FUNCTIONS OF BOUNDED VARIATION & SETS OF

FINITE PERIMETER

XUWEN ZHANG

Contents
Chapter: 1 Functions of bounded variation 2
1. Definition and basic properties 2
2. Approximation and compactness 5
3. Isoperimetric inequalities 9
4. Coarea formula for BV functions 12
Chapter: 2 Sets of finite perimeter 16
5. Definition and basic properties 16
5.1. Definition 16
5.2. Basic properties 18
5.3. Lower semicontinuity of perimeter 19
5.4. Topological boundary and Gauss-Green measure 20
5.5. Regularization 23
6. Geometric variational problems 26
6.1. Geometric variational problems and sets of finite perimeter 26
6.2. Compactness for sets of finite perimeter 28
6.3. Existence of minimizers 30
7. Approximated by open sets with smooth boundary 32
8. Reduced boundary 34
8.1. Basic properties 35
8.2. De Giorgi’s structure theorem 42
8.3. Essential boundary and Federer’s theorem 47
9. First variation of perimeter 52
9.1. First variation formula 52
9.2. Stationary sets of perimeter 57
Chapter: 3 Minimizers 60
10. Perimeter minimizers 60
Appendix A. Covering theorem 65
Appendix B. Vector-valued Radon measures 65
B.1. Weak star convergence 65
B.2. Differentiation of Radon measures 66
Date: December 9, 2024.
1
 Appendix C. Rademacher’s theorem 66
Appendix D. Morse-Sard theorem 66
References 67

Chapter 1: Functions of bounded variation

• Throughout the chapter we will work in the Euclidean space R𝑛 . The
Euclidean scalar product, the corresponding Levi-Civita connection,
and the Lebesgue measure on R𝑛 are denoted respectively by ⟨·, ·⟩,
𝐷, and L 𝑛 .
• For 𝑘 ∈ N, H 𝑘 denotes the 𝑘-dimensional Hausdorff measure on
R𝑛 .
• We always use 𝑈 to denote an open subset ∫ of R . For any function 𝑓
𝑛

defined
∫ on 𝑈, we say that 𝑓 ∈ 𝐿 1 (𝑈) if 𝑈 | 𝑓 |dL 𝑛 < ∞; 𝑓 ∈ 𝐿 1𝑙𝑜𝑐 (𝑈)
if 𝐾 | 𝑓 |dL 𝑛 < ∞ for any compact 𝐾 ⊂ 𝑈.

1. Definition and basic properties

Definition 1.1 (BV functions). A (Lebesgue measurable) function 𝑓 ∈

𝐿 1 (𝑈) is said to have bounded variation in 𝑈 if
∫ 
𝑛 1 𝑛
sup 𝑓 div𝜙dL : 𝜙 ∈ 𝐶𝑐 (𝑈, R ), |𝜙| ≤ 1 < ∞.
𝑈

The set of all such functions is denoted by 𝐵𝑉 (𝑈).

The local version is as follows:
A function 𝑓 ∈ 𝐿 1𝑙𝑜𝑐 (𝑈) is said to have locally bounded variation in 𝑈 if
for each open set 𝑈 ′ ⊂⊂ 𝑈,
∫ 
𝑛 1 ′ 𝑛
sup 𝑓 div𝜙dL : 𝜙 ∈ 𝐶𝑐 (𝑈 , R ), |𝜙| ≤ 1 < ∞.
𝑈′

The set of all such functions is denoted by 𝐵𝑉𝑙𝑜𝑐 (𝑈).

Example 1.2. For 𝑓 ∈ 𝐶 1 (𝑈), integration by parts gives
∫ ∫
𝑛
𝑓 div𝜙dL = − ⟨𝐷 𝑓 , 𝜙⟩ dL 𝑛 ,
𝑈 𝑈

therefore for such a 𝑓 ,

∫  ∫
1
sup 𝑓 div𝜙dL : 𝜙 ∈ 𝐶𝑐 (𝑈, R ), |𝜙| ≤ 1 = |𝐷 𝑓 |dL 𝑛 .
𝑛 𝑛
𝑈 𝑈
2
More generally, if 𝑓 ∈ 𝑊 1,1 (𝑈) then the above integration by parts still
holds and
∫  ∫
1
sup 𝑓 div𝜙dL : 𝜙 ∈ 𝐶𝑐 (𝑈, R ), |𝜙| ≤ 1 = |grad 𝑓 |dL 𝑛 ,
𝑛 𝑛
𝑈 𝑈

where grad 𝑓 denotes the weak differential of 𝑓 ∈ 𝑊 1,1 (𝑈).

It follows immediately
Proposition 1.3. The implications hold:
1,1
𝑊 (𝑙𝑜𝑐) (𝑈) ⊂ 𝐵𝑉(𝑙𝑜𝑐) (𝑈) ⊂ 𝐿 1(𝑙𝑜𝑐) (𝑈).
Moreover, the inclusions are strict, as shown by the following example.
Exercise 1.4. Show that for a domain (bounded open set) 𝐸 ⊂ R𝑛 of class
𝐶 1 , the indicator function
(
1, if 𝑥 ∈ 𝐸,
𝜒𝐸 (𝑥) :=
0, otherwise,
belongs to 𝐵𝑉 (R𝑛 ) with
∫ 
1
sup 𝜒𝐸 div𝜙dL : 𝜙 ∈ 𝐶𝑐 (𝑈, R ), |𝜙| ≤ 1 = H 𝑛−1 (𝜕𝐸).
𝑛 𝑛
R𝑛

Theorem 1.5 (Structure theorem for 𝐵𝑉𝑙𝑜𝑐 functions). For any 𝑓 ∈ 𝐵𝑉𝑙𝑜𝑐 (𝑈),
there exists a Radon measure 𝜇 on 𝑈, and a 𝜇-measurable function 𝜎 : 𝑈 →
R𝑛 such that
(i) |𝜎(𝑥)| = 1 𝜇-a.e.;
(ii) For any 𝜙 ∈ 𝐶𝑐1 (𝑈, R𝑛 ),
∫ ∫
𝑛
𝑓 div𝜙dL = − ⟨𝜎, 𝜙⟩ d𝜇.
𝑈 𝑈

Proof. Define the linear functional 𝐿 : 𝐶𝑐1 (𝑈, R𝑛 ) → R by

∫
𝐿(𝜙) := − 𝑓 div𝜙dL 𝑛 , ∀𝜙 ∈ 𝐶𝑐1 (𝑈, R𝑛 ).
𝑈
By definition of 𝑓 we have for every open 𝑈 ′ ⊂⊂ 𝑈
|𝐿 (𝜙)| ≤ 𝐶 (𝑈 ′)∥𝜙∥ 𝐿 ∞ (𝑈 ′ ) , ∀𝜙 ∈ 𝐶𝑐1 (𝑈 ′, R𝑛 ). (1.1)
Now we extend 𝐿 to 𝐶𝑐 (𝑈, R𝑛 ):
For any 𝜙 ∈ 𝐶𝑐 (𝑈, R𝑛 ),
say spt𝜙 ⊆ 𝐾
′
for some compact 𝐾 ⊂ 𝑈, we can find some open 𝑈 ⊂ 𝑈 such that 𝐾 ⊂
𝑈 ′ ⊂⊂ 𝑈. Approximating 𝜙 by a sequence of functions 𝜙 𝑘 ∈ 𝐶𝑐1 (𝑈 ′, R𝑛 )
(𝑘 = 1, . . .) such that 𝜙 𝑘 → 𝜙 uniformly on 𝑈 ′, then define
¯
𝐿(𝜙) := lim 𝐿 (𝜙 𝑘 ).
𝑘→∞
3
Thanks to (1.1) we have for 𝑗, 𝑘 → ∞
𝐿 (𝜙 𝑗 ) − 𝐿 (𝜙 𝑘 ) = 𝐿(𝜙 𝑗 − 𝜙 𝑘 ) ≤ 𝐶 (𝑈 ′)∥𝜙 𝑗 − 𝜙 𝑘 ∥ 𝐿 ∞ (𝑈 ′ ) → 0,
so 𝐿 (𝜙 𝑘 ) is a Cauchy sequence on R and hence converges to some real
number, which we denote by 𝐿¯ (𝜙). In other words, the limit exists and is
independent of the choice of the converging sequence (so that 𝐿¯ is well-
defined).
Applying the Riesz Representation Theorem, the assertion follows. □
Remark 1.6. In measure theory, the measure 𝜇 obtained above, as a set
¯ In BV theory we
function, is the total variation of the linear functional 𝐿.
write
𝜇 =: ∥𝐷 𝑓 ∥
and call it the variation measure of 𝑓 ∈ 𝐵𝑉𝑙𝑜𝑐 (𝑈). Note that ∥𝐷 𝑓 ∥ is the
Radon measure on 𝑈 which is uniquely characterized by
∫ 
′ 𝑛 1 ′ 𝑛
∥𝐷 𝑓 ∥(𝑈 ) = sup 𝑓 div𝜙dL : 𝜙 ∈ 𝐶𝑐 (𝑈 , R ), |𝜙| ≤ 1 .
𝑈
1,1
Example 1.7. For 𝑓 ∈ 𝑊𝑙𝑜𝑐 (𝑈), and for any open 𝑈 ′ ⊂⊂ 𝑈, there holds
∫ ∫ ∫  
𝑛 𝑛 grad 𝑓
𝑓 div𝜙dL = − ⟨grad 𝑓 , 𝜙⟩ dL = , 𝜙 |grad 𝑓 |dL 𝑛 ,
𝑈 𝑈 𝑈 |grad 𝑓 |
which shows that
∥𝐷 𝑓 ∥ = |grad 𝑓 |L 𝑛 ,
and ( grad 𝑓
|grad 𝑓 | , if grad 𝑓 ≠ 0,
𝜎=
0, if grad 𝑓 = 0.
One of the most important properties of 𝐵𝑉 functions is the following.
Theorem 1.8 (Lower semi-continuity). Suppose 𝑓 𝑘 ∈ 𝐵𝑉 (𝑈) (𝑘 = 1, . . .)
and
𝑓 𝑘 → 𝑓 in 𝐿 1𝑙𝑜𝑐 (𝑈).
Then
∥𝐷 𝑓 ∥(𝑈) ≤ lim inf ∥𝐷 𝑓 𝑘 ∥(𝑈). (1.2)
𝑘→∞

Proof. For any 𝜙 ∈ 𝐶𝑐1 (𝑈, R𝑛 ) with |𝜙| ≤ 1, there holds

∫ ∫
𝑛
𝑓 div𝜙dL = lim 𝑓 𝑘 div𝜙dL 𝑛 ≤ lim inf ∥𝐷 𝑓 𝑘 ∥(𝑈).
𝑈 𝑘→∞ 𝑈 𝑘→∞

Taking supremum over all 𝜙, (1.2) then follows. □

4
Exercise 1.9. For any 𝑓 ∈ 𝐵𝑉 (𝑈), define the 𝐵𝑉 norm as
∥ 𝑓 ∥ 𝐵𝑉 (𝑈) = ∥ 𝑓 ∥ 𝐿 1 (𝑈) + ∥𝐷 𝑓 ∥(𝑈).
Show that 𝐵𝑉 (𝑈) equipped with the 𝐵𝑉 norm is a Banach space.
Note that the equality in (1.2) needs not be achieved, as shown by the
following example.
Example 1.10. Let 𝑈 = (0, 2𝜋) be an interval in R. Consider a sequence
of functions 𝑓 𝑘 (𝑥) = 1𝑘 sin(𝑘𝑥) defined on 𝑈, for 𝑘 ∈ N.
Then 𝑓 𝑘 belongs to 𝐿 1 (𝑈) and in fact
1 2𝜋
∫ ∫
1
| 𝑓 𝑘 |dL = |sin(𝑘𝑥)|d𝑥 → 0,
𝑈 𝑘 0
so that
𝑓 𝑘 → 0 in 𝐿 1 (𝑈).
However, we also have
∫ ∫ 2𝜋 ∫ 2𝑘 𝜋
1 1
|𝐷 𝑓 𝑘 |dL = |cos(𝑘𝑥)|d𝑥 = |cos(𝑥)|d𝑥 = 4.
𝑈 0 𝑘 0

2. Approximation and compactness

Definition 2.1 (Symmetric mollifier). A function 𝜂 on R𝑛 is called a (posi-
tive) symmetric mollifier if
(1) 𝜂 ∈ 𝐶𝑐∞ (R𝑛 ).
(2) spt𝜂
∫ ⊂ 𝐵1 (0).
(3) R𝑛 𝜂dL 𝑛 = 1.
(4) 𝜂(𝑥) = 𝜂(−𝑥).
(5) 𝜂 ≥ 0.
For a function 𝑓 ∈ 𝐿 1𝑙𝑜𝑐 (R𝑛 ), we mollify it as follows: Define for each
𝜖 > 0 the functions
𝑥
𝜂𝜖 (𝑥) = 𝜖 −𝑛 𝜂( ),
𝜖
and ∫
−𝑛 𝑥−𝑧
𝑓𝜖 (𝑥) := ( 𝑓 ∗ 𝜂𝜖 )(𝑥) =𝜖 𝜂( ) 𝑓 (𝑧)d𝑧
R𝑛 𝜖
∫
𝑦
=(−1) 𝑛 𝜖 −𝑛 𝜂( ) 𝑓 (𝑥 − 𝑦)d𝑦
R𝑛 𝜖
∫
= 𝜂(𝑤) 𝑓 (𝑥 + 𝜖 𝑤)d𝑤.
R𝑛
5
Lemma 2.2. For 𝑓 ∈ 𝐿 1𝑙𝑜𝑐 (R𝑛 ), the convolution 𝑓 ∗ 𝜂𝜖 ∈ 𝐶 ∞ (R𝑛 ) and there
holds ∫
𝛼
𝜕 ( 𝑓 ∗ 𝜂𝜖 )(𝑥) = 𝑓 (𝑦)𝜕 𝛼 𝜂𝜖 (𝑥 − 𝑦)d𝑦.
R𝑛
Moreover, if 𝑓 ∈ 𝐶𝑐∞ (R𝑛 ), then
𝜕𝑖 ( 𝑓 ∗ 𝜂𝜖 )(𝑥) = (𝜕𝑖 𝑓 ) ∗ 𝜂𝜖 (𝑥).
Theorem 2.3 (Approximation by smooth functions). If 𝑓 ∈ 𝐵𝑉 (𝑈), then
there exists a sequence of functions { 𝑓 𝑘 } 𝑘∈N ⊂ 𝐵𝑉 (𝑈) ∩ 𝐶 ∞ (𝑈) such that
(i) 𝑓 𝑘 → 𝑓 in 𝐿 1 (𝑈) as 𝑘 → ∞.
(ii) ∥𝐷 𝑓 𝑘 ∥(𝑈) → ∥𝐷 𝑓 ∥(𝑈) as 𝑘 → ∞.
Proof. Step 1. Approximation by virtue of partition of unity.
Let 𝜖 > 0 be fixed, there exists a number 𝑚 >> 1 such that if we write
 
1
𝑈 𝑘 = 𝑥 ∈ 𝑈 : dist(𝑥, 𝜕𝑈) > ∩ 𝐵 𝑘+𝑚 (0), 𝑘 = 0, 1, 2, . . . ,
𝑚+𝑘
then
∥𝐷 𝑓 ∥(𝑈 \ 𝑈1 ) < 𝜖 . (2.1)
Set 𝑈0 = ∅ and define the set
𝑉𝑘 = 𝑈 𝑘+1 \ 𝑈 𝑘−1 , 𝑘 = 1, . . . ,
which is again open for each 𝑘, and clearly the collection of 𝑉𝑘 forms a
covering of 𝑈. Let {𝜁 𝑘 }∞ 𝑘=1 be a partition of the unity subordinate to the
covering {𝑉𝑘 }; that is,
∞
∑︁
∞
𝜁 𝑘 ∈ 𝐶𝑐 (𝑉𝑘 ), 0 ≤ 𝜁 𝑘 ≤ 1(𝑘 = 1, . . .), 𝜁 𝑘 ≡ 1 on 𝑈.
𝑘=1
On each “layer” 𝑉𝑘 , we choose 𝜖 𝑘 > 0 small such that


 spt(( 𝑓 𝜁 𝑘 ) ∗ 𝜂𝜖 𝑘 ) ⊆ 𝑉𝑘
∫

( 𝑓 𝜁 𝑘 ) ∗ 𝜂𝜖 𝑘 − 𝑓 𝜁 𝑘 dL 𝑛 < 2𝜖𝑘 , (2.2)
 ∫𝑈
 ( 𝑓 𝐷𝜁 𝑘 ) ∗ 𝜂𝜖 𝑘 − 𝑓 𝐷𝜁 𝑘 dL 𝑛 < 𝜖𝑘 .

 𝑈 2
Then we define the function
∞
∑︁
𝑓𝜖 = ( 𝑓 𝜁 𝑘 ) ∗ 𝜂𝜖 𝑘 .
𝑘=1
Since at each point 𝑥 ∈ 𝑈 there are only finitely many nonzero terms in the
sum, we have 𝑓𝜖 ∈ 𝐶 ∞ (𝑈).
Since also
∞
∑︁
𝑓 (𝑥) = 𝑓 𝜁 𝑘 (𝑥),
𝑘=1
6
it follows from (2.2) that

∞
∫ ∑︁
∥ 𝑓𝜖 − 𝑓 ∥ 𝐿 1 (𝑈) = ( 𝑓 𝜁 𝑘 ) ∗ 𝜂𝜖 𝑘 − 𝑓 𝜁 𝑘 dL 𝑛
𝑈 𝑘=1
∞
∑︁ ∫
≤ ( 𝑓 𝜁 𝑘 ) ∗ 𝜂𝜖 𝑘 − 𝑓 𝜁 𝑘 dL 𝑛 < 𝜖 .
𝑘=1 𝑈

As the LHS and RHS just depend on 𝜖, we can then send 𝜖 → 0 to find that
𝑓𝜖 → 𝑓 in 𝐿 1 (𝑈).
Step 2. We show that 𝑓𝜖 ∈ 𝐵𝑉 (𝑈).
We verify by definition. Let 𝜙 ∈ 𝐶𝑐1 (𝑈, R𝑛 ) with |𝜙| ≤ 1. Using Fubini’s
theorem and the symmetry of the mollifier 𝜂 it is easily seen that

∫ ∞ ∫
∑︁
𝑛
𝑓𝜖 div𝜙dL = ( 𝑓 𝜁 𝑘 ) ∗ 𝜂𝜖 𝑘 div𝜙dL 𝑛
𝑈 𝑘=1 𝑈
∞ ∫
∑︁
= 𝑓 𝜁 𝑘 div(𝜙 ∗ 𝜂𝜖 𝑘 )dL 𝑛
𝑘=1 𝑈
∞ ∫
∑︁

= 𝑓 div 𝜁 𝑘 (𝜙 ∗ 𝜂𝜖 𝑘 ) dL 𝑛
𝑘=1 𝑈
∞ ∫
∑︁
− 𝑓 𝐷𝜁 𝑘 , 𝜙 ∗ 𝜂𝜖 𝑘 dL 𝑛
𝑘=1 𝑈
∞ ∫
∑︁

= 𝑓 div 𝜁 𝑘 (𝜙 ∗ 𝜂𝜖 𝑘 ) dL 𝑛
𝑘=1 𝑈
∞ ∫
∑︁
− 𝜙, ( 𝑓 𝐷𝜁 𝑘 ) ∗ 𝜂𝜖 𝑘 − 𝑓 𝐷𝜁 𝑘 dL 𝑛
𝑘=1 𝑈
:=𝐼1𝜖 + 𝐼2𝜖 ,

Í∞
where we have used the fact that 𝑘=1 𝐷𝜁 𝑘 ≡ 0 in 𝑈 in the last equality.
Note that

𝜁 𝑘 (𝜙 ∗ 𝜂𝜖 𝑘 ) ≤ 1, ∀𝑘,
7
and that each point in 𝑈 belongs to at most three of the sets {𝑉𝑘 } by
construction, thus
∫ ∞ ∫
∑︁
𝜖
𝐼1 = 𝑛
𝑓 div(𝜁1 (𝜙 ∗ 𝜂𝜖1 ))dL + 𝑓 div(𝜁k 𝜙 ∗ 𝜂𝜖k )dL n
𝑈 𝑘=2 𝑈
∞
∑︁
≤∥𝐷 𝑓 ∥(𝑈) + ∥𝐷 𝑓 ∥(𝑉𝑘 )
𝑘=2
≤∥𝐷 𝑓 ∥(𝑈) + 3∥𝐷 𝑓 ∥(𝑈 \ 𝑈1 )
(2.1)
≤ ∥𝐷 𝑓 ∥(𝑈) + 3𝜖 .
On the other hand, by (2.2) it is easy to see that
𝐼2𝜖 < 𝜖 .
Therefore
∥𝐷 𝑓𝜖 ∥ (𝑈) ≤ ∥𝐷 𝑓 ∥(𝑈) + 4𝜖 < ∞, (2.3)
and hence 𝑓𝜖 ∈ 𝐵𝑉 (𝑈).
Conclusion step.
As 𝑓𝜖 ∈ 𝐵𝑉 (𝑈) for any 𝜖 > 0, and 𝑓𝜖 → 𝑓 in 𝐿 1 (𝑈) as 𝜖 ↘ 0, we can
use lower semi-continuity of 𝐵𝑉 functions (Theorem 1.8) to find
∥𝐷 𝑓 ∥(𝑈) ≤ lim inf ∥𝐷 𝑓𝜖 ∥(𝑈),
𝜖→0

which, together with the estimate (2.3), gives the desired convergence (𝑖𝑖).
□
Exercise 2.4 ((★) Criterion for no loss in limit). For 𝑓 ∈ 𝐵𝑉 (𝑈), if 𝐴 ⊂⊂ 𝑈
is an open set such that
∥𝐷 𝑓 ∥(𝜕 𝐴) = 0.
Then
∥𝐷 𝑓 ∥( 𝐴) = lim ∥𝐷 ( 𝑓 ∗ 𝜂𝜖 )∥( 𝐴).
𝜖↘0

Theorem 2.5 (Compactness of BV functions). Let 𝑈 ⊂ R𝑛 be a bounded

open set with Lipschitz boundary 𝜕𝑈. Assume { 𝑓 𝑘 }∞
𝑘=1 is a sequence in
𝐵𝑉 (𝑈) satisfying
sup ∥ 𝑓 𝑘 ∥ 𝐵𝑉 (𝑈) < ∞.
𝑘
Then there exists a subsequence { 𝑓 𝑘 𝑗 }∞
𝑗=1 and a function 𝑓 ∈ 𝐵𝑉 (𝑈) such
that
𝑓 𝑘 𝑗 → 𝑓 in 𝐿 1 (𝑈) as 𝑗 → ∞.
8
Proof. By virtue of Theorem 2.3, for 𝑘 = 1, 2, . . . , we may choose 𝑔 𝑘 ∈
𝐶 ∞ (𝑈) such that
∫ ∫
𝑛 1
| 𝑓 𝑘 − 𝑔 𝑘 |dL < , sup |𝐷𝑔 𝑘 | dL 𝑛 < ∞.
𝑈 𝑘 𝑘 𝑈

By Rellich-Kondrachov theorem (𝑊 1,1 (𝑈) ⊂⊂ 𝐿 1 (𝑈)), there exists 𝑓 ∈

𝐿 1 (𝑈) and a subsequence {𝑔 𝑘 𝑗 }∞
𝑗=1 such that

𝑔 𝑘 𝑗 → 𝑓 in 𝐿 1 (𝑈),
and hence also
𝑓 𝑘 𝑗 → 𝑓 in 𝐿 1 (𝑈).
Since 𝑓 𝑘 𝑗 ∈ 𝐵𝑉 (𝑈), we conclude the proof by using lower semi-continuity
of BV functions (Theorem 1.8). □

3. Isoperimetric inequalities

Theorem 3.1 (Sobolev inequality for 𝐵𝑉 functions). There exists a constant

𝐶1 = 𝐶1 (𝑛) such that
∥ 𝑓 ∥ 𝐿 1∗ (R𝑛 ) ≤ 𝐶1 ∥𝐷 𝑓 ∥(R𝑛 ), ∀ 𝑓 ∈ 𝐵𝑉 (R𝑛 ), (3.1)
where
𝑛
1∗ = .
𝑛−1
Proof. Since 𝑓 ∈ 𝐵𝑉 (R𝑛 ), we could use Theorem 2.3 to obtain a sequence
of functions { 𝑓 𝑘 } 𝑘∈N ⊂ 𝐵𝑉 (R𝑛 ) ∩ 𝐶 ∞ (R𝑛 ) such that
𝑓 𝑘 → 𝑓 in 𝐿 1 (R𝑛 ), ∥𝐷 𝑓 𝑘 ∥(R𝑛 ) → ∥𝐷 𝑓 ∥(R𝑛 ).
Moreover, by standard cut-off argument we may assume that 𝑓 𝑘 ∈ 𝐶𝑐∞ (𝑈).
Also, the 𝐿 1 -convergence implies
𝑓 𝑘 → 𝑓 L 𝑛 -a.e.

By Fatou’s Lemma and Gagliardo-Nirenberg-Sobolev inequality (𝑊 1,𝑝 (R𝑛 ) ⊂⊂

∗
𝐿 𝑝 ) we obtain
∥ 𝑓 ∥ 𝐿 1∗ (R𝑛 ) ≤ lim inf ∥ 𝑓 𝑘 ∥ 𝐿 1∗ (R𝑛 )
𝑘→∞
≤𝐶1 (𝑛) lim inf ∥𝐷 𝑓 𝑘 ∥ 𝐿 1 (R𝑛 )
𝑘→∞
𝑛
=𝐶1 ∥𝐷 𝑓 ∥(R )
as desired. □
9
Theorem 3.2 (Isoperimetric inequality). Let 𝐸 ⊂ R𝑛 be a bounded set such
that 𝜒𝐸 ∈ 𝐵𝑉 (R𝑛 ), then
𝑛−1
L 𝑛 (𝐸) 𝑛 ≤ 𝐶1 ∥𝐷 𝜒𝐸 ∥(R𝑛 ), (3.2)
where 𝐶1 = 𝐶1 (𝑛) is the resulting constant from Gagliardo-Nirenberg-
Sobolev inequality.
Proof. Take 𝑓 = 𝜒𝐸 in (3.1) yields the result. □
Remark 3.3. (𝑖) The set 𝐸 as above is exactly the so-called set of finite
perimeter/Caccioppoli set.
(𝑖𝑖) In terms of sets of finite perimeter, we shall see in the future that the
RHS of (3.2) is the so-called perimeter of 𝐸 (therefore some math-
ematicians also call the isoperimetric inequality ‘Perimeter bounds
on volume’), and is exactly the (𝑛 − 1)-Hausdorff measure of the
reduced boundary of 𝐸, while it is well-known that in R𝑛 , L 𝑛 = H 𝑛 .
Therefore the isoperimetric inequality is scale-invariant.
(𝑖𝑖𝑖) The optimal constant on the right is given by
1
(𝑛𝜔𝑛𝑛 ) −1 ,
and the optimal inequality is called the (Euclidean) isoperimetric
inequality. Equality is achieved if and only if 𝐸 is, up to a L 𝑛 -null
set, a Euclidean ball.
Theorem 3.4 (Poincaré inequality in balls for 𝐵𝑉 functions). There exists
a constant 𝐶2 = 𝐶2 (𝑛) such that
∥ 𝑓 − ( 𝑓 )𝑥,𝑟 ∥ 𝐿 1∗ (𝐵𝑟 (𝑥)) ≤ 𝐶2 ∥𝐷 𝑓 ∥(𝐵𝑟 (𝑥)), (3.3)
for all balls 𝐵𝑟 (𝑥) ⊂ R𝑛 and 𝑓 ∈ 𝐵𝑉𝑙𝑜𝑐 (R𝑛 ), where
⨏
( 𝑓 )𝑥,𝑟 := 𝑓 dL 𝑛 .
𝐵𝑟 (𝑥)
Proof. We argue as the proof of Theorem 3.1. For a fixed ball 𝐵𝑟 (𝑥), since
𝑓 ∈ 𝐵𝑉𝑙𝑜𝑐 (R𝑛 ) we have 𝑓 ∈ 𝐵𝑉 (𝐵𝑟 (𝑥)). Approximating 𝑓 ∈ 𝐵𝑉 (𝐵𝑟 (𝑥))
by a sequence of functions { 𝑓 𝑘 } 𝑘∈N ⊂ 𝐵𝑉 (𝐵𝑟 (𝑥)) ∩ 𝐶 ∞ (𝐵𝑟 (𝑥)) such that
𝑓 𝑘 → 𝑓 in 𝐿 1 (𝐵𝑟 (𝑥)), 𝑓 𝑘 → 𝑓 L 𝑛 -a.e., ∥𝐷 𝑓 𝑘 ∥(𝐵𝑟 (𝑥)) → ∥𝐷 𝑓 ∥(𝐵𝑟 (𝑥)).
It follows that
( 𝑓 𝑘 )𝑥,𝑟 → ( 𝑓 )𝑥,𝑟 as 𝑘 → ∞.
By Fatou’s Lemma and Poincaré’s inequality for 𝑓 𝑘 ∈ 𝑊 1,1 (𝐵𝑟 (𝑥)), we get
∥ 𝑓 − ( 𝑓 )𝑥,𝑟 ∥ 𝐿 1∗ (𝐵𝑟 (𝑥)) ≤ lim inf ∥ 𝑓 𝑘 − ( 𝑓 𝑘 )𝑥,𝑟 ∥ 𝐿 1∗ (𝐵𝑟 (𝑥))
𝑘→∞
≤𝐶2 lim inf ∥𝐷 𝑓 𝑘 ∥ 𝐿 1 (𝐵𝑟 (𝑥))
𝑘→∞
=𝐶2 ∥𝐷 𝑓 ∥(𝐵𝑟 (𝑥))
10
as desired. □
The following local inequality is very useful in Geometric measure theory
and Calculus of variations.
Theorem 3.5 (Relative isoperimetric inequality in balls). Let 𝐸 ⊂ R𝑛 be a
bounded set such that 𝜒𝐸 ∈ 𝐵𝑉𝑙𝑜𝑐 (R𝑛 ), then for each ball 𝐵𝑟 (𝑥) ⊂ R𝑛 , there
holds
𝑛−1
min{L 𝑛 (𝐵𝑟 (𝑥) ∩ 𝐸), L 𝑛 (𝐵𝑟 (𝑥) \ 𝐸)} ≤ 2𝐶2 ∥𝐷 𝜒𝐵𝑟 (𝑥)∩𝐸 ∥(𝐵𝑟 (𝑥)),
𝑛

(3.4)
where 𝐶2 = 𝐶2 (𝑛) is the constant resulting from Poincaré’s inequality in
balls.
Proof. Take 𝑓 = 𝜒𝐸∩𝐵𝑟 (𝑥) in (3.3), with the observation that in this case
L 𝑛 (𝐵𝑟 (𝑥) ∩ 𝐸)
( 𝑓 )𝑥,𝑟 = ,
L 𝑛 (𝐵𝑟 (𝑥))
we find
∫ ∫
1∗ 𝑛 1∗
𝑓 − ( 𝑓 )𝑥,𝑟 dL = 𝑓 − ( 𝑓 )𝑥,𝑟 dL 𝑛
𝐵𝑟 (𝑥) 𝐵𝑟 (𝑥)∩𝐸
∫
1∗
+ 𝑓 − ( 𝑓 )𝑥,𝑟 dL 𝑛
𝐵𝑟 (𝑥)\𝐸
  1∗
L 𝑛 (𝐵𝑟 (𝑥) \ 𝐸)
= L 𝑛 (𝐵𝑟 (𝑥) ∩ 𝐸)
L (𝐵𝑟 (𝑥))
𝑛
 𝑛  1∗
L (𝐵𝑟 (𝑥) ∩ 𝐸)
+ L 𝑛 (𝐵𝑟 (𝑥) \ 𝐸).
L (𝐵𝑟 (𝑥))
𝑛

Assume WLOG L 𝑛 (𝐵𝑟 (𝑥) ∩ 𝐸) ≤ L 𝑛 (𝐵𝑟 (𝑥) \ 𝐸), then

𝐶2 ∥𝐷 𝜒𝐵𝑟 (𝑥)∩𝐸 ∥(𝐵𝑟 (𝑥)) ≥∥ 𝑓 − ( 𝑓 )𝑥,𝑟 ∥ 𝐿 1∗ (𝐵𝑟 (𝑥))
∫  𝑛−1
1∗ 𝑛
𝑛
= 𝑓 − ( 𝑓 )𝑥,𝑟 dL
𝐵 (𝑥)
 𝑟 
drop the second term L 𝑛 (𝐵𝑟 (𝑥) \ 𝐸) 𝑛−1
𝑛
≥ L (𝐵 𝑟 (𝑥) ∩ 𝐸) 𝑛
L 𝑛 (𝐵𝑟 (𝑥))
1 𝑛−1
≥ min{L 𝑛 (𝐵𝑟 (𝑥) ∩ 𝐸), L 𝑛 (𝐵𝑟 (𝑥) \ 𝐸)} 𝑛
2
as desired. □
Remark 3.6. The term on the RHS of (3.4) is exactly the relative perimeter
of 𝐸 in 𝐵𝑟 (𝑥), which accounts for the perimeter of 𝐸 that lies in the open
ball 𝐵𝑟 (𝑥).
11
4. Coarea formula for BV functions
Definition 4.1 (Super level-sets of functions). For 𝑓 : 𝑈 → R and 𝑡 ∈ R,
define
𝐸 𝑡 := {𝑥 ∈ 𝑈 : 𝑓 (𝑥) > 𝑡}
as the super level-set of 𝑓 (with respect to the value 𝑡).
Lemma 4.2 (Layer-cake formula). If 𝑓 ∈ 𝐿 1 (R𝑛 ), 𝑓 ≥ 0, and 𝑣 ∈ 𝐿 ∞ (R𝑛 ),
then ∫ ∫ ∞ ∫
𝑓 (𝑥)𝑣(𝑥)d𝑥 = d𝑡 𝑣(𝑥)d𝑥. (4.1)
R𝑛 0 𝐸𝑡

Proof. For every 𝑥 ∈ R𝑛 , one has

∫ ∫ ∫ ∞
𝑓 (𝑥) = 𝜒(0, 𝑓 (𝑥)) (𝑡)d𝑡 = 𝜒(0,∞) (𝑡) 𝜒𝐸𝑡 (𝑥)d𝑡 = 𝜒𝐸𝑡 (𝑥)d𝑡,
R R 0
and hence by Fubini’s theorem
∫ ∫  ∫ ∞ 
𝑓 (𝑥)𝑣(𝑥)d𝑥 = 𝑣(𝑥) 𝜒𝐸𝑡 (𝑥)d𝑡 d𝑥
{ 𝑓 ≥0} { 𝑓 ≥0} 0
∫ ∞ ∫
= d𝑡 𝑣(𝑥)d𝑥.
0 𝐸𝑡
□
Lemma 4.3. If 𝑓 ∈ 𝐵𝑉 (𝑈), then the map
𝑡 ↦→ ∥𝐷 𝜒𝐸𝑡 ∥(𝑈), 𝑡 ∈ R,
is L 1 -measurable.
Proof. For any 𝜙 ∈ 𝐶𝑐1 (𝑈, R𝑛 ), by Layer-cake formula with 𝑣 = div𝜙 and
the fact that 𝑓 ∈ 𝐵𝑉 (𝑈), we obtain
∫ ∞ ∫ ∫
d𝑡 𝜒𝐸𝑡 (𝑥)div𝜙(𝑥)d𝑥 = 𝑓 div𝜙(𝑥)d𝑥 < ∞,
0 R𝑛 𝑈
implying that the map
∫ ∫
𝑛
𝑡 ↦→ div𝜙dL = 𝜒𝐸𝑡 div𝜙dL 𝑛
𝐸𝑡 𝑈

isL 1 -measurable. Let 𝐷 denote any countable dense subset of 𝐶𝑐1 (𝑈, R𝑛 ),
then ∫
𝑡 ↦→ ∥𝐷 𝜒𝐸𝑡 ∥(𝑈) = sup div𝜙dL 𝑛
𝜙∈𝐷,|𝜙|≤1 𝐸𝑡

is L 1 -measurable. □
12
Theorem 4.4 (Coarea formula for 𝐵𝑉 functions). If 𝑓 ∈ 𝐵𝑉 (𝑈) then 𝜒𝐸𝑡 ∈
𝐵𝑉 (𝑈) for L 1 -a.e. 𝑡 ∈ R, and
∫ ∞
∥𝐷 𝑓 ∥(𝑈) = ∥𝐷 𝜒𝐸𝑡 ∥(𝑈)d𝑡. (4.2)
−∞
Proof. Let 𝑓 ∈ 𝐿 1 (𝑈)
and 𝜙 ∈ 1
𝐶𝑐 (𝑈, R𝑛 ) with |𝜙| ≤ 1.
Step 1. We prove that there holds
∫ ∫ ∞ ∫ 
𝑛 𝑛
𝑓 div𝜙dL = div𝜙dL d𝑡.
𝑈 −∞ 𝐸𝑡
To see this, notice that 𝑓 = 𝑓 + − −
𝑓 then use
Layer-cake formula (4.1) for
+ −
𝑓 and 𝑣 = div𝜙 also 𝑓 and 𝑣 = div𝜙, respectively. It follows immediately
∫ ∞ ∫
∥𝐷 𝜒𝐸𝑡 ∥(𝑈)d𝑡 ≥ 𝑓 div𝜙d𝑥,
−∞ 𝑈
and hence after taking supremum over all 𝜙,
∫ ∞
∥𝐷 𝜒𝐸𝑡 ∥(𝑈)d𝑡 ≥ ∥𝐷 𝑓 ∥(𝑈).
−∞
Step 2. We prove (4.2) for 𝑓 ∈ 𝐵𝑉 (𝑈) ∩ 𝐶 ∞ (𝑈).
For any 𝑡 ∈ R fixed, write
∫ ∫
𝑛
𝑚(𝑡) := |𝐷 𝑓 |dL = |𝐷 𝑓 |dL 𝑛 .
𝑈\𝐸 𝑡 { 𝑓 ≤𝑡}

Clearly the function 𝑚 is non-decreasing and hence 𝑚′ exists for L 1 -a.e. 𝑡,

with ∫ ∞ ∫
𝑚′ (𝑡)d𝑡 ≤ 𝑚(∞) − 𝑚(−∞) = |𝐷 𝑓 |dL 𝑛 . (4.3)
−∞ 𝑈
Then for a fixed 𝑟 > 0, we define (Lipschitz) cut-off function 𝛾 : R → R
as


 0, if 𝑠 ≤ 𝑡
 𝑠−𝑡

𝛾(𝑠) := 𝑟 if 𝑡 ≤ 𝑠 ≤ 𝑡 + 𝑟

1 if 𝑠 ≥ 𝑡 + 𝑟.


It is then easy to see that
(
1
if 𝑡 < 𝑠 < 𝑡 + 𝑟
𝛾 ′ (𝑠) = 𝑟
0 if 𝑠 < 𝑡 or 𝑠 > 𝑡 + 𝑟.
For any 𝜙 ∈ 𝐶𝑐1 (𝑈, R𝑛 ), a direct computation shows
∫ ∫
− 𝛾( 𝑓 (𝑥))div𝜙(𝑥)d𝑥 = 𝛾 ′ ( 𝑓 (𝑥)) ⟨𝐷 𝑓 , 𝜙⟩ d𝑥
𝑈 𝑈
∫
1
= ⟨𝐷 𝑓 , 𝜙⟩ d𝑥,
𝑟 𝐸𝑡 \𝐸𝑡+𝑟
13
and it follows that
∫ ∫ 
𝑚(𝑡 + 𝑟) − 𝑚(𝑡) 1
= |𝐷 𝑓 |d𝑥 − |𝐷 𝑓 |d𝑥
𝑟 𝑟 𝑈\𝐸𝑡+𝑟 𝑈\𝐸 𝑡
∫
1
= |𝐷 𝑓 |d𝑥
𝑟 𝐸𝑡 \𝐸𝑡+𝑟
∫
1
≥ ⟨𝐷 𝑓 , 𝜙⟩ d𝑥
𝑟 𝐸𝑡 \𝐸𝑡+𝑟
∫
=− 𝛾( 𝑓 (𝑥))div𝜙(𝑥)d𝑥.
𝑈
Fix 𝑡 ∈ R such that 𝑚′ (𝑡)
exists, we let 𝑟 → 0 and obtain
∫
′
𝑚 (𝑡) ≥ − div𝜙(𝑥)d𝑥.
𝐸𝑡

Taking supremum over all 𝜙, we get

𝑚′ (𝑡) ≥ ∥𝐷 𝜒𝐸𝑡 ∥(𝑈).
Combining this with (4.3), we find
∫ ∞ ∫
∥𝐷 𝜒𝐸𝑡 ∥(𝑈)d𝑡 ≤ |𝐷 𝑓 |d𝑥 = ∥𝐷 𝑓 ∥(𝑈)
−∞ 𝑈

as desired. In particular, this proves that for L 1 -a.e. 𝑡,

∥𝐷 𝜒𝐸𝑡 ∥(𝑈) < ∞,
namely, 𝜒𝐸𝑡 ∈ 𝐵𝑉 (𝑈).
Step 3. We prove (4.2) for 𝑓 ∈ 𝐵𝑉 (𝑈).
Using Theorem 2.3, we can approximate 𝑓 ∈ 𝐵𝑉 (𝑈) by a sequence of
smooth functions { 𝑓 𝑘 }∞
𝑘=1 such that

𝑓 𝑘 → 𝑓 in 𝐿 1 (𝑈), ∥𝐷 𝑓 𝑘 ∥(𝑈) → ∥𝐷 𝑓 ∥(𝑈), as 𝑘 → ∞. (4.4)

For any fixed 𝑡 ∈ R and every 𝑘 ∈ N, consider the super level-sets of 𝑓 𝑘
𝐸 𝑡𝑘 := {𝑥 ∈ 𝑈 : 𝑓 𝑘 (𝑥) > 𝑡}.
Since
∫ ∞ ∫ max{ 𝑓 (𝑥), 𝑓 𝑘 (𝑥)}
𝜒𝐸𝑡𝑘 (𝑥) − 𝜒𝐸𝑡 (𝑥) d𝑡 = d𝑡 = | 𝑓 𝑘 (𝑥) − 𝑓 (𝑥)|,
−∞ min{ 𝑓 (𝑥), 𝑓 𝑘 (𝑥)}

we can use Fubini’s theorem to find

∫ ∫ ∞ ∫ 
| 𝑓 𝑘 (𝑥) − 𝑓 (𝑥)| d𝑥 = 𝜒𝐸𝑡𝑘 (𝑥) − 𝜒𝐸𝑡 (𝑥) d𝑥 d𝑡
𝑈 −∞ 𝑈
14
Since 𝑓 𝑘 → 𝑓 in 𝐿 1 (𝑈), we deduce from the above equality that, up to a
subsequence (which we still index by 𝑘), there holds
𝜒𝐸𝑡𝑘 → 𝜒𝐸𝑡 in 𝐿 1 (𝑈)
for L 1 -a.e. 𝑡. In Step 2 we have shown that 𝜒𝐸𝑡𝑘 ∈ 𝐵𝑉 (𝑈), therefore
we may use lower semicontinuity of 𝐵𝑉 functions (Theorem 1.8) to find
𝜒𝐸𝑡 ∈ 𝐵𝑉 (𝑈) with
∥𝐷 𝜒𝐸𝑡 ∥(𝑈) ≤ lim inf ∥𝐷 𝜒𝐸𝑡𝑘 ∥(𝑈).
𝑘→∞
Finally, we use Fatou’s lemma, Step 2, and (4.4) to conclude that
∫ ∞ ∫ ∞
∥𝐷 𝜒𝐸𝑡 ∥(𝑈)d𝑡 ≤ lim inf ∥𝐷 𝜒𝐸𝑡𝑘 ∥(𝑈)d𝑡
−∞ 𝑘→∞ −∞
= lim ∥𝐷 𝑓 𝑘 ∥(𝑈)
𝑘→∞
=∥𝐷 𝑓 ∥(𝑈),
which completes the proof. □
1,∞
Remark 4.5. A similar proof shows that Coarea Theorem holds for 𝑊𝑙𝑜𝑐 (R𝑛 )
functions, which states as follows:
1,∞
If 𝑓 ∈ 𝑊𝑙𝑜𝑐 (R𝑛 ), then for any open set 𝐴 ⊂ R𝑛 , 𝑡 ∈ R ↦→ 𝜒𝐸𝑡 ∩𝐴 is a
Borel function on R with
∫ ∫ ∞
𝑛
|𝐷 𝑓 |dL = ∥𝐷 𝜒𝐸𝑡 ∩𝐴 ∥( 𝐴)d𝑡
𝐴 −∞
as elements of [0, ∞]. (See [2, Theorem 13.1])
Exercise 4.6. Use isoperimetric inequality (Theorem 3.2) to prove Gagliardo-
Nirenberg-Sobolev inequality for non-negative 𝐵𝑉 functions on R𝑛 with
compact support.
(Hint: By Coarea Theorem and isoperimetric inequality,
∫ ∞
𝑛
∥𝐷 𝑓 ∥(R ) = ∥𝐷 𝜒𝐸𝑡 ∥(R𝑛 )d𝑡
−∞
∫ ∞
1 𝑛−1
≥ L 𝑛 (𝐸 𝑡 ) 𝑛 d𝑡,
𝐶1 −∞
then show that ∫ ∞
𝑛−1
∥ 𝑓 ∥ 𝐿 1∗ ≤ L 𝑛 (𝐸 𝑡 ) 𝑛 d𝑡.
0
)
Remark 4.7. The classical Gagliardo-Nirenberg-Sobolev inequality, with
𝑝 = 1, is equivalent to the isoperimetric inequality.

15
Chapter 2: Sets of finite perimeter
We will follow mostly the notations in [2].
• The Euclidean scalar product, the corresponding Levi-Civita con-
nection, and the Lebesgue measure on R𝑛 are denoted respectively
by ⟨·, ·⟩, ∇, and L 𝑛 .
• We adopt the following notations when considering the topology of
R𝑛 : we denote by 𝐸 the topological closure of a set 𝐸, by int(𝐸) the
topological interior of 𝐸, by 𝐸 𝑐 the topological complement of 𝐸,
by 𝜕𝐸 the topological boundary of 𝐸.
• For 𝑘 ∈ N, H 𝑘 denotes the 𝑘-dimensional Hausdorff measure on
R𝑛 .
• We denote by R𝑛 ⊗ R𝑚 the vector space of linear maps from R𝑚 to
R𝑛 .
• We denote by P (R𝑛 ) the space of subsets of R𝑛 , B (R𝑛 ) the Borel
sets in R𝑛 , and M (L 𝑛 ) the Lebesgue measurable sets in R𝑛 .

5. Definition and basic properties

5.1. Definition
Definition 5.1 (Set of finite perimeter/ Caccioppoli set). Let 𝐸 be a Lebesgue
measurable set in R𝑛 . We say that 𝐸 is a set of (locally) finite perimeter in
R𝑛 if 𝜒𝐸 ∈ 𝐵𝑉(𝑙𝑜𝑐) (R𝑛 ).
Proposition 5.2. If 𝐸 is a Lebesgue measurable set in R𝑛 . Then 𝐸 is a set of
locally finite perimeter if and only if there exists a (R𝑛 ) vector-valued Radon
measure 𝜇 𝐸 on R𝑛 such that
∫ ∫
div𝜙dL =𝑛
⟨𝜙, d𝜇 𝐸 ⟩ , ∀𝜙 ∈ 𝐶𝑐1 (R𝑛 , R𝑛 ). (5.1)
𝐸 R𝑛

Proof. The if part follows directly from Theorem 1.5. The only if part is
trivial. □

Note that by virtue of Theorem 1.5, we have the polar decomposition

𝜇 𝐸 = 𝜈 𝐸 |𝜇 𝐸 |,

where |𝜇 𝐸 | is a Radon measure on R𝑛 and is called the total variation

measure, while 𝜈 𝐸 is a |𝜇 𝐸 |-measurable function with |𝜈 𝐸 (𝑥)| = 1 for |𝜇 𝐸 |-
a.e. 𝑥.
If 𝐸 is a bounded open set in R𝑛 with 𝐶 1 -boundary 𝜕𝐸, then 𝜇 𝐸 =
𝜈 𝐸 H 𝑛−1 ⌞𝜕𝐸, where 𝜈 𝐸 is the outer unit normal. In this case formula (5.1)
16
holds because of the classical Gauss-Green Theorem (divergence theorem):
∫ ∫
𝑛
div𝜙dL = ⟨𝜙(𝑥), 𝜈 𝐸 (𝑥)⟩ dH 𝑛−1 (𝑥), ∀𝜙 ∈ 𝐶𝑐1 (R𝑛 , R𝑛 ).
𝐸 𝜕𝐸
Or equivalently,
∫ ∫
∇𝜑dL = 𝑛
𝜑𝜈 𝐸 dH 𝑛−1 , ∀𝜑 ∈ 𝐶𝑐1 (R𝑛 ).
𝐸 𝜕𝐸
Proposition 5.3 (Open sets with Lipschitz boundary). An open set 𝐸 ⊂ R𝑛
has Lipschitz boundary if for every 𝑥 ∈ 𝜕𝐸 there exists 𝑟 > 0 and a Lipschitz
function 𝑢 : R𝑛+1 → R such that up to translation and rotation, 𝜕𝐸 ∩ 𝐵𝑟 (𝑥)
can be written as a graph of 𝑢 over 𝐵𝑟𝑛−1 . If 𝐺 ⊂ 𝐵𝑟𝑛−1 ⊂ R𝑛−1 is the set of
points of differentiability of 𝑢 and 𝐹 (𝑧) = (𝑧, 𝑢(𝑧)), 𝑧 ∈ 𝐵𝑟𝑛−1 , then we may
define on 𝐺 the upwards-pointing unit normal field
(−∇𝑢(𝑧), 1)
𝜈 𝐸 (𝑧) = √︁ , 𝑧 ∈ 𝐺.
1 + |∇𝑢(𝑧)| 2
Note that H 𝑛−1 (𝐺) = 0 by Rademacher’s theorem (Theorem C.2) and hence
H 𝑛−1 (𝐹 (𝐺)) = 0 by area formula. By a (local) covering argument we
conclude that, if 𝐸 is an open set with Lipschitz boundary, then H 𝑛−1 ⌞𝜕𝐸 is
a Gauss-Green measure, meaning that the classical Gauss-Green formula
∫ ∫
∇𝜙(𝑥)d𝑥 = 𝜑𝜈 𝐸 dH 𝑛−1
𝐸 𝜕𝐸
holds true for any 𝜑 ∈ 𝐶𝑐1 (R𝑛 ). This means, 𝐸 as above is a set of locally
finite perimeter, with 𝜇 𝐸 = 𝜈 𝐸 H 𝑛−1 ⌞𝜕𝐸.
Definition 5.4 (Polyhedral). An open set 𝐸 ⊂ R𝑛+1 is said to have polyhedral
boundary if 𝜕𝐸 is finitely piecewise affine.
Corollary 5.5. Proposition 5.3 is satisfied by any open set 𝐸 ⊂ R𝑛+1 with
polyhedral boundary.
Definition 5.6 (Gauss-Green measure, perimeter, relative perimeter). We
call above 𝜇 𝐸 the Gauss-Green measure of 𝐸. The perimeter of 𝐸 is defined
as
𝑃(𝐸) = |𝜇 𝐸 |(R𝑛 ),
and the relative perimeter of 𝐸 in 𝐴 ⊂ R𝑛 is defined as
𝑃(𝐸; 𝐴) = |𝜇 𝐸 |( 𝐴),
which is characterized by
∫ 
𝑛 1 𝑛
𝑃(𝐸; 𝐴) = sup div𝜙dL : 𝜙 ∈ 𝐶𝑐 ( 𝐴, R ), |𝜙| ≤ 1 . (5.2)
𝐸
17
5.2. Basic properties
Definition 5.7 (Set differences). Given Lebesgue measurable sets 𝐸, 𝐹 in
R𝑛 . The set differences of 𝐸 and 𝐹, denoted by 𝐸Δ𝐹, is defined as
𝐸Δ𝐹 = (𝐸 \ 𝐹) ∪ (𝐹 \ 𝐸) .
Proposition 5.8. If 𝐸 and 𝐹 are sets of locally finite perimeter in R𝑛 , then
the following statements are equivalent:
(i) 𝜇 𝐸 = 𝜇 𝐹 as Radon measures.
(ii) |𝐸Δ𝐹 | = 0.
Proof. (𝑖) ⇒ (𝑖𝑖):
Since (5.1) holds for 𝐸 and 𝐹, subtracting the identities we have
∫
( 𝜒𝐸 − 𝜒𝐹 )div𝜙dL 𝑛 = 0, ∀𝜙 ∈ 𝐶𝑐1 (R𝑛 , R𝑛 ).
R𝑛
By vanishing weak gradient, for some 𝑐 ∈ R we have
𝜒𝐸 − 𝜒𝐹 = 𝑐, a.e. in R𝑛 .
Note the possible values of 𝑐 are ±1 and 0, and hence 𝑐 = 0 since 𝐸, 𝐹 are
sets of locally finite perimeter. This implies (𝑖𝑖).
(𝑖) ⇐ (𝑖𝑖):
In this case we have
𝜒𝐸 = 𝜒𝐹 , a.e. in R𝑛 .
Since (5.1) holds for 𝐸 and 𝐹, subtracting the identities we have
∫
⟨𝜙, d(𝜇 𝐸 − 𝜇 𝐹 )⟩ = 0, ∀𝜙 ∈ 𝐶𝑐1 (R𝑛 , R𝑛 ),
R𝑛
which proves (𝑖). □
Remark 5.9. 𝑃(𝐸) is invariant by modifications of 𝐸 on and/or by a set of
measure zero, though these modifications may widely affect the size of its
topological boundary.
See [2, Figure 12.1] for a nice example.
Similarly, one can show that:
Exercise 5.10 (Locality of perimeter). If 𝐸, 𝐹 are set of locally finite perime-
ter in R𝑛 such that
|(𝐸Δ𝐹) ∩ 𝐴| = 0
for some open set 𝐴. Then
𝑃(𝐸; 𝐴) = 𝑃(𝐹; 𝐴).
18
Exercise 5.11 (Scaling and translation). Fix 𝜆 > 0, 𝑥0 ∈ R𝑛 . If 𝐸 is a set of
finite perimeter in R𝑛 , then so is 𝑥0 + 𝜆𝐸, with
𝑃(𝑥 0 + 𝜆𝐸) = 𝜆𝑛−1 𝑃(𝐸).
More generally, if 𝐸 is a set of locally finite perimeter in R𝑛 , then so is
𝑥 0 + 𝜆𝐸, with
𝜇𝑥0 +𝜆𝐸 = 𝜆𝑛−1 Φ# 𝜇 𝐸 ,
where Φ(𝑦) = 𝑥 + 𝜆𝑦 for 𝑦 ∈ R𝑛 . Here Φ♯ 𝜇 𝐸 denotes the push-forward of a
Radon measure, which is defined as
∫ ∫
𝜑(𝑥)d(Φ# 𝜇 𝐸 )(𝑥) = (𝜑 ◦ Φ)(𝑥)d𝜇 𝐸 (𝑥), ∀𝜑 ∈ 𝐶 0 (R𝑛 ).
R𝑛 R𝑛
In particular, choose 𝜑 = 𝜒 𝐴 for some Borel measurable set 𝐴, then this
reads
Φ# 𝜇 𝐸 ( 𝐴) = 𝜇 𝐸 (Φ−1 ( 𝐴)).
[Hint: Using (5.1)]
Exercise 5.12 (Complement). If 𝐸 is a set of locally finite perimeter in R𝑛 ,
then so is 𝐸 𝑐 = R𝑛 \ 𝐸. Moreover,
𝜇R𝑛 \𝐸 = −𝜇 𝐸 , 𝑃(𝐸) = 𝑃(𝐸 𝑐 ).
[Hint: Using (5.1) and note that
∫
∇𝜑dL 𝑛 = 0, ∀𝜑 ∈ 𝐶𝑐1 (R𝑛 ).]
R𝑛
Exercise 5.13. If 𝐸 is a set of finite perimeter in R𝑛 , then so is 𝑄(𝐸) for any
𝑄 ∈ O(𝑛) (rotations in R𝑛 ). Moreover, show that 𝑃(𝐸) = 𝑃(𝑄(𝐸)).

5.3. Lower semicontinuity of perimeter

Definition 5.14 (Convergence as Lebesgue measurable sets). Given Lebesgue
measurable sets {𝐸 ℎ } ℎ∈N , 𝐸 in R𝑛 . We say that 𝐸 ℎ locally converges to 𝐸,
loc
denoted by 𝐸 ℎ → 𝐸, if
lim |𝐾 ∩ (𝐸Δ𝐸 ℎ )| = 0, ∀𝐾 ⊂ R𝑛 compact.
ℎ→∞
We say that 𝐸 ℎ converges to 𝐸, denoted by 𝐸 ℎ → 𝐸, if
lim |𝐸Δ𝐸 ℎ | = 0.
ℎ→∞

Proposition 5.15 (Lower semicontinuity of perimeter). If {𝐸 ℎ } ℎ∈N is a

sequence of sets of locally finite perimeter in R𝑛 with
loc
𝐸 ℎ → 𝐸, lim sup 𝑃(𝐸 ℎ , 𝐾) < ∞,
ℎ→∞
19
for every compact 𝐾 ⊂ R𝑛 . Then 𝐸 is also a set of locally finite perimeter in
R𝑛 , such that 𝜇 𝐸 ℎ → 𝜇 𝐸 as (vector-valued) Radon measure, and for every
open 𝐴 ⊂ R𝑛 , we have
𝑃(𝐸; 𝐴) ≤ lim inf 𝑃(𝐸 ℎ ; 𝐴). (5.3)
ℎ→∞

See [2, Figure 12.2] for a nice example showing that the RHS may not be
attained.
Proof. For 𝐴 open, let 𝜙 ∈ 𝐶𝑐∞ ( 𝐴, R𝑛 ) with |𝜙| ≤ 1. By (5.2),
∫ ∫
𝑛
div𝜙dL = lim div𝜙dL 𝑛 ≤ lim inf 𝑃(𝐸 ℎ ; 𝐴).
𝐸 ℎ→∞ 𝐸ℎ
Taking supremum over all admissible 𝜙, (5.3) is thus proven. It follows
that 𝐸 is also a set of locally finite perimeter. Using (5.1) and the fact that
loc
𝐸 ℎ → 𝐸, we obtain for any 𝜙 ∈ 𝐶𝑐∞ (R𝑛 ),
∫ ∫ ∫ ∫
𝑛 𝑛
lim 𝜙, d𝜇 𝐸 ℎ = lim div𝜙dL = div𝜙dL = ⟨𝜙, d𝜇 𝐸 ⟩ .
ℎ→∞ R𝑛 ℎ→∞ 𝐸ℎ 𝐸 R𝑛

By the density of 𝐶𝑐∞ (R𝑛 ) in 𝐶𝑐0 (R𝑛 ), we have

∫ ∫
lim 𝜙, d𝜇 𝐸 ℎ = ⟨𝜙, d𝜇 𝐸 ⟩ , ∀𝜙 ∈ 𝐶𝑐∞ (R𝑛 ),
ℎ→∞ R𝑛 R𝑛
which proves that 𝜇 𝐸 ℎ → 𝜇 𝐸 as Radon measures. □
An alternative way to prove this is to apply Theorem 1.8 (with necessary
modifications that 𝑓 𝑘 ∈ 𝐵𝑉𝑙𝑜𝑐 (𝑈)).
Exercise 5.16. If 𝐴 is a connected set in R𝑛 , 𝐸 is a set of locally finite
perimeter such that 𝑃(𝐸; 𝐴) = 0. Then, either | 𝐴 \ 𝐸 | = 0, or | 𝐴 ∩ 𝐸 | = 0.

5.4. Topological boundary and Gauss-Green measure

The starting point of this part is Proposition 5.8, in which we prove that
any L 𝑛 -zero modification of a set of locally finite perimeter does not change
the Gauss-Green measure 𝜇 𝐸 . On the other hand, these modifications, as
observed in Remark 5.9, could widely affect the size of the topological
boundary. This motivates us to find a way to modify 𝐸 to “minimize”
the size of its topological boundary, and obtain the following important
properties for sets of locally finite perimeter:
Definition 5.17. An outer measure 𝜇 on R𝑛 is called concentrated on 𝐸 ⊂ R𝑛
if 𝜇(R𝑛 \ 𝐸) = 0. The support of 𝜇, denoted by spt𝜇, is the intersection of
the closed sets 𝐸 which 𝜇 is concentrated on. In particular,
R𝑛 \ spt𝜇 = {𝑥 ∈ R𝑛 : 𝜇(𝐵𝑟 (𝑥)) = 0 for some 𝑟 > 0}.
20
Proposition 5.18. If 𝐸 is a set of locally finite perimeter in R𝑛 , then
spt𝜇 𝐸 = {𝑥 ∈ R𝑛 : 0 < |𝐸 ∩ 𝐵𝑟 (𝑥)| < 𝜔𝑛 𝑟 𝑛 , ∀𝑟 > 0}. (5.4)
Moreover, there exists a Borel set 𝐹 such that
|𝐸Δ𝐹 | = 0, spt𝜇 𝐹 = 𝜕𝐹.
This means, when dealing with set of locally finite perimeter 𝐸, we could
always assume that
spt𝜇 𝐸 = 𝜕𝐸.

Remark 5.19. Later on we will prove that

spt𝜇 𝐸 = 𝜕 ∗ 𝐸,
where 𝜕 ∗ 𝐸 is the so-called reduced boundary of 𝐸, and hence we could
always assume that the topological boundary of a set of finite perimeter 𝐸
is the closure of its reduced boundary:
𝜕𝐸 = 𝜕 ∗ 𝐸.
Proof of Proposition 5.18. Step 1. We characterize the support of 𝜇 𝐸 .
If 𝑥 ∈ R𝑛 is such that |𝐸 ∩ 𝐵𝑟 (𝑥)| = 0 for some 𝑟 > 0, then
∫ ∫
0= ∇𝜙dL = 𝑛
𝜙d𝜇 𝐸 , ∀𝜙 ∈ 𝐶𝑐∞ (𝐵𝑟 (𝑥)).
𝐸 R𝑛
By definition of |𝜇 𝐸 |, we obtain |𝜇 𝐸 |(𝐵𝑟 (𝑥)) = 0, so that 𝑥 ∉ spt𝜇 𝐸 .
On the other hand, if 𝑥 ∈ R𝑛 is such that |𝐸 ∩ 𝐵𝑟 (𝑥)| = 𝜔𝑛 𝑟 𝑛 = |𝐵𝑟 (𝑥)|,
then |𝐵𝑟 (𝑥) \ 𝐸 | = 0, and hence
∫
0= ∇𝜙dL 𝑛
𝐵𝑟 (𝑥)
∫ ∫ ∫
= ∇𝜙dL = 𝑛
∇𝜙dL = 𝑛
𝜙d𝜇 𝐸 , ∀𝜙 ∈ 𝐶𝑐∞ (𝐵𝑟 (𝑥)),
(𝐵𝑟 (𝑥)\𝐸)∪𝐸 𝐸 R𝑛
implying that 𝑥 ∉ spt𝜇 𝐸 . In particular, these prove that
spt𝜇 𝐸 ⊂ {𝑥 ∈ R𝑛 : 0 < |𝐸 ∩ 𝐵𝑟 (𝑥)| < 𝜔𝑛 𝑟 𝑛 , ∀𝑟 > 0}.
Conversely, if 𝑥 ∉ spt𝜇 𝐸 , then there exists some 𝑟 > 0 such that |𝜇 𝐸 |(𝐵𝑟 (𝑥)) =
0, and hence
∫ ∫ ∫
0= 𝜙d𝜇 𝐸 = ∇𝜙dL = 𝑛
𝜒𝐸 ∇𝜙dL 𝑛 , ∀𝜙 ∈ 𝐶𝑐∞ (𝐵𝑟 (𝑥)).
R𝑛 𝐸 R𝑛
This implies, there exists some 𝑐 ∈ R such that 𝜒𝐸 = 𝑐-L 𝑛 a.e. in 𝐵𝑟 (𝑥).
Note that 𝑐 = 0 or 1, and hence
|𝐸 ∩ 𝐵𝑟 (𝑥)| = 0 or 𝜔𝑛 𝑟 𝑛 ,
21
which implies that
{𝑥 ∈ R𝑛 : 0 < |𝐸 ∩ 𝐵𝑟 (𝑥)| < 𝜔𝑛 𝑟 𝑛 , ∀𝑟 > 0} ⊂ spt𝜇 𝐸 .
Combining the above facts, we obtain (5.4).
Step 2. We construct the desired set 𝐹.
First we define two disjoint open sets by setting
𝐴0 ={𝑥 ∈ R𝑛 : there exists 𝑟 > 0, s.t. |𝐸 ∩ 𝐵𝑟 (𝑥)| = 0},
𝐴1 ={𝑥 ∈ R𝑛 : there exists 𝑟 > 0, s.t. |𝐸 ∩ 𝐵𝑟 (𝑥)| = 𝜔𝑛 𝑟 𝑛 }.
To see that these sets are open, for example, take any 𝑥 1 ∈ 𝐴1 with corre-
sponding 𝑟 1 > 0, then by definition
𝐵𝑟1 (𝑥 1 ) \ 𝐸 = 0.
It then follows that for any 𝑦 ∈ 𝐵 𝑟1 (𝑥 1 ), there holds
2

𝐵 𝑟1 (𝑦) \ 𝐸 ≤ 𝐵𝑟1 (𝑥 1 ) \ 𝐸 = 0,
4

so that any 𝑦 ∈ 𝐵 𝑟1 (𝑥1 ) satisfies 𝑦 ∈ 𝐴1 since

2
𝑟1 𝑛
|𝐸 ∩ 𝐵 𝑟1 (𝑦)| = 𝜔𝑛 ( ) .
4 4
This shows that 𝐴1 is open.
Next, note that for every 𝑥 ∈ 𝐴0 , there exists some 𝑟 𝑥 such that |𝐸 ∩
𝐵𝑟 𝑥 (𝑥)| = 0. We suppose 𝑟 𝑥 ≤ 1 otherwise simply replace it by 1. The
union of balls Ø
𝐵 𝑟5𝑥 (𝑥)
𝑥∈𝐴0
forms a covering of 𝐴0 . So by 5-times covering theorem, there exists a
countable subcollection {𝑥 ℎ , 𝑟 ℎ := 𝑟 𝑥 ℎ } such that
Ø Ø
𝐴0 ⊂ 𝐵 𝑟5𝑥 (𝑥) ⊂ 𝐵𝑟 ℎ (𝑥 ℎ ),
𝑥∈𝐴0 ℎ∈N

where |𝐸 ∩ 𝐵𝑟 ℎ (𝑥 ℎ )| = 0 for all ℎ ∈ N. Therefore,

|𝐸 ∩ 𝐴0 | = 0.
On the other hand, if we understand the set 𝐴1 in the way of set complement
as
𝐴1 = {𝑥 ∈ R𝑛 : there exists 𝑟 > 0, s.t. |𝐸 𝑐 ∩ 𝐵𝑟 (𝑥)| = 0}.
Then a similar argument shows that
| 𝐴1 \ 𝐸 | = |𝐸 𝑐 ∩ 𝐴1 | = 0.
Finally, our desired set 𝐹 is defined as
𝐹 := ( 𝐴1 ∪ 𝐸) \ 𝐴0 ,
22
which is of course Borel, with
|𝐹 \ 𝐸 | ≤ | 𝐴1 \ 𝐸 | = 0, |𝐸 \ 𝐹 | ≤ |𝐸 ∩ 𝐴0 | = 0,
that is, |𝐸Δ𝐹 | = 0. Moreover, by construction
𝐴1 ⊂ int(𝐹), 𝐹 ⊂ R𝑛 \ 𝐴 0 .
Thus
𝜕𝐹 = 𝐹 \ int(𝐹) ⊂ R𝑛 ∩ ( 𝐴0𝑐 ) ∩ ( 𝐴1𝑐 ) = R𝑛 ∩ ( 𝐴0 ∪ 𝐴1 ) 𝑐 = R𝑛 \ ( 𝐴0 ∪ 𝐴1 ).

On the other hand, from Step 1 and Proposition 5.8 we know,

R𝑛 \ ( 𝐴0 ∪ 𝐴1 ) = spt𝜇 𝐸 = spt𝜇 𝐹 ⊂ 𝜕𝐹.
In particular, this shows that 𝐹 is the desired set and completes the proof. □

5.5. Regularization
We study in this subsection the properties of convolutions of characteristic
functions of sets of locally finite perimeter with symmetric mollified (recall
Definition 2.1). Namely, given a set of locally finite perimeter 𝐸 in R𝑛 , so
that 𝜒𝐸 ∈ 𝐿 1𝑙𝑜𝑐 (R𝑛 ), for any 𝜖 > 0 one has
∫ ∫
( 𝜒𝐸 ∗ 𝜂𝜖 )(𝑥) = 𝜂𝜖 (𝑥 − 𝑦) 𝜒𝐸 (𝑦)d𝑦 = 𝜂𝜖 (𝑥 − 𝑦)d𝑦, 𝑥 ∈ R𝑛 .
R𝑛 𝐸∩𝐵 𝜖 (𝑥)

Clearly, 0 ≤ ( 𝜒𝐸 ∗ 𝜂𝜖 ) ≤ 1, with
(
1, if |𝐵(𝑥, 𝜖) \ 𝐸 | = 0,
( 𝜒𝐸 ∗ 𝜂𝜖 )(𝑥) =
0, if |𝐵(𝑥, 𝜖) ∩ 𝐸 | = 0.
Think of the case that 𝐸 is a bounded open set with smooth boundary, then
it is essential that
∇( 𝜒𝐸 ∗ 𝜂𝜖 )(𝑥) ≈ −𝜖 −1 𝜈 𝐸 (𝑦 𝑥 ), if dist(𝑥, 𝜕𝐸) < 𝜖,
where 𝑦 𝑥 is the nearest point projection of 𝑥 onto 𝜕𝐸; while ∇( 𝜒𝐸 ∗𝜂𝜖 )(𝑥) = 0
if dist(𝑥, 𝜕𝐸) > 𝜖. Thus, we expect
∫
|{𝑥 ∈ R𝑛 : dist(𝑥, 𝜕𝐸) < 𝜖 }| 𝜖H 𝑛−1 (𝜕𝐸)
|∇( 𝜒𝐸 ∗ 𝜂𝜖 )(𝑥)|d𝑥 ≈ ≈ = 𝑃(𝐸).
R𝑛 𝜖 𝜖
We next prove that this is also valid for sets of locally finite perimeter.
Definition 5.20 (𝜖-regularization of (vector-valued) Radon measures). Given
𝜇 a R𝑚 -valued Radon measure on R𝑛 , then the function (𝜇 ∗ 𝜂𝜖 ) : R𝑛 → R𝑚
is defined as
∫
(𝜇 ∗ 𝜂𝜖 )(𝑥) = 𝜂𝜖 (𝑥 − 𝑦)d𝜇(𝑦), 𝑥 ∈ R𝑛 .
R𝑛
23
As before, (𝜇 ∗ 𝜂𝜖 ) ∈ 𝐶 ∞ (R𝑛 , R𝑚 ), with
∫
∇(𝜇 ∗ 𝜂𝜖 )(𝑥) = ∇𝜂𝜖 (𝑥 − 𝑦)d𝜇(𝑦).
R𝑛
The 𝜖-regularization of 𝜇, denoted by 𝜇𝜖 , is the R𝑚 -valued Radon measure
on R𝑛 defined by
∫
𝜇𝜖 (𝜑) = ⟨𝜑(𝑥), (𝜇 ∗ 𝜂𝜖 )(𝑥)⟩ d𝑥, ∀𝜑 ∈ 𝐶𝑐0 (R𝑛 , R𝑚 ).
R𝑛

Proposition 5.21. If 𝜇 is a R𝑚 -valued Radon measure on R𝑛 , then

𝜇𝜖 → 𝜇, as 𝜖 ↘ 0.
Proof. Note that 𝜑𝜖 → 𝜑 in 𝐶𝑐0 (R𝑛 ) as 𝜖 ↘ 0, and by Fubini’s theorem
∫ ∫
⟨𝜑, d𝜇𝜖 ⟩ = 𝜇𝜖 (𝜑) = ⟨𝜑𝜖 (𝑦), d𝜇(𝑦)⟩ .
R𝑛 R𝑛
Thus ∫ ∫
lim ⟨𝜑, d𝜇𝜖 ⟩ = ⟨𝜑(𝑦), d𝜇(𝑦)⟩ ,
𝜖↘0 R𝑛 R𝑛
which proves that 𝜇𝜖 converges to 𝜇 as measure. □
Exercise 5.22. If 𝜇 is a R𝑚 -valued Radon measure on R𝑚 , prove that
|𝜇𝜖 | → |𝜇|, as 𝜖 ↘ 0.
Proposition 5.23. If 𝐸 is a set of locally finite perimeter in R𝑛 , then
(𝜇 𝐸 )𝜖 = −∇( 𝜒𝐸 ∗ 𝜂𝜖 )L 𝑛 , ∀𝜖 > 0,
and
−∇( 𝜒𝐸 ∗ 𝜂𝜖 )L 𝑛 → 𝜇 𝐸 , |∇( 𝜒𝐸 ∗ 𝜂𝜖 )| L 𝑛 → |𝜇 𝐸 |, as 𝜖 ↘ 0.
Conversely, if 𝐸 is a Lebesgue measurable set in R𝑛 such that
∫
lim sup |∇( 𝜒𝐸 ∗ 𝜂𝜖 )(𝑥)|d𝑥 < ∞, ∀ compact 𝐾,
𝜖↘0 𝐾
then 𝐸 has to be a set of locally finite perimeter.
Proof. The starting point is that, by Morse-Sard theorem (Theorem D.1),
for a.e. 𝑡 > 0 the super-level set { 𝜒𝐸 ∗ 𝜂𝜖 > 𝑡} is an open set with smooth
boundary { 𝜒𝐸 ∗ 𝜂𝜖 = 𝑡}, as 𝜖 ↘ 0 these smooth sets converge to 𝐸.
By (5.1), for every 𝑥 ∈ R𝑛
∫
(𝜇 𝐸 ∗ 𝜂𝜖 )(𝑥) = 𝜂𝜖 (𝑥 − 𝑦)d𝜇 𝐸 (𝑦)
R𝑛
∫
=− ∇𝜂𝜖 (𝑥 − 𝑦)d𝑦 = −∇( 𝜒𝐸 ∗ 𝜂𝜖 )(𝑥).
𝐸
24
Thus
∫
(𝜇 𝐸 )𝜖 (𝜑) = ⟨𝜑(𝑥), (𝜇 𝐸 ∗ 𝜂𝜖 )(𝑥)⟩ d𝑥
R𝑛
∫
=− ⟨𝜑(𝑥), ∇( 𝜒𝐸 ∗ 𝜂𝜖 )(𝑥)⟩ dL 𝑛 (𝑥), ∀𝜑 ∈ 𝐶𝑐0 (R𝑛 , R𝑚 ),
R𝑛

from which the first assertion follows.

The second assertion follows from Proposition 5.21 and Exercise 5.22.
To prove the last assertion, note that by compactness of vector-valued
Radon measures (Lemma B.3), there exists a R𝑛 -valued Radon measure 𝜇
on R𝑛 and a sequence 𝜖 ℎ ↘ 0, such that −∇( 𝜒𝐸 ∗ 𝜂𝜖 )L 𝑛 → 𝜇 as measures.
Thus, we have for any 𝜙 ∈ 𝐶𝑐1 (R𝑛 )
∫ ∫
𝜙d𝜇 = − lim 𝜙(𝑥)∇( 𝜒𝐸 ∗ 𝜂𝜖 ℎ )d𝑥
R𝑛 ℎ→∞ R𝑛
∫ ∫
= − lim 𝜙(𝑥)d𝑥 𝜒𝐸 (𝑦)∇𝜂𝜖 ℎ (𝑥 − 𝑦)d𝑦
ℎ→∞ R𝑛 R𝑛
∫ ∫
= − lim d𝑦 𝜙(𝑥)∇𝜂𝜖 ℎ (𝑥 − 𝑦)d𝑥
ℎ→∞ 𝐸 R𝑛
∫ ∫
= lim d𝑦 𝜂𝜖 ℎ (𝑥 − 𝑦)∇𝜙(𝑥)d𝑥
ℎ→∞ 𝐸 R𝑛
∫ ∫
= lim (∇𝜙)𝜖 ℎ (𝑦)d𝑦 = ∇𝜙dL 𝑛 .
ℎ→∞ 𝐸 𝐸

The last assertion then follows from Proposition 5.2. □

Exercise 5.24. Prove that if 𝐸 and 𝐹 are sets of (locally) finite perimeter in
R𝑛 , then so are 𝐸 ∪ 𝐹 and 𝐸 ∩ 𝐹, and for any open 𝐴 ⊂ R𝑛 ,

𝑃(𝐸 ∪ 𝐹; 𝐴) + 𝑃(𝐸 ∩ 𝐹; 𝐴) ≤ 𝑃(𝐸; 𝐴) + 𝑃(𝐹; 𝐴).

[hint: let 𝑢 𝜖 = 𝜒𝐸 ∗ 𝜂𝜖 , 𝑣 𝜖 = 𝜒𝐹 ∗ 𝜂𝜖 , then

𝑢 𝜖 𝑣 𝜖 → 𝜒𝐸∩𝐹 , 𝑢 𝜖 + 𝑣 𝜖 − 𝑢 𝜖 𝑣 𝜖 → 𝜒𝐸∪𝐹 in 𝐿 1𝑙𝑜𝑐 (R𝑛 ),

Exercise 5.25. Prove that if 𝐸 and 𝐹 are sets of (locally) finite perimeter in
R𝑛 , then so is 𝐸 \ 𝐹, and for any open 𝐴 ⊂ R𝑛 ,

𝑃(𝐸 \ 𝐹; 𝐴) ≤ 𝑃(𝐸; 𝐴) + 𝑃(𝐹; 𝐴).

25
6. Geometric variational problems

6.1. Geometric variational problems and sets of finite perime-

ter
6.1.1. Plateau-type problems. The classical Plateau problem which con-
cerns minimizing area among surfaces with a given curve as boundary, is
one of the archetypical problems in Geometric Measure Theory.
In the sets of finite perimeter settings, the Plateau-type problem can be
formulated as follows:
Given a set 𝐴 ⊂ R𝑛 and a set 𝐸 0 of finite perimeter in R𝑛 , the Plateau-type
problem in 𝐴 with boundary data 𝐸 0 amounts to be minimizing 𝑃(𝐸) among
those sets of finite perimeter 𝐸 that coincide with 𝐸 0 outside 𝐴. Namely,
we consider
𝛾( 𝐴, 𝐸 0 ) := inf{𝑃(𝐸) : 𝐸 \ 𝐴 = 𝐸 0 \ 𝐴}. (6.1)
In this way, 𝐸 0 ∩ 𝜕 𝐴 serves as a “boundary condition”.
6.1.2. (relative) Isoperimetric problem. Given an open set 𝐴 ⊂ R𝑛 , the rel-
ative isoperimetric problem in 𝐴 amounts to the volume-constrained mini-
mization of the relative perimeter in 𝐴, namely
𝛼( 𝐴, 𝑚) = inf{𝑃(𝐸; 𝐴) : 𝐸 ⊂ 𝐴, |𝐸 | = 𝑚}, (6.2)
where 𝑚 ∈ (0, | 𝐴|). A minimizer 𝐸 in (6.2) (whose existence will be
proven later provided that 𝐴 is an open bounded set of finite perimeter),
normalized thanks to Proposition 5.18 that spt𝜇 𝐸 = 𝜕𝐸, is called a relative
isoperimetric set in 𝐴; the corresponding 𝑃(𝐸; 𝐴) is called the relative
isoperimetric profile. In particular, for 𝐴 = R𝑛 , (6.2) is just the classical
Euclidean isoperimetric problem.
6.1.3. Prescribed mean curvature problem. Interesting variational prob-
lems arise from the interaction between perimeter and potential energy,
which is defined as follows:
Definition 6.1 (potential energy). Given a Lebesgue measurable function
𝑔 : R𝑛 → R ∪ {∞}, the potential energy of 𝐸 associated with 𝑔 is defined
as ∫
G(𝐸) := 𝑔(𝑥)d𝑥.
𝐸
Given an open set 𝐴 ⊂ and a Lebesgue measurable function 𝑔 : R𝑛 →
R𝑛
R, the prescribed mean curvature problem amounts to be minimizing the
sum of the perimeter and potential energy among all sets in 𝐴, namely,
inf{𝑃(𝐸) + G(𝐸) : 𝐸 ⊂ 𝐴}. (6.3)
26
The terminology used here comes from the fact that if 𝑔 ∈ 𝐶 0 ( 𝐴), 𝐸 is a
minimizer in (6.3), and 𝐴∩𝜕𝐸 is a 𝐶 2 -hypersurface, then the mean curvature
𝐻 𝐸 of 𝐸 is eqaul to −𝑔 in 𝐴.

6.1.4. (capillarity) Equilibrium shapes. We study the equilibrium shapes

of a liquid confined in a given container. Since the work of Gauss, the
problem has been studied through the introduction of a free energy func-
tional. Precisely, if a liquid occupies a region 𝐸 inside a given container 𝐴
(mathematically, 𝐸 will be a set of finite perimeter and 𝐴 an open set with
sufficiently smooth boundary), then its free energy is given by
∫
𝜎 (𝑃(𝐸; 𝐴) − 𝛽𝑃(𝐸; 𝜕 𝐴)) + 𝑔(𝑥)d𝑥.
𝐸

Here 𝜎 > 0 models for surface tension at the interface between the liquid
and the other medium filling the container 𝐴. The coefficient 𝛽 is called the
relative adhesion coefficient between the fluid and the bounding solid walls
of the recipient. By Young’s law it satisfies |𝛽| ≤ 1. The term −𝜎𝛽𝑃(𝐸; 𝜕 𝐴)
is the so-called wetting energy. In all follows, we shall always set 𝜎 = 1 to
simplify the notation.
The geometric variational problem we should study is formulated as
follows: Given 𝛽 ∈ R, a bounded open set 𝐴 ⊂ R𝑛 , and a set of finite
perimeter 𝐸 ⊂ 𝐴, we set
F𝛽 (𝐸; 𝐴) = 𝑃(𝐸; 𝐴) − 𝛽𝑃(𝐸; 𝜕 𝐴)

for the total surface energy, and (recall) denote by

∫
G(𝐸) = 𝑔(𝑥)d𝑥
𝐸

the potential energy associated with a given Borel function 𝑔 : R𝑛 → R.

The capillary problem amounts to be the (volume-constrained) minimizing
problem of
inf{F𝛽 (𝐸; 𝐴) + G(𝐸) : 𝐸 ⊂ 𝐴, |𝐸 | = 𝑚},
𝑚 ∈ (0, | 𝐴|).
For regular enough minimizer, it satisfies
(𝑖) 𝜕𝐸 ∩ 𝐴 has mean curvature equal to −𝑔 + 𝜆 for some constant 𝜆 ∈ R;
(𝑖𝑖) The Young’s law holds:
⟨𝜈 𝐸 , 𝜈 𝐴 ⟩ = −𝛽 on bdry( 𝐴 ∩ 𝜕𝐸).

Such (hyper)surfaces are called capillary surfaces.

27
6.2. Compactness for sets of finite perimeter
In this subsection we prove the following compactness theorem. These
results are applications of the regularization of sets of finite perimeter, and
play as the key of the so-called Direct Method.
Theorem 6.2. If 𝑅 > 0 and {𝐸 ℎ } are sets of finite perimeter in R𝑛 , with
sup 𝑃(𝐸 ℎ ) < ∞,
ℎ∈N
𝐸 ℎ ⊂ 𝐵𝑅 , ∀ℎ ∈ N,
then there exists a set of finite perimeter 𝐸 ⊂ R𝑛 and ℎ(𝑘) → ∞ as 𝑘 → ∞,
with
𝐸 ℎ(𝑘) → 𝐸, 𝜇 𝐸 ℎ (𝑘 ) → 𝜇 𝐸 , 𝐸 ⊂ 𝐵 𝑅 .
Remark 6.3. (1) One cannot conclude compactness without assuming
the uniform boundness of 𝐸 ℎ . For example, if {𝑥 ℎ } ℎ∈N ⊂ R𝑛 is such
that |𝑥 ℎ | → ∞, then the sequence 𝐸 ℎ := 𝐵1 (𝑥 ℎ ) satisfies |𝐸 ℎ | = 𝜔𝑛
and 𝑃(𝐸 ℎ ) = 𝑛𝜔𝑛 for every ℎ ∈ N, but for any Lebesgue measurable
set 𝐸 with |𝐸 | = 𝜔𝑛 , there holds
|𝐸Δ𝐸 ℎ | → 2𝜔𝑛 .
Thus {𝐸 ℎ } ℎ∈N does not admit any converging subsequence.
(2) Note however, it is clear that {𝐸 ℎ } ℎ∈N locally converges to the empty
set, so that compactness with respect to local convergence still holds.
Proof of Theorem 6.2. Step 1. The poincaré inequality holds: For 𝑢 ∈
𝐶 1 (R𝑛 ),
√
∫ ∫
𝑛
𝑢 − (𝑢)𝑄(𝑥,𝑟) dL ≤ 𝑛𝑟 |∇𝑢|dL 𝑛 ,
𝑄(𝑥,𝑟) 𝑄(𝑥,𝑟)
⨏
where 𝑄(𝑥, 𝑟) = 𝑥 + (0, 𝑟) 𝑛 , (𝑢)𝑄(𝑥,𝑟) = 𝑄(𝑥,𝑟) 𝑢dL 𝑛 .
The proof follows from the classical one.
Step 2. We show that If 𝐸 is a set of finite perimeter in R𝑛 with |𝐸 | < ∞,
then for any 𝑟 > 0 there exists a finite union 𝑇 of disjoint cubes of side
length 𝑟, with
√
|𝐸Δ𝑇 | ≤ 𝑛𝑟 𝑃(𝐸).
Namely, 𝐸 can be approximated by the finite union of disjoint cubes.
To do this we let {𝑄 ℎ } ℎ∈N be a disjoint family of open cubes of side length
Ð
𝑟 such that R𝑛 = ℎ∈N 𝑄 ℎ . For 𝜖 > 0 and 𝑢 = ( 𝜒𝐸 ∗ 𝜂𝜖 ), we have by Step 1
∫ ∑︁ ∫ ∫
1 ∑︁
|∇𝑢| = |∇𝑢| ≥ √ |𝑢 − (𝑢)𝑄 ℎ |.
R𝑛 ℎ∈N 𝑄 ℎ 𝑛𝑟 ℎ∈N 𝑄 ℎ

28
Letting 𝜖 ↘ 0, by virtue of Proposition 5.23 we find
√ ∑︁ ∫ ∑︁ ∫ |𝑄 ℎ ∩ 𝐸 |
𝑛𝑟 𝑃(𝐸) ≥ | 𝜒𝐸 − ( 𝜒𝐸 ) 𝑄 ℎ | = | 𝜒𝐸 − 𝑛
|
ℎ∈N 𝑄 ℎ ℎ∈N 𝑄 ℎ
𝑟
∑︁ ∫ |𝑄 ℎ ∩ 𝐸 |
= | 𝜒𝐸 − |
ℎ∈N (𝑄 ℎ ∩𝐸)∪(𝑄 ℎ \𝐸)
𝑟𝑛
∑︁ |𝑄 ℎ ∩ 𝐸 | 𝑄ℎ ∩ 𝐸
= |𝐸 ∩ 𝑄 ℎ | 1 − 𝑛
+ |𝑄 ℎ \ 𝐸 || |
ℎ∈N
𝑟 𝑟𝑛
∑︁ |𝑄 ℎ ∩ 𝐸 ||𝑄 ℎ \ 𝐸 |
=2 .
ℎ∈N
𝑟𝑛

Since |𝐸 | < ∞, it must be that |𝑄 ℎ ∩ 𝐸 | ≥ |𝑄2ℎ | = 𝑟2 for at most finitely many

𝑛

cubes 𝑄 ℎ . Thus, up to a permutation of index, we may assume that these

cubes are exactly the first 𝑁 elements of the sequence {𝑄 ℎ } ℎ∈N , namely,
𝑟𝑛 𝑟𝑛
|𝑄 ℎ ∩ 𝐸 | ≥ , if 1 ≤ ℎ ≤ 𝑁; |𝑄 ℎ \ 𝐸 | ≥ , if ℎ ≥ 𝑁 + 1.
2 2
Ð𝑁
As a result, if we let 𝑇 = ℎ=1 𝑄 ℎ , then we find, as desired,
𝑁 ∞
√ ∑︁ ∑︁
𝑛𝑟 𝑃(𝐸) ≥ |𝑄 ℎ \ 𝐸 | + |𝑄 ℎ ∩ 𝐸 | = |𝑇 \ 𝐸 | + |𝐸 \ 𝑇 | = |𝐸Δ𝑇 |.
ℎ=1 ℎ=𝑁+1
This completes Step 2.
To proceed, note that the set 𝑋 = {𝐸 ∈ M (L 𝑛 ) : |𝐸 | < ∞} is a complete
metric space endowed with the distance 𝑑 (𝐸, 𝐹) = |𝐸Δ𝐹 | = ∥ 𝜒𝐸 −𝜒𝐹 ∥ 𝐿 1 (R𝑛 )
1
.
Step 3. We prove that each set 𝑌𝑅,𝑝 ⊂ 𝑋, defined as
𝑌𝑅,𝑝 := {𝐸 ∈ M (L 𝑛 ) : 𝐸 ⊂ 𝐵 𝑅 , 𝑃(𝐸) ≤ 𝑝}, 𝑅, 𝑝 ∈ (0, ∞),
is 𝑑-compact.
First note that by Proposition 5.15, the set 𝑌𝑅,𝑝 is closed. By a standard
diagonal argument, we just have to show that for every 𝜎 > 0, there exist
𝑀 ∈ N and {𝑇ℎ } ℎ=1
𝑀 ⊂ 𝑋 with

min 𝑑 (𝐸, 𝑇ℎ ) ≤ 𝜎, ∀𝐸 ∈ 𝑌𝑅,𝑝 ,

1≤ℎ≤𝑀
which shows that 𝑌𝑅,𝑝 is totally bounded. The assertion will then follow
since any closed and totally bounded subset of a complete metric space is
compact.
1Here we identify 𝐸 and 𝐹 if |𝐸Δ𝐹 | = 0 by virtue of Proposition 5.8, that is to say, we
do not distinguish the cases 𝐸 ⊂ 𝐵 𝑅 and 𝐸 ⊂ 𝐵 𝑅 in the context of sets of (locally) finite
perimeter.
29
 √
Indeed, let 𝑟 > 0 be such that 𝑛𝑟 𝑝 ≤ 𝜎, and let {𝑄 ℎ } ℎ∈N be the family
of cubes associated with 𝑟 resulting from Step 2. The family {𝑆 ℎ } ℎ=1 𝑁 of the

cubes from {𝑄 ℎ } ℎ∈N intersecting 𝐵 𝑅 is finite, thus the family {𝑇ℎ } ℎ=1
𝑀 of the

finite union of cubes from {𝑆 ℎ } ℎ=1

𝑁 is finite too. By Step 2, every 𝐸 ∈ 𝑌
𝑅,𝑝
√
admits some 𝑇ℎ such that |𝐸Δ𝑇ℎ | ≤ 𝑛𝑟 𝑝 ≤ 𝜎, as desired.
Concluding step.
By assumption, {𝐸 ℎ } ℎ∈N ⊂ 𝑌𝑅,𝑝 for some 𝑅, 𝑝 > 0. By Step 3, 𝑌𝑅,𝑝 is
𝑑-compact, and hence there exist 𝐸 ⊂ 𝐵 𝑅 and a sequence ℎ(𝑘) → ∞ as
𝑘 → ∞, such that 𝐸 ℎ(𝑘) → 𝐸. By Proposition 5.15, 𝐸 is a set of finite
perimeter in R𝑛 with 𝜇 𝐸 ℎ → 𝜇 𝐸 as (vector-valued) Radon measure. □
As discussed in the previous Remark, and taking into account account
the fact that it is useful to consider sequences of sets that are only of locally
finite perimeter, we give the following compactness result. Before that, we
need the following lemma, which is left as an exercise.
Lemma 6.4. If 𝐸 is a set of locally finite perimeter in R𝑛 , then for any 𝑅 > 0,
𝑃(𝐸 ∩ 𝐵 𝑅 ) ≤ 𝑃(𝐸; 𝐵 𝑅 ) + 𝑃(𝐵 𝑅 ). (6.4)
Corollary 6.5. If {𝐸 ℎ } ℎ∈N are sets of locally finite perimeter in R𝑛 with
sup 𝑃(𝐸 ℎ ; 𝐵 𝑅 ) < ∞, ∀𝑅 > 0,
ℎ∈N
then there exist 𝐸 of locally finite perimeter and ℎ(𝑘) → ∞ as 𝑘 → ∞, with
loc
𝐸 ℎ(𝑘) → 𝐸, 𝜇 𝐸 ℎ (𝑘 ) → 𝜇 𝐸 .
Proof. By the assumption and (6.4), for any fixed 𝑅 > 0, we have 𝑃(𝐸 ℎ ∩
𝐵 𝑅 ) ≤ 𝐶 (𝑛, 𝑅) for every ℎ. Thus we may apply Theorem 6.2 to the sets
{𝐸 ℎ ∩ 𝐵 𝑅 } ℎ∈N such that there exist ℎ(𝑘) → ∞ as 𝑘 → ∞ and 𝐹𝑅 such that
𝐸 ℎ ∩ 𝐵 𝑅 → 𝐹𝑅 .
By a standard diagonal argument, for 𝑗 = 1, 2, . . ., there exists a sequence
of sets of finite perimeter {𝐹 𝑗 } 𝑗 ∈N such that after relabeling,
𝐸 ℎ(𝑘) ∩ 𝐵 𝑗 → 𝐹 𝑗 as 𝑘 → ∞.
Up to null sets, 𝐹 𝑗 ⊂ 𝐹 𝑗+1 , and hence 𝐸 ℎ(𝑘) converges locally to 𝐸 :=
Ð
𝑗 ∈N 𝐹 𝑗 . By Proposition 5.15, we conclude that 𝐸 is also a set of locally
finite perimeter with 𝜇 𝐸 ℎ (𝑘 ) → 𝜇 𝐸 as vector-valued Radon measures. □

6.3. Existence of minimizers

Using compactness theorems, we prove existence of minimizers in the
sets of finite perimeter settings for the geometric variational problems listed
in Section 6.1. These results are achieved by means of the so-called Direct
Method. Roughly speaking, we first prove compactness of an arbitrary
30
minimizing sequence of admissible competitors, then show the minimality
of the limit by lower semi-continuity of convergence.
Proposition 6.6 (Existence of minimizers for the Plateau-type problem).
Let 𝐴 ⊂ R𝑛 be a bounded set and 𝐸 0 be a set of finite perimeter in R𝑛 .
Then there exists a set of finite perimeter 𝐸 such that 𝐸 \ 𝐴 = 𝐸 0 \ 𝐴 and
𝑃(𝐸) ≤ 𝑃(𝐹) for every competitor 𝐹 such that 𝐹 \ 𝐴 = 𝐸 0 \ 𝐴. In particular,
𝐸 is a minimizer in the variational problem (6.1).
Proof. First note that 𝛾 < ∞ since 𝐸 0 is itself an admissible competitor.
Now consider a minimizing sequence {𝐸 ℎ } ℎ∈N in (6.1), i.e.,
𝐸 ℎ \ 𝐴 = 𝐸 0 \ 𝐴, 𝑃(𝐸 ℎ ) ≤ 𝑃(𝐸 0 ), lim 𝑃(𝐸 ℎ ) = 𝛾.
ℎ→∞
If 𝑀ℎ = 𝐸 ℎ Δ𝐸 0 = (𝐸 ℎ \ 𝐸 0 ) ∪ (𝐸 0 \ 𝐸 ℎ ), then by Exercises 5.24, 5.25, 𝑀ℎ
is a set of finite perimeter with
𝑃(𝑀ℎ ) ≤ 2𝑃(𝐸 ℎ ) + 2𝑃(𝐸 0 ) ≤ 4𝑃(𝐸 0 ).
Since 𝐴 is bounded, we may use Theorem 6.2 to deduce that there exists a set
of finite perimeter 𝑀 such that, up to passing to a subsequence, 𝑀ℎ → 𝑀.
Since
𝐸 ℎ = (𝐸 0 ∪ 𝑀ℎ ) \ (𝐸 0 ∩ 𝑀ℎ ),
and 𝑀ℎ → 𝑀, we find 𝐸 ℎ → 𝐸 := (𝐸 0 ∪ 𝑀) \ (𝐸 0 \ 𝑀), which is a set of
finite perimeter thanks again to Exercises 5.24, 5.25. Moreover, we have
𝐸 \ 𝐴 = lim 𝐸 ℎ \ 𝐴 = 𝐸 0 \ 𝐴,
ℎ→∞
and, by Proposition 5.15,
𝛾 ≤ 𝑃(𝐸) ≤ lim inf 𝑃(𝐸 ℎ ) = 𝛾,
ℎ→∞
which shows that 𝐸 is indeed the minimizer. □
Proposition 6.7 (Existence of relative isoperimetric sets). If 𝐴 is an open
bounded set of finite perimeter and 𝑚 ∈ (0, | 𝐴|], then there exists a set of
finite perimeter 𝐸 ⊂ 𝐴 such that 𝑃(𝐸; 𝐴) = 𝛼( 𝐴, 𝑚) and |𝐸 | = 𝑚. In
particular, 𝐸 is a minimizer in the variational problem (6.2).
Proof. Let 𝐸 𝑡 = 𝐴 ∩ {𝑥 ∈ R𝑛 : 𝑥 𝑛 < 𝑡} (𝑡 ∈ R), by continuity there exists
𝑡 0 ∈ R such that |𝐸 𝑡0 | = 𝑚. By Exercise 5.24, 𝐸 𝑡 is a set of finite perimeter
that serves as a competitor so that 𝛼( 𝐴, 𝑚) < ∞.
Now let {𝐸 ℎ } ℎ∈N be a minimizing sequence in (6.2), i.e.,
𝐸 ℎ ⊂ 𝐴, |𝐸 ℎ | = 𝑚, lim 𝑃(𝐸 ℎ ; 𝐴) = 𝛼.
ℎ→∞
Next we use the following fact: If 𝐸 ℎ , 𝐴 are sets of finite perimeter, then so
is 𝐸 ∩ 𝐴, with
𝑃(𝐸 ℎ ∩ 𝐴) ≤ 𝑃(𝐸 ℎ ; 𝐴) + 𝑃( 𝐴).
31
We shall give the proof in due course, and we recall that if 𝐴 = 𝐵 𝑅 , this is
just (6.4). With this fact we have supℎ∈N 𝑃(𝐸 ℎ ) = supℎ∈N 𝑃(𝐸 ℎ ∩ 𝐴) < ∞.
Since 𝐴 is bounded, we may apply Theorem 6.2 to deduce that there exists
a set of finite perimeter 𝐸 ⊂ R𝑛 , such that after passing to a subsequence
𝐸 ℎ → 𝐸. In particular, 𝐸 ⊂ 𝐴 and |𝐸 | = limℎ→∞ |𝐸 ℎ | = 𝑚. Moreover, by
Proposition 5.15,
𝛼 ≤ 𝑃(𝐸; 𝐴) ≤ lim inf 𝑃(𝐸 ℎ ; 𝐴) = 𝛼,
ℎ→∞

which shows that 𝐸 is indeed the minimizer. □

Remark 6.8. Note, however, that the Direct Method can not be used to prove
the existence of minimizers of classical Euclidean isoperimetric problem,
i.e., when 𝐴 = R𝑛 .
Exercise 6.9. If 𝑔 : R𝑛 → [0, ∞) is a Lebesgue measurable function with
𝑔(𝑥) → ∞ as |𝑥| → ∞, and if 𝐴 is a (possibly unbounded) open set with
finite perimeter, then the variational problem
inf{𝑃(𝐸; 𝐴) + G(𝐸) : 𝐸 ⊂ 𝐴, |𝐸 | = 𝑚}
admits minimizers.
We conclude this section by stating the existence of minimizers of capil-
lary problems:
Proposition 6.10. If |𝛽| ≤ 1, 𝑔 ∈ 𝐿 1𝑙𝑜𝑐 (R𝑛 ), 𝐴 is an open bounded set of in
R𝑛 with 𝐶 2 -boundary, and 𝑚 ∈ (0, | 𝐴|), then there exists a minimizer of the
geometric variational problem
𝛾 = inf{F𝛽 (𝐸; 𝐴) + G(𝐸) : 𝐸 ⊂ 𝐴, |𝐸 | = 𝑚}.

The approach to this proposition will be provided in due course.

7. Approximated by open sets with smooth boundary

Theorem 7.1. If 𝐸 is a bounded set of locally finite perimeter in R𝑛 , then

there exists a sequence {𝐸 ℎ } ℎ∈N of open sets with smooth boundary in R𝑛 ,
such that
𝐸 ℎ → 𝐸, 𝑃(𝐸 ℎ ) → 𝑃(𝐸).

Proof. By smooth approximation of 𝐵𝑉 functions (Theorem 2.3) we know

that 𝜒𝐸 can be approximated by a sequence of smooth functions obtained
by mollifying 𝜒𝐸 , which, together with co-area formula of 𝐵𝑉 functions,
yields the desired result.
32
 Precisely, let 𝜖 > 0 and 𝑓𝜖 be the mollified function of 𝜒𝐸 as in the proof
of Theorem 2.3. Then by Theorem 4.4 and note that 0 ≤ 𝑓𝜖 ≤ 1, we have
∫ ∫ 1 ∫
|𝐷 𝑓𝜖 | = d𝑡 |𝐷 𝜒𝐸 𝜖 𝑡 |,
0
where 𝐸 𝜖𝑡 := {𝑥 : 𝑓𝜖 (𝑥) > 𝑡}.
By Exercise 2.4 (with 𝐴 = R𝑛 therein)
∫ ∫
|𝐷 𝜒𝐸 | = lim |𝐷 𝑓𝜖 |dL 𝑛 .
𝜖↘0

Claim. Let 0 < 𝑡 < 1, 𝜖 𝑗 → 0 as 𝑗 → ∞ and 𝐸 𝑗𝑡 = {𝑥 ∈ R𝑛 : 𝑓𝜖 𝑗 (𝑥) >

𝑡}, where 𝑓𝜖 𝑗 = 𝜂𝜖 𝑗 ∗ 𝜒𝐸 , there holds
∫ ∫
𝑛 1
| 𝜒𝐸 𝑗𝑡 − 𝜒𝐸 |dL ≤ | 𝑓𝜖 𝑗 − 𝜒𝐸 |dL 𝑛 .
min(𝑡, 1 − 𝑡)
To see this, note that
𝑓𝜖 𝑗 − 𝜒𝐸 > 𝑡 in 𝐸 𝑗𝑡 \ 𝐸,
𝜒𝐸 − 𝑓𝜖 𝑗 ≥ 1 − 𝑡 in 𝐸 \ 𝐸 𝑗𝑡 ,
so that
∫ ∫ ∫
𝑛 𝑛
| 𝑓𝜖 𝑗 − 𝜒𝐸 |dL ≥ | 𝑓𝜖 𝑗 − 𝜒𝐸 |dL + | 𝑓𝜖 𝑗 − 𝜒𝐸 |dL 𝑛
𝐸 𝑗𝑡 \𝐸 𝐸\𝐸 𝑗𝑡
≥𝑡|𝐸 𝑗𝑡 \ 𝐸 | + (1 − 𝑡)|𝐸 \ 𝐸 𝑗 |
∫
≥ min(𝑡, 1 − 𝑡) | 𝜒𝐸 𝑗 − 𝜒𝐸 |dL 𝑛 ,

which proves the claim.

Using this claim, we have for every 𝑡 ∈ (0, 1) that
𝜒𝐸 𝑗𝑡 → 𝜒𝐸 ,
where, upon possibly rechoosing the sequence 𝜖 𝑗 , 𝐸 𝑗𝑡 are open sets with
smooth boundary by Morse-Sard theorem (Theorem D.1), and hence by
Lower semi continuity
∫ ∫
𝑃(𝐸) = |𝐷 𝜒𝐸 | ≤ lim inf |𝐷 𝜒𝐸 𝑗𝑡 | = lim inf 𝑃(𝐸 𝑗𝑡 ).
𝑗→∞ 𝑗→∞

Therefore
∫ ∫ 1 ∫ ∫
|𝐷 𝜒𝐸 | = lim |𝐷 𝑓𝜖 𝑗 | ≥ d𝑡 lim inf |𝐷 𝜒𝐸 𝑗𝑡 | ≥ |𝐷 𝜒𝐸 |.
𝑗→∞ 0 𝑗→∞

This means, for a.e. 𝑡 ∈ (0, 1),

∫ ∫
lim inf |𝐷 𝜒𝐸 𝑗𝑡 | = |𝐷 𝜒𝐸 |.
𝑗→∞
33
Taking a further subsequence of 𝜖 𝑗 , we can ensure that the above equality
holds with lim rather than lim inf and the theorem is thus proven. □

A refined version of the theorem can be found in [2], and the idea of proof
is essentially the same, namely, using regularization, co-area formula, and
Morse-Sard Theorem.
Theorem 7.2. A Lebesgue measurable set 𝐸 ⊂ R𝑛 is of locally finite perime-
ter if and only if there exists a sequence {𝐸 ℎ } ℎ∈N of open sets with smooth
boundary in R𝑛 , and 𝜖 ℎ ↘ 0, such that
loc
𝐸 ℎ → 𝐸, sup 𝑃(𝐸 ℎ ; 𝐵 𝑅 ) < ∞, ∀𝑅 > 0, |𝜇 𝐸 ℎ | → |𝜇 𝐸 |.
ℎ∈N

In particular, 𝑃(𝐸 ℎ ; 𝐹) → 𝑃(𝐸; 𝐹) whenever 𝑃(𝐸; 𝜕𝐹) = 0. Moreover,

(𝑖) If |𝐸 | < ∞, then 𝐸 ℎ → 𝐸.
(1) If 𝑃(𝐸) < ∞, then 𝑃(𝐸 ℎ ) → 𝑃(𝐸).

8. Reduced boundary

Definition 8.1. The reduced boundary of a set of locally finite perimeter 𝐸

in R𝑛 , denoted by 𝜕 ∗ 𝐸, is the set of those 𝑥 ∈ spt𝜇 𝐸 such that the limit
𝜇 𝐸 (𝐵𝑟 (𝑥))
lim exists and belongs to S𝑛−1 . (8.1)
𝑟↘0 |𝜇 𝐸 |(𝐵𝑟 (𝑥))
A Borel function 𝜈 𝐸 : 𝜕 ∗ 𝐸 → S𝑛−1 , defined by setting
𝜇 𝐸 (𝐵𝑟 (𝑥))
𝜈 𝐸 (𝑥) := lim , 𝑥 ∈ 𝜕 ∗ 𝐸,
𝑟↘0 |𝜇 𝐸 |(𝐵𝑟 (𝑥))
is called the (measure-theoretic) outer unit normal to 𝐸. By Lebesgue point
theorem and Lebesgue-Besicovitch differentiation Theorem (Theorem B.4),
we have
𝜇 𝐸 = 𝜈 𝐸 |𝜇 𝐸 |⌞𝜕 ∗ 𝐸, (8.2)
so that (5.1) takes the form
∫ ∫
div𝜙dL =𝑛
⟨𝜙, 𝜈 𝐸 ⟩ d|𝜇 𝐸 |, ∀𝜙 ∈ 𝐶𝑐1 (R𝑛 , R𝑛 ).
𝐸 𝜕∗ 𝐸

Example 8.2. In the case that 𝐸 is an open set with 𝜕𝐸 of class 𝐶 1 , one
has 𝜕 ∗ 𝐸 = 𝜕𝐸 and the measure-theoretic outer unit normal coincides with
classical notion of outer unit normal.
34
8.1. Basic properties
Proposition 8.3. If 𝐸 is a set of locally finite perimeter and |𝐸Δ𝐹 | = 0, then
𝜇 𝐸 = 𝜇 𝐹 and therefore 𝜕 ∗ 𝐸 = 𝜕 ∗ 𝐹. Indeed, the reduced boundary 𝜕 ∗ 𝐸 is
uniquely determined by the Gauss-Green measure 𝜇 𝐸 of 𝐸.
Proof. This follows directly from Proposition 5.8 and the definition of re-
duced boundary. □
Proposition 8.4. If 𝐸 is a set of locally finite perimeter in R𝑛 , then
(1) 𝜕 ∗ 𝐸 ⊂ spt𝜇 𝐸 ⊂ 𝜕𝐸.
(2) spt𝜇 𝐸 = 𝜕 ∗ 𝐸.
(3) Up to a modification on sets of (Lebesgue) measure zero, we may
always assume that
𝜕 ∗ 𝐸 = 𝜕𝐸. (8.3)
Proof. By Proposition 5.18 and the definition of 𝜕 ∗ 𝐸, (1) follows.
By (8.2), 𝜇 𝐸 is concentrated on 𝜕 ∗ 𝐸, and hence on 𝜕 ∗ 𝐸. By definition of
support, one has spt𝜇 𝐸 ⊂ 𝜕 ∗ 𝐸. (2) then follows.
(3) is a direct consequence of (2) and Proposition 5.18. □
Definition 8.5 (blow-ups of set). For 𝑥 ∈ R𝑛 and 𝑟 > 0, the blow-ups of a
set 𝐸 is defined as
𝐸 −𝑥
𝐸 𝑥,𝑟 := = Φ𝑥,𝑟 (𝐸),
𝑟
where Φ𝑥,𝑟 (𝑦) = 𝑦−𝑥𝑟 for any 𝑦 ∈ R .
𝑛

Definition 8.6 (density). Given 𝐸 ⊂ R𝑛+1 and 𝑥 ∈ R𝑛 , if the limit

|𝐸 ∩ 𝐵𝑟 (𝑥)|
Θ𝑛 (𝐸)(𝑥) := lim
𝑟↘0 𝜔𝑛 𝑟 𝑛
exists, it is then called the 𝑛-dimensional density of 𝐸 at 𝑥. Θ𝑛 (𝐸) defines a
Borel function on R𝑛 .
Given 𝑡 ∈ [0, 1], the set of points of density 𝑡 of 𝐸 is defined as
𝐸 (𝑡) := {𝑥 ∈ R𝑛 : Θ𝑛 (𝐸)(𝑥) = 𝑡}.
By Lebesgue’s points theorem,
loc
𝑥 ∈ 𝐸 (1) iff 𝐸 𝑥,𝑟 → R𝑛 as 𝑟 ↘ 0,
loc
𝑥 ∈ 𝐸 (0) iff 𝐸 𝑥,𝑟 → ∅ as 𝑟 ↘ 0.
Lemma 8.7. If 𝐸 is a set of locally finite perimeter in R𝑛 , then for 𝑥 ∈ R𝑛
and 𝑟 > 0, then 𝐸 𝑥,𝑟 is a set of locally finite perimeter in R𝑛 with
(Φ𝑥,𝑟 )# 𝜇 𝐸
𝜇 𝐸 𝑥,𝑟 = . (8.4)
𝑟 𝑛−1
35
Proof. For test function 𝜑 ∈ 𝐶𝑐1 (R𝑛 ), write 𝜑𝑥,𝑟 = 𝜑 ◦ Φ𝑥,𝑟 , then ∇𝜑𝑥,𝑟 =
𝑟 −1 (∇𝜑 ◦ Φ𝑥,𝑟 ), and hence change of variables gives
∫ ∫ ∫
𝑛 1 𝑛 1
∇𝜑dL = 𝑛 (∇𝜑 ◦ Φ𝑥,𝑟 )dL = 𝑛−1 ∇𝜑𝑥,𝑟 dL 𝑛
𝐸 𝑥,𝑟 𝑟 𝐸 𝑟 𝐸
∫ ∫
1 1
= 𝑛−1 𝜑𝑥,𝑟 d𝜇 𝐸 = 𝑛−1 𝜑d(Φ𝑥,𝑟 )# 𝜇 𝐸 .
𝑟 R𝑛 𝑟 R𝑛
(Φ ) 𝜇
# 𝐸
Since 𝑥,𝑟 𝑟 𝑛−1
is a vector-valued Radon measure, by Proposition 5.2 𝐸 𝑥,𝑟 is
a set of locally finite perimeter with 𝜇 𝐸 𝑥,𝑟 of the desired form. □
Lemma 8.8 (Intersection with ball). If 𝐸 is a set of locally finite perimeter
in R𝑛 and 𝑥 ∈ R𝑛 , then for every 𝑟 > 0, 𝐸 ∩ 𝐵𝑟 (𝑥) is a set of finite perimeter
in R𝑛 . Moreover, for a.e. 𝑟 > 0,
𝜇 𝐸∩𝐵𝑟 (𝑥) =𝜈 𝐵𝑟 (𝑥) H 𝑛−1 ⌞(𝐸 ∩ 𝜕𝐵𝑟 (𝑥)) + 𝜇 𝐸 ⌞𝐵𝑟 (𝑥),
|𝜇 𝐸∩𝐵𝑟 (𝑥) | =H 𝑛−1 ⌞(𝐸 ∩ 𝜕𝐵𝑟 (𝑥)) + |𝜇 𝐸 |⌞𝐵𝑟 (𝑥)
𝑃(𝐸 ∩ 𝐵𝑟 (𝑥)) =H 𝑛−1 (𝐸 ∩ 𝜕𝐵𝑟 (𝑥)) + 𝑃(𝐸; 𝐵𝑟 (𝑥)).
Proof. Step 1. Show that if 𝐸 and 𝐹 are open sets with 𝐶 1 -boundary,
with
𝑃(𝐸; 𝜕𝐹) = 𝑃(𝐹; 𝜕𝐸) = H 𝑛−1 (𝜕𝐸 ∩ 𝜕𝐹) = 0,
then 𝐸 ∩ 𝐹 (which is a set of locally finite perimeter by Exercise 5.24)
satisfies
𝜇 𝐸∩𝐹 = 𝜈 𝐸 H 𝑛−1 ⌞(𝐹 ∩ 𝜕𝐸) + 𝜈 𝐹 H 𝑛−1 ⌞(𝐸 ∩ 𝜕𝐹).
This is geometrically obvious and hence omitted.
Step 2. Approximating 𝐸 by smooth sets.
WLOG we assume 𝑥 = 0. We may assume that |𝐸 | < ∞ so by Theorem
7.2 there exists a sequence {𝐸 ℎ } ℎ∈N of open sets with smooth boundary such
that 𝐸 ℎ → 𝐸, |𝜇 𝐸 ℎ | → |𝜇 𝐸 |. In particular, by co-area formula
∫ ∞
|𝐸Δ𝐸 ℎ | = H 𝑛−1 ((𝐸Δ𝐸 ℎ ) ∩ 𝜕𝐵𝑟 ) d𝑟,
0
and hence (note that {𝜕𝐵𝑟 }𝑟∈R+ are disjoint Borel sets which foliate R𝑛 ) for
a.e. 𝑟 > 0,
lim H 𝑛−1 ((𝐸Δ𝐸 ℎ ) ∩ 𝜕𝐵𝑟 ) = 0, (8.5)
ℎ→∞
and
𝑃(𝐸; 𝜕𝐵𝑟 ) = 𝑃(𝐸 ℎ ; 𝜕𝐵𝑟 ) = 0, ∀ℎ ∈ N. (8.6)
For any 𝑟 satisfies the above properties, by Step 1 we have
𝜇 𝐸 ℎ ∩𝐵𝑟 = 𝜇 𝐸 ℎ ⌞𝐵𝑟 + 𝜇 𝐵𝑟 ⌞𝐸 ℎ , ∀ℎ ∈ N. (8.7)
36
 The fact that 𝜇 𝐵𝑟 ⌞𝐸 ℎ → 𝜇 𝐵𝑟 ⌞𝐸 as ℎ → ∞ is a direct consequence of
(8.5) and that 𝜇 𝐵𝑟 = 𝜈 𝐵𝑟 H 𝑛−1 ⌞𝜕𝐵𝑟 .
The fact that 𝜇 𝐸 ℎ ∩𝐵𝑟 → 𝜇 𝐸∩𝐵𝑟 as ℎ → ∞ follows from the convergence
𝐸 ℎ ∩ 𝐵𝑟 → 𝐸 ∩ 𝐵𝑟 and that, thanks to (8.7),
lim sup 𝑃(𝐸 ℎ ∩ 𝐵𝑟 ) ≤ lim sup 𝑃(𝐸 ℎ ; 𝐵𝑟 ) + 𝑃(𝐵𝑟 ) ≤ 𝑃(𝐸; 𝐵𝑟 ) + 𝑃(𝐵𝑟 ) < ∞.
ℎ→∞ ℎ→∞

Finally we show that 𝜇 𝐸 ℎ ⌞𝐵𝑟 → 𝜇 𝐸 ⌞𝐵𝑟 as ℎ → ∞. For any 𝑠 > 𝑟 let

𝜓 𝑠 ∈ 𝐶𝑐0 (𝐵 𝑠 ; [0, 1]) be such that 𝜓 𝑠 = 1 on 𝐵𝑟 , then by the convergence
|𝜇 𝐸 ℎ | → |𝜇 𝐸 |, for every 𝜑 ∈ 𝐶𝑐0 (R𝑛 ),
∫ ∫
𝜑d𝜇 𝐸 − 𝜑𝜓 𝑠 d𝜇 𝐸 ≤ sup |𝜑||𝜇 𝐸 |(𝐵 𝑠 \ 𝐵𝑟 ),
𝐵𝑟 R𝑛 R𝑛

and hence
∫ ∫
lim sup 𝜑d𝜇 𝐸 ℎ − 𝜑𝜓 𝑠 d𝜇 𝐸 ℎ ≤ sup |𝜑||𝜇 𝐸 |(𝐵 𝑠 \ 𝐵𝑟 ).
ℎ→∞ 𝐵𝑟 R𝑛 R𝑛
∫ ∫
Since R𝑛 𝜑𝜓 𝑠 d𝜇 𝐸 ℎ → R𝑛 𝜑𝜓 𝑠 d𝜇 𝐸 as ℎ → ∞, letting first ℎ → ∞ and then
𝑠 ↘ 𝑟, taking also (8.6) into account, we find as required that
∫ ∫
lim sup 𝜑d𝜇 𝐸 ℎ − 𝜑d𝜇 𝐸 ≤ 2 sup |𝜑||𝜇 𝐸 |(𝜕𝐵𝑟 ) = 0.
ℎ→∞ 𝐵𝑟 𝐵𝑟 R𝑛

By the above three convergences, we are allowed to let ℎ → ∞ in (8.7)

to obtain the first assertion. Thanks again to (8.6), the vector-valued Radon
measures on the Right of the first assertion are mutually singular, so that the
second and the third assertion follows immediately from the first assertion.
□
Exercise 8.9 (Intersection with half-spaces). If 𝐻𝑡 = {𝑥 ∈ R𝑛 : ⟨𝑥, 𝑒⟩ < 𝑡}
for some 𝑒 ∈ S𝑛−1 , 𝑡 ∈ R, and 𝐸 is a set of finite perimeter in R𝑛 with
|𝐸 | < ∞, then 𝐸 ∩ 𝐻𝑡 is a set of finite perimeter in R𝑛 , and for a.e. 𝑡 ∈ R,
𝜇 𝐸∩𝐻𝑡 = 𝜇 𝐸 ⌞𝐻𝑡 + 𝑒H 𝑛−1 ⌞(𝐸 ∩ 𝜕𝐻𝑡 ).
In particular, for a.e. 𝑡 ∈ R
H 𝑛−1 (𝐸 ∩ 𝜕𝐻𝑡 ) ≤ 𝑃(𝐸; 𝐻𝑡 ), 𝑃(𝐸 ∩ 𝐻𝑡 ) ≤ 𝑃(𝐸).
Proposition 8.10 (Characterization of half-spaces). If 𝐹 is a set of locally
finite perimeter in R𝑛 and 𝜈 ∈ S𝑛−1 is such that 𝜈 𝐹 (𝑦) = 𝜈 for |𝜇 𝐹 |-a.e.
𝑦 ∈ 𝜕 ∗ 𝐹, then there exists 𝛼 ∈ R such that 𝐹 is equivalent to the (open)
half-space
{𝑧 ∈ R𝑛 : ⟨𝑧, 𝜈⟩ < 𝛼}.
37
Proof. Let 𝑢 𝜖 = 𝜒𝐹 ∗ 𝜂𝜖 ∈ 𝐶 ∞ (R𝑛 ) for 𝜖 > 0. If 𝑇 ∈ 𝐶𝑐1 (R𝑛 , R𝑛 ) then by
Fubini’s theorem ∫ ∫
𝑢 𝜖 div𝑇 = div(𝑇 ∗ 𝜂𝜖 ).
R𝑛 𝐹
∫
Since R𝑛 div(𝑢 𝜖 𝑇) = 0, we have (recall that 𝜈 𝐹 (𝑦) = 𝑦 on 𝜕 ∗ 𝐹)
∫ ∫ ∫ ∫
− ⟨∇𝑢 𝜖 , 𝑇⟩ = 𝑢 𝜖 div𝑇 = div(𝑇 ∗ 𝜂𝜖 ) = ⟨(𝑇 ∗ 𝜂𝜖 ), d𝜇 𝐹 ⟩
R𝑛 R𝑛 𝐹 R𝑛
∫
= ⟨𝜈, (𝑇 ∗ 𝜂𝜖 )⟩ d|𝜇 𝐹 |.
𝜕∗ 𝐹

If we test this with 𝑇 = 𝜑𝜈′ for 𝜑 ∈ 𝐶𝑐1 (R𝑛 ) and 𝜈′ ∈ 𝑆 𝑛−1 ∩ 𝜈 ⊥ , then
∫
𝜕𝑢 𝜖
− ′
𝜑 = 0,
R𝑛 𝜕𝜈

which implies that 𝜕𝑢

𝜕𝜈 ′ = 0 on R .
𝜖 𝑛

If instead we use 𝑇 = 𝜑𝜈, then

∫ ∫
𝜕𝑢 𝜖
− 𝜑= 𝜑𝜖 d|𝜇 𝐹 |,
R𝑛 𝜕𝜈 𝜕∗ 𝐹
which is non-negative for any 𝜑 ≥ 0 (note that if 𝜑 ≥ 0 on R𝑛 then 𝜑𝜖 ≥ 0
on R𝑛 for every 𝜖 > 0), and hence
𝜕𝑢 𝜖
≤ 0 on R𝑛 .
𝜕𝜈
By these two facts, 𝑢 𝜖 (𝑦) is a decreasing function of the variable ⟨𝑦, 𝜈⟩,
that is, there exists 𝑓𝜖 ∈ 𝐶 ∞ (R; [0, 1]) with 𝑓𝜖′ ≤ 0 on R such that
( 𝜒𝐹 ∗ 𝜂𝜖 )(𝑦)𝑢 𝜖 (𝑦) = 𝑓𝜖 (⟨𝑦, 𝜈⟩), ∀𝑦 ∈ R𝑛 .
Letting 𝜖 ↘ 0 we then find lim𝜖→0 𝑓𝜖 (𝑡) ∈ {0, 1} for a.e. 𝑡 ∈ R; and hence
(recall that 𝑓𝜖′ ≤ 0) there exists 𝛼 ∈ R such that 𝑓𝜖 (𝑡) → 𝜒(−∞,𝛼) (𝑡) for a.e.
𝑡, i.e.,
( 𝜒𝐹 ∗ 𝜂𝜖 )(𝑦)𝑢 𝜖 (𝑦) → 𝜒(−∞,𝑦) (⟨𝑦, 𝜈⟩), for a.e. 𝑦 ∈ R𝑛 .
This proves the assertion. □
Theorem 8.11 (Tangential property of reduced boundary). If 𝐸 is a set of
locally finite perimeter in R𝑛 and 𝑥 ∈ 𝜕 ∗ 𝐸, then
loc
𝐸 𝑥,𝑟 → 𝐻𝑥 = {𝑦 ∈ R𝑛 : ⟨𝑦, 𝜈 𝐸 (𝑥)⟩ ≤ 0} as 𝑟 ↘ 0.
Similarly, if Π𝑥 = 𝜕𝐻𝑥 = 𝜈 𝐸 (𝑥) ⊥ , then as 𝑟 ↘ 0,
𝜇 𝐸 𝑥,𝑟 → 𝜈 𝐸 (𝑥)H 𝑛−1 ⌞Π𝑥 , |𝜇 𝐸 𝑥,𝑟 | → H 𝑛−1 ⌞Π𝑥 . (8.8)
38
Proof. Let 𝑥 ∈ 𝜕 ∗ 𝐸, by Lemma 8.8 we have
𝜇 𝐸∩𝐵𝑟 (𝑥) = 𝜈 𝐵𝑟 (𝑥) H 𝑛−1 ⌞(𝐸 ∩ 𝜕𝐵𝑟 (𝑥)) + 𝜇 𝐸 ⌞𝐵𝑟 (𝑥), (8.9)
𝑃(𝐸 ∩ 𝐵𝑟 (𝑥)) = H 𝑛−1 (𝐸 ∩ 𝜕𝐵𝑟 (𝑥)) + 𝑃(𝐸; 𝐵𝑟 (𝑥)) (8.10)
for a.e. 𝑟 > 0.
Step 1. We prove the existence of 𝑟 (𝑥) and 𝐶 (𝑛) positive such that
𝑃(𝐸 ∩ 𝐵𝑟 (𝑥)) ≤ 3H 𝑛−1 (𝐸 ∩ 𝜕𝐵𝑟 (𝑥)), for a.e. 𝑟 < 𝑟 (𝑥), (8.11)
𝑃(𝐸; 𝐵𝑟 (𝑥)) ≤ 𝐶 (𝑛)𝑟 𝑛−1 , ∀𝑟 < 𝑟 (𝑥). (8.12)
Indeed, by (8.9), if 𝜑 ∈ 𝐶𝑐1 (R𝑛 ) is such that 𝜑 ≡ 1 on 𝐵𝑟 (𝑥), then
∫ ∫
0= ∇𝜑 = 𝜑d𝜇 𝐸∩𝐵𝑟 (𝑥)
𝐸∩𝐵𝑟 (𝑥) R𝑛
∫ ∫
𝑛−1
= 𝜑𝜈 𝐵𝑟 (𝑥) dH + 𝜑d𝜇 𝐸
𝐸∩𝜕𝐵𝑟 (𝑥) 𝐵𝑟 (𝑥)
∫
= 𝜈 𝐵𝑟 (𝑥) dH 𝑛−1 + 𝜇 𝐸 (𝐵𝑟 (𝑥)),
𝐸∩𝜕𝐵𝑟 (𝑥)

this implies |𝜇 𝐸 (𝐵𝑟 (𝑥))| ≤ H 𝑛−1 (𝐸 ∩ 𝜕𝐵𝑟 (𝑥)) for a.e. 𝑟 > 0.
On the other hand, since 𝑥 ∈ 𝜕 ∗ 𝐸, by (8.1) there exists 𝑟 (𝑥) > 0 such that
𝑃(𝐸; 𝐵𝑟 (𝑥)) ≤ 2|𝜇 𝐸 (𝐵𝑟 (𝑥))|, ∀𝑟 < 𝑟 (𝑥).
Combining these two inequalities we find
𝑃(𝐸; 𝐵𝑟 (𝑥)) ≤ 2H 𝑛−1 (𝐸 ∩ 𝜕𝐵𝑟 (𝑥)), for a.e. 𝑟 < 𝑟 (𝑥).
By this and (8.10) we obtain (8.11), also it is trivial that H 𝑛−1 (𝐸 ∩𝜕𝐵𝑟 (𝑥)) ≤
𝑛𝜔𝑛 𝑟 𝑛−1 , and hence we obtain (8.12) for a.e. 𝑟 < 𝑟 (𝑥). It is then easy to see
that this is true for any 𝑟 < 𝑟 (𝑥) since 𝑃(𝐸; 𝐵𝑟 (𝑥)) is monotone on 𝑟.
Step 2. We prove two rough lower bounds on the 𝑛-dimensional density
ratios of 𝐸 and 𝐸 𝑐 at 𝑥; precisely, we show that
|𝐸 ∩ 𝐵𝑟 (𝑥)| 1
≥ , ∀𝑟 < 𝑟 (𝑥), (8.13)
𝑟 𝑛 (3𝑛) 𝑛
|𝐸 𝑐 ∩ 𝐵𝑟 (𝑥)| 1
≥ , ∀𝑟 < 𝑟 (𝑥). (8.14)
𝑟 𝑛 (3𝑛) 𝑛
Since 𝜇 𝐸 = −𝜇 𝐸 𝑐 , it suffices to show (8.13). To this end we define
𝑚 : (0, ∞) → [0, ∞) as
𝑚(𝑟) = |𝐸 ∩ 𝐵𝑟 (𝑥)|, 𝑟 > 0,
which is monotone increasing. By co-area formula
∫ 𝑟
𝑚(𝑟) = |𝐸 ∩ 𝐵𝑟 (𝑥)| = H 𝑛−1 (𝐸 ∩ 𝜕𝐵𝑡 (𝑥))d𝑡,
0
39
and hence
𝑚′ (𝑟) = H 𝑛−1 (𝐸 ∩ 𝜕𝐵𝑟 (𝑥)), for a.e. 𝑟 > 0,
with 𝑚(0) = 0. Moreover, 𝑚 > 0 because 𝑥 ∈ 𝜕 ∗ 𝐸 ⊂ spt𝜇 𝐸 , which implies
|𝜇 𝐸 |(𝐵𝑟 (𝑥)) > 0 and hence 𝑃(𝐸 ∩ 𝐵𝑟 (𝑥)) > 0 thanks to (8.10). This in turn
implies that 𝑚(𝑟) > 0, otherwise a contradiction. By (a rough) perimeter
bounds on volume and (8.11) we find
𝑛−1 𝑛−1
𝑚(𝑟) 𝑛 = |𝐸 ∩ 𝐵𝑟 (𝑥)| 𝑛 ≤ 𝑃(𝐸 ∩ 𝐵𝑟 (𝑥)) ≤ 3H 𝑛−1 (𝐸 ∩ 𝜕𝐵𝑟 (𝑥)) = 3𝑚′ (𝑟),
𝑛−1
for a.e. 𝑟 < 𝑟 (𝑥). Dividing both sides by 𝑚(𝑟) 𝑛 > 0, we obtain
1 1 1
≤ 𝑚′ (𝑟)𝑚(𝑟) −1+ 𝑛 = 𝑛(𝑚 𝑛 ) ′ (𝑟),
3
1 𝑟
which implies 𝑚(𝑟) 𝑛 ≥ 3𝑛 for all 𝑟 < 𝑟 (𝑥), this is exactly (8.13).
loc
Step 3. We prove that 𝐸 𝑥,𝑟 → 𝐻𝑥 as 𝑟 ↘ 0.
It suffices to show that, for any 𝑟 ℎ ↘ 0, up to extracting a subsequence
loc
𝐸 𝑥,𝑟 ℎ → 𝐻𝑥 as ℎ → ∞.
By (8.4), Exercise 5.11, and (8.12), for every 𝑅 > 0
𝑃(𝐸; 𝐵𝑟 𝑅 (𝑥) 𝑟 (𝑥)
𝑃(𝐸 𝑥,𝑟 ; 𝐵 𝑅 ) = 𝑛−1
≤ 𝐶 (𝑛)𝑅 𝑛−1 , ∀𝑟 < .
𝑟 𝑅
Given 𝑟 ℎ ↘ 0 as ℎ → ∞, by Corollary 6.5 up to extracting a subsequence,
there exists a set of locally finite perimeter 𝐹 in R𝑛 such that
loc
𝐸 𝑥,𝑟 ℎ → 𝐹, 𝜇 𝐸 𝑥,𝑟ℎ → 𝜇 𝐹 .
Moreover, up to extracting a further subsequence we can assume that
|𝜇 𝐸 𝑥,𝑟ℎ | → 𝜈 for some Radon measure 𝜈 on R𝑛 . By Lemma B.1 (𝑖), for
a.e. 𝑅 > 0 (indeed, for those 𝑅 > 0 such that 𝜈(𝜕𝐵 𝑅 ) = 0), we have
lim 𝜇 𝐸 𝑥,𝑟ℎ (𝐵 𝑅 ) = 𝜇 𝐹 (𝐵 𝑅 ). (8.15)
ℎ→∞
On the other hand, since 𝑥 ∈ 𝜕 ∗ 𝐸, by (8.4) we find
𝜇 𝐸 𝑥,𝑟 (𝐵 𝑅 ) 𝜇 𝐸 (𝐵𝑟 𝑅 (𝑥))
lim = lim = 𝜈 𝐸 (𝑥),
𝑟↘0 |𝜇 𝐸 𝑥,𝑟 |(𝐵 𝑅 ) 𝑟↘0 |𝜇 𝐸 |(𝐵𝑟 𝑅 (𝑥))

so that, since 𝑃(𝐸 𝑥,𝑟 ; 𝐵 𝑅 ) = |𝜇 𝐸 𝑥,𝑟 |(𝐵 𝑅 ), we find

𝑃(𝐸 𝑥,𝑟 ; 𝐵 𝑅 )
lim = 1. (8.16)
𝑟↘0 𝜈 𝐸 (𝑥), 𝜇 𝐸 𝑥,𝑟 (𝐵 𝑅 )
By lower semi-continuity, (8.16) and (8.15), we conclude for a.e. 𝑅 > 0
𝑃(𝐹; 𝐵 𝑅 ) ≤ lim inf 𝑃(𝐸 𝑥,𝑟 ℎ ; 𝐵 𝑅 ) = lim 𝜈 𝐸 (𝑥), 𝜇 𝐸 𝑥,𝑟 (𝐵 𝑅 )
ℎ→∞ ℎ→∞
= ⟨𝜈 𝐸 (𝑥), 𝜇 𝐹 (𝐵 𝑅 )⟩ ≤ |𝜇 𝐹 (𝐵 𝑅 )| ≤ |𝜇 𝐹 |(𝐵 𝑅 ) = 𝑃(𝐹; 𝐵 𝑅 ).
40
This means, for a.e. 𝑅 > 0,
|𝜇 𝐹 |(𝐵 𝑅 ) = lim |𝜇 𝐸 𝑥,𝑟ℎ |(𝐵 𝑅 ), (8.17)
ℎ→∞
and
|𝜇 𝐹 |(𝐵 𝑅 ) = ⟨𝜈 𝐸 (𝑥), 𝜇 𝐹 (𝐵 𝑅 )⟩ . (8.18)
By (8.17) and Lemma B.2, we deduce that
|𝜇 𝐸 𝑥,𝑟ℎ | → |𝜇 𝐹 |.
By (8.18), we find
∫
0= (1 − ⟨𝜈 𝐸 (𝑥), 𝜈 𝐹 (𝑦)⟩) d|𝜇 𝐹 |(𝑦),
𝐵𝑅
for a.e. 𝑅 > 0. The integrated is non-negative, and hence for |𝜇 𝐹 |-a.e.
𝑦 ∈ 𝜕 ∗ 𝐹,
𝜈 𝐹 (𝑦) = 𝜈 𝐸 (𝑥).
Then we apply Proposition 8.10 to deduce that there exists some 𝛼 ∈ R such
that
|𝐹Δ {𝑦 ∈ R𝑛 : ⟨𝑦, 𝜈 𝐸 (𝑥)⟩ < 𝛼}| = 0. (8.19)
If 𝛼 < 0, then 𝐹 ⊂ 𝐻𝑥 and |𝐹 ∩ 𝐵−𝛼 | = 0, so that
|𝐹 ∩ 𝐵−𝛼 | |𝐸 𝑥,𝑟 ℎ ∩ 𝐵−𝛼 | |𝐸 ∩ 𝐵−𝑟 ℎ 𝛼 (𝑥)|
0= = lim = lim ,
|𝐵−𝛼 | ℎ→∞ |𝐵−𝛼 | ℎ→∞ |𝐵−𝑟 ℎ 𝛼 (𝑥)|
a contradiction to (8.13). A similar argument shows that 𝛼 > 0 would
contradiction to (8.14). Thus 𝛼 = 0 and hence 𝐹 = 𝐻𝑥 by (8.19) as desired.
Concluding step.
loc
We have proven that, as 𝑟 ↘ 0, 𝐸 𝑥,𝑟 → 𝐻𝑥 , 𝜇 𝐸 𝑥,𝑟 → 𝜇 𝐻 𝑥 , and |𝜇 𝐸 𝑥,𝑟 | →
|𝜇 𝐻 𝑥 |, this completes the proof of the theorem since we apparently have by
virtue of the classical Gauss-Green theorem
𝜇 𝐻 𝑥 = 𝜈 𝐸 (𝑥)H 𝑛−1 ⌞Π𝑥 .
□
An immediate and important consequence is
Corollary 8.12. If 𝐸 is a set of locally finite perimeter in R𝑛 and 𝑥 ∈ 𝜕 ∗ 𝐸,
then
|𝐸 ∩ 𝐵𝑟 (𝑥)| 1
lim = , (8.20)
𝑟↘0 𝜔𝑛 𝑟 𝑛 2
and
𝑃(𝐸; 𝐵𝑟 (𝑥))
lim = 1. (8.21)
𝑟↘0 𝜔 𝑛−1 𝑟 𝑛−1

In particular, 𝜕 ∗ 𝐸 ⊂ 𝐸 (1/2) , the set of points of density one-half of 𝐸.

41
Proof. Let 𝑥 ∈ 𝜕 ∗ 𝐸, 𝐻𝑥 and Π𝑥 be as in Theorem 8.11. Since |𝐻𝑥 ∩ 𝐵| = 𝜔𝑛
2 ,
loc
the fact that 𝐸 𝑥,𝑟 → 𝐻𝑥 implies, as 𝑟 ↘ 0
|𝐸 ∩ 𝐵𝑟 (𝑥)| |𝐸 𝑥,𝑟 ∩ 𝐵1 | |𝐻𝑥 ∩ 𝐵1 | 1
= → = ,
𝜔𝑛 𝑟 𝑛 𝜔𝑛 𝜔𝑛 2
which is (8.20). Since Π𝑥 ∩ 𝜕𝐵1 is an (𝑛 − 1)-dim unit sphere, we have
H 𝑛−1 (Π𝑥 ∩𝜕𝐵1 ) = 0, and hence by (8.8) and (the Radon measure counterpart
of) Lemma B.1 (𝑖), as 𝑟 ↘ 0
|𝜇 𝐸 𝑥,𝑟 |(𝐵1 ) → H 𝑛−1 (Π𝑥 ∩ 𝐵1 ) = 𝜔𝑛−1 .
(8.21) then follows from the fact that |𝜇 𝐸 𝑥,𝑟 |(𝐵1 ) = 𝑟 1−𝑛 𝑃(𝐸; 𝐵𝑟 (𝑥)), thanks
to (8.4). □

8.2. De Giorgi’s structure theorem

The following famous theorem is known as De Giorgi’s structure theo-
rem, which says that reduced boundaries have the structure of generalized
hypersurfaces, thus leads to a generalized Gauss-Green theorem.
Theorem 8.13. If 𝐸 is a set of locally finite perimeter in R𝑛 , then the
Gauss-Green measure 𝜇 𝐸 of 𝐸 satisfies
𝜇 𝐸 = 𝜈 𝐸 H 𝑛−1 ⌞𝜕 ∗ 𝐸, |𝜇 𝐸 | = H 𝑛−1 ⌞𝜕 ∗ 𝐸,
and the generalized Gauss-Green formula holds:
∫ ∫
∇𝜑 = 𝜑𝜈 𝐸 dH 𝑛−1 , ∀𝜑 ∈ 𝐶𝑐1 (R𝑛 ).
𝐸 𝜕∗ 𝐸
Moreover, there exist countably many 𝐶 1 -hypersurfaces 𝑀ℎ in R𝑛 , compact
sets 𝐾 ℎ ⊂ 𝑀ℎ , and a Borel set 𝐹 with H 𝑛−1 (𝐹) = 0, such that
Ø
𝜕∗ 𝐸 = 𝐹 ∪ 𝐾ℎ ,
ℎ∈N
and, for every 𝑥 ∈ 𝐾 ℎ , 𝜈 𝐸 (𝑥) ⊥ = 𝑇𝑥 𝑀ℎ , the tangent space to 𝑀ℎ at 𝑥.
Remark 8.14. 𝜕∗ 𝐸 is the so-called countably (𝑛 − 1)-rectifiable set in R𝑛 .
We present two proofs of this theorem, the first one is based on the
following Rectifiability theorem, see [2, Theorem 10.8]:
Theorem 8.15. If 𝜇 is a Radon measure on R𝑛 , 𝑀 is a Borel set in R𝑛 , 𝜇
is concentrated on 𝑀, and for every 𝑥 ∈ 𝑀, there exists a 𝑘-dimensional
plane Π𝑥 in R𝑛 such that the 𝑘-blow ups of 𝜇
(Φ𝑥,𝑟 )# 𝜇
→ H 𝑘 ⌞Π𝑥 , 𝑟 ↘ 0,
𝑟𝑘
then 𝜇 = H 𝑘 ⌞𝑀 and 𝑀 is locally H 𝑘 -rectifiable.
42
Proof of Theorem 8.13 (I). By Theorem 8.11 every the (𝑛 − 1)-blow up of
|𝜇 𝐸 | at every 𝑥 ∈ 𝜕 ∗ 𝐸 takes the form H 𝑛−1 ⌞Π𝑥 , where Π𝑥 = 𝜈 𝐸 (𝑥) ⊥ .
Since |𝜇 𝐸 | is concentrated on 𝜕 ∗ 𝐸, using Theorem 8.15 we find
|𝜇 𝐸 | = H 𝑛−1 ⌞𝜕 ∗ 𝐸.
Taking (8.2) into account, the assertion follows. □
Another proof is based on the following fundamental extension theorem:

Theorem 8.16 (Whitney’s extension theorem). If 𝐶 ⊂ R𝑛 is closed, 𝑢 : 𝐶 →

R, 𝑇 : 𝐶 → R𝑛 are continuous functions, then there exists 𝑣 ∈ 𝐶 1 (R𝑛 ) such
that 𝑢 = 𝑣 and 𝑇 = ∇𝑣 on 𝐶 if and only if, for every compact set 𝐾 ⊂ 𝐶,
 
|𝑢(𝑦) − 𝑢(𝑥) − ⟨𝑇 (𝑥), (𝑦 − 𝑥)⟩|
lim sup : 0 < |𝑥 − 𝑦| < 𝛿, 𝑥, 𝑦 ∈ 𝐾 = 0.
𝛿↘0 |𝑥 − 𝑦|

and its corollary:

Corollary 8.17. If 𝐾 ⊂ R𝑛 is compact, 𝑇 : 𝐾 → R𝑛 is continuous, and
 
|⟨𝑇 (𝑥), (𝑦 − 𝑥)⟩|
lim sup : 0 < |𝑥 − 𝑦| ≤ 𝛿, 𝑥, 𝑦 ∈ 𝐾 = 0, (8.22)
𝛿↘0 |𝑥 − 𝑦|
then there exists 𝑣 ∈ 𝐶 1 (R𝑛 ) such that 𝐾 ⊂ {𝑥 ∈ R𝑛 : 𝑣(𝑥) = 0} and 𝑇 = ∇𝑣
on 𝐾. In particular, if 𝑇 ≠ 0 on 𝐾, then 𝐾 is contained in a 𝐶 1 -hypersurface.
Proof of Theorem 8.13 (II). Fix 𝑥 ∈ 𝜕 ∗ 𝐸 and write
𝐻𝑥− = {𝑦 ∈ R𝑛 : ⟨𝑦 − 𝑥, 𝜈 𝐸 (𝑥)⟩ ≤ 0}, 𝐻𝑥+ = {𝑦 ∈ R𝑛 : ⟨𝑦 − 𝑥, 𝜈 𝐸 (𝑥)⟩ ≥ 0}.

By Theorem 8.11 and Corollary 8.12 we have

|𝐵𝑟 (𝑥) ∩ 𝐻𝑥+ ∩ 𝐸 |
lim = 0, (8.23)
𝑟↘0 𝜔𝑛 𝑟 𝑛
and
|𝐵𝑟 (𝑥) ∩ 𝐻𝑥− ∩ 𝐸 | 1
lim = . (8.24)
𝑟↘0 𝜔𝑛 𝑟 𝑛 2
To prove the theorem we may replace 𝐸 by 𝐸 ∩ 𝐵 𝑅 for any 𝑅 > 0. By
Lemma 8.8 𝐸 ∩ 𝐵 𝑅 is a set of finite perimeter.
Step 1. We prove that there exists a |𝜇 𝐸 |-negligible set 𝐹 such that
Ø
𝜕∗ 𝐸 = 𝐹 ∪ 𝐾ℎ , (8.25)
𝑘∈N

with 𝐾 ℎ compact, 𝐾 ℎ ⊂ 𝑀ℎ , where 𝑀ℎ is a 𝐶 1 -hypersurface 𝑀ℎ and

𝜈 𝐸 (𝑥) ⊥ = 𝑇𝑥 𝑀ℎ for every 𝑥 ∈ 𝐾 ℎ .
43
 For a sequence {𝑟 𝑗 } 𝑗 ∈N such that 𝑟 𝑗 ↘ 0 as 𝑗 → ∞, we write
|𝐵𝑟 𝑗 (𝑥) ∩ 𝐻𝑥+ ∩ 𝐸 |
𝑓 𝑗 (𝑥) = ,
𝜔𝑛 𝑟 𝑗 𝑛
and
|𝐵𝑟 𝑗 (𝑥) ∩ 𝐻𝑥− ∩ 𝐸 |
𝑔 𝑗 (𝑥) = .
𝜔𝑛 𝑟 𝑛𝑗
By (8.23) and (8.24), 𝑓 𝑗 → 0, 𝑔 𝑗 → 12 for |𝜇 𝐸 |-a.e. 𝑥 ∈ 𝜕 ∗ 𝐸 (in fact, for all
𝑥 ∈ 𝜕 ∗ 𝐸).
Since |𝜇 𝐸 |(𝜕 ∗ 𝐸) = 𝑃(𝐸) < ∞, by Egoroff’s theorem, there exists a
compact set 𝐹1 ⊂ 𝜕 ∗ 𝐸 such that 𝑓 𝑗 → 0, 𝑔 𝑗 → 12 uniformly on 𝐹1 and
1
|𝜇 𝐸 |(𝜕 ∗ 𝐸 \ 𝐹1 ) < |𝜇 𝐸 |(𝜕 ∗ 𝐸) = 𝑃(𝐸).
2
For 𝜕 ∗ 𝐸 \ 𝐹1 we can again use Egoroff’s theorem to find a compact set
𝐹2 ⊂ 𝜕 ∗ 𝐸 \ 𝐹1 such that 𝑓 𝑗 → 0, 𝑔 𝑗 → 21 uniformly on 𝐹2 and
1 1
|𝜇 𝐸 |(𝜕 ∗ 𝐸 \ (𝐹1 ∪ 𝐹2 )) ≤ |𝜇 𝐸 |(𝜕 ∗ 𝐸 \ 𝐹1 ) ≤ ( ) 2 𝑃(𝐸).
2 2
An inductive argument then shows that there exists a seuqnce of compact
sets {𝐹𝑖 }𝑖∈N , with |𝜇 𝐸 |(𝜕 𝐸 \ 𝑖∈N 𝐹𝑖 ) = 0, such that 𝑓 𝑗 → 0, 𝑔 𝑗 → 12
∗ Ð
uniformly on each 𝐹𝑖 .
On the other hand, since 𝜈 𝐸 is a Borel function on 𝜕 ∗ , by Lusin’s theorem
(and a similar induction argument) for every 𝑖 ∈ N there exists a sequence
of compact sets {𝐶𝑖,𝑘 } 𝑘∈N such that 𝜈 𝐸 is continuous on 𝐶𝑖,𝑘 and
Ø
|𝜇 𝐸 |(𝐹𝑖 \ 𝐶𝑖,𝑘 ) = 0.
𝑘∈N
Concluding this step, we have proven that, there exists a sequence of compact
sets {𝐾 ℎ } ℎ∈N with 𝐾 ℎ ⊂ 𝜕 ∗ 𝐸, such that
(𝑖) If 𝐹 = 𝜕 ∗ 𝐸 \ ℎ∈N 𝐾 ℎ , then |𝜇 𝐸 |(𝐹) = 0.
Ð
(𝑖𝑖) For every ℎ ∈ N, 𝜈 𝐸 is continuous on 𝐾 ℎ .
(𝑖𝑖𝑖) For every ℎ ∈ N, there exists an increasing function 𝜎ℎ : [0, ∞) →
[0, ∞) with 𝜎ℎ (0+ ) = 0 such that, for every 𝑥 ∈ 𝐾 ℎ ,
|𝐵𝑟 (𝑥) ∩ 𝐻𝑥+ ∩ 𝐸 | |𝐵𝑟 (𝑥) ∩ 𝐻𝑥− ∩ 𝐸 | 1
≤ 𝜎ℎ (𝑟), − ≤ 𝜎ℎ (𝑟).
𝜔𝑛 𝑟 𝑛 𝜔𝑛 𝑟 𝑛 2
Each pair (𝐾 ℎ , 𝜈 𝐸 ) satisfies the criterion (8.22): more precisely, for every
𝜖 ∈ (0, 1) there exists 𝛿 > 0 (depending on 𝜖, 𝑛, ℎ) such that
(
|𝑥 − 𝑦| ≤ 𝛿,
⇒ |⟨𝜈 𝐸 (𝑥), 𝑦 − 𝑥⟩| ≤ 𝜖 |𝑥 − 𝑦|. (8.26)
𝑥, 𝑦 ∈ 𝐾 ℎ
44
To see this, we argue by contradiction: if 𝑥, 𝑦 ∈ 𝐾 ℎ and ⟨𝜈 𝐸 (𝑥), 𝑦 − 𝑥⟩ >
𝜖 |𝑥 − 𝑦|, then we have
𝐵𝜖 |𝑥−𝑦| (𝑦) ⊂ 𝐵2|𝑥−𝑦| (𝑥) ∩ 𝐻𝑥+ . (8.27)
By property (𝑖𝑖𝑖) and (8.27), we find
𝜔 
− 𝜔𝑛 𝜎ℎ (𝜖 |𝑥 − 𝑦|) 𝜖 𝑛 |𝑥 − 𝑦| 𝑛 ≤ 𝐵𝜖 |𝑥−𝑦| (𝑦) ∩ 𝐻 𝑦− ∩ 𝐸
𝑛
2
≤ 𝐵𝜖 |𝑥−𝑦| (𝑦) ∩ 𝐸 ≤ 𝐵2|𝑥−𝑦| (𝑥) ∩ 𝐻𝑥+ ∩ 𝐸
≤𝜎ℎ (2|𝑥 − 𝑦|)2𝑛 |𝑥 − 𝑦| 𝑛 .
If |𝑥 − 𝑦| ≤ 𝛿 then we obtain
(𝜔𝑛 − 2𝜔𝑛 𝜎ℎ (𝜖 |𝑥 − 𝑦|)) ≤ 2𝑛+1 𝜎ℎ (2𝛿),
which leads to a contradiction as soon as 𝛿 ↘ 0 with 𝜖 and ℎ fixed. Thus
for every 𝜖 ∈ (0, 1) there exists 𝛿 > 0 such that
|𝑥 − 𝑦| ≤ 𝛿, 𝑥, 𝑦 ∈ 𝐾 ℎ ,
implies
⟨𝜈 𝐸 (𝑥), 𝑦 − 𝑥⟩ < 𝜖 |𝑥 − 𝑦|.
The complementary estimate is derived similarly, and (8.26) is thus proved.
Hence, by the corollary to Whitney’s extension theorem, for each ℎ ∈ N
there exists a 𝐶 1 -hypersurface 𝑀ℎ such that 𝐾 ℎ ⊂ 𝑀ℎ and 𝜈 𝐸 (𝑥) ⊥ = 𝑇𝑥 𝑀ℎ
for each 𝑥 ∈ 𝐾 ℎ .
Step 2. We prove that for every Borel set 𝐺 ⊂ R𝑛 , there holds
H 𝑛−1 (𝐺 ∩ 𝜕 ∗ 𝐸) ≤ 𝐶 (𝑛)|𝜇 𝐸 |(𝐺), (8.28)
that is, H 𝑛−1 ⌞𝜕 ∗ 𝐸 ≤ 𝐶 (𝑛)|𝜇 𝐸 | on B (R𝑛 ).
Let 𝐴 be an open set with 𝐺 ∩ 𝜕 ∗ 𝐸 ⊂ 𝐴, and consider the family F of
those balls 𝐵𝑟 (𝑥) ⊂ 𝐴 such that
𝜔𝑛−1𝑟 𝑛−1
𝑥 ∈ 𝐺 ∩ 𝜕 ∗ 𝐸, 0 < 𝑟 < 𝛿, |𝜇 𝐸 |(𝐵𝑟 (𝑥)) ≥ .
2
By (8.21), F is a covering of 𝐺 ∩ 𝜕 ∗ 𝐸. By Besicovitch’s covering theorem
𝜉 (𝑛)
(Theorem A.1), there exist finitely many subfamilies {Fℎ } ℎ=1 of F , with
each Fℎ disjoint and at most countable, such that 𝐺 ∩ 𝜕 ∗ 𝐸 is covered by
Ð𝜉 (𝑛)
𝑖=1 Fℎ . Thus, by the construction of F we find
(𝑛)
𝜉∑︁ ∑︁ (𝑛)
𝜉∑︁ ∑︁
∗
H𝛿𝑛−1 (𝐺 ∩ 𝜕 𝐸) ≤ 𝜔𝑛−1𝑟 𝑛−1
≤2 |𝜇 𝐸 |(𝐵𝑟 (𝑥))
ℎ=1 𝐵𝑟 (𝑥)∈Fℎ ℎ=1 𝐵𝑟 (𝑥)∈Fℎ
≤2𝜉 (𝑛)|𝜇 𝐸 |( 𝐴).
Letting first 𝛿 ↘ 0, and then use the arbitrariness of 𝐴 and the fact that |𝜇 𝐸 |
is concentrated on 𝜕 ∗ 𝐸 to infer (8.28) with 𝐶 (𝑛) = 2𝜉 (𝑛).
45
Step 3. We prove that H 𝑛−1 ⌞𝜕 ∗ 𝐸 = |𝜇 𝐸 | on P (R𝑛 ).
By (8.28), H 𝑛−1 ⌞𝜕 ∗ 𝐸 is locally finite, and hence is a Radon measure.
Thus it suffices to prove that H 𝑛−1 ⌞𝜕 ∗ 𝐸 = |𝜇 𝐸 | on B (R𝑛 ). Since |𝜇 𝐸 |(𝐹) =
0, by (8.25) we just have to prove that
H 𝑛−1 ⌞𝐾 ℎ = |𝜇 𝐸 |⌞𝐾 ℎ on B (R𝑛 ), ∀ℎ ∈ N. (8.29)
By Lebesgue-Besicovitch differentiation theorem, there exists Borel sets
𝑌ℎ ⊂ 𝑀ℎ with |𝜇 𝐸 |(𝑌ℎ ) = 0 such that
H 𝑛−1 ⌞𝑀ℎ = 𝜃|𝜇 𝐸 | + H 𝑛−1 ⌞𝑌ℎ , (8.30)

where 𝜃 = 𝐷 |𝜇 𝐸 | (H 𝑛−1 ⌞𝑀ℎ ) : R𝑛 → R satisfies

H 𝑛−1 (𝑀ℎ ∩ 𝐵𝑟 (𝑥))

𝜃 (𝑥) = lim , (8.31)
𝑟↘0 |𝜇 𝐸 |(𝐵𝑟 (𝑥))
for |𝜇 𝐸 |-a.e. 𝑥 ∈ 𝜕 ∗ 𝐸. By area formula it is easy to see that if 𝑥 ∈ 𝑀ℎ then
H 𝑛−1 (𝑀ℎ ∩ 𝐵𝑟 (𝑥))
lim = 1.
𝑟↘0 𝜔𝑛−1𝑟 𝑛−1
Since 𝐾 ℎ ⊂ 𝑀ℎ ∩ 𝜕 ∗ 𝐸, we deduce from (8.31) and (8.21) that 𝜃 = 1 on 𝐾 ℎ .
In particular, if 𝐺 is a Borel set with 𝐺 ⊂ 𝐾 ℎ , then (8.30) implies (8.29),
which can be seen as follows:
H 𝑛−1 (𝐺) = H 𝑛−1 (𝐺 ∩ 𝑀ℎ ) = |𝜇 𝐸 |(𝐺) + H 𝑛−1 (𝑌ℎ ∩ 𝐺) = |𝜇 𝐸 |(𝐺),
where we have used
H 𝑛−1 (𝑌ℎ ∩ 𝐺) ≤ H 𝑛−1 (𝑌ℎ ∩ 𝜕 ∗ 𝐸) ≤ 𝐶 (𝑛)|𝜇 𝐸 |(𝑌ℎ ) = 0,
thanks to (8.28).
The proof is thus completed. □
Exercise 8.18. Show that if 𝐸 is a set of locally finite perimeter and 𝑥 ∈ 𝜕 ∗ 𝐸,
then ∫
1
𝜈 𝐸 (𝑥) = lim 𝜈 𝐸 dH 𝑛−1 .
𝑟↘0 𝜔 𝑛−1 𝑟 𝑛−1 𝐵𝑟 (𝑥)∩𝜕 ∗ 𝐸

Exercise 8.19 (Characterization of balls). Show that a set of locally finite

perimeter 𝐸 in R𝑛 is equivalent to 𝐵𝑟 (𝑥) for some fixed 𝑥 ∈ R𝑛 , 𝑟 > 0, if
and only if
𝑦−𝑥
𝜈 𝐸 (𝑦) = ,
|𝑦 − 𝑥|
for H 𝑛−1 -a.e. 𝑦 ∈ 𝜕 ∗ 𝐸.
46
 With De Giorgi’s structure theorem, we can go back to state the tangential
property of reduced boundary (Theorem 8.11) in a more explicit form as
follows (using (8.4)):
𝜕∗ 𝐸 − 𝑥
𝜈 𝐸 H 𝑛−1 ⌞( ) → 𝜈 𝐸 (𝑥)H 𝑛−1 ⌞Π𝑥 ,
𝑟
and
𝑛−1 𝜕∗ 𝐸 − 𝑥
H ⌞( ) → H 𝑛−1 ⌞Π𝑥 ,
𝑟
as 𝑟 ↘ 0.

8.3. Essential boundary and Federer’s theorem

Definition 8.20 (Essential boundary). Given 𝐸 ∈ M (L 𝑛 ), the essential
boundary of 𝐸, denoted by 𝜕 𝑒 𝐸, is defined as
 
𝜕 𝑒 𝐸 = R𝑛 \ 𝐸 (0) ∪ 𝐸 (1) .
1
It is clear that 𝐸 ( 2 ) ⊂ 𝜕 𝑒 𝐸, also recall that we have shown in Corollary 8.12
1 1
that the reduced boundary 𝜕 ∗ 𝐸 ⊂ 𝐸 ( 2 ) . Hence 𝜕 ∗ 𝐸 ⊂ 𝐸 ( 2 ) ⊂ 𝜕 𝑒 𝐸. The
following Federer’s theorem shows that the reduced boundary and essential
boundary are equivalent in terms of H 𝑛−1 .
Theorem 8.21 (Federer’s theorem). If 𝐸 is a set of locally finite perimeter
1
in R𝑛 , then 𝜕 ∗ 𝐸 ⊂ 𝐸 ( 2 ) , with
H 𝑛−1 (𝜕 𝑒 𝐸 \ 𝜕 ∗ 𝐸) = 0. (8.32)
To prove this theorem we need the following density comparison theorem
for Radon measures and its corollary.
Definition 8.22 (upper/lower 𝑘-dimensional density). Given 𝜇 a Radon
measure on R𝑛 and 𝑘 ∈ (0, 𝑛], we define the upper/lower 𝑘-dimensional
density of 𝜇 as 𝜃 ∗𝑘 (𝜇)(𝑥), 𝜃 ∗𝑘 (𝜇)(𝑥), given by
𝜇(𝐵𝑟 (𝑥))
𝜃 ∗𝑘 (𝜇)(𝑥) = lim sup , 𝑥 ∈ R𝑛 ,
𝑟↘0 𝜔𝑘 𝑟 𝑘
and
𝜇(𝐵𝑟 (𝑥))
𝜃 ∗𝑘 (𝜇)(𝑥) = lim inf 𝑘
, 𝑥 ∈ R𝑛 .
𝑟↘0 𝜔𝑘 𝑟
If 𝜃 ∗𝑘 (𝜇)(𝑥) and 𝜃 ∗𝑘 (𝜇)(𝑥) agree then it is called the 𝑘-dimensional density
of 𝜇 at 𝑥, denoted by 𝜃 𝑘 (𝜇)(𝑥). Note that in this case we have
𝜇(𝐵𝑟 (𝑥))
𝜃 𝑘 (𝜇)(𝑥) = lim .
𝑟↘0 𝜔𝑘 𝑟 𝑘
47
Theorem 8.23 (Upper 𝑘-dimensional density and comparison with H 𝑘 ). If
𝜇 is a Radon measure on R𝑛 , 𝑀 is a Borel set, and 𝑘 ∈ (0, 𝑛), then
1 ≤ 𝜃 ∗𝑘 (𝜇) on 𝑀 ⇒ H 𝑘 (𝑀) ≤ 𝜇(𝑀), (8.33)
and
𝜃 ∗𝑘 (𝜇) ≤ 1 on 𝑀 ⇒ 𝜇(𝑀) ≤ 2 𝑘 H 𝑘 (𝑀).
For the proof, see e.g., [2, Theorem 6.4].
Corollary 8.24 (Density at points not in 𝑀). If 𝑘 ∈ (0, 𝑛) and 𝑀 ⊂ R𝑛 is
a Borel set with H 𝑘 (𝑀 ∩ 𝐾) < ∞ for every compact set 𝐾 ⊂ R𝑛 , then for
H 𝑘 -a.e. 𝑥 ∈ R𝑛 \ 𝑀,
H 𝑘 (𝑀 ∩ 𝐵𝑟 (𝑥))
lim = 0.
𝑟↘0 𝜔𝑘 𝑟 𝑘
Proof. As this is local property we may replace 𝑀 by 𝑀 ∩ 𝐾 for arbitrary
compact 𝐾, therefore we may direct assume that H 𝑘 (𝑀) < ∞. For any
𝜖 > 0, set
𝐹𝜖 = {𝑥 ∈ R𝑛 \ 𝑀 : 𝜃 ∗𝑘 (𝜇)(𝑥) ≥ 𝜖 },
where 𝜇 = H 𝑠 ⌞𝑀. By (8.33)
𝜖H 𝑘 (𝐹𝜖 ) ≤ 𝜇(𝐹𝜖 ) = H 𝑘 (𝑀 ∩ 𝐹𝜖 ) = 0.
Since this is true for arbitrary 𝜖 > 0, the assertion follows. □
See also [1] for another proof.
Proof of Theorem 8.21. By the relative isoperimetric inequality and the triv-
ial fact that
1 𝑛−1
|𝐸 ∩ 𝐵𝑟 (𝑥)| ≤ 𝜔𝑛𝑛 𝑟 |𝐸 ∩ 𝐵𝑟 (𝑥)| 𝑛 ,
we find
𝑃(𝐸; 𝐵𝑟 (𝑥)) |𝐵𝑟 (𝑥) ∩ 𝐸 | |𝐵𝑟 (𝑥) \ 𝐸 |
≥ 𝑐(𝑛) min{ , }.
𝑟 𝑛−1 𝑟𝑛 𝑟𝑛
∗ (H 𝑛−1 ⌞𝜕 ∗ 𝐸)(𝑥) = 0 implies 𝑥 ∈ 𝐸 (0) ∪ 𝐸 (1) . In particular,
Thus, 𝜃 𝑛−1
∗
𝜕 𝑒 𝐸 ⊂ {𝑥 ∈ R𝑛 : 𝜃 𝑛−1 (H 𝑛−1 ⌞𝜕 ∗ 𝐸)(𝑥) > 0},
and hence
𝜕 𝑒 𝐸 \ 𝜕 ∗ 𝐸 ⊂ {𝑥 ∈ R𝑛 \ 𝜕 ∗ 𝐸 : 𝜃 𝑛−1
∗
(H 𝑛−1 ⌞𝜕 ∗ 𝐸)(𝑥) > 0}.
By Corollary 8.24, this last set is H 𝑛−1 -negligible, the assertion follows.
□
48
 In the following we write 𝐴 ≈ 𝐵 if 𝐴, 𝐵 are Borel sets such that
H 𝑛−1 ( 𝐴Δ𝐵) = 0. By (8.32), 𝜕 ∗ 𝐸 ≈ 𝐸 (1/2) ≈ 𝜕 𝑒 𝐸, and
𝑀 ≈ (𝑀 ∩ 𝐸 (1) ) ∪ (𝑀 ∩ 𝐸 (0) ) ∪ (𝑀 ∩ 𝐸 (1/2) ), (8.34)
for every Borel set 𝑀 and for every set of locally finite perimeter 𝐸.
Theorem 8.25 (Set operations on Gauss-Green measure). If 𝐸, 𝐹 are set of
locally finite perimeter, and we let
{𝜈 𝐸 = 𝜈 𝐹 } = {𝑥 ∈ 𝜕 ∗ 𝐸 ∩ 𝜕 ∗ 𝐹 : 𝜈 𝐸 (𝑥) = 𝜈 𝐹 (𝑥)},
{𝜈 𝐸 = −𝜈 𝐹 } = {𝑥 ∈ 𝜕 ∗ 𝐸 ∩ 𝜕 ∗ 𝐹 : 𝜈 𝐸 (𝑥) = −𝜈 𝐹 (𝑥)},
then 𝐸 ∩ 𝐹, 𝐸 \ 𝐹, and 𝐸 ∪ 𝐹 are sets of locally finite perimeter with
𝜇 𝐸∩𝐹 =𝜇 𝐸 ⌞𝐹 (1) + 𝜇 𝐹 ⌞𝐸 (1) + 𝜈 𝐸 H 𝑛−1 ⌞{𝜈 𝐸 = 𝜈 𝐹 }, (8.35)
𝜇 𝐸\𝐹 =𝜇 𝐸 ⌞𝐹 (0) − 𝜇 𝐹 ⌞𝐸 (1) + 𝜈 𝐸 H 𝑛−1 ⌞{𝜈 𝐸 = −𝜈 𝐹 }, (8.36)
𝜇 𝐸∪𝐹 =𝜇 𝐸 ⌞𝐹 (0) + 𝜇 𝐹 ⌞𝐸 (0) + 𝜈 𝐸 H 𝑛−1 ⌞{𝜈 𝐸 = 𝜈 𝐹 }, (8.37)
and
𝜕 ∗ (𝐸 ∩ 𝐹) ≈ (𝐹 (1) ∩ 𝜕 ∗ 𝐸) ∪ (𝐸 (1) ∩ 𝜕 ∗ 𝐹) ∪ {𝜈 𝐸 = 𝜈 𝐹 }, (8.38)
𝜕 ∗ (𝐸 \ 𝐹) ≈ (𝐹 (0) ∩ 𝜕 ∗ 𝐸) ∪ (𝐸 (1) ∩ 𝜕 ∗ 𝐹) ∪ {𝜈 𝐸 = −𝜈 𝐹 }, (8.39)
𝜕 ∗ (𝐸 ∪ 𝐹) ≈ (𝐹 (0) ∩ 𝜕 ∗ 𝐸) ∪ (𝐸 (0) ∩ 𝜕 ∗ 𝐹) ∪ {𝜈 𝐸 = 𝜈 𝐹 }. (8.40)
Moreover, for every Borel set 𝐺 ∈ B (R𝑛 ),
𝑃(𝐸 ∩ 𝐹; 𝐺) = 𝑃(𝐸; 𝐹 (1) ∩ 𝐺) + 𝑃(𝐹; 𝐸 (1) ∩ 𝐺) + H 𝑛−1 ({𝜈 𝐸 = 𝜈 𝐹 } ∩ 𝐺),
(8.41)
(0) (1) 𝑛−1
𝑃(𝐸 \ 𝐹; 𝐺) = 𝑃(𝐸; 𝐹 ∩ 𝐺) + 𝑃(𝐹; 𝐸 ∩ 𝐺) + H ({𝜈 𝐸 = −𝜈 𝐹 } ∩ 𝐺),
(8.42)
𝑃(𝐸 ∪ 𝐹; 𝐺) = 𝑃(𝐸; 𝐹 (0) ∩ 𝐺) + 𝑃(𝐹; 𝐸 (0) ∩ 𝐺) + H 𝑛−1 ({𝜈 𝐸 = 𝜈 𝐹 } ∩ 𝐺).

Proof. By set operations,

|(𝐸 ∪ 𝐹) ∩ 𝐵𝑟 (𝑥)| =|(𝐸 \ 𝐹) ∩ 𝐵𝑟 (𝑥)| + |(𝐹 \ 𝐸) ∩ 𝐵𝑟 (𝑥)| + |(𝐸 ∩ 𝐹) ∩ 𝐵𝑟 (𝑥)|
=|𝐸 ∩ 𝐵𝑟 (𝑥)| + |𝐹 ∩ 𝐵𝑟 (𝑥)| − |(𝐸 ∩ 𝐹) ∩ 𝐵𝑟 (𝑥)|.
Thus at those points such that the densities involved exist, we find
max{𝜃 𝑛 (𝐸), 𝜃 𝑛 (𝐹)} ≤ 𝜃 𝑛 (𝐸 ∪ 𝐹) = 𝜃 𝑛 (𝐸) + 𝜃 𝑛 (𝐹) − 𝜃 𝑛 (𝐸 ∩ 𝐹). (8.43)
Then we claim that
𝜕 (𝐸 ∩ 𝐹) ≈ (𝐹 (1) ∩ 𝜕 ∗ 𝐸) ∪ (𝐸 (1) ∩ 𝜕 ∗ 𝐹) ∪ (𝜕 ∗ (𝐸 ∩ 𝐹) ∩ 𝜕 ∗ 𝐸 ∩ 𝜕 ∗ 𝐹),
∗

(8.44)
with the three sets on the right having H 𝑛−1 -negligible mutual intersec-
tions.
49
 Write 𝑀 = (𝐸 ∩ 𝐹) (1/2) and decompose 𝑀 with respect to 𝐸 by (8.34).
Note that (𝐸 ∩ 𝐹) (1/2) ∩ 𝐸 (0) = ∅ thanks to 𝜃 𝑛 (𝐸 ∩ 𝐹) ≤ 𝜃 𝑛 (𝐸), we find
   
(𝐸 ∩ 𝐹) (1/2) ≈ (𝐸 ∩ 𝐹) (1/2) ∩ 𝐸 (1) ∪ (𝐸 ∩ 𝐹) (1/2) ∩ 𝐸 (1/2) .

Similarly, decomposing 𝑀1 = (𝐸 ∩ 𝐹) (1/2) ∩ 𝐸 (1) and 𝑀2 = (𝐸 ∩ 𝐹) (1/2) ∩

𝐸 (1/2) with respect to 𝐹 by (8.34), and the fact that (𝐸 ∩ 𝐹) (1/2) ∩ 𝐹 (0) = ∅
thanks to 𝜃 𝑛 (𝐸 ∩ 𝐹) ≤ 𝜃 𝑛 (𝐹), we find
   
(1/2) (1/2) (1) (1) (1/2) (1) (1/2)
(𝐸 ∩ 𝐹) ≈ (𝐸 ∩ 𝐹) ∩𝐸 ∩𝐹 ∪ (𝐸 ∩ 𝐹) ∩𝐸 ∩𝐹
   
∪ (𝐸 ∩ 𝐹) (1/2) ∩ 𝐸 (1/2) ∩ 𝐹 (1) ∪ (𝐸 ∩ 𝐹) (1/2) ∩ 𝐸 (1/2) ∩ 𝐹 (1/2) .
By (8.43) we have (𝐸 ∩ 𝐹) (1/2) ∩ 𝐸 (1) ∩ 𝐹 (1) = ∅, and
(𝐸 (1) ∩ 𝐹 (1/2) ) ⊂ (𝐸 ∩ 𝐹) (1/2) , (𝐹 (1) ∩ 𝐸 (1/2) ) ⊂ (𝐸 ∩ 𝐹) (1/2) . (8.45)
Hence
     
(𝐸 ∩ 𝐹) (1/2) ≈ 𝐸 (1) ∩ 𝐹 (1/2) ∪ 𝐸 (1/2) ∩ 𝐹 (1) ∪ (𝐸 ∩ 𝐹) (1/2) ∩ 𝐸 (1/2) ∩ 𝐹 (1/2) ,

where the three sets on the right are mutually disjoint. By Federer’s theorem
(Theorem 8.21), the claim is thus proved.
We are now ready to prove (8.35). For simplicity we write 𝜇 = 𝜇 𝐸∩𝐹 . By
De Giorgi’s structure theorem (Theorem 8.13), |𝜇| = H 𝑛−1 ⌞(𝜕 ∗ (𝐸 ∩ 𝐹)),
(8.44) then implies
𝜇 = 𝜇⌞(𝐹 (1) ∩ 𝜕 ∗ 𝐸) + 𝜇⌞(𝐸 (1) ∩ 𝜕 ∗ 𝐹) + 𝜇⌞(𝜕 ∗ (𝐸 ∩ 𝐹) ∩ 𝜕 ∗ 𝐸 ∩ 𝜕 ∗ 𝐹) .
(8.46)
For 𝜈 ∈ S , we adopt the notation 𝐻 [𝜈] := {𝑦 ∈ R : ⟨𝑦, 𝜈⟩ < 0}. For any
𝑛 𝑛

𝑥 ∈ 𝜕 ∗ (𝐸 ∩ 𝐹) ∩ 𝐹 (1) ∩ 𝜕 ∗ 𝐸, we have by Tangential property of sets of finite

perimeter (Theorem 8.11) that, as 𝑟 ↘ 0,
loc loc loc
(𝐸 ∩ 𝐹)𝑥,𝑟 → 𝐻 [𝜈 𝐸∩𝐹 (𝑥)], 𝐸 𝑥,𝑟 → 𝐻 [𝜈 𝐸 (𝑥)], 𝐹𝑥,𝑟 → R𝑛 ,
loc
(𝐸 ∩ 𝐹)𝑥,𝑟 = 𝐸 𝑥,𝑟 ∩ 𝐹𝑥,𝑟 → 𝐻 [𝜈 𝐸 (𝑥)] ∩ R𝑛 = 𝐻 [𝜈 𝐸 (𝑥)],
and hence 𝜈 𝐸∩𝐹 (𝑥) = 𝜈 𝐸 (𝑥). Thus by De Giorgi’s structure theorem (The-
orem 8.13), (8.45) and Federer’s theorem (Theorem 8.21),
 
𝜇⌞(𝐹 (1) ∩ 𝜕 ∗ 𝐸) =𝜈 𝐸∩𝐹 H 𝑛−1 ⌞ 𝜕 ∗ (𝐸 ∩ 𝐹) ∩ 𝐹 (1) ∩ 𝜕 ∗ 𝐸
(8.47)
=𝜈 𝐸 H 𝑛−1 ⌞(𝐹 (1) ∩ 𝜕 ∗ 𝐸).
Similarly (the same argument with 𝐸 and 𝐹 exchanging),
𝜇⌞(𝐸 (1) ∩ 𝜕 ∗ 𝐹) = 𝜈 𝐹 H 𝑛−1 ⌞(𝐸 (1) ∩ 𝜕 ∗ 𝐹). (8.48)
50
Finally, if 𝑥 ∈ (𝜕 ∗ (𝐸 ∩ 𝐹) ∩ 𝜕 ∗ 𝐸 ∩ 𝜕 ∗ 𝐹), then again by Tangential property
(Theorem 8.11), as 𝑟 ↘ 0,
loc loc loc
(𝐸 ∩ 𝐹)𝑥,𝑟 → 𝐻 [𝜈 𝐸∩𝐹 (𝑥)], 𝐸 𝑥,𝑟 → 𝐻 [𝜈 𝐸 (𝑥)], 𝐹𝑥,𝑟 → 𝐻 [𝜈 𝐹 (𝑥)],
loc
(𝐸 ∩ 𝐹)𝑥,𝑟 = 𝐸 𝑥,𝑟 ∩ 𝐹𝑥,𝑟 → 𝐻 [𝜈 𝐸 (𝑥)] ∩ 𝐻 [𝜈 𝐹 (𝑥)],
which implies 𝜈 𝐸∩𝐹 (𝑥) = 𝜈 𝐸 (𝑥) = 𝜈 𝐹 (𝑥), and hence
(𝜕 ∗ (𝐸 ∩ 𝐹) ∩ 𝜕 ∗ 𝐸 ∩ 𝜕 ∗ 𝐹) ⊂ {𝜈 𝐸 = 𝜈 𝐹 }.
At the same time, if 𝑥 ∈ {𝜈 𝐸 = 𝜈 𝐹 }, then by again by Tangential property of
sets of finite perimeter, as 𝑟 ↘ 0,
(𝐸 ∩ 𝐹)𝑥,𝑟 = 𝐸 𝑥,𝑟 ∩ 𝐹𝑥,𝑟 → 𝐻 [𝜈 𝐸 (𝑥)] ∩ 𝐻 [𝜈 𝐹 (𝑥)] = 𝐻 [𝜈 𝐸 (𝑥)],
which immediately implies 𝑥 ∈ (𝐸 ∩ 𝐹) (1/2) and hence the inclusions
(𝜕 ∗ (𝐸 ∩ 𝐹) ∩ 𝜕 ∗ 𝐸 ∩ 𝜕 ∗ 𝐹) ⊂ {𝜈 𝐸 = 𝜈 𝐹 } ⊂ (𝐸 ∩ 𝐹) (1/2) .
Thus by Federer’s theorem (Theorem 8.21),
(𝜕 ∗ (𝐸 ∩ 𝐹) ∩ 𝜕 ∗ 𝐸 ∩ 𝜕 ∗ 𝐹) ≈ {𝜈 𝐸 = 𝜈 𝐹 },
and hence
𝜇⌞(𝜕 ∗ (𝐸 ∩ 𝐹) ∩ 𝜕 ∗ 𝐸 ∩ 𝜕 ∗ 𝐹) =𝜈 𝐸∩𝐹 H 𝑛−1 ⌞{𝜈 𝐸 = 𝜈 𝐹 }
(8.49)
=𝜈 𝐸 H 𝑛−1 ⌞{𝜈 𝐸 = 𝜈 𝐹 }.
By (8.46), (8.47), (8.48), and (8.49), we conclude the proof of (8.35).
(8.38) follows from (8.35) and the mutual singularity of the three Radon
measures on the right of (8.35). Since (R𝑛 \ 𝐸) (0) = 𝐸 (1) , (R𝑛 \ 𝐸) (1) = 𝐸 (0) ,
𝜇 𝐸 = −𝜇R𝑛 \𝐸 , and 𝐸 \ 𝐹 = 𝐸 ∩ (R𝑛 \ 𝐹), by (8.35) (precisely, using (8.35)
with 𝐹 therein replaced by R𝑛 \𝐹) we deduce (8.36) and consequently (8.39).
Since R𝑛 \ (𝐸 ∪ 𝐹) = (R𝑛 \ 𝐸) ∩ (R𝑛 \ 𝐹), we similarly prove (8.37) and
consequently (8.40). This completes the proof. □
Exercise 8.26. Show that if 𝐸, 𝐹 are sets of locally finite perimeter in R𝑛 ,
then so is 𝐸Δ𝐹, with
𝜇 𝐸Δ𝐹 = 𝜇 𝐸 ⌞𝐹 (0) + 𝜇 𝐹 ⌞𝐸 (0) − 𝜇 𝐸 ⌞𝐹 (1) − 𝜇 𝐹 ⌞𝐸 (1) .
In particular, for any 𝐺 ∈ B (R𝑛 ),
𝑃(𝐸Δ𝐹; 𝐺) = 𝑃(𝐸; 𝐺 \ 𝜕 ∗ 𝐹) + 𝑃(𝐹; 𝐺 \ 𝜕 ∗ 𝐸) ≤ 𝑃(𝐸; 𝐺) + 𝑃(𝐹; 𝐺).
Exercise 8.27. Show that if 𝐸, 𝐹 are sets of locally finite perimeter in R𝑛
with 𝐸 ⊂ 𝐹, then 𝜈 𝐸 = 𝜈 𝐹 on 𝜕 ∗ 𝐸 ∩ 𝜕 ∗ 𝐹, with
𝜇 𝐸 = 𝜇 𝐸 ⌞𝐹 (1) + 𝜈 𝐹 ⌞(𝜕 ∗ 𝐸 ∩ 𝜕 ∗ 𝐹).
In particular, 𝑃(𝐸) = 𝑃(𝐸; 𝐹 (1) ) + 𝑃(𝐸; 𝐹 (1/2) ).
51
 The following definition serves as connectedness in the theory of sets of
finite perimeter.
Definition 8.28 (Indecomposability). A set of finite perimeter 𝐸 is inde-
composable if, whenever 𝐸 = 𝐸 1 ∪ 𝐸 2 with |𝐸 1 ∩ 𝐸 2 | = 0 and 𝑃(𝐸) =
𝑃(𝐸 1 ) + 𝑃(𝐸 2 ), then either |𝐸 1 | = 0 or |𝐸 2 | = 0.
Exercise 8.29. Show that if 𝐸, 𝐹 are sets of finite perimeter in R𝑛 , 𝐸 is
indecomposable, and up to L 𝑛 - and H 𝑛−1 -negligible sets respectively, there
holds
𝐹 ⊂ 𝐸, 𝜕 ∗ 𝐹 ⊂ 𝜕 ∗ 𝐸,
then either |𝐹 | = 0 or |𝐸 \ 𝐹 | = 0.
Exercise 8.30. Show that if 𝐴 is an open, connected set with locally finite
perimeter in R𝑛 , then 𝐴 is indecomposable.

9. First variation of perimeter

9.1. First variation formula

Let 𝐸 be an open set in R𝑛 , let 𝑘 ∈ N. We say that 𝐸 has 𝐶 𝑘 -boundary if
for every 𝑥 ∈ 𝜕𝐸 there exist 𝑟 > 0 and 𝜓 ∈ 𝐶 𝑘 (𝐵𝑟 (𝑥)) with ∇𝜓(𝑦) ≠ 0 for
every 𝑦 ∈ 𝐵𝑟 (𝑥) and
𝐵𝑟 (𝑥) ∩ 𝐸 = {𝑦 ∈ 𝐵𝑟 (𝑥) : 𝜓(𝑦) < 0},

𝐵𝑟 (𝑥) ∩ 𝜕𝐸 = {𝑦 ∈ 𝐵𝑟 (𝑥) : 𝜓(𝑦) = 0}.

The outer unit normal 𝜈 𝐸 to 𝐸 is defined locally as
∇𝜓(𝑦)
𝜈 𝐸 (𝑦) = , ∀𝑦 ∈ 𝐵𝑟 (𝑥) ∩ 𝐸.
|∇𝜓(𝑦)|
Note that the definition is independent of the choice of 𝜓 and 𝑟, so 𝜈 𝐸 can be
viewed as a vector field defined on the whole 𝜕𝐸, with 𝜈 𝐸 ∈ 𝐶 𝑘−1 (𝜕𝐸; S𝑛−1 ).
We say that 𝑓 : R𝑛 → R𝑛 is a diffeomorphism of R𝑛 if 𝑓 is smooth,
bijective, and has a smooth inverse 𝑔 = 𝑓 −1 . Using the above notations, from
𝑓 ({𝑥 : 𝜓(𝑥) = 0}) = {𝑦 : 𝜓 ◦ 𝑔(𝑦) = 0} and the chain rule ∇(𝜓 ◦ 𝑔)(𝑦) =
(∇𝑔(𝑦)) ∗ [(∇𝜓) ◦ 𝑔(𝑦)], we find
(∇𝑔(𝑦)) ∗ [(∇𝜓) ◦ 𝑔(𝑦)]
𝜈 𝑓 (𝐸) (𝑦) = , ∀𝑦 ∈ 𝜕 𝑓 (𝐸) = 𝑓 (𝜕𝐸).
|(∇𝑔(𝑦)) ∗ [(∇𝜓) ◦ 𝑔(𝑦)] |
We next show that similar conclusions hold for sets of locally finite perimeter.

52
Proposition 9.1 (Diffeomorphic images of sets of finite perimeter). If 𝐸 is
a set of locally finite perimeter in R𝑛 and 𝑓 is a diffeomorphism of R𝑛 with
𝑔 = 𝑓 −1 , then 𝑓 (𝐸) is a set of locally finite perimeter in R𝑛 with
H 𝑛−1 ( 𝑓 (𝜕 ∗ 𝐸)Δ𝜕 ∗ 𝑓 (𝐸)) = 0, (9.1)
and
∫ ∫
𝜑𝜈 𝑓 (𝐸) dH 𝑛−1
= (𝜑 ◦ 𝑓 )Jf(∇𝑔 ◦ 𝑓 ) ∗ [𝜈 𝐸 ]dH 𝑛−1 , (9.2)
𝜕 ∗ 𝑓 (𝐸) 𝜕∗ 𝐸

for every 𝜑 ∈ 𝐶𝑐0 (R𝑛 ). In particular, for every 𝐹 ∈ B (R𝑛 ),

∫
∗
𝑛−1
H (𝐹 ∩ 𝜕 𝑓 (𝐸)) = Jf |∇𝑔 ◦ 𝑓 ) ∗ [𝜈 𝐸 ] | dH 𝑛−1 . (9.3)
𝑔(𝐹)∩𝜕 ∗ 𝐸

Note that (9.2) can be equivalently written as

𝜇 𝑓 (𝐸) = 𝑓# (Jf(∇𝑔 ◦ 𝑓 ) ∗ 𝜇 𝐸 ) .

Proof. Note that if 𝑢 ℎ → 𝑢 in 𝐿 1𝑙𝑜𝑐 (R𝑛 ), then 𝑢 ℎ ◦ 𝑔 → 𝑢 ◦ 𝑔 in 𝐿 1𝑙𝑜𝑐 (R𝑛 ) as

well, which follows from the following inequality: for any 𝐾 compact (𝑔(𝐾)
is then also compact), the area formula yields
∫ ∫ ∫
𝑛
|𝑢 ℎ ◦ 𝑔 − 𝑢 ◦ 𝑔| = |𝑢 ℎ − 𝑢|Jf ≤ 𝐿𝑖 𝑝( 𝑓 ; 𝑔(𝐾)) |𝑢 ℎ − 𝑢|.
𝐾 𝑔(𝐾) 𝑔(𝐾)

To proceed, define a tensor field 𝐺 𝑓 ∈ 𝐶 ∞ (R𝑛 ; R𝑛 ⊗ R𝑛 ) as

∗
𝐺 𝑓 = Jf(∇𝑔 ◦ 𝑓 ) .
If 𝑢 ∈ 𝐶 1 (R𝑛 ), 𝑇 ∈ 𝐶𝑐1 (R𝑛 ; R𝑛 ), 𝑣 = 𝑢 ◦ 𝑔, and 𝑆 = 𝑇 ◦ 𝑔, then by area
formula,
∫ ∫ ∫
𝑣div𝑆 = − ⟨𝑆, ∇𝑣⟩ = − ⟨𝑆, (∇𝑔) ∗ [∇𝑢 ◦ 𝑔]⟩
R R
∫R
𝑛 𝑛 𝑛
(9.4)
=− ⟨𝑇 (𝑥), 𝐺 𝑓 (𝑥)∇𝑢(𝑥)⟩ d𝑥.
R𝑛

Letting 𝑢 = ( 𝜒𝐸 ) ∗ 𝜂𝜖 , then 𝑣 = 𝑢 ◦ 𝑔 → 𝜒 𝑓 (𝐸) in 𝐿 1𝑙𝑜𝑐 (R𝑛 ) as 𝜖 ↘ 0. By

(9.4), Proposition 5.23, and De Giorgi’s structure theorem (Theorem 8.13),
∫ ∫
div𝑆 = ⟨𝑇 (𝑥), 𝐺 𝑓 (𝑥)𝜈 𝐸 (𝑥)⟩ dH 𝑛−1 (𝑥),
𝑓 (𝐸) 𝜕∗ 𝐸

which implies (recall Proposition 5.2) that 𝑓 (𝐸) is a set of locally finite
perimeter, with
∫ ∫
𝑛−1
𝑆, 𝜈 𝑓 (𝐸) dH = ⟨𝑇 (𝑥), 𝐺 𝑓 (𝑥)𝜈 𝐸 (𝑥)⟩ dH 𝑛−1 (𝑥).
𝜕 ∗ 𝑓 (𝐸) 𝜕∗ 𝐸
53
Then we choose 𝑆 = (𝜑𝑒) ∗ 𝜂𝜖 for any 𝜑 ∈ 𝐶𝑐0 (R𝑛 ) and 𝑒 ∈ S𝑛−1 , letting
𝜖 ↘ 0 we find (recall 𝑇 = 𝑆 ◦ 𝑓 )
 ∫   ∫ 
𝑛−1 𝑛−1
𝑒, 𝜑𝜈 𝑓 (𝐸) dH = 𝑒, (𝜑 ◦ 𝑔)𝐺 𝑓 𝜈 𝐸 dH ,
𝜕 ∗ 𝑓 (𝐸) 𝜕∗ 𝐸
which implies (9.2).
By (9.2) and an approximation argument,
∫ ∫
𝑛−1
𝜈 𝑓 (𝐸) dH = 𝐺 𝑓 𝜈 𝐸 dH 𝑛−1 , (9.5)
𝐹∩𝜕 ∗ 𝑓 (𝐸) 𝑔(𝐹)∩𝜕 ∗ 𝐸
for every Borel set 𝐹 ⊂ R𝑛 . Taking total variation in (9.5), (9.3) follows.
Now we show that 𝑓 (𝐸 (1) ) = 𝑓 (𝐸) (1) and 𝑓 (𝐸 (0) ) = 𝑓 (𝐸) (0) , so as to
show that 𝜕 𝑒 𝑓 (𝐸) = 𝑓 (𝜕 𝑒 𝐸). Once this is done, (9.1) is then proved by
Federer’s theorem (Theorem 8.21).
To do this, note that (R𝑛 \ 𝐸) (0) = 𝐸 (1) and 𝑓 is a bijection, so we just
have to show 𝑓 (𝐸 (0) ) = 𝑓 (𝐸) (0) . Let us set
𝐿 𝑥 = 𝐿𝑖 𝑝(𝑔; 𝐵1 ( 𝑓 (𝑥))), 𝑀𝑥 = 𝐿𝑖 𝑝( 𝑓 ; 𝐵 𝐿 𝑥 (𝑥)), 𝑥 ∈ R𝑛 ,
then 𝑔(𝐵𝑟 ( 𝑓 (𝑥))) ⊂ 𝐵 𝐿 𝑥 𝑟 (𝑥) for every 𝑟 < 1, and hence by area formula
∫
| 𝑓 (𝐸) ∩ 𝐵𝑟 ( 𝑓 (𝑥))| = Jf ≤ 𝑀𝑥𝑛 𝐸 ∩ 𝐵 𝐿 𝑥 𝑟 (𝑥) ,
𝐸∩𝑔(𝐵𝑟 ( 𝑓 (𝑥)))

for every 𝑥 ∈ R𝑛 , 𝑟 < 1; in particular, if 𝑥 ∈ 𝐸 (0) then 𝑓 (𝑥) ∈ 𝑓 (𝐸) (0) ,

which completes the proof. □
We provide the second order Taylor’s expansion of the determinant close
to the identity. Note that the first order expansion is sufficient to derive
the first variation formula of perimeter, we include here the second order
expansion since it will be needed in the future (second variation formula).
These are linear algebra facts, for the proof, see [2, Lemma 17.4].
Lemma 9.2. If 𝑍 ∈ R𝑛 ⊗ R𝑛 , (write) Id = IdR𝑛 , and 𝑍 2 = 𝑍 ◦ 𝑍, then
(Id + 𝑡𝑍) −1 = Id − 𝑡𝑍 + 𝑡 2 𝑍 2 + 𝑂 (𝑡 3 ),
and
𝑡2  
det(Id + 𝑡𝑍) = 1 + 𝑡tr(𝑍) + tr(𝑍) 2 − tr(𝑍 2 ) + 𝑂 (𝑡 3 ).
2
We need the following further notations to state the first variation formula
of perimeter.
Definition 9.3 (local variation). A one parameter family of diffeomorphisms
of R𝑛 is a smooth function
(𝑥, 𝑡) ∈ R𝑛 × (−𝜖, 𝜖) ↦→ 𝑓 (𝑡, 𝑥) = 𝑓𝑡 (𝑥) ∈ R𝑛 , 𝜖 > 0,
54
such that for each fixed 𝑡 with |𝑡| < 𝜖, 𝑓𝑡 : R𝑛 → R𝑛 is a diffeomorphism of
R𝑛 .
Given an open set 𝐴 in R𝑛 , we say that { 𝑓𝑡 } |𝑡|<𝜖 is a local variation in 𝐴
if it defines as one-parameter family of diffeomorphisms such that
𝑓0 (𝑥) = 𝑥, 𝑥 ∈ R𝑛 ,
and
{𝑥 ∈ R𝑛 : 𝑓𝑡 (𝑥) ≠ 𝑥} ⊂⊂ 𝐴, ∀|𝑡| < 𝜖 .
Under this assumption it is easy to see that
𝑓𝑡 (𝐸)Δ𝐸 ⊂⊂ 𝐴, ∀𝐸 ⊂ R𝑛 ,
and the following Taylor expansions hold uniformly on R𝑛 :
𝑓𝑡 (𝑥) = 𝑥 + 𝑡𝑇 (𝑥) + 𝑂 (𝑡 2 ), ∇ 𝑓𝑡 (𝑥) = Id + 𝑡∇𝑇 (𝑥) + 𝑂 (𝑡 2 ),
where 𝑇 ∈ 𝐶𝑐∞ ( 𝐴; R𝑛 ) is called the initial velocity of the family { 𝑓𝑡 } |𝑡|<𝜖 ,
𝜕 𝑓𝑡
𝑇 (𝑥) = (𝑥, 0), 𝑥 ∈ R𝑛 .
𝜕𝑡
Conversely, starting from 𝑇 ∈ 𝐶𝑐∞ ( 𝐴; R𝑛 ) there are two general ways to
construct a local variation { 𝑓𝑡 } |𝑡|<𝜖 in 𝐴 having 𝑇 as its initial velocity. One
is simply defining
𝑓𝑡 (𝑥) = 𝑥 + 𝑡𝑇 (𝑥), 𝑥 ∈ R𝑛 .
The other relies on solving the (ODE) Cauchy problems
𝜕
𝑓 (𝑡, 𝑥) = 𝑇 ( 𝑓 (𝑡, 𝑥)), 𝑥 ∈ R𝑛 ,
𝜕𝑡
𝑓 (0, 𝑥) = 𝑥, 𝑥 ∈ R𝑛 ,
which is solvable for small values of 𝑡. In both cases, we say that { 𝑓𝑡 } |𝑡|<𝜖 is
a local variation associated with 𝑇. Finally, we define the first variation of
perimeter (relative to 𝐴) with respect to local variation (associated with 𝑇)
{ 𝑓𝑡 } |𝑡|<𝜖 in 𝐴 to be
d
| 𝑡=0 𝑃( 𝑓𝑡 (𝐸); 𝐴), for 𝑇 ∈ 𝐶𝑐∞ ( 𝐴; R𝑛 ) given.
d𝑡

Theorem 9.4 (First variation of perimeter). If 𝐴 is an open set in R𝑛 , 𝐸 is

a set of locally finite perimeter, and { 𝑓𝑡 } |𝑡|<𝜖 is a local variation in 𝐴, then
∫
𝑃( 𝑓𝑡 (𝐸); 𝐴) = 𝑃(𝐸; 𝐴) + 𝑡 div𝐸 𝑇dH 𝑛−1 + 𝑂 (𝑡 2 ),
𝜕∗ 𝐸
where 𝑇 is the initial velocity of { 𝑓𝑡 } |𝑡|<𝜖 and div𝐸 𝑇 : 𝜕 ∗ 𝐸 → R is given by
div𝐸 𝑇 (𝑥) = div𝑇 (𝑥) − ⟨𝜈 𝐸 (𝑥), ∇𝑇 (𝑥) [𝜈 𝐸 (𝑥)]⟩ , 𝑥 ∈∗ 𝐸,
is a Borel function called the boundary divergence of 𝑇 on 𝐸.
55
 In particular,
∫
d
| 𝑡=0 𝑃( 𝑓𝑡 (𝐸); 𝐴) = div𝐸 𝑇dH 𝑛−1 . (9.6)
d𝑡 𝜕∗ 𝐸

Proof. By Proposition 9.1 (in particular, (9.3)),

∫
𝑃( 𝑓𝑡 (𝐸); 𝐴) = Jft |(∇𝑔𝑡 ◦ 𝑓𝑡 ) ∗ 𝜈 𝐸 | dH 𝑛−1 , 𝑔𝑡 = ( 𝑓𝑡 ) −1 ,
𝐴∩𝜕 ∗ 𝐸
so that 𝑃( 𝑓𝑡 (𝐸); 𝐴) is a smooth function on 𝑡 in a neighborhood of 𝑡 = 0.
Since ∇ 𝑓𝑡 = Id + 𝑡∇𝑇 + 𝑂 (𝑡 2 ), by Lemma 9.2 we have
∇𝑔𝑡 ◦ 𝑓𝑡 = (∇ 𝑓𝑡 ) −1 = Id − 𝑡∇𝑇 + 𝑂 (𝑡 2 ),

Jft = 1 + 𝑡div𝑇 + 𝑂 (𝑡 2 ),
uniformly on R𝑛 as 𝑡 → 0. In particular,
|(∇𝑔𝑡 ◦ 𝑓𝑡 ) ∗ [𝜈 𝐸 ]| 2 =|𝜈 𝐸 − 𝑡 (∇𝑇) ∗ [𝜈 𝐸 ]| 2 + 𝑂 (𝑡 2 ) = 1 − 2𝑡 ⟨𝜈 𝐸 , (∇𝑇) ∗ [𝜈 𝐸 ]⟩ + 𝑂 (𝑡 2 )
=1 − 2𝑡 ⟨𝜈 𝐸 , (∇𝑇) [𝜈 𝐸 ]⟩ + 𝑂 (𝑡 2 ),
and hence we conclude
Jft |(∇𝑔𝑡 ◦ 𝑓𝑡 ) ∗ 𝜈 𝐸 | = 1 + 𝑡 (div𝑇 − ⟨𝜈 𝐸 , ∇𝑇 [𝜈 𝐸 ]⟩) + 𝑂 (𝑡 2 )
as required. □
If 𝐸 is an open set with 𝐶 2 -boundary, then by tangential divergence
theorem we find
∫ ∫
div𝜕𝐸 𝑇dH 𝑛−1
= ⟨𝑇, H𝜕𝐸 ⟩ dH 𝑛−1 , ∀𝑇 ∈ 𝐶𝑐1 (R𝑛 ; R𝑛 ),
𝜕𝐸 𝜕𝐸
where div𝜕𝐸 denotes the tangential divergence with respect to 𝜕𝐸, and H𝜕𝐸
is the mean curvature vector to 𝜕𝐸. This motivates us to define
Definition 9.5 (distributional mean curvature vector). If 𝐸 is a set of locally
finite perimeter in R𝑛 , then the distributional mean curvature vector of 𝐸
in an open set 𝐴 is the functional H𝐸 : 𝐶𝑐∞ ( 𝐴; R𝑛 ) → R, defined by the
formula ∫
⟨H𝐸 , 𝑇⟩ = div𝐸 𝑇dH 𝑛−1 , 𝑇 ∈ 𝐶𝑐∞ ( 𝐴; R𝑛 ).
𝜕∗ 𝐸
Note that H𝐸 = HR𝑛 \𝐸 .
We say that 𝐸 has (locally summable) distributional (scalar) mean cur-
vature in 𝐴 if there exists 𝐻 ∈ 𝐿 1𝑙𝑜𝑐 ( 𝐴 ∩ 𝜕 ∗ 𝐸; H 𝑛−1 ), such that
∫ ∫
div𝐸 𝑇dH 𝑛−1
= ⟨𝑇, 𝐻𝜈 𝐸 ⟩ dH 𝑛1 , 𝑇 ∈ 𝐶𝑐∞ ( 𝐴; R𝑛 ).
𝜕∗ 𝐸 𝜕∗ 𝐸
56
In this case, 𝐻 is uniquely defined H 𝑛−1 -a.e. on 𝐴 ∩ 𝜕 ∗ 𝐸 by Lebesgue points
lemma, and H𝐸 is well-defined as
H𝐸 = 𝐻 𝐸 𝜈 𝐸 H 𝑛−1 ⌞( 𝐴 ∩ 𝜕 ∗ 𝐸).
A more general form of the first variation formula of perimeter is as
follows (recall Remark 8.14).
Exercise 9.6. If 𝑀 is a locally H 𝑛−1 -rectifiable set in R𝑛 and { 𝑓𝑡 } |𝑡|<𝜖 is a
local variation with initial velocity 𝑇 ∈ 𝐶𝑐∞ ( 𝐴; R𝑛 ), then
∫
d 𝑛−1
| 𝑡=0 H ( 𝑓𝑡 (𝑀)) = div 𝑀 𝑇dH 𝑛−1 .
d𝑡 𝑀

9.2. Stationary sets of perimeter

Definition 9.7. We say that a set of locally finite perimeter 𝐸 is stationary
for perimeter in an open set 𝐴 ⊂ R𝑛 if spt𝜇 𝐸 = 𝜕𝐸, and
d
| 𝑡=0 𝑃( 𝑓𝑡 (𝐸); 𝐴) = 0, (9.7)
d𝑡
whenever { 𝑓𝑡 } |𝑡|<𝜖 is a local variation in 𝐴.
Remark 9.8. Note that the condition spt𝜇 𝐸 = 𝜕𝐸 is not a real restriction,
since if a set of locally finite perimeter 𝐸 ′ satisfies (9.7), then by Proposition
5.18 we can always find a set of finite perimeter 𝐸, equivalent to 𝐸 ′ in the
sense of sets of locally finite perimeter, such that the required properties are
satisfied.
We have nevertheless included this in the definition, so as to give a precise
geometric meaning to the notion of topological boundary.
Corollary 9.9 (Vanishing mean curvature). A set of locally finite perimeter
𝐸 is stationary for perimeter in the open set 𝐴 if and only if
∫
div𝐸 𝑇dH 𝑛−1 = 0, ∀𝑇 ∈ 𝐶𝑐1 ( 𝐴; R𝑛 ). (9.8)
𝜕∗ 𝐸
In particular, 𝐸 has vanishing distributional mean curvature in 𝐴.
Proof. This follows directly from Theorem 9.4, especially (9.6). □
Exercise 9.10. Show that the cone 𝐸 = {𝑥 ∈ R2 : 𝑥 1 𝑥 2 > 0} is stationary
for perimeter in R2 .
Theorem 9.11 (Monotonicity of density ratios). If 𝐸 is stationary for
perimeter in the open set 𝐴 and 𝑥 0 ∈ 𝐴, then the density ratios
𝑃(𝐸; 𝐵𝑟 (𝑥 0 ))
𝜔𝑛−1𝑟 𝑛−1
are increasing on 𝑟 ∈ (0, dist(𝑥0 , 𝜕 𝐴)).
57
Proof. Write 𝑑 = dist(𝑥 0 , 𝜕 𝐴). Assume WLOG that 𝑥 0 = 0 and write for
simplicity 𝐵𝑟 = 𝐵𝑟 (𝑥 0 ). Let 𝜑 ∈ 𝐶 ∞ (R; [0, 1]) be bump function with
1
𝜑(𝑠) = 1 if 𝑠 ≤ , 𝜑 = 0 if 𝑠 ≥ 1, 𝜑′ ≤ 0 on R. (9.9)
2
Consider the function Φ ∈ 𝐶 ∞ ((0, 𝑑)) defined by
∫
|𝑥|
Φ(𝑟) := 𝜑( )dH 𝑛−1 (𝑥), 𝑟 ∈ (0, 𝑑).
𝜕∗ 𝐸 𝑟
Claim. 𝑟 1−𝑛 Φ(𝑟) is increasing on 𝑟 ∈ (0, 𝑑).
To see this, define
|𝑥|
𝑇𝑟 ∈ 𝐶𝑐1 ( 𝐴; R𝑛 ), 𝑇𝑟 (𝑥) = 𝜑( )𝑥, 𝑥 ∈ R𝑛 ,
𝑟
and test (9.8) by 𝑇𝑟 . Direct computations yield
|𝑥| |𝑥| ′ |𝑥| 𝑥 𝑥
∇𝑇𝑟 = 𝜑( )Id + 𝜑( ) ⊗ , ∀𝑥 ∈ R𝑛 ,
𝑟 𝑟 𝑟 |𝑥| |𝑥|
|𝑥| |𝑥| ′ |𝑥|
div𝑇𝑟 = tr∇𝑇𝑟 = 𝑛𝜑( ) + 𝜑 ( ), ∀𝑥 ∈ R𝑛 ,
𝑟 𝑟 𝑟
|𝑥| |𝑥| ′ |𝑥| ⟨𝑥, 𝜈 𝐸 (𝑥)⟩ 2
⟨𝜈 𝐸 , ∇𝑇𝑟 [𝜈 𝐸 ]⟩ = 𝜑( ) + 𝜑( ) 2
, ∀𝑥 ∈ 𝜕 ∗ 𝐸,
𝑟 𝑟 𝑟 |𝑥|
⟨𝑥, 𝜈 𝐸 (𝑥)⟩ 2
 
|𝑥| |𝑥| ′ |𝑥|
div𝐸 𝑇𝑟 = (𝑛 − 1)𝜑( ) + 𝜑 ( ) 1− , ∀𝑥 ∈ 𝜕 ∗ 𝐸,
𝑟 𝑟 𝑟 |𝑥| 2
and hence we obtain
∫ ∫
|𝑥| 𝑛−1 |𝑥| ′ |𝑥|
(𝑛 − 1) 𝜑( )dH + 𝜑 ( )dH 𝑛−1
𝜕∗ 𝐸 𝑟 𝜕∗ 𝐸 𝑟 𝑟
(9.10)
|𝑥| ′ |𝑥| ⟨𝑥, 𝜈 𝐸 (𝑥)⟩ 2
∫
= 𝜑( ) 2
dH 𝑛−1 .
∗
𝜕 𝐸 𝑟 𝑟 |𝑥|
′
Since 𝜑′ ≤ 0, (9.10) implies (𝑛 − 1)Φ(𝑟) − Φ 𝑟(𝑟) ≤ 0 for 𝑟 ∈ (0, 𝑑), that is,
 ′
1−𝑛
𝑟 Φ(𝑟) ≥ 0 for 𝑟 ∈ (0, 𝑑),
which proves the claim.
To proceed, we select a sequence of functions {𝜑 ℎ } ℎ∈N ⊂ 𝐶 ∞ (R; [0, 1])
with each 𝜑 ℎ satisfying (9.9), and such that 𝜑 ℎ monotonically converges to
𝜒(−∞,1) as ℎ → ∞, then by monotone convergence we find
∫
|𝑥|
lim Φℎ (𝑟) = lim 𝜑 ℎ ( )dH 𝑛−1 (𝑥) = 𝑃(𝐸; 𝐵𝑟 ).
ℎ→∞ ℎ→∞ 𝜕 ∗ 𝐸 𝑟
We thus deduce that 𝑟 1−𝑛 𝑃(𝐸; 𝐵𝑟 ) is increasing on 𝑟 ∈ (0, 𝑑). □
58
Corollary 9.12 (Density estimates for stationary sets). If 𝐸 is stationary for
perimeter in the open set 𝐴, then
𝑃(𝐸; 𝐵𝑟 (𝑥)) ≥ 𝜔𝑛−1𝑟 𝑛−1 , (9.11)
for every 𝑥 ∈ 𝐴 ∩ 𝜕𝐸 and every 𝐵𝑟 (𝑥) ⊂ 𝐴. In particular,
H 𝑛−1 ( 𝐴 ∩ (𝜕𝐸 \ 𝜕 ∗ 𝐸)) = 0, (9.12)
where 𝜃 𝑛−1 (𝜕𝐸) and 𝜃 𝑛−1 (𝜕 ∗ 𝐸) exist and coincide on 𝐴 ∩ 𝜕𝐸, with
𝜃 𝑛−1 (𝜕𝐸) ≥ 1, ∀𝑥 ∈ 𝐴 ∩ 𝜕𝐸,
𝜃 𝑛−1 (𝜕𝐸) = 1, ∀𝑥 ∈ 𝐴 ∩ 𝜕 ∗ 𝐸.
Proof. By Theorem 9.11 we define
𝑃(𝐸; 𝐵𝑟 (𝑥))
𝛾(𝑥, 𝑟) = , 𝑥 ∈ 𝐴, 𝑟 < dist(𝑥, 𝜕 𝐴),
𝜔𝑛−1𝑟 𝑛−1
which is an increasing function on 𝑟 variable.
Since by (8.21), 𝜃 𝑛−1 (𝜕 ∗ 𝐸)(𝑥) = 𝛾(𝑥, 0+ ) = 1 for 𝑥 ∈ 𝐴 ∩ 𝜕 ∗ 𝐸, we find
𝛾(𝑥, 𝑟) ≥ 1, ∀𝑥 ∈ 𝐴 ∩ 𝜕 ∗ 𝐸, 𝐵𝑟 (𝑥) ⊂ 𝐴. (9.13)
If 𝑥 ∈ 𝐴 ∩ (𝜕𝐸 \ 𝜕 ∗ 𝐸), then by (8.3) there exists {𝑥 ℎ } ℎ∈N ⊂ 𝐴 ∩ 𝜕 ∗ 𝐸 such
that 𝑥 ℎ → 𝑥. By (9.13), we conclude for a.e. 𝑟 > 0
𝑃(𝐸; 𝐵𝑟 (𝑥)) = H 𝑛−1 (𝜕 ∗ 𝐸 ∩ 𝐵𝑟 (𝑥)) = lim H 𝑛−1 (𝜕 ∗ 𝐸 ∩ 𝐵𝑟 (𝑥 ℎ )) ≥ 𝜔𝑛−1𝑟 𝑛−1 .
ℎ→∞

By monotonicity of 𝛾(𝑥, 𝑟) for 𝑥 ∈ 𝐴 and 𝑟 ∈ (0, dist(𝑥, 𝜕 𝐴)), we conclude

that 𝛾(𝑥, 𝑟) ≥ 1 for every 𝑥 ∈ 𝐴 ∩ 𝜕𝐸 and 𝐵𝑟 (𝑥) ⊂ 𝐴, which proves (9.11).
Set 𝜇 = H 𝑛−1 ⌞( 𝐴 ∩ 𝜕 ∗ 𝐸), then by (9.11), 𝜃 𝑛−1
∗ (𝜇) ≥ 1 on 𝐴 ∩ 𝜕𝐸. By

Theorem 8.23,
H 𝑛−1 ( 𝐴 ∩ 𝜕 ∗ 𝐸) = 𝜇( 𝐴 ∩ 𝜕𝐸) ≥ H 𝑛−1 ( 𝐴 ∩ 𝜕𝐸),
which implies (9.12).
Notice that Theorem 9.11 gives
inf 𝛾(𝑥, 𝑟) = lim 𝛾(𝑥, 𝑟) = 𝜃 𝑛−1 (𝜕 ∗ 𝐸)(𝑥), ∀𝑥 ∈ 𝐴 ∩ 𝜕𝐸,
𝑟>0 𝑟↘0
implying that 𝜃 𝑛−1 ∗
(𝜕 𝐸)(𝑥) exists at every 𝑥 ∈ 𝐴 ∩ 𝜕𝐸.
By (9.12), 𝜃 𝑛−1 (𝜕𝐸)(𝑥) = 𝜃 𝑛−1 (𝜕 ∗ 𝐸)(𝑥) at every 𝑥 ∈ 𝐴 ∩ 𝜕𝐸, with
𝜃 𝑛−1 (𝜕𝐸) ≥ 1 thanks to (9.11). For 𝑥 ∈ 𝐴 ∩ 𝜕 ∗ 𝐸 we have by corollary of
tangential property, see (8.21), that 𝜃 𝑛−1 (𝜕𝐸)(𝑥) = 1, which completes the
proof. □

59
Chapter 3: Minimizers

10. Perimeter minimizers

Definition 10.1 (Perimeter minimizer). Given an open bounded set 𝐴 ⊂ R𝑛
and a set of locally finite perimeter 𝐸 in R𝑛 , we say that 𝐸 is a perimeter
minimizer in 𝐴, if spt𝜇 𝐸 = 𝜕𝐸, and whenever 𝐸Δ𝐹 ⊂⊂ 𝐴,
𝑃(𝐸; 𝐴) ≤ 𝑃(𝐹; 𝐴). (10.1)
The reason that spt𝜇 𝐸 = 𝜕𝐸 is included in the definition can be found in
Remark 9.8.
Exercise 10.2. Show that every open half-space in R𝑛 is a perimeter mini-
mizer in R𝑛 .
Definition 10.3 (Local perimeter minimizer). Given an open bounded set
𝐴 ⊂ R𝑛 and a set of locally finite perimeter 𝐸 in R𝑛 , we say that 𝐸 is a
local perimeter minimizer in 𝐴 at scale 𝑟 0 , if spt𝜇 𝐸 = 𝜕𝐸, and (10.1) holds
whenever 𝐸Δ𝐹 ⊂⊂ 𝐴 ∩ 𝐵𝑟 (𝑥0 ), 𝑥 ∈ 𝐴.
Exercise 10.4. Show that the cone constructed in Exercise 9.10 is not a local
perimeter minimizer in R2 .
It is easy to see that, if 𝐸 is a local perimeter minimizer in 𝐴, then 𝐸 is
stationary for perimeter in any 𝐵𝑟 (𝑥) ∩ 𝐴 ⊂ 𝐵𝑟0 (𝑥) ∩ 𝐴 ⊂ 𝐴 in the sense
of Definition 9.7. Therefore we have the following direct consequence of
Corollary 9.12.
Theorem 10.5. Given an open bounded set 𝐴 ⊂ R𝑛 and a local perimeter
minimizer 𝐸 in 𝐴, then for every 𝐵𝑟 (𝑥) ⊂⊂ 𝐴 with 𝑥 ∈ 𝐴 ∩ 𝜕𝐸, 𝑟 < 𝑟 0 , the
estimates in Corollary 9.12 hold. In fact, we have the following.
1 |𝐸 ∩ 𝐵𝑟 (𝑥)| 1
≤ ≤ 1 − 𝑛, (10.2)
2 𝑛 𝜔𝑛 𝑟 𝑛 2
𝑃(𝐸; 𝐵𝑟 (𝑥))
𝜔𝑛−1 ≤ ≤ 𝑛𝜔𝑛 , (10.3)
𝑟 𝑛−1
and as in Corollary 9.12,
H 𝑛−1 ( 𝐴 ∩ (𝜕𝐸 ∩ 𝜕 ∗ 𝐸)) = 0.
Proof. By Proposition 5.18, since we are assuming spt𝜇 𝐸 = 𝜕𝐸, we have
0 < |𝐸 ∩ 𝐵𝑟 (𝑥)| < 𝜔𝑛 𝑟 𝑛 , ∀𝑥 ∈ 𝜕𝐸, 𝑟 > 0. (10.4)
As usual, we have
H 𝑛−1 (𝜕 ∗ 𝐸 ∩ 𝜕𝐵𝑟 (𝑥)) = 0, a.e. 𝑟 < min{dist(𝑥, 𝜕 𝐴), 𝑟 0 } =: 𝑑. (10.5)
60
For any such 𝑟, let 𝐹 = 𝐸 ∩ 𝐵𝑟 (𝑥), then 𝐸Δ𝐹 ⊂⊂ 𝐵 𝑠 (𝑥) ⊂ 𝐴 for any
𝑠 ∈ (𝑟, 𝑑). Using local perimeter minimality, (10.5) and (8.42), we find
𝑃(𝐸; 𝐵 𝑠 (𝑥)) ≤𝑃(𝐹; 𝐵 𝑠 (𝑥)) = 𝑃(𝐸 \ 𝐵𝑟 (𝑥); 𝐵 𝑠 (𝑥))
=H 𝑛−1 (𝐸 (1) ∩ 𝜕𝐵𝑟 (𝑥)) + 𝑃(𝐸; 𝐵 𝑠 (𝑥) \ 𝐵𝑟 (𝑥)).
Letting 𝑠 ↘ 𝑟 we obtain
𝑃(𝐸; 𝐵𝑟 (𝑥)) ≤ H 𝑛−1 (𝐸 (1) ∩ 𝜕𝐵𝑟 (𝑥)), (10.6)
which implies the upper bound in (10.3) since H 𝑛−1 (𝐸 (1) ∩ 𝜕𝐵𝑟 (𝑥)) ≤
𝑃(𝐵𝑟 (𝑥)) = 𝑛𝜔𝑛 𝑟 𝑛−1 . On the other hand, by adding the term H 𝑛−1 (𝐸 (1) ∩
𝜕𝐵𝑟 (𝑥)) to both sides of the(10.6) and recall that by (8.41) and (10.5),
𝑃(𝐸 ∩ 𝐵𝑟 (𝑥)) = 𝑃(𝐸; 𝐵𝑟 (𝑥)) + H 𝑛−1 (𝐸 (1) ∩ 𝜕𝐵𝑟 (𝑥)),
we thus find
𝑃(𝐸 ∩ 𝐵𝑟 (𝑥)) ≤ 2H 𝑛−1 (𝐸 (1) ∩ 𝜕𝐵𝑟 (𝑥)),
so that by the Euclidean isoperimetric inequality, applied to 𝐸 ∩ 𝐵𝑟 (𝑥0,
1 𝑛−1
𝑛𝜔𝑛𝑛 |𝐵𝑟 (𝑥) ∩ 𝐸 | 𝑛 ≤ 2H 𝑛−1 (𝐸 (1) ∩ 𝜕𝐵𝑟 (𝑥)). (10.7)
As shown before (using co-area formula for distance function), the function
𝑚 : (0, ∞) → [0, ∞), 𝑠 ↦→ |𝐸 ∩ 𝐵 𝑠 (𝑥)| = |𝐸 (1) ∩ 𝐵 𝑠 (𝑥)| is monotone with
𝑚′ (𝑠) = H 𝑛−1 (𝐸 (1) ∩ 𝜕𝐵 𝑠 (𝑥)) for a.e. 𝑠 > 0. Hence (10.7) gives
1 𝑛−1
𝑛𝜔𝑛𝑛 𝑚(𝑟) 𝑛 ≤ 2𝑚′ (𝑟), a.e. 𝑟 < 𝑑.
By (10.4), 𝑚 > 0 with 𝑚(0+ ) = 0. Hence we may divide the above inequality
𝑛−1
both sides by 𝑚 𝑛 then integrate on (0, 𝑟), to find
1 1
𝜔𝑛𝑛 𝑟 ≤ 2𝑚(𝑟) 𝑛 , ∀𝑟 < 𝑑,
which proves the lower bound in (10.2).
To proceed, note that R𝑛 \ 𝐸 is also a local perimeter minimizer in 𝐴 at the
scale 𝑟 0 , and the upper bound in (10.4) implies that |(R𝑛 \ 𝐸) ∩ 𝐵 𝑠 (𝑥)| > 0
for every 𝑠 > 0, therefore we can repeat the above argument with R𝑛 \ 𝐸 in
place of 𝐸 to deduce the upper bound in (10.2). The required estimates are
now all clear and the proof is completed. □
Theorem 10.6 (Comparison sets by replacement). If 𝐸, 𝐺 are sets of locally
finite perimeter in R𝑛 and 𝐴 is an open set of finite perimeter such that
H 𝑛−1 (𝜕 ∗ 𝐴 ∩ 𝜕 ∗ 𝐸) = H 𝑛−1 (𝜕 ∗ 𝐴 ∩ 𝜕 ∗ 𝐺) = 0, (10.8)
then the set 𝐹 defined as
𝐹 = (𝐺 ∩ 𝐴) ∪ (𝐸 \ 𝐴)
61
is a set of locally finite perimeter. Moreover, if 𝐴 ⊂⊂ 𝐴′ and 𝐴′ is open,
then
 
𝑃(𝐹; 𝐴′) = 𝑃(𝐺; 𝐴) + 𝑃(𝐸; 𝐴′ \ 𝐴) + H 𝑛−1 (𝐸 (1) Δ𝐺 (1) ) ∩ 𝜕 ∗ 𝐴 .
(10.9)
Proof. As we have proven before, 𝐹 is a set of locally finite perimeter. By
Theorem 8.25, in particular, (8.37),
𝜇 𝐹 =𝜇 𝐸\𝐴 ⌞(𝐺 ∩ 𝐴) (0) + 𝜇𝐺∩𝐴 ⌞(𝐸 ∩ 𝐴) (0) + 𝜈𝐺∩𝐴 H 𝑛−1 ⌞{𝜈𝐺∩𝐴 = 𝜈 𝐸\𝐴 }
:=(𝐼) + (𝐼 𝐼) + (𝐼 𝐼 𝐼).
(10.10)
For (𝐼), by (10.8) and (8.36), we have
𝜇 𝐸\𝐴 = 𝜇 𝐸 ⌞(R𝑛 \ 𝐴) − 𝜇 𝐴 ⌞𝐸 (1) . (10.11)

Moreover, by the fact that R𝑛 \ 𝐴 = 𝐴 (0) ⊂ (𝐺 ∩ 𝐴) (0) , we have

(𝜇 𝐸 ⌞(R𝑛 \ 𝐴))⌞(𝐺 ∩ 𝐴) (0) = 𝜇 𝐸 ⌞(R𝑛 \ 𝐴). (10.12)
By the fact that 𝐺 (0) ∩ 𝜕 ∗ 𝐴 ⊂ (𝐺 ∩ 𝐴) (0) ∩ 𝜕 ∗ 𝐴, and by Federer’s theorem
(Theorem 8.21) also (10.8), for H 𝑛−1 -a.e. 𝑧 ∈ (𝐺 ∩ 𝐴) (0) ∩ 𝜕 ∗ 𝐴 we have
𝑧 ∈ 𝐺 (1) ∪ 𝐺 (0) ; now, if 𝑧 ∈ 𝐺 (1) then
𝜃 𝑛 (𝐺 ∩ 𝐴) + 𝜃 𝑛 (𝐺 ∪ 𝐴) = 𝜃 𝑛 (𝐺) + 𝜃 𝑛 ( 𝐴),
we find 𝜃 𝑛 (𝐺 ∪ 𝐴) ≥ 32 , which is impossible. Thus, (𝐺 ∩ 𝐴) (0) ∩ 𝜕 ∗ 𝐴 is
H 𝑛−1 -equivalent to 𝐺 (0) ∩ 𝜕 ∗ 𝐴, which shows
(𝜇 𝐴 ⌞𝐸 (1) )⌞(𝐺 ∩ 𝐴) (0) = 𝜇 𝐴 ⌞(𝐸 (1) ∩ 𝐺 (0) ). (10.13)
From (10.11), (10.12), (10.13) we get
(𝐼) = 𝜇 𝐸\𝐴 ⌞(𝐺 ∩ 𝐴) (0) = 𝜇 𝐸 ⌞(R𝑛 \ 𝐴) − 𝜇 𝐴 ⌞(𝐸 (1) ∩ 𝐺 (0) ).
For (𝐼 𝐼), by (10.8) and (8.35) we have
𝜇𝐺∩𝐴 = 𝜇𝐺 ⌞ 𝐴 + 𝜇 𝐴 ⌞𝐺 (1) . (10.14)
Moreover, by the fact that 𝐴 open and hence 𝐴 ⊂ (𝐸 \ 𝐴) (0) , we obtain
(𝜇𝐺 ⌞ 𝐴)⌞(𝐸 \ 𝐴) (0) = 𝜇𝐺 ⌞ 𝐴. (10.15)
Next, note that on one hand we have 𝐸 (0) ∩ 𝜕 ∗ 𝐴 ⊂ (𝐸 ∩ 𝐴) (0) ∩ 𝜕 ∗ 𝐴. On
the other hand, by Federer’s theorem (Theorem 8.21) and (10.8), we have
for H 𝑛−1 -a.e. 𝑧 ∈ (𝐸 \ 𝐴) (0) ∩ 𝜕 ∗ 𝐴, 𝑧 ∈ 𝐸 (1) ∪ 𝐸 (0) . Again, 𝑧 ∈ 𝐸 (1) is not
possible since
𝜃 𝑛 (𝐸 ∪ (R𝑛 \ 𝐴)) + 𝜃 𝑛 (𝐸 \ 𝐴) = 𝜃 𝑛 (𝐸) + 𝜃 𝑛 (R𝑛 \ 𝐴).
62
Therefore (𝐸 \ 𝐴) (0) ∩ 𝜕 ∗ 𝐴 is H 𝑛−1 -equivalent to 𝐸 (0) ∩ 𝜕 ∗ 𝐴, which shows
that
(𝜇 𝐴 ⌞𝐺 (1) )⌞(𝐸 \ 𝐴) (0) = 𝜇 𝐴 ⌞(𝐺 (1) ∩ 𝐸 (0) ). (10.16)
From (10.14), (10.15), (10.16) we obtain
(𝐼 𝐼) = 𝜇𝐺∩𝐴 ⌞(𝐸 ∩ 𝐴) (0) = 𝜇𝐺 ⌞ 𝐴 + 𝜇 𝐴 ⌞(𝐺 (1) ∩ 𝐸 (0) ).
For (𝐼 𝐼 𝐼) we claim that it is vanishing.
Indeed, by (10.11), (10.14), we know that up to H 𝑛−1 -negligible sets,
𝜕 ∗ (𝐺 ∩ 𝐴) = ( 𝐴 ∩ 𝜕 ∗ 𝐺) ∪ (𝐺 (1) ∩ 𝜕 ∗ 𝐴),
𝜕 ∗ (𝐸 \ 𝐴) = (𝐸 (1) ∩ 𝜕 ∗ 𝐺) ∪ (𝜕 ∗ 𝐸 \ 𝐴).
By Federer’s theorem (Theorem 8.21) and (10.8), we find
H 𝑛−1 (𝜕 ∗ 𝐹 ∩ 𝜕 ∗ (𝐺 ∩ 𝐴) ∩ 𝜕 ∗ (𝐸 \ 𝐴)) = H 𝑛−1 (𝐹 (1/2) ∩ 𝐸 (1) ∩ 𝐺 (1) ∩ 𝜕 ∗ 𝐴).

Note also that 𝐹 (1/2) ∩ 𝐸 (1) ∩ 𝐺 (1) ∩ 𝜕 ∗ 𝐴 = ∅ since

𝜃 𝑛 (𝐹) =𝜃 𝑛 (𝐺 ∩ 𝐴) + 𝜃 𝑛 (𝐸 \ 𝐴) = 𝜃 𝑛 (𝐺) − 𝜃 𝑛 (𝐺 \ 𝐴) + 𝜃 𝑛 (𝐸) − 𝜃 𝑛 (𝐸 ∩ 𝐴)
≥𝜃 𝑛 (𝐺) + 𝜃 𝑛 (𝐸) − 2𝜃 𝑛 ( 𝐴).
This shows that
H 𝑛−1 (𝜕 ∗ 𝐹 ∩ 𝜕 ∗ (𝐺 ∩ 𝐴) ∩ 𝜕 ∗ (𝐸 \ 𝐴)) = 0,
and proves the claim.
In turn, by (𝐼), (𝐼 𝐼), (𝐼 𝐼 𝐼) we find using (10.10) that
𝜇 𝐹 = 𝜇 𝐸 ⌞(R𝑛 \ 𝐴) + 𝜇𝐺 ⌞ 𝐴 + 𝜇 𝐴 ⌞(𝐺 (1) ∩ 𝐸 (0) ) − 𝜇 𝐴 ⌞(𝐸 (1) ∩ 𝐺 (0) ).
(10.17)
Finally, by (10.17) we get
 
′ ′ 𝑛−1 ∗ (1) (0) (1) (0)
𝑃(𝐹; 𝐴 ) = 𝑃(𝐸; 𝐴 \ 𝐴) + 𝑃(𝐺; 𝐴) + H (𝜕 𝐴 ∩ (𝐺 ∩ 𝐸 ) ∪ (𝐸 ∩ 𝐺 ) ).

By (10.8) and note that H 𝑛−1 (R𝑛 \ (𝐸 (0) ∪ 𝐸 (1) ∪ 𝜕 ∗ 𝐸)) = 0 (thanks again
to Federer’s theorem), we get
H 𝑛−1 (𝜕 ∗ 𝐴 ∩ (𝐺 (1) ∩ 𝐸 (0) )) = H 𝑛−1 (𝜕 ∗ 𝐴 ∩ (𝐺 (1) \ 𝐸 (1) )),
and the similar equality for 𝐺 (0) ∩ 𝐸 (1) (by symmetry), which, combined
with (10.10), implies (10.9). □
As said, it is easy to see that, if 𝐸 is a local perimeter minimizer in 𝐴,
then 𝐸 is stationary for perimeter in any 𝐵𝑟 (𝑥) ∩ 𝐴 ⊂ 𝐵𝑟0 (𝑥) ∩ 𝐴 ⊂ 𝐴 in the
sense of Definition 9.7. In fact, 𝐸 is stable for perimeter in the considered
sets, in the sense that the second variation of perimeter is strictly positive
with respect to any initial velocity.
63
Exercise 10.7. Show the second variation formula for perimeter, by using
Proposition 9.1.

64
Appendix A. Covering theorem

Theorem A.1 (Besicovitch’s covering theorem). If 𝑛 ≥ 1, then there exists

a positive constant 𝜉 (𝑛) with the following property. If F is a family of
closed non-degenerate balls of R𝑛 , and either the set 𝐶 of the centers of the
balls in F is bounded or
sup{diam(𝐵) : 𝐵 ∈ F } < ∞,
then there exist F1 , . . . , F𝜉 (𝑛) subfamilies of F such that
(𝑖) Each family F𝑖 is pairwise disjoint (meaning that the balls in F𝑖 are
disjoint, for every fixed 𝑖) and at most countable;
Ð𝜉 (𝑛) Ð
(1) 𝐶 ⊂ 𝑖=1 𝐵∈F𝑖 𝐵.

Appendix B. Vector-valued Radon measures

B.1. Weak star convergence

Lemma B.1. Let {𝜇 ℎ } ℎ∈N be R𝑚 -valued Radon measures on R𝑛 .
(i) If 𝜇 ℎ → 𝜇 and |𝜇 ℎ | → 𝜈, then for every Borel set 𝐸 ⊂ R𝑛 ,
|𝜇|(𝐸) ≤ 𝜈(𝐸).

Furthermore, if 𝐸 is a bounded Borel set with 𝜈(𝜕𝐸) = 0, then

𝜇(𝐸) = lim 𝜇 ℎ (𝐸).
ℎ→∞

(ii) If 𝜇 ℎ → 𝜇, |𝜇 ℎ |(R𝑛 ) → |𝜇|(R𝑛 ), and |𝜇|(R𝑛 ) < ∞ then |𝜇| → |𝜇|.

Lemma B.2. For 𝜇 ℎ → 𝜇 and 𝑟 𝑘 → ∞ as 𝑘 → ∞, if
lim |𝜇 ℎ |(𝐵𝑟 𝑘 ) = |𝜇|(𝐵𝑟 𝑘 ), ∀𝑘 ∈ N
ℎ→∞

then
|𝜇 ℎ | → |𝜇|.

Lemma B.3 (Compactness of vector-valued Radon measures). If {𝜇 ℎ } ℎ∈N

are R𝑚 -valued Radon measures on R𝑛 , with
sup |𝜇 ℎ |(𝐾) < ∞, ∀𝐾 ⊂ R𝑛 compact,
ℎ∈N

then there exists a R𝑚 -valued Radon measure 𝜇 on R𝑛 and ℎ(𝑘) → ∞ as

𝑘 → ∞, such that 𝜇 ℎ(𝑘) → 𝜇 as vector-valued Radon measures.
65
B.2. Differentiation of Radon measures
Theorem B.4 (Lebesgue-Besicovitch differentiation theorem). If 𝜈 is a R𝑚
vector-valued Radon measure on R𝑛 , and 𝜇 is a Radon measure on R𝑛 , then
for 𝜇-a.e. 𝑥 ∈ R𝑛 there exists the limit
𝜈(𝐵𝑟 (𝑥))
𝐷 𝜇 𝜈(𝑥) = lim ∈ R𝑚 ,
𝑟↘0 𝜇(𝐵𝑟 (𝑥))

which defines a Borel vector field 𝐷 𝜇 𝜈 ∈ 𝐿 1𝑙𝑜𝑐 (R𝑛 , 𝜇; R𝑚 ), with the property
that
𝜈 = (𝐷 𝜇 𝜈)𝜇 + 𝜈 𝜇𝑠 on M (𝜇),
where 𝜈 𝑆𝜇 ⊥ 𝜇.

Appendix C. Rademacher’s theorem

Lemma C.1 (Lipschitz continuity and 𝑊 1,∞ ). For 𝑓 : R𝑛 → R𝑚 , the
following statements are equivalent:
(1) 𝑓 is locally Lipschitz continuous in R𝑛 .
1,∞
(2) 𝑓 ∈ 𝑊𝑙𝑜𝑐 (R𝑛 ), that is, 𝑓 admits a weak gradient
∇ 𝑓 ∈ 𝐿 ∞ (R𝑛 , R𝑚 ⊗ R𝑛 ).
Theorem C.2 (Rademacher’s theorem). If 𝑓 : R𝑛 → R𝑚 is a Lipschitz
function and 𝑥 is a Lebesgue point of the weak gradient ∇ 𝑓 , then 𝑓 is
differentiable at 𝑥, with
d 𝑓𝑥 [𝜏] = ∇ 𝑓 (𝑥) [𝜏], ∀𝜏 ∈ R𝑛 .
In particular, 𝑓 is differentiable L 𝑛 -a.e. on R𝑛 thanks to Lebesgue points
theorem.

Appendix D. Morse-Sard theorem

Theorem D.1 (Morse-Sard theorem). For 𝑢 ∈ 𝐶 ∞ (R𝑛 ) define the critical
set
C := {𝑥 ∈ R𝑛 : ∇𝑢(𝑥) = 0}.
Then L 𝑛 (C) = 0. Moreover, the level-set {𝑢 = 𝑡} is a smooth hypersurface
in R𝑛 for a.e. 𝑡 ∈ R.
Remark D.2. See also [3, Theorem 10.4] for a 𝐶 1 Sard-type theorem.

66
References
[1] Lawrence C. Evans and Ronald F. Gariepy, Measure theory and fine properties of
functions, Revised, Textbooks in Mathematics, CRC Press, Boca Raton, FL, 2015.
MR3409135
[2] Francesco Maggi, Sets of finite perimeter and geometric variational problems, Cam-
bridge Studies in Advanced Mathematics, vol. 135, Cambridge University Press, Cam-
bridge, 2012. An introduction to geometric measure theory. MR2976521
[3] Leon Simon, Lectures on geometric measure theory, Proceedings of the Centre for
Mathematical Analysis, Australian National University, vol. 3, Australian National
University, Centre for Mathematical Analysis, Canberra, 1983. MR756417

67