Nonlinear Analysis 193 (2020) 111311

Contents lists available at ScienceDirect

Nonlinear Analysis
www.elsevier.com/locate/na

Semilinear elliptic equations for the fractional Laplacian with

Hardy potential
Mouhamed Moustapha Fall
African Institute for Mathematical Sciences in Senegal, KM2 Route de Joal Mbour, BP 1418, Senegal

article info abstract

Article history: In this paper we study existence and nonexistence of nonnegative distributional
Received 23 May 2018 solutions for a class of semilinear fractional elliptic equations involving the Hardy
Accepted 9 July 2018 potential.
Communicated by Enzo Mitidieri © 2018 Elsevier Ltd. All rights reserved.

Keywords:
Hardy inequality
Critical exponent
Nonexistence
Distributional solutions
Fractional Laplacian

1. Introduction

Let B be a ball of RN , N > 2s, centered at 0. Let s ∈ (0, 1), p > 1 and γ ≥ 0. In this paper, we study
existence and nonexistence of nonnegative functions u ∈ L1s ∩ Lploc (B) satisfying
−2s
(−∆)s u − γ|x| u = up in B, (1.1)

where
{ }
|u|
∫
L1s = u : RN → R : N +2s
<∞ .
RN 1 + |x|
Equality (1.1) is understood in the sense of distributions. The distribution (−∆)s u is defined as
∫
⟨(−∆)s u, φ⟩ = u(−∆)s φdx ∀φ ∈ Cc∞ (B).
RN
s
Here the fractional Laplacian (−∆) is defined via the Fourier transform as
∫
1 2s
(−∆)s φ(x) = N
|ζ| φ̂(ζ)eıζ·x dζ, (1.2)
(2π) 2 RN

E-mail address: mouhamed.m.fall@aims-senegal.org.

https://doi.org/10.1016/j.na.2018.07.008
0362-546X/© 2018 Elsevier Ltd. All rights reserved.
2 M.M. Fall / Nonlinear Analysis 193 (2020) 111311

where
∫
1
φ̂(ζ) = F(φ)(ζ) = N e−ıζ·x φ(x)dx
(2π) 2 RN

is the Fourier transform of φ. The nonlocal structure of the fractional Laplacian (−∆)s can be seen in its
representation in the real space
φ(x) − φ(y)
∫
s
(−∆) φ(x) = Cs,N P.V. N +2s
dy, (1.3)
R N |x − y|
for some positive constant Cs,N . For the equivalence between (1.3) and (1.2), we refer the reader to [38].
Problem (1.1) is related to the relativistic Hardy inequality, which were proved by Herbst in [33] for
s = 1/2 and Yafayev in [51] for all s ∈ (0, 1), and reads as follows
∫ ∫
−2s 2s
γ0 |x| u2 dx ≤ |ζ| û2 dζ ∀u ∈ Cc∞ (RN ), (1.4)
RN RN

where
2
( N +2s )
2s Γ 4 )
γ0 = 2 ( N −2s . (1.5)
Γ2 4
2
The constant γ0 is optimal and converges to the classical Hardy constant (N −2) 4 when s → 1. Here Γ is the
usual gamma function. We should mention that (1.4) is a particular case of the Stein and Weiss inequality,
see [47].
A great deal of work is currently been devoted to the study of the fractional Laplacian as it appears in
many fields such as probability theory, physics and mathematical finance. We refer the reader to papers
[12,45,46,28,5] (and the references there in) for a nice expository. A good reference for the potential theory
−2s
of (−∆)s can be found in the book of Landkof [38]. The operator (−∆)s − γ0 |x| appears in the problem
of stability of relativistic matter in magnetic fields. One can see [29] where a lower bound and a Gagliardo–
Nirenberg-type inequality were proved.
The problem of existence and nonexistence of (1.1), for s = 1, was studied by Brezis–Dupaigne–Tesei
N +2−2β
in [8] where the authors showed that for β ∈ 0, N 2−2 , 1 < p < N
[ ]
−2−2β , the problem

(N − 2)2
( )
−2
− ∆u − − β 2 |x| u = up in D′ (B)
4
has a positive solution u ∈ Lp (B) and does not have any nonnegative and nontrivial supersolution
u ∈ Lploc (B \ {0}) when β ∈ 0, N 2−2 and p ≥ N N +2−2β
[ )
−2−2β . Some related results and problems are
in [7,8,17,19,20,22,49,26,25,23,9,10].
Our results in this paper extends the one of Brezis–Dupaigne–Tesei in [8] to the case s ∈ (0, 1). Before
stating them, we fix the following notation: for α ∈ 0, N −2s
[ )
2 , we put
( N +2s+2α ) ( N +2s−2α )
Γ Γ
γα = 22s ( N −2s−2α
4 4
) ( N −2s+2α ).
Γ 4 Γ 4
N −2s
The mapping α ↦→ γα is monotone decreasing and γα → 0 when α → 2 . We should mention that this
2s−N
−2s
constant appears in the perturbation of the ’ground-state’ |x| 2 for the operator (−∆)s −γ0 |x| . Indeed,
2s−N
+α
letting ϑα = |x| 2 we have
−2s
(−∆)s ϑα − γα |x| ϑα = 0 in RN \ {0},

see Lemma 4.1 in Section 4.

 M.M. Fall / Nonlinear Analysis 193 (2020) 111311 3

Our existence result is the following

Theorem 1.1. Let α ∈ 0, N −2s N +2s−2α

[ ]
2 and 1 < p < N −2s−2α . There exists a function u ∈ L1s ∩ Lp (B) satisfying
u > 0 in B and
−2s
(−∆)s u − γα |x| u = up in D′ (B).

As what concerns nonexistence, we have obtained the

Theorem 1.2. Let α ∈ 0, N −2s and u ∈ L1s ∩ Lploc (B \ {0}) such that u ≥ 0 on RN and
[ )
2
−2s
(−∆)s u − γα |x| u ≥ up in D′ (B \ {0}).
N +2s−2α
If p ≥ N −2s−2α , then u = 0 in B.

The Pohozaev type result in [27] shows that there is no positive energy solutions with zero exterior
N +2s
Dirichlet data, provided p ≥ N −2s . Sharp asymptotic behavior for energy solutions to fractional elliptic
equations involving Hardy potentials was also proved in [24]. Fractional Hardy inequality with boundary
singularity was studied in [6]. We observe that if α = 0 we have p+1 = N2N−2s : the critical Hardy–Littlewood–
N +2s−2α N −2s
Sobolev exponent (see [39]) and that N −2s−2α → +∞ as α → 2 .
The proof of Theorem 1.2 relies on weak comparison principles techniques recently introduced by the
author in [22]. However substantial difficulties have to be overcome due to the nonlocal structure of the
fractional Laplacian. Nonexistence result of nonlinear elliptic problems using comparison principles have
been obtained in [1,2,44,41,36,35,37] and the references therein. Existence and qualitative properties of
solutions to semilinear elliptic equations involving the Hardy potential have been also studied in [27,24].
N +2s−2α N +2s
For the existence result, in the supercritical case N −2s−2α > p ≥ N −2s , we have an explicit solution
2N
constructed via ϑα . In the subcritical case, p + 1 < N −2s , we used standard variational arguments thanks
to the following improved fractional Hardy inequality:
( )
2
Theorem 1.3. Let 2 > q > max 1, 1+2s . Then there exists a constant C > 0 such that for all u ∈ Cc∞ (B),
∫ ∫
2 −2s 2 2s
C∥u∥W τ,q (B) ≤ γ0 |x| u dx − |ζ| û2 dζ, (1.6)
0
B RN
1+2s 1
where τ = 2 − q.

This result, which might be of self interest, is proved in Appendix.

The proof of the results presented above are manly based on a Dirichlet-to-Neumann operator Bs for
which Bs u = (−∆)s ũ in D′ (E) for any Lipschitz bounded open set E of RN and ũ is the null extension
outside E of a function u belonging to some Sobolev space. We start with some notations and definition on
fractional Sobolev spaces before. For s > 0, the Sobolev space H s (RN ) is defined as
s
H s (RN ) = {u ∈ L2 (RN ) : |ζ| û ∈ L2 (RN )}

with norm
s
∥u∥H s (RN ) = ∥û∥L2 (RN ) + ∥|ζ| û∥L2 (RN ) .

Let E be a bounded domain in RN with Lipschitz boundary. See e.g. [32], recall that for q > 1, the fractional
Sobolev space W s,q (E) is given by the set of measurable functions u such that
q
|u(x) − u(y)|
∫ ∫
q q
∥u∥W s,q (E) := ∥u∥Lq (E) + N +qs
dxdy < ∞.
E E |x − y|
4 M.M. Fall / Nonlinear Analysis 193 (2020) 111311

We define W0s,q (E) to be the closure of Cc∞ (E) with respect to the norm ∥·∥W s,q (E) . As a notation convention,
we put H s (E) = W s,2 (E) and H0s (E) = W0s,2 (E) which are Hilbert spaces.
It is well known that if u ∈ H 1 (E) then ũ, its null extension outside E, is in H 1 (RN ) and ∥u∥H 1 (E) =
∥u∥H 1 (RN ) . This is not in general true for functions in H s (E) (s = 1/2 for instance). We shall define a space
of functions in which we recover this defect by imposing integrability of null extensions.
We define the Hilbert space D s,2 (RN ) as the completion of Cc∞ (RN ) with respect to the norm:
∫
2s 2
v ↦→ |ζ| |v̂| dζ. (1.7)
RN

As it will be apparently clear in the remaining of the paper, it will be useful to consider the Hilbert space

H0s (E) := u ∈ H s (E) : ũ ∈ D s,2 (RN ) ,

{ }
(1.8)

where we put here (and hereafter)

{
u in E
ũ =
0 in RN \ E.

The space H0s (E) is endowed with the norm

∫ ∫
2 2s ˆ 2 2s 2
∥u∥H s (E) = |ζ| |ũ| dζ = |ζ| |F(ũ)| dζ. (1.9)
0
RN RN

Note that, since E is bounded, by (1.4) there exists a constant C(E) > 0 such that

C(E)∥ũ∥H s (RN ) ≤ ∥u∥H0s (E) ≤ ∥ũ∥H s (RN ) ∀u ∈ H0s (E). (1.10)

Therefore

H0s (E) = {u ∈ H s (E) : ũ ∈ H s (RN )}.

See for instance [[32], Theorem 1.4.2.2 ], the space Cc∞ (E) is dense in H0s (E).
Given w ∈ H s (RN ), minimization procedure yields the existence of a function H(w) ∈ Hloc
1
(RN ) being
N +1
the harmonic extension of w over the half space R+ :

div(t1−2s ∇H(w)) = 0 in RN +1
{
+ ,
N (1.11)
H(w) = w on R .

In [15], Caffarelli and Silvestre proved that (−∆)s w is given by the Dirichlet-to-Neumann operator
w ↦→ limt→0 t1−2s ∂H(w)
∂t . Namely,

∂H(w)
− lim t1−2s = κs (−∆)s w in D s,2 (RN ),
t→0 ∂t
for some constant κs > 0. In addition,
∫ ∫
2 2s
t1−2s |∇H(w)| dxdt = κs |ζ| ŵdζ. (1.12)
RN
+
+1
RN

We want to provide similar arguments in bounded open sets. For any u ∈ H0s (E), we can consider its
harmonic extension H(ũ) as in (1.11). We define the Dirichlet-to-Neumann operator Bs : H0s (E) → H −s (E)
given by
∂H(ũ)
Bs u = −κ−1
s lim t
1−2s
,
t→0 ∂t
 M.M. Fall / Nonlinear Analysis 193 (2020) 111311 5

where H −s (E) is the dual of H0s (E). This operator turns out to be linear and it is an isometry,

∥Bs u∥H −s (E) = ∥u∥H0s (E) ∀u ∈ H0s (E)

by (1.12). Moreover

Bs u = (−∆)s ũ in D′ (E) ∀u ∈ H0s (E).

In particular a solution u ∈ H0s (E) to the problem

Bs u = f in E

yields a solution v ∈ D s,2 (RN ) to the problem

(−∆)s v = f
{
in E,
v=0 in RN \ E

and conversely. We refer to the next section for more details.

In order to get, say, qualitative informations on the solution to the problem

Bs u = (−∆)s ũ = f,

it is, in some cases, more convenient to work with the (mixed) problem

⎨div(t1−2s ∇H(ũ)) = 0 in RN +1
⎧
+ ,
∂H(ũ) (1.13)
⎩− lim t1−2s = κs f in E.
t→0 ∂t
New difficulties arise here because of the (possible) degeneracy of Eq. (1.13). However the weight t1−2s falls
into the Muckenhoupt class of weights thus regularity results, Harnack inequalities are available (see [21])
and this is enough for our purpose in this paper.
An interesting characterization of H0s (E), see [32], is that H0s (E) is the interpolation space (H02 (E),
L2 (E))s,2 :
⎧ s
⎨H (E) s ∈ (0, 1/2),
H0s (E) = H s (E) s = 1/2, (1.14)
⎩ s 00
H0 (E) s ∈ (1/2, 1),

where
u2 (x)
{ ∫ }
1 1
2
H00 (E) = u ∈ H 2 (E) : dx < ∞ ,
E d(x)

endowed with the natural norm, with d(x) = dist(x, ∂E).

Remark 1.4. Let E be a smooth bounded domain of RN . Recently a pseudo differential operator As of
order 2s was introduced by Cabré and Tan [13] for s = 1/2 (see [16] for every s ̸= 1/2) in the following way:
for any u ∈ H0s (E)
∞
∑
As u = µsk uk φk ,
k=1

where µk is the zero Dirichlet eigenvalues of −∆ with corresponding orthonormal eigenfunctions φk and
uk = E uφk dx is the component of u in the L2 (E) basis {φk }.
∫
6 M.M. Fall / Nonlinear Analysis 193 (2020) 111311

Using (1.14), it was shown in [13,16] that

∞
{ }
∑ 2
H0s (E) = u ∈ L2 (E) : µsk |uk | < ∞ . (1.15)
k=1

The operator As corresponds to the Dirichlet-to-Neumann operator given by the harmonic extension over
1
the cylinder E × (0, ∞). Indeed, let HL,s (E × (0, ∞)) be the set of measurable functions w : E × (0, ∞) → R
1
with w ∈ H (E × (r1 , r2 )), 0 < r1 < r2 < ∞ and w = 0 on ∂E × (0, ∞) such that the following norm
∫
2 2
∥w∥H 1 (E×(0,∞)) = t1−2s |∇w| dxdt < ∞.
L,s E×(0,∞)

In [13,16], the authors showed that for any g ∈ H −s (E) there exists a unique solution u ∈ H0s (E) to
{
As u = g in E,
(1.16)
u = 0 on ∂E.
1
In addition u is the trace of w ∈ HL,s (E × (0, ∞)) which is the unique solution to
⎧ 1−2s
⎨div(t
⎪ ∇w) = 0 in E × (0, ∞),
w = 0 on ∂E × (0, ∞), (1.17)
⎩−t1−2s ∂w = κ g on E,
⎪
N,s
∂t
where κN,s (κN,s = 1 for s = 1/2) is a constant depending only on N and s. Moreover, it holds that, with
the norm in (1.15),
2
∥u∥2 = κN,s ∥w∥H 1 (E×(0,∞)) . (1.18)
L,s

We can compare the operator As with the operator Bs . For simplicity, we consider the case s = 1/2.
Assume that g ∈ Cc∞ (E) is nonnegative and nontrivial and u is a solution to (1.16), which is positive on E.
Take w its extension over the cylinder. Consider H(ũ) which is the harmonic extension of ũ in RN+
+1
given
by (1.13). Clearly

H(ũ) ≥ w̃ in RN
+
+1
.

It follows from Hopf lemma that

∂w ∂H(ũ)
− >− in E.
∂t ∂t
Hence

A1/2 u > B1/2 u in E.

In particular the operator As yields (up to a multiplicative constant) subsolution to the fractional Laplacian
(−∆)s . This is the reason why the use of Bs is more convenient in this paper.

We give here the plan of the paper:

• Section 2: Notations and preliminaries.

• Section 2.1: Dirichlet-to-Neumann operator.
• Section 3: Comparison and maximum principles.
• Section 4: Nonexistence of positive supersolutions.
• Section 5: Existence of positive solutions.
• Appendix A.1: Remainder term for the fractional Hardy inequality.
 M.M. Fall / Nonlinear Analysis 193 (2020) 111311 7

2. Notations

For G an open set of RN and a nonnegative weight function a, we use the standard notations for
p
weighted Lebesgue spaces Lp (G; a(x)) = {u : G → RM : G |u| a(x)dx < ∞}. B N (0, r) is a ball in
∫

RN centered at 0 with radius r > 0 and S N −1 = ∂B N (0, 1). RN +

+1
= {(t, x) : t > 0, x ∈ RN }.
N +1 N +1 N +1 N +1
B+ (0, r) = R+ ∩ B (0, r) and S+ = R+ ∩ S . If there is no confusion, we will put B N = B N (0, 1)
N N
N +1 N +1 1,q N +1 N +1 N
and B+ = B+ (0, 1). The space W0,S (B+ (0, r); a(x)) = {u ∈ W 1,q (B+ (0, r); a(x)) : u = 0 on S+ }.
N +1 1,2 1−2s ∞
If G is an open subset of R+ , we let D (G; t ) be the completion of Cc (G) with respect to the scalar
product (u, v) ↦→ ⟨u, v⟩L2 (G;t1−2s ) .

2.1. Dirichlet-to-Neumann operator

It is well known that the space of Schwartz functions S contains Cc∞ (RN ) and that F is a bijection from
S into itself. In particular (−∆)s φ ∈ C ∞ (RN ) ∩ L∞ (RN ) for every φ ∈ Cc∞ (RN ). In fact we have for any
φ ∈ Cc2 (RN ), see [5,27],
∥φ∥Cc2
|(−∆)s φ(x)| ≤ C N +2s
∀x ∈ RN .
1 + |x|
This motivates the following:

Definition 2.1. Let G be an open subset of RN . Given u ∈ L1s , the distribution (−∆)s u ∈ D′ (G) is defined as
∫
s
⟨(−∆) u, φ⟩ = u(−∆)s φdx ∀φ ∈ Cc∞ (G).
RN

Some recent results concerning s-superharmonic functions in the sense of distributions as above are in [45].
Consider the Poisson kernel of RN+
+1
:= {(t, x) : t > 0, x ∈ RN }
1
P (t, x) = pN,s t2s 2 , (2.1)
(|x| + t2 )(N +2s)/2
where pN,s is a normalization constant, see [12] for an explicit value. Let u ∈ L1s , we can define
∫
u(y)
ū(t, x) = P (t, ·) ∗ u = pN,s t2s ) N +2s dy ∀(t, x) ∈ RN
+
+1
.
RN
( 2
2 2
|y − x| + t

As in [15], it turns out that

div(t1−2s ∇ū) = 0 RN
+
+1
.

Therefore ū is smooth in RN
+
+1
. Moreover if u is continuous at some point x0 then

lim ū(t, x0 ) → u(x0 ).

t→0

By an argument of [15], we have that

∂ ū
− lim t1−2s (t, x0 ) = κs (−∆)s u(x0 ), (2.2)
t→0 ∂t
where
Γ (1 − s)
ks = (2.3)
22s−1 Γ (s)
8 M.M. Fall / Nonlinear Analysis 193 (2020) 111311

was found in [12]. For any w ∈ H s (RN ), we denote by H(w) ∈ D1,2 (RN
+
+1 1−2s
;t ) its unique harmonic
N +1
extension over R+ . Namely (see for instance [15,12]),

div(t1−2s ∇H(w)) = 0 in RN +1
⎧
⎪ + ,
⎪
⎨ N
H(w) = w on R , (2.4)
⎪ 1−2s ∂H(w)
⎪
⎩ −t = κs (−∆)s w on RN .
∂t
In particular if w ∈ Cc2 (RN ) then H(w) = P (t, ·) ∗ w. In addition one can check (see [15]), using integration
by parts and the Parseval identity, that
∫ ∫ ∫
2 2s 2
t 1−2s
|∇H(w)| dxdt = κs |ζ| ŵdζ = κs |(−∆)s/2 w| dx. (2.5)
RN
+
+1
RN RN

Therefore from the definition of the space H0s (E), we have

∫
2 2
κs ∥v∥H s (E) = t1−2s |∇H(ṽ)| dxdt ∀v ∈ H0s (E), (2.6)
0
RN
+
+1

where as usual ṽ is the null extension of v outside E. Moreover by (1.4),

∫ ∫
−2s 2
κs γ0 |x| u2 dx ≤ t1−2s |∇H(ṽ)| dxdt ∀v ∈ H0s (E).
E RN
+
+1

We now introduce a Dirichlet-to-Neumann operator Bs defined on H0s (E).

Proposition 2.2. Let E be a bounded open set with Lipschitz boundary. Denote by H −s (E) the dual of
H0s (E). Then the mapping Bs : H0s (E) → H −s (E) given by
∫
∂H(ṽ)
⟨Bs v, φ⟩H −s (E),H s (E) = −κ−1
s lim t1−2s φ̃dx ∀v, φ ∈ H0s (E)
0
RN t→0 ∂t

is a linear isometry. In addition for any v ∈ H0s (E) we have

Bs v = (−∆)s ṽ in D′ (E). (2.7)

Proof . By definition for any v ∈ H0s (E), ṽ ∈ H s (RN ) thus the operator Bs is well defined and linear.
Consider H(ṽ) which satisfies (2.4). Then an integration by parts yields for every φ ∈ H0s (E)
∫ ∫
1−2s ∂H(ṽ)
t ∇H(ṽ) · ∇H(φ̃)dxdt = lim t1−2s φ̃dx.
R N +1
RN t→0 ∂t
+

This, (2.6) and Hölder inequality imply that

2 2
⟨Bs v, φ⟩H −s (E),H s (E) ≤ ∥v∥H s (E) ∥φ∥H s (E)
0 0 0

while
2
⟨Bs v, v⟩H −s (E),H s (E) = ∥v∥H s (E) .
0 0

Hence

∥Bs v∥H −s (E) = ∥v∥H0s (E) .

 M.M. Fall / Nonlinear Analysis 193 (2020) 111311 9

On the other hand, for any φ ∈ Cc∞ (E), we have by integration by parts
∫ ∫
∂H(ṽ)
t1−2s ∇H(ṽ) · ∇H(φ̃)dxdt = lim t1−2s φdx
N +1
R+ R N t→0 ∂t
∫
∂H(φ)
= lim t1−2s ṽdx
RN t→0 ∂t
∫
= κs ṽ(−∆)s φdx.
RN

This means that

Bs v = (−∆)s ṽ in D′ (E). □

We turn to the characterization of the space H0s (E). As suggested with the fact that H(ṽ) ∈
H (RN
1
+
+1 1−2s
;t ) for every v ∈ H0s (E), we have the converse:

Proposition 2.3. Let E be a bounded open set with Lipschitz boundary. Define
{ ⏐ }
1
H0,T (E; t1−2s ) = w ∈ H 1 (RN +1 1−2s
;t ) : w⏐ ≡ 0 on RN \ E
⏐
+
RN

and

s
(E) = u ∈ Ds,2 (RN ) : u ≡ 0 in RN \ E .
{ }
H0,T

We have the following equalities:

{ ⏐ } { ⏐ }
H0s (E) = u⏐ s
: u ∈ H0,T (E) = w⏐ 1
: w ∈ H0,T (E; t1−2s ) . (2.8)
⏐ ⏐
E E

In particular

(−∆)s u = Bs ǔ in D′ (E), s
∀u ∈ H0,T (E),
⏐
where ǔ = u⏐E .

Proof . The first equality in (2.8) is immediate by definition. The second equality is a consequence of the
1
⏐
trace embedding theorem, see [43]. Indeed, take w ∈ H0,T (E; t1−2s ). Then the null extension of w⏐E outside
E is nothing but w which belongs to H s (RN ) and in addition ∥w∥H s (E) ≤ ∥w∥H s (RN ) . □

Summarizing, we state the following

Proposition 2.4. Pick g ∈ H −s (E). Let v ∈ H0s (E) (given by the Lax–Milligram theorem) be the unique
solution to

Bs v = g in E.

Let w ∈ D1,2 (RN

+
+1 1−2s
;t ) solve the mixed problem
1−2s
∇w) = 0 in RN +1
⎧
⎨div(t
⎪ + ,
N
w = 0 on R \ E,
⎩−t1−2s ∂w = κ g on E.
⎪
s
∂t
10 M.M. Fall / Nonlinear Analysis 193 (2020) 111311

Then v = w in E; for any φ ∈ H0s (E)

∫
2s
⟨Bs v, φ⟩H −s (E),H s (E) = |ζ| F(ṽ)F(φ̃)
N
∫R
= (−∆)s/2 ṽ(−∆)s/2 φ̃dx
RN
= ⟨g, φ⟩H −s (E),H s (E)
∫
−1
= κs t1−2s ∇w · ∇H(φ̃)dxdt
RN
+
+1
∫
= κ−1
s t1−2s ∇H(ṽ) · ∇H(φ̃)dxdt
RN
+
+1

and thus
∫
2 2
κs ∥v∥H s (E) = t1−2s |∇w| dxdt.
0
RN
+
+1

We can extend the above in unbounded domains:

Remark 2.5. Here we consider E any open subset of RN with Lipschitz boundary. Define

H s (E) := u ∈ H s (E) : ũ ∈ H s (RN ) ,

{ }
(2.9)
where as usual ũ stands for the null extension of u outside E. We have that Cc∞ (E) is dense in H s (E),
see [32].
By similar arguments, we have that the operator
1−2s ∂H(ṽ)
B¯s (v) = −κ−1
s t +v
∂t
is a linear isometry form H s (E) → (H s (E))′ , where (H s (E))′ is the dual of H s (E).

3. Comparison and maximum principles

Unless otherwise stated, E is a bounded Lipschitz open set of RN . We have the following technical result
which will be useful in the sequel.

Lemma 3.1. Let En be a sequence of Lipschitz open sets such that En ⊂⊂ En+1 and ∪∞
n=1 En = E. Let
gn ∈ L2 (E) such that gn → g in L2 (E). Consider vn ∈ H0s (En ) solution to
Bs vn = gn in En .
If v ∈ H0s (E) is the unique solution to
Bs v = g in E
2
n → v in L (E).
then v˜

n ∈ H0 (E). In addition we have by

1 s
Proof . Observe that H(˜vn ) ∈ H0,T (E; t1−2s ) thus by Proposition 2.3 v˜
Hardy and Hölder inequality
∥˜
vn ∥H0s (E) = ∥vn ∥H0s (E) ≤ C(E)∥gn ∥L2 (E) .

n is bounded. By assumption it converges weakly to v in H0 (E) and strongly in L (E) because

s 2
Therefore v˜
Cc (E) is dense in H0 (E). □
∞ s
 M.M. Fall / Nonlinear Analysis 193 (2020) 111311 11

The following maximum principle can be found in [[16] Lemma 2.4] or in [21].

Lemma 3.2. Let E be a bounded Lipschitz domain of RN . Let v ∈ H0s (E), with v ≥ 0 on RN , such that

Bs v ≥ 0 in E.

If v ̸= 0 then for any compact set K ⊂ E

ess inf v > 0.

K

Lemma 3.3. Let g ∈ L2 (E), g ≥ 0 on E and let w ∈ L1loc (RN

+
+1
), such that
∫
2
t1−2s |∇w| dxdt < ∞
RN
+
+1

and
∫ ∫ ∫
1−2s
t ∇w · ∇ϕdxdt + c wϕdxdt ≥ κs gϕdx (3.1)
RN
+
+1
E E

1
for every nonnegative ϕ ∈ H0,T (E; t1−2s ), where c ≥ 0. Assume that w ≥ 0 on RN \ E. Then w ≥ 0 in
RN
+
+1
.

1
Proof . Test (3.1) with max(−w, 0) ∈ H0,T (E; t1−2s ). □

Lemma 3.4. Let c ≥ 0 be a constant and let u ∈ L1s , u ≥ 0 on RN and g ∈ L2 (E) such that

(−∆)s u + cu ≥ g in D′ (E). (3.2)

Let v ∈ H0s (E) solves

Bs v + cv = g in E. (3.3)

Then

u≥v in E.

Proof . Recall that (3.2) is equivalent to

∫ ∫ ∫
u(−∆)s φdx ≥ gφdx − c uφdx ∀φ ∈ Cc∞ (E), φ ≥ 0. (3.4)
RN E E

Denote by ρn the standard mollifier (which is symmetric: ρn (−x) = ρn (x)) and put un = ρn ∗ u.
Claim: for any φ ∈ Cc∞ (RN )
∫ ∫
(−∆)s un φ = u(−∆)s (ρn ∗ φ). (3.5)
RN RN

It is easy to check using Fubini’s theorem and the symmetry of ρn that

∫ ∫ ∫
(−∆)s un φdx = un (−∆)s φdx = uρn ∗ (−∆)s φdx. (3.6)
RN RN RN
12 M.M. Fall / Nonlinear Analysis 193 (2020) 111311

Now we notice that, in RN ,

2s
ρn ∗ (−∆)s φ = F(F(ρn ∗ (−∆)s φ)) = F(|ζ| (F(ρn )F(φ))) = (−∆)s (ρn ∗ φ).

Using this in (3.6), we get (3.5) as claimed.

Let En := {x ∈ E : dist(x, ∂E) > 1/n}. We deduce from (3.4) and (3.5) that for all φ ∈ Cc∞ (En ) and
φ≥0
∫ ∫ ∫ ∫
s s
(−∆) un φdx = u(−∆) (ρn ∗ φ)dx ≥ g(ρn ∗ φ)dx − c u(ρn ∗ φ)dx
RN RN ∫E ∫E
= (ρn ∗ g)φdx − c (ρn ∗ u)φdx.
E E
We conclude that

(−∆)s un (x) + cun (x) ≥ ρn ∗ g(x) =: gn (x) for every x ∈ En . (3.7)

We let wn (t, x) = P (t, ·) ∗ un (x) be the harmonic extension of un via the Poisson kernel so that
{
div(t1−2s ∇wn ) = 0 RN+
+1
,
(3.8)
wn = un RN .
It turns out that
∂wn
− t1−2s + cwn = κs (−∆)s un + cun ≥ κs gn on En (3.9)
∂t
2
and in addition t1−2s |∇wn | ∈ L1loc (RN
+
+1
). Let vn ∈ H0s (En ) be the solution to

Bs vn + cvn = gn in En .
1,2 N +1
We take a large R > 0 so that B N (0, R) contains E and we let vn,R ∈ W0,S (B+ (0, R); t1−2s ) be the
unique solution (obtained by minimization) to the problem
N +1
div(t1−2s ∇vn,R ) = 0 B+
⎧
⎪
⎪ (0, R),
∂v
⎨
n,R
−t1−2s + cvn,R = κs gn En , (3.10)
⎪
⎪ ∂t
vn,R = 0 B N (0, R) \ En .
⎩

N +1
By extending vn,R to be zero outside B+ (0, R), it is standard to show that vn,R → vn as R → ∞ in
N +1 1−2s
1
H (R+ ; t ). Since wn ≥ vn,R by Lemma 3.3, it follows that, sending R → ∞, wn ≥ vn in RN . In
2
particular un ≥ vn in En . By Lemma 3.1, v˜
n → v in L (E) and the proof is complete. □

We recall the definition of the s-capacity of a compact set A ⊂ E:

2
Cs (A) = inf {∥ϕ∥H s (E) : φ ≥ 1 in a neighborhood of A}. (3.11)
ϕ∈Cc∞ (E) 0

Note that if Cs (A) = 0 then |A| = 0 by Poincaré inequality (see (1.10)). We have the following comparison
result modulo small sets.

Lemma 3.5. Let A be a compact subset of E with Cs (A) = 0. Let u ∈ L1s , c ≥ 0 and g ∈ L2 (E) such that

(−∆)s u + cu ≥ g in D′ (E \ A).

Let v ∈ H0s (E) solve

Bs v + cv = g in E.

Then u ≥ v in E.
 M.M. Fall / Nonlinear Analysis 193 (2020) 111311 13

Proof . Let Aε be a smooth open ε-neighborhood of A compactly contained in E. Define Dε = {x ∈ D :

dist(x, ∂(E \ Aε )) > ε}. It is clear that

(−∆)s u + cu ≥ g in D′ (Dε ).

Consider vε ∈ H0s (Dε ) solving

Bs vε + cvε = g in Dε .

By Lemma 3.4 we have u ≥ vε in Dε . The same argument as in the proof of Lemma 3.1 yields v˜ε ∈ H0s (E)
for every s ∈ (0, 1) and it is bounded. Hence it converges weakly to some function w in H0s (E) and strongly
in L2 (E). In particular u ≥ w. Moreover for any φ ∈ Cc∞ (E \ A), we can choose ε > 0 so small that suppφ
is contained in Dε thus taking the limit as ε → 0, we get
∫ ∫
⟨w, φ⟩H0s (E) + c wφ = gφdx ∀φ ∈ Cc∞ (E \ A).
E E

From this equality, to conclude the proof (that is v = w), it suffices to show that Cc∞ (E \ A) is dense in
Cc∞ (E) with the H0s (E)-norm because w ∈ H0s (E).
Since Cs (A) = 0, there exists a sequence ψn ∈ Cc∞ (E) such that ψn ≥ 1 in a neighborhood of A and in
addition
2
∥χn ∥H s (E) ≤ ∥ψn ∥H0s (E) → 0, (3.12)
0

where χn = min(ψn , 1). Now take any ϕ ∈ Cc∞ (E) and note that (1 − χn )ϕ ∈ Cc∞ (E \ A) and moreover
(1 − χn )ϕ → ϕ in H0s (E) by (3.12). This concludes the proof. □

We shall define a new space which is more convenient when dealing with the Hardy potential. Namely,
we assume that there exists b ∈ L1loc (E) and a constant C > 0 such that
∫ ∫
2
∥φ∥H s (E) − b(x)φ2 dx ≥ C φ2 dx ∀φ ∈ Cc∞ (E). (3.13)
0
E E

Definition 3.6. Let b ∈ L1loc (E) so that (3.13) holds. The Hilbert space H0,b
s
(E) is the completion of
∞
Cc (E) with respect to the scalar product
∫
⟨φ, ϕ⟩H0s (E) − b(x)φϕdx ∀φ, ϕ ∈ Cc∞ (E).
E

Note that the Lax–Milligram theorem implies that for any f ∈ L2 (E), there exists a unique solution to
the problem
{
Bs v − b(x)v = f in E,
(3.14)
v ∈ H0,b
s
(E),

in the sense that for all ϕ ∈ H0,b

s
(E)
∫ ∫
⟨v, ϕ⟩H0s (E) − b(x)vϕdx = f ϕdx.
E E

Remark 3.7. Let ε > 0. Put dε (x) = b(x)(1 − ε). Then H0s (E) = H0,ds
ε
(E) by Proposition 2.3. This holds
true because if v ∈ H0,dε (E) then by (3.13) we have ṽ ∈ H (R ). By similar argument H0s (E) = H0,b
s s N s
(E)
∞
if b ∈ L (E).
14 M.M. Fall / Nonlinear Analysis 193 (2020) 111311

Lemma 3.8. Let A be a compact subset of E with Cs (A) = 0. Let b ∈ L1loc (E) such that (3.13) holds.
Suppose that u ∈ L1s with u, b ≥ 0 and f ∈ L2 (E), f ≥ 0 such that

(−∆)s u − b(x)u ≥ f in D′ (E \ A). (3.15)

Let v ∈ H0,b
s
(E) be the unique solution to

Bs v − b(x)v = f in E.

Then

u≥v in E.

Proof . Step 1:We first prove the result if b ∈ L∞ (E).

We let v0 ∈ H0s (E) solving

Bs v0 = f in E.

Then 0 ≤ v0 ≤ u in E by Lemma 3.5 and because f ≥ 0. We define inductively the sequence vn ∈ H0s (E) by

Bs v1 = b(x)v0 + f in E, Bs vn = b(x)vn−1 + f in E.

Since b ≥ 0, we have (−∆)s u ≥ b(x)v0 + f in D′ (E \ A). Thus using once again Lemma 3.5, we obtain
v0 ≤ v1 ≤ u in E. By induction, we have

v0 ≤ v1 ≤ · · · ≤ vn ≤ u in E ∀n ∈ N.

Since vn−1 ≤ vn in E, we have

∫ ∫
2 2
∥vn ∥H s (E) − b(x)|vn | ≤ f (x)vn dx.
0
E E

By Hölder inequality and (3.13) (see Remark 3.7) vn is bounded in H0s (E). We conclude that vn ⇀ v in
H0s (E) as n → ∞ which is the unique solution to

Bs v = b(x)v + f in E.

Since vn → v in L2 (E), we get v ≤ u in E.

Step 2: Conclusion of the proof.
We put bk (x) = min(b(x), k) for every k ∈ N. We consider v k ∈ H0s (E) be the unique solution to
∫ ∫
⟨v k , φ⟩H0s (E) − min {b(x), k} v k φ = f φ ∀φ ∈ Cc∞ (E). (3.16)
E E

Since bk ∈ L∞ (E), thanks to Step 1, we have v k ≤ u in E.

Next, we check that such a sequence v k , satisfying (3.16), converges to v in L2 (E) when k → ∞. Indeed,
we have
∫
2 2 2
∥v k ∥H s (E) ≤ ∥v k ∥H s (E) − min{b(x), k} |v k | dx
0,b 0
∫ E

= f v dx ≤ C∥v k ∥H s (E)
k
0,b
E
 M.M. Fall / Nonlinear Analysis 193 (2020) 111311 15

by Hölder inequality and by (3.13), where the constant C depends on f and E but not on k. Therefore the
sequence v k is bounded in H0,b
s
(E). We conclude that there exists ṽ ∈ H0,b
s
(E) such that, for a subsequence,
v ⇀ v in H0,b (E). Now by (3.16), we have
k s

∫ ∫
k k
⟨v , φ⟩H s (E) + (b(x) − min{b(x), k}) v φ = f φ.
0,b
E E

Since for every k ≥ 1 and any φ ∈ Cc∞ (E)

⏐(b(x) − min{b(x), k}) v k φ⏐ ≤ (b(x) − min{b(x), k}) u|φ| ≤ 2b(x)u|φ| ∈ L1 (E),
⏐ ⏐

the dominated convergence theorem implies that

∫
⟨v, φ⟩H s (E) = f φ for any φ ∈ Cc∞ (E). (3.17)
0,b
E

We therefore have that v = v by uniqueness. By (3.17), we have

2 2
∥v − v k ∥H s (E) = ∥v k ∥H s − ⟨v, v k ⟩H s (E) + ⟨v, v − v k ⟩H s (E)
(E)
0,b 0,b 0,b 0,b
∫
2
= ∥v k ∥H s (E) − f v k + ⟨v, v − v k ⟩H s (E)
0,b 0,b
∫ E ∫
k 2 k 2
≤ ∥v ∥H s (E) − min{b(x), k} |v | dx − f v k + ⟨v, v − v k ⟩H s (E)
0 0,b
E E
= ⟨v, v − v k ⟩H s (E) .
0,b

We thus obtain
∫
2
C(E) |v − v k | dx ≤ ⟨v, v − v k ⟩H s (E) →0
0,b
Ω

by (3.13). Hence v k → v pointwise and thus v ≤ u in E. □

Remark 3.9. The same result as in Lemma 3.8 holds if we assumed the coercivity that there exist constants
C, c > 0 such that for all φ ∈ Cc∞ (E)
∫ ∫ ∫
2 2 2
∥φ∥H s (E) + c φ dx − b(x)φ dx ≥ C φ2 dx. (3.18)
0
E E E

We close this section with the following useful lemma and its immediate consequence. Its counterpart,
for s = 1, is in [25].

Lemma 3.10. Let E be a bounded Lipschitz domain of RN . Let A be a compact subset of E with Cs (A) = 0.
Let u ∈ L1s with u ≩ 0 on RN and b ∈ L1loc (E) with b > 0 on E. Assume that

(−∆)s u ≥ b(x)u in D′ (E \ A). (3.19)

Then
∫ ∫
2 2
t1−2s |∇H(φ)| dxdt = κs ∥φ∥H s (E) ≥ κs b(x)φ2 dx ∀φ ∈ Cc∞ (E). (3.20)
RN
+
+1 0
E

Proof . Put gk (x) := min(b(x)u, k) for integers k ≥ 1. We have that gk > 0 on E. Let vk ∈ H0s (E) be the
unique solution to

Bs vk = gk in E.
16 M.M. Fall / Nonlinear Analysis 193 (2020) 111311

By Lemma 3.2, we have 1

vk ∈ L∞
loc (E) and by the standard maximum principle H(v˜k ) > 0. Moreover by
Lemma 3.5, we have

u ≥ vk > 0 in E. (3.21)
H(φ)
Let φ ∈ Cc∞ (E). Put Vk = H(v˜k ) and Vkε = Vk + ε, for ε > 0. Set ψ = Vkε so that Vkε ψ 2 ∈ H0,T
1
(E; t1−2s ).
Simple computations show that
2 2 2
|∇H(φ)| = |Vkε ∇ψ| + ∇Vkε · ∇(Vkε ψ 2 ) = |Vkε ∇ψ| + ∇Vk · ∇(Vkε ψ 2 ).

Thus using integration by parts we have

∫ ∫
2
t1−2s |∇H(φ)| dxdt ≥ t1−2s ∇Vk · ∇(Vkε ψ 2 )dxdt
RN
+
+1
RN
+
+1

φ2
∫
= gk dx.
E (vk + ε)2
Take the limit as ε → 0 to get
∫ ∫
2 gk 2
t1−2s |∇H(φ)| dxdt ≥ κs φ dx
RN +1
E vk
+

by Fatou’s lemma. By (3.21), we infer that

∫ ∫
2 gk 2
t1−2s |∇H(φ)| dxdt ≥ κs φ dx.
RN +1
+
E u

Again by Fatou’s lemma, inequality (3.20) follows immediately by taking k → +∞. □

The following result appeared in [3] in the case s = 1.

Theorem 3.11. Let E be a bounded Lipschitz domain of RN with 0 ∈ E, N > 2s. Then there is no
nonnegative and nontrivial u ∈ L1s satisfying
−2s
(−∆)s u ≥ γ|x| u in D′ (E \ {0}),

with γ > γ0 .

Proof . Note that Cs ({0}) = 0 provided N > 2s (see [42, p. 397]). If such u exists then u > 0 in E by the
maximum principle thus Lemma 3.10 implies that
2
∥φ∥H s (E)
0
inf 2 ≥ γ.
φ∈Cc∞ (E) ∥φ∥L2 (E)
∥φ∥2
H s (E)
0
This contradicts the optimality of the Hardy constant γ0 := inf φ∈Cc∞ (E) ∥φ∥2 2
. □
L (E)

4. Nonexistence of positive supersolutions

We start with the following

2s−N
+α
Lemma 4.1. For every α ∈ (− N2 − s, N2 − s), put ϑα (x) = |x| 2 . Then
−2s
(−∆)s ϑα = γα |x| ϑα in RN \ {0},
 M.M. Fall / Nonlinear Analysis 193 (2020) 111311 17

where
( N +2s+2α ) ( N +2s−2α )
2s Γ 4 Γ 4
γα = 2 ( N −2s−2α ) ( N −2s+2α ). (4.1)
Γ 4 Γ 4

For α ≥ 0, the function α ↦→ γα is continuous

( and decreasing.
)
There exists a positive function Υα ∈ C β RN+
+1
\ {0} such that

div(t1−2s ∇Υα ) = 0 in RN +1
⎧
⎪
⎪ +
Υα = ϑα on ∂RN +1
⎨
+ \ {0} (4.2)
⎪ ∂Υ α −2s
⎩−t1−2s = κs (−∆)s ϑα = κs γα |x| ∂RN +1
\ {0}.
⎪
ϑα on +
∂t
N +1
Moreover if α > 0 then |∇Υα | ∈ L2 (B+ (0, R); t1−2s ) for every R > 0.

Proof . Note that ϑα ∈ L1s . The Fourier transform of radial functions (see [[48] Theorem 4.1]) yields
∫ ∞
1−N 1 N −1
F(ϑα )(ρ) = ρ 2 (rρ) 2 J N −2 (rρ)ϑα r 2 dr,
0 2

where J N −2 is the Bessel function. Then we have

2
∫ ∞
−N
F(ϑα )(ρ) = ρ 2 −s−α (rρ)s+α J N −2 (rρ)d(rρ)
0 2
−N
= mα ρ 2 −s−α ,

where
( N +2s+2α )
s+α Γ 4
mα = 2 ( N −2s−2α ).
Γ 4

Now we notice that (γα = mα m−α )

−N
(−∆)s ϑα = F(F((−∆)s ϑα )) = F(ρ2s F(ϑα )(ρ)) = mα F(ρ 2 +s−α ) = γα r−2s ϑα .

For the proof of the fact that the map α ↦→ γα is continuous and decreasing, we refer to [18].
We define
P (t, ·) ∗ ϑα (t, x) ∀(t, x) ∈ RN +1
{
+
Υα (t, x) = N +1
ϑα (x) ∀x ∈ ∂R+ \ {0},
where P is the Poisson kernel defined in Section 2.1. Clearly Υα is positive. We have that
∂Υα
− t1−2s = κs (−∆)s ϑα in RN \ {0}.
∂t
Hence we get (4.2). ( )
From the regularity theory of [12], we deduce that Υα ∈ C β RN +
+1
\ {0} for some β > 0. In addition
( )
Υα ∈ H 1 Ω × (t1 , t2 ); t1−2s for every Ω ⊂⊂ RN \ {0} and 0 < t1 < t2 < ∞.
2s−N −1
Observe that Υα (λz) = λ 2 +α Υα (z) and thus choosing λ = |z| , we infer that Υα (z) ≤
2s−N 2s−N
−1 +α +α N +1
Υα (z|z| )|z| 2 ≤ C|z| 2 , for every z = (t, x) ∈ B+ (0, R) and R > 0. From this, we deduce
−2 2 N +1 −2s
that t 1−2s
|z| Υα ∈ L1 (B+ (0, R)) for α > 0. We also have |x| Υα2 ∈ L1loc (RN ) for α > 0.
We let φ be a cut-off function such that φ = 0 for |z| < ε, φ = 1 for 2ε < |z| < R, φ = 0 for |z| > 2R and
|∇φ| ≤ Cε−1 for ε < |z| < 2ε. We use φ2 Υα as a test function in (4.2) to get
∫ ∫
−2s
t1−2s ∇Υα ∇(φ2 Υα ) = |x| Υα2 φ2 .
RN
+
+1
∂RN
+
+1
18 M.M. Fall / Nonlinear Analysis 193 (2020) 111311

Integrating by parts and using Young’s inequality, for some constant c > 0 depending only on N, s and R,
we have
∫ ∫ ∫
2 −2s 2
c t1−2s φ2 |∇Υα | ≤ |x| Υα2 φ2 + t1−2s Υα2 |∇φ| .
N +1 N +1
B+ (0,R)\B+ (0,ε) B N (0,2R) RN
+
+1

Therefore
∫ ∫ ∫
1−2s 2 −2s −2
c t |∇Υα | ≤ |x| Υα2 + t1−2s |(t, x)| Υα2 .
N +1 N +1
B+ (0,R)\B+ (0,ε) B N (0,2R) ε<|(t,x)|<2ε

N +1
Letting ε → 0, Fatou’s lemma yields |∇Υα | ∈ L2 (B+ (0, R); t1−2s ) for α > 0. □

The comparison result obtained in Lemma 3.8 allows us to derive the following estimate when the potential
b(x) is the Hardy one.

Lemma 4.2. Let N > 2s, α ∈ [0, (N − 2s)/2) and p > 1. Suppose that u ∈ L1s ∩ Lploc (B N (0, 2) \ {0}), u ≩ 0
on RN such that
−2s
(−∆)s u − γα |x| u ≥ up in D′ (B N (0, 2) \ {0}). (4.3)

Then there exists a constant C > 0 such that

∫ ∫
2 −2s
∥φ∥H s (B N (0,2)) − γα |x| φ2 dx ≥ C φ2 dx ∀φ ∈ Cc∞ (B N (0, 2)). (4.4)
0
B N (0,2) B N (0,2)

Moreover there exists a constant C ′ > 0 such that

2s−N
+α
u ≥ vα ≥ C ′ |x| 2 in B N (0, 1), (4.5)
−2s
where vα ∈ H0,b
s
(B N (0, 2)) (with b(x) = γα |x| ) is the solution to
−2s
Bs vα − γα |x| vα = min(up , 1) in B N (0, 2).

Proof . Inequality (4.4) is trivial for α > 0 so we consider only the case α = 0. Since Cs ({0}) = 0 provided
N > 2s, by Lemmas 3.2 and 3.5, we have u > 0 in B N (0, 2) and

M = ess inf u > 0.

B N (0,1)

Hence u ∈ L1s satisfies

−2s
(−∆)s u − γα |x| u ≥ M p−1 u in D′ (B N (0, 1) \ {0}).

By Lemma 3.10 we obtain (4.4) thanks to the scale invariance of the integrals in the left hand side.
−2s
We put b(x) = γα |x| . By (4.4), for α ≥ 0, we can let vα ∈ H0,b
s
(B N (0, 2)) be the solution to
−2s
Bs vα − γα |x| vα = min(up , 1) in B N (0, 2).

Then by Lemma 3.8, we have u ≥ vα in B N (0, 2). We first consider the case α > 0. Then γα < γ0 and thus
vα ∈ H0s (B N (0, 2)). By the regularity result of [12], we get that Vα = H(˜ N +1
vα ) is continuous in B+ (0, 1)\{0}
and Vα ≥ 0 by Lemma 3.3. Consider ϑα and its harmonic extension Υα given by Lemma 4.1. The maximum
principal (see Lemma 3.2) implies that we can set
min N vα
S+
′
C = > 0. (4.6)
max N Υα
S+
 M.M. Fall / Nonlinear Analysis 193 (2020) 111311 19

Put w = C ′ Υα − Vα . We have that

⎧
N +1
⎪
⎪ div(t1−2s ∇w) = 0 in B+ (0, 1),
⎪
⎨
w≤0 on S+N,
(4.7)
⎪ 1−2s ∂w
⎪
−2s
⎩−t
⎪ − κs γα |x| w ≤ 0 on B N (0, 1).
∂t
N +1
Then w+ := max(w, 0) ∈ H0,S
1
(B+ (0, 1); t1−2s ) and therefore by integration by parts
∫ ∫
1−2s + 2 −2s
t |∇w | dxdt − κs γα |x| (w+ )2 dx ≤ 0.
RN
+
+1
B N (0,1)

In particular
∫
2 −2s
∥w+ ∥H s (B N (0,1)) − γα |x| (w+ )2 dx ≤ 0.
0
B N (0,1)

Hence w+ ≡ 0 by Hardy’s inequality. Hence vα ≥ C ′ ϑα in B N (0, 1) that is (4.5) for α > 0.

For the case α = 0, we put αn = 1/n and we notice that the sequence vαn ∈ H0s (B N (0, 2)) solution to
the problem

−2s
Bs vαn − γαn |x| vαn = min(up , 1) in B N (0, 2)

is monotone increasing to v0 because the mapping α ↦→ γα is decreasing. Therefore, taking into account
(4.6), we readily get (4.5). □

Proof of Theorem 1.2. If u ̸= 0 on E, it then follows from Lemma 4.2 that there exist r, Cr > 0 such
that
2s−N
+α
u(x) ≥ vα (x) ≥ Cr |x| 2 ∀x ∈ B N (0, r) ⊂ E, (4.8)

−2s
where vα ∈ H0,b
s
(E) (with b(x) = γα |x| ) is the solution to

−2s
Bs vα − γα |x| vα = min(up , 1) in E.

−2s
On the other hand, thanks to (4.8), by applying Lemma 3.10 with b(x) = γα |x| u + up−1 , we get for all
φ ∈ Cc∞ (B N (0, r)),
∫ ∫
2 −2s
∥φ∥H s (B N (0,r)) − γα |x| φ2 dx ≥ up−1 φ2 dx. (4.9)
0
E B N (0,r)

We first consider the case α > 0. If r is small, by (4.8) we have, for 0 < α′ < α,

−2s ) −2s
(−∆)s u − γα′ |x| u ≥ −γα′ + γα + Crp−1 |x| u in D′ (B N (0, r) \ {0}).
(

By Lemma 3.8 and using the same arguments as in Lemma 4.2 we get, provided α′ ↗ α,
2s−N
+α′
u(x) ≥ Cr′ |x| 2 ∀x ∈ B N (0, r/2), (4.10)
20 M.M. Fall / Nonlinear Analysis 193 (2020) 111311

for some constant Cr′ > 0. Using the estimate (4.10) in (4.9), we get
∫ ∫
2s−N
+α′ (p−1)
|x|( )
2 −2s 2 ′ p−1
∥φ∥H s (B N (0,r/2)) − γα |x| φ dx ≥ (Cr ) 2 φ2 dx,
0
B N (0,r/2) B N (0,r/2)

N +2s−2α′
for all φ ∈ Cc∞ (B N (0, r/2)). Since p > N −2s−2α′ , we have
( )
2s − N ′
− δ := + α (p − 1) + 2s < 0.
2
Hence for every ρ ∈ (0, r/2)
∫
2 −2s
∥φ∥H s (B N (0,ρ)) ≥ (γα + (Cr′ )p−1 ρ−δ ) |x| φ2 dx ∀φ ∈ Cc∞ (B N (0, ρ)).
0
B N (0,ρ)

This contradicts the sharpness of the Hardy constant thanks to the scale invariance of the inequality.
Finally, for the case α = 0 we note that (4.9) implies, by density, that
∫ ∫
2 −2s
∥vα ∥H s (E) − γα |x| vα2 dx ≥ vαp+1 dx.
0
E E

This also leads to a contradiction because vα ∈ H0,b

s
(E)
while by (4.8)
∫ ∫
2s−N
|x|( 2 )
(p+1)
vαp+1 dx ≥ C ′ dx
E B N (0,r)
∫ ∫ r
≥ C′ t−1 dtdσ = +∞,
S N −1 0
′
for some constant C > 0. □

5. Existence of positive solutions and proof of Theorem 1.1

The proof will be separated into several cases. We put

∫
2 −2s
Eα (u) := ∥u∥H s (B) − γα |x| u2 dx ∀u ∈ Cc∞ (B),
0
B

where B is a ball in RN centered at 0 with N > 2s.

Case 1: α ∈ (0, (N − 2s)/2] and 1 < p < (N + s)/(N − 2s).
Thanks to the Hardy–Littlewood–Sobolev inequality (see e.g. [39]) we have
(∫ )2/(p+1)
Eα (u) ≥ up+1 ∀u ∈ Cc∞ (B). (5.1)
B

Thanks to the compact embedding of H0s (B) into Lp+1 (B), we can minimize Eα over the set
{ ∫ }
u ∈ H0 (B) :
s + p+1
(u ) =1 . (5.2)
B

Let u ∈ H0s (B) be the minimizer. Put u± = max(±u, 0). By Proposition 2.3, u± belongs to H0s (B). We
check rapidly that Eα (u+ ) ≤ Eα (u). Observe that
∫ ∫
2 2 2
κs ∥u± ∥H s (B) = t1−2s |∇H(u± )| dxdt ≤ t1−2s |∇(H(u)± )| dxdt
0
RN
+
+1
RN
+
+1
 M.M. Fall / Nonlinear Analysis 193 (2020) 111311 21

because H(u± ) and H(u)± have the same trace on RN while H(u± ) has minimal Dirichlet energy. Now
using this and Hardy’s inequality we have
∫
+ + 2 −2s
κs Eα (u ) = κs ∥u ∥H s (B) − κs γα |x| (u+ )2 dx
0
∫B ∫
2 −2s −2s
= κs ∥u+ ∥H s (B) − κs γα |x| u2 dx + κs γα |x| (u− )2 dx
0
∫B B
+ 2 −2s 2 − 2
≤ κs ∥u ∥H s (B) − κs γα |x| u dx + κs ∥u ∥H s (B)
0 0
∫ B ∫
2 −2s
≤ t1−2s |∇(H(u)+ )| dxdt − κs γα |x| u2 dx
RN
+
+1
B
∫
2
+ t1−2s |∇(H(u)− )| dxdt
RN
+
+1
∫ ∫
1−2s 2 −2s 2
= t |∇H(u)| dxdt − κs γα |x| u dx
RN
+
+1
B

= κs Eα (u).
Thus we may assume that u = u+ is a nonnegative and nontrivial minimizer therefore there exists a Lagrange
multiplier λ > 0 such that
−2s
Bs u − γα |x| u = λup in B.
1
Hence λ p−1 ũ is a solution of problem (1.1).
Case 2: α = 0 and 1 < p < (N + 2s)/(N
( −(2s). ))
2
Theorem 1.3 yields for every q ∈ 2, max 1, 1+2s
2
E0 (u) ≥ ∥u∥W τ,q (B) u ∈ Cc∞ (B),
0
−2s
with τ = 1+2s
2 − 1q > 0. Therefore H0,bs
(B) is compactly embedded into Lp+1 (B), with b(x) = γ0 |x| .
Hence we can minimize E0 over the set
{ ∫ }
u ∈ H0,b (B) :
s + p+1
(u ) =1 . (5.3)
B
+
We have to check again that E0 (u ) ≤ E0 (u). But this can be done by density and using similar arguments
as above. We skip the details. We get a positive minimizer u = u+ of E0 in the set (5.3). We conclude that
1
λ p−1 ũ is a solution to (1.1) for some Lagrange multiplier λ > 0.
Case 3: α ∈ (0, (N − 2s)/2) and (N + 2s)/(N − 2s) ≤ p < (N + 2s − 2α)/(N − 2s − 2α).
2s−N
+β
Consider ϑβ (x) = |x| 2 given by Lemma 4.1 which satisfies
−2s
(−∆)s ϑβ = γβ |x| ϑβ in RN \ {0}.
−2s
We look for a solution of the form w = µr p−1 with a constant µ > 0 to be determined in a minute. Assume
2s−N −2s
that we can take β ≥ 0 such that r 2 +β = r p−1 then
−2s
(−∆)s w = γβ |x| w + wp − wp−1 w
−2s −2s
= γα |x| w + wp + (γβ − γα − µp−1 )|x| w.
Since β ↦→ γβ is decreasing, we can choose µp−1 = γβ − γα > 0 provided α > β. But note that α > β as soon
as p < (N + 2s − 2α)/(N − 2s − 2α) and the fact that p ≥ (N + 2s)/(N − 2s) implies β ≥ 0. In conclusion
we have, in RN \ {0},
−2s
(−∆)s w − γα |x| w = wp
and w ∈ L1s ∩ Lp (B) is a solution to (1.1). The proof of Theorem 1.1 is thus completed. □
22 M.M. Fall / Nonlinear Analysis 193 (2020) 111311

Acknowledgments

This work is supported by the Alexander-von-Humboldt Foundation. The author would like to thank
Professor Tobias Weth for fruitful discussions and the referee for his/her detailed and valuable comments.

Appendix. Proof of Theorem 1.3

A.1. Remainder term for the fractional hardy inequality

Let E be a bounded open set of RN , N > 2s, with 0 ∈ E. The following (local) Hardy inequality is a
consequence of (1.4)
∫
−2s 2
γ0 u2 |x| dx ≤ ∥u∥H s (E) ∀u ∈ Cc∞ (E). (A.1)
0
E

In addition the constant γ0 is optimal. Our objective, in this section, is to improve inequality (1.4) in bounded
domains of RN .
Many deal of work has been done in improving the classical Hardy inequality starting from the work of
Brezis–Vázquez [11]. We also quote [4,50,31] for related improvements.
We shall prove a Vázquez–Zuazua-type
( ) (see [50]) improvement for the fractional Hardy inequality (1.4).
That is for 2 > q > max 1, 1+2s , there exists a constant C(E) > 0 such that for all u ∈ Cc∞ (E),
2

∫
2 2 −2s 2
C(E)∥u∥W τ,q (E) ≤ ∥u∥H s (E) − γ0 |x| u dx,
0 0
E

where τ = 1+2s 1
2 − q . The proof requires several preliminary lemmata.
Consider the function Υ0 defined in Lemma 4.1 satisfying

div(t1−2s Υ0 ) = 0 in RN +1
⎧
⎪
⎪ + ,
⎪
⎨ 2s−N
N
Υ0 = |x| 2 on R \ {0}, (A.2)
⎩−t1−2s ∂Υ0 = κs γ0 |x|−2s Υ0 on RN \ {0}.
⎪
⎪
⎪
∂t
We have seen in the proof of Lemma 4.1 that
(2s−N )/2 N +1
|Υ0 (z)| ≤ C|z| ∀z = (x, t) ∈ B+ . (A.3)

By scale invariance, we have that

Υ0 (z) = R(N −2s)/2 Υ0 (Rz) ∀R > 0. (A.4)

This implies the estimate

(2s−N )/2 N +1
|Υ0 (z)| ≥ C|z| ∀z ∈ B+ (A.5)

and also
(2s−N )/2−1 N +1
|∇Υ0 (z)| ≤ C|z| ∀z ∈ B+ . (A.6)

We now prove the following result which were proved in [18] when s = 1/2.
 M.M. Fall / Nonlinear Analysis 193 (2020) 111311 23

Lemma A.1. For every q ∈ (1, 2) there exists a constant C > 0 such that for all φ ∈ Cc∞ (B N +1 )
(∫ )2/q ∫ ∫
q q 2 −2s
C t 2 |∇φ| dz ≤ t1−2s |∇φ| dz − κs γ0 |x| φ2 dx. (A.7)
N +1 N +1
B+ B+ BN

φ
Proof . Let φ ∈ Cc∞ (B N +1 \ {0}) and put ψ = Υ0 . Simple computations yield
2 2
|∇φ| = |Υ0 ∇ψ| + ∇Υ0 · ∇(Υ0 ψ 2 ).

Integration by parts and using (A.2) leads to

∫ ∫ ∫
2 −2s 2
t1−2s |∇φ| dz − κs γ0 |x| φ2 dx ≥ t1−2s Υ02 |∇ψ| dz.
N +1 N +1
B+ BN B+

z
By (A.5) and using polar coordinates z = rσ = |z| |z| , we get
∫ ∫ 1 ∫
2 2
t1−2s Υ02 |∇ψ| dz ≥ C (σ1 )1−2s r|∇ψ| dσdr, (A.8)
N +1 N
B+ 0 S+

where σ1 is the component of σ in the t direction. We wish to show that there exists a constant C > 0 such
that
(∫ ∫ )q/2
1 ∫
q(1−2s)
2 q
I := (σ1 ) 1−2s
r|∇ψ| dσdr ≥C t 2 |∇φ| dz ∀φ ∈ Cc∞ (B N +1 ). (A.9)
N N +1
0 S+ B+

We have
∫ ∫ ∫
q(1−2s) q q(1−2s) q q(1−2s) q q
t 2 |∇φ| dz = t 2 |∇ψΥ0 + ψ∇Υ0 | dz ≤ C t 2 (|∇ψΥ0 | + |ψ∇Υ0 | ) dz.
N +1 N +1 N +1
B+ B+ B+

Put
∫
q(1−2s) q
I1 = t 2 |∇ψΥ0 | dz,
N +1
B+
∫
q(1−2s) q
I2 = t 2 |ψ∇Υ0 | dz.
N +1
B+

Using (A.3) and Hölder inequality, we have

∫ ∫ 1
q(1−2s)/2 q
I1 ≤ C (σ1 ) rN +q(1−N )/2 |∇ψ| drdσ
N
S+ 0
∫ 1 ∫
N +q(1−N )/2−q/2 q
=C r (σ1 )q(1−2s)/2 rq/2 |∇ψ| dσdr (A.10)
0 N
S+
(∫ )q/2
∫ 1
2
≤C rN (2−q)/2 (σ1 )(1−2s) r|∇ψ| dσ dr
0 N
S+
(∫ )q/2
∫ 1
2
≤C (σ1 )(1−2s) r|∇ψ| dσ dr
0 N
S+
(∫ )q/2
1 ∫
2
≤C (σ1 )(1−2s) r|∇ψ| dσdr
0 N
S+
≤ CI.
24 M.M. Fall / Nonlinear Analysis 193 (2020) 111311

On the other hand we have using (A.6)

∫ ∫ 1
q
I2 ≤ C (σ1 )q(1−2s)/2 rN (2−q)/2−q/2 |ψ| drdσ
N
S+ 0
1 ⏐ ∂ψ ⏐q
∫ ∫ ⏐ ⏐
≤C (σ1 )q(1−2s)/2rN (2−q)/2−q/2+q ⏐⏐ ⏐ drdσ
N
S+ 0 ∂r ⏐
∫ 1 ∫
q
≤C rN +q(1−N )/2−q/2 (σ1 )q(1−2s)/2 rq/2 |∇ψ| dσdr (A.11)
0 N
S+
≤ CI,
∫1 ∫1 q
where in the second inequality we have used the one dimensional Hardy inequality 0 f q dr ≤ c 0 r−q |f ′ | dr
and observing that (A.10) is just (A.11). The lemma follows because Cc∞ (B N +1 \{0}) is dense in Cc∞ (B N +1 )
with respect to the H 1 (B N +1 ; ta )-norm when N ≥ 3, see [34]. □

The Lion’s interpolation inequality, [[40] Paragraph 5], shows that for a ̸= 0 and − 1q < a2 < 1
q there exists
a constant C > 0 such that
∫ ∫
q aq q aq q
C∥v∥W τ,q (RN ) ≤ t 2 |∇v| dxdt + t 2 |v| dxdt, ∀v ∈ Cc∞ (RN +1 ), (A.12)
RN
+
+1
RN
+
+1

where τ = 2−a 1
2 − q.
We first prove a generalized weighted Poincaré trace inequality. The proof is standard. We recall that the
1,q N +1 qa/2
space W0,S (B+ ;t ) was defined in Section 2.

Lemma A.2. Suppose that q > 1 and τ := 2−a 1

2 − q > 0. Then the following inequality holds
∫ ∫ ∫
q q q 1,q N +1 qa/2
tqa/2 |u| dz ≤ C tqa/2 |∇u| dz + C |u| dz ∀u ∈ W0,S (B+ ;t ). (A.13)
N +1 N +1
B+ B+ BN

Proof . Inequality (A.13) is well known for a = 0. So we restrict ourself to the case a ̸= 0.
1,q N +1 qa/2
Assume by contradiction that (A.13) does not hold. Then there exists a sequence un ∈ W0,S (B+ ;t )
such that
∫ ∫
q q
tqa/2 |∇un | dz + |un | dz = o(1) (A.14)
N +1
B+ BN

and
∫
q
tqa/2 |un | dz > 0 ∀n ∈ N.
N +1
B+

q
tqa/2 |un | dz = 1. But then (A.14) implies that un is
∫
Up to normalization, we may assume that N +1
B+
N +1 qa/2 N +1 qa/2 N +1 qa/2
bounded in W 1,q (B+ ;t ) thus un ⇀ u in W (B+ 1,q
;t ) and un → u in Lq (B+ ;t ) (see [30])
so that
∫
q
tqa/2 |u| dz = 1. (A.15)
N +1
B+

It follows from (A.12) ⏐and the compact embedding of W0τ,q (B N ) into Lq (B N ) , that un → u in Lq (B N ).
N +1
From (A.14) we get u⏐ = 0 and also ∇u = 0. It turns out that u = 0 in B+ , a contradiction with
⏐
RN
(A.15). □
 M.M. Fall / Nonlinear Analysis 193 (2020) 111311 25

By an argument of partition of unity, we have the following

( )
2
Lemma A.3. Let 2 > q > max 1, 1+2s . There exist some constants C, c > 0 such that for all
∞ N +1
φ ∈ Cc (R )
∫ ∫ ∫
2 2 −2s
C∥φ∥W τ,q (B N ) ≤ t1−2s |∇φ| dz − κs γ0 |x| φ2 dx + c φ2 dx, (A.16)
RN
+
+1
RN RN

1+2s
where τ = 2 − 1q .

Proof . We put a = 1 − 2s and define

∫ ∫
2 a−1 2
J(v) := ta |∇v| dz − κs γ0 |x| v dx.
RN
+
+1
RN

N +1 a
Let χ ∈ Cc∞ (B N +1 ), 0 ≤ χ ≤ 1 in B N +1 and such that χ ≡ 1 on B N +1 (0, 1/2). Let η ∈ H 1 (B+ ; t ) be
the minimum of the problem
{∫ }
2 N +1 a
inf ta |∇u| dz : u − χ ∈ H01 (B+ ;t ) .
N +1
B+

Then
⎧ a N +1
⎨div(t ∇η) = 0 in B+
N
η = χ on B
⎩
η = 0 on S+N.

It turns out that 0 ≤ η ≤ 1 in B+ N +1

. In addition, thanks to [14], limt→0 ta ∂η ∞ N
∂t ∈ Lloc (B ). Given
∞ N +1
φ ∈ Cc (R ), simple computations based on integration by parts lead to
∫ ∫ ∫
a 2 a 2
t |∇(φη)| dz ≤ t |∇φ| dz + c φ2 dx,
RN
+
+1
RN
+
+1
BN

where c > 0 depends only on η. On the other hand we have

∫ ∫ ∫
a−1 2 a−1 2 a−1 2
|x| (ηφ) dx = |x| φ dx + (1 − η 2 )|x| φ
B N N N
∫R R
∫
a−1 2
≤ |x| φ dx + c φ2 dx.
RN RN
Therefore we obtain
∫
J(φη) ≤ J(φ) + c φ2 dx.
RN
Applying Lemma A.1, we infer that
(∫ )2/q ∫
qa q
t 2 |∇(ηφ)| dz ≤ J(φ) + c φ2 dx.
RN
+
+1
RN

By Lemma A.2 and Hölder inequality (1 < q < 2)

∫
2
C∥ηφ∥W 1,q (RN +1 ;tqa/2 ) ≤ J(φ) + c φ2 dx.
+ RN
Using Lions’ interpolation inequality (A.12) with a ̸= 0, we obtain
∫
2
C∥ηφ∥W τ,q (RN ) ≤ J(φ) + c φ2 dx.
RN

If a = 0, it is well know that W 1,q

(RN
+
+1
embeds continuously into W 1−1/q,q (RN ). Recalling that η = χ ≡ 1
)
N
on B (0, 1/2), the lemma follows by scaling. □
26 M.M. Fall / Nonlinear Analysis 193 (2020) 111311

Taking advantages to the singular nature of the Hardy potential and the scale invariance, we can now
complete the proof of Theorem 1.3.

Proof of Theorem 1.3. Let u ∈ Cc∞ (B N ) and define U := H(ũ) = H(u). Then U ∈ D1,2 (RN
+
+1 1−2s
;t )
thus by Lemma A.3 and a density argument we get
∫ ∫ ∫
2 2 −2s
C∥U ∥W τ,q (B N ) ≤ t1−2s |∇U | dz − κs γ0 |x| U 2 dx + c U 2 dx.
RN
+
+1
RN RN

Since u = U on B N , it follows that

∫ ∫
2 2 −2s 2
C∥u∥W τ,q (B N ) ≤ ∥u∥H s (B N ) − γ0 |x| u dx + c u2 dx.
0
BN BN

Let r ∈ (0, 1) we derive from the above that for every u ∈ Cc∞ (B N (0, r))
∫ ∫
2 2 −2s
C∥u∥W τ,q (B N (0,r)) ≤ ∥u∥H s (B N (0,r)) − γ0 |x| u2 dx + c u2 dx, (A.17)
0
B N (0,r) B N (0,r)

with c, C > 0 independent on r. This holds because ∥u∥H s (B N (0,r)) = ∥u∥H s (B N (0,1)) as long as
0 0
u ∈ Cc∞ (B N (0, r)) and r < 1.
It is clear from (A.17) that wee need only to show that there exists a constant C2 > 0 such that for every
u ∈ Cc∞ (B N )
∫ ∫
2 2 −2s
C2 u dx ≤ ∥u∥H s (B N ) − γ0 |x| u2 dx. (A.18)
0
BN BN

Before proceeding, we recall that the mapping α ↦→ γα (defined in Lemma 4.1) is decreasing. We will use
this fact and the estimates in Lemma 4.2 to conclude the proof. Pick
( )
N − 2s
α ∈ 0,
2
and let r > 0 be so small that

(γ0 − γα )r−2s − c > 0. (A.19)

−2s
As we did in Section 5, by (A.17), we can define the space H0,b s
(B N (0, r)) with b(x) = γ0 |x| − c.
Letting 2 < p + 1 < N −2s , we can choose q (close to 2) so that W0τ,q (B N (0, r)) is compactly embedded
2N

into Lp+1 (B N (0, r)). Then minimization procedure implies that there exists a nonnegative and nontrivial
ur ∈ H0,b
s
(B N (0, r)) solution to
−2s
Bs ur + cur − γ0 |x| ur = Cupr in B N (0, r).

By density
−2s p
(−∆)s u ur − γ0 |x|
˜r + c˜ u
˜r = C˜
ur in D′ (B N (0, r)). (A.20)

We have, by (A.19),
−2s p p
(−∆)s u
˜r − γα |x| ˜r ≥ ((γ0 − γα )r−2s − c)˜
u ur ≥ C˜
ur + C˜ ur in D′ (B N (0, r)).

Therefore by Lemma 4.2, there exists a constant Cr > 0 such that

− N −2s +α
ur ≥ Cr |x| 2 in B N (0, r/2). (A.21)
 M.M. Fall / Nonlinear Analysis 193 (2020) 111311 27

For every k ∈ N, take v k ∈ H0s (B N (0, r)) as the solution to

−2s
Bs vk + cvk − (γ0 − 1/k)|x| vk = C min(upr , k) in B N (0, r).

Since u
˜r satisfies (A.20), it follows (see Remark 3.9) that

ur ≥ vk > 0 in B N (0, r) (A.22)

by Lemmas 3.2 and 3.3.

Put Vk = H(v˜k ) we have that for any Ψ ∈ H0,T1
(B N (0, r); t1−2s )
∫ ∫ ∫
−1 1−2s −2s
κs t ∇Vk · ∇Ψ dz = (γ0 − 1/k) |x| vk Ψ dx − c vk Ψ dx
RN +1
B N (0,r) B N (0,r)
+ ∫ (A.23)
+C min(upr , k)Ψ dx.
B N (0,r)

Thanks to (A.21), we can choose r′ ∈ (0, r/2) (small) such that

− c + Cup−1
r ≥1 in B N (0, r′ ). (A.24)
H(φ)
For such a fixed r′ , take φ ∈ Cc∞ (B N (0, r′ )). For ε > 0, set Vkε = Vk + ε and put ψ := Vkε . We have

2 2
|∇H(φ)| = |Vkε ∇ψ| + ∇Vk · ∇(Vkε ψ 2 )

and also Vkε ψ 2 ∈ H0,T

1
(B N (0, r); t1−2s ). This together with (A.23) yields
∫ ∫
2 −2s vk
κs−1 t1−2s |∇H(φ)| dxdt ≥ (γ0 − 1/k) |x| φ2 dx
RN +1
B N (0,r ′ ) vk + ε
+
∫
vk
−c φ2 dx
vk + ε
B N (0,r ′ )
min(upr , k) 2
∫
+C φ dx.
B N (0,r ′ ) vk + ε

Thus
∫ ∫ ∫
2 −2s vk
κ−1
s t1−2s |∇H(φ)| dxdt ≥ (γ0 − 1/k) |x| φ2 dx − c φ2 dx
RN +1
B N (0,r ′ ) vk + ε B N (0,r ′ )
+
min(upr , k) 2
∫
+C φ dx.
B N (0,r ′ ) vk + ε

By Fatou’s lemma, when ε → 0, we have

∫ ∫ ∫
2 −2s
κ−1
s t1−2s
|∇H(φ)| dxdt ≥ (γ 0 − 1/k) |x| φ2 dx − c φ2 dx
RN
+
+1
B N (0,r ′ ) B N (0,r ′ )

min(upr , k) 2
∫
+C φ dx.
B N (0,r ′ ) vk

By (A.22) we obtain
∫ ∫ ∫
2 −2s
κ−1
s t1−2s |∇H(φ)| dxdt ≥ (γ0 − 1/k) |x| φ2 dx − c φ2 dx
RN
+
+1
B N (0,r ′ ) B N (0,r ′ )

min(upr , k) 2
∫
+C φ dx.
B N (0,r ′ ) ur
28 M.M. Fall / Nonlinear Analysis 193 (2020) 111311

It follows again from Fatou’s lemma that

∫ ∫ ∫
−1 2 −2s
κs t 1−2s
|∇H(φ)| dxdt ≥ γ0 |x| 2
φ dx + (−c + Curp−1 )φ2 dx.
RN
+
+1
B N (0,r ′ ) B N (0,r ′ )

From the choice of r′ in (A.24), we get immediately, for any φ ∈ Cc∞ (B N (0, r′ )),
∫ ∫ ∫
2 −2s 2
κ−1
s t1−2s
|∇H(φ)| dxdt − γ 0 |x| φ dx ≥ φ2 dx.
RN
+
+1
B N (0,r ′ ) B N (0,r ′ )

By scaling we have (A.18) which was our objective. □

References

[1] S. Armstrong, B. Sirakov, Nonexistence of positive supersolutions of elliptic equations via the maximum principle, Comm.
Partial Differential Equations 36 (11) (2011) 2011–2047.
[2] S. Armstrong, B. Sirakov, A new approach to Liouville theorems for elliptic inequalities, in: Progress in Variational
Problems, Jun 2011, Kyoto, Japan. 1740, 2011, pp. 64–73.
[3] P. Baras, J.A. Goldstein, The heat equation with a singular potential, Trans. Amer. Math. Soc. 284 (1) (1984) 121–139.
[4] G. Barbatis, S. Filippas, A. Tertikas, A unified approach to improved Lp Hardy inequalities with best constants, Trans.
Amer. Math. Soc. 356 (2004) 2169–2196.
[5] K. Bogdan, T. Byczkowski, Potential theory for the α-stable Schrödinger operator on bounded Lipschitz domains, Studia
Math. 133 (1) (1999) 53–92.
[6] K. Bogdan, B. Dyda, The best constant in a fractional Hardy inequality, Math. Nachr. 284 (2011) 629–638. http:
//dx.doi.org/10.1002/mana.200810109.
[7] H. Brezis, X. Cabré, Some simple nonlinear PDEs without solutions, Bull. UMI 1 (1998) 223–262.
[8] H. Brezis, L. Dupaigne, A. Tesei, On a semilinear elliptic equation with inverse-square potential, Selecta Math. (N.S.)
11 (1) (2005) 1–7.
[9] H. Brezis, M. Marcus, Hardy’s inequalities revisited. Dedicated to Ennio De Giorgi, Ann. Scuola Norm. Sup. Pisa Cl. Sci.
(4) 25 (1–2) (1997) 217–237.
[10] H. Brezis, M. Marcus, I. Shafrir, Extermal functions for Hardy’s inequality with weight, J. Funct. Anal. 171 (2000) 177–191.
[11] H. Brezis, J.L. Vázquez, Blow-up solutions of some nonlinear elliptic problems, Rev. Mat. Univ. Complut. Madrid 10 (2)
(1997) 443–469.
[12] X. Cabré, Y. Sire, Nonlinear equations for fractional Laplacians I: Regularity, maximum principles, and Hamiltonian
estimates, Ann. Inst. H. Poincaré 31 (2014) 23–53.
[13] X. Cabré, J. Tan, Positive solutions of nonlinear problems involving the square root of the Laplacian, Adv. Math. 224 (5)
(2010) 2052–2093.
[14] L.A. Caffarelli, S. Salsa, L. Silvestre, Regularity estimates for the solution and the free boundary of the obstacle problem
for the fractional Laplacian, Invent. Math. 171 (2) (2008) 425–461.
[15] L. Caffarelli, L. Silvestre, An extension problem related to the fractional Laplacian, Comm. Partial Differential Equations
32 (7–9) (2007) 1245–1260.
[16] A. Capella, J. Dávila, L. Dupaigne, Y. Sire, Regularity of radial extremal solutions for some non local semilinear equations,
Comm. Partial Differential Equations 36 (8) (2011) 1353–1384.
[17] J. Dávila, L. Dupaigne, Hardy-type inequalities, J. Eur. Math. Soc. (JEMS) 6 (3) (2004) 335–365.
[18] ].J. Dávila, L. Dupaigne, M. Montenegro, The extremal solution of a boundary reaction problem, Commun. Pure Appl.
Anal. 7 (4) (2008) 795–817.
[19] L. Dupaigne, Semilinear elliptic PDEs with a singular potential, J. Anal. Math. 86 (2002) 359–398.
[20] L. Dupaigne, G. Nedev, Semilinear elliptic PDEs with a singular potential, Adv. Differential Equations 7 (2002) 973–1002.
[21] E.B. Fabes, C.E. Kenig, R.P. Serapioni, The local regularity of solutions of degenerate elliptic equations, Comm. Partial
Differential Equations 7 (1) (1982) 77–116.
[22] M.M. Fall, Nonexistence of distributional supersolutions of a semilinear elliptic equation with Hardy potential, J. Funct.
Anal. 263 (8) (2012) 2205–2227.
[23] M.M. Fall, On the Hardy Poincaré inequality with boundary singularities, Commun. Contemp. Math. 14 (3) (2012)
1250019. http://dx.doi.org/10.1155/2011/917201.
[24] M.M. Fall, V. Felli, Unique continuation property and local asymptotics of solutions to fractional elliptic equations, Comm.
Partial Differential Equations 39 (2) (2014) 354–397.
[25] M.M. Fall, R. Musina, Sharp nonexistence results for a linear elliptic inequality involving Hardy and Leray potentials,
J. Inequal. Appl. 2011 (2011).
[26] M.M. Fall, R. Musina, Hardy-Poincaré inequality with boundary singularities, Proc. Roy. Soc. Edinburgh A 142 (2012)
1–18.
 M.M. Fall / Nonlinear Analysis 193 (2020) 111311 29

[27] M.M. Fall, T. Weth, Nonexistence results for a class of fractional elliptic boundary value problems, J. Funct. Anal. 263 (8)
(2012) 2205–2227.
[28] P. Felmer, A. Quaas, J. Tan, Positive solutions of Nonlinear Schrödinger equation with the fractional Laplacian, Proc.
Roy. Soc. Edinburgh Sect. A 142 (6) (2012) 1237–1262.
[29] R.L. Frank, A Simple Proof of Hardy-Lieb–Thirring Inequalities, Comm. Math. Phys. 290 (2009) 789–800.
[30] A. Fröhlich, The Helmholtz decomposition of weighted Lq -spaces for Muckenhoupt weights, Ann. Univ. Ferrara Sez. 46 (1)
(2000) 11–19.
[31] F. Gazzola, H.-C. Grunau, E. Mitidieri, Hardy inequalities with optimal constants and remainder terms, Trans. Amer.
Math. Soc. 356 (2004) 2149–2168.
[32] P. Grisvard, Elliptic Problems in Nonsmooth Domains, in: Monographs and Studies in Mathematics, vol. 24, Pitman
(Advanced Publishing Program), Boston, MA, 1985.
[33] I.W. Herbst, Spectral theory of the operator (p2 + m2 )1/2 − Ze2 /r, Comm. Math. Phys. 53 (3) (1977) 285–294.
[34] T. Kilpeläinen, Weighted Sobolev spaces and capacity, Ann. Acad. Sci. Fenn. AI 19 (1) (1994) 95–113.
[35] V. Kondratieva, V. Liskevich, Z. Sobolb, Second-order semilinear elliptic inequalities in exterior domains, J. Differential
Equations 187 (2003) 429–455.
[36] V. Kondratieva, V. Liskevich, Z. Sobolb, Positive solutions to semi-linear and quasi-linear second-order elliptic equations on
unbounded domains, in: M. Chipot (Ed.), Handbook of Differential Equation, in: Stationary Partial Differential Equations,
vol. 6, Elsevier, 2008, pp. 177–268.
[37] V. Kondratieva, V. Liskevich, Z. Sobolb, Positive supersolutions to semi-linear second-order non-divergence type elliptic
equations in exterior domains, Trans. Amer. Math. Soc. 361 (2009) 697–713.
[38] N.S. Landkof, Foundations of Modern Potential Theory, Springer-Verlag, New York, 1972 Translated from the Russian by
A.P. Doohovskoy, Die Grundlehren der mathematischen Wissenschaften, Band 180.
[39] E.H. Lieb, Sharp constants in the Hardy-Littlewood-Sobolev and related inequalities, Ann. of Math. (2) 118 (2) (1983)
349–374.
[40] J.L. Lions, Théorémes de trace et d’interpolation. I, Ann. Scuola Norm. Sup. Pisa (3) 13 (1959) 389–403.
[41] V. Liskevich, S. Lyakhova, V. Moroz, Positive solutions to semilinear elliptic equations with critical lower order terms,
in: EQUADIFF 2003, World Sci. Publ., Hackensack, NJ, 2005, pp. 549–554.
[42] V.G. Maz’ja, Sobolev Spaces, Springer-Verlag, 1985.
[43] A. Nekvinda, Characterization of traces of the weighted Sobolev space W 1,p (Ω , dεM ) on M , Czechoslovak Math. J.
43(118) (4) (1993) 695–711.
[44] S.I. Pohozaev, A. Tesei, Nonexistence of local solutions to semilinear partial differential inequalities, Ann. Inst. H. Poincaré
Anal. Non Linéaire 21 (2004) 487–502.
[45] L. Silvestre, Regularity of the obstacle problem for a fractional power of the Laplace operator, Comm. Pure Appl. Math.
60 (1) (2007) 67–112.
[46] Y. Sire, E. Valdinoci, Fractional Laplacian phase transitions and boundary reactions: a geometric inequality and a symmetry
result, J. Funct. Anal. 256 (6) (2009) 1842–1864.
[47] E.M. Stein, G. Weiss, Fractional integrals on n-dimensional Euclidean space, J. Math. Mech. 7 (1958) 503–514.
[48] E.M. Stein, G. Weiss, Introduction to Fourier Analysis on Euclidean Spaces, in: Princeton Mathematical Series, vol. 32,
Princeton University Press, Princeton, NJ, 1971.
[49] S. Terracini, On positive entire solutions to a class of equations with a singular coefficient and critical exponent, Adv.
Differential Equations 1 (2) (1996) 241–264.
[50] J.L. Vázquez, E. Zuazua, The Hardy inequality and the asymptotic behavior of the heat equation with an inverse-square
potential, J. Funct. Anal. 173 (2000).
[51] D. Yafaev, Sharp constants in the Hardy-Rellich inequalities, J. Funct. Anal. 168 (1) (1999) 121–144.