Advances in Mathematics 224 (2010) 2052–2093

www.elsevier.com/locate/aim

Positive solutions of nonlinear problems involving the

square root of the Laplacian
Xavier Cabré a,∗ , Jinggang Tan b
a ICREA and Departament de Matemàtica Aplicada I, Universitat Politècnica de Catalunya, Diagonal 647,
08028 Barcelona, Spain
b Departamento de Matemática, Universidad Técnica Federico Santa María, Avda. España 1680, Valparaíso, Chile

Received 11 May 2009; accepted 12 January 2010

Available online 20 February 2010
Communicated by Luis Caffarelli

Abstract
We consider nonlinear elliptic problems involving a nonlocal operator: the square root of the Laplacian in
a bounded domain with zero Dirichlet boundary conditions. For positive solutions to problems with power
nonlinearities, we establish existence and regularity results, as well as a priori estimates of Gidas–Spruck
type. In addition, among other results, we prove a symmetry theorem of Gidas–Ni–Nirenberg type.
© 2010 Elsevier Inc. All rights reserved.

Keywords: Fractional Laplacian; Critical exponent; Nonlinear mixed boundary problem; A priori estimates; Nonlinear
Liouville theorems; Moving planes method

1. Introduction

This paper is concerned with the study of positive solutions to nonlinear problems involving
a nonlocal positive operator: the square root of the Laplacian in a bounded domain with zero
Dirichlet boundary conditions. We look for solutions to the nonlinear problem
⎧
⎨ A1/2 u = f (u) in Ω,
u=0 on ∂Ω, (1.1)
⎩
u>0 in Ω,

* Corresponding author.
E-mail addresses: xavier.cabre@upc.edu (X. Cabré), jinggang.tan@usm.cl (J. Tan).

0001-8708/$ – see front matter © 2010 Elsevier Inc. All rights reserved.
doi:10.1016/j.aim.2010.01.025
 X. Cabré, J. Tan / Advances in Mathematics 224 (2010) 2052–2093 2053

where Ω is a smooth bounded domain of Rn and A1/2 stands for the square root of the Laplacian
operator − in Ω with zero Dirichlet boundary values on ∂Ω.
To define A1/2 , let {λk , ϕk }∞
k=1 be the eigenvalues and corresponding eigenfunctions of the
Laplacian operator − in Ω with zero Dirichlet boundary values on ∂Ω,

−ϕk = λk ϕk in Ω,
ϕk = 0 on Ω,

normalized by ϕk L2 (Ω) = 1. The square root of the Dirichlet Laplacian, that we denote by
A1/2 , is given by

∞
 ∞
 1/2
u= ck ϕk → A1/2 u = ck λk ϕk , (1.2)
k=1 k=1

 ∞
which clearly maps H01 (Ω) = {u = ∞ k=1 ck ϕk | k=1 λk ck < ∞} into L (Ω).
2 2

The fractions of the Laplacian, such as the previous square root A1/2 , are the infinitesimal
generators of Lévy stable diffusion processes and appear in anomalous diffusions in plasmas,
flames propagation and chemical reactions in liquids, population dynamics, geophysical fluid
dynamics, and American options in finance.
Essential to the results in this paper is to realize the nonlocal operator A1/2 in the following
local manner. Given a function u defined in Ω, we consider its harmonic extension v in the cylin-
der C := Ω × (0, ∞), with v vanishing on the lateral boundary ∂L C := ∂Ω × [0, ∞). Then, A1/2
is given by the Dirichlet to Neumann map on Ω, u → ∂ν ∂v
|Ω×{0} , of such harmonic extension
in the cylinder. In this way, we transform problem (1.1) to a local problem in one more dimen-
sion. By studying this problem with classical local techniques, we establish existence of positive
solutions for problems with subcritical power nonlinearities, regularity and an L∞ -estimate of
Brezis–Kato type for weak solutions, a priori estimates of Gidas–Spruck type, and a nonlinear
Liouville type result for the square root of the Laplacian in the half-space. We also obtain a
symmetry theorem of Gidas–Ni–Nirenberg type.
The analogue problem to (1.1) for the Laplacian has been investigated widely in the last
decades. This is the problem
⎧
⎨ −u = f (u) in Ω,
u=0 on ∂Ω, (1.3)
⎩
u>0 in Ω;

see [24] and references therein. Considering the minimization problem min{uH 1 (Ω) |
0
uLp+1 (Ω) = 1}, one obtains a positive solution of (1.3) in the case f (u) = up , 1 < p < n+2
n−2 ,
since the Sobolev embedding is compact. Ambrosetti and Rabinowitz [1] introduced the moun-
tain pass theorem to study problem (1.3) for more general subcritical nonlinearities. Instead, for
n+2
f (u) = u n−2 , Pohozaev identity leads to nonexistence to (1.3) if Ω is star-shaped. In contrast,
Brezis and Nirenberg [4] showed that the nonexistence of solution may be reverted by adding a
small linear perturbation to the critical power nonlinearity.
For the square root A1/2 of the Laplacian, we derive the following result on existence of
positive solutions to problem (1.1).
2054 X. Cabré, J. Tan / Advances in Mathematics 224 (2010) 2052–2093

Theorem 1.1. Let n  1 be an integer and 2 = n−1 2n

when n  2. Suppose that Ω is a smooth
bounded domain in R and f (u) = u . Assume that 1 < p < 2 − 1 = n+1
n p
n−1 if n  2, or that
1 < p < ∞ if n = 1.
Then, problem (1.1) admits at least one solution. This solution (as well as every weak solution)
belongs to C 2,α (Ω) for some 0 < α < 1.

As mentioned before, we realize problem (1.1) through a local problem in one more dimension
by a Dirichlet to Neumann map. This provides a variational structure to the problem, and we
study its corresponding minimization problem. Here the Sobolev trace embedding comes into
play, and its critical exponent 2 = n−1 2n
, n  2, is the power appearing in Theorem 1.1. We
call p critical (respectively, subcritical or supercritical) when p = 2 − 1 = n+1
n−1 (respectively,
p < 2 − 1 or p > 2 − 1). In the subcritical case of Theorem 1.1, the compactness of the Sobolev
 

trace embedding in bounded domains leads to the existence of solution. Its regularity will be
consequence of further results presented later in this introduction.

Remark 1.2. In [26] the second author J. Tan establishes the nonexistence of classical solutions to
(1.1) with f (u) = up in star-shaped domains for the critical and supercritical cases. In addition,
an existence result of Brezis–Nirenberg type [4] for f (u) = up + μu, μ > 0, is also established.

Gidas and Spruck [14] established a priori estimates for positive solutions of problem (1.3)
when f (u) = up and p < n+2 n−2 . Its proof involves the method of blow-up combined with two
important ingredients: nonlinear Liouville type results in all space and in a half-space. The proofs
of such Liouville theorems are based on the Kelvin transform and the moving planes method or
the moving spheres method. Here we establish an analogue: the following a priori estimates of
Gidas–Spruck type for solutions of problem (1.1).

Theorem 1.3. Let n  2 and 2 = n−1 2n

. Assume that Ω ⊂ Rn is a smooth bounded domain and
f (u) = up , 1 < p < 2 − 1 = n+1
n−1 .
Then, there exists a constant C(p, Ω), which depends only on p and Ω, such that every weak
solution of (1.1) satisfies

uL∞ (Ω)  C(p, Ω).

To prove this result, we combine the blow-up method and two useful ingredients: a nonlinear
Liouville theorem for the square root of the Laplacian in all of Rn , and a similar one in the half-
space Rn+ with zero Dirichlet boundary values on ∂Rn+ . The first one in the whole space was
proved by Ou [22] using the moving planes method and by Y.Y. Li, M. Zhu and L. Zhang [18,17]
using the moving spheres method. Its statement is the following.

Theorem 1.4. (See [18,22,17].) For n  2 and 1 < p < 2 − 1 = n+1

n−1 , there exists no weak
solution of the problem


(−)1/2 u = up in Rn ,
(1.4)
u>0 in Rn .
 X. Cabré, J. Tan / Advances in Mathematics 224 (2010) 2052–2093 2055

As we will see later, here (−)1/2 is the usual half-Laplacian in all of Rn , and problem (1.4)
is equivalent to problem v = 0 and v > 0 in Rn+1 n+1
+ , ∂ν v = v on ∂R+ . The corresponding Li-
p

ouville theorem for the square root of the Laplacian in R+ = {x ∈ R | xn > 0} was not available
n n

and we establish it in this paper for bounded solutions.

Theorem 1.5. Let n  2, 2 = 2n

n−1 , and 1 < p  2 − 1 = n+1
n−1 . Then, there exists no bounded
solution u of
⎧
⎨ A1/2 u = u in Rn+ ,
p

u=0 on ∂Rn+ , (1.5)

⎩
u>0 in Rn+ ,

where A1/2 is the square root of the Laplacian in Rn+ = {xn > 0} with zero Dirichlet boundary
conditions on ∂Rn+ .
In an equivalent way, let

Rn+1
++ = z = (x1 , x2 , . . . , xn , y) xn > 0, y > 0 .

If n  2 and 1 < p  2 − 1 = n+1

n−1 , then there exists no bounded solution v ∈ C 2 (Rn+1
++ ) ∩
C(Rn+1
++ ) of
⎧
⎪
⎪ v = 0 in Rn+1
++ ,
⎪
⎪
⎪
⎨v = 0 on {xn = 0, y > 0},
∂v (1.6)
⎪
⎪ = vp on {xn > 0, y = 0},
⎪ ∂ν
⎪
⎪
⎩
v>0 in Rn+1
++ ,

where ν is the unit outer normal to Rn+1

++ at {xn > 0, y = 0}.

The proof of this result combines the Kelvin transform, the moving planes method, and a
Hamiltonian identity for the half-Laplacian found by Cabré and Solà-Morales [5]. The result of
Theorem 1.5 is still open without the assumption of boundedness of the solution.
Gidas, Ni, and Nirenberg [13] established symmetry properties for solutions to problem (1.3)
when f is Lipschitz continuous and Ω has certain symmetries. The proof of these symmetry
results uses the maximum principle and the moving planes method. The moving planes method
was introduced by Alexandroff to study a geometric problem, while in the framework of problem
(1.3) was first used by Serrin. In the improved version of Berestycki and Nirenberg [3], it replaces
the use of Hopf’s lemma by a maximum principle in domains of small measure.
Here we proceed in a similar manner and obtain the following symmetry result of Gidas–Ni–
Nirenberg type for (1.1).

Theorem 1.6. Assume that Ω is a bounded smooth domain of Rn which is convex in the x1
direction and symmetric with respect to the hyperplane {x1 = 0}. Let f be Lipschitz continuous
and u be a C 2 (Ω) solution of (1.1).
Then, u is symmetric with respect to x1 , i.e., u(−x1 , x ) = u(x1 , x ) for all (x1 , x ) ∈ Ω. In
∂u
addition, ∂x 1
< 0 for x1 > 0.
2056 X. Cabré, J. Tan / Advances in Mathematics 224 (2010) 2052–2093

In particular, if Ω = BR (0) is a ball, then u is radially symmetric, u = u(|x|) = u(r) for

r = |x|, and it is decreasing, i.e., ur < 0 for 0 < r < R.

We prove this symmetry result by using the moving planes method combined with the follow-
ing maximum principle for the square root A1/2 of the Laplacian in domains of small measure
(see Proposition 4.4 for a more general statement in nonsmooth domains).

Proposition 1.7. Assume that u ∈ C 2 (Ω) satisfies


A1/2 u + c(x)u  0 in Ω,
u=0 on ∂Ω,

where Ω is a smooth bounded domain in Rn and c ∈ L∞ (Ω). Then, there exists δ > 0 depending
only on n and c− L∞ (Ω) , such that if |Ω ∩ {u < 0}|  δ then u  0 in Ω.

The above maximum principle in “small” domains replaces the use of Hopf’s lemma to prove
symmetry results for A1/2 in Lipschitz domains. We point out that Chipot, Chlebík, Fila, and
Shafrir [9] studied the related problem:
⎧ 
⎪
⎪ −v = g(v) in BR+ = z ∈ Rn+1 |z|  R, zn+1 > 0 ,
⎪
⎪
⎪
⎨v = 0 on ∂BR+ ∩ {zn+1 > 0},
∂v (1.7)
⎪
⎪ = f (v) on ∂B +
∩ {z = 0},
⎪
⎪ ∂ν R n+1
⎪
⎩
v>0 in BR+ ,

where f, g ∈ C 1 (R) and ν is the unit outer normal. They proved existence, nonexistence, and
axial symmetry results for solutions of (1.7). Following one of their proofs, we establish Hopf’s
lemma for A1/2 , Lemma 4.3 below. Finally, let us mention that singular solutions and extremal
solutions of similar problems to (1.7) have been considered by Davila, Dupaigne, and Montene-
gro [10,11].
As we mentioned, crucial to our results is that A1/2 is a nonlocal operator in Ω but which can
be realized through a local problem in Ω × (0, ∞). To explain this, let us start with the square
root of the Laplacian (or half-Laplacian) in Rn . Let u be a bounded continuous function in all
of Rn . There is a unique harmonic extension v of u in the half-space Rn+1+ = R × (0, ∞). That
n

is,

v = 0 in Rn+1
+ = (x, y) ∈ R × (0, ∞) ,
n

v=u on Rn = ∂Rn+1
+ .

Consider the operator T : u → −∂y v(·, 0). Since ∂y v is still a harmonic function, if we apply the
operator T twice, we obtain

(T ◦ T )u = ∂yy v|y=0 = −x v|y=0 = −u in Rn .

Thus, we see that the operator T mapping the Dirichlet data u to the Neumann data −∂y v(·, 0)
is actually a square root of the Laplacian. Indeed it coincides with the usual half-Laplacian,
see [16].
 X. Cabré, J. Tan / Advances in Mathematics 224 (2010) 2052–2093 2057

Here we introduce a new analogue extension problem in a cylinder C := Ω × (0, ∞) in one

more dimension to realize (1.1) by a local problem in C. More precisely, we look for a function v
with v(·, 0) = u in Rn satisfying the following mixed boundary value problem in a half-cylinder:
⎧ v = 0 in C = Ω × (0, ∞),
⎪
⎪
⎪
⎨v = 0 on ∂L C := ∂Ω × [0, ∞),
∂v (1.8)
⎪
⎪ = f (v) on Ω × {0},
⎪
⎩ ∂ν
v>0 in C,
where ν is the unit outer normal to C at Ω × {0}. If v satisfies (1.8), then the trace u on Ω × {0}
of v is a solution of problem (1.1). Indeed, since ∂y v is harmonic and also vanishes on the lateral
boundary ∂Ω × [0, ∞), we see as before that the Dirichlet to Neumann map u → −∂y v(·, 0) is
the unique positive square root A1/2 of the Dirichlet Laplacian in Ω.
The generators of Lévy symmetric stable diffusion processes are the fractional powers of the
Laplacian (−)s in all of Rn , 0 < s < 1. Fractional Laplacians attract nowadays much interest
in physics, biology, finance, as well as in mathematical nonlinear analysis (see [2]). One of the
few nonlinear results which is not recent is due to Sugitani [25], who proved blow up results for
solutions of heat equations ∂t u + (−)s u = f (u) in Rn , for all 0 < s < 1. It is important to note
that the fundamental solution of the fractional heat equation has power decay (or heavy) tails, in
contrast with the exponential decay in case of the classical heat equation. Lévy processes have
also been applied to model American options [2]. As recent nonlinear works for fractional diffu-
sions, let us mention the following. Caffarelli and Silvestre [7] have given a new local realization
of the fractional Laplacian (−)s , for all 0 < s < 1, through the Dirichlet–Neumann map of an
appropriate degenerate elliptic operator. The regularity of the obstacle problem for the fractional
powers of the Laplacian operator was proved by Silvestre [23]. The optimal regularity for such
Signorini problem was improved in [6]. Moreover, the operator (−)s plays an important role
in the study of the quasi-geostrophic equations in geophysical fluid dynamics; see the important
recent paper [8] by Caffarelli and Vasseur. Cabré and Solà-Morales [5] studied layer solutions
(solutions which are monotone with respect to one variable) of (−)1/2 u = f (u) in Rn , where
f is of balanced bistable type.
To prove Theorem 1.1, in view of (1.8) being a local realization of (1.1), we consider the
Sobolev space

1
H0,L (C) = v ∈ H 1 (C) v = 0 a.e. on ∂L C = ∂Ω × [0, ∞) ,

equipped with the norm v = ( C |∇v|2 dx dy)1/2 . Since problem (1.8) has variational structure,
we consider its corresponding minimization problem
  
2 p+1
I0 = inf ∇v(x, y) dx dy v ∈ H0,L
1
(C), v(x, 0) dx = 1 .
C Ω

We will prove that, for subcritical powers, there is a minimizer for this problem. Its trace on
Ω × {0} will provide with a weak solution of (1.1).
Thus, it is important to characterize the space V0 (Ω) of all traces on Ω × {0} of functions in
1 (C). This is stated in the following result—which corresponds to Proposition 2.1 of the next
H0,L
section.
2058 X. Cabré, J. Tan / Advances in Mathematics 224 (2010) 2052–2093

Proposition 1.8. Let V0 (Ω) be the space of all traces on Ω × {0} of functions in H0,L
1 (C). Then,

we have


V0 (Ω) := u = trΩ v v ∈ H0,L
1
(C)
  2 
u (x)
= u ∈ H 1/2 (Ω) dx < +∞
d(x)
Ω
∞ ∞

  1/2
= u ∈ L (Ω) u =
2
bk ϕk satisfying bk2 λk < +∞ ,
k=1 k=1

where d(x) = dist(x, ∂Ω), and {λk , ϕk } is the Dirichlet spectral decomposition of − in Ω as
above, with {ϕk } an orthonormal basis of L2 (Ω).
Furthermore, V0 (Ω) equipped with the norm

  2 1/2
u
uV0 (Ω) = uH 1/2 (Ω) +
2
(1.9)
d
Ω

is a Banach space.

The fact that d −1/2 u ∈ L2 (Ω) if u is the trace of a function in H0,L

1 (C) follows from a trace

boundary Hardy inequality, originally due to Nekvinda [21]; see Lemma 2.6 in the next section
for a simple proof. Thus, in the next section we need to consider the operator A1/2 defined as
in (1.2) but now mapping A1/2 : V0 (Ω) → V0∗ (Ω), where V0∗ (Ω) is the dual space of V0 (Ω). For
 ∞ ∞
u= ∞
1/2
k=1 bk ϕk ∈ V0 (Ω), we will have A1/2 ( k=1 bk ϕk ) = k=1 bk λk ϕk . Moreover, there
will be a unique harmonic extension v ∈ H0,L 1 (C) in C of u, and it is given by the expression

∞
  1/2 
v(x, y) = bk ϕk (x) exp −λk y for all (x, y) ∈ C.
k=1

Thus, the operator A1/2 : V0 (Ω) → V0∗ (Ω) is given by the Dirichlet–Neumann map

∞

∂v 1/2
A1/2 u := = bk λk ϕk .
∂ν Ω×{0} k=1

Note that A1/2 ◦ A1/2 is equal to − in Ω with zero Dirichlet boundary value on ∂Ω. More
precisely, we will have that the inverse B1/2 = A−1 ∗
1/2 —which maps V0 (Ω) into itself, and also
L2 (Ω) into itself—is the unique square root of the inverse Laplacian (−)−1 in Ω with zero
Dirichlet boundary values on ∂Ω; see the next section for details.
To establish the regularity of weak solutions to (1.1) obtained by the previous minimization
technique, we establish the following results of Calderón–Zygmund and of Schauder type for the
linear problem
 X. Cabré, J. Tan / Advances in Mathematics 224 (2010) 2052–2093 2059


A1/2 u = g(x) in Ω,
(1.10)
u=0 on ∂Ω;

see Proposition 3.1 for more details.

Theorem 1.9. Let u ∈ V0 (Ω) be a weak solution of (1.10), where g ∈ V0∗ (Ω) and Ω is a C 2,α
bounded domain in Rn , for some 0 < α < 1.
If g ∈ L2 (Ω), then u ∈ H01 (Ω).
If g ∈ H01 (Ω), then u ∈ H 2 (Ω) ∩ H01 (Ω).
If g ∈ L∞ (Ω), then u ∈ C α (Ω).
If g ∈ C α (Ω) and g|∂Ω ≡ 0, then u ∈ C 1,α (Ω).
If g ∈ C 1,α (Ω) and g|∂Ω ≡ 0, then u ∈ C 2,α (Ω).

In this paper we will give full—and rather simple—proofs of these regularity results, specially
since we could only find references for some of them and, besides, in close statements to ours but
not precisely ours. Our proof of Theorem 1.9 uses the extension problem in Ω × (0, ∞) related
to (1.10), and transforms it to a problem with zero Dirichlet boundary in Ω × {0} by using
an auxiliary function introduced in [5]. Then, by making certain reflections and using classical
interior regularity theory for the Laplacian, we prove Hölder regularity for u and its derivatives.
To apply the previous Hölder regularity linear results to our nonlinear problem (1.1), we first
need to prove that g := f (u) is bounded, i.e., u is bounded. We will see that boundedness of weak
solutions holds for subcritical and critical nonlinearities; we establish this result in Section 5. We
will follow the Brezis–Kato approach using the bootstrap method. In this way, we establish the
following (see Theorem 5.2).

Theorem 1.10. Assume that g0 is a Carathéodory function in Ω × R satisfying

 
g0 (x, s)  C 1 + |s|p for all (x, s) ∈ Ω × R,

for some constant C, 1  p  n+1

n−1 if n  2, or 1  p < ∞ if n = 1, where Ω is a smooth
bounded domain in Rn . Let u ∈ V0 (Ω) be a weak solution of

A1/2 u = g0 (x, u) in Ω,
u=0 on ∂Ω.

Then, u ∈ L∞ (Ω).

The paper is organized as follows. In Section 2, we study the appropriate function spaces
1 (C) and V (Ω), and we give the proof of Proposition 1.8 and other related results. The
H0,L 0
regularity results of Theorem 1.9 can be found in Section 3. Maximum principles, Hopf’s lemma,
and the maximum principle in “small” domains of Proposition 1.7 are proved in Section 4. The
complete proof of Theorem 1.1 is given in Section 5 by studying the minimization problem and
applying the previous results on regularity and maximum principles. We prove Theorem 1.10
also in Section 5, while Theorems 1.3 and 1.5 are established in Section 6, and Theorem 1.6 in
Section 7.
2060 X. Cabré, J. Tan / Advances in Mathematics 224 (2010) 2052–2093

2. Preliminaries: function spaces and the operator A1/2

In this section we collect preliminary facts for future reference. First of all, let us set the
standard notations to be used in the paper. We denote the upper half-space in Rn+1 by

Rn+1
+ = z = (x, y) = (x1 , . . . , xn , y) ∈ R
n+1
y>0 .

Denote by H s (U ) = W s,2 (U ) the Sobolev space in a domain U of Rn or of Rn+1

+ . Letting
U ⊂ Rn and s > 0, H s (U ) is a Banach space with the norm
   1/2
|u(x) − u(x̄)|2 2
uH s (U ) = dx d x̄ + u(x) dx .
|x − x̄|n+2s
U U U

Let Ω be a bounded smooth domain in Rn . Denote the half-cylinder with base Ω by

C = Ω × (0, ∞)

and its lateral boundary by

∂L C = ∂Ω × [0, ∞).

To treat the nonlocal problem (1.1), we will study a corresponding extension problem in one
more dimension, which allows us to investigate (1.1) by studying a local problem via classical
nonlinear variational methods. We consider the Sobolev space of functions in H 1 (C) whose
traces vanish on ∂L C:

1
H0,L (C) = v ∈ H 1 (C) v = 0 a.e. on ∂L C , (2.1)

equipped with the norm

 1/2
v = |∇v| dx dy
2
. (2.2)
C

We denote by trΩ the trace operator on Ω × {0} for functions in H0,L

1 (C):

trΩ v := v(·, 0), for v ∈ H0,L

1
(C).

We have that trΩ v ∈ H 1/2 (Ω), since it is well known that traces of H 1 functions are H 1/2
functions on the boundary.
Recall the well known spectral theory of the Laplacian − in a smooth bounded domain Ω
with zero Dirichlet boundary values. We repeat each eigenvalue of − in Ω with zero Dirichlet
boundary conditions according to its (finite) multiplicity:

0 < λ1 < λ2  · · ·  λk  · · · → ∞, as k → ∞,
 X. Cabré, J. Tan / Advances in Mathematics 224 (2010) 2052–2093 2061

and we denote by ϕk ∈ H01 (Ω) an eigenfunction corresponding to λk for k = 1, 2, . . . . Namely,


−ϕk = λk ϕk in Ω,
(2.3)
ϕk = 0 on Ω.

We can take them to form an orthonormal basis {ϕk } of L2 (Ω), in particular,


ϕk2 dx = 1,
Ω

and to belong to C 2 (Ω) by regularity theory.

Now we can state the main results which we prove in this section.

Proposition 2.1. Let V0 (Ω) be the space of all traces on Ω × {0} of functions in H0,L
1 (C). Then,

we have

V0 (Ω) := u = trΩ v v ∈ H0,L
1
(C)
  2 
u (x)
= u ∈ H (Ω)
1/2
dx < +∞
d(x)
Ω
∞ ∞

  1/2
= u ∈ L2 (Ω) u = bk ϕk satisfying bk2 λk < +∞ ,
k=1 k=1

where d(x) = dist(x, ∂Ω), and {λk , ϕk } is the Dirichlet spectral decomposition of − in Ω as
above, with {ϕk } an orthonormal basis of L2 (Ω).
Furthermore, V0 (Ω) equipped with the norm
  2 1/2
u
uV0 (Ω) = uH 1/2 (Ω) +
2
(2.4)
d
Ω

is a Banach space.

Proposition 2.2. If u ∈ V0 (Ω), then there exists a unique harmonic extension

∞ v in C of u such
1 (C). In particular, if the expansion of u is written by u =
that v ∈ H0,L k=1 k k ∈ V0 (Ω), then
b ϕ

∞
  1/2 
v(x, y) = bk ϕk (x) exp −λk y for all (x, y) ∈ C,
k=1

where {λk , ϕk } is the Dirichlet spectral decomposition of − in Ω as above, with {ϕk } an or-
thonormal basis of L2 (Ω).
The operator A1/2 : V0 (Ω) → V0∗ (Ω) is given by

∂v
A1/2 u := ,
∂ν Ω×{0}
2062 X. Cabré, J. Tan / Advances in Mathematics 224 (2010) 2052–2093

where V0∗ (Ω) is the dual space of V0 (Ω). We have that

∞
 1/2
A1/2 u = bk λk ϕk ,
k=1

and that A1/2 ◦ A1/2 (when A1/2 is acting, for instance, on smooth functions with compact sup-
port in Ω) is equal to − in Ω with zero Dirichlet boundary values on ∂Ω. More precisely, the
inverse B1/2 := A−1
1/2 is the unique positive square root of the inverse Laplacian (−)
−1 in Ω

with zero Dirichlet boundary values on ∂Ω.

The proofs of these two propositions need the development of several tools. First let us
1 (C). Denote by D 1,2 (Rn+1 ) the closure of the set of
give some properties of the space H0,L +
smooth functions compactly supported in Rn+1 + with respect to the norm of wD1,2 (Rn+1 ) =
+
( Rn+1 |∇w|2 dx dy)1/2 . We recall the well known Sobolev trace inequality. For w ∈ D1,2 (Rn+1
+ ),
+
we have
 (n−1)/2n   1/2
2n/(n−1) 2
w(x, 0) dx C ∇w(x, y) dx dy , (2.5)
Rn Rn+1
+

where C depends only on n.

Denote for n  2,

2n n+1
2 = and 2 − 1 = .
n−1 n−1

We say that p is subcritical if 1 < p < 2 − 1 = n+1

n−1 for n  2, and 1 < p < ∞ for n = 1. We also
say that p is critical if p = 2 − 1 = n−1 for n  2, and that p is supercritical if p > 2 − 1 = n+1
 n+1
n−1
for n  2.
Lions [19] showed that

Rn+1 |∇w(x, y)|2 dx dy  n+1 
+
S0 = inf w∈D 1,2
R+ (2.6)
|w(x, 0)|2 dx)2/2
 
( Rn

is achieved. Escobar [12] proved that the extremal functions have all the form

ε (n−1)/2
Uε (x, y) = , (2.7)
|(x − x0 , y + ε)|n−1

where x0 ∈ Rn and ε > 0 are arbitrary. In addition, the best constant is

1/n
(n − 1)σn
S0 = ,
2

where σn denotes the volume of n-dimensional sphere Sn ⊂ Rn+1 .

 X. Cabré, J. Tan / Advances in Mathematics 224 (2010) 2052–2093 2063

The Sobolev trace inequality leads directly to the next three lemmas. For v ∈ H0,L
1 (C), its

extension by zero in Rn+1 n+1

+ \ C can be approximated by functions compactly supported in R+ .
Thus the Sobolev trace inequality (2.5) leads to:

Lemma 2.3. Let n  2 and 2 = 2n

n−1 . Then there exists a constant C, depending only on n, such
that, for all v ∈ H0,L
1 (C),

 1/2  1/2
2 2
v(x, 0) dx C ∇v(x, y) dx dy . (2.8)
Ω C

By Hölder’s inequality, since Ω is bounded, the above lemma leads to:

Lemma 2.4. Let 1  q  2 for n  2. Then, we have that for all v ∈ H0,L
1 (C),

 1/q  1/2
q 2
v(x, 0) dx C ∇v(x, y) dx dy , (2.9)
Ω C

where C depends only on n, q, and the measure of Ω. Moreover, (2.9) also holds for 1  q < ∞
if n = 1.

1 (C)) ⊂ Lq (Ω), where 1  q  2 for n  2 and 1  q < ∞

This lemma states that trΩ (H0,L
for n = 1 (see the proof of Lemma 2.5 for the case n = 1). In addition, we also have the following
compact embedding.

Lemma 2.5. Let 1  q < 2 = n−1

2n
for n  2 and 1  q < ∞ for n = 1. Then trΩ (H0,L
1 (C)) is
q
compactly embedded in L (Ω).

1 (C)) ⊂ H 1/2 (Ω) and that H 1/2 (Ω) ⊂⊂ Lq (Ω) when 1 

Proof. It is well known that trΩ (H0,L
q < 2 = n−1 for n  2 and 1  q < ∞ for n = 1. Here ⊂⊂ denotes compact embedding.
 2n

This completes the proof of the lemma. However, if one wants to avoid the use of the fractional
Sobolev space H 1/2 (Ω), the following is an alternative simple proof.
Considering the restriction of functions in C to Ω × (0, 1), it suffices to show that the embed-
ding is compact with C replaced by Ω × (0, 1). To prove this, let vm ∈ H0,L 1 (Ω × (0, 1)) := {v ∈

H 1 (Ω × (0, 1)) | v = 0 a.e. on ∂Ω × (0, 1)} such that vm 0 weakly in H0,L 1 (Ω × (0, 1)), as

m → ∞. We may assume by the classical Rellich’s theorem in Ω × (0, 1) that vm → 0 strongly

in L2 (Ω × (0, 1)), as m → ∞. We introduce the function wm = (1 − y)vm . It is clear that

wm |Ω×{0} = vm , wm |Ω×{1} = 0.

By direct computations we have

2064 X. Cabré, J. Tan / Advances in Mathematics 224 (2010) 2052–2093

  1 
2 2  2 
vm (x, 0) dx = wm (x, 0) dx = − ∂y wm (x, y) dx dy
Ω Ω 0 Ω
 1  1/2  1  1/2
2
2 2
wm (x, y) dx dy ∇wm (x, y) dx dy .
0 Ω 0 Ω

Therefore, since wm = (1 − y)vm is bounded in H 1 (Ω × (0, 1)) and wm → 0 strongly in L2 (Ω ×

(0, 1)), we find that, as m → ∞,

vm (x, 0) → 0 strongly in L2 (Ω) and hence also in L1 (Ω).

On the other hand, since q is subcritical, the following interpolation inequality,

   θ  1−θ
vm (·, 0)  vm (·, 0)L1 (Ω) vm (·, 0)L2 (Ω)
Lq (Ω)

for some 0 < θ < 1 completes the proof since we already know that vm converges strongly to
zero in L1 (Ω). 2

We also need to establish a trace boundary Hardy inequality, which already appeared in a
work of Nekvinda [21].

Lemma 2.6. We have that

 1 
trΩ H0,L (C) ⊂ H 1/2 (Ω)

is a continuous injection. In addition, for every v ∈ H0,L

1 (C),

 
|v(x, 0)|2 2
dx  C ∇v(x, y) dx dy, (2.10)
d(x)
Ω C

where d(x) = dist(x, ∂Ω) and the constant C depends only on Ω.

Proof. The first statement is clear since the traces of H 1 (C) functions belong to H 1/2 (∂C).
Regarding the second statement, we prove it in two steps.
Step 1. Assume first that n = 1 and Ω = (0, 1). For 0 < x0 < 1/2, consider the segment from
(0, x0 ) to (x0 , 0) in C = (0, 1) × (0, ∞). We have

x0
x0
v(x0 , 0) = v(t, x0 − t) t=0
= (∂x v − ∂y v)(t, x0 − t) dt.
0

Then
 X. Cabré, J. Tan / Advances in Mathematics 224 (2010) 2052–2093 2065

x0
2 2
v(x0 , 0)  x0 2 ∇v(t, x0 − t) dt.
0

Dividing this inequality by x0 and integrating in x0 over (0, 1/2), and making the change of
variables x = t, y = x0 − t, we deduce

1/2 1/2 1/2 

|v(x0 , 0)|2
dx0  2 dx dy|∇v|  2 |∇v|2 dx dy.
2
x0
0 0 0 C

Doing the same on (1/2, 1), this establishes inequality (2.10) of the lemma.
Step 2. In the general case, after straightening a piece of the boundary ∂Ω and rescaling the new
variables, we can consider the inequality in a domain D = {x = (x , xn ) | |x | < 1, 0 < xn < 1/2}
and assume that v = 0 on {xn = 0, |x | < 1} × (0, ∞), since the flatting procedure possesses
equivalent norms. By the argument in Step 1 above, we have

1/2 1/2∞
|v(x, 0)|2
dxn  C |∇v|2 dxn dy,
xn
0 0 0

for all x with |x | < 1. From this, integrating in x we have

  1/2
|v(x, 0)|2 |v(x, 0)|2
dx = dx dxn
xn xn
D D 0

C |∇v|2 dx dy.
D×(0,∞)

Since after flattening of ∂Ω, xn is comparable to d(x) = dist(x, ∂Ω), this is the desired inequal-
ity (2.10). 2

Recall that the fractional Sobolev space H 1/2 (Ω) is a Banach space with the norm
  
|u(x) − u(x̄)|2 2
u2H 1/2 (Ω) = dx d x̄ + u(x) dx. (2.11)
|x − x̄|n+1
Ω Ω Ω

Note that the closure H0 (Ω) of smooth functions with compact support, Cc∞ (Ω), in H 1/2 (Ω)
1/2

is all the space H 1/2 (Ω) (see Theorem 11.1 in [20]). That is, Cc∞ (Ω) is dense in H 1/2 (Ω).
1 (C) “vanish” on ∂Ω in the
However, in contrast with this, the traces in Ω of functions in H0,L
sense given by (2.10).
Recall that we have denoted by V0 (Ω) the space of traces on Ω × {0} of functions in H0,L
1 (C):


V0 (Ω) := u = trΩ v v ∈ H0,L
1
(C) ⊂ H 1/2 (Ω), (2.12)
2066 X. Cabré, J. Tan / Advances in Mathematics 224 (2010) 2052–2093

endowed with the norm (2.4) in Proposition 2.1. The dual space of V0 (Ω) is denoted by V0∗ (Ω),
equipped with the norm

gV0∗ (Ω) = sup u, g u ∈ V0 (Ω), uV0 (Ω)  1 .

Next we give the first characterization of the space V0 (Ω):

Lemma 2.7. Let V0 (Ω) be the space of traces on Ω × {0} of functions in H0,L
1 (C), as in (2.12).

Then, we have
  2 
u (x)
V0 (Ω) = u ∈ H (Ω)
1/2
dx < +∞ ,
d(x)
Ω

where d(x) = dist(x, ∂Ω).

Proof. The inclusion ⊂ follows from Lemma 2.6. Next we show the other inclusion. Let u ∈
H 1/2 (Ω) satisfy Ω u2 /d < ∞. Let ũ be the extension of u in all of Rn assigning ũ ≡ 0 in
Rn \ Ω. The quantity
  
|ũ(x) − ũ(x̄)|2 2
ũ2H 1/2 (Rn ) = dx d x̄ + ũ(x) dx
|x − x̄|n+1
Rn Rn Rn

can be bounded—using ũ ≡ 0 in Rn \ Ω—by a constant times

  2 1/2
u (x)
uH 1/2 (Ω) +
2
dx ,
d(x)
Ω

that we assume to be finite. Hence, ũ ∈ H 1/2 (Rn ) and thus ũ is the trace in Rn = ∂Rn+1
+ of a
function ṽ ∈ H 1 (Rn+1
+ ).
Next, we use a partition of the unity, and local bi-Lipschitz maps (defined below) sending
Rn+1
+ into Ω × [0, ∞) = C being the identity on Ω × {0} and mapping R \ Ω = (∂R+ ) \ Ω
n n+1

into ∂Ω × [0, ∞). By composing these maps with the function ṽ (cut off with the partition of
1 (C) function with u as trace on Ω × {0}, as desired.
unity), we obtain an H0,L
Finally we give a concrete expression for one such bi-Lipschitz map. First, consider the one-
dimensional case Ω = (0, ∞). Then simply take the bi-Lipschitz map

(x, y) ∈ (0, ∞) × (0, ∞) = Ω × (0, ∞)

 2 
x − y2 2xy
→  , ∈ R × (0, ∞),
x2 + y2 x2 + y2

whose Jacobian can be checked to be identically 2. In the general case, we can flatten the bound-
ary ∂Ω and use locally the previous map. 2

Next we consider, for a given function u ∈ V0 (Ω), the minimizing problem:

 X. Cabré, J. Tan / Advances in Mathematics 224 (2010) 2052–2093 2067

 
inf |∇v|2 dx dy v ∈ H0,L
1
(C), v(·, 0) = u in Ω . (2.13)
C

By the definition of V0 (Ω), the set of functions v where we minimize is nonempty. By lower
weak semi-continuity and by Lemma 2.5, we see that there exists a minimizer v. We will prove
next that this minimizer v is unique. We call v a weak solution of the problem
⎧
⎨ v = 0 in C,
v=0 on ∂L C, (2.14)
⎩
v=u on Ω × {0}.

That is, we have

Lemma 2.8. For u ∈ V0 (Ω), there exists a unique minimizer v of (2.13). The function v ∈
1 (C) is the harmonic extension of u (in the weak sense) to C vanishing on ∂ C.
H0,L L

Proof. By the definition of V0 (Ω), we have that, for every u ∈ V0 (Ω), there exists at least one
1 (C) such that tr (w) = u. Then the standard minimization argument gives (using lower
w ∈ H0,L Ω
semi-continuity and Lemma 2.5) the existence of a minimizer. The uniqueness of minimizer
follows automatically from the identity of the parallelogram used for two possible minimizers v1
and v2 ,
   
v1 − v2 1 1 v1 + v2
0J = J (v1 ) + J (v2 ) − J  0,
2 2 2 2

where J (v) = C |∇v|2 dx dy, which leads to v1 = v2 . 2

By Lemma 2.8, there exists a unique function v ∈ H0,L

1 (C) which is the harmonic extension

of u in C vanishing on ∂L C, and that we denote by

v := h-ext(u).

It is easy to see that for every η ∈ C ∞ (C) ∩ H 1 (C) and η ≡ 0 on ∂L C,

 
∂v
∇v∇η dx dy = η dx. (2.15)
∂ν
C Ω

By Lemma 2.6, there exists a constant C such that for every u ∈ V0 (Ω),
 
uV0 (Ω)  C h-ext(u)H 1 (C ) . (2.16)
0,L

Next, note that the h-ext operator is bijective from V0 (Ω) to the subspace H of H0,L
1 (C) formed

by all harmonic functions in H0,L1 (C). Since both V (Ω) and H are Banach spaces, the open
0
mapping theorem gives that we also have the reverse inequality to (2.16), i.e., there exists a
constant C such that
2068 X. Cabré, J. Tan / Advances in Mathematics 224 (2010) 2052–2093

 
h-ext(u) 1 (C )  CuV0 (Ω) , (2.17)
H0,L

for all u ∈ V0 (Ω). From this we deduce the following. Given a smooth ξ ∈ V0 (Ω), consider
the h-ext(ξ ) and call it η. Now, we use (2.15) and (2.17) (for u and ξ ) to obtain | Ω ∂ν
∂v
ξ dx| 
∗
CuV0 (Ω) ξ V0 (Ω) . That is, ∂ν |Ω ∈ V0 (Ω) and there is the bound:
∂v

 
∂ 
 h-ext(u)  CuV0 (Ω) .
 ∂ν 
V0∗ (Ω)

Hence we have

Lemma 2.9. The operator A1/2 : V0 (Ω) → V0∗ (Ω) defined by

∂v
A1/2 u := , (2.18)
∂ν Ω×{0}

where v = h-ext(u) ∈ H0,L

1 (C) is the harmonic extension of u in C vanishing on ∂ C, is linear
L
and bounded from V0 (Ω) to V0∗ (Ω).

We now give the spectral representation of A1/2 and the corresponding structure of the space
V0 (Ω).

Lemma 2.10.

(i) Let {ϕk } be an orthonormal basis of L2 (Ω) forming a spectral decomposition of − in

Ω with Dirichlet boundary conditions as in (2.3), with {λk } the corresponding Dirichlet
eigenvalues of − in Ω. Then, we have

∞ ∞

  1/2
V0 (Ω) = u = bk ϕk ∈ L (Ω)2
bk2 λk < +∞ .
k=1 k=1
∞
(ii) Let u ∈ V0 (Ω). Then we have, if u = k=1 bk ϕk ,

∞

bk λk ϕk ∈ V0∗ (Ω).
1/2
A1/2 u =
k=1

∞ Let u ∈ V0 (Ω), which is contained in L (Ω). Let its expansion be written by u(x) =
Proof. 2

k=1 bk ϕk (x). Consider the function

∞
  1/2 
v(x, y) = bk ϕk (x) exp −λk y , (2.19)
k=1

which is clearly smooth for y > 0. Observe that v(x, 0) = u(x) in Ω and, for y > 0,
 X. Cabré, J. Tan / Advances in Mathematics 224 (2010) 2052–2093 2069

∞
   1/2   1/2 
v(x, y) = bk −λk ϕk (x) exp −λk y + λk ϕk (x) exp −λk y = 0.
k=1

Thus, v is a harmonic extension of u. We will have that v = h-ext(u), by uniqueness, once we

find the condition on {bk } for v to belong to H0,L
1 (C). But such condition is simple. Using (2.19)

and that {ϕk } are eigenfunctions of − and orthonormal in L2 (Ω), we have

∞ ∞

|∇v| dx dy =
2
|∇x v|2 + |∂y v|2 dx dy
0 Ω 0 Ω

∞
 ∞
 1/2 
=2 bk2 λk exp −2λk y dy
k=1 0
∞
 ∞

1 1/2
=2 bk2 λk 1/2
= bk2 λk .
k=1 2λk k=1

∞ 2 1/2
This means that v ∈ H0,L 1 (C) if and only if
k=1 bk λk < ∞. Therefore, this condition on {bk }
is equivalent to u ∈ V0 (Ω).
Assertion (ii) follows from the direct computation of − ∂v
∂y |y=0 using (2.19). 2

In functional analysis, the classical spectral decomposition holds for self-adjoint compact
operators, such as the Dirichlet inverse Laplacian (−)−1 : L2 (Ω) → L2 (Ω). This is the reason
why we now define, with the aid of the Lax–Milgram theorem, a compact operator B1/2 which
will be the inverse of A1/2 .

Definition 2.11. Define the operator B1/2 : V0∗ (Ω) → V0 (Ω), by g → trΩ v, where v is found by
solving the problem:

⎧ v = 0 in C,
⎪
⎨
v=0 on ∂L C, (2.20)
⎪
⎩ ∂v = g(x) on Ω × {0},
∂ν

as we indicate next.

We say that v is a weak solution of (2.20) whenever v ∈ H0,L

1 (C) and


 
∇v∇ξ dx dy = g, ξ(·, 0) (2.21)
C

for all ξ ∈ H0,L

1 (C). We see that there exists a unique weak solution of (2.20) by the Lax–Milgram
1 (C):
theorem, via studying the corresponding functional in H0,L
2070 X. Cabré, J. Tan / Advances in Mathematics 224 (2010) 2052–2093


1  
I (v) = |∇v|2 dx dy − g, v(·, 0) ,
2
C

where g ∈ V0∗ (Ω) is given. Observe that the operator B1/2 is clearly the inverse of the operator
A1/2 .
On the other hand, let us compute B1/2 ◦ B1/2 |L2 (Ω) . Here note that since V0 (Ω) ⊂ L2 (Ω),
we have L2 (Ω) ⊂ V0∗ (Ω). For a given g ∈ L2 (Ω), let ϕ ∈ H01 (Ω) ∩ H 2 (Ω) be the solution of
Poisson’s problem for the Laplacian

−ϕ = g in Ω,
ϕ=0 on ∂Ω.

Since H01 (Ω) ⊂ V0 (Ω) (for instance, by Lemma 2.10), there is a unique harmonic extension
ψ ∈ H0,L
1 (C) of ϕ in C such that

⎧
⎨ ψ = 0 in C,
ψ =0 on ∂L C,
⎩
ψ =ϕ on Ω × {0}.

Moreover, ψ̃(x, y) := ψ(x, y) − ϕ(x) solves

⎧
⎨ −ψ̃ = ϕ = −g(x) in C,
ψ̃ = 0 on ∂L C,
⎩
ψ̃ = 0 on Ω × {0}.

Considering the odd reflection ψ̃od of ψ̃ across Ω × {0}, and the function

g(x), y  0,
god (x, y) =
−g(x), y < 0,

we have

−ψ̃od = −god in Ω × R,
ψ̃od = 0 on ∂Ω × R.

Therefore, since god ∈ L2 (Ω × (−2, 2)), we deduce ψ̃od ∈ H 2 (Ω × (−1, 1)) and hence ψ ∈
H 2 (Ω × (0, 1)). We deduce, by the smoothness of the harmonic function ψ for y > 0 and by its
exponential decay in y—see (2.19)—that ψ ∈ H0,L1 (C) ∩ H 2 (C).

It follows that −∂y ψ ∈ H0,L (C) solves

1


(−∂y ψ) = 0 in C,
−∂y ψ = 0 on ∂L C,

and
∂
(−∂y ψ) = ∂yy ψ = −x ψ = −ϕ = g on Ω × {0}.
∂ν
 X. Cabré, J. Tan / Advances in Mathematics 224 (2010) 2052–2093 2071

Since V0 (Ω) ⊂ L2 (Ω), we have that g ∈ L2 (Ω) ∼ = L2 (Ω)∗ ⊂ V0∗ (Ω), and we deduce that the so-
lution v ∈ V0 (Ω) of (2.20) is v = −∂y ψ , because of the uniqueness of H0,L
1 (C) solution of (2.20).

In particular, B1/2 g = v(·, 0) = −∂y ψ(·, 0). On the other hand, since ψ ∈ H0,L
1 (C) solves

⎧
⎪
⎪ ψ = 0 in C,
⎨
ψ =0 on ∂L C,
⎪
⎪ ∂ψ
⎩ ≡ −∂y ψ(·, 0) = v(·, 0) = B1/2 g on Ω × {0},
∂ν

we conclude that

(B1/2 ◦ B1/2 )g = B1/2 v(·, 0) = ψ(·, 0) = ϕ = (−)−1 g.

Summarizing the above argument, we have:

Proposition 2.12. B1/2 ◦ B1/2 |L2 (Ω) = (−)−1 : L2 (Ω) → L2 (Ω), where (−)−1 is the inverse
Laplacian in Ω with zero Dirichlet boundary conditions.

Note that B1/2 : L2 (Ω) → L2 (Ω) is a self-adjoint operator. In fact, since for v1 , v2 ∈ H0,L
1 (C),

   
∂v1 ∂v2
(v2 v1 − v1 v2 ) dx dy = v2 − v1 dx,
∂ν ∂ν
C Ω

we see
 
B1/2 g2 · g1 dx = B1/2 g1 · g2 dx
Ω Ω

and
 
v2 (x, 0)A1/2 v1 (x, 0) dx = v1 (x, 0)A1/2 v2 (x, 0) dx.
Ω Ω

On the other hand, by using (2.21) with ξ = v and Lemma 2.5, we obtain that B1/2 is a positive
compact operator in L2 (Ω). Hence by the spectral theory for self-adjoint compact operators, we
have that all the eigenvalues of B1/2 are real, positive, and that there are corresponding eigenfunc-
tions which make up an orthonormal basis of L2 (Ω). Furthermore, such basis and eigenvalues
are explicit in terms of those of the Laplacian with Dirichlet boundary conditions, since (−)−1
has B1/2 as unique, positive and self-adjoint square root, by Proposition 2.12. Summarizing:

Proposition 2.13. Let {ϕk } be an orthonormal basis of L2 (Ω) forming a spectral decomposi-
tion of − in Ω with Dirichlet boundary conditions, as in (2.3), with {λk } the corresponding
Dirichlet eigenvalues of − in Ω. Then, for all k  1,
2072 X. Cabré, J. Tan / Advances in Mathematics 224 (2010) 2052–2093

 1/2
A1/2 ϕk = λk ϕk in Ω,
(2.22)
ϕk = 0 on ∂Ω.

1/2
In particular, {ϕk } is also a basis formed by the eigenfunctions of A1/2 , with eigenvalues {λk }.

Proof of Proposition 2.1. It follows from Lemma 2.7 and Lemma 2.10. 2

Proof of Proposition 2.2. It follows from Lemma 2.8, Lemma 2.10 and its proof, and Proposi-
tions 2.12 and 2.13. 2

3. Regularity of solutions

In this section we study the regularity of weak solutions for linear and nonlinear problems
involving A1/2 . First we consider the linear problem

A1/2 u = g(x) in Ω,
(3.1)
u=0 on ∂Ω,

where g ∈ V0∗ (Ω) and Ω is a smooth bounded domain in Rn . By the construction of the previous
section, the precise meaning of (3.1) is that u = trΩ v, where the function v ∈ H0,L
1 (C) with

v(·, 0) = u ∈ V0 (Ω) satisfies

⎧ v = 0 in C,
⎪
⎨
v=0 on ∂L C, (3.2)
⎪
⎩ ∂v = g(x) on Ω × {0}.
∂ν

We will say then that v is a weak solution of (3.2) and that u is a weak solution of (3.1).
Most of this section contains the proof of the following analogues of the W 2,p -estimates of
Calderón–Zygmund and of the Schauder estimates.

Proposition 3.1. Let α ∈ (0, 1), Ω be a C 2,α bounded domain of Rn , g ∈ V0∗ (Ω), v ∈ H0,L
1 (C)

be the weak solution of (3.2), and u = trΩ v be the weak solution of (3.1). Then,

(i) If g ∈ L2 (Ω), then u ∈ H01 (Ω).

(ii) If g ∈ H01 (Ω), then u ∈ H 2 (Ω) ∩ H01 (Ω).
(iii) If g ∈ L∞ (Ω), then v ∈ W 1,q (Ω × (0, R)) for all R > 0 and 1 < q < ∞. In particular,
v ∈ C α (C) and u ∈ C α (Ω).
(iv) If g ∈ C α (Ω) and g|∂Ω ≡ 0, then v ∈ C 1,α (C) and u ∈ C 1,α (Ω).
(v) If g ∈ C 1,α (Ω) and g|∂Ω ≡ 0, then v ∈ C 2,α (C) and u ∈ C 2,α (Ω).

As a consequence, we deduce the regularity of bounded weak solutions to the nonlinear prob-
lem

A1/2 u = f (u) in Ω,
(3.3)
u=0 on ∂Ω.
 X. Cabré, J. Tan / Advances in Mathematics 224 (2010) 2052–2093 2073

As before, the precise meaning for (3.3) is that v ∈ H0,L

1 (C), v(·, 0) = u, and v is a weak solution

of
⎧ v = 0 in C,
⎪
⎨
v=0 on ∂L C, (3.4)
⎩ ∂v = f v(·, 0)
⎪
on Ω × {0}.
∂ν

Here the weak solution u is assumed to be bounded. Regularity results for subcritical and critical
problems and for weak solutions which are not assumed a priori to be bounded will be proved in
Section 5.
By C0 (Ω) we denote the space of continuous functions in Ω vanishing on the boundary ∂Ω.
In the following result, f (0) = 0 is assumed to ensure the C 1 (Ω) regularity of solutions of (3.3).

Proposition 3.2. Let α ∈ (0, 1), Ω be a C 2,α bounded domain of Rn , and f be a C 1,α function
such that f (0) = 0. If u ∈ L∞ (Ω) is a weak solution of (3.3), and thus v ∈ H0,L1 (C) ∩ L∞ (C) is

a weak solution of (3.4), then u ∈ C 2,α (Ω) ∩ C0 (Ω). In addition, v ∈ C 2,α (C).

Proof. By (iii) of Proposition 3.1 we have that u ∈ C α (Ω). Next, by (iv) of Proposition 3.1 and
since on ∂Ω × {0}, g := f (v(·, 0)) = f (0) = 0, we have u ∈ C 1,α (Ω). Finally, v ∈ C 2,α (C) and
u ∈ C 2,α (Ω) from (v) of Proposition 3.1 since g = f (u) vanishes on ∂Ω and it is of class C 1,α ,
since both f and u are C 1,α . 2

Proof of Proposition 3.1. (i) and (ii). Both statements follow immediately from Propositions 2.1
1/2
and 2.2. Simply use that {ϕk } is an orthonormal basis of L2 (Ω) and that {ϕk /λk } is an orthonor-
mal basis of H01 (Ω). For part (ii), note that if A1/2 u = g ∈ H01 (Ω), then we have u ∈ L2 (Ω).
(iii) Let v be a weak solution of (3.2). We proceed with a useful method, introduced by Cabré
and Solà-Morales in [5], which consists of using the auxiliary function

y
w(x, y) = v(x, t) dt for (x, y) ∈ C. (3.5)
0

Since (w)y = 0 in C, we have that w is independent of y. Hence we can compute it on

{y = 0}. On {y = 0}, since w ≡ 0, we have w = wyy = vy . Thus w is a solution of the Dirichlet
problem

−w(x, y) = g(x) in C,
(3.6)
w=0 on ∂C.

We extend w to the whole cylinder Ω × R by odd reflection:


w(x, y) for y  0,
wod (x, y) =
−w(x, −y) for y  0.

Moreover, we put
2074 X. Cabré, J. Tan / Advances in Mathematics 224 (2010) 2052–2093


g(x) for y > 0,
god (x, y) =
−g(x) for y < 0.

Then we obtain

−wod = god in Ω × R,
(3.7)
wod = 0 on ∂Ω × R.

Since god ∈ Lq (Ω × (−2R, 2R)) for all R > 0 and 1 < q < ∞, regularity for the Dirichlet
problem (3.7) gives wod ∈ W 2,q (Ω × (−R, R)) for all R > 0 and 1 < q < ∞. In particular,
w ∈ C 1,α (C). Therefore, v = wy ∈ C α (C) and u ∈ C α (Ω).
(iv) Choose a smooth domain H such that Ω ⊂ H , and let

g in Ω,
gH =
0 in H \ Ω.

We have that gH ∈ C α (H ), since g|∂Ω = 0, by assumption. Consider the weak solution vH of

⎧ v = 0 in H × (0, ∞),
⎪
⎨
H
vH = 0 on ∂H × [0, ∞),
⎪
⎩ ∂vH = g (x)
H on H × {0}.
∂ν
Consider also the auxiliary function

y
wH (x, y) = vH (x, t) dt in H × [0, ∞),
0

which solves problem (3.6) with Ω and g replaced by H and gH .

Using boundary regularity theory (but away from the corners of H × [0, ∞)) for this Dirichlet
problem, we see that wH is C 2,α (H × (0, ∞)) (again, here we do not claim regularity at the
corners ∂H × {0}). Thus, wH ∈ C 2,α (C) (here instead we include the corners ∂Ω × {0} of C).
Consider the difference ϕ = wH − w in C, where w is defined by (3.5). It is clear that
⎧
⎨ ϕ = 0 in C,
ϕ = wH on ∂L C,
⎩
ϕ=0 on Ω × {0}.

We extend ϕ to the whole cylinder Ω × R by odd reflection:


ϕ(x, y) for y  0,
ϕod (x, y) =
−ϕ(x, −y) for y  0.

Moreover, we put

wH (x, y) for y > 0,
wH,od (x, y) =
−wH (x, −y) for y  0.
 X. Cabré, J. Tan / Advances in Mathematics 224 (2010) 2052–2093 2075

Then we have

ϕod = 0 in Ω × R,
(3.8)
ϕod = wH,od on ∂Ω × R.

Since wH ∈ C 2,α (C), wH ≡ 0 on ∂Ω × {0}, and ∂yy wH = ∂y vH = −gH = −g = 0 on ∂Ω ×

{0}, we deduce that wH,od ∈ C 2,α (∂Ω × R). It follows from elliptic regularity for (3.8) that
ϕod ∈ C 2,α (Ω × R). Thus, ϕ ∈ C 2,α (C), w ∈ C 2,α (C) and v = ∂y w ∈ C 1,α (C).
(v) Choose a smooth bounded domain B such that Ω ⊂ B. B could be the same as H in (ii),
for instance a ball, but we change its name for notation clarity. Since g ∈ C 1,α (Ω), there exists
an extension gB ∈ C 1,α (B); see [15]. Consider the solution vB of
⎧ v = 0 in B × (0, ∞),
⎪
⎨
B
vB = 0 on ∂B × [0, ∞),
⎪
⎩ ∂vB = g
B on B × {0}.
∂ν
Consider the auxiliary function

y
wB (x, y) = vB (x, t) dt in B × [0, ∞).
0

As before, from interior boundary regularity for the Dirichlet problem of the type (3.6)
satisfied by wB , we obtain that wB ∈ C 3,α (B × [0, ∞)) since gB ∈ C 1,α (B) (away from the
corners ∂B × {0}). Thus, vB ∈ C 2,α (B × [0, ∞)). Thus, vB ∈ C 2,α (C). Consider the difference
ψ = vB − v in C, where v is a weak solution of (3.2). We have that ψ = vB − v satisfies
⎧
⎪
⎪ ψ = 0 in C,
⎨
ψ = vB on ∂L C,
⎪
⎪ ∂ψ
⎩ = 0 on Ω × {0}.
∂ν

We extend ψ to the whole cylinder Ω × R now by even reflection:


ψ(x, y) for y  0,
ψev (x, y) =
ψ(x, −y) for y  0.

Moreover, we put

vB (x, y) for y > 0,
vB,ev (x, y) =
vB (x, −y) for y  0.

∂ψ
Then, since ∂ν = 0 on Ω × {0}, we have

ψev = 0 in Ω × R,
ψev = vB,ev on ∂Ω × R.
2076 X. Cabré, J. Tan / Advances in Mathematics 224 (2010) 2052–2093

Since vB ∈ C 2,α (C), −∂y vB = gB = g = 0 on ∂Ω × {0}, we deduce that vB,ev ∈ C 2,α (∂Ω × R).
Therefore, it follows from classical regularity that ψev ∈ C 2,α (Ω × R). Thus, ψ ∈ C 2,α (C), and
v ∈ C 2,α (C). 2

4. Maximum principles

In this section we establish several maximum principles for A1/2 . We denote by C0 (Ω) the
space of continuous functions in Ω vanishing on the boundary ∂Ω. For convenience, we state
the results for functions in C0 (Ω) ∩ C 2 (Ω) (a space contained in H01 (Ω) ⊂ V0 (Ω)), but this can
be weakened.
The first statement is the weak maximum principle.

Lemma 4.1. Assume that u ∈ C 2 (Ω) satisfies


A1/2 u + c(x)u  0 in Ω,
u=0 on ∂Ω,

where Ω is a smooth bounded domain in Rn and c  0 in Ω. Then, u  0 in Ω.

Proof. Consider the extension v = h-ext(u). If we prove that v  0 in C, then u  0 in Ω.

Suppose by contradiction that v is negative somewhere in C. Then, since v = 0 in C and v = 0
on ∂L C, we deduce that v is negative somewhere in Ω × {0} and that infC v < 0 is achieved at
some point (x0 , 0) ∈ Ω × {0}. Thus, we have

inf v = v(x0 , 0) < 0.

C

By Hopf’s lemma,

vy (x0 , 0) > 0.

It follows
∂v
= −vy (x0 , 0) = A1/2 v(x0 , 0) < 0.
∂ν
Therefore, since c  0,

A1/2 v(x0 , 0) + c(x0 )v(x0 , 0) < 0.

This is a contradiction with the hypothesis A1/2 u + c(x)u  0. 2

The next statement is the strong maximum principle for A1/2 .

Lemma 4.2. Assume that u ∈ C 2 (Ω) satisfies

⎧
⎨ A1/2 u + c(x)u  0 in Ω,
u0 in Ω,
⎩
u=0 on ∂Ω,
 X. Cabré, J. Tan / Advances in Mathematics 224 (2010) 2052–2093 2077

where Ω is a smooth bounded domain in Rn and c ∈ L∞ (Ω). Then, either u > 0 in Ω, or u ≡ 0

in Ω.

Proof. The proof is similar to that of Lemma 4.1. Consider v = h-ext(u). We observe that v  0
in C. Suppose that v ≡ 0 but u = 0 somewhere in Ω. Then there exists a minimum point (x0 , 0) ∈
Ω × {0} of v where v(x0 , 0) = 0. Then by Hopf’s lemma we see that A1/2 u(x0 ) = −vy (x0 , 0)
< 0. This implies that A1/2 u(x0 ) + c(x0 )u(x0 ) < 0, because of v(x0 , 0) = u(x0 ) = 0. 2

Next we establish Hopf’s lemma for A1/2 , following the proof from [9].

Lemma 4.3. Let Ω be a bounded domain in Rn and c ∈ L∞ (Ω).

(i) Assume that Ω is smooth and that 0 ≡ u ∈ C 2 (Ω) satisfies

A1/2 u + c(x)u  0 in Ω,
u0 in Ω,
u=0 on ∂Ω.
∂u
Then, ∂ν0 < 0 on ∂Ω, where ν0 is the unit outer normal to ∂Ω.
(ii) Assume that P ∈ ∂Ω and that ∂Ω is smooth in a neighborhood of P . Let 0 ≡ v ∈ C 2 (C) ∩
L∞ (C), where C = Ω × (0, ∞), satisfy
⎧ v = 0 in C,
⎪
⎨
v0 on ∂L C,
⎪
⎩ ∂v
+ c(x)v  0 on Ω × {0}.
∂ν
∂v(P ,0)
If v(P , 0) = 0, then ∂ν0 < 0, where ν0 is the unit outer normal in Rn to ∂Ω.

Proof. We follow the proof given in [9]. Note that statement (i) is a particular case of (ii). Thus,
we only need to prove (ii).
Step 1. We shall first prove the lemma in the case c ≡ 0. Without loss of generality we may
assume that (P , 0) = P1 = (b1 , 0, . . . , 0) ∈ ∂Ω × {0}, b1 > 0 and ν0 = (1, 0, . . . , 0). Hence we
need to prove
∂v(P1 )
< 0.
∂x1

Since Ω is smooth in a neighborhood of P , there is a half-ball in Rn+1

+ included in the domain C,
such that P1 is the only point in the closed half-ball belonging also to ∂L C. Let P2 ∈ Ω × {0} and
r > 0 be the center and radius of such ball. Then we have P2 = (b2 , 0, . . . , 0) ∈ Ω × {0}. Denote

Br+ (P2 ) := z = (x, y) |z − P2 | < |P1 − P2 | =: r, y > 0 ⊂ C,
+ 
Br/2 (P2 ) := z = (x, y) |z − P2 | < |P1 − P2 |/2, y > 0 ,
+
A = Br+ (P2 ) \ Br/2 (P2 ).

Recall that P1 ∈ ∂Br+ (P2 ) ∩ (∂Ω × {0}).

2078 X. Cabré, J. Tan / Advances in Mathematics 224 (2010) 2052–2093

Consider the function on A:

   
ϕ(z) = exp −λ|z − P2 |2 − exp −λ|P1 − P2 |2 ,

with λ > 0 to be determined later. Note that

 
ϕ = exp −λ|z − P2 |2 4λ2 |z − P2 |2 − 2(n + 1)λ .

We can choose λ > 0 large enough such that ϕ  0 in A.

On the other hand, by Lemma 4.2, we see that v > 0 in A \ {P1 }. Hence, since ϕ ≡ 0 on
∂Br+ (P2 ) ∩ {y > 0}, we can take ε > 0 such that

v − εϕ  0 on ∂A ∩ {y > 0}.

Since −(v − εϕ)  0 in A, and

∂v
−∂y (v − εϕ) =  0 on ∂A ∩ {y = 0}
∂ν

(recall that c ≡ 0), by the maximum principle as in Lemma 4.1 we obtain

v − εϕ  0 in A.

Thus, from v − εϕ = 0 at P1 we see that ∂x1 (v − εϕ)(P1 )  0. Therefore, ∂x1 v(P1 ) 

ε∂x1 ϕ(P1 ) = −2λ(b1 − b2 )e−λ|P1 −P2 | < 0. Thus we have the desired result.
2

Step 2. In the case c ≡ 0, we define the function w = v exp(−βy) for some β > 0 to be
determined. From a direct calculation, we see that

−w − 2β∂y w = β 2 w  0 in C

and, choosing β  cL∞ (Ω) ,

 
−∂y w  β − c(x) w  0 on Ω × {0}.

Now we can apply to w the same approach as in Step 1, with  replaced by  + 2β∂y , and
obtain the assertion. 2

Finally, we establish a maximum principle for A1/2 in domains of small measure. Note that
in part (ii) of its statement, the hypothesis on small measure is made only on the base Ω of the
cylinder C.
 X. Cabré, J. Tan / Advances in Mathematics 224 (2010) 2052–2093 2079

Proposition 4.4.

(i) Assume that u ∈ C 2 (Ω) satisfies


A1/2 u + c(x)u  0 in Ω,
u=0 on ∂Ω,

where Ω is a smooth bounded domain in Rn and c ∈ L∞ (Ω). Then, there exists δ > 0
depending only on n and c− L∞ (Ω) , such that if |Ω ∩ {u < 0}|  δ, then u  0 in Ω.
(ii) Assume that Ω is a bounded (not necessary smooth) domain of Rn and c ∈ L∞ (Ω). Let
v ∈ C 2 (C) ∩ L∞ (C), where C = Ω × (0, ∞), satisfy
⎧ v = 0 in C,
⎪
⎨
v0 on ∂L C,
⎪
⎩ ∂v + c(x)v  0 on Ω × {0}.
∂ν

Then, there exists δ > 0 depending only on n and c− L∞ (Ω) , such that if

Ω ∩ v(·, 0) < 0  δ

then v  0 in C.

Proof. For part (i) of the theorem, consider v = h-ext(u). We see that v satisfies the assumptions
on part (ii) of the theorem. Hence, it is enough to prove part (ii). For this, let v − = max{0, −v}
 0. Since v − = 0 on ∂Ω × [0, ∞), we see
  
− − ∂v
0 = v v dx dy = v dx + |∇v − |2 dx dy.
∂ν
C Ω×{0} C

Then,
 
∂v
|∇v − |2 dx dy = − v− dx
∂ν
C Ω×{0}
 
−
 2
 v cv dx = −c v − dx
Ω×{0} Ω

 2
 c− v − (·, 0) dx
Ω∩{v − (·,0)>0}
  2
 Ω ∩ v − (·, 0) > 0 c− L∞ (Ω) v − (·, 0)L2n/(n−1) (Ω) .
1/n

Thus, extending v − by 0 outside C we obtain an H 1 (Rn+1 −

+ ) function and thus we have, if v ≡ 0,

|∇v − |2 dx dy − 2
Rn+1
+ C |∇v | dx dy
0 < S0  =
v − (·, 0)2L2n/(n−1) (Rn ) v − (·, 0)2L2n/(n−1) (Ω)
2080 X. Cabré, J. Tan / Advances in Mathematics 224 (2010) 2052–2093


 Ω ∩ v − (·, 0) > 0 c− L∞ (Ω) ,
1/n

where S0 is the best constant of the Sobolev trace inequality in Rn+1 −

+ . If |Ω ∩ {v (·, 0) > 0}| is
small enough, we arrive at a contradiction. 2

5. Subcritical case and L∞ estimate of Brezis–Kato type

In this section, we study the nonlinear problem (1.1) with f (u) = up in the subcritical and
critical cases. In the subcritical case we look for a function v(x, y) satisfying for x ∈ Ω and
y ∈ R+ ,
⎧ v = 0 in C = Ω × (0, ∞),
⎪
⎪
⎪
⎨v = 0 on ∂L C = ∂Ω × [0, ∞),
∂v (5.1)
⎪
⎪ = vp on Ω × {0},
⎪
⎩ ∂ν
v>0 in C,

where ν is the unit outer normal to C at Ω × {0} and 1 < p < 2 − 1 if n  2, or 1 < p < ∞ if
n = 1. If v is a solution of (5.1), then v(x, 0) = u(x) is a solution of (1.1) with the nonlinearity
f (u) = up .
In order to find a solution of (5.1) as stated in Theorem 1.1, we consider the following mini-
mization problem:
  
2 p+1
I0 = inf ∇v(x, y) dx dy v ∈ H0,L
1
(C), v(x, 0) dx = 1 .
C Ω

We show that I0 is achieved.

Proposition 5.1. Assume that 1 < p < 2 − 1 if n  2 or 1 < p < ∞ if n = 1. Then I0 is achieved
1 (C) by a nonnegative function v.
in H0,L

Proof. First, there is a function v ∈ H0,L

1 (C) such that

 
2 p+1
∇v(x, y) dx dy < ∞ and v(x, 0) dx = 1.
C Ω

In fact, it suffices to take any C ∞ function with compact support in Ω × [0, ∞) and not identi-
cally zero on Ω × {0}, and multiply it by an appropriate constant. Next we complete the proof
by weak lower semi-continuity of the Dirichlet integral and by the compact embedding property
in Lemma 2.5. Finally, note that |v|  0 is a nonnegative minimizer if v is a minimizer. 2

To establish the regularity of the minimizer just obtained, we prove an L∞ -estimate of Brezis–
Kato type by the technique of bootstrap for subcritical or critical nonlinear problems. Let g0 be
a Carathéodory function in Ω × R satisfying the growth condition
 
g0 (x, s)  C 1 + |s|p for all (x, s) ∈ Ω × R, (5.2)
 X. Cabré, J. Tan / Advances in Mathematics 224 (2010) 2052–2093 2081

where Ω is a smooth domain in Rn , 1  p  n+1

n−1 if n  2, or 1  p < ∞ if n = 1. We consider
the problem

⎧ v = 0 in C = Ω × (0, ∞),
⎪
⎨
v=0 on ∂L C = ∂Ω × [0, ∞), (5.3)
⎪
⎩ ∂v = g (·, v)
0 on Ω × {0}.
∂ν

Theorem 5.2. Let v ∈ H0,L1 (C) be a weak solution of (5.3) and assume the growth condition (5.2)

for g0 , with 1  p  n+1 ∞

n−1 if n  2, or 1  p < ∞ if n = 1. Then, v(·, 0) ∈ L (Ω).

Proof. The proof follows the one of Brezis–Kato for the Laplacian. First of all, let us rewrite the
condition on g0 as
 
g0 (x, v)  a(x) 1 + v(x, 0)

with the function

|g0 (x, v(x, 0))|

a(x) :=
1 + |v(x, 0)|

which satisfies
 p−1 
0  a  C 1 + v(·, 0) ∈ Ln (Ω),

2n
since v ∈ H0,L
1 (C), v(·, 0) ∈ L n−1 (Ω) and p − 1  2
n−1 .
Denote

Br+ = (x, y) (x, y) < r and y > 0 .

2β
For β  0 and T > 1, let ϕβ,T = vvT ∈ H0,L
1 (C) with v = min{|v|, T }. Denote
T


DT = (x, y) ∈ C v(x, y) < T .

By direct computation, we see

 
 β 2 2β
∇ vvT dx dy = vT |∇v|2 dx dy
C C

 
+ 2β + β 2 |v|2β |∇v|2 dx dy.
DT

Multiplying (5.3) by ϕβ,T and integrating by parts, we obtain

2082 X. Cabré, J. Tan / Advances in Mathematics 224 (2010) 2052–2093

 
2β
vT |∇v|2 dx dy + 2β |v|2β |∇v|2 dx dy
C DT
 
 2β  2β
= ∇v∇ vvT dx dy = g0 (x, v)vvT dx
C Ω×{0}

 2 2β
 a(x) 1 + |v| vT dx.
Ω×{0}

Combining these facts, we have

 
 β 2   2β
∇ vvT dx dy  C(β + 1) a(x) 1 + |v|2 vT dx,
C Ω×{0}

where C denotes different constants independent of T and of β. By Lemma 2.4, we deduce

  2/2 
β 2   2β
vvT dx  C(β + 1) a(x) 1 + |v|2 vT dx. (5.4)
Ω×{0} Ω×{0}

Assume that |v(·, 0)|β+1 ∈ L2 (Ω) for some β  0. Then we obtain that 2 2β
Ω×{0} |v| vT dx and
2β
Ω×{0} vT dx are bounded uniformly in T . In what follows, let C denote constants independent
of T —but that may depend on β and v(·, 0)β+1 L2 (Ω) . Given M0 > 0, we have
  
2β 2β 2β
a|v|2 vT dx  M0 |v|2 vT dx + a|v|2 vT dx
Ω×{0} Ω×{0} {aM0 }
  1/n   2/2
β 2
 CM0 + a n dx vvT dx
{aM0 } Ω×{0}
  2/2
β 2
 CM0 + ε(M0 ) vvT dx ,
Ω×{0}

2β
where ε(M0 ) = ( {aM0 } a n dx)1/n → 0 as M0 → ∞. Note that we can deal with Ω×{0} avT dx
in the analogue procedure. Therefore, we deduce from the last inequalities and (5.4), taking M0
large enough so that C(β + 1)ε(M0 ) = 12 , that

  2/2
β 2
vvT dx  C(1 + M0 ). (5.5)
Ω×{0}


Thus letting T → ∞, since C is independent of T , we obtain that |v(·, 0)|β+1 ∈ L2 (Ω). This
conclusion followed simply from assuming |v(·, 0)|β+1 ∈ L2 (Ω).
 X. Cabré, J. Tan / Advances in Mathematics 224 (2010) 2052–2093 2083

Hence, by iterating β0 = 0, βi + 1 = (βi−1 + 1) n−1

n
if i  1 in (5.5), we conclude that v(·, 0) ∈
Lq (Ω) for all q < ∞. Finally, the proof of part (iii) in Proposition 3.1—which only uses g ∈
Lq (Ω) for all q < ∞ and not g ∈ L∞ (Ω)—applied with g(x) = g0 (x, v(x, 0)), which satisfies
|g|  C(1 + |v(·, 0)|p ) ∈ Lq (Ω) for all q < ∞, leads to v(·, 0) ∈ C α (Ω) ⊂ L∞ (Ω). 2

Proof of Theorem 1.1. Proposition 5.1 gives the existence of a weak nonnegative solution v to
(5.1) after multiplying the nonnegative minimizer of I0 by a constant to take care of the Lagrange
multiplier. Then, Theorem 5.2 gives that v(·, 0) ∈ L∞ (Ω). Next, Proposition 3.2 gives that u ∈
C 2,α (Ω), since f (s) = |s|p is a C 1,α function for some α ∈ (0, 1). Finally, the strong maximum
principle, Lemma 4.2, leads to u > 0 in Ω. 2

6. A priori estimates for positive solutions

In this section we prove Theorem 1.3. Namely, we establish a priori estimates of Gidas–Spruck
type for weak solutions of
⎧
⎪
⎪ v = 0 in C = Ω × (0, ∞) ⊂ Rn+1
+ ,
⎪
⎪
⎨v = 0 on ∂L C = ∂Ω × [0, ∞),
∂v (6.1)
⎪
⎪ = vp on Ω × {0},
⎪
⎪
⎩ ∂ν
v>0 in C,

where Ω ⊂ Rn is a bounded smooth domain, n  2, and 1 < p < n+1 n−1 .

For this, we need two nonlinear Liouville theorems for problems involving the square root of
the Laplacian in unbounded domains—one in the whole space, another in the half-space. The
first one was proved by Y.Y. Li, Zhang and Zhu in [18,17] and Ou in [22]. Its statement is the
following—and it is equivalent to Theorem 1.4 in the Introduction.

Theorem 6.1. (See [18,17,22].) For n  2 and 1 < p < 2 − 1 = n+1

n−1 , there exists no weak
solution of the problem

⎧
⎪
⎪ v = 0 in Rn+1
+ ,
⎨
∂v
= vp on ∂Rn+1
+ , (6.2)
⎪
⎪ ∂ν
⎩
v>0 in Rn+1
+ .

We need to prove an analogue nonlinear Liouville type result involving the square root of
− with Dirichlet boundary value in the half-space. This is Theorem 1.5 of the Introduction and
Proposition 6.3 in this section. As we will see, this nonlinear Liouville theorem in Rn+ will be
first reduced to the one-dimensional case R+ , by using the moving planes method. After this,
we prove that there exists no positive bounded solution for the nonlinear Neumann boundary
problem in the quarter R2++ , which corresponds to the nonlinear Liouville theorem involving
the square root of − with Dirichlet boundary value in the half-line; see Proposition 6.4. To
complete the proof of Theorem 1.5 we will use the following Liouville theorem in dimension
n + 1 = 2.
2084 X. Cabré, J. Tan / Advances in Mathematics 224 (2010) 2052–2093

Proposition 6.2. (See [9].) Suppose that v weakly solves

⎧
⎪
⎪ −v  0 in R2+ ,
⎨
∂v
0 on ∂R2+ , (6.3)
⎪
⎪ ∂ν
⎩
v0 in R2+ .

Then, v is a constant.

As usual, very strong Liouville theorems (but quite simple to prove) hold in low dimensions,
but not in higher ones. Compare (6.3) in low dimensions for supersolutions of the homogeneous
linear problem with (6.2) for solutions of a precise nonlinear problem. The proof of Proposi-
tion 6.2 in [9] compared in an appropriate way the solution v with log(| · |). For completeness,
we give here an alternative proof.

Proof of Proposition 6.2. Replacing v by v − infR2 v  0, we may assume infR2 v = 0. Letting

+ +
w = 1 − v, we have
⎧
⎪
⎪ −w  0 in R2+ ,
⎨
∂w
0 on ∂R2+ , (6.4)
⎪
⎪ ∂ν
⎩
w1 in R2+ .

In addition, supR2 w = 1. Let ξR ∈ C ∞ (R2 ) be a function with compact support in B2R (0), equal
+
to 1 in BR (0), and with |∇ξR |  C
R. Let

+ 
DR,2R := (x, y) ∈ R2 R  (x, y)  2R, y > 0 .

Multiplying the first equation in (6.4) by w + ξR2 , integrating in R2+ and using the Neumann con-
dition and w +  1, we see that
 
ξR2 ∇w + ξR ∇ξR w + ∇w +
2
2 (6.5)
+
R2+ DR,2R
  1/2   1/2
ξR2 ∇w +
2
C |∇ξR |2
+ +
DR,2R DR,2R
  1/2
+ 2
C ξR2 ∇w . (6.6)
+
DR,2R

This leads, letting R ↑ ∞, to R2 |∇w + |2 < ∞. As a consequence of this, the integral in (6.6)
+
tends to zero as R → ∞. Thus, by (6.5) and (6.6),
 X. Cabré, J. Tan / Advances in Mathematics 224 (2010) 2052–2093 2085


∇w +
2
= 0.
R2+

Thus, w + is constant, and since supR2 w = 1, we conclude w ≡ 1. 2

+

Proposition 6.3. Let n  2 and


Rn+1
++ = z = (x1 , x2 , . . . , xn , y) xn > 0, y > 0 .

Assume that v is a classical solution of

⎧
⎪
⎪ v = 0 in Rn+1
++ ,
⎪
⎪
⎪
⎨v = 0 on {xn = 0, y  0},
∂v (6.7)
⎪
⎪ = vp on {xn > 0, y = 0},
⎪
⎪ ∂ν
⎪
⎩
v>0 in Rn+1
++ ,

where 1  p  n+1
n−1 . Then, v depends only on xn and y.

Proof. We shall follow the steps of [14]. Let en = (0, . . . , 0, 1, 0) and N = n + 1. Consider the
conformal transformation

z + en
z̄ = T (z) =
|z + en |2

and the Kelvin transformation w of v

w(z̄) = |z + en |N −2 v(z) = |z̄|2−N v(z).

+ en + en + en
Denote B1/2 ( 2 ) := {z̄ = (x̄, ȳ) | |z̄ − 12 en | < 12 , ȳ > 0}, S1/2 ( 2 ) := ∂B1/2 ( 2 ) ∩ {ȳ > 0},
+ en
Γ0,1/2 := ∂B1/2 ( 2 ) ∩ {ȳ = 0}.
+ en
Note that, through T , Rn+1 ++ = {xn > 0, y > 0} gets mapped into the half-ball B1/2 ( 2 ), the
boundary {xn > 0, y = 0} becomes the ball Γ0,1/2 , {xn = 0, y  0} goes to the half-sphere
+ en
S1/2 ( 2 ), and the infinity goes to z̄ = 0.
We see that w satisfies
⎧  
⎪
⎪ + en
⎪ w = 0
⎪ inB1/2 ,
⎪
⎪ 2
⎪
⎪  
⎪
⎪ en
⎪
⎨w = 0 +
on S1/2 ,
2
⎪
⎪ ∂w(z̄)
⎪
⎪ = |z̄|p(N −2)−N w p (z̄) on Γ0,1/2 ,
⎪
⎪ ∂ν
⎪
⎪  
⎪
⎪ en
⎪
⎩w > 0 +
in B1/2 .
2
2086 X. Cabré, J. Tan / Advances in Mathematics 224 (2010) 2052–2093

Since |z̄|p(n−1)−(n+1) is nonincreasing in the z̄i direction for all i = 1, . . . , n − 1 (in fact, in any
direction orthogonal to the z̄n -axis), the moving planes method used as in [9] gives that w is
symmetric about all the z̄i -axis for i = 1, . . . , n − 1. This leads to w = w(|z̄ |, z̄n , ȳ), where z̄ =
(z̄1 , . . . , z̄n−1 ) and hence v = v(|x |, xn , y). Now, since we may perform the Kelvin’s transform
with respect to any point (−x0 , −1, 0)—and not only with respect to x0 = 0 as before—we
conclude that v = v(xn , y) as claimed. 2

Proposition 6.4. Assume that f is a C 1,α function for some α ∈ (0, 1), such that f > 0 in (0, ∞)
and f (0) = 0. Let C be a positive constant. Then there is no bounded solution of the problem
⎧
⎪ v = 0 in R2++ = {x > 0, y > 0},
⎪
⎪
⎪
⎨v = 0 on {x = 0, y  0},
∂v (6.8)
⎪
⎪ = f (v) on {x > 0, y = 0},
⎪ ∂ν
⎪
⎩
0<vC in R2++ .

Proof. We use some tools developed in [5].

Suppose by contradiction that there is such solution v. First, we claim that v(x, 0) → 0 as
x → ∞. Suppose by contradiction that there exists a sequence am → ∞ (m → ∞) such that
v(am , 0) → α > 0. Let vm (x, y) := v(x + am , y). It is clear that vm is a solution of (6.8) in
Um := {(x, y) | x > −am , y > 0}. Moreover, vm (0, 0) = v(am , 0) → α. Therefore there exists a
subsequence, still denoted by vm , such that vm → w in Cloc2 (R2 ) as m → ∞, and w is a solution
+
of
⎧ 
⎪
⎪ ∂ w + ∂yy w = 0 in R2+ = (x, y) | y > 0 ,
⎨ xx
∂w (6.9)
⎪ = f (w)  0 on {y = 0},
⎪
⎩ ∂ν
0wC in {y > 0}.

Notice that

w(0, 0) = α > 0.

On the other hand, by Proposition 6.2 we know that w is identically constant. This is impossible
due to the nonlinear Neumann condition, since f > 0 in (0, ∞) and f (w(0, 0)) = f (α) > 0. We
conclude the claim, that is, v(x, 0) → 0 as x → +∞.
Note that we can reflect the function v with respect to {x = 0, y > 0}, ṽ(x, y) = −v(−x, y)
for x < 0, and obtain a bounded harmonic function ṽ in all R2+ = {y > 0}, since v ≡ 0 on {x =
0, y > 0}. Applying interior gradient estimates to the bounded harmonic function ṽ in the ball
Bt (x, t) ⊂ R2+ , we obtain

Cv∞ C 1
∇v(x, t)   , for all t > , x > 0.
t t 2

On the other hand, by the results of [5] applied to the solution ṽ in R2+ (or equivalently by the
proof of Proposition 3.2 of this paper; note that f (0) = 0), we have that |∇v| and |D 2 v| are
bounded in R2++ ∩ {0  y  1}. We conclude that |∇v| and |D 2 v| are bounded in R2++ and
 X. Cabré, J. Tan / Advances in Mathematics 224 (2010) 2052–2093 2087

C
∇v(x, t)  , for all t > 0, x > 0.
t +1

Using interior estimates for harmonic functions as before, but now with the partial derivatives of
v instead of v, it follows that

C
D 2 v(x, t)  , for all t > 0, x > 0.
t2 +1

Moreover, we have
 
∂ |∂x v(x, t)|2 − |∂y v(x, t)|2 C
 3 .
∂x 2 t +1

By these facts, we see that the function

+∞
|∂x v(x, t)|2 − |∂y v(x, t)|2
Φ(x) := dt
2
0

is well defined and dΦ

dx is also.
v
Using that limt→∞ |∇v(x, t)| = 0, we obtain, for F (v) = 0 f (s) ds,

d   
Φ(x) + F v(x, 0)
dx
+∞
 
= [∂xx v∂x v − ∂y v∂xy v](x, t) dt + f (v)∂x v (x, 0)
0
 
= ∂y v∂x v + f (v)∂x v (x, 0) = 0,

thanks to the harmonicity of v and the Neumann boundary condition. This leads to the
Hamiltonian-type identity
 
Φ(·) + F v(·, 0) is identically constant in (0, +∞).

Furthermore, using that limx→+∞ v(x, 0) = 0, and that limx→+∞ v(x, y) = 0 uniformly in
compact sets in y (we can prove this by the same previous argument leading to limx→+∞ v(x, 0)
= 0), together with the above bounds for |∇v(x, y)| for y large, we deduce

lim Φ(x) = 0.
x→+∞

From all these we obtain

 
Φ(x) + F v(x, 0) ≡ 0, for x > 0.

Since v = 0 and thus ∂y v = 0 along the y-axis, we see by the definition of Φ(0) that
2088 X. Cabré, J. Tan / Advances in Mathematics 224 (2010) 2052–2093

+∞
  1
0 = Φ(0) + F v(0, 0) = Φ(0) = |∂x v|2 (0, t) dt.
2
0

This implies that ∂x v = 0 on {x = 0, y > 0}, which contradicts Hopf’s lemma. Thus, the contra-
diction means that there is no positive bounded solution of the problem. 2

Before proving Theorems 1.5 and 1.3, let us make some comments.

Remark 6.5. Theorem 1.5 is still open without the boundedness assumption on v.
In this respect, let us give some examples of problems in the quarter plane R2++ . The function
v(x, y) = x is an unbounded solution of the problem
⎧
⎪
⎪ −v = 0, v  0 in R2++ ,
⎨
v=0 on {x = 0, y > 0},
⎪
⎪ ∂v
⎩ =0 on {x > 0, y = 0}.
∂ν
This tells us that the result of Proposition 6.2 (which did not require boundedness of the solution
in the half-plane) does not hold in the quarter plane.
On the other hand, it is clear that v(x, y) = π2 arctan y+1
x
satisfies v = 0 and −∂y v|y=0 =
πx
2(1+x 2 )
 0 for x > 0. Hence, there exists a bounded harmonic function in the quarter plane R2++
such that
⎧
⎪
⎪ −v = 0, v  0 in R2++ ,
⎨
v=0 on {x = 0, y > 0},
⎪
⎪ ∂v
⎩ 0 on {x > 0, y = 0}.
∂ν

Thus the nonlinear condition ∂v

∂ν = v p on {y = 0} is important in Theorem 1.5.

Proof of Theorem 1.5. It follows from Propositions 6.3 and 6.4. 2

Proof of Theorem 1.3. We know by Theorem 5.2 and Proposition 3.2 that all weak solutions u
of (1.1), with f as in Theorem 1.3, belong to C 2 (Ω) ∩ C0 (Ω). Assume by contradiction that the
theorem is not true and hence that there is a sequence um of solutions of (1.1) with

Km = um L∞ (Ω) → ∞.

Since vm = h-ext(um ) is a positive harmonic function in C vanishing on ∂L C, we have that vm

has also Km as maximum in C and that it is attained at a point (xm , 0) ∈ Ω × {0}. Let

p−1
Ωm = Km (Ω − xm )

and define

−1
 1−p 1−p 
ṽm (x, y) = Km v xm + Km x, Km y , x ∈ Ωm , y > 0.
 X. Cabré, J. Tan / Advances in Mathematics 224 (2010) 2052–2093 2089

We have that ṽm L∞ (Ωm ×(0,∞))  1 and

⎧
⎪
⎪ ṽm = 0 in Cm := Ωm × (0, ∞),
⎪
⎪
⎨ ṽm = 0 on ∂Ωm × (0, ∞),
∂ ṽm (6.10)
⎪
⎪
p
= ṽm on Ωm × {0},
⎪
⎪ ∂ν
⎩
ṽm > 0 in Cm .

Notice that

ṽm (0, 0) = 1.

Let

dm = dist(xm , ∂Ω).

Two cases may occur as m → ∞; either case (a):

p−1
Km dm → ∞

for a subsequence still denoted as before, or case (b):

p−1
Km dm is bounded.
p−1 p−1
If case (a) occurs, we have that BK p−1 d (0) = Km Bdm (0) ⊂ Ωm and that Km dm → ∞.
m m
By local compactness (Arzelà–Ascoli) of bounded solutions to (6.10) (recall ṽm L∞ (Ωm )  1),
through a subsequence, we obtain a solution ṽ of problem (6.2) in all of Rn+1 + = R × (0, ∞)—
n

note that ṽm (0, 0) = 1 leads to ṽ(0, 0) = 1 and hence ṽ ≡ 0 and ṽ > 0. This is a contradiction to
Theorem 6.1.
p−1
Assume now that case (b), Km dm is bounded, occurs. Note first that since the right-hand
p p
side of problem (6.1) for vm satisfies |vm |p = vm  Km , we deduce from the proofs of Proposi-
p
tion 3.1 (iii) and (iv) that ∇um L∞ (Ω)  CKm for a constant C independent of m. Now, since
um |∂Ω ≡ 0 (where um = vm (·, 0)), we get
p
Km = vm (xm , 0)  ∇um L∞ (Ω) dist(xm , ∂Ω)  CKm dm .

We deduce that
p−1
0 < c  Km dm

for some positive constant c. Thus, in this case (b), we may assume that, up to a subsequence,

p−1
Km dm → a ∈ (0, ∞) (6.11)

for some constant a > 0.

p−1
We deduce that, up to a certain rotation of Rn for each index m, since we have (6.11), Km →
∞, dm → 0, and that BK p−1 d (0) is tangent to ∂Ωm , the domains Ωm converge to the half-space
m m
2090 X. Cabré, J. Tan / Advances in Mathematics 224 (2010) 2052–2093

Rn+ = {xn > −a}. Thus, through a subsequence of ṽm , we obtain a solution ṽ of problem (6.7) in
Rn+1
++ = {xn > −a, y > 0} with ṽ bounded by 1 and ṽ > 0 (since ṽm (0, 0) = 1 for all m). This is
a contradiction with Theorem 1.5. 2

Remark 6.6. From Theorem 1.3 we have a priori bounds for solutions of problem (1.1) with
f (u) = up , 1 < p < n+1
n−1 . As a consequence, by using blow-up techniques and topological de-
gree theory, one can obtain existence of positive solutions for related problems—for instance,
for nonlinearities f (x, u) of power type, as well as other boundary conditions. See Gidas and
Spruck [14] for some of these applications when the operator is the classical Laplacian.

7. Symmetry of solutions

The goal of this section is to prove a symmetry result of Gidas–Ni–Nirenberg type for positive
solutions of nonlinear problems involving the operator A1/2 , as stated in Theorem 1.6, by using
the moving planes method. For this, we work with the equivalent local problem (1.8) and derive
the following.

Theorem 7.1. Assume that Ω is a bounded smooth domain of Rn which is convex in the x1
direction and symmetric with respect to the hyperplane {x1 = 0}. Let f be Lipschitz continuous
and let v ∈ C 2 (C) be a solution of (1.8), where C = Ω × (0, +∞). Then, v is symmetric with
respect to x1 , i.e., v(−x1 , x , y) = v(x1 , x , y) for all (−x1 , x , y) ∈ C. In addition, ∂x
∂v
1
< 0 for
x1 > 0.

Proof of Theorems 1.6 and 7.1. It suffices to prove Theorem 7.1. From it, Theorem 1.6 follows
immediately.
Let x = (x1 , x ) ∈ Ω and λ > 0. Consider the sets
   
Σλ = x1 , x ∈ Ω x1 > λ and Tλ = x1 , x ∈ Ω x1 = λ .

For x ∈ Σλ , define xλ = (2λ − x1 , x ). By the hypotheses on the domain Ω we see that

{xλ | x ∈ Σλ } ⊂ Ω.

Recall that v ∈ C 2 (C) is a solution of

⎧
⎪ v = 0 in C = Ω × (0, ∞),
⎪
⎪
⎨v = 0 on ∂L C = ∂Ω × [0, ∞),
⎪ ∂v
⎪
⎪ = f (v) on Ω × {0},
⎩ ∂ν
v>0 in C.

For (x, y) ∈ Σλ × [0, ∞), let us define

 
vλ (x, y) = v(xλ , y) = v 2λ − x1 , x , y

and
 X. Cabré, J. Tan / Advances in Mathematics 224 (2010) 2052–2093 2091

wλ (x, y) = (vλ − v)(x, y).

Note that vλ satisfies

⎧ v = 0 in Σλ × (0, ∞),
⎪
⎨
λ
vλ  0 on (∂Ω ∩ Σ λ ) × (0, ∞),
⎪
⎩ ∂vλ = f (v )
λ on Σλ × {0}.
∂ν

Thus, since ∂Σλ = (∂Ω ∩ Σ λ ) ∪ Tλ and wλ ≡ 0 on Tλ , we have that wλ satisfies

⎧ w = 0 in Σλ × (0, ∞),
⎪
⎨
λ
wλ  0 on (∂Σλ ) × (0, ∞), (7.1)
⎪
⎩ ∂w λ
+ cλ (x)wλ = 0 on Σλ × {0},
∂ν
where
f (vλ ) − f (v)
cλ (x, 0) = − .
vλ − v

Note that cλ (x, 0) ∈ L∞ (Σλ ).

Let λ∗ = sup{λ | Σλ = ∅} and let ε > 0 be a small number. If λ ∈ (λ∗ − ε, λ∗ ), then Σλ has
small measure and we have, by part (ii) of Proposition 4.4 (applied with Ω replaced with Σλ ),
that

wλ  0 in Σλ × (0, ∞).

Note here that Σλ is not a smooth domain but that part (ii) of Proposition 4.4 does not require
smoothness of the domain. By the strong maximum principle, Lemma 4.2, for problem (7.1) we
see that wλ is identically equal to zero or strictly positive in Σλ × (0, ∞). Since λ > 0, we have
wλ > 0 in (∂Ω ∩ ∂Σλ ) × (0, ∞), and hence we conclude that wλ > 0 in Σλ × (0, ∞).
Let λ0 = inf{λ > 0 | wλ  0 in Σλ × (0, ∞)}. We are going to prove that λ0 = 0. Suppose that
λ0 > 0 by contradiction. First, by continuity, we have wλ0  0 in Σλ0 × (0, ∞). Then, as before,
we deduce wλ0 > 0 in Σλ0 × (0, ∞). Next, let δ > 0 be a constant and K ⊂ Σλ0 be a compact
set such that |Σλ0 \ K|  δ/2. We have wλ0 (·, 0)  η > 0 in K for some constant η, since K is
compact. Thus, we obtain that wλ0 −ε (·, 0) > 0 in K and that |Σλ0 −ε \ K|  δ for ε small enough.
Now we apply again part (ii) of Proposition 4.4 in Σλ0 −ε × (0, ∞) to the function wλ0 −ε .
We know that wλ0 −ε (·, 0)  0 in K, and hence {wλ0 −ε < 0} ⊂ Σλ0 −ε \ K, which has measure at
most δ. We take δ to be the constant of part (ii) of Proposition 4.4. We deduce that

wλ0 −ε  0 in Σλ0 −ε × (0, ∞).

This is a contradiction to the definition of λ0 . Thus, λ0 = 0.

We have proved, letting λ ↓ λ0 = 0 that
     
v −x1 , x , y  v x1 , x , y in Ω ∩ {x1 > 0} × (0, ∞)

and, since wλ = 0 on Tλ ,
2092 X. Cabré, J. Tan / Advances in Mathematics 224 (2010) 2052–2093

1 ∂wλ
∂x1 v = − <0 for x1 > 0,
2 ∂x1

by Hopf’s lemma. Finally replacing x1 by −x1 , we deduce the desired symmetry v(−x1 , x , y) =
v(x1 , x , y). 2

Acknowledgments

Both authors were supported by the Spain Government grants MTM2005-07660-C02-01 and
MTM2008-06349-C03-01. The first author was supported by the Catalan Government grant
SGR2009-345. The second author was supported by CONICYT Becas de Postgrado of Chile
and the Programa de Recerca del Centre de Recerca Matemàtica, Barcelona, Spain.

References

[1] A. Ambrosetti, P. Rabinowitz, Dual variational methods in critical points theory and applications, J. Funct. Anal. 14
(1973) 349–381.
[2] D. Applebaum, Lévy processes—from probability to finance and quantum groups, Notices Amer. Math. Soc. 51
(2004) 1336–1347.
[3] H. Berestycki, L. Nirenberg, On the method of moving planes and the sliding method, Bol. Soc. Brasil. Mat. 22
(1991) 1–37.
[4] H. Brezis, L. Nirenberg, Positive solutions of nonlinear elliptic equations involving critical Sobolev exponents,
Comm. Pure Appl. Math. 36 (1983) 437–477.
[5] X. Cabré, J. Solà-Morales, Layer solutions in a halfspace for boundary reactions, Comm. Pure Appl. Math. 58
(2005) 1678–1732.
[6] L. 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 (2008) 425–461.
[7] L. Caffarelli, L. Silvestre, An extension problem related to the fractional Laplacian, Comm. Partial Differential
Equations 32 (2007) 1245–1260.
[8] L. Caffarelli, A. Vasseur, Drift diffusion equations with fractional diffusion and the quasi-geostrophic equation,
Ann. of Math., in press.
[9] M. Chipot, M. Chlebík, M. Fila, I. Shafrir, Existence of positive solutions of a semilinear elliptic equation in Rn+
with a nonlinear boundary condition, J. Math. Anal. Appl. 223 (1998) 429–471.
[10] J. Davila, Singular solutions of semi-linear elliptic problems, in: Handbook of Differential Equations, vol. 2, Sta-
tionary Partial Differential Equations, Elsevier Science, 2009, Chapter 2.
[11] J. Davila, L. Dupaigne, M. Montenegro, The extremal solution of a boundary reaction problem, Commun. Pure
Appl. Anal. 7 (2008) 795–817.
[12] J. Escobar, Sharp constant in a Sobolev trace inequality, Indiana Univ. Math. J. 37 (1988) 687–698.
[13] B. Gidas, W.-M. Ni, L. Nirenberg, Symmetry and related properties via the maximum principle, Comm. Math.
Phys. 68 (1979) 209–243.
[14] B. Gidas, J. Spruck, A priori bounds for positive solutions of nonlinear elliptic equations, Comm. Partial Differential
Equations 6 (1981) 883–901.
[15] D. Gilbarg, N.S. Trudinger, Elliptic Partial Differential Equations of Second Order, Classics in Mathematics,
Springer-Verlag, Berlin, 2001.
[16] N.S. Landkof, Foundations of Modern Potential Theory, Springer-Verlag, 1972.
[17] Y.Y. Li, L. Zhang, Liouville-type theorems and Harnack-type inequalities for semilinear elliptic equations, J. Anal.
Math. 90 (2003) 27–87.
[18] Y.Y. Li, M. Zhu, Uniqueness theorems through the method of moving spheres, Duke Math. J. 80 (1995) 383–417.
[19] P.L. Lions, The concentration-compactness principle in the calculus of variations, The limit case II, Rev. Mat.
Iberoamericana 1 (1985) 45–121.
[20] J.L. Lions, E. Magenes, Non-homogeneous Boundary Value Problems and Applications, vol. I, Die Grundlehren
der Math. Wissenschaften, vol. 181, Springer-Verlag, 1972.
[21] A. Nekvinda, Characterization of traces of the weighted Sobolev space W 1,p (Ω, d M ) on M, Czechoslovak Math.
J. 43 (118) (1993) 695–711.
 X. Cabré, J. Tan / Advances in Mathematics 224 (2010) 2052–2093 2093

[22] B. Ou, Positive harmonic functions on the upper half space satisfying a nonlinear boundary condition, Differential
Integral Equations 9 (1996) 1157–1164.
[23] L. Silvestre, Regularity of the obstacle problem for a fractional power of the Laplace operator, Comm. Pure Appl.
Math. 60 (2006) 67–112.
[24] M. Struwe, Variational Methods, Ergebnisse der Mathematik und ihrer Grenzgebiete, vol. 34, Springer-Verlag, 1996.
[25] S. Sugitani, On nonexistence of global solutions for some nonlinear integral equations, Osaka J. Math. 12 (1975)
45–51.
[26] J. Tan, The Brezis–Nirenberg type problem involving the square root of the Laplacian, preprint.