Introduction
In this paper, we study the asymptotics via Γ \Gamma Γ -convergence of a general class of non-local diffuse interface models for phase transitions with infinitely many stable states. Our prototype example is the functional
1 ∣ log ε ∣ ∬ Q × Q ∣ u ( x ) − u ( y ) ∣ 2 ∣ x − y ∣ n + 1 d x d y + 1 ε ∣ log ε ∣ ∫ Q W ε ( u ( x ) ) d x (1) \frac{1}{|\log\varepsilon|}\iint_{Q\times Q}\frac{|u(x)-u(y)|^2}{|x-y|^{n+1}}\,dx\,dy+\frac{1}{\varepsilon|\log\varepsilon|}\int_Q W_\varepsilon(u(x))\,dx \tag*{(1)} ∣ log ε ∣ 1 ∬ Q × Q ∣ x − y ∣ n + 1 ∣ u ( x ) − u ( y ) ∣ 2 d x d y + ε ∣ log ε ∣ 1 ∫ Q W ε ( u ( x )) d x (1)
where n ∈ N ∖ { 0 } n \in\mathbb{N} \setminus\{0\} n ∈ N ∖ { 0 } and Q Q Q is a cube in R n \mathbb{R}^n R n . The first term is the square of the H 1 / 2 H^{1/2} H 1/2 Gagliardo seminorm, and the second one is a potential that vanishes exactly on the integers, i.e. { t : W ε ( t ) = 0 } = Z \{t: W_\varepsilon(t)=0\}=\mathbb{Z} { t : W ε ( t ) = 0 } = Z .
This energy shows the typical structure of variational models for phase transitions. It consists of a singular perturbation of a multi-well potential (the prototype being the Cahn–Hilliard functional [11 ] ) whose global minima represent the pure phases, which are the energetically preferred configurations of the system. After rescaling the Cahn–Hilliard models one obtains the Modica–Mortola functional (see [22 , 23 ] ) and in the limit recovers a sharp interface model. In the limit regime, only pure phases are allowed, and the effective energy concentrates on the jump set: the interface that separates different phases.
This variational scheme applies to a wide range of problems in physics and materials science. An extensive body of literature has explored variants of the classical Modica–Mortola functional to account for anisotropy (see, e.g., [24 , 9 ] ), multiphase vectorial systems [8 ] , and homogenization in perforated or heterogeneous media (see [2 , 12 ] , [14 ] and references therein). Parallel to these gradient-based developments, a significant line of research has focused on phase transition models with non-local regularizations, where the gradient term is replaced by convolution-type [1 ] or fractional interaction kernels ([4 ] ). This framework has been further extended from an analytical perspective to phase transitions governed by arbitrary fractional operators, such as Gagliardo-type seminorms ([27 , 25 ] ).
In this setting one can also formulate the classical Nabarro–Peierls energy for dislocations (see [18 ] ), leading to the development of phase-field models for dislocations driven by non-local H 1 / 2 H^{1/2} H 1/2 -type regularizations, as in (1) , with W ε ( t ) = dist 2 ( t , Z ) W_{\varepsilon}(t)=\operatorname{dist}^{2}(t,\mathbb{Z}) W ε ( t ) = dist 2 ( t , Z ) . Here the periodic potential reflects the underlying crystalline structure: the order parameter u : Q → R u:Q\to\mathbb{R} u : Q → R describes the plastic slips of the form u e 1 u e_{1} u e 1 , while the H 1 / 2 H^{1/2} H 1/2 seminorm represents the elastic energy induced by the slip u u u along a single slip plane. In the limit this model produces a line-tension model for crystal plasticity (see, e.g., [15 ] , see also [19 , 20 ] for the case n = 1 n=1 n = 1 with applications to boundary vortices). Generalization of this model to vector-valued slips and to the case of multiple slip planes can be found in [10 ] .
In this paper, motivated by the study of partial dislocations and the emergence of stacking faults in crystals, we introduce a novel mathematical ingredient into this framework. Specifically, we analyze the asymptotics of a singular, non-local perturbation of a multi-well potential - in the spirit of the aforementioned phase-field models for dislocations [15 ] . A key feature of our model is that the potential exhibits stable local minima. These minima represent “intermediate” metastable states that are energetically permissible, though less favorable than the global ones. In the mechanics literature, the energy profile mapping these configurations is known as the stacking fault energy or γ \gamma γ -surface [Vit68].
For concreteness, we consider a potential W ε : R → [ 0 , + ∞ ) W_{\varepsilon}:\mathbb{R}\to[0,+\infty) W ε : R → [ 0 , + ∞ ) of the form
W ε ( u ) : = min { dist 2 ( u , Z ) , dist 2 ( u , Z + 1 2 ) + δ ε } , (2) W_{\varepsilon}(u):=\min\left\{\operatorname{dist}^{2}(u,\mathbb{Z}),\operatorname{dist}^{2}\left(u,\mathbb{Z}+\frac{1}{2}\right)+\delta_{\varepsilon}\right\}, \tag*{(2)} W ε ( u ) := min { dist 2 ( u , Z ) , dist 2 ( u , Z + 2 1 ) + δ ε } , (2)
where { δ ε } ⊂ R + \{\delta_{\varepsilon}\}\subset\mathbb{R}^{+} { δ ε } ⊂ R + represents the energy gap (or “cost”) associated with the secondary wells. The effect of intermediate wells depends on how fast the gap δ ε \delta_{\varepsilon} δ ε vanishes as ε → 0 \varepsilon\to0 ε → 0 . Intuitively, the asymptotic height of the intermediate well determines whether the limiting energy allows partial slips and, if so, at what energetic cost. The goal of this work is to identify the relevant asymptotic regimes corresponding to different limiting behaviors. In particular, in the case δ ε = ε ∣ log ε ∣ \delta_{\varepsilon}=\varepsilon|\log\varepsilon| δ ε = ε ∣ log ε ∣ , we prove that the limit energy of (1) , choosing W ε W_{\varepsilon} W ε as in (2) , is
ω n − 1 ∫ J u ∣ u + − u − ∣ d H n − 1 + ∣ { u ∈ Z + 1 2 } ∣ for u ∈ B V ( Q ; 1 2 Z ) , \omega_{n-1}\int_{J_{u}}|u^{+}-u^{-}|\,\mathrm{d}\mathcal{H}^{n-1}+\left|\left\{u\in\mathbb{Z}+\frac{1}{2}\right\}\right|
\qquad\text{for }u\in BV\left(Q;\frac{1}{2}\mathbb{Z}\right), ω n − 1 ∫ J u ∣ u + − u − ∣ d H n − 1 + { u ∈ Z + 2 1 } for u ∈ B V ( Q ; 2 1 Z ) ,
where ω n − 1 \omega_{n-1} ω n − 1 is the volume of the unit ball in R n − 1 \mathbb{R}^{n-1} R n − 1 (ω 0 = 1 \omega_{0}=1 ω 0 = 1 ). Therefore, in this regime, the permanence in intermediate states has a cost which is captured by the limit energy and shows the emergence of the stacking fault term, which, in the context of dislocations, represents an energy concentrated on the regions of the slip plane separating partial dislocation lines. This is the main result of this paper.
It is worth noting that variational models involving partial defects have attracted significant interest in recent years. On one hand models similar in spirit to the one proposed here have been considered in the discrete setting for two-dimensional systems (see [7 ] for a generalized XY model and [6 ] for a next-to-nearest neighbor lattice model for partial edge dislocations) where the asymptotic energy concentrates on fractional singularities (vortices or dislocations) connected by line-like string defects. On the other hand, several continuum models share a similar analytical structure, where the coupling between fractional point singularities and line discontinuities is topologically induced, leading to a structure where integer-degree defects split into fractional components connected by penalized interfaces (see [16 ] ).
For completeness we study the asymptotics in terms of Γ \Gamma Γ -convergence of a more general problem in which we consider all possible regimes for the parameter δ ε \delta_\varepsilon δ ε , we assume that the local minima may have a more general structure (not necessarily at half integers) and we also include a more general non-local regularization which covers all fractional seminorms H s H^s H s .
Formulation of the problem and main result
Let Q ⊂ R n Q \subset\mathbb{R}^n Q ⊂ R n be an n n n -dimensional cube with n ≥ 1 n \ge1 n ≥ 1 . For s ∈ [ 1 2 , 1 ) s \in[\frac{1}{2},1) s ∈ [ 2 1 , 1 ) , we investigate the asymptotic behavior, as ε \varepsilon ε goes to 0, of the family of functionals F ε : L 1 ( Q ) → [ 0 , + ∞ ] F_\varepsilon\colon L^1(Q) \to[0,+\infty] F ε : L 1 ( Q ) → [ 0 , + ∞ ] defined as follows:
F ε ( u ) : = { 1 θ ε ( ε 2 s ∫ Q ∫ Q ∣ u ( x ) − u ( y ) ∣ 2 ∣ x − y ∣ n + 2 s d x d y + ∫ Q W ε ( u ( x ) ) d x ) , if u ∈ H s ( Q ) , + ∞ , otherwise , (3) F_\varepsilon(u) :=
\begin{cases}
\frac{1}{\theta_\varepsilon}\left(\varepsilon^{2s}\int_Q\int_Q \frac{|u(x)-u(y)|^2}{|x-y|^{n+2s}}\,dx\,dy+\int_Q W_\varepsilon(u(x))\,dx\right), & \text{if }u\in H^s(Q),\\
+\infty, & \text{otherwise},
\end{cases} \tag*{(3)} F ε ( u ) := { θ ε 1 ( ε 2 s ∫ Q ∫ Q ∣ x − y ∣ n + 2 s ∣ u ( x ) − u ( y ) ∣ 2 d x d y + ∫ Q W ε ( u ( x )) d x ) , + ∞ , if u ∈ H s ( Q ) , otherwise , (3)
where the scaling factor θ ε \theta_\varepsilon θ ε is chosen as in [26 ] :
θ ε : = { ε ∣ log ε ∣ , if s = 1 2 , ε , if s ∈ ( 1 2 , 1 ) . \theta_\varepsilon:=
\begin{cases}
\varepsilon|\log\varepsilon|, & \text{if }s=\frac{1}{2},\\
\varepsilon, & \text{if }s\in\left(\frac{1}{2},1\right).
\end{cases} θ ε := { ε ∣ log ε ∣ , ε , if s = 2 1 , if s ∈ ( 2 1 , 1 ) .
The non-local term represents the square of the Gagliardo H s H^s H s -seminorm of u u u in Q Q Q . For the sake of brevity, we shall denote it simply by
[ u ] H s ( Q ) 2 : = ∫ Q ∫ Q ∣ u ( x ) − u ( y ) ∣ 2 ∣ x − y ∣ n + 2 s d x d y . [u]^2_{H^s(Q)} := \int_Q\int_Q \frac{|u(x)-u(y)|^2}{|x-y|^{n+2s}}\,dx\,dy. [ u ] H s ( Q ) 2 := ∫ Q ∫ Q ∣ x − y ∣ n + 2 s ∣ u ( x ) − u ( y ) ∣ 2 d x d y .
The potential W ε : R → [ 0 , + ∞ ) W_\varepsilon\colon\mathbb{R} \to[0,+\infty) W ε : R → [ 0 , + ∞ ) is designed to model a system with multiple hierarchical states. It vanishes on Z \mathbb{Z} Z and has local minima at S : = { s z } z ∈ Z S := \{s_z\}_{z\in\mathbb{Z}} S := { s z } z ∈ Z representing partial slips or intermediate stable configurations. To ensure that the wells are well-separated, we assume that there exists a constant d ∈ ( 0 , 1 2 ] d \in(0,\frac{1}{2}] d ∈ ( 0 , 2 1 ] such that:
z + d ≤ s z ≤ z + 1 − d for every z ∈ Z . (4) z+d\le s_z\le z+1-d \qquad\text{for every }z\in\mathbb{Z}. \tag*{(4)} z + d ≤ s z ≤ z + 1 − d for every z ∈ Z . (4)
The energetic cost at these intermediate configurations is δ ε ≥ 0 \delta_\varepsilon\ge0 δ ε ≥ 0 with
lim ε → 0 δ ε = 0. \lim_{\varepsilon\to0}\delta_\varepsilon=0. ε → 0 lim δ ε = 0.
For concreteness, the potential W ε W_\varepsilon W ε is defined by
W ε ( u ) : = min { dist 2 ( u , Z ) , dist 2 ( u , S ) + δ ε } . W_\varepsilon(u):=\min\left\{\operatorname{dist}^2(u,\mathbb{Z}),\operatorname{dist}^2(u,S)+\delta_\varepsilon\right\}. W ε ( u ) := min { dist 2 ( u , Z ) , dist 2 ( u , S ) + δ ε } .
The rate at which δ ε \delta_\varepsilon δ ε vanishes plays a crucial role in our analysis giving rise to different variational regimes. We study how the competition between the fractional interaction term and the penalization of the intermediate wells through δ ε \delta_\varepsilon δ ε determines both the set of admissible limit values (the state space) and the form of the Γ \Gamma Γ -limit.
We distinguish two different regimes. We assume that the relevant limits exist, and classify the regimes as follows: (R1) Supercritical/Critical regime: We are in the supercritical or critical regime if
Figure 1. Example of potential W ε W_{\varepsilon} W ε
K : = lim ε → 0 δ ε θ ε ∈ [ 0 , + ∞ ) . (5) K := \lim_{\varepsilon\to0} \frac{\delta_{\varepsilon}}{\theta_{\varepsilon}} \in[0,+\infty). \tag*{(5)} K := ε → 0 lim θ ε δ ε ∈ [ 0 , + ∞ ) . (5)
We refer to the case K = 0 K = 0 K = 0 as the supercritical regime, and to the case K ∈ ( 0 , + ∞ ) K \in(0,+\infty) K ∈ ( 0 , + ∞ ) as the critical one.
(R2) Subcritical regime: We are in the subcritical regime if
lim ε → 0 δ ε θ ε = + ∞ . (6) \lim_{\varepsilon\to0} \frac{\delta_{\varepsilon}}{\theta_{\varepsilon}} = +\infty. \tag*{(6)} ε → 0 lim θ ε δ ε = + ∞. (6)
In the case s = 1 2 s = \frac{1}{2} s = 2 1 and regime (R2), we additionally assume
ℓ : = lim ε → 0 ∣ log δ ε ∣ ∣ log ε ∣ ∈ [ 0 , 1 ] . (7) \ell:= \lim_{\varepsilon\to0} \frac{|\log\delta_{\varepsilon}|}{|\log\varepsilon|} \in[0,1]. \tag*{(7)} ℓ := ε → 0 lim ∣ log ε ∣ ∣ log δ ε ∣ ∈ [ 0 , 1 ] . (7)
As we shall see, these regimes lead to different target spaces for the limit functions. In regime (R1), the intermediate wells are “cheap” enough to be visited in the limit. In this case, the state space of the limit function is B V ( Q ; Z ∪ S ) BV(Q;\mathbb{Z} \cup S) B V ( Q ; Z ∪ S ) . Conversely, in (R2) the intermediate wells become energetically too expensive. Consequently, the limit function u u u is forced to take values in Z \mathbb{Z} Z .
The limiting jump costs
In the sharp interface limit only pure phases are present: the limit u u u takes values in Z ∪ S \mathbb{Z} \cup S Z ∪ S in regime (R1) and in Z \mathbb{Z} Z in (R2). Moreover, the limit energy is concentrated on the jump set J u J_u J u with density Ψ ( u + , u − ) \Psi(u^{+},u^{-}) Ψ ( u + , u − ) , where u + , u − ∈ Z ∪ S u^{+},u^{-} \in\mathbb{Z} \cup S u + , u − ∈ Z ∪ S denote the values of u u u around a jump point. We refer to Ψ ( ⋅ , ⋅ ) \Psi(\cdot,\cdot) Ψ ( ⋅ , ⋅ ) as the jump cost . In all cases, it is the product of three factors
Ψ ( ⋅ , ⋅ ) = λ s , n ⏟ Dimensional const. ⋅ κ s ⏟ Transition profile const. ⋅ Φ ( ⋅ , ⋅ ) ⏟ Position of the wells \Psi(\cdot,\cdot)=\underbrace{\lambda_{s,n}}_{\text{Dimensional const.}}\cdot\underbrace{\kappa_s}_{\text{Transition profile const.}}\cdot\underbrace{\Phi(\cdot,\cdot)}_{\text{Position of the wells}} Ψ ( ⋅ , ⋅ ) = Dimensional const. λ s , n ⋅ Transition profile const. κ s ⋅ Position of the wells Φ ( ⋅ , ⋅ )
that we define as follows.
Definition 2.1 (Position of the wells). Let V \mathcal{V} V be a discrete set in R \mathbb{R} R . For a < b a < b a < b in V \mathcal{V} V , let a = p 0 < p 1 < ⋯ < p N = b a = p_0 < p_1 < \cdots< p_N = b a = p 0 < p 1 < ⋯ < p N = b be the points of V ∩ [ a , b ] \mathcal{V} \cap[a,b] V ∩ [ a , b ] . We define
Φ V ( a , b ) : = ∑ i = 1 N ( p i − p i − 1 ) 2 , \Phi_{\mathcal{V}}(a,b):=\sum_{i=1}^{N}(p_i-p_{i-1})^2, Φ V ( a , b ) := i = 1 ∑ N ( p i − p i − 1 ) 2 , extended by Φ V ( a , b ) = Φ V ( b , a ) \Phi_{\mathcal{V}}(a,b)=\Phi_{\mathcal{V}}(b,a) Φ V ( a , b ) = Φ V ( b , a ) and Φ V ( a , a ) = 0 \Phi_{\mathcal{V}}(a,a)=0 Φ V ( a , a ) = 0 .
Definition 2.2 (Transition profile constant). We set κ 1 / 2 = 2 \kappa_{1/2}=2 κ 1/2 = 2 and for s ∈ ( 1 2 , 1 ) s\in\left(\frac{1}{2},1\right) s ∈ ( 2 1 , 1 ) ,
κ s : = inf { [ w ] H s ( R ) 2 + ∫ R dist 2 ( w , { 0 , 1 } ) d x : w ∈ H l o c s ( R ) , lim x → − ∞ w ( x ) = 0 , lim x → + ∞ w ( x ) = 1 } . (8) \kappa_s:=\inf\left\{[w]_{H^s(\mathbb{R})}^2+\int_{\mathbb{R}}\operatorname{dist}^2(w,\{0,1\})\,dx:w\in H^s_{\mathrm{loc}}(\mathbb{R}),\lim_{x\to-\infty}w(x)=0,\lim_{x\to+\infty}w(x)=1\right\}. \tag*{(8)} κ s := inf { [ w ] H s ( R ) 2 + ∫ R dist 2 ( w , { 0 , 1 }) d x : w ∈ H loc s ( R ) , x → − ∞ lim w ( x ) = 0 , x → + ∞ lim w ( x ) = 1 } . (8) where
[ w ] H s ( R ) 2 = ∫ R ∫ R ∣ w ( x ) − w ( y ) ∣ 2 ∣ x − y ∣ 1 + 2 s d x d y . [w]_{H^s(\mathbb{R})}^2=\int_{\mathbb{R}}\int_{\mathbb{R}}\frac{\lvert w(x)-w(y)\rvert^2}{\lvert x-y\rvert^{1+2s}}\,dx\,dy. [ w ] H s ( R ) 2 = ∫ R ∫ R ∣ x − y ∣ 1 + 2 s ∣ w ( x ) − w ( y ) ∣ 2 d x d y . The constant κ s \kappa_s κ s is the one-dimensional cost of a transition from 0 to 1 for the standard double-well potential dist 2 ( ⋅ , { 0 , 1 } ) \operatorname{dist}^2(\cdot,\{0,1\}) dist 2 ( ⋅ , { 0 , 1 }) : if s > 1 2 s>\frac{1}{2} s > 2 1 , it is given by the optimal profile problem (see, e.g., [25 , 26 ] ); if s = 1 2 s=\frac{1}{2} s = 2 1 , with the logarithmic scale, the cost does not depend on the profile, and it is equal to 2 (see, e.g., [3 , 15 , 19 ] ).
Definition 2.3 (Dimensional constant). For n ≥ 1 n\ge1 n ≥ 1 and s ∈ [ 1 2 , 1 ) s\in\left[\frac{1}{2},1\right) s ∈ [ 2 1 , 1 ) , we set
C s , n : = π n − 1 2 Γ ( s + 1 2 ) Γ ( n + 2 s 2 ) { ∫ R n − 1 ( 1 + ∣ t ∣ 2 ) − n + 2 s 2 d t , if n ≥ 2 , 1 , if n = 1 and λ s , n : = ( C s , n ) 1 2 s . (9) C_{s,n}:=\pi^{\frac{n-1}{2}}\frac{\Gamma\left(s+\frac{1}{2}\right)}{\Gamma\left(\frac{n+2s}{2}\right)}
\begin{cases}
\int_{\mathbb{R}^{n-1}}\left(1+\lvert t\rvert^2\right)^{-\frac{n+2s}{2}}\,dt, & \text{if }n\ge2,\\
1, & \text{if }n=1
\end{cases}
\qquad\text{and}\qquad
\lambda_{s,n}:=(C_{s,n})^{\frac{1}{2s}}. \tag*{(9)} C s , n := π 2 n − 1 Γ ( 2 n + 2 s ) Γ ( s + 2 1 ) { ∫ R n − 1 ( 1 + ∣ t ∣ 2 ) − 2 n + 2 s d t , 1 , if n ≥ 2 , if n = 1 and λ s , n := ( C s , n ) 2 s 1 . (9) In particular, with this definition we have
λ s , 1 = 1 , λ 1 / 2 , n = C 1 / 2 , n = ω n − 1 , \lambda_{s,1}=1,\qquad\lambda_{1/2,n}=C_{1/2,n}=\omega_{n-1}, λ s , 1 = 1 , λ 1/2 , n = C 1/2 , n = ω n − 1 , where ω n − 1 \omega_{n-1} ω n − 1 is the volume of the unit ball of R n − 1 \mathbb{R}^{n-1} R n − 1 (with ω 0 = 1 \omega_0=1 ω 0 = 1 ).
Based on the above definitions, we can define the jump costs.
Definition 2.4 (Jump costs). For n ≥ 1 n\ge1 n ≥ 1 and s ∈ [ 1 2 , 1 ) s\in\left[\frac{1}{2},1\right) s ∈ [ 2 1 , 1 ) , we define
Ψ s u p : ( Z ∪ S ) × ( Z ∪ S ) → [ 0 , + ∞ ) , Ψ s u b : Z × Z → [ 0 , + ∞ ) , \Psi_{\mathrm{sup}}:(\mathbb{Z}\cup S)\times(\mathbb{Z}\cup S)\to[0,+\infty),\qquad\Psi_{\mathrm{sub}}:\mathbb{Z}\times\mathbb{Z}\to[0,+\infty), Ψ sup : ( Z ∪ S ) × ( Z ∪ S ) → [ 0 , + ∞ ) , Ψ sub : Z × Z → [ 0 , + ∞ ) , defined by
Ψ s u p : = λ s , n κ s Φ Z ∪ S , Ψ s u b : = λ s , n κ s { ( 1 − ℓ ) Φ Z + ℓ Φ Z ∪ S , if s = 1 2 , Φ Z ∪ S , if s > 1 2 , \Psi_{\mathrm{sup}}:=\lambda_{s,n}\kappa_s\Phi_{\mathbb{Z}\cup S},\qquad
\Psi_{\mathrm{sub}}:=\lambda_{s,n}\kappa_s
\begin{cases}
(1-\ell)\Phi_{\mathbb{Z}}+\ell\Phi_{\mathbb{Z}\cup S}, & \text{if }s=\frac{1}{2},\\
\Phi_{\mathbb{Z}\cup S}, & \text{if }s>\frac{1}{2},
\end{cases} Ψ sup := λ s , n κ s Φ Z ∪ S , Ψ sub := λ s , n κ s { ( 1 − ℓ ) Φ Z + ℓ Φ Z ∪ S , Φ Z ∪ S , if s = 2 1 , if s > 2 1 , where ℓ ∈ [ 0 , 1 ] \ell\in[0,1] ℓ ∈ [ 0 , 1 ] is the parameter in (7) .
Observe that, for s > 1 2 s>\frac{1}{2} s > 2 1 , the cost Ψ s u b \Psi_{\mathrm{sub}} Ψ sub is the restriction of Ψ s u p \Psi_{\mathrm{sup}} Ψ sup to integer phases; if instead s = 1 2 s=\frac{1}{2} s = 2 1 , it is the interpolation of the cost with the wells in Z \mathbb{Z} Z and in Z ∪ S \mathbb{Z}\cup S Z ∪ S (see Remark 2.9). Moreover, both jump costs are additive with respect to the order of the wells, namely
Ψ s u p ( α , γ ) = Ψ s u p ( α , β ) + Ψ s u p ( β , γ ) for α < β < γ in Z ∪ S , \Psi_{\mathrm{sup}}(\alpha,\gamma)=\Psi_{\mathrm{sup}}(\alpha,\beta)+\Psi_{\mathrm{sup}}(\beta,\gamma)\qquad\text{for }\alpha<\beta<\gamma\text{ in }\mathbb{Z}\cup S, Ψ sup ( α , γ ) = Ψ sup ( α , β ) + Ψ sup ( β , γ ) for α < β < γ in Z ∪ S , Ψ s u b ( α , γ ) = Ψ s u b ( α , β ) + Ψ s u b ( β , γ ) for α < β < γ in Z . \Psi_{\mathrm{sub}}(\alpha,\gamma)=\Psi_{\mathrm{sub}}(\alpha,\beta)+\Psi_{\mathrm{sub}}(\beta,\gamma)\qquad\text{for }\alpha<\beta<\gamma\text{ in }\mathbb{Z}. Ψ sub ( α , γ ) = Ψ sub ( α , β ) + Ψ sub ( β , γ ) for α < β < γ in Z . This means that a jump between non-consecutive phases costs exactly as much as a staircase of elementary (admissible) jumps visiting all intermediate phases. Therefore, it is sufficient to characterize the jump costs between consecutive wells in Z ∪ S \mathbb{Z}\cup S Z ∪ S and Z \mathbb{Z} Z , respectively. Let a < b a<b a < b be two consecutive elements of Z ∪ S \mathbb{Z}\cup S Z ∪ S , and let k ∈ Z k\in\mathbb{Z} k ∈ Z . By Definition 2.4 , we have
Ψ s u p ( a , b ) = { ω n − 1 2 ( b − a ) 2 , if s = 1 2 , λ s , n κ s ( b − a ) 2 , if s > 1 2 . \Psi_{\mathrm{sup}}(a,b)=
\begin{cases}
\omega_{n-1}2(b-a)^2, & \text{if }s=\frac{1}{2},\\
\lambda_{s,n}\kappa_s(b-a)^2, & \text{if }s>\frac{1}{2}.
\end{cases} Ψ sup ( a , b ) = { ω n − 1 2 ( b − a ) 2 , λ s , n κ s ( b − a ) 2 , if s = 2 1 , if s > 2 1 .
Ψ s u b ( k , k + 1 ) = { ω n − 1 2 ( 1 − ℓ 2 w k ( 1 − w k ) ) , if s = 1 2 , λ s , n κ s ( 1 − 2 w k ( 1 − w k ) ) , if s > 1 2 , with w k : = s k − k , \Psi_{\mathrm{sub}}(k,k+1)=
\begin{cases}
\omega_{n-1}2\left(1-\ell\,2w_k(1-w_k)\right), & \text{if }s=\frac{1}{2},\\
\lambda_{s,n}\kappa_s\left(1-2w_k(1-w_k)\right), & \text{if }s>\frac{1}{2},
\end{cases}
\qquad\text{with}\qquad w_k:=s_k-k, Ψ sub ( k , k + 1 ) = { ω n − 1 2 ( 1 − ℓ 2 w k ( 1 − w k ) ) , λ s , n κ s ( 1 − 2 w k ( 1 − w k ) ) , if s = 2 1 , if s > 2 1 , with w k := s k − k ,
where for s = 1 2 s=\frac{1}{2} s = 2 1 we used λ 1 / 2 , n = ω n − 1 \lambda_{1/2,n}=\omega_{n-1} λ 1/2 , n = ω n − 1 .
The Γ \Gamma Γ -convergence theorem
We can now state the main results of the paper, which provide a complete characterization of the Γ \Gamma Γ -limit in the different regimes.
Throughout, for u ∈ B V ( Q ; I ) u\in BV(Q;\mathcal{I}) u ∈ B V ( Q ; I ) with I ⊂ R \mathcal{I}\subset\mathbb{R} I ⊂ R discrete, J u J_u J u denotes the (approximate) jump set, u ± u^\pm u ± the one-sided traces on J u J_u J u with the orientation chosen so that u + > u − u^+>u^- u + > u − , [ u ] : = u + − u − [u]:=u^+-u^- [ u ] := u + − u − the jump amplitude, and H n − 1 \mathcal{H}^{n-1} H n − 1 the ( n − 1 ) (n-1) ( n − 1 ) -dimensional Hausdorff measure (see [5 ] ). We write ∣ A ∣ |A| ∣ A ∣ for the Lebesgue measure of a measurable set A ⊂ R n A\subset\mathbb{R}^n A ⊂ R n .
Theorem 2.5 (Compactness). Let s ∈ [ 1 2 , 1 ) s\in\left[\frac{1}{2},1\right) s ∈ [ 2 1 , 1 ) and let { u ε } ⊂ L 1 ( Q ) \{u_\varepsilon\}\subset L^1(Q) { u ε } ⊂ L 1 ( Q ) satisfy
sup ε F ε ( u ε ) < + ∞ . \sup_{\varepsilon}F_\varepsilon(u_\varepsilon)<+\infty. ε sup F ε ( u ε ) < + ∞. Then there exists a sequence of integers { a ε } ⊂ Z \{a_\varepsilon\}\subset\mathbb{Z} { a ε } ⊂ Z such that the translated sequence { u ε − a ε } \{u_\varepsilon-a_\varepsilon\} { u ε − a ε } is relatively compact in L 1 ( Q ) L^1(Q) L 1 ( Q ) .
Moreover, for small ε \varepsilon ε , the sequence { u ε − a ε } \{u_\varepsilon-a_\varepsilon\} { u ε − a ε } is uniformly bounded in L n n − 1 ( Q ) L^{\frac{n}{n-1}}(Q) L n − 1 n ( Q ) if n ≥ 2 n\geq2 n ≥ 2 , and in L q ( Q ) L^q(Q) L q ( Q ) for every q ∈ [ 1 , + ∞ ) q\in[1,+\infty) q ∈ [ 1 , + ∞ ) if n = 1 n=1 n = 1 ; if n = 1 n=1 n = 1 and s > 1 2 s>\frac{1}{2} s > 2 1 , it is uniformly bounded in L ∞ ( Q ) L^\infty(Q) L ∞ ( Q ) . Finally, if u ε → u u_\varepsilon\to u u ε → u in L 1 ( Q ) L^1(Q) L 1 ( Q ) , then u ∈ B V ( Q ; Z ∪ S ) u\in BV(Q;\mathbb{Z}\cup S) u ∈ B V ( Q ; Z ∪ S ) . In regime (R2), we have u ∈ B V ( Q ; Z ) u\in BV(Q;\mathbb{Z}) u ∈ B V ( Q ; Z ) .
Remark 2.6. The limit points of { u ε − a ε } \{u_\varepsilon-a_\varepsilon\} { u ε − a ε } need not take values in Z ∪ S \mathbb{Z}\cup S Z ∪ S if S S S is not periodic, because F ε F_\varepsilon F ε is not invariant under integer translations. For instance, take u ε ≡ s k ε u_\varepsilon\equiv s k_\varepsilon u ε ≡ s k ε with k ε → ∞ k_\varepsilon\to\infty k ε → ∞ . In general these limit points take values in Z ∪ S ∗ \mathbb{Z}\cup S^* Z ∪ S ∗ , where S ∗ : = { lim ε s z + a ε − a ε : z ∈ Z } S^*:=\{\lim_{\varepsilon}s z+a_\varepsilon-a_\varepsilon:z\in\mathbb{Z}\} S ∗ := { lim ε sz + a ε − a ε : z ∈ Z } (see the proof in § 3.3 ).
Theorem 2.7 (Γ \Gamma Γ -convergence). Let s ∈ [ 1 2 , 1 ) s\in\left[\frac{1}{2},1\right) s ∈ [ 2 1 , 1 ) and Q ⊂ R n Q\subset\mathbb{R}^n Q ⊂ R n be an open cube. As ε → 0 \varepsilon\to0 ε → 0 , the sequence of functionals F ε F_\varepsilon F ε in (3) Γ \Gamma Γ -converges, with respect to the strong L 1 ( Q ) L^1(Q) L 1 ( Q ) topology, to the following limits.
Regime (R1): If δ ε \delta_\varepsilon δ ε satisfies (5) with K ∈ [ 0 , + ∞ ) K\in[0,+\infty) K ∈ [ 0 , + ∞ ) , the Γ \Gamma Γ -limit is
F s u p ( u ) : = { ∫ J u Ψ s u p ( u + , u − ) d H n − 1 + K ∣ { x ∈ Q : u ( x ) ∈ S } ∣ , if u ∈ B V ( Q ; Z ∪ S ) , + ∞ , otherwise , F_{\mathrm{sup}}(u):=
\begin{cases}
\int_{J_u}\Psi_{\mathrm{sup}}(u^+,u^-)\,d\mathcal{H}^{n-1}
+K\left|\{x\in Q:u(x)\in S\}\right|, & \text{if }u\in BV(Q;\mathbb{Z}\cup S),\\
+\infty, & \text{otherwise},
\end{cases} F sup ( u ) := { ∫ J u Ψ sup ( u + , u − ) d H n − 1 + K ∣ { x ∈ Q : u ( x ) ∈ S } ∣ , + ∞ , if u ∈ B V ( Q ; Z ∪ S ) , otherwise , where Ψ s u p \Psi_{\mathrm{sup}} Ψ sup is the jump cost in Definition 2.4 .
F s u b ( u ) : = { ∫ J u Ψ s u b ( u + , u − ) d H n − 1 , if u ∈ B V ( Q ; Z ) , + ∞ , otherwise , F_{\mathrm{sub}}(u):=
\begin{cases}
\int_{J_u}\Psi_{\mathrm{sub}}(u^+,u^-)\,d\mathcal{H}^{n-1}, & \text{if }u\in BV(Q;\mathbb{Z}),\\
+\infty, & \text{otherwise},
\end{cases} F sub ( u ) := { ∫ J u Ψ sub ( u + , u − ) d H n − 1 , + ∞ , if u ∈ B V ( Q ; Z ) , otherwise , where Ψ s u b \Psi_{\mathrm{sub}} Ψ sub is the jump cost in Definition 2.4 .
It should be noted that the two regimes differ in the set of admissible phases and in the role of the secondary wells S S S .
In regime (R1), the phases in S S S are admissible and each has a volume cost K K K ;
In regime (R2), the phases in S S S are not admissible; however, the set S S S affects the jump cost depending on s s s :
If s > 1 2 s > \frac{1}{2} s > 2 1 , the cost of the jump k → k + 1 k \to k+1 k → k + 1 is equal to the cost of a double jump k → s k → k + 1 k \to s_k \to k+1 k → s k → k + 1 , passing through the intermediate well s k s_k s k ;
If s = 1 2 s = \frac{1}{2} s = 2 1 , also the parameter ℓ \ell ℓ plays a role: it represents the interpolation between a jump that ignores s k s_k s k and the one that passes through it (see Remark 2.9).
For future reference, given a discrete set V ⊂ R \mathcal{V} \subset\mathbb{R} V ⊂ R and an open domain A ⊆ R n A \subseteq\mathbb{R}^n A ⊆ R n , we define
F ε V ( u ; A ) : = 1 θ ε ( ε 2 s ∫ A ∫ A ∣ u ( x ) − u ( y ) ∣ 2 ∣ x − y ∣ n + 2 s d x d y + ∫ A dist 2 ( u ( x ) , V ) d x ) , (10) F_{\varepsilon}^{\mathcal{V}}(u;A) := \frac{1}{\theta_{\varepsilon}}\left(\varepsilon^{2s}\int_A\int_A \frac{\lvert u(x)-u(y)\rvert^2}{\lvert x-y\rvert^{n+2s}}\,dx\,dy+\int_A \operatorname{dist}^2(u(x),\mathcal{V})\,dx\right), \tag*{(10)} F ε V ( u ; A ) := θ ε 1 ( ε 2 s ∫ A ∫ A ∣ x − y ∣ n + 2 s ∣ u ( x ) − u ( y ) ∣ 2 d x d y + ∫ A dist 2 ( u ( x ) , V ) d x ) , (10)
and we omit A A A when A = Q A=Q A = Q .
Remark 2.8 (Periodic potential). Theorem 2.7 also covers the Γ \Gamma Γ -limit of the functional F ε Z F_{\varepsilon}^{\mathbb{Z}} F ε Z , obtained from (10) with V = Z \mathcal{V}=\mathbb{Z} V = Z . This functional was studied for s = 1 2 s=\frac{1}{2} s = 2 1 in [19 ] (n = 1 n=1 n = 1 ) and [15 ] (n = 2 n=2 n = 2 ). Its Γ \Gamma Γ -limit is
λ s , n κ s ∫ J u [ u ] d H n − 1 , u ∈ B V ( Q ; Z ) , (11) \lambda_{s,n}\kappa_s\int_{J_u}[u]\,d\mathcal{H}^{n-1}, \qquad u\in BV(Q;\mathbb{Z}), \tag*{(11)} λ s , n κ s ∫ J u [ u ] d H n − 1 , u ∈ B V ( Q ; Z ) , (11)
which, for s = 1 2 s=\frac{1}{2} s = 2 1 , equals 2 ω n − 1 ∫ J u [ u ] d H n − 1 2\omega_{n-1}\int_{J_u}[u]\,d\mathcal{H}^{n-1} 2 ω n − 1 ∫ J u [ u ] d H n − 1 , that is precisely the result already obtained in [19 ] and [15 ] . The functional (11) can be obtained by a simple rescaling; indeed we have
F ε Z ( u ) = 4 F ε 1 2 Z ( u 2 ) , F_{\varepsilon}^{\mathbb{Z}}(u)=4F_{\varepsilon}^{\frac{1}{2}\mathbb{Z}}\left(\frac{u}{2}\right), F ε Z ( u ) = 4 F ε 2 1 Z ( 2 u ) ,
where F ε 1 2 Z F_{\varepsilon}^{\frac{1}{2}\mathbb{Z}} F ε 2 1 Z is the functional in (3) with S = Z + 1 2 S=\mathbb{Z}+\frac{1}{2} S = Z + 2 1 and δ ε = 0 \delta_{\varepsilon}=0 δ ε = 0 , i.e. the regime ( R 1 ) (R1) ( R 1 ) with K = 0 K=0 K = 0 . Hence, the Γ \Gamma Γ -limit of F ε 1 2 Z ( ⋅ / 2 ) F_{\varepsilon}^{\frac{1}{2}\mathbb{Z}}(\cdot/2) F ε 2 1 Z ( ⋅ /2 ) is, by Theorem 2.7 , the following
λ s , n κ s ∫ J u Φ 1 2 Z ( u + / 2 , u − / 2 ) d H n − 1 = λ s , n κ s ∫ J u [ u ] 4 d H n − 1 , u ∈ B V ( Q ; Z ) . \lambda_{s,n}\kappa_s\int_{J_u}\Phi_{\frac{1}{2}\mathbb{Z}}(u^+/2,u^-/2)\,d\mathcal{H}^{n-1}=\lambda_{s,n}\kappa_s\int_{J_u}\frac{[u]}{4}\,d\mathcal{H}^{n-1}, \qquad u\in BV(Q;\mathbb{Z}). λ s , n κ s ∫ J u Φ 2 1 Z ( u + /2 , u − /2 ) d H n − 1 = λ s , n κ s ∫ J u 4 [ u ] d H n − 1 , u ∈ B V ( Q ; Z ) .
Remark 2.9 (Role of ℓ \ell ℓ for s = 1 2 s=\frac{1}{2} s = 2 1 ). In the subcritical regime ( R 2 ) (R2) ( R 2 ) for s = 1 2 s=\frac{1}{2} s = 2 1 , the limit energy F s u b F_{\mathrm{sub}} F sub can be interpreted as an interpolation, through the parameter ℓ \ell ℓ , between the two extreme cases: when the potential wells are only at integers, and when wells occur in Z ∪ S \mathbb{Z}\cup S Z ∪ S . Indeed, by definition, we have
F s u b ( u ) = ( 1 − ℓ ) [ 2 ω n − 1 ∫ J u [ u ] d H n − 1 ] + ℓ [ 2 ω n − 1 ∫ J u Φ Z ∪ S ( u − , u + ) d H n − 1 ] , for u ∈ B V ( Q ; Z ) , F_{\mathrm{sub}}(u)=(1-\ell)\left[2\omega_{n-1}\int_{J_u}[u]\,d\mathcal{H}^{n-1}\right]+\ell\left[2\omega_{n-1}\int_{J_u}\Phi_{\mathbb{Z}\cup S}(u^-,u^+)\,d\mathcal{H}^{n-1}\right], \qquad\text{for }u\in BV(Q;\mathbb{Z}), F sub ( u ) = ( 1 − ℓ ) [ 2 ω n − 1 ∫ J u [ u ] d H n − 1 ] + ℓ [ 2 ω n − 1 ∫ J u Φ Z ∪ S ( u − , u + ) d H n − 1 ] , for u ∈ B V ( Q ; Z ) ,
where the first term is the Γ \Gamma Γ -limit of F ε Z F_{\varepsilon}^{\mathbb{Z}} F ε Z (as we observe in Remark 2.8) and the second is the Γ \Gamma Γ -limit of F ε Z ∪ S F_{\varepsilon}^{\mathbb{Z}\cup S} F ε Z ∪ S (as follows from Theorem 2.7 choosing δ ε = 0 \delta_{\varepsilon}=0 δ ε = 0 ).
Let us notice that the jump cost appearing in the limit functional has a different nature according to s s s .
For s > 1 / 2 s>1/2 s > 1/2 the energy of a transition between two consecutive wells is the minimal energy of a one-dimensional profile connecting them: the cost κ s \kappa_s κ s in (8) is defined by an optimal profile problem on R \mathbb{R} R ;
For s = 1 / 2 s=1/2 s = 1/2 , there is no optimal profile; indeed, because of the logarithmic scaling, the cost depends only on the end states: the cost is constant, κ s = 2 \kappa_s=2 κ s = 2 .
In both cases, the n n n -dimensional problem reduces to a one-dimensional one across the interface, and the dimension only enters through a multiplicative constant λ s , n \lambda_{s,n} λ s , n . The next remark makes this reduction explicit and explains how the constant affects the cost in each case.
Crucial comparison estimate and strategy of the proof
Besides the presence of the secondary wells and their effect in the different regimes, the present model shows two major difficulties: lack of coercivity of the potential and non-locality of the singular perturbation.
The lack of coercivity is clear: every integer constant has zero energy. Hence compactness can only be expected up to integer translations, as in the periodic case studied in [15 , 19 ] (n = 1 , 2 n=1,2 n = 1 , 2 , s = 1 / 2 s=1/2 s = 1/2 ). However, in our setting the potential is not periodic and then the energy F ε F_\varepsilon F ε is not invariant under integer translations, so the limit of a translated sequence need not take values in Z ∪ S \mathbb{Z}\cup S Z ∪ S (see Remark 2.6).
A key ingredient in the analysis is the comparison between our original functional and a simpler reference functional F ε Z ∪ S F_\varepsilon^{\mathbb{Z}\cup S} F ε Z ∪ S , obtained from (10) with V = Z ∪ S \mathcal{V}=\mathbb{Z}\cup S V = Z ∪ S , i.e.
F ε Z ∪ S ( u ) : = 1 θ ε ( ε 2 s ∫ Q ∫ Q ∣ u ( x ) − u ( y ) ∣ 2 ∣ x − y ∣ n + 2 s d x d y + ∫ Q dist 2 ( u ( x ) , Z ∪ S ) d x ) . (12) F_\varepsilon^{\mathbb{Z}\cup S}(u):=\frac{1}{\theta_\varepsilon}\left(\varepsilon^{2s}\int_Q\int_Q\frac{\lvert u(x)-u(y)\rvert^2}{\lvert x-y\rvert^{n+2s}}\,dx\,dy+\int_Q\operatorname{dist}^2(u(x),\mathbb{Z}\cup S)\,dx\right). \tag*{(12)} F ε Z ∪ S ( u ) := θ ε 1 ( ε 2 s ∫ Q ∫ Q ∣ x − y ∣ n + 2 s ∣ u ( x ) − u ( y ) ∣ 2 d x d y + ∫ Q dist 2 ( u ( x ) , Z ∪ S ) d x ) . (12)
Clearly F ε Z ∪ S ( u ) ≤ F ε ( u ) F_\varepsilon^{\mathbb{Z}\cup S}(u)\leq F_\varepsilon(u) F ε Z ∪ S ( u ) ≤ F ε ( u ) . A more careful comparison allows us to isolate and quantify the specific contribution of the energy gap δ ε \delta_\varepsilon δ ε in the asymptotic behavior of F ε F_\varepsilon F ε .
To see this, for any δ ∈ ( 0 , 1 ] \delta\in(0,1] δ ∈ ( 0 , 1 ] we define the set
A δ : = { t ∈ R : dist ( t , S ) < δ dist ( t , Z ) } , A_\delta:=\{t\in\mathbb{R}:\operatorname{dist}(t,S)<\delta\operatorname{dist}(t,\mathbb{Z})\}, A δ := { t ∈ R : dist ( t , S ) < δ dist ( t , Z )} ,
and note that if δ ∈ ( 0 , 1 ) \delta\in(0,1) δ ∈ ( 0 , 1 ) and ε \varepsilon ε is small enough (depending on δ \delta δ and d d d ), it holds
A δ ⊂ { t ∈ R : dist 2 ( t , S ) + δ ε < dist 2 ( t , Z ) } ⊂ A 1 . A_\delta\subset\{t\in\mathbb{R}:\operatorname{dist}^2(t,S)+\delta_\varepsilon<\operatorname{dist}^2(t,\mathbb{Z})\}\subset A_1. A δ ⊂ { t ∈ R : dist 2 ( t , S ) + δ ε < dist 2 ( t , Z )} ⊂ A 1 .
Moreover, by the very definition of the potential W ε W_\varepsilon W ε we have that W ε ( t ) = dist 2 ( t , Z ) = dist 2 ( t , Z ∪ S ) W_\varepsilon(t)=\operatorname{dist}^2(t,\mathbb{Z})=\operatorname{dist}^2(t,\mathbb{Z}\cup S) W ε ( t ) = dist 2 ( t , Z ) = dist 2 ( t , Z ∪ S ) if t ∉ A 1 t\notin A_1 t ∈ / A 1 and W ε ( t ) ≤ dist 2 ( t , S ) + δ ε = dist 2 ( t , Z ∪ S ) + δ ε W_\varepsilon(t)\leq\operatorname{dist}^2(t,S)+\delta_\varepsilon=\operatorname{dist}^2(t,\mathbb{Z}\cup S)+\delta_\varepsilon W ε ( t ) ≤ dist 2 ( t , S ) + δ ε = dist 2 ( t , Z ∪ S ) + δ ε if t ∈ A 1 t\in A_1 t ∈ A 1 , with equality if t ∈ A δ t\in A_\delta t ∈ A δ (for ε \varepsilon ε small). This, for every given δ ∈ ( 0 , 1 ) \delta\in(0,1) δ ∈ ( 0 , 1 ) and ε > 0 \varepsilon>0 ε > 0 small enough, leads to the following two-sided estimate
F ε Z ∪ S ( u ) + δ ε θ ε ∣ A δ ( u ) ∣ ≤ F ε ( u ) ≤ F ε Z ∪ S ( u ) + δ ε θ ε ∣ A 1 ( u ) ∣ , (13) F_\varepsilon^{\mathbb{Z}\cup S}(u)+\frac{\delta_\varepsilon}{\theta_\varepsilon}\lvert A_\delta(u)\rvert\leq F_\varepsilon(u)\leq F_\varepsilon^{\mathbb{Z}\cup S}(u)+\frac{\delta_\varepsilon}{\theta_\varepsilon}\lvert A_1(u)\rvert, \tag*{(13)} F ε Z ∪ S ( u ) + θ ε δ ε ∣ A δ ( u )∣ ≤ F ε ( u ) ≤ F ε Z ∪ S ( u ) + θ ε δ ε ∣ A 1 ( u )∣ , (13)
where
A δ ( u ) : = { x ∈ Q : u ( x ) ∈ A δ } . (14) A_\delta(u):=\{x\in Q:u(x)\in A_\delta\}. \tag*{(14)} A δ ( u ) := { x ∈ Q : u ( x ) ∈ A δ } . (14)
The estimate (13) is a fundamental technical tool in our analysis. It effectively decouples the total energy into two distinct contributions:
A fractional transition energy (F ε Z ∪ S F_\varepsilon^{\mathbb{Z}\cup S} F ε Z ∪ S ), which accounts for the sharp interface energy;
A penalization term proportional to the measure of the set where u u u is close to the secondary wells in S S S .
This decomposition clarifies the emergence of the two regimes: in the subcritical regime (R2), the penalization factor δ ε / θ ε \delta_\varepsilon/\theta_\varepsilon δ ε / θ ε diverges, acting as a constraint that forces the limit function to take values only in Z \mathbb{Z} Z . Conversely, in the critical/supercritical regime (R1), this factor remains bounded or vanishes, allowing the secondary wells to survive in the limit as a bulk contribution.
Another elementary tool, used both in the proof of compactness and in the lower bound, is a truncation inequality. Let V = { v i } i ∈ Z ⊂ R \mathcal{V}=\{v_i\}_{i\in\mathbb{Z}}\subset\mathbb{R} V = { v i } i ∈ Z ⊂ R be a discrete set with v i < v i + 1 v_i<v_{i+1} v i < v i + 1 for all i ∈ Z i\in\mathbb{Z} i ∈ Z , and let T i t : = ( t ∨ v i ) ∧ v i + 1 T_i t:=(t\vee v_i)\wedge v_{i+1} T i t := ( t ∨ v i ) ∧ v i + 1 be the truncation between two consecutive values of V \mathcal{V} V . For all t , t ′ ∈ R t,t'\in\mathbb{R} t , t ′ ∈ R , we have
∑ i ∈ Z ∣ T i t − T i t ′ ∣ 2 ≤ ∣ t − t ′ ∣ 2 and ∑ i ∈ Z W ( T i t ) = W ( t ) for every W ≥ 0 vanishing on V . \sum_{i\in\mathbb{Z}}\lvert T_i t-T_i t'\rvert^2\leq\lvert t-t'\rvert^2
\qquad\text{and}\qquad
\sum_{i\in\mathbb{Z}}W(T_i t)=W(t)\text{ for every }W\geq0\text{ vanishing on }\mathcal{V}. i ∈ Z ∑ ∣ T i t − T i t ′ ∣ 2 ≤ ∣ t − t ′ ∣ 2 and i ∈ Z ∑ W ( T i t ) = W ( t ) for every W ≥ 0 vanishing on V .
Hence, we obtain
∑ i ∈ Z F ε V ( T i u ) ≤ F ε V ( u ) , (15) \sum_{i\in\mathbb{Z}} F_\varepsilon^{\mathcal{V}}(T_i u) \leq F_\varepsilon^{\mathcal{V}}(u), \tag*{(15)} i ∈ Z ∑ F ε V ( T i u ) ≤ F ε V ( u ) , (15)
where F ε V F_\varepsilon^{\mathcal{V}} F ε V is defined in (10) . Thus the energy of u u u controls the sum of the energies of its truncations, and the analysis can often be reduced to a single transition between two consecutive points of V \mathcal{V} V .
Strategy of the proof. The compactness result (Theorem 2.5 ) is proved in § 3 . By the lower estimate in (13) it is enough to establish the compactness result for a sequence uniformly bounded for the energy F ε Z ∪ S F_\varepsilon^{\mathbb{Z}\cup S} F ε Z ∪ S . Here, a bi-Lipschitz change of variables mapping Z ∪ S \mathbb{Z}\cup S Z ∪ S onto Z \mathbb{Z} Z reduces the problem to the periodic energy F ε Z F_\varepsilon^{\mathbb{Z}} F ε Z . For the latter, we prove uniform L q L^q L q bounds up to integer translations, extending [15 , 19 ] , and we deduce compactness by truncating between consecutive integers and using the classical two-well results. Since F ε F_\varepsilon F ε is not invariant under integer translations, the limit takes values in Z ∪ S \mathbb{Z}\cup S Z ∪ S (in Z \mathbb{Z} Z in regime (R2)) only for sequences that converge without translations.
In § 5 we establish the lower bound for the functional F ε F_\varepsilon F ε . By the truncation inequality (15) , it is enough to consider a single transition between two consecutive wells of the limit. In regime (R1), the lower estimate (13) allows us to distinguish the contribution of the fixed wells in Z \mathbb{Z} Z from that of the local minima in S S S . In particular we can study F ε Z ∪ S F_\varepsilon^{\mathbb{Z}\cup S} F ε Z ∪ S , whose potential does not depend on ε \varepsilon ε ; the same works in regime (R2) if s > 1 / 2 s>1/2 s > 1/2 . The result in dimension one (§ 5.1 ) reduces to the classical two-well setting, for which the lower bound is known (see, e.g., [3 , 25 , 26 ] ). The genuinely new case is regime (R2) for s = 1 / 2 s=1/2 s = 1/2 (see Lemma 5.2 ) where an almost optimal transition between k k k and k + 1 k+1 k + 1 rests at s k s_k s k on an interval of length ε / δ ε \varepsilon/\delta_\varepsilon ε / δ ε : this lowers the cost by an amount depending on ℓ \ell ℓ . Then the n n n -dimensional case is obtained in § 5.2 by slicing in all directions splitting the potential among them with a probability density on the sphere; optimizing this density we recover the dimensional constant λ s , n \lambda_{s,n} λ s , n .
In § 6 we establish the upper bound. By density, it is enough to construct a recovery sequence for each function with a polyhedral jump set that jumps between consecutive wells in the limit. The recovery sequence depends on the regime and on s s s . In particular, for s = 1 / 2 s=1/2 s = 1/2 , in regime (R1), it is a transition between consecutive wells that occurs at scale ε \varepsilon ε ; in regime (R2), the sequence additionally rests at the intermediate well on a layer of width of order ε / δ ε \varepsilon/\delta_\varepsilon ε / δ ε . For s ∈ ( 1 / 2 , 1 ) s\in(1/2,1) s ∈ ( 1/2 , 1 ) , each transition is described by an optimal profile problem defined in Section 4. We first define the sequence in one dimension and then extend it to higher dimensions by considering the one-dimensional profile along the normal direction to the jump set.
Finally, in Section 7, we consider the case s < 1 / 2 s<1/2 s < 1/2 with the scaling θ ε : = ε 2 s \theta_\varepsilon:=\varepsilon^{2s} θ ε := ε 2 s . The Γ \Gamma Γ -limit is the non-local energy [ u ] H s ( Q ) 2 [u]_{H^s(Q)}^2 [ u ] H s ( Q ) 2 , plus the bulk term in regime (R1), for functions which take values in Z ∪ S \mathbb{Z}\cup S Z ∪ S (in Z \mathbb{Z} Z for (R2)).
Compactness
In this section we prove Theorem 2.5 . The proof has three steps.
Uniform bounds for the periodic energy. We show that a sequence with sup ε F ε Z ( u ε ) < + ∞ \sup_\varepsilon F_\varepsilon^{\mathbb{Z}}(u_\varepsilon)<+\infty sup ε F ε Z ( u ε ) < + ∞ is bounded in L q ( Q ) L^q(Q) L q ( Q ) , for some q > 1 q>1 q > 1 , after an integer translation.
n ≥ 2 n\geq2 n ≥ 2 , s ∈ [ 1 2 , 1 ) s\in[\frac{1}{2},1) s ∈ [ 2 1 , 1 ) (Proposition 3.7 ): bound in L n n − 1 ( Q ) L^{\frac{n}{n-1}}(Q) L n − 1 n ( Q ) . We extend the argument of [15 ] (n = 2 n=2 n = 2 , s = 1 2 s=\frac{1}{2} s = 2 1 ). Its key ingredient is a lower bound for the non-local interaction between disjoint sets (Lemmas 3.3 and 3.5 ), which gives an isoperimetric-type inequality for the super-level sets of u ε u_\varepsilon u ε .
n = 1 n=1 n = 1 , s > 1 2 s>\frac{1}{2} s > 2 1 (Proposition 3.1 ): bound in L ∞ ( Q ) L^\infty(Q) L ∞ ( Q ) . Every transition between consecutive integers costs at least a fixed amount, so the number of transitions is bounded.
Compactness for the periodic energy (Step 1 in § 3.3 ). By (15) , the truncations of the translated sequence between consecutive integers have bounded energy. Each of them is a two-well problem, hence compact in L 1 L^1 L 1 by [3 ] (n = 1 n=1 n = 1 , s = 1 2 s=\frac{1}{2} s = 2 1 ), [25 ] (n = 1 n=1 n = 1 , s > 1 2 s>\frac{1}{2} s > 2 1 ) and [26 ] (n ≥ 2 n\ge2 n ≥ 2 ). The L q L^q L q bound of 1) controls the tails, and the limit belongs to B V ( Q ; Z ) BV(Q;\mathbb{Z}) B V ( Q ; Z ) .
General potential (Steps 2–5 in § 3.3 ). The piecewise affine homeomorphism ϕ S \phi^S ϕ S , defined in Step 2 , maps Z ∪ S \mathbb{Z}\cup S Z ∪ S onto 1 2 Z \frac{1}{2}\mathbb{Z} 2 1 Z and satisfies F ε ( u ) ≥ d 2 F ε Z ( 2 ϕ S ∘ u ) F_\varepsilon(u)\ge d^2F_\varepsilon^Z(2\phi^S\circ u) F ε ( u ) ≥ d 2 F ε Z ( 2 ϕ S ∘ u ) . Applying compactness to 2 ϕ S ∘ u ε 2\phi^S\circ u_\varepsilon 2 ϕ S ∘ u ε gives compactness of u ε u_\varepsilon u ε up to integer translations. If u ε u_\varepsilon u ε converges, the translations are eventually constant up to a subsequence, as shown in Step 4 , and the limit belongs to B V ( Q ; Z ∪ S ) BV(Q;\mathbb{Z}\cup S) B V ( Q ; Z ∪ S ) . In regime (R2), the lower estimate in (13) forces ∣ { u ∈ S } ∣ = 0 \lvert\{u\in S\}\rvert=0 ∣{ u ∈ S }∣ = 0 .
Uniform L q L^q L q bounds in dimension 1
The uniform bounds in the case s = 1 2 s=\frac{1}{2} s = 2 1 were established by Kurzke in [19 Proposition 2.7] . For the reader’s convenience we recall this result in the following proposition. For s > 1 2 s>\frac{1}{2} s > 2 1 we give a direct proof which yields an L ∞ L^\infty L ∞ estimate.
Proposition 3.1 (Uniform L q L^q L q -estimates with $n=1). Let n = 1 n=1 n = 1 and let I ⊂ R I\subset\mathbb{R} I ⊂ R be a bounded and open interval. Suppose that { u ε } ⊂ L 1 ( I ) \{u_\varepsilon\}\subset L^1(I) { u ε } ⊂ L 1 ( I ) is a sequence with uniformly bounded energy, i.e.,
sup ε F ε Z ( u ε ; I ) < + ∞ . \sup_\varepsilon F_\varepsilon^Z(u_\varepsilon;I)<+\infty. ε sup F ε Z ( u ε ; I ) < + ∞. Then, there exists a sequence of integers { a ε } ⊂ Z \{a_\varepsilon\}\subset\mathbb{Z} { a ε } ⊂ Z such that the translated sequence { u ε − a ε } \{u_\varepsilon-a_\varepsilon\} { u ε − a ε } is uniformly bounded in L q ( I ) L^q(I) L q ( I ) for every 1 ≤ q < ∞ 1\le q<\infty 1 ≤ q < ∞ .
Furthermore, if s ∈ ( 1 2 , 1 ) s\in\left(\frac{1}{2},1\right) s ∈ ( 2 1 , 1 ) , the translated sequence is also uniformly bounded in L ∞ ( I ) L^\infty(I) L ∞ ( I ) .
Proof. We only prove the result for s > 1 2 s>\frac{1}{2} s > 2 1 , since the case s = 1 2 s=\frac{1}{2} s = 2 1 is already proved in [19 ] . For each ε > 0 \varepsilon>0 ε > 0 , we define a translation a ε ∈ Z a_\varepsilon\in\mathbb{Z} a ε ∈ Z as:
a ε : = max { a ∈ Z : ∣ { u ε > a − 1 } ∣ ≥ ∣ I ∣ 4 } . a_\varepsilon:=\max\left\{a\in\mathbb{Z}:\left|\{u_\varepsilon>a-1\}\right|\ge\frac{|I|}{4}\right\}. a ε := max { a ∈ Z : ∣ { u ε > a − 1 } ∣ ≥ 4 ∣ I ∣ } .
This choice ensures that ∣ { u ε ≤ a ε } ∣ ≥ ∣ I ∣ 4 \left|\{u_\varepsilon\le a_\varepsilon\}\right|\ge\frac{|I|}{4} ∣ { u ε ≤ a ε } ∣ ≥ 4 ∣ I ∣ and ∣ { u ε > a ε − 1 } ∣ ≥ ∣ I ∣ 4 \left|\{u_\varepsilon>a_\varepsilon-1\}\right|\ge\frac{|I|}{4} ∣ { u ε > a ε − 1 } ∣ ≥ 4 ∣ I ∣ . Consider the truncated functions
v ε : = ( u ε − a ε ) + and w ε : = ( a ε − 1 − u ε ) + ; v_\varepsilon:=(u_\varepsilon-a_\varepsilon)_+\quad\text{and}\quad w_\varepsilon:=(a_\varepsilon-1-u_\varepsilon)_+; v ε := ( u ε − a ε ) + and w ε := ( a ε − 1 − u ε ) + ;
then ∣ u ε − a ε ∣ ≤ 1 + v ε + w ε |u_\varepsilon-a_\varepsilon|\le1+v_\varepsilon+w_\varepsilon ∣ u ε − a ε ∣ ≤ 1 + v ε + w ε . Since the energy is invariant under integer translations and non-increasing under this truncation operation, we have F ε Z ( v ε ; I ) ≤ F ε Z ( u ε ; I ) F_\varepsilon^Z(v_\varepsilon;I)\le F_\varepsilon^Z(u_\varepsilon;I) F ε Z ( v ε ; I ) ≤ F ε Z ( u ε ; I ) and F ε Z ( w ε ; I ) ≤ F ε Z ( u ε ; I ) F_\varepsilon^Z(w_\varepsilon;I)\le F_\varepsilon^Z(u_\varepsilon;I) F ε Z ( w ε ; I ) ≤ F ε Z ( u ε ; I ) . Thus, { v ε } \{v_\varepsilon\} { v ε } and { w ε } \{w_\varepsilon\} { w ε } are sequences with uniformly bounded energy. Furthermore, we may reduce the problem to the case of monotone functions via rearrangement. Let v ε ∗ v_\varepsilon^* v ε ∗ denote the monotone non-decreasing rearrangement of v ε v_\varepsilon v ε . The potential term ∫ I dist 2 ( v ε , Z ) \int_I\operatorname{dist}^2(v_\varepsilon,\mathbb{Z}) ∫ I dist 2 ( v ε , Z ) is invariant under rearrangement due to equimeasurability, while the fractional Gagliardo seminorm decreases, [ v ε ∗ ] H s ≤ [ v ε ] H s [v_\varepsilon^*]_{H^s}\le[v_\varepsilon]_{H^s} [ v ε ∗ ] H s ≤ [ v ε ] H s (see [17 ] ). Moreover ∥ v ε ∗ ∥ ∞ = ∥ v ε ∥ ∞ \|v_\varepsilon^*\|_\infty=\|v_\varepsilon\|_\infty ∥ v ε ∗ ∥ ∞ = ∥ v ε ∥ ∞ and ∣ { v ε ∗ = 0 } ∣ = ∣ { v ε ≤ a ε } ∣ ≥ ∣ I ∣ / 4 |\{v_\varepsilon^*=0\}|=|\{v_\varepsilon\le a_\varepsilon\}|\ge|I|/4 ∣ { v ε ∗ = 0 } ∣ = ∣ { v ε ≤ a ε } ∣ ≥ ∣ I ∣/4 . An analogous argument applies to w ε w_\varepsilon w ε , with ∣ { w ε ∗ = 0 } ∣ = ∣ { u ε > a ε − 1 } ∣ ≥ ∣ I ∣ / 4 |\{w_\varepsilon^*=0\}|=|\{u_\varepsilon>a_\varepsilon-1\}|\ge|I|/4 ∣ { w ε ∗ = 0 } ∣ = ∣ { u ε > a ε − 1 } ∣ ≥ ∣ I ∣/4 . Therefore, it is sufficient to prove that ∥ u ε ∥ ∞ ≤ C \|u_\varepsilon\|_\infty\le C ∥ u ε ∥ ∞ ≤ C for a sequence { u ε } \{u_\varepsilon\} { u ε } with F ε Z ( u ε ; I ) ≤ M F_\varepsilon^Z(u_\varepsilon;I)\le M F ε Z ( u ε ; I ) ≤ M that satisfies
u ε ≥ 0 , u ε is non-decreasing , and ∣ { u ε = 0 } ∣ ≥ ∣ I ∣ 4 . (16) u_\varepsilon\ge0,\qquad u_\varepsilon\ \text{is non-decreasing},\qquad\text{and}\qquad\left|\{u_\varepsilon=0\}\right|\ge\frac{|I|}{4}. \tag*{(16)} u ε ≥ 0 , u ε is non-decreasing , and ∣ { u ε = 0 } ∣ ≥ 4 ∣ I ∣ . (16)
By the Sobolev embedding, u ε ∈ C 0 , α ( I ‾ ) u_\varepsilon\in C^{0,\alpha}(\overline{I}) u ε ∈ C 0 , α ( I ) with α = s − 1 2 \alpha=s-\frac{1}{2} α = s − 2 1 (see [13 Theorem 8.2] ). For each ε \varepsilon ε we can define
N ε : = ⌊ sup I u ε ⌋ . N_\varepsilon:= \left\lfloor\sup_I u_\varepsilon\right\rfloor. N ε := ⌊ I sup u ε ⌋ .
We want to prove that if F ε Z ( u ε ) ≤ M F_\varepsilon^{\mathbb{Z}}(u_\varepsilon) \le M F ε Z ( u ε ) ≤ M , then N ε ≤ C N_\varepsilon\le C N ε ≤ C . To this end, we show that each transition of u ε u_\varepsilon u ε between two consecutive integers carries a strictly positive cost; hence, if the total energy is finite, such transitions can occur only a finite number of times.
Let u ∈ L ∞ ( I ) u \in L^\infty(I) u ∈ L ∞ ( I ) be a continuous function that satisfies (16) , and let N : = ⌊ sup I u ⌋ N := \lfloor\sup_I u \rfloor N := ⌊ sup I u ⌋ . For each integer 0 ≤ j ≤ N − 1 0 \le j \le N-1 0 ≤ j ≤ N − 1 , define I j = [ x j , y j ] I^j=[x^j,y^j] I j = [ x j , y j ] to be an interval such that
u ( x j ) = j + 1 4 , u ( y j ) = ( j + 1 ) − 1 4 . u(x^j)=j+\frac{1}{4}, \qquad u(y^j)=(j+1)-\frac{1}{4}. u ( x j ) = j + 4 1 , u ( y j ) = ( j + 1 ) − 4 1 .
By monotonicity, we have
j + 1 4 ≤ u ( x ) ≤ ( j + 1 ) − 1 4 , ∀ x ∈ I j , j+\frac{1}{4}\le u(x)\le(j+1)-\frac{1}{4}, \qquad\forall x\in I^j, j + 4 1 ≤ u ( x ) ≤ ( j + 1 ) − 4 1 , ∀ x ∈ I j ,
and the intervals I j I^j I j are pairwise disjoint. Finally, set h j = y j − x j h^j=y^j-x^j h j = y j − x j , the length of the interval I j I^j I j . We have that
[ u ] H s ( I ) 2 ≥ ∑ j = 0 N − 1 ∫ I j ∫ I j ∣ u ( x ) − u ( y ) ∣ 2 ∣ x − y ∣ 1 + 2 s d x d y . [u]^2_{H^s(I)}\ge\sum_{j=0}^{N-1}\int_{I^j}\int_{I^j}\frac{\lvert u(x)-u(y)\rvert^2}{\lvert x-y\rvert^{1+2s}}\,dx\,dy. [ u ] H s ( I ) 2 ≥ j = 0 ∑ N − 1 ∫ I j ∫ I j ∣ x − y ∣ 1 + 2 s ∣ u ( x ) − u ( y ) ∣ 2 d x d y .
We estimate it from below considering the rescaled function
u ^ j ( x ^ ) = u ( x j + h j x ^ ) with x ^ ∈ [ 0 , 1 ] . \widehat{u}_j(\widehat{x})=u(x^j+h^j\widehat{x}) \quad\text{with }\widehat{x}\in[0,1]. u j ( x ) = u ( x j + h j x ) with x ∈ [ 0 , 1 ] .
Observe that
∣ u ^ j ( 1 ) − u ^ j ( 0 ) ∣ = 1 2 . \lvert\widehat{u}_j(1)-\widehat{u}_j(0)\rvert=\frac{1}{2}. ∣ u j ( 1 ) − u j ( 0 )∣ = 2 1 .
By a change of variables we obtain
∫ I j ∫ I j ∣ u ( x ) − u ( y ) ∣ 2 ∣ x − y ∣ 1 + 2 s d x d y = ( h j ) 1 − 2 s [ u ^ j ] H s ( 0 , 1 ) 2 . \int_{I^j}\int_{I^j}\frac{\lvert u(x)-u(y)\rvert^2}{\lvert x-y\rvert^{1+2s}}\,dx\,dy=(h^j)^{1-2s}[\widehat{u}_j]^2_{H^s(0,1)}. ∫ I j ∫ I j ∣ x − y ∣ 1 + 2 s ∣ u ( x ) − u ( y ) ∣ 2 d x d y = ( h j ) 1 − 2 s [ u j ] H s ( 0 , 1 ) 2 .
Now, the function u ^ j \widehat{u}_j u j spans a vertical distance 1 / 2 1/2 1/2 on [ 0 , 1 ] [0,1] [ 0 , 1 ] ; therefore its fractional seminorm is bounded from below by a positive constant C C C depending only on s s s , namely
[ u ^ j ] H s ( 0 , 1 ) 2 ≥ C . [\widehat{u}_j]^2_{H^s(0,1)}\ge C. [ u j ] H s ( 0 , 1 ) 2 ≥ C .
Indeed, by the Sobolev embedding ([13 Theorem 8.2] ) applied to u ^ j − m j \widehat{u}_j-m_j u j − m j , with m j = ∫ 0 1 u ^ j m_j=\int_0^1\widehat{u}_j m j = ∫ 0 1 u j , which has the same increments as u ^ j \widehat{u}_j u j , there exists a constant C H = C H ( s ) > 0 C_H=C_H(s)>0 C H = C H ( s ) > 0 such that
∣ u ^ j ( x ) − u ^ j ( y ) ∣ ≤ C H ∥ u ^ j − m j ∥ H s ( 0 , 1 ) ∣ x − y ∣ s − 1 2 . \lvert\widehat{u}_j(x)-\widehat{u}_j(y)\rvert\le C_H\|\widehat{u}_j-m_j\|_{H^s(0,1)}\lvert x-y\rvert^{s-\frac{1}{2}}. ∣ u j ( x ) − u j ( y )∣ ≤ C H ∥ u j − m j ∥ H s ( 0 , 1 ) ∣ x − y ∣ s − 2 1 .
By the Poincaré inequality for fractional Sobolev spaces (see [21 Theorem 1.17] ), we have
∥ u ^ j − m j ∥ L 2 ( 0 , 1 ) ≤ C P [ u ^ j ] H s ( 0 , 1 ) , \|\widehat{u}_j-m_j\|_{L^2(0,1)}\le C_P[\widehat{u}_j]_{H^s(0,1)}, ∥ u j − m j ∥ L 2 ( 0 , 1 ) ≤ C P [ u j ] H s ( 0 , 1 ) ,
hence, taking x = 1 x=1 x = 1 , y = 0 y=0 y = 0 ,
[ u ^ j ] H s ( 0 , 1 ) 2 ≥ 1 C H 2 ( 1 + C P 2 ) ( 1 2 ) 2 . [\widehat{u}_j]^2_{H^s(0,1)}\ge\frac{1}{C_H^2(1+C_P^2)}\left(\frac{1}{2}\right)^2. [ u j ] H s ( 0 , 1 ) 2 ≥ C H 2 ( 1 + C P 2 ) 1 ( 2 1 ) 2 .
We conclude that
[ u ] H s ( I j ) 2 ≥ c ∗ ( h j ) 1 − 2 s , with c ∗ = 1 4 C H 2 ( 1 + C P 2 ) . [u]^2_{H^s(I^j)}\ge c^*(h^j)^{1-2s}, \qquad\text{with }c^*=\frac{1}{4C_H^2(1+C_P^2)}. [ u ] H s ( I j ) 2 ≥ c ∗ ( h j ) 1 − 2 s , with c ∗ = 4 C H 2 ( 1 + C P 2 ) 1 .
Moreover, since dist ( u , Z ) ≥ 1 4 \operatorname{dist}(u,\mathbb{Z}) \ge\frac{1}{4} dist ( u , Z ) ≥ 4 1 on I j I^j I j , for the potential term we have
1 ε ∫ I dist 2 ( u ( x ) , Z ) d x ≥ 1 16 ε ∑ j = 0 N − 1 h j . \frac{1}{\varepsilon}\int_{I}\operatorname{dist}^{2}(u(x),\mathbb{Z})\,\mathrm{d}x \ge\frac{1}{16\varepsilon}\sum_{j=0}^{N-1}h^j. ε 1 ∫ I dist 2 ( u ( x ) , Z ) d x ≥ 16 ε 1 j = 0 ∑ N − 1 h j .
In conclusion, we obtain
F ε Z ( u ) ≥ ε 2 s − 1 c ∑ j = 0 N − 1 ( h j ) 1 − 2 s + 1 16 ε ∑ j = 0 N − 1 h j ≥ C ( s ) ∑ j = 0 N − 1 ( ε 2 s − 1 ( h j ) 1 − 2 s + 1 ε h j ) . F_{\varepsilon}^{\mathbb{Z}}(u) \ge\varepsilon^{2s-1}c\sum_{j=0}^{N-1}(h^j)^{1-2s}+\frac{1}{16\varepsilon}\sum_{j=0}^{N-1}h^j \ge C(s)\sum_{j=0}^{N-1}\left(\varepsilon^{2s-1}(h^j)^{1-2s}+\frac{1}{\varepsilon}h^j\right). F ε Z ( u ) ≥ ε 2 s − 1 c j = 0 ∑ N − 1 ( h j ) 1 − 2 s + 16 ε 1 j = 0 ∑ N − 1 h j ≥ C ( s ) j = 0 ∑ N − 1 ( ε 2 s − 1 ( h j ) 1 − 2 s + ε 1 h j ) .
By Young’s inequality in the form x 1 2 s y 2 s − 1 2 s ≤ 1 2 s x + 2 s − 1 2 s y ≤ x + y x^{\frac{1}{2s}}y^{\frac{2s-1}{2s}} \le\frac{1}{2s}x+\frac{2s-1}{2s}y \le x+y x 2 s 1 y 2 s 2 s − 1 ≤ 2 s 1 x + 2 s 2 s − 1 y ≤ x + y , we estimate
F ε Z ( u ) ≥ C ( s ) ∑ j = 0 N − 1 ( ε 2 s − 1 ( h j ) 1 − 2 s ) 1 2 s ( h j ε ) 2 s − 1 2 s = C ( s ) N . F_{\varepsilon}^{\mathbb{Z}}(u) \ge C(s)\sum_{j=0}^{N-1}\left(\varepsilon^{2s-1}(h^j)^{1-2s}\right)^{\frac{1}{2s}}\left(\frac{h^j}{\varepsilon}\right)^{\frac{2s-1}{2s}}=C(s)N. F ε Z ( u ) ≥ C ( s ) j = 0 ∑ N − 1 ( ε 2 s − 1 ( h j ) 1 − 2 s ) 2 s 1 ( ε h j ) 2 s 2 s − 1 = C ( s ) N .
Therefore, if F ε Z ( u ε ) ≤ M F_{\varepsilon}^{\mathbb{Z}}(u_\varepsilon)\le M F ε Z ( u ε ) ≤ M , then N ε ≤ C ( M , s ) N_\varepsilon\le C(M,s) N ε ≤ C ( M , s ) and consequently ∥ u ε ∥ ∞ ≤ C ( M , s ) \lVert u_\varepsilon\rVert_\infty\le C(M,s) ∥ u ε ∥ ∞ ≤ C ( M , s ) .
The proof is based on a summability estimate for the super-level sets of the sequence. The heuristic idea is that the energy F ε Z ( u ε ) F_{\varepsilon}^{\mathbb{Z}}(u_\varepsilon) F ε Z ( u ε ) behaves asymptotically like the B V BV B V -seminorm. By comparing consecutive super-level sets, one expects that the perimeter of the set { u ε > k } \{u_\varepsilon>k\} { u ε > k } is controlled by the energy of the truncation T k − 1 u ε : = ( u ε ∨ ( k − 1 ) ) ∧ k T_{k-1}u_\varepsilon:=(u_\varepsilon\vee(k-1))\wedge k T k − 1 u ε := ( u ε ∨ ( k − 1 )) ∧ k . In combination with the isoperimetric inequality, this would formally lead to an estimate of the form
∣ { u ε > k } ∣ n − 1 n ≤ C F ε Z ( T k − 1 u ε ) , \lvert\{u_\varepsilon>k\}\rvert^{\frac{n-1}{n}}\le CF_{\varepsilon}^{\mathbb{Z}}(T_{k-1}u_\varepsilon), ∣{ u ε > k } ∣ n n − 1 ≤ C F ε Z ( T k − 1 u ε ) ,
from which the L n n − 1 L^{\frac{n}{n-1}} L n − 1 n bound would follow. The key tool is a lower bound for the non-local interaction between disjoint sets (Lemmas 3.3 and 3.5), in the spirit of [27 ] .
In what follows B r ( x ) B_r(x) B r ( x ) will denote the ball centered at x x x with radius r r r (we will omit the center when x = 0 x=0 x = 0 and simply write B r B_r B r ). Similarly Q r ( x ) Q_r(x) Q r ( x ) will denote the cube of side r r r centered at x x x . We denote simply by Q r Q_r Q r the cube of side r r r centered at 0 0 0 .
Proposition 3.2 (Proposition 14, [15 ] ). Let n ≥ 2 n\ge2 n ≥ 2 , and let φ ∈ C ∞ ( R n ) \varphi\in C^\infty(\mathbb{R}^n) φ ∈ C ∞ ( R n ) satisfy φ ≥ 0 \varphi\ge0 φ ≥ 0 , φ > 0 \varphi>0 φ > 0 in B 1 / 4 B_{1/4} B 1/4 (and φ \varphi φ even), ∫ φ = 1 \int\varphi=1 ∫ φ = 1 , supp φ ⊆ B 1 / 2 \operatorname{supp}\varphi\subseteq B_{1/2} supp φ ⊆ B 1/2 and set φ δ ( x ) : = δ − n φ ( x / δ ) \varphi_\delta(x):=\delta^{-n}\varphi(x/\delta) φ δ ( x ) := δ − n φ ( x / δ ) .
For every μ ∈ ( 0 , 1 2 ] \mu\in(0,\frac{1}{2}] μ ∈ ( 0 , 2 1 ] there exists c ( μ ) > 0 c(\mu)>0 c ( μ ) > 0 (depending also on n n n and φ \varphi φ ) such that for every measurable set A ⊂ Q 1 A\subset Q_1 A ⊂ Q 1 with μ ≤ ∣ A ∣ ≤ 1 − μ \mu\le\lvert A\rvert\le1-\mu μ ≤ ∣ A ∣ ≤ 1 − μ ,
1 δ ∫ A ∫ Q 1 ∖ A φ δ ( x − y ) d y d x ≥ c ( μ ) , ∀ δ ∈ ( 0 , 1 ] . \frac{1}{\delta}\int_A\int_{Q_1\setminus A}\varphi_\delta(x-y)\,\mathrm{d}y\,\mathrm{d}x\ge c(\mu),\qquad\forall\delta\in(0,1]. δ 1 ∫ A ∫ Q 1 ∖ A φ δ ( x − y ) d y d x ≥ c ( μ ) , ∀ δ ∈ ( 0 , 1 ] . The proof in [15 ] is written for n = 2 n=2 n = 2 but holds verbatim in any dimension.
Lemma 3.3. Let n ≥ 2 n\ge2 n ≥ 2 , and let A A A , B ⊆ Q 1 B\subseteq Q_1 B ⊆ Q 1 be two disjoint subsets. Define E : = Q 1 ∖ ( A ∪ B ) E:=Q_1\setminus(A\cup B) E := Q 1 ∖ ( A ∪ B ) . For every μ ∈ ( 0 , 1 2 ] \mu\in(0,\frac{1}{2}] μ ∈ ( 0 , 2 1 ] , let c ( μ ) c(\mu) c ( μ ) be the constant of Proposition 3.2. There exists C ( μ ) > 0 C(\mu)>0 C ( μ ) > 0 (depending also on φ \varphi φ , the dimension n n n and s s s ) such that, if μ ≤ ∣ A ∣ ≤ 1 − μ \mu\le\lvert A\rvert\le1-\mu μ ≤ ∣ A ∣ ≤ 1 − μ and ∣ E ∣ < c ( μ ) / 2 \lvert E\rvert<c(\mu)/2 ∣ E ∣ < c ( μ ) /2 , then
∫ A ∫ B 1 ∣ x − y ∣ n + 2 s d y d x ≥ { C ( μ ) log c ( μ ) 2 ∣ E ∣ , if s = 1 2 , C ( μ ) ( c ( μ ) ∣ E ∣ ) 2 s − 1 , if s ∈ ( 1 2 , 1 ) . (17) \int_A\int_B\frac{1}{\lvert x-y\rvert^{n+2s}}\,\mathrm{d}y\,\mathrm{d}x\ge
\begin{cases}
C(\mu)\log\frac{c(\mu)}{2\lvert E\rvert}, & \text{if }s=\frac{1}{2},\\
C(\mu)\left(\frac{c(\mu)}{\lvert E\rvert}\right)^{2s-1}, & \text{if }s\in\left(\frac{1}{2},1\right).
\end{cases} \tag*{(17)} ∫ A ∫ B ∣ x − y ∣ n + 2 s 1 d y d x ≥ ⎩ ⎨ ⎧ C ( μ ) log 2 ∣ E ∣ c ( μ ) , C ( μ ) ( ∣ E ∣ c ( μ ) ) 2 s − 1 , if s = 2 1 , if s ∈ ( 2 1 , 1 ) . (17)
Proof. If ∣ E ∣ = 0 |E|=0 ∣ E ∣ = 0 the right-hand side is + ∞ +\infty + ∞ , and the argument below with m = ∞ m=\infty m = ∞ shows that so is the left-hand side. Assume ∣ E ∣ > 0 |E|>0 ∣ E ∣ > 0 . To obtain the estimate (17) , we approximate the singular kernel with a dyadic sum of regular kernels. Let φ δ \varphi_\delta φ δ be as in Proposition 3.2 and assume 2 ∣ E ∣ c ( μ ) ≤ δ ≤ 1 \frac{2|E|}{c(\mu)}\leq\delta\leq1 c ( μ ) 2∣ E ∣ ≤ δ ≤ 1 . Then
1 δ ∬ A × B φ δ ( x − y ) d y d x = 1 δ ∬ A × ( Q 1 ∖ A ) φ δ ( x − y ) d y d x − 1 δ ∬ A × E φ δ ( x − y ) d y d x ≥ c ( μ ) − ∣ E ∣ δ ≥ c ( μ ) 2 . (18) \frac{1}{\delta}\iint_{A\times B}\varphi_\delta(x-y)\,\mathrm{d}y\,\mathrm{d}x=\frac{1}{\delta}\iint_{A\times(Q_1\setminus A)}\varphi_\delta(x-y)\,\mathrm{d}y\,\mathrm{d}x-\frac{1}{\delta}\iint_{A\times E}\varphi_\delta(x-y)\,\mathrm{d}y\,\mathrm{d}x\geq c(\mu)-\frac{|E|}{\delta}\geq\frac{c(\mu)}{2}. \tag*{(18)} δ 1 ∬ A × B φ δ ( x − y ) d y d x = δ 1 ∬ A × ( Q 1 ∖ A ) φ δ ( x − y ) d y d x − δ 1 ∬ A × E φ δ ( x − y ) d y d x ≥ c ( μ ) − δ ∣ E ∣ ≥ 2 c ( μ ) . (18)
Let m ∈ N m\in\mathbb{N} m ∈ N be defined by
m : = ⌊ log 2 c ( μ ) 2 ∣ E ∣ ⌋ , that is 2 − ( m + 1 ) ≤ 2 ∣ E ∣ c ( μ ) ≤ 2 − m , (19) m:=\left\lfloor\log_2\frac{c(\mu)}{2|E|}\right\rfloor,
\qquad\text{that is}\qquad
2^{-(m+1)}\leq\frac{2|E|}{c(\mu)}\leq2^{-m}, \tag*{(19)} m := ⌊ log 2 2∣ E ∣ c ( μ ) ⌋ , that is 2 − ( m + 1 ) ≤ c ( μ ) 2∣ E ∣ ≤ 2 − m , (19)
which is well defined since ∣ E ∣ < c ( μ ) / 2 |E|<c(\mu)/2 ∣ E ∣ < c ( μ ) /2 . Choose δ i = 2 − i \delta_i=2^{-i} δ i = 2 − i for i = 0 , … , m i=0,\ldots,m i = 0 , … , m , so that 2 ∣ E ∣ c ( μ ) ≤ δ m ≤ ⋯ ≤ δ 0 = 1 \frac{2|E|}{c(\mu)}\leq\delta_m\leq\cdots\leq\delta_0=1 c ( μ ) 2∣ E ∣ ≤ δ m ≤ ⋯ ≤ δ 0 = 1 . We claim that there exists a constant C C C (depending on n n n , s s s and φ \varphi φ ) such that
1 ∣ x − y ∣ n + 2 s ≥ C ∑ i = 0 m 1 δ i 2 s φ δ i ( x − y ) , ∀ x , y ∈ Q 1 , x ≠ y . (20) \frac{1}{|x-y|^{n+2s}}\geq C\sum_{i=0}^{m}\frac{1}{\delta_i^{2s}}\varphi_{\delta_i}(x-y),\qquad\forall x,y\in Q_1,\ x\ne y. \tag*{(20)} ∣ x − y ∣ n + 2 s 1 ≥ C i = 0 ∑ m δ i 2 s 1 φ δ i ( x − y ) , ∀ x , y ∈ Q 1 , x = y . (20)
Recall that φ δ ( x − y ) = 0 \varphi_\delta(x-y)=0 φ δ ( x − y ) = 0 whenever ∣ x − y ∣ ≥ δ / 2 |x-y|\geq\delta/2 ∣ x − y ∣ ≥ δ /2 ; hence, if ∣ x − y ∣ ≥ δ 0 = 1 |x-y|\geq\delta_0=1 ∣ x − y ∣ ≥ δ 0 = 1 then every term of the sum vanishes.
If ∣ x − y ∣ < δ m |x-y|<\delta_m ∣ x − y ∣ < δ m , then
∑ i = 0 m 1 δ i 2 s φ δ i ( x − y ) ≤ sup φ ∑ i = 0 m 1 δ i n + 2 s ≤ sup φ 1 δ m n + 2 s ∑ i = 0 ∞ ( 1 2 n + 2 s ) i ≤ C ( n ) sup φ 1 δ m n + 2 s ≤ C ( n , φ ) ∣ x − y ∣ n + 2 s . \sum_{i=0}^{m}\frac{1}{\delta_i^{2s}}\varphi_{\delta_i}(x-y)\leq\sup\varphi\sum_{i=0}^{m}\frac{1}{\delta_i^{n+2s}}\leq\sup\varphi\frac{1}{\delta_m^{n+2s}}\sum_{i=0}^{\infty}\left(\frac{1}{2^{n+2s}}\right)^i\leq C(n)\sup\varphi\frac{1}{\delta_m^{n+2s}}\leq\frac{C(n,\varphi)}{|x-y|^{n+2s}}. i = 0 ∑ m δ i 2 s 1 φ δ i ( x − y ) ≤ sup φ i = 0 ∑ m δ i n + 2 s 1 ≤ sup φ δ m n + 2 s 1 i = 0 ∑ ∞ ( 2 n + 2 s 1 ) i ≤ C ( n ) sup φ δ m n + 2 s 1 ≤ ∣ x − y ∣ n + 2 s C ( n , φ ) .
On the other hand, if δ i 0 + 1 ≤ ∣ x − y ∣ < δ i 0 \delta_{i_0+1}\leq|x-y|<\delta_{i_0} δ i 0 + 1 ≤ ∣ x − y ∣ < δ i 0 for some i 0 ∈ { 0 , … , m − 1 } i_0\in\{0,\ldots,m-1\} i 0 ∈ { 0 , … , m − 1 } , then φ δ i ( x − y ) = 0 \varphi_{\delta_i}(x-y)=0 φ δ i ( x − y ) = 0 for all i ≥ i 0 + 2 i\geq i_0+2 i ≥ i 0 + 2 , hence, since ∣ x − y ∣ < δ i 0 = 2 δ i 0 + 1 |x-y|<\delta_{i_0}=2\delta_{i_0+1} ∣ x − y ∣ < δ i 0 = 2 δ i 0 + 1 ,
∑ i = 0 m 1 δ i 2 s φ δ i ( x − y ) ≤ sup φ ∑ i = 0 i 0 + 1 1 δ i n + 2 s ≤ ∑ i = 0 ∞ ( 1 2 n + 2 s ) i sup φ δ i 0 + 1 n + 2 s ≤ C ( n , φ ) ∣ x − y ∣ n + 2 s . \sum_{i=0}^{m}\frac{1}{\delta_i^{2s}}\varphi_{\delta_i}(x-y)\leq\sup\varphi\sum_{i=0}^{i_0+1}\frac{1}{\delta_i^{n+2s}}\leq\sum_{i=0}^{\infty}\left(\frac{1}{2^{n+2s}}\right)^i\frac{\sup\varphi}{\delta_{i_0+1}^{n+2s}}\leq\frac{C(n,\varphi)}{|x-y|^{n+2s}}. i = 0 ∑ m δ i 2 s 1 φ δ i ( x − y ) ≤ sup φ i = 0 ∑ i 0 + 1 δ i n + 2 s 1 ≤ i = 0 ∑ ∞ ( 2 n + 2 s 1 ) i δ i 0 + 1 n + 2 s sup φ ≤ ∣ x − y ∣ n + 2 s C ( n , φ ) .
This completes the proof of (20) . Integrating (20) over A × B A\times B A × B and using (18) , we obtain
∬ A × B 1 ∣ x − y ∣ n + 2 s d y d x ≥ C ∑ i = 0 m 1 δ i 2 s − 1 ∬ A × B 1 δ i φ δ i ( x − y ) d y d x ≥ C c ( μ ) 2 ∑ i = 0 m ( 2 2 s − 1 ) i . \iint_{A\times B}\frac{1}{|x-y|^{n+2s}}\,\mathrm{d}y\,\mathrm{d}x\geq C\sum_{i=0}^{m}\frac{1}{\delta_i^{2s-1}}\iint_{A\times B}\frac{1}{\delta_i}\varphi_{\delta_i}(x-y)\,\mathrm{d}y\,\mathrm{d}x\geq\frac{Cc(\mu)}{2}\sum_{i=0}^{m}(2^{2s-1})^i. ∬ A × B ∣ x − y ∣ n + 2 s 1 d y d x ≥ C i = 0 ∑ m δ i 2 s − 1 1 ∬ A × B δ i 1 φ δ i ( x − y ) d y d x ≥ 2 C c ( μ ) i = 0 ∑ m ( 2 2 s − 1 ) i .
If s = 1 2 s=\frac12 s = 2 1 the sum equals m + 1 ≥ log 2 c ( μ ) 2 ∣ E ∣ m+1\geq\log_2\frac{c(\mu)}{2|E|} m + 1 ≥ log 2 2∣ E ∣ c ( μ ) , which gives the first line in (17) . If s ∈ ( 1 2 , 1 ) s\in\left(\frac12,1\right) s ∈ ( 2 1 , 1 ) then by (19) ,
∑ i = 0 m ( 2 2 s − 1 ) i ≥ ( 2 m ) 2 s − 1 ≥ ( c ( μ ) 4 ∣ E ∣ ) 2 s − 1 , \sum_{i=0}^{m}(2^{2s-1})^i\geq(2^m)^{2s-1}\geq\left(\frac{c(\mu)}{4|E|}\right)^{2s-1}, i = 0 ∑ m ( 2 2 s − 1 ) i ≥ ( 2 m ) 2 s − 1 ≥ ( 4∣ E ∣ c ( μ ) ) 2 s − 1 ,
which gives the second line of (17) , up to renaming C ( μ ) C(\mu) C ( μ ) . □ \square □
Remark 3.4 (Necessity of the assumption ∣ E ∣ < c ( μ ) / 2 |E|<c(\mu)/2 ∣ E ∣ < c ( μ ) /2 ). For s ∈ ( 1 2 , 1 ) s\in\left(\frac12,1\right) s ∈ ( 2 1 , 1 ) , without the smallness assumption on ∣ E ∣ |E| ∣ E ∣ , the statement is false. Indeed, let A A A be the left half of Q 1 Q_1 Q 1 , let B ⊂ Q 1 B\subset Q_1 B ⊂ Q 1 be a ball of measure η \eta η contained in the right half and at distance at least 1 / 4 1/4 1/4 from A A A , and set E = Q 1 ∖ ( A ∪ B ) E=Q_1\setminus(A\cup B) E = Q 1 ∖ ( A ∪ B ) . Then ∣ E ∣ = 1 / 2 − η |E|=1/2-\eta ∣ E ∣ = 1/2 − η , while
∬ A × B d y d x ∣ x − y ∣ n + 2 s ≤ ∣ A ∣ ∣ B ∣ 4 n + 2 s → η → 0 0. \iint_{A\times B}\frac{\mathrm{d}y\,\mathrm{d}x}{|x-y|^{n+2s}}\leq|A||B|4^{n+2s}\xrightarrow{\eta\to0}0. ∬ A × B ∣ x − y ∣ n + 2 s d y d x ≤ ∣ A ∣∣ B ∣ 4 n + 2 s η → 0 0.
On the other hand, the right-hand side of (17) converges to the positive constant C ( μ ) ( 2 c ( μ ) ) 2 s − 1 C(\mu)(2c(\mu))^{2s-1} C ( μ ) ( 2 c ( μ ) ) 2 s − 1 . Thus, some smallness assumption on ∣ E ∣ |E| ∣ E ∣ (equivalently, a lower bound on ∣ B ∣ |B| ∣ B ∣ ) is necessary. Instead, if s = 1 2 s=\frac{1}{2} s = 2 1 the assumption is unnecessary: if ∣ E ∣ ≥ c ( μ ) / 2 |E|\ge c(\mu)/2 ∣ E ∣ ≥ c ( μ ) /2 , the right-hand side is non-positive and the inequality is trivial. See also the corresponding Lemma 15 in [15 ] .
Lemma 3.5. Let n ≥ 2 n\ge2 n ≥ 2 and s ∈ [ 1 2 , 1 ) s\in\left[\frac{1}{2},1\right) s ∈ [ 2 1 , 1 ) . Let Q Q Q be an open cube in R n \mathbb{R}^{n} R n . For every μ ∈ ( 0 , 1 ) \mu\in(0,1) μ ∈ ( 0 , 1 ) there exist γ 0 > 0 \gamma_{0}>0 γ 0 > 0 and two constants C 1 , C 2 > 0 C_{1},C_{2}>0 C 1 , C 2 > 0 (depending on μ \mu μ , n n n and s s s ) such that for every A , B ⊆ Q A,B\subseteq Q A , B ⊆ Q with A ∩ B = ∅ A\cap B=\varnothing A ∩ B = ∅ , ∣ A ∣ ≤ ( 1 − μ ) ∣ Q ∣ |A|\le(1-\mu)|Q| ∣ A ∣ ≤ ( 1 − μ ) ∣ Q ∣ and
∣ E ∣ ≤ γ 0 ∣ A ∣ , where E : = Q ∖ ( A ∪ B ) , (21) |E|\le\gamma_{0}|A|,\qquad\text{where }E:=Q\setminus(A\cup B), \tag*{(21)} ∣ E ∣ ≤ γ 0 ∣ A ∣ , where E := Q ∖ ( A ∪ B ) , (21) the following inequality holds
∬ A × B 1 ∣ x − y ∣ n + 2 s d y d x ≥ { C 1 ∣ A ∣ n − 1 n log ( C 2 ∣ A ∣ / ∣ E ∣ ) , if s = 1 2 , C 1 ∣ A ∣ n − 2 s n ( ∣ A ∣ / ∣ E ∣ ) 2 s − 1 , if s ∈ ( 1 2 , 1 ) . (22) \iint_{A\times B}\frac{1}{|x-y|^{n+2s}}\,dy\,dx\ge
\begin{cases}
C_{1}|A|^{\frac{n-1}{n}}\log(C_{2}|A|/|E|),&\text{if }s=\frac{1}{2},\\
C_{1}|A|^{\frac{n-2s}{n}}(|A|/|E|)^{2s-1},&\text{if }s\in\left(\frac{1}{2},1\right).
\end{cases} \tag*{(22)} ∬ A × B ∣ x − y ∣ n + 2 s 1 d y d x ≥ { C 1 ∣ A ∣ n n − 1 log ( C 2 ∣ A ∣/∣ E ∣ ) , C 1 ∣ A ∣ n n − 2 s ( ∣ A ∣/∣ E ∣ ) 2 s − 1 , if s = 2 1 , if s ∈ ( 2 1 , 1 ) . (22)
Proof. Assume that ∣ E ∣ , ∣ A ∣ > 0 |E|,|A|>0 ∣ E ∣ , ∣ A ∣ > 0 , otherwise it is trivial. Without loss of generality we may assume Q = ( − 1 / 2 , 1 / 2 ) n Q=(-1/2,1/2)^{n} Q = ( − 1/2 , 1/2 ) n , since both sides of (22) scale as r n − 2 s r^{n-2s} r n − 2 s under x ↦ r x x\mapsto rx x ↦ r x (the ratio ∣ A ∣ / ∣ E ∣ |A|/|E| ∣ A ∣/∣ E ∣ , the volume constraint ∣ A ∣ ≤ ( 1 − μ ) ∣ Q ∣ |A|\le(1-\mu)|Q| ∣ A ∣ ≤ ( 1 − μ ) ∣ Q ∣ and hypothesis (21) being scale invariant); the general case follows by rescaling.
Let Q ′ = Q / 3 Q'=Q/3 Q ′ = Q /3 denote the concentric subcube. We first prove the interior estimate
∬ A × B 1 ∣ x − y ∣ n + 2 s d y d x ≥ { C 1 ∣ A ∩ Q ′ ∣ n − 1 n log ( C 2 ∣ A ∩ Q ′ ∣ / ∣ E ∣ ) , if s = 1 2 , C 1 ∣ A ∩ Q ′ ∣ n − 2 s n ( ∣ A ∩ Q ′ ∣ / ∣ E ∣ ) 2 s − 1 , if s ∈ ( 1 2 , 1 ) , (23) \iint_{A\times B}\frac{1}{|x-y|^{n+2s}}\,dy\,dx\ge
\begin{cases}
C_{1}|A\cap Q'|^{\frac{n-1}{n}}\log(C_{2}|A\cap Q'|/|E|),&\text{if }s=\frac{1}{2},\\
C_{1}|A\cap Q'|^{\frac{n-2s}{n}}(|A\cap Q'|/|E|)^{2s-1},&\text{if }s\in\left(\frac{1}{2},1\right),
\end{cases} \tag*{(23)} ∬ A × B ∣ x − y ∣ n + 2 s 1 d y d x ≥ { C 1 ∣ A ∩ Q ′ ∣ n n − 1 log ( C 2 ∣ A ∩ Q ′ ∣/∣ E ∣ ) , C 1 ∣ A ∩ Q ′ ∣ n n − 2 s ( ∣ A ∩ Q ′ ∣/∣ E ∣ ) 2 s − 1 , if s = 2 1 , if s ∈ ( 2 1 , 1 ) , (23)
under the hypothesis ∣ E ∣ ≤ γ 1 ∣ A ∩ Q ′ ∣ |E|\le\gamma_{1}|A\cap Q'| ∣ E ∣ ≤ γ 1 ∣ A ∩ Q ′ ∣ and ∣ A ∣ ≤ 3 − n − 1 |A|\le3^{-n-1} ∣ A ∣ ≤ 3 − n − 1 . Then we derive the global estimate (22) by a reflection and extension argument.
Set μ ′ : = min { μ , 3 − n − 1 } \mu':=\min\{\mu,3^{-n-1}\} μ ′ := min { μ , 3 − n − 1 } and suppose first that ∣ A ∣ > μ ′ |A|>\mu' ∣ A ∣ > μ ′ . Then μ ′ ≤ ∣ A ∣ ≤ 1 − μ ≤ 1 − μ ′ \mu'\le|A|\le1-\mu\le1-\mu' μ ′ ≤ ∣ A ∣ ≤ 1 − μ ≤ 1 − μ ′ and, by (21) , ∣ E ∣ ≤ γ 0 ∣ A ∣ ≤ γ 0 < c ( μ ′ ) / 2 |E|\le\gamma_{0}|A|\le\gamma_{0}<c(\mu')/2 ∣ E ∣ ≤ γ 0 ∣ A ∣ ≤ γ 0 < c ( μ ′ ) /2 , provided
γ 0 < c ( μ ′ ) 2 . \gamma_{0}<\frac{c(\mu')}{2}. γ 0 < 2 c ( μ ′ ) .
Hence Lemma 3.3 applies and gives (22) at once, since ∣ A ∣ ≤ 1 |A|\le1 ∣ A ∣ ≤ 1 , with C 1 , C 2 C_{1},C_{2} C 1 , C 2 depending only on μ \mu μ (n n n and s s s ).
We may thus assume ∣ A ∣ ≤ 3 − n − 1 |A|\le3^{-n-1} ∣ A ∣ ≤ 3 − n − 1 and we prove (23) by a covering argument.
By scaling the estimate (17) , with μ = 1 4 \mu=\frac{1}{4} μ = 4 1 , we obtain the following: let Q r ⊂ Q Q_{r}\subset Q Q r ⊂ Q be a cube of side r > 0 r>0 r > 0 , and let A ⊆ Q A\subseteq Q A ⊆ Q be such that
1 4 ≤ ∣ A ∩ Q r ∣ ∣ Q r ∣ ≤ 3 4 and ∣ E ∩ Q r ∣ < c ( 1 / 4 ) 2 r n ; \frac{1}{4}\le\frac{|A\cap Q_{r}|}{|Q_{r}|}\le\frac{3}{4}\quad\text{and}\quad|E\cap Q_{r}|<\frac{c(1/4)}{2}r^{n}; 4 1 ≤ ∣ Q r ∣ ∣ A ∩ Q r ∣ ≤ 4 3 and ∣ E ∩ Q r ∣ < 2 c ( 1/4 ) r n ;
then
∬ A ∩ Q r × B ∩ Q r 1 ∣ x − y ∣ n + 2 s d y d x ≥ { C r n − 1 log c r n 2 ∣ E ∩ Q r ∣ , if s = 1 2 , C r n − 2 s ( c r n 2 ∣ E ∩ Q r ∣ ) 2 s − 1 , if s ∈ ( 1 2 , 1 ) , (24) \iint_{A\cap Q_{r}\times B\cap Q_{r}}\frac{1}{|x-y|^{n+2s}}\,dy\,dx\ge
\begin{cases}
Cr^{n-1}\log\frac{cr^{n}}{2|E\cap Q_{r}|},&\text{if }s=\frac{1}{2},\\
Cr^{n-2s}\left(\frac{cr^{n}}{2|E\cap Q_{r}|}\right)^{2s-1},&\text{if }s\in\left(\frac{1}{2},1\right),
\end{cases} \tag*{(24)} ∬ A ∩ Q r × B ∩ Q r ∣ x − y ∣ n + 2 s 1 d y d x ≥ ⎩ ⎨ ⎧ C r n − 1 log 2∣ E ∩ Q r ∣ c r n , C r n − 2 s ( 2∣ E ∩ Q r ∣ c r n ) 2 s − 1 , if s = 2 1 , if s ∈ ( 2 1 , 1 ) , (24)
where c = c ( 1 / 4 ) c=c(1/4) c = c ( 1/4 ) and C > 0 C>0 C > 0 is a constant depending only on s s s and the dimension n n n .
We claim that for almost every x ∈ A ∩ Q ′ x\in A\cap Q' x ∈ A ∩ Q ′ there exists a cube Q r x ( x ) Q_{r_x}(x) Q r x ( x ) , centered at x x x with side length r x ≤ 1 / 3 r_x\le1/3 r x ≤ 1/3 (hence Q r x ( x ) ⊆ Q Q_{r_x}(x)\subseteq Q Q r x ( x ) ⊆ Q ), such that
1 4 ≤ ∣ A ∩ Q r x ( x ) ∣ ∣ Q r x ( x ) ∣ ≤ 3 4 . (25) \frac{1}{4}\le\frac{|A\cap Q_{r_x}(x)|}{|Q_{r_x}(x)|}\le\frac{3}{4}. \tag*{(25)} 4 1 ≤ ∣ Q r x ( x ) ∣ ∣ A ∩ Q r x ( x ) ∣ ≤ 4 3 . (25)
Indeed, for almost every x ∈ A ∩ Q ′ x \in A \cap Q' x ∈ A ∩ Q ′ , the density of A A A at x x x is 1 1 1 , while for the cube of side 1 / 3 1/3 1/3 centered at x x x , i.e., Q 1 / 3 ( x ) ⊂ Q Q_{1/3}(x) \subset Q Q 1/3 ( x ) ⊂ Q , we have
∣ A ∩ Q 1 / 3 ( x ) ∣ ∣ Q 1 / 3 ( x ) ∣ ≤ 3 n ∣ A ∣ ≤ 1 3 . \frac{\lvert A \cap Q_{1/3}(x)\rvert}{\lvert Q_{1/3}(x)\rvert} \le3^n \lvert A\rvert\le\frac{1}{3}. ∣ Q 1/3 ( x )∣ ∣ A ∩ Q 1/3 ( x )∣ ≤ 3 n ∣ A ∣ ≤ 3 1 .
By the continuity of ∣ A ∩ Q r ( x ) ∣ ∣ Q r ( x ) ∣ \frac{\lvert A\cap Q_r(x)\rvert}{\lvert Q_r(x)\rvert} ∣ Q r ( x )∣ ∣ A ∩ Q r ( x )∣ with respect to r r r , this guarantees the existence of an intermediate scale r x r_x r x satisfying (25) .
By the Besicovitch covering theorem we obtain a family of disjoint cubes Q r i ( x i ) Q_{r_i}(x_i) Q r i ( x i ) , i ∈ I i \in I i ∈ I , satisfying (3.10) such that
∑ i ∈ I r i n = ∑ i ∈ I ∣ Q r i ( x i ) ∣ ≥ c ~ ∣ A ∩ Q ′ ∣ , \sum_{i\in I}r_i^n=\sum_{i\in I}\lvert Q_{r_i}(x_i)\rvert\ge\tilde c\lvert A\cap Q'\rvert, i ∈ I ∑ r i n = i ∈ I ∑ ∣ Q r i ( x i )∣ ≥ c ~ ∣ A ∩ Q ′ ∣ ,
where c ~ \tilde c c ~ depends only on n n n . Consider a subfamily J J J corresponding to cubes where E E E has sufficiently low density,
J : = { i ∈ I : ∣ E ∩ Q r i ( x i ) ∣ r i n < 2 c ~ ∣ E ∣ ∣ A ∩ Q ′ ∣ } . (26) J:=\left\{i\in I:\frac{\lvert E\cap Q_{r_i}(x_i)\rvert}{r_i^n}<\frac{2}{\tilde c}\frac{\lvert E\rvert}{\lvert A\cap Q'\rvert}\right\}. \tag*{(26)} J := { i ∈ I : r i n ∣ E ∩ Q r i ( x i )∣ < c ~ 2 ∣ A ∩ Q ′ ∣ ∣ E ∣ } . (26)
We have
∑ i ∈ J r i n = ∑ i ∈ I r i n − ∑ i ∉ J r i n ≥ c ~ ∣ A ∩ Q ′ ∣ − c ~ 2 ∑ i ∈ J ∣ E ∩ Q r i ( x i ) ∣ ∣ E ∣ ∣ A ∩ Q ′ ∣ ≥ c ~ 2 ∣ A ∩ Q ′ ∣ . (27) \sum_{i\in J}r_i^n=\sum_{i\in I}r_i^n-\sum_{i\notin J}r_i^n\ge\tilde c\lvert A\cap Q'\rvert-\frac{\tilde c}{2}\sum_{i\in J}\frac{\lvert E\cap Q_{r_i}(x_i)\rvert}{\lvert E\rvert}\lvert A\cap Q'\rvert\ge\frac{\tilde c}{2}\lvert A\cap Q'\rvert. \tag*{(27)} i ∈ J ∑ r i n = i ∈ I ∑ r i n − i ∈ / J ∑ r i n ≥ c ~ ∣ A ∩ Q ′ ∣ − 2 c ~ i ∈ J ∑ ∣ E ∣ ∣ E ∩ Q r i ( x i )∣ ∣ A ∩ Q ′ ∣ ≥ 2 c ~ ∣ A ∩ Q ′ ∣ . (27)
We now check that (24) is applicable to each cube Q r i ( x i ) Q_{r_i}(x_i) Q r i ( x i ) with i ∈ J i\in J i ∈ J . The density condition is satisfied thanks to (25) ; the estimate on E E E is satisfied by (26) and the hypothesis ∣ E ∣ ≤ γ 1 ∣ A ∩ Q ′ ∣ \lvert E\rvert\le\gamma_1\lvert A\cap Q'\rvert ∣ E ∣ ≤ γ 1 ∣ A ∩ Q ′ ∣ of (23) ; indeed
∣ E ∩ Q r i ∣ r i n < 2 c ~ ∣ E ∣ ∣ A ∩ Q ′ ∣ ≤ 2 γ 1 c ~ ≤ c 2 provided γ 1 ≤ c ~ c 4 . \frac{\lvert E\cap Q_{r_i}\rvert}{r_i^n}<\frac{2}{\tilde c}\frac{\lvert E\rvert}{\lvert A\cap Q'\rvert}\le\frac{2\gamma_1}{\tilde c}\le\frac{c}{2}\quad\text{provided}\quad\gamma_1\le\frac{\tilde c c}{4}. r i n ∣ E ∩ Q r i ∣ < c ~ 2 ∣ A ∩ Q ′ ∣ ∣ E ∣ ≤ c ~ 2 γ 1 ≤ 2 c provided γ 1 ≤ 4 c ~ c .
The same choice guarantees c ~ ∣ A ∩ Q ′ ∣ 4 ∣ E ∣ ≥ 1 \frac{\tilde c\lvert A\cap Q'\rvert}{4\lvert E\rvert}\ge1 4 ∣ E ∣ c ~ ∣ A ∩ Q ′ ∣ ≥ 1 , so that the logarithm below is non-negative. Then we can fix γ 1 : = 3 − n min { c ( μ ′ ) / 4 , c ~ c / 4 } \gamma_1:=3^{-n}\min\{c(\mu')/4,\tilde c c/4\} γ 1 := 3 − n min { c ( μ ′ ) /4 , c ~ c /4 } .
By (24) and (27) we get
∬ A × B d x d y ∣ x − y ∣ n + 2 s ≥ ∑ i ∈ J ∬ ( A ∩ Q r i ) × ( B ∩ Q r i ) d x d y ∣ x − y ∣ n + 2 s ≥ { ∑ i ∈ J C r i n − 1 log ( c r i n 2 ∣ E ∩ Q r i ∣ ) , if s = 1 2 , ∑ i ∈ J C r i n − 2 s ( c r i n 2 ∣ E ∩ Q r i ∣ ) 2 s − 1 , if s ∈ ( 1 2 , 1 ) ≥ { ( ∑ i ∈ J r i n − 1 ) C log ( c ~ ∣ A ∩ Q ′ ∣ 4 ∣ E ∣ ) , ( ∑ i ∈ J r i n − 2 s ) C ( c ~ ∣ A ∩ Q ′ ∣ 4 ∣ E ∣ ) 2 s − 1 ≥ { C ∣ A ∩ Q ′ ∣ n − 1 n log ( c ~ ∣ A ∩ Q ′ ∣ 4 ∣ E ∣ ) , if s = 1 2 , C ∣ A ∩ Q ′ ∣ n − 2 s n ( c ~ ∣ A ∩ Q ′ ∣ 4 ∣ E ∣ ) 2 s − 1 , if s ∈ ( 1 2 , 1 ) , \begin{aligned}
\iint_{A\times B}\frac{dx\,dy}{\lvert x-y\rvert^{n+2s}}
&\ge\sum_{i\in J}\iint_{(A\cap Q_{r_i})\times(B\cap Q_{r_i})}\frac{dx\,dy}{\lvert x-y\rvert^{n+2s}}\\
&\ge
\begin{cases}
\sum_{i\in J}Cr_i^{n-1}\log\left(\frac{cr_i^n}{2\lvert E\cap Q_{r_i}\rvert}\right),&\text{if }s=\frac12,\\[6pt]
\sum_{i\in J}Cr_i^{n-2s}\left(\frac{cr_i^n}{2\lvert E\cap Q_{r_i}\rvert}\right)^{2s-1},&\text{if }s\in\left(\frac12,1\right)
\end{cases}\\
&\ge
\begin{cases}
\left(\sum_{i\in J}r_i^{n-1}\right)C\log\left(\frac{\tilde c\lvert A\cap Q'\rvert}{4\lvert E\rvert}\right),\\[6pt]
\left(\sum_{i\in J}r_i^{n-2s}\right)C\left(\frac{\tilde c\lvert A\cap Q'\rvert}{4\lvert E\rvert}\right)^{2s-1}
\end{cases}\\
&\ge
\begin{cases}
C\lvert A\cap Q'\rvert^{\frac{n-1}{n}}\log\left(\frac{\tilde c\lvert A\cap Q'\rvert}{4\lvert E\rvert}\right),&\text{if }s=\frac12,\\[6pt]
C\lvert A\cap Q'\rvert^{\frac{n-2s}{n}}\left(\frac{\tilde c\lvert A\cap Q'\rvert}{4\lvert E\rvert}\right)^{2s-1},&\text{if }s\in\left(\frac12,1\right),
\end{cases}
\end{aligned} ∬ A × B ∣ x − y ∣ n + 2 s d x d y ≥ i ∈ J ∑ ∬ ( A ∩ Q r i ) × ( B ∩ Q r i ) ∣ x − y ∣ n + 2 s d x d y ≥ ⎩ ⎨ ⎧ ∑ i ∈ J C r i n − 1 log ( 2 ∣ E ∩ Q r i ∣ c r i n ) , ∑ i ∈ J C r i n − 2 s ( 2 ∣ E ∩ Q r i ∣ c r i n ) 2 s − 1 , if s = 2 1 , if s ∈ ( 2 1 , 1 ) ≥ ⎩ ⎨ ⎧ ( ∑ i ∈ J r i n − 1 ) C log ( 4 ∣ E ∣ c ~ ∣ A ∩ Q ′ ∣ ) , ( ∑ i ∈ J r i n − 2 s ) C ( 4 ∣ E ∣ c ~ ∣ A ∩ Q ′ ∣ ) 2 s − 1 ≥ ⎩ ⎨ ⎧ C ∣ A ∩ Q ′ ∣ n n − 1 log ( 4 ∣ E ∣ c ~ ∣ A ∩ Q ′ ∣ ) , C ∣ A ∩ Q ′ ∣ n n − 2 s ( 4 ∣ E ∣ c ~ ∣ A ∩ Q ′ ∣ ) 2 s − 1 , if s = 2 1 , if s ∈ ( 2 1 , 1 ) ,
where we used that
∑ i ∈ J r i k ≥ ( ∑ i ∈ J r i n ) k / n if r i ≥ 0 , 0 < k ≤ n . \sum_{i\in J}r_i^k\ge\left(\sum_{i\in J}r_i^n\right)^{k/n}\quad\text{if }r_i\ge0,\ 0<k\le n. i ∈ J ∑ r i k ≥ ( i ∈ J ∑ r i n ) k / n if r i ≥ 0 , 0 < k ≤ n .
Now, to establish the full estimate (22) , we use a reflection argument as in [15 ] . We extend A A A , B B B , and E E E to the larger cube Q ~ = 3 Q \widetilde{Q}=3Q Q = 3 Q by reflecting them across the hyperplanes x k = ± 1 / 2 x_k=\pm1/2 x k = ± 1/2 for k = 1 , … , n k=1,\ldots,n k = 1 , … , n . Specifically, let I = { − 1 , 0 , 1 } n \mathcal{I}=\{-1,0,1\}^{n} I = { − 1 , 0 , 1 } n be the set of multi-indices. For each α ∈ I \alpha\in\mathcal{I} α ∈ I , we define the reflection π α ( x ) = ( π α 1 ( x 1 ) , … , π α n ( x n ) ) \pi_{\alpha}(x)=(\pi_{\alpha_1}(x_1),\ldots,\pi_{\alpha_n}(x_n)) π α ( x ) = ( π α 1 ( x 1 ) , … , π α n ( x n )) , where π 0 ( t ) = t \pi_0(t)=t π 0 ( t ) = t , π − 1 ( t ) = − 1 − t \pi_{-1}(t)=-1-t π − 1 ( t ) = − 1 − t , and π 1 ( t ) = 1 − t \pi_1(t)=1-t π 1 ( t ) = 1 − t . We then set A ~ = ⋃ α ∈ I π α ( A ) \widetilde{A}=\bigcup_{\alpha\in\mathcal{I}}\pi_{\alpha}(A) A = ⋃ α ∈ I π α ( A ) , and similarly for B ~ \widetilde{B} B and E ~ \widetilde{E} E . Since ∣ π α ( x ) − π β ( y ) ∣ ≥ ∣ x − y ∣ |\pi_{\alpha}(x)-\pi_{\beta}(y)|\geq|x-y| ∣ π α ( x ) − π β ( y ) ∣ ≥ ∣ x − y ∣ for all α , β ∈ I \alpha,\beta\in\mathcal{I} α , β ∈ I and x , y ∈ Q x,y\in Q x , y ∈ Q , the integral over the extended sets satisfies:
∬ A ~ B ~ 1 ∣ x − y ∣ n + 2 s d x d y ≤ 3 2 n ∬ A B 1 ∣ x − y ∣ n + 2 s d x d y . \iint_{\widetilde{A}\,\widetilde{B}}\frac{1}{|x-y|^{n+2s}}\,dx\,dy\leq3^{2n}\iint_{A\,B}\frac{1}{|x-y|^{n+2s}}\,dx\,dy. ∬ A B ∣ x − y ∣ n + 2 s 1 d x d y ≤ 3 2 n ∬ A B ∣ x − y ∣ n + 2 s 1 d x d y .
Moreover, denoting by Q ~ ′ = Q ~ / 3 = Q \widetilde{Q}'=\widetilde{Q}/3=Q Q ′ = Q /3 = Q the subcube concentric to Q ~ \widetilde{Q} Q ,
∣ Q ~ ′ ∩ A ~ ∣ = ∣ A ∣ , ∣ E ~ ∣ = 3 n ∣ E ∣ , ∣ A ~ ∣ = 3 n ∣ A ∣ ≤ 3 n ( 1 − μ ) = ( 1 − μ ) ∣ Q ~ ∣ , |\widetilde{Q}'\cap\widetilde{A}|=|A|,\qquad|\widetilde{E}|=3^n|E|,\qquad|\widetilde{A}|=3^n|A|\leq3^n(1-\mu)=(1-\mu)|\widetilde{Q}|, ∣ Q ′ ∩ A ∣ = ∣ A ∣ , ∣ E ∣ = 3 n ∣ E ∣ , ∣ A ∣ = 3 n ∣ A ∣ ≤ 3 n ( 1 − μ ) = ( 1 − μ ) ∣ Q ∣ ,
so the volume constraints are preserved (∣ A ~ ∣ / ∣ Q ~ ∣ = ∣ A ∣ ≤ 3 − n − 1 |\widetilde{A}|/|\widetilde{Q}|=|A|\leq3^{-n-1} ∣ A ∣/∣ Q ∣ = ∣ A ∣ ≤ 3 − n − 1 ), and the hypothesis ∣ E ∣ ≤ γ 1 ∣ A ∩ Q ′ ∣ |E|\leq\gamma_1|A\cap Q'| ∣ E ∣ ≤ γ 1 ∣ A ∩ Q ′ ∣ , required in (23) , corresponds to 3 − n ∣ E ∣ ≤ γ 1 ∣ A ∣ 3^{-n}|E|\leq\gamma_1|A| 3 − n ∣ E ∣ ≤ γ 1 ∣ A ∣ , which is satisfied thanks to (21) since γ 0 = 3 − n γ 1 \gamma_0=3^{-n}\gamma_1 γ 0 = 3 − n γ 1 . Applying the interior estimate (23) to A ~ , B ~ ⊂ Q ~ \widetilde{A},\widetilde{B}\subset\widetilde{Q} A , B ⊂ Q and rescaling by a factor of 3 3 3 , we obtain
∬ A ~ B ~ d x d y ∣ x − y ∣ n + 2 s ≥ { C ~ 1 ∣ A ~ ∩ Q ~ ∣ n − 1 n log ( C ~ 2 ∣ A ~ ∩ Q ~ ∣ ∣ E ~ ∣ ) s = 1 2 , C ~ 1 ∣ A ~ ∩ Q ~ ∣ n − 2 s n ( ∣ A ~ ∩ Q ~ ∣ ∣ E ~ ∣ ) 2 s − 1 s ∈ ( 1 2 , 1 ) , \iint_{\widetilde{A}\,\widetilde{B}}\frac{dx\,dy}{|x-y|^{n+2s}}\geq
\begin{cases}
\widetilde{C}_1|\widetilde{A}\cap\widetilde{Q}|^{\frac{n-1}{n}}\log\left(\dfrac{\widetilde{C}_2|\widetilde{A}\cap\widetilde{Q}|}{|\widetilde{E}|}\right) & s=\frac{1}{2},\\
\widetilde{C}_1|\widetilde{A}\cap\widetilde{Q}|^{\frac{n-2s}{n}}\left(\dfrac{|\widetilde{A}\cap\widetilde{Q}|}{|\widetilde{E}|}\right)^{2s-1} & s\in\left(\frac{1}{2},1\right),
\end{cases} ∬ A B ∣ x − y ∣ n + 2 s d x d y ≥ ⎩ ⎨ ⎧ C 1 ∣ A ∩ Q ∣ n n − 1 log ( ∣ E ∣ C 2 ∣ A ∩ Q ∣ ) C 1 ∣ A ∩ Q ∣ n n − 2 s ( ∣ E ∣ ∣ A ∩ Q ∣ ) 2 s − 1 s = 2 1 , s ∈ ( 2 1 , 1 ) ,
which, together with the previous inequality, gives (22) with constants C 1 , C 2 C_1,C_2 C 1 , C 2 depending only on n n n , s s s and μ \mu μ . □
Remark 3.6. By Remark 3.4, in Lemma 3.5 , (21) is not necessary for s = 1 2 s=\frac{1}{2} s = 2 1 : it suffices to choose C 2 ≤ γ 0 C_2\leq\gamma_0 C 2 ≤ γ 0 .
We now establish a uniform L n n − 1 L^{\frac{n}{n-1}} L n − 1 n bound for the sequence, up to integer translations. The proof in the case s = 1 2 s=\frac{1}{2} s = 2 1 follows the same lines as the two-dimensional argument presented in [15 Theorem 13] .
Proposition 3.7. Let n ≥ 2 n\geq2 n ≥ 2 and s ∈ [ 1 2 , 1 ) s\in[\frac{1}{2},1) s ∈ [ 2 1 , 1 ) , and let Q ⊂ R n Q\subset\mathbb{R}^n Q ⊂ R n be an open cube. Given M > 0 M>0 M > 0 , there exist ε 0 ∈ ( 0 , 1 ) \varepsilon_0\in(0,1) ε 0 ∈ ( 0 , 1 ) and a constant C ( M , n , Q , s ) C(M,n,Q,s) C ( M , n , Q , s ) such that the following holds. If 0 < ε < ε 0 0<\varepsilon<\varepsilon_0 0 < ε < ε 0 and
F ε Z ( u ε ) ≤ M , F_{\varepsilon}^{\mathbb{Z}}(u_\varepsilon)\leq M, F ε Z ( u ε ) ≤ M , then there exists a ε ∈ Z a_\varepsilon\in\mathbb{Z} a ε ∈ Z such that
∥ u ε − a ε ∥ L n n − 1 ( Q ) ≤ C ( M , n , Q , s ) . \|u_\varepsilon-a_\varepsilon\|_{L^{\frac{n}{n-1}}(Q)}\leq C(M,n,Q,s). ∥ u ε − a ε ∥ L n − 1 n ( Q ) ≤ C ( M , n , Q , s ) .
Proof. Throughout the proof, ε 0 \varepsilon_0 ε 0 is chosen small enough, depending only on M M M , n n n , s s s and σ \sigma σ below. We define a ε = max { a ∈ Z : ∣ { u ε − a > 0 } ∣ ≥ 1 2 ∣ Q ∣ } a_\varepsilon=\max\{a\in\mathbb{Z}:|\{u_\varepsilon-a>0\}|\geq\frac{1}{2}|Q|\} a ε = max { a ∈ Z : ∣ { u ε − a > 0 } ∣ ≥ 2 1 ∣ Q ∣ } . With this choice we have that the translated sequence v ε : = u ε − a ε v_\varepsilon:=u_\varepsilon-a_\varepsilon v ε := u ε − a ε satisfies
∣ { v ε ≥ 0 } ∣ ≥ 1 2 ∣ Q ∣ and ∣ { v ε ≤ 1 } ∣ ≥ 1 2 ∣ Q ∣ . (28) |\{v_\varepsilon\geq0\}|\geq\frac{1}{2}|Q|\qquad\text{and}\qquad|\{v_\varepsilon\leq1\}|\geq\frac{1}{2}|Q|. \tag*{(28)} ∣ { v ε ≥ 0 } ∣ ≥ 2 1 ∣ Q ∣ and ∣ { v ε ≤ 1 } ∣ ≥ 2 1 ∣ Q ∣. (28)
Since F ε Z F_{\varepsilon}^{\mathbb{Z}} F ε Z is invariant under integer translations, F ε Z ( v ε ) = F ε Z ( u ε ) ≤ M F_{\varepsilon}^{\mathbb{Z}}(v_\varepsilon)=F_{\varepsilon}^{\mathbb{Z}}(u_\varepsilon)\leq M F ε Z ( v ε ) = F ε Z ( u ε ) ≤ M . Fix σ ∈ ( 0 , 1 4 ) \sigma\in(0,\frac{1}{4}) σ ∈ ( 0 , 4 1 ) and k ∈ Z k\in\mathbb{Z} k ∈ Z . Fix the following level sets
A k ε = { x ∈ Q : v ε > k + 1 − σ } , B k ε = { x ∈ Q : v ε < k + σ } A_k^\varepsilon=\{x\in Q:v_\varepsilon>k+1-\sigma\},\qquad B_k^\varepsilon=\{x\in Q:v_\varepsilon<k+\sigma\}\qquad A k ε = { x ∈ Q : v ε > k + 1 − σ } , B k ε = { x ∈ Q : v ε < k + σ }
E k ε : = { x ∈ Q : k + σ ≤ v ε ≤ k + 1 − σ } = Q ∖ ( A k ε ∪ B k ε ) . E_k^\varepsilon:= \{x \in Q : k+\sigma\le v_\varepsilon\le k+1-\sigma\}=Q\setminus(A_k^\varepsilon\cup B_k^\varepsilon). E k ε := { x ∈ Q : k + σ ≤ v ε ≤ k + 1 − σ } = Q ∖ ( A k ε ∪ B k ε ) .
Step 1. We prove that there exists C = C ( M , n , Q , σ , s ) C=C(M,n,Q,\sigma,s) C = C ( M , n , Q , σ , s ) such that
∑ k = 0 ∞ ∣ A k ε ∣ n − 1 n ≤ C . (29) \sum_{k=0}^{\infty}|A_k^\varepsilon|^{\frac{n-1}{n}}\le C. \tag*{(29)} k = 0 ∑ ∞ ∣ A k ε ∣ n n − 1 ≤ C . (29)
We treat the positive part, i.e., the case k ≥ 1 k\ge1 k ≥ 1 ; the estimate for the negative part follows by symmetry (applying Lemma 3.5 with the roles of A A A and B B B exchanged, which is legitimate since the kernel is symmetric, although the right-hand side of (22) is not).
Throughout, T k v ε : = ( v ε ∨ k ) ∧ ( k + 1 ) T_kv_\varepsilon:=(v_\varepsilon\vee k)\wedge(k+1) T k v ε := ( v ε ∨ k ) ∧ ( k + 1 ) denotes the truncation at levels k k k and k + 1 k+1 k + 1 , so that ∣ T k v ε ( x ) − T k v ε ( y ) ∣ ≥ 1 − 2 σ |T_kv_\varepsilon(x)-T_kv_\varepsilon(y)|\ge1-2\sigma ∣ T k v ε ( x ) − T k v ε ( y ) ∣ ≥ 1 − 2 σ for x ∈ A k ε x\in A_k^\varepsilon x ∈ A k ε , y ∈ B k ε y\in B_k^\varepsilon y ∈ B k ε . Let us observe that for k ≥ 1 k\ge1 k ≥ 1 one has A k ε ⊂ { v ε > 1 } A_k^\varepsilon\subset\{v_\varepsilon>1\} A k ε ⊂ { v ε > 1 } , so (28) gives ∣ A k ε ∣ ≤ 1 2 ∣ Q ∣ |A_k^\varepsilon|\le\frac12|Q| ∣ A k ε ∣ ≤ 2 1 ∣ Q ∣ : Lemma 3.5 is therefore applicable with μ = 1 / 2 \mu=1/2 μ = 1/2 . (For k ≤ − 1 k\le-1 k ≤ − 1 one argues symmetrically, using B k ε ⊂ { v ε < 0 } B_k^\varepsilon\subset\{v_\varepsilon<0\} B k ε ⊂ { v ε < 0 } .)
Case s = 1 2 s=\frac12 s = 2 1 . By the energy bound, we have ∣ E k ε ∣ ≤ M σ 2 ε ∣ log ε ∣ |E_k^\varepsilon|\le\frac{M}{\sigma^2}\varepsilon|\log\varepsilon| ∣ E k ε ∣ ≤ σ 2 M ε ∣ log ε ∣ . Fix α ∈ ( 0 , 1 / 3 ) \alpha\in(0,1/3) α ∈ ( 0 , 1/3 ) ; then ∣ E k ε ∣ ≤ ε 3 α |E_k^\varepsilon|\le\varepsilon^{3\alpha} ∣ E k ε ∣ ≤ ε 3 α for ε < ε 0 \varepsilon<\varepsilon_0 ε < ε 0 .
∬ A k ε × B k ε d y d x ∣ x − y ∣ n + 1 ≥ C 1 ∣ A k ε ∣ n − 1 n log C 2 ∣ A k ε ∣ ε 3 α . \iint_{A_k^\varepsilon\times B_k^\varepsilon}\frac{dy\,dx}{|x-y|^{n+1}}\ge C_1|A_k^\varepsilon|^{\frac{n-1}{n}}\log\frac{C_2|A_k^\varepsilon|}{\varepsilon^{3\alpha}}. ∬ A k ε × B k ε ∣ x − y ∣ n + 1 d y d x ≥ C 1 ∣ A k ε ∣ n n − 1 log ε 3 α C 2 ∣ A k ε ∣ .
Since ∣ A k ε ∣ > ε 2 α |A_k^\varepsilon|>\varepsilon^{2\alpha} ∣ A k ε ∣ > ε 2 α ,
log C 2 ∣ A k ε ∣ ε 3 α ≥ log C 2 ε 2 α ε 3 α = α ∣ log ε ∣ + log C 2 ≥ α 2 ∣ log ε ∣ \log\frac{C_2|A_k^\varepsilon|}{\varepsilon^{3\alpha}}\ge\log\frac{C_2\varepsilon^{2\alpha}}{\varepsilon^{3\alpha}}=\alpha|\log\varepsilon|+\log C_2\ge\frac{\alpha}{2}|\log\varepsilon| log ε 3 α C 2 ∣ A k ε ∣ ≥ log ε 3 α C 2 ε 2 α = α ∣ log ε ∣ + log C 2 ≥ 2 α ∣ log ε ∣
for ε \varepsilon ε small, whence, we have
∣ A k ε ∣ n − 1 n ≤ C F ε 2 ( T k v ε ) . |A_k^\varepsilon|^{\frac{n-1}{n}}\le C F_\varepsilon^2(T_kv_\varepsilon). ∣ A k ε ∣ n n − 1 ≤ C F ε 2 ( T k v ε ) .
If ∣ A k ε ∣ 1 / 2 ≤ ε α |A_k^\varepsilon|^{1/2}\le\varepsilon^\alpha ∣ A k ε ∣ 1/2 ≤ ε α . Let γ ∈ ( 1 , 2 ) \gamma\in(1,2) γ ∈ ( 1 , 2 ) . We claim that there exists C = C ( n , Q , α , γ , M , ε 0 ) C=C(n,Q,\alpha,\gamma,M,\varepsilon_0) C = C ( n , Q , α , γ , M , ε 0 ) such that
∣ A k ε ∣ n − 1 n ≤ C ( k − σ ) γ . (30) |A_k^\varepsilon|^{\frac{n-1}{n}}\le\frac{C}{(k-\sigma)^\gamma}. \tag*{(30)} ∣ A k ε ∣ n n − 1 ≤ ( k − σ ) γ C . (30)
Indeed by the Sobolev embedding H 1 / 2 ( Q ) ↪ L 2 n n − 1 ( Q ) H^{1/2}(Q)\hookrightarrow L^{\frac{2n}{n-1}}(Q) H 1/2 ( Q ) ↪ L n − 1 2 n ( Q ) (see [13 ] , Theorem 6.7) and the fact that ∣ { v ε ≤ 1 } ∣ ≥ 1 2 ∣ Q ∣ |\{v_\varepsilon\le1\}|\ge\frac12|Q| ∣ { v ε ≤ 1 } ∣ ≥ 2 1 ∣ Q ∣ (that allows us to estimate the L 2 L^2 L 2 norm of ( v ε − 1 ) + (v_\varepsilon-1)_+ ( v ε − 1 ) + with the seminorm), we have
∥ ( v ε − 1 ) + ∥ L 2 n n − 1 ( Q ) ≤ C ( n , Q ) [ ( v ε − 1 ) + ] H 1 / 2 ( Q ) ≤ C ( ∣ log ε ∣ F ε 2 ( v ε ) ) 1 / 2 . \|(v_\varepsilon-1)_+\|_{L^{\frac{2n}{n-1}}(Q)}\le C(n,Q)[(v_\varepsilon-1)_+]_{H^{1/2}(Q)}\le C(|\log\varepsilon|F_\varepsilon^2(v_\varepsilon))^{1/2}. ∥ ( v ε − 1 ) + ∥ L n − 1 2 n ( Q ) ≤ C ( n , Q ) [( v ε − 1 ) + ] H 1/2 ( Q ) ≤ C ( ∣ log ε ∣ F ε 2 ( v ε ) ) 1/2 .
By Hölder’s inequality we get
∣ A k ε ∣ ( k − σ ) ≤ ∫ A k ε ∣ ( v ε − 1 ) + ∣ d x ≤ ∣ A k ε ∣ n + 1 2 n ∥ ( v ε − 1 ) + ∥ L 2 n n − 1 ( Q ) . |A_k^\varepsilon|(k-\sigma)\le\int_{A_k^\varepsilon}|(v_\varepsilon-1)_+|\,dx\le|A_k^\varepsilon|^{\frac{n+1}{2n}}\|(v_\varepsilon-1)_+\|_{L^{\frac{2n}{n-1}}(Q)}. ∣ A k ε ∣ ( k − σ ) ≤ ∫ A k ε ∣ ( v ε − 1 ) + ∣ d x ≤ ∣ A k ε ∣ 2 n n + 1 ∥ ( v ε − 1 ) + ∥ L n − 1 2 n ( Q ) .
Hence
∣ A k ε ∣ n − 1 2 n ≤ C k − σ ( ∣ log ε ∣ F ε 2 ( v ε ) ) 1 / 2 . |A_k^\varepsilon|^{\frac{n-1}{2n}}\le\frac{C}{k-\sigma}(|\log\varepsilon|F_\varepsilon^2(v_\varepsilon))^{1/2}. ∣ A k ε ∣ 2 n n − 1 ≤ k − σ C ( ∣ log ε ∣ F ε 2 ( v ε ) ) 1/2 .
Since ∣ A k ε ∣ 1 / 2 ≤ ε α |A_k^\varepsilon|^{1/2}\le\varepsilon^\alpha ∣ A k ε ∣ 1/2 ≤ ε α and
β : = n − 1 γ n − n − 1 2 n > 0 for γ ∈ ( 1 , 2 ) , \beta:=\frac{n-1}{\gamma n}-\frac{n-1}{2n}>0\qquad\text{for }\gamma\in(1,2), β := γ n n − 1 − 2 n n − 1 > 0 for γ ∈ ( 1 , 2 ) ,
we have that ε 2 α β ∣ log ε ∣ 1 / 2 \varepsilon^{2\alpha\beta}|\log\varepsilon|^{1/2} ε 2 α β ∣ log ε ∣ 1/2 is bounded for ε ≤ ε 0 \varepsilon\le\varepsilon_0 ε ≤ ε 0 . This implies that
∣ A k ε ∣ n − 1 γ n ≤ C ( 1 ( k − σ ) γ F ε Z ( v ε ) ) 1 / 2 , |A_k^\varepsilon|^{\frac{n-1}{\gamma n}} \le C\left(\frac{1}{(k-\sigma)^\gamma}F_\varepsilon^{\mathbb{Z}}(v_\varepsilon)\right)^{1/2}, ∣ A k ε ∣ γ n n − 1 ≤ C ( ( k − σ ) γ 1 F ε Z ( v ε ) ) 1/2 ,
and (30) follows by raising the last inequality to the power γ \gamma γ . In conclusion, fixing α \alpha α and γ \gamma γ as before, we have (3.14) by (2.13); indeed
∑ k = 1 ∞ ∣ A k ε ∣ n − 1 n ≤ C ∑ k = 1 ∞ ( 1 ( k − σ ) γ + F ε Z ( T k v ε ) ) ≤ C ( 1 + F ε Z ( v ε ) ) ≤ C ( 1 + M ) . \sum_{k=1}^{\infty}|A_k^\varepsilon|^{\frac{n-1}{n}} \le C\sum_{k=1}^{\infty}\left(\frac{1}{(k-\sigma)^\gamma}+F_\varepsilon^{\mathbb{Z}}(T_kv_\varepsilon)\right)\le C(1+F_\varepsilon^{\mathbb{Z}}(v_\varepsilon))\le C(1+M). k = 1 ∑ ∞ ∣ A k ε ∣ n n − 1 ≤ C k = 1 ∑ ∞ ( ( k − σ ) γ 1 + F ε Z ( T k v ε ) ) ≤ C ( 1 + F ε Z ( v ε )) ≤ C ( 1 + M ) .
Case s > 1 2 s>\frac{1}{2} s > 2 1 . By the energy bound, we have ∣ E k ε ∣ ≤ M σ 2 ε M σ 2 |E_k^\varepsilon|\le\frac{M}{\sigma^2}\varepsilon^\frac{M}{\sigma^2} ∣ E k ε ∣ ≤ σ 2 M ε σ 2 M .
Fix τ ∈ ( 0 , 1 ) \tau\in(0,1) τ ∈ ( 0 , 1 ) and γ > 1 \gamma>1 γ > 1 such that
τ β s ≥ 2 s − 1 2 , where β s : = n − 1 γ n − n − 2 s 2 n > 0. (31) \tau\beta_s\ge\frac{2s-1}{2},
\qquad\text{where}\qquad
\beta_s:=\frac{n-1}{\gamma n}-\frac{n-2s}{2n}>0. \tag*{(31)} τ β s ≥ 2 2 s − 1 , where β s := γ n n − 1 − 2 n n − 2 s > 0. (31)
Such a choice is possible: as γ → 1 + \gamma\to1^+ γ → 1 + and τ → 1 − \tau\to1^- τ → 1 − one has β s → n − 2 + 2 s 2 n \beta_s\to\frac{n-2+2s}{2n} β s → 2 n n − 2 + 2 s , and
n − 2 + 2 s 2 n > 2 s − 1 2 ⟺ ( n − 2 + 2 s ) − n ( 2 s − 1 ) > 0 ⟺ 2 ( n − 1 ) ( 1 − s ) > 0 , \frac{n-2+2s}{2n}>\frac{2s-1}{2}
\Longleftrightarrow(n-2+2s)-n(2s-1)>0
\Longleftrightarrow2(n-1)(1-s)>0, 2 n n − 2 + 2 s > 2 2 s − 1 ⟺ ( n − 2 + 2 s ) − n ( 2 s − 1 ) > 0 ⟺ 2 ( n − 1 ) ( 1 − s ) > 0 ,
which holds for every n ≥ 2 n\ge2 n ≥ 2 and s ∈ ( 1 2 , 1 ) s\in\left(\frac{1}{2},1\right) s ∈ ( 2 1 , 1 ) .
∣ E k ε ∣ ∣ A k ε ∣ ≤ M σ 2 ε 1 − τ ≤ γ 0 for ε small, \frac{|E_k^\varepsilon|}{|A_k^\varepsilon|}\le\frac{M}{\sigma^2}\varepsilon^{1-\tau}\le\gamma_0
\qquad\text{for }\varepsilon\text{ small,} ∣ A k ε ∣ ∣ E k ε ∣ ≤ σ 2 M ε 1 − τ ≤ γ 0 for ε small,
so (21) holds and Lemma 3.5 applies with μ = 1 2 \mu=\frac{1}{2} μ = 2 1 . Therefore
F ε Z ( T k v ε ) ≥ ( 1 − 2 σ ) 2 ε 2 s − 1 ∫ A k ε ∫ B k ε d y d x ∣ x − y ∣ n + 2 s + 1 ε ∫ E k ε dist 2 ( v ε , Z ) d x ≥ C ε 2 s − 1 ∣ A k ε ∣ n − 2 s n ( ∣ A k ε ∣ ∣ E k ε ∣ ) 2 s − 1 + σ 2 ε ∣ E k ε ∣ = C ε 2 s − 1 ∣ A k ε ∣ 2 s ( n − 1 ) n ∣ E k ε ∣ 1 − 2 s + σ 2 ε ∣ E k ε ∣ . \begin{aligned}
F_\varepsilon^{\mathbb{Z}}(T_kv_\varepsilon)\ge(1-2\sigma)^2\varepsilon^{2s-1}\int_{A_k^\varepsilon}\int_{B_k^\varepsilon}\frac{dy\,dx}{|x-y|^{n+2s}}+\frac{1}{\varepsilon}\int_{E_k^\varepsilon}\operatorname{dist}^2(v_\varepsilon,\mathbb{Z})\,dx \\
&\ge C\varepsilon^{2s-1}|A_k^\varepsilon|^{\frac{n-2s}{n}}\left(\frac{|A_k^\varepsilon|}{|E_k^\varepsilon|}\right)^{2s-1}+\frac{\sigma^2}{\varepsilon}|E_k^\varepsilon| \\
&= C\varepsilon^{2s-1}|A_k^\varepsilon|^{\frac{2s(n-1)}{n}}|E_k^\varepsilon|^{1-2s}+\frac{\sigma^2}{\varepsilon}|E_k^\varepsilon|.
\end{aligned} F ε Z ( T k v ε ) ≥ ( 1 − 2 σ ) 2 ε 2 s − 1 ∫ A k ε ∫ B k ε ∣ x − y ∣ n + 2 s d y d x + ε 1 ∫ E k ε dist 2 ( v ε , Z ) d x ≥ C ε 2 s − 1 ∣ A k ε ∣ n n − 2 s ( ∣ E k ε ∣ ∣ A k ε ∣ ) 2 s − 1 + ε σ 2 ∣ E k ε ∣ = C ε 2 s − 1 ∣ A k ε ∣ n 2 s ( n − 1 ) ∣ E k ε ∣ 1 − 2 s + ε σ 2 ∣ E k ε ∣.
Using Young’s inequality with exponents p = 2 s p=2s p = 2 s and q = 2 s 2 s − 1 q=\frac{2s}{2s-1} q = 2 s − 1 2 s ,
F ε Z ( T k v ε ) ≥ C ( C 1 , σ , s ) ( ε 2 s − 1 ∣ A k ε ∣ 2 s ( n − 1 ) n ∣ E k ε ∣ 1 − 2 s ) 1 2 s ( ∣ E k ε ∣ ε ) 2 s − 1 2 s = C ( C 1 , σ , s ) ∣ A k ε ∣ n − 1 n . F_\varepsilon^{\mathbb{Z}}(T_kv_\varepsilon)\ge C(C_1,\sigma,s)\left(\varepsilon^{2s-1}|A_k^\varepsilon|^{\frac{2s(n-1)}{n}}|E_k^\varepsilon|^{1-2s}\right)^{\frac{1}{2s}}\left(\frac{|E_k^\varepsilon|}{\varepsilon}\right)^{\frac{2s-1}{2s}}=C(C_1,\sigma,s)|A_k^\varepsilon|^{\frac{n-1}{n}}. F ε Z ( T k v ε ) ≥ C ( C 1 , σ , s ) ( ε 2 s − 1 ∣ A k ε ∣ n 2 s ( n − 1 ) ∣ E k ε ∣ 1 − 2 s ) 2 s 1 ( ε ∣ E k ε ∣ ) 2 s 2 s − 1 = C ( C 1 , σ , s ) ∣ A k ε ∣ n n − 1 .
If ∣ A k ε ∣ ≤ ε τ |A_k^\varepsilon|\le\varepsilon^\tau ∣ A k ε ∣ ≤ ε τ , we argue as in the case s = 1 2 s=\frac{1}{2} s = 2 1 . By the Sobolev embedding H s ( Q ) ↪ L 2 n n − 2 s ( Q ) H^s(Q)\hookrightarrow L^{\frac{2n}{n-2s}}(Q) H s ( Q ) ↪ L n − 2 s 2 n ( Q ) (admissible since n ≥ 2 > 2 s n\ge2>2s n ≥ 2 > 2 s ), (28) , the Poincaré inequality for functions vanishing on a set with positive measure, [ ( v ε − 1 ) + ] H s ≤ ε 1 − 2 s F ε Z ( v ε ) [(v_\varepsilon-1)_+]_{H^s}\le\varepsilon^{1-2s}F_\varepsilon^{\mathbb{Z}}(v_\varepsilon) [( v ε − 1 ) + ] H s ≤ ε 1 − 2 s F ε Z ( v ε ) and Hölder’s inequality, we have
∣ A k ε ∣ n − 2 s 2 n ≤ C k − σ ε − 2 s − 1 2 ( F ε Z ( v ε ) ) 1 / 2 . |A_k^\varepsilon|^{\frac{n-2s}{2n}}\le\frac{C}{k-\sigma}\varepsilon^{-\frac{2s-1}{2}}\left(F_\varepsilon^{\mathbb{Z}}(v_\varepsilon)\right)^{1/2}. ∣ A k ε ∣ 2 n n − 2 s ≤ k − σ C ε − 2 2 s − 1 ( F ε Z ( v ε ) ) 1/2 .
Multiplying by ∣ A k ε ∣ β s ≤ ε τ β s |A_k^\varepsilon|^{\beta_s}\le\varepsilon^{\tau\beta_s} ∣ A k ε ∣ β s ≤ ε τ β s and using (31) ,
∣ A k ε ∣ n − 1 γ n ≤ C k − σ ε τ β s − 2 s − 1 2 ( F ε Z ( v ε ) ) 1 / 2 ≤ C M k − σ , |A_k^\varepsilon|^{\frac{n-1}{\gamma n}}\le\frac{C}{k-\sigma}\varepsilon^{\tau\beta_s-\frac{2s-1}{2}}\left(F_\varepsilon^{\mathbb{Z}}(v_\varepsilon)\right)^{1/2}\le\frac{C\sqrt{M}}{k-\sigma}, ∣ A k ε ∣ γ n n − 1 ≤ k − σ C ε τ β s − 2 2 s − 1 ( F ε Z ( v ε ) ) 1/2 ≤ k − σ C M ,
and raising to the power γ \gamma γ we obtain again (3.15).
Combining the two cases and using (15) , we have
∑ k = 1 ∞ ∣ A k ε ∣ n − 1 n ≤ C ∑ k = 1 ∞ ( 1 ( k − σ ) γ + F ε Z ( T k v ε ) ) ≤ C ( 1 + F ε Z ( v ε ) ) ≤ C ( 1 + M ) . \sum_{k=1}^{\infty} \lvert A_k^\varepsilon\rvert^{\frac{n-1}{n}} \le C\sum_{k=1}^{\infty}\left(\frac{1}{(k-\sigma)^\gamma}+F_\varepsilon^{\mathbb{Z}}(T_kv_\varepsilon)\right)\le C\left(1+F_\varepsilon^{\mathbb{Z}}(v_\varepsilon)\right)\le C(1+M). k = 1 ∑ ∞ ∣ A k ε ∣ n n − 1 ≤ C k = 1 ∑ ∞ ( ( k − σ ) γ 1 + F ε Z ( T k v ε ) ) ≤ C ( 1 + F ε Z ( v ε ) ) ≤ C ( 1 + M ) .
Step 2. For any non-increasing sequence { b k } \{b_k\} { b k } of non-negative numbers we have
∑ k = 1 ∞ k 1 n − 1 b k ≤ ( ∑ k = 1 ∞ b k n − 1 n ) n n − 1 . (32) \sum_{k=1}^{\infty} k^{\frac{1}{n-1}}b_k\le\left(\sum_{k=1}^{\infty}b_k^{\frac{n-1}{n}}\right)^{\frac{n}{n-1}}. \tag*{(32)} k = 1 ∑ ∞ k n − 1 1 b k ≤ ( k = 1 ∑ ∞ b k n n − 1 ) n − 1 n . (32)
Indeed
∑ k = 1 ∞ k 1 n − 1 b k = ∑ k = 1 ∞ ( k b k n − 1 n ) 1 n − 1 b k n − 1 n \sum_{k=1}^{\infty}k^{\frac{1}{n-1}}b_k=\sum_{k=1}^{\infty}\left(kb_k^{\frac{n-1}{n}}\right)^{\frac{1}{n-1}}b_k^{\frac{n-1}{n}} k = 1 ∑ ∞ k n − 1 1 b k = k = 1 ∑ ∞ ( k b k n n − 1 ) n − 1 1 b k n n − 1
and
k b k n − 1 n ≤ ∑ j = 1 ∞ b j n − 1 n . kb_k^{\frac{n-1}{n}}\le\sum_{j=1}^{\infty}b_j^{\frac{n-1}{n}}. k b k n n − 1 ≤ j = 1 ∑ ∞ b j n n − 1 .
Thus, we have
∑ k = 1 ∞ k 1 n − 1 b k ≤ ( ∑ j = 1 ∞ b j n − 1 n ) 1 n − 1 ∑ k = 1 ∞ b k n − 1 n = ( ∑ k = 1 ∞ b k n − 1 n ) n n − 1 . \sum_{k=1}^{\infty}k^{\frac{1}{n-1}}b_k\le\left(\sum_{j=1}^{\infty}b_j^{\frac{n-1}{n}}\right)^{\frac{1}{n-1}}\sum_{k=1}^{\infty}b_k^{\frac{n-1}{n}}=\left(\sum_{k=1}^{\infty}b_k^{\frac{n-1}{n}}\right)^{\frac{n}{n-1}}. k = 1 ∑ ∞ k n − 1 1 b k ≤ ( j = 1 ∑ ∞ b j n n − 1 ) n − 1 1 k = 1 ∑ ∞ b k n n − 1 = ( k = 1 ∑ ∞ b k n n − 1 ) n − 1 n .
We apply inequality (32) with b k = ∣ A k − 2 ε ∣ b_k=\lvert A_{k-2}^\varepsilon\rvert b k = ∣ A k − 2 ε ∣ in combination with (29) . This yields
∫ Q ∣ ( v ε ) + ∣ n n − 1 d x = n n − 1 ∫ 0 ∞ t 1 n − 1 ∣ { v ε > t } ∣ d t ≤ 2 ∑ k = 1 ∞ k 1 n − 1 ∣ { v ε > k − 1 } ∣ ≤ 2 ∑ k = 1 ∞ k 1 n − 1 ∣ A k − 2 ε ∣ ≤ C . \int_Q\lvert(v_\varepsilon)_+\rvert^{\frac{n}{n-1}}\,dx=\frac{n}{n-1}\int_0^\infty t^{\frac{1}{n-1}}\lvert\{v_\varepsilon>t\}\rvert\,dt\le2\sum_{k=1}^{\infty}k^{\frac{1}{n-1}}\lvert\{v_\varepsilon>k-1\}\rvert\le2\sum_{k=1}^{\infty}k^{\frac{1}{n-1}}\lvert A_{k-2}^\varepsilon\rvert\le C. ∫ Q ∣( v ε ) + ∣ n − 1 n d x = n − 1 n ∫ 0 ∞ t n − 1 1 ∣{ v ε > t }∣ d t ≤ 2 k = 1 ∑ ∞ k n − 1 1 ∣{ v ε > k − 1 }∣ ≤ 2 k = 1 ∑ ∞ k n − 1 1 ∣ A k − 2 ε ∣ ≤ C .
The claim follows by arguing in a similar way for the negative part of v ε v_\varepsilon v ε .
Proof of the compactness result
The proof follows the scheme described at the beginning of this section.
Proof of Theorem 2.5. Throughout, p = n n − 1 p=\frac{n}{n-1} p = n − 1 n if n ≥ 2 n\ge2 n ≥ 2 and p = 2 p=2 p = 2 if n = 1 n=1 n = 1 .
Step 1. Strong compactness for the periodic case. First we prove that if { u ε } ⊂ L 1 ( Q ) \{u_\varepsilon\}\subset L^1(Q) { u ε } ⊂ L 1 ( Q ) satisfies F ε Z ( u ε ) ≤ M F_\varepsilon^{\mathbb{Z}}(u_\varepsilon)\le M F ε Z ( u ε ) ≤ M , then there exists b ε ∈ Z b_\varepsilon\in\mathbb{Z} b ε ∈ Z such that { u ε − b ε } \{u_\varepsilon-b_\varepsilon\} { u ε − b ε } is precompact in L 1 ( Q ) L^1(Q) L 1 ( Q ) and all its limit points are in B V ( Q ; Z ) BV(Q;\mathbb{Z}) B V ( Q ; Z ) .
Let b ε ∈ Z b_\varepsilon\in\mathbb{Z} b ε ∈ Z be chosen as in Proposition 3.1 if n = 1 n=1 n = 1 and Proposition 3.7 if n ≥ 2 n\ge2 n ≥ 2 , and consider the translated sequence v ε : = u ε − b ε v_\varepsilon:=u_\varepsilon-b_\varepsilon v ε := u ε − b ε . Since F ε Z F_\varepsilon^{\mathbb{Z}} F ε Z is invariant under integer translations, F ε Z ( v ε ) = F ε Z ( u ε ) ≤ M F_\varepsilon^{\mathbb{Z}}(v_\varepsilon)=F_\varepsilon^{\mathbb{Z}}(u_\varepsilon)\le M F ε Z ( v ε ) = F ε Z ( u ε ) ≤ M . This sequence is uniformly bounded in L p L^p L p ; hence, up to extraction of a subsequence, v ε ⇀ v v_\varepsilon\rightharpoonup v v ε ⇀ v weakly in L p ( Q ) L^p(Q) L p ( Q ) . To obtain strong convergence in L 1 L^1 L 1 we use a truncation argument. For all k ∈ Z k\in\mathbb{Z} k ∈ Z , we consider the truncation operator T k u = ( u ∨ k ) ∧ ( k + 1 ) T_ku=(u\vee k)\wedge(k+1) T k u = ( u ∨ k ) ∧ ( k + 1 ) . We have
1 θ ε ( ε 2 s [ T k v ε ] H s ( Q ) 2 + ∫ Q dist 2 ( T k v ε , { k , k + 1 } ) d x ) ≤ F ε Z ( v ε ) ≤ M . \frac{1}{\theta_\varepsilon}\left(\varepsilon^{2s}[T_kv_\varepsilon]^2_{H^s(Q)}+\int_Q\operatorname{dist}^2(T_kv_\varepsilon,\{k,k+1\})\,dx\right)\le F_\varepsilon^{\mathbb{Z}}(v_\varepsilon)\le M. θ ε 1 ( ε 2 s [ T k v ε ] H s ( Q ) 2 + ∫ Q dist 2 ( T k v ε , { k , k + 1 }) d x ) ≤ F ε Z ( v ε ) ≤ M .
By the compactness and lower bound results for double-well potentials (established in [3 ] for s = 1 2 s=\frac{1}{2} s = 2 1 and n = 1 n=1 n = 1 , in [25 Theorem 1.2] for s > 1 2 s>\frac{1}{2} s > 2 1 and n = 1 n=1 n = 1 , and in [26 Theorem 1.3] for n ≥ 2 n\ge2 n ≥ 2 ), for every k ∈ Z k\in\mathbb{Z} k ∈ Z there exists a function u ‾ k ∈ B V ( Q ; { k , k + 1 } ) \overline{u}_k\in BV(Q;\{k,k+1\}) u k ∈ B V ( Q ; { k , k + 1 }) such that, up to the extraction of a subsequence, the sequence T k v ε T_kv_\varepsilon T k v ε converges to u ‾ k \overline{u}_k u k in L 1 ( Q ) L^1(Q) L 1 ( Q ) and
c 0 Per ( { u ‾ k = k + 1 } ; Q ) ≤ lim inf ε → 0 F ε Z ( T k v ε ) , c_0\operatorname{Per}(\{\overline{u}_k=k+1\};Q)\le\liminf_{\varepsilon\to0}F_\varepsilon^{\mathbb{Z}}(T_kv_\varepsilon), c 0 Per ({ u k = k + 1 } ; Q ) ≤ ε → 0 lim inf F ε Z ( T k v ε ) ,
with c 0 > 0 c_0>0 c 0 > 0 depending only on n n n and s s s . By a diagonal argument, the same subsequence works for every k ∈ Z k\in\mathbb{Z} k ∈ Z . Now consider N ∈ N N\in\mathbb{N} N ∈ N and the truncation T N v = ( v ∨ ( − N ) ) ∧ N T^N v=(v\vee(-N))\wedge N T N v = ( v ∨ ( − N )) ∧ N . We can express this new truncation operator in terms of the previous operators T k T_k T k as
T N v = T − N v + ∑ i = − N + 1 N − 1 ( T i v − i ) . T^N v = T_{-N}v + \sum_{i=-N+1}^{N-1}(T_i v-i). T N v = T − N v + i = − N + 1 ∑ N − 1 ( T i v − i ) .
Hence, there exists v N ∈ B V ( Q ; Z ) v^N \in BV(Q;\mathbb{Z}) v N ∈ B V ( Q ; Z ) , namely v N = u ˉ − N + ∑ i = − N + 1 N − 1 ( u ˉ i − i ) v^N=\bar{u}_{-N}+\sum_{i=-N+1}^{N-1}(\bar{u}_i-i) v N = u ˉ − N + ∑ i = − N + 1 N − 1 ( u ˉ i − i ) , such that, up to a subsequence,
T N v ε → v N in L 1 ( Q ) . T^N v_\varepsilon\to v^N \quad\text{in } L^1(Q). T N v ε → v N in L 1 ( Q ) .
Moreover, by Fatou’s lemma and the truncation inequality in (15) ,
∣ D v N ∣ ( Q ) ≤ ∑ k = − N N − 1 Per ( { u ˉ k = k + 1 } ; Q ) ≤ 1 c 0 lim inf ε → 0 ∑ k ∈ Z F ε Z ( T k v ε ) ≤ M c 0 . |Dv^N|(Q) \leq\sum_{k=-N}^{N-1}\operatorname{Per}(\{\bar{u}_k=k+1\};Q) \leq\frac{1}{c_0}\liminf_{\varepsilon\to0}\sum_{k\in\mathbb{Z}}F_\varepsilon^{\mathbb{Z}}(T_kv_\varepsilon)\leq\frac{M}{c_0}. ∣ D v N ∣ ( Q ) ≤ k = − N ∑ N − 1 Per ({ u ˉ k = k + 1 } ; Q ) ≤ c 0 1 ε → 0 lim inf k ∈ Z ∑ F ε Z ( T k v ε ) ≤ c 0 M .
For each fixed N ∈ N N\in\mathbb{N} N ∈ N , we have
∥ v N − v ∥ L 1 ( Q ) ≤ lim inf ε → 0 ∥ T N v ε − v ε ∥ L 1 ( Q ) = lim inf ε → 0 ∫ { ∣ v ε ∣ > N } ( ∣ v ε ∣ − N ) d x ≤ lim inf ε → 0 ∫ { ∣ v ε ∣ > N } ∣ v ε ∣ d x ≤ C N p − 1 . \begin{aligned}
\|v^N-v\|_{L^1(Q)}\leq\liminf_{\varepsilon\to0}\|T^Nv_\varepsilon-v_\varepsilon\|_{L^1(Q)}=\liminf_{\varepsilon\to0}\int_{\{|v_\varepsilon|>N\}}(|v_\varepsilon|-N)\,dx \\
&\leq\liminf_{\varepsilon\to0}\int_{\{|v_\varepsilon|>N\}}|v_\varepsilon|\,dx\leq\frac{C}{N^{p-1}}.
\end{aligned} ∥ v N − v ∥ L 1 ( Q ) ≤ ε → 0 lim inf ∥ T N v ε − v ε ∥ L 1 ( Q ) = ε → 0 lim inf ∫ { ∣ v ε ∣ > N } ( ∣ v ε ∣ − N ) d x ≤ ε → 0 lim inf ∫ { ∣ v ε ∣ > N } ∣ v ε ∣ d x ≤ N p − 1 C .
In the first inequality we used the weak lower semicontinuity of the L 1 L^1 L 1 norm (since T N v ε − v ε ⇀ v N − v T^Nv_\varepsilon-v_\varepsilon\rightharpoonup v^N-v T N v ε − v ε ⇀ v N − v ) and in the last one we used that on the set { ∣ v ε ∣ > N } \{|v_\varepsilon|>N\} { ∣ v ε ∣ > N } we have ∣ v ε ∣ ≤ ∣ v ε ∣ p / N p − 1 |v_\varepsilon|\leq|v_\varepsilon|^p/N^{p-1} ∣ v ε ∣ ≤ ∣ v ε ∣ p / N p − 1 and the L p ( Q ) L^p(Q) L p ( Q ) bound. In conclusion, we have
∥ v ε − v ∥ L 1 ( Q ) ≤ ∥ v ε − T N v ε ∥ L 1 ( Q ) + ∥ T N v ε − v N ∥ L 1 ( Q ) + ∥ v N − v ∥ L 1 ( Q ) ≤ 2 C N p − 1 + ∥ T N v ε − v N ∥ L 1 ( Q ) . \begin{aligned}
\|v_\varepsilon-v\|_{L^1(Q)}\leq\|v_\varepsilon-T^Nv_\varepsilon\|_{L^1(Q)}+\|T^Nv_\varepsilon-v^N\|_{L^1(Q)}+\|v^N-v\|_{L^1(Q)} \\
&\leq\frac{2C}{N^{p-1}}+\|T^Nv_\varepsilon-v^N\|_{L^1(Q)}.
\end{aligned} ∥ v ε − v ∥ L 1 ( Q ) ≤ ∥ v ε − T N v ε ∥ L 1 ( Q ) + ∥ T N v ε − v N ∥ L 1 ( Q ) + ∥ v N − v ∥ L 1 ( Q ) ≤ N p − 1 2 C + ∥ T N v ε − v N ∥ L 1 ( Q ) .
Letting first ε → 0 \varepsilon\to0 ε → 0 and then N → ∞ N\to\infty N → ∞ , we get v ε v_\varepsilon v ε convergent to v v v in L 1 ( Q ) L^1(Q) L 1 ( Q ) . Thus, v N = T N v ∈ B V ( Q ; Z ) v^N=T^Nv\in BV(Q;\mathbb{Z}) v N = T N v ∈ B V ( Q ; Z ) with ∣ D T N v ∣ ( Q ) ≤ M / c 0 |DT^Nv|(Q)\leq M/c_0 ∣ D T N v ∣ ( Q ) ≤ M / c 0 for every N N N , and then v ∈ B V ( Q ; Z ) v\in BV(Q;\mathbb{Z}) v ∈ B V ( Q ; Z ) by lower semicontinuity of the total variation as $N\to\infty.
Step 2. Bi-Lipschitz rectification and extension to the general potential. We now reduce the general potential to the periodic one. Let w z : = s z − z w_z:=s_z-z w z := s z − z . We define the function ϕ S : R → R \phi^S:\mathbb{R}\to\mathbb{R} ϕ S : R → R as follows:
ϕ S ( x ) : = { z + x − z 2 w z if x ∈ [ z , s z ) , z + 1 2 + x − s z 2 ( 1 − w z ) if x ∈ [ s z , z + 1 ) . \phi^S(x):=
\begin{cases}
z+\dfrac{x-z}{2w_z} & \text{if }x\in[z,s_z),\\
z+\dfrac{1}{2}+\dfrac{x-s_z}{2(1-w_z)} & \text{if }x\in[s_z,z+1).
\end{cases} ϕ S ( x ) := ⎩ ⎨ ⎧ z + 2 w z x − z z + 2 1 + 2 ( 1 − w z ) x − s z if x ∈ [ z , s z ) , if x ∈ [ s z , z + 1 ) .
where z = ⌊ x ⌋ z=\lfloor x\rfloor z = ⌊ x ⌋ . Note that ϕ S ( Z ∪ S ) = 1 2 Z \phi^S(\mathbb{Z}\cup S)=\frac{1}{2}\mathbb{Z} ϕ S ( Z ∪ S ) = 2 1 Z and ϕ S \phi^S ϕ S is a strictly increasing continuous function. The function ϕ S \phi^S ϕ S is bi-Lipschitz on R \mathbb{R} R . Specifically by (4) for all x , y ∈ R x,y\in\mathbb{R} x , y ∈ R ,
1 2 ( 1 − d ) ∣ x − y ∣ ≤ ∣ ϕ S ( x ) − ϕ S ( y ) ∣ ≤ 1 2 d ∣ x − y ∣ . \frac{1}{2(1-d)}|x-y|\leq|\phi^S(x)-\phi^S(y)|\leq\frac{1}{2d}|x-y|. 2 ( 1 − d ) 1 ∣ x − y ∣ ≤ ∣ ϕ S ( x ) − ϕ S ( y ) ∣ ≤ 2 d 1 ∣ x − y ∣.
Since 2 ϕ S 2\phi^S 2 ϕ S is 1 d \frac{1}{d} d 1 -Lipschitz and maps Z ∪ S \mathbb{Z}\cup S Z ∪ S onto Z \mathbb{Z} Z , for every t ∈ R t\in\mathbb{R} t ∈ R
dist 2 ( t , Z ∪ S ) ≥ d 2 dist 2 ( 2 ϕ S ( t ) , Z ) , \operatorname{dist}^2(t,\mathbb{Z}\cup S)\geq d^2\operatorname{dist}^2(2\phi^S(t),\mathbb{Z}), dist 2 ( t , Z ∪ S ) ≥ d 2 dist 2 ( 2 ϕ S ( t ) , Z ) ,
and therefore, by the lower estimate in (13) ,
F ε ( u ε ) ≥ F ε Z ∪ S ( u ε ) ≥ d 2 F ε Z ( 2 ϕ S ∘ u ε ) . F_\varepsilon(u_\varepsilon)\geq F_\varepsilon^{\mathbb{Z}\cup S}(u_\varepsilon)\geq d^2F_\varepsilon^{\mathbb{Z}}(2\phi^S\circ u_\varepsilon). F ε ( u ε ) ≥ F ε Z ∪ S ( u ε ) ≥ d 2 F ε Z ( 2 ϕ S ∘ u ε ) .
By Step 1 applied to 2 ϕ S ∘ u ε 2\phi^{S}\circ u_{\varepsilon} 2 ϕ S ∘ u ε , there exists a sequence { b ε } ⊂ Z \{b_{\varepsilon}\}\subset\mathbb{Z} { b ε } ⊂ Z such that, up to a subsequence,
ψ ε : = 2 ϕ S ( u ε ) − b ε ⟶ ψ ∈ B V ( Q ; Z ) strongly in L 1 ( Q ) and a.e. \psi_{\varepsilon}:=2\phi^{S}(u_{\varepsilon})-b_{\varepsilon}\longrightarrow\psi\in BV(Q;\mathbb{Z})\quad\text{strongly in }L^{1}(Q)\text{ and a.e.} ψ ε := 2 ϕ S ( u ε ) − b ε ⟶ ψ ∈ B V ( Q ; Z ) strongly in L 1 ( Q ) and a.e.
Step 3. Compactness up to translations. Write b ε = 2 a ε + r ε b_{\varepsilon}=2a_{\varepsilon}+r_{\varepsilon} b ε = 2 a ε + r ε with a ε ∈ Z a_{\varepsilon}\in\mathbb{Z} a ε ∈ Z and r ε ∈ { 0 , 1 } r_{\varepsilon}\in\{0,1\} r ε ∈ { 0 , 1 } . We prove that { u ε − a ε } \{u_{\varepsilon}-a_{\varepsilon}\} { u ε − a ε } is relatively compact. Up to a subsequence, r ε = r ∈ { 0 , 1 } r_{\varepsilon}=r\in\{0,1\} r ε = r ∈ { 0 , 1 } is constant. For each k ∈ Z k\in\mathbb{Z} k ∈ Z set g k ( y ) : = ( ϕ S ) − 1 ( y + k ) − k g_k(y):=(\phi^{S})^{-1}(y+k)-k g k ( y ) := ( ϕ S ) − 1 ( y + k ) − k . Then
u ε − a ε = g a ε ( y ε ) , y ε : = ψ ε + r 2 ⟶ y : = ψ + r 2 ∈ 1 2 Z a.e. u_{\varepsilon}-a_{\varepsilon}=g_{a_{\varepsilon}}(y_{\varepsilon}),\qquad y_{\varepsilon}:=\frac{\psi_{\varepsilon}+r}{2}\longrightarrow y:=\frac{\psi+r}{2}\in\frac{1}{2}\mathbb{Z}\quad\text{a.e.} u ε − a ε = g a ε ( y ε ) , y ε := 2 ψ ε + r ⟶ y := 2 ψ + r ∈ 2 1 Z a.e.
Note that for every k ∈ Z k\in\mathbb{Z} k ∈ Z , g k ( 0 ) = ( ϕ S ) − 1 ( k ) − k = k − k = 0 g_k(0)=(\phi^{S})^{-1}(k)-k=k-k=0 g k ( 0 ) = ( ϕ S ) − 1 ( k ) − k = k − k = 0 and each g k g_k g k is 2-Lipschitz; hence ∣ g k ( y ) ∣ ≤ 2 ∣ y ∣ |g_k(y)|\leq2|y| ∣ g k ( y ) ∣ ≤ 2∣ y ∣ . In particular ∣ u ε − a ε ∣ ≤ 1 + ∣ ψ ε ∣ |u_{\varepsilon}-a_{\varepsilon}|\leq1+|\psi_{\varepsilon}| ∣ u ε − a ε ∣ ≤ 1 + ∣ ψ ε ∣ , so { u ε − a ε } \{u_{\varepsilon}-a_{\varepsilon}\} { u ε − a ε } satisfies the bounds in the statement. For each m ∈ Z m\in\mathbb{Z} m ∈ Z the sequence { g a ε ( m / 2 ) } \{g_{a_{\varepsilon}}(m/2)\} { g a ε ( m /2 )} is bounded, so up to a subsequence (independent of m m m , by a diagonal argument)
g a ε ( m / 2 ) ⟶ Γ ( m ) ∈ R , for each m ∈ Z . g_{a_{\varepsilon}}(m/2)\longrightarrow\Gamma(m)\in\mathbb{R},\qquad\text{for each }m\in\mathbb{Z}. g a ε ( m /2 ) ⟶ Γ ( m ) ∈ R , for each m ∈ Z .
Thus
u ε − a ε ⟶ Γ ( ψ + r ) strongly in L 1 ( Q ) and a.e. u_{\varepsilon}-a_{\varepsilon}\longrightarrow\Gamma(\psi+r)\quad\text{strongly in }L^{1}(Q)\text{ and a.e.} u ε − a ε ⟶ Γ ( ψ + r ) strongly in L 1 ( Q ) and a.e.
Finally, ∣ Γ ( m + 1 ) − Γ ( m ) ∣ ≤ 1 |\Gamma(m+1)-\Gamma(m)|\leq1 ∣Γ ( m + 1 ) − Γ ( m ) ∣ ≤ 1 because g k g_k g k is 2-Lipschitz. By the coarea formula, since ψ ∈ B V ( Q ; Z ) \psi\in BV(Q;\mathbb{Z}) ψ ∈ B V ( Q ; Z ) , we get
∣ D [ Γ ( ψ + r ) ] ∣ ( Q ) ≤ ∑ j ∈ Z ∣ Γ ( j + 1 + r ) − Γ ( j + r ) ∣ Per ( { ψ > j } ; Q ) ≤ ∣ D ψ ∣ ( Q ) < ∞ . |D[\Gamma(\psi+r)]|(Q)\leq\sum_{j\in\mathbb{Z}}|\Gamma(j+1+r)-\Gamma(j+r)|\operatorname{Per}(\{\psi>j\};Q)\leq|D\psi|(Q)<\infty. ∣ D [ Γ ( ψ + r )] ∣ ( Q ) ≤ j ∈ Z ∑ ∣Γ ( j + 1 + r ) − Γ ( j + r ) ∣ Per ({ ψ > j } ; Q ) ≤ ∣ D ψ ∣ ( Q ) < ∞.
Note that Γ ( 2 z ) = z \Gamma(2z)=z Γ ( 2 z ) = z and Γ ( 2 z + 1 ) = lim ε ( s z + a ε − a ε ) = z + lim ε w z + a ε \Gamma(2z+1)=\lim_{\varepsilon}(s_{z+a_{\varepsilon}}-a_{\varepsilon})=z+\lim_{\varepsilon}w_{z+a_{\varepsilon}} Γ ( 2 z + 1 ) = lim ε ( s z + a ε − a ε ) = z + lim ε w z + a ε , which gives the set S ∗ S^{*} S ∗ of Remark 2.6.
Step 4. Limit in Z ∪ S \mathbb{Z}\cup S Z ∪ S . We want to show that if u ε u_{\varepsilon} u ε converges to u u u in L 1 L^{1} L 1 with F ε ( u ε ) ≤ M F_{\varepsilon}(u_{\varepsilon})\leq M F ε ( u ε ) ≤ M then u ∈ B V ( Q ; Z ∪ S ) u\in BV(Q;\mathbb{Z}\cup S) u ∈ B V ( Q ; Z ∪ S ) . We have
∣ a ε ∣ ∣ Q ∣ ≤ ∥ u ε ∥ L 1 ( Q ) + ∥ u ε − a ε ∥ L 1 ( Q ) . |a_{\varepsilon}||Q|\leq\|u_{\varepsilon}\|_{L^{1}(Q)}+\|u_{\varepsilon}-a_{\varepsilon}\|_{L^{1}(Q)}. ∣ a ε ∣∣ Q ∣ ≤ ∥ u ε ∥ L 1 ( Q ) + ∥ u ε − a ε ∥ L 1 ( Q ) .
and the right-hand side is bounded by the convergence in L 1 L^{1} L 1 of u ε u_{\varepsilon} u ε and Step 3 . Hence, up to a subsequence, a ε = a a_{\varepsilon}=a a ε = a for all ε \varepsilon ε . Then Γ ( m ) = g a ( m / 2 ) \Gamma(m)=g_a(m/2) Γ ( m ) = g a ( m /2 ) and
u = lim ε → 0 u ε = a + g a ( y ) = ( ϕ S ) − 1 ( y + a ) . u=\lim_{\varepsilon\to0}u_{\varepsilon}=a+g_a(y)=(\phi^{S})^{-1}(y+a). u = ε → 0 lim u ε = a + g a ( y ) = ( ϕ S ) − 1 ( y + a ) .
Since y + a ∈ 1 2 Z y+a\in\frac{1}{2}\mathbb{Z} y + a ∈ 2 1 Z a.e. and ( ϕ S ) − 1 ( 1 2 Z ) = Z ∪ S (\phi^{S})^{-1}(\frac{1}{2}\mathbb{Z})=\mathbb{Z}\cup S ( ϕ S ) − 1 ( 2 1 Z ) = Z ∪ S , and u ∈ B V ( Q ) u\in BV(Q) u ∈ B V ( Q ) by Step 3 , we get u ∈ B V ( Q ; Z ∪ S ) u\in BV(Q;\mathbb{Z}\cup S) u ∈ B V ( Q ; Z ∪ S ) .
Step 5. Regime (R2). We want to prove that if u ε → u u_{\varepsilon}\to u u ε → u in L 1 ( Q ) L^{1}(Q) L 1 ( Q ) and δ ε / θ ε → ∞ \delta_{\varepsilon}/\theta_{\varepsilon}\to\infty δ ε / θ ε → ∞ , then u ∈ B V ( Q ; Z ) u\in BV(Q;\mathbb{Z}) u ∈ B V ( Q ; Z ) . Fix δ ∈ ( 0 , 1 ) \delta\in(0,1) δ ∈ ( 0 , 1 ) . By (13) and δ ε / θ ε → ∞ \delta_{\varepsilon}/\theta_{\varepsilon}\to\infty δ ε / θ ε → ∞ we have
∣ A δ ( u ε ) ∣ ≤ M θ ε δ ε ⟶ 0 , |A_{\delta}(u_{\varepsilon})|\leq\frac{M\theta_{\varepsilon}}{\delta_{\varepsilon}}\longrightarrow0, ∣ A δ ( u ε ) ∣ ≤ δ ε M θ ε ⟶ 0 ,
where A δ ( u ε ) A_{\delta}(u_{\varepsilon}) A δ ( u ε ) is defined in (14) . On the other hand, by (4) ,
{ x ∈ Q : u ( x ) ∈ S and ∣ u ε ( x ) − u ( x ) ∣ < δ d / 2 } ⊂ A δ ( u ε ) . \{x\in Q:u(x)\in S\text{ and }|u_{\varepsilon}(x)-u(x)|<\delta d/2\}\subset A_{\delta}(u_{\varepsilon}). { x ∈ Q : u ( x ) ∈ S and ∣ u ε ( x ) − u ( x ) ∣ < δ d /2 } ⊂ A δ ( u ε ) .
Therefore
∣ { u ∈ S } ∣ ≤ ∣ A δ ( u ε ) ∣ + ∣ { ∣ u ε − u ∣ ≥ δ d / 2 } ∣ ⟶ 0 , |\{u\in S\}|\leq|A_{\delta}(u_{\varepsilon})|+|\{|u_{\varepsilon}-u|\geq\delta d/2\}|\longrightarrow0, ∣ { u ∈ S } ∣ ≤ ∣ A δ ( u ε ) ∣ + ∣ { ∣ u ε − u ∣ ≥ δ d /2 } ∣ ⟶ 0 ,
where for the second term we use convergence in measure. Hence ∣ { u ∈ S } ∣ = 0 |\{u\in S\}|=0 ∣ { u ∈ S } ∣ = 0 , and by Step 4 , u ∈ B V ( Q ; Z ) u\in BV(Q;\mathbb{Z}) u ∈ B V ( Q ; Z ) . □ \square □
The optimal profile problem for s > 1 2 s > \frac{1}{2} s > 2 1
In this section, we fix s ∈ ( 1 2 , 1 ) s \in(\frac{1}{2},1) s ∈ ( 2 1 , 1 ) , n ≥ 1 n \ge1 n ≥ 1 and we set W S = dist 2 ( ⋅ , Z ∪ S ) W^{S}=\operatorname{dist}^{2}(\cdot,\mathbb{Z}\cup S) W S = dist 2 ( ⋅ , Z ∪ S ) . We show that the jump cost λ s , n κ s Φ Z ∪ S \lambda_{s,n}\kappa_{s}\Phi_{\mathbb{Z}\cup S} λ s , n κ s Φ Z ∪ S of Definition 2.4 is the cost of the optimal transition for the potential W S W^{S} W S , and introduce the truncated minimizer v T v^{T} v T used for the recovery sequence in Section 6. For v ∈ H l o c s ( R ) v\in H^{s}_{\mathrm{loc}}(\mathbb{R}) v ∈ H loc s ( R ) , let
E s , n ( v ) : = C s , n [ v ] H s ( R ) 2 + ∫ R W S ( v ) d x , \mathcal{E}_{s,n}(v):=C_{s,n}[v]_{H^{s}(\mathbb{R})}^{2}+\int_{\mathbb{R}}W^{S}(v)\,dx, E s , n ( v ) := C s , n [ v ] H s ( R ) 2 + ∫ R W S ( v ) d x ,
where C s , n C_{s,n} C s , n is the constant defined in (9) and [ v ] H s ( R ) [v]_{H^{s}(\mathbb{R})} [ v ] H s ( R ) is the Gagliardo seminorm on R \mathbb{R} R (possibly + ∞ +\infty + ∞ ). For a , b ∈ Z ∪ S a,b\in\mathbb{Z}\cup S a , b ∈ Z ∪ S and T > 0 T>0 T > 0 , we set
Θ s , n ( a , b ) : = inf { E s , n ( v ) : v ∈ H l o c s ( R ) , lim x → − ∞ v ( x ) = a , lim x → + ∞ v ( x ) = b } . (33) \Theta_{s,n}(a,b):=\inf\left\{\mathcal{E}_{s,n}(v):v\in H^{s}_{\mathrm{loc}}(\mathbb{R}),\ \lim_{x\to-\infty}v(x)=a,\ \lim_{x\to+\infty}v(x)=b\right\}. \tag*{(33)} Θ s , n ( a , b ) := inf { E s , n ( v ) : v ∈ H loc s ( R ) , x → − ∞ lim v ( x ) = a , x → + ∞ lim v ( x ) = b } . (33)
and
Θ s , n T ( a , b ) : = inf { E s , n ( v ) : v ∈ H l o c s ( R ) , v = a on ( − ∞ , − T ] , v = b on [ T , + ∞ ) } . (34) \Theta^{T}_{s,n}(a,b):=\inf\left\{\mathcal{E}_{s,n}(v):v\in H^{s}_{\mathrm{loc}}(\mathbb{R}),\ v=a\ \text{on }(-\infty,-T],\ v=b\ \text{on }[T,+\infty)\right\}. \tag*{(34)} Θ s , n T ( a , b ) := inf { E s , n ( v ) : v ∈ H loc s ( R ) , v = a on ( − ∞ , − T ] , v = b on [ T , + ∞ ) } . (34)
We prove the following result, which is the general version of the observation presented in Remark 2.10 for a jump between any two values of Z ∪ S \mathbb{Z}\cup S Z ∪ S .
Proposition 4.1. Let a , b ∈ Z ∪ S a,b\in\mathbb{Z}\cup S a , b ∈ Z ∪ S with a < b a<b a < b . Then
Θ s , n ( b , a ) = Θ s , n ( a , b ) = λ s , n κ s Φ Z ∪ S ( a , b ) , \Theta_{s,n}(b,a)=\Theta_{s,n}(a,b)=\lambda_{s,n}\kappa_{s}\Phi_{\mathbb{Z}\cup S}(a,b), Θ s , n ( b , a ) = Θ s , n ( a , b ) = λ s , n κ s Φ Z ∪ S ( a , b ) , where λ s , n \lambda_{s,n} λ s , n is defined in (9) , κ s \kappa_{s} κ s in (8) and Φ Z ∪ S \Phi_{\mathbb{Z}\cup S} Φ Z ∪ S in Definition 2.1. Moreover, if v T v^{T} v T is a minimizer of (34) (see Lemma 4.2), then
lim T → ∞ E s , n ( v T ) = λ s , n κ s Φ Z ∪ S ( a , b ) . \lim_{T\to\infty}\mathcal{E}_{s,n}(v^{T})=\lambda_{s,n}\kappa_{s}\Phi_{\mathbb{Z}\cup S}(a,b). T → ∞ lim E s , n ( v T ) = λ s , n κ s Φ Z ∪ S ( a , b ) .
The proposition follows from the following properties of Θ s , n \Theta_{s,n} Θ s , n .
Lemma 4.2. Let s ∈ ( 1 2 , 1 ) s\in(\frac{1}{2},1) s ∈ ( 2 1 , 1 ) and a , b ∈ Z ∪ S a,b\in\mathbb{Z}\cup S a , b ∈ Z ∪ S with a < b a<b a < b . The following properties hold.
(1) For any T > 0 T>0 T > 0 , the infimum Θ s , n T ( a , b ) \Theta^{T}_{s,n}(a,b) Θ s , n T ( a , b ) is attained by a non-decreasing minimizer v T ∈ H l o c s ( R ) v^{T}\in H^{s}_{\mathrm{loc}}(\mathbb{R}) v T ∈ H loc s ( R ) , with a ≤ v T ≤ b a\le v^{T}\le b a ≤ v T ≤ b .
(2) The truncated energy recovers the full-line infimum as T → ∞ T\to\infty T → ∞ , that is,
lim T → + ∞ Θ s , n T ( a , b ) = Θ s , n ( a , b ) . \lim_{T\to+\infty}\Theta^{T}_{s,n}(a,b)=\Theta_{s,n}(a,b). T → + ∞ lim Θ s , n T ( a , b ) = Θ s , n ( a , b ) . (3) If there exists c ∈ ( a , b ) ∩ ( Z ∪ S ) c\in(a,b)\cap(\mathbb{Z}\cup S) c ∈ ( a , b ) ∩ ( Z ∪ S ) , then the infimum is additive:
Θ s , n ( a , b ) = Θ s , n ( a , c ) + Θ s , n ( c , b ) . \Theta_{s,n}(a,b)=\Theta_{s,n}(a,c)+\Theta_{s,n}(c,b). Θ s , n ( a , b ) = Θ s , n ( a , c ) + Θ s , n ( c , b ) . (4) If there exists an intermediate well c ∈ ( a , b ) ∩ ( Z ∪ S ) c\in(a,b)\cap(\mathbb{Z}\cup S) c ∈ ( a , b ) ∩ ( Z ∪ S ) , then the infimum Θ s , n ( a , b ) \Theta_{s,n}(a,b) Θ s , n ( a , b ) is not attained, otherwise it is attained.
Proof. (1) Since the truncation ( v ∨ a ) ∧ b (v\vee a)\wedge b ( v ∨ a ) ∧ b does not increase the seminorm and W S ( a ) = W S ( b ) = 0 W^{S}(a)=W^{S}(b)=0 W S ( a ) = W S ( b ) = 0 , we can take a minimizing sequence { v k } \{v_k\} { v k } , for (34) , with a ≤ v k ≤ b a\le v_k\le b a ≤ v k ≤ b . Then { v k } \{v_k\} { v k } is bounded in H s ( ( − 2 T , 2 T ) ) H^{s}((-2T,2T)) H s (( − 2 T , 2 T )) and, since v k v_k v k is fixed outside [ − T , T ] [-T,T] [ − T , T ] , up to a subsequence it converges in L l o c 2 ( R ) L^{2}_{\mathrm{loc}}(\mathbb{R}) L loc 2 ( R ) and a.e. in R \mathbb{R} R . By Fatou’s lemma the limit is a minimizer. Finally, the non-decreasing rearrangement of a minimizer on [ − T , T ] [-T,T] [ − T , T ] is still admissible, has the same potential energy and does not increase the seminorm (see [17 ] ); hence we can take v T v^{T} v T non-decreasing.
(2) Since Θ s , n T ( a , b ) \Theta^{T}_{s,n}(a,b) Θ s , n T ( a , b ) is non-increasing in T T T and is bounded from below by Θ s , n ( a , b ) \Theta_{s,n}(a,b) Θ s , n ( a , b ) , the limit exists and is greater than or equal to Θ s , n ( a , b ) \Theta_{s,n}(a,b) Θ s , n ( a , b ) . The opposite inequality follows by a standard argument for fractional transition layers; we refer to [25 ] , Proposition 3.1] for the detailed construction, which only uses that W S W^S W S vanishes at a a a and b b b and is quadratic near the wells.
(3) Let c ∈ ( a , b ) ∩ ( Z ∪ S ) c \in(a,b) \cap(\mathbb{Z} \cup S) c ∈ ( a , b ) ∩ ( Z ∪ S ) . We first prove that
Θ s , n ( a , b ) ≤ Θ s , n ( a , c ) + Θ s , n ( c , b ) . (35) \Theta_{s,n}(a,b) \le\Theta_{s,n}(a,c) + \Theta_{s,n}(c,b). \tag*{(35)} Θ s , n ( a , b ) ≤ Θ s , n ( a , c ) + Θ s , n ( c , b ) . (35)
Let η > 0 \eta> 0 η > 0 . By definition of the infimum and by the observation in the proof of ( 1 ) (1) ( 1 ) –( 2 ) (2) ( 2 ) , there exist T > 0 T > 0 T > 0 and two non-decreasing functions v 1 , v 2 v_1,v_2 v 1 , v 2 , with a ≤ v 1 ≤ c ≤ v 2 ≤ b a \le v_1 \le c \le v_2 \le b a ≤ v 1 ≤ c ≤ v 2 ≤ b , such that
E s , n ( v 1 ) ≤ Θ s , n ( a , c ) + η , E s , n ( v 2 ) ≤ Θ s , n ( c , b ) + η , \mathcal{E}_{s,n}(v_1) \le\Theta_{s,n}(a,c) + \eta,\qquad\mathcal{E}_{s,n}(v_2) \le\Theta_{s,n}(c,b) + \eta, E s , n ( v 1 ) ≤ Θ s , n ( a , c ) + η , E s , n ( v 2 ) ≤ Θ s , n ( c , b ) + η ,
with
v 1 ( x ) = a for x ≤ − T , v 1 ( x ) = c for x ≥ 0 and v 2 ( x ) = c for x ≤ 0 , v 2 ( x ) = b for x ≥ T . v_1(x)=a\ \text{for }x\le-T,\quad v_1(x)=c\ \text{for }x\ge0\quad\text{and}\quad v_2(x)=c\ \text{for }x\le0,\quad v_2(x)=b\ \text{for }x\ge T. v 1 ( x ) = a for x ≤ − T , v 1 ( x ) = c for x ≥ 0 and v 2 ( x ) = c for x ≤ 0 , v 2 ( x ) = b for x ≥ T .
Given L > 0 L > 0 L > 0 we construct a competitor by placing the two profiles far apart:
g L ( x ) = { v 1 ( x ) , x ≤ 0 , v 2 ( x − L ) , x ≥ 0. g_L(x)=
\begin{cases}
v_1(x), & x\le0,\\
v_2(x-L), & x\ge0.
\end{cases} g L ( x ) = { v 1 ( x ) , v 2 ( x − L ) , x ≤ 0 , x ≥ 0.
The energy decomposes as
E s , n ( g L ) = E s , n ( v 1 ) + E s , n ( v 2 ) + 2 I L , \mathcal{E}_{s,n}(g_L)=\mathcal{E}_{s,n}(v_1)+\mathcal{E}_{s,n}(v_2)+2I_L, E s , n ( g L ) = E s , n ( v 1 ) + E s , n ( v 2 ) + 2 I L ,
where the interaction term between the two profiles is
I L = C s , n ∫ R ∫ R ( v 1 ( x ) − v 1 ( y ) ) ( v 2 ( x − L ) − v 2 ( y − L ) ) ∣ x − y ∣ 1 + 2 s d x d y . I_L=C_{s,n}\int_{\mathbb{R}}\int_{\mathbb{R}}\frac{(v_1(x)-v_1(y))(v_2(x-L)-v_2(y-L))}{\lvert x-y\rvert^{1+2s}}\,dx\,dy. I L = C s , n ∫ R ∫ R ∣ x − y ∣ 1 + 2 s ( v 1 ( x ) − v 1 ( y )) ( v 2 ( x − L ) − v 2 ( y − L )) d x d y .
An easy computation shows that I L I_L I L tends to 0 0 0 as L → ∞ L\to\infty L → ∞ , since s > 1 2 s>\frac{1}{2} s > 2 1 . Since g L g_L g L is admissible for Θ s , n ( a , b ) \Theta_{s,n}(a,b) Θ s , n ( a , b ) , we get
Θ s , n ( a , b ) ≤ E s , n ( g L ) ≤ Θ s , n ( a , c ) + Θ s , n ( c , b ) + 2 η + 2 I L . \Theta_{s,n}(a,b)\le\mathcal{E}_{s,n}(g_L)\le\Theta_{s,n}(a,c)+\Theta_{s,n}(c,b)+2\eta+2I_L. Θ s , n ( a , b ) ≤ E s , n ( g L ) ≤ Θ s , n ( a , c ) + Θ s , n ( c , b ) + 2 η + 2 I L .
Letting L → + ∞ L\to+\infty L → + ∞ and η → 0 \eta\to0 η → 0 we have (35) .
Now we prove the opposite inequality,
Θ s , n ( a , b ) ≥ Θ s , n ( a , c ) + Θ s , n ( c , b ) . (36) \Theta_{s,n}(a,b)\ge\Theta_{s,n}(a,c)+\Theta_{s,n}(c,b). \tag*{(36)} Θ s , n ( a , b ) ≥ Θ s , n ( a , c ) + Θ s , n ( c , b ) . (36)
Let v v v be an admissible function in the definition of Θ s , n ( a , b ) \Theta_{s,n}(a,b) Θ s , n ( a , b ) with finite energy, define
w 1 : = min { v , c } , w 2 : = max { v , c } , w_1:=\min\{v,c\},\qquad w_2:=\max\{v,c\}, w 1 := min { v , c } , w 2 := max { v , c } ,
with w 1 w_1 w 1 admissible for Θ s , n ( a , c ) \Theta_{s,n}(a,c) Θ s , n ( a , c ) and w 2 w_2 w 2 admissible for Θ s , n ( c , b ) \Theta_{s,n}(c,b) Θ s , n ( c , b ) . Moreover, since W S ( c ) = 0 W^S(c)=0 W S ( c ) = 0 ,
∫ R W S ( v ) d x = ∫ R W S ( w 1 ) d x + ∫ R W S ( w 2 ) d x . \int_{\mathbb{R}}W^S(v)\,dx=\int_{\mathbb{R}}W^S(w_1)\,dx+\int_{\mathbb{R}}W^S(w_2)\,dx. ∫ R W S ( v ) d x = ∫ R W S ( w 1 ) d x + ∫ R W S ( w 2 ) d x .
Moreover, for each x , y ∈ R x,y\in\mathbb{R} x , y ∈ R ,
∣ v ( x ) − v ( y ) ∣ 2 = ∣ w 1 ( x ) − w 1 ( y ) ∣ 2 + ∣ w 2 ( x ) − w 2 ( y ) ∣ 2 + 2 ( c − w 1 ( x ) ) ( w 2 ( y ) − c ) + 2 ( c − w 1 ( y ) ) ( w 2 ( x ) − c ) . \lvert v(x)-v(y)\rvert^2=\lvert w_1(x)-w_1(y)\rvert^2+\lvert w_2(x)-w_2(y)\rvert^2+2(c-w_1(x))(w_2(y)-c)+2(c-w_1(y))(w_2(x)-c). ∣ v ( x ) − v ( y ) ∣ 2 = ∣ w 1 ( x ) − w 1 ( y ) ∣ 2 + ∣ w 2 ( x ) − w 2 ( y ) ∣ 2 + 2 ( c − w 1 ( x )) ( w 2 ( y ) − c ) + 2 ( c − w 1 ( y )) ( w 2 ( x ) − c ) .
Hence, we obtain
[ v ] H s 2 = [ w 1 ] H s 2 + [ w 2 ] H s 2 + 4 ∫ R ∫ R ( c − w 1 ( x ) ) ( w 2 ( y ) − c ) ∣ x − y ∣ 1 + 2 s d x d y > [ w 1 ] H s 2 + [ w 2 ] H s 2 , [v]_{H^s}^2=[w_1]_{H^s}^2+[w_2]_{H^s}^2+4\int_{\mathbb{R}}\int_{\mathbb{R}}\frac{(c-w_1(x))(w_2(y)-c)}{\lvert x-y\rvert^{1+2s}}\,dx\,dy>[w_1]_{H^s}^2+[w_2]_{H^s}^2, [ v ] H s 2 = [ w 1 ] H s 2 + [ w 2 ] H s 2 + 4 ∫ R ∫ R ∣ x − y ∣ 1 + 2 s ( c − w 1 ( x )) ( w 2 ( y ) − c ) d x d y > [ w 1 ] H s 2 + [ w 2 ] H s 2 ,
where the strict inequality follows by the fact that c − w 1 > 0 c-w_1>0 c − w 1 > 0 and w 2 − c > 0 w_2-c>0 w 2 − c > 0 on sets with positive measure. Hence
E s , n ( v ) > E s , n ( w 1 ) + E s , n ( w 2 ) ≥ Θ s , n ( a , c ) + Θ s , n ( c , b ) , (37) \mathcal{E}_{s,n}(v)>\mathcal{E}_{s,n}(w_1)+\mathcal{E}_{s,n}(w_2)\ge\Theta_{s,n}(a,c)+\Theta_{s,n}(c,b), \tag*{(37)} E s , n ( v ) > E s , n ( w 1 ) + E s , n ( w 2 ) ≥ Θ s , n ( a , c ) + Θ s , n ( c , b ) , (37)
and taking the infimum over v v v we get (36) .
(4) The strict inequality in (37) shows that the minimum is not achieved if an intermediate well is present. Moreover, if no intermediate wells are present then the infimum Θ s , n ( a , b ) \Theta_{s,n}(a,b) Θ s , n ( a , b ) is attained by a strictly increasing profile (see [25 Proposition 3.3] ). □
Proof of Proposition 4.1. Let a = p 0 < p 1 < ⋯ < p N = b a=p_0<p_1<\cdots<p_N=b a = p 0 < p 1 < ⋯ < p N = b be the points in [ a , b ] ∩ ( Z ∪ S ) [a,b]\cap(\mathbb{Z}\cup S) [ a , b ] ∩ ( Z ∪ S ) ; by Lemma 4.2 (3), we have
Θ s , n ( a , b ) = ∑ i = 0 N − 1 Θ s , n ( p i , p i + 1 ) . \Theta_{s,n}(a,b)=\sum_{i=0}^{N-1}\Theta_{s,n}(p_i,p_{i+1}). Θ s , n ( a , b ) = i = 0 ∑ N − 1 Θ s , n ( p i , p i + 1 ) .
Since truncating at p i p_i p i and p i + 1 p_{i+1} p i + 1 does not increase E s , n \mathcal{E}_{s,n} E s , n , we can rewrite Θ s , n ( p i , p i + 1 ) \Theta_{s,n}(p_i,p_{i+1}) Θ s , n ( p i , p i + 1 ) as
Θ s , n ( p i , p i + 1 ) = inf { C s , n [ v ] H s ( R ) 2 + ∫ R dist 2 ( v , { p i , p i + 1 } ) d x : v ( − ∞ ) = p i , v ( + ∞ ) = p i + 1 } . \Theta_{s,n}(p_i,p_{i+1})=\inf\left\{C_{s,n}[v]^2_{H^s(\mathbb{R})}+\int_{\mathbb{R}}\operatorname{dist}^2(v,\{p_i,p_{i+1}\})\,dx:\ v(-\infty)=p_i,\ v(+\infty)=p_{i+1}\right\}. Θ s , n ( p i , p i + 1 ) = inf { C s , n [ v ] H s ( R ) 2 + ∫ R dist 2 ( v , { p i , p i + 1 }) d x : v ( − ∞ ) = p i , v ( + ∞ ) = p i + 1 } .
Let w ∈ H l o c s ( R ) w\in H^s_{\mathrm{loc}}(\mathbb{R}) w ∈ H loc s ( R ) be an admissible profile for the above problem, we consider the function v ( x ) = p i + ( p i + 1 − p i ) w ( x / λ s , n ) v(x)=p_i+(p_{i+1}-p_i)w(x/\lambda_{s,n}) v ( x ) = p i + ( p i + 1 − p i ) w ( x / λ s , n ) . Since [ w ( ⋅ / λ ) ] H s ( R ) 2 = λ 1 − 2 s [ w ] H s ( R ) 2 [w(\cdot/\lambda)]^2_{H^s(\mathbb{R})}=\lambda^{1-2s}[w]^2_{H^s(\mathbb{R})} [ w ( ⋅ / λ ) ] H s ( R ) 2 = λ 1 − 2 s [ w ] H s ( R ) 2 and C s , n λ s , n 1 − 2 s = λ s , n C_{s,n}\lambda_{s,n}^{1-2s}=\lambda_{s,n} C s , n λ s , n 1 − 2 s = λ s , n , we get
E s , n ( v ) = λ s , n ( p i + 1 − p i ) 2 ( [ w ] H s ( R ) 2 + ∫ R dist 2 ( w , { 0 , 1 } ) d x ) . \mathcal{E}_{s,n}(v)=\lambda_{s,n}(p_{i+1}-p_i)^2\left([w]^2_{H^s(\mathbb{R})}+\int_{\mathbb{R}}\operatorname{dist}^2(w,\{0,1\})\,dx\right). E s , n ( v ) = λ s , n ( p i + 1 − p i ) 2 ( [ w ] H s ( R ) 2 + ∫ R dist 2 ( w , { 0 , 1 }) d x ) .
The map w ↦ v w\mapsto v w ↦ v is a bijection between the admissible profiles from 0 0 0 to 1 1 1 and those from p i p_i p i to p i + 1 p_{i+1} p i + 1 ; hence, by definition of κ s \kappa_s κ s in (8) , Θ s , n ( p i , p i + 1 ) = λ s , n κ s ( p i − p i + 1 ) 2 \Theta_{s,n}(p_i,p_{i+1})=\lambda_{s,n}\kappa_s(p_i-p_{i+1})^2 Θ s , n ( p i , p i + 1 ) = λ s , n κ s ( p i − p i + 1 ) 2 . Finally, E s , n ( v T ) = Θ s , n T ( a , b ) → Θ s , n ( a , b ) \mathcal{E}_{s,n}(v^T)=\Theta^T_{s,n}(a,b)\to\Theta_{s,n}(a,b) E s , n ( v T ) = Θ s , n T ( a , b ) → Θ s , n ( a , b ) by (2) in Lemma 4.2 . □
Lower bound
In this section, we establish the lower bound for the functional F ε F_\varepsilon F ε . In what follows, we set W S : = dist 2 ( ⋅ , Z ∪ S ) W^S:=\operatorname{dist}^2(\cdot,\mathbb{Z}\cup S) W S := dist 2 ( ⋅ , Z ∪ S ) and V \mathcal{V} V the set of wells of the limit (V = Z ∪ S \mathcal{V}=\mathbb{Z}\cup S V = Z ∪ S in regime (R1), V = Z \mathcal{V}=\mathbb{Z} V = Z in regime (R2)). The proof is divided into three steps.
One-dimensional lower bound. For each regime, we prove the lower bound for a single transition between two consecutive points a < b a<b a < b in V \mathcal{V} V . This is the step where the jump cost is computed:
Regime (R1) (Lemma 5.1). By the lowest estimate (13) , it suffices to find the lower bound of the reference energy F ε Z ∪ S F_\varepsilon^{\mathbb{Z}\cup S} F ε Z ∪ S , whose potential W S W^S W S does not depend on ε \varepsilon ε ; this reduction is carried out in § 5.3 . Between two consecutive points of Z ∪ S \mathbb{Z}\cup S Z ∪ S , the potential is the classic double-well potential, whose one-dimensional Γ \Gamma Γ -limit was computed in [3 ] for s = 1 2 s=\frac{1}{2} s = 2 1 and in [26 , 25 ] for s > 1 2 s>\frac{1}{2} s > 2 1 .
Regime (R2), s > 1 2 s>\frac{1}{2} s > 2 1 (Lemma 5.2). By (13) , the analysis reduces to the reference energy F ε Z ∪ S F_\varepsilon^{\mathbb{Z}\cup S} F ε Z ∪ S . A transition from k k k to k + 1 k+1 k + 1 costs as much as a double transition k → s k → k + 1 k\to s_k\to k+1 k → s k → k + 1 . In particular, the height δ ε \delta_\varepsilon δ ε of the well at s k s_k s k plays no role.
Regime (R2), s = 1 2 s=\frac{1}{2} s = 2 1 (Lemma 5.2). This is the genuinely new case. The comparison with F ε Z ∪ S F_\varepsilon^{\mathbb{Z}\cup S} F ε Z ∪ S is still valid, but no longer sharp. Heuristically, an almost optimal transition from k k k to k + 1 k+1 k + 1 rests at s k s_k s k on an interval of length about ε / δ ε \varepsilon/\delta_\varepsilon ε / δ ε . Consequently, the transition cost between the integer minima k k k and k + 1 k+1 k + 1 must be explicitly calculated, leading to a limiting energy that depends on the parameter ℓ \ell ℓ in (7) .
Dimension n n n (Proposition 5.3). We slice in all directions and apply the one-dimensional lemmas on each line. On the slices the constants in front of the seminorm and of the potential are no longer 1 1 1 , so we use a version of the one-dimensional lemmas with arbitrary constants, which follows from them by scaling (Lemma 5.4).
Conclusion (Section 5.3). Truncations between consecutive points of V \mathcal{V} V reduce the general case to a single transition. In regime (R1), the lower estimate in (13) also produces the bulk term K ∣ { u ∈ S } ∣ K\lvert\{u \in S\}\rvert K ∣{ u ∈ S }∣ .
Lower bounds for the one-dimensional case
We first recall the classical lower-bound results for two-well potentials, applied to F ε Z ∪ S F_{\varepsilon}^{\mathbb{Z}\cup S} F ε Z ∪ S .
Lemma 5.1 (Lower bound for F ε Z ∪ S F_{\varepsilon}^{\mathbb{Z}\cup S} F ε Z ∪ S ). Let n = 1 n=1 n = 1 . Let v i v_i v i and v i + 1 v_{i+1} v i + 1 be two consecutive points of Z ∪ S \mathbb{Z}\cup S Z ∪ S . Let I ⊂ R I\subset\mathbb{R} I ⊂ R be a bounded open interval and u ∈ B V ( I ; { v i , v i + 1 } ) u\in BV(I;\{v_i,v_{i+1}\}) u ∈ B V ( I ; { v i , v i + 1 }) . Let { u ε } ⊂ L 1 ( I ; [ v i , v i + 1 ] ) \{u_\varepsilon\}\subset L^1(I;[v_i,v_{i+1}]) { u ε } ⊂ L 1 ( I ; [ v i , v i + 1 ]) be a sequence such that u ε u_\varepsilon u ε converges to u u u in L 1 ( I ) L^1(I) L 1 ( I ) . Then
lim inf ε → 0 F ε Z ∪ S ( u ε ; I ) ≥ Ψ s u p ( v i , v i + 1 ) H 0 ( J u ) = κ s ( v i + 1 − v i ) 2 H 0 ( J u ) . \liminf_{\varepsilon\to0}F_{\varepsilon}^{\mathbb{Z}\cup S}(u_\varepsilon;I)\geq\Psi_{\mathrm{sup}}(v_i,v_{i+1})\mathcal{H}^0(J_u)=\kappa_s(v_{i+1}-v_i)^2\mathcal{H}^0(J_u). ε → 0 lim inf F ε Z ∪ S ( u ε ; I ) ≥ Ψ sup ( v i , v i + 1 ) H 0 ( J u ) = κ s ( v i + 1 − v i ) 2 H 0 ( J u ) . Proof. Since the sequence { u ε } \{u_\varepsilon\} { u ε } takes values in the interval [ v i , v i + 1 ] [v_i,v_{i+1}] [ v i , v i + 1 ] , with v i , v i + 1 v_i,v_{i+1} v i , v i + 1 two consecutive points of Z ∪ S \mathbb{Z}\cup S Z ∪ S , we have dist 2 ( u ε , Z ∪ S ) = dist 2 ( u ε , { v i , v i + 1 } ) \operatorname{dist}^2(u_\varepsilon,\mathbb{Z}\cup S)=\operatorname{dist}^2(u_\varepsilon,\{v_i,v_{i+1}\}) dist 2 ( u ε , Z ∪ S ) = dist 2 ( u ε , { v i , v i + 1 }) . Then, the claim is the classical one-dimensional lower bound for double-well potentials.
Case s > 1 2 s>\frac{1}{2} s > 2 1 . This is the fractional counterpart of the Modica–Mortola theorem, proved in [26 , 25 ] ; the lower bound is given by the optimal transition cost Θ s , 1 ( v i , v i + 1 ) \Theta_{s,1}(v_i,v_{i+1}) Θ s , 1 ( v i , v i + 1 ) defined in (33) , as shown in [25 Proposition 5.1] . Hence the thesis follows by Proposition 4.1.
Case s = 1 2 s=\frac{1}{2} s = 2 1 . The lower bound 2 ( v i + 1 − v i ) 2 H 0 ( J u ) 2(v_{i+1}-v_i)^2\mathcal{H}^0(J_u) 2 ( v i + 1 − v i ) 2 H 0 ( J u ) is proved in [3 ] , and κ 1 / 2 = 2 \kappa_{1/2}=2 κ 1/2 = 2 . Here the energy does not exhibit equipartition, and the limit is determined solely by the H 1 / 2 H^{1/2} H 1/2 interactions between the two wells of the potential.
The following lemma addresses the specific features of regime (R2). For s = 1 2 s=\frac{1}{2} s = 2 1 , the presence of the additional well s k s_k s k modifies the limiting energy cost.
Lemma 5.2 (Regime (R2) – Wells in Z \mathbb{Z} Z ). Let n = 1 n=1 n = 1 and assume we are in regime (R2). Let I ⊂ R I\subset\mathbb{R} I ⊂ R be a bounded open interval, u ∈ B V ( I ; { k , k + 1 } ) u\in BV(I;\{k,k+1\}) u ∈ B V ( I ; { k , k + 1 }) with k ∈ Z k\in\mathbb{Z} k ∈ Z , and { u ε } ⊂ L 1 ( I ; [ k , k + 1 ] ) \{u_\varepsilon\}\subset L^1(I;[k,k+1]) { u ε } ⊂ L 1 ( I ; [ k , k + 1 ]) be a sequence such that u ε u_\varepsilon u ε converges to u u u in L 1 ( I ) L^1(I) L 1 ( I ) . Then
lim inf ε → 0 F ε ( u ε ; I ) ≥ H 0 ( J u ) Ψ s u b ( k , k + 1 ) = H 0 ( J u ) { κ s ( 1 − 2 ( s k − k ) ( k + 1 − s k ) ) , if s > 1 2 , 2 − 4 ℓ ( s k − k ) ( k + 1 − s k ) , if s = 1 2 . \liminf_{\varepsilon\to0}F_\varepsilon(u_\varepsilon;I)\geq\mathcal{H}^0(J_u)\Psi_{\mathrm{sub}}(k,k+1)=\mathcal{H}^0(J_u)
\begin{cases}
\kappa_s(1-2(s_k-k)(k+1-s_k)), & \text{if }s>\frac{1}{2},\\
2-4\ell(s_k-k)(k+1-s_k), & \text{if }s=\frac{1}{2}.
\end{cases} ε → 0 lim inf F ε ( u ε ; I ) ≥ H 0 ( J u ) Ψ sub ( k , k + 1 ) = H 0 ( J u ) { κ s ( 1 − 2 ( s k − k ) ( k + 1 − s k )) , 2 − 4 ℓ ( s k − k ) ( k + 1 − s k ) , if s > 2 1 , if s = 2 1 . Proof. Case s > 1 2 s>\frac{1}{2} s > 2 1 . For s > 1 2 s>\frac{1}{2} s > 2 1 , the thesis follows directly from the estimate F ε ( u ε ) ≥ F ε Z ∪ S ( u ε ) F_\varepsilon(u_\varepsilon)\geq F_{\varepsilon}^{\mathbb{Z}\cup S}(u_\varepsilon) F ε ( u ε ) ≥ F ε Z ∪ S ( u ε ) . Truncating the function in [ k , s k ] [k,s_k] [ k , s k ] and [ s k , k + 1 ] [s_k,k+1] [ s k , k + 1 ] we can apply Lemma 5.1 and conclude that
lim inf ε → 0 F ε Z ∪ S ( u ε ) ≥ Ψ s u p ( k , k + 1 ) H 0 ( J u ) = κ s ( ( s k − k ) 2 + ( k + 1 − s k ) 2 ) H 0 ( J u ) . \liminf_{\varepsilon\to0}F_{\varepsilon}^{\mathbb{Z}\cup S}(u_\varepsilon)\geq\Psi_{\mathrm{sup}}(k,k+1)\mathcal{H}^0(J_u)=\kappa_s((s_k-k)^2+(k+1-s_k)^2)\mathcal{H}^0(J_u). ε → 0 lim inf F ε Z ∪ S ( u ε ) ≥ Ψ sup ( k , k + 1 ) H 0 ( J u ) = κ s (( s k − k ) 2 + ( k + 1 − s k ) 2 ) H 0 ( J u ) . Case s = 1 2 s=\frac{1}{2} s = 2 1 . It is sufficient to prove the result for a jump set consisting of a single point, J u = { x } J_u=\{x\} J u = { x } . Indeed, for a general jump set, we can consider a finite collection of disjoint intervals { I j } \{I_j\} { I j } , each containing exactly one jump point of u u u , and then exploit the superadditivity of the energy:
F ε ( u ε ; I ) ≥ ∑ j F ε ( u ε ; I j ) . F_\varepsilon(u_\varepsilon;I)\geq\sum_j F_\varepsilon(u_\varepsilon;I_j). F ε ( u ε ; I ) ≥ j ∑ F ε ( u ε ; I j ) . We can reduce the problem to the case where u ε u_\varepsilon u ε is a sequence of monotone non-decreasing functions via rearrangement, since the potential term ∫ I W ε ( u ε ) d x \int_I W_\varepsilon(u_\varepsilon)\,dx ∫ I W ε ( u ε ) d x is invariant under rearrangement, while the fractional Gagliardo seminorm decreases by rearrangement (see, e.g., [17 ] ). Let w k : = s k − k w_k:=s_k-k w k := s k − k . Fix δ ∈ ( 0 , d / 4 ) \delta\in(0,d/4) δ ∈ ( 0 , d /4 ) with d d d as in (4) , let us consider the following disjoint sets
A ε : = { u ε ∈ [ k , k + δ ) } , B ε : = { u ε ∈ ( k + 1 − δ , k + 1 ] } , A_\varepsilon:=\{u_\varepsilon\in[k,k+\delta)\},\qquad B_\varepsilon:=\{u_\varepsilon\in(k+1-\delta,k+1]\}, A ε := { u ε ∈ [ k , k + δ )} , B ε := { u ε ∈ ( k + 1 − δ , k + 1 ]} ,
C ε : = { u ε ∈ ( s k − δ , s k + δ ) } , D ε : = I ∖ ( A ε ∪ B ε ∪ C ε ) . C_{\varepsilon} := \{u_{\varepsilon} \in(s_k-\delta,s_k+\delta)\}, \qquad D_{\varepsilon} := I \setminus(A_{\varepsilon} \cup B_{\varepsilon} \cup C_{\varepsilon}). C ε := { u ε ∈ ( s k − δ , s k + δ )} , D ε := I ∖ ( A ε ∪ B ε ∪ C ε ) .
We can easily estimate the potential term as follows
∫ I W ε ( u ε ( x ) ) d x ≥ ∫ D ε W ε ( u ε ( x ) ) d x + ∫ C ε W ε ( u ε ( x ) ) d x ≥ ∣ D ε ∣ inf D ε W ε ∘ u ε + ∣ C ε ∣ inf C ε W ε ∘ u ε . \int_I W_{\varepsilon}(u_{\varepsilon}(x))\,dx \ge\int_{D_{\varepsilon}} W_{\varepsilon}(u_{\varepsilon}(x))\,dx + \int_{C_{\varepsilon}} W_{\varepsilon}(u_{\varepsilon}(x))\,dx \ge|D_{\varepsilon}|\inf_{D_{\varepsilon}} W_{\varepsilon}\circ u_{\varepsilon} + |C_{\varepsilon}|\inf_{C_{\varepsilon}} W_{\varepsilon}\circ u_{\varepsilon}. ∫ I W ε ( u ε ( x )) d x ≥ ∫ D ε W ε ( u ε ( x )) d x + ∫ C ε W ε ( u ε ( x )) d x ≥ ∣ D ε ∣ D ε inf W ε ∘ u ε + ∣ C ε ∣ C ε inf W ε ∘ u ε .
Now, for small ε \varepsilon ε we have
inf D ε W ε ∘ u ε ≥ min { W ε ( x ) : x ∈ [ k + δ , s k − δ ] ∪ [ s k + δ , k + 1 − δ ] } = δ 2 \inf_{D_{\varepsilon}} W_{\varepsilon}\circ u_{\varepsilon} \ge\min\{W_{\varepsilon}(x):x\in[k+\delta,s_k-\delta]\cup[s_k+\delta,k+1-\delta]\}=\delta^2 D ε inf W ε ∘ u ε ≥ min { W ε ( x ) : x ∈ [ k + δ , s k − δ ] ∪ [ s k + δ , k + 1 − δ ]} = δ 2
and
inf C ε W ε ∘ u ε ≥ min { W ε ( x ) : x ∈ [ s k − δ , s k + δ ] } = δ ε . \inf_{C_{\varepsilon}} W_{\varepsilon}\circ u_{\varepsilon} \ge\min\{W_{\varepsilon}(x):x\in[s_k-\delta,s_k+\delta]\}=\delta_{\varepsilon}. C ε inf W ε ∘ u ε ≥ min { W ε ( x ) : x ∈ [ s k − δ , s k + δ ]} = δ ε .
Hence,
∫ I W ε ( u ε ( x ) ) d x ≥ ∣ D ε ∣ δ 2 + ∣ C ε ∣ δ ε . (38) \int_I W_{\varepsilon}(u_{\varepsilon}(x))\,dx \ge|D_{\varepsilon}|\delta^2+|C_{\varepsilon}|\delta_{\varepsilon}. \tag*{(38)} ∫ I W ε ( u ε ( x )) d x ≥ ∣ D ε ∣ δ 2 + ∣ C ε ∣ δ ε . (38)
For the Gagliardo seminorm, we have
1 2 ∫ I ∫ I ∣ u ε ( x ) − u ε ( y ) ∣ 2 ∣ x − y ∣ 2 d x d y ≥ ∫ A ε ∫ B ε ∣ u ε ( x ) − u ε ( y ) ∣ 2 ∣ x − y ∣ 2 d x d y + ∫ C ε ∫ B ε ∣ u ε ( x ) − u ε ( y ) ∣ 2 ∣ x − y ∣ 2 d x d y + ∫ A ε ∫ C ε ∣ u ε ( x ) − u ε ( y ) ∣ 2 ∣ x − y ∣ 2 d x d y ≥ ( 1 − 2 δ ) 2 ∫ A ε ∫ B ε 1 ∣ x − y ∣ 2 d x d y + ( 1 − w k − 2 δ ) 2 ∫ C ε ∫ B ε 1 ∣ x − y ∣ 2 d x d y \begin{aligned}
\frac{1}{2}\int_I\int_I\frac{|u_{\varepsilon}(x)-u_{\varepsilon}(y)|^2}{|x-y|^2}\,dx\,dy \\
&\ge\int_{A_{\varepsilon}}\int_{B_{\varepsilon}}\frac{|u_{\varepsilon}(x)-u_{\varepsilon}(y)|^2}{|x-y|^2}\,dx\,dy+\int_{C_{\varepsilon}}\int_{B_{\varepsilon}}\frac{|u_{\varepsilon}(x)-u_{\varepsilon}(y)|^2}{|x-y|^2}\,dx\,dy+\int_{A_{\varepsilon}}\int_{C_{\varepsilon}}\frac{|u_{\varepsilon}(x)-u_{\varepsilon}(y)|^2}{|x-y|^2}\,dx\,dy \\
&\ge(1-2\delta)^2\int_{A_{\varepsilon}}\int_{B_{\varepsilon}}\frac{1}{|x-y|^2}\,dx\,dy+(1-w_k-2\delta)^2\int_{C_{\varepsilon}}\int_{B_{\varepsilon}}\frac{1}{|x-y|^2}\,dx\,dy
\end{aligned} 2 1 ∫ I ∫ I ∣ x − y ∣ 2 ∣ u ε ( x ) − u ε ( y ) ∣ 2 d x d y ≥ ∫ A ε ∫ B ε ∣ x − y ∣ 2 ∣ u ε ( x ) − u ε ( y ) ∣ 2 d x d y + ∫ C ε ∫ B ε ∣ x − y ∣ 2 ∣ u ε ( x ) − u ε ( y ) ∣ 2 d x d y + ∫ A ε ∫ C ε ∣ x − y ∣ 2 ∣ u ε ( x ) − u ε ( y ) ∣ 2 d x d y ≥ ( 1 − 2 δ ) 2 ∫ A ε ∫ B ε ∣ x − y ∣ 2 1 d x d y + ( 1 − w k − 2 δ ) 2 ∫ C ε ∫ B ε ∣ x − y ∣ 2 1 d x d y
+ ( w k − 2 δ ) 2 ∫ A ε ∫ C ε 1 ∣ x − y ∣ 2 d x d y . \qquad+(w_k-2\delta)^2\int_{A_{\varepsilon}}\int_{C_{\varepsilon}}\frac{1}{|x-y|^2}\,dx\,dy. + ( w k − 2 δ ) 2 ∫ A ε ∫ C ε ∣ x − y ∣ 2 1 d x d y .
In order to estimate each integral we use the explicit one-dimensional computation: if α 1 < β 1 < α 2 < β 2 \alpha_1<\beta_1<\alpha_2<\beta_2 α 1 < β 1 < α 2 < β 2 , then
∫ α 1 β 1 ∫ α 2 β 2 1 ∣ x − y ∣ 2 d x d y = log ( ( β 2 − β 1 ) ( α 2 − α 1 ) ( β 2 − α 1 ) ( α 2 − β 1 ) ) . (39) \int_{\alpha_1}^{\beta_1}\int_{\alpha_2}^{\beta_2}\frac{1}{|x-y|^2}\,dx\,dy=\log\left(\frac{(\beta_2-\beta_1)(\alpha_2-\alpha_1)}{(\beta_2-\alpha_1)(\alpha_2-\beta_1)}\right). \tag*{(39)} ∫ α 1 β 1 ∫ α 2 β 2 ∣ x − y ∣ 2 1 d x d y = log ( ( β 2 − α 1 ) ( α 2 − β 1 ) ( β 2 − β 1 ) ( α 2 − α 1 ) ) . (39)
Therefore we get
∫ A ε ∫ B ε 1 ∣ x − y ∣ 2 d x d y = log ( ( ∣ A ε ∣ + ∣ C ε ∣ + ∣ D ε ∣ ) ( ∣ B ε ∣ + ∣ C ε ∣ + ∣ D ε ∣ ) ( ∣ C ε ∣ + ∣ D ε ∣ ) ∣ I ∣ ) . \int_{A_{\varepsilon}}\int_{B_{\varepsilon}}\frac{1}{|x-y|^2}\,dx\,dy=\log\left(\frac{(|A_{\varepsilon}|+|C_{\varepsilon}|+|D_{\varepsilon}|)(|B_{\varepsilon}|+|C_{\varepsilon}|+|D_{\varepsilon}|)}{(|C_{\varepsilon}|+|D_{\varepsilon}|)|I|}\right). ∫ A ε ∫ B ε ∣ x − y ∣ 2 1 d x d y = log ( ( ∣ C ε ∣ + ∣ D ε ∣ ) ∣ I ∣ ( ∣ A ε ∣ + ∣ C ε ∣ + ∣ D ε ∣ ) ( ∣ B ε ∣ + ∣ C ε ∣ + ∣ D ε ∣ ) ) .
To estimate the last two terms, we rely on a rearrangement argument. Since u ε u_{\varepsilon} u ε transitions from k k k to k + 1 k+1 k + 1 , the corresponding sets are spatially ordered in I I I as A ε A_{\varepsilon} A ε , followed by a portion of D ε D_{\varepsilon} D ε , then C ε C_{\varepsilon} C ε , the remaining part of D ε D_{\varepsilon} D ε , and finally B ε B_{\varepsilon} B ε . Noting that the interaction integral decreases as the distance between the sets increases, a lower bound for the double integral over A ε × C ε A_{\varepsilon}\times C_{\varepsilon} A ε × C ε (and similarly over B ε × C ε B_{\varepsilon}\times C_{\varepsilon} B ε × C ε ) is obtained by placing the entire measure of the transition set D ε D_{\varepsilon} D ε between them. A rigorous justification of this argument, based on the properties of monotone rearrangements, is provided in [3 Lemma 2] . This leads to the following estimates
∫ B ε ∫ C ε 1 ∣ x − y ∣ 2 d x d y ≥ log ( ( ∣ C ε ∣ + ∣ D ε ∣ ) ( ∣ B ε ∣ + ∣ D ε ∣ ) ( ∣ C ε ∣ + ∣ D ε ∣ + ∣ B ε ∣ ) ∣ D ε ∣ ) , \int_{B_{\varepsilon}}\int_{C_{\varepsilon}}\frac{1}{|x-y|^2}\,dx\,dy\ge\log\left(\frac{(|C_{\varepsilon}|+|D_{\varepsilon}|)(|B_{\varepsilon}|+|D_{\varepsilon}|)}{(|C_{\varepsilon}|+|D_{\varepsilon}|+|B_{\varepsilon}|)|D_{\varepsilon}|}\right), ∫ B ε ∫ C ε ∣ x − y ∣ 2 1 d x d y ≥ log ( ( ∣ C ε ∣ + ∣ D ε ∣ + ∣ B ε ∣ ) ∣ D ε ∣ ( ∣ C ε ∣ + ∣ D ε ∣ ) ( ∣ B ε ∣ + ∣ D ε ∣ ) ) ,
and
∫ A ε ∫ C ε 1 ∣ x − y ∣ 2 d x d y ≥ log ( ( ∣ C ε ∣ + ∣ D ε ∣ ) ( ∣ A ε ∣ + ∣ D ε ∣ ) ( ∣ C ε ∣ + ∣ D ε ∣ + ∣ A ε ∣ ) ∣ D ε ∣ ) . \int_{A_{\varepsilon}}\int_{C_{\varepsilon}}\frac{1}{|x-y|^2}\,dx\,dy\ge\log\left(\frac{(|C_{\varepsilon}|+|D_{\varepsilon}|)(|A_{\varepsilon}|+|D_{\varepsilon}|)}{(|C_{\varepsilon}|+|D_{\varepsilon}|+|A_{\varepsilon}|)|D_{\varepsilon}|}\right). ∫ A ε ∫ C ε ∣ x − y ∣ 2 1 d x d y ≥ log ( ( ∣ C ε ∣ + ∣ D ε ∣ + ∣ A ε ∣ ) ∣ D ε ∣ ( ∣ C ε ∣ + ∣ D ε ∣ ) ( ∣ A ε ∣ + ∣ D ε ∣ ) ) .
Now, by the convergence in measure of the sequence u ε u_{\varepsilon} u ε we have
lim inf ε → 0 ∣ A ε ∣ ≥ lim inf ε → 0 ∣ { u = k } ∩ { ∣ u ε − u ∣ < δ } ∣ ≥ ∣ { u = k } ∣ − lim ε → 0 ∣ { ∣ u ε − u ∣ ≥ δ } ∣ = ∣ { u = k } ∣ . \liminf_{\varepsilon\to0} |A_{\varepsilon}| \ge\liminf_{\varepsilon\to0} \left|\{u=k\}\cap\{|u_{\varepsilon}-u|<\delta\}\right| \ge|\{u=k\}|-\lim_{\varepsilon\to0}\left|\{|u_{\varepsilon}-u|\ge\delta\}\right|=|\{u=k\}|. ε → 0 lim inf ∣ A ε ∣ ≥ ε → 0 lim inf ∣ { u = k } ∩ { ∣ u ε − u ∣ < δ } ∣ ≥ ∣ { u = k } ∣ − ε → 0 lim ∣ { ∣ u ε − u ∣ ≥ δ } ∣ = ∣ { u = k } ∣.
Similarly
lim inf ε → 0 ∣ B ε ∣ ≥ lim inf ε → 0 ∣ { u = k + 1 } ∩ { ∣ u ε − u ∣ < δ } ∣ ≥ ∣ { u = k + 1 } ∣ − lim ε → 0 ∣ { ∣ u ε − u ∣ ≥ δ } ∣ = ∣ { u = k + 1 } ∣ , \begin{aligned}
\liminf_{\varepsilon\to0}|B_{\varepsilon}|&\ge\liminf_{\varepsilon\to0}\left|\{u=k+1\}\cap\{|u_{\varepsilon}-u|<\delta\}\right|\\
&\ge|\{u=k+1\}|-\lim_{\varepsilon\to0}\left|\{|u_{\varepsilon}-u|\ge\delta\}\right|=|\{u=k+1\}|,
\end{aligned} ε → 0 lim inf ∣ B ε ∣ ≥ ε → 0 lim inf ∣ { u = k + 1 } ∩ { ∣ u ε − u ∣ < δ } ∣ ≥ ∣ { u = k + 1 } ∣ − ε → 0 lim ∣ { ∣ u ε − u ∣ ≥ δ } ∣ = ∣ { u = k + 1 } ∣ ,
lim ε → 0 ∣ C ε ∣ ≤ lim ε → 0 ∣ { ∣ u ε − u ∣ ≥ δ } ∣ = 0 and lim ε → 0 ∣ D ε ∣ ≤ lim ε → 0 ∣ { ∣ u ε − u ∣ ≥ δ } ∣ = 0. \lim_{\varepsilon\to0}|C_{\varepsilon}|\le\lim_{\varepsilon\to0}\left|\{|u_{\varepsilon}-u|\ge\delta\}\right|=0
\quad\text{and}\quad
\lim_{\varepsilon\to0}|D_{\varepsilon}|\le\lim_{\varepsilon\to0}\left|\{|u_{\varepsilon}-u|\ge\delta\}\right|=0. ε → 0 lim ∣ C ε ∣ ≤ ε → 0 lim ∣ { ∣ u ε − u ∣ ≥ δ } ∣ = 0 and ε → 0 lim ∣ D ε ∣ ≤ ε → 0 lim ∣ { ∣ u ε − u ∣ ≥ δ } ∣ = 0.
Using the estimate from below of the integrals and these limits, we can estimate the seminorm as follows
∬ I × I ∣ u ε ( x ) − u ε ( y ) ∣ 2 ∣ x − y ∣ 2 d x d y ≥ 2 [ ( 1 − 2 δ ) 2 log ( 1 ∣ C ε ∣ + ∣ D ε ∣ ) + ( ( 1 − w k − 2 δ ) 2 + ( w k − 2 δ ) 2 ) log ( ∣ C ε ∣ + ∣ D ε ∣ ∣ D ε ∣ ) ] + O ( 1 ) = − ( 4 w k ( 1 − w k ) − 8 δ 2 ) log ( ∣ C ε ∣ + ∣ D ε ∣ ) − 2 [ ( 1 − w k − 2 δ ) 2 + ( w k − 2 δ ) 2 ] log ∣ D ε ∣ + O ( 1 ) = − ( 4 w k ( 1 − w k ) − 8 δ 2 ) log ( ∣ C ε ∣ + ∣ D ε ∣ ) − ( 2 − 4 w k ( 1 − w k ) − 8 δ + 16 δ 2 ) log ∣ D ε ∣ + O ( 1 ) . (40) \begin{aligned}
\iint_{I\times I}\frac{|u_{\varepsilon}(x)-u_{\varepsilon}(y)|^{2}}{|x-y|^{2}}\,dx\,dy
&\ge2\left[(1-2\delta)^{2}\log\left(\frac{1}{|C_{\varepsilon}|+|D_{\varepsilon}|}\right)\right.\\
&\qquad\left.+\left((1-w_{k}-2\delta)^{2}+(w_{k}-2\delta)^{2}\right)\log\left(\frac{|C_{\varepsilon}|+|D_{\varepsilon}|}{|D_{\varepsilon}|}\right)\right]+O(1)\\
&=-(4w_{k}(1-w_{k})-8\delta^{2})\log(|C_{\varepsilon}|+|D_{\varepsilon}|)\\
&\qquad-2\left[(1-w_{k}-2\delta)^{2}+(w_{k}-2\delta)^{2}\right]\log|D_{\varepsilon}|+O(1)\\
&=-(4w_{k}(1-w_{k})-8\delta^{2})\log(|C_{\varepsilon}|+|D_{\varepsilon}|)\\
&\qquad-(2-4w_{k}(1-w_{k})-8\delta+16\delta^{2})\log|D_{\varepsilon}|+O(1). \tag*{(40)}
\end{aligned} ∬ I × I ∣ x − y ∣ 2 ∣ u ε ( x ) − u ε ( y ) ∣ 2 d x d y ≥ 2 [ ( 1 − 2 δ ) 2 log ( ∣ C ε ∣ + ∣ D ε ∣ 1 ) + ( ( 1 − w k − 2 δ ) 2 + ( w k − 2 δ ) 2 ) log ( ∣ D ε ∣ ∣ C ε ∣ + ∣ D ε ∣ ) ] + O ( 1 ) = − ( 4 w k ( 1 − w k ) − 8 δ 2 ) log ( ∣ C ε ∣ + ∣ D ε ∣ ) − 2 [ ( 1 − w k − 2 δ ) 2 + ( w k − 2 δ ) 2 ] log ∣ D ε ∣ + O ( 1 ) = − ( 4 w k ( 1 − w k ) − 8 δ 2 ) log ( ∣ C ε ∣ + ∣ D ε ∣ ) − ( 2 − 4 w k ( 1 − w k ) − 8 δ + 16 δ 2 ) log ∣ D ε ∣ + O ( 1 ) . (40)
Using estimates (40) and (38) , we obtain
F ε ( u ε ) ≥ 1 ∣ log ε ∣ [ − ( 4 w k ( 1 − w k ) − 8 δ 2 ) log ( ∣ C ε ∣ + ∣ D ε ∣ ) − ( 2 − 4 w k ( 1 − w k ) − 8 δ + 16 δ 2 ) log ∣ D ε ∣ + ∣ D ε ∣ δ 2 ε + ∣ C ε ∣ δ ε ε + o ( 1 ) ] = 4 w k ( 1 − w k ) − 8 δ 2 ∣ log ε ∣ ( − log ( ∣ C ε ∣ + ∣ D ε ∣ ) + ( ∣ C ε ∣ + ∣ D ε ∣ ) δ ε ε ( 4 w k ( 1 − w k ) − 8 δ 2 ) ) + 2 − 4 w k ( 1 − w k ) − 8 δ + 16 δ 2 ∣ log ε ∣ ( − log ∣ D ε ∣ + ∣ D ε ∣ δ 2 − δ ε ε ( 2 − 4 w k ( 1 − w k ) − 8 δ + 16 δ 2 ) ) + o ( 1 ) . \begin{aligned}
F_{\varepsilon}(u_{\varepsilon})\ge\frac{1}{|\log\varepsilon|}\bigg[&
-(4w_{k}(1-w_{k})-8\delta^{2})\log(|C_{\varepsilon}|+|D_{\varepsilon}|)\\
&-(2-4w_{k}(1-w_{k})-8\delta+16\delta^{2})\log|D_{\varepsilon}|\\
&+\frac{|D_{\varepsilon}|\delta^{2}}{\varepsilon}+\frac{|C_{\varepsilon}|\delta_{\varepsilon}}{\varepsilon}+o(1)\bigg]\\
={}&\frac{4w_{k}(1-w_{k})-8\delta^{2}}{|\log\varepsilon|}
\left(-\log(|C_{\varepsilon}|+|D_{\varepsilon}|)+(|C_{\varepsilon}|+|D_{\varepsilon}|)\frac{\delta_{\varepsilon}}{\varepsilon(4w_{k}(1-w_{k})-8\delta^{2})}\right)\\
&+\frac{2-4w_{k}(1-w_{k})-8\delta+16\delta^{2}}{|\log\varepsilon|}
\left(-\log|D_{\varepsilon}|+|D_{\varepsilon}|\frac{\delta^{2}-\delta_{\varepsilon}}{\varepsilon(2-4w_{k}(1-w_{k})-8\delta+16\delta^{2})}\right)+o(1).
\end{aligned} F ε ( u ε ) ≥ ∣ log ε ∣ 1 [ = − ( 4 w k ( 1 − w k ) − 8 δ 2 ) log ( ∣ C ε ∣ + ∣ D ε ∣ ) − ( 2 − 4 w k ( 1 − w k ) − 8 δ + 16 δ 2 ) log ∣ D ε ∣ + ε ∣ D ε ∣ δ 2 + ε ∣ C ε ∣ δ ε + o ( 1 ) ] ∣ log ε ∣ 4 w k ( 1 − w k ) − 8 δ 2 ( − log ( ∣ C ε ∣ + ∣ D ε ∣ ) + ( ∣ C ε ∣ + ∣ D ε ∣ ) ε ( 4 w k ( 1 − w k ) − 8 δ 2 ) δ ε ) + ∣ log ε ∣ 2 − 4 w k ( 1 − w k ) − 8 δ + 16 δ 2 ( − log ∣ D ε ∣ + ∣ D ε ∣ ε ( 2 − 4 w k ( 1 − w k ) − 8 δ + 16 δ 2 ) δ 2 − δ ε ) + o ( 1 ) .
Now, using that if x , M > 0 x,M>0 x , M > 0 then log x ≤ M x − log M \log x\le Mx-\log M log x ≤ M x − log M , we have (using that δ ε ≤ δ 2 \delta_{\varepsilon}\le\delta^{2} δ ε ≤ δ 2 eventually)
F ε ( u ε ) ≥ 4 w k ( 1 − w k ) − 8 δ 2 ∣ log ε ∣ log ( δ ε ε ( 4 w k ( 1 − w k ) − 8 δ 2 ) ) + 2 − 4 w k ( 1 − w k ) − 8 δ + 16 δ 2 ∣ log ε ∣ log ( δ 2 − δ ε ε ( 2 − 4 w k ( 1 − w k ) − 8 δ + 16 δ 2 ) ) + o ( 1 ) = ( 4 w k ( 1 − w k ) − 8 δ 2 ) ( 1 − ∣ log δ ε ∣ ∣ log ε ∣ + O ( 1 ) ∣ log ε ∣ ) + ( 2 − 4 w k ( 1 − w k ) − 8 δ + 16 δ 2 ) ( 1 + O ( 1 ) ∣ log ε ∣ ) + o ( 1 ) \begin{aligned}
F_{\varepsilon}(u_{\varepsilon}) \ge\frac{4w_k(1-w_k)-8\delta^2}{|\log\varepsilon|}\log\left(\frac{\delta_{\varepsilon}}{\varepsilon(4w_k(1-w_k)-8\delta^2)}\right)+\frac{2-4w_k(1-w_k)-8\delta+16\delta^2}{|\log\varepsilon|}\log\left(\frac{\delta^2-\delta_{\varepsilon}}{\varepsilon(2-4w_k(1-w_k)-8\delta+16\delta^2)}\right)+o(1) \\
&=(4w_k(1-w_k)-8\delta^2)\left(1-\frac{|\log\delta_{\varepsilon}|}{|\log\varepsilon|}+\frac{O(1)}{|\log\varepsilon|}\right)+(2-4w_k(1-w_k)-8\delta+16\delta^2)\left(1+\frac{O(1)}{|\log\varepsilon|}\right)+o(1)
\end{aligned} F ε ( u ε ) ≥ ∣ log ε ∣ 4 w k ( 1 − w k ) − 8 δ 2 log ( ε ( 4 w k ( 1 − w k ) − 8 δ 2 ) δ ε ) + ∣ log ε ∣ 2 − 4 w k ( 1 − w k ) − 8 δ + 16 δ 2 log ( ε ( 2 − 4 w k ( 1 − w k ) − 8 δ + 16 δ 2 ) δ 2 − δ ε ) + o ( 1 ) = ( 4 w k ( 1 − w k ) − 8 δ 2 ) ( 1 − ∣ log ε ∣ ∣ log δ ε ∣ + ∣ log ε ∣ O ( 1 ) ) + ( 2 − 4 w k ( 1 − w k ) − 8 δ + 16 δ 2 ) ( 1 + ∣ log ε ∣ O ( 1 ) ) + o ( 1 )
⟶ ε → 0 ( 4 w k ( 1 − w k ) − 8 δ 2 ) ( 1 − ℓ ) + ( 2 − 4 w k ( 1 − w k ) − 8 δ + 16 δ 2 ) . \underset{\varepsilon\to0}{\longrightarrow}(4w_k(1-w_k)-8\delta^2)(1-\ell)+(2-4w_k(1-w_k)-8\delta+16\delta^2). ε → 0 ⟶ ( 4 w k ( 1 − w k ) − 8 δ 2 ) ( 1 − ℓ ) + ( 2 − 4 w k ( 1 − w k ) − 8 δ + 16 δ 2 ) .
Taking the limit as δ → 0 \delta\to0 δ → 0 we have that lim inf ε → 0 F ε ( u ε ) ≥ 2 − 4 ℓ w k ( 1 − w k ) \liminf_{\varepsilon\to0}F_{\varepsilon}(u_{\varepsilon})\ge2-4\ell w_k(1-w_k) lim inf ε → 0 F ε ( u ε ) ≥ 2 − 4 ℓ w k ( 1 − w k ) and therefore the thesis follows. □
Lower bounds for the n n n -dimensional case
In this section, we extend the one-dimensional lower bound to n ≥ 2 n\ge2 n ≥ 2 by slicing. Differently from the local energy where we slice with respect to the normal of the jump set J u J_u J u , the non-local energy is a superposition of one-dimensional H s H^s H s energies along all directions ξ ∈ S n − 1 \xi\in\mathbb{S}^{n-1} ξ ∈ S n − 1 (see (41) ).
We prove the following result.
Proposition 5.3. Let n ≥ 2 n\ge2 n ≥ 2 and s ∈ [ 1 2 , 1 ) s\in[\frac{1}{2},1) s ∈ [ 2 1 , 1 ) .
Regime (R1). Assume (R1) holds and let v i v_i v i and v i + 1 v_{i+1} v i + 1 be two consecutive points in Z ∪ S \mathbb{Z}\cup S Z ∪ S . Let u ∈ B V ( Q ; { v i , v i + 1 } ) u\in BV(Q;\{v_i,v_{i+1}\}) u ∈ B V ( Q ; { v i , v i + 1 }) . Let { u ε } ⊂ L 1 ( Q ; [ v i , v i + 1 ] ) \{u_{\varepsilon}\}\subset L^1(Q;[v_i,v_{i+1}]) { u ε } ⊂ L 1 ( Q ; [ v i , v i + 1 ]) be a sequence such that u ε u_{\varepsilon} u ε converges to u u u in L 1 ( Q ) L^1(Q) L 1 ( Q ) . Then
lim inf ε → 0 F ε Z ∪ S ( u ε ) ≥ Ψ s u p ( v i , v i + 1 ) H n − 1 ( J u ) . \liminf_{\varepsilon\to0}F_{\varepsilon}^{\mathbb{Z}\cup S}(u_{\varepsilon})\ge\Psi_{\mathrm{sup}}(v_i,v_{i+1})\mathcal{H}^{n-1}(J_u). ε → 0 lim inf F ε Z ∪ S ( u ε ) ≥ Ψ sup ( v i , v i + 1 ) H n − 1 ( J u ) . Regime (R2). Assume (R2) holds. Let u ∈ B V ( Q ; { k , k + 1 } ) u\in BV(Q;\{k,k+1\}) u ∈ B V ( Q ; { k , k + 1 }) with k ∈ Z k\in\mathbb{Z} k ∈ Z . Let { u ε } ⊂ L 1 ( Q ; [ k , k + 1 ] ) \{u_{\varepsilon}\}\subset L^1(Q;[k,k+1]) { u ε } ⊂ L 1 ( Q ; [ k , k + 1 ]) be a sequence such that u ε u_{\varepsilon} u ε converges to u u u in L 1 ( Q ) L^1(Q) L 1 ( Q ) . Then
lim inf ε → 0 F ε ( u ε ) ≥ Ψ s u b ( k , k + 1 ) H n − 1 ( J u ) . \liminf_{\varepsilon\to0}F_{\varepsilon}(u_{\varepsilon})\ge\Psi_{\mathrm{sub}}(k,k+1)\mathcal{H}^{n-1}(J_u). ε → 0 lim inf F ε ( u ε ) ≥ Ψ sub ( k , k + 1 ) H n − 1 ( J u ) .
Notation. Fix n ≥ 2 n\ge2 n ≥ 2 and s ∈ [ 1 2 , 1 ) s\in[\frac{1}{2},1) s ∈ [ 2 1 , 1 ) ; to treat both regimes at once, we use the following notation.
In regime (R1): a < b a<b a < b are two consecutive points of Z ∪ S \mathbb{Z}\cup S Z ∪ S , V ε : = W S = dist 2 ( ⋅ , Z ∪ S ) V_{\varepsilon}:=W^S=\operatorname{dist}^2(\cdot,\mathbb{Z}\cup S) V ε := W S = dist 2 ( ⋅ , Z ∪ S ) (independent of ε \varepsilon ε ), F ‾ ε : = F ε Z ∪ S \overline{F}_{\varepsilon}:=F_{\varepsilon}^{\mathbb{Z}\cup S} F ε := F ε Z ∪ S as in (12) and Ψ ( 1 ) : = λ s , n − 1 Ψ s u p \Psi^{(1)}:=\lambda_{s,n}^{-1}\Psi_{\mathrm{sup}} Ψ ( 1 ) := λ s , n − 1 Ψ sup the one-dimensional cost.
In regime (R2): a < b a<b a < b are two consecutive points of Z \mathbb{Z} Z , V ε : = W ε V_{\varepsilon}:=W_{\varepsilon} V ε := W ε , F ‾ ε : = F ε \overline{F}_{\varepsilon}:=F_{\varepsilon} F ε := F ε as in (3) , and Ψ ( 1 ) : = λ s , n − 1 Ψ s u b \Psi^{(1)}:=\lambda_{s,n}^{-1}\Psi_{\mathrm{sub}} Ψ ( 1 ) := λ s , n − 1 Ψ sub the one-dimensional cost.
With this notation, we have
F ‾ ε ( v ; Q ) : = 1 θ ε ( ε 2 s [ v ] H s ( Q ) 2 + ∫ Q V ε ( v ( t ) ) d t ) , for all v ∈ L 1 ( Q ) . \overline{F}_{\varepsilon}(v;Q):=\frac{1}{\theta_{\varepsilon}}\left(\varepsilon^{2s}[v]_{H^s(Q)}^2+\int_QV_{\varepsilon}(v(t))\,dt\right),\qquad\text{for all }v\in L^1(Q). F ε ( v ; Q ) := θ ε 1 ( ε 2 s [ v ] H s ( Q ) 2 + ∫ Q V ε ( v ( t )) d t ) , for all v ∈ L 1 ( Q ) .
On each slice, the constants in front of the seminorm and of the potential are no longer 1. We first need to establish how the multiplicative constants in front of the two terms affect the cost of a unit jump. Let I ⊂ R I \subset\mathbb{R} I ⊂ R be a bounded open interval, V : R → [ 0 , + ∞ ) V:\mathbb{R} \to[0,+\infty) V : R → [ 0 , + ∞ ) a continuous potential and c > 0 c>0 c > 0 . We define
G ε ( v ; I , V , c ) : = 1 θ ε ( ε 2 s c [ v ] H s ( I ) 2 + ∫ I V ( v ( t ) ) d t ) , for all v ∈ L 1 ( I ) . \mathcal{G}_{\varepsilon}(v;I,V,c):=\frac{1}{\theta_{\varepsilon}}\left(\varepsilon^{2s}c[v]_{H^s(I)}^2+\int_I V(v(t))\,dt\right),\qquad\text{for all }v\in L^1(I). G ε ( v ; I , V , c ) := θ ε 1 ( ε 2 s c [ v ] H s ( I ) 2 + ∫ I V ( v ( t )) d t ) , for all v ∈ L 1 ( I ) .
The following lemma extends Lemma 5.1 and Lemma 5.2 to arbitrary constants in front of the two terms.
Lemma 5.4. Let n ≥ 2 n\ge2 n ≥ 2 , s ∈ [ 1 2 , 1 ) s\in\left[\frac{1}{2},1\right) s ∈ [ 2 1 , 1 ) , and assume either (R1) or (R2). Let a , b , V ε , Ψ ( 1 ) a,b,V_\varepsilon,\Psi^{(1)} a , b , V ε , Ψ ( 1 ) be as above. Let I ⊂ R I\subset\mathbb{R} I ⊂ R be a bounded open interval and u ∈ B V ( I ; { a , b } ) u\in BV(I;\{a,b\}) u ∈ B V ( I ; { a , b }) . Let { u ε } ⊂ L 1 ( I ; [ a , b ] ) \{u_\varepsilon\}\subset L^1(I;[a,b]) { u ε } ⊂ L 1 ( I ; [ a , b ]) be a sequence that converges to u u u in L 1 ( I ) L^1(I) L 1 ( I ) . Then for all c , p > 0 c,p>0 c , p > 0 we have
lim inf ε → 0 G ε ( u ε ; I , p V ε , c ) ≥ p 1 − 1 2 s c 1 2 s Ψ ( 1 ) ( a , b ) H 0 ( J u ) . \liminf_{\varepsilon\to0}\mathcal{G}_{\varepsilon}(u_\varepsilon;I,pV_\varepsilon,c)\ge p^{1-\frac{1}{2s}}c^{\frac{1}{2s}}\Psi^{(1)}(a,b)\mathcal{H}^0(J_u). ε → 0 lim inf G ε ( u ε ; I , p V ε , c ) ≥ p 1 − 2 s 1 c 2 s 1 Ψ ( 1 ) ( a , b ) H 0 ( J u ) . For c = p = 1 c=p=1 c = p = 1 this reduces to Lemmas 5.1 and 5.2.
Proof. Since G ε ( v ; I , p V ε , c ) = c G ε ( v ; I , q V ε , 1 ) \mathcal{G}_{\varepsilon}(v;I,pV_\varepsilon,c)=c\mathcal{G}_{\varepsilon}(v;I,qV_\varepsilon,1) G ε ( v ; I , p V ε , c ) = c G ε ( v ; I , q V ε , 1 ) with q : = p / c q:=p/c q := p / c , and c q 1 − 1 2 s = p 1 − 1 2 s c 1 2 s cq^{1-\frac{1}{2s}}=p^{1-\frac{1}{2s}}c^{\frac{1}{2s}} c q 1 − 2 s 1 = p 1 − 2 s 1 c 2 s 1 , we may take c = 1 c=1 c = 1 .
Setting ε ˉ : = q − 1 2 s ε \bar{\varepsilon}:=q^{-\frac{1}{2s}}\varepsilon ε ˉ := q − 2 s 1 ε , we have ε 2 s = q ε ˉ 2 s \varepsilon^{2s}=q\bar{\varepsilon}^{2s} ε 2 s = q ε ˉ 2 s , hence
G ε ( v ; I , q V ε , 1 ) = q θ ε ˉ θ ε 1 θ ε ˉ ( ε ˉ 2 s [ v ] H s ( I ) 2 + ∫ I V ε ( v ) ) , \mathcal{G}_{\varepsilon}(v;I,qV_\varepsilon,1)=\frac{q\theta_{\bar{\varepsilon}}}{\theta_\varepsilon}\frac{1}{\theta_{\bar{\varepsilon}}}\left(\bar{\varepsilon}^{2s}[v]_{H^s(I)}^2+\int_I V_\varepsilon(v)\right), G ε ( v ; I , q V ε , 1 ) = θ ε q θ ε ˉ θ ε ˉ 1 ( ε ˉ 2 s [ v ] H s ( I ) 2 + ∫ I V ε ( v ) ) , where
V ε = { W S , in regime ( R 1 ) , W ε , in regime ( R 2 ) , V_\varepsilon=
\begin{cases}
W^S, & \text{in regime }(R1),\\
W_\varepsilon, & \text{in regime }(R2),
\end{cases} V ε = { W S , W ε , in regime ( R 1 ) , in regime ( R 2 ) , with
q θ ε ˉ θ ε = q 1 − 1 2 s for s > 1 2 and q θ ε ˉ θ ε → q 1 − 1 2 s for s = 1 2 . \frac{q\theta_{\bar{\varepsilon}}}{\theta_\varepsilon}=q^{1-\frac{1}{2s}}\quad\text{for }s>\frac{1}{2}
\qquad\text{and}\qquad
\frac{q\theta_{\bar{\varepsilon}}}{\theta_\varepsilon}\to q^{1-\frac{1}{2s}}\quad\text{for }s=\frac{1}{2}. θ ε q θ ε ˉ = q 1 − 2 s 1 for s > 2 1 and θ ε q θ ε ˉ → q 1 − 2 s 1 for s = 2 1 . Since, in regime (R2),
lim ε → 0 δ ε θ ε = lim ε ˉ → 0 δ ε ˉ θ ε ˉ = + ∞ , \lim_{\varepsilon\to0}\frac{\delta_\varepsilon}{\theta_\varepsilon}
=\lim_{\bar{\varepsilon}\to0}\frac{\delta_{\bar{\varepsilon}}}{\theta_{\bar{\varepsilon}}}=+\infty, ε → 0 lim θ ε δ ε = ε ˉ → 0 lim θ ε ˉ δ ε ˉ = + ∞ , we can apply Lemmas 5.1 and 5.2 along ε ˉ → 0 \bar{\varepsilon}\to0 ε ˉ → 0 and obtain
lim inf ε → 0 G ε ( u ε ; I , p V ε , c ) = c lim inf ε ˉ → 0 G ε ˉ ( u ε ; I , q V ε ˉ , 1 ) ≥ c q 1 − 1 2 s Ψ ( 1 ) ( a , b ) H 0 ( J u ) . \liminf_{\varepsilon\to0}\mathcal{G}_{\varepsilon}(u_\varepsilon;I,pV_\varepsilon,c)
=c\liminf_{\bar{\varepsilon}\to0}\mathcal{G}_{\bar{\varepsilon}}(u_\varepsilon;I,qV_{\bar{\varepsilon}},1)
\ge cq^{1-\frac{1}{2s}}\Psi^{(1)}(a,b)\mathcal{H}^0(J_u). ε → 0 lim inf G ε ( u ε ; I , p V ε , c ) = c ε ˉ → 0 lim inf G ε ˉ ( u ε ; I , q V ε ˉ , 1 ) ≥ c q 1 − 2 s 1 Ψ ( 1 ) ( a , b ) H 0 ( J u ) .
In the following, we slice F ˉ ε \bar{F}_\varepsilon F ˉ ε in all directions.
Lemma 5.5. Let n ≥ 2 n\ge2 n ≥ 2 , s ∈ [ 1 2 , 1 ) s\in\left[\frac{1}{2},1\right) s ∈ [ 2 1 , 1 ) and Q ⊂ R n Q\subset\mathbb{R}^n Q ⊂ R n an open cube, and assume either (R1) or (R2). Let a , b , V ε , Ψ ( 1 ) a,b,V_\varepsilon,\Psi^{(1)} a , b , V ε , Ψ ( 1 ) be as above. Let p p p be a probability density on S n − 1 \mathbb{S}^{n-1} S n − 1 with p > 0 p>0 p > 0 H n − 1 \mathcal{H}^{n-1} H n − 1 -a.e. If { u ε } ⊂ L 1 ( Q ; [ a , b ] ) \{u_\varepsilon\}\subset L^1(Q;[a,b]) { u ε } ⊂ L 1 ( Q ; [ a , b ]) converges in L 1 ( Q ) L^1(Q) L 1 ( Q ) to u ∈ B V ( Q ; { a , b } ) u\in BV(Q;\{a,b\}) u ∈ B V ( Q ; { a , b }) , then
lim inf ε → 0 F ˉ ε ( u ε ; Q ) ≥ 2 − 1 2 s Ψ ( 1 ) ( a , b ) ∫ J u ∩ Q d H n − 1 ∫ S n − 1 d H n − 1 ( ξ ) p ( ξ ) 1 − 1 2 s ∣ ν u ⋅ ξ ∣ , \liminf_{\varepsilon\to0}\bar{F}_\varepsilon(u_\varepsilon;Q)\ge2^{-\frac{1}{2s}}\Psi^{(1)}(a,b)\int_{J_u\cap Q}d\mathcal{H}^{n-1}\int_{\mathbb{S}^{n-1}}d\mathcal{H}^{n-1}(\xi)p(\xi)^{1-\frac{1}{2s}}|\nu_u\cdot\xi|, ε → 0 lim inf F ˉ ε ( u ε ; Q ) ≥ 2 − 2 s 1 Ψ ( 1 ) ( a , b ) ∫ J u ∩ Q d H n − 1 ∫ S n − 1 d H n − 1 ( ξ ) p ( ξ ) 1 − 2 s 1 ∣ ν u ⋅ ξ ∣ , where ν u \nu_u ν u is the unit normal to the jump set J u J_u J u of u u u .
Proof. By the slicing formula for the H s H^s H s -seminorm (see [Leo23, Theorem 6.47]), we have
[ u ] H s ( Q ) 2 = 1 2 ∫ S n − 1 d H n − 1 ( ξ ) ∫ ξ ⊥ d H n − 1 ( y ) ∫ Q ξ , y ∫ Q ξ , y ∣ u ( y + t ξ ) − u ( y + τ ξ ) ∣ 2 ∣ t − τ ∣ 1 + 2 s d t d τ (41) [u]_{H^s(Q)}^2=\frac{1}{2}\int_{\mathbb{S}^{n-1}}d\mathcal{H}^{n-1}(\xi)\int_{\xi^\perp}d\mathcal{H}^{n-1}(y)\int_{Q_{\xi,y}}\int_{Q_{\xi,y}}\frac{|u(y+t\xi)-u(y+\tau\xi)|^2}{|t-\tau|^{1+2s}}\,dt\,d\tau \tag*{(41)} [ u ] H s ( Q ) 2 = 2 1 ∫ S n − 1 d H n − 1 ( ξ ) ∫ ξ ⊥ d H n − 1 ( y ) ∫ Q ξ , y ∫ Q ξ , y ∣ t − τ ∣ 1 + 2 s ∣ u ( y + t ξ ) − u ( y + τ ξ ) ∣ 2 d t d τ (41)
where Q ξ , y : = { t ∈ R : y + t ξ ∈ Q } Q_{\xi,y} := \{t \in\mathbb{R} : y + t\xi\in Q\} Q ξ , y := { t ∈ R : y + t ξ ∈ Q } . For the potential term, by Tonelli’s theorem for every ξ ∈ S n − 1 \xi\in\mathbb{S}^{n-1} ξ ∈ S n − 1 and every measurable V ≥ 0 V \ge0 V ≥ 0 we have
∫ Q V ( u ( x ) ) d x = ∫ S n − 1 d H n − 1 ( y ) ∫ Q ξ , y V ( u ( y + t ξ ) ) d t . \int_Q V(u(x))\,dx = \int_{\mathbb{S}^{n-1}} d\mathcal{H}^{n-1}(y) \int_{Q_{\xi,y}} V(u(y+t\xi))\,dt. ∫ Q V ( u ( x )) d x = ∫ S n − 1 d H n − 1 ( y ) ∫ Q ξ , y V ( u ( y + t ξ )) d t .
Now, since p p p is a probability density on S n − 1 \mathbb{S}^{n-1} S n − 1 , we have
∫ Q V ( u ( x ) ) d x = ∫ S n − 1 d H n − 1 ( ξ ) ∫ ξ ⊥ d H n − 1 ( y ) ( p ( ξ ) ∫ Q ξ , y V ( u ( y + t ξ ) ) d t ) . (42) \int_Q V(u(x))\,dx = \int_{\mathbb{S}^{n-1}} d\mathcal{H}^{n-1}(\xi) \int_{\xi^\perp} d\mathcal{H}^{n-1}(y) \left(p(\xi)\int_{Q_{\xi,y}} V(u(y+t\xi))\,dt\right). \tag*{(42)} ∫ Q V ( u ( x )) d x = ∫ S n − 1 d H n − 1 ( ξ ) ∫ ξ ⊥ d H n − 1 ( y ) ( p ( ξ ) ∫ Q ξ , y V ( u ( y + t ξ )) d t ) . (42)
Using (41) and (42) , we can rewrite the functional as follows
F ‾ ε ( u ε ) = ∫ S n − 1 d H n − 1 ( ξ ) ∫ ξ ⊥ d H n − 1 ( y ) G ε ( u ε ( y + ⋅ ξ ) ; Q ξ , y , p ( ξ ) V ε , 1 / 2 ) , \overline{F}_{\varepsilon}(u_{\varepsilon}) = \int_{\mathbb{S}^{n-1}} d\mathcal{H}^{n-1}(\xi) \int_{\xi^\perp} d\mathcal{H}^{n-1}(y) \mathcal{G}_{\varepsilon}(u_{\varepsilon}(y+\cdot\xi);Q_{\xi,y},p(\xi)V_{\varepsilon},1/2), F ε ( u ε ) = ∫ S n − 1 d H n − 1 ( ξ ) ∫ ξ ⊥ d H n − 1 ( y ) G ε ( u ε ( y + ⋅ ξ ) ; Q ξ , y , p ( ξ ) V ε , 1/2 ) ,
where the integrand vanishes if Q ξ , y = ∅ Q_{\xi,y}=\varnothing Q ξ , y = ∅ .
Now, observe that for a.e. ( ξ , y ) (\xi,y) ( ξ , y ) , we have
( u ε ) ξ , y : = u ε ( y + ⋅ ξ ) → u ( y + ⋅ ξ ) = : u ξ , y ∈ B V ( Q ξ , y ; { a , b } ) in L 1 ( Q ξ , y ) , (u_{\varepsilon})_{\xi,y}:=u_{\varepsilon}(y+\cdot\xi)\to u(y+\cdot\xi)=:u_{\xi,y}\in BV(Q_{\xi,y};\{a,b\}) \quad\text{in }L^1(Q_{\xi,y}), ( u ε ) ξ , y := u ε ( y + ⋅ ξ ) → u ( y + ⋅ ξ ) =: u ξ , y ∈ B V ( Q ξ , y ; { a , b }) in L 1 ( Q ξ , y ) ,
as a consequence of Tonelli’s theorem. Hence Lemma 5.4 gives, for a.e. ( ξ , y ) (\xi,y) ( ξ , y ) ,
lim inf ε → 0 G ε ( ( u ε ) ξ , y ; Q ξ , y , p ( ξ ) V ε , 1 / 2 ) ≥ 2 − 1 2 s p ( ξ ) 1 − 1 2 s Ψ ( 1 ) ( a , b ) H 0 ( J u ξ , y ) . \liminf_{\varepsilon\to0}\mathcal{G}_{\varepsilon}((u_{\varepsilon})_{\xi,y};Q_{\xi,y},p(\xi)V_{\varepsilon},1/2)\ge2^{-\frac{1}{2s}}p(\xi)^{1-\frac{1}{2s}}\Psi^{(1)}(a,b)\mathcal{H}^{0}(J_{u_{\xi,y}}). ε → 0 lim inf G ε (( u ε ) ξ , y ; Q ξ , y , p ( ξ ) V ε , 1/2 ) ≥ 2 − 2 s 1 p ( ξ ) 1 − 2 s 1 Ψ ( 1 ) ( a , b ) H 0 ( J u ξ , y ) .
By [5 Theorem 3.103] , we have
∫ ξ ⊥ H 0 ( J u ( y + ⋅ ξ ) ) d H n − 1 ( y ) = ( b − a ) − 1 sup { ∫ Q u ∂ ϕ ∂ ξ : ϕ ∈ C c ∞ ( Q ) , ∥ ϕ ∥ ∞ ≤ 1 } = ∫ J u ∣ ν u ⋅ ξ ∣ d H n − 1 \int_{\xi^\perp}\mathcal{H}^{0}(J_{u(y+\cdot\xi)})\,d\mathcal{H}^{n-1}(y)=(b-a)^{-1}\sup\left\{\int_Q u\frac{\partial\phi}{\partial\xi}:\phi\in C^\infty_c(Q),\ \|\phi\|_\infty\le1\right\}=\int_{J_u}|\nu_u\cdot\xi|\,d\mathcal{H}^{n-1} ∫ ξ ⊥ H 0 ( J u ( y +⋅ ξ ) ) d H n − 1 ( y ) = ( b − a ) − 1 sup { ∫ Q u ∂ ξ ∂ ϕ : ϕ ∈ C c ∞ ( Q ) , ∥ ϕ ∥ ∞ ≤ 1 } = ∫ J u ∣ ν u ⋅ ξ ∣ d H n − 1
where ν u ( x ) \nu_u(x) ν u ( x ) is the normal to the jump set J u J_u J u at H n − 1 \mathcal{H}^{n-1} H n − 1 -a.e. point x ∈ J u x\in J_u x ∈ J u . Then, by Fatou’s lemma, the thesis follows. □
The best choice of p p p is given in the following result.
Lemma 5.6. Let e ∈ S n − 1 e\in\mathbb{S}^{n-1} e ∈ S n − 1 and let p p p be a probability density on S n − 1 \mathbb{S}^{n-1} S n − 1 with p > 0 p>0 p > 0 H n − 1 \mathcal{H}^{n-1} H n − 1 -a.e. Then
∫ S n − 1 p ( ξ ) 1 − 1 2 s ∣ e ⋅ ξ ∣ d H n − 1 ( ξ ) ≤ ( 2 C s , n ) 1 2 s . (43) \int_{\mathbb{S}^{n-1}}p(\xi)^{1-\frac{1}{2s}}|e\cdot\xi|\,d\mathcal{H}^{n-1}(\xi)\le(2C_{s,n})^{\frac{1}{2s}}. \tag*{(43)} ∫ S n − 1 p ( ξ ) 1 − 2 s 1 ∣ e ⋅ ξ ∣ d H n − 1 ( ξ ) ≤ ( 2 C s , n ) 2 s 1 . (43) Moreover, if
p ( ξ ) = ∣ e ⋅ ξ ∣ 2 s 2 C s , n , p(\xi)=\frac{|e\cdot\xi|^{2s}}{2C_{s,n}}, p ( ξ ) = 2 C s , n ∣ e ⋅ ξ ∣ 2 s , then equality holds in (43) .
Proof. We first prove that, for all e ∈ S n − 1 e\in\mathbb{S}^{n-1} e ∈ S n − 1 ,
∫ S n − 1 ∣ e ⋅ ξ ∣ 2 s d H n − 1 ( ξ ) = 2 C s , n . (44) \int_{\mathbb{S}^{n-1}}|e\cdot\xi|^{2s}\,d\mathcal{H}^{n-1}(\xi)=2C_{s,n}. \tag*{(44)} ∫ S n − 1 ∣ e ⋅ ξ ∣ 2 s d H n − 1 ( ξ ) = 2 C s , n . (44)
Let us decompose R n = e ⊥ ⊕ R e ≃ R n − 1 × R \mathbb{R}^n=e^\perp\oplus\mathbb{R}e\simeq\mathbb{R}^{n-1}\times\mathbb{R} R n = e ⊥ ⊕ R e ≃ R n − 1 × R ; then for each ξ ∈ S n − 1 \xi\in\mathbb{S}^{n-1} ξ ∈ S n − 1 we can write
ξ = ( ξ ′ , ξ n ) ∈ R n − 1 × R with ξ n = ξ ⋅ e and ξ ′ = ξ − ξ n e . \xi=(\xi',\xi_n)\in\mathbb{R}^{n-1}\times\mathbb{R}\quad\text{with}\quad\xi_n=\xi\cdot e\quad\text{and}\quad\xi'=\xi-\xi_n e. ξ = ( ξ ′ , ξ n ) ∈ R n − 1 × R with ξ n = ξ ⋅ e and ξ ′ = ξ − ξ n e .
If ξ n > 0 \xi_n>0 ξ n > 0 (analogously if ξ n < 0 \xi_n<0 ξ n < 0 ), we use the change of variables
ζ : = ξ ′ ξ n ∈ R n − 1 . \zeta:=\frac{\xi'}{\xi_n}\in\mathbb{R}^{n-1}. ζ := ξ n ξ ′ ∈ R n − 1 .
Then
ξ = ( ζ , 1 ) 1 + ∣ ζ ∣ 2 , d H n − 1 = d ζ 1 ( 1 + ∣ ζ ∣ 2 ) n 2 and ∣ ξ ⋅ e ∣ 2 s = 1 ( 1 + ∣ ζ ∣ 2 ) s . \xi=\frac{(\zeta,1)}{\sqrt{1+|\zeta|^2}},\qquad d\mathcal{H}^{n-1}=d\zeta\frac{1}{(1+|\zeta|^2)^{\frac n2}}\qquad\text{and}\qquad|\xi\cdot e|^{2s}=\frac{1}{(1+|\zeta|^2)^s}. ξ = 1 + ∣ ζ ∣ 2 ( ζ , 1 ) , d H n − 1 = d ζ ( 1 + ∣ ζ ∣ 2 ) 2 n 1 and ∣ ξ ⋅ e ∣ 2 s = ( 1 + ∣ ζ ∣ 2 ) s 1 .
In conclusion, rewriting the integral (44) and using the definition of C s , n C_{s,n} C s , n in (9) , we have
∫ S n − 1 ∣ e ⋅ ξ ∣ 2 s d H n − 1 ( ξ ) = 2 ∫ S n − 1 ∩ { ξ n > 0 } ∣ e ⋅ ξ ∣ 2 s d H n − 1 ( ξ ) = 2 ∫ R n − 1 d ζ ( 1 + ∣ ζ ∣ 2 ) n + 2 s 2 = 2 C s , n . \int_{\mathbb{S}^{n-1}} |e\cdot\xi|^{2s}\,d\mathcal{H}^{n-1}(\xi)=2\int_{\mathbb{S}^{n-1}\cap\{\xi_n>0\}} |e\cdot\xi|^{2s}\,d\mathcal{H}^{n-1}(\xi)=2\int_{\mathbb{R}^{n-1}}\frac{d\zeta}{(1+|\zeta|^2)^{\frac{n+2s}{2}}}=2C_{s,n}. ∫ S n − 1 ∣ e ⋅ ξ ∣ 2 s d H n − 1 ( ξ ) = 2 ∫ S n − 1 ∩ { ξ n > 0 } ∣ e ⋅ ξ ∣ 2 s d H n − 1 ( ξ ) = 2 ∫ R n − 1 ( 1 + ∣ ζ ∣ 2 ) 2 n + 2 s d ζ = 2 C s , n .
Now, observe that (43) follows by Hölder’s inequality with exponents 2 s 2s 2 s and 2 s / ( 2 s − 1 ) 2s/(2s-1) 2 s / ( 2 s − 1 ) , indeed
∫ S n − 1 p ( ξ ) 1 − 1 2 s ∣ e ⋅ ξ ∣ ≤ ( ∫ S n − 1 p ( ξ ) ) 1 − 1 2 s ( ∫ S n − 1 ∣ e ⋅ ξ ∣ 2 s ) 1 2 s = ( 2 C s , n ) 1 2 s . \int_{\mathbb{S}^{n-1}}p(\xi)^{1-\frac{1}{2s}}|e\cdot\xi|\leq\left(\int_{\mathbb{S}^{n-1}}p(\xi)\right)^{1-\frac{1}{2s}}\left(\int_{\mathbb{S}^{n-1}}|e\cdot\xi|^{2s}\right)^{\frac{1}{2s}}=(2C_{s,n})^{\frac{1}{2s}}. ∫ S n − 1 p ( ξ ) 1 − 2 s 1 ∣ e ⋅ ξ ∣ ≤ ( ∫ S n − 1 p ( ξ ) ) 1 − 2 s 1 ( ∫ S n − 1 ∣ e ⋅ ξ ∣ 2 s ) 2 s 1 = ( 2 C s , n ) 2 s 1 .
Proof of Proposition 5.3. Using the unified notation defined above we can treat both cases at once. Then, let us assume a , b , V ε , Ψ ( 1 ) a,b,V_\varepsilon,\Psi^{(1)} a , b , V ε , Ψ ( 1 ) as before. We want to prove that: if u ∈ B V ( Q ; { a , b } ) u\in BV(Q;\{a,b\}) u ∈ B V ( Q ; { a , b }) and { u ε } ⊂ L 1 ( Q ; [ a , b ] ) \{u_\varepsilon\}\subset L^1(Q;[a,b]) { u ε } ⊂ L 1 ( Q ; [ a , b ]) converges to u u u in L 1 ( Q ) L^1(Q) L 1 ( Q ) , then
lim inf ε → 0 F ‾ ε ( u ε ) ≥ λ s , n Ψ ( 1 ) ( a , b ) H n − 1 ( J u ) . \liminf_{\varepsilon\to0}\overline{F}_\varepsilon(u_\varepsilon)\geq\lambda_{s,n}\Psi^{(1)}(a,b)\mathcal{H}^{n-1}(J_u). ε → 0 lim inf F ε ( u ε ) ≥ λ s , n Ψ ( 1 ) ( a , b ) H n − 1 ( J u ) .
We define the sequence of Radon measures { μ ε } \{\mu_\varepsilon\} { μ ε } on the cube Q Q Q as follows
μ ε ( A ) : = 1 θ ε ( ε 2 s ∫ Q ∫ A ∣ u ε ( x ) − u ε ( y ) ∣ 2 ∣ x − y ∣ n + 2 s d x d y + ∫ A V ε ( u ε ) ) , for all measurable A ⊂ Q , \mu_\varepsilon(A):=\frac{1}{\theta_\varepsilon}\left(\varepsilon^{2s}\int_Q\int_A\frac{|u_\varepsilon(x)-u_\varepsilon(y)|^2}{|x-y|^{n+2s}}\,dx\,dy+\int_A V_\varepsilon(u_\varepsilon)\right),\qquad\text{for all measurable }A\subset Q, μ ε ( A ) := θ ε 1 ( ε 2 s ∫ Q ∫ A ∣ x − y ∣ n + 2 s ∣ u ε ( x ) − u ε ( y ) ∣ 2 d x d y + ∫ A V ε ( u ε ) ) , for all measurable A ⊂ Q ,
with μ ε ( Q ) = F ‾ ε ( u ε ) \mu_\varepsilon(Q)=\overline{F}_\varepsilon(u_\varepsilon) μ ε ( Q ) = F ε ( u ε ) . By the Besicovitch differentiation theorem, to prove the thesis, it is enough to prove that for H n − 1 \mathcal{H}^{n-1} H n − 1 -a.e. y ∈ J u y\in J_u y ∈ J u , we have
lim inf r → 0 lim inf ε → 0 μ ε ( Q r ) r n − 1 ≥ λ s , n Ψ ( 1 ) ( a , b ) , (45) \liminf_{r\to0}\liminf_{\varepsilon\to0}\frac{\mu_\varepsilon(Q_r)}{r^{n-1}}\geq\lambda_{s,n}\Psi^{(1)}(a,b), \tag*{(45)} r → 0 lim inf ε → 0 lim inf r n − 1 μ ε ( Q r ) ≥ λ s , n Ψ ( 1 ) ( a , b ) , (45)
where, with a little abuse of notation, Q r Q_r Q r denotes the cube centered at y y y , with side r r r , parallel to the normal ν u ( y ) \nu_u(y) ν u ( y ) on J u J_u J u at $y. Observe that F ‾ ε ( u ε ; Q r ) ≤ μ ε ( Q r ) ≤ μ ε ( Q ) ≤ F ε ( u ε ) ≤ sup ε F ε ( u ε ) \overline{F}_\varepsilon(u_\varepsilon;Q_r)\leq\mu_\varepsilon(Q_r)\leq\mu_\varepsilon(Q)\leq F_\varepsilon(u_\varepsilon)\leq\sup_\varepsilon F_\varepsilon(u_\varepsilon) F ε ( u ε ; Q r ) ≤ μ ε ( Q r ) ≤ μ ε ( Q ) ≤ F ε ( u ε ) ≤ sup ε F ε ( u ε ) , where the last term can be assumed finite. Hence, up to a subsequence μ ε ⇀ ∗ μ \mu_\varepsilon\overset{*}{\rightharpoonup}\mu μ ε ⇀ ∗ μ and
μ ( Q ‾ r ) ≥ lim inf ε → 0 μ ε ( Q r ) ≥ lim inf ε → 0 F ‾ ε ( u ε ; Q r ) . \mu(\overline{Q}_r)\geq\liminf_{\varepsilon\to0}\mu_\varepsilon(Q_r)\geq\liminf_{\varepsilon\to0}\overline{F}_\varepsilon(u_\varepsilon;Q_r). μ ( Q r ) ≥ ε → 0 lim inf μ ε ( Q r ) ≥ ε → 0 lim inf F ε ( u ε ; Q r ) .
Now, fixing y ∈ J u y\in J_u y ∈ J u , by Lemma 5.5, for each p ≥ 0 p\geq0 p ≥ 0 with ∫ S n − 1 p ( ξ ) d ξ = 1 \int_{\mathbb{S}^{n-1}}p(\xi)\,d\xi=1 ∫ S n − 1 p ( ξ ) d ξ = 1 and p > 0 p>0 p > 0 H n − 1 \mathcal{H}^{n-1} H n − 1 -a.e., we have
lim inf ε → 0 F ‾ ε ( u ε ; Q r ) ≥ 2 − 1 2 s Ψ ( 1 ) ( a , b ) ∫ J u ∩ Q r d H n − 1 ( x ) ∫ S n − 1 d H n − 1 ( ξ ) p ( ξ ) 1 − 1 2 s ∣ ν u ( x ) ⋅ ξ ∣ ; \liminf_{\varepsilon\to0}\overline{F}_\varepsilon(u_\varepsilon;Q_r)\geq2^{-\frac{1}{2s}}\Psi^{(1)}(a,b)\int_{J_u\cap Q_r}d\mathcal{H}^{n-1}(x)\int_{\mathbb{S}^{n-1}}d\mathcal{H}^{n-1}(\xi)\,p(\xi)^{1-\frac{1}{2s}}|\nu_u(x)\cdot\xi|; ε → 0 lim inf F ε ( u ε ; Q r ) ≥ 2 − 2 s 1 Ψ ( 1 ) ( a , b ) ∫ J u ∩ Q r d H n − 1 ( x ) ∫ S n − 1 d H n − 1 ( ξ ) p ( ξ ) 1 − 2 s 1 ∣ ν u ( x ) ⋅ ξ ∣ ;
in particular, we can choose p ( ξ ) = ∣ ν u ( y ) ⋅ ξ ∣ 2 s ( 2 C s , n ) − 1 p(\xi)=|\nu_u(y)\cdot\xi|^{2s}(2C_{s,n})^{-1} p ( ξ ) = ∣ ν u ( y ) ⋅ ξ ∣ 2 s ( 2 C s , n ) − 1 . Hence, using Fatou’s lemma
lim inf r → 0 lim inf ε → 0 F ‾ ε ( u ε ; Q r ) r n − 1 ≥ Ψ ( 1 ) ( a , b ) ∫ S n − 1 ∣ ν u ( y ) ⋅ ξ ∣ 2 s − 1 2 C s , n 1 − 1 2 s ⋅ ( lim inf r → 0 1 r n − 1 ∫ J u ∩ Q r ∣ ν u ( x ) ⋅ ξ ∣ d H n − 1 ( x ) ) d H n − 1 ( ξ ) . \liminf_{r\to0}\liminf_{\varepsilon\to0}\frac{\overline{F}_\varepsilon(u_\varepsilon;Q_r)}{r^{n-1}}\geq\Psi^{(1)}(a,b)\int_{\mathbb{S}^{n-1}}\frac{|\nu_u(y)\cdot\xi|^{2s-1}}{2C_{s,n}^{1-\frac{1}{2s}}}\cdot\left(\liminf_{r\to0}\frac{1}{r^{n-1}}\int_{J_u\cap Q_r}|\nu_u(x)\cdot\xi|\,d\mathcal{H}^{n-1}(x)\right)d\mathcal{H}^{n-1}(\xi). r → 0 lim inf ε → 0 lim inf r n − 1 F ε ( u ε ; Q r ) ≥ Ψ ( 1 ) ( a , b ) ∫ S n − 1 2 C s , n 1 − 2 s 1 ∣ ν u ( y ) ⋅ ξ ∣ 2 s − 1 ⋅ ( r → 0 lim inf r n − 1 1 ∫ J u ∩ Q r ∣ ν u ( x ) ⋅ ξ ∣ d H n − 1 ( x ) ) d H n − 1 ( ξ ) .
By rectifiability of J u J_u J u , we have
lim r → 0 1 r n − 1 ∫ J u ∩ Q r ∣ ν u ( x ) ⋅ ξ ∣ d H n − 1 ( x ) = ∣ ν u ( y ) ⋅ ξ ∣ . \lim_{r\to0}\frac{1}{r^{n-1}}\int_{J_u\cap Q_r}|\nu_u(x)\cdot\xi|\,d\mathcal{H}^{n-1}(x)=|\nu_u(y)\cdot\xi|. r → 0 lim r n − 1 1 ∫ J u ∩ Q r ∣ ν u ( x ) ⋅ ξ ∣ d H n − 1 ( x ) = ∣ ν u ( y ) ⋅ ξ ∣.
Hence, (45) follows by Lemma 5.6, recalling that λ s , n = ( C s , n ) 1 2 s \lambda_{s,n}=(C_{s,n})^{\frac{1}{2s}} λ s , n = ( C s , n ) 2 s 1 (see (9) ). □
Proof of the lower bound
Proof of Theorem 2.7 (lower bound). The lower bound is an immediate consequence of Proposition 5.3, the truncation inequality (15) and the lower estimate in (13) . We give details for regime (R1); (R2) is similar.
Regime (R1). Let V = Z ∪ S = { v i } i ∈ Z \mathcal{V}=\mathbb{Z}\cup S=\{v_i\}_{i\in\mathbb{Z}} V = Z ∪ S = { v i } i ∈ Z with v i < v i + 1 v_i<v_{i+1} v i < v i + 1 . Let { u ε } ⊂ L 1 ( Q ) \{u_\varepsilon\}\subset L^1(Q) { u ε } ⊂ L 1 ( Q ) be a sequence that converges to u u u in L 1 ( Q ) L^1(Q) L 1 ( Q ) such that F ε ( u ε ) ≤ M F_\varepsilon(u_\varepsilon)\le M F ε ( u ε ) ≤ M . Then, by Theorem 2.5, u ∈ B V ( Q ; V ) u\in BV(Q;\mathcal{V}) u ∈ B V ( Q ; V ) . Let T i T_i T i be the truncations between consecutive points of V \mathcal{V} V in (15) . By (13) and the truncation inequality (15) , for any δ ∈ ( 0 , 1 ) \delta\in(0,1) δ ∈ ( 0 , 1 ) and small ε \varepsilon ε , we have that
F ε ( u ε ) ≥ F ε Z ∪ S ( u ε ) + δ ε θ ε ∣ A δ ( u ε ) ∣ ≥ ∑ i ∈ Z F ε Z ∪ S ( T i u ε ) + δ ε θ ε ∣ A δ ( u ε ) ∣ . (46) F_\varepsilon(u_\varepsilon)\ge F_\varepsilon^{\mathbb{Z}\cup S}(u_\varepsilon)+\frac{\delta_\varepsilon}{\theta_\varepsilon}|A_\delta(u_\varepsilon)|\ge\sum_{i\in\mathbb{Z}}F_\varepsilon^{\mathbb{Z}\cup S}(T_i u_\varepsilon)+\frac{\delta_\varepsilon}{\theta_\varepsilon}|A_\delta(u_\varepsilon)|. \tag*{(46)} F ε ( u ε ) ≥ F ε Z ∪ S ( u ε ) + θ ε δ ε ∣ A δ ( u ε ) ∣ ≥ i ∈ Z ∑ F ε Z ∪ S ( T i u ε ) + θ ε δ ε ∣ A δ ( u ε ) ∣. (46)
Now, for each i ∈ Z i\in\mathbb{Z} i ∈ Z , { T i u ε } \{T_i u_\varepsilon\} { T i u ε } converges in L 1 L^1 L 1 to T i u T_i u T i u and satisfies the hypotheses of Proposition 5.3 (if n ≥ 2 n\ge2 n ≥ 2 ) and Lemma 5.1 (if n = 1 n=1 n = 1 ); hence, we have
lim inf ε → 0 F ε Z ∪ S ( T i u ε ) ≥ Ψ s u p ( v i , v i + 1 ) H n − 1 ( J T i u ) = ∫ J T i u Ψ s u p ( v i , v i + 1 ) d H n − 1 ( x ) \liminf_{\varepsilon\to0}F_\varepsilon^{\mathbb{Z}\cup S}(T_i u_\varepsilon)\ge\Psi_{\mathrm{sup}}(v_i,v_{i+1})\mathcal{H}^{n-1}(J_{T_i u})=\int_{J_{T_i u}}\Psi_{\mathrm{sup}}(v_i,v_{i+1})\,d\mathcal{H}^{n-1}(x) ε → 0 lim inf F ε Z ∪ S ( T i u ε ) ≥ Ψ sup ( v i , v i + 1 ) H n − 1 ( J T i u ) = ∫ J T i u Ψ sup ( v i , v i + 1 ) d H n − 1 ( x )
Now observe that, up to an H n − 1 \mathcal{H}^{n-1} H n − 1 -null set,
J u = ⋃ j ∈ Z J T j u . J_u=\bigcup_{j\in\mathbb{Z}}J_{T_j u}. J u = j ∈ Z ⋃ J T j u .
We can use additivity of Ψ s u p \Psi_{\mathrm{sup}} Ψ sup to conclude that
∑ i ∈ Z Ψ s u p ( v i , v i + 1 ) χ J T i u ( x ) = ∑ i : [ v i , v i + 1 ] ⊂ [ u − ( x ) , u + ( x ) ] Ψ s u p ( v i , v i + 1 ) χ J u ( x ) = Ψ s u p ( u − ( x ) , u + ( x ) ) χ J u ( x ) , \sum_{i\in\mathbb{Z}}\Psi_{\mathrm{sup}}(v_i,v_{i+1})\chi_{J_{T_i u}}(x)=\sum_{i:\,[v_i,v_{i+1}]\subset[u^-(x),u^+(x)]}\Psi_{\mathrm{sup}}(v_i,v_{i+1})\chi_{J_u}(x)
=\Psi_{\mathrm{sup}}(u^-(x),u^+(x))\chi_{J_u}(x), i ∈ Z ∑ Ψ sup ( v i , v i + 1 ) χ J T i u ( x ) = i : [ v i , v i + 1 ] ⊂ [ u − ( x ) , u + ( x )] ∑ Ψ sup ( v i , v i + 1 ) χ J u ( x ) = Ψ sup ( u − ( x ) , u + ( x )) χ J u ( x ) ,
hence
lim inf ε → 0 F ε Z ∪ S ( u ε ) ≥ ∫ J u Ψ s u p ( u + ( x ) , u − ( x ) ) d H n − 1 ( x ) . (47) \liminf_{\varepsilon\to0}F_\varepsilon^{\mathbb{Z}\cup S}(u_\varepsilon)\ge\int_{J_u}\Psi_{\mathrm{sup}}(u^+(x),u^-(x))\,d\mathcal{H}^{n-1}(x). \tag*{(47)} ε → 0 lim inf F ε Z ∪ S ( u ε ) ≥ ∫ J u Ψ sup ( u + ( x ) , u − ( x )) d H n − 1 ( x ) . (47)
Moreover, by the convergence in measure, lim inf ε ∣ A δ ( u ε ) ∣ ≥ ∣ { u ∈ S } ∣ \liminf_{\varepsilon}|A_\delta(u_\varepsilon)|\ge|\{u\in S\}| lim inf ε ∣ A δ ( u ε ) ∣ ≥ ∣ { u ∈ S } ∣ , so that
lim inf ε → 0 δ ε θ ε ∣ A δ ( u ε ) ∣ ≥ K ∣ { u ∈ S } ∣ . (48) \liminf_{\varepsilon\to0}\frac{\delta_\varepsilon}{\theta_\varepsilon}|A_\delta(u_\varepsilon)|\ge K|\{u\in S\}|. \tag*{(48)} ε → 0 lim inf θ ε δ ε ∣ A δ ( u ε ) ∣ ≥ K ∣ { u ∈ S } ∣. (48)
By (46) , using (47) and (48) , we conclude that
lim inf ε → 0 F ε ( u ε ) ≥ F s u p ( u ) . \liminf_{\varepsilon\to0}F_\varepsilon(u_\varepsilon)\ge F_{\mathrm{sup}}(u). ε → 0 lim inf F ε ( u ε ) ≥ F sup ( u ) .
Regime (R2). We repeat the truncation argument with V = Z \mathcal{V}=\mathbb{Z} V = Z , for a sequence { u ε } ⊂ L 1 ( Q ) \{u_\varepsilon\}\subset L^1(Q) { u ε } ⊂ L 1 ( Q ) converging to u u u in L 1 ( Q ) L^1(Q) L 1 ( Q ) , and such that F ε ( u ε ) ≤ M F_\varepsilon(u_\varepsilon)\le M F ε ( u ε ) ≤ M . By Theorem 2.5 we have that u ∈ B V ( Q ; V ) u\in BV(Q;\mathcal{V}) u ∈ B V ( Q ; V ) . By Proposition 5.3 if n ≥ 2 n\ge2 n ≥ 2 and Lemma 5.2 if n = 1 n=1 n = 1 , we obtain
lim inf ε → 0 F ε ( u ε ) ≥ ∫ J u Ψ s u b ( u + ( x ) , u − ( x ) ) d H n − 1 ( x ) = F s u b ( u ) . \liminf_{\varepsilon\to0}F_\varepsilon(u_\varepsilon)\ge\int_{J_u}\Psi_{\mathrm{sub}}(u^+(x),u^-(x))\,d\mathcal{H}^{n-1}(x)=F_{\mathrm{sub}}(u). ε → 0 lim inf F ε ( u ε ) ≥ ∫ J u Ψ sub ( u + ( x ) , u − ( x )) d H n − 1 ( x ) = F sub ( u ) .
Upper bound
The proof of the upper bound is organized according to the fractional parameter s s s and the regime under consideration: s = 1 2 s=\frac{1}{2} s = 2 1 in Section 6.1 and s > 1 2 s>\frac{1}{2} s > 2 1 in Section 6.2. In each case, we first construct a one-dimensional recovery sequence, and then extend the construction to higher dimensions via a flat-interface approximation argument.
We denote by V \mathcal{V} V the set of wells of the limit, i.e. V = Z ∪ S \mathcal{V}=\mathbb{Z}\cup S V = Z ∪ S in regime (R1) and V = Z \mathcal{V}=\mathbb{Z} V = Z in regime (R2), ordered as { v i } i ∈ Z \{v_i\}_{i\in\mathbb{Z}} { v i } i ∈ Z with v i < v i + 1 v_i<v_{i+1} v i < v i + 1 , and by F F F the candidate Γ \Gamma Γ -limit, i.e. F s u p F_{\mathrm{sup}} F sup in regime (R1) and F s u b F_{\mathrm{sub}} F sub in regime (R2). By standard properties of Γ \Gamma Γ -convergence, it suffices to construct, for every u u u in a class D ⊆ B V ( Q ; V ) \mathcal{D}\subseteq BV(Q;\mathcal{V}) D ⊆ B V ( Q ; V ) , dense in energy, a recovery sequence { u ε } \{u_\varepsilon\} { u ε } satisfying
u ε → u in L 1 ( Q ) and lim sup ε → 0 F ε ( u ε ) ≤ F ( u ) . u_\varepsilon\to u\quad\text{in }L^1(Q)\quad\text{and}\quad\limsup_{\varepsilon\to0}F_\varepsilon(u_\varepsilon)\leq F(u). u ε → u in L 1 ( Q ) and ε → 0 lim sup F ε ( u ε ) ≤ F ( u ) .
Let D \mathcal{D} D be the set of functions u ∈ L ∞ ( Q ; V ) ∩ B V ( Q ; V ) u\in L^\infty(Q;\mathcal{V})\cap BV(Q;\mathcal{V}) u ∈ L ∞ ( Q ; V ) ∩ B V ( Q ; V ) whose jump set J u J_u J u is a finite union of ( n − 1 ) (n-1) ( n − 1 ) -dimensional simplices, J u = ⋃ i = 1 m J i J_u=\bigcup_{i=1}^m J^i J u = ⋃ i = 1 m J i , with dim H ( J i ∩ J j ) ≤ n − 2 \dim_{\mathcal{H}}(J^i\cap J^j)\leq n-2 dim H ( J i ∩ J j ) ≤ n − 2 for i ≠ j i\ne j i = j , and such that, for H n − 1 \mathcal{H}^{n-1} H n − 1 -almost every x ∈ J u x\in J_u x ∈ J u , the trace values u + ( x ) u^+(x) u + ( x ) and u − ( x ) u^-(x) u − ( x ) are consecutive elements of V \mathcal{V} V :
{ u + ( x ) , u − ( x ) } = { v j , v j + 1 } for some j ∈ Z . (49) \{u^+(x),u^-(x)\}=\{v_j,v_{j+1}\}\quad\text{for some }j\in\mathbb{Z}. \tag*{(49)} { u + ( x ) , u − ( x )} = { v j , v j + 1 } for some j ∈ Z . (49)
In regime (R2) this means ∣ u + − u − ∣ ( x ) = 1 |u^+-u^-|(x)=1 ∣ u + − u − ∣ ( x ) = 1 ; in regime (R1), the jump amplitude is either s k − k s_k-k s k − k or ( k + 1 ) − s k (k+1)-s_k ( k + 1 ) − s k . The density of the class D \mathcal{D} D is a consequence of [BCG17], which provides an approximation in energy of any u ∈ B V ( Q ; V ) u\in BV(Q;\mathcal{V}) u ∈ B V ( Q ; V ) by functions whose jump sets are finite unions of ( n − 1 ) (n-1) ( n − 1 ) -dimensional simplices. The property (49) can then be obtained by splitting each flat interface into parallel ones which decompose any large jump into jumps between consecutive wells.
In what follows u ∈ D u\in\mathcal{D} u ∈ D . To avoid boundary effects, we extend u u u to a larger cube Q ~ \widetilde{Q} Q containing Q Q Q , reflecting u u u across the faces of Q Q Q ; the extension, still denoted by u u u , has no jumps across ∂ Q \partial Q ∂ Q . Let Q ⋐ Q ′ ⋐ Q ~ Q\Subset Q'\Subset\widetilde{Q} Q ⋐ Q ′ ⋐ Q . In the following, φ ε \varphi_\varepsilon φ ε is the standard mollifier supported in B ε ( 0 ) B_\varepsilon(0) B ε ( 0 ) .
Higher dimension notation. Since the jump set of u u u is a union of simplices, namely J u = ⋃ i = 1 m J i J_u=\bigcup_{i=1}^m J^i J u = ⋃ i = 1 m J i , for each i i i we denote by u i − < u i + u_i^-<u_i^+ u i − < u i + the traces of u u u on J i J^i J i and by ν i \nu^i ν i the unit normal to J i J^i J i pointing towards { u = u i + } \{u=u_i^+\} { u = u i + } ; in regime (R2) we define also the intermediate well s i : = s u i − s^i:=s_{u_i^-} s i := s u i − and w i : = s i − u i − w^i:=s^i-u_i^- w i := s i − u i − . Let
Σ : = ⋃ i ≠ j ( J i ∩ J j ) , N σ : = { x ∈ Q ~ : dist ( x , J u ) < σ } . \Sigma:=\bigcup_{i\ne j}(J^i\cap J^j),\qquad N_\sigma:=\{x\in\widetilde{Q}:\operatorname{dist}(x,J_u)<\sigma\}. Σ := i = j ⋃ ( J i ∩ J j ) , N σ := { x ∈ Q : dist ( x , J u ) < σ } .
Since Σ \Sigma Σ is contained in a finite union of ( n − 2 ) (n-2) ( n − 2 ) -dimensional simplices, for σ \sigma σ small
∣ N σ ∣ ≤ C σ , ∣ { x ∈ Q ~ : dist ( x , Σ ) < σ } ∣ ≤ C σ 2 , (50) |N_\sigma|\leq C\sigma,\qquad\left|\{x\in\widetilde{Q}:\operatorname{dist}(x,\Sigma)<\sigma\}\right|\leq C\sigma^2, \tag*{(50)} ∣ N σ ∣ ≤ C σ , { x ∈ Q : dist ( x , Σ ) < σ } ≤ C σ 2 , (50)
where C C C depends on u u u .
Upper bound for s = 1 2 s=\frac{1}{2} s = 2 1
For s = 1 2 s=\frac{1}{2} s = 2 1 the energy is dominated by the logarithmic divergence of the H 1 / 2 H^{1/2} H 1/2 seminorm. In regime (R1) the recovery sequence is a transition between consecutive wells at scale ε \varepsilon ε . In regime (R2), in addition, it rests at the intermediate well on a layer of width about ε / δ ε \varepsilon/\delta_\varepsilon ε / δ ε .
Dimension one
Proof of Theorem 2.7 . (Upper bound, n = 1 n=1 n = 1 , s = 1 2 s=\frac{1}{2} s = 2 1 , (R2)) We extend u u u to R \mathbb{R} R with a constant without creating a jump point at ∂ Q \partial Q ∂ Q (in the two extremal points of the interval). Let ρ : = min { ∣ x − y ∣ : x , y ∈ J u , x ≠ y } \rho:=\min\{|x-y|:x,y\in J_u,\ x\ne y\} ρ := min { ∣ x − y ∣ : x , y ∈ J u , x = y } (ρ = 1 \rho=1 ρ = 1 if J u = { x } J_u=\{x\} J u = { x } ). Let ε 0 \varepsilon_0 ε 0 be such that δ ε < 3 / 4 \delta_\varepsilon<3/4 δ ε < 3/4 and 4 ε / δ ε < ρ 4\varepsilon/\delta_\varepsilon<\rho 4 ε / δ ε < ρ for ε < ε 0 \varepsilon<\varepsilon_0 ε < ε 0 (this is possible by (6) ). For x 0 ∈ J u x_0\in J_u x 0 ∈ J u , let u − ( x 0 ) < u + ( x 0 ) u^-(x_0)<u^+(x_0) u − ( x 0 ) < u + ( x 0 ) and be the two traces of u u u at x 0 x_0 x 0 , let s ( x 0 ) : = s u − ( x 0 ) s(x_0) := s_{u^-}(x_0) s ( x 0 ) := s u − ( x 0 ) be the intermediate well and w ( x 0 ) = s ( x 0 ) − u − ( x 0 ) w(x_0) = s(x_0)-u^-(x_0) w ( x 0 ) = s ( x 0 ) − u − ( x 0 ) . We define
u ~ ε : = { s ( x 0 ) , in ( x 0 − ε δ ε , x 0 + ε δ ε ) , x 0 ∈ J u , u , otherwise , and u ε : = u ~ ε ∗ φ ε . \widetilde{u}_{\varepsilon}:=
\begin{cases}
s(x_0), & \text{in } \left(x_0-\frac{\varepsilon}{\delta_\varepsilon},x_0+\frac{\varepsilon}{\delta_\varepsilon}\right),\ x_0\in J_u,\\
u, & \text{otherwise},
\end{cases}
\qquad\text{and}\qquad
u_\varepsilon:=\widetilde{u}_\varepsilon* \varphi_\varepsilon. u ε := { s ( x 0 ) , u , in ( x 0 − δ ε ε , x 0 + δ ε ε ) , x 0 ∈ J u , otherwise , and u ε := u ε ∗ φ ε .
By definition, ∥ u − u ε ∥ ∞ ≤ 1 \lVert u-u_\varepsilon\rVert_\infty\leq1 ∥ u − u ε ∥ ∞ ≤ 1 , and ∣ { u ≠ u ε } ∣ ≤ 4 H 0 ( J u ) ε / δ ε \lvert\{u\neq u_\varepsilon\}\rvert\leq4\mathcal{H}^0(J_u)\varepsilon/\delta_\varepsilon ∣{ u = u ε }∣ ≤ 4 H 0 ( J u ) ε / δ ε , so u ε → u u_\varepsilon\to u u ε → u in L 1 ( Q ) L^1(Q) L 1 ( Q ) .
Now we estimate the H 1 / 2 H^{1/2} H 1/2 -seminorm. First of all, we notice that we can ignore the interaction between distant points, i.e.,
[ u ε ] H 1 / 2 ( Q ) 2 = O ( 1 ) + ∑ x 0 ∈ J u [ u ε ] H 1 / 2 ( B ρ / 2 ( x 0 ) ) 2 . [u_\varepsilon]^2_{H^{1/2}(Q)}=O(1)+\sum_{x_0\in J_u}[u_\varepsilon]^2_{H^{1/2}(B_{\rho/2}(x_0))}. [ u ε ] H 1/2 ( Q ) 2 = O ( 1 ) + x 0 ∈ J u ∑ [ u ε ] H 1/2 ( B ρ /2 ( x 0 )) 2 .
Fixing x 0 ∈ J u x_0\in J_u x 0 ∈ J u , in B ρ / 2 ( x 0 ) B_{\rho/2}(x_0) B ρ /2 ( x 0 ) let A ε : = { u ε = u − ( x 0 ) } A_\varepsilon:=\{u_\varepsilon=u^-(x_0)\} A ε := { u ε = u − ( x 0 )} , P ε : = { u ε = s ( x 0 ) } P_\varepsilon:=\{u_\varepsilon=s(x_0)\} P ε := { u ε = s ( x 0 )} and C ε : = { u ε = u + ( x 0 ) } C_\varepsilon:=\{u_\varepsilon=u^+(x_0)\} C ε := { u ε = u + ( x 0 )} . Then, using that u ε u_\varepsilon u ε is C / ε C/\varepsilon C / ε -Lipschitz, the contribution of the remaining interactions is O ( 1 ) O(1) O ( 1 ) . Hence, we have
[ u ε ] H 1 / 2 ( B ρ / 2 ( x 0 ) ) 2 = O ( 1 ) + 2 ∫ A ε ∫ P ε ∣ u ε ( x ) − u ε ( y ) ∣ 2 ∣ x − y ∣ 2 d x d y + 2 ∫ A ε ∫ C ε ∣ u ε ( x ) − u ε ( y ) ∣ 2 ∣ x − y ∣ 2 d x d y + 2 ∫ C ε ∫ P ε ∣ u ε ( x ) − u ε ( y ) ∣ 2 ∣ x − y ∣ 2 d x d y = O ( 1 ) + 2 w ( x 0 ) 2 ∫ A ε ∫ P ε d x d y ∣ x − y ∣ 2 + 2 ∫ A ε ∫ C ε 1 ∣ x − y ∣ 2 d x d y + 2 ( 1 − w ( x 0 ) ) 2 ∫ C ε ∫ P ε d x d y ∣ x − y ∣ 2 . \begin{aligned}
[u_\varepsilon]^2_{H^{1/2}(B_{\rho/2}(x_0))}
={}&O(1)+2\int_{A_\varepsilon}\int_{P_\varepsilon}\frac{\lvert u_\varepsilon(x)-u_\varepsilon(y)\rvert^2}{\lvert x-y\rvert^2}\,dx\,dy\\
&+2\int_{A_\varepsilon}\int_{C_\varepsilon}\frac{\lvert u_\varepsilon(x)-u_\varepsilon(y)\rvert^2}{\lvert x-y\rvert^2}\,dx\,dy
+2\int_{C_\varepsilon}\int_{P_\varepsilon}\frac{\lvert u_\varepsilon(x)-u_\varepsilon(y)\rvert^2}{\lvert x-y\rvert^2}\,dx\,dy\\
={}&O(1)+2w(x_0)^2\int_{A_\varepsilon}\int_{P_\varepsilon}\frac{dx\,dy}{\lvert x-y\rvert^2}\\
&+2\int_{A_\varepsilon}\int_{C_\varepsilon}\frac{1}{\lvert x-y\rvert^2}\,dx\,dy
+2(1-w(x_0))^2\int_{C_\varepsilon}\int_{P_\varepsilon}\frac{dx\,dy}{\lvert x-y\rvert^2}.
\end{aligned} [ u ε ] H 1/2 ( B ρ /2 ( x 0 )) 2 = = O ( 1 ) + 2 ∫ A ε ∫ P ε ∣ x − y ∣ 2 ∣ u ε ( x ) − u ε ( y ) ∣ 2 d x d y + 2 ∫ A ε ∫ C ε ∣ x − y ∣ 2 ∣ u ε ( x ) − u ε ( y ) ∣ 2 d x d y + 2 ∫ C ε ∫ P ε ∣ x − y ∣ 2 ∣ u ε ( x ) − u ε ( y ) ∣ 2 d x d y O ( 1 ) + 2 w ( x 0 ) 2 ∫ A ε ∫ P ε ∣ x − y ∣ 2 d x d y + 2 ∫ A ε ∫ C ε ∣ x − y ∣ 2 1 d x d y + 2 ( 1 − w ( x 0 ) ) 2 ∫ C ε ∫ P ε ∣ x − y ∣ 2 d x d y .
By (39) , since A ε A_\varepsilon A ε , P ε P_\varepsilon P ε , C ε C_\varepsilon C ε are intervals separated by gaps of lengths at least ε \varepsilon ε and P ε P_\varepsilon P ε has length proportional to ε / δ ε − ε \varepsilon/\delta_\varepsilon-\varepsilon ε / δ ε − ε , we have
∫ C ε ∫ P ε d x d y ∣ x − y ∣ 2 = ∫ A ε ∫ P ε d x d y ∣ x − y ∣ 2 = ∣ log δ ε ∣ + O ( 1 ) \int_{C_\varepsilon}\int_{P_\varepsilon}\frac{dx\,dy}{\lvert x-y\rvert^2}
=\int_{A_\varepsilon}\int_{P_\varepsilon}\frac{dx\,dy}{\lvert x-y\rvert^2}
=\lvert\log\delta_\varepsilon\rvert+O(1) ∫ C ε ∫ P ε ∣ x − y ∣ 2 d x d y = ∫ A ε ∫ P ε ∣ x − y ∣ 2 d x d y = ∣ log δ ε ∣ + O ( 1 )
and
∫ C ε ∫ A ε d x d y ∣ x − y ∣ 2 = log δ ε ε + O ( 1 ) . \int_{C_\varepsilon}\int_{A_\varepsilon}\frac{dx\,dy}{\lvert x-y\rvert^2}
=\log\frac{\delta_\varepsilon}{\varepsilon}+O(1). ∫ C ε ∫ A ε ∣ x − y ∣ 2 d x d y = log ε δ ε + O ( 1 ) .
For the potential, we have
∫ Q W ε ( u ε ( x ) ) d x = ∑ y ∈ J u ∫ B ρ / 2 ( y ) W ε ( u ε ( x ) ) d x + C ε ≤ H 0 ( J u ) ∫ B ρ / 2 ( x 0 ) W ε ( u ε ( x ) ) d x + C ε , \int_Q W_\varepsilon(u_\varepsilon(x))\,dx
=\sum_{y\in J_u}\int_{B_{\rho/2}(y)}W_\varepsilon(u_\varepsilon(x))\,dx+C\varepsilon
\leq\mathcal{H}^0(J_u)\int_{B_{\rho/2}(x_0)}W_\varepsilon(u_\varepsilon(x))\,dx+C\varepsilon, ∫ Q W ε ( u ε ( x )) d x = y ∈ J u ∑ ∫ B ρ /2 ( y ) W ε ( u ε ( x )) d x + C ε ≤ H 0 ( J u ) ∫ B ρ /2 ( x 0 ) W ε ( u ε ( x )) d x + C ε ,
with x 0 ∈ J u x_0\in J_u x 0 ∈ J u . Now, since W ε ( s ( x 0 ) ) = δ ε W_\varepsilon(s(x_0))=\delta_\varepsilon W ε ( s ( x 0 )) = δ ε and it is bounded, we have
∫ B ρ / 2 ( x 0 ) W ε ( u ε ( x ) ) d x ≤ ∣ P ε ∣ δ ε + ε sup ε W ε ≤ C ε . \int_{B_{\rho/2}(x_0)}W_\varepsilon(u_\varepsilon(x))\,dx
\leq\lvert P_\varepsilon\rvert\delta_\varepsilon+\varepsilon\sup_\varepsilon W_\varepsilon
\leq C\varepsilon. ∫ B ρ /2 ( x 0 ) W ε ( u ε ( x )) d x ≤ ∣ P ε ∣ δ ε + ε ε sup W ε ≤ C ε .
In conclusion, since θ ε = ε ∣ log ε ∣ \theta_\varepsilon=\varepsilon\lvert\log\varepsilon\rvert θ ε = ε ∣ log ε ∣ and ∣ log δ ε ∣ / ∣ log ε ∣ → ℓ \lvert\log\delta_\varepsilon\rvert/\lvert\log\varepsilon\rvert\to\ell ∣ log δ ε ∣ / ∣ log ε ∣ → ℓ , we have
lim sup ε → 0 F ε ( u ε ) ≤ lim sup ε → 0 1 ∣ log ε ∣ ∑ x 0 ∈ J u [ 2 ∣ log ε ∣ − ( 2 − 2 w ( x 0 ) 2 − 2 ( 1 − w ( x 0 ) ) 2 ) ∣ log δ ε ∣ ] = ∑ x 0 ∈ J u ( 2 − 4 w ( x 0 ) ( 1 − w ( x 0 ) ) ℓ ) = F s u b ( u ) . \begin{aligned}
\limsup_{\varepsilon\to0}F_\varepsilon(u_\varepsilon)
\leq\limsup_{\varepsilon\to0}\frac{1}{\lvert\log\varepsilon\rvert}
\sum_{x_0\in J_u}\left[2\lvert\log\varepsilon\rvert-\left(2-2w(x_0)^2-2(1-w(x_0))^2\right)\lvert\log\delta_\varepsilon\rvert\right] \\
&=\sum_{x_0\in J_u}\left(2-4w(x_0)(1-w(x_0))\ell\right)=F_{\mathrm{sub}}(u).
\end{aligned} ε → 0 lim sup F ε ( u ε ) ≤ ε → 0 lim sup ∣ log ε ∣ 1 x 0 ∈ J u ∑ [ 2 ∣ log ε ∣ − ( 2 − 2 w ( x 0 ) 2 − 2 ( 1 − w ( x 0 ) ) 2 ) ∣ log δ ε ∣ ] = x 0 ∈ J u ∑ ( 2 − 4 w ( x 0 ) ( 1 − w ( x 0 )) ℓ ) = F sub ( u ) .
Remark 6.1. The specific shape of u ε u_\varepsilon u ε near the jump is not relevant for the upper bound, as already observed in [3 ] and [19 ] . Indeed, in the previous computation we only used that, near each jump point x 0 x_0 x 0 , u ε u_\varepsilon u ε is C / ε C/\varepsilon C / ε -Lipschitz, that ∣ { u ε = s ( x 0 ) } ∣ ∼ ε / δ ε \lvert\{u_\varepsilon=s(x_0)\}\rvert\sim\varepsilon/\delta_\varepsilon ∣{ u ε = s ( x 0 )}∣ ∼ ε / δ ε and that u ε u_\varepsilon u ε takes values outside { u − ( x 0 ) , s ( x 0 ) , u + ( x 0 ) } \{u^-(x_0),s(x_0),u^+(x_0)\} { u − ( x 0 ) , s ( x 0 ) , u + ( x 0 )} on a set of measure O ( ε ) O(\varepsilon) O ( ε ) .
Proof of Theorem 2.7 . (Upper bound, n = 1 n=1 n = 1 , s = 1 2 s=\frac{1}{2} s = 2 1 , (R1)) We extend u u u to R \mathbb{R} R with a constant without creating a jump point at ∂ Q \partial Q ∂ Q (in the two extremal points of the interval). Let u ε : = u ∗ φ ε u_\varepsilon:=u\ast\varphi_\varepsilon u ε := u ∗ φ ε . Then { u ε ≠ u } ⊂ { x ∈ Q : dist ( x , J u ) < ε } \{u_\varepsilon\ne u\}\subset\{x\in Q:\operatorname{dist}(x,J_u)<\varepsilon\} { u ε = u } ⊂ { x ∈ Q : dist ( x , J u ) < ε } , so u ε → u u_\varepsilon\to u u ε → u in L 1 ( Q ) L^1(Q) L 1 ( Q ) and ∣ A 1 ( u ε ) ∣ → ∣ { u ∈ S } ∣ \lvert A_1(u_\varepsilon)\rvert\to\lvert\{u\in S\}\rvert ∣ A 1 ( u ε )∣ → ∣{ u ∈ S }∣ . Hence, by the upper estimate (13) ,
lim sup ε → 0 F ε ( u ε ) ≤ lim sup ε → 0 F ε Z U S ( u ε ) + K ∣ { u ∈ S } ∣ . \limsup_{\varepsilon\to0}F_\varepsilon(u_\varepsilon)\le\limsup_{\varepsilon\to0}F_\varepsilon^{ZU\!S}(u_\varepsilon)+K\lvert\{u\in S\}\rvert. ε → 0 lim sup F ε ( u ε ) ≤ ε → 0 lim sup F ε Z U S ( u ε ) + K ∣{ u ∈ S }∣ .
Near each x 0 ∈ J u x_0\in J_u x 0 ∈ J u , the sequence u ε u_\varepsilon u ε performs a transition between two consecutive wells u − ( x 0 ) , u + ( x 0 ) ∈ Z ∪ S u^-(x_0),u^+(x_0)\in Z\cup S u − ( x 0 ) , u + ( x 0 ) ∈ Z ∪ S within a layer of thickness 2 ε 2\varepsilon 2 ε . The computation in the previous proof (with P ε = ∅ P_\varepsilon=\varnothing P ε = ∅ ) gives
[ u ε ] H 1 / 2 ( Q ) 2 = 2 ∑ x 0 ( u + ( x 0 ) − u − ( x 0 ) ) 2 ∣ log ε ∣ + O ( 1 ) and ∫ Q dist 2 ( u ε , Z ∪ S ) ≤ C ε . [u_\varepsilon]_{H^{1/2}(Q)}^2=2\sum_{x_0}\bigl(u^+(x_0)-u^-(x_0)\bigr)^2\lvert\log\varepsilon\rvert+O(1)
\quad\text{and}\quad
\int_Q\operatorname{dist}^2(u_\varepsilon,Z\cup S)\le C\varepsilon. [ u ε ] H 1/2 ( Q ) 2 = 2 x 0 ∑ ( u + ( x 0 ) − u − ( x 0 ) ) 2 ∣ log ε ∣ + O ( 1 ) and ∫ Q dist 2 ( u ε , Z ∪ S ) ≤ C ε .
Hence
lim sup ε → 0 F ε ( u ε ) ≤ 2 ∑ x 0 ( u + ( x 0 ) − u − ( x 0 ) ) 2 + K ∣ { u ∈ S } ∣ = F s u p ( u ) . \limsup_{\varepsilon\to0}F_\varepsilon(u_\varepsilon)\le2\sum_{x_0}\bigl(u^+(x_0)-u^-(x_0)\bigr)^2+K\lvert\{u\in S\}\rvert=F_{\mathrm{sup}}(u). ε → 0 lim sup F ε ( u ε ) ≤ 2 x 0 ∑ ( u + ( x 0 ) − u − ( x 0 ) ) 2 + K ∣{ u ∈ S }∣ = F sup ( u ) .
Higher dimension
The extension to dimension n ≥ 2 n\ge2 n ≥ 2 is based on the concentration of the energy in the jump set J u J_u J u , that we capture by analyzing the asymptotic properties of the measures associated with the fractional seminorm.
For { u ε } ⊂ L 1 ( Q ′ ) \{u_\varepsilon\}\subset L^1(Q') { u ε } ⊂ L 1 ( Q ′ ) , we define the measures μ ε \mu_\varepsilon μ ε on Q ′ ‾ × Q ′ ‾ \overline{Q'}\times\overline{Q'} Q ′ × Q ′ by
μ ε ( A × B ) : = 1 ∣ log ε ∣ ∫ A ∫ B ∣ u ε ( x ) − u ε ( y ) ∣ 2 ∣ x − y ∣ n + 1 d x d y , A , B ⊂ Q ′ measurable . (51) \mu_\varepsilon(A\times B):=\frac{1}{\lvert\log\varepsilon\rvert}\int_A\int_B\frac{\lvert u_\varepsilon(x)-u_\varepsilon(y)\rvert^2}{\lvert x-y\rvert^{n+1}}\,dx\,dy,\qquad A,B\subset Q'\text{ measurable}. \tag*{(51)} μ ε ( A × B ) := ∣ log ε ∣ 1 ∫ A ∫ B ∣ x − y ∣ n + 1 ∣ u ε ( x ) − u ε ( y ) ∣ 2 d x d y , A , B ⊂ Q ′ measurable . (51)
Lemma 6.2. Let { u ε } \{u_\varepsilon\} { u ε } be a sequence such that ∥ u ε ∥ L ∞ ( Q ′ ) ≤ M \lVert u_\varepsilon\rVert_{L^\infty(Q')}\le M ∥ u ε ∥ L ∞ ( Q ′ ) ≤ M and sup ε μ ε ( Q ′ × Q ′ ) < + ∞ \sup_\varepsilon\mu_\varepsilon(Q'\times Q')<+\infty sup ε μ ε ( Q ′ × Q ′ ) < + ∞ with μ ε \mu_\varepsilon μ ε defined by (51) , and let μ \mu μ be any weak* limit of μ ε \mu_\varepsilon μ ε (up to a subsequence). Then μ \mu μ is concentrated on the diagonal D : = { ( x , x ) : x ∈ Q ′ ‾ } D:=\{(x,x):x\in\overline{Q'}\} D := {( x , x ) : x ∈ Q ′ } and λ ( A ) : = μ ( A × A ) \lambda(A):=\mu(A\times A) λ ( A ) := μ ( A × A ) defines a Radon measure on Q ′ ‾ \overline{Q'} Q ′ .
Proof. To prove that μ \mu μ is concentrated on the diagonal D D D , it is enough to show that for any continuous non-negative function φ : Q ′ ‾ × Q ′ ‾ → R \varphi:\overline{Q'}\times\overline{Q'}\to\mathbb{R} φ : Q ′ × Q ′ → R such that supp φ ∩ D = ∅ \operatorname{supp}\varphi\cap D=\varnothing supp φ ∩ D = ∅ , we have
∫ φ d μ = 0. \int\varphi\,d\mu=0. ∫ φ d μ = 0.
Let δ : = dist ( supp φ , D ) > 0 \delta:=\operatorname{dist}(\operatorname{supp}\varphi,D)>0 δ := dist ( supp φ , D ) > 0 , so that ∣ x − y ∣ ≥ δ \lvert x-y\rvert\ge\delta ∣ x − y ∣ ≥ δ for all ( x , y ) ∈ supp φ (x,y)\in\operatorname{supp}\varphi ( x , y ) ∈ supp φ . Since − M ≤ u ε ≤ M -M\le u_\varepsilon\le M − M ≤ u ε ≤ M , we obtain
∫ φ ( x , y ) d μ = lim ε → 0 1 ∣ log ε ∣ ∫ Q ′ ∫ Q ′ φ ( x , y ) ∣ u ε ( x ) − u ε ( y ) ∣ 2 ∣ x − y ∣ n + 1 d x d y ≤ lim ε → 0 4 M 2 ∥ φ ∥ ∞ ∣ Q ′ ∣ 2 ∣ log ε ∣ δ n + 1 = 0. \int\varphi(x,y)\,d\mu=\lim_{\varepsilon\to0}\frac{1}{\lvert\log\varepsilon\rvert}\int_{Q'}\int_{Q'}\varphi(x,y)\frac{\lvert u_\varepsilon(x)-u_\varepsilon(y)\rvert^2}{\lvert x-y\rvert^{n+1}}\,dx\,dy\le\lim_{\varepsilon\to0}\frac{4M^2\lVert\varphi\rVert_\infty\lvert Q'\rvert^2}{\lvert\log\varepsilon\rvert\delta^{n+1}}=0. ∫ φ ( x , y ) d μ = ε → 0 lim ∣ log ε ∣ 1 ∫ Q ′ ∫ Q ′ φ ( x , y ) ∣ x − y ∣ n + 1 ∣ u ε ( x ) − u ε ( y ) ∣ 2 d x d y ≤ ε → 0 lim ∣ log ε ∣ δ n + 1 4 M 2 ∥ φ ∥ ∞ ∣ Q ′ ∣ 2 = 0.
Hence μ \mu μ is concentrated on D D D .
Proof of Theorem 2.7 . (Upper bound, n ≥ 2 n\ge2 n ≥ 2 , s = 1 2 s=\frac{1}{2} s = 2 1 ) Let σ ε : = ε / δ ε \sigma_\varepsilon:=\varepsilon/\delta_\varepsilon σ ε := ε / δ ε in regime (R2) and σ ε : = ε \sigma_\varepsilon:=\varepsilon σ ε := ε in regime (R1).
Step 1. Definition of the recovery sequence. In regime (R1) we set u ε : = u ∗ φ ε u_\varepsilon:=u\ast\varphi_\varepsilon u ε := u ∗ φ ε . In regime (R2) we modify u u u near each unit jump on J i J^i J i by inserting a layer where the function takes the intermediate value s i s^i s i . In order to keep these layers pairwise disjoint and away from Σ \Sigma Σ , we shrink each simplex. For i = 1 , … , m i=1,\ldots,m i = 1 , … , m , let d i ( ⋅ ) : = dist ( ⋅ , J i ) d^{i}(\cdot):=\operatorname{dist}(\cdot,J^{i}) d i ( ⋅ ) := dist ( ⋅ , J i ) and
R ε i : = { x ∈ Q ~ : d i ( x ) ≤ ε δ ε and d j ( x ) > 4 ε δ ε , ∀ j ≠ i } , J ~ ε i : = { x ∈ J i : x + t v i ∈ I i , ∀ ∣ t ∣ ≤ ε δ ε } , R^{i}_{\varepsilon}:=\left\{x\in\widetilde{Q}:d^{i}(x)\le\frac{\varepsilon}{\delta_{\varepsilon}}\text{ and }d^{j}(x)>\frac{4\varepsilon}{\delta_{\varepsilon}},\ \forall j\ne i\right\},\qquad\widetilde{J}^{i}_{\varepsilon}:=\left\{x\in J^{i}:x+tv^{i}\in I^{i},\ \forall|t|\le\frac{\varepsilon}{\delta_{\varepsilon}}\right\}, R ε i := { x ∈ Q : d i ( x ) ≤ δ ε ε and d j ( x ) > δ ε 4 ε , ∀ j = i } , J ε i := { x ∈ J i : x + t v i ∈ I i , ∀∣ t ∣ ≤ δ ε ε } ,
P ε i : = { y + t v i : y ∈ J ~ ε i , ∣ t ∣ ≤ ε δ ε } . P^{i}_{\varepsilon}:=\left\{y+tv^{i}:y\in\widetilde{J}^{i}_{\varepsilon},\ |t|\le\frac{\varepsilon}{\delta_{\varepsilon}}\right\}. P ε i := { y + t v i : y ∈ J ε i , ∣ t ∣ ≤ δ ε ε } .
Then H n − 1 ( J ~ ε i ) → H n − 1 ( J i ) \mathcal{H}^{n-1}(\widetilde{J}^{i}_{\varepsilon})\to\mathcal{H}^{n-1}(J^{i}) H n − 1 ( J ε i ) → H n − 1 ( J i ) , when ε → 0 \varepsilon\to0 ε → 0 and the sets P ε i + B ε / δ ε ( 0 ) P^{i}_{\varepsilon}+B_{\varepsilon/\delta_{\varepsilon}}(0) P ε i + B ε / δ ε ( 0 ) are pairwise disjoint. Now, we define the sequence u ~ ε \widetilde{u}_{\varepsilon} u ε , as follows:
u ~ ε = { s i , in P ε i , u , otherwise in Q ~ and u ε : = u ~ ε ∗ φ ε . \widetilde{u}_{\varepsilon}=
\begin{cases}
s^{i},&\text{in }P^{i}_{\varepsilon},\\
u,&\text{otherwise in }\widetilde{Q}
\end{cases}
\qquad\text{and}\qquad
u_{\varepsilon}:=\widetilde{u}_{\varepsilon}*\varphi_{\varepsilon}. u ε = { s i , u , in P ε i , otherwise in Q and u ε := u ε ∗ φ ε .
In both regimes, we have ∥ u ε ∥ ∞ ≤ M : = ∥ u ∥ ∞ \|u_{\varepsilon}\|_{\infty}\le M:=\|u\|_{\infty} ∥ u ε ∥ ∞ ≤ M := ∥ u ∥ ∞ and u = u ε u=u_{\varepsilon} u = u ε in the set Q ~ ∖ N 4 σ ε \widetilde{Q}\setminus\mathcal{N}_{4\sigma_{\varepsilon}} Q ∖ N 4 σ ε . Then
∥ u ε − u ∥ L 1 ( Q ) ≤ 2 M ∣ N 4 σ ε ∣ → 0 , as ε → 0. \|u_{\varepsilon}-u\|_{L^{1}(Q)}\le2M|\mathcal{N}_{4\sigma_{\varepsilon}}|\to0,\qquad\text{as }\varepsilon\to0. ∥ u ε − u ∥ L 1 ( Q ) ≤ 2 M ∣ N 4 σ ε ∣ → 0 , as ε → 0.
Moreover, ∣ ∇ u ε ∣ ≤ C / ε |\nabla u_{\varepsilon}|\le C/\varepsilon ∣∇ u ε ∣ ≤ C / ε .
Step 2. Potential term. In regime (R1), using the upper estimate (13) , the fact that u ε = u u_{\varepsilon}=u u ε = u outside N 4 ε \mathcal{N}_{4\varepsilon} N 4 ε , the boundedness of the potential, and that ∣ A 1 ( u ε ) ∣ → ∣ { u ∈ S } ∣ |A_{1}(u_{\varepsilon})|\to|\{u\in S\}| ∣ A 1 ( u ε ) ∣ → ∣ { u ∈ S } ∣ by convergence in measure (for the definition of A 1 A_{1} A 1 see (14) ), we have
1 ε ∣ log ε ∣ ∫ Q W ε ( u ε ( x ) ) d x ≤ C ∣ N 4 ε ∣ ε ∣ log ε ∣ + δ ε ε ∣ log ε ∣ ∣ A 1 ( u ε ) ∣ → K ∣ { u ∈ S } ∣ . \frac{1}{\varepsilon|\log\varepsilon|}\int_{Q}W_{\varepsilon}(u_{\varepsilon}(x))\,dx\le\frac{C|\mathcal{N}_{4\varepsilon}|}{\varepsilon|\log\varepsilon|}+\frac{\delta_{\varepsilon}}{\varepsilon|\log\varepsilon|}|A_{1}(u_{\varepsilon})|\to K|\{u\in S\}|. ε ∣ log ε ∣ 1 ∫ Q W ε ( u ε ( x )) d x ≤ ε ∣ log ε ∣ C ∣ N 4 ε ∣ + ε ∣ log ε ∣ δ ε ∣ A 1 ( u ε ) ∣ → K ∣ { u ∈ S } ∣.
In regime (R2), we have that { u ε ∉ Z ∪ S } ⊂ Z ε : = N ε ∪ ⋃ i { dist ( x , ∂ P ε i ) ≤ ε } \{u_{\varepsilon}\notin Z\cup S\}\subset Z_{\varepsilon}:=\mathcal{N}_{\varepsilon}\cup\bigcup_{i}\{\operatorname{dist}(x,\partial P^{i}_{\varepsilon})\le\varepsilon\} { u ε ∈ / Z ∪ S } ⊂ Z ε := N ε ∪ ⋃ i { dist ( x , ∂ P ε i ) ≤ ε } , with ∣ Z ε ∣ ≤ C ε |Z_{\varepsilon}|\le C\varepsilon ∣ Z ε ∣ ≤ C ε . On the other hand, { u ε ∈ S } ⊂ P ε : = ⋃ i P ε i \{u_{\varepsilon}\in S\}\subset P_{\varepsilon}:=\bigcup_{i}P^{i}_{\varepsilon} { u ε ∈ S } ⊂ P ε := ⋃ i P ε i with ∣ P ε ∣ ≤ C ε / δ ε |P_{\varepsilon}|\le C\varepsilon/\delta_{\varepsilon} ∣ P ε ∣ ≤ C ε / δ ε . Hence
1 ε ∣ log ε ∣ ∫ Q W ε ( u ε ( x ) ) d x ≤ C ∣ Z ε ∣ + δ ε ∣ P ε ∣ ε ∣ log ε ∣ ≤ C ∣ log ε ∣ → 0. \frac{1}{\varepsilon|\log\varepsilon|}\int_{Q}W_{\varepsilon}(u_{\varepsilon}(x))\,dx\le\frac{C|Z_{\varepsilon}|+\delta_{\varepsilon}|P_{\varepsilon}|}{\varepsilon|\log\varepsilon|}\le\frac{C}{|\log\varepsilon|}\to0. ε ∣ log ε ∣ 1 ∫ Q W ε ( u ε ( x )) d x ≤ ε ∣ log ε ∣ C ∣ Z ε ∣ + δ ε ∣ P ε ∣ ≤ ∣ log ε ∣ C → 0.
Step 3. Boundedness of the energy measures. Let { μ ε } \{\mu_{\varepsilon}\} { μ ε } be the sequence of measures on Q ‾ ′ × Q ‾ ′ \overline{Q}'\times\overline{Q}' Q ′ × Q ′ defined in (51) . We show that
sup ε < ε 0 μ ε ( Q ′ × Q ′ ) < + ∞ , \sup_{\varepsilon<\varepsilon_{0}}\mu_{\varepsilon}(Q'\times Q')<+\infty, ε < ε 0 sup μ ε ( Q ′ × Q ′ ) < + ∞ ,
and therefore { μ ε } \{\mu_{\varepsilon}\} { μ ε } is precompact in the weak* topology of measures on Q ′ × Q ′ Q'\times Q' Q ′ × Q ′ . We write
∫ Q ′ ∫ Q ′ ∣ u ε ( x ) − u ε ( y ) ∣ 2 ∣ x − y ∣ n + 1 d x d y = ∫ R n 1 ∣ z ∣ n + 1 ∫ Q ′ ∩ ( Q ′ − z ) ∣ u ε ( x + z ) − u ε ( x ) ∣ 2 d x d z \int_{Q'}\int_{Q'}\frac{|u_{\varepsilon}(x)-u_{\varepsilon}(y)|^{2}}{|x-y|^{n+1}}\,dx\,dy=\int_{\mathbb{R}^{n}}\frac{1}{|z|^{n+1}}\int_{Q'\cap(Q'-z)}|u_{\varepsilon}(x+z)-u_{\varepsilon}(x)|^{2}\,dx\,dz ∫ Q ′ ∫ Q ′ ∣ x − y ∣ n + 1 ∣ u ε ( x ) − u ε ( y ) ∣ 2 d x d y = ∫ R n ∣ z ∣ n + 1 1 ∫ Q ′ ∩ ( Q ′ − z ) ∣ u ε ( x + z ) − u ε ( x ) ∣ 2 d x d z
and split the integral into ∣ z ∣ ≤ ε |z|\le\varepsilon ∣ z ∣ ≤ ε , ε < ∣ z ∣ < 3 4 σ ε \varepsilon<|z|<\frac{3}{4}\sigma_{\varepsilon} ε < ∣ z ∣ < 4 3 σ ε and ∣ z ∣ ≥ 3 4 σ ε |z|\ge\frac{3}{4}\sigma_{\varepsilon} ∣ z ∣ ≥ 4 3 σ ε ; for regime (R1) the middle range is empty. For r > 0 r>0 r > 0 , we define
M r : = N ε + r ∪ ⋃ i { x ∈ Q ~ : dist ( x , ∂ P ε i ) ≤ ε + r } , \mathcal{M}_{r}:=\mathcal{N}_{\varepsilon+r}\cup\bigcup_{i}\left\{x\in\widetilde{Q}:\operatorname{dist}(x,\partial P^{i}_{\varepsilon})\le\varepsilon+r\right\}, M r := N ε + r ∪ i ⋃ { x ∈ Q : dist ( x , ∂ P ε i ) ≤ ε + r } ,
which contains all the points that are either close to the uncovered part of J u J_{u} J u or to the boundary of the transition regions (in regime (R1), M r = N ε + r \mathcal{M}_{r}=\mathcal{N}_{\varepsilon+r} M r = N ε + r ). Then
∣ M r ∣ ≤ C ( ε + r ) , (52) |\mathcal{M}_{r}|\le C(\varepsilon+r), \tag*{(52)} ∣ M r ∣ ≤ C ( ε + r ) , (52)
for some constant C > 0 C>0 C > 0 independent of ε \varepsilon ε and r r r . Moreover, u ε ( x + z ) = u ε ( x ) u_{\varepsilon}(x+z)=u_{\varepsilon}(x) u ε ( x + z ) = u ε ( x ) if x ∉ M ∣ z ∣ x\notin\mathcal{M}_{|z|} x ∈ / M ∣ z ∣ , since the segment [ x , x + z ] [x,x+z] [ x , x + z ] is in the region where u ε u_{\varepsilon} u ε is constant, and u ε ( x + z ) = u ε ( x ) u_{\varepsilon}(x+z)=u_{\varepsilon}(x) u ε ( x + z ) = u ε ( x ) if x ∉ N 4 σ ε + ∣ z ∣ x\notin\mathcal{N}_{4\sigma_{\varepsilon}+|z|} x ∈ / N 4 σ ε + ∣ z ∣ .
∫ ∣ z ∣ ≤ ε d z ∣ z ∣ n + 1 ∫ Q ′ ∩ ( Q ′ − z ) ∣ u ε ( x + z ) − u ε ( x ) ∣ 2 d x ≤ C ∫ ∣ z ∣ ≤ ε d z ∣ z ∣ n + 1 ∣ M ε ∣ ∣ z ∣ 2 ε 2 ≤ C ∫ ∣ z ∣ ≤ ε d z ∣ z ∣ n − 1 ε ≤ C . \int_{|z|\leq\varepsilon}\frac{dz}{|z|^{n+1}}\int_{Q'\cap(Q'-z)}|u_\varepsilon(x+z)-u_\varepsilon(x)|^2\,dx\leq C\int_{|z|\leq\varepsilon}\frac{dz}{|z|^{n+1}}|\mathcal{M}_{\varepsilon}|\frac{|z|^2}{\varepsilon^2}\leq C\int_{|z|\leq\varepsilon}\frac{dz}{|z|^{n-1}\varepsilon}\leq C. ∫ ∣ z ∣ ≤ ε ∣ z ∣ n + 1 d z ∫ Q ′ ∩ ( Q ′ − z ) ∣ u ε ( x + z ) − u ε ( x ) ∣ 2 d x ≤ C ∫ ∣ z ∣ ≤ ε ∣ z ∣ n + 1 d z ∣ M ε ∣ ε 2 ∣ z ∣ 2 ≤ C ∫ ∣ z ∣ ≤ ε ∣ z ∣ n − 1 ε d z ≤ C .
∫ ε < ∣ z ∣ < 3 4 σ ε 1 ∣ z ∣ n + 1 ∫ Q ′ ∩ ( Q ′ − z ) ∣ u ε ( x + z ) − u ε ( x ) ∣ 2 d x d z ≤ 4 M 2 ∫ ε < ∣ z ∣ < 3 4 σ ε ∣ M ∣ z ∣ ∣ ∣ z ∣ n + 1 d z \int_{\varepsilon<|z|<\frac{3}{4}\sigma_\varepsilon}\frac{1}{|z|^{n+1}}\int_{Q'\cap(Q'-z)}|u_\varepsilon(x+z)-u_\varepsilon(x)|^2\,dx\,dz\leq4M^2\int_{\varepsilon<|z|<\frac{3}{4}\sigma_\varepsilon}\frac{|\mathcal{M}_{|z|}|}{|z|^{n+1}}\,dz ∫ ε < ∣ z ∣ < 4 3 σ ε ∣ z ∣ n + 1 1 ∫ Q ′ ∩ ( Q ′ − z ) ∣ u ε ( x + z ) − u ε ( x ) ∣ 2 d x d z ≤ 4 M 2 ∫ ε < ∣ z ∣ < 4 3 σ ε ∣ z ∣ n + 1 ∣ M ∣ z ∣ ∣ d z
≤ 8 C M 2 ∫ ε < ∣ z ∣ < 3 4 σ ε d z ∣ z ∣ n ≤ C ( ∣ log δ ε ∣ + 1 ) . \leq8CM^2\int_{\varepsilon<|z|<\frac{3}{4}\sigma_\varepsilon}\frac{dz}{|z|^n}\leq C(|\log\delta_\varepsilon|+1). ≤ 8 C M 2 ∫ ε < ∣ z ∣ < 4 3 σ ε ∣ z ∣ n d z ≤ C ( ∣ log δ ε ∣ + 1 ) .
For ∣ z ∣ ≥ 3 4 σ ε |z|\geq\frac{3}{4}\sigma_\varepsilon ∣ z ∣ ≥ 4 3 σ ε , we have N 4 σ ε + ∣ z ∣ ⊆ N 18 ∣ z ∣ \mathcal{N}_{4\sigma_\varepsilon+|z|}\subseteq\mathcal{N}_{18|z|} N 4 σ ε + ∣ z ∣ ⊆ N 18∣ z ∣ . Hence, using again that u u u is bounded by M M M , we obtain
∫ 3 4 σ ε ≤ ∣ z ∣ ≤ diam ( Q ′ ) 1 ∣ z ∣ n + 1 ∫ Q ′ ∩ ( Q ′ − z ) ∣ u ε ( x + z ) − u ε ( x ) ∣ 2 d x d z ≤ 4 M 2 ∫ 3 4 σ ε ≤ ∣ z ∣ ≤ diam ( Q ′ ) ∣ N 18 ∣ z ∣ ∣ ∣ z ∣ n + 1 d z ≤ C ( ∣ log σ ε ∣ + 1 ) . \begin{aligned}
\int_{\frac{3}{4}\sigma_\varepsilon\leq|z|\leq\operatorname{diam}(Q')}\frac{1}{|z|^{n+1}}\int_{Q'\cap(Q'-z)}|u_\varepsilon(x+z)-u_\varepsilon(x)|^2\,dx\,dz\leq4M^2\int_{\frac{3}{4}\sigma_\varepsilon\leq|z|\leq\operatorname{diam}(Q')}\frac{|\mathcal{N}_{18|z|}|}{|z|^{n+1}}\,dz \\
&\leq C(|\log\sigma_\varepsilon|+1).
\end{aligned} ∫ 4 3 σ ε ≤ ∣ z ∣ ≤ diam ( Q ′ ) ∣ z ∣ n + 1 1 ∫ Q ′ ∩ ( Q ′ − z ) ∣ u ε ( x + z ) − u ε ( x ) ∣ 2 d x d z ≤ 4 M 2 ∫ 4 3 σ ε ≤ ∣ z ∣ ≤ diam ( Q ′ ) ∣ z ∣ n + 1 ∣ N 18∣ z ∣ ∣ d z ≤ C ( ∣ log σ ε ∣ + 1 ) .
Dividing everything by ∣ log ε ∣ |\log\varepsilon| ∣ log ε ∣ , we obtain
μ ε ( Q ′ × Q ′ ) ≤ C + o ( 1 ) \mu_\varepsilon(Q'\times Q')\leq C+o(1) μ ε ( Q ′ × Q ′ ) ≤ C + o ( 1 )
as ε \varepsilon ε goes to 0. In particular, up to a subsequence μ ε ⇀ ∗ μ \mu_\varepsilon\mathrel{\stackrel{*}{\rightharpoonup}} \mu μ ε ⇀ ∗ μ and, by Lemma 6.2, μ \mu μ is concentrated on the diagonal and we can define λ \lambda λ . Moreover, λ \lambda λ is concentrated on J u ∩ Q ′ J_u\cap Q' J u ∩ Q ′ . Indeed, considering U ρ : = { dist ( x , Σ ) ≤ ρ } U_\rho:=\{ \operatorname{dist}(x,\Sigma)\leq\rho\} U ρ := { dist ( x , Σ ) ≤ ρ } , for r < ρ r<\rho r < ρ we have ∣ N r ∩ U ρ ∣ + ∣ M r ∩ U ρ ∣ ≤ C ( ε + r ) ρ |\mathcal{N}_r\cap U_\rho|+|\mathcal{M}_r\cap U_\rho|\leq C(\varepsilon+r)\rho ∣ N r ∩ U ρ ∣ + ∣ M r ∩ U ρ ∣ ≤ C ( ε + r ) ρ and ∣ N r ∩ U ρ ∣ ≤ C ρ 2 |\mathcal{N}_r\cap U_\rho|\leq C\rho^2 ∣ N r ∩ U ρ ∣ ≤ C ρ 2 for every r r r ; hence repeating the same computation as above, we have
μ ε ( U ρ × Q ′ ) ≤ C ρ + o ( 1 ) \mu_\varepsilon(U_\rho\times Q')\leq C\rho+o(1) μ ε ( U ρ × Q ′ ) ≤ C ρ + o ( 1 )
and then λ ( Σ ) = 0 \lambda(\Sigma)=0 λ ( Σ ) = 0 . Moreover, u ε u_\varepsilon u ε is constant away from J u J_u J u , so λ ( Q ′ ∖ J u ) = 0 \lambda(Q'\setminus J_u)=0 λ ( Q ′ ∖ J u ) = 0 .
Step 4. Density of λ \lambda λ . To conclude the proof, it suffices to show that for H n − 1 \mathcal{H}^{n-1} H n − 1 -a.e. x 0 ∈ J i x_0\in J^i x 0 ∈ J i ,
lim sup r → 0 λ ( Q 2 r ν i ( x 0 ) ) ( 2 r ) n − 1 ≤ ω n − 1 Ψ ( 1 ) ( u − i , u + i ) , \limsup_{r\to0}\frac{\lambda(Q_{2r}^{\nu^i}(x_0))}{(2r)^{n-1}}\leq\omega_{n-1}\Psi^{(1)}(u_-^i,u_+^i), r → 0 lim sup ( 2 r ) n − 1 λ ( Q 2 r ν i ( x 0 )) ≤ ω n − 1 Ψ ( 1 ) ( u − i , u + i ) ,
where
Ψ ( 1 ) ( u − i , u + i ) = { 2 ( u + i − u − i ) 2 , in ( R 1 ) , 2 ( 1 − 2 ℓ w i ( 1 − w i ) ) , in ( R 2 ) , (53) \Psi^{(1)}(u_-^i,u_+^i)=
\begin{cases}
2(u_+^i-u_-^i)^2, & \text{in } \mathrm{(R1)},\\
2(1-2\ell w^i(1-w^i)), & \text{in } \mathrm{(R2)},
\end{cases} \tag*{(53)} Ψ ( 1 ) ( u − i , u + i ) = { 2 ( u + i − u − i ) 2 , 2 ( 1 − 2 ℓ w i ( 1 − w i )) , in ( R1 ) , in ( R2 ) , (53)
where Q 2 r ν i ( x 0 ) Q_{2r}^{\nu^i}(x_0) Q 2 r ν i ( x 0 ) is the cube of side 2 r 2r 2 r centered at x 0 x_0 x 0 oriented in the direction ν i \nu^i ν i .
Up to a rigid motion, we can assume that x 0 = 0 x_0=0 x 0 = 0 , ν i = e 1 \nu^i=e_1 ν i = e 1 and J i = { x 1 = 0 } J^i=\{x_1=0\} J i = { x 1 = 0 } in a small cube Q 2 r = ( − r , r ) n Q_{2r}=(-r,r)^n Q 2 r = ( − r , r ) n . By construction, for small ε \varepsilon ε , u ε ( x ) = g ε ( x 1 ) u_\varepsilon(x)=g_\varepsilon(x_1) u ε ( x ) = g ε ( x 1 ) in Q 2 r Q_{2r} Q 2 r , where g ε g_\varepsilon g ε is the one-dimensional recovery sequence of the previous subsection for the jump from u − i u_-^i u − i to u + i u_+^i u + i (see Section 6.1.1 and Remark 6.1). From this, (53) follows easily by reducing the integral to one dimension and applying the one-dimensional Γ \Gamma Γ -lim sup inequality for g ε g_\varepsilon g ε . For clarity, let us show the steps.
1 ( 2 r ) n − 1 ∣ log ϵ ∣ ∬ Q 2 r × Q 2 r ∣ u ϵ ( x ) − u ϵ ( y ) ∣ 2 ∣ x − y ∣ n + 1 d x d y = 1 ( 2 r ) n − 1 ∣ log ϵ ∣ ∬ Q 2 r × Q 2 r ∣ g ϵ ( x 1 ) − g ϵ ( y 1 ) ∣ 2 ∣ x − y ∣ n + 1 d x d y = 1 ( 2 r ) n − 1 ∣ log ϵ ∣ ∫ − r r ∫ − r r ∫ ( − r , r ) n − 1 ∫ ( − r , r ) n − 1 ∣ g ϵ ( x 1 ) − g ϵ ( y 1 ) ∣ 2 ( ( x 1 − y 1 ) 2 + ∣ x ′ − y ′ ∣ 2 ) n + 1 2 d x ′ d y ′ d x 1 d y 1 . \begin{aligned}
\frac{1}{(2r)^{n-1}|\log\epsilon|}\iint_{Q_{2r}\times Q_{2r}}\frac{|u_\epsilon(x)-u_\epsilon(y)|^2}{|x-y|^{n+1}}\,dx\,dy
&=\frac{1}{(2r)^{n-1}|\log\epsilon|}\iint_{Q_{2r}\times Q_{2r}}\frac{|g_\epsilon(x_1)-g_\epsilon(y_1)|^2}{|x-y|^{n+1}}\,dx\,dy\\
&=\frac{1}{(2r)^{n-1}|\log\epsilon|}\int_{-r}^{r}\int_{-r}^{r}\int_{(-r,r)^{n-1}}\int_{(-r,r)^{n-1}}\frac{|g_\epsilon(x_1)-g_\epsilon(y_1)|^2}{\left((x_1-y_1)^2+|x'-y'|^2\right)^{\frac{n+1}{2}}}\,dx'\,dy'\,dx_1\,dy_1.
\end{aligned} ( 2 r ) n − 1 ∣ log ϵ ∣ 1 ∬ Q 2 r × Q 2 r ∣ x − y ∣ n + 1 ∣ u ϵ ( x ) − u ϵ ( y ) ∣ 2 d x d y = ( 2 r ) n − 1 ∣ log ϵ ∣ 1 ∬ Q 2 r × Q 2 r ∣ x − y ∣ n + 1 ∣ g ϵ ( x 1 ) − g ϵ ( y 1 ) ∣ 2 d x d y = ( 2 r ) n − 1 ∣ log ϵ ∣ 1 ∫ − r r ∫ − r r ∫ ( − r , r ) n − 1 ∫ ( − r , r ) n − 1 ( ( x 1 − y 1 ) 2 + ∣ x ′ − y ′ ∣ 2 ) 2 n + 1 ∣ g ϵ ( x 1 ) − g ϵ ( y 1 ) ∣ 2 d x ′ d y ′ d x 1 d y 1 .
Now, note that, for each a > 0 a>0 a > 0 ,
1 2 r ∫ − r r ∫ − r r ( a + ∣ x i − y i ∣ 2 ) − n + 1 2 d x i d y i = 1 2 r ∫ − 2 r 2 r ( a + z i 2 ) − n + 1 2 ( 2 r − ∣ z i ∣ ) d z i ≤ ∫ − ∞ ∞ ( a + z i 2 ) − n + 1 2 d z i . \begin{aligned}
\frac{1}{2r}\int_{-r}^{r}\int_{-r}^{r}\left(a+|x_i-y_i|^2\right)^{-\frac{n+1}{2}}\,dx_i\,dy_i
=\frac{1}{2r}\int_{-2r}^{2r}\left(a+z_i^2\right)^{-\frac{n+1}{2}}(2r-|z_i|)\,dz_i \\
&\leq\int_{-\infty}^{\infty}\left(a+z_i^2\right)^{-\frac{n+1}{2}}\,dz_i.
\end{aligned} 2 r 1 ∫ − r r ∫ − r r ( a + ∣ x i − y i ∣ 2 ) − 2 n + 1 d x i d y i = 2 r 1 ∫ − 2 r 2 r ( a + z i 2 ) − 2 n + 1 ( 2 r − ∣ z i ∣ ) d z i ≤ ∫ − ∞ ∞ ( a + z i 2 ) − 2 n + 1 d z i .
Using this fact, we can estimate from above:
μ ϵ ( Q 2 r × Q 2 r ) ( 2 r ) n − 1 = 1 ( 2 r ) n − 1 ∣ log ϵ ∣ ∬ Q 2 r × Q 2 r ∣ u ϵ ( x ) − u ϵ ( y ) ∣ 2 ∣ x − y ∣ n + 1 d x d y ≤ 1 ∣ log ϵ ∣ ∫ − r r ∫ − r r ∣ g ϵ ( x 1 ) − g ϵ ( y 1 ) ∣ 2 ( ∫ R n − 1 1 ( ∣ x 1 − y 1 ∣ 2 + ∣ z ∣ 2 ) n + 1 2 d z ) d x 1 d y 1 = ω n − 1 ∣ log ϵ ∣ ∫ − r r ∫ − r r ∣ g ϵ ( x 1 ) − g ϵ ( y 1 ) ∣ 2 ∣ x 1 − y 1 ∣ 2 d x 1 d y 1 . \begin{aligned}
\frac{\mu_\epsilon(Q_{2r}\times Q_{2r})}{(2r)^{n-1}}
&=\frac{1}{(2r)^{n-1}|\log\epsilon|}\iint_{Q_{2r}\times Q_{2r}}\frac{|u_\epsilon(x)-u_\epsilon(y)|^2}{|x-y|^{n+1}}\,dx\,dy\\
&\leq\frac{1}{|\log\epsilon|}\int_{-r}^{r}\int_{-r}^{r}|g_\epsilon(x_1)-g_\epsilon(y_1)|^2\left(\int_{\mathbb{R}^{n-1}}\frac{1}{\left(|x_1-y_1|^2+|z|^2\right)^{\frac{n+1}{2}}}\,dz\right)\,dx_1\,dy_1\\
&=\frac{\omega_{n-1}}{|\log\epsilon|}\int_{-r}^{r}\int_{-r}^{r}\frac{|g_\epsilon(x_1)-g_\epsilon(y_1)|^2}{|x_1-y_1|^2}\,dx_1\,dy_1.
\end{aligned} ( 2 r ) n − 1 μ ϵ ( Q 2 r × Q 2 r ) = ( 2 r ) n − 1 ∣ log ϵ ∣ 1 ∬ Q 2 r × Q 2 r ∣ x − y ∣ n + 1 ∣ u ϵ ( x ) − u ϵ ( y ) ∣ 2 d x d y ≤ ∣ log ϵ ∣ 1 ∫ − r r ∫ − r r ∣ g ϵ ( x 1 ) − g ϵ ( y 1 ) ∣ 2 ( ∫ R n − 1 ( ∣ x 1 − y 1 ∣ 2 + ∣ z ∣ 2 ) 2 n + 1 1 d z ) d x 1 d y 1 = ∣ log ϵ ∣ ω n − 1 ∫ − r r ∫ − r r ∣ x 1 − y 1 ∣ 2 ∣ g ϵ ( x 1 ) − g ϵ ( y 1 ) ∣ 2 d x 1 d y 1 .
Thus, we finally get
λ ( Q 2 r ) ( 2 r ) n − 1 ≤ ω n − 1 lim sup ϵ → 0 1 ∣ log ϵ ∣ ∫ − r r ∫ − r r ∣ g ϵ ( x 1 ) − g ϵ ( y 1 ) ∣ 2 ∣ x 1 − y 1 ∣ 2 d x 1 d y 1 ≤ ω n − 1 Ψ ( 1 ) ( u i − , u i + ) . \frac{\lambda(Q_{2r})}{(2r)^{n-1}}\leq\omega_{n-1}\limsup_{\epsilon\to0}\frac{1}{|\log\epsilon|}\int_{-r}^{r}\int_{-r}^{r}\frac{|g_\epsilon(x_1)-g_\epsilon(y_1)|^2}{|x_1-y_1|^2}\,dx_1\,dy_1\leq\omega_{n-1}\Psi^{(1)}(u_i^-,u_i^+). ( 2 r ) n − 1 λ ( Q 2 r ) ≤ ω n − 1 ϵ → 0 lim sup ∣ log ϵ ∣ 1 ∫ − r r ∫ − r r ∣ x 1 − y 1 ∣ 2 ∣ g ϵ ( x 1 ) − g ϵ ( y 1 ) ∣ 2 d x 1 d y 1 ≤ ω n − 1 Ψ ( 1 ) ( u i − , u i + ) .
where we used the one-dimensional upper bound inequality for { g ϵ } \{g_\epsilon\} { g ϵ } (see § 6.1.1 ). In conclusion, by the Besicovitch differentiation theorem,
\lambda\leq\omega_{n-1}\Psi^{(1)}(u^-,u^+)\mathcal{H}^{n-1}\mathbin{\vrule height 1.4ex depth -0.4ex width 0.07ex\vrule height 0.07ex depth -0.02ex width 0.7ex}J_u.
Now, since Q ‾ × Q ‾ ⋐ Q ′ × Q ′ \overline{Q}\times\overline{Q}\Subset Q'\times Q' Q × Q ⋐ Q ′ × Q ′ , by Step 2 we have
lim sup ϵ → 0 F ϵ ( u ϵ ) ≤ μ ( Q ‾ × Q ‾ ) + lim sup ϵ → 0 1 ϵ ∣ log ϵ ∣ ∫ Q W ϵ ( u ϵ ) d x ≤ ∫ J u ∩ Q ω n − 1 Ψ ( 1 ) ( u − , u + ) d H n − 1 + { K ∣ { u ∈ S } ∣ , in ( R 1 ) , 0 , in ( R 2 ) . \begin{aligned}
\limsup_{\epsilon\to0}F_\epsilon(u_\epsilon)
&\leq\mu(\overline{Q}\times\overline{Q})+\limsup_{\epsilon\to0}\frac{1}{\epsilon|\log\epsilon|}\int_Q W_\epsilon(u_\epsilon)\,dx\\
&\leq\int_{J_u\cap Q}\omega_{n-1}\Psi^{(1)}(u^-,u^+)\,d\mathcal{H}^{n-1}
+\begin{cases}
K|\{u\in S\}|,&\text{in }(R1),\\
0,&\text{in }(R2).
\end{cases}
\end{aligned} ϵ → 0 lim sup F ϵ ( u ϵ ) ≤ μ ( Q × Q ) + ϵ → 0 lim sup ϵ ∣ log ϵ ∣ 1 ∫ Q W ϵ ( u ϵ ) d x ≤ ∫ J u ∩ Q ω n − 1 Ψ ( 1 ) ( u − , u + ) d H n − 1 + { K ∣ { u ∈ S } ∣ , 0 , in ( R 1 ) , in ( R 2 ) .
where we used that H n − 1 ( J u ∩ ∂ Q ) = 0 \mathcal{H}^{n-1}(J_u\cap\partial Q)=0 H n − 1 ( J u ∩ ∂ Q ) = 0 . This is exactly F ( u ) F(u) F ( u ) , since ω n − 1 Ψ ( 1 ) \omega_{n-1}\Psi^{(1)} ω n − 1 Ψ ( 1 ) is the jump cost for s = 1 2 s=\frac{1}{2} s = 2 1 .
Upper bound for s > 1 2 s>\frac{1}{2} s > 2 1
Proof of Theorem 2.7 (s > 1 2 s>\frac{1}{2} s > 2 1 ). Here θ ϵ = ϵ \theta_\epsilon=\epsilon θ ϵ = ϵ . We set
T ϵ : = min { ϵ − 1 / 4 , δ ϵ − 1 / 2 } and η ϵ : = ϵ T ϵ . (54) T_\epsilon:=\min\left\{\epsilon^{-1/4},\delta_\epsilon^{-1/2}\right\}\qquad\text{and}\qquad\eta_\epsilon:=\epsilon T_\epsilon. \tag*{(54)} T ϵ := min { ϵ − 1/4 , δ ϵ − 1/2 } and η ϵ := ϵ T ϵ . (54)
so that T ε → + ∞ T_\varepsilon\to+\infty T ε → + ∞ and η ε → 0 \eta_\varepsilon\to0 η ε → 0 for ε → 0 \varepsilon\to0 ε → 0 . Let v i T ε v_i^{T_\varepsilon} v i T ε be a (non-decreasing) minimizer of the optimal profile problem Θ s , n T ε ( u i − , u i + ) \Theta_{s,n}^{T_\varepsilon}(u_i^-,u_i^+) Θ s , n T ε ( u i − , u i + ) as defined in (34) , with C s , n C_{s,n} C s , n in front of the seminorm. We recall that v i T ε v_i^{T_\varepsilon} v i T ε is equal to u i − u_i^- u i − in ( − ∞ , − T ε ) (-\infty,-T_\varepsilon) ( − ∞ , − T ε ) and to u i + u_i^+ u i + in ( T ε , + ∞ ) (T_\varepsilon,+\infty) ( T ε , + ∞ ) and, by Proposition 4.1, Θ s , n T ε ( u i − , u i + ) → Θ s , n ( u i − , u i + ) \Theta_{s,n}^{T_\varepsilon}(u_i^-,u_i^+) \to\Theta_{s,n}(u_i^-,u_i^+) Θ s , n T ε ( u i − , u i + ) → Θ s , n ( u i − , u i + ) for ε → 0 \varepsilon\to0 ε → 0 .
Case n ≥ 2 n \ge2 n ≥ 2
Step 1: Construction of the recovery sequence. For each simplex J i J^i J i , let d i ( x ) d^i(x) d i ( x ) be the signed distance from the hyperplane containing J i J^i J i , positive in the direction ν i \nu^i ν i . We define disjoint tubular neighborhoods of width η ε \eta_\varepsilon η ε :
I ε i : = { x ∈ Q ~ : ∣ d i ( x ) ∣ ≤ 2 η ε and dist ( x , J j ) > 4 η ε , ∀ j ≠ i } , J ~ ε i : = { y ∈ J i : y + t ν i ∈ I ε i , ∀ ∣ t ∣ ≤ 2 η ε } , I_\varepsilon^i := \{x \in\widetilde{Q} : |d^i(x)| \le2\eta_\varepsilon\text{ and } \operatorname{dist}(x,J^j) > 4\eta_\varepsilon,\ \forall j \ne i\}, \qquad\widetilde{J}_\varepsilon^i := \{y \in J^i : y+t\nu^i \in I_\varepsilon^i,\ \forall|t| \le2\eta_\varepsilon\}, I ε i := { x ∈ Q : ∣ d i ( x ) ∣ ≤ 2 η ε and dist ( x , J j ) > 4 η ε , ∀ j = i } , J ε i := { y ∈ J i : y + t ν i ∈ I ε i , ∀∣ t ∣ ≤ 2 η ε } ,
P ε i : = { y + t ν i : y ∈ J ~ ε i , ∣ t ∣ ≤ η ε } , P_\varepsilon^i := \{y+t\nu^i : y \in\widetilde{J}_\varepsilon^i,\ |t| \le\eta_\varepsilon\}, P ε i := { y + t ν i : y ∈ J ε i , ∣ t ∣ ≤ η ε } ,
so that H n − 1 ( J ~ ε i ) → H n − 1 ( J i ) \mathcal{H}^{n-1}(\widetilde{J}_\varepsilon^i) \to\mathcal{H}^{n-1}(J^i) H n − 1 ( J ε i ) → H n − 1 ( J i ) and the sets P ε i + B η ε ( 0 ) P_\varepsilon^i+B_{\eta_\varepsilon}(0) P ε i + B η ε ( 0 ) are pairwise disjoint. We set
u ~ ε ( x ) : = { v i T ε ( d i ( x ) / ε ) if x ∈ P ε i , u ( x ) otherwise in Q ~ . \widetilde{u}_\varepsilon(x) :=
\begin{cases}
v_i^{T_\varepsilon}(d^i(x)/\varepsilon) & \text{if } x \in P_\varepsilon^i,\\
u(x) & \text{otherwise in } \widetilde{Q}.
\end{cases} u ε ( x ) := { v i T ε ( d i ( x ) / ε ) u ( x ) if x ∈ P ε i , otherwise in Q .
The function u ~ ε \widetilde{u}_\varepsilon u ε still has some jumps on J u ∖ ⋃ i J ~ ε i J_u \setminus\bigcup_i \widetilde{J}_\varepsilon^i J u ∖ ⋃ i J ε i and across the lateral boundary of P ε i P_\varepsilon^i P ε i . Now, notice that there exists L ≥ 3 L \ge3 L ≥ 3 , only depending on u u u , such that the jump set of u ~ ε \widetilde{u}_\varepsilon u ε in Q ′ Q' Q ′ is contained in { dist ( ⋅ , Σ ) < ( L − 2 ) η ε } \{\operatorname{dist}(\cdot,\Sigma) < (L-2)\eta_\varepsilon\} { dist ( ⋅ , Σ ) < ( L − 2 ) η ε } . Then, to smooth the function near Σ \Sigma Σ , we introduce a cut-off function ψ ε ∈ C ∞ ( Q ~ ; [ 0 , 1 ] ) \psi_\varepsilon\in C^\infty(\widetilde{Q};[0,1]) ψ ε ∈ C ∞ ( Q ; [ 0 , 1 ]) such that ψ ε = 1 \psi_\varepsilon=1 ψ ε = 1 in { dist ( ⋅ , Σ ) ≤ ( L − 1 ) η ε } \{\operatorname{dist}(\cdot,\Sigma) \le(L-1)\eta_\varepsilon\} { dist ( ⋅ , Σ ) ≤ ( L − 1 ) η ε } and ψ ε = 0 \psi_\varepsilon=0 ψ ε = 0 in { dist ( ⋅ , Σ ) ≥ L η ε } \{\operatorname{dist}(\cdot,\Sigma) \ge L\eta_\varepsilon\} { dist ( ⋅ , Σ ) ≥ L η ε } , with ∣ ∇ ψ ε ∣ ≤ C / η ε |\nabla\psi_\varepsilon| \le C/\eta_\varepsilon ∣∇ ψ ε ∣ ≤ C / η ε . The recovery sequence is
u ε : = ψ ε ( u ~ ε ∗ φ η ε ) + ( 1 − ψ ε ) u ~ ε , u_\varepsilon:= \psi_\varepsilon(\widetilde{u}_\varepsilon* \varphi_{\eta_\varepsilon}) + (1-\psi_\varepsilon)\widetilde{u}_\varepsilon, u ε := ψ ε ( u ε ∗ φ η ε ) + ( 1 − ψ ε ) u ε ,
where φ η ε \varphi_{\eta_\varepsilon} φ η ε is a standard mollifier supported in B η ε ( 0 ) B_{\eta_\varepsilon}(0) B η ε ( 0 ) . With this choice, u ε u_\varepsilon u ε has no jumps and { u ε ≠ u } ⊂ N L η ε \{u_\varepsilon\ne u\} \subset N_{L\eta_\varepsilon} { u ε = u } ⊂ N L η ε , so u ε → u u_\varepsilon\to u u ε → u in L 1 ( Q ) L^1(Q) L 1 ( Q ) . This construction provides a smooth transition in a neighborhood of Σ \Sigma Σ , while preserving the optimal profile elsewhere.
In the following, let M : = ∥ u ∥ ∞ M := \|u\|_\infty M := ∥ u ∥ ∞ and U : = { dist ( ⋅ , Σ ) < ( L + 1 ) η ε } U := \{\operatorname{dist}(\cdot,\Sigma) < (L+1)\eta_\varepsilon\} U := { dist ( ⋅ , Σ ) < ( L + 1 ) η ε } , so that ∣ U ∣ ≤ C η ε 2 |U| \le C\eta_\varepsilon^2 ∣ U ∣ ≤ C η ε 2 by (50) .
Step 2: Estimate of the seminorm. Observe that if ∣ z ∣ ≥ η ε |z| \ge\eta_\varepsilon ∣ z ∣ ≥ η ε and x ∉ N ( L + 1 ) ∣ z ∣ x \notin N_{(L+1)|z|} x ∈ / N ( L + 1 ) ∣ z ∣ , then u ε ( x + z ) = u ε ( x ) u_\varepsilon(x+z)=u_\varepsilon(x) u ε ( x + z ) = u ε ( x ) , since u ε = u u_\varepsilon=u u ε = u outside N L η ε N_{L\eta_\varepsilon} N L η ε . Hence, by (50) , the long-range interaction vanishes: we have
ε 2 s − 1 ∫ η ε ≤ ∣ z ∣ ≤ diam ( Q ′ ) d z ∣ z ∣ n + 2 s ∫ Q ∩ ( Q − z ) ∣ u ε ( x + z ) − u ε ( x ) ∣ 2 d x ≤ ε 2 s − 1 C ∫ η ε ≤ ∣ z ∣ ≤ diam ( Q ′ ) d z ∣ z ∣ n + 2 s − 1 = C T ε 1 − 2 s → 0. \begin{aligned}
\varepsilon^{2s-1}\int_{\eta_\varepsilon\le|z|\le\operatorname{diam}(Q')} \frac{dz}{|z|^{n+2s}}\int_{Q\cap(Q-z)} |u_\varepsilon(x+z)-u_\varepsilon(x)|^2\,dx \\
&\le\varepsilon^{2s-1}C\int_{\eta_\varepsilon\le|z|\le\operatorname{diam}(Q')} \frac{dz}{|z|^{n+2s-1}} = CT_\varepsilon^{1-2s} \to0.
\end{aligned} ε 2 s − 1 ∫ η ε ≤ ∣ z ∣ ≤ diam ( Q ′ ) ∣ z ∣ n + 2 s d z ∫ Q ∩ ( Q − z ) ∣ u ε ( x + z ) − u ε ( x ) ∣ 2 d x ≤ ε 2 s − 1 C ∫ η ε ≤ ∣ z ∣ ≤ diam ( Q ′ ) ∣ z ∣ n + 2 s − 1 d z = C T ε 1 − 2 s → 0.
For the short-range part (∣ x − y ∣ ≤ η ε |x-y| \le\eta_\varepsilon ∣ x − y ∣ ≤ η ε ), we distinguish two cases: when x , y ∉ U x,y \notin U x , y ∈ / U and when x ∈ U x \in U x ∈ U . Let x , y ∉ U x,y \notin U x , y ∈ / U with ∣ x − y ∣ ≤ η ε |x-y| \le\eta_\varepsilon ∣ x − y ∣ ≤ η ε . If x ∉ N 2 η ε x \notin N_{2\eta_\varepsilon} x ∈ / N 2 η ε , the segment [ x , y ] [x,y] [ x , y ] does not meet the layers, so u ε ( x ) = u ε ( y ) u_\varepsilon(x)=u_\varepsilon(y) u ε ( x ) = u ε ( y ) . Otherwise, if x ∈ N 2 η ε x \in N_{2\eta_\varepsilon} x ∈ N 2 η ε , there exists i i i such that x , y ∈ C i : = { p + t ν i : p ∈ J ~ ε i , ∣ t ∣ ≤ 3 η ε } x,y \in C_i := \{p+t\nu^i : p \in\widetilde{J}_\varepsilon^i,\ |t| \le3\eta_\varepsilon\} x , y ∈ C i := { p + t ν i : p ∈ J ε i , ∣ t ∣ ≤ 3 η ε } . In this set, u ε = v i T ε ( d i / ε ) u_\varepsilon=v_i^{T_\varepsilon}(d^i/\varepsilon) u ε = v i T ε ( d i / ε ) . Hence, we have
ε 2 s − 1 ∬ ( Q ∖ U ) 2 ∩ { ∣ x − y ∣ ≤ η ε } ∣ u ε ( x ) − u ε ( y ) ∣ 2 ∣ x − y ∣ n + 2 s d x d y ≤ ε 2 s − 1 ∑ j ∫ J j ∫ − 3 η ε 3 η ε ∫ J j ∫ − 3 η ε 3 η ε ∣ v j T ε ( z ε ) − v j T ε ( w ε ) ∣ 2 ( ∣ σ − τ ∣ 2 + ( z − w ) 2 ) n + 2 s 2 d σ d z d τ d w ≤ ∑ j H n − 1 ( J j ) ∫ R ∫ R ∣ v j T ε ( ξ ) − v j T ε ( ζ ) ∣ 2 ( ∫ R n − 1 d θ ( ∣ θ ∣ 2 + ( ξ − ζ ) 2 ) n + 2 s 2 ) d ξ d ζ , (55) \begin{aligned}
\varepsilon^{2s-1}\iint_{(Q\setminus U)^2\cap\{\lvert x-y\rvert\leq\eta_\varepsilon\}}\frac{\lvert u_\varepsilon(x)-u_\varepsilon(y)\rvert^2}{\lvert x-y\rvert^{n+2s}}\,dx\,dy
\leq\varepsilon^{2s-1}\sum_j\int_{J^j}\int_{-3\eta_\varepsilon}^{3\eta_\varepsilon}\int_{J^j}\int_{-3\eta_\varepsilon}^{3\eta_\varepsilon}\frac{\left\lvert v_j^{T_\varepsilon}\left(\frac{z}{\varepsilon}\right)-v_j^{T_\varepsilon}\left(\frac{w}{\varepsilon}\right)\right\rvert^2}{\left(\lvert\sigma-\tau\rvert^2+(z-w)^2\right)^{\frac{n+2s}{2}}}\,d\sigma\,dz\,d\tau\,dw \\
&\leq\sum_j\mathcal{H}^{n-1}(J^j)\int_{\mathbb{R}}\int_{\mathbb{R}}\left\lvert v_j^{T_\varepsilon}(\xi)-v_j^{T_\varepsilon}(\zeta)\right\rvert^2\left(\int_{\mathbb{R}^{n-1}}\frac{d\theta}{\left(\lvert\theta\rvert^2+(\xi-\zeta)^2\right)^{\frac{n+2s}{2}}}\right)d\xi\,d\zeta, \tag*{(55)}
\end{aligned} ε 2 s − 1 ∬ ( Q ∖ U ) 2 ∩ {∣ x − y ∣ ≤ η ε } ∣ x − y ∣ n + 2 s ∣ u ε ( x ) − u ε ( y ) ∣ 2 d x d y ≤ ε 2 s − 1 j ∑ ∫ J j ∫ − 3 η ε 3 η ε ∫ J j ∫ − 3 η ε 3 η ε ( ∣ σ − τ ∣ 2 + ( z − w ) 2 ) 2 n + 2 s v j T ε ( ε z ) − v j T ε ( ε w ) 2 d σ d z d τ d w ≤ j ∑ H n − 1 ( J j ) ∫ R ∫ R v j T ε ( ξ ) − v j T ε ( ζ ) 2 ( ∫ R n − 1 ( ∣ θ ∣ 2 + ( ξ − ζ ) 2 ) 2 n + 2 s d θ ) d ξ d ζ , (55)
where we used the change of variables
ξ = z ε , ζ = w ε , θ = σ − τ ε . \xi=\frac{z}{\varepsilon},\qquad\zeta=\frac{w}{\varepsilon},\qquad\theta=\frac{\sigma-\tau}{\varepsilon}. ξ = ε z , ζ = ε w , θ = ε σ − τ .
Since x , y ∈ Q x,y\in Q x , y ∈ Q , in (55) the variables σ , τ \sigma,\tau σ , τ range only in J j ∩ ( Q + B 3 η ε ) J^j\cap(Q+B_{3\eta_\varepsilon}) J j ∩ ( Q + B 3 η ε ) ; therefore, in the following, H n − 1 ( J j ) \mathcal{H}^{n-1}(J^j) H n − 1 ( J j ) stands for H n − 1 ( J j ∩ ( Q + B 3 η ε ) ) \mathcal{H}^{n-1}(J^j\cap(Q+B_{3\eta_\varepsilon})) H n − 1 ( J j ∩ ( Q + B 3 η ε )) , which tends to H n − 1 ( J j ∩ Q ‾ ) = H n − 1 ( J j ∩ Q ) \mathcal{H}^{n-1}(J^j\cap\overline{Q})=\mathcal{H}^{n-1}(J^j\cap Q) H n − 1 ( J j ∩ Q ) = H n − 1 ( J j ∩ Q ) as ε → 0 \varepsilon\to0 ε → 0 . Now, notice that
∫ R n − 1 1 ( ∣ θ ∣ 2 + A 2 ) n + 2 s 2 d θ = 1 A 1 + 2 s ∫ R n − 1 1 ( ∣ ω ∣ 2 + 1 ) n + 2 s 2 d ω = C s , n A 1 + 2 s . \int_{\mathbb{R}^{n-1}}\frac{1}{\left(\lvert\theta\rvert^2+A^2\right)^{\frac{n+2s}{2}}}\,d\theta=\frac{1}{A^{1+2s}}\int_{\mathbb{R}^{n-1}}\frac{1}{\left(\lvert\omega\rvert^2+1\right)^{\frac{n+2s}{2}}}\,d\omega=\frac{C_{s,n}}{A^{1+2s}}. ∫ R n − 1 ( ∣ θ ∣ 2 + A 2 ) 2 n + 2 s 1 d θ = A 1 + 2 s 1 ∫ R n − 1 ( ∣ ω ∣ 2 + 1 ) 2 n + 2 s 1 d ω = A 1 + 2 s C s , n .
Then,
ε 2 s − 1 ∬ ( Q ∖ U ) 2 ∩ { ∣ x − y ∣ ≤ η ε } ∣ u ε ( x ) − u ε ( y ) ∣ 2 ∣ x − y ∣ n + 2 s d x d y ≤ ∑ j H n − 1 ( J j ) C s , n [ v j T ε ] H s ( R ) 2 . \varepsilon^{2s-1}\iint_{(Q\setminus U)^2\cap\{\lvert x-y\rvert\leq\eta_\varepsilon\}}\frac{\lvert u_\varepsilon(x)-u_\varepsilon(y)\rvert^2}{\lvert x-y\rvert^{n+2s}}\,dx\,dy\leq\sum_j\mathcal{H}^{n-1}(J^j)C_{s,n}[v_j^{T_\varepsilon}]_{H^s(\mathbb{R})}^2. ε 2 s − 1 ∬ ( Q ∖ U ) 2 ∩ {∣ x − y ∣ ≤ η ε } ∣ x − y ∣ n + 2 s ∣ u ε ( x ) − u ε ( y ) ∣ 2 d x d y ≤ j ∑ H n − 1 ( J j ) C s , n [ v j T ε ] H s ( R ) 2 .
Now, let x ∈ U x\in U x ∈ U with ∣ x − y ∣ ≤ η ε \lvert x-y\rvert\leq\eta_\varepsilon ∣ x − y ∣ ≤ η ε . Let h ε : = u ~ ε ∗ φ η ε h_\varepsilon:=\widetilde{u}_\varepsilon*\varphi_{\eta_\varepsilon} h ε := u ε ∗ φ η ε , which is C / η ε C/\eta_\varepsilon C / η ε -Lipschitz. By definition, we have
u ε ( x ) − u ε ( y ) = ψ ε ( x ) ( h ε ( x ) − h ε ( y ) ) + ( ψ ε ( x ) − ψ ε ( y ) ) ( h ε ( y ) − u ~ ε ( y ) ) + ( 1 − ψ ε ( x ) ) ( u ~ ε ( x ) − u ~ ε ( y ) ) . u_\varepsilon(x)-u_\varepsilon(y)=\psi_\varepsilon(x)(h_\varepsilon(x)-h_\varepsilon(y))+(\psi_\varepsilon(x)-\psi_\varepsilon(y))(h_\varepsilon(y)-\widetilde{u}_\varepsilon(y))+(1-\psi_\varepsilon(x))(\widetilde{u}_\varepsilon(x)-\widetilde{u}_\varepsilon(y)). u ε ( x ) − u ε ( y ) = ψ ε ( x ) ( h ε ( x ) − h ε ( y )) + ( ψ ε ( x ) − ψ ε ( y )) ( h ε ( y ) − u ε ( y )) + ( 1 − ψ ε ( x )) ( u ε ( x ) − u ε ( y )) .
The first two terms are bounded by C ∣ x − y ∣ / η ε C\lvert x-y\rvert/\eta_\varepsilon C ∣ x − y ∣ / η ε , hence their contribution vanishes; indeed
C ε 2 s − 1 ∣ U ∣ η ε 2 ∫ ∣ z ∣ ≤ η ε d z ∣ z ∣ n + 2 s − 2 ≤ C ε T ε 2 − 2 s → 0. C\frac{\varepsilon^{2s-1}\lvert U\rvert}{\eta_\varepsilon^2}\int_{\lvert z\rvert\leq\eta_\varepsilon}\frac{dz}{\lvert z\rvert^{n+2s-2}}\leq C\varepsilon T_\varepsilon^{2-2s}\to0. C η ε 2 ε 2 s − 1 ∣ U ∣ ∫ ∣ z ∣ ≤ η ε ∣ z ∣ n + 2 s − 2 d z ≤ C ε T ε 2 − 2 s → 0.
The third term vanishes in { dist ( x , Σ ) ≤ ( L − 1 ) η ε } \{\operatorname{dist}(x,\Sigma)\leq(L-1)\eta_\varepsilon\} { dist ( x , Σ ) ≤ ( L − 1 ) η ε } ; otherwise the segment [ x , y ] [x,y] [ x , y ] does not meet the jumps of u ~ ε \widetilde{u}_\varepsilon u ε . Hence we can repeat the same computation as in (55) and conclude that the energy contribution is bounded from above by
C H n − 1 ( J u ∩ { dist ( x , Σ ) < ( L + 2 ) η ε } ) ≤ C η ε → 0. C\mathcal{H}^{n-1}\left(J_u\cap\{\operatorname{dist}(x,\Sigma)<(L+2)\eta_\varepsilon\}\right)\leq C\eta_\varepsilon\to0. C H n − 1 ( J u ∩ { dist ( x , Σ ) < ( L + 2 ) η ε } ) ≤ C η ε → 0.
In conclusion,
ε 2 s − 1 [ u ε ] H s ( Q ) 2 ≤ ∑ j = 1 m H n − 1 ( J j ) C s , n [ v j T ε ] H s ( R ) 2 + o ( 1 ) . \varepsilon^{2s-1}[u_\varepsilon]_{H^s(Q)}^2\leq\sum_{j=1}^{m}\mathcal{H}^{n-1}(J^j)C_{s,n}[v_j^{T_\varepsilon}]_{H^s(\mathbb{R})}^2+o(1). ε 2 s − 1 [ u ε ] H s ( Q ) 2 ≤ j = 1 ∑ m H n − 1 ( J j ) C s , n [ v j T ε ] H s ( R ) 2 + o ( 1 ) .
Step 3: Estimate of the potential term. By the upper estimate (13) , W ε ≤ dist 2 ( ⋅ , Z ∪ S ) + δ ε χ A 1 W_\varepsilon\leq\operatorname{dist}^2(\cdot,Z\cup S)+\delta_\varepsilon\chi_{A_1} W ε ≤ dist 2 ( ⋅ , Z ∪ S ) + δ ε χ A 1 . Moreover, outside U U U we use the change of variables t = d i ( x ) / ε t=d^i(x)/\varepsilon t = d i ( x ) / ε and we obtain
1 ε ∫ Q W ε ( u ε ) d x ≤ ∑ i = 1 m H n − 1 ( J i ) ∫ − T ε T ε W S ( v i T ε ( t ) ) d t + δ ε ε ∣ A 1 ( u ε ) ∣ + C ∣ U ∣ ε . \frac{1}{\varepsilon}\int_Q W_\varepsilon(u_\varepsilon)\,dx\leq\sum_{i=1}^{m}\mathcal{H}^{n-1}(J^i)\int_{-T_\varepsilon}^{T_\varepsilon}W^S(v_i^{T_\varepsilon}(t))\,dt+\frac{\delta_\varepsilon}{\varepsilon}\lvert A_1(u_\varepsilon)\rvert+\frac{C\lvert U\rvert}{\varepsilon}. ε 1 ∫ Q W ε ( u ε ) d x ≤ i = 1 ∑ m H n − 1 ( J i ) ∫ − T ε T ε W S ( v i T ε ( t )) d t + ε δ ε ∣ A 1 ( u ε )∣ + ε C ∣ U ∣ .
where C ∣ U ∣ / ε ≤ C η ε 2 / ε ≤ C ε 1 / 2 → 0 C|U|/\varepsilon\le C\eta_\varepsilon^2/\varepsilon\le C\varepsilon^{1/2}\to0 C ∣ U ∣/ ε ≤ C η ε 2 / ε ≤ C ε 1/2 → 0 . In regime (R1), the second term converges to K ∣ { u ∈ S } ∣ K|\{u\in S\}| K ∣ { u ∈ S } ∣ . In regime (R2), u u u takes values in Z \mathbb{Z} Z , so ∣ A 1 ( u ε ) ∣ ≤ ∣ N L η ε ∣ ≤ C η ε |A_1(u_\varepsilon)|\le|\mathcal{N}_{L\eta_\varepsilon}|\le C\eta_\varepsilon ∣ A 1 ( u ε ) ∣ ≤ ∣ N L η ε ∣ ≤ C η ε , and the second term is bounded by C δ ε 1 / 2 → 0 C\delta_\varepsilon^{1/2}\to0 C δ ε 1/2 → 0 .
Step 4: Conclusion. Combining Step 2 and Step 3, and since
C s , n [ v i T ε ] H s ( R ) 2 + ∫ R dist 2 ( v i T ε , Z ∪ S ) = Θ s , n ( u i − , u i + ) , C_{s,n}[v_i^{T_\varepsilon}]_{H^s(\mathbb{R})}^2+\int_{\mathbb{R}}\operatorname{dist}^2(v_i^{T_\varepsilon},\mathbb{Z}\cup S)=\Theta_{s,n}(u_i^-,u_i^+), C s , n [ v i T ε ] H s ( R ) 2 + ∫ R dist 2 ( v i T ε , Z ∪ S ) = Θ s , n ( u i − , u i + ) ,
we have
lim sup ε → 0 F ε ( u ε ) ≤ ∑ j = 1 m H n − 1 ( J j ) Θ s , n ( u j − , u j + ) + { K ∣ { u ∈ S } ∣ in (R1) , 0 in (R2) , \limsup_{\varepsilon\to0}F_\varepsilon(u_\varepsilon)\le\sum_{j=1}^{m}\mathcal{H}^{n-1}(J^j)\Theta_{s,n}(u_j^-,u_j^+)+\begin{cases} K|\{u\in S\}| & \text{in (R1)},\\ 0 & \text{in (R2)}, \end{cases} ε → 0 lim sup F ε ( u ε ) ≤ j = 1 ∑ m H n − 1 ( J j ) Θ s , n ( u j − , u j + ) + { K ∣ { u ∈ S } ∣ 0 in (R1) , in (R2) ,
which is exactly F ( u ) F(u) F ( u ) , since, by Proposition 4.1 , Θ s , n ( u i − , u i + ) = λ s , n κ s Φ Z ∪ S ( u i − , u i + ) \Theta_{s,n}(u_i^-,u_i^+)=\lambda_{s,n}\kappa_s\Phi_{\mathbb{Z}\cup S}(u_i^-,u_i^+) Θ s , n ( u i − , u i + ) = λ s , n κ s Φ Z ∪ S ( u i − , u i + ) .
Case n = 1 n=1 n = 1 .
Now J u = { x 1 , … , x m } J_u=\{x_1,\ldots,x_m\} J u = { x 1 , … , x m } and ν i ∈ { ± 1 } \nu^i\in\{\pm1\} ν i ∈ { ± 1 } is the orientation of the jump. We define the recovery sequence as
u ε ( x ) : = { v i T ε ( ν i ( x − x i ) ε ) , if x ∈ [ x i − η ε , x i + η ε ] = : P ε i , u ( x ) , otherwise in Q , u_\varepsilon(x):=\begin{cases}v_i^{T_\varepsilon}\left(\dfrac{\nu^i(x-x_i)}{\varepsilon}\right), & \text{if }x\in[x_i-\eta_\varepsilon,x_i+\eta_\varepsilon]=:P_\varepsilon^i,\\ u(x), & \text{otherwise in }Q,\end{cases} u ε ( x ) := ⎩ ⎨ ⎧ v i T ε ( ε ν i ( x − x i ) ) , u ( x ) , if x ∈ [ x i − η ε , x i + η ε ] =: P ε i , otherwise in Q ,
where η ε , T ε \eta_\varepsilon,T_\varepsilon η ε , T ε are defined in (54) . For ε \varepsilon ε sufficiently small, the intervals [ x i − η ε , x i + η ε ] [x_i-\eta_\varepsilon,x_i+\eta_\varepsilon] [ x i − η ε , x i + η ε ] are mutually disjoint. By the change of variables ξ = ( x − x i ) / ε \xi=(x-x_i)/\varepsilon ξ = ( x − x i ) / ε and ζ = ( y − x i ) / ε \zeta=(y-x_i)/\varepsilon ζ = ( y − x i ) / ε , we have:
ε 2 s − 1 ∬ P ε i × P ε i ∣ u ε ( x ) − u ε ( y ) ∣ 2 ∣ x − y ∣ 1 + 2 s d x d y = ∫ − T ε T ε ∫ − T ε T ε ∣ v i T ε ( ξ ) − v i T ε ( ζ ) ∣ 2 ∣ ξ − ζ ∣ 1 + 2 s d ξ d ζ . \varepsilon^{2s-1}\iint_{P_\varepsilon^i\times P_\varepsilon^i}\frac{|u_\varepsilon(x)-u_\varepsilon(y)|^2}{|x-y|^{1+2s}}\,dx\,dy=\int_{-T_\varepsilon}^{T_\varepsilon}\int_{-T_\varepsilon}^{T_\varepsilon}\frac{|v_i^{T_\varepsilon}(\xi)-v_i^{T_\varepsilon}(\zeta)|^2}{|\xi-\zeta|^{1+2s}}\,d\xi\,d\zeta. ε 2 s − 1 ∬ P ε i × P ε i ∣ x − y ∣ 1 + 2 s ∣ u ε ( x ) − u ε ( y ) ∣ 2 d x d y = ∫ − T ε T ε ∫ − T ε T ε ∣ ξ − ζ ∣ 1 + 2 s ∣ v i T ε ( ξ ) − v i T ε ( ζ ) ∣ 2 d ξ d ζ .
Similarly, the potential term scales as
1 ε ∫ P ε i W S ( u ε ( x ) ) d x = ∫ − T ε T ε W S ( v i T ε ( ξ ) ) d ξ . \frac{1}{\varepsilon}\int_{P_\varepsilon^i}W^S(u_\varepsilon(x))\,dx=\int_{-T_\varepsilon}^{T_\varepsilon}W^S(v_i^{T_\varepsilon}(\xi))\,d\xi. ε 1 ∫ P ε i W S ( u ε ( x )) d x = ∫ − T ε T ε W S ( v i T ε ( ξ )) d ξ .
Since the interactions between different jumps contribute to the energy as o ( 1 ) o(1) o ( 1 ) , we have
lim sup ε → 0 F ε ( u ε ) ≤ ∑ i = 1 m Θ s , 1 ( u i − , u i + ) + { K ∣ { u ∈ S } ∣ in (R1) , 0 in (R2) . \limsup_{\varepsilon\to0}F_\varepsilon(u_\varepsilon)\le\sum_{i=1}^{m}\Theta_{s,1}(u_i^-,u_i^+)+\begin{cases}K|\{u\in S\}| & \text{in (R1)},\\ 0 & \text{in (R2)}.\end{cases} ε → 0 lim sup F ε ( u ε ) ≤ i = 1 ∑ m Θ s , 1 ( u i − , u i + ) + { K ∣ { u ∈ S } ∣ 0 in (R1) , in (R2) .
This concludes the proof.
Limiting behavior for s < 1 2 s<\frac{1}{2} s < 2 1
In the low-order regime s < 1 2 s<\frac{1}{2} s < 2 1 , the characteristic feature is that bounded B V BV B V functions belong to H s ( Q ) H^s(Q) H s ( Q ) with [ u ] H s ( Q ) 2 ≤ C ∥ u ∥ ∞ ∣ D u ∣ ( Q ) [u]_{H^s(Q)}^2\le C\|u\|_\infty|Du|(Q) [ u ] H s ( Q ) 2 ≤ C ∥ u ∥ ∞ ∣ D u ∣ ( Q ) .
Assuming the scaling parameter θ ε = ε 2 s \theta_\varepsilon=\varepsilon^{2s} θ ε = ε 2 s as in [26 ] , the functional (3) simplifies to
F ε ( u ; Q ) = ∬ Q × Q ∣ u ( x ) − u ( y ) ∣ 2 ∣ x − y ∣ n + 2 s d x d y + 1 ε 2 s ∫ Q W ε ( u ( x ) ) d x . (56) F_\varepsilon(u;Q)=\iint_{Q\times Q}\frac{|u(x)-u(y)|^2}{|x-y|^{n+2s}}\,dx\,dy+\frac{1}{\varepsilon^{2s}}\int_QW_\varepsilon(u(x))\,dx. \tag*{(56)} F ε ( u ; Q ) = ∬ Q × Q ∣ x − y ∣ n + 2 s ∣ u ( x ) − u ( y ) ∣ 2 d x d y + ε 2 s 1 ∫ Q W ε ( u ( x )) d x . (56)
Also in this case, the limiting behavior is determined by the value of K : = lim ε → 0 δ ε / θ ε ∈ [ 0 , + ∞ ] K:=\lim_{\varepsilon\to0}\delta_\varepsilon/\theta_\varepsilon\in[0,+\infty] K := lim ε → 0 δ ε / θ ε ∈ [ 0 , + ∞ ] . Unlike the cases s ≥ 1 2 s\ge\frac{1}{2} s ≥ 2 1 , the limit functional for s < 1 2 s<\frac{1}{2} s < 2 1 is non-local and does not concentrate on the jump set J u J_u J u . The scaling θ ε = ε 2 s \theta_\varepsilon=\varepsilon^{2s} θ ε = ε 2 s is the natural one for this regime since it preserves the fractional seminorm, while the potential term acts as a penalization for intermediate values, either allowing them (if K < ∞ K<\infty K < ∞ ) or forbidding them (if K = ∞ K=\infty K = ∞ ).
For a closed set V ⊂ R \mathcal{V} \subset\mathbb{R} V ⊂ R , we set
H s ( Q ; V ) : = { u ∈ L 1 ( Q ) : [ u ] H s ( Q ) < + ∞ , u ( x ) ∈ V for a.e. x ∈ Q } . H^{s}(Q;\mathcal{V}) := \{u \in L^{1}(Q): [u]_{H^{s}(Q)} < +\infty, u(x) \in\mathcal{V}\text{ for a.e. }x \in Q\}. H s ( Q ; V ) := { u ∈ L 1 ( Q ) : [ u ] H s ( Q ) < + ∞ , u ( x ) ∈ V for a.e. x ∈ Q } .
Theorem 7.1 (Compactness and Γ \Gamma Γ -convergence for s < 1 2 s < \frac{1}{2} s < 2 1 ). Let s ∈ ( 0 , 1 2 ) s \in(0,\frac{1}{2}) s ∈ ( 0 , 2 1 ) and consider the sequence of functionals { F ε } \{F_{\varepsilon}\} { F ε } defined in (56) .
(i) Compactness: Let { u ε } ⊂ L 1 ( Q ) \{u_{\varepsilon}\} \subset L^{1}(Q) { u ε } ⊂ L 1 ( Q ) be a sequence such that sup ε F ε ( u ε ) ≤ M < + ∞ \sup_{\varepsilon} F_{\varepsilon}(u_{\varepsilon}) \le M < +\infty sup ε F ε ( u ε ) ≤ M < + ∞ . Then there exists a sequence of integers { a ε } ⊂ Z \{a_{\varepsilon}\} \subset\mathbb{Z} { a ε } ⊂ Z such that the translated sequence { u ε − a ε } \{u_{\varepsilon}-a_{\varepsilon}\} { u ε − a ε } is relatively compact in L 1 ( Q ) L^{1}(Q) L 1 ( Q ) . If, in addition, u ε → u u_{\varepsilon} \to u u ε → u in L 1 ( Q ) L^{1}(Q) L 1 ( Q ) then u ∈ H s ( Q ; Z ∪ S ) u \in H^{s}(Q;\mathbb{Z}\cup S) u ∈ H s ( Q ; Z ∪ S ) . Moreover, if δ ε ε 2 s → + ∞ \frac{\delta_{\varepsilon}}{\varepsilon^{2s}} \to+\infty ε 2 s δ ε → + ∞ , then u ∈ H s ( Q ; Z ) u \in H^{s}(Q;\mathbb{Z}) u ∈ H s ( Q ; Z ) .
(ii) Γ \Gamma Γ -convergence:
If δ ε ε 2 s → K ∈ [ 0 , + ∞ ) \frac{\delta_{\varepsilon}}{\varepsilon^{2s}} \to K \in[0,+\infty) ε 2 s δ ε → K ∈ [ 0 , + ∞ ) , then F ε F_{\varepsilon} F ε Γ \Gamma Γ -converges in L 1 ( Q ) L^{1}(Q) L 1 ( Q ) to:
G s u p ( u ) = { [ u ] H s ( Q ) 2 + K ∣ { x ∈ Q : u ( x ) ∈ S } ∣ if u ∈ H s ( Q ; Z ∪ S ) , + ∞ otherwise. G_{\mathrm{sup}}(u) =
\begin{cases}
[u]_{H^{s}(Q)}^{2} + K|\{x \in Q:u(x) \in S\}| & \text{if }u \in H^{s}(Q;\mathbb{Z}\cup S),\\
+\infty& \text{otherwise.}
\end{cases} G sup ( u ) = { [ u ] H s ( Q ) 2 + K ∣ { x ∈ Q : u ( x ) ∈ S } ∣ + ∞ if u ∈ H s ( Q ; Z ∪ S ) , otherwise. G s u b ( u ) = { [ u ] H s ( Q ) 2 if u ∈ H s ( Q ; Z ) , + ∞ otherwise. G_{\mathrm{sub}}(u) =
\begin{cases}
[u]_{H^{s}(Q)}^{2} & \text{if }u \in H^{s}(Q;\mathbb{Z}),\\
+\infty& \text{otherwise.}
\end{cases} G sub ( u ) = { [ u ] H s ( Q ) 2 + ∞ if u ∈ H s ( Q ; Z ) , otherwise.
Proof. (i) To prove compactness, we observe that the uniform bound on the H s ( Q ) H^{s}(Q) H s ( Q ) seminorm leads to a uniform bound in H s ( Q ) H^{s}(Q) H s ( Q ) for u ε u_{\varepsilon} u ε minus its mean, due to the fractional Poincaré inequality. Then by the compact embedding of H s ( Q ) H^{s}(Q) H s ( Q ) into L 2 ( Q ) L^{2}(Q) L 2 ( Q ) , the thesis follows. More precisely, from the uniform energy bound sup ε > 0 F ε ( u ε ) ≤ M \sup_{\varepsilon>0} F_{\varepsilon}(u_{\varepsilon}) \le M sup ε > 0 F ε ( u ε ) ≤ M , we immediately obtain the estimates
[ u ε ] H s ( Q ) 2 ≤ M and ∫ Q dist 2 ( u ε , Z ∪ S ) d x ≤ ∫ Q W ε ( u ε ) d x ≤ M ε 2 s ⟶ 0. (57) [u_{\varepsilon}]_{H^{s}(Q)}^{2} \le M \qquad\text{and}\qquad\int_{Q}\operatorname{dist}^{2}(u_{\varepsilon},\mathbb{Z}\cup S)\,dx \le\int_{Q}W_{\varepsilon}(u_{\varepsilon})\,dx \le M\varepsilon^{2s} \longrightarrow0. \tag*{(57)} [ u ε ] H s ( Q ) 2 ≤ M and ∫ Q dist 2 ( u ε , Z ∪ S ) d x ≤ ∫ Q W ε ( u ε ) d x ≤ M ε 2 s ⟶ 0. (57)
By the fractional Poincaré–Wirtinger inequality on the bounded connected domain Q Q Q (see [21 Theorem 6.33] ), we have
\|u_{\varepsilon}-\bar{u}_{\varepsilon}\|_{L^{2}(Q)} \le C[u_{\varepsilon}]_{H^{s}(Q)} \le C\sqrt{M}, \qquad\text{where }\bar{u}_{\varepsilon}:=\fint_{Q}u_{\varepsilon}.
Since [ u ε − u ˉ ε ] H s ( Q ) = [ u ε ] H s ( Q ) [u_{\varepsilon}-\bar{u}_{\varepsilon}]_{H^{s}(Q)}=[u_{\varepsilon}]_{H^{s}(Q)} [ u ε − u ˉ ε ] H s ( Q ) = [ u ε ] H s ( Q ) , the sequence { u ε − u ˉ ε } \{u_{\varepsilon}-\bar{u}_{\varepsilon}\} { u ε − u ˉ ε } is uniformly bounded in H s ( Q ) H^{s}(Q) H s ( Q ) . Since the embedding H s ( Q ) ↪ L 2 ( Q ) H^{s}(Q)\hookrightarrow L^{2}(Q) H s ( Q ) ↪ L 2 ( Q ) is compact (see [13 Theorem 7.1] ), up to a subsequence,
u ε − u ˉ ε ⟶ v strongly in L 2 ( Q ) . u_{\varepsilon}-\bar{u}_{\varepsilon} \longrightarrow v \qquad\text{strongly in }L^{2}(Q). u ε − u ˉ ε ⟶ v strongly in L 2 ( Q ) .
Now, let a ε ∈ Z a_{\varepsilon}\in\mathbb{Z} a ε ∈ Z be an integer closest to u ˉ ε \bar{u}_{\varepsilon} u ˉ ε , so that u ˉ ε − a ε ∈ [ − 1 2 , 1 2 ] \bar{u}_{\varepsilon}-a_{\varepsilon}\in[-\frac{1}{2},\frac{1}{2}] u ˉ ε − a ε ∈ [ − 2 1 , 2 1 ] . Then, up to a subsequence,
u ˉ ε − a ε ⟶ c , for some c ∈ [ − 1 2 , 1 2 ] . \bar{u}_{\varepsilon}-a_{\varepsilon} \longrightarrow c,\qquad\text{for some }c\in[-\frac{1}{2},\frac{1}{2}]. u ˉ ε − a ε ⟶ c , for some c ∈ [ − 2 1 , 2 1 ] .
Consequently,
u ε − a ε ⟶ v + c strongly in L 2 ( Q ) . u_{\varepsilon}-a_{\varepsilon} \longrightarrow v+c \qquad\text{strongly in }L^{2}(Q). u ε − a ε ⟶ v + c strongly in L 2 ( Q ) .
Therefore, the sequence { u ε − a ε } \{u_{\varepsilon}-a_{\varepsilon}\} { u ε − a ε } is relatively compact in L 1 ( Q ) L^{1}(Q) L 1 ( Q ) . If u ε → u u_{\varepsilon}\to u u ε → u , then by (57) for θ ε = ε 2 s \theta_{\varepsilon}=\varepsilon^{2s} θ ε = ε 2 s , we have
u ( x ) ∈ Z ∪ S for a.e. x ∈ Q . u(x)\in\mathbb{Z}\cup S\qquad\text{for a.e. }x\in Q. u ( x ) ∈ Z ∪ S for a.e. x ∈ Q .
Finally, by the lower semicontinuity of the fractional seminorm,
[ u ] H s ( Q ) ≤ lim inf ε → 0 [ u ε ] H s ( Q ) < + ∞ . [u]_{H^{s}(Q)}\le\liminf_{\varepsilon\to0}[u_{\varepsilon}]_{H^{s}(Q)}<+\infty. [ u ] H s ( Q ) ≤ ε → 0 lim inf [ u ε ] H s ( Q ) < + ∞.
Moreover, by (13) , we have that if δ ε / ε 2 s → + ∞ \delta_{\varepsilon}/\varepsilon^{2s} \to+\infty δ ε / ε 2 s → + ∞ then ∣ { u ∈ S } ∣ = 0 \lvert\{u \in S\}\rvert= 0 ∣{ u ∈ S }∣ = 0 by the uniform energy bound. This proves the desired compactness result.
Regarding the Γ \Gamma Γ -convergence, the Γ \Gamma Γ -liminf is a direct consequence of (13) and the lower semicontinuity of the H s H^s H s -seminorm with respect to L 1 L^1 L 1 -convergence; if δ ε / ε 2 s → K < + ∞ \delta_{\varepsilon}/\varepsilon^{2s} \to K < +\infty δ ε / ε 2 s → K < + ∞ , the bulk term follows by convergence in measure, i.e. lim inf ε ∣ A δ ( u ε ) ∣ ≥ ∣ { u ∈ S } ∣ \liminf_{\varepsilon} \lvert A_{\delta}(u_{\varepsilon})\rvert\geq\lvert\{u \in S\}\rvert lim inf ε ∣ A δ ( u ε )∣ ≥ ∣{ u ∈ S }∣ with A δ A_{\delta} A δ the one of (13) . For the Γ \Gamma Γ -limsup, given u ∈ H s ( Q ; Z ∪ S ) u \in H^s(Q;\mathbb{Z}\cup S) u ∈ H s ( Q ; Z ∪ S ) , the constant sequence u ε = u u_{\varepsilon}=u u ε = u is a recovery sequence. □
Use of AI
NO AI tools were used to produce mathematical content; the strategy and the proofs were developed and written exclusively by the authors. Claude Opus 5.5 (Anthropic) was used only for minor language editing and to detect typographical and LaTeX errors on the very last version of the manuscript.
Acknowledgments
A.G. and M.A.P. are members of the Gruppo Nazionale per l’Analisi Matematica, la Probabilità e le loro Applicazioni (GNAMPA) of the Istituto Nazionale di Alta Matematica (INdAM).
References
Adriana Garroni
Dipartimento di Matematica, Università La Sapienza di Roma, Piazzale Aldo Moro 5, 00185 Roma, ITALY
Email address: garroni@mat.uniroma1.it
Maria Antonietta Palladino
Classe di Scienze, Scuola Normale Superiore, Piazza dei Cavalieri 7, I-56126 Pisa, ITALY
Email address: mariaantonietta.palladino@sns.it