Lecture 16

March 11, 20 21

1 Convergence theorems
We are going to prove the pointwise convergence of the classical full Fourier
series.
For a C 1 function f (x) on (−π, π) the Fourier series is
∞
1 X
S(x) = A0 + (An cos nx + B sin nx)
2 n=1

with the coecients

Z π
dy
An = f (y) cos ny (n = 0, 1, 2, · · · )
−π π
Z π
dy
Bn = f (y) sin ny (n = 1, 2, · · · )
−π π

The N th partial sum of the series is

N
1 X
SN (x) = A0 + (An cos nx + Bn sin nx).
2 n=1

We want to prove that SN (x) converges to f (x) as N → ∞. So the Fouries

series S(x) equals the function f (x) in (−π, π). Replacing the formulas An and
Bn into SN (x), we have
Z π N
1 X
SN (x) = [1 + 2 (cos ny cos nx + sin ny sin nx)]f (y)dy
2π −π n=1
Z π N
1
(1)
X
= [1 + 2 cos n(x − y)]f (y)dy.
2π −π n=1

Denote the Dirichlet kernel KN to be

N
(2)
X
KN (θ) = 1+2 cos nθ.
n=1

1
Because of the observation
1 1 1
2 cos nθ sin θ = sin(n + )θ − sin(n − )θ.
2 2 2
Thus we have
N
X sin(n + 21 )θ − sin(n − 21 )θ
KN (θ) = 1+
n=1
sin 21 θ
sin(N + 12 )θ − sin 12 θ
= 1+
sin 12 θ
sin(N + 12 )θ
= .
sin θ2

The graph of the Dirichlet kernel KN is

Figure 1: N = 10
2
Compare this with the heat kernel St (x) = . The graph for St is
x
√1 e− 4t
2 πt

2
 Figure 2: t = 0.001

Letting θ = y − x and using the evenness of KN , formula (1) takes the form

Z π
dy
SN (x) = KN (y − x)f (y) .
−π 2π

Notice that by the denition of KN

Z π Z π N
dy X dy
KN (y − x) = [1 + 2 cos n(y − x)]
−π 2π −π n=1
2π
= 1.

Then
Z π
dy
SN (x) − f (x) = KN (y − x)[f (y) − f (x)]
−π 2π
Z π
1 [f (y) − f (x)] dy
= sin(N + )(y − x) .
−π 2 2 sin (y−x) π 2

We have assumed that f (x) has a dierentiable derivative, so f (x)−f x−y

(y)
and
(y−x) are continuous functions with respect the variable θ .
f (y)−f (x) x−y
h(θ) = x−y 2 sin2

3
Then
Z π
1 1
SN (x) − f (x) = sin(N + )(y − x)h(y − x)dy
π −π 2
Z π−x
1 1
= sin(N + )θh(θ)dθ.
π −π−x 2
Because {Xn (θ)}={sin(n + 12 )θ} are an orthogonal set of functions on the
interval (−x, π−x). Hence they are also orthogonal on the interval (−x−π, −x+
π). Due to the least-Square Approximation theorem, if khk2 = (h, h) < ∞ we
have from Bessel's inequality
∞
X (h, Xn )2
≤ khk2 .
n=1
(Xn , Xn )

By direct calculation
Z π−x
1
(Xn , Xn ) = sin2 (N + )θdθ = π.
−π−x 2
So we have for bigger N ,
(h, XN ) → 0 as N → ∞.
Then we check that khk2 < ∞ which is
π−x π−x
f (x + θ) − f (x) 2
Z Z
h2 (θ)dθ = [ ] dθ < ∞.
−π−x −π−x 2 sin θ2
The above inequality is true because h is a continues function.
Exercise 1. If a period function f (x) itself is only piecewise continuous and
f 0 (x) is also piecewise continuous on −∞ < x < ∞, prove that for any xed x
1
lim |SN (x) − [f (x+) + f (x−)]| = 0.
N →∞ 2
We are going to prove the uniform convergence theorem for classical Fourier
series. We assume again that f (x) and f 0 (x) are continuous functions of period
of 2π .
Denote An and Bn are the Fourier coecients of f (x) and let A0n and Bn0
are the Fourier coecients of f 0 (x).
We integrate by parts to get
Z π
dx
An = f (x) cos nx
−π π
Z π
1 dx
= f (x) sin nx|π−π − f 0 (x) sin nx
nπ −π nπ
Z π
dx
= − f 0 (x) sin nx
−π nπ
= −Bn0 .

4
 Similarly,
1
Bn = − A0n .
n
Due to Bessel's inequality for the functions f 0 (x) that the innite series
∞
X Z π
(|A0n |2 + |Bn0 |2 ) ≤ π |f 0 (x)|2 dx < ∞.
n=1 −π

∞ ∞
1 X 1 X
A0 + (An cos nx + B sin nx) ≤ |A0 | + (|An | + |Bn |)
2 n=1
2 n=1
∞
1 X 1
≤ |A0 | + (|A0n | + |Bn0 |)
2 n=1
n
∞ ∞
1 X 1 1 X 0 1
≤ |A0 | + ( 2
) 2[ (|An | + |Bn0 |)2 ] 2
2 n=1
n n=1
∞ ∞
1 X 1 1 X 0 2 1
≤ |A0 | + ( 2
) 2 [2 (|An | + |Bn0 |2 )] 2
2 n=1
n n=1
< ∞.

Here we used the Schwarz inequality for innite series:

∞ ∞ ∞
1 1
X X X
an bn ≤ ( a2n ) 2 ( b2n ) 2 .
n=1 n=1 n=1

So the Fourier series converges absolutely.

Moreover, we have
∞
X
max|f (x) − SN (x)| ≤ |An cos nx + Bn sin nx|
n=N +1
X∞
≤ (|An | + |Bn |)
n=N +1
(asN → ∞) → 0.

2 The Gibbs phenomenon

Let f (x) be a step function with a jump
(
1 0<x<π
f (x) =
−1 −π < x < 0.

5
 Note from the previous discussion, we have
1
lim |SN (0) − [f (0+) + f (0−)]| = 0.
N →∞ 2
In fact,
π
sin(N + 12 )y 0
sin(N + 21 )y
Z Z
1
SN (0) = [ dy − dy]
2π 0 sin y2 −π sin y2

Then
sin(N + 12 )y
π
sin(N + 21 )y
Z Z 0
1 1 1
|SN (0) − [f (0+) + f (0−)]| = [ y − 1]dy − [ − 1]dy
2 2π 0 sin 2 2π −π sin y2
Z π X N Z 0 X N
1 1
= 2 cos nydy − 2 cos nydy
2π 0 n=1 2π −π n=1
= 0.

But we are going to prove that for some xN → 0

lim SN (xN ) 6= 0.
N →∞

Moreover, this limit is 9 percent higher than the jump of the function f .
Here the jump is 2.
This is called Gibbs phenomenon.
Let xN = N π+ 1 , then the partial sum SN is
2

Z π
dy
SN (xN ) = KN (y − xN )f (y)
−π 2π
Z π−xN −xN Z
1
= [ KN (θ)dθ −
KN (θ)dθ]
2π −xN −π−xN
sin(N + 12 )θ sin(N + 12 )θ
Z π−xN Z xN
1
= [ dθ + dθ]
2π −xN sin θ2 π+xN sin θ2
sin(N + 12 )θ sin(N + 12 )θ
Z π−xN Z xN
1
= [ θ
dθ + dθ].
2π π+xN sin 2 −xN sin θ2

We are going to estimate the above two integrals

π−xN
sin(N + 12 )θ π
sin(N + 21 )θ
Z Z
1 1
dθ → dθ → 0 as N → 0. (3)
2π π+xN sin θ2 2π π sin θ2

6
And
Z π1
xN
sin(N + 21 )θ sin(N + 21 )θ
Z
1 1 N+
2
dθ = dθ
2π −xN sin θ2 2π − π 1 sin θ2
N+
2

1 π
Z
1 sin ϕ
(letϕ = (N + )θ) = dϕ
2 π −π (2N + 1) sin 2Nϕ+1
1 π sin ϕ
Z
→ dϕ
π −π ϕ
≈ 1.179. (4)
Combining (3) and (4), we have
lim SN (xN ) ≈ S20 (x20 ) ≈ 1.179 ≈ 9% ∗ 2 + 1.
N →∞

This is Gibbs's 9 percent overshoot phenomenon.

R 0 sin[(N + 21 )(x−y)] dy
The graph for SN (x) = ( 0π − −π
R
) sin 1 (x−y) 2π
2

Figure 3: N = 20