Embedded disc-type surfaces with large constant mean curvature

and free boundaries

1
FALL Mouhamed Moustapha

Sissa, Via Beirut 2-4, Trieste, 34014 Italy

Abstract
3
Given Ω ⊂ R an open bounded set with smooth boundary ∂Ω and H ∈ R, we prove the existence
of embedded H-surfaces supported by ∂Ω, that is regular surfaces in R3 with constant mean curvature
H at every point, contained in Ω and with boundary intersecting ∂Ω orthogonally. More precisely,
we prove that if Q ∈ ∂Ω is a stable stationary point for the mean curvature of ∂Ω, then there exists
a family of embedded 1ε -surfaces near Q, ε > 0 small, which, once dilated by a factor 1ε and suitably
translated, converges to a hemisphere of radius 1 as ε → 0.

1 Introduction
Let Ω be a bounded open set of R3 with smooth boundary ∂Ω (of class C m , m ≥ 3). In this paper we study
the existence of concentrating disc-type constant mean curvature surfaces inside Ω and supported by ∂Ω.
By an H-surface parametrized by u and supported by ∂Ω, we mean a map u ∈ C 2 (B; R3 ) ∩ C 1 (B̄; R3 ) of
the unit disc
B := {(x, y) ∈ R2 : x2 + y 2 < 1}
into R3 satisfying the following conditions:
(
∆u = 2Hux ∧ uy in B,
(1)
|ux |2 − |uy |2 = 0 = ux · uy in B,

(
u(∂B) ⊂ ∂Ω,
(2) ∂u
∂n (σ) ⊥ Tu(σ) ∂Ω ∀σ ∈ ∂B.

Here H ∈ R is the mean curvature of the image of u at every point, ux = ∂u ∂u

∂x , uy = ∂y , ” ∧ ” denotes the
3
exterior product in R , ” · ” denotes the scalar product, n is the outer unit normal on ∂B, ” ⊥ ” means
orthogonal, and Tq ∂Ω denotes the tangent space to ∂Ω at q ∈ ∂Ω.
Solutions to (1)-(2) arise by considering the following isoperimetric (partition) problem:

Given a domain Ω, determine a rectifiable surface of minimal or stationary area relative to Ω, with
boundary contained in ∂Ω and dividing Ω into two parts Ω1 , Ω2 such that

meas Ω1 = σ meas Ω, meas Ω2 = (1 − σ) meas Ω,

where σ denotes a preassigned constant with 0 < σ < 1.

It is clear that whenever such a minimal separating surface exists will meet ∂Ω orthogonally and will
1 E-mail addresses: fall@sissa.it

1
have constant mean curvature, as confirmed in [12]. A solution of the above problem can be obtained
using geometric measure theory, see [17] and [27]. Nevertheless existence of solutions to this problem
with prescribed topology is widely open. The interior and boundary regularities are treated in [10]-[11]
and in [12], respectively. However the afore stated partition problem, prescribing a disc type topology,
can be reduced to a free boundary problem for the Dirichlet integral with the constraint that the volume
enclosed by admissible surfaces is σ meas Ω. In this way the surface verifies equations (1)-(2) with H
being a Lagrange multiplier (see [12]).
Equation (1)-(2) and its parabolic counterpart have been the subject of several works, see for instance
the paper [24] by M.Struwe. The latter generalizes the existence result in [26] and in some sense extends
Hildebrandt’s work [15] for the Plateau problem for H-surfaces, namely (1) with the following boundary
condition instead of (2)

(3) u|∂B : ∂B → Γ is a parametrization of a given Jordan curve Γ ⊂ R3 .

For H = 0 (1),(3) constitute the classical Plateau problem for minimal surfaces solved by J.Douglas [8]
and T.Radò [19]. Generalizations for H 6= 0 were obtained in [15], where the existence of a stable solution
was proved. For ”small” H, Brezis-Coron [4], K.Steffen [21] and M.Struwe [23], found the existence of
unstable solutions as well. These results were extended in [22] where the following result was established:
for H 6= 0, there is always an unstable solution of (1),(3) provided there is a stable solution.
By analogy, since in the free boundary problem stable solutions (trivial solutions) always exist for any H,
one could expect unstable solutions to exist for any H 6= 0. Using an argument by Sacks-Uhlenbeck [20],
existence result for the problem (1)-(2) with H = 0 were obtained by M.Struwe [26] under the assumption
that the supporting surface ∂Ω is C 4 -diffeomorphic to S 2 but his solutions in general are not embedded
in Ω. This result was improved by Grüter-Jost [13], when Ω is convex, proving the existence of embedded
minimal discs. For the case H ∈ R, an existence result was obtained by M.Struwe [24]. Precisely he
proved that there exists a dense subset L of the interval [− L1 , L1 ] containing 0 such that for any H ∈ L
there exists a non constant H-surface contained in a ball BL (0) provided ∂Ω is contained in BL (0) and
is C 4 -diffeomorphic to S 2 . In view of the existence result by M.Struwe, one may ask whether there are
embedded surfaces with large constant mean curvature supported by ∂Ω. This question is related to the
problem of partitioning Ω into two parts of measures ε meas Ω and (1 − ε) meas Ω with ε small. We
give an answer to these questions in this paper by our main result, which is the following.

Theorem 1.1 Suppose Ω ⊂ R3 , is a domain with smooth boundary ∂Ω of class C m , m ≥ 3. Suppose

Q0 ∈ ∂Ω is a local strict maximum or minimum, or a non-degenerate critical point of the mean curvature
of ∂Ω. Then there exists a family of embedded constant mean curvature surfaces supported by ∂Ω and
concentrating at Q0 . Precisely, there exists uε ∈ C m (B, ∂Ω) ∩ C 1,β (B̄, ∂Ω) for every β ∈ (0, 1), ε  1
such that uε (B) is a 1ε -surface supported by ∂Ω and ||uε − Q0 ||C 1,β (B̄,R3 ) → 0 as ε → 0. Moreover 1ε uε ,
suitably translated, converges to a hemisphere of radius 1.

Our next result concerns multiplicity of solutions depending on the topology of ∂Ω, with no assumptions
on the mean curvature of the boundary of Ω.
Theorem 1.2 Under the assumption of Theorem 1.1, there exists at least cat(∂Ω) geometrically distinct
solutions, where cat(∂Ω) is the Lusternik-Schnierelman category of ∂Ω.

Remark 1.3 1. It is worth noticing that, comparing our result with M.Struwe [24], no assumptions
on Ω are made. Furthermore our result is ’”complementary” to Struwe’s one in the sense that his
admissible mean curvatures are bounded while ours can be arbitrarily large.

2. We believe that it should be possible to extend the result to higher-dimensional H − systems (in this
direction see [14]) with simple modifications. Moreover by an approach of [28], we also believe that
one may generalize our result to curved spaces, namely when Ω is a domain in a Riemann manifold.
For more precise comments see Remark 4.3.

2
 Since we look for solutions with a given asymptotic profile, it is convenient to scale the problem by a
factor 1ε : letting Sε = 1ε ∂Ω = 1ε S, we consider the equivalent problem

∆u = 2ux ∧ uy in B,

|u |2 − |u |2 = 0 = u · u

in B,
x y x y
(4)
u(∂B) ⊂ Sε ,


 ∂u
∂n (σ) ⊥ Tu(σ) Sε ∀σ ∈ ∂B.

At first glance, as ε → 0, in the limit we get a plane as a supporting surface, so one is led to consider the
limit problem



∆u = 2ux ∧ uy in B,
|u |2 − |u |2 = 0 = u · u ,

x y x y
(5) 2


u(∂B) ⊂ R × {0},
 ∂u 2
∂n (σ) ⊥ R × {0} ∀σ ∈ ∂B.

The latter problem admits a fundamental solution Π, the inverse of the stereographic projection (from the
south pole) restricted on B (see (13)), and a family of solutions of the form Π ◦ g + p, where p ∈ R2 × {0}
and g is any conformal diffeomorphism of the unit disc. It turns out that this set of solutions defines a
manifold Z̃ of critical points of the Euler functional J associated to (5). It is clear that Z̃ = G × R2 ' R5 ,
where G is the group of Möbius transformations of dimension 3 (see (14)).
Thanks to some results already known in the literature (see [7], [18], [16]), we are able to prove that Z̃ is
a non-degenerate manifold; that is the tangent space Tz Z̃ of Z̃ at any z ∈ Z̃ equals the kernel of d2 J(z).
Hence by the Fredholm theorem we can solve (4) if we are suitably perpendicular to TΠ Z̃ in a suitable
sense, see Lemma 2.4. This is the key step for a finite dimensional reduction of our problem (see [2], [[3]
Section 2.4], [5], [7], [9], [18], [16], [6], [28] for related methods).
As in [24], we take advantage of the variational structure of (4). While in [24] it was necessary to impose
a topological condition on Ω (in order to define an extension operator on a subclass of Sobolev functions,
see Section 3) we can localize the variational formulation using the smallness of ε, see Lemma 3.1.
Because of the free boundary condition in (4), a natural set to study the problem are maps of B into R3
of class H 1,2 such that ∂B is sent into Sε (which we call admissible functions). The subset of admissible
functions with H 2,2 regularity is a Hilbert manifold, dense in the above set as pointed out by M.Struwe
in [24]. Looking for solutions close to Π, reasoning as for the flat case, we impose suitable constraints on
the tangent plane of the Hilbert manifold, in order to guarantee a (partial) invertibility of the linearized
equation as marked before. Once we have this, we fully solve the equation with a finite-dimensional
reduction.
To begin the procedure, we construct approximate solutions, which are nothing but suitable perturbations
of hemispheres which intersect ∂Ω almost orthogonally. The reduction is done transforming the problem
into finding critical points of a functional Fε defined on Sε , see Proportion 3.10. For ε small, Fε admits
the asymptotic expansions in (58), where we see the role played by the mean curvature of ∂Ω.
A similar technique was used by R.Ye [28] to find constant mean curvature surfaces in manifolds,
and the approximate solutions were perturbations of geodesic spheres. These surfaces concentrate near
non-degenerate critical points of the scalar curvature, see the Remark 4.3 for related comments.
One of the main features of performing the Lyapunov-Schmidt reduction for our problem is the action
of the Möbius group, which generates some extra dimensions in the kernel of the linearized equation.
To deal with this problem, we use the invariance of the functional under this action, and show that the
gradient of the functional has basically no component in the subspace TId G of TΠ Z̃. Another issue is
the regularity of admissible functions: while the variational approach settles naturally in H 1,2 (where we
have coercivity, Fredholm properties, etc...), it is from other points of view convenient to work in H 2,2
since we have stronger embeddings and the functionals involved are more regular. To handle this, we
crucially use the smallness of ε, the smoothness of ∂Ω and elliptic regularity estimates, see Lemma 3.8.

3
Acknowledgments. I wish to thank Professor Andrea Malchiodi for suggesting the problem and for
useful discussions. The author has been supported by M.U.R.S.T within the PRIN 2006 Variational
Methods and Nonlinear Differential Equations.

2 Notations and Preliminaries

In this section we introduce some preliminaries that will be used in the sequel.
• We denote by ei , i = 1, 2, 3 the canonical basis of R3 . We let

S 2 = (x1 , x2 , x3 ) ∈ R3 : x21 + x22 + x23 = 1



2
denote the unit sphere and S+ the upper hemisphere is
2
= (x1 , x2 , x3 ) ∈ S 2 : x3 > 0 .

(6) S+

• For 1 ≤ p ≤ ∞ and m ∈ N, let Lp (B, Rn ), H m,p (B, Rn ), denote the usual Lebesgue- Sobolev spaces
with norms || · ||p , || · ||m,p . In particular we will write || · ||2 = || · ||.
n o
• If r > 0 and X̃0 ∈ R3 , set Br (X̃0 ) = X̃ ∈ R3 : |X̃ − X̃0 | < r . We denote Br the ball of radius
r centered at the origin.
• For every u = (u1 , u2 , u3 ), v = (v 1 , v 2 , v 3 ) ∈ H m,p (B, R3 ), we define
3
X
u·v = ui v i , ∇u · ∇v = ux · vx + uy · vy .
i=1

Also we will write |u|2 = u · u and |∇u|2 = ux · ux + uy · uy .

• Let E ⊂ RN be a domain with C m -boundary ∂E = Γ, m < s < m + 1 (s ∈ R) and 1 ≤ p < ∞: we
recall the definitions of fractional Sobolev spaces:
( )
s,p n m,p n |u(z1 ) − u(z2 )| m,p n
H (E, R ) = u ∈ H (E, R ) : N ∈ H (E × E, R ) ,
|z1 − z2 |s+ p

endowed with the natural norm. Covering ∂E = Γ by coordinate charts, one can define the Sobolev
spaces H s,p (Γ, Rn ) (see [1]; paragraph 7.51). Now if 1 < p < ∞, u ∈ H m,p (E, Rn ) then the trace
1
of u, uΓ belongs to H m− p ,p (Γ, Rn ). As a consequence of the trace theorem there exists a constant
C1 > 0 depending only on E such that

||uΓ || m− 1 ,p
p
≤ C1 ||u||m,p ,
H

1
and conversely if v ∈ H m− p ,p (Γ, Rn ), there exists u ∈ H m,p (E, Rn ) such that uΓ = v on Γ and

||u||m,p ≤ C2 ||v|| m− 1 ,p
p
H

for some C2 > 0 depending only on E (see [1]; paragraph 7.56). For brevity in the sequel we simply
write u(σ) instead of u∂B (σ) for a.e. σ ∈ ∂B if u ∈ H m,p (B, Rn ).

4
Through this section we will identify R2 by R2 × {0} as a subspace of R3 . As said in the previous section,
we shall consider the unperturbed problem:


 ∆u = 2ux ∧ uy in B,
|u |2 − |u |2 = 0 = u · u ,

x y x y
(7)
u(∂B) ⊂ R2 ,


 ∂u 2
∂n (σ) ⊥ R ∀σ ∈ ∂B.

We define the Hilbert subspace H of H 1,2 (B; R3 ) as

H = {u ∈ H 1,2 (B; R3 ) : u(σ) ∈ R2 for a.e. σ ∈ ∂B}.

For every u ∈ H ∩ H 2,2 (B, R3 ), we define the functional:

Z
1
(8) J(u) = |∇u|2 + 2V (u),
2 B

where the volume term V is defined for every u ∈ H 1,2 (B, R3 ) by

Z
1
(9) V (u) = u·(ux ∧ uy ).
3 B

It is clear that (7) is the Euler-Lagrange equation of the functional J in the sense that

u ∈ H ∩ H 2,2 (B, R3 ) solves of equation (7) iff hdJ(u), vi = 0 ∀v ∈ H.

In fact the functional J is smoothly defined on H ∩ H 2,2 (B, R3 ). We have, integrating by parts
Z Z
1 ∂u
(10) hdV (u), vi = (ux ∧ uy )·v − ( ∧ u)·vds, ∀u ∈ H 2,2 (B; R3 ), ∀v ∈ H 1,2 (B; R3 ),
B 3 ∂B ∂t

where t(x, y) = (−y, x) is the tangent vector at (x, y) ∈ ∂B to ∂B and ∂u ∂t is the tangential derivative of
u. When u ∈ H ∩ H 2,2 (B, R3 ) and v ∈ H, one has ( ∂u
∂t ∧ u)·v = 0 a.e. on ∂B since ∂u 2
∂t (σ), u(σ), v(σ) ∈ R ,
2,2 3
for a.e. σ ∈ ∂B, so it turns out that for every u ∈ H ∩ H (B, R )
Z Z
hdJ(u), vi = ∇u·∇v + 2 (ux ∧ uy )·v ∀v ∈ H
ZB B
Z
∂u
= [−∆u + 2ux ∧ uy ]·v + · v.
B ∂B ∂n

Since H01 (B, R3 ) ⊂ H, it follows that a critical point u ∈ H ∩ H 2,2 (B, R3 ) of dJ satisfies the first equation
∂u
of (7) and then ∂n (σ) ⊥ R2 for a.e. σ ∈ ∂B so that

∂u ∂u
(11) (σ) ⊥ (σ) for a.e. σ ∈ ∂B.
∂n ∂t
Now, setting
 2  2 !
 ∂u 
  −  ∂u  − 2i ∂u · ∂u
 
Φ(x, y) =  ∂n   ∂t  for every n = (x, y), t = (−y, x) ∈ B
∂n ∂t

we see that Φ is holomorphic and by (11) is real on ∂B. Therefore by the Cauchy-Riemann equations
Φ is constant in B̄ but since Φ(0, 0) = 0, u is conformal. Boundary regularity and strong orthogonality

5
follow from standard elliptic theory.
Now for every u, w ∈ H ∩ H 2,2 (B, R3 ) and v ∈ H by similar argument, we have
Z Z Z
2
3hd V (u)w, vi = w·(vx ∧ uy + ux ∧ vy ) + v·(wx ∧ uy + ux ∧ wy ) + u·(vx ∧ wy + wx ∧ vy )
B B B
Z Z
∂u ∂w
= 3 (ux ∧ wy + wx ∧ uy )·v − ( ∧w+ ∧ u)·vds
∂t ∂t
ZB ∂B

= 3 (ux ∧ wy + wx ∧ uy )·v
B

and thus by density for every u ∈ H ∩ H 2,2 (B, R3 ) there hold

Z Z
(12) hd2 J(u)w, vi = ∇w·∇v + 2 v·(wx ∧ uy + ux ∧ wy ) ∀w, v ∈ H;
B B
Z Z
hd2 J(u)v, vi = |∇v|2 + 4 u·(vx ∧ vy ) ∀v ∈ H.
B B

Note that equation (7) is invariant under the action of the group of Möbius transformation of the unit
disc and by translation in the direction of vectors in the plane. Following [4], up to a reflection with
respect to the plane, J has a manifold of critical points generated by the inverse of the stereographic
projection from the south pole restricted on B. Namely if we set
 2x 2y 1 − x2 − y 2 
(13) Π(x, y) = , , ∀(x, y) ∈ B
1 + x2 + y 2 1 + x2 + y 2 1 + x2 + y 2
and
X −a
(14) G = {g(X) = eiθ , θ ∈ [−π, π), a ∈ B},
1 − āX
where in complex notations, X = (x, y) = x + iy, then the manifold of critical points is

Z̃ = {Π ◦ g + p̃, g ∈ G, p̃ ∈ R2 }.

We collect some useful properties of the map Π in the following lemma. The proof is just simple compu-
tations.
2
Lemma 2.1 Let X = (x, y), |X|2 = x2 + y 2 and µ(X) = 1+|X|2 so that Π has the form

Π(X) = (X, 1)µ − e3 .

2
Then Π(B) = S+ and it satisfies

1. ∆Π = 2Πx ∧ Πy = −2µ2 Π;
∂Π ∂Π
2. Π(σ) = (σ, 0), ∂n (σ) = −e3 , ∂t (σ) = (t, 0) = (−y, x, 0) ∀σ = (x, y) ∈ ∂B;
1 2
3. 2 |∇Π| = |Πx |2 = |Πy |2 = |Πx ∧ Πy | = µ2 ;
4. Πx ∧ Π = Πy , Πy ∧ Π = −Πx ;

f ∈ H 1,2 (B, R3 );
R R
5. B Π[fx ∧ Πy + Πx ∧ fy ] = − B ∇Π · ∇f ∀
6. B |∇Π|2 = 4π.
R

We prove that the manifold Z̃ is non-degenerate, namely that Tz Z̃ = Ker d2 J(z) for all z ∈ Z̃ where
Tz Z̃ denotes the tangent space of Z̃ at z. We first characterize explicitly TΠ Z̃.

6
Lemma 2.2 In the above notations we have
n o
TΠ Z̃ = span Π ∧ e3 ; (e1 · Π)Π; (e2 · Π)Π; e1 ; e2 .

Proof. By easy computations one finds

∂Π ◦ gθ,(0,0)
= e3 ∧ Π,
∂θ |θ=0
∂Π ◦ g0,(a1 ,0)
= (e1 · Π)Π − e1 ,
∂a1 |a1 =0
∂Π ◦ g0,(0,a2 )
= (e2 · Π)Π − e2 ,
∂a2 |a2 =0
∂(Π + p̃)
= ei , i = 1, 2.
∂ p˜i |pi =0

The lemma then follows immediately.

We fix the following notations

(15) E1 = e3 ∧ Π, E2 = (e1 · Π)Π, E3 = (e2 · Π)Π and GΠ = span {E1 ; E2 ; E3 } .

The above result can be restated in the following way

       
 c1 0 a1 
(16) TΠ Z̃ =  c2  +  0  ∧ Π +  a2  · Π Π, ci , ai , b3 ∈ R, i = 1, 2
0 b3 0
 

We are now ready to prove the non degeneracy condition which plays here a key role, we shall state the
following
Lemma 2.3 The following equality holds

TΠ Z̃ = Ker d2 J(Π).

Proof. Let us first emphasize that, in view of (12), by partial integration w ∈ kerd2 J(Π) if and only
if it satisfies the following equation

∆w = 2(wx ∧ Πy + Πx ∧ wy ) in B,

(17) w(∂B) ⊂ R2 ,
 ∂w
 2
∂n (σ) ⊥ R ∀σ ∈ ∂B.
2
Equivalently after inverse of the stereographic projection on the sphere S+ , equation (17) is

2 2
(18) ∆ g0 w = (wφ ∧ Πθ + Πφ ∧ wθ ) in S+ ,
sin φ

where 0 ≤ θ ≤ 2π, 0 ≤ φ ≤ π2 are the spherical coordinates on the half sphere S+

2
and ∆g0 is the Laplacian
2
with respect to the standard metric on S+ .
We shall extend Π and w to the whole sphere S 2 . We may write Π(φ, θ) = (sin φ cos θ, sin φ sin θ, cos φ) =
(x1 , x2 , x3 ), φ ∈ [0, π2 ) and define:

π
Π̃(x1 , x2 , x3 ) = (x1 , x2 , x3 ) if 0 ≤ φ ≤ ,
2
π
Π̃(x1 , x2 , x3 ) = (x1 , x2 , −x3 ((π − φ), θ)) if ≤ φ ≤ π.
2

7
Π̃ is nothing but the inverse of the stereographic projection. Similarly we also extend w(x1 , x2 , x3 ) =
(w1 (x1 , x2 , x3 ), w2 (x1 , x2 , x3 ), w3 (x1 , x2 , x3 )) on S 2 by
π
w̃ = (w1 (x1 , x2 , x3 ), w2 (x1 , x2 , x3 ), w3 (x1 , x2 , x3 )) if 0 ≤ φ ≤ ,
2
π
w̃ = (w1 (x1 , x2 , −x3 ), w2 (x1 , x2 , −x3 ), −w3 (x1 , x2 , −x3 )) if ≤ φ ≤ π.
2
Clearly w̃ ∈ H 1,2 (S 2 ) and satisfies
2
(19) ∆g0 w̃ = (Π̃φ ∧ w̃θ + w̃φ ∧ Π̃θ ) on S 2 .
sin φ

Now by a result in [7] Lemma 9.2 or [18] Proposition 3.1

w̃ = c + b ∧ Π̃ + (a · Π̃)Π̃, for some a, b, c ∈ R3 .

2
Now since w̃ = w on S+ , returning on the plane, we infer that
      
c1 b1 a1
w =  c2  +  b2  ∧ Π +  a2  · Π Π, on B.
c3 b3 a3

The fact that w ∈ H, implies that c3 = b1 = b2 = 0, as well as the orthogonality condition in (17) implies
that a3 = 0. From this we see that w is of the form as in (16).
As mentioned before, equation (1)-(2) is invariant under the non-compact group of conformal transfor-
mations of the unit disc and therefore it is impossible for the Palais-Smale condition to be satisfied. A
convenient way to factor out the symmetry group could be to impose a three-point-condition on admis-
sible functions, for instance see [25]. In our case the boundary data are allowed to vary freely on ∂Ω, so
we shall normalize the admissible functions by imposing integral constraints, restricting ourselves to the
following Hilbert space
 Z 
1,2 3
Hn = u ∈ H (B; R ) : ∇u·∇Ei = 0, i = 1, 2, 3 .
B

We define I0 = J|H∩Hn , and Z = Z̃ ∩ Hn . It follows that TΠ◦g Z = Ker d2 I0 (Π ◦ g) = R2 for every g ∈ G.

Now let
 Z 
⊥
(20) (TΠ Z) = v ∈ H : vi = 0, i = 1, 2 ; W = (TΠ Z)⊥ ∩ Hn
B

so that we have

(21) H = TΠ Z ⊕ (TΠ Z)⊥ = TΠ Z ⊕ W ⊕ GΠ .

Since every v ∈ (TΠ Z)⊥ satisfies v3 ∈ H01,2 (B, R) thus by Poincaré inequality, the space W endowed with
the norm ||∇v|| is a Hilbert and moreover if we impose orthogonality to Π, d2 I0 (Π) becomes coercive on
W , namely the following result holds true (the proof is similar to the one in [16], Lemma 5.5).
Lemma 2.4 There exists a constant C > 0 such that
Z
2 2
hd I0 (Π)v, vi ≥ C||∇v|| ∀v ∈ W with ∇v·∇Π = 0,
B
hd2 I0 (Π)Π, Πi = −π.

8
3 The abstract method
We start with some preliminaries and notations. Recall that ∂Ω is of class C m , m ≥ 3. Let us set S = ∂Ω
and d(X̃) the (signed) distance function defined by
(
dist(X̃, S) if X̃ ∈ Ω,
d(X̃) :=
−dist(X̃, S) if X̃ ∈ R3 \ Ω.

For some small r0 > 0 depending on S, it is well known that d ∈ C m (Σr0 ), where
n o
Σr0 := X̃ ∈ R3 : |d(X̃)| < 2r0 .

If q ∈ S, then up to a rotation (depending on q), we may assume that Tq (S) coincides with R2 and
e3 with the inner unit normal at q. Also S ∩ Br0 (q) − q is the graph of some function ϕq satisfying
ϕq (0, 0) = 0 and dϕq (0, 0) = 0, with Taylor expansion
1 q
ϕq (X) = hA X, Xi + P q (X) + O(|X|4 ) ∀X = (x, y) with |X| < r0 .
2
Here Aq is the Hessian matrix of ϕq at (0, 0) and P q (X) is a cubic polynomial. Similarly, one also has
1
d(X̃) = e3 ·X̃ + hÃq X̃, X̃i + P̃ q (X̃) + O(|X̃|4 ) ∀X̃ ∈ Br0
2

where again P̃ q (X̃) is a cubic polynomial and

−Aq
 
q 0
(22) Ã = .
0 0

The mean curvature of S at q is given by H(q) = trAq .

Let ϕε,q = 1ε ϕq (εX), so Sε ∩ B rε0 (p) − p is the graph of ϕε,q , with p = 1ε q and dε (X̃) = 1ε d(εX̃). Then
we have
ε q r0
(23) ϕε,q (X) = hA X, Xi + ε2 P q (X) + ε3 O(|X|4 ) ∀X = (x, y) with |X| < ,
2 ε
and moreover dε is of class C m (Σ rε0 ) with

ε
(24) dε (X̃) = e3 ·X̃ + hÃq X̃, X̃i + ε2 P̃ q (X̃) + ε3 O(|X̃|4 ) ∀X̃ ∈ B rε0 .
2
We let
Nε (X) = (−∇ϕε,q , 1) = (−εAq · X, 1) + ε2 O(|X|2 ).
be the inner normal of Sε at the point p + (X, ϕε,q (X)).

Admissible functions. The class on which we will study problem (4) is

M(Sε ) = u ∈ H 1,2 (B, R3 ) : u(∂B) ⊂ Sε a.e. .



For u ∈ M(Sε ), we will also define the Hilbert subspace of H 1,2 (B, R3 ),

Mu (Sε ) = v ∈ H 1,2 (B, R3 ) : v(σ) ∈ Tu(σ) Sε a.e. σ ∈ ∂B .



Note that the subclass of M(Sε ) defined by

M2 (Sε ) = M(Sε ) ∩ H 2,2 (B, R3 )

9
is dense in M(Sε ) and it is a Hilbert manifold with tangent space at u ∈ M2 (Sε ) given by

Tu M2 (Sε ) = v ∈ H 2,2 (B, R3 ) : v(σ) ∈ Tu(σ) Sε ∀σ ∈ ∂B = Mu (Sε ) ∩ H 2,2 (B, R3 ),



which is dense in Mu .
Since we are dealing with free boundary surfaces, in order to have a functional whose Euler-Lagrange
equations are (4), following [24] one can correct the term V (u) by subtracting the volume of some surface
ũ contained in Sε and depending on u. First of all we define

M̃2 (Sε ) = ũ ∈ C(B; R3 ) : ũ(B) ⊂ Sε , and M̃2 (Sε ) = M̃(Sε ) ∩ H 2,2 (B, R3 )


with n o
Tũ M̃(Sε ) = ṽ ∈ H 2,2 (B, R3 ) : ṽ(z) ∈ Tũ (z)M̃2 (Sε ) ∀z ∈ B̄ .

Recall that an extension of u ∈ M(Sε ) is a map ũ ∈ M̃(Sε ) such that u = ũ on ∂B and an extension
operator is a smooth map ηε : D(ηε ) ⊂ M(Sε ) → M̃(Sε ) with open domain D(ηε ) such that ηε (u) is an
extension of u for all u ∈ D(ηε ).
For K > 0 we define the following open subset of M(Sε ) constituted by the elements of M(Sε ) having
Dirichlet energy less than K:
 Z 
M̄(Sε ) = u ∈ M(Sε ) : |∇u|2 < K .
B

We also define
M̄2 (Sε ) = M̄(Sε ) ∩ H 2,2 (B, Sε ).
We now state the following result which is in some sense a localized version of Lemma 2.1 in [24].

Lemma 3.1 If ε is sufficiently small, then for every u ∈ M̄(Sε ) there exists an extension operator ηε
defined in a neighborhood of u.
Proof. Let ū denote the harmonic extension of u,
(
∆ū = 0 in B,
ū = u on ∂B.

By standard elliptic regularity, ū is locally smooth in B and k∇ūk ≤ k∇uk. We first prove that
√
ess sup dist(ū(X), Sε ) ≤ C K,
X∈B

but it will be enough to prove

√
(25) dist(ū(0), Sε ) ≤ C K

because for every X ∈ B, there exists a conformal diffeomorphism (in G, see (14)) g : B → B such that
g(X) = 0 and g(∂B) = ∂B, therefore we may replace ū with ū ◦ g −1 .

By the mean value property of harmonic functions we have

Z
1
ū(0) = ū(σ)dσ,
2π ∂B
so
dist(ū(0), Sε ) ≤ |ū(0) − ū(σ 0 )| + dist(ū(σ 0 ), Sε ) for a.e. σ 0 ∈ ∂B.

10
The second term of the right hand side is zero since ū(∂B) ⊂ Sε , hence by Hölder inequality
Z
0 1
dist(ū(0), Sε ) ≤ ess 0inf |ū(0) − ū(σ )| ≤ |ū(0) − ū(σ 0 )|dσ 0
σ ∈∂B 2π ∂B
Z Z
1
≤ |ū(σ) − ū(σ 0 )|dσdσ 0
(2π)2 ∂B ∂B
 21
|ū(σ) − ū(σ 0 )|2
Z Z
2π 0
≤ dσdσ .
(2π)2 ∂B ∂B |σ − σ 0 |2

Since
|ū(σ) − ū(σ 0 )|2
Z Z Z
dσdσ 0 ≤ |∇ū|2 ,
∂B ∂B |σ − σ 0 |2 B

we have Z
1
dist(ū(0), Sε ) ≤ |∇ū|2 ,
2π B

hence (25) follows.

Consequently if ε is small, we can project ū on Sε to obtain an extension ũ defined by the following
implicit equation

(26) ũ(X) = ū(X) − ν ε (ũ(X))dε (ū(X)),

where ν ε (p) is the inner unit normal of Sε at a point p ∈ Sε . By the smoothness of ν ε and of dε in Σ rε0 ,
we have ũ ∈ M̃(Sε ). Moreover the mapping

u → ū → ũ

is smoothly defined in a neighborhood of u in M̄(Sε ) which gives rise to an extension ηε : D(ηε ) ⊂

M̄(Sε ) → M̃(Sε ) with ηε (u) = ũ.
Observe that if u ∈ M2 (Sε ) ∩ M̄(Sε ) and ηε (u) = ũ we have

ηε (u + ϕ) = ηε (u) ∀ϕ ∈ H01,2 (B, R3 ) ∩ H 2,2 (B, R3 )

and
hd ηε (u), ṽi = ṽ on ∂B ∀ṽ ∈ Tũ M̃2 (Sε ).
∂u ∂ ũ
Moreover since u = ũ on ∂B, =
∂t a.e. on ∂B and so by integration by parts, for every u ∈ M̄2 (Sε )
∂t
we have
(27) Z
1 ∂u
hdV (ηε (u)), vi = − ( ∧ u)·vds, ∀v ∈ Tu M̄2 (Sε ) = H01,2 (B, R3 ) ∩ H 2,2 (B, R3 ) + Tũ M̃2 (Sε ).
3 ∂B ∂t

Now we define for every u ∈ M̄(Sε ),

Z
1
(28) Iε (u) = |∇u|2 + 2[V (u) − V (ηε (u))].
2 B

Remark 3.2 For u smooth, it is clear that the term [V (u) − V (ηε (u))] represents the volume of the set
bounded by the image u and ∂Sε . As already explained in the introduction, M.Struwe in [24] needed to
impose some conditions on Ω to define the extension η. In our case instead, since we look for solutions
with bounded energy as ε → 0, no restriction on Ω is needed.

11
Note that by (27) and (10), the differential of Iε at a point u ∈ M̄2 (Sε ) is independent of ηε , namely we
have by density Z Z
hdIε (u), vi = ∇u·∇v + 2 (ux ∧ uy )·v ∀v ∈ Mu (Sε )
B B
and Z Z
2
hd Iε (u)w, vi = ∇w·∇v + 2 (wx ∧ uy + ux ∧ wy )·v ∀v, w ∈ Mu (Sε ).
B B

Hence Iε is smoothly defined on M̄2 (Sε ) moreover (see [24] Lemma 2.2) it easily follows the
Lemma 3.3 Let u ∈ M̄2 (Sε ), then

hdIε (u), vi = 0, ∀v ∈ Tu M̄2 (Sε ) iff u solves equation (4).

3.1 Construction of approximate solutions

We start by proving the following technical lemma.
Lemma 3.4 Let T = (Tij ) be a 2 × 2 symmetric matrix, and consider the following problem

Lω = (tr T )e3 − 2[(T1 · X)e3 ∧ Πy + (T2 · X)Πx ∧ e3 ]
 in B,
(29) ω(∂B) ⊂ R2 ,
 ∂ω

∂n (X) = (T · X, 0) X ∈ ∂B,

where Ti , i = 1, 2, denote the rows of the matrix T and L is the operator

Lu = −∆u + 2[ux ∧ Πy + Πx ∧ uy ].

Then (29) admits a solution ωT ∈ H which satisfies

(30) ||ωT ||2,2 ≤ C|T |∞ ,

where C is a fixed positive constant.

Proof. Problem (29) can be reformulated as

(31) d2 J(Π)[ω] = hT ,

where hT is ∈ H defined by duality as

Z Z
hhT , vi = (tr T e3 − 2[(T1 · X)e3 ∧ Πy + (T2 · X)Πx ∧ e3 ]) · v − (T · X, 0) · v ∀v ∈ H.
B ∂B

By Lemma 2.3, equation (31) is solvable if and only if hT is orthogonal to the vectors ei , and Ej for
i = 1, 2, j = 1, 2, 3. In this case, the condition can be easily verified from cancellation by oddness of the
integrals. Now the estimate (30) follows by standard elliptic regularity.
We will denote by ωq = (ωq1 , ωq2 , ωq3 ) the solution ωAq to (29) for every q ∈ ∂Ω and ωq0 = (ωq1 , ωq2 ). Now
define our approximate solutions to be

z ε,p (X) = Π(X) + p + εωq (X) + ϕε,q (X + εωq0 (X))e3 for every X ∈ B

and let
Ψε,q = εωq (X) + ϕε,q (X + εωq0 (X))e3 .
Then z ε,p has the following properties
z ε,p ∈ M̄2 (Sε ),

12
 ∂z ε,p
(32) (σ) = −Nε (σ + εωq0 (σ)) + O(ε)e3 + O(ε2 ).
∂n
Moreover we have

(33) z ε,p = Π + p + O(ε),

∂z ε,p
(34) = ei + O(ε) i = 1, 2,
∂pi
ε,p
where ∂z ε,q ε,q
∂pi , i = 1, 2 are the derivatives in the directions (1, 0, ϕx ) and (1, 0, ϕy ) respectively. Recalling
the expressions of Ei , see (15) and of G, see (14), we set
∂Ψε,q ◦gθ,(0,0)
E1ε,q = E1 + ∂θ |θ=0
,
∂Ψε,q ◦g0,(a1 ,0)
(35) E2ε,q = E2 + ∂a1 |a1 =0
,
∂Ψε,q ◦g0,(0,a2 )
E3ε,q = E3 + ∂a2 |a2 =0

and we define
Gzε,p = span {E1ε,q ; E2ε,q ; E3ε,q } .
We also need to provide some expansions of the extension z̃ ε,p of z ε,p . For simplicity, we assume that p
is the origin of R3 and we write z = z ε,p .
From now on it will be understood that Oq (X) denotes a smooth function depending only on q, X and
maybe on ε, uniformly bounded as ε → 0 for every X ∈ B and q ∈ S.
We have
ν ε (z̃(X)) = e3 − ε(Aq z̃ 0 (X), 0) + ε2 Oq (X),
where z̃ 0 stands for the first two components of z̃ : z̃ 0 = (z̃ 1 , z̃ 2 ), and
ε
dε (z̄(X)) = e3 ·z̄(X) + hÃq z̄(X), z̄(X)i + ε2 Oq (X).
2
Using (22) and the fact that the harmonic extension of the maps u(x, y) = x and v(x, y) = y are u and
v respectively, we get
ε
dε (z̄(X)) = ϕ̄ε,q (X) + εω̄q3 (X) − hAq X, Xi + ε2 Oq (X).
2
By (26), we obtain
ε
(36) z̃ ε,p (X) = (X + εω̄q0 (X), hAq X, Xi) + ε2 Oq (X).
2
Now having the approximate solutions z ε,p , we define the manifold

(37) Zε = {z ε,p : p ∈ ∂Ω}

with tangent space at z ε,p

∂z ε,p
 
Tzε,p Zε = span , i = 1, 2 .
∂pi
We let
∂z ε,p
 
⊥
(Tzε,p Zε ) = v ∈ Mzε,p (Sε ) : hv, i = 0 i = 1, 2
∂pi 1,2
so that Mzε,p (Sε ) = Tzε,p Zε ⊕ (Tzε,p Zε )⊥ , where h·, ·i1,2 is the scalar product in H 1,2 (B, R3 ).

13
Remark 3.5 Let v ∈ Mzε,p (Sε ). Since for every σ ∈ ∂B, Tzε,p (σ) Sε is spanned by the vectors (1, 0, ϕε,q
x (σ+
εωq0 (σ))) and (0, 1, ϕε,q
y (σ + εωq
0
(σ))) so for a.e. σ ∈ ∂B we have
0 0
v3 (σ) = v1 (σ)ϕε,q ε,q
x (σ + εωq (σ)) + v2 (σ)ϕy (σ + εωq (σ))

and hence by the trace theorem

(38) ||v3 || 1 ≤ Cε(||v1 ||1,2 + ||v2 ||1,2 ) ≤ Cε||v||1,2 .

H 2 ,2 (∂B)

Secondly we observe that there exists C > 0 depending only on Ω such that for every ε  1
Z Z
(39) |v|2 ≤ C |∇v|2 , ∀v ∈ (Tzε,p Zε )⊥ .
B B

In fact on the one hand letting v ∈ (Tzε,p Zε )⊥ we have by (34), B vi = oε (1)||v||1,2 and by Poincaré
R

inequality we have  Z 
 
||vi || ≤ C ||∇vi || +  vi  i = 1, 2.
B
On the other hand by (38) there holds

||v3 ||L2 (∂B) = oε (1)||v||1,2 ,

so using the following inequality

||v3 || ≤ C ||∇v3 || + ||v3 ||L2 (∂B)

we obtain
||v|| ≤ C||∇v|| + oε (1)||v||1,2 .

3.2 The finite-dimensional reduction

We define
 Z 
(40) W= v ∈ (Tzε,p Zε )⊥ : ∇v·∇Eiε,q = 0 i = 1, 2, 3 ,
B

so that by (39) the following decomposition holds

(41) Mzε,p (Sε ) = Tzε,p Zε ⊕ (Tzε,p Zε )⊥ = Tzε,p Zε ⊕ W ⊕ Gzε,p .

From now on, we will assume that p = 0 corresponds to the origin by replacing Sε with Sε − p. In order
to prove Theorem 1.1, we need to show that ||dIε (z ε,p )|| is small, and that d2 Iε (z ε,p ) is invertible on W.
We are going to prove these two facts.
Lemma 3.6 There exists C1 > 0 such that

|hdIε (z ε,p ), vi| ≤ C1 ε2 ||v||1,2 ∀v ∈ Mzε,p (Sε ).

Proof. First of all note that by (32) and (38), for every v ∈ Mzε,p (Sε )
∂z ε,p
Z 
 ≤ Cε2 ||v||1,2 .
 
(42)  ·vds

∂B ∂n 

We set Ψε,q = εωq (X) + ϕε,q (X + εωq0 (X))e3 then

1
Ψε,q = εωq (X) + ϕε,q (X)e3 + O(ε2 ) = ε(ωq (X) + hAq · X, Xie3 ) + O(ε2 ).
2

14
Then by the construction of ωq , Lemma 3.4, there holds

(43) |hLΨε,q , vi| ≤ Cε2 ||v||1,2 .

Let v ∈ Mzε,p (Sε ) then by integration by parts,

Z Z
ε,p
hdIε (z ), vi = ε,p
∇z ·∇v + 2 (zxε,p ∧ zyε,p )·v
B
Z B
∂z ε,p
Z Z Z
= − ∆Π·v − ∆Ψε,q ·v + ·vds + 2 (Πx ∧ Πy )·v
∂B ∂n
ZB B
Z B
ε,q ε,q
+2 [Ψx ∧ Πy + Πx ∧ Ψy ]·v + 2 (Ψx ∧ Ψε,q ε,q
y )·v.
B B

Now use 1. in Lemma 2.1 and the Hölder inequality to have

∂z ε,p
Z 
ε,p ε,q
·vds + 2||∇Ψε,q ||2 ||v||,
 
|hdIε (z ), vi| ≤ |hLΨ , vi| + 

∂B ∂n

hence from (42) and (43) the lemma follows.

Proposition 3.7 There exists a constant C2 > 0 such that for all ε > 0 small
Z
hd2 Iε (z ε,p )v, vi ≥ C2 ||∇v||2 ∀v ∈ W with ∇v·∇z ε,p = 0,
B
hd2 Iε (z ε,p )z ε,p , z ε,p i = −π + oε (1).

Proof. Since d2 Iε (z ε,p ) ' d2 I0 (Π), we may rely on Lemma 2.4.

By (33) and (34) and recalling that W = (TΠ Z)⊥ ∩ Hn (see (20)), it is enough to consider those v ∈
Mzε,p (Sε ) ∩ Hn satisfying Z Z
vi = 0, i = 1, 2 with ∇v·∇Π = 0.
B B
For such v, we define
(
∆v̄3 = 0 in B,
(44)
v̄3 = v3 on ∂B.

By Remark 3.5

(45) ||v̄3 ||1,2 = O(ε)||∇v||.

Let v̌3 = v3 − v̄3 ∈ H01 (B; R). We then have v = v̌ + v̄ where v̌ = (v1 , v2 , v̌3 ) and v̄ = (0, 0, v̄3 );
||∇v||2 = ||∇v̌||2 + ||∇v̄||2 and v̌ ∈ W . Since
Z Z Z Z
2 1 1 1
hd I0 (Π)v̌, v̄i = ∇v̌·∇v̄+ v̌·(v̄x ∧Πy +Πx ∧v̄y )+ v̄·(v̌x ∧Πy +Πx ∧v̌y )+ Π·(v̌x ∧v̄y +v̄x ∧v̌y ),
B 3 B 3 B 3 B
then
hd2 I0 (Π)v, vi = hd2 I0 (Π)v̌, v̌i + 2hd2 I0 (Π)v̌, v̄i + hd2 I0 (Π)v̄, v̄i
and by (45)
hd2 I0 (Π)v̌, v̄i = oε (1)||∇v||2 .
It is easy to verify that
hd2 I0 (Π)v̄, v̄i = ||∇v̄||2 ,
so we have

(46) hd2 I0 (Π)v, vi = hd2 I0 (Π)v̌, v̌i + oε (1)||v||21,2 + ||∇v̄||2 .

15
Let us estimate hd2 I0 (Π)v̌, v̌i. We define v = v̌ + φ, where
 
∇v̄3 · ∇Π3
φ = 0, 0, Π3 .
||∇Π3 ||2

Clearly φ ∈ H01 (B, R3 ) and φ = oε (1)||∇v||e3 , moreover v ∈ W and satisfies

R
B
∇v ·∇Π = 0. Furthermore
||∇v||2 = ||∇v̌||2 + ||∇φ||2 + oε (1)||∇v||2 hence by Lemma 2.4

hd2 I0 (Π)v, vi ≥ C||∇v||2 = C||∇v̌||2 + oε (1)||∇v||2 .

Now we have
hd2 I0 (Π)v̌, v̌i = hd2 I0 (Π)v, vi − 2hd2 I0 (Π)v, φi + hd2 I0 (Π)φ, φi
and by Hölder inequality

hd2 I0 (Π)v, φi = oε (1)||∇v||||∇v||, hd2 I0 (Π)φ, φi = ||∇φ||2 = oε (1)||∇v||2 ,

thus

(47) hd2 I0 (Π)v̌, v̌i ≥ C||∇v̌||2 + oε (1)||∇v||2 .

Therefore by (46) and (47), there holds

Z
2 2 2
hd I0 (Π)v, vi ≥ C̄2 ||∇v|| + oε (1)||∇v|| ∀v ∈ W, ∇v·∇z ε,p = 0.
B

This ends the proof.

We will also need the following Lemma.
Lemma 3.8 Let f ∈ L2 (B, R3 ), and u = (u1 , u2 , u3 ) ∈ Mzε,p (Sε ) satisfy
Z Z
(48) ∇u·∇v = f ·v ∀v ∈ Mzε,p (Sε ).
B B

Then u ∈ H 2,2 (B, R3 ) and there exists a constant C > 0 such that for every ε  1,

||u||2,2 ≤ C||f ||.

Proof. We may assume that z ε,p (x, y) = (x, y, ϕε,q (x, y)) for every (x, y) ∈ ∂B. Notice that H01,2 (B, R3 ) ⊂
Mzε,p (Sε ) so u satisfies ∆u = f a.e. in B.
Let ψ ∈ H 1,2 (B, R), then considering the test function v = (1, 0, ϕε,q
x )ψ ∈ Mz ε,p (Sε ) and using the fact
that

∇u · ∇v = ∇u1 · ∇ψ + ϕε,q 3 3
x (∇u · ∇ψ) + (∇u · ∇ϕ
ε,q
)ψ
= ∇u1 · ∇ψ + ∇(ϕε,q 3 3 ε,q 3 ε,q
x u ) · ∇ψ − u ∇ϕx · ∇ψ + (∇u · ∇ϕx )ψ

to have Z Z Z
1
∇(u + ϕε,q 3
x u )·∇ψ − u3
∇ϕε,q
x · ∇ψ = f˜ψ,
B B B

where f˜ = f 1 + f 3 ϕε,q 3 ε,q

x − ∇u · ∇ϕx . By the Gauss-Green formula
Z Z Z
u3 ∇ϕε,q
x · ∇ψ = div(u3
∇ϕ ε,q
x )ψ − (∇ϕε,q 3
x · n)u ψds.
B B ∂B

Setting w = u +1
ϕε,q
x u
3
∈ H (B, R), g = f˜ + div(u
1,2 3
∇ϕε,q
x ) ∈ L2 (B, R) and φ = (∇ϕε,q 3
x · n)u ∈
1
H 2 ,2 (∂B, R) then w satisfies
Z Z Z
∇w·∇ψ = − φψds + gψ, ∀ψ ∈ H 1,2 (B, R).
B ∂B B

16
It follows that w is a weak solution of the problem
(
∆w = g in B;
∂w
∂n = φ on ∂B,

hence the following properties hold

(49) u1 + ϕε,q 3
x u ∈H
2,2
(B, R);

(50) ||u1 + ϕε,q 3

x u ||2,2 ≤ C(||f || + ||ϕ
ε,q
||C 2 (B̄) ||u3 ||1,2 ).

By a similar argument testing on v = (0, 1, ϕε,q

y )ψ, we have

(51) u2 + ϕε,q 3
y u ∈H
2,2
(B, R);

(52) ||u2 + ϕε,q 3

y u ||2,2 ≤ C(||f || + ||ϕ
ε,q
||C 2 (B̄) ||u3 ||1,2 ).

Since u ∈ Mzε,p (Sε ), u3 (σ) = u1 (σ)ϕε,q 2 ε,q

x (σ) + u (σ)ϕy (σ) for a.e. σ ∈ ∂B. We multiply equation (49)
ε,q ε,q
by ϕx and (51) by ϕy , and take the sum to have
3
u3 (1 + |∇ϕε,q |2 ) ∈ H 2 ,2 (∂B, R).

We have proved that there exists u0 ∈ H 2,2 (B, R3 ) such that u − u0 ∈ H01 (B, R3 ). Thus u ∈ H 2,2 (B, R3 ).
Now let us estimate the H 2,2 norm of u3 . We write u = ǔ + ū where ū = ū3 e3 and ū3 is the harmonic
extension of u3 as in the proof of Proposition 3.7. We have by the trace theorem

||ū3 ||2,2 ≤ C||u3 || 23 ,2 ≤ C̃ε||u||2,2 .

If we consider the test functions in the form v = ψe3 = (0, 0, ψ) ∈ Mzε,p (Sε ) for every ψ ∈ H01,2 (B, R),
it follows that
||ǔ3 ||2,2 ≤ C(||f || + ||ū3 ||2,2 ).
The two previous inequalities give

||u3 ||2,2 ≤ C(||f || + C̃ε||u||2,2 ).

Hence using (50) and (52) we obtain

||u||2,2 ≤ C(||f || + C̄ε||u||2,2 ),

so we have the result.

We let
P : Mzε,p (Sε ) → W
be the projection onto W. Then the operator

Lε,p := P ◦ d2 Iε (z ε,p ) ◦ P : W → W

is uniformly invertible, namely the following result holds.

Corollary 3.9 For every p ∈ Sε , for every ε  1, the operator Lε,p is invertible and there exists a
constant C̄2 > 0 such that

||L−1
ε,p v||2,2 ≤ C̄2 ||v||1,2 ∀v ∈ W, p ∈ Sε .

17
 ε,p
∇z ε,p ·∇ ∂z
R
Proof. Since B ∂pi = oε (1), one has

||∇z ε,p − ∇P z ε,p || = oε (1), Lε,p z ε,p = −πP z ε,p + oε (1).

Now following [3]-(section 8.4) and according to Remark 3.5-(39), if we decompose W = V1 ⊕ V2 , where
 Z 
ε,p ε,p
V1 = RP z ; V2 = v ∈ Mz (Sε ) :ε,p ∇v · ∇z = 0 ∩ W; V1 ⊥ V2 ,
B

then, in matrix form with respect to V1 and V2 , Lε,p can be written as

 
−πId + oε (1) oε (1)
,
oε (1) Bε,p

where Bε,p satisfies, by Proposition 3.7, ||Bε,p v||1,2 ≥ C2 ||v||21,2 for every v ∈ V2 . Hence there exists
C̃2 > 0 such that for ε  1

(53) ||L−1
ε,p v||1,2 ≤ C̃2 ||v||1,2 ∀v ∈ W.

We set u = L−1 2
ε,p v so that P d Iε (z
ε,p
)[u] = P v, or equivalently d2 Iε (z ε,p )[u] − v ∈ Tzε,p Zε ⊕ Gzε,p . Then
by (41), there exist αi , i = 1, 2 and βj , j = 1, 2, 3 such that
2 3
∂z ε,p ∂z ε,p
Z X Z   X Z
hd2 Iε (z ε,p )u, φi − v·φ = αi ∇ ·∇φ + ·φ + βj ∇Ejε,q ·∇φ ∀φ ∈ Mzε,p (Sε ).
B i=1 B ∂pi ∂pi j=1 B

We first estimate αi and βj . By equation (34) and (35), we may write

∂z ε,p
= ei + εOqi (X) i = 1, 2,
∂pi
Ejε,q = Ej + εOqj (X) j = 1, 2, 3,

therefore for every φ ∈ Mzε,p (Sε )

2
X Z 3
X Z 2
X Z 3
X Z Z
αi ei · φ + βj ∇Ej ·∇φ + ε αi Oqi ·φ+ε βj Oqj 2
· φ = hd Iε (z ε,p
)u, φi − v·φ
i=1 B j=1 B i=1 B j=1 B B

thus by the mutual orthogonality of ei , Ej for i = 1, 2, j = 1, 2, 3 it follows by (53) that

 
X2 3
X X2 3
X
|αi | + C̄ |βj | − εC̃  |αi | + |βj | ≤ C(||u||1,2 + ||v||1,2 ) ≤ C||v||1,2 .
i=1 j=1 i=1 j=1

Finally u satisfies
 
2 ε,p j 2
∂z ε,p
Z Z Z Z
X ∂z X X
βi Ejε,q  ·∇φ = 2 ux ∧ zyε,p + zxε,p ∧ uy ·φ +


∇ u − αi − v·φ + αi ·φ
B i=1
∂pi j=1 B B i=1 B ∂pi

for all φ ∈ Mzε,p (Sε ). Therefore by Lemma 3.8 and (53),

||u||2,2 ≤ C̄2 (||u||1,2 + ||v||1,2 ) ≤ C̄2 ||v||1,2 .

This concludes the proof.

Notice that the manifold Zε is not critical for Iε . Nevertheless by Lemma 3.6 and Corollary 3.9 we will
be able to perform a finite-dimensional reduction of problem (4) on the manifold Zε .

18
Proposition 3.10 Suppose ∂Ω is C m , m ≥ 3. Let Iε be the functional defined in (28) and W in (40).
Then for ε > 0 small, there exists a unique w = w(ε, p) ∈ W ∩ Tzε,p M̄2 (Sε ) such that w + z ε,p ∈ M̄2 (Sε )
and dIε (z ε,p + w) ∈ Tzε,p Zε . The function (ε, p) → w(ε, p) is of class C m with respect to p. Moreover,
the function Fε (p) = Iε (z ε,p + w(ε, p)) is of class C m in p and satisfies

Fε0 (p0 ) = 0 =⇒ hdIε (z ε,p0 + w(ε, p0 )), vi = 0 ∀v ∈ Tzε,p0 +w(ε,p0 ) M̄2 (Sε ).

Proof. We have since Tzε,p Zε and Gzε,p are subspaces of Tzε,p M̄2 (Sε ),

Tzε,p M̄2 (Sε ) = Tzε,p Zε ⊕ W ∩ Tzε,p M̄2 (Sε ) ⊕ Gzε,p .

Let P : Tzε,p M̄2 (Sε ) → W ∩ Tzε,p M̄2 (Sε ) the projection onto W. We want to solve the problem:

ε,p
P dIε (z + w) = 0,

(54) P w = w,

w + z ε,p ∈ M̄2 (Sε ),


which will be done using the contraction mapping theorem. Note that equation (54) is equivalent to

P d2 Iε (z ε,p )[w] = P dIε (z ε,p ) + P Nε (w)

with Nε (w) defined by duality in this way: for every φ ∈ Mzε,p (Sε )
Z
hNε (w), φi = hdIε (z ε,p + w), φi + hd2 Iε (z ε,p )w, φi − hdIε (z ε,p ), φi = 2 wx ∧ wy ·φ
B
Z
= 2 [wx ∧ φy + φx ∧ wy ]·w
B

so we are led to solve the fixed point problem

w = −Lε,p −1 {P dIε (z ε,p ) + P Nε (w)}.

Observe that for every φ ∈ Mzε,p (Sε ) and w ∈ Tzε,p M̄2 (Sε ), we have by the Hölder inequality and
Sobolev embeddings

(55) |hNε (w), φi| ≤ C||w||22,2 ||φ||1,2 .

From this formula if we define the map T (w) = −Lε,p −1 {P dIε (z ε,p ) + P Nε (w)}. Then by Corollary 3.9,
T is defined from W ∩ Tzε,p M̄2 (Sε ) to itself. We are going to find a fixed point of T in the following ball
on Tzε,p M̄2 (Sε ),

BR = w ∈ W ∩ Tzε,p M̄2 (Sε ) : ||w||2,2 ≤ ε2 R, w + z ε,p ∈ M̄2 (Sε )



where R > 0 will be determined later. Notice that BR is not empty because H 2,2 (B, R3 ) ∩ H01,2 ⊂
Tzε,p M̄2 (Sε ). Let w ∈ BR . Using (55), Corollary 3.9 and Lemma 3.6 one has the following estimate

||T (w)||2,2 ≤ C̄2 {ε2 C1 + Cε4 R2 }

and thus, if we choose R sufficiently large and ε small respectively, T (w) ∈ BR .

Now for every w1 , w2 ∈ Tzε,p M̄2 (Sε ) there holds
Z
hNε (w1 ) − Nε (w2 ), φi = 2 [wx1 ∧ φy + φx ∧ wy1 ]·w1 − [wx2 ∧ φy + φx ∧ wy2 ]·w2 ,
ZB Z
= 2 [wx ∧ φy + φx ∧ wy ]·(w − w ) + 2 [(w1 − w2 )x ∧ φy + φx ∧ (w1 − w2 )y ]·w2 .
1 1 1 2
B B

19
Reasoning as for (55) we get,

|hNε (w1 ) − Nε (w2 ), φi| ≤ C(||w1 ||2,2 + ||w2 ||2,2 )||w1 − w2 ||2,2 ||φ||1,2

from this, if ε is small enough, the map T is a contraction in BR , yielding the existence. Moreover since
also the implicit function theorem applies, the C m regularity of w follows from the smoothness of Iε .
To prove the last assertion, we set Z̃ε = {z ε,p + w(ε, p) : p ∈ ∂Ω}. First notice that by construction
Tzε,p M̄2 (Sε ) = Tzε,p +w(ε,p) M̄2 (Sε ). We claim that Z̃ε is a natural constraint for Iε . In fact, suppose that
z ε,p0 + w(ε, p0 ) is a critical point of Iε |Z̃ε . Then dIε (z ε,p0 + w(ε, p0 )) is perpendicular to Tzε,p0 +w(ε,p0 ) Z̃ε
thus almost perpendicular to Tzε,p0 Zε . Since by construction of Z̃ε , dIε (z ε,p0 +w(ε, p0 )) ∈ Tzε,p0 Zε ⊕Gε,p0 ,
therefore dIε (z ε,p0 + w(ε, p0 )) must vanish unless

0 6= dIε (z ε,p0 + w(ε, p0 )) ∈ Gε,p0 .

Following ([6] section 2), let us show that the latter case cannot happen. Indeed, suppose there exist
βj ∈ R, j = 1, 2, 3 such that for every φ ∈ Mzε,p (Sε )
3
X Z
(56) hdIε (z ε,p0 + w(ε, p0 )), φi = βk ∇Ekε,q0 ·∇φ.
k=1 B

Without loss of generality, we may assume that Ejε,q0

⊥ Ekε,q0 for k 6= j, namely
Z
∇Ejε,q0 ·Ekε,q0 = 0.
B

Let g1 (t), g2 (t), g3 (t) ∈ G be smooth paths of conformal transformations such that gj (0) = Id and
d ε,p0
dt z ◦ gj (t)|t=0 = Ejε,q0 for j = 1, 2, 3, where Ej are defined in (15). With an abuse of notation, we
write z = z ε,p0 and w = w(ε, p0 ). We have by conformal invariance of the Dirichlet integral
Z
d
∇(z ◦ gk (s) ◦ gj (t))·∇(w ◦ gj (t)) = 0,
dt B
thus
Z   Z  
d d
− ∇ (z ◦ gk (s)) ·∇ w ◦ gj (t)|t=0 = ∇ z ◦ gk (s) ◦ gj (t)|t=0 ·∇w k, j = 1, 2, 3.
B dt B dt
Consequently, since w ⊥ Gzε,q0
Z   Z  2 
ε,q0 d d
(57) − ∇Ek ·∇ w ◦ gj (t)|t=0 = ∇ z ◦ gk (s) ◦ gj (t)|(t,s)=(0,0) ·∇w = 0.
B dt B dsdt

Now let j ∈ {1, 2, 3}. Then again by invariance of Iε with respect to the conformal group G, we have by
(56) and (57)
d d
0 = Iε (z ◦ gj (t))|t=0 = hdIε (z), Ejε,q0 i + hdIε (z), w ◦ gj (t)|t=0 i
dt dt
3 Z  
ε,q0 2
X ε,q0 d
= k∇Ej k βj + βk ∇Ek ·∇ w ◦ gj (t)|t=0 = k∇Ejε,q0 k2 βj .
B dt
k=1

This shows that βj = 0 for every j ∈ {1, 2, 3}.

Recall the expansion of the distance function given in Section 3:

ε
dε (X̃) = e3 ·X̃ + hÃq X̃, X̃i + ε2 P̃ q (X̃) + ε3 O(|X̃|4 ) ∀X̃ ∈ B rε0 ,
2

20
so by construction of the natural constraint Z̃ε , any solution u := z ε,p + wε,p ∈ Z̃ε of (4) is conformal
and is of class C 1,β (B̄, R3 ) by [12] Proposition 4. It is embedded if and only if dε (u(X)) > 0 for every
X ∈ B. This is the case because by construction we have u(∂B) ⊂ Sε so dε (u(σ)) = 0 for very σ ∈ ∂B.
Moreover
u(σ − tn) − u(σ)
u = Π + εOq (X); lim+ = e3 + O(ε) ∀σ ∈ ∂B
t→0 t
and then
dε (u(σ − tn))
dε (u(X)) = Π3 + O(ε) = µ(X) − 1 + O(ε) ∀X ∈ B; lim+ = 1 + O(ε) ∀σ ∈ ∂B,
t→0 t
where µ is defined in Lemma 2.1, satisfies 1 < µ < 2 in B. Hence it follows that if ε is sufficiently small,
u(B) ⊂ 1ε Ω.

4 Proof of Theorem 1.1 and Theorem 1.2

In view of Proposition 3.10, we can obtain existence of solutions to (4) by finding critical points of the
functional Fε (p). The following lemmas are devoted to the expansions of Fε with respect to p and ε.
Lemma 4.1 For ε small one has
π
V (z̃ ε,p ) = − εH(q) + ε2 G2 (q) + O(ε3 ),
24
where G2 : S → R is smooth depending only on q.
Proof. Recall from (36) that
ε
z̃ ε,p (X) = (X + εω̄q0 (X), hAq X, Xi) + ε2 Oq (X)
2
and thus
ε
= −ε(X + εω̄q0 )·Aq X + hAq X, Xi + ε2 Oq (X) + O(ε3 )
(z̃xε,p ∧ z̃yε,p )·z̃ ε,p
2
ε
= − hAq · X, Xi + ε2 Oq (X) + O(ε3 )
2
ε2
Z Z
ε επ
V (z̃ ε,p ) = − hAq X, Xi + Oq (X) + O(ε3 ) = − H(q) + ε2 G2 (q) + O(ε3 )
6 B 3 B 24
and the Lemma follows.

Lemma 4.2 For ε small, the following expansions holds

2π π
Iε (z ε,p ) = − εH(q) + ε2 G(q) + O(ε3 ),
3 12
where G : S → R is smooth depending only on q.
Proof. We recall that
Z Z
1 2
Iε (z ε,p ) = |∇z ε,p |2 + z ε,p ·(zxε,p ∧ zyε,p ) − 2V (z̃ ε,p ).
2 B 3 B

We set Ψε,q = εωq (X) + ϕε,q (X + εωq0 (X))e3 . Let us expand term by term the right hand side of the
above equality
Z Z Z Z
1 1 1
|∇z ε,p |2 = |∇Π|2 + |∇Ψε,q |2 + ∇Π · ∇Ψε,q ;
2 B 2 B 2 B B
Z Z
1 2
= |∇Π| + ∇Π · ∇Ψε,q + O(ε3 ).
2 B B

21
Z Z
z ε,p ·(zxε,p ∧ zyε,p ) z ε,p · Πx ∧ Πy + Πx ∧ Ψε,q ε,q 3


= y + Ψx ∧ Πy + O(ε )
B B
Z Z Z
ε,q ε,q
Ψε,q ·(Πx ∧ Πy ) + O(ε3 ).


= Π·(Πx ∧ Πy ) + Π· Πx ∧ Ψy + Ψx ∧ Πy +
B B B

Using 1. and 5. of Lemma 2.1 we have

Z Z Z Z Z
1 1 1 ∂Π ε,q
z ε,p ·(zxε,p ∧ zyε,p ) = − |∇Π|2 − ∇Π · ∇Ψε,q − ∇Π · ∇Ψε,q + ·Ψ ds + O(ε3 )
B 2 B B 2 B 2 ∂B ∂n
Z Z Z
1 2 3 ε,q 1 ∂Π ε,q
= − |∇Π| − ∇Π · ∇Ψ + ·Ψ ds + O(ε3 ).
2 B 2 B 2 ∂B ∂n
Hence adding up Z Z
1 1
|∇Π|2 − ϕε,q ds − 2V (z̃ ε,p ) + O(ε3 ).
Iε (z ε,p ) =
6 B 3 ∂B
Now by Lemma 4.1, property 6. of Lemma 2.1 and the following computations
Z
P q (X)ds = 0,
∂B
Z Z Z
ε π
ϕε,q ds = hAq · X, Xids + ε2 hAq · X, ωq0 ids + O(ε3 ) = εH(q) + ε2 G1 (q) + O(ε3 ),
∂B 2 ∂B ∂B 2
we obtain
2π π
Iε (z ε,p ) =
− εH(q) − 2V (ηε (z ε,p )) + ε2 G1 (q) + O(ε3 ),
3 6
1 q 0
R
where G1 (q) = 3 ∂B hA · X, ωq ids. Therefore the conclusion follows form Lemma 4.1.

Proof of Theorem 1.1 and Theorem 1.2

First of all we have

Fε (p) = Iε (z ε,p + w) = Iε (z ε,p ) + hdIε (z ε,p ), wi + O(||w||2 ).
Using Lemma 3.6 and the fact that ||w||2,2 ≤ Rε2 we infer that
Fε (p) = Iε (z ε,p ) + O(ε4 ).
Hence Lemma 4.2 yields
2π π
(58) Fε (p) = − εH(εp) − ε2 G(q) + O(ε3 )
3 12
with H of class C m−2 and m − 2 ≥ 1 by assumption. It follows that if Q0 is a strict local maximum or
minimum of H, Fε will have critical points pε for which εpε → Q0 as ε → 0. Furthermore we have
π
(59) ||Fε + εH||C 1 (Sε ) = O(ε2 ).
12
Consequently if Q0 is a non-degenerate critical point of H then we can compute the degree
∂H ∂H
d( , Bδ (p0 ), 0) = i( , p0 , 0) 6= 0
∂p ∂p
for some small δ > 0. Therefore by (59) d( ∂F ∂H
∂p , Bδ (p0 ), 0) = d( ∂p , Bδ (p0 ), 0) 6= 0, yielding a critical point
ε

of Fε .
The proof of Theorem 1.2 follows immediately by Proposition 3.7 and the fact that Fε is C 1 so it has at
least cat(∂Ω) critical points.

22
Remark 4.3 As mentioned in Section 1, we can compare our result to the one of [28]. The expansions
of the mean curvature of a geodesic sphere of radius ε contains only terms of oder ε2 and higher, see
[28], equation (1.4). If we perform our construction in a manifold, by (58) it is evident that the boundary
mean curvature would determine the main terms in the Lyapunov-Schmidt reduction. Other geometric
quantities, like the scalar curvature as the second fundamental form of the boundary could be relevant for
the location of solutions only when the mean curvature is constant.

References
[1] R. Adams, Sobolev Spaces. Academic Press, New York-San Francisco-London (1975).
[2] A. Ambrosetti and M. Badiale, Variational perturbative methods and bifurcation of bound states
from the essential spectrum, Proc. Roy. Soc. Edinburgh Sect. A 128, (1998), 1131-1161.
[3] A. Ambrosetti and A. Malchiodi, Perturbation Methods and Semilinear Elliptic Problems on Rn .
Progress in Mathematics, Birkhäuser Verlag, Basel-Boston-Berlin (2005).
[4] H. Brezis and J.M. Coron, Convergence of solutions of H-systems or how to blow bubbles, Arch.
Rational Mech. Anal. 89, (1985), 21-56.
[5] P. Caldiroli, H-bubbles with prescribed large mean curvature, Manuscripta Math. 113 (2004), 125-
142.
[6] P. Caldiroli and R. Musina, H-bubbles in a perturbative setting: The finite-dimensional reduction
method, Duke Math. Journal, V. 122, No3, 2004.
[7] S. Chanillo and A. Malchiodi, Asymptotic Morse theory for the equation ∆v = 2vx ∧ vy , Comm.
Anal. Geom., V.13, no. 1, (2005), 187-251.
[8] J. Douglas, Solution of the Problem of Plateau, Trans. AMS 33 (1931), 263-321.
[9] V. Felli, A note on the existence of H-bubbles via perturbation methods. Rev. Mat. Iberoamericana
21 (2005), no. 1, 163-178.
[10] M. Grüter, Über die Regularität schwakher Lösungen des Systems ∆x = 2H(x)xu ∧ xv , Dissertation,
Düsseldorf (1979).
[11] M. Grüter, regularity of weak H-surfaces, J. Reine Angew. Math.329 (1981), 1-15.
[12] M. Grüter, S. Hildebrandt and J.C.C. Nitsche, Regularity for surfaces of constant mean curvature
with free boundaries, Acta. Math. 156, (1986), 119-152.
[13] M. Grüter and J.Jost, On embedded minimal discs in convex bodies, Annales de l’institut Henri
Poincaré (C) Analyse non linaire, 3, no. 5, (1986), 345-390.
[14] F. Duzaar and J. F. Grotowski, Existence and regularity for higher-dimensional H-systems Duke
Math. J. 101, no. 3 (2000), 459485.
[15] S. Hildebrandt, On the Plateau problem for surfaces of constant mean curvature, Comm. Pure Appl.
Math. 23, (1970), 97-114.
[16] T. Isobe, On the asymptotic analysis of H-systems. I. Asymptotic behavior of large solutions. Adv.
Differential Equations 6 (2001), no. 5, 513-546.
[17] U. Massari, Esistenza e regolaritá delle ipersurfici di curvature media assegnata in Rn , Arch. Rat.
Mech. Anal. 55, (1974), 357-382.

23
[18] R. Musina, The role of the spectrum of the Laplace operator on S 2 in the H-bubble problem, J.
Anal. Math. 94, (2004), 265-291.

[19] T. Rado On Plateau’s problem, Ann. of Math. 31, (1930), 457-469.

[20] J. Sacks and K. Uhlenbeck, the existence of minimal immersions of 2-spheres, Ann. Math. 113 (1981),
1-24.
[21] K. Steffen, On the non-uniqueness of surfaces with prescribed constant mean curvature spanning a
given contour, Arch. Rat. Mech. 94, (1986), 101-122.
[22] M. Struwe, Large H-surface via the mountain-pass-lemma, Math. Ann. 270, (1985), 441-459.
[23] M. Struwe, Non-uniqueness in the Plateau problem for surfaces of constant mean curvature, Arch.
Rat. Mech. Anal. 93, (1986), 135-157.

[24] M. Struwe, The existence of surfaces of constant mean curvature with free boundaries, Acta Math.
160 (1988), no. 1-2, 19-64.
[25] M. Struwe, Plateau’s problem and the calculus of variations, Mathematical Notes, 35. Princeton
University Press, Princeton, NJ, 1988.
[26] M. Struwe, On a free boundary problem for minimal surfaces, Inv. Math. 75 (1984), 547-560.

[27] J. E. Taylor, Boundary regularity for solutions to various capillarity and free boundary problems,
Comm. PDE, 2, (1977), 323-357.
[28] R. Ye, Foliation by constant mean curvature spheres. Pacific J. Math. 147 (1991), no. 2, 381–396.

24