MAT 571 REAL ANALYSIS II

FOURIER TRANSFORM

PROFESSOR: JOHN QUIGG

SEMESTER: FALL 2014

Fourier transform

Definition 1. The Fourier transform of f ∈ L1 (R) is the function Ff = fb on R defined by

Z
(Ff )(t) = fb(t) = f (x)e−2πitx dx.

Proposition 2. If f ∈ L1 (R) then

(1) fb is continuous;
(2) kfbk∞ ≤ kf k1 .

Proof. (1) For all x, t ∈ R we have

f (x)e−πitx  = |f (x)|,
 

which is integrable. Since the integrand is continuous in t for every fixed x, the continuity
follows from a routine application of the Dominated Convergence Theorem.

(2) For all t ∈ R we have

Z  Z
−2πitx
   
fb(t) =  f (x)e
 dx ≤ |f (x)| dx = kf k1 .

Taking the sup over t, we get kfbk∞ ≤ kf k1 . 

It quickly follows that F is a bounded linear map from L1 (R) to the Banach space Cb (R)
of continuous bounded functions, which is called the Fourier transform. In fact, the image
lies in a significantly smaller Banach space, which we introduce here.
Definition 3. A function f : R → C vanishes at infinity if limx→±∞ f (x) = 0.
Lemma 4. The set of functions in `∞ (R) that vanish at infinity is a closed subspace.

Proof. By linearity of limits the set is a subspace; the nontrivial part is to show that it is
closed. Let {fn } be a sequence of functions in `∞ (R) that vanish at infinity, and let fn → f
in `∞ (R). Then fn → f uniformly, so limx→∞ f (x) = 0 since limx→∞ fn (x) = 0 for all n, and
similarly limx→−∞ f (x) = 0. 
Date: January 6, 2015.
1
2 PROFESSOR: JOHN QUIGG SEMESTER: FALL 2014

Notation. The set of continuous functions in R that vanish at infinity is denoted by C0 (R).

Lemma 5. C0 (R) is a closed subspace of Cb (R).

Proof. The main thing to observe is that every continuous function on R that vanishes at
infinity is bounded, from which the lemma follows since then C0 (R) is the intersection of two
closed subspaces of `∞ (R). So, let f ∈ C0 (R). Choose a < b in R such that |f (x)| < 1 for all
x < a and for all x > b. On the compact interval [a, b], since f is continuous it is bounded.
Then f is bounded on the three sets (−∞, a), [a, b], and (b, ∞), and hence on their union
R. 

Theorem 6 (Riemann-Lebesgue Lemma). F(L1 (R)) ⊂ C0 (R).

Proof. We have already seen that F(L1 (R)) ⊂ Cb (R). We must show that if f ∈ L1 (R) then
limt→±∞ fb(t) = 0.

First, if f = χ(a,b) , then

b
x=b
−e−2πitx 
Z
−2πitx
fb(t) = e dx =
a 2πit x=a
−2πita
e − e2πitb
=
2πit
t→±∞
−−−−→ 0.

Thus by linearity the the conclusion holds for step functions. Now, F : L1 (R) → Cb (R) is
continuous, the set of step functions is dense in L1 (R), and C0 (R) is a closed subset of Cb (R),
so the result follows. 

The Fourier Transform has many properties, of which we’ll only record a few. The first
involves translations, and for which we introduce a convenient notation:

Notation. If f : R → C and y ∈ R we define τy f : R → C by

(τy f )(x) = f (x − y).

Proposition 7. If f ∈ L1 (R) and y, s ∈ R then

(τy f )b(t) = e−2πity fb(t)

τs fb = e2πisx f b.


 MAT 571 REAL ANALYSIS II FOURIER TRANSFORM 3

Proof. The first follows from a simple change of variables:

Z
(τy f )b(t) = (τy f )(x)e−2πitx dx
Z
= f (x − y)e−2πitx dx
Z
= f (x)e−2πit(x+y) dx
Z
= f (x)e−2πitx e−2πity dx
Z
−2πity
=e f (x)e−2πitx dx

= e−2πity fb(t),
and similarly for the second. 

We pause to discuss translation a little further. Note that a bounded function f : R → C

is uniformly continuous if and only if limy→0 kτy f − f ku = 0. Also, it is an exercise that
every f ∈ Cc (R) is uniformly continuous.

By translation-invariance of Lebesgue measure, translation is an isometric isomorphism

of Lp (R) onto itself, for all 1 ≤ p ≤ ∞, i.e., the linear operator τy on Lp (R) satisfies
kτy f kp = kf kp for all y ∈ R and f ∈ Lp (R). Further, for finite p translation is continuous in
the following sense:
Proposition 8. If p < ∞ and f ∈ Lp (R) then
lim kτy f − τx f kp = 0 for all x ∈ R.
y→x

Proof. Since τs+t = τs τt , it suffices to take x = 0. First let f = χ(a,b) for a < b in R. We have
τy f − f = χ(a,b)4(a+y,b+y) ,
so if |y| < b − a
y→0
kτy f − f kpp = λ (a, b) 4 (a + y, b + y) = 2|y| −−→ 0.

By linearity of limits, if f is a step function in Lp (R) then limy→0 kτy f − f kp = 0.

Now let f ∈ Lp (R) be arbitrary, and let ε > 0. Choose a step function g such that
kf − gkp < ε, and choose δ > 0 such that kτy g − gkp < ε whenever |y| < δ. Then
kτy f − f kp ≤ kτy f − τy gkp + kτy g − gkp + kg − f kp
= kτy (f − g)kp + kτy g − gkp + kg − f kp
= 2kf − gkp + kτy g − gkp
< 3ε. 
4 PROFESSOR: JOHN QUIGG SEMESTER: FALL 2014

Folland gives an alternative proof, approximating by functions in Cc (R). Note that in the
above proof we allowed ourselves to be satisfied with 3ε rather than ε; this sort of flexibility
in estimation is common in analysis.

Now back to properties of the Fourier transform:

Proposition 9. If f ∈ L1 (R), a 6= 0, and g(x) = f (x/a), then

gb(t) = afb(at).

Proof. This is a simple change of variables, as in Proposition 7. 

For the next property, we introduce a notation for another class of functions:
Notation. For n ∈ N, C n (R) denotes the set of n-times continuously differentiable functions
on R.

Sometimes it convenient to exend the above notation to include n = 0, by letting C 0 (R)

be the set of continuous functions on R.
Proposition 10. If f and xf are both integrable, then fb ∈ C 1 (R), and

(fb)0 = (−2πixf )b.

Proof. For all x, t ∈ R,

 
∂
 f (x)e−2πitx  = −2πixf (x)e−2πitx  = 2π|xf (x)|,
  
 ∂t 

which is integrable in x, so by our rule for differentiating under the integral sign we have
Z
0 d
(fb) (t) = f (x)e−2πitx dx
dt
Z
∂
= f (x)e−2πitx dx
∂t
Z
= −2πixf (x)e−2πitx dx
= (−2πixf )b. 

By induction, it is straightforward to extend the above proposition to xn f and C n (R) for

any n ∈ N.
Proposition 11. If f ∈ C 1 (R) ∩ C0 (R) and f 0 ∈ L1 (R), then

(f 0 )b(t) = 2πitfb(t).
 MAT 571 REAL ANALYSIS II FOURIER TRANSFORM 5

Proof. Integrate by parts:

Z
0
(f )b(t) = f 0 (x)e−2πitx dx
x=∞ Z
−2πitx 
= f (x)e  − f (x)(−2πit)e−2πitx dx
x=−∞
Z
= 2πit f (x)e−2πitx dx

= 2πitfb(t). 

The following property exhibits a function that is its own Fourier transform, and conse-
quently is an example of considerable importance.
2 2
Proposition 12. If f (x) = e−πx , then fb(t) = e−πt .

Proof. Since f and xf are both integrable, we have

(fb)0 (t) = (−2πixf )b(t)
= i(f 0 )b(t) (since f ∈ C 1 (R) ∩ C0 (R))
= i(2πit)fb(t)
= −2πtfb(t).
By the elementary theory of differential equations, we have
2
fb(t) = ce−2πt
for some c ∈ R. Letting t = 0, we have
Z
2
c = fb(0) = e−πx dx = 1,
2 √
e−x dx =
R
by a routine change of variable in the well-known integral π. 

Convolutions

Throughout we are mainly interested in Lebesgue measure and integration on R or R2 .

“Measurable” will mean Lebesgue measurable, and “integrable” will mean with respect to
Lebesgue measure, unless otherwise specified. Further, L1 (R) and L1 (R2 ) will by default be
with respect to Lebesgue measure. All functions will be complex-valued unless otherwise
specified.
Lemma 13. Define α : R2 → R by α(x, y) = x − y. Then for every measurable set A ⊂ R
the inverse image α−1 (A) is Lebestue measurable.
6 PROFESSOR: JOHN QUIGG SEMESTER: FALL 2014

Proof. We can write A = B ∪ C, where B is Borel and λ(C) = 0. Since α is continuous,

α−1 (B) is Borel. We can find a Borel set D ⊃ C such that λ(D) = 0. We will show that
α−1 (D) has measure 0, and it will follow that α−1 (C) is measurable since it is contained in
α−1 (D) and λ2 is complete, where for this proof λ2 denotes Lebesgue measure on R2 . We
have
ZZ
2 −1 χα−1 (D) dλ2
λ (α (D)) =
Z Z
= χα−1 (D) (x, y) dx dy (Tonelli)
Z Z
= χD (α(x, y)) dx dy
Z Z
= χD (x − y) dx dy
Z Z
= χD (x) dx dy (translation-invariance)
Z
= λ(D) dy
Z
= 0 dy
= 0.

Thus both α−1 (B) and α−1 (C) are measurable, and hence so is

α−1 (A) = α−1 (B) ∪ α−1 (C). 

Corollary 14. If f is measurable on R then the functions

(x, y) 7→ f (x − y) and (x, y) 7→ f (y)

are measurable on R2 .

Proof. If A ⊂ C is Borel then f −1 (A) is measurable, and hence by Lemma 13 so is

{(x, y) ∈ R2 : f (x − y) ∈ A} = α−1 (f −1 (A)).

The argument for the second function is similar, but easier. 

Lemma 15. Let f, g ∈ L1 (R). Then the function (x, y) 7→ f (x − y)g(y) is integrable on R2 .
 MAT 571 REAL ANALYSIS II FOURIER TRANSFORM 7

Proof. By Corollary 14 the function is measurable, so it suffices to show that its absolute
value is integrable, and we apply Tonelli’s Theorem:
ZZ Z Z
|f (x − y)g(y)| dx dy = |f (x − y)g(y)| dx dy
Z Z
= |g(y)| |f (x − y))| dx dy
Z Z
= |g(y)| |f (x))| dx dy
Z  Z 
= |f (x)| dx |g(y)| dy < ∞. 

Definition 16. The convolution of measurable functions f and g on R is the function f ∗ g

defined by Z
(f ∗ g)(x) = f (x − y)g(y) dy,

for all x ∈ R where the integral makes sense.

Proposition 17. For all f, g ∈ L1 (R) the convolution f ∗ g is integrable.

Proof. This follows from Lemma 15 and Fubini’s Theorem. 

Thus, convolution gives a map

(f, g) 7→ f ∗ g : L1 (R) × L1 (R) → L1 (R).
It is routine to verify that this map is bilinear, i.e., f 7→ f ∗ g and g 7→ f ∗ g are linear, and
satisfies the distributive laws f ∗ (g + h) = f ∗ g + f ∗ h and (f + g) ∗ h = f ∗ h + g ∗ h,
so it should be regarded as a type of product on the vector space L1 (R). Thus L1 (R) is an
algebra. Moreover, this product is commutative and associative:
Theorem 18. If f, g, h ∈ L1 (R), then:

(1) f ∗ g = g ∗ f ;
(2) f ∗ (g ∗ h) = (f ∗ g) ∗ h.

Proof. (1)
Z ∞
f ∗ g(x) = f (x − y)g(y) dy
−∞
Z −∞
= f (u)g(x − u)(−du)
Z∞∞
= g(x − u)f (u) du
−∞
= g ∗ f (x).
8 PROFESSOR: JOHN QUIGG SEMESTER: FALL 2014

(2)
Z
f ∗ (g ∗ h)(x) = f (x − y)g ∗ h(y) dy
Z Z
= f (x − y) g(y − z)h(z) dz dy
ZZ
= f (x − y)g(y − z)h(z) dz dy
ZZ
= f (x − y)g(y − z) dy h(z) dz (Fubini)
ZZ
= f (x − z − u)g(u) du h(z) dz
Z
= f ∗ g(x − z)h(z) dz
= (f ∗ g) ∗ h(x). 
Remark 19. Note that by commativity we can express the convolution formula alternatively
as Z
f ∗ g(x) = f (y)g(x − y) dy.
As Folland points out, this could be used to slightly shorten the proof of (2) above.

For the next property, we need the following concept:

Proposition 20. If f, g ∈ L1 (R), then
{f ∗ g 6= 0} ⊂ {f 6= 0} + {g 6= 0}.

Proof. Let A = {f 6= 0} and B = {g 6= 0}. From the definition, it suffices to show that if
f ∗ g(x) 6= 0 then x ∈ A + B. Since
Z Z
0 6= f ∗ g(x) = f (x − y)g(y) dy = f (x − y)g(y) dy,
B
we can choose y ∈ B such that x − y ∈ A, and then x ∈ A + y ⊂ A + B. 
Proposition 21 (Young’s Inequality). If f ∈ L1 (R) and g ∈ Lp (R), then f ∗ g ∈ Lp (R) and
kf ∗ gkp ≤ kf k1 kgkp .

Proof. This follows immediately from the Integral Operator Theorem with K(x, y) = f (x −
y). 

Letting p = 1, we get an important special case of Young’s Inequality:

kf ∗ gk1 ≤ kf k1 kgk1 for all f, g ∈ L1 (R).
This makes L1 (R) a normed algebra, and since it is complete it is also called a Banach
algebra.

We need one more property of translation:

 MAT 571 REAL ANALYSIS II FOURIER TRANSFORM 9

Lemma 22. Let f, g : R → C be measurable, and let x, y ∈ R. If f ∗ g(x − y) exists then

τy (f ∗ g)(x) = (τy f ) ∗ g(x).

Proof.
τy (f ∗ g)(x) = f ∗ g(x − y)
Z
= f (x − y − z)g(z) dz
Z
= τy f (x − z)g(z) dz
= (τy f ) ∗ g(x). 
Notation. The set of uniformly continuous bounded functions on a metric space X is de-
noted Cucb (X).

It is a routine to verify that Cucb (X) is a closed subspace of Cb (X). Since it contains
Cc (R), it contains C0 (R).
Proposition 23. Let f ∈ Lp and g ∈ Lq . Then f ∗ g ∈ Cucb (R), and
kf ∗ gku ≤ kf kp kgkq .
Moreover, if 1 < p < ∞, then f ∗ g ∈ C0 (R).

Proof. For this proof we introduce the notation fe(x) = f (−x). For all x ∈ R we have
Z Z
 
|f (x − y)g(y)| dy = τx (fe)(y)g(y) dy

≤ kτx (fe)kp kqkq (Hölder)

= kfekp kgkq
= kf kp kgkq .
Thus y 7→ f (x − y)g(y) is integrable, and by the triangle for integrals we have
Z  Z
 
 f (x − y)g(y) dy  ≤ |f (x − y)g(y)| dy ≤ kf kp kgkq .
 

For the uniform continuity, we have

kτy (f ∗ g) − f ∗ gku = k(τy f ) ∗ g − f ∗ gku
= k(τy f − f ) ∗ gku
≤ kτy f − f kp kgkq
y→0
−−→ 0.
10 PROFESSOR: JOHN QUIGG SEMESTER: FALL 2014

Now let 1 < p < ∞. Choose fn , gn ∈ Cc (R) such that fn → f in Lp and gn → g in Lq .

Then fn ∗ gn ∈ Cc (R) ⊂ C0 (R) for every n, and
kfn ∗ gn − f ∗ gku ≤ kfn ∗ gn − f ∗ gn ku + kf ∗ gn − f ∗ gku
= k(fn − f ) ∗ gn ku + kf ∗ (gn − g)ku
≤ kfn − f kp kgn kq + kf kp kgn − gkq
n→∞
−−−→ 0,
because kgn kq → kgkq . The result follows since C0 (R) is closed in Cb (R). 

One of the main reasons that we introduced convolutions is the following:

Theorem 24. If f, g ∈ L1 (R) then
(f ∗ g)b = fbgb.

Proof. For t ∈ R we have

Z
(f ∗ g)b(t) = f ∗ g(x)e−2πitx dx
ZZ
= f (x − y)g(y) dy e−2πitx dx
ZZ
= f (x − y)g(y)e−2πitx dx dy (Fubini)
ZZ
= f (x)g(y)e−2πit(x+y) dx dy
ZZ
= f (x)g(y)e−2πitx e−2πity dx dy
Z Z
−2πtx
= f (x)e dx g(y)e−2πity dy

= fb(t)b
g (t). 

Proposition 25. If f, g ∈ L1 (R) then both fbg and f gb are integrable, and
Z Z
f g = f gb.
b

Proof. fbg is integrable because it’s the product of the bounded measurabe function fb and
the integrable function g, and similarly for f gb. We have
Z ZZ
f (t)g(t) dt =
b f (x)e−2πitx dx g(t) dt
Z Z
= f (x) g(t)d−2πitx dt dx (Fubini)
Z
= f (x)bg (x) dx. 
 MAT 571 REAL ANALYSIS II FOURIER TRANSFORM 11
R
Theorem 26. Let φ ∈ L1 (R) with φ = 1, and for a > 0 define φa : R → C by
1 x
φa (x) = φ .
a a
If p < ∞ then
lim kf ∗ φa − f kp = 0 for all f ∈ Lp (R).
a↓0

Proof. First note that for all a > 0 we have

Z Z Z
1 x
φa (x) dx = φ dx = φ(x) dx = 1.
a a
Thus
Z Z
f ∗ φa (x) − f (x) = f (x − y)φa (y) dy − f (x) φa (y) dy
Z


= f (x − y) − f (x) φa (y) dy
Z


= f (x − ay) − f (x) φ(y) dy
Z


= τay f (x) − f (x) φ(y) dy.

Thus
Z 1/p
p
kf ∗ φa − f kp = |f ∗ φa (x) − f (x)| dx
Z Z p 1/p


=  τay f (x) − f (x) φ(y) dy  dx
 
Z Z p 1/p
   
=  τay f (x) − f (x)|φ(y)| dy  dx
 
Z
≤ kτay f − f kp |φ(y)| dy,

using Minkowski’s Inequality for Integrals with dν(y) = |φ(y)|

R dy. Now, kτay f −f kp ≤ 2kf kp ,
so the Dominated Convergence Theoerem implies that kτa yf − f kp |φ(y)| dy → 0 as a ↓
0. 

Note that in the above proof we used the Dominated Convergence Theorem for a limit as
a real variable a goes to 0. This is valid since we can take any sequence an → 0 (a technique
we have used several times before).

It turns out that the algebra L1 (R) has no multiplicitve identity, i.e., there does not exist
g ∈ L1 (R) such that f ∗ g = f for all f ∈ L1 (R). Perhaps the easiest way to see this is to
note that if there is no continuous function equal to f a.e., e.g., if f = χ(a,b) , then f ∗ g, being
continuous, is not the same element of L1 (R) as f . However, the case p = 1 of Theorem 26
shows that there is something that works almost as well. For this reason {φa }t>0 is called
an approximate identity for L1 (R).
12 PROFESSOR: JOHN QUIGG SEMESTER: FALL 2014

Finally, we can invert the Fourier transform. First, we need a slight variation on the
Fourier transform:
Notation. If f ∈ L1 (R) define f ∨ : R → C by
Z
∨
f (t) = f (x)e2πitx dx.

Note that f ∨ (t) = fb(−t).

Theorem 27 (Fourier Inversion Theroem). If both f and fb are integrable, then
f = (fb)∨ = (f ∨ )b a.e.,
and moreover f is a.e. equal to a continuous function.

Proof. Given a > 0 and x ∈ R define φ : R → C by

2 2
φ(t) = e2πitx e−πa t .
2
Letting g(t) = e−πt and h(t) = g(at), we have
1  y  1 −πy2 /a2
bh(y) = gb = e = ga (y),
a a a
where the notation ga means the same as in Theorem 26. Since φ(t) = e2πitx h(t), we have
φ(y)
b = τxb h(y − x) = ga (y − x) = ga (x − y),
h(y) = b
since ga is an even function.

Since fb and φ are in L1 (R),

Z Z
fb(t)φ(t) dt = f (y)φ(y)
b dy
Z
= f (y)ga (x − y) dy
= f ∗ ga (x).
Since g ∈ L1 (R) and Z Z
2
g= e−πt dt = 1,
we have
lim f ∗ ga = f in L1 (R).
a↓0

On the other hand, since fb is integrable and |φ| ≤ 1, the Dominated Convergence Theorem
implies that for all x ∈ R
Z Z
2πitx −πa2 t2
lim f (t)e
b e dt = fb(t)e2πitx dt = (fb)∨ (x).
a↓0

This gives f = (fb)∨ a.e., and by symmetry we have f = (f ∨ )b a.e. Now the last part of the
theorem follows immediately, since f is a.e.-equal to the Fourier transform of an integrable
function. 
 MAT 571 REAL ANALYSIS II FOURIER TRANSFORM 13

Corollary 28. The Fourier transform F : L1 (R) → C0 (R) is injective.

Proof. Since F is linear, it suffices to show that its kernel is {0}. So, suppose f ∈ L1 (R) and
fb = 0. Then in particular fb is also integrable, so by the Fourier Inversion Theorem, for a.e.
x we have Z
f (x) = fb(t)e2πitx dt = 0.