Scattering Theory

Landelijke Master

Erik Koelink
[email protected]

Spring 2008
ii
Contents

1 Introduction 1
1.1 Scattering theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Inverse scattering method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

2 Schrödinger operators and their spectrum 5

d2
2.1 The operator − dx 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
d2
2.2 The Schrödinger operator − dx 2 + q . . . . . . . . . . . . . . . . . . . . . . . . 8
2.3 The essential spectrum of Schrödinger operators . . . . . . . . . . . . . . . . . 12
2.4 Bound states and discrete spectrum . . . . . . . . . . . . . . . . . . . . . . . . 16
2.5 Explicit examples of potentials . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.5.1 Indicator functions as potential . . . . . . . . . . . . . . . . . . . . . . 18
2.5.2 cosh−2 -potential . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.5.3 Exponential potential . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

3 Scattering and wave operators 27

3.1 Time-dependent Schrödinger equation . . . . . . . . . . . . . . . . . . . . . . . 27
3.2 Scattering and wave operators . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.3 Domains of the wave operators for the Schrödinger operator . . . . . . . . . . 35
3.4 Completeness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

4 Schrödinger operators and scattering data 47

4.1 Jost solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.2 Scattering data: transmission and reflection coefficients . . . . . . . . . . . . . 57
4.3 Fourier transform properties of the Jost solutions . . . . . . . . . . . . . . . . 65
4.4 Gelfand-Levitan-Marchenko integral equation . . . . . . . . . . . . . . . . . . 70
4.5 Reflectionless potentials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

5 Inverse scattering method and the Korteweg-de Vries equation 83

5.1 The Korteweg-de Vries equation and some solutions . . . . . . . . . . . . . . . 83
5.2 The KdV equation related to the Schrödinger operator . . . . . . . . . . . . . 89
5.3 Lax pairs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
5.4 Time evolution of the spectral data . . . . . . . . . . . . . . . . . . . . . . . . 96

iii
iv

5.5 Pure soliton solutions to the KdV-equation . . . . . . . . . . . . . . . . . . . . 98

6 Preliminaries and results from functional analysis 103

6.1 Hilbert spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
6.2 Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
6.3 Fourier transform and Sobolev spaces . . . . . . . . . . . . . . . . . . . . . . . 110
6.4 Spectral theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
6.5 Spectrum and essential spectrum for self-adjoint operators . . . . . . . . . . . 115
6.6 Absolutely continuous and singular spectrum . . . . . . . . . . . . . . . . . . . 120

Bibliography 123

Index 125
Chapter 1

Introduction

According to Reed and Simon [9], scattering theory is the study of an interacting system on
a time and/or distance scale which is large compared to the scale of the actual interaction.
This is a natural phenomenon occuring in several branches of physics; optics (think of the blue
sky), acoustics, x-ray, sonar, particle physics,. . . . In this course we focus on the mathematical
aspects of scattering theory, and on an important application in non-linear partial differential
equations.

1.1 Scattering theory

As an example motivating the first chapters we consider the following situation occuring in
quantum mechanics. Consider a particle of mass m moving in three-dimensional space R3
according to a potential V (x, t), x ∈ R3 the spatial coordinate and R time t ∈ 2R. In quantum
mechanics this is modelled by a wave function ψ(x, t) satisfying R3 |ψ(x, t)| dx = 1 for all
time t, and the wave function is interpreted as a probability distribution for each time t. This
means
R that for each time t, the probability that the particle is in the set A ⊂ R3 is given by
A
|ψ(x, t)|2 dx. Similarly, the probability distribution of the momentum for this particle is
given by |ψ̂(p, t)|2 , where
Z
1
ψ̂(p, t) = √
3 e−ip·x ψ(x, t) dx
2π R3

is the Fourier transform of the wave function with respect to spatial variable x. (Here we
have scaled Planck’s constant ~ to 1.) Using the fact that the Fourier transform interchanges
differentiation and multiplication the expected value for the momentum can be expressed in
terms of a differential operator. The kinetic energy, corresponding to |p|2 /2m, at time t of
−1
the particle can be expressed as 2m h∆ψ, ψi, where we take the inner product corresponding
2 3
to the Hilbert space L (R ) and ∆ is the three-dimensional Laplacian (i.e. with respect to
the spatial coordinate x).

1
2 Chapter 1: Introduction

The potential energy at time t of the particle is described by hV ψ, ψi, so that (total)
energy of the particle at time t can be written as
−1
E = hHψ, ψi, H= ∆ + V,
2m
where we have suppressed the time-dependence. The operator H is known as the energy
operator, or as Hamiltonian, or as Schrödinger operator. In case the potential V is independent
of time and suitably localised in the spatial coordinate, we can view H as a perturbation of
−1
the corresponding free operator H0 = 2m ∆. This can be interpreted that we have a ‘free’
particle scattered by the (time-independent) potential V . The free operator H0 is a well-
known operator, and we can ask how its properties transfer to the perturbed operator H.
Of course, this will depend on the conditions imposed on the potential V . In Chapter 2 we
study this situation in greater detail for the case the spatial dimension is 1, but the general
perturbation techniques apply in more general situations as well. In general the potentials for
which these results apply are called short-range potentials.
In quantum mechanics the time evolution of the wave function is determined by the time-
dependent Schrödinger equation
∂
i ψ(x, t) = Hψ(x, t). (1.1.1)
∂t
We want to consider solutions that behave as solutions to the corresponding free time-
dependent Schrödinger equation, i.e. (1.1.1) with H replaced by H0 , for time to ∞ or −∞.
In case of the spatial dimension being 1 and for a time-independent potential V , we study
this situation more closely in Chapter 3. We do this for more general (possibly unbounded)
self-adjoint operators acting a Hilbert space.
Let us assume that m = 21 , then we can write the solution to the free time-dependent
Schrödinger equation as Z
2
ψ(x, t) = F (p) eip·x e−i|p| t dp,
R3
where F (p) denotes the distribution of the momenta at time t = 0 (up to a constant). This
follows easily since the exponentials eip·x are eigenfunctions of H0 = −∆ for the eigenvalue
|p|2 . For the general case the solution is given by
Z
2
ψ(x, t) = F (p) ψp (x) e−i|p| t dp,
R3

where Hψp = |p|2 ψp and where we can expect ψp (x) ∼ eip·x if the potential V is sufficiently
‘nice’. We study this situation more closely in Chapter 4 in case the spatial dimension is one.

1.2 Inverse scattering method

As indicated in Section 4.2 we give an explicit relation between the potential q in the Schrö-
d2
dinger operator − dx 2 + q and its scattering data consisting of the reflection and transition
Chapter 1: Introduction 3

coefficient R and T . To be specific, in Section 4.2 we discuss the transition q 7→ {R, T },

which is called the direct problem, and in Section 4.4 we discuss the transition {R, T } 7→ q,
the inverse problem. In case we take a time-dependent potential q(x, t), the scattering data
also becomes time-dependent. It has been shown that certain non-linear partial differential
equations (with respect to t and x) for q imply that the corresponding time evolution of the
scattering data R and T is linear! Or, scattering can be used to linearise some non-linear
partial differential equations.
The basic, and most famous, example is the Korteweg-de Vries equation, KdV-equation
for short,
qt (x, t) − 6 q(x, t) qx (x, t) + qxxx (x, t) = 0,
which was introduced in 1894 in order to model the solitary waves encountered by James
Scott Russell in 1834 while riding on horseback along the canal from Edinburgh to Glasgow.
Although the Korteweg-de Vries paper “On the change of form of long waves advancing in
a rectangular canal, and on a new type of long stationary waves”, Philosophical Magazine,
39 (1895), 422–443, is now probably the most cited paper by Dutch mathematicians, the
KdV-equation lay dormant for a long time until it was rediscovered by Kruskal and Zabusky
in 1965 for the Fermi-Pasta-Ulam problem on finite heat conductivity in solids. Kruskal and
Zabusky performed numerical experiments, and found numerical evidence for the solitary
waves as observed by Scott Russell, which were named “solitons” by Kruskal and Zabusky.
The fundamental discovery made by Gardner, Greene, Kruskal and Miura in 1967 is that if q
is a solution to the KdV-equation, then the spectrum of the Schrödinger with time-dependent
potential q is independent of time, and moreover, the time evolution for the scattering data,
i.e. the reflection and transmission coefficient, is a simple linear differential equation. We
discuss this method in Chapter 5. In particular, we discuss some of the soliton solutions.
This gave way to a solution method, nowadays called the inverse scattering method, or
inverse spectral method, for classes of non-linear partial differential equations. Amongst
others, similar methods work for well-known partial differential equations such as the modified
KdV-equation qt − 6q 2 qx + qxxx = 0, the sine-Gordon equation qxt = sin q, the non-linear
Schrödinger equation qt = qxx + |q|2 q, and many others. Nowadays, there are many families
of partial differential equations that can be solved using similar ideas by realising them as
conditions for isospectrality of certain linear problem. For the isospectrality we use the method
of Lax pairs in Section 5.3.
There have been several approaches and extensions to the KdV-equations, notably as
integrable systems with infinitely many conserved Poisson-commuting quantities. We refer to
[1], [8].

1.3 Overview
The contents of these lecture notes are as follows. In Chapter 6 we collect some results from
functional analysis, especially on unbounded self-adjoint operators and the spectral theorem,
Fourier analysis, especially related to Sobolev and Hardy spaces. For these results not many
4 Chapter 1: Introduction

proofs are given, since they occur in other courses. In Chapter 6 we also discuss some results
on the spectrum and the essential spectrum, and for these results explicit proofs are given.
So Chapter 6 is to be considered as an appendix.
d2
In Chapter 2 we study first the Schrödinger operator − dx 2 + q, and we give conditions on

q such that the Schrödinger operator is a unbounded self-adjoint operator with the Sobolev
d2
space as the domain which is also the domain for the unperturbed Schrödinger operator − dx 2.

For this we use a classical result known as Rellich’s perturbation theorem on perturbation of
unbounded self-adjoint operators. We discuss the spectrum and the essential spectrum in this
case. It should be noted that we introduce general results, and that the Schrödinger operator
is merely an elaborate example.
In Chapter 3 the time-dependent Schrödinger operator is discussed. We introduce the
notions of wave operators, scattering states and the scattering operator. Again this is a
general procedure for two self-adjoint operators acting on a Hilbert space, and the Schrödinger
operator is an important example.
In Chapter 4 the discussion is specific for the Schrödinger operator. We show how to
determine the Jost solutions, and from this we discuss the reflection and transmission coeffi-
cient. The Gelfand-Levitan-Marchenko integral equation is derived, and the Gelfand-Levitan-
Marchenko equation is the key step in the inverse scattering method.
In Chapter 5 we study the Korteweg-de Vries equation, and we discuss shortly the original
approach of Gardner, Greene, Kruskal and Miura that triggered an enormous amount of
research. We then discuss the approach by Lax, and we show how to construct the N -soliton
solutions to the KdV-equation. We carry out the calculations for N = 1 and N = 2. This
chapter is not completely rigorous.
The lecture notes end with a short list of references, and an index of important and/or
useful notions which hopefully increases the readability. The main source of information for
Chapter 2 is Schechter [10]. Schechter’s book [10] is also relevant for Chapter 3, but for Chapter
3 also Reed and Simon [9] and Lax [7] have been used. For Chapter 4 several sources have
been used; especially Eckhaus and van Harten [4], Reed and Simon [9] as well as an important
original paper [3] by Deift and Trubowitz. For Chapter 5 there is an enormous amount of
information available; introductory and readable texts are by Calogero and Degasperis [2],
de Jager [6], as well as the original paper by Gardner, Greene, Kruskal and Miura [5] that
has been so influential. As remarked before, Chapter 6 is to be considered as an appendix,
and most of the results that are not proved can be found in general text books on functional
analysis, such as Lax [7], Werner [11], or course notes for a course in Functional Analysis.
Chapter 2

Schrödinger operators and their

spectrum

In Chapter 6 we recall certain terminology, notation and results that are being used in Chapter
2.

d 2
2.1 The operator − dx 2

d 2
2
Theorem 2.1.1. − dx 2 with domain the Sobolev space W (R) is a self-adjoint operator on
2
L (R).
d 2
2
We occasionally denote − dx 2 by L0 , and then its domain by D(L0 ) = W (R). We first
d
consider the operator i dx on its domain W 1 (R).
d
Lemma 2.1.2. i dx with domain the Sobolev space W 1 (R) is self-adjoint.
d
Proof. The Fourier transform, see Section 6.3, intertwines i dx with the multiplication operator
1
M defined by (M f )(λ) = λf (λ). The domain W (R) under the Fourier transform is precisely
d
D(M ) = {f ∈ L2 (R) | λ 7→ λf (λ) ∈ L2 (R)}, see Section 6.3. So (i dx , W 1 (R)) is unitarily
equivalent to (M, D(M )).
Observe that (M, D(M )) is symmetric;
Z
hM f, gi = λ f (λ)g(λ) dλ = hf, M gi, ∀f, g ∈ D(M ).
R
∗
Assume now g ∈ D(M ), or
Z
D(M ) 3 f 7→ hM f, gi = λ f (λ)g(λ) dλ ∈ C
R

is a continuous functional on L2 (R). By taking complex conjugates, this implies the existence
of a constant C such that
Z
 
 λg(λ)f (λ) dλ ≤ Ckf k, ∀f ∈ W 1 (R).
R

5
6 Chapter 2: Schrödinger operators and their spectrum

By the converse Hölder’s inequality, it follows that λ 7→ λg(λ) is square integrable, and
kλ 7→ λg(λ)k ≤ C. Hence, D(M ∗ ) ⊂ D(M ), and since the reverse inclusion holds for any
densely defined symmetric operator, we find D(M ) = D(M ∗ ), and so (M, D(M )) is self-
adjoint. This gives the result.
Note that we actually have that the Fourier transform gives the spectral decomposition
d d
of −i dx . The functions x 7→ eiλx are eigenfunctions to −i dx for the eigenvalue λ, and the
iλ·
Fourier transform of f is just λ 7→ hf, e i, and the inverse Fourier transform states that f is
a continuous linear combination of the eigenvectors, f = √12π R hf, eiλ· ieiλ· dλ.
R

Proof of Theorem 2.1.1. Observe that by Lemma 2.1.2

d ∗ d d2
) (i ) = − 2
(i
dx dx dx
0
as unbounded operators, since {f ∈ W (R) | f ∈ W 1 (R)} = W 2 (R). Now the theorem
1

follows from Lemma 6.2.4.

d 2
2
Theorem 2.1.3. − dx 2 with domain W (R) has spectrum σ = [0, ∞). There is no point
spectrum.
Proof. Let us first check for eigenfunctions, so we look for functions −f 00 = λ f , or f 00 +λ f = 0.
This equation
√ is easily solved; f is linear for λ = 0 and a combination of exponential functions
exp(±x −λ) for λ 6= 0. There is no non-trivial combination that makes an eigenfunction
d2
square-integrable, hence − dx 2 has no eigenvalues.

By the proof of Theorem 2.1.1 and Lemma 6.2.4 it follows that the spectrum is contained
in [0, ∞). In order to show that the spectrum is [0, ∞) we establish that any positive λ is
contained in the spectrum. Since the spectrum is closed, the result then follows.
So pick λ > 0 arbitrary. We use the description of Theorem 6.5.1, so we need to construct
a sequence of functions, say fn ∈ W 2 (R), of norm 1, such that k − fn00 − λ fn k → 0. By the first
paragraph ψ(x) = exp(iγx), with γ 2 = λ, satisfies ψ 00 = λψ, but this function is not in L2 (R).
2
The idea is to approximate this function in Lp (R) using an approximation of the delta-function.
The details
√ are as follows. Put φ(x) = 4
2/π exp(−x2 ), so that kφk = 1. Then define
φn (x) = ( n)−1 φ(x/n) and define fn (x) = φn (x) exp(iγx). Then kfn k = kφn k = kφk = 1,
and since fn is infinitely differentiable and fn and all its derivatives tend to zero rapidly as
x → ±∞ (i.e. fn ∈ S(R), the Schwartz space), it obviously is contained in the domain W 2 (R).
Now
1 x
fn0 (x) = φ0n (x)eiγx + iγφn (x)eiγx = √ φ0 ( )eiγx + iγφn (x)eiγx
n n n
1 x 2iγ x
fn00 (x) = 2 √ φ00 ( )eiγx + √ φ0 ( )eiγx − γ 2 φn (x)eiγx ,
n n n n n n
so that
2|γ| · 1 ·
kfn00 + λfn k ≤ √ kφ0 ( )k + 2 √ kφ00 ( )k
n n n n n n
2|γ| 0 1
= kφ k + 2 kφ00 k → 0, n → ∞.
n n
Chapter 2: Schrödinger operators and their spectrum 7

So this sequence meets the requirements of Theorem 6.5.1, and we are done.
d 2
Again we use the Fourier transform to describe the spectral decomposition of − dx 2 . The
d2
functions x 7→ cos(λx) and x 7→ sin(λx) are the eigenfunctions of − dx2 for the eigenvalue
λ2 ∈ (0, ∞). Observe that the Fourier transform as in Section 6.3 preserves the space of even
functions, and that for even functions we can write the transform pair as
√ Z ∞ √ Z ∞
2 2
fˆ(λ) = √ f (x) cos(λx) dx, f (x) = √ fˆ(λ) cos(λx) dλ,
π 0 π 0

which is known as the (Fourier-)cosine transform.

Exercise 2.1.4. Derive the (Fourier-)sine transform using odd functions. By splitting an
arbitrary element f ∈ L2 (R) into an odd and even part give the spectral decomposition of
d2
− dx 2.

d 2
Exercise 2.1.5. The purpose of this exercise is to describe the resolvent for L0 = − dx 2 . We
2
obtain that for z ∈ C\R, z = γ , =γ > 0, we get

−1
Z
−1
(L0 − z) f (x) = eiγ|x−y| f (y) dy
2γi R

• Check that u(x) defined by the right hand side satisfies −u00 − zu = f , for f in a suitable
dense subspace, e.g. Cc∞ (R). Show that (L0 − z)−1 extends to a bounded operator on
L2 (R) using Young’s inequality for convolution products.

• Instead of checking the result for f ∈ Cc∞ (R), we derive it by factoring the second order
differential operator −u00 − γ 2 u = f into two first order differential equations;

d d
(i + γ)u = v, (i − γ)v = f.
dx dx
– Show that the second first order differential equation is solved by
Z x
v(x) = −i e−iγ(x−y) f (y) dy.

R∞
Argue that requiring v ∈ L2 (R) gives v(x) = i x
e−iγ(x−y) f (y) dy. Show that
1
kvk ≤ |=γ| kf k.
– Treat the other first Rorder differential equation in a similar way to obtain the
x
expression u(x) = −i −∞ eiγ(x−y) v(y) dy.
– Finally, express u in terms of f to find the explicit expression for the resolvent
operator.
8 Chapter 2: Schrödinger operators and their spectrum

d 2
2.2 The Schrödinger operator − dx 2 +q

Consider the multiplication operator

Q : L2 (R) ⊃ D(Q) → L2 (R), Q : f → qf (2.2.1)
with domain D(Q) = {f ∈ L2 (R) | qf ∈ L2 (R)} for some fixed function q. We follow the
convention that the multiplication operator by a function, denoted by a small q, is denoted
by the corresponding capital Q. We always assume that q is a real-valued function and that
(Q, D(Q)) is densely defined, i.e. D(Q) is dense in L2 (R). R
Recall that we call a function q locally square integrable if B |q(x)|2 dx < ∞ for each
bounded measurable subset B of R. So we see that in this case any function in Cc∞ (R),
the space of infinitely many times differentiable functions having compact support, is in the
domain D(Q) of the corresponding multiplication operator Q. Since Cc∞ (R) is dense in L2 (R),
we see that for such a potential q the domain D(Q) is dense.
We are only interested in so-called short range potentials. Loosely speaking, the potential
only affects waves in a short range. Mathematically, we want the potential q to have sufficient
decay.
Lemma 2.2.1. (Q, D(Q)) is a self-adjoint operator.
Exercise 2.2.2. Prove Lemma 2.2.1. Hint: use the converse Hölder’s inequality, cf. proof of
Lemma 2.1.2.
Exercise 2.2.3. 1. Which conditions on the potential q ensure that Q : L2 (R) → L2 (R) is
actually a bounded operator?
2. Determine the spectrum of Q in this case.
d 2
We study the Schrödinger1 operator − dx 2
2 + q which has domain W (R) ∩ D(Q). In this

setting we call q the potential of the Schrödinger operator. The first question to be dealt with
d2
is whether or not − dx 2 + q with this domain is self-adjoint or not. For this we use a 1939

perturbation Theorem 2.2.4 by Rellich2 , and then we give precise conditions on the potential
function q such that the conditions of Rellich’s Perturbation Theorem 2.2.4 in this case are
met.
Theorem 2.2.4 (Rellich’s Perturbation Theorem). Let H be a Hilbert space. Assume T : H ⊃
D(T ) → H is a self-adjoint operator and S : H ⊃ D(S) → H is a symmetric operator, such
that D(T ) ⊂ D(S) and ∃ a < 1, b ∈ R
kSxk ≤ a kT xk + b kxk, ∀ x ∈ D(T ),
then (T + S, D(T )) is a self-adjoint operator.
1
Erwin Rudolf Josef Alexander Schrödinger, (12 August 1887 — 4 January 1961) Austrian physicist, who
played an important role in the development of quantum mechanics, and is renowned for the cat in the box.
2
Franz Rellich (14 September 1906 — 25 September 1955) German mathematician, who made contributions
to mathematical physics and perturbation theory.
Chapter 2: Schrödinger operators and their spectrum 9

The infimum over all possible a is called the T -bound of S.

Note that in case S is a bounded self-adjoint operator, the statement is also valid. In this
case the estimate is valid with a = 0 and b = kSk.
Proof. Note that (T + S, D(T )) is obviously a symmetric operator. Since T is self-adjoint,
Ran(T − z) = H by Lemma 6.2.2 for all z ∈ C\R and R(z; T ) = (T − z)−1 ∈ B(H). In
particular, take z = iλ, λ ∈ R\{0}, then, since hT x, xi ∈ R,

k(T − iλ)xk2 = kT xk2 + 2<(iλhT x, xi) + λ2 kxk2

= kT xk2 + λ2 kxk2 , ∀x ∈ D(T ).

For y ∈ H arbitrary, pick x ∈ D(T ) such that (T − iλ)x = y, so that we get kyk2 =
kT (T − iλ)−1 yk2 + λ2 k(T − iλ)−1 yk2 . This implies the basic estimates

kyk ≥ kT (T − iλ)−1 yk, kyk ≥ |λ| k(T − iλ)−1 yk.

Now (T − iλ)−1 y = x ∈ D(T ) ⊂ D(S) and

kS(T − iλ)−1 yk ≤ akT (T − iλ)−1 yk + bk(T − iλ)−1 yk

b b
≤ a kyk + kyk = (a + ) kyk,
|λ| |λ|

and, since a + b/|λ| < 1 for |λ| sufficiently large, it follows that kS(T − iλ)−1 k < 1. In
particular, 1 + S(T − iλ)−1 is invertible in B(H), see Section 6.2. Now

(S + T ) − iλ = (T − iλ) + S = 1 + S(T − iλ)−1 (T − iλ),

(where the equality also involves the domains!) and we want to conclude that Ran(S + T −
iλ) = H. So pick x ∈ H arbitrary, we need to show that there exists y ∈ H such that

−1
(S + T ) − iλ y = x, and this can be rephrased as (T − iλ)y = 1 + S(T − iλ)−1 x by the

invertibility for |λ| sufficiently large. Since Ran(T − iλ) = H such an element y ∈ H does
exist.
Finally, since S + T with dense domain D(T ) is symmetric, it follows by Lemma 6.2.2 that
it is self-adjoint.
The next exercise considers what happens if a = 1.

Exercise 2.2.5. Prove Wüst’s theorem, which states the following. Assume T : H ⊃ D(T ) →
H is a self-adjoint operator and S : H ⊃ D(S) → H is a symmetric operator, such that
D(T ) ⊂ D(S) and there exists b ∈ R such that

kSxk ≤ kT xk + b kxk, ∀ x ∈ D(T ),

then (T + S, D(T )) is an essentially self-adjoint operator.

Proceed by showing that Ker((T + S)∗ ± i) is trivial using Lemma 6.2.2.
10 Chapter 2: Schrödinger operators and their spectrum

1. Show that we can use Rellich’s Perturbation Theorem 2.2.4 to see that T +tS, 0 < t < 1,
is self-adjoint.

2. Take x ∈ Ker((T + S)∗ − i), and show there exists yt ∈ D(T ) such that kyt k ≤ kxk and
(T + tS + i)yt = x. (Use Lemma 6.2.2.)

3. Define zt = x + (1 − t)Syt . Show that hx, zt i = 0.

4. Show that (1 − t)kT yt k ≤ k(T + tS)yt k + tbkyt k and conclude that (1 − t)kT yt k is
bounded as t % 1. Conclude next that (1 − t)kSyt k is bounded as t % 1, and hence
kzt k is bounded.

5. Show that x is the weak limit of zt as t % 1, and conclude that kxk = 0. (Hint consider
hzt − x, ui for u ∈ D(T ).)

6. Conclude that (T + S, D(T )) is an essentially self-adjoint operator.

7. Give an example to show that in general (T + S, D(T )) is not a self-adjoint operator.

d 2
We next apply Rellich’s Perturbation Theorem 2.2.4 to the case T equal − dx 2 and S equal

to Q. For this we need to rephrase the condition of Rellich’s Theorem 2.2.4, see Theorem
2.2.6(5), into conditions on the potential q.

Theorem 2.2.6. The following statements are equivalent:

1. W 2 (R) ⊂ D(Q),

2. ∃ C > 0 kqf k2 ≤ C kf 00 k2 + kf k2 , ∀ f ∈ W 2 (R),

Z y+1
3. sup |q(x)|2 dx < ∞,
y∈R y

4. ∀ε > 0 ∃K > 0 with kqf k2 ≤ ε kf 00 k2 + K kf k2 , ∀f ∈ W 2 (R),

5. ∀ε > 0 ∃K > 0 with kqf k ≤ ε kf 00 k + K kf k, ∀f ∈ W 2 (R).

d 2
So the − dx 2 -bound of Q is zero.

Corollary 2.2.7. If the potential q satisfies Theorem 2.2.6(3), then the Schrödinger operator
d2 2 2
− dx 2 + q is self-adjoint on its domain W (R). In particular, this is true for q ∈ L (R).

d 2
Proof of Theorem 2.2.6. (1) ⇒ (2): Equip W 2 (R) with the graph norm of L =− dx 2 , denoted
2
by k · kL , see Section 6.2, which in particular means that W (R) with this norm is complete.
We claim that Q : (W 2 (R), k · kL ) → L2 (R) is a closed operator, then the Closed Graph
Theorem 6.2.1 implies that it is bounded, which is precisely (2). To prove the closedness, we
take a sequence {fn }∞ 2 2
n=1 such that fn → f in (W (R), k · kL ) and qfn → g in L (R). We have
to show that f ∈ D(Q) and qf = g. First, since fn → f and (W 2 (R), k · kL ) is complete,
Chapter 2: Schrödinger operators and their spectrum 11

we have f ∈ W 2 (R) ⊂ D(Q) by assumption (1). This assumption and Lemma 2.2.1 also
gives hqfn , hi = hfn , qhi for all h ∈ W 2 (R). Taking limits, hg, hi = hf, qhi = hqf, hi since
convergence in (W 2 (R), k · kL ) implies convergence in L2 (R) by k · k ≤ k · kL . Since W 2 (R) is
dense in L2 (R), we may conclude g = Q(f ).
2
(2) ⇒ (3): Put φ(x) = e1−x , so in particular φ ∈ W 2 (R) and φ(x) ≥ 1 for x ∈ [0, 1]. Put
φy (x) = φ(x − y) for the translated function, then
Z y+1 Z
2
|q(x)φy (x)|2 dx ≤ C kφ00y k2 + kφy k2 = C kφ00 k2 + kφk2 < ∞



|q(x)| dx ≤
y R

since the integral is translation invariant.

(3) ⇒ (4): By Sobolev’s imbedding Lemma 6.3.1 we have W 2 (R) ⊂ C 1 (R), so we can
apply the following lemma.
Lemma 2.2.8. Take f ∈ C 1 (R), then for each ε > 0 and all intervals I = [y, y + 1] we have
Z Z
0 1
2 2
|f (x)| ≤ ε |f (t)| dt + (1 + ) |f (t)|2 dt, x ∈ I.
I ε I
Denote by C the supremum in (3). For f ∈ W 2 (R) ⊂ C 1 (R), we multiply the expression
in Lemma 2.2.8 by |q(x)|2 and integrate over the interval I, which gives
Z Z Z
0 −1
|q(x)f (x)| dx ≤ Cε |f (t)| dt + C(1 + ε ) |f (t)|2 dt.
2 2
I I I

Since R = ∪I, we find by summing

kqf k2 ≤ Cε kf 0 k2 + C(1 + ε−1 ) kf k2 .
In order to get the second (weak) derivative into play, we use the Fourier transform, see Section
6.3. In particular,
Z Z Z
0 2 2 2 1 4 2 1
kf k = λ |(Ff )(λ)| dλ ≤ λ |(Ff )(λ)| dλ + |(Ff )(λ)|2 dλ
R 2 R 2 R
Z Z
1 1
= |f 00 (t)|2 dt + |f (t)|2 dt
2 R 2 R
using 2|λ|2 ≤ |λ|4 + 1. Plugging this back in gives
1 1
kqf k2 ≤ Cε kf 00 k2 + C(1 + ε + ε−1 ) kf k2 ,
2 2
which is the required estimate after renaming the constants.
(4) ⇒ (5): Use
√
kqf k2 ≤ ε kf 00 k2 + K kf k2 ≤ ε kf 00 k2 + 2 εK kf 00 kkf k + K kf k2
√ √
2
= εkf 00 k + K kf k ,
which gives the result after taking square roots and renaming the constants.
(5) ⇒ (1): For f ∈ W 2 (R) the assumption implies qf ∈ L2 (R), or f ∈ D(Q).
12 Chapter 2: Schrödinger operators and their spectrum

It remains to prove Lemma 2.2.8, which is the following exercise.

Exercise 2.2.9. 1. Show that, by reducing to real and imaginary parts, we can restrict to
the case that f is a real-valued function in C 1 (R).
d(f 2 )
2. Using (x) = 2f (x)f 0 (x), and 2ab ≤ εa2 + ε−1 b2 , show that
dx
Z Z
2 2 0 2 1
f (x) − f (s) ≤ ε |f (t)| dt + |f (t)|2 dt
I ε I

for x, s ∈ I.

3. Show that for a suitable choice of s ∈ I Lemma 2.2.8 follows.

Exercise 2.2.10. Consider the differential operator

d2 p0 d p00
− − 2 + q −
dx2 p dx p

for some strictly positive p ∈ C 2 (R). Show that this differential operator for real-valued q
is symmetric for a suitable choice of domain on the Hilbert space L2 (R, p(x)2 dx). Establish
d2 2
that this operator is unitarily equivalent to the Schrödinger operator − dx 2 + q on L (R).

We have gone through some trouble to establish a suitable criterion in Corollary 2.2.7 such
that the Schrödinger operator is a self-adjoint operator. It may happen that the potential is
such that the corresponding Schrödinger operator defined on W 2 (R) ∩ D(Q) is not self-adjoint
but a symmetric densely defined operator, and then one has to look for self-adjoint extensions.
It can also happen that the potential is so ‘bad’, that W 2 (R) ∩ D(Q) is not dense. We do not
go into this subject.

2.3 The essential spectrum of Schrödinger operators

The notion of essential spectrum is recalled in Section 6.5. From Theorem 6.5.5 and Theorem
d2
2.1.3 it follows that the essential spectrum of − dx 2 is equal to its spectrum [0, ∞). In this

section we give a condition on the potential q that ensures that the essential spectrum of
d2 d2
− dx 2 + q is [0, ∞) as well. In this section we assume that − dx2 + q is a self-adjoint operator

with domain the Sobolev space W 2 (R), which is the case if Theorem 2.2.6(3) holds.
We start with the general notion of T -compact operator, and next discuss the influence of
a perturbation of T by a T -compact operator S on the essential spectrum.

Definition 2.3.1. Let (T, D(T )) be a closed operator on a Hilbert space H. The operator
(S, D(S)) on the Hilbert space is compact relative to T , or T -compact, if D(T ) ⊂ D(S) and
for any sequence {xn }∞ ∞
n=1 satisfying kxn k+kT xn k ≤ C the sequence {Sxn }n=1 has a convergent
subsequence.
Chapter 2: Schrödinger operators and their spectrum 13

Note that Definition 2.3.1 can be rephrased equivalently as S : (D(T ), k · kT ) → H being

compact, where (D(T ), k · kT ) is the Hilbert space equipped with the graph norm.

Theorem 2.3.2. Let (T, D(T )) be a self-adjoint operator on a Hilbert space H, and (S, D(S))
a closed symmetric T -compact operator. Then (T + S, D(T )) is a self-adjoint operator, and
σess (T ) = σess (S + T ).

Proof. By Definition 2.3.1, D(T ) ⊂ D(S), so if we can check the condition in Rellich’s Per-
turbation Theorem 2.2.4 we may conclude that (T + S, D(T )) is self-adjoint. We claim that
for all ε > 0 there exists K > 0 such that

kSxk ≤ ε kT xk + K kxk, ∀x ∈ D(T ). (2.3.1)

Then the first statement follows from Rellich’s Theorem 2.2.4 by taking ε < 1.
To prove this claim we first prove a weaker claim, namely that there exists a constanst C
such that
kSxk ≤ C kT xk + kxk), ∀x ∈ D(T ).
Indeed, if this estimate would not hold, we can find a sequence {xn }∞ n=1 such that kT xn k +
∞
kxn k = 1 and kSxn k → ∞. But by T -compactness, {Sxn }n=1 has a convergent subsequence,
say Sxnk → y, so in particular kSxnk k → kyk contradicting kSxn k → ∞.
Now to prove the more refined claim (2.3.1) we argue by contradiction. So assume ∃ε > 0
such that we cannot find a K such that (2.3.1) holds. So we can find a sequence {xn }∞ n=1
in D(T ) such that kSxn k > εkT xn k + nkxn k. By changing xn to xn /(kxn k + kT xn k) we
can assume that this sequence
satisfies kxn k + kT xn k = 1. By the weaker estimate we have
kSxn k ≤ C kT xn k + kxn k ≤ C and C ≥ kSxn k ≥ εkT xn k + nkxn k, so that kxn k →
0 and hence kT xn k → 1 as n → ∞. By T -compactness, the sequence {Sxn }∞ n=1 has a
convergent subsequence, which we also denote by {Sxn }∞ n=1 , say Sx n → y. Since we assume
(S, D(S)) closed, we see that xn → 0 ∈ D(S) and y = S0 = 0. On the other hand kyk =
limn→∞ kSxn k ≥ ε limn→∞ kT xn k = ε > 0. This gives the required contradiction.
In order to prove the equality of the essential spectrum, take λ ∈ σess (T ), so we can take
a sequence {xn }∞ n=1 as in Theorem 6.5.4(3). Recall that this means that the sequence satisfies
kxn k = 1, hxn , yi → 0 for all y ∈ H and k(T − λ)xn k → 0. We will show that λ ∈ σess (T + S).
In particular, kxn k + kT xn k is bounded, and by T -compactness it follows that {Sxn }∞ n=1
has a convergent subsequence, say Sxnk → y. We claim that y = 0. Indeed, for arbitrary
z ∈ D(T ) ⊂ D(S) we have

hy, zi = lim hSxnk , zi = lim hxnk , Szi = 0,

k→∞ k→∞

since S is symmetric and using that xn → 0 weakly. By density of D(T ), we have y = 0.

So (S + T − λ)xnk → 0, and by Theorem 6.5.4(3) it follows that λ ∈ σess (S + T ). Or
σess (T ) ⊂ σess (S + T ).
Conversely, we can use the above reasoning if we can show that −S is (T + S)-compact,
since then σess (S + T ) ⊂ σess (−S + T + S) = σess (T ). (Note that we already have shown that
14 Chapter 2: Schrödinger operators and their spectrum

(T + S, D(T )) is self-adjoint, and of course −S is a closed symmetric operator.) It suffices to

prove that S is (T + S)-compact, since then −S is also (T + S)-compact. We first observe
that S is T -compact implies the existence of a constant C0 > 0 such that


kxk + kT xk ≤ C0 kxk + k(S + T )xk . (2.3.2)

Indeed, arguing by contradiction, if this is not true there exists a sequence {xn }∞ n=1 such
that kxn k + kT xn k = 1 and kxn k + k(S + T )xn k → 0, and by T -compactness of S we have
that there exists a convergent subsequence {Sxnk }∞ ∞
k=1 of {Sxn }n=1 converging to x. Then
necessarily T xnk → −x and xnk → 0, and since T is self-adjoint, hence closed, we have x = 0.
This contradicts the assumption kxk + kT xk = 1.
Now to prove that S is (T + S)-compact, take a sequence {xn }∞ n=1 satisfying kxn k + k(T +
S)xn k ≤ C, then by (2.3.2) we have kxn k + kT xn k ≤ CC0 , so by T -compactness the sequence
{Sxn }∞
n=1 has a convergent subsequence. Or S is (T + S)-compact.

Exercise 2.3.3. The purpose of this exercise is to combine Rellich’s Perturbation Theorem
2.2.4 with Theorem 2.3.2 in order to derive the following statement: Let (T, D(T )) be a
self-adjoint operator, and S1 , S2 symmetric operators such that (i) D(T ) ⊂ D(S1 ), (ii) S2 is
a closed T -compact operator, (iii) ∃a < 1, b ≥ 0 such that kS1 xk ≤ akT xk + bkxk for all
x ∈ D(T ). Then S2 is (T + S1 )-compact and σess (T + S1 + S2 ) = σess (T + S1 ).
Prove this result using the following steps.

• Show kT xk ≤ k(T +S1 )xk+akT xk+bkxk and conclude (1−a)kT xk ≤ k(T +S1 )xk+bkxk.

• Show that S2 is (T + S1 )-compact.

• Conclude, using Theorem 2.3.2, that σess (T + S1 + S2 ) = σess (T + S1 ).

d 2
We want to apply Theorem 2.3.2 to the RSchrödinger operator − dx 2 + q. Recall that we
2
call a function q locally square integrable if B |q(x)| dx < ∞ for each bounded measurable
subset B of R. Note that in particular any square integrable function, which we consider as
an element of L2 (R), is locally square integrable. The function q(x) = 1 is an example of a
locally square integrable function that is not square integrable.
d 2
Theorem 2.3.4. The operator (Q, D(Q)) is − dx 2 -compact if and only if q is a locally square

integrable function satisfying Z y+1

lim |q(x)|2 dx → 0.
|y|→∞ y

d 2
Corollary 2.3.5. If q ∈ L2 (R), then the essential spectrum of − dx 2 + q is [0, ∞).

Proof of Theorem 2.3.4. First assume that the condition on the potential q is not valid,
∞
R yn +1 there 2exists a ε > 0 and a sequence∞ of points {yn }n=1 such that |yn | → ∞ and
then
yn
|q(x)| dx ≥ ε. Pick a function φ ∈ Cc (R) with the properties φ(x) ≥ 1 for x ∈ [0, 1]
and supp(φ) ⊂ [−1, 2]. Define the translated function φn (x) = φ(x − yn ), then obviously
Chapter 2: Schrödinger operators and their spectrum 15

φn ∈ W 2 (R) and kφ00n k + kφn k = kφ00 k + kφk is independent of n. Consequently, the as-
d2 ∞
sumption that Q is − dx 2 -compact implies that {Qφn }n=1 has a convergent subsequence, again
∞
denoted by {Qφn }n=1 , say Qφn → f . Observe that
Z Z yn +1
2
|q(x)φn (x)| dx ≥ |q(x)|2 dx ≥ ε,
R yn
√
so that kf k ≥ ε > 0. On the other hand, for any fixed bounded interval I of length 1 we
have φn |I = 0 for n sufficiently large since |yn | → ∞, so that
Z  12 Z  21 Z  21
2 2 2
|f (x)| dx ≤ |f (x) − q(x)φn (x)| dx + |q(x)φn (x)| dx
I I I
Z  12
2
= |f (x) − q(x)φn (x)| dx → 0, n → ∞.
I
R
This shows I |f (x)|2√dx = 0, hence, by filling R with such intervals, kf k = 0, which is
contradicting kf k ≥ ε > 0.
Now assume that the assumption on q is valid. We start with some general remarks on
W (R). For f ∈ W 2 (R) ⊂ C 1 (R) (using the Sobolev inbedding Lemma 6.3.1) we have, using
2

the Fourier transform and the Cauchy-Schwarz inequality (6.1.1),

Z
1
f (x) = √ eiλx (Ff )(λ) dλ
2π R
1 1 √ 1√ √
⇒ |f (x)| ≤ √ kλ 7→ √ k kλ 7→ 1 + λ2 (Ff )(λ)k = 2kλ 7→ 1 + λ2 (Ff )(λ)k
2π 1 + λ2 2

and similarly
1√ √
|f 0 (x)| ≤
2kλ 7→ 1 + λ2 λ(Ff )(λ)k.
2
Now, reasoning as in the proof of (3) ⇒ (4) for Theorem 2.2.6 we see
√ 2
Z Z
kλ 7→ 1 + λ2 (Ff )(λ)k = |(Ff )(λ)| dλ + λ2 |(Ff )(λ)|2 dλ
2
R R
Z Z
3 1 3 1
≤ |(Ff )(λ)|2 dλ + λ4 |(Ff )(λ)|2 dλ = kf k2 + kf 00 k2 ,
2 R 2 R 2 2

and similarly
√ 1 3
kλ 7→ 1 + λ2 λ(Ff )(λ)k2 ≤ kf k2 + kf 00 k2 ,
2 2
√
so, with C = 3,
|f (x)| + |f 0 (x)| ≤ C (kf 00 k + kf k) .
(So we have actually proved that the Sobolev inbedding W 2 (R) ⊂ C 1 (R) is continuous.)
16 Chapter 2: Schrödinger operators and their spectrum

If {fn }∞ 2 00
n=1 is now a sequence in W (R) such that kfn k + kfn k ≤ C1 , then it follows that
fn and fn0 are uniformly bounded, this in particular implies that M = {fn | n ∈ N} is
uniformly continuous. Since M is also bounded in the supremum norm, it follows by the
Arzelà-Ascoli theorem that the closure of M is compact in C(R). In particular, there is a
convergent subsequence, which is also denoted by {fn }∞n=1 . This result, with the locally square
integrability condition on q, now gives for each fixed N > 0
Z N Z N

2 2
|q(x) fn (x) − fm (x) | dx ≤ kfn − fm k∞ |q(x)|2 dx → 0, (2.3.3)
−N −N

so this can be made arbitrarily small for n, m large enough.

It remains to show that restricting the potential to the interval [−N, N ] does not affect
{qfn }R∞
n=1 being a Cauchy sequence. Take ε > 0 arbitrary, and choose a corresponding N such
y+1
that y |q(x)|2 dx < ε for all |y| > N . Denote by qN the function equal to q for |x| ≤ N and
qN (x) = 0 for |x| > N . In (2.3.3) we have dealt with qN (fn − fm ), the remaining part follows
from the following claim. There exists constants C2 , C3 independent of N such that

k(q − qN )f k2 ≤ ε C2 kf 00 k2 + C3 kf k2 , ∀ f ∈ W 2 (R).

The claims follows from the reasoning after Lemma 2.2.8 by replacing q and C by q − qN
and ε (taking e.g. the ε in the proof of Theorem 2.2.6 equal to 1, so we can take C2 = 21 ,
C3 = 52 ).

Exercise 2.3.6. Consider the Schrödinger operator with a potential of the form q1 + q2 , with
q1 (x) = a, q2 (x) = bχI (x), where χI is the indicator function of the set I (i.e. χI (x) = 1 if
x ∈ I and χI (x) = 0 if x ∈ / I) for I the interval I = [c, d]. Use Exercise 2.3.3 to describe the
essential spectrum, see also Section 2.5.1.

2.4 Bound states and discrete spectrum

2 d
We consider the Schrödinger operator − dx 2 +q for which we assume that the essential spectrum

is [0, ∞). So this is the case if q satisfies the condition in Theorem 2.3.4, which we now assume
throughout this section. This in particular means that any λ < 0 in the spectrum is isolated
and contained in the point spectrum by Theorem 6.5.5. In this case there is an eigenfunction
f ∈ W 2 (R) such that
−f 00 (x) + q(x) f (x) = λ f (x), λ < 0.
d 2
We count eigenvalues according to their multiplicity, i.e. the dimension of Ker(− dx 2 +q−λ). In

quantum mechanical applications an eigenfunction for the Schrödinger operator corresponding

to an eigenvalue λ < 0 is called a bound state.
For a Schrödinger operator with essential spectrum [0, ∞) the dimension of the space of
bound states might be zero, finite or infinite. Note that the number of negative eigenvalues
is denumerable by Theorem 6.5.5. We give a simple criterion that guarantees that there is at
Chapter 2: Schrödinger operators and their spectrum 17

least a one-dimensional space of bound states. There are many more explicit criterions on the
potential that imply explicit results for the dimension of the space of bound states.
We start with a general statement.

Proposition 2.4.1. Let (T, D(T )) be a self-adjoint operator on a Hilbert space H such that
σess (T ) ⊂ [0, ∞). Then T has negative eigenvalues if and only if there exists a non-trivial
subspace M ⊂ D(T ) ⊂ H such hT x, xi < 0 for all x ∈ M \{0}.
L
Proof. If T has negative eigenvalues, then put M = Ker(T − λ), summing over all 0 > λ ∈
σp (T ) and where we take only finite linear combinations.
If T has no negative eigenvalues, then it follows by Theorem 6.5.5(ii) that σ(T ) ⊂ [0, ∞).
By the Spectral Theorem 6.4.1, it follows that
Z
hT x, xi = λ dEx,x (λ), ∀x ∈ D(T ),
σ(T )

and since Ex,x , x 6= 0, is a positive measure and σ(T ) ⊂ [0, ∞) the right hand side is obviously
non-negative. So M is trivial.
This principle can be used to establish a criterion for the occurrence of bound states.

Theorem 2.4.2. Assume there exists α > 0 such that

Z r
1 2 2 π
q(x) exp(−2α (x − a) ) dx < − ,
α R 2
d 2
then the Schrödinger operator − dx 2 + q has a negative eigenvalue.

Proof. Take any functionR f ∈ W 2 (R) ⊂ D(Q), then h−f 00 , f i = hf 0 , f 0 i = kf 0 k2 . Hence,

h−f 00 + qf, f i = kf 0 k2 + R q(x)|f (x)|2 dx and if this expression is (strictly) negative, Propo-
sition 2.4.1 implies the result.
In order to obtain the result, we take f (x) = exp(−α2 (x − a)2 ), then f 0 (x) = −2α2 (x −
a) exp(−α2 (x − a)2 ), and

2α 1 √
r
4α4
Z Z
0 2 4 2 −2α2 (x−a)2 2 −y 2 π
kf k = 4α (x − a) e dx = √ y e dy = √ π=α ,
R
3
2α 2 R 22 2

so for this f we find

Z r Z 
0 2 2 π 1 −2α2 (x−a)2
kf k + q(x)|f (x)| dx = α + q(x)e dx
R 2 α R

which is negative by assumption for suitable α.

By applying this idea to other suitable functions, one can obtain other criteria for the
existence of negative eigenvalues for Schrödinger operators.
18 Chapter 2: Schrödinger operators and their spectrum
R
Corollary 2.4.3. Assume the existence of a (measurable) set B ⊂ R such that B q(x) dx < 0
d2
and q(x) ≤ 0 for x ∈
/ B. Then the Schrödinger operator − dx 2 + q has at least one negative

eigenvalue.

The implicit assumption in Corollary 2.4.3 is that q integrable is over the set B. Note
that the standing assumption in this section is the assumption of Theorem 2.3.4. If B is
R p R
1/2
a bounded set, then the Hölder inequality implies B |q(x)| dx ≤ |B| B |q(x)|2 dx , so
that the locally square integrability already implies this assumption. Here |B| denotes the
(Lebesgue) measure of the set B.

Proof. Note
Z Z Z
−2α2 x2 −2α2 x2
q(x)e dx ≤ q(x)e dx → q(x) dx < 0, as α ↓ 0,
R B B

2 2
so that α1 R q(x)e−2α x dx can be made arbitrarily negative, hence we can apply Theorem
R

2.4.2. Note that interchanging limit and integration is justifiable by the Dominated Conver-
gence Theorem 6.1.3.

In particular, the special case B = R of Corollary 2.4.3 gives the following result.
d2
R
Corollary 2.4.4. For q ∈ L1 (R) such that R q(x) dx < 0 the Schrödinger operator − dx 2 + q

has at least one negative eigenvalue.

We also give without proof an estimate on the number of negative eigenvalues for the
Schrödinger operator.

Theorem
R 2.4.5. Assume that q satisfies the condition in Theorem 2.3.4 and additionally that
R
|x| |q(x)| dx < ∞, then the number N of eigenvalues of the Schrödinger operator is bounded
by Z
N ≤2+ |x| |q(x)| dx.
R

The argument used in the proof of Theorem 2.4.5 is an extension of a comparison argument
for a Schrödinger operator on a finite interval, see [4].

2.5 Explicit examples of potentials

2.5.1 Indicator functions as potential
Consider first the potential q(x) = a ∈ R, then using Exercise 2.3.3 or writing q = aI, I
identity operator in B(L2 (R)), we see that the corresponding Schrödinger operator L has
σ = σess = [a, ∞). A somewhat more general statement is contained in the following exercise.
Chapter 2: Schrödinger operators and their spectrum 19

Exercise 2.5.1. Show that if the potential q satisfies q(x) ≥ a almost everywhere, then
(−∞, a) is contained in the resolvent ρ(L). Rephrased, the spectrum σ(L) is contained in
[a, ∞). Prove this using the following steps.

• Show that h(L − λ)f, f i = h−f 00 + qf − λf, f i = kf 0 k2 + R (q(x) − λ)|f (x)|2 dx, ∀f ∈
R

W 2 (R).

• Conclude that for λ < a, (a − λ)kf k ≤ k(L − λ)f k for all f ∈ W 2 (R). Use Theorem
6.5.1 to conclude that λ ∈ ρ(L).

Exercise 2.5.2. Assume that the potential q ∈ L2 (R) satisfies q(x) ≥ a for a < 0. Show
that [a, 0] contains only finitely many points of the spectrum of the corresponding Schrödinger
operator. (Hint: Use Theorem 6.5.5 and Exercise 2.5.1.)

Now consider q(x) = a + bχI (x), b ∈ R, where we take for convenience I = [0, 1], and
denote the corresponding Schrödinger operator again by L. By Exercise 2.3.3 we know that
σess (L) = [a, ∞). In case b > 0, we have σess (L) = [a, ∞) by combining Exercise 2.3.3 and
Theorem 2.3.4. Then Exercise 2.5.1 implies σ(L) = [a, ∞).
We now consider the case a = 0, b < 0. By Corollary 2.4.3 or Corollary 2.4.4, there is
negative point spectrum. On the other hand, the essential spectrum is [0, ∞) and by Exercise
2.5.1 the spectrum is contained in [b, ∞). We look for negative spectrum, which, by Theorem
6.5.5, is contained in the point spectrum and consists of isolated points. We put λ = −γ 2 ,
p
|b| > γ > 0 and we try to find solutions to Lf = λf , or

00 2
f − γ f = 0,
 x < 0,
00 2
f − (γ + b)f = 0, 0 < x < 1,

 00
f − γ 2 f = 0, x > 1.

The first and last equation imply f (x) = A− exp(γx) + B− exp(−γx), x < 0, and f (x) =
A+ exp(γx) + B+ exp(−γx), x > 1, for constants A± , B± ∈ C. Since we need f ∈ L2 (R),
we see that B− = 0 and A+ = 0, and because an eigenfunction (or bound state) can be
changed by multiplication by a constant, we take A− = 1, or f (x) = exp(γx), x < 0, and
f (x) = B+ exp(−γx), x > 1. Now b + γ 2 < 0, and we put −ω 2 = b + γ 2 , ω > 0. It is left as
an exercise to check that γ 2 = −b, i.e ω = 0, does not lead to an eigenfunction. So the second
equation gives f (x) = A cos(ωx) + B sin(ωx), 0 < x < 1. Since an eigenfunction f ∈ W 2 (R) ⊂
C 1 (R), we need to choose A, B, B+ such that f is C 1 at 0 and at 1. At 0 we need 1 = A
for continuity and γ = ωB for continuity of the derivative. So f (x) = cos(ωx) + ωγ sin(ωx),
0 < x < 1, and we need cos(ω) + ωγ sin(ω) = B+ e−γ for continuity at 1 –this then fixes B+ –
and −ω sin(ω) + γ cos(ω) = −γB+ e−γ for a continuous derivative at 1. In order that this has
a solution we require

γ2 2γω
−ω sin(ω) + γ cos(ω) = −γ cos(ω) − sin(ω) ⇒ tan(ω) = 2
ω ω − γ2
20 Chapter 2: Schrödinger operators and their spectrum

and this gives an implicit requirement on the eigenvalue λ = −γ 2 . Having a solution λ ∈

[b, 0], we see from the above calculation that the corresponding eigenspace is at most one-
dimensional, and that the corresponding eigenfunction or bound state is in C 1 (R). It remains
to check that this eigenfunction is actually an element of W 2 (R) in order to be able to conclude
that the corresponding λ is indeed an eigenvalue.
Exercise 2.5.3. Show that the number N of negative eigenvalues,
√ i.e. the number of solutions
to tan(ω) = ω2 −γ 2 , is determined by the condition N − 1 < π−b < N assuming b ∈
2γω
/ π 2 Z.
Show that the corresponding eigenfunction is indeed an element of the Sobolev space, so that
there are N eigenvalues. E.g. for b = −10, there are two bound states corresponding to
γ = 0.03244216751, γ = 2.547591633.
The assumption in Exercises 2.5.3 is inserted to avoid bound states for the eigenvalue 0.
Exercise 2.5.4. Using the techniques above, discuss the spectrum of the Schrödinger operator
with potential given by q(x) = a for x < x1 , q(x) = b for x1 < x < x2 , q(x) = c for x2 < x for
a, b, c ∈ R and x1 < x2 .

2.5.2 cosh−2 -potential

The potentials discussed consist of an important special case for the theory in Chapter 5.
They are closely related to soliton solutions of the Korteweg-de Vries equation, that are going
to be obtained using the inverse scattering method.
Recall the hyperbolic cosine function cosh(x) = 12 (ex + e−x ), the hyperbolic sine function
sinh(x) = 12 (ex − e−x ) and the relation cosh2 (x) − sinh2 (x) = 1. Obviously, dx d
cosh(x) =
d
sinh(x), dx sinh(x) = cosh(x).

We take the potential q(x) = a cosh−2 (px), a, p ∈ R, see Figure 2.1 for the case a = −2, p =
1. Since q ∈ L2 (R), it follows from Corollary 2.2.7 and Corollary 2.3.5 that the corresponding
d2
Schrödinger operator L =− dx 2 + q is self-adjoint and σess (L) = [0, ∞). Since for a > 0 we
0
R
have hLf, f i = kf k + R q(x)|f (x)|2 dx ≥ 0 for all f ∈ W 2 (R) it follows that σ(L) = [0, ∞) as
well. Since cosh−2 ∈ L1 (R), we can use Corollary 2.4.4 to see that for a < 0 there is at least
one negative eigenvalue.
We consider a first example in the following exercise.
Exercise 2.5.5. Show that f (x) = (cosh(px))−1 satisfies
2p2
−f 00 (x) − f (x) = −p2 f (x),
cosh2 (px)
or −p2 ∈ σp (L) for the case a = −2p2 .
In Exercise 2.5.8 you have to show that in this case σ = {−p2 }∪[0, ∞) and that Ker(L+p2 )
is one-dimensional.
Chapter 2: Schrödinger operators and their spectrum 21

x
-2 -1 0 1 2

-0.5

-1

-1.5

-2

Figure 2.1: The potential −2/ cosh2 (x).

Exercise 2.5.6. Consider a general potential q ∈ L2 (R) and the corresponding Schrödinger
operator L1 . Set fp (x) = f (px) for p 6= 0, and let Lp denote the Schrödinger operator with
potential qp . Show that λ ∈ σ(L1 ) ⇐⇒ p2 λ ∈ σ(Lp ) and similarly for the point spectrum and
the essential spectrum.

Because of Exercise 2.5.6 we can restrict ourselves to the case p = 1. We transform the
Schrödinger eigenvalue equation into the hypergeometric differential equation;


z(1 − z) yzz + c − (1 + a + b)z yz − ab y = 0 (2.5.1)

for complex parameters a, b, c. In Exercise 2.5.7 we describe solutions to (2.5.1). This exercise
is not essential, and is best understood within the context of differential equations on C with
regular singular points.
Put λ = γ 2 with γ real or purely imaginary, and consider −f 00 (x) + a cosh−2 (x)f (x) =
γ 2 f (x). We put z = 21 (1−tanh(x)) = (1+exp(2x))−1 , where tanh(x) = sinh(x)/ cosh(x). Note
dz −1
that −∞ is mapped to 1 and ∞ is mapped to 0, and x 7→ z is invertible with dx = 2 cosh 2
(x)
.
So we have to deal with (2.5.1) on the interval (0, 1).
Before transforming the differential equation, note that tanh(x) = 1 − 2z and cosh−2 (x) =
22 Chapter 2: Schrödinger operators and their spectrum

4z(1 − z). Now we put f (x) = (cosh(x))iγ y(z), then we have

1
f 0 (x) = iγ(cosh(x))iγ−1 sinh(x) y(z) − (cosh(x))iγ−2 yz (z),
2
00 iγ−2 2
f (x) = iγ(iγ − 1)(cosh(x)) sinh (x) y(z) + iγ(cosh(x))iγ y(z)
i 1
− γ(cosh(x))iγ−3 sinh(x) yz (z) − (iγ − 2)(cosh(x))iγ−3 sinh(x) yz (z)
2 2
1 iγ−4
+ (cosh(x)) yzz (z)
4
Plugging this into −f 00 (x) + a cosh−2 (x)f (x) = γ 2 f (x) and multiplying by (cosh(x))2−iγ gives,
−z(1 − z) yzz (z) + (1 − 2z)(iγ − 1) yz (z) + (a − γ 2 − iγ) y(z) = 0.
This equationqis of the hypergeometric
q differential type (2.5.1) if we set (a, b, c) in (2.5.1) equal
to ( 21 − iγ + 1
4
− a, 21 − iγ − 1
4
− a, 1 − iγ).
Note that under this transformation the eigenvalue equation −f 00 (x) + a cosh−2 (x)f (x) =
γ 2 f (x) is not transferred into another eigenvalue equation for the hypergeometric differential
equation, since γ also occurs in the coefficient of yz .
Exercise 2.5.7. 1. Define the Pochhammer symbol
Γ(a + n)
(a)n = a(a + 1)(a + 2) · · · (a + n − 1) = , a∈C
Γ(a)
so that (1)n = n! and (a)0 = 1. Define the hypergeometric function
  X ∞
a, b (a)n (b)n n
2 F1 ;z = z .
c n=0
(c)n n!
Show that the radius of convergence is 1 and that it gives an analytic solution to
(2.5.1). Check that
this is a well-defined function for c ∈
/ {· · · , −3, −2, −1, 0}, and
1 a, b
that Γ(c) F
2 1 c
; z is well-defined for c ∈ C.
2. Show that  
1−c a − c + 1, b − c + 1
z 2 F1 ;z
2−c



is also a solution to (2.5.1). (We define z 1−c = exp (1 − c) ln |z| + i arg(z) with
| arg(z)| < π, so that the complex plane is cut along the negative real axis.) Show that
for c 6∈ Z we have obtained two linearly independent solutions to (2.5.1).
3. Rewrite (2.5.1) by switching to w = 1−z, and observe that this is again a hypergeometric
differential equation. Use this to observe that
   
a, b c−a−b c − a, c − b
2 F1 ;1 − z , (1 − z) 2 F1 ;1 − z
a+b+1−c c+1−a−b
are also solutions to (2.5.1). These are linearly independent solutions in case c−a−b ∈
/ Z.
Chapter 2: Schrödinger operators and their spectrum 23

4. Use Exercise 2.5.5 to obtain an expression for cosh−1 (x) in terms of a hypergeometric
function.
Since we are interested in negative eigenvalues we assume that a < 0 and λ = γ 2 < 0; we
put γ = iβ with β > 0. Then the eigenfunction f satisfies
1 1
|f (x)|2 = 2β
|y( )|2 ,
cosh (x) 1 + e2x
so that square integrability
R∞ for x → ∞ gives a condition on y at 0. We call a function f square
integrable at ∞ if a |f (x)|2 dx < ∞ for some a ∈ R. Note that any f ∈ L2 (R) is square
integrable at ∞. Square integrability at −∞ is defined analogously. Note that 2 F1 a,c b ; z

equals 1 at z = 0, so is bounded and the corresponding eigenfunction is then square integrable

at ∞. For the other solution we get
  2
2 1 
2x 2β  a − c + 1, b − c + 1 1 
|f (x)| = 2β
|(1 + e )|  2 F1 ; 2x

cosh (x) 2−c 1+e 

(with a, b, c of the hypergeometric function related to a, γ = iβ as above) and this is not

square integrable for x → ∞. So we conclude that the eigenfunction for the eigenvalue −β 2
that is square integrable for x → ∞ is a multiple of
1
+ β + ( 14 − a)1/2 , 21 + β − ( 14 − a)1/2

1 2
1
f+ (x) = 2 F1 ; . (2.5.2)
coshβ (x) 1+β 1 + e2x
This already implies that the possible eigenspace is at most one-dimensional, since there is a
one-dimensional space of eigenfunctions that is square integrable at ∞.
Similarly, using the solutions of Exercise 2.5.7(3) we can look at eigenfunctions that are
square integrable for x → −∞, and we see that these are a multiple of
1
+ β + ( 14 − a)1/2 , 21 + β − ( 14 − a)1/2 e2x

1 2
f− (x) = 2 F1 ; . (2.5.3)
coshβ (x) 1+β 1 + e2x
So we actually find an eigenfunction in case f+ is a multiple of f− . Since the four hypergeo-
metric functions in Exercise 2.5.7 are solutions to the same second order differential equation,
there are linear relations between them. We need,
     
a, b a, b 1−c a − c + 1, b − c + 1
2 F1 ; 1 − z = A 2 F1 ; z + B z 2 F1 ;z ,
a+b−c+1 c 2−c
Γ(a + b + 1 − c)Γ(1 − c) Γ(a + b + 1 − c)Γ(c − 1)
A = , B = .
Γ(a − c + 1)Γ(b − c + 1) Γ(a)Γ(b)
(2.5.4)
So we find that f− is a multiple of f+ in case the B in (2.5.4) vanishes, or
Γ(1 + β)Γ(β)
= 0.
Γ( 12 +β+ ( 14 − a)1/2 )Γ( 12 + β − ( 41 − a)1/2 )
24 Chapter 2: Schrödinger operators and their spectrum

This can only happen if the Γ-functions in the denominator have poles. Since the Γ-functions
q at the {· · · , −2, −1, 0} and β > 0 and a < 0 we see q
have poles that this can only happen if
1
2
+ β − 14 − a = −n ≤ 0 for n a non-negative integer, or β = 14 − a − 21 − n.
It remains to check that in this case the eigenfunctions are indeed elements of the Sobolev
space W 2 (R). We leave this to the reader.
Exercise 2.5.8. Show that the number of (strictly) negative eigenvalues of the Schrödinger
operator for the potential a/ cosh2 (x) is given by the integer N satisfying N (N − 1) < −a <
N (N + 1). Show that for the special case −a = m(m + 1), m ∈ N, one of the points β
corresponds to zero, and in this case N = m. Conclude that the spectrum of the Schrödinger
operator in Exercise 2.5.5 has spectrum {−p2 } ∪ [0, ∞) and that there is a one-dimensional
space of eigenfunctions, or bound states, for the eigenvalue −p2 .

2.5.3 Exponential potential

We consider the potential q(x) = a exp(−2|x|), then it is clear from Corollary 2.2.7 and
Corollary 2.3.5 that the Schrödinger operator is self-adjoint and that its essential spectrum is
[0, ∞), and for a > 0 we see from Exercise 2.5.1 that its spectrum is [0, ∞) as well. See Figure
2.2 for the case a = −2, and compare with Figure 2.1.

The eigenvalue equation f 00 (x)+(λ−a exp(−2|x|))f (x) = 0 can be transformed into a well-
known differential equation. For the transformation
√ we consider the intervals√(−∞, 0) and
(0, ∞) separately. For x < 0 we put z = −a exp(x), and for x > 0 we put z = −a exp(−x)
and put y(z) = f (x), then the eigenvalue equation is transformed into the Bessel differential
equation
z 2 yzz + z yz + (z 2 − ν 2 ) y = 0,
2
√
where
√ we have put λ = −ν . Note that (−∞, 0) is transformed into √ −a) and1 (0, ∞) into
(0,
( −a, 0), and so we need to glue solutions together at x = 0 or z = −a in a C -fashion.
Exercise 2.5.9. 1. Check the details of the above transformation.
2. Define the Bessel function
∞ ∞
(z/2)ν X (−1)n z 2n X (−1)n z
ν+2n
Jν (z) = = .
Γ(ν + 1) n=0 (ν + 1)n 4n n! n=0 n! Γ(ν + n + 1) 2
Show that the power series in the middle defines an entire function. Show that Jν (z) is
a solution to Bessel differential equation. Conclude that J−ν is also a solution.
3. Show that for ν ∈ / Z the Bessel functions Jν and J−ν are linearly independent solutions
to the Bessel functions. (Hint: show that for any pair of solutions y1 , y2 of the Bessel
differential equation the Wronskian W (z) = y1 (z)y20 (z) − y10 (z)y2 (z) satisfies a first order
differential equation that leads to W (z) = C/z. Show that for y1 (z) = Jν (z), y2 (z) =
J−ν (z) the constant C = −2 sin(νπ)/π.)
Chapter 2: Schrödinger operators and their spectrum 25

x
-2 -1 0 1 2
0

-0.5

y -1

-1.5

-2

Figure 2.2: The potential −2 exp(−2|x|).

The reduction to the Bessel differential equation works for arbitrary a, but we now stick
to the case a < 0. In case a < 0 we want to discuss the discrete spectrum. We consider
2
eigenfunctions f for √ eigenvalue λ = −ν < 0, so we have ν ∈ R and we assume ν > 0.
We now put c = −a. For negative x it follows |f (x)| = |y(cex )|2 , and since Jν (z) =
2
(z/2)ν

Γ(ν+1)
1 + O(z) as z → 0 for ν ∈ / −N, it follows that f is a multiple of Jν (cex ) for x < 0
in order to be square integrable at ∞. In order to extend f to a function for x > 0, we put
f (x) = A Jν (ce−x ) + B J−ν (ce−x ), so that it is a solution to the eigenvalue equation for x > 0.
Since f has to be in W 2 (R) ⊂ C 1 (R) we need to impose a C 1 -transition at x = 0, this gives
(     
A Jν (c) + B J−ν (c) = Jν (c), Jν (c) J−ν (c) A Jν (c)
=⇒ =
−cA Jν0 (c) − cB J−ν
0
(c) = cJν0 (c) Jν0 (c) J−ν0
(c) B −Jν0 (c)
   0  
A π J−ν (c) −J−ν (c) Jν (c)
=⇒ =− ,
B 2 sin(νπ) −Jν0 (c) Jν (c) −Jν0 (c)

since the determinant of the matrix is precisely the Wronskian of the Bessel functions Jν and
J−ν , see Exercise 2.5.9(3), and where we now assume ν ∈ / Z.
2 −x 2
Similarly, for x > 0 we have |f (x)| = |y(ce )| is square integrable at ∞ if only if y
equals Jν , so that f is a multiple of Jν (ce−x ) for x > 0 in order to be square integrable. So
we can only have a square integrable eigenfunction (or bound state) in case B in the previous
26 Chapter 2: Schrödinger operators and their spectrum

calculation vanishes, or Jν (c)Jν0 (c) = 0. Given a, hence c, we have to solve this equation for ν
yielding the corresponding eigenvalues λ = −ν 2 for the Schrödinger operator. This equation
cannot be solved easily in a direct fashion, and requires knowledge about Bessel functions,
its derivatives and its zeros. Since the negative spectrum has to be contained in [−c2 , 0] we
conclude that the smallest positive zero of ν 7→ Jν (c) and ν 7→ Jν0 (c) is less than c. (This is a
classical result, due to Riemann.)
Chapter 3

Scattering and wave operators

3.1 Time-dependent Schrödinger equation

Consider a self-adjoint Schrödinger operator L acting on L2 (R). The time-dependent Schrö-
dinger equation is
iψ 0 (t) = Lψ(t), ψ : D ⊂ R → L2 (R)
on some time domain D. The derivative with respect to time t is defined by
1
ψ(t + h) − ψ(t) − ψ 0 (t)k = 0


lim k
h→0 h
whenever the limit exists. The solution of the time-dependent Schrödinger equation can be
solved completely using the Spectral Theorem 6.4.1 and the corresponding functional calculus.
Theorem 3.1.1. Assume L is a self-adjoint operator on L2 (R), then the system

iψ 0 (t) = Lψ(t), ψ(0) = ψ0 ∈ D(L) (3.1.1)

has a unique solution ψ(t), t ∈ R, given by ψ(t) = exp(−itL)ψ0 which satisfies kψ(t)k = kψ0 k.
Proof. Let us first show existence. We use the Spectral Theorem 6.4.1 to establish the operator
U (t) = exp(−itL), which is a one-parameter group of unitary operators on the Hilbert space
L2 (R). Define ψ(t) = U (t)ψ0 . This is obviously defined for all t ∈ R and ψ(0) = U (0)ψ0 = ψ0 ,
since U (0) = 1 in B(H). Note that this can be defined for any ψ0 , i.e. for this construction
the requirement ψ0 ∈ D(L) is not needed.
In order to show that the differential equation is fulfilled we consider
1
U (t + h)ψ0 − U (t)ψ0 + iLU (t)ψ0 k2


k
h
and to seeR that U (t)ψ0 ∈ D(L) we note that, by the Spectral Theorem 6.4.1, it suffices to
note that R λ2 dEU (t)ψ0 ,U (t)ψ0 (λ) < ∞ which follows from

hE(A)U (t)ψ0 , U (t)ψ0 i = hU (t)E(A)ψ0 , U (t)ψ0 i = hE(A)ψ0 , ψ0 i,

27
28 Chapter 3: Scattering and wave operators

since ψ0 ∈ D(L), U (t) commutes with L and U (t) is unitary by the Spectral Theorem 6.4.1.
Since all operators are functions of the self-adjoint operator L, it follows that this equals
1 Z
−i(t+h)L −itL −itL
  1 −i(t+h)λ 2
2
− e−itλ + iλe−itλ  dEψ0 ,ψ0 (λ)



k e −e + iLe ψ0 k =  e
h R h
Z Z
 −itλ  1 −ihλ 
2  1 −ihλ
− 1 + iλ|2 dEψ0 ,ψ0 (λ).



=  e e − 1 + iλ dEψ0 ,ψ0 (λ) =
  e
R h R h

Since h1 e−ihλ − 1 + iλ → 0 as h → 0, we only need to show
that we can interchange the limit

with the integration. In order to do so we use | h1 e−iλh R− 1 | ≤ |λ|, so that the integrand can
be estimated by 4|λ|2 independent of h. So we require R |λ|2 dEψ0 ,ψ0 (λ) < ∞, or ψ0 ∈ D(L),
see the Spectral Theorem 6.4.1.
To show uniqueness, assume that ψ(t) and φ(t) are solutions to (3.1.1), then their difference
ϕ(t) = ψ(t) − φ(t) is a solution to iϕ0 (t) = Lϕ(t), ϕ(0) = 0. Consider

d d
kϕ(t)k2 = hϕ(t), ϕ(t)i = hϕ0 (t), ϕ(t)i + hϕ(t), ϕ0 (t)i
dt dt  
= h−iLϕ(t), ϕ(t)i + hϕ(t), −iLϕ(t)i = −i hLϕ(t), ϕ(t)i − hϕ(t), Lϕ(t)i = 0

since L is self-adjoint. So kϕ(t)k is constant and kϕ(0)k = 0, it follows that ϕ(t) = 0 for all t
and uniqueness follows.

It follows that the time-dependent Schrödinger equation is completely determined by the

time-independent self-adjoint Schrödinger operator L.

Exercise 3.1.2. We consider the setting of Theorem 3.1.1. Show that ψ(t) = e−iλt u, u ∈
L2 (R), λ ∈ R fixed, is a solution to the time-dependent Schrödinger equation (3.1.1) if and
only if u ∈ D(L) is an eigenvector of L for the eigenvalue λ.

3.2 Scattering and wave operators

2
d 2
Let L0 be the unperturbed Schrödinger operator − dx 2 with its standard domain W (R) and
d2
let L be the perturbed Schrödinger operator − dx 2 + q. We assume that L is self-adjoint, e.g. if

the potential q satisfies the conditions of Theorem 2.3.4. However, it should be noted that the
set-up in this section is much more general and works for any two operators L and L0 which
are (possibly unbounded) self-adjoint operators on a Hilbert space H, and in this situation
we now continue.

Definition 3.2.1. The solution ψ to (3.1.1) for a general self-adjoint operator (L, D(L)) on a
Hilbert space H has an incoming asymptotic state ψ − (t) = exp(−itL0 )ψ − (0) for a self-adjoint
operator (L0 , D(L0 )) on H if
lim kψ(t) − ψ − (t)k = 0.
t→−∞
Chapter 3: Scattering and wave operators 29

The solution ψ to (3.1.1) has an outgoing asymptotic state ψ + (t) = exp(−itL0 )ψ + (0) if

lim kψ(t) − ψ + (t)k = 0.

t→∞

The solution ψ to (3.1.1) is a scattering state if it has an incoming asymptotic state and an
outgoing asymptotic state.
Define the operator W (t) = eitL e−itL0 , which is an element in B(H). Note that t 7→ eitL and
t 7→ e−itL0 are one-parameter groups of unitary operators. In particular, they are isometries,
i.e. preserve norms. Also note that in general eitL e−itL0 6= eit(L−L0 ) , unless L and L0 commute.
Assuming the solution ψ to (3.1.1) has an incoming state, then we have

lim W (t)ψ − (0) = ψ(0).

t→−∞

To see that this is true, write

kW (t)ψ − (0) − ψ(0)k = keitL e−itL0 ψ − (0) − ψ(0)k = ke−itL0 ψ − (0) − e−itL ψ(0)k
= kψ − (t) − ψ(t)k → 0, t → −∞.

Similarly, if the solution ψ to (3.1.1) has an outgoing state, then we have

lim W (t)ψ + (0) = ψ(0).

t→∞

Definition 3.2.2. The wave operators W ± are defined as the strong limits of W (t) as t →
±∞. So its domain is

D(W ± ) = {f ∈ H | lim W (t)f exists}

t→±∞

and W ± f = limt→±∞ W (t)f for f ∈ D(W± ).

Exercise 3.2.3. In Definition 3.2.2 we take convergence in the strong operator topology. To
see that the operator norm is not suited, show that limt→∞ W (t) = V exists in operator norm
if and only if L = L0 , and in this case V is the identity. Show first that eitL V = V eitL0 for
all t ∈ R. (Hint: if V exists, observe that kei(s+t)L e−i(s+t)L0 − eisL V e−isL0 k = keitL e−itL0 −
V k → 0 as t → ∞, and hence, by uniqueness of the limit, V = eisL V e−isL0 . Conclude that
keitL e−itL0 − V k = k1 − V k.)
Since W (t) is a composition of unitaries, it is in particular an isometry. So kW (t)f k = kf k,
so that the wave operators W ± are partial isometries. The wave operators relate the initial
values of the incoming, respectively outgoing, asympotic states for the unperturbed problem
to the initial value for the perturbed problem. So D(W − ), respectively D(W + ), consists
of initial values for the unperturbed time-dependent Schrödinger equation that occur as in-
coming, respectively outgoing, asymptotic states for solutions to (3.1.1). Its range Ran(W − ),
respectively Ran(W + ), consists of initial values for the perturbed time-dependent Schrödinger
equation (3.1.1) that have an incoming, respectively outgoing, asymptotic state.
30 Chapter 3: Scattering and wave operators

Note that for a scattering state ψ, the wave operators completely determine the incoming
and outgoing asymptotic states. Indeed, assume W ± ψ ± (0) = ψ(0), then we can define ψ ± (t) =
exp(−itL0 )ψ ± (0) and ψ(t) = exp(−itL)ψ(0). It then follows that

lim kψ(t) − ψ ± (t)k = lim kψ(0) − W (t)ψ ± (0)k = 0,

t→±∞ t→±∞

so that ψ ± (t) are the incoming and outgoing asymptotic states. In physical models it is
important to have as many scattering states as possible, preferably Ran(W ± ) is the whole
Hilbert space H. Note that scattering states have initial values in Ran W + ∩ Ran W − .
Proposition 3.2.4. Assume that L has an eigenvalue λ for the eigenvalue u ∈ D(L), and
consider the solution ψ(t) = exp(−iλt)u to (3.1.1), see Exercise 3.1.2. Then ψ has no (incom-
ing or outgoing) asymptotic states unless u is an eigenvector for L0 for the same eigenvalue
λ.
Proof. Assume ψ(t) = exp(−iλt)u has an incoming asymptotic state, so that there exists a
v ∈ H such that W − v = u or limt→−∞ W (t)v = u = ψ(0). This means k exp(−itL0 )v −
exp(−itL)uk = k exp(−itL0 )v − exp(−itλ)uk → 0 as t → −∞, and using the isometry
property this gives k v − exp(−it(λ − L0 ))uk → 0 as t → −∞. For s ∈ R fixed we get

ke−i(t+s)(λ−L0 ) u − e−it(λ−L0 ) uk ≤ ke−i(t+s)(λ−L0 ) u − vk + kv − e−it(λ−L0 ) uk → 0, t → −∞,

and so ke−is(λ−L0 ) u − uk → 0 as t → −∞, but since the expression is independent of t, we

actually have e−is(λ−L0 ) u = u for arbitrary s, so that by Exercise 3.1.2, we have L0 u = λu. A
similar reasoning applies to outgoing states.
Exercise 3.2.5. Put W (L, L0 ) as the strong operator limit of eitL e−itL0 to stress the depen-
dence on the self-adjoint operators L and L0 . Investigate how W ± (L0 , L) and W ± (L, L0 ) are
related. Show also that W ± (A, B) = W ± (A, C)W ± (C, B) for self-adjoint A, B, C ∈ B(H)
assuming that the domains of all wave operators involved equal the Hilbert space H. (See
also Proposition 3.4.4.)
Theorem 3.2.6. D(W ± ) and Ran(W ± ) are closed subspaces of H.
Proof. Consider a convergent sequence {fn }∞ +
n=1 , fn → f in H with fn ∈ D(W ). We need to
+
show that f ∈ D(W ). Recall that W (t) ∈ B(H) and consider


kW (t)f − W (s)f k ≤ kW (t)(f − fn )k + kW (s)(fn − f )k + k W (t) − W (s) fn k


= 2 kf − fn k + k W (t) − W (s) fn k

using that W (t) and W (s) are isometries. For ε > 0 arbitrary, we can find N ∈ N such
that for n ≥ N we have kf − fn k ≤ 2ε , since fn → f in H. And since fN ∈ D(W + ) we
have
limt→∞ W (t)fN exists, so there exists T > 0 such that for s, t ≥ T we have k W (t) −
W (s) fN k ≤ 2ε . So kW (t)f − W (s)f k ≤ ε, hence limt→∞ W (t)f exists, and f ∈ D(W + ).
Similarly, D(W − ) is closed.
Chapter 3: Scattering and wave operators 31

Next take a convergent sequence {fn }∞ + +

n=1 , fn ∈ Ran(W ), fn → f in H. Take gn ∈ D(W )
with W + gn = fn , then

kgn − gm k = kW + (gn − gm )k = kfn − fm k,

since W + is an isometry. So {gn }∞ +

n=1 is a Cauchy sequence in D(W ) ⊂ H, hence convergent
to, say, g ∈ D(W + ), since D(W + ) is closed. Now

kfn − W + gk = kW + (gn − g)k = kgn − gk → 0, n → ∞,

so f = W + g ∈ Ran(W + ). Similarly, Ran(W − ) is closed.

So a scattering state ψ has an incoming state and an outgoing state, and their respective
values for t = 0 are related by the wave operators;

W + ψ + (0) = ψ(0) = W − ψ − (0).

Since the wave operator W − is an isometry from its domain to its range, it is injective, and
we can write
ψ + (0) = Sψ − (0), S = (W + )−1 W − .
S is the scattering operator; it relates the incoming asymptotic state ψ − with the outgoing
asymptotic state ψ + for a scattering state ψ. Note that S = (W + )−1 W − with

D(S) = {f ∈ D(W − ) | W − f ∈ Ran(W + )}.

Theorem 3.2.7. D(S) = D(W − ) if and only if Ran(W − ) ⊂ Ran(W + ) and Ran(S) = D(W + )
if and only if Ran(W + ) ⊂ Ran(W − ).

Corollary 3.2.8. S : D(W − ) → D(W + ) is unitary if and only if Ran(W + ) = Ran(W − ).

So we can rephrase Corollary 3.2.8 as S is unitary if and only if all solutions with an
incoming asymptotic state also have an outgoing asymptotic state and vice versa. Loosely
speaking, there are only scattering states.
Proof of Theorem 3.2.7. Consider the first statement. If Ran(W − ) ⊂ Ran(W + ), then D(S) =
{f ∈ D(W − ) | W − f ∈ Ran(W + )} = D(W − ) since the condition is true. Conversely, if
D(S) = D(W − ), we have by definition W − f ∈ Ran(W + ) for all f ∈ D(W − ), or Ran(W − ) ⊂
Ran(W + ).
For the second statement, note that Sf = g means f ∈ D(S) and W + g = W − f , so
Ran(S) = {g ∈ D(W + ) | W + g ∈ Ran(W − )}. So if Ran(W + ) ⊂ Ran(W − ), the condition is
void and Ran(S) = D(W + ). Conversely, if Ran(S) = D(W + ), we have W + g ∈ Ran(W − ) for
all g ∈ D(W + ) or Ran(W + ) ⊂ Ran(W − ).
The importance of the wave operators is that they can be used to describe possible unitary
equivalences of the operators L and L0 .
32 Chapter 3: Scattering and wave operators

Exercise 3.2.9. Let L, L0 be (possibly unbounded) self-adjoint operators with LU = U L0

for some unitary operator U . Show that all spectra σ, σp , σpp , σess , σac , σsc for L and L0 are
the same. See Chapter 6, and in particular Section 6.6 for the notation.
Such self-adjoint operators which are unitarily equivalent have the same spectrum, knowing
the wave operators plus the spectral decomposition of one of the operators gives the spectral
decomposition of the other operator. However, the situation is not that nice, since the wave
operators may not be defined on the whole Hilbert space or the range of the wave operators
may not be the whole Hilbert space. So we need an additional definition, see also Section 6.2
for unexplained notions and notation.
Definition 3.2.10. A closed subspace V of an Hilbert space H is said to reduce, or to be a
reducing subspace for, the operator (T, D(T )) if the orthogonal projection P onto V preserves
the domain of T , P : D(T ) → D(T ), and commutes with T , P T ⊂ T P .
Note that for a self-adjoint operator (T, D(T )) and a reducing subspace V , the orthocom-
plement V ⊥ is also a reducing subspace for (T, D(T )).
We first discuss reduction of the bounded operators e−itL0 and e−itL , and from this, using
Stone’s Theorem 6.4.2, for the unbounded self-adjoint operators L0 and L.
Theorem 3.2.11. For each t ∈ R we have
e−itL0 : D(W + ) → D(W + ), e−itL0 : D(W − ) → D(W − ),
e−itL : Ran(W + ) → Ran(W + ), e−itL : Ran(W − ) → Ran(W − )
and
W + e−itL0 = e−itL W + , W − e−itL = e−itL0 W − .

∗
∗
Since e−itL0 = eitL0 and e−itL = eitL by the self-adjointness of L and L0 , we obtain

e−itL0 : D(W + )⊥ → D(W + )⊥ , e−itL0 : D(W − )⊥ → D(W − )⊥ ,

e−itL : Ran(W + )⊥ → Ran(W + )⊥ , e−itL : Ran(W − )⊥ → Ran(W − )⊥ .

By Theorem 3.2.6 the subspaces D(W ± ), Ran(W ± ) are closed, and since orthocomple-
ments are closed as well, we can rephrase the first part of Theorem 3.2.11.
Corollary 3.2.12. D(W ± ) and D(W ± )⊥ reduce e−itL0 and Ran(W ± ) and Ran(W ± )⊥ reduce
e−itL .
Proof of Theorem 3.2.11. Take f ∈ D(W + ) and put g = W + f and consider

W (t)e−isL0 f − e−isL g = eitL e−itL0 e−isL0 f − e−isL g = eitL e−i(s+t)L0 f − e−isL g

= e−isL ei(s+t)L e−i(s+t)L0 f − e−isL g = e−isL W (t + s)f − e−isL g = e−isL W (t + s)f − g

and since e−isL is an isometry we find, for s ∈ R fixed,

kW (t)e−isL0 f − e−isL gk = kW (t + s)f − gk → 0, t → ∞.

Chapter 3: Scattering and wave operators 33

In particular, e−isL0 f ∈ D(W + ) and W + e−isL0 f = e−isL g = e−isL W + f . This proves that
e−isL0 preserves D(W + ) and that e−isL preserves Ran(W + ) and W + e−isL0 = e−isL W + .
The statement for W − is proved analogously.
In order to see that these spaces also reduce L and L0 we need to consider the orthogonal
projections on D(W ± ) and Ran(W ± ) in relation to the domains of L and L0 . We use the
characterisation of the domain of L, respectively L0 , as those
f ∈ H for which the limit
limt→0 t exp(−itL)f −f , respectively limt→0 1t exp(−itL0 )f −f , exists, see Stone’s Theorem
1

6.4.2.

Theorem 3.2.13. D(W + ) and D(W − ) reduce L0 and Ran(W + ) and Ran(W − ) reduce L.

Proof. We show that first statement. Let P + : H → H be the orthogonal projections on

D(W + ). Then we have to show first that P + preserves the domain D(L0 ). Take f ∈ D(L0 ),
then for all t ∈ R\{0} we have
1 −itL0 + 1

P f − P +f = P + e−itL0 f − f


e
t t
by Corollary 3.2.12. For f ∈ D(L0 ) the right hand side converges as t → 0 by Stone’s Theorem
6.4.2 and continuity of P + . Hence the left hand side converges. By Stone’s Theorem 6.4.2 it
follows that P + f ∈ D(L0 ).
For f ∈ D(L0 ) we multiply this identity by i and take the limit t → 0. Again by Stone’s
Theorem 6.4.2 it follows that L0 P + f = P + L0 f for all f ∈ D(L0 ). Since D(L0 ) = D(P + L0 ) ⊂
D(L0 P + ) = {f ∈ H | P + f ∈ D(L0 )} we get P + L0 ⊂ L0 P + .

Exercise 3.2.14. Prove the other cases of Theorem 3.2.13.

In view of Theorems 3.2.11 and 3.2.13 we may expect that the wave operators intertwine
L and L0 . This is almost true, and this is the content of the following Theorem 3.2.15.

Theorem 3.2.15. Let P be the orthogonal projection on Ker(L0 )⊥ , then LW + P = W + L0

and LW − P = W − L0 .

Note that this is a statement for generally unbounded operators, so that this statement
also involves the domains of the operators. In case L0 has trivial kernel, we see that W +
and W − intertwine the self-adjoint operators. In particular, in case the wave operators are
unitary, we see that L and L0 are unitarily equivalent, and so by Exercise 3.2.9 have the same
spectrum.
Proof. Observe first that Ker(L0 )⊥ = Ran(L0 ). Indeed, we have for arbitrary f ∈ Ker(L0 )
and for arbitrary g ∈ D(L0 ) the identity 0 = hL0 f, gi = hf, L0 gi, since L0 is self-adjoint.
Next we note that, with the notation P ± for the orthogonal projection on the domains
D(W ± ) for the wave operators as in Theorem 3.2.13, we have P P ± P = P ± P . Indeed, this is
obviously true on Ker(L0 ) since both sides are zero. For any f ∈ Ran(L0 ) put f = L0 g, so that
P P + P f = P P + f = P P + L0 g = P L0 P + g = L0 P + g = P + L0 g = P + f = P + P f by Theorem
34 Chapter 3: Scattering and wave operators

3.2.13 (twice) and the observation P L0 = L0 . So the result follows for any f ∈ Ran(L0 ),
and since orthogonal projections are bounded operators the result follows for Ran(L0 ) by
continuity of the projections. The case for P − is analogous.
We first show that LW + P ⊃ W + L0 . So take f ∈ D(W + L0 ) = {f ∈ D(L0 ) | L0 f ∈
D(W + )}. Then f = P f +(1−P )f , so that f ∈ D(L0 ), (1−P )f ∈ Ker(L0 ) ⊂ D(L0 ) shows that
P f ∈ D(L0 ). Since with D(W + ), also D(W + )⊥ , reduces L0 , we see that (1 − P + )P f ∈ D(L0 )
and
L0 (1 − P + )P f = (1 − P + )L0 P f = (1 − P + )L0 f
since L0 P f = L0 f + L0 (1 − P )f = L0 f as 1 − P projects onto Ker(L0 ). Since f ∈ D(W + L0 ),
we have L0 f ∈ D(W + ), so that (1 − P + )L0 f = 0. We conclude that (1 − P + )P f ∈ Ker(L0 )
and P (1 − P + )P f = 0 or P f = P 2 f = P P + P f = P + P f , where the last equality follows from
the second observation in this proof. Now P f = P + P f says P f ∈ D(W + ) or f ∈ D(W + P ).
As a next step we show that W + P f ∈ D(L), and for this we use Stone’s Theorem 6.4.2.
So, using Theorem 3.2.11,
1 −itL 1
− 1 W + P f = W + e−itL0 − 1 P f.



e (3.2.1)
t t
Since P f ∈ D(L0 ), limt→0 1t e−itL − 1 P f exists and since the domain of W + is closed it

follows that the right hand side has a limit −iW + LP f as t → 0. So the left hand side has a
limit, and by Stone’s Theorem 6.4.2, we have W + P f ∈ D(L) and the limit is −iLW + P f . So
f ∈ D(LW + P ) and LW + P f = W + L0 P f , and since L0 P f = L0 f we have W + L0 ⊂ LW + P .
Conversely, to show LW + P ⊂ W + L0 , take f ∈ D(LW + P ), or P f ∈ D(W + ) and W + P f ∈
D(L). So, again by Stone’s Theorem 6.4.2, limt→0 1t (e−itL − 1)W + P f exists, and since (3.2.1)
is valid, we see that W + 1t (e−itL0 − 1)P f converges to, say, g = −iLW + P f . Since W + is
continuous, g ∈ Ran(W + ) = Ran(W + ) by Theorem 3.2.6. So g = W + h for some h ∈ H, and
1 −itL0 1 1
− 1 P f − hk = kW + e−itL0 − 1 P f − W + hk = kW + e−itL0 − 1 P f − gk → 0,




k e
t t t
as t → ∞. Again, by Stone’s Theorem 6.4.2, P f ∈ D(L0 ) and −iL0 P f = h, and thus
W + L0 P f = iW + h = ig = LW + P f . As before, with P f ∈ D(L0 ) it follows f ∈ D(L0 )
and L0 P f = L0 f , so that we have f ∈ D(W + L0 ) and LW + P f = W + L0 f . This proves
LW + P ⊂ W + L.
The case W − is analogous.
d
Exercise 3.2.16. Consider the operator L0 = i dx with domain W 1 (R), then L0 is an un-
d
bounded self-adjoint operator on L2 (R), see Lemma 2.1.2. For L we take i dx + q, for some
potential function q.


• Show that U (t) = eitL0 is a translation operator, i.e. U (t)f (x) = f (x − t).
• Define M to be the multiplication operator by a function m. Show that for im0 = qm
we have L0 M = M L. What conditions on q imply that M : L2 (R) → L2 (R) is a
unitary operator? Give a domain D(L) such that (L, D(L)) is an unbounded self-adjoint
operator on L2 (R).
Chapter 3: Scattering and wave operators 35

• Assume these conditions on m. Conclude that eitL = M ∗ eitL0 M = M ∗ U (t)M , and

W (t) = M ∗ U (t)M U (−t) is a multiplication operator.

• What conditions on q ensure that W ± exist? What are W ± in this case? Describe the
scattering operator S = (W + )−1 W − as well in this case. (Hint: S is a multiple of the
identity.)

3.3 Domains of the wave operators for the Schrödinger

operator
In Sections 3.1, 3.2 the situation was completely general. In this section we discuss conditions
d2 ±
on the potential q of the Schrödinger equation L =− dx 2 +q, such that D(W ) equal the whole

Hilbert space L2 (R).

The idea of the method, due to Cook, is based on the observation that for a f ∈ C 1 (R)
with integrable derivative, f 0 ∈ L1 (R), the limit limt→∞ f (t) exists. This follows from the
estimate Z t Z t
0
|f (t) − f (s)| = | f (x) dx| ≤ |f 0 (x)| dx → 0
s s

for s < t and s, t tend to ∞. Similarly, the limit limt→−∞ f (t) exists. Cook’s idea is also used
in Theorem 3.4.9. This gives rise to the following description.

Proposition 3.3.1. Assume that exp(−itL0 )f ∈ D(L0 ) ∩ D(L) for all t ≥ a for some a ∈ R,
and that Z ∞
k(L − L0 )e−itL0 f k dt < ∞,
a
+
then f ∈ D(W ). Similarly, if exp(−itL0 )g ∈ D(L0 ) ∩ D(L) for all t ≤ b for some b ∈ R,
and that Z b
k(L − L0 )e−itL0 gk dt < ∞,
−∞

then g ∈ D(W − ).

Proof. Recalling W (t)f = eitL e−itL0 f , we see that

W 0 (t)f = iL eitL e−itL0 f − ieitL L0 e−itL0 f,

where we need exp(−itL0 )f ∈ D(L), f ∈ D(L0 ), cf. Theorem
3.1.1. Since this shows
exp(−itL0 )f ∈ D(L) ∩ D(L0 ) we have W 0 (t)f = i exp(itL) L − L0 exp(−itL0 )f and hence
kW 0 (t)f k = k(L − L0 ) exp(−itL0 )f k.
Apply now the previous observation to get for arbitrary u ∈ L2 (R),
Z t Z t
0
|hW (t)f, ui − hW (s)f, ui| ≤ |hW (x)f, ui| dx ≤ kuk kW 0 (x)f k dx,
s s
36 Chapter 3: Scattering and wave operators

which shows that

Z t Z t
0
kW (t)f − W (s)f k ≤ kW (x)f k dx ≤ k(L − L0 ) exp(−ixL0 )f k dx.
s s

By assumption for t > s ≥ a, the right hand side integrated over [a, ∞) is finite. Hence,
limt→∞ W (t)f exists, or f ∈ D(W + ).
Theorem 3.3.2. Let L be the Schrödinger operator, where the potential q satisfies the con-
d2 2
ditions of Corollary 2.2.7, and L0 is the unperturbed operator − dx 2 with domain W (R). If

x 7→ (1 + |x|)α q(x) is an element of L2 (R) for some α > 12 , then D(W + ) = L2 (R) = D(W − ).
The following gives a nice special case of the theorem.
Corollary 3.3.3. Theorem 3.3.2 holds true if the potential q is locally square integrable and
for some β > 1 the fuction x 7→ |x|β q(x) is bounded for |x| → ∞.
Exercise 3.3.4. Prove Corollary 3.3.3 from Theorem 3.3.2.
Proof of Theorem 3.3.2. The idea is to show that we can use Proposition 3.3.1 for sufficiently
many functions. Consider the function fs (λ) = λ exp(−λ2 − isλ) for s ∈ R. Then fs ∈
d2
L1 (R)∩L2 (R). We want to exploit the fact that the Fourier transform diagonalises L0 = − dx 2,

see Theorem 2.1.1 and its proof. We consider us = F −1 fs , then

Z
−itL0 −1 −itλ2 1 2
eiλ(x−s)−λ (1+it) λdλ



e us (x) = F e fs (λ) (x) = √
2π R
−(x−s)2
Z
exp 4(1+it) 2
 z i(x − s)  dz
= √ e−z √ + √
2π C 1 + it 2(1 + it) 1 + it
√ √ p 1
1 + it λ − 2i 1+it
x−s



by switching to z = . Here the square root is z = |z|ei 2 arg z for
| arg z| < π, and the contour C in the complex plane is a line corresponding to the image of
the real line under this coordinate transformation.
2 √ R 2
Lemma 3.3.5. C e−z dz = π, C e−z zdz = 0.
R

Using Lemma 3.3.5 we see that

i(x − s) −(x − s)2

e−itL0 us (x) = √



3 exp ,
2 + 2it 4(1 + it)
so
 −itL
 |x − s| −(x − s)2

 e 0
us (x) ≤
3/4 exp .
4 + 4t2 4(1 + t2 )
Now we have
|q(x)(x − s)|2 −(x − s)2

Z
−itL0 2 −itL0 2
k(L − L0 )e us k = kqe us k =
3/2 exp dx,
R 4 + 4t2 2(1 + t2 )
Chapter 3: Scattering and wave operators 37

which we have to integrate with respect to t in order to be able to apply Proposition 3.3.1.
Using the estimate e−x ≤ (1 + xa )−a for x ≥ 0 and a ≥ 0, we get

−(x − s)2
 (x − s)2 −a a 2 a 2 2 −a

exp ≤ 1 + = (2a) (1 + t ) 2a(1 + t ) + (x − s)
2(1 + t2 ) 2a(1 + t2 )
≤ (2a) (1 + t2 )a |x − s|−2a
a

so that Z
−itL0 2 2 a− 23
k(L − L0 )e us k ≤ C(1 + t ) |q(x)|2 |x − s|2−2a dx
R

and noting |x − s| ≤ (1 + |x|)(1 + |s|) and inserting this estimate gives, assuming a ≤ 1,
Z
−itL0 2 2 a−3/2
k(L − L0 )e us k ≤ C(1 + t ) |q(x)|2 (1 + |x|)2−2a dx
R

with the constant C independent of x and t, but depending on s and a. Hence,

Z Z Z 1/2
−itL0 2 a/2−3/4
k(L − L0 )e us k dt ≤ C (1 + t ) dt |q(x)|2 (1 + |x|)2−2a dx
R R R

and the integral with respect to t is finite for 23 − a > 1 or a < 12 and the integral with respect
to x is the L2 (R)-norm of x 7→ q(x)(1 + |x|)1−a , and this is finite by assumption if 1 − a ≤ α.
Since we need 0 ≤ a < 21 , we see that the integral is finite for a suitable choice of a if α > 12 .
By Proposition 3.3.1 we see that us ∈ D(W ± ) for arbitrary s ∈ R.
Lemma 3.3.6. {us | s ∈ R} is dense in L2 (R).
Now Lemma 3.3.6 and Theorem 3.2.6 imply that D(W + ) = L2 (R) = D(W − ).

Exercise 3.3.7. Prove Lemma 3.3.5 and Lemma 3.3.6. For the proof of Lemma 3.3.5 use
that the lemma is true if C = R and Cauchy’s theorem from Complex Function Theory on
shifting contours for integrals of analytic functions. For the proof of Lemma 3.3.6 proceed as
follows; take g ∈ L2 (R) orthogonal to any us , then
Z
2
Fg (λ)λe−λ eisλ dλ


0 = hg, us i = hFg, Fus i =
R

2
for any s, i.e. the Fourier inverse of Fg (λ)λe−λ equals zero. Conclude that g has to be

zero in L2 (R).

3.4 Completeness
As stated in Section 3.2 we want as many scattering states as possible. Moreover, by Proposi-
tion 3.2.4, the discrete spectrum is not be expected to be related to scattering states, certainly
d2
not for the Schrödinger operator L and unperturbed Schrödinger operator − dx 2 , since for the
38 Chapter 3: Scattering and wave operators

last one σp = ∅, whereas the discrete spectrum of the Schrödinger operator L may be non-
trivial. For this reason we need to exclude eigenvectors, and we go even one step further by
restricting to the subspace of absolute continuity, see Section 6.6. Again, we phrase the results
for self-adjoint operators L and L0 acting on a Hilbert space H, but the main example is the
case of the Schrödinger operators on L2 (R).
We assume that the domains of the wave operator W ± at least contain the absolute
continuous subspace Hac (L0 ) for the operator L0 .
d 2
Exercise 3.4.1. Show that for the unperturbed Schrödinger operator − dx 2 the subspace of

absolute continuity is equal to H, so Pac = 1.

Definition 3.4.2. Let L and L0 be self-adjoint operators on the Hilbert space H. The gener-
alised wave operators Ω± = Ω± (L, L0 ) exist if Hac (L0 ) ⊂ D(W ± ), and Ω± = W ± Pac (L0 ).

We first translate some of the results of the wave operators obtain in Section 3.2 to the
generalised wave operators Ω± .

Proposition 3.4.3. Assume that Ω± exist, then

1. Ω± are partial isometries with inital space Ran(Pac (L0 )) and final space Ran(Ω± ),

2. Ω± (D(L0 )) ⊂ D(L) and L Ω± = Ω± L0 ,

3. Ran(Ω± ) ⊂ Ran(Pac (L)).

Proof. The first statement follows from the discussion following Definition 3.2.2. The sec-
ond statement follows from Theorem 3.2.13

⊥ and Theorem 3.2.15 and the observation that
Ker(L0 ) ⊂ Ran Ppp (L0 ) ⊂ Ran Pac (L0 ) and Theorem
 6.6.2. By
 the second result we see
that the generalised wave operators intertwine LRan(Ω± ) with L0 Ran(Pac (L0 )) and by the first

statement this equivalence is unitary. So LRan(Ω± ) has only absolutely continuous spectrum
or Ran(Ω± ) ⊂ Pac (L), see Section 6.6.

The next proposition should be compared to Exercise 3.2.5, and because we take into
account the projection onto the absolutely continuous subspace some more care is needed.

Proposition 3.4.4. Let A, B, C be self-adjoint operators, and assume that Ω± (A, B) and
Ω± (B, C) exist, then Ω± (A, C) exist and Ω± (A, C) = Ω± (A, B)Ω± (B, C).

In particular, Ω± (A, A) = Pac (A), which follows by definition.

Proof. Since Ω+ (B, C) exists, it follows from the third statement of Proposition

3.4.3 that
Ran(Ω+ (B, C)) ⊂ Ran(Pac (B)), so for any ψ ∈ H we have 1 − Pac (B) Ω+ (B, C)ψ = 0, or

lim k 1 − Pac (B) eitB e−itC ψk = 0.

t→∞
Chapter 3: Scattering and wave operators 39

Now
eitA e−itC Pac (C)ψ = eitA e−itB eitB e−itC Pac (C)ψ
= eitA e−itB Pac (B)eitB e−itC Pac (C)ψ + eitA e−itB (1 − Pac (B))eitB e−itC Pac (C)ψ.
Put Ω+ (B, C)ψ = φ, and Ω+ (A, B)φ = ξ, then
keitA e−itC Pac (C)ψ − ξk ≤ keitA e−itB Pac (B)eitB e−itC Pac (C)ψ − ξk
+ keitA e−itB (1 − Pac (B))eitB e−itC Pac (C)ψk
≤ keitA e−itB Pac (B)eitB e−itC Pac (C)ψ − eitA e−itB Pac (B)φk + keitA e−itB Pac (B)φ − ξk
+ k(1 − Pac (B))eitB e−itC Pac (C)ψk
≤ keitB e−itC Pac (C)ψ − φk + keitA e−itB Pac (B)φ − ξk + k(1 − Pac (B))eitB e−itC Pac (C)ψk
and since each of these three terms tends to zero as t → ∞, the result follows.
Completeness is related to the fact that we only have scattering states, i.e. any solution
to the time-dependent Schrödinger equation with an incoming asymptotic state also has an
outgoing asymptotic state and vice versa. This is known as weak asymptotic completeness,
and for the generalised wave operators this means Ran(Ω+ (L, L0 )) = Ran(Ω− (L, L0 )). We

⊥
have strong asymptotic completeness if Ran(Ω+ (L, L0 )) = Ran(Ω− (L, L0 )) = Ppp (L)H .
Here Ppp (L) is the projection on the closure of the subspace of consisting of all eigenvectors of
L, see Section 6.6. In this section we have an intermediate notion for completeness, and here
Pac (L) denotes the orthogonal projection onto the absolutely continuous subspace for L, see
Section 6.6.
Definition 3.4.5. Assume Ω± (L, L0 ) exist, then we say that the generalised wave operators
are complete if Ran(Ω+ (L, L0 )) = Ran(Ω− (L, L0 )) = Ran(Pac (L)).
Note that completeness plus empty singular continous spectrum of L, see Section 6.6, is
equivalent to strong asymptotic completeness.
Theorem 3.4.6. Assume Ω± (L, L0 ) exist. Then the generalised wave operators Ω± (L, L0 )
are complete if and only if the generalised wave operators Ω± (L0 , L) exist.
Proof. Assume first that Ω± (L, L0 ) and Ω± (L0 , L) exist. By Proposition 3.4.4 we have
Pac (L) = Ω± (L, L) = Ω± (L, L0 )Ω± (L0 , L),
so that Ran(Pac (L)) ⊂ Ran(Ω± (L, L0 )). Since Proposition 3.4.3 gives the reverse inclusion,
we have Ran(Pac (L)) = Ran(Ω± (L, L0 )), and so the generalised wave operators are complete.
Conversely, assume completeness of the wave operators Ω± (L, L0 ). For φ ∈ Pac (L) =
Ran(Ω+ (L, L0 )) we have ψ such that φ = Ω+ (L, L0 )ψ, or
keitL e−itL0 Pac (L0 )ψ − φk = ke−itL0 Pac (L0 )ψ − e−itL φk = kPac (L0 )ψ − eitL0 e−itL φk
tends to zero as t → ∞. It follows that strong limit of eitL0 e−itL exists on the absolutely
continuous subspace for L, hence Ω+ (L0 , L) exists.
40 Chapter 3: Scattering and wave operators

Theorem 3.4.6 gives a characterisation that is usually hard to establish. The reason is that
L0 is the unperturbed “simple” operator, and we can expect to have control on its absolutely
continuous subspace and its spectral decomposition, but in order to apply Theorem 3.4.6 one
also needs to know these properties of the perturbed operator, which is much harder. There
are many theorems (the so-called Kato-Birman theory) on the existence of the generalised
wave operators if L and L0 (or f (L) and f (L0 ) for a suitable function f ) differ by a trace-class
operator. This is less applicable for the Schrödinger operators.
In the remainder of this Chapter 3 we use the notion of T -smooth operators in order to
study wave operators for the Schrödinger operators. However, the proof is too technical to
consider in detail, and we only discuss the main ingredients.
Definition 3.4.7. Let (T, D(T )) be a self-adjoint operator on a Hilbert space H. A closed
operator (S, D(S)) on H is T -smooth if there exists a constant C such that for all x ∈ H the
element eitT x ∈ D(S) for almost all t ∈ R and
Z
1
kSeitT xk2 dt ≤ C kxk2 .
2π R
Note that the identity operator is never T -smooth for any operator T .
The notion of T -smooth is stronger than the assumption occuring in the Rellich Pertur-
bation Theorem 2.2.4.
Proposition 3.4.8. Let (T, D(T )) be a self-adjoint operator on a Hilbert space H, and the
closed operator (S, D(S)) on H be T -smooth. Then for all ε > 0 there exists a K such that

kSxk ≤ ε kT xk + K kxk, ∀x ∈ D(T ).

Moreover, Ran(S ∗ ) ⊂ Hac (T ).

Proof. Use the functional calculus and the formula
Z ∞
1
=i e−ixt eiat dt, =a > 0,
x−a 0

to find for y ∈ D(S ∗ ), x ∈ H, a = λ + iε, the expression

Z ∞ Z ∞
∗ −itT ∗ iλt −εt
−i hR(λ + iε)x, S yi = he x, S yi e e dt = hSe−itT x, yi eiλt e−εt dt
0 0

for all ε > 0, where R(z) = (T − z)−1 , z ∈ ρ(T ), is the resolvent for T , see Section 6.4. To
estimate this expression we use the Cauchy-Schwarz inequality (6.1.1) in L2 (0, ∞);
Z ∞
∗
|hR(λ + iε)x, S yi| ≤ kyk kSe−itT xke−εt dt
0
 21 Z  21
∞ ∞
kyk √
Z
≤ kyk kSe−itT xk2 dt e−2εt dt ≤√ 2πC kxk,
0 0 2ε
Chapter 3: Scattering and wave operators 41

since S is T -smooth. It follows that y 7→ hS ∗ y, R(λ + iε)xi

qis continuous, so that R(λ + iε)x ∈
D(S ∗∗ ) = D(S), since S is closed, and kSR(λ + iε)xk ≤ πC ε
kxk. Since Ran(R(z)) = D(T ),
see Lemma 6.2.2, we see that D(T ) ⊂ D(S). Now for x ∈ D(T ) write
r r
πC πC  √ 
kSxk = kSR(λ+iε)(T −(λ+iε)) xk ≤ k(T −(λ+iε)) xk ≤ 2 2
kT xk+ λ + ε kxk
ε ε
and take ε big in order to obtain the result.
Next, use the closedness of Hac to reduce to proving Ran(S ∗ ) ⊂ Hac (T ). Take x = S ∗ y ∈
Ran(S ∗ ) and consider almost everywhere
√ Z
2π F (t) = e−itu dEx,x (u) = he−itT x, xi = he−itT x, S ∗ yi = hSe−itT x, yi,
R

where E is the spectral measure for the self-adjoint operator T . Now estimate |F (t)| ≤
kSe−itT xkkyk to see that F ∈ H. Since F is the Fourier transform of the measure Ex,x
we see fromR the Plancherel Theorem, see Section 6.3, that Ex,x is absolutely continuous by
Ex,x (B) = B F −1 F (λ) dλ, B Borel set. Hence, x is in the absolutely continuous subspace

for T .
Theorem 3.4.9. Assume L and L0 are self-adjoint operators such that L = L0 + A∗ B in the
following sense; D(L) ⊂ D(A), D(L0 ) ⊂ D(B) and for all x ∈ D(L), y ∈ D(L0 )
hLx, yi = hx, L0 yi + hAx, Byi.
Assume that A is L-smooth and B is L0 -smooth, then the wave operators W ± exist as unitary
operators.
In particular, Ran W + = Ran W − = H. Note that Theorem 3.4.9 does not deal yet
with generalised wave operators, so it cannot be applied to the Schrödinger operators once
the perturbed Schrödinger operator has discrete spectrum, cf. Proposition 3.2.4. Moreover,
it still has the problem that there is a smoothness condition with respect to the perturbed
operator L. The proof of Theorem 3.4.9 is based on the idea of Cook as for Proposition 3.3.1.
Proof. Take x ∈ D(L0 ) and consider w(t) = eitL e−itL0 x. Take y ∈ D(L) and consider
d d
hw(t), yi = he−itL0 x, e−itL yi = −ihL0 e−itL0 x, e−itL yi + ihe−itL0 x, L e−itL yi
dt dt
= ihB e−itL0 x, A e−itL yi
since e−itL0 x ∈ D(L), eitL y ∈ D(L0 ). Consequently, for t > s
Z t Z t
−iuL0 −iuL
|hw(t) − w(s), yi| ≤ |hB e x, A e yi| du ≤ kB e−iuL0 xk kA e−iuL yk du
s s
Z t  12 Z t  12
≤ kB e−iuL0 xk2 du kA e−iuL yk2 du
s s
Z t  12 Z  12 Z t  12
≤ kB e−iuL0 xk2 du kA e−iuL yk2 du ≤ Ckyk kB e−iuL0 xk2 du
s R s
42 Chapter 3: Scattering and wave operators

for some constant C since A is L-smooth. In particular, we see

Z t  21
−iuL0 2
kw(t) − w(s)k ≤ C kB e xk du (3.4.1)
s

and since the integrand is integrable over R we see that w(s) is Cauchy as s → ∞. So the
domain of W + contains the dense subset D(L0 ), and since D(W + ) is closed by Theorem 3.2.6
it follows that D(W + ) is the whole Hilbert space. Similarly for W − .
Since the problem is symmetric in L and L0 , i.e. L0 = L − B ∗ A it follows that the strong
limits of eitL0 e−itL also exist for t → ±∞. Since these are each others inverses, unitarity
follows.
d 2
Theorem 3.4.9 can be used to discuss completeness for the Schrödinger operator − dx 2 + q,

but the details are complicated. We only discuss globally the proof of Theorem 3.4.13. For
this we first give a condition on the resolvent of T and S for S to be a T -smooth operators
using the Fourier transform. R
First recall, that for F : R → H a a Hilbert space valued function, such
R that R kF (t)k dt <
∞ (and t 7→R hF (t), yi is measurable for all y ∈ H) we can define R F (t) dt ∈ H. In-
deed, yR 7→ R hy, F (t)i
R dt is bounded and now use Riesz’s representation theorem. Note
that
R F (t) dt R≤ R kF (t)k dt. For such a function F we define its Fourier transform

as FF (λ) = √12π R e−iλt F (t) dt as an element of H.

Lemma 3.4.10. Let (S, D(S)) be a closed operator on H, then

Z Z
2
kSF (t)k2 dt,


kS FF (λ)k dλ =
R R

where the left hand side is set to ∞ in case FF (λ) ∈

/ D(S) almost everywhere, and similarly
for the right hand side.

Exercise 3.4.11. Prove Lemma 3.4.10 according to the following steps.

• First take S bounded, then for any x ∈ H we see that hSFF (λ), xi = hFF (λ), S ∗ xi is
the (ordinary) Fourier transform of hSF (t), xi = hF (t), S ∗ xi. Use Plancherel and next
take x elements
R of an orthonormal
R basis of the Hilbert space (we assume H separable)
Deduce R kSFF (λ)k2 dλ = R kSF (t)k2 dt.

• Next assume (S, D(S)) self-adjoint with spectral decomposition E. Use the previous
result for SE(−n, n) and use that if F (t) ∈ D(S) almost everywhere kSE(−n, n) F (t)k2
converges monotonically to kSF (t)k2 as n → ∞. Now use the monotone convergence
theorem.

• For arbitrary (S, D(S)) we take its polar decomposition S = U |S| with D(|S|) = D(S),
(|S|, D(|S|)) self-adjoint and k |S|xk = kSxk for all x ∈ D(S) = D(|S|).
Chapter 3: Scattering and wave operators 43

Apply Lemma 3.4.10 (with the Fourier transform replaced by the inverse Fourier transform)
to F (t) = e−εt e−itT x for t ≥ 0, F (t) = 0 for t < 0. Using the functional calculus as in the
beginning of the proof of Proposition 3.4.8 we get, for ε > 0 and R(z) the resolvent for T ,
Z Z ∞
2
kSR(λ + iε)xk dλ = 2π e−2εt kSe−itT xk2 dt.
R 0

Similarly, we obtain for ε > 0

Z Z 0
2
kSR(λ − iε)xk dλ = 2π e2εt kSe−itT xk2 dt.
R −∞

Using the Parseval version, or by polarizing, one can also derive

Z
hSR(λ + iε)x, SR(λ − iε)yi dλ = 0
R

assuming the integral exists. In particular, we get

Z Z
2 2
kS R(λ + iε) − R(λ − iε) xk2 dλ


4ε kSR(λ + iε)R(λ − iε)xk dλ =
R RZ

= 2π e−2ε|t| kSe−itT xk2 dt.

R

using the resolvent equation R(µ) − R(λ) = (µ − λ)R(µ)R(λ).

Proposition 3.4.12. S is a T -smooth operator if and only if for all x ∈ H we have R(λ ±
iε)x ∈ D(S) for allmost all λ ∈ R and
Z
sup kSR(λ + iε)xk2 + kSR(λ − iε)xk2 dλ < ∞.
kxk=1, ε>0 R

If εkSR(λ + iε)k2 ≤ C or if εkSR(λ − iε)k2 ≤ C independently of ε > 0, λ ∈ R, then S is

T -smooth.

Proof. For the first statement we use the two equations following Exercise 3.4.11 to obtain
Z Z
kSR(λ + iε)xk + kSR(λ − iε)xk dλ = 2π e−2ε|t| kSe−itT xk2 dt.
2 2
R R

In case S is a T -smooth operator the right hand side is bounded for ε ≥ 0, so the supremum
of the left hand side is finite. In case the supremum of the left hand side is finite, it follows
from the monotone convergence theorem that R kSe−itT xk2 dt < ∞.
R

For the second statement we use

Z Z Z
−2ε|t| −itT 2 2
2π e kSe xk dt = 4ε kSR(λ + iε)R(λ − iε)xk dλ ≤ 4εC kR(λ − iε)xk2 dλ
2
R R R
44 Chapter 3: Scattering and wave operators

and now use the Spectral Theorem 6.4.1 to observe

Z
2 1
kR(λ − iε)xk = hR(λ + iε)R(λ − iε)x, xi = dEx,x (t).
R |t − λ − iε|2
So we can rewrite the right hand side as
Z Z Z Z
1 1
4εC 2
dEx,x (t) dλ = 4εC 2
dλ dEx,x (t)
R R |t − λ − iε| R R |t − λ − iε|

(interchanging is valid since the integrand is positive), and the inner integral
Z
1
Z
1 1  λ − t ∞ π
dλ = dλ = arctan = .

2 2 2
R |t − λ − iε| R (t − λ) + ε ε ε ε

λ=−∞

So we get
Z Z Z
1
4εC dEx,x (t) dλ = 4πC dEx,x (t) = 4πCkxk2
R R |t − λ − iε|2 R

by the Spectral Theorem 6.4.1. So this leads to

Z
2π e−2ε|t| kSe−itT xk2 dt ≤ 4πCkxk2 , (3.4.2)
R

or S is T -smooth.
The machinery of T -smooth operators and in particular Theorem 3.4.9 can be applied to
the Schrödinger operator, but the proof is very technical and outside the scope of these lecture
notes.
d d 2 2
Theorem 3.4.13. Assume q ∈ L1 (R) ∩ L2 (R), and let L0 = − dx 2 , L = − dx2 + q, then the
±
wave operators Ω (L, L0 ) are complete
We put (p
|q(x)|, q(x) 6= 0, q(x)
a(x) = −x2
, b(x) = .
e q(x) = 0 a(x)
Then we are in the situation of Theorem 3.4.9, but the space of absolute continuity has
to brought into play. This is done by employing the spectral decomposition and assuming
that there exists an open set Z such that E(Z) corresponds to the orthogonal projection on
the absolutely continuous subspace. Next it turns out that even for R being the resolvent
d2
of L0 = − dx 2 one cannot prove the sufficient condition of Proposition 3.4.12 since there

is no estimate uniformly in λ ∈ R. One first shows that the condition can be relaxed to
εkAR(λ ± iε)k2 ≤ CI independently of ε > 0, λ ∈ I for any interval I with compact closure
in Z, where E(Z) is Pac .
For R the resolvent of L0 and B multiplication by b this is easily checked. Note that in
the application indeed b ∈ L2 (R).
Chapter 3: Scattering and wave operators 45

Lemma 3.4.14. With R the resolvent of L0 = − dxd2 and B the multiplication operator by
b ∈ L2 (R) we have εkBR(λ ± iε)k2 ≤ CI for all intervals I with compact closure in (0, ∞).

Note that BR(z), =z 6= 0, is a compact operator. Indeed, by Theorem 2.3.4 B is L0 -

compact, and if {fn }∞ 2
n=1 is a bounded sequence, say kfn k ≤ M , in L (R) then we have

kR(z)fn k + kL0 R(z)fn k ≤ M kR(z)k + k(L0 − z)R(z)fn k + kzR(z)fn k

≤ M kR(z)k(1 + |z|) + M,

so that the sequence is bounded. Hence, B being L0 -compact we see that BR(z)fn has a
convergent subsequence, so that BR(z) is a compact operator.
Proof. For this we need the resolvent of L0 , which we claim equals
−1
Z
eiγ|x−y| f (y) dy


R(z)f (x) =
2γi R

where z = γ 2 ∈ C\R and =γ > 0, see Exercise 2.1.5.

Assuming this we see using the Cauchy-Schwarz inequality (6.1.1) in L2 (R),

|b(x)|2 |b(x)|2
Z Z Z
−2=γ |x−y|
2 2
e−2=γ |y| dy kf k2


|b(x) R(z)f (x)| ≤ 2
e dy |f (y)| dy = 2
4|γ| R R 4|γ| R
2 2 2
|b(x)| kf k kak
= kf k2 =⇒ kBR(z)f k2 ≤
4|z| =γ 4|z| =γ

giving the required estimate.

The analogous statement to Lemma 3.4.14 with R the resolvent for the perturbed operator
and the multiplication by a ∈ L2 (R) is a much harder statement to prove. The idea is to rewrite
the relation in Theorem 3.4.9 for the resolvent as

∗ 
∗ 
AR(z) = AR0 (z) + A BR0 (z̄) AR(z) =⇒ 1 − A BR0 (z̄) AR(z) = AR0 (z)

where we denote the
resolvent for L by R and the resolvent for L0 by R0 . If we now can show
∗
that 1 − A BR0 (z̄) is invertible and if we have sufficient control on its inverse we can use
Lemma 3.4.14 once more to find the result. In particular one needs the inverse to be uniformly
bounded in a neighbourhood of an interval I (with I as in Lemma 3.4.14). We will not go
into this, but refer to Schechter [10, Ch. 9] and Reed and Simon [9, Ch. XIII].
46 Chapter 3: Scattering and wave operators
Chapter 4

Schrödinger operators and scattering

data

4.1 Jost solutions

d 2
2
The unperturbed Schrödinger operator − dx 2 has eigenvalues λ = γ for the corresponding
eigenfunctions exp(±iγx). If the potential q is of sufficient decay at ±∞ we can expect
d2
that eigenfunctions of the Schrödinger equation − dx 2 + q exist that behave as exp(±iγx) for
1
x → ±∞. These solutions are known as the Jost solutions.
The basic assumption in this chapter on the potential q is that it is a real-valued function
having sufficient decay. This is specified in each of the results.
Definition 4.1.1. The Schrödinger integral equation at ∞ is the integral equation
Z ∞
iγx sin(γ(x − y))
f (x) = e − q(y) f (y) dy
x γ
and the Schrödinger integral equation at −∞ is the integral equation
Z x
−iγx sin(γ(x − y))
f (x) = e + q(y) f (y) dy
−∞ γ
For f a solution to the Schrödinger integral equation at ∞ we can formally calculate
Z ∞
0 iγx sin(γ(x − x))
f (x) = iγe + q(x) f (x) − cos(γ(x − y)) q(y) f (y) dy
γ x
Z ∞
= iγeiγx − cos(γ(x − y)) q(y) f (y) dy =⇒
x
Z ∞
00 2 iγx
f (x) = −γ e + cos(γ(x − x)) q(x) f (x) + γ sin(γ(x − y)) q(y) f (y) dy
x
= −γ 2 f (x) + q(x) f (x),
1
Res Jost, Swiss theoretical physicist.

47
48 Chapter 4: Schrödinger operators and scattering data

so that it gives a solution to the Schrödinger eigenvalue equation −f 00 +qf = γ 2 f . The integral
equation can be derived by viewing f 00 +γ 2 f = u, u = q f , as an inhomogeneous linear second-
order differential equation and apply the method of variation of constants. Similarly, a solution
of the Schrödinger integral equation at −∞ gives rise to an eigenfunction of the Schrödinger
operator for the eigenvalue γ 2 . The Schrödinger integral equation can also be obtained using
the method of variation of constants for f 00 + γ 2 f = g, and then take g = qf .
The three-dimensional analogues of the Schrödinger integral equations in quantum me-
chanics are known as the Lippmann-Schwinger equation.
Theorem 4.1.2. Assume that the potential q is a real-valued integrable function. Let γ ∈ C,
=γ ≥ 0, γ 6= 0. Then the Schrödinger integral equation at ∞ has a unique solution fγ+ which
is continuously differentiable and satisfies the estimates
Z ∞
+ iγx −x=γ 
 1 
|fγ (x) − e | ≤ e exp( |q(y)| dy) − 1
|γ| x
+ Z ∞
dfγ iγx −x=γ 
 1 
| (x) − iγ e | ≤ |γ| e exp( |q(y)| dy) − 1.
dx |γ| x
Moreover, the Schrödinger integral equation at −∞ has a unique solution fγ− which is contin-
uously differentiable and satisfies the estimates
Z x
− −iγx

x=γ  1 
|fγ (x) − e |≤e exp( |q(y)| dy) − 1,
|γ| −∞
− Z x
dfγ iγx

x=γ  1 
| (x) + iγe | ≤ |γ| e exp( |q(y)| dy) − 1.
dx |γ| −∞
Using ex − 1 ≤ xex we can estimate
Z ∞ Z ∞
+ iγx −x=γ 1 1
|fγ (x) − e | ≤ e |q(y)| dy exp( |q(y)| dy),
|γ| x |γ| x
Z ∞ Z ∞ (4.1.1)
dfγ+ iγx −x=γ 1
| (x) − iγ e | ≤ e |q(y)| dy exp( |q(y)| dy),
dx x |γ| x
and a similar estimate for fγ− and its derivative.
+ −
Since q is real-valued it follows that for x ∈ R we have fγ+ (x) = f−γ̄ (x) and fγ− (x) = f−γ̄ (x)
+
as solutions to the Schrödinger integral equations. Since q is an integrable function and fγ ∈
C 1 (R) is bounded for x → ∞, the differentiations in the calculations preceding Theorem 4.1.2
are justifiable using Lebesgue’s Theorems 6.1.3 and 6.1.4 on differentiation and dominated
convergence and hold almost everywhere. The solutions in Theorem 4.1.2 are known as Jost
solutions.
Proof. The Schrödinger integral equations are Volterra2 type integral equations, and a stan-
dard way to find a solution is by successive iteration. Put f0 (x) = e−iγx , and define
Z ∞
sin(γ(x − y))
fn+1 (x) = − q(y) fn (y) dy, n ≥ 0.
x γ
2
Vito Volterra (3 May 1860 — 11 October 1940), Italian mathematician.
Chapter 4: Schrödinger operators and scattering data 49

Put mn (x) = e−iγx fn (x), so that mn is defined inductively by m0 (x) = 1 and

Z ∞
sin(γ(x − y)) iγ(y−x)
mn+1 (x) = − e q(y) mn (y) dy
x γ
Z ∞
1
= (e2iγ(y−x) − 1) q(y) mn (y) dy, n ≥ 0.
x 2iγ

Now
sin(γ(x − y)) eiγ(y−x)  = 1 1 − e2iγ(y−x)  ≤ 1 1 + e−2(y−x)=γ ≤ 1
   

2 2
for y ≥ x and =γ ≥ 0. Hence, assuming =γ ≥ 0 we find
Z ∞
mn+1 (x) ≤ 1
 
|q(y)| |mn (y)| dy,
|γ| x

n
R(x)
and then we have by induction on n the estimate |mn (x)| ≤ , where
|γ|n n!
Z ∞ Z
R(x) = |q(y)| dy ≤ |q(y)| dy = kqk1 .
x R

Note that R is a continuous decreasing bounded function, which is almost everywhere differ-
entiable by Lebesgue’s differentation Theorem 6.1.4. Indeed, for n = 0 this inequality is valid
since m0 (x) = 1, and

∞

n Z R(x)
n+1
R(y) R(x)
Z
mn+1 (x) ≤ 1 1
  n
|q(y)| dy = n+1 s ds = n+1
|γ| x |γ|n n! |γ| n! 0 |γ| (n + 1)!

by putting s = R(y), so ds = −|q(y)| dy and the interval [x, ∞) is mapped to [0, R(x)], since
R is decreasing. P
∞ P∞
So the series n=0 mn (x) converges uniformly, and Pso does the series n=0 fn (x), for
∞
=γ ≥ 0. It is straightforward to check that the series n=0 fn (x) gives a solution to the
Schrödinger equation at ∞, and this solution we denote by fγ+ . Then
∞
X ∞
X
−x=γ
|fγ+ (x) iγx
−e |≤ |fn (x)| ≤ e |mn (x)| ≤ e−x=γ | exp(R(x)/|γ|) − 1| → 0, x→∞
n=1 n=1

for fixed γ ∈ C with γ 6= 0, =γ ≥ 0.

Note that m0 ∈ C ∞ (R) for all γ ∈ C, and since the integrals defining mn (x) in terms of
mn−1 converge absolutely for =γ ≥ 0 we see that each mn is at least in C 1 (R). Consequently,
fn ∈ C 1 (R), and
Z ∞ Z ∞
dfn+1
(x) = − cos(γ(x−y)) q(y) fn (y) dy = − cos(γ(x−y)) eiγy q(y) mn (y) dy (4.1.2)
dx x x
50 Chapter 4: Schrödinger operators and scattering data

and now using

cos(γ(x − y)) eiγy  = |eiγx | 1 1 + e2iγ(y−x) | ≤ e−x=γ
  
2
for y ≥ x, =γ ≥ 0. This gives
  Z ∞ Z ∞
n −x=γ

n+1
 dfn+1  −x=γ R(y) e R(x)
 dx (x) = e
  |q(y)| |mn (y)| dy ≤ e−x=γ |q(y)| dy = .
x x |γ|n n! |γ|n (n + 1)!
It follows that the series ∞ dfn
P
n=0 dx (x) converges uniformly on compact subsets of R, so this is
a continuous function and fγ ∈ C 1 (R). We find for =γ ≥ 0
+

 +  X ∞   X ∞ −x=γ

n
 dfγ dfn e R(x)
= |γ| e−x=γ (exp(R(x)/|γ|) − 1)

iγx 
 
 dx (x) − iγe  ≤  dx (x) ≤
  
|γ| n−1 n!
n=1 n=1

which tends to zero as x → ∞.

In order to see that this solution is unique, let f and g be two C 1 -solutions to the
Schrödinger integral equation at ∞. Then h(x) = e−iγx (f (x) − g(x)) is a solution to
Z ∞
sin(γ(x − y)) iγ(y−x)
h(x) = − e q(y) h(y) dy.
x γ
Let x0 ∈ R be arbitrary, and put L = supx≥x0 |h(x)| < ∞, since f and g remain bounded as
x → ∞. So |h(x)| ≤ L for all x
≥ x0 , and by induction, using the integral representation as
n
R(x) kqkn
before, we find |h(x)| ≤ L |γ|n n! ≤ L |γ|n 1n! for all n ∈ N. Hence, h(x) = 0 for all x ≥ x0 , and
since x0 ∈ R is arbitrary, h = 0 and uniqueness follows.
The statements for the Schrödinger integral equation at −∞ are proved analogously, and
are left as an exercise.
Exercise 4.1.3. With the assumptions in Theorem 4.1.2 and assuming that q is moreover an
element of C k (R). Show that fγ+ , fγ− ∈ C k+2 .
From the proof of Theorem 4.1.2 we obtain the following corollary, since a series of analytic
functions is analytic if the convergence is uniformly on compact sets and each fn is analytic
in γ for =γ > 0.
df + dfγ−
Corollary 4.1.4. γ 7→ fγ+ (x), γ 7→ dxγ and γ 7→ fγ− (x), γ 7→ dx
are analytic in γ for =γ > 0
and continuous in γ for =γ ≥ 0, γ 6= 0.
−iγx +
Note that in the proof of Theorem 4.1.2 we actually deal with m+ γ (x) = e fγ (x) and
− iγx −
mγ (x) = e fγ (x). These functions are solutions to the differential equation
m00 + 2iγ m0 = q m (4.1.3)
as follows by a straightforward calculation, and they are the solutions to an integral equation;
Z ∞ 2iγ(y−x)
e −1
m(x) = 1 + q(y) m(y) dy (4.1.4)
x 2iγ
Chapter 4: Schrödinger operators and scattering data 51

−
for m = m+ γ and a similar one for mγ .
The estimates in Theorem 4.1.2 can be improved by imposing more conditions on the
potential q. We discuss two possible extensions. The first additional assumption also allows
an extension to the case γ = 0.
R
Theorem 4.1.5. Assume the potential q is real-valued and R (1 + |x|)|q(x)| dx < ∞. Then
for γ ∈ C with =γ ≥ 0,
Z ∞
+ iγx C e−x=γ
|fγ (x) − e | ≤ (1 + max(−x, 0)) (1 + |y|)|q(y)| dy
1 + |γ| x
dfγ+ C e−x=γ ∞
Z
iγx
| (x) − iγ e | ≤ (1 + |y|)|q(y)| dy,
dx 1 + |γ| x
where C is a constant only depending on the potential q. Moreover, the Schrödinger integral
equation at −∞ has a unique solution fγ− which is continuously differentiable and satisfies the
estimates
Z x
− −iγx C e−x=γ
|fγ (x) − e |≤ (1 + max(x, 0)) (1 + |y|)|q(y)| dy
1 + |γ| −∞
dfγ− C e−x=γ x
Z
iγx
| (x) + iγe | ≤ (1 + |y|)|q(y)| dy.
dx 1 + |γ| −∞
Proof. In particular, the potential q satisfies the estimate of Theorem 4.1.2, and the estimates
given there hold.
Use the estimate
 1 
 (e2iγ(y−x) − 1) ≤ (y − x)
2iγ
for γ ∈ R, y − x ≥ 0 and iterate to find
ZZ Z
|mn (x)| ≤ ··· (y1 − x)(y2 − y1 ) · · · (yn − yn−1 )|q(y1 )| · · · |q(yn )| dyn · · · dy1
x≤y1 ≤y2 ···≤yn
ZZ Z
≤ ··· (y1 − x)(y2 − x) · · · (yn − x)|q(y1 )| · · · |q(yn )| dyn · · · dy1 .
x≤y1 ≤y2 ···≤yn

Note that the integrand is invariant with respect to permutations of y1 up to yn , but not
the region. Taking all possible orderings gives, where Sn is the group of permutations on n
elements,
n! |mn (x)|
X ZZ Z
≤ ··· (y1 − x)(y2 − x) · · · (yn − x)|q(y1 )| · · · |q(yn )| dyn · · · dy1
w∈Sn x≤yw(1) ≤yw(2) ···≤yw(n)
Z ∞Z ∞ Z ∞
= ··· (y1 − x)(y2 − x) · · · (yn − x)|q(y1 )| · · · |q(yn )| dyn · · · dy1
x x x
Z ∞ n
= (y − x) |q(y)| dy ,
x
52 Chapter 4: Schrödinger operators and scattering data
R∞
so that, with R(x) = x
(y − x) |q(y)| dy, we get
|m+
γ (x)| ≤ exp(R(x)) − 1 ≤ R(x) exp(R(x)).

Note that R(x) ≥ R(y) for x ≤ y. Here we use again ex − 1 ≤ xex .

Next consider the estimate for m+γ by
Z ∞
|m+
γ (x)| ≤ 1 + (y − x)|q(y)| |m+γ (y)| dy
x
Z ∞ Z ∞
+
=1+ y|q(y)| |mγ (y)| dy + (−x) |q(y)| |m+
γ (y)| dy
x x
Z ∞ Z ∞
+
≤1+ y|q(y)| |mγ (y)| dy + (−x) |q(y)| |m+
γ (y)| dy
0 x

and observe that this inequality holds for positive and negative x ∈ R. We estimate the first
term, which is independent of x, by
Z ∞ Z ∞
+
1+ y|q(y)| |mγ (y)| dy ≤ 1 + R(0) exp(R(0)) y|q(y)| dy
0 0

2
= 1 + R(0) exp(R(0)) = C1 < ∞
R∞
where C1 is a finite constant only depending on R(0) = 0 y|q(y)| dy.
γm+ (x)
Now define M (x) = C1 (1+|x|) , p(x) = (1 + |x|)|q(x)|, then we can rewrite the implicit
+
estimate on mγ (x) in terms of M (x) as follows;
Z ∞
C1 (1 + |x|)|M (x)| ≤ C1 + (−x) C1 (1 + |t|) |q(t)| |M (t)| dt
x
Z ∞
1 −x
=⇒ |M (x)| ≤ + p(t) |M (t)| dt (4.1.5)
1 + |x| 1 + |x| x
Z ∞
≤1+ p(t) |M (t)| dt,
x

so that as before, cf. proof of Theorem 4.1.2, or using Gronwall’s3 Lemma, see Exercise 4.1.6,
we get Z ∞ Z



|M (x)| ≤ exp p(y) dy ≤ exp (1 + |y|) |q(y)| dy = C2 < ∞.
x R
So |m+ ≤ C3 (1 + |x|) with C3 = C1 C2 only depending on properties of q. This estimate is
γ (x)|
now being used to estimate
Z ∞ Z ∞
+ +
|mγ (x) − 1| ≤ y|q(y)| |mγ (y)| dy + (−x) |q(y)| |m+
γ (y)| dy
0 x
Z ∞ Z ∞
≤ R(0) exp(R(0)) y|q(y)| dy + (−x)C3 (1 + |y|) |q(y)| dy
0 x
3
Thomas Hakon Grönwall (16 January 1877 — 9 May 1932), Swedish-American mathematician, engineer,
chemist.
Chapter 4: Schrödinger operators and scattering data 53

and for x ≤ 0 we can estimate this by

Z ∞ Z ∞
≤ R(0) exp(R(0)) (1 + |y|) |q(y)| dy + |x|C3 (1 + |y|) |q(y)| dy
x x
Z ∞
≤ C4 (1 + |x|) (1 + |y|) |q(y)| dy
x


with C4 = max R(0) exp(R(0)), C3 only depending on properties of q. For x ≥ 0 we already
have the estimate
Z ∞
+
|mγ (x) − 1| ≤ R(x) exp(R(x)) ≤ R(x) exp(R(0)) = exp(R(0)) (y − x) |q(y)| dy
x
Z ∞ Z ∞
≤ exp(R(0)) y |q(y)| dy ≤ exp(R(0)) (1 + |y|) |q(y)| dy.
x x

Taking C5 = max(C4 , exp(R(0))) then gives

Z ∞
|m+
γ (x) − 1| ≤ C5 (1 + max(0, −x)) (1 + |y|) |q(y)| dy,
x

which is an estimate uniform in γ ∈ R. Combining this estimate in case |γ| ≤ 1 with the
estimate of Theorem 4.1.2 in case |γ| ≥ 1 gives the result for γ ∈ R. By Corollary 4.1.4 and
Theorem 4.1.2 the result is extended to γ ∈ C, =γ ≥ 0.
Differentiating (4.1.4) gives
Z ∞
dm+γ
(x) = − e2iγ(y−x) q(y) m+
γ (y) dy,
dx x

so that
dm+ ∞
(1 + max(0, −y))
Z
γ
| (x)| ≤ C |q(y)| dy,
dx x 1 + |γ|
which gives the result. The statements for fγ− are proved analogously.
Exercise 4.1.6. The following is known as Gronwall’s Lemma. Assume R ∞f , p are positive
continuous functions
R∞ on R such that there exists K ≥ 0 with f (x) ≤ K + x f (t)p(t) dt, then
f (x) ≤ K exp( x p(t) dt). Prove this according to the following steps.
• Observe that
f (x)p(x)
R∞ ≤ p(x)
K + x f (t)p(t) dt
and integrate both sides.
R∞ R∞
• Exponentiate the resulting inequality to get K + x
f (t)p(t) dt ≤ K exp( x p(t) dt),
and deduce the result.
At what cost can we remove the continuity assumption on f and p?
54 Chapter 4: Schrödinger operators and scattering data

Recall the definition of the Hardy space in Section 6.3.

Corollary 4.1.7. Under the assumptions of Theorem 4.1.5 γ 7→ e−iγx fγ+ (x) − 1 is an element
of the Hardy class H2+ for each fixed x ∈ R. Moreover, the H2+ -norm can be estimated
uniformly for x ∈ [x0 , ∞).
Proof. Put γ = α + iβ, with β > 0, and so |γ| ≥ |α| and consider
Z Z 2 Z
−iγx + 2
 1
|e fγ (x) − 1| dα ≤ C(1 + max(−x, 0)) (1 + |y|) |q(y)| dy | |2 dα < ∞
R R R 1 + |α|
independent of β = =γ > 0 by Theorem 4.1.5.
The second extension of Theorem 4.1.2 deals with analytic extensions of the Jost function
with respect to γ under the assumption that the potential decays even faster.

RTheorem
2m|y|
4.1.8. Let q be a real-valued potential. Assume that there exists m > 0 such that
R
e |q(y)| dy < ∞, then
Z ∞
−x=γ  −2mx0
+ iγx
 
e2my |q(y)| dy − 1,


|fγ (x) − e | ≤ e exp C(γ, m)e
x
Z ∞
dfγ+ iγx e −x=γ 
−2mx0

e2my |q(y)| dy − 1,


| (x) − iγe | ≤  exp C(γ, m)e
dx C(γ, m) x

for =γ > −m, x ≥ x0 , where

1
C(γ, m) =
ε + 2|<γ|
dfγ+
with ε = 2 min(m, m + =γ). In particular, γ 7→ fγ+ (x) and γ 7→ dx
(x) are analytic for γ ∈ C
with =γ > −m for fixed x ≥ x0 .
Using ex − 1 ≤ xex we see that
Z ∞ Z ∞
−x=γ −2mx0 −2mx0
|fγ+ (x) − eiγx | 2my
e2my |q(y)| dy ,


≤e C(γ, m)e e |q(y)| dy exp C(γ, m)e
x x

and similarly for its derivative, which is similar to the estimate in Theorem 4.1.5, but note
that the γ-region where this estimate is valid is much larger.
Proof. Since the assumption implies that q is integrable, Theorem 4.1.2 shows that there
is a unique solution fγ+ , and we only need to improve the estimates. Using the estimate
2|z|
| sin z| ≤ 1+|z| exp(|=z|) for z ∈ C (prove this estimate) we can now estimate, for y ≥ x ≥ x0 ,
 
1 
 sin(γ(x − y)) eiγ(y−x)  ≤ 2|x − y| (y−x)|=γ| −(y−x)=γ −2m(y−x) 2my −2mx
γ  1 + |γ| |x − y| e e e e e

2|x − y|


(y−x) |=γ|−=γ−2m −2mx0 2my
≤ e e e ≤ C(γ, m) e−2mx0 e2my
1 + |γ| |x − y|
Chapter 4: Schrödinger operators and scattering data 55

for =γ > −m. Here C(γ, m) is a finite constant independent of x and y, which we discuss
later in the proof.
Under this assumption we now have for x ≥ x0
Z ∞
−2mx0
|mn+1 (x)| ≤ C(γ, m)e e2my |q(y)| |mn (y)| dy,
x

hence we obtain in a similar way the estimate

R∞
n

n e2my |q(y)| dy
|mn (x)| ≤ C(γ, m) e−2mnx0 x
n!
for x ≥ x0 , which gives the result.
It remains to discuss the constant C(γ, m). Let ε = 2(m + =γ) > 0, in case 0 ≥ =γ > −m
and ε = 2m in case =γ ≥ 0, then we see that we can take for C(γ, m) the maximum of the
function
2x
x 7→ e−εx , x ≥ 0.
1 + |γ|x
q
A calculation shows that the maximum is attained for 2|γ| (−1 + 1 + 4|γ|
1
ε
) = q2
4|γ|
.
ε(1+ 1+ ε
)
The maximum is
q
2 exp −2/(1 + 1 + 4|γ|


ε
) 1 1
q ≤ ≤ = C(γ, m).
4|γ| 2 ε + 2|γ| ε + 2|<γ|
ε(1 + 1 + ε )
Next we use (4.1.2) and the estimate


−2mx0 −x=γ (y−x) |=γ|−=γ−2m
| cos(γ(x − y))eiγy
|≤e e e e2my ≤ e−2mx0 e−x=γ e2my
for x ≥ x0 and =γ > −m. This gives
  Z ∞
 dfn+1  −2mx0 −x=γ
 dx (x) ≤ e
  e e2my |q(y)| |mn (y)| dy
x
n

n −2mnx0 1 ∞ 2my
Z Z ∞
−2mx0 −x=γ 2mt
≤e e C(γ, m) e e |q(y)| e |q(t)| dt dy
n! x y
Z ∞ n+1
−x=γ

n −2m(n+1)x0 1 2my
=e C(γ, m) e e |q(y)| dy
(n + 1)! x

and this estimate gives the result as in Theorem 4.1.2.

Exercise 4.1.9. Prove the corresponding statements for the Jost solution fγ− under the as-
sumptions on the potential q in Theorem 4.1.8;
Z x
− −iγx
e−2my |q(y)| dy − 1,

x=γ  2mx0


|fγ (x) − e |≤e exp C(γ, m)e
−∞
dfγ− x=γ  Z x
−iγx e
e−2my |q(y)| dy − 1,
2mx0


| (x) + iγe |≤  exp C(γ, m)e
dx C(γ, m) −∞
56 Chapter 4: Schrödinger operators and scattering data

dfγ−
for =γ > −m, x ≤ x0 with (Cγ, m) as in Theorem 4.1.8. Conclude γ 7→ fγ− (x) and γ 7→ dx
(x)
are analytic for γ ∈ C with =γ > −m for fixed x ≤ x0 .
We also need some properties of the Jost solutions when differentiated with respect to γ.
R ∂m+
Proposition 4.1.10. Assume that R |x|k |q(x)| dx < ∞ for k = 0, 1, 2, then ∂γγ (x) is a
continously differentiable function, and it is the unique solution to the integral equation
Z ∞ 2iγx
e −1
n(x) = M (x) + q(y) n(y) dy,
x 2iγ
Z ∞
y − x 2iγ(y−x) e2iγ(y−x) − 1

M (x) = e − 2
q(y) m+
γ (y) dy,
x γ 2iγ
with limx→∞ n(x) = 0.
∂m+γ
It follows from the proof below that we can also estimate ∂γ
(x) explicitly. Moreover, it
is a solution to
n00 + 2iγ n0 = q n − 2i m+
γ,

i.e. the differential equation (4.1.3) differentiated with respect to γ. Switching back to the
∂f + −iγx ∂mγ
+
Jost solution fγ+ , we see that we find a solution ∂γγ (x) = −ixe−iγx m+
γ (x) + e ∂γ
(x) to

−f 00 + q f = γ 2 f + 2γ fγ+ .
Proof. Rewrite (4.1.4) with m(x, γ) = m+ γ (x),
Z ∞ x
e2iγx − 1
Z
m(x, γ) = 1 + D(y − x, γ) q(y) m(y, γ) dy, D(x, γ) = = e2iγt dt.
x 2iγ 0
∂m
Denoting ṁ(x, γ) = ∂γ
(x, γ),we see that ṁ satisfies the integral equation
Z ∞
ṁ(x, γ) = M (x, γ) + D(y − x, γ) q(y) ṁ(y, γ) dy,
x
Z ∞
M (x, γ) = Ḋ(y − x, γ) q(y) m(y, γ) dy,
x

which we consider as an integral equation for ṁ(x, γ) with known function M and which want
to solve in the same way by an iteration argument;
Z ∞ ∞
X
h0 (x) = M (x, γ), hn+1 (x) = D(y − x, γ) q(y) hn (y) dy, ṁ(x, γ) = hn (x).
x n=0

The iteration scheme is the same as in the proof of Theorem 4.1.2 except for the initial
condition. So we investigate the initial function M (x, γ). Note that
Z x
x 1 1
|Ḋ(x, γ)| = | 2ite2iγt dt| = | e2iγx − D(x, γ)| ≤ 2 (1 + |γ|x)
0 γ γ |γ|
Chapter 4: Schrödinger operators and scattering data 57

for x ≥ 0, =γ ≥ 0, so that
Z ∞
1
|M (x, γ)| ≤ (1 + |γ|(y − x))|q(y)| |m+
γ (y)| dy
x |γ|2
By the estimate |m+
γ (y)|≤ C(1 + |y|), cf. the proof of Theorem 4.1.5, we get
Z ∞ Z ∞
1 1
|M (x, γ)| ≤ 2 (1 + |y|) |q(y)| dy + (y − x)|q(y)| (1 + |y|) dy
|γ| x |γ| x
and the first integral is bounded and tends to zero for x → ∞, and for the second integral we
use
Z ∞ Z ∞ Z ∞
(y − x)|q(y)| (1 + |y|) dy = y|q(y)| (1 + |y|) dy + (−x) |q(y)| (1 + |y|) dy
x x x
Z ∞ Z ∞
2
≤ |q(y)| (|y| + |y| ) dy + (−x) |q(y)| (1 + |y|) dy = K(x),
0 x

cf.
R ∞ proof of Theorem 4.1.5. So RM (x) is bounded, and even tends to zero, for x → ∞ since
∞
x
(y − x)|q(y)|(1 + |y|) dy ≤ x |q(y)|(|y| + |y|2 ) dy → 0 and grows at most linearly for
x → −∞, Z ∞
C C
|M (x, γ)| ≤ 2 (1 + |y|) |q(y)| dy + K(x) = K1 (x)
|γ| x |γ|
Note that K1 (x) ≥ K1 (y) for x ≤ y, and then we find
Z ∞ n
K1 (x)
|hn (x)| ≤ |q(y)| dy
|γ|n n! x

and we find a solution to the integral equation. The remainder of the proof follows the lines
of proofs of Theorems 4.1.2 and 4.1.5 and is left as an exercise.
Exercise 4.1.11. Finish the proof of Proposition 4.1.10, and state and prove the correspond-
ing proposition for m−
γ.

4.2 Scattering data: transmission and reflection coeffi-

cients
We assume that the potential q is a real-valued integrable function, so that Theorem 4.1.2 ap-
plies. We also assume the conditions of Theorems 2.3.4, so that the corresponding Schrödinger
operator is self-adjoint with essential spectrum [0, ∞). So we get several solutions of the eigen-
d2 2
value equation for the operator − dx 2 + q for the eigenvalue λ = γ , γ ∈ R \ {0}. In particular
+ + − −
we get fγ , f−γ , fγ , f−γ all as solutions of the eigenvalue equation, so there are relations
amongst them since the solution space is 2-dimensional. In order to describe these we con-
sider the Wronskian4 W (f, g)(x) = f (x)g 0 (x) − f 0 (x)g(x). The Wronskian of two solutions of
the eigenvalue equation is non-zero if and only if the solutions are linearly independent.
4
Josef Hoëné de Wronski (23 August 1778 — 8 August 1853), Polish-French mathematician.
58 Chapter 4: Schrödinger operators and scattering data

Proposition 4.2.1. Assume q is a continuous real-valued integrable potential. For two so-
lutions f1 , f2 to −f 00 + qf = λf , the Wronskian W (f1 , f2 ) is constant, and in particular
+ −
W (fγ+ , f−γ ) = −2iγ and W (fγ− , f−γ ) = 2iγ, γ ∈ R\{0}.

Proof. Differentiating gives

d
[W (f, g)](x) = f (x)g 00 (x) − f 00 (x)g(x) = (q(x) − λ)f (x)g(x) − (q(x) − λ)f (x)g(x) = 0,
dx
+
since f and g are solutions to the eigenvalue equation. In particular, in case f = fγ+ , g = f−γ
we can evaluate this constant by taking x → ∞ and using the asymptotics of Theorem 4.1.2.
−
Similarly for the Wronskian of f±γ .

Exercise 4.2.2. In case the function q is not continuous, the derivative of the Wronskian has
to be interpreted in the weak sense. Modify Proposition 4.2.1 and its proof accordingly.

Since we now have four solutions, two by two linearly independent for γ 6= 0, to the
Schrödinger eigenvalue equation which has a two-dimensional solution space, for the eigenvalue
λ = γ 2 , γ ∈ R\{0}, we obtain a± ±
γ , bγ for γ ∈ R\{0} such that for all x ∈ R,

fγ− (x) = a+ + + +
γ fγ (x) + bγ f−γ (x),
(4.2.1)
fγ+ (x) = a− − − −
γ fγ (x) + bγ f−γ (x).

± + + −
Note that for γ ∈ R\{0} the relation fγ± (x) = f−γ (x) implies a+ + −
γ = a−γ , bγ = b−γ , aγ = a−γ
−
and b−
γ = b−γ . Now

2iγ = W (fγ− , f−γ

−
) = W (a+ + + + + + + +
γ fγ + bγ f−γ , a−γ f−γ + b−γ fγ )
= a+ + + + + + + + + + + +
γ a−γ W (fγ , f−γ ) + bγ b−γ W (f−γ , fγ ) = −2iγaγ a−γ + 2iγbγ b−γ ,
=⇒ 1 = |b+ 2 + 2
γ | − |aγ |

for γ ∈ R\{0}.
Taking Wronskians in (4.2.1) and using W (f, f ) = 0 and Proposition 4.2.1 we obtain

− 1
a+ + + + +
γ W (fγ , f−γ ) = W (fγ , f−γ ) =⇒ aγ = − W (fγ− , f−γ
+
),
2iγ
− 1
b+ + + + +
γ W (f−γ , fγ ) = W (fγ , fγ ) =⇒ bγ = W (fγ− , fγ+ ).
2iγ

Similarly, we find
1 1
a−
γ =
−
W (fγ+ , f−γ ), b−
γ = − W (fγ+ , fγ− ), 1 = |b− 2 − 2
γ | − |aγ | .
2iγ 2iγ
−
It follows that b+
γ = bγ .
Chapter 4: Schrödinger operators and scattering data 59

Using the Jost solutions we now define the solution ψγ (x), γ ∈ R\{0}, by the boundary
conditions at ±∞;
(
T (γ) exp(−iγx), x → −∞,
ψγ (x) ∼
exp(−iγx) + R(γ) exp(iγx), x → ∞
Note that the first condition determines ψγ up to a multiplicative constant, which is determined
by the requirement that the coefficient of exp(−iγx) is 1 as x → ∞. The function T (γ) is the
transmission coefficient and R(γ) is the reflection coefficient.
A potential q satisfying the assumptions above is a reflectionless potential if R(γ) = 0 for
γ ∈ R\{0}.
Using Theorem 4.1.2 it follows that
ψγ (x) = T (γ) fγ− (x) = f−γ
+
(x) + R(γ) fγ+ (x).
It follows that
1 1 a+
γ a+
γ
T (γ) = +
= −, R(γ) = +
= −
.
bγ bγ bγ bγ
This implies for γ ∈ R\{0}
T (γ) = T (−γ), R(γ) = R(−γ), |T (γ)|2 + |R(γ)|2 = 1,
+
2iγ W (fγ− , f−γ ) (4.2.2)
T (γ) = , R(γ) = −
W (fγ− , fγ+ ) W (fγ− , fγ+ )
Physically, |T (γ)|2 + |R(γ)|2 = 1 can be interpreted as conservation of energy for transmitted
and reflected waves.
Note that the reflection coefficient is related to waves travelling from left to right, and
we could also have (equivalently) studied the reflection coefficient R− (γ) = a− −
γ /bγ for waves
travelling from right to left. Note that the transmission coefficient does not change. Relabeling
the reflection coefficient R+ (γ) = R(γ), we define for γ ∈ R (or γ ∈ R\{0} depending on the
potential q) the scattering matrix
 
T (γ) R− (γ)
S(γ) = ∈ U (2), (4.2.3)
R+ (γ) T (γ),
i.e. S(γ) is a 2 × 2-unitary matrix. To see this we note first that, as for R(γ),
R− (γ) 1
= a−
γ =
−
W (fγ+ , f−γ ), R− (−γ) = R− (γ), |T (γ)|2 + |R− (γ)|2 = 1.
T (γ) 2iγ
From the expressions in terms of Wronskians we find
R+ (γ) R− (−γ)
+ = 0 =⇒ T (γ)R− (−γ) + R+ (γ)T (−γ) = T (γ)R− (γ) + R+ (γ)T (γ) = 0
T (γ) T (−γ)
which shows that the columns of S(γ) are orthogonal vectors. Since we already established
that each column vector has length 1, we obtain S(γ) ∈ U (2).
60 Chapter 4: Schrödinger operators and scattering data

Exercise 4.2.3. Work out the relation between the scattering matrix S(γ) and the scattering
d2 d2
operator S as defined in Section 3.2 in the case L0 = − dx 2 , L = − dx2 + q (with appropriate

assumptions on the potential q). (Hint: use the Fourier transform to describe the spectral
decomposition of L0 in terms of a C2 -vector-valued measure, and describe the action of the
scattering operator in terms of this decomposition.)
Proposition 4.2.4. Assuming the conditions of Theorem 4.1.2 we have
lim T (γ) = 1, lim R(γ) = 0.
γ→±∞ γ→±∞

dfγ−
Proof. Write fγ− (x) = e−iγx + R0− (x), fγ+ (x) = eiγx + R0+ (x), dx
(x) = −iγe−iγx + R1− (x)
dfγ+
dx
= iγeiγx + R1+ (x), where Ri± (x), i = 0, 1, also depend on γ, so that
(x)

W (fγ− , fγ+ ) = e−iγx + R0− (x) iγeiγx + R1+ (x) − −iγe−iγx + R1− (x) eiγx + R0+ (x)

= 2iγ + iγeiγx R0− (x) + e−iγx R1+ (x) + R0− (x)R1+ (x)
+ iγe−iγx R0+ (x) − eiγx R1− (x) − R0+ (x)R1− (x).
Theorem 4.1.2 shows that for real x
1

|R0± | ≤ G(γ), |R1± | ≤ |γ|G(γ), G(γ) = exp(kqk1 /|γ|) − 1 = O .
|γ|
Hence for γ ∈ R\{0}
 W (fγ− , fγ+ )
 
 ≤ 2G(γ) + G(γ) 2 → 0,


 − 1 |γ| → ∞.
 2iγ 

Using (4.2.2) this gives the limit for the transmission coefficient, and from this the limit for
the reflection coefficient follows.
The transmission and reflection can be written as integrals involving the potential function
q, the solution ψγ , and the Jost solution.
Proposition 4.2.5. Assume q is continuous and satisfies the conditions of Theorem 4.1.2,
then for γ ∈ R\{0},
Z
1
R(γ) = q(x) ψγ (x) e−iγx dx,
2iγ R
Z
1
T (γ) = 1 + q(x) ψγ (x) eiγx dx,
2iγ R
Z
1 1
= 1− e−iγx q(x)fγ+ (x) dx.
T (γ) 2iγ R
In the three-dimensional analogue of Proposition 4.2.5, the last type of integrals is precisely
what can be measured in experiments involving particle collission. Again, the continuity of q
is not essential, cf. Exercise 4.2.2.
Chapter 4: Schrödinger operators and scattering data 61

Proof. Take f any solution to −f 00 = γ 2 f and g a solution to −g 00 +qg = γ 2 g. Proceeding as in

d
the proof of Proposition 4.2.1 we see that dx [W (f, g)](x) = q(x)f (x)g(x), so that integrating
gives Z b
W (f, g)(a) − W (f, g)(b) = q(x) f (x) g(x) dx.
a
Now use this result with f (x) = e−iγx , g(x) = ψγ (x), and take the limit a → −∞ and b → ∞.
Using the explicit expression of ψγ in terms of Jost solutions and Theorem 4.1.2, then gives
Z
q(x) ψγ (x) e−iγx dx = lim W (e−iγx , ψγ )(b) − lim W (e−iγx , ψγ )(a)
R b→∞ a→−∞
−iγx iγx
= R(γ)W (e ,e ) = 2iγ R(γ),
which gives the result for the reflection coefficient. The limits of the Wronskian follow from
Theorem 4.1.2, cf. proof of Proposition 4.2.4. The statement for the transmission coefficient
follows similarly with f (x) = eiγx .
To find other integral representations of the reflection and transmission coefficients we
write
Z ∞ 2iγ(y−x)
+ e −1
mγ (x) = 1 + q(y)m+
γ (y) dy
x 2iγ
e−2iγx ∞ 2iγy
Z  Z ∞ 
+ 1 +
= e q(y)mγ (y) dy + 1 − q(y)mγ (y) dy
2iγ x 2iγ x
e−2iγx
Z  Z 
2iγy + 1 +
= e q(y)mγ (y) dy + 1 − q(y)mγ (y) dy + o(1), x → −∞.
2iγ R 2iγ R
Note that the remainder terms
Z x Z x
e2iγy q(y)m+
γ (y) dy, q(y)m+
γ (y) dy
−∞ −∞

are o(1) as x → −∞, since, by Theorem 4.1.2, m+

γ is bounded for fixed γ ∈ R\{0}, and q is
integrable by assumption. So
e−iγx
Z  Z 
+ iγy + iγx 1 −iγy +
fγ (x) = e q(y)fγ (y) dy + e 1− e q(y)fγ (y) dy + o(1), x → −∞,
2iγ R 2iγ R
so b−
R −iγy −
1 1
e q(y)fγ+ (y) dy. Since T (γ) = 1/b−
+
R iγy
γ = 1 − 2iγ R
e q(y)f γ (y) dy and a γ = 2iγ R γ the
result follows.
Using Proposition 4.2.5 and its proof we can refine the asymptotic behaviour of the trans-
mission coefficient. Corollary 4.2.6 and its proof also indicate how to obtain the asymptotic
1
expansion of T (γ) for γ → ±∞.
Corollary 4.2.6. With the assumptions as in Proposition 4.2.5, then for =γ ≥ 0,
Z
1 1 1
=1− q(y) dy + O( 2 ).
T (γ) 2iγ R |γ|
62 Chapter 4: Schrödinger operators and scattering data
P∞
Proof. Using e−iγx fγ+ (x) = m+
γ (x) = 1 +
1
n=1 mn (x) with mn (x) = O( |γ|n ) for γ → ±∞, see
proof of Theorem 4.1.2, in the last integral equation of Proposition 4.2.5 gives the result.
Recall that a function is meromorphic in an open domain of C if it is holomorphic except
at a set without accumulation points where the function has poles. Such a function can always
be written as a quotient of two holomorphic functions.
Theorem 4.2.7. Assume that q satisfies the assumptions of Theorems 4.1.5, 2.3.4 and Propo-
sition 4.1.10 and assume moreover q ∈ L∞ (R). The transmission coefficient T is meromorphic
in the the open upper half plane =γ > 0 with a finite number of simple poles at ipn , pn > 0,
1 ≤ n ≤ N . Moreover, T is continuous in =γ ≥ 0 except for γ = 0, γ = ipn , 1 ≤ n ≤ N .
Then −p2n , 1 ≤ n ≤ N , are the simple eigenvalues of the corresponding Schrödinger operator
with eigenfunction fn = fip+n . Put Cn = limx→∞ epn x fip−n , then
i 1 ρn
Resγ=ipn T = = i . (4.2.4)
Cn kfn k2 Cn
Proof. By Corollary 4.1.4 we see that W (fγ− , fγ+ ) is analytic in γ for =γ > 0. By (4.2.2) T
is an analytic function in the open upper half plane except for poles that can only occur at
zeros of γ 7→ W (fγ− , fγ+ ), =γ > 0.
Next using the integral representation
Z
1 1
=1− q(x) m+γ (x) dx
T (γ) 2iγ R
as in Proposition 4.2.5 and the estimate |m+ γ (x)| ≤ C(1 + |x|), see Theorem 4.1.5 and its
1
proof, we get that T (γ) is continuous in =γ ≥ 0, γ 6= 0, since m+ γ is continuous in =γ ≥ 0. By
(4.2.2) we have |T (γ)|−1 ≥ 1 for γ ∈ R\{0}, so that T (γ) = T (γ)1 −1 is continuous near =γ = 0,
γ 6= 0.
For γ ∈ C, =γ > 0, we put γ = α + iβ, β > 0, and consider, using Proposition 4.2.5,
Z ∞ 2iγ(y−x)
+ e −1
mγ (x) = 1 + q(y) m+γ (y) dy
x 2iγ
Z x Z ∞ 2iγ(y−x)
1 1 + e
= + q(y) mγ (y) dy + q(y) m+
γ (y) dy.
T (γ) 2iγ −∞ x 2iγ
We want to use this expression to establish the behaviour of m+γ (x) as x → −∞, so we assume
x < 0. By Theorem 4.1.2 we see that
 Z x  Z x
 1 +


 2iγ q(y) mγ (y) dy  ≤ C
 |q(y)| dy = o(1), x → −∞,
−∞ −∞

with C depending on q and γ, since q is integrable. Next, with C depending on q and γ,

Z ∞ 2iγ(y−x)  Z x/2 Z ∞
e −2β(y−x)
e−2β(y−x) |q(y)| dy
 +

 q(y) mγ (y) dy  ≤ C
 e |q(y)| dy + C

x 2iγ x x/2
Z x/2 Z
≤C |q(y)| dy + Ceβx |q(y)| dy
x R
Chapter 4: Schrödinger operators and scattering data 63

by Theorem 4.1.2. The first term is o(1) as x → −∞, since q is integrable, and the second
term O(exβ ), x → −∞. So we obtain for γ, =γ > 0, the asymptotic behaviour

1
m+
γ (x) = + o(1), x → −∞.
T (γ)

Similarly, one shows m−γ (x) = (T (γ))

−1
+ o(1), x → ∞, for γ in the open upper half plane.
Suppose that γ0 , =γ0 > 0, is not a pole of T , so T (γ1 0 ) 6= 0. Then fγ+0 and fγ−0 are linearly
independent solutions and hence span the solution space, since the Wronskian is non-zero. By
the above result we have

eiγ0 x
fγ+0 (x) = eiγ0 x m+
γ0 (x) = + o(e−x=γ0 ), x → −∞ =⇒ fγ+0 ∈
/ L2 (R)
T (γ0 )

Similarly, fγ−0 6∈ L2 (R), and since fγ+0 and fγ−0 span the solution space, there is no square
integrable solution, hence γ02 is not an eigenvalue of the corresponding Schrödinger operator.
Assume next that T −1 has a zero at γ0 with =γ0 > 0, so that fγ−0 is a multiple of fγ+0 . By
Theorem 4.1.2 it follows that fγ+0 , respectively fγ−0 , is a square integrable function for x → ∞,
respectively for x → −∞. Since fγ−0 is a multiple of fγ+0 , it follows that fγ+0 ∈ L2 (R). Theorem
df + df −
4.1.2 then also gives dxγ0 = dxγ0 ∈ L2 (R), and since (fγ±0 )00 (x) = −γ02 fγ±0 (x)+q(x)fγ±0 (x) ∈ L2 (R)
since we assume q ∈ L∞ (R). This then gives that the eigenfunction is actually in the domain
W 2 (R). Hence fγ+0 is an eigenfunction for the corresponding Schrödinger equation for the
eigenvalue γ02 . Since the Schrödinger operator is self-adjoint by Corollary 2.2.7, we have
γ02 ∈ R and so γ0 ∈ iR>0 .
So the poles of T are on the positive imaginary axis, and such a point, say ip, p > 0,
corresponds to an eigenvalue −p2 of the corresponding Schrödinger operator. Since q ∈ L∞ (R)
we have that the spectrum is contained in [−kqk∞ , ∞), and by Theorems 2.3.4 and 2.3.2 its
essential spectrum is [0, ∞). By Theorem 6.5.5 it follows that the intersection of the spectrum
with [−kqk∞ , 0) can only have a finite number of points which all correspond to the point
spectrum.
So we can label the zeros of the Wronskian on the positive imaginary axis as ipn , pn >
0, n ∈ {1, 2, · · · , N }, and then fn (x) = fip+n (x) is a square integrable eigenfunction of the
Schrödinger operator. We put
Z
1 1
= |fn (x)|2 dx, kfn k = √ .
ρn R ρn

It remains

−1 to show that the residue of T at ipn is expressible in terms of ρn . Start with
2iγ T (γ) = W (fγ+ , fγ− ), which is an equality for analytic functions in the open upper half
plane. Differentiating with respect to γ gives

2i dT −1 ∂fγ+ − +
∂fγ−
+ 2iγ (γ) = W ( , f )(x) + W (fγ , )(x),
T (γ) dγ ∂γ γ ∂γ
64 Chapter 4: Schrödinger operators and scattering data

where we have emphasized the x-dependence in the Wronskian. In particular, for γ = γ0 a

pole of T , we get

dT −1 ∂fγ+ ∂fγ−
2iγ0 (γ0 ) = W ( |γ=γ0 , fγ−0 )(x) + W (fγ+0 , |γ=γ0 )(x).
dγ ∂γ ∂γ
∂f −
Since fγ+ is a solution to −f 00 + q f = γ 2 f and ∂γγ is a solution to −f 00 + q f = γ 2 f + 2γ fγ− ,
cf. Proposition 4.1.10, we find, cf. the proof of Proposition 4.2.1,

d ∂fγ−
d2 ∂fγ− d2 fγ+ ∂fγ−

+
W (fγ , +
) (x) = fγ (x) 2 (x) − 2
(x) (x) = 2γfγ+ (x) fγ− (x),
dx ∂γ dx ∂γ dx ∂γ
so that
x ∂fγ− ∂fγ−
Z
2γ fγ+ (y) fγ− (y) dy = W (fγ+ , +
)(x) − W (fγ , )(a)
a ∂γ ∂γ
and this we want to consider more closely for γ = γ0 a pole of the transmission coefficient T .
Since we already have established that poles lie on the positive imaginary axis we can take
γ0 = ipn , pn > 0. Then fn (x) = fip+n (x) is a multiple of fip−n (x) and Theorem 4.1.2 implies
that fn (x) and its derivate fn0 (x) behave like epn x as x → −∞. Combined with Proposition
4.1.10, we see that we can take the limit a → −∞ to get

∂fγ−
Z x
2ipn fn (y) fip−n (y) dy = W (fip+n , |γ=−ipn )(x).
−∞ ∂γ

In a similar way we get

Z ∞ ∂fγ+
2ipn fn (y) fip−n (y) dy = W ( |γ=ipn , fip−n )(x)
x ∂γ

so that we obtain
dT −1
Z
−2 pn (ipn ) = 2ipn fn (y) fip−n (y) dy.
dγ R

Since fip±n are real-valued, cf. remark following Theorem 4.1.2, non-zero and multiples of each
other we see that fip−n = Cn fip+n (x) = Cn fn (x) for some real non-zero constant Cn , so that

dT −1
Z
Cn
(ipn ) = |fn (y)|2 dy 6= 0.
dγ i R

Note that Cn = limx→∞ Cn epn x fip+n (x) = limx→∞ epn x fip−n . It follows that the zero of 1
T
at ipn
R −1
is simple, so that T has a simple pole at ipn with residue Cin R |fn (y)|2 dy .

Exercise 4.2.8. The statement on the simplicity of the eigenvalues in Theorem 4.2.7 has not
been proved. Show this.
Chapter 4: Schrödinger operators and scattering data 65

Definition 4.2.9. Assume the conditions as in Theorem 4.2.7, then the transmission coeffi-
cient T , the reflection coefficient R together with the poles on the positive imaginary axis with
the corresponding square norms; {(pn , ρn ) | pn > 0, 1 ≤ n ≤ N } constitute the scattering data.

Given the potential q, the direct scattering problem constitutes of determining T , R and
{(pn , ρn )}.

Remark 4.2.10. It can be shown that it suffices to take the reflection coefficient R to-
gether with {(pn , ρn ) | pn > 0, 1 ≤ n ≤ N } as the scattering data, since the transmis-
sion coefficient
p can be completely recovered from this. Indeed, the norm of T follows from
|T (γ)| = 1 − |R(γ)|2 , see 4.2.2, and the transmission coefficient can be completely recon-
structed using complex function techniques and Hardy spaces.

Exercise 4.2.11. Work out the scattering data for the cosh−2 -potential using the results as
in Section 2.5.2. What can you say about the scattering data for the other two examples in
Sections 2.5.1, 2.5.3?

4.3 Fourier transform properties of the Jost solutions

We assume that the potential q satisfies the conditions of Theorem 4.1.5. By Corollary 4.1.7
+
we see that γ 7→ m+
γ (x) − 1 is in the Hardy class H2 for each x ∈ R, see Section 6.3. So by
the Paley-Wiener Theorem 6.3.2 it is the inverse Fourier transform of an L2 (0, ∞) function;
Z ∞
+
mγ (x) = 1 + B(x, y) e2iγy dy,
0

where we have switched to 2y instead of y in order to have nicer looking formulas in the sequel.
Note that for the Jost solution we have
Z ∞
+ iγx 1 1
fγ (x) = e + A(x, y) eiγy dy, A(x, y) = B(x, (y − x)).
x 2 2

So the Paley-Wiener Theorem 6.3.2 gives B(x, ·) ∈ L2 (0, ∞) for each x ∈ R, but the kernel B
satisfies many more properties. Let us consider the kernel B more closely. Recall the proof of
Theorem 4.1.2 and (4.1.4), and write
Z ∞ ∞
1 2iγ(y−x) X
m+
γ (x) −1= (e − 1)q(y) dy + mn (x; γ),
x 2iγ n=2

where we have emphasized mn (x; γ) = mn (x) to express the dependence of mn on γ. Then

y 7→ 12 B(x, 12 y) is the inverse Fourier transform of m+ γ (x) − 1 with respect to γ. By the
estimates in the proof P of Theorem 4.1.2 it follows that each mn (x; γ), n ≥ 2, is of order
O( |γ|1 2 ), so we see that ∞ 1
n=2 mn (x; γ) is in L (R) as function of γ, so its Fourier transform,
66 Chapter 4: Schrödinger operators and scattering data

say 12 B1 (x, 21 y), is in C0 (R) by the Riemann-Lebesgue Lemma. So it remains to take the
Fourier transform of the integral. First recall, cf. proof of Theorem 4.1.5, that the integrand
is majorized by (y − x)|q(y)| which is assumed to be integrable on [x, ∞) and the integral is
1
of order O( |γ| ), so that it is in L2 (R) with respect to γ. Write
Z ∞ ∞ Z ∞
1 2iγ(y−x) 1 2iγ(y−x)
(e − 1)q(y) dy = − (e − 1)w(y) + w(y) e2iγ(y−x) dy

2iγ 2iγ y=x
Zx ∞ x
1 ∞
Z
2iγy 1 iγy
= w(x + y) e dy = w(x + y) e dy
0 2 0 2
R∞
where we put w(x) = x q(y) dy, so that limx→∞ w(x) = 0. So we have written the integral as
the inverse Fourier transform of 12 H(y)w(x + 21 y), where H is the Heaviside function H(x) = 1
for x ≥ 0 and H(x) = 0 for x < 0. So we obtain
B(x, y) = H(y)w(x + y) + B1 (x, y), B1 (x, ·) ∈ C0 (R).
R∞
In particular, B(x, 0) = w(x) = x q(t) dt.
00 0
Since m+ +
γ (x) is a solution to m + 2iγ m = qm, we see that mγ (x) − 1 is a solution to
dm+
m00 + 2iγ m0 = qm + q. Since limx→∞ m+ γ
γ (x) − 1 = 0 and limx→∞ dx (x) = 0, we can integrate
this over the interval [x, ∞) to find the integral equation for m+γ;
Z ∞ Z ∞
0
−m (x) − 2iγm(x) = q(y) m(y) dy + q(y) dy.
x x

Theorem 4.3.1. Assume q satisfies the conditions of Theorem 4.1.5. The integral equation
Z ∞ Z yZ ∞
B(x, y) = q(t) dt + q(t) B(t, z) dt dz, y ≥ 0,
x+y 0 x+y−z

has a unique real-valued solution B(x, y) satisfying

Z ∞ Z ∞

|B(x, y)| ≤ |q(t)| dt exp (t − x) |q(t)| dt .
x+y x

So in particular, B(x, ·) ∈ L∞ (0, ∞) ∩ L1 (0, ∞) ⊂ Lp (0, ∞), 1 ≤ p ≤ ∞, with

Z ∞ Z ∞


kB(x, ·)k∞ ≤ |q(t)| dt exp (t − x) |q(t)| dt ,
Zx ∞ x
Z ∞


kB(x, ·)k1 ≤ (t − x)|q(t)| dt exp (t − x) |q(t)| dt .
x x

Moreover, B is a solution to
∂

Bx (x, y) − By (x, y) = q(x)B(x, y)
∂x
R∞
with boundary conditions B(x, 0) = Rx q(t) dt and limx→∞ kB(x, ·)k∞ = 0.
∞
Moreover, defining m(x; γ) = 1 + 0 B(x, y) e2iγy dy, then the function e−iγx m(x; γ) is the
Jost solution fγ+ for the corresponding Schrödinger operator.
Chapter 4: Schrödinger operators and scattering data 67

The partial differential equation is a hyperbolic boundary value problem known as a Gour-
sat partial differential equation.
Proof. We solve the integral equation by an iteration as before. Put B(x, y) = ∞
P
n=0 Kn (x, y)
with Kn (x, y) defined recursively
Z ∞ Z yZ ∞
K0 (x, y) = q(t) dt, Kn+1 (x, y) = q(t) Kn (t, z) dt dz.
x+y 0 x+y−z

It is then clear that B(x, y) solves the integral equation provided the series converges. We
claim that

n Z ∞ Z ∞
R(x)
|Kn (x, y)| ≤ S(x+y), R(x) = (t−x)|q(t)| dt, S(x) = |q(t)| dt. (4.3.1)
n! x x
R∞
Note that 0 S(x + y) dy = R(x). Let us first assume the claim is true. Then the series
converges uniformly on compact sets in R2 and the required estimate follows. It is then also
clear that Kn (x, y) ∈ R for all n since q is real-valued, so that B is real-valued. We leave
uniqueness as an exercise, cf. the uniqueness proof in the proof of Theorem 4.1.2.
Observe that S(x + y) ≤ S(x) for y ≥ 0, so that estimate on the L∞ (0, ∞)-norm of B(x, ·)
follows immediately, and this estimate also gives limx→∞ kB(x, ·)k∞ = 0, since S(x) → 0 as
x → ∞. For the L1 (0, ∞)-norm we calculate
Z ∞

∞
Z


|B(x, y)| dy ≤ exp R(x) S(x + y) dy = exp R(x) R(x)
0 0

Differentiating with respect to x and y gives

Z y
∂B
(x, y) = −q(x + y) − q(x + y − z) B(x + y − z, z) dz,
∂x 0
Z ∞ Z y
∂B
(x, y) = −q(x + y) + q(t) B(t, y) dt − q(x + y − z) B(x + y − z, z) dz.
∂y x 0

Here we use Lebesgue’s Differentiation Theorem 6.1.4, so the resulting identities hold almost
everywhere. Unless we impose differentiability conditions on q, weR cannot state anything
∞
about the higher order partial derivatives, but Bx (x, y) − By (x, y) = x q(t) B(t, y) dt can be
differentiated with respect to x, and this gives the required partial differential equation. We
also obtain
Z y
 ∂B 
 (x, y) + q(x + y) ≤
 |q(x + y − z) B(x + y − z, z)| dz,
∂x 0
Z y
≤ |q(x + y − z)| exp(R(x + y − z)) S(x + y) dz (4.3.2)
0
Z x+y
≤ S(x + y) exp(R(x)) |q(t)| dt ≤ S(x + y) exp(R(x)) S(x)
x
68 Chapter 4: Schrödinger operators and scattering data

since R is decreasing. The one but last estimate can also be used to observe that
Z x+y
 ∂B 
 (x, y) + q(x + y) ≤ S(x) exp(R(x))
 |q(t)| dt → 0, y ↓ 0,
∂x x

since S(x + y) ≤ S(x) and q is integrable. This gives the other boundary condition for the
partial differential equation. Similarly,
Z y Z ∞
 ∂B 
 (x, y) + q(x + y) ≤ |q(x + y − z) B(x + y − z, z)| dz + |q(t) B(t, y)| dt,
∂y 0 x
Z y Z ∞
≤ |q(x + y − z)| exp(R(x + y − z)) S(x + y) dz + |q(t)| exp(R(t)) S(t + y) dt
0 x
Z x+y Z ∞
≤ S(x + y) exp(R(x)) |q(t)| dt + S(x + y) exp(R(x)) |q(t)| dt
x x
≤ 2 S(x + y) exp(R(x)) S(x).
R∞
Define n(x; γ) = 0 B(x, y) e2iγy dy, then by (4.3.2) we find | ∂B ∂x
(x, y)| ≤ |q(x + y)| + S(x +
y)S(x) exp(R(x)) and the right hand side is integrable with respect to y ∈ [0, ∞). So we can
differentiate with respect to x in the integrand to get
Z ∞ Z ∞  Z ∞
0 2iγy 2iγy
n (x, γ) = Bx (x, y) e dy = Bx (x, y) − By (x, y) e dy + By (x, y) e2iγy dy
0 0
Z ∞Z ∞ ∞ Z ∞0
= − q(t) B(t, y) dt e2iγy dy + B(x, y)e2iγy  − 2iγ B(x, y) e2iγy dy

y=0
Z0 ∞ Zx ∞ Z ∞ 0

= − q(t) B(t, y) dt e2iγy dy − B(x, 0) − 2iγ B(x, y) e2iγy dy

Z ∞0 x
Z ∞ Z ∞0
= q(t) B(t, y) e2iγy dydt − B(x, 0) − 2iγ B(x, y) e2iγy dy
x 0 0
Z ∞ Z ∞
= − q(t) n(t; γ) dt − q(t) dt − 2iγ n(x; γ)
x x

since limy→∞R ∞B(x, y) = 0 by limy→∞ S(x + y) = 0, and where Fubini’s theorem is applied and
B(x, 0) = x q(t) dt. So n(x; γ) satisfies the integrated version of the differential equation
m00 + 2iγm0 = qm + q. Since |n(x; γ)| ≤ kB(x, ·)k1 → 0 as x → ∞, it follows that n(γ; x) =
m+γ (x) − 1 and the result follows.
It remains to prove the estimate in (4.3.1). This is proved by induction on n. The case
n = 0 is trivially satisfied by definition of S(x + y). For the induction step, we obtain by the
induction hypothesis and the monotonicity of S the estimate
Z yZ ∞
(R(t))n
|Kn+1 (x, y)| ≤ |q(t)| S(t + z) dt dz
0 x+y−z n!
Z yZ ∞
(R(t))n
≤ S(x + y) |q(t)| dt dz
0 x+y−z n!
Chapter 4: Schrödinger operators and scattering data 69

and it remains to estimate the double integral. Interchanging the order of integration, the
integral equals
Z x+y Z ∞
(R(t))n y (R(t))n y
Z Z
|q(t)| dz dt + |q(t)| dz dt
x n! x+y−t x+y n! 0
Z x+y Z ∞
(R(t))n (R(t))n
= (t − x)|q(t)| dt + y|q(t)| dt
x n! x+y n!
Z ∞ Z ∞ Z ∞ n
(R(t))n 1
≤ (t − x)|q(t)| dt = (t − x)|q(t)| (u − t)|q(u)| du dt
x n! x n! t
Z ∞ Z ∞ n
1 (R(x))n+1
≤ (t − x)|q(t)| (u − x)|q(u)| du dt =
x n! t (n + 1)!
using y ≤ t − x in the first inequality and u − t ≤ u − x for x ≥ t in the second inequality.
This proves the induction step.
Exercise 4.3.2. Work out the kernel B for the cosh−2 -potential using the results as in Section
2.5.2, see also Section 4.5. What can you say about the kernel B for the other two examples
in Sections 2.5.1, 2.5.3?
Note that in the above considerations for the Jost solution γ ∈ R, but it is possible to
extend the results to γ in the closed upper half plane, =γ ≥ 0.
Proposition 4.3.3. Assume that q satisfies the assumptions of Theorem 4.1.5. The relations
Z ∞ Z ∞
+ 2iγy + iγx
mγ (x) = 1 + B(x, y) e dy, fγ (x) = e + A(x, y) eiγy dy
0 x

remain valid for =γ ≥ 0.

R∞
Proof. Fix x ∈ R and consider F (γ) = m+ γ (x)−1− 0 B(x, y)e
2iγy
dy, then we know that F is
analytic in the open upper half plane =γ > 0 and continuous on =γ ≥ 0. Moreover, F (γ) = 0
1
R F (z)
for γ ∈ R. We now use for fixed γ in the open upperhalf plane F (γ) = 2πi C z−γ
dz for a
suitable contour C in the open upper half plane. We take C to be a semicircle of sufficiently
large radius R just above the real axis at =z = ε < =γ. This gives the estimate
Z R+iε
|F (z)| |F (z)|
Z
1 1
|F (γ)| ≤ dz + dz
2π |z|=R,=z>0 |z − γ| 2π −R+iε |z − γ|
Z π Z R
1 R iθ 1 |F (t + iε)|
≤ |F (Re )| dθ + dt.
2π dist(γ, |z| = R) 0 2π −R |t + iε − γ|
For γ in the open upper half plane we can take ε ↓ 0 in the second integral by dominated
convergence, which then gives zero. The factor in front of the first integral is bounded as
R → ∞ for γ fixed in the upper half plane, so the result follows from the claim
Z π
lim |F (Reiθ )| dθ = 0.
R→∞ 0
70 Chapter 4: Schrödinger operators and scattering data
Rπ
To prove this first observe that 0 |m+ Reiθ
− 1| dθ → 0 as R → ∞, since the integrand is O( R1 )
by Theorem 4.1.2. So it remains to estimate
Z πZ ∞ Z πZ ∞
2iReiθ y
|B(x, y)e | dy dθ = |B(x, y)|e−2Ry sin θ dy dθ
0 0
! Z0 0
Z Z ε(R) Z π−ε(R) π ∞
= + + |B(x, y)|e−2Ry sin θ dy dθ,
0 ε(R) π−ε(R) 0

where ε(R) ∈ (0, 21 π) is choosen such that sin θ ≥ √1R for all θ ∈ (ε(R), π − ε(R)). We have
sin ε(R) = √1R and ε(R) ∼ √1R as R → ∞. Now the integral can be estimated by
Z ∞ Z ∞ √ π
2ε(R) |B(x, y)| dy + π e−2y R |B(x, y)| dy = 2ε(R)kB(x, ·)k1 + √ kB(x, ·)k2 → 0
0 0 24R
as R → ∞.
Exercise 4.3.4. Work out the proof of the second statement of Proposition 4.3.3.
.

4.4 Gelfand-Levitan-Marchenko integral equation

In this section we consider the inverse scattering problem, namely how to construct the poten-
tial from the scattering data. From Theorem 4.3.1 we see that the potential q can be recovered
from the kernel B for the Jost solution fγ+ . So if we can characterize the kernel B in terms
of scattering data we are done, and this is described by the Gelfand5 -Levitan6 -Marchenko7
integral equation. To see how this comes about we first start with a formal calculation. We
assume that T has no poles in the open upper half plane, i.e. the corresponding Schrödinger
operator has no discrete spectrum. Assume that the reflection coefficient R(γ) has a Fourier
transform R(γ) = R r(y)eiγy dy, where we absorb the factor √12π in r, then using the results
R

in Section 4.3 give

T (γ) fγ− (x) = f−γ
+
(x) + R(γ) fγ+ (x)
Z ∞ Z Z Z ∞
−iγx −iγy
=e + A(x, y)e dy + r(y)e iγ(x+y) iγz
dy + r(z)e dz A(x, y)e−iγy dy
x R R x

and this is rewritten as

Z −x Z Z Z ∞ 
iγy iγy
A(x, −y)e dy + r(y − x)e dy + A(x, z) r(y − z) dz eiγy dy
−∞ R R x
= T (γ) fγ− (x) −e −iγx
.
5
Israil Moiseevic Gelfand (2 September 1913 — ), Russian mathematician, one of the greatest mathemati-
cians of the 20th century.
6
Boris Levitan (1915 — 4 April 2004), Russian mathematician.
7
Vladimir Aleksandrovich Marchenko, (7 July 1922 — ), Ukrainian mathematician.
Chapter 4: Schrödinger operators and scattering data 71

+
Now as in Section 4.3 we have that m− γ (x) − 1 is in the Hardy class H2 , so we similarly get
∞
fγ− (x) = e−iγx + −x A− (x, y)eiγy dy, hence the right hand side equals
R

T (γ) fγ− (x) − e−iγx = T (γ) − 1 e−iγx + T (γ) m−

−iγx
γ (x) − 1 e .
+
Now if γ 7→ T (γ) − 1 and γ 7→ T (γ) m−


γ (x) − 1 are of Hardy class H2 , which can be shown
+
by using that γ 7→ m− γ (x) − 1 is in H2 and by refining the estimates on T in Proposition 4.2.4,
it follows from the Paley-Wiener Theorem 6.3.2 and shifting by x that Fourier transform of
the right hand side is supported on [−x, ∞). But the left hand side is a Fourier transform, so
that Z ∞
A(x, −y) + r(y − x) + A(x, z) r(y − z) dz = 0, y < −x. (4.4.1)
x
So assuming the reflection coefficient, hence its Fourier transform, is known the kernel A,
satisfies (4.4.1). Assuming that the integral equation can be uniquely solved, the potential q
then follows from B by a formula as in Theorem 4.3.1. The integral equation (4.4.1) is known
as the Gelfand-Levitan-Marchenko equation. We see that the Gelfand-Levitan-Marchenko
equation is essentially expressing a fact on the Fourier transform of the fundamental relation
+
T (γ) fγ− (x) = f−γ (x) + R(γ) fγ+ (x) between Jost solutions.
In case the Schrödinger operator has point spectrum, the transmission coefficient has
poles. In this case the Fourier transform is not identically zero for y < −x, but there are
non-zero contributions. The idea is to find these non-zero contributions by shifting the path
in R T (γ)fγ− (x)e−iγx dγ to =γ is a large enough constant, and by picking residues at ipn ,
R

1 ≤ n ≤ N , with notation and residues as in Theorem 4.2.7. This idea shows that the above
approach can be adapted by adding a finite sum to the kernel r in (4.4.1).
In order to make the above approach rigorous we need to investigate the reflection co-
efficient more closely. Instead of dealing with the kernel A we deal with the kernel B as
considered in Theorem 4.3.1.
Proposition 4.4.1. Assume the potential q satisfies the conditions of Theorem 4.1.2, then
the reflection coefficient R ∈ L2 (R), so that its inverse Fourier transform, up to scaling and
1
R
linear transformation of the argument, Kc : y 7→ π R R(γ)e2iγy dy is defined as an element in
L2 (R).
So Kc (x) = π1 R R(γ) e2iγx dγ is considered as an element of L2 (R). Since R(γ) = R(−γ)
R

we see that Kc (x) = π1 R R(−γ) e−2iγx dγ = Kc (x) or Kc is real-valued.

R

Proof. First observe that the reflection coefficient is bounded on R by (4.2.2). From the proof
1 1 1
of Proposition 4.2.4 it follows that T (γ) − 1 = O( |γ| ) as γ → ±∞. So T (γ) = 1 + O( |γ| ) and
p 1 2 1 2
so is |R(γ)| = 1 − |T (γ)|2 = O( |γ| ). So |R(γ)| = O( |γ|2 ), and R ∈ L (R).
Assuming the conditions of Theorem 4.2.7, we can consider the discrete spectrum of the
corresponding Schrödinger operator. With the notation of Theorem 4.2.7 we define
N
X
Kd (x) = 2 ρn e−2pn x
n=1
72 Chapter 4: Schrödinger operators and scattering data

and Kd (x) = 0 if the transmission coefficient has no poles, i.e. the Schrödinger operator has no
discrete eigenvalues. So we let Kc , respectively Kd , correspond to the continuous, respectively
discrete, spectrum of the Schrödinger operator. Now define, K = Kc + Kd as the sum of two
square integrable functions on the interval [a, ∞) for any fixed a ∈ R.
Theorem 4.4.2. Assume the potential q satisfies the conditions of Theorem 4.2.7. Then the
kernel B satisfies the integral equation
Z ∞
K(x + y) + B(x, z) K(x + y + z) dz + B(x, y) = 0,
0
for allmost all y ≥ 0 which for each fixed x ∈ R is an equation that holds for B(x, ·) ∈ L2 (0, ∞).
Moreover, the integral equation has a unique solution B(x, ·) ∈ L2 (0, ∞).
The integral equation is the Gelfand-Levitan-Marchenko equation, and we see that the
kernel B, hence by Theorem 4.3.1 the potential q, is completely determined by the scattering
data. So in this way the inverse scattering problem is solved. In particular, the transmission
coefficient T is not needed, cf. Remark 4.2.10, and hence is determined by the remainder of
the scattering data.
+
Proof. So consider T (γ) fγ− (x) = f−γ (x) + R(γ) fγ+ (x) and rewrite this in terms of m±
±γ (x) to
find for γ ∈ R\{0}
T (γ) m− +
γ (x) = m−γ (x) + R(γ) e
2iγx
m+
γ (x)
Z ∞  Z ∞ 
−2iγy 2iγx
=1+ B(x, y) e dy + R(γ)e 1+ B(x, y) e2iγy dy
Z ∞ 0 0
Z ∞
− −2iγy 2iγx 2iγx
=⇒ T (γ) mγ (x) − 1 = B(x, y) e dy + R(γ)e + R(γ)e B(x, y) e2iγy dy
0 0
using Theorem 4.3.1. As functions of γ, the first term on the right hand side is an element of
L2 (R) by Theorem 4.3.1, and similarly for the last term on the right hand side using that R(γ)
is bounded by 1, see (4.2.2). Proposition 4.4.1 gives that the middle term on the right hand
side is an element of L2 (R). So we see that the function on the left hand side is
an element of
L2 (R), which we can also obtain directly from T (γ) m− −


γ (x)−1 = T (γ) mγ (x)−1 + T (γ)−1 .
By Corollary 4.1.7 and the boundedness of T , the first term is even in the Hardy class H2+
and the second term is in L2 (R) by the estimates in the proof of Proposition 4.2.4.
So we take the inverse Fourier transform of this identity for elements of L2 (R). Since
we have switched to other arguments we use the Fourier transform in the following form, cf.
Section 6.3;
Z Z
−2iλy 1
g(λ) = f (y) e dy, f (y) = (Gg)(y) = g(λ) e2iλy dλ,
R π R
where G is a bounded invertible operator on L2 (R). So applying G we get an identity in L2 (R)
for each fixed x;
G γ 7→ T (γ) m− 2iγx



γ (x) − 1 = B(x, ·) + G γ 7→ R(γ)e
Z ∞
2iγx
B(x, z) e2iγz dz .


+ G γ 7→ R(γ)e
0
Chapter 4: Schrödinger operators and scattering data 73

It remains to calculate the three G-transforms explicitly. Since the identity is not in L2 (R) we
have to employ approximations.

We first calculate G γ 7→ R(γ)e2iγx . Put Rε (γ) = exp(−εγ 2 ) R(γ), so that Rε ∈ L1 (R) ∩
L2 (R), and note that
Z Z
2
|Rε (γ)e2iγx 2iγx 2
− R(γ)e | dγ = |R(γ)|2 |1 − e−εγ |2 dγ → 0, ε ↓ 0,
R R

by the Dominated Convergence Theorem 6.1.3. By continuity of G on L2 (R) we then have

lim G γ 7→ Rε (γ)e2iγx = G γ 7→ R(γ)e2iγx in L2 (R).

ε↓0

Now
Z
1 2
2iγx
R(γ) e−εγ e2iγ(x+y) dy = G γ 7→ Rε (γ) (x + y)



G γ 7→ Rε (γ)e (y) =
π R

and this tends to y 7→ Kc (x + y) as an element of L2 (R) as ε ↓ 0.

Similarly we consider the other G transform on the right hand side;
Z Z ∞ Z ∞ 1 Z
1  2iγx 2iγz

2iγy 2iγ(x+y+z)

Rε (γ)e B(x, z)e dz e dγ = B(x, z) Rε (γ)e dγ dz,
π R 0 0 π R
2
since the integrals converge absolutely by |Rε (γ)| ≤ e−εγ and B(x, ·) ∈ L1 (R) by Theorem
4.3.1. The inner integral converges to Kc (x + y + z) as ε ↓ 0 in L2 (R), and since B(x, ·) ∈
L2 (0, ∞) by TheoremR4.3.1 the Cauchy-Schwarz inequality (6.1.1) for L2 (0, ∞) shows that the
∞
integral converges to 0 B(x, z)Kc (x + y + z) dz as ε ↓ 0.
It remains to calculate the G-transform of the left hand side, and this is where the poles of
the transmission coefficient play a role. By Theorem 4.2.7 the function γ 7→ T (γ)m− γ (x) − 1
is meromorphic in =γ > 0 for fixed x and continuous in =γ ≥ 0. Moreover, it has simple
poles at ipn , 1 ≤ n ≤ N , so we want to employ Cauchy’s Theorem. We first multiply by
1
1−iεγ
, ε > 0, which has a pole at γ = 1/iε = −i/ε in the open lower half plane. (Note that
exp(−εγ 2 ) is not a good modification, because we want to consider the function in the upper
half plane.) Then, for ε > 0,
Z
1 1
Iε (x, y) = (T (γ)m−γ (x) − 1)e
2iγy
dγ
π R 1 − iεγ
has an integrable integrand meromorphic in =γ > 0 and continuous in =γ ≥ 0. We only
consider this integral for y ≥ 0. We claim that
Z
1 1
Iε (x, y) = lim (T (γ)m−
γ (x) − 1)e
2iγy
dγ,
M →∞ π C 1 − iεγ
M

where CM is the closed contour in the complex plane consisting of the interval [−M, M ] on
the real axis and the half circle z = M eiθ , 0 ≤ θ ≤ π. (Note that we first observe this for CM
δ
74 Chapter 4: Schrödinger operators and scattering data

where the interval is taken [−M + iδ, M + iδ] for some δ > 0 and next let δ & 0.) This is a
consequence of

1
T (γ)m−
γ (x) − 1 = O( ), =γ ≥ 0,
|γ|

so that the integrand is O( |γ|1 2 ) and the integrand over the half circle tends to zero.
Now the contour integral can be evaluated, cf. proof of Proposition 4.3.3, by Cauchy’s
Theorem;

N
2πi X 1
Iε (x, y) = Resγ=ipn (T (γ)m−
γ (x) − 1)e
2iγy
π n=1 1 − iεγ
N
X 1
Resγ=ipn T (γ) m− −2pn y


= 2i ipn (x)e .
n=1
1 + εpn

Since fip−n (x) = Cn fip+n (x) = Cn fn (x), cf. proof of Theorem 4.2.7, we find m−
ipn (x) =
−2pn x +
e Cn mipn (x). With the value of the residue in Theorem 4.2.7, we find

N
X 1
lim Iε (x, y) = −2 m+ (x)e−2pn (x+y)
ε↓0
n=1
kfn k2 ipn
N N Z ∞
X 1 −2pn (x+y) X 1
= −2 2
e −2 2
B(x, z)e−2pn z dz e−2pn (x+y)
n=1
kfn k n=1
kf n k 0
Z ∞
= −Kd (x + y) − B(x, z) Kd (x + y + z) dz.
0

Combining the ingredients gives the required Gelfand-Levitan-Marchenko integral equation

for B.
In order to prove the uniqueness of the solution, we consider the operator f 7→ K (x) f with

Z ∞
(x)
K f (y) = K(x + y + z) f (z) dz,
0

then K (x) : L2 (0, ∞) → L2 (0, ∞) is a bounded operator as we show below. For uniqueness we
need to show that K (x) f + f = 0 only has the trivial solution f = 0 in L2 (R).
Chapter 4: Schrödinger operators and scattering data 75

First, to show that K (x) is bounded on L2 (0, ∞) we consider for f, g ∈ L2 (0, ∞)∩L1 (0, ∞),
Z ∞ Z ∞ 
(x)
hK f, gi = K(x + y + z) f (z) dz g(y) dy
0 0
N Z
X ∞ Z ∞
=2 ρn e−2pn (x+y+z) f (z) g(y) dz dy
n=1 0 0
Z ∞ Z ∞ Z
1
+ f (z)g(y)R(γ)e2iγ(x+y+z) dγ dy dz
π 0 0 R
N
X Z
F −1 f (2γ) F −1 ḡ (2γ) R(γ) e2iγx dγ



= hf, en ihen , gi + 2
n=1 R
N Z
X 1
F −1 f (γ) F −1 ḡ (γ) R( γ) eiγx dγ



= hf, en ihen , gi +
n=1 R 2
√
where en (y) = 2ρn e−2pn y−pn x is an element of L2 (0, ∞) and h·, ·i denotes the inner product of
the Hilbert space L2 (0, ∞), which we consider as a subspace of L2 (R) in the obvious way. Note
that interchanging of the integrals is justified because all integrals converge for f, g ∈ L1 (0, ∞).
Next we estimate the right hand side in terms of the L2 (0, ∞)-norm of g by
N
X
kgkL2 (0,∞) kf kL2 (0,∞) ken k2L2 (0,∞) + kF −1 f kL2 (R) kF −1 ḡkL2 (R)
n=1
N
X 
≤ kgkL2 (0,∞) kf kL2 (0,∞) ken k2L2 (0,∞) + 1 ,
n=1

since |R(γ) e2iγx | ≤ 1 and, recall L2 (0, ∞) ⊂ L2 (R), the Fourier transform being an isometry.
So K (x) is a bounded operator, and the above expression for hK (x) f, gi remains valid for
arbitrary f, g ∈ L2 (0, ∞).
So for a real-valued f ∈ L2 (0, ∞)
N Z
X 1
(x)
kf k2L2 (0,∞) 2
| F −1 f (γ)|2 R( γ) eiγx dγ


hf + K f, f i = + |hf, en i| +
n=1 R 2

so for real-valued f ∈ L2 (0, ∞) with K (x) f + f = 0 we get

Z N
−1 21 iγx
X
|hf, en i|2 ≤ 0


| F f (γ)| (1 + R( γ) e ) dγ = −
R 2 n=1

−1
R

using kf k2L2 (0,∞) = R
| F f (γ)|2 dγ. Since f is real-valued, x, γ ∈ R, we have

1 1
| F −1 f (γ)|2 (1 + R( γ) eiγx ) = | F −1 f (−γ)|2 (1 + R(− γ) e−iγx )



2 2
76 Chapter 4: Schrödinger operators and scattering data

by (4.2.2), so
Z ∞ Z ∞
−1 1 1
2 iγx
| F −1 f (γ)|2 (1 + R( γ) eiγx ) dγ ≤ 0.




2 | F f (γ)| < 1 + R( γ) e dγ = 2<
0 2 0 2

Since |R(γ)| ≤ 1 by (4.2.2) it follows that < 1 + R( 12 γ) eiγx ≥ 0, so that the integral is

non-negative and hence it has to be zero and since the integrand
is non-negative it has to be
zero almost everywhere. Note that a zero of < 1 + R( 21 γ) eiγx can only occur if |R( 12 γ)| = 1,
its maximal value. The transmission coefficient is zero, T (γ) = 0, by (4.2.2) and hence
+ + +
f−γ (x) + R(γ) fγ+ (x) = 0, or f−γ and fγ+ are linearly dependent solutions, i.e. W (fγ+ , f−γ ) = 0.
By Proposition 4.2.1 this implies γ = 0. We conclude that F −1 f is zero almost everywhere,
hence f is zero almost everywhere, or f = 0 ∈ L2 (0, ∞).
So any real-valued f ∈ L2 (0, ∞) satisfying f + K (x) f = 0 equals f = 0 ∈ L2 (0, ∞). Since
the kernel K is real-valued, for any f ∈ L2 (0, ∞) with f +K (x) f = 0 we have <f +K (x) <f = 0
and =f + K (x) =f = 0, so then f = 0 ∈ L2 (0, ∞). This proves uniqueness of the solution.

Remark 4.4.3. The Gelfand-Levitan-Marchenko integral equation of Theorem 4.4.2 presents

us with an algorithm to reconstruct uniquely the potential q from the scattering data, and ac-
tually we can even do this just knowing the reflection coefficient and the bound states, i.e. the
eigenvalues plus the corresponding norms. This is actually the only part of information related
to the transmission coefficient that is needed, cf. Theorem 4.2.7, and Remark 4.2.10. Note
moreover that, given the scattering data, the Gelfand-Levitan-Marchenko integral equation is
a linear problem for B.

We do not address the characterisation problem, i.e. given a matrix S(γ) as in (4.2.3),
under which conditions on its matrix elements is S(γ) the scattering matrix corresponding to
a potential q? The characterisation problem depends on the class of potentials that is taken
into account. We refer to the paper [3] of Deift and Trubowitz for these results.

Remark 4.4.4. We briefly sketch another approach to recover the potential function from the
scattering data, which is known as the Fokas-Its Riemann-Hilbert approach. We assume that
the transmission coefficient T has no poles in the upper half plane, but the line of reasoning
can be easily adapted to this case. We write

T (γ)eiγx fγ− (x) − eiγx f−γ

+
(x) = R(γ)fγ+ (x)eiγx , γ ∈ R.
+
Note that T (γ)eiγx fγ− (x) is analytic in the open upper half plane =γ > 0 and that eiγx f−γ (x)
is analytic in the open lower half plane =γ < 0, and both have continuous extensions to the
real axis. Such a problem is a Riemann-Hilbert type problem, and this was solved by Plemelj
in 1908. Taking into account the behaviour of the functions as γ → ±∞, the Plemelj solution
to such a Riemann-Hilbert problem is given by

R(λ)fλ+ (x)eiλx
Z
iγx + 1
e f−γ (x) = 1 + dλ.
2πi R λ−γ
Chapter 4: Schrödinger operators and scattering data 77

Relabelling gives an integral equation for the Jost solution;

R(λ)fλ+ (x)ei(λ+γ)x
Z
+ iγx 1
fγ (x) = e + dλ
2πi R λ+γ

and this determines the Jost solution from the scattering data.
Now the Jost solution fγ+ determines the potential q, as can be seen as follows. Write
fγ+ (x) = eiγx+u(x;γ) with u(x; γ) → 0 as x → ∞, =γ ≥ 0 and u(x; γ) → 0 as γ → ±∞, =γ ≥ 0.
Then the Schrödinger equation for fγ+ translates into the differential equation −2iγux =
uxx + u2x − q, which is a first order equation for ux . Because of its behaviour as γ → ±∞ we
expand ux (x; γ) = ∞ vn (x)
P
n=1 (2iγ)n and this gives v1 (x) = q(x) and recursively defined vn ’s. So in
1
particular, ux (x; γ) = 2iγ q(x) as γ → ±∞. Moreover, fγ+ (x)e−iγx = eu(x;γ) ∼ 1 + u(x; γ), so
that we formally obtain

d + 
q(x) = lim 2iγ fγ (x)e−iγx − 1 .
γ→∞, =γ≥0 dx
So in this way, one can also recover the potential q from the scattering data.

4.5 Reflectionless potentials

We now consider the situation of a reflectionless potential, R(γ) = 0, γ ∈ R. This is a very
special case, and the Gelfand-Levitan-Marchenko equation essentially is a matrix equation.
Of course, we need to assume that the corresponding Schrödinger equation has at least one
eigenvalue, otherwise the kernel K is identically equal to zero, so that B is identically equal
d2
to zero and hence the potential is trivial, and we get the Schrödinger equation − dx 2.

We first consider the easiest non-trival case, N = 1. So that K(x) = Kd (x) = 2ρe−2px ,
ρ = ρ1 , p = p1 , and the Gelfand-Levitan-Marchenko equation becomes
Z ∞
−2p(x+y)
2ρe + B(x, z)2ρe−2p(x+y+z) dz + B(x, y) = 0, y ≥ 0.
0
Z ∞
−2p(x+y) −2p(x+y)
=⇒ B(x, y) = −2ρe − 2ρe B(x, z)e−2pz dz,
0

so that the y-dependence of B(x, y) is 2ρe−2py . Substitute B(x, y) = −2ρe−2py w(x), so that
Z ∞
−2xp −2px −ρ
w(x) = e +e w(x) −2ρe−4pz dz = e−2px + w(x)e−2px
0 2p
ρ −2px
so that w(x) = e−2px /(1 + 2p
e ) and we get

−2ρe−2p(x+y) −2ρe−2px
B(x, y) = ρ −2px =⇒ B(x, 0) = ρ −2px .
1 + 2p e 1 + 2p e
78 Chapter 4: Schrödinger operators and scattering data

Deriving this expression and multiplying by −1 gives the potential by Theorem 4.3.1;
4pρe−2px −2p2
q(x) = − ρ −2px 2 = , (4.5.1)
(1 + 2p e ) cosh2 (px + 21 ln(2p/ρ))

which is, up to a affine scaling of the variable, the potential considered in Section 2.5.2.
So the potential −2p2 cosh−2 (px) is reflectionless, and we can ask whether there are more
values for a such that the potential a cosh−2 (px) is reflectionless. By Exercise 2.5.6 it suffices
to consider the case p = 1. We can elaborate on the discussion in Section 2.5.2 by observing
 q q 
1 1 1 1

iγ − iγ + 4 − a, 2 − iγ − 4 − a 1 
fγ+ (x) = 2iγ cosh(x) 2 F1  2 ; ,
1 − iγ 1 + e2x
 q q 
1 1 1 1

iγ − iγ + 4 − a, 2 − iγ − 4 − a e2x
fγ− (x) = 2iγ cosh(x) 2 F1  2 ; ,
1 − iγ 1 + e2x
 q q 
1 1 1 1
+ − a, − − a 1 
+
f−γ (x) = e−ixγ 2 F1  2 4 2 4
; ,
1 + iγ 1 + e2x

so fγ+ corresponds to the solution 2 F1 (a, b; c, z) of the corresponding hypergeometric differential

+
equation (2.5.1), fγ− corresponds to 2 F1 (a, b; a + b + 1 − c; 1 − z), and f−γ corresponds to
q
z 1−c 2 F1 (a − c + 1, b − c + 1; 2 − c; z) with the values (a, b, c) given by ( 12 − iγ ± 14 − a, 1 − iγ)
as in Section 2.5.2. Note that 1 + a + b − c = c in this case. So the transmission and reflection
+
coefficient defined by T (γ)fγ− (x) = f−γ (x) + R(γ)fγ+ (x) follow from the relation (2.5.4) for
hypergeometric functions; T (γ) is 1/B and R(γ) equals A/B. Plugging in the values of A and
B from (2.5.4) with the values of (a, b, c) in terms of a and γ gives an explicit expression for
the transmission and reflection coefficient;
q q
Γ( 21 − iγ + 14 − a)Γ( 12 − iγ − 14 − a)
T (γ) = ,
Γ(1 − iγ)Γ(−iγ)
r r r
1
1 1 1 1 Γ(iγ)
R(γ) = cos π − a Γ( − iγ + − a)Γ( − iγ − − a) .
4 2 4 2 4 πΓ(−iγ)
q
So R(γ) = 0 can only occur if the cosine vanishes, so we need π 14 − a = 12 π + lπ, l ∈ Z,
and this implies a = −l(l + 1), and in this case

Γ(1 − iγ + l)Γ(−iγ − l) (1 − iγ)l (1 − iγ)l

T (γ) = = = (−1)l .
Γ(1 − iγ)Γ(−iγ) (−iγ − l)l (1 + iγ)l
So T is rational in γ, and we can read off the poles on the positive imaginary axis. There are
l simple poles at ik, k ∈ {1, · · · , l}.
Chapter 4: Schrödinger operators and scattering data 79

Proposition 4.5.1. The potential q(x) = −l(l + 1) cosh−2 (x) is a reflectionless potential.
The corresponding self-adjoint Schrödinger operator has essential spectrum [0, ∞) and discrete
spectrum −l2 , −(l − 1)2 , · · · , −1. The eigenfunction for the eigenvalue −k 2 is given by
 
+ 1 k − l, 1 + k + l 1
fk (x) = fik (x) = 2 F1 ;
(2 cosh(x))k 1+k 1 + e2x

with L2 (R)-norm given by kfk k2 = kl (k−1)!

1
.

Note that the 2 F1 -series in the eigenfunction is a terminating series, since k − l ∈ −N and
(−n)m = 0 for m > n.

Proof. The statements about the spectrum and the precise form of the eigenvalues and eigen-
functions follow from Theorem 4.2.7 and the previous considerations. It remains to calculate
the squared norm, for which we use Theorem 4.2.7. First we calculate

ekx e2x
 
kx − k − l, 1 + k + l
Ck = lim e fik (x) = lim x 2 F1 ;
x→∞ x→∞ (e + e−x )k 1+k 1 + e2x
 
k − l, 1 + k + l (−l)l−k
= 2 F1 ;1 = = (−1)l−k ,
1+k (1 + k)l−k

where we use 2 F1 (−n, b; c; 1) = (c − b)n /(b)n for n ∈ N, which is known as the Chu-Vander-
monde summation formula. In order to apply Theorem 4.2.7 we also need to calculate the
residue of the transmission coefficient at γ = ik;

(1 − iγ)l (1 − iγ)l
Resγ=ik T (γ) = Resγ=ik (−1)l = lim (γ − ik)(−1)l
(1 + iγ)l γ→ik (1 + iγ)l
l
 
−i (−1) (1 + k)l l−k 1 l
= k−1
= i (−1) ,
(−1) (k − 1)! (l − k)! (k − 1)! k

so that Theorem 4.2.7 gives the required result.

R
Exercise 4.5.2. Calculate R |fk (x)|2 dx directly using the substitution z = 1/(1 + e2x ), to
see that this integral equals

1 1
 
k − l, l + k + 1
Z
| 2 F1 ; z |2 z k−1 (1 − z)k−1 dz.
2 0 1+k
R1
Expand the terminating 2 F1 -series, and use the beta-integral 0 z α−1 (1−z)β−1 dz = Γ(α)Γ(β)
Γ(α+β)
, to
rewrite this as a double series. Rephrase the result of Proposition 4.5.1 as a double summation
formula.

Now that we have established the existence of reflectionless potentials, we can try to solve
the Gelfand-Levitan-Marchenko equation for a reflectionless potential with an arbitrary, but
80 Chapter 4: Schrödinger operators and scattering data

finite,
PN number of discrete eigenvalues (or bound states). So we have now K(x) = Kc (x) =
−2pn x
2 n=1 ρn e and we have to solve B(x, y) from
N
X Z ∞ N
X
−2pn (x+y)
2 ρn e + B(x, z) 2 ρn e−2pn (x+y+z) dz + B(x, y) = 0,
n=1 0 n=1

and this shows that we can expand B(x, y) as a linear combination of e−2pn y when considered
as function of y. We put —the form choosen makes the matrix involved of the form I plus a
positive definite matrix—
N
X √
B(x, y) = ρn e−pn (x+2y) wn (x),
n=1

and we need to determine wn (x) from the Gelfand-Levitan-Marchenko equation in this case.
Plugging this expression into the integral equation we see that we get an identity, when
considered as function in y, is a linear combination of e−2pn y . Since the pn ’s are different, each
√
coefficient of e−2pn y has to be zero. This gives, after dividing by ρn e−pn x ,
N ∞
√ X √ Z
2 ρn e−pn x + 2 ρm ρn wm (x)e−(pn +pm )x e−2(pn +pm )z dz + wn (x) = 0
m=1 0

√ √
for n = 1, · · · , N . Put w(x) = (w1 (x), · · · , wN (x))t and v(x) = 2( ρ1 e−p1 x , · · · , ρn e−pn x )t
we can rewrite this as (I + S(x))w(x) + v(x) = 0, where I is the identity N × N -matrix and
√
S(x)nm = ρn ρm e−(pn +pm )x /(pn + pm ), which is also a symmetric matrix. In order to see that
S(x) is also positive definite we take an arbitrary vector ξ ∈ CN ,
N N Z ∞
X X √ √
hS(x)ξ, ξi = S(x)ij ξj ξi = ρj e−pj x ξj e−pj t ρi e−pi x ξi e−pi t dt
i,j=1 i,j=1 0
Z ∞
= |f (x, t)|2 dt ≥ 0,
0
PN √
with f (x, t) = i=1 ρi ξi e−pi (x+t) . It follows that I + S(x) is invertible, as predicted by

−1
Theorem 4.4.2, so that we can solve for w(x) = − I + S(x) v(x). So
N
X √
−1 1
B(x, y) = wn (x) ρn e−pn (x+2y) = −h I + S(x) v(x), v(x + 2y)i
n=1
2
N
1
−1 1 X
−1
=⇒ B(x, 0) = − h I + S(x) v(x), v(x)i = − I + S(x) n,m vm (x)vn (x)
2 2 n,m=1
N
X
−1 d

−1 d

=2 I + S(x) n,m I + S(x) m,n = 2 tr I + S(x) I + S(x)
n,m=1
dx dx
Chapter 4: Schrödinger operators and scattering data 81

observing that

d √ 1
I + S(x) n,m = − ρn ρm e−(pn +pm )x = − vn (x)vm (x).


dx 4

In order to rewrite B(x, 0) in an even more compact way, note that for a N × N -matrix
A(x) depending on a variable x, we can calculate

d
det(A(x))
dx    
a011 (x) a12 (x) · · · a1N (x) a11 (x) a12 (x) · · · a01N (x)
 a0 (x) a22 (x) . . . a2N (x)   a21 (x) a22 (x) · · · a0 (x) 
 21 2N
= det  .. + · · · + det
  
... ..   .. ... .. 
 . .   . . 
0 0
aN 1 (x) aN 2 (x) · · · aN N (x) aN 1 (x) aN 2 (x) · · · aN N (x)
N
X M
X N
X
= a01j (x)A1j (x) + ··· + a0N j (x)AN j (x) = a0ij (x)Aij (x)
j=1 j=1 i,j=1
N
X dA
a0ij (x)A−1 (x)A−1 (x)


= det(A(x)) ji (x) = det(A(x)) tr
i,j=1
dx

by developing according to the columns that have been differentiated, denoting the signed
principal minors by Aij (x), so that A−1 (x)ij = det(A(x))−1 Aji (x).
This observation finally gives

d
d2

B(x, 0) = 2 ln det(I + S(x)) , q(x) = −2 2
ln det(I + S(x)) ,
dx dx

by Theorem 4.3.1. We formalise this result.

Proposition 4.5.3. If q is a reflectionless potential with scattering data R(γ) = 0, {(pn , ρn ) |

pn > 0, 1 ≤ n ≤ N }, then
d2

q(x) = −2 ln det(I + S(x)) ,
dx2
with S(x) the N × N -matrix given by

√
S(x)n,m = ρn ρm e−(pn +pm )x /(pn + pm ).

In Section 5.5 we consider the case N = 2 with an additional time t-dependence of Propo-
82 Chapter 4: Schrödinger operators and scattering data

sition 4.5.3. In case N = 2 we can make Proposition 4.5.3 somewhat more explicit;
√ !
ρ1 −2p1 x ρ1 ρ2 −(p1 +p2 )x
1 + 2p e p1 +p2
e
det(I + S(x)) = det √ρ1 ρ2 −(p 1
1 +p2 )x ρ2 −2p2 x
p1 +p2
e 1 + 2p 2
e
ρ1 −2p1 x ρ2 −2p2 x  1 1 
= 1+ e + e + − ρ1 ρ2 e−2(p1 +p2 )x
2p1 2p2 4p1 p2 (p1 + p2 )2
ρ1 −2p1 x ρ2 −2p2 x ρ1 ρ2 (p2 − p1 )2 −2(p1 +p2 )x
= 1+ e + e + e
2p1 2p2 4p1 p2 (p1 + p2 )2
√ √ √
ρ1 ρ2 −(p1 +p2 )x n p2 − p1  p1 + p2 2 p1 p2 (p1 +p2 )x p2 − p1 ρ1 ρ2 −(p1 +p2 )x 
= √ e √ e + √ e
2 p1 p2 p1 + p2 p2 − p1 ρ1 ρ2 p1 + p2 2 p1 p2
 √ρ p √
p1 ρ2 (p1 −p2 )x o
1 2 (p2 −p1 )x
+ √ e +√ e
p1 ρ2 ρ1 p2
√ √
p2 − p1 ρ1 ρ2 −(p1 +p2 )x n p1 + p2 2 p1 p2

= √ e cosh (p1 + p2 )x + ln( √ )
p1 + p2 p1 p2 p2 − p1 ρ1 ρ2
√
p1 + p2 ρ1 p 2
o
+ cosh (p2 − p1 )x + ln( √ ) .
p2 − p1 p1 ρ2

This expression is used in Section 5.5.

Chapter 5

Inverse scattering method and the

Korteweg-de Vries equation

5.1 The Korteweg-de Vries equation and some solutions

The Korteweg1 -de Vries2 , or KdV, equation is the following partial differential equation

qt (x, t) − 6 q(x, t) qx (x, t) + qxxx (x, t) = 0 or qt − 6qqx + qxxx = 0, (5.1.1)

where q is a function of two variables x, space, and t, time, with x, t ∈ R. Here qt (x, t) =
∂q
∂t
(x, t), etc. This equation can be viewed as a non-linear evolution equation by writing it as
qt = S(q) with S a non-linear map on a suitable function space defined by S(f ) = 6f fx −fxxx .
We also consider the Korteweg-de Vries equation together with an initial condition;

qt (x, t) − 6 q(x, t) qx (x, t) + qxxx (x, t) = 0, q(x, 0) = q0 (x), x ∈ R, t > 0. (5.1.2)

We first discuss some straightforward properties of the KdV-equation (5.1.1), also in rela-
tion to some other types of (partial) differential equations in the following exercises.
Exercise 5.1.1. Show that with q a solution to the Korteweg-de Vries equation (5.1.1) also
q̃(x, t) = q(x − 6Ct, t) − C is a solution. This is known as Galilean invariance.
Exercise 5.1.2. Show that by a change of variables the Korteweg-de Vries equation can be
transformed into qt + aqqx + bqx + cqxxx = 0 for arbitrary real constants a, b, c.
Exercise 5.1.3. The Burgers3 equation qt = qxx + 2qqx looks similar to the KdV-equation.
Show that the Burgers equation can be transformed into a linear equation ut = uxx by the
Hopf-Cole transformation q = (ln u)x .
1
Diederik Johannes Korteweg (31 March 1848 — 10 May 1941), professor of mathematics at the University
of Amsterdam and supervisor of de Vries’s 1894 thesis “Bijdrage tot de kennis der lange golven”.
2
Gustav de Vries (22 January 1866 — 16 December 1934), mainly worked as a high school teacher.
3
Johannes Martinus Burgers (13 January 1895 — 7 June 1981), is one of the founding fathers of research
in fluid dynamics in the Netherlands.

83
84 Chapter 5: Inverse scattering method and the KdV equation

Exercise 5.1.4. The modified KdV-equation, or mKdV-equation, is ut + uxxx − 6u2 ux = 0.

Show that for a solution u of the mKdV-equation, the relation q = ux + u2 gives a solution q
to the KdV-equation. This transformation is known as the Miura transformation. Hint: Show
that qt + qxxx − 6qqx can be written as a first order differential applied to ut + uxxx − 6u2 ux .

With writing down this initial value problem, the question arises on the existence and
uniqueness of solutions. This is an intricate questions and the answer depends, of course, on
smoothness and decay properties of q0 . We do not go into this, but formulate two results
which we will not prove. Our main concern is to establish a method in order to construct
solutions explicitly.
dk q0
Theorem 5.1.5. Assume q0 ∈ C 4 (R) and (x) = O(|x|−M ), M > 10, for k ∈ {0, 1, 2, 3, 4}.
dxk
Then the Korteweg-de Vries initial value problem (5.1.2) has a real-valued unique solution (in
the classical sense).

Exercise 5.1.6. In this exercise we sketch a proof of the unicity statement. Assume q and u
are solutions to (5.1.2), and put w = u − q.

• Show that wt = 6uwx + 6qx w − wxxx using the KdV-equation.

• Multiply by w and integrate to conclude

Z Z Z
1d
w (x, t) dx = 6 u(x, t)w(x, t)wx (x, t) dx + 6 qx (x, t)w2 (x, t) dx
2
2 dt R R R
Z
− w(x, t)wxxx (x, t) dx
R

and argue that the last integral vanishes.

• Integrate by parts in the first integral to obtain

Z Z
d 1
w (x, t) dx = 12 (qx (x, t) − ux (x, t)) w2 (x, t)dx.
2
dt R R 2

• Use |qx (x,R t) − 21 ux (x, t)| ≤ C Rto conclude dtd R w2 (x, t) dx ≤ 12C R w2 (x, t) dx and
R R

conclude R w2 (x, t) dx ≤ e12C t R w2 (x, 0) dx = 0. Hence w(x, t) = 0.

Theorem 5.1.5 is proved using the Inverse Spectral Method, a method that we discuss later,
and it should be noted that even a C ∞ (R)-condition of the initial value q0 is not sufficient to
guarantee a continuous global solution to (5.1.2). Methods from evolution equations have led
to the following existence result.

Theorem 5.1.7. Assume the initial value q0 is an element of the Sobolev space W s (R) for
s ≥ 2. Then the Korteweg-de Vries initial value problem (5.1.2) has a solution q such that
t 7→ q(·, t) is a continuous map from R≥0 into W s (R) and a C 1 -map from R≥0 into L2 (R).
Chapter 5: Inverse scattering method and the KdV equation 85

The last statement means that limh→0 k h1 q(·, t + h) − q(·, t) − qt (·, t)k = 0 in L2 (R), and

moreover that t 7→ qt (·, t) is a continuous map from R≥0 to L2 (R).

Solutions to the KdV-equation R (5.1.1) have many nice properties, such as a large number
of invariants. Let us show that R q(x, t) dx is an invariant, i.e. independent of time t. Write

0 = qt − 6qqx + qxxx = qt + −3q 2 + qxx x .

Assuming that we can differentiate under the integral sign and that q and qxx decay sufficiently
fast we get
Z Z Z
d
q(x, t) dx = qt (x, t) dx = (3q 2 (x, t) − qxx (x, t))x dx = 0.
dt R R R

In physical applications this is considered as conservation of mass. More generally, whenever

we have a relation of the form Tt + Xx = 0, where T and X are depending R on the solution
q of the KdV-equation and its derivatives with respect to x, we have R T dx as a (possible)
invariant. In the above example we have T = q and X = −3q 2 + qxx .
R
Exercise 5.1.8. Show also that under suitable conditions R q 2 (x, t) dx is an invariant by
considering q(qt − 6qqx + qxxx ) = 0 and writing this in the form (q 2 )t equals a derviative with
respect to x. This is considered as conservation of momentum. What are T and X?
In fact there is an infinite number of conserved quantities for the KdV-equation, and this
point of view is exploited in integrable systems.
We now give a class of solutions to (5.1.1) of the form q(x, t) = f (x − ct) for some c ∈ R.
This method is already worked out in the 1895-paper of Korteweg and de Vries. So the KdV
equation gives an ordinary partial differential equation for f ;

−cf 0 − 6f f 0 + f 000 = 0 =⇒ −cf − 3f 2 + f 00 = a,

where a is some integration constant. Multiplying by f 0 shows that we can integrate again,
and
c 1
−cf f 0 − 3f 2 f 0 + f 00 f 0 = af 0 =⇒ − f 2 − f 3 + (f 0 )2 = af + b
2 2
for some integration constant b. This is a non-linear first order (ordinary) differential equation

(f 0 )2 = F (f ), F (x) = 2x3 + cx2 + ax + b,

after rescaling the integrating constants. Since the left hand side is a square, the regions where
the cubic polynomial F is positive are of importance. So the zeros of the F play a role in the
analysis.
Remark 5.1.9. In general the equation y 2 = F (x), with F a cubic polynomial, is an elliptic
curve, so that we can expect elliptic functions involved in the solution. The elliptic curve
is non-singular (i.e. no cusps or self-intersections) if and only if the polynomial F has three
distinct roots.
86 Chapter 5: Inverse scattering method and the KdV equation

The analysis now has to be split up according to the following cases;

1. F has three distinct real roots;
2. F has two real roots, of which one necessarily is of order two, and we have two cases:
the double root is smaller or larger than the simple root;
3. F has one simple root;
4. F has one root of order three.

1 y

0
-1.5 -1 -0.5 0 0.5 1
x

-1

-2

Figure 5.1: The cubic polynomial F in case 2 with double zero larger than the simple zero.

Case 1 can be completely solved in terms of elliptic functions. For our purposes case 2
is the most interesting, since the corresponding solution is related to the so-called soliton
solution of the KdV equation (5.1.1) The case 2 is of interest in case the simple root, say α, is
smaller than the double root, say β, so F (x) = 2(x − α)(x − β)2 as in the situation of Figure
5.1 with α = −1 and β = 0.
If one considers the differential equation (f 0 )2 = F (f ) in this case then we find that a
solution f with initial value α < f (x0 ) < β satisfies α < f (x) < β for all x. So we can expect
bounded solutions in this case.
Chapter 5: Inverse scattering method and the KdV equation 87

Taking square roots and noting that the differential equation is separable gives
Z f
dξ √
√ = ± 2x + C,
α (ξ − β) ξ − α

where C denotes a general arbitrary constant. In the integral we substitute ξ = α + (β −

2
q sin ν, so dξ = (β − α)2 sin ν cosR ν dν and the integration runs over [0, φ] with sin φ =
α)
f −α φ 1
β−α
. So the integral equals √−2
β−α 0 cos ν
d
dν, and now using that dν ln | tan( ν2 )| = sin1 ν by an
elementary computation we find

−2 φ π √
√ ln | tan( + )| = ± 2x + C,
β−α 2 4

and using the addition formula for tan or for sin and cos we can rewrite this as
 1 + tan( φ )  r
β−α
2
ln =± (x − C).
 
φ  2
1 − tan( 2 )

1+tan( φ )
Denoting the right hand side by y, we get 2
1−tan( φ )
= ey which gives tan( φ2 ) = tanh( y2 ). Now
2
we can determine f as a function of x

φ φ
f (x) = α + (β − α) sin2 φ = α + 4(β − α) sin2 ( ) cos2 ( )
2 2
φ φ
= α + 4(β − α) tan2 ( ) cos4 ( )
2 2
tan2 ( φ2 )
= α + 4(β − α)
2
1 + tan2 ( φ2 )
tanh2 ( y2 )
= α + 4(β − α)
2
1 + tanh2 ( y2 )
α−β
= α + (β − α) tanh2 (y) = β +
cosh2 (y)
α−β
= β+ q
cosh2 ( β−α
2
(x − C))

using cos2 φ = 1/(1 + tan2 φ) and 12 1 + tanh2 ( y2 ) tanh y = tanh( y2 ) and using that cosh

is an even function. In this solution the zeros of F are related to c by c = −2α − 4β by

comparing the coefficients of x2 in the cubic polynomial F . We see that lim|x|→∞ f (x) = 0
can only be achieved by choosing β = 0, so that α < 0 and c = −α > 0, then f is essentially a
cosh−2 -function, see Figure 2.1 for the case c = 4. So we have proved the following proposition.
88 Chapter 5: Inverse scattering method and the KdV equation

Proposition 5.1.10. For c > 0, the function

−c
q(x, t) = √
2 cosh2 ( 12 c(x − ct + C))

is a solution to the KdV-equation (5.1.1) for any real constant C.

Note that q(x, t) is negative and its minimum occurs for x = ct, so we can view this as
a solitary wave (directed downwards) of height 12 c travelling to the right at a speed c. Note
that the height and speed are related!
Exercise 5.1.11. Assume now that F has one root of order three. Show that this root is
− 16 c, and that the corresponding solution is f (x) = − 61 c + (x−C)
2
2.

Exercise 5.1.12. Give a qualitative analysis of possible growth/decay behaviour of the solu-
tions to (f 0 )2 = F (f ) for the other listed cases for the cubic polynomial F .
Exercise 5.1.13. Assume now that F has three distinct real roots,
R φ then one can proceed
in the same way, except that now we get the integral of the form 0 √ 1
dν instead of
1−k2 sin2 ν
Rφ 1
0 cos ν
dν. Use the Jacobian elliptic function cn defined by
Z φ
1
v = p dν cos φ = cn(v; k).
0 1 − k 2 sin2 ν
The corresponding solutions to the KdV-equation (5.1.1) are known as cnoidal waves, and
were already obtained by Korteweg and de Vries in their 1895 paper.
In the course of history the KdV-equation (5.1.1) has been linked to other well-known
ordinary differential equations by other suitable substitutions, and two of them are discussed
in the following exercises.
Exercise 5.1.14. Put q(x, t) = t + f (x + 3t2 ), and show that q satisfies the KdV-equation
(5.1.1) if and only if f satisfies 1 − 6f f 0 + f 000 = 0, or by integrating C + z − 3f 2 + f 00 = 0.
This equation is essentially the Painlevé I equation g 00 (x) = 6g 2 (x) + x, which is the sim-
plest equation in the Painlevé4 equations consisting of 6 equations. The Painlevé equations
are essentially the second order differential equations of the form y 00 (t) = R(t, y, y 0 ) with R
a polynomial in y, y 0 with meromorphic coefficients in t together with the Painlevé prop-
erty meaning that there are no movable branch points and no movable essential singularities
exluding linear and integrable differential equations.
Exercise 5.1.15. Put q(x, t) = −(3t)−2/3 f (3t)x1/3 , and show that q is a solution to the KdV-

equation (5.1.1) if and only if f satisfies f 000 + (6f − z)f 0 − 2f = 0. This is related to the
Painlevé II equation.
4
Paul Painlevé (5 December 1863 — 29 October 1933), French mathematician and politician. Painlevé was
Minister in several French Cabinets, as well as Prime-Minister of France.
Chapter 5: Inverse scattering method and the KdV equation 89

5.2 The KdV equation related to the Schrödinger op-

erator
This section is of a motivational and historical nature, and no proofs are given, only heuristical
derivations.
The link between the Korteweg-de Vries equation and the Schrödinger operator has been
pointed out in a three-page paper by Gardner, Greene, Kruskal and Miura of the Plasma
Physics Laboratory of Princeton University in the Physical Review Letters in 1967. This
paper5 has turned out to be a starting point for a whole line of research for solving certain
non-linear partial differential equations. We give a short description of their line of reasoning,
which apparently was strongly motivated by numerical experiments. The approach has been
generalised enormously, and we consider the approach by Lax in Section 5.3 in more detail.
Consider the Schrödinger equation with time-dependent potential,

−fxx (x, t) + q(x, t)f (x, t) = λ(t)f (x, t),

where the eigenfunctions and eigenvalues will also depend on t. Expressing q in terms of
f and λ and plugging this into the KdV-equation (5.1.1) gives a huge equation, that upon
multiplying by f 2 , can be written as

λt f 2 + f ux − fx u x = 0,



u(x, t) = ft (x, t) + fxxx (x, t) − 3 q(x, t) + λ(t) fx (x, t). (5.2.1)

This is a tedious verification and one of the main steps, assuming that all partial derivatives
exist and ftxx = fxxt . Note that the second term in (5.2.1) is a derivative of a Wronskian.
Later it has been pointed out that the KdV-equation can be interpreted as a compatibility
condition ftxx = fxxt for solutions of systems of partial differential equations. This approach
has also led to several other generalisations.
Integrating relation (5.2.1) over R, and assuming that f ∈ L2 (R), so that λ corresponds
to the discrete spectrum of the Schrödinger operator, we get
a
λt kf k2 = − lim f (x, t) ux (x, t) − fx (x, t)ux (x, t)

a→∞ −a

and the right hand side gives zero, so that the (discrete) spectrum is constant. So the im-
portant observation is that the KdV-equation is a description of an isospectral family of
Schrödinger operators. Hence, we can proceed as follows to find a solution to an initial value
problem (5.1.2) for the KdV-equation:
1. Describe the spectral data of the Schrödinger operator with potential q0 ; this is described
in Section 4.2.
2. Solve the evolution of the spectral data when the potential evolves according to the
KdV-equation; this is yet unclear.
5
Gardner, Greene, Kruskal and Miura have received the 2006 AMS Leroy P. Steele Prize for a Seminal
Contribution to Research for one of their follow-up papers.
90 Chapter 5: Inverse scattering method and the KdV equation

3. Determine the potential q(x, t) from the spectral data of the Schrödinger operator at time
t; this can be done essentially using the Gelfand-Levitan-Marchenko integral equation
of Theorem 4.4.2 and Theorem 4.3.1.
We shortly derive heuristically that the evolution in step 2 is linear, and since the Gelfand-
Levitan-Marchenko integral equation is also linear, this procedure gives a mainly linear method
to solve the KdV-inital value problem (5.1.2). Note however that in explicit cases it is hard
to solve the Gelfand-Levitan-Marchenko equation explicitly, except for the reflectionless po-
tentials as discussed in Section 4.5. This method is now known as the Inverse Scattering
Method (ISM), or inverse spectral method, or inverse scattering transformation. The inverse
scattering method can be considered as an analogue of the Fourier transform used to solve
linear differential equations.
Now that we have established λt = 0 for an eigenvalue λ in the discrete spectrum, we use
this in (5.2.1) and carrying out the differentiation gives


0 = (f ux − fx u)x = f uxx − fxx u = f uxx − (q − λ)u ,
or u is also a solution to the Schrödinger equation. Hence, u = C f + D g, with C, D ∈ C and
g a linearly independent solution, i.e.
ft + fxxx − 3(q + λ) fx = C f + D g. (5.2.2)
Assume λ = λn = −p2n a discrete eigenvalue using the notation as in Theorem 4.2.7, so that
f (x, t) = fn (x, t) ∼ e−pn x for x → ∞. So in the region where the potential vanishes, in
particular for x → ∞, we see that left hand side has exponential decay, and since g doesn’t
have exponential decay, it follows that D = 0.
Now fixing n, normalise ψ(x, t) = c(t)fn (x, t), so that kψ(·, t)k = 1 for all t, and since in
the above considerations f was an arbitrary solution to the Schrödinger equation we get
ψt + ψxxx − 3(q + λ)ψx = Cψ.
We claim that C = 0, and in order to see this we multiply this equation R by ψ and integrate
over R with respect to x. Then the left
hand side equals C. Now recall R ψt (x, t)ψ(x, t) dx =
1 1 d
R2 hψt (·, t), ψ(·, t)i + hψ(·, t), ψ1 t (·, t)i 2 = 2 dt
kψ(·, t)k2 = 0, since its norm is constant 1. Also
λ ψ (x, t)ψ(x, t) dx = λn 2 ψ(x, t) |∞
R n x −∞ = 0, since the eigenfunction decays exponentially

2
fast. Similarly, R ψxxx (x, t)ψ(x, t) dx = − R ψxx (x, t)ψx (x, t) dxR = − 12 ψx (x, t) |∞
R R
−∞ = 0,
since
R the derivative of the R eigenfunction decays exponentially. R q(x, t)ψx (x, t)ψ(x, t) dx =
λ ψ (x, t)ψ(x, t) dx + R ψx (x, t)ψxx (x, t) dx = 0, as before. So for the normalised eigen-
R n x
function ψ(·, t) for the constant eigenvalue λn we find
ψt + ψxxx − 3(q + λ)ψx = 0.
Now use ψ(x, t) ∼ c(t)e−pn x as x → ∞, and since we assume that the potential decays
sufficiently fast we get a linear first order differential equation for c(t);
dc 3
(t) + (−pn )3 c(t) − 3(−p2n )(−pn )c(t) = 0 =⇒ c(t) = e4pn t c(0).
dt
Chapter 5: Inverse scattering method and the KdV equation 91

Now that we have determined the evolution of the discrete part of the spectral data,
we consider next the evolution of the continuous part of the spectrum. Let λ = γ 2 , and
consider (5.2.2) for ψ(x, t) = ψγ (x, t) with ψγ as in Section 4.2, where we now assume that
the transmission and reflection coefficient also depend on t; R(γ, t), T (γ, t). Again using that
the potential q vanishes for large enough x, we see that
ψt (x, t) + ψxxx (x, t) − 3(q(x, t) + λ)ψx (x, t) ∼ Tt (γ, t) + 4iγ 3 T (γ, t) e−iγx ,


x → −∞,
3 −iγx 3 iγx


ψt (x, t) + ψxxx (x, t) − 3(q(x, t) + λ)ψx (x, t) ∼ 4iγ e + Rt (γ, t) − 4iγ e , x → ∞.
We conclude that ψt + ψxxx − 3(q + λ)ψx is a constant multiple of ψ, and this constant is 4iγ 3
as follows by looking at the coefficient of e−iγx for x → ∞. Equating gives two linear first
order differential equations for the transmission and reflection coefficients;
Tt (γ, t) + 4iγ 3 T (γ, t) = 4iγ 3 T (γ, t) =⇒ T (γ, t) = T (γ, 0),
3
Rt (γ, t) − 4iγ 3 R(γ, t) = 4iγ 3 R(γ, t) =⇒ R(γ, t) = R(γ, 0)e8iγ t .
Note that this also implies that starting with a reflectionless potential q0 , we stay within the
class of reflectionless potentials.
Exercise 5.2.1. Assume that the √ solution scheme for the Korteweg-de Vries equation with
initial value q0 (x) = − 2c cosh−2 ( 12 cx) of this section is valid. Use the results of Section 4.5
to derive the solution q(x, t) of Proposition 5.1.10. in this way. This solution is known as the
pure 1-soliton solution, see Section 5.5.

5.3 Lax pairs

The method of Lax6 pairs is a more general method to find isospectral time evolutions. In
the general setup, we have a family of operators L(t), t ∈ I with I ⊂ R some interval, acting
on a Hilbert space H. We assume that the domains do not vary with t, so D(L(t)) = D(L)
for all t and that D(L) is dense. We say that the family is strongly continuous if for each
x ∈ D(L) = D(L(t)) the map t 7→ L(t)x, I → H, is continuous. The family L(t) is isospectral
if the spectrum of L(t) is independent of t.
The basic example is the Schrödinger operator with time-dependent potential
d2
L(t) = − + q(·, t), D(L(t)) = W 2 (R),
dx2
assuming the potential q(·, t) satisfies the conditions of Corollary 2.2.7 for all t.
Definition 5.3.1. The derivative dL dt
(t) = Lt (t), t ∈ I, of a strongly continuous family L(t) is
defined with respect to the strong operator topology, i.e. D(Lt (t)) consists of those x ∈ D(L) ⊂
H such that
L(t + h)x − L(t)x
lim
h→0 h
converges in H to, say, y ∈ H, and then Lt (t)x = y.
6
Peter D. Lax (1 May 1926 — ), Hungarian-American mathematician, winner of the 2005 Abelprize.
92 Chapter 5: Inverse scattering method and the KdV equation

Definition 5.3.1 should be compared with Stone’s Theorem 6.4.2.

Assuming that D(Lt (t)) is independent of t and dense in H, then again Lt (t) gives a
family of operators on H. If this family is again strongly continuous, then we say that L(t) is
a strongly C 1 -family of operators.
Return to the example of a family of Schrödinger operators, then for any f ∈ W 2 (R), we
have as elements in L2 (R)

1  00  q(·, t + h) − q(·, t)
−f + q(·, t + h)f + f 00 − q(·, t)f = f .
h h
Assuming that the partial derivative with respect to t of q exists we obtain

 q(x, t + h) − q(x, t)
Z 
1 2
k L(t + h)f − L(t)f − qt (·, t)f k2 = − qt (x, t) |f (x)|2 dx.


h h

R

If we assume moreover that qt (·, t) ∈ L∞ (R) with locally bounded L∞ -norm we can apply
Dominated Convergence Theorem 6.1.3 to see that the right hand tends to zero as h → 0 even
for arbitrary f ∈ L2 (R). So in this case we have established Lt (t) as multiplication operator
by qt (·, t) with domain D(L) = W 2 (R).
Let us assume that we also have a strongly continuous family of bounded operators
V (t) ∈ B(H). As an application we derive a product rule, which is the same as for functions
except that we have to keep track of domains. Note that D(V (t)L(t)) = D(L(t)) = D(L) is
independent of t, and write for x ∈ D(Lt (t))

1 
V (t + h)L(t + h)x − V (t)L(t)x
h
1   1 
= V (t + h)L(t + h)x − V (t + h)L(t)x + V (t + h)L(t)x − V (t)L(t)x
h h
1   1  
= V (t + h) L(t + h)x − L(t)x + V (t + h)L(t)x − V (t)L(t)x
h h

and the first term tends to V (t)Lt (t)x as follows from

V (t + h) 1 L(t + h)x − L(t)x − V (t)Lt (t)x

 
h
1 
≤ V (t + h) L(t + h)x − L(t)x − V (t + h)Lt (t)x + V (t + h)Lt (t)x − V (t)Lt (t)x
h
1 
≤ kV (t + h)k L(t + h)x − L(t)x − Lt (t)x + V (t + h)Lt (t)x − V (t)Lt (t)x
h
and the first term tends to 0 as h → 0 by definition of the derived operator assuming that
kV (t)k is locally bounded and the second tends to zero since V is strongly continuous. More-
over, the second term tends to Vt (t)L(t)x if we assume that L(t)x ∈ D(Vt (t)). So we have
proved Lemma 5.3.2.
Chapter 5: Inverse scattering method and the KdV equation 93

Lemma 5.3.2. (i) Let L(t) be a strongly continuous family of operators on a Hilbert space H
with constant domain D(L) = D(L(t)) and let V (t) be a strongly continuous family of locally
bounded operators on H which is locally
bounded in the operator norm. If x ∈ D(Lt (t)) and
L(t)x ∈ D(Vt (t)), then x ∈ D (V L)t (t) and
d(V L)
(t)x = Vt (t)L(t)x + V (t)Lt (t)x.
dt
∗
∗
(ii) Let V (t) be as in (i), then dVdt (t) = Vt (t) .
Exercise 5.3.3. Prove the remaining statement of Lemma 5.3.2. What happens if we try to
differentiate L(t)V (t) with respect to t?
For any two (possibly) unbounded operators (A, D(A)), and (B, D(B)) the commutator
[A, B] is defined as the operator [A, B] = AB − BA with D([A, B]) = D(AB) ∩ D(BA) =
{x ∈ D(A) ∩ D(B) | Bx ∈ D(A) and Ax ∈ D(B)}. A (possibly unbounded) operator B is
anti-selfadjoint if iB is a self-adjoint operator. So B ∗ = −B, and in particular D(B) = D(B ∗ ).
Theorem 5.3.4 (Lax). Let L(t), t ≥ 0, be a family of self-adjoint operators on a Hilbert
space with D(L(t)) = D(L) independent of t which is strongly continuously differentiable.
Assume that there exists a family of anti-selfadjoint operators B(t), with constant domain
D(B(t)) = D(B), depending continuously on t such that
• Lt = [B, L], i.e. Lt (t) = [B(t), L(t)] for all t ≥ 0,
• there exists a strongly continuous family V (t) ∈ B(H) satisfying Vt (t) = B(t)V (t),
V (0) = 1 ∈ B(H),
• L(t)V (t) is differentiable with respect to t.
Then L(t) is unitarily equivalent to L(0) and in particular, the spectrum of L(t) is independent
of t.
Such pairs L, B are called Lax pairs, and this idea has been fruitful in e.g. integrable
systems, see Exercise 5.3.9 for an easy example.
Note that in case B = B(t) is independent of t, then we can take V (t) = exp(tB). In
particular, this is a family of unitary operators, since B is anti-selfadjoint and so the second
requirement is automatically fulfilled. This remains true in the general time dependent case
under suitable conditions on B(t), but there is not such an easy description of the solution.
This is outside the scope of these notes, but in the case of the KdV-equation as in the sequel
Vt (t) = B(t)V (t), V (0) = 1 has a unitary solution in case q(·, 0) ∈ W 3 (R).
Sketch of proof. We first claim that V (t) is actually unitary for each t. To see this, take
w(t) = V (t)w, v(t) = V (t)v and so dw
dt
(t) = B(t)w(t) and dv
dt
(t) = B(t)v(t), so that
d d d
hw(t), v(t)i = h w(t), v(t)i + hw(t), v(t)i = hB(t)w(t), v(t)i + hw(t), B(t)v(t)i
dt dt dt
= hB(t)w(t), v(t)i − hB(t)w(t), v(t)i = 0.
94 Chapter 5: Inverse scattering method and the KdV equation

Or hV (t)w, V (t)vi = hw, vi, and V (t) is an isometry, and since we also get from this hw, vi =
hV (t)∗ V (t)w, vi for arbitrary v it follows V (t)∗ V (t) = 1. Note that W (t) = V (t)V (t)∗ satisfies
Wt = BV V ∗ + V V ∗ B ∗ = BW − W B = [B, W ] with initial condition W (0) = 1. Since
W (t) = 1 is a solution, we find W (t) = V (t)V (t)∗ = 1, or V (t) is unitary, and the claim
follows.
By Lemma 5.3.2 we see that V (t)∗ is also differentiable, and so is L̃(t) = V (t)∗ L(t)V (t).
We want to show that L̃(t) is independent of t. Assuming this for the moment, it follows
that V (t)∗ L(t)V (t) = L(0), since V (0) = 1, and hence L(t) is unitarily equivalent to L(0) by
L(t) = V (t)L(0)V (t)∗ .
To prove the claim we differentiate with respect to t the relation L(t)V (t) = V (t)L̃(t);

dL dV dV dL̃
(t) V (t) + L(t) (t) = (t) L̃(t) + V (t) (t)
dt dt dt dt
dL dL̃
=⇒ (t) V (t) + L(t) B(t) V (t) = B(t) V (t) L̃(t) + V (t) (t)
dt dt
dL dL̃
=⇒ (t) V (t) + L(t) B(t) V (t) = B(t) V (t) V (t)∗ L(t) V (t) + V (t) (t)
dt dt
dL̃  dL 
=⇒ V (t) (t) = (t) + L(t) B(t) − B(t) L(t) V (t) = 0
dt dt
using the product rule of Lemma 5.3.2, the differential equation for the unitary family V (t).
Since V (t) is unitary we get ddtL̃ (t) = 0.

Exercise 5.3.5. Show formally that a self-adjoint operator of the form L = L0 + Mq , where
L0 is a fixed self-adjoint operator and Mq is multiplication by a t-dependent function q, and
assuming there exists a anti-selfadjoint B such that BL − LB = MK(q) , then L is isospectral
if q satisfies the equation qt = K(q).

Lax’s Theorem 5.3.4 dates from 1968, shortly after the the discovery of Gardner, Greene,
Kruskal and Miura of the isospectral relation of the Schrödinger operator and the KdV-
equation. In order to see how the KdV-equation arises in the context of Lax pairs, we take
L(t) as before as the Schrödinger operator with time-dependent potential q(·, t). We look for
B(t) in the form of a differential operator, and since it has to be anti-self-adjoint we only
allow for odd-order derivatives, i.e. we try for integer m and yet undetermined functions bj ,
j = 0, · · · , m − 1,
m−1
d2m+1 X d2j+1 d2j+1 

Bm (t) = 2m+1 + Mj (t) 2j+1 + 2j+1 Mj (t) , Mj (t)f (x) = bj (x, t)f (x),
dx j=0
dx dx

considered as operator on L2 (R) with domain W 2m+1 (R) for suitable functions bj .
d
In case m = 0 we have B0 (t) is dx independent of t, and [B, L] = qx , i.e. the multiplication
2
operator on L (R) by multiplying by qx . So then the condition Lt = [B, L] is related to the
partial differential equation qt = qx .
Chapter 5: Inverse scattering method and the KdV equation 95

Exercise 5.3.6. Solve qt = qx , and derive directly the isospectral property of the correspond-
ing Schrödinger operator L(t).

In case m = 1 we set B1 = ∂ 3 + b∂ + ∂b, where ∂ denotes derivative with respect to x and

where b = b(x, t), B1 = B1 (t), then

[B1 (t), L(t)] = ∂ 3 + b∂ + ∂b −∂ 2 + q − −∂ 2 + q ∂ 3 + b∂ + ∂b

= [∂ 3 , q] − [∂ 3 , b] + [∂, ∂b∂] + [b∂, q] + [∂b, q]

and using the general commutation [∂, f ] = fx repeatedly we see [∂ 3 , q] = 3∂qx ∂ + qxxx ,
[∂, ∂b∂] = ∂bx ∂, [b∂, q] = bqx , [∂b, q] = bqx , so that we obtain

[B1 (t), L(t)] = ∂(3qx − 2bx )∂ + qxxx − bxxx + 2bqx ,

3
and in order to make this a multiplication operator we need 3qx − 2bx = 0, or bx = q .
2 x
Choosing b = 32 q we see that we require the relation

1
− qxxx + 3qqx = [B1 (t), L(t)] = Lt (t) = qt ,
2
which is the KdV-equation up to a change of variables, see Exercise 5.1.2.
3
d d
Exercise 5.3.7. Check that taking B(t) = −4 dx 3 + 6q(x, t) dx + 3qx (x, t) gives [B(t), L(t)] =

−qxxx (·, t) + 6qx (·, t)q(·, t) corresponding to the KdV-equation (5.1.1).

Exercise 5.3.8. Consider the case m = 2 and derive the corresponding 5-th order partial
differential equation for q. Proceeding in this way, one obtains a family of Korteweg-de Vries
equations.

Exercise 5.3.9. Assume that we have a Lax pair as in Theorem 5.3.4 in the case of a
finite dimensional Hilbert space, e.g. H = C2 . Then we can view Lt = [B, L]
as a system
n
of N = dim H first order coupled differential equations. Show that tr L(t) then gives
invariants for this system. Show also that (Lk )t = [B, Lk ] for k ∈ N. Work this out for the
2
example of the harmonic oscillator; q̈ + ω
 q = 0, where dot denotes derivative with
 respect to
0 − 12 ω
 
p ωq
time t, by checking that L = , B is the constant matrix 1 and q̇ = p
ωq −p 2
ω 0
gives a Lax pair.

The idea of Lax pairs has turned out to be very fruitful in many other occassions; similar
approaches can be used for other nonlinear partial differential equations, such as the sine-
Gordon equation, the nonlinear Schrödinger equations, see Exercise 5.3.5. These equations
have many properties in common, e.g. an infinite number of conserved quantities, so-called
Bäcklund transformations to combine solutions into new solutions –despite the nonlinearity.
We refer to [2], [6] for more information.
96 Chapter 5: Inverse scattering method and the KdV equation

5.4 Time evolution of the spectral data

Assume that we have a Lax pair as in Section 5.3, and assume that L(0) has µ as a discrete
eigenvalue with eigenvector w ∈ H. By Theorem 5.3.4, it follows that µ is an eigenvalue of
L(t), and since L(t) = V (t)L(0)V (t)∗ it follows that L(t)V (t)w = V (t)L(0)w = µ V (t)w, or
w(t) = V (t)w is the eigenvector of L(t) for the eigenvalue µ. So we have an explicit evolution
of the eigenvector for the constant eigenvalue µ. This eigenfunction also satisfies a first order
differential equation, since

d
w(t) = Vt (t)w = B(t)V (t)w = B(t)w(t).
dt
Exercise 5.4.1. Check directly for the case m = 0, cf. Exercise 5.3.6, that the eigenfunction
for a discrete eigenvalue µ satisfies the appropriate differential equation at arbitrary time t.

Assume now that q(·, t) satisfies the conditions of Theorem 4.2.7 for all t, and assume that
the Schrödinger operator at time t = 0 has bound states. Then it follows that itphas the same
eigenvalues at all times t. It follows that ψn (x, 0) = Nn (0)fn (x, 0), Nn (0) = ρn (0), is an
eigenfunction of the Schrödinger operator at t = 0 of length 1. Hence V (t)ψn (·, 0) = ψn (·, t) is
an eigenfunction of the Schrödinger operator at t of length 1, since V (t) is unitary. Moreover,
ψn (·, t) is a multiple of fn (x, t), the solution
p of the Schrödinger integral equation at time t.
So ψn (x, t) = Nn (t)fn (x, t), Nn (t) = ρn (t) with the notation as in Theorem 4.2.7 with
time dependency explicitly denoted up to a phase factor. Let us also assume that q(·, t) is
compactly supported for all (fixed) t, then we know by Theorem 4.1.2, with notation as in
Theorem 4.2.7 that fn (x, t) = e−pn x for x sufficiently large, i.e. outside the support of q(·, t).
d3
For such x we have B(t) = −4 dx 3 with the version as in Exercise 5.3.7, so that the phase

factor has to be 1 and

d d3
ψn (x, t) = B(t)ψn (x, t) = −4 3 N (t)e−pn x = 4p3n N (t)e−pn x ,
dt dx
d d dNn
ψn (x, t) = Nn (t)e−pn x = e−pn x (t)
dt dt dt

for sufficiently large x. So dN dt

n
(t) = 4p3n N (t), and Nn (t) = exp(4p3n t)Nn (0) or ρn (t) =
exp(8p3n t)ρn (0). So this gives the time-development for the discrete part of the scattering
data in terms of a simple linear first order differential equation.
For the time-dependency of the scattering and reflection coefficient we proceed in an anal-
ogous way. Take γ ∈ R and assume that q(·, t) satisfies the conditions of Theorem 4.2.7 for
+
all t. Consider the solution ψγ (x) = T (γ) fγ− (x) = f−γ (x) + R(γ) fγ+ (x) for the eigenvalue
equation for Schrödinger solution at time zero for the eigenvalue γ 2 , see Section 4.2. Then we
know by Lax’s Theorem 5.3.4 that Ψγ (x, t) = V (t) ψγ (x) is a solution of the corresponding
eigenvalue equation for the Schrödinger equation at time t. So we can write
−
+
Ψγ (x, t) = A(t) f−γ (x, t) + E(t) fγ+ (x, t) = C(t) f−γ (x, t) + D(t) fγ− (x, t)
Chapter 5: Inverse scattering method and the KdV equation 97

±
with A(0) = 1, E(0) = R(γ), C(0) = 0, D(0) = T (γ), where f±γ (x, t) denote the Jost
solutions of the corresponding Schrödinger operator at time t. Moreover, assume that q(·, t) is
compactly supported for all (fixed) t, so that Ψγ (x, t) = A(t) e−iγx + E(t) eiγx for x sufficiently
d3
large. Again for such x we have B(t) = −4 dx 3 and so

d
Ψγ (x, t) = B(t) Ψγ (x, t) = −4(−iγ)3 A(t) e−iγx − 4(iγ)3 E(t) eiγx
dt
d dA dE
Ψγ (x, t) = (t) e−iγx + (t) eiγx .
dt dt dt
This gives dt (t) = −4iγ A(t), A(0) = 1 and dE
dA 3
dt
(t) = 4iγ 3 E(t), E(0) = R(γ), hence A(t) =
exp(−4iγ 3 t), E(t) = exp(4iγ 3 t)R(γ).
For x sufficiently negative we find Ψγ (x, t) = C(t) eiγx + D(t) e−iγx , and
d
Ψγ (x, t) = B(t) Ψγ (x, t) = −4(iγ)3 C(t) eiγx − 4(−iγ)3 D(t) e−iγx
dt
d dC dD
Ψγ (x, t) = (t) eiγx + (t) e−iγx .
dt dt dt
This gives dC
dt
(t) = 4iγ 3 C(t), C(0) = 0 and dD dt
(t) = −4iγ 3 D(t), D(0) = T (γ). Hence
3
C(t) = 0, D(t) = exp(−4iγ t)T (γ).
Combining this then gives
3 3 3
Ψγ (x, t) = e−4iγ t T (γ) fγ− (x, t) = e−4iγ t f−γ
+
(x, t) + e4iγ t R(γ) fγ+ (x).
If we now denote the corresponding time-dependent Jost solutions, transmission and reflection
coefficients by
ψγ (x, t) = T (γ, t) fγ− (x, t) = f−γ
+
(x, t) + R(γ, t) fγ+ (x, t),
3
we obtain e4iγ t Ψγ (x, t) = ψγ (x, t). This then immediately gives T (γ, t) = T (γ), or the
3
transmission coefficient is independent of time, and R(γ, t) = e8iγ t R(γ).
Theorem 5.4.2. Assume q(x, t) is a solution to the KdV-equation (5.1.1) such that for each
∂k q
t ≥ 0 the function q(·, t), t ≥ 0, satisfies the conditions of Theorem 4.2.7 and such that ∂x k (x, t)

is bounded for k ∈ {0, 1, 2, 3} as |x| → ∞ and lim|x|→∞ q(x, t) = lim|x|→∞ qx (x, t) = 0. Then
d2
the scattering data of L(t) = − dx 2 + q(·, t) satisfies

3
T (γ, t) = T (γ), R(γ, t) = e8iγ t R(γ), pn (t) = pn (0), ρn (t) = exp(8p3n t)ρn (0).
Note that the theorem in particular shows that the poles of the transmission coefficient T ,
hence the eigenvalues of the Schrödinger operator, are time-independent.
In case q(·, t) is of compact support, then the above argumentation leads to the statements
of Theorem 5.4.2. For the N -soliton solutions this is not the case, and we need Theorem 5.4.2
in the more general form. The idea of the proof is the same.
Note that Theorem 5.4.2 provides a solution scheme for the KdV-equation with initial
condition (5.1.2) as long as the initial condition q(·, 0) satisfies the conditions. Determine the
scattering data at time t, and next use the Gelfand-Levitan-Marchenko integral equation as
in Theorem 4.4.2 to determine q(x, t).
98 Chapter 5: Inverse scattering method and the KdV equation

5.5 Pure soliton solutions to the KdV-equation

We now look for specific solutions of the KdV-equation using Theorem 5.4.2. The pure N -
soliton solution to the KdV-equation (5.1.1) is the solution that is determined using Theorem
5.4.2 starting with a reflectionfless potential and with N discrete points in the spectrum at
time t = 0. Note that by Theorem 5.4.2 the potential remains reflectionless in time, and the
corresponding Schrödinger operator has N discrete points in the spectrum in time. We only
consider the cases N = 1 and N = 2.
We first consider the case N = 1 and R(γ) = 0 at time t = 0. By Proposition 4.5.1 the
potential at time t = 0 equals q(x, 0) = −2 cosh−2 (x + C), C = 12 ln 2, see the derivation in
the beginning of Section 4.5, but the constant C can be scaled away. From Proposition 4.5.1
we see that p = p1 = 1 and ρ = ρ1 = 1 at time t = 0. Using Theorem 5.4.2 we find that at
arbitrary time t we have p(t) = 1 and ρ(t) = e8t . So we can solve the potential q(x, t) at time
t in the same way as in the beginning of Section 4.5, using p = 1, ρ = e8t , or we can just use
(4.5.1) with ρ = e8t . This then gives q(x, t) = −2 cosh−2 (x − 4t + C), which is precisely the
solution for the KdV-equation given in Proposition 5.1.10 for c = 4, and we obtain the pure
1-soliton solution.
Next we consider the case N = 2, so that we require at time t = 0 the conditions R(γ) = 0,
p1 = 1, p2 = 2, ρ1 = 21 , ρ2 = 1, q(x, 0) = −6 cosh−2 (x), as follows from Proposition 4.5.1. Now
the time evolution of the scattering data follows from Theorem 5.4.2, so that R(γ, t) = 0 for
all t, p1 (t) = p1 = 1, p2 (t) = p2 = 2, ρ1 (t) = 21 e8t , ρ2 (t) = e64t . We then need to solve for the
d
kernel B(x, y; t) and the potential q(x, t) = − dx B(x, 0) from the Gelfand-Levitan-Marchenko
equation. For this we use Proposition 4.5.3 in the case N = 2. It follows that the determinant
of the symmetric matrix I + S(x, t), which is now time-dependent, is given by
1  
det(I + S(x, t)) = e36t−3x cosh(3x − 36t + ln(12)) + 3 cosh(x − 28t)
6
using the calculation given after Proposition 4.5.3 for the case N = 2. By replacing x and t
by x + a, t + b for suitable a and b we can reduce to the case that

det(I + S(x, t)) = C e36t−3x cosh(3x − 36t) + 3 cosh(x − 28t)

for some constant C independent of x and t. Take the logarithm and deriving the resulting
0 2 00 (x)f (x)
expression twice and multiplying by −2 to find the potential; q(x, t) = 2 (f (x))(f−f
(x))2
, with
f (x) = det(I + S(x, t)). This shows that the potential q(x, t) is indeed independent of C.
Using a computer algebra system is handy, and in case one uses Maple, we find
5 cosh (−x + 28 t) cosh (−3 x + 36 t) + 3 − 3 sinh (−x + 28 t) sinh (−3 x + 36 t)
q(x, t) = −12 .
(3 cosh (−x + 28 t) + cosh (−3 x + 36 t))2
Using the addition formula cosh(x ± y) = cosh(x) cosh(y) ± sinh(x) sinh(y) we obtain
3 + 4 cosh(2x − 8t) + cosh(4x − 64t)
q(x, t) = −12
2 . (5.5.1)
3 cosh(x − 28t) + cosh(3x − 36t)
Chapter 5: Inverse scattering method and the KdV equation 99

Note that indeed

3 + 4 cosh(2x) + cosh(4x) −6
q(x, 0) = −12
2 = ,
3 cosh(x) + cosh(3x) cosh2 (x)

as can be done easily in e.g. Maple. Compare this with Proposition 4.5.3. The solution (5.5.1)
is known as the pure 2-soliton solution.
So we have proved from Theorem 5.4.2 and the inverse scattering method described in
Sections 4.4, 4.5 the following Proposition.

Proposition 5.5.1. Equation (5.5.1) is a solution to the KdV-equation (5.1.1) with initial
condition q(x, 0) = cosh−62 (x) .

Having the solution (5.5.1) at hand, one can check by a direct computation, e.g. using a
computer algebra system like Maple, that it solves the KdV-equation. However, it should be
clear that such a solution cannot be guessed without prior knowledge.
In Figure 5.2 we have plotted the solution q(x, t) for different times t. It seems that q(·, t)
exists of two solitary waves for t  0 and t  0, and that the ‘biggest’ solitary wave travels
faster than the ‘smallest’ solitary wave, and the ‘biggest’ solitary wave overtakes the ‘smallest’
at t = 0. For t ≈ 0 the waves interact, and then the shapes are not affected by this interaction
when t grows. This phenomenon is typical for soliton solutions.
In order to see this from the explicit form (5.5.1) we introduce new variables x1 = x−4p21 t =
x − 4t and x2 = x − 4p22 t = x − 16t, so Figure 5.2 suggest that the ‘top’ of the ‘biggest’ or
‘fastest travelling’ solitary wave occurs for x2 = 0, and the top for the ‘smallest’ or ‘slowest
travelling’ solitary wave occurs for x1 = 0. So we see the ‘biggest’ solitary wave travel at four
times the speed of the ‘smallest’ solitary wave.
We first consider the ‘fastest travelling’ solitary wave, so we rewrite

3 + 4 cosh(2x2 + 24t) + cosh(4x2 )

q2 (x2 , t) = q(x2 + 16t, t) = −12
2 ,
3 cosh(x2 − 12t) + cosh(3x2 + 12t)

so that for t → ∞ we have

2e2x2 +24t −8
q2 (x2 , t) ∼ −12 =
3 12t−x2 1 3x2 +12t 2 2
cosh (2x2 − 12 ln(3))


2
e + 2
e

and for t → −∞ we have

2e−2x2 −24t −8
q2 (x2 , t) ∼ −12 = .
3 −12t+x2 1 −3x2 −12t 2 2
cosh (2x2 + 12 ln(3))


2
e + 2
e

From these asymptotic considerations, we see that the ‘biggest’ solitray wave undergoes a
phase-shift, meaning that its top is shifted by ln(3) forward during the interaction with the
‘smallest’ solitary wave.
100 Chapter 5: Inverse scattering method and the KdV equation

Similarly, we can study

3 + 4 cosh(2x1 ) + cosh(4x1 − 48t)

q1 (x1 , t) = q(x1 + 4t, t) = −12
2 ,
3 cosh(x1 − 24t) + cosh(3x1 − 24t)

so that
 1 48t−4x1
2
e −2
−12 t → ∞,


 3 24t−x1 = ,
(2e + 12 e24t−3x1 )2 2
cosh (x1 + 12 ln(3))
q1 (x1 , t) ∼ 1 −48t+4x1
2
e −2
−12 , t → −∞.


3 −24t+x1 =
+ 12 e−24t+3x1 )2 2
cosh (x1 − 12 ln(3))

 (2e

So the ‘smallest’ wave undergoes the same phase-shift, but in the opposite direction.
So we can conclude that
−2 −8
q(x, t) ∼ 2 + 2 , t → ±∞.
cosh (x − 4t ± 2 ln(3)) cosh (x − 16t ∓ 12 ln(3))
1

This also suggests that one can build new solutions out of known solutions, even though
the Korteweg-de Vries equation (5.1.1) is non-linear. This is indeed the case, we refer to [2],
[6] for more information.

Remark 5.5.2. It is more generally true that a solution which exhibits a solitary wave in
its solution for t  0, then the speed of this solitary wave equals −4λ for some λ < 0 in the
discrete spectrum of the corresponding Schrödinger operator.

Remark 5.5.3. It is clear that for N ≥ 3 the calculations become more and more cumbersome.
There is a more unified treatment of soliton solutions possible, using the so-called τ -functions.
For an introduction of these aspects of the KdV-equation in relation also to vertex algebras
one can consult [8].
Chapter 5: Inverse scattering method and the KdV equation 101

x x x
-20 -10 0 10 20 -20 -10 0 10 20 -20 -10 0 10 20
0 0 0

-2 -2 -2

y -4 y -4 y -4

-6 -6 -6

-8 -8 -8

x x x
-20 -10 0 10 20 -20 -10 0 10 20 -20 -10 0 10 20
0 0 0

-2 -2 -2

y -4 y -4 y -4

-6 -6 -6

-8 -8 -8

x x x
-20 -10 0 10 20 -20 -10 0 10 20 -20 -10 0 10 20
0 0 0

-2 -2 -2

y -4 y -4 y -4

-6 -6 -6

-8 -8 -8

Figure 5.2: N = 2-soliton solution for t = −1, t = −0.5, t = −0.2 (first row), t = −0.05,
t = 0, t = 0.05 (second row) and t = 0.2, t = 0.5, t = 1 (third row).
102 Chapter 5: Inverse scattering method and the KdV equation
Chapter 6

Preliminaries and results from

functional analysis

In this Chapter we recall several notions from functional analysis in order to fix notations and
to recall standard results. Most of the unproved results can be found in Lax [7], Werner [11] or
in course notes for a course in Functional Analysis. Some of the statements are equipped with
a proof, notably in Section 6.5 for the description of the spectrum and the essential spectrum.

6.1 Hilbert spaces

A Banach1 space X is a vector space (over C) equipped with a norm, i.e. a mapping k·k : X →
R such that
• kαxk = |α| kxk, ∀ α ∈ C, ∀x ∈ X ,

• kx + yk ≤ kxk + kyk, ∀ x, y ∈ X , (triangle inequality)

• kxk = 0 ⇐⇒ x = 0,
such that X is complete with respect to the metric topology induced by d(x, y) = kx − yk.
Completeness means that any Cauchy2 sequence, i.e. a sequence {xn }∞ n=1 in X such that
∀ε > 0 ∃N ∈ N ∀n, m ≥ N kxn − xm k < ε, converges in X to some element x ∈ X , i.e. ∀ε > 0
∃N ∈ N ∀n ≥ N kxn − xk < ε. If xn → x in X , then kxn k → kxk by the reversed triangle
inequality |kxk − kyk| ≤ kx − yk.
An inner product on a vector space H is a mapping h·, ·i : H × H → C such that
• hαx + βy, zi = αhx, zi + βhy, zi ∀α, β ∈ C, ∀x, y, z ∈ H,

• hx, yi = hy, xi, ∀x, y ∈ H,

1
Stefan Banach (30 March 1892 — 31 August 1945), Polish mathematician, who is one of the founding
fathers of functional analysis.
2
Augustin Louis Cauchy (21 August 1789 — 23 May 1857), French mathematician.

103
104 Chapter 6: Preliminaries and results from functional analysis

• hx, xi ≥ 0 ∀x ∈ H and hx, xi = 0 ⇐⇒ x = 0.

p
It can be shown that kxk = hx, xi induces a norm on H, and we call H a Hilbert3 space if
it is complete. The standard inequality is the Cauchy-Schwarz4 inequality;

|hx, yi| ≤ kxk kyk. (6.1.1)

A linear map φ : H → C on a Hilbert space H, also known as a functional, is continuous

if and only if |φ(x)| ≤ Ckxk for some constant C. Note that (6.1.1) implies that x 7→ hx, yi is
continuous. The Riesz5 representation theorem states that any continuous functional φ : H →
C is of this form, i.e. ∃y ∈ H such that φ(x) = hx, yi. A sequence {xn }∞ n=1 in H is weakly
convergent to x ∈ H if limn→∞ hxn , yi = hx, yi, ∀y ∈ H. By the Cauchy-Schwarz inequality
(6.1.1) we see that convergence in H, i.e. with respect to the norm, implies weak convergence.
The converse is not true in general. In general, a bounded sequence in a Hilbert space does
∞
not need have a convergent subsequence (e.g. take any orthonormal sequence √ {xn }n=1 in an
infinite dimensional Hilbert space, since then kxn k = 1 and kxn − xm k = 2, n 6= m, by
the Pythagorean6 theorem), but it has a weakly convergent subsequence (e.g. in the example
xn → 0 weakly). This is known as weak compactness. This is actually weak sequentially
compact, but the Eberlein-Smulian theorem states that a weakly closed, bounded set A in
a Banach space X is weakly compact if and only if any sequence in A has a convergent
subsequence.
Exercise 6.1.1. In this exercise we sketch a proof of the weak sequentially compactness of a
bounded set in a Hilbert space H. Let {xn }∞
n=1 be a bounded sequence, say kxn k ≤ 1. We have
to show that there exists a subsequence {xnk }∞k=1 and a x ∈ H such that limk→∞ hxnk , yi =
hx, yi for all y ∈ H.
• Fix m, and use |hxn , xm i| ≤ 1 to find a subsequence {xni }∞
i=1 such that hxni , xm i con-
verges as i → ∞.

• Find a subsequence, say {yn }∞ n=1 such that hyn , xm i converges for all m as n → ∞.
Conclude that hyn , ui converges for all u in the linear span U of the elements xn , n ∈ N.

• Show that hyn , ui converges for all u in the closure U of the linear span U

• By writing an arbitrary w ∈ H as w = u+z, u ∈ U, z ∈ U ⊥ , show that hyn , wi converges

for all w ∈ H as n → ∞.

• Define φ : H → C as φ(w) = limn→∞ hyn , wi. Show that φ is a continuous linear func-
tional.
3
David Hilbert (23 January 1862 —14 February 1943), German mathematician. Hilbert is well-known for
his list of problems presented at the ICM 1900, some still unsolved.
4
Hermann Amandus Schwarz (25 January 1843 — 30 November 1921), German mathematician.
5
Frigyes Riesz (22 January 1880 — 28 February 1956), Hungarian mathematician, who made many contri-
butions to functional analysis.
6
Pythagoras of Samos (approximately 569 BC – 475 BC), Greek philosopher.
Chapter 6: Preliminaries and results from functional analysis 105

• Use the Riesz representation theorem to find x ∈ H such that the subsequence {yn }∞
n=1
converges weakly to x.

Example 6.1.2. Using the Lebesgue7 integral, we put for 1 ≤ p < ∞

Z
p
L (R) = {f : R → C measurable | |f (x)|p dx < ∞}.
R

R 1/p
This is a Banach space with respect to the norm kf kp = R |f (x)|p dx . The case p = ∞
∞
is L (R) = {f : R → C measurable | ess supx∈R |f (x)| < ∞}, which is a Banach space for
the norm kf k∞ = ess supx∈R |f (x)|. Here we follow the standard convention of identifying two
functions that are equal almost everywhere.
The case p = 2 gives a Hilbert space L2 (R) with inner product given by
Z
hf, gi = f (x) g(x) dx
R

Since we work mainly with the Hilbert space L2 (R), we put kf k = kf k2 . In this case the
Cauchy-Schwarz inequality states
Z Z  21 Z  21
  2 2
 f (x)g(x) dx ≤ |f (x)| dx |g(x)| dx ,
R R R

which is also known as Hölder’s inequality. The converse Hölder’s inequality states for mea-
surable function f such that
Z
 
sup  f (x)g(x) dx = C < ∞,
R
R
where the supremum is taken over all functions g such that kgk ≤ 1 and R
f (x)g(x)dx exists,
we have f ∈ L2 (R) and kf k = C.

We recall some basic facts from Lebesgue’s integration theory. We also recall that for a
convergent sequence {fn }∞ p
n=1 to f in L (R), 1 ≤ p ≤ ∞ there exists a convergent subsequence
{fnk }∞
k=1 that converges pointwise almost everywhere to f .

Theorem 6.1.3 (Lebesgue’s Dominated Convergence Theorem). If {f }∞ n=1 is a sequence of

measurable functions on R, such that fn → f pointwise almost everywhere on a measurable
set E ⊂ R. If there exists a function g integrable on E such that |fn | ≤ g almost everywhere
on E, then Z Z
lim fn (x) dx = f (x) dx.
n→∞ E E
7
Henri Lon Lebesgue (28 June 1875 — 26 July 1941), French mathematician, and inventor of modern
integration theory.
106 Chapter 6: Preliminaries and results from functional analysis

The Dominated Convergence Theorem 6.1.3 is needed to show that the Lp (R) spaces are
complete. Another useful result is the following.

RTheorem 6.1.4 (Lebesgue’s Differentiation Theorem). For f ∈ L1 (R) its indefinite integral
x
−∞
f (y) dy is differentiable with derivative f (x) almost everywhere.

A Banach space is separable if there exists a denumerable dense set. The examples Lp (R),
1 ≤ p < ∞, are separable, but L∞ (R) is not separable. In these lecture notes the Hilbert
spaces are assumed to be separable.
Another example of a Banach space is C(R), the space of continuous functions f : R → C
with respect to the supremum norm, kf k∞ = supx∈R |f (x)|. The Arzelà-Ascoli8 theorem
states that M ⊂ C(R) with the properties (i) M bounded, (ii) M closed, (iii) M is uniformly
continuous (i.e. ∀ε > 0 ∃δ > 0 ∀f ∈ M : |x − y| < δ ⇒ |f (x) − f (y)| < ε), then M is compact
in C(R).

6.2 Operators
An operator is a linear map T from its domain D(T ), a linear subspace of a Banach space X ,
to a Banach space Y, denoted by T : X ⊃ D(T ) → Y or by T : D(T ) → Y or by (T, D(T )) if
X and Y are clear from the context, or simply by T . We say that (T, D(T )) is an extension of
(S, D(S)), in notation S ⊂ T , if D(S) ⊂ D(T ) and Sx = T x for all x ∈ D(S). Two operators
S and T are equal if S ⊂ T and T ⊂ S, so that in particular the domains have to be equal.
By Ker(T ) we denote the kernel of T , Ker(T ) = {x ∈ D(T ) | T x = 0}, and by Ran(T ) we
denote its range, Ran(T ) = {y ∈ Y | ∃x ∈ D(T ) such that T x = y}. The graph norm on
D(T ) is given by kxkT = kxk + kT xk.
The operator (T, D(T )) is densely defined if the closure of its domain is the whole Banach
space, D(T ) = X . We define the sum of two operators S and T by (S + T )x = Sx + T x for
x ∈ D(T + S) = D(T ) ∩ D(S), and the composition is defined as (ST )x = S(T x) with domain
D(ST ) = {x ∈ D(T ) | T x ∈ D(S)}. Note that it might happen that the domains of the sum
or composition are trivial, even if S and T are densely defined.
A linear operator T : X → Y is continuous if and only if T is bounded, i.e. there exists a
constant C such that kT xk ≤ Ckxk for all x ∈ X . The operator norm is defined by

kT xk
kT k = sup
x∈X kxk

and the space B(X , Y) = {T : X → Y | T bounded} is a Banach space with respect to

the operator norm kT k. If (T, D(T )) is densely defined and ∃C such that kT xk ≤ Ckxk
∀x ∈ D(T ), then T can be extended uniquely to a bounded operator T : X → Y with kT k ≤ C.
P B(X n) = B(X , X ). For an operator T ∈ B(X ) with operator norm kT k < 1, we check
We put
that ∞ n=0 T converges with respect to the operator norm and this gives an inverse to 1 − T ,

8
Cesare Arzelà (6 March 1847 — 15 March 1912), Italian mathematician. Ascoli Italian mathematician.
Chapter 6: Preliminaries and results from functional analysis 107

where 1 ∈ B(X ) is the identity operator x 7→ x. For Hilbert spaces we use apart from the
operator norm on B(H) also the strong operator topology, in which Tn →T if Tn x → T x
for all x ∈ H, and the weak operator topology, in which Tn → T if hTn x, yi → hT x, yi for
all x, y ∈ H. Note that convergence in operator norm implies convergence in strong operator
topology, which in turn implies convergence in weak operator topology.
For T ∈ B(H), H Hilbert space, the adjoint operator T ∗ ∈ B(H) is defined by hT x, yi =
hx, T ∗ yi for all x, y ∈ H. We call T ∈ B(H) a self-adjoint (bounded) operator if T ∗ = T . If a
self-adjoint operator T satisfies hT x, xi ≥ 0 for all x ∈ H, then T is a positive operator, T ≥ 0.
T ∈ B(H) a unitary operator if T ∗ T = 1 = T T ∗ . An isometry is an operator T ∈ B(H), which
preserves norms; kT xk = kxk for all x ∈ H. A surjective isometry is unitary. An orthogonal
projection is a self-adjoint operator P ∈ B(H) such that P 2 = P , so that P projects onto
Ran(P ), a closed subspace of H, or P |Ran(P ) is the identity and Ker(P ) = Ran(P )⊥ , the
orthogonal complement. In particular P is a positive operator. A partial isometry U ∈ B(H)
is an element such that U U ∗ and U ∗ U are orthogonal projections. The range of the projection
U ∗ U is the inital subspace, say D, and the range of the projection U U ∗ is the final subspace,
say R, and we can consider U as a unitary map from D to R.
An operator T : H ⊃ D(T ) → H is symmetric if hT x, yi = hx, T yi for all x, y ∈ D(T ). For
a densely defined operator (T, D(T )) we define

D(T ∗ ) = {x ∈ H | y 7→ hT y, xi is continuous on D(T )}.

Since D(T ) is dense in H it follows that y 7→ hT y, xi extends to a continuous functional on H,

which, by the Riesz representation theorem, is equal to y 7→ hT y, xi = hy, zi for some z ∈ H.
Then we define T ∗ x = z. An operator (T, D(T )) is self-adjoint if (T, D(T )) equals (T ∗ , D(T ∗ )).
So in particular, a self-adjoint operator is densely defined. A self-adjoint operator (T, D(T ))
is positive if hT x, xi ≥ 0 for all x ∈ D(T ). Note that a densely defined symmetric operator
satisfies T ⊂ T ∗ , and that a self-adjoint operator is symmetric, but not conversely. For a
self-adjoint operator (T, D(T )) we have hT x, xi ∈ R for all x ∈ D(T ). Note that the definition
of self-adjointness coincides with the definition of self-adjointness in case the operator T is
bounded, since then D(T ) = H = D(T ∗ ).
We now switch back to operators from a Banach space X to a Banach space Y. An operator
K is compact if the closure of the image of the unit ball {x ∈ X | kxk ≤ 1} is compact in Y,
and we denote the compact operators by K(X , Y). It follows that K(X , Y) ⊂ B(X , Y), i.e.
each compact operator is bounded. Compactness can be restated as follows: for each bounded
sequence {xn }∞ ∞
n=1 in X , the sequence {Kxn }n=1 has a convergent subsequence. Any operator
with finite dimensional range is compact, we denote these operators by F (X , Y). A typical
example is the rank one operator x 7→ hx, yi z for fixed y, z ∈ H, For a Hilbert space H we
have that K(H) = K(H, H) is the operator norm closure of F (H) = F (H, H), and K(H) is
a closed two-sided ideal in B(H), i.e. K ∈ K(H) and S ∈ B(H) implies SK, KS ∈ K(H).
An operator T : X ⊃ D(T ) → Y is closed if for a sequence {xn }∞ n=1 in D(T ) with the
properties that xn → x in X and T xn → y in Y we may conclude that x ∈ D(T ) and T x = y.
This is equivalent to the assumption that the graph of T , i.e. G(T ) = {(x, T x) | x ∈ D(T )} is
closed in X × Y. An operator (T, D(T )) is closable if there exists a closed operator (S, D(S))
108 Chapter 6: Preliminaries and results from functional analysis

such that S extends T , T ⊂ S. The smallest (with respect to extensions) closed operator of a
closable operator is its closure, denoted (T̄ , D(T̄ )). For closed operator (T, D(T )) its domain
is complete with respect to the graph norm k · kT .

Theorem 6.2.1 (Closed Graph Theorem). A closed operator (T, D(T )) with D(T ) = X is
bounded, T ∈ B(X ).

A closed operator (T, D(T )) from one Hilbert space H1 to another Hilbert space H2 has
a polar decomposition, i.e. T = U |T |, where D(|T |) = D(T ) and (|T |, D(|T |)) is self-adjoint
operator on H1 and U : H1 → H2 is a partial isometry with initial space (Ker T )⊥ and final
space Ran T . The condition Ker T = Ker |T | determine U and |T | uniquely.
For an operator on a Hilbert space H we have that (T, D(T )) is closable if and only its
adjoint (T ∗ , D(T ∗ )) is densely defined and then its closure is (T ∗∗ , D(T ∗∗ )). In particular,
any densely defined symmetric operator is closable, its closure being T ∗∗ . In particular, if
(T, D(T )) is self-adjoint, then it is a closed operator. For a closed, densely defined operator
(T, D(T )) on a Hilbert space H, the linear space D(T ) is a Hilbert space with respect to
hx, yiT = hx, yi + hT x, T yi. The corresponding norm is the graph norm.
The following lemma gives necessary and sufficient conditions for a densely defined, closed,
symmetric operator (T, D(T )) to be a self-adjoint operator on the Hilbert space H.

Lemma 6.2.2. Let (T, D(T )) be densely defined, symmetric operator, then the following are
equivalent.

1. Its closure (T̄ , D(T̄ )) is a self-adjoint operator,

2. Ker(T ∗ + i) = {0} = Ker(T ∗ − i),

3. Ran(T + i) = H = Ran(T − i),

4. Ker(T ∗ − z) = {0} for all z ∈ C\R,

5. Ran(T − z) = H for all z ∈ C\R.

In case (T, D(T )) is closed, the spaces Ran(T ± i), Ran(T − z) are closed.

In fact for a closed, densely defined, symmetric operator T the dimension of Ker(T ∗ − z)
and Ran(T − z) is constant in the upper half plane =z > 0 and in the lower half plane =z < 0.
A densely defined, symmetric operator (T, D(T )) whose closure is self-adjoint, is known
as an essentially self-adjoint operator.
Sketch of proof. We prove Lemma 6.2.2 in case T is closed and for the equivalence of the first
three assumptions. Note that 5 ⇒ 3 and 4 ⇒ 2.
First, for y ∈ Ran(T + z)⊥ we have hT x, yi + zhx, yi = h(T + z)x, yi = 0 for all x ∈ D(T ),
so that in particular by definition y ∈ D(T ∗ ) and h(T ∗ + z̄)y, xi = 0 for all x ∈ D(T ). Since
D(T ) is dense, it follows that (T ∗ + z̄)y = 0 or y ∈ Ker(T ∗ + z̄). The reverse conclusion can
be obtained by reversing the argument. This shows that 2 ⇔ 3, and also 4 ⇔ 5.
Chapter 6: Preliminaries and results from functional analysis 109

Note that hT x, xi ∈ R, x ∈ D(T ), for T symmetric, so that

k(T + i)xk2 = h(T + i)x, (T + i)xi = kT xk2 + kxk2 + 2<hT x, ixi = kT xk2 + kxk2 ≥ kxk2 .

In particular, (T + i)x = 0 implies x = 0, or Ker(T + i) = {0}. Similarly, Ker(T − i) = {0}, so

1 ⇒ 2. Another consequence of this inequality is that (T + i)−1 : Ran(T + i) → D(T ) exists
and k(T + i)−1 k ≤ 1. Let {yn }∞ n=1 be a Cauchy sequence in Ran(T + i), put yn = (T + i)xn
and let yn → y in H. Then {xn }∞ n=1 is a Cauchy sequence in H, with limit, say, x. Then
T xn = yn − ixn → y − ix. Since T is closed, it follows that y ∈ D(T ) and T x = y − ix or
(T + i)x = y ∈ Ran(T + i). Similarly one establishes the closedness of the other ranges.
To prove 3 ⇒ 1 it suffices to check that D(T ∗ ) ⊂ D(T ), since we already have T ⊂ T ∗ ,
since T is a densely defined symmetric operator. So take y ∈ D(T ∗ ) arbitrary, and consider
(T ∗ − i)y = (T − i)x for some x ∈ D(T ), since Ran(T − i) = H by the previous paragraph.
Since T ⊂ T ∗ , it follows that (T ∗ −i)y = (T ∗ −i)x or y −x ∈ Ker(T ∗ −i) = Ran(T +i)⊥ = {0},
by the already established equivalence 2 ⇔ 3. Hence y = x ∈ D(T ).
d
Exercise 6.2.3. Consider the operator T = i dx .

• Take H = L2 (R) and D(T ) = Cc1 (R), the space of continuously differentiable functions
with compact support. You may assume that Cc1 (R) is dense in L2 (R). Show that
(T, D(T )) is essentially self-adjoint.

– Show that (T, D(T )) is symmetric.

R
– Fix z ∈ C\R arbitrary. Show that R eizx f (x) dx = 0 for f ∈ Ran(T −z) by solving
iu0 − zu = f .
R izx
– Conversely, if f is a continuous function with compact support
R x iz(y−x)and R
e f (x) dx =
0, then f ∈ Ran(T − z). (Hint: consider u(x) = −i −∞ e f (y) dy.)
– ShowRthat the space consisting of continuous functions f with compact support
with R eizx f (x) dx = 0 is dense in L2 (R), and conclude, using Lemma 6.2.2, that
(T, D(T )) is essentially self-adjoint.

• Take H = L2 ([0, ∞]), and D(T ) = Cc1 (0, ∞), the space of continuously differentiable
functions with compact support in (0, ∞). You may assume that Cc1 (0, ∞) is dense in
L2 ([0, ∞]). Show that T is not essentially self-adjoint.

– Show that (T, D(T )) is symmetric.

– Show that R eizx f (x) dx = 0 for f ∈ Ran(T − z) by solving iu0 − zu = f .
R

– Show that for =z > 0, Ran(T − z) is not dense in H.

– Conclude that (T, D(T )) is not essentially self-adjoint

In this case one can show that there doesn’t exist a self-adjoint (S, D(S)) such that
T ⊂ S, or T has no self-adjoint extensions.
110 Chapter 6: Preliminaries and results from functional analysis

There is an abundance of positive self-adjoint operators, where (T, D(T )) is a positive

operator if hT x, xi ≥ 0 for all x ∈ D(T ).

Lemma 6.2.4. Let (T, D(T )) be a closed, densely defined operator on H. Then T ∗ T with its
domain D(T ∗ T ) = {x ∈ D(T ) | T x ∈ D(T ∗ )} is a densely defined self-adjoint operator on H.
Moreover, this operator is positive and its spectrum σ(T ∗ T ) ⊂ [0, ∞).

See Section 6.4 for the definition of the spectrum. We only apply Lemma 6.2.4 in case T
is a self-adjoint operator, and in this case Lemma 6.2.4 follows from the Spectral Theorem
6.4.1.

6.3 Fourier transform and Sobolev spaces

For an integrable function f ∈ L1 (R) we define its Fourier9 transform as
Z
ˆ 1
Ff (λ) = f (λ) = √ f (x) e−iλx dx. (6.3.1)
2π R

If f 0 , f ∈ L1 (R) we have, by integration by parts, F(f 0 )(λ) = iλ(Ff )(λ), so the Fourier
transform intertwines differentiation and multiplication. The Riemann10 -Lebesgue lemma
states that F : L1 (R) → C0 (R), where C0 (R) is the space of continuous functions on R that
vanish at infinity. For f ∈ L1 (R) such that Ff ∈ L1 (R) we have the Fourier inversion formula
1
R
f (x) = √2π R (Ff )(λ) eiλx dλ. We put
Z
1
F f (λ) = fˇ(λ) = √
−1
f (x) eiλx dx.
2π R
Rn
For f ∈ L2 (R) ∩ L1 (R) one can show that −n f (x)e±iλx dx converges in L2 (R)-norm as
n → ∞. This defines F, F −1 : L2 (R) → L2 (R), and then Parseval’s11 identity holds;

hf, gi = hFf, Fgi. (6.3.2)

In particular, F : L2 (R) → L2 (R) is unitary, kFf k = kf k, which is Plancherel’s12 identity.

Using the Fourier transform we define the Sobolev13 space

W m (R) = {f ∈ L2 (R) | λ 7→ (iλ)p (Ff )(λ) ∈ L2 (R), ∀p ∈ {0, 1, 2, · · · , m} } (6.3.3)

9
Jean Baptiste Joseph Fourier (21 March 1768 — 16 May 1830), French mathematician, who invented
Fourier series, and used it to solve the heat equation.
10
Georg Friedrich Bernhard Riemann (17 September 1826 — 20 July 1866), German mathematician, who
made important contributions to many different areas of mathematics, and posed the Riemann hypothesis,
one of the most famous open problems.
11
Marc-Antoine Parseval des Chênes (27 April 1755 — 16 August 1836), French mathematician.
12
Michel Plancherel (16 January 1885 — 4 March 1967), Swiss mathematician.
13
Sergei Lvovich Sobolev (6 October 1908 — 3 January 1989), Russian mathematician.
Chapter 6: Preliminaries and results from functional analysis 111

The Sobolev space is a Hilbert space for the inner product

m
X
hf, giW m (R) = h(iλ)p (Ff ), (iλ)p (Fg)iL2 (R) .
p=0

We can also define the Sobolev space using weak derivatives . We say that f ∈ L2 (R) has
p
a weak derivative of order p ∈ N in case Cc∞ (R) 3 φ 7→ (−1)p hf, ddxφp i is continuous (as a
functional on the Hilbert space L2 (R)). Here Cc∞ (R) is the space of infinitely many times
differentiable functions having compact support. This space is dense in Lp (R), 1 ≤ p < ∞.
p
By the Riesz representation theorem there exists g ∈ L2 (R) such that hg, φi = (−1)p hf, ddxφp i,
and we define the p-th weak derivative of f as Dp f = g. Then one can show that
W m (R) = {f ∈ L2 (R) | Dp f ∈ L2 (R) exists, ∀p ∈ {0, 1, 2, · · · , m}},
m
X
hf, giW m (R) = hDp f, Dp giL2 (R)
p=0

By abuse of notation we put f 0 = Df , f 00 = D2 f , etc. for weak derivatives.

Note that Cc∞ (R) ⊂ W m (R), and since Cc∞ (R) is dense in L2 (R), it follows that the Sobolev
spaces W m (R) are dense in L2 (R).
Lemma 6.3.1 (Sobolev imbedding). For m, k ∈ N such that m > k + 21 we have W m (R) ⊂
C k (R), with C k (R) the space of k-times continuous differentiable functions on R.
Note that this means, that f ∈ W m (R) can be identified with a function in C k after
changing it on a set of measure zero. In fact this inclusion is even continuous, and a proof for
the case m = 2, k = 1 is given in the proof of Theorem 2.3.4.
We also require a classical theorem of Paley and Wiener concerning the Fourier transform.
A function analytic in the open complex upper half plane is an element of the Hardy14 class
H2+ if Z
C = sup |f (x + iy)|2 dx < ∞.
y>0 R
√
Then C is the corresponding H2+ -norm of f .
Theorem 6.3.2 (Paley15 -Wiener16 ). f ∈ H2+ if and only if f is the inverse Fourier transform
of F ∈ L2 (R) with supp(F ) ⊂ [0, ∞),
Z ∞
1
f (z) = √ F (λ)eizλ dλ, z ∈ C, =z > 0.
2π 0
14
Godfrey Harold Hardy (7 February 1877 – 1 December 1947), English mathematician, also famous for his
quote “There is no permanent place in the world for ugly mathematics.”.
15
Raymond Edward Alan Christopher Paley (7 January 1907 — 7 April 1933), English mathematician,
killed by an avalanche while skiing during his visit to Wiener.
16
Norbert Wiener (26 November 1894 — 18 March 1964), American mathematician, who made numerous
contributions to different areas in mathematics.
112 Chapter 6: Preliminaries and results from functional analysis

Moreover, the following Plancherel formula holds;

Z ∞ Z
|F (λ)| dλ = sup |f (x + iy)|2 dx,
2
0 y>0 R

i.e. the L2 -norm of F is the H2+ -norm of f , and limb↓0 f (· + ib) = F −1 F in L2 (R).
Proof. First, if f is the inverse Fourier transform of F supported on [0, ∞), then f is an
analytic function in the open upper half plane since, writing z = a + ib,
Z ∞
1
f (z) = f (a + ib) = √ F (λ)e−bλ eiaλ dλ
2π 0
R
and this integral converges absolutely in the open upper half plane. Then C f (z) dz = 0 for
any closed curve in the open upper half plane by interchanging integrations and eiλz being
analytic. So f is analytic by Morera’s theorem. By the Plancherel identity we have

sup ka 7→ f (a + ib)k = sup kλ 7→ e−bλ F (λ)k = kF k,

b>0 b>0

and
  Z ∞
−1 −1 −bλ
kf (· + ib) − F 2
F k = kF λ 7→ (1 − e 2
)F (λ) k = (1 − e−bλ )2 |F (λ)|2 dλ → 0,
0

as b ↓ 0 by the Dominated Convergence Theorem 6.1.3.

To finish the proof, we have to show that any element f in the Hardy class H2+ can be
written in this way. Define fb (x) = f (x + ib), i.e. the function restricted to the line =z = b.
Then fb ∈ L2 (R), and kfb k ≤ C independent of b. Consider

ebλ
Z Z Z
−iλx 1 −iλ(x+ib) 1
bλ
e F(fb )(λ) = √ e fb (x) dx = √ e fb (x) dx = √ e−iλz f (z) dz
2π R 2π R 2π =z=b
We claim that this expression is independent of b > 0 as functions in L2 (R). Assuming this
claim is true, we define F (λ) = ebλ F(fb )(λ), then
Z
e−2bλ |F (λ)|2 dλ = ke−bλ F k2 = kFfb k2 = kfb k2 ≤ C 2
R

independent of b. From this we can observe the following; (i) F (λ) = 0 for λ < 0 (almost
everywhere) by considering b → ∞, (ii) F ∈ L2 (R) by taking the limit b ↓ 0, and (iii), as
identity in L2 (R),
Z ∞
−1
  1
f (z) = f (a + ib) = fb (a) = F Ffb (a) = √ eiλa (Ffb )(λ) dλ
Z ∞ Z ∞ 2π 0
1 1
=√ eiλa e−bλ F (λ) dλ = √ eiλz F (λ) dλ.
2π 0 2π 0
Chapter 6: Preliminaries and results from functional analysis 113

The proof of theR claim is as follows. Let Cα , α > 0, be the rectangle with vertices at ±α + i
and ±α + ib, then Cα e−iλz f (z) dz = 0 by analyticity of the integrand in the open upper half
R α+ib
plane. Put I(α) = α+i e−iλz f (z) dz in case b > 1 (the other case is similar). Then
Z b Z b Z b
2 −iλ(α+it) 2 2
|I(α)| = |i e f (α + it) dt| ≤ |f (α + it)| dt e2tλ dt,
1 1 1

where the first integral is independent of λ. Since f ∈ H2+ it follows

Z Z b Z bZ
2
|f (α + it)|2 dα dt ≤ C 2 |b − 1|,


|f (α + it)| dt dα =
R 1 1 R
Rb
with C the Hardy space norm of f . Or α 7→ G(α) = 1 |f (α + it)|2 dt is in L1 (R), there exist
sequence βn → ∞ and γn → −∞ such that G(βn ) → 0 and G(γn ) → 0, and hence I(βn ) → 0
and I(γn ) → 0. Note that these sequences are independent of λ by the estimate above. We
now define Z βn
1
gn (b, λ) = √ f (x + ib) e−iλx dx,
2π γn
so that
lim eλb gn (b, λ) − eλ gn (1, λ) = 0.
n→∞

By the Plancherel theorem we have limn→∞ kF(fb ) − gn (b, ·)k = 0, so that we can find a
subsequence of {gn (b, ·)}∞n=1 that converges pointwise to F(fb ) almost everywhere, and by
restricting the previous limit to this subsequence we see that ebλ F(fb )(λ) is independent of b
as claimed.
In the last part of the proof we have used the fact that if {fn } is a Cauchy sequence in
L (R) converging to f ∈ L2 (R), then there exists a subsequence {fnk } such that fnk → f ,
2

k → ∞, pointwise almost everywhere.

6.4 Spectral theorem

We consider densely defined operators (T, D(T )) on a Hilbert space H. Then z ∈ C is in the
resolvent set if T − z is invertible in B(H), i.e. there exists a bounded linear operator R(z; T )
such that R(z; T ) : H → D(T ) ⊂ H such that R(z; T )(T − z) ⊂ (T − z)R(z; T ) = 1. Note
that this implies T − z to be a bijection D(T ) → H. The resolvent set is denoted by ρ(T ),
and its complement σ(T ) is its spectrum. The spectrum is always closed in C. If for λ ∈ C
there exists an element x ∈ D(T ) ⊂ H such that T x = λx, then we say that λ ∈ σp (T ), point
spectrum, and σp (T ) ⊂ σ(T ). Then x is an eigenvector for the eigenvalue λ ∈ σp (T ).
For (T, D(T )) a self-adjoint operator we have σ(T ) ⊂ R, and for a positive self-adjoint
operator we have σ(T ) ⊂ [0, ∞).
Recall that a σ-algebra on a space X is a collection M of subsets of X, such that (i)
X ∈ M, (ii) for A ∈ M its complement Ac = X\A ∈ M (iii) if Ai ∈ M then ∪∞ n=1 Ai ∈ M.
114 Chapter 6: Preliminaries and results from functional analysis

The Borel17 sets on R are the smallest σ-algebra that contain all open intervals. A (positive)
measure is a map µ : M → [0, ∞] such that µ P is countably additive, i.e. for all sets Ai ∈ M
∞
with Ai ∩Aj = ∅ for i 6= j we have µ(∪∞ A
i=1 i ) = i=1 µ(Ai ). A complex measure is a countably
additive map µ : M → C.
For the Spectral Theorem 6.4.1 we need the notion of spectral measure, which is a orthog-
onal projection valued measure on the Borel sets of R. So, denoting the spectral measure by
E, this means that for any Borel set A ⊂ R, E(A) ∈ B(H) is an orthogonal projection such
that E(∅) = 0, E(R) = 1 (the identity element of B(H)) and for pairwise disjoint Borel sets
A1 , A2 , · · · we have
X∞
E(Ai ) x = E(∪∞ i=1 Ai ) x, ∀ x ∈ H.
i=1

This can be restated as that the series ∞

P
i=1 E(Ai ) converges in the strong operator topology
∞
of B(H) to E(∪i=1 Ai ). One can show that e.g. E(A ∩ B) = E(A)E(B), so that spectral
projections commute.
In particular, for E a spectral measure and x, y ∈ H we can define a complex-valued Borel
measure Ex,y on R by Ex,y (A) = hE(A)x, yi for A ⊂ R a Borel set. Note that Ex,x (A) is a
positive Borel measure, since E(A) = E(A)2 = E(A)∗ ;
Ex,x (A) = hE(A)x, xi = hE(A)2 x, xi = hE(A)x, E(A)xi = kE(A)xk2 ≥ 0.
Theorem 6.4.1 (Spectral Theorem). Let (T, D(T )) be a self-adjoint operator on a Hilbert
space H. Then there exists a uniquely determined spectral measure E such that
Z
hT x, yi = λ dEx,y (λ), ∀x ∈ D(T ), ∀y ∈ H,
R

and the support of the complex measure Ex,y is contained in the spectrum σ(T ). For any
bounded measurable function f there is a uniquely defined operator f (T ) defined by
Z
hf (T )x, yi = f (λ) dEx,y (λ),
R

such that the map B(R) 3 f 7→ f (T ) ∈ B(H) is a ∗-algebra homomorphism from the space
B(R) of bounded measurable functions to the space of bounded linear operators, i.e. (f g)(T ) =
f (T )g(T ), (af + bg)(T ) = af (T ) + bg(T ), f¯(T ) = f (T )∗ , for all f, g ∈ B(R), a, b ∈ C, and
where f¯(x) = f (x). Moreover, for a measurable real-valued function f : R → R define
Z
D = {x ∈ H | |f (λ)|2 dEx,x (λ) < ∞},
Z R
hf (T )x, yi = f (λ) dEx,y (λ),
R

then (f (T ), D) is a self-adjoint operator. Moreover, S ∈ B(H) commutes with T , ST ⊂ T S,

if and only if S commutes with all spectral projections E(A), A Borel set of R.
17
Félix Edouard Justin Émile Borel (7 January 1871 — 3 February 1956), French mathematician.
Chapter 6: Preliminaries and results from functional analysis 115

The map f 7→ f (T ) is known as the functional calculus for bounded measurable functions.
This can be extended to unbounded measurable functions. All integrals can be restricted to
the spectrum σ(T ), i.e. the spectral measure is supported on the spectrum.
The Spectral Theorem 6.4.1 in particular characterises the domain in terms of the spectral
measure, since Z
D(T ) = {x ∈ H | λ2 dEx,x (λ) < ∞}.
R
2
R 2
Note also that kf (T )xk = R |f (λ)| dEx,x (λ) ≥ 0 since Ex,x is a positive measure.
In particular, it follows from the Spectral Theorem 6.4.1, that for a self-adjoint (T, D(T ))
on the Hilbert space H and the function exp(−itx), t ∈ R, we get a bounded operator U (t) =
exp(−itT ). From the functional calculus it follows immediately that U (t + s) = U (t)U (s),
U (0) = 1, U (t)∗ = U (−t), so that t 7→ U (t) is a 1-parameter group of unitary operators in
B(H). Stone’s18 Theorem 6.4.2 states that the converse is also valid.

Theorem 6.4.2 (Stone). Assume R 3 t 7→ U (t) is a 1-parameter group of unitary operators,

i.e. U (s)U (t) = U (s + t) for all s, t ∈ R, and U (t) is a unitary operator in B(H) for each
t ∈ R. Assume moreover that R 3 t 7→ U (t)x ∈ H is continuous for all x ∈ H. Then there
exists a unique self-adjoint operator (T, D(T )) such that U (t) = exp(−itT ), and the operator
T is defined by

U (t)x − x U (t)x − x
D(T ) = {x ∈ H | lim converges} T x = i lim , x ∈ D(T ).
t→0 t t→0 t
U (t) as in Stone’s Theorem 6.4.2 is called a strongly continuous 1-parameter group of
unitary operators.

6.5 Spectrum and essential spectrum for self-adjoint

operators
In this subsection we recall some facts that are usually not parts of a standard course in
functional analysis. So we provide some of the statements of proofs.
We start with characterising the spectrum of a self-adjoint operator. Theorem 6.5.1 can
be extended to normal operators, i.e. T ∗ T = T T ∗ (also for unbounded operators T ).

Theorem 6.5.1. Let (T, D(T )) be a self-adjoint operator on a Hilbert space H, and λ ∈ R.
Then λ ∈ σ(T ) if and only if the following condition holds: there exists a sequence {xn }∞
n=1
in D(T ) such that

1. kxn k = 1,

2. lim (T − λ)xn = 0.
n→∞

18
Marshall Harvey Stone (8 April 1903 — 9 January 1989), American mathematician.
116 Chapter 6: Preliminaries and results from functional analysis

In particular, for λ ∈ σp (T ) we can take the sequence constant xn = x with x an eigenvector

of norm one for the eigenvalue λ. In general, for a closed operator (T, D(T )) we say that λ
is in the approximate point spectrum if there exists a sequence {xn }∞ n=1 in D(T ) such that
kxn k = 1 and limn→∞ (T − λ)xn = 0. Theorem 6.5.1 states that for a self-adjoint operator the
approximate point spectrum is its spectrum.

Proof. Let us first assume that such a sequence exists. We have to prove that λ ∈ σ(T ).
Suppose not, then λ is element of the resolvent, and so there exists R(λ; T ) ∈ B(H) such that
R(λ; T )(T − λ) ⊂ 1. So

1 = kxn k = kR(λ; T )(T − λ)xn k ≤ kR(λ; T )k k(T − λ)xn k

and since kR(λ; T )k is independent of n, the right hand side can be made arbitrarily small
since (T − λ)xn → 0. This is a contradiction, so that λ ∈ / ρ(T ), or λ ∈ σ(T ). Note that this
implication is independent of (T, D(T )) being self-adjoint, so we have proved that in general
the approximate point spectrum is part of the spectrum.
We prove the converse statement by negating it. So we assume such a sequence does not
exist, and we have to show that λ ∈ ρ(T ). First we claim that there exists a C > 0 such that

kxk ≤ C k(T − λ)xk ∀x ∈ D(T ). (6.5.1)

To see why (6.5.1) is true, we assume that there exists a sequence {yn }∞
n=1 in D(T ) such that

kyn k
→ ∞, n → ∞.
k(T − λ)yn k

Putting xn = yn /kyn k, noting that yn 6= 0, this gives kxn k = 1 and k(Tkx nk

−λ)xn k
→ ∞, so
k(T − λ)xn k → 0, which is precisely the statement whose negation we assume to be true.
As an immediate corollary to (6.5.1) we see that T − λ is injective. Indeed, assume
(T − λ)x1 = y = (T − λ)x2 , then 0 = ky − yk = Ck(T − λ)(x1 − x2 )k ≥ kx1 − x2 k ≥ 0 and
kx1 − x2 k = 0 or x1 = x2 .
Another consequence of (6.5.1) is Ran(T − λ) is closed. Indeed, take any sequence {yn }∞
n=1
in Ran(T − λ), and assume yn → y in H. We need to show that y ∈ Ran(T − λ). By the
previous observation we can pick unique xn ∈ D(T ) with (T − λ)xn = yn . Then (6.5.1) and
yn → y imply that {xn }∞n=1 is a Cauchy sequence, hence convergent to, say, x ∈ H. For any
z ∈ D(T ) we have

h(T − λ)z, xi = lim h(T − λ)z, xn i = lim hz, (T − λ)xn i = lim hz, yn i = hz, yi
n→∞ n→∞ n→∞

since xn ∈ D(T ∗ ) = D(T ) and λ ∈ R. This implies by definition that x ∈ D((T − λ)∗ ) =
D(T ∗ ) = D(T ) and that (T − λ)∗ x = y, or (T − λ)x = y, or y ∈ Ran(T − λ).
Summarising, T −λ is an injective map onto the closed subspace Ran(T −λ). We now show
that Ran(T − λ) is the whole Hilbert space H. Fix an element z ∈ H. We define a functional
Chapter 6: Preliminaries and results from functional analysis 117

φ : Ran(T − λ) → C by φ(y) = hx, zi with x ∈ D(T ) uniquely defined by (T − λ)x = y. Then

φ is indeed linear. Using (6.5.1) and the Cauchy-Schwarz inequality (6.1.1) we see

|φ(y)| = |hx, zi| ≤ kxk kzk ≤ C k(T − λ)xk kzk = C kyk kzk

or φ is a continuous linear functional on Ran(T −λ) as a closed subspace of H, so that the Riesz
representation theorem gives φ(y) = hy, wi for some w ∈ H. This is h(T − λ)x, wi = hx, zi
and since this is true for all x ∈ D(T ) we see that w ∈ D((T − λ)∗ ) = D(T ∗ ) = D(T ) and
z = (T − λ)∗ w = (T − λ)w since T is self-adjoint and λ ∈ R. This shows that z ∈ Ran(T − λ),
and since z ∈ H is arbitrary, it follows that Ran(T − λ) = H as claimed.
So now we have T − λ as an injective map from D(T ) to H = Ran(T − λ), and we can
define its inverse (T − λ)−1 as a map from H to D(T ). Now (6.5.1) implies that (T − λ)−1 is
bounded, hence (T − λ)−1 ∈ B(H) and so λ ∈ ρ(T ).
The Spectral Theorem 6.4.1 for the case of a compact operator is well-known, and we recall
it.
Theorem 6.5.2 (Spectral theorem for compact operators). Let K ∈ K(H) be self-adjoint.
Then its spectrum is a denumerable set in R with 0 as the only possible point of accumulation.
Each non-zero point in the spectrum is an eigenvalue of K, and the corresponding eigenspace is
finite dimensional. Denoting the spectrum by |λ1 | ≥ |λ2 | ≥ |λ3 | · · · (each λ ∈ σ(K) occurring
as many times as dim(Ker(K − λ))) and {ei }i≥1 , kei k = 1, the corresponding eigenvectors,
then M X
H = Ker(K) ⊕ C ei , Kx = λi hx, ei i ei .
i≥1 i≥1

The essential spectrum of a self-adjoint operator is the part of the spectrum that is not
influenced by compact perturbations. Suppose e.g. that λ ∈ σp (T ) with corresponding eigen-
vector e such that the self-adjoint operator x 7→ T x − λhx, eie has λ in its resolvent, then
we see that this part of the spectrum is influenced by a compact (in this case even rank-one)
perturbation. The definition of the essential spectrum is as follows. The essential spectrum
is also known as the Weyl spectrum.
Definition 6.5.3. Let (T, D(T )) be a self-adjoint operator on a Hilbert space H, then the
essential spectrum is \
σess (T ) = σ(T + K).
K∈K(H), K ∗ =K

Obviously, σess (T ) ⊂ σ(T ) by taking K equal to the zero operator. Also note that the
essential spectrum only makes sense for infinite dimensional Hilbert spaces. For H finite
dimensional the essential spectrum of any self-adjoint operator is empty.
Theorem 6.5.4. Let (T, D(T )) be a self-adjoint operator on a Hilbert space H, and take
λ ∈ R. Then the following statements are equivalent:
1. λ ∈ σess (T );
118 Chapter 6: Preliminaries and results from functional analysis

2. there exists a sequence {xn }∞

n=1 in the domain D(T ) such that

• kxn k = 1,
• {xn }∞
n=1 has no convergent subsequence,
• (T − λ)xn → 0;
3. there exists a sequence {xn }∞
n=1 in the domain D(T ) such that

• kxn k = 1,
• xn → 0 weakly,
• (T − λ)xn → 0.
Proof. (3) ⇒ (2): Assume that {xn }∞ n=1 has a convergent subsequence, we relabel and we can
assume that the sequence converges to x. Since kxn k = 1, we have kxk = 1. Since xn → x
we also have xn → x weakly, and since also xn → 0 weakly we must have x = 0 contradicting
kxk = 1.
(2) ⇒ (3): Since kxn k = 1 is a bounded sequence, it has a weakly convergent subsequence,
which, by relabeling, may be assumed to be the original sequence, see Exercise 6.1.1. Denote
its weak limit by x. Since by assumption there is no convergent subsequence, there exists
δ > 0 so that kxn − xk ≥ δ for all n ∈ N (by switching to a subsequence if necessary). Since
the sequence is from D(T ), the self-adjointness implies
h(T − λ)y, xn i = hy, (T − λ)xn i
for all y ∈ D(T ). Since (T − λ)xn → 0 and xn → x weakly we find by taking n → ∞ that
h(T − λ)y, xi = 0 for all y ∈ D(T ). Since this is obviously continuous as a functional in y, we
see that x ∈ D((T − λ)∗ ) = D(T ), since T is self-adjoint, and that (T − λ)∗ x = (T − λ)x = 0.
We now define zn = (xn − x)/kxn − xk, which can be done since kxn − xk ≥ δ > 0. Then
by construction kzn k = 1 and zn → 0 weakly. Moreover, since x ∈ Ker(T − λ),
1 1
k(T − λ)zn k = k(T − λ)xn k ≤ k(T − λ)xn k → 0.
kxn − xk δ
So the sequence {zn }∞
n=1 satisfies the conditions of (3).
(2) ⇒ (1): We prove the negation of this statement. Assume λ ∈
/ σess (T ), and let K be a
self-adjoint compact operator such that λ ∈ ρ(T + K). So in particular R(λ; T + K) : H →
D(T + K) = D(T ) is bounded, or
kR(λ; T + K)xk ≤ kR(λ; T + K)k kxk
=⇒ kyk ≤ kR(λ; T + K)k k(T + K − λ)yk
by switching x to (T +K −λ)y, y ∈ D(T ) arbitrary. Now take any sequence {xn }∞n=1 satisfying
the first and last condition of (2). Then by the above observation, with C = kR(λ; T + K)k,
kxn − xm k ≤ C k(T + K − λ)(xn − xm )k
≤ C k(T − λ)xn k + C k(T − λ)xm k + C kK(xn − xm )k
Chapter 6: Preliminaries and results from functional analysis 119

and the first two terms on the right hand side tend to zero by assumption. Since K is compact,
and the sequence {x}∞ ∞
n=1 is bounded by 1 there is a convergent subsequence of {Kxn }n=1 . By
∞
relabeling we may assume that the sequence {Kxn }n=1 is convergent, and then the final term
on the right hand side tends to zero. This means that {xn }∞ n=1 is a Cauchy sequence, hence
convergent. So the three conditions in (2) cannot hold.
(1) ⇒ (2) We consider two possibilities; dim Ker(T − λ) < ∞ or dim Ker(T − λ) = ∞.
In the last case we pick an orthonormal sequence {xn }∞ n=1 in Ker(T − λ) ⊂ D(T ). Then
obviously, kxn k = 1 and (T − λ)x√n = 0 for all n, and this sequence cannot have a convergent
subsequence, since kxn − xm k = 2 by the Pythagorean theorem, cf. Section 6.1.
In case Ker(T − λ) is finite-dimensional, we claim that there exists a sequence {xn }∞ n=1
in D(T ) such that xn ⊥ Ker(T − λ), kxn k = 1 and (T − λ)xn → 0. In order to prove the
claim we first observe that D(T ) ∩ Ker(T − λ)⊥ is dense in Ker(T − λ)⊥ . Indeed, let P denote
the orthogonal projection on Ker(T − λ) and take arbitrary x ∈ Ker(T − λ)⊥ , ε > 0, so that
∃y ∈ D(T ) such that kx − yk < ε. Write y = P y + (1 − P )y, then P y ∈ Ker(T − λ) ⊂
D(T ) so that with y ∈ D(T ) and D(T ) being a linear space, also (1 − P )y ∈ D(T ). Now
kx − (1 − P )yk = k(1 − P )(x − y)k ≤ kx − yk < ε, so the density follows.
In order to see why the claim in the previous paragraph is true, note that P is self-adjoint
and, since this is a finite rank operator, P is compact. Note that T +P is a self-adjoint operator.
If such a sequence would not exist, then we can conclude, by restricting to Ker(T − λ)⊥ , as in
the proof of Theorem 6.5.1, cf. (6.5.1), that there exists a C such that for all x ⊥ Ker(T − λ),
x ∈ D(T ) we have kxk ≤ Ck(T − λ)xk. For x ∈ D(T ) ∩ Ker(T − λ)⊥ we have

h(T − λ)x, P xi = hx, (T − λ)P xi = 0,

so that, using this orthogonality in the last equality,

kxk2 = k(1 − P )xk2 + kP xk2 ≤ C 2 k(T − λ)xk2 + kP xk2

≤ max(C 2 , 1) k(T − λ)xk2 + kP xk2 = max(C 2 , 1) k(T + P − λ)xk2 .

This implies that T + P − λ has a bounded inverse, so that λ ∈ ρ(T + P ) or λ ∈ / σess (T ).

Now the claim can be used to finish the proof as follows. We have to show that {xn }∞ n=1
does not have a convergent subsequence. Suppose, that it does have a convergent subsequence,
which we denote by {xn }∞
n=1 as well, and say that xn → x. Then kxk = 1, and x ∈ Ker(T −λ) .
⊥

On the other hand, (T −λ)xn → 0, and since T is self-adjoint, hence closed, we see that T −λ is
closed. This means that x ∈ D(T ) and x ∈ Ker(T −λ). So x ∈ Ker(T −λ)∩Ker(T −λ)⊥ = {0},
which contradicts kxk = 1.
The essential spectrum can be obtained from the spectrum by “throwing out the point
spectrum”. Below we mean by an isolated point λ of a subset σ of R that there does not
exists a sequence {λn }∞
n=1 of points in σ such that λn → λ in R.

Theorem 6.5.5. Let (T, D(T )) be a self-adjoint operator on a Hilbert space H. Then we have

1. If λ ∈ σ(T )\σess (T ), then λ is an isolated point in σ(T ).

120 Chapter 6: Preliminaries and results from functional analysis

2. If λ ∈ σ(T )\σess (T ), then λ ∈ σp (T ).

Proof. To prove the first statement assume that λ is not isolated in σ(T ), so there exists a
sequence {λn }∞
n=1 , λn 6= λ, λn ∈ σ(T ) and λn → λ. By invoking Theorem 6.5.1 we can find
for each n ∈ N an element xn ∈ H such that kxn k = 1 and k(T − λn )xn k < n1 |λ − λn | for
n ∈ N. Then
1
k(T − λ)xn k ≤ k(T − λn )xn k + |λ − λn | kxk < (1 + ) |λ − λn | → 0, n → ∞,
n
In order to show that λ ∈ σess (T ), we can use Theorem 6.5.4. It suffices to show that the
constructed sequence {xn }∞
n=1 has no convergent subsequence. So suppose it does have a
convergent subsequence, again denoted by {xn }∞ n=1 , say xn → x. Then kxk = 1, and by
closedness of T − λ, x ∈ D(T ) and x ∈ Ker(T − λ). On the other hand,

|(λ − λn )| |hxn , xi| = |hxn , λxi − λn hxn , xi| = |hxn , T xi − λn hxn , xi|
1
= |h(T − λn )xn , xi| ≤ |λ − λn |
n
and since λ 6= λn we see that hxn , xi → 0. But by assumption hxn , xi → kxk2 = 1, so this
give the required contradiction.
For the second statement we use Theorem 6.5.1 to get a sequence {xn }∞ n=1 such that
kxn k = 1 and (T − λ)xn → 0. Since λ ∈ / σess (T ), Theorem 6.5.4 implies that this sequence
must have a convergent subsequence, again denoted {xn }∞ n=1 with xn → x. Since T is self-
adjoint, T − λ is closed and hence x ∈ D(T ) and (T − λ)x = 0. Now kxk = 1 shows that
Ker(T − λ) is non-trivial, and hence λ ∈ σp (T ).

6.6 Absolutely continuous and singular spectrum

Let λ denote Lebesgue measure restricted to the σ-algebra of the Borel sets on R. An arbitrary
measure in this Section is a finite complex measure. Recall that a measure µ is absolutely
continuous (with respect to Lebesgue measure λ) if for all Borel sets B with λ(B) = 0 we have
µ(B) = 0, and the measure µ is singular (with respect to Lebesgue measure λ) if there exists
a Borel set Z such that λ(Z) = 0 and µ(Z c ) = 0, where Z c = R\Z is its complement in R. So
the singular measure is supported on Z. The Lebesgue Decomposition Theorem states that
any measure µ can be uniquely decomposed as µ = µac + µs with µac absolutely continuous
and µs singular. Knowing the support Z of the singular measure, the Lebesgue decomposition
is given by µs (B) = µ(B ∩ Z), µac (B) = µ(B ∩ Z c ) for any Borel set B. Also, the sum
of two absolutely continuous, respectively singular, measures is again absolutely continuous,
respectively singular.
Recall that for a self-adjoint operator (T, D(T )) the Spectral Theorem 6.4.1 gives a finite
complex measure Ex,y for any pair x, y ∈ H by Ex,y (B) = hE(B)x, yi, where E is the spectral
measure, and in case x = y we have that Ex,x is a positive measure.
Chapter 6: Preliminaries and results from functional analysis 121

Definition 6.6.1. Given a self-adjoint operator (T, D(T )) we define

Hac = Hac (T ) = {x ∈ H | Ex,x is absolutely continuous},

Hs = Hs (T ) = {x ∈ H | Ex,x is singular}.

Hac is the absolutely continuous subspace and Hs is the singular subspace.

We call x ∈ Hac an absolutely continuous element, and x ∈ Hs a singular element of H.

Theorem 6.6.2. Consider a self-adjoint operator (T, D(T )) on H, then

1. H = Hac ⊕ Hs , so in particular Hac and Hs are closed subspaces,

2. Hac and Hs reduce T .

Recall the definition of reducing spaces as in Definition 3.2.10 for the second statement.
Using Theorem 6.6.2(2) we can define  the absolutely continuous spectrum of a self-adjoint
operator (T, D(T )) as σac (T ) = σ(T Hac ) and the singular spectrum σs (T ) = σ(T Hs ). Using
Theorem 6.6.2 it follows that σ(T ) = σac (T ) ∪ σs (T ).

Proof. First observe that, using hE(B)x, yi = hE(B)x, E(B)yi since E(B) is self-adjoint
projection, we have
|hE(B)x, yi|2 ≤ hE(B)x, xi hE(B)y, yi (6.6.1)
by the Cauchy-Schwarz inequality (6.1.1). Now (6.6.1) shows that for x ∈ Hac and any Borel
set B with λ(B) = 0 we also have hE(B)x, yi = 0, or Ex,y is absolutely continuous. If
x ∈ Hs we have a Borel set B such that λ(B) = 0 and Ex,x (B c ) = 0, and (6.6.1) implies that
Ex,y (B c ) = 0 as well, so Ex,y is singular as well.
Observe that Ecx,cx = |c|2 Ex,x for any c ∈ C, so that with x also cx are in Hac , respectively
Hc . Next let x, y ∈ Hac , then for any Borel set B

Ex+y,x+y (B) = hE(B)(x + y), x + yi = hE(B)x, xi + hE(B)y, yi + hE(B)x, yi + hE(B)y, xi

and the first two measures are absolutely continuous by assumption, and the last two are
absolutely continuous by the reasoning in the previous paragraph. So x + y ∈ Hac . Similarly,
Hs is a linear space.
To see that Hac ⊥ Hs take x ∈ Hac , y ∈ Hs and consider the measure Ex,y . This measure is
absolutely continuous, since x ∈ Hac , and singular, since y ∈ Hs , so it has to be the measure
identically equal to zero. So hE(B)x, yi = 0 for all Borel sets B, and taking B = R and
recalling E(R) = 1 in B(H), we see that hx, yi = 0.
To finish the proof of the first statement we show that any element, say z ∈ H, can be
written as z = x + y with x ∈ Hac and y ∈ Hs . Use the Lebesgue decomposition theorem to
write Ez,z = µac + µs , and let Z be the Borel set such that λ(Z) = 0 and µs (Z c ) = 0. Put
x = E(Z c )z, y = E(Z)z, and it remains to show that these elements are the required ones.
122 Chapter 6: Preliminaries and results from functional analysis

First, x + y = E(Z)z + E(Z c )z = E(R)z = z by additivity of the spectral measure. Next for
an arbitrary Borel set B

Ex,x (B) = hE(B)x, xi = hE(B)E(Z c )z, E(Z c )zi

= hE(Z c )E(B)E(Z c )z, zi = hE(B ∩ Z c )z, zi = Ez,z (B ∩ Z c ),

so that Ex,x = µac and hence x ∈ Hac . Similarly, y ∈ Hs .

To prove the second statement, we take z ∈ D(T ) ⊂ H and write z = x + y, x ∈ Hac ,
y ∈ Hs . We claim that x, y ∈ D(T ). To see this, recall that Ex,x , Ey,y and Ez,z are positive
measures satisfying Ez,z = Ex,x + Ey,y , so that
Z Z
2
λ dEx,x (λ) ≤ λ2 dEz,z < ∞,
R R

implying that x ∈ D(T ) using the Spectral Theorem 6.4.1. Similarly, y ∈ D(T ). For x ∈
D(T ) ∩ Hac , y ∈ H, we have, again by the Spectral Theorem 6.4.1,
Z
ET x,y (B) = λ dEx,y , (6.6.2)
B

so that ET x,y is absolutely continuous with respect to Ex,y , in fact its Radon-Nikodym deriva-
tive is λ. Since Ex,y is absolutely continuous, it follows that ET x,T x –take y = T x– is absolutely
continuous, hence T x ∈ Hac . We can reason in a similar way for the singular subspace Hs , or
we can use the first statement and T being self-adjoint to see that T : D(T ) ∩ Hs → Hs .
In case T has an eigenvector x for the eigenvalue λ, then Ex,x is a measure with Ex,x ({λ}) >
0 and Ex,x (R\{λ}) = 0. So in particular, all eigenspaces are contained in Hs . Define Hpp =
Hpp (T ) as the closed linear span of all eigenvectors of T , so that Hpp ⊂ Hs . Again, Hpp
reduces T . We denote by Hcs = Hcs (T ) the orthocomplement of Hpp in Hs , then Hcs also
reduces T . So we get the decomposition

H = Hpp ⊕ Hcs ⊕ Hac .

Here ‘pp’ stands for ‘pure point’ and ‘cs’ for ‘continuous
 singular’. Since these
 spaces reduce
T , we have corresponding spectra, σpp (T ) = σ(T Hpp ) and σsc (T ) = σ(T Hsc ). In general,

σpp (T ) is not equal to σp (T ), but σpp (T ) = σp (T ).
Bibliography

[1] O. Babelon, D. Bernard, M. Talon, Introduction to Classical Integrable Systems, Cam-

bridge Univ. Press, 2003.

[2] F. Calogero and A. Degasperis, Spectral Transform and Solitons I, North Holland, 1982.

[3] P. Deift and E. Trubowitz, Inverse scattering on the line, Commun. Pure and Applied
Math. 32 (1979), 121–251.

[4] W. Eckhaus and A. van Harten, The Inverse Scattering Transformation and the Theory
of Solitons, North Holland, 1981.

[5] C.S. Gardner, J.M. Greene, M.D. Kruskal, R.M. Miura, Method for solving the Korteweg-
de Vries equation, Physical Review Letters 19 (1967), 1095–1097.

[6] E.M. de Jager, Mathematical introduction to the theory of solitons, pp. 355–617 in Math-
ematical Structures in Continuous Dynamical Systems (E. van Groesen, E.M. de Jager
eds.), Nort Holland, 1994.

[7] P.D. Lax, Functional Analysis, Wiley, 2002.

[8] T. Miwa, M. Jimbo and E. Date, Solitons: Differential Equations, Symmetries and Infi-
nite Dimensional Algebras, Cambridge Univ. Press, 2000.

[9] M. Reed and B. Simon, Methods of Modern Mathematical Physics III: Scattering Theory,
Academic Press, 1979.

[10] M. Schechter, Operator Methods in Quantum Mechanics, North Holland, 1981 (reprinted
by Dover).

[11] D. Werner, Funktionalanalysis, 4. überarbeitete Auflage, Springer, 2002.

123
124
Index

absolutely continuous measure, 120 direct scattering problem, 65

d2
absolutely continuous spectrum, 121 − dx 2, 5

absolutely continuous subspace, 121

approximate point spectrum, 116 eigenvalue, 113
Arzelà-Ascoli theorem, 16, 106 eigenvector, 113
asymptotic state, incoming, 28 essential spectrum, 12, 117
asymptotic state, outgoing, 29 essentially self-adjoint, 108
extension, 106
Banach space, 103
Bessel differential equation, 24 Fourier inversion, 110
Bessel function, 24 Fourier transform, 110
Borel set, 114 functional, 104
bound state, 16
bounded self-adjoint operator, 107 Galilean invariance, 83
Burgers equation, 83 Gelfand-Levitan-Marchenko equation, 71, 72
generalised wave operator, 38
Cc∞ (R), 111 graph norm, 10, 13, 108
C k (R), 111 Gronwall’s Lemma, 52
closable operator, 107
closed graph theorem, 108 Hölder’s inequality, 105
closed operator, 107 Hardy class, 54, 111
closure, 108 Heaviside function, 66
cnoidal waves, 88 Hilbert space, 104
commutator, 93 Hopf-Cole transformation, 83
compact operator, 107 hypergeometric differential equation, 21
compact relative to a closed operator, 12 hypergeometric function, 22
complete, 39
completeness, strong asymptotic, 39 incoming asymptotic state, 28
completeness, weak asymptotic, 39 inner product, 103
converse Hölder’s inequality, 8, 105 inverse scattering method, 90
Cook’s method, 35 inverse scattering problem, 70, 72
cosh(x), 20 isolated point, 119
C(R), 106 isometry, 107
isospectral, 91
densely defined, 8
densely defined operator, 106 Jost solution, 47, 48

125
126 Index

KdV-equation, 83 potential, reflectionless, 59

Korteweg-de Vries equation, 83 pure 1-soliton solution, 91, 98
pure 2-soliton solution, 99
Lax pair, 93
Lebesgue decomposition theorem, 120 reduce, 32
locally square integrable, 8, 14 reducing subspace, 32
Lp (R), 105 reflection coefficient, 59
L2 (R), 105 reflectionless potential, 59, 77
Rellich’s perturbation theorem, 8
Miura transformation, 84 resolvent, 113
mKdV-equation, 84 Riemann-Lebesgue lemma, 66, 110
modified KdV-equation, 84 Riesz representation theorem, 104
multiplication operator, 8
σ-algebra, 113
norm, 103 scattering data, 65
operator, 106 scattering matrix, 59
operator norm, 106 scattering operator, 31
operator, − dxd2 scattering problem, direct, 65
2, 5

operator, closable, 107 scattering problem, inverse, 70

operator, closed, 107 scattering state, 29
operator, compact, 107 Schrödinger integral equation, 47
operator, compact relative to a closed opera- Schrödinger operator, 8
tor, 12 self-adjoint operator, 107
operator, densely defined, 106 self-adjoint Schrödinger operator, 10
operator, multiplication, 8 selfadjoint, essentially, 108
operator, Schrödinger, 8 separable, 106
operator, smooth relative to a self-adjoint op- singular measure, 120
erator, 40 singular spectrum, 121
orthogonal projection, 107 singular subspace, 121
outgoing asymptotic state, 29 sinh(x), 20
Sobolev imbedding lemma, 11, 15, 111
Painlevé equation, 88 Sobolev space, 110
Paley-Wiener theorem, 111 soliton solution, 91, 98, 99
Parseval’s identity, 110 spectral theorem, 114
partial isometry, 107 spectral theorem for compact operators, 117
phase shift, 99 spectrum, 113
Plancherel’s identity, 110 spectrum of self-adjoint operator, 115
point spectrum, 113 spectrum, essential, 12, 117
polar decomposition, 42, 108 square integrable at ∞, 23
positive operator, 110 Stone’s theorem, 115
potential, 8 strong asymptotic completeness, 39
−2
potential, cosh , 20 strong operator topology, 107
potential, indicator function, 18 symmetric operator, 107
Index 127

T -bound, 9
T -compact operator, 12
time-dependent Schrödinger equation, 27
transmission coefficient, 59
T -smooth operator, 40

uniformly continuous, 16, 106

unitary operator, 107

wave operator, 29
wave operator, generalised, 38
weak asymptotic completeness, 39
weak compactness, 104
weak convergence, 104
weak derivative, 111
weak operator topology, 107
Weyl spectrum, 117
Wronskian, 24, 25, 57