# Constructive approximation on graded meshes for the integral fractional Laplacian

We consider the homogeneous Dirichlet problem for the integral fractional Laplacian. We prove optimal Sobolev regularity estimates in Lipschitz domains satisfying an exterior ball condition. We present the construction of graded bisection meshes by a greedy algorithm and derive quasi-optimal convergence rates for approximations to the solution of such a problem by continuous piecewise linear functions. The nonlinear Sobolev scale dictates the relation between regularity and approximability.

## Authors

• 10 publications
• 12 publications
10/06/2021

### Besov regularity for the Dirichlet integral fractional Laplacian in Lipschitz domains

We prove Besov regularity estimates for the solution of the Dirichlet pr...
05/07/2020

### Local energy estimates for the fractional Laplacian

The integral fractional Laplacian of order s ∈ (0,1) is a nonlocal opera...
04/08/2022

### Exponential convergence of hp-FEM for the integral fractional Laplacian in 1D

We prove weighted analytic regularity for the solution of the integral f...
12/15/2021

### Weighted analytic regularity for the integral fractional Laplacian in polygons

We prove weighted analytic regularity of solutions to the Dirichlet prob...
03/23/2021

### Robust BPX preconditioner for the integral fractional Laplacian on bounded domains

We propose and analyze a robust BPX preconditioner for the integral frac...
09/10/2020

### On explicit form of the FEM stiffness matrix for the integral fractional Laplacian on non-uniform meshes

We derive exact form of the piecewise-linear finite element stiffness ma...
05/28/2020

### Local convergence of the FEM for the integral fractional Laplacian

We provide for first order discretizations of the integral fractional La...
##### This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

## 1. introduction

We consider the integral fractional Laplacian on a bounded domain ,

 (1.1) (−Δ)su(x)=Λ(d,s) p.v.∫Rdu(x)−u(y)|x−y|d+2s,Λ(d,s)=22ssΓ(s+d2)πd/2Γ(1−s).

and corresponding homogeneous Dirichlet problem

 (1.2) {(−Δ)su=f in Ω,u=0 in Ωc=Rd∖¯¯¯¯Ω,

where We assume that is a bounded Lipschitz domain satisfying an exterior ball condition and a bounded function. We are interested in analyzing the performance of a greedy algorithm for approximating solutions to (1.2) by continuous piecewise linear functions over graded bisection meshes.

Regardless of the regularity of the domain and the right-hand side , solutions to (1.2) generically develop an algebraic singular layer of the form (cf. e.g. [16])

 (1.3) u(x)≈dist(x,∂Ω)s,

that limits the smoothness of solutions. Heuristically, let us consider that

is the half-line . Thus, we can interpret the behavior (1.3) as , and wonder under what conditions this function belongs to a Sobolev space with differentiability index and integrability index . For that purpose, let us compute (Riemann-Liouville) derivatives of order of the function , :

 ∂rv(x)≃xs−r+,t>0.

We observe that is -integrable near if and only if , namely, if . This heuristic discussion illustrates the natural interplay between the differentiability order and integrability index for membership of solutions to (1.2) in the class , at least for dimension . It turns out that the restriction is needed irrespective of dimension (cf Theorem 3.7 below).

For the sake of approximation, one can find the optimal choice of the indexes by inspecting a DeVore diagram; see Figure 1.1 for an illustration in the two-dimensional setting. Recall the definition of Sobolev number and the Sobolev line corresponding to the nonlinear approximation scale of ,

 {(1p,r):Sob\,(Wrp)=Sob\,(Hs)}={(1p,r):r=s−1+2p}.

In order to have a compact embedding , we require or equivalently to lie above this line. In addition, the regularity restriction , derived heuristically earlier, is depicted in red and intersects the Sobolev line at .

Letting and with arbitrarily small, we have

 Sob\,(Ws+1−ε1+ε)=s−1+ε(1−ε1+ε)>Sob\,(Hs),

while the condition is also satisfied. This yields an optimal choice of parameters in dimension .

One can perform an analogous argument for arbitrary dimension : the optimal approximation space can be found on a DeVore diagram by intersecting the Sobolev line corresponding to with the regularity line . For these two lines are parallel, while for these lines meet at the point with

 p=2(d−1)d,r=s+d2(d−1).

This indicates that the optimal regularity one can expect corresponds to the differentiability order for .

In this paper we justify this heuristic argument rigorously and exploit it to construct suitable graded bisection meshes via a greedy algorithm that delivers quasi-optimal convergence rates for continuous piecewise linear approximation. In Section 2, we introduce some notation regarding fractional-order Sobolev spaces and the weak formulation of (1.2). Section 3 is devoted to providing a rigorous proof of the regularity estimates discussed above. In Section 4 we study the performance of the greedy algorithm. Finally, Section 5 includes some numerical experiments for illustrating our theoretical findings: we observe optimal convergence rates and that the singular boundary layer (1.3) dominates reentrant corner singularities.

## 2. Fractional Sobolev Spaces

In this section we set the notation and review some properties of the spaces involved in the rest of the paper. We start by recalling some function spaces.

Given and , we consider the seminorm

 (2.1) |v|Wσp(Ω):=(Λ(σ,d,p)∬Ω×Ω|v(x)−v(y)|p|x−y|d+σpdxdy)1p.

Above, we set the constant in such a way that in the limits and one recovers the standard integer-order norms. More precisely, by the results in [8] and [19], we require

 (2.2) Λ(σ,d,p)≃σpΓ(d2)4πd/2 for σ≃0,Λ(σ,d,p)≃(1−σ)p∫Sd−1|ω⋅e1|pdω for σ≃1.

We see that these constants vanish linearly in as and as . For we set the constant as in the definition (1.1) of the fractional Laplacian , which is consistent with these requirements.

We adopt the convention that zero-order derivatives correspond to the identity, and write , and . Given , let be the largest integer number smaller or equal than , , and we define

 Wtp(Ω):={v∈Wk,p(Ω):|Dαv|∈Wσp(Ω) ∀α with |α|=k},

with the norm

 ∥v∥Wtp(Ω):=∥v∥Wkp(Ω)+|v|Wtp(Ω),|v|Wtp(Ω)=∑|α|=k|Dαv|Wσp(Ω).

The Sobolev number of is defined to be .

For our purposes, we need to consider zero-extension spaces as well. For , we denote by its extension by zero on . If and , we define to be the space of functions whose trivial extensions are globally in ,

 ˜Wtp(Ω)={v∈Wkp(Ω):˜v∈Wtp(Rd)}.

These spaces characterize the regularity of functions across . It is important to realize that if , , then the spaces and are identical; equivalently, any function in can be extended by zero without changing its regularity (cf. [15, Corollary 1.4.4.5]). In contrast, if , then the notion of trace is well defined in and its subspace of functions with vanishing trace coincides with . Finally, the case is exceptional and corresponds to the so-called Lions-Magenes space [18, Theorem 1.11.7].

From now on, for any given function we will drop the tilde to denote its zero extension, and assume that the domain of is . An important feature of these zero-extension spaces is the following Poincaré inequality: if for and , then . Therefore,

 ∥v∥˜Wtp(Ω):=|v|Wtp(Rd)

defines a norm equivalent to in .

As usual, we denote Sobolev spaces with integrability index by using the letter instead of . Hence , and we define and its duality pairing. For , the weak formulation of (1.2) reads: find such that

 (2.3) a(u,v):=Λ(d,s)2∬Rd×Rd(u(x)−u(y))(v(x)−v(y))|x−y|d+2sdxdy=⟨f,v⟩

for all . Existence and uniqueness of weak solutions in , as well as the stability of the solution map , are straightforward consequences of the Lax-Milgram theorem. We point out that, in the left-hand side of (2.3), the integration region is effectively .

## 3. Regularity of solutions

The purpose of this section is to provide regularity estimates for solutions of (1.2) in terms of fractional Sobolev norms with arbitrary integrability index . In a similar fashion to [2], our starting point shall be the precise weighted Hölder estimates derived by X. Ros-Oton and J. Serra [24]. In [2] these estimates were employed to obtain regularity estimates in weighted Sobolev spaces with differentiability , where the weight is a power of the distance to the boundary of . As we show below, such regularity estimates are optimally suited for the case . Here, we derive optimal regularity estimates for any . Our technique consists in recasting the estimates from [24] in unweighted Sobolev spaces with differentiability index but at the expense of an integrability index .

### 3.1. Hölder regularity.

We start with two important assumptions. We first assume satisfies an exterior ball condition. In [24] this assumption allows the construction of suitable barriers that lead to -regularity of solutions up to the boundary of , or equivalently, exploiting the zero extension [24, Proposition 1.1]

 (3.1) ∥u∥Cs(Rd)≤C∥f∥L∞(Ω).

Secondly, we shall assume possesses certain Hölder continuity. This assumption allows us to derive higher-order regularity estimates on the solution.

We will employ the letter to denote either the distance functions

 δ(x)=dist(x,∂Ω), x∈Ω, δ(x,y)=min{δ(x),δ(y)}, x,y∈Ω.

For , we denote by the -seminorm. If , let us set with integer and . We consider the seminorm

 |v|(θ)β=supx,y∈Ωδ(x,y)β+θ|Dkv(x)−Dkv(y)||x−y|β′,

and the associated norm in the following way: for ,

 ∥v∥(θ)β=k∑ℓ=0(supx∈Ωδ(x)ℓ+θ|Dℓv(x)|)+|v|(θ)β,

while for ,

 ∥v∥(θ)β=∥v∥C−θ(Ω)+k∑ℓ=1(supx∈Ωδ(x)ℓ+θ|Dℓv(x)|)+|v|(θ)β.

The following estimate [24, Proposition 1.4] is essential in what follows. It hinges on (3.1) but does not make use of the exterior ball condition directly.

###### Theorem 3.1 (weighted Hölder regularity).

Let be a bounded Lipschitz domain and be such that neither nor is an integer. Let be such that , and be a solution of (1.2). Then, and

 ∥u∥(−s)β+2s≤C(Ω,s,β)(∥u∥Cs(Rd)+∥f∥(s)β).

We next recast this estimate depending , with the exceptional case . According to (3.1) and the definition of , we have

 ∥u∥Cs(Rd)+∥f∥(s)β≤C∥f∥Cβ(¯¯¯Ω).
###### Corollary 3.2 (pointwise weighted bounds).

Let be a bounded Lipschitz domain that satisfies the exterior ball condition and . If , then

1. Case : we have and

 (3.2) supx,y∈Ωδ(x,y)β+s|u(x)−u(y)||x−y|β+2s≤C∥f∥Cβ(¯¯¯Ω).
2. Case and : we have

 (3.3) supx∈Ωδ(x)1/2|Du(x)|≤C∥f∥Cβ(¯¯¯Ω).
3. Case : we have

 (3.4) supx∈Ωδ(x)1−s|Du(x)|+supx,y∈Ωδ(x,y)β+s|Du(x)−Du(y)||x−y|β+2s−1≤C∥f∥Cβ(¯¯¯Ω).

The following interior Hölder estimate [24, Lemma 2.9] will also be used later.

###### Lemma 3.3 (interior regularity).

If and , then verifies

 (3.5) |u|Cγ(¯¯¯¯¯¯¯¯¯¯¯¯¯BR(x))≤CRs−γ∥f∥L∞(Ω) ∀x∈Ω,

where R = and the constant C depends only on and , and blows up only when .

### 3.2. Sobolev regularity.

Our goal for the remainder of this section is to use the Hölder estimates we have reviewed to derive bounds on Sobolev norms of . We first show that under suitable assumptions on the right-hand side , the first-order derivatives of the solution are -integrable. For such a purpose, we resort to the following result [8], which utilizes the asymptotic behavior as of the scaling factor in the definition (2.1) of the seminorm .

###### Proposition 3.4 (limits of fractional seminorms).

Assume , . Then, it holds that

 (3.6) limε→0|v|W1−εp(Ω)=|v|W1p(Ω).
###### Remark 3.5 (integrability of powers of the distance function to the boundary).

On Lipschitz domains, powers of the distance function to the boundary have the following integrability property: for every it holds that (cf. for example [9, Lemma 2.14])

 (3.7) ∫Ωδ(x)−αdx=O(11−α).
###### Theorem 3.6 (W1p-regularity).

Let be a bounded Lipschitz domain satisfying an exterior ball condition, , such that , and satisfy the following regularity assumptions:

 ⎧⎪ ⎪⎨⎪ ⎪⎩f∈C1−2s(¯¯¯¯Ω)if s<1/2,f∈Cβ(¯¯¯¯Ω)if s=1/2 (for some β>0),f∈L∞(Ω)if s>1/2.

Then, the solution to (1.2) satisfies , which coincides with .

###### Proof.

We shall prove that, for every sufficiently small, and

 (3.8) |u|pW1−εp(Rd)≤C(Ω,d,s,p,f).

From (3.6) and the fact that is compactly supported, (3.8) implies that . Since on it follows that and has a well defined and vanishing trace because . To exploit symmetry of the integrand in the definition of we decompose the domain of integration into

and its complement within and realize that

 ∬Ω×Ω|u(x)−u(y)|p|x−y|d+(1−ε)pdydx=2∬D|u(x)−u(y)|p|x−y|d+(1−ε)pdydx.

Similarly, for the rest of the domain of integration we have

 ∬[Ω×Ωc]∪[Ωc×Ω]|u(x)−u(y)|p|x−y|d+(1−ε)pdydx=2∬Ω×Ωc|u(x)−u(y)|p|x−y|d+(1−ε)pdydx

We further split the effective domain of integration into the sets

 A:={(x,y)∈D:|x−y|<δ(x)2},B:=[D∖A]∪[Ω×Ωc],

and rewrite the seminorm defined in (2.1) as

 |u|pW1−εp(Rd)Λ(1−ε,d,p)=2∬A|u(x)−u(y)|p|x−y|d+(1−ε)pdydx+2∬B|u(x)−u(y)|p|x−y|d+(1−ε)pdydx,

where as according to (2.2). We finally fix if and if , and estimate the contributions on and .

On the set , we use the Hölder estimate (3.1) and integration in polar coordinates together with (3.7) with to obtain

 ∬B|u(x)−u(y)|p|x−y|d+(1−ε)pdydx ≤C∥f∥pL∞(Ω)∫Ω∫B(x,δ(x)/2)c|x−y|−d−(1−ε)p+spdydx ≤Cp(1−ε−s)∥f∥pL∞(Ω)∫Ωδ(x)p(−1+ε+s)dx ≤Cp(1−ε−s)(1−p(1−ε−s))∥f∥pL∞(Ω).

To deal with the set A, we first assume that , note that yields , and employ (3.3) together with (3.7) with to get

 ∫Ω|Du(x)|pdx≤C∥f∥Cβ(¯¯¯Ω)∫Ωδ(x)−p/2dx≤C2−p∥f∥Cβ(¯¯¯Ω).

For and distinguish two cases. In the case , we resort to (3.5)

 |u(x)−u(y)||x−y|γ≤Cδ(x)s−γ∥f∥L∞(Ω)∀|x−y|≤δ(x)2

and to write

 ∬Aδ(x)p(s−γ)|x−y|d+(1−ε)pdydx ≤C∫Ωδ(x)p(s−γ)∫B(x,δ(x)/2)|x−y|−d−(1−ε−γ)pdydx ≤Cp(γ−1+ε)∫Ωδ(x)p(−1+ε+s)dx,

provided . This implies

 ∬A|u(x)−u(y)|p|x−y|d+(1−ε)pdydx≤Cp(γ−1+ε)(1−p(1−ε−s))∥f∥pL∞(Ω),

and simply setting yields

 ∬A|u(x)−u(y)|p|x−y|d+(1−ε)pdydx≤Cε(1−p(1−ε−s))∥f∥pL∞(Ω).

For , we resort to (3.2) with , namely

 |u(x)−u(y)||x−y|β+2s≤Cδ(x)−(β−s)∥f∥C1−2s(¯¯¯Ω)

because . We next integrate in polar coordinates to get

 ∬A |u(x)−u(y)|p|x−y|d+(1−ε)pdydx ≤C∥f∥pC1−2s(¯¯¯Ω)∫Ωδ(x)−p(β+s)∫B(x,δ(x)/2)|x−y|−d−(1−ε−β−2s)pdydx ≤Cp(β+2s−1+ε)∥f∥pC1−2s(¯¯¯Ω)∫Ωδ(x)p(s−1+ε)dx ≤Cp(β+2s−1+ε)(1−p(1−ε−s))∥f∥pC1−2s(¯¯¯Ω).

Thus, letting we obtain

 ∬A|u(x)−u(y)|p|x−y|d+(1−ε)pdydx≤Cε(1−p(1−ε−s))∥f∥pC1−2s(¯¯¯Ω).

Collecting the estimates above and recalling the asymptotic behavior as given in (2.2), we deduce the desired expression (3.8) for . This concludes the proof. ∎

We now aim to prove higher-order Sobolev regularity estimates.

###### Theorem 3.7 (Sobolev regularity).

Let be a bounded Lipschitz domain satisfying an exterior ball condition, , for some . Furthermore, let and . Then, the solution of (2.3) belongs to , with

 (3.9)

where the constant is robust with respect to and .

###### Proof.

Our hypotheses imply that . Therefore, we distinguish between two cases: either or . We shall focus on the case because the case can be dealt with the same arguments, but performed over the function instead of its gradient.

Since , it turns out that , whence . Moreover, we have and consequently we can apply Theorem 3.6 (-regularity) to deduce . This concludes the proof in the case ; if we next aim to bound . Similarly to the proof of Theorem 3.6, we split the domain of integration into the sets

 A={(x,y)∈D:|x−y|<δ(x)=δ(x,y)},B=[D∖A]∪[Ω×Ωc],

and write

 ∥u∥p˜Wrp(Ω)Λ(r−1,d,p)=2∬A|Du(x)−Du(y)|p|x−y|d+(r−1)pdydx+2∬B|Du(x)−Du(y)|p|x−y|d+(r−1)pdydx.

We now proceed as in the case of Theorem 3.6. On the set , we exploit the bound on given in (3.4) to obtain

 ∬A|Du(x)−Du(y)|p|x−y|d+(r−1)pdydx≤C∥f∥pCβ(¯¯¯Ω)∫Ωδ(x)−p(β+s)∫B(x,δ(x))|x−y|−d+p(β+2s−r)dydx≤Cp(β+2s−r)∥f∥pCβ(¯¯¯Ω)∫Ωδ(x)−p(r−s)dx.

because the assumption ensures the convergence of the integral on . In view of (3.7) with , the integral in the right hand side above is convergent and is of order , whence

 ∬A|Du(x)−Du(y)|p|x−y|d+(r−1)pdydx≤Cp(β+2s−r)(1−p(r−s))∥f∥pCβ(¯¯¯Ω).

On the set , we utilize the pointwise bound given in (3.4) to write

 |Du(x)−Du(y)|≤|Du(x)|+|Du(y)|≤Cδ(x)s−1∥f∥Cβ(¯¯¯Ω)

because either , whence , or and . Consequently, since ,

 ∬B |Du(x)−Du(y)|p|x−y|d+(r−1)pdydx ≤C∥f∥pCβ(¯¯¯Ω)∫Ωδ(x)−p(1−s)∫B(x,δ(x))c|x−y|−d−(r−1)pdydx ≤C∥f∥pCβ(¯¯¯Ω)p(r−1)∫Ωδ(x)−p(r−s)dx≤Cp(r−1)(1−p(r−s))∥f∥pCβ(¯¯¯Ω).

Collecting the estimates on and , we infer that

and the behavior of the ratio is robust with respect to and , according to (2.2). This implies that and (3.9) is valid. ∎

For the purposes of approximation, we aim to take as large as possible. On the one hand, we have the limitation from the hypotheses of Theorem 3.7 (Sobolev regularity). On the other hand, one requires in order to have . These two straight lines meet at , whence we deduce the extreme differentiability and integrability indices

 (3.10) r=s+1p,p=2(d−1)d.

This is in agreement with the estimates in weighted spaces from [2, 6]. Let us now specify admissible choices of differentiability parameter and integrability parameter so that is as close to and as close to as possible.

###### Corollary 3.8 (optimal regularity for d≥2).

Let and satisfy . If , then any yields