You are on page 1of 68

Chapter 1 Fourier series

1.1 Appearance of Fourier series


The birth of Fourier series can be traced back to the solutions of wave equation
in the work of Bernoulli and the heat equation in the work of Fourier.

Consider an elastic string of finite length l fixed at the end points x = 0 and
x = l. At time, say t = 0, it is distorted from the equilibrium position and allowed to
vibrate. The problem is to find the vibrations of the string at any point x and any
time t > 0. The vibration of the string is governed by the linear partial differential
equation

∂ 2u 2
2∂ u ∂
= c , u(x, 0) = f (x), u(x, 0) = g(x). (1.1)
∂t2 ∂x2 ∂t

Here c2 denotes a physical constant, f (x) gives the initial position and g(x) gives the
initial velocity.

In 1747, D. Alembert obtained the solution of (1.1) in the form

x+ct
Z
1 1
u(x, t) = [f (x + ct) + f (x − ct)] + g(y)dy.
2 2c
x−ct

However, in 1753, Bernoulli had a different idea of solving this equation which was
based on the observation that the functions sin λct, sin λx, and cos λct, cos λx satisfy

the equation for any λ ∈ R. If we choose λ = l
, n = 1, 2, 3, . . . they also satisfy the

1
boundary conditions. As the equation is linear, he argued that any superposition of
such solutions will also be a solution.

To explain this method further, let us use the method of separation of variables.
Let u(x, t) = F (x)G(t). Substituting for u in (1.1), we get c2 F 00 (x)G(t) = F (x)G00 (t),
which can be rewritten as

F 00 (x) 1 G00 (t)


= 2 . (1.2)
F (x) c G(t)

In (1.2), the left hand side is a function of x and the right hand side is a function of
t. This is possible only if

00
F (x) 1 G00 (t)
= k, 2 = k,
F (x) c G(t)
where k is a constant. Thus we end up with two differential equations,

F 00 (x) − kF (x) = 0, (1.3)

G00 (t) − kc2 G(t) = 0. (1.4)


Since the string is fixed at end points x = 0 and x = l, we have the boundary
conditions

u(0, t) = 0; u(l, t) = 0, t > 0, (1.5)

which are satisfied if we assume

F (0) = F (l) = 0. (1.6)


The vibrations of the string at time t depend upon the initial deflection and initial
velocity. Let f denote the initial deflection and g denote the initial velocity. Thus
the initial conditions are given by

u(x, 0) = f (x), (1.7)

2
∂u
(x, 0) = g(x). (1.8)
∂t

Notice that the constant k which appears in (1.3) and (1.4) is a real number. Hence
it can be positive or negative or zero.
Case 1. If k = 0, then (1.3) and (1.4) become F 00 (x) = 0; G00 (t) = 0. Solving for F ,
we get F (x) = Ax + B, where A and B are constants. The boundary conditions force
A = B = 0 and hence F = 0.
Case 2. Let k = µ2 be positive. Substituting in (1.3), we get F 00 (x) − µ2 F (x) = 0. In
this case, the solution is given by F (x) = Aeµx +Be−µx , where A and B are constants.
Once again the boundary conditions force F to be zero.
Case 3. Let k = −λ2 . Then (1.3) becomes F 00 (x) + λ2 F (x) = 0. On solving, we get
F (x) = A cos λx + B sin λx, where A and B are constants. Applying F (0) = 0, we
get A = 0.

Thus, F (x) = B sin λx. The equation F (l) = 0 leads to B sin λl = 0. Then,
we get either B = 0 or sin λl = 0. In order to get a non-trivial solution, we assume that
B 6= 0. Thus, sin λl = 0, which leads to λl = nπ, n ∈ Z. As sin 0 = 0 and sin (−n) =

− sin n, it is enough to take n ∈ N. Thus we have λn = , n = 1, 2, 3, . . . and
l

Fn (x) = Bn sin λn x.
00
Now (1.4) becomes Gn (t) + λ2n c2 Gn (t) = 0. On solving we get,

Gn (t) = Cn cos (cλn t) + Dn sin (cλn t).

Let un (x, t) = Bn sin λn x(Cn cos (cλn t) + Dn sin (cλn t)). It is clear that for each n,
un (x, t) satisfies the wave equation with the correct boundary conditions.

3
The function un also satisfies the initial conditions

∂un
un (x, 0) = Bn Cn sin λn x, (x, 0) = cBn Dn λn sin λn x.
∂t


P
Consequently, the superpositions u = un solves the wave equation with the bound-
n=1
ary conditions u(0, t) = u(l, t) = 0. However, the initial conditions will be satisfied
only if

X
f (x) = Bn Cn sin λn x,
n=1


P
and g(x) = c Bn Dn λn sin λn x. At this point, it is not clear whether any f or g
n=1
can be expanded in terms of sin λn x as above and hence the solution of the wave
equation obtained by Bernoulli was incomplete. In 1807, Fourier was working on the
initial value problem for the heat equation

∂u ∂ 2u
(x, t) = (x, t), u(x, 0) = f (x).
∂t ∂x2

Let us assume the same boundary conditions as above, namely u(0, t) = u(l, t) = 0
which means the temperature at the end points is kept at zero throughout. Proceeding
as in the case of wave equation, we get the elementary solutions

2
un (x, t) = An (sin λn x) e−λn t .

P
The superposition u(x, t) = un (x, t), which solves the heat equation, satisfies the
n=1
initial condition provided

X
f (x) = An sin λn x.
n=1

4
Fourier asserted that any f satisfying f (0) = f (l) = 0 can be expanded in terms of
sin λn x as above.

1.2 The formal Fourier series and some special ker-

nels
In Section 1.1, we mentioned the idea of expanding a function f (x) satisfying
f (0) = f (l) = 0 in terms of sin nπ
l
x, n = 1, 2, . . . Note that the functions sin nπ
l
x are
periodic with period 2l and hence f also has to be periodic in order to be expanded
as above.

Given f with f (0) = f (l) = 0, we can always consider it as a periodic function


of period 2l defined on the whole of R and hence it is reasonable to expect an expansion
of the above type.

For the sake of simplicity, we take l = π and consider functions f that are 2π-
periodic. Rather than restricting our attention to functions satisfying f (0) = f (π) =
0, we consider all 2π periodic functions. In this case, besides sin nx, we also need to
include cos nx in the expansion of the type


X
f (x) = (an cos nx + bn sin nx).
n=0

Using the formulae, cos nx = 21 (einx + e−inx ), sin nx = 1


2i
(einx − e−inx ), we can rewrite
the above expansion as


X
f (x) = cn einx .
n=−∞

5
This is a reasonable expansion as f is 2π-periodic and so is each einx . The family of
functions einx have the interesting property, which we call orthogonality,

einx e−imx dx = 0 or 2π,
−π
depending on whether n 6= m or n = m. Therefore, if termwise integration is allowed,
then the expansion

X
f (x) = cn einx ,
n=−∞

leads us to

1
f (x) e−inx dx = cn .

−π

This formula, due to Fourier, allows us to calculate cn in the expansion. We can now
formally introduce the Fourier series.
Definition 1.2.1. Let f be a Lebesgue integrable function on [−π, π]. Then the nth
Fourier coefficient of f is defined by


1
fb(n) = f (x) e−inx dx, n ∈ Z.

−π


fb(n) einx .
P
The Fourier series of f is formally defined by f (x) ∼
n=−∞
Example 1.2.2. Let f (x) = x, −π ≤ x ≤ π. Then, the Fourier coefficients of f are
obtained as follows: for n 6= 0


1
fb(n) = x e−inx dx

−π

6

e−inx
 
1
= xd
2π −in
−π

−1
= x d(e−inx )
2πin
−π
 π

−1
Z
π
 
= x e−inx − e−inx dx
2πin  −π 
−π
−1
= {2π(−1)n − 0} , n 6= 0.
2πin

(−1)n+1 1 Rπ
Thus fb(n) = , n 6= 0. But fb(0) = x dx = 0. In this case, the Fourier
in 2π −π
series of f can be formally written as

∞ ∞
X (−1)n+1 inx X (−1)n+1
f (x) ∼ e =2 sin nx.
n=−∞
in n=1
n
n6=0

Example 1.2.3. Let f (x) = x2 , −π ≤ x ≤ π. Then the nth Fourier coefficient of f


is given by


1
fb(n) = x2 e−inx dx

−π

e−inx
 
1 2
= xd .
2π −in
−π

Using integration by parts, we get,

2(−1)n
fb(n) = , n 6= 0.
n2

7
But

1 π2
fb(0) = x2 dx = .
2π 3
−π

In this case, the formal Fourier series of f can be written as


π2 X 2(−1)n inx
f (x) ∼ + 2
e
3 n=−∞
n
n6=0

π2 X4(−1)n
= + cos nx.
3 n=1
n2

Figure 1.1: Fourier series of x2 .

8
Consider the Fourier series

X
f (x) ∼ fb(n) einx
n=−∞

X ∞
X
= fb(0) + fb(−n) e−inx + fb(n) einx
n=1 n=1
X∞
 
= fb(0) + fb(−n) + fb(n) cos nx
n=1
X∞
 
+i fb(n) − fb(−n) sin nx,
n=1

using e±inx = cos nx ± i sin nx. If f is an even function, then fb will also be an even
function. In fact,

1
fb(−n) = f (x)einx dx

−π

1
= f (−x)einx dx

−π

1
= f (−x)e−(−inx) dx

−π

= fb(n),

by applying a change of variable x = −y. In this case, the resulting Fourier series

P
becomes f (x) ∼ fb(0) + 2 fb(n) cos nx. On the other hand, if f is an odd function,
n=1
then fb will also be an odd function and the Fourier series takes the form


X
f (x) ∼ 2i fb(n) sin nx.
n=1

9
1.2.1 Dirichlet kernel

The formal Fourier series will represent the function f provided the partial
sums of the series converges to f . We take the symmetric partial sum

N
X
SN f (x) = fb(n)einx ,
n=−N

which can be written in the form


1
SN f (x) = f (y)DN (x − y)dy,

−π

where the function DN , called the Dirichlet kernel, is given by

N
X
DN (x) = einx .
n=−N
Indeed, from the definition,

1
inx
fb(n)e = f (y)e−in(y−x) dy,

−π
and hence
N Zπ N
X 1 X
inx
e−in(y−x) dy.

fb(n)e = f (y)
n=−N
2π n=−N
−π

The kernel DN can be explicitly calculated:

sin (2N + 1) x2
DN (x) = . (1.9)
sin ( x2 )

10
Figure 1.2: Dirichlet kernel.

11
In order to establish (1.9), consider
XN
DN (x) = einx
n=−N
−1
X N
X
inx
= e + einx + 1
n=−N n=1
N
X N
X
−inx
= −1 + e + einx
n=0 n=0
N N
n n
X X
= −1 + (e−ix ) + (eix ) .
n=0 n=0
Writing w = eix , we have
N
X N
X
DN (x) = −1 + w−n + wn
n=0 n=0
−(N +1)
1−w 1 − w(N +1)
= −1 + +
1 − w−1 1−w
−N
w−w 1 − w(N +1)
= −1 + +
w−1 1−w
w−N − wN +1
=
1−w
−1/2
w (w−N − wN +1 )
=
w−1/2 (1 − w)
w−N −1/2 − wN +1/2
=
w−1/2 − w1/2
sin (2N + 1) x2
= .
sin ( x2 )

Observe that, from the very definition, the Fourier coefficients of DN are given by


 1 if |n| ≤ N
D
b N (n) =
 0 otherwise.

12
1.2.2 Dirichlet problem for the disc and Poisson kernel
Let D denote the unit disc in the complex plane i.e. D = {z ∈ C : |z| < 1}.
Given a continuous function f on the boundary of D we are interested in finding a
function u on D satisfying the Laplace equation ∆u = 0 with the boundary condition
lim u(reiθ ) = f (θ). This is called the Dirichlet problem for the unit disc D. The
r→1

Laplace equation in polar coordinates takes the form

∂ 2 u 1 ∂u 1 ∂ 2u
+ + = 0.
∂r2 r ∂r r2 ∂θ2

Integrating the equation against e−inθ dθ we see that


∂2 ∂ 1 ∂ 2u
r2
un (r) + r u n (r) + (r, θ)e−inθ dθ = 0,
∂r2 ∂r 2π ∂θ2
−π

where

1
un (r) = u(r, θ)e−inθ dθ.

−π

Integrating by parts, the third term becomes −n2 un (r) and hence we get

∂2 ∂
r2 2
un (r) + r un (r) − n2 un (r) = 0.
∂r ∂r

Clearly, un (r) = an r|n| satisfies the above equation and the initial conditions lim u(r, θ)
r→1
= f (θ) leads to lim un (r) = fb(n). Hence, an = fb(n) and the solution u(r, θ) has the
r→1
formal expansion

X
u(r, θ) = fb(n)r|n| einθ .
n=−∞

13
As before, we can write this in a compact form. Let us define a kernel Pr (θ), called
the Poisson kernel, by


X
Pr (θ) = r|n| einθ .
n=−∞
Then u(r, θ) takes the form

1
u(r, θ) = f (ϕ)Pr (θ − ϕ)dϕ.

−π

The kernel Pr (θ) can also be calculated explicitly. Indeed, we have


X
Pr (θ) = r|n| einθ
n=−∞
X∞ ∞
X
=1+ n inθ
r e + rn e−inθ
n=1 n=1

= 1 + reiθ (1 − reiθ )−1 + re−iθ (1 − re−iθ )−1


2r cos θ − 2r2
=1+
1 − 2r cos θ + r2
1 − r2
= .
1 − 2r cos θ + r2
It can be easily shown that the Fourier coefficients of Pr are given by

Pbr (n) = r|n| .

1.2.3 The heat kernel

Returning to the heat equation, let u(x, t) solve the problem

∂u(x, t) ∂ 2 u(x, t)
= ,
∂t ∂x2
14
u(x, 0) = f (x), −π ≤ x ≤ π.

For each n, let



1
un (t) = u(x, t) e−inx dx.

−π
Then we have

d 1 ∂ 2 u(x, t) −inx
(un (t)) = e dx,
dt 2π ∂x2
−π

which after integration by parts, leads to

d
(un (t)) = −n2 un (t).
dt

2
The solution of this equation is given by un (t) = an e−tn . The initial condition
u(x, 0) = f (x) gives an = fb(n) and hence,


2
X
u(x, t) = fb(n)e−tn einx .
n=−∞
As before, we can rewrite this as

1
u(x, t) = f (y)ht (x − y)dy,

−π

where ht (x), called the heat kernel, is given by


2
X
ht (x) = e−tn einx .
n=−∞

Unlike the Poisson or Dirichlet kernel, we do not have a closed form expression for
ht (x).

15
1.3 Uniqueness of Fourier series and some conse-

quences
Let T denote the unit circle. A function on T can be treated as a function on
R which is 2π-periodic. Let C(T) denote collection of all continuous functions on T.
The following theorem gives the uniqueness of Fourier series.
Theorem 1.3.1. Let f ∈ C(T). Suppose that fb(n) = 0 ∀ n ∈ Z then f = 0.

Proof. For 0 < δ < π, to be fixed soon, consider the function pδ (x) = (1+cos x−cosδ).
Then we claim that pδ (x) ≥ 1 on [−δ, δ] and pδ (x) < 1 on [−δ, δ]c . Since cos x
is an increasing function on [−π, 0] and a decreasing function on [0, π] we have for
x ∈ [−δ, 0], cos x − cos δ ≥ 0. Then pδ (x) = 1 + cos x − cos δ ≥ 1 on [−δ, 0].
On the other hand, if x ∈ [0, δ] then cos x ≥ cos δ and hence pδ (x) ≥ 1 on
[0, δ]. However, if −π < x < −δ then cos x < cos (−δ) = cos δ.
Hence, pδ (x) < 1 on [−π, −δ]. Similarly, if δ < x < π, then cos x < cos δ which
leads to pδ < 1 on [δ, π].
Let f = g + ih, where g and h are real valued functions on T. We prove that
f (x) = 0 at every x. Without loss of generality, we take x = 0. Suppose f (0) 6= 0.
Then, again without loss of generality, we can assume that g(0) > 0 and we will arrive
at a contradiction using the continuity of g.
1
Choose δ > 0 such that g(x) > g(0) on [−δ, δ]. Define pδ with this δ and
2
take Pn (x) = pδ (x)n . By the hypothesis,


f (x)Pn (x)dx = 0,
−π

16
which gives

g(x)Pn (x)dx = 0.
−π
R
On the other hand, g(x)Pn (x)dx goes to 0 by dominated convergence theorem
[−δ,δ]c
since pδ (x) < 1. This shows that


lim g(x)Pn (x)dx = 0.
n→∞
−δ

On the other hand, Pn (x) ≥ 1 on this interval and hence,


g(x)Pn (x)dx ≥ δ g(0).
−δ

This contradiction proves our claim that f (0) = 0.

Corollary 1.3.2. Suppose f ∈ L1 (T) and fb(n) = 0 ∀ n ∈ Z. Then f (x) = 0 a. e.


Rx
Proof. Let f ∈ L1 (T). Define g(x) := f (t)dt, −π ≤ x ≤ π, then g is continuous
−π
i fb
and g 0 = f a.e. Further, for all n 6= 0, gb(n) = n
so that gb(n) = 0.
Applying Theorem 1.3.1 to the continuous function g(x) − gb(0), we conclude
that g(x) = gb(0) is a constant. Hence, f = 0 a.e.

Theorem 1.3.3. Let f ∈ C(T) be such that fb ∈ l1 (Z). Then the Fourier series of f
converges uniformly to f on T.
∞ N
Proof. As fb ∈ l1 (Z), fb(n)einx . We claim
P P
|fb(n)| < ∞. Define SN f (x) =
n=−∞ n=−N
that SN (f ) → f uniformly on T.

17
For N > M , we have, for all x ∈ T,

X N M
X
|SN f (x) − SM f (x)| = fb(n)einx − fb(n)einx


n=−N n=−M


X
inx
= f (n)e
b
N ≥|n|>M
X
≤ |fb(n)|.
N ≥|n|>M

Thus by applying Cauchy’s criterion for uniform convergence, one can conclude
that SN (f ) converges to a continuous function g (uniformly) on T. Further gb(n) =
fb(n) ∀ n ∈ Z. In fact, for all n


1
gb(n) = g(x)e−inx dx

−π
Zπ ∞
!
1 X
= fb(m)eimx e−inx dx
2π m=−∞
−π
∞ Z π
1 X
= fb(m)ei(m−n)x dx
2π m=−∞
−π

= fb(n).

Here, changing the order of integral and the sum was possible because the convergence
is uniform. As (f − g)b(n) = 0 ∀ n ∈ Z, we conclude that f = g.

Corollary 1.3.4. If f ∈ C 2 (T), then the Fourier series of f converges uniformly to


f on T.

18
Proof. By the theorem, it is enough to prove that fb ∈ l1 (Z). We shall show that
 
f (n) = O |n|1 2 as |n| → ∞. Consider, for n 6= 0,
b


1
fb(n) = f (x)e−inx dx

−π
 
−inx π Zπ
1  e 1 0 −inx

= f (x) + f (x)e dx

2π  −in −π in


−π

1 0
= f (x)e−inx dx
2πin
−π
 
−inx π Zπ
1 
0 e 1 00

= f (x) + f (x)e−inx dx .

2πin  −in −π in 
−π
Thus

1 00
fb(n) = − f (x)e−inx dx.
2πn2
−π

00 M
Let M = max |f (x)|. Thus |fb(n)| ≤ n2
for all n 6= 0. Consequently,
x∈[−π,π]


X X
|fb(n)| ≤ fb(0) + f (n)
b

n=−∞ n6=0

X 1
≤ f (0) + 2M
b
n=1
n2

< ∞,

which shows that fb ∈ l1 (Z).

19
1.4 Convolution theory
Before defining convolution of two functions, we observe the following fact.
Proposition 1.4.1. If f is a periodic function of period 2 π, then

Zπ Zπ π+a
Z
f (x + a)dx = f (x)dx = f (x)dx.
−π −π −π+a

Rπ Rπ π+a
R
Proof. Consider f (x+a)dx. Taking x 7→ x+a, we get f (x+a)dx = f (x)dx.
−π −π −π+a
If a > 0, then

π+a
Z Zπ π+a
Z
f (x)dx = f (x)dx + f (x)dx
−π+a −π+a π
Zπ −π+a
Z π+a
Z
= f (x)dx − f (x)dx + f (x)dx
−π −π π
Zπ −π+a
Z π+a
Z
= f (x)dx − f (x + 2π)dx + f (x)dx
−π −π π
Zπ π+a
Z π+a
Z Zπ
= f (x)dx − f (y)dy + f (x)dx = f (x)dx.
−π π π −π

Similarly, we can prove the result when a < 0.

Definition 1.4.2. For f, g ∈ L1 (T), the convolution of f and g is defined as


1
(f ∗ g)(x) := f (x − y) g(y) dy.

−π

20
Immediately we observe that f ∗ g = g ∗ f . In fact,

x+π
Z Zπ
1 1
(f ∗ g)(x) = f (t)g(x − t)dt = g(x − t)f (t)dt = (g ∗ f )(x).
2π 2π
x−π −π

(a) χ[−1,1] . (b) χ[−1,1] ∗ χ[−1,1] .

Figure 1.3

Theorem 1.4.3. L1 (T) is a commutative Banach algebra.

Proof. We know that L1 (T) is a Banach space. Let f, g ∈ L1 (T). Consider


1
kf ∗ gk1 = |f ∗ g(x)| dx

−π
Zπ Zπ
1
= f (x − y) g(y) dy dx

(2π)2

−π −π
Zπ  Zπ
1 
≤ |f (x − y)| |g(y)| dy dx.
(2π)2
−π −π

21
Applying Fubini’s theorem we get,
 
Zπ Zπ
1
kf ∗ gk1 ≤  |f (x − y)|dx |g(y)|dy
(2π)2
−π −π

1
= kf k1 |g(y)| dy

−π

= kf k1 kgk1 < ∞.

Thus f ∗g ∈ L1 (T) and kf ∗gk1 ≤ kf k1 kgk1 . We have already proved that f ∗g = g ∗f.
In order to prove associativity, consider, for x ∈ R,


1
(f ∗ g) ∗ h(x) = (f ∗ g) (y) h(x − y) dy

−π
Zπ  1 Zπ
1 
= f (t) g(−t + y) dt h(x − y) dy
2π 2π
−π −π
Zπ 1 Zx+π
1 
= f (t) g(−t − u + x) h(u) du dt
2π 2π
−π x−π

1
= f (t) (g ∗ h) (x − t) dt

−π

= [f ∗ (g ∗ h)] (x).

By the linearity of integral, it follows that (f + g) ∗ h = (f ∗ h) + (g ∗ h) and


αf ∗ g = f ∗ αg. Thus L1 (T) is a commutative Banach algebra.

Theorem 1.4.4. Let f, g ∈ C(T). Then f ∗ g ∈ C(T).

22
Proof. Let f, g ∈ C(T). Let  > 0 be given. Let M = max |g(y)|. For s, t ∈ T, we have
y∈T

Zπ Zπ
1
|(f ∗ g)(s) − (f ∗ g)(t)| = f (s − y) g(y) dy − f (t − y) g(y) dy



−π −π

1
≤ |f (s − y) − f (t − y)| |g(y)| dy

−π

M
≤ |f (s − y) − f (t − y)| dy.

−π

Since f is uniformly continuous on T, there exists δ > 0 such that |s − t| < δ implies

|f (s) − f (t)| < M
. Thus |(f ∗ g) (s) − (f ∗ g) (t)| ≤  if |s − t| < δ, proving that
f ∗ g ∈ C(T).

In general if f, g ∈ L1 (T), then we can show that there exist sequences of


functions {fn }, {gn } in C(T) such that fn ∗ gn converges to f ∗ g in L1 . However, if
we assume that both f and g are bounded, then the convergence becomes uniform,
which is established in the next theorem.
Theorem 1.4.5. Let f, g ∈ L1 (T). If, in addition, f and g are bounded, then there
exist sequences {fn } and {gn } in C(T) such that fn ∗ gn converges to f ∗ g uniformly
on T.

Proof. Let f, g ∈ L1 (T) be bounded. Then there exist {fn } and {gn } in C(T) such
that

kfn − f k1 → 0 as n → ∞ sup |fn (x)| ≤ k1 , ∀ n


x∈T

23
and

k gn − g k1 → 0 as n → ∞ sup |gn (x)| ≤ k2 , ∀ n.


x∈T

Consider

f ∗ g − fn ∗ gn = f ∗ g − fn ∗ g + fn ∗ g − fn ∗ gn

= (f − fn ) ∗ g + fn ∗ (g − gn ).

Let M = sup |g(y)|. For x ∈ T, we have,


y∈T

1
(f − fn ) ∗ g(x) = (f − fn )(x − y) g(y) dy

−π

1
= (f (x − y) − fn (x − y)) g(y) dy.

−π
Thus

1
|(f − fn ) ∗ g(x)| ≤ |f (x − y) − fn (x − y)| |g(y)| dy

−π

1
≤ sup |g(y)| |f (x − y) − fn (x − y)| dy
2π y∈T
−π
M
≤ k fn − f k1 → 0 as n → ∞.

Also π
Z
1
|fn ∗ (g − gn )(x)| = (g − gn ) (x − y) fn (y) dy


−π

1
≤ |g(x − y) − gn (x − y)| |fn (y)| dy

−π

k1
≤ |g(x − y) − gn (x − y)| dy

−π

24
k1
= k gn − g k1 → 0 as n → ∞.

Thus fn ∗ gn − f ∗ g → 0 uniformly on T as n → ∞. In other words, fn ∗ gn → f ∗ g


uniformly on T as n → ∞.

Theorem 1.4.6. Let f, g ∈ L1 (T). Then (f ∗ g)b(n) = fb(n) gb(n), ∀ n ∈ Z.

Proof.


1
(f ∗ g) (n) =
b
(f ∗ g) (x) e−inx dx

−π
 
Zπ Zπ
1 1
= f (x − y) g(y) dy  e−inx dx
2π 2π
−π −π
 
Zπ Zπ
1 1
= f (x − y) e−inx dx g(y) dy
2π 2π
−π −π
 
Zπ Zπ
1  1 e−iny
= f (x − y) e−in(x−y) dx g(y) dy
2π 2π
−π −π

1
= fb(n) g(y) e−iny dy = fb(n)b
g (n).

−π

1.5 An approximate identity


In the previous section, we have shown that L1 (T) is Banach algebra under
convolution. It is therefore natural to ask if L1 (T) admits an identity. In other words,
the question is whether there exists ϕ ∈ L1 (T) such that f ∗ ϕ = f for all f ∈ L1 . It
is easy to see that this is not the case. Indeed, if such a ϕ exists, then by Theorem

25
1.4.6 fb(n)ϕ(n)
b = fb(n) for all n which is possible only if ϕ(n)
b = 1 for all n. But for
ϕ ∈ L1 (T), ϕ(n)
b → 0 as n → ± ∞. (See Corollary 1.7.2.) Hence there is no ϕ ∈ L1 (T)
such that f ∗ ϕ = f for all f .

Thus we look for a sequence {en } such that f ∗en → f as n → ∞ in appropriate


norm. Such a sequence is called an approximate identity.

Towards this end, first we shall prove the following theorem.


Theorem 1.5.1. Let {kn } be a sequence of functions defined on T satisfying


1
1. kn (x) dx = 1 ∀ n.

−π


2. There exists 0 ≤ M < ∞ such that |kn (x)| dx ≤ M, ∀ n ≥ 1.
−π

Z
3. For every δ > 0, |kn (x)| dx → 0 as n → ∞.
δ≤|x|≤π

If f ∈ C(T), then kn ∗ f converges to f uniformly on T .

1 Rπ
Proof. As kn (y) dy = 1,
2π −π


1
f ∗ kn (x) − f (x) = [f (x − y) − f (x)] kn (y) dy,

−π
which gives

1
|f ∗ kn (x) − f (x)| ≤ |f (x − y) − f (x)| |kn (y)| dy. (1.10)

−π

26
Let  > 0 be given. Then by uniform continuity of f there exists δ > 0 such that
|f (s) − f (t)| <  whenever |s − t| < δ. By (1.10),


1
|f ∗ kn (x) − f (x)| ≤ |f (x − y) − f (x)| |kn (y)| dy

−δ
Z
1
+ |f (x − y) − f (x)| |kn (y)| dy

δ≤|y|≤π

=: I1 + I2 .

We have,

1
I1 = |f (x − y) − f (x)| |kn (y)|dy

−δ

 M
≤ |kn (y)| dy ≤ .
2π 2π
−δ
On the other hand
Z
1
I2 = |f (x − y) − f (y)||kn (y)| dy

δ≤|y|≤π

kf k∞
Z
≤ |kn (y)| dy, (1.11)
π
δ≤|y|≤π

which goes to zero as n → ∞. Here kf k∞ = sup |f (x)|. Thus f ∗ kn converges to f


x∈T
uniformly on T.

Remark 1.5.2. Since C(T) is dense in L1 (T) and uniform convergence is stronger
than L1 convergence, it follows from the above theorem that {kn } gives an approxi-
mate identity for L1 (T).

27
Remark 1.5.3. We shall show that the Dirichlet kernel does not satisfy the hypoth-
esis of Theorem 1.5.1. Recall

N
X sin (N + 12 )x
DN (x) = einx = .
n=−N
sin ( x2 )

1 Rπ 1 Rπ inx
We have DN (x) dx = 1. Since e dx = 0 for all n 6= 0. However,
2π −π 2π −π

property (ii) of Theorem 1.5.1 fails. We shall show that |DN (x)| dx ≥ c ln N, ∀ N.
−π

Consider
Zπ Zπ 1
sin (N + 2 )x

|DN (x)|dx = dx
sin ( x2 )

−π −π

|sin (N + 12 )x|
≥2 dx,
|x|
−π

sin x2

since x ≤ 1. Setting y = (N + 21 ) x, we get,
2

(N + 12 )π
Zπ Z
sin y
|DN (x)|dx ≥ 2 y dy

−π −(N + 12 )π

(N + 12 )π
Z
sin y
=4 y dy

0
 
(N + 12 )π
 ZN π sin y
 Z 
sin y 
=4 y dy +

y dy  .


 
0 Nπ

28
Since the second term on the right hand side is non-negative, we get,

Zπ ZN π
sin y
|DN (x)| dx ≥ 4 y dy

−π 0
(k+1)π
N −1 Z
X sin y
=4 y dy.

k=0 kπ

1 1
For k π ≤ y ≤ (k + 1) π, we have ≥ . Hence
|y| (k + 1) π

Zπ (k+1)π
N −1 Z
X 1
|DN (x)|dx ≥ 4 | sin y| dy
k=0
(k + 1)π
−π kπ
N −1 Zπ
4X 1
= | sin(y)|dy,
π k=0 (k + 1)
0
N −1
X 1
≥ c0
k=0
(k + 1)
N
X 1
≥ c0 ≥ c00 log N, as N → ∞.
l=1
l

1.5.1 Fejér kernel

As an example of approximate identity, we introduce the kernel σn (x). This


is defined, using the Dirichlet kernel Dk , by

N −1
1 X
σN (x) = Dk (x).
N k=0

29
Note that
N −1
1 X
f ∗ σN (x) = f ∗ Dk (x)
N k=0
N −1
1 X
= Sn (f (x)), (1.12)
N k=0

and these are called the Fejér means. Fejér proved the following result. In fact, he
has proved for f ∈ Lp (T), 1 ≤ p < ∞. However, we shall give the proof for p = 1
now and give the remaining separately in the solved problem.

Figure 1.4: Fejér kernel.

30
Theorem 1.5.4. Let f ∈ L1 (T). Then f ∗ σN → f in L1 (T).

Proof. The theorem will follow if we can show that {σN } is an approximate identity.
In order to show this result, first we shall calculate N σN (x).

Consider
N −1 N −1
w−n − wn+1
X X  
N σN (x) = Dn (x) =
n=0 n=0
1−w
(N −1 N −1
)
1 X X
= w−n − wn+1
1 − w n=0 n=0
1 − w−N 1 − wN
 
1
= −w
1−w 1 − w1 1−w
−N
(1 − wN )
 
1 (1 − w )
= w −w
1−w w−1 1−w
w  −N
w − 1 − 1 + wN

= 2
(1 − w)
w
= (wN + w−N − 2)
(1 − w)2
w 2
= 2
(wN/2 − w−N/2 ) .
(1 − w)
Thus
 !2
1 sin N x2
σN (x) = .
N sin ( x2 )

Note that σN (x) ≥ 0 for every x. Further

Zπ Zπ N −1
1 1 1 X
σN (x) dx = Dn (x) dx
2π 2π N n=0
−π −π
N −1 Zπ
1 X
= Dn (x) dx = 1.
2πN n=0
−π

31
Fix δ > 0. Let δ ≤ |x| ≤ π. Since sin2 x
≥ sin2 δ
 
2 2
= Cδ > 0 we get σN (x) ≤
sin2 (N x2 )
.
N Cδ

Consider
sin2 N x2
Z Z 
1
|σN (x)|dx =  dx
N sin2 x2
δ≤|x|≤π δ≤|x|≤π
Z
1 1
sin2 N x2 dx.


N Cδ
δ≤|x|≤π
1 2π
≤ → 0 as N → ∞.
N Cδ

Thus σN satisfies the required properties of Theorem 1.5.1. Consequently, we con-


clude that whenever f ∈ C(T), then f ∗ σN (x) converges uniformly to f on T .

1.5.2 Poisson kernel


Recall that for 0 ≤ r < 1,


1 − r2 X
Pr (θ) = 2
= r|n| einθ .
1 − 2r cos θ + r n=−∞

We shall show that {Pr } gives an approximate identity. First, notice that Pr (θ) ≥ 0.
(i)

Zπ Zπ  X

1 1 
Pr (θ) dθ = r|n| einθ dθ
2π 2π n=−∞
−π −π
∞ Zπ
1 X
= r|n| einθ dθ
2π n=−∞
−π

32
∞ Zπ
1 X |n|
= r einθ dθ
2π n=−∞
−π

= 1.

Figure 1.5: Poisson kernel.

33
1
(ii) Fix δ > 0 and consider δ ≤ |θ| ≤ π and 2
≤ r < 1. Then

1 − 2r cos θ + r2 = (1 − r)2 + 2r(1 − cos θ)

= (1 − r)2 + 4r sin2 2θ


≥ 2 sin2 2δ = 2cδ .


Hence,
1 − r2
Pr (θ) = ,
1 − 2r cos θ + r2
1 − r2
≤ ,
2cδ
and consequently
1 − r2
Z
Pr (θ)dθ ≤ · 2π → 0 as r → 1.
2cδ
δ≤|θ|≤π

Thus Pr satisfies the required properties for an approximate identity. Consequently,


we conclude that whenever f ∈ C(T), f ∗ Pr (θ) converges uniformly to f on T.

1.6 Summability of Fourier series



ck . The nth partial sum of
P
Suppose we have a series of complex numbers
k=0
n
P ∞
P
the series is defined as Sn := ck . If Sn → s as n → ∞, then we say that ck
k=0 k=0
1
converges to s. Let σN = N
+ S1 + · · · + SN −1 ). The arithmetic mean σN is
(S0

called the N th
P
partial Cesaro sum of the series ck . If σN converges to a limit σ as
k=0

P
N → ∞, then we say that ck is Cesaro summable to σ.
k=0

P
A series of complex numbers ck is said to be Abel summable to s if for
k=0
∞ ∞
ck rk converges and lim ck rk = s.
P P
every 0 ≤ r < 1, the series
k=0 r→1 k=0

34
Remark 1.6.1. Convergence ⇒ Cesaro summability ⇒ Abel summability. (See more
details in Solved problems.)
Consider

Zπ ∞
!
1 X
f ∗ Pr (θ) = f (x) r|n| ein(θ−x) dx
2π n=−∞
−π
Zπ ∞
1 X
= f (x)r|n| ein(θ−x) dx

−π n=−∞
∞ Zπ
1 X  
= f (x)e−inx r|n| einθ dx
2π n=−∞
−π

1 X
f ∗ Pr (θ) = fb(n)r|n| einθ . (1.13)
2π n=−∞

Thus we see that f ∗ Pr are nothing but the Abel means of f . Similarly, f ∗ σN are
nothing but the Cesaro means. Hence we have the following theorems.
Theorem 1.6.2. If f ∈ C(T), then the Fourier series of f is uniformly Cesaro
summable to f .
Theorem 1.6.3. If f ∈ C(T), then the Fourier series of f is uniformly Abel summable
to f .

1.7 Fourier series of f ∈ L2(T)

Recall that L2 (T) is a Hilbert space with the inner product


1
hf, gi = f (t)g(t)dt for f, g ∈ L2 (T).

−π

35
Theorem 1.7.1. (Bessel’s inequality) If f ∈ L2 (T), then

∞ Zπ
X 1
|fb(n)|2 ≤ |f (t)|2 dt.
n=−∞

−π

Proof. Consider

hf − SN f, f − SN f i = hf, f i − hf, SN f i − hSN f, f i + hSN f, SN f i (1.14)

N
fb(n) einx . If we denote en (x) = einx , x ∈ R, then we can
P
Recall that SN f (x) =
n=−N
N
P
write SN f = fb(n)en and also fb(n) = hf, en i. Thus
n=−N

* N
+
X
hf, SN f i = f, fb(n)en
n=−N
N
X
= fb(n)hf, en i
n=−N
N
X
= fb(n)fb(n)
n=−N
N
X
= |fb(n)|2 .
n=−N

Further * +
N
X N
X
hSN f, SN f i = fb(n)en , fb(m)em
n=−N m=−N
N
X N
X
= fb(n)fb(m)hen , em i.
n=−N m=−N

36
As hen , em i = δn,m , we can conclude that
N
X
hSN f, SN f i = |fb(n)|2 .
n=−N
Hence, (1.14) turns out to be
N
X
hf − SN f, f − SN f i = hf, f i − |fb(n)|2 ≥ 0,
n=−N
from which it follows that
XN
|fb(n)|2 ≤ hf, f i
n=−N

1
= |f (t)|2 dt.

−π
Letting N → ∞, we get
∞ Zπ
X
2 1
|fb(n)| ≤ |f (t)|2 dt.
n=−∞

−π

This leads us to the following important result.

Corollary 1.7.2. (Riemann-Lebesgue lemma) If f ∈ L1 (T), then fb(n) → 0 as


Rπ Rπ
n → ± ∞. Equivalently, if f ∈ L1 (T), then f (x) cos nx dx → 0 and f (x)
−π −π
sin nx dx → 0 as n → ± ∞.

Proof. The result follows immediately from the theorem for f ∈ L2 (T). Let f ∈
L1 (T). Since L2 (T) = L1 (T) ∩ L2 (T) is dense in L1 (T), there exists a sequence of
functions {fn } in L2 (T) such that fn → f in L1 -norm. Now lim fbn (m) = 0 for each
m→∞
n. Consider

Z2π
1
|fbn (m) − fb(m)| ≤ |fn (x) − f (x)| dx

0

37
= kfn − f k1 → 0,

as n → ∞. This shows that fbn converges to fb uniformly on Z.

Hence

lim fb(m) = lim lim fbn (m)


m→∞ m→∞ n→∞

= lim lim fbn (m) = 0.


n→∞ m→∞

Theorem 1.7.3. (Best approximation theorem) Let f ∈ L2 (T). Let SN (f )


N
denote the N th partial sum of the Fourier series, and tN (x) = cn einx be any
P
n=−N
trigonometric polynomial. Then kf − SN f k2 ≤ kf − tN k2 and the equality holds iff
fb(n) = cn .

Proof. Consider

kf − tN k22 = hf − tN , f − tN i = hf, f i − hf, tN i − htN , f i + htN , tN i. (1.15)

But * +
N
X
hf, tN i = f, cn e n
n=−N
N
X
= cn hf, en i
n=−N
N
X
= cn fb(n).
n=−N
Also * +
N
X m
X
htN , tN i = cn e n , cm em
n=−N n=−m

38
N
X m
X
= cn cm hen , em i
n=−N m=−m
N
X m
X
= cn cm δn,m
n=−N n=−m
N
X
= |cn |2 .
n=−N
Hence (1.15) becomes
N
X N
X N
X
kf − tN k22 = hf, f i − cn fb(n) − cn fb(n) + |cn |2
n=−N n=−N n=−N
N
X N
X
= hf, f i − |fb(n)|2 + |fb(n) − cn |2 .
n=−N n=−N
Thus
N
X
kf − tN k22 = kf − SN f k22 + |fb(n) − cn |2 ,
n=−N

proving our assertion. Recall that a complex valued function f on T is said to be


Lipschitz continuous on T if |f (s) − f (t)| < |s − t| for all s, t ∈ T. In the follow-
ing theorem, we show that Lipschitz continuity leads to the uniform convergence of
Fourier series.

Theorem 1.7.4. If f is a Lipschitz continuous function on T, then SN (f ) converges


to f uniformly on T.
 
1 1 it
iN t 2
−it
−iN t 2 t
Proof. Consider sin (N + 2
)t = e e −e e . As f is Lipschitz contin-
2i
f (x−t)−f (x)
uous, the function t is bounded. Now, applying Riemann-Lebesgue lemma
sin
2
i −i
f (x−t)−f (x)
to t e 2 t and f (x−t)−f t
(x)
e 2 ,
t
we can conclude the result.
sin sin
2 2

39
Theorem 1.7.5. Let f ∈ C(T). Suppose f is piecewise continuously differentiable
function on T. Then the Fourier series of f converges uniformly to f on T.

Proof. It is enough to prove that fb ∈ l1 (Z). Given that f is a piecewise continuously


differentiable function on T. We know that T can be identified with [−π, π]. Then
there exists a partition x1 , x2 , . . . , xm−1 ∈ [−π, π], such that −π = x0 ≤ x1 ≤ · · · ≤
xm−1 ≤ xm = π and f |[xj−1 ,xj ] is continuously differentiable for each j = 1, 2, · · · , n.
Let gj denote the continuous derivative of f on [xj−1 , xj ]. Choose a periodic
function g on [−π, π] such that g = gj on [xj−1 , xj ]. Then, clearly, g ∈ L2 (T). Using

g (n)|2 ≤ kgk22 .
P
Bessel’s inequality, we have |b
n=−∞

Consider, for n 6= 0
Zxj −inx xj
Zxj
e 1 0
f (x)e−inx dx = f (x) + e−inx f (x)dx.
−in xj−1 in

xj−1 xj−1
Thus

1
fb(n) = f (x)e−inx dx

−π
x
m Zj
1 X
= f (x)e−inx dx.
2π j=1
xj−1
Now

gb(n)
|fb(n)| =
in
 
1 1 2
≤ + |b
g (n)| .
2 n2
Thus, we get,

X X
|fb(n)| = |fb(0)| + |fb(n)|
n=−∞ n6=0

40
1 X 1 2

≤ |fb(0)| + + |b
g (n)|
2 n6=0 n2

< ∞,

from which it follows that fb ∈ l1 (Z).

Theorem 1.7.6. For any f ∈ L2 (T), SN (f ) converges to f in L2 (T), i.e kSN f − f k2


→ 0 as N → ∞.

Proof. Let  > 0 be given. We know that C(T) is dense in L2 (T). Therefore, given
f ∈ L2 (T), there exists h ∈ C(T) such that

kf − hk2 < . (1.16)

Writing f − SN f = (f − h) + (h − SN h) + SN (h − f ) and noting that kSN (h − f )k2 ≤


kf − hk2 , it is enough to show that kh − SN hk <  for all N ≥ N0 for some N0 .
As h ∈ C(T), there exists a trigonometric polynomial P of degree N0 such that
kh − P k2 < . For N ≥ N0 , SN P = P and hence

kh − SN hk2 ≤ kh − P k2 + kSN (P − h)k2 ≤ 2.

This proves our assertion.

Theorem 1.7.7. (Parseval’s Theorem) For any f, g ∈ L2 (T),

Zπ ∞
1 X
f (x)g(x)dx = g (n).
fb(n)b
2π n=−∞
−π

41
In particular,
Zπ ∞
1 X
|f (x)|2 dx = |fb(n)|2 .
2π n=−∞
−π

Proof. We have shown in Theorem 1.7.6 that SN f → f in L2 . Hence, hSN f, SN gi →


hf, gi which simply means

∞ Zπ
X 1
g (n) =
fb(n)b f (x)g(x)dx.
n=−∞

−π

Corollary 1.7.8. The functions {einx : n ∈ Z} forms an orthonormal basis for L2 (T).
(See Appendix A.1.3.)
Corollary 1.7.9. Let f ∈ L2 (T). Then the Fourier series of f can be integrated term
by term in (a, b) ⊆ [−π, π].

Proof. Let g(t) = χ[a,b] (t). Then g ∈ L2 (T) and hence by Parseval’s theorem,

Zπ ∞
1 X
f (t)g(t)dt = fb(n)b
g (n).
2π n=−∞
−π
This means that  b 
Zb ∞ Z
1 1 X
f (t)dt = fb(n)  eint dt ,
2π 2π n=−∞
a a

which is our assertion.

42
1.8 Existence of a continuous function whose Fourier

series diverges
We shall show in this section that, there exists a continuous function whose
Fourier series diverges at 0. In order to prove the theorem, we make use of the uniform
boundedness principle from functional analysis. (See Appendix A.2.5.)

Let X = C[−π, π], Y = R. Define AN f := SN f (0). Then |AN f | = |SN f (0)| =


|DN ∗ f (0)| ≤ kf k∞ kDN k1 . In other words, the operator norm of AN is given by
kAN k ≤ kDN k1 . We shall show that kAN k = kDN k1 . Take g(x) = sgnDN (x). Then
kgk∞ = 1. Although g is not continuous, it can be approximated by a sequence of
continuous functions gn .

Choose gn such that kgn k∞ = 1. Then AN gn → kDN k1 1, by using Lebesgue


dominated convergence theorem. Thus, kAN k = kDN k1 . But kDN k1 ≥ c log N (see
Remark 1.5.3). Hence, kAN k → ∞ as N → ∞. Hence, it follows from uniform
boundedness principle that there exists a continuous function f in [−π, π] such that
|AN f | = |SN f (0)| → ∞ as N → ∞, which means the Fourier series of f diverges
at 0.

1.9 Brief history of Fourier series


The appearance of Fourier series can be traced back to the works of d ’Alem-
bet(1747), Euler (1748) and D.Bernoulli (1753) in the study of vibrating strings, but
the theory of Fourier series truly began with the profound work of Fourier on heat
conduction at the beginning of the nineteenth century. He studied the problem of
heat flow in a thin wire fixed at two end points with length π and zero temperature

43
at the ends. He showed that the initial temperature can be expressed as the sum of
an infinite series of sines and cosines, now popularly known as Fourier series. Even
though Fourier did not give a convincing proof of convergence of such infinite series,
he offered the conjecture that convergence holds for an arbitrary function. Subsequent
work by Dirichlet, Riemann, Lebesgue, and others, throughout the next two hundred
years was needed to show that any arbitrary periodic function can be expressed in
terms of trigonometric series.

1.10 Solved problems


 
k 1
1.10.1. Suppose that f ∈ C (T). Show that fb(n) = O as n → ∞.
|nk |
Solution.


1
fb(n) = f (x)e−inx dx

−π

e−inx
 
1
= f (x)d .
2π −in
−π

Using Bernoulli’s formula for successive integration by parts, we get

(−1)k k b
fb(n) = (f ) (n).
(−in)k
Then
1
|fb(n)| ≤ |(f k )b(n)|
|n|k
1
≤ k
kf (k) k1 .
|n|

44

1
1.10.2. If f ∈ L1 (T), then show that fb(n) = (f (x) − f (x + π/n)) e−inx dx. In

−π  
1
addition, if f satisfies the Holder’s condition of order α, then fb(n) = O .
|n|α
Solution. We have,


1
fb(n) = f (x)e−inx dx.

−π

Since e = −1, we can write

−1
fb(n) = f (x)e−inx+πi dx

−π

−1 π
= f (x)e−in(x− n ) dx.

−π
π−π/n
−1
Z
= f (y + πn )e−iny dy.

−π−π/n
Hence

−1
fb(n) = f (x + πn )e−inx dx.

−π
Thus
Zπ Zπ
1 −inx 1
2fb(n) = f (x)e − f (x + πn )e−inx dx,
2π 2π
−π −π

from which it follows that,


1
f (x) − f (x + πn ) e−inx dx.

fb(n) =

−π

45
Then

1
f (x) − f (x + πn ) dx

|fb(n)| ≤

−π

1 π α
≤ c · 2π

4π n
−π
k
= ,
|n|α

where k is a constant.
ZR
sin x π
1.10.3. Using Riemann-Lebesgue lemma, prove that lim dx = .
R→∞ x 2
0
ZR
sin x
Solution. Let I = lim dx. Let R = (N + 21 )π, then
R→∞ x
0

1
(N + )π
Z2
sin x
I = lim dx.
N →∞ x
0

Applying change of variable x = (N + 12 )y, we get


sin (N + 12 )y
I = lim dy.
N →∞ y
0

From the property of Dirichlet kernel, we get


sin (N + 12 )y
dy = π
sin ( y2 )
0

46
Hence,
Zπ 
sin (N + 12 )y sin (N + 12 )y

I − π/2 = lim dy − dy
N →∞ y 2 sin ( y2 )
0
Zπ  
1 1
= lim − sin (N + 12 )y dy.
N →∞ y 2 sin ( y2 )
0

1 1
Since, − y
 ∈ L2 (T), from Riemann-Lebesgue lemma, it follows that I = π/2.
y 2 sin 2
n
n 2 2n
P  
1.10.4. Prove the Binomial identity k
= n
.
k=0
Solution. Consider

n  
X n
ek (x) = (1 + e1 (x))n .
k=0
k

By Parsaval’s inequality

n  2 Zπ Zπ
X n 1 2 1
= |f (x)| dx = |(1 + e1 (x))n |dx
k=0
k 2π 2π
−π −π

1
= (1 + e1 (x))n (1 + e1 (x))n dx

−π

1
= (1 + e1 (x))n (1 + e1 (x))n dx

−π

1
= (1 + e1 (x))(1 + e1 (x))n dx

−π

1
= (2 + 2 cos x)n dx

−π

47

1
= · 2n (1 + cos x)n dx

−π

22n
· cos2n x

= 2
dx

−π
 
2n
= .
k

1.10.5. By considering the Fourier series of f (x) = |x|, x ∈ [−π, π], find the sum of
P∞ 1 P∞ 1 P∞ 1
the series 4
and 4
. What is the sum of 2
.
n=0 (2n + 1) n=0 n n=1 n
Solution. We have


1
fb(n) = |x|e−inx dx

−π
 π 
Z Zπ
1  
= |x| cos nxdx − i |x| sin nxdx
2π  
−π −π
 π 
Z
1  
= 2 x cos nxdx
2π  
0
 
π Zπ
1 x sin nx
 sin nx 
= − dx if n 6= 0
π n 0 n 
0

1 1
=− · sin nxdx
π n
0
1 π
= 2 cos nx 0

1
= 2 {(−1)n − 1}.

48
For n 6= 0, we have 
 0 if n is even
fb(n) =
 −2 if n is odd
n2 π


1
fb(0) = |x|dx

−π

1
= xdx
π
0

1 x2

=
π 2 0
π
= .
2

Thus the Fourier Series of f (x) = |x| on [−π, π] is given by


X
|x| = fb(0) + fb(n)einx
n=−∞
n6=0
π X −2
= + einx
2 n=±1,±3.. n2 π

π X 1
= + 2(−2) 2
ei(2n−1)x , x ∈ [−π, π]
2 n=1
π(2n − 1)

π 4X 1
|x| = − 2
ei(2n−1)x , x ∈ [−π, π]
2 π n=1 (2n − 1)

On putting x = 0, we get


π 4X 1
0= −
2 π n=1 (2n − 1)2

49
Hence, it follows that

X 1 π2
2
= .
n=1
(2n − 1) 8

P∞ 1 P∞ 1
Now we shall calculate the sum of the series 2
. Let s = 2
. Then
n=1 n n=1 n

∞ ∞
X 1 X 1
s= 2
+
n=2,4
n n=1,3
n2
X 1 X 1
= 2
+
m∈N
(2m) m∈N
(2m − 1)2

1X 1 π2
= +
4 m=1 m2 8
Thus
1 π2
s= s+ .
4 8
This leads to

X 1 π2
=
n=1
n2 6

Using Parseval’s inequality, we have

Zπ ∞
1 2 π2 X
x dx = + |fb(n)|2 .
2π 4 n=−∞
−π n6=0
Hence,

1 π2 X 4
x2 dx = + ,
π 4 n=±1,±3,...
n π2
4
0
which means that
1 π3 π2 8 X 1
· = + 2 .
π 3 4 π n=1,3,... n4

50
Thus it follows that
π2 π2 8 X 1
− = 2 ,
3 4 π n=0,1,... (2n + 1)4

from which we arrive at

π2 8 X 1
= 2 .
12 π n=0,1,... (2n + 1)4

This implies that


X 1 π4
= .
n=0
(2n + 1)4 96

P∞ 1
Next, we shall calculate the sum 4
.
n=1 n

Let

0
X 1
s = .
n=1
n4
Then
X 1 X 1
s0 = +
n even
n4 n odd n4
∞ ∞
X 1 X 1
= 4
+
n=1
(2n) n=0
(2n + 1)4
∞ ∞
1 X 1 X 1
= 4
+ .
16 n=1 n n=0
(2n + 1)4
Hence
1 0 π4
s0 − s = .
16 96

51
This leads to

X 1 π4
4
= .
n=1
n 90

1.10.6. Compute the Fourier coefficients of the function f given by


 |x| |x| < 1
f (x) =
 0 otherwise.

Solution. We have


1
fb(n) = f (x) e−inx dx

−π
Z1
1
= |x| e−inx dx

−1
 0 
Z Z1
1  
= −x e−inx dx + xe−inx dx
2π  
−1 0
 
−inx 0
Z0 −inx −inx 1
Z1 −inx 
1  xe − e xe + e
=− dx − dx
2π  −in −1 −in −in 0 −in 
−1 0
(  −inx 0  −inx 1 )
in −in
1 e 1 e e −1 e
=− + − +
2π (−in) in −in −1 (−in) in −in 0
ein e−in
 
1 1 −in in
 1 1
=− e −e + 2 − 2 − 2 + 2
2π in n n n n
 
1 1 −in in
 2 1 in −in

=− e −e + 2 − 2 e +e
2π in n n
 
1 2 2 sin n 2 cos n
=− 2
− −
2π n n n2

52
 
1 1 cos n
=− − sin n − .
nπ n n

1.10.7. Compute the Fourier coefficients of 1-periodic function f (t) = 1, 0 ≤ t ≤ 1.


Solution. We have

Z1
fb(n) = e−2πint dt
0
1
e−2πint
=
2πin 0
i  −2πin 
= e −1
2πn
i
= [(−1)n − 1] , n 6= 0
2πn
But
Z1
fb(0) = dt = 1.
0
Hence
i
 [(−1)n − 1] n 6= 0
f (n) =
b 2πn
1 n = 0.

Rt
1.10.8. Let f ∈ L1 (T) such that fb(0) = 0. Define F (t) = f (τ ) dτ . Show that F is
0
fb(n)
continuous, 2π-periodic and Fb(n) = , n 6= 0.
in
Solution. Let  > 0 be given. Then for t > s,
t
Z Zs

|F (t) − F (s)| = f (τ ) dτ − f (τ ) dτ


0 0

53
t
Z

= f (τ ) dτ


s
Zt
≤ |f (τ )| dτ.
s

As f ∈ L1 (T), there exists a δ > 0 such that

Zt
|f (τ ) |dτ ≤  whenever |t − s| < δ.
s
Thus, F is continuous. Consider
t+2π
Z
F (t + 2π) = f (τ ) dτ
0
Z2π 2π+t
Z
= f (τ ) dτ + f (τ ) dτ
0 2π
Zt
= 2π fb(0) + f (τ ) dτ
0

= F (t).

Thus F is 2π-periodic. By Radon-Nikodym theorem, F is differentiable with F 0 (t) =


f (t) − f (0). Thus

Z2π
Fb(n) = F (t)e−int dt
0
Z2π
e−int
 
= F (t) d
−in
0

54
2π Z2π −int
e−int e
= F (t) − F 0 (t)dt
−in 0 −in
0
Z2π Z2π
1
= f (t)e−int dt − f (0) e−int dt
in
0 0
1 b
= f (n), n 6= 0.
in

1.10.9. Calculate χ[−π,0] ∗ χ[0,π] in L1 (T).


Solution. Note that

Z0
1
χ[−π,0] ∗ χ[0,π] (t) = χ[0,π] (t − s) ds.

−π
Now, for this integral to be non-zero, s should belong to [−π, 0] and [−π + t, t].

If −π ≤ t ≤ 0
Zt
1 1
χ[−π,0] ∗ χ[0,π] (t) = dt = [t + π].
2π 2π
−π
If 0 < t ≤ π, then
Z0
1 1
χ[−π,0] ∗ χ[0,π] (t) = dt = [−t + π].
2π 2π
−π+t
Hence
1
χ[−π,0] ∗ χ[0,π] (t) = (π − |t|)χ[−π,π] (t).

1.10.10. Suppose θ : R → T is a continuous group homomorphism then show that


θ(t) = eitx for some real number x. Use this to derive the form of the continuous
group homomorphism θ : T → T.

55
Solution. If θ is the trivial homomorphism, then we are done. If not, there exists a
a > 0 such that

Za
c= θ(t)dt 6= 0.
0
Thus
Za Za+x
cθ(x) = θ(x + t)dt = θ(t)dt.
0 0
So, θ is differentiable and
1 1
θ0 (x) = [θ(a + x) − θ(x)] = [θ(a) − 1]θ(x) = c0 ,
c c

1
where c0 = [θ(a) − 1]. Thus, θ satisfies the ordinary differential equation θ0 = c0 θ.
c
0
Hence, θ(t) = ec t and since |θ(t)| = 1 ∀ t, c0 = 2πix for some x ∈ R.

Since T ' R/Z, any continuous group homomorphism θ : T → T gives rise to


a continuous group homomorphism θ0 : R → T given by θ0 = θ · π, where π is the
canonical quotient map, π : R → R/Z. Also, ker θ0 contains Z. Thus by the form of
the continuous homomorphism from R → T, θ(t) = eint for some integer n.
1.10.11. Let G be a closed proper subgroup of T. Show that G is finite.
Solution. Since G is closed, T/G is an abelian group. Let θ : T/G → T be a
continuous group homomorphism. Then the map θ0 : T → T defined as θ0 = θ · π,
π : T → T/G the canonical quotient map, is a continuous group homomorphism of
T to T whose kernel contains G. But by the previous problem, θ0 (t) = eint for some
integer n. Therefore ker(θ0 ) is finite and hence G is finite.
1.10.12. If {kn } is a sequence of good kernels, show that for 1 ≤ p < ∞, kkn ∗ f −
f kp → 0 as n → ∞.

56
Solution. Suppose that the conclusion is true for any g ∈ C(T) ⊆ Lp (T). Let
1

f ∈ Lp (T). Since {kn } forms a good kernels, 2π |kn (t)|dt = c < ∞. As C(T) is
0

dense in Lp (T), given  > 0, there exists g ∈ C(T) such that kf − gkp < 3c
. By
assumption, there exists an n0 ∈ N such that for all n ≥ n0 kkn ∗ g − gkp < 3 . Thus,

kkn ∗ f − f kp ≤ kkn ∗ f − kn ∗ gkp + kkn ∗ g − gkp + kg − f kp

= kkn ∗ (f − g)kp + kkn ∗ g − gkp + kg − f kp

≤ ck(f − g)kp + kkn ∗ g − gkp + kg − f kp


  
< + +
3 3 3
Thus it is enough to prove for functions in C(T). Let f ∈ C(T). Then
 2π  p1
Z
1 
kkn ∗ f − f kp = |kn ∗ f (x) − f (x)|p dx

0
 2π p  p1
Z Z2π
1  1

= 2π kn (t)f (x − t)dt − f (x) dx


0 0
 2π p  p1
Z Z2π Z2π
1 1 1
=  k n (t)f (x − t)dt − kn (t)f (x)dt dx
2π 2π


0 0 0
 2π  p1
Z Z2π
1  1
≤ |kn (t) [f (x − t) − f (x)]|p dx dt
2π 2π
0 0
Z2π
1
= |kn (t)|kft − f kp dt

0
2π−δ
Z
1
= |kn (t)|kft − f kp dt

−δ

57
Zδ 2π−δ
Z
1
= |kn (t)|kft − f kp dt + |kn (t)|kft − f kp dt,

−δ δ

using Minkowski’s inequality. Now, given  > 0 there exists δ > 0 such that kft −
f kp < , whenever |t| < δ. Now, for this δ there exists an n0 ∈ N such that for all
2π−δ
1
R
n ≥ n0 , 2π |kn (t)|dt < . Thus
δ

 
Zδ 2π−δ
Z
1 
kkn ∗ f − f kp ≤  |kn (t)|dt + 2kf k∞ |kn (t)|dt

−δ δ
Z2π
< |kn (t)|dt + 2kf k∞ 
0

= c.

 
1
1.10.13. If f is of bounded variation on T, then show that fb(n) = O |n|
.
Solution. We have

Z
1 −int

|fb(n)| = e f (t)dt


0

Z  −int 
1 e
= f (t)d
2π −in
0
−int 2π
Z2π −int
1 e − e
= f (t) df (t)
2π −in 0 −in

0

Z
1 −int
V ar(f )

= e df (t) .
2π|n|
|n|
0

58
1.10.14. If the conjugate of Dirichlet kernel is given by

X
D̃N (x) = sgn(n) einx ,
|n|≤N

where 


 1 n>0

sgn(n) = 0 n=0



−1 n < 0


show that |D̃N (x)|dx ≤ c ln N .
−π
Solution. We know that

cos ( x2 ) − cos (N + 12 )x
D̃N (x) = i .
sin ( x2 )
Then
Zπ Zπ
cos ( x2 ) − cos (N + 12 )x

D̃N (x) dx = dx

sin ( x2 )
−π −π

cos ( x2 ) − cos N x cos( x2 ) + sin N x sin ( x2 )

= dx
sin ( x )
2

−π

cos ( x2 )[1 − cos (N x)]


= + sin N x dx
sin ( x )
2

−π


1 − cos (N x)
≤ dx + 2π
sin ( x2 )
−π
Zπ 2 N x 
sin 2
=2 dx + 2π
sin ( x2 )
−π

59
π
Z2 2
sin x
=4 sin x dx + 2π

π

2
π
Z2 2
sin (N x)
=8 dx + 2π
sin x
0
π
Z2 2
sin N x
≤8 dx + 2π

2x
0
π

π
Z2 2
sin N x
= 4π dx + 2π
x
0

Z2
sin2 y
= 4π dy + 2π
|y|
0

Z2
|sin y|
= 4π dy + 2π
|y|
0 π nπ

Z2 N Z 2
 sin x X 1 
= 4π 
 x
dx + dx + 2π 
n=2
x 
0 π
(n−1)
2
N
X nπ
2
=c+ ln x|(n−1) π + 2π
n=2 2
N
X n π2
=c+ ln π + 2π
n=2
(n − 1) 2
N  
X n
=c+ ln
n=2
n−1
 
N n
= c + ln Π
n=2 n − 1

60
= c + ln N

≤ c ln N.

1.10.15. Let f ∈ C(π). For each pair of integers m, n, where 0 ≤ m < n, define
n
1
P
σm,n (f, x) = n−m Si (f, x), where Si (f, x) stands for i-th partial sum of f at x.
i=m+1
(a) Show that σkn,(k+1)n (f ; x) → f (x) uniformly as n → ∞, where k, n are integers.
(b) If k, m, n are positive integers with kn ≤ m ≤ (k +1)n show that |σkn,(k+1)n (f ; x)−
2A A
Sm (f ; x)| ≤ k
, where f is such that |fb(j)| ≤ |j|
for j 6= 0.
(c) Using (a) and (b) show that Sn (f ; x) → f (x) uniformly on T whenever fb(n) =
O( n1 ).
1 n
P
Solution. For integers m and n, we have σm,n (f ; x) = Si (f ; x). Rewrit-
n − m i=m+1
ing this in terms of the kernel, we obtain

1
σm,n (f ; x) = [(n + 1)σn+1 (f ; x) − (m + 1)σm+1 (f ; x)] (1.17)
n−m

Using the Fourier expansion of the kernel, one obtains the following Fourier expansion
for σm,n (f ) as follows:

X n + 1 − |j|
σm,n (f ; x) = Sm (f ; x) + fb(j)eijx . (1.18)
n−m
m<|j|≤n

By (1.17), we have

1
σkn,(k+1)n (f ; x) = [(k + 1)n + 1)σ(k+1)n (f ; x) − (kn + 1)σkn+1 (f ; x)]
n
= (k + 1 + n1 )σ(k+1)n (f ; x) − (k + n1 )σkn+1 (f ; x) (1.19)

61
which converges uniformly to (k +1)f (x)−kf (x) = f (x) as n → ∞, by using problem
1.10.12. This proves (a).
To prove (b), we use (1.18). In fact, from (1.18), we get

(k+1)n
X X A 2A
σkn,(k+1)n (f ; x) − Sm (f ; x) ≤ |fb(j)| ≤ 2 ≤ .
j k
|kn|≤|j|≤(k+1)n j=kn+1

A
For (c), given that fb(n) = O( n1 ), that is |fb(n)| ≤ , for n 6= 0. Let  > 0 be given.
|n|
A 
Choose an integer k > 0 such that < . By (a) we can find n0 ≥ k such that for
k 4

all n ≥ n0 , |σkn,(k+1)n (f ; x) − f (x)| < . Let m ≥ kn0 such that for some n ≥ n0 ,
2
2A 
kn ≤ m < (k + 1)n. By (b), we have |σkn,(k+1)n (f ; x) − Sm (f ; x)| < < . Thus
k 2
by the above two inequalities, (c) follows.

∞ 1
(−1)n−1 is Cesaro summable to .
P
1.10.16. Show that
n=1  2
∞  1 if n is odd
(−1)n−1 . Here an =
P
Solution.
n=1  −1 if n is even.

 1 if n is odd
Here sn is the n-th partial sum of an and sn =
 0 if n is even.
Consider

s1 + s2 + · · · + sn
σn = .
n
s1 + s2 + · · · + s2n
σ2n =
2n
n
=
2n

62
1
= ∀ n ∈ N.
2
1
lim σ2n =
n→∞ 2
s1 + s2 + · · · + s2n−1
σ2n−1 =
2n − 1
n−1+1 n
= = .
2n − 1 2n − 1
1
Therefore σ2n−1 → as n → ∞
2
1
lim σn = .
n→∞ 2

∞ 1
(−1)n−1 is Cesaro summable to
P
Hence .
n=1 2


(−1)n (n + 1) is not Cesaro summable but Able summable to
P
1.10.17. Show that
n=0
1
.
4 ∞
(−1)n (n + 1)rn = 1 − 2r + 3r2 − · · · = (1 + r)−2 for
P
Solution. We know that
n=0
0 < r < 1. Then


X 1
lim (−1)n (n + 1)rn = lim(1 + r)−2 = .
r→1
n=0
r→1 4
∞ 1
(−1)n (n + 1) is Abel summable to . Now consider the series
P
Therefore,
n=0 4
X∞ X∞
(−1)n (n + 1) = (−1)n−1 n.
n=0 n=1

−n

2
if n is even
Here sn =
n+1

2
if n is odd.
Clearly,
s1 + s2 + · · · + s2n
σ2n = = 0 for all n,
2n

63
s1 + s2 + · · · + s2n−1
σ2n−1 =
2n − 1
2n − 1 + 1
=
2(2n − 1)
n
= .
2n − 1

1 ∞
(−1)n (n + 1) is not Cesaro summable.
P
Therefore, σ2n−1 → as n → ∞. Hence
2 n=0

P ∞
P
1.10.18. Prove that if ci converges to s, then ci is Cesaro summable to s.
i=1 i=1
n
P
Solution. Let sn = ci . The assumption is that sn → s as n → ∞. Given  > 0
i=1
there exists m ∈ N such that |sn − s| <  for all n > m. Now for all n > m, we can
write


s1 + s2 + · · · + sn
|σn − s| = − s
n

(s1 − s) + · · · + (sm − s) + (sm+1 − s) + · · · + (sn − s)
=
n
m n
P
(si − s)
P
(si − s)
i=1 i=m+1
≤ +
n n
A n−m
≤ + 
n n
A
≤ + ,
n

m
P A
where A = (si − s). Now we can find m1 ∈ N such that <  for all n ≥ m1 .
i=1 n

P
Choose m0 = max{m, m1 }. Then |σn − s| <  for all n ≥ m0 . The series cn is
n=1
Cesaro summable to s.

64

P
1.10.19. Assume that an is Cesaro summable to l and lim nan = 0. Then show
n=1 n→∞

P
that an converges to l.
n=1

P
Solution. Assume that an is Cesaro summable to l and lim nan = 0. Consider,
n=1 n→∞
n
P + · · · + sn
s1 + s2
sn = ai and σn = .
i=1 n
By the assumption, lim σn = l
n→∞

s1 + s2 + · · · + sn
σn =
n
a1 + (a1 + a2 ) + · · · + (a1 + a2 + · · · + an )
=
n
n  
X i−1
= 1− ai .
i=1
n
Now
n   n
X i−1 X
σn − sn = 1− ai − ai
i=1
n i=1
n n
1X 1X
= ai − i ai
n i=1 n i=1
n
sn 1X
= − i ai
n n i=1
  n
1 1X
1+ sn = σ n + iai (1.20)
n n i=1

1P n
Since lim nan = 0, lim i ai = 0. Taking limit as n → ∞ in 1.20, we have
n→∞ n→∞ n i=1
lim 1 + n1 sn = lim σn = l. Thus lim sn = l.

n→∞ n→∞ n→∞

Here we have used the following result. If lim an = m 6= 0, lim an bn = l then


n→∞ n→∞
l
lim bn exists and equals .
n→∞ m

65

P ∞
P
1.10.20. If Cn is Cesaro summable to l, then prove that Cn is Abel summable
n=1 n=1
to l.

P
Solution. Assume that Cn is Cesaro summable to l. Define sn and σn as before
n=1
then lim σn = l. It can be easily seen that C1 = s1 ; Cn = sn − sn−1 for all n ≥ 2 and
n→∞

s1 = σ1 , sn = nσn − (n − 1)σn−1 for all n ≥ 2. Consider 0 < r < 1. Now it can be



nσn rn−1 converges absolutely for all 0 ≤ r < 1. Then
P
easily seen that
n=1


X ∞
X
2 n−1
(1 − r) nσn r = (1 − r) [nσn rn−1 − nσn rn ]
n=1 n=1
" ∞ ∞
#
X X
= (1 − r) nσn rn−1 − nσn rn
" n=1 ∞
n=1

#
X X
= (1 − r) σ1 + nσn rn−1 − (n − 1)σn−1 rn−1
n=2 n=2
" ∞
#
X
= (1 − r) s1 + (nσn − (n − 1)σn−1 ) rn−1
n=2
" ∞
#
X
= (1 − r) s1 + sn rn−1
n=2

X
= (1 − r) sn rn−1
n=1

X ∞
X
n−1
= sn r − sn r n
n=1 n=1

X ∞
X
= s1 + sn rn−1 − sn−1 rn−1
n=2 n=2

X
= s1 + (sn − sn−1 ) rn−1
n=2
X∞ ∞
X
= c1 + Cn rn−1 = Cn rn−1 .
n=2 n=1

66
∞ ∞
nσn rn−1 converges for 0 ≤ r < 1, it follows that Cn rn−1 converges for
P P
Since
n=1 n=1

Cn rn−1 for all 0 ≤ r < 1. Now we shall show that
P
0 ≤ r < 1. Define f (r) =
n=1
lim f (r) = l. Choose  > 0. Then there exists m ∈ N such that |σn − l| <  for all
r→1
n ≥ m. Now choose δ = . Then for all 1 −  < r < 1, we have

X ∞
|f (r) − l| = Cn rn−1 − l


n=1
X∞ X∞
= (1 − r)2 nσn rn−1 − l(1 − r)2 nrn−1


n=1 n=1
X ∞
= (1 − r)2 n(σn − l)rn−1


n=1

X
≤ (1 − r)2 n |σn − l| rn−1
n=1
"m−1 ∞
#
X X
= (1 − r)2 nrn−1 |σn − l| +  nrn−1
n=1 n=m
"m−1 #
X
= (1 − r)2 n |σn − l| + (1 − r)−2
n=1
m−1
X
2
= k(1 − r) + , where k = n|σn − l|
n=1

≤ k2 + .

Now for any 0 > 0, there exists unique  > 0 (in fact other root is negative) such
that k2 +  = 0 . Hence lim f (r) = l.
r→1

1.10.21. Assume that f is periodic and monotone on [−π, π].Then show that fb(n) =
 
1
O |n| .

67
Solution. Assume that f is monotone. Then by Lebesgue-Young theorem f is dif-
ferentiable a.e and f 0 ∈ L1 [−π, π]. Now


1
fb(n) = f (t) e−int dt

−π
 

−int π −int
1  e + f 0 (t) e
= f (t) dt . (1.21)
2π −in −π in
−π

The first term of the R.H.S of (1.21) is zero since f (π) = f (−π). Then fb(n) =
1 Rπ 0
f (t)e−int dt which implies
2πin −π


1
f (n) ≤ |f 0 (t)| dt
b
2π|n|
−π
K
= , since f 0 ∈ L1 [−π, π]
2π|n|
 
1
f (n) = O
b .
|n|

1.11 Exercises

1.11.1. Let f ∈ L1 (T). Show that lim |f (x + δ) − f (x)| dx = 0.
δ→0 −π
1.11.2. Show that Fourier series of the function

 1 x is rational
f (x) =
 0 x is irrational

is identically 0 and does not converge to f for any rational number x.

68

You might also like