Pólya-Szegő principles in metric spaces

Francesco Nobili

PhD Course - Universitá di Pisa

Academic year 2024/2025, Fall/Winter

The goal of this PhD course is to build a general and flexible machinery of decreasing
rearrangements satisfying fine Pólya-Szegő principles covering a variety of different settings.
In particular, the set-up will be that of functions defined on metric measure spaces. This is
motivated by the growing interest around functional inequalities in different settings such as,
but not limited to, weighted Euclidean and Riemannian manifolds; metric spaces with synthetic
Ricci curvature bounds, sub-Riemannian structures etc.
In the first part of the course, we shall focus on a key tool to build up this machinery, which
is the theory of functions of bounded variation on metric measure spaces. We shall discuss the
notion of total variation in this setting and settle important properties of this functional space.
In the second part, we shall consider notions of decreasing rearrangements into open intervals
of the real line weighted by a positive weight and discuss their basic properties. Subsequently, we
investigate the effect of decreasing rearrangements on Dirichlet and BV energies relying on the
theory of functions of bounded variation. In particular, assuming an underlying isoperimetric
principle, we will derive the so-called Pólya-Szegő inequality in this generality.
Finally, we conclude by discussing relevant classes of examples where this theory applies
and/or comment on the rigidity of the Pólya-Szegő inequality.
Main references.

[1] L. Ambrosio and S. Di Marino, Equivalent definitions of BV space and of total vari-
ation on metric measure spaces, J. Funct. Anal., (2014).

[2] A. Cianchi and N. Fusco, Functions of bounded variation and rearrangements, Arch.
Ration. Mech. Anal., (2002).

[3] M. Miranda, Jr., Functions of bounded variation on “good” metric spaces, J. Math.
Pures Appl, (2003).

[4] F. Nobili and I. Y. Violo, Fine Pólya-Szegő rearrangement inequalities in metric spaces
and applications, arXiv:2409.14182, (2024).

Address: Universitá di Pisa, Dipartimento di Matematica, Largo Bruno Pontecorvo 5, 56127 Pisa, Italy
Email : [email protected]
Lecture 1 - [21/10/2024]

Notation
Let X be a set. Denote by 2X the power set, i.e. the sets of all subsets of X. An outer measure
is a set value map µ : 2X → [0, ∞] so that

µ(E) ≥ 0,
µ(∅) = 0,
µ(E) ≤ µ(F ),
X
µ(∪n En ) ≤ µ(En ),
n

for all E, F, En ∈ F with E ⊂ F and n ∈ N. The latter property is called σ-subadditivity or

countably sub-additivity.
A measurable space is a couple (X, F) where X is a set and F ⊂ 2X is a σ-algebra (or σ-field).
Recall that F is a σ-algebra provided X ∈ F and it holds

E ∈ F =⇒ E c ∈ F,
En ∈ F for n ∈ N =⇒ ∪n En ∈ F.

A measure µ on a measurable space (X, F) is a set function µ : F → [0, ∞] satisfying

µ(E) ≥ 0,
µ(∅) = 0,
X
µ(∪n En ) = µ(En ),
n

for all E, En ∈ F with Ei ∩ Ej = ∅ if i ̸= j and for all i, j, n ∈ N. The latter property is called
σ-additivity or countable additivity. A finite measure is a measure µ so that µ(X) < ∞. The
triple (X, F, µ) is called measure space.
A metric space is a couple (X, d) where X is a set and d : X × X → [0, ∞] is the distance
satisfying

d(x, y) ≥ 0,
d(x, y) = 0 if and only if x = y,
d(y, x) = d(x, y),
d(x, z) ≤ d(x, y) + d(y, z),

for all x, y, z ∈ X. Recall that on a metric space there is a canonical choice of a σ-algebra,
that is called Borel σ-algebra and will be denoted by B. A Borel measure on a metric space is,
therefore, a measure in the previous sense with respect to the Borel σ-algebra. Given A, B ⊂ X,
we denote
dist(A, B) := inf{d(x, y) : x ∈ A, y ∈ B}.

1
A metric space is called complete, provided Cauchy-sequences converge, and it is called separable,
provided there exists a countable set of points (xn ) ⊂ X that is dense in X. Furthermore, it is
called proper, provided bounded closed sets are compact.
A metric measure space is a triple (X, d, m) where

(X, d) is a complete and separable metric space,

m ̸= 0 is boundedly finite Borel measure.

Here, m ̸= 0 means that m(X) ̸= 0, while boundedly finite means that m(B) < ∞ if B ⊂ X is
Borel and bounded (for instance, a metric ball).
If E ⊂ X is Borel and µ is a Borel measure, we denote

µ E(B) := µ(E ∩ B), ∀ B ⊂ X Borel.

Analogously, if u : Ω → R is a Borel function and U ⊂ Ω is a set, we denote u|U the restricted

function on the open set U defined by u|U (x) = u(x) for all x ∈ U .
If ∅ ̸= Ω ⊂ X is open, we denote by Lp (Ω), Lploc (Ω) the equivalence class up to m Ω-a.e.
equality of Borel functions in Ω which are respectively p-integrable in Ω and p-integrable on a
neighborhood of each point in Ω. We shall sometimes also write Lp (m) to stress the measure we
are considering. Also, for u ∈ L1loc (Ω), we recall the notion of essential supremum and essential
infimum, as defined by

esssup u := inf{M ∈ R : u ≤ M m-a.e. on Ω},

essinf u := sup{m ∈ R : u ≥ m m-a.e. on Ω}.

Analogously, we set C(Ω), Cb (Ω), Lip(Ω), Lipbs (Ω), Liploc (Ω) respectively the set of continuous
functions on Ω, the set of continuous and bounded functions on Ω, the set of Lipschitz functions
on Ω, the set of Lipschitz functions having support bounded and contained in Ω and, lastly, the
set of functions which are Lipschitz on a neighborhood of each point in Ω. Recall that u : Ω → R
is called Lipschitz, provided there exist a constant L ∈ (0, ∞) so that

|u(x) − u(y)| ≤ Ld(x, y), ∀x, y ∈ Ω.

In this case, we also say that u is L-Lipschitz, and the smallest positive L for the above to hold
will be denoted by Lip(u).

The total variation

Given a metric measure space (X, d, m) and an open set Ω ⊂ X, let us define the notion of local
Lipschitz constant of u : Ω → R, defined as

|u(x) − u(y)|
lip(u)(x) := lim , ∀ x ∈ Ω,
y→x d(x, y)

where the limit superior is canonically taken to be zero if x is isolated.

Exercise 1.1. Prove that the local Lipschitz constant satisfies the following calculus rules. If
u, v ∈ Liploc (Ω), α, β, c ∈ R and f ∈ Lip(R) then it holds pointwise on Ω:

• Vector space structure.

lip(αu + βu) ≤ |α|lip(u) + |β|lip(v).

2
 • Chain rule.
lip(f ◦ u) ≤ Lip(f )lip(u).

• Leibniz rule.
lip(uv) ≤ |u|lip(v) + |v|lip(u).

• Locality on open sets. if u ≡ c on Ω, then

lip(u) = 0.

• Convexity. for every φ ∈ Lip(Ω) with 0 ≤ φ ≤ 1 we have

lip(φu + (1 − φ)v) ≤ φlip(u) + (1 − φ)lip(v) + |u − v|lip(φ). (1.1)

(Hint)

Let us start by considering the following notion of total variation on open sets.

Definition 1.2 (Total variation). Let (X, d, m) be a metric measure space, let ∅ ̸= Ω ⊂ X be
open and consider u ∈ L1 (Ω). For every U ⊂ Ω open, we define
ˆ
|Du|(U ) := inf lim lip(un ) dm,
n↑∞ U

where the infimum is taken among all (un ) ⊂ Liploc (U ) satisfying un → u in L1 (U ) and it is
canonically set to +∞ if no such sequence exist and to 0 if U = ∅.

Notation. We have just defined |Du|(·) even though we did not speak either of a
distributional derivative Du, nor of a total variation | · | in any sense. The main
task from here is to justify this notation by showing that it shares analogous
properties to classical notions of total variations.

Remark 1.3. The above notion of total variation is well defined for all u ∈ L1 (Ω) and takes
possibly value +∞ by definition. However, if |Du|(Ω) < ∞, then also |Du|(U ) < ∞ for any
U ⊂ Ω is open. Indeed, considering (un ) ⊂ Liploc (Ω) in the definition of |Du|(Ω), we clearly get
for every U ⊂ Ω open and setting vn := un |U :

• vn ∈ Liploc (U );

• vn → u in L1 (U );
´ ´
• limn↑∞ U lip(vn ) dm ≤ limn↑∞ Ω lip(un ) dm < ∞.

Thus, (vn ) is a competitor

´ for the definition of |Du|(U ) and the above estimate implies that
|Du|(U ) ≤ limn→∞ Ω lip(un ) dm < ∞. Finally, by arbitrariness of (un ), we have deduce

|Du|(U ) ≤ |Du|(Ω). (1.2)

3
Remark 1.4 (Total variation for locally integable functions). On can similarly define |Du|(U )
for u ∈ L1loc (Ω) and for every U ⊂ Ω open, by setting
ˆ
|Du|(U ) := inf lim lip(un ) dm,
n↑∞ U

where, this time, the infimum is taken among all (un ) ⊂ Liploc (U ) satisfying un → u in L1loc (U ).
By L1loc -convergence, recall that we mean that for every point x ∈ U there exists an open set
B ∋ x so that un → u in L1 (B). Instead, if (X, d) is also assumed proper (i.e. bounded sets
are compact), then this is the usual notion of convergence in L1 on compact sets. To keep the
course lighter, we stick to L1 -integrable functions in this note. ■

Lecture 2 - [28/11/2024]

We now start discussing basic properties around the notion of total variation.

Exercise 2.5. Prove that the total variation satisfies the following rules: given u, v ∈ L1 (Ω)
and α, β, c ∈ R, it holds:

• Vector space structure.

|D(αu + βv)|(Ω) ≤ |α||Du|(Ω) + |β||Dv|(Ω).

• Locality on open sets. If u = c m-a.e. on Ω, then

|Du|(Ω) = 0.

In particular, |D(αu + c)|(Ω) = |α||Du|(Ω) and, if u = v m-a.e. on Ω, then |Du|(Ω) = |Dv|(Ω).

We next show that an optimal sequence always exists realizing the value |Du|(Ω).

Lemma 2.6. Let (X, d, m) be a metric measure space, let ∅ ̸= Ω ⊂ X be open and let u ∈ L1 (Ω).
If |Du|(Ω) < ∞, then there is (un ) ⊂ Liploc (Ω) so that un → u in L1 (Ω) and
ˆ
lip(un ) dm → |Du|(Ω), as n ↑ ∞.
Ω

Proof. Since |Du|(Ω) < ∞, we have for all k ∈ N a sequence (uk,n )n ⊂ Liploc (Ω) so that
ˆ
1 1
uk,n → u in L (Ω) as n ↑ ∞, lim lip(uk,n ) dm ≤ |Du|(Ω) + .
n↑∞ Ω k

Therefore, there is n(k) ∈ N so that ∥uk,n − u∥L1 (Ω) ≤ k1 for all n ≥ n(k). By definition of limit
inferior and possibly increasing n(k), we can require also n(k) > k and the validity of
ˆ
3
|Du|(Ω) ≤ lip(uk,n(k) ) dm ≤ |Du|(Ω) + .
Ω k

The sequence (uk,n(k) ) does the job.

4
 Next, we show a lower semicontinuity result for the total variation.
Lemma 2.7 (Lower semicontinuity). Let (X, d, m) be a metric measure space, let ∅ ̸= Ω ⊂ X
be open and let (un ) ⊂ L1 (Ω) be so that un → u in L1 (Ω). Then, it holds

|Du|(Ω) ≤ lim |Dun |(Ω).

n↑∞

Proof. If limn↑∞ |Dun |(Ω) = ∞, there is nothing to prove. Let us then suppose it to be finite
and, up to pass to a not relabeled subsequence, let us suppose that |Dun |(Ω) < ∞ for all n ∈ N
and that the limit inferior is a limit. By Lemma 2.6, for every ´ n ∈ N consider an optimal
sequence (un,k ) ⊂ Liploc (Ω) satisfying un,k → un in L1 (Ω) and Ω lip(un,k ) dm → |Dun |(Ω) as
k → ∞. A diagonalization argument gives the conclusion.

The following is a useful property to join locally Lipschitz functions (also called fundamental
estimate in the literature of the calculus of variations). This will be used next to join sequences
arising from the definition of the total variation.
Lemma 2.8 (Joint property). Let (X, d, m) be a metric measure space and M, N ⊂ X be two
open sets so that dist(M \ N, N \ M ) > 0. Let u ∈ Liploc (M ), v ∈ Liploc (N ). Then, there is
w ∈ Liploc (M ∪ N ) satisfying

w = u, on a neighborhood of M \ N, w = v, on a neighborhood of N \ M, (2.3)

and, for some C(M, N ) > 0 and some open set H ⊂ M ∩ N depending only on M, N , we have
ˆ ˆ ˆ ˆ
lip(w) dm ≤ lip(u) dm + lip(v) dm + C(M, N ) |u − v| dm. (2.4)
M ∪N M N H

Furthermore, for every σ ∈ L1 (M ∪ N ) it holds

ˆ ˆ ˆ
|w − σ| dm ≤ |u − σ| dm + |v − σ| dm. (2.5)
M ∪N M N

Proof. Set η := dist(M \ N, N \ M ) and recall that η > 0 by assumption. We shall suppose that
M ∩ N ̸= ∅, otherwise the claims become trivial choosing w = u on M , w = v on N . First,
consider the 3/η-Lipschitz function
3 n η + η o
φ(x) := min d(x, N \ M ) − , , ∀x ∈ X.
η 3 3
Notice that φ ≡ 1 on a neighborhood of M \ N , φ ≡ 0 on a neighborhood of N \ M and
0 ≤ φ ≤ 1. Let us define
w := φu + (1 − φ)v,
and let us rewrite here the convexity inequality estimate (1.1):

lip(w) ≤ φlip(u) + (1 − φ)lip(v) + |u − v|lip(φ).

The above, recalling that lip(φ) ≤ 3/η, in turn guarantees that w ∈ Liploc (M ∪ N ). A direct
check then gives (2.3).
We now prove (2.4). To this aim, we define H := {0 < φ < 1} ∩ (M ∪ N ) which is open and
contained in M ∩ N , and we estimate
ˆ ˆ


lip(w) dm = χInt({φ=1}) + χInt({φ=0}) + χH lip(w) dm
M ∪N ˆM ∪N ˆ ˆ
3
≤ lip(u) dm + lip(v) dm + |u − v| dm,
M N η H

5
where, in the last inequality, we used the convexity estimate for lip(w) and the facts


χM ∪N χInt({φ=1}) + χInt({φ=0}) + χH φlip(u) ≤ χM lip(u),


χM ∪N χInt({φ=1}) + χInt({φ=0}) + χH (1 − φ)lip(v) ≤ χN lip(v),


χM ∪N χInt({φ=1}) + χInt({φ=0}) + χH lip(φ)|u − v| ≤ 3/η χH |u − v|,

thanks to the very definition of φ, H and recalling that lip(φ) = 0 on Int({φ = 1}) and on
Int({φ = 0}) by locality on open sets. In particular, we get (2.4) with C(M, N ) := 3/η. Finally,
property (2.5) follows by convexity and since φ ≤ χM and 1 − φ ≤ χN .

Next, we show the sub-additivity and monotonicity properties of the total variation relying
on the above joint property.

Lemma 2.9. Let (X, d, m) be a metric measure space, let ∅ ̸= Ω ⊂ X be open and let u ∈ L1 (Ω).
Consider two open sets U, W ⊂ Ω. It holds:

i) |Du|(U ) ≤ |Du|(W ) if U ⊂ W ;

ii) |Du|(U ∪ W ) ≤ |Du|(U ) + |Du|(W );

iii) |Du|(U ∪ W ) = |Du|(U ) + |Du|(W ) if dist(U, W ) > 0.

iv) if Un ⊂ Ω for n = {0, 1, 2, ...} are open and increasing Un ↑ U , then

|Du|(U ) = lim |Du|(Un ) = sup |Du|(Un ).

n↑∞ n

Proof. First, we prove i). If |Du|(W ) = +∞, there is nothing to prove, otherwise, this was
already obtained in (1.2).
We now prove iii) and we shall only consider the case when |Du|(U ∪W ), |Du|(U ), |Du|(W ) <
∞, otherwise there is nothing to prove (thanks to i)). For the ≥ inequality, it suffices to consider
an optimal sequence (un ) ⊂ Liploc (U ∪ W ) for |Du|(U ∪ W ) by Lemma 2.6, ´ notice that the
restrictions
´ of un to U, W are competitors
´ for |Du|(U ), |Du|, and use lim n→∞ U ∪W lip(un ) dm ≥
limn→∞ U lip(un |U ) dm + limn→∞ W lip(un |W ) dm to conclude (here, the only relevant fact is
that U ∩ W = ∅). Conversely, to show the ≤ inequality, consider optimal sequences (un ) ⊂
Liploc (U ) and (vn ) ⊂ Liploc (W ) given by Lemma 2.6 respectively for |Du|(U ) and |Du|(W ).
Here we use that the sets U, W are well separated, i.e. dist(U, W ) > 0. If we set wn equal to
un in U and equal to vn in W , we have (wn ) ⊂ Liploc (U ∪ W ), wn → u in L1 (U ∪ W ) and
χU ∪W lip(wn ) ≤ χU lip(un ) + χW lip(vn ). Integrating and sending n to infinity, we complete the
proof of iii).
We now prove iv) using iteratively the joint Lemma 2.8. We also make use of both the
conclusions i),iii), now available. First, since Un ⊂ U , it is enough to show (thanks to i)) that
|Du|(U ) ≤ supn |Du|(Un ) < ∞, being all the other possibilities trivial. Also, notice that by
replacing Un with {x ∈ Un : dist(x, Unc ) > 1/n}, it is not restrictive to further suppose that
dist(Un , Ω \ Un+1 ) > 0, for all n ∈ N. For every k ∈ N, set
(
U2 , if k = 1,
Sk :=
Uk \ U k−2 if k ≥ 2,

and notice that, if 2 ≤ n < n + 3 ≤ m, then by construction

dist(Sn , Sm ) = dist(Un \ Un−2 , Um \ Um−2 ) ≥ dist(Un , Ω \ Un+1 ) > 0.

6
Thus, the families (S3k ), (S3k+1 ), (S2k+2 ) have well separated elements in Ω and we have
∞ n
X X by iii)
|Du|(S3k+j ) = lim |Du|(S3k+j ) = lim |Du|(∪k≤n S3k+j )
n→∞ n→∞
k=1 k=1
by i)
≤ lim |Du|(U3n+j ) ≤ sup |Du|(Un ) < ∞,
n→∞ n∈N
P∞
for all j = {0, 1, 2}. It follows that k=1 |Du|(S3k ) < ∞, hence for ε > 0 there is k̄ ∈ N so that
X
|Du|(Sk ) ≤ ε. (2.6)
k≥k̄

We see now that if we are able to prove the existence of (um ) ⊂ Liploc (U ) so that
ˆ
1
um → u in L (U ), and |Du|(Uk̄ ) + 3ε ≥ lim lip(um ) dm, (2.7)
m↑∞ U

then conclusion iv) will follow, by arbitrariness of ε and definition of |Du|(U ). Fix now m ∈ N
and, for every h ∈ N, define the sets Dh := Sh+k̄ , Bh := Uh+k̄ if h ≥ 1 while D0 = B0 = Uk̄ . For
every h ≥ 0, and by Lemma 2.6 there is (ψk,h )k ⊂ Liploc (Dh ) and a large enough integer k(h)
so that ˆ
1
lip(ψk,h ) dm ≤ |Du|(Dh ) + , ∀k ≥ k(h). (2.8)
Dh m2h
Since ψk,h → u in L1 (Dh ), possibly increasing k(h) > h depending only on h, we can also require
ˆ
1
|ψk(h),h − u| dm ≤ h
. (2.9)
Dh m2

Moreover, possibly enlarging again k(h) we can guarantee that

ˆ ˆ
ε ε
ch |ψk(h),h − u| dm ≤ h
, ch |ψk(h+1),h+1 − u| dm ≤ , (2.10)
Hh 2·2 Hh 2 · 2h

where ch and Hh are respectively the constant and the open set appearing in Lemma 2.8 with
sets M = Bh , N = Dh+1 . Notice in fact that dist(Dh+1 \ Bh , Bh \ Dh+1 ) > 0 and we also used
that H h ⊂ Bh ∩ Dh+1 which in turn is contained in Dh .
For any h ≥ 0, we then define um,h ∈ Liploc (Bh ) iteratively setting um,0 := ψk(0),0 and then,
given um,h , we build um,h+1 joining um,h with ψk(h+1),h+1) thanks to Lemma 2.8, thus satisfying

um,h+1 = um,h , on Bh−1 , um,h+1 = ψk(h+1),h+1 on Bh+1 \ B h ,

(recall property (2.3)). We claim that

ˆ  
1 1
|um,h+1 − u| dm ≤ 1 − h+1 , (2.11)
Bh +1 m 2
ˆ ˆ ˆ
ε
lip(um,h+1 ) dm ≤ lip(um,h ) dm + lip(ψk(h+1),h+1) dm + . (2.12)
Bh+1 Bh Dh+1 2h

Indeed, the latter follows by (2.4) estimating the last term

ˆ ˆ (2.10) ε
ch |ψk(h+1),h+1 − um,h | = ch |ψk(h+1),h+1 − ψk(h),h | dm ≤ ,
Hh Hh 2h

7
since H h ⊂ Bh \ B h−1 and um,h = ψk(h),h on Bh \ B h−1 by strong induction assumption. While,
the former, follows by choosing σ = u in (2.5) and by an induction argument
ˆ ˆ ˆ
|um,h+1 − u| dm ≤ |ψk(h+1),h+1 − u| dm + |um,h − u| dm,
Bh+1 Dh+1 Bh
(2.9)
   
1 1 1 1 1
≤ + 1− h = 1 − h+1 .
m2h+1 m 2 m 2
Here comes the conclusion of iv). For every m ∈ N, we can set
um (x) := um,h (x), if x ∈ Bh−1 .
By construction, this is well defined on U and (um ) ⊂ Liploc (U ). Moreover, ∥u−um ∥L1 (U ) ≤ 1/m
by (2.11) and monotone convergence theorem. By iteration, we also get
ˆ ˆ ˆ
lip(um ) dm = lim lip(um,h+1 ) dm ≤ lim lip(um,h+1 ) dm
U h↑∞ Bh h↑∞ Bh+1
∞
(2.12),(2.8) X ∞
X 1 1
≤ |Du|(Dh ) + ( + ε) h
m 2
h=0 h=0
X 2 (2.6) 2
= |Du|(Uk̄ ) + |Du|(Sk ) + + 2ε ≤ |Du|(Uk̄ ) + + 3ε.
m m
k≥k̄

By sending m ↑ ∞, we thus get (2.7) in turn implying property iv).

We now prove ii). Again, we suppose that |Du|(U ) + |Du|(W ) < ∞, otherwise there is
nothing to prove. Notice that it is sufficient to show
|Du|(Ur ∪ Wr ) ≤ |Du|(U ) + |Du|(W ), ∀r > 0.
where Ur := {x ∈ U : dist(x, U c ) > r}, Wn := {x ∈ W : dist(x, W c )
> r}. Indeed, as Ur ∪ Wr ↑
U ∪ W as r ↓ 0 along a chosen sequence, property iv) would then give the conclusion. Thus, to
conclude, we need to show the above estimate for a given r > 0. Consider (un ) ⊂ Liploc (U ) and
(vn ) ⊂ Liploc (W ) as given by Lemma 2.6 respectively for |Du|(U ) and |Du|(W ). Consider the
sets M = (Ur ∪ Wr ) ∩ U, N = (Ur ∪ Wr ) ∩ W satisfying dist(M \ N, N \ M ) > 0 by construction.
Notice also that M ∪ N = Ur ∪ Wr . We can thus invoke Lemma 2.8 with un , vn to get a joint
sequence (wn ) ⊂ Liploc (Ur ∪ Wr ) satisfying
ˆ (2.4)
ˆ ˆ ˆ
lip(wn ) dm ≤ lip(un ) dm + lip(vn ) dm + C(M, N ) |un − vn | dm,
Ur ∪Wr U W U ∩W
ˆ (2.5)
ˆ ˆ
|wn − u| ≤ |un − u| dm + |vn − u| dm,
Ur ∪Wr U W

having used the inclusions M ⊂ U, N ⊂ W and H n ⊂ U ∩ W . Finally, since un , vn → u

respectively in L1 (U ) and L1 (W ), we also get un − vn → 0 in L1 (U ∩ W ) and wn → u in
L1 (Ur ∪ Wr ) so that sending n → ∞ in the first of the above concludes the proof.

Lecture 3 - [11/12/2024]

Next, we shall extend the total variation in the sense of Carathéodory from open sets to the
whole Borel σ-algebra (recall B is the collection of Borel sets in Ω and 2Ω is the power sets).

8
Theorem 3.10. Let (X, d, m) be a metric measure space, let ∅ ̸= Ω ⊂ X be open and let
u ∈ L1 (Ω). Then 
B ∋ B 7→ inf |Du|(U ) : B ⊂ U, U ⊂ Ω open ,
is a Borel measure.

Proof. Denote by A the class of open subsets of Ω and consider the set value map A ∋ U 7→
α(U ) := |Du|(U ) ∈ [0, +∞]. Let us extend α on 2Ω by setting

2Ω ∋ B 7→ α∗ (B) := inf{α(U ) : B ⊂ U, U ∈ A} ∈ [0, ∞].

We shall use without notice that α∗ (U ) = α(U ) if U ∈ A, and that α∗ (B) ≤ α∗ (B ′ ) if B ⊂ B ′ .

Both properties simply follow from the definition. The conclusion of the proof will be achieved
if we can prove that the restriction of α∗ to B is a Borel measure. We proceed in different steps.
α is countably sub-additive on A. Let (Un ) ⊂ A be a collection of open subsets in Ω and
set U := ∪n Un . For every N ∈ N, consider AN := ∪n≤N Un . Then, by construction, AN ∈ A
and AN ↑ U as N ↑ ∞. Then, by iv) in Lemma 2.9, we know that α(U ) = supN α(AN ). In
particular, for every ϵ > 0, there is N0 ∈ N so that, for all N > N0 it holds
X X
α(U ) − ε ≤ α(AN ) ≤ α(Un ) ≤ α(Un ),
n≤N n∈N

having used, in the central inequality, that α is sub-additive for finite unions (cf. ii) Lemma 2.9).
Sending ε ↓ 0, we get that α is countably sub-additive on A.
α∗ is an outer measure. Since α∗ (∅) = 0 and α∗ is monotone, we only need to show that
α∗ is countably sub-additive on 2Ω . Let then (Bn ) ⊂ 2Ω and set B := ∪n Bn . If α∗ (U ) = +∞,
there is nothing to do, so let us assumeP it to be finite. In this case, let ε > 0 and (ϵn ) be a
∗
sequence of positive numbers so that n εn ≤ ε. By definition of α (Bn ) (and since it is finite
by monotonicity), for every n ∈ N, there is Un ∈ A with Bn ⊂ Un so that α(Un ) ≤ α∗ (Bn ) + ϵn .
Therefore, we know B ⊂ U := ∪n Un and also U ∈ A is a competitor for the definition of α∗ (U )
making possible to estimate:
X X X
α∗ (B) ≤ α(U ) ≤ α(Un ) ≤ α∗ (Bn ) + εn ≤ α∗ (Bn ) + ε,
n n n

where, in the second inequality, we used that α is countably subbadditive on A, as proved in

the previous step. Sending ε ↓ 0, we get that α∗ is an outer measure.
Conclusion. Here we conclude the proof by showing that α∗ is a measure on B. This is a
standard application of Carathéodory’s extension principles for outer measures. Indeed, denote
by M ⊂ 2Ω the collection of α∗ -measurable sets, that is E ∈ M if α∗ (B) ≥ α∗ (E∩B)+α∗ (B\E),
for all B ∈ 2Ω . Recall that M is a σ-algebra (in fact, it is the smallest σ-algebra where α∗ is also
countably additive). Therefore, the proof will be concluded if we are able to show that B ⊂ M,
so that the restriction of α∗ on B will be a measure. In particular, it is enough to show that
A ⊂ M, that is M contains the open subsets of Ω (simply because M is a σ-algebra and B the
smallest one containing open sets). To this aim, we need to verify that if U ∈ A then it holds

α∗ (B) ≥ α∗ (U ∩ B) + α∗ (B \ U ), ∀B ∈ 2Ω .

Clearly, we only need to check the above when α∗ (B) < ∞. We argue by contradiction and
suppose that there are U ∈ A and B ∈ 2Ω so that the -strict- converse inequality holds. By
definition of α∗ (B) (and since it is finite), we can find W ∈ A with B ⊂ W so that

α(W ) < α∗ (U ∩ B) + α∗ (B \ U ) ≤ α(U ∩ W ) + α∗ (W \ U ).

9
Now, since U ∩ W ∈ A and since D(·) := dist(·, (U ∩ W )c ) is continuous, we infer the existence
of sets Cn := {x ∈ U ∩ W : D(x) > 1/n} so that Cn ∈ A for all n ∈ N and Cn ↑ U ∩ W as n ↑ ∞.
In particular, by iv) in Lemma 2.9 there exists k ∈ N large enough so that

α(W ) < α(Ck ) + α∗ (W \ U ).

From the fact Ck ⊂ U ∩ W we deduce W \ U ⊂ W \ Ck+1 ∈ A. Also, since Ck ⊂ W , then

Ck ∪ (W \ Ck+1 ) ⊂ W . Finally, by construction, we have dist(Ck , W \ Ck+1 ) > 0. Therefore, by
iii) in Lemma 2.9 and monotonicity, we can estimate

α(Ck ) + α∗ (W \ U ) ≤ α(Ck ) + α(W \ Ck+1 ) = α(Ck ∪ W \ Ck+1 ) ≤ α(W ).

This leads to a contradiction hence concluding the proof.

Notation. From now on, we will adopt a standard abuse of notation and always
denote by |Du| the Borel measure obtained in Theorem 3.10.

Functions of bounded variation and sets of finite perimeter

Next, we give the definition of functions of bounded variation.
Definition 3.11 (The space BV (Ω)). Let (X, d, m) be a metric measure space, let ∅ ̸= Ω ⊂ X
be open and consider u ∈ L1 (Ω). We say that u is a function of bounded variation on Ω, writing
u ∈ BV (Ω), provided there exist (un ) ⊂ Liploc (Ω) so that un → u in L1 (Ω) and
ˆ
sup lip(un ) dm < ∞.
n∈N Ω

Notice that, thanks to Theorem 3.10, we know that when u ∈ BV (Ω), then |Du| is a finite
Borel measure. As expected, a relevant class of functions of bounded variation are integrable
Lipschitz functions.
Example 3.12. It holds that u ∈ Liploc (Ω) with u, lip(u) ∈ L1 (Ω) is so that u ∈ BV (Ω) and

|Du| ≤ lip(u) m, (3.13)

as Borel measures on Ω. Indeed, directly taking u|U ∈ Liploc (U ) as a competitor constant

sequence in the definition of |Du|(U ) for any U ⊂ Ω open, we deduce
ˆ
|Du|(U ) ≤ lip(u) dm < ∞.
U

In particular, by choosing U = Ω we get that u ∈ BV (Ω) and, more generally, by arbitrariness

of U open and since |Du| is a finite Borel measure thanks to Theorem 3.10, then (3.13) also
follows.
Finally, notice that many distance-like functions can be built satisfying the above, for in-
stance, for any B ⊂ X bounded, we can consider for any c > 0 the cut-off function

η(x) := (1 − c · d(x, B))+ , ∀x ∈ X.

Since supp(η) is bounded and η ∈ L∞ , then u ∈ L1 (X). Thus, since clearly Lip(η) ≤ 1/c,
we have that u ∈ BV (X). This construction (and several similar ones) shows that there is
abundance of BV functions in every metric measure space. ■

10
 The following exercise shows that the space BV can be trivialised if the metric measure
structure is not sufficiently rich.

Exercise 3.13. Consider any complete and separable metric space (X, d). Find m so that

L1 (m) = BV (X) ?

(Spoiler)

Next, we show the existence of an optimal sequence approximating |Du| in the sense of mea-
sures. This is sometimes called BV-strict convergence and is the best “smooth” approximation
we can in get in this generality.

Lemma 3.14. Let (X, d, m) be a metric measure space, let ∅ ̸= Ω ⊂ X be open and let u ∈
BV (Ω). Then, there exists (un ) ⊂ Liploc (Ω) so that
ˆ ˆ
1
un → u in L (Ω), φlip(un ) dm → φ d|Du|, ∀φ ∈ Cb (Ω).

Proof. Consider the optimal sequence (un ) ⊂ Liploc (Ω) given by Lemma 2.6 for |Du|(Ω). Since
un → u in L1 (Ω), we only need to show the last claim. Take thus any φ ∈ Cb (Ω) and, possibly
splitting the positive and negative part, it is not restrictive to suppose φ nonnegative. Consider
the finite positive Borel measures µn := lip(un )m for n ∈ N and let us estimate
ˆ ˆ ∞ ˆ ∞
lim φlip(un ) dm = lim µn ({φ > t}) dt ≥ lim µn ({φ > t}) dt
n→∞ n→∞ 0 0 n→∞
ˆ ∞ ˆ
≥ |Du|({φ > t}) dt = φ d|Du|,
0

having used, Cavalieris’ formula in the first and last step, in the second step Fatou’s lemma,
and in the third step the fact that the set U := {φ > t} is open and that (un |U ) is a competitor
sequence in the definition of the quantity |Du|({φ > t}). Now, repeating the arguments with
ψ := sup φ − φ, we get instead
ˆ ˆ
lim −φlip(un ) dm ≥ − φ d|Du|,
n→∞

concluding the proof by arbitrariness of φ.

The last conclusion is typically referred to as a narrow/weak convergence of finite measures,

shortly written lip(un )m ⇀ |Du| in duality with Cb (Ω). We next show a basic chain rule result.

Lemma 3.15. Let (X, d, m) be a metric measure space, let ∅ ̸= Ω ⊂ X be open and let u ∈
BV (Ω). If f ∈ Lip(R) (with f (0) = 0 if m(Ω) = ∞), then f ◦ u ∈ BV (Ω) and

|D(f ◦ u)| ≤ Lip(f )|Du|,

holds as measure on Ω.

11
Proof. We claim first that f ◦ u ∈ BV (Ω). By assumptions, f ◦ u ∈ L1 (Ω) since
ˆ ˆ ˆ ˆ
|f ◦ u(x)| dm ≤ |f ◦ u − f (0)| dm + |f (0)| dm ≤ Lip(f ) |u| dm + f (0)m(Ω) < ∞,
Ω Ω Ω Ω

where, evidently, the latter term is not present if m(Ω) = ∞. Next, by Lemma 3.14, we can
consider (un ) ⊂ Liploc (Ω) so that un → u in L1 (Ω) and lipun m ⇀ |Du| in duality with Cb (Ω) as
n ↑ ∞. In particular, using that f is Lipschitz, we get that f ◦ un → f ◦ u in L1 (Ω) and
ˆ
lim lip(f ◦ un ) dm ≤ Lip(f )|Du|(Ω) < ∞,
n↑∞

having used the basic Leibniz rule for the local Lipschitz constant. This shows the claim.
Finally, we show the last conclusion and take any nonnegative φ ∈ Cb (Ω). We can estimate
ˆ ˆ ∞ ˆ
φ d|D(f ◦ u)| = |D(f ◦ u)|({φ > t}) dt ≤ lim |D(f ◦ un )|({φ > t}) dt
0 n↑∞
ˆ ˆ
= lim φ d|D(f ◦ un )| ≤ φlip(f ◦ un ) dm
n↑∞
ˆ ˆ
≤ Lip(f ) lim φlip(un ) dm = Lip(f ) φ d|Du|,
n↑∞

having used, in the first and third steps Cavalieri’s formula, in the second step Lemma 2.7
since {φ > t} is open and Fatou’s lemma, in the fourth step (3.13) since f ◦ un ∈ Liploc (Ω) by
assumptions, and in the last equality the convergence lip(un )m ⇀ |Du|. Being φ arbitrary and
nonnegative, the conclusion follows.

An important class of functions of bounded variation is the case of characteristic functions.

Recall that, given a set E ⊂ X with m(E ∩ Ω) < ∞, then χE ∈ L1 (Ω) and therefore we can
consider writing |DχE |(Ω), having possibly infinite value.

Definition 3.16. Let (X, d, m) be a metric measure space, let ∅ ̸= Ω ⊂ X be open and let E ⊂ X
be Borel. We say that E is a set of finite perimeter in Ω, provided

m(E ∩ Ω) < ∞, |DχE |(Ω) < ∞.

Notation. If E is Borel with m(E ∩ Ω) < ∞, then we denote

Per(E, ·) := |DχE |(·),

and call Per(E, ·) the perimeter measure of E in Ω (c.f. Theorem 3.10). The
same notation is adopted even if E is defined only up to m Ω-a.e. equality,
being Per(E, ·) unaffected.
Observe that, if E ⊂ Ω ⊂ Ω′ and χE ∈ BV (Ω), then Per(E, ·) is only known to
be a finite measure on Borel subsets of Ω. In general, we do not know whether
χE ∈ BV (Ω′ ) as well (and this heavily depends on ∂Ω). However, in most parts
of this note Ω will be fixed and no confusion will arise.
Finally, we set Per(∅, Ω) = 0 = Per(Ω, Ω) = Per(X, Ω) and Per(E) := Per(E, X).

12
Remark 3.17. If E is a set of finite perimeter in Ω, then χE ∈ BV (Ω) and, therefore Theorem
3.10 applies to give that |DχE | is a finite Borel measure in Ω.
Notice also that |DχE | does not depend on the Borel representative of E, that is if F is
so that m(E△F ) = 0, then Per(F, ·) = Per(E, ·) as measures. Here, E△F is the symmetric
difference of two sets defined by E△F := (E \ F ) ∪ (F \ E). ■

A first straightforward observation is that the perimeter measure is concentrated on the

topological boundary of the set.

Exercise 3.18. Given a set of finite perimeter E in Ω, prove

Per(E, B) = Per(E, B ∩ ∂E), ∀B ⊂ Ω Borel. (3.14)

(Spoiler)

Let us face a Borel regularity property that will be needed in the forthcoming theorem.

Lemma 3.19. Let (X, d, m) be a metric measure space, let ∅ ̸= Ω ⊂ X be open and let u ∈
L1 (Ω) be satisfying m({u > t}) < ∞ for all t > essinf u. For every U ⊂ Ω open, the map
t 7→ Per({u > t}, U ) is L1 -measurable.

Proof. For a given U ⊂ Ω open, let us write

R ∋ t 7→ F (t) := Per({u > t}, U ) ∈ [0, ∞].

We need to show that F (·) is L1 -measurable. First, notice that R ∋ t 7→ m({u > t}) is
monotone non-increasing and finitely valued in (essinf u, +∞), hence it has at most countably
many discontinuity points. Denote N ⊂ R the countable set of discontinuity point and, possibly
adding the value {essinfu} to N , we observe that for all t ∈ R \ N we have m({u = t}) = 0.
Hence, for every t ∈ R \ N and tn → t, we have χ{u>tn } → χ{u>t} in L1 (Ω) as n ↑ ∞. Therefore,
by lower semicontinuity on open sets (cf. Lemma 2.7), we have

F (t) = Per({u > t}, U ) ≤ lim Per({u > tn }, U ) = lim F (tn ). (3.15)
n↑∞ n→∞

Now, for any s > 0 write F −1 (0, s) = E1 ∪ E2 , for two suitable disjoint sets E1 , E2 so that
E1 ⊂ R \ N and E2 ⊂ N . Since E2 is countable, by construction, the conclusion will follow by
showing that E1 is L1 -measurable. However, if we consider R \ N with the induced subspace
topology, property (3.15) gives that F (t) is lower semicontinuous on R \ N (here, we are using
that lower semicontinuity can be equivalently check with converging sequences in the subspace
topology of R \ N ). In particular, E1 is closed in R \ N as it is a sub-level set of a lower
semicontinuous function. Now, recall that the subspace topology makes the inclusion map
ι : R \ N → R continuous (in fact, it is the coarsest topology with such property). Thus, ι(E1 ) is
an analytic set, and in particular, is L1 -measurable. As explained, this concludes the proof.

13
Lecture 4 - [12/12/2024]

Next, we prove the so-called coarea formula, linking a function of bounded variation with the
perimeter measures of its level sets.

Theorem 4.20 (Coarea formula). Let (X, d, m) be a metric measure space, let ∅ ̸= Ω ⊂ X be
open and let u ∈ L1 (Ω) be satisfying m({u > t}) < ∞ for all t > essinfu. Then, we have
ˆ +∞
u ∈ BV (Ω) if and only if Per({u > t}, Ω) dt < ∞.
−∞

Moreover, if any in the above holds, then t 7→ Per({u > t}, E) is L1 -measurable for every E ⊂ Ω
Borel and the coarea formula holds
ˆ ˆ +∞ ˆ
φ d|Du| = φ dPer({u > t}, ·)dt,
−∞

for every φ : Ω → [0, ∞] Borel.

Proof. We notice that, having assumed that m({u > t}) < ∞ for all t > essinf u we can write
´ ∞t 7→ Per({u > t}, Ω) is
Per({u > t}, Ω) taking possibly value +∞. Moreover, as Ω is open, then
1
L -measurable by Lemma 3.19 and nonnegative. Hence, the integral −∞ Per({u > t}, Ω) dt is
well defined.
Proof of (⇒). We will prove that
ˆ +∞
|Du|(U ) ≤ Per({u > t}, U ) dt, (4.16)
−∞

for all U ⊂ Ω open. Clearly, by choosing U = Ω, this implies that u ∈ BV (Ω) as well.
Assume for the moment that U is bounded and u ∈ L∞ (Ω), say |u| ≤ k m-a.e. in Ω for some
k ∈ N. We shall remove these two extra assumptions later.
 Fix an arbitrary integer n ∈ N, any
j−1 j
index j ∈ {−kn + 1, ..., kn} and consider points tj,n ∈ n , n so that

ˆ j
Per({u > tj,n }, U ) n
≤ Per({u > t}, U ) dt.
n j−1
−1
n

Define
kn
1 X
un (x) := −k + χ{u>tj,n } (x), ∀x ∈ U, n ∈ N,
n
j=−kn+1

observing that, since U is bounded and un is bounded, we have (un ) ⊂ L1 (U ). Using that the
total variation of a sum is controlled by the sum of the total variations, we next see that
kn ˆ +∞
1 X
|Dun |(U ) ≤ Per({u > tj,n }, U ) ≤ Per({u > t}, U ) dt (4.17)
n −∞
j=−kn+1

14
for all n ∈ N. Notice now that {u > tj,n } = ∪kn
i=j Ei,n where we set Ei,n := {ti,n < u ≤ ti+1,n }, for
all i = −kn + 1, ..., kn. Set also E−kn,n := {−k < u ≤ t−kn+1,n } and Ekn,n := {tkn,n < u < k}.
We can then rewrite
kn kn kn
1 X X 1 X
un (x) = −k + χEi,n (x) = −k + (j + kn)χEj,n (x)
n n
j=−kn+1 i=j j=−kn+1
kn
X j
= −k χE−kn,n (x) + χE (x).
n j,n
j=−kn+1

Thanks to the above, we have that un → u in L1 (U ) as n ↑ ∞. This follows noticing that

|u − un | = |u − j/n| ≤ 1/n on Ej,n for all j = −kn, ..., kn, hence
ˆ kn ˆ
X
|u − un | dm = |u − j/n| dm ≤ m(U )/n.
U j=−kn U ∩Ej,n

In particular, by lower semicontinuity for open sets in Lemma 2.7, sending n to infinity in (4.17)
proves (4.16) for U arbitrary open bounded and u ∈ L∞ (U ).
Now, for general u ∈ L1 (U ), define uk = max{−k, min{u, k}} for all k ∈ N which therefore
satisfies uk ∈ L∞ (Ω). Then, by the previous discussion, property (4.16) holds for every k ∈ N.
Since uk → u in L1 (U ) by Markov inequality, a further lower semicontinuity argument (cf.
Lemma 2.7 again) in combination with a chain-rule argument (cf. Lemma 3.15, noticing that
t 7→ min{−k, max{t, k}} is 1-Lipschitz) yields (4.16) for possibly unbounded functions, and for
U bounded open. Finally, for general U open and possibly unbounded, the claim (4.16) follows
by monotonicity and property iv) in Lemma 2.9 by considering the bounded sets Un := U ∩Bn (x)
for x ∈ U, n ∈ N and sending n ↑ ∞.
Proof of (⇐). We now show the converse and we assume u ∈ BV (Ω). We will prove that
ˆ +∞
Per({u > t}, U ) dt ≤ |Du|(U ), (4.18)
−∞

for all U ⊂ Ω open. Again, this implies the conclusion by choosing U = Ω.

First, let us further suppose that u ∈ Liploc (Ω) with u, lip(u) ∈ L1 (Ω) (in particular, it is
BV (Ω), c.f. (3.13)) and assume also that U ⊂ Ω is open and bounded. Define
ˆ
R ∋ t 7→ m(t) := lip(u) dm.
U ∩{u>t}

Observe that m(t) is nonnegative, monotone and bounded. Hence, it is a.e. differentiable. Say
m′ (t) exists for all t except in a negligible set N with m′ (t) ≤ 0 since m(·) is non-increasing.
Fix n ∈ N and t ∈ / N and consider∗

 0,
 if s ≤ t,
gn (s) := n(s − t), if s ∈ (t, t + 1/n],

1, if s > t + 1/n.


Set fn (x) := gn (u(x)). Since U is bounded, then fn ∈ L1 (U ) since gn is bounded. We claim

that fn → χ{u>t} in L1 (U ) as n ↑ ∞. Indeed,
ˆ ˆ
χ
|fn − {u>t} | dm = |fn − 1| dm ≤ m({t < u ≤ t + 1/n} ∩ U ) → 0,
U {t<u≤t+1/n}∩U

 0,
 if s ≤ t,
∗
In the classroom, I wrote gn (s) := n(t − s) − 1, if s ∈ (t, t + 1/n], that is not correct (not even Lipschitz)

1, if s > t + 1/n


15
as n ↑ ∞. Next, we claim that |Dfn |(U ) is uniformly bounded. To this aim, notice that gn
is n-Lipschitz, from which it follows that lip(fn ) ≤ nlip(u)χ{t≤u≤t+1/n} pointwise on U , as on
the complementary set of {t ≤ u ≤ t + 1/n} (which is open), we have that ´ fn is constant and,
therefore, lip(fn ) is zero by locality on open sets. Moreover, notice that {u=t}∩U lip(u) dm = 0
´
(since, at such t, we have ∃m′ (t) = limn→∞ n {t−1/n<u≤t}∩U lip(u) dm). Therefore, we get
ˆ ˆ
lip(fn ) dm ≤ n lip(u) dm ≤ n(m(t) − m(t + 1/n)).
U {t<u≤t+1/n}∩U

In particular, since t is a differentiability point of m(·), the right-hand side is uniformly bounded
in n ∈ N and this together with the lower semicontinuity in Lemma 2.7, shows
(3.13)
ˆ
Per({u > t}, U ) ≤ lim |Dfn |(U ) ≤ lim lip(fn ) dm ≤ −m′ (t) < +∞.
n↑∞ n↑∞ U

Integrating, and since the above holds for a.e. t ∈ R, we thus achieve
ˆ +∞ ˆ +∞ ˆ
′
Per({u > t}, U ) dt ≤ − m (t) dt ≤ lip(u) dm. (4.19)
−∞ −∞ U

Recall that U was an arbitrary bounded open set in Ω, therefore by the continuity property of
U 7→ Per({u > t}, U ) (cf. iv) in Lemma 2.9), the above holds for all U open by application of
monotone convergence in the left-hand and with an exhaustion argument with increasing balls.
Finally, for general nonnegative u ∈ BV (Ω), it is sufficient to argue by an approximation
using Lemma ´2.6. Indeed, consider the optimal sequence un → u in L1 (U ) given by Lemma
2.6 satisfying U lip(un ) dm → |Du|(U ) as n → ∞. Notice next that there is a non relabeled
subsequence in n so that

for a.e. t ∈ R χ{un >t} → χ{u>t} in L1 (U ) as n ↑ ∞. (4.20)

Suppose for the moment that this is true, then by lower semicontinuity and Fatou (4.19), yields
ˆ +∞ ˆ +∞ ˆ
Per({u > t}, U ) dt ≤ lim Per({un > t}, U ) dt ≤ lim lip(un ) dm.
−∞ n↑∞ −∞ n↑∞ U

Recalling that (un ) is optimal in the definition of |Du|(U ), our claim (4.18) follows.
We must still show (4.20) to conclude the current direction of the proof. This can be obtained
by a simple application of Fubini theorem noticing
ˆ +∞ ˆ ˆ ˆ +∞
χ{un >t} − χ{u>t} dmdt = χ{un >t} − χ{u>t} dtdm = ∥un − u∥L1 (U ) ,
−∞ U U −∞

where, in the last equality, we used the identity

ˆ +∞ ˆ max{u(x),un (x)}
χ{un >t} − χ{u>t} dt = dt = |un (x) − u(x)|.
−∞ min{u(x),un (x)}
´
Hence, up to a subsequence, U χ{un >t} − χ{u>t} dm → 0 as n → ∞ for a.e. t ∈ R, as desired.
Conclusion. Let us (equivalently) consider u ∈ BV (Ω). We need first to check that the integral
in the coarea formula makes sense by checking that t 7→ Per({u > t}, E) is L1 -measurable, for
every E ⊂ Ω Borel. Here the assumption u ∈ BV (Ω) will be needed. Recall that Lemma 3.19
guarantees that this is true for any open set U ⊂ Ω. Hence, for general E Borel, the claim
follows by a standard application of the monotone class theorem: denoting F the sub-class of
Borel sets E ⊂ Ω so that t 7→ Per({u > t}, E) is L1 -measurable, we need to show that F is a
monotone class, that is:

16
 • if En ↑ E and En ∈ F, then E ∈ F;

• f Fn ↓ F with Fn ∈ F, then F ∈ F.

We start with the former. Recall that Per({u > t}, ·) is a Borel measure, simply because
χ{u>t} ∈ L1 (Ω) by assumptions and by Theorem 3.10. Hence, it is monotone continuous from
below and it holds supn Per({u > t}, En ) = Per({u > t}, E). This shows that E ∈ F as well, as
the pointwise supremum of L1 -measurable function is L1 -measurable.
We next show the latter property, using here that u ∈ BV (Ω). Equivalently, we know that
Per({u > t}, ·) is a finite Borel measure for a.e. t ∈ (essinf u, +∞). In particular, it holds that
Per({u > t}, F1 ) < ∞ and by continuity from above that inf n Per({u > t}, Fn ) = Per({u > t}, F )
for a.e. t ∈ (essinf u, +∞). We point out that t 7→ inf n Per({u > t}, Fn ) is L1 -measurable, as a
pointwise infimum. Thus, t 7→ Per({u > t}, F ) is a.e. equal to a L1 -measurable functions, from
which F ∈ F follows.
Finally, for every φ : Ω → [0, ∞] Borel, the coarea formula and the conclusion of the proof
follow now by the combination of (4.16),(4.18) and by arbitrariness of U open.

As a corollary, the above theorem gives also the abundance of sets of finite perimeter on
every metric measure space. For instance, metric balls.

Exercise 4.21. Fix x0 ∈ X. Show that

Per(Br (x0 )) < ∞, a.e. r > 0.

(Spoiler)

Remark 4.22. Notice that, in Theorem 4.20 we assumed m({u > t}) < ∞, even though the
validity of the coarea formula for a possibly (locally) integrable BV function is known/expected.
In fact, for a general u ∈ BV (Ω), we can apply our result to the positive and negative parts u+
and u− to obtain
ˆ 0 ˆ +∞
|Du− | = Per({u < t}, ·)dt, |Du+ | = Per({u > t}, ·)dt,
−∞ 0

since u+ , u− ∈ BV (Ω) (c.f. Lemma 3.15), and thus the sets appearing in the above are of finite
measure by Cavalieri’s formula. The problem is that our definition of a set of finite perimeter
requires its mass to be finite. Hence, we are are not allowed to write Per({u > t}, ·) if, for
instance, t < 0. However, for the purposes of this course the assumption m({u > t}) < ∞ for
possibly negative
´ +∞ t > essinfu is completely natural. Hence, we will enforce it from now on so
that writing −∞ Per({u > t}, ·)dt is meaningful. ■

Finer version of coarea formula

In this part, we explore a localized coarea formula between successive level sets and then discuss
the decomposition properties of the total variation measure. The idea is to come up with a
regularity assumption on the function u for a finer coarea formula to hold. This will be key to
studying rearrangement inequalities in the second part of this course. We start with a motivating
example.

17
Example 4.23. Fix u ∈ BV (X) be satisfying m({u > t}) < ∞ for t > essinf u and consider
esse inf u < r < s < ess sup u and φ : X → [0, ∞] Borel. Is it true that
ˆ ˆ sˆ
φ|Du| = φ dPer({u > t}, ·) dt ?
{r<u<s} r

Clearly, simply choosing u = χE for E set of finite perimeter, we see that

ˆ (
0, if r > 1,
r 7→ d|Du| =
{r<u<+∞} Per(E), if r ≤ 1.

Therefore, an identity as in the above clearly cannot hold, as

ˆ ∞
r 7→ Per({u > t}) dt = rPer(E)χ(0,1) (r).
r

What goes wrong in this example is the presence of discontinuity points. ■

In the previous example, we saw that the coarea formula studied in Theorem 4.20 cannot
be in general localized between two successive level-sets of the original function. We next show
that, within this class of continuous functions, it is possible to give an affirmative answer.

Lemma 4.24 (Coarea for continuous functions). Let (X, d, m) be a metric measure space, let
∅ ̸= Ω ⊂ X be open and let u ∈ BV (Ω) ∩ C(Ω) be satisfying m({u > t}) < ∞ for t > essinf u.
Then, for every r < s and for every φ : Ω → [0, ∞] Borel, it holds
ˆ ˆ sˆ
φ d|Du| = φ dPer({u > t}, ·)dt. (4.21)
{r<u<s} r

In particular, |Du|({u = t}) = 0 holds for all t ∈ R.

Proof. The standard coarea formula gives

ˆ ˆ +∞ ˆ
φ d|Du| = φ dPer({u > t}, ·)dt.
{r<u<s} −∞ {r<u<s}

Recalling that the Perimeter measure is concentrated in the topological boundary of the set (c.f.
(3.14)), and since by continuity of u it holds ∂{u > t} ⊂ {u = t}, we deduce

Per({u > t}, {r < u < s}) = χ(r,s) (t) Per({u > t}, Ω), a.e. t ∈ R.

Combining this with the above, (4.21) follows.

In what follows, we would like to export ideas from the above discussions to the case of a
general function of bounded variation.

Goal. Single out a weak assumption on u ∈ BV (Ω) (in particular, weaker than
continuity) for (4.21) to hold.

18
Bibliography

[1] L. Ambrosio and S. Di Marino, Equivalent definitions of BV space and of total variation
on metric measure spaces, J. Funct. Anal., 266 (2014), pp. 4150–4188.

[2] A. Cianchi and N. Fusco, Functions of bounded variation and rearrangements, Arch.
Ration. Mech. Anal., 165 (2002), pp. 1–40.

[3] M. Miranda, Functions of bounded variation on “good” metric spaces, Journal de

Mathématiques Pures et Appliquées, 82 (2003), pp. 975–1004.

[4] F. Nobili and I. Y. Violo, Fine Pólya-Szegő rearrangement inequalities in metric spaces
and applications. arXiv:2409.14182, 2024.

19