TOPICS IN FOURIER ANALYSIS-II

M.T. NAIR

Contents

1. Trigonometric series and Fourier series 2

2. Riemann Lebesgue Lemma 4
3. Dirichlet kernel 6
4. Dirichlet-Dini criterion for convergence 8
5. Ces̀aro summablity of Fourier series 12
6. Divergence of Fourier series 18
7. Uniqueness 20
8. Convolution 23
9. L2 -Theory 26
References 28

NOTES PREPARED FOR PART OF AN ELECTIVE COURSE FOURIER ANALYSIS

FOR M.SC. STUDENTS OF IIT MADRAS, JULY-NOVEMBER, 2014.
1
2 M.T. NAIR

1. Trigonometric series and Fourier series

Definition 1.1. A series of the form

∞
X
(1.1) c0 + (an cos nx + bn sin nx)
n=1

is called a trigonometric series, where c0 , an , bn are real numbers.

• If (1.1) converges on [−π, π] to a an integrable function f and if it can be

integrated term by term, then
f (−π) = f (π),
and
Z π Z π Z π
1 1 1
c0 = f (x) cos nxdx, an = f (x) cos nxdx, bn = f (x) sin nxdx.
2π −π π −π π −π

• If the (1.1) converges (pointwise) on [−π, π] to a function f , then f can be

extended as a 2π-periodic function by defining
f (x + 2nπ) = f (x), n ∈ Z.
∞
X
• If the series (|an | + |bn |) converges, then (1.1) converges uniformly on [−π, π]
n=1
and it can be integrated term by term. We know that if f ∈ L1 [−π, π], then
the function f˜ : [−π, π] → C defined by

˜ f (x), x ∈ [−π, π),
f (x) =
f (−π), x = π
satisfies
f˜(−π) = f˜(π) and f˜ = f a.e.
• The series (1.1) can be written as
∞
X
cn einx .
n=−∞

Definition 1.2. Let f ∈ L1 [−π, π]. The Fourier series of f is the series
∞
a0 X
(1.2) + (an cos nx + bn sin nx),
2 n=1

where
Z π Z π
1 1
(1.3) an = f (x) cos nxdx, bn = f (x) sin nxdx.
π −π π −π
 TOPICS IN FOURIER ANALYSIS 3

The series
∞ Z π
X 1
(1.4) cn e inx
with cn = f (x)e−inx dx
n=−∞
2π −π

is also called the Fourier series of f . The coefficients cn are called the Fourier
coefficient and are usually denoted by fˆ(n), i.e.,
Z π
ˆ 1
f (n) = f (x)e−inx dx, n ∈ Z.
2π −π
The sum
N
X
SN (f, x) := fˆ(n)einx
n=−N

is called the N -th partial sum of the Fourier series (1.4)).

Notation: In the above and in the following, the integral are w.r.t. the Lebesgue
measure.

The fact that (1.2) is the Fourier series of f is usually written as

∞
a0 X
f (x) ≈ + (an cos nx + bn sin nx).
2 n=1

Equivalently,
∞
X
f (x) ≈ fˆ(n)einx .
n=−∞
inx
Since cos nx, sin x, e are 2π-periodic functions, we can talk about Fourier series of
2π-periodic functions. If (1.2) (resp. (1.4)) converges at a point x ∈ [−π, π], then it
converges at x + 2kπ for every k ∈ Z.

• The Fourier series (1.4)) converges at x ∈ [−π, π] if and only if SN (f, x) → f (x)
as N → ∞.
• If f ∈ L1 [−π, π], then fˆ(n) → 0 as |n| → ∞.
X∞ X∞
• If ˆ
|f (n)| converges, then fˆ(n)einx converges uniformly.
n=−∞ n=−∞

Suppose Fourier series of f ∈ L1 [−π, π] converges uniformly, say to g. Then g is

continuous and
Z π ∞ Z π
1 X
ˆ 1
g(x)e−imx
= f (n) ei(n−m)x dx = fˆ(m),
2π −π n=−∞
2π −π
4 M.T. NAIR

i.e., ĝ(m) = fˆ(m) for all m ∈ Z. A natural question would be whether f = g a.e. We
shall answer this affirmatively.
We know that if the Fourier series of f ∈ L1 [−π, π] converges, then
fˆ(n) → 0 as n → ∞.
Can we assert this for every f ∈ L1 [−π, π]? The answer is in the affirmative as proved
in the next section.

2. Riemann Lebesgue Lemma

Theorem 2.1. (Riemann Lebesgue lemma) Let f ∈ L1 [a, b]. Then

Z b Z b
f (t) cos(λt)dt → 0 and f (t) sin(λt)dt → 0 as λ → ∞.
a a

Corollary 2.2. (Riemann Lebesgue lemma) Let f ∈ L1 [a, b]. Then

Z b Z b
f (t) cos(nt)dt → 0 and f (t) sin(nt)dt → 0 as n → ∞.
a a

For the proof of the Theorem 2.1, we shall make use of

LEMMA 2.3. The span of all step functions1 on [a, b] is dense in L1 [a, b].

Proof of Theorem 2.1. First we observe that if for every ε > 0, there exists a func-
tion g ∈ L1 [a, b] such that kf − gk1 < ε and the the result is true for g, then the result
is true for f also.

Indeed,
Z b Z b Z b
f (t) cos(λt)dt ≤ [f (t) − g(t)] cos(λt)dt + g(t) cos(λt)dt
a a a
Z b
≤ ε+ g(t) cos(λt)dt .
a
Rb
Let λ0 > 0 be such that a g(t) cos(λt)dt < ε for all λ ≥ λ0 . Then we
have Z b
f (t) cos(λt)dt < 2ε ∀ λ ≥ λ0
a

1Step
functions are finite linear combinations of characteristic functions. Also, recall that L1 [a, b]
is the vector space of all Lebesgue measurable complex valued functions f such that kf k1 :=
Z b
|f (x)|dx < ∞. Here, dx stands for the Lebesgue measure.
a
 TOPICS IN FOURIER ANALYSIS 5
Rb Rb
so that a f (t) cos(λt)dt → 0 as λ → ∞. Similarly, a f (t) sin(λt)dt → 0
as λ → ∞.

Hence, it is enough to prove the result for step functions. Since every step function is a
finite linear combination of characteristic functions on intervals, it is enough to prove
for f of the form f = χ[c,d] , [c, d] ⊆ [a, b]. Note that
Z b Z d
χ[c,d] cos(λt)dt = cos(λt)dt
a c
sin(λd) − sin(λc)
=
λ
2
≤ → 0 as λ → ∞.
|λ|
Z b
Similarly, χ[c,d] sin(λt)dt → 0 as λ → ∞. 
a

Remark 2.4. If f is Riemann integrable on [a, b], then there exists a sequence of (fn )
of step functions such that kf − fn k1 → 0. Thus, conclusion in Theorem 2.1 holds if f
is Riemann integrable.

Proof of Lemma 2.3. If f ∈ L1 [a, b] with f ≥ 0, then there exists an increasing sequence
of non-negative simple measurable functions ϕn , n ∈ N such that ϕn → f pointwise.
Rb
Hence, by DCT, a |f − ϕn | → 0. From this, for any complex valued f ∈ L1 [a, b], there
exists a sequence (ϕn ) of simple complex measurable functions
Z b
|f − ϕn | → 0.
a

We observe (see [3]):

(1) Every simple real valued measurable function is a finite linear combination of
characteristic function of measurable sets.
(2) For every measurable set E ⊆ (a, b) and ε > 0, there exists an open set G ⊇ E
such that m(G \ E) < ε. Hence,
Z b Z b
|χG − χE | = |χ(G\E) | ≤ m(G \ E) < ε.
a a
S∞
(3) If G ⊆ (a, b) is an open set, then G = k=1 In , where {In } is a countable
disjoint family of open intervals in (a, b);
n
X
χG = lim ψn , ψn = χIk ,
n→∞
k=1
6 M.T. NAIR

Since 0 ≤ ψn ≤ χG , by DCT,
Z
|χG − ψn | → 0.

(4) By (1)-(3), if ϕ is a simple measurable function and ε > 0, there exists a step
function ψ such that
Z b
|ϕ − ψ| < ε.
a

Thus, the lemma is proved. 

3. Dirichlet kernel

Note that
N
1 X inx π
Z Z π
−int 1
SN (f, x) := e f (t)e dt = f (t)DN (x − t)dt,
2π n=−N −π 2π −π
where
N
X
DN (t) := eint .
n=−N
Redefining f at the end-points if necessary, and extending it as a 2π-periodic function,
we can also write (verify!),
Z π
1
SN (f, x) = f (x − t)DN (t)dt.
2π −π

Notation: We denote by T the unit circle T := {eit : −π ≤ t < π}. Note that if
f : T → C and if we define f˜ : R → C by f˜(t) = f (eit ), then
f˜(−π) = f˜(π) and f˜(t + 2nπ) = f (t) for all n ∈ Z.
That is, f˜ is a 2π-periodic function. In the due course, we shall identify 2π-periodic
functions with functions on T . We shall denote L1 (T ) for the space of all 2π-periodic
(complex valued) functions on R (with equality replaced equal a.e.) which are inte-
grable on [−π, π] with norm
Z π
1
f 7→ kf k1 := |f (x)|dx.
2π −π
Analogously, for 1 ≤ p < ∞, Lp (T ) denotes the space of all 2π-periodic (complex
valued) functions f on R such that |f |p is integrable on [−π, π] with norm
 Z π 1/p
1 p
f 7→ kf kp := |f (x)| dx
2π −π
 TOPICS IN FOURIER ANALYSIS 7

The space L2 (T ) is also a Hilbert space with inner product

Z π
1
(f, g) 7→ hf, gi := f (x)g(x) dx.
2π −π

Definition 3.1. The function DN (·) is called the Dirichlet kernel.

We observe that,

• D
Z Nπ (−t) = DN (t) for all t ∈ [−π, π] and
• DN (t)dt = 1.
−π
N
X N
X N
X
int int −int
• DN (t) = e =1+ [e + e ]=1+2 cos nt.
n=−N n=1 n=1
Z π
Remark 3.2. We shall see that |DN (t)|dt → ∞ as N → ∞.
−π

Theorem 3.3.



 2N + 1, t = 0,

DN (t) =
sin(N + 21 )t
, t 6= 0.


sin( 2t )



Proof. Clearly, DN (0) = 2N + 1. So, let t 6= 0. Note that

N
X
it
(e − 1)DN (t) = [ei(n+1)t − eint ] = ei(N +1)t − e−iN t .
n=−N

But,
(eit − 1)DN (t) = eit/2 (eit/2 − e−it/2 )DN (t) = 2ieit/2 sin(t/2)DN (t).

Thus,

2i sin(t/2)DN (t) = e−it/2 [ei(N +1/2)t − e−i(N +1/2)t ] = 2i sin(N + 1/2)t.

i.e.,
sin(N + 21 )t
DN (t) = , t 6= 2kπ.
sin( 2t )

8 M.T. NAIR

4. Dirichlet-Dini criterion for convergence

We investigate the convergence:

SN (f, x) → f (x).
Z π Rπ
1
Since DN (t)dt = 1 and SN (f, x) = 2π −π
f (x − t)DN (t)dt, we have
−π
Z π
1
f (x) − SN (f, x) = [f (x) − f (x − t)]DN (t)dt.
2π −π

Theorem 4.1. (Dirichlet-Dini criterion) Let f ∈ L1 (T ). If f satisfies

Z π
f (x) − f (x − t)
dt < ∞ (∗)
−π t
at a point x ∈ [−π, π], then
SN (f, x) → f (x).
If (∗) hods uniformly for x ∈ [−π, π], then the convergence {SN (f, x)} to f (x) is
uniform.
f (x) − f (x − t)
Remark 4.2. In the above theorem, by , we mean the function
t
(
f (x)−f (x−t)
t
, t 6= 0
ϕ(t) =
0, t = 0.

Proof of Theorem 4.1. We observe that

Z π
1
f (x) − SN (f, x) = [f (x) − f (x − t)]DN (t)dt
2π −π
sin(N + 21 )t
Z π
1
= [f (x) − f (x − t)] dt
2π −π sin( 2t )
1 π f (x) − f (x − t)
Z   
t/2 1
= t sin(N + )tdt
π −π t sin( 2 ) 2

Since (t/2)/[sin(t/2)] is bounded, in view of Riemann Lebesgue lemma, we have the

following. 

The following corollaries are immediate from Theorem 4.1.

Corollary 4.3. Suppose f is Lipschitz at a point2 x ∈ [−π, π]. Then
SN (f, x) → f (x) as N → ∞.
2Afunction ϕ : I → C is said to be Lipschitz at a point x0 ∈ I if there exists K0 > 0 such that
|ϕ(x) − ϕ(x0 )| ≤ K0 |x − x0 | for all x ∈ I.
 TOPICS IN FOURIER ANALYSIS 9

Corollary 4.4. Suppose f is Lipschitz3 on [−π, π]. Then

SN (f, x) → f (x) as N →∞
uniformly on [−π, π].

Notation: We denote by C(T ) the space of all 2π-periodic continuous functions on R,

and by C k (T ) for k ∈ N ∪ {0}, the space of all 2π-periodic functions on R which are
k-times continuously differentiable on R.

Corollary 4.5. If f ∈ C 1 (T ), then

SN (f, x) → f (x) as N →∞
uniformly on R.

Now obtain a more general result.

Theorem 4.6. Suppose f is a 2π-periodic function such that the following limits exist
at a point x ∈ R:
f (x+) := lim f (x + t), f (x−) := lim f (x − t),
t→0+ t→0+

f (x + t) − f (x+) f (x−) − f (x − t)
f 0 (x+) := lim , f 0 (x−) := lim .
t→0+ t t→0+ t
Then
f (x+) + f (x−)
SN (f, x) → as N → ∞.
2

Proof. Since DN (t) = DN (−t), we have

Z π
1
SN (f, x) = f (x − t)DN (t)dt
2π −π
Z 0 Z π
1 1
= f (x − t)DN (t)dt + f (x − t)DN (t)dt
2π −π 2π 0
Z π Z π
1 1
= f (x + t)DN (t)dt + f (x − t)DN (t)dt
2π 0 2π 0
Z π
1
= [f (x + t) + f (x − t)]DN (t)dt
2π 0
and Z π Z π
1 2
DN (t)dt = DN (t)dt.
2π −π 2π 0

3A function ϕ : I → C is said to be Lipschitz on I if there exists K > 0 such that |ϕ(x) − ϕ(x0 )| ≤
K0 |x − x0 | for all x ∈ I.
10 M.T. NAIR

Hence, for any β ∈ R,

Z π
1
SN (f, x) − β = [f (x + t) + f (x − t) − 2β]DN (t)dt.
2π 0
f (x+)+f (x−)
Taking β = 2
, we have
f (x + t) + f (x − t) − 2β = [f (x + t) − f (x+)] − [f (x−) − f (x − t)].
Thus,
SN (f, x) − β = AN + BN ,
where
Z π Z π
1 1
AN = [f (x + t) − f (x+)]DN (t)dt, BN = [f (x−) − f (x − t)]DN (t)dt.
2π 0 2π 0

Note that
Z π
1
A = [f (x + t) − f (x+)]DN (t)dt
2π 0
sin(N + 21 )t
Z π
1
= [f (x + t) − f (x+)] dt
2π 0 sin( 2t )
1 π f (x + t) − f (x+)
Z   
t/2 1
= t sin(N + )tdt
π 0 t sin( 2 ) 2

f (x+t)−f (x+)
Since t
→ f 0 (x+) as t → 0+, there exists δ > 0 such that
f (x + t) − f (x+)
0<t<δ =⇒ − f 0 (x+) ≤ 1
t
f (x + t) − f (x+)
=⇒ ≤ 1 + |f 0 (x+)|.
t
Hence, the function
  
f (x + t) − f (x+) t/2
t 7→ , t 6= 0,
t sin( 2t )
is bounded on (0, δ), and hence, belongs to L1 (T ). Therefore,by Riemann Lebesgue
lemma, AN → 0 as N → ∞. Similarly, we see that, BN → 0 as N → ∞. 

An immediate corollary:

Corollary 4.7. If f ∈ C(T ) and has left and right derivative at a point x, then
SN (f, x) → f (x) as N → ∞.

The following result is known as localization lemma.

 TOPICS IN FOURIER ANALYSIS 11

LEMMA 4.8. For 0 < r < π and x ∈ [−π, π],

Z
f (x − t)DN (t)dt → 0 as N → ∞.
r≤|t|≤π

Proof. Observe that

Z Z
f (x − t)DN (t)dt = g(x, t) sin(N + 1/2)tdt,
r≤|t|≤π r≤|t|≤π

where 
f (x − t)/ sin(t/2), r ≤ |t| ≤ π,
g(x, t) =
0, |t| ≤ r.
Since g(x, ·) is integrable, by Riemann Lebesgue lemma,
Z
g(x, t) sin(N + 1/2)tdt → 0 as N → ∞.
r≤|t|≤π

Proof of Corollary 4.4 using localization lemma. Suppose f is Lipschitz at a point

x ∈ [−π, π] with Lipschitz constant Kx , i.e., there exists δ > 0 such that
|f (x) − f (x − t)| ≤ Kx |t| whenever |t| < δ.
Now,
Z π
1
f (x) − SN (f, x) = [f (x) − f (x − t)]DN (t)dt
2π −π
Z
1
= [f (x) − f (x − t)]DN (t)dt
2π 0≤|t|<δ
Z
1
+ [f (x) − f (x − t)]DN (t)dt
2π δ≤|t|≤π

By Lemma 4.8,
Z
1
[f (x) − f (x − t)]DN (t)dt → 0 as N → ∞.
2π δ≤|t|≤π

Hence, for a given ε > 0, there exists N0 ∈ N such that for all N ≥ N0 ,
Z
1
[f (x) − f (x − t)]DN (t)dt < ε/2.
2π δ≤|t|≤π
Also, But,
Z Z
1 1
[f (x) − f (x − t)]DN (t)dt ≤ |f (x) − f (x − t)| |DN (t)|dt,
2π 0≤|t|<δ 2π 0≤|t|<δ
12 M.T. NAIR
Z Z
1 1
|f (x) − f (x − t)| |DN (t)|dt ≤ Kx |t| |DN (t)|dt,
2π 0≤|t|<δ 2π 0≤|t|<δ

sin(N + 12 )t t/2 1
|t| |DN (t)| = |t| t =2 t | sin(N + )t| ≤ 2M,
sin( 2 ) sin( 2 ) 2
t/2
where M is a bound for | on 0 < |t| ≤ δ. Hence,
sin( 2t )
Z
1 4M Kx δ 2M Kx δ
[f (x) − f (x − t)]DN (t)dt ≤ = .
2π 0≤|t|<δ 2π π
2M Kx δ
We may take δ such that < ε/2. Hence,
π
Z
1
|f (x) − SN (f, x)| ≤ [f (x) − f (x − t)]DN (t)dt
2π 0≤|t|<δ
Z
1
+ [f (x) − f (x − t)]DN (t)dt
2π δ≤|t|≤π
< ε for all N ≥ N0 .


Exercise 4.9. Suppose f is 2π-periodic and Hölder continuous at x, i.e., there exist
M > 0 and α > 0 such that |f (x) − f (y)| ≤ M |x − y|α for all y ∈ [−π, π]. Then show
that SN (f, x) → f (x) as N → ∞.

Exercise 4.10. Suppose f is 2π-periodic and Hölder continuous on [−π, π], i.e., there
exist M > 0 and α > 0 such that |f (x) − f (y)| ≤ M |x − y|α for all x, y ∈ [−π, π].
Then show that SN (f, x) → f (x) uniformly.

5. Ces̀aro summablity of Fourier series

Theorem 5.1. (Fejér’s theorem) If f ∈ C(T ), then the Fourier series of f is

uniformly Ces̀aro summable on [−π, π], that is,
N
1 X
σN (f, x) := Sk (f, x) → f (x) as N →∞
N + 1 k=0
uniformly on [−π, π].

Recall that
k Z π
X 1
Sk (f, x) := fˆ(n)einx = f (x − t)Dk (t)dt.
n=−k
2π −π
 TOPICS IN FOURIER ANALYSIS 13

Hence,
N
( N
)
Z π
1 X 1 X
σN (f, x) = Sk (f, x) = f (x − t) Dk (t) .
N +1 k=0 −π N +1 k=0

Thus, Z π
σN (f, x) = f (x − t)KN (t)dt,
−π
where
N
1 X
KN (t) := Dk (t).
N + 1 k=0
Definition 5.2. The function KN (t) defined above is called the Fejér kernel.

We observe that Z π
1
KN (t)dt = 1.
2π −π
Hence, Z π
1
f (x) − σN (f, x) = [f (x) − f (x − t)]KN (t)dt.
2π −π
For the proof of Theorem 5.1, we shall make use of the following lemma.
LEMMA 5.3. The following results hold.

(1) For t 6= 0,
1 1 − cos(N + 1)t 1 sin2 [(N + 1)t/2]
KN (t) = = .
N +1 1 − cos t N +1 sin2 (t/2)
(2) KN (t) is an even function and KN (t) ≥ 0 for all t ∈ [−π, π].
(3) For 0 < δ ≤ π,  
1 1
KN (t) ≤ .
N + 1 sin2 (δ/2)
In particular, KN is positive and KN (t) → 0 as N → ∞ uniformly on 0 < δ ≤ |t| ≤ π.

Proof of Theorem 5.1. Sine KN (t) is a non-negative function (see Lemma 5.3), we
have Z π
1
|f (x) − σN (f, x)| ≤ |f (x) − f (x − t)|KN (t)dt.
2π −π
Let ε > 0 be given. Since f is uniformly continuous, there exists δ ∈ (0, π] such that
|f (x) − f (y)| < ε whenever |x − y| < δ.
Hence, Z Z
1 ε
|f (x) − f (x − t)|KN (t)dt < KN (t)dt = ε.
2π |t|<δ 2π |t|<δ
14 M.T. NAIR

Also, since f is uniformly bounded there exists M > 0 such that |f (y)| ≤ M for all
x ∈ [−π, π].
Z Z
1 2M
|f (x) − f (x − t)|KN (t)dt ≤ KN (t)dt.
2π |t|≥δ 2π |t|≥δ
We have observed in Lemma 5.3 that KN (t) is an even function and KN (t) → 0 as
N → ∞ uniformly on [δ, π]. Hence, there exists N0 such that
4M π
Z Z
1
|f (x) − f (x − t)|KN (t)dt ≤ KN (t)dt < ε for all N ≥ N0 .
2π |t|≥δ 2π δ
Hence,
Z π
1
|f (x) − σN (f, x)| ≤ |f (x) − f (x − t)|KN (t)dt < 2ε
2π −π
for all N ≥ N0 . Note that N0 is independent of the point x. Thus, we have proved
that SN (f, x) → f (x) as N → ∞ uniformly for x ∈ [−π, π]. 

Remark 5.4. The proof of Theorem 5.1 reveals more:

If f is peace-wise continuous and 2π-periodic, and continuous at x, then

σN (f, x) → f (x) as N → ∞.

Notation:

• un (x) := einx , n ∈ Z.
• AC(T ) denotes the vector space of all 2π-periodic complex valued functions
defined on R which are absolutely continuous.

• span{un : n ∈ Z} is the space (over C) of all trigonometric polynomials.

Corollary 5.5. The space of all trigonometric polynomials is dense in C(T ) with
respect to the uniform norm, and hence dense in Lp (T ) w.r.t. k · kp for 1 ≤ p < ∞.

Proof. By Theorem 5.1, space of all trigonometric polynomials is dense in C(T ) with
respect to the uniform norm k · k∞ . Hence, for any f ∈ C(T ), there exists a sequence
(fn ) of trigonometric polynomials such that
Z π
p
kf − fn kp = |f (x) − fn (x)|p dx ≤ 2πkf − fn kp∞ → 0
−π

as n → ∞. 

Corollary 5.6. If f ∈ L2 (T ) for some 1 ≤ p < ∞ and fˆ(n) = 0 for all n ∈ Z, then
f = 0 a.e.
 TOPICS IN FOURIER ANALYSIS 15

Proof. Suppose f ∈ L2 (T ) for some 1 ≤ p < ∞ and fˆ(n) = 0 for all n ∈ Z, i.e.,
hf, un i = 0 for all n ∈ Z. By Corollary 5.5, it follows that kf kL2 = 0. Hence, f = 0
a.e. 

Corollary 5.7. If f ∈ C(T ) such that fˆ(n) = 0 for all n ∈ Z, then f = 0. In

particular, if f, g ∈ C(T ) such that fˆ(n) = ĝ(n) for all n ∈ Z, then f = g.

Proof. Suppose f ∈ C(T ) such that fˆ(n) = 0 for all n ∈ Z. Thus, hf, un iL2 = 0 for all
n ∈ Z. Since C(T ) ⊆ L2 [−π, π], f ∈ L2 [−π, π]. Hence by Corollary, f = 0 a.e. Since
f is continuous, f = 0. 

The above corollary shows:

The Fourier coefficients of f ∈ C(T ) determines f uniquely.

Corollary 5.8. If f ∈ C 2 (T ), then

fc00 (n) = (in)2 fˆ(n) for all n ∈ Z.

In particular, fˆ(n) = o( n12 ), and the Fourier series of f converges uniformly to f .

Proof. Let f ∈ C 2 (T ). Then, using integration by parts, we obtain,

Z π
ˆ
2π f (n) = f (x)e−inx dx
−π
Z π
h e−inx iπ h e−inx i
= f (x) − f 0 (x) dx
−in −π −π −in
1 π 0
Z
= f (x)e−inx dx
in −π
e−inx iπ 1 π 00 h e−inx i
Z
1h 0
= f (x) − f (x) dx
in −in −π in −π −in
Z π
1
= 2
f 00 (x)e−inx dx.
(in) −π

Hence, fc00 (n) = (in)2 fˆ(n) for all n ∈ Z. In particular, fˆ(n) = o(1/n2 ). Therefore,
P ˆ
n∈Z |f (n)| converges, and hence the Fourier series converges uniformly. Suppose
SN (f, x) → g(x) uniformly. Then it follows that g ∈ C(T ) and ĝ(n) = fˆ(n) for all
n ∈ Z. Therefore, by Corollary 5.7, g = f. 

Following the same arguments as in the proof of Corollary 5.8, we obtain:

16 M.T. NAIR

Corollary 5.9. If f ∈ C 1 (T ) and f 0 is absolutely continuous, then f 00 exists almost

everywhere, f 00 ∈ L1 [−π, π] and

fc00 (n) = (in)2 fˆ(n) for all n ∈ Z,

and the Fourier series of f converges uniformly to f .

More generally,

Theorem 5.10. If f ∈ C k−1 (T ) and f (k−1) is absolutely continuous for some k ∈ N,

then f (k) exists almost everywhere f (k) ∈ L1 (T ) and
(k) (n) = (in)k fˆ(n) for all
fd n ∈ Z.

Proof of Lemma 5.3. We have

N
1 X sin(k + 1/2)t
KN (t) := Dk (t) where Dk (t) =
N + 1 k=0 sin t/2

Hence,
N N
X sin(k + 1/2)t X ei(k+1/2)t − e−i(k+1/2)t
(N + 1)KN (t) = =
k=0
sin t/2 k=0
eit/2 − e−it/2
But,
ei(k+1/2)t − e−i(k+1/2)t ei(k+1)t − e−ikt
= ,
eit/2 − e−it/2 eit − 1
ei(k+1/2)t − e−i(k+1/2)t eikt − e−i(k+1)t
= ,
eit/2 − e−it/2 1 − e−it
Therefore,
N
X
[eit − 1](N + 1)KN (t) = [ei(k+1)t − e−ikt ], (1)
k=0

N
X
−it
[1 − e ](N + 1)KN (t) = [eikt − e−i(k+1)t ] (2)
k=0

Subtracting the (2) from (1),

N
X
[2 cos t − 2](N + 1)KN (t) = 2 [cos(k + 1)t − cos kt] = 2[cos(N + 1)t − 1]
k=0

Thus,
1 cos(N + 1)t − 1 1 sin2 [(N + 1)t/2]
KN (t) = = .
N +1 cos t − 1 N +1 sin2 (t/2)
 TOPICS IN FOURIER ANALYSIS 17

Thus, we have proved (1). It is clear that KN (t) is even and non-negative. Now, for
0 < δ ≤ π, sin2 (t/2) ≥ sin δ/2, so that
Z π Z π Z π
1 sin2 [(N + 1)t/2] 1 1
KN (t)dt = 2 dt ≤ 2 dt.
δ N +1 δ sin (t/2) N + 1 δ sin (δ/2)
Thus,
π
π−δ
Z
KN (t)dt ≤ → 0 as N → ∞.
δ (N + 1) sin2 (δ/2)


Exercise 5.11. Suppose f is piecewise continuous and 2π-periodic. If fˆ(n) = 0 for all
n ∈ Z, then f (x) = 0 for all x at which f is continuous.

Exercise 5.12. If f ∈ C 1 (T ), then fˆ(n) = O(1/n). More generally, f ∈ C k (T ) implies

fˆ(n) = O(1/nk ).

Example 5.13. Let f (x) = x2 , |x| ≤ π. Note that

Z π
ˆ π3
2π f (0) = x2 dx = 2
−π 3
so that fˆ(0) = π 2 /3, and for n 6= 0,
Z π
ˆ
2π f (n) = x2 e−inx dx
−π
−inx iπ π
e−inx
Z
2e
h
= x − 2x dx
−in −π −π −in
−inx iπ
h e h e−inx iπ
= x2 − 2x
−in −π (−in)2 −π
h e−inx iπ h e−inx iπ einx
= − 2x = 2x = 4π
(−in)2 −π n2 −π n2
(−1)n
= 4π 2
n
Hence, for n 6= 0,
n
(−1)
fˆ(n) = 2 2 .
n
Thus,
∞ ∞
π2 X (−1)n inx π 2 X (−1)n
x2 ≈ +2 2
e = + 4 2
cos nx.
3 n6=0
n 3 n=1
n
Since the series of coefficients converges absolutely, we have
∞
π3 X (−1)n
f (x) = +4 2
cos nx.
3 n=1
n
18 M.T. NAIR

Taking x = 0,
∞
π3 X (−1)n
0= +4 2
.
3 n=1
n
Thus,
∞
X (−1)n+1 π2
= .
n=1
n2 12
Taking x = π,
∞ ∞
2π3 X (−1)n n π3 X 1
π = +4 2
(−1) = +4 .
3 n=1
n 3 n=1
n2
Thus,
∞
X 1 π2
2
= .
n=1
n 6

Example 5.14. Let f (x) = x, x ∈ [−π, π]. Note that fˆ(0) = 0 and for n 6= 0,
Z π h e−inx iπ Z π −inx h e−inx iπ
e
2π fˆ(n) = xe−inx
dx = x − dx = x .
−π −in −π −π −in −in −π
Thus,
h e−inx iπ 1 einπ
2π fˆ(n) = x = [πe−inπ + πeinπ ] = 2π
−in −π −in −in
so that
(−1)n (−1)n+1
fˆ(n) = = .
−inπ inπ
Hence,
∞
X (−1)n+1 X (−1)n+1 X (−1)n+1
x= einx = [einx − e−inx ] = 2 sin nx
n6=0
in n=1
in n=1
n

Taking x = π/2 we obtain the Madhava-Nīlakaṅtha series

∞ ∞
π X (−1)n+1 nπ X (−1)n
= sin = . ♦
4 n=1
n 2 n=0
2n + 1

6. Divergence of Fourier series

Theorem 6.1. There exists f ∈ C(T ) such that {SN (f, 0)} is unbounded; in particular,
the Fourier series of f does not converge to f at 0.

For this we shall make use of the Uniform Boundedness Principle from Functional
Analysis:
 TOPICS IN FOURIER ANALYSIS 19

Theorem 6.2. (Uniform Boundedness Principle) Let (Tn ) be a sequence of con-

tinuous linear transformations from a Banach space X to a normed linear space Y . If
for each u ∈ X, the set {kTn uk : n ∈ N} is bounded, then there exists M > 0 such that

sup kTn uk ≤ M ∀ n ∈ N.
kuk≤1

Let
ϕN (f ) := SN (f, 0), f ∈ C(T ).
We see that ϕN : C(T ) → C is a linear functional for each N ∈ N and
Z π  Z π 
1 1
|ϕN (f )| = |SN (f, 0)| = f (−t)DN (t)dt ≤ kf k∞ |DN (t)|dt .
2π −π 2π −π
Hence, each ϕN is a continuous linear functional on C(T ) and
Z π
1
sup |ϕN (f )| ≤ |DN (t)|dt.
kuk∞ ≤1 2π −π
In fact,

Theorem 6.3. Z π
1
sup |ϕN (f )| = |DN (t)|dt
kuk∞ ≤1 2π −π

and
Z π N
8X1
|DN (t)|dt ≥ .
−π π k=1 k

Proof of Theorem 6.1. By Theorem 6.3, there does not exist M > 0 such that
supkuk∞ ≤1 |ϕN (f )| ≤ M for all n ∈ N. Hence, by Theorem 6.2, there exists f ∈ C(T )
such that {|ϕn (f )| : n ∈ N} is unbounded. Hence, there exists f ∈ C(T ) such that
Fourier series of f diverges at 0. 

Remark 6.4. Let D := {f ∈ C(T ) : {SN (f, 0)} does not converge}. Then C(T ) \ D
is a subspace of C(T ), and by Theorem 6.1, C(T ) \ D is a proper subspace. Hence,
C(T ) \ D is nowhere dense, and hence D is dense in C(T ). Thus, we have proved the
following:

There exists a dense subset D of C(T ) such that for each f ∈ D, the
Fourier series of f diverges at 0.

In place of 0, we can take any point in [−π, π] and obtain similar divergence result at
that point.
20 M.T. NAIR

7. Uniqueness

Theorem 7.1. (Uniqueness of Fourier series) Let f ∈ L1 (T ). If fˆ(n) = 0 for all

n ∈ N, then f = 0 a.e.

Proof. Let
Z t
g(t) = f (x)dx, t ∈ [−π, π].
−π

Then, by Fundamental Theorem of Lebesgue Integration (FTLI), g is absolutely con-

tinuous, g 0 exists a.e. and g 0 = f a.e. Note that
Z t+2π Z π
g(t + 2π) − g(t) = f (x)dx = f (x)dx = 2π fˆ(0) = 0.
t −π

Hence g is 2π-periodic. Let

Z t
h(t) = g(x)dx, t ∈ [−π, π].
−π

Then we see that

Z t+2π Z π
h(t + 2π) − h(t) = g(x)dx = g(x)dx = 2πĝ(0).
t −π

Taking
Z t
G(t) = [g(x) − ĝ(0)]dx, t ∈ [−π, π],
−π

we have
Z π
G(t + 2π) − G(t) = [g(x) − ĝ(0)]dx = 2π[ĝ(0) − ĝ(0)] = 0.
−π

Thus, G is 2π-periodic, and G00 = f a.e.Hence,

fˆ(n) = G
c00 (n) = (in)2 Ĝ(n) for all n 6= 0.

Therefore, Ĝ(n) = 0 for all n 6= 0. Hence, by Corollary 5.9, G(x) = Ĝ(0), and hence
G00 = 0, so that f = 0 a.e. 

Recall that for each f ∈ L1 (T ),

fˆ(n) → 0 as |n| → ∞.

Thus, (fˆ(n)) ∈ c0 (Z) for every f ∈ L1 [−π, π].

Notation: c0 (Z) is the set of all sequences ϕ : Z → C such that ϕ(n) → 0 as |n| → ∞.
 TOPICS IN FOURIER ANALYSIS 21

Theorem 7.2. The map F : L1 (T ) → c0 (Z) be defined by

F(f ) = (fˆ(n)), f ∈ L1 (T )
is an injective continuous linear operator which is not onto.

Proof. For f, g ∈ L1 (T ) and α ∈ C, we have

+ g)(n)) = fˆ(n) + ĝ(n) for all n ∈ Z,

(f\
c (n)) = αfˆ(n). for all n ∈ Z,
αf
Thus, F is a linear operator. Note that
Z π Z π
ˆ 1 −inx 1
|f (n)| = f (x)e dx ≤ |f (x)|dx.
2π −π 2π −π
Thus, if we endow L1 [−π, π] with the norm
Z π
1
kf kL1 := |f (x)|dx, f ∈ L1 (T ),
2π −π
then we see that F is a continuous linear operator. By Theorem 7.1, F is injective. So,
it remains to show that F is not onto. If it is onto, then my Bounded Inverse Theorem,
its inverse is also continuous. Note that
F(DN ) = {D
dN (n)}

and
b N (n) = 1 for |n| ≥ N
D
so that
k(F(DN ))k∞ = 1 for all N ∈ N.
If F is onto, then, by Bounded Inverse Theorem4 its inverse F −1 is continuous so that
(kDN k) = {kF −1 (F(DN )k} is bounded, which is not true. 

By the above theorem there exists (cn ) ∈ c0 (Z) such that there is no f ∈ L( T )
satisfying cn = fˆ(n) for all n ∈ N. It is a natural urge to have an example of such a
sequence cn ). We shall show that cn ) with

1/ log(n), n ≥ 2,
cn =
0, n ≤ 1,
is such a sequence. This is a consequence of the first part of the following theorem.
4If X and Y are Banach spaces and T : X → Y is a continuous bijective linear operator, then T −1
is also continuous.
22 M.T. NAIR

X fˆ(n)
Theorem 7.3. Let f ∈ L1 (T ). Then einx converges at every x ∈ R and
n6=0
n
Z b XZ b
f (x)dx = fˆ(n)einx dx.
a n∈Z a

For proving the above theorem we shall make use of the following theorem:

Theorem 7.4. (Jordan) If f ∈ L1 (T ) is of bounded variation5, then for every x ∈ R,

1
SN (f, x) → (f (x+) + f (x−) as N → ∞.
2
In particular, if f ∈ AC(T ), then
SN (f, x) → f (x) as N →∞
for every x ∈ R.

It can be easily shown that:

Every absolutely continuous function is of bounded variation.

Proof of Theorem 7.3. Let

Z t
g(t) = [f (x) − fˆ(0)]dx.
−π

Then g is absolutely continuous and g is 2π-periodic, i.e., g ∈ AC(T ), g 0 ∈ L1 (T ) and

g 0 = f − fˆ(0) a.e. Therefore, ĝ 0 (n) = inĝ(n) for all n 6= 0 so that
fˆ(n)
ĝ(n) = , n 6= 0.
in
By Jordan’s theorem,
X X fˆ(n)
g(x) = ĝ(0) + ĝ(n)einx = ĝ(0) + einx .
n6=0 n6=0
in

X fˆ(n)
In particular, einx converges. Also,
n6=0
n

X fˆ(n) X Z x
g(x) − g(y) = [e inx
−e inx
]= fˆ(n) eint dt.
n6=0
in n6=0 y

5A
function f : [a, b] → C is of bounded variation if there exits κ > 0 such that for every partition
Pn
x0 < x1 < · · · < xn = b, k=1 |f (xk+1 ) − f (xk | ≤ κ.
 TOPICS IN FOURIER ANALYSIS 23

But,
Z x Z x Z x
g(x) − g(y) = 0
g (t)dt = [f (t) − fˆ(0)]dt = f (t)dt − fˆ(0)(x − y)dt.
y y y

Hence,
Z x XZ b
f (t)dt = fˆ(n)eint dt.
y n∈Z a

This competes the theorem. 

Corollary 7.5. Let (cn ) be with


1/ log(n), n ≥ 2,
cn =
0, n ≤ 1,

Then there is no f ∈ L1 (T ) satisfying cn = fˆ(n) for all n ∈ N.

Proof. Suppose f ∈ L1 (T ) satisfying cn = fˆ(n) for all n ∈ N. Then by the first part of
Theorem 7.3, the series ∞
P einx
P∞ 1
n=2 n log n converges. In particular, taking x = 0, n=2 n log n
converges, which is not true (e.g., by integral test).


8. Convolution

Given f, g ∈ L1 (T ), it can be shown that

(x, y) 7→ f (x − y)g(y)

is measurable on R × R, and hence, for each x ∈ [−π, π], the integral

Z π
f (x − y)g(y)dy
−π

converges.

Definition 8.1. The convolution of f, g ∈ L1 (T ) is defined by

Z π
1
(f ∗ g)(x) = f (x − y)g(y)dy, x ∈ [−π, π].
2π −π

We observe the following:

24 M.T. NAIR

(1) f ∗ g ∈ L1 (T ) and kf ∗ gk1 ≤ kf k1 kgk1 :

Z π Z π Z π Z π 
|f (x − y)| |g(y)|dydx = |f (x − y)| dx |g(y)|dy
−π −π −π −π
Z π
= 2πkf k1 |g(y)|dy
−π
= (2π)2 kf k1 kgk1 .

(2) f ∗ g = g ∗ f :
Z π Z π
f (x − y)g(y)dy = f (τ )g(x − τ )dy
−π −π
Z x+π
= f (τ )g(x − τ )dy
x−π
Z π
= f (τ )g(x − τ )dy.
−π

∗ g(n) = ĝ(nfˆ(n) for all n ∈ Z:

(3) f[
Z π
1
∗ g(n) =
f[ (f ∗ g)(x)e−inx dx,
2π −π

Z π
−inx 1
(f ∗ g)(x)e = f (x − y)g(y)e−inx dy
2π −π
Z π
1
= f (x − y)g(y)e−in(x−y) e−iny dy,
2π −π

Z π Z π Z π 
−inx −in(x−y) −iny
(f ∗ g)(x)e dx = f (x − y)g(y)e e dy dx
−π −π −π
Z π Z π 
−in(x−y)
= f (x − y)e dx g(y)e−iny dy
−π −π
Z π
= 2π fˆ(n)g(y)e−iny dy
−π

= (2π)2 fˆ(n)ĝ(n).
 TOPICS IN FOURIER ANALYSIS 25

(4) (f ∗ g) ∗ h = f ∗ (g ∗ h):
Z π Z π Z π 
(f ∗ g)(x − y)h(y)dy = f (x − y − t)g(t)dt h(y)dy
−π −π −π
Z π Z π 
= f (x − τ )g(τ − y)dτ h(y)dy
−π −π
Z π Z π 
= f (x − τ ) g(τ − y)h(y)dy dτ
−π −π
Z π
= 2π f (x − τ )(g ∗ h)(τ )dτ
−π
2
= (2π) [f ∗ (g ∗ h)](x).

Theorem 8.2. With respect to convolution as multiplication, L1 (T ) is a Banach alge-

bra.

• The Banach algebra L1 (T ) does not have a multiplicative identity:

Suppose there exists ϕ ∈ L( T ) such that f ∗ ϕ = f for all f ∈ L1 (T ). Then
fˆ(n)ϕ̂(n) = fˆ(n) for all f ∈ L1 (T ). Hence, ϕ̂(n) = 1 whenever ϕ̂(n) 6= 0. But,
ϕ̂(n) → 0 as |n| → ∞. Hence, there exists N ∈ N such that ϕ̂(n) = 0 for all
n ≥ N . Let f ∈ L1 (T ) be such that fˆ(n) 6= 0 for some n ≥ N . Then for such
n, we obtain

0 = fˆ(n)ϕ̂(n) = fˆ(n) 6= 0,

which is a contradiction.

However,

• There exists (ϕn ) in L1 (T ) such that kf ∗ ϕn − f k1 → 0.

In fact, we have the following.

Theorem 8.3. Let Kn be the Fejér kernel. Then, for every f ∈ L1 (T ),

kf ∗ Kn − f k1 → 0 as N → ∞.

Proof. Recall that if g ∈ C(T ), then kg ∗ ϕn − gk1 → 0. Let f ∈ L1 (T ) and ε > 0

be given. Let g ∈ C(T ) be such that kf − gk1 < ε, and let N ∈ N be such that
26 M.T. NAIR

kg ∗ ϕn − gk1 < ε for all n ≥ N . Then, for n ≥ N , we have

kf ∗ Kn − f k1 ≤ kf ∗ Kn − g ∗ Kn k1 + kg ∗ Kn − gk1 + kg − f k1
≤ k(f − g) ∗ Kn k1 + ε + ε
≤ k(f − g) ∗ Kn k1 + 2ε
≤ kf − gk1 kKn k1 + 2ε
≤ 3ε.
1
Rπ 1
Rπ
The last inequality is due the fact that 2π −π
|Kn (t)|dt = 2π −π
Kn (t)dt = 1. 

9. L2 -Theory

The norm on L2 (T ) is given by

 Z π 1/2
1 2
kf k2 = |f (x)| dx .
2π −π

Observe:

(1) If un (x) := einx , n ∈ Z, then the set {un : n ∈ Z} is an orthonormal set and
span{un : n ∈ Z}, the space of all trigonometric polynomials, is dense in L2 (T ).
(2) Let f ∈ L2 (T ) and SN (f ) := SN (f, ·). Then
X N
(a) SN (f ) = hf, un iun .
n=−N
N
X
(b) kSN (f )k22 = |fˆ(n)|2 .
n=−N
N
X
(c) kf − SN (f )k22 = kf k22 − kSN (f )|22 = kf k22 − |fˆ(n)|2 .
n=−N
N
X
(d) (Bessel’s inequality): |fˆ(n)|2 ≤ kf k22 ∀ N ∈ N. In particular,
n=−N
fˆ(n) → 0 as |n| → ∞.
(e) hf − SN (f ), un i = 0 ∀ |n| ≤ N }.
(f) kf − SN (f )k2 ≤ kf − gk ∀ g ∈ span{un : n ∈ Z, |n| ≤ N }.

Only (2)(f) requires some explanation.

Note that for every g ∈ span{un : n ∈ Z, |n| ≤ N },
kf − gk22 = kf − SN (f )k22 + kSN (f ) − gk22 ,
because, in view of (2)(e), hf − SN (f ), SN (f ) − gi = 0.
 TOPICS IN FOURIER ANALYSIS 27

• The result in (2)(d) gives another proof for the Riemann Lebesgue lemma,
because L2 (T ) is dense in L1 (T ).
• In view of (2)(f),
kf − SN (f )k2 = inf{kf − gk2 : g ∈ XN },
where XN := span{un : n ∈ Z, |n| ≤ N }. In other words, SN (f ) is the (unique!)
best approximation of f from XN . Uniqueness is due to the following: Suppose
ϕ be in XN such that
kf − ϕk2 = inf{kf − gk2 : g ∈ XN }.
Then,
kf − ϕk22 = kf − SN (f )k22 + kSN (f ) − ϕk22
since hf − SN (f ), SN (f ) − ϕi = 0 so that we obtain kSN (f ) − ϕk2 = 0.

Theorem 9.1. Let f ∈ L2 (T ). Then we have the following:

(1) span{un : n ∈ Z} is an orthonormal basis of L2 (T ), i.e., a maximal orthonormal

set in L2 (T ). X
(2) (Fourier expansion) f = fˆ(n)un in L2 (T ).
n∈Z X
(3) (Parseval’s formula) 2
kf k2 = |fˆ(n)|2 .
n∈Z

Proof. (1) It can be seen that hf, un i = 0 for all n ∈ Z implies f = 0 in L2 (T ). Hence,
span{un : n ∈ Z} is a maximal orthonormal set in L2 (T ).
(3) We observe that, for n > m,
X
kSn (f ) − Sm (f )k2 ≤ |fˆ(n)|2 .
n≤|k|≤m

Hence, {Sn (f )} is a cauchy sequence in L2 (T ). Therefore, it converges to some g ∈

L2 (T ). It can be seen that ĝ(n) = fˆ(n) for all n ∈ Z. Therefore, g = f in L2 (T ).
(3) Follows from (2). 

Now, we give another proof for the following theorem:

Theorem 9.2. If f ∈ C 1 (T ), then the Fourier series of f converges absolutely, and

uniformly to f . Further,
 
1
kf − SN (f, ·)k∞ = O √ .
N
28 M.T. NAIR

Proof.
f ∈ C 1 (T ) =⇒ fˆ0 (n) = infˆ(n).
Hence,
!1/2
X X1 X1 X 1 π
|fˆ(n)| = | |infˆ(n)| = | |fˆ0 (n)| ≤ kfˆ0 k2 = √ kfˆ0 k2 .
n6=0 n6=0
n n6=0
n n6=0
n2 3
Hence the Fourier series of f converges absolutely, and uniformly to a continuous
function, say g ∈ C(T ). Since ĝ(n) = fˆ(n) for all n ∈ Z, we obtain g = f . We also
observe that, for all x ∈ R,
 1/2
X X1 X 1 fˆ0 k2
 kfˆ0 k2 ≤ k√
|f (x) − SN (f, x)| ≤ |fˆ(n)| = | |fˆ0 (n)| ≤  2
.
n n N
|n|>N n6=0 |n|>N

This completes the proof. 

References
[1] R. Bhatia, Fourier Series, TRIM, Hindustan Book Agency, 1993 (Second Edition: 2003) .
[2] S. Kesavan, Lectures on Fourier Series, Notes: Third Annual Foundation School, December,
2006.
[3] M.T. Nair, Measure and Integration, Notes for the MSc. Course, Jan-may, 2014.
[4] R. Radha & S. Thangavelu, Fourier Series, Web-Course, NPTEL, IIT Madras, 2013.
[5] B. O. Turesson, Fourier Analysis, Distribution Thoery, and Wavelets, Lecture Notes, March,
2012.

Department of Mathematics, I.I.T. Madras, Chennai-600 036, INDIA

E-mail address: mtnair@iitm.ac.in