Generic non-selfadjoint Zakharov-Shabat operators

arXiv:1204.5200v3 [math.SP] 18 Dec 2012

T. Kappeler ∗, P. Lohrmann †, P. Topalov ‡

December 19, 2012

Abstract
In this paper we develop tools to study families of non-selfadjoint operators L(ϕ), ϕ ∈
P , characterized by the property that the spectrum of L(ϕ) is (partially) simple. As a
case study we consider the Zakharov-Shabat operators L(ϕ) appearing in the Lax pair
of the focusing NLS on the circle. The main result says that the set of potentials ϕ
of Sobolev class H N , N ≥ 0, so that all small eigenvalues of L(ϕ) are simple, is path
connected and dense.

1 Introduction
In this paper we develop tools to study families of non-sefadjoint operators L(ϕ), depending
on a parameter ϕ. To fix ideas assume that the parameter space P is a subset of some real
Hilbert space and for any ϕ ∈ P , L(ϕ) has a discrete spectrum. Ideally, the spectrum
spec L(ϕ) is simple for any ϕ ∈ P , i.e. any eigenvalue of L(ϕ) has algebraic multiplicity
one, and spec L(ϕ) can then be represented, under appropriate regularity assumptions on
the parameter dependence of L(ϕ), ϕ ∈ P , by a family of eigenvalues (λj (ϕ))j∈J with
λj : P → C being real-analytic for any j ∈ J. However, typically, such a situation does not
hold and one is interested in tools to estimate the size of the subset

P ′ = {ϕ ∈ P | spec L(ϕ) simple}.

In particular, it is of interest to know if P ′ is open, dense, or connected. As an illustration

we recall the classical theorem of Neumann and Wigner [18] saying that within the space
P of all real symmetric n × n matrices, n ≥ 2, the ones with multiple eigenvalues form an
algebraic variety of codimensions two. In particular, the set P ′ of all real symmetric n × n
∗
Supported in part by the Swiss National Science Foundation
†
Supported in part by the Swiss National Science Foundation and the European Research Council under
FP7 “New connections between dynamical systems and Hamiltonian PDE with small divisor phenomena”
‡
Supported in part by NSF grant DMS-0901443

1
matrices with simple spectrum is path-wise connected and dense. See also [2, 3], [15], [6],
[4] for related results.
As a case study we consider in this paper the Zakharov-Shabat operators (ZS) appearing
in the Lax pair of the defocussing nonlinear Schrödinger equation (dNLS) and the focusing
one (fNLS). These operators are differential operators of first order of the form (x ∈ R, ∂x =
∂/∂x)    
1 0 0 ϕ1
L(ϕ) = i ∂ + (1.1)
0 −1 x ϕ2 0
where the potential ϕ = (ϕ1 , ϕ2 ) is in L2c = L2 (T, C) × L2 (T, C) and T = R/Z. In the case
of dNLS, ϕ is in the subspace L2r whereas in the case of fNLS, ϕ is in iL2r . Here L2r ⊆ L2c
denotes the real subspace

L2r = {ϕ = (ϕ1 , ϕ2 ) ∈ L2c | ϕ2 = ϕ1 }.

For ϕ ∈ L2c arbitrary we denote by specp L(ϕ) the spectrum of the operator L = L(ϕ) with
domain
1 1
domp L(ϕ) = {F ∈ Hloc × Hloc | F (1) = ±F (0)}.
As L(ϕ) has a compact resolvent specp L(ϕ) is discrete and each of its eigenvalues has finite
algebraic multiplicity. It is referred to as the periodic spectrum of L(ϕ) and its eigenvalues
as periodic eigenvalues of L(ϕ) – or by a slight abuse of terminology as periodic eigenvalues
of ϕ. First let us state the following rough estimate of the periodic eigenvalues of L(ϕ) –
cf. Lemma 1 in [16] as well as Theorem 4.1 and Theorem 4.2 in [17]. For the convenience
of the reader it is proved in Section 2.
Lemma 1.1. [Counting Lemma] For each potential in L2c there exist a neighborhood W ⊆
L2c and an integer R ∈ Z≥0 so that for any ϕ ∈ W , when counted with their algebraic
multiplicities, L(ϕ) has two periodic eigenvalues in each disk

Dn = {λ ∈ C | |λ − nπ| < π/4} (1.2)

with |n| > R and 4R + 2 eigenvalues in the disk

BR = {λ ∈ C | |λ| < Rπ + π/4}. (1.3)

There are no other periodic eigenvalues.

The Counting Lemma shows that given a potential ϕ in L2c , for any |n| > R with R
sufficiently large, the periodic eigenvalues of L(ϕ) come in pairs, located in the disjoint disks
Dn . In case they are equal, one gets an eigenvalue of geometric and algebraic multiplicity
two (cf. Section 2). For ϕ in L2r or iL2r one can say more. Let us first consider the case
ϕ ∈ L2r . Then L(ϕ) is self-adjoint and hence specp L(ϕ) real. It is well known that when

2
listed with their algebraic multiplicities, the periodic eigenvalues are given by two doubly
infinite real sequences, (λ+ − ± 2
n )n∈Z and (λn )n∈Z satisfying λn = nπ + ℓn and

· · · < λ− + − +
n ≤ λn < λn+1 ≤ λn+1 < · · · .

– see e.g. [9] for a proof. In particular, L(ϕ) has a multiple eigenvalue iff there exists n ∈ Z
with λ− + 2 − +
n = λn . The set Zn of potentials in Lr with λn = λn Sis a real-analytic submanifold
2
of codimension two. Hence, for any N ∈ Z≥0 the set Lr \ |n|≤N Zn is open, dense, and
S
connected in L2r . Furthermore, n∈Z Zn is dense in L2r .
For ϕ ∈ iL2r , the periodic spectrum of L(ϕ) is more complicated. If ϕ 6= 0, L(ϕ) is not
selfadjoint and hence its periodic spectrum is not necessarily real. Moreover, besides the
asymptotic properties provided by the Counting Lemma, the spectrum has a symmetry. For
any λ ∈ specp L(ϕ), its complex conjugate λ is also a periodic eigenvalue and its algebraic
and geometric multiplicities are the same as the ones of λ (cf. Section 2). In addition
any real eigenvalue has geometric multiplicity two and its algebraic multiplicity is even. No
further constraints are known for the 4R+2 periodic eigenvalues in the disk BR , given by the
Counting Lemma. It turns out that some of the feature of specp L(ϕ) are still comparable
to the ones in the case where the potential is in L2r . To describe them we introduce the
following notion.

Definition 1. We say that a potential ϕ ∈ iL2r is standard, if any real periodic eigenvalue
of L(ϕ) has algebraic multiplicity two and any periodic eigenvalue in C\R is simple.

Denote by Sp the set of all standard potentials in iL2r . Due to the Counting Lemma, the
property of being a standard potential involves only the 4R + 2 eigenvalues in BR . One can
show in a straightforward way that Sp is open in iL2r and contains the zero potential. To
state our main result we need to introduce some additional notation. For any N ∈ Z≥0 , let
HcN = H N (T, C) × H N (T, C) and iHrN = HcN ∩ iL2r where H N (T, C) denotes the Sobolev
space of functions f : T → C with distributional derivatives up to order N in L2 (T, C).
Note that H 0 (T, C) = L2 (T, C) and Hc0 = L2c .

Theorem 1.2. For any N ∈ Z≥0 , Sp ∩ iHrN is path-wise connected.

Remark 1.3. Concerning the proof of this theorem let us first point out that in contrast
to papers such as [2], the Hilbert spaces iHrN , N ≥ 0, considered in Theorem 1.2 are real.
Therefore one can not apply the standard arguments used to prove that the complement of
a proper algebraic variety in a complex Hilbert space is path-wise connected.

We begin by analyzing potentials with a multiple eigenvalue λ. It turns out that the
case where the geometric multiplicity of λ is equal to 1 and the one where it is 2 have
to be treated differently. In Section 3 we show by general arguments that for any given
ψ ∈ iL2r with a periodic eigenvalue λψ of L(ψ) of geometric multiplicity one and algebraic

3
multiplicity mp ≥ 2 there is a neighborhood W of ψ ∈ iL2r so that the set of potentials in W ,
having a periodic eigenvalue near λψ of algebraic multiplicity mp and geometric multiplicity
one, is contained in a real-analytic submanifold of codimension two (Theorem 3.2). A
corresponding result is proved for a potential ψ in iL2r admitting a periodic eigenvalue of
geometric multiplicity two (Theorem 3.3). The proof of Theorem 3.2 and Theorem 3.3 are
based on a Theorem formulated in general terms and proved in Appendix A, providing a
class of functionals which can be used to construct submanifolds with the properties stated
in Theorem 3.2 and Theorem 3.3. In Section 2 we describe the set-up used throughout the
paper and in Appendix B we illustrate our results for the constant potentials and show an
auxiliary result needed in the proof of Theorem 3.2.
It follows from the proof of Theorem 1.2 that for any N ∈ Z≥0 , Sp ∩ iHrN is dense in
N
iHr . However, there is a much easier way to prove this density result and it turns out
that a stronger result holds. First we need to introduce some more notation. For ϕ ∈ L2c ,
denote by specD L(ϕ) the Dirichlet spectrum of the operator L(ϕ), i.e., the spectrum of the
operator L(ϕ) with the domain

domD L(ϕ) = {f = (f1 , f2 ) ∈ H 1 ([0, 1], C)2 | f1 (0) = f2 (0), f1 (1) = f2 (1)}. (1.4)

The Dirichlet spectrum is discrete and each eigenvalue has finite algebraic multiplicity. Let

SD := {ϕ ∈ iL2r | specD L(ϕ) is simple}. (1.5)

Theorem 1.4. For any N ∈ Z≥0 , Sp ∩ iHrN and SD ∩ iHrN are open and dense in iHrN .

To prove the statement of Theorem 1.4 concerning density, we locally reduce the problem
to one for matrices and then use the discriminant to conclude the theorem.
The basis for the study of the geometry of the phase space of fNLS are the spectral
properties of L(ϕ). Such an analysis was initiated in [1] and later in more detail, taken
up in [16]. However, much remains to be discovered – see also [5]. In a forthcoming
paper we will use Theorem 1.2 to construct action and angle coordinates for the fNLS in a
neighborhood of a standard potential ϕ ∈ iL2r .

2 Set-up
In this section we introduce some more notations, recall several known results needed in
the sequel and establish some auxiliary results. We consider the ZS operator L(ϕ), defined
by (1.1), for ϕ = (ϕ1 , ϕ2 ) in L2c . For any λ ∈ C, let M = M (x, λ, ϕ) be the fundamental
2 × 2 matrix of the equation
L(ϕ)M = λM

4
satisfying the initial condition M (0, λ, ϕ) = Id2×2 ,
 
m1 m2
M= .
m3 m4
Further, we denote by M1 , M2 the first, respectively second column of M . The fundamental
solution M (x, λ, ϕ) is a continuous function on R × C × L2c and for any given x ∈ R, it is
analytic in λ, ϕ on C × L2c – see e.g. Section 1 in [9]. Moreover, the proof of Theorem 1.1
in [9] shows that the following stronger statement holds.
Lemma 2.1. The fundamental matrix M defines an analytic map

M : C × L2c → C([0, 2]), (λ, ϕ) 7→ M (·, λ, ϕ).

For ϕ = 0, the fundamental solution Eλ (x) := M (x, λ, 0) is given by the diagonal matrix
diag(e−iλx , eiλx ). In the sequel we denote by (·). the derivative with respect to λ.
Symmetry: The ZS operator has various symmetries – see e.g. [7]. In this paper, the follow-
ing one is used frequently. For any function f : R → C2 with components f1 , f2 introduce
the functions f˘, fˆ : R → C2 , given by

f˘ = (−f¯2 , f¯1 ) and fˆ = −(f¯2 , f¯1 ).

Note that for any ϕ ∈ L2c , one has ϕ = ϕ̂ iff ϕ ∈ iL2r and that If := f˘ is an anti-involution,
I2 f = −f .
Lemma 2.2. Assume that ϕ ∈ L2c , λ ∈ C, and f in Hloc 1 (R, C2 ) solves (L(ϕ) − λ)n f = 0

for some n ∈ Z≥1 . Then

(L(ϕ̂) − λ̄)n f˘ = 0.
Proof. Introduce the matrices
     
0 1 1 0 0 −1
P = , R= , J= .
1 0 0 −1 1 0
−1 2 2
A direct computation
  that P R = J, P R = −(P R) , P = Id, and R = Id. As
shows
0 ϕ1
L(ϕ) = iR∂x + it then follows that
ϕ2 0

P R(L(ϕ) − λ)(P R)−1 = L(ϕ̂) − λ̄

and hence
P R(L(ϕ) − λ)n (P R)−1 = (L(ϕ̂) − λ̄)n .
As f˘ = P Rf¯ one then concludes for any f ∈ Hloc
1 (R, C) satisfying (L(ϕ) − λ)n f = 0 that

(L(ϕ̂) − λ̄)n f˘ = 0 as claimed.

5
Periodic spectrum: By the definition of the fundamental solution M , any solution f of the
equation L(ϕ)f = λf is given by f (x) = M (x, λ)f (0). Hence, a complex number λ is a
periodic eigenvalue of L(ϕ) iff there exists a non zero solution of L(ϕ)f = λf with

f (1) = M (1, λ)f (0) = ±f (0).

It means that 1 or −1 is an eigenvalue of the Floquet matrix M (1, λ). Denote by ∆(λ) ≡
∆(λ, ϕ) the discriminant of L(ϕ),

∆(λ, ϕ) := m1 (1, λ, ϕ) + m4 (1, λ, ϕ),

i.e., the trace of the fundamental matrix M , evaluated at x = 1. In view of the Wronskian
identity, det M (1, λ) = 1, it then follows that λ is a periodic eigenvalue of L(ϕ) iff ∆(λ) =
±2. For later reference we record the following
Proposition 2.3. For any ϕ ∈ L2c , the periodic spectrum of L(ϕ) coincides as a set with
the zero set of the function

χp (λ) ≡ χp (λ, ϕ) = ∆2 (λ, ϕ) − 4.

The discriminant ∆ and hence the characteristic function χp are analytic on C × L2c .
Actually, more is true. We will see below that for any periodic eigenvalue λϕ of L(ϕ),
the algebraic multiplicity of λϕ coincides with the multiplicity of λϕ as a root of χp (·, ϕ).
Recall that the algebraic multiplicity of a periodic eigenvalue λ of L(ϕ), ϕ ∈ L2c , equals the
dimension of the root space Rλ (ϕ), defined as the following subspace of domp L(ϕ),

Rλ (ϕ)={f ∈ domp L(ϕ) | ∃ n ∈ N ∀1 ≤ k ≤ n, L(ϕ)k f ∈ domp L(ϕ), (λ − L(ϕ))n f = 0}.

First we give the following rough localization of the roots of χp – see Section 6 in [9].
Recall that the disks Dn and BR have been introduced in (1.2) respectively (1.3).
Lemma 2.4. For each potential in L2c there exist a neighborhood W in L2c and R ∈ Z≥0
such that for any ϕ ∈ W the entire function χp (·, ϕ) has exactly two roots in each disk Dn
with |n| > R, and 4R + 2 roots in the disk BR , counted with their multiplicities. There are
no other roots.
Lemma 2.4 leads to the following corollary. To formulate it, denote by kϕk1 the norm
of ϕ ∈ Hc1 ,
kϕk1 := (kϕk2 + k∂x ϕk2 )1/2 .
Corollary 2.5. For any ρ > 0 there exists R ≡ Rρ ≥ 1 so that for any ϕ ∈ Hc1 with
kϕk1 ≤ ρ, the entire function χp (·, ϕ) has exactly two roots in each disk Dn with |n| > R
and exactly 4R + 2 roots in the disk BR , counted with their multiplicities. There are no
other roots.

6
Proof. For any ϕ ∈ Hc1 , let Wϕ := W and Rϕ := R be as in the statement of Lemma 2.4.
By Rellich’s theorem, Hc1 is compactly embedded in L2c . Hence there exist finitely many
potentials (ϕj )j∈J in Hc1 with kϕj k1 ≤ ρ so that (Wϕj )j∈J covers the closed ball of radius
ρ in Hc1 centered at 0. Then R ≡ Rρ := maxj∈J Rϕj has the claimed properties.

We now prove that the algebraic multiplicity of a periodic eigenvalue equals its multi-
plicity as a root of the characteristic function χp . First we note that by functional calculus,
the algebraic multiplicity mp (λ) ≡ mp (λ, ϕ) of a periodic eigenvalue λ of L(ϕ) with ϕ ∈ L2c
is equal to the dimension of the subspace of domp L(ϕ), given by the image of the Riesz
projector Πλ (ϕ), Z
1
Πλ (ϕ) = (z − Lp (ϕ))−1 dz,
2πi ∂B(λ)
where Lp (ϕ) denotes the operator L(ϕ) with domain domp L(ϕ), B(λ) denotes the open
disk centered at λ with sufficiently small radius so that B(λ) ∩ specp L(ϕ) = {λ}, and the
circle ∂B(λ) is counterclockwise oriented. By Proposition 2.3, λ is a root of χp (·, ϕ). Denote
by mr (λ) the multiplicity of λ as a root of χp (·, ϕ).
Lemma 2.6. For any periodic eigenvalue λ of L(ϕ) with ϕ ∈ L2c , mr (λ) = mp (λ).
Proof. First, note that a direct computation shows that the statement of the Lemma holds
for the zero potential ϕ = 0. A simple perturbation argument involving Proposition 2.3,
Lemma 2.4, the argument principle, and the properties of the Riesz projector (see the
arguments below), then shows that the Lemma continues to hold in an open neighborhood
of zero in L2c .
Now, consider the general case. Take ϕ ∈ L2c . As {sϕ | 0 ≤ s ≤ 1} is compact in L2c
there exist a connected open neighborhood W of the line segment [0, ϕ] in L2c so that the
integer R ≥ 1 of Lemma 2.4 can be chosen independently of ψ ∈ W. First consider the
periodic eigenvalues in BR . For ψ ∈ W denote by ΠR (ψ) the Riesz projector
Z
1
ΠR (ψ) = (z − Lp (ψ))−1 dz.
2πi ∂BR
Note that by functional calculus

Image ΠR (ψ) = ⊕λ∈BR ∩specp L(ψ) Rλ (ψ) (2.1)

where Rλ (ϕ) is the root space corresponding to λ. Moreover, standard arguments show
that W → L(L2c , L2c ), ψ 7→ ΠR (ψ), is analytic. In particular, by the general properties of
the projection operators the dimension of Image ΠR (ψ) is independent on ψ ∈ W (see [10],
Chapter III, §3). Consider the operator,
Z
1
A(ψ) = z(z − Lp (ψ))−1 dz .
2πi ∂BR

7
One easily sees that W → L(L2c , L2c ), ψ 7→ A(ψ), is analytic. By functional calculus
Lp (ψ)|Image ΠR (ψ) = A(ψ)|Image ΠR (ψ) , and hence
   
det λ − Lp (ψ)|Image ΠR (ψ) = det λ − A(ψ)|Image ΠR (ψ) .

Hence the polynomial

 
Q(λ, ψ) := det λ − Lp (ψ)|Image ΠR (ψ)

is well defined, analytic in C × W, and has leading coefficient one. By (2.1), the roots of
Q(·, ψ) are precisely the periodic eigenvalues of L(ψ) in BR counted with their multiplicities.
On the other hand, define
Y
P (λ, ψ) := (λ − λ+ −
j )(λ − λj ).
|j|≤R

Note that P (λ, ψ) is a polynomial in λ of degree 4R + 2 with leading coefficient 1. By the

argument principle and the last statement of Proposition 2.3, P (λ, ψ) is analytic in C × W.
Hence, the coefficients of Q(·, ψ) and P (·, ψ) are analytic on W. As Q(·, ψ) = P (·, ψ) in an
open neighborhood of zero in L2c we get by analyticity that
Q(·, ψ) = P (·, ψ)
for any ψ ∈ W. In particular, the Lemma holds also for any λ ∈ BR ∩ specp L(ψ). The same
argument shows that the statement of the Lemma holds also for any λ ∈ Dn ∩ specp L(ψ),
|n| > R.
Now we are ready to prove Lemma 1.1 stated in the introduction.

Proof of Lemma 1.1. By Lemma 2.6, for any ϕ ∈ L2c , the roots of χp (·, ϕ) coincide with
the eigenvalues of Lp (ϕ), together with the corresponding multiplicities. Lemma 1.1 thus
follows from Lemma 2.4.

For potentials ϕ in iL2r , the results discussed so far lead to a convenient description of
the periodic spectrum of L(ϕ). To state it we introduce the following order of C. We say
that two complex numbers a, b are lexicographically ordered, a 4 b, if [Re(a) < Re(b)] or
[Re(a) = Re(b) and Im(a) ≤ Im(b)].
Proposition 2.7. For any ϕ ∈ iL2c , any real periodic eigenvalue of L(ϕ) has geometric
multiplicity two and even algebraic multiplicity. For any periodic eigenvalue λ of L(ϕ) in
C \ R, its complex conjugate λ̄ is also a periodic eigenvalue of L(ϕ) and has the same
algebraic and geometric multiplicity as λ. It then follows that the periodic eigenvalues
of L(ϕ), when counted with their algebraic multiplicities, are given by two doubly infinite
+
sequences (λ+ − − +
n )n∈Z and (λn )n∈Z where λn = λn and Im λn ≥ 0 for any n ∈ Z so that
(λ+
n )n∈Z is lexicographically ordered.

8
Proof. It follows from Lemma 2.2 that for any λ ∈ specp L(ϕ), its complex conjugate λ̄ is
in specp L(ϕ) as well and that λ and λ̄ have the same geometric and the same algebraic
multiplicities. In addition, it follows from Lemma 2.2 that the geometric multiplicity of
each real periodic eigenvalue is two. It remains to show that any real periodic eigenvalue
of L(ϕ) has even algebraic multiplicity. Denote by Rλ (ϕ) the root space of L(ϕ) − λ. By
Lemma 2.2, If = f˘ is a R-linear anti-involution, leaving the finite dimensional vector space
invariant. Hence I defines a complex structure on Rλ (ϕ) and hence dimR Rλ (ϕ) is even.
This means that the algebraic multiplicity of λ is even.

Finally, we state the following well-known result on the asymptotics of the roots of χp
– see e.g. Section 6 in [9].

Proposition 2.8. For any ϕ ∈ L2c , the set of roots of χp (·, ϕ), listed with multiplicities,
consists of a sequence of pairs λ− +
n (ϕ), λn (ϕ), n ∈ Z, of complex numbers satisfying

λ± 2
n (ϕ) = nπ + ℓn

locally uniformly in ϕ, i.e., the sequences (λ± 2

n (ϕ) − nπ)n∈Z are locally bounded in ℓ (Z, C).

Discriminant: Denote by ∆ ˙ the partial derivative of the discriminant ∆(λ, ϕ) with respect
˙
to λ. Then ∆(λ, ϕ) is analytic on C × L2c as well. The following properties of ∆ and ∆ ˙
are well known – see e.g. Section 6 in [9] as well as Proposition 2.7 above. To state them,
introduce
πn := nπ for n ∈ Z \ {0} and π0 := 1.

Proposition 2.9. Let ϕ be an arbitrary element in L2c .

(i) The function λ 7→ ∆2 (λ, ϕ) − 4 is entire and admits the product representation
Y (λ+ (ϕ) − λ) (λ− (ϕ) − λ)
n n
∆2 (λ, ϕ) − 4 = −4 .
πn2
n∈Z

˙
(ii) The function λ 7→ ∆(λ, ϕ) is entire and has countably many roots. They can be listed
when counted with their order in such a way that they are lexicographically ordered and
satisfy the asymptotic estimates
λ̇n = nπ + ℓ2n ,
˙
locally uniformly in ϕ. In addition, ∆(λ, ϕ) admits the product representation

Y λ̇n − λ
˙
∆(λ, ϕ) = 2 .
πn
n∈Z

9
(iii) For any ϕ ∈ iL2r and λ ∈ C,
¯
∆(λ̄, ϕ) = ∆(λ, ϕ) and ¯˙
∆(λ, ˙ λ̄, ϕ).
ϕ) = ∆(
˙ ϕ) is invariant under complex conjugation. In view of the
In particular, the zero set of ∆(·,
asymptotics stated in (ii), for n sufficiently large, λ̇n is real.

Dirichlet spectrum: Recall from the introduction that for ϕ ∈ L2c we denote by specD L(ϕ)
the Dirichlet spectrum of the operator L(ϕ), i.e., the spectrum of the operator L(ϕ) consid-
ered with domain (1.4). As L(ϕ), when viewed as an operator with domain domD (L) has
compact resolvent the Dirichlet spectrum is discrete. For any λ ∈ C and ϕ ∈ L2c , denote
 
m̀1 m̀2
M̀ := = M (1, λ, ϕ) .
m̀3 m̀4
Similarly as in the periodic case, one can show that the operator L(ϕ) with domain domD (L)
admits the entire function
m̀4 + m̀3 − m̀2 − m̀1
χD (λ, ϕ) :=
2i
as a characteristic function and that the following results hold.
Lemma 2.10. For an arbitrary potential in L2c there exist a neighborhood W in L2c and an
integer R ≥ 1 so that when counted with their algebraic multiplicity, for any ϕ ∈ W, there
is exactly one Dirichlet eigenvalue in each disk
Dn := {λ ∈ C | |λ − nπ| < π/4} |n| > R,
and there are exactly 2R + 1 Dirichlet eigenvalues in the disk
BR := {λ ∈ C | |λ| < Rπ + π/4} .
There are no other Dirichlet eigenvalues.
Proposition 2.11. (i) For any ϕ ∈ L2c , the Dirichlet eigenvalues (µn (ϕ))n∈Z of L(ϕ) can be
listed with their algebraic multiplicities in such a way that they are lexicographically ordered
and satisfy the asymptotic estimates
µn (ϕ) = nπ + ℓ2n ,
locally uniformly in ϕ. Moreover, χD (λ, ϕ) admits the product representation
Y µn − λ
χD (λ, ϕ) = − .
πn
n∈Z

(ii) For ϕ ∈ L2r , the Dirichlet eigenvalues are real and for any n ∈ Z
λ− +
n (ϕ) ≤ µn (ϕ) ≤ λn (ϕ).

10
 By Lemma 2.10 for |n| sufficiently large, the Dirichlet eigenvalue µn is simple. Moreover
one has
Lemma 2.12. (i) If for a given potential ϕ ∈ L2c , λ is a periodic eigenvalue of L(ϕ) of
geometric multiplicity 2, then λ is a Dirichlet eigenvalue of L(ϕ). (ii) If for a given potential
ϕ ∈ iL2r , λ is a real periodic eigenvalue of L(ϕ) then it is of geometric multiplicity two and
hence also a Dirichlet eigenvalue of L(ϕ).
Proof. (i) If λ is a periodic eigenvalue of L(ϕ) of geometric multiplicity two, then M1 , M2
and hence M1 + M2 satisfy periodic or anti-periodic boundary conditions. As

m1 (0) + m2 (0) = 1 = m3 (0) + m4 (0)

it then follows that λ is a Dirichlet eigenvalue. (ii) follows from (i) and Proposition 2.7.

L2 -gradients: Let F : V → C be an analytic function on an open set V in L2c . The L2 -

gradient ∂F of F at ψ ∈ V is an element in L2c such that for any h ∈ L2c

dψ F (h) = h∂F, hir

where dψ F denotes the differential of F at ψ and

Z 1


h∂F, hir := (∂1 F )(x)h1 (x) + (∂2 F )(x)h2 (x) dx .
0

Let λϕ be a periodic eigenvalue of L(ϕ), ϕ ∈ L2c , of geometric multiplicity one. Then

M̀ (λϕ , ϕ) 6= ± Id2×2 , and hence m̀2 (λϕ , ϕ) or m̀3 (λϕ , ϕ) is not equal to zero. The proof of
the following lemma can be found e.g. in Section 4 in [9] (cf. Lemma 2 in [16]). To state it
introduce the ∗ product, (f1 , f2 ) ∗ (g1 , g2 ) := (f2 g2 , f1 g1 ).
Lemma 2.13. Under the conditions listed above and if in addition m̀2 (λϕ , ϕ) 6= 0 one has

i∂∆ = m̀2 f ∗ f

where f is the eigenfunction of λϕ normalized so that

 
1
f (x) = M (x, λϕ , ϕ) with ζ := (ξ − m̀1 )/m̀2 ,
ζ

where ξ ∈ {±1} is the eigenvalue of M̀ (λϕ , ϕ). Similarly, if m̀3 (λϕ , ϕ) 6= 0, then at (λϕ , ϕ)

i∂∆ = −m̀3 f ∗ f

where  
ζ
f (x) = M (x, λϕ , ϕ) with ζ := (ξ − m̀4 )/m̀3 .
1

11
 Actually, the formulas for ∂∆ above can be obtained from the following formula for ∂ M̀
(see Section 3 in [9]).

Lemma 2.14. The L2 -gradient of the Floquet matrix M̀ ≡ M (1, λ, ϕ) is given by

 
−m̀1 M1 ∗ M2 + m̀2 M1 ∗ M1 −m̀1 M2 ∗ M2 + m̀2 M1 ∗ M2
i∂ M̀ = (2.2)
−m̀3 M1 ∗ M2 + m̀4 M1 ∗ M1 −m̀3 M2 ∗ M2 + m̀4 M1 ∗ M2

where M1 and M2 are the two column vectors of M and the elements of the matrix in
parentheses are column vectors.

3 Proof of Theorem 1.2

The aim of this section is to prove Theorem 1.2 saying that Sp ∩ iHrN is path connected
for any N ∈ Z≥0 . First we need to analyze multiple periodic eigenvalues of L(ϕ) locally in
iL2r . Recall that the characteristic functions for the Dirichlet and the periodic spectrum of
L(ϕ), ϕ ∈ L2c , denoted by χD and χp respectively, are given by

2iχD (λ, ϕ) = (m̀4 + m̀3 − m̀2 − m̀1 )|λ,ϕ and χp (λ, ϕ) = ((m̀1 + m̀4 )2 − 4) λ,ϕ
.

Assume that λ ∈ C is a periodic eigenvalue of L(ϕ) of geometric multiplicity two,

mg (λ, ϕ) = 2 .1

By Lemma 2.12 it then follows that λ is at the same time a Dirichlet eigenvalue, i.e.,

χD (λ) = 0, χp (λ) = 0, and ∂λ χp (λ) = 0 .

One can easily see that the following more general statement holds.

Lemma 3.1. Let ϕ ∈ L2c and let λ be a periodic eigenvalue of L(ϕ). Then mg (λ, ϕ) = 2 iff
M̀ (λ) ≡ M (1, λ) is diagonalizable or, equivalently, M̀ (λ) ∈ {± Id2×2 }.

A periodic eigenvalue of geometric multiplicity two, mg (λ, ϕ) = 2, is said to be non-

degenerate if the algebraic multiplicity of λ, when viewed as a periodic eigenvalue of L(ϕ), is
two, mp (λ, ϕ) = 2, and degenerate otherwise. Note that for the zero potential, any periodic
eigenvalue is of geometric multiplicity two and non-degenerate. More generally, by Lemma
2.12(ii) any real periodic eigenvalue of L(ϕ) with ϕ ∈ iL2r is of geometric multiplicity two.
It might be degenerate – see Corollary 6.7(iii) in Appendix B. Furthermore note that a
non-degenerate periodic eigenvalue of L(ϕ), ϕ ∈ iL2r , of geometric multiplicity two is not
1
In what follows, mg (λ) ≡ mg (λ, ϕ) denotes the geometric multiplicity of λ ∈ C as a periodic eigenvalue
of L(ϕ).

12
necessarily a simple Dirichlet eigenvalue. Indeed, by Corollary 6.7, forp
the constant potential
ϕa = (a, −ā), a ∈ C, and n ∈ Z≥1 with 0 < nπ < |a|, the points ±i + |a|2 − n2 π 2 are non-
degenerate periodic eigenvalues of geometric multiplicity two. As i Im(a) p is a Dirichlet
eigenvalue of L(ϕa ), the phasepof a can be chosen so that Im(a) equals +
|a| − n2 π 2 – see
2
2 2 2
Corollary 6.8. For such an a, i |a| − n π is a Dirichlet eigenvalue of algebraic multiplicity
+

two.
The first result concerns potentials ψ ∈ iL2r with the property that L(ψ) admits a
periodic eigenvalue λψ with mp (λψ ) ≥ 2 and mg (λψ ) = 1. In this case it is convenient to
distinguish between a periodic eigenvalue in the proper sense, characterized by ∆(λψ , ψ) =
2 and an anti-periodic eigenvalue, characterized by ∆(λψ , ψ) = −2. The corresponding
characteristic functions are
χ±p (λ, ψ) = ∆(λ, ψ) ∓ 2.

Note that χp (λ, ψ) = χ+ − ε

p (λ, ψ)χp (λ, ψ). Finally denote by D (λψ ) ⊆ C the open disk of
radius ε > 0 centered at λψ .
Theorem 3.2. Assume that for ψ ∈ iHrN , N ≥ 0, λψ is a periodic eigenvalue of L(ψ)
in the proper sense [alternatively, anti-periodic eigenvalue of L(ψ)] of algebraic multiplicity
m ≥ 2, and geometric multiplicity one. Then for any ε > 0 sufficiently small there exists
an open neighborhood V ⊆ iHrN of ψ such that the set

X := {ϕ ∈ V | ∃λ ∈ D ε (λψ ) with mp (λ, ϕ) = m and mg (λ, ϕ) = 1}

is contained in a real-analytic submanifold Y of iHrN of (real) codimension two, which is

closed in V. In addition, V can be chosen so that for any ϕ ∈ V, all periodic eigenvalues of
L(ϕ) in D ε (λψ ) have geometric multiplicity one.
Proof. First assume that N = 0. As the cases where λψ is a periodic eigenvalue in the
proper sense and where it is an anti-periodic eigenvalue can be treated in the same way
we concentrate on the first case only. First we remark that due to Proposition 2.7 one
has Im(λψ ) 6= 0. By the first part of Theorem 5.1 applied to the characteristic function
χ+p (λ, ϕ) = ∆(λ, ϕ) − 2, for any ε > 0 sufficiently small there exists an open neighborhood
V ⊆ iL2r of ψ so that for any ϕ ∈ V, L(ϕ) has m periodic eigenvalues λ1 (ϕ), . . . , λm (ϕ),
listed with their algebraic multiplicities, in the the open disk D ε = D ε (λψ ) and none on
the boundary ∂D ε . By the characterization of the geometric multiplicity of Lemma 3.1,
mg (λψ ) = 1 implies that either m̀2 (λψ , ψ) 6= 0 or m̀3 (λψ , ψ) 6= 0. Hence by shrinking V and
ǫ > 0 if necessary it follows that mg (λk (ϕ)) = 1 for any 1 ≤ k ≤ m and ϕ ∈ V.
In order to apply Theorem 5.1(i) we look for an analytic function F : C × L2c → C so
that X – after shrinking V, if necessary – is contained in the zero set of
m
X
Fχ+
p
: V → C, ϕ 7→ F (λj (ϕ), ϕ) .
j=1

13
For any q ≥ 1, take Fq (λ) = (λ − λψ )q . By Theorem 5.1(ii) applied to the pair (Fq , χ+
p ) one
concludes that
m
X
Eq : V → C, ϕ 7→ (λj (ϕ) − λψ )q
j=1

is analytic.2 Note that for any ϕ ∈ X,

Em (ϕ) = m(λϕ − λψ )m and E1 (ϕ) = m(λϕ − λψ )

where for ϕ ∈ X, λϕ denotes the unique periodic eigenvalue of L(ϕ) in D ε (λψ ). To obtain
a functional which vanishes on X we set

G : V → C, ϕ 7→ mm−1 Em (ϕ) − E1 (ϕ)m .

Note that G is analytic and 

GX = 0 . (3.1)
As ∂Fm = 0 and as Fm (λ) has a zero of order m at λ = λψ one concludes from Theo-
rem 5.1(ii) that at (λ, ϕ) = (λψ , ψ),

∂Em = a∂∆, a 6= 0 .

As m ≥ 2 and E1 (ψ) = 0 it follows that

 
 
∂(E1 (ϕ)m ) = mE1 (ϕ)m−1 ∂E1  =0
ϕ=ψ ϕ=ψ

and hence at ϕ = ψ
∂G = a∂∆, a 6= 0.
It remains to show that near ψ the zero set of G is a real-analytic submanifold of codimension
two. Clearly

GR : V → R, ϕ 7→ Re G(ϕ) and GI : V → R, ϕ 7→ Im G(ϕ) (3.2)

are two real-analytic functionals. In view of the implicit function theorem it then remains
to show that the differentials dψ GR and dψ GI as elements in L(iL2r , R) are R-linearly inde-
pendent. Recall that by assumption, λψ has geometric multiplicity one. Then m̀2 (λψ , ψ) or
m̀3 (λψ , ψ) is not equal to zero. Assume for simplicity that m̀2 (λψ , ψ) 6= 0. The case when
m̀3 (λψ , ψ) 6= 0 is treated in the same way. It follows from Lemma 2.13 that at (λψ , ψ)

i∂∆ = m̀2 f ∗ f
2
A function F : V → C, V ⊆ iL2r , is called analytic if it is the restriction to V = Vc ∩ iL2r of an analytic
function F̃ : Vc → C where Vc is an open set in L2c .

14
where f is the appropriately normalized 1-periodic eigenfunction of L(ψ) corresponding to
λψ . Summarizing the computations above, one has in the case where m̀2 (λψ ) 6= 0

∂G = −ia · m̀2 (λψ )f ∗ f, a 6= 0 . (3.3)

In view of Lemma 5.3 (iii) it is to show that the R-linear functionals in iL2r

ℓR (h) := Re(hf ∗ f, hir ) and ℓI (h) := Im(hf ∗ f, hir ) , h ∈ iL2r ,

are R-linearly independent at ψ (see the discussion before Lemma 5.3 in Appendix A). By
Lemma 5.3(iv) we know that ℓR and ℓI are R-linearly dependent iff there exists c ∈ C\{0}
so that
\
cf ∗ cf + cf ∗ cf = 0. (3.4)
Assume that (3.4) holds for some c 6= 0. It is convenient to introduce g := cf and s := ğ =
(−ḡ2 , ḡ1 ). Then equation (3.4) reads

(g12 , g22 ) = (s21 , s22 ) . (3.5)

By Lemma 2.2, s satisfies

L(ψ)s = λψ s .
Hence,

ig1′ + ψ1 g2 = λψ g1 and is′1 + ψ1 s2 = λψ s1 (3.6)

−ig2′ + ψ2 g1 = λψ g2 and − is′2 + ψ2 s1 = λψ s2 . (3.7)
1 (R, C2 ) is a non-zero 1-periodic solution of L(ψ)g = λ g we conclude
As g = (g1 , g2 ) ∈ Hloc ψ
that g(x) 6= 0 for any x ∈ R. This and the periodicity of g imply that there are only four
possible cases:
Case 1: There exists a non-empty finite interval (a, b) ⊆ R such that ∀x ∈ (a, b)

g1 (x)g2 (x) 6= 0, g1 (a)g2 (a) = 0, and g1 (b)g2 (b) = 0;

Case 2: There exists a non-empty finite interval (a, b) ⊆ R such that ∀x ∈ (a, b)

g1 (x) = 0 and g2 (x) 6= 0;

Case 3: There exists a non-empty finite interval (a, b) ⊆ R such that ∀x ∈ (a, b)

g2 (x) = 0 and g1 (x) 6= 0;

Case 4: ∀x ∈ R
g1 (x)g2 (x) 6= 0.

15
First, assume that Case 1 holds. It follows from (3.5) that on (a, b),

g1 = σ1 s1 and g2 = σ2 s2 (3.8)

where σ1 , σ2 ∈ {±1}. If σ1 = σ2 one obtains from (3.6) that Im(λψ )g1 = 0 on (a, b). As
Im(λψ ) 6= 0 we see that g1 = 0 on (a, b). This contradicts one of the assumptions in Case
1. Now, assume that (σ1 , σ2 ) = (1, −1). Summing up the two equations in (3.6) we get
that ig1′ = Re(λψ )g1 on (a, b), or g1 (x) = η1 e−i Re(λψ )x , with constant η1 6= 0. Similarly, one
gets from (3.7) that g2 (x) = η2 ei Re(λψ )x , η2 6= 0. This implies that g1 (x)g2 (x) = η1 η2 6= 0
on (a, b). By continuity, g1 (a)g2 (a) 6= 0, which contradicts again one of the assumptions in
Case 1. The case (σ1 , σ2 ) = (−1, 1) is treated in the same way. Hence, Case 1 does not
occur.
Now, assume that Case 2 holds. Then, it follows from (3.5) that on (a, b)

g1 = s1 = 0 and g2 = σs2

where σ ∈ {±1}. This together with (3.7) implies that Im(λψ )g2 = 0 on (a, b). As Im(λψ ) 6=
0 we see that g2 = 0 on (a, b). This contradicts one of the conditions in Case 2. In the same
way one treats Case 3.
Finally, consider Case 4. Arguing as in Case 1 we see that (3.8) holds and the only
possible cases are (σ1 , σ2 ) = (1, −1) and (σ1 , σ2 ) = (−1, 1). If (σ1 , σ2 ) = (1, −1) one
concludes from (3.6) that

ig1′ = Re(λψ )g1 and ψ1 g2 = i Im(λψ )g1 (3.9)

and from (3.7) that

ψ2 g1 = i Im(λψ )g2 and − ig2′ = Re(λψ )g2 . (3.10)
Hence
g1 (x) = η1 e−i Re(λψ )x , g2 (x) = η2 ei Re(λψ )x
with η1 , η2 in C\{0}. Solving (3.9)-(3.10) for ψ1 , ψ2 one then gets
g1 η1
ψ1 = i Im(λψ ) = i Im(λψ ) e−2i Re(λψ )x
g2 η2
and
g2 η2
ψ2 = i Im(λψ ) = i Im(λψ ) e2i Re(λψ )x .
g1 η1
As ψ ∈ iL2r and thus ψ 1 = −ψ2 one has η1 /η2 = eiα with α ∈ R, and as ψ is 1-periodic it
follows that Re(λψ ) = kπ for some k ∈ Z. Hence

ψ1 (x) = i Im(λψ )eiα e−2kπix and λψ = kπ + i Im(λψ ).

16
By Lemma 6.6, λψ = kπ + i Im(λψ ) has algebraic multiplicity one. This contradicts the
assumption mp (λψ ) = m ≥ 2. The case (σ1 , σ2 ) = (−1, 1) is treated in the same way
as the case (σ1 , σ2 ) = (1, −1). Altogether we have shown that ℓR and ℓI , and hence
the differentials dψ GR and dψ GI , are R-linearly independent. As GX = 0, the claimed
statement concerning X then follows from the implicit function theorem.
Finally, assume that N ≥ 1. Take ψ ∈ iHrN and note that the restrictions GR V∩iH N
r
and GI V∩iHrN
of the functionals (3.2) considered above are real-analytic. Moreover, for any
h ∈ iHrN ,

dψ (GR V∩iHrN
)(h) = h∂GR , hir and dψ (GI V∩iHrN
)(h) = h∂GI , hir .

Assume that there exist α, β ∈ R, (α, β) 6= 0, such that for any h ∈ iHrN , αh∂GR , hir +
βh∂GI , hir = 0. As iHrN is dense in iL2r we see by a continuity argument that the last
equality holds also for any h ∈ iL2r . As this contradicts to the result obtained in the
case N = 0 we get that the differentials dψ (GR V∩iH N ) and dψ (GI V∩iH N ) are R-linearly
r r
independent in L(iHrN , R). Then, arguing as in the case N = 0 we complete the proof of
Theorem 3.2.

The second result deals with potentials ψ ∈ iL2r with the property that L(ψ) admits a
periodic eigenvalue λψ with mg (λψ ) = 2. In this case, M̀ (λψ ) ∈ {±Id2×2 } and hence λψ is
at the same time a Dirichlet eigenvalue. Denote by mD (λψ ) the algebraic multiplicity of λψ
as Dirichlet eigenvalue.

Theorem 3.3. Assume that for ψ in iHrN , N ≥ 0, λψ is a periodic eigenvalue of L(ψ)

in the proper sense [alternatively, anti-periodic eigenvalue of L(ψ)] with mg (λψ ) = 2 and
mD (λψ ) = m ≥ 1. Then for any ε > 0 sufficiently small there exists an open neighborhood
V ⊆ iHrN of ψ such that the set

X := {ϕ ∈ V | ∃λ ∈ D ε (λψ ) with mg (λ, ϕ) = 2, mD (λ, ϕ) = m}

is contained in a real-analytic submanifold Y in iHrN of (real) codimension two, which is

closed in V.

Proof. As the case N ∈ Z≥1 is treated in the same way as N = 0 – see the proof of Theorem
3.2 above – we concentrate on the latter case only. Similarly, as the cases where λψ is a
periodic eigenvalue in the proper sense and where it is an anti-periodic eigenvalue can be
treated in the same way we concentrate on the first case only. It turns out that we have to
distinguish between two different cases. We begin with the case where m = mD (λψ ) ≥ 2.
Case 1: m ≥ 2. By  Theorem 5.1(i), applied to the characteristic function χD (λ, ϕ) =
i 
2 (m̀1 + m̀2 − m̀3 − m̀4 ) λ,ϕ , for any ε > 0 sufficiently small there exists an open neighborhood
V ⊆ iL2r of ψ so that for any ϕ ∈ V, L(ϕ) has m Dirichlet eigenvalues µ1 (ϕ), . . . , µm (ϕ),

17
listed with their algebraic multiplicities, in the open disk D ε ≡ D ε (λψ ) and none on the
boundary ∂D ε . Similarly as in the proof of Theorem 3.2, introduce the functional

G : V → C, ϕ 7→ mm−1 Em (ϕ) − E1 (ϕ)m (3.11)

where here, for any q ≥ 1,

m
X
Eq : V → C, ϕ 7→ (µj (ϕ) − λψ )q .
j=1

By Theorem 5.1(ii), applied to Fq (λ) := (λ − λψ )q and χD (λ, ϕ), one concludes that Eq is
analytic for any q ≥ 1. Note that for any ϕ ∈ X,

Em (ϕ) = m(µϕ − λψ )m and E1 (ϕ) = m(µϕ − λψ )

where for ϕ ∈ X, µϕ denotes the unique Dirichlet eigenvalue of L(ϕ) in D ε . It then follows
that 
GX = 0 .
As ∂Fm = 0 and as Fm (λ) has a zero of order m at λ = λψ one concludes from Theo-
rem 5.1(ii) and (3.11) that at (λ, ϕ) = (λψ , ψ),

∂G = a∂χD , a 6= 0 . (3.12)

By Lemma 2.14, the L2 -gradient of the Floquet matrix M̀ ≡ M (1, λ, ϕ) is given by

 
−m̀1 M1 ∗ M2 + m̀2 M1 ∗ M1 −m̀1 M2 ∗ M2 + m̀2 M1 ∗ M2
i∂ M̀ = (3.13)
−m̀3 M1 ∗ M2 + m̀4 M1 ∗ M1 −m̀3 M2 ∗ M2 + m̀4 M1 ∗ M2

where M1 and M2 are the two column vectors of M and the elements of the matrix in
parentheses are column vectors. Thus

2∂χD = i∂ m̀1 + i∂ m̀2 − i∂ m̀3 − i∂ m̀4

= (m̀2 − m̀4 )M1 ∗ M1 + (m̀3 − m̀1 )M2 ∗ M2
+ (m̀2 + m̀3 − m̀1 − m̀4 )M1 ∗ M2 . (3.14)

As at (λ, ϕ) = (λψ , ψ), M̀ = Id2×2 one gets

2∂χD = −M1 ∗ M1 − M2 ∗ M2 − 2M1 ∗ M2 . (3.15)

By Lemma 2.1, M (·, λψ , ψ) ∈ C([0, 2]). In particular, it can be evaluated at x = 0. One

thus obtains  
1
2∂χD (0, λψ , ψ) = − .
1

18
In view of (3.12),  
 a 1

∂G x=0 = − , a 6= 0 . (3.16)
2 1
In addition to G we need to introduce a second functional, denoted by H,
m
X
H : V → C, ϕ 7→ m̀2 (µj (ϕ), ϕ).
j=1

Note that H X = 0. By Lemma 2.1, m̀2 : C × L2c → C, (λ, ϕ) 7→ m2 (1, λ, ϕ) is analytic. By
Theorem 5.1, applied to (F, χ) = (m̀2 , χD ) it follows that H is analytic and that at (λψ , ψ)
m
X
∂H = m ∂ m̀2 + aj ∂λm−j ∂χD .
j=0

As m̀2 (λψ , ψ) = 0, a0 = 0 by Theorem 5.1 and one gets at (λψ , ψ)

m
X
∂H = m ∂ m̀2 + aj ∂λm−j ∂χD . (3.17)
j=1

Let us first discuss the term m ∂ m̀2 in more detail. By (3.13) one has

i ∂ m̀2 = −m̀1 M2 ∗ M2 + m̀2 M1 ∗ M2 .

As M̀ = Id2×2 at (λψ , ψ) one then gets

     
 1  1
i ∂ m̀2  =− and m ∂ m̀2  = im . (3.18)
x=0 0 x=0 0
Next, let us turn to the second term of the right hand side of formula (3.17). It follows
from (3.14) and Lemma 2.1 that

C → C([0, 2]), λ 7→ ∂χD (·, λ, ψ)

is analytic. This implies that



∂λk ∂χD (·, λ, ψ) = ∂λk (∂χD (0, λ, ψ)) .
x=0

For any λ ∈ C
   
0 1
2∂χD (0, λ, ψ) = (m̀2 − m̀4 ) + (m̀3 − m̀1 )
1 0
   
1 0
= (m̀3 − m̀1 ) + (m̀1 + m̀2 − m̀3 − m̀4 )
1 1

19
or    
1 0
2∂χD (0, λ, ψ) = (m̀3 − m̀1 ) − 2iχD (λ, ψ) . (3.19)
1 1
As by assumption, ∂λk χD (λψ , ψ) = 0 for any 0 ≤ k ≤ m − 1, it then follows from formula
(3.19) that  
m−j m−j 1
2∂λ (∂χD (0, λ, ψ)) λ=λ = ∂λ (m̀3 − m̀1 ) λ=λ
ψ ψ 1
for any 1 ≤ j ≤ m. When combined with (3.17) and (3.18) one has
   
 1 1

∂H x=0 = im +κ (3.20)
0 1

for some κ ∈ C.
Following the notation introduced in Appendix A, denote by ℓ = ℓG : iL2r → C the
R-linear functional induced by ∂G = (∂1 G, ∂2 G)
Z 1
ℓG (h) := h∂G, hir = (∂1 Gh1 + ∂2 Gh2 )dx
0

and let
ℓG G
R (h) := Re(h∂G, hir ) and ℓI (h) := Im(h∂G, hir ).

According to (5.4), one has for h ∈ iL2r

 D ∂G + ∂G
c E

∂s  Re G(ϕ + sh) = ℓG
R (h) = ,h
s=0 2 r

and similarly
 D ∂G − ∂G
c E

∂s  Im G(ϕ + sh) = ℓG
I (h) = ,h
s=0 2i r

where we recall that for f = (f1 , f2 ) ∈ L2c , fˆ is given by fˆ = −(f 2 , f 1 ). By formula (3.16),
for ϕ = ψ,    
1  1 1 i 1
c 
(∂G + ∂G) = (ā − a) = − Im(a) (3.21)
2 x=0 4 1 2 1
and     
1 c  1 1 i 1
(∂G − ∂G) = − (a + a) = Re(a) (3.22)
2i x=0 4i 1 2 1
where a 6= 0. Similarly, we define for H
D ∂H + ∂H
d E d E
D ∂H − ∂H
ℓH
R (h) = , h and ℓH
I (h) = ,h .
2 r 2i r

20
By formula (3.20), at ϕ = ψ,
 
1

1
d
∂H + ∂H = i m/2 + Im(κ) (3.23)
2 x=0 1

and    
1
m 1 1
d
∂H − ∂H = − i Re(κ) . (3.24)
2i x=0 2 −1 1
In view of the identities (3.21) - (3.24) introduce

F1 : V → R, ϕ 7→ Im H(ϕ)

and (
Re G(ϕ), Im(a) 6= 0
F2 : V → R, ϕ 7→ .
Im G(ϕ), Im(a) = 0
1
As a 6= 0, Im(a) = 0 implies that Re(a) 6= 0 and hence according to (3.22), 2i (∂G −


c 
∂G) = − 2i Re(a) 11 6= 0. Now define
x=0

Y = {ϕ ∈ V | F1 (ϕ) = 0, F2 (ϕ) = 0}.

 
By construction, GX = 0, H X = 0 and hence X ⊆ Y . By (3.21), (3.22), and (3.24), ∂F1
and ∂F2 are R-linearly independent at ϕ = ψ. By the implicit function theorem, it then
follows that after shrinking V, if necessary, X is contained in a real-analytic submanifold of
iL2r of codimension two. Hence the claimed result for X is established in Case 1.
Case 2: m = mD (λψ ) = 1 & mg (λψ ) = 2. By Theorem 5.1(i), applied to the char-
acteristic function χD , for any ε > 0 sufficiently small there exists an open neighborhood
V ⊆ iL2r of ψ so that for any ϕ ∈ V, L(ϕ) has precisely one Dirichlet eigenvalue, denoted by
µ(ϕ) in the open disk D ε = D ε (λψ ) and none on the boundary ∂D ε . As µ(ϕ) is simple, it
follows from the inverse function theorem that the mapping µ : V → C is analytic. In view
of Lemma 3.1,
X ⊆ {ϕ ∈ V | m̀2 (µ(ϕ), ϕ) = m̀3 (µ(ϕ), ϕ) = 0} . (3.25)
Consider the functionals,

H1 : V → C, ϕ 7→ m̀2 (µ(ϕ), ϕ)

and
H2 : V → C, ϕ 7→ m̀3 (µ(ϕ), ϕ) .
In view of Lemma 2.1, H1 and H2 are analytic, and by (3.25),

H1 X
= H2 X
= 0.

21
Next, we will compute the L2 -gradients of H1 and H2 at ϕ = ψ. By the chain rule, we have
that at ϕ = ψ
∂H1 = ∂λ m̀2 (λψ , ψ) ∂µ + ∂ m̀2 (λψ , ψ) (3.26)
and
∂H2 = ∂λ m̀3 (λψ , ψ) ∂µ + ∂ m̀3 (λψ , ψ) . (3.27)
Using the identity χD (µ(ϕ), ϕ) = 0 for ϕ ∈ V and that χ̇D (µ(ϕ), ϕ) 6= 0 by the assumed
simplicity of µ(ϕ) we obtain that
1
∂µ = − ∂χD
χ̇D

where ∂χD = ∂χD (λψ , ψ) and χ̇D = χ̇D (λψ , ψ). By (3.15),

−2∂χD = M1 ∗ M1 + M2 ∗ M2 + 2M1 ∗ M2 ,

and hence,
1

∂µ = M1 ∗ M1 + M2 ∗ M2 + 2M1 ∗ M2 . (3.28)
2χ̇D
In particular,  
1 1
∂µ x=0
= . (3.29)
2χ̇D 1
By Lemma 2.14,
∂ m̀2 = im̀1 M2 ∗ M2 − im̀2 M1 ∗ M2 (3.30)
and
∂ m̀3 = im̀3 M1 ∗ M2 − im̀4 M1 ∗ M1 . (3.31)
As at (λ, ϕ) = (λψ , ψ), M̀ = Id2×2 we get that
 
1
∂ m̀2 x=0
=i (3.32)
0

and  
0
∂ m̀3 x=0
= −i . (3.33)
1
Combining (3.26)-(3.33) we then obtain at ϕ = ψ


∂H1 = κ1 M1 ∗ M1 + M2 ∗ M2 + 2M1 ∗ M2 + iM2 ∗ M2 (3.34)


∂H2 = κ2 M1 ∗ M1 + M2 ∗ M2 + 2M1 ∗ M2 − iM1 ∗ M1 (3.35)

22
and    
1 1
∂H1 x=0
= κ1 +i
1 0
   
1 0
∂H2 x=0
= κ2 −i
1 1
where
κ1 := ∂λ m̀2 /2χ̇D and κ2 := ∂λ m̀3 /2χ̇D . (3.36)
Arguing as in Case 1 we compute
 
1

1
d1
∂H1 + ∂H = i 1/2 + Im(κ1 ) (3.37)
2 x=0 1
   
1
1 1 1
d
∂H1 − ∂H1 x=0 = − i Re(κ1 ) (3.38)
2i 2 −1 1
and  
1

1
d2
∂H2 + ∂H = i − 1/2 + Im(κ2 ) (3.39)
2 x=0 1
   
1
1 1 1
d2
∂H2 − ∂H = − i Re(κ2 ) . (3.40)
2i x=0 2 −1 1
For any analytic function F : V → C denote for simplicity
1
1

∂R F := c
∂F + ∂F and ∂I F := c .
∂F − ∂F
2 2i
We now show that
rankR {∂R H1 , ∂I H1 , ∂R H2 , ∂I H2 } ≥ 2 . (3.41)
Assume on the contrary that the rank above is one. Then it follows from (3.37)-(3.40) that
i i
κ1 = a − and κ2 = a + , where a ∈ R . (3.42)
2 2
This together with (3.34) and (3.35) imply that at ϕ = ψ

∂H1 − ∂H2 = −2iM1 ∗ M2 . (3.43)

As the rank in (3.41) is assumed to be one, we get from Lemma 3.4 below that

M1 ∗ M2 ≡ 0 .

It means that at ϕ = ψ, for any 0 ≤ x ≤ 1

m1 (x, λψ )m2 (x, λψ ) = 0 and m3 (x, λψ )m4 (x, λψ ) = 0. (3.44)

23
Multiplying the first row of L(ψ)M1 = λψ M1 by m4 and the second by m2 yields

im′1 m4 = λψ m1 m4 and − im′3 m2 = λψ m3 m2 .

Taking the difference of the two equations and using the Wronskian identity one then gets

i(m′1 m4 + m′3 m2 ) = λψ . (3.45)

In the same way one gets, after multiplying the first row of L(ϕ)M2 = λψ M2 by m3 and
the second by m1

im′2 m3 = λψ m2 m3 and − im′4 m1 = λψ m4 m1

leading to
i(m′2 m3 + m′4 m1 ) = −λψ . (3.46)
Adding (3.45) and (3.46) one obtains

∂x (m1 m4 + m2 m3 ) = 0

or, in view of the Wronskian identity,

∂x (m1 m4 ) = 0 .

As m1 m4 x=0 = 1 one therefore has

m1 (x, λψ )m4 (x, λψ ) = 1 ∀ 0 ≤ x ≤ 1.

This combined with (3.44) leads to

m2 (x, λψ ) = 0 and m3 (x, λψ ) = 0 ∀ 0 ≤ x ≤ 1.

Multiplying the first row of L(ψ)M2 = λψ M2 by m1 and using that m2 = 0 yields

0 = ψ1 m4 m1 = ψ1 .

As ψ is in iL2r one has ψ2 = −ψ 1 and hence

ψ = 0.

A simple computation (cf. Lemma 6.4) shows that

 −iλ 
e 0
M̀ (λ, ψ) = .
ψ=0 0 eiλ

24
Hence,
m̀2 (λ, 0) = m̀3 (λ, 0) ≡ 0 and χD (λ, 0) = sin λ
and by (3.36)
κ1 = κ2 = 0.
This contradicts (3.42). Therefore, (3.41) holds. By the implicit function theorem it then
follows that, after shrinking V if necessary, X is contained in a real-analytic submanifold in
iL2r of codimension two.

Lemma 3.4. If M1 ∗ M2 6≡ 0 then

[∂I H1 and ∂R (H1 − H2 )] or [∂I H1 and ∂I (H1 − H2 )]

are R-linearly independent.

Proof. As M1 ∗M2 6≡ 0 then in view of (3.43) ∂R (H1 − H2 ) 6≡ 0 or ∂I (H1 − H2 ) 6≡ 0. Assume

for example that ∂I (H1 − H2 ) 6≡ 0. Assume that

α∂I H1 + β∂I (H1 − H2 ) = 0

where (α, β) 6= 0, α, β ∈ R. Restricting the equality above at x = 0 and using that by

(3.37)-(3.40) and (3.42), ∂I (H1 − H2 ) x=0 = 0 and ∂I H1 x=0 6= 0, we obtain that α = 0.
Hence, β∂I (H1 − H2 ) ≡ 0. As ∂I (H1 − H1 ) 6≡ 0 we see that β = 0. This shows that ∂I H1
and ∂I (H1 − H2 ) are R-linearly independent. The case ∂R (H1 − H2 ) 6≡ 0 is considered in
the same way.

Theorem 3.2 and Theorem 3.3 are now used to prove Theorem 1.2 stated in the intro-
duction.

Proof of Theorem 1.2. As the case N ∈ Z≥1 is treated in the same way as N = 0 we
concentrate on the latter case only. Let ζ, ξ with ζ 6= ξ be arbitrary elements in Sp . It is
to show that there exists a continuous path γ ∗ : [0, 1] → Sp with γ ∗ (0) = ζ and γ ∗ (1) = ξ.
The path γ ∗ will be constructed by deforming the straight line ℓ, parametrized by

γ 0 : [0, 1] → iL2r , t 7→ (1 − t)ζ + tξ.

First let us observe that as the straight line ℓ is compact, Lemma 1.1 implies that there
exist a tubular neighborhood Uℓ of ℓ,

Uℓ := {ϕ ∈ iL2r | dist(ϕ, ℓ) < δ}

for some δ > 0 and an integer R > 0 so that for any ϕ ∈ Uℓ , the eigenvalues λ+ n and
− +
λn = λn of L(ϕ) with |n| > R are in the disk Dn whereas the 4R + 2 remaining eigenvalues

25
λ±n , |n| ≤ R, are contained in BR . In addition, in view of Lemma 2.10, we can ensure that
for any ϕ ∈ Uℓ and for any |n| > R, µn ∈ Dn , and the remaining 2R+1 Dirichlet eigenvalues
µn ∈ BR , |n| < R. The path γ 0 will be deformed within Uℓ . Note that for any |n| > R,
either λ+ − +
n and λn are both simple periodic eigenvalues or λn is a real periodic eigenvalue
+ +
with mg (λn ) = 2 and mp (λn ) = 2. Hence to verify that a potential ϕ ∈ Uℓ is standard it
suffices to study the eigenvalues λ± n with |n| ≤ R.
As by Theorem 1.4, Sp is open and the endpoints ζ, ξ of γ 0 are assumed to be in Sp
there exist open balls Vζ , Vξ in Sp ∩ Uℓ centered at ζ respectively ξ.
In a first step we apply Proposition 3.5, based on Theorem 3.3, to show that there exists
a path γ 1 : [0, 1] → Uℓ with γ 1 (0) ∈ Vζ and γ 1 (1) ∈ Vξ so that for any ϕ on γ 1 , no periodic
eigenvalue λ± n with |n| ≤ R has geometric multiplicity two. Note that the path γζ : [0, 1] →
Vζ [γξ : [0, 1] → Vξ ], connecting γζ (0) = ζ [γξ (0) = ξ] with γζ (1) = γ 1 (0) [γξ (1) = γ 1 (1)] by
a straight line is in Sp ∩ Uℓ .
Then we apply Proposition 3.7, based on Theorem 3.2, to show that γ 1 can be deformed
within Uℓ to a path γ 2 with the same end points as γ 1 so that for any ϕ on γ 2 , all its
periodic eigenvalues λ± 2
n with |n| ≤ R are simple. In particular, γ is contained in Sp . The
path γ is then defined by concatenating γζ , γ , and γξ , i.e., γ = γξ−1 ◦ γ 2 ◦ γζ .
∗ 2 −1 ∗

To describe our construction of γ 1 in more detail we first introduce some more notation.
Recall that for any Dirichlet eigenvalue µ of L(ϕ) with ϕ ∈ L2c , mD (µ) ≡ mD (µ, ϕ) denotes
its algebraic multiplicity. It is convenient to set mD (µ) = mD (µ, ϕ) = 0 for any µ in C
which is not a Dirichlet eigenvalue of L(ϕ). Note that for any ϕ ∈ Uℓ one has

mD (λ±
n (ϕ)) ≤ 2R + 1 ∀|n| ≤ R.

Furthermore introduce for any ϕ ∈ Uℓ

MϕD := max{mD (λ± ±

n (ϕ)) | |n| ≤ R; mg (λn (ϕ)) = 2}.

We point out that

0 ≤ MϕD ≤ 2R + 1 ∀ϕ ∈ Uℓ .
Finally, for any continuous path γ : [0, 1] → Uℓ set
D 

MγD := max{Mγ(t) 0 ≤ t ≤ 1} .

Note that MγD = 0 implies that for any ϕ ∈ γ there is no periodic eigenvalue λ± n with
|n| ≤ R and mg (λ± n ) = 2. If M D = 0, choose γ 1 to be γ 0 . On the other hand, if M D > 0,
γ0 γ0
then Proposition 3.5 says that there exists a continuous path γ̃ 0 : [0, 1] → Uℓ , connecting
Vζ with Vξ so that Mγ̃D0 < MγD0 . In particular, γ̃ 0 (0) ∈ Vζ and γ̃ 0 (1) ∈ Vξ . This procedure
is iterated till we get a continuous path γ 1 : [0, 1] → Uℓ connecting Vζ with Vξ so that
MγD1 = 0.

26
 To deform γ 1 to γ 2 we have to deal with potentials ϕ with multiple periodic eigenvalues
λ±
n (ϕ) of geometric multiplicity one, i.e., mp (λ± ±
n (ϕ)) ≥ 2 and mg (λn (ϕ)) = 1. To this end
introduce for any ϕ ∈ Uℓ

Mϕp = max{mp (λ±

n (ϕ)) | |n| ≤ R}.

p
As 1 ≤ mp (λ±n (ϕ)) ≤ 4R + 2 for any ϕ in Uℓ it follows that 1 ≤ Mϕ ≤ 4R + 2. Moreover,
by construction
Mγp1 (0) = 1 and Mγp1 (1) = 1.
Finally, for a continuous path γ : [0, 1] → Uℓ define
p
Mγp := max{Mγ(t) | 0 ≤ t ≤ 1}.

We now deform the path γ 1 . If Mγp1 = 1, then γ 1 is already a path in Sp and we set γ 2 := γ 1 .
On the other hand, if Mγp1 ≥ 2, Proposition 3.7 implies that there exists a continuous path
γ̃ 1 : [0, 1] → Uℓ from γ 1 (0) to γ 1 (1) so that Mγ̃p1 < Mγp1 and Mγ̃D1 = 0. This procedure is
iterated till we get a continuous path γ 2 : [0, 1] → Uℓ from γ 1 (0) to γ 1 (1) so that Mγp2 = 1.
Then γ 2 is a path inside Sp connecting γ 1 (0) with γ 1 (1).

It remains to prove the two propositions used in the proof of Theorem 1.2.

Proposition 3.5. Let γ : [0, 1] → Uℓ be a continuous path with standard potentials as end
points, i.e., ζ := γ(0), ξ := γ(1) ∈ Sp , and ζ 6= ξ. Denote by Vζ , Vξ open disjoint balls
in Sp ∩ Uℓ centered at ζ, respectively ξ. If MγD > 0 then there exists a continuous path
γ̃ : [0, 1] → Uℓ with γ̃(0) ∈ Vζ , γ̃(1) ∈ Vξ and Mγ̃D < MγD .

Proof. For any ϕ ∈ γ, denote by

λ1 (ϕ), . . . , λK (ϕ), K ≡ Kϕ ∈ Z≥0

the list of different periodic eigenvalues of L(ϕ) inside BR with mg (λk (ϕ)) = 2 for any
1 ≤ k ≤ K. If MϕD < MγD , then choose an open ball Wϕ ⊆ Uℓ centered at ϕ so that for any
ψ ∈ Wϕ ,
MψD ≤ MϕD (< MγD ) . (3.47)
The existence of such neighborhood follows easily from the second statement of Theorem
3.2 and Theorem 5.1 (i) applied with χ = χD . On the other hand, if MϕD = MγD , let

I ≡ Iϕ := {1 ≤ j ≤ K | mD (λj (ϕ)) = MγD }.

27
By Theorem 3.3, applied to (λj (ϕ), ϕ) for any j ∈ I, there exists a path connected neigh-
borhood Wϕ of ϕ in Uℓ and a union Zϕ = ∪j∈Iϕ Zϕj of submanifolds Zϕj of codimension two
which are closed in Wϕ so that

{ψ ∈ Wϕ | MψD = MγD } ⊆ Zϕ . (3.48)

By shrinking Wϕ , if necessary, we further can assume that

MψD ≤ MϕD (= MγD ) ∀ψ ∈ Wϕ . (3.49)

In addition,
 if ϕ is either ζ or ξ we assume that Wϕ ⊆ Vζ or Wϕ ⊆ Vξ respectively. As
{γ(t)0 ≤ t ≤ 1} is compact the cover (Wϕ )ϕ∈γ([0,1]) admits a finite subcover (Wϕ )ϕ∈Λ
where Λ ⊆ γ([0, 1]) is finite and contains ζ and ξ. We claim that there exists a sequence
Wi ≡ Wϕi , 1 ≤ i ≤ N ≡ Nγ with ϕi ∈ Λ so that Wi−1 ∩ Wi 6= ∅ for any 2 ≤ j ≤ Nγ and
ϕ1 = ζ or ϕN = ξ. Indeed, choose ϕ1 = ζ to begin with. Then γ −1 (W1 ) is an open subset
of [0, 1]. As Vζ ∩ Vξ = ∅ it follows that

t1 := sup γ −1 (W1 ) < 1.

As (Wϕ )ϕ∈Λ covers γ([0, 1]) there exists ϕ2 ∈ Λ with γ(t1 ) ∈ W2 . As W2 is open and
γ : [0, 1] → Uℓ is continuous it then follows that W1 ∩ W2 6= ∅. Continuing in this way one
obtains the sequence Wi with ϕi ∈ Λ, 1 ≤ i ≤ N ≡ Nγ so that ϕ1 = ζ, ϕN = ξ, and Wi−1 ∩
Wi 6= ∅ for any 2 ≤ i ≤ N . As for any 1 ≤ i ≤ N, Zi ≡ Zϕi is a finite union of submanifolds
of codimension two it then follows that for any 2 ≤ i ≤ N, (Wi−1 ∩ Wi )\(Zi−1 ∪ Zi ) 6= 0.
For any 2 ≤ i ≤ N , choose ηi ∈ (Wi−1 ∩ Wi )\(Zi−1 ∪ Zi ). By (3.47)-(3.49) one concludes
that for any 2 ≤ i ≤ N, MηDi < MγD and

ηi , ηi+1 ∈ Wi \Zi .

As Zi is a finite union of submanifolds of codimension two which are closed in Wi , Lemma 3.6
stated below applies repeatedly. Hence for any 2 ≤ i ≤ N − 1, there exists a continuous
path γi : [0, 1] → Wi \Zi such that γi (0) = ηi and γi (1) = ηi+1 . By (3.47)-(3.49) one has
MγDi < MγD . As η2 ∈ W1 ⊆ Vζ and ηN ∈ WN ⊆ Vξ it then follows that the concatena-
tion γ̃ of γ2 , . . . , γN −1 is a continuous curve γ̃ : [0, 1] → Uℓ with the properties listed in
Proposition 3.5.

Let us now state and prove the lemma referred to in the proof of Proposition 3.5.
Lemma 3.6. Let U be an open, path connected set in a Hilbert space E and let Z ⊆ U be
a closed smooth submanifold of codimension two. Then U \Z is open and path connected.
Proof. This lemma is well known. In fact, it can be proved following the line of arguments
used in Proposition 3.5 and by taking into account the following special case, where Z is a
linear subspace of E of codimension two and hence E\Z is obviously path connected.

28
Proposition 3.7. Let γ : [0, 1] → Uℓ be a continuous path with standard potentials as end
points so that MγD = 0. If Mγp ≥ 2, then there exists a continuous path γ̃ : [0, 1] → Uℓ with
the same end points as γ, Mγ̃D = 0, and Mγ̃p < Mγp .

Proof. The assumption MγD = 0 implies that Mγ(0) D = 0 and Mγ(1)D = 0. As γ(0) and
γ(1) are both standard potentials one concludes from Proposition 2.7 that all their periodic
p p
eigenvalues λ±
n with |n| ≤ R are simple. In particular, one has Mγ(0) = 1 and Mγ(1) = 1.
For any ϕ ∈ γ denote by

λ1 (ϕ), . . . , λK (ϕ), K ≡ Kϕ ∈ Z≥0

the list of different multiple periodic eigenvalues of L(ϕ) inside BR . By assumption,

mg (λi (ϕ)) = 1 for 1 ≤ i ≤ K. If Mϕp < Mγp , by the second statement of Theorem 3.2
and Theorem 5.1 (i) applied with χ = χp and χ = χD there exists a neighborhood Wϕ ⊆ Uℓ
of ϕ so that for any ψ ∈ Wϕ

Mψp ≤ Mϕp (< Mγp ) and MψD = 0.

On the other hand, if Mϕp = Mγp , let


I ≡ Iϕ := {1 ≤ j ≤ K mp (λj (ϕ)) = Mγp }.

By Theorem 3.2, applied to (ϕ, λj (ϕ)) for any j ∈ I, there exists an open ball Wϕ in Uℓ ,
centered at ϕ and a union Zϕ = ∪j∈Iϕ Zϕj of submanifolds Zϕj of codimension two which are
closed in Wϕ so that
{ψ ∈ Wϕ | Mψp = Mγp } ⊆ Zϕ .
By shrinking Wϕ , if necessary, we further can assume that for any ψ ∈ Wϕ

Mψp ≤ Mϕp and MψD = 0.

Then argue as in the proof of Proposition 3.5 to conclude that there is a continuous path
γ̃ : [0, 1] → Uℓ with the same end points as γ, Mγ̃D = 0, and Mγ̃p < Mγp .

It turns out that the proof of Theorem 1.2 actually leads to the following additional
result. We say that a potential ϕ ∈ L2c is R-simple, R ∈ Z≥−1 , if µn , λ± n ∈ Dn for any
|n| > R, µn , λ±
n ∈ B R for any |n| ≤ R, and the eigenvalues (λ + , λ− )
n n |n|≤R are all simple.
R
Note that the zero potential is (−1)-simple. Denote by T the set of R-simple potentials
in L2c and by T the set of potentials ϕ ∈ L2c so that specp L(ϕ) is simple.
Inspecting the proof of Theorem 1.2 one sees that at the same time, the following result
has been proved.
Corollary 3.8. For any N ∈ Z≥0 and for any ζ, ξ ∈ T ∩ iHrN there exists R ∈ Z≥−1 such
that ζ and ξ are path connected in T R ∩ iHrN .

29
Proof. First, consider the case N = 0. Take ζ, ξ ∈ T ∩ iL2r and denote by ℓ the straight line
connecting ζ with ξ in iL2r . By the counting lemmas and the fact that ℓ is compact, there
exists R ∈ Z≥−1 so that for any ϕ ∈ ℓ, µn , λ± ±
n ∈ Dn for any |n| > R, and µn , λn ∈ BR , for
any |n| ≤ R. With this choice of R, one can follow the arguments of the proof of Theorem
1.2 to conclude that there exists a path γ in T R ∩iL2r connecting ζ and ξ. The case N ∈ Z≥1
is treated similarly.

For later applications it is useful to consider also a stronger version of the notion of R-
simple potentials. For any ψ ∈ iL2r denote by Iso0 (ψ) the connected component containing
ψ of the isospectral set,

Iso(ψ) := ϕ ∈ iL2r | specp L(ϕ) = specp L(ψ) .3

We say that a potential ψ ∈ iL2r is uniformly R-simple, R ∈ Z≥−1 , if Iso0 (ψ) ⊆ T R . In other
words, we require that for S any ϕ ∈ Iso0 (ψ), the periodic and the Dirichlet eigenvalues of
L(ϕ) are contained in BR ∪ |n|>R Dn and, when counted with their algebraic multiplicities,
satisfy the following conditions:

(S1) #(Dn ∩ specp L(ϕ)) = 2, #(Dn ∩ specD L(ϕ))) = 1 ∀|n| > R;

(S2) #(BR ∩ specp L(ϕ)) = 4R + 2, #(BR ∩ specD L(ϕ)) = 2R + 1;

(S3) The ball BR contains only simple periodic eigenvalues.

S
We denote the set of uniformly R-simple potentials by U R and let U ∗ := R≥−1 U R . In
view of the Counting Lemmas (see Lemma 1.1 and Lemma 2.10) and the compactness of
Iso0 (ψ) (see Lemma 3.11 below) it follows that for any N ∈ Z≥0 ,

T ∩ iHrN ⊆ U ∗ ∩ iHrN . (3.50)

Proposition 3.9. For any N ∈ Z≥0 , the set T ∩ iHrN is dense in iHrN . As a consequence,
U ∗ ∩ iHrN is dense in iHrN .

Proof. Let N ∈ Z≥0 and let ψ ∈ iHrN . In view of the Counting Lemmas there exists
R ∈ Z≥−1 and an open neighborhood U (ψ) of ψ in iHrN such that any ϕ ∈ U (ψ) satisfies
conditions (S1) and (S2). It follows from Theorem 3.2 and Theorem 3.3 that T R ∩ U (ψ)
is open and dense in U (ψ). In view of Proposition 2.7, Lemma 2.12, and Theorem 3.3, for
any |n| > R, the set

Zn := ϕ ∈ U (ψ) λ− +
n (ϕ) = λn (ϕ) ⊆ U (ψ)
3
The periodic eigenvalues are counted with their algebraic multiplicities.

30
is contained in a submanifold in U (ψ) of (real) codimension two. Hence, for any |n| > R,
Zn is closed and nowhere dense in T R ∩ U (ψ), and by the Baire theorem the set
  [
T ∩ U (ψ) = T R ∩ U (ψ) \ Zn
|n|>R

is dense in U (ψ). This completes the proof of the first statement of Proposition 3.9. The
second statement then follows from (3.50).

We finish this appendix by showing that U ∗ ∩ iHrN , N ≥ 1, contains a subset which is

open in iHrN . Here we use that for N ≥ 1, the Counting Lemmas for the periodic and the
Dirichlet spectrum of L(ϕ) with ϕ in iHrN hold uniformly on any bounded subset of iHrN .

Proposition 3.10. ∀N ∈ Z≥1 , U ∗ ∩ iHrN contains a subset which is open and dense in
iHrN .

We first need to make some preliminary R 1considerations. It follows

R 1 from [9, Theorem
13.4 and 13.5] that the quantities J1 (ϕ) := 0 |ϕ1 |2 dx and J2 (ϕ) := 0 (|∂x ϕ1 |2 − |ϕ1 |4 ) dx
are spectral invariants of the periodic spectrum of L(ϕ) for ϕ ∈ iHr1 . By the generalized
Gagliardo-Nirenberg inequality there exist absolute constants C1 , C2 > 0 so that for any
u ∈ L2 (T, C) and ε > 0


kukL4 ≤ C1 k∂x uk1/4 kuk3/4 + C2 kuk ≤ C1 ε2 k∂x uk1/2 + kuk3/2 /ε2 + C2 kuk

– see [19, Theorem 1] with n = 1 (for the case of a circle instead of an interval), j = 0, p =
4, m = 1, r = 2, q = 2, and a = 1/4. By taking ε2 = 1/(3C1 ) in the inequality above one
obtains
1
kuk4L4 ≤ k∂x uk2 + 37 C18 kuk6 + 33 C24 kuk4 .
3
R1 R1 R1
Writing 0 |∂x ϕ1 | dx = J2 + 0 |ϕ1 |4 dx it then follows that 32 0 |∂x ϕ1 |2 dx ≤ J2 + C(J13 +
2

J12 ) where C = max(37 C18 , 33 C24 ). We thus have proved that any ϕ ∈ Iso0 (ψ) ∩ iHr1 , can be
bounded by



kϕ1 k2H 1 ≤ 3J2 + 3C(J13 + J12 ) + J1 ψ ≤ 3kψ1 k2H 1 + 3C kψ1 k6 + kψ1 k4 .

where kψ1 k2H 1 := k∂x ψ1 k2 + kψ1 k2 . This implies that for any ρ > 0 there exists a positive
constant Cρ > 0 such that for any ψ ∈ Bρ1 ,

Iso0 (ψ) ∩ iHr1 ⊆ BC

1
ρ
, (3.51)

where BρN := {ϕ ∈ iHrN | kϕkHcN < ρ}.

31
Proof of Proposition 3.10. Take N ∈ Z≥1 , ρ > 0, and consider the set
[
IρN := Iso0 (ϕ) ∩ iHrN .
ϕ∈BρN

In view of (3.51) we have the following sequence of continuous embedding

IρN ⊆ Iρ1 ⊆ BC
1
ρ
⊆ iL2r

where the last embedding is compact. Hence, IρN is a precompact set in iL2r and, by the
Counting Lemma, there exists R ≡ Rρ ≥ 0 such that any ϕ ∈ IρN satisfies conditions (S1)
and (S2). This implies that T R ∩ BρN ⊆ U ∗ ∩ iHrN . Hence,

T ∩ BρN ⊆ Pρ ⊆ U ∗ ∩ iHrN , (3.52)

where Pρ denotes the open subset T R ∩ BρN of iHrN . By Proposition 3.9, the set T ∩ BρN
is dense in BρN . This together with the first inclusion in (3.52) implies that Pρ is open
S
and dense in BρN . Therefore ρ∈Z≥1 Pρ is open and dense in iHrN and thus is a subset of
U ∗ ∩ iHrN with the claimed properties. 

We complete this section by proving the following result used above to establish (3.50).

Lemma 3.11. For any ψ in iL2r , Iso(ψ) is compact.

Proof. Let (ϕn )n≥1 be a sequence in Iso(ψ). By [9, Theorem 13.4] the L2 -norm is a spectral
invariant of the periodic spectrum of L(ϕ) for ϕ ∈ iL2r . Hence, kϕn k = kψk, for any n ≥ 1,
and therefore there exist ϕ ∈ iL2r and a weakly convergent subsequence, which we again
w
denote by (ϕn )n≥1 , such that ϕn → ϕ as n → ∞. By [9, Theorem 4.1], for any λ ∈ C the
map L2c → C, ϕ 7→ ∆(λ, ϕ), is compact and hence ∆(λ, ψ) = limn→∞ ∆(λ, ϕn ) = ∆(λ, ϕ).
By the definition of Iso(ψ) it then follows that ϕ ∈ Iso(ψ), implying that kϕk = kψk. As a
consequence ϕn → ϕ in iL2r . This shows that Iso(ψ) is a compact subset of iL2r .

4 Proof of Theorem 1.4

The aim of this section is to prove Theorem 1.4.

Proof of Theorem 1.4. We begin by showing that Sp ∩ iHrN and SD ∩ iHrN are open. First
consider the case Sp ∩ iHr0 = Sp . For ψ ∈ Sp arbitrary choose R ∈ Z≥0 as in Lemma 1.1 and
let ε > 0 be smaller than twice the distance between any two different periodic eigenvalues
of L(ψ) in BR . By Lemma 1.1, Proposition 2.3, and Theorem 5.1(i), there exists an open

32


neighborhood Wψ
of ψ in L2c so that for any ϕ ∈ Wψ , # Dn ∩ specp L(ϕ) = 2 and
# BR ∩ specp L(ϕ) = 4R + 2 and for any eigenvalue λ ∈ BR ∩ specp L(ψ),



# D ε (λ) ∩ specp L(ϕ) = # D ε (λ) ∩ specp L(ψ)

where D ε (λ) ⊆ C is the open disk of radius ε centered at λ. By the definition of Sp , one then
concludes that Wψ ∩ iL2r ⊆ Sp . The openness of Sp ∩ iHrN (N ≥ 1) and SD ∩ iHrN (N ≥ 0)
is proved in a similar fashion. Next we show that Sp is dense in iL2r . As iHr1 is dense in
iL2r it suffices to show that Sp ∩ iHr1 is dense in iHr1 . Actually we prove a slightly stronger
statement. Denote by Bρ1 the open ball of radius ρ > 0 in Hc1 , centered at 0. Choose
R ≡ Rρ ≥ 1 as in Corollary 2.5 and introduce

Sp,R := ϕ ∈ Sp | λ±n simple ∀ |n| ≤ R .

We claim that Sp,R ∩ Bρ1 is dense in Bρ1 ∩ iL2r . For any ϕ ∈ Bρ1 and R as above introduce
Y
Qp,R (λ) := (λ+ −
k − λ)(λk − λ).
|k|≤R

Then Qp,R is a polynomial in λ of degree 4R + 2 with coefficients depending analytically in

ϕ on Bρ1 . Indeed, any coefficient of the polynomial Qp,R is a symmetric polynomial in λ±k,
|k| ≤ R, and hence can be written as a polynomial in
X
sn := (λ+ n − n
k ) + (λk ) , 0 ≤ n ≤ 4R + 2.
|k|≤R

To see that each sn is analytic on Bρ1 , note that by the argument principle, sn is given by
Z
1 χ̇p (λ)
sn = λn dλ
2πi |λ|=π(R+1/4) χp (λ)

and hence analytic as χp (λ, ϕ) and χ̇p (λ, ϕ) are analytic on C × L2c and χp (λ, ϕ) does not
vanish for ϕ ∈ Bρ1 and λ in {λ ∈ C | |λ| = Rπ + π/4}. Denote by Dp,R the discriminant of
the polynomial Qp,R . Recall that the discriminant is the resultant of Qp,R and its derivative
∂λ Qp,R (λ). It is given by
 
a0 · · · a4R a4R+1 a4R+2 0 · · · 0
 .. .. .. .. .. .. 
 . . . . . . 
 
 0 · · · a0 a a a · · · a 
Dp,R = det  b0 · · · b4R b4R+1
1 2 3 4R+2 
 0 0 ··· 0  
 .. .. .. .. .. .. 
 . . . . . . 
0 ··· 0 b0 b1 b2 ··· b4R+1

33
where (a0 , a1 , ..., a4R+2 ) is the coefficient vector of Qp,R and repeated 4R + 1 times whereas
(b0 , b1 , ..., b4R+1 ) is the one of Q̇p,R , and repeated 4R+2 times. Note that Dp,R is an analytic
function on Bρ1 . Furthermore, it has the property that it vanishes at an element ϕ ∈ Bρ1 iff
Qp,R (·, ϕ) has at least one multiple zero – see e.g. [14]. In particular, we have

ϕ ∈ Bρ1 ∩ iHr1 | Dp,R (ϕ) 6= 0 = Sp,R ∩ Bρ1 .
To show that Sp,R ∩ Bρ1 is dense in Bρ1 ∩ iHr1 it thus suffices to show that Sp,R ∩ Bρ1 6= ∅
and that Dp,R is real valued on Bρ1 ∩ iHr1 . To see that Dp,R is real valued on Bρ1 ∩ iHr1
note that by Proposition 2.7 for any k ∈ Z and ϕ in Bρ1 ∩ iHr1 , λ− +
k (ϕ) = λ̄k (ϕ). Hence by
the definition of Qp,R , its coefficients are real valued and Dp,R is therefore real valued on
Bρ1 ∩ iHr1 . Finally, to see that Sp,R ∩ Bρ1 6= ∅ we use that elements in iHr1 near 0 can be
represented by Birkhoff coordinates. Indeed, by Theorem 1.1 of [11] and formula (3.8) of
[11], for any element ϕ of iHr1 near 0 with Birkhoff coordinates (xk , yk )k∈Z satisfying for
+ −
any given k ∈ Z, x2k + yk2 6= 0, one has λ Pk 6= λk . Thus by Theorem 1.1 of [11], any sequence
(xk , yk )k∈Z with values in iR × iR and k∈Z (1 + |k|)2 (|xk |2 + |yk |2 ) sufficiently small so that
x2k + yk2 6= 0 for any |k| ≤ R is an element in Sp,R ∩ Bρ1 . In the same way one shows that
Sp ∩ iHrN is dense in iHrN for any N ∈ Z≥0 .
The corresponding density result for SD is proved in a similar fashion. As iHr1 is dense
in iL2r , it suffices to show that SD ∩ iHr1 is dense in iHr1 . For any ρ > 0 choose R ≡ Rρ ≥ 1
so that for any ϕ ∈ Bρ1 the statement of Lemma 2.10 holds. Introduce

SD,R := ϕ ∈ iL2r | µn simple and µn 6= µ̄n ∀|n| ≤ R .
We claim that for any ρ > 0, SD,R ∩ Bρ1 is dense in Bρ1 ∩ iHr1 . Arguing as above one reduces
in a first step the proof of the density of SD,R ∩ Bρ1 in Bρ1 ∩ iHr1 to the proof of SD,R ∩ Bρ1 6= ∅.
Here Y
QD,R := (µk − λ)(µ̄k − λ)
|k|≤R

plays the role of Qp,R . Next we use P again that by Theorem 1.1 in [11], any sequence
(xk , yk )k∈Z with values in iR × iR and k∈Z (1 + |k|)2 (|xk |2 + |yk |2 ) sufficiently small, rep-
resents an element ϕ in Bρ1 close to 0. Together with Proposition 4.1 in [11] it follows that
 + −
if xk 6= 0 andP yk = 0, then λ− +
k 6= λk and µk ∈ λk , λk . Hence µk 6∈ R and µk 6= µ̄k . In
µ k , λ±
addition, for k∈Z (1 + |k|)2 (|xk |2 + |yk |2 ) sufficiently small, P k ∈ Dk for any k ∈ Z.
Hence, any sequence (xk , yk )k∈Z with values in iR × iR and k∈Z (1 + |k|)2 (|xk |2 + |yk |2 )
sufficiently small, so that yk = 0 for any |k| ≤ R, represents an element in SD,R ∩ iHr1 . The
case N ∈ Z≥1 is treated in a similar way.

Inspecting the proof of Theorem 1.4 one sees that actually the following result for the
set of potentials S ∗ introduced at the end of Section 3 has been proved.
Corollary 4.1. For any N ∈ Z≥0 , S ∗ ∩ iHrN is dense in iHrN .

34
5 Appendix A: L2 -gradients of averaging functions
In this appendix, in a quite general set-up, we state and prove a theorem on multiple roots
of characteristic functions applied in the proofs of Theorem 3.2 and Theorem 3.3.

Theorem 5.1. Let F, χ : C × L2c → C be analytic maps and ψ an arbitrary but fixed
element in L2c . Assume that at zψ ∈ C, χ(·, ψ) has a zero of order m ≥ 1. Then the
following statements hold:
(i) For any ε > 0 sufficiently small there exists an open neighborhood V ⊆ L2c of ψ
such that for any ϕ ∈ V, χ(·, ϕ) has exactly m roots z1 (ϕ), . . . , zm (ϕ), listed with their
multiplicities, in the open disk D ε ≡ D ε (zψ ) := {λ ∈ C | |λ − zψ | < ε} and no roots on the
boundary ∂D ε of D ε .
(ii) The functional Fχ : V → C, defined by
m
X
Fχ (ϕ) := F (zj (ϕ), ϕ),
j=1

is analytic and at (ϕ, λ) = (ψ, zψ )

m
X
∂Fχ = m ∂F + aj ∂λm−j ∂χ
j=0

where aj ∈ C, 0 ≤ j ≤ m, and ∂ denotes the L2 -gradient with respect to ϕ and ∂λ denotes

the derivative with respect to λ. If F (·, ψ) has a zero of order k ≥ 1 at zψ , then a0 = . . . =
ak−1 = 0; if k = m, then

1 m (λ − zψ )m+1 ∂λ χ(λ, ψ) 

am = − ∂λ F (λ, ψ)  6= 0.
m! χ(λ, ψ)2 λ=zψ

Remark 5.2. As χ : C × L2c → C is analytic it follows that for any ϕ ∈ L2c , ∂χ : C → L2c ,
λ 7→ ∂χ (ϕ,λ) is analytic and so is ∂λk ∂χ for any k ≥ 1.

Proof. (i) By the analyticity of χ(·, ψ) there exists ε > 0 so that χ(·, ψ) does not vanish on
D ε \ {zψ }. By the analyticity of χ it then follows that there exists a neighborhood V of ψ in
L2c so that for any ϕ ∈ V, χ(·, ϕ) does not vanish in a small tubular neighborhood of ∂D ε
in C. It then follows by the argument principle that for any ϕ ∈ V, χ(·, ϕ) has precisely m
zeros in D ε , when counted with their multiplicities.
(ii) Again by the argument principle, for any ϕ ∈ V one has
Z
1 χ̇(λ, ϕ)
Fχ (ϕ) = F (λ, ϕ) dλ (5.1)
2πi ∂Dε χ(λ, ϕ)

35
where χ̇ = ∂λ χ. Note that the integrand in (5.1) is analytic on ∂D ε × V, whence Fχ
is analytic on V. To compute its L2 -gradient ∂Fχ it is convenient to introduce for g =
(g1 , g2 ), h = (h1 , h2 ) ∈ L2c ,
Z 1
hg, hir = (g1 h1 + g2 h2 )dx.
0
Then, by the definition of the 2
L -gradient, one has at ϕ = ψ,
d 
h∂Fχ , hir =  Fχ (ψ + sh)
ds Z s=0
 
1 d χ̇(λ, ψ + sh)
= F (λ, ψ + sh) dλ.
2πi ∂Dε ds s=0 χ(λ, ψ + sh)
By the product rule one gets at ϕ = ψ
Z   
1 χ̇ 1 1
h∂Fχ , hir = h∂F, hir + F · h∂ χ̇, hir − 2 h∂χ, hir χ̇ dλ.
2πi ∂Dε χ χ χ
Hence ∂Fχ is given by
Z  
1 χ̇ 1 (λ − zψ )m F · 1 (λ − zψ )m+1 F
∂F + · (∂χ) − · χ̇∂χ dλ.
2πi ∂Dε χ (λ − zψ )m χ (λ − zψ )m+1 χ2
(5.2)
Here we used that ∂ χ̇ = (∂χ)· and that ∂F, ∂χ : C → L2c are analytic and hence in particular,
the maps ∂F, ∂χ, (∂χ)· : C → L2c are continuous. Furthermore, as by assumption, χ(·, ψ)
(λ−zψ )m F (λ−zψ )m+1 F χ̇
has a zero of order m at λ = zψ , χ (∂χ)· and χ2
∂χ are both analytic
functions on D ε with values in L2c . Hence by the argument principle, at ϕ = ψ,
Z 
1 χ̇ 
∂F dλ = m ∂F 
2πi ∂Dε χ λ=zψ

and by Cauchy’s integral formula,

Z
1 1 (λ − zψ )m F
(∂χ)· dλ
2πi ∂Dε (λ − zψ )m χ
  
1 m−1  (λ − zψ )m F ·
= ∂  (∂χ)
(m − 1)! λ λ=zψ χ
and
Z
1 1 (λ − zψ )m+1 F χ̇
m+1
∂χdλ
2πi∂D ε (λ − zψ ) χ2
 
1 m  (λ − zψ )m+1 F χ̇
= ∂  ∂χ .
m! λ λ=zψ χ2

36
Thus ∂Fχ at ϕ = ψ is given by
   
 1  (λ − zψ )m F
∂Fχ = m ∂F  + ∂λm−1  ∂λ ∂χ
λ=zψ (m − 1)! λ=zψ χ
  m+1 
1 m (λ − zψ ) F χ̇
− ∂λ  2
∂χ .
m! λ=zψ χ
The claimed formula for ∂Fχ at ϕ = ψ then follows from the Leibniz rule. If F (·, ψ) has a
zero of order k ≥ 1 at λ = zψ , then at (ϕ, λ) = (ψ, zψ )
m
X
∂Fχ = m ∂F + aj ∂λm−j ∂χ,
j=k

i.e., aj = 0 for 0 ≤ j ≤ k − 1. If k = m, then

∂Fχ = m ∂F + am ∂χ
where in this case
1 m (λ − zψ )m+1 χ̇(λ, ψ) 
am = − ∂λ F (λ, ψ) 6= 0.
m! χ(λ, ψ)2 λ=zψ

Finally we record a few simple facts from linear algebra, also needed in Section 3.
Consider f = (f1 , f2 ) in L2c and denote by ℓ ≡ ℓf the R-linear functional on the R-vector
space iL2r induced by f ,
ℓ : iL2r → C, h 7→ hf, hir ,
where Z 1
hf, hir = (f1 h1 + f2 h2 )dx.
0
Write ℓ(h) as ℓR (h) + iℓI (h) where ℓR ≡ ℓf,R and ℓI ≡ ℓf,I are the elements in the dual
L(iL2r , R) of iL2r given by
ℓR (h) = Re(hf, hir ) and ℓI (h) = Im(hf, hir ) . (5.3)
They can be expressed in terms of f and fˆ = −(f 2 , f 1 ) as follows
D f + fˆ E D f − fˆ E
ℓR (h) = , h and ℓI (h) = ,h . (5.4)
2 r 2i r

As the subspace iL2r ⊆ L2c is the subset of all elements ϕ ∈ L2c satisfying ϕ = ϕ̂ it follows
ˆ −fˆ
that f +2 f and f 2i are in iL2r .
Using that f 7→ fˆ is an involution and that for any c in C, (cfd) = cfˆ, the following
lemma can be proved in a straightforward way.

37
 f +fˆ f −fˆ
Lemma 5.3. (i) ℓR , ℓI are R-linearly independent iff 2 , 2i are R-linearly independent.
f +fˆ f −fˆ
(ii) ℓR , ℓI are R-linearly independent iff 2 , 2i are C-linearly independent.
(iii) For any λ ∈ C \ {0}, ℓf,R , ℓf,I are R-linearly independent iff ℓλf,R , ℓλf,I are R-
linearly independent.
(iv) ℓf,R , ℓf,I are R-linearly dependent iff there exists λ ∈ C\{0} so that

d)
λf + (λf
= 0.
2

6 Appendix B: Examples
In this section we consider potentials in iL2r of the form (a ∈ C, k ∈ Z)

ϕa,k (x) = (ae2πikx , −āe−2πikx ). (6.1)

Most of the results presented in this section can be found in [16]. We include them for
the convenience of the reader. First we show that we can easily relate various spectra of
L(ϕa,k ) with the corresponding ones for k = 0. More generally, for an arbitrary potential
ϕ ∈ L2c , various spectra of L(ϕ1 e2πikx , ϕ2 e−2πikx ) are related to the corresponding spectra
of (ϕ1 , ϕ2 ) by the following lemma which can be verified in a straightforward way.
Lemma 6.1. Assume that f = (f1 , f2 ) is a solution of L(ϕ)f = λf where ϕ ∈ L2c is
arbitrary. Then (f1 eiπkx , f2 e−iπkx ) is a solution of

L(ϕ1 e2πikx , ϕ2 e−2πikx )g = (λ − kπ)g.

Corollary 6.2. For any ϕ ∈ L2c and k ∈ Z, the fundamental solution

M̌ (x, λ) ≡ M (x, λ, (ϕ1 e2iπkx , ϕ2 e−2iπkx ))

of L(ϕ1 e2πikx , ϕ2 e−2πikx ) is related to the fundamental solution M (x, λ) of L(ϕ1 , ϕ2 ) by

M̌ = diag (eiπkx , e−iπkx ) · M (x, λ + kπ).

Corollary 6.2 yields the following application.

Proposition 6.3. For any ϕ = (ϕ1 , ϕ2 ) ∈ L2c and any k ∈ Z,

specp (L(ϕ1 e2πikx , ϕ2 e−2πikx )) = specp (L(ϕ1 , ϕ2 )) − kπ

and
specD (L(ϕ1 e2πikx , ϕ2 e−2πikx )) = specD (L(ϕ1 , ϕ2 )) − kπ
(with multiplicities).

38
Proof. Recall that the characteristic functions χp and χD are given by

χp (λ) = (m̀1 (λ) + m̀4 (λ))2 − 4

2iχD (λ) = m̀4 (λ) + m̀3 (λ) − m̀2 (λ) − m̀1 (λ).

By Corollary 6.2,
χp (λ, (ϕ1 e2πikx , ϕ2 e−2πikx )) = χp (λ + kπ, ϕ)
and
χD (λ, (ϕ1 e2πikx , ϕ2 e−2πikx )) = (−1)k χD (λ + kπ, ϕ).
As specp (L(ϕ1 e2πikx , ϕ2 e−2πikx )) and specD (L(ϕ1 e2πikx , ϕ2 e−2πikx )) are the zero sets (with
multiplicities) of χp (λ, (ϕ1 e2πikx , ϕ2 e−2πikx )) respectively χD (λ, (ϕ1 e2πikx , ϕ2 e−2πikx )), the
claimed identities follow.

In view of Proposition 6.3, instead of the potentials ϕa,k defined by (6.1), it suffices to
consider the case k = 0,
ϕa ≡ ϕa,0 = (a, −ā), a ∈ C.
In a straightforward way one verifies the following

Lemma 6.4. For any a ∈ C,

!
cos(κx) − iλ sin(κx)
κ ia sin(κx)
κ
M (x, λ, ϕa ) = (6.2)
iā sin(κx)
κ cos(κx) + iλ sin(κx)
κ

where p
κ ≡ κ(λ, a) = λ2 + |a|2 (6.3)

Remark 6.5. Note that κ depends only on the modulus |a| of a pand that the right hand
side of (6.2) does not depend on the choice of the sign of the root λ2 + |a|2 as cosine is an
even function whereas sine is odd. Furthermore, the right hand side of (6.2) is well defined
at κ = 0 as sin(κx)
κ = x + O(κ2 ) .

Periodic spectrum of L(ϕa ): By Lemma 6.4 one has ∆(λ, ϕa ) = 2 cos κ(λ) and hence the
characteristic function of L(ϕa ), considered on the interval [0, 2] with periodic boundary
conditions, is given by

χp (λ, ϕa ) = ∆2 (λ, ϕa ) − 4 = −4 sin2 (κ(λ)). (6.4)

The periodic eigenvalues of L(ϕa ) are thus given by the λ’s satisfying κ(λ) = nπ for some
n ∈ Z, or
λ2 + |a|2 = n2 π 2 . (6.5)

39
The monodromy matrix M̀ for such a λ is given by
 
(−1)n 0
M̀ = (6.6)
0 (−1)n

when n 6= 0 and by
   
1 − iλ ia 1 ± |a| ia
M̀ = = (6.7)
iā 1 + iλ iā 1 ∓ |a|

when n = 0. It is convenient to list the periodic eigenvalues not in lexicographic ordering,

but rather use the integer n ∈ Z in (6.5) as an index. When listed in this way, we denote
the periodic eigenvalues by λ̂±
n , n ∈ Z, which are defined as follows. For any n ∈ Z with
|nπ| > |a| denote p
λ̂+ −
n = λ̂n = sgn(n) ·
+
n2 π 2 − |a|2 .
In view of (6.6), λ̂+
n defined above is a periodic eigenvalue of L(ϕa ) of geometric multiplicity
two. Using (6.3) and (6.4) one easily sees that λ̂+n has algebraic multiplicity is two. Further,
for any n ∈ Z with 0 < |nπ| < |a| denote
p
λ̂+ −
n = λ̂n = sgn(n) · i
+
|a|2 − n2 π 2 .

Again, in view of (6.6), for n ∈ Z with 0 < |nπ| < |a|, λ̂+ n is a periodic eigenvalue of
L(ϕa ) of geometric multiplicity two and, by (6.3) and (6.4), its algebraic multiplicity is two.
Next note that for n = 0, one has λ̂± 0 = ±i|a|. In view of (6.7), for a 6= 0 the geometric
+ −
multiplicity of λ̂0 as well as of λ̂0 equals one. In view of (6.3) and (6.4) the algebraic
multiplicity of λ̂+ −
0 and the one of λ̂0 is one. For a 6= 0, the eigenfunctions corresponding to
+ −
λ̂0 and λ̂0 are the constant vectors (a, i|a|) resp. (a, −i|a|). We then obtain the following
result, used in the proof of Theorem 3.2.

Lemma 6.6. For any k ∈ Z, consider the potential ϕa,−k = (ae−2iπkx , −ae2iπkx ). Then
λ̂±
0 = kπ ± i|a| are periodic eigenvalues of L(ϕa,−k ) of algebraic multiplicity one.

In the special case where |a| = na π for some na ∈ Z>0 set λ̂±
±na = 0. The above
computations yield

Corollary 6.7. (i) For a ∈ C, ϕa is a standard potential iff |a| < π.

(ii) For a ∈ C, any multiple periodic eigenvalue λ of L(ϕa ) satisfies mp (λ) = 2 and
mg (λ) = 2 iff |a| > π and |a| =
6 πZ.

(iii) If a ∈ C \ {0} satisfies |a| ∈ πZ, then 0 is a periodic eigenvalue of L(ϕa ) of algebraic
multiplicity four.

40
Isospectral set Iso0 (ϕa ): Denote by Iso0 (ϕa ) the connected component containing ϕa of the
set Iso(ϕa ) of all potentials ϕ ∈ iL2r with specp L(ϕ) = specp L(ϕa ). By the computations
above one sees that
{|a|eiα | α ∈ R} ⊆ Iso0 (ϕa ).
For |a| sufficiently small, ϕa is in the domain of the Birkhoff map introduced in Theorem
1.1 in [11]. As the L2 -norm is a spectral invariance it then follows that, for |a| sufficiently
small, all of Iso(ϕa ) is contained in this domain. According to the computations above ϕa
is a 1-gap potential. It then follows from Theorem 1.1 in [11] and its proof that Iso(ϕa ) is
homeomorphic to a circle. As a consequence

Iso(ϕa ) = Iso0 (ϕa ) = {|a|eiα | α ∈ R}.

Most likely the latter identities remain true for any |a| < π, but we have not verified this.
For |a| > π, Li and McLaughlin observed p that Iso0 (ϕa ) is larger than {|a|eiα | α ∈ R}.
+
Indeed, let π < |a| < 2π. Then λ̂±1 = ±i |a|2 − π 2 are periodic eigenvalues of geometric
+

multiplicity two. In subsection 4.3 of [16], using Bäcklund transformation techniques, for-
mulas of solutions of fNLS are presented which evolve on Iso0 (ϕa ) and depend explicitly on
x. They are parametrized by the punctured complex plane C∗ := eρ eiβ with coordinates
(ρ, β) ∈ R × R/2πZ, whereas the angle variable α in {|a|eiα | α ∈ R} is proportional to
the time t. As t → ±∞ these solutions approach the x independent solutions evolving
on {|a|eiα | α ∈ R}. Due to the trace formulas ([16], Section 2.4), on the orbits of these
solutions, the periodic eigenvalues λ̂+
±1 have geometric multiplicity one.
Dirichlet spectrum of L(ϕa ): By Lemma 6.4, the characteristic function of the Dirichlet
spectrum of L(ϕa ) is given by
 
sin κ ā − a
χD (λ, ϕa ) = λ+ . (6.8)
κ 2
The Dirichlet eigenvalues of L(ϕa ) are thus given by the λ’s satisfying

κ(λ) = nπ (6.9)

for some n ∈ Z \ {0} or

ā − a
λ+ = 0. (6.10)
2
Note that by the definition of the Dirichlet boundary conditions, any Dirichlet eigenvalue
is of geometric multiplicity one. It is convenient to list the Dirichlet eigenvalues not in
lexicographic ordering, but rather use the integer n in (6.9) as an index. When listed in
this way, we denote the Dirichlet eigenvalues by µ̂n , n ∈ Z, which are defined as follows.
For all n ∈ Z with |nπ| > |a| denote
p
µ̂n = sgn(n) · + n2 π 2 − |a|2 .

41
From (6.8) it follows that µ̂n has algebraic multiplicity one. For all n ∈ Z with 0 < |nπ| < |a|
denote p
µ̂n = sgn(n) · i + |a|2 − n2 π 2 .
By the same arguments as in the case |nπ| > |a|, the algebraic multiplicity of µ̂n is equal
to one iff µ̂n + ā−a
2 6= 0 and two otherwise. For n = 0 denote

µ̂0 = i Im(a).

Again, by the same arguments, µ̂0 has algebraic multiplicity equal to one if
p
[Im(a) 6= 0 and Im(a) 6= ± + |a|2 − n2 π 2 ∀ 0 < |nπ| < |a|] or [Im(a) = 0 and |a| 6∈ πZ>0 ]

or two if n p o
Im(a) ∈ ± + |a|2 − n2 π 2 | 0 < |nπ| < |a|
or three if
Im(a) = 0 and |a| ∈ πZ>0 .
In the special case where |a| = na π for some na ∈ Z>0 one has µ̂na = µ̂−na = 0. The
algebraic multiplicity of µ̂na is two (Im(a) 6= 0) or three (Im(a) = 0). These computations
lead to the following
Corollary 6.8. Let a ∈ C. Then
(i) If |a| < π, then the Dirichlet spectrum of L(ϕa ) is simple.

(ii) If |a| 6∈ πZ>0 , then the only possible multiple Dirichlet eigenvalue is i Im(a). It is at
most of algebraic multiplicity two.

(iii) If |a| ∈ πZ>0 , then 0 is a Dirichlet eigenvalue of algebraic multiplicity two or three.

(iv) For any 0 < nπ < |a| or nπ > |a|, µ̂n is a periodic eigenvalue of geometric and
algebraic multiplicity two whereas for |a| = nπ ∈ πZ>0 , µ̂n = 0 is a periodic eigenvalue
of algebraic multiplicity 4.

References
[1] M. Ablowitz, Y. Ma, The Periodic Cubic Schrödinger Equation, Studies in Applied
Mathematics, 65(1981), 113-158

[2] V. Arnold, On matrices depending on parameters, Russian Math. Surveys, 26 (1968),

29-44

[3] V. Arnold, Modes and quasimodes, Funct. Anal. Appl., 6(1972), 94-101

42
 [4] D. Burghelea, T. Kappeler, Multiplicities of the eigenvalues of the discrete Schrö-
dinger equation in any dimension, Proc. Amer. Math. Soc., 102(1988), no 2, 255-260

[5] P. Deift, Some open problems in random matrix theory and the theory of integrable
systems, in Integrable Systems and Random Matrices: in Honor of Percy Deift, J.
Baik at al., Editors, Contemporary Mathematics, 458(2008), American Math. Soc.,
Providence, 419-420

[6] S. Friedland, J. Robin, J. Sylvester, On the crossing rule, Comm. Pure Appl. Math.,
37(1984), 19-37

[7] B. Grébert, T. Kappeler, Symmetries of the Nonlinear Schrödinger equation, Bull.

Soc. math. France, 130(2002), no. 4, 603-618

[8] B. Grébert, T. Kappeler, Perturbations of the defocusing Nonlinear Schrödinger

equation, Milan J. Math., 71(2003), 141-174

[9] B. Grébert, T. Kappeler, J. Pöschel, Normal form theory for the nonlinear Schrö-
dinger equation, Lectures in Mathematics, EMS, to appear; ArXiv: 0907.3938v1.

[10] T. Kato, Perturbation theory for linear operators, Reprint of the 1980 Edition,
Springer Verlag, 1966

[11] T. Kappeler, P. Lohrmann, P. Topalov, N. Zung, Birkhoff coordinates for the focus-
ing NLS-equation, Comm. Math. Phys., 285(2009), 1087-1107

[12] T. Kappeler, J. Pöschel, KdV & KAM, 45(2003), Ergeb. der Math. und ihrer Gren-
zgeb., Springer Verlag

[13] T. Kappeler, F. Serier, P. Topalov, On the characterization of the smoothness of

skew-adjoint potentials in periodic Dirac operators, J. Funct. Anal., 256(2009), 2069-
2112

[14] K. Kendig, Elementary algebraic geometry, Springer Verlag, 1977

[15] P. Lax, The multiplicities of eigenvalues, Bull. Amer. Math. Soc., 6(1982), 213-215

[16] Y. Li, D. McLaughlin, Morse and Melnikov Functions for NLS Pde’s, Comm. Math.
Phys., 162, 175-214

[17] B. Mityagin, Spectral expansions of one-dimensional periodic Dirac operators, Dy-

namics of PDE, 1(2004), no. 2, 125-191

[18] J. von Neumann, E. Wigner, Über das Verhalten von Eigenwerten bei adiabatischen
Prozessen, Phys. Zeitschrift, 30(1929), 467-479

43
[19] L. Nirenberg, An extended interpolation inequality, Ann. Sc. Norm. Sup. di Pisa,
Ser. 6, 20(1966), 733-737

[20] V. Zakharov, A. Shabat A scheme for integrating nonlinear equations of mathe-

matical physics by the method of the inverse scattering problem I, Functional Anal.
Appl., 8(1974), 226-235

44