                                          Curvature-induced bound states in Robin
                                          waveguides and their asymptotical properties
                                               Pavel Exner1,2 and Alexander Minakov1,3
                                               1 Doppler Institute for Mathematical Physics and Applied Mathemat-




arXiv:1406.7624v1 [math-ph] 30 Jun 2014
                                               ics, Czech Technical University in Prague, Břehová 7, 11519 Prague,
                                               Czech Republic
                                               2
                                                 Department of Theoretical Physics, Nuclear Physics Institute
                                               ASCR, Hlavn 130, 25068 Řež near Prague, Czech Republic
                                               3
                                                 Department of Physics, Faculty of Nuclear Science and Physical
                                               Engineering, Czech Technical University in Prague, Pohraničn
                                               1288/1, 40501 Děčn, Czech Republic
                                               exner@ujf.cas.cz, minakov.ilt@gmail.com
                                               We analyze bound states of Robin Laplacian in infinite planar
                                               domains with a smooth boundary, in particular, their relations to
                                               the geometry of the latter. The domains considered have locally
                                               straight boundary being, for instance, locally deformed halfplanes
                                               or wedges, or infinite strips, alternatively they are the exterior of
                                               a bounded obstacle. In the situation when the Robin condition is
                                               strongly attractive, we derive a two-term asymptotic formula in
                                               which the next-to-leading term is determined by the extremum
                                               of the boundary curvature. We also discuss the non-asymptotic
                                               case of attractive boundary interaction and show that the discrete
                                               spectrum is nonempty if the domain is a local deformation of a
                                               halfplane or a wedge of angle less than π, and it is void if the
                                               domain is concave.


                                          1    Introduction
                                          The task of control motion of quantum particles guiding them in a desired
                                          direction has both the theoretical and practical significance. Most often this
                                          problem is addressed in the situation when the motion is confined to a strip
                                          or a tube which models real-world objects such as semiconductor quantum
                                          wires, carbon nanotubes, etc. The boundary condition involved are at that
                                          typically Dirichlet, modeling a hard-wall boundary, or Neumann. A single
                                          boundary cannot produce in this situation a guided motion unless an external
                                          field is added [7, 10].

                                                                                1
     The situation changes if the boundary is described by a mixed-type con-
dition as in (2.2) below, conventionally called Robin, representing and at-
tractive contact interaction, in other words, with the parameter β > 0. If a
two-dimensional particle is confined to a halfplane with such a boundary, its
spectrum is purely absolutely continuous covering the interval [−β 2 , ∞) and
contains a component describing states moving along the boundary, in par-
ticular, any states referring to the spectral projection on the interval [−β 2 , 0)
have this property. The spectral simple picture becomes more complicated,
however, when the boundary is geometrically nontrivial and the correspond-
ing Laplacian may have a nonempty discrete spectrum. A discussion of such
eigenvalues is the main topic of this paper.
     Our main result concerns the asymptotic behavior of the eigenvalue in
the strong coupling case, β → ∞. We are going to consider planar domains
the boundary of which is an infinite smooth curve without self-intersections
assuming that its curvature decays fast enough outside a compact. The
method we shall use combines a bracketing argument with spectral analysis
of Laplacians in a curved strip. It was first proposed in Ref. [6] dealing with
curve-supported δ interactions in the plane. Its ‘one-sided’ version suitable
for application to domains with a boundary proved to be different, however,
due to the presence of an extra term which changes substantially the result
[3] leading to asymptotic inequalities only. A two-term asymptotic expansion
can be nevertheless obtain if one combines this conclusion with a variational
estimate devised by K. Pankrashkin [11] for treatment of the ground state.
We note that the smoothness of the boundary is essential, once it is allowed
to have angles the asymptotics changes in the leading order [9].
     In contrast to Ref. [3] where Robin billiards were discussed we consider
here infinite domains where the essential spectrum is nonempty and the dis-
crete one may not exist. The asymptotic formula we are going to derive will
yield sufficient conditions for its existence and nonexistence in the strong
coupling regime. In addition, one is able to show that if such geometri-
cally induced bound states exist, a sufficiently strong boundary attraction
can produce any prescribed finite number of them. Furthermore, with small
modifications the technique will allow us to derive analogous asymptotic ex-
pansion for Robin waveguides where the existence of the discrete spectrum
is known [8], as well for domains being the exterior of a compact obstacle.
     The result concerning a single infinite boundary curve raises naturally
the question about existence of bound states beyond the asymptotic regime.
For simplicity we shall suppose that the parts of the boundary outside a

                                        2
compact are straight and not parallel, on the other hand we weaken the
requirement on the local smoothness. The essential spectrum then coincides
with the interval [−β 2 , ∞) and using variational arguments we are going to
show that the discrete spectrum is nonempty provided the ‘internal’ angle
between the two boundary asymptotes does not exceed π, except for the
trivial case when the domain in question is a halfplane. On the other hand,
while bound states may sometimes exist even for the asymptote angle in
(π, 2π), the discrete spectrum is empty if the domain in question is concave.
    The paper is organized as follows. In the next section we formulate
the state the problem properly and formulate the results about the strong-
coupling asymptotic behavior of the discrete spectrum for the case of a single
infinite boundary. They will be subsequently proved in Section 3; some parts
of the proofs follows closely the analogous arguments in Ref. [3] and will be
presented only briefly. Consequences for a curved Robin waveguide are indi-
cated in Section 4. Then we pass to the non-asymptotic regime and prove
in Section 5 the above mentioned results about the spectrum. Finally, in
Section 6 we shall present the asymptotic result for the Robin problem in
the exterior of a bounded domain.


2    Robin problem in a domain with single in-
     finite boundary
Let us now state our Robin problem properly. Consider a domain Ω ⊂ R2 the
boundary of which is an infinite smooth curve Γ without self-intersections,
and assume that
  (i) there is an a > 0 such that any two points of Γ the arc-length distance
      of which is larger that 3a have disjoint a-neighborhoods in Ω, or equiv-
      alently, there is a a-neighborhood of the boundary in Ω which does not
      intersect itself,
 (ii) Γ : R → R2 is a C 4 -smooth curve, without loss of generality we may
      parametrize itself by its arc length, Γ(s) = (Γ1 , Γ2 ). The orientation is
      chosen in such a way that Ω lies to the left of Γ if we follow it in the
      direction of increasing s,
(iii) the signed curvature γ(s) = Γ′1 (s)Γ′′2 (s) − Γ′2 (s)Γ′′1 (s) of Γ satisfies the
                            −1−ε
      bound |γ(s)|√ ≤ chsi       for some numbers c, ε > 0, where as usual we
      put hsi := 1 + s .  2



                                          3
                                                  ′
(iv) the first two derivatives of γ are bounded, γ+ := max |γ ′ (s)| < ∞ and
                                                            s∈R
      ′′
     γ+  := max |γ ′′ (s)| < ∞.
            s∈R

For future purposes we introduce the following quantities

      γ+ := max |γ(s)| < ∞ ,      γ ∗ := max γ(s) ,   γ∗ := min γ(s) ,    (2.1)
             s∈R                         s∈R                s∈R

which are finite due to the assumptions.

Remark 2.1 The assumption (iii) guarantees that the tangent vector of Γ
has limits as |s| → ∞. On the other hand, it does not require the exis-
tence of asymptotes, as an example one can take a domain delineated by
a parabola, where γ(s) = O(|s|−3/2 ). It will be clear from the discussion
in the next section, however, that the method works even for curves with a
slower curvature decay allowing thus for more ‘exotic’ domains, for instance,
an Ω which is outside a compact (the interior of which contains the origin
of coordinates) bordered by a pair of logarithmic spirals given by parametric
equations r = aj ebθ with a1 < a2 < a1 e2πb .

   Let us now consider the following boundary-value problem in Ω:

                   − ∆f = λf       in Ω ,
                                                                          (2.2)
                      ∂f
                         = βf      on ∂Ω = Γ with β > 0 ,
                      ∂n
                  ∂
where the symbol ∂n  denotes the outside normal derivative. It is straightfor-
ward to check that the quadratic form
                                             Z
                    qβ [f ] = k∇f kL2 (Ω) − β |f (x)|2 ds
                                   2
                                                                         (2.3)
                                               Γ

with Dom(qβ ) = H 1 (Ω) is closed and below bounded; we denote by Hβ the
unique self-adjoint operator associated with it. In general Hβ may not have a
nonempty discrete spectrum — think of the situation when Ω is halfplane —
but we shall assume that σdisc (Hβ ) 6= ∅ and denote by λj its j-th eigenvalue.
The existence of such eigenvalues will be a matter of a further discussion.
Our main goal is to analyze the asymptotic behavior of λj = λj (β) as the


                                        4
parameter β tends to infinity. To state the results we have to introduce the
operator
                               d2    1
                      S = − 2 − γ 2 (s) in L2 (R)                         (2.4)
                              ds     4
with the domain H 2 (R). In view of assumption (iii) its essential spectrum is
(0, ∞) and the discrete spectrum is nonempty unless γ = 0, under the decay
hypothesis we made it is finite. We denote by µj the j-th eigenvalue of S0 ,
provided it exists. Our first main result then reads as follows.

Theorem 2.2 Adopt the assumptions (i)–(iv) and denote #σdisc (S) = M,
the multiplicity taken into account. If γ ∗ > 0, then to any N ∈ N there is a
βN > 0 such that the problem (2.2) has at least N eigenvalues, again counted
with their multiplicity, for all β > βN , satisfying the he following asymptotic
expansion                                               
                        λj (β) = −β 2 − γ ∗ β + O β 2/3 .
The lower bound can be made more precise, specifically
                          2                  
                        γ∗                log β
        λj (β) ≥ − β +        + µj + O            , 1≤j≤M,                 (2.5)
                         2                  β
                          2            
                        γ∗          log β
        λj (β) ≥ − β +        +O             , j>M.                        (2.6)
                         2            β

We see that in the asymptotic regime the condition γ ∗ > 0, in other words,
a local convexity of the boundary, produces a discrete spectrum of any finite
cardinality. Note that this includes situations when the boundary
                                                              R    is a local
deformation of a straight line, since in such a case we have R γ(s) ds = 0,
and consequently, the curvature has to be sign changing.

Remark 2.3 Let us stress that the claim about the existence of any finite
number of bound states for β large enough requires a sufficient smoothness of
the boundary. Not only the proof presented in Sec. 3 needs this assumption,
but the result is not valid without it in general. As an example consider Ω
in the form of a quadrant, {(x, y) ∈ R2 : x > 0, y > 0} with the boundary
consisting of two halflines meeting at the right angle. The spectrum is easily
found by separation of variables: we have σess (Hβ ) = [−β 2 , ∞) and for any
β > 0 there is exactly one simple eigenvalue equal to −2β 2 .



                                       5
3     Proof of Theorem 2.2
Let us start with introducing quadratic forms and operators which we shall
need in the argument. To this aim, we need the following result analogous
to Lemma 2.1 of Ref. [6].

Lemma 3.1 There is an a1 > 0 such that the map Φ,

         R × (0, a) ∋ (s, u) 7→ (Γ1 (s) − uΓ′2 (s), Γ2 (s) + uΓ′1 (s)) ∈ R2 ,

is injective for any a ∈ (0, a1 ].

We skip the proof which is basically the same as in Ref. [6] in checking
that Φ is locally a diffeomorphism. The only new element here is the non-
compactness of Φ(R × (0, a)) which requires the assumption (i) above to
ensure the injectivity.
   Choose now an a ∈ (0, a1 ] to be specified later and denote by Σa the strip
neighborhood of Γ ≡ Γ0 of the width a, in other words

                                 Σa := Φ(R × (0, a)) .

Then Ω \ Σa =: Λa is an unbounded simply connected domain with the
boundary which we denote as Γa . We define the quadratic forms
                           Z
    D                2
                                                
   qa,β [f ] := k∇f kΣa − β |f (x)|2ds for f ∈ f ∈ H 1 (Σa ) : f |Γa = 0 ,
                                Γ0
                                Z
     N
    qa,β [f ] := k∇f k2Σa − β        |f (x)|2ds   for f ∈ H 1 (Σa ),
                                Γ0

which are easily checked to be closed and bounded from below, and denote
by LD       N                                                      D      N
     a,β , La,β the unique self-adjoint operators associated with qa,β , qa,β , re-
spectively. The first key component of the proof is a Dirichlet-Neumann
bracketing — see [12, Sec. XIII.15, Prop. 4] — in our case it consists of
imposing additional boundary condition at Γa . This yields the bounds
                                             D        D
                    LN        N
                     a,β ⊕ (−∆Λa ) ≤ Hβ,0 ≤ La,β ⊕ (−∆Λa )                      (3.7)

in L2 (Ω) = L2 (Σa ) ⊕ L2 (Λa ). Since the estimating operators have a direct-
sum structure and the terms in the inequalities (3.7) referring to Λa are

                                             6
positive, in order to estimate the (negative) eigenvalues of Hβ,0 it is sufficient
to estimate those of LD           N
                        a,β and La,β .
    To achieve this, we bring in the second main ingredient of the proof
introducing a ‘straightening’ transformation in the spirit of Ref. [4] which
produces a pair of operators in L2 (R × (0, a)) that are unitarily equivalent
to LD          N
     a,β and La,β , respectively. This is achieved by introducing the following
change of variables,
                                                     1
                           f (x1 , x2 ) =                     ϕ(s, u) ;                        (3.8)
                                               (1 − uγ(s))1/2

then it is straightforward to check that for any function f ∈ H 2 (Σa ) we have
also ϕ ∈ H 2 (R × (0, a)) and
                                 "
                                                 2       2
                                       1     ∂ϕ      ∂ϕ
         2          2
  (|fx1 | + |fx2 | )(x1 , x2 ) =           2
                                                   +       + Ve |ϕ|2
                                   (1 − uγ) ∂s       ∂u
                                                                     
               uγ ′          ∂ϕ       ∂ϕ         γ         ∂ϕ        ∂ϕ
       +                  ϕ      +ϕ        +            ϕ      +ϕ          (s, u)
           2(1 − uγ)3        ∂s       ∂s     2(1 − uγ)     ∂u        ∂u

with
                                          γ 2 (s)     u2 (γ ′ (s))2
                      Ve (s, u) :=                 +                ,
                                      4(1 − uγ(s))2 4(1 − uγ(s))4
                                            ∂f
where we use the standard shorthands, fxj = ∂x j
                                                 . Then
         ZZ                                                 Z
                       2          2
                                      
               |fx1 | + |fx2 |            dx1 dx2 − β           |f (x)|2 ds
         Σa                                                 Γ
                  Z Za                                  2             Z Za         2
                                1       ∂ϕ                                    ∂ϕ
              =                       2
                                                            du ds +                    du ds
                           (1 − uγ(s)) ∂s                                     ∂u
                  R    0                                              R   0
                  Z Za                                      Z            
                                           2                      γ(s)
              +            V (s, u) |ϕ| du ds −                        + β |ϕ(s, 0)|2 ds
                                                                   2
                  R    0                                    R
                  Z
                           γ(s)
              +                     |ϕ(s, a)|2 ds ,                                            (3.9)
                       2(1 − aγ(s))
                  R




                                                    7
where
                                     ′
                                                                
                            ∂      uγ   (s)      ∂           γ(s)
     V (s, u) = Ve (s, u) −                    −
                            ∂s 2(1 − uγ(s))3     ∂u 2(1 − uγ(s))
                                     ′′
                     2
                    γ (s)         uγ (s)       5 u2 (γ ′ (s))2
         =−                   −              −                  .   (3.10)
              4(1 − uγ(s))2 2(1 − uγ(s))3 4 (1 − uγ(s))4

Armed with these formulæ we can now introduce the sought pair of estimat-
ing operators on L2 (R × (0, a). We consider the domains
        
  QD            1
    a = ϕ ∈ H (R × (0, a)) : ϕ(., a) = 0 ,     QN     1
                                                 a = H (R × (0, a)) , (3.11)

and define on them the quadratic forms bD                N
                                            a,β [ϕ] and ba,β [ϕ], respectively, being
                                             D
equal to the right-hand side of (3.9); for ba,β [ϕ], the last summand is skipped.
    Using these definitions, it is straightforward to check easily the following
claim analogous to Proposition 2.2 of Ref. [6].
                                D     N
Lemma 3.2 The operators Ba,β      , Ba,β  associated with the above quadratic
                                    D     N
forms are unitarily equivalent to La,β , La,β , respectively.
                             D      N
Our aim is to get bounds to Ba,β , Ba,β just introduced using operators with
separated variables. We put
                                                ′′
                             γ 2 (s)          aγ+
              V+ (s) := −               +               ,
                          4(1 + aγ+ )2 2(1 − aγ+ )3
                             γ 2 (s)            ′′
                                              aγ+         5 a2 (γ+′ 2
                                                                   )
              V− (s) := −             2
                                        −             3
                                                        −               ,
                          4(1 − aγ+ )     2(1 − aγ+ )     4 (1 − aγ+ )4

and estimate the right-hand side of the expression (3.9). For an a satisfying
0 < a < γ+ /2 and ϕ belonging to QD         N
                                    a and Qa , respectively, we define

                                       ZZ
                                        a
                                                   2                 Z Za        2
            ebD [ϕ] = (1 − aγ+ )−2           ∂ϕ                             ∂ϕ
              a,β                                      ds du +                       ds du
                                             ∂s                             ∂u
                                       0 R                           0 R
                     Z Za                         γ            Z
                                                       ∗
                 +          V+ (s) |ϕ|2 ds du −            +β         |ϕ(s, 0)|2ds
                                                   2
                     0 R                                         R




                                             8
and
                               ZZ
                                a
                                            2             Z Za       2         ZZa
ebN [ϕ] = (1 + aγ+ )−2                 ∂ϕ                        ∂ϕ
  a,β                                           ds du +                ds du +     V− (s) |ϕ|2 ds du
                                       ∂s                        ∂u
         ∗    Z 0 R                   0 R
                                                                 Z
                                                                               0 R
         γ                           γ ∗
      −     +β   |ϕ(s, 0)|2 ds +                                   |ϕ(s, a)|2 ds ;
          2                      2(1 − aγ∗ )
                       R                                         R

then we have
                                   bD         eD           for f ∈ QD
                                    a,β [ϕ] ≤ ba,β [ϕ]              a,                     (3.12)
                                   bN         eN           for f ∈ QN
                                    a,β [ϕ] ≥ ba,β [ϕ]              a .                    (3.13)
 Let HeD , He N be the self-adjoint operators associated with the forms ebD ,
        a,β  a,β                                                           a,β
ebN , respectively. By T D we denote the self-adjoint operator associated with
  a,β                      a,β
 the form
                               Za            γ      
                                                ∗
                    ta,β [ϕ] = |ϕ′ (u)|2du −
                     D
                                                  + β |ϕ(0)|2
                                              2
                                       0
                           1                      N
defined on {ϕ ∈ H (0, a) : ϕ(a) = 0}. Similarly, Ta,β will be the self-adjoint
operator associated with the form
                Za               ∗    
                                 γ                   γ∗
tN
 a,β (ϕ, ϕ) =
                       ′       2
                     |ϕ (u)| du−    + β |ϕ(0)|2+             |ϕ(a)|2, ϕ ∈ H 1 (0, a) .
                                  2              2(1 − aγ∗ )
                0

Furthermore, we define
                                    d2
 UaD/N = −(1 ∓ aγ+ )−2                  + V± (s)         in L2 (R) with the domain H 2 (R) ,
                                    ds2
where the upper/lower sign refers to the index D/N, respectively. This allows
us to write the estimating operators in the form
         eD = UD ⊗ I + I ⊗ T D ,
         H                                          eN = UN ⊗ I + I ⊗ T N ,
                                                    H                                      (3.14)
           a,β a            a,β                       a,β a            a,β

allowing us to assess contributions coming from the longitudinal and trans-
verse variables separately.
    Consider first the longitudinal part. We denote by µD               N
                                                             j (a) and µj (a) the
                       D         N
j-th eigenvalue of Ua and Ua , respectively, and recall the result of Proposi-
tion 2.3 in Ref. [6] (valid for operators on the line corresponding to an infinite
boundary curve as well [5]):

                                                     9
Lemma 3.3 There exists a constant C > 0 such that the estimates

                                  |µD
                                    j (a) − µj | ≤ Ca                                   (3.15)

and
                                  |µN
                                    j (a) − µj | ≤ Ca                                   (3.16)
hold for any 0 < a < (2γ+ )−1 , where C is independent of a, j.

On the other hand, Lemmata 2.4 and 2.5 of Ref. [3] allow us to estimate the
                          D        N
principal eigenvalue of Ta,β and Ta,β  with an exponential precision. Specifi-
cally, we have the following claims:
                                      
Lemma 3.4 Assume that a β + γ2∗ > 34 . Then Ta,β       D
                                                         has only one negative
                                 D
eigenvalue which we denote by ζa,β . It satisfies the inequalities
          γ∗ 2    D
                              γ∗ 2        γ∗ 2 −a(β+ γ2∗ )
    − β+         ≤ ζa,β ≤− β+        +4 β+         e           .
            2                  2              2
                            ∗
                                      n                   o
                                          γ∗      2 log 5
Lemma 3.5 Assume that β + γ2 > max 2(1−aγ    ∗)
                                                ,   3a
                                                                    N
                                                            . Then Ta,β
                                  N
has a unique negative eigenvalue ζa,β , and moreover, we have
                2                      2                                      2
           γ∗           45          γ∗                ∗
                                                                              γ∗
                                                 
                                               −a β+ γ2       N
      − β+            −          β+            e           ≤ ζa,β ≤−       β+           .
           2             4          2                                         2

Now we are in position to prove the inequalities (2.5) for the first M eigen-
values provided they exist. The bracketing estimate we have squeezes the
eigenvalues in question between those of the operators (3.14). Since the lat-
ter have separated variables, their eigenvalues are sums of eigenvalues of the
longitudinal and transverse component which can be estimated using Lem-
mata 3.3 and 3.5, hence it is sufficient to choose a = β3 log β to arrive at the
inequalities (2.5).
    At the same time, these estimates does not help to establish the existence
of the eigenvalues, because the essential spectrum thresholds of the operators
                                                             D/N
(3.14) are wide apart. Indeed, the essential spectrum of Ua       starts in view
of the assumption (iii) at zero, hence Lemmata 3.3, 3.4, 3.5 yield

                                 e D ) = −β 2 − γ∗ β + O (1) ,
                       inf σdis (H a,β
                                 e N ) = −β 2 − γ ∗ β + O (1) .
                       inf σess (H  a,β


                                               10
Consequently, we cannot be sure that the eigenvalues of H    e D are situated
                                                               a,β
below the essential spectrum threshold of the original operator. To overcome
                                                                           D
this difficulty we need to derive a better upper estimate of the operator Ba,β
from Lemma 3.2. First we note that its essential spectrum threshold is close
to −β 2 .

Lemma 3.6 Under our assumptions we have
                                                 
                           D         2      log β
                inf σess (Ba,β ) = −β + O           .                     (3.17)
                                              β
                                                                     D
Proof: We employ once more Dirichlet-Neumann bracketing and squeeze Ba,β
between a pair of operators on L2 (R × (0, a),
                            DN         D      DD
                           Ba,β,s 0
                                    ≤ Ba,β ≤ Ba,β,s 0
                                                      ,                   (3.18)
                 D
obtained from Ba,β  by adding Neumann and Dirichlet condition, respectively,
at the segments ±s0 × (0, a). Each of the estimating operators is a direct
sum of three parts. The middle one refers to a precompact region, hence it
does not contribute to the essential spectrum, hence it is sufficient to assess
the tail parts. This can be done in a way analogous to (3.14). The difference
is that now we consider the curvature only for |s| > s0 , thus in the transverse
part we modify Lemmata 3.4, 3.5 by replacing γ ∗ and γ∗ by the maximum
and minimum of γ(s) in the tail regions; in view of assumption (iii) the
moduli of these quantities can be made arbitrarily small by choosing s0 large
enough. The same applies to the contribution of the first terms in V± (s) to
the longitudinal part, while the remaining ones are in view of assumption
(iv) proportional to a giving rise to the error term in (3.18).               

Remark 3.7 If we strengthen assumption (iv) requiring, in addition, that
lim|s|→∞ γ (j) = 0 for j = 1, 2, we can localize the essential spectrum threshold
with an exponential precision, however, we do not need such a claim to prove
Theorem 2.2.

Now we can complete the proof by replacing the upper bound coming from
                                                                  D
(3.14) by a more precise variational estimate of the operator Ba,β  . Consider
                                                          2      D
first its principal eigenvalue which satisfies λ1 (β)kϕkL2 ≤ ba,β [ϕ] for any
ϕ ∈ QD  a and construct the following family of trial functions,
                                                        
                      ϕ̂(s, u) = χε (s) e−αu − e−2aα+uα ,

                                       11
 where χε is a smooth function on R with the support located in an ε-
 neighborhood of a point s∗ in which the curvature reaches its maximum,
 γ(s∗ ) = γ ∗ , and ε is a parameter to be determined later. In view of the
 smoothness of γ in combination with assumption (iii), at least one such point
 exists. The function χε used above is supposed to be of the form
                                                   
                                         s − s∗ + ε
                           χε (s) := χ                ,
                                             2ε

 where χ(x) is a fixed smooth function on R with the support in the interval
 (0, 1). It is straightforward to check the scaling relations,

          kχε k2L2 (R) = 2εkχk2L2 (0,1) ,     kχ′ε k2L2 (R) = (2ε)−1 kχ′ k2L2 (0,1) .   (3.19)

 We also note that on the support of χε , i.e. for any s ∈ (s∗ − ε, s∗ + ε) we
 have
                      |γ(s) − γ ∗ | < γ+
                                       ′
                                         |s − s∗ | < γ+
                                                      ′
                                                        ε.
 Computing the terms of the form bD
                                  a,β [ϕ] we get for the longitudinal kinetic
 contribution the estimate
Z aZ                       2             Za sZ∗ +ε             
            1       ∂ ϕ̂                        1                   −αu     −2aα+uα 2
                                                                                       
                         ds du ≤                          + cεu    e     − e
       (1 − uγ(s))2 ∂s                      (1 − uγ ∗ )2
0 R                              0 s∗ −ε
                                                                   
                            1                         1             
            ′    2
         ×(χ (s)) ds du       +O α    −2
                                             + cε        2
                                                           +O α   −3
                                                                          kχk2L2 (R) ,
                           2α                       4α

 where c > 0 is a generic constant independent of β, a, and ε. Similarly,
                  Z aZ          2
                         ∂ ϕ̂                 α             
                                    ds du =     1 + O αe−2aα kχk2L2 (0,L)
                         ∂u                   2
                  0 R




                                                 12
holds for the transverse kinetic term,
     Za Z
           V (s, u) |ϕ̂|2 ds du
     0   R
             Za sZ∗ +ε                                                           !
                            (γ ∗ )2       uγ ′′ (s∗ )   5 u2 (γ ′ (s∗ ))2
         ≤             −              −               −                   + cε
                         4(1 − uγ ∗ ) 2(1 − uγ ∗ )3 4 (1 − uγ ∗ )4
           0 s∗ −ε
                                 2
         × e−αu − e−2aα+uα |χ(s)|2dsdu
                             !
                (γ ∗ )2           1              
         = −            + cε         1 + O α−1 kχk2L2 (R)
                  4             2α
for the potential one, and
      Z                                      
          γ(s)                           γ∗ − ε            2
   −                         2
               + β |ϕ̂(s, 0)| ds ≥ − β +          1 − e−2aα kχk2L2 (R) ,
           2                                2
     R
for the boundary one. Finally, the trial function norm satisfies
            ZZ
             a
                                   1                 
                |ϕ̂(s, u)|2dsdu =      1 + O αe−2aα kχk2L2 (R) .
                                  2α
              0 R

Putting these expressions together and taking (3.19) into account we get
                                ′ 2                                       
          bD
           a,β [ϕ̂]      1 kχ kL2 (0,1)                −1
                                                                   1    −2
                     ≤ 2                    1+O α           + cε      +α
         kϕ̂k2L2 (R)    4ε kχk2L2 (0,1)                            2α
                                                             !
                                                 ∗ 2                    
                                                 (γ )
             +α2 1 + O αe−2aα + −                       + cε 1 + O α−1
                                                   4
                                  
                           γ∗ − ε                       
             −2α β +                  1 + O αe−2aα .
                              2
                              γ∗
Now we choose α = β +              in which case the right-hand side of the last
                               2
inequality becomes
                 ′ 2                                                     2
         1 kχ kL2 (0,1)               −1
                                                    1      −2            γ∗
                            1+O β          + cε         +β         − β+
        4ε2 kχk2L2 (0,1)                            2β                      2
                                                      !
                             γ∗            (γ ∗ )2                    
                   +ε β +          + −             + cε 1 + O α−1 ,
                              2              4

                                         13
and to optimize the last formula with respect to ε we take ε = β −1/3 , which
yields the estimate
                                          2
                      bD
                       a,β [ϕ̂]         γ∗             
                         2
                                 ≤− β+        + O β 2/3 .               (3.20)
                    kϕ̂kL2 (0,L)         2

Since by Lemma 3.6 the window given by this estimate does not overlap with
                                           D
the essential spectrum, the operator Ba,β     has an isolated eigenvalue, and the
same is by Lemma 3.2 and the min-max principle, Ref. [12, Sec. XIII.1], true
for the operator LDa,β , and in turn also for the original operator Hβ .
    The argument for the higher eigenfunctions proceeds in the same way. In
this case we employ slightly modified trial functions of the form
                                                           
                     ϕ̂j (s, u) = χε,j (s) e−αu − e−2aα+uα ,

where the longitudinal part contains the function χε with a shifted argument,
for instance                                          
                                    s − s∗ + (2j − 1)ε
                    χε,j (s) := χ                        .
                                            2ε
The above estimate of the form remains essentially the same, up to the values
of the constants involved. By construction, the functions χε,j with different
values of j have disjoint supports, hence ϕ̂j is orthogonal to ϕ̂i , i = 1, . . . , j−1,
and using once more the min-max principle we conclude that the eigenvalue
λj (β) is again isolated for β large enough and has the upper bound given
by the right-hand side of (3.20). Choosing j = N, we obtain in combination
with (2.5) the first claim of the theorem, the inequality (2.6) follows from
the fact that the eigenvalues are arranged in the ascending order.                    


4     Bound state asymptotics of a curved Robin
      waveguide
Since the analysis performed so far was based on behavior of the solutions
in the vicinity of the boundary, most of the above results can be extended
to situations when the boundary has several disjoint components. A case
of particular interest concerns the case when Ω is a (non-straight) strip of a
constant width d > 0. While in the case when the boundary is a single curve
the discrete spectrum may be empty — for instance, if Ω is strictly concave

                                          14
— a Robin waveguide has always isolated eigenvalues unless it is straight,
similarly to the Dirichlet one treated in Ref. [4]. The claim was proved
in Ref. [8] for repulsive Robin boundary, β < 0, but the argument carries
over without any modification to the attractive case; the only exception is
the Neumann case, β = 0. In fact, the proof in Ref. [8] is done under the
assumption that the strip in straight outside a compact region, however, it
is easily modified to cover situation when it is straight only asymptotically.
    Consider one boundary of the strip as reference one, called Γ0 character-
ized by its curvature γ0 , and suppose that that it satisfies assumptions (ii)–
(iv) of Sec. 2. The strip we are interested in can be regarded as the family
of ‘parallel’ curves Γu : R ∋ s 7→ (Γ0,1 (s) − uΓ′0,2(s), Γ0,2 (s) + uΓ′0,1(s)) ∈ R2
with u ∈ (0, d), while Γd is its other boundary. The curvature of Γu can
                             γ(s)
be expressed as γu (s) = 1−uγ(s)  . We introduce the quantities γ0∗ and γ0,∗
analogous to the extreme of γ in (2.1) and
                                  γ∗                                   γ∗
         γd∗ := max γd (s) =            ,    γd,∗ := min γd (s) =           .
                s∈R            1 − dγ ∗              s∈R            1 − dγ∗
Since the width d introduces a length scale into the problem we replace the
assumption (i) by the following requirement:
 (i’) max{dγ0∗ , dγd∗ } < 1 and the strip Ω does not intersect itself.
The first requirement guarantees the existence of the ‘straightening’ trans-
formation analogous to that of Lemma 3.1 locally, the second one globally.
As before we are interested in the boundary-value problem
                         − ∆f = λf          in Ω
                                                                          (4.21)
                           ∂f
                              = βf on Γ := Γ0 ∪ Γd
                           ∂n
with a parameter β > 0 assumed to be large. The normal has at both part
of the boundary the outward direction; note that the parametrization of Γd
by its arc length has the opposite orientation.
    The argument of the previous section can be copied verbatim for each
boundary component and a < 21 d. The spectrum is then estimated by the
union of the spectra coming from the strip neighborhoods of Γ0 and Γd . Due
to the opposite orientation, the curvature to consider for the latter is in fact
−γ(s). As at least one of the quantities γ0∗ , −γd,∗ is positive, unless Ω is
straight, we get the existence of arbitrarily many bound states in asymptotic
regime, β → ∞. In particular, Theorem 2.2 implies the following claim:

                                            15
Theorem 4.1 Suppose that Ω is not straight and adopt the assumption (i’),
(ii)–(iv). Then to any positive N there is a βN such that for any β > βN
the Robin Laplacian Hβ on Ω has at least N eigenvalues with the following
asymptotic expansion,
                                                      
        λj (β) = −β 2 − max {γ ∗ , −γd,∗ } β + O β 2/3 , j = 1, . . . , N .


5    Bound states in the non-asymptotic regime
Let us return now to our basic example in which the boundary of Ω is a
single infinite curve. As indicated in the introduction, the results obtained
in Sec. 3 motivate us to ask about the existence of bound states beyond
the asymptotic regime. While the general setting is the same as before,
the spectral properties have now a more global character which forces us
to modify the assumptions made in Sec. 2. For the sake of simplicity we
are going to suppose here that the boundary Γ : R → R2 of Ω, assumed
again to be a C 4 -smooth curve, is straight outside a compact which opens
the following three possibilities:

  (i) Γ is a nontrivial local deformation of the straight line, i.e. there is a
      positive s0 such that Γ1 (s) = Γ1 (±s0 ) + (s ∓ s0 ) and Γ2 (s) = 0 holds
      for any ±s ≥ s0 ,
 (ii) Ω is a nontrivial local deformation of a wedge, i.e. there is an α ∈ (0, 21 π)
      and a positive s0 such that
                          
                            Γ1 (s) = Γ1 (±s0 ) + (s ∓ s0 ) cos α
                                                                            (5.22)
                            Γ2 (s) = Γ2 (±s0 ) ± (s ∓ s0 ) sin α

      holds for any ±s ≥ s0 ,
(iii) there is an α ∈ [− 12 π, 0) and an s0 > 0 such that relations (5.22) hold
      for any ±s ≥ s0 .

Note that the assumptions (i) and (iii) partly overlap, the latter with α = 0
covers also the situation when the ‘outer’ components of the boundary are
parallel but not necessarily parts of a single line.
   First we shall identify the essential spectrum of the corresponding Robin
Laplacian Hβ associated with the boundary-value problem (2.2) in a way
which strengthens under the present assumptions the claim of Lemma 3.6.

                                        16
Theorem 5.1 Any of the assumptions (i)–(iii) implies σess (Hβ ) = [−β 2 , ∞).

Proof: Let us check first that [−β 2 , ∞) ⊂ σess (Hβ ). To this end it is sufficient
to construct an appropriate Weyl sequence, i.e. to find fn ∈ D(Hβ ) =

 f ∈ H 2 (Ω) : ∂f
               ∂n
                  = βf such that kfn kL2 (Ω) = 1 and

          kHβ,0 fn − λfn kL2 (Ω) = k − (fn )xx − (fn )yy − λfn kL2 (Ω) → 0

holds as n → ∞ for any fixed λ ∈ [−β 2 , ∞). Spectral properties of Hβ are
certainly invariant w.r.t. the choice of the Cartesian system in the plane. We
thus rotate the domain Ω by the angle α clockwise; then Γ will contain the
segment [x0 , ∞) of the real axis for some number x0 . We choose a function
ϕ ∈ C0∞ such   that supp ϕ ⊂ (−1,21) and 2kϕkL2 (R) = 1 and define ϕn (x) :=
 1    x
√ ϕ
  n   n
         − n . Then supp ϕn ⊂ (n − n, n + n), the function is normalized,
kϕn kL2 (R) = 1, and the norms kϕ′n kL2 (R) and√ kϕ′′n kL2 (R) vanish as n → ∞.
                                   √               2
    Let us now put fn (x, y) = 2β ϕn (x) ei λ+β x e−βy . These functions are
for α ∈ [− 12 π, 41 π] and all n large enough normalized as needed,

                     Z                          nZ2 +nZ∞

    kfn k2L2 (Ω) =       |fn (x, y)|2 dxdy =               2β |ϕn (x)|2 e−2βy dxdy = 1 ,
                     Ω                         n2 −n 0

they belong to D(Hβ ) and satisfy
                             p               p                 √
                                                                         2
  (Hβ,0 fn − λfn ) (x, y) = − 2β ϕ′′n (x) + 2i λ + β 2 ϕ′n (x) ei λ+β x e−βy ,
                                                                           (5.23)
hence kHβ fn − λfn kL2 (R) → 0 holds as n → ∞. If α ∈ ( 14 π, 21 π) the argument
is no longer valid because the above trial function does not satisfy the correct
boundary conditions at the other part of the boundary being far away from
the origin a halfline of the angle θ := π − 2α ∈ (0, π/2) with respect to the
positive x-half-axis. In that case we choose a function g ∈ C0∞ (R) such that
                                 (
                                   1 ... 0 ≤ y ≤ 1
                          g(y) =
                                   0 ...      y≥2

and put                                                                       
                         p            √                                 3y
                                          2
              fn (x, y) = 2β ϕn (x) ei λ+β x e−βy g                   2
                                                                                   ;
                                                                     n tan θ

                                               17
it is obvious that these function satisfy trivially the boundary conditions for
all n large enough. Their squared norms are given by
                    Z 2 βn2 tan θ        Z 4 βn2 tan θ                 2
           2
                       3
                                   −u
                                            3
                                                        −u      3u
      kfn kL2 (Ω) =               e du +               e g      2 tan θ
                                                                           du ,
                     0                    2
                                          3
                                            βn 2 tan θ       2βn
                              2   2      
hence kfn k2L2 (Ω) = 1 − O e− 3 βn tan θ . On the other hand, the relation (5.23)
remains valid for 0 ≤ y ≤ 13 n2 tan θ, while        for 31 n2 tan θ ≤ y ≤ 23 n2 tan θ its
right hand side is multiplied by g n2 3y    tan θ
                                                    and the additional terms
                                                                   
                    p                          3        ′      3y
                   − 2β ϕn (x) − 2β 2                 g
                                          n tan θ           n2 tan θ
                                                    √
                           9                3y                    2
                      + 4     2 g
                                    ′′
                                          2
                                                          ei λ+β x e−βy
                        n tan θ          n tan θ

have appear, and for y ≥ 23 n2 tan θ the expression vanishes; using these
observations it is again easy to construct an appropriate Weyl sequence.
This shows that [−β 2 , ∞) ⊂ σess (Hβ ).
    To prove the opposite inclusion, we use again a bracketing argument
dividing the domain Ω into smaller a union of subdomains Ωj , j = 1, 2, 3, 4,
and their boundaries in two different ways as shown in Figures 1 and 2,
the former referring to α ∈ (0, 12 π), the latter to α ∈ [− 12 π, 0]. Imposing
Neumann conditions at the added boundaries, we estimate Hβ from below,
                                    N
                                               NR
                                                          NR
                                                                
            Hβ,0 ≥ −∆N    R
                         Ω1   ⊕  −∆  Ω2    ⊕ −∆ Ω3   ⊕  −∆ Ω4     .     (5.24)

The spectrum of the operator at the right-hand side is, of course, the union
the component spectra. The domain Ω1 is compact and thus it does not con-
tribute to the essential spectrum, the domain Ω2 does but the corresponding
operator is positive by definition. Hence it is sufficient to find the essen-
tial spectrum for the domains Ω3 and Ω4 being semi-infinite strips or planar
quadrants depending on the sign of α; in the former case it is important that
the strip width b can be made as large as one wishes.
                                             strip
    Denote the corresponding operators by Tβ,b     and Tβquadrant ; the quadratic
forms associated with them are
                         Z Zb                                       Z
               strip                 2           2
                                                     
              qβ,b   =           |fx | + |fy |           dxdy − β       |f (x, 0)|2 dx
                         R   0                                      R


                                             18
                                                       N

                                             b
                                                               W   2
                                                                       N           b



          W                                       W    3
                                                                           W   4

                                     R                             P
                                                                                       R
           R
                                                   N           W   1   N

                                                               R

Robin boundary conditions
                                                 Robin-Neumann b.c.
               Figure 1: Domain splitting for α > 0.




                                                               W
          W                                                N
                                                                   2




                                                                                       N
                                         N
                                                               W   1

           R
                                     W   3
                                                               R                       W   4

                                                                                   R
                                         R


Robin boundary conditions
                                                 Robin-Neumann b.c.
               Figure 2: Domain splitting for α ≤ 0.




                                19
and                     Z Z                                        Z
           quadrant                 2           2
                                                    
          qβ,b      =           |fx | + |fy |           dxdy − β       |f (x, 0)|2 dx ,
                        R   R                                      R

respectively. Since the variables decouple the spectra are easily found, in
particular, we have σ(Tβquadrant ) = σess (Tβquadrant ) = [−β 2 , ∞). In the strip
case we have to find the spectral threshold of −∆NR    β,(0,b) , the Robin-Neumann
Laplacian on (0, b) with the domain
               
                 f ∈ H 2 (0, b) : −f ′ (0) = βf (0), f ′ (b) = 0 .

It is straightforward to check that inf σ(−∆NR            2
                                            β,(0,b) ) = −ζ , where ζ = ζ(b) > 0
is the solution to the equation
                                    ζ −β
                                         = e−2ζb .
                                    ζ +β
Since lim ζ(b) = β, we conclude that inf σess (Tβstrip ) > −β 2 − ε for an
      b→+∞
arbitrary ε > 0, which yields the desired result.                       
   On the other hand, for the existence of a discrete spectrum the sign of
the asymptotic bending angle α is important.

Theorem 5.2 The operator Hβ associated with the eigenvalue problem (2.2)
has a bound state under any of assumptions (i) or (ii).

Proof: We start with assumption (i). The quadratic form (2.3) associated
with the operator Hβ can be written as
                        ZZ                        Z
                                2       2
              qβ [f ] =    |fx | + |fy | dx dy − β |f |2 ds .
                            Ω                                      Γ

By the previous theorem the essential spectrum coincides with the interval
[−β 2 , +∞), hence it is sufficient to find a function f ∈ H 1 (Ω) such that

                        S[f ] := qβ [f ] + β 2 kf k2L2 (Ω0 ) < 0 .

To this aim we choose a function ψ1 ∈ C0∞ (R) with the properties similar to
that of g of the preceding proof, but two-sided,
                               (
                                  1 . . . 0 ≤ |x| ≤ 1
                      ψ1 (x) =
                                  0 ...      |x| ≥ 2

                                           20
and define
                                     fn (x, y) = ψn (x)e−βy ,           (5.25)
                          x             ′
where ψn (x) := ψ1     ; we note that kψn kL2 (R) → 0 and ψn (x) → 1 pointwise
                           n
as n → ∞. Then S[fn ] acquires the form
ZZ                             ZZ                          Z
            2 −2βy                         2 −2βy
      ′
    |ψn (x)| e     dx dy +2β 2
                                   |ψn (x)| e     dx dy −β |ψn (x)|2 e−2βy ds
Ω                                        Ω                               Γ
                                  RR                          R
and using Stokes formula,                −Fy (x, y) dxdy =        F (x, y) dx, we get
                                  Ω                          ∂Ω
                      ZZ                                 Z
                                  2 −2βy             β
              β   2
                           |ψn (x)| e        dx dy =         |ψn (x)|2 e−2βy dx
                                                     2
                      Ω                                  Γ

and, henceforth,
          ZZ                         Z
                     2 −2βy
 S[fn ] =      ′
             |ψn (x)| e     dx dy + β (Γ′1 (s) − 1) |ψn (Γ1 (s))|2 e−2βΓ2 (s) ds .
          Ω                                     R

In the limit n → ∞ the first term tends to zero and the second one to
                         Z
                       β (Γ′1 (s) − 1) e−2βΓ2 (s) ds .
                                 R

The integrand in this expression is non-positive, and since by assumption
the trivial case, Γ = R, is excluded, there is an interval where Γ′1 (s) < 1.
Consequently, S[fn ] < 0 holds for all large enough n.
   Under the assumption (ii) we proceed in a similar way, but we modify
the trial function (5.25) as follows,
                                     fn (x, y) = ψn (x)e−γy ,
where the constant γ will be specified later. Then the form value S[fn ] is
 ZZ                        Z  2                  
            2 −2γy             β + γ2 ′
       ′
    |ψn (x)| e     dx dy +              Γ1 (s) − β |ψn (Γ1 (s))|2 e−2γΓ2 (s) ds .
                                  2γ
 Ω                                   R

As before, the first summand tends to zero as n → ∞, while the limit of the
second one is         Z  2               
                         β + γ2 ′
                                Γ1 (s) − β e−2γΓ2 (s) ds .
                            2γ
                           R


                                                21
In the previous case we had Γ′1 = 1 outside a compact and we had to choose
γ = β to make the integral converge. Now any γ > 0 will do to the presence
of the exponential factor, recall that Γ2 (s) ∼ |s| sin α for large |s|. Choosing
again γ = β we get as in the previous case
                              Z
                  S[fn ] → β (Γ′1 (s) − 1) e−2βΓ2 (s) ds < 0
                              R

as n → ∞. This concludes the proof.                                            
    On the other hand the assumption (iii) alone does guarantee absence of a
discrete spectrum. If Γ is sufficiently smooth, we know from Theorem 2.2 that
a sufficient condition for its existence is a local positivity of the curvature.
One may conjecture that there is no discrete spectrum for γ ∗ ≤ 0. It appears
that this is indeed the case, not only asymptotically.

Theorem 5.3 In addition to (iii), assume that Ω is concave and its bound-
ary is C 2 -smooth, then σdisc (Hβ ) = ∅.

Proof: In view of Theorem 5.1 it is sufficient to check that Hβ + β 2 I ≥ 0.
By assumption we have γ(s) ≤ 0 for any s ∈ R, an furthermore, one can
introduce the curvilinear coordinated of Lemma 3.1 in the entire Ω, that
is, with u running through the interval (0, ∞) because the factor 1 − uγ(s)
vanishes nowhere. The quadratic form associated with Hβ is given by the
expression analogous to (3.9) with the last term missing. For our present
purpose, however, it is more suitable to regard the wave functions as elements
of the space L2 (R×(0, ∞), (1−uγ(s))dsdu), in other words to write the right-
hand side of (3.8) as ψ(s, u). The quadratic form in question can be then
rewritten as
             Z Z ∞"                       2                         2
                                                                      #
                         1      ∂ψ                         ∂ψ
    qβ [ψ] =                       (s, u) + (1 − uγ(s))       (s, u) ds du
              R 0   1 − uγ(s) ∂s                           ∂u
                 Z
             −β |ψ(s, 0)|2ds .
                 R

Note that the expression contains the curvature but not its derivatives, hence
the C 2 smoothness of the boundary is sufficient. The first term on the right-
hand side is obviously non-negative which allows us to estimate the form


                                       22
from below,
                   Z        (Z                                              )
                               ∞             2
                                   ∂ψ
        qβ [ψ] ≥       ds             (s, u) (1 − uγ(s)) du − β|ψ(s, 0)|2
                   R          0    ∂u

for any ψ ∈ L2 (R × (0, ∞), (1 − uγ(s))dsdu). Next we notice that for a fixed
s the expression in the curly bracket is nothing else than the zero angular
momentum part of the quadratic form associated with the Robin problem in
the exterior to the disc of radius −γ(s)−1 . In Example 6.4 below we will show
that the corresponding spectral threshold is larger than −β 2 the saturation
being reached in the case of an infinite radius when the disc becomes a
halpflane. Since γ(s) ≤ 0 holds for any s ∈ R, the claim is proved.          


6     Exterior of a compact
Let us turn now to the other situation mentioned in the introduction. We
consider a compact a simply connected ‘obstacle’ Ω the boundary of which is
a simple close C 4 -smooth curve Γ and ask about the problem in its exterior,
Ωc := R2 \ Ω̄, i.e.

                              − ∆f = λf           in Ωc ,
                                                                                (6.26)
                                   ∂f
                                      = βf       on ∂Ω = Γ.
                                   ∂n
As before the boundary can be parametrized by its arc length, its orientation
being chosen in such a way that the obstacle Ω lies to the left as one moves
                                                              ∂
along its perimeter in the positive direction. The normal ∂n    in (6.26) is the
                                c
outside one with respect to Ω . The associated quadratic form is
                                                  Z
                   qβ,ext [f ] = k∇f kL2 (Ωc ) − β |f (x)|2ds
                                      2
                                                                          (6.27)
                                                       Γ

being defined on with Dom(qβ,ext ) = H 1(Ωc ). It is closed and below bounded;
we denote by Hβ,ext the unique self-adjoint operator associated with this form
and by λj,ext its j-th eigenvalue, numbered in the ascending order with the
multiplicity taken into account. It is straightforward to see that σess (Hβ,ext ) =
[0, ∞); as before we are interested in the asymptotic behavior of the discrete

                                             23
spectrum as β → ∞. To state the result, we employ again a one-dimensional
comparison operator with the curvature-induced potential, this time

                               d2    1
                       S=−        2
                                    − γ 2 (s) in L2 (0, L) ,                   (6.28)
                               ds    4
L being the perimeter of Ω, with the domain
                 
      Dom(S) = f ∈ H 2(0, L) : f (0) = f (L), f ′ (0) = f ′ (L) .              (6.29)

The spectrum of S is purely discrete; we denote by µj the j-th eigenvalue of
S counted with the multiplicity, j ∈ N. The oriented curvature of Γ of the
boundary is denoted by γext (s) := Γ′1 (s)Γ′′2 (s) − Γ′2 (s)Γ′′1 (s), and furthermore,
                           ∗
we introduce the symbols γext  = max γext (s) and γ∗,ext = min γext (s).
                                     [0,L]                      [0,L]


Remark 6.1 There is no need to put the ‘ext’ label to the operator S be-
cause it is invariant with respect to the curvature sign and orientation choice.
On the other hand, attention has to be paid to the curvature. We define it
here in the way consistent with the convention of Sec. 2, so that it is positive
when the curve is turning left in the direction of the parametrization. We
prefer to label it to avoid a confusion when comparing the result to that
of Ref. [3]. In that paper the orientation of Γ is the same, clockwise, but
the curvature is defined with the opposite sign. This Rchoice together with
the opposite orientation of the normal means we have Γ γext (s) ds = 2π, in
particular, that the curvature is non-negative if the obstacle Ω is convex.

Theorem 6.2 Under the stated assumptions, to any fixed integer j there is
a βj > 0 such that the number of negative eigenvalues of Hβ,ext is not smaller
than j. The j-th eigenvalue behaves in the limit β → ∞ as
                                                          
                    λext,j (β) = −β 2 + γ∗,ext β + O β 2/3 ,

where the lower asymptotic bound can be be improved to
                                                           
                                γ∗,ext 2            log β
              λext,j (β) ≥ − β −           + µj + O           .
                                   2                    β

Remark 6.3 As in similar situations we get also an upper bound analogous
                                             ∗
to the last formula with γ∗,ext replaced by γext which is not of much use be-
cause the two squeeze to produce a true asymptotics only if Ω is a circular

                                             24
disc. What is more important, similarly as in Ref. [3] the assumption about
simple connectedness of the boundary was done for simplicity only. If Ω is a
finite family of obstacles which do not touch each other, we have the analo-
gous result with the asymptotics being determined by the external curvature
minimum taken over all obstacle ‘components’.

Proof: The method of Ref. [3], modified in the previous sections, applies
readily; we have just to change signs at appropriate places and sketch the
argument briefly. We employ the map Φext defined as
        (0, L) × (0, a) ∋ (s, u) 7→ (Γ1 (s) + uΓ′2 (s), Γ2 (s) − uΓ′1(s)) ∈ R2 .
It may not be defined for all a, of course, unless Ω is convex, but due to
the smoothness of the boundary it is a diffeomorphism for a small enough.
We choose such an a > 0 to be sufficiently small and denote by Σa,ext the
one-sided strip neighborhood of Γ of width a,
                             Σa,ext := Φext ((0, L) × (0, a)) ,
and as before we impose Dirichlet and Neumann conditions at the curve
Γa := Φext ((0, L) × {a}) which is at the same time the boundary of the
unbounded and simply connected domain Ωc \ Σa,ext =: Λa,ext . Denoting
    D/N
by qa,β,ext [f ] the corresponding quadratic forms supported on Σa,ext , and by
 D/N
La,β,ext , respectively, the associated self-adjoint operators, we have
                                                     D            D
                  LN            N
                   a,β,ext ⊕ (−∆Λa,ext ) ≤ Hβ,ext ≤ La,β,ext ⊕ (−∆Λa,ext )         (6.30)

in L2 (Ωc ) = L2 (Σa,ext ) ⊕ L2 (Λa,ext ), where the parts related to Λa,ext are pos-
itive and can be thus neglected. Next we pass to the curvilinear coordinated
using the formula analogous to (3.8) with the opposite sign in the denomi-
nator and rewrite the quadratic form supported by the strip neighborhood.
For any f ∈ H 1 (Σa,ext ) we also have ϕ ∈ H 1 ((0, L) × (0, a)) and the form
equals
    ZL Za "                   2          2
                                                      #
                1         ∂ϕ       ∂ϕ
                                +          + Vext |ϕ|2 (s, u) du ds
            (1 + uγ)2 ∂s           ∂u
    0   0
              ZL                         ZL
                     γ(s)            2            γ(s)
            −     β−        |ϕ(s, 0)| ds −                 |ϕ(s, a)|2ds ,          (6.31)
                      2                       2(1 + aγ(s))
              0                                  0


                                            25
where
                               γ 2 (s)      uγ ′′ (s)   5 u2 (γ ′ (s))2
        Vext (s, u) = −                 +             −                 .
                           4(1 + uγ(s))2 2(1 + uγ(s))3 4 (1 + uγ(s))4
In analogy with Lemma 3.2 we introduce operators in L2 ((0, L) × (0, a))
                         D/N
unitarily equivalent to La,β,ext and the quadratic forms associated with them.
To pass to estimating operators with separated variables, we introduce
                                           ′
            γ+ := max |γ(.)| ,            γ+ := max |γ ′ (.)| ,            ′′
                                                                          γ+  := max |γ ′′ (.)| ,
                     [0,L]                           [0,L]                            [0,L]
                                      2                            ′′
                               γ (s)                             aγ+
            Vext,+ (s) := −               +               ,
                            4(1 + aγ+ )2 2(1 − aγ+ )3
                               γ 2 (s)            ′′
                                                aγ+         5 a2 (γ+′ 2
                                                                     )
            Vext,− (s) := −             2
                                          −             3
                                                          −               .
                            4(1 − aγ+ )     2(1 − aγ+ )     4 (1 − aγ+ )4
We choose an a satisfying 0 < a < γ+ /2 and for vectors ϕ belonging to the
domains QD        N
         a and Qa defined in analogy with (3.11), respectively, we define

                                               ZaZL              2             ZaZL           2
           ebD                        −2                  ∂ϕ                           ∂ϕ
             a,β,ext [ϕ] = (1 − aγ+ )                                ds du +                      ds du
                                                          ∂s                           ∂u
                                               0 0                             0 0
                    ZaZL                                 ZL
                                         2            γ∗
               +           Vext,+ (s) |ϕ| ds du − β −         |ϕ(s, 0)|2 ds
                                                      2
                    0 0                                                        0

and
                                               ZaZL              2             ZaZL           2
           ebN                        −2                  ∂ϕ                           ∂ϕ
             a,β,ext [ϕ] = (1 + aγ+ )                                ds du +                      ds du
                                                          ∂s                           ∂u
                                               0 0                           0 0
                    ZaZL                                                  ZL
                                           2          γ∗ 
                +          Vext,− (s) |ϕ| ds du − β −                              |ϕ(s, 0)|2 ds
                                                      2
                    0 0                                                    0
                                      ZL
                            γ+
                    −                       |ϕ(s, a)|2 ds .
                        2(1 − aγ+ )
                                      0

The forms ebD               eN
            a,β,ext [ϕ] and ba,β,ext [ϕ] are similar to those denoted by the same
symbols in Ref. [3], the only difference is that γ ∗ is replaced with −γ∗ and γ∗

                                                     26
with −γ ∗ , this following the argument of the said paper we get, in particular,
the lower bound stated in Theorem 6.2.
   To get a better upper bound which allows to get the tow term asymptotic
expansion we employ again a variational estimate choosing trial functions of
the form
                                               
                             s − s∗ + (2j − 1)ε                  
             ϕ̂(s, u) := χ                        e−αu − e−2aα+uα ,
                                     2ε
where χ is again a fixed smooth function with the support in (0, 1) and s∗ is
the point in which the curvature reaches its minimum, γ(s∗ ) = γ∗ . For j = 1
we get
                                  ′ 2                                        
         bD
          a,β,ext [ϕ̂]    1 kχ kL2 (0,1)                 −1
                                                                     1     −2
                       ≤ 2                    1+O α           + Cε       +α
        kϕ̂k2L2 (0,L)    4ε kχk2L2 (0,1)                             2α
                                                              !
                                                       2
                                              (γ ∗ )                      
             +α2 1 + O αe−2aα + −                        + Cε 1 + O α−1
                                                  4
                                  
                          γ∗ + ε                         
             −2α β −                   1 + O αe−2aα
                              2
                          γ∗
and choosing α = β +          the right-hand side of the estimates becomes
                          2
                                                                
                                                                            γ∗ 2
                ′ 2
        1 kχ kL2 (0,1)                 −1
                                                      1      −2
                           1  +  O   β      + Cε          + β       −   β −
       4ε2 kχk2L2 (0,1)                              2β                      2
                                                         !
                            γ∗        − (γ∗ )2                        
                   +ε β −          +             + Cε 1 + O α−1 .
                             2              4

Optimizing with respect to ε by taking ε = β −1/3 , we get the inequality
                     bD
                      a,β [ϕ̂]
                                        γ∗ 2           
                        2
                                ≤ −  β −       + O β 2/3                (6.32)
                   kϕ̂kL2 (0,L)          2
which the result for the first eigenvalue. The argument for j ≥ 2 is the same,
as before we take into account that by construction the used trial functions
are mutually orthogonal.                                                    
    We have mentioned that the Dirichler-Neumann estimates used in the
proof squeeze only in the case when the curvature is constant. Let us look
at this situation more closely.

                                       27
Example 6.4 Suppose that the obstacle Ω is a disc of radius R, for defi-
niteness centered at the origin, hence
                                             1
                                γ(s) ≡ γ =                              (6.33)
                                             R
and the the comparison operator S is just a shifted Laplacian with periodic
boundary conditions and its eigenvalues µj can be expressed explicitly as
                                       2 !
                                  1     j
                       µj = − +               R−2 ,                   (6.34)
                                  4     2

where [y] denotes the maximum integer which less or equal to y. The rota-
tional symmetry makes it natural to employ polar coordinates,
                  
                    x = r cos θ
                                    r ≥ R , 0 ≤ θ < 2π .
                     y = r sin θ

Writing with an abuse of notation f (x, y) = vf (r, θ) we can cast the eigen-
value problem in question with λ = −k 2 into the form
                      2
                       ∂ f    1 ∂f     1 ∂2f
                     
                           +       +          = k 2 f,
                      ∂r 2    r ∂r    r 2 ∂θ2
                                                                       (6.35)
                     
                         ∂f
                     
                      −           = βf.
                          ∂r r=R
Solution to
          Pthe first equation   in (6.35) is conventionally sought in the form
                            imθ
f (r, θ) = m∈Z cm Km (kr)e . Moreover, operator commutes with the an-
                      ∂
gular momentum, −i ∂θ   with periodic boundary conditions, hence they have
common eigenspaces, and we can consider sequence {cm } with nonzero cm
corresponding to a single values of |m|; it goes without saying that the dis-
crete spectrum has multiplicity two except the eigenvalue corresponding to
m = 0 which is simple. The boundary condition in (6.35) can be then rewrit-
ten as
                             ′
                        kKm    (kR) + βKm (kR) = 0.                      (6.36)
for a fixed m ∈ Z. To find its solutions, let us change the variables to u = kR
and α = βR, in which case the condition (6.36) reads
                                     ′
                                   uKm (u)
                               −           = α.                         (6.37)
                                   Km (u)

                                      28
The function at the left-hand side is strictly increasing for u > 0, equal to
m at u = 0, hence (6.37) has a unique solution for any fixed m and α > m.
As α → +∞, so does u in (6.37), and using the well-known asymptotics of
modified Bessel functions, we find
                  ′
               uKm  (u)     1 4m2 − 1
           −            = u+ +        + O(u−2) ,                 u → +∞ .
                Km (u)      2   8u
In combination with the spectral condition (6.37) this yields
                         1 4m2 − 1
                u=α−       −       + O(α−2) ,                α → +∞ .
                         2   8α
This, in turn, implies the asymptotics for u2 , and returning to the original
variables β, k we find
                         2           
                        1             1
          2
      −k = − β −            + m −2
                                          R−2 + O(β −1 ) , β → +∞ .
                       2R             4
This agrees, of course, with the conclusion of Theorem 6.2 taking (6.33) and
(6.34) into account. At the time it shows that there is not much room for
improving the error term in the theorem, because it differs from the one in
this explicitly solvable example by the logarithmic factor only.
    The result also shows that eigenvalues corresponding to the exterior of a
disc are larger that −β 2 . This is true not only asymptotically, which is the
fact we have used in the proof of Theorem 5.3. Indeed, from (6.37) we can
derive that 0 < u < α, or equivalently, 0 < k < β and −β 2 < −k 2 . It follows
from the fact that
                            ′
                           Km (x)
                       −          > 1,        m ∈ N,x > 0.                       (6.38)
                           Km (x)
To check this inequality we use the relations [1, p.79, (26), p.82, (21)]
                                                         Z∞
                           ′
  Km−1 (x) + Km+1 (x) = −2Km (x) ,            Km (x) =        e−x cosh u cosh(mu)du .
                                                         0

Using the first of these formulæ we can rewrite (6.38) as Km+1 (x) − Km (x) >
Km (x) − Km−1 (x). The validity of the last relation, in turn, follows from the
positivity of the second derivative of Km (x) with respect to m, which can be
checked using the second one the above formulæ.

                                         29
Acknowledgments
The research has been supported by the projects “Support of inter-sectoral
mobility and quality enhancement of research teams at Czech Technical Uni-
versity in Prague”, CZ.1.07/2.3.00/30.0034, sponsored by European Social
Fund in the Czech Republic, and 14-06818S of the Czech Science Foundation.
A.M. expresses his gratitude to Stepan Manko and Satoshi Ohya for useful
discussions.


References
[1] A. Erdelyi, W. Magnus, F. Oberhettinger, F.G. Tricomi: Higher tran-
    scendental functions, Vol. I; based, in part, on the notes left by Harry
    Bateman, McGraw-Hill, New York-Toronto-Londo, 1953.

[2] P. Exner: Leaky quantum graphs: a review, in “Analysis on graphs and
    its applications”, Proc. Symp. Pure Math., vol. 77; Amer. Math. Soc.,
    Providence, R.I., 2008., pp. 523–564.

[3] P. Exner, A. Minakov, L. Parnovski: Asymptotic eigenvalue estimates
    for a Robin problem with a large parameter, Portugaliae Math. (2014),
    toappear; arXiv:1312.7293.

[4] P. Exner and P. Šeba: Bound states in curved quantum waveguides, J.
    Math. Phys. 30 (1989), 2574–2580.

[5] P. Exner, K. Yoshitomi: Band gap of the Schrödinger operator with a
    strong δ-interaction on a periodic curve, Ann. H. Poincaré 2 (2001),
    1139–1158.

[6] P. Exner, K. Yoshitomi: Asymptotics of eigenvalues of the Schrödinger
    operator with a strong δ-interaction on a loop, J. Geom. Phys. 41 (4)
    (2002), 344–358.

[7] J. Fröhlich, G.M. Graf, J. Walcher: On the extended nature of edge states
    of quantum Hall Hamiltonians, Ann. Henri Poincaré 1 (2000), 405–442.

[8] M. Jlek: Quantum waveguides with Robin boundary conditions, BSc the-
    sis, Czech Technical University 2006.


                                     30
[9] M. Levitin, L. Parnovski: On the principal eigenvalue of a Robin problem
    with a large parameter. Math. Nachr. 281 (2) (2008), 272–281.

[10] N. Macris, Ph.A. Martin, J. Pulé: On edge states in semi-infinite quan-
    tum Hall systems, J. Phys. A: Math. Gen. 32 (1999), 1985–1996.

[11] 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.

[12] M. Reed, B. Simon: Methods of Modern Mathematical Physics. IV. Anal-
    ysis of Operators, Academic Press, New York 1978.




                                     31
