(February 19, 2005)

Applications to Fourier series

Paul Garrett garrett@math.umn.edu http://www.math.umn.edu/˜garrett/

We showed that the Fourier series of periodic C 1 functions f on R1 converge uniformly (that is, in the C o
metric), and also converge pointwise to the original f . Thus, in the one dimensional case, the Fourier series
of periodic C 1 functions do converge to the functions.
From the general discussion of L2 functions, we know that a sequence of L2 functions has a subsequence
which converges pointwise. (In fact, that is how the limit is constructed in proving completeness of L2 .)
This applies to Fourier series, but pointedly does not say anything about the pointwise convergence of the
whole sequence of partial sums.
Theorem: There is a periodic continuous function f on [0, 1] (periodicity meaning that f (0) = f (1)) such
that the Fourier series X
fˆ(n) e2πinx
n∈Z

of f does not converge at 0, where as usual

Z 1
fˆ(n) = e−2πinx f (x) dx
0

The Banach-Steinhaus (Uniform Boundedness) theorem for Banach spaces V has as a special case that for
a collection of continuous linear functionals {λα : α ∈ A} on V either there is a uniform bound M such that
|λα | ≤ M for all α ∈ A, or else there is v in the unit ball of V such that

sup |λα v| = +∞
α∈A

In fact, the collection of such v is dense in the unit ball, and is an intersection of a countable collection of
dense open sets (called a Gδ ).
Consider the functionals X
λN (f ) = fˆ(n)
|n|≤N

These are partial sums of the Fourier series of f evaluated at 0. Since

 
Z 1 X 
−2πinx 
 
|λN (f )| ≤ 
 e  |f (x)| dx
0 |n|≤N 
 
Z 1  X  X
−2πinx 
e−2πinx |1
 
≤ |f |∞ · 
 e  dx = |f |∞ · |
0 |n|≤N  |n|≤N

The point is to show, rather, that equality holds, namely

X
|λN | = | e−2πinx |1
|n|≤N

and that as n → ∞ the latter L1 -norms go to ∞.

First, summing the finite geometric series and doing a little adjustment gives
1 1
X e−2πiN x − e−2πi(−N −1)x e2πi(N + 2 )x − e−2πi(N + 2 )x sin 2π(N + 21 )x
e−2πinx = = =
e−2πix − 1 eπix − e−πix sin πx
|n|≤N

1
 Paul Garrett: Applications to Fourier series (February 19, 2005)

Applying to this the elementary inequality

| sin t| ≤ |t|
gives
Z 2π(N + 12 )
1  sin 2π(N + 12 )x 
Z   Z 1
 dx ≥ 1 1 1
 | sin 2π(N + )x| dx = | sin x| dx
0
 sin πx 
0 2 x 0 x
N Z 2π` N
X 1 X 1
≥ | sin x| dx ≥ →∞
` 2π(`−1) `
`=1 `=1

as N → ∞. Thus, the L1 -norms do go to ∞.

Let g(x) be the sign of the kernel

X sin 2π(N + 21 )x
e−2πinx =
sin πx
|n|≤N

from above. Let gj be a sequence of periodic continuous functions with |gj | ≤ 1 and going to g pointwise.
Then by the dominated convergence theorem
Z 1 X Z 1 X Z 1 X
lim λN (gj ) = lim gj (x) e−2πinx dx = g(x) e−2πinx dx = | e−2πinx | dx
j j 0 0 0
|n|≤N |n|≤N |n|≤N

That is, the L1 -norm of the kernel really is the norm of the functional.
The Banach-Steinhaus theorem implies that there is f ∈ C o (R/Z) such that

sup |λN (f )| = +∞
N

That is, the Fourier series of f does not converge at 0. ///

The result can be strengthened by using Baire’s theorem again.
Theorem: There is f ∈ C o (R/Z) such that for a dense Gδ of x in [0, 1]
 
 X 
ˆ 2πinx 
 
sup  f (n) e =∞

N 
|n|≤N 

In fact, the set of such f is a dense Gδ in C o (R/Z).

Proof: Take a dense countable set of points xj in the interval. Let λj,N be the continuous linear functionals
on C o (R/Z) defined by X
λj,N (f ) = fˆ(n) e2πinxj
|n|≤N

As in the previous proof, the set Ej of functions f where

sup |λj,N f | = +∞
N

is a dense Gδ , so the intersection

E = ∩j Ej
is a dense Gδ . Also X
s(f, x) = sup | fˆ(n) e2πinxj |
N
|n|≤N

2
 Paul Garrett: Applications to Fourier series (February 19, 2005)

is the sup of continuous functions, so is lower semicontinuous, and so for any f ∈ C o (R/Z)

{x ∈ [0, 1] : s(f, x) = +∞}

is a Gδ in [0, 1]. Then for f ∈ E, the corresponding Gδ contains a dense subset (the xi ). ///

Lemma: In a complete metric space X, d (with no isolated points), a dense Gδ is uncountable.

Proof: Suppose that \
E= Un = {x1 , x2 . . .}
n

is a dense Gδ , where the sets Un are (necessarily dense) open. Let

Vn = Un − {x1 , . . . , xn }
T
Then Vn is still dense (by the assumption that there are no isolated points) and open, but n Vn = φ,
contradicting the Baire theorem. ///

Theorem: (Riemann-Lebesgue Lemma) Let f ∈ L1 [0, 1]. Then

fˆ(n) → 0

Proof: Finite linear combinations of exponentials are dense in C o (R/Z) in sup norm, and C o [0, 1] is dense
in L1 [0, 1]. That is, given f ∈ L1 there is g ∈ C o (R/Z) such that |f − g|1 < ε and a finite linear combination
h of exponentials such that |g − h|∞ < ε. Since the sup-norm dominates the L1 -norm, |f − h|1 < 2ε.
Given such h, for large-enough n the Fourier coefficients are simply 0, by orthogonality of distinct
exponentials. Thus, Z 1 
|fˆ(n)| =  (f (x) − h(x)) e−2πinx dx ≤ |f − h|1 < 2ε
 
0

This proves the Riemann-Lebesgue Lemma. ///

Theorem: (corollary of Baire and Open Mapping) It is not true that every sequence an → 0 occurs as the
collection of Fourier coefficients of an L1 function.
Proof: Let c0 be the collection of sequences {an : n ∈ Z} of complex numbers such that

lim an = 0
n→∞

with norm
|{an }| = sup |an |
n
1
This norm makes c0 a Banach space. Let T : L → c0 by

T f = {fˆ(n) : n ∈ Z} ∈ c0

Since Z 1  Z 1
|fˆ(n)| =  f (x) e−2πinx dx ≤
 
|f (x)| dx = |f |1
0 0

it is clear that |T | ≤ 1. In fact, taking f (x) = 1 shows that |T | = 1.

We check that T is injective: if fˆ(n) = 0 for every n for some L1 function f , then for any trigonometric
polynomial h we have Z 1
f (x) h(x) dx = 0
0

3
 Paul Garrett: Applications to Fourier series (February 19, 2005)

and by dominated convergence (and density of trigonometric polynomials) the same holds for all continuous
h. From Lusin’s theorem and (again) dominated convergence, the same applies with h being a characteristic
function of a measurable set. Thus f = 0, proving injectivity of T .
If T were surjective from L1 to c0 , then the Open Mapping Theorem would assure the existence of δ > 0
such that for every L1 function f
|fˆ|∞ ≥ δ |f |1
Let X
fN (x) = e−2πinx
|n|≤N

On one hand, the sup norm of fˆN is certainly 1. On the other hand, the computation above shows that
the L1 norm of fN goes to ∞ as N →= ∞. Thus, there is no such δ > 0. Thus, T cannot be surjective.
///

Remark: For any locally compact Hausdorff space X we can define

Co (X) = Coo (X) = {f ∈ C o (X) : for each ε > 0 there is compact K such that |f (x)| < ε off K}

and show that with sup norm this space is complete.