Introduction to Geometric

Measure Theory - Lecture Notes

Version: 1.6

Gláucio Terra
IME - USP
E-mail address: [email protected]
 2010 Mathematics Subject Classification. Primary

..
 Contents

General Notations and Conventions iii

Chapter 1. Measure and Integration Theory 1
1.1. Measures 1
1.1.1. Operations on measures 4
1.1.2. Measures on topological spaces 5
1.2. Measurable Maps 14
1.3. Integration Theory 20
1.3.1. Lp spaces 24
1.3.2. Change of variables formula 27
1.4. Product measures and Fubini-Tonelli’s theorem 27
1.5. Signed measures and Lebesgue-Radon-Nikodym theorems 30
1.6. Convolutions 34
1.7. Lusin’s and Egorov’s Theorems 40
Chapter 2. Hausdorff Measures 43
2.1. Carathéodory’s construction 43
2.2. Vitali’s Covering Theorem 48
2.3. Steiner Symmetrization 51
2.4. The isodiametric inequality; Ln = Hn 53
Chapter 3. Differentiation of Measures 57
3.1. Densities 57
3.2. Differentiation Theorems 62
Chapter 4. Rn -valued Radon Measures 81
4.1. Linear functionals on spaces of continuous functions 81
4.2. Operations with Rn -valued Radon measures 104
4.3. Weak-star convergence 111
Chapter 5. Area and Coarea Formulas 121
5.1. Lipschitz maps on Rn 121
5.1.1. Extensions of Lipschitz maps 121
5.1.2. Rademacher’s theorem 122
5.1.3. Linear maps and Jacobians 131
5.2. The area formula 137
i
ii CONTENTS

5.3. The coarea formula 151

5.3.1. Applications of the coarea formula 167
Chapter 6. Sobolev Spaces 169
6.1. Approximation by smooth functions, part I 176
6.2. Product and Chain Rules 183
6.3. Approximation by smooth functions, part II 187
6.4. Lipschitz functions and W1,∞ 195
6.5. Traces and Extensions 195
6.6. Sobolev Inequalities 210
6.6.1. Case 1 ≤ p < n 210
6.6.2. Case n < p 219
6.7. Compactness 223
Chapter 7. Functions of Bounded Variation and Sets of Finite
Perimeter 229
7.1. Gauss-Green Measures and Generalized Divergence
Theorem 233
7.2. Regularization of Radon measures and BV functions 238
7.3. First properties of BV functions 242
7.4. Traces and Extensions 249
7.5. Compactness 268
7.6. Sets of Finite Perimeter and Existence of Minimal Surfaces269
7.6.1. Support of the Gauss-Green measure 269
7.6.2. Operations with Sets of Finite Perimeter, part I 270
7.6.3. Compactness from perimeter bounds 272
7.6.4. Existence of Minimizers 274
Bibliography 277
Index 279
List of Symbols 283
 General Notations and Conventions

We list below some basic notation used for objects which are not
defined in the text. A more detailed list, including the notation used
for objects defined in the text (with references to the pages where each
object was defined) may be found at the list of symbols at the end.

General convention for function spaces: From chapter 4 on, all

function spaces refer to spaces of real-valued functions, unless otherwise
specified.

2X power set of X 283

∪A {y | ∃x ∈ A, y ∈ x} 283
∩A {y | ∀x ∈ A, y ∈ x} 283
∪α∈A Xα the same as ∪{Xα | α ∈ A} 283
∪α∈A Xα the same as ∩{Xα | α ∈ A} 283
R [−∞, ∞] (extended real numbers) 283
U(x, r) open ball of center x and radius r 283
B(x, r) closed ball of center x and radius r 283
U or Un open unit ball in Rn 283
B or Bn closed unit ball in Rn 283
Sn unit sphere in Rn+1 283
C(x, r, h) U(p · x, r) × U(q · x, h) ⊂ Rk × Rn−k , for x = (p · x, q · x) ∈ Rk × Rn−k
283
C(x, r) C(x, r, r) 283
C(x, r, h) C(x, r, h) ⊂ Rk × Rn−k 283
C(x, r) C(x, r, r) 283
A closure of A 283
Ao interior of A 283
c
A complement of A 283
χA characteristic function of A 283

iii
iv Glossary

A∆B ˙
symmetric difference of A and B, i.e. (A \ B) ∪(B \ A) 283
AbX A is a compact subset of X 283
kxk norm of x (in Rn , the euclidean norm, unless otherwise specified) 283
k·ku norm of uniform convergence. 283
x
sgn x kxk
if x 6= 0 and 0 otherwise 283
standard basis of Rn 283
(e1 , . . . , en )
π m/2
α(m) (euclidean volume of Bm if m integer) 283
Γ(m/2 + 1)
lim sup f (x) inf sup f (x) = lim sup f (x) 283
x→x0 δ>0 x∈U(x0 ,r) δ→0 x∈U(x0 ,r)
lim inf f (x) sup inf f (x) = lim inf f (x) 283
x→x0 δ>0 x∈U(x0 ,r) δ→0 x∈U(x0 ,r)
LCH locally compact Hausdorff space 283
LCS locally compact separable metric space 283
Cb bounded continuous functions 283
Ckb Ck functions with bounded derivatives up to order k 283
Cc continuous functions with compact support 283
Ckc Ck functions with compact support 283
C0 continuous functions which vanish at infinity 283
Ck0 Ck functions whose derivatives up to order k vanish at infinity 283
Ck (U ) Ck functions on U whose derivatives up to order k extend continuously
to U 283
gr f graph of f , i.e. {(x, y) | y = f (x)} 283
epi f epigraph of f , i.e. {(x, y) ∈ dom f × R | y ≥ f (x)} 283
epiS f strict epigraph of f , i.e. {(x, y) ∈ dom f × R | y > f (x)} 283
hyp f hypograph of f , i.e. {(x, y) ∈ dom f × R | y ≤ f (x)} 283
hypS f strict hypograph of f , i.e. {(x, y) ∈ dom f × R | y < f (x)} 283
f ≺U f ∈ Cc (U ) and 0 ≤ f ≤ 1 283
δa Dirac measure centered at a. 283
fˇ x 7→ f (−x) 283
τy f x 7→ f (x − y) 283
Lip f Lipschitz constant of f 283 
α1
  αn
∂ αf For a multi-index α ∈ Zn+ , ∂x∂ 1 . . . ∂x∂n f 283
|α| For a multi-index α ∈ Zn , α1 + · · · + αn . 283
Df Fréchet derivative of f 283
Λ(n, m) set of strictly increasing functions {1, . . . , m} → {1, . . . , n} 283
 CHAPTER 1

Measure and Integration Theory

1.1. Measures
Definition 1.1. A measure on a set X is a set function µ : 2X →
[0, ∞] such that:
M1) µ(∅) = 0,
M2) (monotonicity) µ(A) ≤ µ(B) whenever P A ⊂ B,
∞
M3) (countable subadditivity) µ(∪∞ A
n=1 n ) ≤ n=1 µ(An ).

Warning. Our nomenclature is in accordance with the one com-

monly used in Geometric Measure Theory. However, most textbooks
on Real Analysis (see, for instance, [Fol99]) call such a set function an
outer measure, reserving the name measure for a countably additive set
function defined on a σ-algebra M of subsets of X, as defined below in
1.6. We shall use the term “measure” for both types of set functions,
if no confusion arises; if the context does not make it clear, we may
use, for clarification, “measure on a σ-algebra” or “measure on M” for
countably additive set functions on σ-algebras, or “outer measure” for
the set functions introduced in the previous definition.
Definition 1.2. Given a measure µ on a set X, a subset A ⊂
X is called measurable with respect to µ (or µ-measurable, or simply
measurable) if it satisfies the Carathéodory condition:
∀T ⊂ X, µ(T ) = µ(T ∩ A) + µ(T \ A).
We denote by σ(µ) the set of measurable subsets of X with respect
to µ.
Example 1.3. The following are examples of measures:
1) Let X be a set and µ : 2X → [0, ∞] be defined by µ(A) := card (A)
if A is finite and µ(A) := ∞ otherwise. Then µ is a measure on X,
called counting measure, and it can be readily checked that σ(µ) =
2X .
2) Let X be a set, a ∈ X and µ : 2X → [0, ∞] be defined by µ(A) := 1
if a ∈ A and µ(A) := 0 otherwise. Then µ is a measure on X, called
Dirac measure centered at a, denoted by δa (or simply δ if a = 0).
It can be readily checked that σ(δa ) = 2X .
1
2 1. MEASURE AND INTEGRATION THEORY

3) Let X = Rn and µ : 2X → [0, ∞] be defined by µ(A) := inf{ Q∈A vol(Q) |

P
A countable cover of A by cubes with sides parallel to the coordi-
nate axes}, where vol(Q) denotes the euclidean volume of the cube
Q (which is not assumed to be open or closed, i.e. any product of
intervals with the same length is a valid cube). Then µ is a measure
on Rn , called Lebesgue measure. We denote the Lebesgue measure
on Rn by |·| or Ln , and the set σ(Ln ) of Lebesgue-measurable sets
by L Rn (or simply L ).
4) Hausdorff measures, to be defined in section 2.
Exercise 1.4.
a) Ln is invariant by translations, i.e. ∀x ∈ Rn , ∀A ⊂ Rn , Ln (A + x) =
Ln (A).
b) Ln is homogeneous of degree n with respect to homotheties, i.e.
∀λ > 0, ∀A ⊂ Rn , Ln (λA) = λn Ln (A).
Exercise 1.5. Let µ and ν be measures on X and c > 0. Then:
a) µ + ν is a measure on X and σ(µ) ∩ σ(ν) ⊂ σ(µ + ν).
b) cµ is a measure on X and σ(cµ) = σ(µ).
Definition 1.6. Given a set X, M ⊂ 2X is called an algebra of
subsets of X if it contains the empty set, it is closed under comple-
mentation and closed under finite unions. M is called a σ-algebra if it
is an algebra closed under countable unions. The sets in M are called
measurable with respect to M, or M-measurable, or (if clear from the
context) simply measurable.
Given a σ-algebra M ⊂ 2X , we call a set function µ : M → [0, ∞]
a measure on M if it satisfies:
M1) µ(∅) = 0,
∞
M2) (countable additivity) µ(∪˙ n=1 An ) = ∞ ˙
P
n=1 µ(An ) (we use “∪” for
disjoint unions).
A measure µ : M → [0, ∞] is called:
• complete if E ∈ M, µ(E) = 0 and A ⊂ E implies A ∈ M,
• finite if µ(X) < ∞,
• σ-finite if there exists a sequence (En )n∈N in M such that
∪n∈N En = X and ∀n ∈ N, µ(En ) < ∞. More generally, a set
A ⊂ X is said to be σ-finite if it can be covered by countably
many measurable sets of finite measure.
Theorem 1.7 (Carathéodory). If µ is a measure on a set X, then
σ(µ) is a σ-algebra and the restriction of µ to σ(µ) is a complete mea-
sure on σ(µ).
 1.1. MEASURES 3

Thus, each measure determines a measure on a σ-algebra by re-

striction to its measurable sets. Conversely, each measure defined on a
σ-algebra of subsets of X can be extended to a measure on X:
Theorem 1.8. If M is a σ-algebra of subsets of X and µ : M →
[0, ∞] is a measure on M, then the set function:
µ∗ : 2X −→ [0, ∞]
A 7−→ inf{µ(E) | A ⊂ E ∈ M}
is a measure which extends µ and such that M ⊂ σ(µ∗ ).
The above theorem is a corollary of Carathéodory’s extension the-
orem ([Fol99], proposition 1.13). Henceforth, whenever no confusion
arises, we shall drop the “∗” in the notation and denote by the same
symbol both the measure µ on M and its induced measure on X.
Definition 1.9. A measure µ : 2X → [0, ∞] is called:
• regular , if ∀A ⊂ X, ∃E ∈ σ(µ) such that A ⊂ E and µ(A) =
µ(E).
• finite (respectively, σ-finite) if so is its restriction to σ(µ), cf.
definition 1.6. We define similarly sets which are finite or σ-
finite with respect to µ.
Remark 1.10. i) If we depart from a measure µ : M → [0, ∞]
defined on a σ-algebra M ⊂ 2X and take its extension µ∗ : 2X →
[0, ∞] given by theorem 1.8, then the measure µ∗ : σ(µ∗ ) → [0, ∞]
is an extension of µ. It coincides with the completion of µ if µ is
σ-finite (hence it is equal to µ if µ is complete and σ-finite). In
general, this extension coincides with saturation of the completion
of µ (see exercise 1.22 in [Fol99]).
ii) Similarly, if we depart from a measure µ : 2X → [0, ∞], take
the measure on σ(µ) given by the restriction of µ to σ(µ), and
then take the extension µ∗ : 2X → [0, ∞] of the latter measure
given by theorem 1.8, then µ∗ is a regular measure which satisfies
µ ≤ µ∗ . Equality holds iff µ is a regular measure, cf. exercise 1.20
in [Fol99].
Proposition 1.11 (continuity properties of measures). For a mea-
sure µ on X, the following properties hold:
i) (continuity from below) if (En )n∈N is an increasing sequence in
σ(µ), then µ(∪∞n=1 En ) = limn→∞ µ(En ),
ii) (continuity from above) if (En )n∈N is a decreasing sequence in σ(µ)
and µ(E1 ) < ∞, then µ(∩∞ n=1 En ) = limn→∞ µ(En ).
4 1. MEASURE AND INTEGRATION THEORY

Exercise 1.12. If µ is a regular measure on X, property i) in

proposition 1.11 holds for any increasing sequence (En )n∈N of subsets
of X (i.e. the sets need not be measurable).
1.1.1. Operations on measures. Three useful ways of obtaining
new measures from old are restrictions, traces and pushforwards:
Definition 1.13 (Restrictions and traces of measures). Let µ be
a measure on a set X and A ⊂ X. We define the:
x
• restriction of µ to A, denoted by µ A, as the measure 2X →
[0, ∞] given by E 7→ µ(A ∩ E).
• trace of µ on A, denoted by µ|A , as the measure 2A → [0, ∞]
given by E 7→ µ(E), i.e the restriction to 2A ⊂ 2X of the map
µ : 2X → [0, ∞].
x
Note that the restriction µ A is a measure on X, whereas the
trace µ|A is a measure on A. Moreover, we do not assume A to be
µ-measurable.
Definition 1.14 (Pushforward of measures). Let µ be a measure
on the set X and f : X → Y a map into the set Y . We define a
measure 2Y → [0, ∞] on Y by:
A ⊂ Y 7→ µ f −1 (A) ,

called pushforward of µ by f and denoted by f# µ.

Proposition 1.15. Let µ be a measure on the set X, A ⊂ X and
f : X → Y . The following properties hold:
i) σ(µ) ⊂ σ(µ A).x
ii) If E ∈ σ(µ), then E ∩ A ∈ σ(µ|A ). Besides, if A ∈ σ(µ), then
σ(µ|A ) = σ(µ) ∩ 2A = {E ∈ σ(µ) | B ⊂ A}.
iii) For B ⊂ Y , f −1 (B) is µ-measurable iff ∀A ⊂ X, B is f# (µ A)- x
measurable.
Proof.
i) Let B ∈ σ(µ). It follows from Carathéodory’s condition in 1.2 that,
x
for
all T ⊂ X, µ A(T ) = µ(A ∩ T ) = µ(A ∩ T ∩ B) + µ (A ∩ T ) \
x x x
B = µ A(T ∩B)+µ A(T \B), hence B is µ A-measurable.
ii) Let B ∈ σ(µ). It follows from Carathéodory’s condition that, for
all T ⊂ X, µ(T ) = µ(T ∩ B) + µ(T \ B). In particular, for all
T ⊂ A, since T ∩ B = T ∩ B ∩ A and T \ B = T \ (B ∩ A):

µ(T ) = µ(T ∩ B) + µ(T \ B) = µ(T ∩ B ∩ A) + µ T \ (B ∩ A) ,
hence B ∩ A is µ|A -measurable.
 1.1. MEASURES 5

Besides, if A ∈ σ(µ) and B ∈ σ(µ|A ), for all T ⊂ X:

1 2
µ(T ) = µ(T ∩ A) + µ(T \ A) =
3
= µ(T ∩ A ∩ B) + µ(T ∩ A \ B) + µ(T \ A) = µ(T ∩ B) + µ(T \ B),
where we have used the µ-measurability of A in (1), the µ|A -
measurability of B in (2) and again the µ-measurability of A in

us to conclude
that µ(T ∩ A \ B) + µ(T \ A) =
(3), which allows
µ (T \ B) ∩ A + µ (T \ B) \ A = µ(T \ B). Thus B ∈ σ(µ).
iii) Let B ⊂ Y such that f −1 (B) is µ-measurable. For all A ⊂ X, for
all S ⊂ Y :

x x
f# (µ A)(S) = µ A f −1 (S) = µ A ∩ f −1 (S) =

= µ A ∩ f −1 (S) ∩ f −1 (B) + µ A ∩ f −1 (S) \ f −1 (B) =

= µ A ∩ f −1 (S ∩ B) + µ A ∩ f −1 (S \ B) =

xA)(S ∩ B) + f (µ xA)(S \ B),

= f# (µ #

hence B is f (µ xA)-measurable.
#
Conversely, assume that ∀A
xA)-measurable.

⊂ X, B is f (µ #
For all T ⊂ X: µ T ∩ f (B) + µ T \ f (B) = f (µ xT )(B) +
−1 −1
#
f (µ xT )(Y \ B) = f (µ xT )(Y ) = µ xT (X) = µ(T ), hence
# #
f −1 (B) is µ-measurable.

1.1.2. Measures on topological spaces. We now introduce a
topology τ on the set X. We shall consider measures on X which
interact with the topology, in the sense that they have nice regularity
and approximation properties, to be made precise below. This will
allow us to obtain theorems which make an interplay between topology
and measure theory, one of the key ideas of Geometric Measure Theory.
Recall that, given a subset S ⊂ 2X , there exists a smallest σ-algebra
of subsets of X which contains S, that is, the intersection of the family
of σ-algebras that contain S (this family is non-empty, since 2X is such
a σ-algebra). We denote this σ-algebra by σ(S), the so-called σ-algebra
generated by S.
Definition 1.16. For a topological space (X, τ ), we define its Borel
σ-algebra as the σ-algebra generated by τ , i.e. σ(τ ). We denote it by
B X or B(X). The elements of B X are called Borel sets.
6 1. MEASURE AND INTEGRATION THEORY

We say that a measure µ on X is a Borel measure if each Borel set

is µ-measurable, i.e. if B X ⊂ σ(µ). A Borel regular measure on X is
a Borel measure on X which satisfies: ∀A ⊂ X, ∃E ∈ B X such that
A ⊂ E and µ(A) = µ(E).
Exercise 1.17. Let µ and ν be measures on a topological space X
and c > 0.
a) If µ and ν are Borel measures on X, so are µ + ν and cµ.
b) If µ and ν are Borel regular measures on X, so are µ + ν and cµ.
Note that, if S ⊂ 2X and M is a σ-algebra of subsets of X, then
σ(S) ⊂ M iff S ⊂ M. In particular, if (X, τ ) is a topological space
and µ is a measure on X, then µ is a Borel measure iff τ ⊂ σ(µ), i.e. if
each open subset of X is µ-measurable (or, equivalently, if each closed
subset of X is µ-measurable). In case the topology be metrizable by a
metric d, a simple criterion for a measure µ to be Borelian is given by
the theorem below.
Theorem 1.18 (Carathéodory’s criterion). A measure µ on a met-
ric space (X, d) is Borel iff
(Ca) µ(A ∪ B) = µ(A) + µ(B)
whenever A, B ⊂ X satisfy d(A, B) := inf{d(a, b) | a ∈ A, b ∈ B} > 0.
Proof. If µ is Borel and d(A, B) > 0, then A∩B
= ∅ = A∩B and,


by the measurability of A, µ(A∪B) = µ (A∪B)∩A +µ (A∪B)\A =
µ(A) + µ(B).
Conversely, assume that condition (Ca) holds. In order to prove
that µ is Borel, it suffices to prove that every closed set C is measurable.
Equivalently, we must show that, for each T ⊂ X such that µ(T ) < ∞,
µ(T ) ≥ µ(T ∩ C) + µ(T \ C) (what clearly implies Carathéodory’s
condition in definition 1.2, since the same inequality is trivial if µ(T ) =
∞ and the other inequality holds by subadditivity of µ).
For each i ∈ N, let Ci := {x ∈ X | d(x, C) ≤ 1/i}. Then d(T ∩
C, T \ Ci ) ≥ 1/i > 0, so that the monotonicity

of µ and condition (Ca)
imply µ(T ) ≥ µ (T ∩ C) ∪ (T \ Ci ) = µ(T ∩ C) + µ(T \ Ci ). Therefore,
i→∞
it suffices to prove that µ(T \ Ci ) −→ µ(T \ C).
1
For each j ∈ N, let Tj := T ∩ {x ∈ X | j+1 < d(x, C) ≤ 1j }. Due to
the fact that C is closed, d(x, C) > 0 iff x ∈ X\C, henceP∀i ∈ N, T \C =
T \Ci ∪∞ \C) ≤ µ(T \Ci )+ ∞
j=i Tj . Therefore, ∀i ∈ N, µ(TP j=i µ(Tj ). The
thesis then follows if we show that ∞ j=1 µ(Tj ) < ∞, since this implies
P∞ i→∞
j=i µ(Tj ) −→ 0, so that µ(T \ C) ≤ lim inf µ(T \ Ci ) ≤ lim sup µ(T \
Ci ) ≤ µ(T \ C), where the last inequality holds by monotonicity.
 1.1. MEASURES 7

Since d(Ti , Tj ) > 0 if i 6= j are both odd or both even, it follows from
condition (Ca) that, for all k ∈ N, µ(T ) ≥ µ(∪kj=1 T2j ) = kj=1 µ(T2j )
P

and µ(T ) ≥ µ(∪kj=1 T2j+1 ) = kj=1 µ(T2j+1 ), thus ∞

P P
j=1 µ(T 2j ) ≤ µ(T )
P∞ P∞
and j=1 µ(T2j+1 ) ≤ µ(T ), from what we conclude that j=1 µ(Tj ) ≤
2µ(T ) < ∞.

Example 1.19. The Lebesgue measure Ln is a Borel regular mea-
sure on Rn . Indeed:
1) Let d be the euclidean distance in Rn ; we show Ln satisfies the
Carathéodory criterion (CA). Given A, B ⊂ Rn such that d(A, B) =
δ > 0, let A be a countable cover of A∪B by cubes with sides paral-
lel to the coordinate axes. Subdividing the sides of each cube in A, if
necessary, we may take another countable cover A0 P of A ∪ B formed
by cubes of diameter less than δ/2 and such that Q∈A0 vol(Q) =
P
Q∈A vol(Q). Discarding the cubes of the latter family which do
not intersect A or B, we obtain a subfamily A00 which still covers
A ∪ B. In view of the choice of the diameters of the cubes in A0 ,
we can decompose A00 in two disjoint subfamilies A00 = A1 ∪˙ A2 ,
where the cubes in A1 cover A and P those in A2 cover
P B. It then
n n
follows that L (A) + L (B) ≤ Q∈A1 vol(Q) + Q∈A2 vol(Q) =
P P P
Q∈A00 vol(Q) ≤ Q∈A0 vol(Q) = Q∈A vol(Q). By the arbitrari-
ness of the countable cover A of A∪B by cubes with sides parallel to
the coordinate axes, we conclude that Ln (A) + Ln (B) ≤ Ln (A ∪ B),
and the other inequality holds by finite subadditivity of Ln . Thus,
by theorem 1.18, Ln is a Borel measure.
2) Let A ⊂ Rn such that Ln (A) < ∞. We contend that ∃B ∈ B Rn
such that A ⊂ B and Ln (A) = Ln (B) (hence Ln is Borel regular).
As a matter of fact, for each n ∈ N, take a countable cover An
of A by cubes with P sides parallel to the coordinate axes such that
Ln (A) + 1/n > Q∈An vol(Q). Take B := ∩n∈N ∪Q∈An Q, so that
A⊂B∈B PRn . Then, for each n ∈ N, An covers B, so that ∀n ∈ N,
Ln (B) ≤ Q∈An vol(Q) < Ln (A) + 1/n, whence Ln (B) ≤ Ln (A),
and the other inequality holds by monotonicity of Ln .
Remark 1.20. Since the Lebesgue measure of each bounded cube
Q in Rn with sides parallel to the coordinate axes is finite (as it is
≤ vol(Q) < ∞, by definition; it actually coincides with vol(Q), but we
postpone the proof of this fact to example 1.86, after the introduction
of product measures), and since Rn is a countable union of such cubes,
which are Borelian, it follows that the restriction of Ln to B Rn is a
σ-finite measure on B Rn . By the Borel regularity of Ln , cf. example
8 1. MEASURE AND INTEGRATION THEORY

1.19, the extension given by theorem 1.8 of Ln : B Rn → [0, ∞] is Ln

itself. Therefore, by remark 1.10.(i), we conclude that the completion
of Ln : B Rn → [0, ∞]. is Ln : L = σ(Ln ) → [0, ∞].
Remark 1.21. The fact that the inclusions B Rn ⊂ L ⊂ 2R are
n

strict can be seen by cardinality arguments. Indeed, card (B Rn )=c and

card (L ) = 2c , whence B Rn $ L . As to the strictness of the other
inclusion, it holds a more general result – see theorem 2.2.4 in [Fed69].
Definition 1.22. A Borel measure µ on a topological space (X, τ )
is called:
• open σ-finite if there exists a sequence (Un )n∈N of open subsets
of X such that X = ∪n∈N Un and ∀n ∈ N, µ(Un ) < ∞.
• locally finite if, for each x ∈ X, there exists an open neighbor-
hood U of x such that µ(U ) < ∞ .
It is clear that a locally finite Borel measure on a second countable
topological space is open σ-finite.
As a rule of thumb, there are two main classes of measures which in-
teract nicely with the topology: 1) locally finite Borel regular measures
on separable metric spaces and 2) Radon measures (to be introduced in
definition 1.28) on locally compact Hausdorff spaces. For instance, the
approximation theorem below holds in the first case (and, by definition
1.28, similar approximation properties also hold for Radon measures).
In later developments of the theory we shall be mainly interested in
locally compact separable metric spaces (for instance, open subsets of
Rn or, more generally, locally closed subsets of Rn , like embedded sub-
manifolds), for which the aforementioned classes of measures coincide,
cf. exercise 1.32.
Theorem 1.23 (approximation by open and closed sets). Let µ be
an open σ-finite Borel regular measure on a topological space (X, τ )
for which each closed set is a Gδ (i.e. a countable intersection of open
sets). The following approximation properties hold:
i) (approximation by open sets from the outside) ∀A ⊂ X, µ(A) =
inf{µ(U ) | A ⊂ U ∈ τ },
ii) (approximation by closed sets from the inside) ∀A ∈ σ(µ), µ(A) =
sup{µ(C) | C ⊂ A, C closed}.
Remark 1.24. The theorem holds, in particular, for a locally finite
Borel regular measure on a separable metric space.
The proof is a consequence of the following lemmas.
Lemma 1.25. Let X be a set, S ⊂ 2X and F ⊂ 2X such that:
 1.1. MEASURES 9

• F is closed under countable intersections and countable unions.

• If A ∈ S, both A and its complement Ac belong to F.
Then F ⊃ σ(S).
Proof. Let G := {A ∈ F | Ac ∈ F}. Then:
1) S ⊂ G.
2) G is closed under complementation.
3) G is closed under countable unions. Indeed, if (An )n∈N is a sequence
in G, then ∪n∈N An ∈ F and (∪n∈N An )c = ∩n∈N Acn ∈ F.
Therefore, G is a σ-algebra which contains S, i.e. σ(S) ⊂ G ⊂ F. 
Corollary 1.26. If X is a set and S ⊂ 2X , σ(S) is the smallest
family of subsets of X closed under countable unions and countable
intersections, which contains S and the complements of the elements
of S.
Lemma 1.27. Let µ be a Borel measure on a topological space (X, τ )
for which each closed set is a Gδ . If B ∈ B X and µ(B) < ∞, for all
 > 0 there exists a closed set C ⊂ B such that µ(B \ C) < .
x
Proof. Define ν := µ B. By proposition 1.15, ν is a finite Borel
measure.
Let S be the family of all closed subsets of X and F the family of
all ν-measurable sets A ⊂ X such that ∀ > 0, ∃C ⊂ A closed with
ν(A \ C) < . We assert that F ⊃ B X ; in particular, that implies
B ∈ F, whence the thesis. The assertion follows once we show that F
satisfies the hypotheses of lemma 1.25. Indeed:
• Let (An )n∈N be a sequence in F and fix  > 0. For each n ∈ N,
∃Cn ⊂ An closed such that ν(An \ Cn ) < 2−n . Then, since
both ∩n∈N An \ ∩n∈N Cn and ∪n∈N An \ ∪n∈N Cn are contained in
∪n∈N (An \ Cn ), it follows by subadditivity that:
1) ν(∩n∈N An \ ∩n∈N Cn ) <  and ∩n∈N Cn is closed, thus
∩n∈N An ∈ F
2) ν(∪n∈N An \∪n∈N Cn ) < . Since ν is finite and the sequence
of ν-measurable sets (∪n∈N An \ ∪kn=1 Cn )k∈N decreases to
∪n∈N An \ ∪n∈N Cn , by continuity from above 1.11 there
exists k ∈ N such that ν(∪n∈N An \ ∪kn=1 Cn ) < . As
∪kn=1 Cn is closed, this shows that ∪n∈N An ∈ F.
• Since every closed subset of X is a Gδ , taking complements we
conclude that every open subset of X is a Fσ , i.e. a countable
union of closed sets. Thus, if C ∈ S, then C ∈ F and X \ C ∈
F, since F is closed under countable unions by the previous
item.
10 1. MEASURE AND INTEGRATION THEORY

Hence, by lemma 1.25, F ⊃ σ(S) = B X , as asserted. 

Proof of theorem 1.23. Firstly, we prove part (ii). Let A ∈
σ(µ). Assume that µ(A) < ∞. Since µ is Borel regular, ∃B 0 ∈ B X
such that B 0 ⊃ A and µ(B 0 ) = µ(A) < ∞, hence µ(B 0 \ A) = 0. Use
the Borel regularity again to obtain B 00 ∈ B X such that B 00 ⊃ B 0 \ A
and µ(B 00 ) = µ(B 0 \ A) = 0. Then B := B 0 \ B 00 ∈ B X is such that
B ⊂ A and µ(B) = µ(A). Applying lemma 1.27 for a given  > 0,
we obtain a closed set C ⊂ B such that µ(B \ C) = µ(A \ C) < ,
which proves part (ii) in case A has finite measure. If µ(A) = ∞, due
to the fact that µ is σ-finite, there exists a disjoint sequence (An )n∈N
in σ(µ) such that A = ∪˙ n∈N An and ∀n ∈ N, µ(An ) < ∞. Given
 > 0, for each n ∈ N, apply the case just proved to obtain a closed
P∞Cn ⊂ An such that µ(An \ Cn ) = µ(An ) − µ(C
set Pn∞) < 2−n . Since
n=1 µ(An ) = µ(A) = ∞, it then follows that n=1 µ(Cn ) = ∞.
Thus, for every M > 0, there exists N ∈ NP such that the closed
N
subset C = ∪N C
n=0 n of A has measure µ(C) = n=1 µ(Cn ) > M , i.e.
sup{µ(C) | C ⊂ A, C closed} = ∞ = µ(A).
In order to prove part (i), we may assume, by Borel regularity, that
A ∈ B X . Assume that there exists an open set V such that V ⊃ A
and µ(V ) < ∞. The thesis in this case follows from part (ii), passing
to the complements: for
a given  > 0, take a closed set C ⊂ V \ A
such that µ (V \ A) \ C < . Then U = V \ C is an open set which
does the job: U ⊃ A and µ(U \ A) < , since U \ A = (V \ A) \ C.
In the general case, given A ∈ B X , we use the hypothesis of µ being
open σ-finite to obtain a sequence (Vn )n∈N of open sets of finite measure
such that A ⊂ ∪n∈N Vn . Fix  > 0. For each n ∈ N, we may apply the
case just proved to the Borel set A ∩ V
n ⊂ Vn to obtain an open set
Un ⊃ A ∩ Vn such that µ Un \ (A ∩ Vn ) < 2−n . Then U = ∪n∈N
Un is
an open set which cointains A and U \ A ⊂ ∪n∈N Un \ (A ∩ Vn ) , thus
µ(U \ A) <  by countable subadditivity.

Definition 1.28. A Radon measure on a locally compact Hausdorff
topological space (X, τ ) is a Borel measure µ on X such that:
R1) (finiteness on compact sets) if K is a compact subset of X, then
µ(K) < ∞,
R2) (interior regularity for open sets) for all U ⊂ X open, µ(U ) =
sup{µ(K) | K ⊂ U, K compact},
R3) (exterior regularity) for all A ⊂ X, µ(A) = inf{µ(U ) | A ⊂
U, U open}.
Remark 1.29.
 1.1. MEASURES 11

i) Note that, by condition R2 in the definition above, every Radon

measure is Borel regular.
ii) A measure µ : B X → [0, ∞] is called a Radon measure on B X if
its extension µ∗ : 2X → [0, ∞] given by theorem 1.8 is a Radon
measure as defined above. That is equivalent to saying that µ
satisfies R1, R2 and R3 for any Borel set A ⊂ X, what coincides
with the usual definition of Radon measures in most Real Analysis
textbooks (for instance, in [Fol99], section 7.1).
It is clear that, if we depart from a Radon measure µ : 2X →
[0, ∞], its restriction to B X is an Radon measure on B X , whose
extension given by theorem 1.8 is the original measure µ, thanks to
its Borel regularity. We therefore obtain a bijection between the set
of outer Radon measures on X and the set of Radon measures on
B X , which associates each Radon outer measure to its restriction
to B X . By means of this bijection, we may identify Radon outer
measures on X and Radon measures on B X .
Exercise 1.30. If µ and ν are Radon measures on a locally compact
Hausdorff space X and c > 0, then µ + ν and cµ are Radon measures
on X.
Exercise 1.31. If µ is a Radon measure on a locally compact Haus-
dorff space (X, τ ), then µ is inner regular on all σ-finite µ-measurable
sets, i.e. property R2 holds for any σ-finite µ-measurable set A in
place of U . In particular, if µ is σ-finite, property R2 holds for all
µ-measurable sets.
Exercise 1.32. Let X be a locally compact separable metric space.
Then µ is a Radon measure on X iff µ is a locally finite Borel regular
measure on X. Moreover, if µ is such a measure, then µ is σ-finite,
hence it is inner regular on all µ-measurable sets by the previous exer-
cise.
Remark 1.33. It follows from exercise 1.32 that, if X is a locally
compact separable metric space and µ : B X → [0, ∞] is a measure
which is finite on compact subsets of X, then the extension of µ to a
measure on X given by theorem 1.8 is a Radon measure (since it is a
locally finite Borel regular measure on X). In particular, the measure
µ on B X is Radon, cf. remark 1.29.
Definition 1.34 (Support of a measure on a topological space).
Let µ be a measure on a topological space X.
• We say that µ is concentrated on a set A ⊂ X if µ(X \ A) = 0.
• The support of µ, denoted by spt µ, is the complement of the
union of all open sets V ⊂ X such that µ(V ) = 0.
12 1. MEASURE AND INTEGRATION THEORY

In the situation of the definition above, in general it is not true

that µ is concentrated on its support, i.e that µ(X \ spt µ) = 0. The
following proposition gives two sufficient conditions for that property
to hold:
Proposition 1.35. If µ is a measure on a second countable topo-
logical space or if µ is a Radon measure on a locally compact Hausdorff
topological space, then µ is concentrated on its support. Actually, its
support is the smallest closed set on which µ is concentrated.
Proof. If µ is a measure on a second countable topological space
X, by Lindelöf’s theorem we may cover X \ spt µ by countably many
open sets of measure zero, thus µ(X \ spt µ) = 0. If µ is a Radon
measure on a locally compact Hausdorff topological space, for each
compact K ⊂ X \ spt µ, we may cover K with finitely many open sets
of measure zero, hence µ(K) = 0. By interior regularity, it follows that
µ(X \ spt µ) = sup{µ(K) | K ⊂ X \ spt µ, K compact} = 0. In any of
the two cases, its clear that spt µ the smallest closed set on which µ is
concentrated.

In the following propositions, we relate measurability and regularity
properties of a measure µ and those of the measures obtained from µ
by restriction or pushforward operations.
Proposition 1.36. Let µ be a measure on the set X and A ⊂ X.
The following properties hold:
i) If X is a metric space, µ is a Borel regular measure on X and
either 1) A ∈ B X or 2) A ∈ σ(µ) and µ(A) < ∞, then µ A is x
Borel regular.
ii) If X is a locally compact separable metric space, µ a Radon measure
on X and either 1) A ∈ B X or 2) A ∈ σ(µ) and µ(A) < ∞, then
x
µ A is a Radon measure.
Proof.
x
i) In both cases µ A is a Borel measure, by proposition 1.15. We
must show that it is Borel regular.
1) Let A ∈ B X . Given T ⊂ X, we must show that ∃B ∈ B X
x x
such that B ⊃ T and µ A(B) = µ A(T ). Since µ is Borel
regular, ∃B 0 ∈ B X such that B 0 ⊃ A∩T and µ(B 0 ) = µ(A∩T ).
Since B 0 ⊃ A ∩ B 0 ⊃ A ∩ T , by monotonicity it follows µ(A ∩
B 0 ) = µ(A ∩ T ). Then, taking B = B 0 ∪ Ac ∈ B X , we have
x
B ⊃ T and µ A(B) = µ(A ∩ B) = µ(A ∩ B 0 ) = µ(A ∩ T ) =
x
µ A(T ).
 1.1. MEASURES 13

2) Let A ∈ σ(µ) with µ(A) < ∞. Since µ is Borel regular, ∃A0 ∈

B X such that A0 ⊃ A and µ(A0 ) = µ(A). Since A ∈ σ(µ) has µ-
finite measure, by finite additivity it follows that µ(A0 \ A) = 0,
x x
hence µ A = µ A0 is Borel regular by the previous item.
ii) By remark 1.29, µ is Borel regular. Hence, by the previous item,
x x
µ A is Borel regular. Since µ is locally finite, so is µ A. There-
x
fore, from exercise 1.32, we conclude that µ A is Radon.

Proposition 1.37. If both X and Y are separable locally compact
metric spaces, f a continuous proper map and µ a Radon measure on
X, then f# µ is a Radon measure on Y , and spt f# µ = f (spt µ).
Proof.
1) f# µ is a Borel measure. Indeed, if U ⊂ Y is open, so is f −1 (U ),
since f is continuous. In particular, f −1 (U ) ∈ B X ⊂ σ(µ), hence
U ∈ σ(f# µ) by proposition 1.15.
2) f# µ is locally finite. Indeed, if K ⊂ Y is compact, so is f −1 (K),
since f is proper. Hence f# µ(K) = µ f −1 (K) < ∞. As Y is


locally compact, the assertion follows.
3) f# µ is Borel regular (hence Radon, by the previous items and by
exercise 1.32). Indeed, given T ⊂ Y , we apply the exterior regularity
of µ on f −1 (T ) to obtain, for each n ∈
N, Un ⊂ X open such that
Un ⊃ f −1 (T ) and µ(Un ) ≤ µ f −1 (T ) + 1/n. Since f is closed
(because it is proper), Vn = Y \ f (X
\ Un ) is open in Y , T ⊂ Vn
and, noting that f −1 Y \ f (X
\ Un ) = X \ f −1 f (X \ Un ) ⊂ Un ,
f# µ(Vn ) ≤ µ(Un ) ≤ µ f −1 (T ) + 1/n = f# µ(T ) + 1/n. Taking
V = ∩n∈N Vn ∈ B Y , we then have T ⊂ V and f# µ(T ) = f# µ(V ),
what proves the assertion.
4) Finally, we prove that spt f# µ = f (spt
µ). Firstly,−1since 0 =
−1
f# µ(Y \spt f# µ) = µ X\f (spt f# µ) , and since X\f (spt f# µ)
is open in X, it follows that X \ f −1 (spt f# µ) ⊂ X \ spt µ, hence
(taking complements) spt µ ⊂ f −1 (spt f# µ), from what we con-
clude that f (spt µ) ⊂ spt f# µ.
On the other hand, f# µ Y \ f (spt µ) = µ X \ f −1 f (spt µ) ≤



µ(X \spt µ) = 0, hence spt f# µ is concentrated on f (spt µ). Since f
is closed, f (spt µ) is a closed subset of Y , thus spt f# µ ⊂ f (spt µ),
and we had already proved the other inclusion, whence the thesis.

Remark 1.38. Let X and Y be separable locally compact metric
spaces, µ a Radon measure on X and f : X → Y a Borelian map, i.e.
14 1. MEASURE AND INTEGRATION THEORY

such that ∀B ∈ B Y , f −1 (B) ∈ B X . If f is not a continuous proper

map, f# µ may not be a Radon measure on Y (it might not be finite on
compact sets, and even if it is, it might not be a Borel regular measure).
However, if we add the hypothesis that, for all K ⊂ Y , µ f −1 (K) <


∞, we may modify the definition of the pushforward in order to ensure
that f# µ be a Radon measure on Y . Instead of taking the pushforward
by f of the outer measure µ (i.e. the pushforward in the sense of
definition 1.14), we take the pushforward by f of the measure µ : B X →
[0, ∞], i.e. the measure f# µ on B Y given by A ∈ B Y 7→ µ f −1 (A) ,


which is finite on compact sets by the hypothesis assumed on f and µ.
Then, by remarks 1.29 and 1.33, f# µ is a Radon measure on Y . Both
definitions of f# µ coincide if f is a proper continuous map.
1.2. Measurable Maps
Definition 1.39 (Measurable spaces and measurable maps). A
measurable space is a pair (X, M) where X is a set and M is a σ-
algebra of subsets of X. The elements of M are called M-measurable
(or simply measurable, if M is clear from the context) subsets of X.
Given measurable spaces (X, M) and (Y, N ), a map f : X → Y is
called measurable with respect to M and N (or simply measurable, if
M and N are clear from the context) if, ∀A ∈ N , f −1 (A) ∈ M.
If X (or Y ) is a topological space, we shall tacitly assume that the
σ-algebra M is the Borel σ-algebra B X , unless another σ-algebra is
explicitly specified. Thus, for instance:
• For X and Y topological spaces, a map f : X → Y is called
Borelian or Borel measurable if it is measurable with respect
to B X and B Y .
• For X = R and Y a topological space (in particular, for Y = R
or C), a map f : X → Y is called Lebesgue measurable if it
is measurable with respect to L and B Y , where L is the
σ-algebra of Lebesgue measurable subsets of R.
Definition 1.40 (µ-measurable maps). Let µ be a measure on the
set X and Y a topological space. A function f : dom f ⊂ X → Y is
called measurable with respect to µ if the following conditions hold:
i) its domain covers almost all of X, i.e. µ(X \ dom f ) = 0,
ii) for all B ∈ B Y , f −1 (B) is µ-measurable.
Due to the fact that every subset of null measure of X is µ-measurable,
a map f : dom f ⊂ X → Y is measurable with respect to µ in the
sense of the definition above iff any extension of f to a a map X → Y is
measurable with respect to σ(µ) and B Y in the sense of definition 1.39.
 1.2. MEASURABLE MAPS 15

Moreover, if f is µ-measurable, any other function which coincides with

f except for a set of null measure is also µ-measurable.
We list below some of the main properties of measurable maps.
Theorem 1.41 (Properties of measurable maps). Let (X, M), (Y, N ),
(Z, O) be measurable spaces. The following properties hold:
i) f : X → Y is measurable iff given S ⊂ 2Y such that σ(S) = N ,
for all B ∈ S, f −1 (B) ∈ M.
ii) If f : X → Y and g : Y → Z are both measurable maps, so is g ◦ f .
iii) If X and Y are topological spaces and f : X → Y is continuous,
then it is Borelian.
iv) For Y = R, if (fn )n∈N is a sequence of measurable maps X → R,
the following maps X → R are measurable: inf n∈N fn , supn∈N fn ,
lim inf fn , lim sup fn . In particular, if (fn )n∈N is pointwise con-
vergent, the limit function is measurable. More generally, if Y is
a metric space and (fn )n∈N is a pointwise convergent sequence of
measurable maps X → Y , the limit function is measurable.
Proof.
i) The implication (⇒) is clear. On the other hand, N 0 := {T ⊂
Y | f −1 (T ) ∈ M} is clearly a σ-algebra of subsets of Y . Thus, if
S ⊂ N 0 , N = σ(S) ⊂ N 0 , i.e. f is measurable.
ii) ∀A ∈ O, (g ◦ f )−1 (A) = f −1 g −1 (A) ∈ M.

iii) Since f is continuous, ∀U ∈ τY , f −1 (U ) ∈ τX ⊂ σ(τX ) = B X . As

σ(τY ) = B Y , it suffices to apply part i to S = τY .
iv) Let g := inf n∈N fn . For all α ∈ R, g −1 ([α, ∞]) = ∩n∈N fn−1 ([α, ∞]) ∈
M. Since S := {[α, ∞] | α ∈ R} generates B R , it follows from part
i that g is measurable. Similarly, supn∈N fn is measurable, and so
are lim inf fn = supk∈N inf n≥k fn and lim sup fn = inf k∈N supn≥k fn .
Finally, let (fn )n∈N be a sequence of measurable maps X →
Y pointwise convergent to f . Then, for each open set U ⊂ Y ,
f −1 (U ) = ∪i∈N ∪j∈N ∩n≥j fn−1 ({x ∈ U | d(x, Y \ U ) ≥ 1/i}) ∈ M,
hence f is measurable.

Corollary 1.42. If f, g : X → R are both measurable, so are
max{f, g} and min{f, g}. In particular, both f + := max{f, 0} and
f − := max{−f, 0} are measurable.
Definition 1.43 (σ-algebra induced by a family of maps). Let
fα
X be a set, (Yα , Nα )α∈A a family of measurable spaces and (X −→
Yα )α∈A a family of maps defined on X. The smallest σ-algebra on X
for which ∀α ∈ A, fα is measurable (i.e. the intersection of the family
16 1. MEASURE AND INTEGRATION THEORY

of σ-algebras which make all fα ’s measurable

maps) is called σ-algebra
induced by (fα )α∈A , denoted by σ (fα )α∈A .
Proposition
1.44. With the notation from definition 1.43, let
M = σ (fα )α∈A .
i) If ∀α ∈ A, Nα =
σ(Sα ), then M = σ {V ⊂ X | ∃α ∈ A, ∃D ∈
Sα , V = fα−1 (D)} .
ii) If (Z, O) is a measurable space, then a map g : Z → X is measur-
able with respect to O and M iff ∀α ∈ A, fα ◦ g is measurable.
Proof.
i) Let N := σ {V ⊂ X | ∃α ∈ A, ∃D ∈ Sα , V = fα−1 (D)} . It follows


from theorem 1.41.i that ∀α ∈ A, fα is measurable with respect to
N and Nα . Hence, M ⊂ N , and the other inclusion follows from
the fact that {V ⊂ X | ∃α ∈ A, ∃D ∈ Sα , V = fα−1 (D)} ⊂ M.
ii) (⇒) follows from theorem 1.41 part ii. Conversely, if ∀α ∈ A,
fα ◦ g is measurable, then ∀α ∈ A, ∀D ∈ Nα , g −1 fα−1 (D) = (fα ◦
g)−1 (D) ∈ O, thus g is measurable by the previous item and by
theorem 1.41 part i.

Particular cases of the above construction are:
Product σ-algebra: Let (Xα , Mα )α∈AQbe a family of measur-
able spaces. On the product X = α∈A Xα , the σ-algebra
induced by the family of projections (prα : X → Xα )α∈A is
called product σ-algebra, denoted by ⊗α∈A Mα .
Pullback: Let (Y, N ) be a measurable space and f : X → Y .
The σ-agebra on X induced by {f } is called pullback of N ,
denoted by f ∗ N .
Note that f ∗ N = {f −1 (V ) : V ∈ N }. In particular, if
X ⊂ Y e f = i is the inclusion X → Y , the pullback i∗ N
coincides with {B ∩ X : B ∈ N }, called restriction or trace of
N on X, usually denoted by N |X . In this situation, if X ∈ N ,
then N |X = {B ∈ N | B ⊂ X}.
Remark 1.45. As an application of proposition 1.44, note that:
• a map taking values on a product of measurable spaces en-
dowed with the product σ-algebra is measurable iff each of its
components is measurable.
• if a map f takes values on a measurable space (Y, N ) and has
its image contained in a subset X, than f is measurable iff it
is measurable as a map taking values on X endowed with the
trace σ-algebra.
 1.2. MEASURABLE MAPS 17

• If X is a topological space and A ⊂ X, B X |A = B A , i.e. the

trace σ-algebra of B X on A coincides with the Borel σ-algebra
of A endowed with the relative topology.
Exercise 1.46. If X is a locally compact separable metric space,
µ a Radon measure on X and A ⊂ X is locally compact in the relative
topology (i.e. A is a locally closed subspace), then µ|A is a Radon
measure on A.
Proposition 1.47 (Product of Borel σ-algebras).
Q Let (Xα , τα )α∈A
be a family of topological spaces and X = α∈A Xα endowed with the
product topology. Then:
i) ⊗α∈A B Xα ⊂ B X .
ii) Equality holds in the previous item if A is countable and ∀α ∈ A, τα
is second countable.
Q
Proof. For each α ∈ A, the projection α∈A Xα → Xα is con-
tinuous. Hence, by theorem 1.41.iii, it is measurable with respect to
B X and B Xα . It then follows from proposition 1.44.ii that the iden-
tity X → X is measurable with respect to B X and ⊗α∈A B Xα , i.e.
⊗α∈A B Xα ⊂ B X .
On the other hand, assume that A is countable and ∀α ∈ A, τα is
second countable, so that the product topology on X is second count-
able.QWe may take a countable base for this topology formed by rectan-
gles α∈A Uα where each Uα is open in Xα and, except for finetely many
α’s, Uα = Xα . Since each such rectangle is measurable with respect
to ⊗α∈A B Xα , it follows that every open set in the product topology,
being a countable union of such rectangles, is measurable with respect
to ⊗α∈A B Xα , thus B X ⊂ ⊗α∈A B Xα . 
Corollary 1.48. For any n ∈ N, B Rn = ⊗n1 B R . In particular,
if (X, M) is a measure space, a map f = (f1 , . . . , fn ) : X → Rn is
measurable iff each component fi is measurable, 1 ≤ i ≤ n.
It follows from the corollary above, identifying C ≡ R2 as metric
spaces, that a function f : X → C is measurable iff both Re f and Im f
are measurable.
Corollary 1.49. Let (X, M) be a measurable space, Y a topo-
logical space, f1 , . . . , fn : X → R measurable maps and Φ : Rn → Y
Borelian. Then Φ(f1 , . . . , fn ) : X → Y is measurable. In particular,
sums, products and differences of measurable maps X → R are mea-
surable.
So is the quotient of measurable maps, as long as the denominator
is never zero, but see example 1.51, below.
18 1. MEASURE AND INTEGRATION THEORY

The following is a useful criterion for testing measurability in terms

of countable measurable covers.
Proposition 1.50. Let (X, M), (Y, N ) be measurable spaces, and
(An )n∈N a sequence in M such that ∪n∈N An = X. Then a map f :
X → Y is measurable iff ∀n ∈ N, f |An : An → Y is measurable, where
each An is endowed with the trace σ-algebra.
Proof. For each B ∈ N , f −1 (B) = ∪n∈N f |−1
An (B) ∈ M, since
∀n ∈ N, M|An ⊂ M, due to the fact that ∀n ∈ N, An ∈ M. 
The following example shows how the proposition above may be
applied:
Example 1.51.
1) Let sgn : C → C be defined by sgn z:= z/|z| if z 6= 0 and sgn 0 := 0.
Taking X = C, A1 = {0} and A2 = C \ {0} in proposition 1.50, it
is clear that sgn is measurable (note that the trace σ-algebra on A2
coincides with its Borel σ-algebra as a metric subspace of C, and
that sgn |A2 is continuous). Thus, if (X, M) is a measurable space,
each measurable function f : X → C admits a polar decomposition
f = sgn f · |f |, where each factor is measurable.
2) Let + : R×R → R be arbitrarily defined on {(+∞, −∞), (−∞, +∞)}
and in the usual way on the complement of this set. Taking A1 =
{(+∞, −∞), (−∞, +∞)}, A2 = R × {+∞} ∪ {+∞} × R, A3 =
R × {−∞} ∪ {−∞} × R and A4 = R × R in preposition 1.50, is
clear that + is Borelian. Thus, if (X, M) is a measurable space and
f, g : X → R are measurable maps, so is f + g. We can treat sim-
ilarly the difference, product and quotient of extended real valued
measurable maps.
We now focus our attention in a class of measurable maps which
will be used to develop the integration theory of a measure µ on a set
X.
Definition 1.52 (Simple functions). Let X and Y be measurable
spaces. A function ϕ : X → Y is called simple if it is measurable and
has finite image.
In the next section we shall be concerned with simple functions
taking values in R or C.
Proposition 1.53 (properties of simple functions). Let (X, M) be
a measure space.
i) The set of all C-valued simple functions on X is a subalgebra of
CX .
 1.2. MEASURABLE MAPS 19

ii) If f : X → [0, ∞] is measurable, there exists an increasing sequence

(ϕn )n∈N of simple functions X → [0, ∞) which converges pointwise
to f and such that the convergence is uniform on each part where
f is bounded.
iii) If f : X → C is measurable, there exists a sequence (ϕn )n∈N of
simple functions X → C which converges pointwise to f , and such
that ∀n ∈ N, |ϕn | ≤ |ϕn+1 | ≤ |f | and the convergence is uniform
on each part where f is bounded.
Exercise 1.54. Let (X, M) be a measurable space, f : XP→ [0, ∞]
measurable, (rn )n∈N a sequence in (0, ∞) such that rn → 0 eP ∞i=1 rn =
n
∞. Then there exists a sequence (An )n∈N in M such that k=1 rk χAk
increases pointwise to f .
Hint. Define (Ak )k∈N and (gk )k∈N inductively by: 1) A1 := {x ∈
X | r1 ≤ f (x)} and g1 := r1 χA1 ; 2) Ak := {x ∈ X | gk−1 (x)+rk ≤ f (x)}
and gk := gk−1 + rk χAk .
Exercise 1.55. Let (X, M) be a measurable space, Y a separable
metric space and f : X → Y a measurable map. Then there exists a
sequence of simple functions X → Y which converges pointwise to f .
We end this section with a definition of support for measurable
functions on a topological space endowed with a Borel measure which
is often more natural from a measure-theoretic point of view than the
usual definition of support.
Definition 1.56 (Support and essential support). Let X be a topo-
logical space endowed with a Borel measure µ and f a measurable
function on X.
i) The support of f , denoted by spt f , is the complement in X of the
union of all open sets on which f is null1.
ii) The essential support of f , denoted by ess spt f , is the complement
in X of the union of all open sets on which f is µ-a.e. null.
Remark 1.57 (Support and essential support). With the notation
from the previous definition:
1) It is clear that ess spt f ⊂ spt f .
2) If f is continuous on X and spt µ = X, then ess spt f = spt f .
3) If X is second countable, or if X is locally compact Hausdorff and µ
is Radon, then ess spt f is the complement of the biggest open set
on which f is null µ-a.e.
1Actually this definition makes sense for arbitrary functions on topological
spaces, not necessarily endowed with measures.
20 1. MEASURE AND INTEGRATION THEORY

4) We adopt the convention that, henceforth, “support of f ” means

“essential support of f ”, which will be denoted accordingly by “spt f ”.

1.3. Integration Theory

Up to the end of this section, we fix a measure µ on the set X. The
restriction of µ to σ(µ) yields a classical measure space (X, σ(µ), µ),
for which an integration theory is developed in standard Real Analysis
textbooks. For the sake of completeness, we list some definitions and
theorems below and refer the reader elsewhere for more details.
In the theory of integration described below we consider measurable
functions on X taking values in R or C. We denote by L+ (µ) the set
of µ-measurable functions on X taking values in [0, ∞].
Definition 1.58. For a simple function ϕ ∈ L+ (µ), i.e. for ϕ simple
and taking values in P[0, ∞), let Im ϕ = {a1 , . . . , an }, with ai 6= aj if
n
i 6= j, so that ϕ = i=1 ai χϕ−1 (ai ) (which is the so-called standard
form or standard representation of the simple function ϕ). We define
the integral of ϕ with respect to µ by:
ˆ X n
ai µ ϕ−1 (ai ) ∈ [0, ∞],


ϕ dµ :=
i=1

where we use the convention 0 · ∞ := 0.

For an arbitrary f ∈ L+ (µ), we now define:
ˆ ˆ
f dµ := sup{ ϕ dµ | ϕ ∈ L+ simple , ϕ ≤ f } ∈ [0, ∞].
+
´
One can´ check that, ´ ´ f, g ∈ L´ (µ) and c ∈ [0, ∞), (f +
whenever
g) dµ = f dµ + g dµ and cf dµ = c f dµ.
For a µ-measurable function f taking values in R, we consider the
positive and negative parts of f , i.e. f + = max{f, 0} and f − =
max{−f, 0}; according to corollary 1.42, they are both measurable
+ − + −
and satisfy ´ f+ = f − f , ´|f | −= f + f . We say that f´ is inte-
´ + if ´ f + dµ < ∞ or f dµ < ∞; if so, we define
grable ´ f dµ:=
f ´dµ− f dµ ∈ R. We say that f is´summable if both f + dµ < ∞
and f − dµ´ < ∞ (or, equivalently, if |f | dµ < ∞), i.e. if f is inte-
grable and f dµ ∈ R.
As it is usual, henceforth we omit the “µ” in the notation whenever
the measure ´ is clear´ from the context. For a measurable set E ⊂ X,
we define E f := χE f .
´ Finally, a µ-measurable C-valued function f is called summable if
|f | < ∞ (or, equivalently, if both real and imaginary parts of f are
 1.3. INTEGRATION THEORY 21
´ ´ ´
summable). For such a function, we define f := Re f + i Im f ∈
C. We denote by L1 (µ) the set of summable functions f : X → C.
If µ is the counting measure P
on a set X and´f is an integrable func-
tion on X, we use the notation x∈X f (x) for f dµ, called unordered
sum of f .
Exercise 1.59. Let If µ be the counting measure on a set X and
f : X → [0, ∞]. Then
X nX o
f (x) = sup f (x) | F ⊂ X finite .
x∈X x∈F
P
Moreover, if x∈X f (x) < ∞, then {x ∈ X | f (x) > 0 is countable.
Warning. Some authors use the nomenclature “almost integrable”
for what we have called “integrable” and “integrable” for what we have
called “summable”.
We summarize the main properties of the integral defined above
in the theorems that follow. As it is usual, we say that a property P
which refers to points of X holds µ-almost everywhere (or simply almost
everywhere if the measure is clear from the context), with notation “P
µ-a.e.” or “P a.e. [µ]”, if the set of the points at which P does not
hold has measure zero.
Theorem 1.60 (properties of the integral). The following proper-
ties for the integral defined in 1.58 hold:
i) L1 (µ) is a complex vector space and the integral is a linear func-
tional on it. ´
ii) If f ∈ L+ , then f = 0 iff f = 0 µ-a.e. ´ ´
iii) If f and g ´are integrable
´ and f = g a.e., then f = g. If f ≤ g
a.e., then f ≤ g. ´ ´
iv) (integral triangle inequality)
´ If f ∈ L1 , then | f | ≤ |f |.
v) ∀f ∈ L1 (µ), kf k1 := |f | dµ defines a seminorm on L1 (µ).
It follows from ii, above, that the linear subspace N := {f ∈ L1 (µ) |
kf k1 = 0} of L1 (µ) consists of the measurable functions on X which
are null almost everywhere. The elements of the quotient L1 (µ)/N are,
therefore, classes of equivalence of summable functions which coincide
almost everywhere, and k·k1 is a norm on this quotient. The fact
that this norm is complete (so that L1 (µ)/N is a Banach space) is a
consequence of the convergence theorems 1.62 and 1.64 stated below.
Remark 1.61. As it is usual, we shall, henceforth, overload the
notation “L1 (µ)”, which will be used both with its original meaning and
also to denote the aforementioned quotient space. That is, whenever
22 1. MEASURE AND INTEGRATION THEORY

we write “f ∈ L1 (µ)”, it may signify, depending on the context, that f

is a summable function or that f is a class of equivalence of summable
functions which coincide almost everywhere. A similar remark applies
to the Lp spaces, to be introduced in subsection 1.3.1, below.
Theorem 1.62 (monotone convergence theorem). Let (fn )n∈N be
+ +
an increasing
´ ´sequence in L (µ), which converges µ-a.e. to f ∈ L (µ).
Then fn → f .
Theorem´ 1.63 (Fatou’s lemma).
´ Let (fn )n∈N be a sequence in
L+ (µ). Then lim inf fn ≤ lim inf fn .
Theorem 1.64 (dominated convergence theorem). Let (fn )n∈N be
a sequence in L1 (µ) dominated by a summable function g, i.e such that
∀n ∈ N, |fn | ≤ g. If (f´n )n∈N converges
´ pointwise almost everywhere to
1
f , then f ∈ L (µ) and fn → f .
Corollary 1.65. With the same hypothesis, fn → f in L1 (µ).
Proof. |fn −f | converges pointwise almost everywhere´ to zero and
the convergence is dominated by 2g, hence kfn − f k1 = |fn − f | →
0. 

Thus, dominated pointwise almost everywhere convergence implies

convergence in L1 . On the other hand, without additional hypothe-
ses we cannot recover pointwise almost everywhere convergence from
L1 convergence, but we can ensure the pointwise almost everywhere
convergence of a subsequence, i.e. if (fn )n converges to f in L1 (µ),
there exists a subsequence of (fn ) which converges pointwise almost
everywhere to f .
The following improved version of the dominated convergence the-
orem often comes in handy:
Theorem 1.66 (generalized dominated convergence theorem). Let
(fn )n∈N and (gn )n∈N be sequences in L1 (µ) such that:
i) ∀n ∈ N, |fn | ≤ gn µ-a.e.
ii) f´n → f ´pointwise a.e. and gn → g pointwise a.e.
iii) gn → g < ∞.
´ ´
Then f ∈ L1 (µ) and fn → f .
An important application of the dominated convergence theorem
is related to the study of continuity and differentiability of functions
defined by integrals. For instance, the following proposition is a direct
consequence of the dominated convergence theorem:
 1.3. INTEGRATION THEORY 23

Proposition 1.67 (differentiation under the integral sign using

the dominated convergence theorem). Let I ⊂ R be a nondegenerate
interval and f : X × I → R´such that ∀t ∈ I, f (·, t) ∈ L1 (µ). Let
F : I → R defined by F (t) := f (x, t) dµ(x).
i) Let t0 ∈ I and suppose that ∀x ∈ X, ∃ limt→t0 f (x, t) and there
exists g ∈ L1 (µ) such that |f (x, t)| ≤ g(x) for all (x, t). Then
ˆ ˆ
lim f (x, t) dµ(x) = lim f (t, x) dµ(x).
t→t0 t→t0

In particular, F is continuous at t0 if ∀x ∈ X, f (x, ·) is continuous

in t0 .
ii) Suppose that exists ∂f
∂t
and there exists g ∈ L1 (µ) such that ∂f
∂t
(x, t) ≤
g(x) for all (x, t). Then F is differentiable and
ˆ
0 ∂f
F (t) = (x, t) dµ(x).
∂t
Similar statements hold if we replace the parameter interval I
by an open subset of Rk .
Exercise 1.68 (upper and lower integrals). We call ϕ : X → C an
extended simple function if it is µ-measurable and its image is count-
able. For a function f : X → [0, ∞], not necessarily measurable, we
define:
´∗ ´
• (upper integral ) f dµ:= inf{ ϕ dµ | ϕ ∈ L+ extended simple, ϕ ≥
f a.e.} ∈ [0, ∞],´ ´
• (lower integral ) ∗ f dµ:= sup{ ϕ dµ | ϕ ∈ L+ extended simple, ϕ ≤
f a.e.} ∈ [0, ∞].
Prove that:
´ ´∗ ´
a) If ∗ f = f < ∞, then f is µ-measurable and f coincides with
1
both upper and lower integrals
´ ´∗ f ∈
(hence ´ L ).
b) If f is µ-measurable, then ∗ f = f = f.
c) The monotone convergence theorem holds for the upper integral,
i.e. if (fn )n∈N is a sequence of positive functions (not necessarily
measurable) which increases
´∗ µ-a.e.
´∗ to a function f (not necessarily
measurable), then fn → f . Similarly, Fatou’s lemma also
holds for the upper integral.
d) If (fn )n∈N is a sequence
´ ∗ of positive functions (not necessarily mea-
surable) such that fn → 0, there exists a subsequence of (fn )n∈N
which converges pointwise almost everywhere to zero.
Exercise 1.69. Let µ be a measure on a set X and A ⊂ X.
´
x
a) ´If f ∈ L+ (µ), then f ∈ L+ (µ A), f |A ∈ L+ (µ|A ) and f d(µ xA) =
f |A d(µ|A ).
24 1. MEASURE AND INTEGRATION THEORY

b) ´If A ∈ σ(µ), both integrals in the previous item coincide with

A
f dµ.
Exercise 1.70. Let µ be a measure on the set X and f : X →
+ +
´ a map into´ the set Y . If g ∈ L (f# µ), then g ◦ f ∈ L (µ) and
Y
g d(f# µ) = g ◦ f dµ.
Remark 1.71. In the previous exercise, if X and Y are topological
spaces and f : X → Y is a Borelian map, the same same statement
holds for a Borelian function g ≥ 0 on Y if we take the alternative
definition of the pushforward from remark 1.38, i.e. we take the push-
forward by f of the measure µ on B X , which is a measure f# µ on
B Y , and then we take the extension of this measure given by theorem
1.7 (the alternative definition may be more convenient in this situation
because it yields a Borel regular measure). In fact, both definitions of
f# µ coincide on B Y , and the integrals depend only on the measures
on the Borel sets, i.e. they depend only on µ : B X → [0, ∞] and
f# µ : B Y → [0, ∞].
1.3.1. Lp spaces.
Definition 1.72. Let f be a C-valued measurable function on X.
We define:
´
• For real 0 < p < ∞, kf kp := ( |f |p dµ)1/p ∈ [0, ∞].
• For p = ∞, kf kp := inf{C ∈ R | |f | ≤ C µ − a.e. on X} ∈
[0, ∞] (note that inf ∅ = +∞).
For 0 < p ≤ ∞, we define Lp (µ):= {f : X → C µ − measurable |
kf kp < ∞}.
For p ∈ [1, ∞], we define its conjugate exponent p0 ∈ [1, ∞] by
1
p
+ p10 = 1 (thus p0 = ∞ for p = 0 and p0 = 1 for p = ∞).

For each real 0 < p < ∞ or p = ∞, one can readily check that
Lp (µ) is a vector space over C. For 1 ≤ p ≤ ∞, it follows from theorem
1.75, stated below, that k·kp is a seminorm on Lp (µ).
Theorem 1.73 (Hölder’s inequality). For any p ∈ [1, ∞], f, g C-
valued measurable functions on X, the following inequality holds:
kf gk1 ≤ kf kp kgkp0 .
0
In particular, f g ∈ L1 (µ) if f ∈ Lp (µ) and g ∈ Lp (µ).
Theorem 1.74 (Generalized Hölder’s inequality). Let p1 , . . . , pk ∈
[1, ∞] such that ki=1 p1i = 1r ≤ 1 and f1 , . . . , fk C-valued measurable
P
 1.3. INTEGRATION THEORY 25

functions on X. Then
k
Y k
Y
k fi kr ≤ kf kpi .
i=1 i=1
Qk
In particular, i=1 fi ∈ Lr (µ) if fi ∈ Lpi (µ) for 1 ≤ i ≤ k.
Theorem 1.75 (Minkowski’s inequality). For any p ∈ [1, ∞], f, g
C-valued measurable functions on X, the following inequality holds:
kf + gkp ≤ kf kp + kgkp .
For 1 ≤ p ≤ ∞, the linear subspace N := {f ∈ Lp (µ) | kf kp =
0} of Lp (µ) consists of the measurable functions on X which are null
almost everywhere. Therefore, the quotient Lp (µ)/N consists of classes
of equivalence of functions in Lp (µ) which coincide almost everywhere,
and k·kp is a norm on this quotient, which is complete by the following
theorem. As in remark 1.61, we shall henceforth overload the notation
“Lp (µ)”, which will be used both with its original meaning and also to
denote the aforementioned quotient space.
Theorem 1.76. For 1 ≤ p ≤ ∞, Lp (µ) is a Banach space. For
p = 2, it is a Hilbert ´space, since k·k2 is induced by the Hermitian
inner product hf, gi := f ḡ dµ (where ¯· denotes complex conjugation),
whenever f, g ∈ L2 (µ).
For 1 ≤ p < ∞, the theorem above is a consequence of the conver-
gence theorems for the integral 1.62, 1.63, 1.64.
We now state a basic interpolation theorem which may be derived
by a convenient application of Hölder’s inequality.
Theorem 1.77 (Basic interpolation for Lp spaces). If 0 < p < q <
r ≤ ∞, then Lp (µ) ∩ Lr (µ) ⊂ Lq (µ) and, for all measurable f on X,
kf kq ≤ kf kλp kf k1−λ
r , where λ ∈ (0, 1) is defined by
1 λ 1−λ
= + .
q p r
The following density theorem is a consequence of the regularity
properties of Radon measures (1.28, 1.31). For a locally compact Haus-
dorff space X, we denote by Cc (X) the space of continuous functions
on X with compact support.
Proposition 1.78. If µ is a Radon measure on a locally compact
Hausdorff space X and 1 ≤ p < ∞, then Cc (X) is dense in Lp (µ).
26 1. MEASURE AND INTEGRATION THEORY

Proof. Since Lp simple functions are dense in Lp (as it can be read-

ily checked by means of proposition 1.53 and theorem 1.64), it suffices
to prove that such functions may be arbitrarily approximated in the
Lp norm by continuous functions with compact support. Besides, since
any Lp simple function is a finite linear combination of characteristic
functions of measurable sets of finite measure, it suffices to show that,
given E ∈ σ(µ) with µ(E) < ∞ and  > 0, there exists φ ∈ Cc (X)
such that kφ − χE kp < . Indeed, take a compact set K ⊂ E and
an open set U ⊃ E such that µ(U \ K) < δ, with δ > 0 to be cho-
sen later. Applying Urysohn’s lemma, choose φ ∈ Cc (X) such that
0 ≤ φ ≤ 1, φ ≡ 1 on a neighborhood of K and spt φ ⊂ U . Therefore,
χK ≤ φ ≤ χU and χK ≤ χE ≤ χU , so that |φ − χE | ≤ χU − χK , what
implies kφ − χE kp ≤ kχU − χK kp = δ 1/p . Taking δ 1/p < , the thesis is
achieved. 
We now identify the dual space of the Banach space Lp (µ). For
fixed p ∈ [1, ∞], q = p0 the conjugate exponent´ of p and f ∈ Lp (µ),
q
let Φp (f ) : L (µ) → C be defined by g 7→ f g dµ. It follows from
Hölder’s inequality 1.73 that Φp (f ) is well defined, Φp (f ) ∈ Lq (µ)0
and kΦp (f )kq ≤ kf kp . Actually, in “almost all” situations the last
inequality is an equality, and Φp is an isometry of Lp (µ) onto Lq (µ):
Theorem 1.79 (Riesz representation theorem). With the notation
above, if 1 < p < ∞, Φp is an isometry of Lp (µ) onto Lq (µ)0 , so that
we may identify by means of this isometry Lq (µ)0 ≡ Lp (µ).
• For p = ∞, if µ is σ-finite, Φ∞ is an isometry of L∞ (µ) onto
0 0
L1 (µ) , so that L1 (µ) ≡ L∞ (µ).
• For p = 1, Φ1 is an isometry of L1 (µ) into L∞ (µ)0 , but in
general it is not onto, i.e. in general the dual of L∞ (µ) is
bigger than L1 (µ).
We end this subsection with a criterion for compacity in Lp (Ln ).
For a function f defined on Rn and x, y ∈ Rn , we adopt the usual
notation for translations:
τy f (x) := f (x − y)
Theorem 1.80 (Kolmogorov-Riesz-Fréchet). Let 1 ≤ p < ∞ and
F be a bounded subset of Lp (Ln ) such that
lim kτh f − f kp = 0
h→0

uniformly in f ∈ F. Then, for each Ω ∈ σ(Ln ) with finite measure,

the closure of F|Ω in Lp (Ln |Ω ) is compact.
Here, F|Ω := {f |Ω | f ∈ F}.
 1.4. PRODUCT MEASURES AND FUBINI-TONELLI’S THEOREM 27

1.3.2. Change of variables formula. We state in the next two

theorems the version of the change of variables formula for the Lebesgue
integral which is usually presented in Real Analysis textbooks. That
formula will be generalized in chapter 5 by the area and coarea formu-
las.
Theorem 1.81 (linear change of variables for the Lebesgue inte-
gral). Let T ∈ GL(n, R).
i) If A ∈ L Rn , then T · A ∈ L Rn and Ln (T · A) = |det T |Ln (A);
ii) If f : Rn → R is Lebesgue-measurable, so is f ◦ T , and, if f ≥ 0
or f ∈ L1 , ˆ ˆ
f dL = f ◦ T |det T | dLn .
n

Theorem 1.82 (C1 -change of variables formula for the Lebesgue

integral). Let U ⊂ Rn open and φ : U → Rn be a C1 diffeomorphism
onto its image V := φ(U ) (which is an open subset of Rn ). If f is
a Lebesgue-measurable function on V , f ◦ φ is a Lebesgue-measurable
function on U ; besides, if f ≥ 0 or f ∈ L1 , then
ˆ ˆ
n
f dL = f ◦ φ(x)|det Dφ(x)| dLn (x).
V U

1.4. Product measures and Fubini-Tonelli’s theorem

If (X, M, µ) and (Y, N , ν) are measure spaces, there exists a stan-
dard construction, based on Carathéodory’s extension theorem, which
yields a measure µ × ν on the product σ-algebra M ⊗ N ⊂ 2X×Y ,
called product measure of µ and ν. If both µ and ν are σ-finite,
µ × ν is characterized by the property of being the unique measure
on M × N such that , for all measurable rectangles A × B ∈ M × N ,
µ×ν(A×B) = µ(A)ν(B). The main tool used in the study and compu-
tation of integrals with respect to µ × ν is the classical Fubini-Tonelli’s
theorem, which relates integrals with respect to µ × ν to iterated inte-
grals with respect to µ and ν.
We now describe how to make an analogous construction for the
product of outer measures µ on a set X and ν on a set Y . We may
define the product µ × ν as the extension given by theorem 1.8 of the
product (in the sense of the previous paragraph) µ|σ(µ) ⊗ ν|σ(ν) . That
is equivalent to the definition below.
Definition 1.83 P(product measure). We define, for all E ⊂ X ×
Y , µ × ν(E) := inf{ n∈N µ(An )ν(Bn ) | ∀n ∈ N, An ∈ σ(µ), Bn ∈
σ(ν), E ⊂ ∪n∈N An × Bn } ∈ [0, ∞]. Recall that we use the convention
0 · ∞ = 0. We call µ × ν the product measure of µ and ν.
28 1. MEASURE AND INTEGRATION THEORY

We make a similar definition for any finite number of measures.

The theorem below, which may be obtained as a direct consequence
of classical Fubini-Tonelli’s theorem for products of measures on σ-
algebras, ensures that µ × ν it is indeed a measure. Note that µ × ν is
a regular measure, but we do not assume the regularity of µ or ν. We
use the following:
Notation.
• For E ⊂ X ×Y and (x0 , y0 ) ∈ X ×Y , Ex0 := {y ∈ Y | (x0 , y) ∈
E} (the x0 -section of E) and Ey0 := {x ∈ X | (x, y0 ) ∈ E}
(the y0 -section of E).
• For a function f defined on dom f ⊂ X × Y and (x0 , y0 ) ∈
X × Y , fx0 (the x0 -section of f ) and fy0 (the y0 -section of
f ) are the functions defined, respectively, on (dom f )x0 and
(dom f )y0 by y 7→ f (x0 , y) and x 7→ f (x, y0 ).
Theorem 1.84 (Fubini-Tonelli’s for outer measures, [Fed69], [EG91]).
With the notation from the previous definition, µ × ν : 2X×Y → [0, ∞]
is a regular measure. Moreover:
i) If A ∈ σ(µ) and B ∈ σ(ν), then A × B ∈ σ(µ × ν) and µ × ν(A ×
B) = µ(A)ν(B).
ii) If E ∈ σ(µ × ν) is σ-finite with respect to µ × ν, then, for µ-almost
every x ∈ X, Ex ∈ σ(ν), and for ν-almost every y ∈ Y , Ey ∈ σ(µ).
The functions x 7→ ν(Ex ) and y 7→ µ(Ey ) are measurable, and the
measure of E may be computed by:
ˆ ˆ
µ × ν(E) = ν(Ex ) dµ(x) = µ(Ey ) dν(y).

iii) If f is an integrable function defined on dom f ⊂ X × Y such that

{f 6= 0} is σ-finite with respect to µ×ν (what holds, in particular, if
f is summable), then, for µ-almost every x ∈ X, fx is ν-integrable,
and for ν-almost every y ∈ Y , fy ´is µ-integrable. The ´ almost
everywhere defined
´ functions x 7→ fx dν and y 7→ fy dµ are
integrable, and f d(µ × ν) may be computed by iterated integrals:
ˆ ˆ ˆ ˆ ˆ



f d(µ × ν) = fx dν dµ(x) = fy dµ dν(y).

Remark 1.85. If µ and ν are both σ-finite, then so is µ × ν, so that

the σ-finiteness hypotheses in parts ii and iii above are automatically
fulfilled. Moreover, every positive measurable function is integrable,
so that part iii holds for such functions (what corresponds to classical
Tonelli’s theorem).
 1.4. PRODUCT MEASURES AND FUBINI-TONELLI’S THEOREM 29

Example 1.86. We show that the Lebesgue measure Ln on Rn

coincides with the product measure (L1 )n = L1 ×· · ·×L1 . Fix A ⊂ Rn .
1) As defined in example 1.3, Ln (A) = inf{ Q∈A vol(Q) | A countable
P

cover of A by cubes with sides parallel to the coordinate axes}.

Since any cube Q ∈ Rn is a product of intervals (hence a product
of L1 -measurable sets) and since, for each such cube, the euclidean
volume vol(Q) coincides with (L1 )n (Q) (by the fact that the length
of an interval coincides with its Lebesgue measure), we immediately
conclude from definition 1.83 that (L1 )n (A) ≤ Ln (A).
2) In the definition of Ln (A), we may use rectangles (i.e. products of
arbitrary intervals) instead of cubes (products of intervals with the
same side length), without modifying Q Ln (A). Indeed, it suffices to
show that, for any such rectangle R = nj=1 Ij and  > 0, R may be
P
covered by a countable family A of cubes such that Q∈A vol(Q) ≤
vol(R) + . In order to accomplish that, assume vol(R) < ∞ (oth-
erwise we are done) and, for m ∈ N (to be chosen later), cover
k
each interval Ij by countably many disjoint intervals (Ij,m )k∈N with
side lengths equal to 1/m so that k∈N L1 (Ij,m k
) − L1 (Ij ) < 1/m.
P
Q k1 kn


Then Am := I1,m × · · · × In,m n ∈N
is a countable cover of R
P Qkn1 ,...,kP 1 k

m→∞
by cubes and Q∈Am vol(Q) = j=1 k∈N L (Ij,m ) −→ vol(R);
thus, for m sufficiently large, A = Am does the job.
3) In view of the previous item, to prove the remaining Qn inequality
1 n n
(L ) (A) ≥ L (A), it suffices to show that, given B = j=1 Bj with
(∀1 ≤ j ≤ n) Bj ∈ L R , for all  > 0, there exists aP countable fam-
ily A of rectangles which covers B and such that Q∈A vol(Q) ≤
1 1
Q Q
1≤j≤n L (B j ) + . We assume that 1≤j≤n L (Bj ) < ∞, oth-
erwise the inequality is trivial. Recall that L1 is a Borel regu-
lar measure, as we have seen in example 1.19; actually, it is a
Radon measure, by exercise 1.32. It then follows from theorem
1.23 that, for any m ∈ N and for 1 ≤ j ≤ n, there
exist open sets
1
Uj,m ⊂ R such that Bj ⊂ Uj,m and L Uj,m \ Bj < 1/m. Since each
open set in R is a countable disjoint union of open intervals, there
k
exists a countable family (Ij,m )k∈N of disjoint open intervals such
k
Q k1 kn


˙
that Uj,m = ∪k∈N Ij,m ; take Am := I1,m × · · · × In,m .
k1 ,...,kn ∈N
Then A m is a countable family of rectangles which covers B and
m→∞ Q
= nj=1 L1 (Uj,m ) −→ 1
P Q
R∈Am vol(R) 1≤j≤n L (Bj ); thus, for m
sufficiently large, A = Am does the job.

Exercise 1.87 (Layer-cake formula, [LL01]). Let µ be a σ-finite

measure on X and ν a Radon measure on [0, ∞). Define φ : [0, ∞) → R
30 1. MEASURE AND INTEGRATION THEORY


by φ(t) = ν [0, t) . Then, for every f ∈ L+ (µ):
ˆ ˆ
φ ◦ f dµ = µ({f > t}) dν(t).
[0,∞)

In particular, if p > 0 and ν = ptp−1 dt, it follows:

ˆ ˆ ∞
p
f (x) dµ(x) = p µ({f > t})tp−1 dt.
0

Hint. Compute the integral on the first member by means of Fubini-

Tonelli’s theorem.
We close this section with a useful generalization of Minkowski’s
inequality 1.75 which may be obtained as a corollary of Fubini-Tonelli’s
theorem:
Theorem 1.88 (Minkowski’s inequality for integrals). Let (X, M, µ)
and (Y, N , ν) be σ-finite measure spaces and f a M ⊗ N -measurable
function on X × Y .
i) If f ≥ 0 and 1 ≤ p < ∞, then
hˆ ˆ p i1/p ˆ hˆ i1/p
p
f (x, y) dν(y) dµ(x) ≤ f (x, y) dµ(x) dν(y).

ii) If p ∈ [1, ∞], the inequality below holds if the second member
makes sense and is finite. That is, if for ν-a.e. y ∈ Y f (·, y) ∈
Lp (µ) and the a.e. defined ν-measurable function y 7→ kf (·, y)kp
is ν-summable, then for µ-a.e. x ∈ X, the function f´ (x, ·) is ν-
summable, the a.e. defined µ-measurable function x 7→ f (x, y) dν(y)
is in Lp (µ) and
ˆ ˆ
k f (·, y) dν(y)kp ≤ kf (·, y)kp dν(y).

1.5. Signed measures and Lebesgue-Radon-Nikodym

theorems
In this subsection we are concerned with measures on σ-algebras
(i.e. we don’t consider outer measures). We recall the notion of “signed
measure” and some important decomposition theorems which may be
used to relate the properties and the integration theory of one measure
on a σ-algebra to the corresponding properties of another measure on
the same σ-algebra.
Definition 1.89 (signed measures). A charge or signed measure
on a measurable space (X, M) is a set function ν : M → R such that
SM1) ν(∅) = 0;
 1.5. SIGNED MEASURES AND LEBESGUE-RADON-NIKODYM THEOREMS 31

SM2) Im ν ⊂ [−∞, ∞) or Im ν ⊂ (−∞, ∞] (i.e. ν omits −∞ or +∞);

SM3) ν is σ-additive, i.e. for all countable disjoint family (An )n∈N in
M, X
ν(∪n∈N An ) = ν(An ),
n∈N
with the meaning that n 7→ µ(An ) is summable with respect to
the counting measure on N and the sum is µ(∪n∈N An ).
We say that a signed measure ν is finite if Im ν ⊂ R (i.e. if ν omits
both −∞ and +∞). We say that ν is σ-finite if there exists a sequence
(An )n∈N in M such that ∪n∈N An = X and ∀n ∈ N, ν(An ) ∈ R.
Example 1.90.
1) Let µ1 and µ2 be measures on (X, M) such that µ1 or µ2 is finite.
Then ν = µ1 − µ2 is a signed measure on (X, M).
2) Let µ be a measure on (X, M) and f : X → R an integrable function
´ sense of definition 1.58). Then f µ : M → R given by
(in the
A 7→ A f dµ is a signed measure.
Remark 1.91.
1) The second example is a particular case of the first, since f µ =
f + µ − f − µ. We will see in theorem 1.94 and in exercise 1.96 that
every signed measure on (X, M) may be written in both forms 1)
and 2).
2) Note that every measure on (X, M) is a signed measure. As it is
usual, for clarity reasons, sometimes we call a measure on (X, M)
a positive measure, to contrast with “signed measure”.
Definition 1.92 (absolute continuity and mutual singularity). Let
µ and ν be positive measures on a measurable space (X, M). We say
that:
1) µ is absolutely continuous with respect to ν (notation: µ  ν) if
∀A ∈ M, ν(A) = 0 implies µ(A) = 0.
2) µ and ν are mutually singular (notation: µ ⊥ ν) if there exists
A ∈ M such that µ is concentrated on A and ν is concentrated on
X \ A.
Exercise 1.93. Let µ be a finite positive measure and ν a positive
measure on a measurable space (X, M). Then µ  ν iff ∀ > 0,
∃δ > 0, ∀A ∈ M, ν(A) < δ implies µ(A) < .
Theorem 1.94 (Jordan decomposition theorem). Let ν be a signed
measure on a measure space (X, M). Then there are unique positive
measures ν + and ν − such that ν = ν + − ν − and ν + ⊥ ν − .
32 1. MEASURE AND INTEGRATION THEORY

Definition 1.95 (positive part, negative part and total variation

of a signed measure). With the notation from theorem 1.94, we call ν +
the positive part of ν and ν − the negative part of ν.
The positive measure |ν|:= ν + + ν − is called the total variation of
ν.
Exercise 1.96. If ν is a signed measure on a measure space (X, M),
there exists a Borelian |ν|-integrable function f : X → R such that
|f | ≡ 1 and ν = f |ν|.
Definition 1.97 (integration with respect to a signed measure).
Let ν be a signed measure on a measure space (X, M) and f : X → R a
measurable function. We say that f is summable with respect to ν if it is
summable with respect to |ν| and we use the notation L1 (ν) := L1 (|ν|).
For such f , we define
ˆ ˆ ˆ
f dν := f dν − f dν − .
+

Note that the integral defined above satisfies the usual linearity
and convergence properties, which are inherited from the corresponding
properties of the integrals with respect to ν + and ν − .
Definition 1.98 (absolute continuity and mutual singularity, bis).
Let ν be a signed measure and µ a positive measure on a measurable
space (X, M). We say that:
1) ν  µ if |ν|  µ.
2) ν ⊥ µ if |ν| ⊥ µ.
Exercise 1.99. Let ν be a signed measure on a measurable space
(X, M). The following properties hold:
a) For all A ∈ M, |ν(A)| ≤ |ν|(A).
b) ν is finite (respectively, σ-finite) iff ν + and ν − are finite (respectively,
σ-finite) iff |ν| is finite (respectively, σ-finite). If ν is finite, Im ν is
a bounded subset of R.
c) For all A ∈ M,
X
|ν|(A) = sup{ |ν(An )| | ∀n ∈ N, An ∈ M and ∪˙ An = A}.
n∈N
n∈N

d) L1 (|ν|) = L1 (ν + ) ∩ ´L1 (ν − ). ´
e) For all f ∈ L1 (ν), | f dν| ≤ |f | d|ν|.
Exercise 1.100. Let ν1 and ν2 be signed measures on a measurable
space (X, M) such that both omit −∞ or both omit +∞, µ a positive
measure (X, M) and c ∈ R. Then:
 1.5. SIGNED MEASURES AND LEBESGUE-RADON-NIKODYM THEOREMS 33

a) cν1 and ν1 + ν2 are signed measures.

b) |cν1 | = |c||ν1 | and |ν1 + ν2 | ≤ |ν1 | + |ν2 |, with equality if |ν1 | ⊥ |ν2 |.
c) If ν1 ⊥ µ and ν2 ⊥ µ, then ν1 + ν2 ⊥ µ. If both ν1 and ν2 are
positive measures, then ν1 ⊥ µ and ν2 ⊥ µ iff ν1 + ν2 ⊥ µ.
d) If ν1  µ and ν2  µ, then ν1 + ν2  µ. If both ν1 and ν2 are
positive measures, then ν1  µ and ν2  µ iff ν1 + ν2 ´µ.
e) L1 (ν1 )´∩ L1 (ν2 ) ⊂ 1 1 1
´ L (ν1 + ν2 ) and, ∀f ∈ L (ν1 ) ∩ L (ν2 ), f d(ν1 +
ν2 ) = f dν1 + f dν2 . ´ ´
f) If c 6= 0, L1 (cν1 ) = L1 (ν1 ) and, ∀f ∈ L1 (ν1 ), f d(cν1 ) = c f dν1 .
Theorem 1.101 (Lebesgue decomposition theorem). Let ν be a
signed measure and µ a positive measure on a measurable space (X, M),
both σ-finite. Then there exist unique signed measures νs and νa on
(X, M) such that νs ⊥ µ, νa  µ and ν = νs + νa .
Definition 1.102. With the notation from the previous theorem,
we call νs the singular part of ν, νa the absolutely continuous part of ν
and ν = νs + νa the Lebesgue decomposition of ν with respect to µ.
Theorem 1.103 (Radon-Nikodym theorem). Let µ be a positive
measure and ν a σ-finite signed measure on a measurable space (X, M),
such that ν  µ. Then there exists a µ-integrable function f : X → R,
unique up to µ-null sets in M, such that ν = f µ, i.e. for all A ∈ M,
ˆ
ν(A) = f dµ.
A

Definition 1.104 (Radon-Nikodym derivative). With the notation

from the previous theorem, we cal f (or any measurable function which
coincides µ-a..e with f ) the Radon-Nikodym derivative of ν with respect
dν
to µ and denote it by dµ .

Note that, in the situation of example 1.90.2), i.e. if ν = f µ where

µ is a positive measure on (X, M) and f : X → R is µ-integrable,
then ν  µ. Hence, if ν is σ-finite, it follows from the uniqueness of
dν
the Radon-Nikodym derivative stated in theorem 1.103 that f = dµ
(equality here and in similar statements below means that f is in the
dν
equivalence class of dµ modulo µ-a.e. null functions).
The Radon-Nikodym derivative has the properties suggested by the
dν
notation dµ . We list those properties in exercise 1.105 and proposition
1.107 below.
Exercise 1.105. Let (X, M) be a measurable space and µ a posi-
tive measure on (X, M).
34 1. MEASURE AND INTEGRATION THEORY

a) If ν is a σ-finite signed measure on (X, M) and ν  µ, then

d|ν| dν
= .
dµ dµ
In other words, if f : X → R is µ-integrable and ν = f µ, then
dν
|ν| = |f |µ. Moreover, ν is finite iff dµ ∈ L1 (µ).
b) Let ν1 and ν2 be σ-finite signed measures on (X, M) such that both
omit −∞ or both omit +∞ and c ∈ R. Suppose that ν1  µ and
ν2  µ. Then
d(cν1 ) dν1 d(ν1 + ν2 ) dν1 dν2
=c and = + .
dµ dµ dµ dµ dµ
dν
c) If ν is a σ-finite signed measure on (X, M) and ν  µ, then dµ
=
d(ν + ) d(ν − ) d(ν + ) dν + d(ν − ) dν −



dµ
− dµ
, dµ
= dµ
and dµ
= dµ
.
Remark 1.106. Let ν be a signed measure on (X, M). It is imme-
diate from defintion 1.98 that ν  |ν|. If ν is σ-finite, it then follows
dν
from exercise 1.105.a) with |ν| in place of µ that d|ν| = ±1 |ν|-a.e. on
X.
Proposition 1.107 (chain rule for the Radon-Nikodym deriva-
tive). Let λ, ν and µ be positive measures on a measurable space (X, M)
with λ and ν σ-finite and such that λ  ν  µ. Then:
i) For every f : X → [0, ∞] measurable,
ˆ ˆ
dν
f dν = f dµ.
dµ
ii) λ  µ and
dλ dλ dν
= · .
dµ dν dµ
1.6. Convolutions
In this section we consider integrals with respect to the Lebesgue
measure in Rn , which we often denote by dx, dy, etc. We recall the
basic properties of convolutions and mollifiers, which will be extensively
used in subsequent chapters.
Let f, g : Rn → C be Ln -measurable functions. We define the
convolution f ∗ g by:
ˆ
f ∗ g(x) := f (x − y)g(y) dy,

whenever the integral makes sense at least for Ln -a.e. x ∈ Rn . That

occurs mainly in two cases:
 1.6. CONVOLUTIONS 35

• one of the functions is essentially bounded, the other belongs

to L1loc (Ln ) and one of them has compact support;
• one of the functions belong to Lp (Ln ) and the other belongs
to Lq (Ln ), with p1 + 1q ≥ 1, cf. proposition 1.108.g) below.
Broadly speaking, whenever defined, the convolution product is
commutative and associative, and inherits the regularity properties
from both factors. The latter property is widely explored in techniques
which involve approximation of functions by means of mollifiers.
We summarize the main properties of the convolution product in
the propositions below. For a function f defined on Rn and x, y ∈ Rn ,
we use the notation:

τy f (x) := f (x − y) and fˇ(x) := f (−x)

as well as the standard multi-index notation for partial derivatives, i.e.

given α = (α1 , . . . , αn ) ∈ Zn+ ,
 ∂  α1  ∂  αn
∂ α f (x) := ... f (x),
∂x1 ∂xn
|α| := α1 + · · · + αn .

Proposition 1.108 (properties of the convolution product). For

f, g, h : Rn → C such that the convolution products are defined:
a) f ∗ g is Ln -measurable.
b) f ∗ g = g ∗ f.
c) (f ∗ g) ∗ h = f ∗(g ∗ h).
d) spt f ∗ g ⊂ spt f + spt g.
e) (f ˇ∗ g) = fˇ ∗ ǧ.
f) ∀y ∈ Rn , τy (f ∗ g) = (τy f ) ∗ g = f ∗(τy g).
1 1
g) (Young’s inequality) If p, q, r ∈ [1, ∞] with p
+ q
= 1 + 1r , then

kf ∗ gkr ≤ kf kp kgkq .

Thus, f ∗ g ∈ Lr (Ln ) if f ∈ Lp (Ln ) and g ∈ Lq (Ln ). In particular,

for p = 1 and q = r ∈ [1, ∞], kf ∗ gkq ≤ kf k1 kgkq .
If p and q are conjugate exponents, i.e. if we take p, q, r above
with r = ∞, then f ∗ g(x) exists for every x ∈ Rn and f ∗ g is
bounded and uniformly continuous; besides, if both p and q are finite,
then f ∗ g ∈ C0 (Rn , C).
h) If f ∈ L1loc (Ln ) and g ∈ L∞ (Ln ) with spt g compact, then f ∗ g ∈
L1loc (Ln ).
36 1. MEASURE AND INTEGRATION THEORY

i) If f and g satisfy the hypothesis of one of the two previous items

and ϕ ∈ Cc (Rn , C), then
ˆ ¨ ˆ
f ∗ g(x)ϕ(x) dx = f (x)g(y)ϕ(x + y) dx dy = f (x)ǧ ∗ ϕ(x) dx.

j) Let 0 ≤ k ≤ ∞ and α ∈ Zn a multi-index with |α| ≤ k. If f ∈

Ckb (Rn ) and g ∈ L1 (Ln ), or f ∈ Ck (Rn ), g ∈ L1loc (Ln ) and one of them
has compact support, then f ∗ g ∈ Ck (Rn ) and ∂ α (f ∗ g) = (∂ α f ) ∗ g.
Recall our convention adopted in remark 1.57, i.e. for measurable
functions “support” means “essential support”.
The proof of the proposition above can be found in standard real
analysis textbooks, but we offer a proof of part j) as an application of
the dominated convergence theorem.
Proof of part j). We prove the assertion for k = 1 and α = ej ,
1 ≤ j ≤ n. The general case follows by induction using the same
argument.
1) Suppose f ∈ C1b and g ∈ L1 . We have
∂
[f (x − y)g(y)] = ∂xj f (x − y)g(y)
∂xj
hence ∀x, y ∈ Rn , ∂x∂ j [f (x−y)g(y)] ≤ k∂xj f ku |g(x)|. Since g ∈ L1 ,
we may differentiate under the integral sign using the dominated
convergence theorem 1.67.ii:
ˆ
∂xj (f ∗ g)(x) = ∂xj f (x − y)g(y) dy = (∂xj f ) ∗ g(x).

Moreover, since ∂xj f ∈ L∞ and g ∈ L1 , it follows from the last state-

ment in part g) that (∂xj f ) ∗ g is continuous (actually it is uniformly
continuous). Since that holds for all 1 ≤ j ≤ n, we conclude that
f ∗ g ∈ C1 , as asserted.
2) Suppose that f ∈ C1 (Rn ), g ∈ L1loc (Ln ) and one of them, say f , has
compact support (the case spt g b Rn is similar).
Fix x0 ∈ Rn and r > 0. Let K be the compact set B(x0 , r)−spt f
(so that, for x ∈ B(x0 , r), x−y ∈ spt f implies y ∈ K). We have, for
all (x, y) ∈ U(x0 , r) × Rn , ∂x∂ j [f (x − y)g(y)] = |∂xj f (x − y)||g(y)| ≤
k∂xj f ku χK |g|. Since χK |g| ∈ L1 (Ln ), we may apply proposition 1.67
with U(x0 , r) in place of I and (Rn , Ln ) in place of (X, µ), yielding,
for all x ∈ U(x0 , r),
ˆ
∂xj (f ∗ g)(x) = ∂xj f (x − y)g(y) dy = (∂xj f ) ∗ g(x).
 1.6. CONVOLUTIONS 37

Since (∂xj f ) ∗ g coincides on U(x0 , r) with (∂xj f ) ∗(χK g) (where

K = B(x0 , r) − spt f , as above), ∂xj f ∈ L∞ and χK g ∈ L1 , we
conclude that (∂xj f ) ∗ g is continuous on U(x0 , r). As that holds
for 1 ≤ j ≤ n, it follows that f ∗ g is continuously differentiable on
U(x0 , r); since x0 ∈ Rn and r > 0 were arbitrarily taken, we are
done.

Lemma 1.109. If f ∈ C0 (Rn ), then f is uniformly continuous.
Proof. Fix  > 0 and let K ⊂ Rn compact such that |f | ≤  on
K c . Since f |K is uniformly continuous, there exists δ > 0 such that
|f (x) − f (y)| ≤  whenever x, y ∈ K with kx − yk < δ. If x, y ∈ Rn
and kx − yk < δ, we have:
1) if x, y ∈ K, |f (x) − f (y)| ≤ ;
2) if x, y ∈ K c , |f (x) − f (y)| ≤ 2;
3) if x ∈ K and y ∈ K c , the closed segment [x, y] intersects ∂K, hence
there exists z ∈ ∂K ⊂ K such that kz−xk < δ. Since, by continuity,
|f | ≤  in ∂K, we have |f (x)−f (y)| ≤ |f (x)−f (z)|+|f (z)−f (y)| ≤
|f (x) − f (z)| + |f (z)| + |f (y)| ≤ 3.
Hence, for all x, y ∈ Rn with kx − yk < δ, |f (x) − f (y)| ≤ 3. 
Lemma 1.110. If 1 ≤ p < ∞, translation is continuous in the
L norm, i.e. for fixed f ∈ Lp (Ln ), the map Rn → Lp (Ln ) given by
p

y 7→ τy f is continuous.
Proof. Fix z ∈ Rn . We must prove that limy→z kτy f − τz f kp = 0.
Fix  > 0. Since Cc (Rn ) is dense in Lp (Ln ) (by proposition 1.78), there
exists g ∈ Cc (Rn ) such that kf − gkp ≤ . Then kτy f − τz f kp ≤
kτy (f − g)kp + kτy g − τz gkp + kτz (g − f )kp ≤ 2 + kτy g − τz gkp . Since
Cc (Rn ) ⊂ C0 (Rn ), it follows from lemma 1.109 that g is uniformly
continuous; hence
y→z
kτy g − τz gkp ≤ kτy g − τz gku Ln (spt g) → 0.
We therefore conclude that lim supy→z kτy f − τz f kp ≤ 2, whence the
thesis, since  > 0 was arbitrarily taken. 
In the next theorem we use the following notation: for φ : Rn → C
and t > 0, we define φt : Rn → C by
(1.1) φt (x) := t−n φ(t−1 x).
Note that,
´ if φ ∈ L´1 (Ln ), it follows from theorem 1.81 that φt ∈
L1 (Ln ) and φ dLn = φt dLn .
38 1. MEASURE AND INTEGRATION THEORY
´
Theorem 1.111 (mollifiers, part I). Let φ ∈ L1 (Ln ) with φ dLn =
a and f : Rn → C.
t→0
i) If 1 ≤ p < ∞ and f ∈ Lp (Ln ), then φt ∗ f → af in Lp (Ln ).
ii) If f is uniformly continuous and either (1) f is bounded or (2)
t→0
spt φ is compact, then φt ∗ f → af uniformly in Rn .
iii) If f is continuous on an open set U ⊂ Rn and either (1) f ∈
t→0
L∞ (Ln ) or (2) f ∈ L∞ n
loc (L ) and spt φ is compact, then φt ∗ f → af
uniformly on compact subsets of U .
Proof.
i) ∀t > 0, ∀x ∈ Rn ,
ˆ
z=t−1 y
f ∗ φt (x) − af (x) = [f (x − y) − f (x)]φt (y) dy =
ˆ
= [f (x − tz) − f (x)]φ(z) dz =
ˆ
= [τtz f (x) − f (x)]φ(z) dz.

Thus, by Minkowski’s inequality for integrals 1.88,

ˆ
kf ∗ φt − af kp ≤ kτtz f − f kp |φ(z)| dz.

For any sequence tn → 0, kτtn z f − f kp |φ(z)| converges pointwise to

0, by force of lemma 1.110, and kτtn z f − f kp |φ(z)| ≤ 2kf kp |φ(z)|.
Since φ ∈ L1 (Ln ), we may therefore ´apply the dominated conver-
n→∞
gence theorem 1.64 to conclude that kτtn z f − f kp |φ(z)| dz → 0,
hence kf ∗ φtn − af kp → 0. Since the sequence tn → 0 was ar-
bitrarily taken, it follows that kf ∗ φt − af kp → 0 as t → 0, as
asserted.
ii) By the same computation from the previous item, ∀t > 0, ∀x ∈ Rn ,
ˆ
f ∗ φt (x) − af (x) = [τtz f (x) − f (x)]φ(z) dz.
´
Thus kf ∗ φt − af ku ≤ kτtz f − f ku |φ(z)| dz. Since f is uniformly
t→0
continuous, we have kτtz f − f ku → 0 for all z ∈ Rn . If (1)
holds, then kτtz f − f ku |φ(z)| ≤ 2kf ku |φ(z)| for all z ∈ Rn ; if (2)
holds, by the compacity of spt φ and by the uniform continuity of
f we may take δ > 0 such that for all 0 < t < δ and all z ∈ spt φ,
kτtz f −f ku ≤ 1, whence kτtz f −f ku |φ(z)| ≤ |φ(z)| for all 0 < t < δ
and all z ∈ Rn . In either case, the dominated convergence theorem
ensures that kf ∗ φt − af ku → 0 as t → 0, as asserted.
 1.6. CONVOLUTIONS 39

iii) ´Fix  > 0. Let K ⊂ U compact and C ⊂ Rn compact such that

Rn \C
|φ| dLn ≤  (which exists, since φ ∈ L1 ). Take η > 0 such that
η · sup{|x| | x ∈ C} < d(K, U c ); then K 0 := {x − ty | x ∈ K, y ∈
C, |t| ≤ η} is a compact subset of U which contains K. Since
f |U is continuous, by compacity it follows that f |K 0 is uniformly
continuous; therefore, there exists 0 < δ < min{η, 1} such that
sup{|f (x − ty) − f (x)| | x ∈ K, y ∈ C, |t| < δ} < . Hence, if
|t| < δ, ∀x ∈ K:
ˆ
|f ∗ φt (x) − af (x)| = [τty f (x) − f (x)]φ(y) dy ≤
ˆ
≤ |τty f (x) − f (x)||φ(y)| dy =
ˆ ˆ
= |τty f (x) − f (x)||φ(y)| dy + |τty f (x) − f (x)||φ(y)| dy ≤
Rn \C C
≤ 2kf kL∞ (K 00 ) + kφk1 .
where K 00 = Rn in case (1) or K 00 = K + B(0, sup{|y| | y ∈ spt φ})
in case (2) (recall that δ < 1).
Since  > 0 was arbitrarily taken, we therefore conclude that
k(f ∗ φt − af )|K ku = sup{|f ∗ φt (x) − af (x)| | x ∈ K} → 0 as
t → 0.

In the applications of the previous theorem, the most important
cases are a = 1 and a = 0. For a = 1, we call (φt )>0 an approximate
identity or mollifier , since it can be used to approximate a function f
on Rn by means of the convolutions φt ∗ f , which have the same class
of regularity as φ, in the appropriate topology of the function space f
belongs to.
For instance, we may take (φt )>0 given by the following definition:
Definition 1.112 (standard mollifier in Rn ). Let φ : Rn → R be
the smooth function given by
(
c exp kxk12 −1


if kxk < 1
φ(x) :=
0 if kxk ≥ 1,
´
where c is chosen so that Rn φ(x) dx = 1. The family (φt )>0 induced
by φ by means of (1.1) is called standard mollifier in Rn .
Note that spt φ = B(0, 1), so that ∀t > 0, spt φt = B(0, t).
Exercise 1.113. Show that φ in the definition above is a C∞ func-
tion on Rn .
40 1. MEASURE AND INTEGRATION THEORY

Exercise 1.114 (differentiable Urysohn’s lemma). If K ⊂ U ⊂ Rn

with K compact and U open, there exists f ∈ C∞ n
c (R ) such that 0 ≤
f ≤ 1, f ≡ 1 on K and spt f ⊂ U .
Hint. Let K 0 := K + B(0, 12 d(K, U c )) and (φt )>0 the standard
mollifier in Rn . Take f = φt ∗ χK 0 for a convenient choice of t.
Exercise 1.115 (approximation in L1loc ). Let 1 ≤ p < ∞, f ∈
Lploc (Ln ) and
(φt )>0 the standard mollifier in Rn . Then φt ∗ f → f in
Lploc (Ln ), i.e.
for all K ⊂ Rn compact, kφt ∗ f − f kLp (Ln |K ) → 0.
Hint. For each K ⊂ Rn compact, let K 0 := K +B1 and fe := χK 0 ·f .
Then fe ∈ Lp (Ln ) and, by theorem 1.111, φt ∗ fe → fe in Lp (Ln ).
Remark 1.116. The previous exercise means that φt ∗ f converges
to f in the Fréchet topology of Lploc (Ln ).
We will resume this discussion on mollifiers and approximations in
chapter 6.
1.7. Lusin’s and Egorov’s Theorems
Theorem 1.117 (Lusin, [Fed69]). Let µ be a Borel regular mea-
sure on a metric space X (respectively, a Radon measure on a locally
compact Hausdorff space X), Y a separable metric space, f : dom f ⊂
X → Y a µ-measurable map. Then, for each A ∈ σ(µ) with µ(A) < ∞
and for each  > 0, there exists a closed (respectively, compact) set
C ⊂ A such that µ(A \ C) <  and f |C is continuous.
Proof. We may assume dom f = X (otherwise, replace A by
A ∩ dom f and extend f arbitrarily to X). For each i ∈ N, there
exists a countable disjoint sequence (Yi,j )j∈N in B Y such that ∀j ∈ N,
diam Yi,j < 1/i and ∪˙ j∈N Yi,j = Y . To obtain such a sequence, cover
Y with countably many balls (Bi,j ) with diameter less than 1/i (what
is possible, since Y is separable) and then take ∀j ∈ N, Yi,j := Bi,j \
∪j−1
n=1 Bi,n . Let ∀j ∈ N, Ai,j := A ∩ f
−1
(Yi,j ) ∈ σ(µ), so that A =
˙∪j∈N Ai,j . By theorem 1.23 (respectively, exercise 1.31), there exists a
closed (respectively, compact) set Ci,j ⊂ Ai,j such that ∀j ∈ N, µ(Ai,j \
Ci,j ) < 2−i−j . Then µ(A \ ∪j∈N Ci,j ) < 2−i  and, by proposition 1.11.ii
(continuity from above for µ), there exists J(i) ∈ N such that µ(A \
J(i)
∪j=1 Ci,j ) < 2−i . Now, for each 1 ≤ j ≤ J(i), choose yi,j ∈ Yi,j
and define the function gi on the closed (respectively, compact) set
J(i)
Ci := ∪j=1 Ci,j by gi |Ci,j ≡ yi,j . It is then clear that gi : Ci → Y is
continuous and, due to the fact that diam Y
i,j < 1/i, we have (denoting
by d the metric on Y ) sup{d f (x), gi (x) | x ∈ Ci } ≤ 1/i. Take
 1.7. LUSIN’S AND EGOROV’S THEOREMS 41

C := ∩P i∈N Ci , so that C is closed (respectively, compact) and µ(A \

C) ≤ i∈N µ(A
\ Ci ) ≤ . As ∀i ∈ N, C ⊂ Ci , we have ∀i ∈ N,
sup{d f (x), gi (x) | x ∈ C} ≤ 1/i, hence (gi |C )i∈N converges uniformly
to f |C , so f |C is continuous. 
Corollary 1.118. With the same hypotheses, if µ is σ-finite, f
coincides µ-a.e. with a Borelian map X → Y .
Proof. Let (Ai )i∈N be a sequence in σ(µ) of disjoint sets such
that ∀i ∈ N, µ(Ai ) < ∞ and X = ∪˙ i∈N Ai . For each i, j ∈ N, we
may apply theorem 1.117 to obtain a closed set Ci,j ⊂ Ai such that
µ(Ai \ Ci,j ) < 1/j and f |Ci,j is continuous. Then B := ∪i,j∈N Ci,j is a
Borel set such that µ(X \ C) = 0 and, for each E ∈ B Y , (f |B )−1 (E) =
∪i,j∈N (f |Ci,j )−1 (E) ∈ B X . Choose y0 ∈ Y and define F : X → Y by
F ≡ y0 on X \ B and F = f on B. Then F is Borelian and F = f
µ-a.e. 
Corollary 1.119. Let µ be a σ-finite Borel regular measure on
a metric space X (respectively, a σ-finite Radon measure on a locally
compact Hausdorff space X), Y = R or C, f : dom f ⊂ X → Y a
µ-measurable function. Then there exists a sequence (fn )n∈N in C(X)
(respectively, in Cc (X)) which converges pointwise µ-a.e. to f . More-
over, if kf k∞ < ∞, we can take this sequence (fn )n∈N so that ∀n ∈ N,
kfn k∞ ≤ kf k∞ .
Proof. We may assume, modifying f on a set of measure zero if
necessary, that dom f = X and ∀x ∈ X, |f (x)| ≤ kf k∞ . Let (Ai )i∈N
be a sequence in σ(µ) of disjoint sets such that ∀i ∈ N, µ(Ai ) < ∞
and X = ∪˙ i∈N Ai . For each i, j ∈ N, we may apply theorem 1.117 to
obtain a closed (respectively, compact) set Ci,j ⊂ Ai such that µ(Ai \
Ci,j ) < 1/j and f |Ci,j is continuous. For each n ∈ N, let Cn be the
closed (respectively, compact) set ∪ni,j=1 Ci,j . Since the last union is
finite, f |Cn is continuous; use Tietze’s extension theorem to extend
f |Cn to a function fn ∈ C(X) (respectively, fn ∈ Cc (X)). Note that,
if kf k∞ < ∞, we may and do take the extension fn so that kfn k∞ ≤
kf k∞ . The sequence (fn )n∈N thus defined converges pointwise to f on
∪n∈N Cn = ∪i,j∈N Ci,j . Since µ(X \ ∪i,j∈N Ci,j ) = 0, we are done. 
Definition 1.120. Let µ be a measure on the set X, Y a metric
space and A ⊂ X. We say that a sequence (fn )n∈N of µ-measurable
Y -valued functions on X converges almost uniformly on A to a µ-
measurable function f : dom f ⊂ X → Y if, for all  > 0, there exists
B ∈ σ(µ) such that µ(A \ B) <  and (fn )n∈N converges uniformly to
f on B.
42 1. MEASURE AND INTEGRATION THEORY

Note that, both in the definition above and in the theorem below,
we do not assume A to be µ-measurable.
Theorem 1.121 (Egorov). Let µ be a measure on the set X, Y a
separable metric space, A ⊂ X with µ(A) < ∞ and (fn )n∈N a sequence
of Y -valued measurable functions on X which converges pointwise µ-
a.e. on A to a µ-measurable function f : dom f ⊂ X → Y . Then
(fn )n∈N converges almost uniformly to f on A.
Proof. Fix  > 0 and denote by d the
metric on Y . For i, j ∈ N,
define Ci,j := ∪n≥j {x ∈ X | d fn (x), f (x) ≥ 1/i}. Note that each Ci,j
is µ-measurable, since d : Y × Y → R is continuous (hence
Borelian)
and, for fixed n ≥ j, x ∈ dom fn ∩ dom f 7→ fn (x), f (x) ∈ Y × Y
is µ-measurable (it is in this point that we use the separability
of Y ,
what ensures B Y ×Y = B Y ⊗ B Y ), so that x 7→ d fn (x), f (x) is µ-
measurable and Ci,j is a countable union of µ-measurable sets. Hence,
for fixed i ∈ N, it follows from proposition 1.15.i that (Ci,j )j∈N is a
x
decreasing sequence of µ A-measurable sets. Since µ(A) < ∞, and
x
since ∩j∈N Ci,j has µ A-measure zero due to the fact that (fn )n∈N con-
verges µ-a.e. on A to f , we may apply the continuity from above 1.11
x x j→∞
for the measure µ A to conclude that µ A(Ci,j ) −→ 0. Therefore,
x
for fixed i ∈ N, there exists J(i) ∈ N such that µ A(Ci,J(i) ) < 2−i .
Let B :=P X \ ∪i∈N Ci,J(i) . Then B ∈ σ(µ), µ(A \ B) = µ ( ∪i∈Nx
x
Ci,J(i) ) ≤ i∈N µ (Ci,J(i) ) <  and (fn )n∈N converges uniformly to f
on B. 
We close this section with an application of Egorov’s theorem. Re-
call that, for a Banach space X, the weak topology of X is the topology
induced by its dual X 0 , i.e. the weakest topological vector space topol-
ogy on X which makes all elements of X 0 continuous.
Theorem 1.122 (theorem 1.35 in [AFP00]). Let µ be a measure
on the set X and 1 < p < ∞. If (fn )n∈N is a bounded sequence in
Lp (µ) which converges pointwise almost everywhere to a function f ,
then f ∈ Lp (µ) and (fn )n∈N converges weakly to f .
 CHAPTER 2

Hausdorff Measures

2.1. Carathéodory’s construction

Lebesgue measure on Rn is not adequate to study “lower dimen-
sional” objects, such as embedded k-dimensional manifolds or, more
generally, their measure-theoretic cousins, the k-rectifiable sets. For
that purpose, we shall introduce Hausdorff measures and dimension,
a class of Borel regular measures whose origins may be traced back to
[Hau18] and [Car14].
In order to construct Hausdorff measures, we depart from an ab-
stract construction on a metric space X which may be used to generate
a plethora of Borel measures on X with nice geometric flavor. We shall
be concerned only with Hausdorff measures here, but the interested
reader may consult, for instance, [Fed69], [Mat95] or [KP08] for
other such measures as well.
Let X be a metric space, F ⊂ 2X and ζ : F → [0, ∞]. Roughly
speaking, the idea is to “measure” the elements of F by means of
the method or gauge ζ and use that to define a Borel measure on
X, abstracting the geometric idea underlying the construction of the
Lebesgue measure. We define such measure in two steps:
1) For 0 < δ ≤ ∞ we define ∀A ⊂ X,
X
ψδ (A) := inf{ ζ(S) | G ⊂ F∩{S | diam S ≤ δ}, G countable cover of A}.
S∈G

Note that inf ∅ = ∞, so that ψδ (A) = ∞ if there is no countable

cover of A by elements of F with diameter ≤ δ. Moreover, if A = ∅,
G = ∅ ⊂ F is such a cover and, since the sum over the empty family
is zero (as we defined the sum by means of the integral with respect
to counting measure at the end of definition 1.58), we conclude
that ψδ (∅) = 0. Besides, it is straightforward to check that ψδ is
monotone and countably subbaditive, i.e. it is a measure according
to definition 1.1. In general, the measures (ψδ )δ thus defined are not
Borel, but we can fabricate a Borel measure out of them, as we do
in the second step of the construction:
2) Define, for each A ⊂ X, ψ(A) := sup{ψδ (A) | 0 < δ ≤ ∞} ∈ [0, ∞].
43
44 2. HAUSDORFF MEASURES

Note that, for fixed A ⊂ X, {ψδ (A)}δ is decreasing in δ, so that

the sup in the definition above coincides with limδ→0 ψδ (A).
Definition 2.1. With the notation above, we call ψ the result of
Carathéodory’s construction from the gauge ζ on F, and we call ψδ the
size δ approximating measure.
We prove in the next proposition that ψ is actually a Borel measure.
Proposition 2.2. Let X be a metric space and ψ be the result of
Carathéodory’s construction from the gauge ζ on F ⊂ 2X . Then ψ is
a Borel measure. Besides, if F ⊂ B X , ψ is Borel regular.
Proof. We denote by d the metric on X. That ψ is a measure
follows directly from the fact that, for each δ ∈ (0, ∞], ψδ is a measure.
In order to prove that ψ is Borel, we verify the Carathéodory’s criterion
1.18. Let A, B ⊂ X such that d(A, B) = δ > 0 and ψ(A ∪ B) < ∞; we
must show that ψ(A∪B) ≥ ψ(A)+ψ(B). Indeed, for any η < δ/2, cover
A ∪ B by a countable family G ⊂ F whose elements have diameters
≤ η. Note that no element of G intersects both A an B, thanks to the
triangle inequality. Thus, discarding the elements of G which do not
meet A or B, we obtain a subcover G 0 ⊂ G of A∪B that may be written
0
P union G =P
as a disjoint G1 ∪˙ G2 , wherePG1 covers A P and G2 covers
B. Then S∈G ζ(S) ≥ S∈G 0 ζ(S) = S∈G1 ζ(S) + S∈G2 ζ(S) ≥
ψη (A) + ψη (B). By the arbitrariness of G ⊂ F whose elements have
diameters ≤ η, we conclude that ψη (A ∪ B) ≥ ψη (A) + ψη (B), for all
η < δ/2. Hence, taking η → 0, it follows that ψ(A∪B) ≥ ψ(A)+ψ(B),
as asserted.
Assume now that F ⊂ B X . We contend that ψ is Borel regular.
Indeed, let A ⊂ X such that ψ(A) < ∞; we must prove the existence
of B ∈ B X such that B ⊃ A and ψ(B) = ψ(A). For each δ > 0, we
can take Bδ ∈ B X such that Bδ ⊃ A and ψδ (Bδ ) = ψδ (A): choose,
for each n ∈ N, a countable Pcover Gn ⊂ F of A whose elements have
diameters ≤ δ, such that S∈Gn ζ(S) < ψδ (A) + 1/n, and then put
Bδ := ∩n∈N ∪S∈Gn S. Define B := ∩n∈N B1/n ∈ B X . Then B ⊃ A and,
for each n ∈ N, ψ1/n (A) ≤ ψ1/n (B) ≤ ψ1/n (B1/n ) = ψ1/n (A), so that
ψ1/n (A) = ψ1/n (B), and taking n → ∞ yields ψ(A) = ψ(B). 
Definition 2.3. Let X be a metric space and m a nonnegative
real number. Take F = 2X and ζ : 2X → [0, ∞] given by
(diam S)m
ζ(S) := α(m) ,
2m
π m/2
where α(m) = Γ(m/2+1) (i.e. the euclidean volume of Bm if m integer).
The result of Carathéodory’s construction from the gauge ζ on 2X is
 2.1. CARATHÉODORY’S CONSTRUCTION 45

called Hausdorff m-dimensional measure on X, denoted by Hm . We

use the notation Hδm for the size δ approximation of Hm .

Proposition 2.4 (immediate properties of Hausdorff measure).

Let X be a metric space and m a nonnegative real number. The fol-
lowing properties hold for Hm :
1) The Hausdorff measure is compatible with the operation of taking
traces. That is, if X is a metric space and A ⊂ X, the trace of Hm
on A coincides with the m-dimensional Hausdorff measure on A (as
a metric subspace of X).
2) The Hausdorff measure is invariant by isometries. That is, if Y is
another metric space and f : X → Y is an isometry onto Y , then
the pushforward f# Hm coincides with the Hausdorff m-dimensional
measure on Y .
3) If Y is another metric space and f :
X → Y has Lipschitz constant
Lip f < ∞, then ∀A ⊂ X, Hm f (A) ≤ (Lip f )m Hm (A).
4) Hm also coincides with the result of Carathéodory’s construction
from ζ (same gauge as in definition 2.3) on F 0 = {closed subsets of
X} or F 00 = {open subsets of X}. If X is a normed vector space,
we may also take F 000 = {closed convex subsets of X}.
5) Hm is a Borel regular measure on X.
6) H0 coincides with the counting measure on X.

Proof. Properties 1) and 2) are immediate; property 2) is also a

direct consequence of 3) since, if f is an isometry onto Y , then Lip f =
Lip f −1 = 1.
3) Let A ⊂ X, δ ∈ (0, ∞] and G a countable cover of A by subsets of
diameter ≤ δ. Then f (G) := {f (S) | S ∈ G} is a countable cover
of f (A) by subsets of diameter ≤ δ · Lip f since, for each S ⊂ X,
diam f (S) ≤ diam S · Lip f . Hence,

X
m
HδmLip f f (A) ≤ α(m)2−m diam f (S) ≤
S∈G
X
≤ (Lip f )m α(m)2−m (diam S)m .
S∈G

By the arbitrariness

of G, taking the infimum we conclude that
HδmLip f f (A) ≤ (Lip f )m Hδm (A); thus, taking δ → 0, the thesis
follows.
4) It is clear that we may use F 0 = {closed subsets of X} instead of
F = 2X , since the diameters are not affected by taking closures. If X
is a normed vector space, the same argument works for F 000 = {closed
46 2. HAUSDORFF MEASURES

convex subsets of X}, since diameters are not affected by taking

closed convex hulls either.
As to the remaining case, let ψ be the result of Carathéodory’s
construction from ζ on F 00 = {open subsets of X}. Since F 00 ⊂
F = 2X , it is clear that Hm ≤ ψ. To prove the reverse inequal-
ity, define, for each S ⊂ X and δ > 0, Sδ := ∪x∈S U(x, δ/2); note
that Sδ is open (for it is a union of open balls) and diam Sδ ≤
diam S + δ. Given δ > 0 and A ⊂ X such that Hm (A) < ∞,
for any countable cover P G = {Sn | n ∈ N} of A by subsets of
diameter ≤ δ such that n∈N α(m)2−m (diam Sn )m < ∞, and for
any 0 <  ≤ δ, G := {(Sn )2−n  | n ∈ N} ⊂ F 000 is a countable
open
P cover of−mA by subsets of
diameter
m P ≤ 2δ. Therefore, ψ2δ (A) ≤
n∈N α(m)2 diam (Sn )2−n  ≤ n∈N α(m)2 (diam Sn +2−n )m .
−m

Taking  → 0 along any decreasing sequence, we Pmay apply −mtheorem

1.64 to conclude thatP the last sum converges to n∈N α(m)2 (diam Sn )m ,
−m
so that ψ2δ (A) ≤ n∈N α(m)2 (diam Sn )m . By the arbitrariness
of the cover G, taking the infimum we conclude that ψ2δ (A) ≤
Hδm (A); thus, taking δ → 0, it follows that ψ(A) ≤ Hm (A), as
asserted.
5) Since F 0 ⊂ B X , it follows from the previous item and from propo-
sition 2.2 that Hm is a Borel regular measure.
6) It is clear that, ∀x ∈ X, ∀δ ∈ (0, ∞], Hδ0 ({x}) = 1. Thus, H0 ({x}) =
1. Since H0 is a Borel measure, we conclude that the measure of
each finite set coincides with its cardinality and the measure of each
infinity set is ∞, i.e. H0 is the counting measure on X.

Corollary 2.5. If X, Y are metric spaces, m a nonnegative

andm f : X → Y is an isometry into Y , then ∀A ⊂ X,
real number
m
H f (A) = H (A).
Proof. By proposition 2.4.(1), we may substitute Y 0 := Im f for
Y without modifying Hm f (A) . Since f : X → Y 0 is an isometry


onto Y 0 , the thesis follows from proposition 2.4.(2). 
Exercise 2.6 (Hm –null sets). Let X be a metric space, A ⊂ X
and 0 < m < ∞. The following statements are equivalent:
1) Hm (A) = 0.
2) ∃δ ∈ (0, ∞] such that Hδm (A) = 0.
3) ∀ > 0, ∃(En )n∈N cover of A such that n∈N (diam En )m < .
P

The next proposition is a preparation for the introduction of the

notion of Hausdorff dimension.
 2.1. CARATHÉODORY’S CONSTRUCTION 47

Proposition 2.7. Let X be a metric space, A ⊂ X and 0 ≤ s <

t < ∞. If Hs (A) < ∞ then Ht (A) = 0.
Proof. For each δ > 0, since Hδs (A) ≤ Hs (A) < ∞, there ex-
ists
P a countable cover G of A by subsets of diameter ≤ δ such that
−s s s
S∈G α(s)2 (diam S) < H (A) + 1. Then
X α(t)2−t X
Hδt (A) ≤ α(t)2−t (diam S)t = −s
α(s)2−s (diam S)t−s (diam S)s ≤
S∈G
α(s)2 S∈G
α(t) t−s X α(t) t−s s
≤ 2s−t α(s)2−s (diam S)s ≤ 2s−t


δ δ H (A) + 1 ,
α(s) S∈G
α(s)
and taking δ → 0 we conclude that Ht (A) = 0. 
As a corollary, if 0 ≤ s < t < ∞ and Ht (A) > 0, then Hs (A) = ∞.
It then follows that inf{m ∈ [0, ∞) | Hm (A) = 0} = sup{m ∈ [0, ∞) |
Hm (A) = ∞} ∈ [0, ∞].
Definition 2.8. Let X be a metric space and A ⊂ X. The ex-
tended real number inf{m ∈ [0, ∞) | Hm (A) = 0} = sup{m ∈ [0, ∞) |
Hm (A) = ∞} ∈ [0, ∞] is called Hausdorff dimension of A, denoted by
H-dim A.
With the notation above, note that ∀m > H-dim A, Hm (A) = 0,
and ∀m < H-dim A, Hm (A) = ∞. For m = H-dim A, nothing can
be said about Hm (A), i.e. it can be zero, strictly positive or ∞. On
the other hand, if ∃m ∈ [0, ∞) such that 0 < Hm (A) < ∞, then
H-dim A = m.
Exercise 2.9 (properties of Hausdorff dimension). Let X be a
metric space.
a) If Y ⊂ X is a metric subspace of X and A ⊂ Y , the Hausdorff
dimension of A as a subset of the metric space Y is the same for A
as a subset of the metric space X.
b) The Hausdorff dimension is invariant by isometries, i.e. if Y is a
metric space, f : X → Y an isometry into Y and A ⊂ X, then
H-dim A = H-dim f (A).
c) Let X, Y be metric spaces and f : X → Y be a Lipschitz map.
For all A ⊂ X, H-dim f (A) ≤ H-dim A. In particular, if f is
bi-Lipschitz onto its image (i.e. f is Lipschitz and has a Lipschitz
inverse f −1 : Im f → X), then ∀A ⊂ X, H-dim f (A) = H-dim A.
d) (monotonicity) If A ⊂ B ⊂ X, H-dim A ≤ H-dim B.
e) (stability with respect to countable unions) If A = ∪n∈N An ⊂ X,
then H-dim A = sup{H-dim An | n ∈ N}.
48 2. HAUSDORFF MEASURES

We will show next that, for X = Rn , the Hausdorff n-dimensional

measure Hn coincides with the Lebesgue n-dimensional measure Ln . In
particular, that implies H-dim Rn = n (use the stability with respect
to countable unions of the Hausdorff dimension, cf. exercise 2.9, to
Rn = ∪n∈N Cn , where each Cn is a cube with finite Lebesgue measure).
More generally, we will show in exercise 5.42 with the help of the area
formula that, for any smooth embedded k-submanifold M ⊂ Rn , the
measure induced by the Riemannian metric on M coincides with the
trace Hk |M , which implies H-dim M = k.
In order to prove that Hn = Ln in Rn , we need to establish some
preliminaries which are of interest on their own right.

2.2. Vitali’s Covering Theorem

Notation. For a closed ball B = B(x, r) in Rn and 0 < t < ∞, we
define
tB := B(x, tr).
In a general metric space, however, the center and radius of a ball are
not uniquely determined (take, for instance, [0, ∞) as a metric subspace
of R and look at the balls centered at 0). In this case, for a closed ball
B ⊂ X, in order to define tB, we could choose once and for all a center
x and a radius r and proceed like above, i.e. we might consider x and
r as part of the given data when we speak of a ball; instead, we prefer
to proceed like in [Mat95] and define, let us say, for t = 5:
(2.1)
5B := ∪{B 0 ⊂ X closed ball | B 0 ∩ B 6= ∅, diam B 0 ≤ 2 diam B},
which clearly coincides with the previous definition in case X = Rn .
Theorem 2.10 (5–times covering lemma). Let X be a metric space
and F ⊂ 2X a family of nondegenerate closed balls in X such that
sup{diam B | B ∈ F} < ∞. Then there exists a disjoint subfamily
G ⊂ F such that ∪B∈F B ⊂ ∪B∈G 5B.
Proof. Let R := sup{diam B | B ∈ F} < ∞. Since the balls in F
are nondegenerate, i.e. have strictly positive diameter, for any B ∈ F,
diam B ∈ (0, R] = ∪˙ j∈N ( 2Rj , 2j−1
R
]. Thus, putting ∀j ∈ N, Fj := {B ∈
R R
F | diam B ∈ ( 2j , 2j−1 ]}, we have ∪˙ j∈N Fj = F.
We now define inductively (Gj )j∈N by: 1) G1 is a maximal disjoint
subfamily of F1 , obtained by an application of Zorn’s lemma to the
set of all disjoint subfamilies of F1 partially ordered by inclusion; 2)
Once defined G1 ⊂ F1 , . . . , Gj−1 ⊂ Fj−1 , we take a maximal disjoint
subfamily Gj of Fj0 := {B ∈ Fj | ∀B 0 ∈ ∪j−1 0
i=1 Gi , B ∩ B = ∅}, obtained
 2.2. VITALI’S COVERING THEOREM 49

by an application of Zorn’s lemma to the set of all disjoint subfamilies

of Fj0 ⊂ Fj partially ordered by inclusion.
We contend that G := ∪j∈N Gj ⊂ F satisfies the thesis of the theo-
rem. Indeed, it is clear, by construction, that G is a disjoint subfamily
of F. On the other hand, for any B ∈ Fj , there exists B 0 ∈ ∪ji=1 Gi such
that B ∩ B 0 6= ∅, otherwise Gj ∪{B}
˙ % Gj would be a disjoint subfam-
0 R
ily of Fj , violating the maximality of Gj . Since diam B ≤ 2j−1 = 2 2Rj
and 2Rj < diam B 0 , it follows that diam B < 2 diam B 0 , so that B ⊂
5B 0 . 
Remark 2.11. With the notation from theorem 2.10:
1) Note that, if X is separable, then G is countable (since any disjoint
family of sets with nonempty interiors in X is countable).
2) We have actually proved a stronger statement than the thesis: there
exists a disjoint subfamily G ⊂ F such that, for any B ∈ F, ∃B 0 ∈ G
with B ∩ B 0 6= ∅ and diam B < 2 diam B 0 (thus B ⊂ 5B 0 ).
Definition 2.12. Let X be a metric space, F a collection of balls
in X and A ⊂ X. We say that F is a fine cover A, or that F covers
A finely, if F is a cover of A such that, ∀x ∈ A, inf{diam B | x ∈ B ∈
F} = 0.
Corollary 2.13. Let X be a metric space, A ⊂ X, F ⊂ 2X a
family of nondegenerate closed balls of X which covers A finely. Then
there exists a disjoint subfamily G ⊂ F such that, for all F ⊂ F finite,
A \ ∪B∈F B ⊂ ∪B∈G\F 5B.
Proof. Since the cover F is fine, we may assume that sup{diam B |
B ∈ F} ≤ 1; otherwise, discard the balls in F with diameter > 1, so
that the remaining balls still cover A finely. Take G ⊂ F as in remark
2.11.2. Let x ∈ A \ ∪B∈F B. Since F is finite, ∪B∈F B is closed, hence
there exists r > 0 such that U(x, r) ∩ ∪B∈F B = ∅. Since F covers A
finely, there exists B ∈ F such that x ∈ B and diam B < r, so that
B ⊂ U(x, r), thus B ∩ ∪B∈F B = ∅. By remark 2.11.2, there exists
B 0 ∈ G such that B 0 ∩ B 6= ∅ (hence B 0 ∈
/ F ) and diam B < 2 diam B 0 ,
so that x ∈ B ⊂ 5B 0 . Therefore, A \ ∪B∈F B ⊂ ∪B∈G\F 5B, as as-
serted. 
Corollary 2.14 (Vitali’s covering theorem for the Lebesgue mea-
sure). Let A ⊂ Rn and F a collection of nondegenerate closed balls in
Rn which covers A finely. Then, for every  > 0, there exists a disjoint
subfamily G ⊂ F such that Ln (∪G) ≤ Ln (A) +  and Ln (A \ ∪G) = 0.
Note that we do not assume A to be Ln -measurable.
50 2. HAUSDORFF MEASURES

Proof. Assume first that Ln (A) < ∞. We may assume that

L (A) > 0, otherwise the thesis is trivial. Fix 0 < δ < 5−n (so that
n

1−5−n +δ < 1) with δLn (A) < . Since Ln is Borel regular, by theorem
1.23 there exists an open set U ⊃ A such that Ln (U ) < (1 + δ)Ln (A).
We will show that there exists a disjoint subfamily G ⊂ F whose balls
are contained in U and Ln (A \ ∪G) = 0, whence the thesis (since
Ln (∪G) ≤ Ln (U ) < (1 + δ)Ln (A) < Ln (A) + ).
Fix θ ∈ (1 − 5−n + δ, 1).

1) Put FU := {B ∈ F | B ⊂ U, diam B ≤ 1}. Since F covers A finely,

it is clear that FU is still a fine cover of A. Applying theorem 2.10
to FU , there exists a disjoint subfamily GU ⊂ FU such that A ⊂
∪FU ⊂ ∪B∈GU 5B. Since GU is disjoint (hence countable, by remark
n n n
P
2.11), it follows that L (A) ≤ L (∪ B∈GU 5B) ≤ B∈GU L (5B) =
5n B∈GU Ln (B) = 5n Ln (∪GU ). Thus Ln (A\∪GU ) ≤ Ln (U \∪GU ) =
P

Ln (U ) − Ln (∪GU ) ≤ (1 + δ − 5−n )Ln (A) < θLn (A); moreover, since

Ln (A) < ∞, we may apply the continuity from above 1.11 for the
x
Borel measure Ln A to obtain a finite subfamily G1 ⊂ GU such
that Ln (A \ ∪G1 ) < θLn (A).
2) Given 2 ≤ j ∈ N, assume we have defined finite disjoint subfamilies
G1 ⊂ · · · ⊂ Gj−1 ⊂ F such that, for 1 ≤ i ≤ j − 1, the balls of Gi are
contained in U and Ln (A \ ∪Gi ) < θi Ln (A). If Ln (A \ ∪Gj−1 ) = 0,
we stop and the thesis follows with G := Gj−1 ; otherwise, we reapply
the argument of the previous item to the open set U 0 := U \ ∪Gj−1
in place of U and to A0 := A \ ∪Gj−1 ⊂ U 0 in place of A (reducing
the open set U 0 , if necessary, we may assume that Ln (U 0 ) < (1 +
δ)Ln (A0 )): take FU 0 := {B ∈ F | B ⊂ U 0 , diam B ≤ 1}, which
is a fine cover of A0 , and use theorem 2.10 to extract a disjoint
subfamily GU 0 ⊂ FU 0 such that A0 ⊂ ∪FU 0 ⊂ ∪B∈GU 0 5B, so that
Ln (A0 \ ∪GU 0 ) ≤ (1 + δ − 5−n )Ln (A0 ). Then, applying once more the
x
continuity from above for the Borel measure Ln A0 , there exists
a finite set Gj0 ⊂ GU 0 such that Ln (A0 \ ∪Gj0 ) < θLn (A0 ) < θj Ln (A).
Put Gj := Gj−1 ∪ Gj0 . Then Gj ⊂ F is a finite disjoint family whose
balls are contained in U , and A \ ∪Gj = A0 \ ∪Gj0 has Lebesgue
measure < θj Ln (A).
3) We have thus inductively defined an increasing sequence (Gj )j∈N
such that, for each j ∈ N, Gj is a finite disjoint subfamily of F
whose balls are contained in U , with Ln (A\∪Gj ) < θj Ln (A). Define
G := ∪i∈N Gi ; then G is a disjoint subfamily of F whose balls are
j→∞
contained in U , with ∀j ∈ N, Ln (A \ ∪G) ≤ θj Ln (A) −→ 0. Hence
Ln (A \ ∪G) = 0, which concludes the proof in the case Ln (A) < ∞.
 2.3. STEINER SYMMETRIZATION 51

If Ln (A) = ∞, we take ∀k ∈ N, Vk := {x ∈ Rn | k − 1 < kxk < k}.

Given  > 0, for each k ∈ N we take an open set Uk ⊃ A ∩ Vk such that
L Uk \ (A ∩ Vk ) < 2−k ; substituting Uk ∩ Vk for Uk , we may assume
n


Uk ⊂ Vk . We now apply the first part of the proof to find, for each k ∈
N, a disjoint subfamily Gk ⊂ F whose balls are contained in Uk ⊂ Vk
and Ln (A ∩ Vk \ ∪Gk ) = 0. Then, since the Vk ’s are pairwise disjoint,
G := ∪k∈N Gk is a disjoint

subfamily of F such that Ln (A \ ∪G) ≤
L ∪k∈N (A∩Vk \∪Gk ) +L (∪k∈N {x ∈ Rn | kxk = k−1}) = 0. Besides,
n n

Ln (∪G) = k∈N Ln (∪Gk ) < k∈N Ln (A ∩ Vk ) + 2−k  = Ln (A) + . 

P P

Corollary 2.15 (filling open sets with balls with respect to Lebesgue
measure). Let U ⊂ Rn be an open set and F a family of nondegenerate
closed balls contained in U which covers U finely (for instance, if F
is the family of all nondegenerate closed balls contained in U , or the
family of all such balls with diameters bounded by a fixed δ > 0). Then
there exists a disjoint subfamily G ⊂ F such that Ln (U \ ∪G) = 0.
Proof. Apply the previous corollary with A = U . 
2.3. Steiner Symmetrization
We briefly study in this subsection some properties of the oper-
ation called Steiner symmetrization, introduced by Jakob Steiner in
1836 ([Ste38]). This operation will be used to prove the isodiametric
inequality in 2.19, our key ingredient to show that Ln = Hn in Rn .
Definition 2.16. Let (e1 , . . . , en ) be the standard basis of Rn and
identify Rn−1 ≡ he1 , . . . , en−1 i, R ≡ hen i, so that Rn ≡ Rn−1 × R. We
define the Steiner symmetrization with respect to Rn−1 to be the map
n n
Sen : 2R → 2R defined by (see figure 1):
[ 1
Sen (A) := {(x0 , xn ) | |xn | ≤ L1 (Ax0 )},
0 n−1
2
{x ∈R |Ax0 6=∅}

where we have used the notation for sections established in 1.4.

Given a ∈ Sn−1 ⊂ Rn , we define similarly the Steiner symmetriza-
tion Sa with respect to the (n−1)-dimensional subspace hai⊥ : take any
orthogonal map φ ∈ O(n) such that φ(a) = en (hence φ(hai⊥ ) = Rn−1 )
and put Sa := φ−1 ◦ Sen ◦ φ.

Proposition 2.17 (properties of Steiner symmetrization). Let a ∈

n−1
S .
i) ∀A ⊂ Rn , diam Sa (A) ≤ diam A.
n
ii) If A ⊂ R
is Ln -measurable, then so is Sa (A) and Ln (A) =
Ln Sa (A) .
52 2. HAUSDORFF MEASURES

Figure 1. Steiner Symmetrization

Proof. Since diameters and Lebesgue measure are preserved by

isometries, it suffices to prove the proposition for a = en .
i) First we make a reduction: it suffices to prove the thesis for closed
sets. Indeed, assuming the thesis for closed sets, since the Steiner
symmetrization is clearly monotone with respect to set inclusion,
it follows for arbitrary A ⊂ Rn that diam Sen (A) ≤ diam Sen (A) ≤
diam A = diam A.
So, assume that A is closed. Given x, y ∈ Sen (A), we will exhibit
x , y ∈ A such that kx − yk ≤ kx0 − y 0 k, what clearly implies
0 0

diam Sen (A) ≤ diam (A). Indeed, let (see figure 1 for the notation)
x = (b, xn ), y = (c, yn ), r := inf{t | (b, t) ∈ A}, s := sup{t |
(b, t) ∈ A}, u := inf{t | (c, t) ∈ A} and v := sup{t | (c, t) ∈ A}.
Note that, since A is closed, (b, r), (b, s), (c, u), (c, v) ∈ A. Note
also that, since Ab ⊂ [r, s] and Ac ⊂ [u, v], we have s − r ≥ L1 (Ab )
and v − u ≥ L1 (Ac ).
Up to relabeling the points, we may assume that s − u ≥ v − r
(like it is the case in the figure). It then follows that:
1 1
s − u ≥ (s − u) + (v − r) =
2 2
1 1
= (s − r) + (v − u) ≥
2 2
1 1 1
≥ L (Ab ) + L1 (Ac ) ≥
2 2
≥ |xn | + |yn | ≥ |xn − yn |.
 2.4. THE ISODIAMETRIC INEQUALITY; Ln = Hn 53

Thus, kx − yk2 = |xn − yn |2 + kb − ck2 ≤ (s − u)2 + kb − ck2 =

k(b, s) − (c, u)k2 , thus the assertion is proved with x0 = (b, s), y 0 =
(c, u) ∈ A.
ii) If n = 1,
Sen (A) is a closed set and it is clear that L1 (A) =
L1 Sen (A) . If n ≥ 2, let f : Rn−1 → [0, ∞] be given by f (x) =
1 1
2
L (Ax ). It follows from Fubini-Tonelli’s theorem 1.84 that f is
Ln−1 -measurable; hence, by lemma 2.18, S := {(x, t) ∈ Rn−1 × R |
−f (x) ≤ t ≤ f (x)} is Ln -measurable. Then Sen (A) = S \ {(x, 0)
|
Ax = ∅} is Ln -measurable and the fact that Ln (A) = Ln Sen (A)
is a consequence of Fubini-Tonelli’s theorem.

Lemma 2.18. Let f : Rn → [0, ∞] be Ln -measurable. Then hyp f :=
{(x, t) ∈ Rn × [0, ∞) | t ≤ f (x)} ⊂ Rn+1 is Ln+1 -measurable.
Proof. Let θ : R × R → R be defined by θ(x, y) = x − y, with ∞ −
∞ := 0, −∞−(−∞) := 0. Then θ is measurable with respect to B R×R
and B R (see proposition 1.50 and example 1.51).
On the other hand,
(f, ι) : Rn × R → R × R given by (x, y) 7→ f (x), y is measurable with
respect to L Rn+1 and B R ⊗ B R , since each component is measurable
(see remark 1.45). Since B R ⊗ B R = B R×R (by proposition 1.47), it
follows that the composite f − ι is measurable with respect to L Rn+1
and B R , whence hyp f = (f − ι)−1 ([0, ∞]) ∩ Rn × [0, ∞) is Ln+1 -
measurable. 

2.4. The isodiametric inequality; Ln = Hn

Theorem 2.19 (isodiametric inequality). The Lebesgue measure of
any subset of Rn is at most the measure of an euclidean ball with the
same diameter. That is, for all A ⊂ Rn ,
 diam A n
Ln (A) ≤ α(n) .
2
Proof. We assume that diam A < ∞, otherwise the thesis is triv-
ial.
1) Let (e1 , . . . , en ) be the standard basis of Rn . Define S0 := Sen ◦
Sen−1 ◦ · · · ◦ Se1 . Note that the same properties stated in proposition
2.17 for the Steiner symmetrization also hold for S0 (just apply that
proposition n times in a row).
2) We contend that, for all B ⊂ Rn , for 1 ≤ j ≤ n, S0 (B) is symmetric
with respect to the hyperplane hej i⊥ ; that is, denoting by Rj : Rn →
⊥
n


R the reflection with respect to hej i , Rj S0 (B) = S0 (B). Indeed,
for 1 ≤ j ≤ n, let Bj := Sej ◦ · · · ◦ Se1 (B).
54 2. HAUSDORFF MEASURES

i) By definition 2.16, it is clear that B1 = Se1 (B) is invariant by

R1 .
ii) Assume that, given 2 ≤ j ≤ n, Bj−1 is invariant by Ri for
1 ≤ i ≤ j − 1. We will show that Bj = Sej (Bj−1 ) is invariant
by Ri for 1 ≤ i ≤ j. That is clear for i = j, by definition 2.16.
For i < j, since Bj−1 is invariant by Ri , we have, denoting by
Pj the orthogonal projection on Rn−1 ≡ hej i⊥ and by Pj⊥ the
orthogonal projection on R ≡ hej i, ∀x ∈ Rn−1 :
Ri ◦Pj =Pj ◦Ri
Bj−1 ∩ Pj−1 (x) = Ri (Bj−1 ) ∩ Pj−1 (x) =
Pj−1 (Ri−1


= Ri Bj−1 ∩ · x)

As Ri−1 = Ri and Pj⊥ ◦Ri = Pj⊥ , it then follows that, ∀x ∈ Rn−1 :

(Bj−1 )x = Pj⊥ Bj−1 ∩ Pj−1 (x) =

= Pj⊥ ◦ Ri Bj−1 ∩ Pj−1 (Ri · x) =

= Pj⊥ Bj−1 ∩ Pj−1 (Ri · x) =

= (Bj−1 )Ri ·x .
By the arbitrariness of x ∈ Rn−1 , the equality
above implies,
in view of definition 2.16, that Ri Sej (Bj−1 ) = Sej (Bj−1 ), i.e.
Bj = Sej (Bj−1 ) is invariant by Ri , as asserted.
Our contention is therefore proved.
3) From the contention in the previous item, it follows that, given B ⊂
Rn , S0 (B) is invariant by Rn ◦Rn−1 ◦· · ·◦R1 , i.e. S0 (B) is symmetric
with respect to the origin. Thus, ∀x ∈ S0 (B), −x ∈ S0 (B), so that
2kxk ≤ diam S0 (B) ≤ diam B, i.e. S0 (B) ⊂ B(0, diam 2
B
).
4) It follows from the previous item applied to B = A that:
diam A

Ln (A) ≤ Ln (A) = Ln S0 (A) ≤ Ln B(0,


) =
2
 diam A n  diam A n
= α(n) = α(n) .
2 2

Exercise 2.20. Show an example of a set A ⊂ Rn which is not
contained in any ball with diameter diam A.
Theorem 2.21. For all δ ∈ (0, ∞] and n ∈ N, Hn = Hδn = Ln in
n
R .
Proof. Fix δ ∈ (0, ∞] and A ⊂ Rn .
 2.4. THE ISODIAMETRIC INEQUALITY; Ln = Hn 55

1) Claim 1: Hδn (A) ≥ Ln (A). Indeed, let F be a countable cover of A

by subsets of Rn with diameters ≤ δ. For each S ∈ F, it follows from
the isodiametric inequality 2.19 thatPLn (S) ≤ α(n)2−n (diam S)n .
−n
(diam S)n ≥ n n
P
Hence, S∈F α(n)2 S∈F L (S) ≥ L (A), where
the last inequality is due to the countable subadditivity of Ln . Tak-
ing the infimum of all such covers F yields the claim.
2) Claim 2: for all B ⊂ Rn , Ln (B) = 0 implies Hn (B) = 0. To prove
the claim, fix  > 0 and take F a countable cover P of B by cubes
of sides parallel to the coordinate axes such that Q∈F vol(Q) < ;
such a cover exists, in view of the definition of the Lebesgue measure

n
in example 1.3. Since, for each cube Q ∈ F, vol(Q) = diam √ Q
n
, we
n n/2
P
conclude that Q∈F (diam Q) < n . Hence, by the arbitrariness
of the  > 0 fixed, the claim follows from exercise 2.6.
3) Claim 3: Hδn (A) ≤ Ln (A). Assume that Ln (A) < ∞ (otherwise
the claim is trivial) and take a countable cover FPof A by cubes
of sides parallel to the coordinate axes such that Q∈F vol(Q) <
Ln (A) + . For each Q ∈ F, we may apply corollary 2.15 to Qo
k k
to obtain a countable disjoint family (BQ )k∈N such that each BQ
is a nondegenerate closed ball with diameter ≤ δ contained in Qo
and Ln (Qo \ ∪k∈N BQ k
) = 0. Since Ln (∂Q) = 0, it then follows that
L (Q \ ∪k∈N BQ ) = 0; hence, by claim 2, Hn (Q \ ∪k∈N BQ
n k k
) = 0,
n k n
so Hδ (Q \ ∪k∈N BQ ) = 0. Since, by finite subadditivity, Hδ (Q) ≤
Hδn (Q \ ∪k∈N BQ k
) + Hδn (∪k∈N BQk
) = Hδn (∪k∈N BQk
), it follows that
n n k
Hδ (Q) = Hδ (∪k∈N BQ ). Therefore,
X X countable subadditivity
Hδn (A) ≤ Hδn (Q) = Hδn (∪k∈N BQ
k
) ≤
Q∈F Q∈F
k ≤δ
diam BQ
XX
≤ Hδn (BQ
k
) ≤
Q∈F k∈N
XX XX
≤ α(n)2−n (diam BQ
k n
) = Ln (BQ
k
)=
Q∈F k∈N Q∈F k∈N
X X X
= Ln (∪k∈N BQ
k
)= Ln (Q) = vol(Q) < Ln (A) + .
Q∈F Q∈F Q∈F

Thus, by the arbitrariness of , claim 3 is proved.

4) By claims 1 and 3, Hδn (A) = Ln (A). Since that holds for all δ > 0,
it follows that Ln (A) = Hn (A), hence the thesis follows.

Corollary 2.22. H-dim Rn = n.
56 2. HAUSDORFF MEASURES

Proof. Apply the stability with respect to countable unions of

the Hausdorff dimension, cf. exercise 2.9.e), to Rn = ∪k∈N Ck , where
each Ck is a nondegenerate cube with finite Lebesgue measure, i.e.
0 < Hn (Ck ) < ∞, so that ∀k ∈ N, H-dim Ck = n. 
Exercise 2.23. If E is a k-dimensional subspace of a normed space
X, then H-dim E = k.
 CHAPTER 3

Differentiation of Measures

The main reference for this chapter is [Sim83].

3.1. Densities
Up to the end of this section we fix a metric space (X, d).
Definition 3.1 (upper and lower n-dimensional densities). Let
A ⊂ X, x ∈ X, n > 0 real and µ a measure on X. We define:
1) the n-dimensional upper density of A at x with respect to µ:


µ A ∩ B(x, r)
Θ∗n (µ, A, x) := lim sup ∈ [0, ∞].
r→0 α(n)rn
2) the n-dimensional lower density of A at x with respect to µ:


n µ A ∩ B(x, r)
Θ∗ (µ, A, x) := lim inf ∈ [0, ∞].
r→0 α(n)rn
If Θ∗n (µ, A, x) = Θn∗ (µ, A, x), we denote their common value by
Θn (µ, A, x) and call it density of A at x with respect to µ.
For A = X, we use the notations Θ∗n (µ, x), Θn∗ (µ, x) and Θn (µ, x)
for Θ∗n (µ, X, x), Θn∗ (µ, X, x) and Θn (µ, X, x), respectively.
Note that we don’t assume A to be measurable.
Remark 3.2. With the notation above:
x x
1) Note that Θ∗n (µ, A, x) = Θ∗n (µ A, x) and Θn∗ (µ, A, x) = Θn∗ (µ A, x).
x
2) If U ⊂ X is an open set and x ∈ U , Θ∗n (µ, A, x) = Θ∗n (µ U, A, x)
x
and Θn∗ (µ, A, x) = Θn∗ (µ U, A, x).
Lemma 3.3. If µ is a locally finite Borel measure on X, A ⊂ X, x ∈
X and n > 0 real, then, the definitions of Θ∗n (µ, A, x) or Θn∗ (µ, A, x)
do not change if we use open balls instead of closed balls.
Proof. Recall that, for f : (0, ∞) → R,
lim sup f (r) := inf sup f (ρ),
r→0 r>0 0<ρ<r

lim inf f (r) := sup inf f (ρ).

r→0 r>0 0<ρ<r
57
58 3. DIFFERENTIATION OF MEASURES

Put, for r > 0,

µ A ∩ U(x, r) µ A ∩ B(x, r)
f (r) := and g(r) := .
α(n)rn α(n)rn


Since µ is locally finite, there exists r0 > 0 such that µ U(x, r0 ) < ∞.
In order to prove the lemma, it suffices to show that, ∀0 < r < r0 ,
sup f (ρ) = sup g(ρ) and inf f (ρ) = inf g(ρ)
0<ρ<r 0<ρ<r 0<ρ<r 0<ρ<r

That is a consequence of the following claims:

1) Claim 1: ∀ρ ∈ (0, r), g(ρ) may be arbitrarily approximated by el-
ements of {f (ρ) | 0 < ρ < r}. Indeed, for a given ρ ∈ (0, r),
B(x, ρ) = ∩k∈N U(x, ρ +
1/k); for sufficiently large k, ρ + 1/k < r,
hence µ U(x, ρ + 1/k) < ∞. That allows us to apply the con-
tinuity from above 1.11 to
the Borel measure x

µ A, which en-
sures µ A ∩ U(x, ρ + 1/k) → µ A ∩ B(x, ρ) as k → ∞. Thus,
f (ρ + 1/k) → g(ρ), as asserted.
2) Claim 2: ∀ρ ∈ (0, r), f (ρ) may be arbitrarily approximated by el-
ements of {g(ρ) | 0 < ρ < r}. Indeed, for a given ρ ∈ (0, r),
U(x, ρ) = ∪{B(x, ρ − 1/k) | k ∈ N, 1/k < ρ}. Applying the con-
tinuity from below 1.11 to
the Borel measure x

µ A, it follows
that µ A ∩ B(x, ρ − 1/k) → µ A ∩ U(x, ρ) as k → ∞. Thus,
g(ρ − 1/k) → f (ρ), as asserted.

Proposition 3.4. If µ is a locally finite Borel measure on X, A ⊂
X and n > 0 real, then the functions X → [0, ∞] given by x ∈ X 7→
Θ∗n (µ, A, x) and x ∈ X 7→ Θn∗ (µ, A, x) are Borelian.
Proof.
1) Firstly, note that, for fixed

r > 0, the function X → [0, ∞] given
by x 7→ µ A ∩ U(x, r) is lower semicontinuous (hence Borelian).
Indeed, let x ∈ X and (xn )n∈N a sequence in X convergent to x.
For all k ∈ N such that r − 1/k > 0, ∃n0 ∈
N, ∀n ≥ n0 , U(xn , r) ⊃

B(x, r−1/k). Hence ∀n ≥ n0 , µ A∩U(x
n , r) ≥ µ A∩B(x,
r−1/k) ,
whence lim inf n→∞ µ A∩U(xn , r) ≥ µ A∩B(x, r−1/k) . Applying
the continuity from below 1.11
for the Borel measure

x
µ A, we con-
clude that

µ A∩B(x, r−1/k) → µ A∩U(x, r) ≤ lim inf n→∞ µ A∩
U(xn , r) , which shows the asserted lower semicontinuity at x.
2) Claim: Given r > 0, the functions ψr , ψ r : X → [0, ∞] given by:



µ A ∩ U(x, ρ) µ A ∩ U(x, ρ)
x 7→ inf and x 7→ sup ,
0<ρ<r α(n)ρn 0<ρ<r α(n)ρn
 3.1. DENSITIES 59

respectively, are Borelian. Indeed, it suffices to show that, ∀x ∈ X,

µ A ∩ U(x, ρ) µ A ∩ U(x, ρ)
inf = inf{ | 0 < ρ < r, ρ ∈ Q} and
0<ρ<r α(n)ρn α(n)ρn



µ A ∩ U(x, ρ) µ A ∩ U(x, ρ)
sup = sup{ | 0 < ρ < r, ρ ∈ Q},
0<ρ<r α(n)ρn α(n)ρn
in which case the asserted measurability follows from the previous
item and from theorem 1.41.(iv). In order to show the equalities


µ A∩U(x,ρ)
above, it is enough to prove that, for all 0 < ρ < r, α(n)ρn
may


µ A∩U(x,ρ)
be arbitrarily approximated by elements of the set { α(n)ρn |
0 < ρ < r, ρ ∈ Q}. For that purpose, take a sequence of rationals
(ρk )k∈N in (0, ρ) such that ρk ↑ ρ; then α(n)ρnk → α(n)ρn and,
applying the continuity from below to the Borel
measure µ A, x




µ A∩U(x,ρk ) µ A∩U(x,ρ)
µ A ∩ U(x, ρk ) ↑ µ A ∩ U(x, ρ) , hence α(n)ρn
→ α(n)ρn ,
k
as asserted.
3) Due to the fact that µ is a locally finite Borel measure, it follows
from lemma 3.3 that Θ∗n (µ, A, ·) = inf r∈Q∗+ ψ r and Θn∗ (µ, A, ·) =
supr∈Q∗+ ψr . Thus, from the claim in the previous item and from
theorem 1.41.(iv), we conclude that both Θ∗n (µ, A, ·) and Θn∗ (µ, A, ·)
are Borelian.

Corollary 3.5. If µ is a locally finite Borel measure on X, A ⊂ X
and n > 0 real, then the set Y := {x ∈ X | Θ∗n (µ, A, x) = Θn∗ (µ, A, x)}
is Borel measurable and Θn (µ, A, ·) : Y → [0, ∞] is Borelian.
Theorem 3.6 (comparison density theorem). Let µ be a Borel mea-
sure on a metric space X, n > 0 real, t ≥ 0 and A ⊂ A1 ⊂ X. If
∀x ∈ A, Θ∗n (µ, A1 , x) ≥ t then tHn (A) ≤ µ(A1 ).
Proof. We assume that t > 0 and µ(A1 ) < ∞, otherwise the
thesis is trivial.
Fix 0 < τ < t
and δ > 0. Since, ∀x ∈ A, Θ∗n (µ, A1 , x) =
µ A1 ∩B(x,r)
lim supr→0 α(n)rn
≥ t > τ , it follows that ∀x ∈ A, ∀r > 0,


µ A1 ∩B(x,ρ)
∃0 < ρ < r such that α(n)ρn
> τ . It then follows that F :=


µ A1 ∩B(x,r)
{B | ∃x ∈ A, ∃r > 0, B = B(x, r), α(n)rn > τ, 2r ≤ δ} is a fine
cover of A. Take a disjoint subfamily G ⊂ F, given by corollary 2.13,
60 3. DIFFERENTIATION OF MEASURES

such that, for all F ⊂ F finite,

(3.1) A \ ∪B∈F B ⊂ ∪B∈G\F 5B.
Due to x
P the fact that µ A1 is a Borel measure,

for all F ⊂ G fi-
nite, B∈F µ(A1 ∩P B) = µ A1 ∩ (∪B∈F B) ≤ P µ(A 1 < ∞. Therefore,
)
by exercise 1.59, B∈G µ(A1 ∩ B) = sup{ B∈F µ(A1 ∩ B) | F ⊂
G, F finite} ≤ µ(A1 ) < ∞; since, for all B ∈ G ⊂ F, µ(A1 ∩ B) > 0, it
follows that G is countable. Let (Bk )k∈N be an enumeration of G. For
each k ∈ N, Bk ∈ F, hence there exists xk ∈ A and rk > 0 such that
Bk = B(xk , rk ) and τ α(n)rkn < µ(A1 ∩ Bk ), so that
∞
X X∞
n
τ α(n)rk ≤ µ(A1 ∩ Bk ) =
k=1 k=1


= µ A1 ∩ (∪k∈N Bk ) ≤ µ(A1 ) < ∞.
On the other hand, for all N ∈ N, it follows from (3.1) that A ⊂
(∪N
k=1 Bk ) ∪ (∪k≥N +1 5Bk ). Since, for each k ∈ N, diam Bk ≤ 2rk ≤ δ
and diam 5Bk ≤ 5 diam Bk ≤ 10rk ≤ 5δ, we then conclude that
N
X ∞
X
−n
n
H5δ (A) ≤ α(n)2 (diam Bk ) + n
α(n)2−n (diam 5Bk )n ≤
k=1 k=N +1
N
X ∞
X
≤ α(n)rkn + 5n α(n)rkn .
k=1 k=N +1
n
P∞ n n
Thus,
P∞ taking N → ∞, it follows H 5δ (A) ≤ k=1 α(n)rk , hence τ H5δ (A) ≤
τ k=1 α(n)rkn ≤ µ(A1 ). Taking δ → 0, we obtain τ Hn (A) ≤ µ(A1 ).
Finally, since τ ∈ (0, t) was arbitrarily taken, making τ → t in the last
inequality yields the thesis. 
Theorem 3.7 (upper density theorem). Let µ be a Borel regular
measure on a metric space X, n > 0 real and B ∈ σ(µ) with µ(B) < ∞.
Then Θ∗n (µ, B, x) = 0 for Hn -a.e. x ∈ X \ B.
Proof. Let C ⊂ B be a closed set and t > 0. Define At := {x ∈
X \ B | Θ∗n (µ, B, x) ≥ t} and At1 := X \ C ⊃ At . Since At1 = X \ C
is an open set, it follows from remark 3.2 that, for all x ∈ At ⊂ At1 ,
x x
Θ∗n (µ B, At1 , x) = Θ∗n (µ At1 , B, x) = Θ∗n (µ, B, x) ≥ t. Thus, we
may apply theorem 3.6 with the Borel measure µ B in place of µ, Atx
in place of A and At1 in place of A1 , yielding tHn (At ) ≤ µ B(At1 ) = x
µ(B\C). By the arbitrariness of C, it follows that tHn (At ) ≤ inf{µ(B\
C) | C ⊂ B, C closed}. On the other hand, it follows from proposition
x
1.36 that µ B is a finite Borel regular measure, to which theorem
1.23 may be applied to approximate B by closed sets contained in B,
 3.1. DENSITIES 61

which yields inf{µ(B \ C) | C ⊂ B, C closed} = 0. As t > 0 was

arbitrarily taken, it follows that ∀t > 0, Hn (At ) = 0. Since {x ∈
X \ B | Θ∗n (µ, B, x) > 0} = ∪k∈N A1/k , we conclude that Hn ({x ∈
X \ B | Θ∗n (µ, B, x) > 0}) = 0, whence the thesis. 
Exercise 3.8. If µ is an open σ-finite Borel regular measure on a
metric space X, the thesis in theorem 3.7 holds for all B ∈ σ(µ), i.e.
the hypothesis of µ(B) being finite may be dropped.
Corollary 3.9 (density theorem for the Lebesgue measure). If
B ⊂ Rn is Ln -measurable, then Θn (Ln , B, x) exists for Ln -a.e. x ∈ Rn ,
Θn (Ln , B, x) = 1 for Ln -a.e. x ∈ B and Θn (Ln , B, x) = 0 for Ln -a.e.
x ∈ Rn \ B.
Proof. Note that, if f, g : (0, ∞) → [0, ∞], then
lim inf f (r) + lim inf g(r) ≤ lim inf (f + g)(r) ≤
r→0 r→0 r→0
≤ lim inf f (r) + lim sup g(r) ≤
r→0 r→0
≤ lim sup(f + g)(r) ≤ lim sup f (r) + lim sup g(r).
r→0 r→0 r→0


n Ln B∩B(x,r)
Fix x ∈ R . Applying the inequalities above to f (r) = α(n)rn


n n
L (R \B)∩B(x,r)
and g(r) = α(n)rn
, and taking into consideration that f (r) +
g(r) ≡ 1, it follows that, for all x ∈ Rn :
(3.2) Θ∗n (Ln , B, x) + Θn∗ (Ln , Rn \ B, x) = 1,
and the same holds with Rn \ B in place of B.
On the other hand, theorems 2.21, 3.7, exercise 3.8 and the fact
that 0 ≤ Θn∗ (Ln , B, ·) ≤ Θ∗n (Ln , B, ·) ≤ 1 imply that Θn (Ln , B, x) = 0
for Ln -a.e. x ∈ Rn \ B. The same holds for Rn \ B in place of B, i.e.
Θn (Ln , Rn \ B, x) = 0 for Ln -a.e. x ∈ B. The last equality implies, in
view of (3.2), that Θn (Ln , B, x) = 1 for Ln -a.e. x ∈ B. 
Exercise 3.10. Any convex subset X of Rn is Ln -measurable.
Hint. Use corollary 3.9 to prove that ∂X has null Lebesgue mea-
sure.
Exercise 3.11. Let X be a metric space, n > 0 real and A ⊂ X be
H -measurable with Hn (A) < ∞. Then Θ∗n (Hn , A, x) ≤ 1 for Hn -a.e.
n

x ∈ A.
Hint. For each t > 1, put At := {x ∈ A | Θ∗n (Hn , A, x) ≥ t}.
Given  > 0, take an open set U ⊃ At such that Hn (U ∩A) < Hn (At )+
62 3. DIFFERENTIATION OF MEASURES

(why such an open set exists?) and apply theorem 3.6 with Hn xA in
place of µ, At in place of A and A ∩ U in place of A1 .
3.2. Differentiation Theorems
In the first part of this section we extend theorems 3.6 and 3.7 to
the situation in which we define upper and lower densities of a Borel
measure µ on a metric space X with respect to another Borel measure
ν on X, with convenient regularity and finiteness assumptions.
Definition 3.12 (upper and lower densities of a measure relative
another). Let X be a metric space, µ and ν measures on X, and x ∈
X. We define the upper and lower density of µ relative to ν at x by,
respectively:


µ B(x, r)
Θ∗ν (µ, x) := lim sup
∈ [0, ∞],
r→0 ν B(x, r)


ν µ B(x, r)
Θ∗ (µ, x) := lim inf
∈ [0, ∞],
r→0 ν B(x, r)

where we adopt the extended arithmetic rules 00 := 0, ∞ ∞

:= 0. If
∗ν ν
Θ (µ, x) = Θ∗ (µ, x), we say that the density of µ relative to ν at x
exists and denote it by Θν (µ, x) := Θ∗ν (µ, x) = Θν∗ (µ, x).
Note that:
• if ∃r > 0, µ B(x, r)
= 0, then Θ∗ν (µ, x) = Θν∗ (µ, x)



= 0.
• if @r > 0, µ B(x, r) = 0 and ∃r > 0, ν B(x, r) = 0, then
Θ∗ν (µ, x) = Θν∗ (µ, x) = ∞.
In particular, if x ∈
/ spt µ ∩ sup ν, the upper and lower densities at
x assume value 0 or ∞.
Remark 3.13. If X = Rn , A ⊂ Rn , x ∈ Rn and µ a measure
on Rn , the n-dimensional upper and lower densities of A at x with
respect to µ, defined in 3.1, are special cases of the above definition:
n
x n
x
Θ∗n (µ, A, x) = Θ∗L (µ A, x) and Θn∗ (µ, A, x) = ΘL∗ (µ A, x).
Lemma 3.14. If µ and ν are locally finite Borel measures on a met-
ric space X, and x ∈ X, then the definitions of Θ∗ν (µ, x) or Θν∗ (µ, x)
do not change if we use open balls instead of closed balls.
Proof. It is an adaptation of the proof of lemma 3.3, analyzing
separately the case in which x 6= spt µ ∩ spt ν.
Define, for r > 0,



µ U(x, r) µ B(x, r)
f (r) :=
and g(r) :=
.
ν U(x, r) ν B(x, r)
 3.2. DIFFERENTIATION THEOREMS 63


1) If ∃r0 > 0, µ B(x, r0 ) = 0, then ∀0 < r < r0 , f (r) = g(r) = 0 and
the thesis is trivial in
this case.

2) If @r > 0, µ B(x, r) = 0 and ∃r0 > 0, ν B(x, r0 ) = 0, then
∀0 < r < r0 , f (r) = g(r) = ∞ and the thesis is also trivial in this
case.
3) If neither of the previous cases holds, then x ∈ spt µ ∩ spt ν.
Since both
µ and ν are locally finite,

there exists r0 > 0 such that
µ B(x, r0 ) <
∞ and ν B(x, r0 ) < ∞.
That is, for all 0 < r < r0 ,
0 < µ B(x, r) < ∞ and 0 < ν B(x, r) < ∞. In order conclude
the proof of the lemma, it suffices to show that, ∀0 < r < r0 ,
sup f (ρ) = sup g(ρ) and inf f (ρ) = inf g(ρ)
0<ρ<r 0<ρ<r 0<ρ<r 0<ρ<r

That is a consequence of the following claims:

i) Claim 1: ∀ρ ∈ (0, r), g(ρ) may be arbitrarily approximated by
elements of {f (ρ) | 0 < ρ < r}. Indeed, for a given ρ ∈ (0, r),
B(x, ρ) = ∩k∈N U(x, ρ+1/k);

for sufficiently large k,
ρ+1/k < r,
hence µ U(x, ρ + 1/k) < ∞ and ν U(x, ρ + 1/k) < ∞. That
allows us to apply the continuity from above 1.11
to the Borel

measures µ and ν, which

ensures µ U(x,

ρ+1/k) → µ B(x, ρ)
and ν U(x, ρ + 1/k) → ν B(x, ρ) > 0 as k → ∞. Thus,
f (ρ + 1/k) → g(ρ), as asserted.
ii) Claim 2: ∀ρ ∈ (0, r), f (ρ) may be arbitrarily approximated by
elements of {g(ρ) | 0 < ρ < r}. Indeed, for a given ρ ∈ (0, r),
U(x, ρ) = ∪{B(x, ρ − 1/k) | k ∈ N, 1/k < ρ}. Applying the
continuity from below 1.11 to
the Borel measures
µ and ν, it
follows

that µ B(x,
ρ − 1/k) → µ U(x, ρ) and ν B(x, ρ −
1/k) → ν U(x, ρ) > 0 as k → ∞. Thus, g(ρ − 1/k) → f (ρ),
as asserted.

Proposition 3.15. Let µ and ν be locally finite Borel measures
on a metric space X, with ν finite on all closed balls of X. Then the
functions X → [0, ∞] given by x ∈ X 7→ Θ∗ν (µ, x) and x ∈ X 7→
Θν∗ (µ, x) are Borelian.
Proof. We adapt the proof or proposition 3.4.


1) Let U0 := {x ∈ X | ∃r > 0, ν B(x,
r) = 0} = X \ spt ν and
V0 := {x ∈ X | ∃r > 0, µ B(x, r) = 0} = X \ spt µ. We will
apply proposition 1.50 to A1 = V0 , A2 = U0 \ V0 and A3 = X \
(U0 ∪ V0 ). Note that, since U0 and V0 are open sets, (Ai )1≤i≤3 is
a Borel partition of X. As Θ∗ν (µ, ·) and Θν∗ (µ, ·) are constant on
A1 (equal to 0) and A2 (equal to ∞), their restrictions to A1 and
64 3. DIFFERENTIATION OF MEASURES

A2 , endowed with the respective trace σ-algebras, are measurable.

It remains to show that the restrictions of Θ∗ν (µ, ·) and Θν∗ (µ, ·)
to A3 = spt µ ∩ spt ν are measurable, endowing A3 with the trace
σ-algebra B X |A3 .
2) Note that,
for fixed r > 0, the functions

X → [0, ∞] given by x 7→
µ U(x, r) and x 7→ ν U(x, r) are lower semicontinuous (hence
Borelian). Indeed, let x ∈ X and (xn )n∈N a sequence in X conver-
gent to x. For all k ∈ N such that r − 1/k > 0, ∃n0
∈ N, ∀n ≥ n0 ,
U(xn
, r) ⊃ B(x, r − 1/k). Hence ∀n ≥
n0 , µ U(xn , r) ≥
µ B(x, r −
1/k) , whence lim inf n→∞ µ U(xn , r) ≥ µ B(x, r−1/k) . Applying
the continuity from below

1.11 to the

Borel measure µ, we conclude


that µ B(x, r − 1/k) → µ U(x, r) ≤ lim inf n→∞ µ U(xn , r) , what
shows the asserted lower semicontinuity at x for µ, and the same
argument holds for ν.

µ U(·,r)
It then follows that the quotient
: X → [0, ∞] is Bore-
ν U(·,r)
lian (see example 1.51.2), so its restriction to A3 is measurable with
respect to the trace σ-algebra.
3) Claim: Given r > 0, the functions ψr , ψ r : X → [0, ∞] given by:



µ U(x, ρ) µ U(x, ρ)
x 7→ inf
and x 7→ sup
,
0<ρ<r ν U(x, ρ) 0<ρ<r ν U(x, ρ)

respectively, have measurable restrictions to A3 . Indeed, it suffices

to show that, ∀x ∈ A3 ,



µ U(x, ρ) µ U(x, ρ)
inf
= inf{
| 0 < ρ < r, ρ ∈ Q} and
0<ρ<r ν U(x, ρ) ν U(x, ρ)



µ U(x, ρ) µ U(x, ρ)
sup
= sup{
| 0 < ρ < r, ρ ∈ Q},
0<ρ<r ν U(x, ρ) ν U(x, ρ)

in which case the asserted measurability follows from the previous

item and from theorem 1.41.(iv). In order to show the equalities
above,
it is enough to prove that, for all x ∈ A3 and 0 < ρ < r,
µ U(x,ρ)

may be arbitrarily approximated by elements of the set
ν U(x,ρ)


µ U(x,ρ)
{
| 0 < ρ < r, ρ ∈ Q}. For that purpose, take a sequence
ν U(x,ρ)
of rationals (ρk )k∈N in (0, ρ) such that ρk ↑ ρ; then, applying the


continuity
from below to the

Borel measures

µ and ν, µ U(x, ρk ) ↑
µ U(x, ρ) and ν U(x, ρk ) ↑ ν U(x, ρ) . Since x ∈ A3 ⊂ spt ν and
 3.2. DIFFERENTIATION THEOREMS 65



µ U(x,ρk )
ν is finite on balls, we have 0 < ν U(x, ρ) < ∞. Hence
→
ν U(x,ρk )


µ U(x,ρ)

, as asserted.
ν U(x,ρ)
4) Due to the fact that µ and ν are locally finite Borel measures,
it follows from lemma 3.14 that Θ∗ν (µ, ·)|A3 = inf r∈Q∗+ ψ r |A3 and
Θν∗ (µ, ·)|A3 = supr∈Q∗+ ψr |A3 . Thus, from the claim in the previous
item and from theorem 1.41.(iv), we conclude that both Θ∗ν (µ, ·)|A3
and Θν∗ (µ, ·)|A3 are measurable.

Exercise 3.16. Show that, in proposition 3.15, the hypothesis of ν
being finite on all closed balls of X may be replaced by the hypothesis
of X being separable.
Hint. Adapt the argument above. Prove that, for each x ∈ A3 ,
there exists an open neighborhood x ∈ U ⊂ X and r0 > 0 such that,
for all 0 < r < r0 the restrictions of ψr and ψ r to U ∩A3 are measurable.
Corollary 3.17. Let µ and ν be locally finite Borel measures on
a metric space X, with ν finite on all closed balls of X. Then the
set Y := {x ∈ X | Θ∗ν (µ, x) = Θν∗ (µ, x)} is Borel measurable and
Θν (µ, ·) : Y → [0, ∞] is Borelian.
In order to obtain similar versions of the comparison 3.6 and upper
density 3.7 theorems to the situation in which the densities of a Borel
measure µ are taken with respect to another Borel measure ν, we need
ν to satisfy the “symmetric Vitali property” introduced below. The
idea is to abstract the Vitali property of the Lebesgue measure stated
in corollary 2.14.
Definition 3.18. Let X be a metric space, F a collection of balls
in X and A ⊂ X. We say that F is a strongly fine cover A, or that
F covers A finely in the strong sense, if F is a cover of A such that,
∀x ∈ A, inf{r > 0 | B(x, r) ∈ F} = 0.
It is clear that every strongly fine cover of A is a fine cover of A in
the sense of definition 2.12, but the converse does not hold.
Definition 3.19 (symmetric Vitali property (SVP)). We say that
a measure µ on a metric space X satisfies the symmetric Vitali property
if, for all A ⊂ X with µ(A) < ∞ and for all F strongly fine cover of A by
nondegenerate closed balls, there exists a countable disjoint subfamily
G ⊂ F such that µ(A \ ∪G) = 0.
66 3. DIFFERENTIATION OF MEASURES

Note that A is not assumed to be µ-measurable.

Remark 3.20.
1) It is clear that, if a measure µ on a metric space X has SVP, so does
x
any restriction of µ, i.e. ∀Y ⊂ X, µ Y has SVP.
2) If a measure µ on a metric space X is σ-finite and has SVP, then
µ is concentrated on its support, i.e. µ(X \ spt µ) = 0. Indeed, let
X = ∪k∈N Ak , with ∀k ∈ N, Ak ∈ σ(µ) and µ(Ak ) < ∞. For each
k ∈ N, the family of nondegenerate

closed balls F = {B(x, r) | x ∈
X \spt µ, r > 0, µ B(x, r) = 0} covers Ak \spt µ finely in the strong
sense. Hence, there exists a countable

disjoint subfamily Gk ⊂ F
such that µ (Ak \ spt µ) \ ∪Gk = 0; since µ(∪Gk ) = 0 (because
Gk is countable and each B ∈ Gk has null measure), we conclude
that µ(Ak \ spt µ) = 0. Therefore X \ spt µ = ∪k∈N (Ak \ spt µ) has
µ-measure zero.
We list in the propositions below some sufficient conditions in order
for a measure to satisfy the symmetric Vitali property.
Proposition 3.21 (doubling property implies SVP). Let X be a
separable metric space and µ a finite Borel regular measure on X. As-
sume that µ satisfies the doubling property:
∃C > 0, ∀B ⊂ X nondegenerate closed ball, µ(5B) ≤ Cµ(B),
where 5B is given by (2.1). Then µ has the symmetric Vitali property.
Proof. Let A ⊂ X and F a fine cover of A by nondegenerate
closed balls. By corollary 2.13, there exists a disjoint subfamily G ⊂ F
such that, for all F ⊂ F finite, A \ ∪B∈F B ⊂ ∪B∈G\F 5B. Since X is
separable, G is countable; let (Bn )n∈N be an enumeration of G. Then,
for all N ∈ N,
A \ ∪Nn=1 Bn ⊂ ∪n≥N +1 5Bn .
∞ N →∞ P∞
Hence, µ(A \ ∪N
P
n=1 Bn ) ≤ C n=N +1 µ(Bn ) −→ 0, since n=1 µ(Bn ) =
µ(∪G) ≤ µ(X) < ∞. Thus, applying the continuity from above 1.11
to the finite measure µ, it follows that µ(A \ ∪G) = 0. 
Remark 3.22. We have actually proved that, if µ is a finite Borel
regular measure on X with the doubling property, then the symmetric
Vitali property holds in a stronger sense, i.e. given A ⊂ X with µ(A) <
∞, the symmetric Vitali property holds for arbitrary fine covers of A,
not necessarily in the strong sense.
Proposition 3.23 (Borel measures on subsets of Rn satisfy SVP).
Let X be a metric subspace of Rn and µ a Borel measure on X. Then
µ satisfies the symmetric Vitali property.
 3.2. DIFFERENTIATION THEOREMS 67

In order to prove this proposition, we will need the following cov-

ering theorem:
Theorem 3.24 (Besicovitch covering theorem). For each n ∈ N,
there exists a natural constant N = N (n), depending only on n, which
satisfies the following property: if F is any family of nondegenerate
closed balls in Rn with sup{diam B | B ∈ F} < ∞ and A is the set of
centers of the balls in F, then exist G1 , . . . , GN such that, for 1 ≤ i ≤ N ,
Gi is a disjoint subfamily of F and ∪N i=1 Gi covers A.

For the proof of this theorem, we refer, for instance, to [EG91],

[KP08], [Mat95] or [Fed69].
Corollary 3.25. Let µ be a Borel measure in Rn , A ⊂ Rn with
µ(A) < ∞ and F a family of nondegenerate closed balls which covers
A finely in the strong sense. Then, for any open set U ⊃ A, there
exists a countable disjoint subfamily G ⊂ F such that ∪G ⊂ U and
µ(A \ ∪G) = 0.
Proof. It is an adaptation of the argument used to prove corollary
2.14, using Besicovitch covering theorem instead of the 5-times covering
lemma 2.10.
Let N = N (n) be the constant given by theorem 3.24 and fix θ ∈
(1 − N1 , 1). We may assume that µ(A) > 0, otherwise the thesis is
trivial. Let U ⊃ A be an open set.
1) Put FU := {B ∈ F | B ⊂ U, diam B ≤ 1}. Since F covers A finely
in the strong sense, it is clear that FU is still a strongly fine cover of
A; in particular, A is contained in the set of centers of the balls in
F. Applying theorem 3.24 to FU , we may take disjoint subfamilies
GU1 , . . . GUN ⊂ FU such that
A ⊂ ∪N i
i=1 (∪GU ). Hence, by subadditivity,
PN
µ(A) ≤ i=1 µ A∩(∪GUi ) . We
therefore conclude that there exists
1 ≤ i ≤ N such that µ A∩(∪GUi ) ≥ N1 µ(A) > (1−θ)µ(A). Since GUi
is a countable family (by remark 2.11) of closed balls, we may apply
x
the continuity from below 1.11 to the Borel measure
µ A to obtain
a finite subfamily G1 ⊂ GUi such that µ A ∩ (∪G1 ) > (1 − θ)µ(A).
But, since ∪G1 is Borelian (hence µ-measurable), we have


µ(A) = µ A ∩ (∪G1 ) + µ(A \ ∪G1 ),
and the fact that µ(A) < ∞ allows us to conclude that µ(A\∪G1 ) <
θµ(A).
2) Given 2 ≤ j ∈ N, assume we have defined finite disjoint subfamilies
G1 ⊂ · · · ⊂ Gj−1 ⊂ F such that, for 1 ≤ i ≤ j − 1, the balls of
Gi are contained in U and µ(A \ ∪Gi ) < θi µ(A). We reapply the
argument of the previous item to the open set U 0 := U \ ∪Gj−1
68 3. DIFFERENTIATION OF MEASURES

in place of U and to A0 := A \ ∪Gj−1 ⊂ U 0 in place of A: take

FU 0 := {B ∈ F | B ⊂ U 0 , diam B ≤ 1}, which is a strongly fine
cover of A0 , and use theorem 3.24 as in the previous
item to find a
disjoint subfamily GU 0 ⊂ FU 0 such that µ A0 ∩ (∪GU 0 ) ≥ N1 µ(A0 ) >
(1 − θ)µ(A0 ). Then, applying the continuity from below to the Borel
x
measure µ A0 , there exists a finite set Gj0 ⊂ GU 0 such that µ A0 ∩
(∪Gj0 ) > (1 − θ)µ(A0 ). As in the previous item, the µ-measurability

of Gj0 and the fact that µ(A0 ) < ∞ imply that µ(A0 \∪Gj0 ) < θµ(A0 ) <
θj µ(A). Put Gj := Gj−1 ∪ Gj0 . Then Gj ⊂ F is a finite disjoint family
whose balls are contained in U , and A \ ∪Gj = A0 \ ∪Gj0 satisfies
µ(A \ ∪Gj ) < θj µ(A).
3) We have thus inductively defined an increasing sequence (Gj )j∈N
such that, for each j ∈ N, Gj is a finite disjoint subfamily of F
whose balls are contained in U , with µ(A \ ∪Gj ) < θj µ(A). Define
G := ∪i∈N Gi ; then G is a disjoint subfamily of F whose balls are
j→∞
contained in U , with ∀j ∈ N, µ(A \ ∪G) ≤ θj µ(A) −→ 0. Hence
µ(A \ ∪G) = 0, which concludes the proof.

Proof of proposition 3.23. If X = Rn , the thesis follows di-
rectly from corollary 3.25.
In the general case, let ρ denote the metric on X induced by the
euclidean metric d on Rn and ι : X → Rn the inclusion. We use
superscripts ρ and d for balls in X and Rn , respectively, so that ∀x ∈ X,
B(x, r)ρ = B(x, r)d ∩ X. Given A ⊂ X with µ(A) < ∞ and F ⊂ 2X
a strongly fine cover of A by nondegenerate closed balls, let F 0 :=
n
{B nondegenerate closed ball in Rn | B ∩ X ∈ F}. Then F 0 ⊂ 2R
is a strongly fine cover of A in Rn ; indeed, ∀x ∈ A, ∀δ > 0, the fact
that F is a strongly fine cover for A in X ensures the existence of
0 < r < δ such that B(x, r)d ∩ X = B(x, r)ρ ∈ F, hence B(x, r)d ∈ F 0
by definition. Since, by proposition 1.15.(iii), the pushforward measure
ι# µ is a Borel measure on Rn , it follows from the case already proved
that there exists a countable disjoint subfamily G 0 ⊂ F 0 such that
ι# µ(A \ ∪G 0 ) = 0. Define G := {B ∩ X | B ∈ G 0 }; then G is a countable
disjoint subfamily of F and, since ∀B ∈ G 0 , A \ B = A \ (B ∩ X), it
follows that A \ ∪G = A \ ∪G 0 , thus µ(A \ ∪G) = ι# µ(A \ ∪G 0 ) = 0,
which concludes the proof. 
Theorem 3.26 (general comparison density theorem). Let µ and ν
be open σ-finite Borel regular measures on a metric space X such that
ν has the symmetric Vitali property, t ≥ 0 and A ⊂ X. If ∀x ∈ A,
Θ∗ν (µ, x) ≥ t then tν(A) ≤ µ(A).
 3.2. DIFFERENTIATION THEOREMS 69

Proof. We assume that t > 0, otherwise the thesis is trivial.

Firstly, assume that ν(A) < ∞.
Fix 0 < τ < t
and an open set U ⊃ A. Since, ∀x ∈ A, Θ∗ν (µ, x) =
µ B(x,r)
lim supr→0
≥ t > τ , it follows that ∀x ∈ A, ∀r > 0, ∃0 < ρ < r
ν B(x,r)


µ B(x,ρ)


such that
> τ , so that µ B(x, ρ) > τ ν B(x, ρ) (note that, in
ν B(x,ρ)
order for the quotient to be > τ , according to our extended arithmetic
convention in 3.12, the numerator cannot be 0 and the denominator
cannot be ∞). It
then follows that

F := {B | ∃x ∈ A, ∃r > 0, B =
B(x, r), µ B(x, ρ) > τ ν B(x, ρ) , B ⊂ U } is a strongly fine cover of
A. Since ν(A) < ∞ and ν has the symmetric Vitali property, we may
take a countable disjoint subfamily G ⊂ F such that ν(A \ ∪G) = 0.
Therefore, by countable subadditivity,
X 
τ ν(A) ≤ τ ν(B) + ν(A \ ∪G) ≤
B∈G
X
≤ µ(B) = µ(∪G) ≤ µ(U ).
B∈G

Since µ is open σ-finite Borel regular, theorem 1.23 may be applied and
yields µ(A) = inf{µ(U ) | U ⊃ A open} ≥ τ ν(A) and, taking τ → t,
the thesis follows in case ν(A) < ∞.
If ν(A) = ∞, the fact that ν is open σ-finite allows us to take a
countable disjoint family (Bk )k∈N of Borel sets in X such that ∪˙ k∈N Bk =
X and ∀k ∈ N, ν(Bk ) < ∞. Thus, for all k ∈ N, the case already proved
applies to A ∩ Bk , which yields µ(A ∩ Bk ) ≥ tν(A ∩ Bk ). By the fact
that both x
P µ A and ν AP x are Borel measures, it then follows that
µ(A) = k∈N µ(A ∩ Bk ) ≥ t k∈N ν(A ∩ Bk ) = tν(A). 
Corollary 3.27. Let µ and ν be open σ-finite Borel regular mea-
sures on a metric space X such that ν has the symmetric Vitali prop-
erty. Then Θ∗ν (µ, x) < ∞ for ν-a.e. x ∈ X.
Proof. Let I := {x ∈ X | Θ∗ν (µ, x) = ∞}. We must show that
ν(I) = 0. Since µ is open σ-finite, we may take a sequence of open sets
(Uk )k∈N such that ∪k∈N Uk = X and ∀k ∈ N, µ(Uk ) < ∞.
Fix k ∈ N and t > 0, and let Akt := {x ∈ Uk | Θ∗ν (µ, x) ≥ t}.
Applying theorem 3.26 with Akt in place of A, it follows that tν(Akt ) ≤
µ(Akt ) ≤ µ(Uk ) < ∞. Since I ∩ Uk = ∩t>0 Akt , we then conclude that
t→∞
∀t > 0, ν(I ∩ Uk ) ≤ ν(Akt ) ≤ t−1 µ(Uk ) −→ 0, hence ν(I ∩ Uk ) = 0. As
I = ∪k∈N (I ∩ Uk ), the thesis follows from the countable subadditivity
of ν. 
70 3. DIFFERENTIATION OF MEASURES

Theorem 3.28 (general upper density theorem). Let µ be a Borel

regular measure on a metric space X, ν an open σ-finite Borel regular
measure on X with the symmetric Vitali property, and A ∈ σ(µ) with
x
µ(A) < ∞. Then Θ∗ν (µ A, x) = 0 for ν-a.e. x ∈ X \ A.
Proof. For each t > 0, let S := {x ∈ X \ A | Θ (µ xA, x) ≥ t}.
t
∗ν

By proposition 1.36.(i), µ xA is a finite Borel regular measure on X;

hence, we may apply theorem 3.26 with µ xA in place of µ, ν and S t
in place of A, which yields tν(S ) ≤ µ xA(S ) = µ(A ∩ S ) = 0, since
t t t
A ∩ S = ∅. Thus, ν(S ) = 0, whence ν({x ∈ X \ A | Θ (µ xA, x) >
t t
∗ν

0}) = ν(∪n∈N S1/n ) = 0. 

Theorem 3.29 (general density theorem). Let µ be an open σ-
finite Borel regular measure on a metric space X with symmetric Vitali
x
property and A ∈ σ(µ). Then the density Θµ (µ A, ·) coincides µ-a.e.
on X with χA , i.e.

(
µ A ∩ B(x, r)
µ
x
Θ (µ A, x) = lim
r→0 µ B(x, r)

=
1 for µ-a.e. x ∈ A,
0 for µ-a.e. x ∈ X \ A.
Proof. It is an adaptation of the proof of corollary 3.9, using
theorem 3.28 instead of theorem 3.7.
Firstly, we prove the case in which µ(X) < ∞. Since µ is con-
centrated on its support, by remark 3.20.2), it suffices to show that
x
Θµ (µ A, ·) coincides µ-a.e. on spt µ with χ
A . Fix x ∈ spt µ and de-

µ A∩B(x,r) µ (X\A)∩B(x,r)
fine f, g : (0, ∞) → [0, ∞] by f (r) =
and g(r) =
.
µ B(x,r) µ B(x,r)
Due to the fact that f (r) + g(r) ≡ 1 and that lim inf r→0 (f + g)(r) ≤
lim inf r→0 f (r) + lim supr→0 g(r) ≤ lim supr→0 (f + g)(r), it follows that
(3.3) Θ∗µ (µ xA, x) + Θ (µ x(X \ A), x) = 1,
µ
∗
and the same holds with X \ A in place of A.
On the other hand, since µ is a finite Borel regular measure and
A ∈ σ(µ), we may apply theorem 3.28 with ν = µ, which yields
x
Θ∗µ (µ A, x) = 0 for µ-a.e. x ∈ X \ A. Then the fact that 0 ≤
x x
Θµ∗ (µ A, ·) ≤ Θ∗µ (µ A, ·) ≤ 1 implies that Θµ (µ A, x) = 0 for x
µ-a.e. x ∈ X \ A. The same holds for X \ A in place of A, i.e.
x
Θµ (µ (X \ A), x) = 0 for µ-a.e. x ∈ A. The last equality implies, in
x
view of (3.3), that Θµ (µ A, x) = 1 for µ-a.e. x ∈ A ∩ spt µ, which
concludes the proof in case µ(X) < ∞.
In the general case, since µ is open σ-finite, we may cover X with
countably many open sets (Uk )k∈N such that ∀k ∈ N, µ(Uk ) < ∞.
For fixed k ∈ N, it follows from proposition 1.36.(i) and from remark
x
3.20.(1) that µ Uk is a finite Borel regular measure with SVP, to
 3.2. DIFFERENTIATION THEOREMS 71

x x
which the case already proved yields Θµ Uk (µ (Uk ∩ A), ·) = χA
x x x
(µ Uk )-a.e. on X. Since Uk is open, the functions Θµ Uk (µ (Uk ∩ A), ·)
x x
and Θµ (µ A, ·) coincide on Uk ; hence, Θµ (µ A, ·) coincides with χA
x
µ-a.e. on Uk . As ∪k∈N Uk = X, we conclude that Θµ (µ A, ·) coincides
with χA µ-a.e. on X, as asserted. 

Corollary 3.30 (general Lebesgue differentiation theorem). Let

µ be an open σ-finite Borel regular measure on a metric space X with
symmetric Vitali property and f : X → C a µ-measurable function
satisfying one of the following conditions:

i) f ∈ L1 (µ) or ´
ii) X is separable and f ∈ L1loc (µ), i.e. ∀x ∈ X, ∃r > 0, B(x,r) |f | dµ <
∞.

Then, for µ-a.e. x ∈ X:

ˆ
1
(3.4) lim
f dµ = f (x).
r→0 µ B(x, r) B(x,r)

Proof. Note that, since f = [(Re f )+ − (Re f )− ] + i[(Im f )+ −

(Im f )− ], it suffices to prove the thesis for positive functions, i.e. we
may assume f : X → [0, ∞). Moreover, the fact that µ is open σ-
finite ensures the existence of a sequence (Uk )k∈N of open sets such
´that ∪k∈N Uk = X´ and ∀k ∈ N, µ(Uk ) < ∞; we may also assume that
x
f d(µ Uk ) = Uk f dµ < ∞ in case hypothesis (ii) holds. Therefore,
if we prove the thesis for finite Borel regular measures with SVP and
hypothesis (i), it will follow that, for each k ∈ N, (3.4) holds with
x
µ Uk in place of µ. In particular, for all k ∈ N, (3.4) holds (with
µ) for µ-a.e. x ∈ Uk , hence it holds for µ-a.e. x ∈ X = ∪k∈N Uk . We
may then assume that µ(X) < ∞ and that hypothesis (i) holds, i.e.
f ∈ L1 (µ).
Fix k ∈ N. The fact that µ is a finite Borel regular measure allows us
to apply Lusin’s theorem 1.117, which yields a closed subset Fk ⊂ X
such that µ(X \ Fk ) < 1/k and f |Fk is continuous. Since µ has the
x
symmetric Vitali property, so does µ Fk ; hence, by remark 3.20,
x
µ Fk is concentrated on its support. In particular, if Nk := Fk ∩ (X \
x

spt µ Fk ) = {x ∈ Fk | ∃r > 0, µ Fk ∩ B(x, r) = 0}, then µ(Nk ) =
0.
For each x ∈ Fk \ Nk , for each r > 0, we have 0 < µ Fk ∩ B(x, r) ≤
72 3. DIFFERENTIATION OF MEASURES


µ B(x, r) , so that
ˆ
1

f dµ =
µ B(x, r) Fk ∩B(x,r)
(3.5)
ˆ
µ Fk ∩ B(x, r) 1
=
·
f dµ.
µ B(x, r) µ Fk ∩ B(x, r) Fk ∩B(x,r)


µ Fk ∩B(x,r) r→0
From the general density theorem 3.29, for µ-a.e. x ∈ Fk ,
−→
µ B(x,r)
1
´ r→0
1, and the continuity of f |Fk ensures ∀x ∈ Fk ,
F ∩B(x,r) f dµ −→
k
µ Fk ∩B(x,r)
f (x). It then follows from (3.5) that, adjoining a µ-null set to Nk if
necessary, ∀x ∈ Fk \ Nk ,
ˆ
1 r→0
(3.6)
f dµ −→ f (x).
µ B(x, r) Fk ∩B(x,r)
We contend that, for µ-a.e. x ∈ Fk ,
ˆ
1 r→0
(3.7)
f dµ −→ 0.
µ B(x, r) B(x,r)\Fk
Indeed, let ν be the Borel regular measure
´ on X given by f dµ, i.e. the
extension of the measure A ∈ B X 7→ A f dµ given by theorem 1.8.
Since Fkc = X \ Fk has finite ν-measure (because f ∈ L1 (µ)), we can
apply the general upper density theorem 3.28 with ν in place of µ, µ
in place of ν and Fkc in place of A, thus proving our contention.
It then follows from (3.6) and (3.7) that, adjoining another µ-null
set to Nk if necessary, ∀x ∈ Fk \ Nk , (3.4) holds. Therefore, as k ∈ N
was arbitrarily taken and X \ ∪k∈N Fk is µ-null, (3.4) holds for x in the
complement of the µ-null set (∪k∈N Nk )∪(X \∪k∈N Fk ) and we are done.

Corollary 3.31 (Lebesgue Points). Let X be a separable metric
space, µ an open σ-finite Borel regular measure on X with symmetric
Vitali
´ property, 1 ≤ p < ∞ and f ∈ Lploc (µ), i.e. ∀x ∈ X, ∃r >
0, B(x,r) |f |p dµ < ∞. Then, for µ-a.e. x ∈ X,
ˆ
1
(3.8) lim
|f (y) − f (x)|p dµ(y) = 0.
r→0 µ B(x, r) B(x,r)

Proof. Let {ri | i ∈ N} be a countable dense subset of C. It

follows from corollary 3.30 that, for every i ∈ N, there exists a µ-null
set Ai such that, for all x ∈ Aci ,
ˆ
1
lim
|f − ri |p dµ = |f (x) − ri |p .
r→0 µ B(x, r) B(x,r)
 3.2. DIFFERENTIATION THEOREMS 73

Then the above equality holds for all i ∈ N and for all x in the com-
plement of the µ-null set A := ∪i∈N Ai .
Fix x ∈ Ac and  > 0. There exists i ∈ N such that |f (x) − ri | < .
Then
ˆ
1

|f (y) − f (x)|p dµ(y) ≤
µ B(x, r) B(x,r)
ˆ
2p−1
|f (y) − ri |p + |ri − f (x)|p dµ(y) ≤


≤

µ B(x, r) B(x,r)
ˆ
2p−1
≤
|f (y) − ri |p dµ(y) + 2p−1 |f (x) − ri |p ,
µ B(x, r) B(x,r)
so that
ˆ
1
lim sup
|f (y)−f (x)|p dµ(y) ≤ 2·2p−1 |f (x)−ri |p < 2p p .
r→0 µ B(x, r) B(x,r)

Since  > 0 was arbitrarily taken, the thesis follows. 

Definition 3.32 (Lebesgue Points). With the same notation from
the previous corollary, a point x ∈ X for which (3.8) holds is called
Lebesgue point of f with respect to µ.
It is clear that every point of continuity of f is a Lebesgue point of
f.
If X = Rn and µ = Ln , the limit in (3.8) can be taken along
all closed balls B containing x (not necessarily centered at x) with
diam B → 0:
Corollary 3.33 (Lebesgue points with noncentered balls). Let
1 ≤ p < ∞ and f ∈ Lploc (Ln ). Then, for each Lebesgue point x of f
with respect to Ln (in particular, for Ln -a.e. x ∈ Rn ),
ˆ
1
lim |f (y) − f (x)|p dLn (y) = 0,
B↓{x} Ln (B) B

where the limit is taken over all closed balls B containing x with diam B →
0.
Proof. For each closed ball B containing x, we have:
ˆ
1
|f (y) − f (x)|p dLn (y) ≤
Ln (B) B
ˆ
1
≤ n |f (y) − f (x)|p dLn (y) =
L (B) B(x,diam B)
ˆ
n 1
=2 n
|f (y) − f (x)|p dLn (y)
L B(x, diam B) B(x,diam B)
74 3. DIFFERENTIATION OF MEASURES

and then the thesis follows from corollary 3.31. 

Our next step is to prove a version of the general comparison density
theorem 3.26 for lower densities.
Firstly we introduce for Borel outer measures the notions of abso-
lute continuity and mutual singularity which were introduced in 1.92
for measures on a σ-algebra M.
Definition 3.34 (absolute continuity and mutual singularity). Let
µ and ν be Borel measures on a topological space X. We say that:
1) µ is absolutely continuous with respect to ν (notation: µ  ν) if
∀A ⊂ X, ν(A) = 0 implies µ(A) = 0.
2) µ and ν are mutually singular (notation: µ ⊥ ν) if there exists
A ∈ B X such that µ is concentrated on A and ν is concentrated on
X \ A.
Remark 3.35. Note that µ ⊥ ν iff µ|BX ⊥ ν|BX in the sense of
definition 1.92. Besides, it is clear that, if µ is a Borel measure and
ν is a Borel regular measure on a topological space X, then µ  ν
iff ∀A ∈ B X , ν(A) = 0 implies µ(A) = 0. Thus, if ν is Borel regular,
then µ  ν iff µ|BX  ν|BX in the sense of definition 1.92.
We now prove a version of the Lebesgue decomposition theorem
1.101 for outer measures. The lemma below may be obtained as a
direct consequence of the previous remark and theorem 1.101, but we
give a direct proof.
Lemma 3.36 (Lebesgue decomposition theorem). Let µ be a σ-finite
Borel measure and ν a Borel regular measure on a metric space X.
Then there exists B ∈ B X such that ν is concentrated on B c and
x
µ B c  ν, so that
(LD) µ=µ xB + µ xB , c
µ xB ⊥ ν, µ xB c
 ν.
Moreover:
1) B ∈ B X satisfying (LD) is unique up to µ-null sets, i.e. if B 0 ∈ B X
also satisfies (LD), then B ∆ B 0 is µ-null.
2) the decomposition (LD) is unique in the sense that, if µ = µs + µa
x x
with µs ⊥ ν and µa  ν, then µs = µ B and µa = µ B c .
Definition 3.37. With the notation above, we call µ xB the sin-
gular part and µ xB the absolutely continuous part of µ with respect
c

to ν.
Proof. 1) Assume µ finite. Let F := {F ∈ B X | ν(F ) = 0}
and α := sup{µ(F ) | F ∈ F}, so that 0 ≤ α ≤ µ(X) < ∞. We
 3.2. DIFFERENTIATION THEOREMS 75

contend that this sup is attained. Indeed, take a sequence (Fn )n∈N
in F such that µ(Fn ) → α and define B := ∪n∈N Fn . Then B ∈ B X
and ν(B) = 0 (since each Fn is ν-null), so that B ∈ F. Since
∀n ∈ N, µ(Fn ) ≤ µ(B) and µ(Fn ) → α, it follows α ≤ µ(B), hence
α = µ(B).
x
We contend that µ B c  ν. Indeed, if that is not the case,
the Borel regularity of ν and remark 3.35 imply the existence of
x
A ∈ B X such that ν(A) = 0 and µ B c (A) > 0; hence B ∪ A ∈
F and µ(B ∪ A) = µ(B) + µ(B c ∩ A) = α + µ B c (A) > α, x
which contradicts the definition of α. The contention is then proved,
x
so that ν is concentrated on B c and µ B c  ν. Besides, since
x x
µ B is trivially concentrated on B, we have µ B ⊥ ν and the
decomposition (LD) holds.
2) In the general case, since µ is σ-finite, we can take an increasing
sequence (An )n∈N in σ(µ) such that ∪n∈N An = X and ∀n ∈ N,
x
µ(An ) < ∞. For each n ∈ N, µ An is a finite Borel measure,
to which the previous item can be applied, yielding Bn ∈ B X
x
such that ν is concentrated on Bnc and µ An ∩ Bnc  ν. Let
B := ∪n∈N Bn ∈ B X . Then ν(B) = 0, i.e. ν is concentrated on B c ;
x
we contend that µ B c  ν, which yields the validity of decompo-
sition (LD). Indeed, if F ∈ B X is ν-null, ∀n ∈ N, µ A
c c
x c

n ∩ Bn (F ) =
µ(An ∩ Bn ∩ F ) = 0,
hence µ An ∩ (∩n∈N Bn ) ∩ F = 0. Since
c
An ∩ (∩n∈N Bn ) ∩ F n∈N is a sequence in σ(µ) which increases to


(∩n∈N Bnc ) ∩ F = B c ∩ F , the continuity from below 1.11 applied to
x
µ yields µ B c (F ) = µ(B c ∩ F ) = 0, which proves our contention.
3) Let B 0 ∈ B X such that (LD) also holds with B 0 in place of B.
x
Then µ B \ B 0 is concentrated on B and absolutely continuous
x
with respect to ν (which is null on B), hence µ B \ B 0 is the
x x
null measure. Similarly, µ B 0 \ B is null, and so is µ B ∆ B 0 =
x x
µ B \ B 0 + µ B 0 \ B, i.e. µ(B ∆ B 0 ) = 0, as asserted.
4) Let ν = µs + µa with µs ⊥ ν and µa  ν. Let B 0 ∈ B X such that µs
is concentrated on B 0 and ν is concentrated on (B 0 )c . Then µa  ν
is also concentrated on (B 0 )c ; thus, for all A ⊂ X, µs (A)
= µs (A ∩
x
B 0 ) =
µ(A ∩ B 0 ) = µ B 0 (A) and µa (A) = µa A ∩ (B 0 )c = µ A ∩
x x
(B 0 )c = µ (B 0 )c (A). That is, µa = µ B 0 and µs = µ (B 0 )c .x
In particular, (LD) holds with B 0 in place of B; by the previous
x
item, it follows that µ(B ∆ B 0 ) = 0, hence µs = µ B 0 = µ B x
x x
and µa = µ (B 0 )c = µ B c , as asserted.

76 3. DIFFERENTIATION OF MEASURES

Theorem 3.38 (comparison theorem for lower densities). Let µ

and ν be open σ-finite Borel regular measures on a metric space X,
t ≥ 0 and A ⊂ X with ∀x ∈ A, Θν∗ (µ, x) ≤ t.

i) If µ has SVP, then µ(A) ≤ t ν(A).

ii) If ν has SVP and B is given by lemma 3.36, so that (LD) holds,
then µ(A \ B) ≤ t ν(A).

Proof. It is similar to the proof of theorem 3.26.

(1) Firstly, we make a reduction: it is enough to prove the case in

which both µ(A) and ν(A) are finite. Indeed, suppose that the
thesis holds in that case. In the general case, since both µ and
ν are σ-finite, there exists a disjoint sequence (Bk )k∈N of Borel
sets in X such that µ(Bk ) < ∞, ν(Bk ) < ∞ and X = ∪˙ k∈N Bk .
Thus, for each k ∈ N, the thesis holds for Bk ∩ A in place of
A, so that, for all k ∈ N, µ(Bk ∩ A) ≤ t ν(Bk ∩ A) in case i)
and µ(Bk ∩ A \ B) ≤ t ν(Bk ∩ A) in case ii). By the fact that
x x
both µ PA and ν A are Borel P measures, it then follows that
µ(A) = k∈N µ(BkP ∩ A) ≤ t k∈N ν(Bk ∩ A) P= t ν(A) in case
i), and µ(A \ B) = k∈N µ(Bk ∩ A \ B) ≤ t k∈N ν(Bk ∩ A) =
t ν(A) in case ii).
Assume, therefore, µ(A) < ∞ and ν(A) < ∞.
(2) It is enough to prove part i) for A ⊂ spt µ. Indeed, suppose
that the thesis holds in that case. Since µ is σ-finite and
has SVP, it follows from remark 3.20 that µ is concentrated
on its support; thus, for arbitrary A it follows that µ(A) =
µ(A ∩ spt µ) ≤ t ν(A ∩ spt µ) ≤ t ν(A) and we are done.
Assume, therefore, A ⊂ spt µ. Fix τ > t and an open

ν µ B(x,r)
set U ⊃ A. Since, ∀x ∈ A, Θ∗ (µ, x) = lim inf r→0

ν B(x,r)
≤ t < τ , it
follows that ∀x ∈ A, ∀r > 0, ∃0 < ρ < r such
µ B(x,ρ)


that
< τ , so that µ B(x, ρ) < τ ν B(x, ρ) (note
ν B(x,ρ)


that µ B(x, ρ) > 0, since x ∈ spt µ; hence, in order for the
quotient to be < τ , according to our extended arithmetic con-
vention in 3.12, the denominator cannot be 0). It then follows


that F :=
{B | ∃x ∈ A, ∃r > 0, B = B(x, r), µ B(x, r) <
τ ν B(x, r) , B ⊂ U } is a strongly fine cover of A. Since
µ(A) < ∞ and µ has the symmetric Vitali property, we may
take a countable disjoint subfamily G ⊂ F such that µ(A \
 3.2. DIFFERENTIATION THEOREMS 77

∪G) = 0. Therefore, by countable subadditivity,

X X
µ(A) ≤ µ(A \ ∪G) + µ(B) < τ ν(B) =
B∈G B∈G
= τ ν(∪G) ≤ τ ν(U )
Since ν is open σ-finite Borel regular, theorem 1.23 may be
applied and yields τ ν(A) = inf{τ ν(U ) | U ⊃ A open} ≥ µ(A)
and, taking τ → t, part i) is proved.
(3) We now prove ii). Take B ∈ B X given by lemma 3.36, so
x
that (LD) holds. Note that, since the measure µ B c is abso-
lutely continuous with respect to ν, it clearly has SVP; besides,
it is trivially open σ-finite, it is Borel regular by proposition
1.36.(i), and ∀x ∈ A:

x
Θν∗ (µ B c , x) = lim inf
µ x B c
B(x,


r)


≤
r→0 ν B(x, r)


µ B(x, r)
≤ lim inf
= Θν∗ (µ, x) ≤ t.
r→0 ν B(x, r)

x
We may therefore apply part i) with µ B c in place of µ,
x
yielding µ(A \ B) = µ B c (A) ≤ t ν(A), as asserted.

Theorem 3.39 (differentiation theorem for Borel measures on met-
ric spaces). Let µ and ν be open σ-finite Borel regular measures on a
metric space X. Suppose that X is separable or that ν is finite on
closed balls of X.
i) The set Y := {x ∈ X | Θ∗ν (µ, x) = Θν∗ (µ, x)} is Borel measurable
and Θν (µ, ·) : Y → [0, ∞] is Borelian.
ii) If ν has SVP, Yf := {x ∈ Y | Θν (µ, x) < ∞} is a Borel measurable
subset of X whose complement is ν-null.
iii) If both µ and ν have SVP, µ(Y c ) = ν(Y c ) = 0.
Proof.
1) If ν is finite on closed balls, part (i) follows from corollary 3.17; if
X is separable, part (i) is a direct consequence of exercise 3.16.
2) We now prove part (iii). It suffices to prove the case in which
µ(X) < ∞ and ν(X) < ∞. Indeed, assuming that the thesis holds
in this case, in the general case we can take a sequence (Uk )k∈N
of open sets such that X = ∪k∈N Uk and ∀k ∈ N, µ(Uk ) < ∞
and ν(Uk ) < ∞. The thesis then holds, for each k ∈ N, for
x
the finite Borel regular measures with SVP µ Uk and ν Uk x
78 3. DIFFERENTIATION OF MEASURES

in place of µ and ν, respectively. Thus, for each k ∈ N, since

x x x
Θ∗ν (µ, ·) = Θ∗ν Uk (µ Uk , ·) and Θν∗ (µ, ·) = Θν∗ Uk (µ Uk , ·) x
on the open set Uk , it follows that µ(Y c ∩ Uk ) = 0 = ν(Y c ∩ Uk ),
whence µ(Y c ) = 0 = ν(Y c ), as asserted.
Assume, therefore, that both µ and ν are finite. Let 0 < a <
b < ∞ and Ya,b := {x ∈ X | Θν∗ (µ, x) ≤ a and Θ∗ν (µ, x) ≥ b}. Since
both µ and ν are open σ-finite Borel regular measures with SVP,
we may apply theorems 3.26 and 3.38 to conclude that bν(Ya,b ) ≤
µ(Ya,b ) and µ(Ya,b ) ≤ aν(Ya,b ), so that µ(Ya,b ) ≤ aν(Ya,b ) ≤ ab µ(Ya,b ).
Since ab < 1 and µ(Ya,b ) < ∞, it follows that µ(Ya,b ) = 0, hence
ν(Ya,b ) = 0. As Y c = ∪{Ya,b | 0 < a < b < ∞, a ∈ Q, b ∈ Q}, it
follows that µ(Y c ) = ν(Y c ) = 0, as asserted.
3) We prove part (ii). It is clear from part (i) that Yf = {x ∈ Y |
Θν (µ, x) < ∞} is Borel measurable. By the same reduction made
in the previous item, we may assume that both µ and ν are fi-
nite. Take B ∈ B X given by lemma 3.36, so that (LD) holds.
x
Note that µ B c is a finite Borel regular measure on X with SVP.
x
Applying part (iii) with µ B c in place of µ, we conclude that
x x
{x ∈ X | Θ∗ν (µ B c , x) 6= Θν∗ (µ B c , x)} is ν-null. On the other
hand, as µ(B) < ∞, we may apply theorem 3.28 with B in place
x
of A, yielding Θ∗ν (µ B, x) = 0 for ν-a.e. x ∈ B c . That im-
x x
plies Θ∗ν (µ B c , x) = Θ∗ν (µ, x) and Θν∗ (µ B c , x) = Θν∗ (µ, x)
for ν-a.e. x ∈ B c . It then follows that Y c ∩ B c = {x ∈ B c |
x
Θ∗ν (µ, x) 6= Θν∗ (µ, x)} differs from {x ∈ B c | Θ∗ν (µ B c , x) 6=
x
Θν∗ (µ B c , x)} by a ν-null set; thus, since the latter set is ν-null, so
is the former. Since ν is concentrated on B c , we then conclude that
ν(Y c ) = ν(Y c ∩ B c ) = 0. Finally, by corollary 3.27, F := {x ∈ X |
Θ∗ν (µ, x) < ∞} has ν-null complement; as Yf = Y ∩ F , we conclude
that Yfc = Y c ∪ F c is ν-null, as asserted.

Theorem 3.40 (Lebesgue-Besicovitch-Radon-Nikodym differentia-
tion theorem). Let µ and ν be open σ-finite Borel regular measures on
a metric space X. Suppose that X is separable or that ν is finite on
closed balls of X, and that ν has SVP.
i) Let µ = µs + µa be the Lebesgue decomposition of µ with respect to
x x
ν, i.e. µs = µ B and µa = µ B c , where B ∈ B X is given by
lemma 3.36. Then, for all A ∈ B X ,
ˆ
(3.9) µa (A) = Θν (µ, x) dν(x),
A
´
so that, for all A ∈ B X , µ(A) = A Θν (µ, x) dν(x) + µs (A).
 3.2. DIFFERENTIATION THEOREMS 79

ii) If µ also has SVP, in lemma 3.36 we can take B 0 = {x ∈ X |

Θν (µ, x) = ∞} in place of B.
Proof. i) Note that the integral in (3.9) makes sense, since, by
theorem 3.39.(ii), Θν (µ, ·) is a positive Borel measurable function
defined on the complement of a ν-null set.
Let λ = Θν (µ, ·) dν be the Borel regular measure on X defined
by the second´ member in (3.9), i.e. the extension of the measure
A ∈ B X 7→ A Θν (µ, x) dν(x) given by theorem 1.8. We must show
that µa = λ. Since both measures are Borel regular, it suffices to
show that they coincide on Borel sets.
Let A ∈ B X , fix t > 1 and take Yf given by 3.39.(ii), so that
ν(B ∪ Yfc ) = 0. We contend that both λ and µa are concentrated
on S := {x ∈ B c ∩ Yf | Θν (µ, x) > 0} ∈ B X . Indeed, since
S c = B ∪ Yfc ∪ {x ∈ B c ∩ Yf | Θν (µ, x) = 0}, it is clear that
λ(S c ) = 0. On the other hand applying theorem 3.38.(ii) with
{x ∈ B c ∩ Yf | Θν (µ, x) = 0} in place of
A and t = 0, it follows
that µa {x ∈
B c ∩ Yf | Θν (µ, x) = 0} = µ {x ∈ B c ∩ Yf |
Θν (µ, x) = 0} = 0 and, as µa (B ∪ Yfc ) = 0 (because µa  ν), we
conclude that µa (S c ) = 0.
Define, ∀k ∈ Z, Ak := {x ∈ A ∩ S | tk ≤ Θν (µ, x) < tk+1 } ∈
B X . Since A ∩ S = ∪˙ k∈Z Ak , we have:
X
µa (A) = µa (A ∩ S) = µ(A ∩ S) = µ(Ak ),
k∈Z
(3.10) X
λ(A) = λ(A ∩ S) = λ(Ak ).
k∈Z

On the other hand, for all k ∈ Z:

3.26 3.38.(ii)
tk ν(Ak ) ≤ µ(Ak ) ≤ tk+1 ν(Ak ),
(3.11)
tk ν(Ak ) ≤ λ(Ak ) ≤ tk+1 ν(Ak ).
From (3.10) and (3.11) we then conclude that:
X X
µa (A) = µ(Ak ) ≤ t tk ν(Ak ) ≤ tλ(A)
k∈Z k∈Z
X X
λ(A) = λ(Ak ) ≤ t tk ν(Ak ) ≤ tµa (A).
k∈Z k∈Z

Since t > 1 was arbitrarily taken, we can make t ↓ 1 to conclude

µa (A) ≤ λ(A) and λ(A) ≤ µa (A), hence µa (A) = λ(A), as asserted.
ii) Since B 0 ⊂ Yfc , it follows from theorem 3.39.(ii) that ν(B 0 ) = 0,
i.e. ν is concentrated on (B 0 )c . Therefore, it is enough to prove
80 3. DIFFERENTIATION OF MEASURES

x
that µ (B 0 )c  ν. Indeed, let A ⊂ X such that
ν(A) = 0.
x 0 c 0 c
We must show that µ (B ) (A) = µ (B ) ∩ A = 0. Since
(B 0 )c = {x ∈ X | Θν∗ (µ, x) < ∞} = ∪n∈N {x ∈ X | Θν∗ (µ, x)
≤ n},
it suffices to show that ∀n ∈ N, µ A∩{x ∈ X | Θν∗ (µ, x) ≤ n} = 0.
But, as µ has SVP, we may apply
theorem 3.38.(i), which yields
ν ν
µ A
∩ {x ∈ X | Θ∗ (µ, x) ≤ n} ≤ nν A ∩ {x ∈ X | Θ∗ (µ, x) ≤
n} ≤ nν(A) = 0, whence the thesis.

Corollary 3.41. With the same hypothesis from theorem 3.40,
d(µa | )
Θ (µ, ·) coincides ν-a.e. with the Radon-Nikodym derivative d(ν|BBX) .
ν
X

Proof. It is a direct consequence of (3.9) and the uniqueness of

the Radon-Nikodym derivative stated in theorem 1.103. 
 CHAPTER 4

Rn -valued Radon Measures

4.1. Linear functionals on spaces of continuous functions

In this section we fix a locally compact Hausdorff space X, which
will be assumed σ-compact, unless otherwise specified. We aim to study
the representation of continuous linear functionals on certain spaces of
continuous functions on X by means of vector valued Radon measures.
Notation. We denote by
• Cc (X, Rn ) the space of continuous functions f : X → Rn with
spt f compact;
• C0 (X, Rn ) the space of continuous functions f : X → Rn which
vanish at infinity, i.e. such that ∀ > 0, ∃K ⊂ X compact such
that kf k <  on X \ K.
• Cb (X, Rn ) the space of bounded continuous functions f : X →
Rn .
Endowed with the norm of uniform convergence, i.e. kf ku :=
sup{kf (x)k | x ∈ X}, Cb (X, Rn ) is a Banach space. As it can be read-
ily verified by means of Urysohn’s lemma for locally compact Hausdorff
spaces (see lemma 4.5, below), C0 (X, Rn ) is the closure of Cc (X, Rn ) in
Cb (X, Rn ); in particular, C0 (X, Rn ) is itself a Banach space with the
norm of uniform convergence.
Definition 4.1 (Rn -valued Radon measures). We say that a linear
functional µ : Cc (X, Rn ) → R is an Rn -valued Radon measure on X if,
for each compact K ⊂ X, the restriction of µ to CKc (X, Rn ) := {f ∈
Cc (X, Rn ) | spt f ⊂ K}, endowed with k·ku , is linear continuous; that
is, if ∃CK ≥ 0 such that
(LF cont) sup{µ · f | f ∈ CKc (X, Rn ), kf ku ≤ 1} ≤ CK .
If the condition above holds with a constant C ≥ 0 which does not
depend on K, i.e. if µ is linear continuous on Cc (X, Rn ) endowed with
k·ku , we call µ a finite Rn -valued Radon measure on X.
Remark 4.2.
81
82 4. RN -VALUED RADON MEASURES

1) We will identify Rn -valued Radon measures on X with set func-

tions on X, as the name “measure” indicates, after we prove Riesz
representation theorem for Radon measures 4.9.
2) The definition adopted for an Rn -valued Radon measure on X is
equivalent to saying that µ : Cc (X, Rn ) → R is linear contin-
uous with respect to the natural topological vector space topol-
ogy on Cc (X, Rn ), which is an inductive limit of Fréchet spaces
(an LF space for short). It is actually the countable strict in-
ductive limit (thanks to
the σ-compactness of X) of the Banach
spaces CKc (X, Rn ), k·ku | K ⊂ X compact ; its topology is the
strongest locally convex topology on Cc (X, Rn ) which makes all in-
clusions CKc (X, Rn ) → Cc (X, Rn ) continuous, for K ⊂ X compact.
With such a topology, given a locally convex space Y , a linear map
Cc (X, Rn ) → Y is continuous iff ∀K ⊂ X compact, its restriction
to CKc (X, Rn ) is continuous (as we defined in the case Y = R). We
don’t suppose the reader to have any prior knowledge on locally
convex spaces, but if he or she wants to delve into some of the de-
tails which may be left behind the scenes, we suggest: [Con90],
chapter IV, for a brief overview of locally convex spaces; [Osb14],
for a gentle introduction to locally convex spaces; [K6̈9], [SW99],
[Tre06] or [Bou87] for the heavy stuff.
3) For those fluent in locally convex spaces: the LF topology of Cc (X, Rn )
introduced in the previous item coincides with the product topol-
ogy of the LF spaces Cc (X, R), i.e. we may identify Cc (X, Rn ) ≡
Cc (X, R)n as topological vector spaces. Indeed, the continuity of
Cc (X, Rn ) → Cc (X, R)n , f 7→ (f1 , . . . , fn ), is clear; the continuity
of its inverse can be verified using the facts that it maps bounded
sets to bounded sets, Cc (X, Rn ) is locally convex and Cc (X, R)n is
bornological.
If X is an open set in some Euclidean space, C∞ n
c (X, R) with
its LF topology (i.e. the topology induced by the family of Fréchet
spaces {f ∈ C∞ n
c (X, R) | spt f ⊂ K}, for each K ⊂ X compact) has
a continuous dense inclusion in Cc (X, Rn ) ≡ Cc (X, R)n . That means
that the dual of Cc (X, R)n may be identified with a linear subspace
of the dual of C∞ n n
c (X, R) , i.e. every R -valued Radon measure on
n
X is an R -valued Schwartz distribution on X.
Exercise 4.3 (Rn -valued Radon measures on open sets of Eu-
clidean spaces).
a) Let X be a locally compact separable metric space and (Uk )k∈N
be an increasing sequence of relatively compact open subsets of X
such that ∪k∈N Uk = X. Then a linear map µ : Cc (X, Rn ) → R is
 4.1. LINEAR FUNCTIONALS ON SPACES OF CONTINUOUS FUNCTIONS 83

continuous, i.e. it is an Rn -valued Radon measure on X, iff ∀k ∈ N,

µ|
is continuous (we identify C (Uk , Rn ) with the linear
n
Cc (Uk ,R ),k·ku c

subspace of Cc (X, Rn ) formed by the functions with support in Uk ).

m ∞ n
b) Let U be an open subset
of R . Then Cc (U, R ) is sequentially
n
dense in Cc (U, R ), k·ku .
Hint. Use the standard mollifier 1.112, proposition 1.108 and
theorem 1.111.
c) Let X be an open subset of Rm and (Uk )k∈N be an increasing se-
quence of relatively compact open subsets of X such that ∪k∈N Uk =
X. Let µ : C∞ n
c (X, R ) → R be a linear map such that ∀k ∈ N,
µ| ∞ n

is continuous. Then µ may be uniquely extended
Cc (Uk ,R ),k·ku
to a continuous linear map Cc (X, Rn ) → R.
Hint. Use the two previous items.
Remark 4.4 (Rn -valued Radon measures on open sets of Euclidean
spaces). In view of part c) of the previous exercise, we may identify Rn -
valued Radon measures on open subsets X of Euclidean spaces with
linear functionals µ : C∞ n
c (X, R ) → R such that, for each compact
subset K ⊂ X, the restriction of µ to {f ∈ C∞ n
c (X, R) | spt f ⊂ K}
is continuous with respect to the topology of uniform convergence (i.e.
given by the norm k·ku ).
We recall more preliminaries from Real Analysis in order to prove
the version of Riesz representation theorem for Rn -valued Radon mea-
sures 4.9 below.
Notation. Let X be a locally compact Hausdorff space, U ⊂ X
open and f a function on X. The notation f ≺ U means that 0 ≤ f ≤
1, f ∈ Cc (X, R) and spt f ⊂ U .
Lemma 4.5 (Urysohn’s lemma for LCH). If X is a locally compact
Hausdorff space, U ⊂ X open and K ⊂ U compact, then there exists
f ∈ Cc (X, R) such that χK ≤ f ≺ U .
Theorem 4.6 (Tietze’s extension theorem for LCH). If X is a
locally compact space, K ⊂ X compact and f : K → R continuous,
then f admits a continuous extension fe : X → R. Moreover, we may
take fe with compact support and, if f is bounded, we may also take fe
such that kfeku = kf ku .
Theorem 4.7 (Riesz representation theorem for positive linear
functionals). Let X be a locally compact Hausdorff space and L : Cc (X, R) →
84 4. RN -VALUED RADON MEASURES

R a positive linear functional, i.e. L is linear and L · f ≥ 0 whenever

f ≥ 0. Then there exists a unique Radon ´ measure η on X which rep-
resents L, i.e. ∀f ∈ Cc (X, R), L · f = f dη. Moreover, on open sets
η is given by
η(U ) = sup{L · f | f ≺ U }.
For the proof of 4.5, 4.6 (which are direct consequences of the cor-
responding versions of those theorems for normal spaces) and 4.7 we
refer the reader, for instance, to [Fol99] or [Rud87].
Remark 4.8. Every positive linear functional on Cc (X, R) is an
R-valued Radon measure on X, i.e. positivity implies continuity on
Cc (X, R). Indeed, given K ⊂ X compact, take Φ ∈ Cc (X, R) given by
lemma 4.5 such that χK ≤ Φ ≺ X. For all f ∈ CKc (X, R) with f 6= 0,
we have kf|fk|u ≤ Φ, so that Φ± kffku ≥ 0 and Φ± kffku ∈ Cc (X, R). Hence
L(f )
0 ≤ L Φ ± kffku = L(Φ) ± kf


ku
, which implies |L(f )| ≤ L(Φ)kf ku . The
continuity condition (LF cont) is then satisfied with CK := L(Φ).
Theorem 4.9 (Riesz representation theorem for Radon measures).
Let X be a σ-compact locally compact Hausdorff space and µ : Cc (X, Rn ) →
R an Rn -valued Radon measure on X. Then there exists a unique
Radon measure λ on X and a Borel measurable map ν : X → Rn unique
up to λ-null sets such that kνk = 1 λ-a.e. on X and ∀f ∈ Cc (X, Rn ),
ˆ
(4.1) µ · f = hf, νi dλ,

where h·, ·i denotes the Euclidean inner product in Rn . Moreover,

i) ∀U ⊂ X open,
(4.2) λ(U ) = sup{µ · f | f ∈ Cc (X, Rn ), kf k ≺ U }.
ii) µ is a finite Rn -valued Radon measure iff λ is a finite Radon mea-
sure; if that is the case, kµkC0 (X,Rn )∗ = λ(X).
Remark 4.10. Note that, in (4.2), sup{µ·f | f ∈ Cc (X, Rn ), kf k ≺
U } = sup{|µ · f | | f ∈ Cc (X, Rn ), kf k ≺ U }. Indeed, if f ∈ Cc (X, Rn )
and kf k ≺ U , so does −f , and µ · (−f ) = −µ · f , hence either µ · f or
µ · (−f ) coincides with |µ · f |.
Proof.
1) (Existence) Let C+ c (X):= {f ∈ Cc (X, R) | f ≥ 0}. Define L :
Cc (X) → [0, ∞) by f 7→ sup{µ · φ | φ ∈ Cc (X, Rn ), kφk ≤ f }. Note
+

that L is well-defined, i.e. the sup is indeed ≥ 0 (since µ · 0 = 0)

and finite, due to the continuity condition (LF cont): if f 6= 0,
∀φ ∈ Cc (X, Rn ) with kφk ≤ f , we have ψ := kfφku ∈ Csptc
f
(X, Rn )
4.1. LINEAR FUNCTIONALS ON SPACES OF CONTINUOUS FUNCTIONS 85

and kψku = kφk u

kf ku
≤ 1, hence µ · ψ ≤ Cspt f , i.e. µ · φ ≤ Cspt f kf ku ,
showing that L(f ) ≤ Cspt f kf ku .
We contend that L is additive and 1-homogeneous, i.e. ∀f, g ∈
C+c (X), ∀c ≥ 0, L(f + g) = L(f ) + L(g) and L(cf ) = cL(f ). The
1-homogeneity is clear, since, for c = 0 the equality is trivial and for
−1
c > 0 and f ∈ C+ c (X), we have kφk ≤ cf iff kc φk ≤ f , so that {µ ·
φ | φ ∈ Cc (X, Rn ) | kφk ≤ cf } = c{µ · φ | φ ∈ Cc (X, Rn ) | kφk ≤ f }.
To prove the additivity, let f, g ∈ C+ n
c (X). If φ, ψ ∈ Cc (X, R ) satisfy
kφk ≤ f and kψk ≤ g, then φ + ψ ∈ Cc (X, Rn ) and kφ + ψk ≤ f + g,
so that µ · φ + µ · ψ = µ · (φ + ψ) ≤ L(f + g), and taking the sup
over all such φ and ψ we conclude that L(f ) + L(g) ≤ L(f + g). It
remains to prove the reverse inequality. Given φ ∈ Cc (X, Rn ) such
that kφk ≤ f + g, define:
( (
f g
φ iff + g > 0, φ iff + g > 0,
φ1 := f +g and φ2 := f +g
0 iff + g = 0, 0 iff + g = 0.
Note that, for i = 1, 2, φi is clearly continuous on the open set
{f + g > 0}; besides if x0 ∈ X is such that (f + g)(x0 ) = 0 and
 > 0 is given, there exists an open neighborhood V of x0 on which
f + g < , hence kφk <  on V , whence kφi k <  on V , thus proving
the continuity of φi at x0 . Hence φi is continuous and {kφi k > 0} ⊂
{f + g > 0}; taking closures we conclude that spt φi ⊂ spt (f + g) b
X. Then φ1 , φ2 ∈ Cc (X, Rn ), kφ1 k ≤ f , kφ2 k ≤ g and φ1 + φ2 = φ,
so that µ · φ = µ · φ1 + µ · φ2 ≤ L(f ) + L(g). Taking the sup over all
such φ, we conclude that L(f + g) ≤ L(f ) + L(g) and our contention
is proved.
We now extend L to a positive linear functional on Cc (X, R). For
f ∈ Cc (X, R), we may write f = f + − f − , where f + = max{f, 0} ∈
C+ (X) and f − = max{−f, 0} ∈ C+ (X); define L · f := L(f + ) −
L(f − ) ∈ R (which coincides with the original definition if f = f + ∈
C+c (X)). If c ∈ R and f ∈ Cc (X, R), the fact that L · (cf ) = cL · f
follows from the definition of L and from the equalities (cf )+ = cf +
and (cf )− = cf − if c ≥ 0; (cf )+ = −cf − and (cf )− = −cf + if
c < 0. On the other hand, if f, g ∈ Cc (X, R) and h = f + g, then
h+ + f − + g − = h− + f + + g + ; applying L to both members and
using the additivity of L on C+ c (X), it follows that L·h = L·f +L·g,
thus proving the linearity of L. If f ∈ Cc (X, R) and f ≥ 0, then
f = f + ∈ C+ +
c (X), so that L·f = L(f ) ≥ 0; therefore L is a positive
linear functional on Cc (X, R).
We may then apply theorem 4.7 to L, which ensures the existence
of a unique Radon measure η on X which represents L. Thus, for
86 4. RN -VALUED RADON MEASURES

every f ∈ C+
c (X), we have
ˆ
n
(4.3) L · f = sup{µ · φ | φ ∈ Cc (X, R ), kφk ≤ f } = f dη.

For 1 ≤ i ≤ n, define µi : Cc (X, R) → R by µi · f := µ · (f ei ).

Since k±f ei k = |f | ∈ C+ c (X), it follows from (4.3) that, for all
f´ ∈ Cc (X, R), µi · (±f ) ≤ sup{µ
´ · φ | φ ∈ Cc (X, Rn ), kφk ≤ |f |} =
|f | dη, so that |µi · f | ≤ |f | dη. Thus, since Cc (X, R) is dense
on L1 (η) (by proposition 1.78; we consider Lp spaces of real valued
functions), µi extends to a bounded linear function on L1 (η), still de-
noted by µi . As η is σ-finite (because η is a Radon measure on X and
X is σ-compact), we may apply Riesz representation theorem 1.79
for the dual of L1 to conclude that there exists gi ∈ L∞ (η) which rep-
resents µ´i , i.e. such that ∀f ∈ L1 (η) (in particular, ∀f ∈ Cc (X, R)),
µi · f = f gi dη. It follows from corollary 1.118 that gi coincides
η-a.e. with a Borelian function; since this Borelian function is es-
sentially bounded, it may be modified in η-null Borel set, yielding
a bounded Borelian function in the same L∞ equivalence class. We
may therefore assume that gi is a bounded Borelian function.
For all f = ni=1 fi ei ∈ Cc (X, Rn ),
P
n
X n
X
µ·f = µ · (fi ei ) = µi · f i =
i=1 i=1
Xn ˆ ˆ
= fi gi dη = hf, gi dη,
i=1

where g = (g1 , . . . , gn ) : X → Rn is a bounded Borelian map. To

complete the proof of the existence part of the theorem, we now take
g
ν := kgk in the Borel set where g 6= 0 and 0 on its complement, and
λ := kgkη, i.e. the extension given ´ by theorem 1.8 of the measure
on B X defined by A ∈ B X 7→ A kgk dη. Then ν : X → Rn is
Borelian, by proposition 1.50 with A1 := {g 6= 0} and A2 := Ac1 ,
and kνk = 1´ λ-a.e., since λ(A2 ) = 0. The fact that the measure
A ∈ B X 7→ A kgk dη is a Radon measure on B X is a consequence
of lemma 4.11, below, with kgk in place of f and η in place of µ.
It then follows from remark 1.29.(ii) that ´ λ is a Radon
´ measure on
n
X. Since, for all f ∈ Cc (X, R ), µ · f = hf, gi dη = hf, νi dλ, i.e.
(4.1) holds, so the existence part is proved.
2) (Uniqueness and proof of (4.2)) Suppose that (4.1) holds with a
Radon measure λ and a Borel measurable map ν : X → Rn with
kνk = 1 λ-almost everywhere. Modifying ν on a λ-null Borel set,
if necessary, we may assume that equality holds everywhere, i.e.
 4.1. LINEAR FUNCTIONALS ON SPACES OF CONTINUOUS FUNCTIONS 87

∀x ∈ X, kν(x)k = 1. Given U ⊂ X open, denote by |µ|(U ) the

second member of (4.2), i.e.
|µ|(U ) := sup{µ · f | f ∈ Cc (X, Rn ), kf k ≺ U }.
n
´ spt f ⊂U
For
´ all f ∈ Cc (X,
´ R ) such that kf k ≺ U , we have µ·f = hf, νi dλ =
U
hf, νi dλ ≤ U kf k dλ ≤ λ(U ); hence, taking the sup in the first
member, we conclude that
(4.4) |µ|(U ) ≤ λ(U ).
We now prove the reverse inequality (hence the equality) in (4.4),
which then implies (4.2). Firstly, assume that λ(U ) < ∞. Fix
 > 0. We may apply Lusin’s theorem 1.117 to obtain a compact
set K ⊂ U such that λ(U \ K) <  and ν|K continuous. Then we
may apply Tietze’s extension theorem 4.6 to each component of ν|K ,
yielding f : X → Rn continuous such that f |K = ν|K . Multiplying
f by a convenient cut function, we may assume spt f ⊂ U and
kf ku ≤ 1 + . Indeed, since kνk ≡ 1, by continuity of f we may
take an open neighborhood V ⊂ U of K such that kf |V ku ≤ 1 + ,
and then we take φ ∈ Cc (X, R) given by Urysohn’s lemma 4.5 such
that χK ≤ φ ≺ V , so that φf ∈ Cc (X, Rn ), spt φf ⊂ V ⊂ U and
kφf ku ≤ 1 + ; we then substitute f φ for f . It therefore follows
that:´ ´ ´ ´
i) hf, νi dλ = U hf, νi dλ = U \K hf, νi dλ + K hf, νi dλ. Since
´ ´ ´
hf, νi dλ = hν, νi dλ = λ(K) > λ(U )− and | hf, νi dλ| ≤
´K K ´
U \K

U \K
kf k dλ ≤ kf ku λ(U \K) ≤ (1+), it follows that hf, νi dλ ≥
−(1 + ) + λ(U ) − ´= λ(U ) − (2 + ).
ii) On the other hand, hf, νi dλ = µ · f = kf ku µ · kffku ≤ (1 +
)|µ|(U ), since kf ku ≤ 1 +  and kffku is one of the competitors
in the definition of |µ|(U ), i.e. kffku ∈ Cc (X, Rn ) and kfkfkku ≺ U .
From (i) and (ii) above we conclude that λ(U ) ≤ (2 + ) + (1 +
)|µ|(U ). Since  > 0 was arbitrarily taken, we may send  → 0 to
conclude that λ(U ) ≤ |µ|(U ), thus proving the reverse inequality
(hence the equality) in (4.4) if λ(U ) < ∞.
In the general case, for an arbitrary open set U ⊂ X, due to
σ-compactness of X, we may take an increasing sequence (Un )n∈N
of open subsets of X such that ∀n ∈ N, Un b U and ∪n∈N Un = U
(to obtain such a sequence, take a sequence (Kn )n∈N of compact sets
which increases to U , then for each n ∈ N take an open set Vn such
that Kn ⊂ Vn b U and define Un := ∪ni=1 Vi ). Since, for each n ∈ N,
λ(Un ) < ∞ (because Un is compact and λ is Radon, hence finite on
88 4. RN -VALUED RADON MEASURES

compact sets), we may apply the case already proved to conclude

that λ(Un ) = |µ|(Un ). Applying the continuity from below 1.11 to
λ, it then follows that λ(U ) = supn∈N |µ|(Un ) = supn∈N sup{µ · f |
f ∈ Cc (X, Rn ), kf k ≺ Un }. We contend that the second member in
the latter equality is |µ|(U ) = sup{µ · f | f ∈ Cc (X, Rn ), kf k ≺ U },
which yields the asserted equality in (4.4). Indeed, for each n ∈ N,
|µ|(Un ) ≤ |µ|(U ), thus supn∈N |µ|(Un ) ≤ |µ|(U ). On the other
hand, let f be one of the competitors in the definition of |µ|(U ),
i.e. f ∈ Cc (X, Rn ) and kf k ≺ U . Since spt f ⊂ U is compact, it
can be covered by finitely many of the Un ’s; thus, since (Un )n∈N is
increasing, there exists n ∈ N such that spt f ⊂ Un . It therefore fol-
lows that kf k ≺ Un , i.e. f is one of the competitors in the definition
of |µ|(Un ), hence µ · f ≤ |µ|(Un ) ≤ supn∈N |µ|(Un ). Taking the sup
of µ · f over all such f , we conclude that |µ|(U ) ≤ supn∈N |µ|(Un ),
hence the equality holds, thus proving our contention.
We have thus proved that (4.2) holds, so that λ is uniquely de-
termined on open sets by |µ|. Since Radon measures are uniquely
determined by their values on open sets, we have proved the unique-
ness of λ.
We now prove the uniqueness of ν. Suppose that ν 0 : X → Rn
is another Borelian map such that kν 0 k = 1 λ-a.e. and (4.1) holds
with ν 0 in place of ν. Modifying both ν and ν 0 on a λ-null Borel
set, we may assume´ that kνk ≡ 1 and kν 0 k ≡ 1. Then, for all
f ∈ Cc (X, R ), hf, ν −ν 0 i dλ = 0. Let U ⊂ X open with λ(U ) < ∞
n

and fix  > 0. Once again we apply Lusin’s theorem 1.117 to obtain
a compact set K ⊂ U such that λ(U \ K) <  and (ν − ν 0 )|K
continuous. Then we may apply Tietze’s extension theorem 4.6 to
each component of (ν − ν 0 )|K , yielding f : X → Rn continuous such
that f |K = (ν − ν 0 )|K . Since kν − ν 0 ku ≤ 2, as before, multiplying f
by a convenient cut function if necessary, we may assume spt f ⊂ U
and kf ku ≤ 2 + . Thus
ˆ ˆ
0 = hf, ν − ν i dλ = hf, ν − ν 0 i dλ =
0

ˆ U
ˆ
= hf, ν − ν i dλ + kν − ν 0 k2 dλ.
0
U \K K
´
Since | U \K hf, ν − ν 0 i dλ| ≤ kf ku kν − ν 0 ku λ(U \ K) ≤ (2 + ) · 2
´ ´
and
´ K
kν − ν 0 k2 dλ ≥ U kν − ν 0 k2 dλ − 4, it then follows that
U
kν ´− ν 0 k2 dλ ≤ 4 + (2 + ) · 2. Hence, sending  → 0 we conclude
that U kν − ν 0 k2 dλ = 0, which implies ν = ν 0 λ-a.e. on U . As
X is σ-compact, we may cover X with countably many relatively
 4.1. LINEAR FUNCTIONALS ON SPACES OF CONTINUOUS FUNCTIONS 89

compact open sets (Un )n∈N , which are λ-finite, since λ is Radon
(that is, X is open σ-finite). Therefore, as ∀n ∈ N, ν = ν 0 λ-a.e. on
Un , it follows that ν = ν 0 λ-a.e. on ∪n∈N Un = X, thus proving that
ν is unique up to λ-null sets, as asserted.
It remains to prove assertion ii). Indeed, by (4.2), the norm

of µ
as a linear function on the normed space Cc (X, Rn ), k·ku is given
by kµk = sup{µ · f | f ∈ Cc (X, Rn ), kf ku ≤ 1} = λ(X), so that
µ is a bounded linear functional iff λ(X) <
∞. If that is the case,
as Cc (X, Rn ) is dense on C0 (X, Rn ), k·ku , µ extends to a unique
bounded linear functional on C0 (X, Rn ) with kµkC0 (X,Rn )∗ = λ(X),
thus proving ii).

Lemma 4.11. Let X be a locally compact Hausdorff space, f : X →
[0, ∞) bounded Borelian and µ a σ-finite Radon measure on B X (in
the sense
´ of remark 1.29.ii). Then λ := f µ : B X → [0, ∞] given by
A 7→ A f dµ is a Radon measure on B X .
Proof. It is clear that λ = f µ is a measure on B X which is finite
on compact subsets of X, since f is bounded and µ is Radon (hence
finite on such subsets). We must show that λ is outer regular on Borel
sets and inner regular on open subsets of X (actually it is inner regular
on all Borel sets). That is a consequence of the σ-compactness of µ
and of the fact that λ  µ:
1) Let B ∈ B X . Note that, if µ is finite on B, so is λ = f µ, since f is
bounded.
Assume that µ finite on B. For each n ∈ N, since µ(B) =
inf{µ(U ) | U ⊃ B open} and µ(B) < ∞, we may take an open set
Un ⊃ B such that µ(Un \ B) < n1 . Substituting Un with ∩ni=1 Ui , we
may assume that the sequence (Un )n∈N
thus defined is decreasing.
Then ∩n∈N Un ⊃ B and µ (∩n∈N Un ) \ B = 0; since λ  µ, it then
follows that λ is null on (∩n∈N Un ) \ B, so that λ(B) = λ(∩n∈N Un ).
On the other hand, since µ(U1 ) < ∞ (because µ is finite both on
B and U1 \ B), as noted above we also have λ(U1 ) < ∞; apply-
ing the continuity from above 1.11 to λ, we then conclude that
inf n∈N λ(Un ) = lim λ(Un ) = λ(∩n∈N Un ) = λ(B), thus proving the
outer regularity of λ on B if µ(B) < ∞.
In the general case, using the fact that µ is σ-finite, we may
write B = ∪n∈N Bn as a countable union of Borel sets with finite
µ-measure (hence with finite λ-measure). Given  > 0, for each
n ∈ N, we may choose, in view of the fact that λ(Bn ) < ∞ and
that λ is outer regular on Bn by the case proved above, an open set
90 4. RN -VALUED RADON MEASURES

Un ⊃ Bn such that λ(Un \ Bn ) < 2−n . Put U := ∪n∈N Un , so that

U ⊃ B open. As U \ B ⊂ ∪n∈N (Un \ Bn ), it follows by countable
subadditivity that λ(U \ B) < , thus proving the exterior regularity
of λ on B.
2) Let B ∈ B X . We will show that λ is inner regular on B.
Assume that µ(B) < ∞. Since µ is Radon, it follows from exer-
cise 1.31 that µ is inner regular on B; as µ(B) < ∞, for each n ∈ N,
there exists Kn ⊂ B compact such that µ(B \ Kn ) < n1 . Substitut-
ing Kn with ∪ni=1 Ki , we may assume that (Kn )n∈N thus defined is
increasing. Then ∪n∈N Kn ⊂ B and µ(B \ ∪n∈N Kn ) = 0; as λ  µ,
it then follows λ(B \ ∪n∈N Kn ) = 0. Thus, applying the continu-
ity from below 1.11 to λ, we conclude that λ(B) = λ(∪n∈N Kn ) =
lim λ(Kn ) = supn∈N λ(Kn ), which proves the interior regularity of λ
on B if µ(B) < ∞.
In the general case, by the fact that µ is σ-finite, we may take
an increasing sequence (Bn )n∈N in B X such that ∪n∈N Bn = B and
∀n ∈ N, µ(Bn ) < ∞ (thus λ(Bn ) < ∞). By the case proved above,
for each n ∈ N, λ is inner regular on Bn ; hence, since λ(Bn ) < ∞, we
may take a compact Kn ⊂ Bn such that λ(Bn \ Kn ) < n1 . Therefore,
lim λ(Kn ) = lim λ(Bn ) = λ(B), where in the last equality we have
applied the continuity from below to λ, showing that λ is inner
regular on B, as asserted.

In theorem 4.9, we may drop the σ-compactness hypothesis on X if
µ is a finite Rn -valued Radon measure. That is, we obtain the following
version of the theorem:
Theorem 4.12 (Riesz representation theorem for finite Radon mea-
sures). Let X be a locally compact Hausdorff space and µ : Cc (X, Rn ) →
R a finite Rn -valued Radon measure on X. Then there exists a unique
finite Radon measure λ on X and a Borel measurable map ν : X → Rn
unique up to λ-null sets such that kνk = 1 λ-a.e. on X and ∀f ∈
Cc (X, Rn ), ˆ
µ·f = hf, νi dλ,
where h·, ·i denotes the Euclidean inner product in Rn . Moreover,
i) ∀U ⊂ X open,
λ(U ) = sup{µ · f | f ∈ Cc (X, Rn ), kf k ≺ U }.
ii) kµkC0 (X,Rn )∗ = λ(X).
The proof is the same, as:
 4.1. LINEAR FUNCTIONALS ON SPACES OF CONTINUOUS FUNCTIONS 91

i) In the existence part, we only used the σ-compactness condition

to ensure that the restriction of λ to B X is σ-finite (in order to
be able to apply Riesz representation theorem 1.79 for the dual of
L1 (λ) and lemma 4.11), but λ is finite in case of µ finite (since, if
µ is finite, the positive linear functional L defined in the beginning
of the proof is bounded, hence λ(X) = kLk < ∞ by the formula
to compute the measure which represents the linear functional on
open sets given in theorem 4.7).
ii) In the uniqueness part and in the proof of (4.2), we used the σ-
compactness condition only in case λ(U ) = ∞; but, as pointed in
the previous item, λ is finite in case of µ finite, so that we don’t
need the σ-compactness either.
Definition 4.13 (total variation and polar decomposition). Let µ
be an Rn -valued Radon measure on a σ-compact locally compact Haus-
dorff space X. With the same notation of theorem 4.9, λ is called the
total variation of µ, and the pair (ν, λ) is called the polar decomposition
of µ. Henceforth, we will use the notation |µ|:= λ to denote the total
variation of µ, and
µ = ν|µ|
with the meaning that (ν, |µ|) is the polar decomposition of µ.
Example 4.14. Let X be a σ-compact locally compact Hausdorff
space.
1) Let µ be a locally finite Borel measure on X. Then µ induces a
positive linear functional µ̂ on Cc (X, R) (which
´ is necessarily con-
tinuous, by remark 4.8), given by µ̂ · f := f dµ. If µ is a Radon
measure, then µ̂ = 1 · µ is the polar decomposition of µ̂ (by the
uniqueness of the polar decomposition); in particular, µ is the total
variation of µ̂.
2) Similarly, let ν be a signed measure on B X whose total variation
|ν| is locally finite. Then ν induces
´ a continuous linear functional
ν̂ on Cc (X, R) given by ν̂ · f := f dν. Indeed, it is clear that ν̂
is a well-defined linear functional on Cc (X, R), and the continuity
follows from the triangle
K
´ inequality´ 1.99.e): ∀K ⊂ X compact and
∀f ∈ Cc (X), |ν̂ · f | ≤ |f | d|ν| = K |f | d|ν| ≤ |ν|(K)kf ku , hence ν̂
is bounded on CKc (X).
We may take Borelian function h : X → R such that |h| ≡ 1 and
ν = h|ν|. Indeed, since ν + ⊥ ν − , we may take disjoint Borel sets
P and N such that X = P ∪˙ N , ν + is concentrated on P and ν − is
concentrated on N . Thus, ν + = χP |ν| and ν − = χN |ν|, so that ν =
ν + − ν − = (χP − χN )|ν| and we take h := χP − χN . If |ν| is Radon,
92 4. RN -VALUED RADON MEASURES

it follows from the uniqueness of the polar decomposition of ν̂ that

its polar decomposition is ν̂ = h|ν| (we identify |ν| with a Radon
outer measure on X, cf. remark 1.29). In particular, |ν̂| = |ν|.
Besides, it follows from the uniqueness of the Jordan decomposition
1.94 that, as measures on B X , ν + = h+ |ν| and ν − = h− |ν|. Since
either ν + or ν − if finite, we conclude that either h+ ∈ L1 (|ν|) or
h− ∈ L1 (|ν|), i.e. h is |ν|-integrable. Thus, in order for a continuous
linear functional µ on Cc (X, R) to be induced by a signed measure
on B X whose total variation is Radon, it is necessary that µ have
polar decomposition µ = h|µ| with h |µ|-integrable (which means
that not every continuous linear functional on Cc (X, R) is obtained
in this way if X is not compact).
3) Let X = R and I be the positive linear functional defined on
´b
Cc (X, R) by the Riemann integral, i.e. I · f := a f (x) dx for a < b
such that spt f ⊂ [a, b]. The polar decomposition of I is I = 1 · Ln .
In particular, that could have been taken as the definition of the
Lebesgue measure, i.e. it is the total variation of the positive linear
functional induced by the Riemann integral.

Proposition 4.15 (properties of the total variation, part I). Let

µ and ν be Rn -valued Radon measures on a σ-compact locally compact
Hausdorff space X and c ∈ R. Then:
i) |µ + ν| ≤ |µ| + |ν|, with equality if |µ| ⊥ |ν|.
ii) |cµ| = |c||µ|.

That is, the total variation of Rn -valued Radon measures has the
same properties stated in 1.100.b) for the total variation of signed mea-
sures on a σ-algebra of subsets of X.

Proof. Let U ⊂ X open. It follows from (4.2) and remark 4.10

that:
1) |µ + ν|(U ) = sup{µ · f + ν · f | f ∈ Cc (X, Rn ), kf k ≤ 1} ≤ sup{µ · f |
f ∈ Cc (X, Rn ), kf k ≤ 1} + sup{ν · f | f ∈ Cc (X, Rn ), kf k ≤ 1} =
|µ|(U ) + |ν|(U ).
2) |cµ|(U ) = sup{|c||µ · f | | f ∈ Cc (X, Rn ), kf k ≤ 1} = |c| sup{|µ · f | |
f ∈ Cc (X, Rn ), kf k ≤ 1} = |c||µ|(U ).
For an arbitrary set A ⊂ X, we now use the outer regularity on A of
the Radon measures |µ + ν|, |µ| and |ν|. Note that, for arbitrary open
sets U, V containing A, the open set U ∩ V contains A and |µ|(U ) +
 4.1. LINEAR FUNCTIONALS ON SPACES OF CONTINUOUS FUNCTIONS 93

|ν|(V ) ≥ |µ|(U ∩ V ) + |ν|(U ∩ V ), which justifies the equality (∗) below:

by 1)
|µ + ν|(A) = inf{|µ + ν|(U ) | U ⊃ A open} ≤
∗
≤ inf{|µ|(U ) + |ν|(U ) | U ⊃ A open} =
= inf{|µ|(U ) + |ν|(V ) | U, V ⊃ A open} =
= inf{|µ|(U ) | U ⊃ A open} + inf{|ν|(V ) | V ⊃ A open} =
= |µ|(A) + |ν|(A),
which proves the inequality in part i).
Similarly,
by 2)
|cµ|(A) = inf{|cµ|(U ) | U ⊃ A open} =
= |c| inf{|µ|(U ) | U ⊃ A open} = |c||µ|(A),
thus proving part ii).
It remains to prove the equality in part i) if |µ| ⊥ |ν|. Indeed, in that
case, there exist disjoint Borel sets A, B ⊂ X such that X = A ∪˙ B, |µ|
concentrated on A and |ν| concentrated on B. Let (nµ , |µ|) and (nν , |ν|)
be the polar decompositions of µ and ν, respectively. We have, for all
f ∈ Cc (X, Rn ):
ˆ ˆ
(µ + ν) · f = µ · f + ν · f = hf, nµ i d|µ| + hf, nν i d|ν| =
ˆ
= hf, χA nµ + χB nν i d(|µ| + |ν|).

Since kχA nµ + χB nν k = 1 (|µ| + |ν|)-a.e., it follows that the polar

decomposition of µ+ν is (χA nµ +χB nν , |µ|+|ν|); in particular, |µ+ν| =
|µ| + |ν|, as asserted. 
Definition 4.16 (integration with respect to Rn -valued Radon
measures). Let µ be an Rn -valued Radon measure on a σ-compact lo-
cally compact Hausdorff space X, with polar decomposition µ = ν|µ|.
i) A vector Borelian map f : X → Rn is called summable with respect
to µ if it is summable with respect to |µ|, i.e. if f ∈ L1 (|µ|, Rn ) ≡
L1 (|µ|)n . For such f , we define
ˆ ˆ
f · dµ := hf, νi d|µ| ∈ R.

ii) An scalar Borelian map f : X → R is called summable with respect

to µ if it is summable with respect to |µ|, i.e. if f ∈ L1 (|µ|). For
94 4. RN -VALUED RADON MEASURES

such f , we define
ˆ ˆ ˆ ˆ
f ν1 d|µ|, . . . , f νn d|µ| ∈ Rn .


f dµ := f ν d|µ| =

Remark 4.17.
1) Note that both integrals in the definition above make sense, since
|hf, νi| ≤ kf k ∈ L1 (|µ|) if f vector-valued and kf νk = |f | ∈ L1 (|µ|)
if f scalar-valued.
2) Since |µ| is a Radon measure, we have Cc (X, Rn ) ⊂ L1 (|µ|, Rn );
the inclusion is actually dense, in view of proposition 1.78. It is
clear that the integral defined in i) extends µ : Cc (X, Rn ) → R, i.e.
∀f ∈ Cc (X, Rn ), ˆ
f · dµ = µ · f.
3) The integrals defined above satisfy the usual linearity and conver-
gence properties, which are inherited from the corresponding prop-
erties for the integral with respect to |µ|. So are following versions
of the triangle inequality:
ˆ ˆ ˆ ˆ
| f · dµ| ≤ kf k d|µ| and k f dµk ≤ |f | d|µ|,

for f ∈ L1 (|µ|, Rn ) or f ∈ L1 (|µ|), respectively.

We now aim to identify Rn -valued Radon measures with set func-
tions. Firstly we introduce the notion of Rn -valued measures as set
functions.
Definition 4.18 (Rn -valued measure on a σ-algebra). Let X be a
set and M a σ-algebra of subsets of X. We say that a map µ : M → Rn
is an Rn -valued measure on M if
VM1) µ(∅) = 0;
VM2) µ is σ-additive, i.e. for all countable disjoint family (An )n∈N in
M, X
µ(∪n∈N An ) = µ(An ),
n∈N
with the meaning that the series is absolutely convergent (or,
equivalently, that each component of n 7→ µ(An ) is summable
with respect to the counting measure on N) and the sum is
µ(∪n∈N An ).
Definition 4.19 (Rn -valued Radon measures as set functions). Let
X be a σ-compact locally compact Hausdorff space. We denote by B cX
the set of Borel subsets of X which are relatively compact. We define:
 4.1. LINEAR FUNCTIONALS ON SPACES OF CONTINUOUS FUNCTIONS 95

i) a finite Rn -valued Radon measure set function on X is an Rn -

valued measure on B X in the sense of definition 4.18.
ii) an Rn -valued Radon measure set function on X is a set function
µ : B cX → Rn such that, for all K ⊂ X compact, its restriction
to B K ⊂ B cX is a finite Rn -valued Radon measure set function on
K, i.e. µ|BK : B K → Rn is an Rn -valued measure on B K in the
sense of definition 4.18.
We denote by M(X)n or M(X, Rn ) the set of finite Rn -valued
Radon measure set functions on X and by Mloc (X)n or Mloc (X, Rn )
the set of Rn -valued Radon measure set functions on X. It is clear
that those are real linear spaces, i.e. M(X, Rn ) is a linear subspace of
(Rn )BX and Mloc (X, Rn ) is a linear subspace of (Rn )BX .
c

Remark 4.20.
1) The nomenclature established in the previous definition is provi-
sional. That is, for a moment we want to use different names for
Rn -valued Radon measures as linear functionals on spaces of contin-
uous functions and for Rn -valued Radon measures as set functions.
However, we will see shortly that, if X is second countable, i.e. if X
is a locally compact separable metrizable space (which is the case
of interest in subsequent developments), a (finite) Rn -valued Radon
measure set function on X may be canonically identified with a
(finite) Rn -valued Radon measure on X (the latter in the sense of
definition 4.1), and conversely. Making these identifications, we will
treat those objects as one and the same thing, so that we may aban-
don this provisional nomenclature.
2) Each µ ∈ M(X, Rn ) determines an element of Mloc (X, Rn ) by re-
striction of µ : B X → Rn to B cX . The fact that X is σ-compact
allows to decompose each B ∈ B X as a countable disjoint union of
elements of B cX ; thus, by σ-additivity, µ is uniquely determined by
its restriction to B cX , i.e. the association µ ∈ M(X, Rn ) 7→ µ|BcX ∈
Mloc (X, Rn ) is linear 1-1 and allows us to identify M(X, Rn ) with a
linear subspace of Mloc (X, Rn ). By means of this identification, we
consider, henceforth, M(X, Rn ) as a linear subspace of Mloc (X, Rn ).
Definition 4.21 (induced Rn -valued Radon measure set functions).
Let µ be an Rn -valued Radon measure on a σ-compact locally compact
Hausdorff space X. The Rn -valued Radon measure set function in-
duced by µ is the set function µ̂ : B cX → Rn defined, for all A ∈ B cX ,
by ˆ
µ̂(A) := χA dµ ∈ Rn .
If µ is finite, we define µ̂ : B X → Rn by the same formula.
96 4. RN -VALUED RADON MEASURES

Note that the definition above makes sense, since χA ∈ L1 (|µ|) if

A ∈ B cX or if A ∈ B X and µ finite (i.e. if |µ| is finite, by theorem 4.9).
The fact that µ̂ is actually an Rn -valued Radon measure set function
is proved in the proposition below.
Proposition 4.22 (induced Rn -valued Radon measure set func-
tions). With the notation from the previous definition:
i) µ̂ is a (finite) Rn -valued Radon measure set function on X if µ is
a (finite) Rn -valued Radon measure on X.
ii) The maps I : Cc (X, Rn )∗ → Mloc (X, Rn ) and I : C0 (X, Rn )∗ →
M(X, Rn ) defined by µ 7→ µ̂ are linear 1-1 and commute with the
inclusions, i.e. the following diagram is commutative:

Cc (X, Rn )∗ I
Mloc (X, Rn )

C0 (X, Rn )∗ I
M(X, Rn )

Proof. Let K ⊂ X compact. We assert that µ̂ : B K → Rn

is an Rn -valued measure on B K (in the sense of definition 4.18). It
is clear that µ̂(∅) = 0. Let (An )n∈N be a disjoint sequence of Borel
˙
Pn subsets of K, and A = ∪n∈N An . For each n ∈ N, let
measurable
φn := k=1 χAk . Then φn ν → χA ν pointwise, and the convergence is
dominated, since kφn νk ≤ χA ∈ L1 (|µ|) (because A b X and |µ| is finite
on compact sets). Applying the´dominated convergence
´ theorem 1.64
componentwise,
´ Pn ´it follows that φ
Pnnν d|µ| → χ A ν d|µ| = µ̂(A). As
φn ν d|µ| = k=1 χAk ν d|µ| = k=1 µ̂(Ak ), the assertion is proved.
Hence, µ̂ is an Rn -valued Radon measure set function on X if µ is an
Rn -valued Radon measure on X. The same argument may be used to
prove that µ̂ : B X → Rn is an Rn -valued measure on B X (i.e. a finite
Rn -valued Radon measure set function on X) if µ is finite.
We next prove that I is linear. Let µ and ν be Rn -valued Radon
measures on X and c ∈ R.
To prove that I(µ + ν) = I(µ) + I(ν), we must compare all three of
the polar decompositions µ = n1 |µ|, ν = n2 |ν| and µ+ν = N |µ+ν|. In
order to accomplish that, note that λ := |µ|+|ν| is a Radon measure on
X, by exercise 1.30, and all three of the measures |µ+ν|, |µ| and |ν| are
absolutely continuous with respect to λ (recall that |µ + ν| ≤ |µ| + |ν|,
from proposition 4.15). Since the restrictions to B X of all measures
involved are σ-finite (because they are Radon and X is σ-compact),
we may take Radon-Nikodym derivatives of those restrictions (theorem
1.103) and apply the chain rule for such derivatives (proposition 1.107),
 4.1. LINEAR FUNCTIONALS ON SPACES OF CONTINUOUS FUNCTIONS 97

which yields, for all f ∈ Cc (X, Rn ):

ˆ ˆ
1.107 d|µ|
µ·f = f · n1 d|µ| = f · n1
dλ,
dλ
ˆ ˆ
1.107 d|ν|
ν · f = f · n2 d|ν| = f · n2 dλ,
dλ

hence

ˆ
d|µ| d|ν|
(µ + ν) · f = hf, n1 + n2 i dλ =
dλ dλ
ˆ D
n1 d|µ|
dλ
+ n2 d|ν|
dλ
E d|µ| d|ν|
= f, d|µ| d|ν|
n1 + n2 dλ
n1 dλ + n2 dλ dλ dλ

(we define the quotient to be, for instance, 0 where the denominator
is 0). Note that, since |µ| ≤ λ and |ν| ≤ λ, we have d|µ|
dλ
≤ 1 λ-a.e.
d|ν| d|µ| d|ν|
and dλ ≤ 1 λ-a.e., whence n1 dλ + n2 dλ ≤ 2 λ-a.e.; modifying
those Borelian functions, if necessary, on a λ-null Borel set, we may
assume that they are all bounded. Thus, from lemma 4.11, d|µ|
dλ
λ, d|ν|
dλ
λ
and n1 dλ + n2 dλ λ are Radon measures on B X ; therefore, from
d|µ| d|ν|

remark 1.29, the extensions (denoted with the same notation) of those
measures given by theorem 1.8 are outer Radon measures on X. It
then follows from the uniqueness of the polar decompositions of µ, ν
and µ + ν that

d|µ|
|µ| = λ,
dλ
d|ν|
|ν| = λ,
dλ
d|µ| d|ν|
|µ + ν| = n1 + n2 λ,
dλ dλ
n1 d|µ|
dλ
+ n2 d|ν|
dλ
N= d|µ| d|ν|
|µ + ν| − a.e.
n1 dλ + n2 dλ
98 4. RN -VALUED RADON MEASURES

We thus have, for all A ∈ B cX :

ˆ
µ
\ + ν(A) = N d|µ + ν| =
A
ˆ
n1 d|µ|
dλ
+ n2 d|ν|
dλ d|µ| d|ν|
= d|µ|
n1 + n2 dλ =
A n1 dλ + n2 d|ν| dλ dλ
ˆ dλ
d|µ| d|ν|

= n1 + n2 dλ =
dλ dλ
ˆA
ˆ
d|µ| d|ν|
= n1 dλ + n2 dλ =
dλ dλ
ˆA
ˆ A

= n1 d|µ| + n2 d|ν| = µ̂(A) + ν̂(A)

A A

thus showing that I(µ + ν) = I(µ) + I(ν). Similarly, if c 6= 0 and

µ = n1 |µ| is the polar decomposition
´ of ´µ, it follows that, for all
n
f ∈ Cc (X, R ), (cµ) · f = c hf, n1 i d|µ| = hf, sgn (c)n1 i|c||µ|. Thus,
if cµ = n2 |cν| is the polar decomposition of cµ, it follows from the
uniqueness of such decomposition that n2 = sgn (c)n1 and |cν| = |c||ν|
(which had already been proved in proposition 4.15.(ii)). Therefore,
for all A ∈ B cX ,
ˆ ˆ
ccµ(A) = n2 d|cµ| = sgn (c)n1 d(|c||µ|) =
ˆ A A

= sgn (c)|c|n1 d|µ| = cµ̂(A),

A

thus proving that I(cµ) = cI(µ).

The commutativity of the diagram in part ii) is immediate from the
definitions. It remains to prove that I : Cc (X, Rn )∗ → Mloc (X, Rn ) is
1-1 (which then implies that I : C0 (X, Rn )∗ → M(X, Rn ) is 1-1 in view
of the asserted commutativity).
Let µ and ν be Rn -valued Radon measures such that µ̂ = ν̂, with
respective polar decompositions
´ ´µ = n1 |µ| and µ = n2 |ν|. We have,
for all A ∈ B X , χA n1 d|µ| = χA n2 d|ν|. By linearity of the inte-
c

grals, the latter equality

Pk also holds if we substitute χA with a sim-
ple function φ = i=1 ai χAi such that ∀1 ≤ i ≤ k, ai ∈ R and
Ai ∈ B cX if ai 6= 0. For any f ∈ Cc (X, R), we may take a se-
quence (φm )m∈N of such simple functions which converges pointwise
to f and ∀m ∈ N, |φm | ≤ |f |; that follows from proposition 1.53.iii)
(note that, writing
Pn a simple function φ in the standard representa-
tion, i.e. φ = i=1 ai χφ−1 (ai ) for Im φ = {a1 , . . . , an } with ai 6= aj
if i 6= j, then |φ| ≤ |f | implies φ−1 (ai ) ⊂ spt f if ai 6= 0, hence
 4.1. LINEAR FUNCTIONALS ON SPACES OF CONTINUOUS FUNCTIONS 99

φ−1 (ai ) ∈ B cX if ai 6= 0). Hence φm n1 → f n1 and φm n2 → f n2

pointwise, and the convergence is dominated with respect to both |µ|
and |ν|, since kφm ni k = |φm | ≤ |f | ∈ L1 (|µ|) ∩ L1 (|ν|). Applying the
dominated
´ convergence ´ theorem 1.64 componentwise,
´ it follows
´ that
f n1 d|µ| = limm→∞ φm n1 d|µ| = limm→∞ φm n2 d|ν| = f n2 d|ν|.
Writing the latter equality componentwise, if ni = nj=1 nji ej , we con-
P
´ ´
clude that, for all f ∈ Cc (X, R) and 1 ≤ j ≤ n, f nj1 d|µ| = f nj2 d|ν|.
Therefore, for all f ∈ Cc (X, Rn ),
ˆ
µ · f = f · n1 d|µ| =
n ˆ
X
= ( fj nj1 d|µ|)ej =
j=1
Xn ˆ
= ( fj nj2 d|ν|)ej =
j=1
ˆ
= f · n2 d|ν| = ν · f,

thus µ = ν, showing that I : Cc (X, Rn )∗ → Mloc (X, Rn ) is 1-1, as

asserted.

Conversely, every Rn -valued Radon measure set function on a σ-
compact locally compact Hausdorff space X induces an Rn -valued Radon
measure on X.
Definition 4.23 (induced Rn -valued Radon measures). Let µ be
an Rn -valued Radon measure set function on a σ-compact locally com-
pact Hausdorff space X. The Rn -valued Radon measure induced by µ
is the map µ̌ : Cc (X, Rn ) → R defined, for all f ∈ Cc (X, Rn ), by
n ˆ
X
µ̌ · f := fi dνi ,
i=1 K

where the integrals in the second member are taken with respect to
signed measures νi , 1 ≤ i ≤ n, obtained as the the restrictions of the
components µi of µ to B K , where K ∈ B cX contains spt f .
Note that, by definition 4.19, µ|BK = (ν1 , . . . , νn ) : B K → Rn is an
R -valued measure on B K ; thus, for 1 ≤ i ≤ n, νi is a finite signed
n

measure on the measure space (K, B K ) in the sense of definition 1.89.

We may then take the integrals in the sense of definition 1.97, since f
is bounded (hence f |K ∈ L1 (|νi |) for 1 ≤ i ≤ n).
100 4. RN -VALUED RADON MEASURES

That µ̌ · f is well-defined (i.e. the definition does not depend on the

relatively compact Borel set K ⊃ spt f ) and linear on f will be proved
as part of the next proposition.
Proposition 4.24. With the notation from the previous definition:
i) µ̌ : Cc (X, Rn ) → R is well-defined and linear continuous, i.e. it is
an Rn -valued Radon measure on X. Moreover, µ̌ is finite if so is
µ, i.e. µ̌ : C0 (X, Rn ) → R is linear continuous if µ is finite.
ii) The maps J : Mloc (X, Rn ) → Cc (X, Rn )∗ and J : M(X, Rn ) →
C0 (X, Rn )∗ defined by µ 7→ µ̌ are linear, commute with inclusions
and invert I (defined in proposition 4.22) on the left, i.e. ∀µ Rn -
valued Radon measure,
ˇ = µ.
µ̂
Proof. We firstly show that, given f ∈ Cc (X, Rn ), µ̌ · f is well-
defined, i.e. the definition does not depend on the relatively compact
Borel set K ⊃ spt f . Indeed, let, ∀1 ≤ i ≤ n, νi : B K → R and
νi0 : B spt f → R denote the ith components of the restrictions of µ
to B K and B spt f , respectively (which are finite signed measures). It
follows from the uniqueness of the Jordan decomposition 1.94 that
the positive and negative parts of νi0 coincide with the restrictions of
the´positive and´ negative parts of νi ,´ respectively; thus, for 1 ≤ i ≤
n, K f dνi± = spt f f d(νi± )|spt f = spt f f d(νi0 )± , showing that the
definition of µ̌ · f does not depend on K, as asserted.
To show that µ̌ : Cc (X, Rn ) → R is linear continuous, it suffices to
show that, for each K ⊂ X compact, the restriction µ̌ : CKc (X, Rn ) → R
is linear continuous in the normed space (CKc (X, Rn ), k·ku ) (see defini-
tion 4.1). That follows from Pnthe´ fact that, if µ|BK = (ν1 , . . . , νn ),
µ̌|CKc (X,Rn ) is given by f 7→ i=1 K fi dνi , which is clearly linear in
K n
Pn ´
CPc (X, R ) (by the linearity of the integrals) and |µ̌·f | ≤ i=1 K |fi | d|νi | ≤
( ni=1 |νi |(K))kf ku (where we used the triangle inequality from exercise
1.99.(e)), which yields the asserted continuity.
If µ is finite, i.e. if µ = (µ1 , . . . , µn ) : B X → Rn is an Rn -valued
measure on B X , each component µi of µ is a finite signed measure
on B X and the same argument used in the first paragraph of this
proof to show that µ̌ · f P is well-defined
´ may be applied to conclude that
∀f ∈ Cc (X, Rn ), µ̌·f = ni=1 X fi dµi . Therefore, Pnapplying the triangle
inequality once more, it follows that
|µ̌ · f | ≤ ( i=1 |µi |(X))kf ku , thus
proving that µ̌ : Cc (X, Rn ), k·ku → R is linear continuous; hence, by
the density of Cc (X, Rn ) in C0 (X, Rn ), µ̌ may be uniquely extended
n
Pna continuous linear functional µ̌ : C0 (X, R ) → R with norm ≤
to
i=1 |µi |(X).
 4.1. LINEAR FUNCTIONALS ON SPACES OF CONTINUOUS FUNCTIONS 101

The fact that J : Mloc (X, Rn ) → Cc (X, Rn )∗ is linear follows

directly from its definition and from exercise 1.100 parts e) and f).
Since J : M(X, Rn ) → C0 (X, Rn )∗ was defined by restriction of J :
Mloc (X, Rn ) → Cc (X, Rn )∗ to M(X, Rn ), it is also linear and the com-
mutativity with the inclusions follows by definition.
It remains to show that J ◦I coincides with the identity of Cc (X, Rn )
(hence it also coincides with the identity of C0 (X, Rn ), thanks to the
commutativity of I and J with the inclusions).
Let µ ∈ Cc (X, Rn )∗ be an Rn -valued Radon measure with polar
decomposition µ = N |µ|, where N = (N1 , . . . , Nn ). We must show
ˇ = µ. It follows from definition 4.21 that, for each A ∈ B c ,
that µ̂ ´ X
µ̂(A) = A N d|µ|. Thus, for each K ⊂ X compact, for 1 ≤ i ≤ n,
the ith component of the Rn -valued measure µ̂|BK on B K is the finite
signed measure Ni |µ| on (K, B K ). The positive and negative parts of
its Jordan decomposition are Ni+ |µ|, Ni− |µ| : B K → R, respectively
(since Ni |µ| = Ni+ |µ| − Ni− |µ| and Ni+ |µ| ⊥ Ni− |µ| as measures on the
measurable space (K, B K )). It then follows from definition 4.23 that,
if f ∈ Cc (X, Rn ) and spt f ⊂ K,
n ˆ
X
ˇ
µ̂ · f = fi d(Ni |µ|) =
i=1 K
Xn ˆ n ˆ
X
= fi d(Ni+ |µ|) − fi d(Ni− |µ|) =
i=1 K i=1 K

Xn ˆ Xˆ
n
= fi Ni+ d|µ| − fi Ni− d|µ| =
i=1 K i=1 K

Xn ˆ
= fi Ni d|µ| =
K
ˆi=1
= hf, N i d|µ| = µ · f.
ˇ = µ,
Therefore, by the arbitrariness of K and f , we conclude that µ̂
as asserted. 
We will see next that, if X is a locally compact separable metric
space, J defined above also inverts I on the right, i.e. I◦J is the identity
of the corresponding domains. That is, I : Cc (X, Rn )∗ → Mloc (X, Rn )
and I : C0 (X, Rn )∗ → M(X, Rn ) are surjective isomorphisms, so that
we may identify Cc (X, Rn )∗ ≡ Mloc (X, Rn ) and C0 (X, Rn )∗ ≡ M(X, Rn ).
Proposition 4.25. Let X be a locally compact separable metric
space, I and J defined in propositions 4.22 and 4.24, respectively.
102 4. RN -VALUED RADON MEASURES

Then I ◦ J is the identity of Mloc (X, Rn ), and so is its restriction

to M(X, Rn ).
Proof. Take µ : B cX → Rn in Mloc (X, Rn ). We must show that
ˆ = µ. It suffices to show that, for an arbitrary relatively compact
µ̌
open set U ⊂ X, the restrictions of µ and µ̌ˆ coincide on B U ⊂ B cX (for
the union of such B U coincides with B cX , i.e. every element of B cX is
contained in some relatively compact open set). Fix such a relatively
ˆ|BU = (µ̌
compact open set U and let µ|BU = (µ1 , . . . , µn ), µ̌ ˆ1 , . . . , µ̌
ˆn ).
We must therefore prove that, for 1 ≤ i ≤ n, the finite signed measures
µi and µ̌ˆi coincide on the measurable space (U, B U ).
Fix 1 ≤ i ≤ n. Note that, since U is a locally compact separable
metric space, and since the total variations of both µi and µ̌ ˆi are finite
positive Borel measures on (U, B U ) (in particular, they are finite on
compact subsets of U ), it follows from remark 1.33 that |µi | and |µ̌ ˆi | are
positive Radon measures on B U . Thus, in order to prove that µi = µ̌ ˆi ,
it suffices to show, in´ view of lemma
´ 4.26 below with U in place of X,
that ∀f ∈ Cc (U, R), U f dµi = U f dµ̌ ˆi .
Fix f ∈ Cc (U, R) and let F := f ei ∈ Cc (U, Rn ) ⊂ Cc (X, Rn ). It
follows from definition 4.23 that ˆ
(4.5) µ̌ · F = f dµi .
U
On the other hand, let the polar decomposition of µ̌ ∈ Cc (X, Rn )∗
be µ̌ = N |µ̌|, where
´ N = (N1 , . . . , Nn ). By definition 4.21, for each
A ∈ B cX , µ̌
ˆ(A) = N d|µ̌|. Thus, the ith component of the Rn -valued
A
measure µ̌ˆ|BU on B U is the finite signed measure Ni |µ̌| on (U, B U ). It
then follows that ˆ ˆ
µ̌ · F = hF, N i d|µ̌| = f Ni d|µ̌| =
(4.6) ˆ U

= ˆi .
f dµ̌
U
´ ´
Comparing (4.5) and (4.6), we conclude that f dµi = ˆi ,
f dµ̌
U U
as we wanted to show.

Lemma 4.26. Let X be a σ-compact locally compact Hausdorff space
and µ,ν signed measures on B X whose total variations |µ| and
´ |ν| are
Radon
´ measures on B X . Then µ = ν iff ∀f ∈ Cc (X, R), f dµ =
f dν.
Note that both integrals make sense, since both |µ| and |ν| are
Radon, hence Cc (X, R) ⊂ L1 (µ) ∩ L1 (ν).
 4.1. LINEAR FUNCTIONALS ON SPACES OF CONTINUOUS FUNCTIONS 103

Proof. It suffices to prove (⇐), since the reverse implication is

trivial. ´ ´
Suppose that ∀f ∈ Cc (X, R), f dµ = f dν.
1) We contend that there exist Borelian functions hµ , hν : X → R
such that |hµ | = |hν | ≡ 1 and µ = hµ |µ|, ν = hν |ν|. Indeed, since
µ+ ⊥ µ− , we may take disjoint Borel sets P and N such that X =
P ∪˙ N , µ+ is concentrated on P and µ− is concentrated on N . Thus,
µ+ = χP |µ| and µ− = χN |µ|, so that µ = µ+ − µ− = (χP − χN )|µ|,
thus proving the contention for µ with hµ := χP − χN ; we do the
same for ν. ´ ´
2) The linear functional L defined on Cc (X, R) by f 7→ f dµ = f dν
is continuous, since, by the fact that |µ| is finite on compact sets
and by the triangle inequality ´1.99.e), for´ every K ⊂ X compact
and for every f ∈ CKc (X, R), | f dµ| ≤ K |f | d|µ| ≤ |µ|(K)kf ku .
By the previous item, we have, for all f ∈ Cc (X, R),
ˆ ˆ
L · f = f dµ = f · hµ d|µ|,
ˆ ˆ
L · f = f dν = f · hν d|ν|.

Since |µ| and |ν| are positive Radon measures on B X (which, by

remark 1.29, correspond to outer Radon measures on X, denoted
with the same notation), we conclude that both hµ |µ| and hν |ν| are
polar decompositions for L. Hence, by the uniqueness of the polar
decomposition stated in theorem 4.9 (with n = 1), it follows that
|µ| = |ν| and hµ = hν |µ|-a.e., which implies µ = hµ |µ| = hν |ν| = ν.

Corollary 4.27. If X is a locally compact separable metric space,
then I and J defined in propositions 4.22 and 4.24, respectively, are
surjective isomorphisms, inverses of each other.
Proof. It is a consequence of propositions 4.22, 4.24 and 4.25. 
Remark 4.28. If X is a locally compact separable metric space,
we may therefore identify Cc (X, Rn )∗ ≡ Mloc (X, Rn ) and C0 (X, Rn )∗ ≡
M(X, Rn ) by means of the linear isomorphisms of the previous corol-
lary. With these identifications in mind, we will henceforth abandon
our provisional nomenclature and drop the hats and checks from the
notation, treating an Rn -valued Radon measure µ and the correspond-
ing Rn -valued Radon measure set function µ̂ as being one and the same
object.
104 4. RN -VALUED RADON MEASURES

Corollary 4.29. If X is a locally compact separable metric space,

M(X, Rn ) is a Banach space with the norm kµk := |µ|(X).
Proof. It is a consequence of the identification C0 (X, Rn )∗ ≡ M(X, Rn )
and of theorem 4.9.ii), which asserts that the operator norm of µ ∈
C0 (X, Rn )∗ is |µ|(X). 
Exercise 4.30 (properties of the total variation, part II). Let X
be a locally compact separable metric space and µ an Rn -valued Radon
measure on X. Define, for all E ∈ B X :
i) µ1 (E) := sup{ mk=1 kµ(Ek )k | m ∈ N; ∀1 ≤ k ≤ m, Ek ∈ B X , |µ|(Ek ) <
P
m
˙
∞; ∪k=1 Ek ⊂ PE}.
ii) µ2 (E) := sup{ k∈N kµ(Ek )k | ∀k ∈ N, Ek ∈ B X , |µ|(Ek ) < ∞; ∪˙ k∈N Ek =
E}. ´
iii) µ3 (E) := sup{´E f · dµ | f ∈ L1 (|µ|, Rn ), kf k ≤ 1}.
iv) µ4 (E) := sup{ E f · dµ | f ∈ Cc (X, Rn ), kf k ≤ 1}.
Then µ1 (E) = µ2 (E) = µ3 (E) = µ4 (E) = |µ|(E).

4.2. Operations with Rn -valued Radon measures

We generalize to Rn -valued Radon measures the operations for pos-
itive measures introduced in definitions 1.13 and 1.14.
Definition 4.31 (restrictions of Rn -valued Radon measures). Let
X be a locally compact separable metric space, µ ∈ Cc (X, Rn )∗ an Rn -
valued Radon measure and g : X → R a bounded Borelian function on
x
X. We define the restriction of µ to g, denoted by µ g, as
´ the contin-
n
x
uous linear functional on Cc (X, R ) given by µ g · f := hf g, νi d|µ|
if (ν, |µ|) is the polar decomposition of µ.
Notation. If λ is a positive measure on X and h ∈ L+ (λ), we
x
introduce the alternative notation λ h to denote the measure on
X which has been denoted so far by hλ, i.e. the extension
´ given by
theorem 1.8 of the measure on σ(λ) given by A 7→ A h dλ. This
alternative notation is motivated by the following remark.
Remark 4.32. With the notation from the previous definition:
x
1) Note that µ g is indeed a well-defined continuous linear func-
tional: ∀K ⊂ X compact and ∀f ∈ CKc (X, Rn ), f g ∈ L1 (|µ|), µ g x
is clearly linear in f and, by the triangle inequality (remark 4.17.3),
x x
|µ g · f | ≤ |µ|(K)kgku kf ku , hence µ g is continuous.
x
2) The polar decomposition of µ g is ( |g|
gν
gν
, |g||µ|), where we define
|g|
:= 0 on the Borel set {g = 0}. That follows from the fact that
|g||µ| is a positive Radon measure on X (by lemma 4.11) and from
 4.2. OPERATIONS WITH RN -VALUED RADON MEASURES 105

the uniqueness of the polar decomposition. In particular, using the

notation above, we have
|µ xg| = |µ| x|g|.
3) If µ is a positive Radon measure on X (which we identify with
the element of Cc (X, R)∗ whose polar decomposition is (1, µ)) and
x
A ∈ B X , then µ χA coincides with the positive Radon measure
x
µ A (that this positive measure is Radon follows from proposition
1.36). We extend this notation for an arbitrary µ ∈ Cc (X, Rn )∗ , i.e.
x x
we use the notation µ A in place of µ χA . It then follows from
the previous item that
xA| = |µ| xA.
|µ
4) We may similarly define µ xg ∈ C (X, R )
c
n ∗
for µ ∈ Cc (X, R)∗
and g : X → Rn bounded Borelian:
´ that is the continuous linear
n
functional f ∈ Cc (X, R ) 7→ hf, giν d|µ| if (ν, |µ|) is the polar
gν
decomposition of µ. Then ( kgk , kgk|µ|) is the polar decomposition
x
of µ g. In particular,
|µ xg| = |µ| xkgk.
n ∗
Note that, with this definition, if µ ∈ C (X, R ) has polar de-
composition (ν, |µ|), then µ = |µ| xν.
c

5) We may also define µ xg for g ∈ L (|µ|) using the same formula.

1

and for all f ∈ C (X, R ), |µ xg·

loc
K n
´ for all K ⊂ X compact
Note that, ´ c
f| = hf g, νi d|µ| ≤ ( |g| d|µ|)kf k , whence the continuity of
u
µ xg. As before, the polar decomposition of µ xg is ( , |g||µ|),
K K
gν
|g|
which follows from the uniqueness of the polar decomposition and
from the fact that |g||µ| is a positive Radon measure on B X (by the
fact that it is finite on compact subsets of X and by remark 1.33).
Thus, as before,
x x
|µ g| = |µ| |g|.
6) As a final generalization of the restriction operation, we may define

x
µ T ∈ Cc (X, Rm )∗ for ´µ ∈ Cc (X, Rn )∗ and T ∈ L1loc |µ|, L(Rm , Rn )
by f ∈ Cc (X, Rm ) 7→ hT · f, νi d|µ|, where (ν, |µ|) is the polar
decomposition of µ. As before, it follows from the triangle in-
K m
equality that, for´ all K ⊂ X compact and for all f ∈ Cc (X, R ),
x x
|µ T · f | ≤ ( K kT k d|µ|)kf ku , hence µ T is indeed a continu-
ous linear functional. Note that, defining T ∗ : X → L(Rn , Rm ) by
x 7→ T (x)∗ , i.e. the adjoint of T , we have, ∀f ∈ Cc (X, Rm ):
ˆ ˆ
T∗ · ν
x
µ T · f = hT · f, νi d|µ| = f, ∗
kT · νk
kT ∗ · νk d|µ|.
106 4. RN -VALUED RADON MEASURES

Since kT ∗ ·νk d|µ| is a positive Radon measure on X, it follows as be-

∗
fore from the uniqueness of the polar decomposition that kTT ∗ ·ν
·νk
, kT ∗ ·


x
νk|µ| is the polar decomposition of µ T . In particular,
|µ xT | = |µ| xkT ∗
· νk.
Exercise 4.33 (Lebesgue decomposition and Radon-Nikodym de-
rivative for Rn -valued Radon measures). Let X be a locally compact
separable metric space µ ∈ Cc (X, Rn )∗ and λ a positive Radon measure
on X. We say that
• µ ⊥ λ if |µ| ⊥ λ in the sense of definition 3.34;
• µ  λ if |µ|  λ in the sense of 3.34.
Then:
a) (Lebesgue decomposition) There exist unique Rn -valued Radon mea-
sures µa , µs on X such that µs ⊥ λ, µa  λ, µ = µs + µa .
b) (Radon-Nikodym derivative) If µ  λ, there exists a unique (up
to λ-null sets) Borelian map f : X → Rn with f ∈ L1loc (λ) and
x
µ = λ f . We call f the Radon-Nikodym derivative of µ with
respect to λ and denote it by dµ
dλ
.
Exercise 4.34 (fundamental lemma of the Calculus of Variations).
Let X be an open set in Rm . If µ : Cc (X, Rn ) → R is an Rn -valued
Radon measure on X such that µ · f = 0 for all f ∈ C∞ n
c (X, R ), then
1 m n
µ = 0. In particular, if g ∈ Lloc (L |X , R ) and
ˆ
hf, gi dLm = 0
X

for all f ∈ C∞ n m
c (X, R ), then g = 0 L -a.e. on X.

Definition 4.35 (trace of Rn -valued Radon measures). Let X be

a locally compact separable metric space and A ⊂ X a locally compact
subspace of X (i.e the intersection of an open with a closed subset
of X). If µ is an Rn -valued Radon measure on X with polar decom-
position (ν, |µ|), we define an Rn -valued Radon measure µ|A on A by
´
f ∈ Cc (A, Rn ) 7→ hfe, νi d|µ|, where fe : X → Rn is the extension of f
by 0 in the complement of A.
Proposition 4.36. Let X be a locally compact separable metric
space, A ⊂ X a locally compact subspace and µ ∈ Cc (X, Rn )∗ with po-
lar decomposition (ν, |µ|). Then µ|A is a well-defined Rn -valued Radon
measure on A and it is finite if so is µ. Moreover, the polar decomposi-
tion of µ|A is (ν|A , |µ| A ), where |µ| A denotes the trace of |µ| on A in
 4.2. OPERATIONS WITH RN -VALUED RADON MEASURES 107

the sense of definition 1.13. In particular, if µ is a positive Radon mea-

sure on X, the trace of µ on A in the sense of definition 4.35 coincides
with the trace in the sense of definition 1.13.
Proof. Note that fe is a bounded Borel measurable function and
vanishes on the complement of a compact subset of A (which is also
a compact subset of X, since the inclusion is continuous and X is
Hausdorff), hence fe ∈ L1 (|µ|) and the integral makes sense, i.e. µ|A ·
f is well-defined and clearly linear in f . Moreover, for all K ⊂ A
compact and f ∈ CKc (A, Rn ), we have, by the triangle inequality µ|A ·
´
f = K hfe, νi d|µ| ≤ |µ|(K)kf ku , hence µ|A is continuous, i.e. µ|A ∈
Cc (A, Rn )∗ .
For all f ∈ Cc (A, Rn ), it follows from exercise 1.69 that µ|A · f =
´ ´ ´
hfe, νi d|µ| = A hfe, νi d|µ| = hf, ν|A i d|µ| A , where |µ| A denotes
the trace of |µ| on A in the sense of definition 1.13. Since |µ| A is a
positive Radon measure on B A (it is a Borelian measure by proposition
1.15.ii and it is finite on compact subsets of A, hence it is Radon by
remark 1.33), and since ν|A is Borelian with kν|A k = 1 |µ| A -a.e., it
follows that the polar decomposition of µ|A is (ν|A , |µ| A ). In particular,
|µ|A | = |µ| A . Hence, if µ is finite (i.e. if |µ| is a positive finite Radon
measure), so is µ|A (since its total variation is finite). 
We next introduce the pushforward operation for Rn -valued Radon
measures by transposition. For that purpose, and for those who are
not acquainted with locally convex spaces and LF topologies, we make
an ad hoc definition of continuity which generalizes the notion of con-
tinuity introduced in definition 4.1 (and coincides with the notion of
continuity with respect to the locally convex topologies of the spaces
involved, cf. remark 4.2).
Definition 4.37 (continuity of linear maps on Cc (X, Rn )). Let X
and Y be locally compact separable metric spaces.
i) We say that A ⊂ Cc (X, Rn ) is bounded it there exists K ⊂ X
compact such that A ⊂ CKc (X, Rn ) and A is bounded in the latter
space (i.e. it bounded as a subset of the Banach space CKc (X, Rn )).
ii) We say that a sequence (xn )n∈N in Cc (X, Rn ) converges to x ∈
Cc (X, Rn ) if there exists K ⊂ X compact such that the image of
the sequence is contained in CKc (X, Rn ), x ∈ CKc (X, Rn ) and xn → x
in CKc (X, Rn ).
iii) We say that a linear map T : Cc (X, Rn ) → Cc (Y, Rm ) is continuous
if one of the following equivalent conditions hold:
• T (A) is bounded whenever A ⊂ Cc (X, Rn ) is bounded.
108 4. RN -VALUED RADON MEASURES

• T (xn ) → 0 whenever (xn )n∈N is a sequence in Cc (X, Rn ) such

that xn → 0.
Remark 4.38. That the two conditions in part iii) of the above
definition are indeed equivalent to continuity in the LF topology of
the spaces involved, cf. remark 4.2, is a consequence of the fact that
LF spaces are bornological. Actually, in the above definition, we could
replace the codomain of T by any locally convex space; in particular,
those conditions may also be used to characterize continuity of linear
functionals Cc (X, Rn ) → R.
Proposition 4.39. Let X and Y be locally compact separable met-
ric spaces and T : Cc (X, Rn ) → Cc (Y, Rm ) a linear map.
i) If T is continuous and µ is an Rm -valued Radon measure on Y ,
then µ ◦ T is an Rn -valued Radon measure on X.
ii) If T is continuous with respect to the C0 topology (i.e. the topology
induced by k·ku ) on both domain and codomain, and µ is a finite
Rm -valued Radon measure on Y , then µ ◦ T is a finite Rn -valued
Radon measure on X.
Proof. It is immediate from the definitions. 
Definition 4.40. With the notation from the previous proposition,
we define the transpose of T , T t : Cc (Y, Rm )∗ → Cc (X, Rn )∗ in case (i)
or T t : C0 (Y, Rm )∗ → C0 (X, Rn )∗ in case (ii), by T t · µ := µ ◦ T .
Example 4.41. Let X be a locally compact separable metric space.
1) Let T : X → L(Rm , Rn ) be a continuous map. We define T̂ :
Cc (X, Rm ) → Cc (X, Rn ) by (T̂ · f )(x) := T (x) · f (x). Then T̂ is
clearly linear and, for all K ⊂ X compact and f ∈ CKc (X, Rm ), we
have T̂ · f ∈ CKc (X, Rn ) and kT̂ · f ku ≤ kT |K ku kf ku , which implies
the continuity of T̂ . The transpose of T̂ is given by µ 7→ µ T , x
x
where µ T was defined in part 6) of remark 4.32 (but the situation
in that remark is more general, since it the continuity of T is not
required).
2) Let U ⊂ X open. The inclusion Cc (U, Rn ) ⊂ Cc (X, Rn ) (which
maps f ∈ Cc (U, Rn ) to its extension by 0 on the complement of U )
is clearly continuous; its transpose coincides with µ 7→ µ|U , where
µ|U is the trace of µ on U in the sense of definition 4.35. For a
general locally compact subspace A ⊂ X, we cannot define the trace
by means of transposition, since, in general, there is no canonical
inclusion Cc (A, Rn ) ⊂ Cc (X, Rn ).
Proposition 4.42. Let X and Y be locally compact separable met-
ric spaces and f : X → Y a continuous proper map. Then both
 4.2. OPERATIONS WITH RN -VALUED RADON MEASURES 109

(◦f ) : Cc (Y, Rn ) → Cc (X, Rn ) and (◦f ) : C0 (Y, Rn ) → C0 (X, Rn ) given

by g 7→ g ◦ f are well-defined and linear continuous.
Proof. If g ∈ Cc (Y, Rn ), then spt (g ◦ f ) ⊂ f −1 (spt g), which is
compact, since f is proper, hence g ◦ f ∈ Cc (X, Rn ). If g ∈ C0 (Y, Rn )
and  > 0, there exists K ⊂ Y compact such that K c ⊂ {kgk < }.
Since f is proper, f −1 (K) is compact and f −1 (K)c = f −1 (K c ) ⊂
{kg ◦ f k < }, hence g ◦ f ∈ C0 (X, Rn ). Thus, both (◦f ) : Cc (Y, Rn ) →
Cc (X, Rn ) and (◦f ) : C0 (Y, Rn ) → C0 (X, Rn ) are well-defined and
clearly linear. It remains to prove their continuity. Indeed, for all K ⊂
f −1 (K)
Y compact and for all g ∈ CKc (Y, Rn ), we have g ◦ f ∈ Cc (X, Rn )
and kg ◦f ku ≤ kgku , which implies the continuity of (◦f ) : Cc (Y, Rn ) →
Cc (X, Rn ), and the continuity of (◦f ) : C0 (Y, Rn ) → C0 (X, Rn ) follows
by the same argument. 
Definition 4.43. With the notation from definition 4.42, the trans-
poses (◦f )t : Cc (X, Rn )∗ → Cc (Y, Rn )∗ and (◦f )t : C0 (X, Rn )∗ →
C0 (Y, Rn )∗ are called pushforward by f and denoted by f# :µ 7→ f# µ.
Proposition 4.44. Let X and Y be locally compact separable met-
ric spaces, f : X → Y a continuous proper map and µ ∈ Cc (X, Rn )∗
with polar decomposition (νX , |µ|). Suppose that there exists a Borelian
map νY : Y → Rn such that νY ◦ f = νX . Then the polar decomposi-
tion of f# µ is (νY , f# |µ|), where f# |µ| is the pushforward of |µ| by f in
the sense of the definition 1.14. In particular, if µ is a positive Radon
measure on X, the pushforward of µ by f in the sense of definition
4.43 coincides with the pushforward in the sense of definition 1.14.
Proof. For all g ∈ Cc (Y, Rn ), we have:
ˆ
f# µ · g = µ · (g ◦ f ) = hg ◦ f, νX i d|µ| =
ˆ
ex.1.70
= hg ◦ f, νY ◦ f i d|µ| =
ˆ
= hg, νY i df# |µ|,

where f# |µ| is the pushforward of |µ| by f in the sense of the definition

1.14. Since f# |µ| is a positive

Radon measure on Y , by proposition 1.37,
−1
and since f {kνY k 6= 1} ⊂ {kνX k 6= 1} is |µ|-null (i.e. {kνY k 6=
1} is f# |µ|-null), we conclude that the polar decomposition of f# µ is
(νY , f# |µ|), as asserted.
In particular, if µ is a positive Radon measure on X, the polar
decomposition of µ̌ (using the notation of definition 4.23 for clarity) is
110 4. RN -VALUED RADON MEASURES

(1, µ), so that we can take νY ≡ 1, hence the polar decomposition of

f# µ̌ is (1, f# µ), whence the thesis. 
Exercise 4.45. Let X and Y be locally compact separable metric
spaces, T : X → L(Rm , Rn ) continuous, f : X → Y a continuous proper
map and µ an Rn -valued Radon measure on X. Suppose that there
exists S : Y → L(Rm , Rn ) such that S ◦ f = T . Then f# (µ T ) = x
x
f# µ S. In particular, if f is an homeomorphism, then f# (µ T ) = x
x
f# µ f# T , where f# T := T ◦ f −1 .
Remark 4.46. We may define the pushforward with respect to
more general maps. For instance, let X and Y be locally compact
separable metric spaces, µ ∈ Cc (X, Rn )∗ with polar decomposition
(ν, |µ|) and
f : X → Y a Borelian map such that ∀K ⊂ Y compact,
|µ| f −1 (K) < ∞. We define f# µ : Cc (Y, Rn ) → R by
ˆ
g 7→ hg ◦ f, νi d|µ|.

Note that the integral makes sense, since kg ◦f k ∈ L1 (|µ|) (because it is

a bounded Borelian map which vanishes in the complement of the |µ|-
finite set f −1 (spt g)). Moreover, f# µ is clearly linear
´ and, for all K ⊂ Y
K n
compact and for all g ∈ Cc (Y, R ), |f# µ · g| = f −1 (K) hg ◦ f, νi d|µ| ≤
|µ| f −1 (K) kgku , hence f# µ ∈ Cc (Y, Rn )∗ , i.e. f# µ is indeed an Rn -


valued Radon measure on Y . If f is a continuous proper map, the
latter pushforward coincides with the one defined by transposition in
definition 4.43. As before, we may find the polar decomposition of f# µ
from the polar decomposition (νX , |µ|) if there exists a Borelian map
νY : Y → Rn such that νY ◦ f = νX . For that purpose, we take f# |µ| in
the sense of remark 1.38, which is a positive Radon measure on Y , and
repeat the same computations in the proof of the previous proposition,
for g ∈ Cc (Y, Rn ):
ˆ
f# µ · g = hg ◦ f, νX i d|µ| =
ˆ
remark 1.71
= hg ◦ f, νY ◦ f i d|µ| =
ˆ
= hg, νY i df# |µ|,

where f# |µ| is the pushforward of |µ| by f in the sense of remark 1.38.

Since f# |µ| is a Radon measure on Y and kνY k = 1 f# |µ|-a.e. on Y ,
we conclude that the polar decomposition of f# µ is (νY , f# |µ|).
 4.3. WEAK-STAR CONVERGENCE 111

4.3. Weak-star convergence

Definition 4.47. Let X be a locally compact separable metric
space. We say that
i) a sequence (µk )k∈N in Cc (X, Rn )∗ is weakly-star convergent ´ to µ ∈
n ∗ ∗ n
C
´c (X, R ) (notation: µ k *µ) if, for all f ∈ Cc (X, R ), f · dµ k →
f · dµ;
ii) a sequence (µk )k∈N in C0 (X, Rn )∗ is weakly-star convergent in the
∗f
sense of finite measures´ to µ ∈ C0´(X, Rn )∗ (notation: µk *µ) if,
for all f ∈ C0 (X, Rn ), f · dµk → f · dµ.

Remark 4.48.
1) Some authors use the nomenclature locally weakly star convergent
and weakly star convergent for our definitions above in i) and ii),
respectively.
2) We have used different names to distinguish one from the other, but
both types of convergence above are actually the same notion, i.e.
convergence of sequences with respect to weak star topologies: the
first type in the weak-star dual of Cc (X, Rn ) and the second in the
weak-star dual of C0 (X, Rn ). Note that, in general, none of these
weak-star topologies is first-countable, so that we may have to use
nets or filters to handle general topological problems. However, note
that, thanks to the Banach-Alaoglu theorem and to the separability
of C0 (X, Rn ), strongly bounded subsets of C0 (X, Rn )∗ are relatively
compact and metrizable in the weak-star topology of C0 (X, Rn )∗ .

Proposition 4.49 (relation between weak-star convergence and

weak-star convergence in the sense of finite measures). Let X be a lo-
cally compact separable metric space, (µk )k∈N a sequence in Cc (X, Rn )∗
and µ ∈ Cc (X, Rn )∗ . The following conditions are equivalent:
∗
i) µk * µ and supk∈N |µk |(X) < ∞.
∗f
ii) (µk )k∈N is a sequence in C0 (X, Rn )∗ , µ ∈ C0 (X, Rn )∗ and µk * µ.

Proof.
(i ⇒ ii): For every f ∈ Cc (X, Rn ) such that kf ku ≤ 1, µ · f =
lim µk · f ≤ lim inf |µk |(X) ≤ supk∈N |µk |(X) < ∞; taking the
sup over all such f , we conclude that |µ|(X) ≤ lim inf |µk |(X) <
∞, i.e. µ ∈ C0 (X, Rn )∗ . Moreover, given g ∈ C0 (X, Rn ) and
 > 0, we may take f ∈ Cc (X, Rn ) such that kf − gku < ;
112 4. RN -VALUED RADON MEASURES

hence, for all k ∈ N,

|µk · g − µ · g| ≤ |µk · g − µk · f | + |µk · f − µ · f | + |µ · f − µ · g| ≤
≤ |µk |(X)kf − gku + |µk · f − µ · f | + |µ|(X)kf − gku ≤
≤ 2 sup|µk |(X) + |µk · f − µ · f |.
k∈N

Taking k → ∞, it follows that lim sup|µk ·g−µ·g| ≤ 2 supk∈N |µk |(X);

by the arbitrariness of the  > 0 taken, it then follows that
∗f
lim sup|µk · g − µ · g| = 0, i.e. µk · g → µ · g, whence µk * µ.
(ii ⇒ i): For all g ∈ C0 (X, Rn ), µk · g → µ · g; in particular,
that holds for g ∈ Cc (X, Rn ) and, by the principle of uniform
boundedness, supk∈N |µk |(X) < ∞ (recall that the operator
norm of µk ∈ C0 (X, Rn )∗ is |µk |(X)).

Proposition 4.50. Let X and Y be locally compact separable met-
ric spaces and T : Cc (X, Rn ) → Cc (Y, Rm ) linear continuous. Then
T t : Cc (Y, Rm )∗ → Cc (X, Rn )∗ preserves weak-star convergence of se-
quences. The same holds for weak-star convergence in the sense of
finite measures if T is continuous with respect to the C0 topologies.
Proof. It is immediate from the definitions. 
Remark 4.51.
1) Actually, with the same hypothesis from the previous proposition ,
T t : Cc (Y, Rm )∗ → Cc (X, Rn )∗ and T t : C0 (Y, Rm )∗ → C0 (X, Rn )∗
are continuous with respect to the corresponding weak-star topolo-
gies.
2) In particular, this proposition applies to the operations with Rn -
valued Radon measures which may be defined by transpositions:
x
the restriction µ 7→ µ T with T : X → L(Rm , Rn ) continuous
(example 4.41.1), the trace on open sets µ 7→ µ|U with U ⊂ X open
(example 4.41.2) and the pushforward by a continuous proper map
(definition 4.43 and proposition 4.44).
Exercise 4.52.
a) Let (xk )k∈N be a sequence in Rn convergent to x ∈ Rn . Then
∗f
δxk * δx .
b) (concentration of mass) Let (µk )k∈N be the sequence of positive
Radon measures on Rn given by (∀k ∈ N)µk := k n Ln (0, k −1 )n .
∗f
x
Then µk * δ.
c) (spreading of mass) Let (µk )k∈N be the sequence of positive Radon
Pk −1
measures on Rn given by (∀k ∈ N)µk := m=1 k δm/k . Then
∗f
x
µk * L1 (0, 1).
 4.3. WEAK-STAR CONVERGENCE 113

d) Let (µk )k∈N be a sequence of signed Radon measures on Rn (i.e.

signed measures on B Rn whose total variation is Radon) and µ a
∗
signed Radon measure on Rn such that µk * µ. It is not necessarily
∗ − ∗ ∗
true that µ+ + −
k * µ , µk * µ or |µk | *|µ|.

Proposition 4.53 (foliations by Borel sets for positive Radon mea-

sures). Let X be a locally compact separable metric space, µ a positive
Radon measure on X and (Eα )α∈A a disjoint family of Borel sets in X.
Then {α ∈ A | µ(Eα ) > 0} is countable.
Proof. Since X can be covered by countably many relatively com-
pact open sets, it suffices to show that, for each relatively compact open
set U , the thesis holds
0
x
P for the finite Radon measure ν := µ U . For
each A ⊂ A finite, α∈A0 ν(Eα ) = ν(∪˙ α∈A0 Eα ) ≤ ν(X) < ∞. It then
follows that α ∈ A 7→ ν(Eα ) is summable with respect to the counting
measure, hence {α ∈ A | ν(Eα ) > 0} is countable, as asserted. 
Theorem 4.54 (characterization of weak-star convergence for pos-
itive Radon measures). Let X be a locally compact separable metric
space, (µk )k∈N a sequence of positive Radon measures in X and µ a
positive Radon measure in X. The following conditions are equivalent:
∗
i) µk * µ.
ii) For all K ⊂ X compact and for all U ⊂ X open,
µ(K) ≥ lim sup µk (K) and µ(U ) ≤ lim inf µk (U ).
iii) For all E ∈ B cX such that µ(∂E) = 0, µk (E) → µ(E).
∗
Moreover, if µk * µ and x ∈ spt µ, there exists n ∈ N and a se-
quence (xk )k≥n in X such that ∀k ≥ n, xk ∈ spt µk and xk → x.
Proof.
(i⇒ii) For each f ∈ Cc (U, R) with |f | ≤ 1, µ · f = lim µk · f ≤
lim inf µk (U ). Taking the sup over all such f , it follows that µ(U ) ≤
lim inf µk (U ).
Note that µ(K) = inf{µ · f | f ∈ Cc (X, R), χK ≤ f ≤ 1} (the
inequality ≤ is clear and the reverse inequality follows from the outer
regularity of µ in K and from the Urysohn lemma 4.5). For ´ each f ∈
Cc (X, R) such that χK ≤ f ≤ 1, µ · f = lim µk · f ≥ lim sup χK dµk =
lim sup µk (K). Taking the inf over all such f , we conclude that µ(K) ≥
lim sup µk (K).
(ii⇒iii) Let E ∈ B cX such that µ(∂E) = 0. Applying (ii) for the
compact set K = E and for the open set U = E o , it follows that
µ(E) = µ(E) ≥ lim sup µk (E) ≥ lim sup µk (E) and µ(E) = µ(E o ) ≤
lim inf µk (E o ) ≤ lim inf µk (E), whence µ(E) = lim µk (E), as asserted.
114 4. RN -VALUED RADON MEASURES

(iii⇒i) We must show that, for every f ∈ Cc (X, R), µk · f → µ · f .

Since an arbitrary f ∈ Cc (X, R) can be written as f = f + − f − with
f ± ≥ 0 in Cc (X, R), it suffices to consider the case f ≥ 0. So, fix f ≥ 0
in Cc (X, R).
Since f is continuous with compact support, for every t > 0 the
set {f > t} is open relatively compact and ∂{f > t} ⊂ {f = t}.
We may apply proposition 4.53 to the disjoint family of Borel sets
({f = t})t>0 to conclude that there exists a countable set I ⊂ (0, ∞)
such that µ({f = t}) = 0 for t ∈ (0, ∞) \ I; hence µ(∂{f > t}) = 0 for
t ∈ (0, ∞) \ I. It then follows from (iii) that, for every t ∈ (0, ∞) \ I,
µk ({f > t}) → µ({f > t}).
Define ∀k ∈ N, F, Fk : (0, ∞) → [0, ∞) by Fk (t) := µk ({f > t}) and
F (t) := µ({f > t}). Then F, Fk are Borelian (since they are decreasing
functions) and, as we saw in the previous paragraph, Fk → F pointwise
in (0, ∞) \ I, i.e. Fk → F L1 -a.e. (since countable sets have Lebesgue
measure 0). We contend that the convergence is dominated. Indeed,
∀k ∈ N, ∀t > 0, Fk (t) ≤ µk (spt f )χ[0,kf ku ] (t); hence, if we show that
C := supk∈N µk (spt f ) < ∞, then the sequence Fk will be dominated
by the L1 -summable function Cχ[0,kf ku ] , thus proving our contention.
In order to show that supk∈N µk (spt f ) < ∞, take a relatively com-
pact open set U ⊃ K := spt f and, denoting by d the distance in X,
let ∀0 <  < d(K, U c ), U := {x ∈ X | d(x, K) < }; applying proposi-
tion 4.53 to the Radon measure µ and the disjoint family of Borel sets
(∂U )0<<d(K,U c ) , we conclude that there exists 0 <  < d(K, U c ) such
that µ(∂U ) = 0. Since U is open and relatively compact (because it is
contained in U ), it then follows from (iii) that µk (U ) → µ(U ). Hence
supk∈N µk (spt f ) ≤ supk∈N µk (U ) < ∞, as asserted.
Finally, by the dominated convergence theorem 1.64 and by the
layer-cake formula 1.87 (note that µ, µk are σ-finite, since they are
Radon and X is σ-compact), we have:
ˆ ˆ ∞
µk · f = f dµk = Fk dL1 →
ˆ ∞ 0
ˆ
1
→ F dL = f dµ = µ · f,
0

as we wanted to show.
∗
Finally, suppose that µk * µ and let x ∈ spt µ. We claim that ∀ >
0, there exists n ∈ N such that ∀k ≥ n, U(x, )∩spt µk 6= ∅. Arguing by
contradiction, suppose that there exists  > 0 and a sequence (kn )n∈N
in N such that kn →
∞ and U(x, ) ∩ spt
µkn = ∅. It then follows


from ii) that µ U(x, ) ≤ lim inf µk U(x, ) ≤ lim µkn U(x, ) = 0,
 4.3. WEAK-STAR CONVERGENCE 115


hence µ U(x, ) = 0, which contradicts the fact that x ∈ spt µ and
proves the claim. We now apply the claim for  = i ∈ N, yielding a
sequence (ni )∈N , which we may assume to be strictly increasing. Then,
for each i ∈ N, we may choose xni , xni +1 , . . . , xni+1 −1 such that xj ∈
U(x, 1/i) ∩ spt µj for ni ≤ j < ni+1 , thus yielding a sequence (xj )j≥n1
in X such that ∀j ≥ n1 , xj ∈ spt µj and xj → x.


Exercise 4.55. Let X be a locally compact separable metric space,

(µk )k∈N a sequence of positive Radon measures on X and µ a positive
∗
Radon measure on X. Suppose
that µk * µ and for every r > 0,
lim supk→∞ inf{µk U(x, r) | x ∈ spt µk } > 0. If (xk )k∈N is a conver-
gent sequence in X with ∀k ∈ N, xk ∈ spt µk , then lim xk ∈ spt µ.

Exercise 4.56 (narrow convergence). We say that a sequence (µk )k∈N

in C0 (X, Rn )∗ is narrowly convergent
´ to µ ´∈ C0 (X, Rn )∗ (notation:
∗ nc
µk *µ) if, for all f ∈ Cb (X, Rn ), f · dµk → f · dµ, where Cb (X, Rn )
denotes the Banach space of bounded continuous functions X → Rn
∗ nc
(endowed with the norm of uniform convergence k·ku ). That is, µk * µ
if it converges to µ in the weak-star dual of Cb (X, Rn ).
If (µk )k∈N is a sequence of positive finite Radon measures on X and
∗ nc
µ is a positive finite Radon measure on X, then µk * µ iff µk (X) →
µ(X) and ∀A ⊂ X open, µ(A) ≤ lim inf µk (A).
∗ nc
Hint. To prove´ that the ´stated condition implies µk * µ, it suf-
fices to show that g dµk → g dµ for g ∈ Cb (X, R) with 0´≤ g ≤ 1
(since ´Cb (X, R) is the linear span of such g). Prove that g dµ ≤
lim inf g dµk , using the layer-cake formula 1.87 to compute the inte-
grals and Fatou’s lemma. The same holds for 1 − g in the place of
g.

Proposition 4.57 (weak convergence and total variation, part I).

Let X be a locally compact separable metric space and (µk )k∈N a se-
quence in Cc (X, Rn )∗ weakly-star convergent to µ ∈ Cc (X, Rn )∗ . Then,
for every A ⊂ X open, |µ|(A) ≤ lim inf|µk |(A).

Proof. For every f ∈ Cc (X, Rn ) with spt f ⊂ A and kf k ≤ 1, we

have µ · f = lim µk · f ≤ lim inf|µk |(A). Taking the sup over all such f
yields the thesis. 

Proposition 4.58 (weak convergence and total variation, part II).

Let X be a locally compact separable metric space and (µk )k∈N a se-
quence in Cc (X, Rn )∗ weakly-star convergent to µ ∈ Cc (X, Rn )∗ .
116 4. RN -VALUED RADON MEASURES

∗
i) If ν is a positive Radon measure on X and |µk | * ν, then ∀E ⊂ X,
|µ|(E) ≤ ν(E). Moreover, if E ∈ B cX and ν(∂E) = 0, then
µk (E) → µ(E).
∗f ∗ nc
ii) If |µk |(X) → |µ|(X) < ∞, then |µk | *|µ| (actually |µk | *|µ| in
the sense of exercise 4.56).
Proof.
i) It suffices, by outer regularity, to prove the inequality for E ⊂ X
open. Let A ⊂ X open with A b E, and take f ∈ Cc (X, R)
such that χA ≤ f ≤ 1 and spt f ⊂ E (which exists, by Urysohn’s
lemma 4.5). Then
4.57
|µ|(A) ≤ lim inf|µk |(A) ≤
ˆ ∗
ˆ
|µk | * ν
≤ lim inf f d|µk | = lim f dν ≤
≤ ν(E).
Since, by inner regularity, |µ|(E) = sup{|µ|(K) | K ⊂ E compact} =
sup{|µ|(A) | A ⊂ E open, A b E}, taking the sup in the inequality
above yields |µ|(E) ≤ ν(E), as asserted.
Suppose that E ∈ B cX with ν(∂E) = 0. Fix  > 0. It follows
from lemma 4.59 below that there exists K ⊂ X compact and
A ⊂ X open such that A ⊂ E ⊂ K o and ν(K \ A) < . Take
f ∈ Cc (X, R) such that χA ≤ f ≤ 1 and spt f ⊂ K o . We then
have, for all k ∈ N:
ˆ ˆ
f dµk − µk (E) ≤ |f − χE | d|µk | ≤ |µk |(K \ A)
ˆ
f dµ − µ(E) ≤ |µ|(K \ A) ≤ ν(K \ A)

Thus,
ˆ ˆ ˆ ˆ
|µk (E) − µ(E)| ≤ f dµk − µk (E) + f dµ − µ(E) + f dµk − f dµ ≤
ˆ ˆ
≤ |µk |(K \ A) + ν(K \ A) + f dµk − f dµ .
´ ´
Since limk→∞ f dµk − f dµ = 0 and, by the fact that K \ A
∗
is compact, |µk | * ν and theorem 4.54.ii), lim sup|µk |(K \ A) ≤
ν(K \ A), it follows that
lim sup|µk (E) − µ(E)| ≤ 2ν(K \ A) < 2.
 4.3. WEAK-STAR CONVERGENCE 117

By the arbitrariness of the  > 0 taken, we conclude that µk (E) →

µ(E), as asserted.
ii) It follows from proposition 4.57 that, for all A ⊂ X open, |µ|(A) ≤
lim inf|µk |(A). Since |µk |(X) → |µ|(X) < ∞, we may assume, ex-
cluding the first terms of the sequence if necessary, that |µk |(X) <
∗ nc
∞ for all k ∈ N. It then follows from exercise 4.56 that |µk | *|µ|;
∗f
in particular, |µk | *|µ|.


Lemma 4.59. Let X be a locally compact Hausdorff space, µ a pos-

itive Radon measure on X and E ∈ B cX such that µ(∂E) = 0. Then,
for every  > 0, there is a compact set K ⊂ X and an open set A ⊂ X
such that A ⊂ E b K o and µ(K \ A) < .

Proof. Fix  > 0. Since E b X, we may take a compact set

K ⊂ X such that E b K o and ν(K \ E) < /2. Indeed, by outer
regularity, there exists U ⊃ E open such that ν(U \ E) < /2; take
a relatively compact open set V such that E ⊂ V b U (which exists,
since E is compact) and put K := V . Then E b K o and, as ν(∂E) = 0,
we have ν(K \ E) = ν(K \ E) ≤ ν(U \ E) < /2.
Similarly, by inner regularity there exists a compact set C ⊂ E o
such that µ(E o \ C) < /2. Take a relatively compact open set A ⊂ X
such that C ⊂ A b E o . Since µ(∂E) = 0, we have µ(E \ A) =
µ(E o \ A) ≤ µ(E o \ C) < /2. Finally, since K \ A = (K \ E) ∪ (E \ A),
the thesis follows. 

Exercise 4.60. Let X be a locally compact separable metric space,

(µj )j∈N a sequence of Rn -valued Radon measures on X weakly-star
convergent to an Rn -valued Radon measure µ on X and (Vm )m∈N an
increasing sequence of relatively compact open subsets of X such that
X = ∪m∈N Vm . Suppose that ∀m ∈ N, limj→∞ |µj |(Vm ) = |µ|(Vm ).
∗
Then |µj | *|µ|.

Theorem 4.61 (De La Vallée Poussin). Let X be a locally compact

separable metric space and (µk )k∈N be a sequence of finite Rn -valued
Radon measures on X such that sup{|µk |(X) | k ∈ N} < ∞. Then
there exists a finite Rn -valued Radon measure µ on X and a subse-
∗f
quence (µkj )j∈N of (µk )k∈N such that µkj * µ. Moreover, |µ|(X) ≤
lim inf|µkj |(X).

Proof. The first assertion is a direct consequence of the fact that

strongly closed balls in C0 (X, Rn )∗ are compact and metrizable in the
118 4. RN -VALUED RADON MEASURES

weak-star topology (hence sequentially compact): the compactness fol-

lows from Banach-Alaoglu theorem, and the metrizability follows from
the fact that C0 (X, Rn ) is a separable Banach space.
The second assertion is a consequence of the first and of proposition
4.57. 
Remark 4.62. The second assertion in the previous proposition is
also a consequence of the fact that, for any Banach space Y , the norm
k·k of Y ∗ is weakly-star lower semicontinuous, since k·k = sup{hy, ·i |
y ∈ Y, kyk ≤ 1} is the sup of a family of weakly-star continuous func-
tions.
Corollary 4.63. Let X be a locally compact separable metric space
and (µk )k∈N be a sequence of Rn -valued Radon measures on X such
that, for any K ⊂ X compact, sup{|µk |(K) | k ∈ N} < ∞. Then there
exists an Rn -valued Radon measure µ on X and a subsequence (µkj )j∈N
∗
of (µk )k∈N such that µkj * µ.
Proof. Let (Vm )m∈N be an increasing sequence of relatively com-
pact open subsets of X such that ∪m∈N Vm = X. We apply De La
Vallée Poussin’s theorem 4.61 to each of the traces (µk |Vm )k∈N , m ∈ N,
and then we use a diagonal argument:
1) Since ∀k ∈ N, µk |V1 = |µk | V1 (by proposition 4.36) and supk∈N |µk |(V1 ) ≤
supk∈N |µk |(V 1 ) < ∞, there exists ν1 ∈ C0 (V1 , Rn )∗ and a subse-
∗f
quence µ1 = (µ1k )k∈N of (µk )k such that µ1k |V1 * ν1 .
2) Suppose that we have defined subsequences µ1 , . . . , µi of (µk )k∈N
and νj ∈ C0 (Vj , Rn )∗ for 1 ≤ j ≤ i such that µj is a subsequence
∗f
of µj−1 for 2 ≤ j ≤ i and µjk |Vj * νj for 1 ≤ j ≤ i. We reapply
to µi the argument of the previous item to find a subsequence µi+1
∗f
if µi and νi+1 ∈ C0 (Vi+1 , Rn )∗ such that µi+1 k |Vi+1 * νi+1 . Induc-
tively, we have thus defined a sequence (µi )i∈N of subsequences of
the original sequence (µk )k∈N and a sequence (νi )i∈N with ∀i ∈ N,
νi ∈ C0 (Vi , Rn )∗ .
3) Take the subsequence (λk )k∈N of (µk )k∈N given by λk := µkk . For all
∗f
i ∈ N, (λk )k∈N is a subsequence of µi , hence λk |Vi * νi . In particular,
given f ∈ Cc (X, Rn ) and i ∈ N such that spt f b Vi , we have
ˆ ˆ
k→∞
λk · f = f · dλk = f |Vi · dλk |Vi → νi · f,

hence ν : Cc (X, Rn ) → R given by f 7→ lim λk · f (= νi · f for any

i ∈ N such that spt f b Vi ) is a well-defined linear functional. It
is continuous, i.e. ν ∈ Cc (X, Rn )∗ , since, for each K ⊂ X compact,
 4.3. WEAK-STAR CONVERGENCE 119

we can take i ∈ N such that K b Vi , hence ν|CKc (X,Rn ) = νi |CKc (X,Rn ) .

∗
Since λk * ν, the thesis follows.

 CHAPTER 5

Area and Coarea Formulas

In this chapter we study Lipschitz maps Rn → Rm and two gener-

alizations of the change of variables formula 1.82: the area formula, for
n ≤ m, and the coarea formula (which is also an extension of Fubini-
Tonelli’s theorem), for n ≥ m. Both theorems have the same statement
for n = m.

5.1. Lipschitz maps on Rn

Recall that, given metric spaces X and Y , a map f : X → Y is

called Lipschitz if there exists C ≥ 0 such that ∀x, y ∈ X, dY f (x), f (y) ≤
CdX (x, y). If f is Lipschitz, there exists a smallest such constant C,
namely

dY f (x), f (y)
Lip f := sup{ | x 6= y ∈ X},
dX (x, y)
called Lipschitz constant of f .
In this section we derive some basic properties of Lipschitz maps
R → Rm .
n

5.1.1. Extensions of Lipschitz maps. Let A ⊂ Rn and f : A →

m
R a Lipschitz map. As we will see in subsequent developments, it is
useful to be able to extend f to a Lipschitz map with the same Lipschitz
constant defined on all of Rn . The theorems stated below ensure the
existence of such extensions.
Firstly, we consider the case m = 1:
Theorem 5.1 (McShane’s lemma). Let A ⊂ Rn and f : A → R a
Lipschitz map. Define F : Rn → R by:
(5.1) F (x) := inf{f (a) + Lip f · kx − ak | a ∈ A}.
Then F extends f and Lip F = Lip f .
That F is well defined (i.e. the second member in the previous
equality is > −∞, so that F indeed takes values in R) will be seen as
part of the proof.
The geometric idea behind formula (5.1) is the following: g : X → R
is a Lipschitz function on the metric space X iff there exists C ≥ 0
121
122 5. AREA AND COAREA FORMULAS


such that gr g = { x, g(x) | x ∈ X} ⊂ X × R lies in the intersection
of all cones {(x, y) ∈ X × R | |y − g(a)| ≤ Ckx − ak} for a ∈ X. If that
is the case, the least such C is the Lipschitz constant of g.
Proof.
1) The formula in the statement of the theorem defines F : Rn →
[−∞, ∞). We shall prove that Im F ⊂ R.
2) If x ∈ A, it is clear that the infimum in (5.1) is attained for a = x,
since ∀a ∈ A, f (x) − f (a) ≤ |f (x) − f (a)| ≤ Lip f kx − ak, hence
f (x) ≤ inf{f (a) + Lip f · kx − ak | a ∈ A}. Thus, F (x) = f (x), i.e.
F extends f .
3) If x, y ∈ Rn and a ∈ A, F (x) ≤ f (a) + Lip f · kx − ak ≤ f (a) + Lip f ·
ky − ak + Lip f · kx − yk. Taking the infimum of the second member
over all a ∈ A, we conclude that F (x) ≤ F (y) + Lip f · kx − yk. In
particular, if x ∈ A, we conclude that F (y) ≥ f (x)−Lip f ·kx−yk >
−∞ for all y ∈ Rn , hence Im F ⊂ R.
Exchanging x and y, we also have F (y) ≤ F (x) + Lip f · kx − yk,
so that |F (x) − F (y)| ≤ Lip f · kx − yk. Hence, F is Lipschitz with
Lipschitz constant ≤ Lip f ; since it extends f , its Lipschitz constant
must be Lip f .

For a Lipschitz map f = (f1 , . . . , fm ) : A ⊂ Rn → Rm , we may ap-
ply McShane’s lemma to each component of f , yielding a map F =
n
(F1 , . . . , F√
m) : R → Rm which extends f with Lipschitz constant
Lip F ≤ m Lip f . It is possible, however, to obtain an extension
which has the same Lipschitz constant as f :
Theorem 5.2 (Kirszbraun’s theorem). Let A ⊂ Rn and f : A →
Rm a Lipschitz map. Then there exists a Lipschitz extension f : Rn →
Rm of f such that Lip F = Lip f .
Proof. We refer the reader to [Mag12], page 69. 

5.1.2. Rademacher’s theorem. We prove in this subsection that

every Lipschitz function on Rn is differentiable in the sense of Fréchet
Ln -a.e. on Rn . Besides, its a.e. defined partial derivatives coincide
with its weak partial derivatives, introduced below.
If Ω is an open subset of Rn and X ∈ C1c (Ω, Rn ), then a direct
application of the Fundamental Theorem of Calculus combined with
Fubini-Tonelli’s theorem yields
ˆ
div X dLn = 0.
Ω
 5.1. LIPSCHITZ MAPS ON RN 123

If u ∈ C1 (Ω) and ϕ ∈ C1c (Ω, Rn ), the previous equality applied to

X = uϕ yields the elementary Gauss-Green’s identity in divergence
form:
ˆ ˆ
n
h∇u, ϕi dL = − u div ϕ dLn .
Ω Ω

That motivates, for less regular u, let us say u ∈ L1loc (Ln |Ω ), the
introduction of the distributional gradient of u (that is, the gradient
of u in the sense of the theory of Schwartz distributions) as the linear
functional ∇u : C∞ n
c (Ω, R ) → R given by the second member in the
previous equality, i.e.
ˆ
∇u · ϕ := − u div ϕ dLn .
Ω

Similarly, for 1 ≤ i ≤ n, the distributional i-th partial derivative of u

∂u
is the linear functional ∂x i
: C∞
c (Ω) → R given by
ˆ
∂u ∂ϕ
h , ϕi := − u dLn .
∂xi Ω ∂x i

Whenever those linear functionals are representable as integration of ϕ

against an L1loc function on Ω, we say that u admits weak gradient or
weak partial derivatives:
Definition 5.3 (weak derivatives and gradients). Let Ω be an open
subset of Rn and u ∈ L1loc (Ln |Ω ). We say that:
i) For 1 ≤ i ≤ n, u has weak i-th partial derivative vi ∈ L1loc (Ln |Ω ) if
∀ϕ ∈ C∞c (Ω),
ˆ ˆ
n ∂ϕ
vi ϕ dL = − u dLn .
Ω Ω ∂x i

ii) u has weak gradient v ∈ L1loc (Ln |Ω , Rn ) if ∀ϕ ∈ C∞ n

c (Ω, R ),
ˆ ˆ
(5.2) hv, ϕi dL = − u div ϕ dLn .
n
Ω Ω

We denote the weak derivatives by the same notations used for the
∂u
classical derivatives, i.e. ∂x i
for the i-th weak partial derivative and ∇u
for the weak gradient of u, if they exist; if distinction is needed, we use
w
“ ∂∂xui ” or “∇w u” for the weak partial derivatives and gradient.

Exercise 5.4 (weak gradients, bis). Weak gradients may be also

characterized by means of Gauss-Green identity in gradient form. That
124 5. AREA AND COAREA FORMULAS

is, let Ω be an open subset of Rn and u ∈ L1loc (Ln |Ω ); then u admits

weak gradient v ∈ L1loc (Ln |Ω , Rn ) iff ∀ϕ ∈ C∞
c (Ω),

ˆ ˆ
n
(5.3) ϕv dL = − u∇ϕ dLn .
Ω Ω

Exercise 5.5. Let Ω be an open subset of Rn , u ∈ L1loc (Ln |Ω ) and

w
1 ≤ i ≤ n. If there exists ∂∂xui ∈ L1loc (Ln |Ω ), then ∀ϕ ∈ C1c (Ω),

ˆ ˆ
∂wu ∂ϕ
ϕ dLn = − u dLn .
Ω ∂xi Ω ∂xi

Proposition 5.6. Let Ω be an open subset of Rn and u ∈ L1loc (Ln |Ω ).

i) If the weak partial derivatives or weak gradient of u exist, they are
unique up to Ln -null sets.
ii) u has weak gradient v = (v1 , . . . , vn ) ∈ L1loc (Ln |Ω , Rn ) iff ∀1 ≤ i ≤
n, u has i-th weak partial derivative vi ∈ L1loc (Ln |Ω ).

Proof. Part i) follows from the fundamental lemma of the calculus

of variations 4.34 and part ii) is immediate from exercise 5.4. 

It is clear that, if u ∈ C1 (Ω), the classical and weak gradients of u

coincide. The converse holds in the following sense: if u ∈ L1loc (Ln |Ω )
has weak gradient v ∈ C(Ω, Rn ), then u ∈ C1 (Ω). We postpone the
proof of this fact to exercise 6.22 in chapter 6.

Proposition 5.7 (vanishing weak gradient). Let Ω ⊂´Rn be a con-

nected open set and u ∈ L1loc (Ln |Ω ) such that ∀ϕ ∈ C∞ n
c (Ω), Ω u∇ϕ dL =
0. Then u coincides Ln -a.e. with a constant function.

Proof. 1) For each  > 0, let Ω := {x ∈ Rn | B(x, ) ⊂ Ω} =

{x ∈ Rn | d(x, Ωc ) > }, so that (Ω )>0 is a family of open subsets
of Ω which increases to Ω as  ↓ 0.
Let (φt )t>0 be the standard mollifier in Rn . Given  > 0, let
u : Rn → R be given by u = u on Ω/2 and u = 0 on Rn \
Ω/2 . Since Ω/2 ⊂ Ω, it´follows that u´ ∈ L1loc (Ln ) (because, for
each compact K ⊂ Rn , K |u | dLn = K∩Ω/2 |u| dLn < ∞, since
K ∩ Ω/2 is a compact subset of Ω). Take g := φ/2 ∗ u . Then, by
proposition 1.108, g ∈ C∞ (Rn ) and ∇g = (∇φ/2 ) ∗ u . Therefore,
 5.1. LIPSCHITZ MAPS ON RN 125

since spt φ/2 ⊂ B/2 , we have, ∀x ∈ Ω :

ˆ
∇g (x) = ∇φ/2 (x − y)u (y) dLn (y) =
ˆ
= − ∇[φ/2 (x − ·)]u dLn =
ˆ
x+B/2 ⊂Ω/2
=− ∇[φ/2 (x − ·)]u dLn =
x+B/2
ˆ
=− ∇[φ/2 (x − ·)]u dLn = 0,
Ω

where the last equality is justified by the fact that spt φ/2 (x − ·) ⊂
x + B/2 ⊂ Ω, so that φ/2 (x − ·) ∈ C∞ c (Ω). That is, ∇g ≡ 0 in
the open set Ω ; by elementary Calculus, it then follows that g is a
constant function in each connected component of Ω .
2) We contend that g is convergent to u in L1loc (Ln |Ω ) as  → 0, i.e.
for each compact K ⊂ Ω, kg − ukL1 (Ln |K ) → 0. Indeed, given
K ⊂ Ω compact, let 0 := 12 d(K, Ωc ). Then, ∀ 0 <  < 0 , both
g´ = φ/2 ∗ u and φ/2 ∗ u0 coincide in each x ∈ Ω0 ⊃ K with
φ (x − y)u(y) dLn (y). By exercise 1.115,
x+B/2 /2

→0
φ/2 ∗ u0 → u0
in L1loc (Ln ); therefore, we conclude that g |K → u0 |K = u|K in
L1 (Ln |K ), thus proving our contention.
3) Let (n )n∈N be a sequence in (0, ∞) with n ↓ 0. It follows from the
contention in the previous item that (gn := gn )n∈N is convergent
to u in L1loc (Ln |Ω ); therefore, for each compact K ⊂ Ω, there exists
a subsequence of (gn )n∈N which converges Ln -a.e. on K to u|K .
Since Ω is σ-compact, we may take a sequence of compact subsets
which increases to Ω and apply a diagonal argument to obtain a
subsequence of (gn )n∈N which converges Ln -a.e. on Ω to u. We
denote such subsequence with the same notation (gn )n∈N .
4) Let B be an arbitrary open ball with B b Ω. Since B is compact
and (Ωn )n∈N increases to Ω as n ↑ ∞, there exists N ∈ N such
that B ⊂ Ωn for all n ≥ N . Since B is connected, it follows from
part 1) that gn is constant on B for every n ≥ N . We then conclude
from part 3) that (gn )n∈N is pointwise convergent on B to a constant
function. By the arbitrariness of the open ball B b Ω, it follows
that (gn )n∈N converges pointwise on Ω to a locally constant function
g : Ω → R; but, since Ω is connected, g must be a constant function.
As (gn )n∈N converges pointwise to g and converges pointwise Ln -a.e.
126 5. AREA AND COAREA FORMULAS

to u, it finally follows that u = g Ln -a.e., i.e. u coincides Ln -a.e.

with a constant function, as we wanted to show.

Definition 5.8 (Sobolev spaces and functions). Let Ω be an open
subset of Rn , u : Ω → R and 1 ≤ p ≤ ∞. We say that
i) u is a (1, p)-Sobolev function if u ∈ Lp (Ln |Ω ) and, ∀1 ≤ i ≤ n, u
∂u
has weak partial derivatives ∂x i
∈ Lp (Ln |Ω ). We use the notation
W1,p (Ω) to denote the space of (1, p)-Sobolev functions on Ω.
ii) u is a local (1, p)-Sobolev function if u ∈ Lploc (Ln |Ω ) and, ∀1 ≤
i ≤ n, u has weak partial derivatives ∂x ∂u
i
∈ Lploc (Ln |Ω ). We use
1,p
the notation Wloc (Ω) to denote the space of local (1, p)-Sobolev
functions on Ω.
1,p
It is immediate from the definitions that Wloc (Ω) and W1,p (Ω) are
Ω
linear subspaces of R , and the weak partial derivatives and weak gra-
dient are linear on these spaces. We further develop the basic theory of
weak derivatives and Sobolev spaces in chapter 6. For the moment, we
1,∞
prove that Lipschitz functions on Rn belong to Wloc (Rn ), but firstly
we introduce some notation.
Let u : Rn → R and τ ∈ Sn−1 . For h ∈ R \ {0}, we denote by
τh u : Rn → R the incremental ratio of u in the direction τ :
u(x + hτ ) − u(x)
τh u(x) := .
h
Note that, by the invariance of the Lebesgue measure under trans-
lations, if u ∈ L1loc (Rn ), v : Rn → R bounded Ln -measurable with
compact support and h ∈ R \ {0}:
ˆ ˆ
u(x + hτ )v(x) dL (x) = u(x)v(x − hτ ) dLn (x),
n

hence
ˆ ˆ
n
(5.4) τh u(x)v(x) dL (x) = − u(x)τ−h v(x) dLn (x).

Proposition 5.9. Let f : Rn → R be a Lipschitz function. Then

1,∞
f ∈ Wloc (Rn ).
Proof. It is clear that f ∈ L∞ n
loc (L ). We show that f has weak
gradient in L∞ (Ln , Rn ).
Let τ ∈ Sn−1 and (hk )k∈N a sequence in (0, ∞) convergent to 0.
Then ∀k ∈ N, kτhk f k∞ ≤ Lip f . Since L1 (Ln ) is a separable Banach
∗
space with L1 (Ln ) ≡ L∞ (Ln ) (in view of Riesz representation theorem
1.79), it follows from Banach-Alaoglu theorem that the closed balls in
 5.1. LIPSCHITZ MAPS ON RN 127

L∞ (Ln ) are compact and metrizable in the weak-star topology. Hence,

passing to a subsequence, if necessary, we may assume that there exists
∗
gτ ∈ L∞ (Ln ) such that τhk f * gτ , i.e. for all v ∈ L1 (Ln ),
ˆ ˆ
n k→∞
(5.5) v τhk f dL → vgτ dLn .

Note that, ∀ϕ ∈ C∞ n
c (R ), ∀k ∈ N, τ−hk ϕ has support in the compact
set K := spt ϕ + B(0, supk∈N hk ). Thus, ∀x ∈ Rn , f (x)τ−hk ϕ(x) →
f (x)∇ϕ(x) · τ and the convergence is dominated, since, by the mean
value inequality, ∀k ∈ N, |f τ−hk ϕ| ≤ k∇ϕk∞ |f | χK ∈ L1 (Ln ). That
justifies the application of the dominated convergence theorem in the
last equality below, ∀ϕ ∈ C∞ n
c (R ):
ˆ ˆ
n (5.5) (5.4)
gτ ϕ dL = lim τhk f ϕ dLn =
k→∞
ˆ
1.64
= − lim f τ−hk ϕ dLn =
k→∞
ˆ
= − f ∇ϕ · τ dLn .

Taking τ = ei , 1 ≤ i ≤ n, we conclude that f has weak partial deriva-

tives gei ∈ L∞ (Ln ). 
Let U ⊂ Rn open and f : U → R. Recall that f is differentiable at
x0 ∈ U in the sense of Fréchet if there exists A ∈ L(Rn , R) such that
f (x0 + h) − f (x0 ) − A · h
lim = 0.
h→0 khk
If that is the case, f has first order partial derivatives at x0 , A satisfying
the above condition is unique and coincides with h∇f (x0 ), ·i : Rn → R;
A is called Fréchet derivative of f at x0 and denoted by Df (x0 ).
Equivalently, f is differentiable at x0 if it satisfies the condition
stated in the exercise below. We will use the following two exercises in
the proof of Rademacher’s theorem.
Exercise 5.10 (characterization of Fréchet differentiability). Let
U ⊂ Rn open and f : U → R. Then f is differentiable at x0 ∈ U
iff there exists A ∈ L(Rn , R) and there exists r > 0 such that
f (x0 + tv) − f (x0 )
lim+ =A·v
t→0 t
uniformly in v ∈ rSn−1 . If so,
• the above condition holds for all r > 0 (i.e. if it holds for some
r > 0, then it holds for all r > 0);
128 5. AREA AND COAREA FORMULAS

• A = Df (x0 ).

Exercise 5.11 (weak gradients under scaling and translations).

1,1
Let x ∈ Rn , h > 0, T : Rn → Rn given by τ 7→ x+hτ and u ∈ Wloc (Rn ).
1,1
Then u ◦ T ∈ Wloc (Rn ) and ∇w (u ◦ T )(τ ) = h ∇w u(x + hτ ).

Theorem 5.12 (Rademacher’s theorem). Let f : Rn → R be Lip-

schitz. Then f is differentiable in the sense of Fréchet Ln -a.e. and
∇f = ∇w f Ln -a.e.
1,∞
Proof. Recall that f ∈ Wloc , by proposition 5.9, so that ∇w f ∈
∞ n n 1 n n
Lloc (L , R ) ⊂ Lloc (L , R ). Furthermore, by corollary 3.31 applied to
each component of ∇w f ∈ L1loc (Ln , Rn ), Ln -almost every x ∈ Rn is a
Lebesgue point of ∇w f ; fix such a Lebesgue point x ∈ Rn . We will
show that f is differentiable at x and ∇f (x) = ∇w f (x).
For each h > 0, let gh : Rn → R be given by
f (x + hτ ) − f (x)
gh (τ ) := .
h
By exercise 5.10, the thesis follows once we show that gh (τ ) converges
to ∇w f (x) · τ uniformly with respect to τ on Sn−1 .
Note that ∀h > 0, gh is Lipschitz with Lip gh ≤ Lip f and gh (0) = 0.
Let (hk )k∈N be a sequence in (0, ∞) convergent to 0. We have:
1,∞
1) By proposition 5.9, gh ∈ Wloc (Rn ). Besides, by the linearity of the
weak gradient and exercise 5.11, ∀τ ∈ Rn :

∇w gh (τ ) = ∇w f (x + hτ ).

Hence, the fact that x is a Lebesgue point of ∇w f implies that

ˆ ˆ
w w y=x+hτ
|∇ gh (τ ) − ∇ f (x)| dτ = |∇w f (x + hτ ) − ∇w f (x)| dτ =
ˆ
U(0,1) U(0,1)
1 h→0
= n |∇w f (y) − ∇w f (x)| dy → 0,
h U(x,h)

i.e. ∇w gh converges to the constant function ∇w f (x) in L1 (Ln |U(0,1) )

as h → 0.
2) (gh )h>0 is equicontinuous and pointwise bounded. It then follows
from the Arzelà-Ascoli theorem that there exists g : Rn → R and a
subsequence (hkj )j∈N of (hk )k∈N such that gj := ghkj → g uniformly
on compact subsets of Rn . In particular, g is Lipschitz with Lip g ≤
Lip f and g(0) = 0.
 5.1. LIPSCHITZ MAPS ON RN 129

3) ∀ϕ ∈ C∞
c (U(0, 1)),
ˆ ˆ
∇ g ϕ dL = − g ∇ϕ dLn =
w n

ˆ
= − lim gj ∇ϕ dLn =
j→∞
ˆ
= lim ∇w gj ϕ dLn =
j→∞
ˆ
= ∇w f (x) ϕ(y) dy.

Thus, from the fundamental lemma of the Calculus of Variations

4.34, it follows that ∇w g(y) = ∇w f (x) for Ln -a.e. y ∈ U(0, 1).
4) Define g0 : Rn → R by g0 (τ ) := g(τ ) − ∇w f (x) · τ . Then g0 is
Lipschitz and, by the previous item, ∇w g0 = 0 on U(0, 1). Since
U(0, 1) is connected, it follows from proposition 5.7 that g0 coincides
Ln -a.e. on U(0, 1) with a constant function. As g0 is continuous
and g0 (0) = 0, we conclude that g0 is identically null on U(0, 1)
and, by continuity, identically null on B(0, 1). Thus, ∀τ ∈ B(0, 1),
g(τ ) = ∇w f (x) · τ . Hence, gj (τ ) → ∇w f (x) · τ uniformly with
respect to τ ∈ B(0, 1). Since the sequence (hk )k∈N convergent to
0 was arbitrarily taken, we have shown that every such sequence
admits a subsequence (hkj )j∈N such that gj = ghkj converges to g in


the metric space C(Bn ), k·ku , which implies that limh→0 gh = g in
the same metric space. In particular,
f (x + hτ ) − f (x)
lim+ = ∇w f (x) · τ
h→0 h
uniformly with respect to τ ∈ Sn−1 ⊂ Bn .
By exercise 5.10, it follows that f is differentiable at x and ∇f (x) =
∇w f (x), as we wanted to show. 
Exercise 5.13. Let f : Rn → R be Lipschitz. The set Df of points
where f is differentiable in the sense of Fréchet is Borel measurable and
Df : Df → L(Rn , R) is Borelian.
Corollary 5.14. If Ω ⊂ Rn open and f : Ω → R is locally Lips-
chitz, then f is Ln |Ω -a.e. differentiable in the sense of Fréchet.
Proof. We may cover Ω with a countable family (Uk )k∈N of open
subsets of Ω such that ∀k ∈ N, f |Uk is Lipschitz. For each k ∈ N,
we may extend f |Uk to a Lipschitz function fk : Rn → R, which is
differentiable Ln -a.e. on Rn in view of Rademacher’s theorem. As
differentiability is a local notion, we conclude that f |Uk is differentiable
130 5. AREA AND COAREA FORMULAS

on the complement of a Ln -null set Sk ⊂ Uk . Then f is differentiable

on the complement of the Ln -null set S = ∪k∈N Sk . 
Remark 5.15. We postpone to 6.16 in chapter 6, after we prove
the locality of the weak derivative, the proof that that f : Ω → R
locally Lipschitz has weak gradient ∇w f ∈ L∞ n
loc (L |Ω ), which coincides
Ln |Ω -a.e. with ∇f .
Corollary 5.16. If Ω ⊂ Rn open and f : Ω → Rm is locally
Lipschitz, then f is Ln |Ω -a.e. differentiable in the sense of Fréchet.
Proof. Apply the previous corollary to each component of f . 
Corollary 5.17.
i) Let f : Rn → Rm be locally Lipschitz and Zf := {x ∈ Rn | f (x) =
0}. Then Df (x) = 0 for Ln -a.e. x ∈ Zf .
ii) Let f, g
: Rn → Rn be locally

Lipschitz and Y :=n {x ∈ R |
n

g f (x) = x}. Then Dg f (x) ◦ Df (x) = idRn for L -a.e. x ∈ Y .

Proof.
1) It suffices to prove part i) for m = 1 (in the general case, we argue
componentwise).
2) Note that Zf ∈ B Rn . Let x ∈ Zf such that ∃Df (x) and


Ln Z ∩ B(x, r)
lim
= 1.
r→0 Ln B(x, r)
In view of Rademacher’s theorem 5.16 and of theorem 3.29 (with
Ln in place of µ and Zf in place of A), Ln -a.e. x ∈ Zf satisfies the
above conditions. Therefore, part i) will be proved once we show
that ∇f (x) = 0.
Suppose that ∇f (x) = a ∈ Rn \ {0}. Define S := {v ∈ Sn−1 |
ha, vi > 21 kak}. Note that S is an open neighborhood of a/kak in
Sn−1 . For each r > 0, we define Sr := {tv | 0 < t ≤ r, v ∈ S} ⊂
B(0, r), so that Sr = rS1 .
By exercise 5.10,
f (x + tv) − f (x)
lim = ha, vi
t→0 t
uniformly on v ∈ Sn−1 . It then follows, by the definition of S, that
there exists R > 0 such that, ∀0 < t < R and ∀ v ∈ S,
f (x + tv) f (x + tv) − f (x) 1
= > kak > 0.
t t 2
 5.1. LIPSCHITZ MAPS ON RN 131

In particular, ∀0 < r < R, f > 0 on x + Sr , i.e. Zf ∩ B(x, r) ⊂

B(x, r) \ (x + Sr ). Consequently, ∀0 < r < R,



Ln Zf ∩ B(x, r) Ln B(x, r) \ (x + Sr ) 1.4

≤
=
Ln B(x, r) Ln B(x, r)
Ln (S1 )
=1− ,
α(n)
whence

Ln Zf ∩ B(x, r) Ln (S1 )
lim sup
≤1− .
r→0 Ln B(x, r) α(n)
n
In view of our choice of x, the latter inequality implies 1− Lα(n)
(S1 )
≥ 1,
n
hence L (S1 ) = 0. As S1 has nonempty interior, we have reached a
contradiction, thus showing that ∇f (x) 6= 0 cannot occur.
3) To prove part ii), let F := g ◦ f − idRn . Then F is locally Lipschitz
and Y = ZF ; it then follows from part i) that D(g ◦ f )(x) − idRn =
DF (x) = 0 for Ln -a.e. x ∈ Y . Therefore,
part ii) will be proved
n
once we show that D(g◦f )(x) = Dg f (x) ◦Dg(x) for L -a.e. x ∈ Y .
Let Df := {x ∈ Rn | ∃Df (x)}, Dg := {x ∈ Rn | ∃Dg(x)}, and
X := Y ∩ Df ∩ f −1 (Dg ). Then Y \ X = (Y \ Df ) ∪ Y \ f −1 (D
g ) .

If x ∈ Y \ f −1 (Dg ), then f (x) ∈ Rn \ Dg , hence x = g f (x) ∈

g(Rn \ Dg ). Therefore, Y \ f −1 (Dg ) ⊂ g(Rn \ Dg ), so that
Y \ X ⊂ (Rn \ Df ) ∪ g(Rn \ Dg ).
Since both Rn \Df and Rn \Dg are Ln -null sets (in view of Rademacher’s
theorem 5.17), and since the image of a Ln -null set by a locally Lip-
schitz map is Ln -null, it follows that Y
\ X is Ln -null. On the other
hand, ∀x ∈ X, ∃Df (x) and ∃Dg
f (x) , hence the chain rule ensures
that ∃D(g ◦ f )(x) = Dg f (x) ◦ Dg(x).

5.1.3. Linear maps and Jacobians. In this subsection we recall
some linear algebra and introduce pertinent notations that will be used
in the two main theorems which name this chapter.
Definition 5.18. Let V and W be finite-dimensional Hilbert spaces.
i) A linear map O : V → W is called an orthogonal injection if ∀x, y ∈
V, hO · x, O · yi = hx, yi. We denote the set of orthogonal injections
V → W by O(V, W); we abbreviate O(n, m):= O(Rn , Rm ) and
O(n) := O(n, n).
ii) Let T : V → W be a linear map. We denote by T ∗ the adjoint of
T with respect to the inner products on V and W, i.e. the unique
132 5. AREA AND COAREA FORMULAS

linear map such that ∀x ∈ V, ∀y ∈ W, hx, T ∗ · yi = hT · x, yi. If

V = W and T = T ∗ , we call T self-adjoint or symmetric. We
denote by Sym(V) the set of symmetric linear maps in L(V); we
abbreviate Sym(n):= Sym(Rn ).
iii) We say that a linear map T : V → V is positive if it is symmetric
and ∀x ∈ V, hT · x, xi ≥ 0.
Note that O(V, W) = ∅ if dim V > dim W.
Recall that, for any symmetric linear map T on a finite-dimensional
Hilbert space V, there exists an orthonormal basis of V formed by eigen-
vectors of T . Equivalently, there exist unique c1 , . . . , ck ∈ R pairwise
distinct and unique P1 , . . . , Pk ∈ L(V) such that ∀1 ≤ i, j ≤ k, Pi = Pi∗ ,
Pi2 = Pi , Pi Pj = 0 if i 6= j, ki=1 Pi = idV and T = ki=1 ci Pi ; the ci ’s
P P
are the eigenvalues of T and the Pi ’s are the orthogonal projections on
the corresponding eigenspaces. The decomposition T = ki=1 ci Pi is
P
called the spectral resolution of T .
Theorem 5.19 (existence of square roots). If V is a finite-dimensional
Hilbert space and P ∈ L(V) is a positive operator, there exists a unique
positive operator N ∈ L(V) such that N 2 = P .
√
Notation. We denote N by P .
Proof. Let P = ki=1 ci Ei be the spectral resolution of P . The
P
P √
positiveness of P implies ci ≥ 0 for 1 ≤ i ≤ k. Define N := ki=1 ci Ei ;
then N is positive and N 2 = P , thus proving the existence. On the
other hand, suppose that M is another positive operator such that
M 2 = P . Let the spectral resolution of M be M = ji=1 di Fi . Then
P

P = M 2 = ji=1 d2i Fi . By the uniqueness of the spectral resolution

P
of P , it then follows that j = k and, reordering the di ’s if necessary,
ci = d2i for 1 ≤ i ≤ k, thus proving the uniqueness. 
Theorem 5.20 (polar decomposition). Let V and W be finite-dimensional
Hilbert spaces and L : V → W be a linear map.
i) If dim V ≤ dim W, there exists a positive S ∈ Sym(V) and O ∈
O(V, W) such that
L = O ◦ S.
Moreover, in the above decomposition, S ∈ Sym(V) positive is
unique, and so is O ∈ O(V, W) if L is injective.
ii) If dim V ≥ dim W, there exists a positive S ∈ Sym(W) and O ∈
O(W, V) such that
L = S ◦ O∗ .
Moreover, in the above decomposition, S ∈ Sym(W) positive is
unique, and so is O ∈ O(W, V) if L is surjective.
 5.1. LIPSCHITZ MAPS ON RN 133

Proof. Part ii) follows from part i) applied to L∗ : W → V, so it

is enough to prove part i).
1) (uniqueness) Suppose that there exists S ∈ Sym(V) and O ∈ O(V, W)
such that L = O ◦ S. Then, since O∗ O = idV , it follows that
L∗ L = S 2 . As L∗ L is positive, we conclude that S is the (unique)
positive square root of L∗ L given by theorem 5.19. Moreover, if L is
−1
injective, so is S, hence
√ S is invertible and we must have O = L◦S .
2) (existence) Let S := L∗ L. For each v ∈ V, we must have O ·S ·v =
L·v. Thus, define O on the range of S by O·w := L·v if v ∈ V is such
that S · v = w. If v 0 ∈ V is another vector such that S · v 0 = w, we
must have kL·(v−v 0 )k2 = hL∗ L·(v−v 0 ), v−v 0 i = hS 2 ·(v−v 0 ), v−v 0 i =
0, hence L·v = L·v 0 , which shows that O is well-defined on the range
of S. Besides, it is clearly linear and satisfies ∀v ∈ V, L · v = O · S · v.
If w, w0 ∈ Im S and v, v 0 ∈ V are such that S · v = w, S · v 0 = w0 , we
have:
hO · w, O · w0 i = hL · v, L · v 0 i =
= hL∗ L · v, v 0 i = hS 2 · v, v 0 i =
= hS · v, S · v 0 i = hw, w0 i,
hence O : Im S → W is orthogonal. Finally, since dim V ≤ dim W,
we have dim(Im S)⊥ ≤ dim(O · Im S)⊥ , hence we may extend O to
an orthogonal injection on V (take any orthonormal set in (Im S)⊥
and map it to an orthonormal set on (O · Im S)⊥ ), thus yielding
O ∈ O(V, W) such that O ◦ S = L.

Definition 5.21 (Jacobian of a linear map). Let V and W be finite-
dimensional Hilbert spaces and L ∈ L(V, W), with polar decomposition
O ◦ S if dim V ≤ dim W or S ◦ O∗ if dim V > dim W, cf. theorem 5.20.
We define the Jacobian JLK of L by:
JLK := |det S|.
Remark 5.22.
1) Note that JLK is well-defined, by the uniqueness of S in the polar
decomposition.
2) It is clear that
(√
∗ det L∗ L if dim V ≤ dim W
JLK = JL K = √ ∗
det LL if dim V ≥ dim W.
The next theorem provides a useful formula for computing the Ja-
cobian of a linear map L : V → W between finite-dimensional Hilbert
spaces.
134 5. AREA AND COAREA FORMULAS

Theorem 5.23 (Binet-Cauchy formula). Let V and W be finite-

dimensional Hilbert spaces with n = dim V ≤ dim W = m. If L ∈
L(V, W), then

s X
JLK = (det B)2 ,
B∈µ(m,n)

where µ(m, n) is the set of n × n minors in some matrix representation

of L with respect to orthonormal bases on V and W.

Choosing orthonormal bases on V and W, we identify V ≡ Rn and

W ≡ Rm . We will use the following notation:

Notation. Let n ≤ m.

1) We denote by:
• Φ(m, n) the set of all maps {1, . . . , n} → {1, . . . , m}.
• Σ(m, n) := {λ ∈ Φ(m, n) | λ 1-1}. We abbreviate Σn :=
Σ(n, n) (i.e the set of permutaions of {1, . . . , n}).
• Λ(m, n):= {λ ∈ Σ(m, n) | λ strictly increasing}.
2) For λ ∈ Λ(m, n), let Sλ := heλ(i) | 1 ≤ i ≤ ni ⊂ Rm and Pλ ∈
L(Rm , Sλ ) the orthogonal projection onto Sλ , i.e. Pλ (x1 , . . . , xm ) :=
(xλ(1) , . . . , xλ(n) ).

Proof. With the above notation in force, we must prove that

X
JLK2 = (det Pλ ◦ L)2 .
λ∈Λ(m,n)

Let (Lij )m×n be the matrix of L with respect to the standard bases
of Rn and Rm . Then the matrix (Aij )n×n of A := L∗ L ∈ L(Rn ) with
 5.1. LIPSCHITZ MAPS ON RN 135

respect to the standard basis is given by Aij = m

P
k=1 Lki Lkj . Therefore:
X Yn X Yn X m
JLK2 = det A = sgn (σ) aiσ(i) = sgn (σ) Lki Lkσ(i) =
σ∈Σn i=1 σ∈Σn i=1 k=1
n
X X Y (∗)
= sgn (σ) Lϕ(i)i Lϕ(i)σ(i) =
σ∈Σn ϕ∈Φ(m,n) i=1
n
X X Y ˙ λ∈Λ(m,n) ∪θ∈Σn λ◦θ
Σ(m,n)=∪
= sgn (σ) Lϕ(i)i Lϕ(i)σ(i) =
σ∈Σn ϕ∈Σ(m,n) i=1
X X n
XY
= sgn (σ) Lλ◦θ(i),i Lλ◦θ(i),σ(i) =
σ∈Σn λ∈Λ(m,n) θ∈Σn i=1
X X X Y
= sgn (σ) Lλ◦θ(i),i Lλ◦θ(i),σ(i) =
σ∈Σn λ∈Λ(m,n) θ∈Σn {i=θ−1 (j)|1≤j≤n}

X X n
XY
= sgn (σ) Lλ(j),θ−1 (j) Lλ(j),σ◦θ−1 (j) =
σ∈Σn λ∈Λ(m,n) θ∈Σn j=1
X X X n
Y
= sgn (σ) Lλ(i),θ(i) Lλ(i),σ◦θ(i) =
λ∈Λ(m,n) θ∈Σn σ∈Σn i=1
n
X X X Y sgn (σ)=sgn (ρ)·sgn (θ)
= sgn (σ) Lλ(i),θ(i) Lλ(i),σ◦θ(i) =
λ∈Λ(m,n) θ∈Σn {σ=ρ◦θ−1 |ρ∈Σn } i=1

X X X n
Y
= sgn (ρ) · sgn (θ) Lλ(i),θ(i) Lλ(i),ρ(i) =
λ∈Λ(m,n) θ∈Σn ρ∈Σn i=1
n
X X X
2
= sgn (θ) Lλ(i),θ(i) =
λ∈Λ(m,n) θ∈Σn i=1
X
= (det Pλ ◦ L)2 ,
λ∈Λ(m,n)

where the equality (∗) is justified by the fact that, if ϕ ∈ Φ(m, n) is

not injective, then
X Y n
sgn (σ) Lϕ(i)i Lϕ(i)σ(i) = 0.
σ∈Σn i=1


Remark 5.24. Theorem 5.23 may also be obtained as a corollary
of the Pythagorean theorem. Indeed, with the notation preceding the
136 5. AREA AND COAREA FORMULAS

above proof in force, let ∀1 ≤ i ≤ n, vi := L · ei = m m

P
k=1 Lki ek ∈ R .
Consider the n-vector
X n
^
(5.6) v = v1 ∧ · · · ∧ vn = (det Pλ ◦ L)eλ ∈ Rm ,
λ∈Λ(m,n)

where eλ := eλ(1) ∧ · · · ∧ eλ(n) ∈ n Rm .

V

Vn The Euclidean inner product on Rm induces an inner product on

Rm for which {eλ | λ ∈ Λ(m, n)} is an orthonormal basis (cf.
[Fed69], page 32, or [dL65], page 113). Vn mFor decomposable n-vectors
w = w1 ∧ · · · wn , z = z1 ∧ · · · ∧ zn ∈ R , we have


hw, zi = det hwi , zj i 1≤i,j≤n .

Therefore, computing kvk2 by the Pythagorean theorem:

JLK2 = det L∗ L = det hL∗ L · ei , ej i 1≤i,j≤n =




= det hL · ei , L · ej i 1≤i,j≤n =
Pythagoras+(5.6)
= hv, vi =
X
= (det Pλ ◦ L)2 .
λ∈Λ(m,n)

Definition 5.25 (Jacobian of Lipschitz maps). Let f : Rn → Rm

be Lipschitz. It follows from Rademacher’s theorem 5.12 (applied com-
ponentwise) and from exercise 5.13 that f is differentiable in the com-
plement of a Borel set of Ln -null measure and x 7→ Df (x) is Borelian
Ln -a.e. defined.
We define, for each point x where f is differentiable, the Jacobian
of f at x,
Jf (x) := JDf (x)K,
so that Jf is a Borelian function defined on the complement of a Borel
subset of Rn of Ln -null measure.
Exercise 5.26. With the notation above, check that Jf is indeed
Borelian.
Notation. For a Lipschitz map f : Rn → Rm , we will use hence-
forth the following notation:
• Df := {x ∈ Rn | ∃Df (x)};
• Jf+ := {x ∈ Df | Jf (x) > 0};
• Jf0 := {x ∈ Df | Jf (x) = 0}.
 5.2. THE AREA FORMULA 137

5.2. The area formula

In this section we assume n ≤ m.
Lemma 5.27 (Area Formula, linear case). If L : Rn → Rm is linear
and n ≤ m, then ∀A ⊂ Rn ,
Hn L(A) = JLKLn (A).


(5.7)
Proof. Let L = O ◦ S be a polar decomposition of L, cf. theorem
5.20, where S ∈ Sym(n) positive and O ∈ O(n, m). Then JLK = |det S|.
We have:
1) If JLK = 0, then det S = 0, so that dim Im L = dim Im S ≤ n −
1. It then follows
from exercise 2.23 that H-dim
Im L ≤ n − 1,
n n n n
hence H L(R ) = 0, whence ∀A ⊂ R , H L(A) = 0 and both
members of (5.7) are zero.
2) If JLK > 0, then det S > 0 and O : Rn → Rm is a linear isometry
into Rn . It then follows from corollary 2.5 that, for each closed ball
B(x, r) ⊂ Rn :

 

n n
H L B(x, r) H O ◦ S B(x, r)

=
=
Ln B(x, r) Ln B(x, r)


n
H S B(x, r) 2.21
(5.8) =
=
Ln B(x, r)


n
L S B(x, r) 1.81.i)
=
=
Ln B(x, r)
= |det S| = JLK.


Define ∀A ⊂ Rn , ν(A) := Hn L(A) . We contend that ν is a
Radon measure on Rn and ν  Ln . Indeed,
• L : Rn → Rm is a linear isomorphism onto Im L. In partic-
ular, L : Rn → Im L is a homeomorphism (endowing Im L
with the relative topology), hence the pushforward operation
L# defines a bijection between Borel measures on Rn and Borel
measures on Im L, with inverse L−1 # ; moreover, it is clear that
this bijection restricts to a bijection between Borel regular mea-
sures. Since Hn |Im L is a Borel regular measure (which can be
checked directly in view of the Borel regularity of Hn on Rm ,
or from the fact that the trace Hn |Im L coincides with the m-
dimensional Hausdorff measure of Im L as a metric subspace of
Rm , by proposition 2.4.1), it follows that ν = L−1 # (Hn |Im L ) is
138 5. AREA AND COAREA FORMULAS

a Borel regular measure on Rn . Besides, ∀K ⊂ Rn compact,

it follows from proposition 2.4.3) that ν(K) = Hn L(K) ≤
(Lip L)n Hn (K) = (Lip L)n Ln (K) < ∞. That is, ν is a locally
finite Borel regular measure on Rn , hence it is Radon by exer-
cise 1.32.
• If A ⊂ Rn is Ln -null,

then it follows from proposition 2.4.3) that
ν(A) = H L(A) ≤ (Lip L) H (A) = (Lip L)n Ln (A) = 0,
n n n

hence ν  Ln , thus proving our contention.

It follows from (5.8) that ∀x ∈ Rn , Θν (Ln , x) = JLK. Recall that,
from proposition 3.23, every Borel measure on Rn has the symmetric
Vitaly property, so that theorem 3.40 applies to Radon measures on
Rn , from which we conclude that, ∀A ∈ B Rn ,
ˆ
ν(A) = Θν (Ln , x) dLn (x) = JLKLn (A).
A
By Borel regularity,

both members must coincide for all A ⊂ Rn ,
i.e. H L(A) = JLKLn (A), as we wanted to show.
n


Exercise 5.28. Let T ∈ L(Rn , Rm ), n ≤ m.
a) If R ∈ L(Rn ), then JT ◦ RK = JT KJRK.
b) JT K ≤ kT kn . If T is 1-1, then kT −1 k−n ≤ JT K ≤ kT kn .
c) If m ≤ k and R ∈ L(Rm , Rk ), then JR ◦ T K ≤ kRkn JT K. If R is 1-1,
then kR−1 k−n JT K ≤ JR ◦ T K ≤ kRkn JT K.
Lemma 5.29. Let f : Rn → Rm be Lipschitz, with n ≤ m, and
A ⊂ Rn Ln -measurable. Then:
i) f (A) is Hn -measurable.
ii) The function N (f |A ) : Rm → [0, ∞] given by y 7→ H0 (A ∩ f −1 {y})
n
´is H -measurable.
iii) Rm H0 (A ∩ f −1 {y}) dHn (y) ≤ (Lip f )n Ln (A).
Definition 5.30 (multiplicity function). With the notation from
the previous lemma, N (f |A ) : y 7→ H0 (A ∩ f −1 {y}) is called the mul-
tiplicity function of f |A .
Remark 5.31. Concerning part a) of the previous lemma, for con-
tinuous images of Borel sets we have the following theorem. If X is a
complete, separable metric space, Y a Hausdorff topological space, µ a
Borel measure on Y and f : X → Y continuous, then ∀A ∈ B X , f (A)
is µ-measurable — see [Fed69], paragraph 2.2.13. Actually, ∀A ∈ B X ,
f (A) is a Suslin set. This result is pertinent to the so-called descrip-
tive set theory, for which we refer the interested reader to [Sri98] or
[Mos09].
 5.2. THE AREA FORMULA 139

Proof.
i) Since Ln is σ-finite, we may take a sequence (Ak )k∈N in σ(Ln )
such that ∀k ∈ N, Ln (Ak ) < ∞ and ∪k∈N Ak = A. Then f (A) =
∪k∈N f (Ak ), so that f (A) ∈ σ(Hn ) once we show that ∀k ∈ N,
f (Ak ) ∈ σ(Hn ). It is therefore enough to prove the case in which
Ln (A) < ∞. Since Ln is a Radon measure and Rn is σ-compact,
we may take (by exercise 1.31) an increasing sequence (Ki )i∈N of
compact subsets of A such that Ln (Ki ) → Ln (A). Since A ∈
σ(Ln ) and Ln (A) < ∞, it follows that Ln (A \ Ki ) → 0, hence
Ln (A \ ∪i∈N Ki ) = 0. Therefore,

by proposition 2.4.3), we conclude
that H f (A \ ∪i∈N Ki ) ≤ (Lip f )n Ln (A \ ∪i∈N Ki ) = 0. Since
n

∀i ∈ N, f (Ki ) is compact, it follows that ∪i∈N f (Ki ) ∈ B Rm ⊂

σ(Hn ). As f (A) \ ∪i∈N f (Ki ) ⊂ f (A \ ∪i∈N Ki ), we conclude that
f (A) \ ∪i∈N f (Ki ) is Hn -null, i.e. f (A) is the union of a Borel set
with an Hn -null set, thus f (A) ∈ σ(Hn ).
ii) We may take a sequence (Fi )i∈N such that
• ∀i ∈ N, Fi = (Fji )j∈N is a disjoint family of Borel subsets of
Rn with ∀j ∈ N, diam Fji ≤ 1/i and ∪˙ j∈N Fji = Rn ;
• ∀i ∈ N, each Fji+1 is a subset of some Fki (so that each Fki is a
disjoint union of some of the terms of Fi+1 ).
Let (gi )i∈N be the sequence of functions Rm → [0, ∞] defined
by, ∀i ∈ N,
X
gi := χf (A∩Fji )
j∈N

(the idea is that, for each i ∈ N and y ∈ Rm , gi (y) is the number

of terms of Fi which intersect A ∩ f −1 {y}; intuitively, gi increases
pointwise to the multiplicity function). The thesis then follows
once we show that each gi is Hn -measurable and (gi )i∈N increases
pointwise to the multiplicity function N (f |A ) (which implies the
Hn -measurability of the latter function in view of theorem 1.41.iv).
That is done along the following steps:
1) ∀i, j ∈ N, A ∩ Fji ∈ σ(Ln ), hence χf (A∩Fji ) is Hn -measurable by
part i). Thus, ∀i ∈ N, gi := j∈N χf (A∩Fji ) is Hn -measurable by
P

theorem 1.41.iv).
2) (gi )i∈N is pointwise increasing. Indeed, ∀y ∈ Rm and i ∈ N,
for each j ∈ N such that A ∩ f −1 {y} cuts Fji , i.e. such that
χf (A∩Fji ) (y) = 1, the fact that Fji is a union of terms of Fi+1
implies the existence of k = ki (j) ∈ N such that Fki+1 ⊂ Fji
(thus ki (j) 6= ki (j 0 ) if j 6= j 0 , i.e. ki is 1-1) and A ∩ f −1 {y} cuts
Fki+1 , i.e. χf (A∩F i+1 ) (y) = 1. Then, defining Ni := {j ∈ N |
k
140 5. AREA AND COAREA FORMULAS

A ∩ f −1 {y} ∩ Fji 6= ∅}, we have

X X
gi (y) = χf (A∩Fji ) (y) = χf (A∩Fji ) (y) =
j∈N j∈Ni
X ki is 1-1
= χf (A∩F i+1 ) (y) ≤
ki (j)
j∈Ni
X
≤ χf (A∩F i+1 ) (y) = gi+1 (y).
j
j∈N

3) ∀i ∈ N, gi ≤ N (f |A ). Indeed, since Fi is a disjoint family, for

all y ∈ Rm , A ∩ f −1 {y} = ∪˙ j∈N A ∩ f −1 {y} ∩ Fji . As ∀j ∈
N, H0 (A ∩ f −1 {y} ∩ Fji ) ≥ χf (A∩Fji ) (y), it then follows that
N (f |A )(y) = H0 (A ∩ f −1 {y}) = j∈N H0 (A ∩ f −1 {y} ∩ Fji ) ≥
P
P
j∈N χf (A∩Fji ) (y) = gi (y).
4) ∀y ∈ Rm , ∀k ∈ N such that k ≤ N (f |A )(y), there exists i ∈
N such that gi (y) ≥ k. Indeed, since N (f |A )(y) = H0 (A ∩
f −1 {y}) ≥ k, we may choose k distinct points x1 , . . . , xk ∈
A ∩ f −1 {y}. Take i ∈ N such that kxp − xq k > 1/i for 1 ≤
p < q ≤ k. Since the terms of Fi are disjoint with diameters
≤ 1/i, it follows that ∀1 ≤ p ≤ k, xp belong to exactly one of
i
the terms of Fi , say Fj(p) , with p 7→ j(p) 1-1. Then gi (y) =
P P
j∈N χf (A∩Fji ) (y) ≥ 1≤p≤k χf (A∩Fj(p)
i ) (y) = k, as asserted.
iii) Let (gi )i∈N be the same sequence of functions R → [0, ∞] from the
previous item, so that ∀y ∈ Rm , gi (y) ↑ N (f |A )(y). It follows from
the monotone convergence theorem 1.62 that:
ˆ ˆ
n
N (f |A )(y) dH (y) = lim gi (y) dHn (y) =
Rm i→∞ Rm
X
2.4.3)
= lim Hn f (A ∩ Fji ) ≤
i→∞
j∈N
X
≤ lim inf (Lip f )n Ln (A ∩ Fji ) =
i→∞
j∈N
n n
= (Lip f ) L (A).

Definition 5.32. Let f : Rn → Rm be a Lipschitz map with n ≤ m
and t > 1. We say that (E, S) is a t-linearization for f if E ∈ B Rn
and S ∈ Sym(n) ∩ GL(Rn ) satisfy:
i) ∀x ∈ E, f is differentiable at x and Jf (x) > 0;
ii) ∀x, y ∈ E, t−1 kS · x − S · yk ≤ kf (x) − f (y)k ≤ tkS · x − S · yk;
 5.2. THE AREA FORMULA 141

iii) ∀x ∈ E, ∀v ∈ Rn , t−1 kS · vk ≤ kDf (x) · vk ≤ tkS · vk.

Proposition 5.33. Let f : Rn → Rm be a Lipschitz map with
n ≤ m, t > 1, E ∈ B Rn such that condition i) in definition 5.32 holds
and S ∈ Sym(n) ∩ GL(Rn ). Then (E, S) is a t-linearization for f
iff f |E is 1-1 with Lipschitz inverse and satisfies:
ii’) Lip f |E ◦ S −1 ≤ t and Lip S ◦ (f |E )−1 ≤ t;
iii’) ∀x ∈ E, kDf (x) ◦ S −1 k ≤ t and kS ◦ Df (x)−1 k ≤ t.
Proof. If (E, S) is a t-linearization for f , then:
1) f |E is 1-1 in view of the first inequality in ii);
2) f |E ◦ S −1 is Lipschitz, with Lip f |E ◦ S −1 ≤ t, in view of the second
inequality in ii) with S −1 (x0 ) in place of x and S −1 (y 0 ) in place of y;
3) S ◦ (f |E )−1 is Lipschitz, with Lip S ◦ (f |E )−1 ≤ t, in view of the first
inequality in ii) with f −1 (x0 ) in place of x and f −1 (y 0 ) in place of y;
4) similarly, the second inequality in iii) implies kDf (x) ◦ S −1 k ≤ t and
the first inequality in iii) implies kS ◦ Df (x)−1 k ≤ t;
5) since S −1 and −1 −1

S ◦ (f |E ) are both Lipschitz, so is (f |E ) = S ◦
−1
−1
S ◦ (f |E ) .
Thus we have proved that f |E is 1-1 with Lipschitz inverse and satisfies
conditions ii’) and iii’).
With a similar argument, if f |E is 1-1 with Lipschitz inverse, then
conditions ii’) and iii’) imply ii) and iii), respectively, in definition 5.32,
thus proving the converse implication. 
Corollary 5.34. Let f : Rn → Rm be a Lipschitz map with n ≤
m, t > 1 and (E, S) a t-linearization for f . Then ∀x ∈ E,
(5.9) t−n |det S| ≤ Jf (x) ≤ tn |det S|.
Proof.
5.28.a)
Jf (x) = JDf (x)K|det S −1 ||det S| =
= JDf (x) ◦ S −1 K|det S|.
Hence, from exercise 5.28.b) with Df (x) ◦ S −1 in place of T and from
proposition 5.33.iii’), the thesis follows. 
Theorem 5.35 (Lipschitz linearization, [Fed69]). Let f : Rn →
Rm be a Lipschitz map with n ≤ m, t > 1 and Jf+ = {x ∈ Rn |
∃Df (x) and Jf (x) > 0} (which is a Borel set, by exercises 5.13 and
5.26). Then there exists a countable disjoint family (Ek )k∈N in B Rn
such that Jf+ = ∪˙ k∈N Ek and, ∀k ∈ N, there exists Sk ∈ Sym(n) ∩
GL(Rn ) such that (Ek , Sk ) is a t-linearization for f .
142 5. AREA AND COAREA FORMULAS

Proof. Fix  > 0 such that t−1 +  < 1 < t − . Let S be a

countable dense subset of Sym(n) ∩ GL(Rn ) and G a countable dense
subset of Jf+ . For all S ∈ S, k ∈ N and c ∈ G, we define E(S, k, c) :=
1
B(c, 2k ) ∩ F (S, k), where F (S, k) ⊂ Jf+ is the set of all x ∈ Jf+ such
that1:
F1) ∀v ∈ Rn , (t−1 + )kS · vk ≤ kDf (x) · vk ≤ (t − )kS · vk;
F2) ∀v ∈ Rn such that kvk ≤ k −1 , kf (x + v) − f (x) − Df (x) · vk ≤
kS · vk.
Since Df is Borelian (by exercise 5.13), it is clear that F (S, k) ∈
B R , hence E(S, k, c) ∈ B Rn . Furthermore,
n

1) For all S ∈ S, k ∈ N and c ∈ G, ∀x, y ∈ E(S, k, c),

F 2) F 1)
kf (y) − f (x)k ≤ kDf (x) · (y − x)k + kS · (y − x)k ≤ tkS · (y − x)k,
F 2) F 1)
kf (y) − f (x)k ≥ kDf (x) · (y − x)k − kS · (y − x)k ≥ t−1 kS · (y − x)k.


Therefore, the condition ii) in definition 5.32 is satisfied for E(S, k, c), S .
Besides, in view of F1), condition iii)
in the same definition is triv-
ially satisfied, so that E(S, k, c), S is a t-linearization for f .
2) We contend that Jf+ is the union of the countable family {E(S, k, c) |
S ∈ S, k ∈ N, c ∈ G}. Once we prove this contention, we enumerate
this family as (Êk )k∈N and we take the disjoint sequence (Ek )k∈N
given by Ek := Êk \ ∪k−1
i=1 Êi , thus reaching the thesis in view of the
previous item.
To prove the contention, fix x ∈ Jf+ and let the polar decom-
position of Df (x) be Df (x) = Px ◦ Sx , with Px ∈ O(n, m) and
Sx ∈ Sym(n). Note that, since Df (x) is 1-1, so is Sx , i.e. Sx ∈
1 The idea is the following:
• We want to ensure ii) and iii) in definition 5.32. In order to ensure iii)
we might take F1) with t instead of t − ; however, with t −  it will work
as well and, as we shall see, we need a little “space” for the estimate in
the next step.
• To ensure ii), we need to use somehow the differentiability of f . Assume
that x, y ∈ E ⊂ Jf+ with diam (E) ≤ 1/k sufficiently small (to be chosen).
We then have
f (y) − f (x) = Df (x) · (y − x) + Rx (y − x).
Thus, in order to obtain the desired inequalities in ii) to kf (y)−f (x)k, say
the second one, we must control the norms of both terms in the second
member. As kDf (x) · (y − x)k ≤ (t − )kS · x − S · yk if F1) holds, we
must ensure that kRx (y − x)k ≤ kS · x − S · yk = kS · (y − x)k with 1/k
sufficiently small.
 5.2. THE AREA FORMULA 143

GL(Rn ). Moreover, since S is dense in Sym(n) ∩ GL(Rn ), we may

take a sequence (Si )i∈N in S convergent to Sx ; hence, by continuity,
Sx ◦ Si−1 → idRn and Si ◦ Sx−1 → idRn . Taking i sufficiently large, we
then conclude that there exists S ∈ S such that
kSx ◦ S −1 k ≤ t −  and kS ◦ Sx−1 k ≤ (t−1 + )−1 .
It then follows that, ∀v ∈ Rn ,
kDf (x) · vk = kPx · Sx · vk = kSx · vk = k(Sx S −1 )S · vk ≤
≤ kSx S −1 kkS · vk ≤ (t − )kS · vk,
(t−1 + )kS · vk = (t−1 + )k(SSx−1 )Sx · vk ≤
≤ (t−1 + )kSSx−1 kkSx · vk ≤ kSx · vk = kDf (x) · vk.
That is, F1) is satisfied. Moreover, by the differentiability of f at
x, there exists R(x, ·) : Rn → [0, ∞) continuous and null at v = 0,
such that ∀v ∈ Rn :
kf (x + v) − f (x) − Df (x) · vk = R(x, v)kvk = R(x, v)kS −1 S · vk ≤
≤ R(x, v)kS −1 kkS · vk.
Since limv→0 R(x, v) = R(x, 0) = 0, we may take k ∈ N sufficiently
large so that R(x, v)kS −1 k ≤  for kvk ≤ k −1 , hence F2) is satisfied
for this choice of k. We then conclude that x ∈ F (S, k). Finally,
since G is dense in Jf+ , there exists c ∈ G such that c ∈ U(x, 2k 1
)⇔
1 1
x ∈ U(c, 2k ), so that x ∈ E(S, k, c) = B(c, 2k )∩F (S, k), thus proving
our contention.

Theorem 5.36 (Area Formula). Let f : Rn → Rm be Lipschitz,
n ≤ m. Then, for all A ∈ σ(Ln ),
ˆ ˆ
Jf dL =n
H0 (A ∩ f −1 {y}) dHn (y).
A Rm

Proof. If Ln (A) = 0, the first member is trivially null, and so is

the second member in view of lemma 5.29.iii). Therefore, in view of
Rademacher’s theorem 5.12, we may assume that A ⊂ Df = {x ∈ Rn |
∃Df (x)}. Let Jf+ = {x ∈ Df | Jf (x) > 0} and Jf0 = {x ∈ Df | Jf (x) =
0}, so that Df = Jf+ ∪˙ Jf0 .
1) Case 1: A ⊂ Jf+ . Fix t > 1. Let (Ek )k∈N be a sequence in B Rn given
by the Lipschitz linearization theorem 5.35, i.e. such that Jf+ =
∪˙ k∈N Ek and, for each k ∈ N, there exists Sk ∈ Sym(n) ∩ GL(Rn )
144 5. AREA AND COAREA FORMULAS

such that (Ek , Sk ) is a t-linearization for f . Then, ∀k ∈ N,

2.4
Hn f (A ∩ Ek ) = Hn f |Ek ◦ Sk−1 ◦ Sk (A ∩ Ek ) ≤

(5.10)
5.33.ii0 )
≤ (Lip f |Ek ◦ Sk−1 )n Hn Sk (A ∩ Ek ) ≤
≤ tn Hn Sk (A ∩ Ek ) ,

and

2.4
Hn Sk (A ∩ Ek ) = Hn Sk ◦ (f |Ek )−1 ◦ f (A ∩ Ek ) ≤

(5.11) −1 n n

5.33.ii0 )
≤ (Lip Sk ◦ (f |Ek ) ) H f (A ∩ Ek ) ≤
n n


≤ t H f (A ∩ Ek ) ,

On the other hand, it follows from corollary 5.34 that, ∀k ∈ N,

∀x ∈ Ek ,

(5.12) t−n JSk K ≤ Jf (x) ≤ tn JSk K.

Therefore, ∀k ∈ N:

(5.10) −n n
5.27
t−2n Hn f (A ∩ Ek ) ≤ t H Sk (A ∩ Ek ) =
(5.12)
= t−n JSk KLn (A ∩ Ek ) ≤
ˆ (5.12)
≤ Jf (x) dLn (x) ≤
(5.13) A∩Ek
5.27
≤ tn JSk KLn (A ∩ Ek ) =

(5.11)
= tn Hn Sk (A ∩ Ek ) ≤
≤ t2n Hn f (A ∩ Ek )

Since, ∀k ∈ N, f |Ek is 1-1 (by proposition 5.33),

´ we 0have ∀y ∈
m 0 −1
R , H (A ∩ Ek ∩ f {y}) = χf (A∩E
k ) (y), so that Rm H (A ∩ Ek ∩
f −1 {y}) dHn (y) = Hn f (A ∩ Ek ) . Therefore, from (5.13) and from
the monotone convergence theorem 1.62, we conclude that:
 5.2. THE AREA FORMULA 145

ˆ ˆ
1.62
X
−2n −1 −2n
t 0
H (A ∩ f n
{y}) dH (y) = t H0 (A ∩ Ek ∩ f −1 {y}) dHn (y) =
Rm k∈N Rm

X
(5.13)
= t−2n Hn f (A ∩ Ek ) ≤
k∈N
Xˆ 1.62
ˆ (5.13)
n
≤ Jf (x) dL (x) = Jf (x) dLn (x) ≤
k∈N A∩Ek A
X
t2n Hn f (A ∩ Ek ) =


≤
k∈N
ˆ
1.62
X
= t2n
H0 (A ∩ Ek ∩ f −1 {y}) dHn (y) =
k∈N Rm
ˆ
= t2n H0 (A ∩ f −1 {y}) dHn (y),
Rm

´ ´ ´
thus t−2n Rm H0 (A∩f −1 {y}) dHn (y) ≤ A Jf´dLn ≤ t2n Rm H0 (A∩
f´ −1 {y}) dHn (y). Taking t ↓ 1, it follows that Rm H0 (A∩f −1 {y}) dHn (y) =
A
Jf dLn , as asserted. ´ ´
2) Case 2: A ⊂ Jf0 . Then A Jf dLn = 0; we must show that H0 (A ∩
f −1 {y}) dHn (y) = 0. We may assume that Ln (A) < ∞ (since the
general case is obtained from this and from the monotone conver-
gence theorem, writing A = ∪˙ n∈N An , with ∀n ∈ N, An ∈ σ(Ln )
and Ln (An ) < ∞, which is possible thanks to the σ-finiteness of the
Lebesgue measure).
Fix 0
<  < 1. Define g : Rn → Rm+n ≡ Rm × Rn by g(x) :=

g is Lipschitz 1-1 and ∀x ∈ Dg = Df , Dg(x) =
f (x), x . Then
Df (x),  idRn ∈ L(Rn , Rm × Rn ).
We contend that there exists C = C(n, m, Lip f ) > 0 (in partic-
ular, C does not depend on ) such that ∀x ∈ A, 0 < Jg(x) ≤ C.
Assuming this contention, we have, denoting by pr1 : Rm ×Rn → Rm
146 5. AREA AND COAREA FORMULAS

the projection on the first factor:

2.4.3)
Hn f (A) = Hn pr1 ◦ g(A) ≤



g 1-1 and Lip pr1 =1
≤ (Lip pr1 )n Hn g(A) =
ˆ
case 1
= H0 (A ∩ g −1 {y}) dHn (y) =
ˆR
m+n

contention
= Jg dLn ≤
A
≤ CLn (A).
Thus, since 0 <  < 1 was arbitrarily
taken and we assumed
Ln (A) < ∞, it follows that Hn f (A) = 0. As the multiplic-
ity function N (f |A ) : Rn → R, y 7→ ´ H
0
(A ∩ f −1 {y}), is sup-
ported
´ on f (A), it then follows that H0 (A ∩ f −1 {y}) dHn (y) =
f (A)
H (A ∩ f −1 {y}) dHn (y) = 0, as asserted.
0

It remains to prove
the contention. Since ∀x ∈ Dg = Df ,

Dg(x) = Df (x),  idRn ∈ L(Rn , Rm × Rn ), the Jacobian matrix
of Dg(x) is the (m + n) × n matrix written in block form:
 
[Df (x)]
(5.14) [Dg(x)] = .
In

2
By the Binet-Cauchy formula 5.23, Jg(x) is the sum of the squares
of the n × n-minors of the above matrix. In particular taking
the minor corresponding the the last n rows, we conclude that
∀x ∈ Dg = Df , Jg(x) ≥ n > 0. On the other hand, to obtain
an upper bound for that sum:
• Note that the i-th row of the matrix [Df (x)] is ∇f i (x), where f i
is the i-th component of f in the standard basis of Rm ; the norm
of this row is therefore k∇f i (x)k = kDf i (x)k ≤ Lip f i ≤ Lip f .
• The sum of the squares of the n × n minors of [Dg(x)] may be
written as M1 + M2 , where the terms in M1 are the squares
2
of the n × n minors with rows in [Df (x)], i.e. M1 = Jf (x) ,
and the terms in M2 are the squares of the other minors, i.e.
n × n minors which have at least one row in In . Since  < 1
and the rows in [Df (x)] are bounded in norm by Lip f , each
minor of the latter type is
bounded by  · max{1, (Lip f )n−1
}.
m+n m m+n


Since there are n
− n
summands in M2 , M 2 ≤ n
−
m


2 n−1 2
n
 · max{1, (Lip f ) } . Hence, ∀x ∈ Dg = Df :
   

2
2 m+n m
2
Jg(x) ≤ Jf (x) + −  · max{1, (Lip f )n−1 }2 .
n n
 5.2. THE AREA FORMULA 147

In particular, if x ∈ A ⊂ Jf0 , we conclude that Jg(x) ≤ C,

where
s   
m+n m
C := − max{1, (Lip f )n−1 },
n n
thus proving our contention.
3) General case: A ⊂ Df . It is a direct consequence of cases 1 and 2:
ˆ ˆ ˆ
n n
Jf dL = Jf dL + Jf dLn =
A A∩Jf+ A∩Jf0
ˆ ˆ
−1
= 0
H (A ∩ Jf+ ∩f {y}) dH (y) + n
H0 (A ∩ Jf0 ∩ f −1 {y}) dHn (y) =
ˆ
Rm Rm

= H0 (A ∩ f −1 {y}) dHn (y).

Rm


Corollary 5.37. If f : Rn → Rm is Lipschitz, n ≤ m, then for
H -a.e. y ∈ Rm , f −1 {y} is countable.
n

5.28.b)
Proof. Since ∀x ∈ Df , Jf (x) ≤ kDf (x)kn ≤ (Lip´ f )n , it fol-
lows from the area ´formula 5.36 that, ∀K ⊂ Rn compact, Rm H0 (K ∩
f −1 {y}) dHn (y) = K Jf dLn < ∞. Then, ∀K ⊂ Rn compact, for Hn -
a.e. y ∈ Rm , H0 (K ∩ f −1 {y}) < ∞. Since Rn is σ-compact, it then fol-
lows that for Hn -a.e. y ∈ Rn , ∀K ⊂ Rn compact, H0 (K∩f −1 {y}) < ∞.
For such y, f −1 {y}∩K is finite for each compact K ⊂ Rn , hence f −1 {y}
is countable. 
Corollary 5.38 (Change of variables formula). Let f : Rn → Rm
be Lipschitz, n ≤ m. Then for all g : Rn → R Ln -measurable with
g ≥ 0 or g summable,
ˆ ˆ X
n
g(x) dHn (y).


g Jf dL =
Rn Rm x∈f −1 {y}

Proof. Suppose that g ≥ 0. By exercise 1.54, there exists a se-

quence (Ai )i∈N in σ(Ln ) such that
∞
X 1
g= χAi .
i=1
i

Let ψ : Rm → [0, ∞] be given by ψ(y) := x∈f −1 {y} g(x). Given

P

y ∈ Rm , we may compute x∈f −1 {y} g(x) by means of the monotone

P
148 5. AREA AND COAREA FORMULAS

convergence theorem (with respect to the counting measure on f −1 {y}):

X X X1 MCT 1.62
ψ(y) = g(x) = χAi (x) =
−1 −1 i∈N
i
x∈f {y} x∈f {y}
X1 X
= χAi (x) =
i
i∈N x∈f −1 {y}
X1
= H0 (Ai ∩ f −1 {y}).
i∈N
i

Since, for each i ∈ N, the multiplicity function N (f |Ai ) : y 7→ H0 (Ai ∩

f −1 {y}) is Hn -measurable, by lemma 5.29.ii), we therefore conclude
that ψ is Hn -measurable and ≥ 0. Besides, using the monotone con-
vergence theorem once more and the area formula, we have:
ˆ ˆ X
1 0 MCT 1.62
n
ψ(y) dH (y) = H (Ai ∩ f −1 {y}) dHn (y) =
Rm Rm i∈N i
X1ˆ AF 5.36
= H0 (Ai ∩ f −1 {y}) dHn (y) =
i∈N
i Rm
X1ˆ MCT 1.62
= Jf dLn =
i∈N
i Ai
ˆ X ˆ
1 n
= χAi Jf dL = g Jf dLn ,
Rn i∈N i Rn

thus proving the case in which g ≥ 0.

If g : Rn → R is Ln -summable, we write g = g + − g − and apply the
case already proved to g + and g − , from which the thesis follows.

Corollary 5.39. Let f : Rn → Rm be Lipschitz 1-1, n ≤ m.

´
i) ∀A ∈ σ(Ln ), Hn f (A) = A Jf dLn . In particular, we have
(5.15) f# (Ln xJf ) = H xIm f n

(equality as Borel regular outer measures on Rm ).

ii) ´If g : Rn → R is L´n -measurable with g ≥ 0 or g ∈ L1 (Ln ), then
Im f
g ◦f −1 dHn = Rn g Jf dLn . In particular, if g : Im f → [0, ∞]
is Borelian, then
ˆ ˆ
(5.16) g dH = g ◦ f Jf dLn .
n
Im f

Proof. If f is 1-1,
 5.2. THE AREA FORMULA 149

i) ∀A ∈ σ(L

n
), ´∀y ∈ Rm , H0 (A ∩ f −1 {y}) = χf (A) (y). Hence,
Hn f (A) = A Jf dLn as a direct consequence of the area for-
mula 5.36. The proof of (5.15) is done along the following steps:
x
• Hn Im f is a Borel regular outer measure. Indeed, since
Im f = ∪N ∈N f (BN ) is σ-compact, hence Borelian, we may
apply proposition 1.36.i).
x
• f# (Ln Jf ) is a Borel regular outer measure. Indeed, Ln Jf x
is a Radon measure on Rn , in view of lemma 4.11; hence
x
f# (Ln Jf ) is a Borel outer measure on Rm , since ∀U ⊂ Rm
open, f −1 (U ) is open by the continuity of f , thus Ln Jf - x
x
measurable, so that U is f# (Ln Jf )-measurable in view of
proposition 1.15.iii). It remains to prove the Borel regular-
x
ity of f# (Ln Jf ). Given T ⊂ Rm , the fact that Ln Jf is x
Radon ensures the existence of a sequence of open sets (Uk )k∈N
x
in Rn such that ∀k ∈ N, Uk
⊃ f −1 (T ) and inf{Ln Jf (Uk ) |
x
k ∈ N} = Ln Jf f −1 (T ) . Since ∀k ∈ N, Uk is σ-compact,
so is f (Uk ) (because f is continuous, hence it maps compact
sets to compact sets), thus f (Uk ) ∈ B Rm . Take ∀k ∈ N,
Bk := f (Uk ) ∪ (Rm \ Im f ) ∈ B Rm . Then ∀k ∈ N, Bk ⊃ T
x
and, as f −1 (Bk )
= Uk , inf{f# (Ln Jf )(Bk ) | k ∈ N} =
x x
Ln Jf f −1 (T ) = f# (Ln Jf )(T ), which implies the Borel
x
regularity of f# (Ln Jf ), as asserted.
• In view of the two previous items, it suffices to show that
x x
Hn Im f and f# (Ln Jf ) coincide in each B ∈ B Rm . In-
deed,

Hn xIm f (B) = H (Im f ∩ B) = H

n n

(∗)
f [f −1 (B)] =
ˆ
=
f −1 (B)
Jf dLn = f# (Ln xJf )(B),
where the equality (∗) is due to the area formula applied to
(∀y ∈ Rm )χf [f −1 (B)] (y) = H0 f −1 (B) ∩ f −1 {y} (because f is


1-1).
ii) ∀y ∈ Rm ,
(
X g ◦ f −1 (y) y ∈ Im f
g(x) =
x∈f −1 {y}
0 y ∈ Rm \ Im f.
´ ´
It then follows from corollary 5.38 that Im f g◦f −1 dHn = Rn g Jf dLn .
If g : Im f → [0, ∞] is Borelian, we may apply the latter equality
to g ◦ f in place of g, thus yielding (5.16).
150 5. AREA AND COAREA FORMULAS


Example 5.40 (applications of the area formula).
1) (length of a curve) Let −∞ < a < b < ∞ and γ : [a, b] → Rm be
Lipschitz 1-1. We may extend γ to a Lipschitz function on R, which
we still denote by γ. Note that, for all t in the set Dγ of the points
of differentiability of γ,
Jγ(t) = kγ 0 (t)k.
It then follows from the change of variables formula 5.38 with g =
χ[a,b] that
ˆ b ˆ
0 5.38
kγ (t)k dt = χ[a,b] Jγ dL1 =
a
ˆR X
= ( χ[a,b] (x)) dH1 (y) =
Rm x∈γ −1 {y}
ˆ
= χγ([a,b]) (y) dH1 (y) =
Rm
= H1 γ([a, b]) .

2) (area of a graph) Let g : R
n → R be Lipschitz and f : Rn → Rn+1 be

given by f (x) := x, g(x) . Then f is Lipschitz 1-1 and, computing
by means of the Binet-Cauchy formula 5.23, ∀x ∈ Df = Dg ,
p
Jf (x) = 1 + k∇g(x)k2 .
For each U ⊂ Rn open, it follows from corollary 5.39.i) that the

“surface area” of the graph of g over U , Γ = Γ(g; U ) := { x, g(x) |
x ∈ U } = f (U ), is given by:
ˆ ˆ p
n n
H (Γ) = Jf dL = 1 + k∇g(x)k2 dx.
U U

Exercise 5.41 (Area Formula for locally Lipschitz maps). The

area formula and its corollaries remain valid for locally Lipschitz maps
defined on open subsets of Rn . That is, let n ≤ m, Ω ⊂ Rn open and
f : Ω → Rm locally Lipschitz.
a) (area formula) For all Ln -measurable A ⊂ Ω, the multiplicity func-
tion N (f |A ) : Rm → [0, ∞], y 7→ H0 (A ∩ f −1 {y}), is Hn -measurable
and ˆ ˆ
n
Jf dL = H0 (A ∩ f −1 {y}) dHn (y).
A Rm
n m −1
b) For H -a.e. y ∈ R , f {y} is countable.
 5.3. THE COAREA FORMULA 151

c) (change of variables formula) If g : Ω → R is Ln |Ω -measurable and

g ≥ 0 or g ∈ L1 (Ln |Ω ), then
ˆ ˆ  X 
n
g Jf dL = g(x) dHn (y),
Ω Rm x∈f −1 {y}

meaning that the integral in the second member makes´sense and the
´equality holds. In particular, if f is 1-1, it follows that Ω g Jf dLn =
−1
Im f
g ◦ f dHn .

The classical C1 change of variables formula 1.82 may be obtained

as a corollary of part c) of the previous exercise. The classical formula
actually holds with much weaker hypotheses on the change of variables
φ : U → Rn with U ⊂ Rn open; it suffices, for instance, that φ be a 1-1
C1 -map (it need not be a diffeomorphism).
Exercise 5.42 (Hausdorff dimension and Lebesgue measure of a
k-dimensional Riemannian submanifold of Rn ). For any smooth em-
bedded k-Riemannian submanifold M ⊂ Rn , the measure induced by
the Riemannian metric on M (i.e. the Lebesgue measure of M) coin-
cides with the trace Hk |M . Conclude that H-dim M = k and, if M is
x
closed (i.e. topologically closed), Hk M is a Radon measure on Rn .

5.3. The coarea formula

In this section we assume n ≥ m. The coarea formula is a powerful
generalization of Fubini-Tonelli’s theorem 1.84.
Lemma 5.43 (Coarea formula, linear case). Let L : Rn → Rm be
linear, n ≥ m, A ∈ σ(Ln ). Then:
i) N (L|A ) : Rm → [0, ∞] given by N (L|A )(y) := Hn−m (A ∩ L−1 {y})
is Lm -measurable.
ii)
ˆ
(5.17) Hn−m (A ∩ L−1 {y}) dLm (y) = JLKLn (A)
Rm

Proof. Let O ∈ O(m, n) and S ∈ Sym(m) be given by theorem

5.20, i.e. such that L = S ◦ O∗ is a polar decomposition of L.
(1) Case 1: dim Im L < m. Then, for Lm -a.e. y ∈ Rm , L−1 {y} =
∅, thus N (L|A )(y) = Hn−m (A ∩ L−1 {y}) = 0. That is, N (L|A )
is null Lm -a.e., hence it is Lm -measurable. On the other hand,
since Im L = Im S, we have JLK = |det S| = 0. Therefore,
both members of (5.17) are null.
152 5. AREA AND COAREA FORMULAS

(2) Case 2: L = P : Rn ≡ Rm × Rn−m → Rm is the projection on

the first factor (hence O = P ∗ and S = idRm ). Fix y ∈ Rm and
let pr2 : Rn ≡ Rm × Rn−m → Rn−m be the projection on the
second factor. Then the restriction pr2 : P −1 {y} → Rn−m is an
isometry which maps A ∩ P −1 {y} to the y-section Ay ⊂ Rn−m
(see notation preceding Fubini-Tonelli’s theorem 1.84). There-
fore, from proposition 2.4 parts i) and ii) and from theorem
2.21, we conclude that N (P |A )(y) = Hn−m (A ∩ P −1 {y}) =
Hn−m (Ay ) = Ln−m (Ay ). Hence, from Fubini-Tonelli’s theo-
rem 1.84.ii) applied to the product measure Lm ×Lm−n (which
coincides with Ln , in view of example
´ 1.86), we conclude that
N (P |A ) is Lm -measurable and Rm N (P |A ) dLm = Ln (A), thus
proving (5.17) (since JP K = 1).
(3) Case 3: L : Rn → Rm surjective. Note that, since Im L =
Im S, we have S ∈ Sym(m) ∩ GL(Rm ).
We contend that there exists Q ∈ O(n) such that O∗ =
P ◦ Q, where P : Rn ≡ Rm × Rn−m → Rm is the projection
on the first factor, as in the previous item. Indeed, extend
O ∈ O(m, n) to a linear isometry S : Rn ≡ Rm × Rn−m → Rn
and define Q := S ∗ . Since P ∗ : Rm → Rm × Rn−m is the
inclusion on the first factor, we have S ◦ P ∗ = O, hence O∗ =
P ◦ S ∗ = P ◦ Q, as we wanted.
With Q ∈ O(n) given by the contention proved above, we
have, ∀y ∈ Rm ,
N (L|A )(y) = Hn−m (A ∩ L−1 {y}) =
= Hn−m A ∩ (S ◦ P ◦ Q)−1 {y} =


2.4.2)
= Hn−m Q−1 Q(A) ∩ P −1 {S −1 (y)}

=
n−m −1 −1


=H Q(A) ∩ P {S (y)} =
= N (P |Q(A) ) ◦ S −1 (y).
That is, N (L|A ) = N (P |Q(A) ) ◦ S −1 . By the previous item,
N (P |Q(A) ) is Lm -measurable, and S −1 is continuous, hence
Borelian; it then follows that the composition N (L|A ) = N (P |Q(A) )◦
S −1 is Lm -measurable and ≥ 0. Moreover,
ˆ ˆ
1.81.ii)
m
N (f |A)(y) dL (y) = N (P |Q(A) ) ◦ S −1 (y) dLm (y) =
Rm Rm
ˆ
Case 2
= |det S| N (P |Q(A) ) dLm =
Rm

Q∈O(n)
= |det S|Ln Q(A) = JLKLn (A),
 5.3. THE COAREA FORMULA 153

thus proving (5.17).



In the next lemma we make computations with the upper integral

introduced in exercise 1.68.

Lemma 5.44. Let n, m ∈ N, f : Rn → Rm be Lipschitz. Then,

∀k, l ∈ [0, ∞) and ∀A ⊂ Rn ,
ˆ ∗
α(k)α(l)
Hk (A ∩ f −1 {y}) dHl (y) ≤ (Lip f )l Hk+l (A).
Rm α(k + l)

Note that we neither assume n ≥ m nor the measurability of A

in the statement of the lemma above. This is a particular case from
Federer’s theorem 2.10.25 in [Fed69]; the theorem actually holds for
any Lipschitz map f : X → Y between metric spaces X and Y . We
will prove only the case l = m, for which it is possible to make a simpler
argument, adapted from [EG91], thanks to the isodiametric inequality
2.19. Only this case will be needed in the proof of the coarea formula.

Proof for the case l = m. For each j ∈ N, by proposition 2.4.4)

there exists (Bij )i∈N cover of A by closed sets with diameters ≤ 1/j such
that
X  diam B j k+m 1
i k+m
(5.18) α(k + m) ≤ H1/j (A) + .
i∈N
2 j

∀i, j ∈ N, define
 diam B j k
gij := α(k) i
χf (B j ) .
2 i

Since f (Bij ) is σ-compact (because Bij is closed, hence σ-compact, and

f is continuous), hence Borel measurable, gij is Borelian and ≥ 0, and
so is i∈N gij : Rm → [0, ∞]. Moreover, for each y ∈ Rm and j ∈ N,
P

A ∩ f −1 {y} is contained in the union of the balls of (Bij )i∈N which cut
f −1 {y}, i.e. such that y ∈ f (Bij ). It then follows that, ∀y ∈ Rm ,
X
k
H1/j (A ∩ f −1 {y}) ≤ gij (y).
i∈N
154 5. AREA AND COAREA FORMULAS

Therefore,
ˆ ∗
Hk (A ∩ f −1 {y}) dLm (y) =
Rm
ˆ ∗ monotonicity of
´∗
k −1 m
= lim H1/j (A ∩f {y}) dL (y) ≤
m j→∞
ˆR∗ X 1.68.b)
≤ lim inf gij (y) dLm (y) =
Rm j→∞
i∈N
ˆ X Fatou 1.63
= lim inf gij (y) dLm (y) ≤
Rm j→∞
i∈N
ˆ
MCT 1.62
X
≤ lim inf gij (y) dLm (y) =
j→∞ Rm i∈N
Xˆ
= lim inf gij (y) dLm (y) =
j→∞ Rm
i∈N
X  diam B j k
isodiametric 2.19
= lim inf α(k) i
Lm f (Bij ) ≤
j→∞
i∈N
2
X  diam B j k  diam f (B j ) m
i i
≤ lim inf α(k) α(m) ≤
j→∞
i∈N
2 2
α(k)α(m) X  diam B j k+m (5.18)
i
≤ (Lip f )m lim inf α(k + m) ≤
α(k + m) j→∞
i∈N
2
α(k)α(m) m

k+m 1
≤ (Lip f ) lim inf H1/j (A) + =
α(k + m) j→∞ j
α(k)α(m)
= (Lip f )m Hk+m (A).
α(k + m)


Lemma 5.45. Let f : Rn → Rm be Lipschitz, n ≥ m, A ∈ σ(Ln ).

Then:
i) For Lm -a.e. y ∈ Rm , A ∩ f −1 {y} is Hn−m -measurable.
ii) N (f |A ) : Rm → [0, ∞] given by N (f |A )(y) := Hn−m (A ∩ f −1 {y})
is Lm -measurable.

Proof.
1) Case 1: A is compact. Fix t ≥ 0. For each i ∈ N, let Ui be defined
as the set of points y ∈ Rm such that there exist finitely many open
 5.3. THE COAREA FORMULA 155

sets (Sj )1≤j≤k satisfying the following conditions:




 A ∩ f −1 {y} ⊂ ∪kj=1 Sj ,
(5.19) diam Sj ≤ 1i , ∀1 ≤ j ≤ k,
 n−m
 Pk diam Sj
j=1 α(n − m) ≤ t + 1i .


2

2) Claim #1: ∀i ∈ N, Ui is open. Indeed, let y ∈ Ui . Take (Sj )1≤j≤k

satisfying the conditions (5.19). We contend that there exists r > 0
−1

k
such that A ∩ f U(y, r) ⊂ ∪j=1 Sj , from which we conclude that
U(y, r) ⊂ Ui , thus proving the claim. The contention is a direct
consequence of the fact that A is compact and f is continuous: if
there were no such r > 0, we could take a sequence (yh )h∈N in Rm
convergent to y such that ∀h ∈ N, there exists xh ∈ f −1 {yh } ∩
A \ ∪1≤j≤k Sj . Since A \ ∪1≤j≤k Sj is compact, there would be a
subsequence of (xh )h∈N , which we assume to be (xh )h∈N itself up
to changing the notation, such that xh → x ∈ A \ ∪1≤j≤k Sj . By
continuity, we conclude that f (x) = lim f (xh ) = lim yh = y, hence
x ∈ f −1 {y} \ ∪1≤j≤k Sj , thus yielding a contradiction which proves
our contention.
3) Claim #2: {y ∈ Rm | Hn−m (A ∩ f −1 {y}) ≤ t} = ∩i∈N Ui , hence
it is a Borel set. Since {y ∈ Rm | Hn−m (A ∩ f −1 {y}) ≤ t} = ∅
for t < 0, the claim then implies that N (f |A ) is Borelian if A is
compact. Since ∀y ∈ Rm , A ∩ f −1 {y} is compact, hence Borelian,
we achieve the proof of case 1 once we show the claim.
Proof of claim #2:
• Assume that Hn−m (A ∩ f −1 {y}) ≤ t. Then, ∀δ > 0, Hδn−m (A ∩
f −1 {y}) ≤ t. Given i ∈ N, choose δ ∈ (0, 1i ). In view of
proposition 2.4.4), there exists a countable cover G ofPA∩f −1 {y}
by open subsets of Rn with diameters ≤ δ such that S∈G α(n−
 n−m
m) diam 2
S
< t + 1i . Since A ∩ f −1 {y} is compact, we may
take a finite subcover (Sj )1≤j≤k of G satisfying (5.19), so that
y ∈ Ui . As i ∈ N is arbitrary, it then follows that y ∈ ∩i∈N Ui .
n−m
• Conversely, if ∀ i ∈ N, y ∈ Ui , then (5.19) ensures that H1/i (A∩
−1 1
f {y}) ≤ t + i ; hence, taking i → ∞, we conclude that
Hn−m (A ∩ f −1 {y}) ≤ t.
4) Case 2: A is σ-compact (in particular, that holds if A is open).
Then ∀y ∈ Rm , A ∩ f −1 {y} is σ-compact, hence Borelian. More-
over, we may take an increasing sequence (Ki )i∈N of compact subsets
of A whose union is A, so that ∀y ∈ Rm , the sequence of Borel sets
(Ki ∩ f −1 {y})i∈N increases to A ∩ f −1 {y}. Then, applying the con-
tinuity from below 1.11 for Hn−m , it follows that N (f |Ki ) increases
156 5. AREA AND COAREA FORMULAS

pointwise to N (f |A ); from case 1 and from theorem 1.41.iv) we

therefore conclude that N (f |A ) is Borelian.
5) Case 3: Ln (A) = 0. It follows from lemma 5.44 ´ ∗ with k = mn −
m and l = m, and from theorem 2.21, that Rm N (f |A ) dL =
0.
´ Hence, from exercise 1.68.a), N (f |A ) is Lm -measurable and
Rm
N (f |A ) dL = 0, so that N (f |A ) is Lm -a.e. null. That is,
m

for Lm -a.e. y ∈ Rm , Hn−m (A ∩ f −1 {y}) = 0, which implies that

A ∩ f −1 {y} is Hn−m -measurable.
6) Case 4: Ln (A) < ∞. Since Ln is a Radon measure, we may take
a decreasing sequence (Uk )k∈N of open sets containing A such that
inf{Ln (Uk ) | k ∈ N} = Ln (A). Hence, taking B := ∩k∈N Uk ∈ B Rn ,
we have Ln (B) = Ln (A) < ∞; as A is Ln -measurable, we conclude
that Ln (B \ A) = 0. In particular, it follows from case 3 that,
for Lm -a.e. y ∈ Rm , (B \ A) ∩ f −1 {y} is Hn−m
-null. For such y,
A∩f −1 {y} = (B ∩f −1 {y})\ (B \A)∩f −1 {y} is Hn−m -measurable
and Hn−m (B ∩ f −1 {y}) = Hn−m (A ∩ f −1 {y}), thus showing that
N (f |A ) = N (f |B ) Lm -a.e., so that case 4 will be done once we prove
that N (f |B ) is Lm -measurable. Indeed,
• for each y ∈ Rm and k ∈ N, Uk ∩ f −1 {y} is Borelian and the
sequence (Uk ∩ f −1 {y})k∈N decreases to B ∩ f −1 {y};
• since Ln (A) < ∞ and Ln (Uk ) ↓ Ln (A), we may assume that
Ln (U1 ) < ∞ (discarding the first terms of the sequence (Uk )k∈N ,
if necessary). It then´ follows from lemma ´5.44 with k = n −
∗
m and l = m that Rm N (f |U1 ) dLm = Rm N (f |U1 ) dLm <
∞. Hence, for Lm -a.e. y ∈ Rm , N (f |U1 )(y) = Hn−m (U1 ∩
f −1 {y}) < ∞; for such y, we may apply the continuity from
above 1.11 to conclude that N (f |Uk )(y) ↓ N (f |B )(y). That
is N (f |Uk ) decreases Lm -a.e. to N (f |B ). It then follows that
N (f |B ) is Lm -measurable (in view of case 2 and of theorem
1.41.iv), as asserted.
7) General case. By the σ-finiteness of Ln , we may write A = ∪˙ k∈N Ak ,
where ∀k ∈ N, Ak ∈ σ(Ln ) and Ln (Ak ) < ∞. Then A ∩ f −1 {y} =
∪˙ k∈N (Ak ∩ f −1 {y}). It follows from case 4 that, for Lm -a.e. y ∈
Rm , ∀k ∈ N, Ak ∩ f −1 {y} is Hn−m -measurable; hence, for such y,
A ∩ f −1 {y} is Hn−m -measurable and N (f |A )(y) = k∈N N (f |Ak )(y)
P
by the σ-additivity of Hn−m . Then N (f |A ) is Lm -measurable, in
view of case 4 and theorem 1.41.iv).

Lemma 5.46 (Lipschitz linearization, part II). Let t > 1, h : Rn →
Rn Lipschitz and Jh+ = {x ∈ Dh | Jh(x) > 0}. Then there exists a
countable disjoint family (Ek )k∈N in B J + such that Ln (Jh+ \ ∪k∈N Ek ) =
h
 5.3. THE COAREA FORMULA 157

0 and, ∀k ∈ N, h|Ek is 1-1 and there exists Sk ∈ Sym(n) ∩ GL(Rn )

satisfying
i) Lip Sk−1 ◦ h|Ek ≤ t and Lip(h|Ek )−1 ◦ Sk ≤ t;
ii) ∀x ∈ Ek , kSk−1 ◦ Dh(x)k ≤ t and kDh(x)−1 ◦ Sk k ≤ t.
Remark 5.47. With the notation from the previous lemma:
1) Conditions i) and ii) are equivalent to, respectively:
i’) ∀x, y ∈ h(Ek ),
t−1 kSk−1 · (x − y)k ≤ k(h|Ek )−1 (x) − (h|Ek )−1 (y)k ≤ tkSk−1 · (x − y)k;
ii’) ∀x ∈ Ek , ∀v ∈ Rn , t−1 kSk−1 · vk ≤ kDh(x)−1 · vk ≤ tkSk−1 · vk.
The proof is immediate and similar to the argument used in propo-
sition 5.33.
2) Condition i) implies that h|Ek has Lipschitz inverse, since (h|Ek )−1 =
[(h|Ek )−1 ◦ Sk ] ◦ Sk−1 , whence Lip(h|Ek )−1 ≤ tkSk−1 k.
3) Condition ii) implies that, ∀x ∈ Ek :
(5.20) t−n |det Sk | ≤ Jh(x) ≤ tn |det Sk |.
Indeed,
5.28.a)
Jh(x) = |det Sk ||det Sk−1 |JDh(x)K =
= |det Sk |JSk−1 ◦ Dh(x)K,

hence (5.20) follows from ii) and from exercise 5.28.b) with Sk−1 ◦
Dh(x) in place of T .
Proof.
1) Let (Fk )k∈N be a countable disjoint family in B Rn such that Jh+ =
∪˙ k∈N Fk and ∀k ∈ N, h|Fk is 1-1 with Lipschitz inverse. The existence
of such a family follows from theorem 5.35 and proposition 5.33 with
h in place of f .
2) Fix k ∈ N. As (h|Fk )−1 : h(Fk ) → Rn is Lipschitz, by theorem
5.1 (or theorem 5.2) it may be extended to a Lipschitz map hk :
Rn → Rn . Since h(Fk ) ⊂ {x ∈ Rn | h ◦ hk (x) = x}, it follows
from corollary

5.17 with hk in place of f and h in place of g that
Dh hk (x) ◦ Dhk (x) = idRn for Ln -a.e. x ∈ h(Fk ). Thus, defining
Yk := {x ∈ Rn | ∃Dhk (x), ∃Dh hk (x) , Dh hk (x) ◦ Dhk (x) = idRn }

then Yk ∈ B Rn (in view of
exercise 5.13) and h(Fk ) \ Yk is Ln -null.

Besides, ∀x ∈ Yk , Jh hk (x) · Jhk (x) = 1, so that Jhk (x) > 0, hence
Yk ⊂ Jh+k .
158 5. AREA AND COAREA FORMULAS

3) Applying the Lipschitz linearization theorem 5.35 to hk , there exists

a countable disjoint family (Gkj )j∈N in B Rn and a sequence (Rjk )j∈N
in Sym(n) ∩ GL(Rn ) such that Jh+k = ∪˙ j∈N Gkj and ∀j ∈ N, (Gkj , Rjk )
is a t-linearization for hk . Define, for each j ∈ N,
Ejk := Fk ∩h−1 (Gkj ∩Yk ) ∈ B Rn and Sjk := (Rjk )−1 ∈ Sym(n)∩GL(Rn ).
We will prove that the countable family (Ejk , Sjk )k,j∈N satisfies con-
ditions stated in the theorem.
4) It is clear that (Ejk )k,j∈N is a disjoint family in B J + and ∀k, j ∈ N,
f

h|Ejk is 1-1 with Lipschitz inverse, since Ejk ⊂ F k . Namely, (h|Ejk )−1
is the restriction of hk to h(Ejk ) = h(Fk ) ∩ Gkj ∩ Yk .
5) We contend that Jh+ \ ∪k,n∈N Ek,j is Ln -null. Indeed, since Jh+ =
∪˙ k∈N Fk , it suffices to show that, for each k ∈ N, Fk \ ∪j∈N Ejk is
Ln -null. Since h|Fk is bi-Lipschitz onto h(Fk ), the latter condition
is equivalent to h(Fk \ ∪j∈N Ejk ) being Ln -null. As
h(Fk \ ∪j∈N Ejk ) = h Fk \ [Fk ∩ h−1 (Yk ∩ ∪j∈N Gkj )] =

= h Fk \ h−1 (Yk ∩ Jh+k ) =

| {z }
=Yk

= h(Fk ) \ Yk ,
the contention follows from part 2).
6) ∀k, j ∈ N, hk |Gkj extends (h|Ejk )−1 (by part 4), hence (hk |Gkj )−1 ex-
tends h|Ejk . Therefore, ∀k, j ∈ N,

Lip(Sjk )−1 ◦ h|Ejk = Lip Rjk ◦ h|Ejk ≤ Lip Rjk ◦ (hk |Gkj )−1 ≤ t,
Lip(h|Ejk )−1 ◦ Sjk = Lip(h|Ejk )−1 ◦ (Rjk )−1 ≤ Lip hk |Gkj ◦ (Rjk )−1 ≤ t,
where the last inequalities in both lines are justified by the fact that
(Gkj , Rjk ) is a t-linearization for hk and by proposition 5.33. Thus,
∀k, j ∈ N, condition i) in the statement of the lemma is fulfilled by
(Ejk , Rjk ).


7) ∀k, j ∈ N, ∀x ∈ Ejk , we have h(x) ∈ Gkj ∩ Yk and x = hk h(x) ; in


particular, by the definition of Yk , Dh(x) ◦ Dhk h(x) = idRn , i.e.

−1
Dh(x) = Dhk h(x) . It then follows that, ∀k, j ∈ N, ∀x ∈ Ejk :

−1
k(Sjk )−1 ◦ Dh(x)k = kRjk ◦ Dhk h(x) k ≤ t,
kDh(x)−1 ◦ Sjk k = kDhk h(x) ◦ (Rjk )−1 k ≤ t,

where the last inequalities in both lines are justified by the fact that
(Gkj , Rjk ) is a t-linearization for hk and by proposition 5.33. Thus,
 5.3. THE COAREA FORMULA 159

∀k, j ∈ N, condition ii) in the statement of the lemma is fulfilled by

(Ejk , Rjk ), which concludes the proof.

Theorem 5.48 (Coarea formula). Let f : Rn → Rm be Lipschitz,
n ≥ m. Then, for each Ln -measurable A ⊂ Rn ,
ˆ ˆ
(5.21) n
Jf dL = Hn−m (A ∩ f −1 {y}) dLm (y).
A Rm
Remark 5.49.
1) Recall that N (f |A ) : y 7→ Hn−m (A ∩ f −1 {y}) is Lm -measurable, by
lemma 5.45, so that the integral in the second member of the coarea
formula makes sense.
2) If f : Rn ≡ Rm × Rn−m → Rm is the projection on the first factor,
we have Jf ≡ 1 and the coarea formula reduces to Fubini-Tonelli’s
theorem 1.84. The general case may be interpreted, therefore, as a
“curvilinear” generalization of Fubini-Tonelli’s theorem.
3) If n = m, the coarea formula coincides with the area formula 5.36.
4) If we take the Borel set A := (Rn \Df )∪Jf0 = {x ∈ Rn | @Df (x) or Jf (x) =
0} in the coarea formula, we conclude that Hn−m (A ∩ f −1 {y}) = 0
for Lm -a.e. y ∈ Rm . That may be interpreted as a measure theo-
retic version of Morse-Sard’s theorem: Lm -a.e. y ∈ Rm is a measure
theoretic “regular value” of f , in the sense that, up to Hn−m null
sets, f −1 {y} lies in the set Jf+ of points where Df has maximal rank.
Proof.
1) If A ⊂ Rn \ Df , then Ln (A) = 0 by Rademacher’s theorem 5.12,
hence the first member in (5.21) is null, and so is the second in view
of lemma 5.44 with k = n − m and l = m. Therefore, it suffices
to show (5.21) for A ⊂ Df = Jf+ ∪˙ Jf0 . Since both members are
additive on σ(Ln ), it suffices to consider the cases A ⊂ Jf+ and
A ⊂ Jf0 .
2) Case 1: A ⊂ Jf+ . For each λ ∈ Λ(n, n − m), define hλ : Rn →
Rm × Rn−m by

hλ (x) := f (x), Pλ (x) ,
where Pλ is the orthogonal projection onto heλ(1) , . . . , eλ(n−m) i
≡
Rn−m . For all x ∈ Dhλ = Df , we have Dhλ (x) = Df (x), Pλ ∈
L(Rn ); therefore, Jhλ (x) > 0 iff x ∈ Jfλ , where
Jfλ := {x ∈ Jf+ | Pλ |ker Df (x) is 1-1 }.
For each x ∈ Jf+ , there exists λ ∈ Λ(n, n−m) such that ker Df (x)
is transversal to heλ(1) , . . . , eλ(n−m) i ≡ Rn−m (hence x ∈ Jfλ ). That
160 5. AREA AND COAREA FORMULAS

is, Jf+ = ∪λ∈Λ(n,n−m) Jfλ . Therefore, we may decompose A as a dis-

joint union A = ∪˙ λ∈Λ(n,n−m) Aλ , with Aλ ⊂ Jfλ Ln -measurable. By
the additivity of both members of (5.21) on Ln , it then suffices
to show the equality for each Aλ . We may therefore assume that
A ⊂ Jfλ = Jh+λ for a given λ ∈ Λ(n, n − m).
3) For simplicity of notation, we put h := hλ : Rn → Rn . Let q : Rn ≡
Rm × Rn−m → Rm be the projection on the first factor, so that
f = q ◦ h.
Fix t > 1. Apply lemma 5.46 to h in order to obtain a disjoint
countable family (Ek )k∈N in B J + and a sequence (Sk )k∈N ∈ Sym(n)∩
h
GL(Rn ) such that Ln (Jh+ \ ∪˙ k∈N Ek ) = 0 and ∀k ∈ N, h|Ek is 1-1
and conditions i), ii) in the statement of the lemma are fulfilled. Let
∀k ∈ N, Ak := A ∩ Ek ∈ σ(Ln ), so that A \ ∪˙ k∈N Ak is Ln -null (since
A ⊂ Jh+ ).
We contend that, ∀k ∈ N, ∀x ∈ Ak ,

(5.22) t−m Jq ◦ Sk K ≤ Jf (x) ≤ tm Jq ◦ Sk K.

Indeed, since f = q ◦ h, Df (x) = q ◦ Dh(x) = q ◦ Sk ◦ Sk−1 ◦ Dh(x) .

Thus, defining C := Sk−1 ◦ Dh(x), we have q ◦ Sk = Df (x) ◦ C −1 ,

∗
whence (q ◦ Sk )∗ = (C −1 ) ◦ Df (x)∗ . Therefore, applying exercise
∗ ∗
5.28.c) with Df (x) : Rm → Rn in place of T and (C −1 ) ∈ GL(Rn )
∗
in place of R, and noting that (C −1 ) = (C ∗ )−1 and that the transpo-
∗
sition (·) preserves Jacobians and is a linear isometry, we conclude
that

kCk−m Jf (x) ≤ Jq ◦ Sk K ≤ kC −1 km Jf (x),

hence

ii) from lemma 5.46

Jf (x) ≤ kCkm Jq ◦ Sk K ≤ tm Jq ◦ Sk K,
ii) from lemma 5.46
Jf (x) ≥ kC −1 k−m Jq ◦ Sk K ≥ t−m Jq ◦ Sk K,

thus proving our contention.

 5.3. THE COAREA FORMULA 161

4) ∀k ∈ N,

ˆ
−2n
t Hn−m (Ak ∩ f −1 {y}) dLm (y) =
Rm
ˆ
−2n
Hn−m (Sk−1 ◦ h|Ak )−1 ◦ Sk−1 ◦ h|Ak (Ak ∩ h−1 q −1 {y}) dLm (y) =


=t
Rm
ˆ 2.4.3)
= t−2n Hn−m (Sk−1 ◦ h|Ak )−1 [Sk−1 ◦ h(Ak ) ∩ (q ◦ Sk )−1 {y}] dLm (y) ≤


Rm
ˆ 5.46.i)
−1 n−m −2n
Hn−m Sk−1 ◦ h(Ak ) ∩ (q ◦ Sk )−1 {y} dLm (y) ≤


≤ [Lip(h|Ak ) ◦ Sk ] t
ˆ Rm
5.43
≤ t−n−m Hn−m Sk−1 ◦ h(Ak ) ∩ (q ◦ Sk )−1 {y} dLm (y) =


Rm

2.4.3)
= t−n−m Jq ◦ Sk KLn Sk−1 ◦ h(Ak ) ≤
5.46.i)
≤ t−n−m Jq ◦ Sk K(Lip Sk−1 ◦ h|Ak )n Ln (Ak ) ≤
(5.22)
≤ t−m Jq ◦ Sk KLn (Ak ) ≤
ˆ (5.22)
≤ Jf dLn ≤
Ak

2.4.3)
≤ tm Jq ◦ Sk KLn (Ak ) = tm Jq ◦ Sk KLn (h|−1 Ak ◦ S k ) ◦ (Sk
−1
◦ h|Ak )(A k ) ≤

5.46.i)
≤ tm Jq ◦ Sk K(Lip h|−1
Ak ◦ S k ) n n
L Sk
−1
◦ h| A k
(A k ) ≤
5.43
≤ tm+n Jq ◦ Sk KLn Sk−1 ◦ h|Ak (Ak ) =


ˆ
m+n
Hn−m Sk−1 ◦ h(Ak ) ∩ (q ◦ Sk )−1 {y} dLm (y) =


=t
ˆR
m

= tm+n Hn−m (Sk−1 ◦ h|Ak ) ◦ (h|−1 −1 −1


 m
Ak ◦ S k ) S k ◦ h(A k ) ∩ (q ◦ S k ) {y} dL (y) =
ˆ Rm
m+n
Hn−m Sk−1 ◦ h|Ak Ak ∩ h|−1 −1

 m
=t Ak q {y} dL (y) =
Rm
ˆ 2.4.3)
m+n
Hn−m Sk−1 ◦ h|Ak (Ak ∩ f −1 {y}) dLm (y) ≤
 
=t
Rm
ˆ 5.46.i)
−1
≤t m+n
(Lip Sk ◦ h|Ak ) n−m
Hn−m (Ak ∩ f −1 {y}) dLm (y) ≤
ˆ Rm

≤ t2n Hn−m (Ak ∩ f −1 {y}) dLm (y).

Rm
162 5. AREA AND COAREA FORMULAS

In particular, ∀k ∈ N,
ˆ ˆ
−2n n−m −1 m
t H (Ak ∩ f {y}) dL (y) ≤ Jf dLn ≤
m Ak
(5.23) R
ˆ
≤ t2n Hn−m (Ak ∩ f −1 {y}) dLm (y).
Rm

5) It follows from lemma 5.45 that, for Ln -a.e. y ∈ Rm , ∀k ∈ N,

−1
Pk ∩f {y}
A is Hn−m -measurable, so that Hn−m (∪˙ k∈N Ak ∩f −1 {y}) =
k∈N H
n−m
(Ak ∩ f −1 {y}). It then follows from the monotone con-
vergence theorem 1.62 that
(5.24)
ˆ Xˆ
−1
H n−m ˙ m
( ∪ Ak ∩f {y}) dL (y) = Hn−m (Ak ∩f −1 {y}) dLm (y).
Rm k∈N Rm
k∈N

6) We contend that
(5.25)
ˆ ˆ
−1
H n−m
( ∪˙ Ak ∩ f m
{y}) dL (y) = Hn−m (A ∩ f −1 {y}) dLm (y).
Rm k∈N Rm

Indeed, since ∪˙ k∈N Ak ⊂ A, the inequality “≤” trivially holds in

(5.25). On the other hand, by subadditivity we have, ∀y ∈ Rm ,
−1
Hn−m (A∩f {y})´ ≤ Hn−m (∪˙ k∈N Ak ∩f −1 {y})+Hn−m
´ (A\ ∪˙ k∈N Ak )∩
−1 −1

n−m m n−m ˙
f {y} , whence ´Rm H (A∩f {y}) dL (y) ≤
Rm H (∪k∈N Ak ∩
−1 m n−m −1 m
f {y}) dL (y)+ Rm H (A\ ∪˙ k∈N Ak )∩f {y} dL (y). As A\
˙∪k∈N Ak is Ln -null (by part 3), it follows from lemma
5.44 with k =
´
n−m and l = m that Rm Hn−m (A\∪˙ k∈N Ak )∩f −1 {y} dLm (y) = 0,
thus proving the reverse inequality and our contention follows.
It then follows from (5.24) and (5.25) that
(5.26)
ˆ Xˆ
−1
H n−m m
(A ∩ f {y}) dL (y) = Hn−m (Ak ∩ f −1 {y}) dLm (y).
Rm k∈N Rm

7) Since A \ ∪˙ k∈N Ak is Ln -null, we have

ˆ ˆ Xˆ
n n MCT 1.62
(5.27) Jf dL = Jf dL = Jf dLn .
A ˙ k∈N Ak
∪ Ak
k∈N

Thus, from (5.27), (5.26) and (5.23), we finally conclude that

ˆ ˆ
−2n n−m −1 m
t H (A ∩ f {y}) dL (y) ≤ Jf dLn ≤
Rm
ˆ A

≤ t2n Hn−m (A ∩ f −1 {y}) dLm (y).

Rm
 5.3. THE COAREA FORMULA 163

Since t > 1 was arbitrarily taken, making t ↓ 1 it follows that

ˆ ˆ
n−m −1 m
H (A ∩ f {y}) dL (y) = Jf dLn ,
Rm A
which concludes the proof of case 1.
8) Case 2: A ⊂ Jf0 .
Note that both members in (5.21) are σ-additive on σ(Ln ) (for
the second member, that is a consequence of lemma 5.45 and of the
monotone convergence theorem 1.62). By the σ-finiteness of Ln , we
may therefore assume that Ln (A) < ∞.
Fix  > 0 and define g : Rn × Rm → Rm by g(x, y) := f (x) + y,
q : Rn × Rm → Rn and p : Rn × Rm → Rm the projections on
the first and second factors, respectively. Then g is Lipschitz and
∀ (x, y) ∈ Dg = Df × Rm , Dg(x, y) = Df (x) ◦ q + p. Hence, the
transpose of the Jacobian matrix [Dg(x, y)] is the (n+m)×m matrix
written in block form as
[Df (x)∗ ]
 
∗
[Dg(x, y) ] = ,
Im
i.e. it is of the same form of (5.14), exchanging m with n. Since
the i-th row of the matrix [Df (x)∗ ] is the i-th partial derivative of
f at x, i.e. Df (x) · ei , the norm of such row is ≤ Lip f . Therefore,
with exactly the same argument used in page 146 (case 2 of the area
formula), i.e. using Binet-Cauchy formula 5.23, we conclude that,
∀ (x, y) ∈ Dg = Df × Rm ,
Jg(x, y) ≥ m > 0,
   

2
2 n+m n
2
Jg(x, y) ≤ Jf (x) + −  · max{1, (Lip f )m−1 }2 .
m m
In particular, if (x, y) ∈ A × Rm ⊂ Jf0 × Rm , we conclude that
Jg(x, y) ≤ C, where
s   
n+m n
C := − max{1, (Lip f )m−1 }.
m m
Hence, ∀ (x, y) ∈ A×Rm , 0 < Jg(x, y) ≤ C; in particular, A×Rm ⊂
Jg+ , so that we may apply case 1 with g in place of f and any Ln+m -
measurable subset of A × Rm in place of A.
9) Recall that, by lemma 5.45, the map N (f |A ) : Rm → [0, ∞] given
by N (f |A )(y) = Hn−m (A ∩ f −1 {y}) is Lm -measurable. Since η :
Rm × Rm → Rm given by η(y, w) := y − w is linear and surjective,
it is measurable as a map (R2m ≡ Rm × Rm , L R2m ) → (Rm , L Rm )
(because it may be factored as η = q ◦ ψ, where ψ ∈ GL(R2m )
 164 5. AREA AND COAREA FORMULAS

and q : Rm × Rm → Rm is the projection on the first factor, and

linear isomorphisms preserve the Lebesgue σ-algebra). Therefore,
the composite map N (f |A ) ◦ η is ≥ 0 and L2m -measurable, and
so is the map Rm × Rm → [0, ∞] given by (y, w) 7→ χB(0,1) (w) ·
N (f |A )(y−w). As L2m = Lm ×Lm (by example 1.86), that justifies
the application of Fubini-Tonelli’s theorem 1.84 in the computation
below:
(5.28) ˆ
1.4
∀w ∈ Rm , Hn−m (A ∩ f −1 {y}) dLm (y) =
ˆ Rm

= Hn−m (A ∩ f −1 {y − w}) dLm (y) =

Rm
ˆ ˆ
1
= Hn−m (A ∩ f −1 {y − w}) dLm (y) dLm (w) =
α(m) B(0,1) Rm
ˆ ˆ
1 Fubini 1.84
= χB(0,1) (w) · Hn−m (A ∩ f −1 {y − w}) dLm (y) dLm (w) =
α(m) Rm Rm
ˆ ˆ
1
= χB(0,1) (w) · Hn−m (A ∩ f −1 {y − w}) dLm (w) dLm (y) =
α(m) Rm Rm
ˆ ˆ
1
χB(0,1) (w) · Hn−m (A ∩ f −1 {y − w}) × {w} dLm (w) dLm (y),


=
α(m) Rm Rm
where the last equality is due to corollary 2.5 with the isometry
Rn → Rn × Rm given by x 7→ (x, w).
10) Note that, if x ∈ Rn and w, y ∈ Rm , we have (x ∈ A and g(x, w) =
y) iff (x ∈ A and f (x) = y − w) iff x ∈ A ∩ f −1 {y − w}. Therefore,
defining B := A × B(0, 1) ⊂ Rn × Rm , the following equality holds:
(
∅ if w 6∈ B(0, 1)
B ∩ g −1 {y} ∩ p−1 {w} = −1
(A ∩ f {y − w}) × {w} if w ∈ B(0, 1).
Thus, ∀(y, w) ∈ Rm × Rm ,
χB(0,1) (w)·Hn−m (A∩f −1 {y−w})×{w} = Hn−m (B∩g −1 {y}∩p−1 {w}).

It then follows from (5.28) that

ˆ
Hn−m (A ∩ f −1 {y}) dLm (y) =
Rm
ˆ ˆ
1
= Hn−m (B ∩ g −1 {y} ∩ p−1 {w}) dLm (w) dLm (y).
α(m) Rm Rm
To continue this computation, we apply lemma 5.44 to the inner
integral, with B ∩ g −1 {y} ∈ σ(Ln+m ) in place of A, p : Rn × Rm →
 5.3. THE COAREA FORMULA 165

Rm in place of f , k = n − m and l = m, which yields

ˆ
Hn−m (A ∩ f −1 {y}) dLm (y) ≤
Rm
ˆ
1 α(n − m)α(m) by case 1
= Hn (B ∩ g −1 {y}) dLm (y) =
α(m) α(n) Rm
ˆ
α(n − m) part 8
= Jg dLn+m ≤
α(n) B
α(n − m)
≤ C · Ln+m (B) =
α(n)
α(n − m)α(m)
= C · Ln (A).
α(n)
Since Ln (A) < ∞ (by the reduction in the first part of step 8)
and  > 0 was arbitrarily taken, making  ↓ 0 yields
ˆ ˆ
n−m −1 m
H (A ∩ f {y}) dL (y) = 0 = Jf dLn ,
Rm A
which concludes the proof of case 2 and the thesis follows.

Corollary 5.50 (curvilinear Fubini-Tonelli’s theorem). Let f :
Rn → Rm be Lipschitz, n ≥ m. Then for all g : Rn → R Ln -measurable
with g ≥ 0 or g summable,
ˆ ˆ ˆ 
n
(5.29) g Jf dL = g(x) dHn−m (x) dLm (y),
Rn Rm f −1 {y}

meaning that the iterated integrals in second member make sense and
the equality holds.
Proof. Suppose that g ≥ 0. By exercise 1.54, there exists a se-
quence (Ai )i∈N in σ(Ln ) such that
∞
X 1
g= χAi .
i=1
i
m
Hence, for all y ∈ R ,
∞
X 1
g · χf −1 {y} = χAi ∩f −1 {y} .
i=1
i
It follows from lemma 5.45 that, for Lm -a.e. y ∈ Rm , ∀i ∈ N, χAi ∩f −1 {y}
is Hn−m -measurable; for such y, theorem 1.41 ensures that g · χf −1 {y}
is Hn−m -measurable and ≥ 0, so that the inner integral in the second
166 5. AREA AND COAREA FORMULAS

member of (5.29) makes sense. Moreover, still for y satisfying the above
condition, it follows from the monotone convergence theorem 1.62 that
ˆ ∞
X 1 n−m
g(x) · χf −1 {y} (x) dH n−m
(x) = H (Ai ∩ f −1 {y}).
i=1
i
Since the second member of the above equality defines a Lm -measurable
function Rm → [0, ∞] (in view of lemma 5.45 and theorem´ 1.41),
we conclude that the Lm -a.e. defined positive function y 7→ g(x) ·
χf −1 {y} (x) dHn−m (x) is Lm -measurable and
ˆ ˆ 
g(x) dHn−m (x) dLm (y) =
Rm f −1 {y}
ˆ ∞
X 1 n−m MCT 1.62
= H (Ai ∩ f −1 {y}) dLm (y) =
Rm i=1 i
∞ ˆ
X 1 CAF 5.48
= Hn−m (Ai ∩ f −1 {y}) dLm (y) =
i=1
i Rm
∞ ˆ
X 1 MCT 1.62
= χAi Jf dLn =
i=1
i
ˆ X∞ ˆ
1 n
= χAi Jf dL = g Jf dLn ,
i=1
i Rn

thus proving the case in which g ≥ 0. For g : Rn → R Ln -summable,

we apply the case just proved to the positive and negative parts of
g. 
Exercise 5.51 (Coarea Formula for locally Lipschitz maps). The
coarea formula and its corollary remain valid for locally Lipschitz maps
defined on open subsets of Rn . That is, let n ≥ m, Ω ⊂ Rn open and
f : Ω → Rm locally Lipschitz.
a) (coarea formula) For all Ln -measurable A ⊂ Ω,
• for Lm -a.e. y ∈ Rm , f −1 {y} ∩ A is Hn−m -measurable;
• the function N (f |A ) : Rm → [0, ∞], y 7→ Hn−m (A ∩ f −1 {y}), is
Lm -measurable and
ˆ ˆ
n
Jf dL = Hn−m (A ∩ f −1 {y}) dLm (y).
A Rm
b) (curvilinear Fubini-Tonelli’s theorem) If g : Ω → R is Ln |Ω -measurable
and g ≥ 0 or g ∈ L1 (Ln |Ω ), then
ˆ ˆ ˆ 
n n−m
g Jf dL = g(x) dH (x) dLn (y),
Ω Rm f −1 {y}
 5.3. THE COAREA FORMULA 167

meaning that the iterated integrals in the second member make

sense and the equality holds.
5.3.1. Applications of the coarea formula.
Proposition 5.52 (polar coordinates). If g : Rn → R is Ln -
measurable and g ≥ 0 or g ∈ L1 (Ln ), then
ˆ ˆ ∞ ˆ 
n n−1
(5.30) g dL = g dH dr.
Rn 0 ∂B(0,r)

Proof. Let f : Rn → R be given by f (x) = kxk. Then f is

Lipschitz and ∀ x ∈ Df = Rn \ {0}, ∇f (x) = x/kxk, hence Jf (x) =
k∇f (x)k = 1. Since ∀r ∈ R, f −1 {r} = ∂B(0, r) (in particular, = ∅ for
r < 0), (5.30) is a direct consequence of corollary 5.50. 
Proposition 5.53. Let Ω ⊂ Rn open and f : Ω → R be locally
Lipschitz. Then
ˆ ˆ ∞
n
k∇f k dL = Hn−1 ({f = t}) dt.
Ω −∞

Proof. It is a direct consequence of exercise 5.51.a), with A = Ω,

taking into account that Jf = k∇f k. 
 CHAPTER 6

Sobolev Spaces

In this chapter we study some basic theory of Sobolev spaces W1,p (Ω)
1,p
and Wloc (Ω) on open sets Ω ⊂ Rn . Our primary purpose is to develop
some background for the study of functions of bounded variation and
Cacciopoli sets in the subsequent chapters. For a more extensive treat-
ment on this subject, we refer the reader to standard textbooks — for
instance, [Maz11] or [AF03].
Recall the definitions of weak derivatives and sobolev functions in-
troduced in the previous chapter in 5.3 and 5.8. In order to introduce
1,p
vector space topologies on the spaces W1,p (Ω) and Wloc (Ω), we make
the following definition:

1,1
Definition 6.1. Let Ω ⊂ Rn open and f ∈ Wloc (Ω), i.e. f ∈
1 n
Lloc (L |Ω ) admits weak partial derivatives of first order. We define
´
• for 1 ≤ p < ∞, kf kW1,p (Ω) := ( Ω |f |p + k∇f kp dLn )1/p ∈ [0, ∞];
• for p = ∞, kf kW1,∞ (Ω) := |f | + k∇f k L∞ (Ω) ∈ [0, ∞].

Thus, with the notation above, ∀1 ≤ p ≤ ∞, f ∈ W1,p (Ω) iff kf kW1,p (Ω) <
∞, and it is clear that k·kW1,p (Ω) is a seminorm on W1,p (Ω). Similarly
to the discussion on Lp spaces, the linear subspace N := {f ∈ W1,p (Ω) |
kf kW1,p (Ω) = 0} of W1,p (Ω) consists of the measurable functions on Ω
which are null almost everywhere. Therefore, the quotient W1,p (Ω)/N
consists of classes of equivalence of functions in W1,p (Ω) which coincide
almost everywhere, and k·kW1,p (Ω) is a norm on this quotient, which
is complete by the following proposition. As in remark 1.61, we shall
henceforth overload the notation “W1,p (Ω)”, which will be used both
with its original meaning and also to denote the aforementioned quo-
tient space.

Proposition 6.2. Let Ω ⊂ Rn open. For 1 ≤ p ≤ ∞, W1,p (Ω)

is a Banach space (for p = 2, it is a Hilbert space). It is reflexive for
1 < p < ∞ and it is separable for 1 ≤ p < ∞.
169
170 6. SOBOLEV SPACES

Proof. Let H := Lp (Ln |Ω )×Lp (Ln |Ω , Rn ), which is a Banach space

with the norm
 ´ |f |p + kF kp dLn 1/p
 
Ω
for 1 ≤ p < ∞,
k(f, F )k :=
 |f | + kF k for p = ∞.
L∞ (Ω)

The fact that H is indeed a Banach space, reflexive for 1 < p < ∞ and
separable for 1 ≤ p < ∞ follows from the corresponding properties of
the spaces Lp (Ln |Ω ) and Lp (Ln |Ω , Rn ) ≡ Lp (Ln |Ω )n and from the fact
that the topology of H is the product topology. Moreover, for p = 2, the
norm defined above is induced by the inner product h(f, F ), (g, G)i :=
hf, giL2 (Ln |Ω ) + hF, GiL2 (Ln |Ω ,Rn ) , hence it is a Hilbert space in that case.
We contend that the graph of the weak gradient operator ∇w :
W1,p (Ω) → Lp (Ln |Ω , Rn ) is a closed subspace of H. Indeed, let (uk , vk )k∈N
be a sequence in gr ∇w such that (uk , vk ) → (u, v) ∈ H. We must show
that u is weakly differentiable and ∇w u = v. Indeed, ∀ϕ ∈ C∞ n
c (Ω, R ),
∀k ∈ N, ˆ ˆ
uk div ϕ dL = − hvk , ϕi dLn .
n
Ω Ω
Since uk → u in Lp (Ln |Ω ) and vk → v in in Lp (Ln |Ω , Rn ), the above
equality holds with u in place of uk and v in place of vk , thus proving
our contention.
As a closed subspace of H, gr ∇w is also a Banach space (Hilbert
for p = 2), reflexive for 1 < p < ∞ and separable for 1 ≤ p < ∞. Since
the projection on the first factor gr ∇w → W1,p (Ω) is a linear isometry
onto W1,p (Ω) endowed with the norm defined in 6.1 (in other words,
that norm is the “graph norm”), the thesis follows. 
Proposition 6.3. Let 1 ≤ p ≤ ∞, Ω ⊂ Rn open and (uk )k∈N a
sequence in W1,p (Ω).
i) If (uk )k∈N is Lp -convergent to u ∈ Lp (Ln |Ω ) and (∇uk )k∈N is Lp -
convergent to v ∈ Lp (Ln |Ω , Rn ), then u ∈ W1,p (Ω) and ∇w u = v.
ii) If 1 < p ≤ ∞, (uk )k∈N is Lp -convergent to u ∈ Lp (Ln |Ω ) and the
sequence (∇uk )k∈N is bounded in Lp (Ln |Ω , Rn ), then u ∈ W1,p (Ω).
Proof. With the notation from the previous proof, the first asser-
tion is a direct consequence of the fact that gr ∇w is closed in H.
As to the second assertion, let 1 < p ≤ ∞ and q the conjugate
exponent of p. It follows from the Riesz representation theorem 1.79
that Lp (Ln |Ω , Rn ) is the dual of Lq (Ln |Ω , Rn ); hence, by the fact that
Lq (Ln |Ω , Rn ) is separable (since 1 ≤ q < ∞) and by the Banach-Alaoglu
theorem, strongly closed balls in Lp (Ln |Ω , Rn ) are sequentially com-
pact in the weak-star topology. Thus, passing to a subsequence if
 6. SOBOLEV SPACES 171

necessary, we may assume that (∇uk )k∈N is weakly-star convergent to

v ∈ Lp (Ln |Ω , Rn ). In particular, for every ϕ ∈ C∞c (Ω),
ˆ ˆ
u div ϕ dLn = lim uk div ϕ dLn =
k→∞ Ω
Ω
ˆ
= − lim h∇uk , ϕi dLn =
k→∞ Ω
ˆ
= − hv, ϕi dLn
Ω
thus showing that u is weakly differentiable and ∇w u = v ∈ Lp (Ln |Ω , Rn ),
hence u ∈ W1,p (Ω). 
In order to establish the locality of the weak derivative, we recall
from Advanced Calculus the construction of smooth partitions of unity
(skip to theorem 6.13 if you don’t need to recall that stuff).
Definition 6.4 (point-finite and locally finite families). Let X be
a topological space. We say that a family (Fα )α∈A of subsets of X is
• point-finite if ∀x ∈ X, {α ∈ A | x ∈ Fα } is finite;
• locally finite if ∀x ∈ X, there exists a neighborhood V of x
such that {α ∈ A | V ∩ Fα 6= ∅} is finite.
Remark 6.5. Let X be a topological space.
1) If X is second countable and (Fα )α∈A is a locally finite family of
subsets of X, then A is countable, since we may cover X by count-
ably many open sets, each of which intersects subsets in the family
for at most finitely many indices.
2) It is clear that every locally finite family of subsets of X is point-
finite.
3) It is also clear that, if K ⊂ X is compact and (Fα )α∈A is a locally
finite family of subsets of X, then {α ∈ A | K ∩ Fα 6= ∅} is finite.
Definition 6.6 (smooth partitions of unity on open sets of Rn ).
Let Ω ⊂ Rn open. A smooth partition of unity of Ω is a family (ξα )α∈A
such that:
PU1) ∀α ∈ A, ξα ∈ C∞ (Ω, [0, 1]) and (spt ξα )α∈A is a locally finite
family ofPsubsets of Ω;
PU2) ∀x ∈ Ω, α∈A ξα (x) = 1.
If F is an open cover of Ω, we say that a smooth partition of unity
(ξα )α∈A of Ω is subordinate to F if ∀α ∈ A, there exists U ∈ F such
that spt ξα ⊂ U .
With the notation above, note that, ∀x ∈ Ω, the sum in PU2) is
finite. Actually, thanks to PU1), there exists a neighborhood V ⊂ Ω
172 6. SOBOLEV SPACES

PV := {α ∈ A | spt ξα ∩ V 6= ∅} is finite; hence, ∀y ∈ V ,

of x such that A
P
ξ
α∈A α (y) = α∈AV ξα (y) is a finite sum.
We will need the following theorem from General Topology:
Theorem 6.7 (shrinking lemma). Let X be a normal topological
space and (Uα )α∈A a point-finite open cover of X. Then there exists an
open cover (Vα )α∈A of X such that, ∀α ∈ A, Vα ⊂ Uα .
Proof. We refer the reader to [Eng89], theorem 1.5.18, page 44.

Theorem 6.8 (existence of partitions of unity on open sets of Rn ).
Let Ω ⊂ Rn open and (Uα )α∈A a locally finite open cover of Ω with
∀α ∈ A, Uα b Ω. Then there exists a smooth partition of unity (ξα )α∈A
of Ω such that, ∀α ∈ A, spt ξα b Uα .
Proof. Apply the shrinking lemma 6.7 to the locally finite (hence
point-finite) open cover (Uα )α∈A of Ω endowed with the relative topol-
ogy. We obtain an open cover (Vα )α∈A of Ω such that, ∀α ∈ A,
Ω Ω
Vα = Vα ∩ Ω ⊂ Uα ; in particular, since Uα b Ω, Vα = Vα is a
compact subset of Uα . Then, for each α ∈ A, we may apply the differ-
entiable Urysohn’s lemma 1.114 to obtain ψα ∈ C∞ c (Uα , [0, 1]) such that
ψα ≡ 1 onPVα . Since (spt ψα )α∈A is a locally finite family of subsets of
Ω, ψ := α∈A ψα is a real-valued smooth function on Ω; it is strictly
positive, because (Vα )α∈A cover Ω. Define, ∀α ∈ A,
ψα
ξα := .
ψ
Then ∀α ∈ A, ξα ∈ C∞ c (Ω), spt ξα P
b Uα , (spt ξα )α∈A is locally finite
family of subsets of Ω and ∀x ∈ Ω, α∈A ξα (x) = 1. 
Corollary 6.9. Let Ω ⊂ Rn be open and F an open cover of Ω.
Then there exists a partition of unity (ξα )α∈A of Ω subordinate to F
such that ∀α ∈ A, spt ξα is compact.
Proof. Take a refinement of F formed by open sets with compact
closures in Ω and then a locally finite open refinement G of the latter
cover, which exists, due the paracompactness of Ω. Then apply theorem
6.8 to Ω with the cover G. 
Corollary 6.10. Let K ⊂ Rn be compact and (Ui )1≤i≤N a cover
of K by open subsets of Rn . Then there exists (ξi )1≤i≤N such that
PN
∀1 ≤ i ≤ N , ξi ∈ C∞
c (Ui , [0, 1]) and i=1 ξi ≡ 1 on K.

Proof. Let Ω := ∪1≤i≤N Ui and apply the previous corollary to

obtain a smooth partition of unity (ηα )α∈A of Ω subordinate to the
 6. SOBOLEV SPACES 173

cover (Ui )1≤i≤N such that ∀α ∈ A, spt ηα is compact. Since (spt ηα )α∈A
is a locally finite family of subsets of Ω and K is a compact subset of
Ω, the set AK := {α ∈ A | spt ηα ∩ K 6= ∅} is finite. Since the partition
of unity is subordinate to (Ui )1≤i≤N , for each α ∈ AK we may choose
i(α) ∈ {1, . . . , N } such spt ηα ⊂ Ui(α) . Define, for 1 ≤ i ≤ N ,
X
ξi := ηα ,
{α∈AK |i(α)=i}

recalling that the sum over the empty family is 0.


Corollary 6.11. Let Ω ⊂ Rn be open and F an open cover of Ω.
Then there exists a partition of unity (ξV )V ∈F of Ω strictly subordinate
to F, i.e. such that for all V ∈ F, spt ξV ⊂ V .
In general, it is not possible to choose such a strictly subordinate
partition of unit with compact supports, i.e. for each V ∈ F, the
support of ξV may be not compact.
Proof. We may apply corollary 6.9 to obtain a smooth partition
of unity (ηα )α∈A of Ω subordinate to the cover F such that ∀α ∈ A,
spt ηα is compact. For each α ∈ A we may choose V (α) ∈ F such
spt ηα ⊂ V (α). Define, for V ∈ F,
X
ξV := ηα .
{α∈A|V (α)=V }

Since the above sum isP locally finite, for each V ∈ F, ξV is smooth
with 0 ≤ ξV ≤ 1 and V ∈F ξV ≡ 1 on Ω. Moreover, as the family
(spt ηα )α∈A is locally finite, for all V ∈ F, ∪V (α)=V spt ηV (α) ⊂ V is
closed in Ω (because it is the union of a locally finite family of closed
subsets of Ω), hence spt ξV = ∪V (α)=V spt ηV (α) ⊂ V . That is, (ξV )V ∈F
is a smooth partition of unity of Ω with spt ξV ⊂ V for all V ∈ F. 
Exercise 6.12 (differentiable Urysohn’s lemma, part II). Let F0
and F1 be disjoint closed subsets of Rn . There exists a smooth function
ξ ∈ C∞ (Rn ) such that ξ ≡ 0 on F0 and ξ ≡ 1 on F1 .
Theorem 6.13 (locality of the weak derivative). Let Ω ⊂ Rn , f ∈
L1loc (Ln |Ω ) and F ⊂ 2Ω an open cover of Ω. Then f admits weak partial
derivatives of first order on Ω iff ∀U ∈ F, f |U admits weak partial
derivatives of first order on U . Moreover, weak derivatives commute
with restrictions.
Lemma 6.14. Let U ⊂ Rn open, 1 ≤ p ≤ ∞ and ξ ∈ C∞
c (U ).
174 6. SOBOLEV SPACES

i) If f ∈ Lploc (Ln |U ), then ξ · f (defined as 0 on Rn \ U ) belongs to

Lp (Ln ).
1,p
ii) If f ∈ Wloc (U ), then ξ · f ∈ W1,p (Rn ) and, ∀1 ≤ i ≤ n,
∂ w (ξ · f ) ∂ξ ∂wf
= ·f +ξ· .
∂xi ∂xi ∂xi
Proof.
i) ξ · f is clearly ´Ln -measurable (for instance, ´ by proposition 1.50).
If 1 ≤ p < ∞, Rn |ξ · f | dL ≤ kξku spt ξ |f | dLn < ∞; if p = ∞,
p n p p

kξ · f k∞ ≤ kξku kf kL∞ (Ln |spt ξ ) < ∞.

ii) For all 1 ≤ i ≤ n, it follows from the previous item that both ξ · f
∂ξ w
and g := ∂x i
· f + ξ · ∂∂xfi belong to Lp (Ln ). It therefore suffices
to show that ξ · f admits weak i-th partial derivative equal to g.
Indeed, ∀ϕ ∈ C∞ n
c (R ),
ˆ ˆ  ∂(ξ · ϕ)
∂ϕ n ∂ξ 
(ξ · f ) · dL = f· − · ϕ dLn =
∂xi ∂xi ∂xi
Rn Ω
ˆ  w
∂ f ∂ξ 
=− ξ· ·ϕ+f · · ϕ dLn =
∂xi ∂xi
ˆ Ω

=− g · ϕ dLn ,
Rn
as asserted.

Proof of theorem 6.13. The implication “⇒” and the fact that
weak derivatives commute with restrictions are clear. We must prove
the converse implication, i.e. if 1 ≤ i ≤ n and ∀U ∈ F, f |U admits
weak i-th partial derivative ∂i (f |U ) ∈ L1loc (Ln |U ), then f admits weak
i-th partial derivative on Ω.
1) We may assume that F is locally finite and ∀U ∈ F, U b Ω.
Indeed, in the general case, take a locally finite open refinement G
of F such that ∀U ∈ G, U b Ω. For each V ∈ G, there exists U ∈ F
such that V ⊂ U ; since f |U admits weak i-th partial derivative
∂i (f |U ) ∈ L1loc (Ln |U ), it follows that f |V = (f |U )|V admits weak i-th
partial derivative, so that we may replace F by G.
2) Take a smooth partition of unity (ξV )V ∈F of Ω, given by theorem
∞
P such that ∀V ∈1 F,n ξV ∈ Cc (V ). We contend
6.8,
n
that g :=
ξ
V ∈F V i∂ (f |V ) ∈ L loc (L |Ω ). Indeed, g is clearly L -measurable
and, for each compact K ⊂ Ω, there areP finitely many V1 , . . . , VN ∈
F which intersect K, so that |g|χK ≤ N 1 n
j=1 ξVj |∂i (f |Vj )| ∈ L (L )
by lemma 6.14.i), thus proving our contention.
 6. SOBOLEV SPACES 175

3) Let ϕ ∈ C∞ c (Ω). Since spt ϕ is compact, there are finitely many

V1 , . . . , VN ∈ F which intersect K. We then have
ˆ XN ˆ
n
g ϕ dL = ξj ∂i (f |Vj )ϕ dLn =
Ω j=1 Ω

N ˆ
X
= ∂i (f |Vj ) · (ξj ϕ) dLn =
j=1 Vj

N ˆ
X
=− f |Vj · ∂i (ξj ϕ) dLn =
j=1 Vj

N ˆ
X
=− f · ∂i (ξj ϕ) dLn =
j=1 Ω
ˆ XN
=− f · ∂i ( ξj ϕ) dLn =
Ω j=1
ˆ
=− f · ∂i (ϕ) dLn ,
Ω
thus proving that ∂iw f = g on Ω.

Corollary 6.15. Let Ω ⊂ Rn open, 1 ≤ p ≤ ∞ and f : Ω → R
1,p
Lebesgue measurable. Then f ∈ Wloc (Ω) iff for all open V b Ω, f |V ∈
1,p
W (V ).
Proof. The implication “⇒” is clear, in view of the fact that weak
derivatives commute with restrictions. Conversely, assume that for all
open V b Ω, f |V ∈ W1,p (V ). In particular, for all open V b Ω,
f |V ∈ Lp (Ln |V ), hence f ∈ Lploc (Ln |Ω ) ⊂ L1loc (Ln |Ω ). It then follows
from theorem 6.13 that ∃ ∇w f ∈ L1loc (Ln |Ω , Rn ); besides, for all open
V b Ω, (∇w f )|V = ∇w (f |V ) ∈ Lp (Ln |V , Rn ). That is, f ∈ Lploc (Ln |Ω )
and ∇w f ∈ Lploc (Ln |Ω , Rn ), i.e. f ∈ Wloc
1,p
(Ω). 
Corollary 6.16. Let Ω ⊂ Rn open and f : Ω → Rn locally Lip-
1,∞
schitz. Then f ∈ Wloc (Ω), f is differentiable in the sense of Fréchet
L -a.e. on Ω and ∇ f = ∇f Ln -a.e. on Ω.
n w

Proof. We have already proved in corollary 5.14 that f is differ-

entiable in the sense of Fréchet Ln -a.e. on Ω. It therefore suffices to
show, in view of the locality of both the weak derivative 6.13 and of the
classical Fréchet derivative, that for each open V b Ω on which f has
Lipschitz restriction, f |V ∈ W1,∞ (V ) and ∇w f = ∇f Ln -a.e. on V .
176 6. SOBOLEV SPACES

Indeed, by McShane’s theorem 5.1 we may extend f |V to a Lipschitz

map Rn → R, for which proposition 5.9 and theorem 5.12 apply. 

6.1. Approximation by smooth functions, part I

In this section we fix an open set Ω ⊂ Rn . Our goal is to derive
theorems on approximation of Sobolev functions on Ω by smooth func-
tions. In order to accomplish that, it will be convenient to generalize
the operation of convolution with the standard mollifier (φt )t>0 to func-
tions f : Ω → R. One possible approach is to proceed like we did in
the proof of proposition 5.7. Another approach, which we adopt here,
is to define the regularized functions ft = φt ∗ f on smaller subsets Ωt
of Ω, cf. definition 6.17 below. We call the reader’s attention to the
fact that, in general, it is not possible to simply extend f by zero on
Rn \ Ω and then regularize the extension f¯: it might be the case that
f¯ 6∈ L1loc (Rn ), so that “φt ∗ f¯ ” would not be defined.
Definition 6.17. For each t > 0, let
Ωt := {x ∈ Rn | B(x, t) ⊂ Ω} = {x ∈ Rn | d(x, Ωc ) > t},
so that (Ωt )t>0 is a family of open subsets of Ω which increases to Ω as
t ↓ 0.
Let (φt )t>0 be the standard mollifier in Rn . For each t > 0 and
f ∈ L1loc (Ln |Ω ), we define ft : Ωt → R by, ∀x ∈ Ωt ,
ˆ
ft (x) := (φt ∗ f )(x) = f (y)φt (x − y) dLn (y).
B(x,t)

We call ft the t-approximation or t-regularization of f .

Note that: 1) ft (x) is well-defined since, for x ∈ Ωt , B(x, t) ⊂ Ω; 2)
if Ω = Rn , then Ωt = Rn for all t > 0.
Notation. If f : Ω → R, we denote by f¯ : Rn → R the extension
of f by zero on Rn \ Ω.
In order to simplify the notation, sometimes we omit the bar from
the extension, whenever no confusion arises.
Definition 6.18 (convergence in the sense of Lploc and Wloc 1,p
). Let
n
1 ≤ p ≤ ∞, f : Ω → R L -measurable and, for each k ∈ N, let
fk : dom fk ⊂ Ω → R be Ln -measurable.
• We say that (fk )k∈N converges to f in the sense of Lploc (Ln |Ω )
(notation: “fk → f in Lploc (Ln |Ω )”) if, for all open V b Ω, there
exists k0 ∈ N (possibly depending on V ) such that ∀k ≥ k0 ,
V ⊂ dom fk and kfk − f kLp (Ln |V ) → 0.
 6.1. APPROXIMATION BY SMOOTH FUNCTIONS, PART I 177

• If ∀k ∈ N, dom fk is open, f and fk belong to L1loc on their

domains and admit weak partial derivatives of first order, we
1,p
say that (fk )k∈N converges to f in the sense of Wloc (Ω) (no-
1,p
tation: “fk → f in Wloc (Ω)”) if, for all open V b Ω, there
exists k0 ∈ N (possibly depending on V ) such that ∀k ≥ k0 ,
V ⊂ dom fk and kfk − f kW1,p (V ) → 0.
We make similar definitions of convergence in the sense of Lploc or in the
1,p
sense of Wloc for a family (f )>0 in place of (fk )k∈N .

Remark 6.19. Concerning the previous definition:

(1) What we have in mind is the family (ft )t>0 of the regularized
functions of some f ∈ L1loc (Ln |Ω ), cf. definition 6.18.
(2) For a sequence (fk )k∈N in Lploc (Ln |Ω ) and f ∈ Lploc (Ln |Ω ), the
convergence defined above coincides with the convergence in
the natural topology of Lploc (Ln |Ω ), which is a Fréchet space
topology induced by the family of seminorms {k·kLp (Ln |V ) |
V b Ω open}.
1,p 1,p
(3) Similarly, for a sequence (fk )k∈N in Wloc (Ω) and f ∈ Wloc (Ω),
the convergence defined above coincides with the convergence
1,p
in the natural topology of Wloc (Ω), which is a Fréchet space
topology induced by the family of seminorms {k·kW1,p (V ) | V b
Ω open}.

Theorem 6.20 (mollifiers, part II). Let f ∈ L1loc (Ln |Ω ), (φ )>0 the
standard mollifier and f = φ ∗ f : Ω → R the -approximation of f ,
cf. definition 6.17.
i) ∀ > 0, f ∈ C∞ (Ω ). ´ ´
ii) ∀ > 0, ∀ϕ ∈ C0c (Ω ), f ϕ dLn = f ϕ dLn .
iii) lim→0 f (x) = f (x) if x ∈ Ω is a Lebesgue point of f ; in particular,
f → f Ln -a.e. on Ω.
iv) If f ∈ C(Ω), f → f uniformly on compact subsets of Ω.
v) If f ∈ Lploc (Ln |Ω ) for some 1 ≤ p < ∞, then f → f in the sense
of Lploc (Ln |Ω ).
1,p
vi) If f ∈ Wloc (Ω) for some 1 ≤ p ≤ ∞, then ∀ > 0, ∀1 ≤ i ≤ n,

∂f ∂wf  ∂wf 

= φ ∗ =
∂xi ∂xi ∂xi 
on Ω .
1,p
vii) In particular, if f ∈ Wloc (Ω) for some 1 ≤ p < ∞, then f → f in
1,p
the sense of Wloc (Ω).
178 6. SOBOLEV SPACES

We interpret the theorem above by saying that the -regularized

functions of f approximate f in the “natural topology” of its class of
regularity.

Proof.
i) That is a direct consequence of the dominated convergence theo-
rem, like we argued in the proof of proposition 1.108.j). Indeed, let
x0 ∈ Ω and r > 0 such that U(x0 , r) ⊂ Ω . Then, ∀x ∈ U(x0 , r),
ˆ
f (x) = φ (x − y)f (y) dLn (y),
K

where K := B(x0 , r + ) b Ω. It then follows that, for each multi-

index α ∈ Zn+ , the integral which results from the second member
above by derivation under the integral sign (to be justified) is
ˆ
∂ α (φ )(x − y)f (y) dLn (y),
K

whose integrand is dominated in absolute value by k∂ α φ ku |f | K ∈

L1 (Ln |K ). Therefore, we may differentiate successively under the
integral sign by means of the dominated convergence theorem, cf.
proposition 1.67, to conclude that ∀x ∈ U(x0 , r), ∀α ∈ Zn+ ,
ˆ
α
∃∂ (f )(x) = ∂ α (φ )(x − y)f (y) dLn (y).
K

Since x0 ∈ Ω was arbitrarily taken, we have proved that f has

partial derivatives of all orders on all points of Ω .
ii) If  > 0 and ϕ ∈ C0c (Ω ), we have:
ˆ ˆ ˆ
n Fubini 1.84
f ϕ dL = ϕ(x) f (y)φ (x − y) dy dx =
spt ϕ Ω
ˆ ˆ
φ(z)=φ(−z)
= f (y) φ (x − y)ϕ(x) dx dy =
Ω spt ϕ
ˆ ˆ
= f (y) φ (y − x)ϕ(x) dx dy =
Ω spt ϕ
ˆ
= f (y)ϕ (y) dy,
 6.1. APPROXIMATION BY SMOOTH FUNCTIONS, PART I 179

where the application of Fubini’s theorem is justified by the fact

that

ˆ
Tonelli 1.84
|f (y)| · φ (x − y) · |ϕ(x)| d(Ln × Ln )(x, y) =
spt ϕ×Ω
ˆ ˆ
= |f (y)| |ϕ(x)|φ (y − x) dx dy ≤
Ω spt ϕ
ˆ
≤ kϕku kφ ku |f (y)| dy < ∞.
spt ϕ+B(0,)

iii) Let x ∈ Ω be a Lebesgue point of f . Take δ > 0 such that x ∈ Ωδ .

Then, ∀ 0 <  ≤ δ:

ˆ
f (x) − f (x) = φ (x − y)[f (y) − f (x)] dLn (y) =
ˆ
B(x,)
1 x−y
= φ( )[f (y) − f (x)] dLn (y),
B(x,) n 

so that

ˆ
1 →0
|f (x)−f (x)| ≤ kφku α(n)· n
|f (y)−f (x)| dLn (y) → 0.
L B(x, ) B(x,)

iv) Let K be a compact subset of Ω. Take δ > 0 such that Kδ :=

K+B(0, δ) b Ω. Let F := f · χKδ : Rn → R. Since f is bounded on
Kδ (because f is continuous and Kδ ⊂ Ω is compact), F ∈ L∞ (Ln );
moreover, since f coincides with F on Kδ , F is continuous on
(Kδ )o . It then follows from theorem 1.111.iii) that φ ∗ F → F
uniformly on compact subsets of (Kδ )o . On the other hand, as
F |Kδ = f |Kδ , we conclude that, ∀0 <  ≤ δ, φ ∗ F = f on K,
whence f → f uniformly on K.
v) Let K be a compact subset of Ω. Take δ > 0 such that Kδ :=
K + B(0, δ) b Ω. Let F := f · χKδ : Rn → R. Since Kδ is a
compact subset of Ω, f |Kδ ∈ Lp (Ln |Kδ ), which implies F ∈ Lp (Ln ).
It then follows from theorem 1.111.i) that φ ∗ F → F in Lp (Ln ).
On the other hand, as F |Kδ = f |Kδ , we conclude that, ∀0 <  ≤ δ,
φ ∗ F = f on K, whence kf − f kLp (Ln |K ) → 0.
180 6. SOBOLEV SPACES

vi) ∀ > 0, ∀1 ≤ i ≤ n, ∀ϕ ∈ C∞ c (Ω ),

ˆ ˆ
∂ϕ n (ii) ∂ϕ 1.108
f dL = f φ ∗ dLn =
∂xi ∂xi
ˆ
∂(φ ∗ ϕ)
= f dLn =
∂xi
ˆ w
∂ f (ii)
=− ϕ dLn =
∂x
ˆ  wi 
∂ f
=− ϕ dLn ,
∂xi 
 w 
∂f ∂ f
thus showing that ∂x i
= ∂xi
.

One alternative to the above proof is to use the formula ob-
∂f
tained in (i) for ∂x i
, i.e. ∀x ∈ Ω ,
ˆ
∂f ∂(φ )
(x) = (x − y)f (y) dy =
∂xi Ω ∂xi
ˆ
∂ h i φ (x−·)∈C∞ (Ω)
=− φ (x − ·) f (y) dy = c
Ω ∂xi y
ˆ
∂wf
= φ (x − y) (y) dy =
Ω ∂xi
∂wf
= φ ∗ (x).
∂xi
 w 
p
vii) f → f in Lloc (L |Ω ) by part v) and ∀1 ≤ i ≤ n, ∂xi = ∂∂xfi
n ∂f
→

∂wf
∂xi
in Lploc (Ln |Ω ) by parts vi) and v).

Corollary 6.21. Let 1 ≤ p < ∞, (φt )t>0 the standard mollifier
and f ∈ W1,p (Rn ). Then:
i) ∀ > 0, f = φ ∗ f ∈ C∞ (Rn ) ∩ W1,p (Rn ) and f → f in W1,p (Rn )
as  → 0.
ii) There exists a sequence (fk )k∈N in C∞ n
c (R ) such that fk → f in
W1,p (Rn ).
Proof. It follows from theorem 6.20 with Ω = Rn that f ∈
C∞ (Rn ) and ∇(f ) = (∇w f ) . Since f ∈ Lp (Ln ) and ∇w f ∈ Lp (Ln , Rn ),
it follows from Young’s inequality 1.108.g) that f ∈ Lp (Ln ) and (∇w f ) ∈
Lp (Ln , Rn ), hence f ∈ W1,p (Rn ). Finally, from theorem 1.111.i), f →
f in Lp (Rn ) and ∇(f ) = (∇w f ) → ∇w f in Lp (Rn ), thus showing that
f → f in W1,p (Rn ).
 6.1. APPROXIMATION BY SMOOTH FUNCTIONS, PART I 181

It remains to prove the second assertion. Let ∀k ∈ N, gk :=

φ1/k ∗ f ∈ C∞ (Rn ) ∩ W1,p (Rn ), so that gk → f in W1,p (Rn ) by part
i). Choose ζ ∈ C∞ n
c (R , [0, 1]) such that ζ ≡ 1 on B(0, 1) and spt ζ ⊂
U(0, 2) (which exists, thanks to exercise 1.114). Define, ∀k ∈ N,
∞


ζk ∈ Cc U(0, 2k) by ζk (x) := ζ(x/k), and fk := ζk · gk . Then ∀k ∈ N,
fk ∈ C∞ n
c (R ). We will prove that fk → f in W (R ).
1,p n

1) For all u ∈ Lp (Ln ), ζk · u → u in Lp (Ln ). Indeed, |u − ζk · u|p → 0

pointwise and |u − ζk · u|p ≤ 2p |u|p ∈ L1 (Ln ), hence the dominated
convergence theorem 1.64 yields the assertion.
2) Since f − fk = (f − ζk · f ) + ζk · (f − gk ), we have kf − fk kp ≤
kf − ζk · f kp + kζk ku kf − gk kp → 0, since |ζk | ≤ 1, kf − ζk · f kp → 0
by the previous item and kf − gk kp ≤ kf − gk kW1,p (Rn ) → 0.
3) ∀x ∈ Rn ,
∇fk (x) = ∇ζk (x) · gk (x) + ζk (x) · ∇gk (x) =
1
= · ∇ζ(x/k) · gk (x) + ζk (x) · ∇gk (x).
k
Hence,
∇w f (x) − ∇fk (x) = ∇w f (x) − ζk (x) ∇w f (x)+
1
+ ζk (x)[∇w f (x) − ∇gk (x)] − · ∇ζ(x/k) · gk (x),
k
so that
1
k∇w f −∇fk kp ≤ k∇w f −ζk ∇w f kp +kζk ku k∇w f −∇gk kp + k∇ζku kgk kp .
k
w w w
Since k∇ f − ζk ∇ f kp → 0 by item 1), k∇ f − ∇gk kp ≤ kf −
gk kW1,p (Rn ) → 0 and, as kgk kp = kφ1/k ∗ f kp ≤ kf kp by Young’s
inequality 1.108.g), k1 k∇ζku kgk kp → 0, it follows that k∇w f −
∇fk kp → 0. We have thus proved that fk → f in Lp (Rn ) (by
the previous item) and ∇fk → ∇w f in Lp (Ln , Rn ); that is, fk → f
in W1,p (Rn ).

Exercise 6.22. If u ∈ C1 (Ω), the classical and weak gradients of
u coincide. The converse holds in the following sense: if u ∈ L1loc (Ω)
has weak gradient v ∈ C(Ω, Rn ), then u coincides Ln -a.e. in Ω with a
function ũ ∈ C1 (Ω).
Exercise 6.23 (Friedrichs’s theorem). Let 1 ≤ p < ∞. If u ∈
W (Ω), there exists a sequence (uk )k∈N in C∞
1,p n
c (R ) such that uk |Ω → u
p n w p n n
in L (L |Ω ) and ∇uk |Ω → ∇ u in Lloc (L |Ω , R ).
182 6. SOBOLEV SPACES

Theorem 6.24 (Meyers-Serrin’s theorem). Let 1 ≤ p < ∞ and

u ∈ W1,p (Ω). There exists a sequence (uk )k∈N in C∞ (Ω) ∩ W1,p (Ω) such
that uk → u in W1,p (Ω).

Proof.
1) Let (Uk )k∈N be a locally finite open cover of Ω such that ∀k ∈ N,
Uk b Ω. Take a smooth partition of unity (ξk )k∈N of Ω, with ∀k ∈ N,
spt ξk b Uk , given by theorem 6.8.
2) Fix  > 0 and k ∈ N. Let (φt )t>0 be the standard mollifier in Rn . It
follows from lemma 6.14.ii) that ξk · u ∈ W1,p (Rn ), with spt ξk · u ⊂
spt ξk b Uk . We may therefore apply proposition 1.108.d) and
corollary 6.21.i) to obtain tk sufficiently small so that φtk ∗(ξk · u) ∈
C∞ (Rn ) ∩ W1,p (Rn ) has compact support in Uk and kφtk ∗(ξk · u) −
ξk · ukW1,p (Rn ) < 2−k . Define u : Ω → R by

∞
X  
u := φtk ∗(ξk · u) |Ω .
k=1

Note that the above sum is locally finite, since spt φtk ∗(ξk ·u) k∈N is
a locally finite family of subsets of Ω (because ∀k ∈ N, spt φtk ∗(ξk ·
u) b Uk ). Hence, u ∈ C∞ (Ω)1. Similarly, we have locally finite
sums
∞
X

u − u = φtk ∗(ξk · u) − ξk · u
k=1
(6.1) ∞
X
∇w (u − u) = ∇w φtk ∗(ξk · u) − ξk · u ,


k=1

where the second equality holds because both members coincide on

each relatively compact open subset V b Ω (because on V the sum
is finite and the weak gradient is linear).
3) It follows from (6.1), definition 6.1 and from Minkowski’s inequality
for integrals 1.88 (with µ = Ln |Ω and ν the counting measure on N)

1note


that spt φtk ∗(ξk · u) k∈N is a locally finite family of subsets of Ω (with
the relative topology), but not, in general, a locally finite family of subsets of Rn ,
hence we cannot define a smooth function on Rn using the same formula.
 6.2. PRODUCT AND CHAIN RULES 183

that
∞
X
ku − ukLp (Ln |Ω ) ≤ kφtk ∗(ξk · u) − ξk · ukLp (Ln |Ω ) ≤
k=1
∞
X
≤ kφtk ∗(ξk · u) − ξk · ukW1,p (Ω) < ,
k=1
X∞
k∇w (u − u)kLp (Ln |Ω ) ≤ k∇w φtk ∗(ξk · u) − ξk · u kLp (Ln |Ω ) ≤


k=1
X∞
≤ kφtk ∗(ξk · u) − ξk · ukW1,p (Ω) < .
k=1

We therefore conclude that u ∈ W1,p (Ω) and u → u in W1,p (Ω) as

 → 0.

Exercise 6.25. Let Ω be an open subset of Rn , 1 < p ≤ ∞ and
f ∈ Lp (Ln |Ω ). Then the following conditions are equivalent:
i) f ∈ W1,p (Ω).
ii) There exists a constant C ≥ 0 such that, for all ϕ ∈ C∞ c (Ω) and
all 1 ≤ i ≤ n,
ˆ
∂ϕ
f dLn ≤ CkϕkLp0 (Ω) ,
Ω ∂x i

where p0 is the conjugate exponent of p.

iii) There exists a constant C ≥ 0 such that, for all relatively compact
open ω b Ω and all h ∈ Rn with khk < d(ω, Ωc ),
kτh f − f kLp (Ln |ω ) ≤ Ckhk.
Moreover, we can take C = k∇w f kLp (Ln |Ω ) in (ii) and (iii), and if Ω =
Rn , we have
kτh f − f kLp (Ln ) ≤ k∇w f kLp (Ln ) khk
for all h ∈ Rn .

6.2. Product and Chain Rules

Theorem 6.26 (Product rule). Let Ω be an open set in Rn and u, v
be real functions on Ω satisfying one of the following conditions:
1,1
i) u ∈ Wloc (Ω) and v ∈ C1 (Ω);
1,1
ii) u, v ∈ Wloc (Ω) ∩ L∞ n
loc (L |Ω ).
184 6. SOBOLEV SPACES

1,1
Then uv ∈ Wloc (Ω) and
(6.2) ∇w (uv) = u ∇w v + v ∇w u.
Proof. Note that, assuming either i) or ii), the second member in
(6.2) belongs to L1loc (Ln |Ω ). In view of the locality of the weak derivative
6.13, it therefore suffices to show that, for each open V b Ω, uv has
weak gradient on V given by (6.2).
1) Suppose that i) holds. For each 0 <  < d(V, Ωc ), we denote by
u ∈ C∞ (Ω ) the -approximation of u, cf. definition 6.17. Note that
both v and ∇v are bounded on V ; it then follows from theorem 6.20
that:
• u v ∈ L1 (Ln |V ) and u v → uv in L1 (Ln |V ) as  → 0;
• ∇(u v) = u ∇v + v∇u ∈ L1 (Ln |V , Rn ) and ∇(u v) → u∇v +
v∇u in L1 (Ln |V , Rn ).
Hence, applying proposition 6.3.i), we conclude that uv has weak
gradient on V given by (6.2), as asserted.
2) Suppose that ii) holds. For each 0 <  < d(V, Ωc ), we denote by
v ∈ C∞ (Ω ) the -approximation of v. It follows from part i) with
v in place of v that uv ∈ W1,1 (V ) and ∇w (uv ) = u∇v + v ∇w u.
Besides:
• since u ∈ L∞ (Ln |V ) and v → v in L1 (Ln |V ) (by theorem
6.20.v), it follows that uv → uv in L1 (Ln |V );
• since ∇v = (∇w v) → ∇w v in L1 (Ln |V , Rn ) (by theorem
6.20.vi and 6.20.v, respectively) and u ∈ L∞ (Ln |V ), we also
have u∇v → u ∇w v in L1 (Ln |V , Rn );
• v → v Ln -a.e. on V (by 6.20.iii), hence v ∇w u − v ∇w u → 0
Ln -a.e. on V . The latter convergence is dominated, since, fixing
0 < 0 < d(V, Ωc ), for all 0 <  < 0 :
v =φ ∗ v
kv ∇w u − v ∇w uk ≤ kv kL∞ (V ) + kvkL∞ (V ) k∇w uk ≤

≤ 2kvk
k∇w uk ∈ L1 (Ln |V ).

L∞ V +B(0,0 )

It then follows from the dominated convergence theorem 1.64

(taking  → 0 along an arbitrary sequence) that v ∇w u →
v ∇w u in L1 (Ln |V , Rn ).
Therefore, uv → uv in L1 (Ln |V ) and ∇w (uv ) → u ∇w v+v ∇w u
in L1 (Ln |V , Rn ); the thesis then follows from proposition 6.3.i).

Corollary 6.27. Let Ω be an open set in Rn , 1 ≤ p ≤ ∞ and
u, v ∈ W1,p (Ω) ∩ L∞ (Ln |Ω ). Then uv ∈ W1,p (Ω) ∩ L∞ (Ln |Ω ).
 6.2. PRODUCT AND CHAIN RULES 185

Recall that, for a map F defined on an open set of Rn , we use the

notation DF for the set of points in which F is Fréchet-differentiable.
Theorem 6.28 (Chain rule). Let Ω be an open set in Rn , f ∈
1,1
Wloc and F : R → R Lipschitz with R \ DF countable. Define
(Ω)
0 1,1
F ≡ 0 on R \ DF . Then F ◦ f ∈ Wloc (Ω) and
∇w (F ◦ f ) = (F 0 ◦ f ) · ∇w f.
Proof.
1) The thesis holds if f ∈ C1 (Ω).
Since f is locally Lipschitz, so is F ◦ f . It then follows from
corollary 6.16 that F ◦ f is differentiable in the sense of Fréchet Ln -
1,∞ 1,1
a.e. on Ω, F ◦ f ∈ Wloc (Ω) ⊂ Wloc (Ω) and ∇w (F ◦ f ) = ∇(F ◦ f )
n
L -a.e. on Ω. Hence, it suffices to show that
(6.3) ∇(F ◦ f ) = (F 0 ◦ f ) · ∇f
Ln -a.e. on Ω. The latter equality holds on f −1 (DF ) by the classical
chain rule; we must show that it holds Ln -a.e. on f −1 (R \ DF ).
Indeed, for each t ∈ Rn \ DF , the second member of (6.3) is null on
f −1 {t} (since we defined F 0 ≡ 0 on R \ DF ) and the first member
is null Ln -a.e. on f −1 {t} by corollary 5.17, thus showing that (6.3)
holds Ln -a.e. on f −1 {t}. Since Rn \ DF is countable, we conclude
that (6.3) holds Ln -a.e. on f −1 (Rn \ DF ), as asserted.
1,1
2) General case: let f ∈ Wloc (Ω). Note that
• ∀x ∈ Ω, |F ◦ f (x)| ≤ |F ◦ f (x) − F (0)| + |F (0)| ≤ (Lip F ) ·
|f (x)| + |F (0)|, hence F ◦ f ∈ L1loc (Ln |Ω);
• k(F 0 ◦ f ) · ∇w f k ≤ (Lip F ) · k∇w f k ∈ L1loc (Ln |Ω), hence (F 0 ◦
f ) · ∇w f ∈ L1loc (Ln |Ω).
In view of the locality of the weak derivative, it therefore suffices to
show that, for each open V b Ω, F ◦ f has weak derivative on V
given by (F 0 ◦ f ) · ∇w f .
Let (φt )t>0 be the standard mollifier on Rn . Let 0 < 0 <
d(V, Ωc ) and (k )k∈N a sequence in ]0, 0 [ with k ↓ 0. With the
notation from definition 6.17 in force, let ∀k ∈ N, fk := fk =
φk ∗ f ∈ C∞ (Ωk ). It follows from theorem 6.20 that fk → f in
W1,1 (V ), fk → f Ln -a.e. on V and ∇fk = (∇w f )k → ∇w f Ln -a.e.
on V .
Fix ϕ ∈ C∞ 1
c (V ). Since ∀k ∈ N, fk ∈ C (V ), it follows from part
1) of the proof that, ∀k ∈ N:
ˆ ˆ
(6.4) − F ◦ fk ∇ϕ dL = (F 0 ◦ fk ) · ∇fk ϕ dLn .
n
V V
The thesis then follows once we prove the following claims:
186 6. SOBOLEV SPACES
´ ´
• Claim 1: ´V F ◦ fk ∇ϕ dLn → V F ◦´f ∇ϕ dLn .
• Claim 2: V (F 0 ◦ fk ) · ∇fk ϕ dLn → V (F 0 ◦ f ) · ∇f ϕ dLn .

Proof of claim 1: We have:

i) ∀k ∈ N, kF ◦ fk ∇ϕk ≤ [(Lip F )|fk | + |F (0)|]k∇ϕku ;
| {z }
gk :=
ii) F ◦ fk ∇ϕ → F ◦ f ∇ϕ Ln -a.e. on V ;
iii) g´k → g := [(Lip
´ F )|f | + |F (0)|]k∇ϕku Ln -a.e. on V ;
iv) V gk dLn → V g dLn < ∞ (because fk → f in L1 (Ln |V )).
An application of the generalized dominated convergence theorem
1.66 concludes the proof. 

Proof of claim 2: We have:

i) ∀k ∈ N, k(F 0 ◦ fk ) · ∇fk ϕk ≤ (Lip F )kϕku k∇fk k;
| {z }
gk :=
0 0 n
ii) (F ◦ fk ) · ∇fk ϕ → (F ◦ f ) · ∇f ϕ L -a.e. on V ;
w n
iii) g´k → g := (Lip
´ F )kϕk u k∇ f k L -a.e. on V ;
iv) V gk dL → V g dL < ∞ (because ∇fk → ∇w f in L1 (Ln |V , Rn )).
n n

An application of the generalized dominated convergence theorem

1.66 concludes the proof. 


1,1
Corollary 6.29. Let Ω be an open set in Rn , f ∈ Wloc (Ω) and
1
F : R → R sectionally C on each compact subinterval of R, with
1,1
F 0 ∈ L∞ (R). Define F 0 ≡ 0 on R \ DF . Then F ◦ f ∈ Wloc (Ω) and

∇w (F ◦ f ) = (F 0 ◦ f ) · ∇w f.

Proof. The hypothesis on F implies F Lipschitz with R \ DF

countable. 

Corollary 6.30. Let Ω be an open set in Rn , f ∈ W1,p (Ω) and

F : R → R Lipschitz with R \ DF countable. Define F 0 ≡ 0 on R \ DF .
If Ln (Ω) = ∞, assume that F (0) = 0. Then F ◦ f ∈ W1,p (Ω) and

∇w (F ◦ f ) = (F 0 ◦ f ) · ∇w f.

Proof. Since k(F 0 ◦f )·∇w f k ≤ (Lip f )k∇w f k ∈ Lp (Ln |Ω , Rn ) and

|F ◦f | ≤ (Lip F )|f |+|F (0)| ∈ Lp (Ω), it follows that F ◦f ∈ W1,p (Ω). 
 6.3. APPROXIMATION BY SMOOTH FUNCTIONS, PART II 187

1,1
Corollary 6.31. Let Ω be an open set in Rn and f ∈ Wloc (Ω).
+ − 1,1
Then f , f , |f | ∈ Wloc (Ω) and
(
w + ∇w f Ln -a.e. on {f > 0}
∇ f =
0 Ln -a.e. on {f ≤ 0},
(
w − − ∇w f Ln -a.e. on {f < 0}
∇ f =
0 Ln -a.e. on {f ≥ 0},

w
∇ f
 Ln -a.e. on {f > 0}
∇w |f | = 0 Ln -a.e. on {f = 0}
− ∇w f

Ln -a.e. on {f < 0}.
Proof. Apply theorem 6.28 to F ◦ f , where F is given by, respec-
tively, idR ·χ[0,∞) , − idR ·χ(−∞,0] and |·|. 
1,1
Corollary 6.32. Let Ω be an open set in Rn and f ∈ Wloc (Ω).
w n
Then ∇ f = 0 L -a.e. on {f = 0}.
Proof. Apply the previous corollary and the linearity of the weak
derivative to f = f + − f − . 
6.3. Approximation by smooth functions, part II
We resume the discussion started on section 6.1 on the approxi-
mation of Sobolev functions. With regard to Meyers-Serrin’s theorem
6.24, for instance, we may obtain better approximation results if we
impose some regularity on ∂Ω. For instance, if Ω is a Lipschitz do-
main, in the sense of following definition, we will prove that Sobolev
functions on Ω may be approximated by functions in C∞ (Ω).
Notation. We will use the following notation for cylinders on
products of Euclidean spaces Rn ≡ Rk × Rn−k . Let p : Rk × Rn−k → Rk
and q : Rk × Rn−k → Rn−k be the projections on the first and second
factors, respectively. Given x ∈ Rn , 0 < r, h ≤ ∞, we define the open
and closed cylinders with center x, radius r and half-height h:
• C(x, r, h) := U(p · x, r) × U(q · x, h) ⊂ Rk × Rn−k .
• C(x, r, h) := B(p · x, r) × B(q · x, h) = C(x, r, h).
We use abbreviated notations C(x, r) := C(x, r, r) and C(x, r) :=
C(x, r, r).
Definition 6.33 (Lipschitz domains2). Let n ≥ 2, U ⊂ Rn ≡
Rn−1 × R open and Ω ⊂ U an open subset of U . We say that Ω is
2there is a weaker notion of “Lipschitz domain” which we do not consider here;
our definition corresponds to what sometimes is called a strong Lipschitz domain.
188 6. SOBOLEV SPACES

a Lipschitz domain3 in U if for all x ∈ ∂ U Ω = ∂Ω ∩ U (i.e. x in the

topological boundary of Ω in U ), there exist:
1) a rigid motion Φ ∈ SE(n) with Φ(0) = x;
2) f : Rn−1 → R Lipschitz with f (0) = 0;
3) C(0, r, h) ⊂ Rn−1 × R open cylinder
satisfying the following conditions (see figure 1):


• C := Φ C(0, r, h)
⊂ U ;
• Φ gr f ∩ C(0, r, h) =
C ∩ ∂Ω
• Φ epiS f ∩ C(0, r, h) = C ∩ Ω,
where epiS f = {(x, y) ∈ Rn−1 × R | y > f (x)} is the strict epigraph of
f.

Figure 1. Lipschitz Domain

Theorem 6.34 (Global approximation by smooth functions). Let

Ω ⊂ Rn be a Lipschitz domain. If 1 ≤ p < ∞ and f ∈ W1,p (Ω),
there exists (fk )k∈N in W1,p (Ω) ∩ C∞ (Ω) such that fk → f in W1,p (Ω).
Moreover, if f ∈ W1,p (Ω) ∩ C(Ω), the sequence (fk )k∈N may be chosen
so that it also converges to f uniformly on Ω.
We devote the remaining of this section to the proof of the theorem
above, which will be done along the following lemmas.
Lemma 6.35. Let Ω ⊂ Rn open, 1 ≤ p ≤ ∞, f ∈ W1,p (Ω) and
ξ ∈ C∞ (Rn ) with spt ξ ⊂ Ω and kξkW1,∞ (Rn ) < ∞. Then ξ · f (defined
as 0 on Rn \ Ω) belongs to W1,p (Rn ) and ∇w (ξ · f ) = (∇ξ) · f + ξ · ∇w f .
3our
Lipschitz domains need not be connected, despite the usual meaning of
the term “domain”, i.e. an open connected set.
 6.3. APPROXIMATION BY SMOOTH FUNCTIONS, PART II 189

Note that it is not required that spt ξ be compact.

Proof. It is clear that ξ · f ∈ Lp (Ln ) and g := (∇ξ) · f + ξ · ∇w f ∈
L (Ln , Rn ), thanks to the hypothesis kξkW1,∞ (Rn ) < ∞. Then it suffices
p

to show that ∇w (ξ · f ) exists and coincides with g. Indeed, it follows

1,1
from theorem 6.26 that ξ · f ∈ Wloc (Ω) and that its weak gradient on Ω
coincides with g. The same holds for the restriction of ξ · f to the open
set Rn \spt ξ (because ξ ≡ 0 on this open set). Since Ω∪Rn \spt ξ = Rn ,
the thesis follows from the locality of the weak derivative 6.13. 
Lemma 6.36. Let Γ : R
n−1 → R Lipschitz, h > 0 and α := (Lip Γ)+
2. Then d gr Γ, gr (Γ − h) ≥ αh .



Proof. Let P = x, Γ(x) ∈ gr Γ and Q = y, Γ(y) − h ∈ gr (Γ −
h). If kP − Qk = , then kx
− yk ≤ , hence |Γ(x) − Γ(y)| ≤ (Lip Γ).
Thus, putting R = y, Γ(y) , kR − P k ≤ kx − yk + |Γ(x) − Γ(y)| ≤
(1 + Lip Γ). That implies, by the triangle inequality, h = kR − Qk ≤
kR − P k + kP − Qk ≤ (2 + Lip Γ). We therefore conclude that
h
kP − Qk =  ≥ ,
α
which implies, by the
arbitrariness of P ∈ gr Γ and Q ∈ gr (Γ − h),
that d gr Γ, gr (Γ − h) ≥ αh . 
Notation. Given Γ : Rn−1 → R, Ω := epiS Γ = {(y 0 , yn ) ∈ Rn−1 ×
R | yn > Γ(y 0 )} and h > 0, we denote by Ω−h the strict epigraph of
Γ − h, i.e. Ω−h = epiS (Γ − h) = {(y 0 , yn ) ∈ Rn−1 × R | yn > Γ(y 0 ) − h}.
Notation. If A ⊂ Rn and f : A → R, we denote by spt f the
closure in A of {x ∈ A | f (x) 6= 0} (i.e the usual notation for the
support of f ) and by spt f its closure in Rn .
Lemma 6.37. With the notation above in force, let Γ : Rn−1 → R
Lipschitz. For each h > 0, there exists Ψ ∈ C∞ (Rn ) ∩ W1,∞ (Rn ) such
that 0 ≤ Ψ ≤ 1, Ψ ≡ 1 on Ω and spt Ψ ⊂ Ω−h .
h
Proof. Let (φ )>0 be the standard mollifier. Fix 0 <  < 2α and
∞ n
define Ψ := φ ∗ χΩ−h/2 . It is clear that Ψ ∈ C (R ) and 0 ≤ Ψ ≤ 1.
Moreover:
1) For all x ∈ Ω, B(x, ) ⊂ Ω−h/2 by lemma 6.36, which implies Ψ ≡ 1
on Ω;
2) spt Ψ ⊂ spt χΩ−h/2 +B(0,) = Ω−h/2 + B(0, ) ⊂ Ω−h by lemma 6.36
applied to Γ − h2 in place of Γ;
3) It follows from proposition 1.108 parts g) and j) that k∇Ψk∞ =
k(∇φ ) ∗ χΩ−h/2 k ≤ k∇φ k1 < ∞, hence Ψ ∈ W1,∞ (Rn ).
190 6. SOBOLEV SPACES


Lemma 6.38. With the notation above in force, let Γ : Rn−1 → R
Lipschitz, Ω = epiS Γ and 1 ≤ p < ∞. Let f : Ω → R and, for each
h > 0, τ−h f : Ω−h → R be given by x 7→ f (x + h). Then:
i) if f ∈ Lp (Ln |Ω), then τ−h f ∈ Lp (Ln |Ω−h ) and (τ−h f )|Ω → f in
Lp (Ln |Ω ) as h → 0.
ii) if f ∈ W1,p (Ω), then τ−h f ∈ W1,p (Ω−h ) and (τ−h f )|Ω → f in
W1,p (Ω) as h → 0.
iii) if f is uniformly continuous, so is τ−h f and (τ−h f )|Ω → f uni-
formly as h → 0.
Proof. (1) It is clear that τ−h f ∈ Lp (Ln |Ω−h ). Moreover, us-
ing a bar to denote extensions by zero, τ−h f = τ−h f converges
to f in Lp (Ln ) by lemma 1.110, which implies (τ−h f )|Ω → f
in Lp (Ln |Ω ) as h → 0.
1,1 1,1
(2) It is clear that τ−h maps Wloc (Ω) to Wloc (Ω−h ) and that τ−h
commutes with weak derivatives. Hence, by the previous item,
if f ∈ W1,p (Ω), then ∇w τ−h f = τ−h ∇w f → ∇w f in Lp (Ln |Ω , Rn )
as h → 0, from which we conclude that τ−h f ∈ W1,p (Ω−h ) and
(τ−h f )|Ω → f in W1,p (Ω).
(3) It is immediate from the definition of uniform continuity.

Lemma 6.39. With the notation above in force, let Γ : Rn−1 → R
Lipschitz and Ω = epiS Γ. Then, ∀ > 0, ∀1 ≤ p < ∞:
i) If f ∈ W1,p (Ω), there exists g ∈ C∞ n
c (R ) such that kg|Ω −f kW1,p (Ω) <
.
ii) If f ∈ W1,p (Ω) and f is bounded and uniformly continuous, there
exists g ∈ C∞ (Rn ) satisfying both kg|Ω − f kW1,p (Ω) <  and kg|Ω −
f ku < .
In both cases, if U ⊂ Rn is open and the closure in Rn of spt f is a
compact subset of U , we can take g ∈ C∞ c (U ) satisfying the conditions
stated above.
Proof. Let f ∈ W1,p (Ω). Fix  > 0. By lemma 6.38, we may take
h > 0 such that kτ−h f − f kW1,p (Ω) < /2. Besides, if f is uniformly
continuous, by the same lemma we may choose h > 0 so that kτ−h f −
f ku < /2 also holds on Ω.
Apply lemma 6.37 to obtain Ψ ∈ C∞ (Rn ) ∩ W1,∞ (Rn ) such that
0 ≤ Ψ ≤ 1, Ψ ≡ 1 on Ω and spt Ψ ⊂ Ω−h . It follows from lemma
6.35 with Ω−h in place of Ω and τ−h f in place of f that Ψ · τ−h f
belongs to W1,p (Rn ). Since Ψ ≡ 1 on Ω, Ψ · τ−h f and τ−h f have the
 6.3. APPROXIMATION BY SMOOTH FUNCTIONS, PART II 191

same restrictions to Ω; therefore, kΨ · τ−h f − f kW1,p (Ω) < /2 and, if f

uniformly continuous, we also have kΨ · τ−h f − f ku < /2 on Ω.
To prove part i), apply corollary 6.21.ii) to obtain g ∈ C∞ n
c (R ) such
that kg − Ψ · τ−h f kW1,p (Rn ) < /2, which yields kg|Ω − f kW1,p (Ω) < .
To prove part ii), let (φt )t>0 be the standard mollifier. We contend
that Ψ·τ−h f : Rn → R is bounded and uniformly continuous. Assuming
this contention, to be proved below, we may apply theorem 1.111.ii)
to obtain t > 0 sufficiently small so that kφt ∗ Ψ · τ−h f − Ψ · τ−h f ku <
/2; besides, taking a smaller t if necessary, corollary 6.21.i) yields
kφt ∗ Ψ · τ−h f − Ψ · τ−h f kW1,p (Rn ) < /2. We therefore reach the thesis
with g := φt ∗ Ψ · τ−h f ∈ C∞ (Rn ).
Proof of the contention in the previous paragraph: as f is bounded
and uniformly continuous, so is τ−h f : Ω−h → R. Then Ψ · τ−h f is
clearly bounded; it remains to show that it is uniformly continuous.
Note that Ψ is uniformly continuous, since it is smooth with bounded
derivative, hence it is Lipschitz. As Ψ · τ−h f is continuous (because so
are its restrictions to the open sets Ω and Rn \ spt Ψ) with support
contained in spt Ψ, and ∀x, y ∈ spt Ψ,
|Ψ(x) · τ−h f (x) − Ψ(y) · τ−h f (y)| ≤
≤ |Ψ(x)||τ−h f (x) − τ−h f (y)| + |τ−h f (y)||Ψ(x) − Ψ(y)| ≤
≤ kΨku |τ−h f (x) − τ−h f (y)| + kf ku |Ψ(x) − Ψ(y)|,
we conclude that Ψ · τ−h f is uniformly continuous, as asserted.
Finally, if U ⊂ Rn is open and the closure in Rn of spt f is a
compact subset of U , take ge ∈ C∞ (R) satisfying i) or ii) for a given
 > 0 and ζ ∈ C∞ n
c (R ) with 0 ≤ ζ ≤ 1, spt ζ ⊂ U and ζ ≡ 1 on
the closure in Rn of spt f . Define g := ζe g ∈ C∞ c (U ). Since ζf = f ,
we have g − f = ζ(e g − f ), hence (by theorem 6.26) ∇w (g − f ) =
∇ζ · (e g − f ) + ζ · ∇w (e
g − f ) on Ω, which implies
• kg − f kW1,p (Ω) ≤ kζkW1,∞ (Rn ) ke
g − f kW1,p (Ω) < kζkW1,∞ (Rn ) ;
• in case ii), kg|Ω − f ku ≤ kζku ke g |Ω − f ku < .
Since  > 0 was arbitrarily taken, the statements in i) and ii) are
fulfilled with g in place of ge. 
Lemma 6.40. Let 1 ≤ p ≤ ∞, U ⊂ Rn open, Φ ∈ SE(n) a rigid
motion and V = Φ(U ). The map (◦Φ) : f 7→ f ◦ Φ is:
1) a linear isometry Lp (V ) → Lp (U ) with inverse (◦Φ−1 );
2) a linear isometry W1,p (V ) → W1,p (U ) with inverse (◦Φ−1 );
3) a linear isometry Cb (V ) → Cb (U ) with inverse (◦Φ−1 ).
Proof. Part 3) is immediate and part 1) is an immediate con-
sequence of Φ# Ln = Ln (since the Lebesgue measure is invariant by
192 6. SOBOLEV SPACES

translations and rotations). Part 2) follows from part 1) and from the
fact that weak derivatives commute with (◦Φ), as it can be directly
checked.

Lemma 6.41. Let Ω ⊂ Rn open, f ∈ C(Ω) and ξ ∈ Cc (Rn ). Then
ξ · f : Ω → R is uniformly continuous.
Proof. We may extend f to a continuous function Ω → R, which
on its turn may be extended, in view of Tietze’s extension theorem, to
a continuous function fe : Rn → R. Then ξ · fe ∈ Cc (Rn ) ⊂ C0 (Rn ) is
uniformly continuous by lemma 1.109, and so is its restriction (ξ· fe)|Ω =
ξ · f : Ω → R. 
Proof of theorem 6.34. Let (Ui )i≥0 be a countable open cover
of Rn , where U0 = Rn \ Ω, U1 = Ω and, for each i ≥ 2, Ui is obtained
by rigid motion of a cylinder centered at 0 ∈ Rn as in definition 6.33,
i.e. there exists a rigid motion Φi ∈ SE(n) with Φi (0) = xi ∈ ∂Ω and
there exists ri , hi > 0 and Γi
: Rn−1 → R Lipschitz
with Γi (0) = 0
such that Ui = Φi C(0,
ri , hi ) , Φi gr Γi ∩ C(0, ri , hi ) = Ui ∩ ∂Ω and
Φi epiS Γi ∩ C(0, ri , hi ) = Ui ∩ Ω.
Let (Vk )k∈N be a locally finite refinement of (Ui )i∈N formed by rel-
atively compact open subsets of Rn , and (ξk )k∈N a smooth partition of
unity of Rn such that for each k ∈ N, spt ξk b Vk , given by theorem
6.8. Note that, for each k ∈ N, the fact that ξk ∈ C∞ n
c (R ) and the
product rule 6.26 imply that:
• ξk · f ∈ W1,p (Ω) and spt ξk · f ⊂ spt ξk b Vk ;
• if f belongs to C(Ω), it follows from lemma 6.41 that ξk · f :
Ω → R is uniformly continuous.
Fix  > 0.
1) Claim: for each k ∈ N, there exists gk ∈ C∞ c (Vk ) such that kgk |Ω −ξk ·
f kW1,p (Ω) < 2−k . Moreover, if f ∈ C(Ω), we may take gk ∈ C∞ c (Vk )
−k −k
so that kgk |Ω − ξk · f kW1,p (Ω) < 2  and kgk |Ω − f ku < 2 .
Indeed, there exists ik ∈ N such that Vk ⊂ Uik .
• If ik = 0, ξk · f is the null function on Ω and we may take
gk ≡ 0.
• If ik = 1, spt ξk b Vk ⊂ U1 = Ω, hence ξk · f ∈ W1,p (Rn )
by lemma 6.35 and spt ξk · f b Vk . Besides, if f ∈ C(Ω),
ξk · f ∈ Cc (Rn ). Then, if (φt )t>0 is the standard mollifier,
by corollary 6.21 there exists t > 0 sufficiently small so that
φt ∗(ξk · f ) ∈ C∞ −k
c (Vk ) and kφt ∗(ξk · f ) − ξk · f kW1,p (Rn ) < 2 .
If f ∈ C(Ω), the fact that ξk · f ∈ Cc (Rn ) and theorem 1.111.ii)
 6.3. APPROXIMATION BY SMOOTH FUNCTIONS, PART II 193

ensure the existence of a smaller t > 0 such that we also have

kφt ∗(ξk · f ) − ξk · f kL∞ (Rn ) < 2−k . Put gk := φt ∗(ξk · f ).

• If ik ≥ 2, (ξk · f ) ◦ Φik ∈ W1,p C(0, rik , hik ) ∩ epiS Γik (by
lemma 6.40) and spt (ξk · f ) ◦ Φik ⊂ Φ−1 −1
ik (spt ξk ) b Φik (Vk ) ⊂
C(0, rik , hik ). Hence (ξk ·f )◦Φik ∈ W1,p (epiS Γik ). Furthermore,
if f ∈ C(Ω), as we saw above ξk · f is uniformly continuous on
Ω, hence its restriction to Uik ∩ Ω is bounded (because it may
be continuously extended to the compact set Uik ∩ Ω, on which
it is therefore bounded) and uniformly continuous, and so is
(ξk · f ) ◦ Φik : C(0, rik , hik ) ∩ epiS Γik → R.
We contend that (ξk · f ) ◦ Φik : epiS Γik → R is bounded and
uniformly continuous. Indeed, it is clearly bounded, since it
is null on the complement of the cylinder C(0, rik , hik ) and its
restriction to C(0, rik , hik ) ∩ epiS Γik is bounded. Moreover,
since spt (ξk · f ) ◦ Φik b C(0, rik , hik ) and the restriction (ξk ·
f ) ◦ Φik : C(0, rik , hik ) ∩ epiS Γik → R is uniformly continuous,
given  > 0, we may take 0 < δ < δ0 := d(spt (ξk · f ) ◦ Φik , Rn \
C(0, rik , hik )) such that, putting F := (ξk · f ) ◦ Φik , ∀x, y ∈
epiS Γik ∩ C(0, rik , hik ) with kx − yk < δ, we have |F (x) −
F (y)| < . The same holds for all x, y ∈ epiS Γik with kx−yk <
δ because, if x ∈ Rn \ C(0, rik , hik ) and kx − yk < δ, then
y ∈ Rn \ spt (ξk · f ) ◦ Φik , hence F (x) = F (y) = 0. Thus, the
contention is proved.
Applying lemma 6.39 with (ξk · f ) ◦ Φik in place
of f and
Φ−1
ik (Vk ) in place of U , we obtain g ek ∈ C∞ −1
c Φik (Vk ) such that
kegk |epiS Γik − (ξk · f ) ◦ Φik kW1,p (epiS Γik ) < 2−k  and, if f ∈ C(Ω),
kegk |epiS Γik − (ξk · f ) ◦ Φik ku < 2−k . Thus, in view of lemma
6.40, P gk := gek ◦ Φi−1
k
proves the claim.
2) Let g := g
k≥0 k . Since ∀k ∈ N, gk ∈ C∞ c (Vk ) and (Vk )k∈N is a
n
locally finite open cover of R , the sum which Pdefines g is locally
finite and g ∈ C∞ (Rn ). Besides, since f = k≥0 ξk · f , we have
P
kg|Ω − f kW1,p (Ω) P≤ k≥0 kgk |Ω − ξk · f kW (Ω) < 2 and, if f ∈ C(Ω),
1,p

kg|Ω − f ku ≤ k≥0 kgk |Ω − ξk · f ku < 2.



Remark 6.42. With the same hypothesis and notation from theo-
rem 6.34, we have actually proved that there exists (fk )k∈N in W1,p (Ω)∩
C∞ (Rn ) such that fk → f in W1,p (Ω), which also converges to f uni-
formly on Ω if f ∈ W1,p (Ω) ∩ C(Ω).
194 6. SOBOLEV SPACES

Corollary 6.43. Let Ω ⊂ Rn be a Lipschitz domain. If 1 ≤ p < ∞

and f ∈ W1,p (Ω), there exists (fk )k∈N in C∞ n
c (R ) such that fk |Ω → f in
W1,p (Ω). Moreover, if f ∈ W1,p (Ω) ∩ C(Ω), the sequence (fk )k∈N may
be chosen so that it also converges to f uniformly on compact subsets
of Ω.
Proof. As it was noted in remark 6.42, there exists a sequence
(gk )k∈N in C∞ (Rn ) such that (∀k)gk |Ω ∈ W1,p (Ω) and gk |Ω → f in
W1,p (Ω), and such that (gk |Ω ) also converges uniformly to f if f ∈
W1,p (Ω) ∩ C(Ω).
We now adapt the argument from part ii) of corollary 6.21. Choose
ζ ∈ C∞ n
c (R , [0, 1]) such that ζ ≡
1 on B(0, 1) and spt ζ ⊂ U(0, 2).
∞
Define, ∀k ∈ N, ζk ∈ Cc U(0, 2k) by ζk (x) := ζ(x/k), and fk := ζk ·gk .
Then ∀k ∈ N, fk ∈ C∞ n 1,p
c (R ). We will prove that fk |Ω → f in W (Ω).
We omit restrictions for simplicity of notation; the p-norms are taken
with respect to Ln |Ω .
1) For all u ∈ Lp (Ln |Ω ), ζk · u → u in Lp (Ln |Ω ). Indeed, |u − ζk · u|p → 0
pointwise and |u − ζk · u|p ≤ 2p |u|p ∈ L1 (Ln |Ω ), hence the dominated
convergence theorem 1.64 yields the assertion.
2) Since f − fk = (f − ζk · f ) + ζk · (f − gk ), we have kf − fk kp ≤
kf − ζk · f kp + kζk ku kf − gk kp → 0, since |ζk | ≤ 1, kf − ζk · f kp → 0
by the previous item and kf − gk kp ≤ kf − gk kW1,p (Ω) → 0.
3) ∀x ∈ Rn ,
∇fk (x) = ∇ζk (x) · gk (x) + ζk (x) · ∇gk (x) =
1
= · ∇ζ(x/k) · gk (x) + ζk (x) · ∇gk (x).
k
Hence, ∀x ∈ Ω,
∇w f (x) − ∇fk (x) = ∇w f (x) − ζk (x) ∇w f (x)+
1
+ ζk (x)[∇w f (x) − ∇gk (x)] − · ∇ζ(x/k) · gk (x),
k
so that
1
k∇w f −∇fk kp ≤ k∇w f −ζk ∇w f kp +kζk ku k∇w f −∇gk kp + k∇ζku kgk kp .
k
Since k∇w f − ζk ∇w f kp → 0 by item 1), k∇w f − ∇gk kp ≤ kf −
gk kW1,p (Ω) → 0 and, as (gk )k∈N is bounded in Lp (Ln |Ω ) (because it is
convergent in W1,p (Ω), hence in in Lp (Ln |Ω )), k1 k∇ζk ku kgk kp → 0, it
follows that k∇w f − ∇fk kp → 0. We have thus proved that fk → f
in Lp (Ln |Ω ) (by the previous item) and ∇fk → ∇w f in Lp (Ln |Ω , Rn );
that is, fk → f in W1,p (Ω), as asserted.
 6.5. TRACES AND EXTENSIONS 195

Finally, if f ∈ W1,p (Ω) ∩ C(Ω), (fk |Ω )k∈N converges to f uniformly

on compact subsets of Ω, bacause so does (gk )k∈N , and for each k ∈ N,
fk ≡ gk on B(0, k). 

6.4. Lipschitz functions and W1,∞

Theorem 6.44. Let Ω ⊂ Rn be open and f : Ω → R. Then
1,∞
f ∈ Wloc (Ω) if, and only if, f coincides Ln -a.e. on Ω with a locally
Lipschitz function.
Proof. We have already proved in corollary 6.16 that, if f is locally
1,∞
Lipschitz, than f ∈ Wloc (Ω) (moreover, it is Fréchet-differentiable Ln -
a.e. and its Fréchet derivative coincides Ln -a.e. with its weak gradient).
1,∞
It remains to prove the converse. Suppose that f ∈ Wloc (Ω). It
suffices to show that, for each relatively compact convex open subset
V b Ω, f |V coincides Ln -a.e. with a Lipschitz function, since Ω may
be covered by countably many such convex open sets. Let W b Ω
open such that V b W . Let 0 := d(V , W c ) and (φt )t>0 the standard
mollifier. Then, for every 0 <  < 0 , f = φ ∗ f ∈ C∞ (V ) and for each
x∈V,
ˆ
6.20.vi)
k∇f (x)k = k ∇w f (y)φ (x − y) dLn (y)k ≤
B(x,)
w
≤ k∇ f kL∞ (Ln |W ) .
Thus, putting C := k∇w f kL∞ (Ln |W ) , we have sup{k∇f kL∞ (Ln |V ) | 0 <
 < 0 } ≤ C < ∞. Therefore, for all x, y ∈ V and 0 <  < 0 :
ˆ 1
∇f x + t(y − x) · (y − x) dLn (t)| ≤


|f (y) − f (x)| = |
0
≤ Cky − xk.
Hence, denoting by Lf the set of Lebesgue points of f , it follows from
theorem 6.20.iii) that, taking  ↓ 0, for all x, y ∈ V ∩ Lf ,
|f (y) − f (x)| ≤ Cky − xk,
i.e. f |Lf ∩V is Lipschitz. We may therefore extend this restriction to
a Lipschitz function on V (even on Rn ). Since Ln (V \ Lf ) = 0, we
have proved that f coincides Ln -a.e. on V with a Lipschitz function,
as asserted. 

6.5. Traces and Extensions

We prove below a version of the Gauss-Green theorem for epigraphs
of Lipschitz functions which will be needed to prove theorems on traces
and extensions of Sobolev functions. This theorem will be generalized
196 6. SOBOLEV SPACES

in chapter 7, theorem 7.18; it is essentially a consequence of the area

formula.

Theorem 6.45 (Gauss-Green theorem for Lipschitz epigraphs). Let

n ≥ 2, f : Rn−1 → R Lipschitz and Ω := epiS f (hence ∂Ω = gr f ).
Then
x
i) Hn−1 ∂Ω is a Radon measure on Rn ;
ii) there exists a Borel measurable unit vector field ν : ∂Ω → Rn ,
x
unique up to Hn−1 ∂Ω-null sets, such that, for all ϕ ∈ C1c (Rn ),
ˆ ˆ
n
(6.5) ∇ϕ dL = ϕν dHn−1 ,
Ω ∂Ω

or, equivalently, such that, for all ϕ ∈ C1c (Rn , Rn ),

ˆ ˆ
n
(6.6) div ϕ dL = ϕ · ν dHn−1 .
Ω ∂Ω


Proof. Let Γ : Rn−1 → Rn ≡ Rn−1 ×R be given by x 7→ x, f (x) .
Then Γ is Lipschitz 1-1 and Im Γ = gr f = ∂Ω. Thus, from corol-
x
lary 5.39.i), it follows that Hn−1 ∂Ω = Γ# (Ln−1 JΓ), whence x
x
Hn−1 ∂Ω is a Radon measure on Rn (because it is Borel regular and
finite on compact sets, as one can see directly from the above formula).
It remains to prove the existence and uniqueness up to Hn−1 ∂Ω- x
null sets of ν : ∂Ω → Rn Borel measurable with kνk ≡ 1 such that
(6.6) or, equivalently, (6.5) holds. Indeed, for each x = x0 , f (x0 ) ∈

Γ(Df ) ⊂ ∂Ω, where Df ⊂ Rn−1 is the differentiability set of f , let

∇f (x0 ), −1


(6.7) ν(x) = p ,
1 + k∇f (x0 )k2

and let ν be any constant unit vector field on ∂Ω \ Γ(Df ). Since Df ∈

B Rn−1 , ∇f is Borelian on Df and Ln (Rn−1 \ Df ) = 0 (by exercise 5.13
and by Rademacher’s theorem), it follows that ν is Borelian, kνk ≡ 1
x x
and ∂Ω\Γ(Df ) is Hn−1 ∂Ω-null, so that ν is given Hn−1 ∂Ω-a.e. by
(6.7). We will prove that (6.5) holds with such ν. If ν 0 is another such
x
Borel unit vector field, then (ν, Hn−1 ∂Ω) and (ν 0 , Hn−1 ∂Ω) are x
polar decompositions of the same Rn -valued Radon measure, so that
x
ν = ν 0 Hn−1 ∂Ω-a.e. by the uniqueness of the polar decomposition.
Given δ > 0, let Fδ be the open strip of amplitude 2δ along the
graph of f (see figure 2), i.e. Fδ := {x = (x0 , xn ) ∈ Rn | |xn − f (x0 )| <
δ}. We approximate the characteristic function χΩ of Ω = epiS f by a
Lipschitz function fδ defined as 1 on epi (f + δ), 0 on hyp (fδ ), and by
 6.5. TRACES AND EXTENSIONS 197

linear interpolation on Fδ , i.e. for all x = (x0 , xn ) ∈ Rn−1 × R,


1
 xn ≥ f (x0 ) + δ,
fδ (x) := 0 xn ≤ f (x0 ) − δ,
(2δ)−1 x − f (x0 ) − δ


x ∈ Fδ .
n

It is then clear that 0 ≤ fδ ≤ 1, Lip fδ = (2δ)−1 (1+Lip f ) and fδ → χΩ

pointwise on Rn−1 \ gr f as δ ↓ 0; in particular, fδ → χΩ Ln -a.e. on
Rn , since Ln (gr f ) = 0 (for instance, as an immediate consequence of
Fubini’s theorem). Furthermore, by a direct computation, the classical
gradient of fδ exists on Ln -a.e. x = (x0 , xn ) ∈ Rn and is given by:

(
0 if xn > f (x0 ) + δ or xn < f (x0 ) − δ
∇fδ (x) =
(2δ)−1 −∇f (x0 ), 1 if x ∈ Fδ and x0 ∈ Df .

Figure 2. Gauss-Green theorem for epigraphs

198 6. SOBOLEV SPACES

We now compute, given ϕ ∈ C1c (Rn ) and taking a sequence (δk )k∈N
in (0, ∞) with δk ↓ 0:
ˆ ˆ
5.5,fδk Lipschitz
n DCT 1.64
∇ϕ dL = lim fδk ∇ϕ dLn =
k→∞
Ω
ˆ
∇w fδk =∇fδk by 5.12
= − lim ϕ ∇w fδk dLn =
k→∞
ˆ
1 Fubini 1.84
−∇f (x0 ), 1 dLn (x)


= − lim ϕ(x) =
k→∞ F
δk
2δ k
ˆ ˆ f (x0 )+δk
0

1
= lim ∇f (x ), −1 ϕ(x0 , t) dLn (t) dLn−1 (x0 ) =
k→∞ Rn−1 2δk f (x0 )−δk
ˆ ˆ f (x0 )+δk
bounded
}|
0
z
{ 1 DCT 1.64
= lim ∇f (x ), −1 ϕ(x0 , t) dLn (t) dLn−1 (x0 ) =
k→∞ pr n−1 spt ϕ 2δk f (x0 )−δk
| R {z } | {z }
compact ≤kϕku
ˆ √
0

0 0

n−1 0 JΓ(x0 )= 1+k∇f (x0 )k2
= ∇f (x ), −1 ϕ x , f (x ) dL (x ) =
ˆR
n−1

area formula 5.39.ii)

ϕ Γ(y) ν Γ(y) JΓ(y) dLn−1 (y)



= =
ˆR
n−1

= ϕν dHn−1 ,
∂Ω
thus proving (6.5). 
Definition 6.46. With the notation from the previous theorem, ν
is called outer unit normal to ∂Ω.
Remark 6.47. With the notation from the previous theorem, we
x
have actually
proved that, up to Hn ∂Ω-null sets, on each point point
x = x0 , f (x0 ) in ∂Ω = gr f whose abscissa x0 is a differentiability point
of f ,
∇f (x0 ), −1


(6.8) ν(x) = p .
1 + k∇f (x0 )k2
In particular, if f is C1 , ν coincides with the usual outer unit normal
from Differential Geometry.
Theorem 6.48 (Trace theorem for Sobolev functions on Lipschitz
epigraphs). Let n ≥ 2, Γ : Rn−1 → R Lipschitz, Ω := epiS Γ and
1 ≤ p < ∞. Then:
i) There exists a unique bounded linear operator T : W1,p (Ω) →
Lp (Hn−1 |∂Ω ) such that, for all f ∈ C1c (Rn ), T · (f |Ω ) = f |∂Ω .
 6.5. TRACES AND EXTENSIONS 199

ii) The Gauss-Green formula holds for all f ∈ W1,1 (Ω), i.e. denoting
by ν the unit outer normal to ∂Ω,
ˆ ˆ
w n
(6.9) ∇ f dL = T · f ν dHn−1 ,
Ω ∂Ω

with a similar equality in divergence form. Furthermore, for all

f ∈ W1,p (Ω) and ϕ ∈ C1c (Rn , Rn ),
ˆ ˆ ˆ
n w n
(6.10) f div ϕ dL = − h∇ f, ϕi dL + T · f hϕ, νi dHn−1 .
Ω Ω ∂Ω

Proof. 1) Let (e1 , . . . , en ) be the standard basis of Rn ≡ Rn−1 ×R.

Since, for all x ∈ DΓ , k∇Γ(x)k ≤ Lip Γ, it follows from (6.8) that,
for all x ∈ DΓ ,
1
−en · ν ≥ p .
1 + (Lip Γ)2
p
In particular, putting C := 1 + (Lip Γ)2 , we conclude that
(6.11) 1 ≤ C (−en · ν)
x
Hn−1 ∂Ω-a.e. on ∂Ω.
2) Given  > 0, let β : R → R be given by β (t) := (t2 + 2 )1/2 − .
Note that β ∈ C∞ (R), β ≥ 0, β (t) increases to |t| as  ↓ 0 and
|β0 | ≤ 1.
3) Fix f ∈ C1c (Rn ). Then β ◦ f ∈ C1c (Rn ) (since β (0) = 0, hence
spt β ◦ f ⊂ spt f ). We compute:
ˆ (6.11)
ˆ
n−1 (6.6)
β ◦ f dH ≤ −C h β ◦ f en , νi dHn−1 =
∂Ω ∂Ω
| {z }
∈C1c (Rn ,Rn )
ˆ
∂
= −C [ β ◦ f ] dLn ≤
Ω ∂xn | {z }
∂f
=β0 ◦f · ∂x
ˆ n

|β0 f (x) |k∇f (x)k dLn (x)

≤C
ˆΩ
≤C k∇f k dLn .
Ω

Since β ◦ f increases pointwise to |f | as  ↓ 0, we may therefore

apply the monotone convergence theorem 1.62 to conclude that
ˆ ˆ
n−1
(6.12) |f | dH ≤ C k∇f k dLn .
∂Ω Ω
200 6. SOBOLEV SPACES

In particular, since C1c (Rn ) is dense in W1,1 (Ω) by lemma 6.39.i),

f ∈ C1c (Rn ) 7→ f |∂Ω may be uniquely extended to a bounded linear
function W1,1 (Ω) → L1 (Hn−1 |∂Ω ), thus proving part i) for p = 1.
4) For 1 < p < ∞, given f ∈ C1c (Rn ), note that |f |p ∈ C1c (Rn ) (since
|·|p ∈ C1 (R) for p > 1 and it is null on 0) and ∇(|f |p ) = p|f |p−1 ·
sgn f · ∇f by the chain rule. We may therefore apply (6.12) with
|f |p in place of f ; denoting by q ∈ (1, ∞) the conjugate exponent to
p, we compute:
ˆ ˆ
p n−1
|f | dH ≤ C k∇(|f |p )k dLn =
∂Ω
ˆΩ
= pC |f |p−1 k∇f k dLn ≤
Ω | {z }
 1/p  1/q k∇f kp |f |p
= k∇f kp |f |q(p−1) ≤ +
ˆ ˆ
p q

pC
≤C k∇f kp dLn + |f |p dLn .
Ω q Ω
We therefore conclude that kf kLp (Hn−1 |∂Ω ) ≤ C(n, p, Lip Γ)kf kW1,p (Ω) ;
since C1c (Rn ) is dense in W1,p (Ω), by lemma 6.39.i), the linear map
f ∈ C1c (Rn ) 7→ f |∂Ω may be uniquely extended to a bounded linear
function W1,p (Ω) → Lp (Hn−1 |∂Ω ), thus proving part i) for 1 < p <
∞.
5) It remains to prove part ii). Let f ∈ W1,1 (Ω). By lemma 6.39.i),
we may take a sequence (fk )k∈N in C∞ n
c (R ) such that fk → f in
1,1
W (Ω). It then follows, by the continuity of the trace operator,
that fk |∂Ω = T · fk → T · f in L1 (Hn−1 |∂Ω ). On the other hand, for
each k ∈ N, it follows from the Gauss-Green theorem (6.5) that
ˆ ˆ
n
∇fk dL = fk ν dHn−1 .
Ω ∂Ω
Hence, taking k → ∞, we obtain (6.9).
Similarly, if f ∈ W1,p (Ω) and ϕ ∈ C1c (Rn , Rn ), we may apply once
more lemma 6.39.i) to obtain a sequence (fk )k∈N in C∞ n
c (R ) such that
fk → f in W1,p (Ω). It then follows, by the continuity of the trace
operator, that fk |∂Ω = T · fk → T · f in Lp (Hn−1 |∂Ω ). On the other
hand, for each k ∈ N an application of (6.6) to fk · ϕ ∈ C1c (Rn , Rn )
yields
ˆ ˆ ˆ
n n
fk div ϕ dL = − h∇fk , ϕi dL + fk hϕ, νi dHn−1 .
Ω Ω ∂Ω
Since spt ϕ is compact, we have fk div ϕ → f div ϕ, h∇fk , ϕi →
h∇w f, ϕi and fk hϕ, νi → T · f hϕ, νi in L1 ; therefore, taking k → ∞
in the last equality yields (6.10).
 6.5. TRACES AND EXTENSIONS 201


Remark 6.49. With the notation from the previous theorem, for
p = 1 it follows from 6.12 and from the density of C1c (Rn ) in W1,1 (Ω)
that the inequality
ˆ ˆ
n−1
|T · f | dH ≤ C k∇w f k dLn .
∂Ω Ω
1,1
p
holds for all f ∈ W (Ω), where C := 1 + (Lip Γ)2 .
Corollary 6.50 (Trace theorem for Sobolev functions on Lip-
schitz epigraphs). With the hypothesis from the previous theorem, if
f ∈ W1,p (Ω) ∩ C(Ω), then T · f = f |∂Ω .
Proof. It follows from corollary 6.43 that there exists (fk )k∈N in
C∞
c (Rn
) such that fk |Ω → f in W1,p (Ω) and (fk )k∈N converges to f
uniformly on compact subsets of Ω. Then fk |∂Ω = T · fk → T · f in
Lp (Hn−1 |∂Ω ) and fk |∂Ω → f |∂Ω uniformly on compact subsets of ∂Ω,
which implies T · f = f |∂Ω . 
The theorem below, which generalizes theorem 6.48 for Lipschitz
domains on Rn with bounded frontier, may be skipped on a first read-
ing. We shall need theorem 7.18, which ensures that every Lipschitz
domain Ω ⊂ Rn is a set of locally finite perimeter. The material covered
in chapter 7 up to its first section 7.1 is independent of the remaining
parts of this chapter, so that the reader may study it now if he wishes
to better understand the following theorem.
Theorem 6.51 (Trace theorem for Sobolev functions on Lipschitz
domains). Let n ≥ 2, Ω ⊂ Rn a Lipschitz domain with ∂Ω bounded,
and 1 ≤ p < ∞. Then:
i) There exists a unique bounded linear operator T : W1,p (Ω) →
Lp (Hn−1 |∂Ω ) such that, for all f ∈ C1c (Rn ), T · (f |Ω ) = f |∂Ω .
ii) The Gauss-Green formula holds for all f ∈ W1,1 (Ω), i.e. denoting
by ν the unit outer normal to ∂Ω, cf. definition 7.12,
ˆ ˆ
w n
(6.13) ∇ f dL = T · f ν dHn−1 ,
Ω ∂Ω

with a similar equality in divergence form. Furthermore, for all

f ∈ W1,p (Ω) and ϕ ∈ C1c (Rn , Rn ),
ˆ ˆ ˆ
n w n
(6.14) f div ϕ dL = − h∇ f, ϕi dL + T · f hϕ, νi dHn−1 .
Ω Ω ∂Ω

Proof.
202 6. SOBOLEV SPACES

1) For each x ∈ ∂Ω, there exists an open set Ux ⊂ Rn such that x ∈ Ux

and Ux is obtained by rigid motion of a cylinder centered at 0 ∈ Rn
as in definition 6.33, i.e. there exists a rigid motion Φ ∈ SE(n) with
n−1
Φ(0) = x and there exists r, h > 0 and
Γ:R → R Lipschitz

with
Γ(0) = 0 such that Ux = Φ
C(0, r, h) , Φ gr Γ∩C(0, r, h) = Ux ∩∂Ω
and Φ epiS Γ ∩ C(0, r, h) = Ux ∩ Ω.
2) From the open cover (Ux )x∈∂Ω of the compact set ∂Ω ⊂ Rn , we may
extract a finite subcover (Ui )1≤i≤N . For each 1 ≤ i ≤ N , let the
corresponding objects defined in the
previous item be denoted with
a subscript i, so that Φi C(0, ri , hi ) = Ui .
c
Let U0 := Ω and U−1 := Ω , so that (Ui )−1≤i≤N is a finite open
cover of Rn . We may apply corollary 6.11 to obtain a smooth par-
tition of unity (ξi )−1≤i≤N of Rn with spt ξi ⊂ Ui for −1 ≤ i ≤ N .
Besides, for i ≥ 1, as spt ξi ⊂ Ui b Rn , it follows that spt ξi is a
compact subset of Ui .
3) Fix f ∈ C1c (Rn ) and 1 ≤ p < ∞. We contend that, for 1 ≤ i ≤ N ,
there exists Ci = Ci (n, p, Lip Γi ) such that

k(ξi f )|∂Ω kLp (Hn−1 |∂Ω ) ≤ Ci kξi kW1,∞ (Rn ) kf kW1,p (Ω) .

Indeed, we have:
ˆ ˆ
p n−1
|ξi f | dH = |ξi f |p dHn−1 =
∂Ω ∂Ω∩Ui
ˆ
= x∂Φ (epi Γ )
=
|ξi f |p d Hn−1 i S i
lemma 7.17

ˆ
= |ξ f | d Φ H
i
p
x∂ epi Γ

=
i#
n−1
S i
ex. 1.70

ˆ
= |(ξ f ) ◦ Φ | d H
i x∂ epi Γ
≤
i
p n−1
S i
thm. 6.48

lemma 6.40
≤ Cip k(ξi f ) ◦ Φi kp
=
W1,p epiS Γi ∩C(0,ri ,hi )

product rule 6.26

= Cip kξi f kp
= Cip kξi f kpW1,p (Ω) ≤
W1,p Ui ∩Ω

≤ Cip kξi kpW1,∞ (Rn ) kf kpW1,p (Ω) .

That Ci depends only on n, p and Lip Γi follows from part 4) of the

proof of theorem 6.48. Our contention is then proved.
 6.5. TRACES AND EXTENSIONS 203
PN
4) Since f |∂Ω = i=1 (ξi f )|∂Ω , we have
N
X 3)
kf |∂Ω kLp (Hn−1 |∂Ω ) ≤ k(ξi f )|∂Ω kLp (Hn−1 |∂Ω ) ≤
i=1
N
X

≤ Ci kξi kW1,∞ (Rn ) kf kW1,p (Ω) .
i=1

The above inequality shows that the linear map f ∈ {f |Ω | f ∈

C1c (Rn )} 7→ f |∂Ω ∈ Lp (Hn−1 |∂Ω ) is continuous with respect to the
W1,p (Ω) relative topology on {f |Ω | f ∈ C1c (Rn )}. Since the latter
subspace is dense in W1,p (Ω), by corollary 6.43, we conclude that
there exists a unique continuous linear map W1,p (Ω) → Lp (Hn−1 |∂Ω )
which extends the map f 7→ f |∂Ω on {f |Ω | f ∈ C1c (Rn )}. Assertion
i) is therefore proved.
5) In view of theorem 7.18, (6.13) holds for f ∈ C∞ n
c (R ). Given f ∈
W1,1 (Ω), by corollary 6.43 there exists a sequence (fi )i∈N in C∞ n
c (R )
1,1
such that fi |Ω → f in W (Ω); hence, by continuity of the trace
operator, fi |∂Ω = T · fi → T · f in L1 (Hn−1 |∂Ω ). Therefore,
ˆ ˆ
w n
∇ f dL = lim ∇fi dLn =
i→∞
Ω
ˆΩ
= lim fi ν dHn−1 =
i→∞ ∂Ω
ˆ
= T · f ν dHn−1 ,
∂Ω

thus proving assertion ii).

6) Equality (6.13) in divergence form reads
ˆ ˆ
n
div f dL = hT · f, νi dHn−1 ,
Ω ∂Ω

for all f ∈ W (Ω, R ). Therefore, given f ∈ W1,p (Ω) and ϕ ∈

1,1 n

C1c (Rn , Rn ), (6.14) follows from the previous equality and from the
product rule 6.26 applied to f ϕ componentwise, which yields div (f ϕ) =
f div ϕ + h∇w f, ϕi.

Corollary 6.52 (Trace theorem for Sobolev functions on Lip-
schitz domains). With the hypothesis from the previous theorem, if
f ∈ W1,p (Ω) ∩ C(Ω), then T · f = f |∂Ω .
Proof. It is identical to the proof of corollary 6.50. 
204 6. SOBOLEV SPACES

Definition 6.53 (Extension by reflection with respect to Lipschitz

graphs). Let n ≥ 2, Γ : Rn−1 → R Lipschitz and Ω := epiS Γ. We
identify Rn ≡ Rn−1 × R.
n n 0 0 0


ΦΓ0 : R 0 → R
given by (x , xn ) 7→ x , Γ(x ) − xn −
1) The map
0
Γ(x ) = x , 2Γ(x ) − xn is called reflection with respect to Γ.
2) Given f : Rn → R, the function fΓ := f ◦ ΦΓ is said to be obtained
by f by reflection with respect to Γ.
3) Given f : Ω → R, the extension of f by reflection with respect to Γ
is the function EΓ f : Rn → R given by
(
f (x) if x ∈ Ω
EΓ f (x) :=
f ◦ ΦΓ (x) if x ∈ Ωc .

Theorem 6.54 (Extension by reflection for Sobolev functions on

Lipschitz epigraphs). Let n ≥ 2, Γ : Rn−1 → R Lipschitz, Ω := epiS Γ
and 1 ≤ p < ∞. Then there exists a unique extension operator E :
W1,p (Ω) → W1,p (Rn ), i.e. a bounded linear operator with (E f )|Ω = f
for all f ∈ W1,p (Ω), such that, for all f ∈ C1c (Rn ), E(f |Ω ) = EΓ (f |Ω )
(i.e. the extension of f |Ω by reflection with respect to Γ).

Proof.
1) It suffices to prove that the extension by reflection EΓ is a bounded
linear operator defined on the subspace C1c (Rn )|Ω := {f |Ω | f ∈
C1c (Rn )} of W1,p (Ω), taking values in W1,p (Rn ). Indeed, if that is the
case, since C1c (Rn )|Ω is dense in W1,p (Ω) by lemma 6.39.i), EΓ may
be uniquely extended to a bounded linear operator E : W1,p (Ω) →
W1,p (Rn ). Then E satisfies (E f )|Ω = f for all f ∈ W1,p (Ω) because,
as the restriction R : W1,p (Rn ) → W1,p (Ω) (i.e. given by f 7→ f |Ω ) is
linear continuous, the composite R ◦ E : W1,p (Ω) → W1,p (Ω) is linear
continuous and coincides with the identity on the dense subspace
C1c (Rn )|Ω , hence R ◦ E is the identity. We then reach the thesis, i.e.
E is an extension operator which uniquely extends EΓ .
We must therefore prove that there exists C > 0 such that,
for each f ∈ C1c (Rn ), EΓ (f |Ω ) ∈ W1,p (Rn ) and kEΓ (f |Ω )kW1,p (Rn ) ≤
Ckf kW1,p (Ω) ; then EΓ : C1c (Rn )|Ω → W1,p (Rn ) is a well defined map
and its linearity is clear, hence it is a bounded linear operator.
2) For all x = (x0 , xn ), y = (y 0 , yn ) ∈ Rn , we have

kΦΓ (x) − ΦΓ (y)k = x0 − y 0 , 2[Γ(x0 ) − Γ(y 0 )] − (xn − yn ) ≤

≤ kx0 − y 0 k + 2 Lip Γkx0 − y 0 k + |xn − yn | ≤

≤ 2(Lip Γ + 1)kx − yk,
 6.5. TRACES AND EXTENSIONS 205

hence ΦΓ is Lipschitz with Lip Φγ ≤ 2(Lip Γ + 1). It then fol-

lows from Rademacher’s theorem 5.12 that ΦΓ is Ln -a.e. Fréchet-
differentiable. Moreover, for all x ∈ DΦΓ (thus for Ln -a.e. x in
Rn ),

kDΦΓ (x)k ≤ Lip ΦΓ ≤ 2(Lip Γ + 1)

(6.15)
JΦΓ (x) ≤ (Lip ΦΓ )n ≤ 2n (Lip Γ + 1)n .

3) Fix f ∈ C1c (Rn ). Since both f and ΦΓ are Lipschitz, so is fΓ = f ◦ΦΓ .

It then follows that both EΓ (f |Ω )|Ω = f |Ω and EΓ (f |Ω )|Ωc = fΓ |Ωc are
c
Lipschitz. Hence EΓ (f |Ω ) is Lipschitz since, if x ∈ Ω and y ∈ Ω , the
closed segment [x, y], being connected, must intersect ∂Ω at some
point z; therefore

EΓ (f |Ω )(x) − EΓ (f |Ω )(y) ≤
≤ EΓ (f |Ω )(x) − EΓ (f |Ω )(z) + EΓ (f |Ω )(z) − EΓ (f |Ω )(y) ≤
(Lip f )kx − zk + (Lip fΓ )kz − yk ≤


≤ max{Lip f, Lip fΓ } kx − zk + kz − yk =
= max{Lip f, Lip fΓ }kx − yk,

from which we conclude that Lip EΓ (f |Ω ) ≤ max{Lip f, Lip fΓ } <

∞.
In particular, from Rademacher’s theorem it follows that EΓ (f |Ω )
is Ln -a.e. Fréchet-differentiable and its classical gradient coincides
with its weak gradient Ln -almost everywhere. Since ∂Ω = gr Γ is
Ln -null, and since EΓ (f |Ω ) coincides with f on the open set Ω and
c
with fΓ = f ◦ ΦΓ on the open set Ω , we conclude that the weak
gradient of EΓ (f |Ω ) is given Ln -a.e. by
(
∇f (x) if x ∈ Ω,
∇w EΓ (f |Ω ) (x) =
 
∗
c
∇fΓ (x) = DΦΓ (x) · ∇f ΦΓ (x) if x ∈ Ω ∩ DΦΓ ,

where DΦΓ (x)∗ denotes the adjoint of DΦΓ (x) with respect to the
standard inner product of Rn . In particular, since ∀x ∈ DΦΓ ,
kDΦΓ (x)∗ k = kDΦΓ (x)k ≤ 2(Lip Γ + 1) by (6.15), it follows that

(6.16) ∇w EΓ (f |Ω ) ≤ χΩ · k∇f k + χΩc · 2(Lip Γ + 1)k(∇f ) ◦ ΦΓ k

 

Ln -a.e. on Rn .
4) We estimate the Lp norms of both EΓ (f |Ω ) and of its weak gradient:
206 6. SOBOLEV SPACES

(6.17)
ˆ ˆ ˆ
p n p n ΦΓ =Φ−1
Γ and AF 5.39.ii)
|EΓ (f |Ω )| dL = |f | dL + |f ◦ ΦΓ |p dLn =
c
Rn Ω Ω
ˆ ˆ (6.15)
= |f | dL + |f |p JΦΓ dLn ≤
p n
Ω Ω
ˆ
≤ [2 (Lip Γ + 1) + 1] |f |p dLn .
n n
Ω

Similarly, it follows from (6.16) that

(6.18)
ˆ
 p n
∇w EΓ (f |Ω )

dL ≤
Rn
ˆ ˆ
p n p p AF 5.39.ii)
≤ k∇f k dL + 2 (Lip Γ + 1) k(∇f ) ◦ ΦΓ kp dLn =
c
Ω Ω
ˆ ˆ (6.15)
p n p p
= k∇f k dL + 2 (Lip Γ + 1) k∇f kp JΦΓ dLn ≤
Ω
ˆ Ω

≤ [2n+p (Lip Γ + 1)n+p + 1] k∇f kp dLn .

Ω

From (6.17) and (6.18), we therefore conclude that EΓ (f |Ω ) ∈

W (Rn ) and, with C := [2n+p (Lip Γ+1)n+p +1]1/p , kEΓ (f |Ω )kW1,p (Rn ) ≤
1,p

Ckf kW1,p (Ω) .


Recall our convention from remark 1.57, i.e. we consider essential
supports only.
Corollary 6.55 (Extension by reflection for Sobolev functions on
Lipschitz epigraphs). With the notation from the previous theorem, if
f ∈ W1,p (Ω) and the closure of spt f in Rn is a compact subset of an
open set V ⊂ Rn , then spt E f b V ∪ VΓ , where VΓ = ΦΓ (V ).
Proof. Let W ⊂ Rn be a relatively compact open set such that
spt f b W b V . We may take, by lemma 6.39, a sequence (gk )k∈N in
C∞ 1,p 1,p
c (W ) such that gk |Ω → f in W (Ω). Then E gk → E f in W (R ).
n

Since ∀k ∈ N, E gk = EΓ gk has compact

´ support in W ∪´WΓ , it follows
∞ c
that, for all ϕ ∈ Cc (W ∪ WΓ ), E f ϕ dLn = limk→∞ E gk ϕ dLn =
0. It then follows from the fundamental lemma of the Calculus of
c
Variations 4.34 that E f = 0 Ln -a.e. on W ∪ WΓ , so that spt E f ⊂
W ∪ WΓ = W ∪ W Γ b V ∪ VΓ , as asserted. 
 6.5. TRACES AND EXTENSIONS 207

Theorem 6.56 (Extension of Sobolev functions on Lipschitz do-

mains). Let n ≥ 2, Ω ⊂ Rn a Lipschitz domain with ∂Ω bounded and
1 ≤ p < ∞. Then there exists an extension operator EΩ : W1,p (Ω) →
W1,p (Rn ). Moreover, if Ω is bounded and V ⊂ Rn is an open set
such that Ω b V , we may choose EΩ so that, for all f ∈ W1,p (Ω),
spt (EΩ f ) b V .

Proof.

1) For each x ∈ ∂Ω, there exists an open set Ux ⊂ Rn such that x ∈ Ux

and Ux is obtained by rigid motion of a cylinder centered at 0 ∈ Rn
as in definition 6.33, i.e. there exists a rigid motion Φ ∈ SE(n) with
n−1
Φ(0) = x and there exists r, h > 0 and
Γ:R → R Lipschitz

with
Γ(0) = 0 such that Ux = Φ
C(0, r, h) , Φ gr Γ∩C(0, r, h) = Ux ∩∂Ω
and Φ epiS Γ ∩ C(0, r, h) = Ux ∩ Ω. If Ω is bounded and V ⊂ Rn
is an open set such that Ω b V , we may take smaller r and h
so that Ux ⊂ V . Moreover, since Γ is continuous and Γ(0) = 0,
taking smaller r if necessary we may assume that |Γ(y)| < h/4 for
y ∈ U(0, r) ⊂ Rn−1 ; with that assumption, using the notation from
the previous corollary, we have C(0, r/2,
h/2) ∪ C(0, r/2, h/2)Γ ⊂
C(0, r, h). Let Wx := Φ C(0, r/2, h/2) b Ux .
2) From the open cover (Wx )x∈∂Ω of the compact set ∂Ω, we may ex-
tract a finite subcover (Wi )1≤i≤N . For each 1 ≤ i ≤ N , let the cor-
responding objects defined in the previous
item be denoted
with a
subscript i, so that Φi C(0, ri /2, hi /2) = Wi , Φi C(0, ri , hi ) = Ui ,
|Γi | < hi /4 on U(0, ri ) ⊂ Rn−1 .
c
Let W0 := Ω and W−1 := Ω , so that (Wi )−1≤i≤N is an open cover
of Rn . We may apply corollary 6.11 to obtain a smooth partition of
unity (ξi )−1≤i≤N of unity of Rn with spt ξi ⊂ Wi for −1 ≤ i ≤ N .
Besides, for 1 ≤ i ≤ N , as spt ξi ⊂ Wi b Rn , it follows that spt ξi
is a compact subset of Wi .
We now define a sequence (Ei )0≤i≤N of bounded linear opera-
tors W1,p (Ω) → W1,p (Rn ) whose sum will be the desired extension
operator.
3) For i = 0. For each f ∈ W1,p (Ω), let E0 (f ) := ξ0 · f . We contend
that ξ0 ∈ W1,∞ (Rn ). Indeed, ξ0 ∈ L∞ (Rn ) (because 0 ≤ ξ0 ≤ 1)
PN PN
and, since i=0 ξi ≡ 1 on Ω, we have ∇ξ0 = − i=1 ∇ξi on Ω,
∞ ∞
hence ∇ξ0 |Ω ∈ L (Ω, R ) (because ξi ∈ Cc (R ) ⊂ W1,∞ (Rn ) for
n n

1 ≤ i ≤ N ). As spt ξ0 ⊂ Ω, our contention is proved. Therefore,

ξ0 ∈ C∞ (Rn ) ∩ W1,∞ (Rn ), with spt ξ0 ⊂ Ω; an application of lemma
6.35 yields E0 (f ) = ξ0 · f ∈ W1,p (Rn ) and ∇w [E0 (f )] = (∇ξ0 ) · f +
208 6. SOBOLEV SPACES

ξ0 · (∇w f ). Hence
kE0 (f )kLp (Ln ) ≤ kf kLp (Ln |Ω )
w
≤ kξ0 kW1,∞ (Rn ) kf kLp (Ln |Ω ) + k∇w f kLp (Ln |Ω ,Rn ) ,


∇ [E0 (f )] Lp (Ln )

thus showing that E0 : W1,p (Ω) → W1,p (Rn ) is a well defined bounded
linear operator.
4) For 1 ≤ i ≤ N . We define Ei : W1,p (Ω) → W1,p (Rn ) as the composite
of the following sequence of continuous linear maps:

Lξi 1,p (◦Φi ) 1,p

e0
W1,p (Ω) −→ W(C) (Wi ∩ Ω) −→ W(C) C(0, ri /2, hi /2) ∩ epiS Γi −→
e E (◦Φ−1 )
0
−→ W1,p (epiS Γi ) −→ W1,p (Rn ) −→
i
W1,p (Rn ),
1,p
where W(C) (Wi ∩ Ω) := {f ∈ W1,p (Wi ∩ Ω) | spt f b Wi } is a linear
1,p


subspace of W1,p (Wi ∩ Ω), W(C) C(0, ri /2, hi /2) ∩ epiS Γi := {f ∈


W1,p C(0, ri /2, hi /2) ∩ epiS Γi | spt f b C(0,

ri /2, hi /2)} is a linear
1,p
subspace of W C(0, ri /2, hi /2) ∩ epiS Γi , and the linear maps are
described below:
a) Lξi is the multiplication by ξi . The fact that ξi ∈ C∞ n
c (R ) and
1,p
the product rule 6.26 imply that, for each f ∈ W (Ω), ξi · f ∈
W1,p (Ω) and spt ξi · f ⊂ spt ξi b Wi , so that Lξi : W1,p (Ω) →
1,p
W(C) (Wi ∩ Ω) is a well-defined linear map. Since ∇w (ξi · f ) =
(∇ξi ) · f + ξi · ∇w f , we have
kLξi (f )kLp (Wi ∩Ω) ≤ kξi ku kf kLp (Ω)
∇w [Lξi (f )] Lp (Wi ∩Ω,Rn )
≤ k∇ξi ku kf kLp (Ω) + kξi ku k∇w f kLp (Ω,Rn ) ,

hence Lξi is continuous.

b) By lemma 6.40, (◦Φi ) : W1,p (Wi ∩ Ω) → W1,p C(0, ri /2, hi /2) ∩
1,p


epiS Γi is a surjective linear isometry and maps W(C) (Wi ∩ Ω)
1,p −1


onto W(C) C(0, ri /2, hi /2)∩epiS Γi , since spt f ◦Φi = Φi (spt f ).
1,p 1,p


Hence (◦Φi ) : W(C) (Wi ∩ Ω) → W(C) C(0, ri /2, hi /2) ∩ epiS Γi is
a surjective linear isometry.
1,p


c) e0 : W(C) C(0, ri /2, hi /2) ∩ epiS Γi → W1,p (epiS Γi ) is the exten-
1,p


sion by 0. Note that, for each f ∈ W(C) C(0, ri /2, hi /2)∩epiS Γi ,
• e0 (f ) ∈ Lp (epiS Γi ) and
ke0 (f )kLp (epiS Γi ) = kf k
;
Lp C(0,ri /2,hi /2)∩epiS Γi
 6.5. TRACES AND EXTENSIONS 209

• since epiS Γi is the union of the open sets C(0, ri /2, hi /2) ∩
epiS Γi and epiS Γi \ spt f , and since e0 (f ) has weak gra-
dients on both open sets (as on the latter its restriction is
null), by the locality of the weak derivative 6.13 we con-
clude that e0 (f ) has weak gradient given by ∇w [e0 (f )] =
e0 (∇w f ) ∈ Lp (epiS Γi , Rn ) and
k∇w [e0 (f )]kLp (epiS Γi ,Rn ) = k∇w f k
.
Lp C(0,ri /2,hi /2)∩epiS Γi ,Rn

We therefore conclude that e0 is a well defined linear isometry

into W1,p (epiS Γi ).
d) E : W1,p (epiS Γi ) → W1,p (Rn ) is the extension by reflection with
respect to Γ, cf. theorem 6.54, hence it is linear continuous.
e) (◦Φ−1 1,p n 1,p n
i ) : W (R ) → W (R ) is a surjective linear isometry, cf.
lemma 6.40. PN
5) It follows from the two previous items that EΩ := i=0 Ei is a
well defined bounded linear operator W (Ω) → W (Rn ). We
1,p 1,p

shall prove that (a) it is an extension operator, i.e. for each f ∈

W1,p (Ω), (EΩ f )|Ω = f and (b) in the case Ω bounded, for each
f ∈ W1,p (Ω), spt (EΩ f ) b V (recall that V is given in the statement
of the theorem).
Fix 1 ≤ i ≤ N and f ∈ W1,p (Ω). Since the closure in Rn

1,p
of the support of (ξi · f ) ◦ Φi ∈ W(C) C(0, ri /2, hi /2) ∩ epiS Γi
is a compact subset of C(0, ri /2, hi /2), and since C(0, ri /2, hi /2) ∪
C(0, ri /2, hi /2)Γi ⊂ C(0, ri , hi ), cf. the end of part 1) of the proof,
it follows from corollary 6.55 that spt E[(ξ
i · f ) ◦ Φi ] b C(0,
ri , hi ).
Thus, spt (Ei f ) = Φi spt E[(ξi · f ) ◦ Φi ] b Φi C(0, ri , hi ) = Ui .
We then conclude that:
• Ei f = 0 Ln -a.e. on Ω \ Ui ;
• if x ∈ Ui ∩ Ω, Φ−1 i (x) ∈ C(0, ri , hi ) ∩ epiS Γi , hence E[(ξi · f ) ◦
Φi ] ◦ Φi (x) = (ξi · f )) ◦ Φi ◦ Φ−1
−1
i (x) = (ξi · f )(x).
We have thus proved that (Ei f )|Ω = ξi · f Ln -a.e. on Ω. Therefore,
Ln -a.e. on Ω,
N
X
Ω
(E f )|Ω = (Ei f )|Ω =
i=0
N
X
= ξi · f = f,
i=0
PN
since i=0 ξi ≡ 1 on Ω. That is, as elements of W1,p (Ω), (EΩ f )|Ω =
f , as asserted. Furthermore, in the case Ω bounded, spt (E0 f ) =
210 6. SOBOLEV SPACES

spt (ξ0 · f ) b Ω ⊂ V and, for 1 ≤ i ≤ N , spt (Ei f ) b Ui ⊂ V ,

whence spt (EΩ f ) ⊂ ∪N
i=0 spt (Ei f ) b V , which concludes the proof.

Remark 6.57. With the same proof and notation above, in the case
in which Ω is unbounded and V ⊂ Rn is an open set which contains
Ω, we may choose EΩ so that, for all f ∈ W1,p (Ω), spt (EΩ f ) ⊂ V , but
not necessarily compact.

6.6. Sobolev Inequalities

In this section, for 1 ≤ p ≤ ∞ we want to find continuous injections
of W1,p (Rn ) into Lq (Ln ) for some q. We divide the problem into cases:
1 ≤ p < n, n < p ≤ ∞ and the limit case p = n.
6.6.1. Case 1 ≤ p < n.
Definition 6.58. Let 1 ≤ p < n. We define the Sobolev conjugate
exponent p∗ to p by
1 1 1
∗
= − ,
p p n
that is
np
p∗ = .
n−p
Theorem 6.59 (Sobolev-Gagliardo-Nirenberg inequality). Let 1 ≤
p < n. Then there exists a constant C = C(n, p) such that, for all
f ∈ C1c (Rn ),
kf kp∗ ≤ Ck∇f kp .
Remark 6.60. For 1 ≤ p < n, there can exist only one q ∈ [1, ∞]
for which Sobolev’s inequality holds: it is precisely the Sobolev con-
jugate to p. That can be deduced by a scaling argument: suppose
that q ∈ [1, ∞] and that for all f ∈ C1c (Rn ), kf kq ≤ Ck∇f kp for
some constant C = C(n, p). Fix f ∈ C1c (Rn ) and λ > 0. Then
fλ given by x 7→ f (λx) belongs to C1c (Rn ) and ∇fλ (x) = λ∇f (λx),
so that kfλ kq = λ−n/q kf kq and k∇fλ kp = λ1−n/p k∇f kp . Therefore,
kfλ kq ≤ Ck∇(λf )kp is equivalent to kf kq ≤ Cλ1−n/p+n/q k∇f kp . The
latter inequality must hold for all f ∈ C1c (Rn ) and λ > 0; if the expo-
nent of λ is not 0, sending λ to 0 or to ∞ yields a contradiction. Hence
1 − n/p + n/q = 0, i.e. 1/q = 1/p − 1/n.
Notation. Let x = (x1 , . . . , xn ) ∈ Rn . For 1 ≤ i ≤ n, we denote
by x̂i ∈ Rn−1 the point obtained by deleting the i-th coordinate of x,
i.e. x̂i = (x1 , . . . , xi−1 , xi+1 , . . . , xn ) ∈ Rn−1 .
 6.6. SOBOLEV INEQUALITIES 211

Lemma 6.61. Let n ≥ 2 and f1 , . . . , fn : Rn−1 → [0, ∞] Borelian

functions on Rn−1 . Define f : Rn → [0, ∞] by
n
Y
f (x) := fi (x̂i ).
i=1

Then f is Borelian on Rn and

n
Y
kf kL1 (Ln ) ≤ kfi kLn−1 (Ln−1 ) .
i=1

In particular, f ∈ L1 (Ln ) if fi ∈ Ln−1 (Ln−1 ) for 1 ≤ i ≤ n.

Proof. For 1 ≤ i ≤ n, let pri : Rn → Rn−1 be the projection
x 7→ x̂i , which is continuous,
Qnhence Borelian; then fi ◦ pri is Borelian,
and so is the product f = i=1 fi ◦ pri .
We prove the asserted inequality by induction on n:
1) For n = 2,
ˆ
Tonelli 1.84
kf k1 = f1 (x2 )f2 (x1 ) dL2 (x1 , x2 ) =
ˆ ˆ
= f1 (x2 ) dx2 · f2 (x1 ) dx1 = kf1 k1 kf2 k1 .

2) Induction step. Suppose that the inequality holds for n. We identify

Rn+1 ≡ Rn × R and use the notation x = (x0 , xn+1 ) for x ∈ Rn+1 .
Fix xn+1 ∈ R. It follows from Hölder’s inequality 1.73 that
ˆ ˆ n+1
Y Hölder 1.73
n
f (x1 , . . . , xn , xn+1 ) dL (x1 , . . . , xn ) = fi (x̂i ) dLn (x1 , . . . , xn ) ≤
i=1
ˆ n
 Y 0 1/n0
≤ kfn+1 kn fi (x̂i )n dLn (x1 , . . . , xn ) ,
i=1

where n0 = n−1 n
is the conjugate exponent of n. By the induction
0
hypothesis with fin (·, xn+1 ) in place of fi , we have:
ˆ Yn n ˆ
n0 0
Y
n
fi (x̂i ) dL (x1 , . . . , xn ) ≤ [ fi (x̂i )n (n−1) dLn−1 (xb0 i )]1/(n−1) .
i=1 i=1

It then follows from the two previous equalities that

ˆ n ˆ
Y
f (x1 , . . . , xn , xn+1 ) dL (x1 , . . . , xn ) ≤ kfn+1 kn [ fi (x̂i )n dLn−1 (xb0 i )]1/n .
n

i=1
212 6. SOBOLEV SPACES

The equality above holds for an arbitrarily fixed xn+1 ∈ R. There-

fore, integrating both members on xn+1 and applying Tonelli’s the-
orem, we obtain:
ˆ ˆ ˆ
n+1
f dL = f (x1 , . . . , xn , xn+1 ) dLn (x1 , . . . , xn ) dL1 (xn+1 ) ≤
R Rn
ˆ n ˆ
Y
≤ kfn+1 kn [ fi (x̂i )n dLn−1 (xb0 i )]1/n dL1 (xn+1 ) =
R i=1
n ˆ
ˆ Y gen. Hölder 1.74
= kfn+1 kn [ fi (x̂i )n dLn−1 (xb0 i )]1/n dL1 (xn+1 ) ≤
R i=1
n ˆ ˆ
Y 1/n Tonelli
fi (x̂i )n dLn−1 (xb0 i ) dL1 (xn+1 )

≤ kfn+1 kn =
i=1 R Rn
n ˆ
Y 1/n
fi (x̂i )n dLn (x̂i )

= kfn+1 kn =
i=1
n+1
Y
= kfi kn ,
i=1

thus proving the induction step.



Proof of theorem 6.59. Let f ∈ C1c (Rn ). For 1 ≤ i ≤ n and

for all x = (x1 , . . . , xn ) ∈ Rn :
ˆ xi
∂f
f (x) = (x1 , . . . , xi−1 , t, xi+1 , . . . , xn ) dt,
−∞ ∂xi

hence
ˆ ∞
∂f
|f (x)| ≤ (x1 , . . . , xi−1 , t, xi+1 , . . . , xn ) dt.
−∞ ∂xi

It then follows that

n ˆ
n Y  ∂f 1/(n−1)
|f (x)| n−1 ≤ (x1 , . . . , xi−1 , t, xi+1 , . . . , xn ) dt .
i=1 R ∂xi
| {z }
=:fi (x̂i )
 6.6. SOBOLEV INEQUALITIES 213

We therefore conclude from lemma 6.61 that

ˆ n ˆ ˆ
n/(n−1) n
Y  ∂f 1/(n−1) Tonelli
|f | dL ≤ (x1 , . . . , xi−1 , t, xi+1 , . . . , xn ) dt dx̂i =
i=1 Rn−1 R ∂x i
n ˆ
Y ∂f 1/(n−1)
dLn

= ≤
i=1
∂x i
n ˆ ˆ
n 1/(n−1)
Y n/(n−1)
k∇f k dLn
  
≤ k∇f k dL = ,
i=1

which proves the thesis for p = 1 with C = 1.

For 1 < p < ∞, let f ∈ C1c (Rn ) and g := |f |γ , with γ > 1 to be
chosen later. Note that g ∈ C1c (Rn ) and ∇g = γ · sgn f · |f |γ−1 · ∇f .
We may therefore apply to g the inequality already proved, i.e. with
p = 1, which yields
ˆ γn
 n−1
n
ˆ
Hölder
n
|f | n−1 dL ≤ γ |f |(γ−1) k∇f k dLn ≤
ˆ (γ−1)p
 p−1
p
≤γ |f | p−1 k∇f kp .

We choose γ satisfying
γn (γ − 1)p
= ,
n−1 p−1
i.e.
(n − 1)p
γ= > 1.
n−p
Then
np γn (γ − 1)p
p∗ = = = .
n−p n−1 p−1
We then conclude that
ˆ  n−1 ˆ  p−1
p∗ n n p∗ p
|f | dL ≤γ |f | k∇f kp ,

which yields the thesis with C = γ = γ(n, p).


∗
Corollary 6.62. For 1 ≤ p < n, W1,p (Rn ) ⊂ Lp (Rn ) and the
Sobolev-Gagliardo-Nirenberg inequality 6.59 holds for all f ∈ W1,p (Rn ).
∗
In particular, the inclusion W1,p (Rn ) ⊂ Lp (Rn ) is continuous.
214 6. SOBOLEV SPACES

Proof. Let f ∈ W1,p (Rn ). By corollary 6.21, there exists a se-

quence (fk )k∈N in C∞ n 1,p n
c (R ) such that fk → f in f ∈ W (R ). Pass-
ing to a subsequence, we may assume that fk → f Ln -almost ev-
erywhere. On the other hand, by theorem 6.59, for all j, k ∈ N,
kfj − fk kp∗ ≤ Ck∇fj − ∇fk kp ≤ Ckfj − fk kW1,p (Rn ) . That is, (fk )k∈N
∗ ∗
is a Cauchy sequence in Lp (Ln ). Hence it is convergent in Lp (Ln ),
∗
and since it converges Ln -a.e. fo f , we conclude that f ∈ Lp (Ln ) and
∗
fk → f in Lp (Ln ). Therefore, since for all k ∈ N, kfk kp∗ ≤ Ck∇fk kp ,
taking the limit as k → ∞ in both members yields
kf kp∗ ≤ Ck∇w f kp ,
as asserted. 
Corollary 6.63. Let 1 ≤ p < n and Ω ⊂ Rn a Lipschitz do-
∗
main with ∂Ω bounded. Then W1,p (Ω) ⊂ Lp (Ln |Ω ) with continuous
inclusion.
Proof. Let E : W1,p (Ω) → W1,p (Rn ) be an extension operator,
cf. theorem 6.56, and C = C(n, p) given by the Sobolev-Gagliardo-
Nirenberg inequality 6.59. Then, for all f ∈ W1,p (Ω),
kf kLp∗ (Ln |Ω ) ≤ kE f kLp∗ (Ln ) ≤
≤ Ck∇w (E f )kLp (Ln ,Rn ) ≤ CkE f kW1,p (Rn ) ≤
≤ CkEkkf kW1,p (Ω) .

The Poincaré’s inequality, proved below, is a kind of local version
of the Sobolev-Gagliardo-Nirenberg inequality 6.59, 6.62.
Lemma 6.64. Let X, Y be metric spaces and f : X → Y bi-
Lipschitz
with Lip f −1 = (Lip f )−1 . Then, for all s ≥ 0 and A ⊂ X,
Hs f (A) = (Lip f )s Hs (A), i.e. f −1 # Hs = (Lip f )s Hs .
Proof.
Hs f (A) ≤ (Lip f )s Hs (A) ≤

≤ (Lip f )s (Lip f −1 )s Hs f (A) = Hs f (A) .


Lemma 6.65. For each 1 ≤ p < ∞, there exists a constant C =
C(n, p) such that, for all B(x, r) ⊂ Rn , f ∈ C1 (Rn ) and z ∈ B(x, r),
ˆ ˆ
p n+p−1
|f (y) − f (z)| dy ≤ Cr k∇f (y)kp ky − zk1−n dy.
B(x,r) B(x,r)
 6.6. SOBOLEV INEQUALITIES 215

Proof. 1) For y, z ∈ B(x, r), we have

ˆ 1 ˆ 1
d


f (y)−f (z) = f z +t(y −z) dt = ∇f z +t(y −z) ·(y −z) dt.
0 dt 0

Then
ˆ 1 p Hölder
p p


f (y) − f (z) ≤ |y − z| k∇f z + t(y − z) dtk ≤
0
ˆ 1
≤ |y − z|p k∇f z + t(y − z) kp dt.


0

x
For s > 0, Hn−1 ∂B(z, s) is a finite Radon measure - the trace
Hn−1 |∂B(z,s) actually coincides with the usual Lebesgue measure of
the sphere, cf. exercise 5.42. We may therefore apply Fubini-
Tonelli’s theorem to the product measure L1 ⊗ (Hn−1 ∂B(z, s)) x
in equality (∗) of the following computation:
ˆ
p
f (y) − f (z) dHn−1 (y) ≤
B(x,r)∩∂B(x,s)
ˆ ˆ 1
p (∗)
k∇f z + t(y − z) kp dt dHn−1 (y) =


≤ |y − z|
B(x,r)∩∂B(x,s) 0
ˆ 1ˆ
p
k∇f z + t(y − z) kp dHn−1 (y) dt =


=s
0 B(x,r)∩∂B(x,s) | {z }
=:g(y)
ˆ 1 ˆ
= sp χB((1−t)z+tx,tr)∩∂B(z,ts) ◦ g(y)k(∇f ) ◦ g(y)kp dHn−1 (y) dt =
0
ˆ 1 ˆ
p
=s χB((1−t)z+tx,tr)∩∂B(z,ts) k(∇f )kp d( g# Hn−1 ) dt =
0 | {z }
=t1−n Hn−1 by lemma 6.64
ˆ 1 ˆ
1 B((1−t)z+tx,tr)⊂B(x,r)
= sp k∇f (w)kp dHn−1 (w) dt ≤
0 tn−1 B((1−t)z+tx,tr)∩∂B(z,ts)
ˆ 1 ˆ
p 1
≤s k∇f (w)kp dHn−1 (w) dt =
0 tn−1 B(x,r)∩∂B(z,ts)
ˆ 1 ˆ
n+p−1 1
=s n−1
k∇f (w)kp dHn−1 (w) dt =
0 (ts) B(x,r)∩∂B(z,ts)
ˆ 1ˆ
= sn+p−1 k∇f (w)kp kw − zk1−n dHn−1 (w) dt.
0 B(x,r)∩∂B(z,ts)
216 6. SOBOLEV SPACES

The latter integral may now be computed by means of the coarea

formula 5.50 with the Lipschitz map φ : Rn → R given by
kw − zk 1
φ(w) = , Jφ = Ln -q.s., φ−1 {t} = ∂B(z, ts)
s s
which yields, for all s > 0,
ˆ
p
f (y) − f (z) dHn−1 (y) ≤
ˆ
B(x,r)∩∂B(x,s)

n+p−2
≤s k∇f (w)kp kw − zk1−n dLn (w) ≤
ˆ
B(x,r)∩B(z,s)

≤ sn+p−2 k∇f (w)kp kw − zk1−n dLn (w).

B(x,r)

We now integrate both members of the inequality above from s =

0 to s = 2r, applying once more the coarea formula 5.52 for the
integral of the first member, which yields
ˆ
p
f (y) − f (z) dLn (y) ≤
B(x, r) ∩ B(z, 2r)
| {z }
=B(x,r)
ˆ
(2r)n+p−1
≤ k∇f (w)kp kw − zk1−n dLn (w),
n + p − 1 B(x,r)
whence the thesis with
2n−p+1
C= .
n−p+1

ffl
Notation. For f ∈ L1loc (Ln ) we define the average (f )x,r = B(x,r)
f dLn
of f on B(x, r) by
ˆ
n 1
(f )x,r := f dL := n
f dLn .
B(x,r) L B(x, r)
Theorem 6.66 (Poincaré’s inequality). For 1 ≤ p < n, there exists

a constant C = C(n, p) such that, for all B(x, r) and f ∈ W1,p U(x, r) ,
 1/p∗  1/p
p∗ n w p n
|f − (f )x,r | dL ≤ Cr k∇ f k dL .
B(x,r) B(x,r)

Equivalently,
kf − (f )x,r kLp∗ (B(x,r)) ≤ C 0 k∇w f kLp∗ (B(x,r))
for some constant C 0 = C 0 (n, p) (compare with the Sobolev-Gagliardo-
Nirenberg inequality 6.59).
 6.6. SOBOLEV INEQUALITIES 217

Proof. 1) Let f ∈ C∞ n n
c (R ). For all B(x, r) ⊂ R , we compute

|f − (f )x,r |p dLn =
B(x,r)
p Hölder
= [f (y) − f (z)] dz dy ≤
B(x,r) B(x,r)

p lemma 6.65
≤ f (y) − f (z) dz dy ≤
B(x,r) B(x,r)
ˆ
n+p−1
r Fubini
≤ C n
k∇f (z)kp ky − zk1−n dz dy =
B(x,r) α(n)r B(x,r)
p−1 ˆ
Cr
= k∇f (z)kp ky − zk1−n dy dz ≤
α(n) B(x,r) B(x,r)

≤ 2Cnrp k∇f (z)kp dz.

B(x,r)

where, in the last inequality, we have estimated, for z ∈ B(x, r),

ˆ ˆ
1−n
ky − zk dy ≤ ky − zk1−n dy =
B(x,r) B(z,2r)
ˆ 2r
= nα(n) ρ1−n ρn−1 dρ = 2rnα(n).
0

2) Claim: there exists a constant C = C 0 (n, p) such that, for all g ∈

0

C∞ n
c (R ) and B(x, r) ⊂ R ,
n

 1/p∗  1/p
p∗ n 0 p p n p n
|g| dL ≤C r k∇gk dL + |g| dL .
B(x,r) B(x,r) B(x,r)

Indeed:
a) For x = 0 and r = 1, we have, taking g := g|U (0, 1) and E :
W1,p (U(0, 1)) → W1,p (Rn ) an extension operator, cf. theorem
6.56:
ˆ 1/p∗ ˆ 1/p∗ cor. 6.62
p∗ n p∗ n
|g| dL ≤ |E g| dL ≤
Rn
ˆ
B(0,1)
1/p
≤C k∇w (E g)kp dy ≤
Rn
≤ CkEk kgkW1,p (U(0,1)) =
| {z }
=:C 0 =C 0 (n,p)
ˆ 1/p
0 p p n
=C k∇gk + |g| dL
U(0,1)
218 6. SOBOLEV SPACES

b) For arbitrary x ∈ Rn and r > 0, let ge ∈ C∞ n

c (R ) be given by
ge(y) := g(ry + x), so that
∗ ∗
g |p dLn =
|e |g|p dLn
B(0,1) B(x,r)

g |p dLn =
|e |g|p dLn
B(0,1) B(x,r)

g (y)kp dLn (y) =

k∇e kr∇g(ry + x)kp dLn (y) =
B(0,1) B(0,1)

= rp k∇gkp dLn ,
B(x,r)

hence the claim follows from part a) applied to ge in place of g.

3) Applying part 2) to g := f − (f )x,r ∈ C∞ n
c (R ), we obtain
 ∗
1/p∗
|f − (f )x,r |p dLn ≤
B(x,r)
 1/p by 1)
≤ C 0 rp k∇f kp dLn + |f − (f )x,r |p dLn ≤
B(x,r) B(x,r)
 1/p
≤ C 0 rp k∇f kp dLn + 2Cnrp k∇f kp dLn
B(x,r) B(x,r)

 1/p
≤ C 0 + C 0 (2Cn)1/p r k∇f kp dLn ,
| {z } B(x,r)
=C(n,p)

thus reaching the thesis for f ∈ C∞ n

c (R ).

n 1,p
4) Let x ∈ R , r > 0 and f ∈ W U(x, r) . By corollary 6.43,
there exists
a sequence (fi )i∈N in C∞ n
c (R ) such that fi → f in
1,p
W U(x, r) . For each i ∈ N, it follows from the previous step
of the proof that
 1/p∗  1/p
p∗ n p n
|fi − (fi )x,r | dL ≤ Cr k∇fi k dL .
B(x,r) B(x,r)

As i → ∞, the second member of the inequality above has limit

ffl 1/p

w p n
Cr B(x,r) k∇ f k dL , since ∇fi → ∇w f in Lp U(x, r), Rn . We
ffl 1/p∗
∗
contend that the first member has limit B(x,r) |f − (f )x,r |p dLn ,
whence the thesis.
Indeed, applying the previous inequality with fi − fj in place of f
i ,
∗
we conclude that the sequence {fi −(fi )x,r }i∈N is Cauchy in Lp U(x, r) ,
hence convergent in that space. Its limit must be f − (f )x,r because
 6.6. SOBOLEV INEQUALITIES 219

(fi )x,r → (f )x,r and, passing to a subsequence if necessary, fi → f

Ln -a.e., hence fi − (fi )x,r → f − (f )x,r Ln -a.e., thus proving our con-
tention. 

6.6.2. Case n < p.

Definition 6.67 (Hölder spaces). Let Ω be an open subset of Rn
and 0 < γ ≤ 1. We say that f : Ω → R is Hölder continuous with
exponent γ on Ω if there exists a constant C ≥ 0 such that, for all
x, y ∈ Ω,
|f (x) − f (y)| ≤ Ckx − ykγ .
Such functions form a linear subspace of RΩ . We shall denote by
C0,γ (Ω) the linear subspace of RΩ of all bounded Hölder continuous
functions on Ω.
Note that, for γ = 1, the definition above is equivalent to f being
Lipschitz.
Definition 6.68 (Hölder seminorm). Let Ω be an open subset of
Rn , 0 < γ ≤ 1 and f : Ω → Rn . We define the C0,γ seminorm of f by
|f (x) − f (y)|
[f ]C0,γ (Ω) := sup{ | x 6= y ∈ Ω} ∈ [0, ∞].
kx − ykγ
With the notation above, note that f is Hölder continuous with
exponent γ on Ω if, and only if, [f ]C0,γ (Ω) < ∞.
Proposition 6.69 (Hölder spaces are Banach). Let Ω be an open
subset of Rn and 0 < γ ≤ 1. Then C0,γ (Ω) is a Banach space endowed
with the norm
kf kC0,γ (Ω) := kf ku + [f ]C0,γ (Ω)
Definition 6.70. Let Ω be an open subset of Rn and f : Ω → R.
We say that f ∗ : Ω → R is a version of f if f ∗ = f Ln -a.e. on Ω.
Theorem 6.71 (Morrey’s inequality). Fix n < p < ∞.
i) There exists a constant C
= C(n, p) such that, for all B(x, r) ⊂ Rn
and all f ∈ W1,p U(x, r) ,
 1/p
(6.19) |f (y) − f (z)| ≤ Cr k∇w f kp dLn
B(x,r)
n
for L -a.e. y, z in U(x, r).
ii) If f ∈ W1,p (Rn ), then the limit
f ∗ (x) := lim(f )x,r
r→0
220 6. SOBOLEV SPACES

exists for every x ∈ Rn and f ∗ is a Hölder continuous version of

f with exponent γ = 1 − n/p, with

[f ∗ ]C0,γ (Rn ) ≤ Ck∇w f kLp (Rn ) ,

where C = C(n, p) is the constant from part i).

Remark 6.72. See theorem 6.44 for the case p = ∞.

Proof.
1) Let f ∈ C1 (Rn ). Taking C = C(n, 1) given by lemma 6.65 with
p = 1, we compute, for all B(x, r) ⊂ Rn and y, z ∈ U(x, r),

|f (y) − f (z)| = |f (y) − f (z)| dw ≤

B(x,r)

6.65
≤ |f (y) − f (w)| + |f (z) − f (w)| dw ≤
ˆ
B(x,r)
C n 1−n 1−n

Hölder
≤ r k∇f (w)k ky − wk + kz − wk dw ≤
α(n)rn B(x,r)
ˆ  p−1 ˆ 1
C  1−n

p
1−n p−1 p p n p
≤ ky − wk + kz − wk dw k∇f k dL .
α(n) B(x,r) B(x,r)

Since
ˆ p  p−1
p Minkowski
ky − wk1−n + kz − wk1−n

p−1
dw ≤
B(x,r)
ˆ p  p−1
p
ˆ p  p−1
p
1−n
kz − wk1−n

p−1
p−1
≤ ky − wk dw + dw ≤
B(x,r) B(x,r)
ˆ p  p−1
p
ˆ p  p−1
p
1−n 1−n

p−1
p−1
≤ ky − wk dw + kz − wk dw ≤
B(y,2r) B(z,2r)
 ˆ 2r  p−1
(1−n)p p p>n
n−1
≤ 2 α(n) ρ ρ p−1 dρ =
0
| {z }
1−n
=ρ p−1

p−1
 (2r) −n+p
p−1 
p−1
p
= 2α(n) p
−n+p =
p−1
−n+p
= C(n, p)r p ,
 6.6. SOBOLEV INEQUALITIES 221

we obtain
ˆ  p1
1−n/p p n
|f (y) − f (z)| ≤ C(n, p)r k∇f k dL =
B(x,r)
 1/p
= C(n, p)r k∇f kp dLn
B(x,r)

thus proving part i) for f ∈ C1 (Rn ).

2) Let x ∈ Rn , r > 0 and f ∈ W1,p U(x, r) . By corollary 6.43,
there exists
a sequence (fi )i∈N in C∞ n
c (R ) such that fi → f in
W1,p U(x, r) . Passing to a subsequence, if necessary, we may as-
sume that fi → f on the complement of a Ln -null set N ⊂ U(x, r).
With C = C(n, p) obtained in part 1), we have, for all i ∈ N and all
y, z ∈ U(x, r),
 1/p
|fi (y) − fi (z)| ≤ Cr k∇fi kp dLn .
B(x,r)

Taking i → ∞, it follows that (6.19) holds for all y, z ∈ U(x, r) \ N ,

which concludes the proof of part i).
3) Let f ∈ C1 (Rn ) and x 6= y in Rn . We take r = kx − yk and C =
C(n, p) obtained in step 1 of the proof, which yields the estimate
ˆ 1/p
1−n/p p n
|f (x) − f (y)| ≤ Ckx − yk k∇f k dL ≤
B(x,r)

≤ Ck∇f kLp (Ln ,Rn ) kx − yk1−n/p .

4) Let f ∈ W1,p (Rn ). By corollary 6.21, there exists a sequence (fi )i∈N
in C∞ n 1,p n
c (R ) such that fi → f in W (R ). Passing to a subsequence,
if necessary, we may assume that fi → f on the complement of a
Ln -null set N ⊂ Rn . By the previous step, for each i ∈ N and x 6= y
in Rn , we have
|fi (x) − fi (y)| ≤ Ck∇fi kLp (Ln ,Rn ) kx − yk1−n/p .
Therefore, taking i → ∞ in the previous equality, we conclude that,
for x, y ∈ Rn \ N ,
(6.20) |f (x) − f (y)| ≤ Ck∇w f kLp (Ln ,Rn ) kx − yk1−n/p .
In particular, f is uniformly continuous on Rn \ N , which is dense
in Rn because N is Ln -null (thus it has empty interior). Hence
f |Rn \N may be extended to a continuous function fe : Rn → R,
which therefore coincides with f Ln -a.e. on Rn , i.e. fe is a version of
222 6. SOBOLEV SPACES

f . By continuity, fe satisfies (6.20) for all x, y ∈ Rn , i.e. fe is Holder

continuous with exponent γ = 1 − 1/n and
[fe]C0,γ (Rn ) ≤ Ck∇w f kLp (Rn ) .
Finally, for all x ∈ Rn and all r > 0, (f )x,r = (fe)x,r , because
f = fe Ln almost everywhere. Since fe is continuous, it follows that
∃ limr→0 (f )x,r = limr→0 (fe)x,r = fe(x), i.e. f ∗ = fe, which concludes
the proof of part ii).

Remark 6.73. With the notation from the previous theorem, for
n < p the map W1,p (Rn ) → C(Rn ) given by f 7→ f ∗ is injective: if
f ∗ = g ∗ , then f = g Ln -a.e., hence they represent the same equivalence
class in W1,p (Rn ). Thus, identifying each element f of W1,p (Rn ) with
its continuous version f ∗ , we obtain an inclusion W1,p (Rn ) ⊂ C(Rn ).
We shall see in the next corollary that we actually have a continuous
inclusion W1,p (Rn ) ⊂ C0,1−n/p (Rn ).
Corollary 6.74. If n < p < ∞, then W1,p (Rn ) ⊂ C0,γ (Rn ), where
γ = 1 − n/p, with continuous inclusion.
Proof.
1) Let f ∈ C∞ n n
c (R ) and fix x ∈ R . Taking C = C(n, 1) given by
lemma 6.65 with p = 1, we compute,
6.65
|f (x)| ≤ |f (x) − f (y)| dy + |f (y)| dy ≤
ˆ
B(x,1) B(x,1)
C Hölder
≤ k∇f (y)kky − xk1−n dy + C(n, p)kf k p
≤
α(n) B(x,1) L B(x,1)

 ˆ  p−1
(1−n)p p
≤ C(n, p)k∇f kLp (Rn ) ky − xk p−1 dy +C(n, p)kf kLp (Rn ) ≤
B(x,1)
| {z }
(1−n)p
=C(n,p)<∞, since p−1
>−n

≤ C(n, p)kf kW1,p (Rn ) .

Taking the sup over x ∈ Rn on the first member of the above in-
equality, it follows that f is bounded and
kf ku ≤ C(n, p)kf kW1,p (Rn ) .
2) Let f ∈ W1,p (Rn ). By corollary 6.21, there exists a sequence (fi )i∈N
in C∞ n 1,p n
c (R ) such that fi → f in W (R ). Passing to a subsequence,
if necessary, we may assume that fi → f on the complement of a
 6.7. COMPACTNESS 223

Ln -null set N ⊂ Rn . By the previous step, for each i ∈ N and x in

Rn , we have
|fi (x)| ≤ C(n, p)kfi kW1,p (Rn ) .
Therefore, taking i → ∞ in the previous equality, we conclude that,
for x ∈ Rn \ N ,
|f (x)| ≤ C(n, p)kf kW1,p (Rn ) .
In view of part ii) of theorem 6.71, we therefore conclude that f ∗
is bounded and Hölder continuous with exponent γ = 1 − n/p, i.e.
f ∗ ∈ C0,γ (Rn ), with
kf ∗ kC0,γ (Rn ) ≤ C(n, p)kf kW1,p (Rn ) .

Corollary 6.75. If n < p < ∞ and Ω ⊂ Rn is a Lipschitz domain
with ∂Ω bounded, then W1,p (Ω) ⊂ C0,γ (Ω), where γ = 1 − n/p, with
continuous inclusion.
Proof. Let E : W1,p (Ω) → W1,p (Rn ) be an extension operator, cf.
theorem 6.56. The inclusion W1,p (Ω) ⊂ C0,γ (Ω) is the composite of the
following sequence of continuous linear maps:
E
W1,p (Ω) → W1,p (Rn ) → C0,γ (Rn ) → C0,γ (Ω),
where the last arrow is the restriction f 7→ f |Ω and the middle arrow
is the inclusion from the previous corollary. 

6.7. Compactness
Lemma 6.76. Let 1 ≤ p < n and 1 ≤ q < p∗ , where p∗ is the
Sobolev conjugate of p. Let (fi )i∈N be a bounded sequence in W1,p (Rn ).
Suppose that there is a relatively compact open set V ⊂ Rn such that,
for all i ∈ N, spt fi b V . Then there exists a subsequence (fik )k∈N of
(fi )i which is convergent in Lq (Ln ).
Note that, for 1 ≤ q ≤ p∗ and for each i ∈ N, fi ∈ Lq (Ln ), in view
of Sobolev-Gagliardo-Nirenberg inequality 6.62 and of the fact that
spt fi b V . However, we cannot ensure the existence of a subsequence
as in the statement of the lemma for q = p∗ .
We give two proofs for this lemma.
Proof 1. Let (φ )>0 be the standard mollifier on Rn . For each
 > 0 and i ∈ N, we define
fi := φ ∗ fi ∈ C∞ n
c (R ).
224 6. SOBOLEV SPACES

Substituting V with V + U(0, 1), we may assume that, for all i ∈ N

and for all 0 <  < 1,

spt fi b V.

1) Claim 1: fi → fi as  → 0 on Lq (Ln ), uniformly on i ∈ N.

Indeed:
a) Fix 0 <  < 1. For each i ∈ N, by corollary 6.43 we may take
gi ∈ C∞ n
c (R ) such that kfi − gi kW1,p (Rn ) < . Moreover, since
spt fi b V , we may assume spt gi b V .
For each i ∈ N and x ∈ Rn , we have

ˆ
z=−1 y
gi ∗ φ (x) − gi (x) = [gi (x − y) − gi (x)]φ (y) dy =
ˆ
B(0,)

= [gi (x − z) − gi (x)]φ(z) dz =

B(0,1)
ˆ ˆ 1
d
= φ(z) [gi (x − tz)] dt dz =
B(0,1) 0 dt
ˆ ˆ 1
=− φ(z) ∇gi (x − tz) · z dt dz.
B(0,1) 0

Thus

ˆ
kgi − gi k1 = |gi (x) − gi (x)| dx ≤
Rn
ˆ ˆ ˆ 1
Tonelli
≤ φ(z) k∇gi (x − tz)k dt dz dx =
Rn B(0,1) 0
ˆ ˆ 1 ˆ
= k∇gi (x − tz)k dx dt dz dx =
φ(z)
0 Rn
ˆ ˆ
B(0,1)
Hölder
n spt gi ⊂V
= k∇gi k dL =  k∇gi k dLn ≤
Rn V
p−1
n
≤ L (V ) p k∇gi kp ≤
p−1
≤ Ln (V )


p  k∇fi kp +  .
 6.7. COMPACTNESS 225

b) Hence, for each i ∈ N, we have

Hölder, Young 1.108.g) and a)
kfi − fi k1 ≤ kfi − gi k1 + kgi − gi k1 + kgi − fi k1 ≤
p−1 p−1
≤ 2Ln (V ) kfi − gi kp + Ln (V )


p p  k∇fi kp +  ≤
p−1
≤ Ln (V ) p 2 +  + k∇fi kp ≤


p−1
≤  Ln (V )


p 3 + sup{k∇fi kp | i ∈ N} .
| {z }
=:C<∞

c) It then follows from the interpolation inequality 1.77, with λ ∈

(0, 1] given by 1q = λ + 1−λ
p∗
, that, for each i ∈ N,
6.62
kfi − fi kq ≤ kfi − fi kλ1 kfi − fi kp1−λ
∗ ≤
≤ (C)λ k∇(fi − fi )kp1−λ ≤
| {z }

1−λ
≤ 2 sup{k∇fi kp |i∈N}
0 λ
≤C ,

1−λ
where C 0 = C λ 2 sup{k∇fi kp | i ∈ N} < ∞, which concludes
the proof of claim 1.
2) Claim 2: for each 0 <  < 1 fixed, (fi )i∈N is uniformly bounded and
equicontinuous.
Indeed, for all i ∈ N, it follows from Young’s inequality 1.108.g)
that:
C
kfi k∞ = kφ ∗ fi k∞ ≤ kφ k∞ kfi k1 ≤ n < ∞,

C0
k∇fi k∞ = k∇φ ∗ fi k∞ ≤ k∇φ k∞ kfi k1 ≤ n+1 < ∞,

whence the claim.
3) Claim 3: for each δ > 0, there exists a subsequence (fik )k∈N of (fi )i
such that
lim supkfij − fik kLq (Ln ) ≤ δ.
j,k→∞
To prove claim 3, we firstly apply claim 1 to find 0 <  < 1 such
that, for all i ∈ N, kfi − fi kLq (Ln ) ≤ 2δ .
Since spt fi b V b Rn for all i ∈ N, in view of claim 2 we
may apply Arzelà-Ascoli’s theorem to find a subsequence (fik )k∈N
such that fik is uniformly convergent on Rn . Since Ln (V ) < ∞, this
subsequence is also convergent in Lq (Ln ), so that
lim supkfij − fik kLq (Ln ) = 0.
j,k→∞
226 6. SOBOLEV SPACES


Finally, since kfij − fik kq ≤ kfij − fi,j kq + kfij + fik kq + kfik − fik kq ,
the claim follows.
4) We now apply claim 3 for δ = 1/m, m ∈ N, yielding for each m ∈ N
a subsequence f m = (fim )i∈N of (fim−1 )i∈N , with f 0 = (fi )i∈N , such
that, for all m ∈ N,
1
lim supkfjm − fkm kLq (Ln ) ≤ .
j,k→∞ m
m
The diagonal (fm )m∈N is therefore a subsequence of (fi )i which is
q n
Cauchy in L (L ), hence convergent in that space.

Proof 2. We apply the Kolmogorov-Riesz-Fréchet compactness
criterion 1.80.
Note that, in view of Sobolev-Gagliardo-Nirenberg inequality 6.62,
∗
(fi )i∈N is bounded on Lp (Ln ). Thus, since Ln (V ) < ∞, it follows that,
for 1 ≤ q < p∗ ,
(6.21) sup{kfi kLq (Ln ) | i ∈ N} < ∞.
On the other hand, it follows from exercise 6.25 for p > 1 and from
exercise 7.38 for p = 1 that, for all h ∈ Rn and all i ∈ N,
kτh fi − fi kLp (Ln ) ≤ khk sup{k∇w fi kLp (Ln ) | i ∈ N} .
| {z }
=:C<∞

Thus, in view of Hölder’s inequality, for all h ∈ Rn and all i ∈ N,

p−1
kτh fi − fi kL1 (Ln ) ≤ khkCLn (V ) p .
Therefore, applying the interpolation inequality 1.77, it follows that,
for 1 ≤ q < p∗ , h ∈ Rn and i ∈ N,
p−1
λ
kτh fi −fi kLq (Ln ) ≤ khkλ CLn (V ) p sup{kτh fi − fi kLp∗ (Ln ) | i ∈ N}1−λ ,
| {z }
:=C 0

where λ ∈ (0, 1] is given by

1 1−λ
=λ+ .
q p∗
Since
sup{kτh fi − fi kLp∗ (Ln ) | i ∈ N} ≤ 2 sup{kfi kLp∗ (Ln ) | i ∈ N} < ∞,
we have C 0 < ∞ and, since λ > 0, we conclude that
(6.22) lim kτh fi − fi kLq (Ln ) = 0
h→0

uniformly in i ∈ N.
 6.7. COMPACTNESS 227

With (6.21) and (6.22) in force, we may apply the Kolmogorov-

Riesz-Fréchet compactness criterion 1.80 to F := {fi | i ∈ N}. There-
fore, F|V has compact closure in Lq (Ln |V ); since each f ∈ F has sup-
port in V , we conclude that F has compact closure in Lq (Ln ), whence
the thesis.

Theorem 6.77 (Rellich-Kondrachov). Let Ω be a bounded Lipschitz
domain in Rn , 1 ≤ p < n and 1 ≤ q < p∗ , where p∗ is the Sobolev
conjugate of p. Then
W1,p (Ω) b Lq (Ln |Ω ),
i.e. W1,p (Ω) ⊂ Lq (Ln |Ω ) with compact inclusion.
∗
Proof. We have continuous inclusions W1,p (Ω) ⊂ Lp (Ln |Ω ) ⊂
L (Ln |Ω ), the first in view of corollary 6.63 and the second in view
q

of the fact that Ln (Ω) < ∞ and of Hölder’s inequality.

Therefore, it suffices to show that each bounded sequence (fi )i∈N in
W (Ω) has a subsequence which is convergent in Lq (Ln |Ω ).
1,p

Let V be an open relatively compact subset of Rn such that Ω b

V b Rn and E : W1,p (Ω) → W1,p (Rn ) an extension operator, cf. the-
orem 6.56, such that spt E f b V for all f ∈ W1,p (Ω). Then (E fi )i∈N
is a bounded sequence in W1,p (Rn ) with spt fi b V for all i ∈ N. We
may therefore apply lemma 6.76 to obtain a subsequence (fij )j∈N such
that (E fij )j∈N is convergent in Lq (Ln ). Since fi = E fi |Ω for all i ∈ N,
we conclude that (fij )j∈N is convergent in Lq (Ln |Ω ). 
Corollary 6.78. Let Ω be a bounded Lipschitz domain in Rn ,
1 ≤ p < n and (fi )i∈N a bounded sequence in W1,p (Ω). Then there exists
a subsequence (fij )j∈N of (fi )i∈N which is convergent in each Lq (Ω), for
1 ≤ q < p∗ .
Proof. Let (qm )m∈N be a sequence in [1, p∗ ) which increases to
p∗ . For each m ∈ N, we may apply theorem 6.77 to find a subsequence
f m = (fim )i∈N of f m−1 , with f 0 = (fi )i∈N , such that f m is convergent in
Lqm (Ω) (hence on Lq (Ω) for 1 ≤ q ≤ qm , since Ln (Ω) < ∞). The diag-
m
onal (fm )m∈N is therefore a subsequence of (fi )i∈N which is convergent
in each Lq (Ω), for 1 ≤ q < p∗ . 
Corollary 6.79. Let Ω be a bounded Lipschitz domain in Rn ,
1 < p < n and (fi )i∈N a bounded sequence in W1,p (Ω). Then there exists
f ∈ W1,p (Ω) and a subsequence (fij )j∈N of (fi )i∈N such that fij → f in
each Lq (Ω), for 1 ≤ q < p∗ .
228 6. SOBOLEV SPACES

Proof. Let (fij )j∈N be a subsequence of (fi )i∈N which is convergent

in each Lq (Ω), for 1 ≤ q < p∗ , cf. corollary 6.78. We contend that its
limit f belongs to W1,p (Ω). Indeed, since 1 ≤ p < p∗ , f ∈ Lp (Ω) and
fij → f in Lp (Ω). As (∇fij )j∈N is bounded in Lp (Ln |Ω , Rn ), it follows
from proposition 6.3.ii) that f ∈ W1,p (Ω), as asserted. 
Corollary 6.80. Let Ω be a bounded Lipschitz domain in Rn and
∗
(fi )i∈N a bounded sequence in W1,1 (Ω). Then there exists f ∈ L1 (Ln |Ω )
and a subsequence (fij )j∈N of (fi )i∈N such that fij → f in each Lq (Ω),
for 1 ≤ q < 1∗ .
Proof. Let (fij )j∈N be a subsequence of (fi )i∈N which is convergent
in each Lq (Ω), for 1 ≤ q < 1∗ , cf. corollary 6.78. We contend that
∗
its limit f belongs to L1 (Ln |Ω ). Indeed, since (fij )j∈N is bounded in
∗
L1 (Ln |Ω ) (because it is bounded in W1,p (Ω) and corollary 6.63 may be
applied), it follows from Banach-Alaoglu’s theorem that there exists
a subsequence of (fij )j∈N which is weak-star convergent to some g ∈
∗
L1 (Ln |Ω); since it also converges to f ∈ L1 (Ln |Ω ), we conclude that
∗
f = g ∈ L1 (Ln |Ω ), as asserted. 
 CHAPTER 7

Functions of Bounded Variation and Sets of Finite

Perimeter

Let Ω ⊂ Rn open. We define functions of bounded variation on Ω as

L1loc
functions on Ω whose distributional gradient is an Rn -valued Radon
measure on Ω. For that purpose, we make the following generalization
of the notion of weak derivatives introduced in 5.3:
Definition 7.1 (weak derivatives and gradients, bis). Let Ω be an
open subset of Rn and u ∈ L1loc (Ln |Ω ). We say that:
i) For 1 ≤ i ≤ n, u has weak i-th partial derivative µi ∈ Mloc (Ω, R) ≡
Cc (Ω, R)∗ if ∀ϕ ∈ C∞
c (Ω),
ˆ ˆ
∂ϕ n
u dL = − ϕ dµi .
Ω ∂xi Ω

ii) u has weak gradient µ ∈ Mloc (Ω, Rn ) ≡ Cc (Ω, Rn )∗ if ∀ϕ ∈ C∞ n

c (Ω, R ),
ˆ ˆ
n
(7.1) u div ϕ dL = − ϕ · dµ.
Ω Ω

We use the same notations for weak derivatives introduced in definition

5.3.
Remark 7.2. With the notation from the definition above, it fol-
lows from the definition of Rn -valued Radon measures 4.1 and remark
4.4 that:
1) For 1 ≤ i ≤ n, u ∈ L1loc (Ln |Ω ) admits weak i-th partial derivative if,
for each compact K ⊂ Ω, there exists CK < ∞ such that
ˆ
∂ϕ
sup{ u dLn | ϕ ∈ C∞ c (Ω), spt ϕ ⊂ K, kϕku ≤ 1} ≤ CK .
Ω ∂xi

2) u ∈ L1loc (Ln |Ω ) admits weak gradient if, for each compact K ⊂ Ω,

there exists CK < ∞ such that
ˆ
sup{ u div ϕ dLn | ϕ ∈ C∞ n
c (Ω, R ), spt ϕ ⊂ K, kϕku ≤ 1} ≤ CK .
Ω

3) Weak partial derivatives or weak gradients, if exist, are unique.

229
230 7. FUNCTIONS OF BOUNDED VARIATION

4) u ∈ L1loc (Ln |Ω ) has weak gradient µ = (µ1 , . . . , µn ) ∈ Mloc (Ω, Rn )

iff it has weak partial derivatives of first order µi ∈ Mloc (Ω, R) for
1 ≤ i ≤ n.
5) If u ∈ L1loc (Ln |Ω ) has weak i-th partial derivative vi ∈ L1loc (Ln |Ω ) in
the sense of definition 5.3, then it has weak i-th partial derivative
x
Ln vi ∈ Mloc (Ω, R) in the sense of definition 7.1. Thus, consid-
ering the injection L1loc (Ln |Ω ) ⊂ Mloc (Ω, R) given by v 7→ Ln v, x
we see that definition 5.3 may be considered a particular case of
definition 7.1.
6) It is clear that the set of functions u ∈ L1loc (Ln |Ω ) which admit weak
gradient is a linear subspace of L1loc (Ln |Ω ) and that weak derivatives
and weak gradient are linear in this subspace. We denote it by
BVloc (Ω), cf. definition 7.5 below.
Exercises 5.4 and 5.5 admit the following counterparts for the ex-
tended notion of weak derivatives.
Exercise 7.3 (weak gradients, bis). Weak gradients may be also
characterized by means of Gauss-Green identity in gradient form. That
is, let Ω be an open subset of Rn and u ∈ L1loc (Ln |Ω ); then u admits
weak gradient µ ∈ Mloc (Ω, Rn ) iff ∀ϕ ∈ C∞ (Ω),
ˆ ˆc
(7.2) u∇ϕ dLn = − ϕ dµ.
Ω Ω

Exercise 7.4. Let Ω be an open subset of Rn , u ∈ L1loc (Ln |Ω ) and

w
1 ≤ i ≤ n. If there exists µi = ∂∂xui ∈ Mloc (Ω, R), then ∀ϕ ∈ C1c (Ω),
ˆ ˆ
∂ϕ n
u dL = − ϕ dµi
Ω ∂xi Ω
Definition 7.5. Let Ω be an open subset of Rn .
i) We denote by BVloc (Ω) the set of functions u ∈ L1loc (Ln |Ω ) which
admit weak partial gradient ∇w u ∈ Mloc (Ω, Rn ). Such functions
are called of locally bounded variation on Ω.
ii) We say that u is a function of bounded variation on Ω if u ∈
L1 (Ln |Ω ) and u admits weak gradient ∇w u ∈ M(Ω, Rn ), i.e. its
weak gradient is a finite Rn -valued Radon measure on Ω. We
denote by BV(Ω) the set of functions of bounded variation on Ω.
iii) We say that E ⊂ Ω is a set of locally finite perimeter in Ω if
χE ∈ BVloc (Ω). We say that E is a Caccioppoli set or a set of
finite perimeter in Ω if χE ∈ BVloc (Ω) and ∇w χE ∈ M(Ω, Rn ).
1,1
Example 7.6. Let Ω ⊂ Rn open and f ∈ Wloc (Ω). It follows from
remark 7.2.5) that f ∈ BVloc (Ω) and its measure-weak gradient is given
x
by Ln ∇w f ∈ Mloc (Ω, Rn ).
 7. FUNCTIONS OF BOUNDED VARIATION 231

1,1
The inclusion Wloc (Ω) ⊂ BVloc (Ω) is strict; for instance, if u =
w
χ(0,∞) on Ω = R, ∇ u coincides with the Dirac measure δ0 ∈ M(R, R),
1,1
so that ∇w u ⊥ Ln , hence u ∈ BV(R) \ Wloc (R).
Theorem 7.7 (locality of the weak derivative). Let Ω ⊂ Rn open,
f ∈ L1loc (Ln |Ω ) and F ⊂ 2Ω an open cover of Ω. Then f admits weak
partial derivatives of first order on Ω iff ∀U ∈ F, f |U admits weak
partial derivatives of first order on U . Moreover, weak derivatives com-
mute with restrictions (for a Radon measure, “restriction” here means
“trace”).
Lemma 7.8. Let U ⊂ Rn open.
x
i) If µ ∈ Mloc (Ω, Rn ) and f ∈ L1 (|µ|), then µ f ∈ M(Rn , Rn ) and
x x
|µ f | = |µ| |f |.
ii) If ξ ∈ C∞ n
c (U ) and f ∈ BVloc (U ), then ξ · f (defined as 0 on R \ U )
n
belongs to BV(R ) and
∇w (ξ · f ) = Ln x(f ∇ξ) + ∇ f xξ.
w

Proof.
i) Let (ν, |µ|) be the polar decomposition
´ of µ. Then, for all ϕ ∈
x
Cc (Rn , Rn ), |µ f · ϕ| = | Ω hϕf, νi d|µ|| ≤ kϕku kf kL1 (|µ|) , thus
x
showing that µ f ∈ M(Rn , Rn ). Besides, the ´ samef computation
x
shows that, for all ϕ ∈ Cc (Rn , Rn ), µ f ·ϕ = Rn hϕ, |fν
|
x
i d(µ |f |),
x
hence the polar decomposition of µ f is ( |f fν
|
x
, µ |f |).
ii) It follows from lemma 6.14 that f ∇ξ ∈ L (Ln , Rn ). Let µ :=
1

x x
Ln (f ∇ξ) + ∇w f ξ. Then µ ∈ M(Rn , Rn ) by the previous
item; it therefore suffices to show that ξ · f admits weak gradient
equal to µ. Indeed, ∀ϕ ∈ C∞ n
c (R ),
ˆ ˆ  
(ξ · f ) · ∇ϕ dLn = f · ∇(ξ · ϕ) − ∇ξ · ϕ dLn =
Rn Ω
ˆ ˆ
= − ξϕ d ∇ f − f · ∇ξ · ϕ dLn =
w

ˆΩ Ω

=− ϕ dµ,
Rn
as asserted.

Proof of theorem 7.7. The implication “⇒” and the fact that
weak derivatives commute with restrictions are clear. We must prove
the converse implication, i.e. if ∀U ∈ F, f |U admits weak gradient
∇w (f |U ) ∈ Mloc (U, Rn ), then f admits weak gradient on Ω.
232 7. FUNCTIONS OF BOUNDED VARIATION

1) We may assume that F is locally finite and ∀U ∈ F, U b Ω. Indeed,

in the general case, take a locally finite open refinement G of F such
that ∀U ∈ G, U b Ω. For each V ∈ G, there exists U ∈ F such that
V ⊂ U ; since f |U admits weak gradient ∇w (f |U ) ∈ Mloc (U, Rn ), it
follows that f |V = (f |U )|V admits weak gradient, so that we may
replace F by G.
2) Take a smooth partition of unity (ξV )V ∈F of Ω, given by theo-
∞
rem 6.8, P such wthat ∀V ∈ F, ξV ∈ Cnc (V ). We contend that
x
µ := V ∈F ∇ (f |V ) ξV ∈ Mloc (Ω, R ). Indeed, for each com-
pact K ⊂ Ω, there are finitely
PN many V1 , . . . , VN ∈ F which intersect
w
x
K, so that µ|CKc (Ω,Rn ) = j=1 ∇ (f |Vi ) ξVi is linear continuous by
lemma 7.8.i), thus proving our contention.
3) Let ϕ ∈ C∞ n
c (Ω, R ). Since spt ϕ is compact, there are finitely many
V1 , . . . , VN ∈ F which intersect K. We then have
ˆ XN ˆ
ϕ · dµ = ξj ϕ · d ∇w (f |Vj ) =
Ω j=1 Ω

N ˆ
X
= (ξj ϕ) · d ∇w (f |Vj ) =
j=1 Vj

N ˆ
X
=− f |Vj · div (ξj ϕ) dLn =
j=1 Vj

N ˆ
X
=− f · div (ξj ϕ) dLn =
j=1 Ω
ˆ N
X
=− f · div ( ξj ϕ) dLn =
Ω j=1
ˆ
=− f · div ϕ dLn ,
Ω
thus proving that ∇w f = µ on Ω.

n
Corollary 7.9. Let Ω ⊂ R open and f : Ω → R Lebesgue mea-
surable. Then f ∈ BVloc (Ω) iff for all open V b Ω, f |V ∈ BV(V ).
Proof. The implication “⇒” is clear, in view of the fact that weak
derivatives commute with restrictions. Conversely, assume that for
all open V b Ω, f |V ∈ BV(V ). In particular, for all open V b Ω,
f |V ∈ L1 (Ln |V ), hence f ∈ L1loc (Ln |Ω ). It then follows from theorem 7.7
that ∃ ∇w f ∈ Mloc (Ω, Rn ), thus f ∈ BVloc (Ω). 
 7.1. GENERALIZED DIVERGENCE THEOREM 233

Proposition 7.10. Let Ω be an open subset of Rn . Then BV(Ω)

is a Banach space with the norm
(7.3) kf kBV(Ω) := kf kL1 (Ω) + |∇w f |(Ω).
Proof. It suffices to show that the graph of ∇w : BV(Ω) →
M(Ω, Rn ) ≡ C0 (Ω, Rn )∗ is closed in the Banach space H := L1 (Ln |Ω ) ×
M(Ω, Rn ). Indeed, let (uk , vk )k∈N be a sequence in gr ∇w such that
(uk , vk ) → (u, v) ∈ H. We must show that u is weakly differentiable
and ∇w u = v. Indeed, ∀ϕ ∈ C∞ n
c (Ω, R ), ∀k ∈ N,
ˆ ˆ
n
uk div ϕ dL = − ϕ · dvk .
Ω Ω
Since uk → u in L (L |Ω ) and vk → v in in M(Ω, Rn ) (in particular,
1 n
∗f
vk * v), the above equality holds with u in place of uk and v in place
of vk , thus proving our contention. 
Remark 7.11. If Ω is an open subset of Rn , BVloc (Ω) admits a
Fréchet space topology induced by the family of seminorms {k·kBV(V ) |
V b Ω open}.
7.1. Gauss-Green Measures and Generalized Divergence
Theorem
Definition 7.12 (Gauss-Green measure, exterior normal and perime-
ter measure of a set of locally finite perimeter). Let Ω be an open
subset of Rn and E ⊂ Ω be a set of locally finite perimeter in Ω, i.e.
such that χE ∈ BVloc (Ω) (in particular, if E is a Caccioppoli set in Ω,
cf. definition 7.5). The Rn -valued Radon measure µE := − ∇w χE ∈
Mloc (Ω, Rn ) (attention to the minus sign) is called the Gauss-Green
measure of E.
Let (νE , |µE |) be the polar decomposition of µE . We call the positive
Radon measure P(E, ·):= |µE | on Ω the perimeter measure of E and
νE the exterior normal to E.
Remark 7.13. With the notation from the previous definition, let
E be a set of locally finite perimeter in Ω and ∂ Ω E = Ω ∩ ∂E be the
topological boundary of E in Ω.
1) It is clear that spt µE ⊂ ∂ Ω E. Since νE is determined up to |µE |-
null sets, we may and do assume henceforth that νE = 0 on Ω \ ∂ Ω E
and we identify νE with a Borelian map ∂ Ω E → Rn .
2) It follows from the definition of the polar decomposition and from
exercise 7.4 that, for all ϕ ∈ C1c (Ω, Rn ),
ˆ ˆ
n
(7.4) div ϕ dL = ϕ · νE d|µE |.
E ∂ΩE
234 7. FUNCTIONS OF BOUNDED VARIATION

We call the above equality the generalized Gauss-Green theorem.

Exercise 7.14 (Complements of sets of locally finite perimeter).
Let Ω be an open subset of Rn and E ⊂ Ω be a set of locally finite
perimeter in Ω. Then Ω \ E has locally finite perimeter in Ω and
µΩ\E = −µE .
Exercise 7.15 (Sets of finite perimeter under scaling and trans-
lation). Let E be a set of locally finite perimeter in Rn , x ∈ Rn and
λ > 0. Then x + λE is a set of locally finite perimeter in Rn and
µx+λE = Φ# µE ,
where Φ : Rn → Rn is given by y 7→ x + λy. In particular, if E has
finite perimeter, so does x + λE and P (x + λE, Rn ) = λn−1 P (E, Rn ).
Proposition 7.16 (Lipschitz epigraphs have locally finite perime-
ter). Let n ≥ 2, f : Rn−1 → R Lipschitz and Ω := epiS f . Then Ω
x
is a set of locally finite perimeter in Rn , |µΩ | = Hn−1 ∂Ω and νΩ
coincides with the unit outer normal to ∂Ω in the sense of definition
6.46, i.e.
∇f (x0 ), −1


ν(x) = p
1 + k∇f (x0 )k2
on each point point x = x0 , f (x0 ) in ∂Ω = gr f whose abscissa x0 is


a differentiability point of f .
Proof. It follows from theorem 6.45 and remark 6.47 that χE ad-
mits weak gradient ∇w χE = (−ν, Hn−1 ∂E). x 
We next generalize the previous proposition to Lipschitz domains.
Lemma 7.17. Let n ≥ 2, f : Rn−1 → R Lipschitz, U 0 an open
subset of Rn , E 0 := U 0 ∩ epiS f and ν 0 : ∂ epiS f → Rn the unit outer
normal to ∂ epiS f in the sense of definition 6.46 (which, in view of the
previous proposition, coincides with the exterior normal to epiS f in the
sense of definition 7.12); see figure 1. Let Φ ∈ SE(n) be a rigid motion,
U := Φ(U 0 ), E := Φ(E 0 )
and ν := Φ∗
ν 0 , i.e. ν : ∂Φ(epiS f ) → Rn is
given by x 7→ DΦ Φ−1 (x) · ν 0 Φ−1 (x) . Then:
x

x
i) Φ# Hn−1 ∂ epiS f = Hn−1 ∂Φ(epiS f ).
x
ii) E is a set of locally finite perimeter in U , |µE | = Hn−1 ∂ U E and
its exterior normal is given by νE = ν|∂ U E .

0
x
In particular, Hn−1 ∂Φ(epiS f ) is a Radon measure. Note that
0
∂ epiS f = gr f , ∂ U E 0 = U 0 ∩ ∂ epiS f and ∂ U E = Φ(∂ U E 0 ) = U ∩
∂Φ(epiS f ).
 7.1. GENERALIZED DIVERGENCE THEOREM 235

Figure 1. Gauss-Green measure of a Lipschitz Domain

Proof.

i) Since Φ is an isometry onto Rn , Φ# Hn−1 = Hn−1 . Therefore, for

all A ⊂ Rn ,

Φ# Hn−1 x∂ epi S f (A) = Hn−1

x∂ epi S f Φ−1 (A) =

= Hn−1 ∂ epiS f ∩ Φ−1 (A) =

= Hn−1 Φ−1 ∂Φ(epiS f ) ∩ A =




= Φ# Hn−1 ∂Φ(epiS f ) ∩ A =

= Hn−1 x∂Φ(epi S f )(A).

ii)
1) Since χE 0 = χepiS f |U 0 , it follows from proposition 7.16 and from
theorem 7.7 that E 0 is a set of locally finite perimeter in U 0 and
its Gauss-Green measure µE 0 coincides with the trace µepiS f |U 0 .
Moreover, by proposition 4.36, the polar decomposition of µE 0
is (ν 0 |∂ U 0 E 0 , Hn−1 ).
236 7. FUNCTIONS OF BOUNDED VARIATION

2) Let R ∈ SO(n) be the linear part of Φ, so that DΦ(x) =cte.= R.

We have, for all ϕ ∈ C∞ n
c (U, R ):
ˆ ˆ
n AF 5.39,JΦ≡1 (∗)
div ϕ dL = (div ϕ) ◦ Φ dLn =
E E0
ˆ
7.4 and 1)
= div (R−1 ◦ ϕ ◦ Φ) dLn =
E 0 | {z }
∈C∞ 0 n
c (U ,R )
ˆ
= hR−1 ◦ ϕ ◦ Φ, ν 0 i dHn−1 =
∂U 0 E0
ˆ
= hR−1 ◦ ϕ, ν 0 ◦ Φ−1 i ◦ Φ d Hn−1 x∂ U0
E0 =

ˆ
= hR−1 ◦ ϕ, ν 0 ◦ Φ−1 i d Φ# Hn−1
| {z } |
x∂ U0
E0 =

x
{z }
R∈SO(n) by i)
= hϕ,R◦ν 0 ◦Φ−1 i = Hn−1 ∂U E
ˆ
= hϕ, ν|∂ U E i d Hn−1 x∂ U


E ,

where equality (∗) is justified by, for all x ∈ U 0 ,

chain rule
div (R−1 ◦ ϕ ◦ Φ)(x) = tr D(R−1 ◦ ϕ ◦ Φ)(x) =
= tr R−1 ◦ Dϕ(Φ · x) ◦ R =
 

= tr Dϕ(Φ · x) = (div ϕ) ◦ Φ(x).


Theorem 7.18 (Gaus-Green theorem for Lipschitz domains). Let
n ≥ 2 and Ω ⊂ Rn be a Lipschitz domain. Then Ω is a set of locally
finite perimeter in Rn and |µΩ | = Hn−1 ∂Ω. x
Proof.
1) For each x ∈ ∂Ω, there exists an open set Ux ⊂ Rn such that x ∈ Ux
and Ux is obtained by rigid motion of a cylinder centered at 0 ∈ Rn
as in definition 6.33, i.e. there exists a rigid motion Φ ∈ SE(n) with
n−1
Φ(0) = x and there exists r, h > 0 and
Γ:R → R Lipschitz

with
Γ(0) = 0 such that Ux = Φ
C(0, r, h) , Φ gr Γ∩C(0, r, h) = Ux ∩∂Ω
and Φ epiS Γ ∩ C(0, r, h) = Ux ∩ Ω.
2) From the open cover (Ux )x∈∂Ω of ∂Ω, we may extract a countable
subcover (Ui )i∈N by means of Lindelöf’s theorem. For each i ∈ N, let
the corresponding objects defined in the
previous item be denoted
with a subscript i, so that Φi C(0, ri , hi ) = Ui .
 7.1. GENERALIZED DIVERGENCE THEOREM 237
c
Let U0 := Ω and U−1 := Ω , so that (Ui )i≥−1 is a countable
open cover of Rn . We may apply corollary 6.11 to obtain a smooth
partition of unity (ξi )i≥−1 of Rn with spt ξi ⊂ Ui for i ≥ −1. Besides,
for i ≥ 1, as spt ξi ⊂ Ui b Rn , it follows that spt ξi is a compact
subset of Ui .
3) ´Claim 1: for each i ≥ −1, µi : C∞ n n
c (R , R ) → R given by ϕ 7→
Ω
div (ξi ϕ) dL is a finite R -valued Radon measure on Rn . Indeed,
n n

it is clear that µ−1 = µ0 = 0, and for i ≥ 1 and ϕ ∈ C∞ n n

c (R , R ),

ˆ ˆ
n spt ξi bUi lemma 7.17
div (ξi ϕ) dL = div (ξi ϕ) dLn =
Ω Ui ∩Ω
ˆ
= hϕξi , νi i d Hn−1 x∂ Ui


Ω ,

where νi = Φi∗ νi0 , cf. lemma 7.17. It then follows that, for all
ϕ ∈ C∞ n n
c (R , R ),

x
|µi · ϕ| ≤ Hn−1 ∂ Ui Ω (spt ξi )kϕku .

x
Since Hn−1 ∂ Ui Ω is a Radon measure on Ui and spt ξi is a compact

x
subset of Ui , we conclude that kµi kC0 (Rn ,Rn )∗ ≤ Hn−1 ∂ Ui Ω (spt ξi ) <
∞, thus proving the claim.
4) Claim 2: For each compact subset K of Rn , µK : ´C∞ n
c (K, R ) :=
∞
{ϕ ∈ Cc (R , R ) | spt ϕ ⊂ K} → R given by ϕ 7→ Ω div ϕ dLn is
n n

continuous with respect to the topology of uniform convergence.

Indeed, since K is compact and (spt ξi )i≥−1 is a locally finite
family in Rn , K intersects the members of this family for at most
finitely many indices. That is, there exists N ∈ N suchPthat K ∩
N
spt ξi = ∅ for i > N . Thus, for all ϕ ∈ C∞ n
c (K, R ), ϕ = i=−1 ξi ϕ,
PN PN
hence div ϕ = i=−1 div (ξi ϕ), which implies µK = i=−1 µi |C∞ n ,
c (K,R )
where the µi ’s were defined in the previous item. Therefore, claim
2 follows from claim 1. ´
5) It follows from claim 2 that ϕ ∈ C∞ n n
c (R , R ) 7→ Ω div ϕ dL is
n

linear continuous in the LF topology of Cc (Rn , Rn ), i.e. it is an

Rn -valued Radon measure on Rn . We have thus proved that χΩ ∈
BVloc (Rn ), i.e. Ω is a set of locally finite perimeter in Rn .
Let (νΩ , |µΩ |) be the polar decomposition of µΩ .
6) Claim 3: with the notation from claim 1, for i ≥ 1, the trace of µΩ on
x
Ui has polar decomposition (νi , Hn−1 ∂ Ui Ω). In particular, from
the uniqueness of the polar decomposition it follows that |µΩ | Ui =
x
Hn−1 ∂ Ui Ω and νΩ = νi |µΩ |-a.e. on ∂ Ui Ω.
238 7. FUNCTIONS OF BOUNDED VARIATION

Indeed, for each ϕ ∈ C∞ n

c (Ui , R ),
ˆ
spt ϕbU
µΩ · ϕ = div ϕ dLn = i
Ω
ˆ
lemma 7.17
= div ϕ dLn =
Ui
ˆ
= hϕ, νi i d Hn−1 x∂ Ui


Ω ,

x
whence the claim, since Hn−1 ∂ Ui Ω is a Radon measure on Ui and
kνi k = 1 almost everywhere on Ui with respect to Hn−1 ∂ Ui Ω. x
7) For i ≥ 1, ∂Ω ∩ Ui = ∂ Ui Ω. It then follows from claim 3 that the
x
Borel regular measure Hn−1 ∂Ω and the positive Radon measure
x
|µΩ | have the same traces on Ui , namely, Hn−1 ∂ Ui Ω. Since both
measures have support on ∂Ω, and since (Ui )i≥1 is a countable open
x
cover of ∂Ω, we conclude that |µΩ | = Hn−1 ∂Ω (since, for each
A ∈ B Rn , we may write A∩∂Ω as a countable disjoint union ∪˙ i≥1 Ai
with Ai a Borel subset of Ui for each i ≥ 1).

Corollary 7.19. Let n ≥ 2 and Ω ⊂ Rn be a Lipschitz domain.
x
Then Hn−1 ∂Ω is a Radon measure.
Remark 7.20 (outer normal to a Lipschitz domain). With the no-
tation from the proof of the previous theorem, for each i ≥ 1, the exte-
x
rior normal to Ω coincides Hn−1 ∂Ω-a.e. with νi on ∂Ω ∩ Ui = ∂ Ui Ω.
In particular, it follows from remark 6.47 that, if ∂Ω is a C1 hypersurface
on a neighborhood of p ∈ ∂Ω, we may choose νΩ on this neighborhood
as the usual outer unit normal from Differential Geometry.

7.2. Regularization of Radon measures and BV functions

Proposition 7.21. Let (φt )t>0 be the standard mollifier on Rm .
Then, for each  > 0, the convolution with φ defines a continuous
linear map φ ∗ : Cc (Rm , Rn ) → Cc (Rm , Rn ).
Proof. Given K ⊂ Rm compact, φ maps CKc (Rm , Rn ) to CKc  (Rm , Rn ),
where K := K + B(0, ) is the -neighborhood of K. Since, for all
ϕ ∈ CKc (Rm , Rn ), kφ ∗ ϕku ≤ kϕku , the linear map φ ∗ : CKc (Rm , Rn ) →
CKc  (Rm , Rn ) is bounded with respect to the norm of uniform conver-
gence. 
With the same proof, given an open subset Ω ⊂ Rm , the convolution
with φ defines a continuous linear map φ ∗ : Cc (Ω , Rn ) → Cc (Ω, Rn ),
 7.2. REGULARIZATION OF RADON MEASURES AND BV FUNCTIONS 239

where Ω is given by definition 6.17. It then follows from proposition

4.39 that (φ ∗)t : Mloc (Ω, Rn ) → Mloc (Ω , Rn ) is a well defined linear
map. We shall omit the “t” in the notation of this transpose, i.e. we
denote it with the same notation “φ ∗”.
Definition 7.22 (regularization of Rn -valued Radon measures).
Let Ω be an open subset of Rm , µ ∈ Mloc (Ω, Rn ) and (φt )t>0 the stan-
dard mollifier on Rm . We define the t-approximation or t-regularization
of µ by µt := φt ∗ µ ∈ Mloc (Ωt , Rn ).
Remark 7.23. The definition above extends definition 6.17 for
L1loc (Lm |Ω , Rn ). That is, considering the embedding L1loc (Lm |Ω , Rn ) ⊂
x
Mloc (Ω, Rn ) given by f 7→ Lm |Ω f , we have
(Lm |Ω xf ) = L

m
| x(f ) ∈ M
Ω  loc (Ω , R).

Indeed, for all f ∈ L1loc (Lm |Ω , Rn ) and all ϕ ∈ Cc (Ω , Rn ),

ˆ
m
x
(L |Ω f ) · ϕ = (φ ∗ ϕ) · f dLm
Ω
1.108.i),φˇ =φ
=
ˆ
= ϕ · (φ ∗ f ) dLm =
Ω

= Lm |Ω x(f ) · ϕ.


Proposition 7.24. With the notation from the previous definition,

let Ω ⊂ Rm open and µ ∈ Mloc (Ω, Rn ). Define µ : Ω → Rn by
ˆ

µ (x) := φ (x − y) dµ(y).
Ω
 ∞ n
Then µ ∈ C (Ω , R ) and
µ = Lm |Ω xµ . 

In particular, µ  Lm |Ω and |µ | ≤ |µ| .

Proof. 1) Let (ν, |µ|) be the polar decomposition of µ. For each
closed ball B ⊂ Ω and for each multi-index α ∈ Zm + , we have, for
all x ∈ B and all y ∈ Ω,
∂xα φ (· − y)ν(y) = ∂ α φ (x − y)ν(y) ≤ k∂ α φ ku χB+B(0,) (y).

Since χB+B(0,) ∈ L1 (|µ|), we may apply the dominated convergence

theorem to conclude that, for all x ∈ B o ,
ˆ
∃∂ µ (x) = ∂ α φ (x − y) dµ(y).
α 
240 7. FUNCTIONS OF BOUNDED VARIATION

2) For all ϕ ∈ Cc (Ω , Rn ), we have, for all x ∈ Ω and all y ∈ Ω,

φ (x − y)kϕ(x)k ≤ kφ ku kϕku χspt ϕ (x)χspt ϕ+B(0,) (y),
hence (x, y) ∈ Ω × Ω 7→ φ (x − y)ϕ(x) ∈ Rn is summable with
respect to Lm |Ω ⊗ |µ|. That justifies the application of Fubini’s
theorem in the following computation:
ˆ
ϕ(x) · µ (x) dLm (x) =
Ω
ˆ ˆ 
Fubini
= ϕ(x) φ (x − y) · ν(y) d|µ|(y) dLm (x) =
ˆΩˆ Ω
φˇ =φ

= φ (x − y)ϕ(x) dLm (x) · ν(y) d|µ|(y) =
ˆΩ Ω
= φ ∗ ϕ(y) · dµ(y) =
Ω
= µ · ϕ,

x
thus showing that µ = Lm |Ω µ , as asserted. In particular,
since kµ k ≤ |µ| (by the triangle inequality), it follows that |µ | =
x x
Lm |Ω kµ k ≤ Lm |Ω |µ| = |µ| .

Theorem 7.25 (Weak-star convergence of regularized Radon mea-
sures). Let Ω be an open subset of Rm and µ ∈ Mloc (Ω, Rn ). Then, as
 ↓ 0,
∗ ∗
µ * µ and |µ | *|µ|,
in the sense that, for all ϕ ∈ Cc (Ω, Rn ), µ · ϕ → µ · ϕ and similarly for
the total variations. Moreover, for all  > 0 and E ∈ B Ω ,
|µ |(E) ≤ |µ|(E ),
where E := E + U(0, ) is the -neighborhood of E.
Proof.
1) Let ϕ ∈ Cc (Ω, Rn ) and take 0 > 0 such that spt ϕ ⊂ Ω0 . Put
K := spt ϕ + B(0, 0 ) b Ω and ϕ := φ ∗ ϕ, where (φ )>0 is the
standard mollifier on Rm . Then, for all 0 <  < 0 , spt ϕ ⊂ K and
ϕ → ϕ uniformly, by 1.111.ii). It then follows that
µ · ϕ = µ · ϕ → µ · ϕ
∗
as  ↓ 0, thus showing that µ * µ as  ↓ 0.
 7.2. REGULARIZATION OF RADON MEASURES AND BV FUNCTIONS 241

2) For all  > 0, it follows from proposition 7.24 and remark 4.32 that
x
|µ | = Lm |Ω kµ k. On the other hand, it follows

´ from the trian-
gle inequality that, for all x ∈ Ω , kµ (x)k ≤ Ω φ (x − y) d|µ|(y).
Therefore, for all  > 0 and all E ∈ B Ω ,
ˆ
|µ |(E) = kµ (x)k dLm (x) ≤
ˆE ˆ 
Tonelli
≤ χE (x) φ (x − y) d|µ|(y) dLm (x) =
ˆΩˆ Ω

= φ (x − y)χE (x) dLm (x) d|µ|(y) ≤

Ω Ω
| {z }
≤χE (y)

≤ |µ|(E ).
3) Let V be a relatively compact open subset of Ω. Take 0 > 0 such
that V b Ω0 and (k )k∈N a sequence in (0, 0 ) with k ↓ 0.
In view of part 1), (µk |V ) is a sequence in Mloc (V, Rn ) weak-
star convergent to µ|V . Thus, for all U ⊂ V open, it follows from
proposition 4.57 that |µ| V (U ) = µ|V (U ) ≤ lim inf µk |V (U ) =
lim inf|µk | V (U ).
On the other hand, given K ⊂ V compact, in view of part 2) we
have |µk |(K) ≤ |µ|(Kk ) → |µ|(K) as k → ∞, since the sequence
of relatively compact open sets (Kk )k∈N decreases to K. Hence,
lim sup|µk | V (K) = lim sup|µk |(K) ≤ |µ|(K) = |µ| V (K). There-
fore, applying theorem 4.54, we conclude that the sequence of traces
(|µk | V )k∈N is weak-star convergent to |µ| V . Since the decreasing
sequence (k )k∈N in (0,´0 ) was arbitrarily
´ taken, we conclude that,
for all ϕ ∈ Cc (V, R), Ω ϕ d|µ | → Ω ϕ d|µ| as  → 0. Since the
relatively compact open subset V ⊂ Ω was arbitrarily taken, we
∗
conclude that |µ | *|µ| and the thesis follows.

Proposition 7.26 (regularization of BV functions). Let Ω be an
open subset of Rn , f ∈ BVloc (Ω), (φ )>0 the standard mollifier on Rn ,
f := φ ∗ f ∈ C∞ (Ω ) and (∇w f ) := φ ∗ ∇w f ∈ Mloc (Ω , Rn ). Then:
x
i) (∇w f ) = Ln |Ω ∇(f ).
ii) f → f in the sense of L1loc (Ω).
iii) For each open V b Ω,
x
∗f
x
∗f
Ln |Ω ∇(f ) |V *(∇w f )|V and Ln |Ω k∇(f )k |V *|∇w f | V
as  ↓ 0.
242 7. FUNCTIONS OF BOUNDED VARIATION

Proof. For each ϕ ∈ C∞ n

c (Ω , R ),

ˆ
n
L |Ω x∇(f ) · ϕ =
 ϕ · ∇(f ) dLn =
ˆ
prop. 1.108.i)
= − div ϕf dLn =
ˆ
prop. 1.108.j)
=− (div ϕ) f dLn =
ˆ
=− div (ϕ )f dLn =
ˆ
def. 7.22
= ϕ · d ∇w f =
ˆ
= ϕ · d(∇w f ) ,

thus showing assertion i).

Assertion ii) was already proved in 6.20.
To prove assertion iii), let V b Ω open. Take 0 > 0 such that
∗
V b Ω0 . It follows from theorem 7.25 that (∇w f ) |V *(∇w f )|V and
∗
|(∇w f ) | V *|∇w f | V . Since, by theorem 7.25, sup{|(∇w f ) |(V ) | 0 <
 < 0 } ≤ |∇w f |(V0 ) < ∞, the thesis follows from part i) and from
proposition 4.49.


7.3. First properties of BV functions

Proposition 7.27. Let Ω be an open subset of Rn and (fk )k∈N a
sequence in BVloc (Ω).
∗
i) If f ∈ BVloc (Ω) and fk → f in L1loc (Ln |Ω ), then ∇w fk * ∇w f .
ii) If f ∈ L1loc (Ln |Ω ), fk → f in L1loc (Ln |Ω ) and there exists µ ∈
∗
Mloc (Ω, Rn ) such that ∇w fk * µ, then f ∈ BVloc (Ω) and ∇w f =
µ.

Proof.
 7.3. FIRST PROPERTIES OF BV FUNCTIONS 243

i) For all ϕ ∈ C∞ n
c (Ω, R ),
ˆ ˆ
k→∞
ϕ · d ∇ fk = − div ϕfk dLn →
w

ˆ
→ − div ϕf dLn =
ˆ
= ϕ · d ∇w f,
∗
thus showing that ∇w fk * ∇w f .
ii) For all ϕ ∈ C∞ n
c (Ω, R ),
ˆ ˆ
ϕ · dµ = lim ϕ · d ∇ w fk =
k→∞
ˆ
= − lim div ϕfk dLn =
k→∞
ˆ
= − div ϕf dLn ,

hence f admits weak gradient ∇w f = µ ∈ Mloc (Ω, Rn ), i.e. f ∈

BVloc (Ω).

Corollary 7.28. Let Ω be an open subset of Rn , f ∈ L1loc (Ln |Ω )
and (fi )i∈N a sequence in BVloc (Ω) such that fi → f in L1loc (Ln |Ω ).
i) If, for each compact K ⊂ Ω, sup{|∇w fi |(K) | i ∈ N} < ∞, then
∗
f ∈ BVloc (Ω) and ∇w fi * ∇w f .
ii) If sup{|∇w fi |(Ω) | i ∈ N} < ∞, then f ∈ BVloc (Ω) with ∇w f ∈
∗f
M(Ω, Rn ) and ∇w fi * ∇w f .
Proof.
i) By corollary 4.63, there exists a subsequence (fij )j∈N of (fi )i∈N
∗
and µ ∈ Mloc (Ω, Rn ) such that ∇w fij * µ. It then follows from
proposition 7.27.ii) that f ∈ BVloc (Ω) and ∇w f = µ. Then, from
∗
7.27.i) we conclude that ∇w fi * ∇w f , as asserted.
∗
ii) By the previous item, f ∈ BVloc (Ω) and ∇w fi * ∇w f . By proposi-
∗f
tion 4.49, it then follows that ∇w f ∈ M(Ω, Rn ) and ∇w fi * ∇w f .

Proposition 7.29 (Product rule for BV, part I). Let Ω be an open
subset of Rn , f ∈ BVloc (Ω) and g : Ω → R locally Lipschitz. Then
x x
f g ∈ BVloc (Ω) and ∇w (f g) = ∇w f g + Ln f ∇w g.
Proof.
244 7. FUNCTIONS OF BOUNDED VARIATION

1 Case 1: g ∈ C∞ (Ω). Then, for all ϕ ∈ C∞c (Ω),

ˆ ˆ ˆ
∇ϕf g dL = ∇(ϕg) f dL − ϕ∇g f dLn =
n n
| {z }
∈C∞
c (Ω)
ˆ ˆ
w
=− ϕg d ∇ f − ϕf ∇g dLn ,

whence the thesis.

x
2 General case. It is clear that f g ∈ L1loc (Ln |Ω ) and that µ := ∇w f g+
x
Ln f ∇w g ∈ Mloc (Ω, Rn ). We must show that the weak gradient
of f g exists and coincides with µ. By the locality of the weak de-
rivative, cf. theorem 7.7, it suffices to prove the latter assertion for
the restriction of f g to a given V b Ω open. Let 0 > 0 such that
V b Ω0 and (φ )>0 the standard mollifier on Rn . Fix a sequence
(i )i∈N in (0, 0 ) decreasing to 0. Denoting by a subscript “” the
convolutions with φ , as usual, we have:
(a) gi := gi ∈ C∞ (Ω0 ) and, in view of theorem 6.20.iv), gi → g
uniformly on V . Hence f gi → f g in L1loc (V ).
(b) For each i ∈ N, we may apply case 1 with V in place of Ω to f gi
to conclude that f gi ∈ BVloc (V ) and ∇w (f gi ) = ∇w f gi +x
x
Ln f ∇gi ∈ Mloc (V, Rn ).
(c) For each ϕ ∈ C∞ n
c (V, R ), gi ϕ → gϕ pointwise on V and kgi ϕk ≤
sup{kgi |spt ϕ ku | i ∈ N} · kϕk ∈ L1 (|∇w f |), hence we may apply
the dominated convergence theorem to conclude that
ˆ ˆ
x
ϕ · d(∇ f gi ) → ϕ · d(∇w f g),
w
x
x ∗
i.e. ∇w f gi * ∇w
1,∞
f xg on M n
loc (V, R ).
(d) Since g ∈ Wloc (Ω), it follows from theorem 6.20.vi) that, for
all i ∈ N, ∇gi = (∇w g)i . Thus k∇gi k ≤ k∇w gkL∞ (V0 ) and,
by 6.20.iii), ∇gi → ∇w g Ln -a.e. on V . Hence, for each ϕ ∈
C∞ n w n
c (V, R ), ϕf ∇gi → ϕf ∇ g L -a.e. on V , with kϕf ∇gi k ≤
k∇w gkL∞ (V0 ) |f |kϕk ∈´ L1 (Ln |V ); therefore,
´ by the dominated
convergence theorem, ϕ·f ∇gi dL → ϕ·f ∇w g dLn , whence
n

x ∗
x
Ln f ∇gi * Ln f ∇w g on Mloc (V, Rn ).
∗
(e) From the two previous steps we conclude that ∇w (f gi ) * µ on
Mloc (V, Rn ). By proposition 7.27.ii) with V in place of Ω, it
follows that f g ∈ BVloc (V ) and ∇w (f g) = µ, as we wanted to
show.

 7.3. FIRST PROPERTIES OF BV FUNCTIONS 245

Given Ω ⊂ Rn open, it will be useful in the subsequent develop-

ments to consider the variation of a function in L1loc (Ln |Ω ), in the sense
of the definition below, even if it does not belong to BVloc (Ω).
Definition 7.30 (variation of a function in L1loc ). Let Ω ⊂ Rn open
and f ∈ L1loc (Ω). We define, for each open V ⊂ Ω,
ˆ
Var(f, V ) := sup{ f div ϕ dLn | ϕ ∈ C∞ n
c (V, R ), kϕku ≤ 1}.

Exercise 7.31 (variation of a function in L1loc ). Let Ω ⊂ Rn open

and f ∈ L1loc (Ω). Define, for each B ⊂ Ω,
Varf (B) := inf{Var(f, U ) | U open, B ⊂ U }.
Then Varf is a Borel regular measure on U which extends the variation
Var(f, ·). Moreover, f ∈ BVloc (Ω) if, and only if, Varf is a positive
Radon measure on Ω, in which case it coincides with |∇w f |.
We call Varf the variation measure of f .
Proposition 7.32 (lower semicontinuity of the variation). Let Ω ⊂
R open, (fi )i∈N a sequence in L1loc (Ln |Ω ) and f ∈ L1loc (Ln |Ω ) such that
n

fi → f in L1loc (Ln |Ω ). Then, for all V ⊂ Ω open,

Var(f, V ) ≤ lim inf Var(fi , V ).
In particular, if fi ∈ BVloc (Ω) for all i ∈ N and the second member of
the equality above is finite for each open V b Ω, then f ∈ BVloc (Ω).
Proof. For each ϕ ∈ C∞ n
c (V, R ) with kϕku ≤ 1,
ˆ ˆ
n
f div ϕ dL = lim fi div ϕ dLn ≤ lim inf Var(fi , V ),

and taking the sup on the first member yields the thesis. 
We now prove a theorem on approximation of BV functions by
smooth functions.
Theorem 7.33 (Almgren). Let Ω be an open subset of Rn and
f ∈ BV(Ω). There exists a sequence (fi )i∈N ∈ BV(Ω) ∩ C∞ (Ω) such
that fi → f in L1 (Ln |Ω ) and |∇w fi |(Ω) → |∇w f |(Ω).
Proof. 1) Fix  > 0. Choose N ∈ N sufficiently large so that,
putting
U := Ω1 ∩U(0, N ),
N
|{z}
1
={x∈Ω|d(x,Ωc )> N }
246 7. FUNCTIONS OF BOUNDED VARIATION

we have |∇w f |(Ω − U ) < . We can choose such N because |∇w f |

is a finite Radon measure and the second member above increases
to Ω as N → ∞. We now define (Ui )i∈N by
U0 := U
Ui := Ω 1 ∩ U(0, N + i), i ≥ 1.
N +i

Then (Ui )i is an increasing sequence of open relatively compact sub-

sets of Ω which increases to Ω.
Set U−1 := ∅ and define (Vi )i∈N by
Vi := Ui+1 \ Ui−1 , i ≥ 0.
Note that, for each i < j ∈ N, Vi ∩ Vj = ∅ if j − 1 ≥ i + 1, i.e.
if j ≥ i + 2; hence, each Vi meets at most 3 other Vj ’s (including
itself). Thus, (Vi )i∈N is a locally finite open cover of Ω with Vi b Ω
for each i ≥ 0. By theorem 6.8, there exists a smooth partition of
unity (ξi )i∈N of Ω such that, for all i ≥ 0, ξi ∈ C∞ c (Vi ).
2) Let (φt )t>0 be the standard mollifier on R . Note that f ξi ∈ L1 (Ln ),
n

f ∇ξi ∈ L1 (Ln , Rn ) and both functions have compact support con-

tained in spt ξi b Vi . Then, by proposition 1.108.d) and by theorem
1.111.i) we may choose, for each i ≥ 0, i > 0 sufficiently small so
that
spt φi ∗(ξi f ) b Vi
ˆ
|φi ∗(f ξi ) − f ξi | dLn < /2i+1
(7.5)
ˆ
kφi ∗(f ∇ξi ) − f ∇ξi k dLn < /2i+1 .

3) Define
∞
X
f := φi ∗(f ξi ).
i=0
Since spt φi ∗(f ξi ) b Vi for each i ≥ 0 and since (Vi )i≥0 is a locally
finite family of subsets of Ω, the sum above is locally finite, hence
f ∈ C∞ (Ω). P
Since f = ∞ i=0 f ξi , it follows from (7.5) and from the monotone
convergence theorem that
X∞
kf − f kL1 (Ln |Ω ) ≤ kφi ∗(f ξi ) − f ξi kL1 (Ln |Ω ) < ,
i=0
1 n
i.e. f → f in L (L |Ω ) as  → 0.
It remains to show that f ∈ BV(Ω) and |∇f |(Ω) → |∇w f |(Ω)
as  → 0.
 7.3. FIRST PROPERTIES OF BV FUNCTIONS 247

4) It follows from the previous step and proposition 7.32 that

|∇w f |(Ω) = Var(f, Ω) ≤ lim inf Var(f , Ω).
→0

We will then achieve the thesis once we show that lim sup→0 Var(f , Ω) ≤
|∇w f |(Ω).
For all ϕ ∈ C∞
5) P n
c (Ω, R ) with kϕk ≤ 1, we have, noting that f div ϕ =
∞
i=0 φi ∗(f ξi ) div ϕ is a finite sum (because spt ϕ is compact sub-
set of Ω and (spt ξi )i∈N is a locally finite family of subsets of Ω):
ˆ ∞ ˆ
n
X prop. 1.108.i,j)
f div ϕ dL = φi ∗(f ξi ) div ϕ dLn =
Ω i=0 Ω
∞ ˆ
X
= f ξi div (φi ∗ ϕ) dLn =
i=0 Ω
∞ ˆ ∞ ˆ P∞
X
n
X
n i=0 ∇ξi ≡0
= f div (ξi [φi ∗ ϕ]) dL − hf ∇ξi , φi ∗ ϕi dL =
i=0 Ω i=0 Ω| {z }
=hφi ∗(f ∇ξi ),ϕi
∞ ˆ
X ∞ ˆ
X
= f div (ξi [φi ∗ ϕ]) dLn − hϕ, φi ∗(f ∇ξi ) − f ∇ξi i dLn .
Ω Ω
|i=0 {z } |i=0 {z }
=:I1 =:I2

It follows from (7.5) that |I2 | < . On the other hand, since
kξi [φi ∗ ϕ]k ≤ 1 for all i ≥ 0, we have
ˆ ∞ ˆ
X
n
|I1 | = f div (ξ0 [φ0 ∗ ϕ]) dL + f div (ξi [φi ∗ ϕ]) dLn ≤
Ω i=1 Ω
∞
X
w
≤ |∇ f |(Ω) + |∇w f |(Vi ).
i=1
Note that, since Vi does not intersect Vj if j ≥ i + 2, we have
V1 ∪˙ V3 ∪˙ V5 · · · ⊂ Ω \ U
V2 ∪˙ V4 ∪˙ V6 · · · ⊂ Ω \ U,
P∞
whence i=1 |∇w f |(Vi ) ≤ 2|∇w f |(Ω \ U ) < 2 by our choice of U
in part 1). It then follows that |I1 | ≤ |∇w f |(Ω) + 2, whence
Var(f , Ω) ≤ |∇w f |(Ω) + 3.
Therefore, lim sup→0 Var(f Ω) ≤ |∇w f |(Ω), as asserted.

Remark 7.34. With the same hypothesis from theorem 7.33, if
f ∈ BV(Ω)∩L∞ (Ln |Ω ), there exists a sequence (fi )i∈N ∈ BV(Ω)∩C∞ (Ω)
248 7. FUNCTIONS OF BOUNDED VARIATION

such that fi → f in L1 (Ln |Ω ), |∇w fi |(Ω) → |∇w f |(Ω) and, for all
i ∈ N, kfi kL∞ (Ln |Ω ) ≤ 3kf kL∞ (Ln |Ω ) . That follows from the same proof
of theorem 7.33, noting that, for each  > 0 and for each x ∈ Ω, the
sum in step 3 of the proof defining f (x) has at most 3 nonzero terms
(since x belongs to at most 3 of the Vi ’s), each of which bounded by
kf kL∞ (Ln |Ω ) .
Corollary 7.35 (approximation by smooth functions). Let Ω =
Rn or Ω be a Lipschitz domain in Rn , and f ∈ BV(Ω). There ex-
ists a sequence (fi )i∈N ∈ C∞ n 1 n
c (R ) such that fi |Ω → f in L (L |Ω ) and
w w
|∇ fi |(Ω) → |∇ f |(Ω).
Proof. For each i ∈ N, by theorem 7.33 there exists gi ∈ BV(Ω) ∩
∞
C (Ω) such that kgi −f kL1 (Ln |Ω ) < 1i and |∇w gi |(Ω)−|∇w f |(Ω) < 1/i.
Since BV(Ω) ∩ C∞ (Ω) ⊂ W1,1 (Ω), we may apply corollary 6.43 (or
6.21 for Ω = Rn ) to find, for each i ∈ N, fi ∈ C∞ n
c (R ) such that
´
kgi − fi kL1 (Ln |Ω ) < 1i and Ω k∇fi k − k∇gi k dLn < 1i . It then follows

that ˆ
fi |Ω → f ∈ L1 (Ln |Ω ) and |∇w fi |(Ω) = k∇fi k dLn → |∇w f |(Ω).
Ω

Proposition 7.36 (Product rule for BV, part II). Let Ω be an open
subset of Rn . If f, g ∈ BV(Ω) ∩ L∞ (Ln |Ω ), then f g ∈ BV(Ω).
Proof. By theorem 7.33 and remark 7.34, there exist sequences
(fi )i∈N and (gi )i∈N in BV(Ω) ∩ C∞ (Ω) ∩ L∞ (Ln |Ω ) such that
• fi → f in L1 (Ln |Ω ) and gi → g in L1 (Ln |Ω );
• |∇w fi |(Ω) → |∇w f |(Ω) and |∇w gi |(Ω) → |∇w g|(Ω);
• for all i ∈ N, kfi kL∞ (Ln |Ω ) ≤ C and kgi kL∞ (Ln |Ω ) ≤ C, where
C = 3(kf kL∞ (Ln |Ω ) + kgkL∞ (Ln |Ω ) ) < ∞.
It is then clear that, for all i ∈ N, fi gi ∈ L1 (Ln |Ω ), f g ∈ L1 (Ln |Ω )
and fi gi → f g in L1 (Ln |Ω ). Moreover,
ˆ
w
|fi |k∇gi k + |gi |k∇fi k dLn ≤


lim inf|∇ (fi gi )|(Ω) ≤ lim inf
Ω
≤ C |∇ f |(Ω) + |∇w g|(Ω) < ∞.
w

It then
follows from proposition 7.32 that Var(f g, Ω) ≤ C |∇w f |(Ω)+

|∇w g|(Ω) < ∞, so that f g ∈ BV(Ω), as asserted.

Remark 7.37. With the notation from the previous proposition,
x
it is not true, in general, that ∇w (f g) = ∇w f g + ∇w g f . For x
 7.4. TRACES AND EXTENSIONS 249

instance, take Ω = Rn , E a closed subset of Rn such that χE ∈ BV(Rn )

and f = g = χE , so that f g = χ2E = χE . Then ∇w χE = ∇w χE E x
x
does not coincide with 2∇w χE E if 0 < Ln (E) < ∞.
Exercise 7.38. If p = 1 in exercise 6.25, we have (i) ⇒ (ii) ⇔ (iii).
Moreover,
• (ii) or (iii) are equivalent to f ∈ BV(Ω) and we may take
C = Var(f, Ω) = |∇w f |(Ω) in both cases;
• If Ω = Rn , for all h ∈ Rn
kτh f − f kL1 (Ln ) ≤ khk · |∇w f |(Rn ).

7.4. Traces and Extensions

Theorem 7.39 (Trace theorem for BV functions on Lipschitz epigraphs).
Let n ≥ 2, Γ : Rn−1 → R Lipschitz and Ω := epiS Γ. Then:
i) There exists a unique bounded linear operator T : BV(Ω) → L1 (Hn−1 |∂Ω )
such that, for all f ∈ BV(Ω) and all ϕ ∈ C1c (Rn , Rn ),
ˆ ˆ ˆ
n w
(7.6) f div ϕ dL = − ϕ · d ∇ f + T f ϕ · ν dHn−1 ,
Ω Ω ∂Ω
where ν the unit outer normal to ∂Ω.
ii) For all f ∈ BV(Ω) and for Hn−1 -a.e. x ∈ ∂Ω,

(7.7) lim f (y) − T f (x) dLn (y) = 0,

r→0 B(x,r)∩Ω

so that, for such x,

T f (x) = lim f dLn .

r→0 B(x,r)∩Ω

Proof.
As usual, we identify Rn ≡ Rn−1 × R and, by means of this identi-
fication, we write, for each y ∈ Rn , y = (y 0 , yn ).
1) Given f ∈ BV(Ω), suppose that there exist T f, T 0 f ∈ L1 (Hn−1 |∂Ω )
such that (7.6) holds for all ϕ ∈ C1c (Rn , Rn ). Then, for all such ϕ,
ˆ
(T f − T 0 f ) ϕ · ν dHn−1 = 0,
∂Ω

x x
hence the R -valued Radon measure (Hn−1 ∂Ω) (T f − T 0 f )ν
n

x x
is null. Then so is its total variation (Hn−1 ∂Ω) |T f − T 0 f |,
which means that T f = T 0 f Hn−1 -a.e. on ∂Ω.
In particular, if the bounded operator T satisfying (7.6) exists,
it must be unique.
250 7. FUNCTIONS OF BOUNDED VARIATION

2) We define T on C∞ n ∞ n
c (R )|Ω := {f |Ω | f ∈ Cc (R )} by T f := f |∂Ω ∈
1 n−1
Cc (∂Ω) ⊂ L (H |∂Ω ).
Since Ω is the epigraph of a Lipschitz function, for all ϕ ∈
C1c (Rn , Rn ) we may apply the Gauss-Green theorem 6.45 to f ϕ ∈
C1c (Rn , Rn ), which yields
ˆ ˆ
n
div (f ϕ) dL = f ϕ · ν dHn−1 ,
Ω ∂Ω
where ν is the outer unit normal to ∂Ω = gr Γ. Taking into account
that div (f ϕ) = ∇f · ∇ϕ + f div ϕ, we obtain (7.6) for f ∈ C∞ n
c (R ).
3) Fix  > 0 and
f ∈ C∞ n
c (R ). Let f : ∂Ω → R be defined by, for all
0 0
y = y , Γ(y ) ∈ gr Γ = ∂Ω,
f (y) := f y 0 , Γ(y 0 ) +  .

Note that f ∈ L1 (Hn−1 |∂Ω ), because f ∈ Cc (∂Ω) and Hn−1 |∂Ω is a

Radon measure on ∂Ω.
We also define:
Ω := {y = (y 0 , yn ) ∈ Rn ≡ Rn−1 × R | Γ(y 0 ) < yn < Γ(y 0 ) + },
Ω := Ω \ Ω = epiS (Γ + ).
´  ∂f
For all y = y 0 , Γ(y 0 ) ∈ ∂Ω, f (y) − T f (y) = 0 ∂x y 0 , Γ(y 0 ) +



n
t dt, so that
ˆ 
∂f
y 0 , Γ(y 0 ) + t dt ≤


|f (y) − T f (y)| ≤
∂xn
ˆ0 
∇f y 0 , Γ(y 0 ) + t dt.


≤
0
Therefore, computing by means of the area formula,
(7.8)
ˆ
|f (y) − T f (y)| dHn−1 (y)
∂Ω
ˆ ˆ   AF 5.40.2)
≤ ∇f (y + ten ) dt dHn−1 (y) =
ˆ ∂Ω
ˆ 
0
p Tonelli
∇f y 0 , Γ(y 0 ) + t dt 1 + k∇Γ(y 0 )k2 dLn−1 (y 0 ) ≤


=
Rn−1 0 | √ {z }
≤ 1+(Lip Γ)2 =:C
ˆ
≤C k∇f k dLn .
Ω

4) For an arbitrary f ∈ BV(Ω), we may apply corollary 7.35 to obtain

a sequence (fi )i∈N in C∞ n 1 n
c (R ) such that fi |Ω → f in L (L |Ω ) and
 7.4. TRACES AND EXTENSIONS 251

|∇w fi |(Ω) → |∇w f |(Ω). In particular, it follows from propositions

7.27.i), 4.49 and 4.58.ii) that
∗f ∗f
* *
∇w fi |Ω → ∇w f and ∇w fi |Ω → |∇w f |.
Fix  > 0.
For each i ∈ N, let fi : ∂Ω → R be defined by, for all
y = y 0 , Γ(y 0 ) ∈ ∂Ω,
ˆ
1 

fi y 0 , Γ(y 0 ) + t dt =


fi (y) :=
 0
ˆ
1 
= (fi )t (y) dt.
 0
Note that, for all i ∈ N, fi ∈ Cc (∂Ω) and, by (7.8) (applied to
fi ∈ C∞ n
c (R )),
ˆ ˆ ˆ 
 n−1 1 Tonelli
|T fi − fi | dH ≤ |T fi (y) − (fi )t (y)| dHn−1 (y) dt =
 ∂Ω 0
∂Ω
ˆ ˆ
1  7.8
= |T fi (y) − (fi )t (y)| dHn−1 (y) dt ≤
 0 ∂Ω
≤ C|∇w fi |(Ω ).
Hence, for all i, j ∈ N,
(7.9)
ˆ
|T fi − T fj | dHn−1 ≤
ˆ
∂Ω
ˆ ˆ
 n−1   n−1
|T fi − fi | dH + |fi − fj | dH + |T fj − fj | dHn−1 ≤
∂Ω ∂Ω
ˆ ∂Ω

≤ C |∇w fi |(Ω ) + |∇w fj |(Ω ) + |fi − fj | dHn−1 .

∂Ω

We now estimate
(7.10)
ˆ
Tonelli
|fi − fj | dHn−1 ≤
∂Ω
ˆ ˆ
1  AF 5.40.2)
≤ (fi )t − (fj )t dHn−1 dt =
 0 ∂Ω
ˆ ˆ
1 
p Tonelli
fi y 0 , Γ(y 0 ) + t − fj y 0 , Γ(y 0 ) + t 1 + k∇Γ(y 0 )k2 dy 0 dt ≤


=
 0 Rn−1 | √ {z }
2
≤ 1+(Lip Γ) =C
ˆ
C i,j→∞
≤ |fi − fj | dLn → 0,
 Ω
252 7. FUNCTIONS OF BOUNDED VARIATION

since fi → f in L1 (Ln |Ω ). On the other hand,

|∇w fi |(Ω ) ≤ |∇w fi |(Ω ∩ Ω) = |∇w fi |(Ω) − |∇w fi |(Ω ).

Since |∇w fi |(Ω) → |∇w f |(Ω) and, by proposition 4.57, |∇w f |(Ω ) ≤
lim inf|∇w fi |(Ω ), we conclude that
lim sup|∇w fi |(Ω ) = |∇w f |(Ω) − lim inf|∇w fi |(Ω ) ≤
(7.11) ≤ |∇w f |(Ω) − |∇w f |(Ω ) =
= |∇w f |(Ω ∩ Ω).
It then follows from (7.9), (7.10) and (7.11) that
ˆ
lim sup |T fi − T fj | dHn−1 ≤ 2C|∇w f |(Ω ∩ Ω).
i,j→∞ ∂Ω

Therefore, since  > 0 was arbitrarily taken, |∇w f | is a finite

Radon measure and Ω ∩ Ω decreases to ∅ as  → 0, it follows that
(T fi )i∈N is a Cauchy sequence in L1 (Hn−1 |∂Ω ), thus it is convergent
in that space. We define
T f := lim T fi ∈ L1 (Hn−1 |∂Ω ).
As it was seen in step 2 of the proof, since fi ∈ C∞ n
c (R ) for
each i ∈ N, equality (7.6) holds for fi in place of f . Thus, taking
i → ∞ and taking into account that fi → f in L1 (Ω) and T fi →
T f in L1 (Hn−1 |∂Ω ), we conclude that (7.6) also holds for f . In
particular, our definition of T f is independent of the choice of the
sequence (fi )i∈N in C∞ n 1 n
c (R ) such that fi |Ω → f in L (L |Ω ) and
w w 0
|∇ fi |(Ω) → |∇ f |(Ω). Indeed, if (fi )i∈N is another such sequence
and T 0 f := lim T fi0 in L1 (Hn−1 |∂Ω ), then both T 0 f and T f satisfy
(7.6), which implies, in view of step 1 of the proof, that T 0 f = T f
Hn−1 -a.e. on ∂Ω.
The map T : BV(Ω) → L1 (Hn−1 |∂Ω ) is therefore well-defined,
it is clearly linear and (7.6) is verified for all f ∈ BV(Ω) and all
ϕ ∈ C1c (Rn , Rn ).
5) Let (fi )i∈N be a sequence in BV(Ω) and f ∈ BV(Ω). We contend
that, if fi → f in L1 (Ln |Ω ) and |∇w fi |(Ω) → |∇w f |(Ω), then T fi →
T f in L1 (Hn−1 |∂Ω ). In particular, that proves the continuity of
T : BV(Ω) → L1 (Hn−1 |∂Ω ), thus reaching the conclusion of the proof
of part i).
Indeed, for each i ∈ N we may take a sequence (gji )j∈N in C∞ n
c (R )
such that limj→∞ gji |Ω → fi in L1 (Ln |Ω ) and limj→∞ |∇w gji |(Ω) →
|∇w fi |(Ω). By the previous step, it then follows that limj→∞ T gji =
 7.4. TRACES AND EXTENSIONS 253

T fi in L1 (Hn−1 |∂Ω ). Hence, we may take j = j(i) sufficiently large

i
in order that gi := gj(i) satisfy

1
kfi − gi kL1 (Ln |Ω ) < ,
i
1
|∇w fi |(Ω) − |∇w gi |(Ω) < and
i
1
kT fi − T gi kL1 (Hn−1 |∂Ω ) < .
i

Then (gi )i∈N is a sequence in C∞ n 1 n

c (R ) such that gi → f in L (L |Ω )
w w
and |∇ gi |(Ω) → |∇ f |(Ω); by the previous step of the proof, it
follows that T gi → T f in L1 (Hn−1 |∂Ω ). Then lim T fi = lim T gi =
T f in L1 (Hn−1 |∂Ω ), which proves our contention.
6) We now prove part ii). Fix f ∈ BV(Ω) up to the end of the proof.
For all x ∈ ∂Ω and all r > 0, we estimate

(7.12) ˆ
1

f (y) − T f (x) dLn (y) ≤
Ln B(x, r) ∩ Ω B(x,r)∩Ω
ˆ
1
f (y) − T f y 0 , Γ(y 0 ) dLn (y) +


≤ n

L B(x, r) ∩ Ω B(x,r)∩Ω
| {z }
1
ˆ
1
T f y 0 , Γ(y 0 ) − T f (x) dLn (y)


+ n

L B(x, r) ∩ Ω B(x,r)∩Ω
| {z }
2

We estimate 1 and 2 along the following steps.

7) For x0 ∈ Rn−1 and r ∈ (0, ∞], we generalize the estimate (7.8) with
the infinite closed cylinder C(x0 , r, ∞) = B(x0 , r) × R in place of
Rn−1 × R.
For  > 0, we define f , Ω and Ω as in step 3) of the proof.
Note that, assuming f Borelian (which we may assume without loss
of generality — i.e. in each equivalence class of BV(Ω) we may take
a Borelian representative, in view of corollary 1.118), f : ∂Ω → R
is clearly Borelian.
We extend the notation from step 3) to denote intersections with
the closed cylinder C(x0 , r, ∞):
254 7. FUNCTIONS OF BOUNDED VARIATION

Ω(x0 , r) := Ω ∩ C(x0 , r, ∞),

Ω(x0 , r) := Ω ∩ C(x0 , r, ∞),
Ω(x0 , r) := Ω ∩ C(x0 , r, ∞).
Note that Ω(x0 , ∞) = Ω.
Claim 1: for L1 -a.e.  > 0,
ˆ
|T f − f | dHn−1 ≤ C|∇w f | Ω(x0 , r) ,


(7.13)
∂Ω∩C(x0 ,r,∞)
p
where C = 1 + kLip Γk2 .
If f ∈ C∞ n
c (R ), the claim follows from the same argument used
in estimate (7.8), with ∂Ω ∩ C(x0 , r, ∞) in place of ∂Ω, B(x0 , r) in
place of Rn−1 and Ω(x0 , r) in place of Ω .
To prove claim 1 for f ∈ BV(Ω), we shall apply the coarea
formula with the Lipschitz function g : Rn → R given by, for all
y = (y 0 , yn ) ∈ Rn ,
g(y) = yn − Γ(y 0 ),
whose level sets are translations of gr Γ in the en direction. Note

that, for all y = (y 0 , yn ) ∈ Rn such that y 0 ∈ DΓ , ∇g(y) = −∇Γ(y 0 ), 1 ,
p p
hence Jg(y) = 1 + k∇Γ(y)k2 ≤ 1 + (Lip Γ)2 = C.
Take (fi )i∈N in C∞ n 1 n
c (R ) such that fi |Ω → f in L (L |Ω ) and
|∇w fi |(Ω) → |∇w f |(Ω). We have, for all i ∈ N,
ˆ ∞ ˆ
coarea f. 5.50
|fi − f | dHn−1 dt =
0 g −1 {t}∩Ω(x0 ,r)
ˆ
= |fi − f | Jg dLn ≤
Ω(x0 ,r)
ˆ
i→∞
≤C |fi − f | dLn → 0,
Ω(x0 ,r)

since fi → f in L1 (Ln |Ω ). Thus, passing to a subsequence, if neces-

sary, we conclude that, for L1 -a.e. t > 0,
(7.14)
ˆ ˆ
n−1 ∗ i→∞
(fi )t − ft dH = |fi − f | dHn−1 → 0,
∂Ω∩C(x0 ,r,∞) g −1 {t}∩Ω(x0 ,r)

where in equality (∗) we have used the fact that the isometry y 7→
y + ten of ∂Ω = g −1 {0} onto g −1 {t} preserves Hn−1 measure.
 7.4. TRACES AND EXTENSIONS 255

On the other hand, we may estimate, for all  > 0 and all i ∈ N,

(7.15)
ˆ ˆ
n−1
|T f − f | dH ≤ (fi ) − f dHn−1 +
∂Ω∩C(x0 ,r,∞) ∂Ω∩C(x0 ,r,∞)
| {z }
i→∞
→ 0 for a.e. >0, by (7.14)
ˆ ˆ
n−1
+ T fi − (fi ) dH + |T f − T fi | dHn−1 .
∂Ω∩C(x0 ,r,∞) 0
∂Ω∩C(x ,r,∞)
| {z } | {z }
≤C|∇w fi |(Ω(x0 ,r) ) by the case f ∈C∞ n
c (R )
i→∞
→ 0 by step 5

Besides, adapting the estimate from (7.11), we have, for all  > 0,

lim sup|∇w fi |(Ω(x, y) ) ≤ lim sup|∇w fi |(Ω(x, y) ∩ Ω) ≤

≤ |∇w f |(Ω(x, y) ∩ Ω).

But, since |∇w f | is a Radon measure on Ω, it follows from propo-

sition 4.53 that, for Ln -a.e.  > 0, |∇w f |(g −1 {}) = 0; for such
,

|∇w f |(Ω(x, y) ∩ Ω) = |∇w f |(Ω(x, y) ).

The claim therefore follows taking lim supi→∞ on both members of

(7.15).
8) Note that, for all x = x0 , Γ(x0 ) ∈ ∂Ω and all r > 0, if y = (y 0 , yn ) ∈


B(x, r) ∩ Ω, then

0 < g(y) = yn − Γ(y 0 ) = (yn − xn ) + Γ(x0 ) − Γ(y 0 ) ≤

≤ |yn − xn | + (Lip Γ)ky 0 − x0 k ≤ r(1 + Lip Γ).

That is, B(x, r) ∩ Ω ⊂ g −1 ]0, r(1 + Lip Γ)] ∩ C(x0 , r, ∞). Hence,


using the estimate from claim 1 in the previous step and the coarea
formula, we compute:
256 7. FUNCTIONS OF BOUNDED VARIATION

ˆ Jg≥1
f (y) − T f y 0 , Γ(y 0 ) dLn (y) ≤

ˆ
B(x,r)∩Ω

5.50
f (y) − T f y 0 , Γ(y 0 ) Jg(y) dLn (y) =


≤

g −1 ]0,r(1+Lip Γ)] ∩C(x0 ,r,∞)
ˆ r(1+Lip Γ) ˆ
Hn−1 invariant by isometries
f (y) − T f y 0 , Γ(y 0 ) dHn−1 (y) dt


= =
0 g −1 {t}∩C(x0 ,r,∞)
ˆ r(1+Lip Γ) ˆ claim 1
ft (y) − T f y 0 , Γ(y 0 ) dHn−1 (y) dt ≤


=
0 ∂Ω∩C(x0 ,r,∞)

≤ Cr(1 + Lip Γ)|∇w f | Ω(x0 , r)r(1+Lip Γ) ,

p
where C = 1 + (Lip Γ)2 .
On the other hand, if y = (y 0 , yn ) ∈ Ω(x0 , r)r(1+Lip Γ) , then ky 0 −
x k ≤ r and Γ(y 0 ) < yn < Γ(y 0 ) + r(1 + Lip Γ), hence
0

yn − xn = yn − Γ(x0 ) ≤
≤ r(1 + Lip Γ) + Γ(y 0 ) − Γ(x0 ) ≤ r(1 + 2 Lip Γ),
−(yn − xn ) = − yn − Γ(x0 ) ≤ − Γ(y 0 ) − Γ(x0 ) ≤ r Lip Γ,

whence |yn − xn | ≤ r(1 + 2 Lip Γ). Thus ky − xk ≤ r(2 + 2 Lip Γ), i.e.
Ω(x0 , r)r(1+Lip Γ) ⊂ B(x, r(2 + 2 Lip Γ)) ∩ Ω. We therefore conclude
that
ˆ
f (y) − T f y 0 , Γ(y 0 ) dLn (y) ≤


(7.16) B(x,r)∩Ω

≤ Cr(1 + Lip Γ)|∇w f | B(x, r(2 + 2 Lip Γ)) ∩ Ω .

Similarly, for all x = x0 , Γ(x0 ) ∈ ∂Ω and all r > 0,

ˆ Jg≥1
T f y 0 , Γ(y 0 ) − T f (x) dLn (y) ≤

ˆ
B(x,r)∩Ω

5.50
T f y 0 , Γ(y 0 ) − T f (x) Jg(y) dLn (y) =


≤

g −1 ]0,r(1+Lip Γ)] ∩C(x0 ,r,∞)
ˆ r(1+Lip Γ) ˆ
Hn−1 invariant by isometries
T f y 0 , Γ(y 0 ) − T f (x) dHn−1 (y) dt


= =
0 g −1 {t}∩C(x0 ,r,∞)
ˆ r(1+Lip Γ) ˆ
= T f (y) − T f (x) dHn−1 (y) dt =
0 ∂Ω∩C(x0 ,r,∞)
ˆ
= r(1 + Lip Γ) T f (y) − T f (x) dHn−1 (y)
∂Ω∩C(x0 ,r,∞)
 7.4. TRACES AND EXTENSIONS 257

Besides, if y = (y 0 , yn ) ∈ ∂Ω ∩ C(x0 , r, ∞), then ky 0 − x0 k ≤ r and

|yn − xn | = Γ(y 0 ) − Γ(x0 ) ≤ r Lip Γ. Thus, y ∈ B(x, r(1 + Lip Γ)),
which implies ∂Ω ∩ C(x0 , r, ∞) ⊂ B(x, r(1 + Lip Γ)) ∩ ∂Ω. It then
follows that
ˆ
T f y 0 , Γ(y 0 ) − T f (x) dLn (y) ≤

ˆ
B(x,r)∩Ω
(7.17)
≤ r(1 + Lip Γ) T f (y) − T f (x) dHn−1 (y)
B(x,r(1+Lip Γ))∩∂Ω

9) Claim 2: for Hn−1 -a.e. x ∈ ∂Ω,

|∇w f | B(x, r) ∩ Ω
(7.18) lim = 0.
r→0 rn−1
Indeed, it suffices to prove that Hn−1 (Aη ) = 0 for each η > 0,
where


 |∇w f | B(x, r) ∩ Ω
Aη := x ∈ ∂Ω | lim sup >η .
r→0 rn−1
n−1
Fix η > 0 and δ > 0; we shall estimate H10δ (Aη ). For each
x ∈ Aη and each 0 <  < δ, there exists 0 < r <  such that


|∇w f | B(x, r) ∩ Ω
(7.19) > η.
rn−1
It then follows that F := {B(x, r) | x ∈ Aη , 0 < r <  and (7.19) holds}
is a cover of Aη by nondegenerate closed balls with diameters less
than 2 < 2δ. We may therefore apply the 5-times covering lemma
2.10 to obtain a countable disjoint subfamily G ⊂ F such that
A ⊂ ∪F ⊂ ∪B∈G 5B. Hence, denoting by U the open subset of Ω
given by {x ∈ Ω | d(x, Ωc ) < }, we compute
X
n−1
H10δ (Aη ) ≤ α(n − 1)5n−1 rn−1 ≤
B=B(x,r)∈G
X α(n − 1)5n−1 w
r<
≤ |∇ f | B(x, r) ∩ Ω ≤
η
B=B(x,r)∈G

α(n − 1)5n−1 w
≤ |∇ f |(U ).
η
Since  with 0 <  < δ was arbitrarily taken, |∇w f | is a finite
Radon measure and U decreases to ∅ as  → 0, we conclude that
n−1
H10δ (Aη ) = 0, for all δ > 0. It then follows that Hn−1 (Aη ) = 0,
which concludes the proof of the claim.
 258 7. FUNCTIONS OF BOUNDED VARIATION

10) Claim 3: Let K be the cone with vertex at the origin given by
{(y 0 , yn ) ∈ Rn−1 × R | yn > (Lip Γ)ky 0 k} and, for each r > 0,
Kr := B(0, r) ∩ K. Then, for all x ∈ ∂Ω,

1 α(n) 1
(7.20)
≤ n
.
Ln B(x, r) ∩ Ω L (K1 ) L B(x, r)
n

Indeed, in view of the translation invariance of Lebesgue mea-

sure, replacing Γ by Γ(· + x0 ) − xn , we may assume x = 0. If
y = (y 0 , yn ) ∈ hyp |Γ|, i.e. if yn ≤ |Γ(y 0 )|, then

yn ≤ |Γ(y 0 )| = |Γ(y 0 ) − Γ(x0 )| ≤ (Lip Γ)ky 0 − x0 k = (Lip Γ)ky 0 k,

hence y ∈ K c . That is, hyp |Γ| ⊂ K c . Thus, for each r > 0,

Kr = B(x, r) ∩ K ⊂ B(x, r) ∩ epiS |Γ| ⊂

⊂ B(x, r) ∩ epiS Γ = B(x, r) ∩ Ω.

It then follows that

rn Ln (K1 ) = Ln (Kr ) ≤ Ln B(x, r) ∩ Ω

Ln B(x, r) ∩ Ω ≥ Ln (Kr ) =

Ln (K1 ) n
= Ln (K1 )rn =


L B(x, r) ,
α(n)

whence the claim.

11) Fix x ∈ ∂Ω such that


|∇w f | B(x, r) ∩ Ω
lim = 0 and
r→0 rn−1
(7.21)
lim T f − T f (x) dHn−1 = 0.
r→0 B(x,r)∩∂Ω

For such x, we estimate below 1 and 2 from step 6). Note that
(7.21) holds Hn−1 -a.e. on ∂Ω: the first equality holds Hn−1 -a.e. in
view of claim 2 in step 9) of the proof, and the second holds Hn−1 -a.e.
by the Lebesgue differentiation theorem 3.30 (which may be applied
because T f ∈ L1 (Hn−1 |∂Ω ) and Hn−1 |∂Ω is a Radon measure on ∂Ω).
 7.4. TRACES AND EXTENSIONS 259
p
For each r > 0, with C = 1 + (Lip Γ)2 ,
ˆ (7.16)
1
f (y) − T f y 0 , Γ(y 0 ) dLn (y) ≤


1 = n

L B(x, r) ∩ Ω B(x,r)∩Ω
Cr(1 + Lip Γ)
claim 3
≤ n
|∇w f | B(x, r(2 + 2 Lip Γ)) ∩ Ω ≤
L B(x, r) ∩ Ω
Cr(1 + Lip Γ) w

≤ |∇ f | B(x, r(2 + 2 Lip Γ)) ∩ Ω =
Ln (K1 )rn


C(1 + Lip Γ)(2 + 2 Lip Γ)n−1 |∇w f | B(x, r(2 + 2 Lip Γ)) ∩ Ω
= .
Ln (K1 ) rn−1 (2 + 2 Lip Γ)n−1
Thus, in view of (7.21), we conclude that
(7.22) lim 1 = 0.
r→0

Similarly, for each r > 0,

ˆ (7.17)
1
T f y 0 , Γ(y 0 ) − T f (x) dLn (y) ≤


2 = n

L B(x, r) ∩ Ω B(x,r)∩Ω
ˆ
r(1 + Lip Γ) claim 3
≤ n
T f (y) − T f (x) dHn−1 (y) ≤
L B(x, r) ∩ Ω B(x,r(1+Lip Γ))∩∂Ω
ˆ
(1 + Lip Γ)
≤ n T f (y) − T f (x) dHn−1 (y)
L (K1 )rn−1 B(x,r(1+Lip Γ))∩∂Ω


On the other hand, since prRn−1 B(x, r(1 + Lip Γ)) ∩ ∂Ω ⊂
B(x0 , r(1 + Lip Γ)), it follows from the area formula that

5.39
Hn−1 B(x, r(1 + Lip Γ)) ∩ ∂Ω =
ˆ p n−1 0
=
1 + ∇Γ(y 0 ) dL (y ) ≤
prRn−1 B(x,r(1+Lip Γ))∩∂Ω

≤ CLn−1 B(x0 , r(1 + Lip Γ)) = Cα(n − 1)rn−1 (1 + Lip Γ)n−1 ,

whence
Cα(n − 1)(1 + Lip Γ)n−1
1
≤
.
rn−1 Hn−1 B(x, r(1 + Lip Γ)) ∩ ∂Ω
We therefore conclude that
ˆ
C0
2 ≤ n−1
T f (y)−T f (x) dHn−1 (y).
H B(x, r(1 + Lip Γ)) ∩ ∂Ω B(x,r(1+Lip Γ))∩∂Ω
where
p
0 1 + (Lip Γ)2 α(n − 1)(1 + Lip Γ)n
(7.23) C = .
Ln (K1 )
260 7. FUNCTIONS OF BOUNDED VARIATION

That implies, in view of (7.21),

(7.24) lim 2 = 0.
r→0

Finally, from (7.12), (7.22) and (7.24), it follows that (7.7) holds
for x ∈ ∂Ω satisfying (7.21), i.e. it holds Hn−1 -a.e. on ∂Ω, which
concludes the proof.

Corollary 7.40. With the same hypothesis of theorem 7.39, if
f ∈ BV(Ω) ∩ C(Ω), then T f = f |∂Ω .
Remark 7.41. With the notation from theorem 7.39:
1) We have actually proved in step 5) of the proof that the continu-
ity of T : BV(Ω) → L1 (Hn−1 |∂Ω ) holds in a stronger sense, i.e. if
a sequence (fi )i∈N in BV(Ω) and f ∈ BV(Ω) are such that fi →
f in L1 (Ln |Ω ) and |∇w fi |(Ω) → |∇w f |(Ω), then T fi → T f in
L1 (Hn−1 |∂Ω ).
2) The trace operator from theorem 6.48 for W1,1 (Ω) is the restriction
of the trace operator from theorem 7.39.
Lemma 7.42. Let U ⊂ Rn open, Φ ∈ SE(n) a rigid motion and
U 0 = Φ(U ). If f ∈ BVloc (U 0 ), then f ◦ Φ ∈ BVloc (U ). Moreover, if
DΦ = R ∈ SO(n) and if (ν, |∇w f |) is the polar decomposition of ∇w f ,
then the polar decomposition of ∇w (f ◦ Φ) is
Φ−1 −1 w


∗ ν, Φ # |∇ f | ,

where Φ−1
∗ ν = R
−1
◦ ν ◦ Φ. In particular, f ◦ Φ ∈ BV(U ) if f ∈ BV(U 0 ).
Proof. We have, for all ϕ ∈ C∞ n
c (U, R ):
ˆ ˆ
−1
n AF 5.39,JΦ ≡1 (∗)
(f ◦ Φ) div ϕ dL = f (div ϕ) ◦ Φ−1 dLn =
U
ˆ U0
7.4
= f div (R ◦ ϕ ◦ Φ−1 ) dLn =
U0
| {z }
∈C∞ 0 n
c (U ,R )
ˆ
=− hR ◦ ϕ ◦ Φ−1 , νi d|∇w f | =
ˆU0

=− hR ◦ ϕ, ν ◦ Φi ◦ Φ−1 d|∇w f | =
ˆ
hR ◦ ϕ, ν ◦ Φi d Φ−1 # |∇w f | =


=−
| {z }
R∈SO(n)
= hϕ,R−1 ◦ν◦Φi
ˆ
hϕ, Φ−1 −1 w


= ∗ νi d Φ # |∇ f | ,
 7.4. TRACES AND EXTENSIONS 261

where equality (∗) is justified by, for all x ∈ U 0 ,

chain rule
div (R ◦ ϕ ◦ Φ−1 )(x) = tr D(R ◦ ϕ ◦ Φ−1 )(x) =
= tr R ◦ Dϕ(Φ−1 · x) ◦ R−1 =
 

= tr Dϕ(Φ−1 · x) = (div ϕ) ◦ Φ−1 (x).


Theorem 7.43 (Trace theorem for BV functions on Lipschitz do-
mains). Let n ≥ 2 and Ω ⊂ R a Lipschitz domain with ∂Ω bounded.
Then:
i) There exists a unique bounded linear operator T : BV(Ω) → L1 (Hn−1 |∂Ω )
such that, for all f ∈ BV(Ω) and all ϕ ∈ C1c (Rn , Rn ),
ˆ ˆ ˆ
n w
(7.25) f div ϕ dL = − ϕ · d ∇ f + T f ϕ · ν dHn−1 ,
Ω Ω ∂Ω
where ν the unit outer normal to ∂Ω.
ii) For all f ∈ BV(Ω) and for Hn−1 -a.e. x ∈ ∂Ω,

(7.26) lim f (y) − T f (x) dLn (y) = 0,

r→0 B(x,r)∩Ω

so that, for such x,

T f (x) = lim f dLn .

r→0 B(x,r)∩Ω

Proof. We proceed as in the proof of theorem 6.51. Fix f ∈

BV(Ω).
1) For each x ∈ ∂Ω, there exists an open set Ux ⊂ Rn such that x ∈ Ux
and Ux is obtained by rigid motion of a cylinder centered at 0 ∈ Rn
as in definition 6.33, i.e. there exists a rigid motion Φ ∈ SE(n) with
n−1
Φ(0) = x and there exists r, h > 0 and
Γ:R → R Lipschitz

with
Γ(0) = 0 such that Ux = Φ
C(0, r, h) , Φ gr Γ∩C(0, r, h) = Ux ∩∂Ω
and Φ epiS Γ ∩ C(0, r, h) = Ux ∩ Ω.
2) From the open cover (Ux )x∈∂Ω of the compact set ∂Ω ⊂ Rn , we may
extract a finite subcover (Ui )1≤i≤N . For each 1 ≤ i ≤ N , let the
corresponding objects defined in the
previous item be denoted with
a subscript i, so that Φi C(0, ri , hi ) = Ui .
c
Let U0 := Ω and U−1 := Ω , so that (Ui )−1≤i≤N is a finite open
cover of Rn . We may apply corollary 6.11 to obtain a smooth par-
tition of unity (ξi )−1≤i≤N of Rn with spt ξi ⊂ Ui for −1 ≤ i ≤ N .
Besides, for i ≥ 1, as spt ξi ⊂ Ui b Rn , it follows that spt ξi is a
compact subset of Ui .
262 7. FUNCTIONS OF BOUNDED VARIATION

Note that, in view of the product PN rule 7.29, for 0 ≤ i ≤ N ,

fi := ξi f ∈ BV(Ω). Moreover, f = i=0 fi and spt fi ⊂ spt ξi ⊂ Ui .
3) For 1 ≤ i ≤ N , it follows from lemma 7.42 that fi ◦Φi ∈ BV epiS Γi ∩
C(0, ri , hi ) and spt fi ◦ Φi ⊂ Φ−1


i (spt ξi ) b C(0, ri , hi ). Extending
the latter function by 0, we may consider fi ◦ Φi ∈ BV(epiS Γi ).
Denoting by T the trace operator given by theorem 7.39 applied
to epiS Γi , we may take T · (fi ◦ Φi ) ∈ L1 (Hn−1 |∂ epiS Γi ); moreover,
by (7.7), spt T · (fi ◦ Φi ) ⊂ spt (fi ◦ Φi ) b C(0, ri , hi ). Since the
composition with Φi induces a linear isometry of L1 (Hn−1 |∂Ω∩Ui )
onto L1 (Hn−1 |∂ epiS Γi ∩C(0,ri ,hi ) ), it makes sense to define
Ti · f := T · (fi ◦ Φi ) ◦ Φ−1 1
i ∈ L (H
n−1
|∂Ω∩Ui ) ⊂ L1 (Hn−1 |∂Ω ),
where the latter inclusion is given by the extension by 0. Note that
spt Ti f ⊂ spt ξi .
The map Ti : BV(Ω) → L1 (Hn−1 |∂Ω ) is clearly linear continuous,
since it is the composition of the sequence of linear continuous maps
described in its definition above. Actually, the continuity of Ti holds
in a stronger sense: if a sequence (fk )k∈N in BV(Ω) and f ∈ BV(Ω)
are such that fk → f in L1 (Ln |Ω ) and |∇w fk |(Ω) → |∇w f |(Ω), then
Ti fk → Ti f in L1 (Hn−1 |∂Ω ). Indeed,
k→∞
• It is clear that ξi fk → ξi f in L1 (Ln |Ω ). Moreover, in view of
∗ nc
propositions 7.27.i) and 4.58.ii), we have |∇w fk | *|∇w f | (see
∗ nc
exercise ´4.56 for the definition
´ of the narrow convergence *),
so that Ω ξi d|∇w fk | → Ω ξi d|∇w f | (because ξi |Ω ∈ Cb (Ω)).
It then follows that
ˆ ˆ
w product rule 7.29 k→∞
|∇ (ξi fk )|(Ω) = ξi d|∇ fk | + fk k∇ξi k dLn →
w

ˆ Ω
ˆ Ω
k→∞
→ ξi d|∇w f | + f k∇ξi k dLn =
Ω Ω
w
= |∇ (ξi f )|(Ω).
k→∞
• It follows from the previous item that (ξi fk ) ◦ Φi → (ξi f ) ◦ Φi
in L1 (Ln |epiS Γi ) and, since |∇w [(ξi fk ) ◦ Φi ]| = Φ−1 w
i # |∇ (ξi fk )|
by lemma 7.42,
k→∞
|∇w [(ξi fk ) ◦ Φi ]|(epiS Γi ) → |∇w [(ξi f ) ◦ Φi ]|(epiS Γi ).
We then conclude from remark 7.41.1) that
k→∞
T · [(ξi fk ) ◦ Φi ] → T · [(ξi f ) ◦ Φi ]
in L1 (Hn−1 |∂ epiS Γi ), whence Ti fk → Ti f in L1 (Hn−1 |∂Ω ), as as-
serted.
 7.4. TRACES AND EXTENSIONS 263

4) For 1 ≤ i ≤ N , with the definition of Ti f in step 3), we have (recall

that fi = ξi f ):
• For each x ∈ ∂Ω ∩ spt ξi and each r > 0 such that B(x, r) ⊂ Ui ,

fi (y) − Ti f (x) dLn (y) =

B(x,r)∩Ω

fi ◦ Φi (y) − T · (fi ◦ Φi ) Φ−1 dLn (y).

= i (x)
B(Φ−1
i (x),r)∩epiS Γi

Since Φi is a linear isometry of ∂ epiS Γi ∩C(0, ri , hi ) onto ∂Ω∩Ui

(hence it preserves Hn−1 measure), by (7.7) it follows that

(7.27) lim fi (y) − Ti f (x) dLn (y) = 0

r→0 B(x,r)∩Ω

for Hn−1 a.e. x ∈ ∂Ω ∩ spt ξi . Since the above equality holds

trivially if x ∈ ∂Ω \ spt ξi (because spt Ti f ⊂ spt ξi , as it was
noted in step 3), and because B(x, r)∩spt ξi = ∅ for sufficiently
small r > 0), we conclude that the latter equality holds for Hn−1
a.e. x ∈ ∂Ω.
• For all ϕ ∈ C1c (Rn , Rn ), denoting by ν 0 the unit outer normal
to epiS Γi and by (νi , |∇w fi |) the polar decomposition of ∇w fi ,
264 7. FUNCTIONS OF BOUNDED VARIATION

we have:

(7.28)
ˆ ˆ
n put Ri :=DΦi ∈SO(n)
fi div ϕ dL = fi ◦ Φi (div ϕ) ◦ Φi dLn =
Ω epiS Γi
ˆ
(7.7)
= fi ◦ Φi div (Ri−1 ◦ ϕ ◦ Φi ) dLn =
epiS Γi | {z }
∈C1c (Rn ,Rn )
ˆ
=− (Ri−1 ◦ ϕ ◦ Φi ) · d ∇w (fi ◦ Φi )+
epiS Γi
ˆ
7.42
+ T (fi ◦ Φi ) (Ri−1 ◦ ϕ ◦ Φi ) · ν 0 dHn−1 =
∂ epiS Γi
ˆ
hRi−1 ◦ ϕ ◦ Φi , Ri−1 ◦ νi ◦ Φi i d Φ−1 w


=− i # |∇ fi | +
epiS Γi
ˆ
+ T (fi ◦ Φi ) hϕ, Ri ◦ ν 0 ◦ Φ−1
i i ◦ Φi d H
n−1
x∂ epi S Γi

7.17
=
ˆ
=− hϕ, νi i d|∇w fi |+
Ω
ˆ
+

T (fi ◦ Φi ) ◦ Φ−1
i

hϕ, νi d Hn−1 x∂Φ (epi
i S


f) =
ˆ ˆ
w
=− ϕ · d ∇ fi + Ti f hϕ, νi dHn−1 .
Ω ∂Ω

5) We define T := N 1 n−1
P
i=1 Ti : BV(Ω) → L (H |∂Ω ). It follows from
step 3) that T is linear continuous. Besides, we have:
PN
• For each x ∈ ∂Ω and each y ∈ Ω, f (y)−T f (x) = i=0 fi (y)+
PN PN
i=1 Ti f (x) ≤ |f0 (y)| + i=1 fi (y) − Ti f (x) . Thus, for each
x ∈ ∂Ω and each r > 0,

f (y) − T f (x) dLn (y) ≤ f0 (y) dLn (y)+

B(x,r)∩Ω B(x,r)∩Ω
N
X
+ fi (y) − Ti f (x) dLn (y).
i=1 B(x,r)∩Ω
 7.4. TRACES AND EXTENSIONS 265

Since spt f0 ⊂ spt ξ0 ⊂ Ω, for each x ∈ ∂Ω and r > 0 suffi-

ciently small f0 is null on B(x, r) ∩ Ω, hence

lim f0 (y) dLn (y) = 0.

r→0 B(x,r)∩Ω

It then follows from (7.27) that (7.26) holds for Hn−1 -a.e. x ∈
∂Ω.
• Fix ϕ ∈ C∞ n n
c (R , R ).
Since ξ0 ϕ ∈ C∞ n
c (Ω, R ), we have
ˆ ˆ
w
− ξ0 ϕ · d ∇ f = f div (ξ0 ϕ) dLn =
Ω
ˆΩ ˆ
= ξ0 f div ϕ dL + f ∇ξ0 · ϕ dLn .
n
Ω Ω

Thus

(7.29)
ˆ ˆ

Ω
f0 div ϕ dL = − ϕ · d ∇w f
n
Ω
xξ + L xf ∇ξ

0
n
0
product rule 7.29
=
ˆ
=− ϕ · d ∇ w f0 .
Ω

Therefore, from (7.28) and (7.29),

ˆ ˆ N
X
n n
f div ϕ dL = f0 div ϕ dL + fi div ϕ dLn =
Ω Ω i=1
ˆ N ˆ
X
w
=− ϕ · d ∇ f0 − ϕ · d ∇ w fi +
Ω i=1 Ω
N ˆ
X
+ Ti f ϕ · ν dHn−1 =
∂Ω
ˆi=1 ˆ
w
=− ϕ · d∇ f + T f ϕ · ν dHn−1 ,
Ω ∂Ω

thus (7.25) is verified.

6) We have thus proved the existence of a continuous linear map T :
BV(Ω) → L1 (Hn−1 |∂Ω ) satisfying (7.26) and (7.25). It remains to
prove the uniqueness stated in part i), for which we reapply the
argument used in the proof of the same statement for epigraphs:
given f ∈ BV(Ω), suppose that there exist T f, T 0 f ∈ L1 (Hn−1 |∂Ω )
266 7. FUNCTIONS OF BOUNDED VARIATION

such that (7.25) holds for all ϕ ∈ C1c (Rn , Rn ). Then, for all such ϕ,
ˆ
(T f − T 0 f ) ϕ · ν dHn−1 = 0,
∂Ω
x x
hence the R -valued Radon measure (Hn−1 ∂Ω) (T f − T 0 f )ν
n

x x
is null. Then so is its total variation (Hn−1 ∂Ω) |T f − T 0 f |,
which means that T f = T 0 f Hn−1 -a.e. on ∂Ω.

Corollary 7.44. With the same hypothesis of theorem 7.43, if
f ∈ BV(Ω) ∩ C(Ω), then T f = f |∂Ω .
Remark 7.45. With the notation from theorem 7.43:
1) We have actually proved that the continuity of T : BV(Ω) → L1 (Hn−1 |∂Ω )
holds in a stronger sense, i.e. if a sequence (fi )i∈N in BV(Ω) and
f ∈ BV(Ω) are such that fi → f in L1 (Ln |Ω ) and |∇w fi |(Ω) →
|∇w f |(Ω), then T fi → T f in L1 (Hn−1 |∂Ω ). Indeed, that was proved
in step 3) of the proof for each Ti : BV(Ω) → L1 (Hn−1 |∂Ω ), for
1 ≤ i ≤ N , hence it also holds for T = N
P
i=1 Ti .
2) The trace operator from theorem 6.51 for W1,1 (Ω) is the restriction
of the trace operator from theorem 7.43.
Theorem 7.46 (Extension of BV functions on Lipschitz epigraphs
or Lipschitz domains). Let n ≥ 2 and Ω an open subset of Rn which
is a Lipschitz epigraph or a Lipschitz domain with ∂Ω bounded. Given
f ∈ BV(Ω) and g ∈ BV(Rn \ Ω), let F be Ln -measurable function
defined by (
f (x) x∈Ω
F (x) :=
g(x) x ∈ Rn \ Ω.
Then F ∈ BV(Rn ) and
(7.30) ∇w F = i# ∇w f + i# ∇w g − Hn−1 x∂Ω x(T f − T g)ν,
where i# ∇w f and i# ∇w g are the pushforwards of ∇w f ∈ M(Ω, Rn )
c
and ∇w g ∈ M(Ω , Rn ) by the respective inclusions (the pushforward
is taken in the sense of remark 4.46), ν is the unit outer normal of Ω
c
and T denotes both trace operators BV(Ω), BV(Ω ) → L1 (Hn−1 |∂Ω ). In
particular,
|∇w F | = i# |∇w f | + i# |∇w g| + Hn−1 x∂Ω x|T f − T g|.
Note that, since Ln (∂Ω) = 0 (because, as we have already seen,
H-dim ∂Ω = n − 1), F is indeed an almost everywhere defined Ln -
measurable function. Besides, if Ω is a Lipschitz epigraph or a Lipschitz
c
domain with bounded frontier, so is Ω , so that the trace operator
 7.4. TRACES AND EXTENSIONS 267
c
BV(Ω ) → L1 (Hn−1 |∂Ω ) exists. By exercise 7.14 and by the fact that
c
χΩc = χΩc Ln -a.e., the Gauss-Green measure of Ω is −µΩ .
Proof.
1) For all ϕ ∈ C∞ n n
c (R , R ) with kϕku ≤ 1,
(7.31)
ˆ ˆ ˆ
n n (7.6) or (7.25)
F div ϕ dL = f div ϕ dL + g div ϕ dLn =
c
Rn Ω
ˆ ˆΩ
= − ϕ · d ∇w f − ϕ · d ∇w g+
c
ˆΩ Ω

+ (T f − T g)ϕ · ν dHn−1 .
∂Ω
Therefore,
ˆ
n w w c
Var(F, R ) ≤ |∇ f |(Ω) + |∇ g|)(Ω ) + |T f − T g| dHn−1 < ∞,
∂Ω
n
which implies F ∈ BV(R ), as asserted.
2) It remains to prove the formulas for ∇w F and |∇w F |. Since Rn =
c
Ω ∪˙ ∂Ω ∪˙ Ω , we have
∇w F = ∇w F xΩ + ∇ F xΩ + ∇ F x∂Ω.
w c w

We must compute the three measures appearing in the second mem-

c
ber above. Since Ω and Ω are open sets, by the locality of the weak
gradient it is clear that ∇w F |Ω = ∇w f and ∇w F |Ωc = ∇w g, hence
∇w F xΩ = i #∇
w
f and ∇w F xΩ c
= i# ∇w g.
We contend that
∇w F x∂Ω = −H x∂Ω x(T f − T g)ν.
n−1

Indeed, fix  > 0 and let Ω := ∂Ω + U(0, ) be the open

-neighborhood of ∂Ω. By exercise 6.12 (differentiable Urysohn
lemma) We may take ζ ∈ C∞ (Rn ) such that 0 ≤ ζ ≤ 1, ζ ≡ 1 on
∂Ω and ζ ≡ 0 on Ωc ; in particular, spt ζ ⊂ Ω ⊂ ∂Ω + B(0, ).
For each ϕ ∈ C∞ n n
c (R , R ), applying (7.31) with ϕζ in place of
ϕ yields
ˆ ˆ ˆ
w w
− ϕζ · d ∇ F = − ϕζ · d ∇ f − ϕζ · d ∇w g+
c
ˆΩ
Rn Ω

+ (T f − T g)ϕ · ν dHn−1 .
∂Ω
As  → 0, ϕζ converges pointwise to ϕχ∂Ω , and kϕζk ≤ kϕk ∈
L1 (|∇w F |) ∩ L1 (|∇w f |) ∩ L1 (|∇w g|). We may therefore apply the
268 7. FUNCTIONS OF BOUNDED VARIATION

dominated convergence theorem along a sequence convergent to 0,

which yields
ˆ ˆ
w
− ϕ · d∇ F = (T f − T g)ϕ · ν dHn−1 ,
∂Ω ∂Ω

thus proving our contention.

Finally, since ∇w F x Ω, ∇w Fx c
x
Ω and ∇w F ∂Ω are pair-
wise mutually singular, it follows from proposition 4.15 that
F xΩ| + |∇ F xΩ | + |∇ F x∂Ω|
c
|∇w F | = |∇w w w

which yields the stated formula for |∇w F |.


Corollary 7.47 (Extension of BV functions on Lipschitz epigraphs
or Lipschitz domains). Let n ≥ 2 and Ω an open subset of Rn which
is a Lipschitz epigraph or a Lipschitz domain with ∂Ω bounded. The
extension by 0 defines a bounded linear operator BV(Ω) → BV(Rn ).
Proof. For each f ∈ BV(Ω), its extension by 0 f¯ : Rn → R
coincides Ln -a.e. with F defined in the previous theorem by means of
f and g ≡ 0, hence f¯ ∈ BV(Rn ). Moreover, it follows from (7.30) and
from the continuity of the trace operator that kf¯kBV(Rn ) = kf¯kL1 (Ln ) +
|∇w f¯|(Rn ) = kf kL1 (Ln |Ω ) + |∇w f |(Ω) + kT f kL1 (Hn−1 |∂Ω ) ≤ Ckf kBV(Ω) .

Corollary 7.48. Let n ≥ 2 and Ω an open subset of Rn which
is a Lipschitz epigraph or a Lipschitz domain with ∂Ω bounded. Given
f ∈ W1,1 (Ω) and g ∈ W1,1 (Rn \ Ω) such that T f = T g, then F defined
in theorem 7.46 belongs to W1,1 (Rn ).
Proof. We have F ∈ BV(Rn ) and, by (7.30), ∇w F = i# ∇w f +
x x
i# ∇w g = Ln ∇w f + Ln ∇w g ∈ L1 (Ln , Rn ). 

7.5. Compactness
Theorem 7.49 (Compactness theorem for BV). Let Ω ⊂ Rn be a
bounded Lipschitz domain and (fi )i∈N a sequence in BV(Ω) such that
sup{kfi kBV(Ω) | i ∈ N} < ∞.
Then there exists f ∈ BV(Ω) and a subsequence (fij )j∈N of (fi )i such
that fij → f in L1 (Ln |Ω).
We present two proofs for this theorem.
7.6. SETS OF FINITE PERIMETER AND EXISTENCE OF MINIMAL SURFACES
269

Proof 1. For each i ∈ N, we may apply theorem´7.33 to obtain

gi ∈ C∞ (Ω) ∩ BV(Ω) such that kfi − gi kL1 (Ω) ≤ 1/i and Ω k∇gi k dLn ≤
|∇w fi |(Ω) + 1/i. In particular,
ˆ
sup{ k∇gi k dLn | i ∈ N} < ∞.
Ω

It then follows that (gi )i∈N is a bounded sequence in W1,1 (Ω) ⊂ BV(Ω).
We may therefore apply Rellich-Kondrachov’s theorem 6.77 to obtain
a subsequence (gij )j∈N of (gi )i and f ∈ L1 (Ω) 1 such that gij → f in
L1 (Ω). Thus, fij → f in L1 (Ω). Moreover, it follows from proposition
7.32 that
Var(f, Ω) ≤ lim inf Var(fij , Ω) ≤ sup{kfi kBV(Ω) | i ∈ N} < ∞,
| {z }
=|∇w fij |(Ω)

whence f ∈ BV(Ω). 
Proof 2. By means of the extension by 0, cf. corollary 7.47, we
may assume that (fi )i∈N a sequence in BV(Rn ) and spt fi ⊂ Ω b Rn .
It is clear that (fi )i∈N is bounded in L1 (Ln ), since it is bounded in
BV(Ω). Moreover, it follows from exercise 7.38 that
kτh f − f kL1 (Ln ) ≤ khk · sup{|∇w fi |(Rn ) | i ∈ N},
| {z }
<∞

so that limh→0 kτh fi − fi kL1 (Ln ) = 0 uniformly on i ∈ N. The thesis

then follows from the Kolmogorov-Riesz-Fréchet compactness criterion
1.80. 

7.6. Sets of Finite Perimeter and Existence of Minimal

Surfaces
In this section we develop some basic properties of sets of finite
perimeter and we apply the direct method of the Calculus of Vari-
ations to prove the existence of minimizers in some geometric varia-
tional problems. Recall the definitions and notations for sets of finite
perimeter in 7.5 and 7.12.
7.6.1. Support of the Gauss-Green measure. Let Ω be an
open set in Rn , E ⊂ Ω a set of locally finite perimeter in Ω and
µE ∈ Mloc (Ω, Rn ) its Gauss-Green measure. As we have already noted
in 7.13.1), it is clear that spt µE ⊂ ∂ Ω E. Actually, we have the follow-
ing precise description of spt µE . We use |·| to denote the Lebesgue
measure in Rn .
∗
1actually f ∈ L1 (Ω), by corollary 6.80
270 7. FUNCTIONS OF BOUNDED VARIATION

Proposition 7.50. If E ⊂ Ω is a set of locally finite perimeter in

the open subset Ω of Rn , then
spt µE = {x ∈ Ω | ∀r > 0, 0 < |E ∩ U(x, r)| < α(n)rn } ⊂ ∂ Ω E.
Moreover, there exists a Borel set F ⊂ Ω in the same L1loc class of E
such that µF = ∂ Ω F .
Proof.
1) Let x ∈ Ω. If there exists r > 0 such that |E ∩ U(x, r)| = 0 (re-
spectively, such that |E ∩ U(x, r)| = α(n)rn ), then χE = 0 (respec-
tively, χE = 1) Ln -a.e. on the open set Ω ∩ U(x, r), which implies
∇w χE = 0 on Ω ∩ U(x, r) by the locality if the weak derivative 7.7,
hence Ω ∩ U(x, r) ⊂ Ω \ spt µE .
Conversely, if x ∈ Ω \ spt µE , there exists r > 0 such that
U(x, r) ⊂ Ω and ∇w χE = 0 on U(x, r). It then follows from propo-
sition 5.7 that χE coincides Ln -a.e. with a constant function on
U(x, r), hence χE = 0 a.e. on U(x, r) or χE = 1 a.e. on U(x, r),
which implies |E ∩ U(x, r)| = 0 or |E ∩ U(x, r)| = α(n)rn , respec-
tively.
We have thus proved that x ∈ Ω \ spt µE if, and only if, there
exists r > 0 such that |E ∩ U(x, r)| = 0 or |E ∩ U(x, r)| = α(n)rn .
2) Up to modifying E on a Ln -null set, we may assume that E ∈ B Ω .
Define:
A0 := {x ∈ Ω | ∃r > 0, |E ∩ U(x, r)| = 0},
A1 := {x ∈ Ω | ∃r > 0, |E ∩ U(x, r)| = α(n)rn } =
= {x ∈ Ω | ∃r > 0, |(Ω \ E) ∩ U(x, r)| = 0}.
Then A0 and A1 are disjoint open subsets of Ω with |E ∩ A0 | = 0
and |A1 \ E| = 0. Define F := (E ∪ A1 ) \ A0 ∈ B Ω . Then:
• E \ F ⊂ E ∩ A0 and F \ E ⊂ A1 \ E, so that |E 4 F | = 0.
• It follows from the previous item that µF = µE , hence ∂ Ω F ⊃
spt µF = spt µE = Ω \ (A0 ∪ A1 ) by part 1) of the proof.
Ω
• Since A1 ⊂ F o and F ⊂ Ω \ A0 , we conclude that ∂ Ω F ⊂
Ω \ (A0 ∪ A1 ), whence the thesis.

7.6.2. Operations with Sets of Finite Perimeter, part I.
Proposition 7.51. Let Ω be an open subset of Rn . If E, F are
sets of (locally) finite perimeter in Ω, then so are E ∪ F and E ∩ F .
Moreover,
(7.32) |µE∪F | + |µE∩F | ≤ |µE | + |µF |.
 7.6. SETS OF FINITE PERIMETER AND EXISTENCE OF MINIMAL SURFACES
271

Proof. It follows from proposition 7.36 and of the locality of the

weak derivative that both χE∩F = χE χF and χE∪F = χE + χF − χE χF
belong to BVloc (Ω).
In order to prove (7.32), it suffices to show that the inequality holds
when both members are computed in each open A b Ω. Fix such an
open A b Ω, let (φ )>0 be the standard mollifier in Rn and take 0 > 0
such that A b Ω0 .
Define, for 0 <  < 0 , f := φ ∗ χE ∈ C∞ (Ω0 ) and g := φ ∗ χF ∈
C∞ (Ω0 ), so that 0 ≤ f , g ≤ 1, f g → χE∩F in L1loc (Ω0 ) and h :=
f + g − f g → χE∪F in L1loc (Ω0 ). Then:
1) It follows from proposition 7.26 that, for all open V b Ω0 ,
∗ ∗
|∇w f | V
*|µE | V
and |∇w g | V
*|µF | V .
In particular, taking an open set V such that A b V b Ω0 , we
conclude that
4.54.ii)
lim sup|∇w f |(A) ≤ lim sup|∇w f |(A) ≤ |µE |(A),
and, similarly, lim sup|∇w g |(A) ≤ |µF |(A).
2) For 0 <  < 0 ,
ˆ
w
f kg k + g kf k dLn ,


|∇ (f g )|(A) ≤
ˆA
|∇w h |(A) ≤ (1 − g )kf k + (1 − f )kg k dLn ,


A
hence
|∇w (f g )|(A) + |∇w h |(A) ≤ |∇w f |(A) + |∇w g |(A).
3) Taking the lim inf of both members in the previous equality along
the sequence  = 1/k, it follows from step 1) and from the lower
semicontinuity of the variation 7.32 that
|µE∪F |(A) + |µE∩F |(A) ≤ |µE |(A) + |µF |(A).
The previous inequality holds for each open A b Ω. In particular,
given such an open A b Ω, it may be applied to Ak := {x ∈ A |
d(x, Ac ) > k1 }, for each k ∈ N, which yields
|µE∪F |(Ak ) + |µE∩F |(Ak ) ≤ |µE |(Ak ) + |µF |(Ak ) ≤ |µE |(A) + |µF |(A).
Since the sequence (Ak )k∈N increases to A, taking limk→∞ in the first
member of the previous inequality allows us to conclude that
|µE∪F |(A) + |µE∩F |(A) ≤ |µE |(A) + |µF |(A),
which proves (7.32), whence the thesis. 
272 7. FUNCTIONS OF BOUNDED VARIATION

7.6.3. Compactness from perimeter bounds.

Definition 7.52. Let (Ei )i∈N be a sequence of Lebesgue measur-
able sets in Rn and E a Lebesgue measurable set in Rn . We say that
Ei * E
if kχEi − χE kL1 (Ln ) = |Ei 4 E| → 0.
loc
We say that Ei * E if χEi → χE in L1loc (Ln ).
Theorem 7.53 (Compactness from perimeter bounds). Let R > 0
and (Ei )i∈N be a sequence of sets of finite perimeter in Rn such that
sup P(Ei ) < ∞,
i∈N
Ei ⊂ U(0, R) ∀i ∈ N.
Then there exists a set E ⊂ U(0, R) of finite perimeter in Rn and
a subsequence (Eij )j∈N of (Ei )i∈N such that
∗
Eij * E and µEij * µE .

Proof. Let Ω = U(0, R), which is a bounded Lipschitz domain.

Note that, given f ∈ L1 (Ln |Ω ) ⊂ L1 (Ln ), it follows from corollary 7.47
that f ∈ BV(Ω) if, and only if, its extension by 0 belongs to BV(Rn ).
The hypothesis implies that (χEi )i∈N is a bounded sequence in
BV(Ω) since, for all i ∈ N, kχEi kL1 (Ln |Ω ) ≤ α(n)Rn and |∇w (χEi |Ω )|(Ω) =
|∇w χEi | Ω (Ω) ≤ P(Ei ) ≤ supi∈N P(Ei ) < ∞. It then follows from
theorem 7.49 that there exists a subsequence (Eij )j∈N of (Ei )i∈N and
f ∈ BV(Ω) such that χEij → f in L1 (Ln |Ω ). Since there exists a sub-
sequence of χEij which converges Ln -a.e. to f on Ω, we conclude that
there exists E ∈ B Ω such that f = χE Ln -a.e. on Ω, hence χE ∈ BV(Ω)
and Eij * E. By the remark on the first paragraph of the proof, we
have χE ∈ BV(Rn ), i.e. E is a set of finite perimeter in Rn . Finally,
∗
it follows from proposition 7.27.i) that µEij * µE , which completes the
proof. 
Lemma 7.54. Let Ω ⊂ Rn be a bounded Lipschitz domain and E ⊂
Rn be a set of locally finite perimeter. Then E ∩ Ω is a set of finite
perimeter in Rn and
P(E ∩ Ω) ≤ P(E, Ω) + P(Ω).
Proof. We know from proposition 7.51 (with Rn in place of Ω
and Ω in place of F ) that E ∩ Ω is a set of locally finite perimeter
in Rn . It then suffices to prove the asserted inequality, which implies
 7.6. SETS OF FINITE PERIMETER AND EXISTENCE OF MINIMAL SURFACES
273

P(E ∩ Ω) < ∞, since P(E, Ω) < ∞ (because E is a set of locally finite

perimeter in Rn and Ω b Rn ) and P(Ω) < ∞ (because ∂Ω is bounded).
Consider F in theorem 7.46 given by f = χE |Ω and g = 0. As
elements of L1loc (Ln ), we have F = χE∩Ω ; it then follows from theorem
7.46 that ∇w F = ∇w χE∩Ω is given by (7.30). In particular, since
|T f | ≤ 1 by (7.26), it follows that
ˆ
w n w
P(E ∩ Ω) = |∇ F |(R ) ≤ |∇ f |(Ω) + |T f | dHn−1 =
ˆ ∂Ω

= |∇w χE | Ω (Ω) + |T f | dHn−1 ≤

| {z } ∂Ω
=|µE |(Ω)=P(E,Ω)

≤ P(E, Ω) + Hn−1 (∂Ω) = P(E, Ω) + P(Ω),

as asserted. 
Corollary 7.55 (Compactness from perimeter bounds). Let (Ei )i∈N
be a sequence of sets of locally finite perimeter in Rn such that, for all
R > 0,

sup P Ei , U(0, R) < ∞.
i∈N
Then there exists a set E of locally finite perimeter in Rn and a
subsequence (Eij )j∈N of (Ei )i∈N such that
loc ∗
Eij * E and µEij * µE .
Proof. For each N ∈ N, it follows from lemma 7.54 that, for all
i ∈ N, Ei ∩ U(0, N ) is a set of finite perimeter in Rn and




sup P Ei ∩ U(0, N ) ≤ sup P Ei , U(0, N ) + P U(0, N ) < ∞.
i∈N i∈N

We may therefore apply theorem 7.53 to obtain a subsequence

(Ej1 )j∈N of (Ei )i∈N and for each k ≥ 2 a subsequence (Ejk )j∈N of (Ejk−1 )j∈N
such that for all k ∈ N, χEjk ∩U(0,k) converges in L1 (Ln ) to a set of finite
perimeter Ek ⊂ U(0, k) of Rn . The diagonal (Ekk )k∈N is therefore a sub-
sequence of (Ei )i∈N such that χEkk ∩U(0,N ) is L1 (Ln ) convergent for each
N ∈ N. That is, χEkk is a convergent sequence in L1loc (Ln ) and its limit
n
is the characteristic
function of a Borel measurable set F ⊂ R such
that | F ∩ U(0, k) 4 Ek | = 0 for each k ∈ N, i.e. F ∩ U(0, k) is a set

of finite perimeter in Rn for each k ∈ N, hence χF |U(0,k) ∈ BV U(0, k)
for each k ∈ N. We have thus proved that χF ∈ BVloc (Rn ), i.e. F is a
loc
set of locally finite perimeter in Rn , and Ekk * F . It then follows from
∗
proposition 7.27.i) that µEkk * µF , which completes the proof. 
274 7. FUNCTIONS OF BOUNDED VARIATION

7.6.4. Existence of Minimizers. In this subsection we apply

the direct method of the Calculus of Variations to prove the existence
of minimizers of two classes of geometric variational problems. Such
application rests on the compactness theorems 7.53, 7.55 and on the
lower semicontinuity of the perimeter 7.32.
Firstly we consider the Plateau problem in a compact subset K of
n
R with boundary data given by a set M of locally finite perimeter in
Rn . The problem consists in finding a set E0 ⊂ Rn of locally finite
perimeter which has least perimeter in K among the sets E ⊂ Rn with
locally finite perimeter whose boundaries are “fixed” by M , in the sense
that E \ K = M \ K — see figure 2.

Figure 2. Plateau problem in K with boundary data M

Proposition 7.56 (Minimizers for the Plateau problem in K with

boundary data M ). Let K ⊂ Rn be a compact set and M be a set of
locally finite perimeter in Rn . Then there exists E0 ⊂ Rn of locally
finite perimeter which minimizes the functional
E 7→ P(E, K)
in the class E := {E ⊂ Rn | χE ∈ BVloc (Rn ) and E \ K = M \ K}.
Proof. Note that E = 6 ∅, since M ∈ E. Let m := inf{P(E, K) |
E ∈ E} (hence 0 ≤ m < ∞), and (Ei )i∈N a sequence in E such that
P(Ei , K) → m. Take R > 0 such that Ω := U(0, R) ⊃ K.
For all i ∈ N, we have
P(Ei , Ω) = P(Ei , Ω \ K) + P(Ei , K) =
= P(M, Ω \ K) + P(Ei , K) ≤ C(Ω).
7.6. SETS OF FINITE PERIMETER AND EXISTENCE OF MINIMAL SURFACES
275

It then follows from corollary 7.55 that there exists a set of locally
finite perimeter E0 ⊂ Rn and a subsequence (Eij )j∈N of (Ei )i such
loc
that Eij * E0 . Modifying E0 on a Ln -null set, if necessary, me may
assume that E0 \ K = M \ K, so that E0 ∈ E. Besides, by the
lower semicontinuity of the variation 7.32 it follows that P(E0 , Ω) ≤
lim inf P(Eij , Ω), that is
P(M, Ω \ K) + P(E0 , K) ≤


≤ lim inf P(M, Ω \ K) + P(Eij , K) =
= P(M, Ω \ K) + lim inf P(Eij , K),
whence P(E0 , K) ≤ lim inf P(Eij , K) = lim P(Ei , K) = m. Since E0 ∈
E, we also have the opposite inequality m ≤ P(E0 , K), hence m =
P(E0 , K). 

Figure 3. Relative isoperimetric problem in Ω

Given an open set Ω ⊂ Rn , the relative isoperimetric problem in Ω

is the problem of finding sets with least perimeter in Ω with a fixed
prescribed volume — see figure 3. Precisely, given m ∈ (0, |Ω|) (note
that |Ω| is not assumed to be finite), we want to decide whether the
following infimum is realized by a set of finite perimeter in Ω:
α(m, Ω) := inf{P(E, Ω) | E ⊂ Ω, χE ∈ BV(Ω), |E| = m}.
We say that a set E ⊂ Ω of finite perimeter in Ω is a relative isoperi-
metric set in Ω if if is normalized according to proposition 7.50 so
that spt µE = ∂ Ω E and it is a minimizer of the above problem, i.e. if
P(E, Ω) = α(|E|, Ω). If Ω is a bounded Lipschitz domain, the existence
of such minimizers may be proved once more by a direct application of
the direct method of the Calculus of Variations:
Proposition 7.57 (Existence of relative isoperimetric sets on bounded
Lipschitz domains). Let Ω be a bounded Lipschitz domain and m ∈
(0, |Ω|]. Then there exists a set E ⊂ Ω such that χE ∈ BV(Ω), |E| = m
and P(E, Ω) = α(m, Ω).
276 7. FUNCTIONS OF BOUNDED VARIATION

Proof. Let E := {P(E, Ω) | E ⊂ Ω, χE ∈ BV(Ω), |E| = m}.

1) We contend that E is not empty. Indeed, for each t ∈ Rn , define
Ωt := Ω ∩ {x ∈ Rn | x1 < t}, so that χΩt ∈ BV(Ω). By a direct
application of the dominated convergence theorem, t ∈ R 7→ |Ωt | ∈
R is a continuous function which is null in t0 such that Ωt0 = ∅
and |Ω| in t1 such that Ωt1 = Ω (such t0 and t1 exist because Ω
is bounded). Therefore, by the intermediate value theorem, there
exists t ∈ [t0 , t1 ] such that |Ωt | = m, hence Ωt ∈ E.
2) It follows from the previous item that 0 ≤ α(m, Ω) < ∞. Let
(Ei )i∈N be a sequence in E such that P(Ei , Ω) → α(m, Ω). Lemma
7.54 ensures that, for all i ∈ N, Ei is a set of finite perimeter in Rn
and
P(Ei ) ≤ P(Ei , Ω) + P(Ω),
so that sup{P(Ei ) | i ∈ N} < ∞. Since Ω is bounded, we may
therefore apply the compactness criterion 7.53 to obtain E ⊂ Ω
such that χE ∈ BV(Rn ) and such that, passing to a subsequence if
necessary, Ei * E. Then |Ei | → |E|, so that |E| = m, i.e. E ∈ E.
Besides, by the lower semicontinuity of the variation 7.32 it follows
that P(E, Ω) ≤ lim inf P(Ei , Ω) = m; since E ∈ E, we also have the
opposite inequality, hence P(E, Ω) = m and we are done.

 Bibliography

[AF03] Robert A. Adams and John J. F. Fournier, Sobolev spaces, second

ed., Pure and Applied Mathematics (Amsterdam), vol. 140, Else-
vier/Academic Press, Amsterdam, 2003. MR 2424078
[AFP00] Luigi Ambrosio, Nicola Fusco, and Diego Pallara, Functions of bounded
variation and free discontinuity problems, Oxford Mathematical Mono-
graphs, The Clarendon Press, Oxford University Press, New York, 2000.
MR 1857292
[Bou87] N. Bourbaki, Topological vector spaces. Chapters 1–5, Elements of Mathe-
matics (Berlin), Springer-Verlag, Berlin, 1987, Translated from the French
by H. G. Eggleston and S. Madan. MR 910295
[Car14] C. Carathéodory, Über das lineare maß von punktmengen- eine verall-
gemeinerung des längenbegriffs, Nachrichten von der Gesellschaft der
Wissenschaften zu Göttingen, Mathematisch-Physikalische Klasse 1914
(1914), 404–426.
[Con90] John B. Conway, A course in functional analysis, second ed., Gradu-
ate Texts in Mathematics, vol. 96, Springer-Verlag, New York, 1990.
MR 1070713
[dL65] Elon Lages de Lima, Cálculo Tensorial, Notas de Matemática, vol. 32,
Instituto de Matemática Pura e Aplicada, Rio de Janeiro, 1965.
[EG91] L.C. Evans and R.F. Gariepy, Measure theory and fine properties of func-
tions, Studies in Advanced Mathematics, Taylor & Francis, 1991.
[Eng89] Ryszard Engelking, General topology, second ed., Sigma Series in Pure
Mathematics, vol. 6, Heldermann Verlag, Berlin, 1989, Translated from
the Polish by the author. MR 1039321
[Fed69] Herbert Federer, Geometric measure theory, Die Grundlehren der math-
ematischen Wissenschaften, Band 153, Springer-Verlag New York Inc.,
New York, 1969. MR 0257325
[Fol99] Gerald B. Folland, Real analysis, second ed., Pure and Applied Math-
ematics (New York), John Wiley & Sons, Inc., New York, 1999, Mod-
ern techniques and their applications, A Wiley-Interscience Publication.
MR 1681462
[Hau18] Felix Hausdorff, Dimension und äußeres Maß, Math. Ann. 79 (1918),
no. 1-2, 157–179. MR 1511917
[K6̈9] Gottfried Köthe, Topological vector spaces. I, Translated from the Ger-
man by D. J. H. Garling. Die Grundlehren der mathematischen Wis-
senschaften, Band 159, Springer-Verlag New York Inc., New York, 1969.
MR 0248498
[KP08] Steven G. Krantz and Harold R. Parks, Geometric integration theory,
Cornerstones, Birkhäuser Boston, Inc., Boston, MA, 2008. MR 2427002

277
278 BIBLIOGRAPHY

[LL01] Elliott H. Lieb and Michael Loss, Analysis, second ed., Graduate Studies
in Mathematics, vol. 14, American Mathematical Society, Providence, RI,
2001. MR 1817225
[Mag12] F. Maggi, Sets of finite perimeter and geometric variational problems: An
introduction to geometric measure theory, Cambridge Studies in Advanced
Mathematics, Cambridge University Press, 2012.
[Mat95] Pertti Mattila, Geometry of sets and measures in Euclidean spaces, Cam-
bridge Studies in Advanced Mathematics, vol. 44, Cambridge University
Press, Cambridge, 1995, Fractals and rectifiability. MR 1333890
[Maz11] Vladimir Maz’ya, Sobolev spaces with applications to elliptic partial differ-
ential equations, augmented ed., Grundlehren der Mathematischen Wis-
senschaften [Fundamental Principles of Mathematical Sciences], vol. 342,
Springer, Heidelberg, 2011. MR 2777530
[Mos09] Yiannis N. Moschovakis, Descriptive set theory, second ed., Mathemati-
cal Surveys and Monographs, vol. 155, American Mathematical Society,
Providence, RI, 2009. MR 2526093
[Osb14] M. Scott Osborne, Locally convex spaces, Graduate Texts in Mathematics,
vol. 269, Springer, Cham, 2014. MR 3154940
[Rud87] Walter Rudin, Real and complex analysis, third ed., McGraw-Hill Book
Co., New York, 1987. MR 924157
[Sim83] Leon Simon, Lectures on geometric measure theory, Proceedings of the
Centre for Mathematical Analysis, vol. 3, Australian National University,
Centre for Mathematical Analysis, Canberra, 1983. MR 756417
[Sri98] S. M. Srivastava, A course on Borel sets, Graduate Texts in Mathematics,
vol. 180, Springer-Verlag, New York, 1998. MR 1619545
[Ste38] J. Steiner, Einfache Beweise der isoperimetrischen Hauptsätze, J. Reine
Angew. Math. 18 (1838), 281–296. MR 1578194
[SW99] H. H. Schaefer and M. P. Wolff, Topological vector spaces, second ed.,
Graduate Texts in Mathematics, vol. 3, Springer-Verlag, New York, 1999.
MR 1741419
[Tre06] François Treves, Topological vector spaces, distributions and kernels,
Dover Publications, Inc., Mineola, NY, 2006, Unabridged republication
of the 1967 original. MR 2296978
 Index

absolutely continuous, 31, 74 continuity of a measure

absolutely continuous part, 74 from above, 3
algebra of sets, 2 from below, 3
almost everywhere, 21 continuous linear map on Cc (X, Rn ),
approximate identity, 39 107
Area Formula, 143 convergent sequence in Cc (X, Rn ),
107
Besicovitch covering theorem, 67 counting measure, 1
Binet-Cauchy formula, 134 cover
Borel fine, 49
map, 14 strongly fine, 65
measure, 6
locally finite, 8
De La Vallée Poussin theorem, 117
open σ-finite, 8
regular measure, 6 density
set, 5 upper and lower n-dimensional, 57
bounded subset of Cc (X, Rn ), 107 upper and lower density of a
bounded variation function, 230 measure relative to another, 62
density theorem
Caccioppoli set, 230 comparison, 59
Carathéodory general, 68
construction, 43 lower density, 75
criterion, 6 for the Lebesgue measure, 61
extension theorem, 3 general, 70
measurability condition, 1 upper, 60
theorem, 2 general, 70
Chain rule for Sobolev functions, 184 differentiable Urysohn’s lemma, 40
chain rule for the Radon-Nikodym differentiation theorem
derivative, 34 for Borel measures, 77
change of variables formula, 147 Lebesgue-Besicovitch-Radon-
for C1-diffeomorphisms, 27 Nikodym,
Coarea Formula, 159 78
Compactness from perimeter Legesgue, 71
bounds, 272, 273 Dirac measure, 1
compactness theorem for BV, 268 dominated convergence theorem, 22
comparison density theorem, 59 generalized, 22
conjugate exponent, 24 doubling property, 66
279
280 INDEX

Egorov’s theorem, 42 Kolmogorov-Riesz-Fréchet

Lp space, 24 compactness criterion for Lp , 26
epigraph, 188
existence of relative isoperimetric Layer-cake formula, 29
sets, 275 Lebesgue decomposition theorem,
Extension theorem 33, 74
for BV functions, 266 for Rn -valued Radon measures,
for Sobolev functions on Lipschitz 106
domains, 207 Lebesgue differentiation theorem, 71
for Sobolev functions on Lipschitz Lebesgue measure, 2
epigraphs, 204 Lebesgue point, 73
exterior normal, 233 Lebesgue-Besicovitch-Radon-
Nikodym differentiation
Fatou’s lemma, 22 theorem, 78
fine cover, 49 Lipschitz domain, 187
finite perimeter set, 230 Lipschitz linearization theorem, 141
5–times covering lemma, 48 locally finite Borel measure, 8
foliations by Borel sets for positive locally finite family, 171
Radon measures, 113 lower integral, 23
Friedrichs’s theorem, 181 Lusin’s theorem, 40
Fubini-Tonelli’s theorem, 28
curvilinear, 165 McShane’s lemma, 121
Fundamental Lemma of the Calculus measurable
of Variations, 106 map, 14
set, 14
Gauss-Green measure, 233 with respect to a measure, 1
Gauss-Green theorem space, 14
for Lipschitz domains, 236 measure, 1
for Lipschitz epigraphs, 196 σ-finite, 2, 3
generalized, 234 Borel, 6
approximation by open and
Hölder spaces, 219 closed sets, 8
Hölder’s inequality, 24 locally finite, 8
generalized, 24 open σ-finite, 8
Hausdorff Borel regular, 6
dimension, 47 complete, 2
measure, 44 concentrated on a set, 11
finite, 2, 3
integrable function, 20
Hausdorff, 44
integral, 20
on a σ-algebra, 2
with respect to an Rn -valued
Rn -valued, 94
measure, 93
outer, 1
isodiametric inequality, 53
positive, 31
Jacobian product, 27
of linear maps, 133 pushforward, 4
of Lipschitz maps, 136 Radon, 10
Jordan decomposition theorem, 31 Rn -valued, 81
regular, 3
Kirszbraun’s theorem, 122 restriction, 4
 INDEX 281

for Rn -valued measures, 104 for Rn -valued Radon measures,

signed, 30 106
support, 11 theorem, 33
trace, 4 Rellich Kondrachov compactness
of an Rn -valued Radon measure, theorem, 227
106 Riesz representation theorem
Meyers-Serrin’s theorem, 182 for finite Radon measures, 90
Minimizers for the Plateau problem, for positive linear functionals, 83
274 for Radon measures, 84
Minkowski’s inequality, 25 for the dual of Lp , 26
for integrals, 30
mollifier, 39 shrinking lemma, 172
monotone convergence theorem, 22 σ-algebra, 2
Morrey’s inequality, 219 generated by a family of maps, 15
multiplicity function, 138 Borel, 5
mutually singular, 31, 74 generated by a set, 5
product, 16
negative part pullback, 16
of a function, 15 trace, 16
of a signed measure, 32 σ-finite set, 3
signed measure, 30
open σ-finite Borel measure, 8
finite, 31
outer unit normal, 198
σ-finite, 31
partition of unity, 171 simple function, 18
strictly subordinate to an open extended, 23
cover, 173 standard representation, 20
subordinate to an open cover, 171 singular part, 74
perimeter measure, 233 Sobolev conjugate, 210
Poincaré’s inequality, 216 Sobolev function, 126
point-finite family, 171 local, 126
polar decomposition Sobolev-Gagliardo-Nirenberg
of linear maps, 132 inequality, 210
of vector measures, 91 standard mollifier, 39
positive linear functional, 83 Steiner symmetrization, 51
positive measure, 31 strongly fine cover, 65
positive part summable function, 20
of a function, 15 with respect to a signed measure,
of a singed measure, 32 32
Product rule with respect to an Rn -valued
for BV, 243, 248 measure, 93
for Sobolev functions, 183 support
essential, 19
Rademacher’s theorem, 128 of a function, 19
Radon measure, 10 of a measure, 11
Rn -valued, 81 symmetric Vitali property, 65
set functions, 94
Radon-Nikodym t-linearization, 140
derivative, 33 t-regularization
chain rule, 34 of a function, 176
282 INDEX

of a Radon measure, 239

Tietze’s extension theorem for LCH,
83
total variation
of a signed measure, 32
total variation measure, 91
trace of an Rn -valued Radon
measure, 106
Trace theorem
for BV functions on Lipschitz
domains, 261
for BV functions on Lipschitz
epigraphs, 249
for Sobolev functions on Lipschitz
domains, 201
for Sobolev functions on Lipschitz
epigraphs, 198
transpose of a continuous linear map
on Cc (X, Rn ), 108

unordered sum, 21
upper density theorem, 60
general, 70
upper integral, 23
Urysohn’s lemma
differentiable, 40, 173
for LCH, 83

variation
measure, 245
of a function in L1loc , 245
version of a function, 219
Vitali’s covering theorem, 49

weak derivatives and gradients, 123

as Rn -valued Radon measures, 229
weak-star convergence, 111
in the sense of finite measures, 111
narrow, 115

Young’s inequality, 35
 List of Symbols

σ(µ) σ-algebra of µ-measurable sets 1

δa Dirac measure centered at a. 1
Ln Lebesgue measure on Rn 2
L Lebesgue σ-algebra 2
x
µ A restriction of µ to A 4, 105
µ|A trace of µ on A 4, 106
f# µ pushforward of µ by f 4, 109
σ(S) σ-algebra generated by A 5
BX Borel σ-algebra of X 5
c cardinality of the continuum 8
Gδ countable intersection of open sets 8
spt µ support of a measure µ 11
f+ max{f, 0} 15
f−
max{−f, 0} 15
σ (fα )α∈A σ-algebra generated by (fα )α∈A 16
⊗α∈A Mα product σ-algebra 16
f ∗N pullback of N by f 16
N |X trace of N on X 16
z
sgn z |z|
if z 6= 0 and 0 otherwise 18
spt f support of a function f 19
ess spt f essential support of a function f 19
+
L
´ (µ) µ-measurable functions taking values on [0, ∞] 20
f dµ integral of f with respect to µ 20
1
L
P(µ) summable functions with respect to µ 21
x∈X f (x) unordered sum of f 21
µ-a.e.
´∗ almost everywhere with respect to µ 21
´ f dµ upper integral of f with respect to µ 23
∗
f dµ lower integral of f with respect to µ 23
kf kp p-norm of f 24
Lp (µ) p-summable functions with respect to µ 24

283
284 List of Symbols

Cc (X) space of continuous functions with compac support 25

τy f x 7→ f (x − y) 26, 35
µ×ν product measure of µ and ν 27
µν µ is absolutely continuous with respect to ν 31, 74, 106
µ⊥ν µ and ν are mutually singular 31, 74, 106
ν+ positive part of ν 32
ν− negative part of ν 32
|µ| total variation of µ 32, 91
dν
dµ
Radon-Nikodym derivative of ν with respect to µ 33
fˇ x 7→ f (−x) 35   α1   αn
∂ αf For a multi-index α ∈ Zn , ∂x∂ 1 . . . ∂x∂n f 35
n
|α| For a multi-index α ∈ Z , α1 + · · · + αn . 35
Hm Hausdorff m-dimensional measure 45
Hδm size δ approximation of Hausdorff m-dimensional measure 45
H-dim A Hausdorff dimension of A 47
Sa Steiner symmetrization with respect to hai⊥ 51
Θ∗n (µ, A, x) n-dimensional upper density of A at x with respect to µ 57
Θn∗ (µ, A, x) n-dimensional lower density of A at x with respect to µ 57
Θn (µ, A, x) n-dimensional density of A at x with respect to µ 57
Θ∗n (µ, x) n-dimensional upper density at x with respect to µ 57
Θn∗ (µ, x) n-dimensional lower density at x with respect to µ 57
Θn (µ, x) n-dimensional density at x with respect to µ 57
Θ∗ν (µ, x) upper density of µ with respect to ν at x 62
Θν∗ (µ, x) lower density of µ with respect to ν at x 62
Θν (µ, x) density of µ with respect to ν at x 62
SVP symmetric Vitali property 65
L1loc (µ) f : X → C µ-measurable and locally µ-summable 71
f ≺U f ∈ Cc (U ) and 0 ≤ f ≤ 1 83
C+ c (X) {f ∈ Cc (X) | f ≥ 0} 84
B cX the set of Borel subsets of X which are relatively compact 94
M(X)n or M(X, Rn ) finite Rn -valued Radon measures on X 95
Mloc (X)n or Mloc (X, Rn ) Rn -valued Radon measures on X 95
µ g
∗
x restriction of µ to g 104
* weak star convergence 111
∗f
* weak star convergence for finite measures 111
∗ nc
* narrow convergence 115
∂wu
∂xi
i-th weak partial derivative of u 123
 Glossary 285

∇w u weak gradient of u 123

W1,p space of (1, p)-Sobolev functions 126
1,p
Wloc space of local (1, p)-Sobolev functions 126
Df Fréchet derivative of f 127
O(n, m) set of orthogonal injections Rn → Rm 131
Sym(n) set of symmetric linear maps Rn → Rn 132
JLK Jacobian of L 133
Λ(m, n) set of strictly increasing functions {1, . . . , m} → {1, . . . , n} 134
Jf Jacobian of f 136
Df {x | ∃Df (x)} 136
Jf+ {x ∈ Df | ∃ Jf (x) > 0} 136
Jf0 {x
´ ∈D f | ∃ Jf (x) = 0} 136
kf kW1,p (Ω) ( Ω |f | + kgrad f kp )1/p dLn 169
p

kf kW1,∞ (Ω) |f | + k∇f k L∞ (Ω) 169

ffl
−1 ´
(f )x,r = B(x,r) f dLn average of f on B(x, r) ⊂ Rn , i.e. Ln B(x, r) B(x,r)
f dLn 216
C0,γ (Ω) Hölder continuous functions with exponent γ on Ω 219
BVloc (Ω) locally bounded variation functions on Ω 230
BV(Ω) bounded variation functions on Ω 230
P(E, ·) perimeter measure of a set E of locally finite perimeter 233
Var(f, V ) variation of f on V 245