Citation: Luigi Ambrosio, Giuseppe Savaré. Duality properties of metric Sobolev spaces and capacity[J]. Mathematics in Engineering, 2021, 3(1): 1-31. doi: 10.3934/mine.2021001
[1] | Virginia Agostiniani, Lorenzo Mazzieri, Francesca Oronzio . A geometric capacitary inequality for sub-static manifolds with harmonic potentials. Mathematics in Engineering, 2022, 4(2): 1-40. doi: 10.3934/mine.2022013 |
[2] | Patrizia Pucci, Letizia Temperini . On the concentration–compactness principle for Folland–Stein spaces and for fractional horizontal Sobolev spaces. Mathematics in Engineering, 2023, 5(1): 1-21. doi: 10.3934/mine.2023007 |
[3] | Yoshikazu Giga, Hirotoshi Kuroda, Michał Łasica . The fourth-order total variation flow in Rn. Mathematics in Engineering, 2023, 5(6): 1-45. doi: 10.3934/mine.2023091 |
[4] | Lorenzo Brasco . Convex duality for principal frequencies. Mathematics in Engineering, 2022, 4(4): 1-28. doi: 10.3934/mine.2022032 |
[5] | Hitoshi Ishii . The vanishing discount problem for monotone systems of Hamilton-Jacobi equations. Part 1: linear coupling. Mathematics in Engineering, 2021, 3(4): 1-21. doi: 10.3934/mine.2021032 |
[6] | Catharine W. K. Lo, José Francisco Rodrigues . On the obstacle problem in fractional generalised Orlicz spaces. Mathematics in Engineering, 2024, 6(5): 676-704. doi: 10.3934/mine.2024026 |
[7] | Fernando Farroni, Giovanni Scilla, Francesco Solombrino . On some non-local approximation of nonisotropic Griffith-type functionals. Mathematics in Engineering, 2022, 4(4): 1-22. doi: 10.3934/mine.2022031 |
[8] | François Murat, Alessio Porretta . The ergodic limit for weak solutions of elliptic equations with Neumann boundary condition. Mathematics in Engineering, 2021, 3(4): 1-20. doi: 10.3934/mine.2021031 |
[9] | La-Su Mai, Suriguga . Local well-posedness of 1D degenerate drift diffusion equation. Mathematics in Engineering, 2024, 6(1): 155-172. doi: 10.3934/mine.2024007 |
[10] | Gerd Grubb . The principal transmission condition. Mathematics in Engineering, 2022, 4(4): 1-33. doi: 10.3934/mine.2022026 |
Dedicated to Sandro Salsa on the occasion of his 70th birthday.
In this paper we investigate the properties of the duals of the metric Sobolev spaces H1,p(X), where X=(X,τ,d,m) is an extended metric-topological measure space and p∈(1,+∞).
In the simpler case when (X,d) is a complete and separable metric space, τ is the topology induced by the metric and m is a positive and finite Borel (thus Radon) measure on X, H1,p(X) can be defined as the natural domain of the Lp(X,m)-relaxation of the pre-Cheeger energy form
pCEp(f):=∫X(lipf(x))pdm(x),f∈Lipb(X), |
initially defined only for bounded Lipschitz functions. Here lipf(x) defines the asymptotic Lipschitz constant
lipf(x)=lim supy,z→xy≠z|f(y)−f(z)|d(y,z). |
For every function f∈H1,p(X) one can define the Cheeger energy
CEp(f):={lim infn→∞pCEp(fn):fn∈Lipb(X), fn→f strongly in Lp(X,m)} |
and the Sobolev norm
‖f‖H1,p(X):=(‖f‖pLp+CEp(f))1/p, |
thus obtaining a Banach space. It is therefore quite natural to study its dual, which we will denote by H−1,q(X).
In such a general situation, however, when we do not assume any doubling and/or Poincaré assupmptions, H1,p(X) may fail to be reflexive or separable and it is not known if the generating class of bounded Lipschitz functions is strongly dense.
As a first contribution, we will show that it could be more convenient to consider the smaller subspace H−1,qpd(X) of H−1,q(X) obtained by taking the strong closure of Lq(X,m). Linear functionals in H−1,qpd(X) are characterized by their behaviour on Lipb(X) (or on even smaller generating subalgebras) and their dual norm can also be computed by the formula
‖L‖H−1,q(X)=sup{⟨L,f⟩:f∈Lipb(X), pCEp(f)+‖f‖pLp≤1}, | (1.1) |
which is well adapted to be applied to general Borel measures μ on X.
In Sections 3 and 4 we will show that H−1,qpd(X) has three important properties:
(a) it can be identified with the predual of H1,p(X) (thus showing in particular that H1,p(X) is the dual of a separable Banach space);
(b) every positive Borel measure μ satisfying
|∫Xfdμ|≤C(pCEp(f)+‖f‖pLp)1/pfor every f∈Lipb(X) |
can be extended in a unique way to a functional Lμ∈H−1,qpd(X);
(c) every positive functional on Lipb(X) such that the supremum in (1.1) is finite can be represented by a positive Radon measure.
This last property relies on a representation formula of the dual of the Cheeger energy by (nonparametric) dynamic plans (Theorem 4.6) which is interesting by itself. As a further important application of this result, in the final section 5 we will show that negligible sets in E with respect to the Newtonian capacity Capp are also μ-negligible for every positive Borel measure with finite dual energy. As a byproduct, we can express the duality of μ with a function f in H1,p(X) in the integral form
⟨μ,f⟩=∫X˜fdμ, |
where ˜f is any good representative of f in the Newtonian space N1,p(X).
Our last application concerns the variational representation of the Newtonian capacity of a closed set F⊂X
(Capp(F))1/p=sup{μ(F):μ∈M+(X), μ(X∖F)=0, ‖Lμ‖H−1,q(X)≤1}. |
Main notation | |
(X,τ) | Hausdorff topological space |
(X,τ,d) | Extended metric-topological (e.m.t.) space, see §2.2 and Definition 2.2 |
X=(X,τ,d,m) | Extended metric-topological measure (e.m.t.m.) space, see §2.2 |
M+(X) | Positive and finite Radon measures on a Hausdorff topological space X, §2.1 |
B(X) | Borel subsets of X |
f♯μ | Push forward of μ∈M(X) by a (Lusin μ-measurable) map f:X→Y, (2.1) |
Cb(X,τ), Cb(X) | τ-continuous and bounded real functions on X |
Lipb(X,τ,d) | Bounded, τ-continuous and d-Lipschitz real functions on X, (2.2) |
lipf(x) | Asymptotic Lipschitz constant of f at a point x, (2.4) |
A | Compatible unital sub-algebra of Lipb(X,τ,d), Definition 2.3 |
Lq(σ|μ) | q-Entropy functionals on Radon measures, (4.2) |
C([0,1];(X,τ)) | τ-continuous curves defined in [0,1] with values in X, §2.4 |
τC, dC | Compact open topology and extended distance on C([0,1];X), §2.4 |
BVC([0,1];X) | Continuous curves with d-bounded variation, §2.4 |
RA(X) | Continuous and rectifiable arcs, §2.4 |
e(⋅,t),et(⋅),e[⋅] | Evaluation maps along curves and arcs, §2.4 |
τA, dA | Quotient topology and extended distance on RA(X), §2.4 |
Rγ | Arc-length reparametrization of a rectifiable arc γ, §2.4 |
∫γf | Integral of a function f along a rectifiable curve (or arc) γ, §2.4 |
ℓ(γ) | length of γ, §2.4 |
νγ | Radon measure in M+(X) induced by integration along a rectifiable arc γ, (2.9) |
pCEp, CEp, CEp,κ | (pre)Cheeger energy, Definition 3.1 |
H1,p(X) | Metric Sobolev space induced by the Cheeger energy, Definition 3.1 |
|Df|⋆ | Minimal p-relaxed gradient, (3.1)-(3.2) |
\operatorname{Bar}_{q}( {\boldsymbol \pi}) | q -barycentric entropy of a dynamic plan, Definition 4.2 |
{{\mathcal{B}}_q}\left( {{\rm{RA}}\left( X \right)} \right) | Plans with barycenter in L^q(X, \mathfrak m) , Definition 4.2 |
\mathscr{D}_q(\mu_0,\mu_1) | Dual dynamic cost, (4.5) |
\operatorname{Mod}_{p}(\Gamma) | p -Modulus of a collection \Gamma\subset \mathrm{RA}(X) , Definition 5.1 |
N^{1,p}( \mathbb{X}) | Newtonian space, Definition 5.3 |
\operatorname{Cap}_p | Newtonian capacity, (5.2) |
Let (Y,\tau_Y) be a Hausdorff topological space. We will denote by {\mathrm C}_b(Y,\tau_Y) the space of \tau_Y -continuous and bounded real functions defined on Y ; \mathscr{B}(Y,\tau_Y) is the collection of the Borel subsets of Y ; we will often omit the explicit indication of the topology \tau_Y , when it will be clear from the context.
Definition 2.1 (Radon measures [22, Chap. Ⅰ, Sect. 2]). A finite Radon measure \mu: \mathscr{B}(Y,\tau_Y)\to[0,+\infty) is a Borel nonnegative \sigma -additive finite measure satisfying the following inner regularity property:
\begin{equation*} \forall\,B\in \mathscr{B}(Y,\tau_Y):\quad \mu(B) = \sup\Big\{\mu(K):K\subset B,\ {K\; compact}\Big\}. \end{equation*} |
We will denote by \mathcal M_+(Y) the collection of all the finite positive Radon measures on Y .
It is worth mentioning that every Borel measure in a Polish, Lusin, Souslin, or locally compact space with a countable base of open sets is Radon [22, Ch. Ⅱ, Sect. 3]. In particular the notation of \mathcal M_+(Y) is consistent with the standard one adopted e.g., in [4,6,24], where Polish or second countable locally compact spaces are considered.
If (Y,\tau_Y) is completely regular, the weak (or narrow) topology \tau_{ \mathcal M_+} on \mathcal M_+(Y) can be defined as the coarsest topology for which all maps
\begin{equation*} \mu\mapsto \int_Y h\, {\mathrm d}\mu\qquad\text{from $ \mathcal M_+(Y)$ into $ \mathbb{R}$} \end{equation*} |
are continuous as h:Y\to \mathbb{R} varies in {\mathrm C}_b(Y,\tau_Y) [22, p. 370,371].
Recall that a set A\subset Y is \mu -measurable, \mu\in \mathcal M_+(Y) , if there exist Borel sets B_1,B_2\in \mathscr{B}({Y,\tau_Y}) such that B_1\subset A\subset B_2 and \mu(B_2\setminus B_1) = 0 . A set is called universally (Radon) measurable if it is \mu -measurable for every Radon measure \mu\in \mathcal M_+(Y) .
Let (Z,\tau_Z) be another Hausdorff topological space. A map f:Y\to Z is Borel (resp. \mu -measurable) if for every B\in \mathscr{B}(Z) f^{-1}(B)\in \mathscr{B}(Y) (resp. f^{-1}(B) is \mu -measurable). f is Lusin \mu -measurable if for every \varepsilon>0 there exists a compact set K_ \varepsilon\subset Y such that \mu(Y\setminus K_ \varepsilon)\le \varepsilon and the restriction of f to K_ \varepsilon is continuous. A map f:Y\to Z is called universally Lusin measurable if it is Lusin \mu -measurable for every Radon measure \mu\in \mathcal M_+(Y) .
Every Lusin \mu -measurable map is also \mu -measurable. Whenever f is Lusin \mu -measurable, its push-forward
\begin{equation} f_\sharp\mu\in \mathcal M_+(Z),\quad f_\sharp\mu(B): = \mu(f^{-1}(B))\quad \text{for every Borel subset }B\subset \mathscr{B}(Z) \end{equation} | (2.1) |
induces a Radon measure in Z .
Given a power p\in (1,\infty) and a Radon measure \mu in (Y,\tau_Y) we will denote by L^p(Y,\mu) the usual Lebesgue space of class of p -summable \mu -measurable functions defined up to \mu -negligible sets.
Let (X,\tau) be a Hausdorff topological space. An extended distance is a symmetric map {\mathsf d}:X\times X\to[0,\infty] satisfying the triangle inequality and the property {\mathsf d}(x,y) = 0 iff x = y in X : we call (X, {\mathsf d}) an extended metric space. We will omit the adjective "extended" if {\mathsf d} takes real values.
Let {\mathsf d} be an extended distance on X . For every f:X\to \mathbb{R} and A\subset X we set
\begin{equation*} \mathop{{\rm{Lip}}}\nolimits(f,A): = \inf\Big\{L\in [0,\infty]:|f(y)-f(z)|\le L {\mathsf d}(y,z)\quad \text{for every }y,z\in A\Big\}. \end{equation*} |
We adopt the convention to omit the set A when A = X . We consider the class of \tau -continuous and {\mathsf d} -Lipschitz functions
\begin{equation} \begin{aligned} \mathop{{\rm{Lip}}}_b\nolimits(X,\tau, {\mathsf d}): = {}&\Big\{f\in {\mathrm C}_b(X,\tau): \mathop{{\rm{Lip}}}\nolimits(f) \lt \infty\Big\}, \end{aligned} \end{equation} | (2.2) |
and for every \kappa>0 we will also consider the subsets
\begin{equation*} \begin{aligned} \mathop{\rm Lip}_{b,\kappa}\nolimits(X,\tau, {\mathsf d}): = {}& \Big\{f\in {\mathrm C}_b(X,\tau): \mathop{{\rm{Lip}}}\nolimits(f)\le \kappa\Big\}. \end{aligned} \end{equation*} |
A particular role will be played by \mathop{\rm Lip}_{b,1}\nolimits(X,\tau, {\mathsf d}) . It is easy to check that \mathop{{\rm{Lip}}}_b\nolimits(X,\tau, {\mathsf d}) is a real and commutative sub-algebras of {\mathrm C}_b(X,\tau) with unit. According to [2, Definition 4.1], an extended metric-topological space (e.m.t. space) (X,\tau, {\mathsf d}) is characterized by a Hausdorff topology \tau and an extended distance {\mathsf d} satisfying a suitable compatibility condition.
Definition 2.2 (Extended metric-topological spaces). Let (X, {\mathsf d}) be an extended metric space, let \tau be a Hausdorff topology in X . We say that (X,\tau, {\mathsf d}) is an extended metric-topological (e.m.t.) space if:
(X1) the topology \tau is generated by the family of functions \mathop{{\rm{Lip}}}_b\nolimits(X,\tau, {\mathsf d}) ;
(X2) the distance {\mathsf d} can be recovered by the functions in \mathop{\rm Lip}_{b,1}\nolimits(X,\tau, {\mathsf d}) through the formula
\begin{equation} {\mathsf d}(x,y) = \sup\limits_{f\in \mathop{\rm Lip}_{b,1}\nolimits(X,\tau, {\mathsf d})}|f(x)-f(y)| \quad{for\; every }\;x,y\in X. \end{equation} | (2.3) |
We will say that (X,\tau, {\mathsf d}) is complete if {\mathsf d} -Cauchy sequences are {\mathsf d} -convergent. All the other topological properties usually refer to (X,\tau) .
The previous assumptions guarantee that (X,\tau) is completely regular. When an e.m.t. space (X,\tau, {\mathsf d}) is provided by a positive Radon measure \mathfrak m\in \mathcal M_+(X,\tau) we will say that
\begin{equation*} \text{the system $ \mathbb{X} = (X,\tau, {\mathsf d}, \mathfrak m)$ is an extended metric-topological measure (e.m.t.m.) space.} \end{equation*} |
Definition 2.2 yields two important properties linking {\mathsf d} and \tau : first of all
\begin{equation*} {\mathsf d}\text{ is $\tau\times\tau$-lower semicontinuous in $X\times X$}, \end{equation*} |
since it is the supremum of a family of continuous maps by (2.3). On the other hand, every {\mathsf d} -converging net (x_j)_{j\in J} indexed by a directed set J is also \tau -convergent:
\begin{equation*} \lim\limits_{j\in J} {\mathsf d}(x_j,x) = 0\quad\Rightarrow\quad \lim\limits_{j\in J}x_j = x\quad\text{w.r.t. $\tau$.} \end{equation*} |
It is sufficient to observe that \tau is the initial topology generated by \mathop{{\rm{Lip}}}_b\nolimits(X,\tau, {\mathsf d}) so that a net (x_j) is convergent to a point x if and only if
\begin{equation*} \lim\limits_{j\in J}f(x_j) = f(x)\quad\text{for every }f\in \mathop{{\rm{Lip}}}_b\nolimits(X,\tau, {\mathsf d}). \end{equation*} |
In many situations it could be useful to consider smaller subalgebras which are however sufficiently rich to recover the metric properties of an extended metric topological space (X,\tau, {\mathsf d}) .
Definition 2.3 (Compatible algebras of Lipschitz functions). Let \mathscr{A} be a unital subalgebra of \mathop{{\rm{Lip}}}_b\nolimits(X,\tau, {\mathsf d}) and let us set \mathscr{A}_\kappa: = \mathscr{A}\cap \mathop{\rm Lip}_{b,\kappa}\nolimits(X,\tau, {\mathsf d}) .
We say that \mathscr{A} is compatible with the metric-topological structure (X,\tau, {\mathsf d}) if
{\mathsf d}(x,y) = \sup\limits_{f\in \mathscr{A}_1}|f(x)-f(y)|\quad {for\; every }\;x,y\in X. |
In particular, \mathscr{A} separates the points of X .
It is not difficult to show that any compatible algebra \mathscr{A} is dense in L^p(X, \mathfrak m) [21, Lemma 2.27]. If we do not make a different explicit choice, we will always assume that an e.m.t.m. space \mathbb{X} is endowed with the canonical algebra \mathscr{A}( \mathbb{X}): = \mathop{{\rm{Lip}}}_b\nolimits(X,\tau, {\mathsf d}) .
For every f:X\to \mathbb{R} and x\in X , denoting by \mathscr{U}_x the directed set of all the \tau -neighborhoods of x , we set
\mathop{\rm{ lip }}\nolimits f(x): = \lim\limits_{U\in \mathscr{U}_x} \mathop{\rm{ Lip }}\nolimits(f,U) = \inf\limits_{U\in \mathscr{U}_x} \mathop{\rm{ Lip }}\nolimits(f,U)\quad x\in X. |
Notice that \mathop{{\rm{Lip}}}\nolimits(f,\{x\}) = 0 and therefore \mathop{{\rm{lip}}}\nolimits f(x) = 0 if x is an isolated point of X . We can also define \mathop{{\rm{lip}}}\nolimits f as
\begin{equation} \mathop{{\rm{lip}}}\nolimits f(x) = \limsup\limits_{y,z\to x\atop y\neq z}\frac{|f(y)-f(z)|}{ {\mathsf d}(y,z)}, \end{equation} | (2.4) |
where the convergence of y,z to x in (2.4) is intended with respect to the topology \tau . In particular,
\begin{equation} \mathop{{\rm{lip}}}\nolimits f(x)\ge | {\mathrm D} f|(x): = \limsup\limits_{y\to x}\frac{|f(y)-f(x)|}{ {\mathsf d}(x,y)}. \end{equation} | (2.5) |
It is not difficult to check that x\mapsto \mathop{{\rm{lip}}}\nolimits f(x) is a \tau -upper semicontinuous map and f is locally {\mathsf d} -Lipschitz in X iff \mathop{{\rm{lip}}}\nolimits f(x)<\infty for every x\in X . When (X, {\mathsf d}) is a length space, \mathop{{\rm{lip}}}\nolimits f coincides with the upper \tau -semicontinuous envelope of the local Lipschitz constant (2.5).
We collect in the next useful lemma the basic calculus properties of \mathop{{\rm{lip}}}\nolimits f .
Lemma 2.4. For every f,g, {\raise.3ex\hbox{ $\chi$ }}\in {\mathrm C}_b(X) with {\raise.3ex\hbox{ $\chi$ }}(X)\subset [0,1] we have
\begin{align*} \mathop{{\rm{lip}}}\nolimits (\alpha f+\beta g)&\le |\alpha|\, \mathop{{\rm{lip}}}\nolimits f+|\beta|\, \mathop{{\rm{lip}}}\nolimits g\quad\mathit{\text{for every}}\;\alpha,\beta\in \mathbb{R},\\ \mathop{{\rm{lip}}}\nolimits(fg)&\le |f| \mathop{{\rm{lip}}}\nolimits g+|g| \mathop{{\rm{lip}}}\nolimits f,\\ \mathop{{\rm{lip}}}\nolimits((1- {\raise.3ex\hbox{$\chi$}})f+ {\raise.3ex\hbox{$\chi$}} g)&\le (1- {\raise.3ex\hbox{$\chi$}}) \mathop{{\rm{lip}}}\nolimits f+ {\raise.3ex\hbox{$\chi$}} \mathop{{\rm{lip}}}\nolimits g+ \mathop{{\rm{lip}}}\nolimits {\raise.3ex\hbox{$\chi$}}|f-g|. \end{align*} |
Moreover, whenever \phi\in {\mathrm C}^1( \mathbb{R})
\mathop{{\rm{lip}}}\nolimits(\phi\circ f)= |\phi'\circ f| \mathop{{\rm{lip}}}\nolimits f\,. |
We briefly recap some useful results concerning the extended metric-topological structure of the space of rectifiable arcs in an e.m.t. space (X,\tau, {\mathsf d}) . We refer to [21, §3] for a more detailed discussion and for the related proofs.
For every \gamma:[0,1]\to X and t\in [0,1] we set
V_\gamma(t): = \sup\Big\{\sum\limits_{j = 1}^N {\mathsf d}(\gamma(t_j),\gamma(t_{j-1})): 0 = t_0 \lt t_1 \lt \cdots \lt t_N = t\ \Big\}, \quad \ell(\gamma): = V_\gamma(1). |
\mathrm{BVC}([0,1];X) will denote the space of {\mathsf d} -continuous maps \gamma:[0,1]\to X such that \ell(\gamma)<\infty ; notice that if \ell(\gamma) = 0 then \gamma is constant. We will also consider the set of curves with constant velocity
\begin{equation} \mathrm{BVC}_c([0,1];X): = \big\{\gamma\in \mathrm{BVC}([0,1];X): V_\gamma(t) = \ell(\gamma)t\big\}. \end{equation} | (2.6) |
Notice that for every \gamma\in \mathrm{BVC}([0,1];X) the map V_\gamma:[0,1]\to [0, \ell(\gamma)] is continuous and surjective and whenever \ell(\gamma)>0
\begin{equation} \begin{gathered} \text{there exists a unique $ \ell(\gamma)$-Lipschitz map $ R_\gamma\in \mathrm{BVC}_c([0,1];X)$ such that}\\ \gamma(t) = R_\gamma( \ell(\gamma)^{-1}V_{\gamma}(t))\quad\text{for every $t\in [0,1]$}, \end{gathered} \end{equation} | (2.7) |
with | R_\gamma'|(s) = \ell(\gamma) a.e.; when \ell(\gamma) = 0 then R_\gamma(t) = \gamma(t) is constant. We can use R_\gamma to define the integral of a bounded or nonnegative Borel function f:X\to \mathbb{R} along \gamma :
\begin{equation} \int_\gamma f = \int_0^{1} f( R_\gamma(s)) | R_\gamma'|(s)\, {\mathrm d} s = \ell(\gamma)\int_0^{1} f( R_\gamma(s))\, {\mathrm d} s. \end{equation} | (2.8) |
We also notice that (2.8) yields
\begin{equation} \int_\gamma f = \int_X f\, {\mathrm d}\nu_\gamma\quad \text{where}\quad \nu_\gamma: = \ell(\gamma)( R_\gamma)_\sharp({\mathscr L}^{1} \llcorner {[0,1]}). \end{equation} | (2.9) |
We will endow \mathrm{BVC}([0,1];X) with the compact-open topology \tau_ {\mathrm C} induced by \tau . By definition, a subbasis generating \tau_C is given by the collection of sets
S(K,V): = \Big\{\gamma\in {\mathrm C}([0,1];X):\gamma(K)\subset V\Big\},\quad K\subset [0,1]\text{ compact,}\ V\text{ $\tau$-open in $X$.} |
By [19, §46, Thm. 46.8, 46.10] if the topology \tau is induced by a distance \delta , then the topology \tau_ {\mathrm C} is induced by the uniform distance \delta_ {\mathrm C}(\gamma,\gamma'): = \sup_{t\in [a,b]}\delta(\gamma(t),\gamma'(t)) and convergence w.r.t. the compact-open topology coincides with the uniform convergence w.r.t. \delta . If moreover \tau is separable then also \tau_ {\mathrm C} is separable [14, 4.2.18].
We will denote by {\mathsf e}: \mathrm{BVC}([0,1];X)\times [0,1]\to X the evaluation map, which is defined by {\mathsf e}_t(\gamma) = {\mathsf e}(\gamma,t): = \gamma(t) for every t\in [0,1] ; {\mathsf e} is continuous. We will also adopt the notation {\mathsf e}[\gamma]: = {\mathsf e}(\{\gamma\}\times[0,1]) = \{\gamma(t):t\in [0,1]\} for the image of \gamma in X .
The extended distance {\mathsf d}:X\times X\to[0,\infty] induces the extended distance {\mathsf d}_ {\mathrm C} in \mathrm{BVC}([0,1];X) by
{\mathsf d}_ {\mathrm C}(\gamma_1,\gamma_2): = \sup\limits_{t\in [a,b]} {\mathsf d}(\gamma_1(t),\gamma_2(t)) |
and ( {\mathrm C}([0,1];X),\tau_ {\mathrm C}, {\mathsf d}_ {\mathrm C}) is an extended metric-topological space [21, Prop. 3.2].
Let us denote by \Sigma the set of continuous, nondecreasing and surjective maps \sigma:[0,1]\to[0,1] . On \mathrm{BVC}([0,1];X) we introduce the relation
\begin{gathered} \gamma_1\sim \gamma_2\quad\text{if}\quad \exists\,\sigma_i\in \Sigma:\ {\gamma_1}\circ\sigma_1 = {\gamma_2}\circ\sigma_2, \end{gathered} |
and the function
{\mathsf d}_ {\mathrm A}(\gamma_1,\gamma_2): = \inf\limits_{\sigma_i\in \Sigma} {\mathsf d}_ {\mathrm C}(\gamma_1\circ\sigma_1,\gamma_2\circ\sigma_2) \quad\text{ for every $ \gamma_i\in \mathrm{BVC}([0,1];X) $}. |
It is possible to prove that \sim is an equivalence relation [21, §3.2, Cor. 3.5] and {\mathsf d}_ {\mathrm A} satisfies
\begin{align*} {\mathsf d}_ {\mathrm A}(\gamma_1,\gamma_2) = {\mathsf d}_ {\mathrm A}( R_{\gamma_1}, R_{\gamma_2}) & = \inf\limits_{\sigma\in \Sigma\atop \sigma\text{ injective}} {\mathsf d}_ {\mathrm C}({\gamma_1},{\gamma_2}\circ\sigma) = \inf\limits_{\gamma_i'\sim\gamma_i} {\mathsf d}_{ {\mathrm C}}(\gamma_1',\gamma_2'). \end{align*} |
In particular {\mathsf d}_ {\mathrm A} satisfies the triangle inequality, is invariant with respect to \sim and \gamma\sim\gamma' if and only if {\mathsf d}_ {\mathrm A}(\gamma,\gamma') = 0 . We collect a list of useful properties [21, §3.2]:
Lemma 2.5. (a) The space \mathrm{RA}(X): = \mathrm{BVC}([0,1];X)/\sim endowed with the quotient topology \tau_ {\mathrm A} is an Hausdorff space. We will denote by \mathfrak{q}: \mathrm{BVC}([0,1];X)\to \mathrm{RA}(X) the quotient map.
(b) If the topology \tau is induced by the distance \delta then the quotient topology \tau_ {\mathrm A} is induced by \delta_ {\mathrm A} (considered as a distance between equivalence classes of curves).
(c) ( \mathrm{RA}(X),\tau_ {\mathrm A}, {\mathsf d}_ {\mathrm A}) is an extended metric-topological space.
(d) For every \gamma,\gamma'\in \mathrm{BVC}([0,1];X) we have
\gamma\sim\gamma'\quad\Leftrightarrow\quad R_{\gamma} = R_{\gamma'}, |
and all the curves \gamma' equivalent to \gamma can be described as \gamma' = R_\gamma\circ\sigma for some \sigma\in \Sigma . Moreover, if \gamma'\sim\gamma then
\ell(\gamma') = \ell(\gamma),\quad R_{\gamma'} = R_\gamma,\quad \int_\gamma f = \int_{\gamma'} f, |
so that the functions R , \ell , the evaluation maps {\mathsf e}_0, {\mathsf e}_1 , and the integral \int_\gamma f are invariant w.r.t. reparametrizations. We will still denote them by the same symbols.
(e) \ell: \mathrm{RA}(X)\to [0,+\infty] is \tau_ {\mathrm A} -lower semicontinuous and {\mathsf e}_0, {\mathsf e}_1: \mathrm{RA}(X)\to X are continuous. If f:X\to[0,+\infty] is lower semicontinuous then the map \gamma\mapsto \int_\gamma f is lower semicontinuous w.r.t. \tau_ {\mathrm A} in \mathrm{RA}(X) .
We conclude this section with a list of useful properties concerning the compactness in \mathrm{RA}(X) and the measurability of some importants maps, see [21, Thm. 3.13].
Theorem 2.6. (a) If \gamma_i , i\in I , is a converging net in \mathrm{RA}(X) with \gamma = \lim_{i\in I}\gamma_i and \lim_{i\in I} \ell(\gamma_i) = \ell(\gamma) then
\lim\limits_{i\in I} R_{\gamma_i} = R_\gamma\quad{w.r.t. }\;\tau_ {\mathrm C}, |
and for every bounded and continuous function f\in {\mathrm C}_b(X,\tau) we have
\lim\limits_{i\in I}\int_{\gamma_i}f = \int_\gamma f. |
In particular, we have
\lim\limits_{i\in I}\nu_{\gamma_i} = \nu_\gamma\quad{weakly \;in }\; \mathcal M_+(X). |
(b) The map \gamma\mapsto \nu_\gamma from \mathrm{RA}(X) to \mathcal M_+(X) is universally Lusin measurable.
(c) The map \gamma\mapsto R_\gamma is universally Lusin measurable from \mathrm{RA}(X) to \mathrm{BVC}_c([0,1];X) endowed with the topology \tau_ {\mathrm C} and it is also Borel if (X,\tau) is Souslin.
(d) If f:X\to \mathbb{R} is a bounded Borel function (or f:X\to[0,+\infty] Borel) the map \gamma\mapsto \int_\gamma f is Borel. In particular the family of measures \{\nu_\gamma\}_{\gamma\in \mathrm{RA}(X)} is Borel.
(e) If (X,\tau) is compact and \Gamma\subset \mathrm{RA}(X) satisfies \sup_{\gamma\in \Gamma} \ell(\gamma)<+\infty then \Gamma is relatively compact in \mathrm{RA}(X) w.r.t. the \tau_ {\mathrm A} topology.
(f) If (X, {\mathsf d}) is complete and \Gamma\subset \mathrm{RA}(X) satisfies the following conditions:
1) \sup_{\gamma\in \Gamma} \ell(\gamma)<+\infty ;
2) there exists a \tau -compact set K\subset X such that {\mathsf e}[\gamma]\cap K\neq \emptyset for every \gamma\in \Gamma ;
3) \{\nu_\gamma:\gamma\in \Gamma\} is equally tight, i.e. for every \varepsilon>0 there exists a \tau -compact set K_ \varepsilon\subset X such that \nu_\gamma(X\setminus K_ \varepsilon)\le \varepsilon for every \gamma\in \Gamma ,
then \Gamma is relatively compact in \mathrm{RA}(X) w.r.t. the \tau_ {\mathrm A} topology.
Notice that the third condition in the statement (f) of Theorem 2.6 implies the second one whenever \inf_{\gamma\in \Gamma}\ell(\gamma)>0 .
In this section we will always assume that \mathbb{X} = (X,\tau, {\mathsf d}, \mathfrak m) is a complete e.m.t.m. space and \mathscr{A} is a compatible sub-algebra of \mathop{{\rm{Lip}}}\nolimits_b(X,\tau, {\mathsf d}) . We also fix a summability exponent p\in (1,\infty) with conjugate q = p/(p-1) .
Let us first define the notion of Cheeger energy \mathsf{C\kern-1.5pt E}_{p} associated to \mathbb{X} , [3,5,6,10,21].
Definition 3.1 (Cheeger energy). For every \kappa\ge0 and p\in (1,\infty) we define the "pre-Cheeger" energy functionals
\mathsf{p\kern-1pt C\kern-1.5pt E}_{p}(f): = \int_X \big( \mathop{\rm{lip }}\nolimits f(x)\big)^p \, {\mathrm d} \mathfrak m, \qquad {for \;every } \; f\in \mathop{\rm{Lip }}\nolimits_b(X,\tau, {\mathsf d}), |
with \mathsf{p\kern-1pt C\kern-1.5pt E}_p(f) = +\infty if f \in L^p(X)\setminus \mathop{{\rm{Lip}}}\nolimits_b(X,\tau, {\mathsf d}) . The L^p -lower semicontinuous envelope of \mathsf{p\kern-1pt C\kern-1.5pt E}_{p} is the "strong" Cheeger energy
\mathsf{C\kern-1.5pt E}_{p}(f): = \inf \Big\{\liminf\limits_{n\to \infty}\int_X \big( \mathop{\rm{ lip }}\nolimits f_n\big)^p\, {\mathrm d} \mathfrak m: f_n\in \mathop{\rm{Lip }}\nolimits_b(X,\tau, {\mathsf d}),\ f_n\to f \;{ in }\;L^p(X, \mathfrak m)\Big\}. |
For every k\ge0 and f\in L^p(X, \mathfrak m) we also set
\mathsf{p\kern-1pt C\kern-1.5pt E}_{p, \kappa}(f): = \mathsf{p\kern-1pt C\kern-1.5pt E}_p(f)+ \kappa \|f\|_{L^p(X, \mathfrak m)}^p,\qquad \mathsf{C\kern-1.5pt E}_{p, \kappa}(f): = \mathsf{C\kern-1.5pt E}_p(f)+ \kappa \|f\|_{L^p(X, \mathfrak m)}^p. |
We denote by H^{1,p}( \mathbb{X}) the subset of L^p(X, \mathfrak m) whose elements f have finite Cheeger energy \mathsf{C\kern-1.5pt E}_{p}(f)<\infty : it is a Banach space with norm \|f\|_{ H^{1,p}( \mathbb{X})}: = \Big( \mathsf{C\kern-1.5pt E}_{p,1}(f)\Big)^{1/p}.
Remark 3.2 (The notation \mathsf{C\kern-1.5pt E} and H^{1,p} ). We used the symbol \mathsf{C\kern-1.5pt E} instead of \mathsf{Ch} (introduced by [6]) in the previous definition to stress three differences:
● the dependence on the strongest \mathop{{\rm{lip}}}\nolimits f instead of | {\mathrm D} f| ,
● the factor 1 instead of 1/p in front of the energy integral.
In this paper we will mainly adopt the "strong" approach to metric Sobolev spaces and we will use the notation H^{1,p}( \mathbb{X}) to stress this fact. We refer to [5,6] for the equivalent weak definition of W^{1,p}( \mathbb{X}) by test plan. In the final section 5 we will also use a few properties related to the intermediate (but still equivalent) Newtonian point of view, see [8,17].
It is not difficult to check that \mathsf{C\kern-1.5pt E}_p:L^p(X, \mathfrak m)\to [0,+\infty] is a convex, lower semicontinuous and p -homogeneous functional; it is the greatest L^p -lower semicontinuous functional "dominated" by \mathsf{p\kern-1pt C\kern-1.5pt E}_{p} . Notice that when \mathfrak m has not full support, two different elements f_1,f_2\in \mathop{{\rm{Lip}}}\nolimits_b(X,\tau, {\mathsf d}) may give rise to the same equivalence class in L^p(X, \mathfrak m) . In this case, \mathsf{C\kern-1.5pt E}_p can be equivalently defined starting from the functional
\widetilde{ \mathsf{p\kern-1pt C\kern-1.5pt E}}_{p}(f): = \inf\Big\{ \mathsf{p\kern-1pt C\kern-1.5pt E}_p(\tilde f):\tilde f\in \mathop{\rm{Lip }}\nolimits_b(X,\tau, {\mathsf d}),\ \tilde f = f \text{ $ \mathfrak m$-a.e.}\Big\}, |
defined on the quotient space \mathop{{\rm{Lip}}}\nolimits_b(X,\tau, {\mathsf d})/\sim_{ \mathfrak m} .
Whenever \mathsf{C\kern-1.5pt E}_p(f)<\infty one can show [5,6] that the closed convex set
\begin{equation} S_p(f): = \Big\{G\in L^p(X, \mathfrak m): \exists f_n\in \mathop{{\rm{Lip}}}\nolimits_b(X,\tau, {\mathsf d}):f_n\to f,\ \mathop{{\rm{lip}}}\nolimits f_n \rightharpoonup G\text{ in }L^p(X, \mathfrak m)\Big\} \end{equation} | (3.1) |
admits a unique element of minimal norm, the minimal relaxed gradient denoted by | {\mathrm D} f|_\star . | {\mathrm D} f|_\star is also minimal in S_p(f) with respect to the natural order structure, i.e.,
\begin{equation} | {\mathrm D} f|_\star\in S_p(f),\quad | {\mathrm D} f|_\star\le G\quad\text{for every }G\in S_p(f). \end{equation} | (3.2) |
The Cheeger energy \mathsf{C\kern-1.5pt E}_{p} admits an integral representation in terms of the minimal relaxed gradient:
\mathsf{C\kern-1.5pt E}_p(f) = \int_X | {\mathrm D} f|_\star^p(x)\, {\mathrm d} \mathfrak m(x)\quad \text{for every }f\in H^{1,p}( \mathbb{X}), |
and enjoys the following strong approximation result (see [5,6] in the case of bounded Lipschitz functions, [7] for the "metric" algebra generated by truncated distance functions and [21, Thm. 12.1] for the general case):
Theorem 3.3 (Density in energy of compatible algebras). Let \mathscr{A} be a compatible sub-algebra of \mathop{{\rm{Lip}}}\nolimits_b(X,\tau, {\mathsf d}) and let I be a closed (possibly unbounded) interval of \mathbb{R} . For every f\in H^{1,p}( \mathbb{X}) taking values in I there exists a sequence (f_n)\subset \mathscr{A} with values in I such that
f_n\to f,\quad \mathop{\rm{ lip }}\nolimits f_n\to | {\mathrm D} f|_\star\quad{strongly \;in }\;L^p(X, \mathfrak m). |
We collect a list of useful properties [6] of the minimal p -relaxed gradient.
Theorem 3.4. For every f,g\in H^{1,p}( \mathbb{X}) we have
(a) (Pointwise sublinearity) For | {\mathrm D} (\alpha f+\beta g)|_{\star}\le \alpha | {\mathrm D} f|_{\star}+\beta | {\mathrm D} g|_{\star} .
(b) (Leibniz rule) For every f,g\in H^{1,p}( \mathbb{X})\cap L^\infty(X, \mathfrak m) we have fg\in H^{1,p}( \mathbb{X}) and
\begin{equation} | {\mathrm D} {(fg)}|_\star\le |f|\,| {\mathrm D} g|_\star+|g|\,| {\mathrm D} f|_\star. \end{equation} | (3.3) |
(c) (Locality) For any Borel set N\subset \mathbb{R} with {\mathscr L}^{1}(N) = 0 we have
| {\mathrm D} f|_{\star} = 0\quad{ \mathfrak m-a.e. \;on }\;f^{-1}(N). |
In particular for every constant c\in \mathbb{R}
\text{$ | {\mathrm D} f|_{\star} = | {\mathrm D} g|_{\star} \mathfrak m -a.e. \;on \; \{f-g = c\} $}. |
(d) (Chain rule) If \phi\in \mathop{{\rm{Lip}}}\nolimits( \mathbb{R}) then \phi\circ f\in H^{1,p}( \mathbb{X}) with
\begin{equation} | {\mathrm D} (\phi\circ f)|_{\star}\le |\phi'(f)|\,| {\mathrm D} f|_{\star}. \end{equation} | (3.4) |
Equality holds in (3.4) if \phi is monotone or {\mathrm C}^1 .
Let us now study a few important properties of the Legendre transform of the p -Cheeger energy and its relation with the dual of the Sobolev space H^{1,p}( \mathbb{X}) when p\in (1,\infty) ; recall that we denote by q = p' = p/(p-1) the conjugate exponent of p . Let us first recall a simple property of p -homogeneous convex functionals (see e.g., [21, Lemma A.7].
Lemma 3.5 (Dual of positively p -homogeneous functionals). Let C be a convex cone of some vector space V , p>1 , and \phi,\psi:C\to [0,\infty] with \psi = \phi^{1/p} , \phi = \psi^p . We have the following properties:
(a) \phi is convex and positively p -homogeneous (i.e., \phi( \kappa v) = \kappa^p\phi(v) for every \kappa\ge 0 and v\in C ) in C if and only if \psi is convex and positively 1 -homogeneous on C (a seminorm, if C is a vector space and \psi is finite and even).
(b) Under one of the above equivalent assumptions, setting for every linear functional z:V\to \mathbb{R}
\begin{equation} \frac 1q\phi^*(z): = \sup\limits_{v\in C} {\langle} z,v {\rangle}-\frac 1p\phi(v),\quad \psi_*(z): = \sup\Big\{ {\langle} z,v {\rangle}:v\in C,\ \psi(v)\le 1\Big\}, \end{equation} | (3.5) |
we have
\psi_*(z) = \inf\Big\{c\ge 0: {\langle} z,v {\rangle}\le c\,\psi(v)\quad {for \;every }\;v\in C\Big\}, \quad \phi^*(z) = (\psi_*(z))^q, |
where in the first infimum we adopt the convention \inf A = +\infty if A is empty.
(c) An element v\in C attains the first supremum in (3.5) if and only if
\left\langle z,v \right\rangle = (\psi_*(z))^q = (\psi(v))^p. |
We want to study the dual functionals related to \mathsf{C\kern-1.5pt E}_{p, \kappa} and \mathsf{p\kern-1pt C\kern-1.5pt E}_{p, \kappa} . The simplest situation is provided by L^p-L^q -duality:
\begin{aligned} \frac1q \mathsf{C\kern-1.5pt E}_{p, \kappa}^*(w): = &\sup\limits_{u\in L^p} \int_X wu\, {\mathrm d} \mathfrak m-\frac 1p \mathsf{C\kern-1.5pt E}_{p, \kappa}(u) \quad\text{for every }w\in L^q(X, \mathfrak m),\\ \frac1q \mathsf{p\kern-1pt C\kern-1.5pt E}_{p, \kappa}^*(w): = &\sup\limits_{u\in \mathop{\rm{Lip }}\nolimits_b(X,\tau, {\mathsf d})} \int_X wu\, {\mathrm d} \mathfrak m-\frac 1p \mathsf{p\kern-1pt C\kern-1.5pt E}_{p, \kappa}(u) \quad\text{for every }w\in L^q(X, \mathfrak m). \end{aligned} |
By Fenchel-Moreau duality Theorem (see e.g., [9, Theorem 1.11], [13, Chap. Ⅳ]) it is immediate to check that
\begin{align} \mathsf{p\kern-1pt C\kern-1.5pt E}_{p, \kappa}^*(w)& = \mathsf{C\kern-1.5pt E}_{p, \kappa}^*(w)\quad&\text{for every }w\in L^q(X, \mathfrak m),\\ \frac 1p \mathsf{C\kern-1.5pt E}_{p, \kappa}(u)& = \sup\limits_{w\in L^q(X, \mathfrak m)} \int_X uw\, {\mathrm d} \mathfrak m-\frac 1q \mathsf{p\kern-1pt C\kern-1.5pt E}_{p, \kappa}^{*}(w) \quad&\text{for every }u\in L^p(X, \mathfrak m). \end{align} | (3.6) |
The situation is more complicated if one wants to study the dual of \mathsf{C\kern-1.5pt E}_p with respect to the Sobolev duality. Just to clarify all the possibilities we consider three normed vector spaces:
● The separable and reflexive Banach space V: = L^p(X, \mathfrak m) ;
● The vector space \mathscr{A}_p associated to a compatible algebra \mathscr{A} endowed with the norm \mathsf{p\kern-1pt C\kern-1.5pt E}^{1/p}_{p,1} .
● The Banach space W = H^{1,p}( \mathbb{X}) with the norm \mathsf{C\kern-1.5pt E}^{1/p}_{p,1} .
Notice that we do not know any information concerning the separability and the reflexivity of the Banach space H^{1,p}( \mathbb{X}) nor the (strong) density of \mathscr{A} in W . Since both \mathscr{A} and W = H^{1,p}( \mathbb{X}) are dense in V = L^p(X, \mathfrak m) , if we identify V' with L^q(X, \mathfrak m) we clearly have
L^q(X, \mathfrak m) = V'\subset ( \mathscr{A}_p) ',\quad L^q(X, \mathfrak m) = V'\subset W'\quad \text{with continuous inclusions.} |
On the other hand, every element L\in W' can be considered as a bounded linear functional on \mathscr{A}_p and thus induces an element L_{\rm restr} of ( \mathscr{A}_p) ' just by restriction, but it may happen that this identification map is not injective. Finally, since \mathsf{p\kern-1pt C\kern-1.5pt E}_{p,1} may be strictly greater than \mathsf{C\kern-1.5pt E}_{p,1} on \mathscr{A}_p , in general not all the bounded linear functionals on \mathscr{A}_p may admit an extension to W .
Taking all these facts into account, now we want to address the question of the unique extension of a given bounded linear functional L on \mathscr{A}_p to an element of the dual Sobolev space W' . We begin with a precise definition.
Definition 3.6 (The spaces H^{-1,q}( \mathbb{X}) , H^{-1,q}_{\mathrm{pd}}(\mathbb{X}) and \mathscr{A}_{q}' ). We define:
● H^{-1,q}( \mathbb{X}) as the dual W' of H^{1,p}( \mathbb{X}) ;
● H^{-1,q}_{\mathrm{pd}}(\mathbb{X}) as the subset of H^{-1,q}( \mathbb{X}) whose elements L satisfy the following condition: for every choice of f,f_n\in H^{1,p}( \mathbb{X}) , n\in \mathbb{N} , and every constant C>0
\begin{equation} \mathsf{C\kern-1.5pt E}_{p}(f_n)\le C,\quad \lim\limits_{n\to\infty}\|f_n-f\|_{L^p(X, \mathfrak m)} = 0\quad\Rightarrow\quad \lim\limits_{n\to\infty} {\langle} L,f_n {\rangle} = {\langle} L,f {\rangle} . \end{equation} | (3.7) |
● \mathscr{A}_{q}' as the set of linear functionals L on \mathscr{A} satisfying the following two conditions: there exists a constant D>0 such that
\begin{equation} \big| {\langle} L,f {\rangle}\big|\le D\Big( \mathsf{p\kern-1pt C\kern-1.5pt E}_{p,1}(f)\Big)^{1/p}\quad{for \;every }\;f\in \mathscr{A}, \end{equation} | (3.8a) |
and for every sequence f_n\in \mathscr{A} and every constant C>0
\begin{equation} \mathsf{p\kern-1pt C\kern-1.5pt E}_{p}(f_n)\le C,\quad \lim\limits_{n\to\infty}\|f_n\|_{L^p(X, \mathfrak m)} = 0\quad\Rightarrow\quad \liminf\limits_{n\to\infty}\big| {\langle} L,f_n {\rangle}\big| = 0. \end{equation} | (3.8b) |
When \mathscr{A} = \mathscr{A}( \mathbb{X}) = \mathop{{\rm{Lip}}}\nolimits_b(X,\tau, {\mathsf d}) we will write \mathscr{A}_{q}' = \mathscr{A}_q'( \mathbb{X}) .
It would not be difficult to check that if H^{1,p}( \mathbb{X}) is reflexive then \mathscr{A} is strongly dense in H^{1,p}( \mathbb{X}) and H^{-1,q}( \mathbb{X}) = H^{-1,q}_{\mathrm{pd}}(\mathbb{X})\simeq \mathscr{A}_{q}' . In the general case, only a partial result holds and we will show that H^{1,p}( \mathbb{X}) can be identified with the dual of H^{-1,q}_{\mathrm{pd}}(\mathbb{X}) , i.e., H^{-1,q}_{\mathrm{pd}}(\mathbb{X}) is a predual of H^{1,p}( \mathbb{X}) (this property justifies the index _{\rm pd} in the notation). Let us start with a first identification:
Proposition 3.7 ( \mathscr{A}_{q}'\simeq H^{-1,q}_{\mathrm{pd}}(\mathbb{X}) ). The following properties hold:
(a) \mathscr{A}_{q}' is a closed subspace of ( \mathscr{A}_p)' : in particular, it is a Banach space with the norm
\| L\|_{ \mathscr{A}_{q}'}: = \Big( \mathsf{p\kern-1pt C\kern-1.5pt E}_{p,1}^*( L)\Big)^{1/q} = \sup\Big\{ {\langle} L,f {\rangle}:f\in \mathscr{A},\ \mathsf{p\kern-1pt C\kern-1.5pt E}_{p,1}(f)\le 1\Big\}. |
(b) A linear functional L on \mathscr{A} belongs to \mathscr{A}_{q}' if and only if for every \varepsilon>0 there exists a constant \kappa>0 such that
\begin{equation} \big| {\langle} L,f {\rangle}\big|^p\le \varepsilon\, \mathsf{p\kern-1pt C\kern-1.5pt E}_{p}(f)+\kappa \|f\|_{L^p}^p\quad\mathit{\text{for every}}\quad f\in \mathscr{A}. \end{equation} | (3.9) |
In this case (3.8b) holds in the stronger form where \liminf is replaced by \limsup .
(c) Every linear functional L\in \mathscr{A}_{q}' admits a unique extension \tilde L in H^{-1,q}_{\mathrm{pd}}(\mathbb{X}). The map L\mapsto \tilde L is a surjective isometry between \mathscr{A}_{q}' and H^{-1,q}_{\mathrm{pd}}(\mathbb{X}) , which is therefore a closed subspace of H^{-1,q}( \mathbb{X}) . In particular, if L, L'\in H^{-1,q}_{\mathrm{pd}}(\mathbb{X}) coincide on \mathscr{A} then L = L' .
Proof. (a) It is sufficient to prove that \mathscr{A}_{q}' is closed in the Banach space ( \mathscr{A}_p)' . Let L be an element of the closure and for every \varepsilon>0 choose elements L_ \varepsilon\in \mathscr{A}_{q}' such that \| L- L_ \varepsilon\|_{( \mathscr{A}_p)'}\le \varepsilon . For every sequence f_n\in \mathscr{A} as in (3.8b) we have
\begin{align*} \liminf_{n\to\infty} \big| {\langle} L,f_n {\rangle}\big|\le \limsup\limits_{n\to\infty} \big| {\langle} L- L_ \varepsilon,f_n {\rangle}\big|+ \liminf_{n\to\infty}\big| {\langle} L_ \varepsilon,f_n {\rangle}\big|\le C \varepsilon. \end{align*} |
Since \varepsilon>0 is arbitrary we obtain that L\in \mathscr{A}_{q}' .
(b) If L satisfies (3.9) and f_n\in \mathscr{A} is a sequence as in (3.8b) we have
\begin{align*} \limsup\limits_{n\to\infty}\big| {\langle} L,f_n {\rangle}\big|^p \le \limsup\limits_{n\to\infty} \Big( \varepsilon\, \mathsf{p\kern-1pt C\kern-1.5pt E}_{p}(f_n)+ \kappa \|f_n\|_{L^p}^p\Big)\le \varepsilon\, C; \end{align*} |
since \varepsilon is arbitrary we deduce that \limsup\limits_{n\to\infty}\big| {\langle} L,f_n {\rangle}\big| = 0 , thus (3.8b) in the stronger form.
In order to prove the converse implication, we argue by contradiction by assuming that there exists \varepsilon>0 and a sequence f_n\in \mathscr{A} such that
\big| {\langle} L,f_n {\rangle}\big|^p\ge \varepsilon \mathsf{p\kern-1pt C\kern-1.5pt E}_p(f_n)+n \|f_n\|_{L^p}^p \gt 0. |
By possibly replacing f_n with f_n\big( \mathsf{p\kern-1pt C\kern-1.5pt E}_{p,1}(f_n)\big)^{-1/p} , it is not restrictive to assume that \mathsf{p\kern-1pt C\kern-1.5pt E}_{p,1}(f_n) = 1 ; by (3.8a) we have for n> \varepsilon
\varepsilon\le \varepsilon \mathsf{p\kern-1pt C\kern-1.5pt E}_p(f_n)+n\|f_n\|_{L^p}^p\le \big| {\langle} L,f_n {\rangle}\big|^p\le D^p |
so that \lim_{n\to\infty}\|f_n\|_{L^p} = 0 but \liminf_{n\to\infty}\big| {\langle} L,f_n {\rangle}\big|\ge \varepsilon^{1/p}>0.
(c) In order to define \tilde L we fix f\in H^{1,p}( \mathbb{X}) and any sequence f_n\in \mathscr{A} such that f_n\to f in L^p(X, \mathfrak m) with E^p: = \sup \mathsf{p\kern-1pt C\kern-1.5pt E}_p(f_n)<\infty . By (3.9), for every \varepsilon>0 there exists \kappa>0 such that
| {\langle} L,f_n-f_m {\rangle}|\le 2 \varepsilon^{1/p} E+ \kappa^{1/p} \|f_n-f_m\|_{L^p} |
which shows that the sequence n\mapsto {\langle} L,f_n {\rangle} satisfies the Cauchy condition and thus admits a limit which we denote by {\langle} \tilde L,f {\rangle} . This notation is justified by the fact that the limit does not depend on the sequence f_n : in fact, if f_n' is another sequence converging to f in L^p(X, \mathfrak m) with equibounded energy, (3.7) shows that \lim\limits_{n\to\infty} {\langle} L,f_n-f_n' {\rangle} = 0 . It is also easy to check that the map H^{1,p}( \mathbb{X})\ni f\mapsto {\langle} \tilde L,f {\rangle} is a linear functional.
In order to show that \tilde L is bounded, for every f\in H^{1,p}( \mathbb{X}) we select an optimal sequence f_n such that \mathsf{C\kern-1.5pt E}_p(f) = \lim_{n\to\infty} \mathsf{p\kern-1pt C\kern-1.5pt E}_p(f_n) : by construction
\big| {\langle} \tilde L,f {\rangle}\big| = \lim\limits_{n\to\infty} \big| {\langle} L,f_n {\rangle}\big|\le \limsup\limits_{n\to\infty}\| L\|_{ \mathscr{A}_{q}'} \Big( \mathsf{p\kern-1pt C\kern-1.5pt E}_{p,1}(f_n)\Big)^{1/p} = \| L\|_{ \mathscr{A}_{q}'}\Big( \mathsf{C\kern-1.5pt E}_{p,1}(f)\Big)^{1/p} |
so that \|\tilde L\|_{ H^{-1,q}( \mathbb{X})}\le \| L\|_{ \mathscr{A}_{q}'} . On the other hand for every f\in \mathscr{A} with \mathsf{p\kern-1pt C\kern-1.5pt E}_p(f)\le 1 by choosing the constant sequence f_n\equiv f we get
{\langle} L,f {\rangle} = {\langle} \tilde L,f {\rangle} = \|\tilde L\|_{ H^{-1,q}( \mathbb{X})}\Big( \mathsf{C\kern-1.5pt E}_{p,1}(f)\Big)^{1/p} \le \|\tilde L\|_{ H^{-1,q}( \mathbb{X})} |
since \mathsf{C\kern-1.5pt E}_p(f)\le \mathsf{p\kern-1pt C\kern-1.5pt E}_p(f)\le 1 . It follows that \| L\|_{ \mathscr{A}_{q}'}\le \|\tilde L\|_{ H^{-1,q}( \mathbb{X})} so that the extension map \iota: L\mapsto \tilde L is an isometry.
It remains to prove that the image of \iota coincides with H^{-1,q}_{\mathrm{pd}}(\mathbb{X}) . Since it is clear that H^{-1,q}_{\mathrm{pd}}(\mathbb{X})\subset \iota( \mathscr{A}_{q}') , it is sufficient to show the converse inclusion, i.e., that every element \tilde L = \iota( L) satisfies (3.7). By linearity, it is not restrictive to check (3.7) for f = 0 . If f_n\in H^{1,p}( \mathbb{X}) has equibounded Cheeger energy and \lim_{n\to\infty}\|f_n\|_{L^p} = 0 , by the very definition of the Cheeger energy and the definition of \tilde L we can find another sequence g_n\in \mathscr{A} such that
\mathsf{p\kern-1pt C\kern-1.5pt E}_p(g_n)\le \mathsf{C\kern-1.5pt E}_p(f_n)+\frac 1n, \quad \|g_n-f_n\|_{L^p}\le \frac 1n,\quad \big| {\langle} L,g_n {\rangle}- {\langle} \tilde L,f_n {\rangle}\big|\le \frac 1n. |
Since L\in \mathscr{A}_{q}' and \lim_{n\to\infty}\|g_n\|_{L^p} = 0 we have \lim_{n\to\infty} {\langle} L,g_n {\rangle} = 0 so that \lim_{n\to\infty} {\langle}\tilde L,f_n {\rangle} = 0 .
Let us now express the dual functionals by a infimal convolution.
Lemma 3.8. For every L\in H^{-1,q}( \mathbb{X}) and every \alpha \ge0 , \beta>0 we have
\begin{equation} \begin{aligned} \frac1q \mathsf{C\kern-1.5pt E}_{p,\alpha+\beta}^*( L) & = \sup\limits_{g\in H^{1,p}( \mathbb{X})} {\langle} L,g {\rangle}-\frac 1p \mathsf{C\kern-1.5pt E}_{p,\alpha}(g)-\frac {\beta} p\|g\|_{L^p}^p \\& = \min_{f\in L^q(X, \mathfrak m)} \frac 1q \mathsf{C\kern-1.5pt E}_{p,\alpha}^*( L-f)+ \frac 1{q\beta^{q/p} }\|f\|_{L^q}^q. \end{aligned} \end{equation} | (3.10) |
Proof. (3.10) is a particular case of the duality formula for the sum of two convex functions \varphi,\psi:W\to (-\infty,+\infty]
(\varphi+\psi)^*( L) = \min_{f\in W}\varphi^*( L-f)+\psi^*(f)\quad \text{for every } L\in W' |
which holds in every Banach space W whenever there exists a point w_0\in W such that \phi(w_0)<\infty and \psi is finite and continuous at w_0 by Fenchel-Rockafellar Theorem ([20], see also [9, Theorem 1.12]). Here W = H^{1,p}( \mathbb{X}) , \phi(g): = \frac 1p \mathsf{C\kern-1.5pt E}_{p,\alpha}(g) , \psi(g): = \frac \beta p\|g\|_{L^p}^p .
We collect in the next proposition a further list of useful properties. We will denote by {\mathrm J}_p:L^p(X, \mathfrak m)\to L^q(X, \mathfrak m) the duality map
( {\mathrm J}_p u)(x): = |u(x)|^{p-1}u(x),\quad \int_X u\, {\mathrm J}_p u\, {\mathrm d} \mathfrak m = \|u\|_{L^p}^p = \| {\mathrm J}_pu\|_{L^q}^q, |
and by {\mathrm A}_p: H^{1,p}( \mathbb{X})\to \mathfrak P( H^{-1,q}( \mathbb{X})) the subdifferential of the Cheeger energy with respect to the Sobolev duality
\begin{equation} L\in {\mathrm A}_p u\quad\Leftrightarrow\quad u\in H^{1,p}( \mathbb{X}),\quad {\langle} L,v-u {\rangle}\le \frac 1p \mathsf{C\kern-1.5pt E}_p(v)-\frac 1p \mathsf{C\kern-1.5pt E}_p(u)\quad \text{for every }v\in H^{1,p}( \mathbb{X}). \end{equation} | (3.11) |
Since \mathsf{C\kern-1.5pt E}_{p} is continuous in H^{1,p}( \mathbb{X}) , {\mathrm A}_pu\neq\emptyset for every u\in H^{1,p}( \mathbb{X}) [13, Chap. 1, Prop. 5.3] (notice that {\mathrm A}_p is different from the subdifferential of \mathsf{C\kern-1.5pt E}_{p} w.r.t. the L^p - L^q duality pair). The sum
{\mathrm Q}_{p, \kappa}: = {\mathrm A}_p+ \kappa {\mathrm J}_p\quad\text{is the subdifferential in $ H^{1,p}( \mathbb{X}) $ of $ \frac 1p \mathsf{C\kern-1.5pt E}_{p, \kappa} $.} |
Proposition 3.9. We have the following properties
(a) For every L\in H^{-1,q}_{\mathrm{pd}}(\mathbb{X}) and every \kappa\ge0 we have
\begin{align} \frac1q \mathsf{C\kern-1.5pt E}_{p, \kappa}^*( L) = &\sup\limits_{f\in H^{1,p}( \mathbb{X})} {\langle} L,f {\rangle}-\frac 1p \mathsf{C\kern-1.5pt E}_{p, \kappa}(f) \\ = & \sup\limits_{\phantom{f\ \,\,}f\in \mathscr{A}\phantom{( \mathbb{X})}} {\langle} L,f {\rangle}-\frac 1p \mathsf{C\kern-1.5pt E}_{p, \kappa}(f) \end{align} | (3.12) |
\begin{align} = & \sup\limits_{\phantom{f\ \,\,}f\in \mathscr{A}\phantom{( \mathbb{X})}} {\langle} L,f {\rangle}-\frac 1p \mathsf{p\kern-1pt C\kern-1.5pt E}_{p, \kappa}(f) = \frac 1q \mathsf{p\kern-1pt C\kern-1.5pt E}_{p, \kappa}^*( L). \end{align} | (3.13) |
(b) H^{-1,q}_{\mathrm{pd}}(\mathbb{X}) coincides with the (strong) closure of V' = L^q(X, \mathfrak m) in W' = H^{-1,q}( \mathbb{X}) .
(c) For every L\in H^{-1,q}_{\mathrm{pd}}(\mathbb{X}) and \kappa>0 there exists a unique solution u_ \kappa = {\mathrm Q}_{p, \kappa}^{-1}( L)\in H^{1,p}( \mathbb{X}) of the problem
\begin{equation} \min\limits_{u\in H^{1,p}( \mathbb{X})} \frac 1p \mathsf{C\kern-1.5pt E}_{p,\kappa}(u) - {\langle} L,u {\rangle} \end{equation} | (3.14) |
which satisfies
\begin{equation} {\mathrm Q}_{p, \kappa}u_ \kappa = {\mathrm A}_p u_ \kappa+ \kappa {\mathrm J}_p u_ \kappa\ni L,\quad \mathsf{C\kern-1.5pt E}^*_{p, \kappa}( L) = \mathsf{C\kern-1.5pt E}_{p, \kappa}(u_ \kappa) = {\langle} L,u_ \kappa {\rangle}. \end{equation} | (3.15) |
(d) For every L\in H^{-1,q}( \mathbb{X}) and \kappa>0 there exists a unique function f_ \kappa: = {\mathrm R}_ \kappa(L) solving the minimum problem
\begin{align} \frac 1q \mathsf{C\kern-1.5pt E}_{p, \kappa}^*( L) & = \min_{f\in L^q(X, \mathfrak m)} \frac 1q \mathsf{C\kern-1.5pt E}_p^*( L-f)+\frac 1{q \kappa^{q/p}}\|f\|_{L^q}^q \end{align} | (3.16) |
The map {\mathrm R}_ \kappa: H^{-1,q}( \mathbb{X})\to L^q(Z, \mathfrak m) is strongly continuous. Moreover, if L\in H^{-1,q}_{\mathrm{pd}}(\mathbb{X}) then f_ \kappa = {\mathrm R}_\kappa(L) = \kappa {\mathrm J}_p u_ \kappa = \kappa {\mathrm J}_p\circ {\mathrm Q}_{p, \kappa}^{-1}( L).
(e) For every L\in H^{-1,q}( \mathbb{X}) we have
\mathsf{C\kern-1.5pt E}_p^*( L) = \lim\limits_{ \kappa \downarrow0} \mathsf{C\kern-1.5pt E}_{p, \kappa}^*( L) = \sup\limits_{ \kappa \gt 0} \mathsf{C\kern-1.5pt E}_{p, \kappa}^*( L). |
Proof. (a) (3.13) (which implies (3.12)) follows by an easy approximation argument combining the definition of \mathsf{C\kern-1.5pt E}_p and the continuity property (3.7) and it follows by the same argument at the end of the proof of claim (c) of Proposition 3.7.
(b) Since H^{-1,q}_{\mathrm{pd}}(\mathbb{X}) is a closed subspace of H^{-1,q}( \mathbb{X}) and clearly contains L^q(X, \mathfrak m) , it is sufficient to prove that L^q(X, \mathfrak m) is dense in H^{-1,q}_{\mathrm{pd}}(\mathbb{X}) . For every n\in \mathbb{N} we consider the functional {\mathsf G}_n: = \mathsf{C\kern-1.5pt E}_{p,1+n^p}^* and we want to show that
\begin{equation} \limsup\limits_{n \uparrow\infty} {\mathsf G}_n( L) = 0; \end{equation} | (3.17) |
by using (3.10) (with \alpha: = 1,\beta: = n^p ), (3.17) is in fact equivalent to the density of L^q(X, \mathfrak m) in H^{-1,q}_{\mathrm{pd}}(\mathbb{X}) . By the first formula of (3.10), for every \varepsilon>0 we can find g_n\in H^{1,p}( \mathbb{X}) such that
\begin{equation} \frac 1q {\mathsf G}_n( L)\le {\langle} L,g_n {\rangle}- \frac1p \mathsf{C\kern-1.5pt E}_{p,1}(g_n)-\frac {n^p}{p} \|g_n\|_{L^p}^p+ \varepsilon. \end{equation} | (3.18) |
Since
{\langle} L,g_n {\rangle}\le \frac{2^{q/p}}{q} \mathsf{C\kern-1.5pt E}_{p,1}^*( L)+\frac1{2p} \mathsf{C\kern-1.5pt E}_{p,1}(g_n) |
and {\mathsf G}_n( L)\ge0 , we obtain
\frac1{2p} \mathsf{C\kern-1.5pt E}_{p,1}(g_n) +\frac {n^p}{p} \|g_n\|_{L^p}^p\le \varepsilon+ \frac{2^{q/p}}{q} \mathsf{C\kern-1.5pt E}_{p,1}^*( L) |
so that \mathsf{C\kern-1.5pt E}_{p,1}(g_n) is uniformly bounded and \|g_n\|_{L^p}\to0 as n\to\infty . By (3.7) we conclude that \lim_{n\to\infty} {\langle} L,g_n {\rangle} = 0 and therefore (3.18) yields \limsup_{n\to\infty} {\mathsf G}_n( L)\le \varepsilon . Since \varepsilon>0 is arbitrary, we obtain (3.17).
(c) The existence of a solution u_ \kappa\in H^{1,p}( \mathbb{X}) to (3.14) follows by (3.7) and the Direct method of the Calculus of Variations. Let us take a minimizing sequence f_n\in H^{1,p}( \mathbb{X}) such that
\begin{equation} \lim\limits_{n\to\infty} \frac 1p \mathsf{C\kern-1.5pt E}_{p, \kappa}(f_n)- {\langle} L,f_n {\rangle} = M: = \inf\limits_{f\in H^{1,p}( \mathbb{X})}\frac 1p \mathsf{C\kern-1.5pt E}_{p, \kappa}(f_n)- {\langle} L,f_n {\rangle}. \end{equation} | (3.19) |
Since f_n is uniformly bounded in H^{1,p}( \mathbb{X}) , up to extracting a suitable subsequence (still denoted by f_n ), it is not restrictive to assume that f_n is converging to a function f\in H^{1,p}( \mathbb{X}) weakly in L^p(X, \mathfrak m) and
\begin{equation} S = \lim\limits_{n\to\infty}\|f_n\|_{L^p} = \lim\limits_{n\to\infty}\Big[\frac p \kappa \Big(M + {\langle} L,f_n {\rangle}- \frac 1p \mathsf{C\kern-1.5pt E}_{p}(f_n)\Big)\Big]^{1/p}. \end{equation} | (3.20) |
We prove that f_n is a Cauchy sequence: by the uniform convexity of the L^p(X, \mathfrak m) -norm, for every \varepsilon>0 there exist S'<S<S'' such that for every h_1,h_2\in L^p(X, \mathfrak m)
\begin{equation} \|h_1\|_{L^p}\le S'',\ \|h_2\|_{L^p}\le S'',\quad \left\|\frac{h_1+h_2}2\right\|_{L^p}\ge S'\quad\Rightarrow \quad \|h_2-h_1\|\le \varepsilon. \end{equation} | (3.21) |
By (3.20) we can find \bar n\in \mathbb{N} such that for every n\ge \bar n and
\|f_n\|_{L^p}\le S'',\qquad M-\frac 1p \mathsf{C\kern-1.5pt E}_p(f_n)+ {\langle} L,f_n {\rangle}\ge \frac \kappa p (S')^p. |
For every m,n\ge \bar n we thus get
\begin{align*} M&\le \frac 1p \mathsf{C\kern-1.5pt E}_{p, \kappa}\Big(\frac 12(f_n+f_m)\Big)-\frac12 {\langle} L,f_n+f_m {\rangle} \\&\le \frac 12\Big(\frac1p \mathsf{C\kern-1.5pt E}_p(f_n)- {\langle} L,f_n {\rangle}+ \frac1p \mathsf{C\kern-1.5pt E}_p(f_m)- {\langle} L,f_m {\rangle}\Big) +\frac\kappa p\Big\|\frac{f_n+f_m}2\Big\|_{L^p}^p \\&\le M-\frac \kappa p (S')^p+\frac\kappa p\Big\|\frac{f_n+f_m}2\Big\|_{L^p}^p, \end{align*} |
and therefore
\Big\|\frac{f_n+f_m}2\Big\|_{L^p}\ge S' |
so that (3.21) yields \|f_n-f_m\|_{L^p}\le \varepsilon for every n,m\ge \bar n . We deduce that \lim_{n\to\infty} \|f_n-f\|_{L^p} = 0; since f_n is uniformly bounded in H^{1,p}( \mathbb{X}) , (3.7) yields \lim_{n\to\infty } {\langle} L,f_n {\rangle} = {\langle} L,f {\rangle} and the lower semicontinuity of the Cheeger energy yields \mathsf{C\kern-1.5pt E}_p(f)\le \liminf_{n\to\infty} \mathsf{C\kern-1.5pt E}_p(f_n) . By (3.19) we conclude that \frac1p \mathsf{C\kern-1.5pt E}_{p, \kappa}(f)- {\langle} L,f {\rangle} = M so that f is the unique minimizer of (3.14).
(d) (3.16) is an immediate consequence of (3.10) with \alpha = 0 and \beta = \kappa .
In order to prove the continuity of {\mathrm R}_ \kappa , let L_n\in H^{-1,q}( \mathbb{X}) be a sequence strongly converging to L and let f_n = {\mathrm R}_ \kappa(L_n)\in L^q(X, \mathfrak m) . Since \mathsf{C\kern-1.5pt E}_{p, \kappa}^*(L_n) is uniformly bounded, we obviously get a uniform bound for \mathsf{C\kern-1.5pt E}_{p, \kappa}^*(L_n-f_n) and \|f_n\|_{L^q} . Let f\in L^q(X, \mathfrak m) be any weak L^q limit point of f_n , e.g., attained along a subsequence f_{n(j)} . Since \lim_{n\to\infty} \mathsf{C\kern-1.5pt E}_{p, \kappa}^*(L_n) = \mathsf{C\kern-1.5pt E}_{p, \kappa}^*(L) and
\liminf\limits_{j\to\infty} \mathsf{C\kern-1.5pt E}_{p, \kappa}^*(L_{n(j)}-f_{n(j)})\ge \mathsf{C\kern-1.5pt E}_{p, \kappa}^*(L-f),\qquad \liminf\limits_{j\to\infty}\|f_{n(j)}\|_{L^q}^q\ge \|f\|_{L^q}^q |
we deduce that
\mathsf{C\kern-1.5pt E}_{p, \kappa}^*(L)\ge \mathsf{C\kern-1.5pt E}_{p, \kappa}^*(L-f)+\frac1{q \kappa^{q/p}}\|f\|_{L^q}^q |
so that f = {\mathrm R}_ \kappa(L) . Since {\mathrm R}_ \kappa(L) is the unique weak limit point of the sequence f_n in L^q , we conclude that f_n \rightharpoonup {\mathrm R}_ \kappa(L) in L^q(X, \mathfrak m) . The same variational argument also shows that \limsup_{n\to\infty}\|f_n\|_{L^q}\le \|f\|_{L^q} so that the convergence is strong.
Finally, if L\in H^{-1,q}_{\mathrm{pd}}(\mathbb{X}) , f_ \kappa is the (unique) minimizer of (3.10) and u_ \kappa is the (unique) minimizer of (3.14), we get
\begin{align*} \frac1q \mathsf{C\kern-1.5pt E}_{p}( L-f_ \kappa)+\frac 1p \mathsf{C\kern-1.5pt E}_p(u_ \kappa) - {\langle} L-f_ \kappa,u_ \kappa {\rangle}+ \frac1{q \kappa^{q/p}}\|f_ \kappa\|_{L^q}^q+ \frac \kappa p\|u_ \kappa\|_{L^p}^p- {\langle} f_ \kappa,u_ \kappa {\rangle} = 0 \end{align*} |
which yields
{\mathrm A}_p u_ \kappa = L-f_ \kappa,\quad f_ \kappa = \kappa {\mathrm J}_pu_ \kappa. |
(e) Since the map \kappa \mapsto \mathsf{C\kern-1.5pt E}^*_{p, \kappa}( L) is nonincreasing, we have \lim\limits_{ \kappa \downarrow0} \mathsf{C\kern-1.5pt E}^*_{p, \kappa}( L) = \sup\limits_{ \kappa>0} \mathsf{C\kern-1.5pt E}^*_{p, \kappa}( L)\le \mathsf{C\kern-1.5pt E}_p^*( L) . On the other hand, for every f\in H^{1,p}( \mathbb{X}) and \varepsilon>0 , choosing \kappa>0 sufficiently small so that \frac \kappa p\|f\|^p_{L^p}\le \varepsilon we get
\begin{align*} {\langle} L, f {\rangle}-\frac 1p \mathsf{C\kern-1.5pt E}_p(f)- \varepsilon&\le {\langle} L, f {\rangle}-\frac 1p \mathsf{C\kern-1.5pt E}_p(f)-\frac \kappa p\|f\|^p_{L^p} = {\langle} L, f {\rangle}-\frac 1p \mathsf{C\kern-1.5pt E}_{p, \kappa}(f) \\& \le \frac 1q \mathsf{C\kern-1.5pt E}_{p, \kappa}^*( L) \le \frac 1q \sup\limits_{ \kappa \gt 0} \mathsf{C\kern-1.5pt E}_{p, \kappa}^*( L) \end{align*} |
Since the inequality holds for every \varepsilon>0 and every f\in H^{1,p}( \mathbb{X}) , we obtain the converse inequality \mathsf{C\kern-1.5pt E}_p^*( L)\le \sup_{ \kappa>0} \mathsf{C\kern-1.5pt E}_{p, \kappa}^*( L) .
Proposition 3.9 yields the following interesting duality result, which is also related to the theory of derivations discussed in [12].
Corollary 3.10 ( H^{1,p}( \mathbb{X}) is the dual of a Banach space). H^{1,p}( \mathbb{X}) can be isometrically identified with the dual of H^{-1,q}_{\mathrm{pd}}(\mathbb{X}) . In particular, if L^q(X, \mathfrak m) is a separable space, H^{1,p}( \mathbb{X}) is the dual of a separable Banach space.
Proof. Let z be a bounded linear functional on H^{-1,q}_{\mathrm{pd}}(\mathbb{X}) . Since L^q(X, \mathfrak m) is continuously and densely imbedded in H^{-1,q}_{\mathrm{pd}}(\mathbb{X}) , for every f\in L^q(X, \mathfrak m) {\langle} z,f {\rangle}\le \|z\|_{ H^{-1,q}_{\mathrm{pd}}(\mathbb{X})}\|f\|_{L^q} , so that there exists a unique u = \iota(\zeta)\in L^p(X, \mathfrak m) such that
{\langle} z,f {\rangle} = \int_X u f\, {\mathrm d} \mathfrak m\quad\text{for every }f\in L^q(X, \mathfrak m). |
By (3.6) and the strong density of L^q(X, \mathfrak m) in H^{-1,q}_{\mathrm{pd}}(\mathbb{X})
\begin{align*} \frac1p \mathsf{C\kern-1.5pt E}_{p,1}(u) & = \sup\limits_{f\in L^q(X, \mathfrak m)}\int_X uf\, {\mathrm d} \mathfrak m- \frac 1q \mathsf{C\kern-1.5pt E}_{p,1}^*(f) = \sup\limits_{f\in L^q(X, \mathfrak m)} {\langle} z,f {\rangle}- \frac 1q\|f\|_{H^{-1,q}( \mathbb{X})}^q \\& = \sup\limits_{f\in H^{-1,q}_{\mathrm{pd}}(\mathbb{X})} {\langle} z,f {\rangle}- \frac 1q\|f\|_{ H^{-1,q}_{\mathrm{pd}}(\mathbb{X})}^q = \frac 1p\|z\|_{( H^{-1,q}_{\mathrm{pd}}(\mathbb{X}))'}^p. \end{align*} |
It follows that \iota is an isometry from the dual of H^{-1,q}_{\mathrm{pd}}(\mathbb{X}) and H^{1,p}( \mathbb{X}) . Since \iota is clearly surjective, we conclude.
Remark 3.11 ( H^{1,p}( \mathbb{X}) as Gagliardo completion [16]). Recall that if (A,\|\cdot\|_A) is a normed vector space continuously imbedded in a Banach space (V,\|\cdot\|_V) , the Gagliardo completion A^{V,c} is the Banach space defined by
W: = \Big\{v\in V:\exists (a_n)_n\subset A, \ \lim\limits_{n\to\infty}\|a_n-v\|_V = 0,\ \sup\limits_n \|a_n\|_A \lt \infty\Big\} |
with norm
\|v\|_{W}: = \inf\Big\{\liminf\limits_{n\to\infty} \|a_n\|_A:a_n\in A,\ \lim\limits_{n\to\infty}\|a_n-v\|_V = 0\Big\}. |
When \mathop{{\rm{supp}}}\nolimits( \mathfrak m) = X , we can identify \mathscr{A} with a vector space A with the norm induced by \mathsf{p\kern-1pt C\kern-1.5pt E}_{p} imbedded in V: = L^p(X, \mathfrak m) ; it is immediate to check that H^{1,p}( \mathbb{X}) coincides with the Gagliardo completion of A in V . When A (and therefore W ) is strongly dense in V , we can identify the dual V' of V as a subset of the dual W' of W and we can define the set W_{\kern-2pt\mathrm{pd}}' as the closure of V' in W' . If V is uniformly convex, the same statements and characterizations given in Propositions 3.7 and 3.9 hold in this more abstract setting. In particular, W can be isometrically identified with the dual of W_{\kern-2pt\mathrm{pd}}' .
The following result provides a useful criterion to check if a linear functional on \mathscr{A} belongs to \mathscr{A}_{q}' . Let us first recall that a subset F\subset L^1(X, \mathfrak m) is weakly relatively compact in L^1(X, \mathfrak m) if and only if it satisfies one of the following equivalent properties [13, Chap. Ⅷ, Theorem 1.3]:
a) for all \varepsilon>0 there exists m\ge0 such that
\int_{|f(x)|\ge m}|f(x)|\, {\mathrm d} \mathfrak m(x)\le \varepsilon\quad\text{for every }f\in F; |
b) (Equiintegrability) For all \varepsilon>0 there exists \delta>0 such that for every Borel set B\subset X
\int_{B}|f(x)|\, {\mathrm d} \mathfrak m(x)\le \varepsilon\quad\text{whenever $f\in F$ and $ \mathfrak m(B)\le \delta$}. |
c) (Uniform superlinear estimate) There exists a positive, increasing, l.s.c. and convex function \Phi:[0,\infty)\to [0,\infty) such that \lim_{r\to\infty}\Phi(r)/r = +\infty and
\sup\limits_{f\in F} \int_X \Phi(|f(x)|)\, {\mathrm d} \mathfrak m(x) \lt \infty. |
Proposition 3.12. Let L be a linear functional on \mathscr{A} satisfying (3.8a). If for every sequence f_n\in \mathscr{A} satisfying
\begin{equation} -1\le f_n\le 1,\quad \lim\limits_{n\to\infty}\|f_n\|_{L^p(X, \mathfrak m)} = 0,\quad \{( \mathop{{\rm{lip}}}\nolimits f_n)^p:n\in \mathbb{N}\}\quad\mathit{\text{is equiintegrable}} \end{equation} | (3.22) |
one has \liminf_{n\to\infty}| {\langle} L,f_n {\rangle}| = 0 , then L\in \mathscr{A}_{q}' .
Proof. We split the proof in two steps.
Claim 1: if L is a linear functional on \mathscr{A} satisfying (3.8a) and for every sequence ( f_n)_{n\in \mathbb{N}}\subset \mathscr{A}
\begin{equation} \lim\limits_{n\to\infty}\|f_n\|_{L^p} = 0\ {and}\ \{( \mathop{{\rm{lip}}}\nolimits f_n)^p:n\in \mathbb{N}\}\;{ equiintegrable} \quad\Rightarrow\quad \liminf\limits_{n\to\infty}| {\langle} L,f_n {\rangle}| = 0 \end{equation} | (3.23) |
then L\in \mathscr{A}_{q}' .
We argue by contradiction and we assume that there exists a sequence f_n\in \mathscr{A} such that
\mathsf{p\kern-1pt C\kern-1.5pt E}_{p,1}(f_n)\le C,\quad \lim\limits_{n\to\infty}\|f_n\|_{L^p} = 0\quad\text{and}\quad \liminf\limits_{n\to\infty}| {\langle} L,f_n {\rangle}| \gt 0. |
By possibly changing the sign of f_n it is not restrictive to assume that {\langle} L,f_n {\rangle}\ge c>0 for every n\in \mathbb{N} . Applying Mazur Lemma we find coefficients \alpha_{n,m}\ge0 , n\in \mathbb{N} , 0\le m\le M(n) such that g_n: = \sum_{m = 0}^{M(n)} \alpha_{n,m} \mathop{{\rm{lip}}}\nolimits f_{n+m} is strongly converging in L^p(X, \mathfrak m) . Thus n\mapsto g_n^p is strongly converging in L^1(X, \mathfrak m) and it is therefore equi-integrable.
We now consider \tilde f_n: = \sum_{m = 0}^{M(n)} \alpha_{n,m} f_{n+m} . By construction
\begin{equation} {\langle} L,\tilde f_n {\rangle} = \sum\limits_{m = 0}^{M(n)} \alpha_{n,m} {\langle} L,f_{n+m} {\rangle}\ge c \gt 0,\quad \lim\limits_{n\to\infty}\|\tilde f_n\|_{L^p} = 0 \end{equation} | (3.24) |
and
\mathop{\rm{ lip }}\nolimits\tilde f_n\le \sum\limits_{m = 0}^{M(n)} \alpha_{n,m} \mathop{\rm{ lip }}\nolimits f_{n+m} = g_n |
so that (3.23) yields \liminf_{n\to\infty} {\langle} L,\tilde f_n {\rangle} = 0 , which contradicts the first inequality of (3.24).
Claim 2: it is sufficient to prove the implication (3.23) for sequences taking values in [-1,1] . Let f_n\in \mathscr{A} as in (3.23), m_n: = \sup |f_n| , E^p: = \sup\limits_n \mathsf{p\kern-1pt C\kern-1.5pt E}_p(f_n)<\infty , and let \phi\in {\mathrm C}^1( \mathbb{R}) be an odd function such that
\begin{equation} \phi(r) = r\quad \text{if }|r|\le 1/2,\quad -1\le \phi(r)\le 1,\quad 0\le \phi'(r)\le 1\quad\text{for every }r\in \mathbb{R}. \end{equation} | (3.25) |
Let us fix \varepsilon>0 and \delta: = \varepsilon/3E so that E^p\delta^p\le \varepsilon^p/3 . For every choice of n\in \mathbb{N} we can find an odd polynomial P_n such that (see e.g., [21, Lemma 2.23])
-1\le P_n(r)\le 1,\quad 0\le P_n'(r)\le 1,\quad |P_n'(r)-\phi'(r)|\le \delta\quad \text{for every }r\in [-m_n,m_n], |
We set
g_n: = P_n\circ f_n,\quad h_n: = f_n-g_n = Q_n(f_n)\quad\text{where}\quad Q_n(r) = r-P_n(r); |
notice that g_n and h_n belong to \mathscr{A} as well. Since \mathop{{\rm{lip}}}\nolimits g_n\le \mathop{{\rm{lip}}}\nolimits f_n and g_n takes values in [-1,1] , by assumption we have \liminf_{n\to\infty}| {\langle} L,g_{n} {\rangle}| = 0 . On the other hand, \|h_{n}\|_{L^p}\le \|f_n\|_{L^p} , \mathop{{\rm{lip}}}\nolimits h_{n}\le \mathop{{\rm{lip}}}\nolimits f_n , and \mathop{{\rm{lip}}}\nolimits h_{n}(x)\le Q_n'(f_n(x)) \mathop{{\rm{lip}}}\nolimits f_n(x)\le \delta \mathop{{\rm{lip}}}\nolimits f_n(x) whenever |f_n(x)|<1/2 . Since Chebichev inequality yields
\mathfrak m\big\{|f_n|\ge 1/2\big\}\le 2^p\|f_n\|^p_{L^p(X, \mathfrak m)},\quad \lim\limits_{n\to\infty} \mathfrak m\big\{f_n\ge 1/2\big\} = 0, |
we can choose n_0 sufficiently big so that for every n\ge n_0
\int_{\{|f_n|\ge 1/2\}} ( \mathop{\rm{ lip }}\nolimits f_n)^p\, {\mathrm d} \mathfrak m\le \varepsilon^p/3,\quad \int_X |h_n|^p\, {\mathrm d} \mathfrak m\le \varepsilon^p/3, |
and
\begin{align*} \int_X ( \mathop{{\rm{lip}}}\nolimits h_n)^p\, {\mathrm d} \mathfrak m &\le \delta^p\int_{\{|f_n| \lt 1/2\}} ( \mathop{{\rm{lip}}}\nolimits f_n)^p\, {\mathrm d} \mathfrak m +\int_{\{|f_n|\ge 1/2\}} ( \mathop{{\rm{lip}}}\nolimits f_n)^p\, {\mathrm d} \mathfrak m \le \delta^p E^p+\frac 13 \varepsilon^p\le \frac 23 \varepsilon^p. \end{align*} |
(3.8a) then yields | {\langle} L,h_n {\rangle}|\le D \varepsilon for n\ge n_0 and therefore \liminf_{n\to\infty}| {\langle} L,f_n {\rangle}|\le D \varepsilon . Since \varepsilon>0 is arbitrary, we conclude.
Our last criterium concerns positive functionals, i.e., satisfying
\begin{equation} {\langle} L,f {\rangle}\ge0\quad\text{whenever }f\in \mathscr{A},\ f\ge0. \end{equation} | (3.26) |
We will see in Theorem 4.7 that they are always induced by a Radon measure.
Theorem 3.13. If L is a linear functional on \mathscr{A} satisfying (3.8a) and (3.26), then L\in \mathscr{A}_{q}' .
Proof. We apply Proposition 3.12 and refine the last argument of its proof. Let f_n\in \mathscr{A} as in (3.22) with E^p: = \sup\limits_n \mathsf{C\kern-1.5pt E}_p(f_n) We select strictly positive parameters \varepsilon,\kappa>0 , \delta: = \varepsilon/3E , the odd function \phi as in (3.25) with \phi_\kappa(r): = \kappa \phi(r/\kappa) , and odd polynomials \tilde P_{\kappa, \varepsilon} satisfying
-1\le \tilde P_{\kappa, \varepsilon}(r)\le 1,\quad 0\le \tilde P_{\kappa, \varepsilon}'(r)\le 1,\quad |\tilde P_{\kappa, \varepsilon}'(r)-\phi'(r)|\le \delta\quad \text{if }|r|\le \frac 1\kappa. |
We also set P_{\kappa, \varepsilon}(r): = \kappa \tilde P_{ \kappa, \varepsilon}(r/\kappa) , Q_{\kappa, \varepsilon}(r) = r-P_{\kappa, \varepsilon}(r) , g_{n, \kappa}: = P_{\kappa, \varepsilon}(f_n) , h_{n, \kappa}: = f_n-g_{n, \kappa} = Q_{\kappa, \varepsilon}(f_n) . By (3.26) and observing that -\kappa\le g_{n,\kappa}\le \kappa and the constant function 1\in \mathscr{A} has finite Cheeger energy, we have
-\kappa {\langle} L,1 {\rangle} = - {\langle} L,\kappa {\rangle}\le {\langle} L,g_{n, \kappa} {\rangle}\le {\langle} L, \kappa {\rangle} = \kappa {\langle} L,1 {\rangle}. |
On the other hand, since 0\le Q'_{n,\kappa}\le 1 if |r|\le 1 and |Q_{n,\kappa}'(r)|\le \delta if |r|\le \kappa/2 , we have
|h_{n,\kappa}|\le |f_n|,\quad \mathop{\rm{ lip }}\nolimits h_{n,\kappa}\le \mathop{\rm{ lip }}\nolimits f_n,\quad \mathop{\rm{ lip }}\nolimits h_{n,\kappa}\le \delta \mathop{\rm{ lip }}\nolimits f_n\quad\text{if }|f_n| \lt \kappa/2. |
Applying Chebychev inequality
\mathfrak m\big\{|f_n|\ge \kappa/2\big\}\le \frac {2^p\|f_n\|^p_{L^p(X, \mathfrak m)}}{\kappa^p}, |
we can find n_0 (depending on \varepsilon,\kappa ) sufficiently big such that
\int_X |h_{n, \kappa}|^p\, {\mathrm d} \mathfrak m\le \varepsilon^p/3,\quad \int_{\{f_n\ge \kappa/2\}} ( \mathop{\rm{ lip }}\nolimits f_n)^p\, {\mathrm d} \mathfrak m\le \varepsilon^p/3\quad\text{for every }n\ge n_0, |
so that
\int_X ( \mathop{\rm{ lip }}\nolimits h_{n,\kappa})^p\, {\mathrm d} \mathfrak m \le \delta^p\int_{\{|f_n| \lt \kappa/2\}} ( \mathop{\rm{ lip }}\nolimits f_n)^p\, {\mathrm d} \mathfrak m +\int_{\{|f_n|\ge 1/2\}} ( \mathop{\rm{ lip }}\nolimits f_n)^p\, {\mathrm d} \mathfrak m \le \delta^p E^p+\frac 13 \varepsilon^p\le \frac 23 \varepsilon^p. |
By (3.8a) it follows that
\liminf\limits_{n\to\infty}| {\langle} L,f_n {\rangle}|\le \liminf\limits_{n\to\infty}\big(| {\langle} L,g_{n, \kappa} {\rangle}|+| {\langle} L,h_{n, \kappa} {\rangle}|\big)\le \kappa {\langle} L,1 {\rangle}+D \varepsilon. |
Since \varepsilon,\kappa are arbitrary, we get \liminf_{n\to\infty}| {\langle} L,f_n {\rangle}| = 0 .
Definition 3.14 (Measure with finite dual energy). A Radon measure \mu\in {{\mathcal M}}_+(X) has finite energy if there exists a constant D>0 such that
\begin{equation} \int_X f\, {\mathrm d}\mu\le D\Big( \mathsf{p\kern-1pt C\kern-1.5pt E}_{p,1}(f)\Big)^{1/p} \quad{for\; every\; nonnegative }\;f\in \mathop{{\rm{Lip}}}\nolimits_{b}(X,\tau, {\mathsf d}). \end{equation} | (3.27) |
Corollary 3.15 (Measures with finite dual energy belong to \mathscr{A}_q'( \mathbb{X}) ). If \mu\in {{\mathcal M}}_+(X) has finite energy then the linear functional f\mapsto \int_X f\, {\mathrm d}\mu on \mathscr{A}( \mathbb{X}) = \mathop{{\rm{Lip}}}\nolimits_b(X,\tau, {\mathsf d}) belongs to \mathscr{A}_q'( \mathbb{X}) and can be uniquely extended to a functional L_\mu\in H^{-1,q}_{\mathrm{pd}}(\mathbb{X}) satisfying
\begin{equation} \mathsf{C\kern-1.5pt E}_{p, \kappa}^*(L_\mu) = \mathsf{p\kern-1pt C\kern-1.5pt E}_{p, \kappa}^*(\mu)\quad\mathit{\text{for every}}\; \kappa \gt 0. \end{equation} | (3.28) |
In particular there exists a unique element u_ \kappa = {\mathrm Q}_{p, \kappa}^{-1}(L_\mu)\in H^{1,p}( \mathbb{X}) minimizing (3.14) with L = L_\mu . u_ \kappa satisfies the variational inequality
\begin{equation} \int_X f\, {\mathrm d}\mu- \kappa\int_X {\mathrm J}_p(u_ \kappa)f\, {\mathrm d} \mathfrak m\le \frac1p \mathsf{C\kern-1.5pt E}_p(u_ \kappa+f)-\frac 1p \mathsf{C\kern-1.5pt E}(u_ \kappa)\quad \mathit{\text{for every}}\;f\in \mathop{{\rm{Lip}}}\nolimits_b(X,\tau, {\mathsf d}). \end{equation} | (3.29) |
Proof. We can apply Theorem 3.13 with \mathscr{A} = \mathop{{\rm{Lip}}}\nolimits_b(X,\tau, {\mathsf d}) . Clearly (3.26) holds; by decomposing f\in \mathop{{\rm{Lip}}}\nolimits_b(X,\tau, {\mathsf d}) as the difference f = f_+-f_- of its positive and negative part and recalling that \mathsf{p\kern-1pt C\kern-1.5pt E}_p(f_\pm)\le \mathsf{p\kern-1pt C\kern-1.5pt E}_p(f) , (3.27) yields (3.8a) with constant 2D . (3.29) follows by (3.15) and the definition of {\mathrm A}_p given in (3.11).
Definition 4.1 ((Nonparametric) dynamic plans). A (nonparametric) dynamic plan is a Radon measure {\boldsymbol \pi}\in \mathcal M_+( \mathrm{RA}(X)) on \mathrm{RA}(X) such that
\begin{equation} {\boldsymbol \pi}(\ell): = \int_{ \mathrm{RA}(X)}\ell(\gamma)\, {\mathrm d} {\boldsymbol \pi}(\gamma) \lt \infty. \end{equation} | (4.1) |
Using the universally Lusin-measurable map R: \mathrm{RA}(X)\to \mathrm{BVC}_c([0,1];X) in (2.7) we can also lift {\boldsymbol \pi} to a Radon measure \tilde {\boldsymbol \pi} = R_\sharp {\boldsymbol \pi} on the subset \mathrm{BVC}_c([0,1];X) of \mathrm{BVC}([0,1];X) defined in (2.6). Conversely, any Radon measure \tilde {\boldsymbol \pi} on {\mathrm C}([0,1];(X,\tau)) concentrated on \mathrm{BVC}([0,1];X) yields the Radon measure {\boldsymbol \pi}: = \mathfrak{q}_\sharp\tilde {\boldsymbol \pi} on \mathrm{RA}(X) . Notice that \mathfrak{q}_\sharp( R_\sharp {\boldsymbol \pi}) = {\boldsymbol \pi} .
If {\boldsymbol \pi} is a dynamic plan in \mathcal M_+( \mathrm{RA}(X)) , thanks to Theorem 2.6(e) and Fubini's Theorem [11, Chap. Ⅱ-14], we can define the Borel measure \mu_ {\boldsymbol \pi}: = {\rm{Proj}} ({\boldsymbol \pi})\in \mathcal M_+(X) by the formula
\begin{equation*} \int f\, {\mathrm d}\mu_ {\boldsymbol \pi}: = \iint_\gamma f\, {\mathrm d} {\boldsymbol \pi}(\gamma) \quad\text{for every bounded Borel function }f:X\to \mathbb{R}. \end{equation*} |
\mu_ {\boldsymbol \pi} is a Radon measure with total mass {\boldsymbol \pi}(\ell) given by (4.1) [21, §8] and it can also be considered as the integral w.r.t. {\boldsymbol \pi} of the Borel family of measures \nu_\gamma , \gamma\in \mathrm{RA}(X) [11, Chap. Ⅱ-13], in the sense that
\int_X f\, {\mathrm d}\mu_ {\boldsymbol \pi}(x) = \int_{ \mathrm{RA}(X)}\Big(\int_X f\, {\mathrm d}\nu_\gamma\Big)\, {\mathrm d} {\boldsymbol \pi}(\gamma). |
Recall that p,q\in (1,\infty) is a fixed pair of conjugate exponents.
Definition 4.2. We say that {\boldsymbol \pi}\in \mathcal M_+( \mathrm{RA}(X)) has barycenter in L^q(X, \mathfrak m) if there exists h\in L^q(X, \mathfrak m) such that \mu_ {\boldsymbol \pi} = h \mathfrak m , or, equivalently, if
\int \int_\gamma f\, {\mathrm d} {\boldsymbol \pi}(\gamma) = \int fh\, {\mathrm d} \mathfrak m\quad {for \; every \;bounded\; Borel\;function }\;f:X\to \mathbb{R}, |
and we call \operatorname{Bar}_{q} ( {\boldsymbol \pi}): = \|h\|_{L^q(X, \mathfrak m)} the barycentric q -entropy of {\boldsymbol \pi} . We will denote by {{\mathcal{B}}_q}\left( {{\rm{RA}}\left( X \right)} \right) the set of all plans with barycenter in L^q(X, \mathfrak m) and we will set \operatorname{Bar}_{q}( {\boldsymbol \pi}): = +\infty if {\boldsymbol \pi}\not\in {{\mathcal{B}}_q}\left( {{\rm{RA}}\left( X \right)} \right) .
\operatorname{Bar}_{q}: \mathcal M_+( \mathrm{RA}(X))\to[0,+\infty] is a convex and positively 1 -homogeneous functional, which is lower semicontinuous w.r.t. the weak topology of \mathcal M_+( \mathrm{RA}(X)) , since it can also be characterized as the L^q entropy of the projected measure \mu_ {\boldsymbol \pi} = {\rm{Proj}} ({\boldsymbol \pi}) with respect to \mathfrak m :
\operatorname{Bar}_{q}^q( {\boldsymbol \pi}) = \mathscr{L}_q(\mu_ {\boldsymbol \pi}| \mathfrak m), |
where for an arbitrary \sigma\in \mathcal M_+(X)
\begin{equation} \mathscr L_q(\sigma| \mathfrak m): = \begin{cases} \int_X \Big(\frac{ {\mathrm d}\sigma}{ {\mathrm d} \mathfrak m}\Big)^q\, {\mathrm d} \mathfrak m&\text{if $\sigma\ll \mathfrak m$,}\\ +\infty&\text{otherwise.} \end{cases} \end{equation} | (4.2) |
Notice that \operatorname{Bar}_{q}( {\boldsymbol \pi}) = 0 iff {\boldsymbol \pi} is concentrated on the set of constant arcs in \mathrm{RA}(X) .
For every \mu_0,\mu_1\in \mathcal M_+(X) we will consider the (possibly empty) set
\Pi(\mu_0,\mu_1): = \Big\{ {\boldsymbol \pi}\in \mathcal M_+( \mathrm{RA}(X)):\ ( {\mathsf e}_i)_\sharp {\boldsymbol \pi} = \mu_i\Big\} |
and we define
\begin{equation} \mathscr{D}_q(\mu_0,\mu_1): = \inf\Big\{\operatorname{Bar}_{q}^q( {\boldsymbol \pi}): {\boldsymbol \pi}\in \Pi(\mu_0,\mu_1)\Big\}, \end{equation} | (4.3) |
with the usual convention \mathscr{D}_q(\mu_0,\mu_1): = +\infty if \Pi(\mu_0,\mu_1) = \emptyset . Clearly \mathscr{D}_q(\mu_0,\mu_1) = +\infty if \mu_0(X)\neq \mu_1(X) .
Assuming that (X, {\mathsf d}) is complete, it is possible to show ([21, §11]) that whenever \mathscr{D}_q(\mu_0,\mu_1)<+\infty the infimum in (4.3) is attained and the set of optimal plans \Pi_o(\mu_0,\mu_1) is a compact convex subset of \mathcal M_+( \mathrm{RA}(X)) . Morever, for every optimal {\boldsymbol \pi}\in \Pi_o(\mu_0,\mu_1) the measure \mu_o = {\rm{Proj}} ({\boldsymbol \pi}) is independent of {\boldsymbol \pi} .
\mathscr{D}_q provides an important representation for the dual of the p -Cheeger energy.
Theorem 4.3. ([21, Thm. 11.8]). For every \mu_0,\mu_1\in \mathcal M_+(X) we have
\begin{equation} \mathscr{D}_q(\mu_0,\mu_1) = \mathsf{p\kern-1pt C\kern-1.5pt E}_p^*(\mu_0-\mu_1). \end{equation} | (4.4) |
Remark 4.4. Let \mu = \mu_0-\mu_1 with \mu_i\in \mathcal M_+(X) and let \mu_+,\mu_-\in \mathcal M_+(X) be mutually singular Radon measures providing the Jordan decomposition of \mu as \mu_+-\mu_- with \mu' = \mu_0-\mu_+ = \mu_1-\mu_-\in \mathcal M_+(X) . (4.4) shows that
\mathscr{D}_q(\mu_0,\mu_1) = \mathsf{p\kern-1pt C\kern-1.5pt E}_p^*(\mu_0-\mu_1) = \mathsf{p\kern-1pt C\kern-1.5pt E}_p^*(\mu_+-\mu_-) = \mathscr{D}_q(\mu_+,\mu_-). |
Denoting by \mathfrak c:X\to \mathrm{RA}(X) the map that at every point x associates the constant curve taking values x , if {\boldsymbol \pi}_o\in \Pi_o(\mu_+,\mu_-) and {\boldsymbol \pi}': = \mathfrak c_\sharp \mu' , it is easy to check that {\boldsymbol \pi}: = {\boldsymbol \pi}_o+ {\boldsymbol \pi}'\in \Pi_o(\mu_0,\mu_1) .
Definition 4.5. For every nonparametric dynamic plan {\boldsymbol \pi}\in \mathcal M_+( \mathrm{RA}(X)) and \kappa>0 we define
\mathscr{E}_{q, \kappa}( {\boldsymbol \pi}): = \operatorname{Bar}_{q}^q( {\boldsymbol \pi})+ \kappa^{-q/p} \mathscr{L}_q(( {\mathsf e}_1)_\sharp {\boldsymbol \pi}| \mathfrak m). |
For every \mu\in \mathcal M_+(X) we set
\begin{equation} \mathscr{D}_{q, \kappa}(\mu): = \inf\Big\{ \mathscr{E}_{q, \kappa}( {\boldsymbol \pi}): {\boldsymbol \pi}\in \mathcal M_+( \mathrm{RA}(X)),\quad ( {\mathsf e}_{0})_\sharp {\boldsymbol \pi} = \mu\Big\}. \end{equation} | (4.5) |
Theorem 4.6. For every \mu\in \mathcal M_+(X) we have
\mathscr{D}_{q, \kappa}(\mu) = \mathsf{p\kern-1pt C\kern-1.5pt E}_{p, \kappa}^*(\mu). |
Moreover, if one of the above quantities is finite
(a) The infimum in (4.5) is attained and there exists a unique pair of functions f_{ \kappa},g_{ \kappa}\in L^q(X, \mathfrak m) such that for every optimal plan {\boldsymbol \pi}
g_{ \kappa} \mathfrak m = {\rm{Proj}} ({\boldsymbol \pi}),\quad \mu_ \kappa = f_{ \kappa} \mathfrak m = ( {\mathsf e}_1)_\sharp {\boldsymbol \pi},\quad {\boldsymbol \pi}\in \Pi_o(\mu,\mu_ \kappa). |
(b) There exists a unique solution u_ \kappa = {\mathrm Q}_{p, \kappa}^{-1}(L_\mu) of
\mathrm {A}_p u+ \kappa {\mathrm J}_p u\ni L_\mu |
and it satisfies
\begin{equation} {\mathrm J}_p\big(| {\mathrm D} u|_\star\big) = g_{ \kappa},\quad \kappa {\mathrm J}_p u = f_{ \kappa} \end{equation} | (4.6) |
\left\langle L_\mu,u \right\rangle = \mathsf{C\kern-1.5pt E}_{p, \kappa}(u) = \mathsf{C\kern-1.5pt E}_{p, \kappa}^*(L_\mu) = \mathsf{p\kern-1pt C\kern-1.5pt E}_{p, \kappa}^*(\mu). |
Moreover, setting \mu_\pm: = (\mu-\mu_ \kappa)_\pm and \mu': = \mu-\mu_+ = \mu_ \kappa-\mu_- , we can always choose {\boldsymbol \pi} = {\boldsymbol \pi}_o+ {\boldsymbol \pi}' where {\boldsymbol \pi}_o\in \Pi_o(\mu_+,\mu_-) , {\boldsymbol \pi}' = \mathfrak c_\sharp \mu' , \mathsf{p\kern-1pt C\kern-1.5pt E}_{p}^*(\mu-\mu_ \kappa) = \operatorname{Bar}_{q}^q(\pi) = \operatorname{Bar}_{q}^q(\pi_o) .
Proof. By Corollary 3.15 we can extend \mu to a functional L_\mu\in H^{-1,q}_{\mathrm{pd}}(\mathbb{X}) satisfying (3.28). We can then apply Proposition 3.9 and find nonnegative f_ \kappa\in L^q(X, \mathfrak m) and u_ \kappa\in L^p(X, \mathfrak m) such that
\mathsf{C\kern-1.5pt E}_{p, \kappa}^*(L_\mu) = \mathsf{C\kern-1.5pt E}_{p}^*(L_\mu-f_ \kappa)+ \kappa^{-q/p}\|f_ \kappa\|_{L^q(X, \mathfrak m)}^q = \mathsf{p\kern-1pt C\kern-1.5pt E}_{p}^*(\mu-f_ \kappa \mathfrak m)+ \kappa^{-q/p} \mathscr{L}_q(f_ \kappa \mathfrak m| \mathfrak m) |
and u_\kappa satisfies f_ \kappa = \kappa {\mathrm J}_p u_ \kappa and (3.15) with L = L_\mu . Setting \mu_ \kappa: = f_ \kappa \mathfrak m and selecting {\boldsymbol \pi}\in \Pi_o(\mu,\mu_ \kappa) according to Remark 4.4, (4.4) yields
\mathsf{C\kern-1.5pt E}_{p, \kappa}^*(L_\mu) = \operatorname{Bar}_{q}^q( {\boldsymbol \pi})+ \kappa^{-q/p} \mathscr{L}_q(( {\mathsf e}_1)_\sharp {\boldsymbol \pi}| \mathfrak m) = \mathscr{E}_{q, \kappa}( {\boldsymbol \pi})\ge \mathscr{D}_{q, \kappa}(\mu). |
On the other hand, it is easy to check that \mathscr{D}_{q, \kappa}(\mu)\ge \mathsf{p\kern-1pt C\kern-1.5pt E}_{p, \kappa}^*(\mu) , since for every plan {\boldsymbol \pi}\in {{\mathcal M}}_+( \mathrm{RA}(X)) as in (4.5) and every f\in \mathop{{\rm{Lip}}}\nolimits_b(X,\tau, {\mathsf d})
\begin{align*} \int_X f\, {\mathrm d} \mu &\le \int_X f\, {\mathrm d}( {\mathsf e}_0)_\sharp {\boldsymbol \pi} - \int_X f\, {\mathrm d}( {\mathsf e}_1)_\sharp {\boldsymbol \pi}+ \int_X f\, {\mathrm d}( {\mathsf e}_1)_\sharp {\boldsymbol \pi} \\&\le \int_{ \mathrm{RA}(X)}\int_\gamma \mathop{{\rm{lip}}}\nolimits f\, {\mathrm d} {\boldsymbol \pi}(\gamma)+ \|f\|_{L^p}\Big( \mathscr{L}_q(( {\mathsf e}_1)_\sharp {\boldsymbol \pi}| \mathfrak m)\Big)^{1/q} \\&\le \big\| \mathop{{\rm{lip}}}\nolimits f\big\|_{L^p}\operatorname{Bar}_{q}( {\boldsymbol \pi})+ \|f\|_{L^p}\Big( \mathscr{L}_q(( {\mathsf e}_1)_\sharp {\boldsymbol \pi}| \mathfrak m)\Big)^{1/q} \le \Big( \mathsf{p\kern-1pt C\kern-1.5pt E}_{p, \kappa}(f)\Big)^{1/p}\,\Big( \mathscr{E}_{q, \kappa}( {\boldsymbol \pi})\Big)^{1/q}. \end{align*} |
Using now the fact that {\langle} L_\mu-f_ \kappa,u {\rangle} = \operatorname{Bar}_{q}^q( {\boldsymbol \pi}) = \mathsf{C\kern-1.5pt E}_p(u_ \kappa) we get
\begin{equation} \int_X g_ \kappa^q\, {\mathrm d} \mathfrak m = \int_X | {\mathrm D} u_ \kappa|_\star^p\, {\mathrm d} \mathfrak m = {\langle} L_\mu-f_ \kappa,u_ \kappa {\rangle}. \end{equation} | (4.7) |
We can also select a sequence w_n\in \mathop{{\rm{Lip}}}\nolimits_b(X,\tau, {\mathsf d}) such that w_n\to u_ \kappa , \mathop{{\rm{lip}}}\nolimits w_n \to | {\mathrm D} u_ \kappa|_\star strongly in L^p(X, \mathfrak m) , so that
\begin{align*} {\langle} L_\mu-f_ \kappa,u_ \kappa {\rangle} & = \lim\limits_{n\to\infty} {\langle} L_\mu-f_ \kappa,w_n {\rangle} = \lim\limits_{n\to\infty} \Big(\int_X w_n\, {\mathrm d}\mu-\int_X w_n\, {\mathrm d}\mu_1\Big) \\& = \lim\limits_{n\to\infty} \Big(\int_{ \mathrm{RA}(X)} w_n( {\mathsf e}_0(\gamma))-w_n( {\mathsf e}_1(\gamma))\Big)\, {\mathrm d} {\boldsymbol \pi}(\gamma) \\&\le \limsup\limits_{n\to\infty} \int_{ \mathrm{RA}(X)} \int_\gamma \mathop{{\rm{lip}}}\nolimits w_n \, {\mathrm d} {\boldsymbol \pi}(\gamma) = \limsup\limits_{n\to\infty} \int_X g_ \kappa \mathop{{\rm{lip}}}\nolimits w_n\, {\mathrm d} \mathfrak m = \int_X g_ \kappa| {\mathrm D} u|_\star\, {\mathrm d} \mathfrak m. \end{align*} |
Inserting this inquality in (4.7) we obtain the first identity of (4.6).
Let us give a first important application of the above result to the representation of positive functionals.
Theorem 4.7. Let \mathscr{A} be a compatible subalgebra of \mathop{{\rm{Lip}}}\nolimits_b(X,\tau, {\mathsf d}) and let L be functional on \mathscr{A} satisfying
| {\langle} L,f {\rangle}|\le D\Big( \mathsf{p\kern-1pt C\kern-1.5pt E}_{1,p}(f)\Big)^{1/p}\quad{for \; every\; f \in \mathscr{A} ,}\quad {\langle} L, f {\rangle}\ge0\quad {for\; every\; positive }\;f\in \mathscr{A}. |
Then L admits a unique extension \tilde L\in H^{-1,q}_{\mathrm{pd}}(\mathbb{X}) and there exists a unique \mu\in \mathcal M_+(X) representing L as
\begin{equation} {\langle} L,f {\rangle} = \int_X f\, {\mathrm d}\mu\quad\mathit{\text{for every}}\;f\in \mathscr{A}. \end{equation} | (4.8) |
Proof. By Theorem 3.13 and Proposition 3.7(c) we know that L is the restriction to \mathscr{A} of a unique functional \tilde L\in H^{-1,q}_{\mathrm{pd}}(\mathbb{X}) . It is easy to check that \tilde L is also positive on H^{1,p}( \mathbb{X}) and applying Proposition 3.9 we can find a sequence w_n\in L^q(X, \mathfrak m) , w_n\ge0 , strongly converging to \tilde L in H^{-1,q}( \mathbb{X}) . Let \mu_n: = w_n \mathfrak m and \nu_n: = {\mathrm R}_1(w_n) \mathfrak m ; applying Theorem 4.6 we can find optimal dynamic plans {\boldsymbol \pi}_n\in \Pi_o(\mu_n,\nu_n) such that \operatorname{Bar}_{q}( {\boldsymbol \pi}_n) = \mathsf{p\kern-1pt C\kern-1.5pt E}_{1,p}^*(\mu_n-\nu_n)\le C . Since \mathscr{L}_q(\nu_n| \mathfrak m) is also uniformly bounded, the sequence {\boldsymbol \pi}_n satisfies the tightness criterium of [21, Lemma 8.5], so that it admits a subsequence (still denoted by {\boldsymbol \pi}_n ) weakly converging to {\boldsymbol \pi}\in {{\cal B}_q}\left( {{\rm{RA}}\left( X \right)} \right) \subset \mathcal M_+( \mathrm{RA}(X)) .
The Radon measure \mu: = ( {\mathsf e}_0)_\sharp {\boldsymbol \pi} is the weak limit of \mu_n : in particular, for every f\in \mathscr{A}
\left\langle L,f \right\rangle = \left\langle \tilde L,f \right\rangle = \lim\limits_{n\to\infty}\int_X f\, {\mathrm d}\mu_n = \int_X f\, {\mathrm d}\mu. |
In this last section we apply the previous result to prove new properties of the Newtonian capacity. We first recall the basic facts about the Newtonian approach [18,23], based on the notion of p -Modulus which has been introduced by Fuglede [15] in the natural framework of collection of positive measures, as in [1]. We refer to [8,17] for a comprehensive presentation of this topic. As usual, p,q\in (1,\infty) denote a pair of conjugate exponents.
Definition 5.1 ( p -Modulus of a family of rectifiable arcs). The p -Modulus of a collection \Gamma\subset \mathrm{RA}(X) is defined by
\operatorname{Mod}_{p}(\Gamma) : = \inf \left\{ \int_X f^p \, {\mathrm d} \mathfrak m: \, f:X\to[0,\infty] \ \text{is Borel},\ \int_\gamma f \geq 1 \; \; { for \;all }\; \gamma \in \Gamma \right\}. |
\Gamma is said to be \operatorname{Mod}_{p} -negligible if \operatorname{Mod}_{p}(\Gamma) = 0 . We say that a property P on \mathrm{RA}(X) holds \operatorname{Mod}_{p} -a.e. if the set of arcs where P fails is \operatorname{Mod}_{p} -negligible. We say that a property P on X holds quasi everywhere (q.e.) if the set of points E where P fails is \mathfrak m -negligible and satisfies
\operatorname{Mod}_{p}(\Gamma_E) = 0\quad {where} \quad \Gamma_E: = \Big\{\gamma\in \mathrm{RA}(X): \ell(\gamma) \gt 0,\ {\mathsf e}[\gamma]\cap E \ne \emptyset\Big\}. |
Notice that if E is \mathfrak m -negligible then for \operatorname{Mod}_{p} -a.e. arc \gamma the set \{t\in [0,1]: R_\gamma(t)\in E\} is {\mathscr L}^{1} -negligible. It can be shown (see e.g., [8]) that \operatorname{Mod}_{p} is an increasing and subadditive functional which is continuous along increasing sequences. In fact, by [1] and [21, §7.2], \operatorname{Mod}_{p} is also continuous along decreasing sequence of compact sets, so that it is a Choquet capacity for the compact paving in \mathrm{RA}(X) [11, Chap. Ⅲ, 28].
It is not difficult to check that for every dynamic plan {\boldsymbol \pi}\in {{\cal B}_q}\left( {{\rm{RA}}\left( X \right)} \right) and every {\boldsymbol \pi} -measurable set \Gamma\subset \mathrm{RA}(X)
{\boldsymbol \pi}(\Gamma)\le \operatorname{Bar}_{q}( {\boldsymbol \pi}) \operatorname{Mod}_{p}^{1/p}(\Gamma), |
which in particular shows that Borel \operatorname{Mod}_{p} -negligible sets are also {\boldsymbol \pi} -negligible for every {\boldsymbol \pi}\in {{\cal B}_q}\left( {{\rm{RA}}\left( X \right)} \right) . In fact, a much more refined result holds [1,21]:
Theorem 5.2. If \mathbb{X} is a complete e.m.t.m. space and \tau is a Souslin topology for X , then every Borel or Souslin set \Gamma in \mathrm{RA}(X) is \operatorname{Mod}_{p} -capacitable and satisfies
\Big( \operatorname{Mod}_{p}(\Gamma)\Big)^{1/p} = \sup\Big\{ {\boldsymbol \pi}(\Gamma): {\boldsymbol \pi}\in \mathcal M_+( \mathrm{RA}(X)),\ \operatorname{Bar}_{q}( {\boldsymbol \pi})\le 1\Big\}. |
In particular, \Gamma is \operatorname{Mod}_{p} -negligible if and only if {\boldsymbol \pi}(\Gamma) = 0 for every {\boldsymbol \pi}\in {{\cal B}_q}\left( {{\rm{RA}}\left( X \right)} \right) .
Recall that {\mathsf e}_i(\gamma) , i = 0,1 , denote the initial and final points of a rectifiable arc \gamma\in \mathrm{RA}(X) .
Definition 5.3 (Newtonian weak upper gradient). Let f:X\to \mathbb{R} be \mathfrak m -measurable and p -summable function. We say that f belongs to the Newtonian space N^{1,p}( \mathbb{X}) if there exists a nonnegative g\in L^p(X, \mathfrak m) such that
\begin{equation} \Big|f( {\mathsf e}_1(\gamma))-f( {\mathsf e}_0(\gamma))\Big|\le \int_\gamma g\quad{for \; \operatorname{Mod}_{p}-a.e.\; arc \; \gamma\in \mathrm{RA}(X)}. \end{equation} | (5.1) |
In this case, we say that g is a N^{1,p} -weak upper gradient of f .
Functions with \operatorname{Mod}_{p} -weak upper gradient have the important Beppo-Levi property of being absolutely continuous along \operatorname{Mod}_{p} -a.e. arc \gamma (more precisely, this means f\circ R_\gamma is absolutely continuous, see [23, Proposition 3.1]). Notice that functions in N^{1,p}( \mathbb{X}) are everywhere defined. We say that \tilde f\in N^{1,p}( \mathbb{X}) is a good representative of a function f\in L^p(X, \mathfrak m) if \mathfrak m(\{\tilde f\neq f\}) = 0 .
Weak upper gradient enjoys a strong stability property [8, Prop. 2.3], resulting from a refined version of Fuglede's Lemma:
Theorem 5.4. Let f_n\in N^{1,p}( \mathbb{X}), g_n\in L^p(X, \mathfrak m) be two sequences strongly converging to f,g\in L^p(X, \mathfrak m) respectively in L^p(X, \mathfrak m) . If g_n is a N^{1,p} -weak upper gradient of f then there exists a good representative \tilde f\in N^{1,p}( \mathbb{X}) of f and a subsequence k\mapsto n(k) such that f_{n(k)}\to \tilde f quasi everywhere; moreover g is a N^{1,p} -weak upper gradient of \tilde f .
It is clear that a function f\in \mathop{{\rm{Lip}}}\nolimits_b(X,\tau, {\mathsf d}) belongs to N^{1,p}( \mathbb{X}) and \mathop{{\rm{lip}}}\nolimits f is a N^{1,p} -weak upper gradient (it is in fact an upper gradient). By Theorems 5.4 and 3.3 one can easily get that also every f\in H^{1,p}( \mathbb{X}) admits a good representative \tilde f\in N^{1,p}( \mathbb{X}) and | {\mathrm D} f|_\star is a N^{1,p} -weak upper gradient of \tilde f . Equivalently, \tilde f is absolutely continuous along \operatorname{Mod}_{p} -a.e. arc and g satisfies (5.1) \operatorname{Mod}_{p} -a.e.
In fact these two notions are essentially equivalent modulo the choice of a representative in the equivalence class [1,5,6,21]:
Theorem 5.5. Let us suppose that \mathbb{X} is a complete e.m.t.m. space. Every function f\in N^{1,p}( \mathbb{X}) also belongs to H^{1,p}( \mathbb{X}) and every N^{1,p} -weak upper gradient g satisfies g\ge | {\mathrm D} f|_\star \mathfrak m -a.e., so that | {\mathrm D} f|_\star is also the minimal N^{1,p} -weak upper gradient of f .
The Newtonian capacity \operatorname{Cap}_p(E) of a subset E\subset X can be defined as
\begin{equation} \operatorname{Cap}_p(E): = \inf\Big\{ \mathsf{C\kern-1.5pt E}_{p,1}(u):u\in N^{1,p}( \mathbb{X}),\ u\ge 1\text{ on $E$}\Big\}. \end{equation} | (5.2) |
By choosing u as the function taking the constant value 1 it is immediate to check that in our setting the capacity of a set is always finite and
\operatorname{Cap}_p(E)\le \mathfrak m(X) \lt \infty\quad\text{for every }E\subset X. |
It can be proved [8, Prop. 1.48] that
\begin{equation} \text{$E\subset X$ has $0$ capacity if and only if $E$ is $ \mathfrak m$-negligible and } \operatorname{Mod}_{p}(\Gamma_E) = 0, \end{equation} | (5.3) |
so that a property P on X holds quasi everywhere if the set where P fails has 0 capacity. Notice that if \tilde f_i\in N^{1,p}( \mathbb{X}) coincide \mathfrak m -a.e., then \tilde f_1 = \tilde f_2 q.e. in X . Notice moreover that we can use q.e. inequality in (5.2), i.e.,
u\in N^{1,p}( \mathbb{X}),\quad u\ge 1 \text{ q.e. on }E\quad\Rightarrow\quad \operatorname{Cap}_p(E)\le \mathsf{C\kern-1.5pt E}_{p,1}(u). |
We also recall that the capacity satisfies the following properties [8, Thm 1.27, Prop. 2.22, Thm. 6.4]:
\begin{align*} & \operatorname{Cap}_p(\emptyset) = 0 \\& \mathfrak m(E)\le \operatorname{Cap}_p(E) \\& E_1\subset E_2\quad\Rightarrow\quad \operatorname{Cap}_p(E_1)\le \operatorname{Cap}_p(E_2) \\& \operatorname{Cap}_p(E_1\cup E_2)+ \operatorname{Cap}_p(E_1\cap E_2)\le \operatorname{Cap}_p(E_1)+ \operatorname{Cap}_p(E_2) \\& \operatorname{Cap}_p\Big(\bigcup_{n = 1}^\infty E_n)\le \sum\limits_{n = 1}^\infty \operatorname{Cap}_p(E_n) \\& E_n\uparrow E\quad\Rightarrow\quad \operatorname{Cap}_p(E) = \lim\limits_{n\to\infty} \operatorname{Cap}_p(E_n) = \sup\limits_{n \gt 0} \operatorname{Cap}_p(E_n). \end{align*} |
We want now to study the relation between the Newtonian capacity and measures \mu\in {{\mathcal M}}_+(X) with finite energy, according to Definition 3.14. We will denote by
\begin{equation} \mu = \mu^a+\mu^\perp,\quad \mu^a = \varrho \mathfrak m\ll \mathfrak m,\quad \mu^\perp\perp \mathfrak m \end{equation} | (5.4) |
the canonical Lebesgue decomposition of \mu w.r.t. \mathfrak m . Since by a simple truncation argument it is easy to check that
\frac 1q \mathsf{p\kern-1pt C\kern-1.5pt E}_{p,1}^*(\mu) = \sup\Big\{\int_X f\, {\mathrm d}\mu-\frac1p \mathsf{p\kern-1pt C\kern-1.5pt E}_{p,1}(f):f\in \mathop{\rm{Lip }}\nolimits_b(X,\tau, {\mathsf d}),\ f\ge0\Big\} |
we obtain that
\mu\le \nu\quad\Rightarrow\quad \mathsf{p\kern-1pt C\kern-1.5pt E}_{p,1}^*(\mu)\le \mathsf{p\kern-1pt C\kern-1.5pt E}_{p,1}^*(\nu). |
In particular \mathsf{p\kern-1pt C\kern-1.5pt E}_{p,1}^*(\mu^\perp)\le \mathsf{p\kern-1pt C\kern-1.5pt E}_{p,1}^*(\mu)<\infty .
Theorem 5.6. Let \mu\in {{\mathcal M}}_+(X) be a measure with finite energy and let L_\mu\in H^{-1,q}_{\mathrm{pd}}(\mathbb{X}) be the linear functional associated to \mu according to Corollary 3.15.
(a) If E is a universally measurable subset of X with 0 capacity then E is \mu -negligible.
(b) If u\in H^{1,p}( \mathbb{X}) is nonnegative and \tilde u\in N^{1,p}( \mathbb{X}) is a good representative of u , then \tilde u\in L^1(X,\mu) and
\begin{equation} {\langle} L_\mu,u {\rangle} = \int_X \tilde u\, {\mathrm d}\mu\le \Big( \mathsf{p\kern-1pt C\kern-1.5pt E}_{p,1}^*(\mu)\Big)^{1/q} \Big( \mathsf{C\kern-1.5pt E}_{p,1}(u)\Big)^{1/p}. \end{equation} | (5.5) |
Proof. (a) Let E be a set with 0 capacity according to (5.3); since \mathfrak m(E) = 0 , by considering the Lebesgue decomposition (5.4) it is sufficient to show that \mu^\perp(E) = 0 . It is not restrictive to assume \mu^\perp(X)>0 ; by Theorem 4.6 we can find a plan {\boldsymbol \pi}\in {{\cal B}_q}\left( {{\rm{RA}}\left( X \right)} \right) such that
\mu^\perp = ( {\mathsf e}_0)_\sharp {\boldsymbol \pi}. |
It follows that
\mu^\perp(E) = {\boldsymbol \pi}\big\{\gamma\in \mathrm{RA}(X): {\mathsf e}_0(\gamma)\in E\big\} \le {\boldsymbol \pi}(\Gamma_E)\le \operatorname{Mod}_{p}(\Gamma_E)\operatorname{Bar}_{q} ( {\boldsymbol \pi}) = 0. |
(b) Let us first assume that \tilde u\le M for some constant M>0 . We can find a sequence u_n\in \mathop{{\rm{Lip}}}\nolimits_b(X,\tau, {\mathsf d}) taking values in [0,M] , converging to \tilde u q.e. and with u_n\to u, \mathop{{\rm{lip}}}\nolimits u_n\to | {\mathrm D} u|_\star in L^2(X, \mathfrak m) . The uniform bound, the q.e. convergence and the fact that L_\mu\in H^{-1,q}_{\mathrm{pd}}(\mathbb{X}) yield
\lim\limits_{n\to\infty} \int_X u_n\, {\mathrm d}\mu = \int_X \tilde u\, {\mathrm d}\mu = {\langle} L_\mu,\tilde u {\rangle} = {\langle} L_\mu,u {\rangle}\le \Big( \mathsf{p\kern-1pt C\kern-1.5pt E}_{p,1}^*(\mu)\Big)^{1/q} \Big( \mathsf{C\kern-1.5pt E}_{p,1}(u)\Big)^{1/p} |
The case of a general nonnegative u follows by passing to the limit in the sequence u_M: = u\land M as M \uparrow\infty and using monotone convergence.
Theorem 5.7. For every Borel set E\subset X and every measure \mu\in {{\mathcal M}}_+(X) with finite energy we have
\begin{equation} \mu(E)\le \Big( \operatorname{Cap}_p(E)\Big)^{1/p}\Big( \mathsf{p\kern-1pt C\kern-1.5pt E}_{p,1}^*(\mu)\Big)^{1/q}. \end{equation} | (5.6) |
If F\subset X is a closed set then there exists \mu\in {{\mathcal M}}_+(X) supported in F with
\mu(F) = \operatorname{Cap}_p(F) = \mathsf{p\kern-1pt C\kern-1.5pt E}_{p,1}^*(\mu) = \mathsf{C\kern-1.5pt E}_{p,1}(u) |
where u\in \mathbb{N}^{1,p}( \mathbb{X}) realizes the infimum of (5.2) and
L_\mu = {\mathrm J}_pu+ {\mathrm A}_p u\quad\text{in } H^{-1,q}_{\mathrm{pd}}(\mathbb{X}). |
Equivalently, for every closed set F\subset X
\begin{equation} \big( \operatorname{Cap}_p(F)\big)^{1/p} = \max \Big\{\mu(F):\mu\in {{\mathcal M}}_+(X),\ \mathsf{p\kern-1pt C\kern-1.5pt E}_{p,1}^*(\mu)\le 1\Big\}. \end{equation} | (5.7) |
Proof. (5.6) follows easily by (5.5).
Let us now consider the case when F is closed and let us set {{\mathcal K}}: = \{u\in N^{1,2}( \mathbb{X}):\ u\ge 1\; \text{ q.e. on }\;F\} ; {{\mathcal K}} can be identified with a convex subset of H^{1,p}( \mathbb{X}) . Let u_n\in {{\mathcal K}} be a sequence satisfying \lim\limits_{n\to\infty} \mathsf{C\kern-1.5pt E}_{p,1}(u_n) = \operatorname{Cap}_p(F) . By a truncation argument, it is not restrictive to assume that 0\le u_n\le 1 .
By applying Mazur's Theorem and Theorem 5.4 it is not restrictive to assume that there exists u\in N^{1,p}( \mathbb{X}) such that
u_n\to u\quad\text{q.e.},\quad \big\|u_n-u\big\|_{L^2}+\big\| | {\mathrm D} u_n|_\star-| {\mathrm D} u|_\star\big\|_{L^2}\to 0 |
Up to modifying u in a set of 0 capacity, we can thus suppose that u\in {{\mathcal K}} , 0\le u\le 1 , and \mathsf{C\kern-1.5pt E}_{p,1}(u) = \operatorname{Cap}_p(F) . The minimality yields that there exists L\in {\mathrm A}_p(u)+ {\mathrm J}_p(u)\subset H^{-1,p}( \mathbb{X}) such that
{\langle} L,v-u {\rangle}\ge0\quad \text{for every }v\in {{\mathcal K}}. |
In particular, choosing v = u+\phi with \phi nonnegative we get
{\langle} L,\phi {\rangle}\ge0\quad \text{for every $ \phi\in N^{1,p}( \mathbb{X}) $, $ \phi\ge0 $}, |
so that L\in H^{-1,q}_{\mathrm{pd}}(\mathbb{X}) and its action on bounded Lipschitz functions can be represented by a positive Radon measure \mu according to (4.8) thanks to Theorem 4.7
Choosing now \phi\in \mathop{{\rm{Lip}}}\nolimits_b(X,\tau, {\mathsf d}) such that \phi\equiv0 on F we get
{\langle} L,\phi {\rangle} = \int_X \phi\, {\mathrm d}\mu = 0, |
so that \mu(X\setminus F) = 0 and \mu is concentrated on F (recall that \mathop{{\rm{Lip}}}\nolimits_b(X,\tau, {\mathsf d}) generates the \tau topology of X ). Since L has finite energy \mathsf{p\kern-1pt C\kern-1.5pt E}_{p,1}^*(\mu) = \mathsf{C\kern-1.5pt E}_{p,1}^*( L) . Since u\in N^{1,p}( \mathbb{X}) is nonnegative it follows that
\operatorname{Cap}_p(F) = \mathsf{C\kern-1.5pt E}_{p,1}(u) = \mathsf{C\kern-1.5pt E}_{p,1}^*( L) = {\langle} L,u {\rangle} = \int_X u\, {\mathrm d}\mu = \mu(F). |
The renormalization \tilde\mu: = \mu \big( \mathsf{p\kern-1pt C\kern-1.5pt E}_{p,1}^*(\mu)\big)^{-1/q} provides (5.7).
The authors acknowledge support from the MIUR - PRIN project 2017TEXA3H "Gradient flows, Optimal Transport and Metric Measure Structures".
The authors declare no conflict of interest.
[1] | Ambrosio L, Di Marino S, Savaré G (2015) On the duality between p-modulus and probability measures. J Eur Math Soc 17: 1817-1853. |
[2] | Ambrosio L, Erbar M, Savaré G (2016) Optimal transport, Cheeger energies and contractivity of dynamic transport distances in extended spaces. Nonlinear Anal 137: 77-134. |
[3] | Ambrosio L, Ghezzi R (2016) Sobolev and bounded variation functions on metric measure spaces. In: Geometry, analysis and dynamics on sub-Riemannian manifolds. Vol. Ⅱ, Zürich: Eur. Math. Soc., 211-273. |
[4] | Ambrosio L, Gigli N, Savaré G (2008) Gradient Flows in Metric Spaces and in the Space of Probability Measures, 2 Eds., Basel: Birkhäuser Verlag. |
[5] | Ambrosio L, Gigli N, Savaré G (2013) Density of Lipschitz functions and equivalence of weak gradients in metric measure spaces. Rev Mat Iberoam 29: 969-996. |
[6] | Ambrosio L, Gigli N, Savaré G (2014) Calculus and heat flow in metric measure spaces and applications to spaces with Ricci bounds from below. Invent Math 195: 289-391. |
[7] | Ambrosio L, Stra F, Trevisan D (2017) Weak and strong convergence of derivations and stability of flows with respect to MGH convergence. J Funct Anal 272: 1182-1229. |
[8] | Björn A, Björn J (2011) Nonlinear Potential Theory on Metric Spaces, Zürich: European Mathematical Society. |
[9] | Brezis H (2011) Functional Analysis, Sobolev Spaces and Partial Differential Equations, New York: Springer. |
[10] | Cheeger J (1999) Differentiability of Lipschitz functions on metric measure spaces. Geom Funct Anal 9: 428-517. |
[11] | Dellacherie C, Meyer PA (1978) Probabilities and Potential, Amsterdam: North-Holland Publishing Co. |
[12] | Di Marino S (2014) Sobolev and BV spaces on metric measure spaces via derivations and integration by parts. arXiv: 1409.5620. |
[13] | Ekeland I, Temam R (1974) Analyse Convexe et Problèmes Variationnels, Paris: Dunod, GauthierVillars. |
[14] | Engelking R (1989) General Topology, 2 Eds., Berlin: Heldermann Verlag. |
[15] | Fuglede B (1957) Extremal length and functional completion. Acta Math. 98: 171-219. |
[16] | Gagliardo E (1961) A unified structure in various families of function spaces. Compactness and closure theorems, In: Proc. Internat. Sympos. Linear Spaces (Jerusalem, 1960), Jerusalem: Jerusalem Academic Press, 237-241. |
[17] | Heinonen J, Koskela P, Shanmugalingam N, et al. (2015) Sobolev Spaces on Metric Measure Apaces. An Approach Based on Upper Gradients, Cambridge: Cambridge University Press. |
[18] | Koskela P, MacManus P (1998) Quasiconformal mappings and Sobolev spaces. Studia Math 131: 1-17. |
[19] | Munkres JR (2000) Topology, NJ, Upper Saddle River: Prentice Hall Inc. |
[20] | Rockafellar RT (1966) Characterization of the subdifferentials of convex functions. Pacific J Math 17: 497-510. |
[21] | Savaré G (2019) Sobolev spaces in extended metric-measure spaces. arXiv: 1911.04321. |
[22] | Schwartz L (1973) Radon Measures on Arbitrary Topological Spaces and Cylindrical Measures, Tata Institute of Fundamental Research, London: Oxford University Press. |
[23] | Shanmugalingam N (2000) Newtonian spaces: An extension of Sobolev spaces to metric measure spaces. Rev Mat Iberoam 16: 243-279. |
[24] | Villani C (2003) Topics in Optimal Transportation, RI, Providence: American Mathematical Society. |
1. | Luigi Ambrosio, Toni Ikonen, Danka Lučić, Enrico Pasqualetto, Metric Sobolev Spaces I: Equivalence of Definitions, 2024, 1424-9286, 10.1007/s00032-024-00407-7 |