Geometric Measure Theory

Ilkka Holopainen

May 8, 2021
2 Geometric Measure Theory

Contents
1 Review of measure theory 3
1.1 Measures and outer measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.16 Metric outer measure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.20 Regularity of measures, Radon-measures . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.35 Hausdorff measure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.47 Hausdorff dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.52 Hausdorff measures in Rn . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.60 Riesz representation theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.64 Weak convergence of measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.68 Compactness of measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

2 Lipschitz mappings and rectifiable sets 21

2.1 Extension of Lipschitz mappings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.5 Rademacher’s theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.8 Linear maps and Jacobians . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.15 Jacobians of Lipschitz mappings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.17 The area formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.23 The co-area formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.29 Rectifiable sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

3 Varifolds 37
3.1 Basic definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.8 First and second variation formulae . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

4 Currents 46
4.2 m-vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.10 m-covectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.13 m-vector fields, m-covector fields, and smooth differential m-forms . . . . . . . . . . 50
4.20 m-currents; definition and basic notions . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.72 Rectifiable currents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.84 Slicing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.93 Deformation theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.103Rectifiability and compactness theorems . . . . . . . . . . . . . . . . . . . . . . . . . 77

5 Mass minimizing currents 87

6 Appendix 89
6.1 Proof of Riesz’ representation theorem 1.62 . . . . . . . . . . . . . . . . . . . . . . . 89
6.10 Proof of Theorem 1.67 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
6.11 Proof of Theorem 1.69 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
Fall 2016 3

The material is collected mainly from books [EG], [Fe], [LY], [Ma], [Mo], and [Si]
and from the lecture notes ”Currents and varifolds” (fall 2011) by P. Mattila and
”Moderni reaalianalyysi” by I. Holopainen.
The aim of the course is to give an introduction to the theory of varifolds and currents that
are kind of generalized surfaces. They have been used in many geometric variational problems, in
particular, in connections with higher dimensional minimal surfaces.
First we recall some basic notions of geometric measure theory.

1 Review of measure theory

1.1 Measures and outer measures
Let X be a set and let
P(X) = {A : A ⊂ X}
be the power set of X (also denoted by 2X ).
Definition 1.2. The collection M ⊂ P(X) is a σ-algebra “sigma algebra”) in X if
(1) ∅ ∈ M;
(2) A ∈ M ⇒ Ac = X \ A ∈ M;
S
(3) Ai ∈ M, i ∈ N ⇒ ∞
i=1 Ai ∈ M.

Example 1.3. 1. P(X) is the largest σ-algebra in X;

2. {∅, X} is the smallest σ-algebra in X;
3. Leb(Rn ) is the class of Lebesgue measurable sets of Rn .
4. If M is a σ-algebra in X and A ⊂ X, then
M|A = {B ∩ A : B ∈ M}
is a σ-algebra in A.
5. If M is a σ-algebra in X and A ∈ M, then
MA = {B ⊂ X : B ∩ A ∈ M}
is a σ-algebra in X.
Definition 1.4. If F ⊂ P(X) is a family of subsets of X, then
\
σ(F) = {M : M is a σ-algebra in X, F ⊂ M}

is the σ-algebra generated by F. It is the smallest σ-algebra that contains F.

Example 1.5. Recall that the set I ⊂ Rn is an open n-interval if it is of the form
I = {(x1 , . . . , xn ) : aj < xj < bj },
where −∞ ≤ aj < bj ≤ +∞. Then
notat.
σ({I : I n-interval}) = σ({A : A ⊂ Rn open}) = Bor(Rn )
is the σ-algebra of Borel sets of Rn . (Can you prove the left side equality?)
4 Geometric Measure Theory

Observe that all open subsets of Rn , closed sets, Gδ sets (countable intersections of open sets),
Fσ sets (countable unions of closed sets), Fσδ sets, Gδσ sets (etc.) are Borel sets. Thus for example
the set of rational numbers Q is Borel.

Remark 1.6. In every topological space X one can define Borel sets as

Bor(X) = σ({A : A ⊂ X open}).

Definition 1.7. Let M be a σ-algebra in X. A mapping µ : M → [0, +∞] is a measure if there

holds:

(i) µ(∅) = 0,
S
P∞
(ii) µ ∞ i=1 Ai = i=1 µ(Ai ) if the sets Ai ∈ M are disjoint.

The triple (X, M, µ) is called a measure space and the elements of M measurable sets.

The condition (ii) is called countably additivity. It follows from the definition that a measure is
monotone: If A, B ∈ M and A ⊂ B, then µ(A) ≤ µ(B).

Remark 1.8. 1. If µ(X) < ∞, the measure µ is finite.

2. If µ(X) = 1, then µ is a probability measure.

S
3. If X = ∞ i=1 Ai , where µ(Ai ) < ∞ ∀i, the measure µ is σ-finite. Then we shall say that X is
σ-finite with respect to µ.

4. If A ∈ M and µ(A) = 0, then A is of measure zero.

5. If X is a topological space and Bor(X) ⊂ M (i.e. every Borel set is measurable), then µ is a
Borel measure.

Example 1.9. 1. X = Rn , M = Leb Rn = the family of Lebesgue measurable sets and µ =

mn = the Lebesgue measure.

2. X = Rn , M = Bor Rn = the family of Borel sets and µ = mn | Bor Rn = the restriction of

the Lebesgue measure to the family of Borel sets.

3. Let X 6= ∅ be any set. Fix x ∈ X and set for all A ⊂ X

(
1, if x ∈ A;
µ(A) =
0, if x ∈
6 A.

Then µ : P(X) → [0, +∞] is a measure (so called Dirac measure at x ∈ X). We often write
µ = δx .

4. If f : Rn → [0, +∞] is Lebesgue measurable, then µ : Leb(Rn ) → [0, +∞],

Z
µ(E) = f (x)dmn (x),
E

is a measure.
Fall 2016 5

5. If (X, M, µ) is a measure space and A ∈ M, then the mapping µxA : MA → [0, +∞],
(µxA)(B) = µ(B ∩ A),
is a measure. It is called the restriction of µ to A.
Theorem 1.10. Let (X, M, µ) be a measure space and A1 , A2 , . . . ∈ M.
(a) If A1 ⊂ A2 ⊂ A3 · · · , then
∞
[

µ Ai = lim µ(Ai ).
i→∞
i=1

(b) If A1 ⊃ A2 ⊃ A3 ⊃ · · · and µ(Ak ) < ∞ for some k, then

∞
\

µ Ai = lim µ(Ai ).
i→∞
i=1

Proof. Course ”Mitta ja integraali”.

Definition 1.11. A mapping µ̃ : P(X) → [0, +∞] is an outer measure in X if the following holds:
(i) µ̃(∅) = 0;
P S∞
(ii) µ̃(A) ≤ ∞ i=1 µ̃(Ai ) if A ⊂ i=1 Ai ⊂ X.
Remark 1.12. 1. An outer measure is defined for all subsets of X.
2. Condition (ii) (monotone subadditivity) implies that an outer measure is monotone, i.e.
µ̃(A) ≤ µ̃(B) if A ⊂ B ⊂ X.
3. In many books an outer measure is simply called a measure. (Soon we will do so, too.)
4. Let µ̃ be an outer measure in X and A ⊂ X. Then the restriction of µ̃ to A, defined by
(µ̃xA)(B) = µ̃(B ∩ A)
is an outer measure in X.
Every outer measure defines the σ-algebra of “measurable” sets in terms of the Carathéodory
condition.
Definition 1.13. Let µ̃ be an outer measure in X. A set E ⊂ X is µ̃-measurable, or briefly
measurable, if
µ̃(A) = µ̃(A ∩ E) + µ̃(A \ E)
for all A ⊂ X.
Theorem 1.14. Let µ̃ be an outer measure in X and
M = Mµ̃ = {E ⊂ X : E is µ̃-measurable}
Then
(a) M is a σ-algebra and
(b) µ = µ̃|M is a measure (i.e. µ is countably additive).
Proof. Course ”Mitta ja integraali”.
Definition 1.15. We say that an outer measure µ̃ in a topological space X is a Borel outer measure
if every Borel set of X is µ̃-measurable (i.e. if the measure defined by µ̃ is a Borel measure).
6 Geometric Measure Theory

1.16 Metric outer measure

We shall next study the question when an outer measure µ̃ in a topological space X is Borel.
Definition 1.17 (Carathéodory’s criterion). An outer measure µ̃ in a metric space (X, d) is a
metric outer measure if
µ̃(A ∪ B) = µ̃(A) + µ̃(B)
for all A, B ⊂ X, for which dist(A, B) = inf{d(a, b) : a ∈ A, b ∈ B} > 0.
Theorem 1.18. An outer measure µ̃ of a metric space (X, d) is a Borel outer measure if and only
if µ̃ is a metric outer measure.
We first formulate and prove the following lemma.
Lemma 1.19. Let µ̃ be a metric outer measure, A ⊂ X and G an open set such that A ⊂ G. If

Ak = {x ∈ A : dist(x, Gc ) ≥ 1/k}, k ∈ N,

then µ̃(A) = limk→∞ µ̃(Ak ).

S∞ S∞
Proof. Since G is open, A ⊂ k=1 Ak . Thus A = k=1 Ak . Let

Bk = Ak+1 \ Ak .

Then ! !
∞
[ ∞
[
A = A2n ∪ B2k ∪ B2k+1 ,
k=n k=n
and thus
∞
X ∞
X
µ̃(A) ≤ µ̃(A2n ) + µ̃(B2k ) + µ̃(B2k+1 ) .
k=n k=n
| {z } | {z }
=(I) =(II)

Let now n → ∞.
(1) If the sums (I), (II) → 0 as n → ∞, then

µ̃(A) ≤ lim µ̃(A2n ) ≤ µ̃(A)

n→∞

and the claim is true.

(2) If (I) 6→ 0 as n → ∞, then X
µ̃(B2k ) = ∞.
k
On the other hand,
n−1
[
A ⊃ A2n ⊃ B2k ,
k=1
where
1 1
dist(B2k , B2k+2 ) ≥ − > 0.
2k + 1 2k + 2
Because µ̃ is a metric outer measure, we have
n−1
X n−1
[

µ̃(B2k ) = µ̃ B2k ≤ µ̃(A2n ) ≤ µ̃(A).
k=1 k=1
Fall 2016 7

Letting n → ∞ we obtain
µ̃(A) = lim µ̃(Ak ) = ∞.
k→∞

The argument goes in the same way if the sum (II) 6→ 0 as n → ∞.

Proof of Theorem 1.18. Suppose first that µ̃ is a metric outer measure. We want to prove that µ̃
is a Borel outer measure. Because Bor(X) = σ({F : F ⊂ X closed}) and Mµ̃ is a σ-algebra, it is
enough to show that every closed set F ⊂ X is µ̃-measurable.
Let E ⊂ X be an arbitrary test set in the Carathéodory condition. We apply Lemma 1.19 for
the sets A = E \ F and G = X \ F . Let Ak = {x ∈ E \ F : dist(x, Gc ) ≥ 1/k}, k ∈ N. Then

dist(Ak , F ) ≥ 1/k

and
lim µ̃(Ak ) = µ̃(E \ F ).
k→∞
Because µ̃ is metric,

µ̃(E) ≥ µ̃ (E ∩ F ) ∪ Ak = µ̃(E ∩ F ) + µ̃(Ak ).
Letting k → ∞ we get
µ̃(E) ≥ µ̃(E ∩ F ) + µ̃(E \ F ).
On the other hand it follows from the monotonicity of the outer measure that

µ̃(E) ≤ µ̃(E ∩ F ) + µ̃(E \ F ).

Thus F is µ̃-measurable and µ̃ is a Borel outer measure.

The proof of the converse implication is left as an exercise.

1.20 Regularity of measures, Radon-measures

Among outer measures particularly useful are those with a large class of measurable sets. Such
outer measures are called regular.

Definition 1.21. We say that an outer measure µ̃ of X is regular if, for every A ⊂ X, there exists
a µ̃-measurable set E such that A ⊂ E and µ(E) = µ̃(A) (such a set E is called a measurable cover
of A.)

Definition 1.22. Let X be a topological space.

(a) We say that an outer measure µ̃ of X is Borel regular if µ is a Borel measure and for every
A ⊂ X there exists a Borel set B ∈ Bor(X) such that A ⊂ B and µ(B) = µ̃(A).

(b) Let (X, M, µ) be a measure space such that Bor(X) ⊂ M (i.e. µ is a Borel measure). Then
the measure µ is called Borel regular if for every A ∈ M there exists B ∈ Bor(X) such that
A ⊂ B and µ(A) = µ(B).

Lemma 1.23. If µ̃ is a Borel regular outer measure in X and A ⊂ X is µ̃-measurable s.t. µ(A) <
∞, then µ̃xA is Borel regular. If A ∈ Bor(X), then the assumption µ(A) < ∞ is not needed.

Proof. Exercise.

Theorem 1.24. Let µ̃ be a Borel regular outer measure in a metric space X, A ⊂ X µ̃-measurable
and ε > 0.
8 Geometric Measure Theory

(a) If µ(A) < ∞, then there exists a closed set C ⊂ A s.t. µ(A \ C) < ε.
S
(b) If there exist open sets V1 , V2 , . . . ⊂ X s.t. A ⊂ ∞i=1 Vi and µ(Vi ) < ∞ ∀i, then there exists
an open set V ⊂ X s.t. A ⊂ V and µ(V \ A) < ε.
Proof. (a): By replacing µ̃ with a Borel regular outer measure µ̃xA (see Lemma 1.23) we may
assume that µ̃(X) < ∞. We first prove the claim for Borel sets A. Let

D = {A ⊂ X : ∀ε > 0 ∃ closed C ⊂ A and open V ⊃ A s.t. µ(V \ C) < ε}.

We easily see that D satisfies condition (1) and (2) in the definition of a σ-algebra. Suppose that
A1 , A2 , . . . ∈ D and let ε > 0. Then
S there exists closed sets Ci and open sets Vi s.t. Ci ⊂ Ai ⊂ Vi
i
and µ(Vi \ Ci ) < ε/2 . Now V = i Vi is open and
[
X
µ V \ Ci ≤ µ(Vi \ Ci ) < ε.
| S {zi } i

⊂ i (Vi \Ci )

On the other hand, by Theorem 1.10 (b)

n
[ ∞
[



lim µ V \ Ci = µ V \ Ci ,
n→∞
i=1 i=1

and hence there exists n s.t.

n
[

µ V \ Ci < ε.
i=1
Sn
Because i=1 Ci is closed, D satisfies also the condition (3) of a σ-algebra. We next show that D
contains closed sets. Let C be closed and

Vi = {x ∈ X : dist(x, C) < 1/i}.

T
Then Vi is open, V1 ⊃ V2 ⊃ · · · , and C = i Vi . Therefore limi→∞ µ(Vi ) = µ(C) and limi→∞ µ(Vi \
C) = 0. This implies that C ∈ D. Thus D is a σ-algebra containing all closed sets. In particular,
Bor(X) ⊂ D. Therefore part (a) holds for all Borel sets.
Suppose next that A is µ̃-measurable and µ(A) < ∞. Because µ̃ is Borel regular, there exists
a Borel set B ⊃ A s.t. µ(A) = µ(B). Then µ(B \ A) = 0. Furthermore, there exists a Borel set
D ⊃ B \ A s.t. µ(D) = 0. Now E = B \ D is Borel, E ⊂ A, and

µ(A \ E ) = 0.
| {z }
⊂D

Applying the first part of the proof to the Borel set E we conclude that, for each ε > 0, there exists
a closed set C ⊂ E = B \ D (⊂ A) s.t. µ(E \ C) < ε. But then

µ(A \ C) ≤ µ(A \ E) + µ(E \ C) < ε,

and hence (a) holds for the set A.

(b) By applying part (a) to the set Vi \ A we obtain closed sets Ci ⊂ Vi \ A s.t.

µ(Vi \ A \ Ci ) < ε2−i .

S
Then V = i (Vi \ Ci ) is open, A ⊂ V and µ(V \ A) < ε.
Fall 2016 9

Remark 1.25. The Borel regularity of the outer measure was not needed to prove claims (a) and
(b) for Borel sets A. Therefore Theorem 1.24 holds for all Borel outer measures, if we, furthermore,
assume that A is Borel.

We also have the following version of Theorem 1.24

Theorem 1.26. Let µ̃ be a Borel regular outer measure in a metric space X and
∞
[
X= Vj ,
j=1

where Vj is open and µ(Vj ) < ∞ for each j ∈ N. Then

(1.27) µ̃(A) = inf{µ(U ) : U open, A ⊂ U }

for every A ⊂ X, and

(1.28) µ(A) = sup{µ(C) : C closed, C ⊂ A}

for every µ-measurable A ⊂ X.

So called Radon measures will be important in what follows. These will be defined next. Recall
that a topological space X is locally compact, if every point x ∈ X has a neighbourhood with com-
pact closure. A topological space is Hausdorff , if its distinct points have disjoint neighbourhoods.

Definition 1.29. Let X be a locally compact Hausdorff space. We say that a measure µ is a
Radon measure, if µ is a Borel measure and

(a) µ(K) < ∞ for all compact K ⊂ X;

(b) µ(V ) = sup{µ(K) : K ⊂ V compact} for all open V ⊂ X;

(c) µ(B) = inf{µ(V ) : B ⊂ V and V ⊂ X open} for all Borel sets B ∈ Bor(X).

Remark 1.30. 1. In general, a Borel regular measure (in a locally compact Hausdorff space)
need not be a Radon measure.

2. On the other hand a Radon measure need not be Borel regular: Let A ⊂ R be non-Lebesgue
measurable, µ̃ = m∗ xA and

µ = µ̃|{E ⊂ R : E µ̃-measurable}.

Then µ is a Radon measure, but not Borel regular.

In some cases Radon measures can be easily characterised.

Theorem 1.31. Let µ be a Borel measure in Rn . Then µ is a Radon measure, if and only if µ is
locally finite, i.e.


∀x ∈ Rn , µ B(x, r) < ∞, when 0 < r < rx .
10 Geometric Measure Theory

Proof. It follows from Definition 1.29 (a) that all Radon measures are locally finite.
Suppose next that µ is locally finite Borel measure in Rn . If K ⊂ Rn is compact, then choose
for every x ∈ K an open ball with centre at x with a finite measure. Because K is compact, it can
be covered with finitely many such balls. Therefore the measure of K is finite and (a) holds.
We next prove conditions (b) and (c) for every Borel set A ⊂ Rn . By applying part (a) of
Theorem 1.24 (see also Remark 1.25) for Borel sets Ai of finite measure,

Ai = A ∩ B(0, i), B(0, i) = {x ∈ Rn : |x| ≤ i},

we find closed sets Ci ⊂ Ai s.t.

µ(Ai \ Ci ) < 1/i.
Then Ci is a compact set as a closed and bounded set (in Rn ). Now

1.10(a)
µ(A) ≥ µ(Ai ) ≥ µ(Ci ) > µ(Ai ) − 1/i −−−−→ µ(A) .
S
This implies (b). Because A ⊂ i B(0, i) and µ(B(0, i)) < ∞, it follows from Theorem 1.24 part
(b) that there exist open sets Vj ⊂ Rn s.t. A ⊂ Vj and µ(Vj \ A) < 1/j. Then

µ(A) ≤ µ(Vj ) = µ(A) + µ(Vj \ A) < µ(A) + 1/j,

This implies (c).

Corollary 1.32. Let µ̃ be a locally finite metric outer measure in Rn . Then the measure µ =
µ̃|M, M = {A ⊂ Rn : A µ̃-measurable}, determined by µ̃ is a Radon measure.

Remark 1.33. Theorem 1.31 holds also more generally. For example, if X is a locally compact
metric space, whose topology has a countable base.

Convention: From now on we call an outer measure µ̃ simply a measure and (to simplify the
notation) we denote it by µ.
Note that outer measures and measures come in a sense hand in hand. Indeed, an outer measure
µ̃ : P(X) → [0, +∞] defines the measure µ = µ̃|M, where M is the σ-algebra of µ̃-measurable sets
and, on the other hand, every measure µ : M → [0, +∞] defined on a σ-algebra M ⊂ P(X) can be
extended to an outer measure µ̃ : P(X) → [0, +∞] by setting

µ̃(A) = inf{µ(B) : A ⊂ B ∈ M}.

Let µ be regular and Ai ⊂ Ai+1 ⊂ X for i ∈ N. We have the counterpart of Theorem 1.10 (a)
∞
[

µ Ai = lim µ(Ai )
i→∞
i=1

even if the sets Ai are not assumed to be µ-measurable.

Let then X be a locally compact, separable metric space. We say that µ is a Radon (outer)
measure if µ is Borel regular and if µ is finite on compact subsets of X. Then such a measure µ
has the properties
µ(A) = inf{µ(U ) : U open, A ⊂ U }
for every A ⊂ X and
µ(A) = sup{µ(K) : K compact, K ⊂ A}
Fall 2016 11

for every µ-measurable A ⊂ X.

Since µ is finite on compact sets, we can integrate continuous functions with compact support.
In particular, if H is a Hilbert space with inner product (·, ·) and if C0 (X, H) denotes the space of
continuous functions X → H with compact support, then associated to each Radon measure µ and
each µ-measurable H-valued function ν : X → H, with |ν| = 1 µ-a.e., we have the linear functional
L : C0 (X, H) → R defined by Z
L(f ) = (f, ν) dµ.
X
The following Riesz representation theorem gives the converse:
Theorem 1.34. Let H be a finite dimensional Hilbert space and let L : C0 (X, H) → R be a linear
functional such that

sup{L(f ) : f ∈ C0 (X, H), |f | ≤ 1, supp f ⊂ K} < ∞

for each compact K ⊂ X. Then there exist a Radon measure µ and a µ-measurable mapping
ν : X → H such that |ν(x)| = 1 for µ-a.e. x ∈ X and
Z
L(f ) = (f, ν) dµ
X

for every f ∈ C0 (X, H).

We will return to this later.

1.35 Hausdorff measure

The Lebesgue n-dimensional measure mn is well suited for the measurement of the size of “large”
subsets of Rn , but it is too crude for measuring “small” subsets of Rn . For example, m2 cannot
distinguish a singleton of R2 from a line, because both have measure zero.
In this chapter we introduce a whole spectrum of “s-dimensional” measures Hs , 0 ≤ s < ∞,
which are able to see the fine structure of sets, better than the Lebesgue measure. The key idea is
that a set A ⊂ Rn is “s-dimensional”, if 0 < Hs (A) < ∞, even if the geometric structure of A were
very complicated.
These measures can be defined in any metric space (X, d). We suppose, however, that X is
separable, i.e. X has a countable dense subset S = {xi }∞ i=1 , and hence X = S̄. This assumption is
only needed to guarantee that X has so called δ-covering for all δ > 0.
Definition 1.36. 1. The diameter of a nonempty set E ⊂ X is

d(E) = sup d(x, y).

x,y∈E

2. A countable collection {Ei }∞

i=1 of subsets of X is a δ-covering, δ > 0, of A ⊂ X if
∞
[
A⊂ Ei and d(Ei ) ≤ δ ∀i ∈ N.
i=1

We fix a “dimension” s ∈ [0, ∞) and δ > 0. For A ⊂ X, we define

∞
 X
s
(1.37) Hδs (A) = inf ωs d(Ei )/2 : {Ei } is a δ-covering of A ,
i=0
12 Geometric Measure Theory

where ωs is the volume of the unit ball in Rs in case s is a positive integer and

0 otherwise some
convenient

s positive constant, and where we make the convention that d({x})/2 = 1 ∀x ∈ X and
d(∅)/2 = 0 ∀s ≥ 0.
We readily see from the definition that

Hδs1 (A) ≥ Hδs2 (A),

if 0 < δ1 ≤ δ2 . Therefore the following limit (1.39) exists and we can set the definition.

Definition 1.38. The s-dimensional Hausdorff (outer) measure of a set A ⊂ X is

 
s
(1.39) H (A) = lim Hδs (A) s
= sup Hδ (A) .
δ→0 δ>0

Remark 1.40. The constant ωs above is usually chosen as

π s/2
ωs =
,
Γ 2s + 1
R∞
where Γ(t) = 0 e−x xt−1 dx, 0 < t < ∞, is the usual gamma function.
In particular, this guarantees that Hn and the n-dimensional Lebesgue outer measure m∗n co-
incide in Rn , i.e.
Hn (A) = m∗n (A) ∀A ⊂ Rn .

We will not prove this identity. For the proof, see e.g. [Si].

Theorem 1.41. (i) Hδs : P(X) → [0, +∞] is an outer measure for all δ > 0.

(ii) Hs : P(X) → [0, +∞] is a metric outer measure.

Proof. (i) (a) Clearly Hδs (∅) = 0.

S
(b) Let then A ⊂ ∞ s
i=1 Ai ⊂ X. We may suppose that Hδ (Ai ) < ∞ ∀i. Let ε > 0 and
choose for every i a δ-covering {Eji }∞
j=1 of the set Ai s.t.

∞
X
s
ωs d(Eji )/2 ≤ Hδs (Ai ) + ε2−i .
j=1

S S∞
Then i,j Eji is a δ-covering of the union i=1 Ai and thus also of A and therefore
X
s
Hδs (A) ≤ ωs d(Eji )/2
i,j
∞
X

≤ Hδs (Ai ) + ε2−i
i=1
∞
X
≤ε+ Hδs (Ai ).
i=1

Letting ε → 0 the desired conclusion follows.

Fall 2016 13

S∞
(ii) Clearly Hs (∅) = 0. If A ⊂ i=1 Ai ⊂ X, then by part (i) and the definition of Hs we obtain
∞
X ∞
X
Hδs (A) ≤ Hδs (Ai ) ≤ Hs (Ai ).
i=1 i=1

Letting δ → 0 we see that Hs is an outer measure. Let then A1 , A2 ⊂ X be sets, for which
dist(A1 , A2 ) > 0. We wish to show that

Hs (A1 ∪ A2 ) = Hs (A1 ) + Hs (A2 ).

It is enough to show that

(1.42) Hδs (A1 ∪ A2 ) ≥ Hδs (A1 ) + Hδs (A2 ),

if δ ≤ dist(A1 , A2 )/3. We may assume that Hδs (A1 ∪ A2 ) < ∞. Let ε > 0 and choose a
δ-covering {Ei }∞i=1 of the set A1 ∪ A2 such that

∞
X
s
ωs d(Ei )/2 ≤ Hδs (A1 ∪ A2 ) + ε.
i=1

Because δ ≤ dist(A1 , A2 )/3, every Ei intersects at most one of the sets A1 or A2 . Therefore
we may divide the δ-covering {Ei }∞ i=1 of A1 ∪ A2 into two disjoint δ-coverings of A1 and A2
as
{Ei }∞ ′ ∞ ′′ ∞
i=1 = {Ei }i=1 ⊔ {Ei }i=1 ,

where
∞
[ ∞
[
A1 ⊂ Ei′ and A2 ⊂ Ei′′ .
i=1 i=1

Therefore
∞
X ∞
X

s
s
Hδs (A1 ) + Hδs (A2 ) ≤ ωs d(Ei′ )/2 + ωs d(Ei′′ )/2
i=1 i=1
∞
X
s
= ωs d(Ei )/2
i=1
s
≤ Hδ (A1 ∪ A2 ) + ε.

Because ε > 0 was arbitrary, we obtain (1.42).

On the basis of Theorems 1.18 and 1.41 every Borel-set of X is Hs -measurable. We denote the
restriction of Hs to Hs -measurable sets with the same symbol Hs . Now there holds:

Theorem 1.43. Hs is a Borel-measure.

Corollary 1.32 yields now:

Corollary 1.44. If A ⊂ Rn is Hs -measurable and Hs (A) < ∞, then Hs xA is a Radon-measure.

Theorem 1.45. The outer measure Hs of a separable metric space X is Borel-regular.

14 Geometric Measure Theory

Proof. Because by the previous theorem the outer measure defined by Hs is Borel, it is enough to
show that for all A ⊂ X there exists B ∈ Bor(X) s.t. A ⊂ B and Hs (A) = Hs (B).
Let A ⊂ X. If Hs (A) = ∞, we may choose B = X and the claim holds. If Hs (A) < ∞, then
we choose a 1/i-covering {Eji }∞
j=1 of A for each i ∈ N s.t.

∞
X
s
ωs d(Eji )/2 ≤ H1/i
s
(A) + 1/i.
j=1

Because d(E) = d(Ē) for all E ⊂ X, we may suppose that the sets Eji are closed. Then

∞ [
\ ∞
B= Eji
i=1 j=1

is a Borel set and A ⊂ B. Furthermore, {Eji } is a 1/i-covering of B for all i ∈ N, and hence

∞
X
s s

s
H1/i (A) ≤ H1/i (B) ≤ ωs d(Eji )/2 ≤ H1/i
s
(A) + 1/i.
j=1

Letting i → ∞ the claim Hs (A) = Hs (B) follows.

Remark 1.46. 1. H0 is the counting measure.

2. Hδs is not, in general, a metric outer measure.

3. Roughly speaking H1 ∼ is a length measure, H2 ∼ is area, etc.

4. It is easily seen that (e.g.) the plane R2 is not σ-finite with respect to H1 .

1.47 Hausdorff dimension

Let (X, d) be a separable metric space. In this chapter we shall define a dimension for sets A ⊂ X,
which reflects the metric size of the set A. Unlike the topological dimension, this dimension need
not be an integer.

Lemma 1.48. Let A ⊂ X and s ≥ 0.

(i) If Hs (A) < ∞, then Ht (A) = 0 for all t > s.

(ii) If Hs (A) > 0, then Ht (A) = ∞ for all 0 ≤ t < s.

Proof. It is enough to prove (i), because the claim (ii) follows from (i). Let δ > 0 and {Ej }∞
j=1 be
a δ-covering of A s.t.
∞
X
s
ωs d(Ej )/2 ≤ Hδs (A) + 1 ≤ Hs (A) + 1 < ∞.
j=1
Fall 2016 15

Then for all t > s

∞
X
t
Hδt (A) ≤ ωt d(Ej )/2
j=1
X∞

s
t−s
= ωt d(Ej )/2 d(Ej )/2
j=1
∞
ωt t−s
X
s
≤ (δ/2) ωs d(Ej )/2
ωs
j=1
ωt
≤ (δ/2)t−s (Hs (A) + 1) .
ωs

The claim follows by letting δ → 0.

Definition 1.49. The Hausdorff dimension of a subset A ⊂ X is a number

dimH (A) = inf{s > 0 : Hs (A) = 0}.

Summarizing what was said above:

1. If t < dimH (A), then Ht (A) = ∞.

2. If t > dimH (A), then Ht (A) = 0.

In general, about the value Hs (A), for s = dimH (A), we cannot say anything: it can take any value
in [0, ∞]. Nevertheless:

(1.50) 0 < Hs (A) < ∞ ⇒ dimH (A) = s .

A set A ⊂ X, for which 0 < Hs (A) < ∞ holds, is called an s-set.

Lemma 1.51. (i) If A ⊂ B, then dimH (A) ≤ dimH (B).

(ii) If Ak ⊂ X, k ∈ N, then
∞
[

dimH Ak = sup dimH (Ak ).
k=1 k

Proof. (Exerc.)

Thus, for example, dimH (Q) = 0.

1.52 Hausdorff measures in Rn

Next we evaluate (or rather estimate) Hausdorff measures and dimensions of Cantor type fractal
sets in Rn . To this end we study the invariance properties of Hausdorff measures. There are other,
more efficient, methods for the determination of the Hausdorff dimension but these will not be
discussed in this course.
Recall first that a mapping T : Rn → Rn is an isometry, if

|T x − T y| = |x − y| ∀x, y ∈ Rn .
16 Geometric Measure Theory

It is well-known that every isometry of Rn is an affine mapping, i.e. of the form

T x = a0 + U x,

where a0 ∈ Rn and U : Rn → Rn is a linear isometry.

In the same way, a mapping R : Rn → Rn is said to be a similarity, if

|Rx − Ry| = c|x − y| ∀x, y ∈ Rn ,

where c > 0 is a constant (stretching factor, scaling factor, etc.). Then R is of the form

Rx = a0 + c U x,

where U is again a linear isometry.

Theorem 1.53. Let A ⊂ Rn . For the outer measure Hs , s ≥ 0, there holds:

(a) Hs (A + x0 ) = Hs (A) ∀x0 ∈ Rn ,

(b) Hs U (A) = Hs (A) for all linear isometries U : Rn → Rn ,


(c) Hs R(A) = cs Hs (A), if R : Rn → Rn is a similarity map, with scaling factor c > 0.


Proof. The claims follow from the observation that d R(E) = c d(E) ∀E ⊂ Rn , where R is as in
part (c).

Let (X, d1 ) and (Y, d2 ) be metric spaces. Recall next that the mapping f : X → Y is L-Lipschitz
(with a constant L > 0), if

d2 f (x), f (y) ≤ L d1 (x, y)
for all x, y ∈ X. In the same way, a mapping g : X → Y is L-bilipschitz , if
1

d1 (x, y) ≤ d2 f (x), f (y) ≤ L d1 (x, y)
L
for all x, y ∈ X. We observe that an L-bilipschitz mapping is always an injection because of the
inequality on the left hand side.

Lemma 1.54. Let (X, d1 ) and (Y, d2 ) be separable metric spaces.

(i) If f : X → Y is L-Lipschitz, then

Hs (f A) ≤ Ls Hs (A) ∀A ⊂ X.

(ii) If g : X → Y is L-bilipschitz, then

dimH (gA) = dimH (A) ∀A ⊂ X.

Proof. (i) We may suppose that Hs (A) < ∞. Fix ε > 0, δ > 0 and choose a δ-covering {Ej }∞
j=1
of A s.t.
X∞

s
ωs d(Ej )/2 ≤ Hδs (A) + ε.
j=1
Fall 2016 17

Then {f (Ej )}∞

j=1 is a Lδ-covering of f A and hence
∞
X
s


s
HLδ (f A) ≤ ωs d f (Ej ) /2
j=1
∞
X
s

s
≤ L ωs d(Ej )/2
j=1
s s
≤L (Hδ (A) + ε).
The claim follows by letting ε → 0 and δ → 0.
(ii) Applying part (i) to the mapping g −1 : g(A) → X we obtain
L−s Hs (A) ≤ Hs (gA).
Thus
L−s Hs (A) ≤ Hs (gA) ≤ Ls Hs (A),
which yields the claim.

We next construct sets with noninteger Hausdorff dimension. Recall the construction of the
Cantor set from Real Analysis I. (We use slightly different notation and consider only a special
case.)
Let 0 < λ < 1/2. Denote I0,1 = [0, 1], I1,1 = [0, λ] and I1,2 = [1 − λ, 1]. In other words, I1,1
and I1,2 is obtained from I0,1 by removing its middle interval with length 1 − 2λ. Next we remove
open interval of length (1 − 2λ)λ from the middle of closed intervals I1,i and continue the process
inductively. Suppose that the intervals In,i , i = 1, . . . , 2n of step n have been defined. Then the
intervals In+1,j , j = 1, . . . , 2n+1 of the step (n + 1) are obtained by removing an open interval of
length (1 − 2λ)λn from the middle of the intervals of step n. Thus
d(In,i ) = λn , ∀n and ∀i = 1, . . . , 2n .
Denote n
2
[
Cn (λ) = In,i
i=1
(“approximation of the nth step”) and
∞
\
C(λ) = Cn (λ).
n=1
Then C(λ) is
compact, uncountable set, without interior points. Furthermore C(λ) is “selfsimilar”
and m1 C(λ) = 0. Cantor’s 1/3-set C(1/3) is a special case of this construction, recurrent in
literature.
I0,1
C0

I1,1 I1,2
C1

I2,1 I2,4
C2
.. ..
. .
18 Geometric Measure Theory

Theorem 1.55. For all 0 < λ < 1/2

log 2
dimH C(λ) = .
log(1/λ)
In particular, dimH C(λ) can attain all values in the interval (0, 1).
Proof. It is enough to show that


(1.56) 2−1−s ωs ≤ Hs C(λ) ≤ 2−s ωs ,

if
log 2
s= .
log(1/λ)
(i) We first give a heuristic argument for finding the exponent s: Clearly

C(λ) = C1 ∪ C2 ,

where C1 and C2 are disjoint and similar to C(λ) with the scaling factor λ. If C(λ) would
satisfy (1.56), then by part (c) of Theorem 1.53


Hs C(λ) = Hs (C1 ) + Hs (C2 )


= 2λs Hs C(λ) .

Thus
1 = 2λs ,
Solving this for s yields
log 2
s= .
log(1/λ)
(ii) A rigorous proof of (1.56): If δ > 0 is given, then choose n ∈ N so large that λn < δ. Then
n
{In,i }2i=1 is a δ-covering of C(λ) and, furthermore,
2 n 2 n 2 n

X X X
Hδs C(λ) ≤ ωs (λn /2)s = 2−s ωs (λn )s = 2−s ωs (1/2)n = 2−s ωs .
i=1 i=1 i=1

Thus

Hs C(λ) ≤ 2−s ωs .
We give a proof for the lower bound (1.56) only in the special case λ = 1/3. The general case
λ ∈ (0, 1/2) would not bring any essential changes to the proof. Suppose that {Ej }∞ j=1 is a
δ-covering of C(1/3) such that
∞
X
s
log 2
ωs d(Ej )/2 ≤ Hs C(1/3) + δ, s= .
log 3
j=1

For each j choose a closed interval Ij (= [a, b]) s.t. Ej ⊂ int Ij (=]a, b[) and d(Ij ) < (1 +
δ)d(Ej ). Then {int Ij }∞
j=1 is an open covering of C(1/3), and hence by the compactness of
C(1/3) we can choose a finite subcover. By relabelling the intervals Ij , we may suppose that
m
[
C(1/3) ⊂ Ij
j=1
Fall 2016 19

and
∞
X


s
Hs C(1/3) + δ ≥ ωs d(Ej )/2
j=1
m
X
−s −s
≥ ωs 2 (1 + δ) d(Ij )s .
j=1

To prove the lower bound (1.56) it is enough to prove that

m
X 1
(1.57) d(Ij )s ≥ ,
2
j=1

if {Ij }m
j=1 is a covering of C(1/3) with finitely many closed intervals Ij . For each j choose
k = k(j) ∈ N, with
(1.58) 3−(k+1) ≤ d(Ij ) < 3−k .
Let k0 be the largest one of the numbers k(j), j = 1, . . . , m. On the basis of the construction
and the choice of the number k = k(j), each Ij can intersect only one of the intervals Ik,i
of step k. Therefore Ij intersects at most 2k0 −k(j) of the intervals Ik0 ,i . Thus the number of
such intervals of step k0 is at most
m
X
2k0 −k(j) .
j=1

On the other hand, there

S are 2k0
intervals of step k0 . Every one of these contains points of
C(1/3) and C(1/3) ⊂ m I
j=1 j , and hence
m
X
2k0 ≤ 2k0 −k(j) .
j=1

Now we can compute

m
X m
X
k0 k0 −k(j) k0
2 ≤ 2 =2 2−k(j)
j=1 j=1
m
X
= 2k0 (3−k(j) )s
j=1
Xm

s
≤ 2k0 3d(Ij ) .
j=1

Simplification yields
m
X
d(Ij )s ≥ 3−s = 1/2.
j=1

Remark 1.59. Refining the above argument one can show that

log 2
Hs C(λ) = 1, s = ,
log(1/λ)
(cf. Falconer, K. J.: The geometry of fractal sets, Cambridge University Press, 1985, pages 14-15).
20 Geometric Measure Theory

1.60 Riesz representation theorem

Recall the following general form of the Riesz representation theorem. Let X be a locally compact,
separable metric space, and let H be a finite dimensional Hilbert space with the inner product
(·, ·). Denote by C0 (X, H) the space of all continuous mappings X → H with compact support. If
L : C0 (X, H) → R is a linear functional such that

sup{L(f ) : f ∈ C0 (X, H), |f | ≤ 1, supp f ⊂ K} < ∞

for all compact K ⊂ X, there exist a Radon measure µ and a µ-measurable ν : X → H such that
|ν(x)| = 1 for ν-a.e. x ∈ X and Z
L(f ) = (f, ν)dµ
X
for every f ∈ C0 (X, H).
See, for example, [Si, Theorem 4.1]. We will consider the special case X = Rn and H = R in
the home work classes.

Definition 1.61. A mapping Λ : C0 (Rn , R) → R is a positive linear functional if

(i) Λ(αf1 + βf2 ) = αΛ(f1 ) + βΛ(f2 ) for all f1 , f2 ∈ C0 (Rn , R) and all α, β ∈ R.

(ii) Λ(f ) ≥ 0 for all f ∈ C0 (Rn , R), with f (x) ≥ 0 ∀x ∈ Rn .

Theorem 1.62 (Riesz representation theorem). Let Λ : C0 (Rn , R) → R be a positive linear

functional. Then there exists a unique Radon measure µ, more precisely, a measure space
(Rn , Bor(Rn ), µ), such that Z
Λ(f ) = f (x) dµ(x)
Rn
for all f ∈ C0 (Rn , R).

The proof of Riesz representation theorem is based on an auxiliary result.

Lemma 1.63. (a) Let V ⊂ Rn be open and K ⊂ V compact. Then there exists f ∈ C0 (Rn , R)
such that
supp(f ) ⊂ V χK (x) ≤ f (x) ≤ 1 ∀x ∈ Rn .
and
S
(b) Let Vj ⊂ Rn , j = 1, . . . , m, be open and K ⊂ m j=1 Vj compact. Then there exist functions
hj ∈ C0 (Rn , R), with
m
X
0 ≤ hj ≤ 1, supp(hj ) ⊂ Vj and χK ≤ hj ≤ 1.
j=1

1.64 Weak convergence of measures

Definition 1.65. Let µk , k ∈ N, be Radon measures in Rn . We say that a sequence (µk ) converges
weakly to a Radon measure a µ, if
Z Z
lim f dµk = f dµ
k→∞ Rn Rn

w
all f ∈ C0 (Rn , R). This is denoted by µk ⇀ µ or µk −
→ µ.
Fall 2016 21

Example 1.66. (i) Let δx be the Dirac measure at x ∈ R. Then δk ⇀ 0.

1


(ii) Let µk = k δ1/k + δ2/k + · · · + δk/k . Then for all f ∈ C0 (R, R)

Z k Z 1
X 1 k→∞
f dµk = f (j/k) −−−→ f (x) dx,
R k 0
j=1

because the sums are Riemann sums of the function f on [0, 1]. Thus µk ⇀ m1 x[0, 1].

Easy examples show that it is not always true that µk (A) → µ(A), if µk ⇀ µ. However, there
holds:

Theorem 1.67. Let µ, µk , k ∈ N, be Radon measures in Rn such that µk ⇀ µ. Then

(a) lim supk→∞ µk (K) ≤ µ(K) if K ⊂ Rn is compact,

(b) lim inf k→∞ µk (V ) ≥ µ(V ) if V ⊂ Rn is open.

1.68 Compactness of measures

The weak convergence of measures is not merely natural but also a very useful notion. The families
of bounded Radon measures are sequentially compact. In many cases this is the only way to
construct measures (as limiting measures of weakly convergent sequences).

Theorem 1.69. Let (µk ) be a sequence of Radon measures in Rn with

sup µk (K) < ∞

k

for all compact K ⊂ Rn . Then there exists a subsequence (µkj ) and a Radon measure µ with

µkj ⇀ µ.

The proof of of this theorem will be discussed in the home work classes.

2 Lipschitz mappings and rectifiable sets

2.1 Extension of Lipschitz mappings
Next we present a useful extension result of Lipschitz mappings.

Theorem 2.2 (McShane-Whitney extension theorem). Let X be a metric space, A ⊂ X, and

f : A → R L-Lipschitz. Then there exists an L-Lipschitz function F : X → R such that F |A = f .

Proof. For every a ∈ A we define an L-Lipschitz function f a : X → R

f a (x) = f (a) + L|a − x|, x ∈ X.

The function F is then defined by setting

F (x) = inf f a (x), x ∈ X.

a∈A
22 Geometric Measure Theory

Clearly F (x) < ∞ ∀x ∈ X. By fixing a0 ∈ A we see that

f (a) + L|a − x| ≥ f (a) + L|a − a0 | − L|a0 − x|

≥ f (a0 ) − L|a0 − x|.

Hence F (x) > −∞ for all x ∈ X. Since every f a is L-Lipschitz and F (x) > −∞ for all x ∈ X, F
is L-Lipschitz. Moreover, for every x ∈ A

F (x) ≤ f x (x) = f (x) ≤ f (y) + L|x − y| = f y (x) ∀y ∈ A,

and hence F |A = f .

Corollary 2.3. Let X be a metric space, A ⊂ X, and f : A → Rn L-Lipschitz. Then there exists
√
a nL-Lipschitz mapping F : X → Rn such that F |A = f.

Proof. Apply Theorem 2.2 to the coordinate functions of f .

Remark 2.4. 1. Theorem 2.2 holds (as such) in the case X ⊂ Rm , f : X → Rn , but the proof
is much harder. This is so called Kirzbraun’s theorem.

2. It is a topic of quite active current research to study which pairs of metric spaces X, Y have
a Lipschitz extension property (i.e. for every A ⊂ X every Lipschitz mapping f : A has a
Lipschitz extension F : X → Y ).

2.5 Rademacher’s theorem

According to Rademacher’s theorem a Lipschitz mapping Rn → Rm is differentiable mn -a.e. Let
us first recall the following definition.

Definition 2.6. A mapping f : Rn → Rm is differentiable at x ∈ Rn if there exits a linear mapping

L : Rn → Rm such that
|f (y) − f (x) − L(y − x)|
lim =0
y→x |y − x|
or, equivalently,
f (y) = f (x) + L(y − x) + o(|y − x|) as y → x.

If such L exists, it is unique and we denote it by Df (x) and call it the derivative of f at x or the
differential of f at x.

Theorem 2.7 (Rademacher’s theorem). Let f : Rn → Rm be locally Lipschitz, i.e. for each compact
K ⊂ Rn there exist a constant LK < ∞ such that

|f (x) − f (y)| ≤ LK |x − y| ∀x, y ∈ K.

Then f is differentiable mn -a.e. in Rn .

The proof will be discussed in home work sessions.

Fall 2016 23

2.8 Linear maps and Jacobians

Let us start with the following fact: Suppose that L : Rn → Rn is linear. Then

m∗n (LA) = |det L|m∗n (A)

for every A ⊂ Rn . We will not prove this formula (think of the special case Lx =
(c1 x1 , c2 x2 , . . . , cn xn ), where c1 , . . . , cn ∈ R). We want to have a counterpart of this ”area for-
mula” in case L : Rn → Rm is linear. Towards this end, let us first recall the following notions
related to linear algebra (without proofs).
Definition 2.9. (i) A linear map O : Rn → Rm is orthogonal if

Ox · Oy = x · y ∀x, y ∈ Rn .

(ii) A linear map S : Rn → Rn is symmetric if

x · Sy = Sx · y ∀x, y ∈ Rn .

(iii) A linear map D : Rn → Rn is diagonal if there are constants d1 , . . . , dn such that

Dx = (d1 x1 , . . . , dn xn ) ∀x = (x1 , . . . , xn ) ∈ Rn .

(iv) The adjoint of a linear map L : Rn → Rm is the linear map L∗ : Rm → Rn defined by

x · L∗ y = (Lx) · y ∀x ∈ Rn , y ∈ Rm .

Theorem 2.10. (i) L∗∗ = L.

(ii) (AB)∗ = B ∗ A∗ .

(iii) O∗ = O−1 if O : Rn → Rn is orthogonal.

(iv) S ∗ = S if S : Rn → Rn is symmetric.

(v) For every symmetric map S : Rn → Rn there exist an orthogonal map O : Rn → Rn and a
diagonal map D : Rn → Rn such that

S = ODO −1 .

(vi) If O : Rn → Rm is orthogonal, then n ≤ m and

O∗ O = id in Rn ,
OO ∗ = id in ORn .

Theorem 2.11 (Polar decomposition). Let L : Rn → Rm be a linear mapping.

(i) If n ≤ m, there exists a symmetric map S : Rn → Rn and an orthogonal map O : Rn → Rm
such that
L = OS.

(ii) If n ≥ m, there a symmetric map S : Rm → Rm and an orthogonal map O : Rm → Rn such

that
L = SO ∗ .
24 Geometric Measure Theory

For the proof, see e.g. [EG]. We are now ready to define the Jacobian of a linear map L : Rn →
Rm .

Definition 2.12. Let L : Rn → Rm be a linear mapping.

(i) If n ≤ m, let L = OS be as above and define the Jacobian of L as

JLK = |det S|.

(ii) If n ≥ m, let L = SO ∗ be as above and define the Jacobian of L as

JLK = |det S|.

Theorem 2.13. (i) If n ≤ m, then JLK2 = det(L∗ L).

(ii) If n ≥ m, then JLK2 = det(LL∗ ).

(iii) JLK = JL∗ K.

Remark 2.14. The Jacobian JLK is well-defined since it is independent of the choices of S and O
by Theorem 2.13.

2.15 Jacobians of Lipschitz mappings

Let f = (f1 , . . . , fm ) : Rn → Rm be a Lipschitz mapping. By Rademacher’s theorem, f is differen-
tiable at mn -a.e. x ∈ Rn . Hence the derivative Df (x) exists and can be expressed as a matrix
 ∂f1 ∂f1 ∂f1 
∂x1 ∂x2 ··· ∂xn
Df (x) =  ... .. .. 

. . 
∂fm ∂fm ∂fm
∂x1 ∂x2 ··· ∂xn

at mn -a.e. x ∈ Rn .

Definition 2.16. The Jacobian of f at a point x, where f is differentiable, is

Jf (x) = JDf (x)K.

2.17 The area formula

Some details will be discussed in the home work classes.
In this subsection we assume that n ≤ m and that f : Rn → Rm is Lipschitz.

Lemma 2.18 (Area formula for linear maps). Let L : Rn → Rm , n ≤ m, be a linear map. Then

Hn (LA) = JLKm∗n(A) ∀A ⊂ Rn .

Note that we have defined Hn so that Hn = m∗n in Rn .

Lemma 2.19. Let A ⊂ Rn be Lebesgue measurable. Then

(i) f A is Hn -measurable,


(ii) the mapping y 7→ H0 A ∩ f −1 (y) is Hn -measurable, and
Fall 2016 25

(iii) Z


n
H0 A ∩ f −1 (y) dHn (y) ≤ Lip(f ) mn (A).
Rm

Lemma 2.20. Let t > 1 and B = {x ∈ Rn : the derivative Df (x) exists and Jf (x) > 0}. Then
there exists a countable collection Ek ∈ Bor(Rn ), k ∈ N, such that
(i)
∞
[
B= Ek ,
k=1

(ii) f |Ek is one-to-one,

(iii) for every k ∈ N there exists a symmetric automorphism Tk : Rn → Rn such that

Lip (f |Ek ) ◦ Tk−1 ≤ t,


Lip Tk ◦ (f |Ek )−1 ≤ t,
t−n |det Tk | ≤ Jf |Ek ≤ tn |det Tk |.

The message of the lemma is that f can be locally approximated by a symmetric automorphism
as closely as we wish.
Theorem 2.21 (The area formula). Let f : Rn → Rm , n ≤ m, be a Lipschitz mapping. Then for
every mn -measurable set A ⊂ Rn
Z Z


Jf (x)dmn (x) = H0 A ∩ f −1 (y) dHn (y).
A Rm

Corollary 2.22 (Change of variables). Let f : Rn → Rm be a Lipschitz map, n ≤ m. Then for

each mn -integrable g : Rn → R
 
Z Z X
g(x)Jf (x)dmn (x) =  g(x) dHn (y).
Rn Rm
x∈f −1 (y)

As an application let us consider the surface

 area (measure) of the graph of a Lipschitz function


g : Rn → R. Define f : Rn → Rn+1 , f (x) = x, g(x) . Then

 
1 0 ··· 0 0
 .. 
 0 1 0 ··· .   
 = In
 
Df =  ..
 . 0 0  ∇g
 
 0 0 ··· 0 1 
∂g ∂g ∂g
∂x1 ∂x2 ··· ··· ∂xn

and Jf2 = 1 + |∇g|2 . For each open U ⊂ Rn , the graph of g over U is

Γ = Γg,U = {(x, g(x)) : x ∈ U }

and Z p
n
H (Γ) = 1 + |∇g|2 dmn (x).
U
26 Geometric Measure Theory

2.23 The co-area formula

In this subsection we assume that n ≥ m. Some details will be discussed in home work classes.
Let us start with linear maps L : Rn → Rm , n ≥ m. Consider first the special case where
L : Rn = Rm × Rk → Rm is the orthogonal projection onto Rm ,

L(x1 , . . . , xm , xm+1 , . . . , xm+k ) = (x1 , . . . , xm ).

Then for each y ∈ Rm the preimage L−1 (y) is an affine (n−m)-dimensional subspace. The preimages
L−1 (y), y ∈ Rm , decompose Rn into parallel (n − m)-dimensional slices. By Fubini’s theorem
Z


Hn−m L−1 (y) ∩ A dmm (y) = Hn (A) = mn (A)
Rm

whenever A ⊂ Rn is Lebesgue measurable. For a general linear map L : Rn → Rm , n ≥ m, we have

the following.
Lemma 2.24. Let L : Rn → Rm , n ≥ m, be a linear mapping and A ⊂ Rn Lebesgue measurable.
Then


(i) the mapping y 7→ Hn−m A ∩ L−1 (y) is mm -measurable, and

(ii) Z


Hn−m A ∩ L−1 (y) dmm (y) = JLKmn (A).
Rm

Similarly to the case of area formula, we have:

Lemma 2.25. Let f : Rn → Rm , n ≥ m, be a Lipschitz mapping and A ⊂ Rn Lebesgue measurable.
Then
(i) f A is mm -measurable,

(ii) A ∩ f −1 (y) is Hn−m -measurable for mm -a.e. y ∈ Rm ,

(iii) the mapping y 7→ Hn−m A ∩ f −1 (y) is mm -measurable, and

(iv) Z

ωn−m ωm
Hn−m A ∩ f −1 (y) dmm (y) ≤ (Lip f )m mn (A).
R m ω n

Lemma 2.26. Let t > 1 and h : Rn → Rn Lipschitz. Let

B = {x ∈ Rn : Dh(x) exists and Jh (x) > 0}.

Then there exists a countable collection Dk ∈ Bor(Rn ), k ∈ N, such that

(i) mn B \ ∪k Dk = 0,

(ii) h|Dk is one-to-one, and

(iii) for each k there exists a symmetric automorphism Sk : Rn → Rn such that

Lip Sk−1 ◦ (h|Dk ) ≤ t,


Lip (h|Dk )−1 ◦ Sk ≤ t,
t−n |det Sk | ≤ Jh|Ek ≤ tn |det Sk |.
Fall 2016 27

Note that above both the domain and the target of h is Rn . For the proof of Lemma 2.26 we
apply Lemma 2.20 to find sets Ek such that each h|Ek is one-to-one and then we apply Lemma 2.20
again to (h|Ek )−1 in hEk .
Theorem 2.27 (The co-area formula). Let f : Rn → Rm , n ≥ m, be a Lipschitz map. Then for
each mn -measurable set A ⊂ Rn
Z Z


Jf (x)dmn (x) = Hn−m A ∩ f −1 (y) dmm (y).
A Rm

Corollary 2.28 (Change of variables). Let f : Rn → Rm , n ≥ m, be a Lipschitz mapping. Then

for each mn -integrable function g : Rn → R, g|f −1 (y) is Hn−m -integrable for mm -a.e. y ∈ Rm and
Z Z Z !
g(x)Jf (x)dmn (x) = gdHn−m dmm (y).
Rn Rm f −1 (y)

As an application we consider level sets of a Lipschitz function f : Rn → R. Then Jf = |∇f |

and hence Z Z ∞


|∇f |dmn = Hn−1 {f = t} dt.
Rn −∞

2.29 Rectifiable sets

Let us start with the following two examples
Example 2.30. Let J0 = Q01 = [0, 1]2 be the closed unit square of the plane and let J1 be the
union of four closed squares Q1i , i = 1, . . . , 4 , in its corners, each with edge length 1/4. In the
next step each of the four squares Q1i , i = 1, . . . , 4 of J1 , will be replaced with four corner squares
Q2j , j = 1, . . . , 16, each with edge length 1/16. Continuing in this way in the step n we have 4n
squares Qnj , j = 1, . . . , 4n , each with egde length 4−n . Let
4 n
[
Jn = Qnj
j=1

and ∞
\
J= Jn .
n=0
Then J is a set of Cantor type. In fact, J = C(1/4) × C(1/4).
J0 J1 J2

Q11 Q12
Q01
Q13 Q14 Q2j

J3

Q3i
······
28 Geometric Measure Theory

What is the Hausdorff dimension dimH J of J? We find a suitable candidate for dimH J by using
similarities. Observe that
4
G
J= J˜i ,
i=1

where J˜i is similar to J with the scaling factor 1/4, and hence Hs (J˜i ) = (1/4)s Hs (J) and further

Hs (J) = 4(1/4)s Hs (J).

If J is an s-set (i.e. 0 < Hs (J) < ∞), then we get from above that

4(1/4)s = 1
√ n
which gives s = 1. Let us give some further details. Fix δ > 0. Observe first that d(Qnj ) = 2/4 .
√ n
If n ∈ N is so large that 2/4n ≤ δ, then {Qnj }4j=1 is a δ-covering of J and thus

4n
X √ √
Hδ1 (J) ≤ d(Qnj ) ≤ 4n 2/4n = 2.
j=1
√
Therefore H1 (J) ≤ 2 < ∞. By an argument similar to that in the proof of Theorem 1.55 one can
show that H1 (J) > 0. Thus

dimH (J) = 1 and 0 < H1 (J) < ∞,

in other words J is a 1-set. However, its geometric structure is very different from that of a
rectifiable curve.
2
√ P: R →
Remark.: A positive lower bound can also be found by using an orthogonal projection
S onto a line S with slope −2. Then the image set P (J) is a segment
√ with length 3/ 5. Because
the projection 1 1
P is 1-Lipschitz, then H (J) ≥ H (P (J)) = 3/ 5. In fact, it can be shown that
√
H1 (J) = 2.

P (J)

It was pointed out above that J = C(1/4) × C(1/4). Observe that dimH (J) = 1 = 2 log 2/ log 4 =
dimH C(1/4) + dimH C(1/4).

Example 2.31. Let q1 , q2 , . . . be those points of the closed unit disk D̄ = {x ∈ R2 : |x| ≤ 1} whose
both coordinates are rational numbers. These points form a countable dense subset of D̄. Let
∞
[
E= Sj ,
j=1
Fall 2016 29

where
Sj = {x ∈ R2 : |x − qj | = 2−j }.
Now
∞
X
0 < H1 (E) ≤ 2π 2−j = 2π < ∞.
j=1

Thus E is a 1-set and dimH E = 1. However, E is dense in D̄, Ē ∩ D̄ = D̄, and therefore E is “very
big”. In which sense does E resemble a rectifiable arc?

These and other similar examples raise several questions:

• In which sense the Cantor-set of Example 2.30 is different from a rectifiable arc?

• What kind of set is a general 1-set? Is there a way to distinguish between “Cantor-type” and
“rectifiable” parts and how these parts could be defined?

• Rectifiable arcs have tangent lines a.e. Does this property have a counterpart for sets such
as in Example 2.31?

Recall Lebesgue’s density theorem from Real Analysis I: If E ∈ Leb(Rn ), then

mn (E ∩ B(x, r))
lim =1
r→0+ mn (B(x, r))
for a.e. x ∈ E. It is a natural question whether Hausdorff measures have some similar properties.
Recall that we defined the Hausdorff measure so that Hn (B̄(x, 1)) = ωn for B(x, r) ⊂ Rn . Keeping
this in mind we define:

Definition 2.32. Let 0 ≤ s < ∞, A ⊂ Rn and a ∈ Rn . The upper and lower s-densities of A at
the point a are

Hs (A ∩ B̄(a, r))
Θ∗s (A, a) = lim sup ,
r→0+ ωs r s
Hs (A ∩ B̄(a, r))
Θs∗ (A, a) = lim inf .
r→0+ ωs r s
If Θs∗ (A, a) = Θ∗s (A, a), then this value is called the (s-dimensional) density of A at a and denoted
by Θs (A, a).

We study densities using covering theorems. Recall from Real Analysis I the following basic
covering theorem and the notion of a Vitali covering of a set. If B is an open (closed) ball centered
at x with radius r, then 5B is an open (closed) ball centered at x with radius 5r.

Theorem 2.33 (Basic covering theorem). Let F be an arbitrary family of balls of Rn s.t.

D = sup{d(B) : B ∈ F} < ∞.

Then there exists a countable (possibly finite) family G ⊂ F s.t.

Bi ∩ Bj = ∅ ∀Bi , Bj ∈ G, Bi 6= Bj , i.e. the balls of G are disjoint; and

[ [
B⊂ 5B .
B∈F B∈G
30 Geometric Measure Theory

Definition 2.34. Let V be a family of balls in Rn . We say that V is a Vitali covering of a set
E ⊂ Rn if for every x ∈ E and every ε > 0 there exists B ∈ V s.t. x ∈ B and d(B) < ε. The family
V is a closed (open) Vitali covering if every B ∈ V is closed (open) ball.
Theorem 2.35 (Vitali’s covering theorem for Hausdorff measures). Let 0 < s < ∞ and let V be a
closed Vitali covering of a set E ⊂ Rn . Then there exists a countable family of disjoint balls Bi ∈ V
s.t. either
X
(2.36) d(Bi )s = ∞
i=1

or
[

(2.37) Hs E \ Bi = 0.
i=1

Remark 2.38. Applying so called Besicovitch’s covering theorem we obtain a counterpart of

Theorem 2.35 for a general Radon-measure of Rn .
Proof. We may suppose that 0 < d(B) < 1 for all B ∈ V. We choose the balls inductively: Let
B1 ∈ V be arbitrary. Suppose that disjoint balls B1 , . . . , Bm ∈ V have been chosen. Let

dm = sup{d(B) : B ∈ V, B ∩ Bi = ∅ ∀1 ≤ i ≤ m}.

If dm = 0, then
m
[
E⊂ Bi ,
i=1
and the claim is proven ((2.37) holds). Indeed, if there exists x ∈ E \ ∪m
i=1 Bi , then

dist(x, ∪m
i=1 Bi ) > 0,

because ∪mi=1 Bi is compact. Because V is a Vitali covering of E, there would exist B ∈ V s.t. x ∈ B
and B ∩ ∪m i=1 Bi = ∅ and therefore dm > 0.
If dm > 0, then choose Bm+1 ∈ V such that d(Bm+1 ) > dm /2. If this selection process will not
end for any m, we obtain disjoint balls {Bi }∞
i=1 ⊂ V. Therefore we must show: If
∞
X
d(Bi )s < ∞,
i=1

then the condition (2.37) holds. For this purpose we show first that
k
[ ∞
[
(2.39) E\ Bi ⊂ 5Bj ∀k ∈ N.
i=1 j=k+1

Indeed, if
k
[
x∈E\ Bi ,
i=1

then x ∈ B̃ ∈ V, where B̃ ∩ Bi = ∅ for all 1 ≤ j ≤ k. Because d(Bm ) → 0 as m → ∞, then

d(B̃) > 2d(Bm+1 ) for some m. Then B̃ must intersect one of the sets Bk+1 , . . . , Bm , because
otherwise
dm ≥ d(B̃) > 2d(Bm+1 ) > dm .
Fall 2016 31

Let Bj be the first one of the balls Bk+1 , . . . , Bm , which B̃ intersects. Then

B̃ ∩ Bj 6= ∅, d(Bj ) > dj−1 /2 ≥ d(B̃)/2 and k + 1 ≤ j ≤ m.

Thus B̃ ⊂ 5Bj and (2.39) holds.

B̃

Bj

Finally let δ > 0. When k is large enough, then d(5Bj ) ≤ δ for all j ≥ k. Thus

∞
[ k
[



Hδs E\ Bi ≤ Hδs E\ Bi
i=1 i=1
∞
[

≤ Hδs 5Bj
j=k+1
X∞
−s
≤ ωs 2 d(5Bj )s
j=k+1
X∞
s
= ωs (5/2) d(Bj )s → 0
j=k+1

as k → ∞. Thus
∞
[

Hδs E\ Bi = 0
i=1

for all δ > 0 and (2.37) holds.

The next theorem is useful when studying local properties of s-sets.

Theorem 2.40. Let 0 < s < ∞, A ⊂ Rn , and Hs (A) < ∞.

(a) 2−s ≤ Θ∗s (A, a) ≤ 1 for Hs -a.e. a ∈ A.

(b) If A is Hs -measurable, then Θ∗s (A, a) = 0 Hs -a.e. a ∈ Rn \ A .

Remark 2.41. 1. The lower density Θs∗ (A, a) could be zero for every a ∈ Rn even if Hs (A) > 0;
see [Ma, Exerc. 2, p. 99 and 4.12].

2. The upper bound 1 in (a) is sharp for all s > 0. The lower bound 2−s is sharp for 0 < s ≤ 1,
but it is not known whether it is sharp for s > 1; see [Ma, 6.4 (2)].

Proof. We first prove the left inequality of part (a). Observe first that
∞
[
∗s −s
{x ∈ A : Θ (A, x) < 2 }= {x ∈ A : Hs (A ∩ B̄(x, r)) < 2−s ωs (1 − 1/k)r s ∀ 0 < r < 1/k}
| {z }
k=1 =Ck
32 Geometric Measure Theory

and then show that Hs (Ck ) = 0 for all k ∈ N. Fix k ∈ N, ε > 0, and denote C = Ck . Cover C
with the sets Ej , j ∈ N, s.t. 0 < d(Ej ) < 1/k, C ∩ Ej 6= ∅, and
∞
X
2−s ωs d(Ej )s ≤ Hs (C) + ε.
j=1

For every j choose xj ∈ C ∩ Ej and denote rj = d(Ej ). Because C ∩ Ej ⊂ A ∩ B̄(xj , rj ), then

∞
X
Hs (C) ≤ Hs (C ∩ Ej )
j=1
∞
X
≤ Hs (A ∩ B̄(xj , rj ))
j=1
∞
X
≤ 2−s ωs (1 − 1/k)rjs
j=1
∞
X
−s
=2 ωs (1 − 1/k) d(Ej )s
j=1
s
≤ (1 − 1/k)(H (C) + ε).

Letting ε → 0 we see that Hs (C) = 0, because 1 − 1/k < 1 and Hs (C) < ∞.
We next prove the right hand side inequality of part (a). Because Hs is a Borel regular (outer
measure), we may suppose that A is a Borel set. Then Corollary 1.44 shows that Hs xA is a Radon
measure.
For t > 1, let
E = {x ∈ A : Θ∗s (A, x) > t}.
We wish to show that Hs (E) = 0. Let δ > 0 and ε > 0. Because Hs xA is a Radon measure, there
exists an open set U ⊂ Rn s.t. E ⊂ U and

Hs (A ∩ U ) < Hs (E) + ε.

For every x ∈ E there exists a radius rx < δ/2, for which B(x, rx ) ⊂ U and a sequence of radii
ri < rx , ri → 0, s.t.

(2.42) Hs (A ∩ B̄(x, ri )) > tωs ris ∀i ∈ N.

(Note that the sequence ri depends on x hence ri = ri (x).) Next we apply Vitali’s covering theorem
to the Vitali covering V = {B̄(x, ri ) : x ∈ E, i ∈ N} of E. Therefore there exist disjoint closed balls
{Bj } ⊂ V s.t.

(2.43) Hs (E \ ∪j Bj ) = 0.

Note that by (2.42)

X X
tωs 2−s d(Bj )s ≤ Hs (A ∩ Bj ) ≤ Hs (A ∩ U ) ≤ Hs (E) + ε < ∞,
j j

and therefore the option (2.36) does not hold and, consequently, (2.43) follows. Hence
X
Hs (E) + ε ≥ t2−s ωs d(Bj )s ≥ t Hδs (E ∩ ∪j Bj ) ≥ t Hδs (E),
j
Fall 2016 33

where the last inequality follows from (2.43) and the subadditivity of Hδs . Letting ε → 0 and δ → 0,
we obtain t Hs (E) ≤ Hs (E) < ∞. Therefore Hs (E) = 0, because t > 1.
Finally we prove (b): Let t > 0 and

B = {x ∈ Rn \ A : Θ∗s (A, x) > t}.

We prove that Hs (B) = 0. Fix δ > 0 and ε > 0. We apply part (b) of Theorem 1.24 to the Borel
regular outer measure Hs xA (see Lemma 1.23). Because (Hs xA)(B) = 0, then by 1.24 there exists
an open U ⊂ Rn s.t. B ⊂ U and Hs (A ∩ U ) < ε. For every x ∈ B there exists a radius 0 < r(x) < δ
s.t. B̄(x, r(x)) ⊂ U and

s
Hs A ∩ B̄(x, r(x)) > tωs r(x) .
From the basic covering theorem 2.33 it follows that there exist (countably many) disjoint balls
Bi = B̄(xi , r(xi )) s.t. B ⊂ ∪i 5Bi . Thus
X
s
t H10δ (B) ≤ t2−s ωs d(5Bi )s
i
X
s −s
= 5 t2 ωs d(Bi )s
i
X
s s
≤5 H (A ∩ Bi )
i
≤ 5s H (A ∩ U )
s

≤ 5s ε.
s (B) = 0, which implies further that Hs (B) = 0 as δ → 0.
Letting ε → 0 we see that H10δ

Corollary 2.44. Let A, B ⊂ Rn be Hs -measurable s.t. B ⊂ A and Hs (A) < ∞. Then for Hs -a.e.
x ∈ B there holds:
Θ∗s (A, x) = Θ∗s (B, x) and Θs∗ (A, x) = Θs∗ (B, x).
Proof.



Hs A ∩ B̄(x, r) Hs (A \ B) ∩ B̄(x, r) Hs B ∩ B̄(x, r)
= + .
ωs r s ωs r s ωs r s
| {z }
r→0+
−−−−→0 H -a.e. x∈B
s

Definition 2.45. A set E ⊂ Rn is m-rectifiable, m ∈ N, if Hm (E) < ∞ and there exists Lipschitz
maps fi : Rm → Rn , i ∈ N, such that
!
[
m m
H E \ fi R = 0.
i

Usually the finiteness of Hs (E) is not required, in which case E is called countably m-rectifiable.
By the McShane-Whitney extension theorem 2.2 it is equivalent to say that
!
[
Hm E \ fi Ai = 0,
i

where Ai ⊂ Rm and fi : Ai → Rn are Lipschitz. More importantly, by applying Rademacher’s

theorem and (a consequence of) Whitney’s extension theorem, we have
34 Geometric Measure Theory

Lemma 2.46. Let E ⊂ Rn be a Hm -measurable, with Hm (E) < ∞. Then E is m-rectifiable if and
only if there exist m-dimensional C 1 -smooth submanifolds Mi ⊂ Rn , i ∈ N, such that
!
[
Hm E \ Mi = 0.
i

Definition 2.47. A set P ⊂ Rn is purely m-unrectifiable if

Hm (P ∩ R) = 0

for all m-rectifiable R ⊂ Rn .

Remark 2.48. The set E in Example 2.31 is 1-rectifiable whereas the set J in Example 2.30 is
purely 1-unrectifiable.

Theorem 2.49. Let E ⊂ Rn be Hm -measurable, with Hm (E) < ∞ (and m ∈ N). Then there exist
Hm -measurable sets P and R such that R is m-rectifiable, P is purely m-unrectifiable,

E =R∪P and R ∩ P = ∅.

Proof. Set M = sup{Hm (R) : R ⊂ E, R is m-rectifiable}. For each i ∈ N, choose an m-rectifiable

set Ri such that
Hm (Ri ) > M − 1/i.

Then we can choose R = ∪i Ri and P = E \ R.

For m, n ∈ N, with m < n, let G(n, m) denote the (Grassmannian) space of all m-dimensional
(vector) subspaces of Rn .

Definition 2.50. We say that V ∈ G(n, m) is an approximate tangent space of E ⊂ Rn at a ∈ Rn

if
Θ∗m (E, a) > 0

and for all δ > 0

1 m

lim m
H {x ∈ E ∩ B̄(a, r) : dist(x − a, V ) > δ|x − a|} = 0.
r→0 r

If such a space exists, we denote it by Ta E or Tam E.

Remark 2.51. (a) For V ∈ G(n, m), a ∈ Rn , and δ > 0 let

C
Va,δ = {x ∈ Rn : dist(x − a, V ) > δ|x − a|}.

We then notice that V = Tam E if and only if Θ∗m (E, a) > 0 and Θ∗m (E ∩ Va,δ
C , a) = 0 for all
C = ∅ for δ ≥ 1 since dist(x − a, V ) ≤ |x − a|.
1 > δ > 0. Note that Va,δ

(b) If m = 1, the approximate tangent line Ta1 E is unique if exists, but for m ≥ 2 Tam E need
not be unique. However, for Hm -measurable sets E, with Hm (E) < ∞ and m ≥ 2, the
approximate tangent space Tam E is unique at Hm -a.e. point a ∈ E where such a space exists.
Fall 2016 35

(c) The definition above differs from (and is weaker than) that in [LY, 3.3.3] or [Si, 11.2] where
V ∈ G(n, m) is said to be the approximate tangent space of an Hm -measurable subset E ⊂ Rn
(with Hm (E ∩ K) < ∞ for every compact K ⊂ Rn ) at a ∈ Rn if
Z Z
m
lim f (y)dH (y) = f (y)dHm (y) ∀f ∈ C0 (Rn ),
λ→0+ η (E) V
a,λ

where ηa,λ : Rn → Rn is defined as ηa,λ (y) = (y − a)/λ for a, y ∈ Rn , λ > 0.

From Corollary 2.44 (see also Remark 2.51 (a)) we get:

Theorem 2.52. Let A ⊂ B ⊂ Rn be Hm -measurable with Hm (B) < ∞. Then for Hm -a.e. x ∈ A,
Txm A exists if and only if Txm B exists. Furthermore, if exist, they are equal Hm -a.e.

In particular, if E is m-rectifiable and Mi ’s are m-dimensional C 1 -submanifolds as in

Lemma 2.46, then at Hm -a.e. x ∈ E ∩ Mi the approximate tangent space of E is the same as
the usual tangent space of Mi .
The following theorem characterizes rectifiable sets in terms of approximate tangent spaces; see
[Ma, Chapter 15]. (This might be discussed in the home work classes.)

Theorem 2.53. Let E ⊂ Rn be Hm -measurable with Hm (E) < ∞. Then E is m-rectifiable if and
only if E has the approximate tangent space Ta E ∈ G(n, m) for Hm -a.e. a ∈ E.

As a corollary, we have a characterization of purely unrectifiability.

Lemma 2.54. Let E ⊂ Rn be Hm -measurable with Hm (E) < ∞. Then E is purely m-unrectifiable
if and only if the set of those points a ∈ E for which Tam E exists is of Hm -measure zero.

Another deep characterization of purely unrectifiable sets is the following Besicovitch-Federer

structure theorem.

Theorem 2.55. Let Q be a countable union of sets with finite Mm -measure. Then Q is purely
m-unrectifiable if and only if Hm (PV Q) = 0 for almost all V ∈ G(n, m). Here PV : Rn → V is
the orthogonal projection and ”almost all” refers to a natural probability Radon measure γn,m on
G(n, m).

For the proof; see e.g. [Ma, Theorem 18.1]. Remark: There is a natural probability Radon
measure γn,m on G(n, m) that can be obtained from the general theory of Haar measures. Indeed,
the group O(n) of orthogonal linear maps Rn → Rn is compact and hence there exists a unique
invariant Radon measure (Haar measure) θn such that θn (O(n)) = 1 and



θn (A) = θn {gh : h ∈ A} = θn {hg : h ∈ A}

for all A ⊂ O(n) and g ∈ O(n). The measure γn,m is then obtained by fixing V ∈ G(n, m) and
setting

γn,m (A) = θn {g : gV ∈ A} , A ⊂ G(n, m).
Being uniformly distributed γn,m is independent of the choice of V .
Suppose that E ⊂ Rn is (countably) m-rectifiable. Theorem 2.53 enables us to define the
gradient ∇E f of a Lipschitz function f : Rn → R at Hm -a.e. x ∈ E as
m
X
(2.56) ∇E f (x) = ∂vi f (x)vi ,
i=1
36 Geometric Measure Theory

where (v1 , . . . , vm ) is an orthonormal basis of Txm E and ∂vi f (x) denotes the directional derivative
of f in the direction vi . Note that we can write
∞
G
E = E0 ⊔ Ej ,
j=1

where Hm (E0 ) = 0 and Ej ⊂ Mj , with Mj an m-dimensional C 1 -submanifold of Rn . Then

∇E f (x) = ∇Mj f (x) whenever x ∈ Ej and f |Mj is differentiable at x (which holds Hm -a.e. in Mj
by Rademacher’s theorem).
Having defined ∇E f (x), we can define the linear map dE fx : Txm E → R by

dE fx (v) = hv, ∇E f (x)i, v ∈ Txm E,

at all points where Txm E and ∇E f (x) exist. Above h·, ·i is the standard inner product in Rn .
If f = (f1 , . . . , fN ) : Rn → RN is Lipschitz, we define a linear map dE fx : Txm E → RN by
N
X
E
d fx (v) = hv, ∇E fj (x)iej ,
j=1

where e1 , . . . , eN is the standard basis of RN . If N ≥ m, we define the Jacobian of f , denoted by

JfE (x), for Hm -a.e. x ∈ E by
q
(2.57) JfE (x) = det(dE fx )∗ ◦ dE fx .

Then we have the general area formula

Z Z


(2.58) JfE dHm = H0 A ∩ f −1 (y) dHm (y)
A RN

for every Hm -measurable A ⊂ E. Similarly, in the case N < m, we can define

q
JfE (x) = det(dE fx ) ◦ (dE fx )∗

and obtain the general co-area formula

Z Z
E m


Jf (x)dH (x) = Hm−N A ∩ f −1 (y) dHN (y)
A RN

for every Hm -measurable set. The following theorem will be useful in studying ”slices” of currents.
Theorem 2.59. Let E ⊂ Rn be m-rectifiable and f : Rn → R Lipschitz. Then for m1 -a.e. t ∈ R,
(1) Et := f −1 (t) ∩ E is (m − 1)-rectifiable and

(2) for Hm−1 -a.e. x ∈ Et , tangent spaces Txm−1 Et and Txm E exist, Txm−1 ⊂ Txm E, and

Txm E = {y + λ∇E f (x) : y ∈ Txm−1 Et , λ ∈ R}.

(3) For every nonnegative Hm -measurable g : E → R, we have (the co-area formula)

Z ∞Z Z
m−1
g dH dt = |∇E f |g dHm .
−∞ Et E
Fall 2016 37

For the proof; see [Si, p. 68-69 and 28.1]. Here we just sketch the proof:
The finiteness of Hm−1 Et for a.e. t ∈ R follows from Lemma 2.25 (iv). We can write
∞
G
E = E0 ⊔ Ej ,
j=1

where Hm (E0 ) =
0 and Ej ⊂ Mj , with Mj an m-dimensional C 1 -submanifold of Rn . Then

Hm−1 E0 ∩ f −1 (t) = 0 for a.e. t ∈ R. Hence it is enough to prove the claims for E = M ,
where M is an m-dimensional C 1 -submanifold of Rn , with Hm (M ) < ∞. Applying the implicit
function theorem (and using local coordinates), we may assume that M ⊂ Rm , with mm (M ) < ∞.
Rademacher’s theorem and Whitney extension theorem imply that, for every ε > 0, there exists
gε ∈ C 1 such that

mm ({x ∈ M : f (x) 6= gε (x) or ∇f (x) 6= ∇gε (x)}) < ε.

Applying this with ε = 1/i, i ∈ N, the problems are reduced to the case f ∈ C 1 . Sard’s theorem
implies that
m1 ({f (x) : |∇f (x)| = 0}) = 0.
Thus we may assume that ∇f (x) 6= 0 for every x ∈ M . Now the implicit function theorem implies
that the level sets Mt = {x ∈ M : f (x) = t} are locally (m − 1)-dimensional C 1 submanifolds, hence
(m − 1)-rectifiable. This proves (1). The claim (2) follows from the facts that ∇M f (x) ∈ Txm M
and ∇M f (x) ⊥ Txm−1 Mt . Finally, (3) is a generalization of the co-area formula.

3 Varifolds
From Wikipedia: Varifolds were first introduced by L.C. Young in 1951, under the name ”gen-
eralized surfaces”. Frederick Almgren slightly modified the definition in his mimeographed notes
(Almgren 1965) and coined the name varifold: he wanted to emphasize that these objects are sub-
stitutes for ordinary manifolds in problems of the calculus of variations. The modern approach to
the theory was based on Almgren’s notes and laid down by William Allard (Allard 1972).
Varifolds can be interpreted as measure-theoretic generalizations of smooth manifolds and they
generalize the idea of rectifiable currents.

3.1 Basic definitions

We start with introducing a metric (and hence a topology) on the Grassmannian space

G(n, m) = {V ⊂ Rn : V m-dimensional subspace of Rn }.

For V, W ∈ G(n, m), define

d(V, W ) = kPV − PW k = sup{|PV x − PW x| : x ∈ Rn , |x| = 1},

where PV : Rn → V is the orthogonal projection onto V . With this metric G(n, m) is a compact
metric space.

Definition 3.2. Let U ⊂ Rn be open and 0 ≤ m ≤ n integers. A Radon (outer) measure on

U × G(n, m) is called an m-dimensional varifold (or m-varifold) in U . The set of m-dimensional
varifolds in U is denoted by Vm (U ).
38 Geometric Measure Theory

Hence

Vm (U ) = {V : V a Radon outer measure on U × G(n, m)}

= {µ : µ Borel regular outer measure on U × G(n, m),
µ(K × G(n, m)) < ∞ ∀ compact K ⊂ U }.

We equip Vm (U ) with the weak topology (the following is just the rephrase of Definition 1.65 for
Radon outer measures):

Definition 3.3. The sequence Vi ∈ Vm (U ) is said to converge to V ∈ Vm (U ) (as varifolds), denoted

by Vi → V , if Vi ⇀ V as Radon (outer) measures, i.e.
Z Z
f dVi → f dV ∀f ∈ C0 (U × G(n, m)).
U ×G(n,m) U ×G(n,m)

Definition 3.4. For each V ∈ Vm (U ) we define the measure kV k and its m-dimensional density
d(V, ·) in U by setting


kV k(A) = V A × G(n, m) for Borel sets A ⊂ U,


kV k B̄(a, r)
d(V, a, r) = , r > 0,
ωm r m
d(V, a) = lim d(V, a, r) for a ∈ U if the limit exists.
r→0

The measure kV k is also called the weight (measure) of V and denoted by µV . The mass of V is
defined as MV = kV k(U ).

We abbreviate
Gn,m (U ) = U × G(n, m), Gn,m = Gn,m (Rn ).

Example 3.5. Let E ⊂ Rn be an Hm -measurable m-rectifiable set. Then E has the approximate
tangent space Txm E ∈ G(n, m) for Hm -a.e. x ∈ E. Define


VE (A) = Hm {x ∈ E : (x, Txm E) ∈ A}

for A ⊂ Gn,m . Then VE is an m-varifold, kVE k = Hm xE, and MVE = Hm (E). Moreover,
Z Z
f dVE = f (x, Txm E) dHm (x)
Gn,m E

for all f ∈ C0 (Gn,m ).

Definition 3.6. Let E and Ẽ be Hm -measurable and (countably) m-rectifiable subsets of Rn , and
let θ (resp. θ̃) be nonnegative and locally Hm -integrable in E (resp. Ẽ). We say that (E, θ) and
(Ẽ, θ̃) are equivalent if


Hm (E \ Ẽ) ∪ (Ẽ \ E) = 0

and θ = θ̃ Hm -a.e. in E ∩ Ẽ. A (countably) rectifiable m-varifold VE,θ = V (E, θ) is the equivalence
class of a pair (E, θ) as above and (E, θ) is called a representative for V . If θ is integer valued,
V (E, θ) is called an integer multiplicity rectifiable m-varifold, or briefly an integer m-varifold.
Fall 2016 39

We adopt the convention that θ ≡ 0 in Rn \E. Associated to a rectifiable m-varifold V = V (E, θ)

there is a Radon measure µV , called the weight measure of V , defined by

(3.7) µV = Hm xθ,

that is Z
µV (A) = θ dHm
A∩E

for Hm -measurable sets A. The mass of V = V (E, θ) is

Z
n
MV = µV (R ) = θ dHm = (Hm xθ)(Rn ).
E∩Rn

Every countably rectifiable m-varifold V (E, θ) induces an m-varifold VE,θ by

Z
VE,θ (A) = θ dHm (x), A ∈ Gn,m .
{x∈E : (x,Txm E)∈A}

3.8 First and second variation formulae

Next we will study how the mass MV of an m-varifold V ∈ Vm (U ) (resp. of a rectifiable m-varifold
V = (E, θ)) behaves under a perturbation by a 1-parameter family of diffeomorphisms. To get
an idea, let us consider first (a less abstract setting of) an m-dimensional C 1 -smooth submanifold
M ⊂ Rn .

Remark 3.9. Let M be an m-dimensional C 1 -smooth submanifold of Rn . For every point x ∈ M

there exist an open neighborhood A ⊂ Rn of x and a C 1 -diffeomorphism ϕ : A → A′ onto an
open set A′ ⊂ Rn such that ϕ(A ∩ M ) is an open subset of Rm × {0} ⊂ Rm × Rn−m . Note that
Tx M = (dϕx )−1 Rm .

Let then U ⊂ Rn be open such that U ∩ M 6= ∅ and Hm (C ∩ M ) < ∞ for every compact C ⊂ U .
Let {φt }, −1 < t < 1, be a 1-parameter family of diffeomorphisms φt : U → U such that

φ : (−1, 1) × U → U, φ(t, x) = φt (x), is C 2 ,

(3.10) φ0 (x) = x ∀x ∈ U, and
φt (x) = x ∀x ∈ U \ K and t ∈ (−1, 1),

for some compact K ⊂ U . Define mappings X = (X 1 , . . . , X n ) : U → Rn and Z =

(Z 1 , . . . , Z n ) : U → Rn by

∂φ(t, x) ∂ 2 φ(t, x)
(3.11) X(x) = and Z(x) = .
∂t |t=0 ∂t2 |t=0

Then

t2
(3.12) φt (x) = x + tX(x) + Z(x) + O(t3 ),
2

where O(t3 ) ∈ Rn , with |O(t3 )| ≤ c|t|3 . Since φt (x) = x for x ∈ U \ K, the maps X and Z are
compactly supported.
40 Geometric Measure Theory

Definition 3.13. Let Mt = φt (M ∩ K). The first and second variations of M (with respect to a
1-parameter family {φt }) are defined as
d m d2 m
H (Mt )|t=0 and H (Mt )|t=0 ,
dt dt2
respectively.
By the area formula
Z
m m


H (Mt ) = H φt (M ∩ K) = Jψt dHm ,
M ∩K
where ψt = φt |M ∩ U . Since we can change the order of integration and differentiation, the
computation of the first and second variations reduces to calculating
∂ ∂2
Jψt |t=0 and J .
∂t ∂t2 ψt |t=0
For that purpose, let us fix orthonormal bases τ1 , . . . , τm of Tx M for x ∈ M and e1 , . . . , en of Rn .
We define the (induced) linear map dψt,x : Tx M → Rn of ψt at x ∈ M by
dψt,x (τ ) = ∂τ φt (x) = ∂τ ψt (x), τ ∈ Tx M.
By (3.12), we have
t2
dψt,x (τ ) = τ + t∂τ X(x) + ∂τ Z(x) + O(t3 ).
2
Writing the basis vectors τj , j = 1, . . . , m, as
n
X
τj = τji ei ,
i=1

we can express the matrix (aij )n×m of dψt,x w.r.t. bases τ1 , . . . , τm of Tx M, x ∈ M , and e1 , . . . , en
of Rn as
t2
aij = τji + t∂τj X i + ∂τj Z i + O(t3 ).
2
Consequently, the matrix of (dψt,x )∗ ◦ (dψt,x ) is (bij )m×m , where
n
X
bij = aki akj
k=1
 

1

= δij + t hτi , ∂τj Xi + hτj , ∂τi Xi + t2 hτi , ∂τj Zi + hτj , ∂τi Zi + h∂τi X, ∂τj Xi + O(t3 ).
2
Next we apply the formula



det I + tA + t2 B = exp log det I + tA + t2 B



= exp Tr log I + tA + t2 B
 
2 1 2 2 3
= exp Tr tA + t B − (tA + t B) + O(t )
2
 
2 1 2 2 3 4 2

3
= exp t Tr A + t Tr B − Tr t A + 2t AB + t B + O(t )
2
 
2 1 2 2 3
= exp t Tr A + t Tr B − t Tr A + O(t )
2
1 1
= 1 + t Tr A + t2 Tr B − t2 Tr A2 + t2 (Tr A)2 + O(t3 )
2 2
Fall 2016 41

for symmetric square matrices I = (δij ) = the identity matrix, A = (Aij ), and B = (Bij ), where

Aij = hτi , ∂τj Xi + hτj , ∂τi Xi = Aji and

1

Bij = hτi , ∂τj Zi + hτj , ∂τi Zi + h∂τi X, ∂τj Xi,
2
to obtain

Jψ2t (x) = det(dψt,x )∗ ◦ (dψt,x ) = det(bij )

m m  m
!2
X X  X
= 1 + 2t hτi , ∂τi Xi + t2
hτi , ∂τi Zi + |∂τi X|2 + 2t2 hτi , ∂τi Xi
i=1 i=1 i=1
m
1 X
2
− t2 hτi , ∂τj Xi + hτj , ∂τi Xi + O(t3 )
2
i,j=1
m
X
= 1 + 2t divM X + t divM Z + t2 2
|∂τi X|2 + 2t2 (divM X)2
i=1
m
X m
X
− t2 hτi , ∂τj Xi2 − t2 hτi , ∂τj Xihτj , ∂τi Xi + O(t3 )
i,j=1 i,j=1
 
m m
X
⊥ 2 X
= 1 + 2t divM X + t2 divM Z + 2 (divM X)2 + ∂τi X − hτi , ∂τj Xihτj , ∂τi Xi
i=1 i,j=1
3
+ O(t ),

where
m
X

⊥
∂τi X = ∂τi X − hτj , ∂τi Xiτj
j=1

is the normal component of ∂τi X (normal to M ). Above divM X is the divergence of X (at x ∈ M )
with respect to M defined as
Xm
divM X = hτi , ∂τi Xi.
i=1

Finally, using
√ 1 1
1 + s = 1 + s − s2 + O(s3 ),
2 8
we get
 
m m
t2 divM Z + 2 (divM X)2 +
X
⊥ 2 X
Jψt (x) = 1 + t divM X + ∂τi X − hτi , ∂τj Xihτj , ∂τi Xi
2
i=1 i,j=1

t2
− (2 divM X)2 + O(t3 )
8  
2 m m
t  2
X
⊥ 2 X
= 1 + t divM X + divM Z + (divM X) + ∂τi X − hτi , ∂τj Xihτj , ∂τi Xi
2
i=1 i,j=1
3
+ O(t ).
42 Geometric Measure Theory

Hence
∂
J = divM X,
∂t ψt |t=0
and therefore, by the area formula, we obtain the first variation formula
Z
d m ∂
H (Mt )|t=0 = J dHm
dt M ∩K ∂t ψt |t=0
Z
(3.14) = divM X dHm
Z M ∩K

= divM X dHm ,
M

where the last equality holds since X ≡ 0 in M \ K. Similarly, we get the second variation formula

d2 m
H (Mt )|t=0
dt2
Z  m
X m
X 

⊥ 2
(3.15) = divM Z + (divM X)2 + ∂τi X − hτi , ∂τj Xihτj , ∂τi Xi dHm
M i=1 i,j=1

Definition 3.16. An m-dimensional C 1 -smooth submanifold M ⊂ Rn is stationary in an open set

U ⊂ Rn if Hm (M ∩ C) < ∞ for every compact C ⊂ U and if
d m
H (Mt )|t=0 = 0
dt
for Mt = φt (M ∩ K) whenever φt and K are as in (3.10).

By the first variation formula (3.14), M is stationary in U if and only if

Z
divM X dHm = 0
M

for every C 1 -smooth X : U → Rn with compact support in U . Indeed, every such X generates a
1-parameter family of C 2 -diffeomorphisms {φt } satisfying (3.10), with K = supp X, as the flow of
X. More precisely, for every x ∈ U , t 7→ φt (x) is the integral curve of X starting at x, that is
φ0 (x) = x and
d

φt (x) = X φt (x) .
dt
Remark 3.17. If M is an m-dimensional C 2 -smooth submanifold of Rn , m < n, and U ⊂ Rn is
open such that Ū ∩ M is compact, then M is stationary in U if and only if H ≡ 0 in M ∩ U , where
H is the mean curvature vector of M . The mean curvature of M will be discussed in a home work
session.

Next we will generalize the first variation formula for rectifiable m-varifolds. Let V = V (E, θ)
be a rectifiable m-varifold in an open set U ⊂ Rn . We suppose for simplicity that

(3.18) θ(x) ≥ 1

for Hm -a.e. x ∈ E. This restriction is made to avoid discussions on approximate tangent spaces
(and hence Jacobians) with respect to multiplicity θ. We conclude from Theorem 2.52 that Txm E
and Txm Ẽ exists and are equal for Hm -a.e. x ∈ E ∩ Ẽ if (Ẽ, θ̃) is another representative for V .
Fall 2016 43

Therefore we can define the approximate tangent space of V at x by setting Tx V = Txm E. Suppose
then that f : U → U ′ is a Lipschitz mapping to an open set U ′ ⊂ RN , N ≥ n, with the Jacobian
JfE defined in (2.57). We notice that JfE (x) = JfẼ (x) for Hm -a.e. x ∈ E ∩ Ẽ, and hence we may
denote it by JfV . By the general area formula (2.58), we have
Z Z Z Z !
X
(3.19) g JfE dHm = g(x) dHm = gH0 dHm
A f E x∈A∩f −1 (y) fE A∩f −1 (y)

for every nonnegative Hm -measurable g on E and Hm -measurable A ⊂ E. Clearly f E is an m-

rectifiable subset of U ′ . We assume, moreover, that f : U → U ′ is proper , that is f −1 K ⊂ U is
compact for every compact K ⊂ U ′ . Then we define θ ′ on U ′ by setting
X Z
′
θ (y) = θ(x) = θ dH0
E∩f −1 (y)
x∈E∩f −1

and the image (or push-forward)

f# V = V (f E, θ ′ ).
Since Z Z Z
θ ′ dHm = θ ′ dHm = θJfE dHm
K f E∩K E∩f −1 K

for every compact K ⊂ U ′,

we see that θ ′ is locally Hm -integrable in U ′ . Hence f# V is a rectifiable
m-varifold in U ′ with multiplicity θ ′ . Moreover,
Z Z
′ m
Mf# V = θ dH = JfE θ dHm .
fE E

Now we are ready to define the first variation of V . Let {φt } be a 1-parameter family of diffeomor-
phisms φt : U → U as in (3.10). We denote V xK = V (E ∩ K, θ|K), where K ⊂ U is the compact
set in (3.10). Then Z
Mφt# (V xK) = JφEt θ dHm
E∩K
and we can compute the first variation

d
M
dt φt# (V xK)|t=0
exactly as in the case of C 1 -submanifolds and obtain
Z
d
(3.20) Mφt# (V xK)|t=0 = divE X dµV ,
dt E

where X is as in (3.11) and divE X is the divergence of X with respect to E, defined as

m
X
divE X(x) = hτi , ∂τi X(x)i,
i=1

with τ1 , . . . , τm an orthonormal bases of Txm E.

As in the case of C 1 -submanifolds, we define
44 Geometric Measure Theory

Definition 3.21. A rectifiable m-varifold V = V (E, θ) is stationary in an open set U ⊂ Rn if

Z
divE X dµV = 0
E

for any C 1 -smooth X : U → Rn with compact support in U .

We also generalize the notion of mean curvature as follows:

Definition 3.22. Let V = V (E, θ) be a rectifiable m-varifold in an open set U ⊂ Rn . Suppose

H : E ∩ U → Rn is locally µV -integrable. We say that V = V (E, θ) has generalized mean curvature
H in U if Z Z
divE X dµV = − hX, Hi dµV
U U

whenever X : U → Rn is a C1 with compact support in U .

Hence a rectifiable m-varifold V = V (E, θ) is stationary in an open set U ⊂ Rn if and only if it

has zero generalized mean curvature in U .
Next we will introduce the variation of a (general) varifold. For that purpose, let U, U ′ ⊂ Rn be
open, V ∈ Vm (U ) an m-varifold in U , and suppose that f : U → U ′ is a C 1 -diffeomorphism. Recall
that an m-varifold in an open set U ⊂ Rn is a Radon (outer) measure on Gn,m (U ) = U × G(n, m).
First we define the push forward of V under f by setting for Borel sets B ⊂ Gn,m (U ′ )
Z
f# V (B) = Jf (x, E)dV (x, E),
F −1 (B)

where F : Gn,m (U ) → Gn,m (U ′ ) is defined by

F (x, E) = f (x), dfx E

and
∗

1/2
Jf (x, E) = det dfx |E ◦ dfx |E , (x, E) ∈ Gn,m (U ).
Note that dfx : Rn → Rn is an invertible linear map for all x ∈ U since f : U → U ′ is a C 1 -
diffeomorphism. In particular, dfx |E : E → dfx E ∈ G(n, m) is invertible. For a Borel set A ⊂ U ,
the restriction V xGn,m (A) is the Radon measure in Gn,m (U ) defined as



V xGn,m (A) (B) = V B ∩ Gn,m (A) , B ⊂ Gn,m (U ).

Definition 3.23. Let V be an m-varifold in an open set U ∈ Rn and let C01 (U, Rn ) be the space of
C 1 -mappings X : U → Rn with compact support in U . Then the first variation of V is the linear
functional δV : C01 (U, Rn ) → R,

d
δV (X) = M ,
dt φt# (V xGn,m (K))|t=0
where {φt } is a 1-parameter family of diffeomorphisms U → U associated to X ∈ C01 (U, Rn ) as in
(3.10) and (3.11), that is φ = φ(·, ·) is the flow of X.

Again, exactly the same computation as in smooth case gives

Z
(3.24) δV (X) = divS X(x) dV (x, S),
Gn,m (U )
Fall 2016 45

where, for any (x, S) ∈ G(n,m) (U ), divS X is the divergence of X with respect to S, defined as
m
X
divS X(x) = hτi , ∂τi X(x)i,
i=1

with τ1 , . . . , τm an orthonormal bases of S.

Definition 3.25. A varifold V ∈ Vm (U ) is said to be stationary if δV (X) = 0 for every X ∈
C01 (U, Rn ).

More generally, V is said to have locally bounded first variation if for each W ⋐ U there exists
a constant c < ∞ such that

|δV (X)| ≤ c sup|X| ∀X ∈ C01 (U, Rn ), with supp X ⊂ W.

U

Definition 3.26. For any V ∈ Vm (U ), we define the set function kδV k : P(U ) → [0, ∞] by

kδV k(U ′ ) = sup{δV (X) : X ∈ C01 (U, Rn ), |X| ≤ 1, supp X ⊂ U ′ }

for open sets U ′ ⊂ U , and then

kδV k(A) = inf{kδV k(U ′ ) : U ′ ⊂ U open, A ⊂ U ′ }

for A ⊂ U .

We note that kδV k is a metric outer measure. If V has locally bounded first variation, then
kδV k is a Radon measure by Theorem 1.31 and, moreover, by the general Riesz representation
theorem 1.34, there exists a kδV k-measurable mapping ηV : U → Sn−1 such that
Z
(3.27) δV (X) = − hX, ηV i dkδV k
U

for all X ∈ C01 (U, Rn ). [Use the Hahn-Banach theorem to extend δV : C01 (U, Rn ) to a linear func-
tional on C0 (U, Rn ) and remember the construction of the Radon measure µ in the proof of the
Riesz representation theorem to note that µ is, in fact, kδV k.]
Recall that the weight (measure) µV = kV k is defined as


µV (A) = V A × G(n, m)

for all Borel sets A ⊂ U . By the Radon-Nikodym theorem (see e.g. [Ma, 2.17], [Si, 4.7], or [Ho,
5.31]), the Radon-Nikodym derivative


kδV k B̄(x, r)
DµV kδV k(x) = lim

r→0 µV B̄(x, r)

exists for µV -a.e. x ∈ U and

Z Z Z
hX, ηV i dkδV k = hX, HV i dµV + hX, ηV i dσ,
U U U

where σ = kδV kxN, N = {x ∈ U : DµV kδV k(x) = ∞}, and

HV (x) = DµV kδV k(x)ηV (x).

46 Geometric Measure Theory

Hence we can write (3.27) as

Z
δV (X) = divS X(x) dV (x, S)
Gn,m (U )
Z Z
= − hX, HV i dµV − hX, ηV i dσ
ZU ZU
= − hX, HV i dµV − hX, ηV i dσ.
U N

We call HV the generalized mean curvature of V , N the generalized boundary of V , σ the generalized
boundary measure of V , and ηV |N the generalized unit co-normal of V .

Remark 3.28. If M ⊂ Rn is a C 2 -smooth m-dimensional submanifold with smooth boundary ∂M

and X ∈ C 1 (U, Rn ), with M ⊂ U, U ⊂ Rn open, then
Z Z Z
m m
divM X dH =− hX, HidH − hX, ηidHm−1 ,
M M ∂M

where H is the mean curvature (vector) of M and η the inward pointing unit co-normal of ∂M ,
that is, |η| ≡ 1, η is normal to ∂M , tangential to M , and points inwards to M .

The first variation formula is applied with certain specific choices of the vector field X. Most
importantly, we obtain the so-called monotonicity formula and its applications to the regularity
theory of varifolds. These will be discussed in a series of presentations in home work sessions. If
the time permits, we will return to these topics in context of currents.

4 Currents
In this Section we introduce and study some basic notions in the theory of currents which (like
varifolds and m-rectifiable varifolds) are kind of generalized surfaces.
Let us start with the following motivating example.

Example 4.1. Let M be a smooth oriented m-dimensional submanifold of Rn . We can integrate

smooth differential m-forms α (with compact support) over M and thus consider M as a linear
functional
[M ] : {smooth differential m-forms with compact support} → R,
Z
[M ](α) = α.
M

Currents are, by definition, such continuous linear functionals on the space of smooth differential
m-forms with compact support; see Definition 4.21.

4.2 m-vectors
In this subsection, we discuss briefly about m-vectors which are kind of ”products” of vectors.
Given v1 , v2 ∈ Rn , a geometric interpretation of the 2-vector v1 ∧ v2 is the oriented parallelogram
spanned by vectors v1 and v2 .
Fall 2016 47

(−v1 ) ∧ v2
v2
v1 ∧ v2 v2
v1 −v1

If v1 = λv2 , the parallelogram is degenerate, and we have v1 ∧ v2 = 0.

Similarly, for a 3-vector v1 ∧ v2 ∧ v3 can be interpreted as an oriented parallelepiped spanned
by vectors v1 , v2 , v3 ∈ Rn .

v3 v2
v1 ∧ v2
v1 ∧ v3 v1
v1 ∧ v2 ∧ v3

Formally, the quickest (but not necessarily the most elegant) way to define the vector space of
m-vectors ^
n
m (R ), m = 0, . . . , n,
is as the space of all (real) linear combinations
X
a ···im ei1 ∧ · · · ∧ eim ,
| i1{z }
1≤i1 <···<im ≤n
∈R

standard (ordered) basis of Rn . The basis (m-)vectors ei1 ∧ · · · ∧ eim ,

where (e1 , . . . , en ) is the V
1 ≤ i1 < · · · < im ≤ n, of m (Rn ) can be defined as the strictly increasing sequences i1 < · · · < im .
Thus we may identify ei1 ∧ · · · ∧ eim with the m-tuple (i1 , . . . , im ) if i1 < · · · < im . Hence
 
^
n n
dim m (R ) = .
m
If m = n, e1 ∧ · · · ∧ en is the only basis vector, and therefore
^
dim n (Rn ) = 1.
Hence we may identify ^
n
n (R ) = R.
Similarly, ^
n
1 (R ) = span(e1 , . . . , en ) = Rn .
We also define ^ ^
n n
0 (R )=R and k (R ) = {0} for k > n.
We want to ”multiply” k-vectors and m-vectors and hence to give ei1 ∧ · · · ∧ eim a meaning as a
”product”Vof vectors
V ei1 , . . . , eim
V. Since the desired properties of the wedge product (or exterior
product) k (Rn ) × m (Rn ) → k+m (Rn ) are
48 Geometric Measure Theory

(a) multilinearity:
^ ^
n n
(au + bv) ∧ cw = ac(u ∧ w) + bc(v ∧ w), a, b, c ∈ R, u, v ∈ k (R ), w ∈ m (R );
^ ^
n n
au ∧ (bv + cw) = ab(u ∧ v) + ac(u ∧ w), a, b, c ∈ R, u ∈ k (R ), v, w ∈ m (R ),

(b) associativity:
u ∧ (v ∧ w) = (u ∧ v) ∧ w, and

(c) anticommutativity:
^ ^
u ∧ v = (−1)km v ∧ u, u∈ k (R
n
), v ∈ m (R
n
),

it is enough to define wedge products

ei1 ∧ ei2 ∧ · · · ∧ eim ∧ ej = (ei1 ∧ ei2 ∧ · · · ∧ eim ) ∧ ej

V
for 1 ≤ i1 < · · · < im and j ∈ {1, . . . , n}. We have already defined ei ∧ej ∈ 2 (Rn ) for 1 ≤ i < j ≤ n
(as the oriented pair (i, j) or the positively oriented unit square in Rn spanned by ei and ej ). We
define
ej ∧ ei = −ei ∧ ej for i < j,
and
ei ∧ ei = 0.
We also have defined already ei ∧ ej ∧ ek if i < j < k , i.e. the positively oriented unit cube spanned
by ei , ej , ek (or the oriented 3-tuple (i, j, k)). So, we define for i < j < k

ei ∧ ek ∧ ej = ei ∧ (ek ∧ ej )
= ei ∧ (−ej ∧ ek )
= −ei ∧ ej ∧ ek ,
ek ∧ ei ∧ ej = −ei ∧ ek ∧ ej )
= −(−ei ∧ ej ∧ ek )
= ei ∧ ej ∧ ek ,

and so on. Also ei ∧ eiV∧ ek = 0 if i = j V

or i = k, or j =Vk. Continuing this way we have the
wedge product u ∧ v ∈ k+m (R ) for u ∈ k (Rn ) and v ∈ m (Rn ). If k + m > n, u ∧ v = 0 and
V
n
n
k+m (R ) = {0}. Let us summarize the discussion above:

Proposition 4.3. The wedge product has the properties:

(a) multilinearity:

(4.4) (au + bv) ∧ cw = ac(u ∧ w) + bc(v ∧ w)

V V
for a, b, c ∈ R, u, v ∈ k (Rn ), w ∈ m (Rn ),
(b) associativity:

(4.5) u ∧ (v ∧ w) = (u ∧ v) ∧ w,

and
Fall 2016 49

(c) anticommutativity:

(4.6) u ∧ v = (−1)km v ∧ u
V n ),
V n ).
for u ∈ k (R v∈ m (R
V n
V thenm-vectors ei1 ∧ · · · ∧ eim , 1 ≤ i1 < · · · < im ≤ n, form a basis of m (R ), we may
Since
equip m (R ) with an inner product h·, ·i such that these m-vectors form an orthonormal basis.
More precisely, denote
^
(n, m) = {(i1 , . . . , im ) ∈ Nm : 1 ≤ i1 < · · · < im ≤ n}
V
and eI = ei1 ∧ · · · ∧ eim for I = (i1 , . . . , im ) ∈ (n, m). Then
* +
X X X
(4.7) aI eI , bJ eJ = a I bI .
V V V
I∈ (n,m) J∈ (n,m) I∈ (n,m)

n
In fact, identifying m (Rn ) and R(m) isomorphically, i.e. by identifying the basis vectors eI , I ∈
V
n
(n, m), with the standard basis vectors of R(m) , the inner product in (4.7) becomes the standard
V
n
inner product in R(m) .
We define the norm
p
(4.8) |v| = hv, vi
V
for v ∈ m (Rn ). If v is a simple m-vector, that is

v = v1 ∧ · · · ∧ v m

for some vectors v1 , . . . , vm ∈ Rn , then

(4.9) |v| = |v1 ∧ · · · ∧ vm |

is the (m-dimensional) volume of the parallelepiped spanned by v1 , . . . , vm . In particular,

|v1 ∧ · · · ∧ vm | = 0

if and only if v1 , . . . , vm are linearly dependent.

4.10 m-covectors
V1 V1
Let (Rn ) denote the dual of Rn (thus (Rn ) = (Rn )∗ ) and let dx1 , . . . , dxn denote the dual
basis of e1 , . . . , en . That is, (
i 1, if i = j;
dx (ej ) = δji =
0, 6 j.
if i =
Then we define the vector space
^ ^ ^

m
(4.11) (Rn ) = m
1
(Rn )

as above by replacing ei with dxi . The elements

X X
(4.12) α= ai1 ···im dxi1 ∧ · · · ∧ dxim = aI dxI
i1 <···<im
V
I∈ (n,m)
50 Geometric Measure Theory

Vm V
of (Rn ) are called m-covectors. The space m (Rn ) has the induced) inner product
* +
X X X
aI dxI , bJ dxJ = a I bI
V V V
I∈ (n,m) J∈ (n,m) I∈ (n,m)

i1 im
Vm n dx ∧ · · · ∧ dx , 1 ≤ i1 V
such that the m-covectors < · · · < im ≤ n, form an
n ). Again we have
V0 orthonormal
basis. Moreover,
Vn n V1 n (R ) is the
Vm n dual vector space of m (R (Rn ) = R =
(R ), n
(R ) = R , and (R ) = {0} if m > n.

4.13 m-vector fields, m-covector fields, and smooth differential m-forms

V
Definition 4.14. If U ⊂ Rn , the mappings U → m (R
n ),

X
x 7→ aI (x)eI ,
V
I∈ (n,m)
Vm
and U → (Rn ), X
x 7→ aI (x)dxI ,
V
I∈ (n,m)

are called m-vector fields and m-covector fields in U , respectively.

V
Definition 4.15. The mappings U → m (Rn ),
X
x 7→ aI (x)dxI ,
V
I∈ (n,m)

are also called (differential) m-forms in U .

If U ⊂ Rn is open and X
α= αI (x)dxI ,
V
I∈ (n,m)

where the functions αI are C ∞ -smooth, we say that α is a C ∞ -smooth differential m-form in U .
The spaceVof all C ∞ -smooth differential m-forms in U will be denoted by Am (U ).
Since 0 (Rn ) = R,V we have A0 (U ) = C ∞ (U, R). If f : U → R is C ∞ , i.e. f ∈ A0 (U ), its
differential df : U → 1 (Rn ) is a C ∞ -smooth differential 1-form such that at a point x ∈ U ,
df (x) : Rn → R is the linear mapping defined by

df (x)v = h∇f (x), vi, v ∈ Rn .

Since, on the other hand,

 
n
X
dxi (v) = dxi  vj ej  = vi
j=1

and hence
* n
+
X ∂f (x)
df (x)v = h∇f (x), vi = ei , v
∂xi
i=1
n n
X ∂f (x) X ∂f (x) i
= vi = dx (v),
∂xi ∂xi
i=1 i=1
Fall 2016 51

we notice that
n
X ∂f i
(4.16) df = dx .
∂xi
i=1

Moreover, dxi is the differential of the ith coordinate function x 7→ xi .

Definition 4.17. Let X

α= αI dxI
V
I∈ (n,m)

be a C ∞ -smooth differential m-form. The exterior derivative of α is the (m + 1)-form

n
X
I
X X ∂αI
dα = dαI ∧ dx = dxi ∧ dxI .
∂xi
I∈ (n,m) i=1
V V
I∈ (n,m)

In particular, df is the exterior derivative of a 0-form f .

Using the facts that

∂ 2 αI ∂ 2 αI
=
∂xi ∂xj ∂xj ∂xi
and
dxi ∧ dxj = −dxj ∧ dxi ,
we obtain
d2 α = d(dα) = 0.

Definition 4.18. Let U ⊂ Rn and V ⊂ Rd be open sets and f = (f 1 , . . . , f d ) : U → V a C ∞ -smooth

mapping. The pull-back of a differential m-form α in V ,
X
α= αi1 ···im dxi1 ∧ · · · ∧ dxim ,
1≤i1 <···<im ≤d

is the differential m-form f ∗ α in U defined by

X

f ∗α = αi1 ···im ◦ f df i1 ∧ · · · ∧ df im ,
1≤i1 <···<im ≤d

where
n
X ∂f j
df j = dxi .
∂xi
i=1

Notice that we do not require α being smooth. The pull-back and the exterior derivative
commute, that is

(4.19) f ∗ (dα) = df ∗ α

for smooth α.
Let D m (U ) ⊂ Am (U ) denote the space of all C ∞ -smooth differential m-forms in U with compact
support, that is, if X
α= αI dxI ,
I
52 Geometric Measure Theory

then each αI is C ∞ -smooth and there exists a compact set K ⊂ U such that supp αI ⊂ K for every I.
We endow D m (U ) with the locally convex topology by saying that a sequence αk ∈ D m (U ), k ∈ N,
X
αk = αkI dxI
I

converges to X
α= αI dxI ∈ D m (U )
I
if there exists a compact set K ⊂ U such that
[
supp αk := supp αkI ⊂ K ∀k
I

and
∂ |J| αkI ∂ |J| αI
→
∂xJ ∂xJ
uniformly as k → ∞ for every multi-index J = i1 · · · iℓ .

4.20 m-currents; definition and basic notions

Definition 4.21. An m-current in an open set U ⊂ Rn is a continuous (w.r.t. the locally convex
topology described above) linear functional
T : D m (U ) → R.
The space of m-currents in U is denoted by Dm (U ).
Definition 4.22. The boundary of an m-current T ∈ Dm (U ) is the (m − 1)-current ∂T ∈ Dm−1 (U )
defined by
∂T (ω) = T (dω)
for all ω ∈ D m−1 (U ). Since d2 = 0, we have ∂ 2 T = ∂(∂T ) = 0.
Example 4.23. Let M ⊂ Rn be a smooth oriented m-dimensional submanifold with smooth
boundary ∂M . Let U ⊂ Rn be an open set such that M ∪ ∂M ⊂ U . Then M and ∂M define
currents
Z
[M ] ∈ Dm (U ) : [M ](ω) = ω, ω ∈ D m (U ),
M
Z
[∂M ] ∈ Dm−1 (U ) : [∂M ](α) = α, α ∈ D m−1 (U ).
∂M

By Stokes’ theorem Z Z
[∂M ](α) = α= dα = [M ](dα)
∂M M
for all α ∈ D m−1 (U ). Hence ∂[M ] = [∂M ].
Remark 4.24. For the definitions of the integrals
Z Z
dα and α
M ∂M

we refer to literature on differential geometry (e.g. [Lee], [Ho2]). However, since we will later
integrate
R differential m-forms over ”oriented” m-rectifiable sets, we will explain below the meaning
of M ω even in this more general setting.
Fall 2016 53

Let V ∈ G(n, m) and let L : Rm → V be a linear isometric isomorphism (the restriction to Rm

of an orthogonal mapping O : Rn → Rn ). Now
^ X
m (V )={ ai1 ···im vi1 ∧ · · · ∧ vim : vij ∈ V }
= {λ(Le1 ) ∧ · · · ∧ (Lem ) : λ ∈ R}
V V
is 1-dimensional. Hence, if v ∈ m (V ), with |v| = 1, then the only other w ∈ m (V ), with |w| = 1,
is w = −v. Let Tx M be the tangent space of M at x (here, first, M is an oriented m-dimensional
n
smooth submanifold ofVR ). Then M being ”oriented” means that we have chosen, for every x ∈ M ,
~ (x) ∈ ~ ~
an m-vector M m (Tx M ) such that |M (x)| ≡ 1 and x 7→ M (x) ∈ G(n, m) is continuous. We
then define Z Z
ω= ~ (x), ω(x)idHm (x).
hM
M M


Here hM~ (x), ω(x)i = ω M ~ (x) (x) ∈ R is the ”dual pairing”. In the general case, M will be
m-rectifiable, Txm M the approximate tangent space of M , and M~ will be replaced by a ”Borel
orientation”.

Example 4.25. 1. m = 0: For a ∈ Rn , let [a] = δa ∈ D0 (Rn ),

[a](ϕ) = ϕ(a), ϕ ∈ D 0 (Rn ).

2. m = 1: Let Γ ⊂ Rn be a C 1 -curve, ~Γ(x) the unit tangent vector to Γ such that x 7→ ~Γ(x) is
continuous. Then Z
[Γ](ω) = h~Γ(x), ω(x)idH1 (x), ω ∈ D 1 (Rn ).
Γ

3. m = n: Let U ⊂ Rn be open with smooth boundary ∂U . Then

Z
[U ](ω) = he1 ∧ · · · ∧ en , ω(x)idmn (x), ω ∈ D n (Rn ).
U

4. Let Q = [0, 1] × [0, 1] and let T ∈ D1 (R2 ) be defined as

Z
T (ω) = he1 , ω(x)idm2 (x), ω ∈ D 1 (R2 ).
Q

Writing ω = ω1 dx1 + ω2 dx2 , we see that

Z
T (ω) = he1 , ω(x)idm2 (x)
Q
Z Z
1
= ω1 he1 , dx i dm2 (x) + ω2 he1 , dx2 i dm2 (x)
Q | {z } Q | {z }
≡1 ≡0
Z
= ω1 dm2 (x).
Q
54 Geometric Measure Theory

On the other, ∂T ∈ D0 (R2 ), so it operates on smooth functions (0-forms) as

 
∂ϕ 1 ∂ϕ 2
∂T (ϕ) = T (dϕ) = T dx + dx
∂x1 ∂x2
Z
∂ϕ
= dm2
Q ∂x1
Z 1Z 1
∂ϕ
= dx1 dx2
0 0 ∂x1
Z 1
= (ϕ(1, x2 ) − ϕ(0, x2 )) dx2
Z0 Z
= ϕ− ϕ,
I1 I0
   
where I0 and I1 are the line segments I0 = (0, 0), (0, 1) , I1 = (1, 0), (1, 1) . So,

∂T = H1 xI1 − H1 xI0 .

∂T ∂T

I0 I1

Notice that T is a 1-dimensional current but its ”support” is 2-dimensional.

Definition 4.26. We define the mass of T ∈ Dm (U ) by

(4.27) M(T ) = sup{T (ω) : ω ∈ D m (U ), |ω(x)| ≤ 1 ∀x ∈ U }.

If W ⊂ U is open, we define

MW (T ) = sup{T (ω) : ω ∈ D m (U ), |ω(x)| ≤ 1 ∀x ∈ W, supp ω ⊂ W }.

Remark 4.28. (a) There is another slightlyVdifferent definition of the mass: Indeed, one first
define the co-mass of an m-covector η ∈ m (Rn ) by
^
n
kηk = sup{hζ, ηi : |ζ| ≤ 1, ζ ∈ m (R ) simple}

and then
M (T ) = sup{T (ω) : ω ∈ D m (U ), kω(x)k ≤ 1 ∀x ∈ U }.
Since kω(x)k ≤ |ω(x)|, it is possible that M (T ) > M(T ).

(b) Suppose that L : D m (U ) → R is a linear map that is continuous with respect to the norm
topology of D m (U ), that is L(ωi ) → L(ω) if ωi , ω ∈ D m (U ), with |ωi − ω| → 0. Since the
convergence in the locally convex topology of D m (U ) implies the convergence in the norm
topology, we notice that L is continuous with respect to the locally convex topology, too.
Fall 2016 55

Hence L ∈ Dm (U ). On the other hand, since the norm topology of D m (U ) is coarser than
the locally convex topology, there can be m-currents with infinite mass. In other words, each
m-current T ∈ Dm (U ) is a linear mapping T : Dm (U ) → R that is continuous with respect to
the locally convex topology but not necessarily with respect to the norm topology of D m (U ).
(c) Since (D m (U ), | · |) is a normed space, its dual space {T ∈ Dm (U ) : M(T ) < ∞} is a Banach
space.
Applying the Hahn-Banach theorem and the Riesz representation theorem we obtain the fol-
lowing:
Theorem 4.29. Suppose that T ∈ Dm (U ) such that MW (T ) < ∞ for every WV⋐ U . Then there
exists a Radon measure µT on Rn and a µT -measurable mapping T~ : Rn → m (Rn ) such that
|T~ (x)| = 1 for µT -a.e. x ∈ Rn and
Z
T (ω) = hT~ (x), ω(x)idµT (x) ∀ω ∈ D m (U ).

The total variation measure µT associated to T is characterized by

µT (W ) = sup{T (ω) : ω ∈ D m (U ), |ω| ≤ 1, supp ω ⊂ W } = MW (T )

for every open W ⊂ U . In particular,

µT (U ) = µT (Rn ) = M(T ).

Definition 4.30 (Restrictions of currents). If T ∈ Dm (U ), M(T ) < ∞, and A ⊂ Rn is Borel, then

the restriction of T to A is the m-current T xA ∈ Dm (U ),
Z
(T xA)(ω) = hT~ (x), ω(x)idµT (x), ω ∈ D m (U ),
A

where T~ and µT are as in Theorem 4.29. Similarly, if g is a µT -integrable function, we define

T xg ∈ Dm (U ), the interior multiplication by g, by
Z
(T xg)(ω) = g(x)hT~ (x), ω(x)idµT (x), ω ∈ D m (U ).

Definition 4.31. The support of T ∈ Dm (U ) is the set

[
supp T = U \ {V : V ⊂ Rn open, T (ω) = 0 ∀ω ∈ D m (U ), supp ω ⊂ V }.

If M(T ) < ∞, and hence µT exists, then supp T = U ∩ supp µT . Recall that the support of the
measure µT is the set
[
supp µT = Rn \ {V : V ⊂ Rn open, µT (V ) = 0}.

Definition 4.32. Let Ti , T ∈ Dm (U ). We say that the sequence Ti converges to T and write

Ti → T

if
lim Ti (ω) = T (ω)
i→∞
w⋆
for every ω ∈ D m (U ). Hence, in fact, Ti −−→ T .
56 Geometric Measure Theory

Proposition 4.33. Suppose Ti → T . Then

∂Ti → ∂T and M(T ) ≤ lim inf M(Ti ).

i→∞

Remark 4.34. The lower semicontinuity of the mass is very important and useful property in
mass minimizing problem.
Remark 4.35. We notice that the normed space (D m (U ), | · |) is separable, and hence the closed
unit ball of its dual {T ∈ Dm (U ) : M(T ) < ∞} is sequentially compact in the weak∗ topology by
the (sequential) Banach-Anaoglu theorem.
By applying the (sequential) Banach-Alaoglu theorem for the Banach space {T ∈
Dm (U ) : M(T ) < ∞} we obtain the following:
Theorem 4.36. Let Ti ∈ Dm (U ) with

sup M(Ti ) < ∞.

i

Then there exist a subsequence Tij and T ∈ Dm (U ) such that

Tij → T.

Next we define the cartesian product of currents Ti ∈ Dmi (Ui ), Ui ⊂ Rni open, i = 1, 2. Any
differential (m1 + m2 )-form ω in U1 × U2 can be written in the form
X
ω(x, y) = ωIJ dxI ∧ dy J , (x, y) ∈ U1 × U2 .
V V
(I, J) ∈ (n1 , m′1 ) × (n2 , m′2 )
′ ′
m1 +m2 =m1 +m2

Then we define:
Definition 4.37. Let Ti ∈ Dmi (Ui ), Ui ⊂ Rni open, i = 1, 2. The cartesian product T1 × T2 ∈
Dm1 +m2 (U1 × U2 ) is defined by
   
X X
(T1 × T2 )(ω) = T1  T2  ωIJ dy J  dxI 
V V
I∈ (n1 ,m1 ) J∈ (n2 ,m2 )

for ω ∈ D m1 +m2 (U1 × U2 ).

V V
Notice that T1 ×T2 ignores the terms dxI ∧dy J , where I ∈ (n1 , m′1 ), J ∈ (n2 , m′2 ), m′1 +m′2 =
m1 + m2 but (m′1 , m′2 ) 6= (m1 , m2 ).
Remark 4.38. (a) The motivation of the definition is, of course, that we want

[M1 × M2 ] = [M1 ] × [M2 ]

if M1 and M2 are smooth submanifolds.

(b) Since
d(α ∧ β) = dα ∧ β + (−1)m α ∧ dβ
for m-forms α, we have

∂(T1 × T2 ) = (∂T1 ) × T2 + (−1)m1 T1 × (∂T2 ).

Fall 2016 57

As an important special case we consider the following example:

 
Example 4.39. Let T1 = [0, 1] ∈ D1 (R) and T2 = T ∈ Dm (Rn ).

[1] × T
T1

 
[0, 1] × T

T [0] × T
Rn Rn

Then
 

∂ [0, 1] × T = (∂T1 ) × T − T1 × ∂T


= [1] − [0] × T − T1 × ∂T
 
= [1] × T − [0] × T − [0, 1] × ∂T.

[1] × T

∂T

(−) (+)
 
[0, 1] × T
T

[0] × T

Next we define the push-forward of a current under a smooth mapping.

Definition 4.40. Suppose that U : Rn and V ⊂ Rd are open sets and f : U → V a C ∞ -mapping.
Let T ∈ Dm (U ) be such that f | supp T is proper, i.e. f −1 K ∩ supp T is compact for every compact
K ⊂ V . We define f♯ T ∈ Dm (V ), the push-forward of T under f , by

f♯ T (ω) = T (ϕf ∗ ω), ω ∈ D m (V ),

where ϕ ∈ C0∞ (U ) is any function such that ϕ ≡ 1 in the compact set supp T ∩ f −1 supp ω ⊂ U .
Notice that ϕf ∗ ω ∈ D m (U ) but it is possible that f ∗ ω ∈
/ D m (U ) since supp f ∗ ω need not be
compact.
Remark 4.41. 1. If f and T are as above, then ∂f♯ T = f♯ ∂T .

2. If Ti → T and f |(supp Ti ∪ supp T ) is proper, then f♯ Ti → f♯ T .

3. Suppose that MW (T ) < ∞ for every W ⋐ U , and hence

Z
T (ω) = hT~ (x), ω(x)idµT (x) ∀ω ∈ D m (U ),

where T~ and µT are given by Theorem 4.29. Then the push-forward f♯ T is given by
Z Z

^
f♯ T (ω) = f ∗ ω, T~ dµT = ω f (x) , m dfx T~ (x) dµT (x).
58 Geometric Measure Theory

1
V
Notice thatV
the formula
V makes
n
V ifdf is C , with f | supp T proper. Above
sense m dfx is the
linear map m dfx : m (R ) → m (R ) defined by
^


m dfx (ei1 ∧ · · · ∧ eim ) = dfx ei1 ∧ · · · ∧ dfx eim
V
for every (i1 , . . . , im ) ∈ (n, m).

Now we can define the homotopy formula for currents. For that purpose let V ⊂ Rd be open
and let f, g : U → V be smooth mappings. Furthermore, suppose that h : [0, 1] × U → V is smooth
such that
h(0, x) = f (x) and h(1, x) = g(x) ∀x ∈ U.
Since (see Example 4.39)
 
 

∂h♯ [0, 1] × T = h♯ ∂ [0, 1] × T
 

= h♯ [1] × T − [0] × T − [0, 1] × ∂T


 

= h♯ [1] × T − h♯ [0] × T − h♯ [0, 1] × ∂T
 

= g♯ T − f♯ T − h♯ [0, 1] × ∂T ,

we have
 
 

(4.42) g♯ T − f♯ T = ∂h♯ [0, 1] × T + h♯ [0, 1] × ∂T .

g♯ T

[1] × T
h
1
 

  h♯ [0, 1] × T
[0, 1] × T
0
Rn
[0] × T   −f♯
T
−h♯ [0, 1] × ∂T
∂T

(−) T (+)

An important special case is the affine homotopy

h(t, x) = tg(x) + (1 − t)f (x).

Definition 4.43 (Cone). Let T ∈ Dm (U ) with supp T compact. The cone over T is
 

0 ⊳ T = h♯ [0, 1] × T ∈ Dm+1 (Rn ),

where h(t, x) = tx.

We notice that
∂(0 ⊳ T ) = T − 0 ⊳ ∂T.
In particular, if T has no boundary, then T itself is a boundary:

T = ∂(0 ⊳ T ).
Fall 2016 59

∂T

T
T

0⊳T ∂T = ∅
V
For a linear mapping L : Rn → Rd we denote by m L the linear mapping
^ ^ ^
n d
mL : m (R ) → m (R )

defined by ^
m L(ei1 ∧ · · · ∧ eim ) = Lei1 ∧ · · · ∧ Leim
V
for every (i1 , . . . , im ) ∈ (n, m). If f : U → V is smooth (V ⊂ Rd open), we see that
^
hv, f ∗ ω(x)i = m dfx (v), ω(x)
V n ),
for all v ∈ m (R ω ∈ D m (V ), and x ∈ U . Hence we can state:
Proposition 4.44. If T ∈ Dm (U ), with supp T compact and M(T ) < ∞ and if f : U → V is C ∞ ,
with f | supp T proper, then
Z

^
f♯ T (ω) = ω f (x) , m dfx T~ (x) dµT (x)

and

(4.45) M(f♯ T ) ≤ Lip(f | supp T )m M(T ).

Recall that  
|g(x) − g(y)|
Lip(g) := sup : x 6= y .
|x − y|
The inequality (4.45) follows from the estimate
^

~ m
m dfx T (x) ≤ Lip(f | supp T ) , x ∈ supp T,

which, in turn, is a consequence of

^
m L(ei1 ∧ · · · ∧ eim ) ≤ kLkm .

Suppose that h : [0, 1] × U → V is the affine homotopy h(t, x) = tg(x) + (1 − t)f (x) between
smooth mappings f, g : U → V . If T ∈ Dm (U ), with M(T ) < ∞, we have
 


m
(4.46) M h♯ [0, 1] × T ≤ sup |f − g| sup |dfx | + |dgx | M(T ).
supp T x∈supp T

This follows from the integral representation (Theorem 4.29) since

 −→ 
[0, 1] × T = e1 ∧ T~ and µ  = (m1 x[0, 1]) × µT ,
[0,1] ×T
60 Geometric Measure Theory

and therefore
Z
 

^

h♯ [0, 1] × T (ω) = ω h(t, x) , m+1 dh(t,x) e1 ∧ T~ (x) dµ 
[0,1] ×T
Z 1 Z ^ 



= ω h(t, x) , g(x) − f (x) e1 ∧ ~
m (tdgx + (1 − t)dfx )T (x) dµT dt.
0

Next we state a couple of further consequences of the homotopy formula.

Lemma 4.47. Let T ∈ Dm (U ), with MW (T ) < ∞ and MW (∂T ) < ∞ for every W ⋐ U . If
f, g : U → V ⊂ Rd are C 1 smooth with f | supp T = g| supp T proper, then f♯ T = g♯ T .
Proof. Applying the homotopy formula (4.42) with h(t, x) = tg(x) + (1 − t)f (x) we obtain
 
 

g♯ T (ω) − f♯ T (ω) = ∂h♯ [0, 1] × T (ω) + h♯ [0, 1] × ∂T (ω)
 
 

= h♯ [0, 1] × T (dω) + h♯ [0, 1] × ∂T (ω),

and therefore, by (4.46),

 
 

g♯ T (ω) − f♯ T (ω) = h♯ [0, 1] × T (dω) + h♯ [0, 1] × ∂T (ω)
 
 

≤ h♯ [0, 1] × T (dω) + h♯ [0, 1] × ∂T (ω)
 

 


≤ M h♯ [0, 1] × T |dω| + M h♯ [0, 1] × ∂T |ω|
 
≤ c M(T )|dω| + M(∂T )|ω| sup |g − f | = 0
supp T

since, by assumption, f = g in supp T .

With help of the homotopy formula we can define f♯ T for a Lipschitz mapping f : U → V ⊂ Rd
provided f | supp T is proper and MW (T ) < ∞, MW (∂T ) < ∞ for every W ⋐ U . For that purpose,
let ηε , ε > 0, be a standard mollifier;

ηε (x) = ε−n η(x/ε),

R
where η : Rn → [0, ∞) is C ∞ , with supp η ⊂ B(0, 1) and η = 1.
Given a Lipschitz map f : U → V we define the C ∞ mapping f (ε) = f ∗ ηε .
Lemma 4.48. Let T ∈ Dm (U ), with MW (T ) < ∞ and MW (∂T ) < ∞ for every W ⋐ U . Let
f : U → V ⊂ Rd be Lipschitz with f | supp T proper. Then the limit
(ε)
f♯ T (ω) := lim f♯ T (ω)
ε→0

exists for every ω ∈ D m (V ). Moreover, supp f♯ T ⊂ f (supp T ) and

m
MW (f♯ T ) ≤ ess sup |dfx | Mf −1 W (T )
f −1 W

for every W ⋐ V .
Proof. Fix ω ∈ D m (V ). If ε > 0 and σ > 0 are sufficiently small (depending on ω ∈ D m (V )), the
homotopy formula with h(t, x) = tf (ε) (x) + (1 − t)f (σ) (x) implies
(ε) (σ)  
 

f♯ T (ω) − f♯ T (ω) = ∂h♯ [0, 1] × T (ω) + h♯ [0, 1] × ∂T (ω)
 
 

= h♯ [0, 1] × T (dω) + h♯ [0, 1] × ∂T (ω).
Fall 2016 61

For sufficiently small ε > 0 and σ > 0 we get from (4.46)

(ε) (σ)
f♯ T (ω) − f♯ T (ω) ≤ c sup f (ε) − f (σ) Lip(f )m ,
f −1 K∩supp T

ε→0
where K ⊂ V is a compact set containing supp ω in its interior. Since f (ε) −−−→ f uniformly on
compact subsets of U , the claims follow.

Theorem 4.49 (Constancy theorem). Let U ⊂ Rn be a domain (i.e. open and connected). If
T ∈ Dn (U ), with ∂T = 0 and MW (T ) < ∞ for all W ⋐ U , then there exists a constant c such that

T = c[U ],

that is Z
1 n


T ϕdx ∧ · · · ∧ dx =c ϕdmn
U

for every ϕ ∈ C0∞ (U ).

Note that m = n above.

Proof. By Theorem 4.29 there exist a Radon measure µT and a µT -measurable function σ : U →
{−1, 1} such that
Z Z Z Z
T (ω) = hω(x), σ(x)e1 ∧ · · · ∧ en idµT (x) = σϕ dµT = ϕ dµT − ϕ dµ−
+
T

for every ω = ϕdx1 ∧ · · · ∧ dxn ∈ D n (U ), where µ+ −

T = µT x{σ = 1} and µT = µT x{σ = −1}. Let
ηε , ε > 0, be as above. Define
Tε (ω) := T (ηε ∗ ω)
V
for 0 < ε < dist(supp ω, ∂U ) and for continuous n-forms ω ∈ C0 (U, n (Rn )) with compact support
in U . Here ηε ∗ ω = ηε ∗ ϕ dx1 ∧ · · · ∧ dxn if ω = ϕ dx1 ∧ · · · ∧ dxn . We first observe that, for fixed
W ⋐ U and ε > 0, the set
^ Z
n n
S = {ηε ∗ ω : ω ∈ C0 (U, (R ), supp ω ⊂ W̄ , |ω|dmn ≤ 1}
U
Vn
is compact in C0 (U, (Rn )) with respect to the norm (| · |) topology. Hence, by continuity of T
also with respect to the norm topology, there exists a constant c = c(T, W, ε) such that
Z
(4.50) |Tε (ω)| ≤ c |ω|dmn
U
Vn
for every ω ∈ C0 (U, (Rn ), with supp ω ⊂ W̄ . On the other hand,
Z Z
Tε (ω) = T (ηε ∗ ω) = ηε ∗ ϕ dµT − ηε ∗ ϕ dµ−
+
T

if ω = ϕdx1 ∧ · · · dxn , ϕ ∈ C0 (W ). Applying the Riesz representation theorem to positive linear

functionals Z
ϕ 7→ ηε ∗ ϕ dµ± T, ϕ ∈ C0 (W ),
62 Geometric Measure Theory

we get Radon measures µ+ −

ε and µε such that
Z Z Z Z
ηε ∗ ϕ dµ+
T = ϕ dµ+
ε and ηε ∗ ϕ dµ−
T = ϕ dµ−
ε .

Hence, by (4.50),
Z Z Z Z
ϕ dµ+
ε − ϕ dµ−
ε = |Tε (ω)| ≤ c |ω|dmn = c |ϕ|dmn ,
U supp ϕ

and therefore µ+ −
ε , µε ≪ mn . The Radon-Nikodym theorem then implies that there exists gε ∈
1
L (mn ) such that
Z
(4.51) Tε (ω) = ϕgε dmn

for ω = ϕdx1 ∧ · · · ∧ dxn , ϕ ∈ C0 (W ). On the other hand, since ∂T = 0 by assumption, we have

(4.52) Tε (dω) = T (ηε ∗ dω) = T d(ηε ∗ ω) = ∂T (ηε ∗ ω) = 0
Vn−1
if ω ∈ C01 (U, (Rn )), with supp ω ⊂ W . Applying this to

ω = ϕdx1 ∧ · · · ∧ dxj−1 ∧ dxj+1 ∧ · · · ∧ dxn ,

for which
∂ϕ 1
dω = (−1)j−1 dx ∧ · · · ∧ dxn ,
∂xj
we get
Z
j−1 ∂ϕ
(4.53) Tε (dω) = (−1) gε dmn = 0
∂xj

for all ϕ ∈ C01 (W ) and for all j ∈ {1, . . . , n}. It follows that the distributional gradient of gε
vanishes mn -a.e. and therefore gε = cε mn -a.e., where cε is a constant1 . Letting then ε → 0 and
W ր U , we obtain (by continuity of T )
Z Z
T (ω) = lim Tε (ω) = lim cε ϕdmn = c ϕdmn = c[U ](ϕ)
ε→0 ε→0 U U

for all ω = ϕdx1 ∧ · · · ∧ dxn ∈ D n (U ), where the limit

lim cε = c
ε→0

exists since the limit Z

lim Tε (ω) = lim cε ϕdmn
ε→0 ε→0 U
exists.

Next we want to weaken the assumption ∂T = 0 to M(∂T ) < ∞. Before we state the theorem
(Theorem 4.65), which a generalization of the Constancy theorem, we first discuss about functions
of bounded variation. We refer to e.g. [EG], [Si], [Ho3] for more details.
1
This follows from Poincaré’s inequality for W 1,1 -functions.
Fall 2016 63

Definition 4.54. Let U ⊂ Rn be open and u ∈ L1loc (U ). Define

Z Z 
1 n
|Du| := sup u div g : g = (g1 , . . . , gn ) ∈ C0 (U ; R ), |g| ≤ 1 .
U U
R
Above U |Du| should be understood just as a notation (not an integral). Furthermore,
n
X ∂gi
div g =
∂xi
i=1

is the usual divergence.

Example 4.55. (a) If u ∈ C 1 (U ), then integration by parts implies that

Z Z
u div g = − ∇u · g ∀g ∈ C01 (U ; Rn ),
U U

and so Z Z
|Du| = |∇u|.
U U

1,1
(b) More generally, if u belongs to the Sobolev space Wloc (U ), then again
Z Z
|Du| = |∇u|,
U U

where ∇u is the distributional gradient of u.

Definition 4.56. A function u ∈ L1loc (U ) is said to have bounded variation in U if

Z
|Du| < ∞.
U

We denote by BV(U ) the vector space of all functions u ∈ L1 (U ) with bounded variation in U .

Definition 4.57. Similarly, a function u ∈ L1loc (U ) has locally bounded variation and belongs to
BVloc (U ) if Z
|Du| < ∞
V
for every relatively compact open set V ⋐ U .

The proof of the following theorem is an application of the Riesz representation theorem.

Theorem 4.58. For every u ∈ BVloc (U ) there exists a Radon measure µ on U and a µ-measurable
mapping σ : U → Rn such that

(i) |σ(x)| = 1 for µ-a.e. x ∈ U ;

(ii) Z Z
u div g dx = − g · σ dµ
U U

for every g ∈ C01 (U ; Rn ).

64 Geometric Measure Theory

Remark 4.59. 1. If u ∈ BVloc (U ), we denote by kDuk the Radon measure µ given by Theo-
rem 4.58 and by
[Du] = kDukxσ
the vector valued measure d[Du] = σ dkDuk. Hence
Z Z Z
u div g = − g · σ dkDuk = − g · d[Du]
U U U

for g ∈ C01 (U ; Rn ).
2. If u ∈ BV(U ) and V ⋐ U is an open subset, then
Z 
1 n
kDuk(V ) = sup u div g dx : g ∈ C0 (U ; R ), |g| ≤ 1 .
V
Hence, using our earlier notation,
Z
|Du| = kDuk(V ).
V
Theorem 4.60 (Lower semicontinuity). Let U ⊂ Rn be open and uj ∈ BV(U ), j ∈ N such that
uj → u in L1loc (U ). Then
Z Z
(4.61) |Du| ≤ lim inf |Duj |.
U j→∞ U
Theorem 4.62. The vector space BV(U ) equipped with the BV-norm
Z
kukBV := kukL1 (U ) + |Du|
U
is a Banach space.
Functions in Sobolev spaces W 1,p (U ), 1 ≤ p < ∞, can be approximated by C ∞ (U ) functions
in the Sobolev norm
kuk1,p := kukp + k|∇u|kp .
In fact, W 1,p (U ) is the completion of C ∞ (U ) in the Sobolev norm and since BV(U ) 6= W 1,1 (U ),
functions in BV(U ) can not be approximated in the BV-norm. However,
Theorem 4.63 (Approximation). Let u ∈ BV(U ). Then there exists a sequence uj ∈ C ∞ (U ), j ∈
N, such that
Z
lim |uj − u| = 0,
j→∞ U
Z Z
lim |∇uj | = |Du|.
j→∞ U U
Suppose that u ∈ BV(U ) and uj ∈ C ∞ (U ) are as above. For each j ∈ N let µj be the vector-
valued Radon-measure defined by
Z
µj (B) = ∇uj dx
B∩U
for Borel sets B ⊂ Rn . Furthermore, let µ be the vector-valued Radon measure
Z Z
µ(B) = d[Du] = σ dkDuk.
B∩U B∩U
Then µj ⇀ µ.
Fall 2016 65

Theorem 4.64 (Compactness). Let U ⊂ Rn be an open set. If uj ∈ BVloc (U ), j ∈ N, is a sequence

such that  Z 
sup kuj kL1 (W ) + |Duj | < ∞
j W

for every W ⋐ U , there exist a subsequence (ujk ) and u ∈ BVloc (U ) such that ujk → u in L1loc (U )
and Z Z
|Du| ≤ lim inf |Dujk |
W jk →∞ W
for every W ⋐ U .

Let us now return to consider n-currents. In the next theorem, which is a generalization of the
Constancy theorem, we weaken the assumption ∂T = 0 to M(∂T ) < ∞.

Theorem 4.65. Let T ∈ Dn (U ) such that M(∂T ) < ∞ and MW (T ) < ∞ for every W ⋐ U . Then
there exists g ∈ BVloc (U ) such that
Z
(4.66) T (ω) = ϕg dmn ,

for all ω = ϕdx1 ∧ · · · ∧ dxn ∈ D n (U ).

The proof is a modification of the proof of the Constancy theorem. Instead of equality (4.52)
we now have an estimate
Z
∂ϕ
(4.67) gε dmn = |Tε (dω)| ≤ sup |ηε ∗ ϕ|M(∂T ) ≤ cε M(∂T )
∂xj

if ω = ϕdx1 ∧ · · · dxj−1 ∧ dxj+1 ∧ · · · ∧ dxn , with ϕ ∈ C01 (W ), |ϕ| ≤ 1. Here cε is a constant that
depends on ε and cε → 1 as ε → 0 since ηε ∗ ϕ → ϕ uniformly. We apply (4.67) with

ω = (−1)j Φj dx1 ∧ · · · ∧ dxj−1 ∧ dxj+1 ∧ · · · dxn ,

where Φj ∈ C01 (U ), supp Φj ⊂ W, is the jth-coordinate function of Φ = (Φ1 , . . . , Φn ) ∈ C01 (U, Rn ),

with |Φ| ≤ 1. We obtain
Z Z Xn
∂Φj
(div Φ)gε dmn = gε dmn ≤ ncε M(∂T ) ≤ 2nM(∂T )
U U ∂xj
j=1

for all Φ ∈ C01 (U, Rn ), with |Φ| ≤ 1, and 0 < ε < dist(W, ∂U ) small enough. Hence gε ∈ BV(W ). It
follows from the Poincaré’s inequality for BV -functions (see e.g. [EG, 5.6.1], [Si, Lemma 6.4]) that
gε is locally uniformly bounded in L1 (U ). We conclude (using Theorem 4.64) that there exists a
sequence εk ց 0 such that gεk → g in L1loc (U ) with g ∈ BVloc (U ). Moreover, it follows from (4.51)
that Z
T (ω) = ϕg dmn
U

for ω = ϕdx1 ∧ ··· ∧ dxn

∈ D n (U ).
Writing an arbitrary α ∈ D n−1 (U ) as
n
X
α= (−1)j Φj dx1 ∧ · · · ∧ dxj−1 ∧ dxj+1 ∧ · · · dxn
j=1
66 Geometric Measure Theory

we first observe that

dα = (div Φ)dx1 ∧ · · · ∧ dxn , Φ = (Φ1 , . . . , Φn ),
and therefore
Z
1 n



∂T (α) = T (dα) = T (div Φ)dx ∧ · · · ∧ dx = div Φ g dmn
U

by (4.66). Finally, it follows directly from definitions that

Z
MW (T ) = |g|dmn
W
and Z
MW (∂T ) = |Dg| = kDgk(W )
W

for every W ⋐ U .
The last theorem in this subsection deals with restrictions of m-currents to subsets of Rn
with ”small” orthogonal V projections onto Rm . To state the result, we define for each multi-index
I = (i1 , . . . , im ) ∈ (n, m) the orthogonal projection PI : Rn → Rm by

PI (x) = PI (x1 , . . . , xn ) = (xi1 , . . . , xim ) ∈ Rm .

n n
Theorem
m

4.68. Suppose thatVE ⊂ R is a closed subset of an open set U ⊂ R such that
H PI E = 0 for every I ∈ (n, m). Then T xE = 0 for all T ∈ Dm (U ), with MW (T ) < ∞
and MW (∂T ) < ∞.

Proof. Let ω ∈ D m (U ). We can write

X
ω= ωI dxI , dxI = dxi1 ∧ · · · ∧ dxim , I = (i1 , . . . , im ).
V
I∈ (n,m)

Hence
X X
T (ω) = T (ωI dxI ) = (T xωI )(dxI )
I I
X

(4.69) = (T xωI ) PI∗ (dy 1 ∧ · · · ∧ dy m )
I
X
= PI♯ (T xωI )(dy 1 ∧ · · · ∧ dy m ),
I

where we have denoted by dy 1 ∧ · · · ∧ dy m the standard basis m-form in Rm . The push-forward

makes sense since supp(T xωI ) is a subset of supp ωI which is a compact subset of U .
For any β ∈ D m−1 (U )


∂(T xωI )(β) = (T xωI )(dβ) = T (ωI dβ) = T d(ωI β) − T (dωI ∧ β)
= ∂T (ωI β) − T (dωI ∧ β),

and so


(4.70) MW ∂(T xωI ) ≤ MW (∂T )|ωI | + MW (T )|dωI |.
Fall 2016 67

We obtain



M ∂PI♯ (T xωI ) = M PI♯ ∂(T xωI ) ≤ c(n, m)M ∂(T xωI ) < ∞
by (4.45), (4.70), and the assumptions MW (T ), MW (∂T ) < ∞ ∀W ⋐ U . Therefore, by Theo-
rem 4.65, there exists g ∈ BV(PI U ) such that
Z
PI♯ (T xωI )(β) = hβ, e1 ∧ · · · ∧ em ig dmm ,
PI U

and hence
PI♯ (T xωI )xPI E = 0
since mm (PI E) = 0. Assuming, without loss of generality, that E is compact, we have


PI♯ (T xωI ) = PI♯ (T xωI )x(Rm \ PI E) = PI♯ (T xωI )x Rn \ PI−1 (PI E) .

This implies




M PI♯ (T xωI ) ≤ M (T xωI )x Rn \ PI−1 (PI E)


(4.71) ≤ M (T xωI )x(Rn \ E)


≤ MW T x(Rn \ E) )|ωI |

for every open W such that supp ω ⊂ W ⋐ U . Combining (4.69) and (4.71) we get


MW (T ) ≤ cMW T x(Rn \ E)

for all open W ⋐ U . In particular,

M(T xE) = MW (T xE) ≤ cMW T x(Rn \ E)

for all W ⋐ U , with E ⊂ W . Choosing a descending sequence of open sets Wi ⋐ U such that
E = ∩i Wi , we get

M(T xE) ≤ cMWi T x(Rn \ E) → 0
which implies T xE = 0.

4.72 Rectifiable currents

Definition 4.73. An m-current T ∈ Dm (U ) in an open set U ⊂ Rn is called a rectifiable m-current
if there exist

1. an m-rectifiable Borel set E ⊂ U , with Hm (E) < ∞;

2. an Hm -integrable positive function θ : E → (0, ∞), and

V
3. an Hm -measurable mapping T~ : E → m (Rn ) such that, for Hm -a.e. x ∈ E, T~ (x) = v1 (x) ∧
· · · ∧ vm (x) where v1 (x), . . . , vm (x) is an orthonormal basis of the approximate tangent space
Txm E, and that Z
T (ω) = ω(x), T~ (x) θ(x)dHm (x)
E

for all ω ∈ D m (U ). Note that |T~ (x)| = 1 for Hm -a.e. x ∈ E.

68 Geometric Measure Theory

The function θ is called the multiplicity of T and T~ is called the orientation for T . We write
T = τ (E, θ, T~ ). Such a current T is called an integer multiplicity (rectifiable) m-current, denoted
T ∈ Rm (U ), if θ is integer valued.
Example 4.74. (1) If T1 , T2 ∈ Rm (U ) and p1 , p2 ∈ N, then p1 T1 + p2 T2 ∈ Rm (U ).

(2) If T1 = τ (E1 , θ1 , T~1 ) ∈ Rm (U ) and T2 = τ (E2 , θ2 , T~2 ) ∈ Rk (V ), then

T1 × T2 = τ (E1 × E2 , θ1 θ2 , T~1 ∧ T~2 ) ∈ Rm+k (U × V ).

(3) If f : U → V is Lipschitz, T = τ (E, θ, T~ ) ∈ Rm (U ), and f | supp T is proper, we can define

f♯ T ∈ Dm (V ) by
Z

^

f♯ T (ω) = ω f (x) , m dE fx T~ (x) θ f (x) dHm (x)
E

for ω ∈ D m (V ). Since
^
d fx T~ (x) = JfE (x),
m E

we get from the area formula that

Z * Vm +
X dE fx T~ (x)
(4.75) f♯ T (ω) = ω(y), θ(x) V dHm (y),
fE m dE f T ~
x (x)
x∈f −1 (y)∩E+

where E+ = {x ∈ E : JfE (x) > 0}. Notice that f E is m-rectifiable, and therefore the
approximate tangent space Tym f E exists at Hm -a.e. x ∈ f E. Hence at points y ∈ f E where
Tym f E exists and for which Txm E and dE fx exist for all x ∈ f −1 (y) ∩ E+ , we have
Vm
dE fx T~ (x)
Vm = ±τ1 ∧ · · · ∧ τm ,
dE fx T~ (x)

where τ1 , . . . , τm is an orthonormal basis of Tym f E. Hence we obtain from (4.75)

Z
f♯ T (ω) = ~
hω(y), S(y)iN (y)dHm (y),
fE

~
where S(y) is an orientation of Tym f E and N (y) is a positive integer satisfying
Vm
X dE fx T~ (x) ~
θ(x) V = N (y)S(y).
x∈f −1 (y)∩E+
m dE fx T~ (x)

In conclusion, f♯ T ∈ Rm (V ).
Definition 4.76. An m-current P ∈ Dm (U ) is a polyhedral (m-)chain if there exist m-dimensional
oriented simplices π1 , . . . , πk ⊂ U and p1 , . . . , pk ∈ R such that
k
X
P = pi [πi ].
i=1

If p1 , . . . , pk ∈ Z, P is called an integral polyhedral chain and we denote

P ∈ Pm (U ).
Fall 2016 69

Recall that an m-simplex π is the convex hull of its m + 1 affinely independent vertices
a0 , . . . , am ∈ Rm , that is a1 − a0 , a2 − a0 , . . . , am − a0 are linearly independent and
(m m
)
X X
π= λi ai : λi = 1, λi ≥ 0 ∀i .
i=0 i=0

Theorem 4.77. If Ti ∈ Rm (U ) is a sequence (of integer multiplicity rectifiable m-currents) with

sup (MW (Ti ) + MW (∂Ti )) < ∞

i∈N

for all W ⋐ U , then there exist a subsequence Tij and T ∈ Rm (U ) such that Tij → T .

Note that the existence of a subsequence and an m-current T ∈ Dm (U ) such that Tij → T
follows from the Banach-Alaoglu theorem; see Theorem 4.36. The difficulty is to prove that T is
an integer multiplicity rectifiable current; we will return to this later.
The next theorem gives a criterion of rectifiability.

Theorem 4.78. Let T ∈ Dm (U ) with M(T ) < ∞. Then T ∈ Rm (U ) if and only if for every ε > 0
there exist P ∈ Pm (Rd ), d ≥ m, and a Lipschitz map f : Rd → Rn such that

(4.79) M(T − f♯ P ) < ε.

Proof. Idea: ⇐ Let T ∈ Dm (U ) with M(T ) < ∞. Each m-simplex is a subset of Rm ⊂ Rd

and hence f♯ P is an m-rectifiable integer multiplicity current. Apply (4.79) with εi ց 0, i.e. let
Pi ∈ Pm (Rd ) such that
M(T − f♯ Pi ) < εi .
Then, for every ω ∈ D m (U ),

|T (ω) − f♯ Pi (ω)| = |(T − f♯ Pi )(ω)| ≤ M(T − f♯ Pi )|ω| → 0.

Hence T , as a limit of integer multiplicity m-currents f♯ Pi ∈ Rm (Rn ), is an integer multiplicity

m-current; see Lemma 4.80.
⇒ Let ε > 0 and T = τ (E, θ, T~ ) ∈ Rm (U ). We may assume (ignoring a set of Hm -measure
zero) that E is a countable union of Lipschitz images fi Ai of subsets Ai ⊂ Rm . Furthermore, we
may assume that the sets Ai are disjoint and that θ|fi Ai takes a constant value θi ∈ N. Then we
take θi copies Ai,j , j = 1, . . . , θi of Ai such that all the sets Ai,j , i ∈ N, j = 1, . . . , θi , are disjoint.
Now we can define a Lipschitz map (after applying the corollary of the McShane-Whitney extension
theorem) f : Rm → Rn such that f Ai,j = f Ai and that f preserves orientation. On the other hand,
each Ai,j can be approximated by finitely many m-simplices and hence T can be approximated (in
mass) by an integral polyhedral chain.

In the above proof, the step M(T − f♯ Pi ) → 0 ⇒ f♯ Pi → T is relatively easy.

Lemma 4.80. The set of integer multiplicity rectifiable currents in Dm (U ) is complete with respect
to the family of seminorms {MW : W ⋐ U }.

Proof. Let Ti = τ (Ei , θi , T~i ) ∈ Rm (U ), i ∈ N be a Cauchy sequence with respect to the family
{MW : W ⋐ U }. Then
Z
(4.81) MW (Ti − Tj ) = |θi T~i − θj T~j |dHm < ε(W, j)
W
70 Geometric Measure Theory

for i ≥ j, where ε(W, j) ց 0 as j → ∞, and we have made a convention θk = 0 and T~k = 0 in

U \ Ek . Since |T~k (x)| ≡ 1 in Ek , we get
Z
(4.82) |θi − θj |dHm < ε(W, j), i ≥ j.
W

Hence θi → θ in L1loc (U, Hm ), where θ is integer valued. From (4.82), we get

(4.83) Hm (E+ \ Ej ) ∪ (Ej \ E+ ) < ε(W, j),

where E+ = {x ∈ U : θ(x) > 0}. Since

θi |T~i − T~j | = |θi T~i − θj T~j + (θj − θi )T~j | ≤ |θi T~i − θj T~j | + |θj − θi ||T~j |,

we have Z
θi |T~i − T~j |dHm < 2ε(W, j), i ≥ j,
W
V
and therefore T~i converges in L1loc (Hm ) to T~ : U → m (Rn ), where T~ is simple and |T~ | = 1 in
V
E+ . Since T~j ∈ m (Tx Ej ) for Hm -a.e. x ∈ Ej and Tx Ej = Tx E+ in Ej ∩ E+ except a set of
V
Hm -measure ≤ ε(W, j) by (4.83), we conclude that T~ ∈ m (Tx E+ ), and so M(T − Tj ) → 0, with
T = τ (E+ , θ, T~ ) ∈ Rm (U ).

4.84 Slicing
In this subsection we introduce the slicing of a current by level sets of a Lipschitz function. [Recall
the co-area formula and, in particular, Theorem 2.59, where we ”sliced” an m-rectifiable set E by
level sets of a Lipschitz function.]
Definition 4.85. A current T ∈ Dm (U ) is normal, denoted by T ∈ Nm (U ), if supp T is compact
and
M(T ) + M(∂T ) < ∞.
Definition 4.86. Let T ∈ Nm (U ) be normal and f : Rn → R a Lipschitz map. The slice of T with
f and t ∈ R is


hT, f, ti := (∂T )x{x : f (x) > t} − ∂ T x{x : f (x) > t} ∈ Dm−1 (U ).

∂T x{f > t}
f≤t f>t
hT, f, ti
T x{f > t}

Theorem 4.87. The slices have the properties:

(1)

hT, f, ti = ∂ T x{x : f (x) < t} − (∂T )x{x : f (x) < t}
except at most countably many t;
Fall 2016 71

(2)
supphT, f, ti ⊂ f −1 (t) ∩ supp T ;

(3)

1

M hT, f, ti ≤ Lip(f ) lim inf µT {x : t ≤ f (x) ≤ t + h} ;
hց0 h

(4)
Z b


M hT, f, ti dt ≤ Lip(f )µT {x : a < f (x) < b} ;
a

(5)
∂hT, f, ti = −h∂T, f, ti;

(6) hT, f, ti is normal for almost every t.

Proof. Idea of (some) proofs: (1) holds for every t for which


(µT + µ∂T ) {x : f (t) = t} = 0.

(2) is easy. To prove (3), we approximate the characteristic function

χ{x : f (x)>t}

by a sequence of C ∞ functions g such that g(x) = 0 if f (x) ≤ t, f (x) = 1 if f (x) ≥ t + h, and

λ Lip(f )
Lip(g) ≤ ,
h
where λ > 1, λ ≈ 1. Then



M hT, f, ti ≈ M (∂T )xg − ∂(T xg)
= M (T xdg)


≤ Lip(g)µT {x : t ≤ f (x) ≤ t + h} .

(4) follows from (3) by integration, (5) is clear, and finally (6) follows from (4) and (5).

Next we slice integer multiplicity rectifiable currents.

Theorem 4.88. Let T = τ (E, θ, T~ ) ∈ Rm (U ), with M(∂T ) < ∞ and let f : Rn → R be a Lipschitz
function. Then for a.e. t ∈ R:
(1) hT, f, ti = τ (Et , θt , T~t ), where

Et = E ∩ f −1 (t),
(
θ(x), if x ∈ Et and ∇E f (x) 6= 0;
θt (x) =
0, if x ∈ Et and ∇E f (x) = 0,
∇E f (x)
T~t (x) = T~ (x)x E ,
|∇ f (x)|

(2) Z Z
∞

M hT, f, ti dt = |∇E f ||θ|dHm ≤ Lip(f )M(T ),
−∞ E
72 Geometric Measure Theory

(3) hT, f, ti ∈ Rm−1 (U ), with M ∂hT, f, ti < ∞ and ∂hT, f, ti = −h∂T, f, ti.
V V V
The interior multiplication x : q (V ) × p (V ) → q−p (V ) is characterized by the condition

hvxα, βi = hv, α ∧ βi
V V Vq−p
whenever v ∈ q (V ), α ∈ p (V ), β ∈ V (V ). Moreover,
Vm there is the standard biduality
Vm between
finite dimensional inner
V product spaces m (V ) and (V ). That is, for every η ∈ (V ) there
exists a unique w ∈ m (V ) such that
^
(4.89) hv, wi = hη, vi ∀v ∈ m (V ).

V
Hence, in particular, T~ (x)x∇E f (x) ∈ m−1 (Txm−1 Et ) is characterized by the property

hT~ (x)x∇E f (x), ηi = hT~ (x), dE fx ∧ ηi

Vm−1
for all η ∈ (Txm−1 Et ).

Proof of Theorem 4.88. Let h : Rn → R be Lipschitz and let hε = ηε ∗ h be as before. Then hε → h

locally uniformly as ε → 0. Now


∂T (hε ω) = T d(hε ω) = T (dhε ∧ ω) + T (hε dω)

for all ω ∈ D m (U ). Here

Z →
Z →
∂T (hε ω) = h∂T , hε ωidµ∂T → h∂T , hωidµ∂T = (∂T xh)(ω)

and Z Z
T (hε dω) = hT~ , hε dωidµT → hT~ , hdωidµT = (T xh)(dω) = ∂(T xh)(ω)

as ε → 0. So,
(∂T xh)(ω) = lim T (dhε ∧ ω) + ∂(T xh)(ω),
ε→0

where
Z
T (dhε ∧ ω) = hT~ (x), dhε (x) ∧ ωiθ(x)dHm (x)
ZE
= hT~ (x), dE hε (x) ∧ ωiθ(x)dHm (x)
E
Z
(4.89)
= hT~ (x)x∇E hε (x), ω(x)iθ(x)dHm (x)
Z E
→ hT~ (x)x∇E h(x), ω(x)iθ(x)dHm (x).
E

2 Hence we get from the convergences above that

Z
(∂T xh)(ω) = hT~ (x)x∇E h(x), ω(x)iθ(x)dHm (x) + ∂(T xh)(ω)
E
2
The last convergence holds since ∇E hε → ∇E h weakly in L2 (Hm xθ) which, in turn, can be proven by noticing
m 1
that E = ⊔∞ i=0 Ei , H (E0 ) = 0, and Ei ⊂ Mi , with Mi a C -smooth submanifold.
Fall 2016 73

for Lipschitz functions h : Rn → R.

Let then f : Rn → R be Lipschitz, t ∈ R, and ε > 0. Define a continuous function γε : R → R
by 
0, if s < t − ε,

γε (s) = linear, if t − ε ≤ s ≤ t,


1, if s > t
and gε = γε ◦ f . We then have
Z
(4.90) (∂T xgε )(ω) = hT~ (x)x∇E gε (x), ω(x)iθ(x)dHm (x) + ∂(T xgε )(ω).
E

Now
ε→0
gε (x) −−−→ χ{f >t} (x),
so
Z →
Z →
(4.91) (∂T xgε )(ω) = h∂T , gε ωidµ∂T → h∂T , χ{f >t} ωidµ∂T = ∂T x{f > t}(ω).

Similarly,
Z Z


(4.92) (T xgε )(dω) = hT~ , gε dωidµT → hT~ , χ{f >t} dωidµT = ∂ T x{f > t} (ω).

By the chain rule

∇E gε (x) = ∇E (γε ◦ f )(x) = γε′ f (x) ∇E f (x)
(
0, if f (x) < t − ε or f (x) > t,
= 1 E
ε ∇ f (x), if t − ε < f (x) < t,
so
Z Z
1
hT~ (x)x∇E gε (x), ω(x)iθ(x)dHm (x) = hT~ (x)x∇E f (x), ω(x)iθ(x)dHm (x)
E ε
{x∈E : t−ε<f (x)<t}
Z  
1 ~ ∇E f
= T x E , ω |∇E f |θdHm
ε |∇ f |
{x∈E : t−ε<f (x)<t}
Z Z 
1 t
= hT~s , ωiθs dH m−1
ds
ε t−ε Es
Z
→ hT~t , ωiθt dHm−1
Et

for a.e. t ∈ R. Recalling (4.90)-(4.92) and the definition of hT, f, ti we get

hT, f, ti = (∂T )x{f > t} − ∂(T x{f > t})
= lim (∂T xgε ) − lim ∂(T xgε )
ε→0 ε→0
Z
= lim T x∇E gε , · θdHm
ε→0 E
Z
= hT~t , ·iθt dHm−1
Et
= τ (Et , θt , T~t ),
74 Geometric Measure Theory

and therefore (1) holds. (2) follows from (1) and the co-area formula. (3) follows from (1), Theo-
rem 2.59, and Theorem 4.87 (5).

It is possible to slice an current T ∈ Dm (U ) with a Lipschitz map f = (f1 , . . . , fk )) : Rn →

Rk , k ≤ m, and a value y = (y1 , . . . , yk ) ∈ Rk by iterating the slicing with fi and yi :

hT, f, yi = hh· · · hhT, f1 , y1 i, f2 , y2 i · · · i , fk , yk i ∈ Dm−k (U ).

4.93 Deformation theory

The deformation theorem is one of the fundamental results and it provides a useful approximation
of a normal current T by a polyhedral chain P lying on a certain m-skeleton such that the error
is of the form T − P = ∂R + S. The main error term is ∂R, where R is the (m + 1)-dimensional
surface through which T is deformed to P . The other error term S arises in moving ∂T into the
skeleton.
We will only state the result and sketch the (long and technical) proof. First we introduce some
notation: Fix k, m, n ∈ N, 0 < m < n, and ε > 0. We denote by

Qε = [0, ε]n ⊂ Rn

the closed n-dimensional cube of side length ε and by

k
[
Lε,k = Lε,j = {π : π j-dimen. closed face of some Qε + pε, p ∈ Zn }
j=1

the k-skeleton of mesh ε. Thus the elements of

• Lε,0 are singletons (vertices),

• Lε,1 are closed line segments (edges) of length ε,

• Lε,2 are closed squares of side length ε, · · ·

• Lε,n are the closed n-cubes Qε + pε, p ∈ Zn , of side length ε.

n

Moreover, we denote by Vε,1 , . . . , Vε,N , N = m+1 the (m + 1)-dimensional affine subspaces of Rn
that contain some (m + 1)-face of Qε . Finally,

Pε,j : Rn → Vε,j

denotes the orthogonal projection onto Vε,j .

Theorem 4.94 (Deformation theorem). Let ε > 0 and T ∈ Dm (Rn ), with M(T ) + M(∂T ) < ∞.
Then there are P, S ∈ Dm (Rn ) and R ∈ Dm+1 (Rn ) such that

T − P = ∂R + S,

where P, R, and S satisfy the following:

X
(4.95) P = απ [π], απ ∈ R,
π∈Lε,m

(4.96) M(P ) ≤ cM(T ), M(∂P ) ≤ cM(∂T ),

Fall 2016 75

(4.97) M(R) ≤ cεM(T ), M(S) ≤ cεM(∂T ),

with c = c(n, m), and

√
supp P ∪ supp R ⊂ {x ∈ Rn : dist(x, supp T ) < 2ε n},
√
supp ∂P ∪ supp S ⊂ {x ∈ Rn : dist(x, supp ∂T ) < 2ε n}.

If T ∈ Rm (Rn ), also P and R can be chosen to be integer multiplicity with απ ∈ Z. If, in addition,
∂T ∈ Rm−1 (Rn ), also S can be chosen to be integer multiplicity.
For the proof, we may assume that ε = 1. Indeed, the ”scaled version” 4.94 follows from the
”unscaled” one where ε = 1 by first applying the homothety x 7→ x/ε, then applying the ”unscaled
version” and then scaling back by x 7→ εx. In particular, the linear dependence of the constant cε
in (4.97) on ε is then obvious.
The main tool in the proof of the deformation theorem is the following lemma that provides a
suitable class of retractions to push-forward T into the m-skeleton L1,m (in the unscaled version).
We denote by q = (1/2, . . . , 1/2) the center of the unit cube Q = Q1 and abbreviate Lk = L1,k , Lk =
L1,k , and Pj = P1,j . Given a point a ∈ B(q, 1/4), we denote

Ln−m−1 (a) = a + Ln−m−1 (shifted skeleton)

and

Ln−m−1 (a; ρ) = {x ∈ Rn : dist x, Ln−m−1 (a) < ρ}, ρ ∈ (0, 1/4).
Then

dist Ln−m−1 (a), Lm ≥ 1/4 ∀a ∈ B(q, 1/4).
Lemma 4.98. For every a ∈ B(q, 1/4) there is a locally Lipschitz map

ψ : Rn \ Ln−m−1 (a) → Rn \ Ln−m−1 (a)

such that


ψ Q \ Ln−m−1 (a) = Q ∩ Lm ,
ψ|Q ∩ Lm = idQ∩Lm ,
|Df (x)| ≤ c/ρ

for mn -a.e. x ∈ Q \ Ln−m−1 (a; ρ), ρ ∈ (0, 1/4), with c = c(n, m) and that

ψ(z + x) = z + ψ(x)

for all x ∈ Rn \ Ln−m−1 (a) and z ∈ Zn .

Thus ψ is a Zn -periodic retraction of Rn \ Ln−m−1 (a) onto Lm . The rough idea is then to define

P̃ = ψ♯ T,
 

R = h♯ [0, 1] × T ,
and
 

S1 = h♯ [0, 1] × ∂T ,

where h(t, x) = tx + (1 − t)ψ(x), so that the homotopy formula gives

T = P̃ + ∂R + S1 .
76 Geometric Measure Theory

Choosing the point a ∈ B(q, 1/4) properly (depending on T ) we may get estimates
M(P̃ ) ≤ cM(T ),
M(∂ P̃ ) ≤ cM(∂T ),
M(R) ≤ cM(T ),
and
M(S1 ) ≤ cM(∂T ).

We notice that P̃ need not be a polyhedral chain. It is used to choose appropriate multiplicities of
the m-faces in the m-skeleton. For each m-face F ∈ Lm , P̃ xF corresponds by Theorem 4.65 to a
BV -function θF so that
Z Z


M(P̃ xF ) = |θF |dHm , M (∂ P̃ )xF = |DθF |dHm .
F F

Letting then Z
1
mF = θF dHm ,
Hm (F ) F
we define X
P = mF [F ]
F ∈Lm
and
S = S1 + (P̃ − P ).
In the proof of the mass estimates, for instance, slicing is used.
Next we give some applications of the deformation theorem.
Theorem 4.99 (Isoperimetric inequality). If T ∈ Rm (Rn ) with supp(T ) compact and ∂T = 0,
there exists R ∈ Rm+1 (Rn ), with supp(R) compact, ∂R = T and
M(R)m/(m+1) ≤ Cn,m M(T ).
Proof. We may assume T 6= 0. Choose ε > 0 so that εm = 2cM(T ), where c = c(n, n) is the
constant in the deformation theorem. By the deformation theorem, there are P, R, and S such that
T = P + ∂R + S,
where R ∈ Rm+1 (Rn ), with compact support,
X
P = απ [π], απ ∈ Z,
π∈Lε,m

M(P ) ≤ cM(T ),
M(S) ≤ cεM(∂T ),
and
M(R) ≤ cεM(T ) = c(2c)1/m M(T )(m+1)/m .
Since ∂T = 0, we obtain from above that S = 0. On the other hand,
X X X
M(P ) = |απ |Hm (π) = εm |απ | = 2cM(T ) |απ | ≤ cM(T ),
|{z}
π∈Lε,m π π
∈N

so απ = 0 for all π, and therefore P = 0. Finally, since P = S = 0, we have ∂R = T .

Fall 2016 77

To state the other application, we first give a definition.

Definition 4.100. The flat distance between m-currents T1 , T2 ∈ Dm (Rn ) is

F (T1 , T2 ) = inf{M(S) + M(R) : T1 − T2 = ∂R + S, R ∈ Dm+1 (Rn ), S ∈ Dm (Rn )}.

Remark 4.101. F (·, ·) is a metric in {T ∈ Dm (Rn ) : M(T ) < ∞} and a convergence with respect
to F is stronger than the weak convergence (i.e. convergence as currents):

F (Ti , T ) → 0 ⇒ Ti → T,

but weaker than the mass convergence:

M(Ti − T ) → 0 ⇒ F (Ti , T ) → 0.

Theorem 4.102 (Polyhedral approximation theorem). If T ∈ Dm (Rn ) with M(T ) + M(∂T ) < ∞,
there exists a sequence Pk of the form
X
Pk = απ [π], απ ∈ R,
π∈Lεk ,m

such that F (T, Pk ) → 0 as k → ∞. If T ∈ Rm (Rn ), we may choose απ ∈ Z, so that Pk ∈ Pm (Rn ).

Proof. Applying the deformation theorem with εk ց 0, we get

T − Pk = ∂Rk + Sk ,

where

M(Rk ) ≤ cεk M(T ) → 0

and
M(Sk ) ≤ cεk M(∂T ) → 0,

and therefore

F (T, Pk ) ≤ cεk M(T ) + M(∂T ) → 0
as k → ∞.

4.103 Rectifiability and compactness theorems

We say that a subset D ⊂ X of a metric space X is ε-dense, ε > 0, if
[
X= B(x, ε).
x∈D

Furthermore, X is totally bounded if, for every ε > 0 there exists a finite ε-dense set D ⊂ X.
Finally, recall that a metric space is compact if and only if it is complete and totally bounded.
We define the flat norm F (T ) = F (T, 0) for T ∈ Dm (Rn ), that is

F (T ) = inf{M(S) + M(R) : T = ∂R + S, R ∈ Dm+1 (Rn ), S ∈ Dm (Rn )}.

Thus
F (Ti ) → 0 ⇒ Ti → 0.
The following converse holds for (integer multiplicity) rectifiable currents.
78 Geometric Measure Theory

Theorem 4.104. Suppose that T0 , Tj ∈ Rm (Rn ), with supp Tj ⊂ K ⊂ Rn and K compact, and
that
sup{M(Tj ) + M(∂Tj )} < ∞.
j

Then
Tj → T0 ⇐⇒ F (Tj − T0 ) → 0.
Before the proof we first established the totally boundedness property: For every ε > 0 and
M > 0 there exists N = N (n, m, ε, M, K) ∈ N such that
N
[
n
(4.105) {T ∈ Rm (R ) : supp(T ) ⊂ K, M(T ) + M(∂T ) < M } ⊂ BF (Rj , ε)
j=1

for some R1 , . . . , RN ∈ Rm (Rn ), where

BF (R, ε) = {T ∈ Rm (Rn ) : F (T − R) ≤ ε}.

Let δ > 0 to be fixed later. By the deformation theorem there are P, S ∈ Rm (Rn ), R ∈ Rm+1 (Rn )
such that
T − P = ∂R + S,
where
X
(4.106) P = απ [π], απ ∈ Z,
π∈Lδ,m

X
(4.107) M(P ) = |απ |δm ≤ cM(T ) ≤ cM,
π

√
(4.108) supp(P ) ⊂ {x : dist(x, K) < 2δ n},

(4.109) M(R) ≤ cδM, M(S) ≤ cδM.

Then
F (T − P ) ≤ M(S) + M(R) ≤ 2cδM < ε
by choosing δ < ε/(2cM ). On the other hand, there can be only finitely many, say at most N ,
currents P satisfying (4.106)-(4.108), where N depends only on K and δ = δ(n, m, ε, M ). This
proves the local boundedness property (4.105).

Proof of Theorem 4.104. We need to prove the implication Tj → T ⇒ F (Tj − T ) → 0. First

we claim that the total boundedness property (4.105) implies that there is a subsequence Tij F -
converging to T0′ ∈ Rm (Rn ), i.e. F (Tij − T0′ ) → 0. Since {Tj } belongs to a union of finitely many
F -balls of radius 1, there exists R ∈ Rm (Rn ) such that BF (R, 1) contains infinitely many Tj ’s,
call them T1,j . Similarly, there exists another R ∈ Rm (Rn ) such that BF (R, 1) contains infinitely
many T1,j ’s, call them T2,j , and so on. Then the diagonal sequence (Tj,j ) is a Cauchy sequence
with respect to the F -norm. Passing to a subsequence, still denoted by (Tj,j ), we may assume that
∞
X
F (Tj,j − Tj−1,j−1) < ∞,
j=2
Fall 2016 79

where
Tj,j − Tj−1,j−1 = ∂Rj + Sj , Rj ∈ Rm+1 (Rn ), Sj ∈ Rm (Rn ),
with
∞
X  
M(Rj ) + M(Sj ) < ∞.
j=2

By Lemma 4.80
∞
X ∞
X
n
Rj ∈ Rm+1 (R ), Sj ∈ Rm (Rn )
j=2 j=2

as limits of Cauchy sequences in the mass norm. Then for

∞
X ∞
X
T0′ := T1,1 + Sj + ∂ Rj ,
j=2 j=2

it holds that
∞
X  
F (T0′ − Tj,j ) ≤ M(Rj ) + M(Sj ) → 0
i=j+1

as j → ∞. Hence Tj → T0′ , and so T0′ = T0 and F (Tj,j − T0 ) → 0. Supposing that there exists
a subsequence (Tji ) such that lim inf F (Tji − T0 ) > 0, we get a contradiction by repeating the
argument above. Hence F (Tj − T0 ) → 0.

Next we prove a rectifiability result whose proof uses the Besicovitch-Federer structure theorem
(Theorem 2.55) on the characterization of purely m-unrectifiable set in terms of projections onto
m-dimensional subspaces of Rn . First we state the following consequence of Theorem 2.55, the
proof is left as an exercise.

Lemma 4.110. Let E ⊂ Rn be Hm -measurable, with Hm (E) < ∞. Suppose that E is purely V m-
m
unrectifiable. Then we can choose the coordinate axis such that H (PI E) = 0 for all I ∈ (n, m).

From Theorem 4.68 we then obtain the following:

Theorem 4.111. Let E be as above and let T ∈ Dm (Rn ), with supp T compact and M(T ) +
M(∂T ) < ∞. Then µT (E) = 0.

Theorem 4.112 (Rectifiability theorem). Let T ∈ Dm (Rn ), with supp T compact and M(T ) +
M(∂T ) < ∞. If

∗m µT B̄(x, r)
Θ (µT , x) = lim sup >0
rց0 ωm r m
for µT -a.e. x ∈ Rn , then there exist a countably m-rectifiable Borel set E and a Borel function
θ : Rn → [0, +∞] such that θ = 0 on Rn \ E,
Z
T (ω) = ω, T~ θ dHm
E

for ω ∈ D m (Rn ) and, for Hm -a.e. x ∈ E, T~ (x) is a unit m-vector associated with the approximate
(µT , m)-tangent space Vx ∈ G(n, m) of E at x.
80 Geometric Measure Theory

Proof. It follows from TheoremV 4.29 that there exists a Radon measure µT on Rn and a µT -
measurable mapping T~ : R → m (Rn ) such that |T~ (x)| = 1 for µT -a.e. x ∈ Rn and
n

Z
T (ω) = hT~ (x), ω(x)idµT (x) ∀ω ∈ D m (Rn ).

Then main steps of the proof then are to establish that

(1) the set {x ∈ Rn : Θ∗m (µT , x) > 0} is countably m-rectifiable,

(2) µT ≪ Hm xE, and that

V
(3) T~ : E → m (Rn ) is a Borel orientation, i.e. T~ (x) = τ1 ∧ · · · ∧ τm for Hm -a.e. x ∈ E, where
τ1 , . . . , τm is an orthonormal basis of the approximate (µT , m)-tangent space of E at x.

Using the Besicovitch covering theorem we can compare an arbitrary Radon measure µ and Hm
(see e.g. [Si, p. 26], [Ma, 2.13], [Ho, 5.23]). Indeed, for all A ⊂ Rn and λ > 0


(4.113) Hm {x ∈ A : Θ∗m (µ, x) > λ} ≤ λ−1 µ(A) ≤ λ−1 µ(Rn )

and


(4.114) µ {x ∈ A : Θ∗m (µ, x) < λ} ≤ λHm (A).

From (4.113) and the assumption M(T ) + M(∂T ) < ∞ we then obtain


(4.115) Hm {x ∈ Rn : Θ∗m (µT , x) = ∞} = 0 = Hm {x ∈ Rn : Θ∗m (µ∂T , x) = ∞}.

This together with Theorem 4.68 then implies that

(4.116) µT {x ∈ Rn : Θ∗m (µT , x) = ∞} = 0 = µ∂T {x ∈ Rn : Θ∗m (µ∂T , x) = ∞}.

Notice that since projections PI are 1-Lipschitz, Hm (PI A) = 0 if Hm (A) = 0. Define

E = {x ∈ Rn : Θ∗m (µT , x) > 0}.

By (4.113), E has a σ-finite Hm -measure. To prove that E is countably m-rectifiable, let P ⊂ E

be purely m-unrectifiable. By Lemma 4.110 and Theorem 4.111, we get µT (P ) = 0, and hence
Hm (P ) = 0 by (4.113). So, E is countably m-rectifiable. By the definition of E, µT (Rn \ E) = 0,
hence T = T xE, that is Z
T (ω) = hω, T~ idµT , ∀ω ∈ D m (Rn ).
E

By (4.114) and (4.116), we then conclude that µT ≪ Hm xE, and therefore there exists, by the
Radon-Nikodym theorem, a Borel function θ : E → [0, +∞] such that
Z
µT (A) = θdHm
A

for every Borel set A ⊂ Rn . Hence

Z
T (ω) = hω, T~ iθdHm , ∀ω ∈ D m (Rn ).
E
Fall 2016 81

It remains to show that T~ is associated with the approximate (µT , m)-tangent space Vx of E at
Hm -a.e. x ∈ E. The approximate (µT , m)-tangent space at x is the m-dimensional subspace
Vx ∈ G(n, m) such that for every δ > 0


lim r −m µT E ∩ B̄(x, r) \ {y : dist(y − x, Vx ) < δ|y − x|} = 0.
rց0

We write E as a disjoint union

∞
G
E = E0 ⊔ Ej ,
j=1

where Hm (E0 ) = 0 and Ej ⊂ Mj , Mj being an m-dimensional C 1 -smooth submanifold of Rn .

Then, in fact, Vx = Txm E = Tx Mj for Hm -a.e. x ∈ E ∩ Mj . For a ∈ Rn and λ > 0 we define (as in
Remark 2.51 (c)) ηa,λ : Rn → Rn ,
y−a
ηa,λ (y) = .
λ
Then for Hm -a.e. x ∈ E, we have

λ−m ηx,λ♯ (Hm xE) ⇀ Hm xVx

as λ ց 0. This follows from the area formula since

−1

−1


λ−m ηx,λ♯ (Hm xE)(A) := λ−m (Hm xE) ηx,λ A = λ−m Hm E ∩ ηx,λ A
Z


= JηEx,λ (y)dHm (y) = Hm ηx,λ E ∩ A
−1
E∩ηx,λ A

= (Hm xηx,λ )(A).

More generally, for Hm -a.e. x ∈ E and for every Borel function ψ : E → [0, +∞], with
Z
ψdHm < ∞,
E

we get
λ−m ηx,λ♯ (Hm xψ) ⇀ ψ(x)Hm xVx ,
that is,
Z Z
−m


(4.117) λ ϕ ηx,λ (y) ψ(y)dHm (y) → ψ(x) ϕdHm
E Vx

for every ϕ ∈ C0 (Rn ). We apply (4.117) with ψ(y) = hdxI , T~ (y)iθ(y) to obtain
Z Z
−m

I ~ m I ~
λ ωI ηx,λ (y) dx , T (y) θ(y)dH (y) → θ(x) dx , T (x) ωI (y)dHm (y)
E Vx
Z
= θ(x) ωI (y)dxI , T~ (x) dHm (y)
Vx

for Hm -a.e. x ∈ E and for all component functions ωI of

X
ω= ωI dxI ∈ D m (Rn ).
V
I∈ (n,m)
82 Geometric Measure Theory

Next we observe that λ−m is the Jacobian determinant of the mapping ηx,λ |Mj : Mj → ηx,λ Mj
between m-dimensional C 1 -smooth submanifolds. Hence
∗



ηx,λ ω|ηx,λ Mj (y) = λ−m (ω|Mj ) ηx,λ (y) ,

and therefore
Z Z
∗
ηx,λ ω(y), T~ (y) m
dH (y) → θ(x) ω(y), T~ (x) dHm (y).
E∩Mj Vx

We define Sx ∈ Dm (Rn ) by
Z
Sx (ω) = θ(x) ω(y), T~ (x) dHm (y), ∀ω ∈ D m (Rn ),
Vx

and claim that ∂Sx = 0. For that purpose let ω ∈ D m−1 (Rn ) and R > 0 such that supp ω ⊂
B n (0, R). Then supp ηx,λ
∗ ω ⊂ B n (x, λR), and therefore

Z  ^ →

∂ηx,λ♯ T (ω) = ηx,λ♯ ∂T (ω) = ω ◦ ηx,λ , m dηx,λ ∂T dµ∂T


≤ λ1−m kωk∞ µ∂T B(x, λR) → 0

as λ → 0 if Θ∗m (µ∂T , x) < ∞ which happens for Hm -a.e. x ∈ E by (4.115). We have proven that

ηx,λ♯ T → Sx and ∂ηx,λ♯ T → 0

as λ → 0, and therefore ∂Sx = 0. Finally, to show that T~ (x) orients Vx , we may assume V without
m
loss of generality that Vx = R × {0}. For j ∈ {m + 1, . . . , n} and I = (i1 , . . . , im−1 ) ∈ (n, m − 1),
let
ω(y) = y j ϕ(y)dy I = y j ϕ(y)dy i1 ∧ · · · ∧ dy im−1 ,
where ϕ ∈ C0∞ (Rn ) is arbitrary. Then


dω = d y j ϕ(y) ∧ dy I = ϕ(y)dy j ∧ dy I + y j dϕ ∧ dy I

and y j ≡ 0 in Vx = Rm × {0}, and hence

Z
0 = ∂Sx (ω) = Sx (dω) = θ(x) ϕ(y) T~ (x), dy j ∧ dy I dHm (y)
Vx
Z
= θ(x) ϕ(y) T~ (x), ej ∧ e1i ∧ · · · ∧ eim −1 dHm (y).
Vx

Since ϕ ∈ C0∞ (Rn ) is arbitrary, we conclude that

hT~ (x), ej ∧ eI i = 0
V
for every j ∈ {m + 1, . . . , n} and I = (i1 , . . . , im−1 ) ∈ (n, m − 1). As |T~ (x)| = 1, this proves that

T~ (x) = ±e1 ∧ · · · ∧ em .
Fall 2016 83

Remark 4.118. 1. We notice that the approximate (µT , m)-tangent space Vx coincides with
the approximate tangent space Txm E for Hm -a.e. x ∈ E. Hence T = τ (E, θ, T~ ).

2. The compactness assumption on supp T is not necessary. It suffices only to assume that
MW (T ) + MW (∂T ) < ∞ for all W ⋐ Rn .

The next lemma is a step towards the compactness theorem.

Lemma 4.119. Suppose Tj ∈ Rm (Rn ), ∂Tj ∈ Rm−1 (Rn ), supp Tj ⊂ K, with K ⊂ Rn compact,
and that


sup M(Tj ) + M(∂Tj ) < ∞.
j

If Tj → T , then T ∈ Rm (Rn ).

Proof. We prove the lemma by induction on m. The case m = 0 is trivial. Suppose that the lemma
holds for rectifiable (m − 1)-currents with integer multiplicity. Let Tj ∈ Rm (Rn ), Tj → T , be a
sequence satisfying the assumptions.
First we prove by using Theorem 4.112 that T is a rectifiable m-current. For that purpose we
will show that Θ(µT , x) > 0 for µT -a.e. x ∈ Rn . For every x ∈ Rn , let ρx be the 1-Lipschitz
function ρx (y) = |y − x|. By Theorem 4.88 (3),

(4.120) hTj , ρx , ti ∈ Rm−1 (Rn )

for a.e. t ∈ R. Clearly,

hTj , ρx , ti = (∂Tj )x Rn \ B̄(x, t) − ∂ Tj x(Rn \ B̄(x, t))



(4.121) → (∂T )x Rn \ B̄(x, t) − ∂ T x(Rn \ B̄(x, t))
= hT, ρx , ti.

Theorem 4.88 (2) and Fatou’s lemma imply that, for all δ > 0,
Z ∞ Z ∞
lim inf M(hTj , ρx , ti)dt ≤ lim inf M(hTj , ρx , ti)dt ≤ lim inf M(Tj ),
δ j→∞ j→∞ δ j→∞

and similarly,
Z ∞ Z ∞
lim inf M(∂hTj , ρx , ti)dt = lim inf M(h∂Tj , ρx , ti)dt
δ j→∞ δ j→∞
Z ∞
≤ lim inf M(h∂Tj , ρx , ti)dt
j→∞ δ
≤ lim inf M(∂Tj ).
j→∞

Hence for a.e. t ∈ R there exists a subsequence such that

(4.122) sup M(hTji , ρx , ti) + M(∂hTji , ρx , ti) < ∞.
ji

The induction hypothesis together with (4.120), (4.121), and (4.122) imply that

(4.123) hT, ρx , ti ∈ Rm−1 (Rn )

84 Geometric Measure Theory

for a.e. t. On the other hand, since ∂Tj → ∂T , we get from the induction hypothesis that

(4.124) (∂T )xB(x, t) ∈ Rm−1 (Rn ),

and since


(4.125) hT, ρx , ti = ∂ T xB(x, t) − (∂T )xB(x, t)

for a.e. t by Theorem 4.87 (1), we obtain

(4.126) ∂ T xB(x, t) = hT, ρx , ti + (∂T )xB(x, t) ∈ Rm−1 (Rn )

for a.e. t.
Next we want to reduce the proof to the case ∂T = 0. Combining Example 4.74 (2) and (3)
we conclude that the cone over a rectifiable current with integer multiplicity is a rectifiable current
with integer multiplicity, that is
 

S ∈ Rd (Rn ) ⇒ 0 ⊳ S = h♯ [0, 1] × S ∈ Rd+1 (Rn ), h(t, x) = tx.

Hence 0 ⊳ ∂T ∈ Rm (Rn ) since ∂T ∈ Rm−1 (Rn ) by the induction hypothesis. We also notice that
(by the homotopy formula)


∂ 0 ⊳ (∂T ) − T = ∂(0 ⊳ ∂T ) − ∂T = ∂T − 0 ⊳ ∂(∂T ) − ∂T = 0.

Hence we may assume without loss of generality that ∂T = 0. Indeed, otherwise we may consider
the sequence T̃j = Tj − 0 ⊳ ∂Tj ∈ Rm (Rn ), with properties

T̃j = Tj − 0 ⊳ ∂Tj → T̃ := T − 0 ⊳ ∂T,

∂ T̃j = 0,
M(T̃j ) ≤ M(Tj ) + M(∂Tj ).

Define, for a fixed x ∈ Rn ,

f (r) = µT B̄(x, r) .
Using the assumption ∂T = 0 we obtain from Theorem 4.87 (3) that





(4.127) M ∂ T xB(x, r) = M hT, ρx , ri ≤ lim inf h−1 f (r + h) − f (r) = f ′ (r)
hց0

for a.e. r > 0. Suppose then that Θ∗m (µT , x) < η < 1, so that
f (s)
lim sup <η
s→0 ωm sm
and that Z δ
1 d  1/m  1
f (r) dr ≤ f 1/m (δ) ≤ ωm
1/m 1/m
η
δ 0 dr δ
for sufficiently small δ > 0. Hence we have
d  1/m  1
f (r) = m f −1 (r)f ′ (r) ≤ 2ωm
1 m 1/m 1/m
η ,
dr
or equivalently
m−1
(4.128) f ′ (r) ≤ 2mωm
1/m 1/m
η f m (r)
Fall 2016 85

for all r in a subset of [0, δ] of positive m1 -measure. Suppose from now on that m > 1 (see
Remark 4.134
for the case m = 1). By the isoperimetric inequality (Theorem 4.99) applied to
∂ T xB(x, r) ∈ Rm−1 (R ), there exists Sr ∈ Rm (Rn ) with the properties
n


∂Sr = ∂ T xB(x, r)
and

m−1
m−1
M(Sr ) m ≤ CM ∂(T xB(x, r) ≤ cη 1/m M T xB̄(x, r) m ,
where also (4.127) and(4.128) were used. Thus there exists a sequence ri ց 0 such that


∂Sri = ∂ T xB̄(x, ri ) ∈ Rm−1 (Rn )
and 1

M(Sri ) ≤ cη m−1 M T xB̄(x, ri ) .
Let then C ⊂ {x : Θ∗m (µT , x) < η} be compact. By Vitali’s covering theorem for the Radon
measure µT , we find, for all ̺ > 0, disjoint balls Bj̺ = B̄(xj , rj ) such that xj ∈ C, rj < ̺,
[ ̺

(4.129) µT C \ Bj = 0,
j

(4.130) Bj̺
⊂ {x : dist(x, C) < ̺},
1

(4.131) M(Sj̺ ) ≤ cη m−1 M T xBj̺
for some Sj̺ ∈ Rm (Rn ) with


∂Sj̺ = ∂ T xBj̺ .
By the homotopy formula, with h(t, x) = tx + (1 − t)xj , we then have


Sj̺ − T xBj̺ = ∂ xj ⊳ (Sj̺ − T xBj̺ )
and hence by (4.46) and (4.131)



Sj̺ − T xBj̺ (ω) = xj ⊳ (Sj̺ − T xBj̺ ) (dω)
(4.46)

≤ ̺M Sj̺ − T xBj̺ kdωk∞
1
≤ c̺(η m−1 + 1)M(T xBj̺ )kdωk∞ .
Since the balls Bj̺ are disjoint and M(T ) < ∞, we get
X ̺

(4.132) Sj − T xBj̺ → 0 as ̺ → 0,
j

and so X X
T xC = lim T xBj̺ = lim Sj̺
̺→0 ̺→0
j j
by (4.129), (4.130), and (4.132). It then follows that
X ̺
X
µT (C) = M(T xC) ≤ lim inf M Sj ≤ lim inf M(Sj̺ )
̺→0 ̺→0
j j
1 X
1 X
≤ lim inf cη m−1 M T xBj̺ ≤ lim inf cη m−1 µT (Bj̺ )
̺→0 ̺→0
j j
1 [

= lim inf cη m−1 µT Bj̺
̺→0
j
1
= cη m−1 µT (C).
86 Geometric Measure Theory

1
If cη m−1 < 1, we obtain µT (C) = 0. Hence

Θ∗m (µT , x) > 0

for µT -a.e. x ∈ Rn . By Theorem 4.112, T = τ (E, θ, T~ ), where E ⊂ Rn is a countably m-rectifiable

Borel set and θ : E → [0, ∞] is a Borel function such that
Z
T (ω) = hω, T~ iθ dHm , ∀ω ∈ D m (Rn ).
E

It remains to show that θ is integer valued. Then Hm (E) < ∞ since

Z
M(T ) = θ dHm < ∞.
E

As in the proof of Theorem 4.112, we have

ηx,r♯ T → θ(x)[Vx ] as r → 0

for Hm -a.e. x ∈ E, where ηx,r (y) = (y − x)/r and Vx = Txm E is the approximate (µT , m)-tangent
space of E at x. Fixing such x, we may assume that

Vx = Rm × {0} = Rm .

Let P : Rn → Rm be the projection P (x, y) = x. Since

P♯ ∂(ηx,r♯ Tj ) = P♯ ηx,r♯ (∂Tj ) ∈ Rm−1 (Rn )

by the assumption ∂Tj ∈ Rm−1 (Rn ) and Example 4.74 (3), and since



P♯ ∂(ηx,r♯ Tj ) → P♯ ∂(ηx,r♯ T ) ,

we have


∂P♯ ηx,r♯ T = P♯ ∂(ηx,r♯ T ) ∈ Rm−1 (Rn )
by the induction hypothesis. We conclude (see Lemma 4.133 below) that


P♯ ηx,r♯ T ∈ Rm (Rm ).

By Theorem 4.65 there exist integer valued functions gr ∈ BV (Rm ) such that
Z Z


P♯ ηx,r♯ T (ω) = ϕgr dHm = hω, e1 ∧ · · · ∧ em igr dHm
Rm Rm

for ω = ϕdx1 ∧ · · · ∧ dxm ∈ D m (Rm ). But

P ∗ (ω) = (ϕ ◦ P )dx1 ∧ · · · ∧ dxm ∈ D m (Rn ),

and so Z Z
m ∗ m ∗
ϕgr dH = ηx,r♯ T (P ω) → θ(x)[R ](P ω) = θ(x) ϕdHm
Rm Rm
as r → 0. Since all gr ’s are integer valued, we conclude that θ(x) ∈ Z which proves the lemma.

Lemma 4.133. If S ∈ Dm (Rm ), supp S is compact, and ∂S ∈ Rm−1 (Rm ), then S ∈ Rm (Rm ).
Fall 2016 87

Proof. Since 0 ⊳ S ∈ Dm+1 (Rm ) = {0}, we have

 

s = 0 ⊳ ∂S + ∂(0 ⊳ S) = 0 ⊳ ∂S = h♯ [0, 1] × ∂S ∈ Rm (Rm )

by Example 4.74 (2), (3).

Remark 4.134. In the case m = 1 (in Lemma 4.119), we have

∂ T xB(x, t) ∈ R0 (Rn )


for a.e. t by (4.126). Assuming ∂ T xB(x, t) 6= 0, we get a contradiction



1 ≤ M ∂ T xB(x, r) ≤ f ′ (r) ≤ 4η < 1


by (4.127) and (4.128), with η < 1/4. Hence ∂ T xB(x, t) = 0 for a.e. r and we may take Sr = 0
(and thus Sj̺ = 0 in (4.131)). It follows that µT (C) = 0.

Theorem 4.135 (Boundary rectifiability theorem). Let T ∈ Rm (Rn ) with supp T compact and
M(∂T ) < ∞. Then ∂T ∈ Rm−1 (Rn ).

Proof. By the polyhedral approximation theorem 4.102, there exists a sequence Pk ∈ Pm (Rn ) of
the form X
Pk = απ [π], απ ∈ Z,
π∈Lεk ,m

such that Pk → T and ∂Pk → ∂T as k → ∞. Since ∂Pk ∈ Rm−1 (Rn ), and ∂(∂Pk ) = 0, we conclude
from Lemma 4.119 that
∂T = lim ∂Pk ∈ Rm−1 (Rn ).
k

Theorem 4.136 (Compactness theorem). Suppose that Tj ∈ Rm (Rn ), with supp(Tj ) ⊂ K and
K ⊂ Rn compact, and that

sup M(Tj ) + M(∂Tj ) < ∞.
j

Then there exist a subsequence Tji and T ∈ Rm (Rn ) such that Tji → T .

Proof. By Theorem 4.135, ∂Tj ∈ Rm−1 (Rn ) and the claim then follows from Lemma 4.119.

5 Mass minimizing currents

In this final section we discuss briefly mass (area) minimizing currents that provide a tool to attack
the general Plateau problem. In particular, we prove the existence of a mass minimizing integer
multiplicity rectifiable m-current given a rectifiable (m − 1)-cycle (of integer multiplicity).

Definition 5.1. An m-current S ∈ Rm (Rn ) is mass minimizing if

M(S) ≤ M(T )

for every T ∈ Rm (Rn ) with ∂T = ∂S.

88 Geometric Measure Theory

Theorem 5.2 (Existence theorem). If T ∈ Rm−1 (Rn ) with ∂T = 0 and supp(T ) is compact, there
exists S0 ∈ Rm (Rn ) such that ∂S0 = T and
M(S0 ) = min{M(S) : S ∈ Rm (Rn ), ∂S = T }.
Hence S0 is mass minimizing.
Proof. By the isoperimetric inequality (Theorem 4.99) there is a m-current S ∈ Rm (Rn ) with
supp(S) compact such that ∂S = T and
M(S)m/(m+1) ≤ Cn,m M(T ) < ∞.
Hence the set S = {S ∈ Rm (Rn ) : ∂S = T } is non-empty (in fact, also 0 ⊳ T would do) and we
may find a minimizing sequence Sj ∈ Rm (Rn ), ∂Sj = T , such that
M(Sj ) → I := inf{M(S) : S ∈ Rm (Rn ), ∂S = T }.
Let R > 0 be so large that supp(T ) ⊂ B̄(0, R) and let f : Rn → B̄(0, R),
(
Rx/|x|, if |x| ≥ R;
f (x) =
x, if |x| ≤ R,
be the 1-Lipschitz retract onto B̄(0, R). Then M(f♯ S) ≤ M(S) for every S ∈ Dm (Rn ) by (4.45).
Hence we may assume that supp(Sj ) ⊂ B̄(0, R) for every j. Moreover
 
sup M(Sj ) + M(∂Sj ) = sup M(Sj ) + M(T ) < ∞.
j j

By the compactness theorem (Theorem 4.136) there exists a subsequence Sji and S0 ∈ Rm (Rn )
such that Sji → S0 . Then T = ∂Sji → ∂S0 , and therefore ∂S0 = T and
I ≤ M(S0 ) ≤ lim inf M(Sji ) = I.
ji →∞

References
[EG] L. Evans and R. Gariepy, Measure theory and fine properties of functions, CRC Press,
1992.
[Fe] H. Federer, Geometric Measure Theory, Springer, 1969.
[Ho] I. Holopainen, Moderni reaalianalyysi, Syksy 2005.
[Ho2] I. Holopainen, Johdatus differentiaaligeometriaan, Syksy 2004.
[Ho3] I. Holopainen, Minimal surfaces.
[Lee] J. M. Lee, Introduction to smooth manifolds, Springer, 2003.
[LY] F. Lin and X. Yang, Geometric Measure Theory: An Introduction, International Press,
2002.
[Ma] P. Mattila, Geometry of sets and measures in Euclidean spaces, Cambridge University
Press, 1995.
[Mo] F. Morgan, Geometric Measure Theory, A Beginner’s Guide, Academic Press, 1987.
[Si] L. Simon, Lectures on Geometric Measure Theory, Australian National University, 1983.
Fall 2016 89

6 Appendix
6.1 Proof of Riesz’ representation theorem 1.62
First we prove the auxiliary lemma (Lemma 1.63).

Proof of Lemma 1.63. (a) Write δ = dist(K, ∂V ). Because K is compact, it follows that δ > 0.
Then the function
2

f (x) = max 0, 1 − dist(x, K)
δ
satisfies the conditions of part (a).

(b) For every x ∈ K there exists a ball B(x, rx ), with B(x, 2rx ) ⊂ Vj for some j. Because K is
compact, it can be covered by finitely many such balls, i.e.

k
[
K⊂ B(xi , rxi ).
i=1

Let Aj be the union of those closed balls B̄(xi , rxi ) for which B(xi , 2rxi ) ⊂ Vj . Then
m
[
Aj ⊂ Vj and K ⊂ Aj .
j=1

By part (a) we choose functions gj ∈ C0 (Rn ) s.t.

χAj ≤ gj ≤ 1 and supp(gj ) ⊂ Vj .

Then define

h1 = g1 ,
h2 = (1 − g1 )g2 ,
..
.
hm = (1 − g1 ) · · · (1 − gm−1 )gm .

Then clearly
0 ≤ hj ≤ 1 and supp(hj ) ⊂ Vj .
P
Induction with respect to m shows that mj=1 hj = 1 − (1 − g1 ) · · · (1 − gm ). Furthermore,

m
X
χK ≤ hj ≤ 1,
j=1

Pm
because if x ∈ K, then x ∈ Aj for some j and consequently 1 − gj (x) = 0 and j=1 hj (x) = 1.

Remark 6.2. In the case of a locally compact Hausdorff space Lemma 1.63 is (a version of)
Urysohn’s lemma.
90 Geometric Measure Theory

Proof of Riesz’ representation theorem 1.62. Define

µ̃(∅) = 0

and set
µ̃(V ) = sup{Λ(f ) : f ∈ C0 (Rn ), 0 ≤ f ≤ 1 and supp(f ) ⊂ V }
for every open set V ⊂ Rn . Then it follows from the definition that

(6.3) 0 ≤ µ̃(V1 ) ≤ µ̃(V2 ),

if V1 , V2 ⊂ Rn are open and V1 ⊂ V2 .

Next define

(6.4) µ̃(A) = inf{µ̃(V ) : A ⊂ V ⊂ Rn , V open}

for all A ⊂ Rn . We show that µ̃ is a metric outer measure. Then all Borel-sets of Rn will be
µ̃-measurable by Theorem 1.18.
1. Monotonicity
µ̃(A1 ) ≤ µ̃(A2 ), if A1 ⊂ A2 ,
follows directly from (6.3) and the definition (6.4).

2. We prove first subadditivity for open sets. In other words, if Vj ⊂ Rn , j ∈ N, are open, then
∞
[ ∞
X


(6.5) µ̃ Vj ≤ µ̃(Vj ).
j=1 j=1
S∞
To prove this let f ∈ C0 (Rn ) s.t. 0 ≤ f ≤ 1 and supp(f ) ⊂ j=1 Vj . Because of the
compactness of supp(f )
m
[
supp(f ) ⊂ Vj .
j=1

Write K = supp(f ). Lemma 1.63, part (b), implies that there exist functions hj ∈ C0 (Rn )
with
m
X
0 ≤ hj ≤ 1, supp(hj ) ⊂ Vj and χK ≤ hj ≤ 1.
j=1

Then
m
X
f= hj f,
j=1

supp(hj f ) ⊂ Vj and 0 ≤ hj f ≤ 1 ∀ j = 1, . . . , m,

and hence
m
X m
X ∞
X
Λ(f ) = Λ(hj f ) ≤ µ̃(Vj ) ≤ µ̃(Vj ).
j=1 j=1 j=1


Taking sup over all “admissible” functions f in the definition of µ̃ ∪j Vj we obtain (6.5). Let
then Aj ⊂ Rn , j ∈ N, be arbitrary sets. Fix ε > 0 and choose open sets Vj ⊂ Rn s.t. Aj ⊂ Vj
and
µ̃(Vj ) ≤ µ̃(Aj ) + ε/2j .
Fall 2016 91

Then
∞
[ ∞
[
Aj ⊂ Vj ,
j=1 j=1

and hence by monotonicity and (6.5) we obtain

∞
[ ∞
[ ∞ ∞


X X
µ̃ Aj ≤ µ̃ Vj ≤ µ̃(Vj ) ≤ µ̃(Aj ) + ε,
j=1 j=1 j=1 j=1

which implies subadditivity for all sets by letting ε → 0. We have proved that µ̃ is an outer
measure.
3. Let V1 , V2 ⊂ Rn be open sets and dist(V1 , V2 ) > 0. Let further fj ∈ C0 (Rn ) s.t. 0 ≤ fj ≤ 1
and supp(fj ) ⊂ Vj , j = 1, 2. Then 0 ≤ f1 + f2 ≤ 1 and supp(f1 + f2 ) ⊂ V1 ∪ V2 , and hence
Λ(f1 ) + Λ(f2 ) = Λ(f1 + f2 ) ≤ µ̃(V1 ∪ V2 ).
Taking sup over all admissible functions f1 and f2 we obtain
(6.5)
(6.6) µ̃(V1 ) + µ̃(V2 ) ≤ µ̃(V1 ∪ V2 ) ≤ µ̃(V1 ) + µ̃(V2 ).
Let then A1 , A2 ⊂ Rn be arbitrary sets with dist(A1 , A2 ) > 0. Fix ε > 0 and choose an open
set V ⊂ Rn s.t. A1 ∪ A2 ⊂ V and
µ̃(V ) ≤ µ̃(A1 ∪ A2 ) + ε.
Choose then open sets Vj ⊂ Rn , j = 1, 2, s.t. Aj ⊂ Vj and dist(V1 , V2 ) > 0. (We may choose
for instance Vj = {x ∈ Rn : dist(x, Aj ) < 31 dist(A1 , A2 )}.) Now Aj ⊂ Vj ∩ V and
dist(V1 ∩ V, V2 ∩ V ) > 0,
and hence by (6.6)
µ̃(A1 ) + µ̃(A2 ) ≤ µ̃(V1 ∩ V ) + µ̃(V2 ∩ V )
= µ̃(V ∩ (V1 ∪ V2 ))
≤ µ̃(V )
≤ µ̃(A1 ∪ A2 ) + ε.
Letting now ε → 0 we see that µ̃ is a metric outer measure.
4. We prove next that µ̃ is locally finite: If B(x, r) ⊂ Rn , then choose f0 ∈ C0 (Rn ) s.t. 0 ≤ f0 ≤ 1
and
χB(x,r) ≤ f0 .


Then f ≤ f0 for all functions f admissible in the definition of µ̃ B(x, r) . Because Λ(f0 −f ) ≥
0, then
Λ(f ) ≤ Λ(f0 ).
Taking sup over all such functions f we obtain


µ̃ B(x, r) = sup Λ(f ) ≤ Λ(f0 ) < ∞.
f

Corollary 1.32 implies that

µ = µ̃| Bor(Rn )
is a Radon measure.
92 Geometric Measure Theory

5. We must still show that Z

Λ(f ) = f dµ
Rn
for all f ∈ C0 (Rn ).
Let f ∈ C0 (Rn ). We may suppose that f ≥ 0, because f = f+ − f− , where f+ = max(0, f ) ∈
C0 (Rn ) and f− = max(0, −f ) ∈ C0 (Rn ). Fix ε > 0 and set for all k ∈ N


fk (x) = max (k − 1)ε, min(f (x), kε) − (k − 1)ε.

kε
(k − 1)ε
fk

Clearly 0 ≤ fk ≤ ε and fk ∈ C0 (Rn ) for all k. Because fk ≡ 0, if (k − 1)ε ≥ kf k∞ , then

m
X
f= fk
k=1

for some m ∈ N. Let K(k) = {x ∈ Rn : f (x) ≥ kε}, k ∈ N and K(0) = supp(f ). Then

(6.7) εχK(k) ≤ fk ≤ εχK(k−1) ,

and hence
m
X m Z Z m

X X

(6.8) ε µ K(k) ≤ fk dµ = f dµ ≤ ε µ K(k − 1) .
k=1 k=1 Rn Rn k=1

On the other hand, if δ > 0, then by (6.7)

1
(1 + δ)fk ≥ 1
ε
in some neighbourhood W of K(k). In particular,
1
(1 + δ)fk ≥ g
ε
for every function g ∈ C0 (Rn ) admissible in the definition of µ(W ). Thus



Λ (1 + δ)fk ≥ εµ(W ) ≥ εµ K(k) ,

and further

Λ(fk ) ≥ εµ K(k)
letting δ → 0. In the same way, fk /ε is admissible in the definition of µ(V ) for every
neighbourhood V of K(k − 1), and hence Λ(fk ) ≤ εµ(V ). Then by the definition


Λ(fk ) ≤ εµ K(k − 1) .
Fall 2016 93

Combining these inequalities we obtain

m
X m
X



(6.9) εµ K(k) ≤ Λ(f ) ≤ ε µ K(k − 1) .
k=1 k=1

The inequalities (6.8) and (6.9) imply that

Z Xm




Λ(f ) − f dµ ≤ ε µ K(k − 1) − µ K(k)
Rn k=1




= ε µ K(0) − µ K(m)


≤ ε µ supp(f ) .
| {z }
<∞

Letting ε → 0 we see that Z

Λ(f ) = f dµ
Rn
and thus µ is the desired Radon measure.
6. Finally, we prove the uniqueness of µ. Let µ1 also be a Radon-measure, for which
Z
Λ(f ) = f dµ1
Rn
for all f ∈ C0 (Rn ). Let V ⊂ Rn be open and bounded. By Lemma 1.63 there exists a
sequence fj ∈ C0 (Rn ) s.t.
0 ≤ f1 (x) ≤ f2 (x) ≤ · · · ≤ fj (x) → χV (x)
for all x ∈ Rn . By the Monotone Convergence Theorem
Z
µ1 (V ) = lim fj dµ1
j→∞

= lim Λ(fj )
j→∞
Z
= lim fj dµ
j→∞

= µ(V ).
Because Bor(Rn ) is a σ-algebra spanned by open and bounded sets, then µ1 = µ (in Bor(Rn )).

6.10 Proof of Theorem 1.67

Proof. (a) Let K ⊂ Rn be compact and V ⊂ Rn be open s.t. K ⊂ V . Choose by part (a) of
Lemma 1.63 a function f ∈ C0 (Rn ) with χK ≤ f ≤ 1 and supp(f ) ⊂ V . Then
Z
µ(V ) ≥ f dµ
Rn Z

= lim f dµk
k→∞ Rn

≥ lim sup µk (K).

k→∞

Because this holds for all open V ⊃ K, the claim in part (a) is proven.
94 Geometric Measure Theory

(b) If V ⊂ Rn is open, then let K ⊂ V compact. In the same way as above we obtain

µ(K) ≤ lim inf µk (V ).

k→∞

Because µ is a Radon measure,

µ(V ) = sup{µ(K) : K ⊂ V compact},

and part (b) is proven.

6.11 Proof of Theorem 1.69

For the proof we need the following auxiliary result.


Lemma 6.12. The norm space C0 (Rn ), k·k , where kf k = sup{|f (x)| : x ∈ Rn }, is separable, i.e.
there is a countable dense set F = {fj }∞ n n
j=1 ⊂ C0 (R ). In other words, if f ∈ C0 (R ) and ε > 0,
then kf − fj k < ε for some fj ∈ F.
Proof. (Exerc.)

Proof of Theorem 1.69. Suppose first that

(6.13) sup µk (Rn ) = A < ∞.

k

Let {fj }∞ n
j=1 be a dense set in C0 (R ). It follows from the assumption (6.13) that
Z

f1 dµk : k ∈ N
Rn

is a bounded subset of R, and hence there is a subsequence (µ1k ) of (µk ) s.t.

Z
k→∞
f1 dµ1k −−−→ a1
Rn

for some a1 ∈ R. Choose inductively for all j ≥ 2 a subsequence {µjk } of the sequence {µkj−1 } s.t.
Z
k→∞
fj dµjk −−−→ aj
Rn

for some aj ∈ R. Then the diagonal sequence {µkk }∞

k=1 satisfies
Z
(6.14) lim fj dµkk = aj
k→∞ Rn

for all j ≥ 1. Let L be the vector space spanned by the functions fj ,

m
X

L= g= λj fj : λj ∈ R, m ∈ N .
j=1

Set X
Λ(g) = λj aj ,
j
Fall 2016 95

when X
g= λj fj .
j

Then by (6.14) we see that Z

Λ(g) = lim g dµkk
k→∞ Rn

for all g ∈ L. In particular, Λ is well defined (i.e. Λ(g) is independent of the particular choice of
the linear combination for g), positive and linear functional in L. Moreover, it follows from (6.13)
that

(6.15) |Λ(g)| ≤ A kgk

for all g ∈ L. If f ∈ C0 (Rn ) is arbitrary, then choose a sequence (hj ), hj ∈ L, s.t.

j→∞
kf − hj k −−−→ 0,

and set
Λ(f ) = lim Λ(hj ).
j→∞

Then it follows from (6.15) that Λ is well defined in C0 (Rn ) (Λ(f ) independent of the choice of the
sequence (hj )) and (6.15) holds for all g ∈ C0 (Rn ). Furthermore, Λ is a positive linear functional
in C0 (Rn ). In fact, if f ≥ 0 and kf − hj k → 0, then lim inf j→∞(min hj ) ≥ 0, and hence
Z
Λ(f ) = lim lim hj dµkk ≥ 0,
j→∞ k→∞ Rn

R
because hj dµkk ≥ A min(0, min hj ). By Riesz’ representation theorem there exists a Radon-
measure µ s.t. Z
Λ(f ) = f dµ
Rn

for all f ∈ C0 (Rn ). We prove next that µkk ⇀ µ. Let ε > 0. For f ∈ C0 (Rn ), we choose g ∈ L such
ε
that kf − gk ≤ 2A . Then for large values of k
Z Z Z
Λ(f ) − f dµkk ≤ |Λ(f − g)| + Λ(g) − g dµkk + (g − f ) dµkk
Rn n Rn
| {zR }
≤ε

≤ Akf − gk + ε + Akf − gk
≤ 2ε.

Therefore µkk ⇀ µ. Finally, we give up the hypothesis (6.13). From the assumption that

sup µk (K) < ∞

k

for all compact K ⊂ Rn and the above argument there follows that for every m ∈ N there exists a
m−1
subsequence (µm m
k ) of (µk ) s.t. {µk : k ∈ N} ⊂ {µk : k ∈ N} and

µm m
k xB(0, m) ⇀ ν ,
96 Geometric Measure Theory

where ν m is a Radon-measure with supp(ν m ) ⊂ B̄(0, m). Then for the diagonal sequence µkk there
holds
µkk xB(0, m) ⇀ ν m
for all m ∈ N. Thus ν m xB(0, ℓ) = ν ℓ , when ℓ ≤ m. Therefore we may define a Radon-measure µ
in Rn setting

X m


µ(E) = ν 1 E ∩ B(0, 1) + ν E ∩ B(0, m) \ B(0, m − 1) , E ∈ Bor(Rn ).
m≥2

Because supp(f ) ⊂ B(0, m0 ), it follows that

Z Z Z
m0
f dµ = f dν = lim f dµkk ,
Rn Rn k→∞ Rn

and hence µkk ⇀ µ.