Topics in analysis (MATS5170)

Analysis in metric measure spaces

– Lecture notes –

Enrico Pasqualetto∗

January 14, 2025

Contents
1 Preliminaries 1
1.1 Riesz spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Measure theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Continuous and Lipschitz maps . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4 Curves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.5 Metric measure spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

2 Metric Sobolev spaces 11

2.1 The space H 1,p (via relaxation) . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2 The space W 1,p (via derivations) . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3 The space N 1,p (via modulus) . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.4 The space B 1,p (via plans) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.5 Calculus rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.6 Equivalence of metric Sobolev spaces . . . . . . . . . . . . . . . . . . . . . . . 29

3 Bibliographical notes 31

1 Preliminaries
1.1 Riesz spaces
Let (P, ⪯) be a poset (i.e. a partially ordered set). For a non-empty set S ⊆ P , we say that:

• p ∈ P is an upper bound for S provided s ⪯ p for every s ∈ S.

∗
enrico.e.pasqualetto@jyu.fi, Department of Mathematics and Statistics, P.O. Box 35 (MaD), FI-40014
University of Jyväskylä, Finland

1
 • q ∈ P is a lower bound for S provided s ⪰ q for every s ∈ S.

• S is order bounded if it has both an upper bound and a lower bound.

• p̄ ∈ P is the supremum of S if it is an upper bound for S and it holds that p̄ ⪯ p

whenever p ∈ P is an upper bound for S.

• q̄ ∈ P is the infimum of S if it is a lower bound for S and it holds that q̄ ⪰ q whenever

q ∈ P is a lower bound for S.

If the supremum (resp. infimum) of S exists, then it is uniquely determined, thus we can
unambiguously denote it by sup S (resp. inf S). Moreover, we give the following definitions:

• A poset (P, ⪯) is called a lattice if p ∨ q := sup{p, q} and p ∧ q := inf{p, q} exist for

every p, q ∈ P . A set S ⊆ P is called a sublattice of P if it is closed under ∨ and ∧,
i.e. if it satisfies p ∨ q, p ∧ q ∈ S for every p, q ∈ S.

• A lattice (P, ⪯) is said to be Dedekind complete if every non-empty subset of P with

an upper bound has a supremum (or, equivalently, if every non-empty subset of P with
a lower bound has an infimum).

• A lattice (P, ⪯) has the countable sup property if for every non-empty subset S ⊆ P
having a supremum there exists a countable subset C ⊆ S such that sup S = sup C.

• A lattice (P, ⪯) has the countable inf property if for every non-empty subset S ⊆ P
having an infimum there exists a countable subset C ⊆ S such that inf S = inf C.
W
If (P, ⪯) is a lattice and S = {pi }i∈I ⊆ P has a supremum, we write i∈I pi := sup S.
V
Similarly, if S has an infimum, then we write i∈I pi := inf S. Given p1 , . . . , pn ∈ P , we write
_ ^
p1 ∨ . . . ∨ pn := pi , p1 ∧ . . . ∧ pn := pi .
i∈{1,...,n} i∈{1,...,n}

Definition 1.1 (Riesz space). A partially ordered linear space (V, ⪯) is a real vector
space V , together with a partial order ⪯ on V satisfying the following conditions:

v+z ⪯w+z for every v, w, z ∈ V with v ⪯ w,

λv ⪰ 0 for every λ ∈ R and v ∈ V with λ ≥ 0 and v ⪰ 0.

A Riesz space is a partially ordered linear space that is a also lattice.

The positive cone V + of a Riesz space (V, ⪯) is defined as

V + := {v ∈ V | v ⪰ 0}.

2
1.2 Measure theory
Let (X, Σ) be a measurable space (i.e. a set X together with a σ-algebra Σ). Then we define

L0ext (Σ) := f : X → R ∪ {±∞} f is Σ-measurable ,



L0 (Σ) := f ∈ L0ext (Σ) f (x) ∈ R for every x ∈ X ,



L∞ (Σ) := f ∈ L0 (Σ) supX |f | < +∞ .



Moreover, if µ ≥ 0 is a measure on (X, Σ) and p ∈ [1, ∞) is a given exponent, then we define

 ˆ 
p 0 p
L (µ) := f ∈ L (Σ) |f | dµ < +∞ .

We also have that µ induces an equivalence relation ∼µ on L0ext (Σ), which is defined as follows:
given any f, g ∈ L0ext (Σ), we declare that f ∼µ g if and only if f = g holds µ-a.e. on X, i.e.


µ x ∈ X f (x) ̸= g(x) = 0.

We then give the following definitions:

• We denote by L0ext (µ) := L0ext (Σ)/ ∼µ the quotient space induced by ∼µ and by
πµ : L0ext (Σ) → L0ext (µ) the canonical projection map. The elements of L0ext (µ) can be
regarded as Σ-measurable functions f : X → R ∪ {±∞} considered up to µ-a.e. equality.

• We define L0 (µ) := πµ (L0 (Σ)) = L0 (Σ)/ ∼µ . The space L0 (µ) is a vector space with
respect to the usual pointwise operations. Sometimes we will use the shorthand notation

f µ := πµ (f ) ∈ L0 (µ) for f ∈ L0 (Σ).

In particular, we write 1µE ∈ L0 (µ) for E ∈ Σ, where 1E ∈ L0 (Σ) is the characteristic

function of E, i.e. 1E (x) := 1 for every x ∈ E and 1E (x) := 0 for every x ∈ X \ E.

• The ∞-Lebesgue space is given by L∞ (µ) := πµ (L∞ (Σ)) = L∞ (Σ)/ ∼µ . The space
L∞ (µ) is a vector subspace of L0 (µ) and it is a Banach space if endowed with the norm

∥f ∥L∞ (µ) := ess supX |f | := inf λ ≥ 0 |f | ≤ λ holds µ-a.e. on X .

• For p ∈ [1, ∞), the p-Lebesgue space is defined as Lp (µ) := πµ (Lp (µ)) = Lp (µ)/ ∼µ .
It is a vector subspace of L0 (µ) and it is a Banach space if endowed with the norm
ˆ 1/p
p
∥f ∥Lp (µ) := |f | dµ .

We equip the space L0ext (µ) with the following partial order relation: given any f, g ∈ L0ext (µ),
we declare that f ≤ g if and only if f¯(x) ≤ ḡ(x) holds for µ-a.e. x ∈ X whenever f¯ ∈ L0ext (Σ)
(resp. ḡ ∈ L0ext (Σ)) is a representative of f (resp. of g). Then (L0ext (µ), ≤) is a lattice. Notice
also that L0 (µ) is sublattice of L0ext (µ), and Lp (µ) is a sublattice of L0 (µ) for all p ∈ [1, ∞].

3
Remark 1.2. Given a σ-finite measure space (X, Σ, µ), there exists a finite measure µ̃ ≥ 0
on (X, Σ) such that µ ≪ µ̃ ≤ µ. Indeed, we can find a partition {En }n∈C ⊆ Σ of X (where
either C = {1, . . . , n̄} for some n̄ ∈ N or C = N) into sets satisfying 0 < µ(En ) < +∞ for
every n ∈ C. Let us then consider the finite measure µ̃ on (X, Σ), which is given by
X 1
µ̃ := µ|En .
2n µ(En )
n∈C

It can be readily checked that µ ≪ µ̃ ≤ µ. ■

Proposition 1.3. Let (X, Σ, µ) be a σ-finite measure space and p ∈ {0} ∪ [1, ∞]. Then Lp (µ)
is a Dedekind complete Riesz space having the countable sup and inf properties.

Proof. It can be easily checked that Lp (µ) is a Riesz space. Let us now fix a subset S = {fi }i∈I
of L0 (µ) having an upper bound g ∈ L0 (µ). Fix a finite measure µ̃ on (X, Σ) with µ ≪ µ̃ ≤ µ,
see Remark 1.2. Define f˜i := arctan ◦fi ∈ L0 (µ̃) for every i ∈ I and g̃ := arctan ◦g ∈ L0 (µ̃).
Let us now consider the family F := { i∈F f˜i : F ⊆ I finite} ⊆ L0 (µ̃) and define
W

ˆ 
M := sup ˜ ˜
f dµ̃ f ∈ F .

´
Since f˜ dµ̃ ≤ π2 µ̃(X) for every f˜ ∈ F, we have that M < +∞. Next, take a maximising
´
sequence (f˜n )n∈N ⊆ F for the above problem, i.e. f˜n dµ̃ → M as n → ∞. Up to replacing
each f˜n with f˜1 ∨ . . . ∨ f˜n , we can assume that (f˜n )n∈N is a non-decreasing sequence. Define

h̃(x) := sup f˜n (x) = lim f˜n (x) for µ̃-a.e. x ∈ X.

n∈N n→∞

Given that f˜n ≤ g̃ for every n ∈ N, we have that h̃ ≤ g̃, thus in particular h̃ ∈ L0 (µ̃). Notice
that h̃ = supi∈C f˜i for some countable set C ⊆ I. Moreover, by applying the monotone
´
convergence theorem, we get that M = h̃ dµ̃. We claim that h := tan ◦h̃ = sup S. Indeed:

• Assume that h is not an upper bound for S. Then there exist a set P ∈ Σ and j ∈ I
such that µ(P ) > 0 and fj > h holds µ-a.e. on P . Then µ̃(P ) > 0 and f˜j > h̃ holds
µ̃-a.e. on P , whence it follows (thanks to the monotone convergence theorem) that
ˆ ˆ ˆ
˜
M = h̃ dµ̃ < h̃ ∨ fj dµ̃ = lim f˜n ∨ f˜j dµ̃ ≤ M,
n→∞

which leads to a contradiction. Therefore, h ∈ L0 (µ) is an upper bound for S.

• Let f ∈ L0 (µ) be an arbitrary upper bound for S. In particular, f˜ := arctan ◦f ∈ L0 (µ̃)

satisfies f˜n ≤ f˜ for every n ∈ N, so that h̃ ≤ f˜ and thus accordingly h ≤ f .

All in all, h = sup S = supi∈C fi . Hence, L0 (µ) is Dedekind complete and it has the countable
sup property. Similar arguments show that L0 (µ) has also the countable inf property. It is
then easy to deduce that Lp (µ), for any p ∈ [1, ∞], has the same properties as well.

4
1.3 Continuous and Lipschitz maps
Let (X, τX ) and (Y, τY ) be topological spaces. Then we denote by C(X; Y) the space of all
continuous maps u : X → Y. If Y = R, we write C(X) := C(X; R). The support supp(f ) ⊆ X
of a function f : X → R is defined as the closure of {f ̸= 0} := {x ∈ X : f (x) ̸= 0} in X.
The space C(X) is a vector space with respect to the usual pointwise operations. Some
distinguished vector subspaces of C(X) are the following:

Cb (X) := f ∈ C(X) f is a bounded function ,

Cc (X) := f ∈ C(X) supp(f ) is compact .

Notice that Cb (X), Cc (X) are vector subspaces of C(X) and that Cc (X) ⊆ Cb (X). The norm

∥f ∥Cb (X) := sup |f | for every f ∈ Cb (X)

X

is complete, so that (Cb (X), ∥ · ∥Cb (X) ) is a Banach space. Furthermore, (Cb (X), ≤) is a Riesz
space, where by ≤ we mean the usual pointwise partial order on Cb (X).
Assuming (X, d) is a metric space, we also consider the space Cbs (X), which is given by

Cbs (X) := f ∈ C(X) supp(f ) is bounded .

We have that Cc (X) ⊆ Cbs (X) and that Cbs (X) is a vector subspace of C(X), but it can
happen that Cbs (X) is not contained in Cb (X). The space (Cbs (X), ≤) is a Riesz space.

1.3.1 Lipschitz maps

Definition 1.4 (Lipschitz map). Let (X, dX ) and (Y, dY ) be metric spaces. Then we say that
a map u : X → Y is L-Lipschitz, for some real number L ≥ 0, provided it satisfies

dY (u(x), u(y)) ≤ L dX (x, y) for every x, y ∈ X. (1.1)

A map u : X → Y is said to be Lipschitz if it is L-Lipschitz for some L ≥ 0. We denote by

LIP(X; Y) the set of all Lipschitz maps from X to Y. If Y = R, we write LIP(X) := LIP(X; R).

Notice that LIP(X; Y) ⊆ C(X; Y). Lipschitz maps are associated with several quantities:

Definition 1.5. Let (X, dX ) and (Y, dY ) be metric spaces. Let u ∈ LIP(X; Y) be given. Then:

i) For any set E ⊆ X, we define the Lipschitz constant of u on E as

 
dY (u(x), u(y))
Lip(u; E) := sup x, y ∈ E, x ̸= y ∈ [0, +∞).
dX (x, y)

The Lipschitz constant of u is given by Lip(u) := Lip(u; X).

5
 ii) The slope of u is the function lip(u) : X → [0, +∞] defined as

dY (u(x), u(y))
lip(u)(x) := lim sup for every accumulation point x ∈ X
y→x dX (x, y)

and lip(u)(x) := 0 for every isolated point x ∈ X.

iii) The asymptotic Lipschitz constant of u at a point x ∈ X is defined as

lipa (u)(x) := inf Lip(u; Br (x)).

r>0

The resulting function lipa (u) : X → [0, +∞) is called the asymptotic slope of u.

Notice that for any given u ∈ LIP(X; Y) we have that

lip(u)(x) ≤ lipa (u)(x) ≤ Lip(u) for every x ∈ X.

We mostly focus on the space LIP(X) of real-valued Lipschitz functions, which is a vector
subspace of C(X). Some distinguished vector subspaces of LIP(X) are the following:

LIPb (X) := f ∈ LIP(X) f is a bounded function = LIP(X) ∩ Cb (X),

LIPbs (X) := f ∈ LIP(X) supp(f ) is bounded = LIP(X) ∩ Cbs (X),

LIPc (X) := f ∈ LIP(X) supp(f ) is compact = LIP(X) ∩ Cc (X).

Notice that LIPc (X) ⊆ LIPbs (X) ⊆ LIPb (X). We also endow LIPb (X) with the norm

∥f ∥LIPb (X) := ∥f ∥Cb (X) + Lip(f ).

It holds that (LIPb (X), ∥ · ∥LIPb (X) ) is a Banach space and it is a Riesz subspace of Cb (X).
Moreover, LIPbs (X) and LIPc (X) are (possibly non-complete) Riesz subspaces of LIPb (X).
For any f, g ∈ LIPbs (X) and λ ∈ R, it can be readily checked that λf, f g ∈ LIPbs (X) and

lipa (λf )(x) = |λ|lipa (f )(x), (1.2a)

lipa (f + g)(x) ≤ lipa (f )(x) + lipa (g)(x), (1.2b)
lipa (f g)(x) ≤ |f (x)| lipa (g)(x) + |g(x)| lipa (f )(x) (1.2c)

for every x ∈ X.

Remark 1.6. Furthermore, (LIPb (X), ∥·∥LIPb (X) ) is a unital, commutative Banach algebra.
Videlicet, the space LIPb (X) is a unital, commutative, associative algebra that is endowed
with a complete norm ∥ · ∥LIPb (X) satisfying

∥f g∥LIPb (X) ≤ ∥f ∥LIPb (X) ∥g∥LIPb (X) for every f, g ∈ LIPb (X).

The verification of the above claim is left as an exercise. ■

6
1.4 Curves
Let (X, d) be a metric space. By a curve in X we mean a continuous map γ : [a, b] → X, for
some a, b ∈ R with a ≤ b. We denote by C (X) the set of all curves in X. Namely, we set
[
C (X) := C([a, b]; X).
a,b∈R:
a≤b

Given any curve γ ∈ C (X), we denote by [aγ , bγ ] its domain of definition. For any t ∈ [aγ , bγ ],
we will often write γt instead of γ(t). By a subcurve of γ ∈ C (X) we mean a curve of the
form γ|[ã,b̃] ∈ C (X), for some ã, b̃ ∈ [aγ , bγ ] with ã ≤ b̃.
Given any a, b ∈ R with a < b, we endow C([a, b]; X) with the supremum distance:

dC([a,b];X) (γ, σ) := sup d(γt , σt ) for every γ, σ ∈ C([a, b]; X).

t∈[a,b]

It holds that


(X, d) is complete ⇐⇒ C([a, b]; X), dC([a,b];X) is complete,


(X, d) is separable ⇐⇒ C([a, b]; X), dC([a,b];X) is separable.

Definition 1.7 (Evaluation maps). Let (X, d) be a metric space and t ∈ [0, 1]. Then we
define the evaluation map et : C([0, 1]; X) → X at time t as

et (γ) := γt for every γ ∈ C([0, 1]; X).

It holds that et is 1-Lipschitz for every t ∈ [0, 1].

1.4.1 Rectifiable curves

Definition 1.8 (Rectifiable curve). Let (X, d) be a metric space. Given any curve γ ∈ C (X),
we define its length ℓ(γ) = ℓd (γ) ∈ [0, +∞] as
Xn 
ℓ(γ) := sup d(γti , γti−1 ) n ∈ N, aγ = t0 < t1 < . . . < tn = bγ .
i=1

We say that γ is rectifiable provided ℓ(γ) < +∞. We denote by R(X) the set of all rectifiable
curves in X.

To any γ ∈ R(X) we associate its length function sγ : [aγ , bγ ] → [0, ℓ(γ)], given by

sγ (t) := ℓ(γ|[aγ ,t] ) for every t ∈ [aγ , bγ ].

The length function sγ is non-decreasing and continuous (Exercise). We define the right
inverse tγ : [0, ℓ(γ)] → [aγ , bγ ] of the function sγ as

tγ (s) := sup t ∈ [aγ , bγ ] sγ (t) = s for every s ∈ [0, ℓ(γ)].

Then tγ is non-decreasing and continuous from the right (Exercise).

7
Definition 1.9 (Arc length parameterisation). Let (X, d) be a metric space and γ ∈ R(X).
Then we define the arc length parameterisation γ̂ : [0, ℓ(γ)] → X of γ as

γ̂(s) := γ(tγ (s)) for every s ∈ [0, ℓ(γ)].

It holds that γ̂ : [0, ℓ(γ)] → X is the unique curve satisfying γt = γ̂sγ (t) for every t ∈ [aγ , bγ ].
It can be also readily checked that

ℓ(γ̂|[s,t] ) = t − s for every s, t ∈ [0, ℓ(γ)] with s < t.

In particular, ℓ(γ̂) = ℓ(γ) and the curve γ̂ is 1-Lipschitz.

Definition 1.10 (Line integral). Let (X, d) be a metric space, γ ∈ R(X) and ρ : X → [0, +∞]
´
a Borel function. Then we define the line integral γ ρ ∈ [0, +∞] of ρ over γ as
ˆ ˆ ℓ(γ)
ρ := ρ(γ̂s ) ds.
γ 0

Definition 1.11 (Length space). A metric space (X, d) is called a length space provided

d(x, y) = inf ℓ(γ) γ : [0, 1] → X rectifiable, γ0 = x, γ1 = y for every x, y ∈ X.

1.4.2 Absolutely continuous curves

Definition 1.12 (Absolutely continuous curve). Let (X, d) be a metric space and q ∈ [1, ∞].
Then we say that a given curve γ : [a, b] → X is q-absolutely continuous provided there
exists a function g ∈ Lq (a, b)+ such that
ˆ t
d(γs , γt ) ≤ g(r) dr for every s, t ∈ [a, b] with s < t. (1.3)
s

We denote by AC q ([a, b]; X) the set of all q-absolutely continuous curves in X defined on [a, b].

Remark 1.13. Notice that any function γ : [a, b] → X satisfying (1.3) for some g ∈ Lq (a, b)+
is continuous (thus, a curve), thanks to the dominated convergence theorem. ■

We shorten AC([a, b]; X) := AC 1 ([a, b]; X). Note that AC ∞ ([a, b]; X) = LIP([a, b]; X) and

AC q2 ([a, b]; X) ⊆ AC q1 ([a, b]; X) for every q1 , q2 ∈ [1, ∞] with q1 < q2 .

Theorem 1.14 (Metric speed). Let (X, d) be a metric space and γ ∈ AC([a, b]; X). Define

d(γt+h , γt )
|γ̇t | := lim sup for every t ∈ [a, b].
h→0 |h|

Then the resulting function |γ̇| : [a, b] → [0, +∞] satisfies the following properties:
d(γt+h ,γt )
i) |γ̇| ∈ L1 ([a, b]) and |γ̇t | = limh→0 |h| for L 1 -a.e. t ∈ [a, b].

8
 ´t
ii) |γ̇| verifies (1.3), i.e. d(γs , γt ) ≤ s |γ̇r | dr holds for every s, t ∈ [a, b] with s < t.

iii) |γ̇| is the L 1 -a.e. minimal function g ∈ L1 (a, b)+ satisfying (1.3). Videlicet, the L 1 -a.e.
g ∈ L1 (a, b)+ : g satisfies (1.3) ∈ L1 (a, b)+ .
V
equivalence class of |γ̇| coincides with
We say that |γ̇| is the metric speed of the curve γ.
Proof. Fix any function g ∈ L1 (a, b)+ for which (1.3) holds. Since [a, b] is compact and γ is
continuous, we have that γ([a, b]) is a compact (thus, a separable) subset of X. Then we can
find a sequence (xn )n∈N ⊆ γ([a, b]) that is dense in γ([a, b]) with respect to the distance d.
For any n ∈ N, define hn : [a, b] → R as hn (t) := d(γt , xn ) for every t ∈ [a, b]. Note that
ˆ t
|hn (s) − hn (t)| ≤ d(γs , γt ) ≤ g(r) dr for every s, t ∈ [a, b] with s < t (1.4)
s
by the reverse triangle inequality, thus hn is an absolutely continuous function. In particular,
hn is a.e. differentiable. Fix an equivalence class vn ∈ L1 (a, b) of its a.e. derivative h′n and let

v(t) := sup vn (t) for every t ∈ (a, b).

n∈N

Dividing (1.4) by t − s and letting s → t, we obtain (thanks to the Lebesgue differentiation

theorem) that |vn (t)| ≤ g(t) for a.e. t ∈ (a, b). It follows that |v(t)| ≤ g(t) for a.e. t ∈ (a, b),
so that v ∈ L1 (a, b). Now, observe that


d(γs , γt ) = sup hn (s) − hn (t) for every s, t ∈ [a, b]. (1.5)
n∈N

Indeed, the inequality ≥ follows from (1.4), while the inequality ≤ holds because for any ε > 0
there exists n ∈ N such that d(γt , xn ) < ε, so that hn (s) − hn (t) ≥ d(γs , γt ) − 2ε. Therefore,
ˆ t ˆ t
d(γs , γt ) = sup vn (r) dr ≤ v(r) dr for every s, t ∈ [a, b] with s < t, (1.6)
n∈N s s
´t
thanks to (1.5) and to hn (t) − hn (s) = s vn (r) dr (due to the absolute continuity of hn ). This
shows that v is the a.e. minimal integrable function satisfying (1.3). To conclude, it remains
d(γ ,γt )
to check that limh→0 t+h |h| = v(t) for a.e. t ∈ (a, b). On the one hand, (1.6) yields
d(γt+h , γt )
lim sup ≤ v(t) for a.e. t ∈ (a, b) (1.7)
h→0 |h|
by the Lebesgue differentiation theorem. On the other hand, for any a < s < t < b it holds
´t
d(γs , γt ) ≥ hn (t)−hn (s) = s vn (r) dr, so that (again by the Lebesgue differentiation theorem)
d(γ ,γt )
we have that lim inf h→0 t+h |h| ≥ supn∈N vn (t) = v(t) for a.e. t ∈ (a, b). Combining this fact
d(γt+h ,γt )
with (1.7), we conclude that limh→0 |h| = v(t) for a.e. t ∈ (a, b), as we claimed.
´1
Every absolutely continuous curve γ ∈ AC([0, 1]; X) is rectifiable and ℓ(γ) = 0 |γ̇t | dt.
More generally, for any Borel function ρ : X → [0, +∞] we have that
ˆ ˆ 1
ρ= ρ(γt )|γ̇t | dt.
γ 0

We omit the proof of these claims.

9
Remark 1.15. There exist rectifiable curves that are not absolutely continuous. For example,
consider the Cantor function c : [0, 1] → [0, 1]. Since it is continuous and non-decreasing, it is
rectifiable, but it is not absolutely continuous (Exercise). ■

1.5 Metric measure spaces

Let (X, τ ) be a topological space. Then we denote by B(X), or B(X, τ ), the Borel σ-algebra
of (X, τ ), i.e. the smallest σ-algebra on X containing τ . By a Borel measure on X we mean a
(σ-additive) measure µ : B(X) → [0, +∞]. We say that µ is locally finite if for every x ∈ X
there exists U ∈ τ such that x ∈ U and µ(U ) < +∞.
We are mostly interested to a subclass of Borel measures, called Radon measures. Recall
that if (X, τ ) is Hausdorff, then every compact subset of X is closed (and, thus, it is Borel).

Definition 1.16 (Radon measure). Let (X, τ ) be a Hausdorff topological space. Then by a
Radon measure on X we mean a Borel measure µ on X such that µ(K) < +∞ for every
compact set K ⊆ X and

µ(E) = inf µ(U ) U ⊇ E open for every Borel set E ⊆ X, (1.8a)

µ(U ) = sup µ(K) K ⊆ U compact for every open set U ⊆ X. (1.8b)

The properties in (1.8a) and (1.8b) are called the outer regularity and the inner regularity
of µ, respectively.

A Borel measure µ on a metric space (X, d) is said to be boundedly finite if µ(B) < +∞
for every bounded Borel set B ⊆ X. Note that every boundedly-finite Borel measure is both
σ-finite and locally finite.

Definition 1.17 (Metric measure space). We say that (X, d, m) is a metric measure space
provided (X, d) is a metric space and m is a boundedly-finite Radon measure on X.

Given any metric measure space (X, d, m), we have (Exercise) that the inner regularity
property (1.8b) extends to all Borel sets. Videlicet, it holds that

m(E) = sup m(K) K ⊆ E compact for every Borel set E ⊆ X.

Proposition 1.18 (Separability of Lp ). Let (X, d, m) be a metric measure space. Then the
p-Lebesgue space Lp (m) is separable for every p ∈ [1, ∞).

Proof. Left as an exercise.

Lemma 1.19 (Density of Lipschitz functions in Lp ). Let (X, d, m) be a metric measure space
and p ∈ [1, ∞). Then {f m : f ∈ LIPbs (X)} is dense in Lp (m).

Proof. By definition of the Lebesgue integral, the vector subspace of Lp (m) that is generated
by {1mE : E ∈ B(X), m(E) < +∞} ⊆ L (m) is dense in L (m). Therefore, since LIPbs (X) is
p p

10
a vector space, it is sufficient to show that for any E ∈ B(X) with m(E) < +∞ and ε > 0
there exists a function f ∈ LIPbs (X) such that ∥f m − 1m E ∥Lp (m) < ε. To prove it, take a
p p
compact subset K of E satisfying m(E \ K) < ε /2 , whose existence is guaranteed by the
inner regularity of m. Now define the function fn : X → [0, 1], for any n ∈ N with n ≥ 1, as


fn (x) := 1 − n d(x, K) ∨ 0 for every x ∈ X,

where d(x, K) := min{d(x, y) : y ∈ K}. One can check that each fn is a boundedly-supported
n-Lipschitz function (Exercise). Moreover, letting C := {x ∈ X : d(x, K) ≤ 1}, we have that
C is a bounded closed set (so that m(C) < +∞) and that fn ≤ 1C on X for every n ≥ 1.
In particular, since fn (x) → 1K (x) as n → ∞ for every x ∈ X (Exercise), by applying the
´
dominated convergence theorem we obtain that |fn − 1K |p dm → 0 as n → ∞. Choosing n
so large that ∥fnm − 1m
K ∥Lp (m) < ε/2 and letting f := fn ∈ LIPbs (X), we thus conclude that

ε
∥f m − 1m
E ∥Lp (m) ≤ ∥f − 1K ∥Lp (m) + ∥1K − 1E ∥Lp (m) <
m m m m
+ m(E \ K)1/p < ε,
2
whence the statement follows.

Corollary 1.20. Let (X, d, m) be a metric measure space. Then the set {f m : f ∈ LIPbs (X)}
is weakly∗ sequentially dense in L∞ (m).

Proof. Fix some finite Borel measure m̃ on X satisfying m ≪ m̃ ≤ m; recall Remark 1.2. Let
f ∈ L∞ (m) = L∞ (m̃) ⊆ L1 (m̃) be given. By Lemma 1.19, we can find (fn )n ⊆ LIPbs (X) such
that fnm̃ → f in L1 (m̃). Up to replacing fn with (fn ∧∥f ∥L∞ (m) )∨(−∥f ∥L∞ (m) ), we can assume
that |fn | ≤ ∥f ∥L∞ (m) on X. Up to a non-relabeled subsequence, we can also assume that
fnm̃ (x) → f (x) for m̃-a.e. x ∈ X. Therefore, for any g ∈ L1 (m) we have |fnm g| ≤ ∥f ∥L∞ (m) |g|
´ ´
for every n ∈ N and fnm g → f g in the m-a.e. sense, thus fnm g dm → f g dm as n → ∞ by
∗
the dominated convergence theorem. This proves that fnm ⇀ f weakly∗ in L∞ (m).

2 Metric Sobolev spaces

2.1 The space H 1,p (via relaxation)
Definition 2.1 (Cheeger energy). Let (X, d, m) be a metric measure space and p ∈ (1, ∞).
Given any function f ∈ Lp (m), we define its Cheeger p-energy Chp (f ) ∈ [0, +∞] as
 ˆ 
1
Chp (f ) := inf lim inf lipa (fn )p dm (fn )n ⊆ LIPbs (X), fnm → f in Lp (m) .
n→∞ p

Definition 2.2 (Sobolev space via relaxation). Let (X, d, m) be a metric measure space. Let
p ∈ (1, ∞) be a given exponent. Then we define the Sobolev space H 1,p (X) as

H 1,p (X) := f ∈ Lp (m) Chp (f ) < +∞ .



11
 The set H 1,p (X) is a vector subspace of Lp (m). We endow H 1,p (X) with the norm

1/p
∥f ∥H 1,p (X) := ∥f ∥pLp (m) + p Chp (f ) for every f ∈ H 1,p (X).

It holds that (H 1,p (X), ∥ · ∥H 1,p (X) ) is a Banach space (Exercise).

Definition 2.3 (Asymptotic relaxed slope). Let (X, d, m) be a metric measure space and
p ∈ (1, ∞). Let f ∈ Lp (m) be given. Then we say that a function G ∈ Lp (m)+ is an
asymptotic p-relaxed slope of f provided there exist a sequence (fn )n ⊆ LIPbs (X) and a
function G̃ ∈ Lp (m)+ with G̃ ≤ G such that fnm ⇀ f and lipa (fn )m ⇀ G̃ weakly in Lp (m).
We denote by RSa,p (f ) the set of all asymptotic p-relaxed slopes of f .

Lemma 2.4. Let (X, d, m) be a metric measure space and p ∈ (1, ∞). Then it holds that

H 1,p (X) = f ∈ Lp (m) RSa,p (f ) ̸= ∅ .


(2.1)

Moreover, the set RSa,p (f ) is a closed convex subset of Lp (m) for every f ∈ H 1,p (X).

Proof. Let f ∈ H 1,p (X) be given. Then there exists a sequence (fn )n ⊆ LIPbs (X) such
that fnm → f in Lp (m) and supn∈N ∥lipa (fn )m ∥Lp (m) < +∞. Since Lp (m) is reflexive, we
have that (up to a non-relabelled subsequence) lipa (fn )m ⇀ G weakly in Lp (m) for some
G ∈ Lp (m)+ , so that G ∈ RSa,p (f ) and thus RSa,p (f ) ̸= ∅. Conversely, if f ∈ Lp (m) is a
given function satisfying RSa,p (f ) ̸= ∅, then we can find a sequence (fn )n∈N ⊆ LIPbs (X) and
a function G ∈ Lp (m)+ such that fnm ⇀ f and lipa (fn )m ⇀ G weakly in Lp (m). Since every
weakly converging sequence is bounded (by the Banach–Steinhaus theorem), we deduce that
´
M := supn∈N lipa (fn )p dm < +∞. Thanks to the Mazur lemma, we can find a sequence
(gn )n∈N ⊆ LIPbs (X) of convex combinations of elements of {fn : n ∈ N} such that gnm → f
strongly in Lp (m) as n → ∞. Using (1.2a) and (1.2b), we deduce that ∥lipa (gn )m ∥Lp (m) ≤ M
for every n ∈ N, thus accordingly f ∈ H 1,p (X). All in all, (2.1) is proved.
Next, let us check that RSa,p (f ) is convex for every f ∈ H 1,p (X). Fix G1 , G2 ∈ RSa,p (f )
and λ ∈ [0, 1]. Then fj,n m ⇀ f and lip (f )m ⇀ G′ weakly in Lp (m) for j = 1, 2, for some
a j,n j
(f1,n )n , (f2,n )n ⊆ LIPbs (X) and G1 , G2 ∈ L (m) with G′1 ≤ G1 and G′2 ≤ G2 . Observe that
′ ′ p +

m
λf1,n + (1 − λ)f2,n ⇀ f, Hn := λ lipa (f1,n )m + (1 − λ) lipa (f2,n )m ⇀ λG′1 + (1 − λ)G′2

weakly in Lp (m) as n → ∞, thus in particular supn∈N ∥Hn ∥Lp (m) < +∞. We infer from (1.2a)

m
and (1.2b) that lipa λf1,n + (1 − λ)f2,n ≤ Hn for all n ∈ N, so that (up to a non-relabelled

m
subsequence) we have that lipa λf1,n + (1 − λ)f2,n ⇀ H for some function H ∈ Lp (m)+
satisfying H ≤ λG′1 +(1−λ)G′2 ≤ λG1 +(1−λ)G2 . This proves that λG1 +(1−λ)G2 ∈ RSa,p (f ).
Finally, we prove that RSa,p (f ) is a closed set in Lp (m). Fix a sequence (Gk )k ⊆ RSa,p (f )
that strongly converges to some limit function G ∈ Lp (m). We aim to show that G ∈ RSa,p (f ).
For any k ∈ N, take a sequence (fk,n )n ⊆ LIPbs (X) and a function G′k ∈ Lp (m)+ with G′k ≤ Gk
m ⇀ f and lip (f m ′ p
such that fk,n a k,n ) ⇀ Gk weakly in L (m) as n → ∞. Now fix k ∈ N.
Using the Mazur lemma, (1.2a) and (1.2b), we find a convex combination gk ∈ LIPbs (X) of

12
{fk,n : n ∈ N} and a function Rk ∈ Lp (m) with lipa (gk )m ≤ Rk such that ∥gkm − f ∥Lp (m) ≤ 1/k
and ∥Rk − G′k ∥Lp (m) ≤ 1/k. In particular, gkm → f strongly in Lp (m) as k → ∞ and

lipa (gk )m ≤ Rk = G′k + (Rk − G′k ) ≤ Gk + (Rk − G′k ) → G strongly in Lp (m),

so that supk∈N ∥lipa (gk )m ∥Lp (m) < +∞. Up to a non-relabelled subsequence, we thus have that
lipa (gk )m ⇀ G′ weakly in Lp (m) for some G′ ∈ Lp (m) with G′ ≤ G. Hence, G ∈ RSa,p (f ).

Due to the uniform convexity of Lp (m), every non-empty closed convex subset of Lp (m)
has a unique element of minimal Lp (m)-norm. Thus, the following definition is well-posed:

Definition 2.5 (Minimal asymptotic relaxed slope). Let (X, d, m) be a metric measure space
and p ∈ (1, ∞). Let f ∈ H 1,p (X) be given. Then we denote by |Df |H the element of RSa,p (f )
of minimal Lp (m)-norm. We call |Df |H the minimal asymptotic p-relaxed slope of f .

It can be readily checked that {f m : f ∈ LIPbs (X)} ⊆ H 1,p (X) and

|Df m |H ≤ lipa (f )m for every f ∈ LIPbs (X).

Remark 2.6. Differently from the classical Euclidean Sobolev spaces, the metric Sobolev
space might depend on the chosen exponent p. Videlicet, there are examples of metric measure
spaces (X, d, m) where, for two different exponents p1 , p2 ∈ (1, ∞), there exists a function
f ∈ H 1,p1 (X) ∩ H 1,p2 (X) whose minimal asymptotic p1 - and p2 -relaxed slopes are different.
Nevertheless, for simplicity of notation we write |Df |H (omitting the dependence on p). ■

Proposition 2.7. Let (X, d, m) be a metric measure space, p ∈ (1, ∞) and f ∈ H 1,p (X).
Then the following properties are satisfied:

i) RSa,p (f ) is a sublattice of Lp (m).

ii) |Df |H ≤ G holds for every G ∈ RSa,p (f ).

´
iii) It holds that Chp (f ) = p1 |Df |pH dm.

iv) There exists a sequence (fn )n ⊆ LIPbs (X) such that fnm → f and lipa (fn )m → |Df |H
strongly in Lp (m).

Proof.
i). Let G1 , G2 ∈ RSa,p (f ) be given. Our aim is to show that G1 ∨ G2 , G1 ∧ G2 ∈ RSa,p (f ),
whence it follows that RSa,p (f ) is a sublattice of Lp (m). The fact that G1 ∨ G2 ∈ RSa,p (f ) is
obvious, thus let us consider G1 ∧ G2 . Fix a Borel representative E of {G1 < G2 }. Thanks
to the inner regularity of m, it is possible to find (Exercise) a sequence (Kn )n of compact
subsets of E such that 1m Kn G1 + 1X\Kn G2 → 1E G1 + 1X\E G2 = G1 ∧ G2 strongly in L (m)
m m m p

as n → ∞. In view of Lemma 2.4, it then suffices to show that 1Kn G1 + 1X\Kn G2 ∈ RSa,p (f )
m m


for any fixed n ∈ N. For any k ∈ N, define ηk : X → [0, 1] as ηk (x) := 1 − k d(x, Kn ) ∨ 0 for
every x ∈ X. Note that ηk ∈ LIPbs (X). For each j = 1, 2, take a sequence (fj,i )i ⊆ LIPbs (X)

13
and G′j ∈ Lp (m)+ with G′j ≤ Gj such that fj,i m ⇀ f and lip (f )m ⇀ G′ weakly in Lp (m).
a j,i j
Define hk,i := ηk f1,i + (1 − ηk )f2,i ∈ LIPbs (X) for all i ∈ N. Then hm k,i ⇀ f weakly in Lp (m).
Moreover, since we can write hk,i = f1,i + (1 − ηk )(f2,i − f1,i ) = f2,i + ηk (f1,i − f2,i ), it follows
from (1.2b) and (1.2c) that


lipa (hk,i ) ≤ lipa (f1,i ) + (1 − ηk ) lipa (f1,i ) + lipa (f2,i ) + |f1,i − f2,i |lipa (1 − ηk ), (2.2a)


lipa (hk,i ) ≤ lipa (f2,i ) + ηk lipa (f1,i ) + lipa (f2,i ) + |f1,i − f2,i |lipa (ηk ). (2.2b)

The right-hand side of (2.2a) (resp. of (2.2b)) weakly converges to G′1 + (1 − ηk )m (G′1 + G′2 )
(resp. to G′2 +ηkm (G′1 +G′2 )) in Lp (m) as i → ∞. Therefore, up to a non-relabelled subsequence,
we have that lipa (hk,i )m ⇀ Hk′ weakly in Lp (m) in i → ∞, for some Hk′ ∈ Lp (m)+ satisfying

Hk′ ≤ G′1 + (1 − ηk )m (G′1 + G′2 ) ∧ G′2 + ηkm (G′1 + G′2 ) .

(2.3)

Hence, Hk′ ∈ RSa,p (f ) holds for every k ∈ N. Letting Vk := {x ∈ X : d(x, Kn ) < 1/k}, we
have that ηk = 1 on Kn and ηk = 0 on X \ Vk , thus (2.3) implies that

Hk′ ≤ 1m
Kn G1 + 1X\Vk G2 + 21Vk \Kn (G1 + G2 ) =: Hk ,
′ m ′ m ′ ′

so that Hk ∈ RSa,p (f ). Note that Hk ⇀ 1m Kn G1 + 1X\Kn G2 weakly in L (m) as k → ∞.

′ m ′ p

Since the set RSa,p (f ) is closed and convex by Lemma 2.4, it follows from Mazur’s lemma
that 1m Kn G1 + 1X\Kn G2 ∈ RSa,p (f ), as desired.
′ m ′

ii). We argue by contradiction: assume that m({G < |Df |H }) > 0 for some G ∈ RSa,p (f ).
Then the function H := G ∧ |Df |H ∈ Lp (m)+ , which belongs to RSa,p (f ) by item i), satisfies
∥H∥Lp (m) < ∥|Df |H ∥Lp (m) , thus leading to a contradiction with the minimality of |Df |H .
iii), iv). Left as an exercise.

2.2 The space W 1,p (via derivations)

Definition 2.8 (Derivation). Let (X, d, m) be a metric measure space and q ∈ (1, ∞). Then
we say that a linear operator b : LIPbs (X) → L1 (m) is a q-derivation on X provided:

i) b satisfies the Leibniz rule, i.e. b(f g) = f m b(g) + g m b(f ) for every f, g ∈ LIPbs (X).

ii) There exists a function H ∈ Lq (m)+ such that

|b(f )| ≤ H lipa (f )m for every f ∈ LIPbs (X). (2.4)

We denote by Derq (X) the set of all q-derivations on X.

The set Derq (X) is a vector space if equipped with the following pointwise operations:

(b1 + b2 )(f ) := b1 (f ) + b2 (f ) for every b1 , b2 ∈ Derq (X) and f ∈ LIPbs (X),

(λb)(f ) := λ b(f ) for every b ∈ Derq (X), λ ∈ R and f ∈ LIPbs (X).

14
To any given q-derivation b ∈ Derq (X), we associate the function |b| ∈ Lq (m)+ , given by
^
|b| := H ∈ Lq (m)+ H satisfies (2.4) .

It holds that |b| satisfies (2.4), i.e. |b(f )| ≤ |b|lipa (f )m for every f ∈ LIPbs (X). Moreover,

∥b∥Derq (X) := ∥|b|∥Lq (m) for every b ∈ Derq (X)

defines a norm ∥ · ∥Derq (X) and (Derq (X), ∥ · ∥Derq (X) ) is a Banach space (Exercise). In addition,
the space Derq (X) is a module over the ring L∞ (m) if endowed with the multiplication

(h · b)(f ) := h b(f ) for every b ∈ Derq (X), h ∈ L∞ (m) and f ∈ LIPbs (X).

Definition 2.9 (Divergence). Let (X, d, m) be a metric measure space and q ∈ (1, ∞). Then
we say that a q-derivation b ∈ Derq (X) has q-divergence div(b) ∈ Lq (m) provided
ˆ ˆ
b(f ) dm = − f m div(b) dm for every f ∈ LIPbs (X). (2.5)

We denote by Derqq (X) the set of all q-derivations on X having q-divergence.

Notice that the q-divergence div(b) of a derivation b ∈ Derqq (X) is uniquely determined
by the identity in (2.5), thanks to the density of LIPbs (X)m in Lp (m) (recall Lemma 1.19).
Moreover, it can be readily checked (Exercise) that hm · b ∈ Derqq (X) for every b ∈ Derqq (X)
and h ∈ LIPbs (X), and the Leibniz rule for the divergence is satisfied. Videlicet,

div(hm · b) = hm div(b) + b(h) for every b ∈ Derqq (X) and h ∈ LIPbs (X).

Definition 2.10 (Sobolev space via derivations). Let (X, d, m) be a metric measure space.
Let p, q ∈ (1, ∞) be conjugate exponents. Let f ∈ Lp (m) be given. Then we declare that f
belongs to W 1,p (X) if there exists a linear operator Lf : Derqq (X) → L1 (m) such that:
´ ´
i) Lf (b) dm = − f div(b) dm for every b ∈ Derqq (X).

ii) Lf (hm · b) = hm Lf (b) for every h ∈ LIPbs (X) and b ∈ Derqq (X).

iii) There exists a function G ∈ Lp (m)+ such that |Lf (b)| ≤ G|b| for every b ∈ Derqq (X).

Remark 2.11 (Uniqueness of Lf ). We claim that, given any f ∈ W 1,p (X), the operator Lf
is uniquely determined. Indeed, if L̃f : Derqq (X) → L1 (m) having the same properties, then
ˆ ˆ ˆ ˆ ˆ
h L̃f (b) dm = L̃f (h · b) dm = − f div(h · b) dm = Lf (h · b) dm = hm Lf (b) dm
m m m m

for every h ∈ LIPbs (X) and b ∈ Derqq (X). Since πm LIPbs (X) is weakly∗ dense in L∞ (m)

(Corollary 1.20), we conclude that L̃f (b) = Lf (b) for every b ∈ Derqq (X), thus L̃f = Lf . ■

15
 To any f ∈ W 1,p (X), we associate the function |Df |W ∈ Lp (m)+ , which we define as
^n o
|Df |W := G ∈ Lp (m)+ |Lf (b)| ≤ G|b| for every b ∈ Derqq (X) .

It holds that |Lf (b)| ≤ |Df |W |b| for every b ∈ Derqq (X). Moreover, we have that

1/p
∥f ∥W 1,p (X) := ∥f ∥pLp (m) + ∥|Df |W ∥pLp (m) for every f ∈ W 1,p (X)

defines a norm ∥ · ∥W 1,p (X) and (W 1,p (X), ∥ · ∥W 1,p (X) ) is a Banach space (Exercise).

Remark 2.12. Fix f ∈ LIPbs (X). Then it holds that f m ∈ W 1,p (X) and |Df m |W ≤ lipa (f )m ,
as it can be readily checked by letting Lf m (b) := b(f ) for every b ∈ Derqq (X). ■

2.3 The space N 1,p (via modulus)

Definition 2.13 (Modulus). Let (X, d, m) be a metric measure space and p ∈ (1, ∞). Given
any curve family Γ ⊆ C (X), we define the set Adm(Γ) of p-admissible densities for Γ as
 ˆ 
Adm(Γ) := ρ : X → [0, +∞] Borel ρ ≥ 1 for every γ ∈ Γ ∩ R(X) .
γ

Moreover, we define the p-modulus Modp (Γ) ∈ [0, +∞] of the curve family Γ as
ˆ 
p
Modp (Γ) := inf ρ dm ρ ∈ Adm(Γ) .

Some comments on the p-modulus are in order:

• Modp (Γ) = Modp Γ ∩ R(X) holds for every Γ ⊆ C (X).

• If Γ contains constant curves, then Modp (Γ) = +∞.

• If Γ, Γ̃ ⊆ C (X) are curve families with the property that each curve of Γ has a subcurve
in Γ̃, then Modp (Γ) ≤ Modp (Γ̃).

• If Γ ⊆ R(X), then the curve family Γ̂ := {γ̂ : γ ∈ Γ} satisfies Modp (Γ̂) = Modp (Γ),
where γ̂ denotes the arc length parameterisation of γ.

• Modp is an outer measure (Exercise).

Theorem 2.14 (Fuglede’s lemma). Let (X, d, m) be a metric measure space and p ∈ (1, ∞).
Let (fn )n ⊆ Lp (m)+ and f ∈ Lp (m)+ be such that fnm → f m in Lp (m). Then there exists a
subsequence (fnk )k such that
ˆ
lim |fnk − f | = 0 for Modp -a.e. γ ∈ R(X).
k→∞ γ

16
 ´
Proof. We extract a subsequence (fnk )k such that |fnk − f |p dm ≤ 2−(p+1)k for all k ∈ N.
´
Define ρk := |fnk − f | for every k ∈ N, as well as Γ := γ ∈ R(X) : γ ρk ↛ 0 as k → ∞ and

´
Γk := γ ∈ R(X) : γ ρk > 2−k for every k ∈ N. Since 2k ρk ∈ Adm(Γk ) and Γ ⊆ j>k Γj
 S

for every k ∈ N, we deduce that for any k ∈ N it holds that

ˆ
2pj ρpk dm ≤
X X X X
Modp (Γ) ≤ Modp (Γj ) ≤ 2pj 2−(p+1)j = 2−j = 2−k .
j>k j>k j>k j>k

Letting k → ∞, we conclude that Modp (Γ) = 0, whence the statement follows.

As a consequence of Fuglede’s lemma, one can prove (Exercise) that

[ 
Modp Γn = lim Modp (Γn ) for every Γ1 ⊆ Γ2 ⊆ . . . contained in C (X).
n→∞
n∈N

Definition 2.15 (Weak upper gradient). Let (X, d, m) be a metric measure space, p ∈ (1, ∞),
f ∈ Lp (m) and G ∈ Lp (m)+ . Then we say that G is a p-weak upper gradient of f if
ˆ
|f (γbγ ) − f (γaγ )| ≤ G for Modp -a.e. γ ∈ R(X).
γ

We denote by WUGp (f ) the set of all p-weak upper gradients of a function f ∈ Lp (m).

Remark 2.16. We claim that for any given G ∈ WUGp (f ) it holds that
ˆ
G < +∞ for Modp -a.e. γ ∈ R(X).
γ
´
Indeed, letting Γ := γ ∈ R(X) : γ G = +∞ , we have that εG ∈ Adm(Γ) for every ε > 0,

´
thus accordingly Modp (Γ) ≤ inf ε>0 εp Gp dm = 0. ■

Lemma 2.17. Let (X, d, m) be a metric measure space and p ∈ (1, ∞). Fix any f ∈ Lp (m)
and G ∈ WUGp (f ). Then for Modp -a.e. curve γ ∈ R(X) satisfying γ ∈ AC([aγ , bγ ]; X) we
have that the function [aγ , bγ ] ∋ t 7→ (f ◦ γ)(t) ∈ R is absolutely continuous and

|(f ◦ γ)′t | ≤ G(γt )|γ̇t | for a.e. t ∈ [aγ , bγ ].

Proof. We define the curve families Γ, Γ0 and Γ1 as

n o
Γ := γ ∈ R(X) γ ∈ AC([aγ , bγ ]; X) and the statement fails ,
 ˆ 
Γ0 := γ ∈ R(X) G = +∞ ,
γ
 ˆ 
Γ1 := γ ∈ R(X) |f (γbγ ) − f (γaγ )| > G .
γ
´t
If γ ∈ Γ\Γ0 , then there exist s, t ∈ [aγ , bγ ] such that s < t and |f (γt )−f (γs )| > s G(γr )|γ̇r | dr,
which means that γ|[s,t] ∈ Γ1 . This shows that every curve in Γ \ Γ0 has a subcurve in Γ1 ,

17
whence it follows that Modp (Γ \ Γ0 ) ≤ Modp (Γ1 ). Given that Modp (Γ0 ) = 0 (by Remark
2.16) and Modp (Γ1 ) = 0 (as G is a p-weak upper gradient of f ), we can conclude that

Modp (Γ) ≤ Modp (Γ \ Γ0 ) + Modp (Γ ∩ Γ0 ) ≤ Modp (Γ1 ) + Modp (Γ0 ) = 0,

which yields the statement.

Proposition 2.18 (Stability of weak upper gradients). Let (X, d, m) be a metric measure
space and p ∈ (1, ∞). Let (fn )n ⊆ Lp (m) and f ∈ Lp (m) be such that fnm → f m in Lp (m).
Let (Gn )n ⊆ Lp (m)+ and G ∈ Lp (m)+ be such that Gm m p
n → G in L (m). Assume that Gn is
a p-weak upper gradient of fn for every n ∈ N. Then there exists a function f¯ ∈ Lp (m) such
that f¯ = f holds in the m-a.e. sense and G is a p-weak upper gradient of f¯.

Proof. Due to Theorem 2.14, we have (up to a non-relabeled subsequence) that fn (x) → f (x)
´
for m-a.e. x ∈ X and γ |Gn − G| → 0 for Modp -a.e. γ. Let us define f˜: X → [−∞, +∞] as

f˜(x) := lim fn (x) for every x ∈ X.

n→∞

Note that f˜ ∼m f . Moreover, we define the curve families Γ, N1 , N2 and Γ̃ as

\ γ ∈ AC([aγ , bγ ]; X), fn ◦ γ is absolutely continuous

Γ := γ ∈ R(X) ′ | ≤ G (γ )|γ̇ | holds for a.e. t ∈ [a , b ]
,
n∈N
and |(fn ◦ γ) t n t t γ γ
 ˆ 
N1 := γ ∈ R(X) |Gn − G| ↛ 0 as n → ∞ ,
γ

N2 := γ ∈ R(X) |f˜(γt )| = +∞ for every t ∈ [aγ , bγ ] ,



Γ̃ := γ ∈ R(X) either |f˜(γaγ )| < +∞ or |f˜(γbγ )| < +∞ .



Since Gn ∈ WUGp (fn ) for all n ∈ N, we have that Modp (C (X) \ Γ) = 0. Also, Modp (N1 ) = 0
by Theorem 2.14, and ε|f˜| ∈ Adm(N2 ) for every ε > 0, thus accordingly
ˆ ˆ
Modp (N2 ) ≤ inf ε p ˜ p
|f | dm = inf ε p
|f |p dm = 0,
ε>0 ε>0

so that Modp (N2 ) = 0. In particular, N := N1 ∪ N2 satisfies Modp (N ) = 0. Next, note that

ˆ
|f˜(γbγ ) − f˜(γaγ )| ≤ G for every γ ∈ (Γ ∩ Γ̃) \ N1 . (2.6)
γ
´ ´
Indeed, we have that |f˜(γbγ ) − f˜(γaγ )| ≤ limn |fn (γbγ ) − fn (γaγ )| ≤ limn γ Gn = γ G. It
follows that the inequality in (2.6) holds also for every γ ∈ Γ \ N . Indeed, if γ ∈ Γ \ N is
given, then |f˜(γt0 )| < +∞ holds for some t0 ∈ [aγ , bγ ], so that the subcurves γ 1 := γ|[aγ ,t0 ]
and γ 2 := γ|[t0 ,bγ ] belong to (Γ ∩ Γ̃) \ N1 , thus (2.6) ensures that
ˆ ˆ ˆ
˜ ˜ ˜ ˜ ˜ ˜
|f (γbγ ) − f (γaγ )| ≤ |f (γbγ ) − f (γt0 )| + |f (γt0 ) − f (γaγ )| ≤ G+ G = G.
γ1 γ2 γ

Therefore, recalling also Remark 2.16, we can conclude that G is a p-weak upper gradient of
the function f¯ := 1{|f˜|<+∞} f˜ ∈ Lp (m), which satisfies f¯ ∼m f˜ ∼m f .

18
Corollary 2.19. Let (X, d, m) be a metric measure space and p ∈ (1, ∞). Let f ∈ Lp (m) be
given. Then WUGm m p
p (f ) := {G : G ∈ WUGp (f )} is a closed sublattice of L (m).

Proof. The closedness part of the statement is a consequence of Proposition 2.18. Let us now
show that WUGp (f ) is a sublattice of Lp (m). To this aim, fix any G1 , G2 ∈ WUGp (f ). It is
obvious that G1 ∨ G2 ∈ WUGp (f ), so it remains to check that G1 ∧ G2 ∈ WUGp (f ). Denote
 
γ ∈ AC([aγ , bγ ]; X), f ◦ γ is absolutely continuous
Γi := γ ∈ R(X)
and |(f ◦ γ)′t | ≤ Gi (γt )|γ̇t | holds for a.e. t ∈ [aγ , bγ ]

for i = 1, 2. Hence, for every curve γ ∈ Γ1 ∩ Γ2 we have that γ and f ◦ γ are absolutely
continuous, and |(f ◦γ)′t | ≤ (G1 ∧G2 )(γt )|γ̇t | for a.e. t ∈ [aγ , bγ ], thus G1 ∧G2 ∈ WUGp (f ).

It follows from Corollary 1.20 that, given any f ∈ Lp (m) with WUGp (f ) ̸= ∅, there exists
a unique element Gf ∈ WUGm p
p (f ) of minimal L (m)-norm, and that Gf is unique also in the
m-a.e. sense (i.e. Gf ≤ G for every G ∈ WUGm p (f )). Furthermore, it holds that Gf = Gf˜
whenever f, f ∈ L (m) satisfy WUGp (f ) ̸= ∅ ̸= WUGp (f˜) and f ∼m f˜; we omit to prove it.
˜ p

Definition 2.20 (Sobolev space via modulus). Let (X, d, m) be a metric measure space,
p ∈ (1, ∞) and f ∈ Lp (m). Then we declare that f belongs to N 1,p (X) provided there exists
an m-a.e. representative f¯ ∈ Lp (m) such that WUGp (f¯) ̸= ∅.

The set N 1,p (X) is a vector subspace of Lp (m). Moreover, given f ∈ N 1,p (X), we know from
the discussion preceding Definition 2.20 that there exists a (unique) function |Df |N ∈ Lp (m)+
such that |Df |N = Gf¯ for every f¯ ∈ πm −1 (f ) with WUG (f¯) ̸= ∅. We call |Df | the minimal
p N
1,p
p-weak upper gradient of f . We endow N (X) with the norm

1/p
∥f ∥N 1,p (X) := ∥f ∥pLp (m) + ∥|Df |N ∥pLp (m) for every f ∈ N 1,p (X).

It holds that (N 1,p (X), ∥ · ∥N 1,p (X) ) is a Banach space (Exercise).

2.4 The space B 1,p (via plans)

Definition 2.21 (Plan with barycenter). Let (X, d, m) be a metric measure space. Then we
say that a finite Radon measure π on C([0, 1]; X) is a plan with q-barycenter, for some
exponent q ∈ (1, ∞), if it satisfies the following properties:

i) π is concentrated on AC q ([0, 1]; X).

ii) (e0 )# π, (e1 )# π ≤ Cπ m for some constant Cπ > 0, where ej , j = 0, 1 are the evaluation
maps (Definition 1.7) and (ej )# π denotes the pushforward measure, i.e. we set

(ej )# π(E) := π(e−1

j (F )) for every E ∈ B(X).

iii) There exists a function Bar(π) ∈ Lq (m)+ , called the barycenter of π, such that
ˆˆ 1 ˆ
f (γt )|γ̇t | dt dπ(γ) = f Bar(π) dm for every f : X → [0, +∞) Borel. (2.7)
0

19
 The barycenter Bar(π) of π is uniquely determined by (2.7). We denote by Bq (X) the
collection of all plans with q-barycenter on (X, d, m).

Remark 2.22. If π ∈ Bq (X) and f, f˜ ∈ Lp (m) are such that f ∼m f˜, then it holds that

f (γ0 ) = f˜(γ0 ), f (γ1 ) = f˜(γ1 ) for π-a.e. γ ∈ AC q ([0, 1]; X).

Indeed, Definition 2.21 ii) ensures that

γ f (γj ) ̸= f˜(γj ) = (ej )# π({f ̸= f˜}) ≤ Cπ m({f ̸= f˜}) = 0



π

for all j = 0, 1. Moreover, we claim that if G, G̃ ∈ Lp (m)+ satisfy G ∼m G̃, then

G(γt )|γ̇t | = G̃(γt )|γ̇t | for π-a.e. γ ∈ AC q ([0, 1]; X) and a.e. t ∈ [0, 1].

Indeed, |G − G̃| ∈ Lp (m)+ is m-a.e. equal to 0, thus the definition of barycenter gives
ˆˆ 1 ˆˆ 1
G(γt )|γ̇t | − G̃(γt )|γ̇t | dt dπ(γ) ≤ G − G̃|(γt )|γ̇t | dt dπ(γ)
0 ˆ 0

= |G − G̃|Bar(π) dm = 0,

which implies G(γt )|γ̇t | − G̃(γt )|γ̇t | = 0 for π-a.e. γ ∈ AC q ([0, 1]; X) and a.e. t ∈ [0, 1]. ■

In light of Remark 2.22, the following definitions are results are meaningful.

Definition 2.23 (Bq -weak upper gradient). Let (X, d, m) be a metric measure space. Let
p, q ∈ (1, ∞) be conjugate exponents. Let f ∈ Lp (m) and G ∈ Lp (m)+ be given. Then we say
that G is a Bq -weak upper gradient of f provided it holds that
ˆ ˆ
|f (γ1 ) − f (γ0 )| dπ(γ) ≤ G Bar(π) dm for every π ∈ Bq (X).

Remark 2.24. The set of all Bq -weak upper gradients of a given function f ∈ Lp (m) is a
closed subset of Lp (m). Indeed, if (Gn )n is a sequence of Bq -weak upper gradients of f and
Gn → G in Lp (m) for some G ∈ Lp (m)+ , then for every π ∈ Bq (X) we have that
ˆ ˆ ˆ
|f (γ1 ) − f (γ0 )| dπ(γ) ≤ lim Gn Bar(π) dm = G Bar(π) dm
n→∞

by Hölder’s inequality, thus accordingly G is a Bq -weak upper gradient of f . ■

We will also need the following equivalent characterisations of Bq -weak upper gradients:

Proposition 2.25. Let (X, d, m) be a metric measure space. Let p, q ∈ (1, ∞) be conjugate
exponents. Fix any f ∈ Lp (m) and G ∈ Lp (m)+ . Then the following are equivalent:

i) G is a Bq -weak upper gradient of f .

20
 ii) For every π ∈ Bq (X), it holds that
ˆ 1
|f (γ1 ) − f (γ0 )| ≤ G(γt )|γ̇t | dt for π-a.e. γ ∈ AC q ([0, 1]; X). (2.8)
0

iii) For every m-a.e. representative f¯ ∈ Lp (m) of f and π ∈ Bq (X), it holds that f¯ ◦ γ has
an absolutely continuous representative f¯γ : [0, 1] → R for π-a.e. γ ∈ AC q ([0, 1]; X) and

|f¯γ′ (t)| ≤ G(γt )|γ̇t | for π-a.e. γ ∈ AC q ([0, 1]; X) and a.e. t ∈ [0, 1]. (2.9)

Proof.
i) =⇒ ii). Assume that i) holds. We argue by contradiction: suppose that (2.8) fails for
some π ∈ Bq (X). Then there exists a Borel set Γ ⊆ AC q ([0, 1]; X) with π(Γ) > 0 such that
´1
the inequality |f (γ1 ) − f (γ0 )| > 0 G(γt )|γ̇t | dt holds for every γ ∈ Γ, so that accordingly
ˆ ˆ ˆ 1
|f (γ1 ) − f (γ0 )| dπ(γ) > G(γt )|γ̇t | dt dπ(γ). (2.10)
Γ Γ 0

On the other hand, the restriction π̃ := π|Γ of π to Γ is a plan with q-barycenter (Exercise),
thus (as G is a Bq -weak upper gradient of f ) we have that
ˆ ˆ ˆˆ 1
|f (γ1 ) − f (γ0 )| dπ(γ) = |f (γ1 ) − f (γ0 )| dπ̃(γ) ≤ G(γt )|γ̇t | dt dπ̃(γ)
Γ 0
ˆ ˆ 1
= G(γt )|γ̇t | dt dπ(γ),
Γ 0

in contradiction with (2.10). Therefore, we proved that ii) holds.

ii) =⇒ iii). Omitted.
iii) =⇒ i). Assume that iii) holds. Fix any π ∈ Bq (X). It holds (Exercise) that f¯(γ0 ) = f¯γ (0)
and f¯(γ1 ) = f¯γ (1) for π-a.e. γ, thus integrating (2.9) with respect to π we obtain that
ˆ ˆ ˆ ˆ 1
¯ ¯
|f (γ1 ) − f (γ0 )| dπ(γ) = |fγ (1) − fγ (0)| dπ(γ) = f¯γ′ (t) dt dπ(γ)
0
ˆˆ 1 ˆˆ 1 ˆ
¯′
≤ |fγ (t)| dt dπ(γ) ≤ G(γt )|γ̇t | dt dπ(γ) = G Bar(π) dm.
0 0

Therefore, we proved that G is a Bq -weak upper gradient of f , thus yielding i).

Definition 2.26 (Sobolev space via plans). Let (X, d, m) be a metric measure space. Let
p, q ∈ (1, ∞) be conjugate exponents and f ∈ Lp (m). Then we declare that the function f
belongs to B 1,p (X) provided it has a Bq -weak upper gradient G ∈ Lp (m)+ .
The set B 1,p (X) is a vector subspace of Lp (m). Moreover, given f ∈ B 1,p (X), it follows
from Remark 2.22 and Proposition 2.25 iii) that the set of all Bq -weak upper gradients of f
is a closed sublattice of Lp (m), thus it admits an m-a.e. minimal element |Df |B ∈ Lp (m)+ ,
called the minimal Bq -weak upper gradient of f . We endow B 1,p (X) with the norm

1/p
∥f ∥B 1,p (X) := ∥f ∥pLp (m) + ∥|Df |B ∥pLp (m) for every f ∈ B 1,p (X).

It holds that (B 1,p (X), ∥ · ∥B 1,p (X) ) is a Banach space (Exercise).

21
2.5 Calculus rules
Theorem 2.27 (‘Axiomatic’ calculus rules). Let (X, d, m) be a metric measure space and let
p ∈ (1, ∞). Assume that V ⊆ L0 (m) and ρ : V → Lp (m)+ satisfy the following conditions:

a) V is a vector subspace of L0 (m) and

ρ(f + g) ≤ ρ(f ) + ρ(g) for every f, g ∈ V,

(2.11)
ρ(λf ) = |λ|ρ(f ) for every f ∈ V and λ ∈ R.

b) If (fn )n ⊆ V, f ∈ L0 (m), G ∈ Lp (m)+ satisfy L1 (m|B ) ∋ 1m 1

B (fn − f ) → 0 in L (m|B ) for
every B ⊆ X bounded Borel and ρ(fn ) ⇀ G weakly in Lp (m), then f ∈ V and ρ(f ) ≤ G.

c) It holds that ϕ ◦ f ∈ V and ρ(ϕ ◦ f ) ≤ Lip(ϕ)ρ(f ) for every f ∈ V and ϕ ∈ LIP(R).

Then the following properties are equivalent:

i) Weak locality. For any f, g ∈ V, it holds that ρ(f ) = ρ(g) m-a.e. on {f = g}.

ii) Strong locality. For any f ∈ V and N ⊆ R Borel with L 1 (N ) = 0, it holds that

ρ(f ) = 0 m-a.e. on f −1 (N ). (2.12)

iii) Chain rule. For any f ∈ V and ϕ ∈ LIP(R), it holds that

ρ(ϕ ◦ f ) ≤ (lip(ϕ) ◦ f )ρ(f ) m-a.e. on X. (2.13)

iv) Lattice property. The set V is a sublattice of L0 (m) and it holds that

ρ(f ∨ g) = 1m
{f ≥g} ρ(f ) + 1{f <g} ρ(g),
m
ρ(f ∧ g) = 1m
{f ≤g} ρ(f ) + 1{f >g} ρ(g)
m

for every f, g ∈ V.

v) Leibniz rule. For any f, g ∈ V ∩ L∞ (m), it holds that f g ∈ V and

ρ(f g) ≤ |f |ρ(g) + |g|ρ(f ) m-a.e. on X.

Proof. First of all, we claim that

item i) =⇒ ρ(ϕ ◦ f ) ≤ (|ϕ′ | ◦ f )ρ(f ) for every f ∈ V and ϕ ∈ C 1 (R) ∩ LIP(R). (2.14)

To prove it, assume that i) holds, and fix any f ∈ V and ϕ ∈ C 1 (R) ∩ LIP(R). For any n ∈ N,
we denote by ϕn the piecewise affine function obtained by interpolating ϕ at {k/n : k ∈ Z}.


Videlicet, for any k ∈ Z we have that ϕn (t) = ϕ(k/n) + n ϕ((k + 1)/n) − ϕ(k/n) (t − k/n) for

22
all t ∈ [k/n, (k + 1)/n]. Note that ϕn is Lip(ϕ)-Lipschitz, ϕn → ϕ uniformly and lip(ϕn ) → ϕ′
on R. It follows from i) that ρ(ϕ ◦ f ) = 0 holds m-a.e. on f −1 ({k/n : k ∈ Z}), thus accordingly
X ϕ((k + 1)/n) − ϕ(k/n)
1m{k/n<f <(k+1)/n} ρ(ϕn ◦ f ) ≤ 1m{k/n<f <(k+1)/n} ρ(f )
X
ρ(ϕn ◦ f ) =
1/n
k∈Z k∈Z

1
X
m
= {k/n<f <(k+1)/n} (lip(ϕn ) ◦ f )ρ(f ) = (lip(ϕn ) ◦ f )ρ(f ) holds m-a.e. on X
k∈Z

thanks to a). Since (lip(ϕn )◦f )ρ(f ) ≤ Lip(ϕ)ρ(f ) for every n ∈ N, the dominated convergence
theorem ensures that (lip(ϕn ) ◦ f )ρ(f ) → (|ϕ′ | ◦ f )ρ(f ) in Lp (m). In particular, up to a non-
relabeled subsequence we have that ρ(ϕn ◦ f ) ⇀ G weakly in Lp (m) for some G ∈ Lp (m)+
with G ≤ (|ϕ′ | ◦ f )ρ(f ). Given that ϕn ◦ f converges to ϕ ◦ f in the sense of b), we deduce
that ρ(ϕ ◦ f ) ≤ G ≤ (|ϕ′ | ◦ f )ρ(f ). Therefore, the claimed implication (2.14) is proved.
Let us now pass to the verification of the equivalence of the items i), ii), iii), iv) and v).
i) =⇒ ii). Fix a finite measure m̃ on (X, Σ) such that m ≪ m̃ ≤ m (recall Remark 1.2).
Consider the pushforward measure µ := f# m̃, which is a finite Borel measure (thus, a Radon
measure) on R. In particular, µ is inner regular, thus we can find a sequence (Kn )n of compact
S

subsets of N such that µ N \ n∈N Kn = 0. It follows that
 [ 
−1 −1
m f (N ) \ f (Kn ) = 0. (2.15)
n∈N

Now fix any n ∈ N and define ψk := k dist(·, Kn ) ∧ 1 ∈ LIPc (R) for every k ∈ N. Letting
´t
ϕk (t) := 0 ψk (s) ds for every t ∈ R, we have that ϕk ∈ C 1 (R) ∩ LIP(R). Observe that

L 1 ({ψk < 1}) = L 1



t ∈ R dist(t, Kn ) < 1/k →0 as k → ∞

by the dominated convergence theorem, whence it follows that for any t ∈ R

ˆ t ˆ t ˆ t
|ϕk (t) − t| = ψk (s) ds − ds ≤ |ψk (s) − 1| ds ≤ L 1 ({ψk < 1}) → 0 as k → ∞.
0 0 0

In particular, ϕk ◦ f → f in the sense of b). Moreover, we have that |ϕ′k | ≤ ∥ψk ∥L∞ (R) = 1 on
R and ϕ′k = ψk = 0 on Kn , thus (2.14) implies that

ρ(ϕk ◦ f ) ≤ (|ϕ′k | ◦ f )ρ(f ) ≤ 1m

f −1 (X\Kn ) ρ(f ) for every k ∈ N.

Therefore, we deduce from b) and the reflexivity of Lp (m) that, up to non-relabeled sub-
sequence, it holds ρ(f ) ≤ 1m f −1 (X\Kn ) ρ(f ), which implies that ρ(f ) = 0 m-a.e. on f
−1 (K ).
n
Recalling (2.15), we finally conclude that ρ(f ) = 0 m-a.e. on f −1 (N ), thus proving ii).
ii) =⇒ iii). For any n ∈ N, we consider the convolution ϕn := ϕ ∗ η1/n ∈ C 1 (R) ∩ LIP(R),
where {ηε }ε>0 is some family of mollifiers in R. Recall that ϕ is a.e. differentiable (by
Rademacher’s theorem). We also have that ϕn → ϕ uniformly and ϕ′n → ϕ′ a.e. on R. Letting
N be the complement of the set where ϕ is differentiable and ϕ′n → ϕ′ , we have L 1 (N ) = 0,
so that ρ(f ) = 0 m-a.e. on f −1 (N ) by ii). In particular, (|ϕ′n | ◦ f )ρ(f ) → (|ϕ′ | ◦ f )ρ(f ) in

23
Lp (m) by the dominated convergence theorem (as each ϕn is Lip(ϕ)-Lipschitz). We also know
from (2.14) that ρ(ϕn ◦ f ) ≤ (|ϕ′n | ◦ f )ρ(f ) for every n ∈ N. Since ϕn ◦ f → ϕ ◦ f in the sense
of b), we deduce that ρ(ϕ ◦ f ) ≤ (|ϕ′ | ◦ f )ρ(f ) = (lip(ϕ) ◦ f )ρ(f ), thus proving iii).
iii) =⇒ i). Let f, g ∈ V be given. Define ϕ ∈ LIP(R) as ϕ(t) := t ∨ 0 for every t ∈ R. Also,
define ϕn ∈ LIP(R) for every n ∈ N as ϕn (t) := t− n1 ∨0 for every t ∈ R. Since lip(ϕn )(0) = 0,





by using iii) we obtain that ρ ϕn ◦ (f − g) ≤ lip(ϕn ) ◦ (f − g) ρ(f − g) = 0 holds m-a.e. on
(f − g)−1 ({0}) = {f = g}. Thanks to ρ ϕn ◦ (f − g) ≤ ρ(f − g), we also have (up to a non-

relabeled subsequence) that ρ ϕn ◦(f −g) has a weak limit G ∈ Lp (m)+ as n → ∞, and G = 0

m-a.e. on {f = g}. Since ϕn → ϕ uniformly, we have ϕn ◦ (f − g) → ϕ ◦ (f − g) = (f − g) ∨ 0

as in b), so that ρ (f − g) ∨ 0 ≤ G. In particular, ρ (f − g) ∨ 0 = 0 m-a.e. on {f = g}.


Similarly, one can show that ρ (g − f ) ∨ 0 = 0 holds m-a.e. on {f = g}, so that




|ρ(f ) − ρ(g)| ≤ ρ(f − g) = ρ (f − g) ∨ 0 − (g − f ) ∨ 0 ≤ ρ (f − g) ∨ 0 + ρ (g − f ) ∨ 0

implies that ρ(f ) = ρ(g) m-a.e. on {f = g}, which gives i).

iii) =⇒ iv). Let f, g ∈ V be given. Note that f ∨ g = 21 | · | ◦ (f − g) + f + g ∈ V, since | · | is

a Lipschitz function. We also know from i) (which follows from iii), as we proved above) that

1m{f ≥g} ρ(f ∨ g) = 1m{f ≥g} ρ(f ), 1m{f <g} ρ(f ∨ g) = 1m{f <g} ρ(g).

Summing these two identities, we obtain that ρ(f ∨ g) = 1m {f ≥g} ρ(f ) + 1{f <g} ρ(g). A similar
m

argument gives f ∧ g ∈ V and ρ(f ∧ g) = 1m{f ≤g} ρ(f ) + 1{f >g} ρ(g). All in all, iv) is proved.
m

iv) =⇒ i). Let f, g ∈ V be given. Define hn := (f − g) ∨ n1 for every n ∈ N. Since λ1m X ∈ V

and ρ(λ1X ) = 0 for every λ ∈ R by a) and c), we have hn ∈ V and ρ(hn ) = 0 m-a.e. on
m

{f − g < 1/n} by iv). In particular, ρ(hn ) = 0 m-a.e. on {f = g} for every n ∈ N. Since

ρ(hn ) ≤ ρ(f − g) for every n ∈ N, we have (up to a non-relabeled subsequence) that ρ(hn )
weakly converges to some G ∈ Lp (m)+ with G = 0 m-a.e. on {f = g}. Also, hn → (f − g) ∨ 0



as in b), so that ρ (f − g) ∨ 0 ≤ G and thus ρ (f − g) ∨ 0 = 0 m-a.e. on {f = g}.


Similarly, one can show that ρ (g − f ) ∨ 0 = 0 m-a.e. on {f = g}. Therefore, the inequality



|ρ(f ) − ρ(g)| ≤ ρ (f − g) ∨ 0 + ρ (g − f ) ∨ 0 implies that ρ(f ) = ρ(g) m-a.e. on {f = g}.
iii) =⇒ v). First of all, let us make some preliminary observations:

• We have that f + λ ∈ V and ρ(f + λ) = ρ(f ) for every f ∈ V and λ ∈ R, thanks to a).

• The lattice property iv) ensures that both the positive part f + := f ∨0 and the negative
part f − := −(f ∧ 0) of any given function f ∈ V belong to V, with

ρ(f + ) = 1m
{f ≥0} ρ(f ) = 1{f >0} ρ(f ),
m
ρ(f − ) = 1m
{f ≤0} ρ(f ) = 1{f <0} ρ(f ).
m
(2.16)

In particular, |f | = f + + f − ∈ V and ρ(|f |) = ρ(f + ) + ρ(f − ) = ρ(f ) for every f ∈ V.

24
 • The Leibniz rule for f, g ∈ V ∩ L∞ (m) with f, g ≥ 0 together with the lattice property
imply v). Indeed, using (2.16) we deduce that for any f, g ∈ V ∩ L∞ (m) one has that

f g = (f + − f − )(g + − g − ) = f + g + − f + g − − f − g + + f − g − ∈ V,
ρ(f g) ≤ ρ(f + g + ) + ρ(f + g − ) + ρ(f − g + ) + ρ(f − g − )
≤ (f + + f − )ρ(g + ) + (g + + g − )ρ(f + ) + (f + + f − )ρ(g − ) + (g + + g − )ρ(f − )
= |f | ρ(g + ) + ρ(g − ) + |g| ρ(f + ) + ρ(f − ) = |f |ρ(g) + |g|ρ(f ).

Accordingly, it remains to check that iii) implies the Leibniz rule for every f, g ∈ V ∩ L∞ (m)
with f, g ≥ 0. To this aim, define fn := f ∨ n1 ∈ V and gn := g ∨ n1 ∈ V for every n ∈ N.
Note that fn = ηn ◦ f and gn = ηn ◦ g, where ηn ∈ LIP(R) is defined as ηn (t) := t ∨ n1 for
every t ∈ R. Now fix n ∈ N and take a Lipschitz function ϕ : R → R with ϕ(t) = log(t) for
all t ≥ 1/n. Since fn ≥ 1/n, we have that ρ(ϕ ◦ fn ) ≤ (ϕ′ ◦ fn )ρ(fn ) = ρ(fn )/fn by iii).
Similarly, ρ(ϕ ◦ gn ) ≤ ρ(gn )/gn . Next, take a Lipschitz function ψ : R → R with ψ(t) = et for


all t ≤ log ∥f ∥L∞ (m) ∥g∥L∞ (m) . We then have that fn gn = ψ(ϕ ◦ fn + ϕ ◦ gn ) ∈ V and

ρ(fn gn ) ≤ ψ ′ ◦ (ϕ ◦ fn − ϕ ◦ gn ) ρ(ϕ ◦ fn + ϕ ◦ gn )





≤ ψ ◦ (ϕ ◦ fn − ϕ ◦ gn ) ρ(ϕ ◦ fn ) + ρ(ϕ ◦ gn )
 
ρ(fn ) ρ(gn )
≤ fn gn + = gn ρ(fn ) + fn ρ(gn ).
fn gn

Since gn ρ(fn ) + fn ρ(gn ) = g 1m

{fn ,gn ≥1/n} ρ(f ) + fn 1{fn ,gn ≥1/n} ρ(g) → gρ(f ) + f ρ(g) in L (m)
m p

(by the dominated convergence theorem) and fn gn → f g in the sense of b), the reflexivity of
Lp (m) and b) ensure that f g ∈ V and ρ(f g) ≤ gρ(f ) + f ρ(g), as desired.
v) =⇒ i). Fix f, g ∈ V and denote h := f − g ∈ V. For n ∈ N, we define ϕn , ψn ∈ LIP(R) as



 −n + n−1 for every t ∈ (−∞, −n],
 −1 for every t ∈ (−n, −n−1 ),
 t+n


ϕn (t) := 0 for every t ∈ [−n−1 , n−1 ],
t − n−1 for every t ∈ (n−1 , n),





 n − n−1

for every t ∈ [n, +∞)

and ψn := (n−1 − | · |) ∨ 0. It follows from v) and the identity (ψn ϕn ) ◦ h = 0 that

ρ(ϕn ◦ h) ≤ ρ (1 − ψn ◦ h)(ϕn ◦ h) + ρ (ψn ◦ h)(ϕn ◦ h)
≤ (1 − ψn ◦ h)ρ(ϕn ◦ h) + (|ϕn | ◦ h)ρ(1 − ψn ◦ h).

In particular, as ψn (0) = n−1 and ϕn (0) = 0, we get that ρ(ϕn ◦ h) ≤ (1 − n−1 )ρ(ϕn ◦ h) m-a.e.
on h−1 (0), so that ρ(ϕn ◦ h) = 0 m-a.e. on h−1 (0). Furthermore, ϕn → idR uniformly on each
compact subset of R, so that ϕn ◦h → h as in b), and ρ(ϕn ◦h) ≤ ρ(h). Accordingly, b) and the
reflexivity of Lp (m) ensure that ρ(h) ≤ G for some weak sublimit G ∈ Lp (m)+ of ρ(ϕn ◦ h) n ,

which satisfies G = 0 m-a.e. on h−1 (0). We then conclude that |ρ(f ) − ρ(g)| ≤ ρ(h) = 0 m-a.e.
on h−1 (0) = {f = g}, proving i).

25
Remark 2.28. Under the assumptions of Theorem 2.27 and assuming i), ii), iii), iv), v) are
satisfied, the chain rule (2.13) self-improves to an equality, i.e. it actually holds that

ρ(ϕ ◦ f ) = (|ϕ′ | ◦ f )ρ(f ) for every f ∈ V and ϕ ∈ LIP(R). (2.17)

We leave the verification of this assertion as an exercise. Let us only point out that (2.17)
is meaningful because ϕ′ exists a.e. in R by Rademacher’s theorem, and ρ(f ) = 0 m-a.e. on
f −1 (N ) by ii), where N ⊆ R denotes the (null) set of non-differentiability points of ϕ. ■
Lemma 2.29. Let (X, d, m) be a metric measure space and p ∈ (1, ∞). Let V ⊆ Lp (m) and
ρ : V → Lp (m)+ be such that the following conditions are satisfied:
a′ ) V is a vector subspace of Lp (m) and (2.11) holds.

b′ ) If (fn )n ⊆ V, f ∈ Lp (m) and G ∈ Lp (m)+ are such that fn → f strongly in Lp (m) and
ρ(fn ) ⇀ G weakly in Lp (m), then f ∈ V and ρ(f ) ≤ G.

c′ ) ϕ ◦ f ∈ V and ρ(ϕ ◦ f ) ≤ (lip(ϕ) ◦ f )ρ(f ) for every f ∈ V and ϕ ∈ LIP(R) with ϕ(0) = 0.

d′ ) η m ∈ V and ρ(η m ) ≤ Lip(η) holds m-a.e. on X for every η ∈ LIPbs (X).

e′ ) η m f ∈ V and ρ(η m f ) ≤ |η m |ρ(f ) + |f |ρ(η m ) for every η ∈ LIPbs (X) and f ∈ V.

Then ρ is weakly local, i.e. for every f, g ∈ V it holds that

ρ(f ) = ρ(g) m-a.e. on {f = g}.

Proof. Let f, g ∈ V and h := f −g ∈ V. Fix x̄ ∈ X. Let ηk := 1−d(·, B(x̄, k)) ∨0 ∈ LIPbs (X)
for every k ∈ N, so that hk := ηkm h ∈ V by e’). Now fix k ∈ N. Define ϕ ∈ LIP(R) as
ϕ(t) := t ∨ 0 for every t ∈ R. Also, define ϕn ∈ LIP(R) for every n ∈ N as ϕn (t) := t − n1 ∨ 0

for every t ∈ R. Since ϕn (0) = lip(ϕn )(0) = 0, by using c’) we obtain that ϕn ◦ hk ∈ V and
ρ(ϕn ◦ hk ) ≤ (lip(ϕn ) ◦ hk )ρ(hk ) = 0 holds m-a.e. on {hk = 0} for every n ∈ N. Thanks to
ρ(ϕn ◦ hk ) ≤ ρ(hk ), we also have (up to a non-relabelled subsequence) that ρ(ϕn ◦ hk ) has a
weak limit G ∈ Lp (m)+ as n → ∞, and G = 0 m-a.e. on {hk = 0}. As |ϕn − ϕ| ≤ 1/n on R,
ˆ ˆ
p m(B̄(x̄, k + 1))
|ϕn ◦ hk − ϕ ◦ hk | dm = |ϕn − ϕ|p ◦ hk dm ≤ → 0 as n → ∞.
B̄(x̄,k+1) np
Therefore, ρ(hk ∨ 0) = ρ(ϕ ◦ hk ) ≤ G by b’). In particular, ρ(hk ∨ 0) = 0 m-a.e. on {hk = 0}.
Since hk = ηkm h → h in Lp (m) as k → ∞ by the dominated convergence theorem, we have
that hk ∨ 0 → h ∨ 0 in Lp (m). Moreover, by applying c’), d’) and e’) we obtain that
ρ(hk ∨ 0) = ρ(ϕ ◦ hk ) ≤ (lip(ϕ) ◦ hk )ρ(hk ) = 1m m m m
{hk ≥0} ρ(ηk h) ≤ ηk ρ(h) + |h|ρ(ηk )

≤ ρ(h) + |h| ∈ Lp (m)+ for every k ∈ N.

T
Hence, thanks to b’) we deduce that ρ(h ∨ 0) = 0 holds m-a.e. on k∈N {hk = 0} = {h = 0}.
Similarly, one can show that ρ((−h) ∨ 0) = 0 holds m-a.e. on {h = 0}, so that


|ρ(f ) − ρ(g)| ≤ ρ(h) = ρ h ∨ 0 − (−h) ∨ 0 ≤ ρ(h ∨ 0) + ρ((−h) ∨ 0)

implies that ρ(f ) = ρ(g) holds m-a.e. on {h = 0} = {f = g}. This gives the statement.

26
Corollary 2.30. Let (X, d, m) be a metric measure space and let p ∈ (1, ∞). Assume that
V ⊆ Lp (m) and ρ : V → Lp (m)+ satisfy items a’), b’), c’), d’) and e’) of Lemma 2.29. Then
there exist a unique subset V̂ of L0 (m) and a unique map ρ̂ : V̂ → Lp (m)+ such that

η m f ∈ V, 1m{η=1} ρ̂(f ) = 1m{η=1} ρ(ηm f ) for every f ∈ V̂ and η ∈ LIPbs (X). (2.18)

Moreover, V̂ and ρ̂ satisfy items a), b), c), i), ii), iii), iv) and v) of Theorem 2.27.


Proof. Fix some x̄ ∈ X. Let us define ηn := 1 − d(·, B(x̄, n)) ∨ 0 ∈ LIPbs (X) for every n ∈ N.
Let us then define the subset V̂ of L0 (m) as
 
1B(x̄,n) ρ(ηn f ) ∈ L (m) .
_
0 m m m p
V̂ := f ∈ L (m) ηn f ∈ V for every n ∈ N, ρ̂(f ) :=
n∈N

Given any f ∈ V̂ and n, m ∈ N with n ≤ m, we have that ηnm f = ηm m f = f on B(x̄, n), so

that 1mB(x̄,n) ρ(ηn f ) = 1B(x̄,n) ρ(ηm f ) by weak locality and thus 1B(x̄,n) ρ̂(f ) = 1B(x̄,n) ρ(ηn f )
m m m m m m

for every n ∈ N. Moreover, for any η ∈ LIPbs (X) we have that {η ̸= 0} ⊆ B(x̄, n) for some
n ∈ N, so that η m f = η m (ηnm f ) ∈ V by e’), and ρ̂(f ) = ρ(η m f ) m-a.e. on {η = 1} ⊆ {η ̸= 0}
by weak locality, thus proving that V̂ and ρ̂ satisfy (2.18). Now, let us prove only item b) of
Theorem 2.27, leaving the verification of the remaining part of the statement as an exercise.
Fix (fk )k ⊆ V, f ∈ L0 (m) and G ∈ Lp (m)+ such that L1 (m|B ) ∋ 1m 1
B (fk − f ) → 0 in L (m|B )
for every bounded Borel set B ⊆ X and ρ̂(fk ) ⇀ G weakly in Lp (m). Fix any n ∈ N. Then
ηnm fk → ηnm f strongly in Lp (m) as k → ∞. Also, by d’), e’) and weak locality, we get that

ρ(ηnm fk ) = 1m
B̄(x̄,n+1) ρ(ηn fk ) ≤ ρ̂(fk ) + 1B̄(x̄,n+1)\B(x̄,n) |fk | =: Gk
m n
for all k ∈ N.

Since Gnk ⇀ G+ 1m B̄(x̄,n+1)\B(x̄,n)

|f | weakly in Lp (m) as k → ∞, we deduce from b’) that (up to
a non-relabelled subsequence in k) it holds that ηnm f ∈ V and ρ(ηnm f ) ≤ G+ 1m B̄(x̄,n+1)\B(x̄,n)
|f |.
Therefore, we have that f ∈ V̂ and 1m B(x̄,n) ρ̂(f ) = 1B(x̄,n) ρ(ηn f ) ≤ 1B(x̄,n) G for every n ∈ N,
m m m

so that ρ̂(f ) ≤ G. This proves the validity of item b) of Theorem 2.27, as desired.

Theorem 2.31 (Calculus rules for metric Sobolev spaces). Let (X, d, m) be a metric measure
space and p ∈ (1, ∞). Fix any S ∈ {H, W, N, B}. Then the following properties are satified:
i) Weak locality. For f, g ∈ S 1,p (X), it holds that |Df |S = |Dg|S m-a.e. on {f = g}.

ii) Strong locality. For f ∈ S 1,p (X) and N ⊆ R Borel with L 1 (N ) = 0, it holds that

|Df |S = 0 m-a.e. on f −1 (N ).

iii) Chain rule. For f ∈ S 1,p (X) and ϕ ∈ LIP(R), it holds that ϕ ◦ f − ϕ(0) ∈ S 1,p (X) and


D ϕ ◦ f − ϕ(0) S ≤ (lip(ϕ) ◦ f )|Df |S m-a.e. on X.

iv) Lattice property. The space S 1,p (X) is a sublattice of Lp (m) and it holds that

|D(f ∨ g)|S = 1m
{f ≥g} |Df |S + 1{f <g} |Dg|S ,
m
|D(f ∧ g)|S = 1m
{f ≤g} |Df |S + 1{f >g} |Dg|S
m

for every f, g ∈ S 1,p (X).

27
 v) Leibniz rule. For any f, g ∈ S 1,p (X) ∩ L∞ (m), it holds that f g ∈ S 1,p (X) and

|D(f g)|S ≤ |f ||Dg|S + |g||Df |S m-a.e. on X.

Proof. Define VS := S 1,p (X) and ρS (f ) := |Df |S for every f ∈ S 1,p (X). To prove the state-
ment, it suffices to show that VS and ρS satisfy the assumptions a’), b’), c’), d’) and e’) of
Lemma 2.29. Indeed, if this is the case, then we know that the space V̂S and the map ρ̂S
given by Corollary 2.30 satisfy all the properties listed in Theorem 2.27, whence the weak
locality, strong locality, chain rule, lattice property and Leibniz rule for (VS , ρS ) easily follow.
Accordingly, we are left with showing that VS and ρS satisfy the assumptions of Lemma
2.29. We only deal with S = B, leaving the other three cases as an exercise. It can be readily
checked that (VB , ρB ) satisfies items a’) and d’) of Lemma 2.29. Furthermore:
b’). Let (fn )n ⊆ B 1,p (X), f ∈ Lp (m) and G ∈ Lp (m)+ be such that fn → f strongly and
|Dfn |B ⇀ G weakly in Lp (m). Fix any π ∈ Bq (X). Using Hölder’s inequality, can estimate
ˆ ˆ
|fn (γ1 ) − fn (γ0 )| dπ(γ) − |f (γ1 ) − f (γ0 )| dπ(γ)
ˆ ˆ
≤ |fn (γ1 ) − f (γ1 )| dπ(γ) + |fn (γ0 ) − f (γ0 )| dπ(γ)
ˆ ˆ
≤ |fn − f | d(e1 )# π + |fn − f | d(e0 )# π
ˆ 1/p ˆ 1/p
p p 1/p
≤ |fn − f | d(e1 )# π + |fn − f | d(e0 )# π ≤ 2Cπ ∥fn − f ∥Lp (m) dm → 0
´ ´
as n → ∞. Moreover, we have that |Dfn |B Bar(π) dm → G Bar(π) dm as n → ∞. Hence,
´ ´
letting n → ∞ in the inequality |fn (γ1 ) − fn (γ0 )| dπ(γ) ≤ |Dfn |B Bar(π) dm, we get that
ˆ ˆ
|f (γ1 ) − f (γ0 )| dπ(γ) ≤ G Bar(π) dm for every π ∈ Bq (X),

which means that f ∈ B 1,p (X) and |Df |B ≤ G. This proves the validity of b’).
c’). Fix f ∈ B 1,p (X) and ϕ ∈ LIP(R) with ϕ(0) = 0. Take a representative f¯ ∈ Lp (m) of f .
Fix also π ∈ Bq (X). Then, letting f¯γ be as in Proposition 2.25 iii), for π-a.e. γ ∈ AC q ([0, 1]; X)
we have that f¯γ is absolutely continuous and |f¯γ′ (t)| ≤ |Df |B (γt )|γ̇t | for a.e. t ∈ [0, 1], whence
it follows that ϕ ◦ f¯γ is absolutely continuous and, as one can readily estimate,

|(ϕ ◦ f¯γ )′ (t)| ≤ lip(ϕ) f¯γ (t) |f¯γ′ (t)| ≤ (lip(ϕ) ◦ f¯)(γt )|Df |B (γt )|γ̇t |


for a.e. t ∈ [0, 1].

Thanks to Proposition 2.25, this gives ϕ ◦ f ∈ B 1,p (X) and |D(ϕ ◦ f )|B ≤ (lip(ϕ) ◦ f )|Df |B .
e’). Fix f ∈ B 1,p (X) and η ∈ LIPbs (X). Take a representative f¯ ∈ Lp (m) of f . One can check
that η m ∈ B 1,p (X) and |Dη m |B ≤ lip(η)m ∈ L∞ (m). Fix any π ∈ Bq (X). Then, letting f¯γ be
as in Proposition 2.25 iii), for π-a.e. curve γ ∈ AC q ([0, 1]; X) we have that f¯γ and ηγ := η ◦ γ
are absolutely continuous, and that |(f¯γ )′ (t)| ≤ |Df |B (γt )|γ̇t | and |ηγ′ (t)| ≤ |Dη m |B (γt )|γ̇t |
hold for a.e. t ∈ [0, 1]. It follows that the function ηγ f¯γ is absolutely continuous and

|(ηγ f¯γ )′ (t)| = ηγ (t)f¯γ′ (t) + f¯γ (t)ηγ′ (t) ≤ η m |Df |B + |f ||Dη m |B (γt )|γ̇t |


for a.e. t ∈ [0, 1].

28
By Proposition 2.25 and the fact that η m f ∈ Lp (m) and η m |Df |B + |f ||Dη m |B ∈ Lp (m), we
conclude that η m f ∈ B 1,p (X) and |D(η m f )|B ≤ η m |Df |B + |f ||Dη m |B , thus proving e’).

2.6 Equivalence of metric Sobolev spaces

Theorem 2.32 (Equivalence of metric Sobolev spaces). Let (X, d, m) be a metric measure
space and p ∈ (1, ∞). Then it holds that

H 1,p (X) = W 1,p (X) ⊆ N 1,p (X) ⊆ B 1,p (X),

with |Df |N ≤ |Df |W = |Df |H for all f ∈ H 1,p (X) and |Df |B ≤ |Df |N for all f ∈ N 1,p (X).
If in addition the metric space (X, d) is complete, then it also holds that

H 1,p (X) = N 1,p (X) = B 1,p (X),

with |Df |B = |Df |N = |Df |H for all f ∈ H 1,p (X).

The full proof of Theorem 2.32 is rather involved, thus we will prove only some of the
implications. Theorem 2.32 follows from the ensuing results.

Theorem 2.33 (H 1,p = W 1,p ). Let (X, d, m) be a metric measure space and p ∈ (1, ∞). Then
H 1,p (X) = W 1,p (X) and |Df |W = |Df |H for every f ∈ H 1,p (X).

Partial proof. We will prove only the inclusion H 1,p (X) ⊆ W 1,p (X) and the validity of the
inequality |Df |W ≤ |Df |H for every f ∈ H 1,p (X). Let f ∈ H 1,p (X) be given.

Proposition 2.34. Let (X, d, m) be a metric measure space. Let p, q ∈ (1, ∞) be conjugate
exponents. Fix any π ∈ Bq (X). Then there exists a unique derivation bπ ∈ Derqq (X) such that
ˆ ˆˆ 1
g bπ (f ) dm = g(γt )(f ◦ γ)′ (t) dt dπ(γ) for every f ∈ LIPbs (X) and g ∈ Lp (m)
0

Moreover, it holds that

ˆ ˆˆ 1
g|bπ | dm ≤ g(γt )|γ̇t | dt dπ(γ) for every g ∈ Lp (m)+ ,
ˆ ˆ 0
h div(bπ ) dm = h(γ1 ) − h(γ0 ) dπ(γ) for every h ∈ Lp (m).

Proof. Left as an exercise.

Corollary 2.35 (W 1,p ⊆ B 1,p ). Let (X, d, m) be a metric measure space and p ∈ (1, ∞).
Then it holds that W 1,p (X) ⊆ B 1,p (X) and |Df |B ≤ |Df |W for every f ∈ W 1,p (X).

Proof. Fix f ∈ W 1,p (X) and π ∈ Bq (X). Define Γ := γ ∈ AC q ([0, 1]; X) : f (γ1 ) ≥ f (γ0 ) .


Consider the plans π ± ∈ Bq (X), that are given by π + := π|Γ and π − := π|AC q ([0,1];X)\Γ . Let

29
bπ± ∈ Derqq (X) be the derivation associated to π ± as in Proposition 2.34. Using the fact that
f ∈ W 1,p (X) and Proposition 2.34, we obtain that
ˆ ˆ ˆ ˆ
f (γ1 ) − f (γ0 ) dπ + (γ) = f div(bπ+ ) dm = − Lf (bπ+ ) dm ≤ |Df |W |bπ+ | dm
ˆˆ 1
≤ |Df |W (γt )|γ̇t | dt dπ + (γ),
0
´ ´´ 1
and similarly f (γ0 ) − f (γ1 ) dπ − (γ) ≤ 0 |Df |W (γt )|γ̇t | dt dπ − (γ). It follows that
ˆ ˆ ˆ
|f (γ1 ) − f (γ0 )| dπ(γ) = f (γ1 ) − f (γ0 ) dπ + (γ) + f (γ0 ) − f (γ1 ) dπ − (γ)
ˆˆ 1 ˆˆ 1
≤ |Df |W (γt )|γ̇t | dt dπ + (γ) + |Df |W (γt )|γ̇t | dt dπ − (γ)
0 0
ˆˆ 1
= |Df |W (γt )|γ̇t | dt dπ(γ),
0

which shows that f ∈ B 1,p (X) and |Df |B ≤ |Df |W .

Lemma 2.36. Let (X, d, m) be a metric measure space. Let p, q ∈ (1, ∞) be conjugate expo-
nents. Let Γ be a Borel subset of AC q ([0, 1]; X) such that Modp (Γ) = 0. Then

π(Γ) = 0 for every π ∈ Bq (X).

´1
Proof. Fix π ∈ Bq (X) and ρ ∈ Adm(Γ). Since Γ̃ := γ ∈ AC q ([0, 1]; X) : 0 ρ(γt )|γ̇t | dt ≥ 1


is a Borel subset of C([0, 1]; X) containing Γ, using Hölder’s inequality we obtain that
ˆ ˆˆ 1 ˆ
π(Γ) ≤ π(Γ̃) = dπ ≤ ρ(γt )|γ̇t | dt dπ(γ) = ρ Bar(π) dm
Γ̃ 0
ˆ 1/p  ˆ 1/q
p q
≤ ρ dm Bar(π) dm .

Taking the infimum over all ρ ∈ Adm(Γ), we get π(Γ) ≤ ∥Bar(π)∥Lq (m) Modp (Γ)1/p = 0.

Remark 2.37. The proof of Lemma 2.36 shows that

π(Γ)p
 
Modp (Γ) ≥ sup π ∈ Bq (X), Bar(π) ̸= 0
∥Bar(π)∥pLq (m)

for every Borel subset Γ of AC q ([0, 1]; X). ■

Corollary 2.38 (N 1,p ⊆ B 1,p ). Let (X, d, m) be a metric measure space and p ∈ (1, ∞). Then
it holds that N 1,p (X) ⊆ B 1,p (X) and |Df |B ≤ |Df |N for every f ∈ N 1,p (X).

Proof. Let f ∈ N 1,p (X) be given. Thanks to Proposition 2.18, we can find an m-a.e. repre-
sentative f¯ ∈ Lp (m) of f and a function G ∈ WUGp (f¯) such that G ∼m |Df |N . Hence,
 ˆ 1 
q ¯ ¯
Γ := γ ∈ AC ([0, 1]; X) |f (γ1 ) − f (γ0 )| > G(γt )|γ̇t | dt
0

30
is a Borel subset of C([0, 1]; X) satisfying Modp (Γ) = 0. In particular, given any π ∈ B(X),
´1
we know from Lemma 2.36 that π(Γ) = 0. This means that |f¯(γ1 ) − f¯(γ0 )| ≤ 0 G(γt )|γ̇t | dt
holds for π-a.e. γ. Integrating with respect to π, we deduce that
ˆ ˆˆ 1
|f (γ1 ) − f (γ0 )| dπ(γ) ≤ |Df |N (γt )|γ̇t | dt dπ(γ) for every π ∈ Bq (X),
0

thus proving that f ∈ B 1,p (X) and |Df |B ≤ |Df |N .

Theorem 2.39 (B 1,p ⊆ H 1,p ). Let (X, d, m) be a metric measure space, with (X, d) complete.
Let p ∈ (1, ∞) be given. Then B 1,p (X) ⊆ H 1,p (X) and |Df |H ≤ |Df |B for every f ∈ B 1,p (X).

Proof. Omitted.

3 Bibliographical notes
Some general references for the material that is discussed in these notes are [18, 6, 17, 5].
Some more specific references are the following:

• See [15, 14] for more about the notions treated in Section 1.1. A standard reference for
Section 1.2 is the monograph [7]. For the material of Section 1.3, see for example [8].
See [8] or [12] also for the topics of Section 1.4; the proof of Theorem 1.14 is taken from
[2, Theorem 1.1.2]. The presentation in Section 1.5 is taken essentially from [18, 17].

• The approach H 1,p via relaxation discussed in Section 2.1 is due to [9]. The notion we
presented here is a variant of it that was introduced and studied starting from [3].

• The notion of derivation we considered in Section 2.2 is due to [11, 10], in turn inspired
by [22]. The corresponding notion of Sobolev space W 1,p was studied in [11, 5].

• The theory of the so-called Newtonian–Sobolev spaces N 1,p discussed in Section 2.3 is
due to [21]. See also [1].

• The Sobolev spaces B 1,p from Section 2.4 were first introduced in [4, 3], with a different
notion of plans. The concept of plan with barycenter (up to some small technical
differences) was introduced in [1] and later used in [20] to define a metric Sobolev space.

• The technical results of Section 2.5 seem to be new in this form, but they are just an
adaptation of similar arguments that can be found in the proof of [16, Theorem 2.2.6].

• The various equivalence results of Section 2.6 can be obtained by combining several
theorems from [3] and [20, 13, 19].

31
References
[1] L. Ambrosio, S. Di Marino, and G. Savaré. On the duality between p-modulus and
probability measures. J. Eur. Math. Soc. (JEMS), 17(8):1817–1853, 2015.

[2] L. Ambrosio, N. Gigli, and G. Savaré. Gradient flows in metric spaces and in the space
of probability measures. Lectures in Mathematics ETH Zürich. Birkhäuser Verlag, Basel,
second edition, 2008.

[3] L. Ambrosio, N. Gigli, and G. Savaré. Density of Lipschitz functions and equivalence of
weak gradients in metric measure spaces. Rev. Mat. Iberoam., 29(3):969–996, 2013.

[4] L. Ambrosio, N. Gigli, and G. Savaré. Calculus and heat flow in metric measure spaces
and applications to spaces with Ricci bounds from below. Invent. Math., 195(2):289–391,
2014.

[5] L. Ambrosio, T. Ikonen, D. Lučić, and E. Pasqualetto. Metric Sobolev spaces I: equiva-
lence of definitions. Milan Journal of Mathematics, pages 1–93, 2024.

[6] A. Björn and J. Björn. Nonlinear potential theory on metric spaces, volume 17 of EMS
Tracts in Mathematics. European Mathematical Society (EMS), Zürich, 2011.

[7] V. I. Bogachev. Measure theory. Vol. I, II. Springer-Verlag, Berlin, 2007.

[8] D. Burago, Y. Burago, and S. Ivanov. A course in metric geometry, volume 33 of

Graduate Studies in Mathematics. American Mathematical Society, Providence, RI,
2001.

[9] J. Cheeger. Differentiability of Lipschitz functions on metric measure spaces. Geom.

Funct. Anal., 9(3):428–517, 1999.

[10] S. Di Marino. Recent advances on BV and Sobolev spaces in metric measure spaces. PhD
thesis, Scuola Normale Superiore (Pisa), 2014.

[11] S. Di Marino. Sobolev and BV spaces on metric measure spaces via derivations and
integration by parts. arXiv preprint, arXiv:1409.5620, 2014.

[12] J. Duda. Absolutely continuous functions with values in a metric space. Real Anal.
Exchange, 32(2):569–581, 2007.

[13] S. Eriksson-Bique. Density of Lipschitz functions in energy. Calc. Var. Partial Differential
Equations, 62(2):23, 2023. Paper No. 60.

[14] D. Fremlin. Measure Theory: Measure algebras. Volume 3. Measure Theory. Torres
Fremlin, 2011.

[15] D. H. Fremlin. Topological Riesz Spaces and Measure Theory. Cambridge: University
Press, 1974.

32
[16] N. Gigli. Nonsmooth differential geometry—an approach tailored for spaces with Ricci
curvature bounded from below. Mem. Amer. Math. Soc., 251(1196):v+161, 2018.

[17] N. Gigli and E. Pasqualetto. Lectures on Nonsmooth Differential Geometry. Springer

International Publishing, Cham, Switzerland, 2020.

[18] J. Heinonen, P. Koskela, N. Shanmugalingam, and J. T. Tyson. Sobolev spaces on metric

measure spaces. An approach based on upper gradients, volume 27 of New Mathematical
Monographs. Cambridge University Press, Cambridge, 2015.

[19] D. Lučić and E. Pasqualetto. Yet another proof of the density in energy of Lipschitz
functions. Manuscripta Mathematica, 175:421–438, 2024.

[20] G. Savaré. Sobolev spaces in extended metric-measure spaces. In New trends on analysis
and geometry in metric spaces, volume 2296 of Lecture Notes in Math., pages 117–276.
Springer, Cham, [2022].

[21] N. Shanmugalingam. Newtonian spaces: an extension of Sobolev spaces to metric mea-

sure spaces. Rev. Mat. Iberoamericana, 16(2):243–279, 2000.

[22] N. Weaver. Lipschitz algebras and derivations. II. Exterior differentiation. J. Funct.
Anal., 178(1):64–112, 2000.

33