1 Pointwise convergence of Fourier series

This is the proof of pointwise convergence of the Fourier series of C 1 functions, f (x) on the
interval −L ≤ x < L. The theorem is that, for such functions,

lim |f˜(x) − SN (x)| → 0 (1)

N →∞

for all x in the interval. SN (x) is the partial sum of the Fourier Series, i.e.

F S f (x) = lim SN (x) (2)

N →∞

The adjusted function f˜(x) = f (x) at all points where f (x) is continuous and at points of
discontinuity, xk , it equals the average of the values of f (x) on either side of the discontinuity
−
f (x+
k ) + f (xk )
f˜(xk ) = . (3)
2
Outside of the interval, f (x) is its periodized counterpart.

Proof:
We begin by writing out the Fourier series WITH the expression for the coefficients
N
a0 X  nπx   nπx 
SN (x) = + an cos + bn sin
2 n=1
L L
Z L N
1 L Xh
Z
1  nπy   nπx   nπy   nπx i
= f (y) dy + f (y) cos cos + f (y) sin sin dy
2L −L L −L n=1 L L L L
Z L N
1 L X
 
nπ (x − y)
Z
1
= f (y) dy + f (y) cos dy
2L −L L −L n=1 L
" N #
1 L

nπ (x − y)
Z
1 X
= f (y) + cos dy.
L −L 2 n=1 L
(4)

This motivates the definition of the function

N
1 X  nπz 
DN (z) = + cos (5)
2 n=1 L

1
which is called the Dirichlet kernel. Let’s look at the related series
N N
1 X 1 1 X inθ
+ cos (nθ) = + e + e−inθ
2 n=1 2 2 n=1
1 X
= N einθ
2 n=−N
2N (6)
e−iN θ X
= einθ
2 n=0
2N
ξ −N X
= ξn where ξ ≡ eiθ .
2 n=0

The last series in (6) is the geometric series

σM = 1 + ξ + ξ 2 + ξ 3 + ... + ξ M (7)

and notice the two relations

σM +1 = σM + ξ M +1 , and σM +1 = 1 + ξ σM . (8)

Eliminating σM +1 we find that the sum of the geometric series is

1 − ξ M +1
σM = . (9)
1−ξ
Substituting this into the cosine series we find
N
1 X ξ −N 1 − ξ 2N +1
+ cos (nθ) =
2 n=1 2 1−ξ
1 1 1
ξ −N ξ N + 2 ξ N + 2 − ξ −N − 2
= 1 1 1
2 ξ2 ξ 2 − ξ− 2
1 1
1 ξ N + 2 − ξ −N − 2 (10)
= 1 1
2 ξ 2 − ξ− 2
1 1
1 eiθ(N + 2 ) − e−iθ(N + 2 )
= θ θ
2 ei 2
− e−i 2
sin N + 12 θ

=
2 sin 2θ
 

which is the form that we will need to analyze the Dirichlet Kernel.
Using the series form of the Dirichlet Kernel, the partial sum of the Fourier series is

2
written as
1 L
Z
SN (x) = f (y)DN (x − y) dy
L −L
1 L−x
Z
= f (z + x)DN (−z) dz
L −L−x
1 L−x
Z
= f (z + x)DN (z) dz by symmetry of Dn (z)
L −L−x
1 L 1 −L 1 L−x
Z Z Z
= f (z + x)DN (z) dz + f (z + x)DN (z) dz + f (z + x)DN (z) dz
L −L L −L−x L L
(11)
For the last two integrals let p(z) = f (z + x)DN (z), which is a periodic function of x with
period 2L. These last two integrals become
1 −L 1 L−x
Z Z
= p(z) dz + p(z) dz
L −L−x L L
1 −L 1 L−x
Z Z
= p(z + 2L) dz + p(z) dz
L −L−x L L (12)
Z L Z L−x
1 1
= p(u) du + p(z) dz
L L−x L L
= 0,
so that the partial sum simplifies to
Z L
1
SN (x) = f (z + x)DN (z) dz. (13)
L −L

A study of SN (x) has been reduced to a study of DN (z). First note the integral
1 L
Z
DN (z) dz = 1 ∀ N (14)
L −L
since the integral of all the cosine terms vanish. Multiplying both sides of this expression by
f (x) yields
1 L
Z
f (x) = f (x) DN (z) dz ∀ N. (15)
L −L
Consider the difference between the adjusted function and the partial sum of the Fourier
Series and substitute the rational function form of the Dirichlet Kernel
1 L h˜
Z i
˜
f (x) − SN (x) = f (x) − f (x + z) DN (z) dz
L −L
1 L h˜
Z i sin  N + 1
πz 
= f (x) − f (x + z)  πz2  L dz (16)
L −L 2 sin 2L
Z Lh i   N πz   
1 ˜
 πz  N πz
= f (x) − f (x + z) sin cot + cos dz
2L −L L 2L L

3
 Finally, we need the Riemann-Lebesgue Lemma. Let f (x) be a piecewise C 1 function on
the interval. Then
1 L
Z  Z L 
i NLπz 1 iN π −iN π 0 i NLπz
lim f (z)e dz = lim f (L)e − f (−L)e + f (z)e dz
N →∞ L −L N →∞ iN π −L
 Z L 
1 iN π −iN π 0 i NLπz
≤ lim |f (L)| |e | + |f (−L)| e + f (z)e dz
N →∞ iN π −L
 Z L 
1 0
≤ lim |f (L)| + |f (−L)| + |f (z)| dz
N →∞ iN π −L
1
≤ lim [|f (L)| + |f (−L)| + 2LM ]
N →∞ iN π
(17)

where M is the maximum of the first derivative on the interval. Integration by parts is
allowed for piecewise C 1 functions, f (x). Clearly this expression goes to zero in the limit.
Now returning to the expression for the partial sum and the adjusted function
Z Lh i   N πz   
1  πz  N πz
lim f˜(x) − SN (x) = lim f˜(x) − f (x + z) sin cot + cos dz
N →∞ N →∞ 2L −L L 2L L
Z Lh  
1 ˜
i N πz  πz 
≤ lim f (x) − f (x + z) sin cot dz
2L N →∞ −L L 2L
Z Lh   
i N πz
+ f˜(x) − f (x + z) cos dz .
−L L
(18)

Since f˜(x) − f (x + z) is a piecewise C 1 function, then the second integral goes to zero in the
limit. In the first integral, the function

f˜(x) − f (x + z)
πz

(19)
tan 2L

is piecewise C 1 everywhere except, maybe at z = 0. However, using L’Hopital’s rule as z → 0

we find that
f˜(x) − f (x + z) df (x) 2L
lim πz

→ , (20)
z→0 tan 2L dx π
which is a bounded function since f (x) is C 1 . Therefore the first integral also goes to zero
in the limit.

2 Uniform Convergence of Fourier Series

The Fourier series of continuous, piecewise C 1 functions on the interval converge uniformly
to the function. This means

lim max |f (x) − SN (x)| → 0. (21)

N →∞ −L≤x≤L

4
 Proof:
First we note that, we have already proved that the Fourier series of continuous functions
f (x) converge pointwise to the function f (x), since the adjusted function f˜(x) is equal to
f (x) for continuous functions.
Pointwise convergence means that F S f (x) = f (x) so we can write

lim max |f (x) − SN (x)| = lim max |F Sf (x) − SN (x)|

N →∞ −L≤x≤L N →∞ −L≤x≤L
∞
X  nπx   nπx 
= lim max an cos + bn sin
N →∞ −L≤x≤L
n=N +1
L L
∞
X  nπx   nπx  (22)
≤ lim max |an | cos + |bn | sin
N →∞ −L≤x≤L
n=N +1
L L
∞
X
≤ lim |an | + |bn |
N →∞
n=N +1

If we can prove that this limit (the tail of series) vanishes then we have shown that the error
goes to zero everywhere. The vanishing of this limit is the same as having
∞
X
|an | + |bn | < ∞ (23)
n=1

i.e. the series converges if and only if the tail vanishes.

The coefficients of the Fourier series are
1 L
Z  nπx 
an = f (x) cos dx
L −L L
1 L  nπx  L 1 L
Z L  nπx 
0
= f (x) sin − f (x) sin dx
L nπ L −L L nπ −L L (24)
Z L
1  nπx 
=− f 0 (x) sin dx
nπ −L L
1
|an | ≤ An
nπ
where An are the Fourier cosine coefficients of the derivative function, f (x). Since f (x) is
continuous and C 1 , then f 0 (x) is piecewise continuous. Similarly for bn
1
|bn | ≤ Bn (25)
nπ
where Bn are the Fourier sine coefficients of the derivative function, f (x).
Now, let f 0 (x) = g(x) which is a piecewise continuous function. Therefore the coefficients
An , Bn exist and the function g(x) is L2 integrable, meaning
Z L
[g(x)]2 dx < ∞. (26)
−L

5
Bessel’s inequality (which is very straightforward to derive) says that
∞ Z L
A20 X 2
+ 2
An + Bn ≤ [g(x)]2 dx (27)
2 n=0 −L

and therefore ∞
X
A2n + Bn2 < ∞. (28)
n=0

The Cauchy Schwarz Inequality:

For two sequences α = (α1 , α2 ....αn ) and β = (β1 , β2 ....βn ), define

α · β = α1 β1 + .... + αn βn (29)

and √
|α| = α · α. (30)
The Cauchy Schwarz inequality states that
p
|α · β| ≤ |α| |β| (31)

Proof:
First, note that α · α ≥ 0 ∀α. Let r be any real number and notice that
0 ≤ |α + rβ|2
= (α1 + rβ1 )2 + .... + (αn + rβn )2
(32)
= α12 + .... + αn2 + 2 (α1 β1 + ... + αn βn ) r + β12 + ... + βn2 r2

0 ≤ α · α + 2α · β r + β · β r2
for all r. The right hand side of this expression describes a parabola in r which opens
upward (β · β) and whose interior is always greater than zero. Therefore the parabola can
only intersect the vertical axis at zero or one location. If it intersected at two locations
then there would be an interval between the two roots r1 < r < r2 where the interior of the
parabola would have values < 0 - which is not possible.
The condition that there be only one or no roots to a quadratic is that the discriminant
of the quadratic formula be negative. The expression for the roots of this parabola is
q
r = −α · β ± (α · β)2 − |α|2 |β|2 (33)

and the discriminant must be zero or negative. Therefore

(α · β)2 − |α|2 |β|2 ≤ 0

(α · β)2 ≤ |α|2 |β|2 (34)
α · β ≤ |α| |β|

6
which is the Cauchy-Schwarz inequality.

Now let’s substitute the value of an , bn into the expression

∞ N
X 1 X |An | |Bn |
|an | + |bn | = lim +
N →∞ π n n
n=1 n=1
 v v v v 
1u N u N
uX 1
u N u N
uX
2
1u X
2
uX 1 
≤ lim t |An | t
2
+ t |Bn | t
N →∞  π
n=1 n=1
n π n=1 n=1
n2 
v
u N
v
u N  v (35)
 u N
1 uX
2
uX
2
uX 1
≤ lim t |An | + t |Bn | t
π N →∞  n=1 n=1

n=1
n2
v
u N
1 uX 1
≤ C lim t .
π N →∞ n=1 n2

The logic of the last line is that the sequences An and Bn are each square integrable - which
means that the sum of the square roots can be pulled in front of the limit. It remains to
prove that the limit of the remaining sum exists,
N
X 1
lim <∞ (36)
N →∞
n=1
n2

which we will do by actually evaluating the sum. Consider the function

f (x) = x; (37)

its Fourier Sine series has coefficients

1 L
Z  nπx 
bn = x sin dx
L −L L
      Z L 
1 L nπL nπL  nπx 
= − L cos − (−L) cos − − cos dx
L nπ L L −L L
  (38)
2L
=− cos (nπ)
nπ
2L (−1)n+1
=
π n
and is written ∞
2L X (−1)n+1  nπx 
F S f (x) = sin L (39)
π n=1 n .

7
Now Bessel’s inequality applied to f (x) implies
∞
4L2 X 1 1 L 2
Z
≤ x dx
π 2 n=1 n2 L −L
∞
4L2 X 1 2L3
≤ (40)
π 2 n=1 n2 3L
∞
X 1 π2
2
≤ .
n=1
n 6

So we conclude that ∞
X π
|an | + |bn | ≤ C . (41)
n=1
6
Since this series coverges then its tail must vanish, i.e
∞
X
lim |an | + |bn | → 0. (42)
N →∞
n=N +1

This completes the proof of uniform convergence.