and Y , there is a definition of what it means for a function X → Y to be
continuous. The definitions for Rn , T, C, etc. are all special cases of this general
definition.
Roughly speaking, a topological group is a topological space that is also
a group. We are most interested in those topological groups that are abelian
(that is, the multiplication is commutative) and ‘locally compact’ (a condition
that I will not explain; it is satisfied by all the examples I will mention). So,
what we are interested in is locally compact, abelian topological groups, which
are usually just called ‘locally compact abelian groups’ or LCAGs for short.
Examples D1.1 i. Rn is a LCAG (with addition as the group operation),
for any n ≥ 0.
ii. The circle T is a LCAG too. Recall that T is the quotient group R/Z, so
that its group operation is addition too.
iii. Z is a LCAG, with + as the group operation and the discrete topology.
iv. Any finite abelian group is a LCAG, with the discrete topology.
v. S = {z ∈ C : |z| = 1} is a LCAG, with · as the group operation. In fact,
T∼= S, via the map t 7→ e2πit .
Definition D1.2 Let G be a LCAG. A character of G is a continuous group
homomorphism G → S.
Examples D1.3 i. For each ξ ∈ Rn , there is a character eξ of Rn defined
by
eξ (x) = e2πiξ.x
(x ∈ Rn ), where ξ.x is the dot product of ξ and x. (This is the ordinary
product in the case n = 1.) You know that eξ is continuous, and you can
easily check that eξ is a group homomorphism (try it!). Fact: these are
the only characters of Rn .
ii. For each k ∈ Z, there is a character ek of T defined in the usual way.
Lemma A7.2 states exactly that each ek is a character in the sense of Def-
inition D1.2. Fact: these are the only characters of T. So, the characters
of T in our new sense are precisely the characters of T in our old sense.

67
 iii. For each t ∈ T, there is a character εt of Z defined by

εt (k) = e2πikt

(t ∈ T). Again, you can easily check that each εt really is a character,
that is, a continuous homomorphism. Fact: these are the only characters
of Z.

The set
b = {characters of G}
G
forms a LCAG in a natural way. I won’t describe the topology, but the group
structure is given ‘pointwise’: if e1 and e2 are characters of G then the character
e1 · e2 is defined by (e1 · e2 )(x) = e1 (x) · e2 (x) (x ∈ G). The hat notation is
related to, but different from, the hat notation fˆ for the Fourier coefficients or
Fourier transform of a function f .
Examples D1.3 tell us what G b is for various LCAGs G:

G G
b
Rn Rn
T Z
Z T
b∼
For instance, T = Z because we have an isomorphism

Z → T b
k 7→ ek .

We saw in Lemma A7.3 that this map is a homomorphism, and in Exam-

ple D1.3(ii) that this map is surjective. It is not hard to see that it is also
injective (that is, ek and e` are different if k 6= `). So, it is an isomorphism of
groups, and in fact an isomorphism of topological groups (that is, a homeomor-
phism too).
In each of the three cases in the table, G b ∼
b
= G. In fact, this is a general
phenomenon:

b∼
Theorem D1.4 (Pontryagin duality) G
b
= G for all LCAGs G.

The group G b can be thought of as the ‘mirror image’ of the group G (Fig-
ure D.1). Pontryagin duality states that the mirror image of the mirror image
of G is G itself.

We now look briefly at Fourier analysis on locally compact abelian groups.

Let G be a LCAG. ‘Nice’ functions on G can be integrated. For example,
you’re used to integrating nice functions on Rn or T. In the case of Z, a function
Z → C is just a double sequence (ck )∞ k=−∞ in C, and integrating it means taking
its sum (which is possible if it’s ‘nice’ enough).
We can define the Fourier transform of a nice function on G, and also the
inverse Fourier transform of a nice function on G: b

(ˆ)
{nice functions on G} o / b ,
{nice functions on G}
(ˇ)

68
 ordinary Fourier transforms

R
R2
R3

T Z

Fourier series
G G
b

finite
abelian
rest of Part D groups

Locally compact abelian groups

Figure D.1: The world of locally compact abelian groups. Think of the dotted
line down the middle as a mirror. Reflecting in the mirror swaps G with G.
b

as follows: for a ‘nice’ function f : G → C, we define its Fourier transform

fˆ: G
b → C by
Z
ˆ
f (e) = f (x)e(x) dx
G

(e ∈ G), b → C, we define its inverse Fourier

b and for a ‘nice’ function φ : G
transform φ̌ : G → C by Z
φ̌(x) = φ(e)e(x) de
G
b

(x ∈ G). The terminology suggests that the two processes should be mutually
inverse, which they are if the functions are ‘nice’ enough.
When G = Rn (and so G b = Rn too), fˆ is the usual Fourier transform
n
of a function f : R → C (see Example D1.3(i)), and φ̌ is the inverse Fourier
transform of a function φ : Rn → C.
When G = T (and so G b = Z), what is here called fˆ(ek ) is what we usually
call fˆ(k), the kth Fourier Pcoefficient of f . Given a double sequence φ : Z → C,
what is here called φ̌(x) is k∈Z φ(k)ek (x). So, this is nothing but Fantasy A8.7.
Both Fourier transforms and Fourier series therefore arise as special cases
of the general theory of Fourier transforms on locally compact abelian groups.
Developing this theory is a major undertaking, requiring some of the theory of
topological groups and also some measure theory. But there is a special case in
which all the complications disappear, and that is what we will study for the
rest of the course. It is the theory of Fourier transforms on finite abelian groups.

69
D2 The dual of a finite abelian group
For the lecture of Monday 24 March. (Lecture of Thu 20 March is cancelled)

Maybe you’re fed up with fussy analytic conditions: this function is continu-
ously differentiable, that series is absolutely summable, and so on. If so, the last
few lectures are for you. They will show you a world in which Fourier analysis
works beautifully without any analytic conditions at all.
Everything will be developed from the ground up, assuming only some ba-
sic group theory. I will not assume any definitions, notation or results from
Section D1.
Definition D2.1 We denote by S the multiplicative group {z ∈ C : |z| = 1}.

Note that z = z −1 for z ∈ S.

When we do Fourier analysis on finite abelian groups, we will always use
multiplicative notation for our groups (even though the locally compact abelian
groups in Section D1 were written additively). It’s only notation.
Lemma D2.2 Let G be a group. Then the set
b = {homomorphisms G → S}
G

is a group under the following operations:

• the product of e1 , e2 ∈ G
b is the product function e1 · e2 (defined as in
Notation A3.3);

• the inverse of e ∈ G
b is e (again, defined as in Notation A3.3);

• the identity is the constant function 1.

Proof The set of all functions G → S is certainly a group under these opera-
tions. We show that G
b is a subgroup of this group, that is:

• if e1 , e2 : G → S are homomorphisms then so is e1 · e2 ;

• if e : G → S is a homomorphism then so is e;
• the constant function 1 : G → S is a homomorphism.
For the first one, if e1 , e2 ∈ G
b then for all x, y ∈ G,

(e1 · e2 )(xy) = e1 (xy)e2 (xy) (by definition of e1 · e2 )

= e1 (x)e1 (y)e2 (x)e2 (y) (since e1 and e2 are homomorphisms)
= e1 (x)e2 (x)e1 (y)e2 (y) (since S is abelian)
= (e1 · e2 )(x) · (e1 · e2 )(y) (by definition of e1 · e2 ),

so e1 · e2 is a homomorphism. The arguments for e and 1 are similar (but

simpler). 

Definition D2.3 Let G be a finite abelian group. A character of G is a

homomorphism G → S. The character group or Pontryagin dual of G is
the group G.
b

70
 ω2
ω = exp(2πi/7)
ω3
1

ω4
S
ω6
ω5

Figure D.2: The complex nth roots of unity are exactly the powers of
exp(2πi/n). Shown: n = 7.

Let us pause to compare this with the definitions in Section D1 (although

this is not logically necessary for anything that follows). In Definition D1.2, a
character of a topological group G was defined to be a continuous homomor-
phism G → S. In Definition D2.3, continuity isn’t mentioned. What’s going
on?
Implicitly, we are putting the discrete topology on our finite abelian groups
(as in Example D1.1(iv)). The discrete topology has the special property that
all maps out of a discrete space are automatically continuous. Thus, when G
is given the discrete topology, a continuous homomorphism G → S is the same
thing as an ordinary homomorphism G → S. One of the nice things about
Fourier theory for finite abelian groups is that the topological and analytic
conditions vanish entirely.
Coming back to the main story, let Cn denote the cyclic group of order n.
cn ∼
Lemma D2.4 C = Cn for all n ≥ 1.
Proof Let n ≥ 1. We prove that C cn is a cyclic group of order n.
Choose a generator x of Cn . Let ω = exp(2πi/n), and note that the complex
numbers z satisfying z n = 1 are exactly the integer powers of ω (Fig. D.2).
Define d : Cn → S by d(xr ) = ω r (r ∈ Z). This is well-defined, as if xr1 = xr2
then r1 ≡ r2 (mod n), so ω r1 = ω r2 (using the fact that ω n = 1). It is also a
homomorphism, since for r, s ∈ Z,
d(xr · xs ) = d(xr+s ) = ω r+s = ω r · ω s = d(xr ) · d(xs ).
So d is a character of Cn .
Now let e be any character of Cn . We have e(x)n = e(xn ) = e(1) = 1, so
e(x) = ω r for some r ∈ Z. I claim that e = dr . Indeed, given y ∈ Cn , we have
y = xs for some s ∈ Z, and then
e(y) = e(xs ) = e(x)s = ω rs = d(x)rs = d(xs )r = d(y)r = dr (y),

proving the claim. So Ccn is cyclic, generated by d.

It remains to find the order of d. Certainly dn = 1, since every element of
Cn is of the form xr for some r ∈ Z, and dn (xr ) = d(x)nr = ω nr = 1. On the
other hand, if 1 ≤ m < n then dm (x) = ω m 6= 1, so dm 6= 1. So d has order n.
We have shown that C cn is a cyclic group generated by an element of order
c ∼
n. Thus, Cn = Cn . 

71
 Next recall that any two groups G1 and G2 have a product (also called a
‘direct product’) G1 × G2 .
∼c c
1 × G2 = G1 × G2 for all finite abelian groups G1 and G2 .
Lemma D2.5 G\

Proof Sheet 5. 

Fact D2.6 Every finite abelian group is isomorphic to Cn1 × Cn2 × · · · × Cnk
for some k, n1 , n2 , . . . , nk ≥ 1.

This is part of the classification theorem for finite abelian groups, which you
may have met in other courses.
Putting together the last three results gives:
b∼
Proposition D2.7 G = G for every finite abelian group G.

Proof This follows from Lemma D2.4, Lemma D2.5 and Fact D2.6. 

This proposition is the reason why the finite abelian groups were drawn on
the central dotted line (the ‘mirror’) of Figure D.1.
Remark D2.8 For a given G, there is usually no canonical (i.e. God-given)
isomorphism G → G. b In order to construct an isomorphism, you have to make
some arbitrary choice, of the same kind you make when tossing a coin or choosing
a basis for a vector space.

72
D3 Fourier transforms on a finite abelian group
For the lecture of Thursday 27 March

First we’ll make some definitions analogous to the definitions for Fourier series.
Then we’ll prove some results analogous to results on Fourier series. We’ll
discover that life is much easier on a finite abelian group than on the circle.
For the rest of this section, let G be a finite abelian group.

Definition D3.1 i. Given a set X, write Fn(X) = {functions X → C}.

1
R P
ii. For f ∈ Fn(G), write G f (x) dx = #G x∈G f (x).
R
iii. For f, g ∈ Fn(G), write hf, gi = G
f (x)g(x) dx.

In (i), Fn(X) is a vector space over C, via the usual addition and scalar
multiplication of functions (as in Notation
R A3.3).
1
R
In (ii), the factor #G ensures that G 1 dx = 1, just as T 1 dx = 1. So the
integral of a function on G can be thought of as its mean value, just as for
functions on T.
We now establish some elementary properties of integration, directly analo-
gous to those for ordinary integration stated in Lemma A3.4.
Lemma D3.2 i. For any functions f, g : G → C,
Z Z Z
(f + g)(x) dx = f (x) dx + g(x) dx.
G G G

ii. For any function f : G → C and c ∈ C,

Z Z
(cf )(x) dx = c f (x) dx.
G G

iii. For any function f : G → C,

Z Z
¯
f (x) dx = f (x) dx.
G G

Proof For (i), let f, g ∈ Fn(G). Then

Z Z
(f + g)(x) dx = (f (x) + g(x)) dx (by definition of f + g)
G G
Z
1 X
= (f (x) + g(x)) (by definition of )
#G G
x∈G
1 X 1 X
= f (x) + g(x)
#G #G
x∈G x∈G
Z Z Z
= f (x) dx + g(x) dx (by definition of ).
G G G

Similar arguments prove parts (ii) and (iii). 

73
Lemma D3.3 h·, ·i is an inner product on Fn(G).

Proof This is straightforward, using Lemma D3.2. (Compare your solution to

Sheet 1, q.6(i).) 

Note that h·, ·i is a genuine inner product; that is, if hf, f i = 0 then f = 0.
(Contrast the comments after Lemma A6.2.) This is because for f ∈ Fn(G),
1 X
hf, f i = |f (x)|2 ,
#G
x∈G

and if this is 0 then f (x) = 0 for all x ∈ G.

In the proof of Lemma A7.4 (which stated R that the characters of T are
orthonormal), it was shown in passing that T ek (x) dx is 1 if k = 0 and 0
otherwise. Something very similar holds for finite abelian groups.
Lemma D3.4 Let e ∈ G.
b Then
Z (
1 if e = 1
e(x) dx =
G 0 otherwise.

Remember that the elements of G b are homomorphisms, so the ‘1’ in ‘e = 1’

means the constant function 1 from G to C.
R
Proof If e = 1 then G e(x) dx = 1 (as noted after Definition D3.1).
If e 6= 1, choose y ∈ G such that e(y) 6= 1. Then
Z
1 X 1 X
e(x) dx = e(x) = e(yz)
G #G #G
x∈G z∈G

since as x runs over G, z = y −1 x runs over G too. (Morally, this is integration

by substitution; we’re substituting x = yz.) So
Z Z
1 X
e(x) dx = e(y)e(z) = e(y) e(x) dx.
G #G G
z∈G
R
But e(y) 6= 1, so G
e(x) dx = 0. 

Proposition D3.5 The characters of G are orthonormal: for e1 , e2 ∈ G,

b
(
1 if e1 = e2
he1 , e2 i =
0 otherwise.

(Compare Lemma A7.4, the analogous result on T.)

Proof Let e1 , e2 ∈ G. b we have e1 e2 ∈ G.

b Since e2 is the inverse of e2 in G, b By
Lemma D3.4,
Z ( (
1 if e1 e2 = 1 1 if e1 = e2
he1 , e2 i = (e1 e2 )(x) dx = =
G 0 otherwise 0 otherwise. 

74
Definition D3.6 Let f ∈ Fn(G). The Fourier transform of f is the function
fˆ ∈ Fn(G)
b defined by fˆ(e) = hf, ei (e ∈ G).
b

(Compare: for Fourier series, fˆ(k) = hf, ek i.)

Note that we’ve defined the Fourier transform of any function on G! In our
simple world of finite groups, there’s no integrability to worry about.
Proposition D3.7 For all φ ∈ Fn(G),
b
X ∧
φ= φ(e) · e
e∈G
b

(The right-hand side means fˆ where f = e∈Gb φ(e) · e. This alternative

P
notation is simply for typesetting convenience, to avoid very wide hats.)
The sum on the right-hand side is a finite sum, since G b∼ = G (by Proposi-
tion D2.7) and G is finite. Unlike the circle, G has only finitely many characters.

Proof Both sides of this equation are functions on G,b so to prove equality, we
need to take an arbitrary element e0 of G
b and show that evaluating each side at
e0 gives the same result.
Let e0 ∈ G.
b Then
X ∧ DX E
φ(e) · e (e0 ) = φ(e) · e, e0
e∈G
b e
X
= φ(e)he, e0 i (by Lemma D3.3)
e
= φ(e0 ) (by orthonormality). 

This result is similar to the ‘Moreover’

Ppart of Proposition

∧ C2.1, which stated
that for ‘nice’ c ∈ Fn(Z), we have c = c e
k∈Z k k .
Recall that Fn(G) is a vector space over C.

Lemma D3.8 dim(Fn(G)) = #G.

Proof For each x ∈ G, define δx ∈ Fn(G) by

(
1 if y = x
δx (y) =
0 otherwise

P ∈ G). The family (δx )x∈G spans Fn(G), since if f ∈ Fn(G) then f =
(y
x∈G f (x)δx (check!). It is also linearly
P independent, since if (cx )x∈G is a
family of complex numbers such that x∈G cx δx = 0, then for each y ∈ G we
have X  X
0= cx δx (y) = cx δx (y) = cy .
x∈G x∈G

So it is a basis. 

Lemma D3.8 doesn’t really have an analogue in the world of Fourier series
(that is, Fourier theory on the circle group T).

75
Theorem D3.9 The characters of G form an orthonormal basis of Fn(G).

Proof By Proposition D3.5, the characters are orthonormal. In particular, they

are linearly independent, so the subspace of Fn(G) that they span has dimension
#G.
b But #G b = #G = dim(Fn(G)) by Proposition D2.7 and Lemma D3.8, so
their span is Fn(G). 

Hence every function on G is a ‘trigonometric polynomial’ (a finite linear

combination of characters). This makes Fourier analysis on finite groups con-
siderably simpler than Fourier analysis on the circle.
P For ˆa function f : G → C, the analogue of the Fourier series of f is the sum
e∈Gb f (e)e. Since this sum is finite, there is no question of convergence; either
it is equal to f , or it is not. The next result says that it is.

Corollary D3.10 For all f ∈ Fn(G), f = e∈Gb fˆ(e)e.

P

but fˆ(e) = hf, ei by definition. 

P
Proof By Theorem D3.9, f = b hf, eie;
e∈G

So we can reconstruct a function from its Fourier transform. For Fourier

series, we had to work hard to prove that!
Since we can reconstruct a function from its Fourier transform, if two func-
tions have the same Fourier transform then they must in fact be the same:

Corollary D3.11 Let f, g ∈ Fn(G). If fˆ = ĝ then f = g.

Proof Follows from Corollary D3.10. 

This is the (simpler) analogue of Corollary B8.5; see also Remark C4.3.

76
D4 Fourier inversion on a finite abelian group
For the lecture of Monday 31 March

Throughout this section, let G be a finite abelian group.

b define φ̌ ∈ Fn(G) by φ̌ = P b φ(e)e.
Definition D4.1 For φ ∈ Fn(G), e∈G

The upside-down hat is meant to suggest the opposite of the Fourier trans-
form. It is pronounced ‘check’ (as in ‘phi-check’).
Remark D4.2 Suppose we were dealing with the circle T rather than a finite
abelian group G. We have T b = Z (as in Section D1), so φ would be an element
of Fn(Z), that is, aPdouble sequence c = (ck )k∈Z . Then č is notation for the
familiar expression k∈Z ck ek .

Fantasy A8.7 comes true in the setting of finite abelian groups. (No fussy
analytical details!) This is the content of the following theorem.
Theorem D4.3 The maps

(ˆ)
Fn(G) o /
Fn(G)
b
(ˇ)

are linear and mutually inverse. In particular, the vector spaces Fn(G) and
Fn(G)
b are isomorphic.

ˇ
Proof Corollary D3.10 states that fˆ = f for all f ∈ Fn(G), and Proposi-
tion D3.7 states that φ̌ˆ = φ for all φ ∈ Fn(G). b So (ˆ) and (ˇ) are mutually
ˆ
inverse. The map ( ) is linear since h·, ·i is linear in the first argument; so its
inverse (ˇ) is linear too. Hence (ˆ) and (ˇ) define an isomorphism of vector
spaces. 
p
Definition D4.4 For f ∈ Fn(G), define kf k2 = hf, f i.

This defines a norm k · k2 on Fn(G), since h·, ·i is an inner product.

Theorem D4.5 (Parseval for finite abelian groups) For all f ∈ Fn(G),
sX
kf k2 = |fˆ(e)|2 .
e∈G
b

Proof By Corollary D3.10, f = e∈Gb fˆ(e)e. So by orthogonality of the char-

P
acters, X
kf k22 = hf, f i = |fˆ(e)|2 .
e∈G
b 

We know from Theorem D4.3 that the function (ˆ) : Fn(G) → Fn(G) b is an
isomorphism of vector spaces. We’d like to say that it’s also an isomorphism of
metric spaces, that is, distance-preserving: kf k2 = kfˆk2 for all f ∈ Fn(G). But
the right-hand side is not yet defined. So we make some definitions:

77
 R
Definition D4.6 b define b φ(e) de = P b φ(e) ∈ C.
i. For φ ∈ Fn(G), G e∈G
R
ii. For φ, ψ ∈ Fn(G),
b define hφ, ψi = b φ(e)ψ(e) de ∈ C.
G
p
iii. For φ ∈ Fn(G),
b define kφk2 = hφ, φi.

Compare and contrast (i) with Definition D3.1(ii): here, there is no factor
1
of #G . This is to make the following true:

Corollary D4.7 (Parseval, restated) For all f ∈ Fn(G),

kf k2 = kfˆk2 .

qP is a restatement of Theorem D4.5, since for φ ∈ Fn(G) we have

Proof This b
kφk2 = 2
b |φ(e)| .
e∈G 

The characters of G are functions G → C with certain properties. Are there

enough such functions to tell different points of G apart? That is, if x and y
are two different points in G, can we find a character e such that e(x) and e(y)
are different?
The answer is yes, and it’s now easy for us to prove. But it’s not so easy to
prove from scratch: try it!
Proposition D4.8 (Characters separate points) Let x, y ∈ G with x 6= y.
Then there is some e ∈ G
b such that e(x) 6= e(y).

Proof Suppose that e(x) = e(y) for all e ∈ G. b Since x 6= y, there is some
f ∈ Fn(G) such that f (x) 6= f (y). (E.g. define f by f (x) = 1 and f (z) = 0 for
all z 6= x.) Then by Corollary D3.10,
X X
f (x) = fˆ(e)e(x) = fˆ(e)e(y) = f (y),
e∈G
b e∈G
b

a contradiction. 

Remark D4.9 We know from Proposition D2.7 that G ∼ =G b∼ b so G ∼

= G, = G.
b b
b
But for all we know so far, there is no canonical choice of isomorphism G → G.
b
b
(Recall Remark D2.8.) However, the next result tells us that although there
is no canonical isomorphism between a group and its single dual, there is a
canonical isomorphism between a group and its double dual.

Theorem D4.10 For x ∈ G, define evx ∈ Fn(G)

b by

evx (e) = e(x)

(e ∈ G).
b Then evx ∈ G
b for all x ∈ G, and the map
b

G −→ G
b
b
x 7−→ evx

is an isomorphism of groups.

78
Proof For each x ∈ G, the function evx is a homomorphism G
b → S: indeed,
for all e1 , e2 ∈ G,
b we have

evx (e1 · e2 ) = (e1 · e2 )(x) (by definition of evx )

= e1 (x)e2 (x) (by definition of e1 · e2 )
= evx (e1 ) evx (e2 ) (by definition of evx ).

This means we can define a function ev : G → G

b by ev(x) = evx (x ∈ G).
b
This function ev is a homomorphism. Indeed,

ev is a homomorphism
⇐⇒ ∀x, y ∈ G, ev(xy) = ev(x) · ev(y)
⇐⇒ ∀x, y ∈ G, evxy = evx · evy
⇐⇒ ∀x, y ∈ G, ∀e ∈ G,
b evxy (e) = (evx · evy )(e)
⇐⇒ ∀x, y ∈ G, ∀e ∈ G,
b evxy (e) = evx (e) · evy (e)
⇐⇒ ∀x, y ∈ G, ∀e ∈ G,
b e(xy) = e(x) · e(y),

which is true. So reading the implications from bottom to top, ev is a homo-

morphism.
Next, ev : G → G b is injective. For let x, y ∈ G and suppose that ev(x) =
b
ev(y). Then evx = evy , that is, evx (e) = evy (e) for all e ∈ G.
b But characters
separate points (Proposition D4.8), so x = y.
Finally, ev : G → Gb is surjective, since ev is injective, #G = #G b = #G b by
b b
Proposition D2.7, and an injection from one finite set to another set of the same
cardinality is always surjective, by the pigeonhole principle. 

This is Pontryagin’s duality theorem (Theorem D1.4) in the special case of

finite abelian groups.
We have now proved analogues of all the main results on Fourier series in
the setting of finite abelian groups. We have also proved the central theorem
of abstract Fourier analysis (Pontryagin duality) in this simplified setting. And
we have done all this in just a few pages, calling on nothing more than basic
group theory.

∗ ∗ ∗

79