Introduction
Poincaré inequalities are a basic principle of first-order analysis: the oscillation of a function on a ball is controlled by its energy near that ball. On metric measure spaces they provide a way to study Sobolev functions without using coordinates or derivatives, see [22 ] . A macroscopic Poincaré inequality expresses a related principle using only oscillations: variation at a given scale is controlled by averaged variation at any smaller scale. Hytönen and Korte recently established such an inequality on doubling Poincaré spaces, comparing an L 1 L^1 L 1 -mean oscillation with an L p L^p L p -average of smaller-scale oscillations, normalized by the usual power scale r p r^p r p [23 Theorem 3.4] . This raises the question of whether the same principle holds for energies and geometries whose natural scale is different from r p r^p r p .
Regular strongly local p p p -energy forms provide a robust newly developed framework for this question. For 1 < p < ∞ 1 < p < \infty 1 < p < ∞ , such a form ( E , F ) (\mathcal{E},\mathcal{F}) ( E , F ) consists of a Sobolev domain F \mathcal{F} F and a p p p -energy E \mathcal{E} E ; see [26 , 27 , 41 ] . This framework includes the usual gradient energies on Riemannian manifolds and Carnot groups, as well as energies on fractals. In the classical examples the natural scale is Ψ ( r ) = r p \Psi(r) = r^p Ψ ( r ) = r p , whereas on fractals it may be a different power or even a nonpower function. Moreover, an energy measure may be singular with respect to the underlying measure, so that there need not be a gradient density to integrate. These features make it useful to formulate Poincaré inequalities directly in terms of energy measures and a scale function Ψ \Psi Ψ .
Our main result is a macroscopic Poincaré inequality for p p p -energy forms at general scales. We work on a complete doubling metric measure space, with a doubling scale function Ψ \Psi Ψ and an energy form satisfying the standing Assumption 2.1 of Section 2. We assume the usual Ψ \Psi Ψ -Poincaré inequality and a capacity upper bound, which says that cutoff functions for a ball can be chosen with energy controlled by the volume of the ball divided by its scale. For f ∈ L l o c p ( X ; m ) f \in L^p_{\mathrm{loc}}(X;m) f ∈ L loc p ( X ; m ) , write
M f ( x , r ) : = 1 Ψ ( r ) ∫ B ( x , r ) ∣ f − f B ( x , r ) ∣ p d m . M_f(x,r) := \frac{1}{\Psi(r)} \int_{B(x,r)} |f-f_{B(x,r)}|^p \,dm. M f ( x , r ) := Ψ ( r ) 1 ∫ B ( x , r ) ∣ f − f B ( x , r ) ∣ p d m .
Theorem 2.2 gives constants C M > 0 C_M > 0 C M > 0 and A M ≥ 1 A_M \ge1 A M ≥ 1 such that
M f ( x , s ) ≤ C M ∫ B ( x , A M s ) M f ( y , r ) d m ( y ) , x ∈ X , 0 < r < s . M_f(x,s) \le C_M \int_{B(x,A_Ms)} M_f(y,r)\,dm(y), \qquad x \in X,\quad0 < r < s. M f ( x , s ) ≤ C M ∫ B ( x , A M s ) M f ( y , r ) d m ( y ) , x ∈ X , 0 < r < s .
Thus normalized oscillation at a given scale is controlled by its average at every smaller scale, with constants independent of f f f , r r r , and s s s . For Ψ ( r ) = r p \Psi(r)=r^p Ψ ( r ) = r p it is close in spirit to [23 Theorem 3.4] , but the two statements are not directly comparable: the left-hand side there is an L 1 L^1 L 1 -mean oscillation, whereas both sides of our inequality involve L p L^p L p -oscillations.
Thus an inequality initially involving the energy yields a comparison that depends only on the metric, the measure, and the scale function. In particular, the function need not belong to the energy domain for the macroscopic inequality to apply. The proof uses partitions of unity constructed from capacitary cutoffs, allowing the argument to work with singular energy measures. Related partitions and discrete approximations in p p p -Dirichlet spaces appear in [17 Lemmas 4.2 and 4.5] . Only doubling of Ψ \Psi Ψ is assumed at this stage; no lower power bound is required as an input. A lower p p p -scaling bound is derived afterwards as a consequence (Corollary 3.12).
Our first application is a uniform comparison of the Korevaar–Schoen functionals [29 ] . For f ∈ L p ( X ; m ) f \in L^p(X;m) f ∈ L p ( X ; m ) , set
J r ( f ) : = 1 Ψ ( r ) ∫ X ∫ B ( x , r ) ∣ f ( x ) − f ( y ) ∣ p d m ( y ) d m ( x ) . (1) J_r(f) := \frac{1}{\Psi(r)} \int_X \int_{B(x,r)} |f(x)-f(y)|^p\,dm(y)\,dm(x). \tag*{(1)} J r ( f ) := Ψ ( r ) 1 ∫ X ∫ B ( x , r ) ∣ f ( x ) − f ( y ) ∣ p d m ( y ) d m ( x ) . (1)
Integrating the macroscopic inequality gives the two-scale estimate J s ( f ) ≲ J r ( f ) J_s(f) \lesssim J_r(f) J s ( f ) ≲ J r ( f ) whenever 0 < r < s 0<r<s 0 < r < s , with a constant independent of both scales. Consequently, control along any sequence of radii tending to zero gives control at every radius. This strengthens the weak monotonicity property
sup r > 0 J r ( f ) ≲ lim inf r ↓ 0 J r ( f ) . \sup_{r>0} J_r(f) \lesssim\liminf_{r\downarrow0} J_r(f). r > 0 sup J r ( f ) ≲ r ↓ 0 lim inf J r ( f ) .
A version of the latter for p = 1 p=1 p = 1 appears in the study of B V BV B V functions in [2 ] ; for general p p p it was formulated as property P ( p , α ) P(p,\alpha) P ( p , α ) in [4 Definition 4.5] . The terminology weak monotonicity is used in [18 ] and [27 Definition 6.3.5] . Related results on trees appear in [5 ] , and on self-similar Laakso-type spaces in [3 Theorem 7.1] . To our knowledge, the available results on fractals and in Dirichlet-form or p p p -energy settings, for instance [18 , 10 , 11 , 27 ] and [38 Theorem 1.4] , establish weak monotonicity in this sup–lim inf form. For Ψ ( r ) = r p \Psi(r)=r^p Ψ ( r ) = r p on doubling Poincaré spaces, a uniform two-scale comparison of the corresponding L 1 L^1 L 1 -oscillation functionals follows by integrating [23 Theorem 3.4] ; the two-scale comparison of the Korevaar–Schoen functionals at general scales Ψ \Psi Ψ appears to be new.
A second application is geometric. Proposition 2.4 shows that volume doubling and the macroscopic Poincaré inequality imply that X X X is connected and force the scale function to satisfy
Ψ ( r ) Ψ ( R ) ≲ ( r R ) p , 0 < r ≤ R ≤ diam X . \frac{\Psi(r)}{\Psi(R)} \lesssim\left(\frac{r}{R}\right)^p,\qquad0<r\le R\le\operatorname{diam} X. Ψ ( R ) Ψ ( r ) ≲ ( R r ) p , 0 < r ≤ R ≤ diam X .
When X X X is unbounded, there is no upper restriction on R R R . The proposition requires no energy form, chain condition, or doubling assumption on Ψ \Psi Ψ . It obtains the lower p p p -scaling restriction of [44 ] directly from an inequality involving only oscillations. Together with doubling of Ψ \Psi Ψ , this bound shows that every scale function admissible in our forward results is of two-sided power type (Remark 3.13); such scale functions need not be powers, as the tree and Laakso-type examples show. The lower bound will also be used in the weak-type characterization below.
Under our hypotheses, the domain characterization of Shimizu [43 Corollaries 1.14 and 1.15] gives
F = { f ∈ L p ( X ; m ) : lim inf r ↓ 0 J r ( f ) < ∞ } = B Ψ , B Ψ : = { f ∈ L p ( X ; m ) : sup r > 0 J r ( f ) < ∞ } , \mathcal{F}=\left\{f\in L^p(X;m):\liminf_{r\downarrow0}J_r(f)<\infty\right\}=\mathcal{B}_{\Psi},\qquad\mathcal{B}_{\Psi}:=\left\{f\in L^p(X;m):\sup_{r>0}J_r(f)<\infty\right\}, F = { f ∈ L p ( X ; m ) : r ↓ 0 lim inf J r ( f ) < ∞ } = B Ψ , B Ψ := { f ∈ L p ( X ; m ) : r > 0 sup J r ( f ) < ∞ } ,
and, for f ∈ F f\in\mathcal{F} f ∈ F ,
E ( f ) ≳ lim inf r ↓ 0 J r ( f ) ≳ sup r > 0 J r ( f ) . \mathcal{E}(f)\gtrsim\liminf_{r\downarrow0}J_r(f)\gtrsim\sup_{r>0}J_r(f). E ( f ) ≳ r ↓ 0 lim inf J r ( f ) ≳ r > 0 sup J r ( f ) .
We recall this essentially known characterization without proof in Subsection 3.8, with comments on its applicability under the present hypotheses because t \mathrm{t} t will be used in the reconstruction of the energy and in the weak-type characterization below. For connections with other Sobolev theories on metric spaces, see [30 , 1 ] .
The converse asks whether an energy can be constructed from the macroscopic inequality itself. Starting with a complete doubling metric measure space and a doubling scale function satisfying this inequality, we assume that B Ψ ∩ C c ( X ) \mathcal{B}_{\Psi}\cap C_c(X) B Ψ ∩ C c ( X ) is uniformly dense in C c ( X ) C_c(X) C c ( X ) . This assumption cannot be dropped: the macroscopic inequality persists when Ψ \Psi Ψ is multiplied by a nondecreasing function (Remark 4.2), whereas Korevaar–Schoen classes at too fast a scale contain only constants. A common subsequential limit of the functionals J r J_r J r then yields a regular strongly local p p p -energy on the domain F K S \mathcal{F}_{\mathrm{KS}} F KS obtained by closing B Ψ ∩ C c ( X ) \mathcal{B}_{\Psi}\cap C_c(X) B Ψ ∩ C c ( X ) in its natural Besov norm. The energy is comparable to sup r > 0 J r \sup_{r>0}J_r sup r > 0 J r on this domain and satisfies a Ψ \Psi Ψ -Poincaré inequality with a fixed dilation of balls. The existence of the form follows from the Korevaar–Schoen theory of p p p -energy forms [27 , 41 ] once weak monotonicity is available; the new point is the Poincaré inequality, which we recover from the macroscopic inequality. For related constructions of p p p -energies as limits of discrete or nonlocal energies, see [28 , 9 , 38 ] and, for p = 2 p=2 p = 2 , [31 ] . Under the forward hypotheses, F K S = F \mathcal{F}_{\mathrm{KS}}=\mathcal{F} F KS = F and the constructed energy is equivalent to the original one. For a preassigned energy form, the converse takes a related form: Poincaré is equivalent to the macroscopic inequality together with a local bound on small-scale oscillations by that energy, under volume doubling and the capacity upper bound (Proposition 3.10).
Our final result gives a second description of the same energy domain, using the distribution of pairwise differences instead of their averages at a fixed radius. In Euclidean space, the Brezis–Van Schaftingen–Yung formula recovers the Sobolev energy from the weak L p L^p L p size of the difference quotient ∣ f ( x ) − f ( y ) ∣ / ∣ x − y ∣ 1 + n / p |f(x)-f(y)|/|x-y|^{1+n/p} ∣ f ( x ) − f ( y ) ∣/∣ x − y ∣ 1 + n / p [8 , 7 , 15 ] . Poliakovsky’s characterization further shows that, for an L p L^p L p function on a Euclidean Lipschitz domain, the behavior of the distribution at high levels already determines Sobolev membership [39 Theorem 1.3] . Metric versions for Lipschitz test functions were developed in [13 , 12 ] , and Hytönen–Yang–Yuan–Zhao characterize the full homogeneous Sobolev space defined by p p p -weak upper gradients on complete doubling p p p -Poincaré spaces [24 Theorem 1.1(ii)] . Related nonlocal characterizations of Sobolev and B V BV B V functions appear in [14 , 32 , 33 , 6 ] .
In our setting, the scale and volume determine the appropriate quotient. Writing V ( x , r ) = m ( B ( x , r ) ) V(x,r)=m(B(x,r)) V ( x , r ) = m ( B ( x , r )) , set
K Ψ , f ( x , y ) : = ∣ f ( x ) − f ( y ) ∣ p Ψ ( d ( x , y ) ) V ( x , d ( x , y ) ) , x ≠ y , K_{\Psi,f}(x,y) := \frac{|f(x)-f(y)|^p}{\Psi(d(x,y))V(x,d(x,y))}, \qquad x\ne y, K Ψ , f ( x , y ) := Ψ ( d ( x , y )) V ( x , d ( x , y )) ∣ f ( x ) − f ( y ) ∣ p , x = y ,
and set K Ψ , f = 0 K_{\Psi,f}=0 K Ψ , f = 0 on the diagonal. Under the forward hypotheses, Theorem 2.6 shows that an L p L^p L p function belongs to F \mathcal{F} F if and only if either of the following two quantities is finite; for f ∈ F f\in\mathcal{F} f ∈ F , both are equivalent to the energy:
E ( f ) ≍ sup λ > 0 λ ( m ⊗ m ) ( { K Ψ , f > λ } ) ≍ lim sup λ → ∞ λ ( m ⊗ m ) ( { K Ψ , f > λ } ) . \mathcal{E}(f) \asymp\sup_{\lambda>0}\lambda(m\otimes m)(\{K_{\Psi,f}>\lambda\}) \asymp\limsup_{\lambda\to\infty}\lambda(m\otimes m)(\{K_{\Psi,f}>\lambda\}). E ( f ) ≍ λ > 0 sup λ ( m ⊗ m ) ({ K Ψ , f > λ }) ≍ λ → ∞ lim sup λ ( m ⊗ m ) ({ K Ψ , f > λ }) .
For Ψ ( r ) = r p \Psi(r)=r^p Ψ ( r ) = r p and upper-gradient energies, the weak-norm characterization is covered by the metric theory cited above. Here the same description extends to doubling scale functions other than r p r^p r p and to possibly singular energy measures.
The link with the preceding results is the macroscopic control of oscillations. A bound on the high-level tail of K Ψ , f K_{\Psi,f} K Ψ , f yields control of localized Korevaar–Schoen functionals along small scales. The macroscopic inequality propagates this control, and the domain characterization then gives membership in F \mathcal{F} F . In the other direction, Poincaré estimates and a covering argument bound the weak quantity by the energy, using the energy measure directly even when it is singular. The Korevaar–Schoen and Brezis–Van Schaftingen–Yung descriptions thus recover the same first-order structure from two different ways of measuring pairwise differences.
Section 2 gives the standing assumptions, states the main results, and presents examples. Section 3 proves the macroscopic Poincaré inequality and its relation with microscopic energy control, derives the two-scale comparison of the Korevaar–Schoen functionals and lower p p p -scaling, and recalls the identification of the energy domain. Section 4 constructs a Korevaar–Schoen p p p -energy from the macroscopic inequality. Finally, Section 5 proves the Brezis–Van Schaftingen–Yung-type characterization.
Standing framework and main results
In this section, we present the framework and our main results. Proofs are given in the later sections.
Metric measure space and scale function
Fix p ∈ ( 1 , ∞ ) p\in(1,\infty) p ∈ ( 1 , ∞ ) . Let ( X , d ) (X,d) ( X , d ) be a nontrivial complete metric space, that is, a complete metric space with at least two points. For x ∈ X x\in X x ∈ X and r > 0 r>0 r > 0 write
B ( x , r ) = { y ∈ X : d ( x , y ) < r } . B(x,r)=\{y\in X:d(x,y)<r\}. B ( x , r ) = { y ∈ X : d ( x , y ) < r } .
Let m m m be a positive Radon measure on X X X with full support such that
0 < m ( B ( x , r ) ) < ∞ , x ∈ X , r > 0. 0 < m(B(x,r)) < \infty,\qquad x \in X,\ r > 0. 0 < m ( B ( x , r )) < ∞ , x ∈ X , r > 0.
Set V ( x , r ) = m ( B ( x , r ) ) V(x,r)=m(B(x,r)) V ( x , r ) = m ( B ( x , r )) . If B = B ( x , r ) B=B(x,r) B = B ( x , r ) and λ > 0 \lambda>0 λ > 0 , write λ B = B ( x , λ r ) \lambda B=B(x,\lambda r) λ B = B ( x , λ r ) . For a measurable set A A A with 0 < m ( A ) < ∞ 0<m(A)<\infty 0 < m ( A ) < ∞ put
h A = ∫ A h d m = 1 m ( A ) ∫ A h d m . h_A=\int_A h\,\mathrm{d}m=\frac{1}{m(A)}\int_A h\,\mathrm{d}m. h A = ∫ A h d m = m ( A ) 1 ∫ A h d m .
Since X = ⋃ n ≥ 1 B ( o , n ) X=\bigcup_{n\geq1}B(o,n) X = ⋃ n ≥ 1 B ( o , n ) for any o ∈ X o\in X o ∈ X , the measure m m m is σ \sigma σ -finite.
The volume doubling condition is
V ( x , 2 r ) ≤ C D V ( x , r ) , x ∈ X , r > 0 , V(x,2r)\leq C_DV(x,r),\qquad x\in X,\ r>0, V ( x , 2 r ) ≤ C D V ( x , r ) , x ∈ X , r > 0 ,
VD
for some C D ≥ 1 C_D\geq1 C D ≥ 1 .
For a ≥ 1 a\geq1 a ≥ 1 , set
C D ( a ) : = C D ⌈ log 2 a ⌉ . (2) C_D(a):=C_D^{\lceil\log_2 a\rceil}. \tag*{(2)} C D ( a ) := C D ⌈ l o g 2 a ⌉ . (2)
It is well known that (VD) implies V ( x , a r ) ≤ C D ( a ) V ( x , r ) V(x,ar)\leq C_D(a)V(x,r) V ( x , a r ) ≤ C D ( a ) V ( x , r ) whenever a ≥ 1 a\geq1 a ≥ 1 .
By the classical packing argument that we recall in Lemma 3.1, (VD) implies that ( X , d ) (X,d) ( X , d ) is metrically doubling. Completeness then implies properness. In particular, X X X is locally compact and separable, and every ball has compact closure. We use these consequences whenever (VD) is assumed.
Let Ψ : [ 0 , ∞ ) → [ 0 , ∞ ) \Psi:[0,\infty)\to[0,\infty) Ψ : [ 0 , ∞ ) → [ 0 , ∞ ) be a doubling homeomorphism: there is C Ψ ≥ 1 C_\Psi\geq1 C Ψ ≥ 1 such that
Ψ ( 2 r ) ≤ C Ψ Ψ ( r ) , r > 0. (3) \Psi(2r)\leq C_\Psi\Psi(r),\qquad r>0. \tag*{(3)} Ψ ( 2 r ) ≤ C Ψ Ψ ( r ) , r > 0. (3)
For a ≥ 1 a\geq1 a ≥ 1 , set C Ψ ( a ) : = C Ψ ⌈ log 2 a ⌉ C_\Psi(a):=C_\Psi^{\lceil\log_2 a\rceil} C Ψ ( a ) := C Ψ ⌈ l o g 2 a ⌉ . Then
Ψ ( a r ) ≤ C Ψ ( a ) Ψ ( r ) , r > 0. (4) \Psi(ar)\leq C_\Psi(a)\Psi(r),\qquad r>0. \tag*{(4)} Ψ ( a r ) ≤ C Ψ ( a ) Ψ ( r ) , r > 0. (4)
Throughout, m ⊗ m m\otimes m m ⊗ m denotes the corresponding product measure. For nonnegative quantities A A A and B B B , the notation A ≲ B A\lesssim B A ≲ B means A ≤ C B A\leq CB A ≤ C B , A ≳ B A\gtrsim B A ≳ B means B ≲ A B\lesssim A B ≲ A , and A ≍ B A\asymp B A ≍ B means that both inequalities hold. Unless otherwise stated, C C C denotes a positive constant that may change from line to line and depends only on the data explicitly named in the statement under consideration.
For every positive σ \sigma σ -finite Radon measure μ \mu μ and every Borel, locally μ \mu μ -integrable function h h h , parameter-dependent integration shows that the maps ( x , r ) ↦ μ ( B ( x , r ) ) (x,r)\mapsto\mu(B(x,r)) ( x , r ) ↦ μ ( B ( x , r )) and ( x , r ) ↦ ∫ B ( x , r ) h d μ (x,r)\mapsto\int_{B(x,r)}h\,\mathrm{d}\mu ( x , r ) ↦ ∫ B ( x , r ) h d μ are Borel. The same holds for ball averages whenever their denominators are positive and finite.
For u ∈ L l o c 1 ( X ; m ) u\in L^1_{\mathrm{loc}}(X;m) u ∈ L loc 1 ( X ; m ) we use the representative-independent measure-theoretic support
supp m u : = supp ( ∣ u ∣ m ) . (5) \operatorname{supp}_m u:=\operatorname{supp}(|u|m). \tag*{(5)} supp m u := supp ( ∣ u ∣ m ) . (5)
Thus u = 0 u=0 u = 0 m m m -a.e. on X ∖ supp m u X\setminus\operatorname{supp}_m u X ∖ supp m u . This convention is useful because the energy measures below need not be absolutely continuous with respect to m m m .
p-Energy forms and analytic assumptions
Let ( E , F ) (\mathcal{E},\mathcal{F}) ( E , F ) be a p p p -energy form on L p ( X ; m ) L^p(X;m) L p ( X ; m ) with domain F ⊂ L p ( X ; m ) \mathcal{F}\subset L^p(X;m) F ⊂ L p ( X ; m ) . The forward results are stated under the following standing assumption.
Assumption 2.1. (E1) ( E , F ) (\mathcal{E},\mathcal{F}) ( E , F ) is a strongly local regular p p p -energy form in the sense of Kajino–Shimizu and Sasaya [26 , 41 ] : F \mathcal{F} F is dense in L p ( X ; m ) L^p(X;m) L p ( X ; m ) , E 1 / p \mathcal{E}^{1/p} E 1/ p is a seminorm, and the form is closed, Markovian, regular, and strongly local.
(E2) The p p p -Clarkson inequalities hold: with q = p / ( p − 1 ) q=p/(p-1) q = p / ( p − 1 ) , for every f , g ∈ F f,g\in\mathcal{F} f , g ∈ F ,
2 ( E ( f ) q / p + E ( g ) q / p ) p − 1 ≤ E ( f + g ) + E ( f − g ) ≤ 2 ( E ( f ) + E ( g ) ) , 1 < p ≤ 2. (6) \begin{aligned}
2\bigl(\mathcal{E}(f)^{q/p}+\mathcal{E}(g)^{q/p}\bigr)^{p-1} &\leq\mathcal{E}(f+g)+\mathcal{E}(f-g)\\
&\leq2\bigl(\mathcal{E}(f)+\mathcal{E}(g)\bigr), \qquad1<p\leq2. \tag*{(6)}
\end{aligned} 2 ( E ( f ) q / p + E ( g ) q / p ) p − 1 ≤ E ( f + g ) + E ( f − g ) ≤ 2 ( E ( f ) + E ( g ) ) , 1 < p ≤ 2. (6)
and
2 ( E ( f ) + E ( g ) ) ≤ E ( f + g ) + E ( f − g ) ≤ 2 ( E ( f ) q / p + E ( g ) q / p ) p − 1 , 2 ≤ p < ∞ . (7) \begin{aligned}
2\bigl(\mathcal{E}(f)+\mathcal{E}(g)\bigr) &\leq\mathcal{E}(f+g)+\mathcal{E}(f-g)\\
&\leq2\bigl(\mathcal{E}(f)^{q/p}+\mathcal{E}(g)^{q/p}\bigr)^{p-1}, \qquad2\leq p<\infty. \tag*{(7)}
\end{aligned} 2 ( E ( f ) + E ( g ) ) ≤ E ( f + g ) + E ( f − g ) ≤ 2 ( E ( f ) q / p + E ( g ) q / p ) p − 1 , 2 ≤ p < ∞. (7)
(E3) F ∩ L ∞ ( X ; m ) \mathcal{F}\cap L^\infty(X;m) F ∩ L ∞ ( X ; m ) is an algebra and
E ( f g ) 1 / p ≤ ∥ f ∥ ∞ E ( g ) 1 / p + ∥ g ∥ ∞ E ( f ) 1 / p , f , g ∈ F ∩ L ∞ ( X ; m ) . (8) \mathcal{E}(fg)^{1/p}\leq\|f\|_\infty\mathcal{E}(g)^{1/p}+\|g\|_\infty\mathcal{E}(f)^{1/p},\qquad f,g\in\mathcal{F}\cap L^\infty(X;m). \tag*{(8)} E ( f g ) 1/ p ≤ ∥ f ∥ ∞ E ( g ) 1/ p + ∥ g ∥ ∞ E ( f ) 1/ p , f , g ∈ F ∩ L ∞ ( X ; m ) . (8)
(E4) The canonical p p p -energy measures satisfy the localization property (G4) below.
At p = 2 p=2 p = 2 the two Clarkson inequalities agree. Under (E1), the normal-contraction property
T ∘ f ∈ F , E ( T ∘ f ) ≤ E ( f ) (9) T\circ f\in\mathcal{F},\qquad\mathcal{E}(T\circ f)\leq\mathcal{E}(f) \tag*{(9)} T ∘ f ∈ F , E ( T ∘ f ) ≤ E ( f ) (9)
whenever T : R → R T:\mathbb{R}\to\mathbb{R} T : R → R is 1-Lipschitz, T ( 0 ) = 0 T(0)=0 T ( 0 ) = 0 , and f ∈ F f\in\mathcal{F} f ∈ F , follows from the energy-form axioms; see Sasaya [41, Proposition 4.5].
For f ∈ F f\in\mathcal{F} f ∈ F , let Γ ( f ) \Gamma(f) Γ ( f ) denote its canonical p p p -energy measure, which exists under (E1) by Sasaya [41, Theorem 2.4]. We also introduce the local domain
F l o c = { u ∈ L l o c p ( X ; m ) : for every relatively compact open U ⊂ X , there is u # ∈ F with u = u # m -a.e. on U } . (10) \mathcal{F}_{\mathrm{loc}}=\left\{u\in L^p_{\mathrm{loc}}(X;m):\begin{array}{l}
\text{for every relatively compact open }U\subset X,\\
\text{there is }u^\#\in\mathcal{F}\text{ with }u=u^\#\ \text{$m$-a.e. on }U
\end{array}\right\}. \tag*{(10)} F loc = { u ∈ L loc p ( X ; m ) : for every relatively compact open U ⊂ X , there is u # ∈ F with u = u # m -a.e. on U } . (10)
We use the following properties of the energy measures:
(G1) Γ ( f ) \Gamma(f) Γ ( f ) is a positive Radon measure and Γ ( f ) ( X ) = E ( f ) \Gamma(f)(X)=\mathcal{E}(f) Γ ( f ) ( X ) = E ( f ) for every f ∈ F f\in\mathcal{F} f ∈ F ;
(G2) for every Borel set A A A , the map f ↦ Γ ( f ) ( A ) 1 / p f\mapsto\Gamma(f)(A)^{1/p} f ↦ Γ ( f ) ( A ) 1/ p is a seminorm on F \mathcal{F} F ;
(G3) if f , g ∈ F ∩ C c ( X ) f,g\in\mathcal{F}\cap C_c(X) f , g ∈ F ∩ C c ( X ) and f − g f-g f − g is constant on a Borel set A A A , then Γ ( f ) ( A ) = Γ ( g ) ( A ) \Gamma(f)(A)=\Gamma(g)(A) Γ ( f ) ( A ) = Γ ( g ) ( A ) ;
(G4) if f , g ∈ F f,g\in\mathcal{F} f , g ∈ F agree m m m -a.e. on an open set U U U , then Γ ( f ) ∣ U = Γ ( g ) ∣ U \left.\Gamma(f)\right|_U=\left.\Gamma(g)\right|_U Γ ( f ) ∣ U = Γ ( g ) ∣ U . Consequently the map f ↦ Γ ( f ) f\mapsto\Gamma(f) f ↦ Γ ( f ) localizes consistently to F l o c \mathcal{F}_{\mathrm{loc}} F loc .
The global properties (G1)–(G3) follow from Sasaya [41, Theorem 2.4]. The localization property (G4), that is, (E4), is imposed separately as the local-domain compatibility needed to define Γ ( u ) \Gamma(u) Γ ( u ) for u ∈ F l o c u\in\mathcal{F}_{\mathrm{loc}} u ∈ F loc ; compare Yang [44, paragraph preceding Lemma 3.4].
The Poincaré inequality PI ( Ψ ) \operatorname{PI}(\Psi) PI ( Ψ ) means that there are C P > 0 C_P>0 C P > 0 and A P ≥ 1 A_P\geq1 A P ≥ 1 such that
∫ B ∣ u − u B ∣ p d m ≤ C P Ψ ( R ) Γ ( u ) ( A P B ) \int_B |u-u_B|^p\,dm\leq C_P\Psi(R)\Gamma(u)(A_PB) ∫ B ∣ u − u B ∣ p d m ≤ C P Ψ ( R ) Γ ( u ) ( A P B )
for every ball B = B ( x , R ) B=B(x,R) B = B ( x , R ) and u ∈ F u\in\mathcal{F} u ∈ F . For Borel sets A 1 , A 2 ⊂ X A_1,A_2\subset X A 1 , A 2 ⊂ X define
cap ( A 1 , A 2 ) = inf { E ( ϕ ) : ϕ ∈ F , ϕ = 1 m -a.e. on an open neighborhood of A 1 , ϕ = 0 m -a.e. on an open neighborhood of A 2 } . \operatorname{cap}(A_1,A_2)=\inf\left\{\mathcal{E}(\phi):\begin{array}{l}
\phi\in\mathcal{F},\ \phi=1\ m\text{-a.e. on an open neighborhood of }A_1,\\
\phi=0\ m\text{-a.e. on an open neighborhood of }A_2
\end{array}\right\}. cap ( A 1 , A 2 ) = inf { E ( ϕ ) : ϕ ∈ F , ϕ = 1 m -a.e. on an open neighborhood of A 1 , ϕ = 0 m -a.e. on an open neighborhood of A 2 } .
The capacity upper bound cap ( Ψ ) ≤ \operatorname{cap}(\Psi)_{\leq} cap ( Ψ ) ≤ means that there are C c > 0 C_c>0 C c > 0 and A c > 1 A_c>1 A c > 1 such that
cap ( B ( x , R ) , X ∖ B ( x , A c R ) ) ≤ C c V ( x , R ) Ψ ( R ) , x ∈ X , 0 < R < diam X . \operatorname{cap}\bigl(B(x,R),X\setminus B(x,A_cR)\bigr)\leq C_c\frac{V(x,R)}{\Psi(R)},\qquad x\in X,\quad0<R<\operatorname{diam}X. cap ( B ( x , R ) , X ∖ B ( x , A c R ) ) ≤ C c Ψ ( R ) V ( x , R ) , x ∈ X , 0 < R < diam X .
Thus, when X X X is bounded, no capacity condition at or above its diameter is included in cap ( Ψ ) ≤ \operatorname{cap}(\Psi)_{\leq} cap ( Ψ ) ≤ . At an admissible radius for which B ( x , A c R ) = X B(x,A_cR)=X B ( x , A c R ) = X , the estimate is automatic. Indeed, boundedness and the preceding properness observation make X X X compact; the regularity and normal-contraction argument used in the proof of Lemma 3.2 gives 1 ∈ F 1\in\mathcal{F} 1 ∈ F , and (G1) and (G3) give E ( 1 ) = 0 \mathcal{E}(1)=0 E ( 1 ) = 0 .
When the corresponding assumptions are in force, the structural data are
p , C D , C Ψ , C P , A P , C c , A c . p,\ C_D,\ C_\Psi,\ C_P,\ A_P,\ C_c,\ A_c. p , C D , C Ψ , C P , A P , C c , A c .
Unless otherwise specified, constants in the forward results depend only on these data.
Macroscopic inequality
For f ∈ L l o c p ( X ; m ) f \in L^p_{\mathrm{loc}}(X;m) f ∈ L loc p ( X ; m ) set
M f ( x , r ) = 1 Ψ ( r ) ∫ B ( x , r ) ∣ f − f B ( x , r ) ∣ p d m . (11) M_f(x,r)=\frac{1}{\Psi(r)}\int_{B(x,r)}\lvert f-f_{B(x,r)}\rvert^p\,\mathrm{d}m. \tag*{(11)} M f ( x , r ) = Ψ ( r ) 1 ∫ B ( x , r ) ∣ f − f B ( x , r ) ∣ p d m . (11)
The preceding measurability observation shows that ( x , r ) ↦ M f ( x , r ) (x,r)\mapsto M_f(x,r) ( x , r ) ↦ M f ( x , r ) is Borel. It is unchanged by altering f f f on an m m m -null set, so the averages below are independent of the representative.
We say that the macroscopic Poincaré inequality ( m P I ( Ψ ) ) (\mathrm{mPI}(\Psi)) ( mPI ( Ψ )) holds if there exist C M > 0 C_M>0 C M > 0 and A M ≥ 1 A_M\ge1 A M ≥ 1 such that for every f ∈ L l o c p ( X ; m ) f\in L^p_{\mathrm{loc}}(X;m) f ∈ L loc p ( X ; m ) and x ∈ X x\in X x ∈ X ,
M f ( x , s ) ≤ C M ∫ B ( x , A M s ) M f ( y , r ) d m ( y ) , 0 < r < s . M_f(x,s)\le C_M\int_{B(x,A_Ms)}M_f(y,r)\,\mathrm{d}m(y),\qquad0<r<s. M f ( x , s ) ≤ C M ∫ B ( x , A M s ) M f ( y , r ) d m ( y ) , 0 < r < s .
m P I ( Ψ ) \mathrm{mPI}(\Psi) mPI ( Ψ )
Theorem 2.2. Suppose that Assumption 2.1 holds, together with (VD), ( P I ( Ψ ) ) (\mathrm{PI}(\Psi)) ( PI ( Ψ )) , and ( c a p ( Ψ ) ≤ ) (\mathrm{cap}(\Psi)_{\le}) ( cap ( Ψ ) ≤ ) . Then ( m P I ( Ψ ) ) (\mathrm{mPI}(\Psi)) ( mPI ( Ψ )) holds, with constants depending only on the structural data.
Consequences of the macroscopic inequality
Define the Korevaar–Schoen functional
J r ( f ) : = 1 Ψ ( r ) ∫ X ∫ B ( x , r ) ∣ f ( x ) − f ( y ) ∣ p d m ( y ) d m ( x ) , (12) J_r(f):=\frac{1}{\Psi(r)}\int_X\int_{B(x,r)}\lvert f(x)-f(y)\rvert^p\,\mathrm{d}m(y)\,\mathrm{d}m(x), \tag*{(12)} J r ( f ) := Ψ ( r ) 1 ∫ X ∫ B ( x , r ) ∣ f ( x ) − f ( y ) ∣ p d m ( y ) d m ( x ) , (12)
and the Besov–Korevaar–Schoen space
B Ψ : = { f ∈ L p ( X ; m ) : sup r > 0 J r ( f ) < ∞ } , B_\Psi:=\{f\in L^p(X;m):\sup_{r>0}J_r(f)<\infty\}, B Ψ := { f ∈ L p ( X ; m ) : r > 0 sup J r ( f ) < ∞ } ,
∥ f ∥ B Ψ : = ∥ f ∥ L p + ( sup r > 0 J r ( f ) ) 1 / p . (13) \lVert f\rVert_{B_\Psi}:=\lVert f\rVert_{L^p}+\left(\sup_{r>0}J_r(f)\right)^{1/p}. \tag*{(13)} ∥ f ∥ B Ψ := ∥ f ∥ L p + ( r > 0 sup J r ( f ) ) 1/ p . (13)
The macroscopic inequality integrates to a uniform two-scale comparison of the Korevaar–Schoen functionals; the proof is given in Subsection 3.6.
Theorem 2.3. Assume (VD), (2.2), and ( m P I ( Ψ ) ) (\mathrm{mPI}(\Psi)) ( mPI ( Ψ )) . Then for every f ∈ L p ( X ; m ) f\in L^p(X;m) f ∈ L p ( X ; m ) and 0 < r < s 0<r<s 0 < r < s ,
J s ( f ) ≤ C J r ( f ) , (14) J_s(f)\le C J_r(f), \tag*{(14)} J s ( f ) ≤ C J r ( f ) , (14) where C C C depends only on p p p , C D C_D C D , C Ψ C_\Psi C Ψ , C M C_M C M , and A M A_M A M . In particular,
sup s > 0 J s ( f ) ≤ C lim inf r ↓ 0 J r ( f ) . (15) \sup_{s>0}J_s(f)\le C\liminf_{r\downarrow0}J_r(f). \tag*{(15)} s > 0 sup J s ( f ) ≤ C r ↓ 0 lim inf J r ( f ) . (15)
The macroscopic inequality also has a geometric consequence, which requires neither an energy form nor doubling of Ψ \Psi Ψ . For β > 0 \beta>0 β > 0 , we say that Ψ \Psi Ψ satisfies L S ( β ) \mathrm{LS}(\beta) LS ( β ) if there exists C L ≥ 1 C_L\ge1 C L ≥ 1 such that
Ψ ( r ) Ψ ( R ) ≤ C L ( r R ) β , 0 < r ≤ R < ∞ . (16) \frac{\Psi(r)}{\Psi(R)}\le C_L\left(\frac{r}{R}\right)^\beta,\qquad0<r\le R<\infty. \tag*{(16)} Ψ ( R ) Ψ ( r ) ≤ C L ( R r ) β , 0 < r ≤ R < ∞. (16)
When diam X < ∞ \operatorname{diam}X<\infty diam X < ∞ , the condition is required only for R ≤ diam X R\le\operatorname{diam}X R ≤ diam X . The proof of the following proposition is given in Subsection 3.7.
Proposition 2.4. Let p ∈ ( 1 , ∞ ) p\in(1,\infty) p ∈ ( 1 , ∞ ) , let ( X , d ) (X,d) ( X , d ) be a nontrivial complete metric space, and let m m m be a positive Radon measure of full support satisfying
0 < m ( B ( x , r ) ) < ∞ , x ∈ X , r > 0. 0<m(B(x,r))<\infty,\qquad x\in X,\ r>0. 0 < m ( B ( x , r )) < ∞ , x ∈ X , r > 0. Let Ψ : [ 0 , ∞ ) → [ 0 , ∞ ) \Psi:[0,\infty)\to[0,\infty) Ψ : [ 0 , ∞ ) → [ 0 , ∞ ) be an increasing homeomorphism. Assume (VD) and ( m P I ( Ψ ) ) (\mathrm{mPI}(\Psi)) ( mPI ( Ψ )) . Then X X X is connected and Ψ \Psi Ψ satisfies L S ( p ) \mathrm{LS}(p) LS ( p ) :
Ψ ( r ) Ψ ( R ) ≤ C ( r R ) p (17) \frac{\Psi(r)}{\Psi(R)}\le C\left(\frac{r}{R}\right)^p \tag*{(17)} Ψ ( R ) Ψ ( r ) ≤ C ( R r ) p (17) whenever 0 < r ≤ R < ∞ 0<r\le R<\infty 0 < r ≤ R < ∞ and, if diam X < ∞ \operatorname{diam}X<\infty diam X < ∞ , R ≤ diam X R\le\operatorname{diam}X R ≤ diam X . The constant depends only on p p p , C D C_D C D , C M C_M C M , and A M A_M A M .
Under the hypotheses of Theorem 2.2, the domain of the given energy form is identified by (64) , with the energy comparison (65) . This is the characterization of Shimizu [43 Corollaries 1.14 and 1.15] ; its applicability under the present hypotheses is discussed in Subsection 3.8.
Converse construction
We now assume only the metric-measure hypotheses and construct an energy from (mPI(Ψ \Psi Ψ )). The construction is subsequential and does not assert uniqueness of the resulting form.
Theorem 2.5. Assume the standing metric-measure hypotheses, (VD), (2.2), and (mPI(Ψ \Psi Ψ )). Fix admissible constants C M C_M C M and A M A_M A M in (mPI(Ψ \Psi Ψ )). Suppose in addition that
B Ψ ∩ C c ( X ) is dense in ( C c ( X ) , ∥ ⋅ ∥ ∞ ) . (18) \mathcal B_{\Psi} \cap C_c(X) \text{ is dense in } (C_c(X),\lVert\cdot\rVert_{\infty}). \tag*{(18)} B Ψ ∩ C c ( X ) is dense in ( C c ( X ) , ∥ ⋅ ∥ ∞ ) . (18) Then there exist a sequence r n ↓ 0 r_n \downarrow0 r n ↓ 0 , a functional E ~ K S \widetilde{\mathcal E}_{\mathrm{KS}} E KS on B Ψ \mathcal B_{\Psi} B Ψ , and a regular strongly local p p p -energy form ( E K S , F K S ) (\mathcal E_{\mathrm{KS}},\mathcal F_{\mathrm{KS}}) ( E KS , F KS ) such that
E ~ K S ( f ) = lim n → ∞ J r n ( f ) ≍ sup r > 0 J r ( f ) , f ∈ B Ψ , (19) \widetilde{\mathcal E}_{\mathrm{KS}}(f)=\lim_{n\to\infty}J_{r_n}(f)\asymp\sup_{r>0}J_r(f),\qquad f\in\mathcal B_{\Psi}, \tag*{(19)} E KS ( f ) = n → ∞ lim J r n ( f ) ≍ r > 0 sup J r ( f ) , f ∈ B Ψ , (19) and
F K S = B Ψ ∩ C c ( X ) ‾ ∥ ⋅ ∥ B Ψ . \mathcal F_{\mathrm{KS}}=\overline{\mathcal B_{\Psi}\cap C_c(X)}^{\lVert\cdot\rVert_{\mathcal B_{\Psi}}}. F KS = B Ψ ∩ C c ( X ) ∥ ⋅ ∥ B Ψ . Here E K S \mathcal E_{\mathrm{KS}} E KS is the restriction of E ~ K S \widetilde{\mathcal E}_{\mathrm{KS}} E KS to F K S \mathcal F_{\mathrm{KS}} F KS . The form satisfies the full p p p -Clarkson inequalities and the generalized p p p -contraction property. Its canonical p p p -energy measure is denoted by Γ K S ( u ) \Gamma^{\mathrm{KS}}(u) Γ KS ( u ) for u ∈ F K S u\in\mathcal F_{\mathrm{KS}} u ∈ F KS . All constants in this theorem depend only on p p p , C D C_D C D , C Ψ C_{\Psi} C Ψ , C M C_M C M , A M A_M A M .
Moreover, with A : = A M + 2 A:=A_M+2 A := A M + 2 ,
∫ B ( x , R ) ∣ u − u B ( x , R ) ∣ p d m ≤ C Ψ ( R ) Γ K S ( u ) ( B ( x , A R ) ) (20) \int_{B(x,R)}\lvert u-u_{B(x,R)}\rvert^p\,dm\le C\Psi(R)\Gamma^{\mathrm{KS}}(u)(B(x,AR)) \tag*{(20)} ∫ B ( x , R ) ∣ u − u B ( x , R ) ∣ p d m ≤ C Ψ ( R ) Γ KS ( u ) ( B ( x , A R )) (20) for every u ∈ F K S u\in\mathcal F_{\mathrm{KS}} u ∈ F KS , x ∈ X x\in X x ∈ X , and R > 0 R>0 R > 0 .
The existence of the form follows from Sasaya’s Korevaar–Schoen construction [41 Theorem 8.3] , based on [27 Theorems 6.3.8 and 6.4.2] . The new conclusion is the Poincaré inequality (20) , which we recover from (mPI(Ψ \Psi Ψ )) in Section 4. No capacity upper bound cap ( Ψ ) ≤ \operatorname{cap}(\Psi)\leq cap ( Ψ ) ≤ is asserted for the reconstructed form; see Remark 4.5.
Weak-type characterizations
For f ∈ L p ( X ; m ) f\in L^p(X;m) f ∈ L p ( X ; m ) choose a Borel representative and define, off the diagonal,
K Ψ , f ( x , y ) : = ∣ f ( x ) − f ( y ) ∣ p Ψ ( d ( x , y ) ) V ( x , d ( x , y ) ) , (21) K_{\Psi,f}(x,y):=\frac{\lvert f(x)-f(y)\rvert^p}{\Psi(d(x,y))V(x,d(x,y))}, \tag*{(21)} K Ψ , f ( x , y ) := Ψ ( d ( x , y )) V ( x , d ( x , y )) ∣ f ( x ) − f ( y ) ∣ p , (21)
and set K Ψ , f = 0 K_{\Psi,f}=0 K Ψ , f = 0 on the diagonal. Its m ⊗ m m\otimes m m ⊗ m -equivalence class and the quantities below are independent of the representative: a change on an m m m -null set N N N affects only ( N × X ) ∪ ( X × N ) (N\times X)\cup(X\times N) ( N × X ) ∪ ( X × N ) . Put
W Ψ ( f ) : = ∥ K Ψ , f ∥ L 1 , ∞ ( m ⊗ m ) = sup λ > 0 λ ( m ⊗ m ) ( { K Ψ , f > λ } ) , (22) \mathcal W_{\Psi}(f):=\lVert K_{\Psi,f}\rVert_{L^{1,\infty}(m\otimes m)}
=\sup_{\lambda>0}\lambda\,(m\otimes m)(\{K_{\Psi,f}>\lambda\}), \tag*{(22)} W Ψ ( f ) := ∥ K Ψ , f ∥ L 1 , ∞ ( m ⊗ m ) = λ > 0 sup λ ( m ⊗ m ) ({ K Ψ , f > λ }) , (22)
W Ψ , ∞ ( f ) : = lim sup λ → ∞ λ ( m ⊗ m ) ( { K Ψ , f > λ } ) . (23) \mathcal W_{\Psi,\infty}(f):=\limsup_{\lambda\to\infty}\lambda\,(m\otimes m)(\{K_{\Psi,f}>\lambda\}). \tag*{(23)} W Ψ , ∞ ( f ) := λ → ∞ lim sup λ ( m ⊗ m ) ({ K Ψ , f > λ }) . (23)
Under (VD), if R = d ( x , y ) > 0 R=d(x,y)>0 R = d ( x , y ) > 0 , then C D − 1 V ( x , R ) ≤ V ( y , R ) ≤ C D V ( x , R ) C_D^{-1}V(x,R)\leq V(y,R)\leq C_DV(x,R) C D − 1 V ( x , R ) ≤ V ( y , R ) ≤ C D V ( x , R ) . Thus replacing the first-coordinate volume in (21) by the second-coordinate volume changes the functionals by factors depending only on C D C_D C D .
Theorem 2.6. Suppose that Assumption 2.1, (VD), (PI(Ψ \Psi Ψ )), and (cap(Ψ \Psi Ψ )≤ _{\leq} ≤ ) hold. Then, for every f ∈ L p ( X ; m ) f \in L^{p}(X;m) f ∈ L p ( X ; m ) ,
f ∈ F ⟺ W Ψ , ∞ ( f ) < ∞ ⟺ W Ψ ( f ) < ∞ . (24) f \in\mathcal{F} \quad\Longleftrightarrow\quad\mathcal{W}_{\Psi,\infty}(f) < \infty\quad\Longleftrightarrow\quad\mathcal{W}_{\Psi}(f) < \infty. \tag*{(24)} f ∈ F ⟺ W Ψ , ∞ ( f ) < ∞ ⟺ W Ψ ( f ) < ∞. (24) Moreover, for f ∈ F f \in\mathcal{F} f ∈ F ,
E ( f ) ≍ W Ψ , ∞ ( f ) ≍ W Ψ ( f ) . (25) \mathcal{E}(f) \asymp\mathcal{W}_{\Psi,\infty}(f) \asymp\mathcal{W}_{\Psi}(f). \tag*{(25)} E ( f ) ≍ W Ψ , ∞ ( f ) ≍ W Ψ ( f ) . (25) The constants depend only on the structural data.
Remark 2.7. Under the hypotheses of Theorem 2.6, suppose that V ( x , r ) ≍ r Q V(x,r) \asymp r^{Q} V ( x , r ) ≍ r Q and Ψ ( r ) ≍ r α \Psi(r) \asymp r^{\alpha} Ψ ( r ) ≍ r α for 0 < r ≤ diam X 0 < r \leq\operatorname{diam} X 0 < r ≤ diam X , with the upper restriction omitted when diam X = ∞ \operatorname{diam} X = \infty diam X = ∞ . Then Corollary 3.12 forces α ≥ p \alpha\geq p α ≥ p . For X = R Q X = \mathbb{R}^{Q} X = R Q and α = p \alpha= p α = p , (2.24) is the Brezis–Van Schaftingen–Yung formula, up to the normalization of ball volume [8 , 7 ] . See also Theorem 1.3 for Besov–Sobolev characterizations, including the weak-type endpoint with Lorentz index ∞ \infty ∞ [16 ] .
Examples
We now describe several classes of spaces satisfying Assumption 2.1 together with (VD), (PI(Ψ \Psi Ψ )), and (cap(Ψ \Psi Ψ )≤ _{\leq} ≤ ). In each case, the scale function is understood up to multiplication by constants.
For the Riemannian, Carnot, tree, and Laakso energies below, linearity of the gradient operators and the classical scalar or Hilbert-valued L p L^{p} L p Clarkson inequalities give (E2). The chain and Leibniz rules give normal contractions and (E3), while the energy-measure formulas give (G1)–(G2). Locality, regularity, and strong locality are justified in the individual examples.
Riemannian and sub-Riemannian spaces. Let ( X , d , m ) (X,d,m) ( X , d , m ) be complete, essentially nonbranching and infinitesimally Hilbertian, and satisfy the measure contraction property MCP( 0 , N ) (0,N) ( 0 , N ) for some 1 < N < ∞ 1 < N < \infty 1 < N < ∞ . For 1 < p < ∞ 1 < p < \infty 1 < p < ∞ , set
F = W 1 , p ( X ) , E p ( f ) = ∫ X ∣ D f ∣ p d m , Γ ( f ) = ∣ D f ∣ p m , \mathcal{F} = W^{1,p}(X), \qquad\mathcal{E}_{p}(f) = \int_{X} |Df|^{p}\,dm, \qquad\Gamma(f) = |Df|^{p}m, F = W 1 , p ( X ) , E p ( f ) = ∫ X ∣ D f ∣ p d m , Γ ( f ) = ∣ D f ∣ p m ,
where ∣ D f ∣ |Df| ∣ D f ∣ denotes the minimal p p p -weak upper gradient. The MCP condition gives volume doubling and, under essential nonbranching, a scale-invariant weak (1, 1)-Poincaré inequality; see [20 Section 7] . Hence (PI(Ψ \Psi Ψ )) holds with Ψ ( r ) = r p \Psi(r) = r^{p} Ψ ( r ) = r p , and Lipschitz distance cutoffs give (cap(Ψ \Psi Ψ )≤ _{\leq} ≤ ). The weak gradient is independent of p p p , and the Hilbertian differential gives the Clarkson inequalities [19 Theorem 3.3 and Proposition 4.4] ; Sobolev locality and Lipschitz density give the remaining parts of Assumption 2.1 [22 ] . This class includes complete Riemannian manifolds with nonnegative Ricci curvature and Riemannian volume, and complete Sasakian manifolds with nonnegative Tanaka–Webster Ricci curvature, equipped with the Carnot–Carathéodory metric and Popp measure [35 Theorem 1.1] , including Heisenberg groups and standard CR spheres [36 ] . Sub-Riemannian manifolds are infinitesimally Hilbertian [34 ] , and ∣ D f ∣ |Df| ∣ D f ∣ is the horizontal gradient norm for smooth reference measures. The same standing assumptions also hold on complete doubling Riemannian manifolds supporting the scale-invariant L p L^{p} L p -Poincaré inequality, including compact Riemannian manifolds [40 ] , and on arbitrary Carnot groups with their Carnot–Carathéodory metric and Haar measure [25 ] .
Post-critically finite self-similar sets. Let L = ( K , S , { F i } i ∈ S ) \mathcal{L} = (K,S,\{F_{i}\}_{i\in S}) L = ( K , S , { F i } i ∈ S ) be a connected p.c.f. self-similar structure carrying a regular strongly local self-similar p p p -resistance form ( E p , F p ) (\mathcal{E}_{p},\mathcal{F}_{p}) ( E p , F p ) with weights ( ρ p , i ) i ∈ S ∈ ( 1 , ∞ ) S (\rho_{p,i})_{i\in S} \in(1,\infty)^{S} ( ρ p , i ) i ∈ S ∈ ( 1 , ∞ ) S . Set
R ^ p ( x , y ) = R E p ( x , y ) 1 / ( p − 1 ) , ∑ i ∈ S ρ p , i − d f , p / ( p − 1 ) = 1 , \widehat{R}_{p}(x,y) = R_{\mathcal{E}_{p}}(x,y)^{1/(p-1)}, \qquad\sum_{i\in S} \rho_{p,i}^{-d_{f,p}/(p-1)} = 1, R p ( x , y ) = R E p ( x , y ) 1/ ( p − 1 ) , i ∈ S ∑ ρ p , i − d f , p / ( p − 1 ) = 1 ,
where R E p R_{\mathcal{E}_p} R E p denotes the effective p p p -resistance, and let m m m be the self-similar probability measure with weights ρ p , i − d f , p / ( p − 1 ) \rho_{p,i}^{-d_{f,p}/(p-1)} ρ p , i − d f , p / ( p − 1 ) . Then m m m is d f , p d_{f,p} d f , p -Ahlfors regular with respect to R ^ p \widehat{R}_p R p , and the standing Poincaré and capacity assumptions hold with
Ψ ( r ) = r d f , p + p − 1 , \Psi(r)=r^{d_{f,p}+p-1}, Ψ ( r ) = r d f , p + p − 1 ,
as follows from [27 Assumption 6.6.1, Lemma 6.6.8, Corollary 6.6.11, and Propositions 6.6.9, 6.6.12, and 6.6.14] . The associated canonical energy measures satisfy (G1)–(G3), while the open-set locality property (G4) follows from [26 Lemma 5.19] . The generalized p p p -contraction property yields the remaining Clarkson, normal-contraction, and algebra properties by [26 Proposition 2.3] . For the cited self-similar constructions, strong locality follows from [26 Proposition 5.24] ; in the Cao–Gu–Qiu construction it is proved in [9 Theorem 5.2] . Such forms exist on strongly symmetric p.c.f. structures, including affine nested fractals and the standard Sierpiński gasket; see [26 Theorem 8.43, Proposition 8.49, and Theorems 8.50–8.51] and [9 Theorems 4.2, 5.1, 5.2, and 6.3] . This statement uses the p p p -resistance metric and the corresponding self-similar measure. It does not assert the same power scale for an arbitrary ambient Euclidean metric.
The planar Sierpiński carpet. For every p ∈ ( 1 , ∞ ) p \in(1,\infty) p ∈ ( 1 , ∞ ) , Murugan and Shimizu construct on the standard planar carpet, with its Euclidean metric and uniform self-similar measure, a regular strongly local self-similar p p p -energy and its energy measures, which satisfy (G1)–(G3); see [38 Theorems 1.1 and 1.2] , building on the construction of p p p -energies and energy measures on carpets in [42 ] . If ρ ( p ) \rho(p) ρ ( p ) is their renormalization constant, then the standing Poincaré and capacity estimates hold with
Ψ ( r ) = r d w ( p ) , d w ( p ) = log ( 8 ρ ( p ) ) log 3 ; \Psi(r)=r^{d_w(p)},\qquad d_w(p)=\frac{\log(8\rho(p))}{\log3}; Ψ ( r ) = r d w ( p ) , d w ( p ) = log 3 log ( 8 ρ ( p )) ;
see [38 Theorem 1.5] . The generalized p p p -contraction property, and hence the required Clarkson, normal-contraction, and algebra properties, is established in [26 Remark 8.20] ; the open-set locality property (G4) follows from [26 Lemma 5.19] .
Metric trees and cable systems. Let ( T , d ) (T,d) ( T , d ) be a complete real tree, let ν \nu ν be its length measure, and let m m m be a full-support Radon measure such that
m ( B ( x , r ) ) ≍ Φ ( r ) , x ∈ T , r > 0 , m(B(x,r))\asymp\Phi(r),\qquad x\in T,\ r>0, m ( B ( x , r )) ≍ Φ ( r ) , x ∈ T , r > 0 ,
where Φ \Phi Φ is a doubling homeomorphism satisfying, for some α ≥ 1 \alpha\ge1 α ≥ 1 ,
c R r ≤ Φ ( R ) Φ ( r ) ≤ C ( R r ) α , 0 < r ≤ R . c\frac{R}{r}\le\frac{\Phi(R)}{\Phi(r)}\le C\left(\frac{R}{r}\right)^\alpha,\qquad0<r\le R. c r R ≤ Φ ( r ) Φ ( R ) ≤ C ( r R ) α , 0 < r ≤ R .
Let F p \mathcal{F}_p F p be the space of functions in L p ( T ; m ) L^p(T;m) L p ( T ; m ) that are absolutely continuous along compact geodesic arcs and whose arc derivative ∂ f \partial f ∂ f belongs to L p ( T ; ν ) L^p(T;\nu) L p ( T ; ν ) , and consider the length-gradient energy
E p ( f ) = ∫ T ∣ ∂ f ∣ p d ν , Γ ( f ) = ∣ ∂ f ∣ p ν . \mathcal{E}_p(f)=\int_T|\partial f|^p\,d\nu,\qquad\Gamma(f)=|\partial f|^p\nu. E p ( f ) = ∫ T ∣ ∂ f ∣ p d ν , Γ ( f ) = ∣ ∂ f ∣ p ν .
Then the standing assumptions hold with
Ψ ( r ) = r p − 1 Φ ( r ) . \Psi(r)=r^{p-1}\Phi(r). Ψ ( r ) = r p − 1 Φ ( r ) .
Locality of arc derivatives gives (G3)–(G4); compactly supported arcwise absolutely continuous functions provide a regular core, and the energy is strongly local. The geodesic Morrey estimate gives PI(Ψ \Psi Ψ ), while tree-adapted cutoffs have energy ≲ r 1 − p ≍ V ( x , r ) / Ψ ( r ) \lesssim r^{1-p}\asymp V(x,r)/\Psi(r) ≲ r 1 − p ≍ V ( x , r ) /Ψ ( r ) ; see [5 Lemma 2.10, Definitions 2.11 and 2.13, Theorem 2.12, Propositions 2.14–2.15, Corollary 2.16, Lemma 3.3, Theorem 3.5, and Proposition 3.6] . The global volume setting is treated in [5 Section 5] .
Yang’s scale-irregular real trees give explicit unbounded instances with V ( x , r ) ≍ V b ( r ) V(x,r) \asymp V_b(r) V ( x , r ) ≍ V b ( r ) and Ψ ( r ) = r p − 1 V b ( r ) \Psi(r)=r^{p-1}V_b(r) Ψ ( r ) = r p − 1 V b ( r ) ; see [44 Section 4, the paragraph preceding Lemma 4.1, Lemma 4.2, and Propositions 4.3–4.4] . Finite connected metric graphs, regarded as cable systems with length measure and edgewise p p p -energy, give compact examples with Ψ ( r ) = r p \Psi(r)=r^p Ψ ( r ) = r p . The same edgewise calculus verifies Assumption 2.1. Discrete edge forms on simplicial trees are not strongly local and hence do not satisfy the standing assumptions; the cable realization is the appropriate model in this setting.
Laakso-type spaces with prescribed scale functions. Let Φ \Phi Φ and Ψ \Psi Ψ be doubling homeomorphisms satisfying
C − 1 ( R r ) p ≤ Ψ ( R ) Ψ ( r ) ≤ C ( R r ) p − 1 Φ ( R ) Φ ( r ) , 0 < r ≤ R . C^{-1}\left(\frac{R}{r}\right)^p \le\frac{\Psi(R)}{\Psi(r)} \le C\left(\frac{R}{r}\right)^{p-1}\frac{\Phi(R)}{\Phi(r)}, \qquad0<r\le R. C − 1 ( r R ) p ≤ Ψ ( r ) Ψ ( R ) ≤ C ( r R ) p − 1 Φ ( r ) Φ ( R ) , 0 < r ≤ R .
Yang [44, Theorem 2.3] constructs an unbounded geodesic Laakso-type metric measure space with V ( x , r ) ≍ Φ ( r ) V(x,r)\asymp\Phi(r) V ( x , r ) ≍ Φ ( r ) and a regular strongly local p p p -energy satisfying PI ( Ψ ) \operatorname{PI}(\Psi) PI ( Ψ ) and the cutoff Sobolev inequality at scale Ψ \Psi Ψ ; the latter implies cap ( Ψ ) ≲ 1 \operatorname{cap}(\Psi)\lesssim1 cap ( Ψ ) ≲ 1 . These constructions also allow scale functions other than powers. By [44 Lemma 6.1] , the form and its energy measure have the gradient representations E p ( f ) = ∫ ∣ ∇ L f ∣ p d λ \mathcal{E}_p(f)=\int|\nabla_{\!L}f|^p\,d\lambda E p ( f ) = ∫ ∣ ∇ L f ∣ p d λ and Γ ( f ) = ∣ ∇ L f ∣ p λ \Gamma(f)=|\nabla_{\!L}f|^p\lambda Γ ( f ) = ∣ ∇ L f ∣ p λ , with λ \lambda λ and ∇ L \nabla_{\!L} ∇ L as in that construction. The local-domain compatibility in [44 paragraph preceding Lemma 3.4] gives (G4). In the Ahlfors-regular specialization Φ ( r ) = r d h \Phi(r)=r^{d_h} Φ ( r ) = r d h and Ψ ( r ) = r β p \Psi(r)=r^{\beta_p} Ψ ( r ) = r β p , every exponent in the sharp range p ≤ β p ≤ d h + p − 1 p\le\beta_p\le d_h+p-1 p ≤ β p ≤ d h + p − 1 occurs [44 Theorem 2.3 and Corollary 2.6] . Self-similar Laakso-type energies are also constructed in [3 ] ; those constructions are distinct from the prescribed-scale spaces used here.
Macroscopic inequalities and their consequences
We first prove Theorem 2.2 by capacitary discrete approximation. We then prove the two-scale comparison of the Korevaar–Schoen functionals (Theorem 2.3) and the connectedness and lower scaling of Ψ \Psi Ψ (Proposition 2.4), before recalling the essentially known identification of the energy domain.
Packing and localization
We first recall the following lemma.
Lemma 3.1. Assume (VD). For every λ ≥ 1 \lambda\ge1 λ ≥ 1 there exists N D ( λ ) < ∞ N_D(\lambda)<\infty N D ( λ ) < ∞ , depending only on C D C_D C D and λ \lambda λ , with the following property. If { x i } i ∈ I \{x_i\}_{i\in I} { x i } i ∈ I is t t t -separated, i.e. d ( x i , x j ) ≥ t d(x_i,x_j)\ge t d ( x i , x j ) ≥ t for i ≠ j i\ne j i = j , then
# { i : x i ∈ B ( x , λ t ) } ≤ N D ( λ ) \#\{i:x_i\in B(x,\lambda t)\}\le N_D(\lambda) # { i : x i ∈ B ( x , λ t )} ≤ N D ( λ ) for every x ∈ X x\in X x ∈ X and $t>0.
Proof. The balls B ( x i , t / 2 ) B(x_i,t/2) B ( x i , t /2 ) are disjoint and lie in B ( x , ( λ + 1 / 2 ) t ) B(x,(\lambda+1/2)t) B ( x , ( λ + 1/2 ) t ) . Iterated doubling compares the measure of this ball with that of every B ( x i , t / 2 ) B(x_i,t/2) B ( x i , t /2 ) , uniformly in i i i ; summing the disjoint measures proves the claim. This is the standard packing argument; see, e.g., [21 , 22 ] .
For u ∈ F l o c u\in\mathcal{F}_{\mathrm{loc}} u ∈ F loc and a relatively compact open set U U U , define
Γ ( u ) ∣ U : = Γ ( u # ) ∣ U (26) \left.\Gamma(u)\right|_U:=\left.\Gamma(u^\#)\right|_U \tag*{(26)} Γ ( u ) ∣ U := Γ ( u # ) U (26)
where u # ∈ F u^\#\in\mathcal{F} u # ∈ F is any representative agreeing with u u u on U U U . Property (G4) makes this definition independent of the representative and yields a positive Radon measure on X X X . This localization is also used for locally finite sums of cutoffs in Yang [44, paragraph preceding Lemma 3.4].
We first record invariance under addition of constants.
Lemma 3.2. Let u ∈ F l o c u \in\mathcal{F}_{\mathrm{loc}} u ∈ F loc , let U ⊂ X U \subset X U ⊂ X be open, and let c ∈ R c \in\mathbb{R} c ∈ R . Then
Γ ( u + c ) ∣ U = Γ ( u ) ∣ U . \Gamma(u+c)|_{U}=\Gamma(u)|_{U}. Γ ( u + c ) ∣ U = Γ ( u ) ∣ U . Proof. Choose u # ∈ F u^{\#} \in\mathcal{F} u # ∈ F agreeing with u u u near U ‾ \overline{U} U . Regularity, together with normal contraction, provides χ ∈ F ∩ C c ( X ) \chi\in\mathcal{F} \cap C_{\mathrm{c}}(X) χ ∈ F ∩ C c ( X ) with χ = 1 \chi=1 χ = 1 near U ‾ \overline{U} U : choose h 0 ∈ C c ( X ) h_{0} \in C_{\mathrm{c}}(X) h 0 ∈ C c ( X ) with h 0 = 1 h_{0}=1 h 0 = 1 near U ‾ \overline{U} U and 0 ≤ h 0 ≤ 1 0 \le h_{0} \le1 0 ≤ h 0 ≤ 1 , then use regularity to find h ∈ F ∩ C c ( X ) h \in\mathcal{F} \cap C_{\mathrm{c}}(X) h ∈ F ∩ C c ( X ) with ∥ h − h 0 ∥ ∞ < 1 / 4 \lVert h-h_{0}\rVert_{\infty}<1/4 ∥ h − h 0 ∥ ∞ < 1/4 . If T T T is zero on ( − ∞ , 1 / 4 ] (-\infty,1/4] ( − ∞ , 1/4 ] , linear on [ 1 / 4 , 3 / 4 ] [1/4,3/4] [ 1/4 , 3/4 ] , and one on [ 3 / 4 , ∞ ) [3/4,\infty) [ 3/4 , ∞ ) , then T / 2 T/2 T /2 is a normal contraction. Hence χ : = T ∘ h ∈ F ∩ C c ( X ) \chi:=T\circ h \in\mathcal{F} \cap C_{\mathrm{c}}(X) χ := T ∘ h ∈ F ∩ C c ( X ) by (9) and homogeneity, and χ = 1 \chi=1 χ = 1 near U ‾ \overline{U} U . The function u # + c χ u^{\#}+c\chi u # + cχ represents u + c u+c u + c on U U U . For every Borel A ⊂ U A \subset U A ⊂ U , (G3) gives Γ ( c χ ) ( A ) = 0 \Gamma(c\chi)(A)=0 Γ ( cχ ) ( A ) = 0 , and the two triangle inequalities furnished by (G2) yield Γ ( u # + c χ ) ( A ) 1 / p = Γ ( u # ) ( A ) 1 / p \Gamma(u^{\#}+c\chi)(A)^{1/p}=\Gamma(u^{\#})(A)^{1/p} Γ ( u # + cχ ) ( A ) 1/ p = Γ ( u # ) ( A ) 1/ p . Hence the restrictions coincide.
Lemma 3.3. Under (PI(Ψ \Psi Ψ )), the same inequality holds for u ∈ F l o c u \in\mathcal{F}_{\mathrm{loc}} u ∈ F loc : for every ball B = B ( x , R ) B=B(x,R) B = B ( x , R ) ,
∫ B ∣ u − u B ∣ p d m ≤ C P Ψ ( R ) Γ ( u ) ( A P B ) . \int_{B}|u-u_{B}|^{p}\,dm \le C_{\mathrm{P}}\Psi(R)\Gamma(u)(A_{\mathrm{P}}B). ∫ B ∣ u − u B ∣ p d m ≤ C P Ψ ( R ) Γ ( u ) ( A P B ) . Proof. Because all balls are relatively compact, choose u # ∈ F u^{\#} \in\mathcal{F} u # ∈ F agreeing with u u u on a relatively compact neighborhood of A P B A_{\mathrm{P}}B A P B . Apply (PI(Ψ \Psi Ψ )) to $u^{#} and use the definition (26) .
A controlled partition of unity
The next lemma supplies the capacitary partition used in the proof. Controlled partitions of this kind were used by Murugan [37 proof of Lemma 2.5] , Baudoin [4 proof of Theorem 5.2] , and Shimizu [43 Lemma 3.10] ; see also Yang [44 Lemma 3.2] and Eriksson-Bique [17 Lemma 4.2] . For Lipschitz partitions of unity subordinate to separated nets in doubling spaces, see Hytönen–Korte [23 Lemma 3.1] . We give the construction under the present assumptions, using the normal-contraction property (9) and the algebra estimate (8) .
Note that since X X X is separable, for every t > 0 t>0 t > 0 there is a countable maximal t t t -separated set { x i } i ∈ I \{x_{i}\}_{i\in I} { x i } i ∈ I . By maximality, X = ⋃ i ∈ I B ( x i , t ) X=\bigcup_{i\in I}B(x_{i},t) X = ⋃ i ∈ I B ( x i , t ) , and by Lemma 3.1 every fixed enlargement { B ( x i , λ t ) } i \{B(x_{i},\lambda t)\}_{i} { B ( x i , λ t ) } i has uniformly bounded multiplicity. In particular, such families are locally finite.
Lemma 3.4. Assume (VD) and (cap(Ψ \Psi Ψ )≤ _{\le} ≤ ). Set
L : = 2 A c . L:=2A_{\mathrm{c}}. L := 2 A c . There is C c u t < ∞ C_{\mathrm{cut}}<\infty C cut < ∞ , depending only on p p p , C D C_{\mathrm{D}} C D , C Ψ C_{\Psi} C Ψ , C c C_{\mathrm{c}} C c , A c A_{\mathrm{c}} A c , such that the following holds. For every t > 0 t>0 t > 0 satisfying
L t < 1 2 diam X if diam X < ∞ , Lt<\frac{1}{2}\operatorname{diam}X \qquad\text{if }\operatorname{diam}X<\infty, L t < 2 1 diam X if diam X < ∞ , and every maximal t t t -separated set { x i } i ∈ I \{x_{i}\}_{i\in I} { x i } i ∈ I , there are functions ψ i ∈ F ∩ L ∞ ( X ; m ) \psi_{i}\in\mathcal{F}\cap L^{\infty}(X;m) ψ i ∈ F ∩ L ∞ ( X ; m ) such that
0 ≤ ψ i ≤ 1 , (27) 0\le\psi_{i}\le1, \tag*{(27)} 0 ≤ ψ i ≤ 1 , (27) supp m ψ i ⊂ B ( x i , L t ) , (28) \operatorname{supp}_{m}\psi_{i}\subset B(x_{i},Lt), \tag*{(28)} supp m ψ i ⊂ B ( x i , L t ) , (28) ∑ i ∈ I ψ i = 1 m -a.e. on X , (29) \sum_{i\in I}\psi_{i}=1 \quad m\text{-a.e. on }X, \tag*{(29)} i ∈ I ∑ ψ i = 1 m -a.e. on X , (29) E ( ψ i ) ≤ C c u t V ( x i , t ) Ψ ( t ) . (30) \mathcal{E}(\psi_{i})\le C_{\mathrm{cut}}\frac{V(x_{i},t)}{\Psi(t)}. \tag*{(30)} E ( ψ i ) ≤ C cut Ψ ( t ) V ( x i , t ) . (30) The sum in (29) is locally finite, and every family { B ( x i , λ t ) } i \{B(x_{i},\lambda t)\}_{i} { B ( x i , λ t ) } i , for fixed λ \lambda λ , has uniformly bounded multiplicity.
Proof. The scale restriction gives 2 t < diam X 2t < \operatorname{diam} X 2 t < diam X , so ( cap ( Ψ ) ≤ ) (\operatorname{cap}(\Psi)\leq) ( cap ( Ψ ) ≤ ) is available at radius 2 t 2t 2 t ; moreover, B ( x i , L t ) B(x_i,Lt) B ( x i , L t ) is a proper subset of X X X . This is immediate when diam X = ∞ \operatorname{diam} X = \infty diam X = ∞ ; when diam X < ∞ \operatorname{diam} X < \infty diam X < ∞ , it follows from sup y ∈ X d ( x i , y ) ≥ 1 2 diam X > L t \sup_{y\in X} d(x_i,y) \geq\frac{1}{2}\operatorname{diam} X > Lt sup y ∈ X d ( x i , y ) ≥ 2 1 diam X > L t . Choose an admissible near-minimizer at radius 2 t 2t 2 t and truncate it by normal contractions. Doubling then gives η i ∈ F \eta_i \in\mathcal{F} η i ∈ F , 0 ≤ η i ≤ 1 0 \leq\eta_i \leq1 0 ≤ η i ≤ 1 , such that
η i = 1 m -a.e. on B ( x i , 2 t ) , supp m η i ⊂ B ( x i , L t ) , E ( η i ) ≲ V ( x i , t ) Ψ ( t ) . \eta_i = 1 \quad m\text{-a.e. on }B(x_i,2t), \qquad\operatorname{supp}_{m}\eta_i \subset B(x_i,Lt), \qquad\mathcal{E}(\eta_i) \lesssim\frac{V(x_i,t)}{\Psi(t)}. η i = 1 m -a.e. on B ( x i , 2 t ) , supp m η i ⊂ B ( x i , L t ) , E ( η i ) ≲ Ψ ( t ) V ( x i , t ) .
The net is countable. After changing representatives on one common null set, we may assume that every η i \eta_i η i vanishes off supp m η i \operatorname{supp}_{m}\eta_i supp m η i and has the stated value on its inner ball. Put J i = { j : d ( x i , x j ) < 2 L t } J_i=\{j:d(x_i,x_j)<2Lt\} J i = { j : d ( x i , x j ) < 2 L t } and S i = ∑ j ∈ J i η j S_i=\sum_{j\in J_i}\eta_j S i = ∑ j ∈ J i η j . Lemma 3.1, nearby-ball volume comparison, and the triangle inequality for E 1 / p \mathcal{E}^{1/p} E 1/ p give
# J i ≲ 1 , E ( S i ) 1 / p ≲ ( V ( x i , t ) Ψ ( t ) ) 1 / p . \#J_i \lesssim1, \qquad\mathcal{E}(S_i)^{1/p}\lesssim\left(\frac{V(x_i,t)}{\Psi(t)}\right)^{1/p}. # J i ≲ 1 , E ( S i ) 1/ p ≲ ( Ψ ( t ) V ( x i , t ) ) 1/ p .
Define the 1-Lipschitz normal contraction H : R → [ 0 , 1 ] H:\mathbb{R}\to[0,1] H : R → [ 0 , 1 ] by
H ( s ) = { 0 , s ≤ 0 , s , 0 ≤ s ≤ 1 , s − 1 , s ≥ 1. H(s)=
\begin{cases}
0, & s\leq0,\\
s, & 0\leq s\leq1,\\
s^{-1}, & s\geq1.
\end{cases} H ( s ) = ⎩ ⎨ ⎧ 0 , s , s − 1 , s ≤ 0 , 0 ≤ s ≤ 1 , s ≥ 1.
Set ψ i = η i H ( S i ) \psi_i=\eta_iH(S_i) ψ i = η i H ( S i ) . Normal contraction and (2.7) give
E ( ψ i ) 1 / p ≲ E ( S i ) 1 / p + E ( η i ) 1 / p ≲ ( V ( x i , t ) Ψ ( t ) ) 1 / p . \mathcal{E}(\psi_i)^{1/p}\lesssim\mathcal{E}(S_i)^{1/p}+\mathcal{E}(\eta_i)^{1/p}\lesssim\left(\frac{V(x_i,t)}{\Psi(t)}\right)^{1/p}. E ( ψ i ) 1/ p ≲ E ( S i ) 1/ p + E ( η i ) 1/ p ≲ ( Ψ ( t ) V ( x i , t ) ) 1/ p .
The family { η i } \{\eta_i\} { η i } is locally finite. Hence
S ( x ) : = ∑ i η i ( x ) S(x):=\sum_i\eta_i(x) S ( x ) := i ∑ η i ( x )
is defined off the common null set and satisfies S ≥ 1 S\geq1 S ≥ 1 , because the B ( x i , t ) B(x_i,t) B ( x i , t ) cover X X X . On { η i ≠ 0 } \{\eta_i\ne0\} { η i = 0 } every nonzero η j \eta_j η j has j ∈ J i j\in J_i j ∈ J i ; therefore S i = S S_i=S S i = S and ψ i = η i / S \psi_i=\eta_i/S ψ i = η i / S . Summation proves (3.4); the other assertions follow from the support bounds and Lemma 3.1. □
Discrete approximation at a fixed scale
Fix f ∈ L l o c p ( X ; m ) f\in L^p_{\mathrm{loc}}(X;m) f ∈ L loc p ( X ; m ) and 0 < r < s 0<r<s 0 < r < s . If diam X < ∞ \operatorname{diam} X<\infty diam X < ∞ and r ≥ diam X r\geq\operatorname{diam} X r ≥ diam X , then every ball B ( y , r ) B(y,r) B ( y , r ) and B ( x , s ) B(x,s) B ( x , s ) is X X X , so
M f ( x , s ) = Ψ ( r ) Ψ ( s ) M f ( y , r ) ≤ M f ( y , r ) M_f(x,s)=\frac{\Psi(r)}{\Psi(s)}M_f(y,r)\leq M_f(y,r) M f ( x , s ) = Ψ ( s ) Ψ ( r ) M f ( y , r ) ≤ M f ( y , r )
for all x , y x,y x , y . We therefore construct the approximation below whenever r ≤ diam X r\leq\operatorname{diam} X r ≤ diam X .
With L = 2 A c L=2A_c L = 2 A c as in Lemma 3.4, let
θ : = 1 16 ( L + 2 ) , t : = θ r . (31) \theta:=\frac{1}{16(L+2)},\qquad t:=\theta r. \tag*{(31)} θ := 16 ( L + 2 ) 1 , t := θ r . (31)
If diam X < ∞ \operatorname{diam} X<\infty diam X < ∞ , then
2 A c t = L t = L θ r < r 16 ≤ 1 16 diam X < 1 2 diam X . 2A_ct=Lt=L\theta r<\frac{r}{16}\leq\frac{1}{16}\operatorname{diam} X<\frac{1}{2}\operatorname{diam} X. 2 A c t = L t = L θ r < 16 r ≤ 16 1 diam X < 2 1 diam X .
Hence the scale hypothesis of Lemma 3.4 is satisfied. Choose a maximal t t t -separated set { x i } \{x_i\} { x i } and let { ψ i } \{\psi_i\} { ψ i } be the partition furnished by Lemma 3.4. Put
B i : = B ( x i , t ) , D i : = B ( x i , r / 4 ) , a i : = f B i . (32) B_i:=B(x_i,t),\qquad D_i:=B(x_i,r/4),\qquad a_i:=f_{B_i}. \tag*{(32)} B i := B ( x i , t ) , D i := B ( x i , r /4 ) , a i := f B i . (32)
Define the locally finite discrete approximation
P r f : = ∑ i a i ψ i . (33) P_r f := \sum_i a_i\psi_i. \tag*{(33)} P r f := i ∑ a i ψ i . (33)
Because the sum is locally finite and every ψ i ∈ F \psi_i \in\mathcal{F} ψ i ∈ F , P r f ∈ F l o c P_r f \in\mathcal{F}_{\mathrm{loc}} P r f ∈ F loc .
This is the capacitary counterpart of formula (27) in Hytönen–Korte [23 ] . Related discrete approximations appear in Shimizu [43, Lemma 3.11] and Yang [44, Section 3]; compare also the discrete convolutions in Eriksson-Bique [17, proof of Lemma 4.5]. The coefficient estimate below parallels the one in the proof of Hytönen–Korte [23, Lemma 3.3], with localized energy measures in place of pointwise Lipschitz constants.
The choice of θ \theta θ gives
1 4 + ( L + 2 ) θ = 1 4 + 1 16 = 5 16 < 1. (34) \frac{1}{4}+(L+2)\theta=\frac{1}{4}+\frac{1}{16}=\frac{5}{16}<1. \tag*{(34)} 4 1 + ( L + 2 ) θ = 4 1 + 16 1 = 16 5 < 1. (34)
In this subsection, constants depend only on p p p , C D C_D C D , C Ψ C_\Psi C Ψ , C c C_c C c , and A c A_c A c , unless an additional dependence is specified. The constants in Lemma 3.7 may also depend on A P A_P A P .
Lemma 3.5. Let
K i : = { j : B i ∩ supp m ψ j ≠ ∅ } . K_i:=\{j:B_i\cap\operatorname{supp}_m\psi_j\ne\varnothing\}. K i := { j : B i ∩ supp m ψ j = ∅ } . Then # K i ≤ N 1 \#K_i\le N_1 # K i ≤ N 1 uniformly in i i i , and for every j ∈ K i j\in K_i j ∈ K i ,
∣ a i − a j ∣ p E ( ψ j ) ≤ C ∫ D i M f ( y , r ) d m ( y ) , (35) |a_i-a_j|^p\mathcal{E}(\psi_j)\le C\int_{D_i}M_f(y,r)\,dm(y), \tag*{(35)} ∣ a i − a j ∣ p E ( ψ j ) ≤ C ∫ D i M f ( y , r ) d m ( y ) , (35) where N 1 N_1 N 1 and C C C depend only on p p p , C D C_D C D , C Ψ C_\Psi C Ψ , C c C_c C c , and A c A_c A c .
Proof. If j ∈ K i j\in K_i j ∈ K i , choose z ∈ B i ∩ supp m ψ j z\in B_i\cap\operatorname{supp}_m\psi_j z ∈ B i ∩ supp m ψ j . By (28) ,
d ( x i , x j ) ≤ d ( x i , z ) + d ( z , x j ) < ( L + 1 ) t . d(x_i,x_j)\le d(x_i,z)+d(z,x_j)<(L+1)t. d ( x i , x j ) ≤ d ( x i , z ) + d ( z , x j ) < ( L + 1 ) t .
Lemma 3.1 therefore gives a uniform bound on # K i \#K_i # K i .
Fix j ∈ K i j\in K_i j ∈ K i and y ∈ D i y\in D_i y ∈ D i , and set Q y = B ( y , r ) Q_y=B(y,r) Q y = B ( y , r ) . If z ∈ B j z\in B_j z ∈ B j , then by (34)
d ( z , y ) ≤ d ( z , x j ) + d ( x j , x i ) + d ( x i , y ) < t + ( L + 1 ) t + r 4 = ( L + 2 ) t + r 4 < r . d(z,y)\le d(z,x_j)+d(x_j,x_i)+d(x_i,y)<t+(L+1)t+\frac{r}{4}=(L+2)t+\frac{r}{4}<r. d ( z , y ) ≤ d ( z , x j ) + d ( x j , x i ) + d ( x i , y ) < t + ( L + 1 ) t + 4 r = ( L + 2 ) t + 4 r < r .
Thus B j ⊂ Q y B_j\subset Q_y B j ⊂ Q y . The same argument, with j = i j=i j = i , gives B i ⊂ Q y B_i\subset Q_y B i ⊂ Q y .
By Jensen’s inequality,
∣ a i − f Q y ∣ p ≤ 1 V ( x i , t ) ∫ B i ∣ f − f Q y ∣ p d m ≤ V ( y , r ) V ( x i , t ) Ψ ( r ) M f ( y , r ) . |a_i-f_{Q_y}|^p\le\frac{1}{V(x_i,t)}\int_{B_i}|f-f_{Q_y}|^p\,dm
\le\frac{V(y,r)}{V(x_i,t)}\Psi(r)M_f(y,r). ∣ a i − f Q y ∣ p ≤ V ( x i , t ) 1 ∫ B i ∣ f − f Q y ∣ p d m ≤ V ( x i , t ) V ( y , r ) Ψ ( r ) M f ( y , r ) .
Since y ∈ B ( x i , r / 4 ) y\in B(x_i,r/4) y ∈ B ( x i , r /4 ) and r / t = θ − 1 r/t=\theta^{-1} r / t = θ − 1 is a fixed structural constant, volume doubling gives
V ( y , r ) V ( x i , t ) ≤ C . (36) \frac{V(y,r)}{V(x_i,t)}\le C. \tag*{(36)} V ( x i , t ) V ( y , r ) ≤ C . (36)
Similarly, ∣ a j − f Q y ∣ p ≤ C Ψ ( r ) M f ( y , r ) |a_j-f_{Q_y}|^p\le C\Psi(r)M_f(y,r) ∣ a j − f Q y ∣ p ≤ C Ψ ( r ) M f ( y , r ) . Therefore
∣ a i − a j ∣ p ≤ C Ψ ( r ) M f ( y , r ) , y ∈ D i . (37) |a_i-a_j|^p\le C\Psi(r)M_f(y,r),\qquad y\in D_i. \tag*{(37)} ∣ a i − a j ∣ p ≤ C Ψ ( r ) M f ( y , r ) , y ∈ D i . (37)
Averaging (37) over D i D_i D i gives
∣ a i − a j ∣ p ≤ C Ψ ( r ) V ( x i , r / 4 ) ∫ D i M f ( y , r ) d m ( y ) . (38) |a_i-a_j|^p\le C\frac{\Psi(r)}{V(x_i,r/4)}\int_{D_i}M_f(y,r)\,dm(y). \tag*{(38)} ∣ a i − a j ∣ p ≤ C V ( x i , r /4 ) Ψ ( r ) ∫ D i M f ( y , r ) d m ( y ) . (38)
By (30) ,
E ( ψ j ) ≤ C c u t V ( x j , t ) Ψ ( t ) . \mathcal{E}(\psi_j) \le C_{\mathrm{cut}} \frac{V(x_j,t)}{\Psi(t)}. E ( ψ j ) ≤ C cut Ψ ( t ) V ( x j , t ) .
Since j ∈ K i j \in K_i j ∈ K i ,
B ( x j , t ) ⊂ B ( x i , ( L + 2 ) t ) ⊂ B ( x i , r / 4 ) = D i , B(x_j,t) \subset B(x_i,(L+2)t) \subset B(x_i,r/4)=D_i, B ( x j , t ) ⊂ B ( x i , ( L + 2 ) t ) ⊂ B ( x i , r /4 ) = D i ,
where the second inclusion follows from ( L + 2 ) t = r / 16 < r / 4 (L+2)t=r/16<r/4 ( L + 2 ) t = r /16 < r /4 . Hence V ( x j , t ) ≤ V ( x i , r / 4 ) V(x_j,t)\le V(x_i,r/4) V ( x j , t ) ≤ V ( x i , r /4 ) . Finally, because r = θ − 1 t r=\theta^{-1}t r = θ − 1 t and θ \theta θ is fixed, the doubling property of Ψ \Psi Ψ gives
Ψ ( r ) ≤ C θ Ψ ( t ) (39) \Psi(r)\le C_\theta\Psi(t) \tag*{(39)} Ψ ( r ) ≤ C θ Ψ ( t ) (39)
where C θ C_\theta C θ depends only on C Ψ C_\Psi C Ψ and A c A_c A c . Multiplying (38) by the energy bound for ψ j \psi_j ψ j and using these two observations proves (35) .
Lemma 3.6. For every ball B = B ( x , s ) B=B(x,s) B = B ( x , s ) with s > r s>r s > r ,
∫ B ∣ f − P r f ∣ p d m ≤ C Ψ ( r ) ∫ B ( x , 2 s ) M f ( y , r ) d m ( y ) . (40) \int_B |f-P_r f|^p\,dm \le C\Psi(r)\int_{B(x,2s)} M_f(y,r)\,dm(y). \tag*{(40)} ∫ B ∣ f − P r f ∣ p d m ≤ C Ψ ( r ) ∫ B ( x , 2 s ) M f ( y , r ) d m ( y ) . (40) Proof. By (29) , non-negativity, and convexity,
∣ f − P r f ∣ p = ∣ ∑ i ψ i ( f − a i ) ∣ p ≤ ∑ i ψ i ∣ f − a i ∣ p . (41) |f-P_r f|^p=\left|\sum_i\psi_i(f-a_i)\right|^p\le\sum_i\psi_i|f-a_i|^p. \tag*{(41)} ∣ f − P r f ∣ p = i ∑ ψ i ( f − a i ) p ≤ i ∑ ψ i ∣ f − a i ∣ p . (41) Fix i i i and y ∈ D i y\in D_i y ∈ D i . By (28) and (34) ,
supp m ψ i ∪ B i ⊂ B ( y , r ) = Q y . \operatorname{supp}_m\psi_i\cup B_i\subset B(y,r)=Q_y. supp m ψ i ∪ B i ⊂ B ( y , r ) = Q y . Therefore
∫ supp m ψ i ∣ f − a i ∣ p d m ≤ 2 p − 1 ∫ Q y ∣ f − f Q y ∣ p d m + 2 p − 1 m ( supp m ψ i ) ∣ a i − f Q y ∣ p . (42) \int_{\operatorname{supp}_m\psi_i}|f-a_i|^p\,dm\le2^{p-1}\int_{Q_y}|f-f_{Q_y}|^p\,dm+2^{p-1}m(\operatorname{supp}_m\psi_i)|a_i-f_{Q_y}|^p. \tag*{(42)} ∫ supp m ψ i ∣ f − a i ∣ p d m ≤ 2 p − 1 ∫ Q y ∣ f − f Q y ∣ p d m + 2 p − 1 m ( supp m ψ i ) ∣ a i − f Q y ∣ p . (42) The first term on the right is 2 p − 1 Ψ ( r ) V ( y , r ) M f ( y , r ) 2^{p-1}\Psi(r)V(y,r)M_f(y,r) 2 p − 1 Ψ ( r ) V ( y , r ) M f ( y , r ) . For the second term, Jensen’s inequality and (36) yield
∣ a i − f Q y ∣ p ≤ C Ψ ( r ) M f ( y , r ) . |a_i-f_{Q_y}|^p\le C\Psi(r)M_f(y,r). ∣ a i − f Q y ∣ p ≤ C Ψ ( r ) M f ( y , r ) . Furthermore,
m ( supp m ψ i ) ≤ V ( x i , L t ) ≤ C D , L V ( x i , t ) ≤ C D , L ′ V ( y , r ) , m(\operatorname{supp}_m\psi_i)\le V(x_i,Lt)\le C_{D,L}V(x_i,t)\le C'_{D,L}V(y,r), m ( supp m ψ i ) ≤ V ( x i , L t ) ≤ C D , L V ( x i , t ) ≤ C D , L ′ V ( y , r ) , so (42) gives
∫ supp m ψ i ∣ f − a i ∣ p d m ≤ C Ψ ( r ) V ( y , r ) M f ( y , r ) . (43) \int_{\operatorname{supp}_m\psi_i}|f-a_i|^p\,dm\le C\Psi(r)V(y,r)M_f(y,r). \tag*{(43)} ∫ supp m ψ i ∣ f − a i ∣ p d m ≤ C Ψ ( r ) V ( y , r ) M f ( y , r ) . (43) Since y ∈ D i y\in D_i y ∈ D i , volume doubling gives V ( y , r ) ≤ C V ( x i , r / 4 ) = C m ( D i ) V(y,r)\le CV(x_i,r/4)=Cm(D_i) V ( y , r ) ≤ C V ( x i , r /4 ) = C m ( D i ) . Averaging (43) over y ∈ D i y\in D_i y ∈ D i therefore yields
∫ supp m ψ i ∣ f − a i ∣ p d m ≤ C Ψ ( r ) ∫ D i M f ( y , r ) d m ( y ) . (44) \int_{\operatorname{supp}_m\psi_i}|f-a_i|^p\,dm\le C\Psi(r)\int_{D_i}M_f(y,r)\,dm(y). \tag*{(44)} ∫ supp m ψ i ∣ f − a i ∣ p d m ≤ C Ψ ( r ) ∫ D i M f ( y , r ) d m ( y ) . (44) Integrating (41) over B B B and using (44) ,
∫ B ∣ f − P r f ∣ p d m ≤ C Ψ ( r ) ∑ i : supp m ψ i ∩ B ≠ ∅ ∫ D i M f ( y , r ) d m ( y ) . (45) \int_B|f-P_r f|^p\,dm\le C\Psi(r)\sum_{i:\operatorname{supp}_m\psi_i\cap B\ne\varnothing}\int_{D_i}M_f(y,r)\,dm(y). \tag*{(45)} ∫ B ∣ f − P r f ∣ p d m ≤ C Ψ ( r ) i : supp m ψ i ∩ B = ∅ ∑ ∫ D i M f ( y , r ) d m ( y ) . (45) The family { D i } \{D_i\} { D i } has uniformly bounded multiplicity by Lemma 3.1, because r / 4 = ( 4 θ ) − 1 t r/4=(4\theta)^{-1}t r /4 = ( 4 θ ) − 1 t is a fixed multiple of t t t . If supp m ψ i ∩ B ( x , s ) ≠ ∅ \operatorname{supp}_m\psi_i\cap B(x,s)\ne\varnothing supp m ψ i ∩ B ( x , s ) = ∅ , then d ( x i , x ) < s + L t d(x_i,x)<s+Lt d ( x i , x ) < s + L t ; hence, for y ∈ D i y\in D_i y ∈ D i ,
d ( y , x ) < s + L t + r / 4 ≤ ( 1 + L θ + 1 4 ) s < 2 s . d(y,x)<s+Lt+r/4\le\left(1+L\theta+\frac14\right)s<2s. d ( y , x ) < s + L t + r /4 ≤ ( 1 + L θ + 4 1 ) s < 2 s . Thus all the relevant D i D_i D i are contained in B ( x , 2 s ) B(x,2s) B ( x , 2 s ) . Using the bounded multiplicity in (45) proves (40) .
Lemma 3.7. For every ball B = B ( x , s ) B = B(x,s) B = B ( x , s ) with s > r s > r s > r ,
Γ ( P r f ) ( A P B ) ≤ C ∫ B ( x , ( A P + 1 ) s ) M f ( y , r ) d m ( y ) . (46) \Gamma(P_{r}f)(A_{P}B) \le C \int_{B(x,(A_{P}+1)s)} M_{f}(y,r)\,dm(y). \tag*{(46)} Γ ( P r f ) ( A P B ) ≤ C ∫ B ( x , ( A P + 1 ) s ) M f ( y , r ) d m ( y ) . (46) Proof. The balls B i B_i B i cover X X X . Let
I B : = { i : B i ∩ A P B ≠ ∅ } . I_{B} := \{i : B_i \cap A_{P}B \ne\varnothing\}. I B := { i : B i ∩ A P B = ∅ } . By local finiteness and relative compactness, I B I_{B} I B is finite, and positivity of Γ ( P r f ) \Gamma(P_{r}f) Γ ( P r f ) gives
Γ ( P r f ) ( A P B ) ≤ ∑ i ∈ I B Γ ( P r f ) ( B i ) . (47) \Gamma(P_{r}f)(A_{P}B) \le\sum_{i\in I_{B}} \Gamma(P_{r}f)(B_i). \tag*{(47)} Γ ( P r f ) ( A P B ) ≤ i ∈ I B ∑ Γ ( P r f ) ( B i ) . (47) Fix i ∈ I B i \in I_{B} i ∈ I B . Recall K i = { j : B i ∩ supp m ψ j ≠ ∅ } K_i = \{j : B_i \cap\operatorname{supp}_{m}\psi_j \ne\varnothing\} K i = { j : B i ∩ supp m ψ j = ∅ } . On B i B_i B i , local finiteness and the partition identity give
P r f − a i = ∑ j ∈ K i ( a j − a i ) ψ j . (48) P_{r}f-a_i = \sum_{j\in K_i}(a_j-a_i)\psi_j. \tag*{(48)} P r f − a i = j ∈ K i ∑ ( a j − a i ) ψ j . (48) The right-hand side is a finite sum belonging to F \mathcal{F} F . By the definition of the localized energy measure and Lemma 3.2,
Γ ( P r f ) ( B i ) = Γ ( P r f − a i ) ( B i ) = Γ ( ∑ j ∈ K i ( a j − a i ) ψ j ) ( B i ) . (49) \Gamma(P_{r}f)(B_i) = \Gamma(P_{r}f-a_i)(B_i) = \Gamma\left(\sum_{j\in K_i}(a_j-a_i)\psi_j\right)(B_i). \tag*{(49)} Γ ( P r f ) ( B i ) = Γ ( P r f − a i ) ( B i ) = Γ j ∈ K i ∑ ( a j − a i ) ψ j ( B i ) . (49) Using the seminorm property (G2), followed by ( ∑ j = 1 N b j ) p ≤ N p − 1 ∑ j = 1 N b j p (\sum_{j=1}^{N} b_j)^p \le N^{p-1}\sum_{j=1}^{N} b_j^p ( ∑ j = 1 N b j ) p ≤ N p − 1 ∑ j = 1 N b j p , we obtain
Γ ( P r f ) ( B i ) ≤ ( # K i ) p − 1 ∑ j ∈ K i ∣ a j − a i ∣ p Γ ( ψ j ) ( B i ) ≤ N 1 p − 1 ∑ j ∈ K i ∣ a j − a i ∣ p E ( ψ j ) . (50) \begin{aligned}
\Gamma(P_{r}f)(B_i) &\le(\#K_i)^{p-1}\sum_{j\in K_i}|a_j-a_i|^p\Gamma(\psi_j)(B_i)\\
&\le N_1^{p-1}\sum_{j\in K_i}|a_j-a_i|^p\mathcal{E}(\psi_j). \tag*{(50)}
\end{aligned} Γ ( P r f ) ( B i ) ≤ ( # K i ) p − 1 j ∈ K i ∑ ∣ a j − a i ∣ p Γ ( ψ j ) ( B i ) ≤ N 1 p − 1 j ∈ K i ∑ ∣ a j − a i ∣ p E ( ψ j ) . (50) Lemma 3.5 and # K i ≤ N 1 \#K_i \le N_1 # K i ≤ N 1 give
Γ ( P r f ) ( B i ) ≤ C ∫ D i M f ( y , r ) d m ( y ) . (51) \Gamma(P_{r}f)(B_i) \le C\int_{D_i} M_f(y,r)\,dm(y). \tag*{(51)} Γ ( P r f ) ( B i ) ≤ C ∫ D i M f ( y , r ) d m ( y ) . (51) The family { D i } \{D_i\} { D i } has uniformly bounded multiplicity. Moreover, if i ∈ I B i \in I_B i ∈ I B , choose z ∈ B i ∩ B ( x , A P s ) z \in B_i \cap B(x,A_Ps) z ∈ B i ∩ B ( x , A P s ) . Then d ( x i , x ) < A P s + t d(x_i,x) < A_Ps+t d ( x i , x ) < A P s + t , and for y ∈ D i y \in D_i y ∈ D i ,
d ( y , x ) < A P s + t + r / 4 ≤ ( A P + θ + 1 / 4 ) s < ( A P + 1 ) s . d(y,x) < A_Ps+t+r/4 \le(A_P+\theta+1/4)s < (A_P+1)s. d ( y , x ) < A P s + t + r /4 ≤ ( A P + θ + 1/4 ) s < ( A P + 1 ) s . Thus
⋃ i ∈ I B D i ⊂ B ( x , ( A P + 1 ) s ) . \bigcup_{i\in I_B}D_i \subset B(x,(A_P+1)s). i ∈ I B ⋃ D i ⊂ B ( x , ( A P + 1 ) s ) . Summing (3.26) in (3.22) and using the bounded multiplicity proves (3.21).
Proof of the macroscopic inequality
The proof follows the decomposition into approximation error and Poincaré oscillation used by Hytönen–Korte [23 proof of Theorem 3.4] . Lemmas 3.6 and 3.7 provide the corresponding estimates in terms of the energy measure.
Proof of Theorem 2.2 . Fix f ∈ L l o c p ( X ; m ) f \in L^p_{\mathrm{loc}}(X;m) f ∈ L loc p ( X ; m ) , x ∈ X x \in X x ∈ X , and 0 < r < s 0<r<s 0 < r < s . When X X X has finite diameter, the case r > diam X r>\operatorname{diam} X r > diam X was settled above. In the remaining cases, let u = P r f u=P_rf u = P r f be the approximation constructed in Subsection 3.3.
Set B = B ( x , s ) B=B(x,s) B = B ( x , s ) . We first compare the oscillation of f f f with the approximation error and the oscillation of u u u . Since
∣ f B − u B ∣ = ∣ ( f − u ) B ∣ ≤ V ( x , s ) − 1 / p ∥ f − u ∥ L p ( B ) , \lvert f_B-u_B\rvert=\lvert(f-u)_B\rvert\le V(x,s)^{-1/p}\lVert f-u\rVert_{L^p(B)}, ∣ f B − u B ∣ = ∣( f − u ) B ∣ ≤ V ( x , s ) − 1/ p ∥ f − u ∥ L p ( B ) ,
we have
∥ f − f B ∥ L p ( B ) ≤ ∥ f − u ∥ L p ( B ) + ∥ u − u B ∥ L p ( B ) + V ( x , s ) 1 / p ∣ u B − f B ∣ \lVert f-f_B\rVert_{L^p(B)}\le\lVert f-u\rVert_{L^p(B)}+\lVert u-u_B\rVert_{L^p(B)}+V(x,s)^{1/p}\lvert u_B-f_B\rvert ∥ f − f B ∥ L p ( B ) ≤ ∥ f − u ∥ L p ( B ) + ∥ u − u B ∥ L p ( B ) + V ( x , s ) 1/ p ∣ u B − f B ∣
and therefore
∫ B ∣ f − f B ∣ p d m ≤ C p ( ∫ B ∣ f − u ∣ p d m + ∫ B ∣ u − u B ∣ p d m ) . (52) \int_B\lvert f-f_B\rvert^p\,\mathrm{d}m\le C_p\left(\int_B\lvert f-u\rvert^p\,\mathrm{d}m+\int_B\lvert u-u_B\rvert^p\,\mathrm{d}m\right). \tag*{(52)} ∫ B ∣ f − f B ∣ p d m ≤ C p ( ∫ B ∣ f − u ∣ p d m + ∫ B ∣ u − u B ∣ p d m ) . (52)
where C p C_p C p depends only on p p p .
By Lemma 3.6 ,
∫ B ∣ f − u ∣ p d m ≤ C Ψ ( r ) ∫ B ( x , 2 s ) M r f ( y , r ) d m ( y ) . (53) \int_B\lvert f-u\rvert^p\,\mathrm{d}m\le C\Psi(r)\int_{B(x,2s)}M_rf(y,r)\,\mathrm{d}m(y). \tag*{(53)} ∫ B ∣ f − u ∣ p d m ≤ C Ψ ( r ) ∫ B ( x , 2 s ) M r f ( y , r ) d m ( y ) . (53)
By Lemmas 3.3 and 3.7 ,
∫ B ∣ u − u B ∣ p d m ≤ C P Ψ ( s ) Γ ( u ) ( A P B ) ≤ C Ψ ( s ) ∫ B ( x , ( A P + 1 ) s ) M r f ( y , r ) d m ( y ) . (54) \begin{aligned}
\int_B\lvert u-u_B\rvert^p\,\mathrm{d}m\le C_P\Psi(s)\Gamma(u)(A_PB) \\
&\le C\Psi(s)\int_{B(x,(A_P+1)s)}M_rf(y,r)\,\mathrm{d}m(y). \tag*{(54)}
\end{aligned} ∫ B ∣ u − u B ∣ p d m ≤ C P Ψ ( s ) Γ ( u ) ( A P B ) ≤ C Ψ ( s ) ∫ B ( x , ( A P + 1 ) s ) M r f ( y , r ) d m ( y ) . (54)
Since r < s r<s r < s and Ψ \Psi Ψ is increasing, Ψ ( r ) ≤ Ψ ( s ) \Psi(r)\le\Psi(s) Ψ ( r ) ≤ Ψ ( s ) , and since A P ≥ 1 A_P\ge1 A P ≥ 1 , B ( x , 2 s ) ⊂ B ( x , ( A P + 1 ) s ) B(x,2s)\subset B(x,(A_P+1)s) B ( x , 2 s ) ⊂ B ( x , ( A P + 1 ) s ) . Combining (52) –(54) and setting
A M : = A P + 1 , A_M:=A_P+1, A M := A P + 1 ,
we obtain
∫ B ( x , s ) ∣ f − f B ( x , s ) ∣ p d m ≤ C Ψ ( s ) ∫ B ( x , A M s ) M r f ( y , r ) d m ( y ) . (55) \int_{B(x,s)}\lvert f-f_{B(x,s)}\rvert^p\,\mathrm{d}m\le C\Psi(s)\int_{B(x,A_Ms)}M_rf(y,r)\,\mathrm{d}m(y). \tag*{(55)} ∫ B ( x , s ) ∣ f − f B ( x , s ) ∣ p d m ≤ C Ψ ( s ) ∫ B ( x , A M s ) M r f ( y , r ) d m ( y ) . (55)
Divide by Ψ ( s ) V ( x , s ) \Psi(s)V(x,s) Ψ ( s ) V ( x , s ) . Using volume doubling once more at the fixed scale A M A_M A M ,
V ( x , A M s ) V ( x , s ) ≤ C D ( A M ) , \frac{V(x,A_Ms)}{V(x,s)}\le C_D(A_M), V ( x , s ) V ( x , A M s ) ≤ C D ( A M ) ,
we obtain from (55)
M f ( x , s ) ≤ C C D ( A M ) ∫ B ( x , A M s ) M f ( y , r ) d m ( y ) . M_f(x,s)\le CC_D(A_M)\int_{B(x,A_Ms)}M_f(y,r)\,\mathrm{d}m(y). M f ( x , s ) ≤ C C D ( A M ) ∫ B ( x , A M s ) M f ( y , r ) d m ( y ) .
This is ( mPI ( Ψ ) ) (\operatorname{mPI}(\Psi)) ( mPI ( Ψ )) with C M = C C D ( A M ) C_M=CC_D(A_M) C M = C C D ( A M ) . □
Remark 3.8. The capacity bound, normal contractions, and the algebra property enter only through the controlled partition of unity. The remaining argument uses volume doubling, PI ( Ψ ) \operatorname{PI}(\Psi) PI ( Ψ ) , the localized properties (G1), (G2), (G4), invariance under addition of constants, and fixed-ratio upper comparisons of Ψ \Psi Ψ .
Remark 3.9. An alternative proof of Theorem 2.2 may be obtained by adapting Yang [45, Proposition 4.5]. We sketch the argument. The result [45, Proposition 4.5] bounds the oscillation on a ball by localized pairwise differences at an arbitrary smaller scale. Applying it at scale r / 2 r/2 r /2 and averaging the pairwise differences over nearby centers yields ( mPI ( Ψ ) ) (\operatorname{mPI}(\Psi)) ( mPI ( Ψ )) , by volume doubling and doubling of Ψ \Psi Ψ . Yang works on unbounded spaces with a chain condition and two-sided power control of Ψ \Psi Ψ ; however, for this estimate, his argument extends to our setting upon using Lemma 3.4 , and only fixed-ratio comparisons of Ψ \Psi Ψ are needed. The passage from L p ( X ; m ) L^p(X;m) L p ( X ; m ) to L l o c p ( X ; m ) L^p_{\mathrm{loc}}(X;m) L loc p ( X ; m ) follows by localization. We did not follow that route to keep the paper self-contained.
Microscopic control
We say that the microscopic control condition ( M C ( Ψ ) ) (\mathrm{MC}(\Psi)) ( MC ( Ψ )) holds if there exist C M C > 0 C_{MC}>0 C M C > 0 and A M C ≥ 1 A_{MC}\geq1 A M C ≥ 1 such that, for every ball B ( x 0 , R ) B(x_0,R) B ( x 0 , R ) and every f ∈ F f\in\mathcal{F} f ∈ F ,
lim inf r ↓ 0 ∫ B ( x 0 , R ) M f ( x , r ) d m ( x ) ≤ C M C Γ ( f ) ( B ( x 0 , A M C R ) ) . \liminf_{r\downarrow0}\int_{B(x_0,R)}M_f(x,r)\,dm(x)\leq C_{MC}\Gamma(f)(B(x_0,A_{MC}R)). r ↓ 0 lim inf ∫ B ( x 0 , R ) M f ( x , r ) d m ( x ) ≤ C M C Γ ( f ) ( B ( x 0 , A M C R )) .
The following proposition relates the macroscopic inequality to the energy measure of the given form.
Proposition 3.10. Suppose that Assumption 2.1 holds.
(i) If (VD) and ( P I ( Ψ ) ) (\mathrm{PI}(\Psi)) ( PI ( Ψ )) hold, then ( M C ( Ψ ) ) (\mathrm{MC}(\Psi)) ( MC ( Ψ )) holds with A M C = 2 A P A_{MC}=2A_P A M C = 2 A P and C M C = C P C D ( 2 A P ) C_{MC}=C_PC_D(2A_P) C M C = C P C D ( 2 A P ) . In fact, the estimate holds with lim sup \limsup lim sup in place of lim inf \liminf lim inf .
(ii) If ( m P I ( Ψ ) ) (\mathrm{mPI}(\Psi)) ( mPI ( Ψ )) and ( M C ( Ψ ) ) (\mathrm{MC}(\Psi)) ( MC ( Ψ )) hold, then ( P I ( Ψ ) ) (\mathrm{PI}(\Psi)) ( PI ( Ψ )) holds with C P ′ = C M C M C C'_P=C_MC_{MC} C P ′ = C M C M C and A P ′ = A M A M C A'_P=A_MA_{MC} A P ′ = A M A M C , without assuming (VD) or ( cap ( Ψ ) ≤ ) (\operatorname{cap}(\Psi)_{\leq}) ( cap ( Ψ ) ≤ ) .
Consequently, under Assumption 2.1, (VD), and ( cap ( Ψ ) ≤ ) (\operatorname{cap}(\Psi)_{\leq}) ( cap ( Ψ ) ≤ ) ,
( P I ( Ψ ) ) ⟺ ( m P I ( Ψ ) ) and ( M C ( Ψ ) ) . (\mathrm{PI}(\Psi))\quad\Longleftrightarrow\quad(\mathrm{mPI}(\Psi))\text{ and }(\mathrm{MC}(\Psi)). ( PI ( Ψ )) ⟺ ( mPI ( Ψ )) and ( MC ( Ψ )) .
Proof of Proposition 3.10. Assume (VD) and ( P I ( Ψ ) ) (\mathrm{PI}(\Psi)) ( PI ( Ψ )) . Fix B = B ( x 0 , R ) B=B(x_0,R) B = B ( x 0 , R ) , f ∈ F f\in\mathcal{F} f ∈ F , and 0 < r < R 0<r<R 0 < r < R . By ( P I ( Ψ ) ) (\mathrm{PI}(\Psi)) ( PI ( Ψ )) ,
M f ( x , r ) ≤ C P Γ ( f ) ( B ( x , A P r ) ) V ( x , r ) . M_f(x,r)\leq C_P\frac{\Gamma(f)(B(x,A_Pr))}{V(x,r)}. M f ( x , r ) ≤ C P V ( x , r ) Γ ( f ) ( B ( x , A P r )) .
If x ∈ B ( z , A P r ) x\in B(z,A_Pr) x ∈ B ( z , A P r ) , then
V ( z , A P r ) ≤ V ( x , 2 A P r ) ≤ C D ( 2 A P ) V ( x , r ) . V(z,A_Pr)\leq V(x,2A_Pr)\leq C_D(2A_P)V(x,r). V ( z , A P r ) ≤ V ( x , 2 A P r ) ≤ C D ( 2 A P ) V ( x , r ) .
Tonelli’s theorem therefore gives
∫ B M f ( x , r ) d m ( x ) ≤ C P ∫ X ( ∫ B ∩ B ( z , A P r ) d m ( x ) V ( x , r ) ) d Γ ( f ) ( z ) ≤ C P C D ( 2 A P ) Γ ( f ) ( B ( x 0 , R + A P r ) ) ≤ C P C D ( 2 A P ) Γ ( f ) ( B ( x 0 , 2 A P R ) ) . \begin{aligned}
\int_B M_f(x,r)\,dm(x)
&\leq C_P\int_X\left(\int_{B\cap B(z,A_Pr)}\frac{dm(x)}{V(x,r)}\right)d\Gamma(f)(z)\\
&\leq C_PC_D(2A_P)\Gamma(f)(B(x_0,R+A_Pr))\\
&\leq C_PC_D(2A_P)\Gamma(f)(B(x_0,2A_PR)).
\end{aligned} ∫ B M f ( x , r ) d m ( x ) ≤ C P ∫ X ( ∫ B ∩ B ( z , A P r ) V ( x , r ) d m ( x ) ) d Γ ( f ) ( z ) ≤ C P C D ( 2 A P ) Γ ( f ) ( B ( x 0 , R + A P r )) ≤ C P C D ( 2 A P ) Γ ( f ) ( B ( x 0 , 2 A P R )) .
Taking the upper limit as r ↓ 0 r\downarrow0 r ↓ 0 proves the stated stronger estimate, and hence ( M C ( Ψ ) ) (\mathrm{MC}(\Psi)) ( MC ( Ψ )) .
Conversely, assume ( m P I ( Ψ ) ) (\mathrm{mPI}(\Psi)) ( mPI ( Ψ )) and ( M C ( Ψ ) ) (\mathrm{MC}(\Psi)) ( MC ( Ψ )) . Fix B = B ( x 0 , R ) B=B(x_0,R) B = B ( x 0 , R ) and f ∈ F f\in\mathcal{F} f ∈ F . For every 0 < r < R 0<r<R 0 < r < R , ( m P I ( Ψ ) ) (\mathrm{mPI}(\Psi)) ( mPI ( Ψ )) and A M ≥ 1 A_M\geq1 A M ≥ 1 give
∫ B ∣ f − f B ∣ p d m ≤ C M Ψ ( R ) V ( x 0 , R ) V ( x 0 , A M R ) ∫ B ( x 0 , A M R ) M f ( x , r ) d m ( x ) ≤ C M Ψ ( R ) ∫ B ( x 0 , A M R ) M f ( x , r ) d m ( x ) . \begin{aligned}
\int_B|f-f_B|^p\,dm
&\leq C_M\Psi(R)\frac{V(x_0,R)}{V(x_0,A_MR)}
\int_{B(x_0,A_MR)}M_f(x,r)\,dm(x)\\
&\leq C_M\Psi(R)\int_{B(x_0,A_MR)}M_f(x,r)\,dm(x).
\end{aligned} ∫ B ∣ f − f B ∣ p d m ≤ C M Ψ ( R ) V ( x 0 , A M R ) V ( x 0 , R ) ∫ B ( x 0 , A M R ) M f ( x , r ) d m ( x ) ≤ C M Ψ ( R ) ∫ B ( x 0 , A M R ) M f ( x , r ) d m ( x ) .
Taking the lower limit as r ↓ 0 r\downarrow0 r ↓ 0 and applying ( M C ( Ψ ) ) (\mathrm{MC}(\Psi)) ( MC ( Ψ )) on B ( x 0 , A M R ) B(x_0,A_MR) B ( x 0 , A M R ) yields
∫ B ∣ f − f B ∣ p d m ≤ C M C M C Ψ ( R ) Γ ( f ) ( B ( x 0 , A M A M C R ) ) . \int_B|f-f_B|^p\,dm\leq C_MC_{MC}\Psi(R)\Gamma(f)(B(x_0,A_MA_{MC}R)). ∫ B ∣ f − f B ∣ p d m ≤ C M C M C Ψ ( R ) Γ ( f ) ( B ( x 0 , A M A M C R )) .
This is ( P I ( Ψ ) ) (\mathrm{PI}(\Psi)) ( PI ( Ψ )) with C P ′ = C M C M C C'_P=C_MC_{MC} C P ′ = C M C M C and A P ′ = A M A M C A'_P=A_MA_{MC} A P ′ = A M A M C . No doubling or capacity estimate is used in this implication. The final equivalence follows from these two implications and Theorem 2.2.
Two-scale comparison of the Korevaar–Schoen functionals
We compare integrated ball oscillations with the Korevaar–Schoen functionals (2.11), then integrate (mPI(Ψ \Psi Ψ )) to prove Theorem 2.3 .
For f ∈ L p ( X ; m ) f \in L^p(X;m) f ∈ L p ( X ; m ) and r > 0 r>0 r > 0 , set
O r ( f ) : = ∫ X M f ( x , r ) d m ( x ) . (56) \mathcal{O}_r(f) := \int_X M_f(x,r)\,dm(x). \tag*{(56)} O r ( f ) := ∫ X M f ( x , r ) d m ( x ) . (56)
Lemma 3.11. Assume (VD) and (2.2). There is C ≥ 1 C \ge1 C ≥ 1 , depending only on p p p , C D C_D C D , C Ψ C_\Psi C Ψ , such that for every f ∈ L p ( X ; m ) f \in L^p(X;m) f ∈ L p ( X ; m ) and r > 0 r>0 r > 0 ,
O r ( f ) ≤ C J 2 r ( f ) , (57) \mathcal{O}_r(f) \le C J_{2r}(f), \tag*{(57)} O r ( f ) ≤ C J 2 r ( f ) , (57) and
J r ( f ) ≤ C O 2 r ( f ) . (58) J_r(f) \le C\mathcal{O}_{2r}(f). \tag*{(58)} J r ( f ) ≤ C O 2 r ( f ) . (58)
Proof. Set
D r ( f ) : = 1 Ψ ( r ) ∫ X ∫ B ( x , r ) ∫ B ( x , r ) ∣ f ( y ) − f ( z ) ∣ p d m ( z ) d m ( y ) d m ( x ) . D_r(f) := \frac{1}{\Psi(r)}\int_X\int_{B(x,r)}\int_{B(x,r)} |f(y)-f(z)|^p\,dm(z)\,dm(y)\,dm(x). D r ( f ) := Ψ ( r ) 1 ∫ X ∫ B ( x , r ) ∫ B ( x , r ) ∣ f ( y ) − f ( z ) ∣ p d m ( z ) d m ( y ) d m ( x ) .
Jensen’s inequality and the inequality ∣ a + b ∣ p ≤ 2 p − 1 ( ∣ a ∣ p + ∣ b ∣ p ) |a+b|^p \le2^{p-1}(|a|^p+|b|^p) ∣ a + b ∣ p ≤ 2 p − 1 ( ∣ a ∣ p + ∣ b ∣ p ) give
O r ( f ) ≤ D r ( f ) ≤ 2 p O r ( f ) . (59) \mathcal{O}_r(f) \le D_r(f) \le2^p\mathcal{O}_r(f). \tag*{(59)} O r ( f ) ≤ D r ( f ) ≤ 2 p O r ( f ) . (59)
By Tonelli’s theorem,
D r ( f ) = 1 Ψ ( r ) ∬ X × X ∣ f ( y ) − f ( z ) ∣ p H r ( y , z ) d m ( z ) d m ( y ) , D_r(f) = \frac{1}{\Psi(r)}\iint_{X\times X}|f(y)-f(z)|^p H_r(y,z)\,dm(z)\,dm(y), D r ( f ) = Ψ ( r ) 1 ∬ X × X ∣ f ( y ) − f ( z ) ∣ p H r ( y , z ) d m ( z ) d m ( y ) ,
where
H r ( y , z ) : = ∫ X 1 B ( x , r ) ( y ) 1 B ( x , r ) ( z ) V ( x , r ) 2 d m ( x ) . H_r(y,z) := \int_X \frac{\mathbf{1}_{B(x,r)}(y)\mathbf{1}_{B(x,r)}(z)}{V(x,r)^2}\,dm(x). H r ( y , z ) := ∫ X V ( x , r ) 2 1 B ( x , r ) ( y ) 1 B ( x , r ) ( z ) d m ( x ) .
The kernel vanishes when d ( y , z ) ≥ 2 r d(y,z)\ge2r d ( y , z ) ≥ 2 r . For x ∈ B ( y , r ) x\in B(y,r) x ∈ B ( y , r ) , volume doubling gives V ( x , r ) ≃ V ( y , r ) V(x,r)\simeq V(y,r) V ( x , r ) ≃ V ( y , r ) , and hence
H r ( y , z ) ≲ 1 B ( y , 2 r ) ( z ) V ( y , 2 r ) . H_r(y,z) \lesssim\frac{\mathbf{1}_{B(y,2r)}(z)}{V(y,2r)}. H r ( y , z ) ≲ V ( y , 2 r ) 1 B ( y , 2 r ) ( z ) .
Together with Ψ ( 2 r ) ≲ Ψ ( r ) \Psi(2r)\lesssim\Psi(r) Ψ ( 2 r ) ≲ Ψ ( r ) , this proves D r ( f ) ≲ J 2 r ( f ) D_r(f)\lesssim J_{2r}(f) D r ( f ) ≲ J 2 r ( f ) and therefore (3.32). Conversely, if d ( y , z ) < r d(y,z)<r d ( y , z ) < r , then every x ∈ B ( y , r / 2 ) x\in B(y,r/2) x ∈ B ( y , r /2 ) satisfies y , z ∈ B ( x , 2 r ) y,z\in B(x,2r) y , z ∈ B ( x , 2 r ) , and fixed-scale volume comparison gives H 2 r ( y , z ) ≳ V ( y , r ) − 1 H_{2r}(y,z)\gtrsim V(y,r)^{-1} H 2 r ( y , z ) ≳ V ( y , r ) − 1 . Therefore
J r ( f ) ≲ Ψ ( 2 r ) Ψ ( r ) D 2 r ( f ) ≲ O 2 r ( f ) , J_r(f)\lesssim\frac{\Psi(2r)}{\Psi(r)}D_{2r}(f)\lesssim\mathcal{O}_{2r}(f), J r ( f ) ≲ Ψ ( r ) Ψ ( 2 r ) D 2 r ( f ) ≲ O 2 r ( f ) ,
which proves (3.33). □
Proof of Theorem 2.3. Let f ∈ L p ( X ; m ) f\in L^p(X;m) f ∈ L p ( X ; m ) and 0 < r < s 0<r<s 0 < r < s . Integrating (mPI(Ψ \Psi Ψ )) in x x x and applying Tonelli’s theorem gives
O s ( f ) ≤ C M ∫ X M f ( y , r ) ( ∫ B ( y , A M s ) d m ( x ) V ( x , A M s ) ) d m ( y ) . \mathcal{O}_s(f)\le C_M\int_X M_f(y,r)\left(\int_{B(y,A_Ms)}\frac{dm(x)}{V(x,A_Ms)}\right)dm(y). O s ( f ) ≤ C M ∫ X M f ( y , r ) ( ∫ B ( y , A M s ) V ( x , A M s ) d m ( x ) ) d m ( y ) .
For x ∈ B ( y , A M s ) x\in B(y,A_Ms) x ∈ B ( y , A M s ) , volume doubling gives V ( y , A M s ) ≲ V ( x , A M s ) V(y,A_Ms)\lesssim V(x,A_Ms) V ( y , A M s ) ≲ V ( x , A M s ) , so the inner integral is uniformly bounded. This proves
O s ( f ) ≤ C O r ( f ) , (60) \mathcal{O}_s(f)\le C\mathcal{O}_r(f), \tag*{(60)} O s ( f ) ≤ C O r ( f ) , (60)
with C C C depending only on C D C_D C D , C M C_M C M , and A M A_M A M . Lemma 3.11, applied at scales s s s and r / 2 r/2 r /2 , now yields
J s ( f ) ≲ O 2 s ( f ) ≲ O r / 2 ( f ) ≲ J r ( f ) , J_s(f)\lesssim\mathcal{O}_{2s}(f)\lesssim\mathcal{O}_{r/2}(f)\lesssim J_r(f), J s ( f ) ≲ O 2 s ( f ) ≲ O r /2 ( f ) ≲ J r ( f ) ,
which is (2.13). For each fixed s > 0 s>0 s > 0 , let r ↓ 0 r\downarrow0 r ↓ 0 in (2.13), and then take the supremum over s s s to obtain (2.14). □
Connectedness and lower scaling
We now prove Proposition 2.4, which requires no energy form, chain condition, or doubling assumption on Ψ \Psi Ψ . Yang [44, Proposition 2.1 and Remark 2.2] proved the lower p p p -scaling (2.16) directly from a chain condition, volume doubling, PI ( Ψ ) \operatorname{PI}(\Psi) PI ( Ψ ) , and the capacity bound. Proposition 2.4 derives the same conclusion from ( mPI ( Ψ ) ) (\operatorname{mPI}(\Psi)) ( mPI ( Ψ )) , with connectedness also obtained as a consequence. To our knowledge, this implication from the macroscopic Poincaré inequality is new in this generality. The proof first establishes connectedness and then applies ( mPI ( Ψ ) ) (\operatorname{mPI}(\Psi)) ( mPI ( Ψ )) to a Lipschitz tent function.
Proof of Proposition 2.4. By the metric-measure observation in Subsection 2.1, X X X is proper. We first show that X X X is connected. If X = U ∪ ˙ W X=U\mathbin{\dot\cup}W X = U ∪ ˙ W were a nontrivial clopen separation, choose x ∈ U x\in U x ∈ U , y ∈ W y\in W y ∈ W , and s > d ( x , y ) s>d(x,y) s > d ( x , y ) . Put
g : = 1 U ∩ B ( x , ( A M + 2 ) s ) ∈ L p ( X ; m ) . g:=\mathbf{1}_{U\cap B(x,(A_M+2)s)}\in L^p(X;m). g := 1 U ∩ B ( x , ( A M + 2 ) s ) ∈ L p ( X ; m ) .
The sets
U ∩ B ( x , ( A M + 1 ) s ) ‾ and W ∩ B ( x , ( A M + 1 ) s ) ‾ U\cap\overline{B(x,(A_M+1)s)}\qquad\text{and}\qquad W\cap\overline{B(x,(A_M+1)s)} U ∩ B ( x , ( A M + 1 ) s ) and W ∩ B ( x , ( A M + 1 ) s )
are nonempty disjoint compact sets. Let Δ > 0 \Delta>0 Δ > 0 be their distance. For 0 < r < min { s , Δ / 2 } 0<r<\min\{s,\Delta/2\} 0 < r < min { s , Δ/2 } , we have
M g ( z , r ) = 0 , z ∈ B ( x , A M s ) . M_g(z,r)=0,\qquad z\in B(x,A_Ms). M g ( z , r ) = 0 , z ∈ B ( x , A M s ) .
On the other hand, B ( x , s ) B(x,s) B ( x , s ) meets both U U U and W W W in sets of positive measure, by full support, so M g ( x , s ) > 0 M_g(x,s)>0 M g ( x , s ) > 0 . This contradicts ( mPI ( Ψ ) ) (\operatorname{mPI}(\Psi)) ( mPI ( Ψ )) . Thus X X X is connected.
Fix 0 < r < R < diam X 0<r<R<\operatorname{diam}X 0 < r < R < diam X . Choose x , y ∈ X x,y\in X x , y ∈ X with d ( x , y ) > R d(x,y)>R d ( x , y ) > R . Since the image of w ↦ d ( x , w ) w\mapsto d(x,w) w ↦ d ( x , w ) is connected, there is a point z z z with d ( x , z ) = R / 2 d(x,z)=R/2 d ( x , z ) = R /2 . Consider the compactly supported 1 1 1 -Lipschitz function
f R ( w ) : = ( R / 8 − d ( x , w ) ) + ∈ L p ( X ; m ) . f_R(w):=(R/8-d(x,w))_+\in L^p(X;m). f R ( w ) := ( R /8 − d ( x , w ) ) + ∈ L p ( X ; m ) .
On E = B ( x , R / 32 ) E=B(x,R/32) E = B ( x , R /32 ) one has f R ≥ 3 R / 32 f_R\ge3R/32 f R ≥ 3 R /32 , whereas f R = 0 f_R=0 f R = 0 on F = B ( z , R / 32 ) F=B(z,R/32) F = B ( z , R /32 ) . Both balls lie in B = B ( x , R ) B=B(x,R) B = B ( x , R ) , and volume doubling gives
m ( E ) , m ( F ) ≥ c V ( x , R ) . (61) m(E),m(F)\ge cV(x,R). \tag*{(61)} m ( E ) , m ( F ) ≥ c V ( x , R ) . (61)
If ( f R ) B ≤ 3 R / 64 (f_R)_B\le3R/64 ( f R ) B ≤ 3 R /64 , integrate over E E E ; otherwise, integrate over F F F . In either case, (3.36) gives
M f R ( x , R ) ≥ c R p Ψ ( R ) . (62) M_{f_R}(x,R)\ge c\frac{R^p}{\Psi(R)}. \tag*{(62)} M f R ( x , R ) ≥ c Ψ ( R ) R p . (62)
On the other hand, the Lipschitz bound implies, for every v ∈ X v\in X v ∈ X ,
M f R ( v , r ) ≤ C r p Ψ ( r ) . (63) M_{f_R}(v,r)\le C\frac{r^p}{\Psi(r)}. \tag*{(63)} M f R ( v , r ) ≤ C Ψ ( r ) r p . (63)
Indeed, the oscillation of f R f_R f R on a ball of radius r r r is at most 2 r 2r 2 r . Applying ( mPI ( Ψ ) ) (\operatorname{mPI}(\Psi)) ( mPI ( Ψ )) to f R f_R f R at scales r < R r<R r < R and using (3.37)–(3.38), we obtain
R p Ψ ( R ) ≤ C r p Ψ ( r ) . \frac{R^p}{\Psi(R)}\le C\frac{r^p}{\Psi(r)}. Ψ ( R ) R p ≤ C Ψ ( r ) r p .
which is (2.16) for R < diam X R<\operatorname{diam}X R < diam X . If diam X < ∞ \operatorname{diam}X<\infty diam X < ∞ , let R ↑ diam X R\uparrow\operatorname{diam}X R ↑ diam X and use continuity of Ψ \Psi Ψ . The case r = R r=R r = R is immediate. □ \square □
Corollary 3.12. Under the hypotheses of Theorem 2.2,
Ψ ( r ) Ψ ( R ) ≤ C ( r R ) p \frac{\Psi(r)}{\Psi(R)}\le C\left(\frac{r}{R}\right)^p Ψ ( R ) Ψ ( r ) ≤ C ( R r ) p
*whenever 0 < r ≤ R < ∞ 0<r\le R<\infty 0 < r ≤ R < ∞ and, if diam X < ∞ \operatorname{diam}X<\infty diam X < ∞ , R ≤ diam X R\le\operatorname{diam}X R ≤ diam X . The constant depends only on p p p , C D C_D C D , C Ψ C_\Psi C Ψ , C P C_P C P , A P A_P A P , C c C_c C c , and A c A_c A c .
Proof. Combine Theorem 2.2 with Proposition 2.4.
Remark 3.13. Under the hypotheses of Theorem 2.2, Corollary 3.12 and (4) give
C Ψ − 1 ( R r ) p ≤ Ψ ( R ) Ψ ( r ) ≤ C Ψ ( R r ) log 2 C Ψ C_{\Psi}^{-1}\left(\frac{R}{r}\right)^{p}\leq\frac{\Psi(R)}{\Psi(r)}\leq C_{\Psi}\left(\frac{R}{r}\right)^{\log_{2}C_{\Psi}} C Ψ − 1 ( r R ) p ≤ Ψ ( r ) Ψ ( R ) ≤ C Ψ ( r R ) l o g 2 C Ψ
whenever 0 < r ≤ R < ∞ 0<r\leq R<\infty 0 < r ≤ R < ∞ and, if diam X < ∞ \operatorname{diam}X<\infty diam X < ∞ , R ≤ diam X R\leq\operatorname{diam}X R ≤ diam X . The upper bound follows from C Ψ ( R / r ) ≤ C Ψ 1 + log 2 ( R / r ) C_{\Psi}(R/r)\leq C_{\Psi}^{1+\log_{2}(R/r)} C Ψ ( R / r ) ≤ C Ψ 1 + l o g 2 ( R / r ) . Thus every scale function admissible in the forward results is of two-sided power type, with lower exponent p p p and upper exponent log 2 C Ψ \log_{2}C_{\Psi} log 2 C Ψ ; letting r ↓ 0 r\downarrow0 r ↓ 0 with R R R fixed shows in particular that C Ψ ≥ 2 p C_{\Psi}\geq2^{p} C Ψ ≥ 2 p . These bounds allow nonpower scales, including the tree and Laakso-type examples of Subsection 2.7. For Ψ ( r ) = r β \Psi(r)=r^{\beta} Ψ ( r ) = r β , combining Shimizu [43 Theorem 4.6] with Baudoin [4 Theorem 4.1] also gives β ≥ p \beta\geq p β ≥ p ; Proposition 2.4 gives relative scaling for a general Ψ \Psi Ψ .
Identification of the energy domain
We recall the domain characterization of Shimizu [43 Corollaries 1.14 and 1.15] , whose argument applies under the hypotheses of Theorem 2.2 with the adjustments described below. It gives
F = { f ∈ L p ( X ; m ) : lim inf r ↓ 0 J r ( f ) < ∞ } = B Ψ , (64) \mathcal{F}=\left\{f\in L^{p}(X;m):\liminf_{r\downarrow0}J_{r}(f)<\infty\right\}=B_{\Psi}, \tag*{(64)} F = { f ∈ L p ( X ; m ) : r ↓ 0 lim inf J r ( f ) < ∞ } = B Ψ , (64)
and
E ( f ) ≍ lim inf r ↓ 0 J r ( f ) ≍ sup r > 0 J r ( f ) , f ∈ F . (65) \mathcal{E}(f)\asymp\liminf_{r\downarrow0}J_{r}(f)\asymp\sup_{r>0}J_{r}(f),\qquad f\in\mathcal{F}. \tag*{(65)} E ( f ) ≍ r ↓ 0 lim inf J r ( f ) ≍ r > 0 sup J r ( f ) , f ∈ F . (65)
The comparison constants depend only on p p p , C D C_{D} C D , C Ψ C_{\Psi} C Ψ , C P C_{P} C P , A P A_{P} A P , C c C_{c} C c , and A c A_{c} A c .
Shimizu’s framework allows noncompact spaces, singular energy measures, and nonpower scale functions. His two-sided power control of Ψ \Psi Ψ follows here from doubling and Corollary 3.12, at scales not exceeding diam X \operatorname{diam}X diam X ; see Remark 3.13. Proposition 2.4 gives connectedness, and hence uniform perfectness, while PI ( Ψ ) \operatorname{PI}(\Psi) PI ( Ψ ) gives E − 1 ( 0 ) ⊂ R 1 \mathcal{E}^{-1}(0)\subset\mathbb{R}1 E − 1 ( 0 ) ⊂ R 1 . Closedness and the Clarkson inequalities give reflexivity of the energy domain [26 Proposition 3.13] .
For the discrete convolution argument in [43 Lemmas 3.10–3.11 and proof of Corollary 1.15] , Lemma 3.4 supplies the required partition under our normal-contraction and algebra assumptions. The localized seminorm property and (G4) permit its use with measure-theoretic supports; continuity of the partition functions is not needed for the approximation and weak-compactness argument. If X X X is bounded, only small radii enter the domain identification, and PI ( Ψ ) \operatorname{PI}(\Psi) PI ( Ψ ) controls the supremum over the remaining radii. We use (64) –(65) without repeating the proof. On the Sierpiński carpet, the corresponding identification is also given in [38 Theorem 1.4] .
Construction of a p p p -energy from the macroscopic inequality
For the construction we assume the metric-measure hypotheses of Theorem 2.5, volume doubling, doubling of Ψ \Psi Ψ , (mPI(Ψ \Psi Ψ )), and the core condition (18) . We construct an energy form from these data.
The ball kernel and the Korevaar–Schoen energy
For r > 0 r>0 r > 0 , define
k r ( x , y ) : = 1 B ( x , r ) ( y ) Ψ ( r ) V ( x , r ) . (66) k_{r}(x,y):=\frac{\mathbf{1}_{B(x,r)}(y)}{\Psi(r)V(x,r)}. \tag*{(66)} k r ( x , y ) := Ψ ( r ) V ( x , r ) 1 B ( x , r ) ( y ) . (66)
The measurability observation in Subsection 2.1 shows that k r k_{r} k r is Borel. The standing volume assumptions give 0 < V ( x , r ) < ∞ 0<V(x,r)<\infty 0 < V ( x , r ) < ∞ . The associated nonlocal functional is J r J_{r} J r , and the kernel is asymptotically local with δ ( r ) = r \delta(r)=r δ ( r ) = r .
By Theorem 2.3,
sup s > 0 J s ( f ) ≤ C lim inf r ↓ 0 J r ( f ) , f ∈ B Ψ . \sup_{s>0} J_s(f) \le C \liminf_{r\downarrow0} J_r(f), \qquad f \in B_\Psi. s > 0 sup J s ( f ) ≤ C r ↓ 0 lim inf J r ( f ) , f ∈ B Ψ .
This is the weak monotonicity condition in Sasaya [41 Definition 8.1] . By Sasaya [41 Theorem 8.3] , based on the construction of Kajino–Shimizu [27 Theorems 6.3.8 and 6.4.2] , whose hypotheses (local compactness and separability, a full-support Radon measure, asymptotic locality of the kernel, and weak monotonicity) are satisfied here, there exists a sequence r n ↓ 0 r_n \downarrow0 r n ↓ 0 for which
E ~ K S ( f ) : = lim n → ∞ J r n ( f ) , f ∈ B Ψ , \widetilde{\mathcal{E}}_{\mathrm{KS}}(f) := \lim_{n\to\infty} J_{r_n}(f), \qquad f \in B_\Psi, E KS ( f ) := n → ∞ lim J r n ( f ) , f ∈ B Ψ ,
is well defined and satisfies
C − 1 sup r > 0 J r ( f ) ≤ E ~ K S ( f ) ≤ C lim inf r ↓ 0 J r ( f ) . (67) C^{-1}\sup_{r>0} J_r(f) \le\widetilde{\mathcal{E}}_{\mathrm{KS}}(f) \le C\liminf_{r\downarrow0} J_r(f). \tag*{(67)} C − 1 r > 0 sup J r ( f ) ≤ E KS ( f ) ≤ C r ↓ 0 lim inf J r ( f ) . (67)
Under (2.17), Sasaya’s theorem shows that
F K S : = B Ψ ∩ C c ( X ) ‾ ∥ ⋅ ∥ B Ψ \mathcal{F}_{\mathrm{KS}} := \overline{B_\Psi\cap C_c(X)}^{\|\cdot\|_{B_\Psi}} F KS := B Ψ ∩ C c ( X ) ∥ ⋅ ∥ B Ψ
and
E K S : = E ~ K S ∣ F K S \mathcal{E}_{\mathrm{KS}} := \left.\widetilde{\mathcal{E}}_{\mathrm{KS}}\right|_{\mathcal{F}_{\mathrm{KS}}} E KS := E KS F KS
form a regular strongly local p p p -energy with the full p p p -Clarkson properties [41 Theorem 8.3] . The generalized p p p -contraction property follows from [27 Theorem 6.3.8(b)] . The Markov and algebra properties follow from Kajino–Shimizu [27 Proposition 6.2.3, items (2) and (4)] . The associated measure Γ K S ( u ) \Gamma^{\mathrm{KS}}(u) Γ KS ( u ) is identified in [41 Corollary 8.4] .
Lemma 4.1. For a measurable set U ⊂ X U \subset X U ⊂ X and f ∈ L p ( X ; m ) f \in L^p(X;m) f ∈ L p ( X ; m ) , define, for r > 0 r>0 r > 0 ,
J r ( f ; U ) : = 1 Ψ ( r ) ∫ U ∫ B ( y , r ) ∣ f ( y ) − f ( z ) ∣ p d m ( z ) d m ( y ) , J_r(f;U) := \frac{1}{\Psi(r)} \int_U \int_{B(y,r)} |f(y)-f(z)|^p\,dm(z)\,dm(y), J r ( f ; U ) := Ψ ( r ) 1 ∫ U ∫ B ( y , r ) ∣ f ( y ) − f ( z ) ∣ p d m ( z ) d m ( y ) , and, for q > 0 q>0 q > 0 , put U q : = { y ∈ X : d ( y , U ) < q } U_q := \{y \in X : d(y,U)<q\} U q := { y ∈ X : d ( y , U ) < q } . Then for every q > 0 q>0 q > 0 ,
∫ U M f ( x , q ) d m ( x ) ≤ C J 2 q ( f ; U q ) , (68) \int_U M_f(x,q)\,dm(x) \le C J_{2q}(f;U_q), \tag*{(68)} ∫ U M f ( x , q ) d m ( x ) ≤ C J 2 q ( f ; U q ) , (68) where C C C depends only on p p p , C D C_D C D , C Ψ C_\Psi C Ψ .
Proof. Restrict the outer variable in the kernel argument of Lemma 3.11 to U U U . The resulting kernel vanishes unless y ∈ U q y \in U_q y ∈ U q and d ( y , z ) < 2 q d(y,z)<2q d ( y , z ) < 2 q . Volume comparison bounds it by
C 1 U q ( y ) 1 B ( y , 2 q ) ( z ) V ( y , 2 q ) . C\frac{\mathbf{1}_{U_q}(y)\mathbf{1}_{B(y,2q)}(z)}{V(y,2q)}. C V ( y , 2 q ) 1 U q ( y ) 1 B ( y , 2 q ) ( z ) .
The doubling of Ψ \Psi Ψ gives (4.3). ∎
Recovery of the Poincaré inequality
Proof of Theorem 2.5. It remains to prove (2.19). First let u ∈ B Ψ ∩ C c ( X ) u \in B_\Psi\cap C_c(X) u ∈ B Ψ ∩ C c ( X ) , fix x ∈ X x \in X x ∈ X , s > 0 s>0 s > 0 , and put U = B ( x , A M s ) U=B(x,A_Ms) U = B ( x , A M s ) , where A M A_M A M is the dilation in (mPI(Ψ \Psi Ψ )). Set q n = r n / 2 q_n=r_n/2 q n = r n /2 . For all sufficiently large n n n , q n < s q_n<s q n < s , hence
M u ( x , s ) ≤ C V ( x , A M s ) ∫ U M u ( y , q n ) d m ( y ) ≤ C V ( x , A M s ) J r n ( u ; U q n ) . (69) M_u(x,s) \le\frac{C}{V(x,A_Ms)}\int_U M_u(y,q_n)\,dm(y) \le\frac{C}{V(x,A_Ms)}J_{r_n}(u;U_{q_n}). \tag*{(69)} M u ( x , s ) ≤ V ( x , A M s ) C ∫ U M u ( y , q n ) d m ( y ) ≤ V ( x , A M s ) C J r n ( u ; U q n ) . (69)
by (mPI(Ψ \Psi Ψ )) and Lemma 4.1. For large n n n , U q n ⊂ B ( x , ( A M + 1 ) s ) U_{q_n} \subset B(x,(A_M+1)s) U q n ⊂ B ( x , ( A M + 1 ) s ) . Since this ball has compact closure, choose ϕ ∈ C c ( X ) \phi\in C_c(X) ϕ ∈ C c ( X ) with 0 ≤ ϕ ≤ 1 0 \le\phi\le1 0 ≤ ϕ ≤ 1 ,
ϕ = 1 on B ( x , ( A M + 1 ) s ) , supp ϕ ⊂ B ( x , ( A M + 2 ) s ) . \phi=1\ \text{on }B(x,(A_M+1)s), \qquad\operatorname{supp}\phi\subset B(x,(A_M+2)s). ϕ = 1 on B ( x , ( A M + 1 ) s ) , supp ϕ ⊂ B ( x , ( A M + 2 ) s ) .
Then
J_{r_n}(u;U_{q_n}) \le\frac{1}{\Psi(r_n)}\int_X \phi(y)\fint_{B(y,r_n)} |u(y)-u(z)|^p\,dm(z)\,dm(y). \tag*{(70)}
Define the measures
d\nu_n^u(y) := \frac{1}{\Psi(r_n)}\fint_{B(y,r_n)} |u(y)-u(z)|^p\,dm(z)\,dm(y).
Kajino–Shimizu [27, Theorem 6.4.2, formula (6.4.7)] show that, along the same sequence ( r n ) (r_n) ( r n ) as in the construction of E K S \mathcal{E}_{\mathrm{KS}} E KS , the integrals ∫ X ϕ d ν n u \int_X \phi\,d\nu_n^u ∫ X ϕ d ν n u converge whenever u ∈ B Ψ ∩ C b ( X ) u\in\mathcal{B}_{\Psi}\cap C_b(X) u ∈ B Ψ ∩ C b ( X ) , ϕ ∈ B Ψ ∩ L ∞ ( X ; m ) \phi\in\mathcal{B}_{\Psi}\cap L^\infty(X;m) ϕ ∈ B Ψ ∩ L ∞ ( X ; m ) , and one of u , ϕ u,\phi u , ϕ has compact support; for ϕ ∈ B Ψ ∩ C c ( X ) \phi\in\mathcal{B}_{\Psi}\cap C_c(X) ϕ ∈ B Ψ ∩ C c ( X ) , the limit is the integral of ϕ \phi ϕ against the p p p -energy measure of u u u . Since u ∈ B Ψ ∩ C c ( X ) u\in\mathcal{B}_{\Psi}\cap C_c(X) u ∈ B Ψ ∩ C c ( X ) , this gives
∫ X ϕ d ν n u ⟶ ∫ X ϕ d Γ K S ( u ) , ϕ ∈ B Ψ ∩ C c ( X ) , \int_X \phi\,d\nu_n^u \longrightarrow\int_X \phi\,d\Gamma^{\mathrm{KS}}(u),\qquad\phi\in\mathcal{B}_{\Psi}\cap C_c(X), ∫ X ϕ d ν n u ⟶ ∫ X ϕ d Γ KS ( u ) , ϕ ∈ B Ψ ∩ C c ( X ) ,
where the identification of the limiting measure with the canonical energy measure Γ K S ( u ) \Gamma^{\mathrm{KS}}(u) Γ KS ( u ) is [41 Theorem 8.3 and Corollary 8.4] . By (2.17), choose ϕ ℓ ∈ B Ψ ∩ C c ( X ) \phi_\ell\in\mathcal{B}_{\Psi}\cap C_c(X) ϕ ℓ ∈ B Ψ ∩ C c ( X ) with ∥ ϕ ℓ − ϕ ∥ ∞ → 0 \|\phi_\ell-\phi\|_\infty\to0 ∥ ϕ ℓ − ϕ ∥ ∞ → 0 . Since
sup n ν n u ( X ) = sup n J r n ( u ) < ∞ , Γ K S ( u ) ( X ) = E K S ( u ) < ∞ , \sup_n\nu_n^u(X)=\sup_n J_{r_n}(u)<\infty,\qquad\Gamma^{\mathrm{KS}}(u)(X)=\mathcal{E}_{\mathrm{KS}}(u)<\infty, n sup ν n u ( X ) = n sup J r n ( u ) < ∞ , Γ KS ( u ) ( X ) = E KS ( u ) < ∞ ,
uniform approximation extends the preceding convergence to ϕ \phi ϕ . Passing to the limit in (4.4) and (4.5) gives
M u ( x , s ) ≤ C V ( x , A M s ) Γ K S ( u ) ( B ( x , ( A M + 2 ) s ) ) . M_u(x,s)\le\frac{C}{V(x,A_Ms)}\Gamma^{\mathrm{KS}}(u)(B(x,(A_M+2)s)). M u ( x , s ) ≤ V ( x , A M s ) C Γ KS ( u ) ( B ( x , ( A M + 2 ) s )) .
Multiplying by Ψ ( s ) V ( x , s ) \Psi(s)V(x,s) Ψ ( s ) V ( x , s ) and using V ( x , s ) ≤ V ( x , A M s ) V(x,s)\le V(x,A_Ms) V ( x , s ) ≤ V ( x , A M s ) yields
∫ B ( x , s ) ∣ u − u B ( x , s ) ∣ p d m ≤ C Ψ ( s ) Γ K S ( u ) ( B ( x , ( A M + 2 ) s ) ) . (71) \int_{B(x,s)} |u-u_{B(x,s)}|^p\,dm\le C\Psi(s)\Gamma^{\mathrm{KS}}(u)(B(x,(A_M+2)s)). \tag*{(71)} ∫ B ( x , s ) ∣ u − u B ( x , s ) ∣ p d m ≤ C Ψ ( s ) Γ KS ( u ) ( B ( x , ( A M + 2 ) s )) . (71)
Finally, let u ∈ F K S u\in\mathcal{F}_{\mathrm{KS}} u ∈ F KS and choose u j ∈ B Ψ ∩ C c ( X ) u_j\in\mathcal{B}_{\Psi}\cap C_c(X) u j ∈ B Ψ ∩ C c ( X ) converging to u u u in ∥ ⋅ ∥ B Ψ \|\cdot\|_{\mathcal{B}_{\Psi}} ∥ ⋅ ∥ B Ψ . By (4.2), u j → u u_j\to u u j → u in L p L^p L p and E K S ( u j − u ) → 0 \mathcal{E}_{\mathrm{KS}}(u_j-u)\to0 E KS ( u j − u ) → 0 . The left-hand side of (4.6) converges because
∥ ( u j − u ) − ( u j − u ) B ( x , s ) ∥ L p ( B ( x , s ) ) ≤ 2 ∥ u j − u ∥ L p ( B ( x , s ) ) . \|(u_j-u)-(u_j-u)_{B(x,s)}\|_{L^p(B(x,s))}\le2\|u_j-u\|_{L^p(B(x,s))}. ∥ ( u j − u ) − ( u j − u ) B ( x , s ) ∥ L p ( B ( x , s )) ≤ 2∥ u j − u ∥ L p ( B ( x , s )) .
For any Borel set E E E , the seminorm property of the canonical energy measure gives
∣ Γ K S ( u j ) ( E ) 1 / p − Γ K S ( u ) ( E ) 1 / p ∣ ≤ Γ K S ( u j − u ) ( E ) 1 / p ≤ E K S ( u j − u ) 1 / p → 0. \left|\Gamma^{\mathrm{KS}}(u_j)(E)^{1/p}-\Gamma^{\mathrm{KS}}(u)(E)^{1/p}\right|
\le\Gamma^{\mathrm{KS}}(u_j-u)(E)^{1/p}
\le\mathcal{E}_{\mathrm{KS}}(u_j-u)^{1/p}\to0. Γ KS ( u j ) ( E ) 1/ p − Γ KS ( u ) ( E ) 1/ p ≤ Γ KS ( u j − u ) ( E ) 1/ p ≤ E KS ( u j − u ) 1/ p → 0.
Thus (4.6) passes to the limit and proves (2.19) with A = A M + 2 A=A_M+2 A = A M + 2 .
The core condition and reconstruction
The macroscopic inequality is monotone in the scale function.
Remark 4.2. Let Φ : [ 0 , ∞ ) → [ 0 , ∞ ) \Phi:[0,\infty)\to[0,\infty) Φ : [ 0 , ∞ ) → [ 0 , ∞ ) be nondecreasing with Φ > 0 \Phi>0 Φ > 0 on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) , and write M Φ Ψ M_{\Phi}^{\Psi} M Φ Ψ for the quantity (2.10) computed with the scale Ψ \Psi Ψ . If (mPI( Ψ ) (\Psi) ( Ψ ) ) holds for Ψ \Psi Ψ , then it holds for Ψ Φ \Psi\Phi ΨΦ with the same constants C M C_M C M and A M A_M A M . Indeed, for 0 < r < s 0<r<s 0 < r < s , since Φ ( r ) ≤ Φ ( s ) \Phi(r)\le\Phi(s) Φ ( r ) ≤ Φ ( s ) ,
M Φ Ψ Φ ( x , s ) = M f Ψ ( x , s ) Φ ( s ) ≤ C M ∫ B ( x , A M s ) M f Ψ ( y , r ) Φ ( s ) d m ( y ) ≤ C M ∫ B ( x , A M s ) M f Ψ Φ ( y , r ) d m ( y ) . M_{\Phi}^{\Psi\Phi}(x,s)=\frac{M^{\Psi}_{f}(x,s)}{\Phi(s)}
\le C_M\int_{B(x,A_Ms)}\frac{M^{\Psi}_{f}(y,r)}{\Phi(s)}\,dm(y)
\le C_M\int_{B(x,A_Ms)}M^{\Psi\Phi}_{f}(y,r)\,dm(y). M Φ ΨΦ ( x , s ) = Φ ( s ) M f Ψ ( x , s ) ≤ C M ∫ B ( x , A M s ) Φ ( s ) M f Ψ ( y , r ) d m ( y ) ≤ C M ∫ B ( x , A M s ) M f ΨΦ ( y , r ) d m ( y ) .
Hence (mPI( Ψ ) (\Psi) ( Ψ ) ) alone does not determine the scale: it only forces Ψ \Psi Ψ to grow at least like r p r^p r p , by Proposition 2.4. In Theorem 2.5, it is the density condition (2.17) that excludes scale functions which grow too fast.
In particular, the condition (2.17) does not follow from (mPI(Ψ \Psi Ψ )) alone. On [ 0 , 1 ] [0,1] [ 0 , 1 ] with Lebesgue measure, mPI(r p r^p r p ) holds by Theorem 2.2 applied to the standard p p p -energy, and Remark 4.2 gives mPI(r β r^{\beta} r β ) for every β ≥ p \beta\ge p β ≥ p . For β > p \beta> p β > p , however, the corresponding supercritical Korevaar–Schoen class contains only constants, since [ 0 , 1 ] [0,1] [ 0 , 1 ] is a connected Poincaré space; see Kajino–Shimizu [27, Introduction] and Baudoin [4 ] . Thus its continuous finite-energy core is not uniformly dense in C ( [ 0 , 1 ] ) C([0,1]) C ([ 0 , 1 ]) .
Remark 4.3. Note that the core condition (2.17) holds if, for every x ∈ X x \in X x ∈ X and every relatively compact open neighborhood U U U of x x x there exists η ∈ C c ( X ) \eta\in C_c(X) η ∈ C c ( X ) such that
0 ≤ η ≤ 1 , η = 1 on an open neighborhood of x , supp η ⊂ U , sup r > 0 J r ( η ) < ∞ . (72) \begin{gathered}
0 \le\eta\le1,\\
\eta= 1 \text{ on an open neighborhood of }x,\qquad\operatorname{supp}\eta\subset U,\\
\sup_{r>0} J_r(\eta) < \infty. \tag*{(72)}
\end{gathered} 0 ≤ η ≤ 1 , η = 1 on an open neighborhood of x , supp η ⊂ U , r > 0 sup J r ( η ) < ∞. (72)
Indeed, for every r > 0 r > 0 r > 0 , Minkowski’s inequality gives
J r ( f g ) 1 / p ≤ ∥ f ∥ ∞ J r ( g ) 1 / p + ∥ g ∥ ∞ J r ( f ) 1 / p , f , g ∈ B Ψ ∩ C c ( X ) . J_r(fg)^{1/p} \le\lVert f\rVert_{\infty}J_r(g)^{1/p}+\lVert g\rVert_{\infty}J_r(f)^{1/p},\qquad f,g \in\mathcal{B}_{\Psi}\cap C_c(X). J r ( f g ) 1/ p ≤ ∥ f ∥ ∞ J r ( g ) 1/ p + ∥ g ∥ ∞ J r ( f ) 1/ p , f , g ∈ B Ψ ∩ C c ( X ) .
Thus B Ψ ∩ C c ( X ) \mathcal{B}_{\Psi}\cap C_c(X) B Ψ ∩ C c ( X ) is a real algebra. The cutoffs make it vanish nowhere and separate points: for distinct x , y x,y x , y , choose U U U with y ∉ U y \notin U y ∈ / U , so that η ( x ) = 1 \eta(x)=1 η ( x ) = 1 and η ( y ) = 0 \eta(y)=0 η ( y ) = 0 . The locally compact real Stone–Weierstrass theorem gives (2.17). Under the hypotheses of Theorem 2.3, the last condition in (4.7) may be replaced by lim inf r ↓ 0 J r ( η ) < ∞ \liminf_{r\downarrow0}J_r(\eta)<\infty lim inf r ↓ 0 J r ( η ) < ∞ , by (2.14).
Corollary 4.4. Under the hypotheses of Theorem 2.2, the core condition (2.17) holds and the energy in Theorem 2.5 may be constructed on the original domain F \mathcal{F} F . Moreover,
E K S ( f ) ≍ E ( f ) , f ∈ F . \mathcal{E}_{\mathrm{KS}}(f)\asymp\mathcal{E}(f),\qquad f\in\mathcal{F}. E KS ( f ) ≍ E ( f ) , f ∈ F .
Proof. Equations (3.39) and (3.40) give B Ψ = F \mathcal{B}_{\Psi}=\mathcal{F} B Ψ = F and equivalence of ∥ ⋅ ∥ B Ψ \lVert\cdot\rVert_{\mathcal{B}_{\Psi}} ∥ ⋅ ∥ B Ψ with the graph norm. Regularity of the original energy implies that F ∩ C c ( X ) \mathcal{F}\cap C_c(X) F ∩ C c ( X ) is uniformly dense in C c ( X ) C_c(X) C c ( X ) and graph-norm dense in F \mathcal{F} F . Hence (2.17) holds, F K S = F \mathcal{F}_{\mathrm{KS}}=\mathcal{F} F KS = F , and the energy equivalence follows from (4.2) and (3.40).
Remark 4.5. Theorem 2.5 gives PI(Ψ \Psi Ψ ) for the reconstructed form, but does not assert cap ( Ψ ) ≤ \operatorname{cap}(\Psi)_{\le} cap ( Ψ ) ≤ . The cutoff criterion in Remark 4.3 gives finite reconstructed energy by (4.2), without the uniform bound V ( x , R ) / Ψ ( R ) V(x,R)/\Psi(R) V ( x , R ) /Ψ ( R ) required by cap ( Ψ ) ≤ \operatorname{cap}(\Psi)_{\le} cap ( Ψ ) ≤ .
Weak-type characterization of the energy domain
The pairwise upper estimate
The next lemma refines the usual dyadic Poincaré telescoping argument by retaining the energy measure on each ball. Compare Yang [44, Lemma 3.4] and, for the point-to-mean step (5.3), Dai–Lin–Yang–Yuan–Zhang [13, arXiv v1, Lemma 2.5].
Lemma 5.1. Assume (VD), (2.2), (PI(Ψ \Psi Ψ )), and LS(β \beta β ) for some β > 0 \beta>0 β > 0 , with constant C L C_L C L . There are A ≥ 1 A\ge1 A ≥ 1 , γ > 0 \gamma>0 γ > 0 , and C < ∞ C<\infty C < ∞ such that, for every f ∈ F f\in\mathcal{F} f ∈ F and for m ⊗ m m\otimes m m ⊗ m -almost every ( x , y ) (x,y) ( x , y ) with R : = d ( x , y ) > 0 R:=d(x,y)>0 R := d ( x , y ) > 0 ,
∣ f ( x ) − f ( y ) ∣ p ≤ C Ψ ( R ) { S f ( x , R ) + S f ( y , R ) } . (73) \lvert f(x)-f(y)\rvert^p \le C\Psi(R)\{S_f(x,R)+S_f(y,R)\}. \tag*{(73)} ∣ f ( x ) − f ( y ) ∣ p ≤ C Ψ ( R ) { S f ( x , R ) + S f ( y , R )} . (73) where
S f ( x , R ) : = ∑ j = 0 ∞ 2 − j γ Γ ( f ) ( B ( x , A 2 − j R ) ) V ( x , 2 − j R ) . (74) S_f(x,R):=\sum_{j=0}^{\infty}2^{-j\gamma}\frac{\Gamma(f)(B(x,A2^{-j}R))}{V(x,2^{-j}R)}. \tag*{(74)} S f ( x , R ) := j = 0 ∑ ∞ 2 − j γ V ( x , 2 − j R ) Γ ( f ) ( B ( x , A 2 − j R )) . (74) One may take γ = β / 2 \gamma=\beta/2 γ = β /2 . The constants depend only on p , C D , C Ψ , C P , A P , C L p,C_D,C_{\Psi},C_P,A_P,C_L p , C D , C Ψ , C P , A P , C L , and β \beta β .
Proof. Use the Lebesgue representative of f f f , which exists almost everywhere by the differentiation theorem for doubling measures. Fix a Lebesgue point x x x , put r j = 2 − j R r_j = 2^{-j}R r j = 2 − j R and B j = B ( x , r j ) B_j = B(x,r_j) B j = B ( x , r j ) . Volume doubling, Hölder’s inequality, and ( P I ( Ψ ) ) (\mathrm{PI}(\Psi)) ( PI ( Ψ )) give
∣ f B j + 1 − f B j ∣ ≤ C ( ∫ B j ∣ f − f B j ∣ p d m ) 1 / p ≤ C Ψ ( r j ) 1 / p ( Γ ( f ) ( A P B j ) V ( x , r j ) ) 1 / p . \left|f_{B_{j+1}}-f_{B_j}\right| \le C\left(\int_{B_j}\left|f-f_{B_j}\right|^p\,\mathrm{d}m\right)^{1/p}
\le C\Psi(r_j)^{1/p}\left(\frac{\Gamma(f)(A_PB_j)}{V(x,r_j)}\right)^{1/p}. f B j + 1 − f B j ≤ C ( ∫ B j f − f B j p d m ) 1/ p ≤ C Ψ ( r j ) 1/ p ( V ( x , r j ) Γ ( f ) ( A P B j ) ) 1/ p .
Since f B j → f ( x ) f_{B_j}\to f(x) f B j → f ( x ) and (16) gives Ψ ( r j ) ≤ C L 2 − j β Ψ ( R ) \Psi(r_j)\le C L 2^{-j\beta}\Psi(R) Ψ ( r j ) ≤ C L 2 − j β Ψ ( R ) ,
∣ f ( x ) − f B ( x , R ) ∣ ≤ C Ψ ( R ) 1 / p ∑ j = 0 ∞ 2 − j β / p ( Γ ( f ) ( A P B j ) V ( x , r j ) ) 1 / p . \left|f(x)-f_{B(x,R)}\right|\le C\Psi(R)^{1/p}\sum_{j=0}^{\infty}2^{-j\beta/p}\left(\frac{\Gamma(f)(A_PB_j)}{V(x,r_j)}\right)^{1/p}. f ( x ) − f B ( x , R ) ≤ C Ψ ( R ) 1/ p j = 0 ∑ ∞ 2 − j β / p ( V ( x , r j ) Γ ( f ) ( A P B j ) ) 1/ p .
Apply Hölder’s inequality to the last series after writing 2 − j β / p = 2 − j β / ( 2 p ) 2 − j β / ( 2 p ) 2^{-j\beta/p}=2^{-j\beta/(2p)}2^{-j\beta/(2p)} 2 − j β / p = 2 − j β / ( 2 p ) 2 − j β / ( 2 p ) . This yields
∣ f ( x ) − f B ( x , R ) ∣ p ≤ C Ψ ( R ) ∑ j = 0 ∞ 2 − j β / 2 Γ ( f ) ( A P B j ) V ( x , r j ) . (75) \left|f(x)-f_{B(x,R)}\right|^p\le C\Psi(R)\sum_{j=0}^{\infty}2^{-j\beta/2}\frac{\Gamma(f)(A_PB_j)}{V(x,r_j)}. \tag*{(75)} f ( x ) − f B ( x , R ) p ≤ C Ψ ( R ) j = 0 ∑ ∞ 2 − j β /2 V ( x , r j ) Γ ( f ) ( A P B j ) . (75)
For Lebesgue points x , y x,y x , y at distance R R R , compare the two means through B ( x , 3 R ) B(x,3R) B ( x , 3 R ) . Applying ( P I ( Ψ ) ) (\mathrm{PI}(\Psi)) ( PI ( Ψ )) on this ball, using nearby-ball volume comparison and the fixed-scale estimate (4) , gives
∣ f B ( x , R ) − f B ( y , R ) ∣ p ≤ C Ψ ( R ) Γ ( f ) ( B ( x , 3 A P R ) ) V ( x , R ) . \left|f_{B(x,R)}-f_{B(y,R)}\right|^p\le C\Psi(R)\frac{\Gamma(f)(B(x,3A_PR))}{V(x,R)}. f B ( x , R ) − f B ( y , R ) p ≤ C Ψ ( R ) V ( x , R ) Γ ( f ) ( B ( x , 3 A P R )) .
Combine this estimate with (75) at x x x and y y y , enlarge A A A by a fixed factor, and use the p p p -triangle inequality. This proves (73) with γ = β / 2 \gamma=\beta/2 γ = β /2 .
Proposition 5.2. Under the hypotheses of Lemma 5.1,
W Ψ ( f ) ≤ C E ( f ) , f ∈ F . (76) \mathcal{W}_{\Psi}(f)\le C\mathcal{E}(f),\qquad f\in\mathcal{F}. \tag*{(76)} W Ψ ( f ) ≤ C E ( f ) , f ∈ F . (76)
The constant depends only on p p p , C D C_D C D , C Ψ C_{\Psi} C Ψ , C P C_P C P , A P A_P A P , C L C_L C L , and β \beta β .
Proof. The anisotropic product metric, dyadic index extraction, and 5-covering argument follow the scheme of Dai–Lin–Yang–Yuan–Zhang [13, arXiv v1, proof of Theorem 2.4, in particular (2.19)–(2.24); see also the covering argument of Brezis–Van Schaftingen–Yung [8 Propositions 2.1–2.2] . The essential modification is that the disjoint first-coordinate balls allow us to sum the possibly singular measure Γ ( f ) \Gamma(f) Γ ( f ) directly.
Fix a Borel representative of f ∈ F f\in\mathcal{F} f ∈ F and λ > 0 \lambda>0 λ > 0 , and write D λ = { K Ψ , f > λ } D_\lambda=\{K_{\Psi,f}>\lambda\} D λ = { K Ψ , f > λ } . By volume doubling,
V ( x , R ) ≍ V ( y , R ) when R = d ( x , y ) > 0. (77) V(x,R)\asymp V(y,R)\quad\text{when }R=d(x,y)>0. \tag*{(77)} V ( x , R ) ≍ V ( y , R ) when R = d ( x , y ) > 0. (77)
Lemma 5.1 shows, up to a null set, that D λ D_\lambda D λ is contained in the union of the two sets
S f ( x , R ) > c λ V ( x , R ) , S f ( y , R ) > c λ V ( y , R ) . S_f(x,R)>c\lambda V(x,R),\qquad S_f(y,R)>c\lambda V(y,R). S f ( x , R ) > c λV ( x , R ) , S f ( y , R ) > c λV ( y , R ) .
Interchanging x x x and y y y gives the same estimate for the second set, so it suffices to consider the first.
Put δ = γ / 2 \delta=\gamma/2 δ = γ /2 . If S f ( x , R ) > c λ V ( x , R ) S_f(x,R)>c\lambda V(x,R) S f ( x , R ) > c λV ( x , R ) , then for some j ≥ 0 j\ge0 j ≥ 0 ,
Γ ( f ) ( B ( x , A 2 − j R ) ) > c 1 λ 2 j δ V ( x , 2 − j R ) V ( x , R ) . (78) \Gamma(f)(B(x,A2^{-j}R))>c_1\lambda2^{j\delta}V(x,2^{-j}R)V(x,R). \tag*{(78)} Γ ( f ) ( B ( x , A 2 − j R )) > c 1 λ 2 j δ V ( x , 2 − j R ) V ( x , R ) . (78)
Indeed, if (78) failed for every j j j , then (74) would give S f ( x , R ) ≤ c λ V ( x , R ) S_f(x,R)\le c\lambda V(x,R) S f ( x , R ) ≤ c λV ( x , R ) once c 1 c_1 c 1 is chosen sufficiently small, since ∑ j ≥ 0 2 − j ( γ − δ ) < ∞ \sum_{j\ge0}2^{-j(\gamma-\delta)}<\infty ∑ j ≥ 0 2 − j ( γ − δ ) < ∞ . The measurability observation in Subsection 2.1, applied to Γ ( f ) \Gamma(f) Γ ( f ) and m m m , shows that the active-pair sets below are Borel.
For j ≥ 0 j \ge0 j ≥ 0 let E λ , j E_{\lambda,j} E λ , j be the set of pairs ( x , y ) (x,y) ( x , y ) , with R = d ( x , y ) > 0 R=d(x,y)>0 R = d ( x , y ) > 0 , for which (5.6) holds. We claim that
( m ⊗ m ) ( E λ , j ) ≤ C 2 − j δ λ E ( f ) . (79) (m \otimes m)(E_{\lambda,j}) \le\frac{C2^{-j\delta}}{\lambda}\mathcal{E}(f). \tag*{(79)} ( m ⊗ m ) ( E λ , j ) ≤ λ C 2 − j δ E ( f ) . (79)
To prove the claim, fix o ∈ X o \in X o ∈ X and first restrict x , y x,y x , y to B ( o , N ) B(o,N) B ( o , N ) . Choose C 0 > max { A , 1 } C_0 > \max\{A,1\} C 0 > max { A , 1 } and equip X × X X \times X X × X with the metric
ρ j ( ( u , v ) , ( u ′ , v ′ ) ) : = max { 2 j A − 1 d ( u , u ′ ) , C 0 − 1 d ( v , v ′ ) } . \rho_j((u,v),(u',v')) := \max\{2^j A^{-1}d(u,u'), C_0^{-1}d(v,v')\}. ρ j (( u , v ) , ( u ′ , v ′ )) := max { 2 j A − 1 d ( u , u ′ ) , C 0 − 1 d ( v , v ′ )} .
For every active pair ( x , y ) (x,y) ( x , y ) let B x , y B_{x,y} B x , y be the ρ j \rho_j ρ j -ball with center ( x , x ) (x,x) ( x , x ) and radius R = d ( x , y ) R=d(x,y) R = d ( x , y ) . Thus
B x , y = B ( x , A 2 − j R ) × B ( x , C 0 R ) . B_{x,y}=B(x,A2^{-j}R)\times B(x,C_0R). B x , y = B ( x , A 2 − j R ) × B ( x , C 0 R ) .
The active pair ( x , y ) (x,y) ( x , y ) belongs to this product ball, so these balls cover the active set. Their radii are bounded by 2 N 2N 2 N . The countable 5-covering lemma in the separable metric space ( X × X , ρ j ) (X \times X,\rho_j) ( X × X , ρ j ) gives a countable disjoint subfamily { B i } \{B_i\} { B i } whose fivefold dilates cover all the active pairs in B ( o , N ) 2 B(o,N)^2 B ( o , N ) 2 ; see [21 Theorem 1.2] . Fixed-ratio volume comparison and (5.6) imply
( m ⊗ m ) ( 5 B i ) ≤ C V ( x i , 2 − j R i ) V ( x i , R i ) ≤ C 2 − j δ λ Γ ( f ) ( B ( x i , A 2 − j R i ) ) . (m \otimes m)(5B_i) \le CV(x_i,2^{-j}R_i)V(x_i,R_i) \le\frac{C2^{-j\delta}}{\lambda}\Gamma(f)(B(x_i,A2^{-j}R_i)). ( m ⊗ m ) ( 5 B i ) ≤ C V ( x i , 2 − j R i ) V ( x i , R i ) ≤ λ C 2 − j δ Γ ( f ) ( B ( x i , A 2 − j R i )) .
The first-coordinate balls B ( x i , A 2 − j R i ) B(x_i,A2^{-j}R_i) B ( x i , A 2 − j R i ) are pairwise disjoint. Indeed, each is contained in the corresponding second-coordinate ball, and a common point in two first-coordinate balls would put its diagonal pair in both product balls. Consequently their Γ ( f ) \Gamma(f) Γ ( f ) -masses sum to at most E ( f ) \mathcal{E}(f) E ( f ) . We obtain (5.7) on B ( o , N ) 2 B(o,N)^2 B ( o , N ) 2 ; monotone convergence as N → ∞ N \to\infty N → ∞ proves the claim.
Sum (5.7) over j j j . The geometric series converges, and the same argument applies to the set with x , y x,y x , y interchanged. Hence
( m ⊗ m ) ( D λ ) ≤ C λ E ( f ) . (m \otimes m)(D_\lambda) \le\frac{C}{\lambda}\mathcal{E}(f). ( m ⊗ m ) ( D λ ) ≤ λ C E ( f ) .
Taking the supremum over λ \lambda λ proves (5.4).
Recovery from the high-level pairwise tail
Lemma 5.3. Assume (VD) and LS(β \beta β ) for some β > 0 \beta>0 β > 0 , with constant C L C_L C L . Let f ∈ L p ( X ; m ) f \in L^p(X;m) f ∈ L p ( X ; m ) with W Ψ , ∞ ( f ) < ∞ \mathbb{W}_{\Psi,\infty}(f)<\infty W Ψ , ∞ ( f ) < ∞ . For k > 0 k>0 k > 0 put f k = ( − k ) ∨ ( f ∧ k ) f_k=(-k)\vee(f\wedge k) f k = ( − k ) ∨ ( f ∧ k ) . Then for every metric ball U ⊂ X U\subset X U ⊂ X ,
lim inf r ↓ 0 J r ( f k ; U ) ≤ C W Ψ , ∞ ( f ) , (80) \liminf_{r\downarrow0} J_r(f_k;U) \le C\mathbb{W}_{\Psi,\infty}(f), \tag*{(80)} r ↓ 0 lim inf J r ( f k ; U ) ≤ C W Ψ , ∞ ( f ) , (80) where J r ( ⋅ ; U ) J_r(\mathord{\cdot};U) J r ( ⋅ ; U ) is the localized functional from Lemma 4.1. The constant C C C depends only on C D C_D C D , C L C_L C L , and β \beta β , and is independent of U U U , k k k , and f f f .
Proof. Write g = f k g=f_k g = f k and fix 0 < ε < δ 0<\varepsilon<\delta 0 < ε < δ , with δ < diam X \delta<\operatorname{diam}X δ < diam X when diam X < ∞ \operatorname{diam}X<\infty diam X < ∞ . The following logarithmic averaging over scales is the ball-kernel analogue of the high-level tail arguments in [7 Section 4] and [39 Theorem 1.5 and Section 2] ; compare also [24 Proposition 1.3] for a localized BBM characterization on Poincaré spaces. Let d Ψ d\Psi d Ψ denote the Lebesgue–Stieltjes measure of the continuous increasing function Ψ \Psi Ψ . Tonelli’s theorem gives
I ε , δ : = ∫ ( ε , δ ] J r ( g ; U ) d Ψ ( r ) Ψ ( r ) = ∬ X × X L ε , δ ( x , y ) d m ( y ) d m ( x ) , (81) \begin{aligned}
I_{\varepsilon,\delta}:=\int_{(\varepsilon,\delta]}J_r(g;U)\frac{d\Psi(r)}{\Psi(r)} \\
&=\iint_{X\times X}L_{\varepsilon,\delta}(x,y)\,dm(y)\,dm(x), \tag*{(81)}
\end{aligned} I ε , δ := ∫ ( ε , δ ] J r ( g ; U ) Ψ ( r ) d Ψ ( r ) = ∬ X × X L ε , δ ( x , y ) d m ( y ) d m ( x ) , (81)
where, with a = max { ε , d ( x , y ) } a=\max\{\varepsilon,d(x,y)\} a = max { ε , d ( x , y )} ,
L ε , δ ( x , y ) : = 1 U ( x ) 1 { d ( x , y ) < δ } ∣ g ( x ) − g ( y ) ∣ p ∫ ( a , δ ] d Ψ ( r ) Ψ ( r ) 2 V ( x , r ) . (82) L_{\varepsilon,\delta}(x,y):=\mathbf{1}_{U}(x)\mathbf{1}_{\{d(x,y)<\delta\}}|g(x)-g(y)|^{p}\int_{(a,\delta]}\frac{\mathrm{d}\Psi(r)}{\Psi(r)^{2}V(x,r)}. \tag*{(82)} L ε , δ ( x , y ) := 1 U ( x ) 1 { d ( x , y ) < δ } ∣ g ( x ) − g ( y ) ∣ p ∫ ( a , δ ] Ψ ( r ) 2 V ( x , r ) d Ψ ( r ) . (82)
The integral is understood as zero when a ≥ δ a\ge\delta a ≥ δ . Monotonicity of r ↦ V ( x , r ) r\mapsto V(x,r) r ↦ V ( x , r ) and the identity
∫ ( a , δ ] d Ψ ( r ) Ψ ( r ) 2 = 1 Ψ ( a ) − 1 Ψ ( δ ) \int_{(a,\delta]}\frac{\mathrm{d}\Psi(r)}{\Psi(r)^{2}}=\frac{1}{\Psi(a)}-\frac{1}{\Psi(\delta)} ∫ ( a , δ ] Ψ ( r ) 2 d Ψ ( r ) = Ψ ( a ) 1 − Ψ ( δ ) 1
give
0 ≤ L ε , δ ≤ K Ψ , g ≤ K Ψ , f . (83) 0\le L_{\varepsilon,\delta}\le K_{\Psi,g}\le K_{\Psi,f}. \tag*{(83)} 0 ≤ L ε , δ ≤ K Ψ , g ≤ K Ψ , f . (83)
The last inequality follows because truncation is 1-Lipschitz.
The kernel L ε , δ L_{\varepsilon,\delta} L ε , δ vanishes outside U × U δ U\times U_{\delta} U × U δ , a set of finite m ⊗ m m\otimes m m ⊗ m -measure. To quantify the lower volume bound, choose o o o and R 0 > 0 R_{0}>0 R 0 > 0 with U ⊂ B ( o , R 0 ) U\subset B(o,R_{0}) U ⊂ B ( o , R 0 ) . Put ε U : = min { 1 , 2 R 0 } \varepsilon_{U}:=\min\{1,2R_{0}\} ε U := min { 1 , 2 R 0 } . Then B ( o , R 0 ) ⊂ B ( x , 2 R 0 ) B(o,R_{0})\subset B(x,2R_{0}) B ( o , R 0 ) ⊂ B ( x , 2 R 0 ) for x ∈ U x\in U x ∈ U , and iterating (VD) backwards from 2 R 0 2R_{0} 2 R 0 to ε \varepsilon ε gives
inf x ∈ U V ( x , ε ) ≥ c U ε Q D , 0 < ε < ε U , \inf_{x\in U}V(x,\varepsilon)\ge c_{U}\varepsilon^{Q_{D}},\qquad0<\varepsilon<\varepsilon_{U}, x ∈ U inf V ( x , ε ) ≥ c U ε Q D , 0 < ε < ε U ,
where Q D : = log 2 C D Q_{D}:=\log_{2}C_{D} Q D := log 2 C D . The constant c U > 0 c_{U}>0 c U > 0 absorbs the rounding of the number of iterations. Since ∣ g ∣ ≤ k |g|\le k ∣ g ∣ ≤ k , (5.10) therefore yields
∥ L ε , δ ∥ ∞ ≤ C U , k Ψ ( ε ) ε Q D = : M ε . (84) \|L_{\varepsilon,\delta}\|_{\infty}\le\frac{C_{U,k}}{\Psi(\varepsilon)\varepsilon^{Q_{D}}}=:M_{\varepsilon}. \tag*{(84)} ∥ L ε , δ ∥ ∞ ≤ Ψ ( ε ) ε Q D C U , k =: M ε . (84)
Fix b > W Ψ , ∞ ( f ) b>\mathcal{W}_{\Psi,\infty}(f) b > W Ψ , ∞ ( f ) . There is 0 < λ 0 < ∞ 0<\lambda_{0}<\infty 0 < λ 0 < ∞ such that
( m ⊗ m ) ( { K Ψ , f > λ } ) ≤ b λ , λ ≥ λ 0 . (m\otimes m)(\{K_{\Psi,f}>\lambda\})\le\frac{b}{\lambda},\qquad\lambda\ge\lambda_{0}. ( m ⊗ m ) ({ K Ψ , f > λ }) ≤ λ b , λ ≥ λ 0 .
Write log + t : = max { log t , 0 } \log^{+}t:=\max\{\log t,0\} log + t := max { log t , 0 } for t > 0 t>0 t > 0 . The layer-cake formula, (5.11), and (5.12) show that
I ε , δ ≤ λ 0 ( m ⊗ m ) ( U × U δ ) + b log + M ε λ 0 . (85) I_{\varepsilon,\delta}\le\lambda_{0}(m\otimes m)(U\times U_{\delta})+b\log^{+}\frac{M_{\varepsilon}}{\lambda_{0}}. \tag*{(85)} I ε , δ ≤ λ 0 ( m ⊗ m ) ( U × U δ ) + b log + λ 0 M ε . (85)
On the other hand,
( inf ε < r ≤ δ J r ( g ; U ) ) log Ψ ( δ ) Ψ ( ε ) ≤ I ε , δ . (86) \left(\inf_{\varepsilon<r\le\delta}J_{r}(g;U)\right)\log\frac{\Psi(\delta)}{\Psi(\varepsilon)}\le I_{\varepsilon,\delta}. \tag*{(86)} ( ε < r ≤ δ inf J r ( g ; U ) ) log Ψ ( ε ) Ψ ( δ ) ≤ I ε , δ . (86)
The lower scaling assumption implies, for fixed δ \delta δ ,
log Ψ ( δ ) Ψ ( ε ) ≥ β log δ ε − log C L . \log\frac{\Psi(\delta)}{\Psi(\varepsilon)}\ge\beta\log\frac{\delta}{\varepsilon}-\log C_{L}. log Ψ ( ε ) Ψ ( δ ) ≥ β log ε δ − log C L .
Together with (5.12), this gives
lim sup ε ↓ 0 log + ( M ε / λ 0 ) log ( Ψ ( δ ) / Ψ ( ε ) ) ≤ 1 + Q D β . (87) \limsup_{\varepsilon\downarrow0}\frac{\log^{+}(M_{\varepsilon}/\lambda_{0})}{\log(\Psi(\delta)/\Psi(\varepsilon))}\le1+\frac{Q_{D}}{\beta}. \tag*{(87)} ε ↓ 0 lim sup log ( Ψ ( δ ) /Ψ ( ε )) log + ( M ε / λ 0 ) ≤ 1 + β Q D . (87)
Continuity of Ψ \Psi Ψ makes d Ψ \mathrm{d}\Psi d Ψ atomless, so no endpoint terms arise. The dependence on U U U , k k k , λ 0 \lambda_{0} λ 0 , and the fixed δ \delta δ occurs only in the finite first term of (5.13) or in additive logarithmic terms. These contributions vanish after division by log ( Ψ ( δ ) / Ψ ( ε ) ) \log(\Psi(\delta)/\Psi(\varepsilon)) log ( Ψ ( δ ) /Ψ ( ε )) and passage to the limit. Divide (5.13) by the logarithm in (5.14) and let ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 . Since
lim ε ↓ 0 inf ε < r ≤ δ J r ( g ; U ) = inf 0 < r ≤ δ J r ( g ; U ) , \lim_{\varepsilon\downarrow0}\inf_{\varepsilon<r\le\delta}J_{r}(g;U)=\inf_{0<r\le\delta}J_{r}(g;U), ε ↓ 0 lim ε < r ≤ δ inf J r ( g ; U ) = 0 < r ≤ δ inf J r ( g ; U ) ,
letting next δ ↓ 0 \delta\downarrow0 δ ↓ 0 gives lim inf r ↓ 0 J r ( g ; U ) \liminf_{r\downarrow0}J_{r}(g;U) lim inf r ↓ 0 J r ( g ; U ) . Finally, letting b ↓ W Ψ , ∞ ( f ) b\downarrow\mathcal{W}_{\Psi,\infty}(f) b ↓ W Ψ , ∞ ( f ) proves (5.8).
Lemma 5.4. Assume the hypotheses of Theorem 2.2 . Let f ∈ L p ( X ; m ) f \in L^{p}(X;m) f ∈ L p ( X ; m ) , L < ∞ L < \infty L < ∞ , and f k = ( − k ) ∨ ( f ∧ k ) f_k = (-k) \vee(f \wedge k) f k = ( − k ) ∨ ( f ∧ k ) . If, for every ball U U U and every k > 0 k > 0 k > 0 ,
lim inf r ↓ 0 ∫ U M f k ( x , r ) d m ( x ) ≤ L , (88) \liminf_{r \downarrow0} \int_U M f_k(x,r)\,dm(x) \leq L, \tag*{(88)} r ↓ 0 lim inf ∫ U M f k ( x , r ) d m ( x ) ≤ L , (88) then f ∈ F f \in\mathcal{F} f ∈ F and E ( f ) ≤ C L \mathcal{E}(f) \leq CL E ( f ) ≤ C L . The constant C C C depends only on p p p , C D C_D C D , C Ψ C_\Psi C Ψ , C P C_P C P , A P A_P A P , C c C_c C c , and A c A_c A c .
Proof. Fix x 0 ∈ X x_0 \in X x 0 ∈ X and integrate ( m P I ( Ψ ) ) (\mathrm{mPI}(\Psi)) ( mPI ( Ψ )) over B ( x 0 , R ) B(x_0,R) B ( x 0 , R ) . Tonelli’s theorem and nearby-ball volume comparison give, for 0 < r < s 0 < r < s 0 < r < s ,
∫ B ( x 0 , R ) M f k ( x , s ) d m ( x ) ≤ C ∫ B ( x 0 , R + A M s ) M f k ( y , r ) d m ( y ) . \int_{B(x_0,R)} M f_k(x,s)\,dm(x) \leq C \int_{B(x_0,R+A_Ms)} M f_k(y,r)\,dm(y). ∫ B ( x 0 , R ) M f k ( x , s ) d m ( x ) ≤ C ∫ B ( x 0 , R + A M s ) M f k ( y , r ) d m ( y ) .
Take the lower limit as r ↓ 0 r \downarrow0 r ↓ 0 , use (88) , and let R ↑ ∞ R \uparrow\infty R ↑ ∞ . Thus O s ( f k ) ≤ C L \mathcal{O}_s(f_k) \leq CL O s ( f k ) ≤ C L for every s , k > 0 s,k > 0 s , k > 0 . Lemma 3.11 gives sup s J s ( f k ) ≤ C L \sup_s J_s(f_k) \leq CL sup s J s ( f k ) ≤ C L ; Fatou’s lemma in (12) gives the same bound for f f f . The domain identification (64) and the energy comparison (65) complete the proof. □
Proof of Theorem 2.6. By Corollary 3.12, Ψ \Psi Ψ satisfies LS ( p ) \operatorname{LS}(p) LS ( p ) . Proposition 5.2 gives
W Ψ , ∞ ( f ) ≤ W Ψ ( f ) ≤ C E ( f ) , f ∈ F . \mathcal{W}_{\Psi,\infty}(f) \leq\mathcal{W}_{\Psi}(f) \leq C\mathcal{E}(f), \qquad f \in\mathcal{F}. W Ψ , ∞ ( f ) ≤ W Ψ ( f ) ≤ C E ( f ) , f ∈ F .
Conversely, let f ∈ L p ( X ; m ) f \in L^{p}(X;m) f ∈ L p ( X ; m ) with W Ψ , ∞ ( f ) < ∞ \mathcal{W}_{\Psi,\infty}(f) < \infty W Ψ , ∞ ( f ) < ∞ . Put L = W Ψ , ∞ ( f ) L = \mathcal{W}_{\Psi,\infty}(f) L = W Ψ , ∞ ( f ) and f k = ( − k ) ∨ ( f ∧ k ) f_k = (-k) \vee(f \wedge k) f k = ( − k ) ∨ ( f ∧ k ) . Fix a ball U U U and a larger ball U ~ \widetilde{U} U containing U r U_r U r for all sufficiently small r r r . The localized comparison (68) and Lemma 5.3 give
lim inf r ↓ 0 ∫ U M f k ( x , r ) d m ( x ) ≤ C lim inf r ↓ 0 J 2 r ( f k ; U ~ ) ≤ C L , \liminf_{r \downarrow0} \int_U M f_k(x,r)\,dm(x) \leq C\liminf_{r \downarrow0} J_{2r}(f_k;\widetilde{U}) \leq CL, r ↓ 0 lim inf ∫ U M f k ( x , r ) d m ( x ) ≤ C r ↓ 0 lim inf J 2 r ( f k ; U ) ≤ C L ,
uniformly in U U U and k k k . Lemma 5.4 gives f ∈ F f \in\mathcal{F} f ∈ F and
E ( f ) ≤ C L = C W Ψ , ∞ ( f ) . (89) \mathcal{E}(f) \leq CL = C\mathcal{W}_{\Psi,\infty}(f). \tag*{(89)} E ( f ) ≤ C L = C W Ψ , ∞ ( f ) . (89)
Combining these estimates proves (24) and (25) . □
Acknowledgments
F.B. is partially supported by grant 10.46540/4283-00175B from the Independent Research Fund Denmark, by the Villum Investigator grant Stochastic Analysis in Aarhus , and by the European Research Council (ERC) under the European Union’s Horizon Europe research and innovation programme. The ERC support is provided through the RanGe project (Grant Agreement No. 101199772).
Declaration on the use of generative AI
The authors are responsible for the mathematical ideas and the structure of the paper. ChatGPT (Astra) and Claude (Fable) were used to assist with proofreading, proof presentation, editorial revision, and bibliographical research. The authors reviewed and edited the resulting content and take full responsibility for the publication.