                                                 AN EFFECTIVE HAMILTONIAN FOR THE EIGENVALUE
                                               ASYMPTOTICS OF THE ROBIN LAPLACIAN WITH A LARGE
                                                                 PARAMETER

                                                               KONSTANTIN PANKRASHKIN AND NICOLAS POPOFF


                                                  Abstract. We consider the Laplacian on a class of smooth domains Ω ⊂ Rν , ν ≥ 2,
                                                  with attractive Robin boundary conditions:
                                                                                            ∂u
                                                                          QΩα u = −∆u,         = αu on ∂Ω, α > 0,




arXiv:1502.00877v2 [math.SP] 28 Apr 2015
                                                                                            ∂n
                                                  where n is the outer unit normal, and study the asymptotics of its eigenvalues Ej (QΩ
                                                                                                                                      α ) as
                                                  well as some other spectral properties for α → +∞ We work with both compact domains
                                                  and non-compact ones with a suitable behavior at infinity. For domains with compact
                                                  C 2 boundaries and fixed j, we show that
                                                                            Ej (QΩ       2
                                                                                 α ) = −α + µj (α) + O(log α),

                                                  where µj (α) is the j th eigenvalue, as soon as it exists, of −∆S −(ν−1)αH with (−∆S ) and
                                                  H being respectively the positive Laplace-Beltrami operator and the mean curvature on
                                                  ∂Ω. Analogous results are obtained for a class of domains with non-compact boundaries.
                                                  In particular, we discuss the existence of eigenvalues in non-compact domains and the
                                                  existence of spectral gaps for periodic domains. We also show that the remainder estimate
                                                  can be improved under stronger regularity assumptions.
                                                     The effective Hamiltonian −∆S − (ν − 1)αH enters the framework of semi-classical
                                                  Schrödinger operators on manifolds, and we provide the asymptotics of its eigenvalues
                                                  in the limit α → +∞ under various geometrical assumptions. In particular, we describe
                                                  several cases for which our asymptotics provides gaps between the eigenvalues of QΩ  α for
                                                  large α.
                                                       2010 Mathematics Subject Classification: 35P15, 35J05, 49R05, 58C40.




                                                                                   1. Introduction
                                              Let Ω ⊂ Rν , ν ≥ 2, be an open set with a sufficiently regular boundary S := ∂Ω. For
                                           α ∈ R, denote by QΩ                      Ω
                                                                α the operator Qα u = −∆u on the functions u defined in Ω and
                                           satisfying the Robin boundary condition
                                                                                   ∂u
                                                                                      = αu on S,
                                                                                   ∂n
                                           where n is the outer unit normal at S. More precisely, QΩ α is the self-adjoint operator in
                                           L2 (Ω) associated with the quadratic form qαΩ defined on the domain D(qαΩ ) = H 1 (Ω) by
                                                                                   Z              Z
                                                                      qαΩ (u, u) =    |∇u|2 dx − α u2 dS,
                                                                                        Ω                  S
                                           where dS stands for the (ν − 1)-dimensional Hausdorff measure on S, which is closed
                                           and semibounded from below under suitable assumptions (e.g. if S is compact or with a
                                           suitable behavior at infinity, see below), and we denote by EjΩ (QΩ         th eigenvalue of
                                                                                                             α ) the j
                                             Ω
                                           Qα below the bottom of the essential spectrum, as soon as it exists. The aim of the paper
                                           is to obtain new results on the asymptotics of the eigenvalues as α tends to +∞.
                                              The problem appears in various applications, such as reaction-diffusion processes [25]
                                           and the enhanced surface superconductivity [14], and the related questions were already
                                                                                              1
                                                                                              2

discussed in the previous works by various authors. Let us present briefly the state of art
for compact domains. It was shown in [25,26] that for piecewise smooth Liptschotz domain
one has E1 (QΩ             2      2
                α ) = −CΩ α + o(α ) as α → +∞, where CΩ ≥ 1 is a constant depending
on the geometric properties of Ω. In particular, CΩ = 1 for C 1 domains, see [5, 27]. More
detailed asymptotic expansions for some specific non-smooth domains were considered
in [17, 26, 31]. As for smooth domains, a more detailed result was obtained first in [9, 30]
for ν = 2 and then in [32] for any ν ≤ 2: if the domain is C 3 and j ∈ N is fixed, then
                        Ej (QΩ       2
                             α ) = −α − (ν − 1)Hmax α + O(α
                                                           2/3
                                                               ),                        (1.1)
where Hmax is the maximum of the mean curvature H of the boundary (the exact defini-
tion will be recalled below). Remark that this asymptotics together with isompermetric
inequalities for the mean curvature have played an important role for the so-called reverse
Faber-Krahn inequality, see [11,32]. The result (1.1) was also obtained in [8] for a class of
non-compact planar domains.
   Although the asymptotics (1.1) shows the influence of the geometry on first orders, it
is not sufficient to distinguish the influence of the number j of the eigenvalue, and to
estimate the gap between eigenvalues. For ν = 2, a complete asymptotic expansion of the
eigenvalues of the form
                                       s
                                                           N
                                          H ′′ (s0 ) 1/2 X            j       N
 Ej (QΩα ) = −α  2
                   − H max α + (2j − 1)             α   +     γj,k α− 2 + o(α− 2 ), γj,k ∈ R,
                                              2
                                                          k=0
                                                                                          (1.2)
where j ∈ N is arbitrary but fixed, was proved in [15] proved under the assumption that
the curvature s 7→ H(s) admits a unique non-degenerated maximum at s0 and the second
derivative is taken with respect to the arc-length. Such a hypothesis is reminiscent of
several works about the first eigenvalues of the magnetic Laplacian in the semi-classical
limit, see [16], which involves the localization of the eigenfunctions at the boundary and
allow expansion of the associated eigenvalues. More precisely, such a maximum of the
curvature acts as a potential well for Schrödinger operators in the harmonic approximation,
see [7].
   In order to describe our results, let us introduce the necessary notation and the class
of domains we consider in this article. Recall that s 7→ n(s) is the Gauss map on S, i.e.
n(s) is the outward pointing unit normal vector at s ∈ S. Consider the shape operator
Ls at s ∈ S, which is defined by Ls := dn(s) : Ts S → Ts S, and let κ1 (s), . . . , κν−1 (s) be
its eigenvalues, called the principal curvatures. Our results will be valid for the so-called
C k -admissible domains defined as follows:
Definition 1.1. Let k ≥ 2. A domain Ω ⊂ Rν is called C k -admissible, if its boundary is
C k , and, in addition, the following holds:
  (H1) There exists δ > 0 such that the map Φ defined by
                                                                               
                Σ := S × (0, δ) ∋ (s, t) 7→ Φ(s, t) := s − tn(s) ∈ Φ S × (0, δ)    (1.3)
       is a diffeomorphism and its image is contained in Ω.
  (H2) The curvatures s 7→ κi (s) are in L∞ (S). Moreover, if k ≥ 3, their gradients are
       also bounded.
   The assumption (H1) is quite standard if one deals with non-compact domains, and it
is sometimes called the non-overlap condition, cf. e.g. [3]. We remark that any C k domain
with a compact boundary is C k -admissible.
   In what follows we denote by K(s) the sum of the principal curvatures:
                           K(s) = κ1 (s) + · · · + κν−1 (s) ≡ tr Ls .
                                                                                                 3

Remark that the quantity H := K/(ν − 1) is exactly the mean curvature on S. We will
denote
                               Kmax := sup K(s).
                                           s∈S
                        Ω
Note that the operator Qα can have a non-empty essential spectrum, and it is more conve-
nient to work with the Rayleigh quotients instead of the eigenvalues. For this purpose, re-
call the min-max principle for the eigenvalues of self-adjoint operators, see e.g. [6, Sec. 4.5]:
Let Q be a lower semibounded self-adjoint operator in a Hilbert space H and q be its qua-
dratic form. Denote
                E(Q) := inf specess Q      (we use the convention inf ∅ = +∞),
                                         q(u, u)
               Ej (Q) := min max                 ,   j ∈ N,                                  (1.4)
                          L⊂D(q), u∈L,   hu, ui
                          dim L=j u6=0
then:
      • if Ej (Q) < E(Q), then Ej (Q) is the j th eigenvalue of Q,
      • if Ej (Q) ≥ E(Q), then Ek (Q) = E(Q) for all k ≥ j.
   Our main result gives a comparison between the Rayleigh quotients of QΩ
                                                                         α with those
of an auxiliary Schrödinger operator acting on the boundary and in which the Robin
coefficient α appears as a coupling constant:
Theorem 1.2. Let ν ≥ 2 and Ω ⊂ Rν be a C 2 -admissible domain. Furthermore, let −∆S
denote the positive Laplace-Beltrami operator on S viewed as a self-adjoint operator in
L2 (S, dS). For any fixed j ∈ N one has
                 Ej (QΩ       2
                      α ) = −α + Ej (−∆S − αK) + O(log α),            α → +∞.                (1.5)
  The proof is presented in Sections 4 and 5.1. Furthermore, in Section 6 we show that
the remainder estimate can be improved under additional assumptions:
Theorem 1.3. Let the assumptions of Theorem 1.2 hold. In addition, assume that Ω is
C 3 -admissible and that K reaches its maximum, then for any fixed j ∈ N one has
                   Ej (QΩ       2
                        α ) = −α + Ej (−∆S − αK) + O(1),            α → +∞.                  (1.6)
   Let us emphasize on the fact that no assumptions are done on the behavior of K near
the set K −1 (Kmax ) for the above results. Using various methods available for the study
of the effective Hamiltonian −∆S − αK one can deduce more precise asymptotics under
various assumptions, which improve the results of preceding works of weaken the respective
assumptions. In particular, as a generalization of (1.1) in Section 7 we obtain:
Corollary 1.4. Let Ω ⊂ Rν be a C 2 -admissible domain, then for each fixed j ∈ N we have
Ej (QΩ       2
     α ) = −α − Kmax α + o(α).
Remark 1.5. If the boundary S is compact, then either specess QΩ           α = ∅ (if Ω is bounded)
or specess QΩ
            α  =  [0, +∞)   (if Ω is  unbounded).    In  the latter  case, it is standard to check
       Ω
that Qα has at most finitely many negative eigenvalues. On the other hand, for any fixed
j one has Ej (QΩ                                                          Ω          Ω
                α ) < 0 if α is sufficiently large, in particular, Ej (Qα ) < E(Qα ). Therefore,
                                      Ω                             Ω
by the min-max principle, each Ej (α) is an eigenvalue of Qα if α is sufficiently large.
   The preceding observation does not hold for domain with non-compact boundaries. In
particular, in [8] one can find various examples of domains Ω with curved non-compact
boundaries such that the respective operators QΩ        α have a purely essential spectrum for
any α > 0. Nevertheless, the existence of eigenvalues can be guaranteed by an additional
assumption:
                             (H3): K∞ := lim sup K(s) < Kmax ,                                (1.7)
                                                 s→∞
                                                                                               4

which allows one to prove the following result extending several estimates of [8]:
Corollary 1.6. Let Ω ⊂ Rν be C k -admissible with non-compact boundary and satisfy
(1.7), then for any N ∈ N there exists αN > 0 such that for α > αN the operator QΩ
                                                                                 α has
                                                                          th
at least N eigenvalues below the essential spectrum. The behavior of the j eigenvalue
EjΩ (α) with a fixed j is given by (1.5) or, if Ω is C 3 , by (1.6).
  The proof is given in Section 7.
   More detailed asymptotic expansions for the eigenvalues can be deduced by using the
toolbox of the semi-classical analysis of Schrödinger operators on manifolds, where the
mean curvature acts as a potential. In Section 8 we describe the results involved by
standard hypotheses on the potential K. In particular, the following hold:
Corollary 1.7. Let Ω ⊂ Rν be a C 5 -admissible domain. If ∂Ω is non-compact, assume
(1.7). Furthermore, assume that K admits a unique global maximum at s0 and that the
Hessian of (−K) at s0 is positive-definite. Denote by µk its eigenvalues and set
                               Xν−1 r                       
                                        µk         
                          E=                2nk − 1 , nk ∈ N .
                                         2
                                   k=1

Then for each j ∈ N there holds, as α → +∞:
                       Ej (QΩ       2
                            α ) = −α − Kmax α + ej α
                                                    1/2
                                                        + O(α1/4 ),

where ej is the j th element of E, counted with multiplicity. Moreover, if ej is of multiplicity
one, the remainder estimate can be improved to O(1).
Corollary 1.8. Let Ω ⊂ R2 be a C 2p+3 -admissible domain with some integer p > 1, and
assume (1.7) if ∂Ω is unbounded. Assume that the curvature of the boundary admits a
unique global maximum at s0 , which is degenerated in the following sense:
                                                               
               K(s) = K(s0 ) − Cp (s − s0 )2p + O (s − s0 )2p+1 , Cp > 0
where s denotes the arc length of the connected component Γ of the boundary where K is
maximal, then for each j ∈ N there holds, as α → +∞:
                                                     1          1  
                  Ej (−∆S − αK) = −Kmax α + ej α p+1 + O α 2(p+1) ,

where ej is the j th eigenvalue of the operator −∂s2 + Cp s2p acting on L2 (R). If ∂Ω is C 2p+4
smooth, then the remainder can be replaced by O(1).
   Finally, in Section 9 we consider the case when Ω is periodic with a compact elementary
cell. In that case, the above main results show that the spectral bands of QΩ  α are deter-
mined, up to a error term, by the spectral bands of the periodic operator −∆S − αK. In
particular, we prove some sufficient conditions guaranteeing the existence of gaps in the
spectrum of QΩ  α.
   The machinery used for the proof of the main results is quite different from all the
previous papers on the Robin eigenvalues and is based on a detailed analysis of the qua-
dratic form and appears to be ideologically very close to the one for the Laplacians in thin
domains, cf. [12, 13, 20, 21]. The reduced operator −∆S − αK appeared already in [21] in
the study of suitable Laplacians in thin neighbordnood of hypersurfaces, and the results
from Section 8 provide improvements of the asymptotics given in [21, Theorem 1.1], under
the respective geometric assumptions.
                                                                                                    5

                                   2. Auxiliary estimates
  We remark first that, as we deal with real-valued operators only, we will work everywhere
with real Hilbert spaces. Let us prove some technical estimates which will be used in the
proof of the main results.
Lemma 2.1. For α > 0 and δ > 0, denote by T D the operator f 7→ −f ′′ acting in L2 (0, δ)
on the domain
                        
               D(T D ) = f ∈ H 2 (0, δ) : f ′ (0) = −αf (0), f (δ) = 0}.
Then, as δα tends to +∞, the operator T D has a unique negative eigenvalue E D , which
satisfies
                              E D = −α2 + O(α2 e−δα ).                            (2.1)
                 D
Furthermore, if ψ is an associated normalized eigenfunction, then
                                   ψ D (0)2 = 2α + O(αe−δα ).
Proof. The assertion was partially proven in Lemma A.2 of [17] by direct computations:
it was shown that the operator T D has a unique negative eigenvalue, that E D = −k2
with k = α + O(αe−δα ), and, finally, that ψ D (t) = C(ek(t−δ) − e−k(t−δ) ), where C is a
normalizing constant. We have then
                                   e2δk − e−2δk                   2ke−2δk
        1 = kψ D k2L2 (0,δ) = C 2                − 2δ , C 2 =                      ,
                                         2k                   1 − 4δke−2δk − e−4δk
which gives
                            (1 − e−2δk )2
        ψ D (0)2 = 2k                        = 2k + O(δk 2 e−2δk ) = 2α + O(αe−δα ).               
                        1 − 4δke−2δk − e−4δk
Lemma 2.2. Let β ≥ 0 be fixed. For α > 0 and δ > 0, denote by T β the operator f 7→ −f ′′
acting in L2 (0, δ) on the domain
                           
                  D(T β ) = f ∈ H 2 (0, δ) : f ′ (0) = −αf (0), f ′ (δ) = βf (δ)}.
Then, as δα tends to +∞, the operator T β has a unique negative eigenvalue E β , which
satisfies
                               E β = −α2 + O(α2 e−δα ).                           (2.2)
                 β
Furthermore, if ψ is an associated normalized eigenfunction, then
                                    ψ β (0)2 = 2α + O(αe−δα ),                                  (2.3)
                                    ψ β (δ)2 = 4αe−2δα + O(αe−3δα ),                            (2.4)
                            k(ψ β )′ k2L2 (0,δ) = α2 + O(α2 e−δα ).                             (2.5)
In addition,
   kf ′ k2L2 (0,δ) − αf (0)2 − βf (δ)2 ≥ 0 for any f ∈ H 1 (0, δ) with hf, ψ β iL2 (0,δ) = 0.   (2.6)
Proof. Once again E β is clearly negative, and we denote by k the positive number such
that E β = −k2 , so that
                                                                
                      β              β  k(t−δ)      β  −k(t−δ)
                     ψ (t) = C 1 +      e      + 1−      e          ,
                                     k                k
with C a normalizing constant. Then the condition ψ ′ (0) = −αψ(0) is equivalent to
                                         β
                               sinh(δk) − cosh(δk)
                            δk           k            = δα.
                                         β
                               cosh(δk) − sinh(δk)
                                         k
                                                                                                  6

Following literally the proof of [17, Lemma A.1] treating the case β = 0, we get the
existence of a unique solution as α gets large, which satisfies k = α + O(αe−δα ), which
gives the asymptotics of E β = −k2 . Moreover, the other eigenvalues of T β are positive,
and since the quadratic form for T β is
               tβ (f, f ) = kf ′ k2L2 (0,δ) − αf (0)2 − βf (δ)2 ,   D(tβ ) = H 1 (0, δ),
the assertion (2.6) follows from the spectral theorem for self-adjoint operators. Then (2.3)
and (2.4) are obtained as in the the proof of Lemma 2.1. Finally, substituting this estimate
into the equality tβ (ψ, ψ) = E β we obtain (2.5).                                        
  Finally, we will need a suitable form of the Sobolev inequality, see e.g. Lemma 8 in [24]:
Lemma 2.3. For any 0 < ℓ ≤ a and f ∈ H 1 (0, a) there holds, with ξ ∈ {0, a},
                                Z a              Z
                                         2     2 a
                           2
                      f (ξ) ≤ ℓ      ′
                                    f (t) dt +       f (t)2 dt.
                                 0             ℓ 0

                3. Reduction to the analysis near the boundary
3.1. Dirichlet-Neumann bracketing. The first steps of the analysis are essentially the
same as in [32]. For δ > 0 denote
                           
                     Ωδ := x ∈ Ω : inf |x − s| < δ}, Θδ := Ω \ Ωδ ,
                                          s∈S

and let qαΩ,N,δ and qαΩ,D,δ be the quadratic forms given by the same expression as qαΩ but
acting on the domains
            D(qαΩ,N,δ ) = H 1 (Ωδ ) ⊕ H 1 (Θδ ),                    e 01 (Ωδ ) ⊕ H01 (Θδ ),
                                                      D(qαΩ,D,δ ) = H
                        e 01 (Ωδ ) := {f ∈ H 1 (Ωδ ) : f = 0 at ∂Ωδ \ S},
                        H

and denote by QΩ,N,δ
                  α     and QΩ,D,δ
                                α   the associated self-adjoint operators in L2 (Ω). The
inclusions D(qαΩ,D,δ ) ⊂ D(qαΩ ) ⊂ D(qαΩ,N,δ ) and the min-max principle imply, for each
j ∈ N, the inequalities
                             Ej (QΩ,N,δ
                                  α     ) ≤ Ej (QΩ          Ω,D,δ
                                                 α ) ≤ Ej (Qα     ).
                                           Ω,D,δ                            Ω,N,δ
Furthermore, QΩ,N,δ
              α     = BαΩ,N,δ ⊕(−∆)N
                                   Θδ and Qα     = BαΩ,D,δ ⊕(−∆)D
                                                                Θδ , where Bα     and
BαΩ,D,δ are the self-adjoint operators in L2 (Ωδ ) associated respectively with the quadratic
forms
                                    Z               Z
                    bΩ,⋆,δ
                     α     (u, u) =   |∇u|2
                                            dx −  α    u2 dS, ⋆ ∈ {N, D},
                                    Ωδ                   S
                            Ω,N,δ
                         D(bα     ) = H 1 (Ωδ ),      D(bΩ,D,δ
                                                         α
                                                                 e 01 (Ωδ ),
                                                               )=H
and (−∆)N             D
           Θδ and (−∆)Θδ denote the Neumann and the Dirichlet Laplacian in Θδ , re-
spectively. As both Neumann and Dirichlet Laplacians are non-negative, we have the
inequalities
            Ej (BαΩ,N,δ ) ≤ Ej (QΩ          Ω,D,δ
                                 α ) ≤ Ej (Bα     ) for all j with Ej (BαΩ,D,δ ) < 0.         (3.1)
The preceding inequalities are valid for any value of δ > 0, but for the rest of the paper
we assume that δ depends on α in a special way:
   the value of δ tends to 0 and the value of δα tends to +∞ as α tends to +∞,                (3.2)
and the precise dependence will be chosen later.
                                                                                                  7


3.2. Change of variables. In order to study the eigenvalues of the operators BαΩ,N,δ and
BαΩ,D,δ we proceed first with a change of variables in Ωδ with small δ. The computations
below are very similar to those performed in [3] for a different problem.
   By assumption, for δ > 0 sufficiently small, the map Φ defined in (1.3) is a diffeomor-
phism between Σ and Ωδ . The metric G on Σ induced by this embedding is
                                        G = g ◦ (Is − tLs )2 + dt2 ,                          (3.3)
where Is : Ts S → Ts S is the identity map, and g is the metric on S induced by the
embedding in Rν . The associated volume form dΣ on Σ is
                    dΣ = | det G|1/2 ds dt = ϕ(s, t)| det g|1/2 ds dt = ϕ dS dt,              (3.4)
where
                                    dS = | det g|1/2 ds
is the induced (ν − 1)-dimensional volume form on S, and the weight ϕ is given by
        ϕ(s, t) := det(Is − tLs ) = 1 − t tr Ls + p(s, t)t2 ≡ 1 − K(s)t + p(s, t)t2 ,         (3.5)
with p being a polynomial in t with coefficients which are bounded and continuous in s.
In particular,
                           |ϕ(s, t) − 1| ≤ k∂t ϕk∞ δ for all (s, t) ∈ Σ.             (3.6)
Let us recall that for a function f : S 7→ R, the boundedness of the gradient ∇s f , as
stated in Definition 1.1, is understood for the norm on the tangent spaces: k∇s f k∞ =
sups∈S k∇s f (s)kTs S , with
                                                              !                    !
                              X            X                       X
                        2                        ρk                    µℓ
            k∇s f (s)kTs S =       gρµ (s)      g (s)∂k f (s)         g (s)∂ℓ f (s)
                                  ρ,µ             k                      ℓ
                                  X
                                         ρµ                             ρµ
                              =         g (s)∂ρ f (s)∂µ f (s) with (g ) = g −1 .
                                  ρ,µ

For future uses, we summarize some obvious properties of ϕ:
Lemma 3.1. Let Ω be a C 2 -admissible domain, the for δ small, the functions Ls , K are
bounded on S, and the functions ∂t ϕ, ∂t2 ϕ, ∂t ϕ−1/2 , (∂t ϕ−1/2 )ϕ1/2 , and ∂t (∂t ϕ−1/2 )ϕ1/2
are bounded on Σ. If, in addition, Ω is C 3 -admissible, then ∇s ∂t (ϕ−1/2 ) is bounded on Σ.
  In particular, for some C > 0 we have
                  X
                      gρµ (s)∂ρ ∂t (ϕ−1/2 )(s, t)∂µ ∂t (ϕ−1/2 ) ≤ C,         (s, t) ∈ Σ.      (3.7)
                    ρ,µ

  Now consider the unitary map
                              U : L2 (Ωδ ) → L2 (Σ, dΣ),       U f = f ◦ Φ,
where Φ is the map from (1.3), and the quadratic forms
            h⋆α (f, f ) = bΩ,⋆,δ
                           α     (U −1 f, U −1 f ),   D(h⋆α ) = U D(bΩ,⋆,δ
                                                                     α     ),   ⋆ ∈ {N, D}.
We have then, using the Einstien summation rule for indices,
                     Z                    Z
           N
         hα (u, u) =    G ∂j u ∂k u dΣ − α |u(s, 0)|2 dS, D(hN
                          jk                                        1
                                                             α ) = H (Σ),
                         Σ                      S
             D                              N
            hα (u, u) = the restriction of hα to D(hD     e1
                                                    α ) = H0 (Σ)
with                           
                     e 1 (Σ) := u ∈ H 1 (Σ) : u(·, δ) = 0 ,
                     H                                              (Gjk ) := G−1 .
                       0
                                                                                                               8

Due to (3.3) we can estimate, with some Cg > 0, depending only on kLs k∞ :

                      (1 − Cg δ)g−1 + dt2 ≤ G−1 ≤ (1 + Cg δ)g−1 + dt2 .

Therefore, we have the form inequalities

                                         N
                                   h−
                                    α ≤ hα             and hD    +
                                                            α ≤ hα                                         (3.8)

with
                                   Z                            Z                       Z
         h−
          α (u, u) := (1 − Cg δ)
                                       ρµ
                                 g ∂ρ u ∂µ u dΣ +                           2
                                                                        |∂t u| dΣ − α       u(s, 0)2 dS,
                               Σ                                    Σ                   S
                                N      1
                    D(h−
                       α ) = D(hα ) = H (Σ),
                                   Z                            Z                       Z
         h+
          α (u, u) := (1 + Cg δ)
                                       ρµ
                                 g ∂ρ u ∂µ u dΣ +                           2
                                                                        |∂t u| dΣ − α       u(s, 0)2 dS,
                               Σ                                    Σ                   S
                    D(h+        D     e1
                       α ) = D(hα ) = H0 (Σ),


where, as usually, (gρµ ) = g −1 . In particular, if Hα− and Hα+ are the self-adjoint operators
acting in L2 (Σ, dΣ) and associated with the forms h−           +
                                                         α and hα respectively, then it follows
from (3.1) and (3.8) that

                 Ej (Hα− ) ≤ Ej (QΩ          +                       +
                                  α ) ≤ Ej (Hα ) for all j with Ej (Hα ) < 0.                              (3.9)



                       4. Proof of Theorem 1.2: upper bound
   Recall that the operator T D has been defined in Lemma 2.1. We have denoted by E D
its lowest eigenvalue, and in this section we denote for shortness ψ := ψ D an associated
normalized eigenfunction. The function ψ will be used to construct test functions for Hα+ .


4.1. An estimate for product functions. Recall that everywhere we assume that δ is
a function of α satisfying (3.2). We have the following estimate:

Lemma 4.1. For v ∈ H 1 (S), consider a function u defined by u(s, t) = v(s)ψ(t), which
                                                   +      +
belongs to D(h+
              α ). There exist positive constants c0 and c1 such that, as α → +∞, for
           1
any v ∈ H (S) there holds

   h+
    α (u, u)
     2         − ED
  kukL2 (Σ,dΣ)
                                             Z
                                (1 + c+
                                      1 δ)           gρµ ∂ρ v ∂µ v ds − αhv, KviL2 (S,dS)
                 ≤ (1 + c+
                         0 δ)
                                                 S
                                                          kvk2L2 (S,dS)
                                                                                    + O(1 + αe−δα ), (4.1)

Moreover, the remainder depends only on kLs k∞ , kKk∞ , k∂t ϕk∞ , and k∂t2 ϕk∞ , and it is
independent of v.
                                                                                                                        9

Proof. Through the estimates we denote by Cj various positive constants. Using (3.6), a
direct evaluation provides
                                    Z
                +
              hα (u, u) =(1 + Cg δ)    ψ(t)2 gρµ ∂ρ v(s) ∂µ v(s) ϕ(s, t)dSdt
                                                  S×(0,δ)
                                  Z               Z δ                                             Z
                                                2                 2                           2
                             +            v(s)          ′
                                                             ψ (t) ϕ(s, t)dt dS − αψ(0)                   v 2 dS
                                      S                  0                                            S
                                                 Z Z δ
                          ≤(1 + C1 δ)                         ψ(t)2 gρµ ∂ρ v(s) ∂µ v(s)dtdS
                                                     S  0                                                           (4.2)
                                  Z                  Z δ                                          Z
                             +   v(s)2   ψ ′ (t)2 ϕ(s, t)dt dS − αψ(0)2                                   v 2 dS
                                      S0                                                              S
                                     Z
                          =(1 + C1 δ) gρµ ∂ρ v ∂µ v dS
                                                     S
                                  Z                  Z δ                                          Z
                                                2                 2                           2
                             +            v(s)                ′
                                                             ψ (t) ϕ(s, t)dt dS − αψ(0)                   v 2 dS.
                                      S                  0                                            S

Moreover, the constant C1 depends only on kLs k∞ and k∂t ϕk∞ . Using a repeated inte-
gration by parts together with the boundary conditions satisfied by ψ, we obtain for all
s ∈ S:
  Z δ
       ψ ′ (t)2 ϕ(s, t)dt
    0
   h               it=δ Z δ                          Z δ
                                        
            ′
 = ψ(t)ψ (t)ϕ(s, t)    +            ′′
                            ψ(t) − ψ (t) ϕ(s, t)dt −     ψ(t)ψ ′ (t)∂t ϕ(s, t)dt
                       t=0        0                                               0
                                      Z δ                               Z δ
                                                1              
 = − ψ(0)ψ ′ (0)ϕ(s, 0) + E D ψ(t)2 ϕ(s, t)dt −       ∂t ψ(t)2 ∂t ϕ(s, t)dt
                            0                   2 0
              Z δ                    h               it=δ Z δ                        (4.3)
       2    D         2            1        2                           2 2
 =αψ(0) + E       ψ(t) ϕ(s, t)dt −     ψ(t) ∂t ϕ(s, t)     −       ψ(t) ∂t ϕ(s, t)dt
               0                   2                   t=0      0
              Z δ                                   Z
       2    D         2            K(s)       2   1 δ
 =αψ(0) + E       ψ(t) ϕ(s, t)dt −       ψ(0) +         ψ(t)2 ∂t2 ϕ(s, t)dt.
               0                     2            2  0

The substitution of (4.3) into (4.2) gives
                         Z
 h+
  α (u, u) ≤ (1 + C 1 δ)   g ρµ ∂ρ v ∂µ v dS
                             S
                                                                         Z Z δ
                                 ψ(0)2                    1
        + E D kuk2L2 (Σ,dΣ) −          hv, KviL2 (S,dS) +                             v(s)2 ψ(t)2 ∂t2 ϕ(s, t)dt dS. (4.4)
                                   2                      2               S   0

As the functions ∂t2 ϕ and K are bounded, we estimate with the help of Lemma 2.1:

  h+           D   2
   α (u, u) − E kukL2 (Σ,dΣ)
                                            Z
                       ≤ (1 + C1 δ)                  gρµ ∂ρ v ∂µ v dS − αhv, KviL2 (S,dS)
                                                 S
                                                                              + O(1 + αe−δα )kvk2L2 (S,dS) , (4.5)

where the O-coefficient depends only on k∂t2 ϕk∞ and kKk∞ . Furthermore, due to (3.6), we
have the estimate kuk2L2 (Σ,dΣ) ≥ (1 − C2 δ)kvk2L2 (S,dS) , where C2 depends only on k∂t ϕk∞ .
                                                                                                           10

This gives
   h+
    α (u, u)
                − ED
  kuk2L2 (Σ,dΣ)
                          Z
             (1 + C1 δ)           g ρµ ∂ρ v ∂µ v dS − αhv, KviL2 (S,dS) + O(1 + αe−δα )kvk2L2 (S,dS)
                              S
         ≤                                                                                             ,
                                                (1 − C2 δ)kvk2L2 (S,dS)
and we deduce the lemma by choosing c+               +
                                       1 = C1 , and c0 > 0 so that (1 − C2 δ)
                                                                             −1 ≤ 1 + c+ δ,
                                                                                       0
which is possible since δ becomes small as α tends to +∞.                                
4.2. Proof of the upper bound. Now for each j we can use the definition (1.4) by
testing on the subspaces L ⊂ D(h+
                                α ) of the form
                                   L = {u : u(s, t) = v(s)ψ(t) with v ∈ Λ},
where Λ are the j th dimensional subspaces of H 1 (S), which is the form domain of −∆S −
αK. Lemma 4.1 then implies
                                                               
         Ej (Hα+ ) − E D ≤ (1 + c+               +
                                 0 δ)Ej − (1 + c1 δ)∆S − αK + O(1 + αe
                                                                           −δα
                                                                               ).   (4.6)
The right-hand side can be estimated as follows:
Lemma 4.2. For any fixed j ∈ N there holds, as α → +∞,
                              +
                                            
         (1 + c+
               0 δ)Ej − (1 + c1 δ)∆S − αK ≤ Ej (−∆S − αK) + O(δα),
where δ is a function of α satisfying (3.2). The constants depend only on kLs k∞ , kKk∞ ,
k∂t ϕk∞ , k∂t2 ϕk∞ .
Proof. We have
                       +
                                    
  (1 + c+
        0 δ)Ej − (1 + c1 δ)∆S − αK
                                                            
                               +
       =(1 + c+
              0 δ) Ej − (1 + c1 δ)∆S + α(Kmax − K) − αKmax
                             +
                                                    
       =(1 + c+
              0 δ)Ej − (1 + c1 δ)∆S + α(Kmax − K) − αKmax + O(δα)
                             +                +
                                                            
       ≤(1 + c+
              0 δ)Ej − (1 + c1 δ)∆S + (1 + c1 δ)α(Kmax − K) − αKmax + O(δα)
                                            
       ≤(1 + Cδ)Ej (−∆S + α(Kmax − K) − αKmax + O(δα)                       (4.7)
                                  
       =Ej − ∆S + α(Kmax − K) − αKmax
                                             
              + O δEj (−∆S + α(Kmax − K) + O(δα)
                                                       
       =Ej (−∆S − αK) + O δEj − ∆S + α(Kmax − K) + O(δα).
                                                                 
As K is bounded, we have the rough estimate Ej − ∆S + α(Kmax − K) = O(α), and the
remainder depends on the constants c+      +
                                      0 , c1 and kKk∞ only.                    
  Finally, combining Lemma 4.2 with (4.6) and Lemma 2.1, we get
  Ej (Hα+ ) ≤ E D + Ej (−∆S − αK) + O(1 + δα + αe−δα )
                                            = −α2 + Ej (−∆S − αK) + O(1 + δα + α2 e−δα ). (4.8)
In order to have an optimal remainder we take
                                       b log α
                                  δ=           , b ≥ 2,
                                          α
then Ej (Hα ) ≤ Ej (Hα+ ) ≤ −α2 + Ej (−∆S − αK) + O(log α).
                                                                                                 11

                                5. Proof of Theorem 1.2: lower bound
5.1. Minoration of the quadratic form. The operator T β of Lemma 2.2 with β = 0
will play a special role and it will be denoted by T N . The first eigenvalue and the first
normalized eigenfunction will be denoted in this section by E N and ψ respectively. Recall
again that δ and α obey (3.2).
   We represent any function u ∈ D(h−  α ) in the form

                                             u(s, t) = v(s)ψ(t) + w(s, t)                     (5.1)
with
                                               Z δ
                                    v(s) :=          ψ(t)u(s, t)dt,   v ∈ H 1 (S).            (5.2)
                                                0
Remark that the both functions (s, t) 7→ v(s)ψ(t) and w are in D(h−  α ). The following
                                                   −           N   2
proposition gives a lower bound on the expression hα (u, u) − E kukL2 (Σ,dΣ) in terms of
this decomposition.
Proposition 5.1. There exist positive constants c−      −
                                                 0 and c1 such that, as α → +∞,

               N      2
  h−
   α (u, u) − E kukL2 (Σ,dΣ)
                    Z
       ≥ (1 − c−
               0 δ)     gρµ ∂ρ v ∂µ v dS − αhv, KviL2 (S,dS) − c−
                                                                1 (1 + αe
                                                                         −δα
                                                                             )kvk2L2 (S,dS)
                            S
                 Z    Z δ                                                                     (5.3)
              α2                     2
         +                  w(s, t) dt dS
              2   S     0

for any u ∈ D(h−                                                        2
               α ). The constants depend only on kLs k∞ , k∂t ϕk∞ and k∂t ϕk∞ .

  The rest of this subsection is devoted to the proof of Proposition 5.1. Using the decom-
position (5.1), we clearly have
                              Z                   Z               Z
        −
       hα (u, u) = (1 − Cg δ)    ρµ
                                g ∂ρ u ∂µ u dΣ +    |∂t u| dΣ − α u(s, 0)2 dS
                                                          2
                                                                                      (5.4)
                                         Σ                        Σ                  S
                     =: I1 + I2 + I3 + I4 ,                                                   (5.5)
where we have set
                        Z δZ
      
       I1 = (1 − Cg δ)           gρµ ∂ρ u ∂µ u ϕ dSdt,
      
      
      
                         0     S
      
            Z         Z δ
      
      
      
                v(s)2
                           ψ ′ (t)2 ϕ(s, t)dt dS,
       I2 =
               S        0
              Z         Z δ                                      Z             
      
      
      
       I3 = 2      v(s)        ′
                             ψ (s)∂t w(s, t)ϕ(s, t)dt dS − αψ(0) v(s)w(s, 0)dS
      
      
      
                 S       0                                         S
      
            Z Z δ                                        Z            Z
      
                                 2
      I =4           ∂ w(s, t) ϕ(s, t)dtdS − αψ(0)
                                t
                                                        2
                                                            v(s) dS − α w(s, 0)2 dS.
                                                                2
                  S    0                                               S                 S

We estimate the four terms separately.
Lemma 5.2. There exists C1 > 0 such that, as α → +∞,
                                        Z
                         I1 ≥ (1 − C1 δ) g ρµ ∂ρ v ∂µ v dS.                                   (5.6)
                                                             S

Moreover, the constant C1 depends only on k∂t ϕk∞ and kLs k∞ and is independent of u.
                                                                                                                                                     12

Proof. Following the decomposition (5.1), we get by using (3.6) a constant C > 0 such
that
Z δZ
              gρµ ∂ρ u(s, t) ∂µ u(s, t) ϕ(s, t)dSdt
 0        S
              Z δZ                                                                               Z δZ
       =                       ψ(t) g  2 ρµ
                                                   ∂ρ v(s) ∂µ v(s) ϕ(s, t)dSdt + 2                               ψ(t)g ρµ ∂ρ v(s) ∂µ w(s, t) ϕdSdt
               0           S                                                                         0       S
                   Z δZ
              +                    gρµ ∂ρ w(s, t) ∂µ w(s, t) ϕdSdt
                       0       S
                                   Z δZ                                                              Z δZ
       ≥(1 − Cδ)                                   ψ(t) g   2 ρµ
                                                                   ∂ρ v(s) ∂µ v(s)dSdt + 2                           ψ(t)gρµ ∂ρ v(s) ∂µ w(s, t) ϕ(s, t)dSdt
                                       0     S                                                           0       S
                                           Z δZ
              + (1 − Cδ)                                   gρµ ∂ρ w(s, t) ∂µ w(s, t)dSdt
                                               0       S
                                   Z                                           Z δZ
       =(1 − Cδ)                           gρµ ∂ρ v(s) ∂µ v(s)dS + 2                         ψ(t)gρµ ∂ρ v(s) ∂µ w(s, t) ϕ(s, t)dSdt
                                       S                                        0        S
                                           Z δZ
              + (1 − Cδ)                                   gρµ ∂ρ w(s, t) ∂µ w(s, t) dSdt,
                                               0       S
                                                                                                                                                (5.7)

where the constant C depends only on k∂t ϕk∞ . Remark that for the function w we have
                           Z δ                                                       Z δ
                                   ψ(t)w(s, t)dt = 0 and, hence,                             ψ(t)∂ρ w(s, t)dt = 0,                s ∈ S.        (5.8)
                               0                                                         0

We deduce:

     Z δZ
                   ψg ρµ ∂ρ v ∂µ w ϕdSdt
      0        S
                                   Z δZ                                             Z δZ
                                                            ρµ
                               =                       ψg        ∂ρ v ∂µ w dSdt +                ψgρµ ∂ρ v ∂µ w (ϕ − 1)dSdt
                                           0       S                                 0       S
                                                                                              Z δZ
                                                                                         =                   ψg ρµ ∂ρ v ∂µ w (ϕ − 1)dSdt. (5.9)
                                                                                                 0       S


Using again (3.6), we estimate with the same constant C, using the Cauchy-Schwarz
inequality for the metric (gρµ ),

      Z δZ
                    ψ(t)gρµ ∂ρ v(s) ∂µ w(s, t) (ϕ(s, t) − 1)dSdt
       0        S
                Z δZ
     ≤ Cδ                          gρµ ψ(t)∂ρ v(s) ∂µ w(s, t) dSdt
                   0           S
                   Z δZ                           Z Z
       Cδ                                      Cδ δ
                                               2 ρµ
     ≤           ψ(t) g ∂ρ v(s) ∂µ v(s)dSdt +              gρµ ∂ρ w(s, t) ∂µ w(s, t)dSdt (5.10)
       2   0   S                                2 0 S
          Z                      Z Z
       Cδ     ρµ              Cδ δ
     =       g ∂ρ v ∂µ v dS +           gρµ ∂ρ w(s, t) ∂µ w(s, t)dSdt
       2 S                    2 0 S
                                                                                                                 13

which gives
  Z δ Z                                         
                 ρµ
 2          ψ(t)g ∂ρ v(s) ∂µ w(s, t) ϕ(s, t)dSdt
        0       S
                           Z                                         Z δZ
                    ≤ Cδ           g   ρµ
                                            ∂ρ v(s) ∂µ v(s)dS + Cδ            gρµ ∂ρ w(s, t) ∂µ w(s, t)dSdt. (5.11)
                               S                                      0   S
Substituting the last inequality into (5.7) we obtain,
  Z δZ                                   Z                 Z δZ
          ρµ                                 ρµ
         g ∂ρ u ∂µ u ϕdSdt ≥ (1−2Cδ) g ∂ρ v ∂µ vdS+(1−2Cδ)      g ρµ ∂ρ w ∂µ w dSdt
    0       S                                              S                                  0   S
and, therefore, for sufficiently small δ:
                  Z δZ                                           Z
       (1 − Cg δ)         g ∂ρ u ∂µ u ϕdSdt ≥ (1 − Cg δ)(1 − 2Cδ) gρµ ∂ρ v ∂µ vdS.
                           ρµ
                         0         S                                                      S
The result follows as Cg depends only on kLs k∞ , and C depends only on k∂t ϕk∞ .                                
Lemma 5.3. There exists C2 > 0 such that, as α → +∞:
             Z
           2
 I2 ≥ αψ(0)    v(s)2 dS + E N ku − wk2L2 (Σ,dΣ)
                     S
                                                    − αhv, KviL2 (S,dS) − C2 (1 + αe−δα )kvk2L2 (S,dS) . (5.12)
The constant C2 depends only on kKk∞ , k∂t ϕk∞ and k∂t2 ϕk∞ and is independent of u.
Proof. As in (4.3), an integration by part leads to
  Z δ                                         Z δ
           2                        2       N                        K(s)
        ′
      ψ (t) ϕ(s, t)dt = αψ(0) + E                 ψ(t)2 ϕ(s, t)dt −       ψ(0)2
   0                                           0                       2
                                                                                 Z
                                                          ∂t ϕ(s, δ)           1 δ
                                                       −             ψ(δ)2 +          ψ(t)2 ∂t2 ϕ(s, t)dt.
                                                               2               2 0
The additional term in comparison with (4.3) comes from the fact that ψ(δ) 6= 0. We
deduce:
             Z         Z δ
       I2 =      v(s)2       ψ ′ (t)2 ϕ(s, t)dt dS
               S         0
                     Z
                                                                ψ(0)2
          = αψ(0)2        v 2 dS + E N ku − wk2L2 (Σ,dΣ) −            hv, KviL2 (S,dS)                (5.13)
                        S                                         2
                                                       Z         Z δ
                ψ(δ)2                                1         2
             −         hv, ∂t ϕ(·, δ)viL2 (S,dS) +        v(s)       ψ(t)2 ∂t2 ϕ(s, t)dtdS.
                  2                                  2 S          0
Due to (2.4), ψ(δ)2 = O(αe−2δα ), and there exists C > 0 such that for α large enough one
has
                    ψ(δ)2
                          hv, ∂t ϕ(·, δ)viL2 (S,dS) ≤ Cαe−2δα kvk2L2 (S,dS) ,       (5.14)
                      2
where the constant C depends only on k∂t ϕk∞ . We also have:
                              Z δ
                                   ψ(t)2 ∂t2 ϕ(s, t)dt ≤ k∂t2 ϕk∞ .                 (5.15)
                               0
                          ′
Moreover, (2.3) provides C > 0, depending only on kKk∞ , such that for α large enough:
             ψ(0)2
                   hv, KviL2 (S,dS) ≤ αhv, KviL2 (S,dS) + C ′ αe−δα kvk2L2 (S,dS) . (5.16)
               2
The lemma follows by combining (5.14)–(5.16) with (5.13).                                                        
                                                                                                                          14

  The crossed term I3 needs a parametric estimate:
Lemma 5.4. There exists C3 > 0 such that for any r > 0 for large α one has
                                                            Z Z δ
                                                          1
            N                          2    2
    I3 ≥ 2E hu − w, wiL2 (Σ,dΣ) − C3 rα kvkL2 (S,dS) − C3         w(s, t)2 dtdS.
                                                          r S 0
Moreover, the constant C3 depends only on k∂t ϕk∞ and does not depend on u.
Proof. Using the integration by parts we have:
         Z      Z δ                                           Z
   I3 = 2 v(s)      ψ ′ (s)∂t w(s, t)ϕ(s, t)dt dS − 2αψ(0) v(s)w(s, 0)dS
           S      0                                             S
         Z      h                     it=δ Z δ
      = 2 v(s) ψ ′ (t)w(s, t)ϕ(s, t)         −     ψ ′′ (t)w(s, t)ϕ(s, t)dt
             S                                          t=0      0
             Z δ                                                          Z
         −           ψ ′ (t)w(s, t)∂t ϕ(s, t)dt dS − 2αψ(0)                     v(s)w(s, 0)dS
              0                                                             S
                         Z                                           Z    Z δ                                        (5.17)
                                                                 N
      = 2αψ(0)               v(s)w(s, 0)ϕ(s, 0)dS + 2E                          v(s)ψ(t)w(s, t)ϕ(s, t)dtdS
                        S                                             S     0
              Z       Z δ                                                          Z
                                    ′
         −2                  v(s)ψ (t)w(s, t)∂t ϕ(s, t)dtdS − 2αψ(0)                       v(s)w(s, t)dS
                  S      0                                                             S
                                                  Z Z δ
             N
      = 2E hu − w, wiL2 (Σ,dΣ) − 2                            v(s)ψ ′ (t)w(s, t)∂t ϕ(s, t)dtdS,
                                                    S    0
where we have used the boundary conditions ψ ′ (δ) = 0 and ψ ′ (0) = −αψ(0).
  We estimate now, with any r > 0:
    Z Z δ
  2       v(s)ψ ′ (t)w(s, t)∂t ϕ(s, t)dtdS
     S   0
                                        Z Z δ
                          ≤ k∂t ϕk∞              2 v(s)ψ ′ (t)w(s, t) dtdS
                                         S   0
                                        Z Z δ                            Z Z δ
                                                         2       k∂t ϕk∞
                                                                 2
                          ≤ k∂t ϕk∞ r              ′
                                            ψ (t) v(s) dtdS +                    w(s, t)2 dtdS (5.18)
                                      S 0                            r    S 0
                                                                 Z Z δ
                                                         k∂t ϕk∞
                          ≤ 2k∂t ϕk∞ rα2 kvk2L2 (S,dS) +               w(s, t)2 dtdS,
                                                             r    S 0
where we have used kψ ′ k2L2 (0,δ) ≤ 2α2 for α large enough, see (2.5). The substitution of
(5.18) into (5.17) gives the lemma by choosing C3 = 2k∂t ϕk∞ .                           
  We are now able to finish the proof of Proposition 5.1. We use Lemmas 5.2–5.4 in (5.5)
and deduce
                        Z
h−
 α (u, u) ≥ (1 − C 1 δ)   gρµ ∂ρ v ∂µ v dS − αhv, KviL2 (S,dS)
                                S
                                                                                    Z Z δ
                                                                                                              2
         + E N ku − wk2L2 (Σ,dΣ) + 2E N hu − w, wiL2 (Σ,dΣ) +                                    ∂t w(s, t) ϕ(s, t)dtdS
                                                                                       S     0
                 Z                                                                                   Z Z δ
                                    2                     2                                 C3                       2
         −α              |w(s, 0)| dS − C4 (1 + rα + αe              −δα
                                                                           )kvk2L2 (S,dS) −                   w(s, t) dt dS,
                     S                                                                           r    S   0
                                                                                                                     (5.19)
where C4 = max(C2 , C3 ). We have the equality
         ku − wk2L2 (Σ,dΣ) + 2hu − w, wiL2 (Σ,dΣ) = kuk2L2 (Σ,dΣ) − kwk2L2 (Σ,dΣ) .                                  (5.20)
                                                                                                       15

Due to (5.8), we can use (2.6), so that
                     Z Z δ                       Z
                                       2
                             ∂t w(s, t) dt dS − α w(s, 0)2 dS ≥ 0,
                        S   0                              S

and, therefore, substituting (5.20) into (5.19), we get
                                          Z
  hα (u, u) − E kukL2 (Σ,dΣ) ≥ (1 − C1 δ) gρµ ∂ρ v ∂µ v dS − αhv, KviL2 (S,dS)
   −           N     2
                                                S
                                                                          Z Z δ
                                                                     C3
   − C4 (1 + rα2 + αe−δα )kvk2L2 (S,dS) − E N kwk2L2 (Σ,dΣ) −                      w(s, t)2 dt dS. (5.21)
                                                                     r     S   0

Due to (3.6), we have
                                                          Z Z δ
                     kwk2L2 (Σ,dΣ) ≥ (1 − k∂t ϕk∞ δ)               w(s, t)2 dt dS.
                                                           S   0

We choose r = 3C3 /α2 in (5.21), so that the asymptotic expansion (2.2) for E N provides
a constant C5 > 0 such that for α large enough:
                                            Z
                                                                 ψ(0)2
  h−           N   2
   α (u, u) − E kukL2 (Σ,dΣ) ≥ (1 − C1 δ)           gρµ ∂ρ v ∂µ v dS −  hv, KviL2 (S,dS)
                                            S                      2
                                                                         Z Z δ
                                                        2            α2
                                    − C5 (1 + αe−δα
                                                    )kvkL2 (S,dS) +            w(s, t)2 dt dS.
                                                                     2 S 0
Therefore, the proof is concluded by setting c−            −
                                              0 = C1 and c1 = C5 . Noticing that C4 and
C5 express with C2 and C3 , we deduce that the constants depends only on kLs k∞ , k∂t ϕk∞
and k∂t2 ϕk∞ .

5.2. Asymptotics of Ej . The expression on the right-hand side of (5.3) can be viewed as
a lower semibounded quadratic form defined on D(h−         2
                                                   α ) ⊂ L (Σ, dSdt). Denote its closure
in L (Σ, dSdt) by q, and let Q be the associated self-adjoint operator in L2 (Σ, dSdt) ≡
     2

L2 (S, dS) ⊗ L2 (0, δ). It writes as
                h                                       i    α2
             Q = − (1 − c−
                         0 δ)∆ S − αK − c−
                                         1 (1 + αe−2δα
                                                       )  P+    (1 − P ),
                                                             2
where P : L2 (Σ, dSdt) → L2 (S, dS) ⊗ ψ is the orthogonal projector (P u)(s, t) := v(s)ψ(t)
with v defined in (5.2). For each fixed j and large α we have
                                                                  α2
               Ej − (1 − c−
                          0 δ)∆ S − αK − c−
                                          1 (1 + αe−2δα
                                                        )  = O(α) <    ,
                                                                    2
hence,
                                                                  
                  Ej (Q) = Ej − (1 − c−              −
                                      0 δ)∆S − αK − c1 (1 + αe
                                                              −δα
                                                                  ) .

Furthermore, using
                         kuk2L2 (Σ,dΣ) ≤ (1 + k∂t ϕk∞ δ)kuk2L2 (Σ,dSdt)
we have a positive constant Cϕ , depending only on k∂t ϕk∞ such that
                         h−
                          α (u, u)                         q(u, u)
                           2         − E N ≥ (1 − Cϕ δ)                 .
                        kukL2 (Σ,dΣ)                    kuk2L2 (Σ,dSdt)
                                                                                                      16

As the identification operator f 7→ f defines an injection of D(h−      2
                                                                 α ) ⊂ L (Σ, dΣ) in D(q) ⊂
 2
L (Σ, dSdt), it follows that
  Ej (Hα− ) ≥ (1 − Cϕ δ)Ej (Q) + E N
                                                         
                                                                    2 −δα
                  = −α2 + (1 − Cϕ δ)Ej − (1 − c−
                                               0 δ)∆S − αK + O(1 + α e    ), (5.22)

where we have used the asymptotics (2.2) for E N . In addition, by Lemma 4.2, we get
                                             
           (1 − Cϕ δ)Ej − (1 − c−
                                0 δ)∆ S − αK    = Ej (−∆S − αK) + O(δα).
Hence, by substituting in (5.22),
                 Ej (Hα− ) ≥ −α2 + Ej (−∆S − αK) + O(1 + α2 e−δα + δα),
and the constants depend only on kLs k∞ , kKk∞ , k∂t ϕk∞ and k∂t2 ϕk∞ . Choosing
                                      b log α
                                 δ=           , b ≥ 2,                           (5.23)
                                         α
we arrive at the result.

                                  6. Proof of Theorem 1.3
   The main idea for improving the remainder estimate is to work in unweighted spaces
from the very beginning. The weight ϕ is indeed C 1 with respect to the s variable now,
and this allows for more precise Taylor expansions of ϕ in Σ, so that the comparison
between the Robin Laplacian and the decoupled operator becomes more precise. We start
with the following simple result:
Lemma 6.1. Under the assumption of Theorem 1.3, for any fixed j ∈ N one has
                      Ej (−∆S − αK) = −αKmax + O(α2/3 ) as α → +∞.
Proof. Due to (−∆S ) ≥ 0 we have the obvious lower bound Ej (−∆S − αK) ≥ −αKmax .
Let us prove the upper bound. For s, s0 ∈ S, let d(s, s0 ) denote the geodesic distance
between s and s0 . Let s0 ∈ S be such that K(s0 ) = Kmax . As K is at least C 1 , there
exist ε > 0 and C > 0 such that
                          K(s) ≥ Kmax − Cd(s, s0 ) as d(s, s0 ) < ε.                           (6.1)
Now let us choose j functions f1 , . . . , fj ∈ Cc∞ (R+ ) having disjoint supports, non identi-
cally zero, and set vi (s) = fi r −1 d(s, s0 ) , where r > 0 is small and will be chosen later.
For small r, the functions vi have pairwise disjoint supports and belong to the domain of
∆S . In particular, they are linearly independent, and
                                vi , (−∆S − αK)vl = 0 for i 6= l.
On the other hand,
                 Z             Z
                                                  2
       θi (r) :=         2
                   vi (s) dS =   fi r −1 d(s, s0 ) dS = ai r ν−1 + o(r ν−1 ),       ai > 0.
                  S                   S
Using (6.1) we have
                           Z                           Z
   vi , (−∆S − αK)vi =           ρµ
                                g ∂ρ vi ∂µ vi dS − α       Kvi2 dS
                            S                          S
                                            ≤ bi r ν−3 − αKmax θi (r) + ci αrθi (r),   bi , ci > 0,
which gives
                 vi , (−∆S − αK)vi
                                   ≤ −αKmax + Ai (r −2 + αr),             Ai > 0.
                        kvi k2
                                                                                                  17

Now it is sufficient to take r := α−1/3 and to test in (1.4) on the subspace L spanned by
v1 , . . . , v j .                                                                      
6.1. Toward unweighted spaces. In order to remove the weight ϕ, we perform the
unitary transform
                       Θ : L2 (Σ, dSdt) ∋ u 7→ ϕ−1/2 u ∈ L2 (Σ, dΣ),
                                                                          
                                                                 −1 D(h± ) ⊂ L2 (Σ, dSdt).
and consider the quadratic forms u 7→ h±α (Θu, Θu) defined on Θ        α
In order to reduce the analysis to decoupled operators, we prove approximation lemmas:
Lemma 6.2. There exists δ0 > 0 and positive constants C and C ′ such that for all
δ ∈ (0, δ0 ) and u ∈ Θ−1 D(h±  α ) one has
   Z                                       Z
         ρµ
       g ∂ρ (ϕ  −1/2
                     u)∂µ (ϕ−1/2
                                 u)ϕdSdt −     gρµ ∂ρ u∂µ udSdt
     Σ                                       Σ
                                               Z
                                         ≤ Cδ      gρµ ∂ρ u∂µ udSdt + C ′ δkuk2L2 (Σ,dSdt) , (6.2)
                                                  Σ
and the constants depend only on kϕk∞ and k∇s ∂t (ϕ−1/2 )k∞ .
Proof. We compute
  Z                                       Z
     gρµ ∂ρ (ϕ−1/2 u)∂µ (ϕ−1/2 u)ϕdSdt −      gρµ ∂ρ u∂µ udSdt
   Σ                                        Σ
                    Z                                     Z
                         ρµ
                 =     g ∂ρ ϕ −1/2
                                   ∂µ ϕ−1/2
                                            ϕu dSdt + 2 gρµ ∂ρ ϕ−1/2 uϕ1/2 ∂µ udSdt. (6.3)
                                              2
                     Σ
Due to (3.5), we have the expansion
                                   ϕ(s, t)−1/2 = 1 + tA(s, t),
where A and its gradient are bounded. In particular, there exists C0 > 0 with
                                       k∇s ϕ−1/2 k∞ ≤ C0 δ,                                 (6.4)
where C0 is controlled by k∂t ∇s ϕ−1/2 k∞ , see (3.7). We deduce that
                  Z
                      gρµ ∂ρ ϕ−1/2 ∂µ ϕ−1/2 ϕu2 dSdt ≤ C1 δ2 kuk2L2 (Σ,dSdt) ,              (6.5)
                     Σ

where the constant C1 is controlled by k∇s ∂t ϕ−1/2 k∞ and kϕk∞ . Using the Cauchy-
Schwarz inequality for the metric (gρµ ), we get
     Z                        
   2   gρµ ∂ρ ϕ−1/2 u ϕ1/2 ∂µ u dSdt
      Σ
                                  Z                                   Z                  1/2
                                         ρµ    −1/2      −1/2 2            ρµ
                             ≤2          g ∂ρ ϕ       ∂µ ϕ    u dSdt    g ∂ρ u∂µ uϕdSdt
                                     Σ                                 Σ
                                   Z                                      Z
                             ≤ δ−1     gρµ ∂ρ ϕ−1/2 ∂µ ϕ−1/2 u2 dSdt + δ      gρµ ∂ρ u∂µ uϕdSdt
                                     Σ                                      Σ
                                                  Z               Z
                                −1       −1/2 2        2               ρµ
                             ≤ δ k∇s ϕ        k∞     u dSdt + δ       g ∂ρ u∂µ uϕdSdt
                                                   Σ                Σ
                                                          Z
                             ≤ C1 δkuk2L2 (Σ,dSdt) + C2 δ    gρµ ∂ρ u∂µ udSdt;
                                                             Σ
on the last step we have used (6.4). We deduce the lemma by combining the last inequality
with (6.3) and (6.5). Since C2 is controlled by kϕk∞ , we deduce that the constants depends
only on kϕk∞ and k∇s ∂t (ϕ−1/2 )k∞ .                                                     
                                                                                                                18

Lemma 6.3. There exists δ0 > 0 and positive constants C and β such that for all δ ∈
(0, δ0 ) there holds
  Z                                           Z                       Z
                                                                         K(s)
          |∂t (ϕ  −1/2        2
                         u)| ϕdSdt ≥                      2
                                                      |∂t u| dSdt −           u(s, 0)2 dS
                                                                       S  2
      Σ
                                            ZΣ
                                                                                                  
                                         −β   u(s, δ)2 dS − Ckuk2L2 (Σ,dSdt) for all u ∈ Θ−1 D(h−
                                                                                                α)  (6.6)
                                               S

and
  Z                                           Z                       Z
                                  2                       2          K(s)
              ∂t (ϕ−1/2
                          u) ϕdSdt ≤                  |∂t u| dSdt −         u(s, 0)2 dS
      Σ                                           Σ                S   2
                                                                                                        
                                                               + Ckuk2L2 (Σ,dSdt) for all u ∈ Θ−1 D(h+
                                                                                                     α ) , (6.7)

and the constants depend on k∂t (ϕ−1/2 )k∞ , kϕ1/2 k∞ and k∂t (ϕ1/2 ∂t ϕ−1/2 )k∞ only.

Proof. We have
  Z                         Z
     |∂t (ϕ−1/2    2
                u)| ϕdSdt −    |∂t u|2 dSdt
   Σ                         Σ
                            Z                              Z
                          =    |∂t (ϕ−1/2 )|2 u2 ϕdSdt + 2   ∂t (ϕ−1/2 )ϕ1/2 u∂t udSdt, (6.8)
                                               Σ                                    Σ

and there exists C0 > 0 such that
                       Z
                                      2
                           ∂t (ϕ−1/2 ) u2 ϕdSdt ≤ C0 kuk2L2 (Σ,dSdt) .                                      (6.9)
                                          Σ

The second term is treated by an integration by parts:
      Z                                                Z Z δ
  2           ∂t (ϕ   −1/2    1/2
                             )ϕ       u∂t udSdt =         ∂t (ϕ−1/2 )ϕ1/2 ∂t (u2 )dtdS
          Σ                                         S 0
                                         Z h                   it=δ      Z
                                      =      ∂t (ϕ−1/2   1/2 2
                                                       )ϕ u         dS −      ∂t (∂t (ϕ−1/2 )ϕ1/2 )u2 dSdt. (6.10)
                                          S                           t=0       Σ

Due to (3.5), we have the expansion
                                                                      K(s)
                                              ∂t (ϕ−1/2 )(s, t) =          + tQ(s, t),
                                                                       2
where Q is bounded in Σ, so that ∂t (ϕ−1/2 )(s, 0) = K(s)/2, and (6.10) provides
   Z                            Z                                          Z
                   1/2             K(s)
 2     ∂t (ϕ−1/2
                 )ϕ u∂t udSdt +                 2             2
                                        u(s, 0) dS ≤ C1 kukL2 (Σ,dsdt) + β    u(s, δ)2 dS,
     Σ                           S   2                                      S

where
                                                   β = sup (ϕ1/2 ∂t ϕ−1/2 )(s, δ) .
                                                         s∈S

By combining this with (6.8) and (6.9), we deduce the lower          bound (6.6), and also the
upper bound (6.7) since u(s, δ) = 0 for u ∈ Θ−1 D(h+         α )  .  Moreover, the constant C0
is controlled by k∂t ϕ   −1/2  2
                              k∞ , the constant β by k∂t ϕ−1/2   1/2
                                                               ϕ k∞ and the constant C1 by
k∂t (∂t ϕ−1/2 ϕ1/2 )k∞ .                                                                     
                                                                                                19

   We deduce by combining the last two lemmas that there exist positive constants c0 and
c1 such that for all δ ∈ (0, δ0 ):
                               Z                        Z
      −
     hα (Θu, Θu) ≥ (1 − c0 δ)        ρµ
                                    g ∂ρ u∂µ udSdt +        |∂t u|2 dSdt
                                 Σ                        Σ
                       Z                              Z
                                   K                                                        (6.11)
                    −       α+          u(s, 0)2 dS − β     u(s, δ)2 dS − c1 kuk2L2 (Σ,dSdt
                         S         2                     S
                                            
                     for u ∈ Θ−1 (D(h−   α )  ,
and
                            Z                       Z
  h+
   α (Θu, Θu) ≤ (1 +  c0 δ)     g ρµ
                                     ∂ρ u∂µ udSdt +     |∂t u|2 dSdt
                              Σ                       Σ
                   Z                                                                    (6.12)
                               K            2            2                    −1    +
                                                                                       
                −       α+           u(s, 0) dS + c1 kukL2 (Σ,dSdt) for u ∈ Θ    D(hα ) .
                    S          2
We denote by qα± the quadratic forms on the right-hand side of (6.11) and (6.12) respec-
tively, defined on the form domains D(qα± ) := Θ−1 (D(h±α )). The associated self-adjoint
operators, both acting on the unweighted space L2 (Σ, dSdt), will be denoted by Q±α . Due
to (3.9) one has
                                  Ω          +                       +
                  Ej (Q−
                       α ) ≤ Ej (Qα ) ≤ Ej (Qα ) for all j with Ej (Qα ) < 0.              (6.13)
6.2. Upper bound. Once again we estimate the quadratic form qα+ evaluated on the
functions u that write as a product u(s, t) = v(s)ψ(t), where ψ is a normalized eigenfunc-
tion of T D associated with E D (see Lemma 2.1) and v ∈ H 1 (S). Here we have simply
kukL2 (Σ,dSdt) = kvk2L2 (S,dS) , and
                                         Z                      Z
                                                          ψ(0)2
 +           D    2                          ρµ
qα (u, u) − E kukL2 (Σ,dSdt) = (1 + c0 δ) g ∂ρ v∂µ vdS −            Kv 2 dS + c1 kvk2L2 (S,dS) .
                                           S                2     S
Using (2.1) we obtain
                                                     
               Ej (Q+       2
                    α ) ≤ −α + Ej − (1 + c0 δ)∆S − αK + O(1 + αe
                                                                −δα
                                                                    ).                     (6.14)
To estimate the right-hand side of (6.14) we need an additional assertion:
Lemma 6.4. For any j ∈ N there exist positive constants C, α0 and δ0 such that for
δ ∈ (0, δ0 ) and α ≥ α0 the following inequalities hold:
                                                        
                  Ej − (1 + c0 δ)∆S − αK ≤ Ej − ∆S − αK + Cδα2/3 ,          (6.15)
                                                        
                  Ej − (1 − c0 δ)∆S − αK ≥ Ej − ∆S − αK − Cδα2/3 .          (6.16)
Proof. We only prove the upper bound, the lower bound being symmetric. We have
                                                    1                
   Ej − (1 + c0 δ)∆S − αK = (1 + c0 δ)Ej − ∆S +            α(Kmax − K) − αKmax
                                                  1 + c0 δ
                                                               
                            ≤ (1 + c0 δ)Ej − ∆S + α(Kmax − K) − αKmax
                                                                          
                            = Ej − ∆S − αK + c0 δEj − ∆S + α(Kmax − K) ,
and it is sufficient to apply Lemma 6.1 to the last term.                                       
   Now let us assume that δ is a function of α satisfying (3.2). Applying Lemma 6.4 to
(6.14) we deduce, for α → +∞,
                                                        2/3
              Ej (Q+       2
                   α ) ≤ −α + Ej − ∆S − αK + O(δα            + 1 + α2 e−δα ).   (6.17)
Choosing δ = α−κ with κ ∈ [2/3, 1) and using (6.13) we obtain the result.
                                                                                                                20

6.3. Lower bound. Similarly to Section 5.1, we decompose any function u ∈ D(qα− ) as

                                                 u(s, t) = v(s)ψ(t) + w(s, t),

where ψ = ψ β is a normalized eigenfunction of the operator T β associated with the first
eigenvalue E β , see Lemma 2.2, and
                                                              Z δ
                                                     v(s) =               ψ(t)u(s, t)dt.
                                                                  0

It follows that
                                                Z δ
                                                      ψ(t)w(s, t)dt = 0,             s ∈ S,                 (6.18)
                                                 0

which provides
       Z                              Z
          v(s)ψ(t)w(s, t)dSdt = 0 and   gρµ ∂ρ v(s)ψ(t)∂µ w(s, t)dSdt = 0.                                  (6.19)
           Σ                                                      Σ

A direct computation provides
                            Z                             Z
   qα− (u, u) = (1 − c0 δ) gρµ ∂ρ v∂µ vdS + (1 − c0 δ)        gρµ ∂ρ w(s, t)∂µ w(s, t)dSdt
                              S                             Σ
                   Z                         Z
                            2 ′    2                   ′
                +      v(s) ψ (t) dSdt + 2 v(s)ψ (t)∂t w(s, t)dSdt
                                               Σ
                   ZΣ                       Z 
                                                     K(s) 
                +      |∂t w(s, t)|2 dSdt −      α+          v(s)2 ψ(0)2 dS
                     Σ                       S          2                                                   (6.20)
                     Z                                       Z 
                                 K(s)                                  K(s) 
                −2         α+          v(s)ψ(0)w(s, 0)dS −         α+           w(s, 0)2 dS
                                  2                            S           2
                     ZS                        Z                            Z
                −β             2     2
                         v(s) ψ(δ) dS − 2β        v(s)ψ(δ)w(s, δ)dS − β        w(s, δ)2 dS
                                S                                 S                                 S
                       − c1 kuk2L2 (Σ,dSdt) ,

where we have used (6.19) to get rid of the crossed terms. We also have, using an integra-
tion by part:
  Z                                              Z           h               it=δ Z δ               
               ′                                                      ′                 ′′
        v(s)ψ (t)∂t w(s, t)dSdt =                     v(s)        ψ (t)w(s, t)    −    ψ (t)w(s, t)dt dS
    Σ                                             S                                 t=0         0
                       Z                                                                 Z δ           
                                        ′                     ′                       β
                   =           v(s) ψ (δ)w(s, δ) − ψ (0)w(s, 0) + E     ψ(t)w(s, t)dt dS
                           S                                          0
                                                 Z                                     
                                              =      βv(s)ψ(δ)w(s, δ) + αv(s)ψ(0)w(s, 0) dS, (6.21)
                                                         S

where we have used (6.19) and the boundary condition for the eigenfunction ψ. Moreover,
by the definition of ψ we have
                                        Z δ
                                                ψ ′ (t)2 dt − αψ(0)2 − βψ(δ)2 = E β .
                                            0
                                                                                              21

Inserting the last inequality and (6.21) into (6.20), we arrive at
                                 Z                              Z
        qα− (u, u) = (1 − c0 δ) gρµ ∂ρ v∂µ vdS + (1 − c0 δ)         gρµ ∂ρ w∂µ wdSdt
                                   S                              Σ
                        Z                  Z
                     +      |∂t w|2 dSdt −      K(s)v(s)ψ(0)w(s, 0)dS
                          Σ                  S
                        Z                            Z                                 (6.22)
                                   K                    K
                     −       α+               2
                                      w(s, 0) dS −          v(s)2 ψ(0)2 dS
                          S         2                  S 2
                          Z
                     −β       w(s, δ)2 dS + E β kvk2L2 (S,dS) − c1 kuk2L2 (Σ,dSdt) .
                           S

Lemma 6.5. There exist R > 0 and α0 > 0 such that for all α ≥ α0 there holds
  Z                        Z
     |∂t w(s, t)|2 dSdt −      K(s)v(s)ψ(0)w(s, 0)dS
   Σ                         S
            Z                              Z                         Z
                      K(s)                     K(s)
          −       α+                  2
                               w(s, 0) dS −              2    2
                                                     v(s) ψ(0) dS − β    w(s, δ)2 dS
              S          2                    S  2                     S
                                                             Z 
                                     α2                          K(s) R 
                                ≥ − kwk2L2 (Σ,dSdt) − ψ(0)2            +    v(s)2 dS (6.23)
                                     2                        S    2     α
for all u ∈ D(qα− ). Moreover, the constant R depends only on kKk∞ .
Proof. Denote by J the term on the left-hand side of (6.23). For any ε > 0 we have
       Z                                   Z                Z
                                         2         2     1
           K(s)v(s)ψ(0)w(s, 0)dS ≤ εψ(0)       v(s) dS +       K(s)2 w(s, 0)2 dS,
         S                                   S           4ε S
and there holds, for sufficiently small ε,
      Z                                Z
                                   B
  J≥      |∂t w(s, t)|2 dSdt − α +         w(s, 0)2 dS
        Σ                           ε    S
                                           Z                       Z
                                         2     K(s)
                                  − ψ(0)              + ε v(s) dS − β w(s, δ)2 dS, (6.24)
                                                              2
                                            S     2                   S

with B = supS (K 2 + |K|). Due to (6.18) and to the inequality (2.6) of Lemma 2.2 we
have
                Z δ
                              2
                    ∂t w(s, t) dt − αw(s, 0)2 − βw(s, δ)2 ≥ 0, s ∈ S.
                     0
It follows that for any η ∈ (0, 1) we can estimate
   Z
       |∂t w(s, t)|2 dSdt
    Σ
                      Z                               Z                      Z
                  ≥η                  2                       2
                          |∂t w(s, t)| dSdt + (1 − η)α w(s, 0) dS + (1 − η)β   w(s, δ)2 dS,
                     Σ                                 S                          S

and the substitution into (6.24) gives
        Z                              Z                    Z
                                     B
  J ≥η     |∂t w(s, t)|2 dSdt − ηα +        w(s, 0)2 dS − ηβ     w(s, δ)2 dS
         Σ                            ε   S                    S
                                                                   Z           
                                                                 2     K(s)
                                                         − ψ(0)              + ε v(s)2 dS.
                                                                    S    2
                                                                                                      22

Therefore, choosing ε = R/α with R > 0 and then using Lemma 2.3 we obtain, for any
ℓ ∈ (0, δ),
         Z                                     Z                        Z                 
                        2
                                      B                     2       2             2
  J ≥η      |∂t w(s, t)| dSdt − η +         α ℓ |∂t w(s, t)| dSdt +        |w(s, t)| dSdt
          Σ                            R           Σ                  ℓ Σ
              Z                            Z                       Z 
                               2          2                                K(s) R 
       − ηβ ℓ |∂t w(s, t)| dSdt +                     2
                                               w(s, t) dSdt − ψ(0) 2
                                                                                +      v(s)2 dS
                  Σ                       ℓ Σ                          S    2       α
       h             B         iZ                       2      Bα        Z
                                                  2
    = η − ℓα η +           − ℓβη      |∂t w(s, t)| dSdt −    ηα +     + ηβ       w(s, t)2 dSdt
                       R            Σ                     ℓ       R            Σ
                 Z 
               2      K(s) R          2
       − ψ(0)               +      v(s) dS.
                  S      2     α
Choose ℓ = ρ(α + β)−1 with ρ ∈ (0, 1/2) and R ≥ B(1 − ρ)−1 , then the choice
                                 ℓBα              ρBα
                       η=                 =                 ∈ (0, 1),
                            R 1 − ℓ(α + β)   (α + β)R(1 − ρ)
implies
                                              B
                                    η − ℓα η +    − ℓβη = 0,
                                               R
and
                                                           Z 
              2Bα(α + β) 2Bα(α + β)                               K(s) R 
  J ≥−                  +                         kwk2L2 (Σ,dSdt) − ψ(0)2
                                                                      +    v(s)2 dS.
               R(1 − ρ)     ρR                                 S    2   α
                                                                                (6.25)
As R can be taken arbirary large, we may choose it in order to have
                                                   
                                           1      1
                                R > 4B         +      ,
                                          1−ρ ρ
then there exists α0 > 0 such that for α ≥ α0 we have
                                                     
                            2Bα(α + β) 2Bα(α + β)         α2
                                         +              <    ,
                             R(1 − ν)          Rν         2
which gives the result.                                                                               
  Substituting the result of Lemma 6.5 into (6.22) and using the equality
                          kuk2L2 (Σ,dSdt) = kvk2L2 (S,dS) + kwk2L2 (Σ,dSdt)
we deduce
                                                  Z
  qα− (u, u) − E β kuk2L2 (Σ,dSdt) ≥ (1 − c0 δ)           gρµ ∂ρ v(s)∂µ v(s)dS
                                                      S
              Z 
                  K(s) R                 α2 
   − ψ(0)2            +   v(s)2 dS − E β +      kwkL2 (Σ,dSdt) − c1 kuk2L2 (Σ,dSdt) . (6.26)
               S   2    α                  2
We choose δ = α−κ , κ ∈ [2/3, 1), then Lemma 2.2 provides
                            E β = −α2 + o(1),              ψ(0)2 = 2α + o(1).
Using rough estimates, we deduce, as α → +∞:
                                           Z
   −          β   2
 qα (u, u) − E kukL2 (Σ,dSdt) ≥ (1 − c0 δ)     gρµ ∂ρ v∂µ v dS
                                             Σ
                         Z
                                                          α2
                      − (αK + 2R + 1)v(s)2 dS +              kwk2L2 (Σ,dSdt) − c1 kuk2L2 (Σ,dSdt) .
                           Σ                               4
                                                                                         23

Following the arguments of Section 5.2, we get
                               β
                                                        
                   Ej (Q−
                        α ) − E ≥ Ej − (1 − c0 δ)∆S − αK + O(1).
Finally, we get the desired lower bound of Theorem 1.3 by using (6.16).

                        7. Proof of Corollaries 1.4 and 1.6
  In this section, we assume that Ω is C 2 -admissible. Then the function K is bounded, and
an easy adaptation of [10, Proposition 1] to the non-euclidean setting gives the following:
Lemma 7.1. For any fixed j ∈ N there holds
                     Ej (−∆S − αK) = −Kmax α + o(α),       α → +∞.                    (7.1)
Proof of Corollary 1.4. It is sufficient to substitute the estimate (7.1) into the asymp-
totics (1.5) of Theorem 1.2.                                                           
   To prove Corollary 1.6 we need a rough estimate for the essential spectrum of QΩα.
Recall that for a self-adjoint operator Q, we have denoted by E(Q) the infimum of its
essential spectrum. Then there holds:
Lemma 7.2. Assume that ∂Ω is non-compact and denote K∞ := lim sups→∞ K(s), then
E(QΩ       2
   α ) ≥ −α − K∞ α + o(α) for large α.

Proof. Lets us modify a bit the construction of Subsection 3.1. By assumption, for any
K0 > K∞ there exists a compact domain S0 ⊂ S such that K(s) ≤ K0 for s ∈ S \ S0 . Set
S1 := S \ S0 . Now let qα′ denote the quadratic form given by the same expression as qαΩ
but acting on the domain D(qα′ ) := H 1 (Ω0δ ) ⊕ H 1 (Ω1δ ) ⊕ H 1 (Θδ ) with
                                                         
              Ω0δ := Φ S0 , (0, δ) , Ω1δ := Φ S1 , (0, δ) , Θδ := Ω \ Ω0δ ∪ Ω1δ
and δ is sufficiently small. Let Q′α be the self-adjoint operator associated with qα′ and
acting in L2 (Ω). Due to the form inequality QΩ            ′               Ω      ′
                                                    α ≥ Qα we have E(Qα ) ≥ E(Qα ). On
                                     ′      0     1          N           j
the other hand, one represents Qα = Qα ⊕ Qα ⊕ (−∆)Θδ , where Qα , j ∈ {0, 1}, is the
self-adjoint operator in L2 (Ωjδ ) generated by the quadratic form
                           Z                  Z
                 j
               qα (u, u) =           2
                                |∇u| dx − α          u2 dS, D(qαj ) = H 1 (Ωjδ ),
                           Ωjδ              ∂Ωjδ ∩S

and (−∆)N                                                          0
         Θδ is the Neumann Laplacian in Θδ . Note that the domain Ωδ is bounded, hence,
the operator Q0α has an empty essential spectrum. It follows that
                                                                        
            E(QΩ          ′               1          N               1
                 α ) ≥ E(Qα ) = min E(Qα ), E((−∆)Θδ ) ≥ min E(Qα ), 0 .

On the other hand, the analysis of Section 5.1 can be applied to the operator Q1α . In
particular, the choice (5.23) for δ gives
                     E1 (Q1α ) ≥ −α2 + E1 (−∆N
                                             S1 − αK) + O(log α),

where −∆N
        S 1 is the Neumann realization of the positive Laplace-Beltrami operator on S1 .
As K ≤ K0 in S1 , we have
                       E(Q1α ) ≥ E1 (Q1α ) ≥ −α2 − K0 α + O(log α)
and, subsequently, E(QΩ       2
                      α ) ≥ −α − K0 α + O(log α). As K0 > K∞ is arbitrary, the result
follows.                                                                           
Proof of Corollary 1.6. Let N ∈ N be fixed. Due to Corollary 1.4 and Lemma 7.2 for
large α we have EN (QΩ        Ω             Ω            th eigenvalue of QΩ due to the
                     α ) < E(Qα ), and EN (Qα ) is the N                   α
min-max principle.                                                                   
                                                                                          24

             8. Analysis of the reduced operator on the boundary
   In this section we gather various standard estimates on the low-lying eigenvalues of
−∆S − αK, depending on the hypotheses on the dimension ν and on K. In this section,
for the cas of an unbounded ∂Ω we assume that the assumption (1.7) holds.
   Now note that by setting h = α−1/2 and V = −K, the operator −∆S − αK writes as
                                                    
                                  −h−2 −h2 ∆S + V
and enters naturally the framework of Schrödinger operators in the semi-classical limit
h → 0. The assumption (1.7) writes now
                                 lim inf V (s) > inf V (s),
                                  s→∞            s∈S
and ensures that the asymptotics of the low-lying eigenvalues of the reduced operator can
be determined by the behavior of V near its minima (that are the maxima of K), under
suitable hypotheses.
Remark 8.1. Assume that the measure of the set K −1 ({Kmax }) is 0. Then the word-
by-word adaptation of [1, Lemma 3.2] to the non-euclidean setting gives, for any fixed
j ∈ N,                                      
                   Ej − ∆S + α(Kmax − K) → +∞ as α → +∞                          (8.1)
If, in addition, Ω is C 3 -admissible, then (1.6) can be decomposed as
                                                                    
               Ej (QΩ         2
                    α ) = −α − Kmax α + Ej − ∆S + α(Kmax − K) + O(1),
                                         
and the term Ej −∆S +α(Kmax −K) has a lower order with respect to α, see Lemma 7.1,
but is large with respect to the remainder O(1), see (8.1), and hence provides a refinement
with respect to the first order asymptotics (1.1).
  The aim is now to describe more precise asymptotics on −∆S − αK, in order to see
the possible gap between eigenvalues, in particular we want to compare Ej (−∆S − αK) +
Kmax α to the remainders in Theorems 1.2 and 1.3. The most commonly studied case is
when the maxima of K are non-degenerate, see [33, Theorem 5.1] or [18]:
Proposition 8.2. Assume that the boundary of Ω is C 5 and, if non-compact, satisfies
(1.7). Furthemore, assume that the function K admits a unique maximum at s0 ∈ S and
that the Hessian of (−K) at s0 is positive-definite. Denote by µk the eigenvalues of the
Hessian and
                             nX ν−1 r                       o
                                      µk
                        E :=             (2nk − 1), nk ∈ N ,                       (8.2)
                                       2
                                 k=1
then for each fixed j ∈ N there holds:
              Ej (−∆S − αK) = −Kmax α + ej α1/2 + O(α1/4 ) as α → +∞,
where ej is the j th element of E, counted with multiplicity. Moreover, if Ω is C 6 , and if
ej is of multiplicity one, the remainder can be replaced by O(1).
   By combining Proposition 8.2 with Theorem 1.2 we obtain Corollary 1.7. Remark that
for ν = 2 one is reduced to
                                r
                              n −K ′′ (s )                  o
                                          0
                           E=               (2n − 1), n ≥ 1
                                       2
and all the elements are of multiplicity one. Therefore, by combining Theorem 1.2 and
Proposition 8.2, we recover the first terms of the asymptotic expansion (1.2), see [15,
Theorem 1.1].
                                                                                            25

   Other cases of extrema are harder to handle, due to the different notions of degeneracy
for the maxima of K, and to the possible interactions with the metric near the maxima.
However, in the case ν = 2, we have the following:
Proposition 8.3. Let ν = 2 and Ω be C 2p+3 -admissible with p ≥ 2 and, if ∂Ω is non-
compact, such that the assumption (1.7) is satisfied. Furthermore, assume that the curva-
ture K admits a unique global maximum at s0 with
                                                               
               K(s) = K(s0 ) − Cp (s − s0 )2p + O (s − s0 )2p+1 , s → s0 ,
where Cp > 0 and s is an arc-length of the connected component Γ of the boundary at
which K takes the maximal value. Then we have the following expansion
                                                        1           1
                    Ej (−∆S − αK) = −Kmax α + ej α p+1 + O(α 2(p+1) ),
where ej is the j th eigenvalue of the operator −∂s2 + Cp s2p in L2 (R). Moreover, if ∂Ω is
C 2p+4 , then the remainder can be replaced by O(1).
Proof. Since we are not interested in exponentially small terms, it suffices by standard
arguments to reduce the analysis to a neighborhood of the minimizer in Γ, denoted by
Γ0 , with Dirichlet boundary conditions at the ends, see [18]. Let γ : R/|Γ0 |Z → Γ0
be an arc-length parametrization of Γ0 . Since the parametrization is normalized and
the metrics in local coordinates is g = kγ ′ k, we only have to consider −∆ − αK on
the interval (s0 − η, s0 + η), with η > 0 fixed, and Dirichlet boundary condition. The
following asymptotics is then a simple consequence of [28, Theorem 2.1] applied with the
semi-classical parameter h = α−1/2 :
                                                              
                                     1       X         −   k
               Ej (−∆S − αK) = α p+1 ej +       βj,k α 2(p+1)  , βj,k ∈ R.
                                               k≥1

If ∂Ω is C 2p+4 , then the curvature is C 2p+2 , and we have the Taylor expansion
                                                                    2p+2
      K(s) = K(s0 ) − Cp (s − s0 )2p + Cp′ (s − s0 )2p+1 + O (s − s0       ), Cp′ ∈ R.

Then, by combining the simplicity of the eigenvalues (ej )j≥1 , the parity of the eigenvectors
of −∂s2 + Cp s2p , and the oddness of the remainder Cp′ (s − s0 )2p+1 in the asymptotic expan-
sion of K, it is standard to show that βj,2 = 0 for all j ≥ 1, see for example [7, Theorem
4.23] for the case p = 1.                                                                   
  The combination of Proposition 8.3 with Theorem 1.2 gives Corollary 1.8.
Remark 8.4. The above statements can be adapted easily to the case where K has several
maxima by using the principle that “each well creates its own series of eigenvalues”.
Corollary 8.5. Let j ∈ N, and assume one of the two following:
    • The hypotheses of Proposition 8.2 hold, and ej is of multiplicity 1 in the set E.
    • The hypotheses of Proposition 8.3 hold.
Then, for α large enough, Ej (QΩ
                               α ) is a simple eigenvalue.

   When we are not in the hypotheses of Remark 8.1, few results exist on the asymptotics
of the first eigenvalues. For example, we can show
Proposition 8.6. Assume that the interior of K −1 ({Kmax }) is not empty. Then, for any
fixed j ∈ N,
                  Ej (−∆S − αK) = −αKmax + O(1) as α → +∞.
                                                                                             26

Proof. Denote by ω ⊂ ∂Ω an open subset of the interior of K −1 {Kmax } with a smooth
boundary. Introduce −∆D  ω , the Laplace-Beltrami operator in ω with the Dirichlet bound-
ary condition. This operator has compact resolvent and we denote by EjD (ω), j ∈ N, its
eigenvalues, and by uj associated normalized eigenfunctions. Denote by Uj the extensions
of uj to ∂Ω by zero, then
                 Z                       Z
                    g ∂ρ Uj ∂µ Uj dS − α KUj2 dS = −αKmax + EjD (ω).
                     ρµ
                   S                    S
As Uj                              2
      are mutually ortohogonal in L (S, dS), we deduce from the min-max principle that
Ej (−∆S − αK) ≤ −αKmax + EjD (ω), and the sought estimate follows.                           
   In particular, in the situation of Proposition 8.6 Theorems 1.2 and (1.3) does not provide
the gap between the eigenvalues of QΩ    α as α → +∞.
   We remark that a particular case of a piecewise constant curvature was recently studied
in [31], and the eigenvalue gaps appear to have finite limits.

                                       9. Periodic case
   The preceding analysis can also be applied to periodic problems. Namely, assume
that there exist linearly independent vectors a1 , . . . , am , m ≤ ν, such that Ω is invariant
under the shifts x 7→ x + aj , j ∈ {1, . . . , m}, and that the quotient (elementary cell)
ω := Ω/(Za1 +· · ·+Zam ) is compact, and then the quotient surface σ := S/(Za1 +· · ·+Zam )
is also compact. Such a situation is covered by the Floquet theory [22]. Namely, for
θ = (θ1 , . . . , θm ) ∈ Tm , T := R/2πZ denote by QΩ α (θ) the self-adjoint operator generated
by the quadratic form
                                             Z              Z
                                 Ω,θ
                                qα (u, u) :=    |∇u| dx − α u2 dS,
                                                    2
                                              ω               σ
                                   n                                                 o
          D(qα ) = Hθ (ω) := u ∈ Hloc (Ω) : u(· + aj ) = eiθj u(·), j = 1, . . . , m .
                 Ω,θ      1                1


It can be easily checked that the operators QΩ (θ) are with compact resolvents, and it is a
standard fact of the Floquet theory that for each fixed j ∈ N the so-called band function
                                                              
                            Tm ∋ θ 7→ Ej (θ, α) := Ej QΩ α (θ)
is continuous, and that
                               [                          
                  spec QΩ
                        α :=         Bj (α),     Bj (α) := Ej (θ, α) : θ ∈ Tm .
                               j∈N

The segment Bj (α) is usually called the j th spectral band of QΩ
                                                                α.
  An analogous representation of the spectrum holds for the reduced operator −∆S −αK.
Namely, denote by Tα (θ) the self-adjoint operator acting in L2 (σ) associated with the
quadratic form
                                    Z                    Z
                        θ
                       tα (v, v) :=     g ∂ρ v∂µ v dS − α Kv 2 dS,
                                         ρµ
                                     σ                    σ
                             n                                                  o
             θ     1                  1                  iθj
         D(tα ) = Hθ (σ) := u ∈ Hloc (S) : u(· + aj ) = e u(·), j = 1, . . . , m .
Again, one checks that Tα (θ) have compact resolvents and the band functions
                                                             
                             Tm ∋ θ 7→ εj (θ, α) := Ej Tα (θ)
are continuous and
                                       [                        
              spec(−∆S − αK) :=              βj (α),   βj (α) := εj (θ, α) : θ ∈ Tm ,
                                       j∈N
                                                                                                      27


and the segment βj (α) will be called the j th spectral band of −∆S − αK.
   One can easily see that the proofs of Theorems 1.2 and 1.3 also work for the operators
  Ω
Qα (θ), which gives the following results:
Theorem 9.1. For any fixed j ∈ N there holds, as α → +∞,
                        Ej (θ, α) = −α2 + εj (θ, α) + Rj (θ, α),       θ ∈ T,
where Rj (θ, α) = O(log α) if Ω is C 2 and Rj (θ, α) = O(1) if Ω is C 3 , and the remainder
estimate is uniform in θ ∈ T.
Corollary 9.2. If j ∈ N is fixed and α → +∞, then the j th spectral band of QΩ       2
                                                                               α + α and
                                                           
the j th spectral band of −∆S − αK are located in a O R(α) -neighborhood of each other,
where R(α) = log α for the C 2 -admissible case and R(α) = 1 for the C 3 -admissible one.
   The result of Theorem 9.1 can be used to study some spectral properties specific for
periodic operators. Recall that a non-empty interval (a, b) ⊂ R is called a (spectral) gap
of a self-adjoint operator A if (a, b) ∩ spec A = ∅ but a, b ∈ spec A. The existence of
spectral gaps is one of the principal questions in the spectral theory of periodic operators,
cf. [2, 4, 19, 23]. In view of Theorem 9.1, the existence of sufficiently large gaps for the
reduced operator −∆ − αK (i.e. having the length of order ακ with some κ > 0) implies
the existence of gaps for the Robin Laplacian QΩ α , and the reduced operator was studied in
numerous preceding works, cf. [29, 34]. For, example the semiclassical analysis of periodic
operators of the form −h∆S + V /h carried out in [34, Theorem 1.1] gives the following
result:
Corollary 9.3. Assume that Ω is C ∞ and periodic as described above. Furthermore,
assume that the function σ ∋ s 7→ K(s) admits a unique maximum at s0 , and that the
Hessian of (−K) at s0 is positive-definite. Let µj be the eigenvalues of the Hessian and
the numbers ej be defined as in Proposition 8.2, then
    (1) for each j ∈ N there exists C > 0 such that
                                                                        
             spec(QΩ      2
                    α + α + Kmax α) ∩ ej α
                                            1/2
                                                − Cα2/5 , ej α1/2 + Cα2/5 6= ∅
       for large α, and
   (2) for each C1 > 0 there exist C2 , C3 > 0 such that
                                                  
  spec(QΩ      2
         α + α + Kmax α) ∩ − C1 α
                                    1/2
                                         , C1 α1/2
                                                    [                                        
                                               ⊂        ej α1/2 − C2 α2/5 , ej α1/2 + C2 α2/5
                                                      ej ≤C3

        as α → +∞.
In particular, for any N ∈ N there exists αN > 0 such that the operator QΩ
                                                                         α has at least
N gaps for α > αN .
  The localization of the spectrum given in the preceding corollary is not expected to be
optimal for periodic domains. Furthermore, it would be interesting to understand some
questions related to the location of the extrema of the band functions, cf. [2]. We hope to
analyze the periodic case in greater detail in subsequent works.

                                           References
     [1] Y. Belaud, B. Helffer, L. Véron: Long-time vanishing properties of solutions of some semilinear
         parabolic equations. Ann. Inst. Henri Poincaré Anal. nonlinear 18:1 (2001) 43–68.
     [2] D. Borisov, K. Pankrashkin: Quantum waveguides with small periodic perturbations: gaps and
         edges of Brillouin zones. J. Phys. A 46 (2013) 235203.
                                                                                                     28

 [3] G. Carron, P. Exner, D. Krejčiřı́k: Topologically nontrivial quantum layers. J. Math. Phys. 45:2
     (2004) 774–784.
 [4] V. Chiadò Piat, S. A. Nazarov, K. Ruotsalainen: Spectral gaps for water waves above a corrugated
     bottom. Proc. Roy. Soc. A 469 (2013) 20120545.
 [5] D. Daners, J. B Kennedy: On the asymptotic behaviour of the eigenvalues of a Robin problem.
     Differ. Integr. Eq. 23:7/8 (2010) 659–669.
 [6] E. B. Davies: Spectral theory and differential operators. Cambridge University Press, 1995.
 [7] M. Dimassi, J. Sjöstrand: Spectral asymptotics in the semi-classical limit, volume 268 of London
     Mathematical Society Lecture Note Series, Cambridge University Press (1999).
 [8] P. Exner, A. Minakov: Curvature-induced bound states in Robin waveguides and their asymptot-
     ical properties. J. Math. Phys. 55 (2014) 122101.
 [9] P. Exner, A. Minakov, L. Parnovski: Asymptotic eigenvalue estimates for a Robin problem with
     a large parameter. Portugal. Math. 71:2 (2014) 141–156.
[10] P. Freitas, D. Krejčiřı́k: Instability results for the damped wave equation in unbounded domains.
     J. Differential Eq. 211:1 (2005) 168–186.
[11] P. Freitas, D. Krejčiřı́k: The first Robin eigenvalue with negative boundary parameter. Adv.
     Math. (to appear), preprint arXiv:1403.6666 (2014)
[12] L. Friedlander, M. Solomyak: On the spectrum of the Dirichlet Laplacian in a narrow infinite
     strip. in T. Suslina , D. Yafaev (Eds.): Spectral Theory of Differential Operators: M. Sh. Birman
     80th Anniversary Collection. (Amer. Math. Soc. Transl. Ser. 2, Vol. 225 of Advances in the
     Mathematical Sciences, Amer. Math. Soc., 2008) 103–116.
[13] L. Friedlander, M. Solomyak: On the spectrum of the Dirichlet Laplacian in a narrow strip.
     Israel J. Math. 170 (2009) 337–354.
[14] T. Giorgi, R. Smits: Eigenvalue estimates and critical temperature in zero fields for enhanced
     surface superconductivity. Z. Angew. Math. Phys. 58:2 (2007) 224–245.
[15] B. Helffer, A. Kachmar: Eigenvalues for the Robin Laplacian in domains with variable curvature.
     Arch. Rat. Mech. Anal. (to appear), preprint arXiv:1411.2700 (2014).
[16] B. Helffer, A. Morame: Magnetic bottles in connection with superconductivity. J. Funct. Anal.
     185:2 (2001) 604–680.
[17] B. Helffer, K. Pankrashkin: Tunneling between corners for Robin Laplacians. J. London Math.
     Soc. 91 (2015) 225–248.
[18] B. Helffer, J. Sjöstrand: Multiple wells in the semi-classical limit I. Commun. PDE 9:4 (1984)
     337–408.
[19] R. Hempel, O. Post: Spectral gaps for periodic elliptic operators with high contrast: an overview.
     In H. G. W. Begehr, R. P. Gilbert, M. W. Wong (Eds.): Progress in Analysis. Proceedings of the
     3rd International ISAAC Congress Berlin 2001. Vol. 1 (World Scientific Publishing Co., Inc.,
     River Edge, NJ, 2003) 577–587.
[20] D. Krejčiřı́k: Spectrum of the Laplacian in a narrow curved strip with combined Dirichlet and
     Neumann boundary conditions ESAIM: Control, Optimisation and Calculus of Variations 15
     (2009) 555-568.
[21] D. Krejčiřı́k: Spectrum of the Laplacian in narrow tubular neighbourhoods of hypersurfaces with
     combined Dirichlet and Neumann boundary conditions. Math. Bohem. 139 (2014) 185–193.
[22] P. Kuchment: Floquet Theory For Partial Differential Equations (Vol. 60 of Operator Theory:
     Advances and Applications, Birkhäuser, 1993).
[23] A. Khrabustovskyi: Opening up and control of spectral gaps of the Laplacian in periodic domains.
     J. Math. Phys. 55 (2014) 121502.
[24] P. Kuchment: Quantum graphs I. Some basic structures. Waves Random Media 14 (2004) S107–
     S128.
[25] A. A. Lacey, J. R. Ockendon, J. Sabina: Multidimensional reaction diffusion equations with
     nonlinear boundary conditions. SIAM J. Appl. Math. 58 (1998) 1622–1647.
[26] M. Levitin, L. Parnovski: On the principal eigenvalue of a Robin problem with a large parameter.
     Math. Nachr. 281 (2008) 272–281.
[27] Y. Lou, M. Zhu: A singularly perturbed linear eigenvalue problem in C 1 domains. Pacific J.
     Math. 214:2 (2004) 323–334.
[28] A. Martinez, M. Rouleux: Effet tunnel entre puits dégénérés. Comm. Partial Differential Equa-
     tions 13:9 (1988) 1157–1187.
[29] A. Outassourt: Comportement semiclassique pour l’opérateur de Schrödinger à potentiel
     périodique. J. Funct. Anal. 72:1 (1987) 65–93.
                                                                                                     29

    [30] K. Pankrashkin: On the asymptotics of the principal eigenvalue for a Robin problem with a large
         parameter in planar domains. Nanosystems: Phys. Chem. Math. 4:4 (2013) 474–483.
    [31] K. Pankrashkin: On the Robin eigenvalues of the Laplacian in the exterior of a convex polygon.
         Nanosystems: Phys. Chem. Math. 6 (2015) 46–56.
    [32] K. Pankrashkin, N. Popoff: Mean curvature bounds and eigenvalues of Robin Laplacians. Calc.
         Var. PDE. (to appear), preprint arXiv:1407.3087 (2014).
    [33] B. Simon: Semiclassical analysis of low lying eigenvalues. I. Non-degenerate minima: Asymptotic
         expansions. Ann. Inst. H. Poincaré A 38:3 (1983) 295-308.
    [34] M. A. Shubin: Semiclassical asymptotics on covering manifolds and Morse inequalities. Geom.
         Funct. Anal. 6:2 (1996) 370–409.

  Laboratoire de mathématiques (UMR 8628 du CNRS), Université Paris-Sud, Bâtiment 425,
91405 Orsay Cedex, France
  E-mail address: konstantin.pankrashkin@math.u-psud.fr
  URL: http://www.math.u-psud.fr/~pankrash/

  Institut mathématique de Bordeaux, Université Bordeaux 1, 351 cours de la libération,
33405 Talence Cedex, France
  E-mail address: nicolas.popoff@math.u-bordeaux1.fr
  URL: http://www.math.u-bordeaux1.fr/~npopoff/
