Real Analysis II

Benjamin Smith∗

This is based on a course by Dmitry Jakobson found at http://www.math.mcgill.ca/jakobson/. All errors

are the responsibility of the author.

Contents

Contents 1

1 Complex Measures 2
1.1 Total variation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Absolute continuity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Raydon-Nikodym Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4 Riesz Representation Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2 Differentiation 7
2.1 Derivatives of (Lebesgue) measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Fundamental Theorem of Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3 Absolutely continuous functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.4 Differentiable transformations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

3 Limits of eigen functions on a torus 17

4 Fourier Transformations 20
4.1 L2 Fourier Transformations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.2 Inversion formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.3 Interpolation Theorems in Lp . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

5 Distributions 28
5.1 Locally summable functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
5.2 Derivative of distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
5.3 Sobolev Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
5.4 Multiplication and Convolution of distributions by C ∞ functions . . . . . . . . . . . . . . . . 31

Index 33

Bibliography 34
∗ [email protected]

1
Chapter 1

Complex Measures

1.1 Total variation

Let Σ be a σ algebra on a set X, a complex measure µ on Σ is a comlpex valued function on Σ such that
∞
X
µ(E) = µ(Ei )
i=1

for all countable partitions {Ei }(⊂Σ) of E(∈ Σ). Notice that, unlike for positive measures, convergence of
this series is required and in fact every rearrangement of this series is required to converge. This implies
also, with a bit of work, that the series is absolutely convergent.
We seek to find a positive real measure λ which dominates µ in the sense that |µ(E)| ≤ λ(E) for all
E ∈ Σ. Keeping λ as small as possible, it certainly must satisfy
∞
X ∞
X
|µ(Ei )| ≤ λ(Ei ) = λ(E), ∀{Ei }`E.
i=1 i=1

Hence, we define the total variation measure of µ by

∞
X
|µ|(E) := sup |µ(Ei )|.
{Ei }`E i=1

This is indeed a positive bounded real measure as the following two theorems will provide.
Theorem 1.1. The total variation measure |µ| of a complex measure µ on Σ is a positive measure on Σ.
This next Lemma about complex numbers is useful in proving the boundedness of |µ|.
Lemma 1.2. If z1 , . . . , zN ∈ C then there is a subset S⊂{1, . . . , N } such that
  N
X  1X
zk  ≥ |zk |.
 
 π


k∈S k=1

Theorem 1.3. If µ is a complex measure on (X, Σ) then |µ|(X) < ∞.

The collection of al complex measures on a a sigma algebra (X, Σ) is a normed linear vector space under
the obvious operations and ||µ|| := |µ|(X).

Positive and negative measures

If µ is a bounded real measure on (X, Σ), define the positive and negative variations of µ by
1 1
µ+ = (|µ| + µ), µ− = (|µ| − µ).
2 2

2
CHAPTER 1. COMPLEX MEASURES 3

Both of these can be shown to be positive bounded measures and

µ = µ+ − µ− , |µ| = µ+ + µ− .

This is known as the Jordan decomposition of µ.

1.2 Absolute continuity

Let µ be a positive and λ an arbitrary (positive or complex) measure on (X, Σ). Note that positive measures
are not a sub class of complex measures because they can be infinite on some sets. λ is absolutely continuous
with respect to µ (written λ  µ) if λ(E) = 0 whenever µ(E) = 0.
If there is a set A ∈ Σ such that λ(E) = λ(A ∩ E) for all E ∈ Σ then λ is said to be concentrated on A.
This is equivalent to saying λ(E) = 0 whenever E ∩ A = φ.
If two measures λ1 , λ2 are concentrated on disjoint sets A, B then they are said to be mutually singular
and we write λ1 ⊥ λ2 .
Properties 1. Let µ, λ, λ1 , λ2 be measures on (X, Σ) with µ positive.
(a) λ concentrated on A, then so is |λ|,
(b) λ1 ⊥ λ2 ⇒ |λ1 | ⊥ |λ2 |,
(c) λ1 ⊥ µ and λ2 ⊥ µ ⇒ λ1 + λ2 ⊥ µ,
(d) λ1  µ and λ2  µ ⇒ λ1 + λ2  µ,
(e) λ  µ ⇒ |λ|  µ
(f) λ1  µ and λ2 ⊥ µ ⇒ λ1 ⊥ λ2 , and
(g) λ  µ and λ ⊥ µ then λ = 0.

Proof. (a) If E ∩ A = 0 and Ej is any partition of E then λ(Ej ) = 0 for all j and hence also |λ|(E) = 0.
(b) Follows from (a).
(c) We have disjoint A1 , B1 and A2 , B2 separating the weight of λ1 from µ and λ2 from µ. Now, λ1 + λ2
is concentrated on A1 ∪ A2 and µ on B1 ∩ B2 which are disjoint.
(d) easy.
P
(e) If µ(E) = 0 and {Ej }`E, then µ(Ej ) = 0 and since λ  µ we have λ(Ej ) = 0 hence |λ(Ej )| = 0.
That is |λ|(E) = 0.
(f) Let A and B disjoint so that λ2 is concentrated on A and µ concentrated on B. By absolute continuity,
λ1 (E) = 0 for all E⊂A hence λ1 is concentrated on Ac .
(g) By (f), we have λ ⊥ λ so λ = 0.

One of the most significant properties of sigma-finite measures is given by this lemma
Lemma 1.4. If µ is a positive σ-finite measure on a σ-algebra (X, Σ), then there is a function w ∈ L1 (µ)
such that 0 < w(x) < 1 for all x ∈ X.

The purpose of this lemma is that µ can be replaced by a finite measure µ̃ (namely dµ̃ = wdµ) which,
by the strict positivity of w, has the same null sets as µ.
CHAPTER 1. COMPLEX MEASURES 4

1.3 Raydon-Nikodym Theorem

Theorem 1.5. [Lebesgue-Raydon-Nikodym Theorem] Let µ be a positive σ-finite measure and λ a complex
measure on (X, Σ).
(a) The is a unique pair of complex measures λa , λs on Σ such that

λ = λa + λs , λa  µ, λs ⊥ µ.

If λ is positive and finite then so are λa and λs .

(b) There is a unique h ∈ L1 (µ) such that
Z
λa (E) = hdµ
E

for every E ∈ Σ.

The pair λa , λs is called the Lebesgue decomposition of λ relative to µ and h is called the Raydon-Nikodym
derivative of λa with respect to µ.
Uniqueness of this decomposition is easily shown by suppose there was another pair, say λ0a , λ0s . Then
λa − λ0a = λs − λ0s satisfying that λa − λ0a  µ and λs − λ0s ⊥ µ so part (g) of our properties imply both
sides are zero giving uniqueness.
If λ is positive and sigma finite, this still holds except h ∈ L1loc . That is ∀x there is U ⊆ X, such that
h|U ∈ L1 (U, dµ).
Existence.....

1.4 Riesz Representation Theorem

If λ is not sigma finite, then this can fail. For example if µ is the Lesbegue measure on [0, 1], λ the counting
measure.

Theorem 1.6. If µ and λ are complex measures on (X, Σ), then the following are equivalent:

1. λ  µ
2. For all  > 0 there is a δ > 0 such that if |µ(E)| < δ then |λ(E)| < .

Proof. Assuming 2, If µ(E) = 0, then µ(E) < δ for all δ implying λ(E) <  for all  > 0, hence |λ(E)| = 0
Assuming 1, suppose 2 fails. That is, there is an  > 0 and {En } a partition of E such that, for all
n, |µ(En )| ≤ 2−n , but |λ(En )| ≥  so also |λ|(En ) ≥ . Let

An = ∪∞
i=n Ei , A = ∩∞
i=1 An .

Then µ(An ) < 2−n+1 , An ⊃ An+1 ⊃ · · · ⊃ A hence µ(A) = 0. However,

|λ|(A) = lim |λ|(An ) ≥  > 0

n→∞

so that |λ|(A) ≥ , hence |λ| is not  µ implying λ is not  µ.

Remark R1. If λ is unbounded, then 1 does not imply 2. For example µ = Lebesgue measure on (0, 1) and
λ(E) = E dtt .

Theorem 1.7. If µ is a complex measure on (X, Σ) then there exists an h : X → C with |h(x)| = 1 for all
x such that dµ = hd|µ|.
CHAPTER 1. COMPLEX MEASURES 5

Proof. µ  kµ| so 1.5 implies there is an h ∈ L1 (|µ|) such that dµ = hd|µ|. Let Ar = {x : |h(x)| < r} and
{Ei } a partition of Ar then
X X Z  X

|µ(Ei )| = 
 hd|µ| ≤ r · |µ|(Ei ) = r|µ|(Ar )
i i Ei i

So µ(Ar ) ≤ r|µ|(Ar ) for all r. If r < 1 then we must have that |µ|(Ar ) = 0 and so |h| ≥ 1 almost everywhere.
If |µ|(E) > 0, then  
 |µ(E)|
Z
 1
0 ≤  hd|µ| = ≤1
|µ|(E) |µ|(E)
R
hd|µ|
Apply Theorem 1.40 of [2] gives E
|mu|(E) ∈ B1 (0)⊆C. So |h(x)| ≤ 1 almost everywhere and or result is
shown.

TheoremR 1.8. µ a positive measure on (X, Σ), g ∈ L1 (µ). If λ(E) :=

R
E
gdµ for all E ∈ Σ, then
|λ|(E) = E |g|dµ.

Proof. There is an h of unit norm such that dλ = hd|λ| and we have that dλ = gdµ, so hd|λ| = gdµ and
d|λ| = h̄gdµ. With |λ| ≥ 0, µ ≥ 0, then h̄g ≥ 0 almost everywhere. Hence, h̄g = |g| almost everywhere with
respect to µ.

Theorem 1.9. [Hahn decomposition Theorem] If µ is a real measure on (X, Σ), then there exist disjoint
A, B ∈ Σ with X = A ∪ B such that the positive and negative variations of µ satisfy

µ+ (E) = µ(A ∩ E); µ− (E) = −µ(B ∩ E)

for all E ∈ Σ. This is called the Hahn decomposition of µ

−1 −1
Proof. µ = hd|µ| with |h| = 1, ( h ∈ R which implies h(x) = ±1. Let A = h ({1}), B = h ({−1}) so
h, on A
µ+ = 21 (|µ| + µ) = 12 (1 + h) = . Now,
0, on B
Z Z
1
µ+ (E) = (1 + h)d|µ| = hd|µ| = µ(E ∩ A)
2 E E∩A

and similarly for µ− .

Corollary 1.10. If µ = λ1 − λ2 with λ1 , λ2 ≥ 0, then λ1 ≥ µ+ , λ2 ≥ µ− .

Proof. µ ≤ λ1 ; µ+ (E) = µ(E ∩ A) ≤ λ1 (E ∩ A)λ1 (E).

Theorem 1.11 (Riesz Representation). X a locally compact Hausdorff space, Λ a positive linear functional
on Cc (X). Then there exists a σ-algebra, Σ in X that contains Borel sets and ∃! positive measure µ on Σ
that represents Λ. That is Z
Λf = f dµ, ∀f ∈ Cc (X)

and such that

(i) µ(K) < ∞, ∀K⊆X compact

(ii) µ is outer regular i.e.
∀E ∈ Σ, µ(E) = inf{µ(V ) : V open, E⊆V }

(iii) µ is inner regular, i.e.

∀E ∈ Σ, µ(E) = sup{µ(K) : K⊆E compact}
CHAPTER 1. COMPLEX MEASURES 6

(iv) µ is complete.
For a locally compact Haudorff space X, f ∈ C0 (X) if for all  > 0 there is a compact set K such that
supX\K |f (x)| < . Note that Cc (X) is dense in C0 (X).
We shall look at bounded (not positive) linear functionals on C0 (X). If µ is a complex Borel measure,
we saw
dµ = hd|µ|
with |h| ≡ 1, h Borel, h ∈ L1 (µ). Z Z
f dµ := f hd|µ|;
Z
χE d(µ + λ) = (µ + λ)(E) = µ(E) + λ(E)
X
where µ, λ are two bounded complex measures. So for f bounded and measurable,
Z Z Z
f d(µ + λ) = f dµ = f dλ
X X X

A complex measure µ is regular if |µ| is, meaning |µ| satisfies (ii) and (iii).
Theorem 1.12 (Riesz Representation). X locally compact Hausdorff. The any bounded linear functional ,
Φ on C0 (X) is represented by a unique, complex, Borel measure µ. That is
Z
Φ(f ) = f dµ.
X

Moreover, ||Φ|| = |µ|(X)(< ∞).

Proof. (sketch:)
Uniqueness:
Suppose µ is complex regular Borel measure. For all f ∈ C0 (X) write µ = µ1 − µ2 so that
Z
f dµ = 0.

Want to see that µ = 0. We know that dµ = hd|µ| and |h| ≡ 1, so for all sequences {fn } in C0 (X), we have
Z Z
|µ|(X) = (h̄ − fn )hd|µ| ≤ |h̄ − fn |d|µ|
X X

1
since, Cc (X) is dense in L (|µ|) we can choose our sequence so that the right hand side converges to 0 in n.
This means that the absolute value of the left hand side is less than  for all  > 0 and hence |µ|(X) = 0 so
µ = 0.
Existence:
Let Φ be a bounded linear functional on C0 (X) with ||Φ|| = 1. Construct a positive linear functional ,
Λ, on Cc (X) such that,
|Φ(f )| ≤ Λ|f | ≤ ||f ||.
Suppose we have constructed Λ, then by Riesz Representation theorem for positive measures, there is a
regular Borel measure λ (λ(X) < ∞)

λ(X) = sup{Λf : 0 ≤ f ≤ 1, f ∈ Cc (X)}

R
It follows that λ(X) ≤ 1, also |Φ(f
R )| ≤ Λ|f | = X |f |dλ implying that |Φ(f )| ≤ ||f ||L1 (λ) . So Φ is a linear
functional on Cc (X) such that if |f |dλ ≤ 1 then |Φ(f )| ≤ 1.
We can extend Φ to a linear functional on L1R(λ), so that it has norm 1. Since (L1 )∗ = L∞ , there is a
Borel function, g, with |g| ≤ 1 such that Φ(f ) = X f gdλ for all f ∈ Cc (X), dµ = gdλ.
Chapter 2

Differentiation

Many important fact about derivatives of integrals and vice versa are obtained simply by examining deriva-
tives of measures and their associated maximal functions. For this, the Raydon-Nikodym theorem and
Lebesgue decomposition will be used to our advantage.

2.1 Derivatives of (Lebesgue) measures

Throughout, m denotes Lebesgue measure, and µ a complex Borel measure on Rk .
Theorem 2.1. For a complex Borel measure µ on R and the function
f (x) = µ(−∞, x),
if x ∈ R1 , A ∈ C then the following are equivalent
(a) f is differentiable at x with value A.
(b) For each  > 0 there is a δ > 0 such that
 
 µ(I) 
 m(I) − A < 
 

for all open intervals I containing x with length < δ.

Proof. Easy, considering (b) as the definition of derivative.
The following definition is motivated through the previous theorem expecting to extend this idea to Rk .
Definition 2.2. The symmetric derivative of µ at x ∈ Rk is
 
dµ µ(Br (x))
(Dµ)(x) = := lim .
dm r→0 m(Br (x))
This will be examined by means of the maximal function M µ of µ, defined by
(M µ)(x) := sup (Qr |µ|)(x)
0<r<∞

µ(Br (x))
where we have shortened our expression by writting (Qr µ)(x) as the quotient m(Br (x)) . Notice, when µ is a
positive measure, that the total variation in this definition is not necessary.
The maximal function M µ is lower semi-continuous and hence measurable. See [2] page 136 for proof.
We seek to prove the Maximal theorem whose proof is simplified via the following lemma:
Lemma 2.3. If W = ∪N i=1 Bri (xi ) then there is a subset S⊆{1, . . . , N } such that
(i) The balls Bi , Bj are mutually disjoint for all i 6= j ∈ S,
(ii) W ⊂ ∪i∈S B3rP i
(xi ), and
(iii) M (W ) ≤ 3k i∈S m(Bri (xi ))

7
CHAPTER 2. DIFFERENTIATION 8

Now,
Theorem 2.4. [Maximal theorem] If µ is a complex Borel measure on Rk and λ > 0 ∈ R, then

m{M µ > λ} ≤ 3k ||µ||/λ.

Proof. Fix µ, λ. For any compact subset K of Eλ = {M µ > λ}, each x ∈ K is the center of an open ball
B with |µ|(B) > λm(B). Some finite collection {B1 , . . . , BN } of the balls covers K and the previous lemma
implies
XN N
X
m(K) ≤ 3k m(Bi ) ≤ 3k λ−1 |µ|(Bi ) ≤ 3k ||µ||/λ.
1 1

The last inequality follows from disjointness of the collection provided by the lemma and our statement
follows by taking the supremum over all compact subsets of Eλ .

Weak L1
For any L1 (Rk ) function f and any λ > 0 we have
Z Z
λm{|f | > λ} ≤ |f |dm ≤ |f |dm = ||f ||1 . (2.1)
|f |>λ Rk

Accordingly, we define any measurable f such that λm{|f | > λ} is a bounded function of λ on (0, ∞) to be
weak L1 .
Weak L1 contains regular L1 and is strictly larger (consider 1/x on (0, ∞)).
To any f ∈ L1 (R), associate the maximal function M f : Rk → [0, ∞] by
Z
1
M f (x) = sup |f |dm.
0<r<∞ m(Br ) Br (x)

If we associate a measure µ given by dµ = f dm, this definition agrees with the perviously defined M µ.
Theorem 2.4 ensures the operator M sends L1 to weak L1 with the bound (3k ) depending only on Rk and
not λ.
That is; for all f ∈ L1 (Rk ), and λ > 0

m{M f > λ} ≤ 3k λ−1 ||f ||1 .

Lebesgue points
If f ∈ L1 (Rk ), then x ∈ Rk is called a Lebesgue point of f if
Z
1
lim |f (y) − f (x)|dm(y) = 0.
r→0 m(Br ) B (x)
r

This holds, for example, if f is continuous at x. In general, this means that the averages of |f − f (x)| are
small on small balls about x. That is Lebesgue points are points in the domain where the function does not
oscillate too much on average.
It is not obvious that all L1 functions have Lebesgue points, however the following theorem shows that
almost every point of an L1 function is a Lebesgue point.
Theorem 2.5. If f ∈ L1 (Rk ), then almost every x ∈ Rk is a Lebesgue point of f .

Proof. Define, for x ∈ Rk , r > 0,

Z
1
(Tr f )(x) := |f − f (x)|dm
m(Br ) Br (x)
CHAPTER 2. DIFFERENTIATION 9

and
(T f )(x) = lim sup (Tr f )(x)
r→0

Our claim is proven if T f = 0, a.e[m].

Pick y > 0 and let n be a positive integer. Since Cc (Rk ) is dense in L1 (Rk ), there is a g ∈ C c (Rk ) such
that ||f − g||1 < 1/n. Set h := f − g. Since g is continuous T g = 0 because every point is a Lebesgue point
of g.
Now, Z
1
(Tr h)(x) ≤ |h|dm + |h(x)|
m(Br ) Br (x)
so we have,
T h ≤ M h + |h|
and since Tr f ≤ Tr g + Tr h, it follows that
Tf ≤ M h + |h|.
Thus,
{T f > 2y}⊂{M h > y} ∪ {|h| > y} =: E(y, n).
Since h ∈ L1 with 1-norm < 1/n then Theorem 2.4 and equation (2.1) show that
m(E(y, n)) ≤ (3k + 1)/yn.
The left hand side of this inequality, is independent of n so that
{T f > 2y}⊂ ∩∞
n=1 E(y, n).

This intersection has Lebesgue measure 0 so that {T f > 2y} is a subset of a set of measure zero. Since
Lebesgue measure is complete, {T f > 2y} is Lebesgue measurable or measure 0. This holds for arbitrary
positive y, implying that T f = 0, a.e.[m].

This theorem yields information about

(a) differentiation of absolutely continuous measures,

(b) differentiation using sets other than balls,
(c) differentiation of indefinite integrals in R1 , and
(d) the metric density of measurable sets
which we shall now indulge.
Theorem 2.6. If µ is a complex Borel measure on Rk , µ  m and f is the Raydon-Nikodym derivative of
µ with respect to m. Then Dµ = f a.e.[m], and
Z
µ(E) = (Dµ)dm
E

for all Borel sets E⊂Rk .

In other words, the Raydon-Nikodym derivative can also be obtained as a limit of quotients Qr µ.
R
Proof. Theorem 1.5 implies µ(E) = E f dm. If x is a Lebesgue point of f then
Z
1 µ(Br (x))
f (x) = lim f dm = lim = Dµ(x)
r→0 m(Br ) B (x) r→0 m(Br (x))
r

which is our result.

CHAPTER 2. DIFFERENTIATION 10

Nicely shrinking sets

Let {Ei } be a family of Borel subsets of Rk is said to shrink nicely to a point x ∈ Rk if there is an
α > 0 such that there is a sequence of balls of {Bri (x)} with limn→∞ ri = 0 with each Ei ⊆Bri (x) so that
m(Ei ) ≥ αm(Bri (x)).
Notice that, x is not required to lie in any of the Ei of even its closure.
Theorem 2.7. for all x ∈ Rk , let {Ei (x)} be a family of sets which shrinks nicely to x. If f ∈ L1 (Rk ), then
Z
1
f (x) = lim f dm
i→∞ m(Ei (x)) E (x)
i

at every Lebesgue point of f (i.e. almost everywhere wrt m).

Rx
Theorem 2.8. [Fundamental theorem of calculus for L1 ] Let f ∈ L1 (R1 ) and F (x) = −∞
f dm. Then
F 0 (x) = f (x) for all Lebesgue points (i.e. a.e.[m]).

Proof. Let δi > 0 converging to 0. The intervals Ei = [x, x + δi ] shrinks nicely to x − δi . By Theorem 2.7,
f (x) = limi→∞ m(E1i (x)) Ei f dm at every Lebesgue point; hence
R

Z
F (x + δi ) − F (x) = f dm
Ei (x)

and m(Ei (x)) = δi so that

F (x + δi ) − F (x)
lim = f (x)
δi →0 δi
implying F (x) has right derivative at each Lebesgue point of x. Proof for left derivative is similar.

Metric density
If E is a Lebesgue measurable subset of Rk , the metric density of E at x is

m(E ∩ Br (x))
lim .
r→0 m(Br (x))

Example 2.9. If E = [0, ∞), the metric density at 0 is 1/2, it is 1 everywhere else in E and 0 outside of E.

When f = χE , Theorem 2.7 gives

m(E ∩ Ei (x))
Z
1
χE (x) = f (x) = lim f dm = lim .
i→∞ m(Ei (x)) Ei (x) i→∞ m(Ei (x))

If the Ei (x)0 s are B1/i (x), the metric density of E is 1 at almost every point in E and 0 on the complement.

Corollary 2.10. This reveals, the property that for all  > 0, there is no E⊆R1 , such that

m(E ∩ I)
< < 1 − , ∀ segments I.
m(I)

Theorem 2.11. For all x ∈ Rk and {Ei (x)} shrinking nicely to x, if µ is a complex Borel measure mutually
singular to m then
µ(Ei (x))
lim =0
i→∞ m(Ei (x))

almost everywhere.
CHAPTER 2. DIFFERENTIATION 11

Proof. By the Jordan decomposition theorem, it sufficed to show for positive measures, µ. Having
α(x)µ(Ei (x)) µ(Ei (x)) µ(Bri )
≤ ≤
m(Ei (x) m(Br1 (x)) m(Bri )
meaning this result is a consequence of the fact that (Dµ)(x) = 0 a.e.[m].
upper derivative
" #
D̄µ(x) lim sup Qr µ(x)
n→∞ 0≤r≤1/n

is a Borel function since it decreases in and is lower semi-continuous.

Now, for λ,  > 0, since µ ⊥ m, µ is concentrated on a set of Lebesque measure 0. Regularity of µ implies
the existence of a compact K such that m(K) = 0 and µ(K) = ||µ|| − .
Define µ1 (E) := µ(K ∩ E) and µ2 := µ − µ1 . The ||µ2 || <  and for all x 6∈ K
D̄µ(x) = D̄µ2 (x) ≤ M µ2 (x)
so that {D̄µ > λ}⊂K ∪ {M µ2 > λ}. Now, Theorem 2.4 implies that
m{D̄µ > λ} ≤ m(K) + m{M µ2 > λ} ≤ 3k λ−1 ||µ2 || = 3k /λ
so D̄µ = 0 a.e. which means Dµ = 0 a.e.
This theorem along with Theorem 2.7 can be combined to form the following:
Theorem 2.12. For x ∈ Rk and {Ei (x)} shrinking nicely to x. If µ is a complex Borel measure on Rk
having Lebesgue decomposition
µ = f dm + µs
with respect to m. Then
µ(Ei (x))
lim = f (x), a.e.[m]
i→∞ m(Ei (x))
In particular, µ ⊥ m iff Dµ(x) = 0, a.e.[m].
In contrast to Theorem 2.11, we have
Theorem 2.13. Any positive Borel measure µ on Rk satisfying that µ ⊥ m, then
Dµ(x) = ∞, a.e.[µ]
Proof. There is a Borel set S with m(S) = 0 = µ(Rk − S) and there are open sets Vj containing S with
m(Vj ) < 1/j for each j ≥ 1.
Let N > 0 and define En to be the set of x ∈ S such that there is a sequence ri (x) converging to 0 and
µ(Bri (x)) < N · m(Bri (x)). Then, Dµ = ∞, ∀x ∈ S − ∪N EN .
Fixing N and j, each x ∈ EN is in the center of a ball Bx containing Vj such that the condition for the
set EN holds. Let |bex be the ball of radius 1/3 the radius of Bx . Now,
Vj ⊃ Wj,N := ∪x∈EN βx ⊃ EN
and µ(Wj,N ) < 3k N/j because, for some compact K⊂Wj,N the βi ’s admit a finite sub cover of K. The
covering Lemma 2.3 now ensures there is a finite collection F ⊂EN with {βx : x ∈ F } are al disjoint and
K⊂ ∪x∈F βx . This implies
X X
µ(K) ≤ µ(Bx ) < N m(Bx )
x∈F x∈F
X
k
= 3 N m(βx ) ≤ 3k N m(Vj ) < 3k N/j.
x∈F

Since K was an arbitrarily chosen compact subset of Wj,N , this golds for all of them and regularity of Wj,N
implies our claim.
To complete proof, set ΩN := ∩j Wj,N . Then EN ⊂ΩN which is a Gδ -set (= countable intersection of
open sets). µ(ΩN ) = 0. If x ∈ S − ∪N ΩN then Dµ(x) = ∞.
CHAPTER 2. DIFFERENTIATION 12

2.2 Fundamental Theorem of Calculus

We have already shown the first part of the fundamental theorem in Theorem 2.8. The second part looks
like Z
f (x) − f (a) = f 0 (t)dt (2.2)
[a,x]

The usual version of this theorem assumes f is differentiable and f 0 is continuous (this has an easy proof).
Before, knowing exactly how to weaken these conditions consider the following examples where this result
fails.

(a) f (x) = x2 sin(x−2 ) on R∗ and 0 at the origin.f is differentiable everywhere but [0,1] |f 0 (t)|dt = ∞ so
R

f 0 is not L1 .
If the integral in (2.2) interpreted, with [0, 1] instead of [a, b], as the limit, as  → 0, of the integrals
over [, 1] then (2.2) still holds.
(b) If f is continuous on [a, b], differentiable almost everywhere and f 0 ∈ L1 . Even these assumptions do
not imply (2.2). Consider the Cantor staircase function.

2.3 Absolutely continuous functions

A function f : I = [a, b] → C is absolutely continuous (AC) if for all  > 0 there Pn exists a δ > 0 such
that
Pn if (α 1 , β1 , ), (α2 , β2 ), . . . , (αn , βn ) is a collection of disjoint intervals in I with i=1 (βi − αi ) < δ then
i=1 |f (βi ) − f (αi )| < .
Notice that AC implies continuity and even uniform continuity on I so this condition is stronger.
Theorem 2.14. I = [a, b], f : I → R continuous and non-decreasing then the following are equivalent:
(a) f is AC on I,
(b) f maps sets of measure 0 to sets of measure 0,
(c) f is differentiable a.e on I, f 0 ∈ L1 (I) and
Z x
f (x) − f (a) = f 0 (t)dt, a ≤ x ≤ b
a

A good example of not absolutely continuous is the Cantor staircase.

Proof. (a) → (b) → (c) → (a)

Let Σ be a Lebesgue sigma algebra on R and assume f is AC. Let E⊂I be of measure 0. We want to
show that f (E) ∈ Σ with m(f (E)) = 0. Let a, b 6∈ E,  > 0 and δ correspond to  as in the definition of AC.
There is an open V such that m(V ) < δ, E⊆V ⊆I. Let (αiP , βi )i∈J be a disjoint collection of open intervals
whose union is V . By Absolute continuity of f this implies i∈J (f (βi ) − f (αi )) ≤ . This proves (a) implies
(b) since f (E)⊆ ∪i∈J (f (αi ), f (βi )).
For (b) implies (c); Let g(x) = x + f (x). Then g satisfies (b) if f satisfies (b). Indeed if |f (β) − f (α)| ≤ ν 0
with |β − α| ≤ ν then |g(β) − g(α)| ≤ ν + ν 0 so that g satisfies (b) when f does. For E⊂I in Σ, then
Theorem 2.20 of [2] ensures we may write E = E1 ∪ E0 where m(E0 ) = 0 and E1 is an Fσ set (that is a
countable union of compact sets). Similarly, since g is continuous g(E1 ) is also an Fσ set, m(g(E0 )) = 0
hence, g(E) = g(E0 ) ∪ g(E1 ) ∈ Σ. Define µ(E) := m(g(E)) and, since g has be chosen to be injective,
disjoint sets have disjoint images implying the countable additivity of µ. We now have that µ is a positive,
bounded measure on Σ and that µ  m (since g satisfies (b)). By Theorem 1.5, there exists an h ∈ L1 (m)
such that dµ = hdm.
Let E = [a, x] so g(E) = [g(a), g(x)] and we have
Z Z x
g(x) − g(a) = m(g(E)) = µ(E) = hdm = h(t)dt.
E a
CHAPTER 2. DIFFERENTIATION 13

Now, Z x Z x
f (x) − f (a) = g(x) − x − g(a) + a = h(t)dt − (x − a) = [h(t) − 1]dt
a a

and Theorem ?? implies f 0 (x) = h(x) − 1 a.e.[m].

Finally, assuming (c), define a measure µ by dµ = f 0 dm and since µ  m, Theorem ?? shows that for
any  > 0, there is a δ > 0 such that whenever E is a union of disjoint segments with total length less than
δ then |µ|(E) < . Since f (y) − f (x) = µ([x, y]) the absolute continuity of f follows.

Theorem 2.15. If f : I → R is absolutely continuous on I = [a, b] and

N
X
F (x) := sup |f (ti ) − f (ti−1 )|
i=1

with supremum taken over all N and a = t0 < t1 < · · · < tN = x. Then F, F + f and F − f are absolutely
continuous and non decreasing on I.
The function F is the total variation function of f . If f is any (complex) function, AC or not, and
F (b) < ∞, then f is said to have bounded variation and F (b) is called the total variation of f on I.
PN
Proof. For x < y ≤ b we have F (y) ≥ |f (y) − f (x)| + i=1 |f (ti ) − f (ti−1 )|. Taking the supremum gives
F (y) ≥ |f (y) − f (x)| + F (x) which is greater than both f (y) − f (x) + F (x) and f (x) − f (y) + F (x) and
implies all three of these are nondecreasing.
It remains to show that F is AC. For (α, β)⊂I,
N
X
F (β) − F (α) = sup |f (ti ) − f (ti−1 )| ?
i=1

PN
with i=1 (ti − ti−1 )0
P= β − α. Let  > 0 and δ be as
Pin the definition of AC for f . Let (αi , βi )⊂I be disjoint
segments such that j (βj − αj ) < δ and estimate j (F (βj ) − F (αj )) using ?.

Theorem 2.16. If f is any complex valued AC function on I = [a, b], then f is differentiable a.e. on I,
f 0 ∈ L1 (m) and Z x
f (x) − f (a) = f 0 (t)dt.
a

Proof. It suffices to show for real valued f . Let F be the total variation of f and define
1 1
f1 = (F + f ), f2 = (F − f ).
2 2
By previous Theorem, f1 , f2 are AC and nondecreasing on I. By (a) implies (c) in Theorem 2.14, this result
holds for each fi and hence also hold for f = f1 − f2 .

Theorem 2.17. If f : [a, b] → R is differentiable on [a, b] and f 0 ∈ L1 ([a, b]) then

Z x
f (x) − f (a) = f 0 (t)dt.
a

Proof. Let  > 0; By Vitali-Caratheodory (Theorem 2.25 in [2]), there exist functions u, v bounded above
and below respectively such that u ≤ f 0 ≤ v, (v − u) <  with u, v upper and lower semi-continuous
R

respectively.
Note: χclosed is upper s.c. and χopen is lower.
Let g = v be our lower semicontinuous upper bound for f . We can also take g > f 0 by adding a small
constant and Z b Z b
g(t)dt < f 0 (t)dt + 
a a
CHAPTER 2. DIFFERENTIATION 14

Let η > 0; and define Z x

Fη (x) = g(t)dt = f (x) + f (a) + η(x − a).
a
For each x ∈ [a, b], ∃, δx > 0 : ∀t ∈ (x, x + δx ) we have

f (t) − f (x)
g(t) > f 0 (x) and < f 0 (x) + η
t−x
where we have used the lower semi-continuity of g (since g −1 (f 0 (x), ∞) is open).
For any such t,
Z t
Fη (t) − Fη (x) = g(s)ds − [f (t) − f (x)] + η(t − x)
x
> (t − x)f 0 (x) − (t − x)(f 0 (x) + η) + η(t − x) = 0.

Since Fη (a) = 0 and Fη is continuous (exercise!) there is a last point x in [a, b] for which Fη (x) = 0. If
x < b then Fη (t) > 0 for t ∈ (x, b]. Either way, Fη (b) ≥ 0 and this holds for all η > 0 implying that
Z b Z b
f (b) − f (a) ≤ g(t)dt ≤ f 0 (t)dt + .
a a

This hold for all positive  meaning

Z b
f (b) − f (a) ≤ f 0 (t)dt.
a
If f satisfies the hypothesis of this theorem, so does −f which yields the reverse inequality.

In summary, either of the following two conditions imply equation 2.2

(i) f : [a, b] → R differentiable everywhere with f 0 ∈ L1
(ii)f is absolutely continuous.

2.4 Differentiable transformations

Definition 2.18. Let T : V ⊂Rk → Rk (V open) if there is a linear operator A : Rk → Rk such that

T (x + h) − T (x) − Ah
lim =0
h→0 |h|

then T is differentiable at x and T 0 (x) = A is the derivative of T at x.

If A : Rk → Rk is linear, then m(A(E))∆(A)·m(E). It is reasonable to conjecture that if T is differentiable
at x, T 0 (x) = A, E a measurable set ’close’ to x , then

m(T (E))
∼ ∆(T 0 (x))
m(E)

where T 0 (x) is the Jacobian of T at x

Notation: ∆(T 0 (x)) := |JT (x)|
Lemma 2.19. Let S = {|x| = 1, x ∈ Rk } = ∂(B1 (0) and F : B̄ → Rk be continuous, 0 <  < 1 such that
for every x ∈ S, |F (x) − x| <  then F (B) ⊃ B1− (0).

Proof. For a contradiction suppose there is an a ∈ B1− (0) such that a 6∈ F (B). That is a 6= F (x) for all
a−F (x)
x ∈ B̄. Let G(x) = |a−F (x)| if x ∈ S, then x · (a − F (x)) = x · ax (x − F (x)) − 1 < |a| +  − 1 < 0. This
means that xG(x) < 0so x 6= G(x). If x ∈ B, then x 6= G(x), since |G(x)| = 1 implying G(x) ∈ S. G is a
continuous map from B̄ to B̄ with no fixed points. A contradiction.
CHAPTER 2. DIFFERENTIATION 15

Theorem 2.20. If V is open in Rk , T : V → Rk is continuous and differentiable at some x ∈ V , then

m(T (Br (x)))

lim = ∆(T (x))
r→0 m(Br (x))

Proof. WLOG, let x = T (x) = 0 and A = T 0 (0)

Case 1: If A is injective (implying surjective since range has same dimension) Let F (x) = A−1 T (x), then
F (0) = A−1 T 0 (0) = A−1 A = Id
0

m(F (Br (0)))

lim =1
r→0 m(Br (0)
with T = AF, m(T (B)) = ∆(A)m(F (B)). Let  > 0, F (0) = 0, F 0 (0) = Id then there is a δ > 0 such that
0 < |x| < δ, we have
|F (x) − x| < |x| ?
If 0 < r < δ, then we claim that

B(0, (1 − )r)⊆1 F (B(0, r))⊆2 B(0, (1 + )r)

For 1, apply lemma with B(0, 1) replaced with B(0, r) and 2, follows from ?. Finally our claim about the
limit of measures follows from these inclusions.
Case 2: given  > 0, there is a η > 0 such that if Eη = {x ∈ Rk : d(x, A(B1 (0)) < 1} then mRk (Eη ) < .
a = T 0 (0) implies there is δ > 0 such that ∀|x| < δ, |T (x) − Ax| ≤ η|x|.
If r < δ, then T (Br (0)) = r · Eη and m(T (Br (0))) <  · rk which shows

m(T (Br (0)))

lim = 0.
r→0 m(Br (0))

T (y)−T (x)
Lemma 2.21. E⊆Rk , m(E) = 0, T : E → Rk and limy→x |y−x| < ∞. Then

m(T (E)) = 0.

Proof. Let p, n be positive integers;

F = Fn,p = {x ∈ E : ∀y ∈ E ∩ B1/p (x), |T (y) − T (x)| ≤ n · |y − x|}

P
the measure of F is zero so can be covered by Bi = Bri (xi ) where xi ∈ Fri < 1/p such that m(Bi ) < .
Let x ∈ F ∩ B, the |x − xi | < ri .

|T (xi ) − T (x)| < n|xi − x| < nri

now T (F ∩ Bi )⊂Bnri (T (xi )) thus T (F )⊂ ∪i Bnri (T (xi )) so
X
m(T (F )) < nk m(Bi ) < nk · 
i

Hence, m(T (F )) = 0 since  was arbitrary.

Now E = supn,p>0 Fn,p ⇒ m(E) = 0.

Theorem 2.22. Let X⊂V ⊂Rk , V open, Y : V → Rk continuous, X Lebesgue measurable, T is injective
and differentiable on X. If m(T (V \X)) = 0, then if Y := T (X), we have, for all f : Rk → [0, ∞) that
Z Z
f dm = (f ◦ T (x)) · |JT (x)|dm(x)
Y X

This is the change of variables formula for T .

CHAPTER 2. DIFFERENTIATION 16

Proof. We prove in three steps

(i) If E⊂V is Lebesgue measurabe then so is T (E).
(ii) For each Lebesgue measurable E,
Z
m(T (E ∩ X)) = χE |JT |dm
X

(iii) For every Lebesgue measurable E,

Z Z
χA dm = (χa ◦ T )|JT |dm
Y X

Indeed, if E0 ⊂V is Lebesgue measurable then m(T (E0 − X)) = 0 by assumption and m(T (E0 ∩ X)) = 0 by
Chapter 3

Limits of eigen functions on a torus

On any Riemannian manifold (M, g), the Laplacian is given by ∆f = div(grad(f )).
L2 (M has a basis of eigen functions satisfying
∆φj (x) + λj φj (x) = 0
and the same holds for bounded domains with nice boundry and measurable boundry conditions (Dirichlet
- φ|∂B = 0 and Neumann - ∂n φ|∂B = 0).
∂2 ∂2 ∂2
On the n−torus, Tn = Rn /2πZn we have ∆ = ∂x 2 + ∂x2 + ∂x2 .
1 2 3

Example 3.1. n = 1 Have the second order differential equation

∂2
φ(x) + λφ(x) = 0
∂x2
λ0 = 0; φ0 = 1. If M is connected, then the next eigen vectors are
φ1n (x) = sin(nx), φ2n (x) = cos(nx)
with λ1n = λ2n = n2 .
The Rayleigh quotient
|∆φ|2
R
λk = inf sup R
Vk φ∈Vk |φ|2
has something to do with the and Vk is something.
On the d torus, we have eigen functions ei(n1 x1 +···+nd xd ) with eigenvalues λ~n = n21 + · · · + n2d .√ The
multiplicities of such eigen values are equal to the number of points on the integer lattice with norm λ.
In the case d = 2; mult(λ) < C λ for all  > 0. This is also related to the number of divisors m of λ
such that m ≡ 1 mod 4. For example, when λ = 5 we have the 8 solutions 5 = (±1)2 + (±2)2 corresponding
to the Gaussian primes. The multiplicity in 2-dimensions is unbounded. The multiplicity of λ can grow as
λd/2−1 . If d ≥ 5, 1/c ≤ mult(λ)
λd/2−1
≤ C as λ → ∞.
So, eigen vectors on the d dimensional torus are φλ = ζ∈Zd ∩S√ (0) cζ eihx,ζi .
P
λ
What are the weak * limits of φλ |2 as λ → ∞? 2
R |φλ (x)| 2dx is the probability density for the particle φλ
and the probability that φλ is in E is given by E |phiλ (x)| dx.
√
Example 3.2. In the case n = 1; φk (x) = 1/ π sin(kx);
φ2k (x) = 1/π sin2 (kx) = 1/2π(1 − cos(2kx))
 
Z 2π Z 2π Z 2π
f (x)φ2k (x)dx = 1/2π 
 
 f (x)dx − f (x) cos(2kx)dx.
0 0 0
| {z }
→0
2
That is, 1/π sin (kx) becomes uniformly distributed as k → ∞

17
CHAPTER 3. LIMITS OF EIGEN FUNCTIONS ON A TORUS 18

Theorem 3.3 (A. Shnirelman, Zelditch, Colin de Verdiere). If M is a compact manfild with ergodic geodesic
flow (negative curvature-ish), for example the torus with > 2 holes, then |φλ (x)|2 dx → dx as λ → ∞ for
almost all φλ (x).
R 2π
Define, for a functions f the Fourier coefficients fˆ(n) := 1/2π 0 f (x)e−inx dx

Lemma 3.4. [Riemann-Lebesgue] If f ∈ L1 (T) then |fˆ(n)| → 0 as |n| → ∞.

Proof. For all  > 0 there is a trig polynomial P (x) such that ||f − P ||1 <  (Stone-Weirstrauss). Fix  > 0
ak eikx be as described of degree N . Let n > N , then fˆ(n) = f[
P
and let p(x) = − p(n) since p̂(n) = 0.
Now, for n > N we have
Z 2π X N
! N Z 2π
X
1/2π ak e ikx
e−inx dx = 1/2π ei(k−n)x dx = 0
0 −N −N 0

implying that |fˆ(n)| = |(f\

− p)(n)| ≤ ||f − p||1 < 

The same result holds for f ∈ L1 (R) with the idea of approximating f by g ∈ Cc∞ (R), then fˆ(t) =
R∞
−∞
f (x)dx = (f\
− g)(t) + ĝ(t). Now, ||f − g||1 <  implying |(f\
− g)(t)| <  and it is enough to show that
|ĝ(t)| < 

Example 3.5. in the case d = 2, fn2 (x, y) = 4 sin2 (nx) cos2 (y)dxdy → (1 + cos(2y))dxdy in n since
2 sin2 (nx)dx → dx and cos1 (y) = 1/2(1 + cos(2y))

If we have ϕλ (x) = |ζ|2 =λ aζ eihx,ζi , then

P

X X
ϕλ (x)ϕ̄λ (x) = aζ eihx,ζi · āη eihx,ηi
|ζ|2 =λ |η|2 =λ
X
= aζ āη eihx,ζ−ηi
|ζ|2 =|η|2 =λ
 
X X X
= |aζ |2 +  aζ āη 
ζ 06=τ ∈Zd ζ−η=τ
CHAPTER 3. LIMITS OF EIGEN FUNCTIONS ON A TORUS 19
Chapter 4

Fourier Transformations

For x ∈ Rn , k ∈ (Rn )∗ , f ∈ L1 (Rn ); fˆ(k) = Rn e−2πihx,ki f (x)dx. The map f 7→ fˆ is linear in f .

R

If τh f (x) := f (x − h); then

Z Z
τh f (k) =
d f (x − h)e −2πihx,ki
dx = f (y)e−2πihy+h,ki dy = e−2πihh,ki fˆ(k)
Rn Rn

and if (δλ f )(x) = f (x/λ);

Z Z Z
δd
λ f (k) = f (x/λ)e−2πihx,ki dx = f (y)e−2πihλy,ki λn dy = λn f (y)e−2πihy,λki dy = λn fˆ(λk)
Rn Rn Rn

−2πihh,ki ˆ nˆ
In summary, τdk f (k) = e f (k) and δd λ f (k) = λ f (λk).
The Fourier transform is continuous; If ||k1 − k2 || < δ ⇒ ||e2πihx,k1 i − ehx,k2 i || <  so that
Z Z
|fˆ(k1 ) − fˆ(k2 )| ≤ |f (x)| · |e−2πihx,k1 i − e−2πihx,k2 i |dx ≤ /||f ||1 |f (x)|dx = 
Rn Rn

Recall, the Riemann Lebesgue Lemma 3.4, and the convolution of f, g ∈ L1 (Rn ) is given by
Z
(f ∗ g)(x) := f (x − y)g(y)dy
Rn

whose Fourier transform is

Z Z 
∗ g(k)
f[ = e2πihx,ki f (x − y)g(y)dy dx
Rn n
 R 
Z Z 
−2πihk,x−yi
 
= g(y)dy 
 e f (x − y)
 dy
R n n
| R {z }
=fˆ(k)
Z
= fˆ(k) g(y)e−2πihy,ki dy = fˆ(k)ĝ(k)
Rn
2
Example 4.1 (Gaussians). If λ > 0 and gλ (x) = e−πλ|x| , |x|2 = x21 + . . . x2n is a Gaussian distribution then
2
ĝλ (k) = λ−k/2 e−π|k| /λ

meaning that Fourier transforms send Gaussians to Gaussians.

√
Proof. By the Homogeneity in λ it suffices to prove for λ = 1 (since λ|x|2 = | λx|2 ) and Since g1 (x) =
Qn −πx2
i=1 e
i , it suffices by Fubini, to prove for n = 1.

20
CHAPTER 4. FOURIER TRANSFORMATIONS 21

Z
2
ĝ1 (x) = e2πihx,ki e−πx
Rn
and completing the square we find
Z
2 2
e − π(x2 + 2ikx − k 2 + k 2 ) = e−πk −π(x+ik)
|e {z } dx
R
:=f (k)

As an exercise, we can differentiate under the integral to find

Z Z
∂f (k) 2 ∂ −π(x+ik)2 2 ∞
= −(2πi)(x + ik)e−π(x+ik) dx = i e dx = ie−π(x+ik) −∞ = 0
∂k R R ∂x

2
e−πx = 1 which yields our result.
R
so that f (k) is constant and f (0) = R

We now define the Fourier transformation on L2 (Rn ). Note that on any compact manifold M L2 (M )⊂L1 (M )
so the idea is to approximate L2 by functions in L1 ∩ L2 .
Theorem 4.2 (Plancherel’s Theorem). f ∈ L1 (Rn ) ∩ L2 (Rn ) implies fˆ ∈ L2 (Rn ) and

||fˆ||2 = ||f |||2 .

Also, the map f 7→ fˆ has a unique extension to a continuous linear map on L2 (Rn ) to itself which is an
isometry (meaning the norm condition still holds). If f, g ∈ L2 (Rn ) then we have Parsevel’s formula:
Z n Z n
¯
hf, gi = f¯g = fˆ(k)ĝ(k)dk = hfˆ, ĝi ?
R R

Proof. If f ∈ L1 ∩ L2 then certainly, by Theorem 3.4 fˆ(k) is bounded hence

Z
2
|fˆ(k)|2 e−(π|k|) dk ? ?.
Rn

2
Since f ∈ L1 it follows by Fubini that f¯(x)f (y)e−π|k| ∈ L1 (R3n ) Now,
Z Z
π(x−y)2
¯ 2πihk,x−yi −π|k|2
f (x)f (y)e e dxdydk = f¯(x)f (y)−n/2 e−  dxdy
R3n R3n

2
R exp [− π(x−y) ]
We claim that Rn n/2

f (y)dy → f (x) as x → 0. This is proven using Theorem 2.16 from [1].
Equation ?? now converges to Rn |f (x)|2 dx implying uniform boundedness and ||fˆ||2 = ||f ||2 by mono-
R

tone convergence.
Now consider a function f ∈ L2 \(L1 ∩ L2 ). Since L1 ∩ L2 is dense in L2 we can obtain a sequence fj here
which converges strongly to f . As we have shown, ||fi − fj ||2 = ||fˆi − fˆj ||2 making the Fourier transforms a
Cauchy sequence in L2 which converges to some fˆ ∈ L2 . Also, using what we have already shown, we have

||fˆ||2 = lim ||fˆj ||2 = lim ||fj ||2 = ||f ||2

j→∞ j→∞

followed by continuity and linearity which is left as an exercise.

Finally, we get ? from ?? by the polarization identity
1
||f + g||22 − i||f + ig||22 − (1 − i)||f ||22 − (1 − i)||g||22


hf, gi =
2
along with our norm condition.
CHAPTER 4. FOURIER TRANSFORMATIONS 22

4.1 L2 Fourier Transformations

If f ∈ L2 (Rn ), the Plancherel’s Theorem guarantees that its Fourier transform is as well. This is remarkable
because it is independent of approximation.
2
For example, the sequence fˆj (k) = |x|<j e−2πihx,ki f (x)dx along with the sequence ĥj (k) = Rn cos(|x|2 /j)e−|x| /j e−2πihk,xi f
R R

both satisfy that there is an fˆ ∈ L2 such that

||fˆj − fˆ||2 = ||ĥj − fˆ||2 → 0

and also ||ĥj − fˆj ||2 → 0.

We remark that the map f 7→ fˆ is not only an isometry, but a unitary transformation. In particular,
invertible.

4.2 Inversion formula

Theorem 4.3 (Inversion). If f ∈ L2 (Rn ) and f ∨ (x) := fˆ(−x) then f = fˆ∨ and this is well defined by
Theorem 4.2

Proof. For f ∈ L2 we have

Z Z
ĝλ (y − x)f (y)dy = gλ (k)fˆ(k)e2πihx,ki dk ?
Rn Rn

2 2
where gλ (k) = e−λπ|k| implying ĝ(y − x) = λ−n/2 e−π|x−y| /λ .
To verify ?, we approximate f by fj ∈ L1 ∩ L2 . Now, since fj → f ∈ L2 , Plancherel’s Theorem says that
fˆj → fˆ ∈ L2 implying ? holds in general. As λto0, the left hand side approaches f (x) by Theorem 2.16 of
[1].
Now, the dominated convergence Theorem implies that gλ fˆ → fˆ so by 4.2 we have (gλ fˆ)∨ → fˆ∨ and
this convergence gives our result.

4.3 Interpolation Theorems in Lp

Proposition 4.4. Let 0 < p < q < r ≤ ∞. Then Lp ∩ Lr ⊆Lq and

||f ||q ≤ ||f ||λp ||f ||1−λ

r

where 1/q = λ/p + (1 − λ)/r.

This proposition was proven on a homework last semester. Taking logs gives

log ||f ||q ≤ λ log ||f ||p + (1 − λ) log ||f ||r

This is a concavity/convexity statement about ||f ||p .

Proposition 4.5 (Three lines lemma - Hadamard?). If ϕ is a bounded continuous function on 0 ≤ Re(z) ≤ 1,
holomorphic in 0 < Re(z) < 1. Suppose |ϕ(z)| ≤ M0 on Re(z) = 0 and |ϕ(z)| ≤ M1 on Re(z) = 1. Then

|ϕ(z)| ≤ M01−t M1t

on Re(z) = t for 0 < t < 1

The proof of this uses the maximum modulus principle.
CHAPTER 4. FOURIER TRANSFORMATIONS 23

Theorem 4.6 (Riesz-Thorin interpolation). Let (X, M, µ), (Y, N, ν) semi-finite measure spaces (∀E : µ(E) =
∞, ∃F ⊆E such that 0 < µ(F ) < ∞) Let p0 , p1 , q0 , q1 ∈ [1, ∞]; 0 < t < 1 and define

1/pt = (1 − t)/p0 + t/p1

similarly for q.
Let T : X → Y be a linear map such that ||T f ||q0 ≤ M0 ||f ||p0 for all f ∈ Lp0 (X) and ||T f ||q1 ≤ M1 ||f ||p1
for all f ∈ Lp1 (X). Then
||T f ||qt ≤ M01−t M1t ||f ||pt
for all f ∈ Lpt (µ), 0 < t < 1.

Proof. Case 0: p0 = p1 ; then pt = p0 = p1 for each t

1−t 1−t
||T f ||qt ≤ ||T f ||1−t t
q0 ||T f ||q1 ≤ M0 ||f ||1−t t t
p0 M1 ||f ||p1 = M0 M1t ||f ||p0

Assuming now that p0 6= p1 and pt < ∞ for allt ∈ (0, 1). Le tSX , SY be the simple functions on
X and Y . We have that SX ⊆Lp (X < µ) is dense for p 6= ∞. We want to see that if f ∈ SX , then
||T f ||qt ≤ M01−t M1t ||f ||pt . By results about the dual Lqt ,
Z 
||T f ||tqt = sup (T f )gdν : g ∈ SY , ||g||qt0 ≤ 1
Y

It suffices to prove the main lemma: f ∈ SX , ||f ||pt = 1, then

Z
(T f )gdν ≤ M01−t M1t
Y

for all g ∈ SY with ||g||qt0 = 1.

Pm Pn
Let f = j=1 aj χEj ; g = k=1 bk χFk . Write aj = |aj |eiθj , bk = |bk |eiψk . Let αj = 1/pj , βj = 1/qj and

α(z) = (1 − t)αo + zα1 , β(z) = (1 − t)β0 + zβ1 .

Note that, by definition, α(t) = 1/pt , β(t) = 1/qt .

Assuming that β(t) < 1 because the case β(t) = 1 is an exercise.
Z n
m X
X 1−β(z)
ϕ(z) = (T fz )gz ν = |aj |α(z)/α(t) |bj | 1−β(t) Cjk
Y j=1 k=1

where Cjk = eiθj +iψk Y (T χEj )χFk dν

R
Pm α(z) Pn 1−β(z)
Note that fz = j=1 |aj | α(t) eiθj χEj and similar for gz = k=1 |bk | 1−β(t) eiψk χFk .
R
ϕ(z) is an entire holomorphic function in 0 ≤ Re(z) ≤ 1. Now, ϕ(t) = Y (T f )gdν and we want to
apply the three line lemma. It would suffice to show that |vf (z)| ≤ M0 for Re(z) = 0 and |ϕ(z)| ≤ M1 for
Re(z) = 1. We will show the first bound.
Let z = is; α(is) = α0 + is(α1 − α0 ); 1 − β(is) = (1 − β0 ) − is(β1 − β0 ) so that |fis | = |f |α0 /αt and
1−β0 0 0
|gis | = |g| 1−βt = |g|qt /q0 since 1 − βt = 1 − 1/qt = 1/qt0 for each t. Apply Hölder’s inequality:
Z
|ϕ(is)| = (T fis )gis dν ≤ ||T fis ||q0 ||gis ||q00 ≤ M0 · ||fis ||p0 · ||gis ||q00
Y

but,
Z  p0 1/p0
"Z 1/pt #pt /p0
pt /p0 pt
||fis| ||p0 = |f | = |f | = (||f ||pt )pt /p0 = 1
X X

and ||gis ||q00 = 1. If z = is, then |ϕ(is)| ≤ M0 which proves the claim.
CHAPTER 4. FOURIER TRANSFORMATIONS 24

Similarly, if Re(z) = 1,, then |ϕ(1 + is)| ≤ M1 (Exercise). This means that for simple functions f ∈
SX , g ∈ SY ,
||T f ||qt ≤ M01−t M1t ||f ||pt
The rest of the proof, proceeds by approximating arbitrary f ∈ Lpt (X) by simple functions.
Let f ∈ Lpt (X, µ) and fn a sequence of simple functions with |fn | ≤ |f | and fn → f pointwise. Let
E = {x : |f (x)| > 1} and g = f · χE as well as gn = fn χE , h = f − g, hn = fn − gn . Assume p0 < p1 , then
g ∈ Lp0 (X < µ), h ∈ Lp1 . Now ||fn − f ||pt → 0 by Dominated convergence; ||gn − g||p0 → 0; ||hn − h||p1 → 0.
Apply T to the second two convergences to find ||T gn − T g||q0 → 0 and ||T hn − T h||q1 → 0. Passing
to a subsequence where T gn → T g and T hn → T h almost everywhere which implies T fn → T f almost
everywhere. Now, by Fatou’s lemma
||T f ||qt ≤ lim inf ||T fn ||qt ≤ M01−t M1t ||f ||pt
n

A slightly stronger formulation of Riesz-Thorin: Let M (t) be the operator norm from Lpt (µ) to Lqt (ν)
we have M (t) ≤ M01−t M1t . If 0 < s < t < u < 1, t = (1 − τ )s + τ u. Apply Riesz Thorin again to conclude
M (t) ≤ M (s)1−τ M (u)τ
so log M (t) is a convex function on [0, 1].
Theorem 4.7 (Hausdorff-Young on torus (non-sharp)). Suppose 1/p + 1/q = 1, 1 ≤ p ≤ 2. Then fˆ ∈ lq (Zn )
and ||fˆ||q ≤ ||f ||p .

Proof. f ∈ L1 implies ||fˆ||∞ ≤ ||f ||1 and f ∈ L2 implies ||fˆ||2 = ||f ||2 . Apply Riesz-Thorin with T f =
fˆ, p0 = 1, p1 = 2 making q0 = ∞, q1 = 2 and 1/pt = 1−t t 1−t t
p0 + p1 , 1/qt = q0 + q1 .
√
From 2 weeks ago, f ∈ L2 (T ) (fˆ(ζ) 6= 0 ↔ ζ ∈ Z2 ∩ {|ζ| = r > 0}), then ||f ||4 ≤ 4 5||f ||2 since
|||fˆ|2 ||l2 ≤ C · ||f ||2 . If f ∈ L2 (T n ) then |||fˆ|2 ||ln ≤ Cn ||f ||2 . This does not imply ||f ||2 p1 < ∞ for any
p1 > 1.
There is no converse to Hausdorff-Young: an open problem on T 2 ,
Is there a bound
||f ||p
<C<∞
||f ||2
for ∆f + λf = 0 and some p > 4?
Remark 2. If M ult(λ) < C < ∞, then there is a uniform L∞ bound. This is only interesting for R2 /Z2
Question 2: On T n , n ≥ 3 is it true for any pn > 2

||f ||pn
<C<∞
||f ||2

Fourier transform in Lp
Want to mimic the extension of the Fourier transform from L1 ∩ Lp to Lp as already done for p = 2 above.
This will only work for 1 ≤ p ≤ 2 if we want fˆ ∈ Lq . Observe f ∈ L1 ⇒ fˆ ∈ L∞ ; ||fˆ||∞ ≤ ||f ||1 and we
recall that, the map L1 → c0 ; f 7→ fˆ is not onto. We had that f ∈ L2 ⇒ fˆ ∈ L2 ; ||fˆ||2 = ||f ||2 .
To define F.T. on Lp , we can mimic the L2 construction by approximating general f ∈ Lp by a sequence
fj ∈ L1 ∩ Lp . We will need to show that that fˆj is a Cauchy sequence in Lq . To do this we need

||fˆ||q ≤ Cp,q ||f ||p . (4.1)

nˆ 1 p
For which p, q can this work? δ\ λ f (k) = λ f (k). For f ∈ L ∩ L we have
Z 1/p  Z 1/p
p n p
||f (x/λ)||p = ||δλ f ||p = |f (x/λ)| dx = λ |f (y)| dy = ||f ||p λn/p
Rn Rn
CHAPTER 4. FOURIER TRANSFORMATIONS 25

What happens to ||fˆ||q ?

Z 1/q Z 1/q
||δd
λ f ||q = |λ fˆ(λk)|q dk
n
=λ n
|fˆ(z)|q dz/λn = λn(1−1/q) ||fˆ||q
Rn Rn

For 4.1 to hold, we need λn/p = λn (1 − 1/q). This is seen by contradiction since if n/p 6= n(1 − 1/q) then
4.1 can’t hold uniformly for all λ
Thus, we must have that p, q are conjugate exponents.
Remark 3. 4.1 can’t hold for p > 2 By assignment 3. and if 1 < p < 2 then 4.1 holds by Hausdorff-Young
inequality.
Last time, we stated the non-sharp Hausdorff-Young inequality in Rn .
Theorem 4.8 (Hausdorff-Young inequality). For 1 < p < 2, f ∈ Lp ∩ L1 with p0 conjugate to p. Then

||fˆ||p0 ≤ Cpn ||f ||p ?

0
where Cp2 = p1/p p0−1/p where equality holds iff

f (x) = A exp −hx, M xi + hB, xi

with A ∈ C, B ∈ Cn and M a symmetric, real, positive definite n × n.

This theorem allows us to extend FT’s to Lp with 1 < p < 2
0
Corollary 4.9. fˆ can be extended to Lp (Rn ), but it is not onto Lp .
Theorem 4.10. If f ∈ Lp , g ∈ Lq with 1 + 1/r = 1/p + 1/q. Assume that p, q, r ∈ [1, 2]. Then

∗ g(k) = fˆ(k)ĝ(k)
f[

General form of Young’s inequality

Theorem 4.11. Suppose 1 ≤ p, q, r ≤ 2; 1/p + 1/q = 1 + 1/r with f ∈ Lp , g ∈ Lq , then

f ∗ g ∈ Lr and ||f ∗ g||r ≤ ||f ||p ||g||q .

From [1] Theorem 4.2 remark (2), statement with sharp constant.
Theorem 4.12. More generally, if 1/p + 1/q + 1/r = 2 then
Z 
 

 n f (x)(g ∗ h)(x)dx  ≤ Cp,q,r,n ||f ||p ||g||q ||h||r

R

p1/p
with Cp,q,r,n = (Cp Cq Cr )n an Cp2 = p01/p
where 1/p + 1/p0 = 1.

Last semester we sketched a proof of Theorem 2 without sharp constants. We now give a different proof
of the first Theorem using Riesz-Thorin.

Proof. Fix q; let p = 1 so 1/q = 1/r i.e. r = q.

claim: f ∈ L1 , g ∈ Lq ⇒ (f ∗ g) ∈ Lq ; ||f ∗ g||q ≤ ||f ||1 ||g||q .
The proof of this claim makes use of Minkowski’s inequality:
Z Z p 1/p Z
|f (y)|||g y ||p dy = ||g||p ||f ||1
 
||f ∗ g||q = 
 [f (y)g(x − y)dy] dx ≤
R n Rn Rn

where g g (x) := g(x − y); ||g y ||p = ||g||p

If r = ∞
CHAPTER 4. FOURIER TRANSFORMATIONS 26

claim 2: 1/p + 1/q = 1 ⇒ ||f ∗ g||∞ ≤ ||f ||p ||g||q by Hölder. Now, we use Riesz-Thorin:
q fixed, g ∈ Lq fixed, T f = f ∗ g 1/p + 1/q = 1 + 1/r, f ∈ Lp , g ∈ Lq , then f ∈ L1 implies T f ∈ Lq with
q−1
||T f ||q ≤ ||f ||1 ||g||q and if f ∈ L q then T f ∈ L∞ ; ||T f ||∞ ≤ ||f || q−1 ||g||q .
q

1 − 1/q ≤ r ≤ 1; 1/rt = (1−t)q

q−1 + t; s0 = ∞, s1 = 1; 1/st =
1−t
s0 + t/s1 = t/1 = t
1−t
f ∈ L ⇒ ||T f ||rt ≤ M0 M1t ||f ||t
r

In Riesz-Thorin, M0 = M1 = ||g||q ... this shit is gross.

As one application of this dirty shit, suppose f ∈ Lp , g ∈ Lq with 1/p + 1/q = 1 + 1/r and 1 ≤ p, q, r ≤ 2.
Then
\
(f ∗ g)(k) = fˆ(k)ĝ(k)
0 0 0
This is proven by seeing that f ∗ g ∈ Lr so fˆ ∈ Lp , ĝ ∈ Lq ⇒ fˆĝ ∈ Lr , by Holder’s h := f ∗ g ∈ Lr
0
implying by Hausdorff Young ĥ ∈ Lr . If f, g ∈ L1 then this holds by Fubini. In general use an approximation
argument.
Theorem 4.13 (Fourier transform of |x|α−n ). If 0 < α < n, f ∈ Cc∞ (Rn ) and cα = Γ(α/2)
π α/2
, then
Z
cα F −1 (|k|−α F(f )(k))(x) = cn−α |x − y|Y n−α f (y)dy = cn−α |x|n−α ∗ f (x).


Rn

Remark 4.
F : Cc∞ (Rn ) → C ω (Rn )
F(f ) and derivatives decay faster than the inverse of any polynomial in k. Also, |k|−m F(f )(k) ∈ L1 (Rn ).
The image of F is called Schwartz space.
There exists a Fourier transform such that RHS ∈ C ∞ (Rn ) decays as |x|α−n and in general not in Lp
for p ≤ 2 unless α ≤ n/2
Proof. Z ∞  
2
−α
cα |k| = e−π|k| λ
λα/2−1 dλ ?
0
|k|−α F(f )(k) ∈ L1 (Rn ) so
cα F −1 (|k|−α F(f )(k))
Z Z ∞   
2
= e2πihk,xi e−π|k| λ λα/2−1 dλ F(f )(k)dk
R 0
Z ∞ Z 
2πihk,xi −π|k|2 λ
= e e F(f )(k)dk λα/2−1 dλ
0 Rn
Z ∞ Z Z  
2
= e2πihk,xi−π|k| λ e−2πihk,yi f (y)dy dk λα/2−1 dλ
0 Rn Rn
Z ∞ Z Z  
2
= e2πihk,xi−π|k| λ−2πihk,yi dk f (y)dy λα/2−1 dλ
0 Rn Rn
2πihk, x − yi − π|k|2 λ + π/λ|x − y|2 −π/λ|x−y|2
Z ∞ Z Z |
√
{z }
α/2−1 −π | πk−i(x−y)/λ|2
= λ f (y) e dkdydλ
n n
Z0 ∞ Z  R  √ R 2
  
−π −π | πk−i(x−y)/λ|  2
= e dk e−π/λ|x−y| f (y)dy λα/2−1 dλ
0 Rn
Z ∞ Z 
n/2 α/2−1 −π/λ|x−y|2
= λ λ  f (y)dy dλ
0 bRn
Z Z ∞  Z
α−n
−1 −π/λ|x−y|2
= λ 2 e dλ = cn−α |x − y|α−n f (y)dy
Rn 0 Rn
CHAPTER 4. FOURIER TRANSFORMATIONS 27

2n
Corollary 4.14 (Extension of 4.13). If 0 < α < n/2, f ∈ Lp (Rn ), p = n+2α , then F(f ) exists. Also

g = cn−α |x|α−n ∗ f ∈ L2 (Rn )

and F(g) exists with cα F −1 (|k|−α F(f )(k)) = g(k).

Moreover,
cα |k|−α F(f )(k) = F(g)(k) ?
and Z Z Z
c2α |k| −2α 2
|F(f )(k)| dk = cn−2α f¯(x)f (y)|x − y|2 dxdy
bRn Rn Rn

Proof. Let fn ∈ Cc∞ (Rn ) such that fn → f ∈ Lp , and gn := cn−α |x|α−n ∗ fn ∈ L2 (Rn ). Since fn → f in Lp
then also fˆn → fˆ ∈ Lq with q = n−2α
2n
something about H.L.S and gn → g in L2 so by Plancherel ĝn → ĝ
in L . By previous Theorem, ĝN (k) = cα |k|−α fˆN (x) there exists a subsequence ĝN → ĝ and fˆN → fˆ both
2

pointwise almost everywhere.ĝ(k) = limN →∞ cα |k|−α fˆN (k) = cα |k|−α fˆ(k) for almost every k.
For the second equation, use Plancherel, the first equation, Fubini and
cn−α−β cα cβ
|x|α−n ∗ |x|β−n (y) = |y|α+β−n


cα+β cn−α cn−β
Chapter 5

Distributions

For an open subset of Rn , let D(Ω) be the space of functions f ∈ Cc∞ (Ω) and φn ⊂D(Ω) converges to φ if
(i) supp(φn − φ)⊂K⊂Ω and
(ii) the derivative of φn converges to the derivative of φ uniformly on K.
Definition 5.1. A distribution T : D(Ω) → C is a linear continuous functional. The space of distributions
is D∗ (Ω), the dual space of D(Ω).
A sequence of distributions Tn converges to T if Tn (φ) → T (φ) for all φ ∈ D(Ω).

5.1 Locally summable functions

For 1 ≤ p ≤ ∞, define the locally Lp functions

Lploc = {f − measurable : ∀K⊂Ω, ||f ||Lp (K) < ∞}

A sequence fn converges to f in Lploc if ||f − fn ||Lp (K) → 0 for all compact K⊂Ω.
Note that, if r > p then Lrloc (Ω)⊂Lploc (Ω) by Hölder. This holds in regular Lp whenever Ω has finite
measure.
If f ∈ L1loc , we can define a linear functional Tf ∈ D∗ (Ω) by Tf φ := Ω f φ which is finite since φ is
R

compactly supported. We need to check the continuity of this operator: Indeed,

Z  Z
 
|Tf φn − Tf φ| =  (φn − φ)f  ≤ ||φn − φ||∞ |f | <  · ||f ||L1 (K) .
Ω K

∗ ∗
We have now that L1loc (Ω)
,→ D (Ω) but D (Ω)\L1loc (Ω)
6= φ.
The following theorem ensures, our inclusion is actually injective.
Example 5.2. The dirac “delta-function” satisfies δx (φ) = φ(x) which is certainly in D∗ (Ω).
Theorem 5.3 (Functions uniquely determined by distributions). If f, g ∈ L1loc (Ω) with Tf (φ) = Tg (φ) for
all φ ∈ D(Ω), then f (x) = g(x) almost everywhere.

Proof. Let Ωm = {x ∈ Ω : x + y ∈ Ω, ∀|y| < 1/m}. Consider j ∈ Cc∞ (Rn ) with supp(j)⊂D1 and Rn j = 1.
R

Define a sequence jm (x) = mn j(mx) which converges to the dirac-delta function δ0 . For fixed M , if m ≥ M
and φny (x) = jn (x − y) ∈ D(Ω) we have that (jn ∗ f )(y) = Tf (φny ) = Tg (φny ) = (jn ∗ g)(y) for all x in ΩM .
We know that jn ∗ f → f and jn ∗ g → g in L1loc . Hence f = g almost everywhere in L1loc (ΩM ) and result
hold as M → ∞ so f ∼ g on Ω.

5.2 Derivative of distributions

For α = (α1 , α2 , . . . , αn ) ∈ Nn , T ∈ D∗ (Ω), write Dα for ( ∂x
∂ α1 ∂ α2
1
) ( ∂x2 ) · · · ( ∂x∂n )αn and
define

28
CHAPTER 5. DISTRIBUTIONS 29

Dα T (φ) := (−1)|α| T (Dα φ)

If f ∈ C |α| (Ω) then Dα Tf (φ) = ±Tf (Dα φ) = ± Ω f Dα φ = Ω (Dα f )φ = TDα f φ
R R

We can show that Dα T is actually a distribution. That is, Dα T ∈ D∗ (Ω) because Dα is a continuous
linear operator on D∗ (Ω). To show this, let Tn → T so that ∀φ ∈ D(Ω) we have Tn φ → T φ. Now,

(Dα Tn )φ = (−1)|α| Tn (Dα φ) → (−1)|α| T (Dα φ) = Dα T φ

hence continuity.

5.3 Sobolev Spaces

Define  
1,1 ∂Tf
Wloc (Ω) = f ∈ L1loc (Ω) : = Tg , g ∈ L1loc (Ω), i = 1, . . . , n
∂xi
 
1,p ∂Tf
Wloc (Ω) = f∈ Lploc (Ω) : = Tg , g ∈ Lploc (Ω), i = 1, . . . , n
∂xi
and we have ∇f = (g1 , . . . , gn ) such that gi ∈ Lploc and can also write
Z Z
∂φ
f i =− gi φ
Ω ∂x Ω

similarly Z Z
f ∇φ = − f~g .
Ω Ω
1,p
Wloc (Ω) is a vector space without norm. If p ≤ r then W 1,r ⊂W 1,p and we can unambiguously define

∂Tf
W 1,p (Ω) = {f ∈ Lp (Ω) : = Tgi , gi ∈ Lp (Ω), i = 1, . . . , n}
∂xi
with a norm ( )1/p
n
X
||f ||W 1,p := ||f ||pp + ||∂i f ||pp
i=1
1,p
making W into a complete Banach space.
For notation, if φ ∈ D(Ω) we write φy (x) = φ(x − y).
Lemma 5.4 (Interchange convolution with distribution). If Ω⊂Rn , φ ∈ D(Ω) define

Oφ := {y ∈ Rn : supp(φy )⊂Ω}

(this is open and non-empty).

If T ∈ D0 (Ω), then
(i) the mapping y 7→ Tφy is smooth on Oφ ,
(ii) Dyα T (φy ) = (−1)|α| T (Dα φ)y = (Dα T )(φy ), and
(iii) If ψ ∈ L1 (Oφ ) has compact support then
Z
ψ(y)T (φy )dy = T (ψ ∗ φ).
Oφ

In other words, this is stronger than linearity. That is, T is continuously linear (probably by continuity).
CHAPTER 5. DISTRIBUTIONS 30

Proof. if y ∈ Oφ choose  > 0 such that y + z ∈ Oφ for all |z| < . For all x ∈ Ω, we have

|φy (x) − φy+z (x) = |φ(x − y) − φ(x − y − z)| < C ?

and this also holds for all derivatives of φ.

We have that φy+z → φy as z → 0 in D(Ω) which implies that T φy+z → T φy so the map y 7→ T φy is
continuous on Oφ .
Now,  
 φ(x + δz) − φ(x)
− (∇φ(x)) · z  ≤ C 0 δ|z|


 δ
and
 
T φy+δz − T φy φy+δz − φy
lim =T lim = −T ((∇φ)y · z)
δ→0 δ δ→0 δ
Rest of proof in [1] Lemma 6.8

Theorem 5.5 (Fundamental theorem of calculus for distributions). If Ω⊂Rn is open, φ ∈ D(Ω) a test
function, T ∈ D∗ (Ω). Suppose for some y ∈ Rn that φty ∈ D(Ω) for 0 ≤ t ≤ 1. Then
Z n
1X
T φt − T φ = yj (∂j T )φty dt
0 j=0

1,1
In particular, if f ∈ Wloc (Rn ), then for all y ∈ Rn and almost every x ∈ Rn ,
Z 1
f (x + y) − f (x) = y · ∇f (x + ty)dt
0

Theorem 5.6 (Equivalence of classical and distributional derivatives). Ω⊆Rn open, T ∈ D∗ (Ω) and set
Gi = ∂i T ∈ D∗ (Ω) for all i, then TFAE:
(1) T = f ∈ C 1 (Ω)
(2) Gi = gi ∈ C 0 (Ω) for all i.
∂f
In each case, gi = ∂x i
, the classical derivative.

Proof. (1) → (2)

(2) → (1); Fix small enough R > 0 such that ω := {x ∈ Ω : dist(x, Ωc ) ≥ R} is non-empty. Let φ ∈ D(Ω)
and |y| < R so that φty ∈ D(Ω) for all t ∈ [−1, 1]. By 5.5,
 
Z 1Xn Z Z Z 1X n
T φy − T φ = yj gj (x)ψ(x − ty)dxdt =  gj (x + ty)yj dt ψ(x)dx
0 j=1 ω ω 0 j=1

by change ofR variable. Let ψ ∈ Cc∞ (Rn ) non negative with support in B := BR (0) and ψ = 1. The
R

convolution B ψ(y)φ(x − y)dy with φ ∈ D(ω) is a function in D(Ω). Integrating our equation against ψ
gives, by Fubini,
 
Z Z X n Z Z 1
ψ(y)T (φy )dy − T φ =  ψ(y) yj gj (x + ty)dtdy  φ(x)dx
B ω j=1 B 0

R
with the first term on the left equal to ω φ(x)T (ψx )dx (by 5.4 and noting that ψx ∈ D(Ω)). Hence,
 
Z n Z
X Z 1
Tφ = T (ψx ) − ψ(y) yj gj (x + ty)dtdy  φ(x)dx.
ω j=1 B 0
| {z }
=:f
CHAPTER 5. DISTRIBUTIONS 31

Finally, by the second part of 5.5, we have

Z n
1X
f (x + y) − f (x) = gj (x + ty)yj dt
0 j=1

Pn
for x ∈ ω and |y| < R. The right hand side is j=1 gj (x)yj + O(|y|) and this proves that f ∈ C 1 (ω) with
derivatives gi . Since x can be chosen arbitrarily in Ω and and choosing small enough R this suffices.

Theorem 5.7 (Distributions with zero derivatives are constants). Let Ω⊆Rn be a connected open set and
T ∈ D∗ (Ω). If ∂i T = 0 for all i then there exists a constant C such that
Z
Tφ = C φ
Ω

for all φ ∈ D(Ω).

Proof. Theorem 5.6 implies T = f ∈ C 1 (Ω) with ∂i f = 0 for each i, hence f is constant.

5.4 Multiplication and Convolution of distributions by C ∞ functions

Multiplication
Let T ∈ D∗ (Ω), ψ ∈ C ∞ (Ω) with
(ψT )(φ) := T (ψφ)
for each test function φ. This is a distribution because ψφ has compact support because φ does. Moreover,
if φn → φ, then ψφn → ψφ. To differentiate ψT , apply the product rule

∂i (ψT )(φ) = ψ(∂i T )(φ) + (∂i ψ)T (φ)

which follows from ∂i (ψT )(φ) = −(ψT )(∂i φ) = −T (ψ∂i φ) = T (∂i (ψφ) − (∂i ψ)φ)...
1,p
When T = Tf for some f ∈ L1loc (Ω), then ψTf = Tψf . Moreover, if f ∈ Wloc then so does ψf and our
product rule reads
∂i (f ψ)(x) = f (x)(∂i ψ)(x) + ψ(x)(∂i f )(x)
k,p
for almost every x (because of ∂i f ). These results extend to W k,p and Wloc .

Convolutions
Let j ∈ Cc∞ (Rn ), T ∈ D∗ (Ω) then we define
Z 
(j ∗ T )(φ) := T (jR ∗ φ) = T j(y)φ−y (x)dy
Rn

where jR (x) = j(−x). Note that jR ∗ φ ∈ Cc∞ (Rn ) so this convolution makes sense and j ∗ T ∈ D∗ (Rn ).
When T is a function f , then j ∗ Tf = Tj∗f .
Theorem 5.8 (Approximation of distributions by C ∞ -funcitons). Let T ∈ D∗ (Rn ) and j ∈ Cc∞ (Rn ). Then
there exists a function ψ ∈ C ∞ (Rn ) (depending on only T and j) such that
Z
(j ∗ T )(φ) = ψ(y)φ(y)dy
bRn

for every φ ∈ D(Rn ).

Further, if we assume that j = 1 and we set j (x) = −n j(x/) for  > 0, then j ∗ T converges to T in
R

D∗ (Rn ) as  → 0.

Proof.
CHAPTER 5. DISTRIBUTIONS 32

The Kernel of a distribution T ∈ D∗ (Ω) is

NT = {φ ∈ D(Ω) : T (φ) = 0}

it forms a closed linear subspace of D(Ω).

Theorem 5.9 (Linear dependence of distributions). Let S1 , . . . , SN ∈ D(Ω) and T ∈ D∗ (Ω) satisfies the
T (φ) = 0 for all φ ∈ ∩N
i=1 NSi . Then there exists c1 , . . . , cN ∈ C such that

N
X
T = ci S i .
i=1
Index

Gδ -set, 11 Rayleigh quotient, 17

Real and Complex Analysis, 34
absolutely continuous, 3, 12 regular, 6
Analysis, 34 Riemann-Lebesgue, 18
Riesz Representation, 5, 6
bounded variation, 13 Riesz-Thorin interpolation, 23
complex measure, 2
Schwartz space, 26
concentrated, 3
shrink nicely, 10
convolution, 20
symmetric derivative, 7
differentiable, 14
total variation, 13
dirac “delta-function”, 28
total variation function, 13
distribution, 28
total variation measure, 2
Gaussian distribution, 20
weak L1 , 8
Hahn decomposition, 5
Hahn decomposition Theorem, 5
Hausdorff-Young inequality, 25
Hausdorff-Young on torus, 24

inner regular, 5

Jordan decomposition, 3

Kernel, 32

Laplacian, 17
Lebesgue decomposition, 4
Lebesgue point, 8
locally Lp functions, 28

maximal function, 7, 8
Maximal theorem, 7
metric density, 10
mutually singular, 3

outer regular, 5

Parsevel’s formula, 21
Plancherel’s Theorem, 21
polarization identity, 21
positive and negative variations, 2

Raydon-Nikodym derivative, 4

33
Bibliography

[1] E. H. Lieb, M. Loss, Analysis, American Mathematical Society, 2001.

[2] Walter Rudin, Real and Complex Analysis, McGraw-Hill 1987.

34