CONFLUENTES MATHEMATICI
Pierre BOUSQUET, Augusto C. PONCE, and Jean VAN SCHAFTINGEN Density of smooth maps for fractional Sobolev spacesWs,pinto`simply connected manifolds whens>1
Tome 5, no2 (2013), p. 3-22.
<http://cml.cedram.org/item?id=CML_2013__5_2_3_0>
© Les auteurs et Confluentes Mathematici, 2013.
Tous droits réservés.
L’accès aux articles de la revue « Confluentes Mathematici » (http://cml.cedram.org/), implique l’accord avec les condi- tions générales d’utilisation (http://cml.cedram.org/legal/).
Toute reproduction en tout ou partie de cet article sous quelque forme que ce soit pour tout usage autre que l’utilisation á fin strictement personnelle du copiste est constitutive d’une infrac- tion pénale. Toute copie ou impression de ce fichier doit contenir la présente mention de copyright.
cedram
Article mis en ligne dans le cadre du
Centre de diffusion des revues académiques de mathématiques http://www.cedram.org/
Confluentes Math.
5, 2 (2013) 3-22
DENSITY OF SMOOTH MAPS FOR FRACTIONAL SOBOLEV SPACES Ws,p INTO ` SIMPLY CONNECTED
MANIFOLDS WHEN s>1
PIERRE BOUSQUET, AUGUSTO C. PONCE, AND JEAN VAN SCHAFTINGEN
Abstract. Given a compact manifoldNn⊂Rνand real numberss>1 and 16p <∞, we prove that the classC∞(Qm;Nn) of smooth maps on the cube with values into Nn is strongly dense in the fractional Sobolev space Ws,p(Qm;Nn) whenNn isbspcsimply connected. Forspinteger, we prove weak sequential density ofC∞(Qm;Nn) whenNn is sp−1 simply connected. The proofs are based on the existence of a retraction ofRν onto Nnexcept for a small subset ofNnand on a pointwise estimate of fractional derivatives of composition of maps inWs,p∩W1,sp.
1. Introduction
In this paper we discuss results and open questions related to the density of smooth maps in Sobolev spaces with values into a manifold. For this purpose, let Nn be a compact manifold of dimension nimbedded in the Euclidean space Rν. For anys >0 and 16p <+∞, we define the class of Sobolev maps defined on the unitmdimensional cubeQmwith values intoNn,
Ws,p(Qm;Nn) =
u∈Ws,p(Qm;Rν) :u∈Nn a.e. .
Whens =k is an integer,Ws,p(Qm;Rν) is the standard Sobolev space equipped with the norm
kukWs,p(Qm)=kukLp(Qm)+
k
X
j=1
kDjukLp(Qm).
Whens is not an integer,s =k+σwith k ∈Nand 0 < σ <1. In this case, by u∈Ws,p(Qm;Rν) we mean thatu∈Wk,p(Qm;Rν) and
[Dku]Wσ,p(Qm)= Z
Qm
Z
Qm
|Dku(x)−Dku(y)|p
|x−y|m+σp dxdy 1/p
<+∞, and the associated norm is given by
kukWs,p(Qm)=kukLp(Qm)+
k
X
j=1
kDjukLp(Qm)+ [Dku]Wσ,p(Qm).
The fractional Sobolev spacesWs,p(Qm;Rν) arise in the trace theory of Sobolev spaces of integer order. For example, the trace is a continuous linear operator from W1,p(Qm;Rν) onto W1−1p,p(∂Qm;Rν) [12, Theorem 1.I].
We first address the question of strong density of smooth maps: given u ∈ Ws,p(Qm;Nn), does there exist a sequence inC∞(Qm;Nn) which converges tou with respect to the strong topology induced by theWs,p norm?
A naive approach consists in applying a standard regularization argument. This works well for maps inWs,p(Qm;Rν) and shows thatC∞(Qm;Rν) is strongly dense in that space. When Rν is replaced by Nn, the conclusion is less clear since the convolution of a mapu∈Ws,p(Qm;Nn) with a smooth kernelϕtyields a map with
Math. classification:58D15, 46E35, 46T20.
Keywords: Strong density; weak sequential density; Sobolev maps; fractional Sobolev spaces;
simply connectedness.
3
values in the convex hull ofNn. In this case, one might try to projectϕt∗uinto the manifoldNn. This is indeed possible for sp>m.
Ifsp > m, then by the Morrey-Sobolev imbedding,Ws,p is continuously imbed- ded intoC0. Thus, every mapu∈Ws,p(Qm;Nn) has a continuous representative andϕt∗uconverges uniformly ast tends to zero. In particular,
t→0lim sup
x∈Qm
dist (ϕt∗u(x), Nn) = 0. (1.1) Hence, one may project ϕt∗uback toNn since the nearest point projection Π is well defined and smooth on a neighborhood ofNn.
Ifsp=m, then the Morrey-Sobolev imbedding fails but property (1.1) remains true [31] since Ws,p injects continuously into the space VMO of functions with vanishing mean oscillation. This fact has been observed by Brezis and Nirenberg [9].
We may summarize as follows:
Theorem 1.1. — C∞(Qm;Nn)is strongly dense inWs,p(Qm;Nn)forsp>m.
The case wheresp < mis more subtle and the answer depends on the topology ofNn. Even whenNnis the unit sphereSnthe approximation problem is not fully understood. For instance, consider the mapu: B3 →S2 defined in the unit ball B3⊂R3by
u(x) = x
|x|.
Then, u ∈ Ws,p(B3;S2) for every s > 0 and p > 1 such that sp < 3, but u cannot be strongly approximated in Ws,p by smooth maps with values into S2 when 2 6sp <3. This example originally due to Schoen and Uhlenbeck [31] for s = 1 can be adapted to the case whereS2 is replaced by any compact manifold Nn and for any value ofs[10, Theorem 3; 23, Theorem 4.4]:
Theorem 1.2. — If sp < m and πbspc(Nn) 6= {0}, then C∞(Qm;Nn) is not strongly dense inWs,p(Qm;Nn).
It seems that the topological conditionπbspc(Nn)6={0}is the only obstruction to the strong density of smooth maps inWs,p(Qm;Nn). This is indeed true when sis an integer by a remarkable result of Bethuel [3, Theorem 1; 17] fors= 1 which has been recently generalized by the authors [6, Theorem 4] for anys∈N(see also [14]):
Theorem 1.3. — For any s ∈ N∗, if sp < m and πbspc(Nn) = {0}, then C∞(Qm;Nn)is strongly dense inWs,p(Qm;Nn).
Some cases of non-integer values have been investigated. For instance when s= 1−1/pin the setting of trace spaces [4, 24] and also whens>1 and Nn =Sn [5, 10]. Brezis and Mironescu [8] have announced in a personal communication a solution to the question of strong density for any 0< s <1.
All these cases give an affirmative answer to the following:
Open Problem 1.4. — Let s6∈N∗. Whensp < m and πbspc(Nn) ={0}, is it true thatC∞(Qm;Nn) is strongly dense inWs,p(Qm;Nn)?
In this paper, we investigate Open problem 1.4 for`simply connected manifolds Nn:
π0(Nn) =· · ·=π`(Nn) ={0}. (1.2) We prove the following:
Theorem 1.5. — Let s >1. If sp < m and if Nn is bspcsimply connected, thenC∞(Qm;Nn)is strongly dense inWs,p(Qm;Nn).
Even in the case wheresis an integer — which is covered in full generality by Theorem 1.3 — the proof is simpler and has its own interest. We have been inspired by Hajłasz [15] who has proved Theorem 1.5 fors= 1.
Our proof of Theorem 1.5 is based on two main ingredients. The geometric tool (Proposition 2.1) gives a smooth retraction of the ambient space Rν onto Nn except for a small subset of Nn. The analytic tool (Proposition 2.6) gives a pointwise estimate of the fractional derivative of η◦u, where η is a smooth map anduis aWs,p map.
The counterpart of Theorem 1.5 for 0< s <1 requires different tools and will be investigated in a subsequent paper.
The second problem we adress in this paper concerns theweak sequential density of C∞(Qm;Nn) in Ws,p(Qm;Nn): given u ∈ Ws,p(Qm;Nn), does there exist a sequence in C∞(Qm;Nn) which is bounded in Ws,p(Qm;Nn) and converges tou in measure?
The casesp>mhas an affirmative answer due to the strong density of smooth maps. When sp < m, we find the same topological obstruction as for the strong density problem whenspis not an integer [3, Theorem 3]:
Theorem 1.6. — Ifsp < mis such thatsp6∈Nand ifC∞(Qm;Nn)is weakly sequentially dense inWs,p(Qm;Nn), thenπbspc(Nn) ={0}.
From Theorem 1.3, it follows that for everys∈N∗such thatsp6∈Nthe problems of weak and strong density of smooth maps inWs,p(Qm;Nn) are equivalent. We expect the same is true fors6∈N; the missing ingredient would be an affirmative answer to Open Problem 1.4.
The conclusion of Theorem 1.6 need not be true when sp is an integer. For instance, by a result of Bethuel [2, Theorem 3],C∞(Q3;S2) is weakly sequentially dense inW1,2(Q3;S2), even though it is not strongly dense by Theorem 1.2.
As a byproduct of the tools we use to prove Theorem 1.5, we establish the following:
Theorem 1.7. — Lets>1. Ifsp < mis such thatsp∈Nand ifNn issp−1 simply connected, thenC∞(Qm;Nn)is weakly sequentially dense inWs,p(Qm;Nn).
This result is due to Hajłasz [15, Corollary 1] whens= 1; Hajłasz’s argument still applies forp= 1 although it is not explicitly stated in his paper. More recently, Hang and Lin [18, Corollary 8.6] proved an analogue of Theorem 1.7 under a weaker topological assumption for s = 1. To our knowledge, the only result concerning weak sequential density of smooth maps for non-integer values ofsdeals with the cases= 1/2,p= 2 andN =S1 and is due to Rivière [29, Theorem 1.2].
Combining Theorem 1.2 and Theorem 1.7 we deduce thatC∞(Qm;Sn) is weakly sequentially dense but not strongly dense inWs,p(Qm;Sn) forn < mandsp=n.
Whensp∈N, we do not know whetherC∞(Qm;Nn) is weakly sequentially dense inWs,p(Qm;Nn) with no extra assumption on the compact manifoldNn. The only results which are known in this sense concerns = 1: forp= 1 [16, Theorem 1.3;
27, Theorem I] and forp= 2 [28, Theorem I].
2. Main tools
2.1. Geometric tool. Our first tool is the construction of a retraction ofRν onto Nn except for a small subset of Nn. This is the only place where the topological assumption (1.2) concerning the`simply connectedness of the manifoldNn comes into place.
Proposition 2.1. — If Nn is ` simply connected, then for every 0 < 6 1 there exist a smooth functionη:Rν→Nn and a compact setK⊂Nn such that
(i) for everyx∈Nn\K,η(x) =x,
(ii) Hn(K)6C`+1, for some constantC >0 depending onNn andν, (iii) for everyj ∈N∗,
kDjηkL∞(Rν)6 C0 j,
for some constantC0 >0depending onNn,ν andj.
The setKis chosen as theneighborhood of ann−`−1 dimensional dual skeleton of Nn. This proposition is the smooth counterpart of Hajłasz’s construction of a Lipschitz continuous mapη [15, Section 4].
The proof of Proposition 2.1 relies on the existence of a triangulation of the manifoldNn. It is more convenient to use a variant of the triangulation based on the decomposition ofNn in terms of cubes rather than simplices.
A cubicationT ofNn is a finite collection of closed sets coveringNnof the form Φ(σ) withσ∈ Qsuch that
(a) Φ : S
σ∈Q
σ→Nn is a biLipschitz map,
(b) Qis a finite collection of cubes of dimensionmin some Euclidean spaceRµ, such that two elements ofQare either disjoint or intersect along a common face of dimension`for some`∈ {0, . . . , n}.
Given ` ∈ {0, . . . , n}, we denote by T` the union of all ` dimensional faces of elements ofT; we callT` the`dimensional skeleton ofT.
We recall the following lemma [15, Lemma 1]:
Lemma 2.2. — LetT be a cubication of Nn and let T` be the ` dimensional skeleton ofT. IfNnis`simply connected, then there exists a Lipschitz continuous functionη:Rν→Nn such that for every x∈T`,η(x) =x.
Proof. —LetCT`⊂R×Rν denote the cone
(λ, λx)∈R×Rν :λ∈[0,1] andx∈T` .
Since CT` is contractible, there exists a continuous map ξ:Rν →CT` such that for everyx∈T`,ξ(x) = (1, x).
We may choose ξ to be uniformly continuous. Indeed, ifp : Rν → Rν is any Lipschitz function such that pcoincides with the identity onT` andpis constant outside some ball containing T`, then for every x∈ T`, ξ◦p(x) = (1, x) and, in addition,ξ◦pis uniformly continuous. Replacingξbyξ◦pif necessary, we assume in the sequel thatξitself is uniformly continuous.
Since Nn is ` simply connected, the identity map in Nn is homotopic to a continuous map in Nn which is constant on T` [33, Section 6]. More precisely, there exist a continuous mapH : [0,1]×Nn→Nn anda∈Nn such that
(a) for everyx∈T`, H(0, x) =a, (b) for everyx∈Nn,H(1, x) =x.
SinceH is constant on{0} ×T`,H induces a continuous quotient mapH :CT`→ Nn defined for every (λ, λx) ∈ CT` by H(λ, λx) = H(λ, x). Then, H ◦ξ is a uniformly continuous map with values into Nn which coincides with the identity map onT`.
Using a standard approximation argument, we may construct a Lipschitz map having the same properties. We present the argument for the sake of completeness.
Givenι >0, let θ:Rν →[0,1] be a Lipschitz continuous function supported in a neighborhood ofT` such that
(a0) for everyx∈T`,θ(x) = 1,
(b0) for everyx∈suppθ,|x−H◦ξ(x)|6ι.
SinceH◦ξis uniformly continuous, there exists a Lipschitz approximationh:Rν → Rν such that for every x∈Rν,
|h(x)−H◦ξ(x)|6ι.
Then, for everyx∈Rν,
H◦ξ(x)− θ(x)x+ (1−θ(x))h(x) 6ι.
SinceH◦ξ(x)∈Nn, it follows that
θ(x)x+ (1−θ(x))h(x)∈Nn+Bνι,
where Bνι is the closed ball inRν of radiusι centered at 0. Choosingι such that the nearest point projection Π :Nn+Bνι →Nn is well-defined and smooth, then we have the conclusion by takingη:Rν→Nn defined forx∈Rν by
η(x) = Π θ(x)x+ (1−θ(x))h(x) .
The proof is complete.
We shall also use dual skeletons associated to a cubication T given by a map Φ : S
σ∈Q
σ→Nn.We first define dual skeletons for a cube inRn. Letj∈ {0, . . . , n}.
When the center of the cube is 0 and the faces are parallel to the coordinate axes, the dual skeleton of dimension j is the set of points in the cube which have at least n−j components equal to zero. By using an isometry, we can define the dual skeleton of a cube of dimensionn in Rµ in general position. Then, the dual skeleton of dimension j of a family Q of cubes as above is simply the union of the dual skeletons of dimension j of each cube. Finally, the dual skeleton Lj of dimensionj of the cubicationT ofNnis the image by Φ of thej dimensional dual skeleton ofQ.
The following lemma implies the homotopy equivalence between the skeletonT` of the manifold Nn and the complement of the dual skeleton Ln−`−1 in Nn. We are particularly interested in the pointwise estimates of the homotopyf:
Lemma 2.3. — Let ` ∈ {0, . . . , n−1}, let T be a cubication of Nn and let Ln−`−1be then−`−1dimensional dual skeleton ofT. Then, there exists a locally Lipschitz continuous function
f : [0,1]×(Nn\Ln−`−1)→Nn such that
(i) for everyt∈[0,1]and for everyx∈T`,f(t, x) =x, (ii) for everyx∈Nn\Ln−`−1,f(0, x) =xandf(1, x)∈T`, (iii) for everyt∈[0,1]and for everyx∈Nn\Ln−`−1,
|∂tf(t, x)|6C, and
|∂xf(t, x)|6 C0 dist (x, Ln−`−1),
for some constants C, C0>0depending onn,`, Nn andT.
Proof. —We first establish the result when the manifold Nn is replaced by the cube [−1,1]n andLn−`−1 is the dual skeleton of dimension n−`−1 of [−1,1]n. Following [33], we consider for everyx∈[−1,1]n,
|x|`= min
S⊂{1,...,n}
|S|=`+1
maxi∈S |xi|.
In particular, for every x ∈ [−1,1]n, x ∈ Ln−`−1 if and only if |x|` = 0. The functionx∈[−1,1]n7→ |x|` is Lipschitz continuous of constant 1.
Letφ`: [−1,1]n\Ln−`−1→T`be defined for every x∈[−1,1]n by φ`(x) = (y1, . . . , yn),
where
yi=
(sgnxi if|xi|>|x|`, xi/|x|` if|xi|<|x|`.
The homotopyf : [0,1]×([−1,1]n\Ln−`−1)→[−1,1]n defined by f(t, x) = (1−t)x+tφ`(x)
has the required properties.
In order to prove the existence of the homotopyf for a general compact manifold Nn, we perform the above construction in every cube of a given cubication Φ :
S
σ∈Q
σ→Nn. If two cubesσ1andσ2 inQhave a non empty intersection, then the corresponding mapsφ`,1 andφ`,2 coincide on the common faceσ1∩σ2. Hence, we can glue together the locally Lipschitz continuous maps obtained for each cube so as to obtain a global map f0 which is defined on the entire collection of cubes in Q. The conclusion follows by taking
f(t, x) = Φ f0(t,Φ−1(x))
.
We now prove a counterpart of Proposition 2.1 for a Lipschitz continuous map η:
Lemma 2.4. — Let`∈ {0, . . . , n−1}, let T be a cubication ofNnandLn−`−1 be then−`−1dimensional dual skeleton ofT and letι >0be such that the nearest point projectionΠ ontoNn is smooth on Nn+Bν2ι. IfNn is `simply connected, then for every0< 61there exists a Lipschitz continuous mapη:Rν →Nnsuch that
(i) η= Πon(Nn+Bιν)\Π−1(Ln−`−1+Bν), (ii) for everyx∈Rν,
|Dη(x)|6 C00 .
for some constantC00>0depending onNn,T andν.
Proof. —Letf be the map given by Lemma 2.3. The extension f¯: {0} ×Ln−`−1
∪ [0,1]×(Nn\Ln−`−1)
→Nn defined by
f¯(t, x) =
(x ift= 0, f(t, x) if 0< t61, is continuous.
Let Π be the nearest point projection onto Nn and denote by Π :Rν → Rν a smooth extension of Π. The image of Π need not be contained in the manifoldNn.
Letθ:Rν →[0,2] be a Lipschitz continuous function such that (a) for everyx∈Nn+Bιν, θ(x) = 2,
(b) for everyx∈Rν\(Nn+B2ιν),θ(x) = 0.
Given 0< 61, letd:Nn+B2ιν →Rbe defined by d(x) = 1
dist (Π(x), Ln−`−1).
Letλ: [0,+∞)→[0,1] be a Lipschitz continuous function such that (a0) for everyt612 and for everyt>2,λ(t) = 0,
(b0) λ(1) = 1.
Denote byη:Rν →Nn the function given by Lemma 2.2. Let η:Rν →Nn be the map defined by
η(x) =
f λ(θ(x)d(x)),Π(x)
ifx∈Nn+B2ιν and θ(x)d(x)>1, η◦f λ(θ(x)d(x)),Π(x)
ifx∈Nn+B2ιν and θ(x)d(x)61, η Π(x)
ifx6∈Nn+B2ιν.
We first check thatη is continuous. For this purpose we only need to consider the borderline cases:
(1) x∈Nn+Bν2ι andθ(x)d(x) = 1,
(2) x∈∂(Nn+B2ιν).
In the first case, since λ(1) = 1, f(1,·)∈ T` andη is the identity map on T`, we have
f λ(θ(x)d(x)),Π(x)
=f(1,Π(x))
=η f(1,Π(x))
=η◦f λ(θ(x)d(x)),Π(x) . In the second case,θ(x) = 0. Sinceλ(0) = 0 andf(0,·) is the identity map onNn,
f λ(θ(x)d(x)),Π(x)
= Π(x) = Π(x), whence
η◦f λ(θ(x)d(x)),Π(x)
=η Π(x) . We now check that property (i) holds. Indeed, if
x∈(Nn+Bνι)\Π−1(Ln−`−1+Bν), thenθ(x) = 2 andd(x)>1. Thus,
λ(θ(x)d(x)) = 0.
We then have
η(x) =f(0,Π(x)) = Π(x).
It remains to establish property (ii). Indeed, if x 6∈ Nn +B2ιν, then η(x) = η Π(x)
and the conclusion follows sinceηand Π(x) are both Lipschitz continuous, with Lipschitz constants independent of . If x ∈ Nn +Bν2ι and θ(x)d(x) < 12, then η(x) =η◦Π(x) and the estimate follows similarly. Finally, ifx∈Nn+B2ιν andθ(x)d(x)> 12, then
dist (Π(x), Ln−`−1)>
4. By the chain rule and the estimates given by Lemma 2.3,
|Dη(x)|6C1
1
+ 1
dist (Π(x), Ln−`−1)
.
Combining both estimates, we get the conclusion. The proof of the lemma is
complete.
We now have all tools to prove Proposition 2.1.
Proof of Proposition 2.1. — Letϕ:Rν →Rbe a smooth map supported in the unit ballBν1. For everyt >0, letϕt:Rν →R be the function defined forx∈Rν byϕt(x) = t1νϕ(xt). Letι >0 as in the previous lemma.
Given 0< 61, letζ:Rν→[0,1] be a smooth function such that (a) for everyx∈Nn\(Ln−`−1+B2ν),ζ(x) = 1,
(b) for everyx6∈(Nn+Bιν)\Π−1(Ln−`−1+Bν),ζ(x) = 0, (c) for everyj∈N∗,
kDjζkL∞(Rν)6 C1 j , whereC1>0 depends onj.
Letη:Rν→Nnbe the Lipschitz continuous map given by the previous lemma and lett >0 to be chosen below.
By property (b) and by Lemma 2.4 (i), the function ζη+ (1−ζ)ϕt∗η
is smooth inRν and for every j∈N∗ there existsC2>0 such that Dj ζη+ (1−ζ)ϕt∗η
L∞(
Rν)6C2 1 + 1
tj + 1 j
. (2.1)
Moreover, by property (a) and by Lemma 2.4 (i), for everyx∈Nn\(Ln−`−1+B2ν), ζη+ (1−ζ)ϕt∗η
(x) =η(x) = Π(x) =x.
By Lemma 2.4 (ii) we have for everyt >0,
kϕt∗η−ηkL∞(Rν)6tkDηkL∞(Rν)6tC3
. Taking
t= ι C3,
it follows from the previous estimate that the image ofζη+(1−ζ)ϕt∗ηis contained inNn+Bιν. Hence, the functionη:Rν→Nn,
η= Π◦ ζη+ (1−ζ)ϕt∗η
, is well-defined and smooth. Property (i) holds with
K=Nn∩(Ln−`−1+B2ν).
Property (ii) also holds sinceKis a neighborhood ofLn−`−1inNnwhose radius is of the order of. By estimate (2.1), property (iii) is also satisfied. This completes
the proof of the proposition.
2.2. Analytic tool. In this section we establish pointwise estimates of derivatives and fractional derivatives of the map η◦u, where η is a smooth function and u belongs toWs,p∩L∞. In the case wheresis an integer, this estimate follows from the classical chain rule for higher order derivatives:
Proposition 2.5. — Letk∈N∗. Ifu∈Wk,p(Qm;Rν)∩W1,kp(Qm;Rν), then for everyj∈ {1, . . . , k}there exists a measurable functionGj∈Lp(Qm)such that for every smooth mapη:Rν→Rν,
|Dj(η◦u)|6[η]Cj(Rν)Gj.
Moreover,kGjkLp(Qm)is bounded from above by a constant depending onk,p,m, kukWk,p(Qm)andkukW1,kp(Qm).
We use the following notation:
[η]Cj(Rν)=
j
X
i=1
kDiηkL∞(Rν).
Proof. —We first observe that η◦u∈Wk,p(Qm;Rν). By the chain rule,
|Dj(η◦u)(x)|6C1 j
X
i=1
|Diη(u(x))| X
16t16···6ti6j, t1+···+ti=j
|Dt1u(x)| · · · |Dtiu(x)|
6C1[η]Cj(Rν) j
X
i=1
X
16t16···6ti6j, t1+···+ti=j
|Dt1u(x)| · · · |Dtiu(x)|.
Let
Gj =C1 j
X
i=1
X
16t16···6ti6j, t1+···+ti=j
|Dt1u| · · · |Dtiu|
Since the map u in the statement belongs toWk,p(Qm;Rν)∩W1,kp(Qm;Rν), it follows from the Gagliardo-Nirenberg interpolation inequality [13, 25] that
Diu∈Ljpi(Qm).
By Hölder’s inequality, we deduce thatGj ∈Lp(Qm).
We now establish a counterpart of the previous proposition for the fractional derivative introduced by Maz0ya and Shaposhnikova [22]. More precisely, given 0< σ <1, 16p <+∞, a domain Ω⊂Rmand a measurable functionu: Ω→Rν, define forx∈Ω,
Dσ,pu(x) = Z
Ω
|u(x)−u(y)|p
|x−y|m+σp dy 1/p
.
We extend this definition for anys >0 such thats /∈Nas follows:
Ds,pu=Dσ,p(Dku),
where k=bscis the integral part of sand σ=s− bscis the fractional part of s.
Using this notation, we have
[Dku]Wσ,p(Ω)=kDs,pukLp(Ω).
Proposition 2.6. — Let s > 1 be such thats 6∈ N. If u ∈ Ws,p(Qm;Rν)∩ W1,sp(Qm;Rν), then there exists a measurable functionH∈Lp(Qm)such that for every smooth mapη :Rν→Rν,
|Ds,p(η◦u)|6[η]σCk+1(Rν)[η]1−σCk(
Rν)H.
Moreover,kHkLp(Qm)is bounded from above by some constant depending ons,p, m,kukWs,p(Qm) andkukW1,sp(Qm).
This proposition implies a theorem of Brezis and Mironescu [8, Theorem 1.1]
concerning the boundedness of the composition operator from Ws,p∩W1,sp into Ws,p. A more elementary proof of the same result has been provided by Maz0ya and Shaposhnikova [22]; our proof of Proposition 2.6 is based on their strategy.
We begin with the following pointwise estimate of Maz0ya and Shaposhnikova [22, Lemma]:
Lemma 2.7. — Letq>1. Ifv∈Wloc1,q(Rm;Rν), then forx∈Rm, Dσ,qv(x)q
6C M|Dv|q(x)σ
M|v|q(x)1−σ
, for some constantC >0 depending onmandq.
The maximal function associated to a nonnegative function f ∈ L1loc(Rm) is defined forx∈Rmby
Mf(x) = sup
ρ>0
1
|Bρm| Z
Bρm(x)
f.
For completeness, we prove Lemma 2.7 using a property of the maximal function due to Hedberg [20, Lemma]:
Lemma 2.8. — Letf ∈L1loc(Rm)be a nonnegative function and letδ >0. For everyx∈Rmandρ >0,
Z
Bmρ(x)
f(y)
|y−x|m−δ dy6CρδMf(x), Z
Rm\Bρm(x)
f(y)
|y−x|m+δ dy6 C
ρδMf(x), for some constantC >0 depending onmandδ.
Proof. —We briefly sketch the proof of Hedberg for the first estimate. The proof of the second one is similar. One has
Z
Bmρ(x)
f(y)
|y−x|m−δ dy=
∞
X
i=0
Z
Bm
ρ2−i(x)\Bm
ρ2−i−1(x)
f(y)
|x−y|m−δ dy
6C1ρδ
∞
X
i=0
2−δiMf(x)6C2ρδMf(x).
Proof of Lemma 2.7. — Letρ >0. By Hardy’s inequality [21, Section 1.3], Z
Bρm(x)
|v(x)−v(y)|q
|x−y|m+σq dy6C1
Z
Bmρ(x)
|Dv(y)|q
|x−y|m−(1−σ)q dy.
Thus, by Hedberg’s lemma, Z
Bmρ(x)
|v(x)−v(y)|q
|x−y|m+σq dy6C2ρ(1−σ)qM|Dv|q(x).
Since
|v(x)−v(y)|q 6C3 |v(x)|q+|v(y)|q , by an explicit integral computation and by Hedberg’s lemma,
Z
Rm\Bmρ(x)
|v(x)−v(y)|q
|x−y|m+σq dy6 C4
ρσq |v(x)|q+M|v|q(x) 6 C5
ρσqM|v|q(x).
We conclude that
Dσ,qv(x)q
6C2ρ(1−σ)qM|Dv|q(x) + C5
ρσqM|v|q(x).
Minimizing the right-hand side with respect toρ, we deduce the pointwise estimate.
The following lemma is implicitly proved in [22, Section 2]:
Lemma 2.9. — Let 0 < σ < 1, 1 6 p < +∞ and i ∈ N∗. If for every α ∈ {1, . . . , i},vα∈Lqα(Rm)andDvα∈Lrα(Rm), where 1< rα< qαand
1−σ qα + σ
rα +
i
X
β=1 β6=α
1 qβ = 1
p,
then Qi
α=1
vα
Wσ,p(Rm)6C
i
X
α=1
kvαk1−σLqα(
Rm)kDvαkσLrα(Rm) i
Y
β=1 β6=α
kvβkLqβ(Rm)
,
for some constantC >0 depending onm, σ,r1, . . . ,ri,q1, . . . ,qi. Proof. —We first consider the case of dimensionm= 1. Note that
i
Q
α=1
vα(x)−
i
Q
α=1
vα(y) 6
i
X
α=1
v1(x)· · ·vα−1(x) vα(x)−vα(y)
vα+1(y)· · ·vi(y) . (2.2) Thus, the left-hand side is bounded from above by a sum of functions of the form
fα(x)|vα(x)−vα(y)|fα(y).
By the Fundamental theorem of Calculus, for everyx, y∈R,
|vα(x)−vα(y)|62|x−y|M|vα0|(x).
Thus, for everyρ >0, Z
B1ρ(x)
|vα(x)−vα(y)|p
|x−y|1+σp (fα(y))pdy6C1 M|v0α|(x)p Z
Bρ1(x)
(fα(y))p
|x−y|1−(1−σ)pdy.
By Hedberg’s lemma, we get Z
Bρ1(x)
|vα(x)−vα(y)|p
|x−y|1+σp (fα(y))pdy6C2ρ(1−σ)p M|v0α|(x)p
M(fα)p(x).
Next, we write
|vα(x)−vα(y)|p(fα(y))p6C3 |vα(x)|p(fα(y))p+|vα(y)|p(fα(y))p . By Hedberg’s lemma, we also have
Z
R\B1ρ(x)
|vα(x)−vα(y)|p
|x−y|1+σp (fα(y))pdy6 C4
ρσp |vα(x)|pM(fα)p(x) +M|vαfα|p(x) .
We conclude that Z
R
|vα(x)−vα(y)|p
|x−y|1+σp (fα(y))pdy
6C2ρ(1−σ)p M|v0α|(x)p
M(fα)p(x) + C4
ρσp |vα(x)|pM(fα)p(x) +M|vαfα|p(x) . Minimizing the right hand side with respect toρ, we then get
Z
R
|vα(x)−vα(y)|p
|x−y|1+σp (fα(y))pdy 6C5 M|vα0|(x)σp
(M(fα)p(x))σ |vα(x)|pM(fα)p(x) +M|vαfα|p(x)1−σ
. Thus,
Z
R
Z
R
(fα(x))p|vα(x)−vα(y)|p
|x−y|1+σp (fα(y))pdxdy 6C5
Z
R
(fα)p M|vα0|σp
M(fα)pσ
|vα|pM(fα)p+M|vαfα|p1−σ
. (2.3)
Let q1
α
=
α−1
P
β=1 1
qβ and q1
α =
i
P
β=α+1 1
qβ, so that by assumption, 1
qα
+ σ rα
+1−σ qα
+ 1 qα =1
p. By Hölder’s inequality,
Z
R
(fα)p M|v0α|σp
|vα|(1−σ)pM(fα)p 6kfαkp
Lqα(R)kM|v0α|kσpLrα(
R)kvαk(1−σ)pLqα(
R)k(M(fα)p)1/pkp
Lqα(R). We estimate the right hand side as follows. By Hölder’s inequality,
kfαkLqα(
R)6
α−1
Y
β=1
kvβkLqβ(R).